On 05/06/2014 07:36 PM, Andres Freund wrote:
On 2014-05-06 13:33:01 +0300, Heikki Linnakangas wrote:
On 03/31/2014 09:08 PM, Robert Haas wrote:
On Wed, Mar 26, 2014 at 9:45 PM, Peter Geoghegan <p...@heroku.com> wrote:
On Wed, Nov 27, 2013 at 9:10 AM, Noah Misch <n...@leadboat.com> wrote:
The threat is that rounding the read size up to the next MAXALIGN would cross
into an unreadable memory page, resulting in a SIGSEGV.  Every palloc chunk
has MAXALIGN'd size under the hood, so the excess read of "toDelete" cannot
cause a SIGSEGV.  For a stack variable, it depends on the ABI.  I'm not aware
of an ABI where the four bytes past the end of this stack variable could be
unreadable, which is not to claim I'm well-read on the topic.  We should fix
this in due course on code hygiene grounds, but I would not back-patch it.


Attached patch silences the "Invalid read of size n" complaints of
Valgrind. I agree with your general thoughts around backpatching. Note
that the patch addresses a distinct complaint from Kevin's, as
Valgrind doesn't take issue with the invalid reads past the end of
spgxlogPickSplit variables on the stack.

Is the needless zeroing this patch introduces apt to cause a
performance problem?

This function is actually pretty wacky.  If we're stuffing bytes with
undefined contents into the WAL record, maybe the answer isn't to
force the contents of those bytes to be defined, but rather to elide
them from the WAL record.

Agreed. I propose the attached, which removes the MAXALIGNs. It's not
suitable for backpatching, though, as it changes the format of the WAL
record.
y
Not knowing anything about spgist this looks sane to me.

Since we did a catversion bump anyway, I revisited this. Committed an expanded version of the patch I posted earlier. I went through all the SP-GiST WAL record types and removed the alignment requirements of tuples from all of them. Most of them didn't have a problem because the structs happened to have suitable alignment by accident, but seems best to not rely on that, for the sake of consistency and robustness if the structs are modified later.

I ran this through my testing tool that compares page image on master and standby after every WAL record replay, with full_page_writes on and off, and found no errors.

Do we need a
backpatchable solution given that we seem to agree that these bugs
aren't likely to cause harm?

I don't think we do.

- Heikki


--
Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-hackers

Reply via email to