[PATCH] drm/ttm: Set CPU caching mode to cached for BOs being swapped out

[Michel Dänzer]

On 06.11.2015 22:02, Thomas Hellstrom wrote:
> On 11/06/2015 03:33 AM, Michel Dänzer wrote:
>> On 05.11.2015 17:47, Thomas Hellstrom wrote:
>>> Hi, Michel,
>>>
>>> On 11/05/2015 09:08 AM, Michel Dänzer wrote:
>>>> From: Michel Dänzer <michel.daenzer at amd.com>
>>>>
>>>> I ran into the BUG_ON in ttm_tt_swapout, presumably the BO being swapped
>>>> out was using a write-combined CPU mapping.
>>>>
>>>> Instead of BUGging out, just set the caching mode to what's needed.
>>>>
>>>> Signed-off-by: Michel Dänzer <michel.daenzer at amd.com>
>>>> ---
>>>>  drivers/gpu/drm/ttm/ttm_tt.c | 3 ++-
>>>>  1 file changed, 2 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/drivers/gpu/drm/ttm/ttm_tt.c b/drivers/gpu/drm/ttm/ttm_tt.c
>>>> index 4e19d0f..c2794eb 100644
>>>> --- a/drivers/gpu/drm/ttm/ttm_tt.c
>>>> +++ b/drivers/gpu/drm/ttm/ttm_tt.c
>>>> @@ -334,7 +334,8 @@ int ttm_tt_swapout(struct ttm_tt *ttm, struct file *persistent_swap_storage)
>>>>  	int ret = -ENOMEM;
>>>>  
>>>>  	BUG_ON(ttm->state != tt_unbound && ttm->state != tt_unpopulated);
>>>> -	BUG_ON(ttm->caching_state != tt_cached);
>>>> +
>>>> +	ttm_tt_set_caching(ttm, tt_cached);
>>>>  
>>>>  	if (!persistent_swap_storage) {
>>>>  		swap_storage = shmem_file_setup("ttm swap",
>>> This *is* actually a bug somewhere, since before ttm_tt_swapout,
>>> ttm_bo_swapout should have moved out the bo to system and set
>>> the correct caching.
>> Maybe ttm_bo_swapout needs to check ttm->caching_state explicitly?
>> AFAICT it only checks the placement flags, but we allow all caching
>> modes for GTT.
>>
> 
> bo->mem.placement should only have two flags set, one for the current
> memory type corresponding to bo->mem.mem_type, the other for the current
> caching. If there are more than one caching flag set at this point,
> that's a bug. The preferred caching mode is chosen from a set of modes
> in ttm_bo_select_caching(), and bo->mem is reassigned to incorporate the
> current caching mode after a successful move.
> 
> I quickly eyed through the TTM default move functions and they seem to
> do the correct thing from what I can tell. Is the radeon ttm driver
> assigning bo->mem in a move callback perhaps?

Only via ttm_bo_move_ttm / to the values passed into the move callback
AFAICT.


-- 
Earthling Michel Dänzer               |               http://www.amd.com
Libre software enthusiast             |             Mesa and X developer