Re: [RFC PATCH 4/5] drm/ttm: Document and optimize ttm_bo_pipeline_gutting()

["Thomas Hellström" <thomas.hellstrom@linux.intel.com>]

On 5/21/21 10:21 AM, Christian König wrote:
> Am 20.05.21 um 17:09 schrieb Thomas Hellström:
>> If the bo is idle when calling ttm_bo_pipeline_gutting(), we 
>> unnecessarily
>> create a ghost object and push it out to delayed destroy.
>> Fix this by adding a path for idle, and document the function.
>>
>> Also avoid having the bo end up in a bad state vulnerable to user-space
>> triggered kernel BUGs if the call to ttm_tt_create() fails.
>>
>> Finally reuse ttm_bo_pipeline_gutting() in ttm_bo_evict().
>>
>> Cc: Christian König <christian.koenig@amd.com>
>> Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
>> ---
>>   drivers/gpu/drm/ttm/ttm_bo.c      | 20 +++++-----
>>   drivers/gpu/drm/ttm/ttm_bo_util.c | 63 ++++++++++++++++++++++++-------
>>   drivers/gpu/drm/ttm/ttm_tt.c      |  5 +++
>>   include/drm/ttm/ttm_tt.h          | 10 +++++
>>   4 files changed, 75 insertions(+), 23 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
>> index ca1b098b6a56..a8fa3375b8aa 100644
>> --- a/drivers/gpu/drm/ttm/ttm_bo.c
>> +++ b/drivers/gpu/drm/ttm/ttm_bo.c
>> @@ -501,10 +501,15 @@ static int ttm_bo_evict(struct 
>> ttm_buffer_object *bo,
>>       bdev->funcs->evict_flags(bo, &placement);
>>         if (!placement.num_placement && !placement.num_busy_placement) {
>> -        ttm_bo_wait(bo, false, false);
>> +        ret = ttm_bo_wait(bo, true, false);
>> +        if (ret)
>> +            return ret;
>>   -        ttm_bo_cleanup_memtype_use(bo);
>> -        return ttm_tt_create(bo, false);
>> +        /*
>> +         * Since we've already synced, this frees backing store
>> +         * immediately.
>> +         */
>> +        return ttm_bo_pipeline_gutting(bo);
>
> Yeah, we tried to avoid pipeline_gutting here because of eviction. But 
> I think when you wait before that should work.
>
>>       }
>>         ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx);
>> @@ -974,13 +979,8 @@ int ttm_bo_validate(struct ttm_buffer_object *bo,
>>       /*
>>        * Remove the backing store if no placement is given.
>>        */
>> -    if (!placement->num_placement && !placement->num_busy_placement) {
>> -        ret = ttm_bo_pipeline_gutting(bo);
>> -        if (ret)
>> -            return ret;
>> -
>> -        return ttm_tt_create(bo, false);
>> -    }
>> +    if (!placement->num_placement && !placement->num_busy_placement)
>> +        return ttm_bo_pipeline_gutting(bo);
>>         /*
>>        * Check whether we need to move buffer.
>> diff --git a/drivers/gpu/drm/ttm/ttm_bo_util.c 
>> b/drivers/gpu/drm/ttm/ttm_bo_util.c
>> index 919ee03f7eb3..1860e2e7563f 100644
>> --- a/drivers/gpu/drm/ttm/ttm_bo_util.c
>> +++ b/drivers/gpu/drm/ttm/ttm_bo_util.c
>> @@ -479,7 +479,8 @@ static void ttm_transfered_destroy(struct 
>> ttm_buffer_object *bo)
>>    */
>>     static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo,
>> -                      struct ttm_buffer_object **new_obj)
>> +                      struct ttm_buffer_object **new_obj,
>> +                      bool realloc_tt)
>>   {
>>       struct ttm_transfer_obj *fbo;
>>       int ret;
>> @@ -493,6 +494,17 @@ static int ttm_buffer_object_transfer(struct 
>> ttm_buffer_object *bo,
>>       ttm_bo_get(bo);
>>       fbo->bo = bo;
>>   +    if (realloc_tt) {
>> +        bo->ttm = NULL;
>> +        ret = ttm_tt_create(bo, true);
>> +        if (ret) {
>> +            bo->ttm = fbo->base.ttm;
>> +            kfree(fbo);
>> +            ttm_bo_put(bo);
>> +            return ret;
>> +        }
>> +    }
>> +
>
> Can't we keep that logic in the caller? I think that would be cleaner.

Indeed, let me see if we can do that without breaking anything.

>
>>       /**
>>        * Fix up members that we shouldn't copy directly:
>>        * TODO: Explicit member copy would probably be better here.
>> @@ -763,7 +775,7 @@ static int ttm_bo_move_to_ghost(struct 
>> ttm_buffer_object *bo,
>>       dma_fence_put(bo->moving);
>>       bo->moving = dma_fence_get(fence);
>>   -    ret = ttm_buffer_object_transfer(bo, &ghost_obj);
>> +    ret = ttm_buffer_object_transfer(bo, &ghost_obj, false);
>>       if (ret)
>>           return ret;
>>   @@ -836,26 +848,51 @@ int ttm_bo_move_accel_cleanup(struct 
>> ttm_buffer_object *bo,
>>   }
>>   EXPORT_SYMBOL(ttm_bo_move_accel_cleanup);
>>   +/**
>> + * ttm_bo_pipeline_gutting - purge the contents of a bo
>> + * @bo: The buffer object
>> + *
>> + * Purge the contents of a bo, async if the bo is not idle.
>> + * After a successful call, the bo is left unpopulated in
>> + * system placement. The function may wait uninterruptible
>> + * for idle on OOM.
>> + *
>> + * Return: 0 if successful, negative error code on failure.
>> + */
>>   int ttm_bo_pipeline_gutting(struct ttm_buffer_object *bo)
>>   {
>>       static const struct ttm_place sys_mem = { .mem_type = 
>> TTM_PL_SYSTEM };
>>       struct ttm_buffer_object *ghost;
>>       int ret;
>>   -    ret = ttm_buffer_object_transfer(bo, &ghost);
>> -    if (ret)
>> -        return ret;
>> +    /* If already idle, no need for ghost object dance. */
>> +    ret = ttm_bo_wait(bo, false, true);
>> +    if (ret == -EBUSY) {
>> +        ret = ttm_buffer_object_transfer(bo, &ghost, true);
>> +        if (ret)
>> +            return ret;
>
> When this is a shortcout to avoid work we should rather use the 
> inverse notation.
>
> In other words something like that:
>
> if (ret != -EBUSY) {
>     ttm_resource_free(bo, &bo->mem);
>     ttm_resource_alloc(bo, &sys_mem, &bo->mem);
>     ttm_tt_create()...
>     return ret;
> }

OK.

>
>>   -    ret = dma_resv_copy_fences(&ghost->base._resv, bo->base.resv);
>> -    /* Last resort, wait for the BO to be idle when we are OOM */
>> -    if (ret)
>> -        ttm_bo_wait(bo, false, false);
>> +        ret = dma_resv_copy_fences(&ghost->base._resv, bo->base.resv);
>> +        /* Last resort, wait for the BO to be idle when we are OOM */
>> +        if (ret)
>> +            ttm_bo_wait(bo, false, false);
>>   -    ttm_resource_alloc(bo, &sys_mem, &bo->mem);
>> -    bo->ttm = NULL;
>> +        dma_resv_unlock(&ghost->base._resv);
>> +        ttm_bo_put(ghost);
>> +    } else {
>> +        if (!bo->ttm) {
>> +            ret = ttm_tt_create(bo, true);
>> +            if (ret)
>> +                return ret;
>> +        } else {
>> +            ttm_tt_unpopulate(bo->bdev, bo->ttm);
>> +            if (bo->type == ttm_bo_type_device)
>> +                ttm_tt_mark_for_clear(bo->ttm);
>
> That's not legal, you can't unpopulate it when the BO is busy.
>
> Instead the TT object must be destroyed with the ghost and a new one 
> created.

We've already verified that the bo is idle here, so we should be fine.

/Thomas


>
> Christian.
>
>> +        }
>> +        ttm_resource_free(bo, &bo->mem);
>> +    }
>>   -    dma_resv_unlock(&ghost->base._resv);
>> -    ttm_bo_put(ghost);
>> +    ttm_resource_alloc(bo, &sys_mem, &bo->mem);
>>         return 0;
>>   }
>> diff --git a/drivers/gpu/drm/ttm/ttm_tt.c b/drivers/gpu/drm/ttm/ttm_tt.c
>> index 539e0232cb3b..0b1053e93db2 100644
>> --- a/drivers/gpu/drm/ttm/ttm_tt.c
>> +++ b/drivers/gpu/drm/ttm/ttm_tt.c
>> @@ -134,6 +134,11 @@ void ttm_tt_destroy_common(struct ttm_device 
>> *bdev, struct ttm_tt *ttm)
>>   }
>>   EXPORT_SYMBOL(ttm_tt_destroy_common);
>>   +void ttm_tt_mark_for_clear(struct ttm_tt *ttm)
>> +{
>> +    ttm->page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
>> +}
>> +
>>   void ttm_tt_destroy(struct ttm_device *bdev, struct ttm_tt *ttm)
>>   {
>>       bdev->funcs->ttm_tt_destroy(bdev, ttm);
>> diff --git a/include/drm/ttm/ttm_tt.h b/include/drm/ttm/ttm_tt.h
>> index 134d09ef7766..91552c83ac79 100644
>> --- a/include/drm/ttm/ttm_tt.h
>> +++ b/include/drm/ttm/ttm_tt.h
>> @@ -157,6 +157,16 @@ int ttm_tt_populate(struct ttm_device *bdev, 
>> struct ttm_tt *ttm, struct ttm_oper
>>    */
>>   void ttm_tt_unpopulate(struct ttm_device *bdev, struct ttm_tt *ttm);
>>   +/**
>> + * ttm_tt_mark_for_clear - Mark pages for clearing on populate.
>> + *
>> + * @ttm: Pointer to the ttm_tt structure
>> + *
>> + * Marks pages for clearing so that the next time the page vector is
>> + * populated, the pages will be cleared.
>> + */
>> +void ttm_tt_mark_for_clear(struct ttm_tt *ttm);
>> +
>>   void ttm_tt_mgr_init(unsigned long num_pages, unsigned long 
>> num_dma32_pages);
>>     #if IS_ENABLED(CONFIG_AGP)
>