Re: Block regression since 3.1-rc3

[Jeff Moyer <jmoyer@redhat.com>]

Tejun Heo <tj@kernel.org> writes:

> Hello, Mike.
>
> On Tue, Oct 11, 2011 at 03:56:12PM -0400, Mike Snitzer wrote:
>> > I don't object to the immediate fix but think that adding such special
>> > case is gonna make the thing even more brittle and make future changes
>> > even more difficult.  Those one off cases tend to cause pretty severe
>> > headache when someone wants to evolve common code, so let's please
>> > find out what went wrong and fix it properly so that everyone follows
>> > the same set of rules.
>> 
>> Are you referring to Jeff's fix as "the immediate fix"?  Christophe
>> seems to have had success with it after all.
>
> I meant reverting the previous commit.  Oops... it seems like I
> misread Jeff's patch.  Please read on.
>
>> As for the special case that you're suggesting makes the code more
>> brittle, etc.  If you could be more specific that'd be awesome.
>
> I was still talking about the previous attempt of making dm treated
> special by flush machinery.  (the purity thing someone was talking
> about)
>
>> Jeff asked a question about the need to kick the queue in this case (as
>> he didn't feel he had a proper justification for why it was needed).
>> 
>> If we can get a proper patch header together to justify Jeff's patch
>> that'd be great.  And then revisit any of the special casing you'd like
>> us to avoid in >= 3.2?
>> 
>> (we're obviously _very_ short on time for a 3.1 fix right now).
> ...
>> > Hmmm... another rather nasty assumption the current flush code makes
>> > is that every flush request has either zero or single bio attached to
>> > it.  The assumption has always been there for quite some time now.
>> 
>> OK.
>> 
>> > That somehow seems broken by request based dm (either that or wrong
>> > request is taking INSERT_FLUSH path).
>> 
>> Where was this issue of a flush having multiple bios reported?
>
> I was misreading Jeff's patch, so the problem is request w/o bio
> reaching INSERT_FLUSH, not rq's with multiple bio's.  Sorry about
> that.  Having another look...
>
> Ah, okay, so, blk-flush on the lower layer device is seeing
> q->flush_rq of the upper layer which doesn't have bio.  Yes, the
> BUG_ON() change looks correct to me.  That or we can do
>
>   BUG_ON(rq->bio != rq->bio_tail); /* assumes zero or single bio rq */
>
> As for the blk_run_queue_async(), it's a bit confusing.  Currently,
> the block layer isn't clear about who's responsible kicking the queue
> after putting a request onto elevator and I suppose Jeff put it there
> because blk_insert_cloned_request() doesn't kick the queue.
>
> Hmm... Jeff, you also added blk_run_queue_async() call in
> 4853abaae7e4a too.  Is there a reason why blk_insert_cloned_request()
> isn't calling __blk_run_queue() or async variant of it like
> blk_insert_request() does?
>
> At any rate, the queue kicking is a different issue.  Let's not mix
> the two here.  The BUG_ON() change looks good to me.

I can submit a two part series, sure.  I'll have to get back to you on
where I think the right place for the queue kick is.  I'll look at it in
detail tomorrow.

-Jeff