Thank you for your contribution! Sashiko AI review found 1 potential issue(s) 
to consider:
- [Medium] Concurrent modification and unannotated lockless reads of 
`gpu->reset_counter` and `ctx->reset_counter` lead to data races and lost 
increments.
--

--- Patch [2]: [PATCH v2 2/2] drm/etnaviv: Add GPU reset counters ---
commit 0f536ab6e10b2d00764a49473b750d556dcda8ee
Author: Christian Gmeiner <[email protected]>

drm/etnaviv: Add GPU reset counters

This commit adds per-GPU and per-context reset counters to etnaviv to allow
userspace to detect GPU resets and determine if their context was guilty.

> --- a/drivers/gpu/drm/etnaviv/etnaviv_drv.c
> +++ b/drivers/gpu/drm/etnaviv/etnaviv_drv.c
[ ... ]
> +static int etnaviv_ioctl_reset_query(struct drm_device *dev, void *data,
> +             struct drm_file *file)
> +{
> +     struct etnaviv_drm_private *priv = dev->dev_private;
> +     struct etnaviv_file_private *ctx = file->driver_priv;
> +     struct drm_etnaviv_reset_query *args = data;
> +     struct etnaviv_gpu *gpu;
> +
> +     if (args->flags)
> +             return -EINVAL;
> +
> +     if (args->pipe >= ETNA_MAX_PIPES)
> +             return -EINVAL;
> +
> +     gpu = priv->gpu[args->pipe];
> +     if (!gpu)
> +             return -ENXIO;
> +
> +     args->global_reset_counter = gpu->reset_counter;
> +     args->context_reset_counter = ctx->reset_counter;

[Severity: Medium]
Is it possible for these counters to be updated concurrently while being
read here? 

Since etnaviv_sched_timedout_job() can increment the counters on another CPU,
it seems this lockless read might need READ_ONCE() to prevent load tearing and
ensure the compiler doesn't optimize the read incorrectly.

> +
> +     return 0;
> +}
[ ... ]
> --- a/drivers/gpu/drm/etnaviv/etnaviv_sched.c
> +++ b/drivers/gpu/drm/etnaviv/etnaviv_sched.c
[ ... ]
> @@ -79,6 +79,9 @@ static enum drm_gpu_sched_stat 
> etnaviv_sched_timedout_job(struct drm_sched_job
>       if(sched_job)
>               drm_sched_increase_karma(sched_job);
>  
> +     gpu->reset_counter++;
> +     submit->ctx->reset_counter++;

[Severity: Medium]
Can a single userspace context submit jobs to multiple GPU pipes that
might hang at the same time? 

If so, could multiple per-GPU scheduler timeout workers execute
etnaviv_sched_timedout_job() simultaneously? If they do, both workers would
evaluate submit->ctx->reset_counter++ concurrently, which could lead to a
multi-writer data race and lost increments. 

Would it make sense to use atomic_t for the context reset counter, or at
least WRITE_ONCE() if exact counts are less critical than preventing data
races?

>       /* get the GPU back into the init state */
>       etnaviv_core_dump(submit);

-- 
Sashiko AI review ยท 
https://sashiko.dev/#/patchset/[email protected]?part=2

Reply via email to