Re: [RFC] block_device_operations prototype changes

[Neil Brown <neilb@suse.de>]

On Monday August 27, viro@ftp.linux.org.uk wrote:
> 	* a bug (AFAICT) in md.c - we open raid components r/w and if it
> fails, it fails.  Good for our purposes, but... how about raid0 on read-only
> devices?  In any case, we have a ready place to store mode_t, so it's not a
> problem for getting the right value to blkdev_put().  FWIW, I think that
> allowing fallback to r/o (and making the entire array r/o, of course) would
> be a good thing to have.  Any comments from drivers/md folks?

I've never heard any suggestion of anybody wanting to include a
readonly device in an md array -  I guess one could imagine an array
of CDROMs, but I doubt anyone would actually create one.
However I agree that falling back to read-only (particular if the
error value indicated that this was the only problem) would be a
sensible thing to do in some circumstances.

More interestingly:  one could argue that when the md array is
switched to read-only, each component device should be reopened as
read-only.  So I would really like the interface to have a way to
switch a device between read-only and read-write without closing and
re-opening.

Though I guess opening and then closing as you suggest below would be OK.

> 	* open_bdev_excl()/close_bdev_excl().  Needs an extra argument for
> the latter.  Two interesting callers:
> 	* kill_block_super() - need to store relevant mode_t in superblock,
> along with reference to bdev.  Note that just looking at sb->s_flags is *not*
> enough - some filesystems go read-only on errors and that changes ->s_flags.
> Another side of that is explicit remount from r/o to r/w and there we have
> a bug - we do _not_ tell the driver that something had happened.  Proper
> fix is simple enough - bdget() + blkdev_get() for write + blkdev_put() with
> old mode_t (i.e. FMODE_READ) once we are done remounting (and similar for
> explicit remount to r/o).


I would *really* like to see this.  When the root filesystem gets
mounted read-only, I want md to know so that it can mark the array as
'clean' straight away.  I really don't like having a reboot-notifier
to mark all arrays as 'clean' on shutdown.

So each device would be responsible for keeping a count of 'readonly'
and 'read-write' accesses?  Or would that be in common code?

I look forward to this all getting cleaned up!

NeilBrown
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/