On Sun, Feb 12, 2017 at 02:00:19PM +0000, Chris Wilson wrote:
> Disabling a tasklet causes it not to run during tasklet_action, but is
> put back onto the runnable tasklet list, and a new softirq raised. As
> the softirq is raised from within __do_softirq() this causing
> __do_softirq() to loop constantly until its timeslice expires and is
> transferred to the ksoftirq thread. ksoftirq then permanently spins,
> as on each action, the disabled tasklet keeps reraising the softirq.
>
> Break this vicious cycle by moving the softirq from the action to the
> final tasklet_enable().
>
> This behaviour appears to be historic (since the first git import).
> However, the looping until timeslice duration (to a max of 2ms) was
> first introduced in commit c10d73671ad3 ("softirq: reduce latencies"),
> with the restart limit restored in commit 34376a50fb1f ("Fix lockup
> related to stop_machine being stuck in __do_softirq.")
>
> Reported-by: Tvrtko Ursulin <[email protected]>
> Signed-off-by: Chris Wilson <[email protected]>
> Cc: Tvrtko Ursulin <[email protected]>
> Cc: Thomas Gleixner <[email protected]>
> Cc: Hannes Reinecke <[email protected]>
> Cc: Jens Axboe <[email protected]>
> Cc: Bjorn Helgaas <[email protected]>
> Cc: Alexander Potapenko <[email protected]>
> Cc: Chen Fan <[email protected]>
> Cc: Ingo Molnar <[email protected]>
> Cc: "Peter Zijlstra (Intel)" <[email protected]>
> Cc: Sebastian Andrzej Siewior <[email protected]>
> Cc: Johannes Thumshirn <[email protected]>
> Cc: Emese Revfy <[email protected]>
> Cc: Sagi Grimberg <[email protected]>
> Cc: Eric Dumazet <[email protected]>
> Cc: Tom Herbert <[email protected]>
> Cc: Ben Hutchings <[email protected]>
> ---
> include/linux/interrupt.h | 7 +++++--
> kernel/softirq.c | 2 --
> 2 files changed, 5 insertions(+), 4 deletions(-)
>
> diff --git a/include/linux/interrupt.h b/include/linux/interrupt.h
> index 53144e78a369..12750f00d00d 100644
> --- a/include/linux/interrupt.h
> +++ b/include/linux/interrupt.h
> @@ -613,8 +613,11 @@ static inline void tasklet_disable(struct tasklet_struct
> *t)
>
> static inline void tasklet_enable(struct tasklet_struct *t)
> {
> - smp_mb__before_atomic();
> - atomic_dec(&t->count);
> + if (!atomic_dec_and_test(&t->count))
> + return;
> +
> + if (test_bit(TASKLET_STATE_SCHED, &t->state))
> + raise_softirq(HI_SOFTIRQ | TASKLET_SOFTIRQ);
raise_softirq is not exported, so let's move tasklet_enable()
out-of-line and export that.
-Chris
--
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
