Re: [committed][gdb/symtab] Fix data race in ~charset_vector

[Tom de Vries <tdevries@suse.de>]

On 7/14/22 19:31, Tom Tromey wrote:
>>>>>> "Tom" == Tom de Vries via Gdb-patches <gdb-patches@sourceware.org> writes:
> 
> Tom> The problem is that we're freeing the charset_vector elements in the destructor,
> Tom> which may still be used by a worker thread.
> 
> Could you say how this happens?
> What is the worker thread doing?
> 

As I understand it, the worker thread is executing do_finalize which 
uses a string (host_charset_name or some such) contained in the 
charset_vector, which is being destroyed in the main thread.

But that's a guess, thread sanitizer didn't report what the worker 
thread was doing: "[failed to restore the stack]".

> Normally I think gdb should be waiting for background tasks to finish
> before proceeding with this sort of destruction.  If this is coming from
> the parallel DWARF reader, maybe we need some higher level API to
> cleanly shut down those threads first.

Well, the worker threads are detached, so they run until process exit. 
We can not detach them, and do a join, but that postpones process exit.

We could try to do pthread_cancel, but still there might be some waiting 
involved.

This was sofar the only case I found of this type of race, so I thought 
it reasonable to fix like this.  If there are more cases, that might not 
be practical.

Thanks,
- Tom