On 16/01/19 09:42, Antonio Argenziano wrote:
On 16/01/19 08:15, Tvrtko Ursulin wrote:
On 11/01/2019 21:28, John Harrison wrote:
On 1/11/2019 09:31, Antonio Argenziano wrote:
On 11/01/19 00:22, Tvrtko Ursulin wrote:
On 11/01/2019 00:47, Antonio Argenziano wrote:
On 07/01/19 08:58, Tvrtko Ursulin wrote:
On 07/01/2019 13:57, Chris Wilson wrote:
Quoting Tvrtko Ursulin (2019-01-07 13:43:29)
On 07/01/2019 11:58, Tvrtko Ursulin wrote:
[snip]
Note about future interaction with preemption: Preemption
could happen
in a command sequence prior to watchdog counter getting
disabled,
resulting in watchdog being triggered following preemption
(e.g. when
watchdog had been enabled in the low priority batch). The
driver will
need to explicitly disable the watchdog counter as part of the
preemption sequence.
Does the series take care of preemption?
I did not find that it does.
Oh. I hoped that the watchdog was saved as part of the
context... Then
despite preemption, the timeout would resume from where we left
off as
soon as it was back on the gpu.
If the timeout remaining was context saved it would be much
simpler (at
least on first glance), please say it is.
I made my comments going only by the text from the commit message
and the absence of any preemption special handling.
Having read the spec, the situation seems like this:
* Watchdog control and threshold register are context saved and
restored.
* On a context switch watchdog counter is reset to zero and
automatically disabled until enabled by a context restore or
explicitly.
So it sounds the commit message could be wrong that special
handling is needed from this direction. But read till the end on
the restriction listed.
* Watchdog counter is reset to zero and is not accumulated
across multiple submission of the same context (due preemption).
I read this as - after preemption contexts gets a new full
timeout allocation. Or in other words, if a context is preempted
N times, it's cumulative watchdog timeout will be N * set value.
This could be theoretically exploitable to bypass the timeout. If
a client sets up two contexts with prio -1 and -2, and keeps
submitting periodical no-op batches against prio -1 context,
while prio -2 is it's own hog, then prio -2 context defeats the
watchdog timer. I think.. would appreciate is someone challenged
this conclusion.
I think you are right that is a possibility but, is that a
problem? The client can just not set the threshold to bypass the
timeout. Also because you need the hanging batch to be simply
preemptible, you cannot disrupt any work from another client that
is higher priority. This is
But I think higher priority client can have the same effect on the
lower priority purely by accident, no?
As a real world example, user kicks off an background transcoding
job, which happens to use prio -2, and uses the watchdog timer.
At the same time user watches a video from a player of normal
priority. This causes periodic, say 24Hz, preemption events, due
frame decoding activity on the same engine as the transcoding client.
Does this defeat the watchdog timer for the former is the question?
Then the questions of can we do something about it and whether it
really isn't a problem?
I guess it depends if you consider that timeout as the maximum
lifespan a workload can have or max contiguous active time.
I believe the intended purpose of the watchdog is to prevent broken
bitstreams hanging the transcoder/player. That is, it is a form of
error detection used by the media driver to handle bad user input. So
if there is a way for the watchdog to be extended indefinitely under
normal situations, that would be a problem. It means the transcoder
will not detect the broken input data in a timely manner and
effectively hang rather than skip over to the next packet. And note
that broken input data can be caused by something as innocent as a
dropped packet due to high network contention. No need for any
malicious activity at all.
My understanding of the intended purpose is the same. And it would be
a very useful feature.
I'm not familiar enough with the application but, in the scenario above,
what if the batch that is being preempted is not stuck but just nice
enough to be preempted enough times so that it wouldn't complete in the
given wall clock time but would be fast enough by itself.
Ignore me, re-reading this I now get you are trying to advocate for an
active-time timeout not pure wall clock time.
Chris mentioned the other day that until hardware is fixed to context
save/restore the watchdog counter this could simply be implemented
using timers. And I have to say I agree. Shouldn't be too hard to
prototype it using hrtimers - start on context in, stop on context
out and kick forward on user interrupts. More or less.
Would this implement the feature on the driver side just like it would
for the HW? I mean have the same IOCTL and silently discard workload
that hit the timeout. Also, would it discard batches while they are in
the queue (not active)?
Antonio
Then if the cost of these hrtimer manipulations wouldn't show in
profiles significantly we would have a solution. At least in execlists
mode. :) But in parallel we could file a feature request to fix the
hardware implementation and then could just switch the timer "backend"
from hrtimers to GPU.
Regards,
Tvrtko
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx