Thread: Valgrind errors on 32-bit OS

Valgrind errors on 32-bit OS

From
o.tselebrovskiy@postgrespro.ru
Date:
Greetings, everyone!

While working on an extension, I've found myself using valgrind on a 
32-bit OS (Debian 11)
and after executing any query (even 'select 1;') under valgrind I've 
been shown the same error
everytime:

==00:00:00:18.109 2528== VALGRINDERROR-BEGIN
==00:00:00:18.109 2528== Use of uninitialised value of size 4
==00:00:00:18.109 2528==    at 0x645B52: pg_comp_crc32c_sb8 
(pg_crc32c_sb8.c:79)
==00:00:00:18.109 2528==    by 0x24295F: XLogRecordAssemble 
(xloginsert.c:780)
==00:00:00:18.109 2528==    by 0x24295F: XLogInsert (xloginsert.c:459)
==00:00:00:18.109 2528==    by 0x4B693B: LogCurrentRunningXacts 
(standby.c:1099)
==00:00:00:18.109 2528==    by 0x4B693B: LogStandbySnapshot 
(standby.c:1055)
==00:00:00:18.109 2528==    by 0x43BECF: BackgroundWriterMain 
(bgwriter.c:336)
==00:00:00:18.109 2528==    by 0x249D8A: AuxiliaryProcessMain 
(bootstrap.c:446)
==00:00:00:18.109 2528==    by 0x448B11: StartChildProcess 
(postmaster.c:5445)
==00:00:00:18.109 2528==    by 0x449F57: reaper (postmaster.c:2907)
==00:00:00:18.109 2528==    by 0x486D587: ??? (in 
/usr/lib/i386-linux-gnu/libpthread-2.31.so)
==00:00:00:18.109 2528==  Uninitialised value was created by a stack 
allocation
==00:00:00:18.109 2528==    at 0x4B682F: LogStandbySnapshot 
(standby.c:1018)

I've been able to reproduce this on branches from REL_11_STABLE up to 
current master.
Valgrind version is 3.21.0.

I was wondering, why does this error occur only on 32-bit OS and not 
64-bit?

I found three threads:
About issues with valgrind and padding -
https://www.postgresql.org/message-id/flat/1082573393.7010.27.camel%40tokyo
About sanitizers (and valgrind) in general -
https://www.postgresql.org/message-id/flat/20160321130850.6ed6f598%40fujitsu
About valgrind suppressions -
https://www.postgresql.org/message-id/flat/4dfabec2-a3ad-0546-2d62-f816c97edd0c%402ndQuadrant.com
and after reading those decided to look for existing valgrind 
suppressions.

And I've found just what I was looking for:

{
    padding_XLogRecData_CRC
    Memcheck:Value8

    fun:pg_comp_crc32c*
    fun:XLogRecordAssemble
}

Supression for this (group of) function(s), but for 8-byte chunks of 
memory.

So the suggestion is to add a suppression for 4-byte values in this 
function.

The patch is attached.

Oleg Tselebrovskiy, Postgres Pro
Attachment