On 11/16/21 09:20, Christian König wrote:
Am 16.11.21 um 08:43 schrieb Thomas Hellström:
On 11/16/21 08:19, Christian König wrote:
Am 13.11.21 um 12:26 schrieb Thomas Hellström:
Hi, Zack,
On 11/11/21 17:44, Zack Rusin wrote:
On Wed, 2021-11-10 at 09:50 -0500, Zack Rusin wrote:
TTM takes full control over TTM_PL_SYSTEM placed buffers. This makes
driver internal usage of TTM_PL_SYSTEM prone to errors because it
requires the drivers to manually handle all interactions between TTM
which can swap out those buffers whenever it thinks it's the right
thing to do and driver.
CPU buffers which need to be fenced and shared with accelerators
should
be placed in driver specific placements that can explicitly handle
CPU/accelerator buffer fencing.
Currently, apart, from things silently failing nothing is enforcing
that requirement which means that it's easy for drivers and new
developers to get this wrong. To avoid the confusion we can document
this requirement and clarify the solution.
This came up during a discussion on dri-devel:
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.kernel.org%2Fdri-devel%2F232f45e9-8748-1243-09bf-56763e6668b3%40amd.com&data=04%7C01%7Cchristian.koenig%40amd.com%7C55e15a3b151b401993ca08d9a8d4c878%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637726454113422983%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=HSg2rZf1yFsCCOUOcoG5Y0ogGE%2FsUymh3UqJYvZ1%2BDM%3D&reserved=0
I took a slightly deeper look into this. I think we need to
formalize this a bit more to understand pros and cons and what the
restrictions are really all about. Anybody looking at the prevous
discussion will mostly see arguments similar to "this is stupid and
difficult" and "it has always been this way" which are not really
constructive.
First disregarding all accounting stuff, I think this all boils
down to TTM_PL_SYSTEM having three distinct states:
1) POPULATED
2) LIMBO (Or whatever we want to call it. No pages present)
3) SWAPPED.
The ttm_bo_move_memcpy() helper understands these, and any
standalone driver implementation of the move() callback _currently_
needs to understand these as well, unless using the
ttm_bo_move_memcpy() helper.
Now using a bounce domain to proxy SYSTEM means that the driver can
forget about the SWAPPED state, it's automatically handled by the
move setup code. However, another pitfall is LIMBO, in that if when
you move from SYSTEM/LIMBO to your bounce domain, the BO will be
populated. So any naive accelerated move() implementation creating
a 1GB BO in fixed memory, like VRAM, will needlessly allocate and
free 1GB of system memory in the process instead of just performing
a clear operation. Looks like amdgpu suffers from this?
I think what is really needed is either
a) A TTM helper that helps move callback implementations resolve
the issues populating system from LIMBO or SWAP, and then also
formalize driver notification for swapping. At a minimum, I think
the swap_notify() callback needs to be able to return a late error.
b) Make LIMBO and SWAPPED distinct memory regions. (I think I'd
vote for this without looking into it in detail).
In both these cases, we should really make SYSTEM bindable by GPU,
otherwise we'd just be trading one pitfall for another related
without really resolving the root problem.
As for fencing not being supported by SYSTEM, I'm not sure why we
don't want this, because it would for example prohibit async
ttm_move_memcpy(), and also, async unbinding of ttm_tt memory like
MOB on vmgfx. (I think it's still sync).
There might be an accounting issue related to this as well, but I
guess Christian would need to chime in on this. If so, I think it
needs to be well understood and documented (in TTM, not in AMD
drivers).
I think the problem goes deeper than what has been mentioned here so
far.
Having fences attached to BOs in the system domain is probably ok,
but the key point is that the BOs in the system domain are under
TTMs control and should not be touched by the driver.
What we have now is that TTMs internals like the allocation state of
BOs in system memory (the populated, limbo, swapped you mentioned
above) is leaking into the drivers and I think exactly that is the
part which doesn't work reliable here. You can of course can get
that working, but that requires knowledge of the internal state
which in my eyes was always illegal.
Well, I tend to agree to some extent, but then, like said above even
disregarding swap will cause trouble with the limbo state, because
the driver's move callback would need knowledge of that to implement
moves limbo -> vram efficiently.
Well my long term plan is to audit the code base once more and remove
the limbo state from the SYSTEM domain.
E.g. instead of a SYSTEM BO without pages you allocate a BO without a
resource in general which is now possible since bo->resource is a
pointer.
This would still allow us to allocate "empty shell" BOs. But a
validation of those BOs doesn't cause a move, but rather just
allocates the resource for the first time.
The problem so far was just that we access bo->resource way to often
without checking it.
So the driver would then at least need to be aware of these empty shell
bos without resource for their move callbacks? (Again thinking of the
move from empty shell -> VRAM).
Thanks,
/Thomas