Re: [PATCH 2/2] mm: add PageWaiters indicating tasks are waiting for a page bit
From: Nicholas Piggin
Date: Sun Dec 25 2016 - 20:17:40 EST
On Sun, 25 Dec 2016 13:51:17 -0800
Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> On Sat, Dec 24, 2016 at 7:00 PM, Nicholas Piggin <npiggin@xxxxxxxxx> wrote:
> > Add a new page flag, PageWaiters, to indicate the page waitqueue has
> > tasks waiting. This can be tested rather than testing waitqueue_active
> > which requires another cacheline load.
> Ok, I applied this one too. I think there's room for improvement, but
> I don't think it's going to help to just wait another release cycle
> and hope something happens.
> Example room for improvement from a profile of unlock_page():
> 46.44 â lock andb $0xfe,(%rdi)
> 34.22 â mov (%rdi),%rax
> this has the old "do atomic op on a byte, then load the whole word"
> issue that we used to have with the nasty zone lookup code too. And it
> causes a horrible pipeline hickup because the load will not forward
> the data from the (partial) store.
> Its' really a misfeature of our asm optimizations of the atomic bit
> ops. Using "andb" is slightly smaller, but in this case in particular,
> an "andq" would be a ton faster, and the mask still fits in an imm8,
> so it's not even hugely larger.
I did actually play around with that. I could not get my skylake
to forward the result from a lock op to a subsequent load (the
latency was the same whether you use lock ; andb or lock ; andl
(32 cycles for my test loop) whereas with non-atomic versions I
was getting about 15 cycles for andb vs 2 for andl.
I guess the lock op drains the store queue to coherency and does
not allow forwarding so as to provide the memory ordering
> But it might also be a good idea to simply use a "cmpxchg" loop here.
> That also gives atomicity guarantees that we don't have with the
> "clear bit and then load the value".
cmpxchg ends up at 19 cycles including the initial load, so it
may be worthwhile. Powerpc has a similar problem with doing a
clear_bit; test_bit (not the size mismatch, but forwarding from
atomic ops being less capable).