Thank you for your contribution! Sashiko AI review found 3 potential issue(s) to consider: - [Critical] accel: ethosu: ethosu_job_sram_size() clobbers region 2 GEM BO with SRAM addr - [High] accel: ethosu: job->base.entity is always NULL in ethosu_job_set_in_flight() - [Medium] accel: ethosu: ethosu_job_close() clears sram_entity before killing entity --
commit a7a26abf2b6252cbc62a4f8adc9daddbc6d61be4 Author: Rob Herring (Arm) <[email protected]> accel: ethosu: Clear SRAM before submission This patch tracks the scheduler entity that last used the on-chip SRAM and clears the SRAM before submitting a job from a different entity. > diff --git a/drivers/accel/ethosu/ethosu_job.c > b/drivers/accel/ethosu/ethosu_job.c > index ef95a8abd61e2..3976115902e83 100644 > --- a/drivers/accel/ethosu/ethosu_job.c > +++ b/drivers/accel/ethosu/ethosu_job.c > @@ -45,11 +45,27 @@ static const struct dma_fence_ops ethosu_fence_ops = { > .get_timeline_name = ethosu_fence_get_timeline_name, > }; > > -static void ethosu_job_set_in_flight(struct ethosu_device *dev, > - struct ethosu_job *job) > +static u32 ethosu_job_sram_size(struct ethosu_job *job) > { > + struct ethosu_validated_cmdstream_info *cmd_info = > + to_ethosu_bo(job->cmd_bo)->info; > + > + return cmd_info->region_size[ETHOSU_SRAM_REGION]; > +} > + > +static bool ethosu_job_set_in_flight(struct ethosu_device *dev, > + struct ethosu_job *job, u32 sram_size) > +{ > + bool clear_sram = false; > + > guard(spinlock)(&dev->job_lock); > + if (sram_size && dev->sram_entity != job->base.entity) { > + dev->sram_entity = job->base.entity; > + clear_sram = true; > + } [Severity: High] Isn't job->base.entity always NULL when ethosu_job_set_in_flight() is called from ethosu_job_run()? In the DRM scheduler, drm_sched_run_job_work() pops the job via drm_sched_entity_pop_job() before invoking sched->ops->run_job(sched_job): drivers/gpu/drm/scheduler/sched_entity.c:drm_sched_entity_pop_job() { ... /* Jobs and entities might have different lifecycles. Since we're * removing the job from the entities queue, set the jobs entity pointer * to NULL to prevent any future access of the entity through this job. */ sched_job->entity = NULL; return sched_job; } Because job->base.entity is set to NULL before ethosu_job_run() runs, and dev->sram_entity starts at NULL, won't dev->sram_entity != job->base.entity always evaluate to false (NULL != NULL), leaving clear_sram false so that memset_io() in ethosu_job_hw_submit() never executes? Would tracking job->base.s_fence->drm_client_id (or job->base.s_fence->scheduled.context) instead of job->base.entity avoid this? > dev->in_flight_job = job; > + > + return clear_sram; > } > > static struct ethosu_job * > @@ -64,7 +80,8 @@ ethosu_job_clear_in_flight(struct ethosu_device *dev) > return job; > } > > -static void ethosu_job_hw_submit(struct ethosu_device *dev, struct > ethosu_job *job) > +static void ethosu_job_hw_submit(struct ethosu_device *dev, struct > ethosu_job *job, > + u32 sram_size, bool clear_sram) > { > struct drm_gem_dma_object *cmd_bo = to_drm_gem_dma_obj(job->cmd_bo); > struct ethosu_validated_cmdstream_info *cmd_info = > to_ethosu_bo(job->cmd_bo)->info; > @@ -79,7 +96,12 @@ static void ethosu_job_hw_submit(struct ethosu_device > *dev, struct ethosu_job *j > dev_dbg(dev->base.dev, "Region %d base addr = %pad\n", region, > &bo->dma_addr); > } > > - if (job->sram_size) { > + if (sram_size) { > + if (clear_sram) { > + memset_io(dev->sram, 0, dev->npu_info.sram_size); > + /* Order the SRAM clear before starting the NPU. */ > + wmb(); > + } > writel_relaxed(lower_32_bits(dev->sramphys), > dev->regs + NPU_REG_BASEP(ETHOSU_SRAM_REGION)); > writel_relaxed(upper_32_bits(dev->sramphys), [Severity: Critical] When a job backs region 2 (ETHOSU_SRAM_REGION) with a GEM BO and sets job->sram_size to 0, won't replacing the job->sram_size check with ethosu_job_sram_size() overwrite the region 2 GEM BO DMA address with dev->sramphys? In ethosu_ioctl_submit_job(), region 2 is only reserved for on-chip SRAM when job->sram_size is non-zero: /* BO region 2 is reserved if SRAM is used */ if (job->region_bo_handles[ETHOSU_SRAM_REGION] && job->sram_size) return -EINVAL; if (edev->npu_info.sram_size < job->sram_size) return -EINVAL; When job->sram_size is 0 and job->region_bo_handles[ETHOSU_SRAM_REGION] is non-zero, ethosu_ioctl_submit_job() only validates cmd_info->region_size[ETHOSU_SRAM_REGION] against gem->size, not against edev->npu_info.sram_size. Since ethosu_job_sram_size() unconditionally returns cmd_info->region_size[ETHOSU_SRAM_REGION] (truncated to u32) without checking job->sram_size, ethosu_job_hw_submit() will clobber NPU_REG_BASEP(ETHOSU_SRAM_REGION) and NPU_REG_BASEP_HI(ETHOSU_SRAM_REGION) with dev->sramphys, causing the NPU to perform unchecked DMA beyond the bounds of on-chip SRAM (or at physical address 0 when no SRAM is configured). Should this continue checking job->sram_size instead of cmd_info->region_size[ETHOSU_SRAM_REGION]? [ ... ] > @@ -482,12 +507,22 @@ int ethosu_job_open(struct ethosu_file_priv > *ethosu_priv) > return ret; > } > > +static void ethosu_job_clear_sram_entity(struct ethosu_device *dev, > + struct drm_sched_entity *entity) > +{ > + guard(spinlock)(&dev->job_lock); > + if (dev->sram_entity == entity) { > + dev->sram_entity = NULL; > + } > +} > + > void ethosu_job_close(struct ethosu_file_priv *ethosu_priv) > { > struct ethosu_device *dev = ethosu_priv->edev; > struct drm_sched_entity *entity = ðosu_priv->sched_entity; > > scoped_guard(mutex, &dev->sched_lock) { > + ethosu_job_clear_sram_entity(dev, entity); > drm_sched_entity_kill(entity); > list_del(ðosu_priv->sched_entity_list); > } [Severity: Medium] Can ethosu_job_clear_sram_entity() race with ethosu_job_run() if it is called before drm_sched_entity_kill()? drm_sched_run_job_work() runs on the scheduler workqueue without holding dev->sched_lock, invokes ethosu_job_run() -> ethosu_job_set_in_flight(), and only signals entity->entity_idle after run_job() returns. If ethosu_job_clear_sram_entity() runs before drm_sched_entity_kill() stops the entity and waits on entity->entity_idle, a concurrent ethosu_job_run() can repopulate dev->sram_entity right before ethosu_priv is freed in ethosu_postclose(). If a subsequent open() reallocates ethosu_file_priv at the same address, could the stale dev->sram_entity pointer match the new entity and skip clearing SRAM? -- Sashiko AI review ยท https://sashiko.dev/#/patchset/[email protected]?part=11
