On Fri, Jan 12, 2018 at 06:35:54AM +0100, Frederic Weisbecker wrote:
> Some softirq vectors can be more CPU hungry than others. Especially
> networking may sometimes deal with packet storm and need more CPU than
> IRQ tail can offer without inducing scheduler latencies. In this case
> the current code defers to ksoftirqd that behaves nicer. Now this nice
> behaviour can be bad for other IRQ vectors that usually need quick
> processing.
>
> To solve this we only defer to threading the vectors that outreached the
> time limit on IRQ tail processing and leave the others inline on real
> Soft-IRQs service. This is achieved using workqueues with
> per-CPU/per-vector worklets.
>
> Note ksoftirqd is not removed as it is still needed for threaded IRQs
> mode.
>
> Suggested-by: Linus Torvalds <torva...@linux-foundation.org>
> Signed-off-by: Frederic Weisbecker <frede...@kernel.org>
> Cc: Dmitry Safonov <d...@arista.com>
> Cc: Eric Dumazet <eduma...@google.com>
> Cc: Linus Torvalds <torva...@linux-foundation.org>
> Cc: Peter Zijlstra <pet...@infradead.org>
> Cc: Andrew Morton <a...@linux-foundation.org>
> Cc: David Miller <da...@davemloft.net>
> Cc: Hannes Frederic Sowa <han...@stressinduktion.org>
> Cc: Ingo Molnar <mi...@kernel.org>
> Cc: Levin Alexander <alexander.le...@verizon.com>
> Cc: Paolo Abeni <pab...@redhat.com>
> Cc: Paul E. McKenney <paul...@linux.vnet.ibm.com>
> Cc: Radu Rendec <rren...@arista.com>
> Cc: Rik van Riel <r...@redhat.com>
> Cc: Stanislaw Gruszka <sgrus...@redhat.com>
> Cc: Thomas Gleixner <t...@linutronix.de>
> Cc: Wanpeng Li <wanpeng...@hotmail.com>
> ---
> kernel/softirq.c | 90
> ++++++++++++++++++++++++++++++++++++++++++++++++++++++--
> 1 file changed, 87 insertions(+), 3 deletions(-)
>
> diff --git a/kernel/softirq.c b/kernel/softirq.c
> index fa267f7..0c817ec6 100644
> --- a/kernel/softirq.c
> +++ b/kernel/softirq.c
> @@ -74,6 +74,13 @@ struct softirq_stat {
>
> static DEFINE_PER_CPU(struct softirq_stat, softirq_stat_cpu);
>
> +struct vector_work {
> + int vec;
> + struct work_struct work;
> +};
> +
> +static DEFINE_PER_CPU(struct vector_work[NR_SOFTIRQS], vector_work_cpu);
> +
> /*
> * we cannot loop indefinitely here to avoid userspace starvation,
> * but we also don't want to introduce a worst case 1/HZ latency
> @@ -251,6 +258,70 @@ static inline bool lockdep_softirq_start(void) { return
> false; }
> static inline void lockdep_softirq_end(bool in_hardirq) { }
> #endif
>
> +static void vector_work_func(struct work_struct *work)
> +{
> + struct vector_work *vector_work;
> + u32 pending;
> + int vec;
> +
> + vector_work = container_of(work, struct vector_work, work);
> + vec = vector_work->vec;
> +
> + local_irq_disable();
> + pending = local_softirq_pending();
> + account_irq_enter_time(current);
> + __local_bh_disable_ip(_RET_IP_, SOFTIRQ_OFFSET);
> + lockdep_softirq_enter();
> + set_softirq_pending(pending & ~(1 << vec));
> + local_irq_enable();
> +
> + if (pending & (1 << vec)) {
Ah I see the problem. Say in do_softirq() we had pending VECTOR 1 and 2.
And we had overrun only VECTOR 1 so VECTOR 1 is enqueued to workqueue.
Right after that we go back to the restart loop in do_softirq in order to
handle pending VECTOR 2 but we erase the local_softirqs_pending state. So
when the workqueue runs, it doesn't see anymore VECTOR 1 pending and we lose
it.
So I need to remove the above condition and make the vector work
unconditionally execute the vector callback.
Now I can go to sleep...
> + struct softirq_action *sa = &softirq_vec[vec];
> +
> + kstat_incr_softirqs_this_cpu(vec);
> + trace_softirq_entry(vec);
> + sa->action(sa);
> + trace_softirq_exit(vec);
> + }
