On Thu, Oct 04, 2018 at 12:34:07PM -0700, Paul E. McKenney wrote:

On Thu, Oct 04, 2018 at 11:59:49AM -0700, Dmitry Torokhov wrote:

quoted

Hi Eric,

On Thu, Oct 04, 2018 at 08:47:49AM -0700, Eric Dumazet wrote:

quoted

syzbot was able to trigger rcu stalls by calling write()
with large number of bytes.

Add a cond_resched() in the loop to avoid this.

I think this simply masks a deeper issue. The code fetches characters
from userspace in a loop, takes a lock, quickly places response in an
output buffer, and releases interrupt. I do not see why this should
cause stalls as we do not hold spinlock/interrupts off for extended
period of time.

Adding Paul so he can straighten me out...

If you are running a !PREEMPT kernel, then you need the cond_resched()
to allow the scheduler to choose someone else to run if needed and
to let RCU know that grace periods can end.  Without the cond_resched(),
if you stay in that loop long enough you will get excessive scheduling
latencies and eventually even RCU CPU stall warning splats.

In a PREEMPT (instead of !PREEMPT) kernel, you would be right.  When
preemption is enabled, the scheduler can preempt and RCU can sense
lack of readers from the scheduling-clock interrupt handler.  Which
is why cond_resched() is nothingness in a PREEMPT kernel.

But because people run !PREEMPT as well as PREEMPT kernels, if that loop
can run for a long time, you need that cond_resched().

OK, I see. I'll apply the patch then.

I think evdev.c needs similar treatment as it will keep looping while
there is data...

Thanks.

-- 
Dmitry