[Intel-gfx] [PATCH 08/10] drm/i915: Move the irq_counter inside the spinlock

[Tvrtko Ursulin]

On 15/01/2018 21:24, Chris Wilson wrote:
> Rather than have multiple locked instructions inside the notify_ring()
> irq handler, move them inside the spinlock and reduce their intrinsic
> locking.
> 
> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> ---
>   drivers/gpu/drm/i915/i915_gem_request.c  |  4 ++--
>   drivers/gpu/drm/i915/i915_irq.c          |  6 +++---
>   drivers/gpu/drm/i915/intel_breadcrumbs.c | 13 ++++++++-----
>   drivers/gpu/drm/i915/intel_ringbuffer.h  |  2 +-
>   4 files changed, 14 insertions(+), 11 deletions(-)
> 
> diff --git a/drivers/gpu/drm/i915/i915_gem_request.c b/drivers/gpu/drm/i915/i915_gem_request.c
> index 836db90ef81b..08bbd56277e5 100644
> --- a/drivers/gpu/drm/i915/i915_gem_request.c
> +++ b/drivers/gpu/drm/i915/i915_gem_request.c
> @@ -1128,7 +1128,7 @@ static bool __i915_spin_request(const struct drm_i915_gem_request *req,
>   	 * takes to sleep on a request, on the order of a microsecond.
>   	 */
>   
> -	irq = atomic_read(&engine->irq_count);
> +	irq = READ_ONCE(engine->breadcrumbs.irq_count);
>   	timeout_us += local_clock_us(&cpu);
>   	do {
>   		if (i915_seqno_passed(intel_engine_get_seqno(engine), seqno))
> @@ -1139,7 +1139,7 @@ static bool __i915_spin_request(const struct drm_i915_gem_request *req,
>   		 * assume we won't see one in the near future but require
>   		 * the engine->seqno_barrier() to fixup coherency.
>   		 */
> -		if (atomic_read(&engine->irq_count) != irq)
> +		if (READ_ONCE(engine->breadcrumbs.irq_count) != irq)
>   			break;
>   
>   		if (signal_pending_state(state, current))
> diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c
> index 0b272501b738..e5f76d580010 100644
> --- a/drivers/gpu/drm/i915/i915_irq.c
> +++ b/drivers/gpu/drm/i915/i915_irq.c
> @@ -1073,9 +1073,6 @@ static void notify_ring(struct intel_engine_cs *engine)
>   	if (unlikely(!engine->breadcrumbs.irq_armed))
>   		return;
>   
> -	atomic_inc(&engine->irq_count);
> -	set_bit(ENGINE_IRQ_BREADCRUMB, &engine->irq_posted);
> -
>   	rcu_read_lock();
>   
>   	spin_lock(&engine->breadcrumbs.irq_lock);
> @@ -1107,6 +1104,9 @@ static void notify_ring(struct intel_engine_cs *engine)
>   			    i915_seqno_passed(seqno, wait->seqno - 1))
>   				tsk = wait->tsk;
>   		}
> +
> +		engine->breadcrumbs.irq_count++;
> +		__set_bit(ENGINE_IRQ_BREADCRUMB, &engine->irq_posted);

I'm nervous about moving the ENGINE_IRQ_BREADCRUMB setting to be 
conditional. __i915_request_irq_complete documents the ordering of these 
things is crucial so I worry we don't miss a wakeup. Once bitten twice 
shy? Don't know..

irq_count change looks safe, so can I, once again, suggest to split into 
two patches? :/

Regards,

Tvrtko

>   	} else {
>   		if (engine->breadcrumbs.irq_armed)
>   			__intel_engine_disarm_breadcrumbs(engine);
> diff --git a/drivers/gpu/drm/i915/intel_breadcrumbs.c b/drivers/gpu/drm/i915/intel_breadcrumbs.c
> index e3667dc1e96d..7c82cfe23922 100644
> --- a/drivers/gpu/drm/i915/intel_breadcrumbs.c
> +++ b/drivers/gpu/drm/i915/intel_breadcrumbs.c
> @@ -98,12 +98,14 @@ static void intel_breadcrumbs_hangcheck(struct timer_list *t)
>   	struct intel_engine_cs *engine =
>   		from_timer(engine, t, breadcrumbs.hangcheck);
>   	struct intel_breadcrumbs *b = &engine->breadcrumbs;
> +	unsigned int irq_count;
>   
>   	if (!b->irq_armed)
>   		return;
>   
> -	if (b->hangcheck_interrupts != atomic_read(&engine->irq_count)) {
> -		b->hangcheck_interrupts = atomic_read(&engine->irq_count);
> +	irq_count = READ_ONCE(b->irq_count);
> +	if (b->hangcheck_interrupts != irq_count) {
> +		b->hangcheck_interrupts = irq_count;
>   		mod_timer(&b->hangcheck, wait_timeout());
>   		return;
>   	}
> @@ -176,7 +178,7 @@ static void irq_enable(struct intel_engine_cs *engine)
>   	 * we still need to force the barrier before reading the seqno,
>   	 * just in case.
>   	 */
> -	set_bit(ENGINE_IRQ_BREADCRUMB, &engine->irq_posted);
> +	__set_bit(ENGINE_IRQ_BREADCRUMB, &engine->irq_posted);
>   
>   	/* Caller disables interrupts */
>   	spin_lock(&engine->i915->irq_lock);
> @@ -270,13 +272,14 @@ static bool use_fake_irq(const struct intel_breadcrumbs *b)
>   	if (!test_bit(engine->id, &engine->i915->gpu_error.missed_irq_rings))
>   		return false;
>   
> -	/* Only start with the heavy weight fake irq timer if we have not
> +	/*
> +	 * Only start with the heavy weight fake irq timer if we have not
>   	 * seen any interrupts since enabling it the first time. If the
>   	 * interrupts are still arriving, it means we made a mistake in our
>   	 * engine->seqno_barrier(), a timing error that should be transient
>   	 * and unlikely to reoccur.
>   	 */
> -	return atomic_read(&engine->irq_count) == b->hangcheck_interrupts;
> +	return READ_ONCE(b->irq_count) == b->hangcheck_interrupts;
>   }
>   
>   static void enable_fake_irq(struct intel_breadcrumbs *b)
> diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.h b/drivers/gpu/drm/i915/intel_ringbuffer.h
> index c5ff203e42d6..f406d0ff4612 100644
> --- a/drivers/gpu/drm/i915/intel_ringbuffer.h
> +++ b/drivers/gpu/drm/i915/intel_ringbuffer.h
> @@ -305,7 +305,6 @@ struct intel_engine_cs {
>   
>   	struct drm_i915_gem_object *default_state;
>   
> -	atomic_t irq_count;
>   	unsigned long irq_posted;
>   #define ENGINE_IRQ_BREADCRUMB 0
>   #define ENGINE_IRQ_EXECLIST 1
> @@ -340,6 +339,7 @@ struct intel_engine_cs {
>   
>   		unsigned int hangcheck_interrupts;
>   		unsigned int irq_enabled;
> +		unsigned int irq_count;
>   
>   		bool irq_armed : 1;
>   		I915_SELFTEST_DECLARE(bool mock : 1);
>