|
On 01/17/2019 10:29 AM, Koenig, Christian wrote:
Am 17.01.19 um 16:22 schrieb Grodzovsky, Andrey:
On 01/17/2019 02:45 AM, Christian König wrote:
Am 16.01.19 um 18:17 schrieb Grodzovsky, Andrey:
On 01/16/2019 11:02 AM, Koenig, Christian wrote:
Am 16.01.19 um 16:45 schrieb Grodzovsky, Andrey:
On 01/16/2019 02:46 AM, Christian König wrote:
Am 15.01.19 um 23:01 schrieb Grodzovsky, Andrey:
On 01/11/2019 05:03 PM, Andrey Grodzovsky wrote:
On 01/11/2019 02:11 PM, Koenig, Christian wrote:
Am 11.01.19 um 16:37 schrieb Grodzovsky, Andrey:
On 01/11/2019 04:42 AM, Koenig, Christian wrote:
Am 10.01.19 um 16:56 schrieb Grodzovsky, Andrey:
[SNIP]
But we will not be adding the cb back in drm_sched_stop
anymore, now we
are only going to add back the cb in drm_sched_startr after
rerunning
those jobs in drm_sched_resubmit_jobs and assign them a new parent
there
anyway.
Yeah, but when we find that we don't need to reset anything anymore
then adding the callbacks again won't be possible any more.
Christian.
I am not sure I understand it, can u point me to example of how this
will happen ? I am attaching my latest patches with waiting only for
the last job's fence here just so we are on same page regarding
the code.
Well the whole idea is to prepare all schedulers, then check once more
if the offending job hasn't completed in the meantime.
If the job completed we need to be able to rollback everything and
continue as if nothing had happened.
Christian.
Oh, but this piece of functionality - skipping HW ASIC reset in case
the
guilty job done is totally missing form this patch series and so needs
to be added. So what you say actually is that for the case were we skip
HW asic reset because the guilty job did complete we also need to skip
resubmitting the jobs in drm_sched_resubmit_jobs and hence preserve the
old parent fence pointer for reuse ? If so I would like to add all this
functionality as a third patch since the first 2 patches are more about
resolving race condition with jobs in flight while doing reset -
what do
you think ?
Yeah, sounds perfectly fine to me.
Christian.
I realized there is another complication now for XGMI hive use case,
we currently skip gpu recover for adev in case another gpu recover for
different adev in same hive is running, under the assumption that we
are going to reset all devices in hive anyway because that should
cover our own dev too. But if we chose to skip now HW asic reset if
our guilty job did finish we will aslo not HW reset any other devices
in the hive even if one of them might actually had a bad job, wanted
to do gpu recover but skipped it because our recover was in progress
in that time.
My general idea on that is to keep a list of guilty jobs per hive,
when you start gpu recover you first add you guilty job to the hive
and trylock hive->reset_lock. Any owner of hive->reset_lock (gpu
recovery in progress) once he finished his recovery and released
hive->reset_lock should go over hive->bad_jobs_list and if at least
one of them is still not signaled (not done) trigger another gpu
recovery and so on. If you do manage to trylock you also go over the
list, clean it and perform recovery. This list probably needs to be
protected with per hive lock.
I also think we can for now at least finish reviewing the first 2
patches and submit them since as I said before they are not dealing
with this topic and fixing existing race conditions. If you are OK
with that I can send for review the last iteration of the first 2
patches where I wait for the last fence in ring mirror list.
Andrey
I implemented HW reset avoidance including XGMI use case according to
the plan i specified. Patch is attached but I can't test it yet due to
XGMI regression in PSP which is supposed to be fixed soon. Please take a
look.
Looks a bit too complicated on first glance. In general we should probably get away from handling a hive in any special way.
Yes, I guess i can do it the same way as the generic handling in amdgpu_device_gpu_recover - there is a list of devices to process which is of size 1 for non xgmi use case or more then 1 for XGMI.
Multiple timeout jobs in a hive are identical to multiple timeout jobs on different engines on a single device.
How about the following handling:
1. Add the timeout job to a list.
2. Try to grab a lock to handle the reset, if that doesn't work because there is already a reset underway return immediately.
3. Stop all schedulers on all affected devices including stopping the timeouts and detaching all callbacks.
4. Double check the list of timed out jobs, if all hw fences of all jobs are completed now we actually don't need to do anything.
What if all the jobs on the timed out list did complete but other job (or jobs) for which we removed the time out timer became hanged ? Wouldn't we miss a required reset in this case and wouldn't even have any indication of their hang ?
If the timeout triggers before we disable it we will have the job on the list of jobs which are hanging.
If we found that we don't reset and re-enable the timeout it will trigger a bit later and we can do the check again.
Thinking about this a bit more we actually don't need to change the handling in any way. The idea with the list is just overkill.
See when the we re-start the scheduler we will also re-arm the timeout, so it will trigger again. It is just a matter of clever timeout re-arming.
Christian.
This implies that hanged job might take in worst case twice (or more if there are recurrent resets) then the TO parameter value to trigger - is this OK ?
Re-arming the timeout should probably have a much reduced value when the job hasn't changed. E.g. something like a few ms.
Now i got thinking about non hanged job in progress (job A) and let's say it's a long job , it just started executing but due to time out of another job (job B) on another (or this scheduler) it's parent cb got disconnected, we disarmed the tdr timer for the
job's scheduler, meanwhile the timed out job did manage to complete before HW reset check and hence we skip HW reset, attach back the cb and rearm job's A tdr timer with a future value of few ms only - aren't we going to get false tdr triggered on job B now
because we didn't let it enough time to run and complete ? I would prefer the other extreme of longer time for time out to trigger then false TDR. Optimally we would have per job timer and rearm to exactly the reminder of it's time out value - but we gave
up on per job tdr work long ago.
In general the more i think about it (correct me if I am wrong) I am less sure how much the optimization feature is useful - if job's time out did trigger what are the chances that the little more time we give it between beginning of tdr function and the time
we do start the actual HW reset will be exactly what it needed to complete. Also, this is still not water proof as the job might complete and signal it's HW fence exactly after we checked for completion but before starting the HW reset code.
Andrey
By unchanged you mean when we didn't resubmit the job because of the optimized non HW reset, right ?
Correct, yes.
About flushing tdr jobs in progress from .free_job cb - looks like drm_sched_job_finish->cancel_delayed_work_sync is not enough, we still need to take care of flushing all sced->work_tdr
for a device and for all devices in hive for XGMI.
What do you think ?
Why should that be necessary? We only wait for the delayed work to make sure that the job is not destroyed while dealing with it.
Christian.
But we might not be waiting for the correct sched->work_tdr, we do the reset routine for all schedulers in a device accessing their jobs too and not only for the scheduler to which the job belongs. For XGMI not only that, we reset all the devices in the hive.
That is harmless you only need to wait for the work_tdr of the current scheduler, not for all of them.
I was thinking, amdgpu driver is not even interested in allowing multiple sced->tdr to execute together - we have to serialize all of them anyway with the trylock mutex (even without XGMI), v3d in v3d_job_timedout seems also to reset all of his schedulers from
the tdr work. Would it make sense to provide the sched->work_td as init parameter to scheduler (same one for all schedulers) so we can enforce serialization by disallowing more then 1 tdr work to execute in the same time ? Other drivers interested to do in
parallel can provide unique sched->work_tdr per scheduler. This does imply drm_sched_job_timedout has to removed and delegated to specific driver implementation as probably other code dealing with sched->work_tdr... Maybe even move tdr handling to the driver
all together ?
Yeah, I was thinking something similar. The problem with this approach is that a delayed work item can have only one delay, but for multiple engines we need multiple delays.
What we could do is to make it a timer instead and raise the work item from the device specific callback.
But that doesn't really saves us the stop all schedulers trouble, so it doesn't buy us much in the end if I see this correctly.
Christian.
Andrey
Andrey
5. Do the reset on all affected devices.
6. Drop the lock.
7. Add callbacks again and restart the schedulers.
I see your steps don't include flushing any tdr in progress from drm_sched_job_finish (or as I did it from amdgpu_job_free_cb) does it mean you don't think we need to flush in order to avoid freeing job while it still might be accessed from time out list ?
Andrey
Regards,
Christian.
Andrey
Andrey
Andrey
_______________________________________________
amd-gfx mailing list
amd-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
amd-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
amd-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
amd-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
|
