Re: [PATCH v8 17/27] scsi-multipath: revalidate paths upon device unblock

[sashiko-bot@kernel.org]

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
- [High] Concurrent invocation of device block/unblock during initialization dereferences a NULL scsi_mpath_head.
- [High] TOCTOU race in scsi_device_block and scsi_internal_device_unblock accessing multipath state outside state_mutex, causing a NULL pointer dereference during device teardown.
--

commit d9e004c51ebcdd24da98702a7fb3f90971c90665
Author: John Garry <john.garry@linux.dev>

    scsi-multipath: revalidate paths upon device unblock

This patch adds a revalidation step for multipath devices when a scsi_device
is unblocked, clearing paths and requeuing any queued bios.

> diff --git a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c
> index 56200a7f5c748..1a6f5e1fd80cb 100644
> --- a/drivers/scsi/scsi_lib.c
> +++ b/drivers/scsi/scsi_lib.c
> @@ -3093,6 +3093,8 @@ static int scsi_internal_device_unblock(struct scsi_device *sdev,
>  	mutex_lock(&sdev->state_mutex);
>  	ret = scsi_internal_device_unblock_nowait(sdev, new_state);
>  	mutex_unlock(&sdev->state_mutex);
> +	if ((ret == 0) && sdev->scsi_mpath_dev)
> +		scsi_mpath_revalidate_paths(sdev->scsi_mpath_dev);

[Severity: High]
Are there race conditions involving multipath state modification outside of
the sdev->state_mutex?

It appears a concurrent __scsi_remove_device() could acquire state_mutex,
transition the device to SDEV_CANCEL or SDEV_DEL, and call
scsi_mpath_remove_device(). This sets scsi_mpath_dev->scsi_mpath_head to
NULL. When scsi_internal_device_unblock() resumes, it could call
scsi_mpath_revalidate_paths() which then attempts to use the NULL pointer.

Is it also possible for this code to execute on a partially initialized
multipath device?

During device discovery, scsi_mpath_dev_alloc() assigns sdev->scsi_mpath_dev
but initially leaves scsi_mpath_dev->scsi_mpath_head as NULL while
performing a blocking INQUIRY. If a transport event (like a link flap)
concurrently triggers scsi_target_block() and scsi_target_unblock(),
scsi_internal_device_unblock() could be called and operate on the still-NULL
scsi_mpath_head.

The same unprotected call pattern appears to exist in scsi_device_block(),
which was introduced in a prior patch in this series.

[ ... ]
> diff --git a/drivers/scsi/scsi_multipath.c b/drivers/scsi/scsi_multipath.c
> index b16fe6a55fa73..456351fee4b98 100644
> --- a/drivers/scsi/scsi_multipath.c
> +++ b/drivers/scsi/scsi_multipath.c
> @@ -238,6 +238,14 @@ static int scsi_multipath_sdev_init(struct scsi_device *sdev)
>  	return 0;
>  }
>  
> +void scsi_mpath_revalidate_paths(struct scsi_mpath_device *scsi_mpath_dev)
> +{
> +       struct scsi_mpath_head *scsi_mpath_head = scsi_mpath_dev->scsi_mpath_head;
> +       struct mpath_head *mpath_head = &scsi_mpath_head->mpath_head;

[Severity: High]
If scsi_mpath_head is NULL due to the races mentioned above during unblock
or block operations, will this unconditionally dereference the NULL pointer
when calculating the offset for mpath_head?

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260907105021.1763236-1-john.g.garry@oracle.com?part=17