On 05/17/2014 12:28 AM, Jeff Janes wrote:
> More fun with my torn page injection test program on 9.4.
>
> 24171 2014-05-16 14:00:44.934 PDT:WARNING: 01000: page verification
> failed, calculated checksum 21100 but expected 3356
> 24171 2014-05-16 14:00:44.934 PDT:CONTEXT: xlog redo split_l: rel
> 1663/16384/16405 left 35191, right 35652, next 34666, level 0, firstright
> 192
> 24171 2014-05-16 14:00:44.934 PDT:LOCATION: PageIsVerified, bufpage.c:145
> 24171 2014-05-16 14:00:44.934 PDT:FATAL: XX001: invalid page in block
> 34666 of relation base/16384/16405
> 24171 2014-05-16 14:00:44.934 PDT:CONTEXT: xlog redo split_l: rel
> 1663/16384/16405 left 35191, right 35652, next 34666, level 0, firstright
> 192
> 24171 2014-05-16 14:00:44.934 PDT:LOCATION: ReadBuffer_common,
> bufmgr.c:483
>
>
> I've seen this twice now, the checksum failure was both times for the block
> labelled "next" in the redo record. Is this another case where the block
> needs to be reinitialized upon replay?
Hmm, it looks like I fumbled the numbering of the backup blocks in the
b-tree split WAL record (in 9.4). I blame the comments; the comments
where the record is generated numbers the backup blocks starting from 1,
but XLR_BKP_BLOCK(x) and RestoreBackupBlock(...) used in replay number
them starting from 0.
Attached is a patch that I think fixes them. In addition to the
rnext-reference, clearing the incomplete-split flag in the child page,
had a similar numbering mishap.
- Heikki
