diff options
Diffstat (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c')
| -rw-r--r-- | drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 3117 |
1 files changed, 2056 insertions, 1061 deletions
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c index be1ab43473c6..b1c24c8fa686 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c @@ -1,3 +1,4 @@ +// SPDX-License-Identifier: MIT /* * Copyright 2014-2018 Advanced Micro Devices, Inc. * @@ -19,30 +20,37 @@ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR * OTHER DEALINGS IN THE SOFTWARE. */ - -#define pr_fmt(fmt) "kfd2kgd: " fmt - +#include <linux/dma-buf.h> #include <linux/list.h> #include <linux/pagemap.h> #include <linux/sched/mm.h> -#include <linux/dma-buf.h> -#include <drm/drmP.h> +#include <linux/sched/task.h> +#include <drm/ttm/ttm_tt.h> + +#include <drm/drm_exec.h> + #include "amdgpu_object.h" +#include "amdgpu_gem.h" #include "amdgpu_vm.h" +#include "amdgpu_hmm.h" #include "amdgpu_amdkfd.h" - -/* Special VM and GART address alignment needed for VI pre-Fiji due to - * a HW bug. - */ -#define VI_BO_SIZE_ALIGN (0x8000) - -/* BO flag to indicate a KFD userptr BO */ -#define AMDGPU_AMDKFD_USERPTR_BO (1ULL << 63) +#include "amdgpu_dma_buf.h" +#include <uapi/linux/kfd_ioctl.h> +#include "amdgpu_xgmi.h" +#include "kfd_priv.h" +#include "kfd_smi_events.h" /* Userptr restore delay, just long enough to allow consecutive VM * changes to accumulate */ #define AMDGPU_USERPTR_RESTORE_DELAY_MS 1 +#define AMDGPU_RESERVE_MEM_LIMIT (3UL << 29) + +/* + * Align VRAM availability to 2MB to avoid fragmentation caused by 4K allocations in the tail 2MB + * BO chunk + */ +#define VRAM_AVAILABLITY_ALIGN (1 << 21) /* Impose limit on how much memory KFD can use */ static struct { @@ -53,12 +61,6 @@ static struct { spinlock_t mem_limit_lock; } kfd_mem_limit; -/* Struct used for amdgpu_amdkfd_bo_validate */ -struct amdgpu_vm_parser { - uint32_t domain; - bool wait; -}; - static const char * const domain_bit_to_string[] = { "CPU", "GTT", @@ -72,26 +74,39 @@ static const char * const domain_bit_to_string[] = { static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work); - -static inline struct amdgpu_device *get_amdgpu_device(struct kgd_dev *kgd) -{ - return (struct amdgpu_device *)kgd; -} - -static bool check_if_add_bo_to_vm(struct amdgpu_vm *avm, +static bool kfd_mem_is_attached(struct amdgpu_vm *avm, struct kgd_mem *mem) { - struct kfd_bo_va_list *entry; + struct kfd_mem_attachment *entry; - list_for_each_entry(entry, &mem->bo_va_list, bo_list) + list_for_each_entry(entry, &mem->attachments, list) if (entry->bo_va->base.vm == avm) - return false; + return true; - return true; + return false; +} + +/** + * reuse_dmamap() - Check whether adev can share the original + * userptr BO + * + * If both adev and bo_adev are in direct mapping or + * in the same iommu group, they can share the original BO. + * + * @adev: Device to which can or cannot share the original BO + * @bo_adev: Device to which allocated BO belongs to + * + * Return: returns true if adev can share original userptr BO, + * false otherwise. + */ +static bool reuse_dmamap(struct amdgpu_device *adev, struct amdgpu_device *bo_adev) +{ + return (adev->ram_is_direct_mapped && bo_adev->ram_is_direct_mapped) || + (adev->dev->iommu_group == bo_adev->dev->iommu_group); } /* Set memory usage limits. Current, limits are - * System (TTM + userptr) memory - 3/4th System RAM + * System (TTM + userptr) memory - 15/16th System RAM * TTM memory - 3/8th System RAM */ void amdgpu_amdkfd_gpuvm_init_mem_limits(void) @@ -99,236 +114,304 @@ void amdgpu_amdkfd_gpuvm_init_mem_limits(void) struct sysinfo si; uint64_t mem; + if (kfd_mem_limit.max_system_mem_limit) + return; + si_meminfo(&si); mem = si.totalram - si.totalhigh; mem *= si.mem_unit; spin_lock_init(&kfd_mem_limit.mem_limit_lock); - kfd_mem_limit.max_system_mem_limit = (mem >> 1) + (mem >> 2); - kfd_mem_limit.max_ttm_mem_limit = (mem >> 1) - (mem >> 3); + kfd_mem_limit.max_system_mem_limit = mem - (mem >> 6); + if (kfd_mem_limit.max_system_mem_limit < 2 * AMDGPU_RESERVE_MEM_LIMIT) + kfd_mem_limit.max_system_mem_limit >>= 1; + else + kfd_mem_limit.max_system_mem_limit -= AMDGPU_RESERVE_MEM_LIMIT; + + kfd_mem_limit.max_ttm_mem_limit = ttm_tt_pages_limit() << PAGE_SHIFT; pr_debug("Kernel memory limit %lluM, TTM limit %lluM\n", (kfd_mem_limit.max_system_mem_limit >> 20), (kfd_mem_limit.max_ttm_mem_limit >> 20)); } -static int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev, - uint64_t size, u32 domain, bool sg) +void amdgpu_amdkfd_reserve_system_mem(uint64_t size) { - size_t acc_size, system_mem_needed, ttm_mem_needed, vram_needed; - uint64_t reserved_for_pt = amdgpu_amdkfd_total_mem_size >> 9; - int ret = 0; + kfd_mem_limit.system_mem_used += size; +} + +/* Estimate page table size needed to represent a given memory size + * + * With 4KB pages, we need one 8 byte PTE for each 4KB of memory + * (factor 512, >> 9). With 2MB pages, we need one 8 byte PTE for 2MB + * of memory (factor 256K, >> 18). ROCm user mode tries to optimize + * for 2MB pages for TLB efficiency. However, small allocations and + * fragmented system memory still need some 4KB pages. We choose a + * compromise that should work in most cases without reserving too + * much memory for page tables unnecessarily (factor 16K, >> 14). + */ - acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size, - sizeof(struct amdgpu_bo)); +#define ESTIMATE_PT_SIZE(mem_size) max(((mem_size) >> 14), AMDGPU_VM_RESERVED_VRAM) + +/** + * amdgpu_amdkfd_reserve_mem_limit() - Decrease available memory by size + * of buffer. + * + * @adev: Device to which allocated BO belongs to + * @size: Size of buffer, in bytes, encapsulated by B0. This should be + * equivalent to amdgpu_bo_size(BO) + * @alloc_flag: Flag used in allocating a BO as noted above + * @xcp_id: xcp_id is used to get xcp from xcp manager, one xcp is + * managed as one compute node in driver for app + * + * Return: + * returns -ENOMEM in case of error, ZERO otherwise + */ +int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev, + uint64_t size, u32 alloc_flag, int8_t xcp_id) +{ + uint64_t reserved_for_pt = + ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size); + struct amdgpu_ras *con = amdgpu_ras_get_context(adev); + uint64_t reserved_for_ras = (con ? con->reserved_pages_in_bytes : 0); + size_t system_mem_needed, ttm_mem_needed, vram_needed; + int ret = 0; + uint64_t vram_size = 0; + system_mem_needed = 0; + ttm_mem_needed = 0; vram_needed = 0; - if (domain == AMDGPU_GEM_DOMAIN_GTT) { - /* TTM GTT memory */ - system_mem_needed = acc_size + size; - ttm_mem_needed = acc_size + size; - } else if (domain == AMDGPU_GEM_DOMAIN_CPU && !sg) { - /* Userptr */ - system_mem_needed = acc_size + size; - ttm_mem_needed = acc_size; - } else { - /* VRAM and SG */ - system_mem_needed = acc_size; - ttm_mem_needed = acc_size; - if (domain == AMDGPU_GEM_DOMAIN_VRAM) - vram_needed = size; + if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_GTT) { + system_mem_needed = size; + ttm_mem_needed = size; + } else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) { + /* + * Conservatively round up the allocation requirement to 2 MB + * to avoid fragmentation caused by 4K allocations in the tail + * 2M BO chunk. + */ + vram_needed = size; + /* + * For GFX 9.4.3, get the VRAM size from XCP structs + */ + if (WARN_ONCE(xcp_id < 0, "invalid XCP ID %d", xcp_id)) + return -EINVAL; + + vram_size = KFD_XCP_MEMORY_SIZE(adev, xcp_id); + if (adev->apu_prefer_gtt) { + system_mem_needed = size; + ttm_mem_needed = size; + } + } else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) { + system_mem_needed = size; + } else if (!(alloc_flag & + (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL | + KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) { + pr_err("%s: Invalid BO type %#x\n", __func__, alloc_flag); + return -ENOMEM; } spin_lock(&kfd_mem_limit.mem_limit_lock); - if ((kfd_mem_limit.system_mem_used + system_mem_needed > - kfd_mem_limit.max_system_mem_limit) || - (kfd_mem_limit.ttm_mem_used + ttm_mem_needed > - kfd_mem_limit.max_ttm_mem_limit) || - (adev->kfd.vram_used + vram_needed > - adev->gmc.real_vram_size - reserved_for_pt)) { + if (kfd_mem_limit.system_mem_used + system_mem_needed > + kfd_mem_limit.max_system_mem_limit) { + pr_debug("Set no_system_mem_limit=1 if using shared memory\n"); + if (!no_system_mem_limit) { + ret = -ENOMEM; + goto release; + } + } + + if (kfd_mem_limit.ttm_mem_used + ttm_mem_needed > + kfd_mem_limit.max_ttm_mem_limit) { ret = -ENOMEM; - } else { - kfd_mem_limit.system_mem_used += system_mem_needed; - kfd_mem_limit.ttm_mem_used += ttm_mem_needed; - adev->kfd.vram_used += vram_needed; + goto release; } + /*if is_app_apu is false and apu_prefer_gtt is true, it is an APU with + * carve out < gtt. In that case, VRAM allocation will go to gtt domain, skip + * VRAM check since ttm_mem_limit check already cover this allocation + */ + + if (adev && xcp_id >= 0 && (!adev->apu_prefer_gtt || adev->gmc.is_app_apu)) { + uint64_t vram_available = + vram_size - reserved_for_pt - reserved_for_ras - + atomic64_read(&adev->vram_pin_size); + if (adev->kfd.vram_used[xcp_id] + vram_needed > vram_available) { + ret = -ENOMEM; + goto release; + } + } + + /* Update memory accounting by decreasing available system + * memory, TTM memory and GPU memory as computed above + */ + WARN_ONCE(vram_needed && !adev, + "adev reference can't be null when vram is used"); + if (adev && xcp_id >= 0) { + adev->kfd.vram_used[xcp_id] += vram_needed; + adev->kfd.vram_used_aligned[xcp_id] += + adev->apu_prefer_gtt ? + vram_needed : + ALIGN(vram_needed, VRAM_AVAILABLITY_ALIGN); + } + kfd_mem_limit.system_mem_used += system_mem_needed; + kfd_mem_limit.ttm_mem_used += ttm_mem_needed; + +release: spin_unlock(&kfd_mem_limit.mem_limit_lock); return ret; } -static void unreserve_mem_limit(struct amdgpu_device *adev, - uint64_t size, u32 domain, bool sg) +void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev, + uint64_t size, u32 alloc_flag, int8_t xcp_id) { - size_t acc_size; - - acc_size = ttm_bo_dma_acc_size(&adev->mman.bdev, size, - sizeof(struct amdgpu_bo)); - spin_lock(&kfd_mem_limit.mem_limit_lock); - if (domain == AMDGPU_GEM_DOMAIN_GTT) { - kfd_mem_limit.system_mem_used -= (acc_size + size); - kfd_mem_limit.ttm_mem_used -= (acc_size + size); - } else if (domain == AMDGPU_GEM_DOMAIN_CPU && !sg) { - kfd_mem_limit.system_mem_used -= (acc_size + size); - kfd_mem_limit.ttm_mem_used -= acc_size; - } else { - kfd_mem_limit.system_mem_used -= acc_size; - kfd_mem_limit.ttm_mem_used -= acc_size; - if (domain == AMDGPU_GEM_DOMAIN_VRAM) { - adev->kfd.vram_used -= size; - WARN_ONCE(adev->kfd.vram_used < 0, - "kfd VRAM memory accounting unbalanced"); + + if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_GTT) { + kfd_mem_limit.system_mem_used -= size; + kfd_mem_limit.ttm_mem_used -= size; + } else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) { + WARN_ONCE(!adev, + "adev reference can't be null when alloc mem flags vram is set"); + if (WARN_ONCE(xcp_id < 0, "invalid XCP ID %d", xcp_id)) + goto release; + + if (adev) { + adev->kfd.vram_used[xcp_id] -= size; + if (adev->apu_prefer_gtt) { + adev->kfd.vram_used_aligned[xcp_id] -= size; + kfd_mem_limit.system_mem_used -= size; + kfd_mem_limit.ttm_mem_used -= size; + } else { + adev->kfd.vram_used_aligned[xcp_id] -= + ALIGN(size, VRAM_AVAILABLITY_ALIGN); + } } - } - WARN_ONCE(kfd_mem_limit.system_mem_used < 0, - "kfd system memory accounting unbalanced"); + } else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) { + kfd_mem_limit.system_mem_used -= size; + } else if (!(alloc_flag & + (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL | + KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) { + pr_err("%s: Invalid BO type %#x\n", __func__, alloc_flag); + goto release; + } + WARN_ONCE(adev && xcp_id >= 0 && adev->kfd.vram_used[xcp_id] < 0, + "KFD VRAM memory accounting unbalanced for xcp: %d", xcp_id); WARN_ONCE(kfd_mem_limit.ttm_mem_used < 0, - "kfd TTM memory accounting unbalanced"); + "KFD TTM memory accounting unbalanced"); + WARN_ONCE(kfd_mem_limit.system_mem_used < 0, + "KFD system memory accounting unbalanced"); +release: spin_unlock(&kfd_mem_limit.mem_limit_lock); } -void amdgpu_amdkfd_unreserve_memory_limit(struct amdgpu_bo *bo) +void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo) { struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); - u32 domain = bo->preferred_domains; - bool sg = (bo->preferred_domains == AMDGPU_GEM_DOMAIN_CPU); + u32 alloc_flags = bo->kfd_bo->alloc_flags; + u64 size = amdgpu_bo_size(bo); - if (bo->flags & AMDGPU_AMDKFD_USERPTR_BO) { - domain = AMDGPU_GEM_DOMAIN_CPU; - sg = false; - } + amdgpu_amdkfd_unreserve_mem_limit(adev, size, alloc_flags, + bo->xcp_id); - unreserve_mem_limit(adev, amdgpu_bo_size(bo), domain, sg); + kfree(bo->kfd_bo); } +/** + * create_dmamap_sg_bo() - Creates a amdgpu_bo object to reflect information + * about USERPTR or DOOREBELL or MMIO BO. + * + * @adev: Device for which dmamap BO is being created + * @mem: BO of peer device that is being DMA mapped. Provides parameters + * in building the dmamap BO + * @bo_out: Output parameter updated with handle of dmamap BO + */ +static int +create_dmamap_sg_bo(struct amdgpu_device *adev, + struct kgd_mem *mem, struct amdgpu_bo **bo_out) +{ + struct drm_gem_object *gem_obj; + int ret; + uint64_t flags = 0; + + ret = amdgpu_bo_reserve(mem->bo, false); + if (ret) + return ret; + + if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) + flags |= mem->bo->flags & (AMDGPU_GEM_CREATE_COHERENT | + AMDGPU_GEM_CREATE_UNCACHED); + + ret = amdgpu_gem_object_create(adev, mem->bo->tbo.base.size, 1, + AMDGPU_GEM_DOMAIN_CPU, AMDGPU_GEM_CREATE_PREEMPTIBLE | flags, + ttm_bo_type_sg, mem->bo->tbo.base.resv, &gem_obj, 0); -/* amdgpu_amdkfd_remove_eviction_fence - Removes eviction fence(s) from BO's + amdgpu_bo_unreserve(mem->bo); + + if (ret) { + pr_err("Error in creating DMA mappable SG BO on domain: %d\n", ret); + return -EINVAL; + } + + *bo_out = gem_to_amdgpu_bo(gem_obj); + (*bo_out)->parent = amdgpu_bo_ref(mem->bo); + return ret; +} + +/* amdgpu_amdkfd_remove_eviction_fence - Removes eviction fence from BO's * reservation object. * * @bo: [IN] Remove eviction fence(s) from this BO - * @ef: [IN] If ef is specified, then this eviction fence is removed if it + * @ef: [IN] This eviction fence is removed if it * is present in the shared list. - * @ef_list: [OUT] Returns list of eviction fences. These fences are removed - * from BO's reservation object shared list. - * @ef_count: [OUT] Number of fences in ef_list. * - * NOTE: If called with ef_list, then amdgpu_amdkfd_add_eviction_fence must be - * called to restore the eviction fences and to avoid memory leak. This is - * useful for shared BOs. * NOTE: Must be called with BO reserved i.e. bo->tbo.resv->lock held. */ static int amdgpu_amdkfd_remove_eviction_fence(struct amdgpu_bo *bo, - struct amdgpu_amdkfd_fence *ef, - struct amdgpu_amdkfd_fence ***ef_list, - unsigned int *ef_count) + struct amdgpu_amdkfd_fence *ef) { - struct reservation_object *resv = bo->tbo.resv; - struct reservation_object_list *old, *new; - unsigned int i, j, k; + struct dma_fence *replacement; - if (!ef && !ef_list) + if (!ef) return -EINVAL; - if (ef_list) { - *ef_list = NULL; - *ef_count = 0; - } - - old = reservation_object_get_list(resv); - if (!old) - return 0; - - new = kmalloc(offsetof(typeof(*new), shared[old->shared_max]), - GFP_KERNEL); - if (!new) - return -ENOMEM; - - /* Go through all the shared fences in the resevation object and sort - * the interesting ones to the end of the list. + /* TODO: Instead of block before we should use the fence of the page + * table update and TLB flush here directly. */ - for (i = 0, j = old->shared_count, k = 0; i < old->shared_count; ++i) { - struct dma_fence *f; - - f = rcu_dereference_protected(old->shared[i], - reservation_object_held(resv)); - - if ((ef && f->context == ef->base.context) || - (!ef && to_amdgpu_amdkfd_fence(f))) - RCU_INIT_POINTER(new->shared[--j], f); - else - RCU_INIT_POINTER(new->shared[k++], f); - } - new->shared_max = old->shared_max; - new->shared_count = k; - - if (!ef) { - unsigned int count = old->shared_count - j; - - /* Alloc memory for count number of eviction fence pointers. - * Fill the ef_list array and ef_count - */ - *ef_list = kcalloc(count, sizeof(**ef_list), GFP_KERNEL); - *ef_count = count; - - if (!*ef_list) { - kfree(new); - return -ENOMEM; - } - } - - /* Install the new fence list, seqcount provides the barriers */ - preempt_disable(); - write_seqcount_begin(&resv->seq); - RCU_INIT_POINTER(resv->fence, new); - write_seqcount_end(&resv->seq); - preempt_enable(); - - /* Drop the references to the removed fences or move them to ef_list */ - for (i = j, k = 0; i < old->shared_count; ++i) { - struct dma_fence *f; - - f = rcu_dereference_protected(new->shared[i], - reservation_object_held(resv)); - if (!ef) - (*ef_list)[k++] = to_amdgpu_amdkfd_fence(f); - else - dma_fence_put(f); - } - kfree_rcu(old, rcu); - + replacement = dma_fence_get_stub(); + dma_resv_replace_fences(bo->tbo.base.resv, ef->base.context, + replacement, DMA_RESV_USAGE_BOOKKEEP); + dma_fence_put(replacement); return 0; } -/* amdgpu_amdkfd_add_eviction_fence - Adds eviction fence(s) back into BO's - * reservation object. - * - * @bo: [IN] Add eviction fences to this BO - * @ef_list: [IN] List of eviction fences to be added - * @ef_count: [IN] Number of fences in ef_list. +/** + * amdgpu_amdkfd_remove_all_eviction_fences - Remove all eviction fences + * @bo: the BO where to remove the evictions fences from. * - * NOTE: Must call amdgpu_amdkfd_remove_eviction_fence before calling this - * function. + * This functions should only be used on release when all references to the BO + * are already dropped. We remove the eviction fence from the private copy of + * the dma_resv object here since that is what is used during release to + * determine of the BO is idle or not. */ -static void amdgpu_amdkfd_add_eviction_fence(struct amdgpu_bo *bo, - struct amdgpu_amdkfd_fence **ef_list, - unsigned int ef_count) +void amdgpu_amdkfd_remove_all_eviction_fences(struct amdgpu_bo *bo) { - int i; + struct dma_resv *resv = &bo->tbo.base._resv; + struct dma_fence *fence, *stub; + struct dma_resv_iter cursor; - if (!ef_list || !ef_count) - return; + dma_resv_assert_held(resv); - for (i = 0; i < ef_count; i++) { - amdgpu_bo_fence(bo, &ef_list[i]->base, true); - /* Re-adding the fence takes an additional reference. Drop that - * reference. - */ - dma_fence_put(&ef_list[i]->base); - } + stub = dma_fence_get_stub(); + dma_resv_for_each_fence(&cursor, resv, DMA_RESV_USAGE_BOOKKEEP, fence) { + if (!to_amdgpu_amdkfd_fence(fence)) + continue; - kfree(ef_list); + dma_resv_replace_fences(resv, fence->context, stub, + DMA_RESV_USAGE_BOOKKEEP); + } + dma_fence_put(stub); } static int amdgpu_amdkfd_bo_validate(struct amdgpu_bo *bo, uint32_t domain, @@ -341,33 +424,51 @@ static int amdgpu_amdkfd_bo_validate(struct amdgpu_bo *bo, uint32_t domain, "Called with userptr BO")) return -EINVAL; + /* bo has been pinned, not need validate it */ + if (bo->tbo.pin_count) + return 0; + amdgpu_bo_placement_from_domain(bo, domain); ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); if (ret) goto validate_fail; - if (wait) { - struct amdgpu_amdkfd_fence **ef_list; - unsigned int ef_count; - - ret = amdgpu_amdkfd_remove_eviction_fence(bo, NULL, &ef_list, - &ef_count); - if (ret) - goto validate_fail; - - ttm_bo_wait(&bo->tbo, false, false); - amdgpu_amdkfd_add_eviction_fence(bo, ef_list, ef_count); - } + if (wait) + amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false); validate_fail: return ret; } -static int amdgpu_amdkfd_validate(void *param, struct amdgpu_bo *bo) +int amdgpu_amdkfd_bo_validate_and_fence(struct amdgpu_bo *bo, + uint32_t domain, + struct dma_fence *fence) { - struct amdgpu_vm_parser *p = param; + int ret = amdgpu_bo_reserve(bo, false); + + if (ret) + return ret; + + ret = amdgpu_amdkfd_bo_validate(bo, domain, true); + if (ret) + goto unreserve_out; + + ret = dma_resv_reserve_fences(bo->tbo.base.resv, 1); + if (ret) + goto unreserve_out; - return amdgpu_amdkfd_bo_validate(bo, p->domain, p->wait); + dma_resv_add_fence(bo->tbo.base.resv, fence, + DMA_RESV_USAGE_BOOKKEEP); + +unreserve_out: + amdgpu_bo_unreserve(bo); + + return ret; +} + +static int amdgpu_amdkfd_validate_vm_bo(void *_unused, struct amdgpu_bo *bo) +{ + return amdgpu_amdkfd_bo_validate(bo, bo->allowed_domains, false); } /* vm_validate_pt_pd_bos - Validate page table and directory BOs @@ -377,173 +478,564 @@ static int amdgpu_amdkfd_validate(void *param, struct amdgpu_bo *bo) * again. Page directories are only updated after updating page * tables. */ -static int vm_validate_pt_pd_bos(struct amdgpu_vm *vm) +static int vm_validate_pt_pd_bos(struct amdgpu_vm *vm, + struct ww_acquire_ctx *ticket) { - struct amdgpu_bo *pd = vm->root.base.bo; + struct amdgpu_bo *pd = vm->root.bo; struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev); - struct amdgpu_vm_parser param; int ret; - param.domain = AMDGPU_GEM_DOMAIN_VRAM; - param.wait = false; - - ret = amdgpu_vm_validate_pt_bos(adev, vm, amdgpu_amdkfd_validate, - ¶m); + ret = amdgpu_vm_validate(adev, vm, ticket, + amdgpu_amdkfd_validate_vm_bo, NULL); if (ret) { - pr_err("amdgpu: failed to validate PT BOs\n"); + pr_err("failed to validate PT BOs\n"); return ret; } - ret = amdgpu_amdkfd_validate(¶m, pd); - if (ret) { - pr_err("amdgpu: failed to validate PD\n"); + vm->pd_phys_addr = amdgpu_gmc_pd_addr(vm->root.bo); + + return 0; +} + +static int vm_update_pds(struct amdgpu_vm *vm, struct amdgpu_sync *sync) +{ + struct amdgpu_bo *pd = vm->root.bo; + struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev); + int ret; + + ret = amdgpu_vm_update_pdes(adev, vm, false); + if (ret) return ret; + + return amdgpu_sync_fence(sync, vm->last_update, GFP_KERNEL); +} + +static uint64_t get_pte_flags(struct amdgpu_device *adev, struct amdgpu_vm *vm, + struct kgd_mem *mem) +{ + uint32_t mapping_flags = AMDGPU_VM_PAGE_READABLE | + AMDGPU_VM_MTYPE_DEFAULT; + + if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE) + mapping_flags |= AMDGPU_VM_PAGE_WRITEABLE; + if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE) + mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE; + + return mapping_flags; +} + +/** + * create_sg_table() - Create an sg_table for a contiguous DMA addr range + * @addr: The starting address to point to + * @size: Size of memory area in bytes being pointed to + * + * Allocates an instance of sg_table and initializes it to point to memory + * area specified by input parameters. The address used to build is assumed + * to be DMA mapped, if needed. + * + * DOORBELL or MMIO BOs use only one scatterlist node in their sg_table + * because they are physically contiguous. + * + * Return: Initialized instance of SG Table or NULL + */ +static struct sg_table *create_sg_table(uint64_t addr, uint32_t size) +{ + struct sg_table *sg = kmalloc(sizeof(*sg), GFP_KERNEL); + + if (!sg) + return NULL; + if (sg_alloc_table(sg, 1, GFP_KERNEL)) { + kfree(sg); + return NULL; } + sg_dma_address(sg->sgl) = addr; + sg->sgl->length = size; +#ifdef CONFIG_NEED_SG_DMA_LENGTH + sg->sgl->dma_length = size; +#endif + return sg; +} + +static int +kfd_mem_dmamap_userptr(struct kgd_mem *mem, + struct kfd_mem_attachment *attachment) +{ + enum dma_data_direction direction = + mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? + DMA_BIDIRECTIONAL : DMA_TO_DEVICE; + struct ttm_operation_ctx ctx = {.interruptible = true}; + struct amdgpu_bo *bo = attachment->bo_va->base.bo; + struct amdgpu_device *adev = attachment->adev; + struct ttm_tt *src_ttm = mem->bo->tbo.ttm; + struct ttm_tt *ttm = bo->tbo.ttm; + int ret; - vm->pd_phys_addr = amdgpu_gmc_pd_addr(vm->root.base.bo); + if (WARN_ON(ttm->num_pages != src_ttm->num_pages)) + return -EINVAL; - if (vm->use_cpu_for_update) { - ret = amdgpu_bo_kmap(pd, NULL); - if (ret) { - pr_err("amdgpu: failed to kmap PD, ret=%d\n", ret); - return ret; - } + ttm->sg = kmalloc(sizeof(*ttm->sg), GFP_KERNEL); + if (unlikely(!ttm->sg)) + return -ENOMEM; + + /* Same sequence as in amdgpu_ttm_tt_pin_userptr */ + ret = sg_alloc_table_from_pages(ttm->sg, src_ttm->pages, + ttm->num_pages, 0, + (u64)ttm->num_pages << PAGE_SHIFT, + GFP_KERNEL); + if (unlikely(ret)) + goto free_sg; + + ret = dma_map_sgtable(adev->dev, ttm->sg, direction, 0); + if (unlikely(ret)) + goto release_sg; + + amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT); + ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); + if (ret) + goto unmap_sg; + + return 0; + +unmap_sg: + dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0); +release_sg: + pr_err("DMA map userptr failed: %d\n", ret); + sg_free_table(ttm->sg); +free_sg: + kfree(ttm->sg); + ttm->sg = NULL; + return ret; +} + +static int +kfd_mem_dmamap_dmabuf(struct kfd_mem_attachment *attachment) +{ + struct ttm_operation_ctx ctx = {.interruptible = true}; + struct amdgpu_bo *bo = attachment->bo_va->base.bo; + + amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT); + return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); +} + +/** + * kfd_mem_dmamap_sg_bo() - Create DMA mapped sg_table to access DOORBELL or MMIO BO + * @mem: SG BO of the DOORBELL or MMIO resource on the owning device + * @attachment: Virtual address attachment of the BO on accessing device + * + * An access request from the device that owns DOORBELL does not require DMA mapping. + * This is because the request doesn't go through PCIe root complex i.e. it instead + * loops back. The need to DMA map arises only when accessing peer device's DOORBELL + * + * In contrast, all access requests for MMIO need to be DMA mapped without regard to + * device ownership. This is because access requests for MMIO go through PCIe root + * complex. + * + * This is accomplished in two steps: + * - Obtain DMA mapped address of DOORBELL or MMIO memory that could be used + * in updating requesting device's page table + * - Signal TTM to mark memory pointed to by requesting device's BO as GPU + * accessible. This allows an update of requesting device's page table + * with entries associated with DOOREBELL or MMIO memory + * + * This method is invoked in the following contexts: + * - Mapping of DOORBELL or MMIO BO of same or peer device + * - Validating an evicted DOOREBELL or MMIO BO on device seeking access + * + * Return: ZERO if successful, NON-ZERO otherwise + */ +static int +kfd_mem_dmamap_sg_bo(struct kgd_mem *mem, + struct kfd_mem_attachment *attachment) +{ + struct ttm_operation_ctx ctx = {.interruptible = true}; + struct amdgpu_bo *bo = attachment->bo_va->base.bo; + struct amdgpu_device *adev = attachment->adev; + struct ttm_tt *ttm = bo->tbo.ttm; + enum dma_data_direction dir; + dma_addr_t dma_addr; + bool mmio; + int ret; + + /* Expect SG Table of dmapmap BO to be NULL */ + mmio = (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP); + if (unlikely(ttm->sg)) { + pr_err("SG Table of %d BO for peer device is UNEXPECTEDLY NON-NULL", mmio); + return -EINVAL; + } + + dir = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? + DMA_BIDIRECTIONAL : DMA_TO_DEVICE; + dma_addr = mem->bo->tbo.sg->sgl->dma_address; + pr_debug("%d BO size: %d\n", mmio, mem->bo->tbo.sg->sgl->length); + pr_debug("%d BO address before DMA mapping: %llx\n", mmio, dma_addr); + dma_addr = dma_map_resource(adev->dev, dma_addr, + mem->bo->tbo.sg->sgl->length, dir, DMA_ATTR_SKIP_CPU_SYNC); + ret = dma_mapping_error(adev->dev, dma_addr); + if (unlikely(ret)) + return ret; + pr_debug("%d BO address after DMA mapping: %llx\n", mmio, dma_addr); + + ttm->sg = create_sg_table(dma_addr, mem->bo->tbo.sg->sgl->length); + if (unlikely(!ttm->sg)) { + ret = -ENOMEM; + goto unmap_sg; + } + + amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT); + ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); + if (unlikely(ret)) + goto free_sg; + + return ret; + +free_sg: + sg_free_table(ttm->sg); + kfree(ttm->sg); + ttm->sg = NULL; +unmap_sg: + dma_unmap_resource(adev->dev, dma_addr, mem->bo->tbo.sg->sgl->length, + dir, DMA_ATTR_SKIP_CPU_SYNC); + return ret; +} + +static int +kfd_mem_dmamap_attachment(struct kgd_mem *mem, + struct kfd_mem_attachment *attachment) +{ + switch (attachment->type) { + case KFD_MEM_ATT_SHARED: + return 0; + case KFD_MEM_ATT_USERPTR: + return kfd_mem_dmamap_userptr(mem, attachment); + case KFD_MEM_ATT_DMABUF: + return kfd_mem_dmamap_dmabuf(attachment); + case KFD_MEM_ATT_SG: + return kfd_mem_dmamap_sg_bo(mem, attachment); + default: + WARN_ON_ONCE(1); + } + return -EINVAL; +} + +static void +kfd_mem_dmaunmap_userptr(struct kgd_mem *mem, + struct kfd_mem_attachment *attachment) +{ + enum dma_data_direction direction = + mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? + DMA_BIDIRECTIONAL : DMA_TO_DEVICE; + struct ttm_operation_ctx ctx = {.interruptible = false}; + struct amdgpu_bo *bo = attachment->bo_va->base.bo; + struct amdgpu_device *adev = attachment->adev; + struct ttm_tt *ttm = bo->tbo.ttm; + + if (unlikely(!ttm->sg)) + return; + + amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU); + (void)ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); + + dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0); + sg_free_table(ttm->sg); + kfree(ttm->sg); + ttm->sg = NULL; +} + +static void +kfd_mem_dmaunmap_dmabuf(struct kfd_mem_attachment *attachment) +{ + /* This is a no-op. We don't want to trigger eviction fences when + * unmapping DMABufs. Therefore the invalidation (moving to system + * domain) is done in kfd_mem_dmamap_dmabuf. + */ +} + +/** + * kfd_mem_dmaunmap_sg_bo() - Free DMA mapped sg_table of DOORBELL or MMIO BO + * @mem: SG BO of the DOORBELL or MMIO resource on the owning device + * @attachment: Virtual address attachment of the BO on accessing device + * + * The method performs following steps: + * - Signal TTM to mark memory pointed to by BO as GPU inaccessible + * - Free SG Table that is used to encapsulate DMA mapped memory of + * peer device's DOORBELL or MMIO memory + * + * This method is invoked in the following contexts: + * UNMapping of DOORBELL or MMIO BO on a device having access to its memory + * Eviction of DOOREBELL or MMIO BO on device having access to its memory + * + * Return: void + */ +static void +kfd_mem_dmaunmap_sg_bo(struct kgd_mem *mem, + struct kfd_mem_attachment *attachment) +{ + struct ttm_operation_ctx ctx = {.interruptible = true}; + struct amdgpu_bo *bo = attachment->bo_va->base.bo; + struct amdgpu_device *adev = attachment->adev; + struct ttm_tt *ttm = bo->tbo.ttm; + enum dma_data_direction dir; + + if (unlikely(!ttm->sg)) { + pr_debug("SG Table of BO is NULL"); + return; + } + + amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU); + (void)ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); + + dir = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? + DMA_BIDIRECTIONAL : DMA_TO_DEVICE; + dma_unmap_resource(adev->dev, ttm->sg->sgl->dma_address, + ttm->sg->sgl->length, dir, DMA_ATTR_SKIP_CPU_SYNC); + sg_free_table(ttm->sg); + kfree(ttm->sg); + ttm->sg = NULL; + bo->tbo.sg = NULL; +} + +static void +kfd_mem_dmaunmap_attachment(struct kgd_mem *mem, + struct kfd_mem_attachment *attachment) +{ + switch (attachment->type) { + case KFD_MEM_ATT_SHARED: + break; + case KFD_MEM_ATT_USERPTR: + kfd_mem_dmaunmap_userptr(mem, attachment); + break; + case KFD_MEM_ATT_DMABUF: + kfd_mem_dmaunmap_dmabuf(attachment); + break; + case KFD_MEM_ATT_SG: + kfd_mem_dmaunmap_sg_bo(mem, attachment); + break; + default: + WARN_ON_ONCE(1); + } +} + +static int kfd_mem_export_dmabuf(struct kgd_mem *mem) +{ + if (!mem->dmabuf) { + struct amdgpu_device *bo_adev; + struct dma_buf *dmabuf; + + bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev); + dmabuf = drm_gem_prime_handle_to_dmabuf(&bo_adev->ddev, bo_adev->kfd.client.file, + mem->gem_handle, + mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? + DRM_RDWR : 0); + if (IS_ERR(dmabuf)) + return PTR_ERR(dmabuf); + mem->dmabuf = dmabuf; } return 0; } -static int vm_update_pds(struct amdgpu_vm *vm, struct amdgpu_sync *sync) +static int +kfd_mem_attach_dmabuf(struct amdgpu_device *adev, struct kgd_mem *mem, + struct amdgpu_bo **bo) { - struct amdgpu_bo *pd = vm->root.base.bo; - struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev); + struct drm_gem_object *gobj; int ret; - ret = amdgpu_vm_update_directories(adev, vm); + ret = kfd_mem_export_dmabuf(mem); if (ret) return ret; - return amdgpu_sync_fence(NULL, sync, vm->last_update, false); + gobj = amdgpu_gem_prime_import(adev_to_drm(adev), mem->dmabuf); + if (IS_ERR(gobj)) + return PTR_ERR(gobj); + + *bo = gem_to_amdgpu_bo(gobj); + (*bo)->flags |= AMDGPU_GEM_CREATE_PREEMPTIBLE; + + return 0; } -/* add_bo_to_vm - Add a BO to a VM +/* kfd_mem_attach - Add a BO to a VM * * Everything that needs to bo done only once when a BO is first added * to a VM. It can later be mapped and unmapped many times without * repeating these steps. * + * 0. Create BO for DMA mapping, if needed * 1. Allocate and initialize BO VA entry data structure * 2. Add BO to the VM * 3. Determine ASIC-specific PTE flags * 4. Alloc page tables and directories if needed * 4a. Validate new page tables and directories */ -static int add_bo_to_vm(struct amdgpu_device *adev, struct kgd_mem *mem, - struct amdgpu_vm *vm, bool is_aql, - struct kfd_bo_va_list **p_bo_va_entry) +static int kfd_mem_attach(struct amdgpu_device *adev, struct kgd_mem *mem, + struct amdgpu_vm *vm, bool is_aql) { - int ret; - struct kfd_bo_va_list *bo_va_entry; - struct amdgpu_bo *pd = vm->root.base.bo; - struct amdgpu_bo *bo = mem->bo; + struct amdgpu_device *bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev); + unsigned long bo_size = mem->bo->tbo.base.size; uint64_t va = mem->va; - struct list_head *list_bo_va = &mem->bo_va_list; - unsigned long bo_size = bo->tbo.mem.size; + struct kfd_mem_attachment *attachment[2] = {NULL, NULL}; + struct amdgpu_bo *bo[2] = {NULL, NULL}; + struct amdgpu_bo_va *bo_va; + bool same_hive = false; + int i, ret; if (!va) { pr_err("Invalid VA when adding BO to VM\n"); return -EINVAL; } - if (is_aql) - va += bo_size; - - bo_va_entry = kzalloc(sizeof(*bo_va_entry), GFP_KERNEL); - if (!bo_va_entry) - return -ENOMEM; - - pr_debug("\t add VA 0x%llx - 0x%llx to vm %p\n", va, - va + bo_size, vm); - - /* Add BO to VM internal data structures*/ - bo_va_entry->bo_va = amdgpu_vm_bo_add(adev, vm, bo); - if (!bo_va_entry->bo_va) { - ret = -EINVAL; - pr_err("Failed to add BO object to VM. ret == %d\n", - ret); - goto err_vmadd; + /* Determine access to VRAM, MMIO and DOORBELL BOs of peer devices + * + * The access path of MMIO and DOORBELL BOs of is always over PCIe. + * In contrast the access path of VRAM BOs depens upon the type of + * link that connects the peer device. Access over PCIe is allowed + * if peer device has large BAR. In contrast, access over xGMI is + * allowed for both small and large BAR configurations of peer device + */ + if ((adev != bo_adev && !adev->apu_prefer_gtt) && + ((mem->domain == AMDGPU_GEM_DOMAIN_VRAM) || + (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) || + (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) { + if (mem->domain == AMDGPU_GEM_DOMAIN_VRAM) + same_hive = amdgpu_xgmi_same_hive(adev, bo_adev); + if (!same_hive && !amdgpu_device_is_peer_accessible(bo_adev, adev)) + return -EINVAL; } - bo_va_entry->va = va; - bo_va_entry->pte_flags = amdgpu_gmc_get_pte_flags(adev, - mem->mapping_flags); - bo_va_entry->kgd_dev = (void *)adev; - list_add(&bo_va_entry->bo_list, list_bo_va); + for (i = 0; i <= is_aql; i++) { + attachment[i] = kzalloc(sizeof(*attachment[i]), GFP_KERNEL); + if (unlikely(!attachment[i])) { + ret = -ENOMEM; + goto unwind; + } - if (p_bo_va_entry) - *p_bo_va_entry = bo_va_entry; + pr_debug("\t add VA 0x%llx - 0x%llx to vm %p\n", va, + va + bo_size, vm); - /* Allocate new page tables if needed and validate - * them. Clearing of new page tables and validate need to wait - * on move fences. We don't want that to trigger the eviction - * fence, so remove it temporarily. - */ - amdgpu_amdkfd_remove_eviction_fence(pd, - vm->process_info->eviction_fence, - NULL, NULL); + if ((adev == bo_adev && !(mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) || + (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) && reuse_dmamap(adev, bo_adev)) || + (mem->domain == AMDGPU_GEM_DOMAIN_GTT && reuse_dmamap(adev, bo_adev)) || + same_hive) { + /* Mappings on the local GPU, or VRAM mappings in the + * local hive, or userptr, or GTT mapping can reuse dma map + * address space share the original BO + */ + attachment[i]->type = KFD_MEM_ATT_SHARED; + bo[i] = mem->bo; + drm_gem_object_get(&bo[i]->tbo.base); + } else if (i > 0) { + /* Multiple mappings on the same GPU share the BO */ + attachment[i]->type = KFD_MEM_ATT_SHARED; + bo[i] = bo[0]; + drm_gem_object_get(&bo[i]->tbo.base); + } else if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) { + /* Create an SG BO to DMA-map userptrs on other GPUs */ + attachment[i]->type = KFD_MEM_ATT_USERPTR; + ret = create_dmamap_sg_bo(adev, mem, &bo[i]); + if (ret) + goto unwind; + /* Handle DOORBELL BOs of peer devices and MMIO BOs of local and peer devices */ + } else if (mem->bo->tbo.type == ttm_bo_type_sg) { + WARN_ONCE(!(mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL || + mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP), + "Handing invalid SG BO in ATTACH request"); + attachment[i]->type = KFD_MEM_ATT_SG; + ret = create_dmamap_sg_bo(adev, mem, &bo[i]); + if (ret) + goto unwind; + /* Enable acces to GTT and VRAM BOs of peer devices */ + } else if (mem->domain == AMDGPU_GEM_DOMAIN_GTT || + mem->domain == AMDGPU_GEM_DOMAIN_VRAM) { + attachment[i]->type = KFD_MEM_ATT_DMABUF; + ret = kfd_mem_attach_dmabuf(adev, mem, &bo[i]); + if (ret) + goto unwind; + pr_debug("Employ DMABUF mechanism to enable peer GPU access\n"); + } else { + WARN_ONCE(true, "Handling invalid ATTACH request"); + ret = -EINVAL; + goto unwind; + } - ret = amdgpu_vm_alloc_pts(adev, vm, va, amdgpu_bo_size(bo)); - if (ret) { - pr_err("Failed to allocate pts, err=%d\n", ret); - goto err_alloc_pts; - } + /* Add BO to VM internal data structures */ + ret = amdgpu_bo_reserve(bo[i], false); + if (ret) { + pr_debug("Unable to reserve BO during memory attach"); + goto unwind; + } + bo_va = amdgpu_vm_bo_find(vm, bo[i]); + if (!bo_va) + bo_va = amdgpu_vm_bo_add(adev, vm, bo[i]); + else + ++bo_va->ref_count; + attachment[i]->bo_va = bo_va; + amdgpu_bo_unreserve(bo[i]); + if (unlikely(!attachment[i]->bo_va)) { + ret = -ENOMEM; + pr_err("Failed to add BO object to VM. ret == %d\n", + ret); + goto unwind; + } + attachment[i]->va = va; + attachment[i]->pte_flags = get_pte_flags(adev, vm, mem); + attachment[i]->adev = adev; + list_add(&attachment[i]->list, &mem->attachments); - ret = vm_validate_pt_pd_bos(vm); - if (ret) { - pr_err("validate_pt_pd_bos() failed\n"); - goto err_alloc_pts; + va += bo_size; } - /* Add the eviction fence back */ - amdgpu_bo_fence(pd, &vm->process_info->eviction_fence->base, true); - return 0; -err_alloc_pts: - amdgpu_bo_fence(pd, &vm->process_info->eviction_fence->base, true); - amdgpu_vm_bo_rmv(adev, bo_va_entry->bo_va); - list_del(&bo_va_entry->bo_list); -err_vmadd: - kfree(bo_va_entry); +unwind: + for (; i >= 0; i--) { + if (!attachment[i]) + continue; + if (attachment[i]->bo_va) { + (void)amdgpu_bo_reserve(bo[i], true); + if (--attachment[i]->bo_va->ref_count == 0) + amdgpu_vm_bo_del(adev, attachment[i]->bo_va); + amdgpu_bo_unreserve(bo[i]); + list_del(&attachment[i]->list); + } + if (bo[i]) + drm_gem_object_put(&bo[i]->tbo.base); + kfree(attachment[i]); + } return ret; } -static void remove_bo_from_vm(struct amdgpu_device *adev, - struct kfd_bo_va_list *entry, unsigned long size) +static void kfd_mem_detach(struct kfd_mem_attachment *attachment) { - pr_debug("\t remove VA 0x%llx - 0x%llx in entry %p\n", - entry->va, - entry->va + size, entry); - amdgpu_vm_bo_rmv(adev, entry->bo_va); - list_del(&entry->bo_list); - kfree(entry); + struct amdgpu_bo *bo = attachment->bo_va->base.bo; + + pr_debug("\t remove VA 0x%llx in entry %p\n", + attachment->va, attachment); + if (--attachment->bo_va->ref_count == 0) + amdgpu_vm_bo_del(attachment->adev, attachment->bo_va); + drm_gem_object_put(&bo->tbo.base); + list_del(&attachment->list); + kfree(attachment); } static void add_kgd_mem_to_kfd_bo_list(struct kgd_mem *mem, struct amdkfd_process_info *process_info, bool userptr) { - struct ttm_validate_buffer *entry = &mem->validate_list; - struct amdgpu_bo *bo = mem->bo; - - INIT_LIST_HEAD(&entry->head); - entry->num_shared = 1; - entry->bo = &bo->tbo; mutex_lock(&process_info->lock); if (userptr) - list_add_tail(&entry->head, &process_info->userptr_valid_list); + list_add_tail(&mem->validate_list, + &process_info->userptr_valid_list); else - list_add_tail(&entry->head, &process_info->kfd_bo_list); + list_add_tail(&mem->validate_list, &process_info->kfd_bo_list); + mutex_unlock(&process_info->lock); +} + +static void remove_kgd_mem_from_kfd_bo_list(struct kgd_mem *mem, + struct amdkfd_process_info *process_info) +{ + mutex_lock(&process_info->lock); + list_del(&mem->validate_list); mutex_unlock(&process_info->lock); } @@ -559,56 +1051,68 @@ static void add_kgd_mem_to_kfd_bo_list(struct kgd_mem *mem, * * Returns 0 for success, negative errno for errors. */ -static int init_user_pages(struct kgd_mem *mem, struct mm_struct *mm, - uint64_t user_addr) +static int init_user_pages(struct kgd_mem *mem, uint64_t user_addr, + bool criu_resume) { struct amdkfd_process_info *process_info = mem->process_info; struct amdgpu_bo *bo = mem->bo; struct ttm_operation_ctx ctx = { true, false }; + struct amdgpu_hmm_range *range; int ret = 0; mutex_lock(&process_info->lock); - ret = amdgpu_ttm_tt_set_userptr(bo->tbo.ttm, user_addr, 0); + ret = amdgpu_ttm_tt_set_userptr(&bo->tbo, user_addr, 0); if (ret) { pr_err("%s: Failed to set userptr: %d\n", __func__, ret); goto out; } - ret = amdgpu_mn_register(bo, user_addr); + ret = amdgpu_hmm_register(bo, user_addr); if (ret) { pr_err("%s: Failed to register MMU notifier: %d\n", __func__, ret); goto out; } - /* If no restore worker is running concurrently, user_pages - * should not be allocated - */ - WARN(mem->user_pages, "Leaking user_pages array"); + if (criu_resume) { + /* + * During a CRIU restore operation, the userptr buffer objects + * will be validated in the restore_userptr_work worker at a + * later stage when it is scheduled by another ioctl called by + * CRIU master process for the target pid for restore. + */ + mutex_lock(&process_info->notifier_lock); + mem->invalid++; + mutex_unlock(&process_info->notifier_lock); + mutex_unlock(&process_info->lock); + return 0; + } - mem->user_pages = kvmalloc_array(bo->tbo.ttm->num_pages, - sizeof(struct page *), - GFP_KERNEL | __GFP_ZERO); - if (!mem->user_pages) { - pr_err("%s: Failed to allocate pages array\n", __func__); + range = amdgpu_hmm_range_alloc(NULL); + if (unlikely(!range)) { ret = -ENOMEM; goto unregister_out; } - ret = amdgpu_ttm_tt_get_user_pages(bo->tbo.ttm, mem->user_pages); + ret = amdgpu_ttm_tt_get_user_pages(bo, range); if (ret) { - pr_err("%s: Failed to get user pages: %d\n", __func__, ret); - goto free_out; + amdgpu_hmm_range_free(range); + if (ret == -EAGAIN) + pr_debug("Failed to get user pages, try again\n"); + else + pr_err("%s: Failed to get user pages: %d\n", __func__, ret); + goto unregister_out; } - amdgpu_ttm_tt_set_user_pages(bo->tbo.ttm, mem->user_pages); - ret = amdgpu_bo_reserve(bo, true); if (ret) { pr_err("%s: Failed to reserve BO\n", __func__); goto release_out; } + + amdgpu_ttm_tt_set_user_pages(bo->tbo.ttm, range); + amdgpu_bo_placement_from_domain(bo, mem->domain); ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); if (ret) @@ -616,14 +1120,10 @@ static int init_user_pages(struct kgd_mem *mem, struct mm_struct *mm, amdgpu_bo_unreserve(bo); release_out: - if (ret) - release_pages(mem->user_pages, bo->tbo.ttm->num_pages); -free_out: - kvfree(mem->user_pages); - mem->user_pages = NULL; + amdgpu_hmm_range_free(range); unregister_out: if (ret) - amdgpu_mn_unregister(bo); + amdgpu_hmm_unregister(bo); out: mutex_unlock(&process_info->lock); return ret; @@ -635,13 +1135,12 @@ out: * object can track VM updates. */ struct bo_vm_reservation_context { - struct amdgpu_bo_list_entry kfd_bo; /* BO list entry for the KFD BO */ - unsigned int n_vms; /* Number of VMs reserved */ - struct amdgpu_bo_list_entry *vm_pd; /* Array of VM BO list entries */ - struct ww_acquire_ctx ticket; /* Reservation ticket */ - struct list_head list, duplicates; /* BO lists */ - struct amdgpu_sync *sync; /* Pointer to sync object */ - bool reserved; /* Whether BOs are reserved */ + /* DRM execution context for the reservation */ + struct drm_exec exec; + /* Number of VMs reserved */ + unsigned int n_vms; + /* Pointer to sync object */ + struct amdgpu_sync *sync; }; enum bo_vm_match { @@ -665,35 +1164,25 @@ static int reserve_bo_and_vm(struct kgd_mem *mem, WARN_ON(!vm); - ctx->reserved = false; ctx->n_vms = 1; ctx->sync = &mem->sync; + drm_exec_init(&ctx->exec, DRM_EXEC_INTERRUPTIBLE_WAIT, 0); + drm_exec_until_all_locked(&ctx->exec) { + ret = amdgpu_vm_lock_pd(vm, &ctx->exec, 2); + drm_exec_retry_on_contention(&ctx->exec); + if (unlikely(ret)) + goto error; - INIT_LIST_HEAD(&ctx->list); - INIT_LIST_HEAD(&ctx->duplicates); - - ctx->vm_pd = kcalloc(ctx->n_vms, sizeof(*ctx->vm_pd), GFP_KERNEL); - if (!ctx->vm_pd) - return -ENOMEM; - - ctx->kfd_bo.priority = 0; - ctx->kfd_bo.tv.bo = &bo->tbo; - ctx->kfd_bo.tv.num_shared = 1; - ctx->kfd_bo.user_pages = NULL; - list_add(&ctx->kfd_bo.tv.head, &ctx->list); - - amdgpu_vm_get_pd_bo(vm, &ctx->list, &ctx->vm_pd[0]); - - ret = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->list, - false, &ctx->duplicates); - if (!ret) - ctx->reserved = true; - else { - pr_err("Failed to reserve buffers in ttm\n"); - kfree(ctx->vm_pd); - ctx->vm_pd = NULL; + ret = drm_exec_prepare_obj(&ctx->exec, &bo->tbo.base, 1); + drm_exec_retry_on_contention(&ctx->exec); + if (unlikely(ret)) + goto error; } + return 0; +error: + pr_err("Failed to reserve buffers in ttm.\n"); + drm_exec_fini(&ctx->exec); return ret; } @@ -711,65 +1200,39 @@ static int reserve_bo_and_cond_vms(struct kgd_mem *mem, struct amdgpu_vm *vm, enum bo_vm_match map_type, struct bo_vm_reservation_context *ctx) { + struct kfd_mem_attachment *entry; struct amdgpu_bo *bo = mem->bo; - struct kfd_bo_va_list *entry; - unsigned int i; int ret; - ctx->reserved = false; - ctx->n_vms = 0; - ctx->vm_pd = NULL; ctx->sync = &mem->sync; + drm_exec_init(&ctx->exec, DRM_EXEC_INTERRUPTIBLE_WAIT | + DRM_EXEC_IGNORE_DUPLICATES, 0); + drm_exec_until_all_locked(&ctx->exec) { + ctx->n_vms = 0; + list_for_each_entry(entry, &mem->attachments, list) { + if ((vm && vm != entry->bo_va->base.vm) || + (entry->is_mapped != map_type + && map_type != BO_VM_ALL)) + continue; - INIT_LIST_HEAD(&ctx->list); - INIT_LIST_HEAD(&ctx->duplicates); - - list_for_each_entry(entry, &mem->bo_va_list, bo_list) { - if ((vm && vm != entry->bo_va->base.vm) || - (entry->is_mapped != map_type - && map_type != BO_VM_ALL)) - continue; - - ctx->n_vms++; - } - - if (ctx->n_vms != 0) { - ctx->vm_pd = kcalloc(ctx->n_vms, sizeof(*ctx->vm_pd), - GFP_KERNEL); - if (!ctx->vm_pd) - return -ENOMEM; - } - - ctx->kfd_bo.priority = 0; - ctx->kfd_bo.tv.bo = &bo->tbo; - ctx->kfd_bo.tv.num_shared = 1; - ctx->kfd_bo.user_pages = NULL; - list_add(&ctx->kfd_bo.tv.head, &ctx->list); - - i = 0; - list_for_each_entry(entry, &mem->bo_va_list, bo_list) { - if ((vm && vm != entry->bo_va->base.vm) || - (entry->is_mapped != map_type - && map_type != BO_VM_ALL)) - continue; - - amdgpu_vm_get_pd_bo(entry->bo_va->base.vm, &ctx->list, - &ctx->vm_pd[i]); - i++; - } - - ret = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->list, - false, &ctx->duplicates); - if (!ret) - ctx->reserved = true; - else - pr_err("Failed to reserve buffers in ttm.\n"); + ret = amdgpu_vm_lock_pd(entry->bo_va->base.vm, + &ctx->exec, 2); + drm_exec_retry_on_contention(&ctx->exec); + if (unlikely(ret)) + goto error; + ++ctx->n_vms; + } - if (ret) { - kfree(ctx->vm_pd); - ctx->vm_pd = NULL; + ret = drm_exec_prepare_obj(&ctx->exec, &bo->tbo.base, 1); + drm_exec_retry_on_contention(&ctx->exec); + if (unlikely(ret)) + goto error; } + return 0; +error: + pr_err("Failed to reserve buffers in ttm.\n"); + drm_exec_fini(&ctx->exec); return ret; } @@ -791,59 +1254,48 @@ static int unreserve_bo_and_vms(struct bo_vm_reservation_context *ctx, if (wait) ret = amdgpu_sync_wait(ctx->sync, intr); - if (ctx->reserved) - ttm_eu_backoff_reservation(&ctx->ticket, &ctx->list); - kfree(ctx->vm_pd); - + drm_exec_fini(&ctx->exec); ctx->sync = NULL; - - ctx->reserved = false; - ctx->vm_pd = NULL; - return ret; } -static int unmap_bo_from_gpuvm(struct amdgpu_device *adev, - struct kfd_bo_va_list *entry, +static int unmap_bo_from_gpuvm(struct kgd_mem *mem, + struct kfd_mem_attachment *entry, struct amdgpu_sync *sync) { struct amdgpu_bo_va *bo_va = entry->bo_va; + struct amdgpu_device *adev = entry->adev; struct amdgpu_vm *vm = bo_va->base.vm; - struct amdgpu_bo *pd = vm->root.base.bo; - /* Remove eviction fence from PD (and thereby from PTs too as - * they share the resv. object). Otherwise during PT update - * job (see amdgpu_vm_bo_update_mapping), eviction fence would - * get added to job->sync object and job execution would - * trigger the eviction fence. - */ - amdgpu_amdkfd_remove_eviction_fence(pd, - vm->process_info->eviction_fence, - NULL, NULL); - amdgpu_vm_bo_unmap(adev, bo_va, entry->va); + if (bo_va->queue_refcount) { + pr_debug("bo_va->queue_refcount %d\n", bo_va->queue_refcount); + return -EBUSY; + } - amdgpu_vm_clear_freed(adev, vm, &bo_va->last_pt_update); + (void)amdgpu_vm_bo_unmap(adev, bo_va, entry->va); - /* Add the eviction fence back */ - amdgpu_bo_fence(pd, &vm->process_info->eviction_fence->base, true); + /* VM entity stopped if process killed, don't clear freed pt bo */ + if (!amdgpu_vm_ready(vm)) + return 0; - amdgpu_sync_fence(NULL, sync, bo_va->last_pt_update, false); + (void)amdgpu_vm_clear_freed(adev, vm, &bo_va->last_pt_update); + + (void)amdgpu_sync_fence(sync, bo_va->last_pt_update, GFP_KERNEL); return 0; } -static int update_gpuvm_pte(struct amdgpu_device *adev, - struct kfd_bo_va_list *entry, - struct amdgpu_sync *sync) +static int update_gpuvm_pte(struct kgd_mem *mem, + struct kfd_mem_attachment *entry, + struct amdgpu_sync *sync) { + struct amdgpu_bo_va *bo_va = entry->bo_va; + struct amdgpu_device *adev = entry->adev; int ret; - struct amdgpu_vm *vm; - struct amdgpu_bo_va *bo_va; - struct amdgpu_bo *bo; - bo_va = entry->bo_va; - vm = bo_va->base.vm; - bo = bo_va->base.bo; + ret = kfd_mem_dmamap_attachment(mem, entry); + if (ret) + return ret; /* Update the page tables */ ret = amdgpu_vm_bo_update(adev, bo_va, false); @@ -852,17 +1304,18 @@ static int update_gpuvm_pte(struct amdgpu_device *adev, return ret; } - return amdgpu_sync_fence(NULL, sync, bo_va->last_pt_update, false); + return amdgpu_sync_fence(sync, bo_va->last_pt_update, GFP_KERNEL); } -static int map_bo_to_gpuvm(struct amdgpu_device *adev, - struct kfd_bo_va_list *entry, struct amdgpu_sync *sync, - bool no_update_pte) +static int map_bo_to_gpuvm(struct kgd_mem *mem, + struct kfd_mem_attachment *entry, + struct amdgpu_sync *sync, + bool no_update_pte) { int ret; /* Set virtual address for the allocation */ - ret = amdgpu_vm_bo_map(adev, entry->bo_va, entry->va, 0, + ret = amdgpu_vm_bo_map(entry->adev, entry->bo_va, entry->va, 0, amdgpu_bo_size(entry->bo_va->base.bo), entry->pte_flags); if (ret) { @@ -874,7 +1327,7 @@ static int map_bo_to_gpuvm(struct amdgpu_device *adev, if (no_update_pte) return 0; - ret = update_gpuvm_pte(adev, entry, sync); + ret = update_gpuvm_pte(mem, entry, sync); if (ret) { pr_err("update_gpuvm_pte() failed\n"); goto update_gpuvm_pte_failed; @@ -883,36 +1336,20 @@ static int map_bo_to_gpuvm(struct amdgpu_device *adev, return 0; update_gpuvm_pte_failed: - unmap_bo_from_gpuvm(adev, entry, sync); + unmap_bo_from_gpuvm(mem, entry, sync); + kfd_mem_dmaunmap_attachment(mem, entry); return ret; } -static struct sg_table *create_doorbell_sg(uint64_t addr, uint32_t size) -{ - struct sg_table *sg = kmalloc(sizeof(*sg), GFP_KERNEL); - - if (!sg) - return NULL; - if (sg_alloc_table(sg, 1, GFP_KERNEL)) { - kfree(sg); - return NULL; - } - sg->sgl->dma_address = addr; - sg->sgl->length = size; -#ifdef CONFIG_NEED_SG_DMA_LENGTH - sg->sgl->dma_length = size; -#endif - return sg; -} - -static int process_validate_vms(struct amdkfd_process_info *process_info) +static int process_validate_vms(struct amdkfd_process_info *process_info, + struct ww_acquire_ctx *ticket) { struct amdgpu_vm *peer_vm; int ret; list_for_each_entry(peer_vm, &process_info->vm_list_head, vm_list_node) { - ret = vm_validate_pt_pd_bos(peer_vm); + ret = vm_validate_pt_pd_bos(peer_vm, ticket); if (ret) return ret; } @@ -928,11 +1365,11 @@ static int process_sync_pds_resv(struct amdkfd_process_info *process_info, list_for_each_entry(peer_vm, &process_info->vm_list_head, vm_list_node) { - struct amdgpu_bo *pd = peer_vm->root.base.bo; + struct amdgpu_bo *pd = peer_vm->root.bo; - ret = amdgpu_sync_resv(NULL, - sync, pd->tbo.resv, - AMDGPU_FENCE_OWNER_UNDEFINED, false); + ret = amdgpu_sync_resv(NULL, sync, pd->tbo.base.resv, + AMDGPU_SYNC_NE_OWNER, + AMDGPU_FENCE_OWNER_KFD); if (ret) return ret; } @@ -968,6 +1405,7 @@ static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info, return -ENOMEM; mutex_init(&info->lock); + mutex_init(&info->notifier_lock); INIT_LIST_HEAD(&info->vm_list_head); INIT_LIST_HEAD(&info->kfd_bo_list); INIT_LIST_HEAD(&info->userptr_valid_list); @@ -975,7 +1413,8 @@ static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info, info->eviction_fence = amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1), - current->mm); + current->mm, + NULL); if (!info->eviction_fence) { pr_err("Failed to create eviction fence\n"); ret = -ENOMEM; @@ -983,104 +1422,135 @@ static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info, } info->pid = get_task_pid(current->group_leader, PIDTYPE_PID); - atomic_set(&info->evicted_bos, 0); INIT_DELAYED_WORK(&info->restore_userptr_work, amdgpu_amdkfd_restore_userptr_worker); *process_info = info; - *ef = dma_fence_get(&info->eviction_fence->base); } vm->process_info = *process_info; /* Validate page directory and attach eviction fence */ - ret = amdgpu_bo_reserve(vm->root.base.bo, true); + ret = amdgpu_bo_reserve(vm->root.bo, true); if (ret) goto reserve_pd_fail; - ret = vm_validate_pt_pd_bos(vm); + ret = vm_validate_pt_pd_bos(vm, NULL); if (ret) { pr_err("validate_pt_pd_bos() failed\n"); goto validate_pd_fail; } - ret = ttm_bo_wait(&vm->root.base.bo->tbo, false, false); + ret = amdgpu_bo_sync_wait(vm->root.bo, + AMDGPU_FENCE_OWNER_KFD, false); if (ret) goto wait_pd_fail; - amdgpu_bo_fence(vm->root.base.bo, - &vm->process_info->eviction_fence->base, true); - amdgpu_bo_unreserve(vm->root.base.bo); + ret = dma_resv_reserve_fences(vm->root.bo->tbo.base.resv, 1); + if (ret) + goto reserve_shared_fail; + dma_resv_add_fence(vm->root.bo->tbo.base.resv, + &vm->process_info->eviction_fence->base, + DMA_RESV_USAGE_BOOKKEEP); + amdgpu_bo_unreserve(vm->root.bo); /* Update process info */ mutex_lock(&vm->process_info->lock); list_add_tail(&vm->vm_list_node, &(vm->process_info->vm_list_head)); vm->process_info->n_vms++; + if (ef) + *ef = dma_fence_get(&vm->process_info->eviction_fence->base); mutex_unlock(&vm->process_info->lock); return 0; +reserve_shared_fail: wait_pd_fail: validate_pd_fail: - amdgpu_bo_unreserve(vm->root.base.bo); + amdgpu_bo_unreserve(vm->root.bo); reserve_pd_fail: vm->process_info = NULL; if (info) { - /* Two fence references: one in info and one in *ef */ dma_fence_put(&info->eviction_fence->base); - dma_fence_put(*ef); - *ef = NULL; *process_info = NULL; put_pid(info->pid); create_evict_fence_fail: mutex_destroy(&info->lock); + mutex_destroy(&info->notifier_lock); kfree(info); } return ret; } -int amdgpu_amdkfd_gpuvm_create_process_vm(struct kgd_dev *kgd, unsigned int pasid, - void **vm, void **process_info, - struct dma_fence **ef) +/** + * amdgpu_amdkfd_gpuvm_pin_bo() - Pins a BO using following criteria + * @bo: Handle of buffer object being pinned + * @domain: Domain into which BO should be pinned + * + * - USERPTR BOs are UNPINNABLE and will return error + * - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their + * PIN count incremented. It is valid to PIN a BO multiple times + * + * Return: ZERO if successful in pinning, Non-Zero in case of error. + */ +static int amdgpu_amdkfd_gpuvm_pin_bo(struct amdgpu_bo *bo, u32 domain) { - struct amdgpu_device *adev = get_amdgpu_device(kgd); - struct amdgpu_vm *new_vm; - int ret; + int ret = 0; - new_vm = kzalloc(sizeof(*new_vm), GFP_KERNEL); - if (!new_vm) - return -ENOMEM; + ret = amdgpu_bo_reserve(bo, false); + if (unlikely(ret)) + return ret; - /* Initialize AMDGPU part of the VM */ - ret = amdgpu_vm_init(adev, new_vm, AMDGPU_VM_CONTEXT_COMPUTE, pasid); - if (ret) { - pr_err("Failed init vm ret %d\n", ret); - goto amdgpu_vm_init_fail; + if (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS) { + /* + * If bo is not contiguous on VRAM, move to system memory first to ensure + * we can get contiguous VRAM space after evicting other BOs. + */ + if (!(bo->tbo.resource->placement & TTM_PL_FLAG_CONTIGUOUS)) { + struct ttm_operation_ctx ctx = { true, false }; + + amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT); + ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); + if (unlikely(ret)) { + pr_debug("validate bo 0x%p to GTT failed %d\n", &bo->tbo, ret); + goto out; + } + } } - /* Initialize KFD part of the VM and process info */ - ret = init_kfd_vm(new_vm, process_info, ef); + ret = amdgpu_bo_pin(bo, domain); if (ret) - goto init_kfd_vm_fail; + pr_err("Error in Pinning BO to domain: %d\n", domain); + + amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false); +out: + amdgpu_bo_unreserve(bo); + return ret; +} - *vm = (void *) new_vm; +/** + * amdgpu_amdkfd_gpuvm_unpin_bo() - Unpins BO using following criteria + * @bo: Handle of buffer object being unpinned + * + * - Is a illegal request for USERPTR BOs and is ignored + * - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their + * PIN count decremented. Calls to UNPIN must balance calls to PIN + */ +static void amdgpu_amdkfd_gpuvm_unpin_bo(struct amdgpu_bo *bo) +{ + int ret = 0; - return 0; + ret = amdgpu_bo_reserve(bo, false); + if (unlikely(ret)) + return; -init_kfd_vm_fail: - amdgpu_vm_fini(adev, new_vm); -amdgpu_vm_init_fail: - kfree(new_vm); - return ret; + amdgpu_bo_unpin(bo); + amdgpu_bo_unreserve(bo); } -int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct kgd_dev *kgd, - struct file *filp, unsigned int pasid, - void **vm, void **process_info, +int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct amdgpu_device *adev, + struct amdgpu_vm *avm, + void **process_info, struct dma_fence **ef) { - struct amdgpu_device *adev = get_amdgpu_device(kgd); - struct drm_file *drm_priv = filp->private_data; - struct amdgpu_fpriv *drv_priv = drm_priv->driver_priv; - struct amdgpu_vm *avm = &drv_priv->vm; int ret; /* Already a compute VM? */ @@ -1088,7 +1558,7 @@ int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct kgd_dev *kgd, return -EINVAL; /* Convert VM into a compute VM */ - ret = amdgpu_vm_make_compute(adev, avm, pasid); + ret = amdgpu_vm_make_compute(adev, avm); if (ret) return ret; @@ -1097,7 +1567,7 @@ int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct kgd_dev *kgd, if (ret) return ret; - *vm = (void *)avm; + amdgpu_vm_set_task_info(avm); return 0; } @@ -1106,22 +1576,18 @@ void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev, struct amdgpu_vm *vm) { struct amdkfd_process_info *process_info = vm->process_info; - struct amdgpu_bo *pd = vm->root.base.bo; if (!process_info) return; - /* Release eviction fence from PD */ - amdgpu_bo_reserve(pd, false); - amdgpu_bo_fence(pd, NULL, false); - amdgpu_bo_unreserve(pd); - /* Update process info */ mutex_lock(&process_info->lock); process_info->n_vms--; list_del(&vm->vm_list_node); mutex_unlock(&process_info->lock); + vm->process_info = NULL; + /* Release per-process resources when last compute VM is destroyed */ if (!process_info->n_vms) { WARN_ON(!list_empty(&process_info->kfd_bo_list)); @@ -1132,165 +1598,218 @@ void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev, cancel_delayed_work_sync(&process_info->restore_userptr_work); put_pid(process_info->pid); mutex_destroy(&process_info->lock); + mutex_destroy(&process_info->notifier_lock); kfree(process_info); } } -void amdgpu_amdkfd_gpuvm_destroy_process_vm(struct kgd_dev *kgd, void *vm) +uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *drm_priv) { - struct amdgpu_device *adev = get_amdgpu_device(kgd); - struct amdgpu_vm *avm = (struct amdgpu_vm *)vm; - - if (WARN_ON(!kgd || !vm)) - return; - - pr_debug("Destroying process vm %p\n", vm); + struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv); + struct amdgpu_bo *pd = avm->root.bo; + struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev); - /* Release the VM context */ - amdgpu_vm_fini(adev, avm); - kfree(vm); + if (adev->asic_type < CHIP_VEGA10) + return avm->pd_phys_addr >> AMDGPU_GPU_PAGE_SHIFT; + return avm->pd_phys_addr; } -void amdgpu_amdkfd_gpuvm_release_process_vm(struct kgd_dev *kgd, void *vm) +void amdgpu_amdkfd_block_mmu_notifications(void *p) { - struct amdgpu_device *adev = get_amdgpu_device(kgd); - struct amdgpu_vm *avm = (struct amdgpu_vm *)vm; + struct amdkfd_process_info *pinfo = (struct amdkfd_process_info *)p; - if (WARN_ON(!kgd || !vm)) - return; + mutex_lock(&pinfo->lock); + WRITE_ONCE(pinfo->block_mmu_notifications, true); + mutex_unlock(&pinfo->lock); +} - pr_debug("Releasing process vm %p\n", vm); +int amdgpu_amdkfd_criu_resume(void *p) +{ + int ret = 0; + struct amdkfd_process_info *pinfo = (struct amdkfd_process_info *)p; + + mutex_lock(&pinfo->lock); + pr_debug("scheduling work\n"); + mutex_lock(&pinfo->notifier_lock); + pinfo->evicted_bos++; + mutex_unlock(&pinfo->notifier_lock); + if (!READ_ONCE(pinfo->block_mmu_notifications)) { + ret = -EINVAL; + goto out_unlock; + } + WRITE_ONCE(pinfo->block_mmu_notifications, false); + queue_delayed_work(system_freezable_wq, + &pinfo->restore_userptr_work, 0); - /* The original pasid of amdgpu vm has already been - * released during making a amdgpu vm to a compute vm - * The current pasid is managed by kfd and will be - * released on kfd process destroy. Set amdgpu pasid - * to 0 to avoid duplicate release. - */ - amdgpu_vm_release_compute(adev, avm); +out_unlock: + mutex_unlock(&pinfo->lock); + return ret; } -uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *vm) +size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev, + uint8_t xcp_id) { - struct amdgpu_vm *avm = (struct amdgpu_vm *)vm; - struct amdgpu_bo *pd = avm->root.base.bo; - struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev); + uint64_t reserved_for_pt = + ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size); + struct amdgpu_ras *con = amdgpu_ras_get_context(adev); + uint64_t reserved_for_ras = (con ? con->reserved_pages_in_bytes : 0); + ssize_t available; + uint64_t vram_available, system_mem_available, ttm_mem_available; - if (adev->asic_type < CHIP_VEGA10) - return avm->pd_phys_addr >> AMDGPU_GPU_PAGE_SHIFT; - return avm->pd_phys_addr; + spin_lock(&kfd_mem_limit.mem_limit_lock); + if (adev->apu_prefer_gtt && !adev->gmc.is_app_apu) + vram_available = KFD_XCP_MEMORY_SIZE(adev, xcp_id) + - adev->kfd.vram_used_aligned[xcp_id]; + else + vram_available = KFD_XCP_MEMORY_SIZE(adev, xcp_id) + - adev->kfd.vram_used_aligned[xcp_id] + - atomic64_read(&adev->vram_pin_size) + - reserved_for_pt + - reserved_for_ras; + + if (adev->apu_prefer_gtt) { + system_mem_available = no_system_mem_limit ? + kfd_mem_limit.max_system_mem_limit : + kfd_mem_limit.max_system_mem_limit - + kfd_mem_limit.system_mem_used; + + ttm_mem_available = kfd_mem_limit.max_ttm_mem_limit - + kfd_mem_limit.ttm_mem_used; + + available = min3(system_mem_available, ttm_mem_available, + vram_available); + available = ALIGN_DOWN(available, PAGE_SIZE); + } else { + available = ALIGN_DOWN(vram_available, VRAM_AVAILABLITY_ALIGN); + } + + spin_unlock(&kfd_mem_limit.mem_limit_lock); + + if (available < 0) + available = 0; + + return available; } int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( - struct kgd_dev *kgd, uint64_t va, uint64_t size, - void *vm, struct kgd_mem **mem, - uint64_t *offset, uint32_t flags) + struct amdgpu_device *adev, uint64_t va, uint64_t size, + void *drm_priv, struct kgd_mem **mem, + uint64_t *offset, uint32_t flags, bool criu_resume) { - struct amdgpu_device *adev = get_amdgpu_device(kgd); - struct amdgpu_vm *avm = (struct amdgpu_vm *)vm; + struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv); + struct amdgpu_fpriv *fpriv = container_of(avm, struct amdgpu_fpriv, vm); enum ttm_bo_type bo_type = ttm_bo_type_device; struct sg_table *sg = NULL; uint64_t user_addr = 0; struct amdgpu_bo *bo; - struct amdgpu_bo_param bp; - int byte_align; + struct drm_gem_object *gobj = NULL; u32 domain, alloc_domain; + uint64_t aligned_size; + int8_t xcp_id = -1; u64 alloc_flags; - uint32_t mapping_flags; int ret; /* * Check on which domain to allocate BO */ - if (flags & ALLOC_MEM_FLAGS_VRAM) { + if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) { domain = alloc_domain = AMDGPU_GEM_DOMAIN_VRAM; - alloc_flags = AMDGPU_GEM_CREATE_VRAM_CLEARED; - alloc_flags |= (flags & ALLOC_MEM_FLAGS_PUBLIC) ? - AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED : - AMDGPU_GEM_CREATE_NO_CPU_ACCESS; - } else if (flags & ALLOC_MEM_FLAGS_GTT) { + + if (adev->apu_prefer_gtt) { + domain = AMDGPU_GEM_DOMAIN_GTT; + alloc_domain = AMDGPU_GEM_DOMAIN_GTT; + alloc_flags = 0; + } else { + alloc_flags = AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE; + alloc_flags |= (flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) ? + AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED : 0; + + /* For contiguous VRAM allocation */ + if (flags & KFD_IOC_ALLOC_MEM_FLAGS_CONTIGUOUS) + alloc_flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS; + } + xcp_id = fpriv->xcp_id == AMDGPU_XCP_NO_PARTITION ? + 0 : fpriv->xcp_id; + } else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) { domain = alloc_domain = AMDGPU_GEM_DOMAIN_GTT; alloc_flags = 0; - } else if (flags & ALLOC_MEM_FLAGS_USERPTR) { - domain = AMDGPU_GEM_DOMAIN_GTT; - alloc_domain = AMDGPU_GEM_DOMAIN_CPU; - alloc_flags = 0; - if (!offset || !*offset) - return -EINVAL; - user_addr = *offset; - } else if (flags & ALLOC_MEM_FLAGS_DOORBELL) { + } else { domain = AMDGPU_GEM_DOMAIN_GTT; alloc_domain = AMDGPU_GEM_DOMAIN_CPU; - bo_type = ttm_bo_type_sg; - alloc_flags = 0; - if (size > UINT_MAX) + alloc_flags = AMDGPU_GEM_CREATE_PREEMPTIBLE; + + if (flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) { + if (!offset || !*offset) + return -EINVAL; + user_addr = untagged_addr(*offset); + } else if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL | + KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) { + bo_type = ttm_bo_type_sg; + if (size > UINT_MAX) + return -EINVAL; + sg = create_sg_table(*offset, size); + if (!sg) + return -ENOMEM; + } else { return -EINVAL; - sg = create_doorbell_sg(*offset, size); - if (!sg) - return -ENOMEM; - } else { - return -EINVAL; + } } + if (flags & KFD_IOC_ALLOC_MEM_FLAGS_COHERENT) + alloc_flags |= AMDGPU_GEM_CREATE_COHERENT; + if (flags & KFD_IOC_ALLOC_MEM_FLAGS_EXT_COHERENT) + alloc_flags |= AMDGPU_GEM_CREATE_EXT_COHERENT; + if (flags & KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED) + alloc_flags |= AMDGPU_GEM_CREATE_UNCACHED; + *mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL); if (!*mem) { ret = -ENOMEM; goto err; } - INIT_LIST_HEAD(&(*mem)->bo_va_list); + INIT_LIST_HEAD(&(*mem)->attachments); mutex_init(&(*mem)->lock); - (*mem)->aql_queue = !!(flags & ALLOC_MEM_FLAGS_AQL_QUEUE_MEM); + (*mem)->aql_queue = !!(flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM); /* Workaround for AQL queue wraparound bug. Map the same * memory twice. That means we only actually allocate half * the memory. */ if ((*mem)->aql_queue) - size = size >> 1; - - /* Workaround for TLB bug on older VI chips */ - byte_align = (adev->family == AMDGPU_FAMILY_VI && - adev->asic_type != CHIP_FIJI && - adev->asic_type != CHIP_POLARIS10 && - adev->asic_type != CHIP_POLARIS11 && - adev->asic_type != CHIP_POLARIS12) ? - VI_BO_SIZE_ALIGN : 1; - - mapping_flags = AMDGPU_VM_PAGE_READABLE; - if (flags & ALLOC_MEM_FLAGS_WRITABLE) - mapping_flags |= AMDGPU_VM_PAGE_WRITEABLE; - if (flags & ALLOC_MEM_FLAGS_EXECUTABLE) - mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE; - if (flags & ALLOC_MEM_FLAGS_COHERENT) - mapping_flags |= AMDGPU_VM_MTYPE_UC; - else - mapping_flags |= AMDGPU_VM_MTYPE_NC; - (*mem)->mapping_flags = mapping_flags; + size >>= 1; + aligned_size = PAGE_ALIGN(size); + + (*mem)->alloc_flags = flags; amdgpu_sync_create(&(*mem)->sync); - ret = amdgpu_amdkfd_reserve_mem_limit(adev, size, alloc_domain, !!sg); + ret = amdgpu_amdkfd_reserve_mem_limit(adev, aligned_size, flags, + xcp_id); if (ret) { - pr_debug("Insufficient system memory\n"); + pr_debug("Insufficient memory\n"); goto err_reserve_limit; } - pr_debug("\tcreate BO VA 0x%llx size 0x%llx domain %s\n", - va, size, domain_string(alloc_domain)); + pr_debug("\tcreate BO VA 0x%llx size 0x%llx domain %s xcp_id %d\n", + va, (*mem)->aql_queue ? size << 1 : size, + domain_string(alloc_domain), xcp_id); - memset(&bp, 0, sizeof(bp)); - bp.size = size; - bp.byte_align = byte_align; - bp.domain = alloc_domain; - bp.flags = alloc_flags; - bp.type = bo_type; - bp.resv = NULL; - ret = amdgpu_bo_create(adev, &bp, &bo); + ret = amdgpu_gem_object_create(adev, aligned_size, 1, alloc_domain, alloc_flags, + bo_type, NULL, &gobj, xcp_id + 1); if (ret) { pr_debug("Failed to create BO on domain %s. ret %d\n", - domain_string(alloc_domain), ret); + domain_string(alloc_domain), ret); goto err_bo_create; } + ret = drm_vma_node_allow(&gobj->vma_node, drm_priv); + if (ret) { + pr_debug("Failed to allow vma node access. ret %d\n", ret); + goto err_node_allow; + } + ret = drm_gem_handle_create(adev->kfd.client.file, gobj, &(*mem)->gem_handle); + if (ret) + goto err_gem_handle_create; + bo = gem_to_amdgpu_bo(gobj); if (bo_type == ttm_bo_type_sg) { bo->tbo.sg = sg; bo->tbo.ttm->sg = sg; @@ -1298,22 +1817,38 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( bo->kfd_bo = *mem; (*mem)->bo = bo; if (user_addr) - bo->flags |= AMDGPU_AMDKFD_USERPTR_BO; + bo->flags |= AMDGPU_AMDKFD_CREATE_USERPTR_BO; (*mem)->va = va; (*mem)->domain = domain; (*mem)->mapped_to_gpu_memory = 0; (*mem)->process_info = avm->process_info; + add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, user_addr); if (user_addr) { - ret = init_user_pages(*mem, current->mm, user_addr); - if (ret) { - mutex_lock(&avm->process_info->lock); - list_del(&(*mem)->validate_list.head); - mutex_unlock(&avm->process_info->lock); + pr_debug("creating userptr BO for user_addr = %llx\n", user_addr); + ret = init_user_pages(*mem, user_addr, criu_resume); + if (ret) goto allocate_init_user_pages_failed; + } else if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL | + KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) { + ret = amdgpu_amdkfd_gpuvm_pin_bo(bo, AMDGPU_GEM_DOMAIN_GTT); + if (ret) { + pr_err("Pinning MMIO/DOORBELL BO during ALLOC FAILED\n"); + goto err_pin_bo; } + bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT; + bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT; + } else { + mutex_lock(&avm->process_info->lock); + if (avm->process_info->eviction_fence && + !dma_fence_is_signaled(&avm->process_info->eviction_fence->base)) + ret = amdgpu_amdkfd_bo_validate_and_fence(bo, domain, + &avm->process_info->eviction_fence->base); + mutex_unlock(&avm->process_info->lock); + if (ret) + goto err_validate_bo; } if (offset) @@ -1322,14 +1857,24 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( return 0; allocate_init_user_pages_failed: - amdgpu_bo_unref(&bo); +err_pin_bo: +err_validate_bo: + remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info); + drm_gem_handle_delete(adev->kfd.client.file, (*mem)->gem_handle); +err_gem_handle_create: + drm_vma_node_revoke(&gobj->vma_node, drm_priv); +err_node_allow: /* Don't unreserve system mem limit twice */ goto err_reserve_limit; err_bo_create: - unreserve_mem_limit(adev, size, alloc_domain, !!sg); + amdgpu_amdkfd_unreserve_mem_limit(adev, aligned_size, flags, xcp_id); err_reserve_limit: + amdgpu_sync_free(&(*mem)->sync); mutex_destroy(&(*mem)->lock); - kfree(*mem); + if (gobj) + drm_gem_object_put(gobj); + else + kfree(*mem); err: if (sg) { sg_free_table(sg); @@ -1339,99 +1884,128 @@ err: } int amdgpu_amdkfd_gpuvm_free_memory_of_gpu( - struct kgd_dev *kgd, struct kgd_mem *mem) + struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv, + uint64_t *size) { struct amdkfd_process_info *process_info = mem->process_info; - unsigned long bo_size = mem->bo->tbo.mem.size; - struct kfd_bo_va_list *entry, *tmp; + unsigned long bo_size = mem->bo->tbo.base.size; + bool use_release_notifier = (mem->bo->kfd_bo == mem); + struct kfd_mem_attachment *entry, *tmp; struct bo_vm_reservation_context ctx; - struct ttm_validate_buffer *bo_list_entry; + unsigned int mapped_to_gpu_memory; int ret; + bool is_imported = false; mutex_lock(&mem->lock); - if (mem->mapped_to_gpu_memory > 0) { - pr_debug("BO VA 0x%llx size 0x%lx is still mapped.\n", - mem->va, bo_size); - mutex_unlock(&mem->lock); - return -EBUSY; + /* Unpin MMIO/DOORBELL BO's that were pinned during allocation */ + if (mem->alloc_flags & + (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL | + KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) { + amdgpu_amdkfd_gpuvm_unpin_bo(mem->bo); } + mapped_to_gpu_memory = mem->mapped_to_gpu_memory; + is_imported = mem->is_imported; mutex_unlock(&mem->lock); /* lock is not needed after this, since mem is unused and will * be freed anyway */ - /* No more MMU notifiers */ - amdgpu_mn_unregister(mem->bo); + if (mapped_to_gpu_memory > 0) { + pr_debug("BO VA 0x%llx size 0x%lx is still mapped.\n", + mem->va, bo_size); + return -EBUSY; + } /* Make sure restore workers don't access the BO any more */ - bo_list_entry = &mem->validate_list; mutex_lock(&process_info->lock); - list_del(&bo_list_entry->head); + list_del(&mem->validate_list); mutex_unlock(&process_info->lock); - /* Free user pages if necessary */ - if (mem->user_pages) { - pr_debug("%s: Freeing user_pages array\n", __func__); - if (mem->user_pages[0]) - release_pages(mem->user_pages, - mem->bo->tbo.ttm->num_pages); - kvfree(mem->user_pages); + /* Cleanup user pages and MMU notifiers */ + if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) { + amdgpu_hmm_unregister(mem->bo); + mutex_lock(&process_info->notifier_lock); + amdgpu_hmm_range_free(mem->range); + mutex_unlock(&process_info->notifier_lock); } ret = reserve_bo_and_cond_vms(mem, NULL, BO_VM_ALL, &ctx); if (unlikely(ret)) return ret; - /* The eviction fence should be removed by the last unmap. - * TODO: Log an error condition if the bo still has the eviction fence - * attached - */ amdgpu_amdkfd_remove_eviction_fence(mem->bo, - process_info->eviction_fence, - NULL, NULL); + process_info->eviction_fence); pr_debug("Release VA 0x%llx - 0x%llx\n", mem->va, mem->va + bo_size * (1 + mem->aql_queue)); /* Remove from VM internal data structures */ - list_for_each_entry_safe(entry, tmp, &mem->bo_va_list, bo_list) - remove_bo_from_vm((struct amdgpu_device *)entry->kgd_dev, - entry, bo_size); + list_for_each_entry_safe(entry, tmp, &mem->attachments, list) { + kfd_mem_dmaunmap_attachment(mem, entry); + kfd_mem_detach(entry); + } ret = unreserve_bo_and_vms(&ctx, false, false); /* Free the sync object */ amdgpu_sync_free(&mem->sync); - /* If the SG is not NULL, it's one we created for a doorbell - * BO. We need to free it. + /* If the SG is not NULL, it's one we created for a doorbell or mmio + * remap BO. We need to free it. */ if (mem->bo->tbo.sg) { sg_free_table(mem->bo->tbo.sg); kfree(mem->bo->tbo.sg); } + /* Update the size of the BO being freed if it was allocated from + * VRAM and is not imported. For APP APU VRAM allocations are done + * in GTT domain + */ + if (size) { + if (!is_imported && + mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) + *size = bo_size; + else + *size = 0; + } + /* Free the BO*/ - amdgpu_bo_unref(&mem->bo); + drm_vma_node_revoke(&mem->bo->tbo.base.vma_node, drm_priv); + drm_gem_handle_delete(adev->kfd.client.file, mem->gem_handle); + if (mem->dmabuf) { + dma_buf_put(mem->dmabuf); + mem->dmabuf = NULL; + } mutex_destroy(&mem->lock); - kfree(mem); + + /* If this releases the last reference, it will end up calling + * amdgpu_amdkfd_release_notify and kfree the mem struct. That's why + * this needs to be the last call here. + */ + drm_gem_object_put(&mem->bo->tbo.base); + + /* + * For kgd_mem allocated in amdgpu_amdkfd_gpuvm_import_dmabuf(), + * explicitly free it here. + */ + if (!use_release_notifier) + kfree(mem); return ret; } int amdgpu_amdkfd_gpuvm_map_memory_to_gpu( - struct kgd_dev *kgd, struct kgd_mem *mem, void *vm) + struct amdgpu_device *adev, struct kgd_mem *mem, + void *drm_priv) { - struct amdgpu_device *adev = get_amdgpu_device(kgd); - struct amdgpu_vm *avm = (struct amdgpu_vm *)vm; + struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv); int ret; struct amdgpu_bo *bo; uint32_t domain; - struct kfd_bo_va_list *entry; + struct kfd_mem_attachment *entry; struct bo_vm_reservation_context ctx; - struct kfd_bo_va_list *bo_va_entry = NULL; - struct kfd_bo_va_list *bo_va_entry_aql = NULL; unsigned long bo_size; bool is_invalid_userptr = false; @@ -1447,27 +2021,33 @@ int amdgpu_amdkfd_gpuvm_map_memory_to_gpu( */ mutex_lock(&mem->process_info->lock); - /* Lock mmap-sem. If we find an invalid userptr BO, we can be + /* Lock notifier lock. If we find an invalid userptr BO, we can be * sure that the MMU notifier is no longer running * concurrently and the queues are actually stopped */ if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) { - down_write(¤t->mm->mmap_sem); - is_invalid_userptr = atomic_read(&mem->invalid); - up_write(¤t->mm->mmap_sem); + mutex_lock(&mem->process_info->notifier_lock); + is_invalid_userptr = !!mem->invalid; + mutex_unlock(&mem->process_info->notifier_lock); } mutex_lock(&mem->lock); domain = mem->domain; - bo_size = bo->tbo.mem.size; + bo_size = bo->tbo.base.size; pr_debug("Map VA 0x%llx - 0x%llx to vm %p domain %s\n", mem->va, mem->va + bo_size * (1 + mem->aql_queue), - vm, domain_string(domain)); + avm, domain_string(domain)); - ret = reserve_bo_and_vm(mem, vm, &ctx); + if (!kfd_mem_is_attached(avm, mem)) { + ret = kfd_mem_attach(adev, mem, avm, mem->aql_queue); + if (ret) + goto out; + } + + ret = reserve_bo_and_vm(mem, avm, &ctx); if (unlikely(ret)) goto out; @@ -1477,80 +2057,44 @@ int amdgpu_amdkfd_gpuvm_map_memory_to_gpu( * the next restore worker */ if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) && - bo->tbo.mem.mem_type == TTM_PL_SYSTEM) + bo->tbo.resource->mem_type == TTM_PL_SYSTEM) is_invalid_userptr = true; - if (check_if_add_bo_to_vm(avm, mem)) { - ret = add_bo_to_vm(adev, mem, avm, false, - &bo_va_entry); - if (ret) - goto add_bo_to_vm_failed; - if (mem->aql_queue) { - ret = add_bo_to_vm(adev, mem, avm, - true, &bo_va_entry_aql); - if (ret) - goto add_bo_to_vm_failed_aql; - } - } else { - ret = vm_validate_pt_pd_bos(avm); - if (unlikely(ret)) - goto add_bo_to_vm_failed; - } - - if (mem->mapped_to_gpu_memory == 0 && - !amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) { - /* Validate BO only once. The eviction fence gets added to BO - * the first time it is mapped. Validate will wait for all - * background evictions to complete. - */ - ret = amdgpu_amdkfd_bo_validate(bo, domain, true); - if (ret) { - pr_debug("Validate failed\n"); - goto map_bo_to_gpuvm_failed; - } - } + ret = vm_validate_pt_pd_bos(avm, NULL); + if (unlikely(ret)) + goto out_unreserve; - list_for_each_entry(entry, &mem->bo_va_list, bo_list) { - if (entry->bo_va->base.vm == vm && !entry->is_mapped) { - pr_debug("\t map VA 0x%llx - 0x%llx in entry %p\n", - entry->va, entry->va + bo_size, - entry); + list_for_each_entry(entry, &mem->attachments, list) { + if (entry->bo_va->base.vm != avm || entry->is_mapped) + continue; - ret = map_bo_to_gpuvm(adev, entry, ctx.sync, - is_invalid_userptr); - if (ret) { - pr_err("Failed to map radeon bo to gpuvm\n"); - goto map_bo_to_gpuvm_failed; - } + pr_debug("\t map VA 0x%llx - 0x%llx in entry %p\n", + entry->va, entry->va + bo_size, entry); - ret = vm_update_pds(vm, ctx.sync); - if (ret) { - pr_err("Failed to update page directories\n"); - goto map_bo_to_gpuvm_failed; - } + ret = map_bo_to_gpuvm(mem, entry, ctx.sync, + is_invalid_userptr); + if (ret) { + pr_err("Failed to map bo to gpuvm\n"); + goto out_unreserve; + } - entry->is_mapped = true; - mem->mapped_to_gpu_memory++; - pr_debug("\t INC mapping count %d\n", - mem->mapped_to_gpu_memory); + ret = vm_update_pds(avm, ctx.sync); + if (ret) { + pr_err("Failed to update page directories\n"); + goto out_unreserve; } + + entry->is_mapped = true; + mem->mapped_to_gpu_memory++; + pr_debug("\t INC mapping count %d\n", + mem->mapped_to_gpu_memory); } - if (!amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) && !bo->pin_count) - amdgpu_bo_fence(bo, - &avm->process_info->eviction_fence->base, - true); ret = unreserve_bo_and_vms(&ctx, false, false); goto out; -map_bo_to_gpuvm_failed: - if (bo_va_entry_aql) - remove_bo_from_vm(adev, bo_va_entry_aql, bo_size); -add_bo_to_vm_failed_aql: - if (bo_va_entry) - remove_bo_from_vm(adev, bo_va_entry, bo_size); -add_bo_to_vm_failed: +out_unreserve: unreserve_bo_and_vms(&ctx, false, false); out: mutex_unlock(&mem->process_info->lock); @@ -1558,20 +2102,49 @@ out: return ret; } +int amdgpu_amdkfd_gpuvm_dmaunmap_mem(struct kgd_mem *mem, void *drm_priv) +{ + struct kfd_mem_attachment *entry; + struct amdgpu_vm *vm; + int ret; + + vm = drm_priv_to_vm(drm_priv); + + mutex_lock(&mem->lock); + + ret = amdgpu_bo_reserve(mem->bo, true); + if (ret) + goto out; + + list_for_each_entry(entry, &mem->attachments, list) { + if (entry->bo_va->base.vm != vm) + continue; + if (entry->bo_va->base.bo->tbo.ttm && + !entry->bo_va->base.bo->tbo.ttm->sg) + continue; + + kfd_mem_dmaunmap_attachment(mem, entry); + } + + amdgpu_bo_unreserve(mem->bo); +out: + mutex_unlock(&mem->lock); + + return ret; +} + int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu( - struct kgd_dev *kgd, struct kgd_mem *mem, void *vm) + struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv) { - struct amdgpu_device *adev = get_amdgpu_device(kgd); - struct amdkfd_process_info *process_info = - ((struct amdgpu_vm *)vm)->process_info; - unsigned long bo_size = mem->bo->tbo.mem.size; - struct kfd_bo_va_list *entry; + struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv); + unsigned long bo_size = mem->bo->tbo.base.size; + struct kfd_mem_attachment *entry; struct bo_vm_reservation_context ctx; int ret; mutex_lock(&mem->lock); - ret = reserve_bo_and_cond_vms(mem, vm, BO_VM_MAPPED, &ctx); + ret = reserve_bo_and_cond_vms(mem, avm, BO_VM_MAPPED, &ctx); if (unlikely(ret)) goto out; /* If no VMs were reserved, it means the BO wasn't actually mapped */ @@ -1580,45 +2153,32 @@ int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu( goto unreserve_out; } - ret = vm_validate_pt_pd_bos((struct amdgpu_vm *)vm); + ret = vm_validate_pt_pd_bos(avm, NULL); if (unlikely(ret)) goto unreserve_out; pr_debug("Unmap VA 0x%llx - 0x%llx from vm %p\n", mem->va, mem->va + bo_size * (1 + mem->aql_queue), - vm); - - list_for_each_entry(entry, &mem->bo_va_list, bo_list) { - if (entry->bo_va->base.vm == vm && entry->is_mapped) { - pr_debug("\t unmap VA 0x%llx - 0x%llx from entry %p\n", - entry->va, - entry->va + bo_size, - entry); - - ret = unmap_bo_from_gpuvm(adev, entry, ctx.sync); - if (ret == 0) { - entry->is_mapped = false; - } else { - pr_err("failed to unmap VA 0x%llx\n", - mem->va); - goto unreserve_out; - } + avm); - mem->mapped_to_gpu_memory--; - pr_debug("\t DEC mapping count %d\n", - mem->mapped_to_gpu_memory); - } - } + list_for_each_entry(entry, &mem->attachments, list) { + if (entry->bo_va->base.vm != avm || !entry->is_mapped) + continue; - /* If BO is unmapped from all VMs, unfence it. It can be evicted if - * required. - */ - if (mem->mapped_to_gpu_memory == 0 && - !amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) && !mem->bo->pin_count) - amdgpu_amdkfd_remove_eviction_fence(mem->bo, - process_info->eviction_fence, - NULL, NULL); + pr_debug("\t unmap VA 0x%llx - 0x%llx from entry %p\n", + entry->va, entry->va + bo_size, entry); + + ret = unmap_bo_from_gpuvm(mem, entry, ctx.sync); + if (ret) + goto unreserve_out; + + entry->is_mapped = false; + + mem->mapped_to_gpu_memory--; + pr_debug("\t DEC mapping count %d\n", + mem->mapped_to_gpu_memory); + } unreserve_out: unreserve_bo_and_vms(&ctx, false, false); @@ -1628,7 +2188,7 @@ out: } int amdgpu_amdkfd_gpuvm_sync_memory( - struct kgd_dev *kgd, struct kgd_mem *mem, bool intr) + struct amdgpu_device *adev, struct kgd_mem *mem, bool intr) { struct amdgpu_sync sync; int ret; @@ -1644,8 +2204,69 @@ int amdgpu_amdkfd_gpuvm_sync_memory( return ret; } -int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_dev *kgd, - struct kgd_mem *mem, void **kptr, uint64_t *size) +/** + * amdgpu_amdkfd_map_gtt_bo_to_gart - Map BO to GART and increment reference count + * @bo: Buffer object to be mapped + * @bo_gart: Return bo reference + * + * Before return, bo reference count is incremented. To release the reference and unpin/ + * unmap the BO, call amdgpu_amdkfd_free_gtt_mem. + */ +int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_bo *bo, struct amdgpu_bo **bo_gart) +{ + int ret; + + ret = amdgpu_bo_reserve(bo, true); + if (ret) { + pr_err("Failed to reserve bo. ret %d\n", ret); + goto err_reserve_bo_failed; + } + + ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT); + if (ret) { + pr_err("Failed to pin bo. ret %d\n", ret); + goto err_pin_bo_failed; + } + + ret = amdgpu_ttm_alloc_gart(&bo->tbo); + if (ret) { + pr_err("Failed to bind bo to GART. ret %d\n", ret); + goto err_map_bo_gart_failed; + } + + amdgpu_amdkfd_remove_eviction_fence( + bo, bo->vm_bo->vm->process_info->eviction_fence); + + amdgpu_bo_unreserve(bo); + + *bo_gart = amdgpu_bo_ref(bo); + + return 0; + +err_map_bo_gart_failed: + amdgpu_bo_unpin(bo); +err_pin_bo_failed: + amdgpu_bo_unreserve(bo); +err_reserve_bo_failed: + + return ret; +} + +/** amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel() - Map a GTT BO for kernel CPU access + * + * @mem: Buffer object to be mapped for CPU access + * @kptr[out]: pointer in kernel CPU address space + * @size[out]: size of the buffer + * + * Pins the BO and maps it for kernel CPU access. The eviction fence is removed + * from the BO, since pinned BOs cannot be evicted. The bo must remain on the + * validate_list, so the GPU mapping can be restored after a page table was + * evicted. + * + * Return: 0 on success, error code on failure + */ +int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_mem *mem, + void **kptr, uint64_t *size) { int ret; struct amdgpu_bo *bo = mem->bo; @@ -1655,9 +2276,6 @@ int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_dev *kgd, return -EINVAL; } - /* delete kgd_mem from kfd_bo_list to avoid re-validating - * this BO in BO's restoring after eviction. - */ mutex_lock(&mem->process_info->lock); ret = amdgpu_bo_reserve(bo, true); @@ -1679,8 +2297,7 @@ int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_dev *kgd, } amdgpu_amdkfd_remove_eviction_fence( - bo, mem->process_info->eviction_fence, NULL, NULL); - list_del_init(&mem->validate_list.head); + bo, mem->process_info->eviction_fence); if (size) *size = amdgpu_bo_size(bo); @@ -1700,39 +2317,44 @@ bo_reserve_failed: return ret; } -int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct kgd_dev *kgd, - struct kfd_vm_fault_info *mem) +/** amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel() - Unmap a GTT BO for kernel CPU access + * + * @mem: Buffer object to be unmapped for CPU access + * + * Removes the kernel CPU mapping and unpins the BO. It does not restore the + * eviction fence, so this function should only be used for cleanup before the + * BO is destroyed. + */ +void amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(struct kgd_mem *mem) { - struct amdgpu_device *adev; + struct amdgpu_bo *bo = mem->bo; - adev = (struct amdgpu_device *)kgd; - if (atomic_read(&adev->gmc.vm_fault_info_updated) == 1) { + (void)amdgpu_bo_reserve(bo, true); + amdgpu_bo_kunmap(bo); + amdgpu_bo_unpin(bo); + amdgpu_bo_unreserve(bo); +} + +int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct amdgpu_device *adev, + struct kfd_vm_fault_info *mem) +{ + if (atomic_read_acquire(&adev->gmc.vm_fault_info_updated) == 1) { *mem = *adev->gmc.vm_fault_info; - mb(); - atomic_set(&adev->gmc.vm_fault_info_updated, 0); + atomic_set_release(&adev->gmc.vm_fault_info_updated, 0); } return 0; } -int amdgpu_amdkfd_gpuvm_import_dmabuf(struct kgd_dev *kgd, - struct dma_buf *dma_buf, - uint64_t va, void *vm, - struct kgd_mem **mem, uint64_t *size, - uint64_t *mmap_offset) +static int import_obj_create(struct amdgpu_device *adev, + struct dma_buf *dma_buf, + struct drm_gem_object *obj, + uint64_t va, void *drm_priv, + struct kgd_mem **mem, uint64_t *size, + uint64_t *mmap_offset) { - struct amdgpu_device *adev = (struct amdgpu_device *)kgd; - struct drm_gem_object *obj; + struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv); struct amdgpu_bo *bo; - struct amdgpu_vm *avm = (struct amdgpu_vm *)vm; - - if (dma_buf->ops != &amdgpu_dmabuf_ops) - /* Can't handle non-graphics buffers */ - return -EINVAL; - - obj = dma_buf->priv; - if (obj->dev->dev_private != adev) - /* Can't handle buffers from other devices */ - return -EINVAL; + int ret; bo = gem_to_amdgpu_bo(obj); if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM | @@ -1744,58 +2366,150 @@ int amdgpu_amdkfd_gpuvm_import_dmabuf(struct kgd_dev *kgd, if (!*mem) return -ENOMEM; + ret = drm_vma_node_allow(&obj->vma_node, drm_priv); + if (ret) + goto err_free_mem; + if (size) *size = amdgpu_bo_size(bo); if (mmap_offset) *mmap_offset = amdgpu_bo_mmap_offset(bo); - INIT_LIST_HEAD(&(*mem)->bo_va_list); + INIT_LIST_HEAD(&(*mem)->attachments); mutex_init(&(*mem)->lock); - (*mem)->mapping_flags = - AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_PAGE_WRITEABLE | - AMDGPU_VM_PAGE_EXECUTABLE | AMDGPU_VM_MTYPE_NC; - (*mem)->bo = amdgpu_bo_ref(bo); + (*mem)->alloc_flags = + ((bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ? + KFD_IOC_ALLOC_MEM_FLAGS_VRAM : KFD_IOC_ALLOC_MEM_FLAGS_GTT) + | KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE + | KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE; + + get_dma_buf(dma_buf); + (*mem)->dmabuf = dma_buf; + (*mem)->bo = bo; (*mem)->va = va; - (*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ? - AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT; + (*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) && + !adev->apu_prefer_gtt ? + AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT; + (*mem)->mapped_to_gpu_memory = 0; (*mem)->process_info = avm->process_info; add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, false); amdgpu_sync_create(&(*mem)->sync); + (*mem)->is_imported = true; + + mutex_lock(&avm->process_info->lock); + if (avm->process_info->eviction_fence && + !dma_fence_is_signaled(&avm->process_info->eviction_fence->base)) + ret = amdgpu_amdkfd_bo_validate_and_fence(bo, (*mem)->domain, + &avm->process_info->eviction_fence->base); + mutex_unlock(&avm->process_info->lock); + if (ret) + goto err_remove_mem; + + return 0; + +err_remove_mem: + remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info); + drm_vma_node_revoke(&obj->vma_node, drm_priv); +err_free_mem: + kfree(*mem); + return ret; +} + +int amdgpu_amdkfd_gpuvm_import_dmabuf_fd(struct amdgpu_device *adev, int fd, + uint64_t va, void *drm_priv, + struct kgd_mem **mem, uint64_t *size, + uint64_t *mmap_offset) +{ + struct drm_gem_object *obj; + uint32_t handle; + int ret; + + ret = drm_gem_prime_fd_to_handle(&adev->ddev, adev->kfd.client.file, fd, + &handle); + if (ret) + return ret; + obj = drm_gem_object_lookup(adev->kfd.client.file, handle); + if (!obj) { + ret = -EINVAL; + goto err_release_handle; + } + + ret = import_obj_create(adev, obj->dma_buf, obj, va, drm_priv, mem, size, + mmap_offset); + if (ret) + goto err_put_obj; + + (*mem)->gem_handle = handle; return 0; + +err_put_obj: + drm_gem_object_put(obj); +err_release_handle: + drm_gem_handle_delete(adev->kfd.client.file, handle); + return ret; +} + +int amdgpu_amdkfd_gpuvm_export_dmabuf(struct kgd_mem *mem, + struct dma_buf **dma_buf) +{ + int ret; + + mutex_lock(&mem->lock); + ret = kfd_mem_export_dmabuf(mem); + if (ret) + goto out; + + get_dma_buf(mem->dmabuf); + *dma_buf = mem->dmabuf; +out: + mutex_unlock(&mem->lock); + return ret; } /* Evict a userptr BO by stopping the queues if necessary * * Runs in MMU notifier, may be in RECLAIM_FS context. This means it * cannot do any memory allocations, and cannot take any locks that - * are held elsewhere while allocating memory. Therefore this is as - * simple as possible, using atomic counters. + * are held elsewhere while allocating memory. * * It doesn't do anything to the BO itself. The real work happens in * restore, where we get updated page addresses. This function only * ensures that GPU access to the BO is stopped. */ -int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem, - struct mm_struct *mm) +int amdgpu_amdkfd_evict_userptr(struct mmu_interval_notifier *mni, + unsigned long cur_seq, struct kgd_mem *mem) { struct amdkfd_process_info *process_info = mem->process_info; - int invalid, evicted_bos; int r = 0; - invalid = atomic_inc_return(&mem->invalid); - evicted_bos = atomic_inc_return(&process_info->evicted_bos); - if (evicted_bos == 1) { + /* Do not process MMU notifications during CRIU restore until + * KFD_CRIU_OP_RESUME IOCTL is received + */ + if (READ_ONCE(process_info->block_mmu_notifications)) + return 0; + + mutex_lock(&process_info->notifier_lock); + mmu_interval_set_seq(mni, cur_seq); + + mem->invalid++; + if (++process_info->evicted_bos == 1) { /* First eviction, stop the queues */ - r = kgd2kfd->quiesce_mm(mm); - if (r) + r = kgd2kfd_quiesce_mm(mni->mm, + KFD_QUEUE_EVICTION_TRIGGER_USERPTR); + + if (r && r != -ESRCH) pr_err("Failed to quiesce KFD\n"); - schedule_delayed_work(&process_info->restore_userptr_work, - msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS)); + + if (r != -ESRCH) + queue_delayed_work(system_freezable_wq, + &process_info->restore_userptr_work, + msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS)); } + mutex_unlock(&process_info->notifier_lock); return r; } @@ -1812,165 +2526,174 @@ static int update_invalid_user_pages(struct amdkfd_process_info *process_info, struct kgd_mem *mem, *tmp_mem; struct amdgpu_bo *bo; struct ttm_operation_ctx ctx = { false, false }; - int invalid, ret; + uint32_t invalid; + int ret = 0; - /* Move all invalidated BOs to the userptr_inval_list and - * release their user pages by migration to the CPU domain - */ + mutex_lock(&process_info->notifier_lock); + + /* Move all invalidated BOs to the userptr_inval_list */ list_for_each_entry_safe(mem, tmp_mem, &process_info->userptr_valid_list, - validate_list.head) { - if (!atomic_read(&mem->invalid)) - continue; /* BO is still valid */ - - bo = mem->bo; - - if (amdgpu_bo_reserve(bo, true)) - return -EAGAIN; - amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU); - ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); - amdgpu_bo_unreserve(bo); - if (ret) { - pr_err("%s: Failed to invalidate userptr BO\n", - __func__); - return -EAGAIN; - } - - list_move_tail(&mem->validate_list.head, - &process_info->userptr_inval_list); - } - - if (list_empty(&process_info->userptr_inval_list)) - return 0; /* All evicted userptr BOs were freed */ + validate_list) + if (mem->invalid) + list_move_tail(&mem->validate_list, + &process_info->userptr_inval_list); /* Go through userptr_inval_list and update any invalid user_pages */ list_for_each_entry(mem, &process_info->userptr_inval_list, - validate_list.head) { - invalid = atomic_read(&mem->invalid); + validate_list) { + invalid = mem->invalid; if (!invalid) /* BO hasn't been invalidated since the last - * revalidation attempt. Keep its BO list. + * revalidation attempt. Keep its page list. */ continue; bo = mem->bo; - if (!mem->user_pages) { - mem->user_pages = - kvmalloc_array(bo->tbo.ttm->num_pages, - sizeof(struct page *), - GFP_KERNEL | __GFP_ZERO); - if (!mem->user_pages) { - pr_err("%s: Failed to allocate pages array\n", + amdgpu_hmm_range_free(mem->range); + mem->range = NULL; + + /* BO reservations and getting user pages (hmm_range_fault) + * must happen outside the notifier lock + */ + mutex_unlock(&process_info->notifier_lock); + + /* Move the BO to system (CPU) domain if necessary to unmap + * and free the SG table + */ + if (bo->tbo.resource->mem_type != TTM_PL_SYSTEM) { + if (amdgpu_bo_reserve(bo, true)) + return -EAGAIN; + amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU); + ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); + amdgpu_bo_unreserve(bo); + if (ret) { + pr_err("%s: Failed to invalidate userptr BO\n", __func__); - return -ENOMEM; + return -EAGAIN; } - } else if (mem->user_pages[0]) { - release_pages(mem->user_pages, bo->tbo.ttm->num_pages); } + mem->range = amdgpu_hmm_range_alloc(NULL); + if (unlikely(!mem->range)) + return -ENOMEM; /* Get updated user pages */ - ret = amdgpu_ttm_tt_get_user_pages(bo->tbo.ttm, - mem->user_pages); + ret = amdgpu_ttm_tt_get_user_pages(bo, mem->range); if (ret) { - mem->user_pages[0] = NULL; - pr_info("%s: Failed to get user pages: %d\n", - __func__, ret); - /* Pretend it succeeded. It will fail later - * with a VM fault if the GPU tries to access - * it. Better than hanging indefinitely with - * stalled user mode queues. + amdgpu_hmm_range_free(mem->range); + mem->range = NULL; + pr_debug("Failed %d to get user pages\n", ret); + + /* Return -EFAULT bad address error as success. It will + * fail later with a VM fault if the GPU tries to access + * it. Better than hanging indefinitely with stalled + * user mode queues. + * + * Return other error -EBUSY or -ENOMEM to retry restore */ + if (ret != -EFAULT) + return ret; + + /* If applications unmap memory before destroying the userptr + * from the KFD, trigger a segmentation fault in VM debug mode. + */ + if (amdgpu_ttm_adev(bo->tbo.bdev)->debug_vm_userptr) { + struct kfd_process *p; + + pr_err("Pid %d unmapped memory before destroying userptr at GPU addr 0x%llx\n", + pid_nr(process_info->pid), mem->va); + + // Send GPU VM fault to user space + p = kfd_lookup_process_by_pid(process_info->pid); + if (p) { + kfd_signal_vm_fault_event_with_userptr(p, mem->va); + kfd_unref_process(p); + } + } + + ret = 0; } + amdgpu_ttm_tt_set_user_pages(bo->tbo.ttm, mem->range); + + mutex_lock(&process_info->notifier_lock); + /* Mark the BO as valid unless it was invalidated - * again concurrently + * again concurrently. */ - if (atomic_cmpxchg(&mem->invalid, invalid, 0) != invalid) - return -EAGAIN; + if (mem->invalid != invalid) { + ret = -EAGAIN; + goto unlock_out; + } + /* set mem valid if mem has hmm range associated */ + if (mem->range) + mem->invalid = 0; } - return 0; +unlock_out: + mutex_unlock(&process_info->notifier_lock); + + return ret; } /* Validate invalid userptr BOs * - * Validates BOs on the userptr_inval_list, and moves them back to the - * userptr_valid_list. Also updates GPUVM page tables with new page - * addresses and waits for the page table updates to complete. + * Validates BOs on the userptr_inval_list. Also updates GPUVM page tables + * with new page addresses and waits for the page table updates to complete. */ static int validate_invalid_user_pages(struct amdkfd_process_info *process_info) { - struct amdgpu_bo_list_entry *pd_bo_list_entries; - struct list_head resv_list, duplicates; - struct ww_acquire_ctx ticket; + struct ttm_operation_ctx ctx = { false, false }; struct amdgpu_sync sync; + struct drm_exec exec; struct amdgpu_vm *peer_vm; struct kgd_mem *mem, *tmp_mem; struct amdgpu_bo *bo; - struct ttm_operation_ctx ctx = { false, false }; - int i, ret; - - pd_bo_list_entries = kcalloc(process_info->n_vms, - sizeof(struct amdgpu_bo_list_entry), - GFP_KERNEL); - if (!pd_bo_list_entries) { - pr_err("%s: Failed to allocate PD BO list entries\n", __func__); - return -ENOMEM; - } - - INIT_LIST_HEAD(&resv_list); - INIT_LIST_HEAD(&duplicates); + int ret; - /* Get all the page directory BOs that need to be reserved */ - i = 0; - list_for_each_entry(peer_vm, &process_info->vm_list_head, - vm_list_node) - amdgpu_vm_get_pd_bo(peer_vm, &resv_list, - &pd_bo_list_entries[i++]); - /* Add the userptr_inval_list entries to resv_list */ - list_for_each_entry(mem, &process_info->userptr_inval_list, - validate_list.head) { - list_add_tail(&mem->resv_list.head, &resv_list); - mem->resv_list.bo = mem->validate_list.bo; - mem->resv_list.num_shared = mem->validate_list.num_shared; - } + amdgpu_sync_create(&sync); + drm_exec_init(&exec, 0, 0); /* Reserve all BOs and page tables for validation */ - ret = ttm_eu_reserve_buffers(&ticket, &resv_list, false, &duplicates); - WARN(!list_empty(&duplicates), "Duplicates should be empty"); - if (ret) - goto out; + drm_exec_until_all_locked(&exec) { + /* Reserve all the page directories */ + list_for_each_entry(peer_vm, &process_info->vm_list_head, + vm_list_node) { + ret = amdgpu_vm_lock_pd(peer_vm, &exec, 2); + drm_exec_retry_on_contention(&exec); + if (unlikely(ret)) + goto unreserve_out; + } - amdgpu_sync_create(&sync); + /* Reserve the userptr_inval_list entries to resv_list */ + list_for_each_entry(mem, &process_info->userptr_inval_list, + validate_list) { + struct drm_gem_object *gobj; - /* Avoid triggering eviction fences when unmapping invalid - * userptr BOs (waits for all fences, doesn't use - * FENCE_OWNER_VM) - */ - list_for_each_entry(peer_vm, &process_info->vm_list_head, - vm_list_node) - amdgpu_amdkfd_remove_eviction_fence(peer_vm->root.base.bo, - process_info->eviction_fence, - NULL, NULL); + gobj = &mem->bo->tbo.base; + ret = drm_exec_prepare_obj(&exec, gobj, 1); + drm_exec_retry_on_contention(&exec); + if (unlikely(ret)) + goto unreserve_out; + } + } - ret = process_validate_vms(process_info); + ret = process_validate_vms(process_info, NULL); if (ret) goto unreserve_out; /* Validate BOs and update GPUVM page tables */ list_for_each_entry_safe(mem, tmp_mem, &process_info->userptr_inval_list, - validate_list.head) { - struct kfd_bo_va_list *bo_va_entry; + validate_list) { + struct kfd_mem_attachment *attachment; bo = mem->bo; - /* Copy pages array and validate the BO if we got user pages */ - if (mem->user_pages[0]) { - amdgpu_ttm_tt_set_user_pages(bo->tbo.ttm, - mem->user_pages); + /* Validate the BO if we got user pages */ + if (bo->tbo.ttm->pages[0]) { amdgpu_bo_placement_from_domain(bo, mem->domain); ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); if (ret) { @@ -1979,33 +2702,24 @@ static int validate_invalid_user_pages(struct amdkfd_process_info *process_info) } } - /* Validate succeeded, now the BO owns the pages, free - * our copy of the pointer array. Put this BO back on - * the userptr_valid_list. If we need to revalidate - * it, we need to start from scratch. - */ - kvfree(mem->user_pages); - mem->user_pages = NULL; - list_move_tail(&mem->validate_list.head, - &process_info->userptr_valid_list); - /* Update mapping. If the BO was not validated * (because we couldn't get user pages), this will * clear the page table entries, which will result in * VM faults if the GPU tries to access the invalid * memory. */ - list_for_each_entry(bo_va_entry, &mem->bo_va_list, bo_list) { - if (!bo_va_entry->is_mapped) + list_for_each_entry(attachment, &mem->attachments, list) { + if (!attachment->is_mapped) continue; - ret = update_gpuvm_pte((struct amdgpu_device *) - bo_va_entry->kgd_dev, - bo_va_entry, &sync); + kfd_mem_dmaunmap_attachment(mem, attachment); + ret = update_gpuvm_pte(mem, attachment, &sync); if (ret) { pr_err("%s: update PTE failed\n", __func__); /* make sure this gets validated again */ - atomic_inc(&mem->invalid); + mutex_lock(&process_info->notifier_lock); + mem->invalid++; + mutex_unlock(&process_info->notifier_lock); goto unreserve_out; } } @@ -2015,15 +2729,51 @@ static int validate_invalid_user_pages(struct amdkfd_process_info *process_info) ret = process_update_pds(process_info, &sync); unreserve_out: - list_for_each_entry(peer_vm, &process_info->vm_list_head, - vm_list_node) - amdgpu_bo_fence(peer_vm->root.base.bo, - &process_info->eviction_fence->base, true); - ttm_eu_backoff_reservation(&ticket, &resv_list); + drm_exec_fini(&exec); amdgpu_sync_wait(&sync, false); amdgpu_sync_free(&sync); -out: - kfree(pd_bo_list_entries); + + return ret; +} + +/* Confirm that all user pages are valid while holding the notifier lock + * + * Moves valid BOs from the userptr_inval_list back to userptr_val_list. + */ +static int confirm_valid_user_pages_locked(struct amdkfd_process_info *process_info) +{ + struct kgd_mem *mem, *tmp_mem; + int ret = 0; + + list_for_each_entry_safe(mem, tmp_mem, + &process_info->userptr_inval_list, + validate_list) { + bool valid; + + /* keep mem without hmm range at userptr_inval_list */ + if (!mem->range) + continue; + + /* Only check mem with hmm range associated */ + valid = amdgpu_hmm_range_valid(mem->range); + amdgpu_hmm_range_free(mem->range); + + mem->range = NULL; + if (!valid) { + WARN(!mem->invalid, "Invalid BO not marked invalid"); + ret = -EAGAIN; + continue; + } + + if (mem->invalid) { + WARN(1, "Valid BO is marked invalid"); + ret = -EAGAIN; + continue; + } + + list_move_tail(&mem->validate_list, + &process_info->userptr_valid_list); + } return ret; } @@ -2042,9 +2792,11 @@ static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work) restore_userptr_work); struct task_struct *usertask; struct mm_struct *mm; - int evicted_bos; + uint32_t evicted_bos; - evicted_bos = atomic_read(&process_info->evicted_bos); + mutex_lock(&process_info->notifier_lock); + evicted_bos = process_info->evicted_bos; + mutex_unlock(&process_info->notifier_lock); if (!evicted_bos) return; @@ -2067,9 +2819,6 @@ static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work) * and we can just restart the queues. */ if (!list_empty(&process_info->userptr_inval_list)) { - if (atomic_read(&process_info->evicted_bos) != evicted_bos) - goto unlock_out; /* Concurrent eviction, try again */ - if (validate_invalid_user_pages(process_info)) goto unlock_out; } @@ -2078,25 +2827,56 @@ static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work) * be a first eviction that calls quiesce_mm. The eviction * reference counting inside KFD will handle this case. */ - if (atomic_cmpxchg(&process_info->evicted_bos, evicted_bos, 0) != - evicted_bos) - goto unlock_out; - evicted_bos = 0; - if (kgd2kfd->resume_mm(mm)) { + mutex_lock(&process_info->notifier_lock); + if (process_info->evicted_bos != evicted_bos) + goto unlock_notifier_out; + + if (confirm_valid_user_pages_locked(process_info)) { + WARN(1, "User pages unexpectedly invalid"); + goto unlock_notifier_out; + } + + process_info->evicted_bos = evicted_bos = 0; + + if (kgd2kfd_resume_mm(mm)) { pr_err("%s: Failed to resume KFD\n", __func__); /* No recovery from this failure. Probably the CP is * hanging. No point trying again. */ } + +unlock_notifier_out: + mutex_unlock(&process_info->notifier_lock); unlock_out: mutex_unlock(&process_info->lock); - mmput(mm); - put_task_struct(usertask); /* If validation failed, reschedule another attempt */ - if (evicted_bos) - schedule_delayed_work(&process_info->restore_userptr_work, + if (evicted_bos) { + queue_delayed_work(system_freezable_wq, + &process_info->restore_userptr_work, msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS)); + + kfd_smi_event_queue_restore_rescheduled(mm); + } + mmput(mm); + put_task_struct(usertask); +} + +static void replace_eviction_fence(struct dma_fence __rcu **ef, + struct dma_fence *new_ef) +{ + struct dma_fence *old_ef = rcu_replace_pointer(*ef, new_ef, true + /* protected by process_info->lock */); + + /* If we're replacing an unsignaled eviction fence, that fence will + * never be signaled, and if anyone is still waiting on that fence, + * they will hang forever. This should never happen. We should only + * replace the fence in restore_work that only gets scheduled after + * eviction work signaled the fence. + */ + WARN_ONCE(!dma_fence_is_signaled(old_ef), + "Replacing unsignaled eviction fence"); + dma_fence_put(old_ef); } /** amdgpu_amdkfd_gpuvm_restore_process_bos - Restore all BOs for the given @@ -2117,89 +2897,108 @@ unlock_out: * 7. Add fence to all PD and PT BOs. * 8. Unreserve all BOs */ -int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef) +int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence __rcu **ef) { - struct amdgpu_bo_list_entry *pd_bo_list; struct amdkfd_process_info *process_info = info; struct amdgpu_vm *peer_vm; struct kgd_mem *mem; - struct bo_vm_reservation_context ctx; - struct amdgpu_amdkfd_fence *new_fence; - int ret = 0, i; struct list_head duplicate_save; struct amdgpu_sync sync_obj; + unsigned long failed_size = 0; + unsigned long total_size = 0; + struct drm_exec exec; + int ret; INIT_LIST_HEAD(&duplicate_save); - INIT_LIST_HEAD(&ctx.list); - INIT_LIST_HEAD(&ctx.duplicates); - pd_bo_list = kcalloc(process_info->n_vms, - sizeof(struct amdgpu_bo_list_entry), - GFP_KERNEL); - if (!pd_bo_list) - return -ENOMEM; - - i = 0; mutex_lock(&process_info->lock); - list_for_each_entry(peer_vm, &process_info->vm_list_head, - vm_list_node) - amdgpu_vm_get_pd_bo(peer_vm, &ctx.list, &pd_bo_list[i++]); - - /* Reserve all BOs and page tables/directory. Add all BOs from - * kfd_bo_list to ctx.list - */ - list_for_each_entry(mem, &process_info->kfd_bo_list, - validate_list.head) { - list_add_tail(&mem->resv_list.head, &ctx.list); - mem->resv_list.bo = mem->validate_list.bo; - mem->resv_list.num_shared = mem->validate_list.num_shared; - } + drm_exec_init(&exec, DRM_EXEC_IGNORE_DUPLICATES, 0); + drm_exec_until_all_locked(&exec) { + list_for_each_entry(peer_vm, &process_info->vm_list_head, + vm_list_node) { + ret = amdgpu_vm_lock_pd(peer_vm, &exec, 2); + drm_exec_retry_on_contention(&exec); + if (unlikely(ret)) { + pr_err("Locking VM PD failed, ret: %d\n", ret); + goto ttm_reserve_fail; + } + } - ret = ttm_eu_reserve_buffers(&ctx.ticket, &ctx.list, - false, &duplicate_save); - if (ret) { - pr_debug("Memory eviction: TTM Reserve Failed. Try again\n"); - goto ttm_reserve_fail; + /* Reserve all BOs and page tables/directory. Add all BOs from + * kfd_bo_list to ctx.list + */ + list_for_each_entry(mem, &process_info->kfd_bo_list, + validate_list) { + struct drm_gem_object *gobj; + + gobj = &mem->bo->tbo.base; + ret = drm_exec_prepare_obj(&exec, gobj, 1); + drm_exec_retry_on_contention(&exec); + if (unlikely(ret)) { + pr_err("drm_exec_prepare_obj failed, ret: %d\n", ret); + goto ttm_reserve_fail; + } + } } amdgpu_sync_create(&sync_obj); - /* Validate PDs and PTs */ - ret = process_validate_vms(process_info); - if (ret) - goto validate_map_fail; - - ret = process_sync_pds_resv(process_info, &sync_obj); - if (ret) { - pr_debug("Memory eviction: Failed to sync to PD BO moving fence. Try again\n"); - goto validate_map_fail; - } - - /* Validate BOs and map them to GPUVM (update VM page tables). */ + /* Validate BOs managed by KFD */ list_for_each_entry(mem, &process_info->kfd_bo_list, - validate_list.head) { + validate_list) { struct amdgpu_bo *bo = mem->bo; uint32_t domain = mem->domain; - struct kfd_bo_va_list *bo_va_entry; + struct dma_resv_iter cursor; + struct dma_fence *fence; + + total_size += amdgpu_bo_size(bo); ret = amdgpu_amdkfd_bo_validate(bo, domain, false); if (ret) { - pr_debug("Memory eviction: Validate BOs failed. Try again\n"); - goto validate_map_fail; + pr_debug("Memory eviction: Validate BOs failed\n"); + failed_size += amdgpu_bo_size(bo); + ret = amdgpu_amdkfd_bo_validate(bo, + AMDGPU_GEM_DOMAIN_GTT, false); + if (ret) { + pr_debug("Memory eviction: Try again\n"); + goto validate_map_fail; + } } - ret = amdgpu_sync_fence(NULL, &sync_obj, bo->tbo.moving, false); - if (ret) { - pr_debug("Memory eviction: Sync BO fence failed. Try again\n"); - goto validate_map_fail; + dma_resv_for_each_fence(&cursor, bo->tbo.base.resv, + DMA_RESV_USAGE_KERNEL, fence) { + ret = amdgpu_sync_fence(&sync_obj, fence, GFP_KERNEL); + if (ret) { + pr_debug("Memory eviction: Sync BO fence failed. Try again\n"); + goto validate_map_fail; + } } - list_for_each_entry(bo_va_entry, &mem->bo_va_list, - bo_list) { - ret = update_gpuvm_pte((struct amdgpu_device *) - bo_va_entry->kgd_dev, - bo_va_entry, - &sync_obj); + } + + if (failed_size) + pr_debug("0x%lx/0x%lx in system\n", failed_size, total_size); + + /* Validate PDs, PTs and evicted DMABuf imports last. Otherwise BO + * validations above would invalidate DMABuf imports again. + */ + ret = process_validate_vms(process_info, &exec.ticket); + if (ret) { + pr_debug("Validating VMs failed, ret: %d\n", ret); + goto validate_map_fail; + } + + /* Update mappings managed by KFD. */ + list_for_each_entry(mem, &process_info->kfd_bo_list, + validate_list) { + struct kfd_mem_attachment *attachment; + + list_for_each_entry(attachment, &mem->attachments, list) { + if (!attachment->is_mapped) + continue; + + kfd_mem_dmaunmap_attachment(mem, attachment); + ret = update_gpuvm_pte(mem, attachment, &sync_obj); if (ret) { pr_debug("Memory eviction: update PTE failed. Try again\n"); goto validate_map_fail; @@ -2207,6 +3006,32 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef) } } + /* Update mappings not managed by KFD */ + list_for_each_entry(peer_vm, &process_info->vm_list_head, + vm_list_node) { + struct amdgpu_device *adev = amdgpu_ttm_adev( + peer_vm->root.bo->tbo.bdev); + + struct amdgpu_fpriv *fpriv = + container_of(peer_vm, struct amdgpu_fpriv, vm); + + ret = amdgpu_vm_bo_update(adev, fpriv->prt_va, false); + if (ret) { + dev_dbg(adev->dev, + "Memory eviction: handle PRT moved failed, pid %8d. Try again.\n", + pid_nr(process_info->pid)); + goto validate_map_fail; + } + + ret = amdgpu_vm_handle_moved(adev, peer_vm, &exec.ticket); + if (ret) { + dev_dbg(adev->dev, + "Memory eviction: handle moved failed, pid %8d. Try again.\n", + pid_nr(process_info->pid)); + goto validate_map_fail; + } + } + /* Update page directories */ ret = process_update_pds(process_info, &sync_obj); if (ret) { @@ -2214,44 +3039,214 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef) goto validate_map_fail; } + /* Sync with fences on all the page tables. They implicitly depend on any + * move fences from amdgpu_vm_handle_moved above. + */ + ret = process_sync_pds_resv(process_info, &sync_obj); + if (ret) { + pr_debug("Memory eviction: Failed to sync to PD BO moving fence. Try again\n"); + goto validate_map_fail; + } + /* Wait for validate and PT updates to finish */ amdgpu_sync_wait(&sync_obj, false); - /* Release old eviction fence and create new one, because fence only - * goes from unsignaled to signaled, fence cannot be reused. - * Use context and mm from the old fence. + /* The old eviction fence may be unsignaled if restore happens + * after a GPU reset or suspend/resume. Keep the old fence in that + * case. Otherwise release the old eviction fence and create new + * one, because fence only goes from unsignaled to signaled once + * and cannot be reused. Use context and mm from the old fence. + * + * If an old eviction fence signals after this check, that's OK. + * Anyone signaling an eviction fence must stop the queues first + * and schedule another restore worker. */ - new_fence = amdgpu_amdkfd_fence_create( + if (dma_fence_is_signaled(&process_info->eviction_fence->base)) { + struct amdgpu_amdkfd_fence *new_fence = + amdgpu_amdkfd_fence_create( process_info->eviction_fence->base.context, - process_info->eviction_fence->mm); - if (!new_fence) { - pr_err("Failed to create eviction fence\n"); - ret = -ENOMEM; - goto validate_map_fail; + process_info->eviction_fence->mm, + NULL); + + if (!new_fence) { + pr_err("Failed to create eviction fence\n"); + ret = -ENOMEM; + goto validate_map_fail; + } + dma_fence_put(&process_info->eviction_fence->base); + process_info->eviction_fence = new_fence; + replace_eviction_fence(ef, dma_fence_get(&new_fence->base)); + } else { + WARN_ONCE(*ef != &process_info->eviction_fence->base, + "KFD eviction fence doesn't match KGD process_info"); } - dma_fence_put(&process_info->eviction_fence->base); - process_info->eviction_fence = new_fence; - *ef = dma_fence_get(&new_fence->base); - /* Attach new eviction fence to all BOs */ - list_for_each_entry(mem, &process_info->kfd_bo_list, - validate_list.head) - amdgpu_bo_fence(mem->bo, - &process_info->eviction_fence->base, true); + /* Attach new eviction fence to all BOs except pinned ones */ + list_for_each_entry(mem, &process_info->kfd_bo_list, validate_list) { + if (mem->bo->tbo.pin_count) + continue; - /* Attach eviction fence to PD / PT BOs */ + dma_resv_add_fence(mem->bo->tbo.base.resv, + &process_info->eviction_fence->base, + DMA_RESV_USAGE_BOOKKEEP); + } + /* Attach eviction fence to PD / PT BOs and DMABuf imports */ list_for_each_entry(peer_vm, &process_info->vm_list_head, vm_list_node) { - struct amdgpu_bo *bo = peer_vm->root.base.bo; + struct amdgpu_bo *bo = peer_vm->root.bo; - amdgpu_bo_fence(bo, &process_info->eviction_fence->base, true); + dma_resv_add_fence(bo->tbo.base.resv, + &process_info->eviction_fence->base, + DMA_RESV_USAGE_BOOKKEEP); } validate_map_fail: - ttm_eu_backoff_reservation(&ctx.ticket, &ctx.list); amdgpu_sync_free(&sync_obj); ttm_reserve_fail: + drm_exec_fini(&exec); mutex_unlock(&process_info->lock); - kfree(pd_bo_list); return ret; } + +int amdgpu_amdkfd_add_gws_to_process(void *info, void *gws, struct kgd_mem **mem) +{ + struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info; + struct amdgpu_bo *gws_bo = (struct amdgpu_bo *)gws; + int ret; + + if (!info || !gws) + return -EINVAL; + + *mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL); + if (!*mem) + return -ENOMEM; + + mutex_init(&(*mem)->lock); + INIT_LIST_HEAD(&(*mem)->attachments); + (*mem)->bo = amdgpu_bo_ref(gws_bo); + (*mem)->domain = AMDGPU_GEM_DOMAIN_GWS; + (*mem)->process_info = process_info; + add_kgd_mem_to_kfd_bo_list(*mem, process_info, false); + amdgpu_sync_create(&(*mem)->sync); + + + /* Validate gws bo the first time it is added to process */ + mutex_lock(&(*mem)->process_info->lock); + ret = amdgpu_bo_reserve(gws_bo, false); + if (unlikely(ret)) { + pr_err("Reserve gws bo failed %d\n", ret); + goto bo_reservation_failure; + } + + ret = amdgpu_amdkfd_bo_validate(gws_bo, AMDGPU_GEM_DOMAIN_GWS, true); + if (ret) { + pr_err("GWS BO validate failed %d\n", ret); + goto bo_validation_failure; + } + /* GWS resource is shared b/t amdgpu and amdkfd + * Add process eviction fence to bo so they can + * evict each other. + */ + ret = dma_resv_reserve_fences(gws_bo->tbo.base.resv, 1); + if (ret) + goto reserve_shared_fail; + dma_resv_add_fence(gws_bo->tbo.base.resv, + &process_info->eviction_fence->base, + DMA_RESV_USAGE_BOOKKEEP); + amdgpu_bo_unreserve(gws_bo); + mutex_unlock(&(*mem)->process_info->lock); + + return ret; + +reserve_shared_fail: +bo_validation_failure: + amdgpu_bo_unreserve(gws_bo); +bo_reservation_failure: + mutex_unlock(&(*mem)->process_info->lock); + amdgpu_sync_free(&(*mem)->sync); + remove_kgd_mem_from_kfd_bo_list(*mem, process_info); + amdgpu_bo_unref(&gws_bo); + mutex_destroy(&(*mem)->lock); + kfree(*mem); + *mem = NULL; + return ret; +} + +int amdgpu_amdkfd_remove_gws_from_process(void *info, void *mem) +{ + int ret; + struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info; + struct kgd_mem *kgd_mem = (struct kgd_mem *)mem; + struct amdgpu_bo *gws_bo = kgd_mem->bo; + + /* Remove BO from process's validate list so restore worker won't touch + * it anymore + */ + remove_kgd_mem_from_kfd_bo_list(kgd_mem, process_info); + + ret = amdgpu_bo_reserve(gws_bo, false); + if (unlikely(ret)) { + pr_err("Reserve gws bo failed %d\n", ret); + //TODO add BO back to validate_list? + return ret; + } + amdgpu_amdkfd_remove_eviction_fence(gws_bo, + process_info->eviction_fence); + amdgpu_bo_unreserve(gws_bo); + amdgpu_sync_free(&kgd_mem->sync); + amdgpu_bo_unref(&gws_bo); + mutex_destroy(&kgd_mem->lock); + kfree(mem); + return 0; +} + +/* Returns GPU-specific tiling mode information */ +int amdgpu_amdkfd_get_tile_config(struct amdgpu_device *adev, + struct tile_config *config) +{ + config->gb_addr_config = adev->gfx.config.gb_addr_config; + config->tile_config_ptr = adev->gfx.config.tile_mode_array; + config->num_tile_configs = + ARRAY_SIZE(adev->gfx.config.tile_mode_array); + config->macro_tile_config_ptr = + adev->gfx.config.macrotile_mode_array; + config->num_macro_tile_configs = + ARRAY_SIZE(adev->gfx.config.macrotile_mode_array); + + /* Those values are not set from GFX9 onwards */ + config->num_banks = adev->gfx.config.num_banks; + config->num_ranks = adev->gfx.config.num_ranks; + + return 0; +} + +bool amdgpu_amdkfd_bo_mapped_to_dev(void *drm_priv, struct kgd_mem *mem) +{ + struct amdgpu_vm *vm = drm_priv_to_vm(drm_priv); + struct kfd_mem_attachment *entry; + + list_for_each_entry(entry, &mem->attachments, list) { + if (entry->is_mapped && entry->bo_va->base.vm == vm) + return true; + } + return false; +} + +#if defined(CONFIG_DEBUG_FS) + +int kfd_debugfs_kfd_mem_limits(struct seq_file *m, void *data) +{ + + spin_lock(&kfd_mem_limit.mem_limit_lock); + seq_printf(m, "System mem used %lldM out of %lluM\n", + (kfd_mem_limit.system_mem_used >> 20), + (kfd_mem_limit.max_system_mem_limit >> 20)); + seq_printf(m, "TTM mem used %lldM out of %lluM\n", + (kfd_mem_limit.ttm_mem_used >> 20), + (kfd_mem_limit.max_ttm_mem_limit >> 20)); + spin_unlock(&kfd_mem_limit.mem_limit_lock); + + return 0; +} + +#endif |