[PATCH v2] drm/amdgpu: Unlocked unmap only clear page table leaves

[Christian König]

Am 23.01.25 um 21:39 schrieb Philip Yang:
> SVM migration unmap pages from GPU and then update mapping to GPU to
> recover page fault. Currently unmap clears the PDE entry for range
> length >= huge page and may free PTB bo. Then update mapping should
> alloc new PT bo, but there is race bug that the freed entry bo maybe
> still on the pt_free list, reused when updating mapping and then freed,
> leave invalid PDE entry and cause GPU page fault.
>
> By setting the update fragment size to 2MB or 1GB, update will clear
> only one PDE entry or clear PTB, to avoid unmap to free PTE bo. This
> fixes the race bug and also improve the unmap and map to GPU
> performance. Update mapping to huge page will still free the PTB bo.
>
> With this change, the vm->pt_freed list and work is not needed. Add
> WARN_ON(unlocked) in amdgpu_vm_pt_add_list to catch if unmap to free the
> PTB.
>
> v2: Limit update fragment size, not hack entry_end (Christian)
>
> Signed-off-by: Philip Yang <Philip.Yang at amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c    |  4 --
>   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h    |  4 --
>   drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c | 54 +++++++++--------------
>   3 files changed, 21 insertions(+), 41 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> index c9c48b782ec1..48b2c0b3b315 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> @@ -2440,8 +2440,6 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm,
>   	spin_lock_init(&vm->status_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);
>   	INIT_KFIFO(vm->faults);
>   
>   	r = amdgpu_vm_init_entities(adev, vm);
> @@ -2613,8 +2611,6 @@ 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_put_task_info(vm->task_info);
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> index 5d119ac26c4f..160889e5e64d 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> @@ -369,10 +369,6 @@ 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;
> -
>   	/* contains the page directory */
>   	struct amdgpu_vm_bo_base     root;
>   	struct dma_fence	*last_update;
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c
> index f78a0434a48f..063d0e0a9f29 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c
> @@ -546,27 +546,6 @@ 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->status_lock);
> -	list_splice_init(&vm->pt_freed, &pt_freed);
> -	spin_unlock(&vm->status_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_list - free PD/PT levels
>    *
> @@ -579,20 +558,9 @@ void amdgpu_vm_pt_free_list(struct amdgpu_device *adev,
>   			    struct amdgpu_vm_update_params *params)
>   {
>   	struct amdgpu_vm_bo_base *entry, *next;
> -	struct amdgpu_vm *vm = params->vm;
> -	bool unlocked = params->unlocked;
>   
>   	if (list_empty(&params->tlb_flush_waitlist))
>   		return;
> -
> -	if (unlocked) {
> -		spin_lock(&vm->status_lock);
> -		list_splice_init(&params->tlb_flush_waitlist, &vm->pt_freed);
> -		spin_unlock(&vm->status_lock);
> -		schedule_work(&vm->pt_free_work);
> -		return;
> -	}
> -
>   	list_for_each_entry_safe(entry, next, &params->tlb_flush_waitlist, vm_status)
>   		amdgpu_vm_pt_free(entry);
>   }
> @@ -611,6 +579,11 @@ static void amdgpu_vm_pt_add_list(struct amdgpu_vm_update_params *params,
>   	struct amdgpu_vm_pt_cursor seek;
>   	struct amdgpu_vm_bo_base *entry;
>   
> +	/*
> +	 * unlocked unmap only clear page table leaves, warning to free the page table entry.
> +	 */
> +	WARN_ON(params->unlocked);
> +
>   	spin_lock(&params->vm->status_lock);
>   	for_each_amdgpu_vm_pt_dfs_safe(params->adev, params->vm, cursor, seek, entry) {
>   		if (entry && entry->bo)
> @@ -794,6 +767,21 @@ static void amdgpu_vm_pte_fragment(struct amdgpu_vm_update_params *params,
>   
>   	/* This intentionally wraps around if no bit is set */
>   	*frag = min_t(unsigned int, ffs(start) - 1, fls64(end - start) - 1);
> +
> +	/*
> +	 * MMU notifier callback unlocked unmap only clear PDE or PTE leaves by setting fragment
> +	 * size to 2MB, 1GB, 512GB. If leave is PDE entry, only clear one entry, next fragment entry
> +	 * will search again for PDE or PTE leaves.
> +	 */
> +	if (params->unlocked) {
> +		if (*frag > 9 && *frag < 18)
> +			*frag = 9;
> +		else if (*frag > 18 && *frag < 27)
> +			*frag = 18;
> +		else if (*frag > 27)
> +			*frag = 27;
> +	}
> +

That here looks unnecessary, the higher level handling is able to pick 
the right PD/PT level based on the fragment anyway.

Apart from that the patch looks good to me.

Regards,
Christian.

>   	if (*frag >= max_frag) {
>   		*frag = max_frag;
>   		*frag_end = end & ~((1ULL << max_frag) - 1);
> @@ -931,7 +919,7 @@ int amdgpu_vm_ptes_update(struct amdgpu_vm_update_params *params,
>   				/* figure out the next fragment */
>   				amdgpu_vm_pte_fragment(params, frag_start, end,
>   						       flags, &frag, &frag_end);
> -				if (frag < shift)
> +				if (frag < shift || (params->unlocked && shift))
>   					break;
>   			}
>   		} while (frag_start < entry_end);