On 16/05/2018 16:12, Chris Wilson wrote:
Catch up with the inflight CSB events, after disabling the tasklet
before deciding which request was truly guilty of hanging the GPU.
Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
Cc: Michał Winiarski <michal.winiarski@xxxxxxxxx>
CC: Michel Thierry <michel.thierry@xxxxxxxxx>
Cc: Jeff McGee <jeff.mcgee@xxxxxxxxx>
Cc: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxxxxxxxx>
---
drivers/gpu/drm/i915/intel_lrc.c | 36 +++++++++++++++++++++++++++++++-
1 file changed, 35 insertions(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c
index e70d8d624899..3b889bb4352a 100644
--- a/drivers/gpu/drm/i915/intel_lrc.c
+++ b/drivers/gpu/drm/i915/intel_lrc.c
@@ -1843,6 +1843,7 @@ static struct i915_request *
execlists_reset_prepare(struct intel_engine_cs *engine)
{
struct intel_engine_execlists * const execlists = &engine->execlists;
+ struct i915_request *request, *active;
GEM_TRACE("%s\n", engine->name);
@@ -1857,7 +1858,40 @@ execlists_reset_prepare(struct intel_engine_cs *engine)
*/
__tasklet_disable_sync_once(&execlists->tasklet);
- return i915_gem_find_active_request(engine);
+ /*
+ * We want to flush the pending context switches, having disabled
+ * the tasklet above, we can assume exclusive access to the execlists.
+ * For this allows us to catch up with an inflight preemption event,
+ * and avoid blaming an innocent request if the stall was due to the
+ * preemption itself.
+ */
+ if (test_bit(ENGINE_IRQ_EXECLIST, &engine->irq_posted))
+ process_csb(engine);
+
+ /*
+ * The last active request can then be no later than the last request
+ * now in ELSP[0]. So search backwards from there, so that if the GPU
+ * has advanced beyond the last CSB update, it will be pardoned.
+ */
+ active = NULL;
+ request = port_request(execlists->port);
+ if (request) {
+ unsigned long flags;
+
+ spin_lock_irqsave(&engine->timeline.lock, flags);
+ list_for_each_entry_from_reverse(request,
+ &engine->timeline.requests,
+ link) {
+ if (__i915_request_completed(request,
+ request->global_seqno))
+ break;
+
+ active = request;
+ }
+ spin_unlock_irqrestore(&engine->timeline.lock, flags);
+ }
+
+ return active;
}
static void execlists_reset(struct intel_engine_cs *engine,
Sounds sensible to me, just don't ask me to describe preemption and
reset flow where things go bad without it. :)
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx>
Regards,
Tvrtko
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
