Re: 回复: [RFC PATCH v2] drm/amdgpu: Remove kfd eviction fence before release bo
Christian König
christian.koenig at amd.com
Fri Feb 7 13:52:47 UTC 2020
Am 07.02.20 um 14:50 schrieb Pan, Xinhui:
> [AMD Official Use Only - Internal Distribution Only]
>
> ________________________________________
> 发件人: Koenig, Christian <Christian.Koenig at amd.com>
> 发送时间: 2020年2月7日 21:46
> 收件人: Pan, Xinhui; amd-gfx at lists.freedesktop.org
> 抄送: Kuehling, Felix; Deucher, Alexander
> 主题: Re: [RFC PATCH v2] drm/amdgpu: Remove kfd eviction fence before release bo
>
> Am 07.02.20 um 14:42 schrieb Pan, Xinhui:
>> No need to trigger eviction as the memory mapping will not be used anymore.
>>
>> All pt/pd bos share same resv, hence the same shared eviction fence. Everytime page table is freed, the fence will be signled and that cuases kfd unexcepted evictions.
>>
>> Signed-off-by: xinhui pab <xinhui.pan at example.com>
>> ---
>> drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 1 +
>> .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 78 ++++++++++++++++++-
>> drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 10 ++-
>> drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 5 +-
>> drivers/gpu/drm/ttm/ttm_bo.c | 38 +++++----
>> 5 files changed, 111 insertions(+), 21 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
>> index 47b0f2957d1f..265b1ed7264c 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
>> @@ -96,6 +96,7 @@ struct amdgpu_amdkfd_fence *amdgpu_amdkfd_fence_create(u64 context,
>> struct mm_struct *mm);
>> bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm);
>> struct amdgpu_amdkfd_fence *to_amdgpu_amdkfd_fence(struct dma_fence *f);
>> +int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo);
>>
>> struct amdkfd_process_info {
>> /* List head of all VMs that belong to a KFD process */
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
>> index ef721cb65868..11315095c29b 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
>> @@ -223,7 +223,7 @@ void amdgpu_amdkfd_unreserve_memory_limit(struct amdgpu_bo *bo)
>> static int amdgpu_amdkfd_remove_eviction_fence(struct amdgpu_bo *bo,
>> struct amdgpu_amdkfd_fence *ef)
>> {
>> - struct dma_resv *resv = bo->tbo.base.resv;
>> + struct dma_resv *resv = &bo->tbo.base._resv;
> That won't work either and probably break a bunch of other cases.
>
> [xh] kfd bos which are allocated explicitly use _resv AFAIK.
> only pt/pd bos share the root._resv
No that is not correct. Imports potentially don't use _resv here.
Anyway please see the patchset I've pointed out in one of my last mails.
With that we can assign bo->tbo.base.resv to bo->tbo._base for deleted BOs.
Christian.
>
> Christian.
>
>> struct dma_resv_list *old, *new;
>> unsigned int i, j, k;
>>
>> @@ -276,6 +276,78 @@ static int amdgpu_amdkfd_remove_eviction_fence(struct amdgpu_bo *bo,
>> return 0;
>> }
>>
>> +int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo)
>> +{
>> + struct amdgpu_vm_bo_base *vm_bo = bo->vm_bo;
>> + struct amdgpu_vm *vm;
>> + struct amdkfd_process_info *info;
>> + struct amdgpu_amdkfd_fence *ef;
>> + struct amdgpu_bo *parent;
>> + int locked;
>> + int ret;
>> + struct ttm_bo_global *glob = &ttm_bo_glob;
>> +
>> + if (vm_bo == NULL)
>> + return 0;
>> +
>> + /* Page table bo has a reference of the parent bo.
>> + * BO itself can't guarntee the vm it points to is alive.
>> + * for example, VM is going to free page tables, and the pt/pd bo might be
>> + * freed by a workqueue. In that case, the vm might be freed already,
>> + * leaving the bo->vm_bo points to vm.root.
>> + *
>> + * so to avoid that, when kfd free its vms,
>> + * 1) set vm->process_info to NULL if this is the last vm.
>> + * 2) set root_bo->vm_bo to NULL.
>> + *
>> + * but there are still races, just like
>> + * cpu 1 cpu 2
>> + * !vm_bo
>> + * ->info = NULL
>> + * free(info)
>> + * ->vm_bo = NULL
>> + * free (vm)
>> + * info = vm->info //invalid vm
>> + *
>> + * So to avoid the race, use ttm_bo_glob lru_lock.
>> + * generally speaking, adding a new lock is accceptable.
>> + * But reusing this lock is simple.
>> + */
>> + parent = bo;
>> + while (parent->parent)
>> + parent = parent->parent;
>> +
>> + spin_lock(&glob->lru_lock);
>> + vm_bo = parent->vm_bo;
>> + if (!vm_bo) {
>> + spin_unlock(&glob->lru_lock);
>> + return 0;
>> + }
>> +
>> + vm = vm_bo->vm;
>> + if (!vm) {
>> + spin_unlock(&glob->lru_lock);
>> + return 0;
>> + }
>> +
>> + info = vm->process_info;
>> + if (!info || !info->eviction_fence) {
>> + spin_unlock(&glob->lru_lock);
>> + return 0;
>> + }
>> +
>> + ef = container_of(dma_fence_get(&info->eviction_fence->base),
>> + struct amdgpu_amdkfd_fence, base);
>> + spin_unlock(&glob->lru_lock);
>> +
>> + locked = dma_resv_trylock(&bo->tbo.base._resv);
>> + ret = amdgpu_amdkfd_remove_eviction_fence(bo, ef);
>> + dma_fence_put(&ef->base);
>> + if (locked)
>> + dma_resv_unlock(&bo->tbo.base._resv);
>> + return ret;
>> +}
>> +
>> static int amdgpu_amdkfd_bo_validate(struct amdgpu_bo *bo, uint32_t domain,
>> bool wait)
>> {
>> @@ -1030,6 +1102,7 @@ void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
>> {
>> struct amdkfd_process_info *process_info = vm->process_info;
>> struct amdgpu_bo *pd = vm->root.base.bo;
>> + struct ttm_bo_global *glob = &ttm_bo_glob;
>>
>> if (!process_info)
>> return;
>> @@ -1051,6 +1124,9 @@ void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
>> WARN_ON(!list_empty(&process_info->userptr_valid_list));
>> WARN_ON(!list_empty(&process_info->userptr_inval_list));
>>
>> + spin_lock(&glob->lru_lock);
>> + vm->process_info = NULL;
>> + spin_unlock(&glob->lru_lock);
>> dma_fence_put(&process_info->eviction_fence->base);
>> cancel_delayed_work_sync(&process_info->restore_userptr_work);
>> put_pid(process_info->pid);
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
>> index 6f60a581e3ba..6ad12298fa84 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
>> @@ -1307,19 +1307,23 @@ void amdgpu_bo_release_notify(struct ttm_buffer_object *bo)
>> if (abo->kfd_bo)
>> amdgpu_amdkfd_unreserve_memory_limit(abo);
>>
>> + amdgpu_amdkfd_remove_fence_on_pt_pd_bos(abo);
>> + abo->vm_bo = NULL;
>> +
>> if (bo->mem.mem_type != TTM_PL_VRAM || !bo->mem.mm_node ||
>> !(abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE))
>> return;
>>
>> - dma_resv_lock(bo->base.resv, NULL);
>> + /* Only kfd bo wipe vram on release. The resv is &_resv. */
>> + dma_resv_lock(&bo->base._resv, NULL);
>>
>> - r = amdgpu_fill_buffer(abo, AMDGPU_POISON, bo->base.resv, &fence);
>> + r = amdgpu_fill_buffer(abo, AMDGPU_POISON, &bo->base._resv, &fence);
>> if (!WARN_ON(r)) {
>> amdgpu_bo_fence(abo, fence, false);
>> dma_fence_put(fence);
>> }
>>
>> - dma_resv_unlock(bo->base.resv);
>> + dma_resv_unlock(&bo->base._resv);
>> }
>>
>> /**
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>> index cc56eaba1911..2b96447e30e5 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>> @@ -945,7 +945,6 @@ static int amdgpu_vm_alloc_pts(struct amdgpu_device *adev,
>> static void amdgpu_vm_free_table(struct amdgpu_vm_pt *entry)
>> {
>> if (entry->base.bo) {
>> - entry->base.bo->vm_bo = NULL;
>> list_del(&entry->base.vm_status);
>> amdgpu_bo_unref(&entry->base.bo->shadow);
>> amdgpu_bo_unref(&entry->base.bo);
>> @@ -3074,6 +3073,7 @@ void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
>> struct amdgpu_bo_va_mapping *mapping, *tmp;
>> bool prt_fini_needed = !!adev->gmc.gmc_funcs->set_prt;
>> struct amdgpu_bo *root;
>> + struct ttm_bo_global *glob = &ttm_bo_glob;
>> int i;
>>
>> amdgpu_amdkfd_gpuvm_destroy_cb(adev, vm);
>> @@ -3105,6 +3105,9 @@ void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
>> }
>>
>> amdgpu_vm_free_pts(adev, vm, NULL);
>> + spin_lock(&glob->lru_lock);
>> + root->vm_bo = NULL;
>> + spin_unlock(&glob->lru_lock);
>> amdgpu_bo_unreserve(root);
>> amdgpu_bo_unref(&root);
>> WARN_ON(vm->root.base.bo);
>> diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
>> index 1494aebb8128..706cd60eb9e0 100644
>> --- a/drivers/gpu/drm/ttm/ttm_bo.c
>> +++ b/drivers/gpu/drm/ttm/ttm_bo.c
>> @@ -429,8 +429,8 @@ static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo)
>> BUG_ON(!dma_resv_trylock(&bo->base._resv));
>>
>> r = dma_resv_copy_fences(&bo->base._resv, bo->base.resv);
>> - if (r)
>> - dma_resv_unlock(&bo->base._resv);
>> +
>> + dma_resv_unlock(&bo->base._resv);
>>
>> return r;
>> }
>> @@ -455,23 +455,19 @@ static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
>> }
>> }
>>
>> -static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
>> +static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo, bool enqueue)
>> {
>> struct ttm_bo_device *bdev = bo->bdev;
>> int ret;
>>
>> - ret = ttm_bo_individualize_resv(bo);
>> - if (ret) {
>> - /* Last resort, if we fail to allocate memory for the
>> - * fences block for the BO to become idle
>> - */
>> - dma_resv_wait_timeout_rcu(bo->base.resv, true, false,
>> - 30 * HZ);
>> - spin_lock(&ttm_bo_glob.lru_lock);
>> - goto error;
>> - }
>> -
>> spin_lock(&ttm_bo_glob.lru_lock);
>> +
>> + if (enqueue)
>> + goto queue;
>> +
>> + if (bo->base.resv != &bo->base._resv)
>> + BUG_ON(!dma_resv_trylock(&bo->base._resv));
>> +
>> ret = dma_resv_trylock(bo->base.resv) ? 0 : -EBUSY;
>> if (!ret) {
>> if (dma_resv_test_signaled_rcu(&bo->base._resv, true)) {
>> @@ -504,7 +500,7 @@ static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
>> dma_resv_unlock(&bo->base._resv);
>> }
>>
>> -error:
>> +queue:
>> kref_get(&bo->list_kref);
>> list_add_tail(&bo->ddestroy, &bdev->ddestroy);
>> spin_unlock(&ttm_bo_glob.lru_lock);
>> @@ -655,6 +651,16 @@ static void ttm_bo_release(struct kref *kref)
>> container_of(kref, struct ttm_buffer_object, kref);
>> struct ttm_bo_device *bdev = bo->bdev;
>> struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
>> + int ret;
>> +
>> + ret = ttm_bo_individualize_resv(bo);
>> + if (ret) {
>> + /* Last resort, if we fail to allocate memory for the
>> + * fences block for the BO to become idle
>> + */
>> + dma_resv_wait_timeout_rcu(bo->base.resv, true, false,
>> + 30 * HZ);
>> + }
>>
>> if (bo->bdev->driver->release_notify)
>> bo->bdev->driver->release_notify(bo);
>> @@ -663,7 +669,7 @@ static void ttm_bo_release(struct kref *kref)
>> ttm_mem_io_lock(man, false);
>> ttm_mem_io_free_vm(bo);
>> ttm_mem_io_unlock(man);
>> - ttm_bo_cleanup_refs_or_queue(bo);
>> + ttm_bo_cleanup_refs_or_queue(bo, !!ret);
>> kref_put(&bo->list_kref, ttm_bo_release_list);
>> }
>>
More information about the amd-gfx
mailing list