Thank you for your contribution! Sashiko AI review found no new issues
in this patch, but noticed 1 related pre-existing issue it does not
introduce:
- [Critical] accel/amdxdna: hardware use-after-free in aie2_hwctx_fini()
  [1]
--

commit 3c1fe55acb7bc743b1d9dd435d5bd52e64f6bdb4
Author: Lizhi Hou <[email protected]>

accel/amdxdna: Destroy the context when hwctx restart fails

This commit updates aie2_hwctx_restart() to stop and destroy the hardware
context when mapping host buffers, adding heaps, or configuring CUs fails.

> diff --git a/drivers/accel/amdxdna/aie2_ctx.c 
> b/drivers/accel/amdxdna/aie2_ctx.c
> index 74d7019268cb6..11f5421e497aa 100644
> --- a/drivers/accel/amdxdna/aie2_ctx.c
> +++ b/drivers/accel/amdxdna/aie2_ctx.c
> @@ -122,7 +122,7 @@ static int aie2_hwctx_restart(struct amdxdna_dev *xdna, 
> struct amdxdna_hwctx *hw
>                               heap->mem.size);
>       if (ret) {
>               XDNA_ERR(xdna, "Map host buf failed, ret %d", ret);
> -             goto out;
> +             goto destroy;
>       }
>  
>       xa_for_each_range(&hwctx->client->dev_heap_xa, heap_id, heap, 1,
> @@ -132,16 +132,20 @@ static int aie2_hwctx_restart(struct amdxdna_dev *xdna, 
> struct amdxdna_hwctx *hw
>                                       heap->mem.size);
>               if (ret) {
>                       XDNA_ERR(xdna, "Add heap %ld failed ret %d", heap_id, 
> ret);
> -                     goto out;
> +                     goto destroy;
>               }
>       }
>  
>       ret = aie2_config_cu(hwctx, NULL);
>       if (ret) {
>               XDNA_ERR(xdna, "Config cu failed, ret %d", ret);
> -             goto out;
> +             goto destroy;
>       }
>  
> +     return 0;
> +
> +destroy:
> +     aie2_hwctx_stop(xdna, hwctx, NULL);
>  out:
>       XDNA_DBG(xdna, "%s restarted, ret %d", hwctx->name, ret);
>       return ret;

[Severity: Critical]
This problem wasn't introduced by this patch, but can restarting the
scheduler in aie2_hwctx_fini() before destroying the entity issue pending
jobs to a destroyed hardware context [1]?

drivers/accel/amdxdna/aie2_ctx.c:aie2_hwctx_fini() {
    ...
        drm_sched_stop(hwctx->priv->sched, NULL);
        aie2_release_resource(hwctx);
        drm_sched_start(hwctx->priv->sched, 0);

        mutex_unlock(&xdna->dev_lock);
        drm_sched_entity_destroy(&hwctx->priv->entity);
    ...
}

drivers/accel/amdxdna/aie2_ctx.c:aie2_sched_job_run() {
    ...
        if (!hwctx->priv->mbox_chann)
                return NULL;
    ...
}

Because drm_sched_start() runs before drm_sched_entity_destroy(), the
scheduler thread pulls remaining pending jobs and passes them to
aie2_sched_job_run() after aie2_release_resource() has freed the hardware
context.  Since aie2_release_resource() clears hwctx->priv->mbox_chann,
aie2_sched_job_run() returns NULL instead of an error pointer such as
ERR_PTR(-ECANCELED), which the DRM scheduler treats as successful job
completion and silently drops the jobs while signaling success to
userspace.

[1] https://sashiko.dev/#/bug/linux-97541525-e808-4942-a13a-978ecaacc230

-- 
Sashiko AI review ยท 
https://sashiko.dev/#/patchset/[email protected]?part=1

Reply via email to