Re: [RFC PATCH v1 4/4] ufs: mcq: Added ufshcd_mcq_abort()

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

On 3/8/23 14:37, Bao D. Nguyen wrote:
> On 3/8/2023 11:02 AM, Bart Van Assche wrote:
>> On 3/7/23 20:01, Bao D. Nguyen wrote:
>>> +    if (ufshcd_mcq_cqe_search(hba, hwq, tag)) {
>>> +        dev_err(hba->dev, "%s: cmd found in cq. hwq=%d, tag=%d\n",
>>> +                __func__, hwq->id, tag);
>>> +        /*
>>> +         * The command should not be 'stuck' in the CQ for such a 
>>> long time.
>>> +         * Is interrupt missing? Process the CQEs here. If the 
>>> interrupt is
>>> +         * invoked at a later time, the CQ will be empty because the 
>>> CQEs
>>> +         * are already processed here.
>>> +         */
>>> +        ufshcd_mcq_poll_cqe_lock(hba, hwq);
>>> +        err = SUCCESS;
>>> +        goto out;
>>> +    }
>>
>> Please remove the above code and also the definition of the 
>> ufshcd_mcq_cqe_search() function. The SCSI error handler submits an 
>> abort to deal with command processing timeouts. 
>> ufshcd_mcq_cqe_search() can only return true in case of a software bug 
>> at the host side. Addressing such bugs is out of scope for the SCSI 
>> error handler.
> 
> This is an attempt to handle the error case similar to SDB mode where it 
> prints "%s: cmd was completed, but without a notifying intr, tag = %d" 
> in the ufshcd_abort() function.
> 
> In this case the command has been completed by the hardware, but some 
> reasons the software has not processed it. We have seen this print 
> happened during debug sessions, so the error case does happen in SBL mode.
> 
> Are you suggesting we should return error in this case without calling 
> ufshcd_mcq_poll_cqe_lock()?

What I am asking is to remove ufshcd_mcq_poll_cqe_lock() and all code 
that depends on that function returning true. Although such code might 
be useful for SoC debugging, helping with SoC debugging is out of scope 
for Linux kernel drivers.

Thanks,

Bart.