[PATCH v5 1/1] drm/amdgpu: Fix amdgpu_vm_pt_free warning

Christian König christian.koenig at amd.com
Thu Sep 15 06:21:36 UTC 2022


Am 14.09.22 um 19:45 schrieb Felix Kuehling:
> Am 2022-09-14 um 12:08 schrieb Philip Yang:
>> Free page table BO from vm resv unlocked context generate below
>> warnings.
>>
>> Add a pt_free_work in vm to free page table BO from vm->pt_freed list.
>> pass vm resv unlock status from page table update caller, and add vm_bo
>> entry to vm->pt_freed list and schedule the pt_free_work if calling with
>> vm resv unlocked.
>>
>> WARNING: CPU: 12 PID: 3238 at
>> drivers/gpu/drm/ttm/ttm_bo.c:106 ttm_bo_set_bulk_move+0xa1/0xc0
>> Call Trace:
>>   amdgpu_vm_pt_free+0x42/0xd0 [amdgpu]
>>   amdgpu_vm_pt_free_dfs+0xb3/0xf0 [amdgpu]
>>   amdgpu_vm_ptes_update+0x52d/0x850 [amdgpu]
>>   amdgpu_vm_update_range+0x2a6/0x640 [amdgpu]
>>   svm_range_unmap_from_gpus+0x110/0x300 [amdgpu]
>>   svm_range_cpu_invalidate_pagetables+0x535/0x600 [amdgpu]
>>   __mmu_notifier_invalidate_range_start+0x1cd/0x230
>>   unmap_vmas+0x9d/0x140
>>   unmap_region+0xa8/0x110
>>
>> Signed-off-by: Philip Yang <Philip.Yang at amd.com>
>> ---
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c    |  5 +++
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h    |  6 ++++
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c | 41 +++++++++++++++++++++--
>>   3 files changed, 49 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c 
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>> index 59cac347baa3..20cfc8c9635b 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>> @@ -2022,6 +2022,9 @@ int amdgpu_vm_init(struct amdgpu_device *adev, 
>> struct amdgpu_vm *vm)
>>       spin_lock_init(&vm->invalidated_lock);
>>       INIT_LIST_HEAD(&vm->freed);
>>       INIT_LIST_HEAD(&vm->done);
>> +    INIT_LIST_HEAD(&vm->pt_freed);
>> +    INIT_WORK(&vm->pt_free_work, amdgpu_vm_pt_free_work);
>> +    spin_lock_init(&vm->pt_free_lock);
>>         /* create scheduler entities for page table updates */
>>       r = drm_sched_entity_init(&vm->immediate, 
>> DRM_SCHED_PRIORITY_NORMAL,
>> @@ -2223,6 +2226,8 @@ void amdgpu_vm_fini(struct amdgpu_device *adev, 
>> struct amdgpu_vm *vm)
>>         amdgpu_amdkfd_gpuvm_destroy_cb(adev, vm);
>>   +    flush_work(&vm->pt_free_work);
>> +
>>       root = amdgpu_bo_ref(vm->root.bo);
>>       amdgpu_bo_reserve(root, true);
>>       amdgpu_vm_set_pasid(adev, vm, 0);
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h 
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
>> index 9ecb7f663e19..b77fe838c327 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
>> @@ -276,6 +276,11 @@ struct amdgpu_vm {
>>       /* BOs which are invalidated, has been updated in the PTs */
>>       struct list_head        done;
>>   +    /* PT BOs scheduled to free and fill with zero if vm_resv is 
>> not hold */
>> +    struct list_head    pt_freed;
>> +    struct work_struct    pt_free_work;
>> +    spinlock_t        pt_free_lock;
>> +
>>       /* contains the page directory */
>>       struct amdgpu_vm_bo_base     root;
>>       struct dma_fence    *last_update;
>> @@ -471,6 +476,7 @@ int amdgpu_vm_pde_update(struct 
>> amdgpu_vm_update_params *params,
>>   int amdgpu_vm_ptes_update(struct amdgpu_vm_update_params *params,
>>                 uint64_t start, uint64_t end,
>>                 uint64_t dst, uint64_t flags);
>> +void amdgpu_vm_pt_free_work(struct work_struct *work);
>>     #if defined(CONFIG_DEBUG_FS)
>>   void amdgpu_debugfs_vm_bo_info(struct amdgpu_vm *vm, struct 
>> seq_file *m);
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c 
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c
>> index 88de9f0d4728..c16579471f22 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c
>> @@ -641,6 +641,27 @@ static void amdgpu_vm_pt_free(struct 
>> amdgpu_vm_bo_base *entry)
>>       amdgpu_bo_unref(&entry->bo);
>>   }
>>   +void amdgpu_vm_pt_free_work(struct work_struct *work)
>> +{
>> +    struct amdgpu_vm_bo_base *entry, *next;
>> +    struct amdgpu_vm *vm;
>> +    LIST_HEAD(pt_freed);
>> +
>> +    vm = container_of(work, struct amdgpu_vm, pt_free_work);
>> +
>> +    spin_lock(&vm->pt_free_lock);
>> +    list_splice_init(&vm->pt_freed, &pt_freed);
>> +    spin_unlock(&vm->pt_free_lock);
>> +
>> +    // flush_work in amdgpu_vm_fini ensure vm->root.bo is valid
>> +    amdgpu_bo_reserve(vm->root.bo, true);
>> +
>> +    list_for_each_entry_safe(entry, next, &pt_freed, vm_status)
>> +        amdgpu_vm_pt_free(entry);
>> +
>> +    amdgpu_bo_unreserve(vm->root.bo);
>> +}
>> +
>>   /**
>>    * amdgpu_vm_pt_free_dfs - free PD/PT levels
>>    *
>> @@ -652,11 +673,24 @@ static void amdgpu_vm_pt_free(struct 
>> amdgpu_vm_bo_base *entry)
>>    */
>>   static void amdgpu_vm_pt_free_dfs(struct amdgpu_device *adev,
>>                     struct amdgpu_vm *vm,
>> -                  struct amdgpu_vm_pt_cursor *start)
>> +                  struct amdgpu_vm_pt_cursor *start,
>> +                  bool unlocked)
>>   {
>>       struct amdgpu_vm_pt_cursor cursor;
>>       struct amdgpu_vm_bo_base *entry;
>>   +    if (unlocked) {
>> +        spin_lock(&vm->pt_free_lock);
>> +        for_each_amdgpu_vm_pt_dfs_safe(adev, vm, start, cursor, entry)
>> +            list_move(&entry->vm_status, &vm->pt_freed);
>> +
>> +        if (start)
>> +            list_move(&start->entry->vm_status, &vm->pt_freed);
>> +        spin_unlock(&vm->pt_free_lock);
>
> Question for Christian: list_move will take the entry off another 
> vm_status list (evicted, idle, invalidated, relocated). I don't think 
> this is safe without holding a reservation lock in most cases. Since 
> the point here is to maintain a list of PTs to be freed without 
> locking a reservation, is the vm_status really the right way to do this?

Oh, good point. This means we need to protect the whole state machine 
with a spinlock or use a separate deleted list.

I don't see how we can easily use a separate list here, but we already 
have the invalidated_lock, maybe we need to rename this one to status 
lock and use it for all cases.

Regards,
Christian.

>
> Regards,
>   Felix
>
>
>> + schedule_work(&vm->pt_free_work);
>> +        return;
>> +    }
>> +
>>       for_each_amdgpu_vm_pt_dfs_safe(adev, vm, start, cursor, entry)
>>           amdgpu_vm_pt_free(entry);
>>   @@ -673,7 +707,7 @@ static void amdgpu_vm_pt_free_dfs(struct 
>> amdgpu_device *adev,
>>    */
>>   void amdgpu_vm_pt_free_root(struct amdgpu_device *adev, struct 
>> amdgpu_vm *vm)
>>   {
>> -    amdgpu_vm_pt_free_dfs(adev, vm, NULL);
>> +    amdgpu_vm_pt_free_dfs(adev, vm, NULL, false);
>>   }
>>     /**
>> @@ -966,7 +1000,8 @@ int amdgpu_vm_ptes_update(struct 
>> amdgpu_vm_update_params *params,
>>                   if (cursor.entry->bo) {
>>                       params->table_freed = true;
>>                       amdgpu_vm_pt_free_dfs(adev, params->vm,
>> -                                  &cursor);
>> +                                  &cursor,
>> +                                  params->unlocked);
>>                   }
>>                   amdgpu_vm_pt_next(adev, &cursor);
>>               }



More information about the amd-gfx mailing list