Re: [PATCH 2/2] scsi: scsi_error: Fix device reset is not triggered

[Bart Van Assche]

On 9/22/23 02:36, Wenchao Hao wrote:
> Fix the issue of skipping scsi_try_bus_device_reset() for devices
> which is in progress of removing in following order:
>
> T1:					T2:scsi_error_handle
> __scsi_remove_device
>    scsi_device_set_state(sdev, SDEV_DEL)
> 					// would skip device with SDEV_DEL state
>    					shost_for_each_device()
> 					  scsi_try_bus_device_reset
> 					flush all commands
>   ...
>   scsi_device is released
>
> Some drivers like smartpqi only implement eh_device_reset_handler,
> if device reset is skipped, the commands which had been sent to
> firmware or devices hardware are not cleared. The error handle
> would flush all these commands in scsi_unjam_host().
>
> When the commands are finished by hardware, use after free issue is
> triggered.
>
> Add parameter "check_state" to macro shost_for_each_device() to
> determine if check device status when traversal scsi_device
> of Scsi_Host, and set this parameter to false when traversal
> in scsi_error_handle to address this issue.

The above is incomprehensible to me. Please explain more clearly why 
this change is needed.

> diff --git a/drivers/scsi/scsi.c b/drivers/scsi/scsi.c
> index d0911bc28663..db8b9e42267c 100644
> --- a/drivers/scsi/scsi.c
> +++ b/drivers/scsi/scsi.c
> @@ -704,6 +704,23 @@ int scsi_cdl_enable(struct scsi_device *sdev, bool enable)
>   	return 0;
>   }
>   
> +static int __scsi_device_get(struct scsi_device *sdev, bool check_state)

"check_state" is a bad argument name because it does not clearly explain 
the purpose of this argument. Would "include_deleted" perhaps be a 
better name?

> +{
> +	if (check_state &&
> +	    (sdev->sdev_state == SDEV_DEL || sdev->sdev_state == SDEV_CANCEL))
> +		goto fail;
> +	if (!try_module_get(sdev->host->hostt->module))
> +		goto fail;
> +	if (!get_device(&sdev->sdev_gendev))
> +		goto fail_put_module;
> +	return 0;
> +
> +fail_put_module:
> +	module_put(sdev->host->hostt->module);
> +fail:
> +	return -ENXIO;
> +}

Looking at the above code, I think we need two functions: one that does 
not include the sdev->sdev_state check and a second function that 
includes the sdev->sdev_state check (scsi_device_get()) and calls the 
first. That will result in code that is easier to read than calls to a 
function with a boolean argument.

> diff --git a/include/scsi/scsi_device.h b/include/scsi/scsi_device.h
> index c498a12f7715..e166d053c839 100644
> --- a/include/scsi/scsi_device.h
> +++ b/include/scsi/scsi_device.h
> @@ -389,21 +389,25 @@ extern void __starget_for_each_device(struct scsi_target *, void *,
>   
>   /* only exposed to implement shost_for_each_device */
>   extern struct scsi_device *__scsi_iterate_devices(struct Scsi_Host *,
> -						  struct scsi_device *);
> +						  struct scsi_device *,
> +						  bool);
>   
>   /**
>    * shost_for_each_device - iterate over all devices of a host
>    * @sdev: the &struct scsi_device to use as a cursor
>    * @shost: the &struct scsi_host to iterate over
> + * @check_state: if skip check scsi_device's state to skip some devices
> + *               scsi_device with SDEV_DEL or SDEV_CANCEL would be skipped
> + *               if this is true
>    *
>    * Iterator that returns each device attached to @shost.  This loop
>    * takes a reference on each device and releases it at the end.  If
>    * you break out of the loop, you must call scsi_device_put(sdev).
>    */
> -#define shost_for_each_device(sdev, shost) \
> -	for ((sdev) = __scsi_iterate_devices((shost), NULL); \
> +#define shost_for_each_device(sdev, shost, check_state) \
> +	for ((sdev) = __scsi_iterate_devices((shost), NULL, check_state); \
>   	     (sdev); \
> -	     (sdev) = __scsi_iterate_devices((shost), (sdev)))
> +	     (sdev) = __scsi_iterate_devices((shost), (sdev), check_state))
>   
>   /**
>    * __shost_for_each_device - iterate over all devices of a host (UNLOCKED)

Since only the SCSI error handler passes 0 as 'check_state' argument to 
shost_for_each_device(), instead of adding a boolean argument to that 
macro, please do the following:
* Introduce a new macro for the check_state = 1 case.
* Keep the semantics for shost_for_each_device().

With this approach no SCSI LLDs will have to be modified.

Thanks,

Bart.