Re: Dereference before NULL check (src/backend/storage/ipc/latch.c)

At Mon, 02 Nov 2020 17:25:04 +0900 (JST), Kyotaro Horiguchi <horikyota(dot)ntt(at)gmail(dot)com> wrote in
> I called SetLatch() during WaitLatch(NULL, ) but that doesn't fire
> WL_LATCH_SET event for me on Windows. (I got it fired on Linux..) On
> Windows, the latch is detected after exiting the WaitLatch()
> call. Seems like MyLatch of waiter is different from
> peerPGPROC->procLatch. And... an update for Visual Studio broke my
> environment... I will investigate this further but everything feel
> cumbersome on Windows...

I managed to reproduce the issue. FWIW the attached modifies
pg_backend_pid() to call "WaitLatch(NULL," and
pg_terminate_backend(pid) to SetLatch() to the process latch of the
pid. (It's minunderstanding that I could reproduce this on Linux.)

Session A:
=# select pg_backend_pid(); -- sleeps for 10 seconds.

Session B:
=# select pg_terminate_backend(A-pid);

[11628] LOG: server process (PID 14568) was terminated by exception 0xC0000005
[11628] DETAIL: Failed process was running: select pg_backend_pid();
[11628] HINT: See C include file "ntstatus.h" for a description of the hexadecimal value.
[11628] LOG: terminating any other active server processes
[2948] WARNING: terminating connection because of crash of another server process
2

With the fix patch, it changes to:

[16632] LOG: FALSE LATCH: 0000000000000000

rebards.

--
Kyotaro Horiguchi
NTT Open Source Software Center