Op 18-10-12 13:02, Thomas Hellstrom schreef:
On 10/18/2012 10:37 AM, Maarten Lankhorst wrote:
Hey,
Op 18-10-12 09:59, Thomas Hellstrom schreef:
On 10/18/2012 09:28 AM, Thomas Hellstrom wrote:
Hi, Maarten,
As you know I have been having my doubts about this change. To me it seems insane to be forced to read the fence pointer under a reserved lock, simply because when you take the reserve lock, another process may have it and there is a substantial chance that that process will also be waiting for idle while holding the reserve lock.
I think it makes perfect sense, the only times you want to read the fence is when you want to change the members protected by the reservation.
No, that's not true. A typical case (or the only case) is where you want to do a map with no_wait semantics. You will want to be able to access a buffer's results even if the eviction code is about to schedule an unbind from the GPU, and have the buffer reserved?
Well either block on reserve or return -EBUSY if reserved, presumably the former..
ttm_bo_vm_fault does the latter already, anyway.
You don't need to hold the reservation while performing the wait itself though, you could check if ttm_bo_wait(no_wait_gpu = true) returns -EBUSY, and if so take a ref to the sync_obj member and then wait after unreserving. You won't reset sync_obj member to NULL in that case, but that should be harmless. This will allow you to keep the reservations fast and short. Maybe a helper would be appropriate for this since radeon and nouveau both seem to do this.
The next time someone wants to do a wait it will go through the fastpath and unset the sync_obj member, since it's already signaled, or it's removed when ttm_execbuffer_util is used.
~Maarten