Thank you for your contribution! Sashiko AI review found 1 potential
issue to consider:
- [Medium] accel/amdxdna: premature return in aie2_cmd_wait() when
  syncobj is reset
--

commit f05d19edddd799b35af693f49ef6b95d8743aad9
Author: Lizhi Hou <[email protected]>

accel/amdxdna: Detect a driver command without firmware response

This commit initializes cmd->result to U32_MAX before submitting debug BO
driver commands so that commands which never receive a firmware response are
reported as failures.

> diff --git a/drivers/accel/amdxdna/aie2_ctx.c 
> b/drivers/accel/amdxdna/aie2_ctx.c
> index 74d7019268cb6..7525fefa6f1ef 100644
> --- a/drivers/accel/amdxdna/aie2_ctx.c
> +++ b/drivers/accel/amdxdna/aie2_ctx.c
> @@ -985,10 +985,14 @@ static void aie2_cmd_wait(struct amdxdna_hwctx *hwctx, 
> u64 seq)
>       struct dma_fence *out_fence = aie2_cmd_get_out_fence(hwctx, seq);
>       struct amdxdna_dev *xdna = hwctx->client->xdna;
>  
> -     if (!out_fence) {
> -             XDNA_ERR(xdna, "Failed to get fence");
> +     /*
> +      * No fence to wait for. Either seq has already signaled and a later
> +      * point pruned it out of the timeline chain, or the syncobj was
> +      * reset. Neither means the command failed, so leave the verdict to
> +      * the caller, which inspects cmd->result.
> +      */
> +     if (!out_fence)
>               return;
> -     }
>  
>       mutex_unlock(&xdna->dev_lock);
>       dma_fence_wait_timeout(out_fence, false, MAX_SCHEDULE_TIMEOUT);
> @@ -1018,6 +1022,7 @@ static int aie2_hwctx_cfg_debug_bo(struct amdxdna_hwctx 
> *hwctx, u32 bo_hdl,
>               goto put_obj;
>       }
>       kref_init(&cmd->refcnt);
> +     cmd->result = U32_MAX;

[Severity: Medium]
Can aie2_hwctx_cfg_debug_bo() and aie2_hwctx_sync_debug_bo() observe
cmd->result == U32_MAX while the job is still running if the syncobj was
reset?

aie2_ctx_syncobj_create() exports hwctx->priv->syncobj to userspace as
hwctx->syncobj_hdl.  If userspace resets that syncobj while a debug BO
command is in flight, aie2_cmd_get_out_fence() returns NULL and
aie2_cmd_wait() returns immediately without waiting for the job to
finish:

CPU0                              CPU1
aie2_hwctx_cfg_debug_bo()
  cmd->result = U32_MAX
  amdxdna_cmd_submit()
                                  DRM_IOCTL_SYNCOBJ_RESET
                                    drm_syncobj_replace_fence(..., NULL)
  aie2_cmd_wait()
    out_fence = NULL
    return
  if (cmd->result)                <-- sees U32_MAX, returns -EINVAL
                                  aie2_sched_drvcmd_resp_handler()
                                    job->drv_cmd->result = 0

Unlike the pruned timeline point case where seq has already signaled,
resetting the syncobj drops the fence before the firmware responds.
aie2_hwctx_cfg_debug_bo() then sees cmd->result == U32_MAX and skips
updating abo->assigned_hwctx even though firmware still executes the
attach or detach command, leaving abo->assigned_hwctx out of sync with
firmware.

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

Reply via email to