Re: PANIC in GIN code - Mailing list pgsql-hackers

On 06/26/2015 08:02 PM, Jeff Janes wrote:

Under high load against f7bb7f0625771bc71869cda, I occasionally get:

PANIC:  XLogBeginInsert was not called

It seems to only come from vacuuming.

Here is an example back-trace:

#0  0x0000003dcb632625 in raise () from /lib64/libc.so.6
#1  0x0000003dcb633e05 in abort () from /lib64/libc.so.6
#2  0x000000000079a39c in errfinish (dummy=<value optimized out>) at
elog.c:551
#3  0x000000000079b0e4 in elog_finish (elevel=<value optimized out>,
fmt=<value optimized out>) at elog.c:1368
#4  0x00000000004dd1fe in XLogInsert (rmid=13 '\r', info=48 '0') at
xloginsert.c:412
#5  0x0000000000478e13 in ginPlaceToPage (btree=0x7fff1fbb7f60,
stack=0x2878760, insertdata=<value optimized out>, updateblkno=<value
optimized out>,
     childbuf=0, buildStats=<value optimized out>) at ginbtree.c:582
#6  0x0000000000479a33 in ginInsertValue (btree=0x7fff1fbb7f60,
stack=0x2878760, insertdata=0x7fff1fbb7fe0, buildStats=0x0) at
ginbtree.c:751
#7  0x000000000047364a in ginEntryInsert (ginstate=0x7fff1fbb8280,
attnum=54624, key=1, category=<value optimized out>, items=0x287d3c0,
nitem=1,
     buildStats=0x0) at gininsert.c:234
#8  0x000000000047ef4b in ginInsertCleanup (ginstate=0x7fff1fbb8280,
vac_delay=<value optimized out>, stats=0x2868d90) at ginfast.c:843
#9  0x000000000047e024 in ginbulkdelete (fcinfo=<value optimized out>) at
ginvacuum.c:547


 From the code, it seems obvious that XLogBeginInsert is getting called at
ginbtree.c line 373, so I think that some obscure code path of
btree->placeToPage must be either consuming or resetting that
XLogBeginInsert before it returns control to ginbtree.c

Here's a theory:

First of all, you have checksums or wal_log_hints enabled.

The page is being split (that's evident from "info=48" above). ginPlaceToPage calls GinNewBuffer, which calls GetFreeIndexPage(). That finds a page that can be recycled, and marks it as used. RecordUsedIndexPage calls MarkBufferDirtyHint(), which in turn calls XLogSaveBufferForHint() to create a full-page image record of the page. That calls XLogBeginInsert() + XLogInsert(), and leaves the begininsert_called == false.

If you had assertions enabled, you'd see the assertion in XLogBeginInsert() to fail.

I'll look into that over the weekend..