[dm-devel] Re: question about BIO/request ordering / barriers

[Jens Axboe]

On Wed, Dec 31 2003, Christophe Saout wrote:
> Am Mi, den 31.12.2003 schrieb Jens Axboe um 16:56:
> 
> [Sorry for quoting the whole thing again, but I think the DM developers
> might know about the issue too]
> 
> > On Wed, Dec 31 2003, Christophe Saout wrote:
> > > Hi!
> > > 
> > > I'm just digging through the device-mapper code and a question came up:
> > > 
> > > Are "intermediate block device drivers" (like device-mapper) allowed to
> > > reorder BIOs?
> > > 
> > > I'm not talking about BIOs submitted from different threads at the same
> > > time but BIOs submitted from the same thread sequentially, especially
> > > writes.
> > > 
> > > That would mean that BIOs might be reordered around barriers which would
> > > break potential users.
> > 
> > That would not be a good idea. Reordering around a barrier is forbidden,
> > you may reorder as you please otherwise. Consider yourself the io
> > scheduler, that is essentially the function you are performing. The io
> > scheduler will honor the barrier as well.
> > 
> > > At the moment I suppose this shouldn't be an issue because I didn't find
> > > a single user in the whole kernel that actually submits BIOs with
> > > BIO_RW_BARRIER set via submit_bio/generic_make_request (journaling
> > > filesystems are simply waiting until all writes are finished before
> > > continueing, right?).
> > 
> > Right, there are some missing bits still.
> 
> Ah, ok. So things are probably going to break in the future if they
> aren't fixed now. That's what I wanted do know.

Yep

> > > There are same cases (in device-mapper) where
> > > 
> > > a) writes get get suspended and queued for later submission where it is
> > > not ensured that those writes are submitted before any other writes that
> > > could possibly occur after the device gets resumed (generic dm code)
> > > b) a stack (instead of a fifo) is used to queue requests and submit them
> > > later (not yet included code)
> > > c) writes can get queued but reads are directly passed through
> > > (snapshotting code too)
> > > 
> > > Also, if DM recevices a barrier shouldn't this barrier be somehow sent
> > > to all real devices instead of the one that the request is actually sent
> > > to?
> > 
> > Yes, the driver must take whatever precautions necessary...
> 
> Ok, thanks. Let's see what can be done about that.
> 
> Is it possible to create empty BIOs that just act as barrier? 

Unfortunately no, the barrier bit currently has to be tied to a bio with
content. I'd be willing to accept a patch to send down empty barrier
bio's though, it's a useful feature in this context.

> Because I think when device-mapper encounters a BIO with BIO_RW_BARRIER
> set it then should also create barriers for the other devices.

It's pretty tricky. Search the list archives for past discussions on
this.

-- 
Jens Axboe