lkml.org 
[lkml]   [2011]   [Mar]   [27]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [GIT PULL] Core block IO bits for 2.6.39
On 2011-03-27 13:49, Avi Kivity wrote:
> On 03/24/2011 03:43 PM, Jens Axboe wrote:
>> Hi Linus,
>>
>> This is the main pull request for the block IO layer and friends for
>> 2.6.39.
>>
>> There are two major things in this tree:
>>
>> - The removal of the per-device plugging state for disks. On fast
>> devices, it ended up hammering the queue lock quite hard. The new
>> scheme puts the plugging state on the stack and allows an IO submitter
>> to finish his batch of IO before pushing it to the queue. Once that
>> push starts, we'll insert/merge with the existing queue.
>>
>> A pointer to this plugging context is stored in the task structure. If
>> a task ends up blocking before it has submitted it's IO (usual cause
>> would be memory allocation of some sort), the plugged list is
>> auto-submitted before the task goes to sleep.
>
> This is the fourth "do something if preempted" hook (the other three are
> kvm, cmwq, and perf). Why not use sched notifiers for this?

It's a 'if preempted', it's 'if going to sleep'. Two issues with the
preempt notifiers for this, I did look into it:

- Not unconditionally available. Not a big problem, we could just do
that.

- Sched out is called under the runqueue lock, and it's not safe to drop
it.

But it's not a bad point, it would be nice if we could unify all these
use cases and provide a notifier that works for all.

--
Jens Axboe



\
 
 \ /
  Last update: 2011-03-27 14:03    [from the cache]
©2003-2011 Jasper Spaans