[PATCH 3.4.62-rt77] fix irq thread scheduling

[PATCH 3.4.62-rt77] fix irq thread scheduling

[Thomas Pfaff]

If an irq is requested during device open, for example by opening a serial 
device, then request_irq is called in the context of the opening process.
If the process lacks off CAP_SYS_NICE the irq thread is running with SCHED_OTHER 
priority.

The problem exists in other real time patches as well, 3.10 seems to be ok.

Signed-off-by: Thomas Pfaff <tpfaff@pcs.com>
---
diff -urp linux-3.4.62-rt.old/kernel/irq/manage.c linux-3.4.62-rt/kernel/irq/manage.c
--- linux-3.4.62-rt.old/kernel/irq/manage.c	2013-09-16 10:13:15.365559288 +0200
+++ linux-3.4.62-rt/kernel/irq/manage.c	2013-09-19 10:22:40.092831656 +0200
@@ -1024,7 +1024,7 @@ __setup_irq(unsigned int irq, struct irq
 			goto out_mput;
 		}
 
-		sched_setscheduler(t, SCHED_FIFO, &param);
+		sched_setscheduler_nocheck(t, SCHED_FIFO, &param);
 
 		/*
 		 * We keep the reference to the task struct even if

Re: [PATCH 3.4.62-rt77] fix irq thread scheduling

[Sebastian Andrzej Siewior]

* Thomas Pfaff | 2013-09-19 11:02:57 [+0200]:

>If an irq is requested during device open, for example by opening a serial 
>device, then request_irq is called in the context of the opening process.
>If the process lacks off CAP_SYS_NICE the irq thread is running with SCHED_OTHER 
>priority.
>
>The problem exists in other real time patches as well, 3.10 seems to be ok.

Hmmm. I see the same check in v3.10 compared to v3.6 which so I *think*
v3.10 has the same problem.
However this does not look to be -RT specific, you should see the same
problem in mainline, too. Could you please verify whether this does
affect v3.10? And if it does, we should try to fix up it upstream, too

Sebastian

Re: [PATCH 3.4.62-rt77] fix irq thread scheduling

[Thomas Pfaff]

On Fri, 4 Oct 2013, Sebastian Andrzej Siewior wrote:

> * Thomas Pfaff | 2013-09-19 11:02:57 [+0200]:
> 
> >If an irq is requested during device open, for example by opening a serial 
> >device, then request_irq is called in the context of the opening process.
> >If the process lacks off CAP_SYS_NICE the irq thread is running with SCHED_OTHER 
> >priority.
> >
> >The problem exists in other real time patches as well, 3.10 seems to be ok.
> 
> Hmmm. I see the same check in v3.10 compared to v3.6 which so I *think*
> v3.10 has the same problem.
> However this does not look to be -RT specific, you should see the same
> problem in mainline, too. Could you please verify whether this does
> affect v3.10? And if it does, we should try to fix up it upstream, too
> 

I did a quick check with 3.10.14-rt9 and it is not affected, because 
sched_setscheduler is called in the context of the irq thread as the mainline 
kernel does, but the 3.6 rt patch moved this into __setup_irq.

Thomas

Re: [PATCH 3.4.62-rt77] fix irq thread scheduling

[Thomas Pfaff]

On Tue, 8 Oct 2013, Pfaff Thomas wrote:

> 
> 
> On Fri, 4 Oct 2013, Sebastian Andrzej Siewior wrote:
> 
> > * Thomas Pfaff | 2013-09-19 11:02:57 [+0200]:
> > 
> > >If an irq is requested during device open, for example by opening a serial 
> > >device, then request_irq is called in the context of the opening process.
> > >If the process lacks off CAP_SYS_NICE the irq thread is running with SCHED_OTHER 
> > >priority.
> > >
> > >The problem exists in other real time patches as well, 3.10 seems to be ok.
> > 
> > Hmmm. I see the same check in v3.10 compared to v3.6 which so I *think*
> > v3.10 has the same problem.
> > However this does not look to be -RT specific, you should see the same
> > problem in mainline, too. Could you please verify whether this does
> > affect v3.10? And if it does, we should try to fix up it upstream, too
> > 
> 
> I did a quick check with 3.10.14-rt9 and it is not affected, because 
> sched_setscheduler is called in the context of the irq thread as the mainline 
> kernel does, but the 3.6 rt patch moved this into __setup_irq.

Sorry, should have been 3.4 rt patch.

Thomas

Re: [PATCH 3.4.62-rt77] fix irq thread scheduling

[Sebastian Andrzej Siewior]

* Thomas Pfaff | 2013-10-08 15:58:50 [+0200]:

>I did a quick check with 3.10.14-rt9 and it is not affected, because 
>sched_setscheduler is called in the context of the irq thread as the mainline 
>kernel does, but the 3.6 rt patch moved this into __setup_irq.

Thanks for that report. It looks that this is broken upstream as well.
I'm going to merge the patch that broke it, your fix and try to do the
same for upstream.

>Thomas

Sebastian