Am 01.06.22 um 15:22 schrieb Daniel Vetter:
On Wed, Jun 01, 2022 at 02:45:42PM +0200, Christian König wrote:
Am 31.05.22 um 04:51 schrieb Sergey Senozhatsky:
On (22/05/30 16:55), Christian König wrote:
Am 30.05.22 um 16:22 schrieb Sergey Senozhatsky:
[SNIP]
So the `lock` should have at least same lifespan as the DMA fence
that borrows it, which is impossible to guarantee in our case.
Nope, that's not correct. The lock should have at least same lifespan as the
context of the DMA fence.
How does one know when it's safe to release the context? DMA fence
objects are still transparently refcount-ed and "live their own lives",
how does one synchronize lifespans?
Well, you don't.
If you have a dynamic context structure you need to reference count that as
well. In other words every time you create a fence in your context you need
to increment the reference count and every time a fence is release you
decrement it.
If you have a static context structure like most drivers have then you must
make sure that all fences at least signal before you unload your driver. We
still somewhat have a race when you try to unload a driver and the fence_ops
structure suddenly disappear, but we currently live with that.
Apart from that you are right, fences can live forever and we need to deal
with that.
Yeah this entire thing is a bit an "oops we might have screwed up" moment.
I think the cleanest way is to essentially do what the drm/sched codes
does, which is split the gpu job into the public dma_fence (which can live
forever) and the internal job fence (which has to deal with all the
resource refcounting issues). And then make sure that only ever the public
fence escapes to places where the fence can live forever (dma_resv,
drm_syncobj, sync_file as our uapi container objects are the prominent
cases really).
It sucks a bit.
It's actually not that bad.
See after signaling the dma_fence_ops is mostly used for debugging I
think, e.g. timeline name etc...
Christian.
-Daniel
