drm/amdgpu: patch cond exec for SDMA

More ground work for conditional execution on SDMA
necessary for preemption.

Signed-off-by: Monk Liu <monk.liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

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

@@ -302,6 +302,8 @@ struct amdgpu_ring_funcs {
 	void (*insert_nop)(struct amdgpu_ring *ring, uint32_t count);
 	/* pad the indirect buffer to the necessary number of dw */
 	void (*pad_ib)(struct amdgpu_ring *ring, struct amdgpu_ib *ib);
+	unsigned (*init_cond_exec)(struct amdgpu_ring *ring);
+	void (*patch_cond_exec)(struct amdgpu_ring *ring, unsigned offset);
 };
 
 /*
@@ -2182,6 +2184,8 @@ amdgpu_get_sdma_instance(struct amdgpu_ring *ring)
 #define amdgpu_ring_emit_hdp_flush(r) (r)->funcs->emit_hdp_flush((r))
 #define amdgpu_ring_emit_hdp_invalidate(r) (r)->funcs->emit_hdp_invalidate((r))
 #define amdgpu_ring_pad_ib(r, ib) ((r)->funcs->pad_ib((r), (ib)))
+#define amdgpu_ring_init_cond_exec(r) (r)->funcs->init_cond_exec((r))
+#define amdgpu_ring_patch_cond_exec(r,o) (r)->funcs->patch_cond_exec((r),(o))
 #define amdgpu_ih_get_wptr(adev) (adev)->irq.ih_funcs->get_wptr((adev))
 #define amdgpu_ih_decode_iv(adev, iv) (adev)->irq.ih_funcs->decode_iv((adev), (iv))
 #define amdgpu_ih_set_rptr(adev) (adev)->irq.ih_funcs->set_rptr((adev))

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

@@ -124,7 +124,8 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
 	struct amdgpu_ctx *ctx, *old_ctx;
 	struct amdgpu_vm *vm;
 	struct fence *hwf;
-	unsigned i;
+	unsigned i, patch_offset = ~0;
+
 	int r = 0;
 
 	if (num_ibs == 0)
@@ -149,6 +150,9 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
 		return r;
 	}
 
+	if (ring->type == AMDGPU_RING_TYPE_SDMA && ring->funcs->init_cond_exec)
+		patch_offset = amdgpu_ring_init_cond_exec(ring);
+
 	if (vm) {
 		/* do context switch */
 		amdgpu_vm_flush(ring, ib->vm_id, ib->vm_pd_addr,
@@ -204,6 +208,9 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
 	if (f)
 		*f = fence_get(hwf);
 
+	if (patch_offset != ~0 && ring->funcs->patch_cond_exec)
+		amdgpu_ring_patch_cond_exec(ring, patch_offset);
+
 	amdgpu_ring_commit(ring);
 	return 0;
 }

drivers/gpu/drm/amd/amdgpu/sdma_v3_0.c

@@ -452,6 +452,31 @@ static void sdma_v3_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 se
 	amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0));
 }
 
+unsigned init_cond_exec(struct amdgpu_ring *ring)
+{
+	unsigned ret;
+	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_COND_EXE));
+	amdgpu_ring_write(ring, lower_32_bits(ring->cond_exe_gpu_addr));
+	amdgpu_ring_write(ring, upper_32_bits(ring->cond_exe_gpu_addr));
+	amdgpu_ring_write(ring, 1);
+	ret = ring->wptr;/* this is the offset we need patch later */
+	amdgpu_ring_write(ring, 0x55aa55aa);/* insert dummy here and patch it later */
+	return ret;
+}
+
+void patch_cond_exec(struct amdgpu_ring *ring, unsigned offset)
+{
+	unsigned cur;
+	BUG_ON(ring->ring[offset] != 0x55aa55aa);
+
+	cur = ring->wptr - 1;
+	if (likely(cur > offset))
+		ring->ring[offset] = cur - offset;
+	else
+		ring->ring[offset] = (ring->ring_size>>2) - offset + cur;
+}
+
+
 /**
  * sdma_v3_0_gfx_stop - stop the gfx async dma engines
  *