Re: [RESEND RFC PATCH v1] scsi: ufs: add retries for SSU

[Bart Van Assche <bvanassche@acm.org>]

On 2020-07-22 02:14, Stanley Chu wrote:
> Hi Bart,
> 
> On Sat, 2020-07-18 at 13:30 -0700, Bart Van Assche wrote:
>> On 2020-07-17 00:39, Lee Sang Hyun wrote:
>>> -	ret = scsi_execute(sdp, cmd, DMA_NONE, NULL, 0, NULL, &sshdr,
>>> -			START_STOP_TIMEOUT, 0, 0, RQF_PM, NULL);
>>> -	if (ret) {
>>> -		sdev_printk(KERN_WARNING, sdp,
>>> -			    "START_STOP failed for power mode: %d, result %x\n",
>>> -			    pwr_mode, ret);
>>> -		if (driver_byte(ret) == DRIVER_SENSE)
>>> -			scsi_print_sense_hdr(sdp, NULL, &sshdr);
>>> +	for (retries = 0; retries < SSU_RETRIES; retries++) {
>>> +		ret = scsi_execute(sdp, cmd, DMA_NONE, NULL, 0, NULL, &sshdr,
>>> +				START_STOP_TIMEOUT, 0, 0, RQF_PM, NULL);
>>> +		if (ret) {
>>> +			sdev_printk(KERN_WARNING, sdp,
>>> +				    "START_STOP failed for power mode: %d, result %x\n",
>>> +				    pwr_mode, ret);
>>> +			if (driver_byte(ret) == DRIVER_SENSE)
>>> +				scsi_print_sense_hdr(sdp, NULL, &sshdr);
>>> +		} else {
>>> +			break;
>>> +		}
>>
>> The ninth argument of scsi_execute() is called 'retries'. Wouldn't it be
>> better to pass a nonzero value as the 'retries' argument of
>> scsi_execute() instead of adding a loop around the scsi_execute() call?
> 
> If a SCSI command issued via scsi_execute() encounters "timeout" or
> "check condition", scsi_noretry_cmd() will return 1 (true) because
> blk_rq_is_passthrough() is true due to REQ_OP_SCSI_IN or REQ_OP_SCSI_OUT
> flag was set to this SCSI command by scsi_execute(). Therefore even a
> non-zero 'retries' value is assigned while calling scsi_execute(), the
> failed command has no chance to be retried since the decision will be
> no-retry by scsi_noretry_cmd().
> 
> (Take command timeout as example)
> scsi_times_out()->scsi_abort_command()->scmd_eh_abort_handler(), here
> scsi_noretry_cmd() returns 1, and then command will be finished (with
> error code) without retry.
> 
> In scsi_noretry_cmd(), there is a comment message in section
> "check_type" as below
> 
> 	/*
> 	 * assume caller has checked sense and determined
> 	 * the check condition was retryable.
> 	 */
> 
> I am not sure if "timeout" and "check condition" cases in such SCSI
> commands issued via scsi_execute() are specially designed to be unable
> to retry.
> 
> Would you have any suggestions if LLD drivers would like to retry these
> kinds of SCSI commands?

How about summarizing the above explanation of why the 'retry' argument
of scsi_execute() cannot be used in a two or three line comment and to
add that comment above the loop introduced by this patch?

Thanks,

Bart.