[lkml]   [2015]   [Oct]   [5]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [REGRESSION] 998ef75ddb and aio-dio-invalidate-failure w/ data=journal
On 10/05/2015 01:22 PM, Linus Torvalds wrote:
> On Mon, Oct 5, 2015 at 5:23 PM, Dave Hansen <> wrote:
>> One thing I've been noticing on Skylake is that barriers (implicit and
>> explicit) are showing up more in profiles.
> Ahh, you're on skylake?


> It's entirely possible that the issue is that the whole
> "stac/mov/clac" is much more expensive because skylake actually ends
> up supporting those AC instructions. That would make sense.
> We could probably do them outside the loop, rather than tightly around
> the actual move instructions. Peter (hpa), is there some sane
> interface to try to do that?

iov_iter_fault_in_readable() is just going and touching a single word in
the page so that it is faulted in, or a pair of words if it manages to
cross a page boundary (which isn't happening here). I'm not sure
there's a loop to move them out of here (for the prefaulting part).

We could theoretically expand the stac/clac to be around the pair of
__get_user()s in fault_in_pages_readable() but that would only help the
case where we are crossing a page boundary.

Although I was probably wrong about the source of the overhead, the
point still remains that the prefaulting is eating cycles for no
practical benefit.

>> What we're seeing here
>> probably isn't actually stac/clac overhead, but the cost of finishing
>> some other operations that are outstanding before we can proceed through
>> here.
> I suspect it actually _is_ stac/clac overhead. It might well be that
> clac/stac ends up serializing loads some way. Last I heard, they were
> reasonably cheap but certainly not free - and when we're talking about
> something that just loops over bringing the line into cache, it might
> be relatively expensive.
> How did you do the profile? Use "-e cycles:pp" to get the precise
> profile information, which should actually attribute the cost to the
> instruction that really causes it.

It reduced the skid a bit.

Plain (no -e"):
> │ stac
> 24.57 │ mov (%rcx),%sil
> 15.70 │ clac
> 28.77 │ test %eax,%eax
> 2.15 │ mov %sil,-0x1(%rbp)
> 8.93 │ ↓ jne 66
> 2.31 │ movslq %edx,%rdx

With "-e cycles:pp":
> │ sub $0x8,%rsp
> 24.57 │ stac
> 15.49 │ mov (%rcx),%sil
> 29.06 │ clac
> 2.24 │ test %eax,%eax
> 8.77 │ mov %sil,-0x1(%rbp)
> 2.22 │ ↓ jne 66
> │ movslq %edx,%rdx

 \ /
  Last update: 2015-10-05 23:01    [W:0.375 / U:0.096 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site