On Thursday November 30, [email protected] wrote:
> [cc'ed to maintainers of md and lvm]
>
> hi,
> 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.
>
> ganesh
Good catch, thanks. Below is a patch to fix md drivers (md.c,
linear.c and raid0.c).
NeilBrown
--- ./drivers/md/md.c 2000/11/30 08:23:51 1.2
+++ ./drivers/md/md.c 2000/11/30 09:20:05 1.3
@@ -179,7 +179,7 @@
return mddev->pers->make_request(mddev, rw, bh);
else {
buffer_IO_error(bh);
- return -1;
+ return 0;
}
}
--- ./drivers/md/linear.c 2000/11/30 08:23:51 1.2
+++ ./drivers/md/linear.c 2000/11/30 09:20:05 1.3
@@ -136,7 +136,8 @@
if (!hash->dev1) {
printk ("linear_make_request : hash->dev1==NULL for block %ld\n",
block);
- return -1;
+ buffer_IO_error(bh);
+ return 0;
}
tmp_dev = hash->dev1;
} else
@@ -145,7 +146,8 @@
if (block >= (tmp_dev->size + tmp_dev->offset)
|| block < tmp_dev->offset) {
printk ("linear_make_request: Block %ld out of bounds on dev %s size %ld offset %ld\n", block, kdevname(tmp_dev->dev), tmp_dev->size, tmp_dev->offset);
- return -1;
+ buffer_IO_error(bh);
+ return 0;
}
bh->b_rdev = tmp_dev->dev;
bh->b_rsector = bh->b_rsector - (tmp_dev->offset << 1);
--- ./drivers/md/raid0.c 2000/11/30 08:23:51 1.2
+++ ./drivers/md/raid0.c 2000/11/30 09:20:05 1.3
@@ -275,16 +275,18 @@
bad_map:
printk ("raid0_make_request bug: can't convert block across chunks or bigger than %dk %ld %d\n", chunk_size, bh->b_rsector, bh->b_size >> 10);
- return -1;
+ goto outerr;
bad_hash:
printk("raid0_make_request bug: hash==NULL for block %ld\n", block);
- return -1;
+ goto outerr;
bad_zone0:
printk ("raid0_make_request bug: hash->zone0==NULL for block %ld\n", block);
- return -1;
+ goto outerr;
bad_zone1:
printk ("raid0_make_request bug: hash->zone1==NULL for block %ld\n", block);
- return -1;
+ outerr:
+ buffer_IO_error(bh);
+ return 0;
}
static int raid0_status (char *page, mddev_t *mddev)
Neil Brown write:
> On Thursday November 30, [email protected] 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
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
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
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