Re: [PATCH] per-backing dev unplugging #2

[Jens Axboe <axboe@suse.de>]

On Sun, Mar 14 2004, Chris Mason wrote:
> On Fri, 2004-03-12 at 15:51, Jens Axboe wrote:
> > On Fri, Mar 12 2004, Chris Mason wrote:
> > > On Fri, 2004-03-12 at 15:34, Jens Axboe wrote:
> > > 
> > > > 
> > > > I don't see how this can make too much of a difference, aside of perhaps
> > > > just moving the window a little. If page->mapping can disappear here,
> > > > that's still a possibility.
> > > 
> > > As Andrew pointed out, the mapping struct won't disappear, but
> > > page->mapping may go null.  So the idea is to use barriers to get a
> > > trusted copy of page->mapping, and use the copy everywhere.
> > 
> > So trusting an atomic assignment of mapping = page->mapping, it should
> > work. It feels a bit icky, though.
> 
> I reproduced on 2.6.4-mm1 + backing dev, but 2.6.4-mm1 alone ran fine. 
> To make a long story short, the swap address space and backing dev don't
> define an unplug_io_fn.  I was able to reproduce quickly with a swap
> heavy workload.  The patch below should fix the oops, but probably isn't
> correct solution since no queues will get unplugged while waiting on
> swap pages.

Duh of course, that's pretty silly actually. So the question is if we
want to keep assigning a dummy unplug_io_fn (default_backing_dev already
has it), or just keep the check. I propose to check like Chris added,
and just kill the default_unplug_io_fn() from readahead.c

Thanks for fixing this Chris, I wonder why your back trace from this
oops was so screwy (->unplug_io_fn() for the swap space was zero-filled,
no?)

-- 
Jens Axboe