Horton

Thread

Commits

  1. Avoid concurrent calls to bindtextdomain().

  2. Clean up Windows-specific mutex code in libpq and ecpglib.

  1. Avoiding concurrent calls to bindtextdomain()

    Tom Lane <tgl@sss.pgh.pa.us> — 2024-02-05T20:03:36Z

    According to the discussion in [1], it's not as safe as we supposed
to allow different threads to call bindtextdomain() concurrently.
Here is a patchset to prevent that by acquiring a mutex around
the libpq and ecpglib calls that are at risk.

In libpq, this would've required yet a third copy of the
Windows-specific ugliness in default_threadlock() and pgtls_init().
I didn't particularly want to do that, so I stole some ideas
from ecpglib to create a more POSIX-compliant emulation of
pthread_mutex_lock().  0001 attached is the refactoring needed
to make that happen, and then 0002 is the actual bug fix.

0001 also gets rid of the possibility that pthread_mutex_init/
pthread_mutex_lock could fail due to malloc failure.  This seems
important since default_threadlock() assumes that pthread_mutex_lock
cannot fail in practice.  I observe that ecpglib also assumes that,
although it's using CreateMutex() which has documented failure
conditions.  So I wonder if we ought to copy this implementation
back into ecpglib; but I've not done that here.

			regards, tom lane

[1] https://www.postgresql.org/message-id/flat/18312-bbbabc8113592b78%40postgresql.org

  2. Re: Avoiding concurrent calls to bindtextdomain()

    Tom Lane <tgl@sss.pgh.pa.us> — 2024-02-08T18:08:21Z

    I wrote:
> 0001 also gets rid of the possibility that pthread_mutex_init/
> pthread_mutex_lock could fail due to malloc failure.  This seems
> important since default_threadlock() assumes that pthread_mutex_lock
> cannot fail in practice.  I observe that ecpglib also assumes that,
> although it's using CreateMutex() which has documented failure
> conditions.  So I wonder if we ought to copy this implementation
> back into ecpglib; but I've not done that here.

The cfbot seemed happy with v1, so here's a v2 that does copy that
code back into ecpglib.  (I kind of wonder why this code exists in
libpq + ecpglib at all, rather than in src/port/; but that seems
like a refactoring exercise for another day.)

I double-checked that all the pthread_mutex_t variables in libpq
and ecpglib are static, so the change in that struct should not
pose an ABI hazard for back-patching.

Barring objections, I plan to push this pretty soon.

			regards, tom lane