On 01/09/2016 11:46 AM, Christian König wrote:
It's correct that exported buffers can't be moved to another domain or swapped to disk while another device is using it.
But for imports that's a different story:
an imported object should never end up on a LRU list in TTM because TTM wouldn't know how to evict it.
Even currently the imported objects are actually added to the LRU. The problem is just that they are not added immediately during creation, but only with the first command submission.
Hmm. The fact that they are added to LRU at all sounds weird. What happens when TTM tries to evict or even swap out an imported buffer?
/Thomas
Regards, Christian.
Am 09.01.2016 um 08:05 schrieb Thomas Hellstrom:
Hi!
I might be misunderderstanding the use-case here, but IIRC the discussion with TTM vs imported / exported buffer objects is that a buffer object needs to be marked NO_EVICT in TTM before it's exported and an imported object should never end up on a LRU list in TTM because TTM wouldn't know how to evict it.
/Thomas
On 01/08/2016 02:41 PM, Christian König wrote:
From: Christian König christian.koenig@amd.com
If we import a BO with an eternal reservation object we don't reserve/unreserve it. So we never add it to the LRU causing a possible deny of service.
Signed-off-by: Christian König christian.koenig@amd.com
drivers/gpu/drm/ttm/ttm_bo.c | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c index 745e996..a98a5d5 100644 --- a/drivers/gpu/drm/ttm/ttm_bo.c +++ b/drivers/gpu/drm/ttm/ttm_bo.c @@ -1170,9 +1170,15 @@ int ttm_bo_init(struct ttm_bo_device *bdev, if (likely(!ret)) ret = ttm_bo_validate(bo, placement, interruptible, false);
- if (!resv)
- if (!resv) { ttm_bo_unreserve(bo);
- } else if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
spin_lock(&bo->glob->lru_lock);ttm_bo_add_to_lru(bo);spin_unlock(&bo->glob->lru_lock);- }
if (unlikely(ret)) ttm_bo_unref(&bo);