As soon as we leave the spinlock after the job has been added to the job
queue, we can no longer rely on the job's data to be available.
I have seen a null-pointer dereference due to sched == NULL in
amd_sched_wakeup via amd_sched_entity_push_job and
amd_sched_ib_submit_kernel_helper. Since the latter initializes
sched_job->sched with the address of the ring scheduler, which is
guaranteed to be non-NULL, this race appears to be a likely culprit.
Signed-off-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Bugzilla: https://bugs.freedesktop.org/attachment.cgi?bugid=93079
Reviewed-by: Christian König <christian.koenig@amd.com>
This way the driver isn't limited in the dependency handling callback.
v2: remove extra check in amd_sched_entity_pop_job()
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
We only need to wait for jobs to be scheduled when
the dependency is from the same scheduler.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
Before this patch the scheduler fence was created when we push the job
into the queue, so we could only get the fence after pushing it.
The mutex now was necessary to prevent the thread pushing the jobs to
the hardware from running faster than the thread pushing the jobs into
the queue.
Otherwise the thread pushing jobs into the queue would have accessed
possible freed up memory when it tries to get a reference to the fence.
So what you get in the end is thread A:
mutex_lock(&job->lock);
...
Kick of thread B.
...
mutex_unlock(&job->lock);
And thread B:
mutex_lock(&job->lock);
....
mutex_unlock(&job->lock);
kfree(job);
I'm actually not sure if I'm still up to date on this, but this usage
pattern used to be not allowed with mutexes. See here as well
https://lwn.net/Articles/575460/.
v2: remove unrelated changes, fix missing owner
v3: rebased, add more commit message
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
The code was correct, but getting two references when the ownership
is linearly moved on is a bit awkward and just overhead.
Signed: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
In two places amdgpu tries to tear down something it hasn't
initalised when failing. This is what happens when you
enable experimental support on topaz which then fails in
ring init.
This patch allows it to fail cleanly.
agd: Split out from from the original patch since the
scheduler is a driver independent.
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org
The first lockup fence will lock the fence list of scheduler.
Then cancel the delayed workqueues for all clean pending fences
without waiting the workqueues to finish.
Change-Id: I9bec826de1aa49d587b0662f3fb4a95333979429
Signed-off-by: Junwei Zhang <Jerry.Zhang@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Embed the scheduler into the ring structure instead of allocating it.
Use the ring name directly instead of the id.
v2: rebased, whitespace cleanup
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Junwei Zhang <Jerry.Zhang@amd.com>
Reviewed-by: Chunming Zhou<david1.zhou@amd.com>
track sched job status like the length of job queue and hw job queue.
v2: fix build after rebase
Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Removing the entity from scheduling can deadlock the whole system.
Wait forever till the remaining IBs are scheduled.
v2: fix comment as well
Signed-off-by: Christian König <christian.koenig@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com> (v1)
Cleanup function name, stop checking scheduler ready twice, but
check if kernel thread should stop instead.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
Remove active_hw_rq and it's protecting queue_lock, they are unused.
User 32bit atomic for hw_rq_count, 64bits for counting to three is a bit
overkill.
Cleanup the function name and remove incorrect comments.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
Rework run queue implementation, especially remove the odd list handling.
v2: cleanup the code only, no algorithem change.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
fix fence is released when pass to **fence sometimes.
add reference for it.
Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>
