Jens Axboe wrote:
> On Mon, Apr 30 2007, Benny Halevy wrote:
>> Jens Axboe wrote:
>>> On Sun, Apr 29 2007, James Bottomley wrote:
>>>> On Sun, 2007-04-29 at 18:48 +0300, Boaz Harrosh wrote:
>>>>> FUJITA Tomonori wrote:
>>>>>> From: Boaz Harrosh <bharrosh@xxxxxxxxxxx>
>>>>>> Subject: [PATCH 4/4] bidi support: bidirectional request
>>>>>> Date: Sun, 15 Apr 2007 20:33:28 +0300
>>>>>>
>>>>>>> diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
>>>>>>> index 645d24b..16a02ee 100644
>>>>>>> --- a/include/linux/blkdev.h
>>>>>>> +++ b/include/linux/blkdev.h
>>>>>>> @@ -322,6 +322,7 @@ struct request {
>>>>>>> void *end_io_data;
>>>>>>>
>>>>>>> struct request_io_part uni;
>>>>>>> + struct request_io_part bidi_read;
>>>>>>> };
>>>>>> Would be more straightforward to have:
>>>>>>
>>>>>> struct request_io_part in;
>>>>>> struct request_io_part out;
>>>>>>
>>>>> Yes I wish I could do that. For bidi supporting drivers this is the most logical.
>>>>> But for the 99.9% of uni-directional drivers, calling rq_uni(), and being some what on
>>>>> the hotish paths, this means we will need a pointer to a uni request_io_part.
>>>>> This is bad because:
>>>>> 1st- There is no defined stage in a request life where to definitely set that pointer,
>>>>> specially in the preparation stages.
>>>>> 2nd- hacks like scsi_error.c/scsi_send_eh_cmnd() will not work at all. Now this is a
>>>>> very bad spot already, and I have a short term fix for it in the SCSI-bidi patches
>>>>> (not sent yet) but a more long term solution is needed. Once such hacks are
>>>>> cleaned up we can do what you say. This is exactly why I use the access functions
>>>>> rq_uni/rq_io/rq_in/rq_out and not open code access.
>>>> I'm still not really convinced about this approach. The primary job of
>>>> the block layer is to manage and merge READ and WRITE requests. It
>>>> serves a beautiful secondary function of queueing for arbitrary requests
>>>> it doesn't understand (REQ_TYPE_BLOCK_PC or REQ_TYPE_SPECIAL ... or
>>>> indeed any non REQ_TYPE_FS).
>>>>
>>>> bidirectional requests fall into the latter category (there's nothing
>>>> really we can do to merge them ... they're just transported by the block
>>>> layer). The only unusual feature is that they carry two bios. I think
>>>> the drivers that actually support bidirectional will be a rarity, so it
>>>> might even be advisable to add it to the queue capability (refuse
>>>> bidirectional requests at the top rather than perturbing all the drivers
>>>> to process them).
>>>>
>>>> So, what about REQ_TYPE_BIDIRECTIONAL rather than REQ_BIDI? That will
>>>> remove it from the standard path and put it on the special command type
>>>> path where we can process it specially. Additionally, if you take this
>>>> approach, you can probably simply chain the second bio through
>>>> req->special as an additional request in the stream. The only thing
>>>> that would then need modification would be the dequeue of the block
>>>> driver (it would have to dequeue both requests and prepare them) and
>>>> that needs to be done only for drivers handling bidirectional requests.
>>> I agree, I'm really not crazy about shuffling the entire request setup
>>> around just for something as exotic as bidirection commands. How about
>>> just keeping it simple - have a second request linked off the first one
>>> for the second data phase? So keep it completely seperate, not just
>>> overload ->special for 2nd bio list.
>>>
>>> So basically just add a struct request pointer, so you can do rq =
>>> rq->next_rq or something for the next data phase. I bet this would be a
>>> LOT less invasive as well, and we can get by with a few helpers to
>>> support it.
>>>
>>> And it should definitely be a request type.
>>>
>> I'm a bit confused since what you both suggest is very similar to what we've
>> proposed back in October 2006 and the impression we got was that it will be
>> better to support bidirectional block requests natively (yet to be honest,
>> James, you wanted a linked request all along).
>
> It still has to be implemented natively at the block layer, just
> differently like described above. So instead of messing all over the
> block layer adding rq_uni() stuff, just add that struct request pointer
> to the request structure for the 2nd data phase. You can relatively easy
> then modify the block layer helpers to support mapping and setup of such
> requests.
>
>> Before we go on that route again, how do you see the support for bidi
>> at the scsi mid-layer done? Again, we prefer to support that officially
>> using two struct scsi_cmnd_buff instances in struct scsi_cmnd and not as
>> a one-off feature, using special-purpose state and logic (e.g. a linked
>> struct scsi_cmd for the bidi_read sg list).
>
> The SCSI part is up to James, that can be done as either inside a single
> scsi command, or as linked scsi commands as well. I don't care too much
> about that bit, just the block layer parts :-). And the proposed block
> layer design can be used both ways by the scsi layer.
Linked SCSI commands have been obsolete since SPC-4 rev 6
(18 July 2006) after proposal 06-259r1 was accepted. That
proposal starts: "The reasons for linked commands have been
overtaken by time and events." I haven't see anyone mourning
their demise on the t10 reflector.
Mapping two requests to one bidi SCSI command might make error
handling more of a challenge.
Doug Gilbert
-
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
