Re: end_request error procedure in 2.5?

[Jens Axboe]

On Mon, Sep 16 2002, Peter T. Breuer wrote:
> "A month of sundays ago Jens Axboe wrote:"
> > >  end_request( req, (req->errors == 0) ? 1 : 0 );
> > >  ..
> > > 
> > >  static void end_request(struct request *req, int uptodate) {
> > >  struct bio *bio;
> > >  while ((bio = req->bio) != NULL) {
> > >              blk_finished_io(bio_sectors(bio));
> > >              req->bio = bio->bi_next;
> > >              bio->bi_next = NULL;
> > >              bio_endio(bio, uptodate);
> > >      }
> > >      blk_put_request(req);
> > >  }
> > > 
> > > 
> > > It works fine except on error.  Kernel 2.5.31.  I understand that
> > > put_request adds the request back to a free list (if gotten from there
> > > via get_request).  The request is ordinary, except out of range ...
> > > it's produced by an e2fsck of the device when the device itself is
> > > unformatted, and the out of range request gets passed to the driver and
> > > is errored there, and "kapow" ..
> > 
> > The error is most likely in the driver calling end_that_request_first(),
> 
> Hmmm ... it's not called. The above is exactly all that is called
> and LOCAL_END_REQUEST is set. OK. I see what you are saying. Yes, I
> will direct my attention to that function instead ...
> 
>  ... and yes, I see a possible path in which the queue spinlock may be
> taken twice. OK!
> 
> Thanks!
> 
> > not the function itself. Maybe you can try to do at least some
> > debugging, I hope you are not expecting anyone to be able to help you
> > from the above report.
> 
> !! :-)
> 
> Thanks, yes I know! However, it's taken me about 4 days to get it this
> far. As you know, complete lockups are hard to debug! There's a race
> condition between the printk appearing on the console and the machine
> stopping :-(. 

NMI watchdog is invaluable for this sort of thing, beats printk by a
wide margin :-). Of course that only works of you have the hardware it
works on, but that's what test machines are for.

-- 
Jens Axboe