Re: observing process blocked on find_or_construct

From: Andrey Semashev (andrey.semashev_at_[hidden])
Date: 2025-01-28 12:56:16

On 1/28/25 15:16, Murali Kishore via Boost wrote:
> Thanks Peter Dimov, i have added logic unlock in signal handler, now i am
> not seeing this issue.
> 
> I am seeing one more issue, if i call construct object and do work and
> clear in loop of ~30000, i see following error while construct call.

It's not enough to just unlock the mutex (or recover it if it was
abandoned). You also need to restore the state it was protecting to a
consistent state. Which is often unrealistic to do since you don't know
which part of the state is corrupted and how to restore it to a
consistent state. For example, you don't have the means to repair the
segment manager, if its internal object tree is left corrupted, and you
don't know whether any of the objects stored in it are half-constructed
or otherwise inconsistent.

Typically, your best course of action when you detect an abandoned mutex
is to scrap the data it protects and start from scratch. And also try
hard to not abandon mutexes as much as possible, e.g. don't just kill
the process on a signal and let it finish its work on the shared memory
first.