[Intel-gfx] [PATCH 01/45] drm/i915: Seal races between async GPU cancellation, retirement and signaling
Tvrtko Ursulin
tvrtko.ursulin at linux.intel.com
Thu Apr 25 10:35:01 UTC 2019
On 25/04/2019 10:19, Chris Wilson wrote:
> Currently there is an underlying assumption that i915_request_unsubmit()
> is synchronous wrt the GPU -- that is the request is no longer in flight
> as we remove it. In the near future that may change, and this may upset
> our signaling as we can process an interrupt for that request while it
> is no longer in flight.
>
> CPU0 CPU1
> intel_engine_breadcrumbs_irq
> (queue request completion)
> i915_request_cancel_signaling
> ... ...
> i915_request_enable_signaling
> dma_fence_signal
>
> Hence in the time it took us to drop the lock to signal the request, a
> preemption event may have occurred and re-queued the request. In the
> process, that request would have seen I915_FENCE_FLAG_SIGNAL clear and
> so reused the rq->signal_link that was in use on CPU0, leading to bad
> pointer chasing in intel_engine_breadcrumbs_irq.
>
> A related issue was that if someone started listening for a signal on a
> completed but no longer in-flight request, we missed the opportunity to
> immediately signal that request.
>
> Furthermore, as intel_contexts may be immediately released during
> request retirement, in order to be entirely sure that
> intel_engine_breadcrumbs_irq may no longer dereference the intel_context
> (ce->signals and ce->signal_link), we must wait for irq spinlock.
>
> In order to prevent the race, we use a bit in the fence.flags to signal
> the transfer onto the signal list inside intel_engine_breadcrumbs_irq.
> For simplicity, we use the DMA_FENCE_FLAG_SIGNALED_BIT as it then
> quickly signals to any outside observer that the fence is indeed signaled.
>
> Fixes: 52c0fdb25c7c ("drm/i915: Replace global breadcrumbs with per-context interrupt tracking")
> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> ---
> drivers/dma-buf/dma-fence.c | 1 +
> drivers/gpu/drm/i915/gt/intel_breadcrumbs.c | 58 +++++++++++++--------
> drivers/gpu/drm/i915/i915_request.c | 1 +
> 3 files changed, 39 insertions(+), 21 deletions(-)
>
> diff --git a/drivers/dma-buf/dma-fence.c b/drivers/dma-buf/dma-fence.c
> index 3aa8733f832a..9bf06042619a 100644
> --- a/drivers/dma-buf/dma-fence.c
> +++ b/drivers/dma-buf/dma-fence.c
> @@ -29,6 +29,7 @@
>
> EXPORT_TRACEPOINT_SYMBOL(dma_fence_emit);
> EXPORT_TRACEPOINT_SYMBOL(dma_fence_enable_signal);
> +EXPORT_TRACEPOINT_SYMBOL(dma_fence_signaled);
>
> static DEFINE_SPINLOCK(dma_fence_stub_lock);
> static struct dma_fence dma_fence_stub;
> diff --git a/drivers/gpu/drm/i915/gt/intel_breadcrumbs.c b/drivers/gpu/drm/i915/gt/intel_breadcrumbs.c
> index 3cbffd400b1b..4283224249d4 100644
> --- a/drivers/gpu/drm/i915/gt/intel_breadcrumbs.c
> +++ b/drivers/gpu/drm/i915/gt/intel_breadcrumbs.c
> @@ -23,6 +23,7 @@
> */
>
> #include <linux/kthread.h>
> +#include <trace/events/dma_fence.h>
> #include <uapi/linux/sched/types.h>
>
> #include "i915_drv.h"
> @@ -83,6 +84,7 @@ static inline bool __request_completed(const struct i915_request *rq)
> void intel_engine_breadcrumbs_irq(struct intel_engine_cs *engine)
> {
> struct intel_breadcrumbs *b = &engine->breadcrumbs;
> + const ktime_t timestamp = ktime_get();
> struct intel_context *ce, *cn;
> struct list_head *pos, *next;
> LIST_HEAD(signal);
> @@ -104,6 +106,11 @@ void intel_engine_breadcrumbs_irq(struct intel_engine_cs *engine)
>
> GEM_BUG_ON(!test_bit(I915_FENCE_FLAG_SIGNAL,
> &rq->fence.flags));
> + clear_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags);
> +
> + if (test_and_set_bit(DMA_FENCE_FLAG_SIGNALED_BIT,
> + &rq->fence.flags))
> + continue;
>
> /*
> * Queue for execution after dropping the signaling
> @@ -111,14 +118,6 @@ void intel_engine_breadcrumbs_irq(struct intel_engine_cs *engine)
> * more signalers to the same context or engine.
> */
> i915_request_get(rq);
> -
> - /*
> - * We may race with direct invocation of
> - * dma_fence_signal(), e.g. i915_request_retire(),
> - * so we need to acquire our reference to the request
> - * before we cancel the breadcrumb.
> - */
> - clear_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags);
> list_add_tail(&rq->signal_link, &signal);
> }
>
> @@ -140,8 +139,21 @@ void intel_engine_breadcrumbs_irq(struct intel_engine_cs *engine)
> list_for_each_safe(pos, next, &signal) {
> struct i915_request *rq =
> list_entry(pos, typeof(*rq), signal_link);
> + struct dma_fence_cb *cur, *tmp;
> +
> + trace_dma_fence_signaled(&rq->fence);
> +
> + rq->fence.timestamp = timestamp;
> + set_bit(DMA_FENCE_FLAG_TIMESTAMP_BIT, &rq->fence.flags);
> +
> + spin_lock(&rq->lock);
> + list_for_each_entry_safe(cur, tmp, &rq->fence.cb_list, node) {
> + INIT_LIST_HEAD(&cur->node);
> + cur->func(&rq->fence, cur);
> + }
> + INIT_LIST_HEAD(&rq->fence.cb_list);
> + spin_unlock(&rq->lock);
>
> - dma_fence_signal(&rq->fence);
I posted some comments on this patch already. In essence it was a
suggestion to not open-code-and-optimize dma_fence_signal, but split it
into two low-level helpers and export from the parent location. Like
__dma_fence_maybe/start_signal and __dma_fence_finish_signal.
bool __dma_fence_start_signal(...)
{
return test_and_set_bit(DMA_FENCE_FLAG_SIGNALED_BIT,
fence->flags);
}
EXPORT_SYMBOL(...)
void __dma_fence_finish_signal(...)
{
trace_dma_fence_signaled...
timestamp...
callbacks...
}
EXPORT_SYMBOL(...)
This way we don't add coupling to low level implementation details in i915.
Regards,
Tvrtko
> i915_request_put(rq);
> }
> }
> @@ -243,19 +255,17 @@ void intel_engine_fini_breadcrumbs(struct intel_engine_cs *engine)
>
> bool i915_request_enable_breadcrumb(struct i915_request *rq)
> {
> - struct intel_breadcrumbs *b = &rq->engine->breadcrumbs;
> -
> - GEM_BUG_ON(test_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags));
> -
> - if (!test_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags))
> - return true;
> + lockdep_assert_held(&rq->lock);
> + lockdep_assert_irqs_disabled();
>
> - spin_lock(&b->irq_lock);
> - if (test_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags) &&
> - !__request_completed(rq)) {
> + if (test_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags)) {
> + struct intel_breadcrumbs *b = &rq->engine->breadcrumbs;
> struct intel_context *ce = rq->hw_context;
> struct list_head *pos;
>
> + spin_lock(&b->irq_lock);
> + GEM_BUG_ON(test_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags));
> +
> __intel_breadcrumbs_arm_irq(b);
>
> /*
> @@ -284,8 +294,8 @@ bool i915_request_enable_breadcrumb(struct i915_request *rq)
> list_move_tail(&ce->signal_link, &b->signalers);
>
> set_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags);
> + spin_unlock(&b->irq_lock);
> }
> - spin_unlock(&b->irq_lock);
>
> return !__request_completed(rq);
> }
> @@ -294,9 +304,15 @@ void i915_request_cancel_breadcrumb(struct i915_request *rq)
> {
> struct intel_breadcrumbs *b = &rq->engine->breadcrumbs;
>
> - if (!test_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags))
> - return;
> + lockdep_assert_held(&rq->lock);
> + lockdep_assert_irqs_disabled();
>
> + /*
> + * We must wait for b->irq_lock so that we know the interrupt handler
> + * has released its reference to the intel_context and has completed
> + * the DMA_FENCE_FLAG_SIGNALED_BIT/I915_FENCE_FLAG_SIGNAL dance (if
> + * required).
> + */
> spin_lock(&b->irq_lock);
> if (test_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags)) {
> struct intel_context *ce = rq->hw_context;
> diff --git a/drivers/gpu/drm/i915/i915_request.c b/drivers/gpu/drm/i915/i915_request.c
> index 5869c37a35e1..fa9037173f1d 100644
> --- a/drivers/gpu/drm/i915/i915_request.c
> +++ b/drivers/gpu/drm/i915/i915_request.c
> @@ -417,6 +417,7 @@ void __i915_request_submit(struct i915_request *request)
> set_bit(I915_FENCE_FLAG_ACTIVE, &request->fence.flags);
>
> if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &request->fence.flags) &&
> + !test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &request->fence.flags) &&
> !i915_request_enable_breadcrumb(request))
> intel_engine_queue_breadcrumbs(engine);
>
>
More information about the Intel-gfx
mailing list