buggy/weird behavior in ttm
thellstrom at vmware.com
Thu Oct 11 09:57:28 PDT 2012
On 10/11/2012 04:50 PM, Maarten Lankhorst wrote:
> I was trying to clean ttm up a little so my changes would be less invasive, and simplify
> the code for debuggability. During testing I noticed the following weirdnesses:
> - ttm_mem_evict_first ignores no_wait_gpu if the buffer is on the ddestroy list.
> If you follow the code, it will effectively spin in ttm_mem_evict_first if a bo
> is on the list and no_wait_gpu is true.
Yes, you're right. This is a bug. At a first glance it looks like the
return unconditionally after kref_put().
> I was working on a commit that removes fence_lock since I was killing off the
> fence lock, but that requires some kind of defined behavior for this. Unless
> we leave this in place as expected behavior..
I don't quite follow you? If you mean defined behavior for the fence
lock it is
that it protects the bo::sync_obj and bo::sync_obj_arg of *all* buffers.
prevously one lock per buffer. The locking order is that it should be
the lru lock, but looking at the code it seems it could be quite
simplified the other
Anyway, if you plan to remove the fence lock and protect it with
reserve, you must
make sure that a waiting reserve is never done in a destruction path. I
mostly concerns the nvidia driver.
> - no_wait_reserve is ignored if no_wait_gpu is false
> ttm_bo_reserve_locked can only return true if no_wait_reserve is true, but
> subsequently it will do a wait_unreserved if no_wait_gpu is false.
> I'm planning on removing this argument and act like it is always true, since
> nothing on the lru list should fail to reserve currently.
Yes, since all buffers that are reserved are removed from the LRU list,
should never be a waiting reserve on them, so no_wait_reserve can be removed
from ttm_mem_evict_first, ttm_bo_evict and possibly other functions in
the call chain.
> - effectively unlimited callchain between some functions that all go through
> ttm_mem_evict_first - ttm_bo_evict - -ttm_bo_mem_space - ttm_bo_mem_force_space - ttm_mem_evict_first
> \ ttm_bo_handle_move_mem /
> I'm not surprised that there was a deadlock before, it seems to me it would
> be pretty suicidal to ever do a blocking reserve on any of those lists,
> lockdep would be all over you for this.
Well, at first this may look worse than it actually is. The driver's
eviction memory order determines the recursion depth
and typically it's 0 or 1, since subsequent ttm_mem_evict_first should
never touch the same LRU lists as the first one.
What would typically happen is that a BO is evicted from VRAM to TT, and
if there is no space in TT, another BO is evicted
to system memory, and the chain is terminated. However a driver could
set up any eviction order but that would be
But in essence, as you say, even with a small recursion depth, a waiting
reserve could cause a deadlock.
But there should be no waiting reserves in the eviction path currently.
> Also it seems ttm_bo_move_ttm, ttm_bo_move_memcpy and ttm_bo_move_accel_cleanup
> don't use some of their arguments, so could those be dropped?
Yes, but they are designed so that they could be plugged into the
driver::move interface, so
if you change the argument list you should do it on driver::move as
make sure they all have the same argument list.
More information about the dri-devel