From: Fabien COELHO <coelho@cri.ensmp.fr>
To: Tom Lane <tgl@sss.pgh.pa.us>
Cc: Alastair McKinley <a.mckinley@analyticsengines.com>
Cc: PostgreSQL Bugs <pgsql-bugs@lists.postgresql.org>
Cc: Peter Eisentraut <peter.eisentraut@enterprisedb.com>
Subject: Re: psql 15beta1 does not print notices on the console until transaction completes
Date: Fri, 3 Jun 2022 20:25:23 +0200 (CEST)
Message-ID: <alpine.DEB.2.22.394.2206032001450.622878@pseudo> (raw)
In-Reply-To: <alpine.DEB.2.22.394.2205311331140.1672363@pseudo>
References: <PAXPR02MB760039506C87A2083AD85575E3DA9@PAXPR02MB7600.eurprd02.prod.outlook.com>
<3662994.1653856025@sss.pgh.pa.us>
<alpine.DEB.2.22.394.2205300849090.1166441@pseudo>
<alpine.DEB.2.22.394.2205301921520.1376708@pseudo>
<alpine.DEB.2.22.394.2205311331140.1672363@pseudo>
>> I've looked into it. This behavior is triggered by the fact that "psql" is
>> trying to keep notices and results in some logical order.
>>
>> The issue did not arise before because intermediate results where coldly
>> discarded, so it did not show whether they were discarded before of after
>> some notice was already shown:-)
>>
>> If notices are to be shown "right away", it means that some notice may be
>> shown *before* the previous result has been shown, just because psql may
>> not have had time to process it and they are just sent through the protocol
>> and arriving for the next result, as shown with the attached patch on
>> non-regression tests:
>>
>> SELECT 1 AS one \; SELECT warn('1.5') \; SELECT 2 AS two ;
>> +NOTICE: warn 1.5
>> +CONTEXT: PL/pgSQL function warn(text) line 2 at RAISE
>> one
>> -----
>> 1
>> (1 row)
>>
>> -NOTICE: warn 1.5
>> -CONTEXT: PL/pgSQL function warn(text) line 2 at RAISE
>> warn
>> ------
>> t
>>
>> Now the second statement warning is shown before the first statement
>> output. I'm pretty sure that the initial order shoul show up from time to
>> time, just because of the underlying race condition.
>>
>> Would this be okay? It simplifies the code significantly, but introduces
>> possible race conditions in the output, which could be hidden from the non
>> regression tests by removing the tests…
>
> This would be the simplest solution.
>
>> If notices are to be shown "right away" when processing a result, but not
>> quite "right away" when the previous result is not yet finished, this is
>> probably achievable but would require some effort and not so clear code.
>
>> Maybe psql could not get the next result when we do not need to know which
>> is last, but the execution would have somehow 2 different paths. I have to
>> try to see how terrible or not the resulting code would be.
>
> I've tried that (see v2 attached), but it does not work as I hoped: the fact
> that a result has not yet been extracted does not preclude the server from
> sending the notices relating to a statement, so delaying the getResult does
> not fix the display order problem in anyway. I was quite naïve.
>
>> Another approach could be do drop the "only show the last result"
>> compatibility, because then "psql" would not need to know that a result is
>> the last, which requires to trigger the next one before closing the
>> previous one, hence opening the race condition.
>
> That would not work, either, for the reason stated above.
>
> In order to have the best of both world, this requires more hopus-pocus:
> accumulating the notices *but* displaying them as soon as dealing with a
> statement is done *and* also while waiting for getResult. I see how I can try
> to reduce a lot the out-of-order window… but not how to fully get rid of it.
After more investigation and quite a few variants, the end result is that
you cannot have the best of both words in all cases, i.e. instantaneous
notices while keeping the order of results and their notices. The
underlying reason for that is that some cases (gexec and others,) need to
know which result is last because they need to to something special about
it, which always requires to wait for the next result before doing
anything with the previous one, so if the next one takes time, then the
previous one is somehow suspended.
In the previous setting, because intermediate results where coldly
skipped, the user could not see whether notices for one statement where
coming before or after some result, so all was well hidden, you just got
the notices as they were sent and the final result displayed. The
underlying truth was that it depended:-)
My short term recommendation is to remove the attempt at keeping the
notice/result order for a tradeoff of getting the notices as soon as they
are emitted, i.e. basically patch 1 and the acceptance of the diffs,
called 1-bis in the attached. Alternatively, the attached patch 3 keeps
the order in the simple case, but not in the gset case, at the price of
some complexity. I would not bet too much that these codes are immune to
race conditions.
--
Fabien.
Attachments:download all (2 files, .tar.gz)psql-notices-1-bis.patch (text/x-diff, 5.1K) downloaddiff --git a/src/bin/psql/common.c b/src/bin/psql/common.cindex 03e6d9ce8f..974959c595 100644--- a/src/bin/psql/common.c+++ b/src/bin/psql/common.c@@ -1061,44 +1061,6 @@ PrintQueryResult(PGresult *result, bool last, bool is_watch, const printQueryOpt
return success;
}
-/*- * Data structure and functions to record notices while they are- * emitted, so that they can be shown later.- *- * We need to know which result is last, which requires to extract- * one result in advance, hence two buffers are needed.- */-struct t_notice_messages-{- PQExpBufferData messages[2];- int current;-};--/*- * Store notices in appropriate buffer, for later display.- */-static void-AppendNoticeMessage(void *arg, const char *msg)-{- struct t_notice_messages *notices = arg;-- appendPQExpBufferStr(¬ices->messages[notices->current], msg);-}--/*- * Show notices stored in buffer, which is then reset.- */-static void-ShowNoticeMessage(struct t_notice_messages *notices)-{- PQExpBufferData *current = ¬ices->messages[notices->current];-- if (*current->data != '\0')- pg_log_info("%s", current->data);- resetPQExpBuffer(current);-}--
/*
* SendQuery: send the query string to the backend
* (and print out result)
@@ -1483,7 +1445,6 @@ ExecQueryAndProcessResults(const char *query, double *elapsed_msec, bool *svpt_g
instr_time before,
after;
PGresult *result;
- struct t_notice_messages notices;
if (timing)
INSTR_TIME_SET_CURRENT(before);
@@ -1513,12 +1474,6 @@ ExecQueryAndProcessResults(const char *query, double *elapsed_msec, bool *svpt_g
return 0;
}
- /* intercept notices */- notices.current = 0;- initPQExpBuffer(¬ices.messages[0]);- initPQExpBuffer(¬ices.messages[1]);- PQsetNoticeProcessor(pset.db, AppendNoticeMessage, ¬ices);-
/* first result */
result = PQgetResult(pset.db);
@@ -1536,7 +1491,6 @@ ExecQueryAndProcessResults(const char *query, double *elapsed_msec, bool *svpt_g
*/
const char *error = PQresultErrorMessage(result);
- ShowNoticeMessage(¬ices);
if (strlen(error))
pg_log_info("%s", error);
@@ -1601,8 +1555,6 @@ ExecQueryAndProcessResults(const char *query, double *elapsed_msec, bool *svpt_g
if (result_status == PGRES_COPY_IN ||
result_status == PGRES_COPY_OUT)
{
- ShowNoticeMessage(¬ices);-
if (is_watch)
{
ClearOrSaveAllResults();
@@ -1610,12 +1562,7 @@ ExecQueryAndProcessResults(const char *query, double *elapsed_msec, bool *svpt_g
return -1;
}
- /* use normal notice processor during COPY */- PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL);-
success &= HandleCopyResult(&result);
-- PQsetNoticeProcessor(pset.db, AppendNoticeMessage, ¬ices);
}
/*
@@ -1623,9 +1570,7 @@ ExecQueryAndProcessResults(const char *query, double *elapsed_msec, bool *svpt_g
* string, it will return NULL. Otherwise, we'll have other results
* to process. We need to do that to check whether this is the last.
*/
- notices.current ^= 1;
next_result = PQgetResult(pset.db);
- notices.current ^= 1;
last = (next_result == NULL);
/*
@@ -1647,9 +1592,6 @@ ExecQueryAndProcessResults(const char *query, double *elapsed_msec, bool *svpt_g
*elapsed_msec = INSTR_TIME_GET_MILLISEC(after);
}
- /* notices already shown above for copy */- ShowNoticeMessage(¬ices);-
/* this may or may not print something depending on settings */
if (result != NULL)
success &= PrintQueryResult(result, last, false, opt, printQueryFout);
@@ -1659,7 +1601,6 @@ ExecQueryAndProcessResults(const char *query, double *elapsed_msec, bool *svpt_g
SetResultVariables(result, true);
ClearOrSaveResult(result);
- notices.current ^= 1;
result = next_result;
if (cancel_pressed)
@@ -1669,11 +1610,6 @@ ExecQueryAndProcessResults(const char *query, double *elapsed_msec, bool *svpt_g
}
}
- /* reset notice hook */- PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL);- termPQExpBuffer(¬ices.messages[0]);- termPQExpBuffer(¬ices.messages[1]);-
/* may need this to recover from conn loss during COPY */
if (!CheckConnection())
return -1;
diff --git a/src/test/regress/expected/psql.out b/src/test/regress/expected/psql.outindex 2a38a93a3b..60acbd1241 100644--- a/src/test/regress/expected/psql.out+++ b/src/test/regress/expected/psql.out@@ -5316,13 +5316,13 @@ AS $$
$$;
-- show both
SELECT 1 AS one \; SELECT warn('1.5') \; SELECT 2 AS two ;
+NOTICE: warn 1.5+CONTEXT: PL/pgSQL function warn(text) line 2 at RAISE
one
-----
1
(1 row)
-NOTICE: warn 1.5-CONTEXT: PL/pgSQL function warn(text) line 2 at RAISE
warn
------
t
@@ -5335,13 +5335,13 @@ CONTEXT: PL/pgSQL function warn(text) line 2 at RAISE
-- \gset applies to last query only
SELECT 3 AS three \; SELECT warn('3.5') \; SELECT 4 AS four \gset
+NOTICE: warn 3.5+CONTEXT: PL/pgSQL function warn(text) line 2 at RAISE
three
-------
3
(1 row)
-NOTICE: warn 3.5-CONTEXT: PL/pgSQL function warn(text) line 2 at RAISE
warn
------
t
psql-notices-3.patch (text/x-diff, 8.1K) downloaddiff --git a/src/bin/psql/common.c b/src/bin/psql/common.cindex 03e6d9ce8f..3a525839a7 100644--- a/src/bin/psql/common.c+++ b/src/bin/psql/common.c@@ -1063,26 +1063,28 @@ PrintQueryResult(PGresult *result, bool last, bool is_watch, const printQueryOpt
/*
* Data structure and functions to record notices while they are
- * emitted, so that they can be shown later.- *- * We need to know which result is last, which requires to extract- * one result in advance, hence two buffers are needed.+ * emitted, so that they can be shown later. "print" tells whether+ * to print or to store the messages.
*/
struct t_notice_messages
{
- PQExpBufferData messages[2];- int current;+ PQExpBufferData messages;+ volatile bool print;
};
/*
* Store notices in appropriate buffer, for later display.
+ * This function is called by a signal handler.
*/
static void
-AppendNoticeMessage(void *arg, const char *msg)+AppendOrPrintNoticeMessage(void *arg, const char *msg)
{
struct t_notice_messages *notices = arg;
- appendPQExpBufferStr(¬ices->messages[notices->current], msg);+ if (notices->print)+ pg_log_info("%s", msg);+ else+ appendPQExpBufferStr(¬ices->messages, msg);
}
/*
@@ -1091,11 +1093,18 @@ AppendNoticeMessage(void *arg, const char *msg)
static void
ShowNoticeMessage(struct t_notice_messages *notices)
{
- PQExpBufferData *current = ¬ices->messages[notices->current];+ bool print = notices->print;+ PQExpBufferData *shown = ¬ices->messages;- if (*current->data != '\0')- pg_log_info("%s", current->data);- resetPQExpBuffer(current);+ /* avoid race condition on the buffer which could lose messages */+ notices->print = true;++ if (*shown->data != '\0')+ pg_log_info("%s", shown->data);+ resetPQExpBuffer(shown);++ /* restore previous printing status */+ notices->print = print;
}
@@ -1478,6 +1487,8 @@ static int
ExecQueryAndProcessResults(const char *query, double *elapsed_msec, bool *svpt_gone_p,
bool is_watch, const printQueryOpt *opt, FILE *printQueryFout)
{
+ /* whether we need to know whether a result is the last early */+ bool need_last = pset.gset_prefix || pset.gexec_flag || pset.crosstab_flag || !pset.show_all_results;
bool timing = pset.timing;
bool success;
instr_time before,
@@ -1513,11 +1524,10 @@ ExecQueryAndProcessResults(const char *query, double *elapsed_msec, bool *svpt_g
return 0;
}
- /* intercept notices */- notices.current = 0;- initPQExpBuffer(¬ices.messages[0]);- initPQExpBuffer(¬ices.messages[1]);- PQsetNoticeProcessor(pset.db, AppendNoticeMessage, ¬ices);+ /* intercept notices, or not, depending on show mode */+ notices.print = true;+ initPQExpBuffer(¬ices.messages);+ PQsetNoticeProcessor(pset.db, AppendOrPrintNoticeMessage, ¬ices);
/* first result */
result = PQgetResult(pset.db);
@@ -1528,6 +1538,13 @@ ExecQueryAndProcessResults(const char *query, double *elapsed_msec, bool *svpt_g
PGresult *next_result;
bool last;
+ /*+ * Notices coming in from now on are attributed to the next (future) result,+ * so direct printing is suspended till this current result is displayed, but+ * only if it is really going to be displayed.+ */+ notices.print = !pset.show_all_results;+
if (!AcceptResult(result, false))
{
/*
@@ -1536,10 +1553,13 @@ ExecQueryAndProcessResults(const char *query, double *elapsed_msec, bool *svpt_g
*/
const char *error = PQresultErrorMessage(result);
- ShowNoticeMessage(¬ices);
if (strlen(error))
pg_log_info("%s", error);
+ /* then reactivate notice flow till next result */+ notices.print = true;+ ShowNoticeMessage(¬ices);+
CheckConnection();
if (!is_watch)
SetResultVariables(result, false);
@@ -1566,6 +1586,9 @@ ExecQueryAndProcessResults(const char *query, double *elapsed_msec, bool *svpt_g
else
result = PQgetResult(pset.db);
+ /* suspend notice flow again after the result, if needed */+ notices.print = !pset.show_all_results;+
/*
* Get current timing measure in case an error occurs
*/
@@ -1601,8 +1624,6 @@ ExecQueryAndProcessResults(const char *query, double *elapsed_msec, bool *svpt_g
if (result_status == PGRES_COPY_IN ||
result_status == PGRES_COPY_OUT)
{
- ShowNoticeMessage(¬ices);-
if (is_watch)
{
ClearOrSaveAllResults();
@@ -1610,22 +1631,38 @@ ExecQueryAndProcessResults(const char *query, double *elapsed_msec, bool *svpt_g
return -1;
}
- /* use normal notice processor during COPY */- PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL);+ /* back to normal notice processor during COPY */+ notices.print = true;+ ShowNoticeMessage(¬ices);
success &= HandleCopyResult(&result);
- PQsetNoticeProcessor(pset.db, AppendNoticeMessage, ¬ices);+ notices.print = !pset.show_all_results;
}
/*
* Check PQgetResult() again. In the typical case of a single-command
* string, it will return NULL. Otherwise, we'll have other results
* to process. We need to do that to check whether this is the last.
+ *+ * Note that notices are possibly suspended while waiting for the next+ * result because the previous result has not been displayed yet.+ *+ * If the next queries takes some time, the output of the current query+ * is suspended, eg "SELECT 0 \; SELECT take_long_time() \gset" will wait+ * before showing the initial 0. The notices are shown anyway, possibly out+ * of order, as a tradeoff between keeping the order vs delaying the notices.
*/
- notices.current ^= 1;- next_result = PQgetResult(pset.db);- notices.current ^= 1;+ if (need_last)+ {+ notices.print = true;+ ShowNoticeMessage(¬ices);+ next_result = PQgetResult(pset.db);+ notices.print = !pset.show_all_results;+ }+ else+ next_result = NULL;+
last = (next_result == NULL);
/*
@@ -1647,19 +1684,28 @@ ExecQueryAndProcessResults(const char *query, double *elapsed_msec, bool *svpt_g
*elapsed_msec = INSTR_TIME_GET_MILLISEC(after);
}
- /* notices already shown above for copy */- ShowNoticeMessage(¬ices);-
/* this may or may not print something depending on settings */
if (result != NULL)
+ {
success &= PrintQueryResult(result, last, false, opt, printQueryFout);
+ /* reactivate notice flow after printing the result */+ notices.print = true;+ ShowNoticeMessage(¬ices);+ }++ /* we need to know whether it was the last now anyway */+ if (!need_last)+ {+ next_result = PQgetResult(pset.db);+ last = (next_result == NULL);+ }+
/* set variables on last result if all went well */
if (!is_watch && last && success)
SetResultVariables(result, true);
ClearOrSaveResult(result);
- notices.current ^= 1;
result = next_result;
if (cancel_pressed)
@@ -1669,10 +1715,13 @@ ExecQueryAndProcessResults(const char *query, double *elapsed_msec, bool *svpt_g
}
}
- /* reset notice hook */+ /* reset standard notice hook */
PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL);
- termPQExpBuffer(¬ices.messages[0]);- termPQExpBuffer(¬ices.messages[1]);++ /* flush notices if any, before cleaning up */+ notices.print = true;+ ShowNoticeMessage(¬ices);+ termPQExpBuffer(¬ices.messages);
/* may need this to recover from conn loss during COPY */
if (!CheckConnection())
diff --git a/src/test/regress/expected/psql.out b/src/test/regress/expected/psql.outindex 2a38a93a3b..baedf04ba5 100644--- a/src/test/regress/expected/psql.out+++ b/src/test/regress/expected/psql.out@@ -5335,13 +5335,13 @@ CONTEXT: PL/pgSQL function warn(text) line 2 at RAISE
-- \gset applies to last query only
SELECT 3 AS three \; SELECT warn('3.5') \; SELECT 4 AS four \gset
+NOTICE: warn 3.5+CONTEXT: PL/pgSQL function warn(text) line 2 at RAISE
three
-------
3
(1 row)
-NOTICE: warn 3.5-CONTEXT: PL/pgSQL function warn(text) line 2 at RAISE
warn
------
t
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: coelho@cri.ensmp.fr, tgl@sss.pgh.pa.us, a.mckinley@analyticsengines.com, pgsql-bugs@lists.postgresql.org, peter.eisentraut@enterprisedb.com
Subject: Re: psql 15beta1 does not print notices on the console until transaction completes
In-Reply-To: <alpine.DEB.2.22.394.2206032001450.622878@pseudo>
* 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