Обсуждение: any suggestions to detect memory corruption

Alex <zhihui.fan1213@gmail.com> writes:
> I can get the following log randomly and I am not which commit caused it.

> 2019-05-08 21:37:46.692 CST [60110] WARNING:  problem in alloc set index
> info: req size > alloc size for chunk 0x2a33a78 in block 0x2a33a18

I've had success in finding memory stomp causes fairly quickly by setting
a hardware watchpoint in gdb on the affected location.  Then you just let
it run to see when the value changes, and check whether that's a "legit"
or "not legit" modification point.

The hard part of that, of course, is to know in advance where the affected
location is.  You may be able to make things sufficiently repeatable by
doing the problem query in a fresh session each time.

            regards, tom lane

On Wed, May 8, 2019 at 10:34 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Alex <zhihui.fan1213@gmail.com> writes:
> > I can get the following log randomly and I am not which commit caused it.
>
> > 2019-05-08 21:37:46.692 CST [60110] WARNING:  problem in alloc set index
> > info: req size > alloc size for chunk 0x2a33a78 in block 0x2a33a18
>
> I've had success in finding memory stomp causes fairly quickly by setting
> a hardware watchpoint in gdb on the affected location.  Then you just let
> it run to see when the value changes, and check whether that's a "legit"
> or "not legit" modification point.
>
> The hard part of that, of course, is to know in advance where the affected
> location is.  You may be able to make things sufficiently repeatable by
> doing the problem query in a fresh session each time.

valgrind might also be a possibility, although that has a lot of overhead.

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

Alex <zhihui.fan1213@gmail.com> writes:
> Someone add some code during backend init which used palloc. but at that
> time,  the CurrentMemoryContext is PostmasterContext.   at the end of
> backend initialization, the PostmasterContext is deleted, then the error
> happens.  the reason why it happens randomly is before the palloc, there
> are some other if clause which may skip the palloc.

> I still can't explain why PostmasterContext may have impact "index info"
> MemoryContext sometime,  but now I just can't reproduce it (before the
> fix,  it may happen in 30% cases).

Well, once the context is deleted, that memory is available for reuse.
Everything will seem fine until it *is* reused, and then boom!

The error would have been a lot more obvious if you'd enabled
MEMORY_CONTEXT_CHECKING, which would overwrite freed data with garbage.
That is normally turned on in --enable-cassert builds.  Anybody who's been
hacking Postgres for more than a week does backend code development in
--enable-cassert mode as a matter of course; it turns on a *lot* of
helpful cross-checks.

            regards, tom lane