diff options
Diffstat (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c')
| -rw-r--r-- | drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 2134 |
1 files changed, 1534 insertions, 600 deletions
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c index 5afbc5e714d0..dad0e2342df9 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c @@ -32,9 +32,12 @@ #include <drm/amdgpu_drm.h> #include "amdgpu.h" #include "amdgpu_trace.h" +#include "amdgpu_amdkfd.h" +#include "amdgpu_gmc.h" -/* - * GPUVM +/** + * DOC: GPUVM + * * GPUVM is similar to the legacy gart on older asics, however * rather than there being a single global gart table * for the entire GPU, there are multiple VM page tables active @@ -62,33 +65,71 @@ INTERVAL_TREE_DEFINE(struct amdgpu_bo_va_mapping, rb, uint64_t, __subtree_last, #undef START #undef LAST -/* Local structure. Encapsulate some VM table update parameters to reduce +/** + * struct amdgpu_pte_update_params - Local structure + * + * Encapsulate some VM table update parameters to reduce * the number of function parameters + * */ struct amdgpu_pte_update_params { - /* amdgpu device we do this update for */ + + /** + * @adev: amdgpu device we do this update for + */ struct amdgpu_device *adev; - /* optional amdgpu_vm we do this update for */ + + /** + * @vm: optional amdgpu_vm we do this update for + */ struct amdgpu_vm *vm; - /* address where to copy page table entries from */ + + /** + * @src: address where to copy page table entries from + */ uint64_t src; - /* indirect buffer to fill with commands */ + + /** + * @ib: indirect buffer to fill with commands + */ struct amdgpu_ib *ib; - /* Function which actually does the update */ - void (*func)(struct amdgpu_pte_update_params *params, uint64_t pe, + + /** + * @func: Function which actually does the update + */ + void (*func)(struct amdgpu_pte_update_params *params, + struct amdgpu_bo *bo, uint64_t pe, uint64_t addr, unsigned count, uint32_t incr, uint64_t flags); - /* The next two are used during VM update by CPU - * DMA addresses to use for mapping - * Kernel pointer of PD/PT BO that needs to be updated + /** + * @pages_addr: + * + * DMA addresses to use for mapping, used during VM update by CPU */ dma_addr_t *pages_addr; + + /** + * @kptr: + * + * Kernel pointer of PD/PT BO that needs to be updated, + * used during VM update by CPU + */ void *kptr; }; -/* Helper to disable partial resident texture feature from a fence callback */ +/** + * struct amdgpu_prt_cb - Helper to disable partial resident texture feature from a fence callback + */ struct amdgpu_prt_cb { + + /** + * @adev: amdgpu device + */ struct amdgpu_device *adev; + + /** + * @cb: callback + */ struct dma_fence_cb cb; }; @@ -96,8 +137,10 @@ struct amdgpu_prt_cb { * amdgpu_vm_level_shift - return the addr shift for each level * * @adev: amdgpu_device pointer + * @level: VMPT level * - * Returns the number of bits the pfn needs to be right shifted for a level. + * Returns: + * The number of bits the pfn needs to be right shifted for a level. */ static unsigned amdgpu_vm_level_shift(struct amdgpu_device *adev, unsigned level) @@ -125,8 +168,10 @@ static unsigned amdgpu_vm_level_shift(struct amdgpu_device *adev, * amdgpu_vm_num_entries - return the number of entries in a PD/PT * * @adev: amdgpu_device pointer + * @level: VMPT level * - * Calculate the number of entries in a page directory or page table. + * Returns: + * The number of entries in a page directory or page table. */ static unsigned amdgpu_vm_num_entries(struct amdgpu_device *adev, unsigned level) @@ -146,11 +191,33 @@ static unsigned amdgpu_vm_num_entries(struct amdgpu_device *adev, } /** + * amdgpu_vm_entries_mask - the mask to get the entry number of a PD/PT + * + * @adev: amdgpu_device pointer + * @level: VMPT level + * + * Returns: + * The mask to extract the entry number of a PD/PT from an address. + */ +static uint32_t amdgpu_vm_entries_mask(struct amdgpu_device *adev, + unsigned int level) +{ + if (level <= adev->vm_manager.root_level) + return 0xffffffff; + else if (level != AMDGPU_VM_PTB) + return 0x1ff; + else + return AMDGPU_VM_PTE_COUNT(adev) - 1; +} + +/** * amdgpu_vm_bo_size - returns the size of the BOs in bytes * * @adev: amdgpu_device pointer + * @level: VMPT level * - * Calculate the size of the BO for a page directory or page table in bytes. + * Returns: + * The size of the BO for a page directory or page table in bytes. */ static unsigned amdgpu_vm_bo_size(struct amdgpu_device *adev, unsigned level) { @@ -158,6 +225,383 @@ static unsigned amdgpu_vm_bo_size(struct amdgpu_device *adev, unsigned level) } /** + * amdgpu_vm_bo_evicted - vm_bo is evicted + * + * @vm_bo: vm_bo which is evicted + * + * State for PDs/PTs and per VM BOs which are not at the location they should + * be. + */ +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; + + vm_bo->moved = true; + 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); +} + +/** + * 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. + */ +static void amdgpu_vm_bo_relocated(struct amdgpu_vm_bo_base *vm_bo) +{ + list_move(&vm_bo->vm_status, &vm_bo->vm->relocated); +} + +/** + * 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. + */ +static void amdgpu_vm_bo_moved(struct amdgpu_vm_bo_base *vm_bo) +{ + list_move(&vm_bo->vm_status, &vm_bo->vm->moved); +} + +/** + * amdgpu_vm_bo_idle - vm_bo is idle + * + * @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. + */ +static void amdgpu_vm_bo_idle(struct amdgpu_vm_bo_base *vm_bo) +{ + list_move(&vm_bo->vm_status, &vm_bo->vm->idle); + vm_bo->moved = false; +} + +/** + * 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->invalidated_lock); + list_move(&vm_bo->vm_status, &vm_bo->vm->invalidated); + spin_unlock(&vm_bo->vm->invalidated_lock); +} + +/** + * 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) +{ + spin_lock(&vm_bo->vm->invalidated_lock); + list_del_init(&vm_bo->vm_status); + spin_unlock(&vm_bo->vm->invalidated_lock); +} + +/** + * amdgpu_vm_bo_base_init - Adds bo to the list of bos associated with the vm + * + * @base: base structure for tracking BO usage in a VM + * @vm: vm to which bo is to be added + * @bo: amdgpu buffer object + * + * Initialize a bo_va_base structure and add it to the appropriate lists + * + */ +static void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base, + struct amdgpu_vm *vm, + struct amdgpu_bo *bo) +{ + base->vm = vm; + base->bo = bo; + base->next = NULL; + INIT_LIST_HEAD(&base->vm_status); + + if (!bo) + return; + base->next = bo->vm_bo; + bo->vm_bo = base; + + if (bo->tbo.resv != vm->root.base.bo->tbo.resv) + return; + + vm->bulk_moveable = false; + if (bo->tbo.type == ttm_bo_type_kernel) + amdgpu_vm_bo_relocated(base); + else + amdgpu_vm_bo_idle(base); + + if (bo->preferred_domains & + amdgpu_mem_type_to_domain(bo->tbo.mem.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_pt_parent - get the parent page directory + * + * @pt: child page table + * + * Helper to get the parent entry for the child page table. NULL if we are at + * the root page directory. + */ +static struct amdgpu_vm_pt *amdgpu_vm_pt_parent(struct amdgpu_vm_pt *pt) +{ + struct amdgpu_bo *parent = pt->base.bo->parent; + + if (!parent) + return NULL; + + return container_of(parent->vm_bo, struct amdgpu_vm_pt, base); +} + +/** + * amdgpu_vm_pt_cursor - state for for_each_amdgpu_vm_pt + */ +struct amdgpu_vm_pt_cursor { + uint64_t pfn; + struct amdgpu_vm_pt *parent; + struct amdgpu_vm_pt *entry; + unsigned level; +}; + +/** + * amdgpu_vm_pt_start - start PD/PT walk + * + * @adev: amdgpu_device pointer + * @vm: amdgpu_vm structure + * @start: start address of the walk + * @cursor: state to initialize + * + * Initialize a amdgpu_vm_pt_cursor to start a walk. + */ +static void amdgpu_vm_pt_start(struct amdgpu_device *adev, + struct amdgpu_vm *vm, uint64_t start, + struct amdgpu_vm_pt_cursor *cursor) +{ + cursor->pfn = start; + cursor->parent = NULL; + cursor->entry = &vm->root; + cursor->level = adev->vm_manager.root_level; +} + +/** + * amdgpu_vm_pt_descendant - go to child node + * + * @adev: amdgpu_device pointer + * @cursor: current state + * + * Walk to the child node of the current node. + * Returns: + * True if the walk was possible, false otherwise. + */ +static bool amdgpu_vm_pt_descendant(struct amdgpu_device *adev, + struct amdgpu_vm_pt_cursor *cursor) +{ + unsigned mask, shift, idx; + + if (!cursor->entry->entries) + return false; + + BUG_ON(!cursor->entry->base.bo); + mask = amdgpu_vm_entries_mask(adev, cursor->level); + shift = amdgpu_vm_level_shift(adev, cursor->level); + + ++cursor->level; + idx = (cursor->pfn >> shift) & mask; + cursor->parent = cursor->entry; + cursor->entry = &cursor->entry->entries[idx]; + return true; +} + +/** + * amdgpu_vm_pt_sibling - go to sibling node + * + * @adev: amdgpu_device pointer + * @cursor: current state + * + * Walk to the sibling node of the current node. + * Returns: + * True if the walk was possible, false otherwise. + */ +static bool amdgpu_vm_pt_sibling(struct amdgpu_device *adev, + struct amdgpu_vm_pt_cursor *cursor) +{ + unsigned shift, num_entries; + + /* Root doesn't have a sibling */ + if (!cursor->parent) + return false; + + /* Go to our parents and see if we got a sibling */ + shift = amdgpu_vm_level_shift(adev, cursor->level - 1); + num_entries = amdgpu_vm_num_entries(adev, cursor->level - 1); + + if (cursor->entry == &cursor->parent->entries[num_entries - 1]) + return false; + + cursor->pfn += 1ULL << shift; + cursor->pfn &= ~((1ULL << shift) - 1); + ++cursor->entry; + return true; +} + +/** + * amdgpu_vm_pt_ancestor - go to parent node + * + * @cursor: current state + * + * Walk to the parent node of the current node. + * Returns: + * True if the walk was possible, false otherwise. + */ +static bool amdgpu_vm_pt_ancestor(struct amdgpu_vm_pt_cursor *cursor) +{ + if (!cursor->parent) + return false; + + --cursor->level; + cursor->entry = cursor->parent; + cursor->parent = amdgpu_vm_pt_parent(cursor->parent); + return true; +} + +/** + * amdgpu_vm_pt_next - get next PD/PT in hieratchy + * + * @adev: amdgpu_device pointer + * @cursor: current state + * + * Walk the PD/PT tree to the next node. + */ +static void amdgpu_vm_pt_next(struct amdgpu_device *adev, + struct amdgpu_vm_pt_cursor *cursor) +{ + /* First try a newborn child */ + if (amdgpu_vm_pt_descendant(adev, cursor)) + return; + + /* If that didn't worked try to find a sibling */ + while (!amdgpu_vm_pt_sibling(adev, cursor)) { + /* No sibling, go to our parents and grandparents */ + if (!amdgpu_vm_pt_ancestor(cursor)) { + cursor->pfn = ~0ll; + return; + } + } +} + +/** + * amdgpu_vm_pt_first_leaf - get first leaf PD/PT + * + * @adev: amdgpu_device pointer + * @vm: amdgpu_vm structure + * @start: start addr of the walk + * @cursor: state to initialize + * + * Start a walk and go directly to the leaf node. + */ +static void amdgpu_vm_pt_first_leaf(struct amdgpu_device *adev, + struct amdgpu_vm *vm, uint64_t start, + struct amdgpu_vm_pt_cursor *cursor) +{ + amdgpu_vm_pt_start(adev, vm, start, cursor); + while (amdgpu_vm_pt_descendant(adev, cursor)); +} + +/** + * amdgpu_vm_pt_next_leaf - get next leaf PD/PT + * + * @adev: amdgpu_device pointer + * @cursor: current state + * + * Walk the PD/PT tree to the next leaf node. + */ +static void amdgpu_vm_pt_next_leaf(struct amdgpu_device *adev, + struct amdgpu_vm_pt_cursor *cursor) +{ + amdgpu_vm_pt_next(adev, cursor); + if (cursor->pfn != ~0ll) + while (amdgpu_vm_pt_descendant(adev, cursor)); +} + +/** + * for_each_amdgpu_vm_pt_leaf - walk over all leaf PDs/PTs in the hierarchy + */ +#define for_each_amdgpu_vm_pt_leaf(adev, vm, start, end, cursor) \ + for (amdgpu_vm_pt_first_leaf((adev), (vm), (start), &(cursor)); \ + (cursor).pfn <= end; amdgpu_vm_pt_next_leaf((adev), &(cursor))) + +/** + * amdgpu_vm_pt_first_dfs - start a deep first search + * + * @adev: amdgpu_device structure + * @vm: amdgpu_vm structure + * @cursor: state to initialize + * + * Starts a deep first traversal of the PD/PT tree. + */ +static void amdgpu_vm_pt_first_dfs(struct amdgpu_device *adev, + struct amdgpu_vm *vm, + struct amdgpu_vm_pt_cursor *cursor) +{ + amdgpu_vm_pt_start(adev, vm, 0, cursor); + while (amdgpu_vm_pt_descendant(adev, cursor)); +} + +/** + * amdgpu_vm_pt_next_dfs - get the next node for a deep first search + * + * @adev: amdgpu_device structure + * @cursor: current state + * + * Move the cursor to the next node in a deep first search. + */ +static void amdgpu_vm_pt_next_dfs(struct amdgpu_device *adev, + struct amdgpu_vm_pt_cursor *cursor) +{ + if (!cursor->entry) + return; + + if (!cursor->parent) + cursor->entry = NULL; + else if (amdgpu_vm_pt_sibling(adev, cursor)) + while (amdgpu_vm_pt_descendant(adev, cursor)); + else + amdgpu_vm_pt_ancestor(cursor); +} + +/** + * for_each_amdgpu_vm_pt_dfs_safe - safe deep first search of all PDs/PTs + */ +#define for_each_amdgpu_vm_pt_dfs_safe(adev, vm, cursor, entry) \ + for (amdgpu_vm_pt_first_dfs((adev), (vm), &(cursor)), \ + (entry) = (cursor).entry, amdgpu_vm_pt_next_dfs((adev), &(cursor));\ + (entry); (entry) = (cursor).entry, \ + amdgpu_vm_pt_next_dfs((adev), &(cursor))) + +/** * amdgpu_vm_get_pd_bo - add the VM PD to a validation list * * @vm: vm providing the BOs @@ -171,15 +615,55 @@ void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm, struct list_head *validated, struct amdgpu_bo_list_entry *entry) { - entry->robj = vm->root.base.bo; entry->priority = 0; - entry->tv.bo = &entry->robj->tbo; + entry->tv.bo = &vm->root.base.bo->tbo; entry->tv.shared = true; entry->user_pages = NULL; list_add(&entry->tv.head, validated); } /** + * amdgpu_vm_move_to_lru_tail - move all BOs to the end of LRU + * + * @adev: amdgpu device pointer + * @vm: vm providing the BOs + * + * Move all BOs to the end of LRU and remember their positions to put them + * together. + */ +void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev, + struct amdgpu_vm *vm) +{ + struct ttm_bo_global *glob = adev->mman.bdev.glob; + struct amdgpu_vm_bo_base *bo_base; + + if (vm->bulk_moveable) { + spin_lock(&glob->lru_lock); + ttm_bo_bulk_move_lru_tail(&vm->lru_bulk_move); + spin_unlock(&glob->lru_lock); + return; + } + + memset(&vm->lru_bulk_move, 0, sizeof(vm->lru_bulk_move)); + + spin_lock(&glob->lru_lock); + list_for_each_entry(bo_base, &vm->idle, vm_status) { + struct amdgpu_bo *bo = bo_base->bo; + + if (!bo->parent) + continue; + + ttm_bo_move_to_lru_tail(&bo->tbo, &vm->lru_bulk_move); + if (bo->shadow) + ttm_bo_move_to_lru_tail(&bo->shadow->tbo, + &vm->lru_bulk_move); + } + spin_unlock(&glob->lru_lock); + + vm->bulk_moveable = true; +} + +/** * amdgpu_vm_validate_pt_bos - validate the page table BOs * * @adev: amdgpu device pointer @@ -188,54 +672,45 @@ void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm, * @param: parameter for the validation callback * * Validate the page table BOs on command submission if neccessary. + * + * Returns: + * Validation result. */ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm, int (*validate)(void *p, struct amdgpu_bo *bo), void *param) { - 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; + struct amdgpu_vm_bo_base *bo_base, *tmp; + int r = 0; - bo_base = list_first_entry(&vm->evicted, - struct amdgpu_vm_bo_base, - vm_status); - spin_unlock(&vm->status_lock); + vm->bulk_moveable &= list_empty(&vm->evicted); - bo = bo_base->bo; - BUG_ON(!bo); - if (bo->parent) { - r = validate(param, bo); - if (r) - return r; + list_for_each_entry_safe(bo_base, tmp, &vm->evicted, vm_status) { + struct amdgpu_bo *bo = bo_base->bo; - spin_lock(&glob->lru_lock); - ttm_bo_move_to_lru_tail(&bo->tbo); - if (bo->shadow) - ttm_bo_move_to_lru_tail(&bo->shadow->tbo); - spin_unlock(&glob->lru_lock); - } + r = validate(param, bo); + if (r) + break; - if (bo->tbo.type == ttm_bo_type_kernel && - vm->use_cpu_for_update) { - r = amdgpu_bo_kmap(bo, NULL); + if (bo->tbo.type != ttm_bo_type_kernel) { + amdgpu_vm_bo_moved(bo_base); + } else { + if (vm->use_cpu_for_update) + r = amdgpu_bo_kmap(bo, NULL); + else + r = amdgpu_ttm_alloc_gart(&bo->tbo); if (r) - return r; + break; + if (bo->shadow) { + r = amdgpu_ttm_alloc_gart(&bo->shadow->tbo); + if (r) + break; + } + amdgpu_vm_bo_relocated(bo_base); } - - spin_lock(&vm->status_lock); - if (bo->tbo.type != ttm_bo_type_kernel) - list_move(&bo_base->vm_status, &vm->moved); - else - list_move(&bo_base->vm_status, &vm->relocated); } - spin_unlock(&vm->status_lock); - return 0; + return r; } /** @@ -244,125 +719,150 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm, * @vm: VM to check * * Check if all VM PDs/PTs are ready for updates + * + * Returns: + * True if eviction list is empty. */ 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); } /** - * amdgpu_vm_alloc_levels - allocate the PD/PT levels + * amdgpu_vm_clear_bo - initially clear the PDs/PTs * * @adev: amdgpu_device pointer - * @vm: requested vm - * @saddr: start of the address range - * @eaddr: end of the address range + * @vm: VM to clear BO from + * @bo: BO to clear + * @level: level this BO is at + * @pte_support_ats: indicate ATS support from PTE * - * Make sure the page directories and page tables are allocated + * Root PD needs to be reserved when calling this. + * + * Returns: + * 0 on success, errno otherwise. */ -static int amdgpu_vm_alloc_levels(struct amdgpu_device *adev, - struct amdgpu_vm *vm, - struct amdgpu_vm_pt *parent, - uint64_t saddr, uint64_t eaddr, - unsigned level) +static int amdgpu_vm_clear_bo(struct amdgpu_device *adev, + struct amdgpu_vm *vm, struct amdgpu_bo *bo, + unsigned level, bool pte_support_ats) { - unsigned shift = amdgpu_vm_level_shift(adev, level); - unsigned pt_idx, from, to; + struct ttm_operation_ctx ctx = { true, false }; + struct dma_fence *fence = NULL; + unsigned entries, ats_entries; + struct amdgpu_ring *ring; + struct amdgpu_job *job; + uint64_t addr; int r; - u64 flags; - uint64_t init_value = 0; - if (!parent->entries) { - unsigned num_entries = amdgpu_vm_num_entries(adev, level); + entries = amdgpu_bo_size(bo) / 8; - parent->entries = kvmalloc_array(num_entries, - sizeof(struct amdgpu_vm_pt), - GFP_KERNEL | __GFP_ZERO); - if (!parent->entries) - return -ENOMEM; - memset(parent->entries, 0 , sizeof(struct amdgpu_vm_pt)); + if (pte_support_ats) { + if (level == adev->vm_manager.root_level) { + ats_entries = amdgpu_vm_level_shift(adev, level); + ats_entries += AMDGPU_GPU_PAGE_SHIFT; + ats_entries = AMDGPU_GMC_HOLE_START >> ats_entries; + ats_entries = min(ats_entries, entries); + entries -= ats_entries; + } else { + ats_entries = entries; + entries = 0; + } + } else { + ats_entries = 0; } - from = saddr >> shift; - to = eaddr >> shift; - if (from >= amdgpu_vm_num_entries(adev, level) || - to >= amdgpu_vm_num_entries(adev, level)) - return -EINVAL; + ring = container_of(vm->entity.rq->sched, struct amdgpu_ring, sched); - ++level; - saddr = saddr & ((1 << shift) - 1); - eaddr = eaddr & ((1 << shift) - 1); + r = reservation_object_reserve_shared(bo->tbo.resv); + if (r) + return r; - flags = AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS | - AMDGPU_GEM_CREATE_VRAM_CLEARED; - if (vm->use_cpu_for_update) - flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED; - else - flags |= (AMDGPU_GEM_CREATE_NO_CPU_ACCESS | - AMDGPU_GEM_CREATE_SHADOW); + r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); + if (r) + goto error; + + r = amdgpu_ttm_alloc_gart(&bo->tbo); + if (r) + return r; + + r = amdgpu_job_alloc_with_ib(adev, 64, &job); + if (r) + goto error; - if (vm->pte_support_ats) { - init_value = AMDGPU_PTE_DEFAULT_ATC; + addr = amdgpu_bo_gpu_offset(bo); + if (ats_entries) { + uint64_t ats_value; + + ats_value = AMDGPU_PTE_DEFAULT_ATC; if (level != AMDGPU_VM_PTB) - init_value |= AMDGPU_PDE_PTE; + ats_value |= AMDGPU_PDE_PTE; + amdgpu_vm_set_pte_pde(adev, &job->ibs[0], addr, 0, + ats_entries, 0, ats_value); + addr += ats_entries * 8; } - /* walk over the address space and allocate the page tables */ - for (pt_idx = from; pt_idx <= to; ++pt_idx) { - struct reservation_object *resv = vm->root.base.bo->tbo.resv; - struct amdgpu_vm_pt *entry = &parent->entries[pt_idx]; - struct amdgpu_bo *pt; + if (entries) + amdgpu_vm_set_pte_pde(adev, &job->ibs[0], addr, 0, + entries, 0, 0); - if (!entry->base.bo) { - r = amdgpu_bo_create(adev, - amdgpu_vm_bo_size(adev, level), - AMDGPU_GPU_PAGE_SIZE, true, - AMDGPU_GEM_DOMAIN_VRAM, - flags, - NULL, resv, init_value, &pt); - if (r) - return r; - - if (vm->use_cpu_for_update) { - r = amdgpu_bo_kmap(pt, NULL); - if (r) { - amdgpu_bo_unref(&pt); - return r; - } - } + amdgpu_ring_pad_ib(ring, &job->ibs[0]); - /* Keep a reference to the root directory to avoid - * freeing them up in the wrong order. - */ - pt->parent = amdgpu_bo_ref(parent->base.bo); - - entry->base.vm = vm; - entry->base.bo = pt; - list_add_tail(&entry->base.bo_list, &pt->va); - spin_lock(&vm->status_lock); - list_add(&entry->base.vm_status, &vm->relocated); - spin_unlock(&vm->status_lock); - } + WARN_ON(job->ibs[0].length_dw > 64); + r = amdgpu_sync_resv(adev, &job->sync, bo->tbo.resv, + AMDGPU_FENCE_OWNER_UNDEFINED, false); + if (r) + goto error_free; - if (level < AMDGPU_VM_PTB) { - uint64_t sub_saddr = (pt_idx == from) ? saddr : 0; - uint64_t sub_eaddr = (pt_idx == to) ? eaddr : - ((1 << shift) - 1); - r = amdgpu_vm_alloc_levels(adev, vm, entry, sub_saddr, - sub_eaddr, level); - if (r) - return r; - } - } + r = amdgpu_job_submit(job, &vm->entity, AMDGPU_FENCE_OWNER_UNDEFINED, + &fence); + if (r) + goto error_free; + + amdgpu_bo_fence(bo, fence, true); + dma_fence_put(fence); + + if (bo->shadow) + return amdgpu_vm_clear_bo(adev, vm, bo->shadow, + level, pte_support_ats); return 0; + +error_free: + amdgpu_job_free(job); + +error: + return r; +} + +/** + * amdgpu_vm_bo_param - fill in parameters for PD/PT allocation + * + * @adev: amdgpu_device pointer + * @vm: requesting vm + * @bp: resulting BO allocation parameters + */ +static void amdgpu_vm_bo_param(struct amdgpu_device *adev, struct amdgpu_vm *vm, + int level, struct amdgpu_bo_param *bp) +{ + memset(bp, 0, sizeof(*bp)); + + bp->size = amdgpu_vm_bo_size(adev, level); + bp->byte_align = AMDGPU_GPU_PAGE_SIZE; + bp->domain = AMDGPU_GEM_DOMAIN_VRAM; + if (bp->size <= PAGE_SIZE && adev->asic_type >= CHIP_VEGA10 && + adev->flags & AMD_IS_APU) + bp->domain |= AMDGPU_GEM_DOMAIN_GTT; + bp->domain = amdgpu_bo_get_preferred_pin_domain(adev, bp->domain); + bp->flags = AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS | + AMDGPU_GEM_CREATE_CPU_GTT_USWC; + if (vm->use_cpu_for_update) + bp->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED; + else if (!vm->root.base.bo || vm->root.base.bo->shadow) + bp->flags |= AMDGPU_GEM_CREATE_SHADOW; + bp->type = ttm_bo_type_kernel; + if (vm->root.base.bo) + bp->resv = vm->root.base.bo->tbo.resv; } /** @@ -373,32 +873,117 @@ static int amdgpu_vm_alloc_levels(struct amdgpu_device *adev, * @saddr: Start address which needs to be allocated * @size: Size from start address we need. * - * Make sure the page tables are allocated. + * Make sure the page directories and page tables are allocated + * + * Returns: + * 0 on success, errno otherwise. */ int amdgpu_vm_alloc_pts(struct amdgpu_device *adev, struct amdgpu_vm *vm, uint64_t saddr, uint64_t size) { - uint64_t last_pfn; + struct amdgpu_vm_pt_cursor cursor; + struct amdgpu_bo *pt; + bool ats = false; uint64_t eaddr; + int r; /* validate the parameters */ if (saddr & AMDGPU_GPU_PAGE_MASK || size & AMDGPU_GPU_PAGE_MASK) return -EINVAL; eaddr = saddr + size - 1; - last_pfn = eaddr / AMDGPU_GPU_PAGE_SIZE; - if (last_pfn >= adev->vm_manager.max_pfn) { + + if (vm->pte_support_ats) + ats = saddr < AMDGPU_GMC_HOLE_START; + + saddr /= AMDGPU_GPU_PAGE_SIZE; + eaddr /= AMDGPU_GPU_PAGE_SIZE; + + if (eaddr >= adev->vm_manager.max_pfn) { dev_err(adev->dev, "va above limit (0x%08llX >= 0x%08llX)\n", - last_pfn, adev->vm_manager.max_pfn); + eaddr, adev->vm_manager.max_pfn); return -EINVAL; } - saddr /= AMDGPU_GPU_PAGE_SIZE; - eaddr /= AMDGPU_GPU_PAGE_SIZE; + for_each_amdgpu_vm_pt_leaf(adev, vm, saddr, eaddr, cursor) { + struct amdgpu_vm_pt *entry = cursor.entry; + struct amdgpu_bo_param bp; + + if (cursor.level < AMDGPU_VM_PTB) { + unsigned num_entries; + + num_entries = amdgpu_vm_num_entries(adev, cursor.level); + entry->entries = kvmalloc_array(num_entries, + sizeof(*entry->entries), + GFP_KERNEL | + __GFP_ZERO); + if (!entry->entries) + return -ENOMEM; + } + + + if (entry->base.bo) + continue; + + amdgpu_vm_bo_param(adev, vm, cursor.level, &bp); - return amdgpu_vm_alloc_levels(adev, vm, &vm->root, saddr, eaddr, - adev->vm_manager.root_level); + r = amdgpu_bo_create(adev, &bp, &pt); + if (r) + return r; + + r = amdgpu_vm_clear_bo(adev, vm, pt, cursor.level, ats); + if (r) + goto error_free_pt; + + if (vm->use_cpu_for_update) { + r = amdgpu_bo_kmap(pt, NULL); + if (r) + goto error_free_pt; + } + + /* Keep a reference to the root directory to avoid + * freeing them up in the wrong order. + */ + pt->parent = amdgpu_bo_ref(cursor.parent->base.bo); + + amdgpu_vm_bo_base_init(&entry->base, vm, pt); + } + + return 0; + +error_free_pt: + amdgpu_bo_unref(&pt->shadow); + amdgpu_bo_unref(&pt); + return r; +} + +/** + * amdgpu_vm_free_pts - free PD/PT levels + * + * @adev: amdgpu device structure + * @vm: amdgpu vm structure + * + * Free the page directory or page table level and all sub levels. + */ +static void amdgpu_vm_free_pts(struct amdgpu_device *adev, + struct amdgpu_vm *vm) +{ + struct amdgpu_vm_pt_cursor cursor; + struct amdgpu_vm_pt *entry; + + for_each_amdgpu_vm_pt_dfs_safe(adev, vm, cursor, entry) { + + if (entry->base.bo) { + entry->base.bo->vm_bo = NULL; + list_del(&entry->base.vm_status); + amdgpu_bo_unref(&entry->base.bo->shadow); + amdgpu_bo_unref(&entry->base.bo); + } + kvfree(entry->entries); + } + + BUG_ON(vm->root.base.bo); } /** @@ -436,6 +1021,15 @@ void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev) } } +/** + * amdgpu_vm_need_pipeline_sync - Check if pipe sync is needed for job. + * + * @ring: ring on which the job will be submitted + * @job: job to submit + * + * Returns: + * True if sync is needed. + */ bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring, struct amdgpu_job *job) { @@ -463,19 +1057,17 @@ bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring, return vm_flush_needed || gds_switch_needed; } -static bool amdgpu_vm_is_large_bar(struct amdgpu_device *adev) -{ - return (adev->mc.real_vram_size == adev->mc.visible_vram_size); -} - /** * amdgpu_vm_flush - hardware flush the vm * * @ring: ring to use for flush - * @vmid: vmid number to use - * @pd_addr: address of the page directory + * @job: related job + * @need_pipe_sync: is pipe sync needed * * Emit a VM flush when it is necessary. + * + * Returns: + * 0 on success, errno otherwise. */ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, bool need_pipe_sync) { @@ -491,14 +1083,25 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, bool need_ id->oa_base != job->oa_base || id->oa_size != job->oa_size); bool vm_flush_needed = job->vm_needs_flush; + bool pasid_mapping_needed = id->pasid != job->pasid || + !id->pasid_mapping || + !dma_fence_is_signaled(id->pasid_mapping); + struct dma_fence *fence = NULL; unsigned patch_offset = 0; int r; if (amdgpu_vmid_had_gpu_reset(adev, id)) { gds_switch_needed = true; vm_flush_needed = true; + pasid_mapping_needed = true; } + gds_switch_needed &= !!ring->funcs->emit_gds_switch; + vm_flush_needed &= !!ring->funcs->emit_vm_flush && + job->vm_pd_addr != AMDGPU_BO_INVALID_OFFSET; + pasid_mapping_needed &= adev->gmc.gmc_funcs->emit_pasid_mapping && + ring->funcs->emit_wreg; + if (!vm_flush_needed && !gds_switch_needed && !need_pipe_sync) return 0; @@ -508,23 +1111,36 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, bool need_ if (need_pipe_sync) amdgpu_ring_emit_pipeline_sync(ring); - if (ring->funcs->emit_vm_flush && vm_flush_needed) { - struct dma_fence *fence; - + if (vm_flush_needed) { trace_amdgpu_vm_flush(ring, job->vmid, job->vm_pd_addr); amdgpu_ring_emit_vm_flush(ring, job->vmid, job->vm_pd_addr); + } + + if (pasid_mapping_needed) + amdgpu_gmc_emit_pasid_mapping(ring, job->vmid, job->pasid); - r = amdgpu_fence_emit(ring, &fence); + if (vm_flush_needed || pasid_mapping_needed) { + r = amdgpu_fence_emit(ring, &fence, 0); if (r) return r; + } + if (vm_flush_needed) { mutex_lock(&id_mgr->lock); dma_fence_put(id->last_flush); - id->last_flush = fence; - id->current_gpu_reset_count = atomic_read(&adev->gpu_reset_counter); + id->last_flush = dma_fence_get(fence); + id->current_gpu_reset_count = + atomic_read(&adev->gpu_reset_counter); mutex_unlock(&id_mgr->lock); } + if (pasid_mapping_needed) { + id->pasid = job->pasid; + dma_fence_put(id->pasid_mapping); + id->pasid_mapping = dma_fence_get(fence); + } + dma_fence_put(fence); + if (ring->funcs->emit_gds_switch && gds_switch_needed) { id->gds_base = job->gds_base; id->gds_size = job->gds_size; @@ -560,16 +1176,20 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, bool need_ * Returns the found bo_va or NULL if none is found * * Object has to be reserved! + * + * Returns: + * Found bo_va or NULL. */ struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm, struct amdgpu_bo *bo) { - struct amdgpu_bo_va *bo_va; + struct amdgpu_vm_bo_base *base; - list_for_each_entry(bo_va, &bo->va, base.bo_list) { - if (bo_va->base.vm == vm) { - return bo_va; - } + for (base = bo->vm_bo; base; base = base->next) { + if (base->vm != vm) + continue; + + return container_of(base, struct amdgpu_bo_va, base); } return NULL; } @@ -578,6 +1198,7 @@ struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm, * amdgpu_vm_do_set_ptes - helper to call the right asic function * * @params: see amdgpu_pte_update_params definition + * @bo: PD/PT to update * @pe: addr of the page entry * @addr: dst addr to write into pe * @count: number of page entries to update @@ -588,10 +1209,12 @@ struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm, * to setup the page table using the DMA. */ static void amdgpu_vm_do_set_ptes(struct amdgpu_pte_update_params *params, + struct amdgpu_bo *bo, uint64_t pe, uint64_t addr, unsigned count, uint32_t incr, uint64_t flags) { + pe += amdgpu_bo_gpu_offset(bo); trace_amdgpu_vm_set_ptes(pe, addr, count, incr, flags); if (count < 3) { @@ -608,6 +1231,7 @@ static void amdgpu_vm_do_set_ptes(struct amdgpu_pte_update_params *params, * amdgpu_vm_do_copy_ptes - copy the PTEs from the GART * * @params: see amdgpu_pte_update_params definition + * @bo: PD/PT to update * @pe: addr of the page entry * @addr: dst addr to write into pe * @count: number of page entries to update @@ -617,13 +1241,14 @@ static void amdgpu_vm_do_set_ptes(struct amdgpu_pte_update_params *params, * Traces the parameters and calls the DMA function to copy the PTEs. */ static void amdgpu_vm_do_copy_ptes(struct amdgpu_pte_update_params *params, + struct amdgpu_bo *bo, uint64_t pe, uint64_t addr, unsigned count, uint32_t incr, uint64_t flags) { uint64_t src = (params->src + (addr >> 12) * 8); - + pe += amdgpu_bo_gpu_offset(bo); trace_amdgpu_vm_copy_ptes(pe, src, count); amdgpu_vm_copy_pte(params->adev, params->ib, pe, src, count); @@ -636,7 +1261,10 @@ static void amdgpu_vm_do_copy_ptes(struct amdgpu_pte_update_params *params, * @addr: the unmapped addr * * Look up the physical address of the page that the pte resolves - * to and return the pointer for the page table entry. + * to. + * + * Returns: + * The pointer for the page table entry. */ static uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr) { @@ -657,6 +1285,7 @@ static uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr) * amdgpu_vm_cpu_set_ptes - helper to update page tables via CPU * * @params: see amdgpu_pte_update_params definition + * @bo: PD/PT to update * @pe: kmap addr of the page entry * @addr: dst addr to write into pe * @count: number of page entries to update @@ -666,6 +1295,7 @@ static uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr) * Write count number of PT/PD entries directly. */ static void amdgpu_vm_cpu_set_ptes(struct amdgpu_pte_update_params *params, + struct amdgpu_bo *bo, uint64_t pe, uint64_t addr, unsigned count, uint32_t incr, uint64_t flags) @@ -673,18 +1303,31 @@ static void amdgpu_vm_cpu_set_ptes(struct amdgpu_pte_update_params *params, unsigned int i; uint64_t value; + pe += (unsigned long)amdgpu_bo_kptr(bo); + trace_amdgpu_vm_set_ptes(pe, addr, count, incr, flags); for (i = 0; i < count; i++) { value = params->pages_addr ? amdgpu_vm_map_gart(params->pages_addr, addr) : addr; - amdgpu_gart_set_pte_pde(params->adev, (void *)(uintptr_t)pe, - i, value, flags); + amdgpu_gmc_set_pte_pde(params->adev, (void *)(uintptr_t)pe, + i, value, flags); addr += incr; } } + +/** + * amdgpu_vm_wait_pd - Wait for PT BOs to be free. + * + * @adev: amdgpu_device pointer + * @vm: related vm + * @owner: fence owner + * + * Returns: + * 0 on success, errno otherwise. + */ static int amdgpu_vm_wait_pd(struct amdgpu_device *adev, struct amdgpu_vm *vm, void *owner) { @@ -699,6 +1342,22 @@ static int amdgpu_vm_wait_pd(struct amdgpu_device *adev, struct amdgpu_vm *vm, return r; } +/** + * amdgpu_vm_update_func - helper to call update function + * + * Calls the update function for both the given BO as well as its shadow. + */ +static void amdgpu_vm_update_func(struct amdgpu_pte_update_params *params, + struct amdgpu_bo *bo, + uint64_t pe, uint64_t addr, + unsigned count, uint32_t incr, + uint64_t flags) +{ + if (bo->shadow) + params->func(params, bo->shadow, pe, addr, count, incr, flags); + params->func(params, bo, pe, addr, count, incr, flags); +} + /* * amdgpu_vm_update_pde - update a single level in the hierarchy * @@ -714,8 +1373,7 @@ static void amdgpu_vm_update_pde(struct amdgpu_pte_update_params *params, struct amdgpu_vm_pt *parent, struct amdgpu_vm_pt *entry) { - struct amdgpu_bo *bo = entry->base.bo, *shadow = NULL, *pbo; - uint64_t pd_addr, shadow_addr = 0; + struct amdgpu_bo *bo = parent->base.bo, *pbo; uint64_t pde, pt, flags; unsigned level; @@ -723,62 +1381,32 @@ static void amdgpu_vm_update_pde(struct amdgpu_pte_update_params *params, if (entry->huge) return; - if (vm->use_cpu_for_update) { - pd_addr = (unsigned long)amdgpu_bo_kptr(parent->base.bo); - } else { - pd_addr = amdgpu_bo_gpu_offset(parent->base.bo); - shadow = parent->base.bo->shadow; - if (shadow) - shadow_addr = amdgpu_bo_gpu_offset(shadow); - } - - for (level = 0, pbo = parent->base.bo->parent; pbo; ++level) + for (level = 0, pbo = bo->parent; pbo; ++level) pbo = pbo->parent; level += params->adev->vm_manager.root_level; - pt = amdgpu_bo_gpu_offset(bo); - flags = AMDGPU_PTE_VALID; - amdgpu_gart_get_vm_pde(params->adev, level, &pt, &flags); - if (shadow) { - pde = shadow_addr + (entry - parent->entries) * 8; - params->func(params, pde, pt, 1, 0, flags); - } - - pde = pd_addr + (entry - parent->entries) * 8; - params->func(params, pde, pt, 1, 0, flags); + amdgpu_gmc_get_pde_for_bo(entry->base.bo, level, &pt, &flags); + pde = (entry - parent->entries) * 8; + amdgpu_vm_update_func(params, bo, pde, pt, 1, 0, flags); } /* - * amdgpu_vm_invalidate_level - mark all PD levels as invalid + * amdgpu_vm_invalidate_pds - mark all PDs as invalid * - * @parent: parent PD + * @adev: amdgpu_device pointer + * @vm: related vm * * Mark all PD level as invalid after an error. */ -static void amdgpu_vm_invalidate_level(struct amdgpu_device *adev, - struct amdgpu_vm *vm, - struct amdgpu_vm_pt *parent, - unsigned level) +static void amdgpu_vm_invalidate_pds(struct amdgpu_device *adev, + struct amdgpu_vm *vm) { - unsigned pt_idx, num_entries; - - /* - * Recurse into the subdirectories. This recursion is harmless because - * we only have a maximum of 5 layers. - */ - num_entries = amdgpu_vm_num_entries(adev, level); - for (pt_idx = 0; pt_idx < num_entries; ++pt_idx) { - struct amdgpu_vm_pt *entry = &parent->entries[pt_idx]; - - if (!entry->base.bo) - continue; + struct amdgpu_vm_pt_cursor cursor; + struct amdgpu_vm_pt *entry; - 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); - } + for_each_amdgpu_vm_pt_dfs_safe(adev, vm, cursor, entry) + if (entry->base.bo && !entry->base.moved) + amdgpu_vm_bo_relocated(&entry->base); } /* @@ -788,7 +1416,9 @@ static void amdgpu_vm_invalidate_level(struct amdgpu_device *adev, * @vm: requested vm * * Makes sure all directories are up to date. - * Returns 0 for success, error for failure. + * + * Returns: + * 0 for success, error for failure. */ int amdgpu_vm_update_directories(struct amdgpu_device *adev, struct amdgpu_vm *vm) @@ -821,42 +1451,28 @@ 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; - struct amdgpu_bo *bo; - bo_base = list_first_entry(&vm->relocated, - struct amdgpu_vm_bo_base, - vm_status); - list_del_init(&bo_base->vm_status); - spin_unlock(&vm->status_lock); + entry = list_first_entry(&vm->relocated, struct amdgpu_vm_pt, + base.vm_status); + amdgpu_vm_bo_idle(&entry->base); - bo = bo_base->bo->parent; - if (!bo) { - spin_lock(&vm->status_lock); + pt = amdgpu_vm_pt_parent(entry); + if (!pt) continue; - } - - parent = list_first_entry(&bo->va, struct amdgpu_vm_bo_base, - bo_list); - pt = container_of(parent, struct amdgpu_vm_pt, base); - entry = container_of(bo_base, struct amdgpu_vm_pt, base); amdgpu_vm_update_pde(¶ms, 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 */ mb(); - amdgpu_gart_flush_gpu_tlb(adev, 0); + amdgpu_asic_flush_hdp(adev, NULL); } else if (params.ib->length_dw == 0) { amdgpu_job_free(job); } else { @@ -864,20 +1480,15 @@ restart: struct amdgpu_ring *ring; struct dma_fence *fence; - ring = container_of(vm->entity.sched, struct amdgpu_ring, + ring = container_of(vm->entity.rq->sched, struct amdgpu_ring, sched); amdgpu_ring_pad_ib(ring, params.ib); amdgpu_sync_resv(adev, &job->sync, root->tbo.resv, AMDGPU_FENCE_OWNER_VM, false); - if (root->shadow) - amdgpu_sync_resv(adev, &job->sync, - root->shadow->tbo.resv, - AMDGPU_FENCE_OWNER_VM, false); - WARN_ON(params.ib->length_dw > ndw); - r = amdgpu_job_submit(job, ring, &vm->entity, - AMDGPU_FENCE_OWNER_VM, &fence); + r = amdgpu_job_submit(job, &vm->entity, AMDGPU_FENCE_OWNER_VM, + &fence); if (r) goto error; @@ -892,226 +1503,208 @@ restart: return 0; error: - amdgpu_vm_invalidate_level(adev, vm, &vm->root, - adev->vm_manager.root_level); + amdgpu_vm_invalidate_pds(adev, vm); amdgpu_job_free(job); return r; } /** - * amdgpu_vm_find_entry - find the entry for an address - * - * @p: see amdgpu_pte_update_params definition - * @addr: virtual address in question - * @entry: resulting entry or NULL - * @parent: parent entry + * amdgpu_vm_update_huge - figure out parameters for PTE updates * - * Find the vm_pt entry and it's parent for the given address. + * Make sure to set the right flags for the PTEs at the desired level. */ -void amdgpu_vm_get_entry(struct amdgpu_pte_update_params *p, uint64_t addr, - struct amdgpu_vm_pt **entry, - struct amdgpu_vm_pt **parent) -{ - unsigned level = p->adev->vm_manager.root_level; - - *parent = NULL; - *entry = &p->vm->root; - while ((*entry)->entries) { - unsigned shift = amdgpu_vm_level_shift(p->adev, level++); +static void amdgpu_vm_update_huge(struct amdgpu_pte_update_params *params, + struct amdgpu_bo *bo, unsigned level, + uint64_t pe, uint64_t addr, + unsigned count, uint32_t incr, + uint64_t flags) - *parent = *entry; - *entry = &(*entry)->entries[addr >> shift]; - addr &= (1ULL << shift) - 1; +{ + if (level != AMDGPU_VM_PTB) { + flags |= AMDGPU_PDE_PTE; + amdgpu_gmc_get_vm_pde(params->adev, level, &addr, &flags); } - if (level != AMDGPU_VM_PTB) - *entry = NULL; + amdgpu_vm_update_func(params, bo, pe, addr, count, incr, flags); } /** - * amdgpu_vm_handle_huge_pages - handle updating the PD with huge pages + * amdgpu_vm_fragment - get fragment for PTEs * - * @p: see amdgpu_pte_update_params definition - * @entry: vm_pt entry to check - * @parent: parent entry - * @nptes: number of PTEs updated with this operation - * @dst: destination address where the PTEs should point to - * @flags: access flags fro the PTEs + * @params: see amdgpu_pte_update_params definition + * @start: first PTE to handle + * @end: last PTE to handle + * @flags: hw mapping flags + * @frag: resulting fragment size + * @frag_end: end of this fragment * - * Check if we can update the PD with a huge page. + * Returns the first possible fragment for the start and end address. */ -static void amdgpu_vm_handle_huge_pages(struct amdgpu_pte_update_params *p, - struct amdgpu_vm_pt *entry, - struct amdgpu_vm_pt *parent, - unsigned nptes, uint64_t dst, - uint64_t flags) +static void amdgpu_vm_fragment(struct amdgpu_pte_update_params *params, + uint64_t start, uint64_t end, uint64_t flags, + unsigned int *frag, uint64_t *frag_end) { - uint64_t pd_addr, pde; + /** + * The MC L1 TLB supports variable sized pages, based on a fragment + * field in the PTE. When this field is set to a non-zero value, page + * granularity is increased from 4KB to (1 << (12 + frag)). The PTE + * flags are considered valid for all PTEs within the fragment range + * and corresponding mappings are assumed to be physically contiguous. + * + * The L1 TLB can store a single PTE for the whole fragment, + * significantly increasing the space available for translation + * caching. This leads to large improvements in throughput when the + * TLB is under pressure. + * + * The L2 TLB distributes small and large fragments into two + * asymmetric partitions. The large fragment cache is significantly + * larger. Thus, we try to use large fragments wherever possible. + * Userspace can support this by aligning virtual base address and + * allocation size to the fragment size. + * + * Starting with Vega10 the fragment size only controls the L1. The L2 + * is now directly feed with small/huge/giant pages from the walker. + */ + unsigned max_frag; - /* In the case of a mixed PT the PDE must point to it*/ - if (p->adev->asic_type >= CHIP_VEGA10 && !p->src && - nptes == AMDGPU_VM_PTE_COUNT(p->adev)) { - /* Set the huge page flag to stop scanning at this PDE */ - flags |= AMDGPU_PDE_PTE; - } + if (params->adev->asic_type < CHIP_VEGA10) + max_frag = params->adev->vm_manager.fragment_size; + else + max_frag = 31; - if (!(flags & AMDGPU_PDE_PTE)) { - 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); - } + /* system pages are non continuously */ + if (params->src) { + *frag = 0; + *frag_end = end; return; } - entry->huge = true; - amdgpu_gart_get_vm_pde(p->adev, AMDGPU_VM_PDB0, - &dst, &flags); - - if (p->func == amdgpu_vm_cpu_set_ptes) { - pd_addr = (unsigned long)amdgpu_bo_kptr(parent->base.bo); + /* This intentionally wraps around if no bit is set */ + *frag = min((unsigned)ffs(start) - 1, (unsigned)fls64(end - start) - 1); + if (*frag >= max_frag) { + *frag = max_frag; + *frag_end = end & ~((1ULL << max_frag) - 1); } else { - if (parent->base.bo->shadow) { - pd_addr = amdgpu_bo_gpu_offset(parent->base.bo->shadow); - pde = pd_addr + (entry - parent->entries) * 8; - p->func(p, pde, dst, 1, 0, flags); - } - pd_addr = amdgpu_bo_gpu_offset(parent->base.bo); + *frag_end = start + (1 << *frag); } - pde = pd_addr + (entry - parent->entries) * 8; - p->func(p, pde, dst, 1, 0, flags); } /** * amdgpu_vm_update_ptes - make sure that page tables are valid * * @params: see amdgpu_pte_update_params definition - * @vm: requested vm * @start: start of GPU address range * @end: end of GPU address range * @dst: destination address to map to, the next dst inside the function * @flags: mapping flags * * Update the page tables in the range @start - @end. - * Returns 0 for success, -EINVAL for failure. + * + * Returns: + * 0 for success, -EINVAL for failure. */ static int amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params, - uint64_t start, uint64_t end, - uint64_t dst, uint64_t flags) + uint64_t start, uint64_t end, + uint64_t dst, uint64_t flags) { struct amdgpu_device *adev = params->adev; - const uint64_t mask = AMDGPU_VM_PTE_COUNT(adev) - 1; + struct amdgpu_vm_pt_cursor cursor; + uint64_t frag_start = start, frag_end; + unsigned int frag; - uint64_t addr, pe_start; - struct amdgpu_bo *pt; - unsigned nptes; - bool use_cpu_update = (params->func == amdgpu_vm_cpu_set_ptes); + /* figure out the initial fragment */ + amdgpu_vm_fragment(params, frag_start, end, flags, &frag, &frag_end); - /* walk over the address space and update the page tables */ - for (addr = start; addr < end; addr += nptes, - dst += nptes * AMDGPU_GPU_PAGE_SIZE) { - struct amdgpu_vm_pt *entry, *parent; + /* walk over the address space and update the PTs */ + amdgpu_vm_pt_start(adev, params->vm, start, &cursor); + while (cursor.pfn < end) { + struct amdgpu_bo *pt = cursor.entry->base.bo; + unsigned shift, parent_shift, mask; + uint64_t incr, entry_end, pe_start; - amdgpu_vm_get_entry(params, addr, &entry, &parent); - if (!entry) + if (!pt) return -ENOENT; - if ((addr & ~mask) == (end & ~mask)) - nptes = end - addr; - else - nptes = AMDGPU_VM_PTE_COUNT(adev) - (addr & mask); - - amdgpu_vm_handle_huge_pages(params, entry, parent, - nptes, dst, flags); - /* We don't need to update PTEs for huge pages */ - if (entry->huge) + /* The root level can't be a huge page */ + if (cursor.level == adev->vm_manager.root_level) { + if (!amdgpu_vm_pt_descendant(adev, &cursor)) + return -ENOENT; continue; + } - pt = entry->base.bo; - if (use_cpu_update) { - pe_start = (unsigned long)amdgpu_bo_kptr(pt); - } else { - if (pt->shadow) { - pe_start = amdgpu_bo_gpu_offset(pt->shadow); - pe_start += (addr & mask) * 8; - params->func(params, pe_start, dst, nptes, - AMDGPU_GPU_PAGE_SIZE, flags); - } - pe_start = amdgpu_bo_gpu_offset(pt); + /* If it isn't already handled it can't be a huge page */ + if (cursor.entry->huge) { + /* Add the entry to the relocated list to update it. */ + cursor.entry->huge = false; + amdgpu_vm_bo_relocated(&cursor.entry->base); } - pe_start += (addr & mask) * 8; - params->func(params, pe_start, dst, nptes, - AMDGPU_GPU_PAGE_SIZE, flags); - } + shift = amdgpu_vm_level_shift(adev, cursor.level); + parent_shift = amdgpu_vm_level_shift(adev, cursor.level - 1); + if (adev->asic_type < CHIP_VEGA10) { + /* No huge page support before GMC v9 */ + if (cursor.level != AMDGPU_VM_PTB) { + if (!amdgpu_vm_pt_descendant(adev, &cursor)) + return -ENOENT; + continue; + } + } else if (frag < shift) { + /* We can't use this level when the fragment size is + * smaller than the address shift. Go to the next + * child entry and try again. + */ + if (!amdgpu_vm_pt_descendant(adev, &cursor)) + return -ENOENT; + continue; + } else if (frag >= parent_shift) { + /* If the fragment size is even larger than the parent + * shift we should go up one level and check it again. + */ + if (!amdgpu_vm_pt_ancestor(&cursor)) + return -ENOENT; + continue; + } - return 0; -} + /* Looks good so far, calculate parameters for the update */ + incr = AMDGPU_GPU_PAGE_SIZE << shift; + mask = amdgpu_vm_entries_mask(adev, cursor.level); + pe_start = ((cursor.pfn >> shift) & mask) * 8; + entry_end = (mask + 1) << shift; + entry_end += cursor.pfn & ~(entry_end - 1); + entry_end = min(entry_end, end); + + do { + uint64_t upd_end = min(entry_end, frag_end); + unsigned nptes = (upd_end - frag_start) >> shift; + + amdgpu_vm_update_huge(params, pt, cursor.level, + pe_start, dst, nptes, incr, + flags | AMDGPU_PTE_FRAG(frag)); + + pe_start += nptes * 8; + dst += nptes * AMDGPU_GPU_PAGE_SIZE << shift; + + frag_start = upd_end; + if (frag_start >= frag_end) { + /* figure out the next fragment */ + amdgpu_vm_fragment(params, frag_start, end, + flags, &frag, &frag_end); + if (frag < shift) + break; + } + } while (frag_start < entry_end); -/* - * amdgpu_vm_frag_ptes - add fragment information to PTEs - * - * @params: see amdgpu_pte_update_params definition - * @vm: requested vm - * @start: first PTE to handle - * @end: last PTE to handle - * @dst: addr those PTEs should point to - * @flags: hw mapping flags - * Returns 0 for success, -EINVAL for failure. - */ -static int amdgpu_vm_frag_ptes(struct amdgpu_pte_update_params *params, - uint64_t start, uint64_t end, - uint64_t dst, uint64_t flags) -{ - /** - * The MC L1 TLB supports variable sized pages, based on a fragment - * field in the PTE. When this field is set to a non-zero value, page - * granularity is increased from 4KB to (1 << (12 + frag)). The PTE - * flags are considered valid for all PTEs within the fragment range - * and corresponding mappings are assumed to be physically contiguous. - * - * The L1 TLB can store a single PTE for the whole fragment, - * significantly increasing the space available for translation - * caching. This leads to large improvements in throughput when the - * TLB is under pressure. - * - * The L2 TLB distributes small and large fragments into two - * asymmetric partitions. The large fragment cache is significantly - * larger. Thus, we try to use large fragments wherever possible. - * Userspace can support this by aligning virtual base address and - * allocation size to the fragment size. - */ - unsigned max_frag = params->adev->vm_manager.fragment_size; - int r; + if (amdgpu_vm_pt_descendant(adev, &cursor)) { + /* Mark all child entries as huge */ + while (cursor.pfn < frag_start) { + cursor.entry->huge = true; + amdgpu_vm_pt_next(adev, &cursor); + } - /* system pages are non continuously */ - if (params->src || !(flags & AMDGPU_PTE_VALID)) - return amdgpu_vm_update_ptes(params, start, end, dst, flags); - - while (start != end) { - uint64_t frag_flags, frag_end; - unsigned frag; - - /* This intentionally wraps around if no bit is set */ - frag = min((unsigned)ffs(start) - 1, - (unsigned)fls64(end - start) - 1); - if (frag >= max_frag) { - frag_flags = AMDGPU_PTE_FRAG(max_frag); - frag_end = end & ~((1ULL << max_frag) - 1); - } else { - frag_flags = AMDGPU_PTE_FRAG(frag); - frag_end = start + (1 << frag); + } else if (frag >= shift) { + /* or just move on to the next on the same level. */ + amdgpu_vm_pt_next(adev, &cursor); } - - r = amdgpu_vm_update_ptes(params, start, frag_end, dst, - flags | frag_flags); - if (r) - return r; - - dst += (frag_end - start) * AMDGPU_GPU_PAGE_SIZE; - start = frag_end; } return 0; @@ -1131,7 +1724,9 @@ static int amdgpu_vm_frag_ptes(struct amdgpu_pte_update_params *params, * @fence: optional resulting fence * * Fill in the page table entries between @start and @last. - * Returns 0 for success, -EINVAL for failure. + * + * Returns: + * 0 for success, -EINVAL for failure. */ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev, struct dma_fence *exclusive, @@ -1171,11 +1766,11 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev, params.func = amdgpu_vm_cpu_set_ptes; params.pages_addr = pages_addr; - return amdgpu_vm_frag_ptes(¶ms, start, last + 1, - addr, flags); + return amdgpu_vm_update_ptes(¶ms, start, last + 1, + addr, flags); } - ring = container_of(vm->entity.sched, struct amdgpu_ring, sched); + ring = container_of(vm->entity.rq->sched, struct amdgpu_ring, sched); nptes = last - start + 1; @@ -1204,11 +1799,13 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev, } else { /* set page commands needed */ - ndw += ncmds * adev->vm_manager.vm_pte_funcs->set_pte_pde_num_dw; + ndw += ncmds * 10; /* extra commands for begin/end fragments */ - ndw += 2 * adev->vm_manager.vm_pte_funcs->set_pte_pde_num_dw - * adev->vm_manager.fragment_size; + if (vm->root.base.bo->shadow) + ndw += 2 * 10 * adev->vm_manager.fragment_size * 2; + else + ndw += 2 * 10 * adev->vm_manager.fragment_size; params.func = amdgpu_vm_do_set_ptes; } @@ -1249,14 +1846,13 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev, if (r) goto error_free; - r = amdgpu_vm_frag_ptes(¶ms, start, last + 1, addr, flags); + r = amdgpu_vm_update_ptes(¶ms, start, last + 1, addr, flags); if (r) goto error_free; amdgpu_ring_pad_ib(ring, params.ib); WARN_ON(params.ib->length_dw > ndw); - r = amdgpu_job_submit(job, ring, &vm->entity, - AMDGPU_FENCE_OWNER_VM, &f); + r = amdgpu_job_submit(job, &vm->entity, AMDGPU_FENCE_OWNER_VM, &f); if (r) goto error_free; @@ -1284,7 +1880,9 @@ error_free: * * Split the mapping into smaller chunks so that each update fits * into a SDMA IB. - * Returns 0 for success, -EINVAL for failure. + * + * Returns: + * 0 for success, -EINVAL for failure. */ static int amdgpu_vm_bo_split_mapping(struct amdgpu_device *adev, struct dma_fence *exclusive, @@ -1337,7 +1935,7 @@ static int amdgpu_vm_bo_split_mapping(struct amdgpu_device *adev, if (nodes) { addr = nodes->start << PAGE_SHIFT; max_entries = (nodes->size - pfn) * - (PAGE_SIZE / AMDGPU_GPU_PAGE_SIZE); + AMDGPU_GPU_PAGES_IN_CPU_PAGE; } else { addr = 0; max_entries = S64_MAX; @@ -1347,7 +1945,9 @@ static int amdgpu_vm_bo_split_mapping(struct amdgpu_device *adev, uint64_t count; max_entries = min(max_entries, 16ull * 1024ull); - for (count = 1; count < max_entries; ++count) { + for (count = 1; + count < max_entries / AMDGPU_GPU_PAGES_IN_CPU_PAGE; + ++count) { uint64_t idx = pfn + count; if (pages_addr[idx] != @@ -1360,7 +1960,7 @@ static int amdgpu_vm_bo_split_mapping(struct amdgpu_device *adev, dma_addr = pages_addr; } else { addr = pages_addr[pfn]; - max_entries = count; + max_entries = count * AMDGPU_GPU_PAGES_IN_CPU_PAGE; } } else if (flags & AMDGPU_PTE_VALID) { @@ -1375,7 +1975,7 @@ static int amdgpu_vm_bo_split_mapping(struct amdgpu_device *adev, if (r) return r; - pfn += last - start + 1; + pfn += (last - start + 1) / AMDGPU_GPU_PAGES_IN_CPU_PAGE; if (nodes && nodes->size == pfn) { pfn = 0; ++nodes; @@ -1395,7 +1995,9 @@ static int amdgpu_vm_bo_split_mapping(struct amdgpu_device *adev, * @clear: if true clear the entries * * Fill in the page table entries for @bo_va. - * Returns 0 for success, -EINVAL for failure. + * + * Returns: + * 0 for success, -EINVAL for failure. */ int amdgpu_vm_bo_update(struct amdgpu_device *adev, struct amdgpu_bo_va *bo_va, @@ -1411,18 +2013,17 @@ int amdgpu_vm_bo_update(struct amdgpu_device *adev, uint64_t flags; int r; - if (clear || !bo_va->base.bo) { + if (clear || !bo) { mem = NULL; nodes = NULL; exclusive = NULL; } else { struct ttm_dma_tt *ttm; - mem = &bo_va->base.bo->tbo.mem; + mem = &bo->tbo.mem; nodes = mem->mm_node; if (mem->mem_type == TTM_PL_TT) { - ttm = container_of(bo_va->base.bo->tbo.ttm, - struct ttm_dma_tt, ttm); + ttm = container_of(bo->tbo.ttm, struct ttm_dma_tt, ttm); pages_addr = ttm->dma_address; } exclusive = reservation_object_get_excl(bo->tbo.resv); @@ -1457,12 +2058,23 @@ int amdgpu_vm_bo_update(struct amdgpu_device *adev, if (vm->use_cpu_for_update) { /* Flush HDP */ mb(); - amdgpu_gart_flush_gpu_tlb(adev, 0); + amdgpu_asic_flush_hdp(adev, NULL); } - spin_lock(&vm->status_lock); - list_del_init(&bo_va->base.vm_status); - spin_unlock(&vm->status_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 + * next command submission. + */ + if (bo && bo->tbo.resv == vm->root.base.bo->tbo.resv) { + uint32_t mem_type = bo->tbo.mem.mem_type; + + if (!(bo->preferred_domains & amdgpu_mem_type_to_domain(mem_type))) + amdgpu_vm_bo_evicted(&bo_va->base); + else + amdgpu_vm_bo_idle(&bo_va->base); + } else { + amdgpu_vm_bo_done(&bo_va->base); + } list_splice_init(&bo_va->invalids, &bo_va->valids); bo_va->cleared = clear; @@ -1477,6 +2089,8 @@ int amdgpu_vm_bo_update(struct amdgpu_device *adev, /** * amdgpu_vm_update_prt_state - update the global PRT state + * + * @adev: amdgpu_device pointer */ static void amdgpu_vm_update_prt_state(struct amdgpu_device *adev) { @@ -1485,16 +2099,18 @@ static void amdgpu_vm_update_prt_state(struct amdgpu_device *adev) spin_lock_irqsave(&adev->vm_manager.prt_lock, flags); enable = !!atomic_read(&adev->vm_manager.num_prt_users); - adev->gart.gart_funcs->set_prt(adev, enable); + adev->gmc.gmc_funcs->set_prt(adev, enable); spin_unlock_irqrestore(&adev->vm_manager.prt_lock, flags); } /** * amdgpu_vm_prt_get - add a PRT user + * + * @adev: amdgpu_device pointer */ static void amdgpu_vm_prt_get(struct amdgpu_device *adev) { - if (!adev->gart.gart_funcs->set_prt) + if (!adev->gmc.gmc_funcs->set_prt) return; if (atomic_inc_return(&adev->vm_manager.num_prt_users) == 1) @@ -1503,6 +2119,8 @@ static void amdgpu_vm_prt_get(struct amdgpu_device *adev) /** * amdgpu_vm_prt_put - drop a PRT user + * + * @adev: amdgpu_device pointer */ static void amdgpu_vm_prt_put(struct amdgpu_device *adev) { @@ -1512,6 +2130,9 @@ static void amdgpu_vm_prt_put(struct amdgpu_device *adev) /** * amdgpu_vm_prt_cb - callback for updating the PRT status + * + * @fence: fence for the callback + * @_cb: the callback function */ static void amdgpu_vm_prt_cb(struct dma_fence *fence, struct dma_fence_cb *_cb) { @@ -1523,13 +2144,16 @@ static void amdgpu_vm_prt_cb(struct dma_fence *fence, struct dma_fence_cb *_cb) /** * amdgpu_vm_add_prt_cb - add callback for updating the PRT status + * + * @adev: amdgpu_device pointer + * @fence: fence for the callback */ static void amdgpu_vm_add_prt_cb(struct amdgpu_device *adev, struct dma_fence *fence) { struct amdgpu_prt_cb *cb; - if (!adev->gart.gart_funcs->set_prt) + if (!adev->gmc.gmc_funcs->set_prt) return; cb = kmalloc(sizeof(struct amdgpu_prt_cb), GFP_KERNEL); @@ -1614,25 +2238,28 @@ static void amdgpu_vm_prt_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm) * or if an error occurred) * * Make sure all freed BOs are cleared in the PT. - * Returns 0 for success. - * * PTs have to be reserved and mutex must be locked! + * + * Returns: + * 0 for success. + * */ int amdgpu_vm_clear_freed(struct amdgpu_device *adev, struct amdgpu_vm *vm, struct dma_fence **fence) { struct amdgpu_bo_va_mapping *mapping; + uint64_t init_pte_value = 0; struct dma_fence *f = NULL; int r; - uint64_t init_pte_value = 0; while (!list_empty(&vm->freed)) { mapping = list_first_entry(&vm->freed, struct amdgpu_bo_va_mapping, list); list_del(&mapping->list); - if (vm->pte_support_ats) + if (vm->pte_support_ats && + mapping->start < AMDGPU_GMC_HOLE_START) init_pte_value = AMDGPU_PTE_DEFAULT_ATC; r = amdgpu_vm_bo_update_mapping(adev, NULL, NULL, vm, @@ -1661,35 +2288,38 @@ int amdgpu_vm_clear_freed(struct amdgpu_device *adev, * * @adev: amdgpu_device pointer * @vm: requested vm - * @sync: sync object to add fences to * * Make sure all BOs which are moved are updated in the PTs. - * Returns 0 for success. + * + * Returns: + * 0 for success. * * PTs have to be reserved! */ int amdgpu_vm_handle_moved(struct amdgpu_device *adev, struct amdgpu_vm *vm) { + struct amdgpu_bo_va *bo_va, *tmp; + struct reservation_object *resv; bool clear; - int r = 0; - - spin_lock(&vm->status_lock); - while (!list_empty(&vm->moved)) { - struct amdgpu_bo_va *bo_va; - struct reservation_object *resv; + int r; - 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->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->invalidated_lock); + while (!list_empty(&vm->invalidated)) { + bo_va = list_first_entry(&vm->invalidated, struct amdgpu_bo_va, + base.vm_status); resv = bo_va->base.bo->tbo.resv; + spin_unlock(&vm->invalidated_lock); - /* Per VM BOs never need to bo cleared in the page tables */ - if (resv == vm->root.base.bo->tbo.resv) - clear = false; /* Try to reserve the BO to avoid clearing its ptes */ - else if (!amdgpu_vm_debug && reservation_object_trylock(resv)) + if (!amdgpu_vm_debug && reservation_object_trylock(resv)) clear = false; /* Somebody else is using the BO right now */ else @@ -1699,14 +2329,13 @@ int amdgpu_vm_handle_moved(struct amdgpu_device *adev, if (r) return r; - if (!clear && resv != vm->root.base.bo->tbo.resv) + if (!clear) reservation_object_unlock(resv); - - spin_lock(&vm->status_lock); + spin_lock(&vm->invalidated_lock); } - spin_unlock(&vm->status_lock); + spin_unlock(&vm->invalidated_lock); - return r; + return 0; } /** @@ -1718,7 +2347,9 @@ int amdgpu_vm_handle_moved(struct amdgpu_device *adev, * * Add @bo into the requested vm. * Add @bo to the list of bos associated with the vm - * Returns newly added bo_va or NULL for failure + * + * Returns: + * Newly added bo_va or NULL for failure * * Object has to be reserved! */ @@ -1732,36 +2363,12 @@ struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev, if (bo_va == NULL) { return NULL; } - bo_va->base.vm = vm; - bo_va->base.bo = bo; - INIT_LIST_HEAD(&bo_va->base.bo_list); - INIT_LIST_HEAD(&bo_va->base.vm_status); + amdgpu_vm_bo_base_init(&bo_va->base, vm, bo); bo_va->ref_count = 1; INIT_LIST_HEAD(&bo_va->valids); INIT_LIST_HEAD(&bo_va->invalids); - if (!bo) - return bo_va; - - list_add_tail(&bo_va->base.bo_list, &bo->va); - - if (bo->tbo.resv != vm->root.base.bo->tbo.resv) - return bo_va; - - if (bo->preferred_domains & - amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type)) - return bo_va; - - /* - * 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. - * */ - spin_lock(&vm->status_lock); - list_move_tail(&bo_va->base.vm_status, &vm->evicted); - spin_unlock(&vm->status_lock); - return bo_va; } @@ -1789,11 +2396,9 @@ static void amdgpu_vm_bo_insert_map(struct amdgpu_device *adev, if (mapping->flags & AMDGPU_PTE_PRT) amdgpu_vm_prt_get(adev); - if (bo && bo->tbo.resv == vm->root.base.bo->tbo.resv) { - spin_lock(&vm->status_lock); - if (list_empty(&bo_va->base.vm_status)) - list_add(&bo_va->base.vm_status, &vm->moved); - spin_unlock(&vm->status_lock); + if (bo && bo->tbo.resv == vm->root.base.bo->tbo.resv && + !bo_va->base.moved) { + list_move(&bo_va->base.vm_status, &vm->moved); } trace_amdgpu_vm_bo_map(bo_va, mapping); } @@ -1805,10 +2410,13 @@ static void amdgpu_vm_bo_insert_map(struct amdgpu_device *adev, * @bo_va: bo_va to store the address * @saddr: where to map the BO * @offset: requested offset in the BO + * @size: BO size in bytes * @flags: attributes of pages (read/write/valid/etc.) * * Add a mapping of the BO at the specefied addr into the VM. - * Returns 0 for success, error for failure. + * + * Returns: + * 0 for success, error for failure. * * Object has to be reserved and unreserved outside! */ @@ -1866,11 +2474,14 @@ int amdgpu_vm_bo_map(struct amdgpu_device *adev, * @bo_va: bo_va to store the address * @saddr: where to map the BO * @offset: requested offset in the BO + * @size: BO size in bytes * @flags: attributes of pages (read/write/valid/etc.) * * Add a mapping of the BO at the specefied addr into the VM. Replace existing * mappings as we do so. - * Returns 0 for success, error for failure. + * + * Returns: + * 0 for success, error for failure. * * Object has to be reserved and unreserved outside! */ @@ -1927,7 +2538,9 @@ int amdgpu_vm_bo_replace_map(struct amdgpu_device *adev, * @saddr: where to the BO is mapped * * Remove a mapping of the BO at the specefied addr from the VM. - * Returns 0 for success, error for failure. + * + * Returns: + * 0 for success, error for failure. * * Object has to be reserved and unreserved outside! */ @@ -1981,7 +2594,9 @@ int amdgpu_vm_bo_unmap(struct amdgpu_device *adev, * @size: size of the range * * Remove all mappings in a range, split them as appropriate. - * Returns 0 for success, error for failure. + * + * Returns: + * 0 for success, error for failure. */ int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev, struct amdgpu_vm *vm, @@ -2017,7 +2632,8 @@ int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev, before->last = saddr - 1; before->offset = tmp->offset; before->flags = tmp->flags; - list_add(&before->list, &tmp->list); + before->bo_va = tmp->bo_va; + list_add(&before->list, &tmp->bo_va->invalids); } /* Remember mapping split at the end */ @@ -2027,7 +2643,8 @@ int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev, after->offset = tmp->offset; after->offset += after->start - tmp->start; after->flags = tmp->flags; - list_add(&after->list, &tmp->list); + after->bo_va = tmp->bo_va; + list_add(&after->list, &tmp->bo_va->invalids); } list_del(&tmp->list); @@ -2076,8 +2693,13 @@ int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev, * amdgpu_vm_bo_lookup_mapping - find mapping by address * * @vm: the requested VM + * @addr: the address * * Find a mapping by it's address. + * + * Returns: + * The amdgpu_bo_va_mapping matching for addr or NULL + * */ struct amdgpu_bo_va_mapping *amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm, uint64_t addr) @@ -2086,6 +2708,35 @@ struct amdgpu_bo_va_mapping *amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm, } /** + * amdgpu_vm_bo_trace_cs - trace all reserved mappings + * + * @vm: the requested vm + * @ticket: CS ticket + * + * Trace all mappings of BOs reserved during a command submission. + */ +void amdgpu_vm_bo_trace_cs(struct amdgpu_vm *vm, struct ww_acquire_ctx *ticket) +{ + struct amdgpu_bo_va_mapping *mapping; + + if (!trace_amdgpu_vm_bo_cs_enabled()) + return; + + for (mapping = amdgpu_vm_it_iter_first(&vm->va, 0, U64_MAX); mapping; + mapping = amdgpu_vm_it_iter_next(mapping, 0, U64_MAX)) { + if (mapping->bo_va && mapping->bo_va->base.bo) { + struct amdgpu_bo *bo; + + bo = mapping->bo_va->base.bo; + if (READ_ONCE(bo->tbo.resv->lock.ctx) != ticket) + continue; + } + + trace_amdgpu_vm_bo_cs(mapping); + } +} + +/** * amdgpu_vm_bo_rmv - remove a bo to a specific vm * * @adev: amdgpu_device pointer @@ -2099,13 +2750,27 @@ void amdgpu_vm_bo_rmv(struct amdgpu_device *adev, struct amdgpu_bo_va *bo_va) { struct amdgpu_bo_va_mapping *mapping, *next; + struct amdgpu_bo *bo = bo_va->base.bo; struct amdgpu_vm *vm = bo_va->base.vm; + struct amdgpu_vm_bo_base **base; - list_del(&bo_va->base.bo_list); + if (bo) { + if (bo->tbo.resv == vm->root.base.bo->tbo.resv) + vm->bulk_moveable = false; - spin_lock(&vm->status_lock); + for (base = &bo_va->base.bo->vm_bo; *base; + base = &(*base)->next) { + if (*base != &bo_va->base) + continue; + + *base = bo_va->base.next; + break; + } + } + + spin_lock(&vm->invalidated_lock); list_del(&bo_va->base.vm_status); - spin_unlock(&vm->status_lock); + spin_unlock(&vm->invalidated_lock); list_for_each_entry_safe(mapping, next, &bo_va->valids, list) { list_del(&mapping->list); @@ -2129,8 +2794,8 @@ void amdgpu_vm_bo_rmv(struct amdgpu_device *adev, * amdgpu_vm_bo_invalidate - mark the bo as invalid * * @adev: amdgpu_device pointer - * @vm: requested vm * @bo: amdgpu buffer object + * @evicted: is the BO evicted * * Mark @bo as invalid. */ @@ -2139,36 +2804,39 @@ void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev, { struct amdgpu_vm_bo_base *bo_base; - list_for_each_entry(bo_base, &bo->va, bo_list) { + /* shadow bo doesn't have bo base, its validation needs its parent */ + if (bo->parent && bo->parent->shadow == bo) + bo = bo->parent; + + for (bo_base = bo->vm_bo; bo_base; bo_base = bo_base->next) { struct amdgpu_vm *vm = bo_base->vm; - 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); + amdgpu_vm_bo_evicted(bo_base); 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); + if (bo_base->moved) continue; - } + bo_base->moved = true; - spin_lock(&bo_base->vm->status_lock); - if (list_empty(&bo_base->vm_status)) - list_add(&bo_base->vm_status, &vm->moved); - spin_unlock(&bo_base->vm->status_lock); + if (bo->tbo.type == ttm_bo_type_kernel) + amdgpu_vm_bo_relocated(bo_base); + else if (bo->tbo.resv == vm->root.base.bo->tbo.resv) + amdgpu_vm_bo_moved(bo_base); + else + amdgpu_vm_bo_invalidated(bo_base); } } +/** + * amdgpu_vm_get_block_size - calculate VM page table size as power of two + * + * @vm_size: VM size + * + * Returns: + * VM page table as power of two + */ static uint32_t amdgpu_vm_get_block_size(uint64_t vm_size) { /* Total bits covered by PD + PTs */ @@ -2186,24 +2854,52 @@ static uint32_t amdgpu_vm_get_block_size(uint64_t vm_size) * amdgpu_vm_adjust_size - adjust vm size, block size and fragment size * * @adev: amdgpu_device pointer - * @vm_size: the default vm size if it's set auto + * @min_vm_size: the minimum vm size in GB if it's set auto + * @fragment_size_default: Default PTE fragment size + * @max_level: max VMPT level + * @max_bits: max address space size in bits + * */ -void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t vm_size, +void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t min_vm_size, uint32_t fragment_size_default, unsigned max_level, unsigned max_bits) { + unsigned int max_size = 1 << (max_bits - 30); + unsigned int vm_size; uint64_t tmp; /* adjust vm size first */ if (amdgpu_vm_size != -1) { - unsigned max_size = 1 << (max_bits - 30); - vm_size = amdgpu_vm_size; if (vm_size > max_size) { dev_warn(adev->dev, "VM size (%d) too large, max is %u GB\n", amdgpu_vm_size, max_size); vm_size = max_size; } + } else { + struct sysinfo si; + unsigned int phys_ram_gb; + + /* Optimal VM size depends on the amount of physical + * RAM available. Underlying requirements and + * assumptions: + * + * - Need to map system memory and VRAM from all GPUs + * - VRAM from other GPUs not known here + * - Assume VRAM <= system memory + * - On GFX8 and older, VM space can be segmented for + * different MTYPEs + * - Need to allow room for fragmentation, guard pages etc. + * + * This adds up to a rough guess of system memory x3. + * Round up to power of two to maximize the available + * VM size with the given page table size. + */ + si_meminfo(&si); + phys_ram_gb = ((uint64_t)si.totalram * si.mem_unit + + (1 << 30) - 1) >> 30; + vm_size = roundup_pow_of_two( + min(max(phys_ram_gb * 3, min_vm_size), max_size)); } adev->vm_manager.max_pfn = (uint64_t)vm_size << 18; @@ -2248,44 +2944,56 @@ void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t vm_size, adev->vm_manager.fragment_size); } +static struct amdgpu_retryfault_hashtable *init_fault_hash(void) +{ + struct amdgpu_retryfault_hashtable *fault_hash; + + fault_hash = kmalloc(sizeof(*fault_hash), GFP_KERNEL); + if (!fault_hash) + return fault_hash; + + INIT_CHASH_TABLE(fault_hash->hash, + AMDGPU_PAGEFAULT_HASH_BITS, 8, 0); + spin_lock_init(&fault_hash->lock); + fault_hash->count = 0; + + return fault_hash; +} + /** * amdgpu_vm_init - initialize a vm instance * * @adev: amdgpu_device pointer * @vm: requested vm * @vm_context: Indicates if it GFX or Compute context + * @pasid: Process address space identifier * * Init @vm fields. + * + * Returns: + * 0 for success, error for failure. */ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm, int vm_context, unsigned int pasid) { - const unsigned align = min(AMDGPU_VM_PTB_ALIGN_SIZE, - AMDGPU_VM_PTE_COUNT(adev) * 8); - uint64_t init_pde_value = 0, flags; - unsigned ring_instance; - struct amdgpu_ring *ring; - struct drm_sched_rq *rq; - unsigned long size; + struct amdgpu_bo_param bp; + struct amdgpu_bo *root; int r, i; 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); INIT_LIST_HEAD(&vm->moved); + INIT_LIST_HEAD(&vm->idle); + INIT_LIST_HEAD(&vm->invalidated); + spin_lock_init(&vm->invalidated_lock); INIT_LIST_HEAD(&vm->freed); /* create scheduler entity for page table updates */ - - ring_instance = atomic_inc_return(&adev->vm_manager.vm_pte_next_ring); - ring_instance %= adev->vm_manager.vm_pte_num_rings; - ring = adev->vm_manager.vm_pte_rings[ring_instance]; - rq = &ring->sched.sched_rq[DRM_SCHED_PRIORITY_KERNEL]; - r = drm_sched_entity_init(&ring->sched, &vm->entity, - rq, amdgpu_sched_jobs, NULL); + r = drm_sched_entity_init(&vm->entity, adev->vm_manager.vm_pte_rqs, + adev->vm_manager.vm_pte_num_rqs, NULL); if (r) return r; @@ -2295,43 +3003,36 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm, vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode & AMDGPU_VM_USE_CPU_FOR_COMPUTE); - if (adev->asic_type == CHIP_RAVEN) { + if (adev->asic_type == CHIP_RAVEN) vm->pte_support_ats = true; - init_pde_value = AMDGPU_PTE_DEFAULT_ATC - | AMDGPU_PDE_PTE; - - } - } else + } else { vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode & AMDGPU_VM_USE_CPU_FOR_GFX); + } DRM_DEBUG_DRIVER("VM update mode is %s\n", vm->use_cpu_for_update ? "CPU" : "SDMA"); - WARN_ONCE((vm->use_cpu_for_update & !amdgpu_vm_is_large_bar(adev)), + WARN_ONCE((vm->use_cpu_for_update & !amdgpu_gmc_vram_full_visible(&adev->gmc)), "CPU update of VM recommended only for large BAR system\n"); vm->last_update = NULL; - flags = AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS | - AMDGPU_GEM_CREATE_VRAM_CLEARED; - if (vm->use_cpu_for_update) - flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED; - else - flags |= (AMDGPU_GEM_CREATE_NO_CPU_ACCESS | - AMDGPU_GEM_CREATE_SHADOW); - - size = amdgpu_vm_bo_size(adev, adev->vm_manager.root_level); - r = amdgpu_bo_create(adev, size, align, true, AMDGPU_GEM_DOMAIN_VRAM, - flags, NULL, NULL, init_pde_value, - &vm->root.base.bo); + amdgpu_vm_bo_param(adev, vm, adev->vm_manager.root_level, &bp); + if (vm_context == AMDGPU_VM_CONTEXT_COMPUTE) + bp.flags &= ~AMDGPU_GEM_CREATE_SHADOW; + r = amdgpu_bo_create(adev, &bp, &root); if (r) goto error_free_sched_entity; - r = amdgpu_bo_reserve(vm->root.base.bo, true); + r = amdgpu_bo_reserve(root, true); if (r) goto error_free_root; - vm->root.base.vm = vm; - list_add_tail(&vm->root.base.bo_list, &vm->root.base.bo->va); - list_add_tail(&vm->root.base.vm_status, &vm->evicted); + r = amdgpu_vm_clear_bo(adev, vm, root, + adev->vm_manager.root_level, + vm->pte_support_ats); + if (r) + goto error_unreserve; + + amdgpu_vm_bo_base_init(&vm->root.base, vm, root); amdgpu_bo_unreserve(vm->root.base.bo); if (pasid) { @@ -2347,50 +3048,152 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm, vm->pasid = pasid; } + vm->fault_hash = init_fault_hash(); + if (!vm->fault_hash) { + r = -ENOMEM; + goto error_free_root; + } + INIT_KFIFO(vm->faults); vm->fault_credit = 16; return 0; +error_unreserve: + amdgpu_bo_unreserve(vm->root.base.bo); + error_free_root: amdgpu_bo_unref(&vm->root.base.bo->shadow); amdgpu_bo_unref(&vm->root.base.bo); vm->root.base.bo = NULL; error_free_sched_entity: - drm_sched_entity_fini(&ring->sched, &vm->entity); + drm_sched_entity_destroy(&vm->entity); return r; } /** - * amdgpu_vm_free_levels - free PD/PT levels + * amdgpu_vm_make_compute - Turn a GFX VM into a compute VM * - * @adev: amdgpu device structure - * @parent: PD/PT starting level to free - * @level: level of parent structure + * @adev: amdgpu_device pointer + * @vm: requested vm * - * Free the page directory or page table level and all sub levels. + * This only works on GFX VMs that don't have any BOs added and no + * page tables allocated yet. + * + * Changes the following VM parameters: + * - use_cpu_for_update + * - pte_supports_ats + * - pasid (old PASID is released, because compute manages its own PASIDs) + * + * Reinitializes the page directory to reflect the changed ATS + * setting. + * + * Returns: + * 0 for success, -errno for errors. */ -static void amdgpu_vm_free_levels(struct amdgpu_device *adev, - struct amdgpu_vm_pt *parent, - unsigned level) +int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm, unsigned int pasid) { - unsigned i, num_entries = amdgpu_vm_num_entries(adev, level); + bool pte_support_ats = (adev->asic_type == CHIP_RAVEN); + int r; + + r = amdgpu_bo_reserve(vm->root.base.bo, true); + if (r) + return r; + + /* Sanity checks */ + if (!RB_EMPTY_ROOT(&vm->va.rb_root) || vm->root.entries) { + r = -EINVAL; + goto unreserve_bo; + } + + if (pasid) { + unsigned long flags; + + spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags); + r = idr_alloc(&adev->vm_manager.pasid_idr, vm, pasid, pasid + 1, + GFP_ATOMIC); + spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags); + + if (r == -ENOSPC) + goto unreserve_bo; + r = 0; + } + + /* Check if PD needs to be reinitialized and do it before + * changing any other state, in case it fails. + */ + if (pte_support_ats != vm->pte_support_ats) { + r = amdgpu_vm_clear_bo(adev, vm, vm->root.base.bo, + adev->vm_manager.root_level, + pte_support_ats); + if (r) + goto free_idr; + } + + /* Update VM state */ + vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode & + AMDGPU_VM_USE_CPU_FOR_COMPUTE); + vm->pte_support_ats = pte_support_ats; + DRM_DEBUG_DRIVER("VM update mode is %s\n", + vm->use_cpu_for_update ? "CPU" : "SDMA"); + WARN_ONCE((vm->use_cpu_for_update & !amdgpu_gmc_vram_full_visible(&adev->gmc)), + "CPU update of VM recommended only for large BAR system\n"); + + if (vm->pasid) { + unsigned long flags; + + spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags); + idr_remove(&adev->vm_manager.pasid_idr, vm->pasid); + spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags); + + /* Free the original amdgpu allocated pasid + * Will be replaced with kfd allocated pasid + */ + amdgpu_pasid_free(vm->pasid); + vm->pasid = 0; + } + + /* Free the shadow bo for compute VM */ + amdgpu_bo_unref(&vm->root.base.bo->shadow); + + if (pasid) + vm->pasid = pasid; + + goto unreserve_bo; + +free_idr: + if (pasid) { + unsigned long flags; - if (parent->base.bo) { - list_del(&parent->base.bo_list); - list_del(&parent->base.vm_status); - amdgpu_bo_unref(&parent->base.bo->shadow); - amdgpu_bo_unref(&parent->base.bo); + spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags); + idr_remove(&adev->vm_manager.pasid_idr, pasid); + spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags); } +unreserve_bo: + amdgpu_bo_unreserve(vm->root.base.bo); + return r; +} - if (parent->entries) - for (i = 0; i < num_entries; i++) - amdgpu_vm_free_levels(adev, &parent->entries[i], - level + 1); +/** + * amdgpu_vm_release_compute - release a compute vm + * @adev: amdgpu_device pointer + * @vm: a vm turned into compute vm by calling amdgpu_vm_make_compute + * + * This is a correspondant of amdgpu_vm_make_compute. It decouples compute + * pasid from vm. Compute should stop use of vm after this call. + */ +void amdgpu_vm_release_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm) +{ + if (vm->pasid) { + unsigned long flags; - kvfree(parent->entries); + spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags); + idr_remove(&adev->vm_manager.pasid_idr, vm->pasid); + spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags); + } + vm->pasid = 0; } /** @@ -2405,14 +3208,16 @@ static void amdgpu_vm_free_levels(struct amdgpu_device *adev, void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm) { struct amdgpu_bo_va_mapping *mapping, *tmp; - bool prt_fini_needed = !!adev->gart.gart_funcs->set_prt; + bool prt_fini_needed = !!adev->gmc.gmc_funcs->set_prt; struct amdgpu_bo *root; u64 fault; int i, r; + amdgpu_amdkfd_gpuvm_destroy_cb(adev, vm); + /* Clear pending page faults from IH when the VM is destroyed */ while (kfifo_get(&vm->faults, &fault)) - amdgpu_ih_clear_fault(adev, fault); + amdgpu_vm_clear_fault(vm->fault_hash, fault); if (vm->pasid) { unsigned long flags; @@ -2422,15 +3227,20 @@ void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm) spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags); } - drm_sched_entity_fini(vm->entity.sched, &vm->entity); + kfree(vm->fault_hash); + vm->fault_hash = NULL; + + drm_sched_entity_destroy(&vm->entity); if (!RB_EMPTY_ROOT(&vm->va.rb_root)) { dev_err(adev->dev, "still active bo inside vm\n"); } rbtree_postorder_for_each_entry_safe(mapping, tmp, &vm->va.rb_root, rb) { + /* Don't remove the mapping here, we don't want to trigger a + * rebalance and the tree is about to be destroyed anyway. + */ list_del(&mapping->list); - amdgpu_vm_it_remove(mapping, &vm->va); kfree(mapping); } list_for_each_entry_safe(mapping, tmp, &vm->freed, list) { @@ -2448,8 +3258,7 @@ void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm) if (r) { dev_err(adev->dev, "Leaking page tables because BO reservation failed\n"); } else { - amdgpu_vm_free_levels(adev, &vm->root, - adev->vm_manager.root_level); + amdgpu_vm_free_pts(adev, vm); amdgpu_bo_unreserve(root); } amdgpu_bo_unref(&root); @@ -2464,8 +3273,10 @@ void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm) * @adev: amdgpu_device pointer * @pasid: PASID do identify the VM * - * This function is expected to be called in interrupt context. Returns - * true if there was fault credit, false otherwise + * This function is expected to be called in interrupt context. + * + * Returns: + * True if there was fault credit, false otherwise */ bool amdgpu_vm_pasid_fault_credit(struct amdgpu_device *adev, unsigned int pasid) @@ -2510,7 +3321,6 @@ void amdgpu_vm_manager_init(struct amdgpu_device *adev) for (i = 0; i < AMDGPU_MAX_RINGS; ++i) adev->vm_manager.seqno[i] = 0; - atomic_set(&adev->vm_manager.vm_pte_next_ring, 0); spin_lock_init(&adev->vm_manager.prt_lock); atomic_set(&adev->vm_manager.num_prt_users, 0); @@ -2519,7 +3329,7 @@ void amdgpu_vm_manager_init(struct amdgpu_device *adev) */ #ifdef CONFIG_X86_64 if (amdgpu_vm_update_mode == -1) { - if (amdgpu_vm_is_large_bar(adev)) + if (amdgpu_gmc_vram_full_visible(&adev->gmc)) adev->vm_manager.vm_update_mode = AMDGPU_VM_USE_CPU_FOR_COMPUTE; else @@ -2549,6 +3359,16 @@ void amdgpu_vm_manager_fini(struct amdgpu_device *adev) amdgpu_vmid_mgr_fini(adev); } +/** + * amdgpu_vm_ioctl - Manages VMID reservation for vm hubs. + * + * @dev: drm device pointer + * @data: drm_amdgpu_vm + * @filp: drm file pointer + * + * Returns: + * 0 for success, -errno for errors. + */ int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) { union drm_amdgpu_vm *args = data; @@ -2572,3 +3392,117 @@ int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) return 0; } + +/** + * amdgpu_vm_get_task_info - Extracts task info for a PASID. + * + * @adev: drm device pointer + * @pasid: PASID identifier for VM + * @task_info: task_info to fill. + */ +void amdgpu_vm_get_task_info(struct amdgpu_device *adev, unsigned int pasid, + struct amdgpu_task_info *task_info) +{ + struct amdgpu_vm *vm; + + spin_lock(&adev->vm_manager.pasid_lock); + + vm = idr_find(&adev->vm_manager.pasid_idr, pasid); + if (vm) + *task_info = vm->task_info; + + spin_unlock(&adev->vm_manager.pasid_lock); +} + +/** + * amdgpu_vm_set_task_info - Sets VMs task info. + * + * @vm: vm for which to set the info + */ +void amdgpu_vm_set_task_info(struct amdgpu_vm *vm) +{ + if (!vm->task_info.pid) { + vm->task_info.pid = current->pid; + get_task_comm(vm->task_info.task_name, current); + + if (current->group_leader->mm == current->mm) { + vm->task_info.tgid = current->group_leader->pid; + get_task_comm(vm->task_info.process_name, current->group_leader); + } + } +} + +/** + * amdgpu_vm_add_fault - Add a page fault record to fault hash table + * + * @fault_hash: fault hash table + * @key: 64-bit encoding of PASID and address + * + * This should be called when a retry page fault interrupt is + * received. If this is a new page fault, it will be added to a hash + * table. The return value indicates whether this is a new fault, or + * a fault that was already known and is already being handled. + * + * If there are too many pending page faults, this will fail. Retry + * interrupts should be ignored in this case until there is enough + * free space. + * + * Returns 0 if the fault was added, 1 if the fault was already known, + * -ENOSPC if there are too many pending faults. + */ +int amdgpu_vm_add_fault(struct amdgpu_retryfault_hashtable *fault_hash, u64 key) +{ + unsigned long flags; + int r = -ENOSPC; + + if (WARN_ON_ONCE(!fault_hash)) + /* Should be allocated in amdgpu_vm_init + */ + return r; + + spin_lock_irqsave(&fault_hash->lock, flags); + + /* Only let the hash table fill up to 50% for best performance */ + if (fault_hash->count >= (1 << (AMDGPU_PAGEFAULT_HASH_BITS-1))) + goto unlock_out; + + r = chash_table_copy_in(&fault_hash->hash, key, NULL); + if (!r) + fault_hash->count++; + + /* chash_table_copy_in should never fail unless we're losing count */ + WARN_ON_ONCE(r < 0); + +unlock_out: + spin_unlock_irqrestore(&fault_hash->lock, flags); + return r; +} + +/** + * amdgpu_vm_clear_fault - Remove a page fault record + * + * @fault_hash: fault hash table + * @key: 64-bit encoding of PASID and address + * + * This should be called when a page fault has been handled. Any + * future interrupt with this key will be processed as a new + * page fault. + */ +void amdgpu_vm_clear_fault(struct amdgpu_retryfault_hashtable *fault_hash, u64 key) +{ + unsigned long flags; + int r; + + if (!fault_hash) + return; + + spin_lock_irqsave(&fault_hash->lock, flags); + + r = chash_table_remove(&fault_hash->hash, key, NULL); + if (!WARN_ON_ONCE(r < 0)) { + fault_hash->count--; + WARN_ON_ONCE(fault_hash->count < 0); + } + + spin_unlock_irqrestore(&fault_hash->lock, flags); +} |