[linux-lvm] Re: [PATCH] Re: [bug] infinite loop in generic_make_request()

[linux-lvm] Re: [PATCH] Re: [bug] infinite loop in generic_make_request()

[Andreas Dilger]

Neil Brown write:
> On Thursday November 30, ganesh@veritas.com wrote:
> > in generic_make_request(), the following code handles stacking:
> > 
> > do {
> >        q = blk_get_queue(bh->b_rdev);
> >        if (!q) {
> >                         printk(...)
> >                         buffer_IO_error(bh);
> >                         break;
> >        }
> > } while (q->make_request_fn(q, rw, bh));
> > 
> > however, make_request_fn may return -1 in case of errors. one such fn is
> > raid0_make_request(). this causes generic_make_request() to loop endlessly.
> > lvm returns 1 unconditionally, so it would also loop if an error occured in
> > lvm_map(). other bdevs might have the same problem.
> > I think a better mechanism would be to mandate that make_request_fn ought
> > to call bh->b_end_io() in case of errors and return 0.
> 
> Good catch, thanks.  Below is a patch to fix md drivers (md.c,
> linear.c and raid0.c).
> 
> NeilBrown

You need to also patch the lvm.c driver at the same time...  However, this
change now makes it very counter-intuitive compared to other kernel
functions.  Normally, we return 0 on success, and -ve on error.  Maybe
the RAID and LVM make_request functions should be changed to do that
instead (i.e. 0 on success, -ve on error, and maybe "1" if they do their
own recursion to break the loop)?

Cheers, Andreas

PS - I fixed the LVM mailing list address...
-- 
Andreas Dilger  \ "If a man ate a pound of pasta and a pound of antipasto,
                 \  would they cancel out, leaving him still hungry?"
http://www-mddsp.enel.ucalgary.ca/People/adilger/               -- Dogbert

[linux-lvm] Re: [PATCH] Re: [bug] infinite loop in generic_make_request()

[Andrea Arcangeli]

On Thu, Nov 30, 2000 at 01:05:53PM -0700, Andreas Dilger wrote:
> the RAID and LVM make_request functions should be changed to do that
> instead (i.e. 0 on success, -ve on error, and maybe "1" if they do their
> own recursion to break the loop)?

We preferred to let the lowlevel drivers to handle error themselfs to
simplify the interface. The lowlevel driver needs to call buffer_IO_error
before returning in case of error.

Andrea

[linux-lvm] Re: [PATCH] Re: [bug] infinite loop in generic_make_request()

[Andreas Dilger]

Andrea writes:
> On Thu, Nov 30, 2000 at 01:05:53PM -0700, Andreas Dilger wrote:
> > the RAID and LVM make_request functions should be changed to do that
> > instead (i.e. 0 on success, -ve on error, and maybe "1" if they do their
> > own recursion to break the loop)?
> 
> We preferred to let the lowlevel drivers to handle error themselfs to
> simplify the interface. The lowlevel driver needs to call buffer_IO_error
> before returning in case of error.

Even if the lowlevel driver handles the error case, it would still
make more sense to stick with the normal kernel practise of -ERROR,
and 0 for success.  Then, if in the future we can do something with the
error codes, at least we don't have to change the interface yet again.
Also, it is a bit confusing, because the lvm (and md, I suppose) driver
returned "0" for success in 2.2, so now you need to special-case the
return value depending on kernel version, if you want to keep the same
code for 2.2 and 2.4.

Cheers, Andreas
-- 
Andreas Dilger  \ "If a man ate a pound of pasta and a pound of antipasto,
                 \  would they cancel out, leaving him still hungry?"
http://www-mddsp.enel.ucalgary.ca/People/adilger/               -- Dogbert

[linux-lvm] Re: [PATCH] Re: [bug] infinite loop in generic_make_request()

[Andrea Arcangeli]

On Thu, Nov 30, 2000 at 02:54:19PM -0700, Andreas Dilger wrote:
> Andrea writes:
> > On Thu, Nov 30, 2000 at 01:05:53PM -0700, Andreas Dilger wrote:
> > > the RAID and LVM make_request functions should be changed to do that
> > > instead (i.e. 0 on success, -ve on error, and maybe "1" if they do their
> > > own recursion to break the loop)?
> > 
> > We preferred to let the lowlevel drivers to handle error themselfs to
> > simplify the interface. The lowlevel driver needs to call buffer_IO_error
> > before returning in case of error.
> 
> Even if the lowlevel driver handles the error case, it would still
> make more sense to stick with the normal kernel practise of -ERROR,
> and 0 for success.  Then, if in the future we can do something with the
> error codes, at least we don't have to change the interface yet again.

You shouldn't see the fact that a storage management driver returns 0 as an
error. It has a different semantic, it only means: "the make_request callback
completed the request, it wasn't a remap". That's all. As said the highlevel
layer doesn't know anything about errors anymore, it only need to know when the
request is completed.

Of course if there's an error during a remap you can't continue so you have to
say "this request is completed" and to tell this you currently have to return
0. But 0 from the point of view of the highlevel layer doesn't mean "error".

I'd suggest to take a third way that is to define only two retvals:

	BLKDEV_IO_REQ_COMPLETED
	BLKDEV_IO_REQ_REMAPPED

Then it doesn't matter anymore what number they're defined to.
generic_make_request simply keeps looping as far as it gets
BLKDEV_IO_REQ_REMAPPED as retval.

Andrea