On 26/03/2021 09:10, Daniel Vetter wrote:
On Wed, Mar 24, 2021 at 12:13:28PM +0000, Tvrtko Ursulin wrote:
From: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx>
"Watchdog" aka "restoring hangcheck" aka default request/fence expiry - second
post of a somewhat controversial feature, now upgraded to patch status.
I quote the "watchdog" becuase in classical sense watchdog would allow userspace
to ping it and so remain alive.
I quote "restoring hangcheck" because this series, contrary to the old
hangcheck, is not looking at whether the workload is making any progress from
the kernel side either. (Although disclaimer my memory may be leaky - Daniel
suspects old hangcheck had some stricter, more indiscriminatory, angles to it.
But apart from being prone to both false negatives and false positives I can't
remember that myself.)
Short version - ask is to fail any user submissions after a set time period. In
this RFC that time is twelve seconds.
Time counts from the moment user submission is "runnable" (implicit and explicit
dependencies have been cleared) and keeps counting regardless of the GPU
contetion caused by other users of the system.
So semantics are really a bit weak, but again, I understand this is really
really wanted by the DRM core even if I am not convinced it is a good idea.
There are some dangers with doing this - text borrowed from a patch in the
series:
This can have an effect that workloads which used to work fine will
suddenly start failing. Even workloads comprised of short batches but in
long dependency chains can be terminated.
And becuase of lack of agreement on usefulness and safety of fence error
propagation this partial execution can be invisible to userspace even if
it is "listening" to returned fence status.
Another interaction is with hangcheck where care needs to be taken timeout
is not set lower or close to three times the heartbeat interval. Otherwise
a hang in any application can cause complete termination of all
submissions from unrelated clients. Any users modifying the per engine
heartbeat intervals therefore need to be aware of this potential denial of
service to avoid inadvertently enabling it.
Given all this I am personally not convinced the scheme is a good idea.
Intuitively it feels object importers would be better positioned to
enforce the time they are willing to wait for something to complete.
v2:
* Dropped context param.
* Improved commit messages and Kconfig text.
v3:
* Log timeouts.
* Bump timeout to 20s to see if it helps Tigerlake.
I think 20s is a bit much, and seems like problem is still there in igt. I
think we need look at that and figure out what to do with it. And then go
back down with the timeout somewhat again since 20s is quite a long time.
Irrespective of all the additional gaps/opens around watchdog timeout.
1)
The relationship with the hearbeat is the first issue. There we have 3x
heartbeat period (each rounded to full second) before sending a
high-prio pulse which can cause a preempt timeout and hence a
reset/kicking out of a non-compliant request.
Defaults for those values mean default expiry shouldn't be lower than 3x
rounded hearbeat interval + preempt timeout, currently ~9.75s. In
practice even 12s which I tried initially was too aggressive due slacks
on some platforms.
2)
20s seems to work apart that it shows the general regression
unconditional default expiry adds. Either some existing IGTs which
create long runnable chains, or the far-fence test which explicitly
demonstrates this. AFAIK, and apart from the can_merge_rq yet
unexplained oops, this is the only class of IGT failures which can appear.
So you could tweak it lower, if you also decide to make real hang
detection stricter. But doing that also worsens the regression with
loaded systems.
I only can have a large shrug/dontknow here since I wish we went more
towards my suggestion of emulating setrlimit(RLIMIT_CPU). Meaning at
least going with GPU time instead of elapsed time and possibly even
leaving the policy of setting it to sysadmins. That would fit much
better with our hangcheck, but, doesn't fit the drm core mandate.. hence
I really don't know.
Regards,
Tvrtko
-Daniel
* Fix sentinel assert.
v4:
* A round of review feedback applied.
Chris Wilson (1):
drm/i915: Individual request cancellation
Tvrtko Ursulin (6):
drm/i915: Extract active lookup engine to a helper
drm/i915: Restrict sentinel requests further
drm/i915: Handle async cancellation in sentinel assert
drm/i915: Request watchdog infrastructure
drm/i915: Fail too long user submissions by default
drm/i915: Allow configuring default request expiry via modparam
drivers/gpu/drm/i915/Kconfig.profile | 14 ++
drivers/gpu/drm/i915/gem/i915_gem_context.c | 73 ++++---
.../gpu/drm/i915/gem/i915_gem_context_types.h | 4 +
drivers/gpu/drm/i915/gt/intel_context_param.h | 11 +-
drivers/gpu/drm/i915/gt/intel_context_types.h | 4 +
.../gpu/drm/i915/gt/intel_engine_heartbeat.c | 1 +
.../drm/i915/gt/intel_execlists_submission.c | 23 +-
.../drm/i915/gt/intel_execlists_submission.h | 2 +
drivers/gpu/drm/i915/gt/intel_gt.c | 3 +
drivers/gpu/drm/i915/gt/intel_gt.h | 2 +
drivers/gpu/drm/i915/gt/intel_gt_requests.c | 28 +++
drivers/gpu/drm/i915/gt/intel_gt_types.h | 7 +
drivers/gpu/drm/i915/i915_params.c | 5 +
drivers/gpu/drm/i915/i915_params.h | 1 +
drivers/gpu/drm/i915/i915_request.c | 129 ++++++++++-
drivers/gpu/drm/i915/i915_request.h | 16 +-
drivers/gpu/drm/i915/selftests/i915_request.c | 201 ++++++++++++++++++
17 files changed, 479 insertions(+), 45 deletions(-)
--
2.27.0
_______________________________________________
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
