[PATCH 7/7] drm/syncobj: use the timeline point in drm_syncobj_find_fence

Chunming Zhou zhoucm1 at amd.com
Fri Nov 23 13:42:11 UTC 2018



在 2018/11/23 21:27, Koenig, Christian 写道:

Am 23.11.18 um 14:15 schrieb Zhou, David(ChunMing):



在 2018/11/23 20:02, Koenig, Christian 写道:


Am 23.11.18 um 12:03 schrieb Christian König:


Am 23.11.18 um 11:56 schrieb zhoucm1:


On 2018年11月23日 18:10, Koenig, Christian wrote:


Am 23.11.18 um 03:36 schrieb zhoucm1:


On 2018年11月22日 19:30, Christian König wrote:


Am 22.11.18 um 07:52 schrieb zhoucm1:


On 2018年11月15日 19:12, Christian König wrote:


Implement finding the right timeline point in
drm_syncobj_find_fence.

Signed-off-by: Christian König <christian.koenig at amd.com><mailto:christian.koenig at amd.com>
---
     drivers/gpu/drm/drm_syncobj.c | 10 +++++++++-
     1 file changed, 9 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/drm_syncobj.c
b/drivers/gpu/drm/drm_syncobj.c
index 589d884ccd58..d42c51520da4 100644
--- a/drivers/gpu/drm/drm_syncobj.c
+++ b/drivers/gpu/drm/drm_syncobj.c
@@ -307,9 +307,17 @@ int drm_syncobj_find_fence(struct drm_file
*file_private,
             return -ENOENT;
           *fence = drm_syncobj_fence_get(syncobj);
-    if (!*fence) {
+    if (!*fence)
             ret = -EINVAL;
+
+    if (!ret && point) {
+        dma_fence_chain_for_each(*fence) {
+            if (!to_dma_fence_chain(*fence) ||
+                (*fence)->seqno <= point)
+                break;


This condition isn't enough to find proper point.
For two examples:
a. No garbage collection happens, the points in chain are
1----3----6----9----12---18---20, if user wants to get point17, then
we should return node 18.


And that is exactly what's wrong in the original logic. In this case
we need to return 12, not 18 because point 17 could have already been
garbage collected.


I don't think so, the 'a' case I already assume there isn't garbage
collection. If user wants to get point17, then we should return
node 18.
timeline means point[N]  must be signaled later than point[N-1].
Point[12] just can make sure point[1] ~point[12] are signaled.
Point[18] signal can make sure point[17] is signaled.
So this case we need to return 18, not 12, which is key timeline
concept.


No, exactly that's incorrect. When we ask for 17 and can't find it then
this means it either never existed or that it is signaled already.

Returning a lower number in this case or even a stub fence is perfectly
fine since we only need to wait for that one in this case.

If we return 18 in this case then we add incorrect synchronization when
there shouldn't be any.


No, That will make timeline not work at all and break timeline
semantics totally.

If there aren't point18 and point20, the chain is
1----3----6----9----12, if user wants to get point 17, you also
return 12? if yes, which absolutely is incorrect. The answer should
be NO, right? point17 should be waited on there until a bigger point
is coming.


Correct, but this is a different case. In this situation we either
return an error or wait for point 17 (or something >=17) to show up.

The key difference is if point 17 shows up then we return point 17,
but if point 18 shows up then we need to return point 12.



For chain is 1----3----6----9----12---18---20, if user wants to wait
on any one of points 13,14,15,16,17,18, we must wait for point 18,
this is timeline semantic.


Ah, now I understand. You are still sticking with the assumption of a
fence number, right?

In other words what you imply here is that we have the same semantic
as when somebody waits for a memory location to be written by number
17, right? In this case the semantics you describe here indeed applies.

But that is certainly not what we want to implement or otherwise we
will never be able to garbage collect the numbers in between.

So if Vulkan has this requirement then we need to reject that.


Backing of and reconsidering this I came to the conclusion that what you
suggest here is actually the most defensive solution.

In other words it is the solution where it's most likely that nothing
goes wrong because the worst thing that can happen is that we
synchronize to much, but never to less.

Going to think about it how we can bring that into alignment with the
proposed garbage collection.


Yeah, for garbae collection, I came up an idea this morning, we can pass
the signaled stub fence when chain node is created, when you walk out
all chain node, you can replace chain->fence with stub fence, that way,
there is no redudant fence referenced in chain node, and we can keep the
signal point in there.



That would still not allow to garbage collect the chain node itself.

only middle chain nodes are need, so the chain shouldn't be too long, right?




But I've came up with something which should work. Assume the original
chain is:

1----3----6----9----12---18

And we garbage collect everything but 6 and 18 then all we need to know
to return the correct node is what the original previous sequence number
was.

6 (3)----18 (12)

When then somebody asks for 17 we can still return 18 and if somebody
asks for 9 we would return 6.

then what point we return when somebody asks for 11?






Another use case, I'm not sure if you considered:
if chain is  1----3----6----9----12---18, a wait operation is on point
17, then we return 18, another signal point 17 comes, then we still wait
on 18(assume 18 takes very long time), that looks not reseonable, but
this is just performance problem potientially. Seems the way of timeline
sw_sync.c with comparing point for signal status can sovle it.



Well I thought that we declared that signaling lower numbers is illegal?

Sorry, I forgot it, quote from spec: "
*RESOLVED*: A 64-bit unsigned integer that can only be set to monotonically
increasing values by signal operations and is not changed by wait operations."

Can we think signaling lower numbers is forbidden?

If that's true, we can directly ignore lower number and return without error, keep the larger signal point.

Thanks,
David



My current solution to that is when userspace messes up the sequence
numbers and submit 1-3-6-9-12-18-17 we start a new chain with 17 and
never look back.

E.g. when somebody then asks for anything below 17 we always return 17
and if somebody asks for 18 we return an error because that is handled
as not signaled yet.

Regards,
Christian.




-David



Thanks,
Christian.



Regards,
Christian.



You can also check sw_sync.c for timeline meaning.

-David


Christian.



-David


b. garbage collection happens on point6, chain would be updated to
1---3---9---12---18---20, if user wants to get point5, then we
should return node 3, but if user wants to get point 7, then we
should return node 9.


Why? That doesn't seem to make any sense to me.



I still have no idea how to satisfy all these requirements with your
current chain-fence. All these logic just are same we encountered
before, we're walking them again. After solving these problems, I
guess all design is similar as before.

In fact, I don't know what problem previous design has, maybe there
are some bugs, can't we fix these bugs by time going? Who can make
sure his implementation never have bugs?


Well there where numerous problems with the original design. For
example we need to reject the requirement that timeline fences are in
order because that doesn't make sense in the kernel.

When userspace does something like submitting fences in the order 1,
5, 3 then it is broken and can keep the pieces. In other words the
kernel should not care about that, but rather make sure that it never
looses any synchronization no matter what.

Regards,
Christian.



-David


+        }
         }
+
         drm_syncobj_put(syncobj);
         return ret;
     }


_______________________________________________
dri-devel mailing list
dri-devel at lists.freedesktop.org<mailto:dri-devel at lists.freedesktop.org>
https://lists.freedesktop.org/mailman/listinfo/dri-devel





-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/dri-devel/attachments/20181123/5c2d12a9/attachment-0001.html>


More information about the dri-devel mailing list