nv-tegra-mirror/linux-nvgpu
1
0
Fork 0
mirror of git://nv-tegra.nvidia.com/linux-nvgpu.git synced 2025-12-22 17:36:20 +03:00
Code Issues Packages Projects Releases Wiki Activity

gpu: nvgpu: skip post_process() operations for single domain.

Browse Source 
There is a possible deadlock that gets triggered when
device is being resumed() and NVS worker thread tries to submit
the data as part of the post_process() operation.

The NVS worker thread works asynchronously in the post_process() part
w.r.t the USER threads and thus an initial implementation requires
acquiring the busy lock() arriving at a deadlock scenario.

This quick change shall disallow post_process() from executing
during the case where we have only one scheduling domain present(legacy)
Any submits meant to be updated are handled via the synchronous
wakeup_process_item() callback.

This implementation is being modified to allow the worker thread
to be suspended/resumed during GPU railgate/unrailgate in upcoming
releases and currently is in a state of flux.

Bug 3723127

Signed-off-by: Debarshi Dutta <ddutta@nvidia.com>
Change-Id: I318cda0fbdd5651884cf21f748c86687679e6fdb
Reviewed-on: https://git-master.nvidia.com/r/c/linux-nvgpu/+/2750293
Reviewed-by: Prateek Sethi <prsethi@nvidia.com>
Reviewed-by: Vijayakumar Subbu <vsubbu@nvidia.com>
GVS: Gerrit_Virtual_Submit
This commit is contained in:
Debarshi Dutta
2022-07-25 11:37:10 +05:30
committed by Richard Zhao
parent ee5053f7be
commit 44b6bfbc1d
1 changed files with 19 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
drivers/gpu/nvgpu/common/nvs/nvs_sched.c
View File

@@ -117,6 +117,25 @@ static u64 nvgpu_nvs_tick(struct gk20a *g)
nvs_next = sched->shadow_domain->parent;
}
if (nvs_next->priv == domain) {
/*
* This entire thread is going to be changed soon.
* The above check ensures that there are no other domain,
* besides the active domain. So, its safe to simply return here.
* Any active domain updates shall are taken care of during
* nvgpu_nvs_worker_wakeup_process_item().
*
* This is a temporary hack for legacy cases where we donot have
* any active domains available. This needs to be relooked at
* during implementation of manual mode.
*
* A better fix is to ensure this thread is suspended during Railgate.
*/
timeslice = nvs_next->timeslice_ns;
nvgpu_mutex_release(&g->sched_mutex);
return timeslice;
}
timeslice = nvs_next->timeslice_ns;
nvgpu_domain_next = nvs_next->priv;