Re: [PATCH 6/6] x86/ticketlock: make __ticket_spin_trylock common

[Jeremy Fitzhardinge <jeremy@goop.org>]

On 01/24/2011 05:58 PM, Nick Piggin wrote:
>>
>> The C version can't take advantage of the fact that the cmpxchg directly
>> sets the flags, so it ends up re-comparing the old and swapped-out
>> values to set the return.  And it doesn't re-use the same sete to set
>> the return value in the quick failed-to-acquire path.
> Hm.

Adding a "cmpxchg_flag" which does its own sete and returns a boolean
"success" flag, the whole thing goes to:

__ticket_spin_trylock:
        movw    (%rdi), %ax     # lock_2(D)->D.5950.tickets, old
        xorl    %edx, %edx      # D.13949
        movzbl  %ah, %ecx       # old, tmp69
        cmpb    %al, %cl        # old, tmp69
        jne     .L5     #,
        leal    256(%rax), %edx #, D.13951
        lock; cmpxchgw %dx,(%rdi); sete %al     # D.13951,* lock, __ret
        movzbl  %al, %edx       # __ret, D.13949
.L5:
        movl    %edx, %eax      # D.13949,
        ret




The eax/edx shuffle is a bit unfortunate I can't see it hurting very much.


>> It might be worth having a generic cmpxchg() variant which returns a
>> succeed/fail flag rather than the fetched value, to avoid comparison in
>> this case - since many (most?) cmpxchg() callers end up doing that
>> comparison.
>>
>> How performance critical is trylock?  I guess the ones in fs/dcache.c
>> are the ones looming large in your mind.
> Well they are on on the reclaim/free path rather than the _hottest_
> paths, but yes they are performance critical.

I think that code looks pretty reasonable.

    J