Re: [PATCH v4 4/4] scsi: ufs: inject errors to verify error handling

[Akinobu Mita]

2015-03-10 19:20 GMT+09:00 Gilad Broner <gbroner@xxxxxxxxxxxxxx>:
>>> +static bool inject_cmd_hang_tr(struct ufs_hba *hba)
>>> +{
>>> +       int tag;
>>> +
>>> +       tag = find_first_bit(&hba->outstanding_reqs, hba->nutrs);
>>> +       if (tag == hba->nutrs)
>>> +               return 0;
>>> +
>>> +       __clear_bit(tag, &hba->outstanding_reqs);
>>> +       hba->lrb[tag].cmd = NULL;
>>> +       __clear_bit(tag, &hba->lrb_in_use);
>>
>> hba->lrb_in_use is a bitmap set by test_and_set_bit_lock().  So
>> this should be cleared by clear_bit_unlock().
>
> You are correct. Thanks.
>
>>
>> And as soon as the bit corresponds to this slot in hba->lrb_in_use is
>> cleared, this slot could be reused.  But if the request corresponds
>> to the slot is not completed yet, it could sacrifice the new request,
>> too.  So should we only inject into the commands which have been
>> completed like this?
>
> Please note that we only clear the bit in hba->lrb_in_use. scsi_done is
> not called for this request. Therefore, the tag is not yet free in the
> block layer and next calls for queuecommand will not pass down this tag to
> be used in the UFS driver. So there is no danger of a new request being
> sacrificed.

OK, I see there is no danger as far as the commands are comming
through queuecommand().  But what about the query requests?
PATCH 1/4 in this series has added ioctl interface for query
request which also acquires a tag in hba->lrb_in_use through
ufshcd_get_dev_cmd_tag().  Although it is very difficult to
happen, is it possible for new query requests to be clashed by
inject_cmd_hang_tr() in the same way I described earlier?
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/