drm/radeon: rework VMID handling

Move binding onto the ring, simplifying handling a bit.
Signed-off-by: Christian König <deathsimple@vodafone.de>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>

drivers/gpu/drm/radeon/ni.c

@@ -1497,14 +1497,6 @@ void cayman_vm_fini(struct radeon_device *rdev)
 {
 }
 
-int cayman_vm_bind(struct radeon_device *rdev, struct radeon_vm *vm, int id)
-{
-	WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR + (id << 2), 0);
-	WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR + (id << 2), vm->last_pfn);
-	WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (id << 2), vm->pt_gpu_addr >> 12);
-	return 0;
-}
-
 #define R600_PTE_VALID     (1 << 0)
 #define R600_PTE_SYSTEM    (1 << 1)
 #define R600_PTE_SNOOPED   (1 << 2)
@@ -1540,10 +1532,20 @@ void cayman_vm_set_page(struct radeon_device *rdev, struct radeon_vm *vm,
 void cayman_vm_flush(struct radeon_device *rdev, struct radeon_ib *ib)
 {
 	struct radeon_ring *ring = &rdev->ring[ib->ring];
+	struct radeon_vm *vm = ib->vm;
 
-	if (!ib->vm || ib->vm->id == -1)
+	if (vm == NULL)
 		return;
 
+	radeon_ring_write(ring, PACKET0(VM_CONTEXT0_PAGE_TABLE_START_ADDR + (vm->id << 2), 0));
+	radeon_ring_write(ring, 0);
+
+	radeon_ring_write(ring, PACKET0(VM_CONTEXT0_PAGE_TABLE_END_ADDR + (vm->id << 2), 0));
+	radeon_ring_write(ring, vm->last_pfn);
+
+	radeon_ring_write(ring, PACKET0(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (vm->id << 2), 0));
+	radeon_ring_write(ring, vm->pt_gpu_addr >> 12);
+
 	/* flush hdp cache */
 	radeon_ring_write(ring, PACKET0(HDP_MEM_COHERENCY_FLUSH_CNTL, 0));
 	radeon_ring_write(ring, 0x1);

drivers/gpu/drm/radeon/radeon.h

@@ -253,6 +253,22 @@ static inline struct radeon_fence *radeon_fence_later(struct radeon_fence *a,
 	}
 }
 
+static inline bool radeon_fence_is_earlier(struct radeon_fence *a,
+					   struct radeon_fence *b)
+{
+	if (!a) {
+		return false;
+	}
+
+	if (!b) {
+		return true;
+	}
+
+	BUG_ON(a->ring != b->ring);
+
+	return a->seq < b->seq;
+}
+
 /*
  * Tiling registers
  */
@@ -628,10 +644,13 @@ struct radeon_ring {
 /*
  * VM
  */
+
+#define RADEON_NUM_VM	16
+
 struct radeon_vm {
 	struct list_head		list;
 	struct list_head		va;
-	int				id;
+	unsigned			id;
 	unsigned			last_pfn;
 	u64				pt_gpu_addr;
 	u64				*pt;
@@ -646,7 +665,7 @@ struct radeon_vm {
 struct radeon_vm_manager {
 	struct mutex			lock;
 	struct list_head		lru_vm;
-	uint32_t			use_bitmap;
+	struct radeon_fence		*active[RADEON_NUM_VM];
 	struct radeon_sa_manager	sa_manager;
 	uint32_t			max_pfn;
 	/* number of VMIDs */
@@ -1117,7 +1136,6 @@ struct radeon_asic {
 	struct {
 		int (*init)(struct radeon_device *rdev);
 		void (*fini)(struct radeon_device *rdev);
-		int (*bind)(struct radeon_device *rdev, struct radeon_vm *vm, int id);
 		uint32_t (*page_flags)(struct radeon_device *rdev,
 				       struct radeon_vm *vm,
 				       uint32_t flags);
@@ -1734,7 +1752,6 @@ void radeon_ring_write(struct radeon_ring *ring, uint32_t v);
 #define radeon_gart_set_page(rdev, i, p) (rdev)->asic->gart.set_page((rdev), (i), (p))
 #define radeon_asic_vm_init(rdev) (rdev)->asic->vm.init((rdev))
 #define radeon_asic_vm_fini(rdev) (rdev)->asic->vm.fini((rdev))
-#define radeon_asic_vm_bind(rdev, v, id) (rdev)->asic->vm.bind((rdev), (v), (id))
 #define radeon_asic_vm_page_flags(rdev, v, flags) (rdev)->asic->vm.page_flags((rdev), (v), (flags))
 #define radeon_asic_vm_set_page(rdev, v, pfn, addr, flags) (rdev)->asic->vm.set_page((rdev), (v), (pfn), (addr), (flags))
 #define radeon_ring_start(rdev, r, cp) (rdev)->asic->ring[(r)].ring_start((rdev), (cp))
@@ -1817,6 +1834,11 @@ int radeon_vm_init(struct radeon_device *rdev, struct radeon_vm *vm);
 void radeon_vm_fini(struct radeon_device *rdev, struct radeon_vm *vm);
 int radeon_vm_bind(struct radeon_device *rdev, struct radeon_vm *vm);
 void radeon_vm_unbind(struct radeon_device *rdev, struct radeon_vm *vm);
+struct radeon_fence *radeon_vm_grab_id(struct radeon_device *rdev,
+				       struct radeon_vm *vm, int ring);
+void radeon_vm_fence(struct radeon_device *rdev,
+		     struct radeon_vm *vm,
+		     struct radeon_fence *fence);
 int radeon_vm_bo_update_pte(struct radeon_device *rdev,
 			    struct radeon_vm *vm,
 			    struct radeon_bo *bo,

drivers/gpu/drm/radeon/radeon_asic.c

@@ -1375,7 +1375,6 @@ static struct radeon_asic cayman_asic = {
 	.vm = {
 		.init = &cayman_vm_init,
 		.fini = &cayman_vm_fini,
-		.bind = &cayman_vm_bind,
 		.page_flags = &cayman_vm_page_flags,
 		.set_page = &cayman_vm_set_page,
 	},
@@ -1480,7 +1479,6 @@ static struct radeon_asic trinity_asic = {
 	.vm = {
 		.init = &cayman_vm_init,
 		.fini = &cayman_vm_fini,
-		.bind = &cayman_vm_bind,
 		.page_flags = &cayman_vm_page_flags,
 		.set_page = &cayman_vm_set_page,
 	},
@@ -1585,7 +1583,6 @@ static struct radeon_asic si_asic = {
 	.vm = {
 		.init = &si_vm_init,
 		.fini = &si_vm_fini,
-		.bind = &si_vm_bind,
 		.page_flags = &cayman_vm_page_flags,
 		.set_page = &cayman_vm_set_page,
 	},
@@ -1599,7 +1596,7 @@ static struct radeon_asic si_asic = {
 			.ring_test = &r600_ring_test,
 			.ib_test = &r600_ib_test,
 			.is_lockup = &si_gpu_is_lockup,
-			.vm_flush = &cayman_vm_flush,
+			.vm_flush = &si_vm_flush,
 		},
 		[CAYMAN_RING_TYPE_CP1_INDEX] = {
 			.ib_execute = &si_ring_ib_execute,
@@ -1610,7 +1607,7 @@ static struct radeon_asic si_asic = {
 			.ring_test = &r600_ring_test,
 			.ib_test = &r600_ib_test,
 			.is_lockup = &si_gpu_is_lockup,
-			.vm_flush = &cayman_vm_flush,
+			.vm_flush = &si_vm_flush,
 		},
 		[CAYMAN_RING_TYPE_CP2_INDEX] = {
 			.ib_execute = &si_ring_ib_execute,
@@ -1621,7 +1618,7 @@ static struct radeon_asic si_asic = {
 			.ring_test = &r600_ring_test,
 			.ib_test = &r600_ib_test,
 			.is_lockup = &si_gpu_is_lockup,
-			.vm_flush = &cayman_vm_flush,
+			.vm_flush = &si_vm_flush,
 		}
 	},
 	.irq = {

drivers/gpu/drm/radeon/radeon_asic.h

@@ -440,7 +440,6 @@ int cayman_asic_reset(struct radeon_device *rdev);
 void cayman_ring_ib_execute(struct radeon_device *rdev, struct radeon_ib *ib);
 int cayman_vm_init(struct radeon_device *rdev);
 void cayman_vm_fini(struct radeon_device *rdev);
-int cayman_vm_bind(struct radeon_device *rdev, struct radeon_vm *vm, int id);
 void cayman_vm_unbind(struct radeon_device *rdev, struct radeon_vm *vm);
 void cayman_vm_flush(struct radeon_device *rdev, struct radeon_ib *ib);
 uint32_t cayman_vm_page_flags(struct radeon_device *rdev,
@@ -470,8 +469,7 @@ int si_irq_set(struct radeon_device *rdev);
 int si_irq_process(struct radeon_device *rdev);
 int si_vm_init(struct radeon_device *rdev);
 void si_vm_fini(struct radeon_device *rdev);
-int si_vm_bind(struct radeon_device *rdev, struct radeon_vm *vm, int id);
-void si_vm_unbind(struct radeon_device *rdev, struct radeon_vm *vm);
+void si_vm_flush(struct radeon_device *rdev, struct radeon_ib *ib);
 int si_ib_parse(struct radeon_device *rdev, struct radeon_ib *ib);
 uint64_t si_get_gpu_clock(struct radeon_device *rdev);
 

drivers/gpu/drm/radeon/radeon_cs.c

@@ -485,6 +485,7 @@ static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev,
 	}
 	radeon_cs_sync_rings(parser);
 	radeon_cs_sync_to(parser, vm->last_flush);
+	radeon_cs_sync_to(parser, radeon_vm_grab_id(rdev, vm, parser->ring));
 
 	if ((rdev->family >= CHIP_TAHITI) &&
 	    (parser->chunk_const_ib_idx != -1)) {
@@ -493,13 +494,11 @@ static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev,
 		r = radeon_ib_schedule(rdev, &parser->ib, NULL);
 	}
 
-out:
 	if (!r) {
-		if (vm->fence) {
-			radeon_fence_unref(&vm->fence);
-		}
-		vm->fence = radeon_fence_ref(parser->ib.fence);
+		radeon_vm_fence(rdev, vm, parser->ib.fence);
 	}
+
+out:
 	mutex_unlock(&vm->mutex);
 	mutex_unlock(&rdev->vm_manager.lock);
 	return r;

drivers/gpu/drm/radeon/radeon_device.c

@@ -1018,7 +1018,6 @@ int radeon_device_init(struct radeon_device *rdev,
 		return r;
 	/* initialize vm here */
 	mutex_init(&rdev->vm_manager.lock);
-	rdev->vm_manager.use_bitmap = 1;
 	rdev->vm_manager.max_pfn = 1 << 20;
 	INIT_LIST_HEAD(&rdev->vm_manager.lru_vm);
 

drivers/gpu/drm/radeon/radeon_gart.c

@@ -437,7 +437,6 @@ int radeon_vm_manager_init(struct radeon_device *rdev)
 	int r;
 
 	if (!rdev->vm_manager.enabled) {
-		/* mark first vm as always in use, it's the system one */
 		/* allocate enough for 2 full VM pts */
 		r = radeon_sa_bo_manager_init(rdev, &rdev->vm_manager.sa_manager,
 					      rdev->vm_manager.max_pfn * 8 * 2,
@@ -461,7 +460,7 @@ int radeon_vm_manager_init(struct radeon_device *rdev)
 
 	/* restore page table */
 	list_for_each_entry(vm, &rdev->vm_manager.lru_vm, list) {
-		if (vm->id == -1)
+		if (vm->sa_bo == NULL)
 			continue;
 
 		list_for_each_entry(bo_va, &vm->va, vm_list) {
@@ -475,11 +474,6 @@ int radeon_vm_manager_init(struct radeon_device *rdev)
 				DRM_ERROR("Failed to update pte for vm %d!\n", vm->id);
 			}
 		}
-
-		r = radeon_asic_vm_bind(rdev, vm, vm->id);
-		if (r) {
-			DRM_ERROR("Failed to bind vm %d!\n", vm->id);
-		}
 	}
 	return 0;
 }
@@ -500,10 +494,6 @@ static void radeon_vm_unbind_locked(struct radeon_device *rdev,
 {
 	struct radeon_bo_va *bo_va;
 
-	if (vm->id == -1) {
-		return;
-	}
-
 	/* wait for vm use to end */
 	while (vm->fence) {
 		int r;
@@ -523,9 +513,7 @@ static void radeon_vm_unbind_locked(struct radeon_device *rdev,
 	radeon_fence_unref(&vm->last_flush);
 
 	/* hw unbind */
-	rdev->vm_manager.use_bitmap &= ~(1 << vm->id);
 	list_del_init(&vm->list);
-	vm->id = -1;
 	radeon_sa_bo_free(rdev, &vm->sa_bo, NULL);
 	vm->pt = NULL;
 
@@ -544,6 +532,7 @@ static void radeon_vm_unbind_locked(struct radeon_device *rdev,
 void radeon_vm_manager_fini(struct radeon_device *rdev)
 {
 	struct radeon_vm *vm, *tmp;
+	int i;
 
 	if (!rdev->vm_manager.enabled)
 		return;
@@ -553,6 +542,9 @@ void radeon_vm_manager_fini(struct radeon_device *rdev)
 	list_for_each_entry_safe(vm, tmp, &rdev->vm_manager.lru_vm, list) {
 		radeon_vm_unbind_locked(rdev, vm);
 	}
+	for (i = 0; i < RADEON_NUM_VM; ++i) {
+		radeon_fence_unref(&rdev->vm_manager.active[i]);
+	}
 	radeon_asic_vm_fini(rdev);
 	mutex_unlock(&rdev->vm_manager.lock);
 
@@ -593,14 +585,13 @@ void radeon_vm_unbind(struct radeon_device *rdev, struct radeon_vm *vm)
 int radeon_vm_bind(struct radeon_device *rdev, struct radeon_vm *vm)
 {
 	struct radeon_vm *vm_evict;
-	unsigned i;
-	int id = -1, r;
+	int r;
 
 	if (vm == NULL) {
 		return -EINVAL;
 	}
 
-	if (vm->id != -1) {
+	if (vm->sa_bo != NULL) {
 		/* update lru */
 		list_del_init(&vm->list);
 		list_add_tail(&vm->list, &rdev->vm_manager.lru_vm);
@@ -623,33 +614,86 @@ int radeon_vm_bind(struct radeon_device *rdev, struct radeon_vm *vm)
 	vm->pt_gpu_addr = radeon_sa_bo_gpu_addr(vm->sa_bo);
 	memset(vm->pt, 0, RADEON_GPU_PAGE_ALIGN(vm->last_pfn * 8));
 
-retry_id:
-	/* search for free vm */
-	for (i = 0; i < rdev->vm_manager.nvm; i++) {
-		if (!(rdev->vm_manager.use_bitmap & (1 << i))) {
-			id = i;
-			break;
+	list_add_tail(&vm->list, &rdev->vm_manager.lru_vm);
+	return radeon_vm_bo_update_pte(rdev, vm, rdev->ring_tmp_bo.bo,
+				       &rdev->ring_tmp_bo.bo->tbo.mem);
+}
+
+/**
+ * radeon_vm_grab_id - allocate the next free VMID
+ *
+ * @rdev: radeon_device pointer
+ * @vm: vm to allocate id for
+ * @ring: ring we want to submit job to
+ *
+ * Allocate an id for the vm (cayman+).
+ * Returns the fence we need to sync to (if any).
+ *
+ * Global and local mutex must be locked!
+ */
+struct radeon_fence *radeon_vm_grab_id(struct radeon_device *rdev,
+				       struct radeon_vm *vm, int ring)
+{
+	struct radeon_fence *best[RADEON_NUM_RINGS] = {};
+	unsigned choices[2] = {};
+	unsigned i;
+
+	/* check if the id is still valid */
+	if (vm->fence && vm->fence == rdev->vm_manager.active[vm->id])
+		return NULL;
+
+	/* we definately need to flush */
+	radeon_fence_unref(&vm->last_flush);
+
+	/* skip over VMID 0, since it is the system VM */
+	for (i = 1; i < rdev->vm_manager.nvm; ++i) {
+		struct radeon_fence *fence = rdev->vm_manager.active[i];
+
+		if (fence == NULL) {
+			/* found a free one */
+			vm->id = i;
+			return NULL;
+		}
+
+		if (radeon_fence_is_earlier(fence, best[fence->ring])) {
+			best[fence->ring] = fence;
+			choices[fence->ring == ring ? 0 : 1] = i;
 		}
-	}
-	/* evict vm if necessary */
-	if (id == -1) {
-		vm_evict = list_first_entry(&rdev->vm_manager.lru_vm, struct radeon_vm, list);
-		radeon_vm_unbind(rdev, vm_evict);
-		goto retry_id;
 	}
 
-	/* do hw bind */
-	r = radeon_asic_vm_bind(rdev, vm, id);
-	radeon_fence_unref(&vm->last_flush);
-	if (r) {
-		radeon_sa_bo_free(rdev, &vm->sa_bo, NULL);
-		return r;
+	for (i = 0; i < 2; ++i) {
+		if (choices[i]) {
+			vm->id = choices[i];
+			return rdev->vm_manager.active[choices[i]];
+		}
 	}
-	rdev->vm_manager.use_bitmap |= 1 << id;
-	vm->id = id;
-	list_add_tail(&vm->list, &rdev->vm_manager.lru_vm);
-	return radeon_vm_bo_update_pte(rdev, vm, rdev->ring_tmp_bo.bo,
-				       &rdev->ring_tmp_bo.bo->tbo.mem);
+
+	/* should never happen */
+	BUG();
+	return NULL;
+}
+
+/**
+ * radeon_vm_fence - remember fence for vm
+ *
+ * @rdev: radeon_device pointer
+ * @vm: vm we want to fence
+ * @fence: fence to remember
+ *
+ * Fence the vm (cayman+).
+ * Set the fence used to protect page table and id.
+ *
+ * Global and local mutex must be locked!
+ */
+void radeon_vm_fence(struct radeon_device *rdev,
+		     struct radeon_vm *vm,
+		     struct radeon_fence *fence)
+{
+	radeon_fence_unref(&rdev->vm_manager.active[vm->id]);
+	rdev->vm_manager.active[vm->id] = radeon_fence_ref(fence);
+
+	radeon_fence_unref(&vm->fence);
+	vm->fence = radeon_fence_ref(fence);
 }
 
 /* object have to be reserved */
@@ -806,7 +850,7 @@ int radeon_vm_bo_update_pte(struct radeon_device *rdev,
 	uint32_t flags;
 
 	/* nothing to do if vm isn't bound */
-	if (vm->id == -1)
+	if (vm->sa_bo == NULL)
 		return 0;
 
 	bo_va = radeon_bo_va(bo, vm);
@@ -928,7 +972,7 @@ int radeon_vm_init(struct radeon_device *rdev, struct radeon_vm *vm)
 {
 	int r;
 
-	vm->id = -1;
+	vm->id = 0;
 	vm->fence = NULL;
 	mutex_init(&vm->mutex);
 	INIT_LIST_HEAD(&vm->list);

drivers/gpu/drm/radeon/si.c

@@ -2789,14 +2789,30 @@ void si_vm_fini(struct radeon_device *rdev)
 {
 }
 
-int si_vm_bind(struct radeon_device *rdev, struct radeon_vm *vm, int id)
+void si_vm_flush(struct radeon_device *rdev, struct radeon_ib *ib)
 {
-	if (id < 8)
-		WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (id << 2), vm->pt_gpu_addr >> 12);
-	else
-		WREG32(VM_CONTEXT8_PAGE_TABLE_BASE_ADDR + ((id - 8) << 2),
-		       vm->pt_gpu_addr >> 12);
-	return 0;
+	struct radeon_ring *ring = &rdev->ring[ib->ring];
+	struct radeon_vm *vm = ib->vm;
+
+	if (vm == NULL)
+		return;
+
+	if (vm->id < 8) {
+		radeon_ring_write(ring, PACKET0(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR
+						+ (vm->id << 2), 0));
+	} else {
+		radeon_ring_write(ring, PACKET0(VM_CONTEXT8_PAGE_TABLE_BASE_ADDR
+						+ ((vm->id - 8) << 2), 0));
+	}
+	radeon_ring_write(ring, vm->pt_gpu_addr >> 12);
+
+	/* flush hdp cache */
+	radeon_ring_write(ring, PACKET0(HDP_MEM_COHERENCY_FLUSH_CNTL, 0));
+	radeon_ring_write(ring, 0x1);
+
+	/* bits 0-7 are the VM contexts0-7 */
+	radeon_ring_write(ring, PACKET0(VM_INVALIDATE_REQUEST, 0));
+	radeon_ring_write(ring, 1 << ib->vm->id);
 }
 
 /*