[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20100825230442.GH2738@quack.suse.cz>
Date: Thu, 26 Aug 2010 01:04:42 +0200
From: Jan Kara <jack@...e.cz>
To: Masayoshi MIZUMA <m.mizuma@...fujitsu.com>
Cc: Jan Kara <jack@...e.cz>, Andreas Dilger <adilger.kernel@...ger.ca>,
Andrew Morton <akpm@...ux-foundation.org>,
linux-ext4 <linux-ext4@...r.kernel.org>
Subject: Re: [PATCH] [RESEND] ext3: set i_extra_isize of 11th inode
On Wed 25-08-10 09:05:52, Masayoshi MIZUMA wrote:
> On Tue, 24 Aug 2010 15:17:36 +0200
> Jan Kara <jack@...e.cz> wrote:
> > On Mon 23-08-10 09:50:56, Masayoshi MIZUMA wrote:
> > > In ext3 filesystem, if following conditions 1., 2., 3. and 4. is satisfied,
> > > getfattr can't search the extended attribute (EA) after remount.
> > >
> > > Condition:
> > > 1. the inode size is over 128 byte
> > > 2. "lost+found" whose inode number is 11 was removed
> > > 3. the 11th inode is used for a file.
> > > 4. the EA locates in-inode
> > >
> > > This happens because of following logic:
> > > i_extra_isize is set to over 0 by ext3_new_inode() when we create
> > > a file whose inode number is 11 after removing "lost+found".
> > > Therefore setfattr creates the EA in-inode.
> > > After remount, i_extra_isize of 11th inode is set to 0 by ext3_iget()
> > > when we lookup the file, so getfattr tries to search the EA out-inode.
> > > However, the EA locates in-inode, so getfattr can't search the EA.
> > >
> > > How to reproduce:
> > > 1. mkfs.ext3 -I 256 /dev/sdXX
> > > 2. mount -o acl,user_xattr /dev/sdXX /TEST
> > > 3. rm -rf /TEST/*
> > > 4. touch /TEST/file (whose inode number is 11)
> > > 5. cd /TEST; setfattr -n user.foo0 -v bar0 file
> > > 6. cd /TEST; getfattr -d file
> > > -> can see foo0/bar0
> > > 7. umount /dev/sdXX
> > > 8. mount -o acl,user_xattr /dev/sdXX /TEST
> > > 9. cd /TEST; getfattr -d file
> > > -> can't see foo0/bar0
> > >
> > > Though the 11th inode is used for "lost+found" normally, the other
> > > file can also use it. Therefore, i_extra_isize of 11th inode should be set
> > > to the suitable value by ext3_iget().
> > Hmm, with which kernel have you tested that? Because my 2.6.32 kernel
> > works just fine (and looking into the code, all should be handled well).
> I tested at 2.6.35.
>
> > Look:
> > mount -o loop,user_xattr ~jack/fs-images/ext3-image /mnt/
> > quack:/crypted/home/jack # cd /mnt/
> > quack:/mnt # touch file
> > quack:/mnt # ls -i file
> > 11 file
> > quack:/mnt # setfattr -n user.foo0 -v bar0 file
> > quack:/mnt # getfattr -d file
> > # file: file
> > user.foo0="bar0"
> >
> > quack:/mnt # cd
> > quack:~ # umount /mnt
> > quack:~ # mount -o loop,user_xattr ~jack/fs-images/ext3-image /mnt/
> > quack:~ # getfattr -d /mnt/file
> > getfattr: Removing leading '/' from absolute path names
> > # file: mnt/file
> > user.foo0="bar0"
> What size is the inode ? This problem happens if the size is larger than 128 byte.
> (I tested at 256 byte inode.)
Ah, that was the reason. Thanks. But looking at the implications, I'm a bit
reluctant to do the change you propose. If someone has a filesystem created
by old mkfs, he could suddently see corrupted xattrs in his lost+found
directory with the new kernel. Not that there would be a big chance this
happens but people run various strange environments...
So I'd rather choose a safer approach for ext3 - see attached patch - it
fixes the problem for me. For ext4 your patch is definitely a way to go,
so please port it to ext4 and send it to Ted Tso. Thanks.
Honza
--
Jan Kara <jack@...e.cz>
SUSE Labs, CR
View attachment "0001-ext3-Fix-lost-extented-attributes-for-inode-with-ino.patch" of type "text/x-patch" (1722 bytes)
Powered by blists - more mailing lists