Thank you for your contribution! Sashiko AI review found 2 potential issue(s) 
to consider:

New issues:
- [Medium] guc_ct_send_recv() still checks xe_device_wedged() instead of 
xe_device_io_blocked() when returning cancellation errors, causing state 
inconsistency.

Pre-existing issues:
- [High] Directly calling probe during AER slot reset bypasses driver core 
error handling, leading to double-remove and crashes if it fails.
--

commit 1cbbd9c3fc1b9bad858f6195cfff20e523edc2ca
Author: Arvind Yadav <[email protected]>

drm/xe: Separate AER reset state from device wedging

This commit modifies the PCI error recovery path (AER) to use a temporary
in_reset state instead of the permanent wedged flag to block driver I/O.
It introduces the xe_device_io_blocked() helper to check both conditions,
replacing xe_device_wedged() in multiple places where temporary blocks
should also apply.

> diff --git a/drivers/gpu/drm/xe/xe_guc_ct.c b/drivers/gpu/drm/xe/xe_guc_ct.c
> index 5c4733da385c5..3c3fe4928fa20 100644
> --- a/drivers/gpu/drm/xe/xe_guc_ct.c
> +++ b/drivers/gpu/drm/xe/xe_guc_ct.c
> @@ -1062,7 +1062,7 @@ static int __guc_ct_send_locked(struct xe_guc_ct *ct, 
> const u32 *action,
>       xe_gt_assert(gt, g2h_len || !num_g2h);
>       lockdep_assert_held(&ct->lock);
>  
> -     if (xe_device_wedged(ct_to_xe(ct))) {
> +     if (xe_device_io_blocked(ct_to_xe(ct))) {
>               ret = -ENOTRECOVERABLE;
>               goto out;
>       }

[ ... ]

[Severity: Medium]
In guc_ct_send_recv(), the code still checks the old xe_device_wedged() state
when returning cancellation errors:

        if (g2h_fence.fail) {
                if (g2h_fence.cancel) {
                        xe_gt_dbg(gt, "H2G request %#x canceled!\n", action[0]);
                        ret = xe_device_wedged(ct_to_xe(ct)) ? -ENOTRECOVERABLE 
: -ECANCELED;
                        goto unlock;
                }

Does this leave a state inconsistency during a PCI AER reset? If an in-flight
H2G request is cancelled because of a temporary IO block, xe_device_wedged()
will evaluate to false. This means the caller receives -ECANCELED instead of
-ENOTRECOVERABLE. Should this check be updated to use xe_device_io_blocked()
as well?

[ ... ]

> diff --git a/drivers/gpu/drm/xe/xe_pci_error.c 
> b/drivers/gpu/drm/xe/xe_pci_error.c
> index 79ce0c671549a..d82256d8721f3 100644
> --- a/drivers/gpu/drm/xe/xe_pci_error.c
> +++ b/drivers/gpu/drm/xe/xe_pci_error.c
[ ... ]
> @@ -116,7 +116,6 @@ static pci_ers_result_t xe_pci_error_slot_reset(struct 
> pci_dev *pdev)
>        * TODO: optimize by re-initializing only the hardware state and 
> re-creating
>        * kernel BOs.
>        */
> -     xe_device_clear_in_reset(xe);
>       pdev->driver->remove(pdev);
>       devres_release_group(&pdev->dev, xe->devres_group);
>  

[Severity: High]
This is a pre-existing issue, but does manually calling probe here bypass
driver core error handling?

If pdev->driver->probe(pdev, ent) fails during xe_pci_error_slot_reset(), it
returns PCI_ERS_RESULT_DISCONNECT. The PCI AER core will then unbind the
driver and call pdev->driver->remove(pdev). 

Since the failing probe likely already called xe_device_remove(xe) on its
error path, wouldn't this lead to a double remove and potential crashes?

> @@ -125,8 +124,8 @@ static pci_ers_result_t xe_pci_error_slot_reset(struct 
> pci_dev *pdev)
>  
>       xe = pdev_to_xe_device(pdev);
>  
> -     /* Wedge the device to prevent I/O operations till the resume callback 
> */
> -     atomic_set(&xe->wedged.flag, 1);
> +     /* Block the new instance until the resume callback. */
> +     xe_device_set_in_reset(xe);
>  
>       return PCI_ERS_RESULT_RECOVERED;
>  }

-- 
Sashiko AI review ยท 
https://sashiko.dev/#/patchset/[email protected]?part=2

Reply via email to