drm/amdgpu: rework VM state machine lock handling v2
authorChristian König <christian.koenig@amd.com>
Thu, 19 Apr 2018 08:56:02 +0000 (10:56 +0200)
committerAlex Deucher <alexander.deucher@amd.com>
Thu, 24 May 2018 15:07:52 +0000 (10:07 -0500)
Only the moved state needs a separate spin lock protection. All other
states are protected by reserving the VM anyway.

v2: fix some more incorrect cases

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h

index 1a8f4e0dd02394a759ec191258a6cbee2d76ba0a..f0deedcaf1c93e7b7585ff21f12c64f90f0d8d91 100644 (file)
@@ -119,9 +119,7 @@ static void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base,
         * is currently evicted. add the bo to the evicted list to make sure it
         * is validated on next vm use to avoid fault.
         * */
-       spin_lock(&vm->status_lock);
        list_move_tail(&base->vm_status, &vm->evicted);
-       spin_unlock(&vm->status_lock);
 }
 
 /**
@@ -228,7 +226,6 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
        struct ttm_bo_global *glob = adev->mman.bdev.glob;
        int r;
 
-       spin_lock(&vm->status_lock);
        while (!list_empty(&vm->evicted)) {
                struct amdgpu_vm_bo_base *bo_base;
                struct amdgpu_bo *bo;
@@ -236,10 +233,8 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
                bo_base = list_first_entry(&vm->evicted,
                                           struct amdgpu_vm_bo_base,
                                           vm_status);
-               spin_unlock(&vm->status_lock);
 
                bo = bo_base->bo;
-               BUG_ON(!bo);
                if (bo->parent) {
                        r = validate(param, bo);
                        if (r)
@@ -259,13 +254,14 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
                                return r;
                }
 
-               spin_lock(&vm->status_lock);
-               if (bo->tbo.type != ttm_bo_type_kernel)
+               if (bo->tbo.type != ttm_bo_type_kernel) {
+                       spin_lock(&vm->moved_lock);
                        list_move(&bo_base->vm_status, &vm->moved);
-               else
+                       spin_unlock(&vm->moved_lock);
+               } else {
                        list_move(&bo_base->vm_status, &vm->relocated);
+               }
        }
-       spin_unlock(&vm->status_lock);
 
        return 0;
 }
@@ -279,13 +275,7 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
  */
 bool amdgpu_vm_ready(struct amdgpu_vm *vm)
 {
-       bool ready;
-
-       spin_lock(&vm->status_lock);
-       ready = list_empty(&vm->evicted);
-       spin_unlock(&vm->status_lock);
-
-       return ready;
+       return list_empty(&vm->evicted);
 }
 
 /**
@@ -477,9 +467,7 @@ static int amdgpu_vm_alloc_levels(struct amdgpu_device *adev,
                        pt->parent = amdgpu_bo_ref(parent->base.bo);
 
                        amdgpu_vm_bo_base_init(&entry->base, vm, pt);
-                       spin_lock(&vm->status_lock);
                        list_move(&entry->base.vm_status, &vm->relocated);
-                       spin_unlock(&vm->status_lock);
                }
 
                if (level < AMDGPU_VM_PTB) {
@@ -926,10 +914,8 @@ static void amdgpu_vm_invalidate_level(struct amdgpu_device *adev,
                if (!entry->base.bo)
                        continue;
 
-               spin_lock(&vm->status_lock);
                if (list_empty(&entry->base.vm_status))
                        list_add(&entry->base.vm_status, &vm->relocated);
-               spin_unlock(&vm->status_lock);
                amdgpu_vm_invalidate_level(adev, vm, entry, level + 1);
        }
 }
@@ -974,7 +960,6 @@ restart:
                params.func = amdgpu_vm_do_set_ptes;
        }
 
-       spin_lock(&vm->status_lock);
        while (!list_empty(&vm->relocated)) {
                struct amdgpu_vm_bo_base *bo_base, *parent;
                struct amdgpu_vm_pt *pt, *entry;
@@ -984,13 +969,10 @@ restart:
                                           struct amdgpu_vm_bo_base,
                                           vm_status);
                list_del_init(&bo_base->vm_status);
-               spin_unlock(&vm->status_lock);
 
                bo = bo_base->bo->parent;
-               if (!bo) {
-                       spin_lock(&vm->status_lock);
+               if (!bo)
                        continue;
-               }
 
                parent = list_first_entry(&bo->va, struct amdgpu_vm_bo_base,
                                          bo_list);
@@ -999,12 +981,10 @@ restart:
 
                amdgpu_vm_update_pde(&params, vm, pt, entry);
 
-               spin_lock(&vm->status_lock);
                if (!vm->use_cpu_for_update &&
                    (ndw - params.ib->length_dw) < 32)
                        break;
        }
-       spin_unlock(&vm->status_lock);
 
        if (vm->use_cpu_for_update) {
                /* Flush HDP */
@@ -1107,9 +1087,7 @@ static void amdgpu_vm_handle_huge_pages(struct amdgpu_pte_update_params *p,
                if (entry->huge) {
                        /* Add the entry to the relocated list to update it. */
                        entry->huge = false;
-                       spin_lock(&p->vm->status_lock);
                        list_move(&entry->base.vm_status, &p->vm->relocated);
-                       spin_unlock(&p->vm->status_lock);
                }
                return;
        }
@@ -1588,8 +1566,9 @@ int amdgpu_vm_bo_update(struct amdgpu_device *adev,
                amdgpu_asic_flush_hdp(adev, NULL);
        }
 
-       spin_lock(&vm->status_lock);
+       spin_lock(&vm->moved_lock);
        list_del_init(&bo_va->base.vm_status);
+       spin_unlock(&vm->moved_lock);
 
        /* If the BO is not in its preferred location add it back to
         * the evicted list so that it gets validated again on the
@@ -1599,7 +1578,6 @@ int amdgpu_vm_bo_update(struct amdgpu_device *adev,
            !(bo->preferred_domains &
            amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type)))
                list_add_tail(&bo_va->base.vm_status, &vm->evicted);
-       spin_unlock(&vm->status_lock);
 
        list_splice_init(&bo_va->invalids, &bo_va->valids);
        bo_va->cleared = clear;
@@ -1811,14 +1789,14 @@ int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
        bool clear;
        int r = 0;
 
-       spin_lock(&vm->status_lock);
+       spin_lock(&vm->moved_lock);
        while (!list_empty(&vm->moved)) {
                struct amdgpu_bo_va *bo_va;
                struct reservation_object *resv;
 
                bo_va = list_first_entry(&vm->moved,
                        struct amdgpu_bo_va, base.vm_status);
-               spin_unlock(&vm->status_lock);
+               spin_unlock(&vm->moved_lock);
 
                resv = bo_va->base.bo->tbo.resv;
 
@@ -1839,9 +1817,9 @@ int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
                if (!clear && resv != vm->root.base.bo->tbo.resv)
                        reservation_object_unlock(resv);
 
-               spin_lock(&vm->status_lock);
+               spin_lock(&vm->moved_lock);
        }
-       spin_unlock(&vm->status_lock);
+       spin_unlock(&vm->moved_lock);
 
        return r;
 }
@@ -1903,10 +1881,10 @@ static void amdgpu_vm_bo_insert_map(struct amdgpu_device *adev,
                amdgpu_vm_prt_get(adev);
 
        if (bo && bo->tbo.resv == vm->root.base.bo->tbo.resv) {
-               spin_lock(&vm->status_lock);
+               spin_lock(&vm->moved_lock);
                if (list_empty(&bo_va->base.vm_status))
                        list_add(&bo_va->base.vm_status, &vm->moved);
-               spin_unlock(&vm->status_lock);
+               spin_unlock(&vm->moved_lock);
        }
        trace_amdgpu_vm_bo_map(bo_va, mapping);
 }
@@ -2216,9 +2194,9 @@ void amdgpu_vm_bo_rmv(struct amdgpu_device *adev,
 
        list_del(&bo_va->base.bo_list);
 
-       spin_lock(&vm->status_lock);
+       spin_lock(&vm->moved_lock);
        list_del(&bo_va->base.vm_status);
-       spin_unlock(&vm->status_lock);
+       spin_unlock(&vm->moved_lock);
 
        list_for_each_entry_safe(mapping, next, &bo_va->valids, list) {
                list_del(&mapping->list);
@@ -2261,28 +2239,24 @@ void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
 
                bo_base->moved = true;
                if (evicted && bo->tbo.resv == vm->root.base.bo->tbo.resv) {
-                       spin_lock(&bo_base->vm->status_lock);
                        if (bo->tbo.type == ttm_bo_type_kernel)
                                list_move(&bo_base->vm_status, &vm->evicted);
                        else
                                list_move_tail(&bo_base->vm_status,
                                               &vm->evicted);
-                       spin_unlock(&bo_base->vm->status_lock);
                        continue;
                }
 
                if (bo->tbo.type == ttm_bo_type_kernel) {
-                       spin_lock(&bo_base->vm->status_lock);
                        if (list_empty(&bo_base->vm_status))
                                list_add(&bo_base->vm_status, &vm->relocated);
-                       spin_unlock(&bo_base->vm->status_lock);
                        continue;
                }
 
-               spin_lock(&bo_base->vm->status_lock);
+               spin_lock(&bo_base->vm->moved_lock);
                if (list_empty(&bo_base->vm_status))
                        list_add(&bo_base->vm_status, &vm->moved);
-               spin_unlock(&bo_base->vm->status_lock);
+               spin_unlock(&bo_base->vm->moved_lock);
        }
 }
 
@@ -2391,9 +2365,9 @@ 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;
-       spin_lock_init(&vm->status_lock);
        INIT_LIST_HEAD(&vm->evicted);
        INIT_LIST_HEAD(&vm->relocated);
+       spin_lock_init(&vm->moved_lock);
        INIT_LIST_HEAD(&vm->moved);
        INIT_LIST_HEAD(&vm->freed);
 
index d6827083572a08c3f67f33ab5fdd62e8cec6feff..0196b9a782f2ecad609de3680d336f7c1e9e7914 100644 (file)
@@ -168,9 +168,6 @@ struct amdgpu_vm {
        /* tree of virtual addresses mapped */
        struct rb_root_cached   va;
 
-       /* protecting invalidated */
-       spinlock_t              status_lock;
-
        /* BOs who needs a validation */
        struct list_head        evicted;
 
@@ -179,6 +176,7 @@ struct amdgpu_vm {
 
        /* BOs moved, but not yet updated in the PT */
        struct list_head        moved;
+       spinlock_t              moved_lock;
 
        /* BO mappings freed, but not yet updated in the PT */
        struct list_head        freed;