Re: bio_chain: proposed solution for bio_alloc failure and large IO simplification

[Jens Axboe]

On Fri, Jun 14 2002, Adam J. Richter wrote:
> On Fri, 14 Jun 2002 Jens Axboe wrote:
> >On Thu, Jun 13 2002, Adam J. Richter wrote:
> [...]
> >> 	newbio = bio_chain(oldbio);
> [...]
> >> 	I realize there may be locking issues in implementing this.
> 
> >but I think that statement is very naive :-)
> 
> >Essentially you are keeping a cookie to the old bio submitted, so that
> >you can merge new bio with it fairly cheaply (it's still not very cheap,
> >although you do eliminate the biggest offender, the list merge scan).
> >The first problem is that once you submit the bio, there are some things
> >that are not yours to touch anymore. You cannot change bi_next for
> >instance, you have no idea what state the bio is in wrt I/O completion.
> >Sure you can hold a reference to the bio, but all that really gets you
> >in this situation is a handle to it, nothing more. How do you decide
> >when to invalidate this cookie that you have?
> 
> 	At any time, there could be only one "hinted" bio in a
> request: the last bio in the request.  So you only have to
> clear the hint when:
> 
> 		1. you merge bio's,
> 		2. elv_next_request is called,
> 		3. newbio is submitted.
> 
> 	In all three cases q->queue_lock gets taken, so we should
> not need to add any additional spin_lock_irq's, and the two lines
> to clear the hint pointers should be trivial.

This logic is flawed. As I said, once you pass the bio to submit_bio,
you can't maintain a pointer to it for these purposes. Grabbing the
queue_lock guarentees absolutely nothing in this regard. Consider loop,
for instance. I/O could be completed by the time bio_submit returns.
Sure you can grab a reference to the bio with bio_get(), but what would
that buy you? Just that the bio at least won't be freed by the time
bio_submit returns, but that's it.

-- 
Jens Axboe