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 <Christian.Koenig@amd.com>
Cc: Luben Tuikov <Luben.Tuikov@amd.com>
Cc: Andrey Grodzovsky <Andrey.Grodzovsky@amd.com>
Cc: Michel Dänzer <michel@daenzer.net>
Signed-off-by: Jiadong.Zhu <Jiadong.Zhu@amd.com>
Acked-by: Luben Tuikov <luben.tuikov@amd.com>
Acked-by: Huang Rui <ray.huang@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

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);

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

@@ -29,6 +29,7 @@
 #include "amdgpu_ring.h"
 
 struct amdgpu_ring;
+
 /**
  * struct amdgpu_mux_entry - the entry recording software rings copying information.
  * @ring: the pointer to the software ring.
@@ -37,6 +38,7 @@ struct amdgpu_ring;
  * @sw_cptr: the position of the copy pointer in the sw ring.
  * @sw_rptr: the read pointer in software ring.
  * @sw_wptr: the write pointer in software ring.
+ * @list: list head for amdgpu_mux_chunk
  */
 struct amdgpu_mux_entry {
 	struct amdgpu_ring      *ring;
@@ -45,6 +47,7 @@ struct amdgpu_mux_entry {
 	u64                     sw_cptr;
 	u64                     sw_rptr;
 	u64                     sw_wptr;
+	struct list_head        list;
 };
 
 struct amdgpu_ring_mux {
@@ -55,6 +58,26 @@ struct amdgpu_ring_mux {
 	unsigned int            ring_entry_size;
 	/*the lock for copy data from different software rings*/
 	spinlock_t              lock;
+	bool                    s_resubmit;
+	uint32_t                seqno_to_resubmit;
+	u64                     wptr_resubmit;
+	struct timer_list       resubmit_timer;
+
+	bool                    pending_trailing_fence_signaled;
+};
+
+/**
+ * struct amdgpu_mux_chunk - save the location of indirect buffer's package on softare rings.
+ * @entry: the list entry.
+ * @sync_seq: the fence seqno related with the saved IB.
+ * @start:- start location on the software ring.
+ * @end:- end location on the software ring.
+ */
+struct amdgpu_mux_chunk {
+	struct list_head        entry;
+	uint32_t                sync_seq;
+	u64                     start;
+	u64                     end;
 };
 
 int amdgpu_ring_mux_init(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring,
@@ -64,15 +87,17 @@ int amdgpu_ring_mux_add_sw_ring(struct amdgpu_ring_mux *mux, struct amdgpu_ring
 void amdgpu_ring_mux_set_wptr(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring, u64 wptr);
 u64 amdgpu_ring_mux_get_wptr(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring);
 u64 amdgpu_ring_mux_get_rptr(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring);
+void amdgpu_ring_mux_start_ib(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring);
+void amdgpu_ring_mux_end_ib(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring);
+bool amdgpu_mcbp_handle_trailing_fence_irq(struct amdgpu_ring_mux *mux);
 
 u64 amdgpu_sw_ring_get_rptr_gfx(struct amdgpu_ring *ring);
 u64 amdgpu_sw_ring_get_wptr_gfx(struct amdgpu_ring *ring);
 void amdgpu_sw_ring_set_wptr_gfx(struct amdgpu_ring *ring);
-
 void amdgpu_sw_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count);
 void amdgpu_sw_ring_ib_begin(struct amdgpu_ring *ring);
 void amdgpu_sw_ring_ib_end(struct amdgpu_ring *ring);
-
 const char *amdgpu_sw_ring_name(int idx);
 unsigned int amdgpu_sw_ring_priority(int idx);
+
 #endif

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

@@ -541,6 +541,7 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job,
 	if (!vm_flush_needed && !gds_switch_needed && !need_pipe_sync)
 		return 0;
 
+	amdgpu_ring_ib_begin(ring);
 	if (ring->funcs->init_cond_exec)
 		patch_offset = amdgpu_ring_init_cond_exec(ring);
 
@@ -601,6 +602,7 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job,
 		amdgpu_ring_emit_switch_buffer(ring);
 		amdgpu_ring_emit_switch_buffer(ring);
 	}
+	amdgpu_ring_ib_end(ring);
 	return 0;
 }
 

drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c

@@ -5456,7 +5456,7 @@ static int gfx_v9_0_ring_preempt_ib(struct amdgpu_ring *ring)
 	ring->trail_seq += 1;
 	amdgpu_ring_alloc(ring, 13);
 	gfx_v9_0_ring_emit_fence(ring, ring->trail_fence_gpu_addr,
-				 ring->trail_seq, AMDGPU_FENCE_FLAG_EXEC);
+				 ring->trail_seq, AMDGPU_FENCE_FLAG_EXEC | AMDGPU_FENCE_FLAG_INT);
 	/*reset the CP_VMID_PREEMPT after trailing fence*/
 	amdgpu_ring_emit_wreg(ring,
 			      SOC15_REG_OFFSET(GC, 0, mmCP_VMID_PREEMPT),
@@ -5882,8 +5882,9 @@ static int gfx_v9_0_eop_irq(struct amdgpu_device *adev,
 
 	switch (me_id) {
 	case 0:
-		/* Fence signals are handled on the software rings*/
-		if (adev->gfx.num_gfx_rings) {
+		if (adev->gfx.num_gfx_rings &&
+		    !amdgpu_mcbp_handle_trailing_fence_irq(&adev->gfx.muxer)) {
+			/* Fence signals are handled on the software rings*/
 			for (i = 0; i < GFX9_NUM_SW_GFX_RINGS; i++)
 				amdgpu_fence_process(&adev->gfx.sw_gfx_ring[i]);
 		}