On 17/04/2019 08:56, 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@xxxxxxxxxxxxxxxxxx>
Cc: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx>
---
drivers/dma-buf/dma-fence.c | 1 +
drivers/gpu/drm/i915/i915_request.c | 1 +
drivers/gpu/drm/i915/intel_breadcrumbs.c | 52 ++++++++++++++----------
3 files changed, 33 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/i915_request.c b/drivers/gpu/drm/i915/i915_request.c
index b836721d3b13..e0efc334463b 100644
--- a/drivers/gpu/drm/i915/i915_request.c
+++ b/drivers/gpu/drm/i915/i915_request.c
@@ -432,6 +432,7 @@ 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);
diff --git a/drivers/gpu/drm/i915/intel_breadcrumbs.c b/drivers/gpu/drm/i915/intel_breadcrumbs.c
index 3cbffd400b1b..e19f84b006cc 100644
--- a/drivers/gpu/drm/i915/intel_breadcrumbs.c
+++ b/drivers/gpu/drm/i915/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;
From here to below is intimate coupling with the dma_fence_signal
implementation, via open-coding it (with some optimizations as well).
I am thinking about this solution.. here we put:
if (!__dma_fence_maybe_signal(&rq->fence))
continue;
Add the low-level helper to dma-fence.c and export it.
And below..
/*
* 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);
..we do:
__dma_fence_complete/force/finish_signal(&rq->fence, timestamp));
Also export etc, instead of the whole above block.
This way it is both self-documenting in our code and we remove the
intimate coupling with dma-fence implementation details.
No need to export the tracepoint then either.
(You can have a prequel patch to optimise the list_del_init in
dma_fence_signal.)
Thoughts?
Regards,
Tvrtko
- dma_fence_signal(&rq->fence);
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);
}
@@ -294,8 +304,8 @@ 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();
spin_lock(&b->irq_lock);
if (test_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags)) {
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
