agora inbox for [email protected]
help / color / mirror / Atom feedTypo in src/backend/access/hash/README
34+ messages / 9 participants
[nested] [flat]
* Typo in src/backend/access/hash/README
@ 2018-03-05 03:30 Thomas Munro <[email protected]>
0 siblings, 1 reply; 34+ messages in thread
From: Thomas Munro @ 2018-03-05 03:30 UTC (permalink / raw)
To: pgsql-hackers
Hi,
Not sure what word was missed here, but I guess "count":
our the number of buckets stored in our cached copy of the metapage. If
-so, the bucket has certainly been split, because the must originally
+so, the bucket has certainly been split, because the count must originally
have been less than the number of buckets that existed at that time and
--
Thomas Munro
http://www.enterprisedb.com
Attachments:
[application/octet-stream] somebody-accidentally-a-word.patch (873B, ../../CAEepm=1_682z-09DNHj4GkCJAqWK-D6h9Oq5ea84T1oqq1-Utg@mail.gmail.com/2-somebody-accidentally-a-word.patch)
download | inline diff:
diff --git a/src/backend/access/hash/README b/src/backend/access/hash/README
index bb90722778f..fa5a9c38455 100644
--- a/src/backend/access/hash/README
+++ b/src/backend/access/hash/README
@@ -190,7 +190,7 @@ After computing the ostensibly-correct bucket number based on our cached
copy of the metapage, we lock the corresponding primary bucket page and
check whether the bucket count stored in hasho_prevblkno is greater than
our the number of buckets stored in our cached copy of the metapage. If
-so, the bucket has certainly been split, because the must originally
+so, the bucket has certainly been split, because the count must originally
have been less than the number of buckets that existed at that time and
can't have increased except due to a split. If not, the bucket can't have
been split, because a split would have created a new bucket with a higher
^ permalink raw reply [nested|flat] 34+ messages in thread
* Re: Typo in src/backend/access/hash/README
@ 2018-03-05 03:40 Justin Pryzby <[email protected]>
parent: Thomas Munro <[email protected]>
0 siblings, 1 reply; 34+ messages in thread
From: Justin Pryzby @ 2018-03-05 03:40 UTC (permalink / raw)
To: Thomas Munro <[email protected]>; +Cc: pgsql-hackers; Robert Haas <[email protected]>
On Mon, Mar 05, 2018 at 04:30:28PM +1300, Thomas Munro wrote:
> Hi,
>
> Not sure what word was missed here, but I guess "count":
>
> our the number of buckets stored in our cached copy of the metapage. If
> -so, the bucket has certainly been split, because the must originally
> +so, the bucket has certainly been split, because the count must originally
> have been less than the number of buckets that existed at that time and
I think there's a 2nd typo:
|After computing the ostensibly-correct bucket number based on our cached
|copy of the metapage, we lock the corresponding primary bucket page and
|check whether the bucket count stored in hasho_prevblkno is greater than
|OUR the number of buckets stored in our cached copy of the metapage. If
remove "our"?
> --94eb2c0d4c02fe3e960566a1f43e
> Content-Type: application/octet-stream; name="somebody-accidentally-a-word.patch"
:)
Justin
^ permalink raw reply [nested|flat] 34+ messages in thread
* Re: Typo in src/backend/access/hash/README
@ 2018-03-05 06:53 Amit Kapila <[email protected]>
parent: Justin Pryzby <[email protected]>
0 siblings, 1 reply; 34+ messages in thread
From: Amit Kapila @ 2018-03-05 06:53 UTC (permalink / raw)
To: Justin Pryzby <[email protected]>; +Cc: Thomas Munro <[email protected]>; pgsql-hackers; Robert Haas <[email protected]>
On Mon, Mar 5, 2018 at 9:10 AM, Justin Pryzby <[email protected]> wrote:
> On Mon, Mar 05, 2018 at 04:30:28PM +1300, Thomas Munro wrote:
>> Hi,
>>
>> Not sure what word was missed here, but I guess "count":
>>
>> our the number of buckets stored in our cached copy of the metapage. If
>> -so, the bucket has certainly been split, because the must originally
>> +so, the bucket has certainly been split, because the count must originally
>> have been less than the number of buckets that existed at that time and
>
> I think there's a 2nd typo:
>
> |After computing the ostensibly-correct bucket number based on our cached
> |copy of the metapage, we lock the corresponding primary bucket page and
> |check whether the bucket count stored in hasho_prevblkno is greater than
> |OUR the number of buckets stored in our cached copy of the metapage. If
>
> remove "our"?
>
Both your and Thomas's proposed change looks good to me.
--
With Regards,
Amit Kapila.
EnterpriseDB: http://www.enterprisedb.com
^ permalink raw reply [nested|flat] 34+ messages in thread
* Re: Typo in src/backend/access/hash/README
@ 2018-03-29 18:38 Bruce Momjian <[email protected]>
parent: Amit Kapila <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Bruce Momjian @ 2018-03-29 18:38 UTC (permalink / raw)
To: Amit Kapila <[email protected]>; +Cc: Justin Pryzby <[email protected]>; Thomas Munro <[email protected]>; pgsql-hackers; Robert Haas <[email protected]>
On Mon, Mar 5, 2018 at 12:23:31PM +0530, Amit Kapila wrote:
> On Mon, Mar 5, 2018 at 9:10 AM, Justin Pryzby <[email protected]> wrote:
> > On Mon, Mar 05, 2018 at 04:30:28PM +1300, Thomas Munro wrote:
> >> Hi,
> >>
> >> Not sure what word was missed here, but I guess "count":
> >>
> >> our the number of buckets stored in our cached copy of the metapage. If
> >> -so, the bucket has certainly been split, because the must originally
> >> +so, the bucket has certainly been split, because the count must originally
> >> have been less than the number of buckets that existed at that time and
> >
> > I think there's a 2nd typo:
> >
> > |After computing the ostensibly-correct bucket number based on our cached
> > |copy of the metapage, we lock the corresponding primary bucket page and
> > |check whether the bucket count stored in hasho_prevblkno is greater than
> > |OUR the number of buckets stored in our cached copy of the metapage. If
> >
> > remove "our"?
> >
>
> Both your and Thomas's proposed change looks good to me.
Fixed in the attached, applied patch.
--
Bruce Momjian <[email protected]> http://momjian.us
EnterpriseDB http://enterprisedb.com
+ As you are, so once was I. As I am, so you will be. +
+ Ancient Roman grave inscription +
Attachments:
[text/x-diff] readme.diff (1.4K, ../../[email protected]/2-readme.diff)
download | inline diff:
diff --git a/src/backend/access/hash/README b/src/backend/access/hash/README
new file mode 100644
index bb90722..21b4a82
*** a/src/backend/access/hash/README
--- b/src/backend/access/hash/README
*************** reality, InvalidBlockNumber.
*** 189,196 ****
After computing the ostensibly-correct bucket number based on our cached
copy of the metapage, we lock the corresponding primary bucket page and
check whether the bucket count stored in hasho_prevblkno is greater than
! our the number of buckets stored in our cached copy of the metapage. If
! so, the bucket has certainly been split, because the must originally
have been less than the number of buckets that existed at that time and
can't have increased except due to a split. If not, the bucket can't have
been split, because a split would have created a new bucket with a higher
--- 189,196 ----
After computing the ostensibly-correct bucket number based on our cached
copy of the metapage, we lock the corresponding primary bucket page and
check whether the bucket count stored in hasho_prevblkno is greater than
! the number of buckets stored in our cached copy of the metapage. If
! so, the bucket has certainly been split, because the count must originally
have been less than the number of buckets that existed at that time and
can't have increased except due to a split. If not, the bucket can't have
been split, because a split would have created a new bucket with a higher
^ permalink raw reply [nested|flat] 34+ messages in thread
* refactoring relation extension and BufferAlloc(), faster COPY
@ 2022-10-29 02:54 Andres Freund <[email protected]>
0 siblings, 1 reply; 34+ messages in thread
From: Andres Freund @ 2022-10-29 02:54 UTC (permalink / raw)
To: pgsql-hackers; Thomas Munro <[email protected]>; Melanie Plageman <[email protected]>; +Cc: Yura Sokolov <[email protected]>; Robert Haas <[email protected]>
Hi,
I'm working to extract independently useful bits from my AIO work, to reduce
the size of that patchset. This is one of those pieces.
In workloads that extend relations a lot, we end up being extremely contended
on the relation extension lock. We've attempted to address that to some degree
by using batching, which helps, but only so much.
The fundamental issue, in my opinion, is that we do *way* too much while
holding the relation extension lock. We acquire a victim buffer, if that
buffer is dirty, we potentially flush the WAL, then write out that
buffer. Then we zero out the buffer contents. Call smgrextend().
Most of that work does not actually need to happen while holding the relation
extension lock. As far as I can tell, the minimum that needs to be covered by
the extension lock is the following:
1) call smgrnblocks()
2) insert buffer[s] into the buffer mapping table at the location returned by
smgrnblocks
3) mark buffer[s] as IO_IN_PROGRESS
1) obviously has to happen with the relation extension lock held because
otherwise we might miss another relation extension. 2+3) need to happen with
the lock held, because otherwise another backend not doing an extension could
read the block before we're done extending, dirty it, write it out, and then
have it overwritten by the extending backend.
The reason we currently do so much work while holding the relation extension
lock is that bufmgr.c does not know about the relation lock and that relation
extension happens entirely within ReadBuffer* - there's no way to use a
narrower scope for the lock.
My fix for that is to add a dedicated function for extending relations, that
can acquire the extension lock if necessary (callers can tell it to skip that,
e.g., when initially creating an init fork). This routine is called by
ReadBuffer_common() when P_NEW is passed in, to provide backward
compatibility.
To be able to acquire victim buffers outside of the extension lock, victim
buffers are now acquired separately from inserting the new buffer mapping
entry. Victim buffer are pinned, cleaned, removed from the buffer mapping
table and marked invalid. Because they are pinned, clock sweeps in other
backends won't return them. This is done in a new function,
[Local]BufferAlloc().
This is similar to Yuri's patch at [0], but not that similar to earlier or
later approaches in that thread. I don't really understand why that thread
went on to ever more complicated approaches, when the basic approach shows
plenty gains, with no issues around the number of buffer mapping entries that
can exist etc.
Other interesting bits I found:
a) For workloads that [mostly] fit into s_b, the smgwrite() that BufferAlloc()
does, nearly doubles the amount of writes. First the kernel ends up writing
out all the zeroed out buffers after a while, then when we write out the
actual buffer contents.
The best fix for that seems to be to optionally use posix_fallocate() to
reserve space, without dirtying pages in the kernel page cache. However, it
looks like that's only beneficial when extending by multiple pages at once,
because it ends up causing one filesystem-journal entry for each extension
on at least some filesystems.
I added 'smgrzeroextend()' that can extend by multiple blocks, without the
caller providing a buffer to write out. When extending by 8 or more blocks,
posix_fallocate() is used if available, otherwise pg_pwritev_with_retry() is
used to extend the file.
b) I found that is quite beneficial to bulk-extend the relation with
smgrextend() even without concurrency. The reason for that is the primarily
the aforementioned dirty buffers that our current extension method causes.
One bit that stumped me for quite a while is to know how much to extend the
relation by. RelationGetBufferForTuple() drives the decision whether / how
much to bulk extend purely on the contention on the extension lock, which
obviously does not work for non-concurrent workloads.
After quite a while I figured out that we actually have good information on
how much to extend by, at least for COPY /
heap_multi_insert(). heap_multi_insert() can compute how much space is
needed to store all tuples, and pass that on to
RelationGetBufferForTuple().
For that to be accurate we need to recompute that number whenever we use an
already partially filled page. That's not great, but doesn't appear to be a
measurable overhead.
c) Contention on the FSM and the pages returned by it is a serious bottleneck
after a) and b).
The biggest issue is that the current bulk insertion logic in hio.c enters
all but one of the new pages into the freespacemap. That will immediately
cause all the other backends to contend on the first few pages returned the
FSM, and cause contention on the FSM pages itself.
I've, partially, addressed that by using the information about the required
number of pages from b). Whether we bulk insert or not, the number of pages
we know we're going to need for one heap_multi_insert() don't need to be
added to the FSM - we're going to use them anyway.
I've stashed the number of free blocks in the BulkInsertState for now, but
I'm not convinced that that's the right place.
If I revert just this part, the "concurrent COPY into unlogged table"
benchmark goes from ~240 tps to ~190 tps.
Even after that change the FSM is a major bottleneck. Below I included
benchmarks showing this by just removing the use of the FSM, but I haven't
done anything further about it. The contention seems to be both from
updating the FSM, as well as thundering-herd like symptoms from accessing
the FSM.
The update part could likely be addressed to some degree with a batch
update operation updating the state for multiple pages.
The access part could perhaps be addressed by adding an operation that gets
a page and immediately marks it as fully used, so other backends won't also
try to access it.
d) doing
/* new buffers are zero-filled */
MemSet((char *) bufBlock, 0, BLCKSZ);
under the extension lock is surprisingly expensive on my two socket
workstation (but much less noticable on my laptop).
If I move the MemSet back under the extension lock, the "concurrent COPY
into unlogged table" benchmark goes from ~240 tps to ~200 tps.
e) When running a few benchmarks for this email, I noticed that there was a
sharp performance dropoff for the patched code for a pgbench -S -s100 on a
database with 1GB s_b, start between 512 and 1024 clients. This started with
the patch only acquiring one buffer partition lock at a time. Lots of
debugging ensued, resulting in [3].
The problem isn't actually related to the change, it just makes it more
visible, because the "lock chains" between two partitions reduce the
average length of the wait queues substantially, by distribution them
between more partitions. [3] has a reproducer that's entirely independent
of this patchset.
Bulk extension acquires a number of victim buffers, acquires the extension
lock, inserts the buffers into the buffer mapping table and marks them as
io-in-progress, calls smgrextend and releases the extension lock. After that
buffer[s] are locked (depending on mode and an argument indicating the number
of blocks to be locked), and TerminateBufferIO() is called.
This requires two new pieces of infrastructure:
First, pinning multiple buffers opens up the obvious danger that we might run
of non-pinned buffers. I added LimitAdditional[Local]Pins() that allows each
backend to pin a proportional share of buffers (although always allowing one,
as we do today).
Second, having multiple IOs in progress at the same time isn't possible with
the InProgressBuf mechanism. I added a ResourceOwnerRememberBufferIO() etc to
deal with that instead. I like that this ends up removing a lot of
AbortBufferIO() calls from the loops of various aux processes (now released
inside ReleaseAuxProcessResources()).
In very extreme workloads (single backend doing a pgbench -S -s 100 against a
s_b=64MB database) the memory allocations triggered by StartBufferIO() are
*just about* visible, not sure if that's worth worrying about - we do such
allocations for the much more common pinning of buffers as well.
The new [Bulk]ExtendSharedRelationBuffered() currently have both a Relation
and a SMgrRelation argument, requiring at least one of them to be set. The
reason for that is on the one hand that LockRelationForExtension() requires a
relation and on the other hand, redo routines typically don't have a Relation
around (recovery doesn't require an extension lock). That's not pretty, but
seems a tad better than the ReadBufferExtended() vs
ReadBufferWithoutRelcache() mess.
I've done a fair bit of benchmarking of this patchset. For COPY it comes out
ahead everywhere. It's possible that there's a very small regression for
extremly IO miss heavy workloads, more below.
server "base" configuration:
max_wal_size=150GB
shared_buffers=24GB
huge_pages=on
autovacuum=0
backend_flush_after=2MB
max_connections=5000
wal_buffers=128MB
wal_segment_size=1GB
benchmark: pgbench running COPY into a single table. pgbench -t is set
according to the client count, so that the same amount of data is inserted.
This is done oth using small files ([1], ringbuffer not effective, no dirty
data to write out within the benchmark window) and a bit larger files ([2],
lots of data to write out due to ringbuffer).
To make it a fair comparison HEAD includes the lwlock-waitqueue fix as well.
s_b=24GB
test: unlogged_small_files, format: text, files: 1024, 9015MB total
seconds tbl-MBs seconds tbl-MBs seconds tbl-MBs
clients HEAD HEAD patch patch no_fsm no_fsm
1 58.63 207 50.22 242 54.35 224
2 32.67 372 25.82 472 27.30 446
4 22.53 540 13.30 916 14.33 851
8 15.14 804 7.43 1640 7.48 1632
16 14.69 829 4.79 2544 4.50 2718
32 15.28 797 4.41 2763 3.32 3710
64 15.34 794 5.22 2334 3.06 4061
128 15.49 786 4.97 2452 3.13 3926
256 15.85 768 5.02 2427 3.26 3769
512 16.02 760 5.29 2303 3.54 3471
test: logged_small_files, format: text, files: 1024, 9018MB total
seconds tbl-MBs seconds tbl-MBs seconds tbl-MBs
clients HEAD HEAD patch patch no_fsm no_fsm
1 68.18 178 59.41 205 63.43 192
2 39.71 306 33.10 368 34.99 348
4 27.26 446 19.75 617 20.09 607
8 18.84 646 12.86 947 12.68 962
16 15.96 763 9.62 1266 8.51 1436
32 15.43 789 8.20 1486 7.77 1579
64 16.11 756 8.91 1367 8.90 1383
128 16.41 742 10.00 1218 9.74 1269
256 17.33 702 11.91 1023 10.89 1136
512 18.46 659 14.07 866 11.82 1049
test: unlogged_medium_files, format: text, files: 64, 9018MB total
seconds tbl-MBs seconds tbl-MBs seconds tbl-MBs
clients HEAD HEAD patch patch no_fsm no_fsm
1 63.27s 192 56.14 217 59.25 205
2 40.17s 303 29.88 407 31.50 386
4 27.57s 442 16.16 754 17.18 709
8 21.26s 573 11.89 1025 11.09 1099
16 21.25s 573 10.68 1141 10.22 1192
32 21.00s 580 10.72 1136 10.35 1178
64 20.64s 590 10.15 1200 9.76 1249
128 skipped
256 skipped
512 skipped
test: logged_medium_files, format: text, files: 64, 9018MB total
seconds tbl-MBs seconds tbl-MBs seconds tbl-MBs
clients HEAD HEAD patch patch no_fsm no_fsm
1 71.89s 169 65.57 217 69.09 69.09
2 47.36s 257 36.22 407 38.71 38.71
4 33.10s 368 21.76 754 22.78 22.78
8 26.62s 457 15.89 1025 15.30 15.30
16 24.89s 489 17.08 1141 15.20 15.20
32 25.15s 484 17.41 1136 16.14 16.14
64 26.11s 466 17.89 1200 16.76 16.76
128 skipped
256 skipped
512 skipped
Just to see how far it can be pushed, with binary format we can now get to
nearly 6GB/s into a table when disabling the FSM - note the 2x difference
between patch and patch+no-fsm at 32 clients.
test: unlogged_small_files, format: binary, files: 1024, 9508MB total
seconds tbl-MBs seconds tbl-MBs seconds tbl-MBs
clients HEAD HEAD patch patch no_fsm no_fsm
1 34.14 357 28.04 434 29.46 413
2 22.67 537 14.42 845 14.75 826
4 16.63 732 7.62 1599 7.69 1587
8 13.48 904 4.36 2795 4.13 2959
16 14.37 848 3.78 3224 2.74 4493
32 14.79 823 4.20 2902 2.07 5974
64 14.76 825 5.03 2423 2.21 5561
128 14.95 815 4.36 2796 2.30 5343
256 15.18 802 4.31 2828 2.49 4935
512 15.41 790 4.59 2656 2.84 4327
s_b=4GB
test: unlogged_small_files, format: text, files: 1024, 9018MB total
seconds tbl-MBs seconds tbl-MBs
clients HEAD HEAD patch patch
1 62.55 194 54.22 224
2 37.11 328 28.94 421
4 25.97 469 16.42 742
8 20.01 609 11.92 1022
16 19.55 623 11.05 1102
32 19.34 630 11.27 1081
64 19.07 639 12.04 1012
128 19.22 634 12.27 993
256 19.34 630 12.28 992
512 19.60 621 11.74 1038
test: logged_small_files, format: text, files: 1024, 9018MB total
seconds tbl-MBs seconds tbl-MBs
clients HEAD HEAD patch patch
1 71.71 169 63.63 191
2 46.93 259 36.31 335
4 30.37 401 22.41 543
8 22.86 533 16.90 721
16 20.18 604 14.07 866
32 19.64 620 13.06 933
64 19.71 618 15.08 808
128 19.95 610 15.47 787
256 20.48 595 16.53 737
512 21.56 565 16.86 722
test: unlogged_medium_files, format: text, files: 64, 9018MB total
seconds tbl-MBs seconds tbl-MBs
clients HEAD HEAD patch patch
1 62.65 194 55.74 218
2 40.25 302 29.45 413
4 27.37 445 16.26 749
8 22.07 552 11.75 1037
16 21.29 572 10.64 1145
32 20.98 580 10.70 1139
64 20.65 590 10.21 1193
128 skipped
256 skipped
512 skipped
test: logged_medium_files, format: text, files: 64, 9018MB total
seconds tbl-MBs seconds tbl-MBs
clients HEAD HEAD patch patch
1 71.72 169 65.12 187
2 46.46 262 35.74 341
4 32.61 373 21.60 564
8 26.69 456 16.30 747
16 25.31 481 17.00 716
32 24.96 488 17.47 697
64 26.05 467 17.90 680
128 skipped
256 skipped
512 skipped
test: unlogged_small_files, format: binary, files: 1024, 9505MB total
seconds tbl-MBs seconds tbl-MBs
clients HEAD HEAD patch patch
1 37.62 323 32.77 371
2 28.35 429 18.89 645
4 20.87 583 12.18 1000
8 19.37 629 10.38 1173
16 19.41 627 10.36 1176
32 18.62 654 11.04 1103
64 18.33 664 11.89 1024
128 18.41 661 11.91 1023
256 18.52 658 12.10 1007
512 18.78 648 11.49 1060
benchmark: Run a pgbench -S workload with scale 100, so it doesn't fit into
s_b, thereby exercising BufferAlloc()'s buffer replacement path heavily.
The run-to-run variance on my workstation is high for this workload (both
before/after my changes). I also found that the ramp-up time at higher client
counts is very significant:
progress: 2.1 s, 5816.8 tps, lat 1.835 ms stddev 4.450, 0 failed
progress: 3.0 s, 666729.4 tps, lat 5.755 ms stddev 16.753, 0 failed
progress: 4.0 s, 899260.1 tps, lat 3.619 ms stddev 41.108, 0 failed
...
One would need to run pgbench for impractically long to make that effect
vanish.
My not great solution for these was to run with -T21 -P5 and use the best 5s
as the tps.
s_b=1GB
tps tps
clients master patched
1 49541 48805
2 85342 90010
4 167340 168918
8 308194 303222
16 524294 523678
32 649516 649100
64 932547 937702
128 908249 906281
256 856496 903979
512 764254 934702
1024 653886 925113
2048 569695 917262
4096 526782 903258
s_b=128MB:
tps tps
clients master patched
1 40407 39854
2 73180 72252
4 143334 140860
8 240982 245331
16 429265 420810
32 544593 540127
64 706408 726678
128 713142 718087
256 611030 695582
512 552751 686290
1024 508248 666370
2048 474108 656735
4096 448582 633040
As there might be a small regression at the smallest end, I ran a more extreme
version of the above. Using a pipelined pgbench -S, with a single client, for
longer. With s_b=8MB.
To further reduce noise I pinned the server to one cpu, the client to another
and disabled turbo mode on the CPU.
master "total" tps: 61.52
master "best 5s" tps: 61.8
patch "total" tps: 61.20
patch "best 5s" tps: 61.4
Hardly conclusive, but it does look like there's a small effect. It could be
code layout or such.
My guess however is that it's the resource owner for in-progress IO that I
added - that adds an additional allocation inside the resowner machinery. I
commented those out (that's obviously incorrect!) just to see whether that
changes anything:
no-resowner "total" tps: 62.03
no-resowner "best 5s" tps: 62.2
So it looks like indeed, it's the resowner. I am a bit surprised, because
obviously we already use that mechanism for pins, which obviously is more
frequent.
I'm not sure it's worth worrying about - this is a pretty absurd workload. But
if we decide it is, I can think of a few ways to address this. E.g.:
- We could preallocate an initial element inside the ResourceArray struct, so
that a newly created resowner won't need to allocate immediately
- We could only use resowners if there's more than one IO in progress at the
same time - but I don't like that idea much
- We could try to store the "in-progress"-ness of a buffer inside the 'bufferpin'
resowner entry - on 64bit system there's plenty space for that. But on 32bit systems...
The patches here aren't fully polished (as will be evident). But they should
be more than good enough to discuss whether this is a sane direction.
Greetings,
Andres Freund
[0] https://postgr.es/m/3b108afd19fa52ed20c464a69f64d545e4a14772.camel%40postgrespro.ru
[1] COPY (SELECT repeat(random()::text, 5) FROM generate_series(1, 100000)) TO '/tmp/copytest_data_text.copy' WITH (FORMAT test);
[2] COPY (SELECT repeat(random()::text, 5) FROM generate_series(1, 6*100000)) TO '/tmp/copytest_data_text.copy' WITH (FORMAT text);
[3] https://postgr.es/m/[email protected]
^ permalink raw reply [nested|flat] 34+ messages in thread
* Re: refactoring relation extension and BufferAlloc(), faster COPY
@ 2022-10-29 13:03 Bharath Rupireddy <[email protected]>
parent: Andres Freund <[email protected]>
0 siblings, 1 reply; 34+ messages in thread
From: Bharath Rupireddy @ 2022-10-29 13:03 UTC (permalink / raw)
To: Andres Freund <[email protected]>; +Cc: pgsql-hackers; Thomas Munro <[email protected]>; Melanie Plageman <[email protected]>; Yura Sokolov <[email protected]>; Robert Haas <[email protected]>
On Sat, Oct 29, 2022 at 8:24 AM Andres Freund <[email protected]> wrote:
>
> Hi,
>
> I'm working to extract independently useful bits from my AIO work, to reduce
> the size of that patchset. This is one of those pieces.
Thanks a lot for this great work. There are 12 patches in this thread,
I believe each of these patches is trying to solve separate problems
and can be reviewed and get committed separately, am I correct?
> In workloads that extend relations a lot, we end up being extremely contended
> on the relation extension lock. We've attempted to address that to some degree
> by using batching, which helps, but only so much.
Yes, I too have observed this in the past for parallel inserts in CTAS
work - https://www.postgresql.org/message-id/CALj2ACW9BUoFqWkmTSeHjFD-W7_00s3orqSvtvUk%2BKD2H7ZmRg%40mail.g....
Tackling bulk relation extension problems will unblock the parallel
inserts (in CTAS, COPY) work I believe.
> The fundamental issue, in my opinion, is that we do *way* too much while
> holding the relation extension lock. We acquire a victim buffer, if that
> buffer is dirty, we potentially flush the WAL, then write out that
> buffer. Then we zero out the buffer contents. Call smgrextend().
>
> Most of that work does not actually need to happen while holding the relation
> extension lock. As far as I can tell, the minimum that needs to be covered by
> the extension lock is the following:
>
> 1) call smgrnblocks()
> 2) insert buffer[s] into the buffer mapping table at the location returned by
> smgrnblocks
> 3) mark buffer[s] as IO_IN_PROGRESS
Makes sense.
I will try to understand and review each patch separately.
Firstly, 0001 avoids extra loop over waiters and looks a reasonable
change, some comments on the patch:
1)
+ int lwWaiting; /* 0 if not waiting, 1 if on
waitlist, 2 if
+ * waiting to be woken */
Use macros instead of hard-coded values for better readability?
#define PROC_LW_LOCK_NOT_WAITING 0
#define PROC_LW_LOCK_ON_WAITLIST 1
#define PROC_LW_LOCK_WAITING_TO_BE_WOKEN 2
2) Missing initialization of lwWaiting to 0 or the macro in twophase.c
and proc.c.
proc->lwWaiting = false;
MyProc->lwWaiting = false;
3)
+ proclist_delete(&lock->waiters, MyProc->pgprocno, lwWaitLink);
+ found = true;
I guess 'found' is a bit meaningless here as we are doing away with
the proclist_foreach_modify loop. We can directly use
MyProc->lwWaiting == 1 and remove 'found'.
4)
if (!MyProc->lwWaiting)
if (!proc->lwWaiting)
Can we modify the above conditions in lwlock.c to MyProc->lwWaiting !=
1 or PROC_LW_LOCK_ON_WAITLIST or the macro?
5) Is there any specific test case that I can see benefit of this
patch? If yes, can you please share it here?
--
Bharath Rupireddy
PostgreSQL Contributors Team
RDS Open Source Databases
Amazon Web Services: https://aws.amazon.com
^ permalink raw reply [nested|flat] 34+ messages in thread
* Re: refactoring relation extension and BufferAlloc(), faster COPY
@ 2022-10-29 18:39 Andres Freund <[email protected]>
parent: Bharath Rupireddy <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Andres Freund @ 2022-10-29 18:39 UTC (permalink / raw)
To: Bharath Rupireddy <[email protected]>; +Cc: pgsql-hackers; Thomas Munro <[email protected]>; Melanie Plageman <[email protected]>; Yura Sokolov <[email protected]>; Robert Haas <[email protected]>
Hi,
On 2022-10-29 18:33:53 +0530, Bharath Rupireddy wrote:
> On Sat, Oct 29, 2022 at 8:24 AM Andres Freund <[email protected]> wrote:
> >
> > Hi,
> >
> > I'm working to extract independently useful bits from my AIO work, to reduce
> > the size of that patchset. This is one of those pieces.
>
> Thanks a lot for this great work. There are 12 patches in this thread,
> I believe each of these patches is trying to solve separate problems
> and can be reviewed and get committed separately, am I correct?
Mostly, yes.
For 0001 I already started
https://www.postgresql.org/message-id/20221027165914.2hofzp4cvutj6gin%40awork3.anarazel.de
to discuss the specific issue.
We don't strictly need v1-0002-aio-Add-some-error-checking-around-pinning.patch
but I did find it useful.
v1-0012-bufmgr-debug-Add-PrintBuffer-Desc.patch is not used in the patch
series, but I found it quite useful when debugging issues with the patch. A
heck of a lot easier to interpret page flags when they can be printed.
I also think there's some architectural questions that'll influence the number
of patches. E.g. I'm not convinced
v1-0010-heapam-Add-num_pages-to-RelationGetBufferForTuple.patch is quite the
right spot to track which additional pages should be used. It could very well
instead be alongside ->smgr_targblock. Possibly the best path would instead
be to return the additional pages explicitly to callers of
RelationGetBufferForTuple, but RelationGetBufferForTuple does a bunch of work
around pinning that potentially would need to be repeated in heap_multi_insert().
> > In workloads that extend relations a lot, we end up being extremely contended
> > on the relation extension lock. We've attempted to address that to some degree
> > by using batching, which helps, but only so much.
>
> Yes, I too have observed this in the past for parallel inserts in CTAS
> work - https://www.postgresql.org/message-id/CALj2ACW9BUoFqWkmTSeHjFD-W7_00s3orqSvtvUk%2BKD2H7ZmRg%40mail.g....
> Tackling bulk relation extension problems will unblock the parallel
> inserts (in CTAS, COPY) work I believe.
Yea. There's a lot of places the current approach ended up being a bottleneck.
> Firstly, 0001 avoids extra loop over waiters and looks a reasonable
> change, some comments on the patch:
> 1)
> + int lwWaiting; /* 0 if not waiting, 1 if on
> waitlist, 2 if
> + * waiting to be woken */
> Use macros instead of hard-coded values for better readability?
>
> #define PROC_LW_LOCK_NOT_WAITING 0
> #define PROC_LW_LOCK_ON_WAITLIST 1
> #define PROC_LW_LOCK_WAITING_TO_BE_WOKEN 2
Yea - this was really more of a prototype patch - I noted that we'd want to
use defines for this in
https://www.postgresql.org/message-id/20221027165914.2hofzp4cvutj6gin%40awork3.anarazel.de
> 3)
> + proclist_delete(&lock->waiters, MyProc->pgprocno, lwWaitLink);
> + found = true;
> I guess 'found' is a bit meaningless here as we are doing away with
> the proclist_foreach_modify loop. We can directly use
> MyProc->lwWaiting == 1 and remove 'found'.
We can rename it, but I think we still do need it, it's easier to analyze the
logic if the relevant check happens on a value from while we held the wait
list lock. Probably should do the reset inside the locked section as well.
> 4)
> if (!MyProc->lwWaiting)
> if (!proc->lwWaiting)
> Can we modify the above conditions in lwlock.c to MyProc->lwWaiting !=
> 1 or PROC_LW_LOCK_ON_WAITLIST or the macro?
I think it's better to check it's 0, rather than just != 1.
> 5) Is there any specific test case that I can see benefit of this
> patch? If yes, can you please share it here?
Yep, see the other thread, there's a pretty easy case there. You can also see
it at extreme client counts with a pgbench -S against a cluster with a smaller
shared_buffers. But the difference is not huge before something like 2048-4096
clients, and then it only occurs occasionally (because you need to end up with
most connections waiting for one of the partitions). So the test case from the
other thread is a lot better.
Greetings,
Andres Freund
^ permalink raw reply [nested|flat] 34+ messages in thread
* Re: Set arbitrary GUC options during initdb
@ 2023-01-27 20:05 Robert Haas <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Robert Haas @ 2023-01-27 20:05 UTC (permalink / raw)
To: Tom Lane <[email protected]>; +Cc: David G. Johnston <[email protected]>; Peter Eisentraut <[email protected]>; [email protected]
On Fri, Jan 27, 2023 at 3:02 PM Tom Lane <[email protected]> wrote:
> The string-hacking was fully as tedious as I expected. However, the
> output looks pretty nice, and this does have the advantage that the
> pre-programmed substitutions become a lot more robust: they are no
> longer dependent on the initdb code exactly matching what is in
> postgresql.conf.sample.
Awesome! Thank you very much.
--
Robert Haas
EDB: http://www.enterprisedb.com
^ permalink raw reply [nested|flat] 34+ messages in thread
* [PATCH v45 7/8] Error out any process that would block at REPACK
@ 2026-03-25 19:35 Álvaro Herrera <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Álvaro Herrera @ 2026-03-25 19:35 UTC (permalink / raw)
Any process waiting on REPACK to release its lock would actually cause
it to deadlock when it tries to upgrade its lock to AEL, losing all work
done to that point. We avoid this by teaching the deadlock detector to
raise an error when this condition is detected.
---
src/backend/commands/repack.c | 53 ++++++++----
src/backend/storage/lmgr/deadlock.c | 15 ++++
src/include/storage/proc.h | 6 +-
src/test/modules/injection_points/Makefile | 1 +
.../expected/repack_deadlock.out | 63 ++++++++++++++
src/test/modules/injection_points/meson.build | 1 +
.../specs/repack_deadlock.spec | 83 +++++++++++++++++++
7 files changed, 202 insertions(+), 20 deletions(-)
create mode 100644 src/test/modules/injection_points/expected/repack_deadlock.out
create mode 100644 src/test/modules/injection_points/specs/repack_deadlock.spec
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index d25ab43da7b..7291b4180ab 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -276,6 +276,16 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
/* Determine the lock mode to use. */
lockmode = RepackLockLevel((params.options & CLUOPT_CONCURRENT) != 0);
+ /*
+ * If in concurrent mode, set the PROC_IN_CONCURRENT_REPACK flag. This
+ * makes the deadlock checker cause anyone that would conflict with us
+ * to error out. It's important to set this flag ahead of actually locking
+ * the relation; it won't of course affect anyone until we do have a lock
+ * that others can conflict with.
+ */
+ if ((params.options & CLUOPT_CONCURRENT) != 0)
+ MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+
/*
* If a single relation is specified, process it and we're done ... unless
* the relation is a partitioned table, in which case we fall through.
@@ -476,11 +486,8 @@ RepackLockLevel(bool concurrent)
* If indexOid is InvalidOid, the table will be rewritten in physical order
* instead of index order.
*
- * Note that, in the concurrent case, the function releases the lock at some
- * point, in order to get AccessExclusiveLock for the final steps (i.e. to
- * swap the relation files). To make things simpler, the caller should expect
- * OldHeap to be closed on return, regardless CLUOPT_CONCURRENT. (The
- * AccessExclusiveLock is kept till the end of the transaction.)
+ * On return, OldHeap is closed but locked with AccessExclusiveLock - the lock
+ * will be released at end of the transaction.
*
* 'cmd' indicates which command is being executed, to be used for error
* messages.
@@ -512,10 +519,12 @@ cluster_rel(RepackCommand cmd, Relation OldHeap, Oid indexOid,
/*
* Make sure we're not in a transaction block.
*
- * The reason is that repack_setup_logical_decoding() could deadlock
- * if there's an XID already assigned. It would be possible to run in
- * a transaction block if we had no XID, but this restriction is
- * simpler for users to understand and we don't lose anything.
+ * The reason is that repack_setup_logical_decoding() could wait
+ * indefinitely for our XID to complete. (The deadlock detector would
+ * not recognize it because we'd be waiting for ourselves, i.e. no
+ * real lock conflict.) It would be possible to run in a transaction
+ * block if we had no XID, but this restriction is simpler for users
+ * to understand and we don't lose anything.
*/
PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
@@ -998,10 +1007,8 @@ rebuild_relation(Relation OldHeap, Relation index, bool verbose,
* Note that the worker has to wait for all transactions with XID
* already assigned to finish. If some of those transactions is
* waiting for a lock conflicting with ShareUpdateExclusiveLock on our
- * table (e.g. it runs CREATE INDEX), we can end up in a deadlock.
- * Not sure this risk is worth unlocking/locking the table (and its
- * clustering index) and checking again if it's still eligible for
- * REPACK CONCURRENTLY.
+ * table (e.g. it runs CREATE INDEX), it should encounter ERROR in the
+ * deadlock checking code.
*/
start_repack_decoding_worker(tableOid);
@@ -3119,10 +3126,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
/*
- * Tuples and pages of the old heap will be gone, but the heap will stay.
+ * Now that we have all access-exclusive locks on all relations, we no
+ * longer want other processes to error out when trying to acquire a
+ * conflicting lock. Therefore, unset our flag.
+ */
+ MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+
+ /*
+ * Tuples and pages of the old heap will be gone, but the heap itself will
+ * stay. In order for predicate locks to continue to work, convert them
+ * to relation-level locks. We do this both for table and indexes.
*/
TransferPredicateLocksToHeapRelation(OldHeap);
- /* The same for indexes. */
for (int i = 0; i < nind; i++)
{
Relation index = ind_refs[i];
@@ -3342,9 +3357,11 @@ start_repack_decoding_worker(Oid relid)
/*
* The decoding setup must be done before the caller can have XID assigned
- * for any reason, otherwise the worker might end up in a deadlock,
- * waiting for the caller's transaction to end. Therefore wait here until
- * the worker indicates that it has the logical decoding initialized.
+ * for any reason, otherwise the worker might end up waiting for the
+ * caller's transaction to end. (Deadlock detector does not consider this
+ * a conflict because the worker is in the same locking group as the
+ * backend that launched it.) Therefore wait here until the worker
+ * indicates that it has the logical decoding initialized.
*/
ConditionVariablePrepareToSleep(&shared->cv);
for (;;)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3d3ec0da111 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -620,6 +620,21 @@ FindLockCycleRecurseMember(PGPROC *checkProc,
proc->statusFlags & PROC_IS_AUTOVACUUM)
blocking_autovacuum_proc = proc;
+ /*
+ * Similarly, if we note that we're blocked by some process
+ * running REPACK (CONCURRENTLY), just fail. That process
+ * is going to upgrade its lock at some point, and it would
+ * be inappropriate for any other process to cause that
+ * to fail.
+ */
+ if (checkProc == MyProc &&
+ proc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+ ereport(ERROR,
+ errcode(ERRCODE_OBJECT_IN_USE),
+ errmsg("could not wait for concurrent REPACK"),
+ errdetail("Process %d waits for REPACK running on process %d",
+ MyProc->pid, proc->pid));
+
/* We're done looking at this proclock */
break;
}
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 1dad125706e..8ad9718f3d6 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -69,10 +69,12 @@ struct XidCache
#define PROC_AFFECTS_ALL_HORIZONS 0x20 /* this proc's xmin must be
* included in vacuum horizons
* in all databases */
+#define PROC_IN_CONCURRENT_REPACK 0x40 /* REPACK (CONCURRENTLY) */
-/* flags reset at EOXact */
+/* flags reset at EOXact. A bit of a misnomer ... */
#define PROC_VACUUM_STATE_MASK \
- (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+ (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+ PROC_IN_CONCURRENT_REPACK)
/*
* Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/Makefile b/src/test/modules/injection_points/Makefile
index 2cd7d87c533..f7663859fe2 100644
--- a/src/test/modules/injection_points/Makefile
+++ b/src/test/modules/injection_points/Makefile
@@ -15,6 +15,7 @@ REGRESS_OPTS = --dlpath=$(top_builddir)/src/test/regress
ISOLATION = basic \
inplace \
repack \
+ repack_deadlock \
repack_toast \
syscache-update-pruned \
heap_lock_update
diff --git a/src/test/modules/injection_points/expected/repack_deadlock.out b/src/test/modules/injection_points/expected/repack_deadlock.out
new file mode 100644
index 00000000000..a86e4767536
--- /dev/null
+++ b/src/test/modules/injection_points/expected/repack_deadlock.out
@@ -0,0 +1,63 @@
+Parsed test spec with 2 sessions
+
+starting permutation: wait_before_lock add_column wakeup_before_lock check1
+injection_points_attach
+-----------------------
+
+(1 row)
+
+step wait_before_lock:
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+ <waiting ...>
+step add_column:
+ alter table repack_deadlock add column noise text;
+ <waiting ...>
+step add_column: <... completed>
+ERROR: could not wait for concurrent REPACK
+step wakeup_before_lock:
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1:
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+
+count
+-----
+ 1
+(1 row)
+
+i|j
+-+-
+1|1
+2|2
+3|3
+4|4
+(4 rows)
+
+count
+-----
+ 4
+(1 row)
+
+injection_points_detach
+-----------------------
+
+(1 row)
+
diff --git a/src/test/modules/injection_points/meson.build b/src/test/modules/injection_points/meson.build
index a414abb924b..1cd88d6db65 100644
--- a/src/test/modules/injection_points/meson.build
+++ b/src/test/modules/injection_points/meson.build
@@ -46,6 +46,7 @@ tests += {
'basic',
'inplace',
'repack',
+ 'repack_deadlock',
'repack_toast',
'syscache-update-pruned',
'heap_lock_update',
diff --git a/src/test/modules/injection_points/specs/repack_deadlock.spec b/src/test/modules/injection_points/specs/repack_deadlock.spec
new file mode 100644
index 00000000000..9d23a6588c2
--- /dev/null
+++ b/src/test/modules/injection_points/specs/repack_deadlock.spec
@@ -0,0 +1,83 @@
+# Test REPACK with a concurrent transaction that would cause a deadlock
+setup
+{
+ CREATE EXTENSION injection_points;
+
+ CREATE TABLE repack_deadlock(i int PRIMARY KEY, j int);
+ INSERT INTO repack_deadlock(i, j) VALUES (1, 1), (2, 2), (3, 3), (4, 4);
+
+ CREATE TABLE relfilenodes(node oid);
+
+ CREATE TABLE data_s1(i int, j int);
+ CREATE TABLE data_s2(i int, j int);
+}
+
+teardown
+{
+ DROP TABLE repack_deadlock;
+ DROP EXTENSION injection_points;
+
+ DROP TABLE relfilenodes;
+ DROP TABLE data_s1;
+ DROP TABLE data_s2;
+}
+
+session s1
+setup
+{
+ SELECT injection_points_set_local();
+ SELECT injection_points_attach('repack-concurrently-before-lock', 'wait');
+}
+# Perform the initial load and wait for s2 to do some data changes.
+step wait_before_lock
+{
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+}
+# Check the table from the perspective of s1.
+#
+# Besides the contents, we also check that relfilenode has changed.
+
+# Have each session write the contents into a table and use FULL JOIN to check
+# if the outputs are identical.
+step check1
+{
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+}
+teardown
+{
+ SELECT injection_points_detach('repack-concurrently-before-lock');
+}
+
+session s2
+# Change the existing data. UPDATE changes both key and non-key columns. Also
+# update one row twice to test whether tuple version generated by this session
+# can be found.
+step add_column
+{
+ alter table repack_deadlock add column noise text;
+}
+
+step wakeup_before_lock
+{
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+}
+
+# Test if data changes introduced while one session is performing REPACK
+# CONCURRENTLY find their way into the table.
+permutation
+ wait_before_lock
+ add_column
+ wakeup_before_lock
+ check1
--
2.47.3
--cldir6umqisr673d
Content-Type: text/x-diff; charset=utf-8
Content-Disposition: attachment;
filename="v45-0008-Teach-snapshot-builder-to-skip-transactions-runn.patch"
^ permalink raw reply [nested|flat] 34+ messages in thread
* [PATCH v45 7/8] Error out any process that would block at REPACK
@ 2026-03-25 19:35 Álvaro Herrera <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Álvaro Herrera @ 2026-03-25 19:35 UTC (permalink / raw)
Any process waiting on REPACK to release its lock would actually cause
it to deadlock when it tries to upgrade its lock to AEL, losing all work
done to that point. We avoid this by teaching the deadlock detector to
raise an error when this condition is detected.
---
src/backend/commands/repack.c | 53 ++++++++----
src/backend/storage/lmgr/deadlock.c | 15 ++++
src/include/storage/proc.h | 6 +-
src/test/modules/injection_points/Makefile | 1 +
.../expected/repack_deadlock.out | 63 ++++++++++++++
src/test/modules/injection_points/meson.build | 1 +
.../specs/repack_deadlock.spec | 83 +++++++++++++++++++
7 files changed, 202 insertions(+), 20 deletions(-)
create mode 100644 src/test/modules/injection_points/expected/repack_deadlock.out
create mode 100644 src/test/modules/injection_points/specs/repack_deadlock.spec
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index d25ab43da7b..7291b4180ab 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -276,6 +276,16 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
/* Determine the lock mode to use. */
lockmode = RepackLockLevel((params.options & CLUOPT_CONCURRENT) != 0);
+ /*
+ * If in concurrent mode, set the PROC_IN_CONCURRENT_REPACK flag. This
+ * makes the deadlock checker cause anyone that would conflict with us
+ * to error out. It's important to set this flag ahead of actually locking
+ * the relation; it won't of course affect anyone until we do have a lock
+ * that others can conflict with.
+ */
+ if ((params.options & CLUOPT_CONCURRENT) != 0)
+ MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+
/*
* If a single relation is specified, process it and we're done ... unless
* the relation is a partitioned table, in which case we fall through.
@@ -476,11 +486,8 @@ RepackLockLevel(bool concurrent)
* If indexOid is InvalidOid, the table will be rewritten in physical order
* instead of index order.
*
- * Note that, in the concurrent case, the function releases the lock at some
- * point, in order to get AccessExclusiveLock for the final steps (i.e. to
- * swap the relation files). To make things simpler, the caller should expect
- * OldHeap to be closed on return, regardless CLUOPT_CONCURRENT. (The
- * AccessExclusiveLock is kept till the end of the transaction.)
+ * On return, OldHeap is closed but locked with AccessExclusiveLock - the lock
+ * will be released at end of the transaction.
*
* 'cmd' indicates which command is being executed, to be used for error
* messages.
@@ -512,10 +519,12 @@ cluster_rel(RepackCommand cmd, Relation OldHeap, Oid indexOid,
/*
* Make sure we're not in a transaction block.
*
- * The reason is that repack_setup_logical_decoding() could deadlock
- * if there's an XID already assigned. It would be possible to run in
- * a transaction block if we had no XID, but this restriction is
- * simpler for users to understand and we don't lose anything.
+ * The reason is that repack_setup_logical_decoding() could wait
+ * indefinitely for our XID to complete. (The deadlock detector would
+ * not recognize it because we'd be waiting for ourselves, i.e. no
+ * real lock conflict.) It would be possible to run in a transaction
+ * block if we had no XID, but this restriction is simpler for users
+ * to understand and we don't lose anything.
*/
PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
@@ -998,10 +1007,8 @@ rebuild_relation(Relation OldHeap, Relation index, bool verbose,
* Note that the worker has to wait for all transactions with XID
* already assigned to finish. If some of those transactions is
* waiting for a lock conflicting with ShareUpdateExclusiveLock on our
- * table (e.g. it runs CREATE INDEX), we can end up in a deadlock.
- * Not sure this risk is worth unlocking/locking the table (and its
- * clustering index) and checking again if it's still eligible for
- * REPACK CONCURRENTLY.
+ * table (e.g. it runs CREATE INDEX), it should encounter ERROR in the
+ * deadlock checking code.
*/
start_repack_decoding_worker(tableOid);
@@ -3119,10 +3126,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
/*
- * Tuples and pages of the old heap will be gone, but the heap will stay.
+ * Now that we have all access-exclusive locks on all relations, we no
+ * longer want other processes to error out when trying to acquire a
+ * conflicting lock. Therefore, unset our flag.
+ */
+ MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+
+ /*
+ * Tuples and pages of the old heap will be gone, but the heap itself will
+ * stay. In order for predicate locks to continue to work, convert them
+ * to relation-level locks. We do this both for table and indexes.
*/
TransferPredicateLocksToHeapRelation(OldHeap);
- /* The same for indexes. */
for (int i = 0; i < nind; i++)
{
Relation index = ind_refs[i];
@@ -3342,9 +3357,11 @@ start_repack_decoding_worker(Oid relid)
/*
* The decoding setup must be done before the caller can have XID assigned
- * for any reason, otherwise the worker might end up in a deadlock,
- * waiting for the caller's transaction to end. Therefore wait here until
- * the worker indicates that it has the logical decoding initialized.
+ * for any reason, otherwise the worker might end up waiting for the
+ * caller's transaction to end. (Deadlock detector does not consider this
+ * a conflict because the worker is in the same locking group as the
+ * backend that launched it.) Therefore wait here until the worker
+ * indicates that it has the logical decoding initialized.
*/
ConditionVariablePrepareToSleep(&shared->cv);
for (;;)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3d3ec0da111 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -620,6 +620,21 @@ FindLockCycleRecurseMember(PGPROC *checkProc,
proc->statusFlags & PROC_IS_AUTOVACUUM)
blocking_autovacuum_proc = proc;
+ /*
+ * Similarly, if we note that we're blocked by some process
+ * running REPACK (CONCURRENTLY), just fail. That process
+ * is going to upgrade its lock at some point, and it would
+ * be inappropriate for any other process to cause that
+ * to fail.
+ */
+ if (checkProc == MyProc &&
+ proc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+ ereport(ERROR,
+ errcode(ERRCODE_OBJECT_IN_USE),
+ errmsg("could not wait for concurrent REPACK"),
+ errdetail("Process %d waits for REPACK running on process %d",
+ MyProc->pid, proc->pid));
+
/* We're done looking at this proclock */
break;
}
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 1dad125706e..8ad9718f3d6 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -69,10 +69,12 @@ struct XidCache
#define PROC_AFFECTS_ALL_HORIZONS 0x20 /* this proc's xmin must be
* included in vacuum horizons
* in all databases */
+#define PROC_IN_CONCURRENT_REPACK 0x40 /* REPACK (CONCURRENTLY) */
-/* flags reset at EOXact */
+/* flags reset at EOXact. A bit of a misnomer ... */
#define PROC_VACUUM_STATE_MASK \
- (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+ (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+ PROC_IN_CONCURRENT_REPACK)
/*
* Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/Makefile b/src/test/modules/injection_points/Makefile
index 2cd7d87c533..f7663859fe2 100644
--- a/src/test/modules/injection_points/Makefile
+++ b/src/test/modules/injection_points/Makefile
@@ -15,6 +15,7 @@ REGRESS_OPTS = --dlpath=$(top_builddir)/src/test/regress
ISOLATION = basic \
inplace \
repack \
+ repack_deadlock \
repack_toast \
syscache-update-pruned \
heap_lock_update
diff --git a/src/test/modules/injection_points/expected/repack_deadlock.out b/src/test/modules/injection_points/expected/repack_deadlock.out
new file mode 100644
index 00000000000..a86e4767536
--- /dev/null
+++ b/src/test/modules/injection_points/expected/repack_deadlock.out
@@ -0,0 +1,63 @@
+Parsed test spec with 2 sessions
+
+starting permutation: wait_before_lock add_column wakeup_before_lock check1
+injection_points_attach
+-----------------------
+
+(1 row)
+
+step wait_before_lock:
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+ <waiting ...>
+step add_column:
+ alter table repack_deadlock add column noise text;
+ <waiting ...>
+step add_column: <... completed>
+ERROR: could not wait for concurrent REPACK
+step wakeup_before_lock:
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1:
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+
+count
+-----
+ 1
+(1 row)
+
+i|j
+-+-
+1|1
+2|2
+3|3
+4|4
+(4 rows)
+
+count
+-----
+ 4
+(1 row)
+
+injection_points_detach
+-----------------------
+
+(1 row)
+
diff --git a/src/test/modules/injection_points/meson.build b/src/test/modules/injection_points/meson.build
index a414abb924b..1cd88d6db65 100644
--- a/src/test/modules/injection_points/meson.build
+++ b/src/test/modules/injection_points/meson.build
@@ -46,6 +46,7 @@ tests += {
'basic',
'inplace',
'repack',
+ 'repack_deadlock',
'repack_toast',
'syscache-update-pruned',
'heap_lock_update',
diff --git a/src/test/modules/injection_points/specs/repack_deadlock.spec b/src/test/modules/injection_points/specs/repack_deadlock.spec
new file mode 100644
index 00000000000..9d23a6588c2
--- /dev/null
+++ b/src/test/modules/injection_points/specs/repack_deadlock.spec
@@ -0,0 +1,83 @@
+# Test REPACK with a concurrent transaction that would cause a deadlock
+setup
+{
+ CREATE EXTENSION injection_points;
+
+ CREATE TABLE repack_deadlock(i int PRIMARY KEY, j int);
+ INSERT INTO repack_deadlock(i, j) VALUES (1, 1), (2, 2), (3, 3), (4, 4);
+
+ CREATE TABLE relfilenodes(node oid);
+
+ CREATE TABLE data_s1(i int, j int);
+ CREATE TABLE data_s2(i int, j int);
+}
+
+teardown
+{
+ DROP TABLE repack_deadlock;
+ DROP EXTENSION injection_points;
+
+ DROP TABLE relfilenodes;
+ DROP TABLE data_s1;
+ DROP TABLE data_s2;
+}
+
+session s1
+setup
+{
+ SELECT injection_points_set_local();
+ SELECT injection_points_attach('repack-concurrently-before-lock', 'wait');
+}
+# Perform the initial load and wait for s2 to do some data changes.
+step wait_before_lock
+{
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+}
+# Check the table from the perspective of s1.
+#
+# Besides the contents, we also check that relfilenode has changed.
+
+# Have each session write the contents into a table and use FULL JOIN to check
+# if the outputs are identical.
+step check1
+{
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+}
+teardown
+{
+ SELECT injection_points_detach('repack-concurrently-before-lock');
+}
+
+session s2
+# Change the existing data. UPDATE changes both key and non-key columns. Also
+# update one row twice to test whether tuple version generated by this session
+# can be found.
+step add_column
+{
+ alter table repack_deadlock add column noise text;
+}
+
+step wakeup_before_lock
+{
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+}
+
+# Test if data changes introduced while one session is performing REPACK
+# CONCURRENTLY find their way into the table.
+permutation
+ wait_before_lock
+ add_column
+ wakeup_before_lock
+ check1
--
2.47.3
--cldir6umqisr673d
Content-Type: text/x-diff; charset=utf-8
Content-Disposition: attachment;
filename="v45-0008-Teach-snapshot-builder-to-skip-transactions-runn.patch"
^ permalink raw reply [nested|flat] 34+ messages in thread
* [PATCH v46 6/7] Error out any process that would block at REPACK
@ 2026-03-25 19:35 Álvaro Herrera <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Álvaro Herrera @ 2026-03-25 19:35 UTC (permalink / raw)
Any process waiting on REPACK to release its lock would actually cause
it to deadlock when it tries to upgrade its lock to AEL, losing all work
done to that point. We avoid this by teaching the deadlock detector to
raise an error when this condition is detected.
---
src/backend/commands/repack.c | 53 ++++++++----
src/backend/storage/lmgr/deadlock.c | 15 ++++
src/include/storage/proc.h | 6 +-
src/test/modules/injection_points/Makefile | 1 +
.../expected/repack_deadlock.out | 63 ++++++++++++++
src/test/modules/injection_points/meson.build | 1 +
.../specs/repack_deadlock.spec | 83 +++++++++++++++++++
7 files changed, 202 insertions(+), 20 deletions(-)
create mode 100644 src/test/modules/injection_points/expected/repack_deadlock.out
create mode 100644 src/test/modules/injection_points/specs/repack_deadlock.spec
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index a58277bb6c1..21ab81c6f7f 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -276,6 +276,16 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
/* Determine the lock mode to use. */
lockmode = RepackLockLevel((params.options & CLUOPT_CONCURRENT) != 0);
+ /*
+ * If in concurrent mode, set the PROC_IN_CONCURRENT_REPACK flag. This
+ * makes the deadlock checker cause anyone that would conflict with us
+ * to error out. It's important to set this flag ahead of actually locking
+ * the relation; it won't of course affect anyone until we do have a lock
+ * that others can conflict with.
+ */
+ if ((params.options & CLUOPT_CONCURRENT) != 0)
+ MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+
/*
* If a single relation is specified, process it and we're done ... unless
* the relation is a partitioned table, in which case we fall through.
@@ -476,11 +486,8 @@ RepackLockLevel(bool concurrent)
* If indexOid is InvalidOid, the table will be rewritten in physical order
* instead of index order.
*
- * Note that, in the concurrent case, the function releases the lock at some
- * point, in order to get AccessExclusiveLock for the final steps (i.e. to
- * swap the relation files). To make things simpler, the caller should expect
- * OldHeap to be closed on return, regardless CLUOPT_CONCURRENT. (The
- * AccessExclusiveLock is kept till the end of the transaction.)
+ * On return, OldHeap is closed but locked with AccessExclusiveLock - the lock
+ * will be released at end of the transaction.
*
* 'cmd' indicates which command is being executed, to be used for error
* messages.
@@ -512,10 +519,12 @@ cluster_rel(RepackCommand cmd, Relation OldHeap, Oid indexOid,
/*
* Make sure we're not in a transaction block.
*
- * The reason is that repack_setup_logical_decoding() could deadlock
- * if there's an XID already assigned. It would be possible to run in
- * a transaction block if we had no XID, but this restriction is
- * simpler for users to understand and we don't lose anything.
+ * The reason is that repack_setup_logical_decoding() could wait
+ * indefinitely for our XID to complete. (The deadlock detector would
+ * not recognize it because we'd be waiting for ourselves, i.e. no
+ * real lock conflict.) It would be possible to run in a transaction
+ * block if we had no XID, but this restriction is simpler for users
+ * to understand and we don't lose anything.
*/
PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
@@ -998,10 +1007,8 @@ rebuild_relation(Relation OldHeap, Relation index, bool verbose,
* Note that the worker has to wait for all transactions with XID
* already assigned to finish. If some of those transactions is
* waiting for a lock conflicting with ShareUpdateExclusiveLock on our
- * table (e.g. it runs CREATE INDEX), we can end up in a deadlock.
- * Not sure this risk is worth unlocking/locking the table (and its
- * clustering index) and checking again if it's still eligible for
- * REPACK CONCURRENTLY.
+ * table (e.g. it runs CREATE INDEX), it should encounter ERROR in the
+ * deadlock checking code.
*/
start_repack_decoding_worker(tableOid);
@@ -3123,10 +3130,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
/*
- * Tuples and pages of the old heap will be gone, but the heap will stay.
+ * Now that we have all access-exclusive locks on all relations, we no
+ * longer want other processes to error out when trying to acquire a
+ * conflicting lock. Therefore, unset our flag.
+ */
+ MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+
+ /*
+ * Tuples and pages of the old heap will be gone, but the heap itself will
+ * stay. In order for predicate locks to continue to work, convert them
+ * to relation-level locks. We do this both for table and indexes.
*/
TransferPredicateLocksToHeapRelation(OldHeap);
- /* The same for indexes. */
for (int i = 0; i < nind; i++)
{
Relation index = ind_refs[i];
@@ -3346,9 +3361,11 @@ start_repack_decoding_worker(Oid relid)
/*
* The decoding setup must be done before the caller can have XID assigned
- * for any reason, otherwise the worker might end up in a deadlock,
- * waiting for the caller's transaction to end. Therefore wait here until
- * the worker indicates that it has the logical decoding initialized.
+ * for any reason, otherwise the worker might end up waiting for the
+ * caller's transaction to end. (Deadlock detector does not consider this
+ * a conflict because the worker is in the same locking group as the
+ * backend that launched it.) Therefore wait here until the worker
+ * indicates that it has the logical decoding initialized.
*/
ConditionVariablePrepareToSleep(&shared->cv);
for (;;)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3d3ec0da111 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -620,6 +620,21 @@ FindLockCycleRecurseMember(PGPROC *checkProc,
proc->statusFlags & PROC_IS_AUTOVACUUM)
blocking_autovacuum_proc = proc;
+ /*
+ * Similarly, if we note that we're blocked by some process
+ * running REPACK (CONCURRENTLY), just fail. That process
+ * is going to upgrade its lock at some point, and it would
+ * be inappropriate for any other process to cause that
+ * to fail.
+ */
+ if (checkProc == MyProc &&
+ proc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+ ereport(ERROR,
+ errcode(ERRCODE_OBJECT_IN_USE),
+ errmsg("could not wait for concurrent REPACK"),
+ errdetail("Process %d waits for REPACK running on process %d",
+ MyProc->pid, proc->pid));
+
/* We're done looking at this proclock */
break;
}
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 1dad125706e..8ad9718f3d6 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -69,10 +69,12 @@ struct XidCache
#define PROC_AFFECTS_ALL_HORIZONS 0x20 /* this proc's xmin must be
* included in vacuum horizons
* in all databases */
+#define PROC_IN_CONCURRENT_REPACK 0x40 /* REPACK (CONCURRENTLY) */
-/* flags reset at EOXact */
+/* flags reset at EOXact. A bit of a misnomer ... */
#define PROC_VACUUM_STATE_MASK \
- (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+ (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+ PROC_IN_CONCURRENT_REPACK)
/*
* Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/Makefile b/src/test/modules/injection_points/Makefile
index 2cd7d87c533..f7663859fe2 100644
--- a/src/test/modules/injection_points/Makefile
+++ b/src/test/modules/injection_points/Makefile
@@ -15,6 +15,7 @@ REGRESS_OPTS = --dlpath=$(top_builddir)/src/test/regress
ISOLATION = basic \
inplace \
repack \
+ repack_deadlock \
repack_toast \
syscache-update-pruned \
heap_lock_update
diff --git a/src/test/modules/injection_points/expected/repack_deadlock.out b/src/test/modules/injection_points/expected/repack_deadlock.out
new file mode 100644
index 00000000000..a86e4767536
--- /dev/null
+++ b/src/test/modules/injection_points/expected/repack_deadlock.out
@@ -0,0 +1,63 @@
+Parsed test spec with 2 sessions
+
+starting permutation: wait_before_lock add_column wakeup_before_lock check1
+injection_points_attach
+-----------------------
+
+(1 row)
+
+step wait_before_lock:
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+ <waiting ...>
+step add_column:
+ alter table repack_deadlock add column noise text;
+ <waiting ...>
+step add_column: <... completed>
+ERROR: could not wait for concurrent REPACK
+step wakeup_before_lock:
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1:
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+
+count
+-----
+ 1
+(1 row)
+
+i|j
+-+-
+1|1
+2|2
+3|3
+4|4
+(4 rows)
+
+count
+-----
+ 4
+(1 row)
+
+injection_points_detach
+-----------------------
+
+(1 row)
+
diff --git a/src/test/modules/injection_points/meson.build b/src/test/modules/injection_points/meson.build
index a414abb924b..1cd88d6db65 100644
--- a/src/test/modules/injection_points/meson.build
+++ b/src/test/modules/injection_points/meson.build
@@ -46,6 +46,7 @@ tests += {
'basic',
'inplace',
'repack',
+ 'repack_deadlock',
'repack_toast',
'syscache-update-pruned',
'heap_lock_update',
diff --git a/src/test/modules/injection_points/specs/repack_deadlock.spec b/src/test/modules/injection_points/specs/repack_deadlock.spec
new file mode 100644
index 00000000000..9d23a6588c2
--- /dev/null
+++ b/src/test/modules/injection_points/specs/repack_deadlock.spec
@@ -0,0 +1,83 @@
+# Test REPACK with a concurrent transaction that would cause a deadlock
+setup
+{
+ CREATE EXTENSION injection_points;
+
+ CREATE TABLE repack_deadlock(i int PRIMARY KEY, j int);
+ INSERT INTO repack_deadlock(i, j) VALUES (1, 1), (2, 2), (3, 3), (4, 4);
+
+ CREATE TABLE relfilenodes(node oid);
+
+ CREATE TABLE data_s1(i int, j int);
+ CREATE TABLE data_s2(i int, j int);
+}
+
+teardown
+{
+ DROP TABLE repack_deadlock;
+ DROP EXTENSION injection_points;
+
+ DROP TABLE relfilenodes;
+ DROP TABLE data_s1;
+ DROP TABLE data_s2;
+}
+
+session s1
+setup
+{
+ SELECT injection_points_set_local();
+ SELECT injection_points_attach('repack-concurrently-before-lock', 'wait');
+}
+# Perform the initial load and wait for s2 to do some data changes.
+step wait_before_lock
+{
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+}
+# Check the table from the perspective of s1.
+#
+# Besides the contents, we also check that relfilenode has changed.
+
+# Have each session write the contents into a table and use FULL JOIN to check
+# if the outputs are identical.
+step check1
+{
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+}
+teardown
+{
+ SELECT injection_points_detach('repack-concurrently-before-lock');
+}
+
+session s2
+# Change the existing data. UPDATE changes both key and non-key columns. Also
+# update one row twice to test whether tuple version generated by this session
+# can be found.
+step add_column
+{
+ alter table repack_deadlock add column noise text;
+}
+
+step wakeup_before_lock
+{
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+}
+
+# Test if data changes introduced while one session is performing REPACK
+# CONCURRENTLY find their way into the table.
+permutation
+ wait_before_lock
+ add_column
+ wakeup_before_lock
+ check1
--
2.47.3
--mdsv5bdgfjjj6d77
Content-Type: text/x-diff; charset=utf-8
Content-Disposition: attachment;
filename="v46-0007-Teach-snapshot-builder-to-skip-transactions-runn.patch"
^ permalink raw reply [nested|flat] 34+ messages in thread
* [PATCH v47 8/9] Error out any process that would block at REPACK
@ 2026-03-25 19:35 Álvaro Herrera <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Álvaro Herrera @ 2026-03-25 19:35 UTC (permalink / raw)
Any process waiting on REPACK to release its lock would actually cause
it to deadlock when it tries to upgrade its lock to AEL, losing all work
done to that point. We avoid this by teaching the deadlock detector to
raise an error when this condition is detected.
---
src/backend/commands/repack.c | 52 ++++++++----
src/backend/storage/lmgr/deadlock.c | 15 ++++
src/include/storage/proc.h | 6 +-
src/test/modules/injection_points/Makefile | 1 +
.../expected/repack_deadlock.out | 63 ++++++++++++++
src/test/modules/injection_points/meson.build | 1 +
.../specs/repack_deadlock.spec | 83 +++++++++++++++++++
7 files changed, 202 insertions(+), 19 deletions(-)
create mode 100644 src/test/modules/injection_points/expected/repack_deadlock.out
create mode 100644 src/test/modules/injection_points/specs/repack_deadlock.spec
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 9c807f75d71..3acb2e1d2ba 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -276,6 +276,16 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
/* Determine the lock mode to use. */
lockmode = RepackLockLevel((params.options & CLUOPT_CONCURRENT) != 0);
+ /*
+ * If in concurrent mode, set the PROC_IN_CONCURRENT_REPACK flag. This
+ * makes the deadlock checker cause anyone that would conflict with us
+ * to error out. It's important to set this flag ahead of actually locking
+ * the relation; it won't of course affect anyone until we do have a lock
+ * that others can conflict with.
+ */
+ if ((params.options & CLUOPT_CONCURRENT) != 0)
+ MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+
/*
* If a single relation is specified, process it and we're done ... unless
* the relation is a partitioned table, in which case we fall through.
@@ -476,11 +486,8 @@ RepackLockLevel(bool concurrent)
* If indexOid is InvalidOid, the table will be rewritten in physical order
* instead of index order.
*
- * Note that, in the concurrent case, the function releases the lock at some
- * point, in order to get AccessExclusiveLock for the final steps (i.e. to
- * swap the relation files). To make things simpler, the caller should expect
- * OldHeap to be closed on return, regardless CLUOPT_CONCURRENT. (The
- * AccessExclusiveLock is kept till the end of the transaction.)
+ * On return, OldHeap is closed but locked with AccessExclusiveLock - the lock
+ * will be released at end of the transaction.
*
* 'cmd' indicates which command is being executed, to be used for error
* messages.
@@ -512,10 +519,12 @@ cluster_rel(RepackCommand cmd, Relation OldHeap, Oid indexOid,
/*
* Make sure we're not in a transaction block.
*
- * The reason is that repack_setup_logical_decoding() could deadlock
- * if there's an XID already assigned. It would be possible to run in
- * a transaction block if we had no XID, but this restriction is
- * simpler for users to understand and we don't lose anything.
+ * The reason is that repack_setup_logical_decoding() could wait
+ * indefinitely for our XID to complete. (The deadlock detector would
+ * not recognize it because we'd be waiting for ourselves, i.e. no
+ * real lock conflict.) It would be possible to run in a transaction
+ * block if we had no XID, but this restriction is simpler for users
+ * to understand and we don't lose anything.
*/
PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
@@ -998,10 +1007,8 @@ rebuild_relation(Relation OldHeap, Relation index, bool verbose,
* Note that the worker has to wait for all transactions with XID
* already assigned to finish. If some of those transactions is
* waiting for a lock conflicting with ShareUpdateExclusiveLock on our
- * table (e.g. it runs CREATE INDEX), we can end up in a deadlock.
- * Not sure this risk is worth unlocking/locking the table (and its
- * clustering index) and checking again if it's still eligible for
- * REPACK CONCURRENTLY.
+ * table (e.g. it runs CREATE INDEX), it should encounter ERROR in the
+ * deadlock checking code.
*/
start_repack_decoding_worker(tableOid);
@@ -3090,7 +3097,16 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
/*
- * Tuples and pages of the old heap will be gone, but the heap will stay.
+ * Now that we have all access-exclusive locks on all relations, we no
+ * longer want other processes to error out when trying to acquire a
+ * conflicting lock. Therefore, unset our flag.
+ */
+ MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+
+ /*
+ * Tuples and pages of the old heap will be gone, but the heap itself will
+ * stay. In order for predicate locks to continue to work, convert them
+ * to relation-level locks. We do this both for table and indexes.
*/
TransferPredicateLocksToHeapRelation(OldHeap);
foreach_ptr(RelationData, index, indexrels)
@@ -3287,9 +3303,11 @@ start_repack_decoding_worker(Oid relid)
/*
* The decoding setup must be done before the caller can have XID assigned
- * for any reason, otherwise the worker might end up in a deadlock,
- * waiting for the caller's transaction to end. Therefore wait here until
- * the worker indicates that it has the logical decoding initialized.
+ * for any reason, otherwise the worker might end up waiting for the
+ * caller's transaction to end. (Deadlock detector does not consider this
+ * a conflict because the worker is in the same locking group as the
+ * backend that launched it.) Therefore wait here until the worker
+ * indicates that it has the logical decoding initialized.
*/
ConditionVariablePrepareToSleep(&shared->cv);
for (;;)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3d3ec0da111 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -620,6 +620,21 @@ FindLockCycleRecurseMember(PGPROC *checkProc,
proc->statusFlags & PROC_IS_AUTOVACUUM)
blocking_autovacuum_proc = proc;
+ /*
+ * Similarly, if we note that we're blocked by some process
+ * running REPACK (CONCURRENTLY), just fail. That process
+ * is going to upgrade its lock at some point, and it would
+ * be inappropriate for any other process to cause that
+ * to fail.
+ */
+ if (checkProc == MyProc &&
+ proc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+ ereport(ERROR,
+ errcode(ERRCODE_OBJECT_IN_USE),
+ errmsg("could not wait for concurrent REPACK"),
+ errdetail("Process %d waits for REPACK running on process %d",
+ MyProc->pid, proc->pid));
+
/* We're done looking at this proclock */
break;
}
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 1dad125706e..8ad9718f3d6 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -69,10 +69,12 @@ struct XidCache
#define PROC_AFFECTS_ALL_HORIZONS 0x20 /* this proc's xmin must be
* included in vacuum horizons
* in all databases */
+#define PROC_IN_CONCURRENT_REPACK 0x40 /* REPACK (CONCURRENTLY) */
-/* flags reset at EOXact */
+/* flags reset at EOXact. A bit of a misnomer ... */
#define PROC_VACUUM_STATE_MASK \
- (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+ (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+ PROC_IN_CONCURRENT_REPACK)
/*
* Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/Makefile b/src/test/modules/injection_points/Makefile
index 2cd7d87c533..f7663859fe2 100644
--- a/src/test/modules/injection_points/Makefile
+++ b/src/test/modules/injection_points/Makefile
@@ -15,6 +15,7 @@ REGRESS_OPTS = --dlpath=$(top_builddir)/src/test/regress
ISOLATION = basic \
inplace \
repack \
+ repack_deadlock \
repack_toast \
syscache-update-pruned \
heap_lock_update
diff --git a/src/test/modules/injection_points/expected/repack_deadlock.out b/src/test/modules/injection_points/expected/repack_deadlock.out
new file mode 100644
index 00000000000..a86e4767536
--- /dev/null
+++ b/src/test/modules/injection_points/expected/repack_deadlock.out
@@ -0,0 +1,63 @@
+Parsed test spec with 2 sessions
+
+starting permutation: wait_before_lock add_column wakeup_before_lock check1
+injection_points_attach
+-----------------------
+
+(1 row)
+
+step wait_before_lock:
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+ <waiting ...>
+step add_column:
+ alter table repack_deadlock add column noise text;
+ <waiting ...>
+step add_column: <... completed>
+ERROR: could not wait for concurrent REPACK
+step wakeup_before_lock:
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1:
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+
+count
+-----
+ 1
+(1 row)
+
+i|j
+-+-
+1|1
+2|2
+3|3
+4|4
+(4 rows)
+
+count
+-----
+ 4
+(1 row)
+
+injection_points_detach
+-----------------------
+
+(1 row)
+
diff --git a/src/test/modules/injection_points/meson.build b/src/test/modules/injection_points/meson.build
index a414abb924b..1cd88d6db65 100644
--- a/src/test/modules/injection_points/meson.build
+++ b/src/test/modules/injection_points/meson.build
@@ -46,6 +46,7 @@ tests += {
'basic',
'inplace',
'repack',
+ 'repack_deadlock',
'repack_toast',
'syscache-update-pruned',
'heap_lock_update',
diff --git a/src/test/modules/injection_points/specs/repack_deadlock.spec b/src/test/modules/injection_points/specs/repack_deadlock.spec
new file mode 100644
index 00000000000..9d23a6588c2
--- /dev/null
+++ b/src/test/modules/injection_points/specs/repack_deadlock.spec
@@ -0,0 +1,83 @@
+# Test REPACK with a concurrent transaction that would cause a deadlock
+setup
+{
+ CREATE EXTENSION injection_points;
+
+ CREATE TABLE repack_deadlock(i int PRIMARY KEY, j int);
+ INSERT INTO repack_deadlock(i, j) VALUES (1, 1), (2, 2), (3, 3), (4, 4);
+
+ CREATE TABLE relfilenodes(node oid);
+
+ CREATE TABLE data_s1(i int, j int);
+ CREATE TABLE data_s2(i int, j int);
+}
+
+teardown
+{
+ DROP TABLE repack_deadlock;
+ DROP EXTENSION injection_points;
+
+ DROP TABLE relfilenodes;
+ DROP TABLE data_s1;
+ DROP TABLE data_s2;
+}
+
+session s1
+setup
+{
+ SELECT injection_points_set_local();
+ SELECT injection_points_attach('repack-concurrently-before-lock', 'wait');
+}
+# Perform the initial load and wait for s2 to do some data changes.
+step wait_before_lock
+{
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+}
+# Check the table from the perspective of s1.
+#
+# Besides the contents, we also check that relfilenode has changed.
+
+# Have each session write the contents into a table and use FULL JOIN to check
+# if the outputs are identical.
+step check1
+{
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+}
+teardown
+{
+ SELECT injection_points_detach('repack-concurrently-before-lock');
+}
+
+session s2
+# Change the existing data. UPDATE changes both key and non-key columns. Also
+# update one row twice to test whether tuple version generated by this session
+# can be found.
+step add_column
+{
+ alter table repack_deadlock add column noise text;
+}
+
+step wakeup_before_lock
+{
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+}
+
+# Test if data changes introduced while one session is performing REPACK
+# CONCURRENTLY find their way into the table.
+permutation
+ wait_before_lock
+ add_column
+ wakeup_before_lock
+ check1
--
2.47.3
--2fxmamo6mu2qbxgv
Content-Type: text/x-diff; charset=utf-8
Content-Disposition: attachment;
filename="v47-0009-Teach-snapshot-builder-to-skip-transactions-runn.patch"
^ permalink raw reply [nested|flat] 34+ messages in thread
* [PATCH v48 6/7] Error out any process that would block at REPACK
@ 2026-03-25 19:35 Álvaro Herrera <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Álvaro Herrera @ 2026-03-25 19:35 UTC (permalink / raw)
Any process waiting on REPACK to release its lock would actually cause
it to deadlock when it tries to upgrade its lock to AEL, losing all work
done to that point. We avoid this by teaching the deadlock detector to
raise an error when this condition is detected.
---
src/backend/commands/repack.c | 52 ++++++++----
src/backend/storage/lmgr/deadlock.c | 15 ++++
src/include/storage/proc.h | 6 +-
src/test/modules/injection_points/Makefile | 1 +
.../expected/repack_deadlock.out | 63 ++++++++++++++
src/test/modules/injection_points/meson.build | 1 +
.../specs/repack_deadlock.spec | 83 +++++++++++++++++++
7 files changed, 202 insertions(+), 19 deletions(-)
create mode 100644 src/test/modules/injection_points/expected/repack_deadlock.out
create mode 100644 src/test/modules/injection_points/specs/repack_deadlock.spec
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index a514053b777..781a471fc1d 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -276,6 +276,16 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
/* Determine the lock mode to use. */
lockmode = RepackLockLevel((params.options & CLUOPT_CONCURRENT) != 0);
+ /*
+ * If in concurrent mode, set the PROC_IN_CONCURRENT_REPACK flag. This
+ * makes the deadlock checker cause anyone that would conflict with us to
+ * error out. It's important to set this flag ahead of actually locking
+ * the relation; it won't of course affect anyone until we do have a lock
+ * that others can conflict with.
+ */
+ if ((params.options & CLUOPT_CONCURRENT) != 0)
+ MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+
/*
* If a single relation is specified, process it and we're done ... unless
* the relation is a partitioned table, in which case we fall through.
@@ -476,11 +486,8 @@ RepackLockLevel(bool concurrent)
* If indexOid is InvalidOid, the table will be rewritten in physical order
* instead of index order.
*
- * Note that, in the concurrent case, the function releases the lock at some
- * point, in order to get AccessExclusiveLock for the final steps (i.e. to
- * swap the relation files). To make things simpler, the caller should expect
- * OldHeap to be closed on return, regardless CLUOPT_CONCURRENT. (The
- * AccessExclusiveLock is kept till the end of the transaction.)
+ * On return, OldHeap is closed but locked with AccessExclusiveLock - the lock
+ * will be released at end of the transaction.
*
* 'cmd' indicates which command is being executed, to be used for error
* messages.
@@ -512,10 +519,12 @@ cluster_rel(RepackCommand cmd, Relation OldHeap, Oid indexOid,
/*
* Make sure we're not in a transaction block.
*
- * The reason is that repack_setup_logical_decoding() could deadlock
- * if there's an XID already assigned. It would be possible to run in
- * a transaction block if we had no XID, but this restriction is
- * simpler for users to understand and we don't lose anything.
+ * The reason is that repack_setup_logical_decoding() could wait
+ * indefinitely for our XID to complete. (The deadlock detector would
+ * not recognize it because we'd be waiting for ourselves, i.e. no
+ * real lock conflict.) It would be possible to run in a transaction
+ * block if we had no XID, but this restriction is simpler for users
+ * to understand and we don't lose anything.
*/
PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
@@ -998,10 +1007,8 @@ rebuild_relation(Relation OldHeap, Relation index, bool verbose,
* Note that the worker has to wait for all transactions with XID
* already assigned to finish. If some of those transactions is
* waiting for a lock conflicting with ShareUpdateExclusiveLock on our
- * table (e.g. it runs CREATE INDEX), we can end up in a deadlock.
- * Not sure this risk is worth unlocking/locking the table (and its
- * clustering index) and checking again if it's still eligible for
- * REPACK CONCURRENTLY.
+ * table (e.g. it runs CREATE INDEX), it should encounter ERROR in the
+ * deadlock checking code.
*/
start_repack_decoding_worker(tableOid);
@@ -3090,7 +3097,16 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
/*
- * Tuples and pages of the old heap will be gone, but the heap will stay.
+ * Now that we have all access-exclusive locks on all relations, we no
+ * longer want other processes to error out when trying to acquire a
+ * conflicting lock. Therefore, unset our flag.
+ */
+ MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+
+ /*
+ * Tuples and pages of the old heap will be gone, but the heap itself will
+ * stay. In order for predicate locks to continue to work, convert them
+ * to relation-level locks. We do this both for table and indexes.
*/
TransferPredicateLocksToHeapRelation(OldHeap);
foreach_ptr(RelationData, index, indexrels)
@@ -3288,9 +3304,11 @@ start_repack_decoding_worker(Oid relid)
/*
* The decoding setup must be done before the caller can have XID assigned
- * for any reason, otherwise the worker might end up in a deadlock,
- * waiting for the caller's transaction to end. Therefore wait here until
- * the worker indicates that it has the logical decoding initialized.
+ * for any reason, otherwise the worker might end up waiting for the
+ * caller's transaction to end. (Deadlock detector does not consider this
+ * a conflict because the worker is in the same locking group as the
+ * backend that launched it.) Therefore wait here until the worker
+ * indicates that it has the logical decoding initialized.
*/
ConditionVariablePrepareToSleep(&shared->cv);
for (;;)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..c20ac682b0d 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -620,6 +620,21 @@ FindLockCycleRecurseMember(PGPROC *checkProc,
proc->statusFlags & PROC_IS_AUTOVACUUM)
blocking_autovacuum_proc = proc;
+ /*
+ * Similarly, if we note that we're blocked by some
+ * process running REPACK (CONCURRENTLY), just fail. That
+ * process is going to upgrade its lock at some point, and
+ * it would be inappropriate for any other process to
+ * cause that to fail.
+ */
+ if (checkProc == MyProc &&
+ proc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+ ereport(ERROR,
+ errcode(ERRCODE_OBJECT_IN_USE),
+ errmsg("could not wait for concurrent REPACK"),
+ errdetail("Process %d waits for REPACK running on process %d",
+ MyProc->pid, proc->pid));
+
/* We're done looking at this proclock */
break;
}
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 1dad125706e..8ad9718f3d6 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -69,10 +69,12 @@ struct XidCache
#define PROC_AFFECTS_ALL_HORIZONS 0x20 /* this proc's xmin must be
* included in vacuum horizons
* in all databases */
+#define PROC_IN_CONCURRENT_REPACK 0x40 /* REPACK (CONCURRENTLY) */
-/* flags reset at EOXact */
+/* flags reset at EOXact. A bit of a misnomer ... */
#define PROC_VACUUM_STATE_MASK \
- (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+ (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+ PROC_IN_CONCURRENT_REPACK)
/*
* Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/Makefile b/src/test/modules/injection_points/Makefile
index 2cd7d87c533..f7663859fe2 100644
--- a/src/test/modules/injection_points/Makefile
+++ b/src/test/modules/injection_points/Makefile
@@ -15,6 +15,7 @@ REGRESS_OPTS = --dlpath=$(top_builddir)/src/test/regress
ISOLATION = basic \
inplace \
repack \
+ repack_deadlock \
repack_toast \
syscache-update-pruned \
heap_lock_update
diff --git a/src/test/modules/injection_points/expected/repack_deadlock.out b/src/test/modules/injection_points/expected/repack_deadlock.out
new file mode 100644
index 00000000000..a86e4767536
--- /dev/null
+++ b/src/test/modules/injection_points/expected/repack_deadlock.out
@@ -0,0 +1,63 @@
+Parsed test spec with 2 sessions
+
+starting permutation: wait_before_lock add_column wakeup_before_lock check1
+injection_points_attach
+-----------------------
+
+(1 row)
+
+step wait_before_lock:
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+ <waiting ...>
+step add_column:
+ alter table repack_deadlock add column noise text;
+ <waiting ...>
+step add_column: <... completed>
+ERROR: could not wait for concurrent REPACK
+step wakeup_before_lock:
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1:
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+
+count
+-----
+ 1
+(1 row)
+
+i|j
+-+-
+1|1
+2|2
+3|3
+4|4
+(4 rows)
+
+count
+-----
+ 4
+(1 row)
+
+injection_points_detach
+-----------------------
+
+(1 row)
+
diff --git a/src/test/modules/injection_points/meson.build b/src/test/modules/injection_points/meson.build
index a414abb924b..1cd88d6db65 100644
--- a/src/test/modules/injection_points/meson.build
+++ b/src/test/modules/injection_points/meson.build
@@ -46,6 +46,7 @@ tests += {
'basic',
'inplace',
'repack',
+ 'repack_deadlock',
'repack_toast',
'syscache-update-pruned',
'heap_lock_update',
diff --git a/src/test/modules/injection_points/specs/repack_deadlock.spec b/src/test/modules/injection_points/specs/repack_deadlock.spec
new file mode 100644
index 00000000000..9d23a6588c2
--- /dev/null
+++ b/src/test/modules/injection_points/specs/repack_deadlock.spec
@@ -0,0 +1,83 @@
+# Test REPACK with a concurrent transaction that would cause a deadlock
+setup
+{
+ CREATE EXTENSION injection_points;
+
+ CREATE TABLE repack_deadlock(i int PRIMARY KEY, j int);
+ INSERT INTO repack_deadlock(i, j) VALUES (1, 1), (2, 2), (3, 3), (4, 4);
+
+ CREATE TABLE relfilenodes(node oid);
+
+ CREATE TABLE data_s1(i int, j int);
+ CREATE TABLE data_s2(i int, j int);
+}
+
+teardown
+{
+ DROP TABLE repack_deadlock;
+ DROP EXTENSION injection_points;
+
+ DROP TABLE relfilenodes;
+ DROP TABLE data_s1;
+ DROP TABLE data_s2;
+}
+
+session s1
+setup
+{
+ SELECT injection_points_set_local();
+ SELECT injection_points_attach('repack-concurrently-before-lock', 'wait');
+}
+# Perform the initial load and wait for s2 to do some data changes.
+step wait_before_lock
+{
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+}
+# Check the table from the perspective of s1.
+#
+# Besides the contents, we also check that relfilenode has changed.
+
+# Have each session write the contents into a table and use FULL JOIN to check
+# if the outputs are identical.
+step check1
+{
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+}
+teardown
+{
+ SELECT injection_points_detach('repack-concurrently-before-lock');
+}
+
+session s2
+# Change the existing data. UPDATE changes both key and non-key columns. Also
+# update one row twice to test whether tuple version generated by this session
+# can be found.
+step add_column
+{
+ alter table repack_deadlock add column noise text;
+}
+
+step wakeup_before_lock
+{
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+}
+
+# Test if data changes introduced while one session is performing REPACK
+# CONCURRENTLY find their way into the table.
+permutation
+ wait_before_lock
+ add_column
+ wakeup_before_lock
+ check1
--
2.47.3
--qfkt2ktdpcfeypib
Content-Type: text/x-diff; charset=utf-8
Content-Disposition: attachment;
filename="v48-0007-Teach-snapshot-builder-to-skip-transactions-runn.patch"
^ permalink raw reply [nested|flat] 34+ messages in thread
* [PATCH v47 8/9] Error out any process that would block at REPACK
@ 2026-03-25 19:35 Álvaro Herrera <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Álvaro Herrera @ 2026-03-25 19:35 UTC (permalink / raw)
Any process waiting on REPACK to release its lock would actually cause
it to deadlock when it tries to upgrade its lock to AEL, losing all work
done to that point. We avoid this by teaching the deadlock detector to
raise an error when this condition is detected.
---
src/backend/commands/repack.c | 52 ++++++++----
src/backend/storage/lmgr/deadlock.c | 15 ++++
src/include/storage/proc.h | 6 +-
src/test/modules/injection_points/Makefile | 1 +
.../expected/repack_deadlock.out | 63 ++++++++++++++
src/test/modules/injection_points/meson.build | 1 +
.../specs/repack_deadlock.spec | 83 +++++++++++++++++++
7 files changed, 202 insertions(+), 19 deletions(-)
create mode 100644 src/test/modules/injection_points/expected/repack_deadlock.out
create mode 100644 src/test/modules/injection_points/specs/repack_deadlock.spec
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 9c807f75d71..3acb2e1d2ba 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -276,6 +276,16 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
/* Determine the lock mode to use. */
lockmode = RepackLockLevel((params.options & CLUOPT_CONCURRENT) != 0);
+ /*
+ * If in concurrent mode, set the PROC_IN_CONCURRENT_REPACK flag. This
+ * makes the deadlock checker cause anyone that would conflict with us
+ * to error out. It's important to set this flag ahead of actually locking
+ * the relation; it won't of course affect anyone until we do have a lock
+ * that others can conflict with.
+ */
+ if ((params.options & CLUOPT_CONCURRENT) != 0)
+ MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+
/*
* If a single relation is specified, process it and we're done ... unless
* the relation is a partitioned table, in which case we fall through.
@@ -476,11 +486,8 @@ RepackLockLevel(bool concurrent)
* If indexOid is InvalidOid, the table will be rewritten in physical order
* instead of index order.
*
- * Note that, in the concurrent case, the function releases the lock at some
- * point, in order to get AccessExclusiveLock for the final steps (i.e. to
- * swap the relation files). To make things simpler, the caller should expect
- * OldHeap to be closed on return, regardless CLUOPT_CONCURRENT. (The
- * AccessExclusiveLock is kept till the end of the transaction.)
+ * On return, OldHeap is closed but locked with AccessExclusiveLock - the lock
+ * will be released at end of the transaction.
*
* 'cmd' indicates which command is being executed, to be used for error
* messages.
@@ -512,10 +519,12 @@ cluster_rel(RepackCommand cmd, Relation OldHeap, Oid indexOid,
/*
* Make sure we're not in a transaction block.
*
- * The reason is that repack_setup_logical_decoding() could deadlock
- * if there's an XID already assigned. It would be possible to run in
- * a transaction block if we had no XID, but this restriction is
- * simpler for users to understand and we don't lose anything.
+ * The reason is that repack_setup_logical_decoding() could wait
+ * indefinitely for our XID to complete. (The deadlock detector would
+ * not recognize it because we'd be waiting for ourselves, i.e. no
+ * real lock conflict.) It would be possible to run in a transaction
+ * block if we had no XID, but this restriction is simpler for users
+ * to understand and we don't lose anything.
*/
PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
@@ -998,10 +1007,8 @@ rebuild_relation(Relation OldHeap, Relation index, bool verbose,
* Note that the worker has to wait for all transactions with XID
* already assigned to finish. If some of those transactions is
* waiting for a lock conflicting with ShareUpdateExclusiveLock on our
- * table (e.g. it runs CREATE INDEX), we can end up in a deadlock.
- * Not sure this risk is worth unlocking/locking the table (and its
- * clustering index) and checking again if it's still eligible for
- * REPACK CONCURRENTLY.
+ * table (e.g. it runs CREATE INDEX), it should encounter ERROR in the
+ * deadlock checking code.
*/
start_repack_decoding_worker(tableOid);
@@ -3090,7 +3097,16 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
/*
- * Tuples and pages of the old heap will be gone, but the heap will stay.
+ * Now that we have all access-exclusive locks on all relations, we no
+ * longer want other processes to error out when trying to acquire a
+ * conflicting lock. Therefore, unset our flag.
+ */
+ MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+
+ /*
+ * Tuples and pages of the old heap will be gone, but the heap itself will
+ * stay. In order for predicate locks to continue to work, convert them
+ * to relation-level locks. We do this both for table and indexes.
*/
TransferPredicateLocksToHeapRelation(OldHeap);
foreach_ptr(RelationData, index, indexrels)
@@ -3287,9 +3303,11 @@ start_repack_decoding_worker(Oid relid)
/*
* The decoding setup must be done before the caller can have XID assigned
- * for any reason, otherwise the worker might end up in a deadlock,
- * waiting for the caller's transaction to end. Therefore wait here until
- * the worker indicates that it has the logical decoding initialized.
+ * for any reason, otherwise the worker might end up waiting for the
+ * caller's transaction to end. (Deadlock detector does not consider this
+ * a conflict because the worker is in the same locking group as the
+ * backend that launched it.) Therefore wait here until the worker
+ * indicates that it has the logical decoding initialized.
*/
ConditionVariablePrepareToSleep(&shared->cv);
for (;;)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3d3ec0da111 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -620,6 +620,21 @@ FindLockCycleRecurseMember(PGPROC *checkProc,
proc->statusFlags & PROC_IS_AUTOVACUUM)
blocking_autovacuum_proc = proc;
+ /*
+ * Similarly, if we note that we're blocked by some process
+ * running REPACK (CONCURRENTLY), just fail. That process
+ * is going to upgrade its lock at some point, and it would
+ * be inappropriate for any other process to cause that
+ * to fail.
+ */
+ if (checkProc == MyProc &&
+ proc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+ ereport(ERROR,
+ errcode(ERRCODE_OBJECT_IN_USE),
+ errmsg("could not wait for concurrent REPACK"),
+ errdetail("Process %d waits for REPACK running on process %d",
+ MyProc->pid, proc->pid));
+
/* We're done looking at this proclock */
break;
}
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 1dad125706e..8ad9718f3d6 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -69,10 +69,12 @@ struct XidCache
#define PROC_AFFECTS_ALL_HORIZONS 0x20 /* this proc's xmin must be
* included in vacuum horizons
* in all databases */
+#define PROC_IN_CONCURRENT_REPACK 0x40 /* REPACK (CONCURRENTLY) */
-/* flags reset at EOXact */
+/* flags reset at EOXact. A bit of a misnomer ... */
#define PROC_VACUUM_STATE_MASK \
- (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+ (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+ PROC_IN_CONCURRENT_REPACK)
/*
* Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/Makefile b/src/test/modules/injection_points/Makefile
index 2cd7d87c533..f7663859fe2 100644
--- a/src/test/modules/injection_points/Makefile
+++ b/src/test/modules/injection_points/Makefile
@@ -15,6 +15,7 @@ REGRESS_OPTS = --dlpath=$(top_builddir)/src/test/regress
ISOLATION = basic \
inplace \
repack \
+ repack_deadlock \
repack_toast \
syscache-update-pruned \
heap_lock_update
diff --git a/src/test/modules/injection_points/expected/repack_deadlock.out b/src/test/modules/injection_points/expected/repack_deadlock.out
new file mode 100644
index 00000000000..a86e4767536
--- /dev/null
+++ b/src/test/modules/injection_points/expected/repack_deadlock.out
@@ -0,0 +1,63 @@
+Parsed test spec with 2 sessions
+
+starting permutation: wait_before_lock add_column wakeup_before_lock check1
+injection_points_attach
+-----------------------
+
+(1 row)
+
+step wait_before_lock:
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+ <waiting ...>
+step add_column:
+ alter table repack_deadlock add column noise text;
+ <waiting ...>
+step add_column: <... completed>
+ERROR: could not wait for concurrent REPACK
+step wakeup_before_lock:
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1:
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+
+count
+-----
+ 1
+(1 row)
+
+i|j
+-+-
+1|1
+2|2
+3|3
+4|4
+(4 rows)
+
+count
+-----
+ 4
+(1 row)
+
+injection_points_detach
+-----------------------
+
+(1 row)
+
diff --git a/src/test/modules/injection_points/meson.build b/src/test/modules/injection_points/meson.build
index a414abb924b..1cd88d6db65 100644
--- a/src/test/modules/injection_points/meson.build
+++ b/src/test/modules/injection_points/meson.build
@@ -46,6 +46,7 @@ tests += {
'basic',
'inplace',
'repack',
+ 'repack_deadlock',
'repack_toast',
'syscache-update-pruned',
'heap_lock_update',
diff --git a/src/test/modules/injection_points/specs/repack_deadlock.spec b/src/test/modules/injection_points/specs/repack_deadlock.spec
new file mode 100644
index 00000000000..9d23a6588c2
--- /dev/null
+++ b/src/test/modules/injection_points/specs/repack_deadlock.spec
@@ -0,0 +1,83 @@
+# Test REPACK with a concurrent transaction that would cause a deadlock
+setup
+{
+ CREATE EXTENSION injection_points;
+
+ CREATE TABLE repack_deadlock(i int PRIMARY KEY, j int);
+ INSERT INTO repack_deadlock(i, j) VALUES (1, 1), (2, 2), (3, 3), (4, 4);
+
+ CREATE TABLE relfilenodes(node oid);
+
+ CREATE TABLE data_s1(i int, j int);
+ CREATE TABLE data_s2(i int, j int);
+}
+
+teardown
+{
+ DROP TABLE repack_deadlock;
+ DROP EXTENSION injection_points;
+
+ DROP TABLE relfilenodes;
+ DROP TABLE data_s1;
+ DROP TABLE data_s2;
+}
+
+session s1
+setup
+{
+ SELECT injection_points_set_local();
+ SELECT injection_points_attach('repack-concurrently-before-lock', 'wait');
+}
+# Perform the initial load and wait for s2 to do some data changes.
+step wait_before_lock
+{
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+}
+# Check the table from the perspective of s1.
+#
+# Besides the contents, we also check that relfilenode has changed.
+
+# Have each session write the contents into a table and use FULL JOIN to check
+# if the outputs are identical.
+step check1
+{
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+}
+teardown
+{
+ SELECT injection_points_detach('repack-concurrently-before-lock');
+}
+
+session s2
+# Change the existing data. UPDATE changes both key and non-key columns. Also
+# update one row twice to test whether tuple version generated by this session
+# can be found.
+step add_column
+{
+ alter table repack_deadlock add column noise text;
+}
+
+step wakeup_before_lock
+{
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+}
+
+# Test if data changes introduced while one session is performing REPACK
+# CONCURRENTLY find their way into the table.
+permutation
+ wait_before_lock
+ add_column
+ wakeup_before_lock
+ check1
--
2.47.3
--2fxmamo6mu2qbxgv
Content-Type: text/x-diff; charset=utf-8
Content-Disposition: attachment;
filename="v47-0009-Teach-snapshot-builder-to-skip-transactions-runn.patch"
^ permalink raw reply [nested|flat] 34+ messages in thread
* [PATCH v46 6/7] Error out any process that would block at REPACK
@ 2026-03-25 19:35 Álvaro Herrera <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Álvaro Herrera @ 2026-03-25 19:35 UTC (permalink / raw)
Any process waiting on REPACK to release its lock would actually cause
it to deadlock when it tries to upgrade its lock to AEL, losing all work
done to that point. We avoid this by teaching the deadlock detector to
raise an error when this condition is detected.
---
src/backend/commands/repack.c | 53 ++++++++----
src/backend/storage/lmgr/deadlock.c | 15 ++++
src/include/storage/proc.h | 6 +-
src/test/modules/injection_points/Makefile | 1 +
.../expected/repack_deadlock.out | 63 ++++++++++++++
src/test/modules/injection_points/meson.build | 1 +
.../specs/repack_deadlock.spec | 83 +++++++++++++++++++
7 files changed, 202 insertions(+), 20 deletions(-)
create mode 100644 src/test/modules/injection_points/expected/repack_deadlock.out
create mode 100644 src/test/modules/injection_points/specs/repack_deadlock.spec
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index a58277bb6c1..21ab81c6f7f 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -276,6 +276,16 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
/* Determine the lock mode to use. */
lockmode = RepackLockLevel((params.options & CLUOPT_CONCURRENT) != 0);
+ /*
+ * If in concurrent mode, set the PROC_IN_CONCURRENT_REPACK flag. This
+ * makes the deadlock checker cause anyone that would conflict with us
+ * to error out. It's important to set this flag ahead of actually locking
+ * the relation; it won't of course affect anyone until we do have a lock
+ * that others can conflict with.
+ */
+ if ((params.options & CLUOPT_CONCURRENT) != 0)
+ MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+
/*
* If a single relation is specified, process it and we're done ... unless
* the relation is a partitioned table, in which case we fall through.
@@ -476,11 +486,8 @@ RepackLockLevel(bool concurrent)
* If indexOid is InvalidOid, the table will be rewritten in physical order
* instead of index order.
*
- * Note that, in the concurrent case, the function releases the lock at some
- * point, in order to get AccessExclusiveLock for the final steps (i.e. to
- * swap the relation files). To make things simpler, the caller should expect
- * OldHeap to be closed on return, regardless CLUOPT_CONCURRENT. (The
- * AccessExclusiveLock is kept till the end of the transaction.)
+ * On return, OldHeap is closed but locked with AccessExclusiveLock - the lock
+ * will be released at end of the transaction.
*
* 'cmd' indicates which command is being executed, to be used for error
* messages.
@@ -512,10 +519,12 @@ cluster_rel(RepackCommand cmd, Relation OldHeap, Oid indexOid,
/*
* Make sure we're not in a transaction block.
*
- * The reason is that repack_setup_logical_decoding() could deadlock
- * if there's an XID already assigned. It would be possible to run in
- * a transaction block if we had no XID, but this restriction is
- * simpler for users to understand and we don't lose anything.
+ * The reason is that repack_setup_logical_decoding() could wait
+ * indefinitely for our XID to complete. (The deadlock detector would
+ * not recognize it because we'd be waiting for ourselves, i.e. no
+ * real lock conflict.) It would be possible to run in a transaction
+ * block if we had no XID, but this restriction is simpler for users
+ * to understand and we don't lose anything.
*/
PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
@@ -998,10 +1007,8 @@ rebuild_relation(Relation OldHeap, Relation index, bool verbose,
* Note that the worker has to wait for all transactions with XID
* already assigned to finish. If some of those transactions is
* waiting for a lock conflicting with ShareUpdateExclusiveLock on our
- * table (e.g. it runs CREATE INDEX), we can end up in a deadlock.
- * Not sure this risk is worth unlocking/locking the table (and its
- * clustering index) and checking again if it's still eligible for
- * REPACK CONCURRENTLY.
+ * table (e.g. it runs CREATE INDEX), it should encounter ERROR in the
+ * deadlock checking code.
*/
start_repack_decoding_worker(tableOid);
@@ -3123,10 +3130,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
/*
- * Tuples and pages of the old heap will be gone, but the heap will stay.
+ * Now that we have all access-exclusive locks on all relations, we no
+ * longer want other processes to error out when trying to acquire a
+ * conflicting lock. Therefore, unset our flag.
+ */
+ MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+
+ /*
+ * Tuples and pages of the old heap will be gone, but the heap itself will
+ * stay. In order for predicate locks to continue to work, convert them
+ * to relation-level locks. We do this both for table and indexes.
*/
TransferPredicateLocksToHeapRelation(OldHeap);
- /* The same for indexes. */
for (int i = 0; i < nind; i++)
{
Relation index = ind_refs[i];
@@ -3346,9 +3361,11 @@ start_repack_decoding_worker(Oid relid)
/*
* The decoding setup must be done before the caller can have XID assigned
- * for any reason, otherwise the worker might end up in a deadlock,
- * waiting for the caller's transaction to end. Therefore wait here until
- * the worker indicates that it has the logical decoding initialized.
+ * for any reason, otherwise the worker might end up waiting for the
+ * caller's transaction to end. (Deadlock detector does not consider this
+ * a conflict because the worker is in the same locking group as the
+ * backend that launched it.) Therefore wait here until the worker
+ * indicates that it has the logical decoding initialized.
*/
ConditionVariablePrepareToSleep(&shared->cv);
for (;;)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3d3ec0da111 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -620,6 +620,21 @@ FindLockCycleRecurseMember(PGPROC *checkProc,
proc->statusFlags & PROC_IS_AUTOVACUUM)
blocking_autovacuum_proc = proc;
+ /*
+ * Similarly, if we note that we're blocked by some process
+ * running REPACK (CONCURRENTLY), just fail. That process
+ * is going to upgrade its lock at some point, and it would
+ * be inappropriate for any other process to cause that
+ * to fail.
+ */
+ if (checkProc == MyProc &&
+ proc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+ ereport(ERROR,
+ errcode(ERRCODE_OBJECT_IN_USE),
+ errmsg("could not wait for concurrent REPACK"),
+ errdetail("Process %d waits for REPACK running on process %d",
+ MyProc->pid, proc->pid));
+
/* We're done looking at this proclock */
break;
}
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 1dad125706e..8ad9718f3d6 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -69,10 +69,12 @@ struct XidCache
#define PROC_AFFECTS_ALL_HORIZONS 0x20 /* this proc's xmin must be
* included in vacuum horizons
* in all databases */
+#define PROC_IN_CONCURRENT_REPACK 0x40 /* REPACK (CONCURRENTLY) */
-/* flags reset at EOXact */
+/* flags reset at EOXact. A bit of a misnomer ... */
#define PROC_VACUUM_STATE_MASK \
- (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+ (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+ PROC_IN_CONCURRENT_REPACK)
/*
* Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/Makefile b/src/test/modules/injection_points/Makefile
index 2cd7d87c533..f7663859fe2 100644
--- a/src/test/modules/injection_points/Makefile
+++ b/src/test/modules/injection_points/Makefile
@@ -15,6 +15,7 @@ REGRESS_OPTS = --dlpath=$(top_builddir)/src/test/regress
ISOLATION = basic \
inplace \
repack \
+ repack_deadlock \
repack_toast \
syscache-update-pruned \
heap_lock_update
diff --git a/src/test/modules/injection_points/expected/repack_deadlock.out b/src/test/modules/injection_points/expected/repack_deadlock.out
new file mode 100644
index 00000000000..a86e4767536
--- /dev/null
+++ b/src/test/modules/injection_points/expected/repack_deadlock.out
@@ -0,0 +1,63 @@
+Parsed test spec with 2 sessions
+
+starting permutation: wait_before_lock add_column wakeup_before_lock check1
+injection_points_attach
+-----------------------
+
+(1 row)
+
+step wait_before_lock:
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+ <waiting ...>
+step add_column:
+ alter table repack_deadlock add column noise text;
+ <waiting ...>
+step add_column: <... completed>
+ERROR: could not wait for concurrent REPACK
+step wakeup_before_lock:
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1:
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+
+count
+-----
+ 1
+(1 row)
+
+i|j
+-+-
+1|1
+2|2
+3|3
+4|4
+(4 rows)
+
+count
+-----
+ 4
+(1 row)
+
+injection_points_detach
+-----------------------
+
+(1 row)
+
diff --git a/src/test/modules/injection_points/meson.build b/src/test/modules/injection_points/meson.build
index a414abb924b..1cd88d6db65 100644
--- a/src/test/modules/injection_points/meson.build
+++ b/src/test/modules/injection_points/meson.build
@@ -46,6 +46,7 @@ tests += {
'basic',
'inplace',
'repack',
+ 'repack_deadlock',
'repack_toast',
'syscache-update-pruned',
'heap_lock_update',
diff --git a/src/test/modules/injection_points/specs/repack_deadlock.spec b/src/test/modules/injection_points/specs/repack_deadlock.spec
new file mode 100644
index 00000000000..9d23a6588c2
--- /dev/null
+++ b/src/test/modules/injection_points/specs/repack_deadlock.spec
@@ -0,0 +1,83 @@
+# Test REPACK with a concurrent transaction that would cause a deadlock
+setup
+{
+ CREATE EXTENSION injection_points;
+
+ CREATE TABLE repack_deadlock(i int PRIMARY KEY, j int);
+ INSERT INTO repack_deadlock(i, j) VALUES (1, 1), (2, 2), (3, 3), (4, 4);
+
+ CREATE TABLE relfilenodes(node oid);
+
+ CREATE TABLE data_s1(i int, j int);
+ CREATE TABLE data_s2(i int, j int);
+}
+
+teardown
+{
+ DROP TABLE repack_deadlock;
+ DROP EXTENSION injection_points;
+
+ DROP TABLE relfilenodes;
+ DROP TABLE data_s1;
+ DROP TABLE data_s2;
+}
+
+session s1
+setup
+{
+ SELECT injection_points_set_local();
+ SELECT injection_points_attach('repack-concurrently-before-lock', 'wait');
+}
+# Perform the initial load and wait for s2 to do some data changes.
+step wait_before_lock
+{
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+}
+# Check the table from the perspective of s1.
+#
+# Besides the contents, we also check that relfilenode has changed.
+
+# Have each session write the contents into a table and use FULL JOIN to check
+# if the outputs are identical.
+step check1
+{
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+}
+teardown
+{
+ SELECT injection_points_detach('repack-concurrently-before-lock');
+}
+
+session s2
+# Change the existing data. UPDATE changes both key and non-key columns. Also
+# update one row twice to test whether tuple version generated by this session
+# can be found.
+step add_column
+{
+ alter table repack_deadlock add column noise text;
+}
+
+step wakeup_before_lock
+{
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+}
+
+# Test if data changes introduced while one session is performing REPACK
+# CONCURRENTLY find their way into the table.
+permutation
+ wait_before_lock
+ add_column
+ wakeup_before_lock
+ check1
--
2.47.3
--mdsv5bdgfjjj6d77
Content-Type: text/x-diff; charset=utf-8
Content-Disposition: attachment;
filename="v46-0007-Teach-snapshot-builder-to-skip-transactions-runn.patch"
^ permalink raw reply [nested|flat] 34+ messages in thread
* [PATCH v48 6/7] Error out any process that would block at REPACK
@ 2026-03-25 19:35 Álvaro Herrera <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Álvaro Herrera @ 2026-03-25 19:35 UTC (permalink / raw)
Any process waiting on REPACK to release its lock would actually cause
it to deadlock when it tries to upgrade its lock to AEL, losing all work
done to that point. We avoid this by teaching the deadlock detector to
raise an error when this condition is detected.
---
src/backend/commands/repack.c | 52 ++++++++----
src/backend/storage/lmgr/deadlock.c | 15 ++++
src/include/storage/proc.h | 6 +-
src/test/modules/injection_points/Makefile | 1 +
.../expected/repack_deadlock.out | 63 ++++++++++++++
src/test/modules/injection_points/meson.build | 1 +
.../specs/repack_deadlock.spec | 83 +++++++++++++++++++
7 files changed, 202 insertions(+), 19 deletions(-)
create mode 100644 src/test/modules/injection_points/expected/repack_deadlock.out
create mode 100644 src/test/modules/injection_points/specs/repack_deadlock.spec
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index a514053b777..781a471fc1d 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -276,6 +276,16 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
/* Determine the lock mode to use. */
lockmode = RepackLockLevel((params.options & CLUOPT_CONCURRENT) != 0);
+ /*
+ * If in concurrent mode, set the PROC_IN_CONCURRENT_REPACK flag. This
+ * makes the deadlock checker cause anyone that would conflict with us to
+ * error out. It's important to set this flag ahead of actually locking
+ * the relation; it won't of course affect anyone until we do have a lock
+ * that others can conflict with.
+ */
+ if ((params.options & CLUOPT_CONCURRENT) != 0)
+ MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+
/*
* If a single relation is specified, process it and we're done ... unless
* the relation is a partitioned table, in which case we fall through.
@@ -476,11 +486,8 @@ RepackLockLevel(bool concurrent)
* If indexOid is InvalidOid, the table will be rewritten in physical order
* instead of index order.
*
- * Note that, in the concurrent case, the function releases the lock at some
- * point, in order to get AccessExclusiveLock for the final steps (i.e. to
- * swap the relation files). To make things simpler, the caller should expect
- * OldHeap to be closed on return, regardless CLUOPT_CONCURRENT. (The
- * AccessExclusiveLock is kept till the end of the transaction.)
+ * On return, OldHeap is closed but locked with AccessExclusiveLock - the lock
+ * will be released at end of the transaction.
*
* 'cmd' indicates which command is being executed, to be used for error
* messages.
@@ -512,10 +519,12 @@ cluster_rel(RepackCommand cmd, Relation OldHeap, Oid indexOid,
/*
* Make sure we're not in a transaction block.
*
- * The reason is that repack_setup_logical_decoding() could deadlock
- * if there's an XID already assigned. It would be possible to run in
- * a transaction block if we had no XID, but this restriction is
- * simpler for users to understand and we don't lose anything.
+ * The reason is that repack_setup_logical_decoding() could wait
+ * indefinitely for our XID to complete. (The deadlock detector would
+ * not recognize it because we'd be waiting for ourselves, i.e. no
+ * real lock conflict.) It would be possible to run in a transaction
+ * block if we had no XID, but this restriction is simpler for users
+ * to understand and we don't lose anything.
*/
PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
@@ -998,10 +1007,8 @@ rebuild_relation(Relation OldHeap, Relation index, bool verbose,
* Note that the worker has to wait for all transactions with XID
* already assigned to finish. If some of those transactions is
* waiting for a lock conflicting with ShareUpdateExclusiveLock on our
- * table (e.g. it runs CREATE INDEX), we can end up in a deadlock.
- * Not sure this risk is worth unlocking/locking the table (and its
- * clustering index) and checking again if it's still eligible for
- * REPACK CONCURRENTLY.
+ * table (e.g. it runs CREATE INDEX), it should encounter ERROR in the
+ * deadlock checking code.
*/
start_repack_decoding_worker(tableOid);
@@ -3090,7 +3097,16 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
/*
- * Tuples and pages of the old heap will be gone, but the heap will stay.
+ * Now that we have all access-exclusive locks on all relations, we no
+ * longer want other processes to error out when trying to acquire a
+ * conflicting lock. Therefore, unset our flag.
+ */
+ MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+
+ /*
+ * Tuples and pages of the old heap will be gone, but the heap itself will
+ * stay. In order for predicate locks to continue to work, convert them
+ * to relation-level locks. We do this both for table and indexes.
*/
TransferPredicateLocksToHeapRelation(OldHeap);
foreach_ptr(RelationData, index, indexrels)
@@ -3288,9 +3304,11 @@ start_repack_decoding_worker(Oid relid)
/*
* The decoding setup must be done before the caller can have XID assigned
- * for any reason, otherwise the worker might end up in a deadlock,
- * waiting for the caller's transaction to end. Therefore wait here until
- * the worker indicates that it has the logical decoding initialized.
+ * for any reason, otherwise the worker might end up waiting for the
+ * caller's transaction to end. (Deadlock detector does not consider this
+ * a conflict because the worker is in the same locking group as the
+ * backend that launched it.) Therefore wait here until the worker
+ * indicates that it has the logical decoding initialized.
*/
ConditionVariablePrepareToSleep(&shared->cv);
for (;;)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..c20ac682b0d 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -620,6 +620,21 @@ FindLockCycleRecurseMember(PGPROC *checkProc,
proc->statusFlags & PROC_IS_AUTOVACUUM)
blocking_autovacuum_proc = proc;
+ /*
+ * Similarly, if we note that we're blocked by some
+ * process running REPACK (CONCURRENTLY), just fail. That
+ * process is going to upgrade its lock at some point, and
+ * it would be inappropriate for any other process to
+ * cause that to fail.
+ */
+ if (checkProc == MyProc &&
+ proc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+ ereport(ERROR,
+ errcode(ERRCODE_OBJECT_IN_USE),
+ errmsg("could not wait for concurrent REPACK"),
+ errdetail("Process %d waits for REPACK running on process %d",
+ MyProc->pid, proc->pid));
+
/* We're done looking at this proclock */
break;
}
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 1dad125706e..8ad9718f3d6 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -69,10 +69,12 @@ struct XidCache
#define PROC_AFFECTS_ALL_HORIZONS 0x20 /* this proc's xmin must be
* included in vacuum horizons
* in all databases */
+#define PROC_IN_CONCURRENT_REPACK 0x40 /* REPACK (CONCURRENTLY) */
-/* flags reset at EOXact */
+/* flags reset at EOXact. A bit of a misnomer ... */
#define PROC_VACUUM_STATE_MASK \
- (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+ (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+ PROC_IN_CONCURRENT_REPACK)
/*
* Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/Makefile b/src/test/modules/injection_points/Makefile
index 2cd7d87c533..f7663859fe2 100644
--- a/src/test/modules/injection_points/Makefile
+++ b/src/test/modules/injection_points/Makefile
@@ -15,6 +15,7 @@ REGRESS_OPTS = --dlpath=$(top_builddir)/src/test/regress
ISOLATION = basic \
inplace \
repack \
+ repack_deadlock \
repack_toast \
syscache-update-pruned \
heap_lock_update
diff --git a/src/test/modules/injection_points/expected/repack_deadlock.out b/src/test/modules/injection_points/expected/repack_deadlock.out
new file mode 100644
index 00000000000..a86e4767536
--- /dev/null
+++ b/src/test/modules/injection_points/expected/repack_deadlock.out
@@ -0,0 +1,63 @@
+Parsed test spec with 2 sessions
+
+starting permutation: wait_before_lock add_column wakeup_before_lock check1
+injection_points_attach
+-----------------------
+
+(1 row)
+
+step wait_before_lock:
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+ <waiting ...>
+step add_column:
+ alter table repack_deadlock add column noise text;
+ <waiting ...>
+step add_column: <... completed>
+ERROR: could not wait for concurrent REPACK
+step wakeup_before_lock:
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1:
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+
+count
+-----
+ 1
+(1 row)
+
+i|j
+-+-
+1|1
+2|2
+3|3
+4|4
+(4 rows)
+
+count
+-----
+ 4
+(1 row)
+
+injection_points_detach
+-----------------------
+
+(1 row)
+
diff --git a/src/test/modules/injection_points/meson.build b/src/test/modules/injection_points/meson.build
index a414abb924b..1cd88d6db65 100644
--- a/src/test/modules/injection_points/meson.build
+++ b/src/test/modules/injection_points/meson.build
@@ -46,6 +46,7 @@ tests += {
'basic',
'inplace',
'repack',
+ 'repack_deadlock',
'repack_toast',
'syscache-update-pruned',
'heap_lock_update',
diff --git a/src/test/modules/injection_points/specs/repack_deadlock.spec b/src/test/modules/injection_points/specs/repack_deadlock.spec
new file mode 100644
index 00000000000..9d23a6588c2
--- /dev/null
+++ b/src/test/modules/injection_points/specs/repack_deadlock.spec
@@ -0,0 +1,83 @@
+# Test REPACK with a concurrent transaction that would cause a deadlock
+setup
+{
+ CREATE EXTENSION injection_points;
+
+ CREATE TABLE repack_deadlock(i int PRIMARY KEY, j int);
+ INSERT INTO repack_deadlock(i, j) VALUES (1, 1), (2, 2), (3, 3), (4, 4);
+
+ CREATE TABLE relfilenodes(node oid);
+
+ CREATE TABLE data_s1(i int, j int);
+ CREATE TABLE data_s2(i int, j int);
+}
+
+teardown
+{
+ DROP TABLE repack_deadlock;
+ DROP EXTENSION injection_points;
+
+ DROP TABLE relfilenodes;
+ DROP TABLE data_s1;
+ DROP TABLE data_s2;
+}
+
+session s1
+setup
+{
+ SELECT injection_points_set_local();
+ SELECT injection_points_attach('repack-concurrently-before-lock', 'wait');
+}
+# Perform the initial load and wait for s2 to do some data changes.
+step wait_before_lock
+{
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+}
+# Check the table from the perspective of s1.
+#
+# Besides the contents, we also check that relfilenode has changed.
+
+# Have each session write the contents into a table and use FULL JOIN to check
+# if the outputs are identical.
+step check1
+{
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+}
+teardown
+{
+ SELECT injection_points_detach('repack-concurrently-before-lock');
+}
+
+session s2
+# Change the existing data. UPDATE changes both key and non-key columns. Also
+# update one row twice to test whether tuple version generated by this session
+# can be found.
+step add_column
+{
+ alter table repack_deadlock add column noise text;
+}
+
+step wakeup_before_lock
+{
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+}
+
+# Test if data changes introduced while one session is performing REPACK
+# CONCURRENTLY find their way into the table.
+permutation
+ wait_before_lock
+ add_column
+ wakeup_before_lock
+ check1
--
2.47.3
--qfkt2ktdpcfeypib
Content-Type: text/x-diff; charset=utf-8
Content-Disposition: attachment;
filename="v48-0007-Teach-snapshot-builder-to-skip-transactions-runn.patch"
^ permalink raw reply [nested|flat] 34+ messages in thread
* [PATCH v56 3/3] Error out any process that would block at REPACK
@ 2026-04-01 15:35 Antonin Houska <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Antonin Houska @ 2026-04-01 15:35 UTC (permalink / raw)
Any process waiting on REPACK to release its lock would actually cause
it to deadlock when it tries to upgrade its lock to AEL, losing all work
done to that point. We avoid this by teaching the deadlock detector to
raise an error when this condition is detected.
---
src/backend/commands/repack.c | 47 ++++++++---
src/backend/storage/lmgr/deadlock.c | 15 ++++
src/include/storage/proc.h | 6 +-
src/test/modules/injection_points/Makefile | 1 +
.../expected/repack_deadlock.out | 63 ++++++++++++++
src/test/modules/injection_points/meson.build | 1 +
.../specs/repack_deadlock.spec | 83 +++++++++++++++++++
7 files changed, 201 insertions(+), 15 deletions(-)
create mode 100644 src/test/modules/injection_points/expected/repack_deadlock.out
create mode 100644 src/test/modules/injection_points/specs/repack_deadlock.spec
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 5e97fcf3818..4dae0e8ebe0 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
* to understand and we don't lose any functionality.
*/
PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+ /*
+ * Also set the PROC_IN_CONCURRENT_REPACK flag. This makes the
+ * deadlock checker cause anyone that would conflict with us to error
+ * out. It's important to set this flag ahead of actually locking the
+ * relation; it won't of course affect anyone until we do have a lock
+ * that others can conflict with.
+ */
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
}
/*
@@ -489,11 +501,8 @@ RepackLockLevel(bool concurrent)
* If indexOid is InvalidOid, the table will be rewritten in physical order
* instead of index order.
*
- * Note that, in the concurrent case, the function releases the lock at some
- * point, in order to get AccessExclusiveLock for the final steps (i.e. to
- * swap the relation files). To make things simpler, the caller should expect
- * OldHeap to be closed on return, regardless CLUOPT_CONCURRENT. (The
- * AccessExclusiveLock is kept till the end of the transaction.)
+ * On return, OldHeap is closed but locked with AccessExclusiveLock - the lock
+ * will be released at end of the transaction.
*
* 'cmd' indicates which command is being executed, to be used for error
* messages.
@@ -1002,10 +1011,8 @@ rebuild_relation(Relation OldHeap, Relation index, bool verbose,
* Note that the worker has to wait for all transactions with XID
* already assigned to finish. If some of those transactions is
* waiting for a lock conflicting with ShareUpdateExclusiveLock on our
- * table (e.g. it runs CREATE INDEX), we can end up in a deadlock.
- * Not sure this risk is worth unlocking/locking the table (and its
- * clustering index) and checking again if it's still eligible for
- * REPACK CONCURRENTLY.
+ * table (e.g. it runs CREATE INDEX), it should encounter ERROR in the
+ * deadlock checking code.
*/
start_repack_decoding_worker(tableOid);
@@ -3090,7 +3097,19 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
/*
- * Tuples and pages of the old heap will be gone, but the heap will stay.
+ * Now that we have all access-exclusive locks on all relations, we no
+ * longer want other processes to error out when trying to acquire a
+ * conflicting lock. Therefore, unset our flag.
+ */
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+
+ /*
+ * Tuples and pages of the old heap will be gone, but the heap itself will
+ * stay. In order for predicate locks to continue to work, convert them
+ * to relation-level locks. We do this both for table and indexes.
*/
TransferPredicateLocksToHeapRelation(OldHeap);
foreach_ptr(RelationData, index, indexrels)
@@ -3364,9 +3383,11 @@ start_repack_decoding_worker(Oid relid)
/*
* The decoding setup must be done before the caller can have XID assigned
- * for any reason, otherwise the worker might end up in a deadlock,
- * waiting for the caller's transaction to end. Therefore wait here until
- * the worker indicates that it has the logical decoding initialized.
+ * for any reason, otherwise the worker might end up waiting for the
+ * caller's transaction to end. (Deadlock detector does not consider this
+ * a conflict because the worker is in the same locking group as the
+ * backend that launched it.) Therefore wait here until the worker
+ * indicates that it has the logical decoding initialized.
*/
ConditionVariablePrepareToSleep(&shared->cv);
for (;;)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..c20ac682b0d 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -620,6 +620,21 @@ FindLockCycleRecurseMember(PGPROC *checkProc,
proc->statusFlags & PROC_IS_AUTOVACUUM)
blocking_autovacuum_proc = proc;
+ /*
+ * Similarly, if we note that we're blocked by some
+ * process running REPACK (CONCURRENTLY), just fail. That
+ * process is going to upgrade its lock at some point, and
+ * it would be inappropriate for any other process to
+ * cause that to fail.
+ */
+ if (checkProc == MyProc &&
+ proc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+ ereport(ERROR,
+ errcode(ERRCODE_OBJECT_IN_USE),
+ errmsg("could not wait for concurrent REPACK"),
+ errdetail("Process %d waits for REPACK running on process %d",
+ MyProc->pid, proc->pid));
+
/* We're done looking at this proclock */
break;
}
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
#define PROC_AFFECTS_ALL_HORIZONS 0x20 /* this proc's xmin must be
* included in vacuum horizons
* in all databases */
+#define PROC_IN_CONCURRENT_REPACK 0x40 /* REPACK (CONCURRENTLY) */
-/* flags reset at EOXact */
+/* flags reset at EOXact. A bit of a misnomer ... */
#define PROC_VACUUM_STATE_MASK \
- (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+ (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+ PROC_IN_CONCURRENT_REPACK)
/*
* Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/Makefile b/src/test/modules/injection_points/Makefile
index 2cd7d87c533..f7663859fe2 100644
--- a/src/test/modules/injection_points/Makefile
+++ b/src/test/modules/injection_points/Makefile
@@ -15,6 +15,7 @@ REGRESS_OPTS = --dlpath=$(top_builddir)/src/test/regress
ISOLATION = basic \
inplace \
repack \
+ repack_deadlock \
repack_toast \
syscache-update-pruned \
heap_lock_update
diff --git a/src/test/modules/injection_points/expected/repack_deadlock.out b/src/test/modules/injection_points/expected/repack_deadlock.out
new file mode 100644
index 00000000000..a86e4767536
--- /dev/null
+++ b/src/test/modules/injection_points/expected/repack_deadlock.out
@@ -0,0 +1,63 @@
+Parsed test spec with 2 sessions
+
+starting permutation: wait_before_lock add_column wakeup_before_lock check1
+injection_points_attach
+-----------------------
+
+(1 row)
+
+step wait_before_lock:
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+ <waiting ...>
+step add_column:
+ alter table repack_deadlock add column noise text;
+ <waiting ...>
+step add_column: <... completed>
+ERROR: could not wait for concurrent REPACK
+step wakeup_before_lock:
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1:
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+
+count
+-----
+ 1
+(1 row)
+
+i|j
+-+-
+1|1
+2|2
+3|3
+4|4
+(4 rows)
+
+count
+-----
+ 4
+(1 row)
+
+injection_points_detach
+-----------------------
+
+(1 row)
+
diff --git a/src/test/modules/injection_points/meson.build b/src/test/modules/injection_points/meson.build
index a414abb924b..1cd88d6db65 100644
--- a/src/test/modules/injection_points/meson.build
+++ b/src/test/modules/injection_points/meson.build
@@ -46,6 +46,7 @@ tests += {
'basic',
'inplace',
'repack',
+ 'repack_deadlock',
'repack_toast',
'syscache-update-pruned',
'heap_lock_update',
diff --git a/src/test/modules/injection_points/specs/repack_deadlock.spec b/src/test/modules/injection_points/specs/repack_deadlock.spec
new file mode 100644
index 00000000000..9d23a6588c2
--- /dev/null
+++ b/src/test/modules/injection_points/specs/repack_deadlock.spec
@@ -0,0 +1,83 @@
+# Test REPACK with a concurrent transaction that would cause a deadlock
+setup
+{
+ CREATE EXTENSION injection_points;
+
+ CREATE TABLE repack_deadlock(i int PRIMARY KEY, j int);
+ INSERT INTO repack_deadlock(i, j) VALUES (1, 1), (2, 2), (3, 3), (4, 4);
+
+ CREATE TABLE relfilenodes(node oid);
+
+ CREATE TABLE data_s1(i int, j int);
+ CREATE TABLE data_s2(i int, j int);
+}
+
+teardown
+{
+ DROP TABLE repack_deadlock;
+ DROP EXTENSION injection_points;
+
+ DROP TABLE relfilenodes;
+ DROP TABLE data_s1;
+ DROP TABLE data_s2;
+}
+
+session s1
+setup
+{
+ SELECT injection_points_set_local();
+ SELECT injection_points_attach('repack-concurrently-before-lock', 'wait');
+}
+# Perform the initial load and wait for s2 to do some data changes.
+step wait_before_lock
+{
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+}
+# Check the table from the perspective of s1.
+#
+# Besides the contents, we also check that relfilenode has changed.
+
+# Have each session write the contents into a table and use FULL JOIN to check
+# if the outputs are identical.
+step check1
+{
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+}
+teardown
+{
+ SELECT injection_points_detach('repack-concurrently-before-lock');
+}
+
+session s2
+# Change the existing data. UPDATE changes both key and non-key columns. Also
+# update one row twice to test whether tuple version generated by this session
+# can be found.
+step add_column
+{
+ alter table repack_deadlock add column noise text;
+}
+
+step wakeup_before_lock
+{
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+}
+
+# Test if data changes introduced while one session is performing REPACK
+# CONCURRENTLY find their way into the table.
+permutation
+ wait_before_lock
+ add_column
+ wakeup_before_lock
+ check1
--
2.47.3
--qs77q6fpwolne4ds--
^ permalink raw reply [nested|flat] 34+ messages in thread
* [PATCH 1/2] Error out any process that would block at REPACK
@ 2026-04-01 15:35 Antonin Houska <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Antonin Houska @ 2026-04-01 15:35 UTC (permalink / raw)
Any process waiting on REPACK to release its lock would actually cause
it to deadlock when it tries to upgrade its lock to AEL, losing all work
done to that point. We avoid this by teaching the deadlock detector to
raise an error when this condition is detected.
---
src/backend/commands/repack.c | 47 ++++++++---
src/backend/storage/lmgr/deadlock.c | 15 ++++
src/include/storage/proc.h | 6 +-
src/test/modules/injection_points/Makefile | 1 +
.../expected/repack_deadlock.out | 63 ++++++++++++++
src/test/modules/injection_points/meson.build | 1 +
.../specs/repack_deadlock.spec | 83 +++++++++++++++++++
7 files changed, 201 insertions(+), 15 deletions(-)
create mode 100644 src/test/modules/injection_points/expected/repack_deadlock.out
create mode 100644 src/test/modules/injection_points/specs/repack_deadlock.spec
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 20dad22c4b7..a5f5df77291 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
* to understand and we don't lose any functionality.
*/
PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+ /*
+ * Also set the PROC_IN_CONCURRENT_REPACK flag. This makes the
+ * deadlock checker cause anyone that would conflict with us to error
+ * out. It's important to set this flag ahead of actually locking the
+ * relation; it won't of course affect anyone until we do have a lock
+ * that others can conflict with.
+ */
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
}
/*
@@ -489,11 +501,8 @@ RepackLockLevel(bool concurrent)
* If indexOid is InvalidOid, the table will be rewritten in physical order
* instead of index order.
*
- * Note that, in the concurrent case, the function releases the lock at some
- * point, in order to get AccessExclusiveLock for the final steps (i.e. to
- * swap the relation files). To make things simpler, the caller should expect
- * OldHeap to be closed on return, regardless CLUOPT_CONCURRENT. (The
- * AccessExclusiveLock is kept till the end of the transaction.)
+ * On return, OldHeap is closed but locked with AccessExclusiveLock - the lock
+ * will be released at end of the transaction.
*
* 'cmd' indicates which command is being executed, to be used for error
* messages.
@@ -1002,10 +1011,8 @@ rebuild_relation(Relation OldHeap, Relation index, bool verbose,
* Note that the worker has to wait for all transactions with XID
* already assigned to finish. If some of those transactions is
* waiting for a lock conflicting with ShareUpdateExclusiveLock on our
- * table (e.g. it runs CREATE INDEX), we can end up in a deadlock.
- * Not sure this risk is worth unlocking/locking the table (and its
- * clustering index) and checking again if it's still eligible for
- * REPACK CONCURRENTLY.
+ * table (e.g. it runs CREATE INDEX), it should encounter ERROR in the
+ * deadlock checking code.
*/
start_repack_decoding_worker(tableOid);
@@ -3090,7 +3097,19 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
/*
- * Tuples and pages of the old heap will be gone, but the heap will stay.
+ * Now that we have all access-exclusive locks on all relations, we no
+ * longer want other processes to error out when trying to acquire a
+ * conflicting lock. Therefore, unset our flag.
+ */
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+
+ /*
+ * Tuples and pages of the old heap will be gone, but the heap itself will
+ * stay. In order for predicate locks to continue to work, convert them
+ * to relation-level locks. We do this both for table and indexes.
*/
TransferPredicateLocksToHeapRelation(OldHeap);
foreach_ptr(RelationData, index, indexrels)
@@ -3364,9 +3383,11 @@ start_repack_decoding_worker(Oid relid)
/*
* The decoding setup must be done before the caller can have XID assigned
- * for any reason, otherwise the worker might end up in a deadlock,
- * waiting for the caller's transaction to end. Therefore wait here until
- * the worker indicates that it has the logical decoding initialized.
+ * for any reason, otherwise the worker might end up waiting for the
+ * caller's transaction to end. (Deadlock detector does not consider this
+ * a conflict because the worker is in the same locking group as the
+ * backend that launched it.) Therefore wait here until the worker
+ * indicates that it has the logical decoding initialized.
*/
ConditionVariablePrepareToSleep(&shared->cv);
for (;;)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..c20ac682b0d 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -620,6 +620,21 @@ FindLockCycleRecurseMember(PGPROC *checkProc,
proc->statusFlags & PROC_IS_AUTOVACUUM)
blocking_autovacuum_proc = proc;
+ /*
+ * Similarly, if we note that we're blocked by some
+ * process running REPACK (CONCURRENTLY), just fail. That
+ * process is going to upgrade its lock at some point, and
+ * it would be inappropriate for any other process to
+ * cause that to fail.
+ */
+ if (checkProc == MyProc &&
+ proc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+ ereport(ERROR,
+ errcode(ERRCODE_OBJECT_IN_USE),
+ errmsg("could not wait for concurrent REPACK"),
+ errdetail("Process %d waits for REPACK running on process %d",
+ MyProc->pid, proc->pid));
+
/* We're done looking at this proclock */
break;
}
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
#define PROC_AFFECTS_ALL_HORIZONS 0x20 /* this proc's xmin must be
* included in vacuum horizons
* in all databases */
+#define PROC_IN_CONCURRENT_REPACK 0x40 /* REPACK (CONCURRENTLY) */
-/* flags reset at EOXact */
+/* flags reset at EOXact. A bit of a misnomer ... */
#define PROC_VACUUM_STATE_MASK \
- (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+ (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+ PROC_IN_CONCURRENT_REPACK)
/*
* Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/Makefile b/src/test/modules/injection_points/Makefile
index f057d143d1a..13c873969d1 100644
--- a/src/test/modules/injection_points/Makefile
+++ b/src/test/modules/injection_points/Makefile
@@ -15,6 +15,7 @@ REGRESS_OPTS = --dlpath=$(top_builddir)/src/test/regress
ISOLATION = basic \
inplace \
repack \
+ repack_deadlock \
repack_toast \
syscache-update-pruned \
heap_lock_update
diff --git a/src/test/modules/injection_points/expected/repack_deadlock.out b/src/test/modules/injection_points/expected/repack_deadlock.out
new file mode 100644
index 00000000000..a86e4767536
--- /dev/null
+++ b/src/test/modules/injection_points/expected/repack_deadlock.out
@@ -0,0 +1,63 @@
+Parsed test spec with 2 sessions
+
+starting permutation: wait_before_lock add_column wakeup_before_lock check1
+injection_points_attach
+-----------------------
+
+(1 row)
+
+step wait_before_lock:
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+ <waiting ...>
+step add_column:
+ alter table repack_deadlock add column noise text;
+ <waiting ...>
+step add_column: <... completed>
+ERROR: could not wait for concurrent REPACK
+step wakeup_before_lock:
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1:
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+
+count
+-----
+ 1
+(1 row)
+
+i|j
+-+-
+1|1
+2|2
+3|3
+4|4
+(4 rows)
+
+count
+-----
+ 4
+(1 row)
+
+injection_points_detach
+-----------------------
+
+(1 row)
+
diff --git a/src/test/modules/injection_points/meson.build b/src/test/modules/injection_points/meson.build
index fb1418e2caa..ead18818c83 100644
--- a/src/test/modules/injection_points/meson.build
+++ b/src/test/modules/injection_points/meson.build
@@ -46,6 +46,7 @@ tests += {
'basic',
'inplace',
'repack',
+ 'repack_deadlock',
'repack_toast',
'syscache-update-pruned',
'heap_lock_update',
diff --git a/src/test/modules/injection_points/specs/repack_deadlock.spec b/src/test/modules/injection_points/specs/repack_deadlock.spec
new file mode 100644
index 00000000000..9d23a6588c2
--- /dev/null
+++ b/src/test/modules/injection_points/specs/repack_deadlock.spec
@@ -0,0 +1,83 @@
+# Test REPACK with a concurrent transaction that would cause a deadlock
+setup
+{
+ CREATE EXTENSION injection_points;
+
+ CREATE TABLE repack_deadlock(i int PRIMARY KEY, j int);
+ INSERT INTO repack_deadlock(i, j) VALUES (1, 1), (2, 2), (3, 3), (4, 4);
+
+ CREATE TABLE relfilenodes(node oid);
+
+ CREATE TABLE data_s1(i int, j int);
+ CREATE TABLE data_s2(i int, j int);
+}
+
+teardown
+{
+ DROP TABLE repack_deadlock;
+ DROP EXTENSION injection_points;
+
+ DROP TABLE relfilenodes;
+ DROP TABLE data_s1;
+ DROP TABLE data_s2;
+}
+
+session s1
+setup
+{
+ SELECT injection_points_set_local();
+ SELECT injection_points_attach('repack-concurrently-before-lock', 'wait');
+}
+# Perform the initial load and wait for s2 to do some data changes.
+step wait_before_lock
+{
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+}
+# Check the table from the perspective of s1.
+#
+# Besides the contents, we also check that relfilenode has changed.
+
+# Have each session write the contents into a table and use FULL JOIN to check
+# if the outputs are identical.
+step check1
+{
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+}
+teardown
+{
+ SELECT injection_points_detach('repack-concurrently-before-lock');
+}
+
+session s2
+# Change the existing data. UPDATE changes both key and non-key columns. Also
+# update one row twice to test whether tuple version generated by this session
+# can be found.
+step add_column
+{
+ alter table repack_deadlock add column noise text;
+}
+
+step wakeup_before_lock
+{
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+}
+
+# Test if data changes introduced while one session is performing REPACK
+# CONCURRENTLY find their way into the table.
+permutation
+ wait_before_lock
+ add_column
+ wakeup_before_lock
+ check1
--
2.47.3
--kdrcpfmkbkc4lqhu
Content-Type: text/x-diff; charset=utf-8
Content-Disposition: attachment;
filename="0002-Publish-list-of-tables-being-repacked-in-shared-memo.patch"
Content-Transfer-Encoding: 8bit
^ permalink raw reply [nested|flat] 34+ messages in thread
* [PATCH v53 6/7] Error out any process that would block at REPACK
@ 2026-04-01 15:35 Antonin Houska <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Antonin Houska @ 2026-04-01 15:35 UTC (permalink / raw)
Any process waiting on REPACK to release its lock would actually cause
it to deadlock when it tries to upgrade its lock to AEL, losing all work
done to that point. We avoid this by teaching the deadlock detector to
raise an error when this condition is detected.
---
src/backend/commands/repack.c | 47 ++++++++---
src/backend/storage/lmgr/deadlock.c | 15 ++++
src/include/storage/proc.h | 6 +-
src/test/modules/injection_points/Makefile | 1 +
.../expected/repack_deadlock.out | 63 ++++++++++++++
src/test/modules/injection_points/meson.build | 1 +
.../specs/repack_deadlock.spec | 83 +++++++++++++++++++
7 files changed, 201 insertions(+), 15 deletions(-)
create mode 100644 src/test/modules/injection_points/expected/repack_deadlock.out
create mode 100644 src/test/modules/injection_points/specs/repack_deadlock.spec
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 8d9b2b2e370..92759f33969 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -292,6 +292,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
* to understand and we don't lose any functionality.
*/
PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+ /*
+ * Also set the PROC_IN_CONCURRENT_REPACK flag. This makes the
+ * deadlock checker cause anyone that would conflict with us to error
+ * out. It's important to set this flag ahead of actually locking the
+ * relation; it won't of course affect anyone until we do have a lock
+ * that others can conflict with.
+ */
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
}
/*
@@ -494,11 +506,8 @@ RepackLockLevel(bool concurrent)
* If indexOid is InvalidOid, the table will be rewritten in physical order
* instead of index order.
*
- * Note that, in the concurrent case, the function releases the lock at some
- * point, in order to get AccessExclusiveLock for the final steps (i.e. to
- * swap the relation files). To make things simpler, the caller should expect
- * OldHeap to be closed on return, regardless CLUOPT_CONCURRENT. (The
- * AccessExclusiveLock is kept till the end of the transaction.)
+ * On return, OldHeap is closed but locked with AccessExclusiveLock - the lock
+ * will be released at end of the transaction.
*
* 'cmd' indicates which command is being executed, to be used for error
* messages.
@@ -1004,10 +1013,8 @@ rebuild_relation(Relation OldHeap, Relation index, bool verbose,
* Note that the worker has to wait for all transactions with XID
* already assigned to finish. If some of those transactions is
* waiting for a lock conflicting with ShareUpdateExclusiveLock on our
- * table (e.g. it runs CREATE INDEX), we can end up in a deadlock.
- * Not sure this risk is worth unlocking/locking the table (and its
- * clustering index) and checking again if it's still eligible for
- * REPACK CONCURRENTLY.
+ * table (e.g. it runs CREATE INDEX), it should encounter ERROR in the
+ * deadlock checking code.
*/
start_repack_decoding_worker(tableOid);
@@ -3096,7 +3103,19 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
/*
- * Tuples and pages of the old heap will be gone, but the heap will stay.
+ * Now that we have all access-exclusive locks on all relations, we no
+ * longer want other processes to error out when trying to acquire a
+ * conflicting lock. Therefore, unset our flag.
+ */
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+
+ /*
+ * Tuples and pages of the old heap will be gone, but the heap itself will
+ * stay. In order for predicate locks to continue to work, convert them
+ * to relation-level locks. We do this both for table and indexes.
*/
TransferPredicateLocksToHeapRelation(OldHeap);
foreach_ptr(RelationData, index, indexrels)
@@ -3369,9 +3388,11 @@ start_repack_decoding_worker(Oid relid)
/*
* The decoding setup must be done before the caller can have XID assigned
- * for any reason, otherwise the worker might end up in a deadlock,
- * waiting for the caller's transaction to end. Therefore wait here until
- * the worker indicates that it has the logical decoding initialized.
+ * for any reason, otherwise the worker might end up waiting for the
+ * caller's transaction to end. (Deadlock detector does not consider this
+ * a conflict because the worker is in the same locking group as the
+ * backend that launched it.) Therefore wait here until the worker
+ * indicates that it has the logical decoding initialized.
*/
ConditionVariablePrepareToSleep(&shared->cv);
for (;;)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..c20ac682b0d 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -620,6 +620,21 @@ FindLockCycleRecurseMember(PGPROC *checkProc,
proc->statusFlags & PROC_IS_AUTOVACUUM)
blocking_autovacuum_proc = proc;
+ /*
+ * Similarly, if we note that we're blocked by some
+ * process running REPACK (CONCURRENTLY), just fail. That
+ * process is going to upgrade its lock at some point, and
+ * it would be inappropriate for any other process to
+ * cause that to fail.
+ */
+ if (checkProc == MyProc &&
+ proc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+ ereport(ERROR,
+ errcode(ERRCODE_OBJECT_IN_USE),
+ errmsg("could not wait for concurrent REPACK"),
+ errdetail("Process %d waits for REPACK running on process %d",
+ MyProc->pid, proc->pid));
+
/* We're done looking at this proclock */
break;
}
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 22822fc68d7..a43cfe53558 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
#define PROC_AFFECTS_ALL_HORIZONS 0x20 /* this proc's xmin must be
* included in vacuum horizons
* in all databases */
+#define PROC_IN_CONCURRENT_REPACK 0x40 /* REPACK (CONCURRENTLY) */
-/* flags reset at EOXact */
+/* flags reset at EOXact. A bit of a misnomer ... */
#define PROC_VACUUM_STATE_MASK \
- (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+ (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+ PROC_IN_CONCURRENT_REPACK)
/*
* Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/Makefile b/src/test/modules/injection_points/Makefile
index 2cd7d87c533..f7663859fe2 100644
--- a/src/test/modules/injection_points/Makefile
+++ b/src/test/modules/injection_points/Makefile
@@ -15,6 +15,7 @@ REGRESS_OPTS = --dlpath=$(top_builddir)/src/test/regress
ISOLATION = basic \
inplace \
repack \
+ repack_deadlock \
repack_toast \
syscache-update-pruned \
heap_lock_update
diff --git a/src/test/modules/injection_points/expected/repack_deadlock.out b/src/test/modules/injection_points/expected/repack_deadlock.out
new file mode 100644
index 00000000000..a86e4767536
--- /dev/null
+++ b/src/test/modules/injection_points/expected/repack_deadlock.out
@@ -0,0 +1,63 @@
+Parsed test spec with 2 sessions
+
+starting permutation: wait_before_lock add_column wakeup_before_lock check1
+injection_points_attach
+-----------------------
+
+(1 row)
+
+step wait_before_lock:
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+ <waiting ...>
+step add_column:
+ alter table repack_deadlock add column noise text;
+ <waiting ...>
+step add_column: <... completed>
+ERROR: could not wait for concurrent REPACK
+step wakeup_before_lock:
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1:
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+
+count
+-----
+ 1
+(1 row)
+
+i|j
+-+-
+1|1
+2|2
+3|3
+4|4
+(4 rows)
+
+count
+-----
+ 4
+(1 row)
+
+injection_points_detach
+-----------------------
+
+(1 row)
+
diff --git a/src/test/modules/injection_points/meson.build b/src/test/modules/injection_points/meson.build
index a414abb924b..1cd88d6db65 100644
--- a/src/test/modules/injection_points/meson.build
+++ b/src/test/modules/injection_points/meson.build
@@ -46,6 +46,7 @@ tests += {
'basic',
'inplace',
'repack',
+ 'repack_deadlock',
'repack_toast',
'syscache-update-pruned',
'heap_lock_update',
diff --git a/src/test/modules/injection_points/specs/repack_deadlock.spec b/src/test/modules/injection_points/specs/repack_deadlock.spec
new file mode 100644
index 00000000000..9d23a6588c2
--- /dev/null
+++ b/src/test/modules/injection_points/specs/repack_deadlock.spec
@@ -0,0 +1,83 @@
+# Test REPACK with a concurrent transaction that would cause a deadlock
+setup
+{
+ CREATE EXTENSION injection_points;
+
+ CREATE TABLE repack_deadlock(i int PRIMARY KEY, j int);
+ INSERT INTO repack_deadlock(i, j) VALUES (1, 1), (2, 2), (3, 3), (4, 4);
+
+ CREATE TABLE relfilenodes(node oid);
+
+ CREATE TABLE data_s1(i int, j int);
+ CREATE TABLE data_s2(i int, j int);
+}
+
+teardown
+{
+ DROP TABLE repack_deadlock;
+ DROP EXTENSION injection_points;
+
+ DROP TABLE relfilenodes;
+ DROP TABLE data_s1;
+ DROP TABLE data_s2;
+}
+
+session s1
+setup
+{
+ SELECT injection_points_set_local();
+ SELECT injection_points_attach('repack-concurrently-before-lock', 'wait');
+}
+# Perform the initial load and wait for s2 to do some data changes.
+step wait_before_lock
+{
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+}
+# Check the table from the perspective of s1.
+#
+# Besides the contents, we also check that relfilenode has changed.
+
+# Have each session write the contents into a table and use FULL JOIN to check
+# if the outputs are identical.
+step check1
+{
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+}
+teardown
+{
+ SELECT injection_points_detach('repack-concurrently-before-lock');
+}
+
+session s2
+# Change the existing data. UPDATE changes both key and non-key columns. Also
+# update one row twice to test whether tuple version generated by this session
+# can be found.
+step add_column
+{
+ alter table repack_deadlock add column noise text;
+}
+
+step wakeup_before_lock
+{
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+}
+
+# Test if data changes introduced while one session is performing REPACK
+# CONCURRENTLY find their way into the table.
+permutation
+ wait_before_lock
+ add_column
+ wakeup_before_lock
+ check1
--
2.47.3
--qr3jlalmmcpkiodg
Content-Type: text/x-diff; charset=utf-8
Content-Disposition: attachment;
filename="v53-0007-Reserve-replication-slots-specifically-for-REPAC.patch"
^ permalink raw reply [nested|flat] 34+ messages in thread
* [PATCH v54 4/5] Error out any process that would block at REPACK
@ 2026-04-01 15:35 Antonin Houska <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Antonin Houska @ 2026-04-01 15:35 UTC (permalink / raw)
Any process waiting on REPACK to release its lock would actually cause
it to deadlock when it tries to upgrade its lock to AEL, losing all work
done to that point. We avoid this by teaching the deadlock detector to
raise an error when this condition is detected.
---
src/backend/commands/repack.c | 47 ++++++++---
src/backend/storage/lmgr/deadlock.c | 15 ++++
src/include/storage/proc.h | 6 +-
src/test/modules/injection_points/Makefile | 1 +
.../expected/repack_deadlock.out | 63 ++++++++++++++
src/test/modules/injection_points/meson.build | 1 +
.../specs/repack_deadlock.spec | 83 +++++++++++++++++++
7 files changed, 201 insertions(+), 15 deletions(-)
create mode 100644 src/test/modules/injection_points/expected/repack_deadlock.out
create mode 100644 src/test/modules/injection_points/specs/repack_deadlock.spec
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index b0898ce52d9..d1be6c60e88 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -292,6 +292,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
* to understand and we don't lose any functionality.
*/
PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+ /*
+ * Also set the PROC_IN_CONCURRENT_REPACK flag. This makes the
+ * deadlock checker cause anyone that would conflict with us to error
+ * out. It's important to set this flag ahead of actually locking the
+ * relation; it won't of course affect anyone until we do have a lock
+ * that others can conflict with.
+ */
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
}
/*
@@ -494,11 +506,8 @@ RepackLockLevel(bool concurrent)
* If indexOid is InvalidOid, the table will be rewritten in physical order
* instead of index order.
*
- * Note that, in the concurrent case, the function releases the lock at some
- * point, in order to get AccessExclusiveLock for the final steps (i.e. to
- * swap the relation files). To make things simpler, the caller should expect
- * OldHeap to be closed on return, regardless CLUOPT_CONCURRENT. (The
- * AccessExclusiveLock is kept till the end of the transaction.)
+ * On return, OldHeap is closed but locked with AccessExclusiveLock - the lock
+ * will be released at end of the transaction.
*
* 'cmd' indicates which command is being executed, to be used for error
* messages.
@@ -1004,10 +1013,8 @@ rebuild_relation(Relation OldHeap, Relation index, bool verbose,
* Note that the worker has to wait for all transactions with XID
* already assigned to finish. If some of those transactions is
* waiting for a lock conflicting with ShareUpdateExclusiveLock on our
- * table (e.g. it runs CREATE INDEX), we can end up in a deadlock.
- * Not sure this risk is worth unlocking/locking the table (and its
- * clustering index) and checking again if it's still eligible for
- * REPACK CONCURRENTLY.
+ * table (e.g. it runs CREATE INDEX), it should encounter ERROR in the
+ * deadlock checking code.
*/
start_repack_decoding_worker(tableOid);
@@ -3096,7 +3103,19 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
/*
- * Tuples and pages of the old heap will be gone, but the heap will stay.
+ * Now that we have all access-exclusive locks on all relations, we no
+ * longer want other processes to error out when trying to acquire a
+ * conflicting lock. Therefore, unset our flag.
+ */
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+
+ /*
+ * Tuples and pages of the old heap will be gone, but the heap itself will
+ * stay. In order for predicate locks to continue to work, convert them
+ * to relation-level locks. We do this both for table and indexes.
*/
TransferPredicateLocksToHeapRelation(OldHeap);
foreach_ptr(RelationData, index, indexrels)
@@ -3370,9 +3389,11 @@ start_repack_decoding_worker(Oid relid)
/*
* The decoding setup must be done before the caller can have XID assigned
- * for any reason, otherwise the worker might end up in a deadlock,
- * waiting for the caller's transaction to end. Therefore wait here until
- * the worker indicates that it has the logical decoding initialized.
+ * for any reason, otherwise the worker might end up waiting for the
+ * caller's transaction to end. (Deadlock detector does not consider this
+ * a conflict because the worker is in the same locking group as the
+ * backend that launched it.) Therefore wait here until the worker
+ * indicates that it has the logical decoding initialized.
*/
ConditionVariablePrepareToSleep(&shared->cv);
for (;;)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..c20ac682b0d 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -620,6 +620,21 @@ FindLockCycleRecurseMember(PGPROC *checkProc,
proc->statusFlags & PROC_IS_AUTOVACUUM)
blocking_autovacuum_proc = proc;
+ /*
+ * Similarly, if we note that we're blocked by some
+ * process running REPACK (CONCURRENTLY), just fail. That
+ * process is going to upgrade its lock at some point, and
+ * it would be inappropriate for any other process to
+ * cause that to fail.
+ */
+ if (checkProc == MyProc &&
+ proc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+ ereport(ERROR,
+ errcode(ERRCODE_OBJECT_IN_USE),
+ errmsg("could not wait for concurrent REPACK"),
+ errdetail("Process %d waits for REPACK running on process %d",
+ MyProc->pid, proc->pid));
+
/* We're done looking at this proclock */
break;
}
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 22822fc68d7..a43cfe53558 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
#define PROC_AFFECTS_ALL_HORIZONS 0x20 /* this proc's xmin must be
* included in vacuum horizons
* in all databases */
+#define PROC_IN_CONCURRENT_REPACK 0x40 /* REPACK (CONCURRENTLY) */
-/* flags reset at EOXact */
+/* flags reset at EOXact. A bit of a misnomer ... */
#define PROC_VACUUM_STATE_MASK \
- (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+ (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+ PROC_IN_CONCURRENT_REPACK)
/*
* Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/Makefile b/src/test/modules/injection_points/Makefile
index 2cd7d87c533..f7663859fe2 100644
--- a/src/test/modules/injection_points/Makefile
+++ b/src/test/modules/injection_points/Makefile
@@ -15,6 +15,7 @@ REGRESS_OPTS = --dlpath=$(top_builddir)/src/test/regress
ISOLATION = basic \
inplace \
repack \
+ repack_deadlock \
repack_toast \
syscache-update-pruned \
heap_lock_update
diff --git a/src/test/modules/injection_points/expected/repack_deadlock.out b/src/test/modules/injection_points/expected/repack_deadlock.out
new file mode 100644
index 00000000000..a86e4767536
--- /dev/null
+++ b/src/test/modules/injection_points/expected/repack_deadlock.out
@@ -0,0 +1,63 @@
+Parsed test spec with 2 sessions
+
+starting permutation: wait_before_lock add_column wakeup_before_lock check1
+injection_points_attach
+-----------------------
+
+(1 row)
+
+step wait_before_lock:
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+ <waiting ...>
+step add_column:
+ alter table repack_deadlock add column noise text;
+ <waiting ...>
+step add_column: <... completed>
+ERROR: could not wait for concurrent REPACK
+step wakeup_before_lock:
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1:
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+
+count
+-----
+ 1
+(1 row)
+
+i|j
+-+-
+1|1
+2|2
+3|3
+4|4
+(4 rows)
+
+count
+-----
+ 4
+(1 row)
+
+injection_points_detach
+-----------------------
+
+(1 row)
+
diff --git a/src/test/modules/injection_points/meson.build b/src/test/modules/injection_points/meson.build
index a414abb924b..1cd88d6db65 100644
--- a/src/test/modules/injection_points/meson.build
+++ b/src/test/modules/injection_points/meson.build
@@ -46,6 +46,7 @@ tests += {
'basic',
'inplace',
'repack',
+ 'repack_deadlock',
'repack_toast',
'syscache-update-pruned',
'heap_lock_update',
diff --git a/src/test/modules/injection_points/specs/repack_deadlock.spec b/src/test/modules/injection_points/specs/repack_deadlock.spec
new file mode 100644
index 00000000000..9d23a6588c2
--- /dev/null
+++ b/src/test/modules/injection_points/specs/repack_deadlock.spec
@@ -0,0 +1,83 @@
+# Test REPACK with a concurrent transaction that would cause a deadlock
+setup
+{
+ CREATE EXTENSION injection_points;
+
+ CREATE TABLE repack_deadlock(i int PRIMARY KEY, j int);
+ INSERT INTO repack_deadlock(i, j) VALUES (1, 1), (2, 2), (3, 3), (4, 4);
+
+ CREATE TABLE relfilenodes(node oid);
+
+ CREATE TABLE data_s1(i int, j int);
+ CREATE TABLE data_s2(i int, j int);
+}
+
+teardown
+{
+ DROP TABLE repack_deadlock;
+ DROP EXTENSION injection_points;
+
+ DROP TABLE relfilenodes;
+ DROP TABLE data_s1;
+ DROP TABLE data_s2;
+}
+
+session s1
+setup
+{
+ SELECT injection_points_set_local();
+ SELECT injection_points_attach('repack-concurrently-before-lock', 'wait');
+}
+# Perform the initial load and wait for s2 to do some data changes.
+step wait_before_lock
+{
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+}
+# Check the table from the perspective of s1.
+#
+# Besides the contents, we also check that relfilenode has changed.
+
+# Have each session write the contents into a table and use FULL JOIN to check
+# if the outputs are identical.
+step check1
+{
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+}
+teardown
+{
+ SELECT injection_points_detach('repack-concurrently-before-lock');
+}
+
+session s2
+# Change the existing data. UPDATE changes both key and non-key columns. Also
+# update one row twice to test whether tuple version generated by this session
+# can be found.
+step add_column
+{
+ alter table repack_deadlock add column noise text;
+}
+
+step wakeup_before_lock
+{
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+}
+
+# Test if data changes introduced while one session is performing REPACK
+# CONCURRENTLY find their way into the table.
+permutation
+ wait_before_lock
+ add_column
+ wakeup_before_lock
+ check1
--
2.47.3
--cdhh4t7ukb6tnjoq
Content-Type: text/x-diff; charset=utf-8
Content-Disposition: attachment;
filename="v54-0005-Reserve-replication-slots-specifically-for-REPAC.patch"
^ permalink raw reply [nested|flat] 34+ messages in thread
* [PATCH v56 3/3] Error out any process that would block at REPACK
@ 2026-04-01 15:35 Antonin Houska <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Antonin Houska @ 2026-04-01 15:35 UTC (permalink / raw)
Any process waiting on REPACK to release its lock would actually cause
it to deadlock when it tries to upgrade its lock to AEL, losing all work
done to that point. We avoid this by teaching the deadlock detector to
raise an error when this condition is detected.
---
src/backend/commands/repack.c | 47 ++++++++---
src/backend/storage/lmgr/deadlock.c | 15 ++++
src/include/storage/proc.h | 6 +-
src/test/modules/injection_points/Makefile | 1 +
.../expected/repack_deadlock.out | 63 ++++++++++++++
src/test/modules/injection_points/meson.build | 1 +
.../specs/repack_deadlock.spec | 83 +++++++++++++++++++
7 files changed, 201 insertions(+), 15 deletions(-)
create mode 100644 src/test/modules/injection_points/expected/repack_deadlock.out
create mode 100644 src/test/modules/injection_points/specs/repack_deadlock.spec
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 5e97fcf3818..4dae0e8ebe0 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
* to understand and we don't lose any functionality.
*/
PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+ /*
+ * Also set the PROC_IN_CONCURRENT_REPACK flag. This makes the
+ * deadlock checker cause anyone that would conflict with us to error
+ * out. It's important to set this flag ahead of actually locking the
+ * relation; it won't of course affect anyone until we do have a lock
+ * that others can conflict with.
+ */
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
}
/*
@@ -489,11 +501,8 @@ RepackLockLevel(bool concurrent)
* If indexOid is InvalidOid, the table will be rewritten in physical order
* instead of index order.
*
- * Note that, in the concurrent case, the function releases the lock at some
- * point, in order to get AccessExclusiveLock for the final steps (i.e. to
- * swap the relation files). To make things simpler, the caller should expect
- * OldHeap to be closed on return, regardless CLUOPT_CONCURRENT. (The
- * AccessExclusiveLock is kept till the end of the transaction.)
+ * On return, OldHeap is closed but locked with AccessExclusiveLock - the lock
+ * will be released at end of the transaction.
*
* 'cmd' indicates which command is being executed, to be used for error
* messages.
@@ -1002,10 +1011,8 @@ rebuild_relation(Relation OldHeap, Relation index, bool verbose,
* Note that the worker has to wait for all transactions with XID
* already assigned to finish. If some of those transactions is
* waiting for a lock conflicting with ShareUpdateExclusiveLock on our
- * table (e.g. it runs CREATE INDEX), we can end up in a deadlock.
- * Not sure this risk is worth unlocking/locking the table (and its
- * clustering index) and checking again if it's still eligible for
- * REPACK CONCURRENTLY.
+ * table (e.g. it runs CREATE INDEX), it should encounter ERROR in the
+ * deadlock checking code.
*/
start_repack_decoding_worker(tableOid);
@@ -3090,7 +3097,19 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
/*
- * Tuples and pages of the old heap will be gone, but the heap will stay.
+ * Now that we have all access-exclusive locks on all relations, we no
+ * longer want other processes to error out when trying to acquire a
+ * conflicting lock. Therefore, unset our flag.
+ */
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+
+ /*
+ * Tuples and pages of the old heap will be gone, but the heap itself will
+ * stay. In order for predicate locks to continue to work, convert them
+ * to relation-level locks. We do this both for table and indexes.
*/
TransferPredicateLocksToHeapRelation(OldHeap);
foreach_ptr(RelationData, index, indexrels)
@@ -3364,9 +3383,11 @@ start_repack_decoding_worker(Oid relid)
/*
* The decoding setup must be done before the caller can have XID assigned
- * for any reason, otherwise the worker might end up in a deadlock,
- * waiting for the caller's transaction to end. Therefore wait here until
- * the worker indicates that it has the logical decoding initialized.
+ * for any reason, otherwise the worker might end up waiting for the
+ * caller's transaction to end. (Deadlock detector does not consider this
+ * a conflict because the worker is in the same locking group as the
+ * backend that launched it.) Therefore wait here until the worker
+ * indicates that it has the logical decoding initialized.
*/
ConditionVariablePrepareToSleep(&shared->cv);
for (;;)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..c20ac682b0d 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -620,6 +620,21 @@ FindLockCycleRecurseMember(PGPROC *checkProc,
proc->statusFlags & PROC_IS_AUTOVACUUM)
blocking_autovacuum_proc = proc;
+ /*
+ * Similarly, if we note that we're blocked by some
+ * process running REPACK (CONCURRENTLY), just fail. That
+ * process is going to upgrade its lock at some point, and
+ * it would be inappropriate for any other process to
+ * cause that to fail.
+ */
+ if (checkProc == MyProc &&
+ proc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+ ereport(ERROR,
+ errcode(ERRCODE_OBJECT_IN_USE),
+ errmsg("could not wait for concurrent REPACK"),
+ errdetail("Process %d waits for REPACK running on process %d",
+ MyProc->pid, proc->pid));
+
/* We're done looking at this proclock */
break;
}
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
#define PROC_AFFECTS_ALL_HORIZONS 0x20 /* this proc's xmin must be
* included in vacuum horizons
* in all databases */
+#define PROC_IN_CONCURRENT_REPACK 0x40 /* REPACK (CONCURRENTLY) */
-/* flags reset at EOXact */
+/* flags reset at EOXact. A bit of a misnomer ... */
#define PROC_VACUUM_STATE_MASK \
- (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+ (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+ PROC_IN_CONCURRENT_REPACK)
/*
* Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/Makefile b/src/test/modules/injection_points/Makefile
index 2cd7d87c533..f7663859fe2 100644
--- a/src/test/modules/injection_points/Makefile
+++ b/src/test/modules/injection_points/Makefile
@@ -15,6 +15,7 @@ REGRESS_OPTS = --dlpath=$(top_builddir)/src/test/regress
ISOLATION = basic \
inplace \
repack \
+ repack_deadlock \
repack_toast \
syscache-update-pruned \
heap_lock_update
diff --git a/src/test/modules/injection_points/expected/repack_deadlock.out b/src/test/modules/injection_points/expected/repack_deadlock.out
new file mode 100644
index 00000000000..a86e4767536
--- /dev/null
+++ b/src/test/modules/injection_points/expected/repack_deadlock.out
@@ -0,0 +1,63 @@
+Parsed test spec with 2 sessions
+
+starting permutation: wait_before_lock add_column wakeup_before_lock check1
+injection_points_attach
+-----------------------
+
+(1 row)
+
+step wait_before_lock:
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+ <waiting ...>
+step add_column:
+ alter table repack_deadlock add column noise text;
+ <waiting ...>
+step add_column: <... completed>
+ERROR: could not wait for concurrent REPACK
+step wakeup_before_lock:
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1:
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+
+count
+-----
+ 1
+(1 row)
+
+i|j
+-+-
+1|1
+2|2
+3|3
+4|4
+(4 rows)
+
+count
+-----
+ 4
+(1 row)
+
+injection_points_detach
+-----------------------
+
+(1 row)
+
diff --git a/src/test/modules/injection_points/meson.build b/src/test/modules/injection_points/meson.build
index a414abb924b..1cd88d6db65 100644
--- a/src/test/modules/injection_points/meson.build
+++ b/src/test/modules/injection_points/meson.build
@@ -46,6 +46,7 @@ tests += {
'basic',
'inplace',
'repack',
+ 'repack_deadlock',
'repack_toast',
'syscache-update-pruned',
'heap_lock_update',
diff --git a/src/test/modules/injection_points/specs/repack_deadlock.spec b/src/test/modules/injection_points/specs/repack_deadlock.spec
new file mode 100644
index 00000000000..9d23a6588c2
--- /dev/null
+++ b/src/test/modules/injection_points/specs/repack_deadlock.spec
@@ -0,0 +1,83 @@
+# Test REPACK with a concurrent transaction that would cause a deadlock
+setup
+{
+ CREATE EXTENSION injection_points;
+
+ CREATE TABLE repack_deadlock(i int PRIMARY KEY, j int);
+ INSERT INTO repack_deadlock(i, j) VALUES (1, 1), (2, 2), (3, 3), (4, 4);
+
+ CREATE TABLE relfilenodes(node oid);
+
+ CREATE TABLE data_s1(i int, j int);
+ CREATE TABLE data_s2(i int, j int);
+}
+
+teardown
+{
+ DROP TABLE repack_deadlock;
+ DROP EXTENSION injection_points;
+
+ DROP TABLE relfilenodes;
+ DROP TABLE data_s1;
+ DROP TABLE data_s2;
+}
+
+session s1
+setup
+{
+ SELECT injection_points_set_local();
+ SELECT injection_points_attach('repack-concurrently-before-lock', 'wait');
+}
+# Perform the initial load and wait for s2 to do some data changes.
+step wait_before_lock
+{
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+}
+# Check the table from the perspective of s1.
+#
+# Besides the contents, we also check that relfilenode has changed.
+
+# Have each session write the contents into a table and use FULL JOIN to check
+# if the outputs are identical.
+step check1
+{
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+}
+teardown
+{
+ SELECT injection_points_detach('repack-concurrently-before-lock');
+}
+
+session s2
+# Change the existing data. UPDATE changes both key and non-key columns. Also
+# update one row twice to test whether tuple version generated by this session
+# can be found.
+step add_column
+{
+ alter table repack_deadlock add column noise text;
+}
+
+step wakeup_before_lock
+{
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+}
+
+# Test if data changes introduced while one session is performing REPACK
+# CONCURRENTLY find their way into the table.
+permutation
+ wait_before_lock
+ add_column
+ wakeup_before_lock
+ check1
--
2.47.3
--qs77q6fpwolne4ds--
^ permalink raw reply [nested|flat] 34+ messages in thread
* [PATCH 1/2] Error out any process that would block at REPACK
@ 2026-04-01 15:35 Antonin Houska <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Antonin Houska @ 2026-04-01 15:35 UTC (permalink / raw)
Any process waiting on REPACK to release its lock would actually cause
it to deadlock when it tries to upgrade its lock to AEL, losing all work
done to that point. We avoid this by teaching the deadlock detector to
raise an error when this condition is detected.
---
src/backend/commands/repack.c | 47 ++++++++---
src/backend/storage/lmgr/deadlock.c | 15 ++++
src/include/storage/proc.h | 6 +-
src/test/modules/injection_points/Makefile | 1 +
.../expected/repack_deadlock.out | 63 ++++++++++++++
src/test/modules/injection_points/meson.build | 1 +
.../specs/repack_deadlock.spec | 83 +++++++++++++++++++
7 files changed, 201 insertions(+), 15 deletions(-)
create mode 100644 src/test/modules/injection_points/expected/repack_deadlock.out
create mode 100644 src/test/modules/injection_points/specs/repack_deadlock.spec
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 20dad22c4b7..a5f5df77291 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
* to understand and we don't lose any functionality.
*/
PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+ /*
+ * Also set the PROC_IN_CONCURRENT_REPACK flag. This makes the
+ * deadlock checker cause anyone that would conflict with us to error
+ * out. It's important to set this flag ahead of actually locking the
+ * relation; it won't of course affect anyone until we do have a lock
+ * that others can conflict with.
+ */
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
}
/*
@@ -489,11 +501,8 @@ RepackLockLevel(bool concurrent)
* If indexOid is InvalidOid, the table will be rewritten in physical order
* instead of index order.
*
- * Note that, in the concurrent case, the function releases the lock at some
- * point, in order to get AccessExclusiveLock for the final steps (i.e. to
- * swap the relation files). To make things simpler, the caller should expect
- * OldHeap to be closed on return, regardless CLUOPT_CONCURRENT. (The
- * AccessExclusiveLock is kept till the end of the transaction.)
+ * On return, OldHeap is closed but locked with AccessExclusiveLock - the lock
+ * will be released at end of the transaction.
*
* 'cmd' indicates which command is being executed, to be used for error
* messages.
@@ -1002,10 +1011,8 @@ rebuild_relation(Relation OldHeap, Relation index, bool verbose,
* Note that the worker has to wait for all transactions with XID
* already assigned to finish. If some of those transactions is
* waiting for a lock conflicting with ShareUpdateExclusiveLock on our
- * table (e.g. it runs CREATE INDEX), we can end up in a deadlock.
- * Not sure this risk is worth unlocking/locking the table (and its
- * clustering index) and checking again if it's still eligible for
- * REPACK CONCURRENTLY.
+ * table (e.g. it runs CREATE INDEX), it should encounter ERROR in the
+ * deadlock checking code.
*/
start_repack_decoding_worker(tableOid);
@@ -3090,7 +3097,19 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
/*
- * Tuples and pages of the old heap will be gone, but the heap will stay.
+ * Now that we have all access-exclusive locks on all relations, we no
+ * longer want other processes to error out when trying to acquire a
+ * conflicting lock. Therefore, unset our flag.
+ */
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+
+ /*
+ * Tuples and pages of the old heap will be gone, but the heap itself will
+ * stay. In order for predicate locks to continue to work, convert them
+ * to relation-level locks. We do this both for table and indexes.
*/
TransferPredicateLocksToHeapRelation(OldHeap);
foreach_ptr(RelationData, index, indexrels)
@@ -3364,9 +3383,11 @@ start_repack_decoding_worker(Oid relid)
/*
* The decoding setup must be done before the caller can have XID assigned
- * for any reason, otherwise the worker might end up in a deadlock,
- * waiting for the caller's transaction to end. Therefore wait here until
- * the worker indicates that it has the logical decoding initialized.
+ * for any reason, otherwise the worker might end up waiting for the
+ * caller's transaction to end. (Deadlock detector does not consider this
+ * a conflict because the worker is in the same locking group as the
+ * backend that launched it.) Therefore wait here until the worker
+ * indicates that it has the logical decoding initialized.
*/
ConditionVariablePrepareToSleep(&shared->cv);
for (;;)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..c20ac682b0d 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -620,6 +620,21 @@ FindLockCycleRecurseMember(PGPROC *checkProc,
proc->statusFlags & PROC_IS_AUTOVACUUM)
blocking_autovacuum_proc = proc;
+ /*
+ * Similarly, if we note that we're blocked by some
+ * process running REPACK (CONCURRENTLY), just fail. That
+ * process is going to upgrade its lock at some point, and
+ * it would be inappropriate for any other process to
+ * cause that to fail.
+ */
+ if (checkProc == MyProc &&
+ proc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+ ereport(ERROR,
+ errcode(ERRCODE_OBJECT_IN_USE),
+ errmsg("could not wait for concurrent REPACK"),
+ errdetail("Process %d waits for REPACK running on process %d",
+ MyProc->pid, proc->pid));
+
/* We're done looking at this proclock */
break;
}
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
#define PROC_AFFECTS_ALL_HORIZONS 0x20 /* this proc's xmin must be
* included in vacuum horizons
* in all databases */
+#define PROC_IN_CONCURRENT_REPACK 0x40 /* REPACK (CONCURRENTLY) */
-/* flags reset at EOXact */
+/* flags reset at EOXact. A bit of a misnomer ... */
#define PROC_VACUUM_STATE_MASK \
- (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+ (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+ PROC_IN_CONCURRENT_REPACK)
/*
* Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/Makefile b/src/test/modules/injection_points/Makefile
index f057d143d1a..13c873969d1 100644
--- a/src/test/modules/injection_points/Makefile
+++ b/src/test/modules/injection_points/Makefile
@@ -15,6 +15,7 @@ REGRESS_OPTS = --dlpath=$(top_builddir)/src/test/regress
ISOLATION = basic \
inplace \
repack \
+ repack_deadlock \
repack_toast \
syscache-update-pruned \
heap_lock_update
diff --git a/src/test/modules/injection_points/expected/repack_deadlock.out b/src/test/modules/injection_points/expected/repack_deadlock.out
new file mode 100644
index 00000000000..a86e4767536
--- /dev/null
+++ b/src/test/modules/injection_points/expected/repack_deadlock.out
@@ -0,0 +1,63 @@
+Parsed test spec with 2 sessions
+
+starting permutation: wait_before_lock add_column wakeup_before_lock check1
+injection_points_attach
+-----------------------
+
+(1 row)
+
+step wait_before_lock:
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+ <waiting ...>
+step add_column:
+ alter table repack_deadlock add column noise text;
+ <waiting ...>
+step add_column: <... completed>
+ERROR: could not wait for concurrent REPACK
+step wakeup_before_lock:
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1:
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+
+count
+-----
+ 1
+(1 row)
+
+i|j
+-+-
+1|1
+2|2
+3|3
+4|4
+(4 rows)
+
+count
+-----
+ 4
+(1 row)
+
+injection_points_detach
+-----------------------
+
+(1 row)
+
diff --git a/src/test/modules/injection_points/meson.build b/src/test/modules/injection_points/meson.build
index fb1418e2caa..ead18818c83 100644
--- a/src/test/modules/injection_points/meson.build
+++ b/src/test/modules/injection_points/meson.build
@@ -46,6 +46,7 @@ tests += {
'basic',
'inplace',
'repack',
+ 'repack_deadlock',
'repack_toast',
'syscache-update-pruned',
'heap_lock_update',
diff --git a/src/test/modules/injection_points/specs/repack_deadlock.spec b/src/test/modules/injection_points/specs/repack_deadlock.spec
new file mode 100644
index 00000000000..9d23a6588c2
--- /dev/null
+++ b/src/test/modules/injection_points/specs/repack_deadlock.spec
@@ -0,0 +1,83 @@
+# Test REPACK with a concurrent transaction that would cause a deadlock
+setup
+{
+ CREATE EXTENSION injection_points;
+
+ CREATE TABLE repack_deadlock(i int PRIMARY KEY, j int);
+ INSERT INTO repack_deadlock(i, j) VALUES (1, 1), (2, 2), (3, 3), (4, 4);
+
+ CREATE TABLE relfilenodes(node oid);
+
+ CREATE TABLE data_s1(i int, j int);
+ CREATE TABLE data_s2(i int, j int);
+}
+
+teardown
+{
+ DROP TABLE repack_deadlock;
+ DROP EXTENSION injection_points;
+
+ DROP TABLE relfilenodes;
+ DROP TABLE data_s1;
+ DROP TABLE data_s2;
+}
+
+session s1
+setup
+{
+ SELECT injection_points_set_local();
+ SELECT injection_points_attach('repack-concurrently-before-lock', 'wait');
+}
+# Perform the initial load and wait for s2 to do some data changes.
+step wait_before_lock
+{
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+}
+# Check the table from the perspective of s1.
+#
+# Besides the contents, we also check that relfilenode has changed.
+
+# Have each session write the contents into a table and use FULL JOIN to check
+# if the outputs are identical.
+step check1
+{
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+}
+teardown
+{
+ SELECT injection_points_detach('repack-concurrently-before-lock');
+}
+
+session s2
+# Change the existing data. UPDATE changes both key and non-key columns. Also
+# update one row twice to test whether tuple version generated by this session
+# can be found.
+step add_column
+{
+ alter table repack_deadlock add column noise text;
+}
+
+step wakeup_before_lock
+{
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+}
+
+# Test if data changes introduced while one session is performing REPACK
+# CONCURRENTLY find their way into the table.
+permutation
+ wait_before_lock
+ add_column
+ wakeup_before_lock
+ check1
--
2.47.3
--kdrcpfmkbkc4lqhu
Content-Type: text/x-diff; charset=utf-8
Content-Disposition: attachment;
filename="0002-Publish-list-of-tables-being-repacked-in-shared-memo.patch"
Content-Transfer-Encoding: 8bit
^ permalink raw reply [nested|flat] 34+ messages in thread
* [PATCH] Error out any process that would block at REPACK
@ 2026-04-01 15:35 Antonin Houska <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Antonin Houska @ 2026-04-01 15:35 UTC (permalink / raw)
Any process waiting on REPACK to release its lock would actually cause
it to deadlock when it tries to upgrade its lock to AEL, losing all work
done to that point. We avoid this by teaching the deadlock detector to
raise an error when this condition is detected.
---
src/backend/commands/repack.c | 60 ++++++++++----
src/backend/storage/lmgr/deadlock.c | 15 ++++
src/include/storage/proc.h | 6 +-
src/test/modules/injection_points/Makefile | 1 +
.../expected/repack_deadlock.out | 63 ++++++++++++++
src/test/modules/injection_points/meson.build | 1 +
.../specs/repack_deadlock.spec | 83 +++++++++++++++++++
7 files changed, 210 insertions(+), 19 deletions(-)
create mode 100644 src/test/modules/injection_points/expected/repack_deadlock.out
create mode 100644 src/test/modules/injection_points/specs/repack_deadlock.spec
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index a514053b777..274ad900c65 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -276,6 +276,21 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
/* Determine the lock mode to use. */
lockmode = RepackLockLevel((params.options & CLUOPT_CONCURRENT) != 0);
+ /*
+ * If in concurrent mode, set the PROC_IN_CONCURRENT_REPACK flag. This
+ * makes the deadlock checker cause anyone that would conflict with us to
+ * error out. It's important to set this flag ahead of actually locking
+ * the relation; it won't of course affect anyone until we do have a lock
+ * that others can conflict with.
+ */
+ if ((params.options & CLUOPT_CONCURRENT) != 0)
+ {
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+ }
+
/*
* If a single relation is specified, process it and we're done ... unless
* the relation is a partitioned table, in which case we fall through.
@@ -476,11 +491,8 @@ RepackLockLevel(bool concurrent)
* If indexOid is InvalidOid, the table will be rewritten in physical order
* instead of index order.
*
- * Note that, in the concurrent case, the function releases the lock at some
- * point, in order to get AccessExclusiveLock for the final steps (i.e. to
- * swap the relation files). To make things simpler, the caller should expect
- * OldHeap to be closed on return, regardless CLUOPT_CONCURRENT. (The
- * AccessExclusiveLock is kept till the end of the transaction.)
+ * On return, OldHeap is closed but locked with AccessExclusiveLock - the lock
+ * will be released at end of the transaction.
*
* 'cmd' indicates which command is being executed, to be used for error
* messages.
@@ -512,10 +524,12 @@ cluster_rel(RepackCommand cmd, Relation OldHeap, Oid indexOid,
/*
* Make sure we're not in a transaction block.
*
- * The reason is that repack_setup_logical_decoding() could deadlock
- * if there's an XID already assigned. It would be possible to run in
- * a transaction block if we had no XID, but this restriction is
- * simpler for users to understand and we don't lose anything.
+ * The reason is that repack_setup_logical_decoding() could wait
+ * indefinitely for our XID to complete. (The deadlock detector would
+ * not recognize it because we'd be waiting for ourselves, i.e. no
+ * real lock conflict.) It would be possible to run in a transaction
+ * block if we had no XID, but this restriction is simpler for users
+ * to understand and we don't lose anything.
*/
PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
@@ -998,10 +1012,8 @@ rebuild_relation(Relation OldHeap, Relation index, bool verbose,
* Note that the worker has to wait for all transactions with XID
* already assigned to finish. If some of those transactions is
* waiting for a lock conflicting with ShareUpdateExclusiveLock on our
- * table (e.g. it runs CREATE INDEX), we can end up in a deadlock.
- * Not sure this risk is worth unlocking/locking the table (and its
- * clustering index) and checking again if it's still eligible for
- * REPACK CONCURRENTLY.
+ * table (e.g. it runs CREATE INDEX), it should encounter ERROR in the
+ * deadlock checking code.
*/
start_repack_decoding_worker(tableOid);
@@ -3090,7 +3102,19 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
/*
- * Tuples and pages of the old heap will be gone, but the heap will stay.
+ * Now that we have all access-exclusive locks on all relations, we no
+ * longer want other processes to error out when trying to acquire a
+ * conflicting lock. Therefore, unset our flag.
+ */
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+
+ /*
+ * Tuples and pages of the old heap will be gone, but the heap itself will
+ * stay. In order for predicate locks to continue to work, convert them
+ * to relation-level locks. We do this both for table and indexes.
*/
TransferPredicateLocksToHeapRelation(OldHeap);
foreach_ptr(RelationData, index, indexrels)
@@ -3288,9 +3312,11 @@ start_repack_decoding_worker(Oid relid)
/*
* The decoding setup must be done before the caller can have XID assigned
- * for any reason, otherwise the worker might end up in a deadlock,
- * waiting for the caller's transaction to end. Therefore wait here until
- * the worker indicates that it has the logical decoding initialized.
+ * for any reason, otherwise the worker might end up waiting for the
+ * caller's transaction to end. (Deadlock detector does not consider this
+ * a conflict because the worker is in the same locking group as the
+ * backend that launched it.) Therefore wait here until the worker
+ * indicates that it has the logical decoding initialized.
*/
ConditionVariablePrepareToSleep(&shared->cv);
for (;;)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..c20ac682b0d 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -620,6 +620,21 @@ FindLockCycleRecurseMember(PGPROC *checkProc,
proc->statusFlags & PROC_IS_AUTOVACUUM)
blocking_autovacuum_proc = proc;
+ /*
+ * Similarly, if we note that we're blocked by some
+ * process running REPACK (CONCURRENTLY), just fail. That
+ * process is going to upgrade its lock at some point, and
+ * it would be inappropriate for any other process to
+ * cause that to fail.
+ */
+ if (checkProc == MyProc &&
+ proc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+ ereport(ERROR,
+ errcode(ERRCODE_OBJECT_IN_USE),
+ errmsg("could not wait for concurrent REPACK"),
+ errdetail("Process %d waits for REPACK running on process %d",
+ MyProc->pid, proc->pid));
+
/* We're done looking at this proclock */
break;
}
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 1dad125706e..8ad9718f3d6 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -69,10 +69,12 @@ struct XidCache
#define PROC_AFFECTS_ALL_HORIZONS 0x20 /* this proc's xmin must be
* included in vacuum horizons
* in all databases */
+#define PROC_IN_CONCURRENT_REPACK 0x40 /* REPACK (CONCURRENTLY) */
-/* flags reset at EOXact */
+/* flags reset at EOXact. A bit of a misnomer ... */
#define PROC_VACUUM_STATE_MASK \
- (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+ (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+ PROC_IN_CONCURRENT_REPACK)
/*
* Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/Makefile b/src/test/modules/injection_points/Makefile
index 2cd7d87c533..f7663859fe2 100644
--- a/src/test/modules/injection_points/Makefile
+++ b/src/test/modules/injection_points/Makefile
@@ -15,6 +15,7 @@ REGRESS_OPTS = --dlpath=$(top_builddir)/src/test/regress
ISOLATION = basic \
inplace \
repack \
+ repack_deadlock \
repack_toast \
syscache-update-pruned \
heap_lock_update
diff --git a/src/test/modules/injection_points/expected/repack_deadlock.out b/src/test/modules/injection_points/expected/repack_deadlock.out
new file mode 100644
index 00000000000..a86e4767536
--- /dev/null
+++ b/src/test/modules/injection_points/expected/repack_deadlock.out
@@ -0,0 +1,63 @@
+Parsed test spec with 2 sessions
+
+starting permutation: wait_before_lock add_column wakeup_before_lock check1
+injection_points_attach
+-----------------------
+
+(1 row)
+
+step wait_before_lock:
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+ <waiting ...>
+step add_column:
+ alter table repack_deadlock add column noise text;
+ <waiting ...>
+step add_column: <... completed>
+ERROR: could not wait for concurrent REPACK
+step wakeup_before_lock:
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1:
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+
+count
+-----
+ 1
+(1 row)
+
+i|j
+-+-
+1|1
+2|2
+3|3
+4|4
+(4 rows)
+
+count
+-----
+ 4
+(1 row)
+
+injection_points_detach
+-----------------------
+
+(1 row)
+
diff --git a/src/test/modules/injection_points/meson.build b/src/test/modules/injection_points/meson.build
index a414abb924b..1cd88d6db65 100644
--- a/src/test/modules/injection_points/meson.build
+++ b/src/test/modules/injection_points/meson.build
@@ -46,6 +46,7 @@ tests += {
'basic',
'inplace',
'repack',
+ 'repack_deadlock',
'repack_toast',
'syscache-update-pruned',
'heap_lock_update',
diff --git a/src/test/modules/injection_points/specs/repack_deadlock.spec b/src/test/modules/injection_points/specs/repack_deadlock.spec
new file mode 100644
index 00000000000..9d23a6588c2
--- /dev/null
+++ b/src/test/modules/injection_points/specs/repack_deadlock.spec
@@ -0,0 +1,83 @@
+# Test REPACK with a concurrent transaction that would cause a deadlock
+setup
+{
+ CREATE EXTENSION injection_points;
+
+ CREATE TABLE repack_deadlock(i int PRIMARY KEY, j int);
+ INSERT INTO repack_deadlock(i, j) VALUES (1, 1), (2, 2), (3, 3), (4, 4);
+
+ CREATE TABLE relfilenodes(node oid);
+
+ CREATE TABLE data_s1(i int, j int);
+ CREATE TABLE data_s2(i int, j int);
+}
+
+teardown
+{
+ DROP TABLE repack_deadlock;
+ DROP EXTENSION injection_points;
+
+ DROP TABLE relfilenodes;
+ DROP TABLE data_s1;
+ DROP TABLE data_s2;
+}
+
+session s1
+setup
+{
+ SELECT injection_points_set_local();
+ SELECT injection_points_attach('repack-concurrently-before-lock', 'wait');
+}
+# Perform the initial load and wait for s2 to do some data changes.
+step wait_before_lock
+{
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+}
+# Check the table from the perspective of s1.
+#
+# Besides the contents, we also check that relfilenode has changed.
+
+# Have each session write the contents into a table and use FULL JOIN to check
+# if the outputs are identical.
+step check1
+{
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+}
+teardown
+{
+ SELECT injection_points_detach('repack-concurrently-before-lock');
+}
+
+session s2
+# Change the existing data. UPDATE changes both key and non-key columns. Also
+# update one row twice to test whether tuple version generated by this session
+# can be found.
+step add_column
+{
+ alter table repack_deadlock add column noise text;
+}
+
+step wakeup_before_lock
+{
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+}
+
+# Test if data changes introduced while one session is performing REPACK
+# CONCURRENTLY find their way into the table.
+permutation
+ wait_before_lock
+ add_column
+ wakeup_before_lock
+ check1
--
2.47.3
--=-=-=--
^ permalink raw reply [nested|flat] 34+ messages in thread
* [PATCH v49 5/7] Error out any process that would block at REPACK
@ 2026-04-01 15:35 Antonin Houska <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Antonin Houska @ 2026-04-01 15:35 UTC (permalink / raw)
Any process waiting on REPACK to release its lock would actually cause
it to deadlock when it tries to upgrade its lock to AEL, losing all work
done to that point. We avoid this by teaching the deadlock detector to
raise an error when this condition is detected.
---
src/backend/commands/repack.c | 60 ++++++++++----
src/backend/storage/lmgr/deadlock.c | 15 ++++
src/include/storage/proc.h | 6 +-
src/test/modules/injection_points/Makefile | 1 +
.../expected/repack_deadlock.out | 63 ++++++++++++++
src/test/modules/injection_points/meson.build | 1 +
.../specs/repack_deadlock.spec | 83 +++++++++++++++++++
7 files changed, 210 insertions(+), 19 deletions(-)
create mode 100644 src/test/modules/injection_points/expected/repack_deadlock.out
create mode 100644 src/test/modules/injection_points/specs/repack_deadlock.spec
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index a514053b777..274ad900c65 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -276,6 +276,21 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
/* Determine the lock mode to use. */
lockmode = RepackLockLevel((params.options & CLUOPT_CONCURRENT) != 0);
+ /*
+ * If in concurrent mode, set the PROC_IN_CONCURRENT_REPACK flag. This
+ * makes the deadlock checker cause anyone that would conflict with us to
+ * error out. It's important to set this flag ahead of actually locking
+ * the relation; it won't of course affect anyone until we do have a lock
+ * that others can conflict with.
+ */
+ if ((params.options & CLUOPT_CONCURRENT) != 0)
+ {
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+ }
+
/*
* If a single relation is specified, process it and we're done ... unless
* the relation is a partitioned table, in which case we fall through.
@@ -476,11 +491,8 @@ RepackLockLevel(bool concurrent)
* If indexOid is InvalidOid, the table will be rewritten in physical order
* instead of index order.
*
- * Note that, in the concurrent case, the function releases the lock at some
- * point, in order to get AccessExclusiveLock for the final steps (i.e. to
- * swap the relation files). To make things simpler, the caller should expect
- * OldHeap to be closed on return, regardless CLUOPT_CONCURRENT. (The
- * AccessExclusiveLock is kept till the end of the transaction.)
+ * On return, OldHeap is closed but locked with AccessExclusiveLock - the lock
+ * will be released at end of the transaction.
*
* 'cmd' indicates which command is being executed, to be used for error
* messages.
@@ -512,10 +524,12 @@ cluster_rel(RepackCommand cmd, Relation OldHeap, Oid indexOid,
/*
* Make sure we're not in a transaction block.
*
- * The reason is that repack_setup_logical_decoding() could deadlock
- * if there's an XID already assigned. It would be possible to run in
- * a transaction block if we had no XID, but this restriction is
- * simpler for users to understand and we don't lose anything.
+ * The reason is that repack_setup_logical_decoding() could wait
+ * indefinitely for our XID to complete. (The deadlock detector would
+ * not recognize it because we'd be waiting for ourselves, i.e. no
+ * real lock conflict.) It would be possible to run in a transaction
+ * block if we had no XID, but this restriction is simpler for users
+ * to understand and we don't lose anything.
*/
PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
@@ -998,10 +1012,8 @@ rebuild_relation(Relation OldHeap, Relation index, bool verbose,
* Note that the worker has to wait for all transactions with XID
* already assigned to finish. If some of those transactions is
* waiting for a lock conflicting with ShareUpdateExclusiveLock on our
- * table (e.g. it runs CREATE INDEX), we can end up in a deadlock.
- * Not sure this risk is worth unlocking/locking the table (and its
- * clustering index) and checking again if it's still eligible for
- * REPACK CONCURRENTLY.
+ * table (e.g. it runs CREATE INDEX), it should encounter ERROR in the
+ * deadlock checking code.
*/
start_repack_decoding_worker(tableOid);
@@ -3090,7 +3102,19 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
/*
- * Tuples and pages of the old heap will be gone, but the heap will stay.
+ * Now that we have all access-exclusive locks on all relations, we no
+ * longer want other processes to error out when trying to acquire a
+ * conflicting lock. Therefore, unset our flag.
+ */
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+
+ /*
+ * Tuples and pages of the old heap will be gone, but the heap itself will
+ * stay. In order for predicate locks to continue to work, convert them
+ * to relation-level locks. We do this both for table and indexes.
*/
TransferPredicateLocksToHeapRelation(OldHeap);
foreach_ptr(RelationData, index, indexrels)
@@ -3288,9 +3312,11 @@ start_repack_decoding_worker(Oid relid)
/*
* The decoding setup must be done before the caller can have XID assigned
- * for any reason, otherwise the worker might end up in a deadlock,
- * waiting for the caller's transaction to end. Therefore wait here until
- * the worker indicates that it has the logical decoding initialized.
+ * for any reason, otherwise the worker might end up waiting for the
+ * caller's transaction to end. (Deadlock detector does not consider this
+ * a conflict because the worker is in the same locking group as the
+ * backend that launched it.) Therefore wait here until the worker
+ * indicates that it has the logical decoding initialized.
*/
ConditionVariablePrepareToSleep(&shared->cv);
for (;;)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..c20ac682b0d 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -620,6 +620,21 @@ FindLockCycleRecurseMember(PGPROC *checkProc,
proc->statusFlags & PROC_IS_AUTOVACUUM)
blocking_autovacuum_proc = proc;
+ /*
+ * Similarly, if we note that we're blocked by some
+ * process running REPACK (CONCURRENTLY), just fail. That
+ * process is going to upgrade its lock at some point, and
+ * it would be inappropriate for any other process to
+ * cause that to fail.
+ */
+ if (checkProc == MyProc &&
+ proc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+ ereport(ERROR,
+ errcode(ERRCODE_OBJECT_IN_USE),
+ errmsg("could not wait for concurrent REPACK"),
+ errdetail("Process %d waits for REPACK running on process %d",
+ MyProc->pid, proc->pid));
+
/* We're done looking at this proclock */
break;
}
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 1dad125706e..8ad9718f3d6 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -69,10 +69,12 @@ struct XidCache
#define PROC_AFFECTS_ALL_HORIZONS 0x20 /* this proc's xmin must be
* included in vacuum horizons
* in all databases */
+#define PROC_IN_CONCURRENT_REPACK 0x40 /* REPACK (CONCURRENTLY) */
-/* flags reset at EOXact */
+/* flags reset at EOXact. A bit of a misnomer ... */
#define PROC_VACUUM_STATE_MASK \
- (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+ (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+ PROC_IN_CONCURRENT_REPACK)
/*
* Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/Makefile b/src/test/modules/injection_points/Makefile
index 2cd7d87c533..f7663859fe2 100644
--- a/src/test/modules/injection_points/Makefile
+++ b/src/test/modules/injection_points/Makefile
@@ -15,6 +15,7 @@ REGRESS_OPTS = --dlpath=$(top_builddir)/src/test/regress
ISOLATION = basic \
inplace \
repack \
+ repack_deadlock \
repack_toast \
syscache-update-pruned \
heap_lock_update
diff --git a/src/test/modules/injection_points/expected/repack_deadlock.out b/src/test/modules/injection_points/expected/repack_deadlock.out
new file mode 100644
index 00000000000..a86e4767536
--- /dev/null
+++ b/src/test/modules/injection_points/expected/repack_deadlock.out
@@ -0,0 +1,63 @@
+Parsed test spec with 2 sessions
+
+starting permutation: wait_before_lock add_column wakeup_before_lock check1
+injection_points_attach
+-----------------------
+
+(1 row)
+
+step wait_before_lock:
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+ <waiting ...>
+step add_column:
+ alter table repack_deadlock add column noise text;
+ <waiting ...>
+step add_column: <... completed>
+ERROR: could not wait for concurrent REPACK
+step wakeup_before_lock:
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1:
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+
+count
+-----
+ 1
+(1 row)
+
+i|j
+-+-
+1|1
+2|2
+3|3
+4|4
+(4 rows)
+
+count
+-----
+ 4
+(1 row)
+
+injection_points_detach
+-----------------------
+
+(1 row)
+
diff --git a/src/test/modules/injection_points/meson.build b/src/test/modules/injection_points/meson.build
index a414abb924b..1cd88d6db65 100644
--- a/src/test/modules/injection_points/meson.build
+++ b/src/test/modules/injection_points/meson.build
@@ -46,6 +46,7 @@ tests += {
'basic',
'inplace',
'repack',
+ 'repack_deadlock',
'repack_toast',
'syscache-update-pruned',
'heap_lock_update',
diff --git a/src/test/modules/injection_points/specs/repack_deadlock.spec b/src/test/modules/injection_points/specs/repack_deadlock.spec
new file mode 100644
index 00000000000..9d23a6588c2
--- /dev/null
+++ b/src/test/modules/injection_points/specs/repack_deadlock.spec
@@ -0,0 +1,83 @@
+# Test REPACK with a concurrent transaction that would cause a deadlock
+setup
+{
+ CREATE EXTENSION injection_points;
+
+ CREATE TABLE repack_deadlock(i int PRIMARY KEY, j int);
+ INSERT INTO repack_deadlock(i, j) VALUES (1, 1), (2, 2), (3, 3), (4, 4);
+
+ CREATE TABLE relfilenodes(node oid);
+
+ CREATE TABLE data_s1(i int, j int);
+ CREATE TABLE data_s2(i int, j int);
+}
+
+teardown
+{
+ DROP TABLE repack_deadlock;
+ DROP EXTENSION injection_points;
+
+ DROP TABLE relfilenodes;
+ DROP TABLE data_s1;
+ DROP TABLE data_s2;
+}
+
+session s1
+setup
+{
+ SELECT injection_points_set_local();
+ SELECT injection_points_attach('repack-concurrently-before-lock', 'wait');
+}
+# Perform the initial load and wait for s2 to do some data changes.
+step wait_before_lock
+{
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+}
+# Check the table from the perspective of s1.
+#
+# Besides the contents, we also check that relfilenode has changed.
+
+# Have each session write the contents into a table and use FULL JOIN to check
+# if the outputs are identical.
+step check1
+{
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+}
+teardown
+{
+ SELECT injection_points_detach('repack-concurrently-before-lock');
+}
+
+session s2
+# Change the existing data. UPDATE changes both key and non-key columns. Also
+# update one row twice to test whether tuple version generated by this session
+# can be found.
+step add_column
+{
+ alter table repack_deadlock add column noise text;
+}
+
+step wakeup_before_lock
+{
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+}
+
+# Test if data changes introduced while one session is performing REPACK
+# CONCURRENTLY find their way into the table.
+permutation
+ wait_before_lock
+ add_column
+ wakeup_before_lock
+ check1
--
2.47.3
--3n3vqr5y2jhknrqj
Content-Type: text/x-diff; charset=utf-8
Content-Disposition: attachment;
filename="v49-0006-Teach-snapshot-builder-to-skip-transactions-runn.patch"
^ permalink raw reply [nested|flat] 34+ messages in thread
* [PATCH v50 6/8] Error out any process that would block at REPACK
@ 2026-04-01 15:35 Antonin Houska <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Antonin Houska @ 2026-04-01 15:35 UTC (permalink / raw)
Any process waiting on REPACK to release its lock would actually cause
it to deadlock when it tries to upgrade its lock to AEL, losing all work
done to that point. We avoid this by teaching the deadlock detector to
raise an error when this condition is detected.
---
src/backend/commands/repack.c | 60 ++++++++++----
src/backend/storage/lmgr/deadlock.c | 15 ++++
src/include/storage/proc.h | 6 +-
src/test/modules/injection_points/Makefile | 1 +
.../expected/repack_deadlock.out | 63 ++++++++++++++
src/test/modules/injection_points/meson.build | 1 +
.../specs/repack_deadlock.spec | 83 +++++++++++++++++++
7 files changed, 210 insertions(+), 19 deletions(-)
create mode 100644 src/test/modules/injection_points/expected/repack_deadlock.out
create mode 100644 src/test/modules/injection_points/specs/repack_deadlock.spec
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 03829892d57..d6e446d582d 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -279,6 +279,21 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
/* Determine the lock mode to use. */
lockmode = RepackLockLevel((params.options & CLUOPT_CONCURRENT) != 0);
+ /*
+ * If in concurrent mode, set the PROC_IN_CONCURRENT_REPACK flag. This
+ * makes the deadlock checker cause anyone that would conflict with us to
+ * error out. It's important to set this flag ahead of actually locking
+ * the relation; it won't of course affect anyone until we do have a lock
+ * that others can conflict with.
+ */
+ if ((params.options & CLUOPT_CONCURRENT) != 0)
+ {
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+ }
+
/*
* If a single relation is specified, process it and we're done ... unless
* the relation is a partitioned table, in which case we fall through.
@@ -479,11 +494,8 @@ RepackLockLevel(bool concurrent)
* If indexOid is InvalidOid, the table will be rewritten in physical order
* instead of index order.
*
- * Note that, in the concurrent case, the function releases the lock at some
- * point, in order to get AccessExclusiveLock for the final steps (i.e. to
- * swap the relation files). To make things simpler, the caller should expect
- * OldHeap to be closed on return, regardless CLUOPT_CONCURRENT. (The
- * AccessExclusiveLock is kept till the end of the transaction.)
+ * On return, OldHeap is closed but locked with AccessExclusiveLock - the lock
+ * will be released at end of the transaction.
*
* 'cmd' indicates which command is being executed, to be used for error
* messages.
@@ -515,10 +527,12 @@ cluster_rel(RepackCommand cmd, Relation OldHeap, Oid indexOid,
/*
* Make sure we're not in a transaction block.
*
- * The reason is that repack_setup_logical_decoding() could deadlock
- * if there's an XID already assigned. It would be possible to run in
- * a transaction block if we had no XID, but this restriction is
- * simpler for users to understand and we don't lose anything.
+ * The reason is that repack_setup_logical_decoding() could wait
+ * indefinitely for our XID to complete. (The deadlock detector would
+ * not recognize it because we'd be waiting for ourselves, i.e. no
+ * real lock conflict.) It would be possible to run in a transaction
+ * block if we had no XID, but this restriction is simpler for users
+ * to understand and we don't lose anything.
*/
PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
@@ -1001,10 +1015,8 @@ rebuild_relation(Relation OldHeap, Relation index, bool verbose,
* Note that the worker has to wait for all transactions with XID
* already assigned to finish. If some of those transactions is
* waiting for a lock conflicting with ShareUpdateExclusiveLock on our
- * table (e.g. it runs CREATE INDEX), we can end up in a deadlock.
- * Not sure this risk is worth unlocking/locking the table (and its
- * clustering index) and checking again if it's still eligible for
- * REPACK CONCURRENTLY.
+ * table (e.g. it runs CREATE INDEX), it should encounter ERROR in the
+ * deadlock checking code.
*/
start_repack_decoding_worker(tableOid);
@@ -3093,7 +3105,19 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
/*
- * Tuples and pages of the old heap will be gone, but the heap will stay.
+ * Now that we have all access-exclusive locks on all relations, we no
+ * longer want other processes to error out when trying to acquire a
+ * conflicting lock. Therefore, unset our flag.
+ */
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+
+ /*
+ * Tuples and pages of the old heap will be gone, but the heap itself will
+ * stay. In order for predicate locks to continue to work, convert them
+ * to relation-level locks. We do this both for table and indexes.
*/
TransferPredicateLocksToHeapRelation(OldHeap);
foreach_ptr(RelationData, index, indexrels)
@@ -3366,9 +3390,11 @@ start_repack_decoding_worker(Oid relid)
/*
* The decoding setup must be done before the caller can have XID assigned
- * for any reason, otherwise the worker might end up in a deadlock,
- * waiting for the caller's transaction to end. Therefore wait here until
- * the worker indicates that it has the logical decoding initialized.
+ * for any reason, otherwise the worker might end up waiting for the
+ * caller's transaction to end. (Deadlock detector does not consider this
+ * a conflict because the worker is in the same locking group as the
+ * backend that launched it.) Therefore wait here until the worker
+ * indicates that it has the logical decoding initialized.
*/
ConditionVariablePrepareToSleep(&shared->cv);
for (;;)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..c20ac682b0d 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -620,6 +620,21 @@ FindLockCycleRecurseMember(PGPROC *checkProc,
proc->statusFlags & PROC_IS_AUTOVACUUM)
blocking_autovacuum_proc = proc;
+ /*
+ * Similarly, if we note that we're blocked by some
+ * process running REPACK (CONCURRENTLY), just fail. That
+ * process is going to upgrade its lock at some point, and
+ * it would be inappropriate for any other process to
+ * cause that to fail.
+ */
+ if (checkProc == MyProc &&
+ proc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+ ereport(ERROR,
+ errcode(ERRCODE_OBJECT_IN_USE),
+ errmsg("could not wait for concurrent REPACK"),
+ errdetail("Process %d waits for REPACK running on process %d",
+ MyProc->pid, proc->pid));
+
/* We're done looking at this proclock */
break;
}
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 1dad125706e..8ad9718f3d6 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -69,10 +69,12 @@ struct XidCache
#define PROC_AFFECTS_ALL_HORIZONS 0x20 /* this proc's xmin must be
* included in vacuum horizons
* in all databases */
+#define PROC_IN_CONCURRENT_REPACK 0x40 /* REPACK (CONCURRENTLY) */
-/* flags reset at EOXact */
+/* flags reset at EOXact. A bit of a misnomer ... */
#define PROC_VACUUM_STATE_MASK \
- (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+ (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+ PROC_IN_CONCURRENT_REPACK)
/*
* Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/Makefile b/src/test/modules/injection_points/Makefile
index 2cd7d87c533..f7663859fe2 100644
--- a/src/test/modules/injection_points/Makefile
+++ b/src/test/modules/injection_points/Makefile
@@ -15,6 +15,7 @@ REGRESS_OPTS = --dlpath=$(top_builddir)/src/test/regress
ISOLATION = basic \
inplace \
repack \
+ repack_deadlock \
repack_toast \
syscache-update-pruned \
heap_lock_update
diff --git a/src/test/modules/injection_points/expected/repack_deadlock.out b/src/test/modules/injection_points/expected/repack_deadlock.out
new file mode 100644
index 00000000000..a86e4767536
--- /dev/null
+++ b/src/test/modules/injection_points/expected/repack_deadlock.out
@@ -0,0 +1,63 @@
+Parsed test spec with 2 sessions
+
+starting permutation: wait_before_lock add_column wakeup_before_lock check1
+injection_points_attach
+-----------------------
+
+(1 row)
+
+step wait_before_lock:
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+ <waiting ...>
+step add_column:
+ alter table repack_deadlock add column noise text;
+ <waiting ...>
+step add_column: <... completed>
+ERROR: could not wait for concurrent REPACK
+step wakeup_before_lock:
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1:
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+
+count
+-----
+ 1
+(1 row)
+
+i|j
+-+-
+1|1
+2|2
+3|3
+4|4
+(4 rows)
+
+count
+-----
+ 4
+(1 row)
+
+injection_points_detach
+-----------------------
+
+(1 row)
+
diff --git a/src/test/modules/injection_points/meson.build b/src/test/modules/injection_points/meson.build
index a414abb924b..1cd88d6db65 100644
--- a/src/test/modules/injection_points/meson.build
+++ b/src/test/modules/injection_points/meson.build
@@ -46,6 +46,7 @@ tests += {
'basic',
'inplace',
'repack',
+ 'repack_deadlock',
'repack_toast',
'syscache-update-pruned',
'heap_lock_update',
diff --git a/src/test/modules/injection_points/specs/repack_deadlock.spec b/src/test/modules/injection_points/specs/repack_deadlock.spec
new file mode 100644
index 00000000000..9d23a6588c2
--- /dev/null
+++ b/src/test/modules/injection_points/specs/repack_deadlock.spec
@@ -0,0 +1,83 @@
+# Test REPACK with a concurrent transaction that would cause a deadlock
+setup
+{
+ CREATE EXTENSION injection_points;
+
+ CREATE TABLE repack_deadlock(i int PRIMARY KEY, j int);
+ INSERT INTO repack_deadlock(i, j) VALUES (1, 1), (2, 2), (3, 3), (4, 4);
+
+ CREATE TABLE relfilenodes(node oid);
+
+ CREATE TABLE data_s1(i int, j int);
+ CREATE TABLE data_s2(i int, j int);
+}
+
+teardown
+{
+ DROP TABLE repack_deadlock;
+ DROP EXTENSION injection_points;
+
+ DROP TABLE relfilenodes;
+ DROP TABLE data_s1;
+ DROP TABLE data_s2;
+}
+
+session s1
+setup
+{
+ SELECT injection_points_set_local();
+ SELECT injection_points_attach('repack-concurrently-before-lock', 'wait');
+}
+# Perform the initial load and wait for s2 to do some data changes.
+step wait_before_lock
+{
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+}
+# Check the table from the perspective of s1.
+#
+# Besides the contents, we also check that relfilenode has changed.
+
+# Have each session write the contents into a table and use FULL JOIN to check
+# if the outputs are identical.
+step check1
+{
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+}
+teardown
+{
+ SELECT injection_points_detach('repack-concurrently-before-lock');
+}
+
+session s2
+# Change the existing data. UPDATE changes both key and non-key columns. Also
+# update one row twice to test whether tuple version generated by this session
+# can be found.
+step add_column
+{
+ alter table repack_deadlock add column noise text;
+}
+
+step wakeup_before_lock
+{
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+}
+
+# Test if data changes introduced while one session is performing REPACK
+# CONCURRENTLY find their way into the table.
+permutation
+ wait_before_lock
+ add_column
+ wakeup_before_lock
+ check1
--
2.47.3
--brnevsqjnuzpyok4
Content-Type: text/x-diff; charset=utf-8
Content-Disposition: attachment;
filename="v50-0007-Introduce-an-option-to-make-logical-replication-.patch"
^ permalink raw reply [nested|flat] 34+ messages in thread
* [PATCH v51 06/10] Error out any process that would block at REPACK
@ 2026-04-01 15:35 Antonin Houska <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Antonin Houska @ 2026-04-01 15:35 UTC (permalink / raw)
Any process waiting on REPACK to release its lock would actually cause
it to deadlock when it tries to upgrade its lock to AEL, losing all work
done to that point. We avoid this by teaching the deadlock detector to
raise an error when this condition is detected.
---
src/backend/commands/repack.c | 60 ++++++++++----
src/backend/storage/lmgr/deadlock.c | 15 ++++
src/include/storage/proc.h | 6 +-
src/test/modules/injection_points/Makefile | 1 +
.../expected/repack_deadlock.out | 63 ++++++++++++++
src/test/modules/injection_points/meson.build | 1 +
.../specs/repack_deadlock.spec | 83 +++++++++++++++++++
7 files changed, 210 insertions(+), 19 deletions(-)
create mode 100644 src/test/modules/injection_points/expected/repack_deadlock.out
create mode 100644 src/test/modules/injection_points/specs/repack_deadlock.spec
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 03829892d57..d6e446d582d 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -279,6 +279,21 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
/* Determine the lock mode to use. */
lockmode = RepackLockLevel((params.options & CLUOPT_CONCURRENT) != 0);
+ /*
+ * If in concurrent mode, set the PROC_IN_CONCURRENT_REPACK flag. This
+ * makes the deadlock checker cause anyone that would conflict with us to
+ * error out. It's important to set this flag ahead of actually locking
+ * the relation; it won't of course affect anyone until we do have a lock
+ * that others can conflict with.
+ */
+ if ((params.options & CLUOPT_CONCURRENT) != 0)
+ {
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+ }
+
/*
* If a single relation is specified, process it and we're done ... unless
* the relation is a partitioned table, in which case we fall through.
@@ -479,11 +494,8 @@ RepackLockLevel(bool concurrent)
* If indexOid is InvalidOid, the table will be rewritten in physical order
* instead of index order.
*
- * Note that, in the concurrent case, the function releases the lock at some
- * point, in order to get AccessExclusiveLock for the final steps (i.e. to
- * swap the relation files). To make things simpler, the caller should expect
- * OldHeap to be closed on return, regardless CLUOPT_CONCURRENT. (The
- * AccessExclusiveLock is kept till the end of the transaction.)
+ * On return, OldHeap is closed but locked with AccessExclusiveLock - the lock
+ * will be released at end of the transaction.
*
* 'cmd' indicates which command is being executed, to be used for error
* messages.
@@ -515,10 +527,12 @@ cluster_rel(RepackCommand cmd, Relation OldHeap, Oid indexOid,
/*
* Make sure we're not in a transaction block.
*
- * The reason is that repack_setup_logical_decoding() could deadlock
- * if there's an XID already assigned. It would be possible to run in
- * a transaction block if we had no XID, but this restriction is
- * simpler for users to understand and we don't lose anything.
+ * The reason is that repack_setup_logical_decoding() could wait
+ * indefinitely for our XID to complete. (The deadlock detector would
+ * not recognize it because we'd be waiting for ourselves, i.e. no
+ * real lock conflict.) It would be possible to run in a transaction
+ * block if we had no XID, but this restriction is simpler for users
+ * to understand and we don't lose anything.
*/
PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
@@ -1001,10 +1015,8 @@ rebuild_relation(Relation OldHeap, Relation index, bool verbose,
* Note that the worker has to wait for all transactions with XID
* already assigned to finish. If some of those transactions is
* waiting for a lock conflicting with ShareUpdateExclusiveLock on our
- * table (e.g. it runs CREATE INDEX), we can end up in a deadlock.
- * Not sure this risk is worth unlocking/locking the table (and its
- * clustering index) and checking again if it's still eligible for
- * REPACK CONCURRENTLY.
+ * table (e.g. it runs CREATE INDEX), it should encounter ERROR in the
+ * deadlock checking code.
*/
start_repack_decoding_worker(tableOid);
@@ -3093,7 +3105,19 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
/*
- * Tuples and pages of the old heap will be gone, but the heap will stay.
+ * Now that we have all access-exclusive locks on all relations, we no
+ * longer want other processes to error out when trying to acquire a
+ * conflicting lock. Therefore, unset our flag.
+ */
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+
+ /*
+ * Tuples and pages of the old heap will be gone, but the heap itself will
+ * stay. In order for predicate locks to continue to work, convert them
+ * to relation-level locks. We do this both for table and indexes.
*/
TransferPredicateLocksToHeapRelation(OldHeap);
foreach_ptr(RelationData, index, indexrels)
@@ -3366,9 +3390,11 @@ start_repack_decoding_worker(Oid relid)
/*
* The decoding setup must be done before the caller can have XID assigned
- * for any reason, otherwise the worker might end up in a deadlock,
- * waiting for the caller's transaction to end. Therefore wait here until
- * the worker indicates that it has the logical decoding initialized.
+ * for any reason, otherwise the worker might end up waiting for the
+ * caller's transaction to end. (Deadlock detector does not consider this
+ * a conflict because the worker is in the same locking group as the
+ * backend that launched it.) Therefore wait here until the worker
+ * indicates that it has the logical decoding initialized.
*/
ConditionVariablePrepareToSleep(&shared->cv);
for (;;)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..c20ac682b0d 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -620,6 +620,21 @@ FindLockCycleRecurseMember(PGPROC *checkProc,
proc->statusFlags & PROC_IS_AUTOVACUUM)
blocking_autovacuum_proc = proc;
+ /*
+ * Similarly, if we note that we're blocked by some
+ * process running REPACK (CONCURRENTLY), just fail. That
+ * process is going to upgrade its lock at some point, and
+ * it would be inappropriate for any other process to
+ * cause that to fail.
+ */
+ if (checkProc == MyProc &&
+ proc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+ ereport(ERROR,
+ errcode(ERRCODE_OBJECT_IN_USE),
+ errmsg("could not wait for concurrent REPACK"),
+ errdetail("Process %d waits for REPACK running on process %d",
+ MyProc->pid, proc->pid));
+
/* We're done looking at this proclock */
break;
}
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 1dad125706e..8ad9718f3d6 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -69,10 +69,12 @@ struct XidCache
#define PROC_AFFECTS_ALL_HORIZONS 0x20 /* this proc's xmin must be
* included in vacuum horizons
* in all databases */
+#define PROC_IN_CONCURRENT_REPACK 0x40 /* REPACK (CONCURRENTLY) */
-/* flags reset at EOXact */
+/* flags reset at EOXact. A bit of a misnomer ... */
#define PROC_VACUUM_STATE_MASK \
- (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+ (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+ PROC_IN_CONCURRENT_REPACK)
/*
* Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/Makefile b/src/test/modules/injection_points/Makefile
index 2cd7d87c533..f7663859fe2 100644
--- a/src/test/modules/injection_points/Makefile
+++ b/src/test/modules/injection_points/Makefile
@@ -15,6 +15,7 @@ REGRESS_OPTS = --dlpath=$(top_builddir)/src/test/regress
ISOLATION = basic \
inplace \
repack \
+ repack_deadlock \
repack_toast \
syscache-update-pruned \
heap_lock_update
diff --git a/src/test/modules/injection_points/expected/repack_deadlock.out b/src/test/modules/injection_points/expected/repack_deadlock.out
new file mode 100644
index 00000000000..a86e4767536
--- /dev/null
+++ b/src/test/modules/injection_points/expected/repack_deadlock.out
@@ -0,0 +1,63 @@
+Parsed test spec with 2 sessions
+
+starting permutation: wait_before_lock add_column wakeup_before_lock check1
+injection_points_attach
+-----------------------
+
+(1 row)
+
+step wait_before_lock:
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+ <waiting ...>
+step add_column:
+ alter table repack_deadlock add column noise text;
+ <waiting ...>
+step add_column: <... completed>
+ERROR: could not wait for concurrent REPACK
+step wakeup_before_lock:
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1:
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+
+count
+-----
+ 1
+(1 row)
+
+i|j
+-+-
+1|1
+2|2
+3|3
+4|4
+(4 rows)
+
+count
+-----
+ 4
+(1 row)
+
+injection_points_detach
+-----------------------
+
+(1 row)
+
diff --git a/src/test/modules/injection_points/meson.build b/src/test/modules/injection_points/meson.build
index a414abb924b..1cd88d6db65 100644
--- a/src/test/modules/injection_points/meson.build
+++ b/src/test/modules/injection_points/meson.build
@@ -46,6 +46,7 @@ tests += {
'basic',
'inplace',
'repack',
+ 'repack_deadlock',
'repack_toast',
'syscache-update-pruned',
'heap_lock_update',
diff --git a/src/test/modules/injection_points/specs/repack_deadlock.spec b/src/test/modules/injection_points/specs/repack_deadlock.spec
new file mode 100644
index 00000000000..9d23a6588c2
--- /dev/null
+++ b/src/test/modules/injection_points/specs/repack_deadlock.spec
@@ -0,0 +1,83 @@
+# Test REPACK with a concurrent transaction that would cause a deadlock
+setup
+{
+ CREATE EXTENSION injection_points;
+
+ CREATE TABLE repack_deadlock(i int PRIMARY KEY, j int);
+ INSERT INTO repack_deadlock(i, j) VALUES (1, 1), (2, 2), (3, 3), (4, 4);
+
+ CREATE TABLE relfilenodes(node oid);
+
+ CREATE TABLE data_s1(i int, j int);
+ CREATE TABLE data_s2(i int, j int);
+}
+
+teardown
+{
+ DROP TABLE repack_deadlock;
+ DROP EXTENSION injection_points;
+
+ DROP TABLE relfilenodes;
+ DROP TABLE data_s1;
+ DROP TABLE data_s2;
+}
+
+session s1
+setup
+{
+ SELECT injection_points_set_local();
+ SELECT injection_points_attach('repack-concurrently-before-lock', 'wait');
+}
+# Perform the initial load and wait for s2 to do some data changes.
+step wait_before_lock
+{
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+}
+# Check the table from the perspective of s1.
+#
+# Besides the contents, we also check that relfilenode has changed.
+
+# Have each session write the contents into a table and use FULL JOIN to check
+# if the outputs are identical.
+step check1
+{
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+}
+teardown
+{
+ SELECT injection_points_detach('repack-concurrently-before-lock');
+}
+
+session s2
+# Change the existing data. UPDATE changes both key and non-key columns. Also
+# update one row twice to test whether tuple version generated by this session
+# can be found.
+step add_column
+{
+ alter table repack_deadlock add column noise text;
+}
+
+step wakeup_before_lock
+{
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+}
+
+# Test if data changes introduced while one session is performing REPACK
+# CONCURRENTLY find their way into the table.
+permutation
+ wait_before_lock
+ add_column
+ wakeup_before_lock
+ check1
--
2.47.3
--r2slln3zpmilwu22
Content-Type: text/x-diff; charset=utf-8
Content-Disposition: attachment;
filename="v51-0007-Check-for-transaction-block-early-in-ExecRepack.patch"
^ permalink raw reply [nested|flat] 34+ messages in thread
* [PATCH v52 06/10] Error out any process that would block at REPACK
@ 2026-04-01 15:35 Antonin Houska <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Antonin Houska @ 2026-04-01 15:35 UTC (permalink / raw)
Any process waiting on REPACK to release its lock would actually cause
it to deadlock when it tries to upgrade its lock to AEL, losing all work
done to that point. We avoid this by teaching the deadlock detector to
raise an error when this condition is detected.
---
src/backend/commands/repack.c | 60 ++++++++++----
src/backend/storage/lmgr/deadlock.c | 15 ++++
src/include/storage/proc.h | 6 +-
src/test/modules/injection_points/Makefile | 1 +
.../expected/repack_deadlock.out | 63 ++++++++++++++
src/test/modules/injection_points/meson.build | 1 +
.../specs/repack_deadlock.spec | 83 +++++++++++++++++++
7 files changed, 210 insertions(+), 19 deletions(-)
create mode 100644 src/test/modules/injection_points/expected/repack_deadlock.out
create mode 100644 src/test/modules/injection_points/specs/repack_deadlock.spec
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 03829892d57..d6e446d582d 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -279,6 +279,21 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
/* Determine the lock mode to use. */
lockmode = RepackLockLevel((params.options & CLUOPT_CONCURRENT) != 0);
+ /*
+ * If in concurrent mode, set the PROC_IN_CONCURRENT_REPACK flag. This
+ * makes the deadlock checker cause anyone that would conflict with us to
+ * error out. It's important to set this flag ahead of actually locking
+ * the relation; it won't of course affect anyone until we do have a lock
+ * that others can conflict with.
+ */
+ if ((params.options & CLUOPT_CONCURRENT) != 0)
+ {
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+ }
+
/*
* If a single relation is specified, process it and we're done ... unless
* the relation is a partitioned table, in which case we fall through.
@@ -479,11 +494,8 @@ RepackLockLevel(bool concurrent)
* If indexOid is InvalidOid, the table will be rewritten in physical order
* instead of index order.
*
- * Note that, in the concurrent case, the function releases the lock at some
- * point, in order to get AccessExclusiveLock for the final steps (i.e. to
- * swap the relation files). To make things simpler, the caller should expect
- * OldHeap to be closed on return, regardless CLUOPT_CONCURRENT. (The
- * AccessExclusiveLock is kept till the end of the transaction.)
+ * On return, OldHeap is closed but locked with AccessExclusiveLock - the lock
+ * will be released at end of the transaction.
*
* 'cmd' indicates which command is being executed, to be used for error
* messages.
@@ -515,10 +527,12 @@ cluster_rel(RepackCommand cmd, Relation OldHeap, Oid indexOid,
/*
* Make sure we're not in a transaction block.
*
- * The reason is that repack_setup_logical_decoding() could deadlock
- * if there's an XID already assigned. It would be possible to run in
- * a transaction block if we had no XID, but this restriction is
- * simpler for users to understand and we don't lose anything.
+ * The reason is that repack_setup_logical_decoding() could wait
+ * indefinitely for our XID to complete. (The deadlock detector would
+ * not recognize it because we'd be waiting for ourselves, i.e. no
+ * real lock conflict.) It would be possible to run in a transaction
+ * block if we had no XID, but this restriction is simpler for users
+ * to understand and we don't lose anything.
*/
PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
@@ -1001,10 +1015,8 @@ rebuild_relation(Relation OldHeap, Relation index, bool verbose,
* Note that the worker has to wait for all transactions with XID
* already assigned to finish. If some of those transactions is
* waiting for a lock conflicting with ShareUpdateExclusiveLock on our
- * table (e.g. it runs CREATE INDEX), we can end up in a deadlock.
- * Not sure this risk is worth unlocking/locking the table (and its
- * clustering index) and checking again if it's still eligible for
- * REPACK CONCURRENTLY.
+ * table (e.g. it runs CREATE INDEX), it should encounter ERROR in the
+ * deadlock checking code.
*/
start_repack_decoding_worker(tableOid);
@@ -3093,7 +3105,19 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
/*
- * Tuples and pages of the old heap will be gone, but the heap will stay.
+ * Now that we have all access-exclusive locks on all relations, we no
+ * longer want other processes to error out when trying to acquire a
+ * conflicting lock. Therefore, unset our flag.
+ */
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+
+ /*
+ * Tuples and pages of the old heap will be gone, but the heap itself will
+ * stay. In order for predicate locks to continue to work, convert them
+ * to relation-level locks. We do this both for table and indexes.
*/
TransferPredicateLocksToHeapRelation(OldHeap);
foreach_ptr(RelationData, index, indexrels)
@@ -3366,9 +3390,11 @@ start_repack_decoding_worker(Oid relid)
/*
* The decoding setup must be done before the caller can have XID assigned
- * for any reason, otherwise the worker might end up in a deadlock,
- * waiting for the caller's transaction to end. Therefore wait here until
- * the worker indicates that it has the logical decoding initialized.
+ * for any reason, otherwise the worker might end up waiting for the
+ * caller's transaction to end. (Deadlock detector does not consider this
+ * a conflict because the worker is in the same locking group as the
+ * backend that launched it.) Therefore wait here until the worker
+ * indicates that it has the logical decoding initialized.
*/
ConditionVariablePrepareToSleep(&shared->cv);
for (;;)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..c20ac682b0d 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -620,6 +620,21 @@ FindLockCycleRecurseMember(PGPROC *checkProc,
proc->statusFlags & PROC_IS_AUTOVACUUM)
blocking_autovacuum_proc = proc;
+ /*
+ * Similarly, if we note that we're blocked by some
+ * process running REPACK (CONCURRENTLY), just fail. That
+ * process is going to upgrade its lock at some point, and
+ * it would be inappropriate for any other process to
+ * cause that to fail.
+ */
+ if (checkProc == MyProc &&
+ proc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+ ereport(ERROR,
+ errcode(ERRCODE_OBJECT_IN_USE),
+ errmsg("could not wait for concurrent REPACK"),
+ errdetail("Process %d waits for REPACK running on process %d",
+ MyProc->pid, proc->pid));
+
/* We're done looking at this proclock */
break;
}
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 1dad125706e..8ad9718f3d6 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -69,10 +69,12 @@ struct XidCache
#define PROC_AFFECTS_ALL_HORIZONS 0x20 /* this proc's xmin must be
* included in vacuum horizons
* in all databases */
+#define PROC_IN_CONCURRENT_REPACK 0x40 /* REPACK (CONCURRENTLY) */
-/* flags reset at EOXact */
+/* flags reset at EOXact. A bit of a misnomer ... */
#define PROC_VACUUM_STATE_MASK \
- (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+ (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+ PROC_IN_CONCURRENT_REPACK)
/*
* Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/Makefile b/src/test/modules/injection_points/Makefile
index 2cd7d87c533..f7663859fe2 100644
--- a/src/test/modules/injection_points/Makefile
+++ b/src/test/modules/injection_points/Makefile
@@ -15,6 +15,7 @@ REGRESS_OPTS = --dlpath=$(top_builddir)/src/test/regress
ISOLATION = basic \
inplace \
repack \
+ repack_deadlock \
repack_toast \
syscache-update-pruned \
heap_lock_update
diff --git a/src/test/modules/injection_points/expected/repack_deadlock.out b/src/test/modules/injection_points/expected/repack_deadlock.out
new file mode 100644
index 00000000000..a86e4767536
--- /dev/null
+++ b/src/test/modules/injection_points/expected/repack_deadlock.out
@@ -0,0 +1,63 @@
+Parsed test spec with 2 sessions
+
+starting permutation: wait_before_lock add_column wakeup_before_lock check1
+injection_points_attach
+-----------------------
+
+(1 row)
+
+step wait_before_lock:
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+ <waiting ...>
+step add_column:
+ alter table repack_deadlock add column noise text;
+ <waiting ...>
+step add_column: <... completed>
+ERROR: could not wait for concurrent REPACK
+step wakeup_before_lock:
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1:
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+
+count
+-----
+ 1
+(1 row)
+
+i|j
+-+-
+1|1
+2|2
+3|3
+4|4
+(4 rows)
+
+count
+-----
+ 4
+(1 row)
+
+injection_points_detach
+-----------------------
+
+(1 row)
+
diff --git a/src/test/modules/injection_points/meson.build b/src/test/modules/injection_points/meson.build
index a414abb924b..1cd88d6db65 100644
--- a/src/test/modules/injection_points/meson.build
+++ b/src/test/modules/injection_points/meson.build
@@ -46,6 +46,7 @@ tests += {
'basic',
'inplace',
'repack',
+ 'repack_deadlock',
'repack_toast',
'syscache-update-pruned',
'heap_lock_update',
diff --git a/src/test/modules/injection_points/specs/repack_deadlock.spec b/src/test/modules/injection_points/specs/repack_deadlock.spec
new file mode 100644
index 00000000000..9d23a6588c2
--- /dev/null
+++ b/src/test/modules/injection_points/specs/repack_deadlock.spec
@@ -0,0 +1,83 @@
+# Test REPACK with a concurrent transaction that would cause a deadlock
+setup
+{
+ CREATE EXTENSION injection_points;
+
+ CREATE TABLE repack_deadlock(i int PRIMARY KEY, j int);
+ INSERT INTO repack_deadlock(i, j) VALUES (1, 1), (2, 2), (3, 3), (4, 4);
+
+ CREATE TABLE relfilenodes(node oid);
+
+ CREATE TABLE data_s1(i int, j int);
+ CREATE TABLE data_s2(i int, j int);
+}
+
+teardown
+{
+ DROP TABLE repack_deadlock;
+ DROP EXTENSION injection_points;
+
+ DROP TABLE relfilenodes;
+ DROP TABLE data_s1;
+ DROP TABLE data_s2;
+}
+
+session s1
+setup
+{
+ SELECT injection_points_set_local();
+ SELECT injection_points_attach('repack-concurrently-before-lock', 'wait');
+}
+# Perform the initial load and wait for s2 to do some data changes.
+step wait_before_lock
+{
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+}
+# Check the table from the perspective of s1.
+#
+# Besides the contents, we also check that relfilenode has changed.
+
+# Have each session write the contents into a table and use FULL JOIN to check
+# if the outputs are identical.
+step check1
+{
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+}
+teardown
+{
+ SELECT injection_points_detach('repack-concurrently-before-lock');
+}
+
+session s2
+# Change the existing data. UPDATE changes both key and non-key columns. Also
+# update one row twice to test whether tuple version generated by this session
+# can be found.
+step add_column
+{
+ alter table repack_deadlock add column noise text;
+}
+
+step wakeup_before_lock
+{
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+}
+
+# Test if data changes introduced while one session is performing REPACK
+# CONCURRENTLY find their way into the table.
+permutation
+ wait_before_lock
+ add_column
+ wakeup_before_lock
+ check1
--
2.47.3
--gp2pyozrd5pweboh
Content-Type: text/x-diff; charset=utf-8
Content-Disposition: attachment;
filename="v52-0007-Check-for-transaction-block-early-in-ExecRepack.patch"
^ permalink raw reply [nested|flat] 34+ messages in thread
* [PATCH] Error out any process that would block at REPACK
@ 2026-04-01 15:35 Antonin Houska <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Antonin Houska @ 2026-04-01 15:35 UTC (permalink / raw)
Any process waiting on REPACK to release its lock would actually cause
it to deadlock when it tries to upgrade its lock to AEL, losing all work
done to that point. We avoid this by teaching the deadlock detector to
raise an error when this condition is detected.
---
src/backend/commands/repack.c | 60 ++++++++++----
src/backend/storage/lmgr/deadlock.c | 15 ++++
src/include/storage/proc.h | 6 +-
src/test/modules/injection_points/Makefile | 1 +
.../expected/repack_deadlock.out | 63 ++++++++++++++
src/test/modules/injection_points/meson.build | 1 +
.../specs/repack_deadlock.spec | 83 +++++++++++++++++++
7 files changed, 210 insertions(+), 19 deletions(-)
create mode 100644 src/test/modules/injection_points/expected/repack_deadlock.out
create mode 100644 src/test/modules/injection_points/specs/repack_deadlock.spec
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index a514053b777..274ad900c65 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -276,6 +276,21 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
/* Determine the lock mode to use. */
lockmode = RepackLockLevel((params.options & CLUOPT_CONCURRENT) != 0);
+ /*
+ * If in concurrent mode, set the PROC_IN_CONCURRENT_REPACK flag. This
+ * makes the deadlock checker cause anyone that would conflict with us to
+ * error out. It's important to set this flag ahead of actually locking
+ * the relation; it won't of course affect anyone until we do have a lock
+ * that others can conflict with.
+ */
+ if ((params.options & CLUOPT_CONCURRENT) != 0)
+ {
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+ }
+
/*
* If a single relation is specified, process it and we're done ... unless
* the relation is a partitioned table, in which case we fall through.
@@ -476,11 +491,8 @@ RepackLockLevel(bool concurrent)
* If indexOid is InvalidOid, the table will be rewritten in physical order
* instead of index order.
*
- * Note that, in the concurrent case, the function releases the lock at some
- * point, in order to get AccessExclusiveLock for the final steps (i.e. to
- * swap the relation files). To make things simpler, the caller should expect
- * OldHeap to be closed on return, regardless CLUOPT_CONCURRENT. (The
- * AccessExclusiveLock is kept till the end of the transaction.)
+ * On return, OldHeap is closed but locked with AccessExclusiveLock - the lock
+ * will be released at end of the transaction.
*
* 'cmd' indicates which command is being executed, to be used for error
* messages.
@@ -512,10 +524,12 @@ cluster_rel(RepackCommand cmd, Relation OldHeap, Oid indexOid,
/*
* Make sure we're not in a transaction block.
*
- * The reason is that repack_setup_logical_decoding() could deadlock
- * if there's an XID already assigned. It would be possible to run in
- * a transaction block if we had no XID, but this restriction is
- * simpler for users to understand and we don't lose anything.
+ * The reason is that repack_setup_logical_decoding() could wait
+ * indefinitely for our XID to complete. (The deadlock detector would
+ * not recognize it because we'd be waiting for ourselves, i.e. no
+ * real lock conflict.) It would be possible to run in a transaction
+ * block if we had no XID, but this restriction is simpler for users
+ * to understand and we don't lose anything.
*/
PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
@@ -998,10 +1012,8 @@ rebuild_relation(Relation OldHeap, Relation index, bool verbose,
* Note that the worker has to wait for all transactions with XID
* already assigned to finish. If some of those transactions is
* waiting for a lock conflicting with ShareUpdateExclusiveLock on our
- * table (e.g. it runs CREATE INDEX), we can end up in a deadlock.
- * Not sure this risk is worth unlocking/locking the table (and its
- * clustering index) and checking again if it's still eligible for
- * REPACK CONCURRENTLY.
+ * table (e.g. it runs CREATE INDEX), it should encounter ERROR in the
+ * deadlock checking code.
*/
start_repack_decoding_worker(tableOid);
@@ -3090,7 +3102,19 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
/*
- * Tuples and pages of the old heap will be gone, but the heap will stay.
+ * Now that we have all access-exclusive locks on all relations, we no
+ * longer want other processes to error out when trying to acquire a
+ * conflicting lock. Therefore, unset our flag.
+ */
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+
+ /*
+ * Tuples and pages of the old heap will be gone, but the heap itself will
+ * stay. In order for predicate locks to continue to work, convert them
+ * to relation-level locks. We do this both for table and indexes.
*/
TransferPredicateLocksToHeapRelation(OldHeap);
foreach_ptr(RelationData, index, indexrels)
@@ -3288,9 +3312,11 @@ start_repack_decoding_worker(Oid relid)
/*
* The decoding setup must be done before the caller can have XID assigned
- * for any reason, otherwise the worker might end up in a deadlock,
- * waiting for the caller's transaction to end. Therefore wait here until
- * the worker indicates that it has the logical decoding initialized.
+ * for any reason, otherwise the worker might end up waiting for the
+ * caller's transaction to end. (Deadlock detector does not consider this
+ * a conflict because the worker is in the same locking group as the
+ * backend that launched it.) Therefore wait here until the worker
+ * indicates that it has the logical decoding initialized.
*/
ConditionVariablePrepareToSleep(&shared->cv);
for (;;)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..c20ac682b0d 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -620,6 +620,21 @@ FindLockCycleRecurseMember(PGPROC *checkProc,
proc->statusFlags & PROC_IS_AUTOVACUUM)
blocking_autovacuum_proc = proc;
+ /*
+ * Similarly, if we note that we're blocked by some
+ * process running REPACK (CONCURRENTLY), just fail. That
+ * process is going to upgrade its lock at some point, and
+ * it would be inappropriate for any other process to
+ * cause that to fail.
+ */
+ if (checkProc == MyProc &&
+ proc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+ ereport(ERROR,
+ errcode(ERRCODE_OBJECT_IN_USE),
+ errmsg("could not wait for concurrent REPACK"),
+ errdetail("Process %d waits for REPACK running on process %d",
+ MyProc->pid, proc->pid));
+
/* We're done looking at this proclock */
break;
}
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 1dad125706e..8ad9718f3d6 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -69,10 +69,12 @@ struct XidCache
#define PROC_AFFECTS_ALL_HORIZONS 0x20 /* this proc's xmin must be
* included in vacuum horizons
* in all databases */
+#define PROC_IN_CONCURRENT_REPACK 0x40 /* REPACK (CONCURRENTLY) */
-/* flags reset at EOXact */
+/* flags reset at EOXact. A bit of a misnomer ... */
#define PROC_VACUUM_STATE_MASK \
- (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+ (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+ PROC_IN_CONCURRENT_REPACK)
/*
* Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/Makefile b/src/test/modules/injection_points/Makefile
index 2cd7d87c533..f7663859fe2 100644
--- a/src/test/modules/injection_points/Makefile
+++ b/src/test/modules/injection_points/Makefile
@@ -15,6 +15,7 @@ REGRESS_OPTS = --dlpath=$(top_builddir)/src/test/regress
ISOLATION = basic \
inplace \
repack \
+ repack_deadlock \
repack_toast \
syscache-update-pruned \
heap_lock_update
diff --git a/src/test/modules/injection_points/expected/repack_deadlock.out b/src/test/modules/injection_points/expected/repack_deadlock.out
new file mode 100644
index 00000000000..a86e4767536
--- /dev/null
+++ b/src/test/modules/injection_points/expected/repack_deadlock.out
@@ -0,0 +1,63 @@
+Parsed test spec with 2 sessions
+
+starting permutation: wait_before_lock add_column wakeup_before_lock check1
+injection_points_attach
+-----------------------
+
+(1 row)
+
+step wait_before_lock:
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+ <waiting ...>
+step add_column:
+ alter table repack_deadlock add column noise text;
+ <waiting ...>
+step add_column: <... completed>
+ERROR: could not wait for concurrent REPACK
+step wakeup_before_lock:
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1:
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+
+count
+-----
+ 1
+(1 row)
+
+i|j
+-+-
+1|1
+2|2
+3|3
+4|4
+(4 rows)
+
+count
+-----
+ 4
+(1 row)
+
+injection_points_detach
+-----------------------
+
+(1 row)
+
diff --git a/src/test/modules/injection_points/meson.build b/src/test/modules/injection_points/meson.build
index a414abb924b..1cd88d6db65 100644
--- a/src/test/modules/injection_points/meson.build
+++ b/src/test/modules/injection_points/meson.build
@@ -46,6 +46,7 @@ tests += {
'basic',
'inplace',
'repack',
+ 'repack_deadlock',
'repack_toast',
'syscache-update-pruned',
'heap_lock_update',
diff --git a/src/test/modules/injection_points/specs/repack_deadlock.spec b/src/test/modules/injection_points/specs/repack_deadlock.spec
new file mode 100644
index 00000000000..9d23a6588c2
--- /dev/null
+++ b/src/test/modules/injection_points/specs/repack_deadlock.spec
@@ -0,0 +1,83 @@
+# Test REPACK with a concurrent transaction that would cause a deadlock
+setup
+{
+ CREATE EXTENSION injection_points;
+
+ CREATE TABLE repack_deadlock(i int PRIMARY KEY, j int);
+ INSERT INTO repack_deadlock(i, j) VALUES (1, 1), (2, 2), (3, 3), (4, 4);
+
+ CREATE TABLE relfilenodes(node oid);
+
+ CREATE TABLE data_s1(i int, j int);
+ CREATE TABLE data_s2(i int, j int);
+}
+
+teardown
+{
+ DROP TABLE repack_deadlock;
+ DROP EXTENSION injection_points;
+
+ DROP TABLE relfilenodes;
+ DROP TABLE data_s1;
+ DROP TABLE data_s2;
+}
+
+session s1
+setup
+{
+ SELECT injection_points_set_local();
+ SELECT injection_points_attach('repack-concurrently-before-lock', 'wait');
+}
+# Perform the initial load and wait for s2 to do some data changes.
+step wait_before_lock
+{
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+}
+# Check the table from the perspective of s1.
+#
+# Besides the contents, we also check that relfilenode has changed.
+
+# Have each session write the contents into a table and use FULL JOIN to check
+# if the outputs are identical.
+step check1
+{
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+}
+teardown
+{
+ SELECT injection_points_detach('repack-concurrently-before-lock');
+}
+
+session s2
+# Change the existing data. UPDATE changes both key and non-key columns. Also
+# update one row twice to test whether tuple version generated by this session
+# can be found.
+step add_column
+{
+ alter table repack_deadlock add column noise text;
+}
+
+step wakeup_before_lock
+{
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+}
+
+# Test if data changes introduced while one session is performing REPACK
+# CONCURRENTLY find their way into the table.
+permutation
+ wait_before_lock
+ add_column
+ wakeup_before_lock
+ check1
--
2.47.3
--=-=-=--
^ permalink raw reply [nested|flat] 34+ messages in thread
* [PATCH v49 5/7] Error out any process that would block at REPACK
@ 2026-04-01 15:35 Antonin Houska <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Antonin Houska @ 2026-04-01 15:35 UTC (permalink / raw)
Any process waiting on REPACK to release its lock would actually cause
it to deadlock when it tries to upgrade its lock to AEL, losing all work
done to that point. We avoid this by teaching the deadlock detector to
raise an error when this condition is detected.
---
src/backend/commands/repack.c | 60 ++++++++++----
src/backend/storage/lmgr/deadlock.c | 15 ++++
src/include/storage/proc.h | 6 +-
src/test/modules/injection_points/Makefile | 1 +
.../expected/repack_deadlock.out | 63 ++++++++++++++
src/test/modules/injection_points/meson.build | 1 +
.../specs/repack_deadlock.spec | 83 +++++++++++++++++++
7 files changed, 210 insertions(+), 19 deletions(-)
create mode 100644 src/test/modules/injection_points/expected/repack_deadlock.out
create mode 100644 src/test/modules/injection_points/specs/repack_deadlock.spec
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index a514053b777..274ad900c65 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -276,6 +276,21 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
/* Determine the lock mode to use. */
lockmode = RepackLockLevel((params.options & CLUOPT_CONCURRENT) != 0);
+ /*
+ * If in concurrent mode, set the PROC_IN_CONCURRENT_REPACK flag. This
+ * makes the deadlock checker cause anyone that would conflict with us to
+ * error out. It's important to set this flag ahead of actually locking
+ * the relation; it won't of course affect anyone until we do have a lock
+ * that others can conflict with.
+ */
+ if ((params.options & CLUOPT_CONCURRENT) != 0)
+ {
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+ }
+
/*
* If a single relation is specified, process it and we're done ... unless
* the relation is a partitioned table, in which case we fall through.
@@ -476,11 +491,8 @@ RepackLockLevel(bool concurrent)
* If indexOid is InvalidOid, the table will be rewritten in physical order
* instead of index order.
*
- * Note that, in the concurrent case, the function releases the lock at some
- * point, in order to get AccessExclusiveLock for the final steps (i.e. to
- * swap the relation files). To make things simpler, the caller should expect
- * OldHeap to be closed on return, regardless CLUOPT_CONCURRENT. (The
- * AccessExclusiveLock is kept till the end of the transaction.)
+ * On return, OldHeap is closed but locked with AccessExclusiveLock - the lock
+ * will be released at end of the transaction.
*
* 'cmd' indicates which command is being executed, to be used for error
* messages.
@@ -512,10 +524,12 @@ cluster_rel(RepackCommand cmd, Relation OldHeap, Oid indexOid,
/*
* Make sure we're not in a transaction block.
*
- * The reason is that repack_setup_logical_decoding() could deadlock
- * if there's an XID already assigned. It would be possible to run in
- * a transaction block if we had no XID, but this restriction is
- * simpler for users to understand and we don't lose anything.
+ * The reason is that repack_setup_logical_decoding() could wait
+ * indefinitely for our XID to complete. (The deadlock detector would
+ * not recognize it because we'd be waiting for ourselves, i.e. no
+ * real lock conflict.) It would be possible to run in a transaction
+ * block if we had no XID, but this restriction is simpler for users
+ * to understand and we don't lose anything.
*/
PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
@@ -998,10 +1012,8 @@ rebuild_relation(Relation OldHeap, Relation index, bool verbose,
* Note that the worker has to wait for all transactions with XID
* already assigned to finish. If some of those transactions is
* waiting for a lock conflicting with ShareUpdateExclusiveLock on our
- * table (e.g. it runs CREATE INDEX), we can end up in a deadlock.
- * Not sure this risk is worth unlocking/locking the table (and its
- * clustering index) and checking again if it's still eligible for
- * REPACK CONCURRENTLY.
+ * table (e.g. it runs CREATE INDEX), it should encounter ERROR in the
+ * deadlock checking code.
*/
start_repack_decoding_worker(tableOid);
@@ -3090,7 +3102,19 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
/*
- * Tuples and pages of the old heap will be gone, but the heap will stay.
+ * Now that we have all access-exclusive locks on all relations, we no
+ * longer want other processes to error out when trying to acquire a
+ * conflicting lock. Therefore, unset our flag.
+ */
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+
+ /*
+ * Tuples and pages of the old heap will be gone, but the heap itself will
+ * stay. In order for predicate locks to continue to work, convert them
+ * to relation-level locks. We do this both for table and indexes.
*/
TransferPredicateLocksToHeapRelation(OldHeap);
foreach_ptr(RelationData, index, indexrels)
@@ -3288,9 +3312,11 @@ start_repack_decoding_worker(Oid relid)
/*
* The decoding setup must be done before the caller can have XID assigned
- * for any reason, otherwise the worker might end up in a deadlock,
- * waiting for the caller's transaction to end. Therefore wait here until
- * the worker indicates that it has the logical decoding initialized.
+ * for any reason, otherwise the worker might end up waiting for the
+ * caller's transaction to end. (Deadlock detector does not consider this
+ * a conflict because the worker is in the same locking group as the
+ * backend that launched it.) Therefore wait here until the worker
+ * indicates that it has the logical decoding initialized.
*/
ConditionVariablePrepareToSleep(&shared->cv);
for (;;)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..c20ac682b0d 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -620,6 +620,21 @@ FindLockCycleRecurseMember(PGPROC *checkProc,
proc->statusFlags & PROC_IS_AUTOVACUUM)
blocking_autovacuum_proc = proc;
+ /*
+ * Similarly, if we note that we're blocked by some
+ * process running REPACK (CONCURRENTLY), just fail. That
+ * process is going to upgrade its lock at some point, and
+ * it would be inappropriate for any other process to
+ * cause that to fail.
+ */
+ if (checkProc == MyProc &&
+ proc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+ ereport(ERROR,
+ errcode(ERRCODE_OBJECT_IN_USE),
+ errmsg("could not wait for concurrent REPACK"),
+ errdetail("Process %d waits for REPACK running on process %d",
+ MyProc->pid, proc->pid));
+
/* We're done looking at this proclock */
break;
}
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 1dad125706e..8ad9718f3d6 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -69,10 +69,12 @@ struct XidCache
#define PROC_AFFECTS_ALL_HORIZONS 0x20 /* this proc's xmin must be
* included in vacuum horizons
* in all databases */
+#define PROC_IN_CONCURRENT_REPACK 0x40 /* REPACK (CONCURRENTLY) */
-/* flags reset at EOXact */
+/* flags reset at EOXact. A bit of a misnomer ... */
#define PROC_VACUUM_STATE_MASK \
- (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+ (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+ PROC_IN_CONCURRENT_REPACK)
/*
* Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/Makefile b/src/test/modules/injection_points/Makefile
index 2cd7d87c533..f7663859fe2 100644
--- a/src/test/modules/injection_points/Makefile
+++ b/src/test/modules/injection_points/Makefile
@@ -15,6 +15,7 @@ REGRESS_OPTS = --dlpath=$(top_builddir)/src/test/regress
ISOLATION = basic \
inplace \
repack \
+ repack_deadlock \
repack_toast \
syscache-update-pruned \
heap_lock_update
diff --git a/src/test/modules/injection_points/expected/repack_deadlock.out b/src/test/modules/injection_points/expected/repack_deadlock.out
new file mode 100644
index 00000000000..a86e4767536
--- /dev/null
+++ b/src/test/modules/injection_points/expected/repack_deadlock.out
@@ -0,0 +1,63 @@
+Parsed test spec with 2 sessions
+
+starting permutation: wait_before_lock add_column wakeup_before_lock check1
+injection_points_attach
+-----------------------
+
+(1 row)
+
+step wait_before_lock:
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+ <waiting ...>
+step add_column:
+ alter table repack_deadlock add column noise text;
+ <waiting ...>
+step add_column: <... completed>
+ERROR: could not wait for concurrent REPACK
+step wakeup_before_lock:
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1:
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+
+count
+-----
+ 1
+(1 row)
+
+i|j
+-+-
+1|1
+2|2
+3|3
+4|4
+(4 rows)
+
+count
+-----
+ 4
+(1 row)
+
+injection_points_detach
+-----------------------
+
+(1 row)
+
diff --git a/src/test/modules/injection_points/meson.build b/src/test/modules/injection_points/meson.build
index a414abb924b..1cd88d6db65 100644
--- a/src/test/modules/injection_points/meson.build
+++ b/src/test/modules/injection_points/meson.build
@@ -46,6 +46,7 @@ tests += {
'basic',
'inplace',
'repack',
+ 'repack_deadlock',
'repack_toast',
'syscache-update-pruned',
'heap_lock_update',
diff --git a/src/test/modules/injection_points/specs/repack_deadlock.spec b/src/test/modules/injection_points/specs/repack_deadlock.spec
new file mode 100644
index 00000000000..9d23a6588c2
--- /dev/null
+++ b/src/test/modules/injection_points/specs/repack_deadlock.spec
@@ -0,0 +1,83 @@
+# Test REPACK with a concurrent transaction that would cause a deadlock
+setup
+{
+ CREATE EXTENSION injection_points;
+
+ CREATE TABLE repack_deadlock(i int PRIMARY KEY, j int);
+ INSERT INTO repack_deadlock(i, j) VALUES (1, 1), (2, 2), (3, 3), (4, 4);
+
+ CREATE TABLE relfilenodes(node oid);
+
+ CREATE TABLE data_s1(i int, j int);
+ CREATE TABLE data_s2(i int, j int);
+}
+
+teardown
+{
+ DROP TABLE repack_deadlock;
+ DROP EXTENSION injection_points;
+
+ DROP TABLE relfilenodes;
+ DROP TABLE data_s1;
+ DROP TABLE data_s2;
+}
+
+session s1
+setup
+{
+ SELECT injection_points_set_local();
+ SELECT injection_points_attach('repack-concurrently-before-lock', 'wait');
+}
+# Perform the initial load and wait for s2 to do some data changes.
+step wait_before_lock
+{
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+}
+# Check the table from the perspective of s1.
+#
+# Besides the contents, we also check that relfilenode has changed.
+
+# Have each session write the contents into a table and use FULL JOIN to check
+# if the outputs are identical.
+step check1
+{
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+}
+teardown
+{
+ SELECT injection_points_detach('repack-concurrently-before-lock');
+}
+
+session s2
+# Change the existing data. UPDATE changes both key and non-key columns. Also
+# update one row twice to test whether tuple version generated by this session
+# can be found.
+step add_column
+{
+ alter table repack_deadlock add column noise text;
+}
+
+step wakeup_before_lock
+{
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+}
+
+# Test if data changes introduced while one session is performing REPACK
+# CONCURRENTLY find their way into the table.
+permutation
+ wait_before_lock
+ add_column
+ wakeup_before_lock
+ check1
--
2.47.3
--3n3vqr5y2jhknrqj
Content-Type: text/x-diff; charset=utf-8
Content-Disposition: attachment;
filename="v49-0006-Teach-snapshot-builder-to-skip-transactions-runn.patch"
^ permalink raw reply [nested|flat] 34+ messages in thread
* [PATCH v50 6/8] Error out any process that would block at REPACK
@ 2026-04-01 15:35 Antonin Houska <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Antonin Houska @ 2026-04-01 15:35 UTC (permalink / raw)
Any process waiting on REPACK to release its lock would actually cause
it to deadlock when it tries to upgrade its lock to AEL, losing all work
done to that point. We avoid this by teaching the deadlock detector to
raise an error when this condition is detected.
---
src/backend/commands/repack.c | 60 ++++++++++----
src/backend/storage/lmgr/deadlock.c | 15 ++++
src/include/storage/proc.h | 6 +-
src/test/modules/injection_points/Makefile | 1 +
.../expected/repack_deadlock.out | 63 ++++++++++++++
src/test/modules/injection_points/meson.build | 1 +
.../specs/repack_deadlock.spec | 83 +++++++++++++++++++
7 files changed, 210 insertions(+), 19 deletions(-)
create mode 100644 src/test/modules/injection_points/expected/repack_deadlock.out
create mode 100644 src/test/modules/injection_points/specs/repack_deadlock.spec
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 03829892d57..d6e446d582d 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -279,6 +279,21 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
/* Determine the lock mode to use. */
lockmode = RepackLockLevel((params.options & CLUOPT_CONCURRENT) != 0);
+ /*
+ * If in concurrent mode, set the PROC_IN_CONCURRENT_REPACK flag. This
+ * makes the deadlock checker cause anyone that would conflict with us to
+ * error out. It's important to set this flag ahead of actually locking
+ * the relation; it won't of course affect anyone until we do have a lock
+ * that others can conflict with.
+ */
+ if ((params.options & CLUOPT_CONCURRENT) != 0)
+ {
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+ }
+
/*
* If a single relation is specified, process it and we're done ... unless
* the relation is a partitioned table, in which case we fall through.
@@ -479,11 +494,8 @@ RepackLockLevel(bool concurrent)
* If indexOid is InvalidOid, the table will be rewritten in physical order
* instead of index order.
*
- * Note that, in the concurrent case, the function releases the lock at some
- * point, in order to get AccessExclusiveLock for the final steps (i.e. to
- * swap the relation files). To make things simpler, the caller should expect
- * OldHeap to be closed on return, regardless CLUOPT_CONCURRENT. (The
- * AccessExclusiveLock is kept till the end of the transaction.)
+ * On return, OldHeap is closed but locked with AccessExclusiveLock - the lock
+ * will be released at end of the transaction.
*
* 'cmd' indicates which command is being executed, to be used for error
* messages.
@@ -515,10 +527,12 @@ cluster_rel(RepackCommand cmd, Relation OldHeap, Oid indexOid,
/*
* Make sure we're not in a transaction block.
*
- * The reason is that repack_setup_logical_decoding() could deadlock
- * if there's an XID already assigned. It would be possible to run in
- * a transaction block if we had no XID, but this restriction is
- * simpler for users to understand and we don't lose anything.
+ * The reason is that repack_setup_logical_decoding() could wait
+ * indefinitely for our XID to complete. (The deadlock detector would
+ * not recognize it because we'd be waiting for ourselves, i.e. no
+ * real lock conflict.) It would be possible to run in a transaction
+ * block if we had no XID, but this restriction is simpler for users
+ * to understand and we don't lose anything.
*/
PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
@@ -1001,10 +1015,8 @@ rebuild_relation(Relation OldHeap, Relation index, bool verbose,
* Note that the worker has to wait for all transactions with XID
* already assigned to finish. If some of those transactions is
* waiting for a lock conflicting with ShareUpdateExclusiveLock on our
- * table (e.g. it runs CREATE INDEX), we can end up in a deadlock.
- * Not sure this risk is worth unlocking/locking the table (and its
- * clustering index) and checking again if it's still eligible for
- * REPACK CONCURRENTLY.
+ * table (e.g. it runs CREATE INDEX), it should encounter ERROR in the
+ * deadlock checking code.
*/
start_repack_decoding_worker(tableOid);
@@ -3093,7 +3105,19 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
/*
- * Tuples and pages of the old heap will be gone, but the heap will stay.
+ * Now that we have all access-exclusive locks on all relations, we no
+ * longer want other processes to error out when trying to acquire a
+ * conflicting lock. Therefore, unset our flag.
+ */
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+
+ /*
+ * Tuples and pages of the old heap will be gone, but the heap itself will
+ * stay. In order for predicate locks to continue to work, convert them
+ * to relation-level locks. We do this both for table and indexes.
*/
TransferPredicateLocksToHeapRelation(OldHeap);
foreach_ptr(RelationData, index, indexrels)
@@ -3366,9 +3390,11 @@ start_repack_decoding_worker(Oid relid)
/*
* The decoding setup must be done before the caller can have XID assigned
- * for any reason, otherwise the worker might end up in a deadlock,
- * waiting for the caller's transaction to end. Therefore wait here until
- * the worker indicates that it has the logical decoding initialized.
+ * for any reason, otherwise the worker might end up waiting for the
+ * caller's transaction to end. (Deadlock detector does not consider this
+ * a conflict because the worker is in the same locking group as the
+ * backend that launched it.) Therefore wait here until the worker
+ * indicates that it has the logical decoding initialized.
*/
ConditionVariablePrepareToSleep(&shared->cv);
for (;;)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..c20ac682b0d 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -620,6 +620,21 @@ FindLockCycleRecurseMember(PGPROC *checkProc,
proc->statusFlags & PROC_IS_AUTOVACUUM)
blocking_autovacuum_proc = proc;
+ /*
+ * Similarly, if we note that we're blocked by some
+ * process running REPACK (CONCURRENTLY), just fail. That
+ * process is going to upgrade its lock at some point, and
+ * it would be inappropriate for any other process to
+ * cause that to fail.
+ */
+ if (checkProc == MyProc &&
+ proc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+ ereport(ERROR,
+ errcode(ERRCODE_OBJECT_IN_USE),
+ errmsg("could not wait for concurrent REPACK"),
+ errdetail("Process %d waits for REPACK running on process %d",
+ MyProc->pid, proc->pid));
+
/* We're done looking at this proclock */
break;
}
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 1dad125706e..8ad9718f3d6 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -69,10 +69,12 @@ struct XidCache
#define PROC_AFFECTS_ALL_HORIZONS 0x20 /* this proc's xmin must be
* included in vacuum horizons
* in all databases */
+#define PROC_IN_CONCURRENT_REPACK 0x40 /* REPACK (CONCURRENTLY) */
-/* flags reset at EOXact */
+/* flags reset at EOXact. A bit of a misnomer ... */
#define PROC_VACUUM_STATE_MASK \
- (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+ (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+ PROC_IN_CONCURRENT_REPACK)
/*
* Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/Makefile b/src/test/modules/injection_points/Makefile
index 2cd7d87c533..f7663859fe2 100644
--- a/src/test/modules/injection_points/Makefile
+++ b/src/test/modules/injection_points/Makefile
@@ -15,6 +15,7 @@ REGRESS_OPTS = --dlpath=$(top_builddir)/src/test/regress
ISOLATION = basic \
inplace \
repack \
+ repack_deadlock \
repack_toast \
syscache-update-pruned \
heap_lock_update
diff --git a/src/test/modules/injection_points/expected/repack_deadlock.out b/src/test/modules/injection_points/expected/repack_deadlock.out
new file mode 100644
index 00000000000..a86e4767536
--- /dev/null
+++ b/src/test/modules/injection_points/expected/repack_deadlock.out
@@ -0,0 +1,63 @@
+Parsed test spec with 2 sessions
+
+starting permutation: wait_before_lock add_column wakeup_before_lock check1
+injection_points_attach
+-----------------------
+
+(1 row)
+
+step wait_before_lock:
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+ <waiting ...>
+step add_column:
+ alter table repack_deadlock add column noise text;
+ <waiting ...>
+step add_column: <... completed>
+ERROR: could not wait for concurrent REPACK
+step wakeup_before_lock:
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1:
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+
+count
+-----
+ 1
+(1 row)
+
+i|j
+-+-
+1|1
+2|2
+3|3
+4|4
+(4 rows)
+
+count
+-----
+ 4
+(1 row)
+
+injection_points_detach
+-----------------------
+
+(1 row)
+
diff --git a/src/test/modules/injection_points/meson.build b/src/test/modules/injection_points/meson.build
index a414abb924b..1cd88d6db65 100644
--- a/src/test/modules/injection_points/meson.build
+++ b/src/test/modules/injection_points/meson.build
@@ -46,6 +46,7 @@ tests += {
'basic',
'inplace',
'repack',
+ 'repack_deadlock',
'repack_toast',
'syscache-update-pruned',
'heap_lock_update',
diff --git a/src/test/modules/injection_points/specs/repack_deadlock.spec b/src/test/modules/injection_points/specs/repack_deadlock.spec
new file mode 100644
index 00000000000..9d23a6588c2
--- /dev/null
+++ b/src/test/modules/injection_points/specs/repack_deadlock.spec
@@ -0,0 +1,83 @@
+# Test REPACK with a concurrent transaction that would cause a deadlock
+setup
+{
+ CREATE EXTENSION injection_points;
+
+ CREATE TABLE repack_deadlock(i int PRIMARY KEY, j int);
+ INSERT INTO repack_deadlock(i, j) VALUES (1, 1), (2, 2), (3, 3), (4, 4);
+
+ CREATE TABLE relfilenodes(node oid);
+
+ CREATE TABLE data_s1(i int, j int);
+ CREATE TABLE data_s2(i int, j int);
+}
+
+teardown
+{
+ DROP TABLE repack_deadlock;
+ DROP EXTENSION injection_points;
+
+ DROP TABLE relfilenodes;
+ DROP TABLE data_s1;
+ DROP TABLE data_s2;
+}
+
+session s1
+setup
+{
+ SELECT injection_points_set_local();
+ SELECT injection_points_attach('repack-concurrently-before-lock', 'wait');
+}
+# Perform the initial load and wait for s2 to do some data changes.
+step wait_before_lock
+{
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+}
+# Check the table from the perspective of s1.
+#
+# Besides the contents, we also check that relfilenode has changed.
+
+# Have each session write the contents into a table and use FULL JOIN to check
+# if the outputs are identical.
+step check1
+{
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+}
+teardown
+{
+ SELECT injection_points_detach('repack-concurrently-before-lock');
+}
+
+session s2
+# Change the existing data. UPDATE changes both key and non-key columns. Also
+# update one row twice to test whether tuple version generated by this session
+# can be found.
+step add_column
+{
+ alter table repack_deadlock add column noise text;
+}
+
+step wakeup_before_lock
+{
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+}
+
+# Test if data changes introduced while one session is performing REPACK
+# CONCURRENTLY find their way into the table.
+permutation
+ wait_before_lock
+ add_column
+ wakeup_before_lock
+ check1
--
2.47.3
--brnevsqjnuzpyok4
Content-Type: text/x-diff; charset=utf-8
Content-Disposition: attachment;
filename="v50-0007-Introduce-an-option-to-make-logical-replication-.patch"
^ permalink raw reply [nested|flat] 34+ messages in thread
* [PATCH v52 06/10] Error out any process that would block at REPACK
@ 2026-04-01 15:35 Antonin Houska <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Antonin Houska @ 2026-04-01 15:35 UTC (permalink / raw)
Any process waiting on REPACK to release its lock would actually cause
it to deadlock when it tries to upgrade its lock to AEL, losing all work
done to that point. We avoid this by teaching the deadlock detector to
raise an error when this condition is detected.
---
src/backend/commands/repack.c | 60 ++++++++++----
src/backend/storage/lmgr/deadlock.c | 15 ++++
src/include/storage/proc.h | 6 +-
src/test/modules/injection_points/Makefile | 1 +
.../expected/repack_deadlock.out | 63 ++++++++++++++
src/test/modules/injection_points/meson.build | 1 +
.../specs/repack_deadlock.spec | 83 +++++++++++++++++++
7 files changed, 210 insertions(+), 19 deletions(-)
create mode 100644 src/test/modules/injection_points/expected/repack_deadlock.out
create mode 100644 src/test/modules/injection_points/specs/repack_deadlock.spec
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 03829892d57..d6e446d582d 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -279,6 +279,21 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
/* Determine the lock mode to use. */
lockmode = RepackLockLevel((params.options & CLUOPT_CONCURRENT) != 0);
+ /*
+ * If in concurrent mode, set the PROC_IN_CONCURRENT_REPACK flag. This
+ * makes the deadlock checker cause anyone that would conflict with us to
+ * error out. It's important to set this flag ahead of actually locking
+ * the relation; it won't of course affect anyone until we do have a lock
+ * that others can conflict with.
+ */
+ if ((params.options & CLUOPT_CONCURRENT) != 0)
+ {
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+ }
+
/*
* If a single relation is specified, process it and we're done ... unless
* the relation is a partitioned table, in which case we fall through.
@@ -479,11 +494,8 @@ RepackLockLevel(bool concurrent)
* If indexOid is InvalidOid, the table will be rewritten in physical order
* instead of index order.
*
- * Note that, in the concurrent case, the function releases the lock at some
- * point, in order to get AccessExclusiveLock for the final steps (i.e. to
- * swap the relation files). To make things simpler, the caller should expect
- * OldHeap to be closed on return, regardless CLUOPT_CONCURRENT. (The
- * AccessExclusiveLock is kept till the end of the transaction.)
+ * On return, OldHeap is closed but locked with AccessExclusiveLock - the lock
+ * will be released at end of the transaction.
*
* 'cmd' indicates which command is being executed, to be used for error
* messages.
@@ -515,10 +527,12 @@ cluster_rel(RepackCommand cmd, Relation OldHeap, Oid indexOid,
/*
* Make sure we're not in a transaction block.
*
- * The reason is that repack_setup_logical_decoding() could deadlock
- * if there's an XID already assigned. It would be possible to run in
- * a transaction block if we had no XID, but this restriction is
- * simpler for users to understand and we don't lose anything.
+ * The reason is that repack_setup_logical_decoding() could wait
+ * indefinitely for our XID to complete. (The deadlock detector would
+ * not recognize it because we'd be waiting for ourselves, i.e. no
+ * real lock conflict.) It would be possible to run in a transaction
+ * block if we had no XID, but this restriction is simpler for users
+ * to understand and we don't lose anything.
*/
PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
@@ -1001,10 +1015,8 @@ rebuild_relation(Relation OldHeap, Relation index, bool verbose,
* Note that the worker has to wait for all transactions with XID
* already assigned to finish. If some of those transactions is
* waiting for a lock conflicting with ShareUpdateExclusiveLock on our
- * table (e.g. it runs CREATE INDEX), we can end up in a deadlock.
- * Not sure this risk is worth unlocking/locking the table (and its
- * clustering index) and checking again if it's still eligible for
- * REPACK CONCURRENTLY.
+ * table (e.g. it runs CREATE INDEX), it should encounter ERROR in the
+ * deadlock checking code.
*/
start_repack_decoding_worker(tableOid);
@@ -3093,7 +3105,19 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
/*
- * Tuples and pages of the old heap will be gone, but the heap will stay.
+ * Now that we have all access-exclusive locks on all relations, we no
+ * longer want other processes to error out when trying to acquire a
+ * conflicting lock. Therefore, unset our flag.
+ */
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+
+ /*
+ * Tuples and pages of the old heap will be gone, but the heap itself will
+ * stay. In order for predicate locks to continue to work, convert them
+ * to relation-level locks. We do this both for table and indexes.
*/
TransferPredicateLocksToHeapRelation(OldHeap);
foreach_ptr(RelationData, index, indexrels)
@@ -3366,9 +3390,11 @@ start_repack_decoding_worker(Oid relid)
/*
* The decoding setup must be done before the caller can have XID assigned
- * for any reason, otherwise the worker might end up in a deadlock,
- * waiting for the caller's transaction to end. Therefore wait here until
- * the worker indicates that it has the logical decoding initialized.
+ * for any reason, otherwise the worker might end up waiting for the
+ * caller's transaction to end. (Deadlock detector does not consider this
+ * a conflict because the worker is in the same locking group as the
+ * backend that launched it.) Therefore wait here until the worker
+ * indicates that it has the logical decoding initialized.
*/
ConditionVariablePrepareToSleep(&shared->cv);
for (;;)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..c20ac682b0d 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -620,6 +620,21 @@ FindLockCycleRecurseMember(PGPROC *checkProc,
proc->statusFlags & PROC_IS_AUTOVACUUM)
blocking_autovacuum_proc = proc;
+ /*
+ * Similarly, if we note that we're blocked by some
+ * process running REPACK (CONCURRENTLY), just fail. That
+ * process is going to upgrade its lock at some point, and
+ * it would be inappropriate for any other process to
+ * cause that to fail.
+ */
+ if (checkProc == MyProc &&
+ proc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+ ereport(ERROR,
+ errcode(ERRCODE_OBJECT_IN_USE),
+ errmsg("could not wait for concurrent REPACK"),
+ errdetail("Process %d waits for REPACK running on process %d",
+ MyProc->pid, proc->pid));
+
/* We're done looking at this proclock */
break;
}
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 1dad125706e..8ad9718f3d6 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -69,10 +69,12 @@ struct XidCache
#define PROC_AFFECTS_ALL_HORIZONS 0x20 /* this proc's xmin must be
* included in vacuum horizons
* in all databases */
+#define PROC_IN_CONCURRENT_REPACK 0x40 /* REPACK (CONCURRENTLY) */
-/* flags reset at EOXact */
+/* flags reset at EOXact. A bit of a misnomer ... */
#define PROC_VACUUM_STATE_MASK \
- (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+ (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+ PROC_IN_CONCURRENT_REPACK)
/*
* Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/Makefile b/src/test/modules/injection_points/Makefile
index 2cd7d87c533..f7663859fe2 100644
--- a/src/test/modules/injection_points/Makefile
+++ b/src/test/modules/injection_points/Makefile
@@ -15,6 +15,7 @@ REGRESS_OPTS = --dlpath=$(top_builddir)/src/test/regress
ISOLATION = basic \
inplace \
repack \
+ repack_deadlock \
repack_toast \
syscache-update-pruned \
heap_lock_update
diff --git a/src/test/modules/injection_points/expected/repack_deadlock.out b/src/test/modules/injection_points/expected/repack_deadlock.out
new file mode 100644
index 00000000000..a86e4767536
--- /dev/null
+++ b/src/test/modules/injection_points/expected/repack_deadlock.out
@@ -0,0 +1,63 @@
+Parsed test spec with 2 sessions
+
+starting permutation: wait_before_lock add_column wakeup_before_lock check1
+injection_points_attach
+-----------------------
+
+(1 row)
+
+step wait_before_lock:
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+ <waiting ...>
+step add_column:
+ alter table repack_deadlock add column noise text;
+ <waiting ...>
+step add_column: <... completed>
+ERROR: could not wait for concurrent REPACK
+step wakeup_before_lock:
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1:
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+
+count
+-----
+ 1
+(1 row)
+
+i|j
+-+-
+1|1
+2|2
+3|3
+4|4
+(4 rows)
+
+count
+-----
+ 4
+(1 row)
+
+injection_points_detach
+-----------------------
+
+(1 row)
+
diff --git a/src/test/modules/injection_points/meson.build b/src/test/modules/injection_points/meson.build
index a414abb924b..1cd88d6db65 100644
--- a/src/test/modules/injection_points/meson.build
+++ b/src/test/modules/injection_points/meson.build
@@ -46,6 +46,7 @@ tests += {
'basic',
'inplace',
'repack',
+ 'repack_deadlock',
'repack_toast',
'syscache-update-pruned',
'heap_lock_update',
diff --git a/src/test/modules/injection_points/specs/repack_deadlock.spec b/src/test/modules/injection_points/specs/repack_deadlock.spec
new file mode 100644
index 00000000000..9d23a6588c2
--- /dev/null
+++ b/src/test/modules/injection_points/specs/repack_deadlock.spec
@@ -0,0 +1,83 @@
+# Test REPACK with a concurrent transaction that would cause a deadlock
+setup
+{
+ CREATE EXTENSION injection_points;
+
+ CREATE TABLE repack_deadlock(i int PRIMARY KEY, j int);
+ INSERT INTO repack_deadlock(i, j) VALUES (1, 1), (2, 2), (3, 3), (4, 4);
+
+ CREATE TABLE relfilenodes(node oid);
+
+ CREATE TABLE data_s1(i int, j int);
+ CREATE TABLE data_s2(i int, j int);
+}
+
+teardown
+{
+ DROP TABLE repack_deadlock;
+ DROP EXTENSION injection_points;
+
+ DROP TABLE relfilenodes;
+ DROP TABLE data_s1;
+ DROP TABLE data_s2;
+}
+
+session s1
+setup
+{
+ SELECT injection_points_set_local();
+ SELECT injection_points_attach('repack-concurrently-before-lock', 'wait');
+}
+# Perform the initial load and wait for s2 to do some data changes.
+step wait_before_lock
+{
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+}
+# Check the table from the perspective of s1.
+#
+# Besides the contents, we also check that relfilenode has changed.
+
+# Have each session write the contents into a table and use FULL JOIN to check
+# if the outputs are identical.
+step check1
+{
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+}
+teardown
+{
+ SELECT injection_points_detach('repack-concurrently-before-lock');
+}
+
+session s2
+# Change the existing data. UPDATE changes both key and non-key columns. Also
+# update one row twice to test whether tuple version generated by this session
+# can be found.
+step add_column
+{
+ alter table repack_deadlock add column noise text;
+}
+
+step wakeup_before_lock
+{
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+}
+
+# Test if data changes introduced while one session is performing REPACK
+# CONCURRENTLY find their way into the table.
+permutation
+ wait_before_lock
+ add_column
+ wakeup_before_lock
+ check1
--
2.47.3
--gp2pyozrd5pweboh
Content-Type: text/x-diff; charset=utf-8
Content-Disposition: attachment;
filename="v52-0007-Check-for-transaction-block-early-in-ExecRepack.patch"
^ permalink raw reply [nested|flat] 34+ messages in thread
* [PATCH v53 6/7] Error out any process that would block at REPACK
@ 2026-04-01 15:35 Antonin Houska <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Antonin Houska @ 2026-04-01 15:35 UTC (permalink / raw)
Any process waiting on REPACK to release its lock would actually cause
it to deadlock when it tries to upgrade its lock to AEL, losing all work
done to that point. We avoid this by teaching the deadlock detector to
raise an error when this condition is detected.
---
src/backend/commands/repack.c | 47 ++++++++---
src/backend/storage/lmgr/deadlock.c | 15 ++++
src/include/storage/proc.h | 6 +-
src/test/modules/injection_points/Makefile | 1 +
.../expected/repack_deadlock.out | 63 ++++++++++++++
src/test/modules/injection_points/meson.build | 1 +
.../specs/repack_deadlock.spec | 83 +++++++++++++++++++
7 files changed, 201 insertions(+), 15 deletions(-)
create mode 100644 src/test/modules/injection_points/expected/repack_deadlock.out
create mode 100644 src/test/modules/injection_points/specs/repack_deadlock.spec
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 8d9b2b2e370..92759f33969 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -292,6 +292,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
* to understand and we don't lose any functionality.
*/
PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+ /*
+ * Also set the PROC_IN_CONCURRENT_REPACK flag. This makes the
+ * deadlock checker cause anyone that would conflict with us to error
+ * out. It's important to set this flag ahead of actually locking the
+ * relation; it won't of course affect anyone until we do have a lock
+ * that others can conflict with.
+ */
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
}
/*
@@ -494,11 +506,8 @@ RepackLockLevel(bool concurrent)
* If indexOid is InvalidOid, the table will be rewritten in physical order
* instead of index order.
*
- * Note that, in the concurrent case, the function releases the lock at some
- * point, in order to get AccessExclusiveLock for the final steps (i.e. to
- * swap the relation files). To make things simpler, the caller should expect
- * OldHeap to be closed on return, regardless CLUOPT_CONCURRENT. (The
- * AccessExclusiveLock is kept till the end of the transaction.)
+ * On return, OldHeap is closed but locked with AccessExclusiveLock - the lock
+ * will be released at end of the transaction.
*
* 'cmd' indicates which command is being executed, to be used for error
* messages.
@@ -1004,10 +1013,8 @@ rebuild_relation(Relation OldHeap, Relation index, bool verbose,
* Note that the worker has to wait for all transactions with XID
* already assigned to finish. If some of those transactions is
* waiting for a lock conflicting with ShareUpdateExclusiveLock on our
- * table (e.g. it runs CREATE INDEX), we can end up in a deadlock.
- * Not sure this risk is worth unlocking/locking the table (and its
- * clustering index) and checking again if it's still eligible for
- * REPACK CONCURRENTLY.
+ * table (e.g. it runs CREATE INDEX), it should encounter ERROR in the
+ * deadlock checking code.
*/
start_repack_decoding_worker(tableOid);
@@ -3096,7 +3103,19 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
/*
- * Tuples and pages of the old heap will be gone, but the heap will stay.
+ * Now that we have all access-exclusive locks on all relations, we no
+ * longer want other processes to error out when trying to acquire a
+ * conflicting lock. Therefore, unset our flag.
+ */
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+
+ /*
+ * Tuples and pages of the old heap will be gone, but the heap itself will
+ * stay. In order for predicate locks to continue to work, convert them
+ * to relation-level locks. We do this both for table and indexes.
*/
TransferPredicateLocksToHeapRelation(OldHeap);
foreach_ptr(RelationData, index, indexrels)
@@ -3369,9 +3388,11 @@ start_repack_decoding_worker(Oid relid)
/*
* The decoding setup must be done before the caller can have XID assigned
- * for any reason, otherwise the worker might end up in a deadlock,
- * waiting for the caller's transaction to end. Therefore wait here until
- * the worker indicates that it has the logical decoding initialized.
+ * for any reason, otherwise the worker might end up waiting for the
+ * caller's transaction to end. (Deadlock detector does not consider this
+ * a conflict because the worker is in the same locking group as the
+ * backend that launched it.) Therefore wait here until the worker
+ * indicates that it has the logical decoding initialized.
*/
ConditionVariablePrepareToSleep(&shared->cv);
for (;;)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..c20ac682b0d 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -620,6 +620,21 @@ FindLockCycleRecurseMember(PGPROC *checkProc,
proc->statusFlags & PROC_IS_AUTOVACUUM)
blocking_autovacuum_proc = proc;
+ /*
+ * Similarly, if we note that we're blocked by some
+ * process running REPACK (CONCURRENTLY), just fail. That
+ * process is going to upgrade its lock at some point, and
+ * it would be inappropriate for any other process to
+ * cause that to fail.
+ */
+ if (checkProc == MyProc &&
+ proc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+ ereport(ERROR,
+ errcode(ERRCODE_OBJECT_IN_USE),
+ errmsg("could not wait for concurrent REPACK"),
+ errdetail("Process %d waits for REPACK running on process %d",
+ MyProc->pid, proc->pid));
+
/* We're done looking at this proclock */
break;
}
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 22822fc68d7..a43cfe53558 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
#define PROC_AFFECTS_ALL_HORIZONS 0x20 /* this proc's xmin must be
* included in vacuum horizons
* in all databases */
+#define PROC_IN_CONCURRENT_REPACK 0x40 /* REPACK (CONCURRENTLY) */
-/* flags reset at EOXact */
+/* flags reset at EOXact. A bit of a misnomer ... */
#define PROC_VACUUM_STATE_MASK \
- (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+ (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+ PROC_IN_CONCURRENT_REPACK)
/*
* Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/Makefile b/src/test/modules/injection_points/Makefile
index 2cd7d87c533..f7663859fe2 100644
--- a/src/test/modules/injection_points/Makefile
+++ b/src/test/modules/injection_points/Makefile
@@ -15,6 +15,7 @@ REGRESS_OPTS = --dlpath=$(top_builddir)/src/test/regress
ISOLATION = basic \
inplace \
repack \
+ repack_deadlock \
repack_toast \
syscache-update-pruned \
heap_lock_update
diff --git a/src/test/modules/injection_points/expected/repack_deadlock.out b/src/test/modules/injection_points/expected/repack_deadlock.out
new file mode 100644
index 00000000000..a86e4767536
--- /dev/null
+++ b/src/test/modules/injection_points/expected/repack_deadlock.out
@@ -0,0 +1,63 @@
+Parsed test spec with 2 sessions
+
+starting permutation: wait_before_lock add_column wakeup_before_lock check1
+injection_points_attach
+-----------------------
+
+(1 row)
+
+step wait_before_lock:
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+ <waiting ...>
+step add_column:
+ alter table repack_deadlock add column noise text;
+ <waiting ...>
+step add_column: <... completed>
+ERROR: could not wait for concurrent REPACK
+step wakeup_before_lock:
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1:
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+
+count
+-----
+ 1
+(1 row)
+
+i|j
+-+-
+1|1
+2|2
+3|3
+4|4
+(4 rows)
+
+count
+-----
+ 4
+(1 row)
+
+injection_points_detach
+-----------------------
+
+(1 row)
+
diff --git a/src/test/modules/injection_points/meson.build b/src/test/modules/injection_points/meson.build
index a414abb924b..1cd88d6db65 100644
--- a/src/test/modules/injection_points/meson.build
+++ b/src/test/modules/injection_points/meson.build
@@ -46,6 +46,7 @@ tests += {
'basic',
'inplace',
'repack',
+ 'repack_deadlock',
'repack_toast',
'syscache-update-pruned',
'heap_lock_update',
diff --git a/src/test/modules/injection_points/specs/repack_deadlock.spec b/src/test/modules/injection_points/specs/repack_deadlock.spec
new file mode 100644
index 00000000000..9d23a6588c2
--- /dev/null
+++ b/src/test/modules/injection_points/specs/repack_deadlock.spec
@@ -0,0 +1,83 @@
+# Test REPACK with a concurrent transaction that would cause a deadlock
+setup
+{
+ CREATE EXTENSION injection_points;
+
+ CREATE TABLE repack_deadlock(i int PRIMARY KEY, j int);
+ INSERT INTO repack_deadlock(i, j) VALUES (1, 1), (2, 2), (3, 3), (4, 4);
+
+ CREATE TABLE relfilenodes(node oid);
+
+ CREATE TABLE data_s1(i int, j int);
+ CREATE TABLE data_s2(i int, j int);
+}
+
+teardown
+{
+ DROP TABLE repack_deadlock;
+ DROP EXTENSION injection_points;
+
+ DROP TABLE relfilenodes;
+ DROP TABLE data_s1;
+ DROP TABLE data_s2;
+}
+
+session s1
+setup
+{
+ SELECT injection_points_set_local();
+ SELECT injection_points_attach('repack-concurrently-before-lock', 'wait');
+}
+# Perform the initial load and wait for s2 to do some data changes.
+step wait_before_lock
+{
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+}
+# Check the table from the perspective of s1.
+#
+# Besides the contents, we also check that relfilenode has changed.
+
+# Have each session write the contents into a table and use FULL JOIN to check
+# if the outputs are identical.
+step check1
+{
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+}
+teardown
+{
+ SELECT injection_points_detach('repack-concurrently-before-lock');
+}
+
+session s2
+# Change the existing data. UPDATE changes both key and non-key columns. Also
+# update one row twice to test whether tuple version generated by this session
+# can be found.
+step add_column
+{
+ alter table repack_deadlock add column noise text;
+}
+
+step wakeup_before_lock
+{
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+}
+
+# Test if data changes introduced while one session is performing REPACK
+# CONCURRENTLY find their way into the table.
+permutation
+ wait_before_lock
+ add_column
+ wakeup_before_lock
+ check1
--
2.47.3
--qr3jlalmmcpkiodg
Content-Type: text/x-diff; charset=utf-8
Content-Disposition: attachment;
filename="v53-0007-Reserve-replication-slots-specifically-for-REPAC.patch"
^ permalink raw reply [nested|flat] 34+ messages in thread
* [PATCH v51 06/10] Error out any process that would block at REPACK
@ 2026-04-01 15:35 Antonin Houska <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Antonin Houska @ 2026-04-01 15:35 UTC (permalink / raw)
Any process waiting on REPACK to release its lock would actually cause
it to deadlock when it tries to upgrade its lock to AEL, losing all work
done to that point. We avoid this by teaching the deadlock detector to
raise an error when this condition is detected.
---
src/backend/commands/repack.c | 60 ++++++++++----
src/backend/storage/lmgr/deadlock.c | 15 ++++
src/include/storage/proc.h | 6 +-
src/test/modules/injection_points/Makefile | 1 +
.../expected/repack_deadlock.out | 63 ++++++++++++++
src/test/modules/injection_points/meson.build | 1 +
.../specs/repack_deadlock.spec | 83 +++++++++++++++++++
7 files changed, 210 insertions(+), 19 deletions(-)
create mode 100644 src/test/modules/injection_points/expected/repack_deadlock.out
create mode 100644 src/test/modules/injection_points/specs/repack_deadlock.spec
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 03829892d57..d6e446d582d 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -279,6 +279,21 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
/* Determine the lock mode to use. */
lockmode = RepackLockLevel((params.options & CLUOPT_CONCURRENT) != 0);
+ /*
+ * If in concurrent mode, set the PROC_IN_CONCURRENT_REPACK flag. This
+ * makes the deadlock checker cause anyone that would conflict with us to
+ * error out. It's important to set this flag ahead of actually locking
+ * the relation; it won't of course affect anyone until we do have a lock
+ * that others can conflict with.
+ */
+ if ((params.options & CLUOPT_CONCURRENT) != 0)
+ {
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+ }
+
/*
* If a single relation is specified, process it and we're done ... unless
* the relation is a partitioned table, in which case we fall through.
@@ -479,11 +494,8 @@ RepackLockLevel(bool concurrent)
* If indexOid is InvalidOid, the table will be rewritten in physical order
* instead of index order.
*
- * Note that, in the concurrent case, the function releases the lock at some
- * point, in order to get AccessExclusiveLock for the final steps (i.e. to
- * swap the relation files). To make things simpler, the caller should expect
- * OldHeap to be closed on return, regardless CLUOPT_CONCURRENT. (The
- * AccessExclusiveLock is kept till the end of the transaction.)
+ * On return, OldHeap is closed but locked with AccessExclusiveLock - the lock
+ * will be released at end of the transaction.
*
* 'cmd' indicates which command is being executed, to be used for error
* messages.
@@ -515,10 +527,12 @@ cluster_rel(RepackCommand cmd, Relation OldHeap, Oid indexOid,
/*
* Make sure we're not in a transaction block.
*
- * The reason is that repack_setup_logical_decoding() could deadlock
- * if there's an XID already assigned. It would be possible to run in
- * a transaction block if we had no XID, but this restriction is
- * simpler for users to understand and we don't lose anything.
+ * The reason is that repack_setup_logical_decoding() could wait
+ * indefinitely for our XID to complete. (The deadlock detector would
+ * not recognize it because we'd be waiting for ourselves, i.e. no
+ * real lock conflict.) It would be possible to run in a transaction
+ * block if we had no XID, but this restriction is simpler for users
+ * to understand and we don't lose anything.
*/
PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
@@ -1001,10 +1015,8 @@ rebuild_relation(Relation OldHeap, Relation index, bool verbose,
* Note that the worker has to wait for all transactions with XID
* already assigned to finish. If some of those transactions is
* waiting for a lock conflicting with ShareUpdateExclusiveLock on our
- * table (e.g. it runs CREATE INDEX), we can end up in a deadlock.
- * Not sure this risk is worth unlocking/locking the table (and its
- * clustering index) and checking again if it's still eligible for
- * REPACK CONCURRENTLY.
+ * table (e.g. it runs CREATE INDEX), it should encounter ERROR in the
+ * deadlock checking code.
*/
start_repack_decoding_worker(tableOid);
@@ -3093,7 +3105,19 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
/*
- * Tuples and pages of the old heap will be gone, but the heap will stay.
+ * Now that we have all access-exclusive locks on all relations, we no
+ * longer want other processes to error out when trying to acquire a
+ * conflicting lock. Therefore, unset our flag.
+ */
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+
+ /*
+ * Tuples and pages of the old heap will be gone, but the heap itself will
+ * stay. In order for predicate locks to continue to work, convert them
+ * to relation-level locks. We do this both for table and indexes.
*/
TransferPredicateLocksToHeapRelation(OldHeap);
foreach_ptr(RelationData, index, indexrels)
@@ -3366,9 +3390,11 @@ start_repack_decoding_worker(Oid relid)
/*
* The decoding setup must be done before the caller can have XID assigned
- * for any reason, otherwise the worker might end up in a deadlock,
- * waiting for the caller's transaction to end. Therefore wait here until
- * the worker indicates that it has the logical decoding initialized.
+ * for any reason, otherwise the worker might end up waiting for the
+ * caller's transaction to end. (Deadlock detector does not consider this
+ * a conflict because the worker is in the same locking group as the
+ * backend that launched it.) Therefore wait here until the worker
+ * indicates that it has the logical decoding initialized.
*/
ConditionVariablePrepareToSleep(&shared->cv);
for (;;)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..c20ac682b0d 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -620,6 +620,21 @@ FindLockCycleRecurseMember(PGPROC *checkProc,
proc->statusFlags & PROC_IS_AUTOVACUUM)
blocking_autovacuum_proc = proc;
+ /*
+ * Similarly, if we note that we're blocked by some
+ * process running REPACK (CONCURRENTLY), just fail. That
+ * process is going to upgrade its lock at some point, and
+ * it would be inappropriate for any other process to
+ * cause that to fail.
+ */
+ if (checkProc == MyProc &&
+ proc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+ ereport(ERROR,
+ errcode(ERRCODE_OBJECT_IN_USE),
+ errmsg("could not wait for concurrent REPACK"),
+ errdetail("Process %d waits for REPACK running on process %d",
+ MyProc->pid, proc->pid));
+
/* We're done looking at this proclock */
break;
}
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 1dad125706e..8ad9718f3d6 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -69,10 +69,12 @@ struct XidCache
#define PROC_AFFECTS_ALL_HORIZONS 0x20 /* this proc's xmin must be
* included in vacuum horizons
* in all databases */
+#define PROC_IN_CONCURRENT_REPACK 0x40 /* REPACK (CONCURRENTLY) */
-/* flags reset at EOXact */
+/* flags reset at EOXact. A bit of a misnomer ... */
#define PROC_VACUUM_STATE_MASK \
- (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+ (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+ PROC_IN_CONCURRENT_REPACK)
/*
* Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/Makefile b/src/test/modules/injection_points/Makefile
index 2cd7d87c533..f7663859fe2 100644
--- a/src/test/modules/injection_points/Makefile
+++ b/src/test/modules/injection_points/Makefile
@@ -15,6 +15,7 @@ REGRESS_OPTS = --dlpath=$(top_builddir)/src/test/regress
ISOLATION = basic \
inplace \
repack \
+ repack_deadlock \
repack_toast \
syscache-update-pruned \
heap_lock_update
diff --git a/src/test/modules/injection_points/expected/repack_deadlock.out b/src/test/modules/injection_points/expected/repack_deadlock.out
new file mode 100644
index 00000000000..a86e4767536
--- /dev/null
+++ b/src/test/modules/injection_points/expected/repack_deadlock.out
@@ -0,0 +1,63 @@
+Parsed test spec with 2 sessions
+
+starting permutation: wait_before_lock add_column wakeup_before_lock check1
+injection_points_attach
+-----------------------
+
+(1 row)
+
+step wait_before_lock:
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+ <waiting ...>
+step add_column:
+ alter table repack_deadlock add column noise text;
+ <waiting ...>
+step add_column: <... completed>
+ERROR: could not wait for concurrent REPACK
+step wakeup_before_lock:
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1:
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+
+count
+-----
+ 1
+(1 row)
+
+i|j
+-+-
+1|1
+2|2
+3|3
+4|4
+(4 rows)
+
+count
+-----
+ 4
+(1 row)
+
+injection_points_detach
+-----------------------
+
+(1 row)
+
diff --git a/src/test/modules/injection_points/meson.build b/src/test/modules/injection_points/meson.build
index a414abb924b..1cd88d6db65 100644
--- a/src/test/modules/injection_points/meson.build
+++ b/src/test/modules/injection_points/meson.build
@@ -46,6 +46,7 @@ tests += {
'basic',
'inplace',
'repack',
+ 'repack_deadlock',
'repack_toast',
'syscache-update-pruned',
'heap_lock_update',
diff --git a/src/test/modules/injection_points/specs/repack_deadlock.spec b/src/test/modules/injection_points/specs/repack_deadlock.spec
new file mode 100644
index 00000000000..9d23a6588c2
--- /dev/null
+++ b/src/test/modules/injection_points/specs/repack_deadlock.spec
@@ -0,0 +1,83 @@
+# Test REPACK with a concurrent transaction that would cause a deadlock
+setup
+{
+ CREATE EXTENSION injection_points;
+
+ CREATE TABLE repack_deadlock(i int PRIMARY KEY, j int);
+ INSERT INTO repack_deadlock(i, j) VALUES (1, 1), (2, 2), (3, 3), (4, 4);
+
+ CREATE TABLE relfilenodes(node oid);
+
+ CREATE TABLE data_s1(i int, j int);
+ CREATE TABLE data_s2(i int, j int);
+}
+
+teardown
+{
+ DROP TABLE repack_deadlock;
+ DROP EXTENSION injection_points;
+
+ DROP TABLE relfilenodes;
+ DROP TABLE data_s1;
+ DROP TABLE data_s2;
+}
+
+session s1
+setup
+{
+ SELECT injection_points_set_local();
+ SELECT injection_points_attach('repack-concurrently-before-lock', 'wait');
+}
+# Perform the initial load and wait for s2 to do some data changes.
+step wait_before_lock
+{
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+}
+# Check the table from the perspective of s1.
+#
+# Besides the contents, we also check that relfilenode has changed.
+
+# Have each session write the contents into a table and use FULL JOIN to check
+# if the outputs are identical.
+step check1
+{
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+}
+teardown
+{
+ SELECT injection_points_detach('repack-concurrently-before-lock');
+}
+
+session s2
+# Change the existing data. UPDATE changes both key and non-key columns. Also
+# update one row twice to test whether tuple version generated by this session
+# can be found.
+step add_column
+{
+ alter table repack_deadlock add column noise text;
+}
+
+step wakeup_before_lock
+{
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+}
+
+# Test if data changes introduced while one session is performing REPACK
+# CONCURRENTLY find their way into the table.
+permutation
+ wait_before_lock
+ add_column
+ wakeup_before_lock
+ check1
--
2.47.3
--r2slln3zpmilwu22
Content-Type: text/x-diff; charset=utf-8
Content-Disposition: attachment;
filename="v51-0007-Check-for-transaction-block-early-in-ExecRepack.patch"
^ permalink raw reply [nested|flat] 34+ messages in thread
* [PATCH v54 4/5] Error out any process that would block at REPACK
@ 2026-04-01 15:35 Antonin Houska <[email protected]>
0 siblings, 0 replies; 34+ messages in thread
From: Antonin Houska @ 2026-04-01 15:35 UTC (permalink / raw)
Any process waiting on REPACK to release its lock would actually cause
it to deadlock when it tries to upgrade its lock to AEL, losing all work
done to that point. We avoid this by teaching the deadlock detector to
raise an error when this condition is detected.
---
src/backend/commands/repack.c | 47 ++++++++---
src/backend/storage/lmgr/deadlock.c | 15 ++++
src/include/storage/proc.h | 6 +-
src/test/modules/injection_points/Makefile | 1 +
.../expected/repack_deadlock.out | 63 ++++++++++++++
src/test/modules/injection_points/meson.build | 1 +
.../specs/repack_deadlock.spec | 83 +++++++++++++++++++
7 files changed, 201 insertions(+), 15 deletions(-)
create mode 100644 src/test/modules/injection_points/expected/repack_deadlock.out
create mode 100644 src/test/modules/injection_points/specs/repack_deadlock.spec
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index b0898ce52d9..d1be6c60e88 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -292,6 +292,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
* to understand and we don't lose any functionality.
*/
PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+ /*
+ * Also set the PROC_IN_CONCURRENT_REPACK flag. This makes the
+ * deadlock checker cause anyone that would conflict with us to error
+ * out. It's important to set this flag ahead of actually locking the
+ * relation; it won't of course affect anyone until we do have a lock
+ * that others can conflict with.
+ */
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
}
/*
@@ -494,11 +506,8 @@ RepackLockLevel(bool concurrent)
* If indexOid is InvalidOid, the table will be rewritten in physical order
* instead of index order.
*
- * Note that, in the concurrent case, the function releases the lock at some
- * point, in order to get AccessExclusiveLock for the final steps (i.e. to
- * swap the relation files). To make things simpler, the caller should expect
- * OldHeap to be closed on return, regardless CLUOPT_CONCURRENT. (The
- * AccessExclusiveLock is kept till the end of the transaction.)
+ * On return, OldHeap is closed but locked with AccessExclusiveLock - the lock
+ * will be released at end of the transaction.
*
* 'cmd' indicates which command is being executed, to be used for error
* messages.
@@ -1004,10 +1013,8 @@ rebuild_relation(Relation OldHeap, Relation index, bool verbose,
* Note that the worker has to wait for all transactions with XID
* already assigned to finish. If some of those transactions is
* waiting for a lock conflicting with ShareUpdateExclusiveLock on our
- * table (e.g. it runs CREATE INDEX), we can end up in a deadlock.
- * Not sure this risk is worth unlocking/locking the table (and its
- * clustering index) and checking again if it's still eligible for
- * REPACK CONCURRENTLY.
+ * table (e.g. it runs CREATE INDEX), it should encounter ERROR in the
+ * deadlock checking code.
*/
start_repack_decoding_worker(tableOid);
@@ -3096,7 +3103,19 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
/*
- * Tuples and pages of the old heap will be gone, but the heap will stay.
+ * Now that we have all access-exclusive locks on all relations, we no
+ * longer want other processes to error out when trying to acquire a
+ * conflicting lock. Therefore, unset our flag.
+ */
+ LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+ MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+ ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+ LWLockRelease(ProcArrayLock);
+
+ /*
+ * Tuples and pages of the old heap will be gone, but the heap itself will
+ * stay. In order for predicate locks to continue to work, convert them
+ * to relation-level locks. We do this both for table and indexes.
*/
TransferPredicateLocksToHeapRelation(OldHeap);
foreach_ptr(RelationData, index, indexrels)
@@ -3370,9 +3389,11 @@ start_repack_decoding_worker(Oid relid)
/*
* The decoding setup must be done before the caller can have XID assigned
- * for any reason, otherwise the worker might end up in a deadlock,
- * waiting for the caller's transaction to end. Therefore wait here until
- * the worker indicates that it has the logical decoding initialized.
+ * for any reason, otherwise the worker might end up waiting for the
+ * caller's transaction to end. (Deadlock detector does not consider this
+ * a conflict because the worker is in the same locking group as the
+ * backend that launched it.) Therefore wait here until the worker
+ * indicates that it has the logical decoding initialized.
*/
ConditionVariablePrepareToSleep(&shared->cv);
for (;;)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..c20ac682b0d 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -620,6 +620,21 @@ FindLockCycleRecurseMember(PGPROC *checkProc,
proc->statusFlags & PROC_IS_AUTOVACUUM)
blocking_autovacuum_proc = proc;
+ /*
+ * Similarly, if we note that we're blocked by some
+ * process running REPACK (CONCURRENTLY), just fail. That
+ * process is going to upgrade its lock at some point, and
+ * it would be inappropriate for any other process to
+ * cause that to fail.
+ */
+ if (checkProc == MyProc &&
+ proc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+ ereport(ERROR,
+ errcode(ERRCODE_OBJECT_IN_USE),
+ errmsg("could not wait for concurrent REPACK"),
+ errdetail("Process %d waits for REPACK running on process %d",
+ MyProc->pid, proc->pid));
+
/* We're done looking at this proclock */
break;
}
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 22822fc68d7..a43cfe53558 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
#define PROC_AFFECTS_ALL_HORIZONS 0x20 /* this proc's xmin must be
* included in vacuum horizons
* in all databases */
+#define PROC_IN_CONCURRENT_REPACK 0x40 /* REPACK (CONCURRENTLY) */
-/* flags reset at EOXact */
+/* flags reset at EOXact. A bit of a misnomer ... */
#define PROC_VACUUM_STATE_MASK \
- (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+ (PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+ PROC_IN_CONCURRENT_REPACK)
/*
* Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/Makefile b/src/test/modules/injection_points/Makefile
index 2cd7d87c533..f7663859fe2 100644
--- a/src/test/modules/injection_points/Makefile
+++ b/src/test/modules/injection_points/Makefile
@@ -15,6 +15,7 @@ REGRESS_OPTS = --dlpath=$(top_builddir)/src/test/regress
ISOLATION = basic \
inplace \
repack \
+ repack_deadlock \
repack_toast \
syscache-update-pruned \
heap_lock_update
diff --git a/src/test/modules/injection_points/expected/repack_deadlock.out b/src/test/modules/injection_points/expected/repack_deadlock.out
new file mode 100644
index 00000000000..a86e4767536
--- /dev/null
+++ b/src/test/modules/injection_points/expected/repack_deadlock.out
@@ -0,0 +1,63 @@
+Parsed test spec with 2 sessions
+
+starting permutation: wait_before_lock add_column wakeup_before_lock check1
+injection_points_attach
+-----------------------
+
+(1 row)
+
+step wait_before_lock:
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+ <waiting ...>
+step add_column:
+ alter table repack_deadlock add column noise text;
+ <waiting ...>
+step add_column: <... completed>
+ERROR: could not wait for concurrent REPACK
+step wakeup_before_lock:
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1:
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+
+count
+-----
+ 1
+(1 row)
+
+i|j
+-+-
+1|1
+2|2
+3|3
+4|4
+(4 rows)
+
+count
+-----
+ 4
+(1 row)
+
+injection_points_detach
+-----------------------
+
+(1 row)
+
diff --git a/src/test/modules/injection_points/meson.build b/src/test/modules/injection_points/meson.build
index a414abb924b..1cd88d6db65 100644
--- a/src/test/modules/injection_points/meson.build
+++ b/src/test/modules/injection_points/meson.build
@@ -46,6 +46,7 @@ tests += {
'basic',
'inplace',
'repack',
+ 'repack_deadlock',
'repack_toast',
'syscache-update-pruned',
'heap_lock_update',
diff --git a/src/test/modules/injection_points/specs/repack_deadlock.spec b/src/test/modules/injection_points/specs/repack_deadlock.spec
new file mode 100644
index 00000000000..9d23a6588c2
--- /dev/null
+++ b/src/test/modules/injection_points/specs/repack_deadlock.spec
@@ -0,0 +1,83 @@
+# Test REPACK with a concurrent transaction that would cause a deadlock
+setup
+{
+ CREATE EXTENSION injection_points;
+
+ CREATE TABLE repack_deadlock(i int PRIMARY KEY, j int);
+ INSERT INTO repack_deadlock(i, j) VALUES (1, 1), (2, 2), (3, 3), (4, 4);
+
+ CREATE TABLE relfilenodes(node oid);
+
+ CREATE TABLE data_s1(i int, j int);
+ CREATE TABLE data_s2(i int, j int);
+}
+
+teardown
+{
+ DROP TABLE repack_deadlock;
+ DROP EXTENSION injection_points;
+
+ DROP TABLE relfilenodes;
+ DROP TABLE data_s1;
+ DROP TABLE data_s2;
+}
+
+session s1
+setup
+{
+ SELECT injection_points_set_local();
+ SELECT injection_points_attach('repack-concurrently-before-lock', 'wait');
+}
+# Perform the initial load and wait for s2 to do some data changes.
+step wait_before_lock
+{
+ REPACK (CONCURRENTLY) repack_deadlock USING INDEX repack_deadlock_pkey;
+}
+# Check the table from the perspective of s1.
+#
+# Besides the contents, we also check that relfilenode has changed.
+
+# Have each session write the contents into a table and use FULL JOIN to check
+# if the outputs are identical.
+step check1
+{
+ INSERT INTO relfilenodes(node)
+ SELECT relfilenode FROM pg_class WHERE relname='repack_deadlock';
+
+ SELECT count(DISTINCT node) FROM relfilenodes;
+
+ SELECT i, j FROM repack_deadlock ORDER BY i, j;
+
+ INSERT INTO data_s1(i, j)
+ SELECT i, j FROM repack_deadlock;
+
+ SELECT count(*)
+ FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
+ WHERE d1.i ISNULL OR d2.i ISNULL;
+}
+teardown
+{
+ SELECT injection_points_detach('repack-concurrently-before-lock');
+}
+
+session s2
+# Change the existing data. UPDATE changes both key and non-key columns. Also
+# update one row twice to test whether tuple version generated by this session
+# can be found.
+step add_column
+{
+ alter table repack_deadlock add column noise text;
+}
+
+step wakeup_before_lock
+{
+ SELECT injection_points_wakeup('repack-concurrently-before-lock');
+}
+
+# Test if data changes introduced while one session is performing REPACK
+# CONCURRENTLY find their way into the table.
+permutation
+ wait_before_lock
+ add_column
+ wakeup_before_lock
+ check1
--
2.47.3
--cdhh4t7ukb6tnjoq
Content-Type: text/x-diff; charset=utf-8
Content-Disposition: attachment;
filename="v54-0005-Reserve-replication-slots-specifically-for-REPAC.patch"
^ permalink raw reply [nested|flat] 34+ messages in thread
end of thread, other threads:[~2026-04-01 15:35 UTC | newest]
Thread overview: 34+ messages (download: mbox mbox.gz follow: Atom feed)
-- links below jump to the message on this page --
2018-03-05 03:30 Typo in src/backend/access/hash/README Thomas Munro <[email protected]>
2018-03-05 03:40 ` Justin Pryzby <[email protected]>
2018-03-05 06:53 ` Amit Kapila <[email protected]>
2018-03-29 18:38 ` Bruce Momjian <[email protected]>
2022-10-29 02:54 refactoring relation extension and BufferAlloc(), faster COPY Andres Freund <[email protected]>
2022-10-29 13:03 ` Re: refactoring relation extension and BufferAlloc(), faster COPY Bharath Rupireddy <[email protected]>
2022-10-29 18:39 ` Re: refactoring relation extension and BufferAlloc(), faster COPY Andres Freund <[email protected]>
2023-01-27 20:05 Re: Set arbitrary GUC options during initdb Robert Haas <[email protected]>
2026-03-25 19:35 [PATCH v47 8/9] Error out any process that would block at REPACK Álvaro Herrera <[email protected]>
2026-03-25 19:35 [PATCH v45 7/8] Error out any process that would block at REPACK Álvaro Herrera <[email protected]>
2026-03-25 19:35 [PATCH v48 6/7] Error out any process that would block at REPACK Álvaro Herrera <[email protected]>
2026-03-25 19:35 [PATCH v46 6/7] Error out any process that would block at REPACK Álvaro Herrera <[email protected]>
2026-03-25 19:35 [PATCH v47 8/9] Error out any process that would block at REPACK Álvaro Herrera <[email protected]>
2026-03-25 19:35 [PATCH v46 6/7] Error out any process that would block at REPACK Álvaro Herrera <[email protected]>
2026-03-25 19:35 [PATCH v45 7/8] Error out any process that would block at REPACK Álvaro Herrera <[email protected]>
2026-03-25 19:35 [PATCH v48 6/7] Error out any process that would block at REPACK Álvaro Herrera <[email protected]>
2026-04-01 15:35 [PATCH v52 06/10] Error out any process that would block at REPACK Antonin Houska <[email protected]>
2026-04-01 15:35 [PATCH v49 5/7] Error out any process that would block at REPACK Antonin Houska <[email protected]>
2026-04-01 15:35 [PATCH v51 06/10] Error out any process that would block at REPACK Antonin Houska <[email protected]>
2026-04-01 15:35 [PATCH v51 06/10] Error out any process that would block at REPACK Antonin Houska <[email protected]>
2026-04-01 15:35 [PATCH v56 3/3] Error out any process that would block at REPACK Antonin Houska <[email protected]>
2026-04-01 15:35 [PATCH v56 3/3] Error out any process that would block at REPACK Antonin Houska <[email protected]>
2026-04-01 15:35 [PATCH 1/2] Error out any process that would block at REPACK Antonin Houska <[email protected]>
2026-04-01 15:35 [PATCH v50 6/8] Error out any process that would block at REPACK Antonin Houska <[email protected]>
2026-04-01 15:35 [PATCH] Error out any process that would block at REPACK Antonin Houska <[email protected]>
2026-04-01 15:35 [PATCH v54 4/5] Error out any process that would block at REPACK Antonin Houska <[email protected]>
2026-04-01 15:35 [PATCH v54 4/5] Error out any process that would block at REPACK Antonin Houska <[email protected]>
2026-04-01 15:35 [PATCH v50 6/8] Error out any process that would block at REPACK Antonin Houska <[email protected]>
2026-04-01 15:35 [PATCH v53 6/7] Error out any process that would block at REPACK Antonin Houska <[email protected]>
2026-04-01 15:35 [PATCH 1/2] Error out any process that would block at REPACK Antonin Houska <[email protected]>
2026-04-01 15:35 [PATCH v53 6/7] Error out any process that would block at REPACK Antonin Houska <[email protected]>
2026-04-01 15:35 [PATCH] Error out any process that would block at REPACK Antonin Houska <[email protected]>
2026-04-01 15:35 [PATCH v49 5/7] Error out any process that would block at REPACK Antonin Houska <[email protected]>
2026-04-01 15:35 [PATCH v52 06/10] Error out any process that would block at REPACK Antonin Houska <[email protected]>
This inbox is served by agora; see mirroring instructions
for how to clone and mirror all data and code used for this inbox