[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

[dm-devel] Re: [RFC PATCH 1/8] rqbased-dm: allow blk_get_request() to be called from interrupt context



Hi Jens,

OK, I understand that.
But I think that the block layer assumption (depending on "current")
is not ideal.
Anyway, thank you for the information.

Thanks,
Kiyoshi Ueda


On Thu, 21 Dec 2006 08:53:05 +0100, Jens Axboe <jens axboe oracle com> wrote:
> On Wed, Dec 20 2006, Kiyoshi Ueda wrote:
> > Hi Jens,
> > 
> > On Wed, 20 Dec 2006 19:49:17 +0100, Jens Axboe <jens axboe oracle com> wrote:
> > > > > Big NACK on this - it's not only really ugly, it's also buggy to pass
> > > > > interrupt flags as function arguments. As you also mention in the 0/1
> > > > > mail, this also breaks CFQ.
> > > > > 
> > > > > Why do you need in-interrupt request allocation?
> > > >  
> > > > Because I'd like to use blk_get_request() in q->request_fn()
> > > > which can be called from interrupt context like below:
> > > >   scsi_io_completion -> scsi_end_request -> scsi_next_command
> > > >   -> scsi_run_queue -> blk_run_queue -> q->request_fn
> > > > 
> > > > Generally, device-mapper (dm) clones an original I/O and dispatches
> > > > the clones to underlying destination devices.
> > > > In the request-based dm patch, the clone creation and the dispatch
> > > > are done in q->request_fn().  To create the clone, blk_get_request()
> > > > is used to get a request from underlying destination device's queue.
> > > > By doing that in q->request_fn(), dm can deal with struct request
> > > > after bios are merged by __make_request().
> > > > 
> > > > Do you think creating another function like blk_get_request_nowait()
> > > > is acceptable?
> > > > Or request should not be allocated in q->request_fn() anyway?
> > > 
> > > You should not be allocating requests from that path, for a number of
> > > reasons.
> > 
> > Could I hear the reasons for my further work if possible?
> > Because of breaking current CFQ?  And is there any reason?
> 
> Mainly I just don't like the design, there are better ways to achieve
> what you need. The block layer has certain assumptions on the context
> from which rq allocation happens, and this breaks it. As I also
> mentioned, you cannot pass flags around as arguments. So the patch is
> even broken as-is.
> 
> -- 
> Jens Axboe


[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]