Thank you for your contribution! Sashiko AI review found 5 potential issue(s) 
to consider:

Pre-existing issues:
- [Critical] Error path cleans up DRM scheduler jobs that have already been 
pushed, causing Use-After-Free and scheduler corruption.
- [Critical] Uninitialized locking context is passed to fini on the success 
path for CPU jobs without BOs, leading to a kernel panic.
- [High] BO reservation locks acquired for `clean_job` are leaked if subsequent 
validation fails.
- [High] Userspace can trigger a kernel deadlock by specifying the same BO 
handle in both the CPU job and the CSD job.
- [High] GFP_KERNEL allocation while holding dma_resv locks causes an AB-BA 
deadlock with memory reclaim.
--

commit e7875ac22f6ebc1c0172f365953ce0df35e9465b
Author: Maíra Canal <[email protected]>

drm/v3d: Make v3d_get_cpu_indirect_csd_params() a pure parser

This commit modifies v3d_get_cpu_indirect_csd_params() to only parse the
indirect CSD extension and stash the arguments. The creation of CSD and
clean jobs, along with their BO locking, is moved into
v3d_submit_cpu_ioctl() to prepare for the unified submission chain.

> diff --git a/drivers/gpu/drm/v3d/v3d_submit.c 
> b/drivers/gpu/drm/v3d/v3d_submit.c
> index fcaf3a6cfddc6..64eba912dc646 100644
> --- a/drivers/gpu/drm/v3d/v3d_submit.c
> +++ b/drivers/gpu/drm/v3d/v3d_submit.c
[ ... ]
> @@ -1404,6 +1403,17 @@ v3d_submit_cpu_ioctl(struct drm_device *dev, void 
> *data,
>               goto fail;
>       }
>  
> +     if (cpu_job->job_type == V3D_CPU_JOB_TYPE_INDIRECT_CSD) {
> +             ret = v3d_setup_csd_jobs_and_bos(file_priv, v3d,
> +                                              &cpu_job->indirect_csd.args,
> +                                              &cpu_job->indirect_csd.job,
> +                                              
> &cpu_job->indirect_csd.clean_job,
> +                                              NULL,
> +                                              &cpu_job->indirect_csd.exec);
> +             if (ret)
> +                     goto fail;
> +     }
> +

[Severity: High]
This isn't a bug introduced by this patch, but does this create a lock leak
on error paths?

If v3d_setup_csd_jobs_and_bos() successfully locks BO reservations using
cpu_job->indirect_csd.exec, and later in v3d_submit_cpu_ioctl() a failure
occurs (for example, in v3d_lookup_bos() or v3d_lock_bo_reservations()),
the error handling jumps to fail:

v3d_submit_cpu_ioctl() {
    ...
    if (ret)
        goto fail;
    ...
}

This appears to bypass the fail_unreserve: label, leaving the ww_mutex and
dma_resv locks permanently held. Could this cause subsequent GPU operations
attempting to use these BOs to stall indefinitely?

[Severity: High]
This is a pre-existing issue, but could this lead to a deadlock if userspace
specifies the same BO handle in both the CPU job and the CSD job?

When v3d_submit_cpu_ioctl() uses two separate locking contexts in the same
thread, the first context acquires the WW mutexes for the CSD job's BOs:

    ret = v3d_setup_csd_jobs_and_bos(..., &cpu_job->indirect_csd.exec);

The second locking context then attempts to acquire WW mutexes for the CPU
job's BOs:

    ret = v3d_lock_bo_reservations(&cpu_job->base, &acquire_ctx);

If the same BO is specified in both jobs, would the second context (younger
transaction) block indefinitely attempting to acquire a WW mutex already held
by the first context, leading to an unrecoverable self-deadlock?

[Severity: Critical]
This is a pre-existing issue, but is it safe to clean up DRM scheduler jobs
that have already been pushed?

In v3d_submit_cpu_ioctl(), once jobs are pushed to the DRM scheduler via
v3d_push_job(), the scheduler takes ownership of their lifecycle.

If a subsequent operation fails:

    ret = drm_sched_job_add_dependency(&clean_job->base, ...);
    if (ret)
        goto fail_unreserve;

The error path jumps to fail_unreserve: and then fail:, calling
v3d_job_cleanup() on the already-pushed jobs.

Does calling drm_sched_job_cleanup() on a pushed job destroy its s_fence and
dependencies xarray while the background scheduler thread might still be
accessing them, potentially causing a use-after-free and scheduler state
corruption?

[Severity: Critical]
This isn't a bug introduced by this patch, but is it possible for an
uninitialized locking context to be passed to the cleanup functions?

In v3d_submit_cpu_ioctl(), the local locking context is bypassed if the CPU
job has no BOs:

    if (args->bo_handle_count) {
        ...
        ret = v3d_lock_bo_reservations(&cpu_job->base, &acquire_ctx);
    }

The function then proceeds to the success path and calls:

    drm_gem_unlock_reservations(job->bo, job->bo_count, acquire_ctx);

Could this result in passing uninitialized stack memory to drm_exec_fini() or
ww_acquire_fini(), causing the cleanup function to dereference and free
garbage pointers?

[Severity: High]
This is a pre-existing issue, but does this allocation trigger an AB-BA
deadlock with memory reclaim?

In v3d_submit_cpu_ioctl(), dependencies are added while BOs are locked:

    ret = drm_sched_job_add_dependency(&csd_job->base.base,
                                       dma_fence_get(cpu_job->base.done_fence));

This internally uses xa_alloc() with the GFP_KERNEL flag while the thread
holds dma_resv locks acquired by v3d_lock_bo_reservations().

If system memory is tight, could the memory allocator trigger the shrinker,
which would attempt to evict GEM objects and acquire dma_resv locks, leading
to a deadlock?

>       clean_job = cpu_job->indirect_csd.clean_job;
>       csd_job = cpu_job->indirect_csd.job;
>

-- 
Sashiko AI review · 
https://sashiko.dev/#/patchset/[email protected]?part=9

Reply via email to