On Mon, Aug 14, 2017 at 8:08 AM, Jason Ekstrand <jason@xxxxxxxxxxxxxx> wrote:
On Mon, Aug 14, 2017 at 12:36 AM, Christian König <christian.koenig@xxxxxxx> wrote:Am 14.08.2017 um 01:14 schrieb Jason Ekstrand:
On August 13, 2017 8:52:21 AM Christian König <christian.koenig@xxxxxxx> wrote:
Am 13.08.2017 um 17:26 schrieb Jason Ekstrand:
On August 13, 2017 6:19:53 AM Christian König
<christian.koenig@xxxxxxx> wrote:
Patches #1-#4 are Acked-by: Christian König <christian.koenig@xxxxxxx>.
Patch #5: NAK, that will break radeon.
On radeon we need the non-default wait or otherwise we can run into a
situation where we never signal a fence.
The general question is why do you need this?
Because i915 sets a non-default wait function so calling wait_any just
bails with fences from i915 immediately bails with -EINVAL. This makes
it work even with non-default waits.
Ok well, let me refine the question: Why does i915 sets a non-default
wait function?
I have no idea.
Can you figure that out? I'm not completely against removing that limitation, but it would be a lot cleaner if we can just fix i915 to not set a non-default wait function.I asked on IRC how bad it would be and chris replied with "a lot of work". He didn't say it's impossible but, apparently, it is a pile of work.In radeon we have it because we need to handle 10+ years of different
hardware generation, each which a bunch of separate bugs in their fence
handling (and some even not solved by today).
So how does wait_any returning -EINVAL for non-default waits help radeon?
When the wait function detects a problem it reports -EDEADLK to the caller to signal that the hardware is stuck.
The Caller then goes up the call chain and ultimately reports this to the IOCTL where the error is handled with a hardware reset. And yeah, I perfectly know that this design sucks badly.Could we use dma_fence::err for this? i.e., could the radeon driver set the error bit and then have wait_any check for errors on wakeup and report the first one it sees?The point is we should try to prevent that the wait for any function is even used together with a radeon fence.Does radeon not support sync_file? Because this is the same mechanism it uses for poll.
Patch #6: Yes, please. Patch is Reviewed-by: Christian König
<christian.koenig@xxxxxxx>.
Patch #7: Already gave my rb on the patch Chris send out earlier.
Patch #8: NAK to the whole approach.
IIRC we discussed a very similar thing during the initial fence bringup
and also during the fence_array development.
The problem is that you can easily build ring dependencies and so
deadlocks with it.
For what it's worth, my proxy-based implementation does not suffer from the same issues as the one suggested by Chris. I was very careful to ensure that drm_syncobj_fence_get will never return a proxy - if the current fence is a proxy, it returns NULL. It should look no different to drivers than any of the other interfaces I've proposed from that perspective.
Yeah, more or less. What I'm suggesting is to use one wait_queue_head_t per drm_syncobj.Yes, that would be a start.I would really prefer an approach which is completely contained inside
the syncobj code base.
Are you use to the approach of internally making a proxy so long as
all the proxy code is inside syncobj?
In general if possible I would rather like to avoid the whole handling
with the proxy/callback altogether, but that possible only works with
wait for any if the waitqueue is global and that wouldn't be ideal either.
I'm happy to get rid of the proxies. They did work nicely but I'm not really attached to them
Is also be happy to go back to the original approach with v3 of thev3 looked like it should work as well, I would just drop abusing the
last patch.
fence callback structure for the signaling.
Ideally we would finally come up with an interface to wait for multiple
waitqueue at the same time, but that probably goes a bit to far.
For now just use a single linked list to start all processes waiting for
a fence to arrive or something like this.
I don't really know what you're suggesting. Patch v3 has a single waitqueue per process. Are you suggesting one per fence?
See a wait_queue is a callback mechanism anyway, so you are wrapping a callback mechanism inside another callback mechanism and that makes not really much sense.Fair enough. There is one little snag though: We need to wait on sync objects and fences at the same time in order for WAIT_ANY | WAIT_FOR_SUBMIT to work. I see two options here:1) Convert dma-fence to use waitqueue instead of its callback mechanism and add a wait_queue_any. A quick grep for dma_fence_add_callback says that this would affect four drivers.
The more I think about it, the less sense using waitqueues makes. The fundamental problem here is that the event we are waiting on is actually the concatenation of two events: submit and signal. Since we are waiting on several of these pairs of concatenated events simultaneously, the only two options we have are to either combine them into one event (the proxy approach) or to implement a wait which is capable of handling both at the same time. I don't see a way to do the latter with wait queues.
--Jason
2) Drop waitqueues and go back and fix patch v2 so that it does the wait correctly.--JasonThe problem is that we don't have a wait_event_* variant which can wait for multiple events. Somebody should really add something like that :)
Regards,
Christian.
--Jason
Regards,
Christian.
Regards,
Christian.
Am 12.08.2017 um 00:39 schrieb Jason Ekstrand:
This series does the same thing as my earlier series in that it adds
a sync
object wait interface complete with WAIT_FOR_SUBMIT flag. While the
uapi
remains unchanged, the guts look a bit different. Instead of adding a
callback mechanism to drm_syncobj that fired whenever replace_fence was
called, it's now using proxy fences. The drm_syncobj_fence_get still
returns NULL whenever the sync object is in an unsubmitted state but
there
is a new drm_syncobj_fence_proxy_get which returns either the real
fence or
a proxy fence that will be triggered the next time replace_fence is
called
with a non-NULL replacement. This does make both
drm_syncobj_fence_get and
drm_syncobj_replace_fence a tiny bit more expensive, but it lets us
do it
all without locking.
This series can be found as a branch here:
https://cgit.freedesktop.org/~jekstrand/linux/log/?h=drm-syn cobj-wait-submit-v4
IGT tests for DRM_IOCTL_SYNCOBJ_WAIT and DRM_IOCTL_SYNCOBJ_RESET can be
found on patchwork here:
https://patchwork.freedesktop.org/series/28666/
Patches to the Intel Vulkan driver to implement
VK_KHR_external_fence on
top of this kernel interface can be found here:
https://cgit.freedesktop.org/~jekstrand/mesa/log/?h=wip/anv- external-fence
Cc: Dave Airlie <airlied@xxxxxxxxxx>
Cc: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
Cc: Christian König <christian.koenig@xxxxxxx>
Chris Wilson (2):
dma-buf/dma-fence: Signal all callbacks from dma_fence_release()
dma-buf/dma-fence: Add a mechanism for proxy fences
Dave Airlie (1):
drm/syncobj: add sync obj wait interface. (v8)
Jason Ekstrand (6):
drm/syncobj: Rename fence_get to find_fence
drm/syncobj: Add a race-free drm_syncobj_fence_get helper
i915: Add support for drm syncobjs
dma-buf/dma-fence: Allow wait_any_timeout without default_wait (v2)
drm/syncobj: Add a reset ioctl
drm/syncobj: Allow wait for submit and signal behavior (v4)
drivers/dma-buf/Makefile | 4 +-
drivers/dma-buf/dma-fence-proxy.c | 186 +++++++++++++++++++
drivers/dma-buf/dma-fence.c | 34 ++--
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 2 +-
drivers/gpu/drm/drm_internal.h| 4 +
drivers/gpu/drm/drm_ioctl.c | 4 +
drivers/gpu/drm/drm_syncobj.c | 275
+++++++++++++++++++++++++++--
drivers/gpu/drm/i915/i915_drv.c | 3 +-
drivers/gpu/drm/i915/i915_gem_execbuffer.c | 146 ++++++++++++++-
include/drm/drm_syncobj.h | 15 +-
include/linux/dma-fence-proxy.h | 25 +++
include/uapi/drm/drm.h | 19 ++
include/uapi/drm/i915_drm.h | 30 +++-
13 files changed, 710 insertions(+), 37 deletions(-)
create mode 100644 drivers/dma-buf/dma-fence-proxy.c
create mode 100644 include/linux/dma-fence-proxy.h
_______________________________________________ dri-devel mailing list dri-devel@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/dri-devel
