Op 19-11-12 14:33, Thomas Hellstrom schreef: > On 11/12/2012 03:00 PM, Maarten Lankhorst wrote: >> Just use the return error from ttm_mem_evict_first instead. > > Here driver need to be able to evict a memory type completely, because they > might shut down > the memory type or clear it for some legacy usage, suspending or whatever, so > returning 0 on -EBUSY isn't sufficient, > we need at least a list empty check, and a shared reservation at this point > is illegal. > > This is a point where the mechanism to exclude other reservers is needed, as > we discussed previously. > vmwgfx is using the ttm lock, but will of course adapt if a new mechanism is > emerging. Normally ttm_mem_evict_first only returns -EBUSY if the list is empty and no_wait = false, so I thought using the return code would be equivalent.
We could do spin_lock(&glob->lru_lock); WARN_ON(!list_empty(&man->lru_lock)); spin_unlock(&glob->lru_lock); to handle this after -EBUSY. With a lot of objects on the lru list, this would save taking lru_lock twice for each object. ~Maarten
