Softirq preemption

Softirq preemption

[Lee Revell]

[-- Attachment #1: Type: text/plain, Size: 653 bytes --]

Ingo,

Is there any chance that the softirq preemption feature from the -rt
kernel can be pushed upstream?

Currently I am testing 2.6.14-rt21 with PREEMPT_DESKTOP and hard/softirq
preemption disabled in order to simulate the behavior of 2.6.14
mainline.  The network TX softirq is the longest non-preemptible code
path, it can run for 5-6ms.  However I can see that the preempt_count
drops to zero on many occasions, which IIRC means that no locks are held
and we could easily preempt at that point.

Without softirq	preemption I don't see how we have any chance of meeting
the 1ms goal for desktop responsiveness that was decided on at OLS 2004.

Lee

[-- Attachment #2: softirq-tx.trace.gz --]
[-- Type: application/x-gzip, Size: 36670 bytes --]

Re: Softirq preemption

[Lee Revell]

[-- Attachment #1: Type: text/plain, Size: 605 bytes --]

On Thu, 2005-12-01 at 14:08 -0500, Lee Revell wrote:
> Ingo,
> 
> Is there any chance that the softirq preemption feature from the -rt
> kernel can be pushed upstream?
> 

FWIW, here is the longest latency I see with 2.6.14-rt22, with the same
preemption settings as 2.6.14 + PREEMPT.  AFAICT a longish network
softirq gets interrupted by several disk and network interrupts which
adds up to ~1ms.  Is there any easy way to break up this code path that
could be acceptable for mainline, or does this seem like a reasonable
lower limit on achievable latency (until/unless PREEMPT_RT gets merged
etc)?

Lee

[-- Attachment #2: 2.6.14-rt22-trace.gz --]
[-- Type: application/x-gzip, Size: 7015 bytes --]

Re: Softirq preemption

[Ingo Molnar]

* Lee Revell <rlrevell@joe-job.com> wrote:

> Is there any chance that the softirq preemption feature from the -rt 
> kernel can be pushed upstream?

yeah, will do it, once the current hrtimer push has been completed. The 
softirq/hardirq preemption feature is dependent on the genirq patch too.

	Ingo