]> git.ipfire.org Git - thirdparty/linux.git/commitdiff
drm/amdgpu: restructure VM state machine v4
authorChristian König <christian.koenig@amd.com>
Tue, 20 Jan 2026 12:09:52 +0000 (13:09 +0100)
committerAlex Deucher <alexander.deucher@amd.com>
Tue, 19 May 2026 15:54:32 +0000 (11:54 -0400)
Instead of coming up with more sophisticated names for states a VM BO
can be in, group them by the type of BO first and then by the state.

So we end with BO type kernel, always_valid and individual and then states
evicted, moved and idle.

Not much functional change, except that evicted_user is moved back
together with the other BOs again which makes the handling in
amdgpu_vm_validate() a bit more complex.

Also fixes a problem with user queues and amdgpu_vm_ready(). We didn't
considered the VM ready when user BOs were not ideally placed, harmless
performance impact for kernel queues but a complete show stopper for
userqueues.

v2: fix a few typos in comments, rename the BO types to make them more
    descriptive, fix a couple of bugs found during testing
v3: squashed together with revert to old status lock handling, looks
    like the first patch still had some bug which this one here should fix.
    Fix a missing lock around debugfs printing.
v4: fix merge clash pointed out by Prike

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Sunil Khatri <sunil.khatri@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c

index f070ea37d9188def17f2bcc928966be76650a409..b634defb0be8bebd8dfde5c9d9e30f34e6211aab 100644 (file)
@@ -1028,12 +1028,12 @@ amdgpu_userq_bo_validate(struct amdgpu_device *adev, struct drm_exec *exec,
        struct amdgpu_bo *bo;
        int ret;
 
-       spin_lock(&vm->status_lock);
-       while (!list_empty(&vm->invalidated)) {
-               bo_va = list_first_entry(&vm->invalidated,
+       spin_lock(&vm->individual_lock);
+       while (!list_empty(&vm->always_valid.evicted)) {
+               bo_va = list_first_entry(&vm->always_valid.evicted,
                                         struct amdgpu_bo_va,
                                         base.vm_status);
-               spin_unlock(&vm->status_lock);
+               spin_unlock(&vm->individual_lock);
 
                bo = bo_va->base.bo;
                ret = drm_exec_prepare_obj(exec, &bo->tbo.base, 2);
@@ -1045,14 +1045,14 @@ amdgpu_userq_bo_validate(struct amdgpu_device *adev, struct drm_exec *exec,
                if (ret)
                        return ret;
 
-               /* This moves the bo_va to the done list */
+               /* This moves the bo_va to the idle list */
                ret = amdgpu_vm_bo_update(adev, bo_va, false);
                if (ret)
                        return ret;
 
-               spin_lock(&vm->status_lock);
+               spin_lock(&vm->individual_lock);
        }
-       spin_unlock(&vm->status_lock);
+       spin_unlock(&vm->individual_lock);
 
        return 0;
 }
@@ -1084,7 +1084,7 @@ retry_lock:
                if (unlikely(ret))
                        goto unlock_all;
 
-               ret = amdgpu_vm_lock_done_list(vm, &exec, 1);
+               ret = amdgpu_vm_lock_individual(vm, &exec, 1);
                drm_exec_retry_on_contention(&exec);
                if (unlikely(ret))
                        goto unlock_all;
@@ -1127,7 +1127,7 @@ retry_lock:
 
        key = 0;
        /* Validate User Ptr BOs */
-       list_for_each_entry(bo_va, &vm->done, base.vm_status) {
+       list_for_each_entry(bo_va, &vm->always_valid.idle, base.vm_status) {
                bo = bo_va->base.bo;
                if (!bo)
                        continue;
@@ -1177,10 +1177,10 @@ retry_lock:
 
        /*
         * We need to wait for all VM updates to finish before restarting the
-        * queues. Using the done list like that is now ok since everything is
+        * queues. Using the idle list like that is now ok since everything is
         * locked in place.
         */
-       list_for_each_entry(bo_va, &vm->done, base.vm_status)
+       list_for_each_entry(bo_va, &vm->always_valid.idle, base.vm_status)
                dma_fence_wait(bo_va->last_pt_update, false);
        dma_fence_wait(vm->last_update, false);
 
index 58437001e05a384c5afedf97b356ca41f1fd3be9..a99d01f9dc7503d49ef7b013c359b0595ec2f6ec 100644 (file)
@@ -138,6 +138,47 @@ static void amdgpu_vm_assert_locked(struct amdgpu_vm *vm)
        dma_resv_assert_held(vm->root.bo->tbo.base.resv);
 }
 
+/* Initialize the amdgpu_vm_bo_status object */
+static void amdgpu_vm_bo_status_init(struct amdgpu_vm_bo_status *lists)
+{
+       INIT_LIST_HEAD(&lists->evicted);
+       INIT_LIST_HEAD(&lists->moved);
+       INIT_LIST_HEAD(&lists->idle);
+}
+
+/*
+ * Make sure we have the lock to modify the vm_bo status and return the object
+ * with the status lists.
+ */
+static struct amdgpu_vm_bo_status *
+amdgpu_vm_bo_lock_lists(struct amdgpu_vm_bo_base *vm_bo)
+{
+       struct amdgpu_vm *vm = vm_bo->vm;
+       struct amdgpu_bo *bo = vm_bo->bo;
+
+       if (amdgpu_vm_is_bo_always_valid(vm, bo)) {
+               /* No extra locking needed, protected by the root PD resv lock */
+               amdgpu_vm_assert_locked(vm);
+
+               if (bo->tbo.type == ttm_bo_type_kernel)
+                       return &vm->kernel;
+
+               return &vm->always_valid;
+       }
+
+       spin_lock(&vm_bo->vm->individual_lock);
+       return &vm->individual;
+}
+
+/* Eventually unlock the status list lock again */
+static void amdgpu_vm_bo_unlock_lists(struct amdgpu_vm_bo_base *vm_bo)
+{
+       if (amdgpu_vm_is_bo_always_valid(vm_bo->vm, vm_bo->bo))
+               amdgpu_vm_assert_locked(vm_bo->vm);
+       else
+               spin_unlock(&vm_bo->vm->individual_lock);
+}
+
 /**
  * amdgpu_vm_is_bo_always_valid - check if the BO is VM always valid
  *
@@ -157,37 +198,44 @@ bool amdgpu_vm_is_bo_always_valid(struct amdgpu_vm *vm, struct amdgpu_bo *bo)
  *
  * @vm_bo: vm_bo which is evicted
  *
- * State for PDs/PTs and per VM BOs which are not at the location they should
- * be.
+ * State for vm_bo objects meaning the underlying BO was evicted and need to
+ * move in place again.
  */
 static void amdgpu_vm_bo_evicted(struct amdgpu_vm_bo_base *vm_bo)
 {
-       struct amdgpu_vm *vm = vm_bo->vm;
-       struct amdgpu_bo *bo = vm_bo->bo;
+       struct amdgpu_vm_bo_status *lists;
 
+       lists = amdgpu_vm_bo_lock_lists(vm_bo);
        vm_bo->moved = true;
-       amdgpu_vm_assert_locked(vm);
-       spin_lock(&vm_bo->vm->status_lock);
-       if (bo->tbo.type == ttm_bo_type_kernel)
-               list_move(&vm_bo->vm_status, &vm->evicted);
-       else
-               list_move_tail(&vm_bo->vm_status, &vm->evicted);
-       spin_unlock(&vm_bo->vm->status_lock);
+       list_move(&vm_bo->vm_status, &lists->evicted);
+       amdgpu_vm_bo_unlock_lists(vm_bo);
 }
 /**
  * amdgpu_vm_bo_moved - vm_bo is moved
  *
  * @vm_bo: vm_bo which is moved
  *
- * State for per VM BOs which are moved, but that change is not yet reflected
- * in the page tables.
+ * State for vm_bo objects meaning the underlying BO was moved but the new
+ * location not yet reflected in the page tables.
  */
 static void amdgpu_vm_bo_moved(struct amdgpu_vm_bo_base *vm_bo)
 {
-       amdgpu_vm_assert_locked(vm_bo->vm);
-       spin_lock(&vm_bo->vm->status_lock);
-       list_move(&vm_bo->vm_status, &vm_bo->vm->moved);
-       spin_unlock(&vm_bo->vm->status_lock);
+       struct amdgpu_vm_bo_status *lists;
+       struct amdgpu_bo *bo = vm_bo->bo;
+
+       /*
+        * The root PD doesn't have a parent PDE and goes directly into the
+        * idle state.
+        */
+       lists = amdgpu_vm_bo_lock_lists(vm_bo);
+       if (bo && bo->tbo.type == ttm_bo_type_kernel && !bo->parent) {
+               vm_bo->moved = false;
+               list_move(&vm_bo->vm_status, &lists->idle);
+       } else {
+               vm_bo->moved = true;
+               list_move(&vm_bo->vm_status, &lists->moved);
+       }
+       amdgpu_vm_bo_unlock_lists(vm_bo);
 }
 
 /**
@@ -195,120 +243,43 @@ static void amdgpu_vm_bo_moved(struct amdgpu_vm_bo_base *vm_bo)
  *
  * @vm_bo: vm_bo which is now idle
  *
- * State for PDs/PTs and per VM BOs which have gone through the state machine
- * and are now idle.
+ * State for vm_bo objects meaning we are done with the state machine and no
+ * further action is necessary.
  */
 static void amdgpu_vm_bo_idle(struct amdgpu_vm_bo_base *vm_bo)
 {
-       amdgpu_vm_assert_locked(vm_bo->vm);
-       spin_lock(&vm_bo->vm->status_lock);
-       list_move(&vm_bo->vm_status, &vm_bo->vm->idle);
-       spin_unlock(&vm_bo->vm->status_lock);
-       vm_bo->moved = false;
-}
+       struct amdgpu_vm_bo_status *lists;
 
-/**
- * amdgpu_vm_bo_invalidated - vm_bo is invalidated
- *
- * @vm_bo: vm_bo which is now invalidated
- *
- * State for normal BOs which are invalidated and that change not yet reflected
- * in the PTs.
- */
-static void amdgpu_vm_bo_invalidated(struct amdgpu_vm_bo_base *vm_bo)
-{
-       spin_lock(&vm_bo->vm->status_lock);
-       list_move(&vm_bo->vm_status, &vm_bo->vm->invalidated);
-       spin_unlock(&vm_bo->vm->status_lock);
-}
-
-/**
- * amdgpu_vm_bo_evicted_user - vm_bo is evicted
- *
- * @vm_bo: vm_bo which is evicted
- *
- * State for BOs used by user mode queues which are not at the location they
- * should be.
- */
-static void amdgpu_vm_bo_evicted_user(struct amdgpu_vm_bo_base *vm_bo)
-{
-       vm_bo->moved = true;
-       spin_lock(&vm_bo->vm->status_lock);
-       list_move(&vm_bo->vm_status, &vm_bo->vm->evicted_user);
-       spin_unlock(&vm_bo->vm->status_lock);
-}
-
-/**
- * amdgpu_vm_bo_relocated - vm_bo is reloacted
- *
- * @vm_bo: vm_bo which is relocated
- *
- * State for PDs/PTs which needs to update their parent PD.
- * For the root PD, just move to idle state.
- */
-static void amdgpu_vm_bo_relocated(struct amdgpu_vm_bo_base *vm_bo)
-{
-       amdgpu_vm_assert_locked(vm_bo->vm);
-       if (vm_bo->bo->parent) {
-               spin_lock(&vm_bo->vm->status_lock);
-               list_move(&vm_bo->vm_status, &vm_bo->vm->relocated);
-               spin_unlock(&vm_bo->vm->status_lock);
-       } else {
-               amdgpu_vm_bo_idle(vm_bo);
-       }
-}
-
-/**
- * amdgpu_vm_bo_done - vm_bo is done
- *
- * @vm_bo: vm_bo which is now done
- *
- * State for normal BOs which are invalidated and that change has been updated
- * in the PTs.
- */
-static void amdgpu_vm_bo_done(struct amdgpu_vm_bo_base *vm_bo)
-{
-       amdgpu_vm_assert_locked(vm_bo->vm);
-       spin_lock(&vm_bo->vm->status_lock);
-       list_move(&vm_bo->vm_status, &vm_bo->vm->done);
-       spin_unlock(&vm_bo->vm->status_lock);
+       lists = amdgpu_vm_bo_lock_lists(vm_bo);
+       if (!amdgpu_vm_is_bo_always_valid(vm_bo->vm, vm_bo->bo))
+               vm_bo->moved = false;
+       list_move(&vm_bo->vm_status, &lists->idle);
+       amdgpu_vm_bo_unlock_lists(vm_bo);
 }
 
 /**
  * amdgpu_vm_bo_reset_state_machine - reset the vm_bo state machine
  * @vm: the VM which state machine to reset
  *
- * Move all vm_bo object in the VM into a state where they will be updated
- * again during validation.
+ * Move all vm_bo object in the VM into a state where their location will be
+ * updated in the page tables again.
  */
 static void amdgpu_vm_bo_reset_state_machine(struct amdgpu_vm *vm)
 {
-       struct amdgpu_vm_bo_base *vm_bo, *tmp;
-
        amdgpu_vm_assert_locked(vm);
+       list_splice_init(&vm->kernel.idle, &vm->kernel.moved);
+       list_splice_init(&vm->always_valid.idle, &vm->always_valid.moved);
 
-       spin_lock(&vm->status_lock);
-       list_splice_init(&vm->done, &vm->invalidated);
-       list_for_each_entry(vm_bo, &vm->invalidated, vm_status)
-               vm_bo->moved = true;
-
-       list_for_each_entry_safe(vm_bo, tmp, &vm->idle, vm_status) {
-               struct amdgpu_bo *bo = vm_bo->bo;
-
-               vm_bo->moved = true;
-               if (!bo || bo->tbo.type != ttm_bo_type_kernel)
-                       list_move(&vm_bo->vm_status, &vm_bo->vm->moved);
-               else if (bo->parent)
-                       list_move(&vm_bo->vm_status, &vm_bo->vm->relocated);
-       }
-       spin_unlock(&vm->status_lock);
+       spin_lock(&vm->individual_lock);
+       list_splice_init(&vm->individual.idle, &vm->individual.moved);
+       spin_unlock(&vm->individual_lock);
 }
 
 /**
  * amdgpu_vm_update_shared - helper to update shared memory stat
  * @base: base structure for tracking BO usage in a VM
  *
- * Takes the vm status_lock and updates the shared memory stat. If the basic
+ * Takes the vm stats_lock and updates the shared memory stat. If the basic
  * stat changed (e.g. buffer was moved) amdgpu_vm_update_stats need to be called
  * as well.
  */
@@ -321,7 +292,7 @@ static void amdgpu_vm_update_shared(struct amdgpu_vm_bo_base *base)
        bool shared;
 
        dma_resv_assert_held(bo->tbo.base.resv);
-       spin_lock(&vm->status_lock);
+       spin_lock(&vm->stats_lock);
        shared = drm_gem_object_is_shared_for_memory_stats(&bo->tbo.base);
        if (base->shared != shared) {
                base->shared = shared;
@@ -333,7 +304,7 @@ static void amdgpu_vm_update_shared(struct amdgpu_vm_bo_base *base)
                        vm->stats[bo_memtype].drm.private += size;
                }
        }
-       spin_unlock(&vm->status_lock);
+       spin_unlock(&vm->stats_lock);
 }
 
 /**
@@ -358,11 +329,11 @@ void amdgpu_vm_bo_update_shared(struct amdgpu_bo *bo)
  *        be bo->tbo.resource
  * @sign: if we should add (+1) or subtract (-1) from the stat
  *
- * Caller need to have the vm status_lock held. Useful for when multiple update
+ * Caller need to have the vm stats_lock held. Useful for when multiple update
  * need to happen at the same time.
  */
 static void amdgpu_vm_update_stats_locked(struct amdgpu_vm_bo_base *base,
-                           struct ttm_resource *res, int sign)
+                                         struct ttm_resource *res, int sign)
 {
        struct amdgpu_vm *vm = base->vm;
        struct amdgpu_bo *bo = base->bo;
@@ -386,7 +357,8 @@ static void amdgpu_vm_update_stats_locked(struct amdgpu_vm_bo_base *base,
                 */
                if (bo->flags & AMDGPU_GEM_CREATE_DISCARDABLE)
                        vm->stats[res_memtype].drm.purgeable += size;
-               if (!(bo->preferred_domains & amdgpu_mem_type_to_domain(res_memtype)))
+               if (!(bo->preferred_domains &
+                     amdgpu_mem_type_to_domain(res_memtype)))
                        vm->stats[bo_memtype].evicted += size;
        }
 }
@@ -405,9 +377,9 @@ void amdgpu_vm_update_stats(struct amdgpu_vm_bo_base *base,
 {
        struct amdgpu_vm *vm = base->vm;
 
-       spin_lock(&vm->status_lock);
+       spin_lock(&vm->stats_lock);
        amdgpu_vm_update_stats_locked(base, res, sign);
-       spin_unlock(&vm->status_lock);
+       spin_unlock(&vm->stats_lock);
 }
 
 /**
@@ -428,37 +400,34 @@ void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base,
        base->next = NULL;
        INIT_LIST_HEAD(&base->vm_status);
 
+       dma_resv_assert_held(vm->root.bo->tbo.base.resv);
        if (!bo)
                return;
+
        base->next = bo->vm_bo;
        bo->vm_bo = base;
 
-       spin_lock(&vm->status_lock);
+       spin_lock(&vm->stats_lock);
        base->shared = drm_gem_object_is_shared_for_memory_stats(&bo->tbo.base);
        amdgpu_vm_update_stats_locked(base, bo->tbo.resource, +1);
-       spin_unlock(&vm->status_lock);
+       spin_unlock(&vm->stats_lock);
 
-       if (!amdgpu_vm_is_bo_always_valid(vm, bo))
+       if (!amdgpu_vm_is_bo_always_valid(vm, bo)) {
+               amdgpu_vm_bo_idle(base);
                return;
-
-       dma_resv_assert_held(vm->root.bo->tbo.base.resv);
+       }
 
        ttm_bo_set_bulk_move(&bo->tbo, &vm->lru_bulk_move);
-       if (bo->tbo.type == ttm_bo_type_kernel && bo->parent)
-               amdgpu_vm_bo_relocated(base);
-       else
-               amdgpu_vm_bo_idle(base);
 
+       /*
+        * When a per VM isn't in the desired domain put it into the evicted
+        * state to make sure that it gets validated on the next best occasion.
+        */
        if (bo->preferred_domains &
            amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type))
-               return;
-
-       /*
-        * we checked all the prerequisites, but it looks like this per vm bo
-        * is currently evicted. add the bo to the evicted list to make sure it
-        * is validated on next vm use to avoid fault.
-        * */
-       amdgpu_vm_bo_evicted(base);
+               amdgpu_vm_bo_moved(base);
+       else
+               amdgpu_vm_bo_evicted(base);
 }
 
 /**
@@ -479,41 +448,41 @@ int amdgpu_vm_lock_pd(struct amdgpu_vm *vm, struct drm_exec *exec,
 }
 
 /**
- * amdgpu_vm_lock_done_list - lock all BOs on the done list
+ * amdgpu_vm_lock_individual - lock all BOs on the individual idle list
  * @vm: vm providing the BOs
  * @exec: drm execution context
  * @num_fences: number of extra fences to reserve
  *
- * Lock the BOs on the done list in the DRM execution context.
+ * Lock the BOs on the individual idle list in the DRM execution context.
  */
-int amdgpu_vm_lock_done_list(struct amdgpu_vm *vm, struct drm_exec *exec,
-                            unsigned int num_fences)
+int amdgpu_vm_lock_individual(struct amdgpu_vm *vm, struct drm_exec *exec,
+                             unsigned int num_fences)
 {
-       struct list_head *prev = &vm->done;
+       struct list_head *prev = &vm->individual.idle;
        struct amdgpu_bo_va *bo_va;
        struct amdgpu_bo *bo;
        int ret;
 
        /* We can only trust prev->next while holding the lock */
-       spin_lock(&vm->status_lock);
-       while (!list_is_head(prev->next, &vm->done)) {
+       spin_lock(&vm->individual_lock);
+       while (!list_is_head(prev->next, &vm->individual.idle)) {
                bo_va = list_entry(prev->next, typeof(*bo_va), base.vm_status);
 
                bo = bo_va->base.bo;
                if (bo) {
                        amdgpu_bo_ref(bo);
-                       spin_unlock(&vm->status_lock);
+                       spin_unlock(&vm->individual_lock);
 
                        ret = drm_exec_prepare_obj(exec, &bo->tbo.base, 1);
                        amdgpu_bo_unref(&bo);
                        if (unlikely(ret))
                                return ret;
 
-                       spin_lock(&vm->status_lock);
+                       spin_lock(&vm->individual_lock);
                }
                prev = prev->next;
        }
-       spin_unlock(&vm->status_lock);
+       spin_unlock(&vm->individual_lock);
 
        return 0;
 }
@@ -609,10 +578,10 @@ int amdgpu_vm_validate(struct amdgpu_device *adev, struct amdgpu_vm *vm,
                       void *param)
 {
        uint64_t new_vm_generation = amdgpu_vm_generation(adev, vm);
-       struct amdgpu_vm_bo_base *bo_base;
-       struct amdgpu_bo *bo;
+       struct amdgpu_vm_bo_base *bo_base, *tmp;
        int r;
 
+       dma_resv_assert_held(vm->root.bo->tbo.base.resv);
        if (vm->generation != new_vm_generation) {
                vm->generation = new_vm_generation;
                amdgpu_vm_bo_reset_state_machine(vm);
@@ -622,49 +591,59 @@ int amdgpu_vm_validate(struct amdgpu_device *adev, struct amdgpu_vm *vm,
                        return r;
        }
 
-       spin_lock(&vm->status_lock);
-       while (!list_empty(&vm->evicted)) {
-               bo_base = list_first_entry(&vm->evicted,
-                                          struct amdgpu_vm_bo_base,
-                                          vm_status);
-               spin_unlock(&vm->status_lock);
+       list_for_each_entry_safe(bo_base, tmp, &vm->kernel.evicted, vm_status) {
+               r = validate(param, bo_base->bo);
+               if (r)
+                       return r;
 
-               bo = bo_base->bo;
+               vm->update_funcs->map_table(to_amdgpu_bo_vm(bo_base->bo));
+               amdgpu_vm_bo_moved(bo_base);
+       }
 
-               r = validate(param, bo);
+       /*
+        * As soon as all page tables are in place we can start updating them
+        * again.
+        */
+       amdgpu_vm_eviction_lock(vm);
+       vm->evicting = false;
+       amdgpu_vm_eviction_unlock(vm);
+
+       list_for_each_entry_safe(bo_base, tmp, &vm->always_valid.evicted,
+                                vm_status) {
+               r = validate(param, bo_base->bo);
                if (r)
                        return r;
 
-               if (bo->tbo.type != ttm_bo_type_kernel) {
-                       amdgpu_vm_bo_moved(bo_base);
-               } else {
-                       vm->update_funcs->map_table(to_amdgpu_bo_vm(bo));
-                       amdgpu_vm_bo_relocated(bo_base);
-               }
-               spin_lock(&vm->status_lock);
+               amdgpu_vm_bo_moved(bo_base);
        }
-       while (ticket && !list_empty(&vm->evicted_user)) {
-               bo_base = list_first_entry(&vm->evicted_user,
-                                          struct amdgpu_vm_bo_base,
-                                          vm_status);
-               spin_unlock(&vm->status_lock);
 
-               bo = bo_base->bo;
-               dma_resv_assert_held(bo->tbo.base.resv);
+       if (!ticket)
+               return 0;
+
+       spin_lock(&vm->individual_lock);
+restart:
+       list_for_each_entry(bo_base, &vm->individual.evicted, vm_status) {
+               struct amdgpu_bo *bo = bo_base->bo;
+
+               if (dma_resv_locking_ctx(bo->tbo.base.resv) != ticket)
+                       continue;
+
+               spin_unlock(&vm->individual_lock);
 
                r = validate(param, bo);
                if (r)
                        return r;
 
-               amdgpu_vm_bo_invalidated(bo_base);
+               amdgpu_vm_bo_moved(bo_base);
 
-               spin_lock(&vm->status_lock);
+               /* It's a bit inefficient to always jump back to the start, but
+                * we would need to re-structure the KFD for properly fixing
+                * that.
+                */
+               spin_lock(&vm->individual_lock);
+               goto restart;
        }
-       spin_unlock(&vm->status_lock);
-
-       amdgpu_vm_eviction_lock(vm);
-       vm->evicting = false;
-       amdgpu_vm_eviction_unlock(vm);
+       spin_unlock(&vm->individual_lock);
 
        return 0;
 }
@@ -689,9 +668,7 @@ bool amdgpu_vm_ready(struct amdgpu_vm *vm)
        ret = !vm->evicting;
        amdgpu_vm_eviction_unlock(vm);
 
-       spin_lock(&vm->status_lock);
-       ret &= list_empty(&vm->evicted);
-       spin_unlock(&vm->status_lock);
+       ret &= list_empty(&vm->kernel.evicted);
 
        spin_lock(&vm->immediate.lock);
        ret &= !vm->immediate.stopped;
@@ -985,18 +962,13 @@ int amdgpu_vm_update_pdes(struct amdgpu_device *adev,
                          struct amdgpu_vm *vm, bool immediate)
 {
        struct amdgpu_vm_update_params params;
-       struct amdgpu_vm_bo_base *entry;
+       struct amdgpu_vm_bo_base *entry, *tmp;
        bool flush_tlb_needed = false;
-       LIST_HEAD(relocated);
        int r, idx;
 
        amdgpu_vm_assert_locked(vm);
 
-       spin_lock(&vm->status_lock);
-       list_splice_init(&vm->relocated, &relocated);
-       spin_unlock(&vm->status_lock);
-
-       if (list_empty(&relocated))
+       if (list_empty(&vm->kernel.moved))
                return 0;
 
        if (!drm_dev_enter(adev_to_drm(adev), &idx))
@@ -1012,7 +984,7 @@ int amdgpu_vm_update_pdes(struct amdgpu_device *adev,
        if (r)
                goto error;
 
-       list_for_each_entry(entry, &relocated, vm_status) {
+       list_for_each_entry(entry, &vm->kernel.moved, vm_status) {
                /* vm_flush_needed after updating moved PDEs */
                flush_tlb_needed |= entry->moved;
 
@@ -1028,11 +1000,8 @@ int amdgpu_vm_update_pdes(struct amdgpu_device *adev,
        if (flush_tlb_needed)
                atomic64_inc(&vm->tlb_seq);
 
-       while (!list_empty(&relocated)) {
-               entry = list_first_entry(&relocated, struct amdgpu_vm_bo_base,
-                                        vm_status);
+       list_for_each_entry_safe(entry, tmp, &vm->kernel.moved, vm_status)
                amdgpu_vm_bo_idle(entry);
-       }
 
 error:
        drm_dev_exit(idx);
@@ -1260,9 +1229,9 @@ error_free:
 void amdgpu_vm_get_memory(struct amdgpu_vm *vm,
                          struct amdgpu_mem_stats stats[__AMDGPU_PL_NUM])
 {
-       spin_lock(&vm->status_lock);
+       spin_lock(&vm->stats_lock);
        memcpy(stats, vm->stats, sizeof(*stats) * __AMDGPU_PL_NUM);
-       spin_unlock(&vm->status_lock);
+       spin_unlock(&vm->stats_lock);
 }
 
 /**
@@ -1406,7 +1375,7 @@ int amdgpu_vm_bo_update(struct amdgpu_device *adev, struct amdgpu_bo_va *bo_va,
                else
                        amdgpu_vm_bo_idle(&bo_va->base);
        } else {
-               amdgpu_vm_bo_done(&bo_va->base);
+               amdgpu_vm_bo_idle(&bo_va->base);
        }
 
        list_splice_init(&bo_va->invalids, &bo_va->valids);
@@ -1629,29 +1598,25 @@ int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
                           struct amdgpu_vm *vm,
                           struct ww_acquire_ctx *ticket)
 {
-       struct amdgpu_bo_va *bo_va;
+       struct amdgpu_bo_va *bo_va, *tmp;
        struct dma_resv *resv;
        bool clear, unlock;
        int r;
 
-       spin_lock(&vm->status_lock);
-       while (!list_empty(&vm->moved)) {
-               bo_va = list_first_entry(&vm->moved, struct amdgpu_bo_va,
-                                        base.vm_status);
-               spin_unlock(&vm->status_lock);
-
+       list_for_each_entry_safe(bo_va, tmp, &vm->always_valid.moved,
+                                base.vm_status) {
                /* Per VM BOs never need to bo cleared in the page tables */
                r = amdgpu_vm_bo_update(adev, bo_va, false);
                if (r)
                        return r;
-               spin_lock(&vm->status_lock);
        }
 
-       while (!list_empty(&vm->invalidated)) {
-               bo_va = list_first_entry(&vm->invalidated, struct amdgpu_bo_va,
-                                        base.vm_status);
+       spin_lock(&vm->individual_lock);
+       while (!list_empty(&vm->individual.moved)) {
+               bo_va = list_first_entry(&vm->individual.moved,
+                                        typeof(*bo_va), base.vm_status);
                resv = bo_va->base.bo->tbo.base.resv;
-               spin_unlock(&vm->status_lock);
+               spin_unlock(&vm->individual_lock);
 
                /* Try to reserve the BO to avoid clearing its ptes */
                if (!adev->debug_vm && dma_resv_trylock(resv)) {
@@ -1681,11 +1646,11 @@ int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
                    drm_gem_is_imported(&bo_va->base.bo->tbo.base) &&
                    (!bo_va->base.bo->tbo.resource ||
                     bo_va->base.bo->tbo.resource->mem_type == TTM_PL_SYSTEM))
-                       amdgpu_vm_bo_evicted_user(&bo_va->base);
+                       amdgpu_vm_bo_evicted(&bo_va->base);
 
-               spin_lock(&vm->status_lock);
+               spin_lock(&vm->individual_lock);
        }
-       spin_unlock(&vm->status_lock);
+       spin_unlock(&vm->individual_lock);
 
        return 0;
 }
@@ -2223,9 +2188,9 @@ void amdgpu_vm_bo_del(struct amdgpu_device *adev,
                }
        }
 
-       spin_lock(&vm->status_lock);
+       spin_lock(&vm->individual_lock);
        list_del(&bo_va->base.vm_status);
-       spin_unlock(&vm->status_lock);
+       spin_unlock(&vm->individual_lock);
 
        list_for_each_entry_safe(mapping, next, &bo_va->valids, list) {
                list_del(&mapping->list);
@@ -2305,14 +2270,7 @@ void amdgpu_vm_bo_invalidate(struct amdgpu_bo *bo, bool evicted)
 
                if (bo_base->moved)
                        continue;
-               bo_base->moved = true;
-
-               if (bo->tbo.type == ttm_bo_type_kernel)
-                       amdgpu_vm_bo_relocated(bo_base);
-               else if (amdgpu_vm_is_bo_always_valid(vm, bo))
-                       amdgpu_vm_bo_moved(bo_base);
-               else
-                       amdgpu_vm_bo_invalidated(bo_base);
+               amdgpu_vm_bo_moved(bo_base);
        }
 }
 
@@ -2333,10 +2291,10 @@ void amdgpu_vm_bo_move(struct amdgpu_bo *bo, struct ttm_resource *new_mem,
        for (bo_base = bo->vm_bo; bo_base; bo_base = bo_base->next) {
                struct amdgpu_vm *vm = bo_base->vm;
 
-               spin_lock(&vm->status_lock);
+               spin_lock(&vm->stats_lock);
                amdgpu_vm_update_stats_locked(bo_base, bo->tbo.resource, -1);
                amdgpu_vm_update_stats_locked(bo_base, new_mem, +1);
-               spin_unlock(&vm->status_lock);
+               spin_unlock(&vm->stats_lock);
        }
 
        amdgpu_vm_bo_invalidate(bo, evicted);
@@ -2600,16 +2558,14 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm,
        vm->va = RB_ROOT_CACHED;
        for (i = 0; i < AMDGPU_MAX_VMHUBS; i++)
                vm->reserved_vmid[i] = NULL;
-       INIT_LIST_HEAD(&vm->evicted);
-       INIT_LIST_HEAD(&vm->evicted_user);
-       INIT_LIST_HEAD(&vm->relocated);
-       INIT_LIST_HEAD(&vm->moved);
-       INIT_LIST_HEAD(&vm->idle);
-       INIT_LIST_HEAD(&vm->invalidated);
-       spin_lock_init(&vm->status_lock);
+
+       amdgpu_vm_bo_status_init(&vm->kernel);
+       amdgpu_vm_bo_status_init(&vm->always_valid);
+       spin_lock_init(&vm->individual_lock);
+       amdgpu_vm_bo_status_init(&vm->individual);
        INIT_LIST_HEAD(&vm->freed);
-       INIT_LIST_HEAD(&vm->done);
        INIT_KFIFO(vm->faults);
+       spin_lock_init(&vm->stats_lock);
 
        r = amdgpu_vm_init_entities(adev, vm);
        if (r)
@@ -3086,100 +3042,64 @@ error_unlock:
 }
 
 #if defined(CONFIG_DEBUG_FS)
-/**
- * amdgpu_debugfs_vm_bo_info  - print BO info for the VM
- *
- * @vm: Requested VM for printing BO info
- * @m: debugfs file
- *
- * Print BO information in debugfs file for the VM
- */
-void amdgpu_debugfs_vm_bo_info(struct amdgpu_vm *vm, struct seq_file *m)
-{
-       struct amdgpu_bo_va *bo_va, *tmp;
-       u64 total_idle = 0;
-       u64 total_evicted = 0;
-       u64 total_relocated = 0;
-       u64 total_moved = 0;
-       u64 total_invalidated = 0;
-       u64 total_done = 0;
-       unsigned int total_idle_objs = 0;
-       unsigned int total_evicted_objs = 0;
-       unsigned int total_relocated_objs = 0;
-       unsigned int total_moved_objs = 0;
-       unsigned int total_invalidated_objs = 0;
-       unsigned int total_done_objs = 0;
-       unsigned int id = 0;
 
-       amdgpu_vm_assert_locked(vm);
+/* print the debug info for a specific set of status lists */
+static void amdgpu_debugfs_vm_bo_status_info(struct seq_file *m,
+                                            struct amdgpu_vm_bo_status *lists)
+{
+       struct amdgpu_vm_bo_base *base;
+       unsigned int id;
 
-       spin_lock(&vm->status_lock);
-       seq_puts(m, "\tIdle BOs:\n");
-       list_for_each_entry_safe(bo_va, tmp, &vm->idle, base.vm_status) {
-               if (!bo_va->base.bo)
-                       continue;
-               total_idle += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
-       }
-       total_idle_objs = id;
        id = 0;
-
        seq_puts(m, "\tEvicted BOs:\n");
-       list_for_each_entry_safe(bo_va, tmp, &vm->evicted, base.vm_status) {
-               if (!bo_va->base.bo)
+       list_for_each_entry(base, &lists->evicted, vm_status) {
+               if (!base->bo)
                        continue;
-               total_evicted += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
-       }
-       total_evicted_objs = id;
-       id = 0;
 
-       seq_puts(m, "\tRelocated BOs:\n");
-       list_for_each_entry_safe(bo_va, tmp, &vm->relocated, base.vm_status) {
-               if (!bo_va->base.bo)
-                       continue;
-               total_relocated += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
+               amdgpu_bo_print_info(id++, base->bo, m);
        }
-       total_relocated_objs = id;
-       id = 0;
 
+       id = 0;
        seq_puts(m, "\tMoved BOs:\n");
-       list_for_each_entry_safe(bo_va, tmp, &vm->moved, base.vm_status) {
-               if (!bo_va->base.bo)
+       list_for_each_entry(base, &lists->moved, vm_status) {
+               if (!base->bo)
                        continue;
-               total_moved += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
+
+               amdgpu_bo_print_info(id++, base->bo, m);
        }
-       total_moved_objs = id;
-       id = 0;
 
-       seq_puts(m, "\tInvalidated BOs:\n");
-       list_for_each_entry_safe(bo_va, tmp, &vm->invalidated, base.vm_status) {
-               if (!bo_va->base.bo)
+       id = 0;
+       seq_puts(m, "\tIdle BOs:\n");
+       list_for_each_entry(base, &lists->moved, vm_status) {
+               if (!base->bo)
                        continue;
-               total_invalidated += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
+
+               amdgpu_bo_print_info(id++, base->bo, m);
        }
-       total_invalidated_objs = id;
-       id = 0;
+}
 
-       seq_puts(m, "\tDone BOs:\n");
-       list_for_each_entry_safe(bo_va, tmp, &vm->done, base.vm_status) {
-               if (!bo_va->base.bo)
-                       continue;
-               total_done += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
-       }
-       spin_unlock(&vm->status_lock);
-       total_done_objs = id;
-
-       seq_printf(m, "\tTotal idle size:        %12lld\tobjs:\t%d\n", total_idle,
-                  total_idle_objs);
-       seq_printf(m, "\tTotal evicted size:     %12lld\tobjs:\t%d\n", total_evicted,
-                  total_evicted_objs);
-       seq_printf(m, "\tTotal relocated size:   %12lld\tobjs:\t%d\n", total_relocated,
-                  total_relocated_objs);
-       seq_printf(m, "\tTotal moved size:       %12lld\tobjs:\t%d\n", total_moved,
-                  total_moved_objs);
-       seq_printf(m, "\tTotal invalidated size: %12lld\tobjs:\t%d\n", total_invalidated,
-                  total_invalidated_objs);
-       seq_printf(m, "\tTotal done size:        %12lld\tobjs:\t%d\n", total_done,
-                  total_done_objs);
+/**
+ * amdgpu_debugfs_vm_bo_info  - print BO info for the VM
+ *
+ * @vm: Requested VM for printing BO info
+ * @m: debugfs file
+ *
+ * Print BO information in debugfs file for the VM
+ */
+void amdgpu_debugfs_vm_bo_info(struct amdgpu_vm *vm, struct seq_file *m)
+{
+       amdgpu_vm_assert_locked(vm);
+
+       seq_puts(m, "\tKernel PT/PDs:\n");
+       amdgpu_debugfs_vm_bo_status_info(m, &vm->kernel);
+
+       seq_puts(m, "\tPer VM BOs:\n");
+       amdgpu_debugfs_vm_bo_status_info(m, &vm->always_valid);
+
+       seq_puts(m, "\tIndividual BOs:\n");
+       spin_lock(&vm->individual_lock);
+       amdgpu_debugfs_vm_bo_status_info(m, &vm->individual);
+       spin_unlock(&vm->individual_lock);
 }
 #endif
 
index f4b0e9da3fa2c4595a473911f937239f66515e9c..3695299f1a0396ce630a240de5bce41c590b0759 100644 (file)
@@ -205,17 +205,34 @@ struct amdgpu_vm_bo_base {
        /* protected by bo being reserved */
        struct amdgpu_vm_bo_base        *next;
 
-       /* protected by vm status_lock */
+       /* protected by vm reservation and invalidated_lock */
        struct list_head                vm_status;
 
        /* if the bo is counted as shared in mem stats
-        * protected by vm status_lock */
+        * protected by vm BO being reserved */
        bool                            shared;
 
        /* protected by the BO being reserved */
        bool                            moved;
 };
 
+/*
+ * The following status lists contain amdgpu_vm_bo_base objects for
+ * either PD/PTs, per VM BOs or BOs with individual resv object.
+ *
+ * The state transits are: evicted -> moved -> idle
+ */
+struct amdgpu_vm_bo_status {
+       /* BOs evicted which need to move into place again */
+       struct list_head                evicted;
+
+       /* BOs which moved but new location hasn't been updated in the PDs/PTs */
+       struct list_head                moved;
+
+       /* BOs done with the state machine and need no further action */
+       struct list_head                idle;
+};
+
 /* provided by hw blocks that can write ptes, e.g., sdma */
 struct amdgpu_vm_pte_funcs {
        /* number of dw to reserve per operation */
@@ -345,47 +362,29 @@ struct amdgpu_vm {
        bool                    evicting;
        unsigned int            saved_flags;
 
-       /* Lock to protect vm_bo add/del/move on all lists of vm */
-       spinlock_t              status_lock;
-
-       /* Memory statistics for this vm, protected by status_lock */
+       /* Memory statistics for this vm, protected by stats_lock */
+       spinlock_t              stats_lock;
        struct amdgpu_mem_stats stats[__AMDGPU_PL_NUM];
 
+       /* BO's belonging to PD/PT which are internal to the kernel. */
+       struct amdgpu_vm_bo_status      kernel;
+
        /*
-        * The following lists contain amdgpu_vm_bo_base objects for either
-        * PDs, PTs or per VM BOs. The state transits are:
-        *
-        * evicted -> relocated (PDs, PTs) or moved (per VM BOs) -> idle
+        * BOs allocated by userspace where the dma_resv is shared with the
+        * root PD
         */
-
-       /* Per-VM and PT BOs who needs a validation */
-       struct list_head        evicted;
-
-       /* PT BOs which relocated and their parent need an update */
-       struct list_head        relocated;
-
-       /* per VM BOs moved, but not yet updated in the PT */
-       struct list_head        moved;
-
-       /* All BOs of this VM not currently in the state machine */
-       struct list_head        idle;
+       struct amdgpu_vm_bo_status      always_valid;
 
        /*
         * The following lists contain amdgpu_vm_bo_base objects for BOs which
-        * have their own dma_resv object and not depend on the root PD. Their
-        * state transits are:
+        * have their own dma_resv object and not depend on the root PD.
         *
-        * evicted_user or invalidated -> done
+        * Lists are protected by the individual_lock.
         */
+       spinlock_t                      individual_lock;
 
-       /* BOs for user mode queues that need a validation */
-       struct list_head        evicted_user;
-
-       /* regular invalidated BOs, but not yet updated in the PT */
-       struct list_head        invalidated;
-
-       /* BOs which are invalidated, has been updated in the PTs */
-       struct list_head        done;
+       /* Userspace BOs with individual resv object */
+       struct amdgpu_vm_bo_status      individual;
 
        /*
         * This list contains amdgpu_bo_va_mapping objects which have been freed
@@ -507,8 +506,8 @@ int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm);
 void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm);
 int amdgpu_vm_lock_pd(struct amdgpu_vm *vm, struct drm_exec *exec,
                      unsigned int num_fences);
-int amdgpu_vm_lock_done_list(struct amdgpu_vm *vm, struct drm_exec *exec,
-                            unsigned int num_fences);
+int amdgpu_vm_lock_individual(struct amdgpu_vm *vm, struct drm_exec *exec,
+                             unsigned int num_fences);
 bool amdgpu_vm_ready(struct amdgpu_vm *vm);
 uint64_t amdgpu_vm_generation(struct amdgpu_device *adev, struct amdgpu_vm *vm);
 int amdgpu_vm_validate(struct amdgpu_device *adev, struct amdgpu_vm *vm,
index b2494381b98a1056cdfbe22ae9a5afc937cb7343..e43a60d09808288a3620ec7417c5b5d4c609ea1d 100644 (file)
@@ -544,9 +544,7 @@ static void amdgpu_vm_pt_free(struct amdgpu_vm_bo_base *entry)
        entry->bo->vm_bo = NULL;
        ttm_bo_set_bulk_move(&entry->bo->tbo, NULL);
 
-       spin_lock(&entry->vm->status_lock);
        list_del(&entry->vm_status);
-       spin_unlock(&entry->vm->status_lock);
        amdgpu_bo_unref(&entry->bo);
 }
 
@@ -590,7 +588,6 @@ static void amdgpu_vm_pt_add_list(struct amdgpu_vm_update_params *params,
        struct amdgpu_vm_pt_cursor seek;
        struct amdgpu_vm_bo_base *entry;
 
-       spin_lock(&params->vm->status_lock);
        for_each_amdgpu_vm_pt_dfs_safe(params->adev, params->vm, cursor, seek, entry) {
                if (entry && entry->bo)
                        list_move(&entry->vm_status, &params->tlb_flush_waitlist);
@@ -598,7 +595,6 @@ static void amdgpu_vm_pt_add_list(struct amdgpu_vm_update_params *params,
 
        /* enter start node now */
        list_move(&cursor->entry->vm_status, &params->tlb_flush_waitlist);
-       spin_unlock(&params->vm->status_lock);
 }
 
 /**