Corrupt WAL production possible in gistxlog.c

Yoichi Hirai <yh@lyon.is.s.u-tokyo.ac.jp>

From: Yoichi Hirai <yh@lyon.is.s.u-tokyo.ac.jp>
To: pgsql-hackers@postgresql.org
Date: 2009-12-24T04:29:22Z
Lists: pgsql-hackers

Attachments

Hello,

I was reading GiST core codes when I found an XLogInsert()
call that can produce a corrupt WAL record.

== Summary ==
There is an execution path that produces a WAL record whose
xl_info indicates XLOG_GIST_PAGE_UPDATE while the record
actually contains a gistxlogPageSplit structure.

== Details ==
(Line numbers are for HEAD as of Wed Dec 23 19:42:15 2009 +0000.)

The problematic XLogInsert() call is on gistxlog.c, line 770:
  recptr = XLogInsert(RM_GIST_ID, XLOG_GIST_PAGE_UPDATE, rdata);
where the last argument rdata has a pointer assigned either
on line 741 or on line 752.

When rdata comes from formSplitRdata() on line 741,
rdata contains a reference to a gistxlogPageSplit structure.
This is inconsistent with the second argument XLOG_GIST_PAGE_UPDATE.

== Importance ==
I think this poses possible data loss under multiple consecutive crashes.

== Fix ==
I attach a simple patch (for HEAD as of the datetime above)
that, I suppose, prevents the corrupt WAL production.
I would be glad if you liked it.

Please note that the execution path exists at least in current HEAD,
REL8_2_STABLE and the branches in between.

Sincerely,

Yoichi Hirai
  yh@lyon.is.s.u-tokyo.ac.jp
  Dept. of Computer Science, The University of Tokyo