drm/i915: Copy across scheduler behaviour flags across submit fences
authorChris Wilson <chris@chris-wilson.co.uk>
Tue, 10 Dec 2019 15:13:32 +0000 (15:13 +0000)
committerChris Wilson <chris@chris-wilson.co.uk>
Wed, 11 Dec 2019 10:28:56 +0000 (10:28 +0000)
We want the bonded request to have the same scheduler properties as its
master so that it is placed at the same depth in the queue. For example,
consider we have requests A, B and B', where B & B' are a bonded pair to
run in parallel on two engines.

A -> B
           \- B'

B will run after A and so may be scheduled on an idle engine and wait on
A using a semaphore. B' sees B being executed and so enters the queue on
the same engine as A. As B' did not inherit the semaphore-chain from B,
it may have higher precedence than A and so preempts execution. However,
B' then sits on a semaphore waiting for B, who is waiting for A, who is
blocked by B.

Ergo B' needs to inherit the scheduler properties from B (i.e. the
semaphore chain) so that it is scheduled with the same priority as B and
will not be executed ahead of Bs dependencies.

Furthermore, to prevent the priorities changing via the expose fence on
B', we need to couple in the dependencies for PI. This requires us to
relax our sanity-checks that dependencies are strictly in order.

v2: Synchronise (B, B') execution on all platforms, regardless of using
a scheduler, any no-op syncs should be elided.

Fixes: ee1136908e9b ("drm/i915/execlists: Virtual engine bonding")
Closes: https://gitlab.freedesktop.org/drm/intel/issues/464
Testcase: igt/gem_exec_balancer/bonded-chain
Testcase: igt/gem_exec_balancer/bonded-semaphore
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191210151332.3902215-1-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/i915_request.c
drivers/gpu/drm/i915/i915_scheduler.c

index ddc6c311349c18072a404a2ecb93f63743ebcdf5..a6238c626a1658c3449d71f6a2c517327dfa2463 100644 (file)
@@ -300,11 +300,11 @@ void i915_request_retire_upto(struct i915_request *rq)
 }
 
 static int
-__i915_request_await_execution(struct i915_request *rq,
-                              struct i915_request *signal,
-                              void (*hook)(struct i915_request *rq,
-                                           struct dma_fence *signal),
-                              gfp_t gfp)
+__await_execution(struct i915_request *rq,
+                 struct i915_request *signal,
+                 void (*hook)(struct i915_request *rq,
+                              struct dma_fence *signal),
+                 gfp_t gfp)
 {
        struct execute_cb *cb;
 
@@ -341,6 +341,8 @@ __i915_request_await_execution(struct i915_request *rq,
        }
        spin_unlock_irq(&signal->lock);
 
+       /* Copy across semaphore status as we need the same behaviour */
+       rq->sched.flags |= signal->sched.flags;
        return 0;
 }
 
@@ -824,31 +826,21 @@ already_busywaiting(struct i915_request *rq)
 }
 
 static int
-emit_semaphore_wait(struct i915_request *to,
-                   struct i915_request *from,
-                   gfp_t gfp)
+__emit_semaphore_wait(struct i915_request *to,
+                     struct i915_request *from,
+                     u32 seqno)
 {
        const int has_token = INTEL_GEN(to->i915) >= 12;
        u32 hwsp_offset;
-       int len;
+       int len, err;
        u32 *cs;
 
        GEM_BUG_ON(INTEL_GEN(to->i915) < 8);
 
-       /* Just emit the first semaphore we see as request space is limited. */
-       if (already_busywaiting(to) & from->engine->mask)
-               goto await_fence;
-
-       if (i915_request_await_start(to, from) < 0)
-               goto await_fence;
-
-       /* Only submit our spinner after the signaler is running! */
-       if (__i915_request_await_execution(to, from, NULL, gfp))
-               goto await_fence;
-
        /* We need to pin the signaler's HWSP until we are finished reading. */
-       if (intel_timeline_read_hwsp(from, to, &hwsp_offset))
-               goto await_fence;
+       err = intel_timeline_read_hwsp(from, to, &hwsp_offset);
+       if (err)
+               return err;
 
        len = 4;
        if (has_token)
@@ -871,7 +863,7 @@ emit_semaphore_wait(struct i915_request *to,
                 MI_SEMAPHORE_POLL |
                 MI_SEMAPHORE_SAD_GTE_SDD) +
                has_token;
-       *cs++ = from->fence.seqno;
+       *cs++ = seqno;
        *cs++ = hwsp_offset;
        *cs++ = 0;
        if (has_token) {
@@ -880,6 +872,28 @@ emit_semaphore_wait(struct i915_request *to,
        }
 
        intel_ring_advance(to, cs);
+       return 0;
+}
+
+static int
+emit_semaphore_wait(struct i915_request *to,
+                   struct i915_request *from,
+                   gfp_t gfp)
+{
+       /* Just emit the first semaphore we see as request space is limited. */
+       if (already_busywaiting(to) & from->engine->mask)
+               goto await_fence;
+
+       if (i915_request_await_start(to, from) < 0)
+               goto await_fence;
+
+       /* Only submit our spinner after the signaler is running! */
+       if (__await_execution(to, from, NULL, gfp))
+               goto await_fence;
+
+       if (__emit_semaphore_wait(to, from, from->fence.seqno))
+               goto await_fence;
+
        to->sched.semaphores |= from->engine->mask;
        to->sched.flags |= I915_SCHED_HAS_SEMAPHORE_CHAIN;
        return 0;
@@ -995,6 +1009,57 @@ i915_request_await_dma_fence(struct i915_request *rq, struct dma_fence *fence)
        return 0;
 }
 
+static bool intel_timeline_sync_has_start(struct intel_timeline *tl,
+                                         struct dma_fence *fence)
+{
+       return __intel_timeline_sync_is_later(tl,
+                                             fence->context,
+                                             fence->seqno - 1);
+}
+
+static int intel_timeline_sync_set_start(struct intel_timeline *tl,
+                                        const struct dma_fence *fence)
+{
+       return __intel_timeline_sync_set(tl, fence->context, fence->seqno - 1);
+}
+
+static int
+__i915_request_await_execution(struct i915_request *to,
+                              struct i915_request *from,
+                              void (*hook)(struct i915_request *rq,
+                                           struct dma_fence *signal))
+{
+       int err;
+
+       /* Submit both requests at the same time */
+       err = __await_execution(to, from, hook, I915_FENCE_GFP);
+       if (err)
+               return err;
+
+       /* Squash repeated depenendices to the same timelines */
+       if (intel_timeline_sync_has_start(i915_request_timeline(to),
+                                         &from->fence))
+               return 0;
+
+       /* Ensure both start together [after all semaphores in signal] */
+       if (intel_engine_has_semaphores(to->engine))
+               err = __emit_semaphore_wait(to, from, from->fence.seqno - 1);
+       else
+               err = i915_request_await_start(to, from);
+       if (err < 0)
+               return err;
+
+       /* Couple the dependency tree for PI on this exposed to->fence */
+       if (to->engine->schedule) {
+               err = i915_sched_node_add_dependency(&to->sched, &from->sched);
+               if (err < 0)
+                       return err;
+       }
+
+       return intel_timeline_sync_set_start(i915_request_timeline(to),
+                                            &from->fence);
+}
+
 int
 i915_request_await_execution(struct i915_request *rq,
                             struct dma_fence *fence,
@@ -1030,8 +1095,7 @@ i915_request_await_execution(struct i915_request *rq,
                if (dma_fence_is_i915(fence))
                        ret = __i915_request_await_execution(rq,
                                                             to_request(fence),
-                                                            hook,
-                                                            I915_FENCE_GFP);
+                                                            hook);
                else
                        ret = i915_sw_fence_await_dma_fence(&rq->submit, fence,
                                                            I915_FENCE_TIMEOUT,
index 1937a26d412ffb125746d84c14dd75230f5468f5..2bc2aa46a1b92d32123af6c989ae7d916fb80a01 100644 (file)
@@ -484,7 +484,6 @@ void i915_sched_node_fini(struct i915_sched_node *node)
         * so we may be called out-of-order.
         */
        list_for_each_entry_safe(dep, tmp, &node->signalers_list, signal_link) {
-               GEM_BUG_ON(!node_signaled(dep->signaler));
                GEM_BUG_ON(!list_empty(&dep->dfs_link));
 
                list_del(&dep->wait_link);