Hi.
I'm sorry for my late response. I took vacations till yesterday.
(2011/08/04 1:25), Jan Kara wrote:
Hello,
On Wed 03-08-11 22:25:48, Toshiyuki Okajima wrote:
On Wed, 3 Aug 2011 11:57:54 +0200
Jan Kara<jack@xxxxxxx> wrote:
On Wed 03-08-11 11:42:03, Toshiyuki Okajima wrote:
(2011/08/01 18:57), Jan Kara wrote:
On Mon 01-08-11 18:45:58, Toshiyuki Okajima wrote:
(2011/08/01 17:45), Jan Kara wrote:
On Mon 01-08-11 13:54:51, Toshiyuki Okajima wrote:
If there are some inodes in orphan list while a filesystem is being
read-only mounted, we should recommend that pepole umount and then
mount it when they try to remount with read-write. But the current
message/comment recommends that they umount and then remount it.
<SNIP>
the most... BTW, I guess you didn't really see this message in practice, did
you?
No.
I have seen this message in practice while quotacheck command was repeatedly
executed per an hour.
Interesting. Are you able to reproduce this? Quotacheck does remount
read-only + remount read-write but you cannot really remount the filesystem
read-only when it has orphan inodes and so you should not see those when
you remount read-write again. Possibly there's race between remounting and
unlinking...
Yes. I can reproduce it. However, it is not frequently reproduced
by using the original procedure (qutacheck per an hour). So, I made a
reproducer.
To tell the truth, I think the race creates the message:
-----------------------------------------------------------------------
EXT3-fs:<dev>: couldn't remount RDWR because of
unprocessed orphan inode list. Please umount/remount instead.
-----------------------------------------------------------------------
which hides a serious problem.
I've inquired about this at linux-fsdevel (I think you were in CC unless
I forgot). It's a race in VFS remount code as you properly analyzed below.
People are working on fixing it but it's not trivial. Filesystem is really
a wrong place to fix such problem. If there is a trivial fix for ext3 to
workaround the issue, I can take it but I'm not willing to push anything
complex - effort should better be spent working on a generic fix.
I also think read-only remount race in VFS layer should be fixed.
However, I think this race depends on ext3/ext4 filesystem
implementation. (Orphan inode list)
So, we should modify ext3/ext4(jbd/jbd2) to fix it.
Umm, I don't understand here. If VFS makes sure that there are no
After I saw the following messages, I thought we must fix EXT3-fs error
at first. So, I created the fix patch.
(1) kernel: EXT3-fs: <dev>: couldn't remount RDWR because of
unprocessed orphan inode list. Please umount/remount instead.
(2) kernel: EXT3-fs error (device <dev>) in start_transaction: Readonly filesystem
I wasn't aware that by fixing the race between "ro-remount" and "unlink",
that EXT3-fs error can be also fixed then.
files open for writing, no unfinished operations changing the filesystem (e.g.
unlink), and no open-but-unlinked files, what remains for ext3 to check?
OK.
Now, I also think we need not modify ext3 to fix these problems.
If we can prevent to add an inode into the orphan list (to start unlinking)
while ro-remounting, we can also prevent (1) and (2).
However, new mechanism to confirm whether "no open-but-unlinked" files
exist while ro-remounting is required, isn't it?
Thanks,
Toshiyuki Okajima
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
