Re: [PATCH] elf: Simplify software TM implementation in _dl_find_object

[Florian Weimer <fweimer@redhat.com>]

* Szabolcs Nagy:

> The 01/07/2022 16:56, Florian Weimer wrote:
>> * Szabolcs Nagy:
>> 
>> > The 01/07/2022 14:41, Florian Weimer wrote:
>> >> With the current set of fences, the version update at the start
>> >> of the TM write operation is redundant.  Also use relaxed MO stores
>> >> during the dlclose update, and skip any version changes there.
>> >> 
>> >> Suggested-by: Szabolcs Nagy <szabolcs.nagy@arm.com>
>> >> 
>> >> ---
>> >> Initial testing looks good, but I'll keep the aarch64 and powerpc64le
>> >> tests running in a loop for a while.
>> >
>> > looks good to me except
>> >
>> >> @@ -282,15 +279,6 @@ _dlfo_mappings_end_update (void)
>> >>    atomic_thread_fence_release ();
>> >>    __atomic_wide_counter_fetch_add_relaxed (&_dlfo_loaded_mappings_version, 1);
>> >>  }
>> >
>> > i think the v += 1 does not have to be atomic so
>> >
>> >  start_version = load_relaxed (&ver);
>> >  store_relaxed (&ver, start_version + 1);
>> >
>> > or if we can change the api to pass start_version
>> > that may be clearer.
>> 
>> The 32-bit wide counter needs to be atomic.  We can't use a straight
>> store in that case.
>> 
>> Should I break abstraction and write through for the 64-bit case?
>
> i didn't think about that.
>
> but it seems there is a
>
> __atomic_wide_counter_add_relaxed
>
> which seems to do the right thing.

Good point.  I'll restart my tests with that.

Thanks,
Flroian