On the x86-64 architecture even a failing cmpxchg grants exclusive
access to the cacheline, making it preferable to retry the failed op
immediately instead of stalling with the pause instruction.
To illustrate the impact, below are benchmark results obtained by
running various will-it-scale tests on top of the 6.2-rc3 kernel and
Cascade Lake (2 sockets * 24 cores * 2 threads) CPU.
All results in ops/s. Note there is some variance in re-runs, but the
code is consistently faster when contention is present.
open3 ("Same file open/close"):
proc stock no-pause
1 805603 814942 (+%1)
2 1054980 1054781 (-0%)
8 1544802 1822858 (+18%)
24 1191064 2199665 (+84%)
48 851582 1469860 (+72%)
96 609481 1427170 (+134%)
fstat2 ("Same file fstat"):
proc stock no-pause
1 3013872 3047636 (+1%)
2 4284687 4400421 (+2%)
8 3257721 5530156 (+69%)
24 2239819 5466127 (+144%)
48 1701072 5256609 (+209%)
96 1269157 6649326 (+423%)
Additionally, a kernel with a private patch to help access() scalability:
access2 ("Same file access"):
proc stock patched patched
+nopause
24 2378041 2005501 5370335 (-15% / +125%)
That is, fixing the problems in access itself *reduces* scalability
after the cacheline ping-pong only happens in lockref with the pause
instruction.
Note that fstat and access benchmarks are not currently integrated into
will-it-scale, but interested parties can find them in pull requests to
said project.
Code at hand has a rather tortured history. First modification showed
up in commit d472d9d98b ("lockref: Relax in cmpxchg loop"), written
with Itanium in mind. Later it got patched up to use an arch-dependent
macro to stop doing it on s390 where it caused a significant regression.
Said macro had undergone revisions and was ultimately eliminated later,
going back to cpu_relax.
While I intended to only remove cpu_relax for x86-64, I got the
following comment from Linus:
I would actually prefer just removing it entirely and see if
somebody else hollers. You have the numbers to prove it hurts on
real hardware, and I don't think we have any numbers to the
contrary.
So I think it's better to trust the numbers and remove it as a
failure, than say "let's just remove it on x86-64 and leave
everybody else with the potentially broken code"
Additionally, Will Deacon (maintainer of the arm64 port, one of the
architectures previously benchmarked):
So, from the arm64 side of the fence, I'm perfectly happy just
removing the cpu_relax() calls from lockref.
As such, come back full circle in history and whack it altogether.
Signed-off-by: Mateusz Guzik <mjguzik@gmail.com>
Link: https://lore.kernel.org/all/CAGudoHHx0Nqg6DE70zAVA75eV-HXfWyhVMWZ-aSeOofkA_=WdA@mail.gmail.com/
Acked-by: Tony Luck <tony.luck@intel.com> # ia64
Acked-by: Nicholas Piggin <npiggin@gmail.com> # powerpc
Acked-by: Will Deacon <will@kernel.org> # arm64
Acked-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
f5fe24ef17