gpu: nvgpu: Resolve failed cond init.

Following changes are added to fix the issue.

1) Threads having higher priority e.g. RT may preempt
threads with sched-normal priority. As a consequence, higher priority
threads might not still see initialization of data in another thread
resulting in failures such as accessing a condition value before initialization.

Any initialization in the parent thread must be accompanied by a barrier
to make it visible in other thread. Added appropriate barriers to prevent
reordering of the initialization in the thread construction path.

2) There is a race condition between nvgpu_cond_signal() and
nvgpu_cond_destroy() in the asynchronous submit code and corresponding
worker thread's process_item callback for NVS. This may lead to
data corruption and resulting in the above errors as well. Fixed
that by adding a refcount based mechanism for ownership sharing
of the struct nvgpu_nvs_worker_item between the two threads.

Bug 3778235

Signed-off-by: Debarshi Dutta <ddutta@nvidia.com>
Change-Id: Ie9b9ba57bc1dcbb8780801be79863adc39690f72
Reviewed-on: https://git-master.nvidia.com/r/c/linux-nvgpu/+/2771535
Reviewed-by: svc-mobile-coverity <svc-mobile-coverity@nvidia.com>
Reviewed-by: svc-mobile-cert <svc-mobile-cert@nvidia.com>
Reviewed-by: Prateek Sethi <prsethi@nvidia.com>
Reviewed-by: Ketan Patil <ketanp@nvidia.com>
GVS: Gerrit_Virtual_Submit <buildbot_gerritrpt@nvidia.com>
This commit is contained in:
Debarshi Dutta
2022-09-05 09:36:58 +05:30
committed by Ketan Patil
parent 038005986e
commit 667867a199
5 changed files with 63 additions and 12 deletions

View File

@@ -18,14 +18,22 @@
#include <linux/sched.h>
#include <linux/version.h>
#include <nvgpu/barrier.h>
#include <nvgpu/thread.h>
#include <nvgpu/timers.h>
int nvgpu_thread_proxy(void *threaddata)
{
struct nvgpu_thread *thread = threaddata;
int ret = thread->fn(thread->data);
bool was_running;
int ret;
/* Ensure any initialization required for this thread is completed.
* The corresponding write barrier lies at the end of nvgpu_worker_init_common.
*/
nvgpu_smp_rmb();
ret = thread->fn(thread->data);
was_running = nvgpu_atomic_xchg(&thread->running, false);