Re: [PATCH] osdblk: a Linux block device for OSD objects

[Jens Axboe <jens.axboe@oracle.com>]

On Tue, Apr 07 2009, Jeff Garzik wrote:
> Boaz Harrosh wrote:
>> On 04/03/2009 12:58 PM, Jeff Garzik wrote:
>>> Jens Axboe wrote:
>>>> This wont work, GFP_NOIO inside the queue lock. You are also only
>>>> cloning the front bio, what happens if you have > 1 bio on the request?
>>>> You seem to dequeue the request and complete all of it, regardless of
>>>> whether bio->bi_size == blk_rq_bytes(rq). I'm assuming you have to clone
>>>> because of how the osd_req_{read,write} works, so I'd suggest storing
>>>> the byte size in your osdblk_request and only completing that in
>>>> osdblk_end_request(). Then do a rq_for_each_bio() look in there, and
>>>> only dequeue if you manage to start an osd request for each of them,
>>>> THEN moving on to the next request.
>>
>> There is nothing preventing from issuing a linked bio list. The only thing
>> is that osd_read/write looks at the first bio for total size.
>> If the first bio->bi_size does not specify the full length of the chain
>> then we should add another parameter to osd_read/write for that.
>>
>> The original idea was to specifically allow chained bios.
>>
>> Please advise?
>
> As passed to us from the block layer, there is nothing special about the  
> size of the first bio, AFAIK.
>
> This seems like a libosd bug?  If you want to support chained bio's, I  
> would presume you would either walk the list and sum all sizes, or in  
> some other way input the total request size?

Completely agree, if you want to support passing in a chain, you better
make the first bio just part of the chain (not some header bio).

And Jeff, you still have that bio_clone() bug in your v2 posting.

-- 
Jens Axboe