drm/i915: Extend LRC pinning to cover GPU context writeback
authorNick Hoath <nicholas.hoath@intel.com>
Tue, 1 Dec 2015 14:48:57 +0000 (14:48 +0000)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Thu, 3 Dec 2015 14:11:55 +0000 (15:11 +0100)
Use the first retired request on a new context to unpin
the old context. This ensures that the hw context remains
bound until it has been written back to by the GPU.
Now that the context is pinned until later in the request/context
lifecycle, it no longer needs to be pinned from context_queue to
retire_requests.
This fixes an issue with GuC submission where the GPU might not
have finished writing back the context before it is unpinned. This
results in a GPU hang.

v2: Moved the new pin to cover GuC submission (Alex Dai)
    Moved the new unpin to request_retire to fix coverage leak
v3: Added switch to default context if freeing a still pinned
    context just in case the hw was actually still using it
v4: Unwrapped context unpin to allow calling without a request
v5: Only create a switch to idle context if the ring doesn't
    already have a request pending on it (Alex Dai)
    Rename unsaved to dirty to avoid double negatives (Dave Gordon)
    Changed _no_req postfix to __ prefix for consistency (Dave Gordon)
    Split out per engine cleanup from context_free as it
    was getting unwieldy
    Corrected locking (Dave Gordon)
v6: Removed some bikeshedding (Mika Kuoppala)
    Added explanation of the GuC hang that this fixes (Daniel Vetter)
v7: Removed extra per request pinning from ring reset code (Alex Dai)
    Added forced ring unpin/clean in error case in context free (Alex Dai)

Signed-off-by: Nick Hoath <nicholas.hoath@intel.com>
Issue: VIZ-4277
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: David Gordon <david.s.gordon@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Alex Dai <yu.dai@intel.com>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Reviewed-by: Alex Dai <yu.dai@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
drivers/gpu/drm/i915/i915_drv.h
drivers/gpu/drm/i915/i915_gem.c
drivers/gpu/drm/i915/intel_lrc.c
drivers/gpu/drm/i915/intel_lrc.h

index 5e276f5d61da69d4085c297342c97e3dd693f190..f95f3676764970cf86216fea0a6b048ea5ce2a0e 100644 (file)
@@ -884,6 +884,7 @@ struct intel_context {
        struct {
                struct drm_i915_gem_object *state;
                struct intel_ringbuffer *ringbuf;
+               bool dirty;
                int pin_count;
        } engine[I915_NUM_RINGS];
 
index 33adc8f8ab20391cf1fcfd0b11db672c9c0d3f8c..6356c347ede0e8dcd993d9909a68f74a18dd5e8a 100644 (file)
@@ -1354,6 +1354,9 @@ static void i915_gem_request_retire(struct drm_i915_gem_request *request)
 {
        trace_i915_gem_request_retire(request);
 
+       if (i915.enable_execlists)
+               intel_lr_context_complete_check(request);
+
        /* We know the GPU must have read the request to have
         * sent us the seqno + interrupt, so use the position
         * of tail of the request to update the last known position
@@ -2764,10 +2767,6 @@ static void i915_gem_reset_ring_cleanup(struct drm_i915_private *dev_priv,
                                        struct drm_i915_gem_request,
                                        execlist_link);
                        list_del(&submit_req->execlist_link);
-
-                       if (submit_req->ctx != ring->default_context)
-                               intel_lr_context_unpin(submit_req);
-
                        i915_gem_request_unreference(submit_req);
                }
                spin_unlock_irq(&ring->execlist_lock);
index 4ebafab53f30e6492e99c0bcac6256f1f5b1c2b9..c3504a09340c301a047fa686eb50a3ea4129113e 100644 (file)
@@ -571,9 +571,6 @@ static int execlists_context_queue(struct drm_i915_gem_request *request)
        struct drm_i915_gem_request *cursor;
        int num_elements = 0;
 
-       if (request->ctx != ring->default_context)
-               intel_lr_context_pin(request);
-
        i915_gem_request_reference(request);
 
        spin_lock_irq(&ring->execlist_lock);
@@ -737,6 +734,13 @@ intel_logical_ring_advance_and_submit(struct drm_i915_gem_request *request)
        if (intel_ring_stopped(ring))
                return;
 
+       if (request->ctx != ring->default_context) {
+               if (!request->ctx->engine[ring->id].dirty) {
+                       intel_lr_context_pin(request);
+                       request->ctx->engine[ring->id].dirty = true;
+               }
+       }
+
        if (dev_priv->guc.execbuf_client)
                i915_guc_submit(dev_priv->guc.execbuf_client, request);
        else
@@ -963,12 +967,6 @@ void intel_execlists_retire_requests(struct intel_engine_cs *ring)
        spin_unlock_irq(&ring->execlist_lock);
 
        list_for_each_entry_safe(req, tmp, &retired_list, execlist_link) {
-               struct intel_context *ctx = req->ctx;
-               struct drm_i915_gem_object *ctx_obj =
-                               ctx->engine[ring->id].state;
-
-               if (ctx_obj && (ctx != ring->default_context))
-                       intel_lr_context_unpin(req);
                list_del(&req->execlist_link);
                i915_gem_request_unreference(req);
        }
@@ -1063,21 +1061,39 @@ reset_pin_count:
        return ret;
 }
 
-void intel_lr_context_unpin(struct drm_i915_gem_request *rq)
+static void __intel_lr_context_unpin(struct intel_engine_cs *ring,
+               struct intel_context *ctx)
 {
-       struct intel_engine_cs *ring = rq->ring;
-       struct drm_i915_gem_object *ctx_obj = rq->ctx->engine[ring->id].state;
-       struct intel_ringbuffer *ringbuf = rq->ringbuf;
-
+       struct drm_i915_gem_object *ctx_obj = ctx->engine[ring->id].state;
+       struct intel_ringbuffer *ringbuf = ctx->engine[ring->id].ringbuf;
        if (ctx_obj) {
                WARN_ON(!mutex_is_locked(&ring->dev->struct_mutex));
-               if (--rq->ctx->engine[ring->id].pin_count == 0) {
+               if (--ctx->engine[ring->id].pin_count == 0) {
                        intel_unpin_ringbuffer_obj(ringbuf);
                        i915_gem_object_ggtt_unpin(ctx_obj);
                }
        }
 }
 
+void intel_lr_context_unpin(struct drm_i915_gem_request *rq)
+{
+       __intel_lr_context_unpin(rq->ring, rq->ctx);
+}
+
+void intel_lr_context_complete_check(struct drm_i915_gem_request *req)
+{
+       struct intel_engine_cs *ring = req->ring;
+
+       if (ring->last_context && ring->last_context != req->ctx &&
+                       ring->last_context->engine[ring->id].dirty) {
+               __intel_lr_context_unpin(
+                               ring,
+                               ring->last_context);
+               ring->last_context->engine[ring->id].dirty = false;
+       }
+       ring->last_context = req->ctx;
+}
+
 static int intel_logical_ring_workarounds_emit(struct drm_i915_gem_request *req)
 {
        int ret, i;
@@ -2351,6 +2367,76 @@ populate_lr_context(struct intel_context *ctx, struct drm_i915_gem_object *ctx_o
        return 0;
 }
 
+/**
+ * intel_lr_context_clean_ring() - clean the ring specific parts of an LRC
+ * @ctx: the LR context being freed.
+ * @ring: the engine being cleaned
+ * @ctx_obj: the hw context being unreferenced
+ * @ringbuf: the ringbuf being freed
+ *
+ * Take care of cleaning up the per-engine backing
+ * objects and the logical ringbuffer.
+ */
+static void
+intel_lr_context_clean_ring(struct intel_context *ctx,
+                           struct intel_engine_cs *ring,
+                           struct drm_i915_gem_object *ctx_obj,
+                           struct intel_ringbuffer *ringbuf)
+{
+       int ret;
+
+       WARN_ON(!mutex_is_locked(&ring->dev->struct_mutex));
+
+       if (ctx == ring->default_context) {
+               intel_unpin_ringbuffer_obj(ringbuf);
+               i915_gem_object_ggtt_unpin(ctx_obj);
+       }
+
+       if (ctx->engine[ring->id].dirty) {
+               struct drm_i915_gem_request *req = NULL;
+
+               /**
+                * If there is already a request pending on
+                * this ring, wait for that to complete,
+                * otherwise create a switch to idle request
+                */
+               if (list_empty(&ring->request_list)) {
+                       int ret;
+
+                       ret = i915_gem_request_alloc(
+                                       ring,
+                                       ring->default_context,
+                                       &req);
+                       if (!ret)
+                               i915_add_request(req);
+                       else
+                               DRM_DEBUG("Failed to ensure context saved");
+               } else {
+                       req = list_first_entry(
+                                       &ring->request_list,
+                                       typeof(*req), list);
+               }
+               if (req) {
+                       ret = i915_wait_request(req);
+                       if (ret != 0) {
+                               /**
+                                * If we get here, there's probably been a ring
+                                * reset, so we just clean up the dirty flag.&
+                                * pin count.
+                                */
+                               ctx->engine[ring->id].dirty = false;
+                               __intel_lr_context_unpin(
+                                       ring,
+                                       ctx);
+                       }
+               }
+       }
+
+       WARN_ON(ctx->engine[ring->id].pin_count);
+       intel_ringbuffer_free(ringbuf);
+       drm_gem_object_unreference(&ctx_obj->base);
+}
+
 /**
  * intel_lr_context_free() - free the LRC specific bits of a context
  * @ctx: the LR context to free.
@@ -2363,7 +2449,7 @@ void intel_lr_context_free(struct intel_context *ctx)
 {
        int i;
 
-       for (i = 0; i < I915_NUM_RINGS; i++) {
+       for (i = 0; i < I915_NUM_RINGS; ++i) {
                struct drm_i915_gem_object *ctx_obj = ctx->engine[i].state;
 
                if (ctx_obj) {
@@ -2371,13 +2457,10 @@ void intel_lr_context_free(struct intel_context *ctx)
                                        ctx->engine[i].ringbuf;
                        struct intel_engine_cs *ring = ringbuf->ring;
 
-                       if (ctx == ring->default_context) {
-                               intel_unpin_ringbuffer_obj(ringbuf);
-                               i915_gem_object_ggtt_unpin(ctx_obj);
-                       }
-                       WARN_ON(ctx->engine[ring->id].pin_count);
-                       intel_ringbuffer_free(ringbuf);
-                       drm_gem_object_unreference(&ctx_obj->base);
+                       intel_lr_context_clean_ring(ctx,
+                                                   ring,
+                                                   ctx_obj,
+                                                   ringbuf);
                }
        }
 }
@@ -2539,5 +2622,12 @@ void intel_lr_context_reset(struct drm_device *dev,
 
                ringbuf->head = 0;
                ringbuf->tail = 0;
+
+               if (ctx->engine[ring->id].dirty) {
+                       __intel_lr_context_unpin(
+                                       ring,
+                                       ctx);
+                       ctx->engine[ring->id].dirty = false;
+               }
        }
 }
index 0b821b91723a61a4766e761ef4975989e9eea61c..fd1b6b43bacc75007c1bd7f94794c479e5e69973 100644 (file)
@@ -91,6 +91,7 @@ void intel_lr_context_reset(struct drm_device *dev,
                        struct intel_context *ctx);
 uint64_t intel_lr_context_descriptor(struct intel_context *ctx,
                                     struct intel_engine_cs *ring);
+void intel_lr_context_complete_check(struct drm_i915_gem_request *req);
 
 /* Execlists */
 int intel_sanitize_enable_execlists(struct drm_device *dev, int enable_execlists);