lkml.org 
[lkml]   [2015]   [May]   [26]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [RFC PATCH] percpu system call: fast userspace percpu critical sections
On Tue, May 26, 2015 at 2:18 PM, Andy Lutomirski <luto@amacapital.net> wrote:
> On Tue, May 26, 2015 at 2:04 PM, Mathieu Desnoyers
>>
>>>
>>> It's too bad that not all architectures have a single-instruction
>>> unlocked compare-and-exchange.
>>
>> Based on my benchmarks, it's not clear that single-instruction
>> unlocked CAS is actually faster than doing the same with many
>> instructions.
>
> True, but with a single instruction the user can't get preempted in the middle.
>
> Looking at your code, it looks like percpu_user_sched_in has some
> potentially nasty issues with page faults. Avoiding touching user
> memory from the scheduler would be quite nice from an implementation
> POV, and the x86-specific gs hack wins in that regard.

ARM has "TLB lockdown entries" which could, I think, be used to
implement per-cpu or per-thread mappings. I'm actually rather
surprised that Linux doesn't already use a TLB lockdown entry for TLS.
(Hmm. Maybe it's because the interface to write the entries requires
actually touching the page. Maybe not -- the ARM docs, in general,
seem to be much less clear than the Intel and AMD docs.)

ARM doesn't seem to have any single-instruction compare-exchange or
similar instruction, though, so this might be all that useful. On the
other hand, ARM can probably do reasonably efficient per-cpu memory
allocation and such with a single ldrex/strex pair.

--Andy


\
 
 \ /
  Last update: 2015-05-27 00:01    [W:0.084 / U:0.028 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site