From: Heikki Linnakangas <hlinnaka@iki.fi>
To: exclusion@gmail.com
To: pgsql-bugs@lists.postgresql.org
Subject: Re: BUG #18129: GiST index produces incorrect query results
Date: Sun, 24 Sep 2023 01:06:35 +0300
Message-ID: <80e4ef74-4a00-4631-963f-3dd11e64c6fe@iki.fi> (raw)
In-Reply-To: <18129-caca016eaf0c3702@postgresql.org>
References: <18129-caca016eaf0c3702@postgresql.org>
On 22/09/2023 09:00, PG Bug reporting form wrote:
> First bad commit is 9155580fd, on which three different outcomes are
> possible: test passes/test fails/assertion fails.
> Assertion failures, that we had observed after 9155580fd, were fixed by
> a7ee7c851 and 741b88435, but apparently there is another defect induced
> by that gist implementation change.
This is indeed a similar bug to that fixed in 741b88435 and a7ee7c851.
In those commits, we fixed the insertion code so that whenever we insert
a downlink to the parent page and that causes the parent page to split,
we invalidate the 'downlinkoffnum' we had memorized for the child.
Because when the parent page is split, the downlink for the child likely
moves.
However, the downlink can move even when we *don't* split the parent.
Splitting a page not only inserts the new tuple for the new right page,
but also updates the old downlink, because the key representing what is
now the left page covers less keys than the original page. And updating
a tuple is implemented as delete+insert. The delete moves existing
tuples on the page! In other words, this code and comment in
gistfinishsplit():
> if (gistinserttuples(state, stack->parent, giststate,
> tuples, 2,
> stack->downlinkoffnum,
> left->buf, right->buf,
> true, /* Unlock parent */
> unlockbuf /* Unlock stack->buffer if caller wants
> * that */
> ))
> {
> /*
> * If the parent page was split, the downlink might have moved.
> */
> stack->downlinkoffnum = InvalidOffsetNumber;
> }
is wrong. We must invalidate 'downlinkoffnum' whether or not the parent
page was split, because updating an existing tuple can also move other
tuples on the page.
Attached is a set of patches to fix this:
1. To catch the problem earlier - already during the index build - I
modified getFindCorrectParent() to always find the parent the hard way
by scanning the parent page, even if 'downlinkoffnum' is set. Once the
downlink is found, it compares it with the 'downlinkoffnum', and PANICs
if it wasn't where we expected. That defeats the point of memorizing the
'downlinkoffnum' in the first place, but it's a useful cross-check.
I don't intend to commit this, but it was useful during debugging.
2. A straightforward fix for the bug.
3. Looking at gistFindCorrectParent, when it steps right, it also
doesn't update 'downlinkoffnum' of the parent page. Surely if we step to
a different block altogether, the memorized downlink position of the
previous block no longer applies.
Worryingly, I haven't been able to get a test failure caused by that.
Maybe there are some reasons why we never need to split the parent after
that, not sure. But it sure looks wrong.
4. Given how many bugs we've already had with this, I'd like to make
this more robust. Currently, we meticulously track if we have made any
inserts that invalidate the memorized location of the downlink. Instead,
gistFindCorrectParent() could treat the memorized location as just a
hint, and always check if the downlink is still at the memorized
location. If it is, great, and if it's not, find it the hard way. That
makes it unnecessary to clear 'downlinkoffnum' at the right places. We
still need the 'retry_from_parent' flag, though.
5. Re-indent the code after previous commit. Separated for easier review.
I'd also like to add something to our regression test suite for better
coverage of this. The intarray test case is very useful, but it'd be
nice to have something in the main regression suite, as this is a
general GiST issue, not related to intarray. I haven't spent any time on
that yet.
--
Heikki Linnakangas
Neon (https://neon.tech)
Attachments:
[text/x-patch] 0001-Add-assertion-that-downlinkoffnum-is-set-correctly.patch (1.5K, ../80e4ef74-4a00-4631-963f-3dd11e64c6fe@iki.fi/2-0001-Add-assertion-that-downlinkoffnum-is-set-correctly.patch)
download | inline diff:
From adca01fd7dd566f6da905c800480176978b5663f Mon Sep 17 00:00:00 2001
From: Heikki Linnakangas <heikki.linnakangas@iki.fi>
Date: Sun, 24 Sep 2023 00:52:20 +0300
Subject: [PATCH 1/5] Add assertion that 'downlinkoffnum' is set correctly.
I don't intend to push this, but it was useful while debugging.
---
src/backend/access/gist/gist.c | 16 ++++++++++++++++
1 file changed, 16 insertions(+)
diff --git a/src/backend/access/gist/gist.c b/src/backend/access/gist/gist.cindex b1f19f6a8e6..10405239027 100644--- a/src/backend/access/gist/gist.c+++ b/src/backend/access/gist/gist.c@@ -1022,6 +1022,13 @@ gistFindCorrectParent(Relation r, GISTInsertStack *child)
{
GISTInsertStack *parent = child->parent;
+#ifdef USE_ASSERT_CHECKING+ int save_downlinkoffnum = child->downlinkoffnum;+ BlockNumber save_parent_blkno = parent->blkno;++ child->downlinkoffnum = InvalidOffsetNumber;+#endif+
gistcheckpage(r, parent->buffer);
parent->page = (Page) BufferGetPage(parent->buffer);
@@ -1047,6 +1054,15 @@ gistFindCorrectParent(Relation r, GISTInsertStack *child)
{
/* yes!!, found */
child->downlinkoffnum = i;
+#ifdef USE_ASSERT_CHECKING+ if (parent->lsn == PageGetLSN(parent->page) && save_downlinkoffnum != InvalidOffsetNumber) {+ if (i != save_downlinkoffnum) {+ elog(PANIC, "expected to find downlink for %u on blk %u offset %u, but it was found on blk %u offset %u",+ child->blkno, save_parent_blkno, save_downlinkoffnum,+ parent->blkno, i);+ }+ }+#endif
return;
}
}
--
2.39.2
[text/x-patch] 0002-Fix-another-bug-in-GiST-index-build.patch (3.3K, ../80e4ef74-4a00-4631-963f-3dd11e64c6fe@iki.fi/3-0002-Fix-another-bug-in-GiST-index-build.patch)
download | inline diff:
From f405db22b6ea2cb39a9b063e4759c0f97fa281a6 Mon Sep 17 00:00:00 2001
From: Heikki Linnakangas <heikki.linnakangas@iki.fi>
Date: Sun, 24 Sep 2023 00:22:30 +0300
Subject: [PATCH 2/5] Fix another bug in GiST index build.
Yet another bug in the ilk of commits a7ee7c851 and 741b88435. In
741b88435, we took care to clear the memorized location of the
downlink when we split the parent page, because splitting the parent
page can move the downlink. But we missed that even *updating* a tuple
on the parent can move it, because updating a tuple on a gist page is
implemented as a delete+insert, so the updated tuple gets moved to the
end of the page. To fix, clear the downlink when we updatie a tuple on
the parent page, even if it's not split.
Backpatch XXX
Reported-by: Alexander Lakhin
Discussion: https://www.postgresql.org/message-id/18129-caca016eaf0c3702@postgresql.org
---
src/backend/access/gist/gist.c | 37 +++++++++++++++++++---------------
1 file changed, 21 insertions(+), 16 deletions(-)
diff --git a/src/backend/access/gist/gist.c b/src/backend/access/gist/gist.cindex 10405239027..d4469eb7cc4 100644--- a/src/backend/access/gist/gist.c+++ b/src/backend/access/gist/gist.c@@ -1272,7 +1272,10 @@ gistinserttuple(GISTInsertState *state, GISTInsertStack *stack,
*
* Returns 'true' if the page had to be split. Note that if the page was
* split, the inserted/updated tuples might've been inserted to a right
- * sibling of stack->buffer instead of stack->buffer itself.+ * sibling of stack->buffer instead of stack->buffer itself. Even if the page+ * was not split but an old tuple was updated, any existing tuples on the page+ * including the one that's updated might move to a different offset on th+ * epage.
*/
static bool
gistinserttuples(GISTInsertState *state, GISTInsertStack *stack,
@@ -1371,7 +1374,8 @@ gistfinishsplit(GISTInsertState *state, GISTInsertStack *stack,
{
/*
* If the parent page was split, the existing downlink might have
- * moved.+ * moved. This assumes that gistinsettuples() doesn't move the+ * tuples around unless the page is split.
*/
stack->downlinkoffnum = InvalidOffsetNumber;
}
@@ -1389,20 +1393,21 @@ gistfinishsplit(GISTInsertState *state, GISTInsertStack *stack,
tuples[0] = left->downlink;
tuples[1] = right->downlink;
gistFindCorrectParent(state->r, stack);
- if (gistinserttuples(state, stack->parent, giststate,- tuples, 2,- stack->downlinkoffnum,- left->buf, right->buf,- true, /* Unlock parent */- unlockbuf /* Unlock stack->buffer if caller wants- * that */- ))- {- /*- * If the parent page was split, the downlink might have moved.- */- stack->downlinkoffnum = InvalidOffsetNumber;- }+ (void) gistinserttuples(state, stack->parent, giststate,+ tuples, 2,+ stack->downlinkoffnum,+ left->buf, right->buf,+ true, /* Unlock parent */+ unlockbuf /* Unlock stack->buffer if caller+ * wants that */+ );++ /*+ * The downlink might have moved when we updated it. Even if the page+ * wasn't split, because gistinserttuples() implements updating the old+ * tuple by removing and re-inserting it!+ */+ stack->downlinkoffnum = InvalidOffsetNumber;
Assert(left->buf == stack->buffer);
--
2.39.2
[text/x-patch] 0003-Also-clear-downlinkoffnum-when-the-parent-block-is-c.patch (1.0K, ../80e4ef74-4a00-4631-963f-3dd11e64c6fe@iki.fi/4-0003-Also-clear-downlinkoffnum-when-the-parent-block-is-c.patch)
download | inline diff:
From 99a71e9543ba2cae0bf91403f78c57671a6aebf8 Mon Sep 17 00:00:00 2001
From: Heikki Linnakangas <heikki.linnakangas@iki.fi>
Date: Sat, 23 Sep 2023 20:50:38 +0300
Subject: [PATCH 3/5] Also clear 'downlinkoffnum' when the parent block is
changed.
While looking at bug #18129, I noticed this spot where we update the
parent block in the stack, but we don't clear the
'downlinkoffnum'. This case wasn't hit by the test case, but it could
lead to similar bugs.
---
src/backend/access/gist/gist.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/src/backend/access/gist/gist.c b/src/backend/access/gist/gist.cindex d4469eb7cc4..b3749bbb432 100644--- a/src/backend/access/gist/gist.c+++ b/src/backend/access/gist/gist.c@@ -1068,6 +1068,7 @@ gistFindCorrectParent(Relation r, GISTInsertStack *child)
}
parent->blkno = GistPageGetOpaque(parent->page)->rightlink;
+ parent->downlinkoffnum = InvalidOffsetNumber;
UnlockReleaseBuffer(parent->buffer);
if (parent->blkno == InvalidBlockNumber)
{
--
2.39.2
[text/x-patch] 0004-Treat-downlinkoffnum-only-as-a-hint-in-gistFindCorre.patch (5.9K, ../80e4ef74-4a00-4631-963f-3dd11e64c6fe@iki.fi/5-0004-Treat-downlinkoffnum-only-as-a-hint-in-gistFindCorre.patch)
download | inline diff:
From 661bd1bbb029f71e0dbe7d37a731cae50afb862b Mon Sep 17 00:00:00 2001
From: Heikki Linnakangas <heikki.linnakangas@iki.fi>
Date: Sun, 24 Sep 2023 00:49:08 +0300
Subject: [PATCH 4/5] Treat 'downlinkoffnum' only as a hint in
gistFindCorrectParent.
We have had a number of bugs in trying to meticulously clear
'downlinkoffnum' when we update a parent page. But instead of trying
hard to clear it all the right places, we can treat is merely as a
hint, so that if we forget to clear it, it still works.
This makes the fixes in commit XXX and a7ee7c8513 unnecessary. It
doesn't hurt to still clear 'downlinkoffnum', but everything would
work without it now.
---
src/backend/access/gist/gist.c | 69 ++++++++++++++++++----------------
1 file changed, 36 insertions(+), 33 deletions(-)
diff --git a/src/backend/access/gist/gist.c b/src/backend/access/gist/gist.cindex b3749bbb432..42f9408becb 100644--- a/src/backend/access/gist/gist.c+++ b/src/backend/access/gist/gist.c@@ -1018,34 +1018,46 @@ gistFindPath(Relation r, BlockNumber child, OffsetNumber *downlinkoffnum)
* remain so at exit, but it might not be the same page anymore.
*/
static void
-gistFindCorrectParent(Relation r, GISTInsertStack *child)+gistFindCorrectParent(Relation r, GISTInsertStack *child, bool is_build)
{
GISTInsertStack *parent = child->parent;
--#ifdef USE_ASSERT_CHECKING- int save_downlinkoffnum = child->downlinkoffnum;- BlockNumber save_parent_blkno = parent->blkno;-- child->downlinkoffnum = InvalidOffsetNumber;-#endif+ ItemId iid;+ IndexTuple idxtuple;+ OffsetNumber i,+ maxoff;+ GISTInsertStack *ptr;
gistcheckpage(r, parent->buffer);
parent->page = (Page) BufferGetPage(parent->buffer);
+ maxoff = PageGetMaxOffsetNumber(parent->page);- /* here we don't need to distinguish between split and page update */- if (child->downlinkoffnum == InvalidOffsetNumber ||- parent->lsn != PageGetLSN(parent->page))+ /* Check if the downlink is still where it was before */+ if (child->downlinkoffnum != InvalidOffsetNumber && child->downlinkoffnum <= maxoff)
{
- /* parent is changed, look child in right links until found */- OffsetNumber i,- maxoff;- ItemId iid;- IndexTuple idxtuple;- GISTInsertStack *ptr;+ iid = PageGetItemId(parent->page, child->downlinkoffnum);+ idxtuple = (IndexTuple) PageGetItem(parent->page, iid);+ if (ItemPointerGetBlockNumber(&(idxtuple->t_tid)) == child->blkno)+ return; /* still there */+ }+ /*+ * The page has changed since we looked. During normal operation, every+ * update of a page changes its LSN, so the LSN we memorized should have+ * changed too. During index build, however, we don't WAL-log the changes+ * until we have built the index, so the LSN doesn't change. There is no+ * concurrent activity during index build, but we might have changed the+ * parent ourselves.+ */+ Assert(parent->lsn != PageGetLSN(parent->page) || is_build);++ /*+ * Scan the page to re-find the downlink. If the page was split, it might+ * have moved to a different page, so follow the right links until we find+ * it.+ */+ {
while (true)
{
- maxoff = PageGetMaxOffsetNumber(parent->page);
for (i = FirstOffsetNumber; i <= maxoff; i = OffsetNumberNext(i))
{
iid = PageGetItemId(parent->page, i);
@@ -1054,15 +1066,6 @@ gistFindCorrectParent(Relation r, GISTInsertStack *child)
{
/* yes!!, found */
child->downlinkoffnum = i;
-#ifdef USE_ASSERT_CHECKING- if (parent->lsn == PageGetLSN(parent->page) && save_downlinkoffnum != InvalidOffsetNumber) {- if (i != save_downlinkoffnum) {- elog(PANIC, "expected to find downlink for %u on blk %u offset %u, but it was found on blk %u offset %u",- child->blkno, save_parent_blkno, save_downlinkoffnum,- parent->blkno, i);- }- }-#endif
return;
}
}
@@ -1114,7 +1117,7 @@ gistFindCorrectParent(Relation r, GISTInsertStack *child)
/* make recursive call to normal processing */
LockBuffer(child->parent->buffer, GIST_EXCLUSIVE);
- gistFindCorrectParent(r, child);+ gistFindCorrectParent(r, child, is_build);
}
}
@@ -1123,7 +1126,7 @@ gistFindCorrectParent(Relation r, GISTInsertStack *child)
*/
static IndexTuple
gistformdownlink(Relation rel, Buffer buf, GISTSTATE *giststate,
- GISTInsertStack *stack)+ GISTInsertStack *stack, bool is_build)
{
Page page = BufferGetPage(buf);
OffsetNumber maxoff;
@@ -1164,7 +1167,7 @@ gistformdownlink(Relation rel, Buffer buf, GISTSTATE *giststate,
ItemId iid;
LockBuffer(stack->parent->buffer, GIST_EXCLUSIVE);
- gistFindCorrectParent(rel, stack);+ gistFindCorrectParent(rel, stack, is_build);
iid = PageGetItemId(stack->parent->page, stack->downlinkoffnum);
downlink = (IndexTuple) PageGetItem(stack->parent->page, iid);
downlink = CopyIndexTuple(downlink);
@@ -1210,7 +1213,7 @@ gistfixsplit(GISTInsertState *state, GISTSTATE *giststate)
page = BufferGetPage(buf);
/* Form the new downlink tuples to insert to parent */
- downlink = gistformdownlink(state->r, buf, giststate, stack);+ downlink = gistformdownlink(state->r, buf, giststate, stack, state->is_build);
si->buf = buf;
si->downlink = downlink;
@@ -1367,7 +1370,7 @@ gistfinishsplit(GISTInsertState *state, GISTInsertStack *stack,
right = (GISTPageSplitInfo *) list_nth(splitinfo, pos);
left = (GISTPageSplitInfo *) list_nth(splitinfo, pos - 1);
- gistFindCorrectParent(state->r, stack);+ gistFindCorrectParent(state->r, stack, state->is_build);
if (gistinserttuples(state, stack->parent, giststate,
&right->downlink, 1,
InvalidOffsetNumber,
@@ -1393,7 +1396,7 @@ gistfinishsplit(GISTInsertState *state, GISTInsertStack *stack,
*/
tuples[0] = left->downlink;
tuples[1] = right->downlink;
- gistFindCorrectParent(state->r, stack);+ gistFindCorrectParent(state->r, stack, state->is_build);
(void) gistinserttuples(state, stack->parent, giststate,
tuples, 2,
stack->downlinkoffnum,
--
2.39.2
[text/x-patch] 0005-Re-indent-the-code.patch (4.3K, ../80e4ef74-4a00-4631-963f-3dd11e64c6fe@iki.fi/6-0005-Re-indent-the-code.patch)
download | inline diff:
From 2de2283bba14674a66eecb46fc99d95b26bb1002 Mon Sep 17 00:00:00 2001
From: Heikki Linnakangas <heikki.linnakangas@iki.fi>
Date: Sun, 24 Sep 2023 00:58:24 +0300
Subject: [PATCH 5/5] Re-indent the code.
Remove the stray braces and re-indent the code. This should be
squashed with the previous commit, but it makes it easier to review
when there's not so much whitespace change.
---
src/backend/access/gist/gist.c | 99 +++++++++++++++++-----------------
1 file changed, 48 insertions(+), 51 deletions(-)
diff --git a/src/backend/access/gist/gist.c b/src/backend/access/gist/gist.cindex 42f9408becb..ec9d7035468 100644--- a/src/backend/access/gist/gist.c+++ b/src/backend/access/gist/gist.c@@ -1055,70 +1055,67 @@ gistFindCorrectParent(Relation r, GISTInsertStack *child, bool is_build)
* have moved to a different page, so follow the right links until we find
* it.
*/
+ while (true)
{
- while (true)+ for (i = FirstOffsetNumber; i <= maxoff; i = OffsetNumberNext(i))
{
- for (i = FirstOffsetNumber; i <= maxoff; i = OffsetNumberNext(i))- {- iid = PageGetItemId(parent->page, i);- idxtuple = (IndexTuple) PageGetItem(parent->page, iid);- if (ItemPointerGetBlockNumber(&(idxtuple->t_tid)) == child->blkno)- {- /* yes!!, found */- child->downlinkoffnum = i;- return;- }- }-- parent->blkno = GistPageGetOpaque(parent->page)->rightlink;- parent->downlinkoffnum = InvalidOffsetNumber;- UnlockReleaseBuffer(parent->buffer);- if (parent->blkno == InvalidBlockNumber)+ iid = PageGetItemId(parent->page, i);+ idxtuple = (IndexTuple) PageGetItem(parent->page, iid);+ if (ItemPointerGetBlockNumber(&(idxtuple->t_tid)) == child->blkno)
{
- /*- * End of chain and still didn't find parent. It's a very-very- * rare situation when root splitted.- */- break;+ /* yes!!, found */+ child->downlinkoffnum = i;+ return;
}
- parent->buffer = ReadBuffer(r, parent->blkno);- LockBuffer(parent->buffer, GIST_EXCLUSIVE);- gistcheckpage(r, parent->buffer);- parent->page = (Page) BufferGetPage(parent->buffer);
}
- /*- * awful!!, we need search tree to find parent ... , but before we- * should release all old parent- */-- ptr = child->parent->parent; /* child->parent already released- * above */- while (ptr)+ parent->blkno = GistPageGetOpaque(parent->page)->rightlink;+ parent->downlinkoffnum = InvalidOffsetNumber;+ UnlockReleaseBuffer(parent->buffer);+ if (parent->blkno == InvalidBlockNumber)
{
- ReleaseBuffer(ptr->buffer);- ptr = ptr->parent;+ /*+ * End of chain and still didn't find parent. It's a very-very+ * rare situation when root splitted.+ */+ break;
}
+ parent->buffer = ReadBuffer(r, parent->blkno);+ LockBuffer(parent->buffer, GIST_EXCLUSIVE);+ gistcheckpage(r, parent->buffer);+ parent->page = (Page) BufferGetPage(parent->buffer);+ }- /* ok, find new path */- ptr = parent = gistFindPath(r, child->blkno, &child->downlinkoffnum);+ /*+ * awful!!, we need search tree to find parent ... , but before we should+ * release all old parent+ */- /* read all buffers as expected by caller */- /* note we don't lock them or gistcheckpage them here! */- while (ptr)- {- ptr->buffer = ReadBuffer(r, ptr->blkno);- ptr->page = (Page) BufferGetPage(ptr->buffer);- ptr = ptr->parent;- }+ ptr = child->parent->parent; /* child->parent already released above */+ while (ptr)+ {+ ReleaseBuffer(ptr->buffer);+ ptr = ptr->parent;+ }- /* install new chain of parents to stack */- child->parent = parent;+ /* ok, find new path */+ ptr = parent = gistFindPath(r, child->blkno, &child->downlinkoffnum);- /* make recursive call to normal processing */- LockBuffer(child->parent->buffer, GIST_EXCLUSIVE);- gistFindCorrectParent(r, child, is_build);+ /* read all buffers as expected by caller */+ /* note we don't lock them or gistcheckpage them here! */+ while (ptr)+ {+ ptr->buffer = ReadBuffer(r, ptr->blkno);+ ptr->page = (Page) BufferGetPage(ptr->buffer);+ ptr = ptr->parent;
}
++ /* install new chain of parents to stack */+ child->parent = parent;++ /* make recursive call to normal processing */+ LockBuffer(child->parent->buffer, GIST_EXCLUSIVE);+ gistFindCorrectParent(r, child, is_build);
}
/*
--
2.39.2
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Reply to all the recipients using the --to and --cc options:
reply via email
To: pgsql-bugs@postgresql.org
Cc: hlinnaka@iki.fi, exclusion@gmail.com, pgsql-bugs@lists.postgresql.org
Subject: Re: BUG #18129: GiST index produces incorrect query results
In-Reply-To: <80e4ef74-4a00-4631-963f-3dd11e64c6fe@iki.fi>
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
This inbox is served by DDX for PostgreSQL; see mirroring instructions
for how to clone and mirror all data and code used for this inbox