Re: Issue in GIN fast-insert: XLogBeginInsert + Read/LockBuffer ordering - Mailing list pgsql-hackers

From Michael Paquier
Subject Re: Issue in GIN fast-insert: XLogBeginInsert + Read/LockBuffer ordering
Date
Msg-id Y0Yh5Xarpb894cRH@paquier.xyz
Whole thread Raw
In response to Re: Issue in GIN fast-insert: XLogBeginInsert + Read/LockBuffer ordering  (Zhang Mingli <zmlpostgres@gmail.com>)
Responses Re: Issue in GIN fast-insert: XLogBeginInsert + Read/LockBuffer ordering  (Zhang Mingli <zmlpostgres@gmail.com>)
List pgsql-hackers
On Fri, Sep 30, 2022 at 06:22:02PM +0800, Zhang Mingli wrote:
> In the same function, there is disorder of XLogBeginInsert and START_CRIT_SECTION.
>
> ```
> collectordata = ptr = (char *) palloc(collector->sumsize);
>
>  data.ntuples = collector->ntuples;
>
>  if (needWal)
>    XLogBeginInsert();
>
>  START_CRIT_SECTION();
> ```
>
> Shall we adjust that too?

Nice catches, both of you.  Let's adjust everything spotted in one
shot.  Matthias' patch makes the ordering right, but the
initialization path a bit harder to follow when using a separate list.
The code is short so it does not strike me as a big deal, and it comes
from the fact that we need to lock an existing buffer when merging two
lists.  For the branch where we insert into a tail page, the pages are
already locked but it looks to be enough to move XLogBeginInsert()
before the first XLogRegisterBuffer() call.

Would any of you like to write a patch?
--
Michael

Attachment

pgsql-hackers by date:

Previous
From: David Rowley
Date:
Subject: Re: shadow variables - pg15 edition
Next
From: "wangw.fnst@fujitsu.com"
Date:
Subject: RE: Perform streaming logical transactions by background workers and parallel apply