On 03/22/2019 03:30 PM, Davidlohr Bueso wrote:
On Fri, 22 Mar 2019, Linus Torvalds wrote:
Some of them _might_ be performance-critical. There's the one on
mmap_sem in the fault handling path, for example. And yes, I'd expect
the normal case to very much be "no other readers or writers" for that
one.
Yeah, the mmap_sem case in the fault path is really expecting an unlocked
state. To the point that four archs have added branch predictions, ie:
92181f190b6 (x86: optimise x86's do_page_fault (C entry point for the
page fault path))
b15021d994f (powerpc/mm: Add a bunch of (un)likely annotations to
do_page_fault)
And using PROFILE_ANNOTATED_BRANCHES shows pretty clearly:
(without resetting the counters)
correct incorrect % Function File Line
------- --------- - -------- ---- ----
4603685 34 0 do_user_addr_fault fault.c 1416
(bootup)
382327745 449 0 do_user_addr_fault fault.c
1416 (kernel build)
399446159 461 0 do_user_addr_fault fault.c
1416 (redis benchmark)
It would probably wouldn't harm doing the unlikely() for all archs, or
alternatively, add likely() to the atomic_long_try_cmpxchg_acquire in
patch 3 and do it implicitly but maybe that would be less flexible(?)
Thanks,
Davidlohr
I had used the my lock event counting code to count the number of
contended and uncontended trylocks. I tested both bootup and kernel
build. I think I saw less than 1% were contended, the rests were all
uncontended. That is similar to what you got. I thought I had sent the
data out previously, but I couldn't find the email. That was the main
reason why I took Linus' suggestion to optimize it for the uncontended case.
Thanks,
Longman
