Re: Random memory related errors on live postgres 14.13 instance on Ubuntu 22.04 LTS

On Wed, 30 Oct 2024 at 13:04, Ian J Cottee <ian(at)cottee(dot)org> wrote:

> Hello everyone, I’ve been using postgres for over 25 years now and never
> had any major issues which were not caused by my own stupidity. In the last
> 24 hours however I’ve had a number of issues on one client's server which I
> assume are a bug in postgres or a possible hardware issue (they are running
> on a Linode) but I need some clarification and would welcome advice on how
> to proceed. I will also forward this mail to Linode support to ask them to
> check for any memory issues they can detect.
>
>
>
> This particular Postgres is running on Ubuntu LTS 22.04 and has the
> following version information:
>
>
>
> ```
>
> PostgreSQL 14.13 (Ubuntu 14.13-0ubuntu0.22.04.1) on x86_64-pc-linux-gnu,
> compiled by gcc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0, 64-bit
>
> ```
>
>
> The quick summary is that over a 24 hour period I had the following errors
> appear in the postgres logs at different times causing the system processes
> to restart:
>
>
> - stuck spinlock detected
> - free(): corrupted unsorted chunks
> - double free or corruption (!prev)
> - corrupted size vs. prev_size
> - corrupted double-linked list
> - *** stack smashing detected ***: terminated
> - Segmentation fault
>
>
>

are you using the postgresql setup compiled from source ?
listing the output of pg_config may give the details

are there any extensions installed, can you list those extensions.

if you have access to source packages ,
Getting a stack trace of a running PostgreSQL backend on Linux/BSD -
PostgreSQL wiki
<https://wiki.postgresql.org/wiki/Getting_a_stack_trace_of_a_running_PostgreSQL_backend_on_Linux/BSD>
can you generate a stacktrace from the process that is crashing or if it
dumped a core, then backtrace from the core dump.

will it be possible to share the actual logs both postgresql and kernel
around the incident ...

Do you have access to core dumps which these crashes may have generated ? i
think ABRT / segmentation faults would generate one.

do you collect stats of system. around the time of crash do you any
abnormal usage of io or cpu or memory , along with locks held in postgresql
setup etc.

> Here’s the more detailed breakdown.
>
>
>
> On Monday evening this week, the following event occurred on the server
>
>
>
> ```
>
> 2024-10-28 18:12:47.145 GMT [575437] xxx(at)xxx PANIC: stuck spinlock
> detected at LWLockWaitListLock,
> ./build/../src/backend/storage/lmgr/lwlock.c:913
>
```
>
>
>
I think a backtrace here would help what part of call stack led to this.
this alone does not look like any bug.

> Followed by:
>
>
>
> ```
>
> 2024-10-28 18:12:47.249 GMT [1880289] LOG: terminating any other active
> server processes
>
> 2024-10-28 18:12:47.284 GMT [1880289] LOG: all server processes
> terminated; reinitializing
>
> ```
>
> And eventually
>
>
> ```
>
> 2024-10-28 18:12:48.474 GMT [575566] xxx(at)xxx FATAL: the database system
> is in recovery mode
>
> 2024-10-28 18:12:48.476 GMT [575550] LOG: database system was not properly
> shut down; automatic recovery in progress
>
> 2024-10-28 18:12:48.487 GMT [575550] LOG: redo starts at DD/405E83A8
>
> 2024-10-28 18:12:48.487 GMT [575550] LOG: invalid record length at
> DD/405EF818: wanted 24, got 0
>
> 2024-10-28 18:12:48.487 GMT [575550] LOG: redo done at DD/405EF7E0 system
> usage: CPU: user: 0.00 s, system: 0.00 s, elapsed: 0.00 s
>
> 2024-10-28 18:12:48.515 GMT [1880289] LOG: database system is ready to
> accept connections
>
> ```
>
> This wasn’t noticed by myself or any users as they tend to all be finished
> by 17:30. However later,
>
>
>

> ```
>
> 2024-10-28 20:27:15.258 GMT [611459] xxx(at)xxx LOG: unexpected EOF on
> client connection with an open transaction
>
> 2024-10-28 21:01:05.934 GMT [620373] xxx(at)xxxx LOG: unexpected EOF on
> client connection with an open transaction
>
> free(): corrupted unsorted chunks
>

it all seems like memory corruption or some leak ... valgrind ? to get more
details if leak ...

> 2024-10-28 21:15:02.203 GMT [1880289] LOG: server process (PID 623803) was
> terminated by signal 6: Aborted
>
> 2024-10-28 21:15:02.204 GMT [1880289] LOG: terminating any other active
> server processes
>
> ```
>
>
>
> This time it could not recover and I didn’t notice until early the next
> morning whilst doing some routine checks.
>
>
>
> ```
>
> 2024-10-28 21:15:03.643 GMT [623807] LOG: database system was not properly
> shut down; automatic recovery in progress
>
> 2024-10-28 21:15:03.655 GMT [623807] LOG: redo starts at DD/47366740
>
> 2024-10-28 21:15:03.663 GMT [623807] LOG: invalid record length at
> DD/475452A0: wanted 24, got 0
>
> 2024-10-28 21:15:03.663 GMT [623807] LOG: redo done at DD/47545268 system
> usage: CPU: user: 0.00 s, system: 0.00 s, elapsed: 0.00 s
>
> 2024-10-28 21:15:03.682 GMT [623829] xxx(at)xxx FATAL: the database system
> is in recovery mode
>
> double free or corruption (!prev)
>
> 2024-10-28 21:15:03.832 GMT [1880289] LOG: startup process (PID 623807)
> was terminated by signal 6: Aborted
>
> 2024-10-28 21:15:03.832 GMT [1880289] LOG: aborting startup due to startup
> process failure
>
> 2024-10-28 21:15:03.835 GMT [1880289] LOG: database system is shut down
>
> ```
>
>
>
> When I noticed in the morning it was able to start without an issue. From
> googling it appeared to be a memory issue and I wondered if the problem was
> sorted now the server process had stopped completely and restarted. The
> problem was not sorted although all the above errors were recovered from
> automatically without any input from myself or the client’s noticing.
>
>
>
> ```
>
> corrupted size vs. prev_size
>
> 2024-10-29 09:55:24.417 GMT [894747] LOG: background worker "parallel
> worker" (PID 947642) was terminated by signal 6: Aborted
>
> ```
>
>
>
> ```
>
> corrupted double-linked list
>
> 2024-10-29 13:14:28.322 GMT [894747] LOG: background worker "parallel
> worker" (PID 1019071) was terminated by signal 6: Aborted
>
> ```
>
>
>
> ```
>
> *** stack smashing detected ***: terminated
>
> 2024-10-28 15:24:30.331 GMT [1880289] LOG: background worker "parallel
> worker" (PID 528630) was terminated by signal 6: A\ borted
>
> ```
>
>
>
> ```
>
> 2024-10-28 15:40:26.617 GMT [1880289] LOG: background worker "parallel
> worker" (PID 533515) was terminated by signal 11: \
>
> Segmentation fault
>
> 2024-10-28 15:40:26.617 GMT [1880289] DETAIL: Failed process was running:
> SELECT "formula_line".id FROM "formul\
>
> ```
>
>

idk why it crashed with sigabrt instead of sigkill if it was indeed a
memory leak and not a bug ... so not sure memory overcommiting can be of
use here ...

how much is the concurrency at peak and with what work mem .... any theory
of excessive work mem and too many concurrent processes holding some locks
for long ...
it should not crash even if that is the case, but just asking ...

lastly.... is it possible to memcheck run on the machine just to ensure no
memory scares ... if this is running on a vm, or bare metal ,,,, any
hardware errors around that time ?

most likely it looks like a h/w issue, we used to see things like this on
bare metals .... which only happened occasionally and then frequently till
we moved away from that setup.

also, does it happen only when the optimiser picks a plan involving
parallel workers for a query?
If you set max_parallel_workers_per_gather to 0, to not
parallelize anything , do you still see the issue ?

Just insights, if not useful, pls ignore.

> Best regards
>
>
> Ian Cottee
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>

--
Thanks,
Vijay

Open to work
Resume - Vijaykumar Jain <https://github.com/cabecada>