On Wed, Apr 5, 2023 at 6:37 PM Dave Hansen [off-list ref] wrote:

On 4/5/23 07:17, Uros Bizjak wrote:

quoted

Add generic and target specific support for local{,64}_try_cmpxchg
and wire up support for all targets that use local_t infrastructure.

I feel like I'm missing some context.

What are the actual end user visible effects of this series?  Is there a
measurable decrease in perf overhead?  Why go to all this trouble for
perf?  Who else will use local_try_cmpxchg()?

This functionality was requested by perf people [1], so perhaps Steven
can give us some concrete examples. In general, apart from the removal
of unneeded compare instruction on x86, usage of try_cmpxchg also
results in slightly better code on non-x86 targets [2], since the code
now correctly identifies fast-path through the cmpxchg loop.

Also important is that try_cmpxchg code reuses the result of cmpxchg
instruction in the loop, so a read from the memory in the loop is
eliminated. When reviewing the cmpxchg usage sites, I found numerous
places where unnecessary read from memory was present in the loop, two
examples can be seen in the last patch of this series.

Also, using try_cmpxchg prevents inconsistencies of the cmpxchg loop,
where the result of the cmpxchg is compared with the wrong "old" value
- one such bug is still lurking in x86 APIC code, please see [3].

Please note that apart from perf subsystem, event subsystem can also
be improved by using local_try_cmpxchg. This is the reason that the
last patch includes a change in events/core.c.

I'm all for improving things, and perf is an important user.  But, if
the goal here is improving performance, it would be nice to see at least
a stab at quantifying the performance delta.

[1] https://lore.kernel.org/lkml/20230301131831.6c8d4ff5@gandalf.local.home/ (local)
[2] https://lore.kernel.org/lkml/Yo91omfDZtTgXhyn@FVFF77S0Q05N.cambridge.arm.com/ (local)
[3] https://lore.kernel.org/lkml/20230227160917.107820-1-ubizjak@gmail.com/ (local)

Uros.