drm/amdgpu: MCBP based on DRM scheduler (v9)

Trigger Mid-Command Buffer Preemption according to the priority of the software
rings and the hw fence signalling condition.

The muxer saves the locations of the indirect buffer frames from the software
ring together with the fence sequence number in its fifo queue, and pops out
those records when the fences are signalled. The locations are used to resubmit
packages in preemption scenarios by coping the chunks from the software ring.

v2: Update comment style.
v3: Fix conflict caused by previous modifications.
v4: Remove unnecessary prints.
v5: Fix corner cases for resubmission cases.
v6: Refactor functions for resubmission, calling fence_process in irq handler.
v7: Solve conflict for removing amdgpu_sw_ring.c.
v8: Add time threshold to judge if preemption request is needed.
v9: Correct comment spelling. Set fence emit timestamp before rsu assignment.

Cc: Christian Koenig <[email protected]>
Cc: Luben Tuikov <[email protected]>
Cc: Andrey Grodzovsky <[email protected]>
Cc: Michel Dänzer <[email protected]>
Signed-off-by: Jiadong.Zhu <[email protected]>
Acked-by: Luben Tuikov <[email protected]>
Acked-by: Huang Rui <[email protected]>
Acked-by: Christian König <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

Author: jiadozhu

drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c

@@ -55,6 +55,7 @@ struct amdgpu_fence {
 
 	/* RB, DMA, etc. */
 	struct amdgpu_ring		*ring;
+	ktime_t				start_timestamp;
 };
 
 static struct kmem_cache *amdgpu_fence_slab;
@@ -199,6 +200,8 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, struct amd
 		}
 	}
 
+	to_amdgpu_fence(fence)->start_timestamp = ktime_get();
+
 	/* This function can't be called concurrently anyway, otherwise
 	 * emitting the fence would mess up the hardware ring buffer.
 	 */
@@ -406,6 +409,57 @@ unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring)
 	return lower_32_bits(emitted);
 }
 
+/**
+ * amdgpu_fence_last_unsignaled_time_us - the time fence emitted until now
+ * @ring: ring the fence is associated with
+ *
+ * Find the earliest fence unsignaled until now, calculate the time delta
+ * between the time fence emitted and now.
+ */
+u64 amdgpu_fence_last_unsignaled_time_us(struct amdgpu_ring *ring)
+{
+	struct amdgpu_fence_driver *drv = &ring->fence_drv;
+	struct dma_fence *fence;
+	uint32_t last_seq, sync_seq;
+
+	last_seq = atomic_read(&ring->fence_drv.last_seq);
+	sync_seq = READ_ONCE(ring->fence_drv.sync_seq);
+	if (last_seq == sync_seq)
+		return 0;
+
+	++last_seq;
+	last_seq &= drv->num_fences_mask;
+	fence = drv->fences[last_seq];
+	if (!fence)
+		return 0;
+
+	return ktime_us_delta(ktime_get(),
+		to_amdgpu_fence(fence)->start_timestamp);
+}
+
+/**
+ * amdgpu_fence_update_start_timestamp - update the timestamp of the fence
+ * @ring: ring the fence is associated with
+ * @seq: the fence seq number to update.
+ * @timestamp: the start timestamp to update.
+ *
+ * The function called at the time the fence and related ib is about to
+ * resubmit to gpu in MCBP scenario. Thus we do not consider race condition
+ * with amdgpu_fence_process to modify the same fence.
+ */
+void amdgpu_fence_update_start_timestamp(struct amdgpu_ring *ring, uint32_t seq, ktime_t timestamp)
+{
+	struct amdgpu_fence_driver *drv = &ring->fence_drv;
+	struct dma_fence *fence;
+
+	seq &= drv->num_fences_mask;
+	fence = drv->fences[seq];
+	if (!fence)
+		return;
+
+	to_amdgpu_fence(fence)->start_timestamp = timestamp;
+}
+
 /**
  * amdgpu_fence_driver_start_ring - make the fence driver
  * ready for use on the requested ring.

drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c

@@ -211,6 +211,7 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
 		}
 	}
 
+	amdgpu_ring_ib_begin(ring);
 	if (job && ring->funcs->init_cond_exec)
 		patch_offset = amdgpu_ring_init_cond_exec(ring);
 
@@ -285,6 +286,7 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
 	    ring->hw_prio == AMDGPU_GFX_PIPE_PRIO_HIGH)
 		ring->funcs->emit_wave_limit(ring, false);
 
+	amdgpu_ring_ib_end(ring);
 	amdgpu_ring_commit(ring);
 	return 0;
 }

drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c

@@ -569,3 +569,15 @@ int amdgpu_ring_init_mqd(struct amdgpu_ring *ring)
 
 	return mqd_mgr->init_mqd(adev, ring->mqd_ptr, &prop);
 }
+
+void amdgpu_ring_ib_begin(struct amdgpu_ring *ring)
+{
+	if (ring->is_sw_ring)
+		amdgpu_sw_ring_ib_begin(ring);
+}
+
+void amdgpu_ring_ib_end(struct amdgpu_ring *ring)
+{
+	if (ring->is_sw_ring)
+		amdgpu_sw_ring_ib_end(ring);
+}

drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h

@@ -145,8 +145,13 @@ signed long amdgpu_fence_wait_polling(struct amdgpu_ring *ring,
 				      uint32_t wait_seq,
 				      signed long timeout);
 unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring);
+
 void amdgpu_fence_driver_isr_toggle(struct amdgpu_device *adev, bool stop);
 
+u64 amdgpu_fence_last_unsignaled_time_us(struct amdgpu_ring *ring);
+void amdgpu_fence_update_start_timestamp(struct amdgpu_ring *ring, uint32_t seq,
+					 ktime_t timestamp);
+
 /*
  * Rings.
  */
@@ -313,6 +318,9 @@ struct amdgpu_ring {
 #define amdgpu_ring_preempt_ib(r) (r)->funcs->preempt_ib(r)
 
 int amdgpu_ring_alloc(struct amdgpu_ring *ring, unsigned ndw);
+void amdgpu_ring_ib_begin(struct amdgpu_ring *ring);
+void amdgpu_ring_ib_end(struct amdgpu_ring *ring);
+
 void amdgpu_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count);
 void amdgpu_ring_generic_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib);
 void amdgpu_ring_commit(struct amdgpu_ring *ring);