Thank you for your contribution! Sashiko AI review found 4 potential issue(s) to consider:
New issues: - [High] Hardware reset notifications are permanently lost on interrupted ioctl waits. - [High] Unprotected GPU register writes to GPU_CMD race with device reset. - [Low] Mixing goto-based error handling with scoped_guard() cleanup in the same function. Pre-existing issues: - [High] Spurious WARN_ON due to an off-by-one boundary check in panfrost_mmu_as_get. -- commit f478a9bf05ba523f5eb295d8a3b40bf995a8444e Author: Adrián Larumbe <[email protected]> drm/panfrost: Fix races between perfcnt and reset sequence This commit updates the reset sequence to use the reset r/w semaphore to govern hardware access. It also expands the DRM uAPI for the perfcnt dump operation to inform userspace when a reset occurs, which resets the counters back to zero. > diff --git a/drivers/gpu/drm/panfrost/panfrost_perfcnt.c > b/drivers/gpu/drm/panfrost/panfrost_perfcnt.c > --- a/drivers/gpu/drm/panfrost/panfrost_perfcnt.c > +++ b/drivers/gpu/drm/panfrost/panfrost_perfcnt.c [ ... ] > @@ -58,25 +63,107 @@ void panfrost_perfcnt_sample_done(struct panfrost_device > *pfdev) [ ... ] > +static int panfrost_perfcnt_dump_locked(struct panfrost_device *pfdev, > + u64 *reset_happened) > +{ [ ... ] > +dump_retry: > + retry = false; > + scoped_guard(rwsem_read, &pfdev->reset.lock) { > + *reset_happened = perfcnt->reset_happened; > + perfcnt->reset_happened = false; [Severity: High] If this wait is interrupted by a signal, wait_for_completion_interruptible_timeout() returns -ERESTARTSYS. Because we already cleared perfcnt->reset_happened to false above, when the syscall is automatically restarted, will it see the cleared flag and falsely report hw_reset = 0 to userspace? Could this cause userspace to permanently lose notifications about hardware resets? > + if (perfcnt->reset_failed) { > + ret = panfrost_perfcnt_hw_enable(pfdev); > + if (ret) > + return ret; > + perfcnt->reset_failed = false; > + } > + > + reinit_completion(&pfdev->perfcnt->dump_comp); > + > + gpu_write(pfdev, GPU_PERFCNT_BASE_LO, lower_32_bits(gpuva)); > + gpu_write(pfdev, GPU_PERFCNT_BASE_HI, upper_32_bits(gpuva)); > + gpu_write(pfdev, GPU_INT_CLEAR, GPU_IRQ_CLEAN_CACHES_COMPLETED | > + > GPU_IRQ_PERFCNT_SAMPLE_COMPLETED); > + gpu_write(pfdev, GPU_CMD, GPU_CMD_PERFCNT_SAMPLE); > + } > + > + /* > + * Here we release the reset semaphore because perfcnt should not get > in the way > + * of a HW reset. The downside is in the gap between the completion is > flagged > + * by the GPU IRQ handler and the sempahore taken once again, yet > another reset > + * could happen, which would lead to a valid sample being discarded. > This is > + * a solution of compromise between ensuring synchronisation with the > reset > + * sequence and letting it happen as quickly as possible for jobs to be > rerun. > + */ > + > ret = > wait_for_completion_interruptible_timeout(&pfdev->perfcnt->dump_comp, > msecs_to_jiffies(1000)); [ ... ] > @@ -132,44 +220,19 @@ static int panfrost_perfcnt_enable_locked(struct > panfrost_device *pfdev, [ ... ] > /* > * Clear the counters to start from a fresh state. > */ > gpu_write(pfdev, GPU_INT_CLEAR, GPU_IRQ_PERFCNT_SAMPLE_COMPLETED); > gpu_write(pfdev, GPU_CMD, GPU_CMD_PERFCNT_CLEAR); [Severity: High] Are these register writes exposed to the exact race condition this patch intends to fix? Executing gpu_write(pfdev, GPU_CMD, GPU_CMD_PERFCNT_CLEAR) without holding pfdev->reset.lock means a concurrent soft reset (which writes GPU_CMD_SOFT_RESET) could race on the GPU_CMD register. > - ret = panfrost_mmu_as_get(pfdev, perfcnt->mapping->mmu); > - if (ret < 0) > - goto err_vunmap; > - [ ... ] > + scoped_guard(rwsem_read, &pfdev->reset.lock) { > + ret = panfrost_perfcnt_hw_enable(pfdev); [Severity: High] This is a pre-existing issue, but does panfrost_mmu_as_get() contain an off-by-one boundary check that could trigger a spurious kernel warning? Looking at panfrost_mmu_as_get() in drivers/gpu/drm/panfrost/panfrost_mmu.c: /* * AS can be retained by active jobs or a perfcnt context, * hence the '+ 1' here. */ WARN_ON(en >= (NUM_JOB_SLOTS + 1)); If a user process submits jobs to all 3 available slots (NUM_JOB_SLOTS = 3) and simultaneously enables a performance counter session, there are 4 references. atomic_inc_return() returns 4, satisfying 4 >= 4, which erroneously triggers the WARN_ON. > + if (ret) > + goto err_vunmap; [Severity: Low] Does mixing goto with scoped_guard() here violate the subsystem cleanup guidelines? Jumping to a manual cleanup label from inside a scoped_guard creates confusing ownership semantics. -- Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=16
