On Wed, Sep 12, 2007 at 10:26:50AM +0100, Steve McIntyre wrote:
> We're creating a small ext2 filesystem in a file to be used as an
> initrd for an embedded device. We do the following:
>
> * dd to create the empty file
> * mke2fs to create the filesystem
> * loopback mount the file
> * copy files in
> * unmount
> * fsck to make sure it will be clean when the embedded box tries to
> mount it
>
> We've triggered different behvaiour as the contents of /dev inside the
> initrd get bigger. We've just crossed a threshold, it seems. Whereas
> previously fsck would always quietly succeed, now it's doing a
> directory re-org on /dev and exiting non-zero to say it has modified
> the filesystem. This causes our Makefile to stop and report
> failure. Obviously, the Makefile can be easily fixed to be less
> paranoid but there still seems to be a problem here.
>
> If I add -O ^dir_index to the options given to mke2fs, this all works
> fine again and reports no changes from e2fsck. Equally, using older
> versions of the e2fs tools on older machines works fine; I'm assuming
> that's because they don't support/enable dir_index by default. Hence
> my suggestion that the dir_index flag is responsible. If e2fsck will
> only re-org corrupt directories then that suggests we're seeing real
> corruption.
Actually, I had forgotten that e2fsck will also re-index (or more
properly, add the htree index) to a directory which is not indexed at
all AND indexing would help it (the number of blocks is >= 3) AND if
the dir_index feature is enabled AND if the filesystem is opened
read/write (i.e., if -n is not specified). This isn't documented in
the man page, but neither is it contradicted by the man page, which
merely states that -D requests that ALL of the directories be optimised.
I'll add some additional explanatory text to the man page, but my
suggestion to you would be change the script to recognize that exit
status less than 4 does not mean that the filesystem is corrupt.
Also, if you're really never going to be using the filesystem on an
ext3 filesystem, you might as well add -O ^dir_index to mke2fs. I
actually do want to rethink how mke2fs asks users to specify ext2
vs. ext3. These days, ext3 should be the default, so it's silly to
require most users to specify -j. But maybe if they really want an
ext2 filesystem we shouldn't be enabling dir_index. And of course, if
I make any change to mke2fs's command line system, it needs to be
backwards compatible with current scripts and user expectations. Heh;
this isn't going to trivial to design.
My current thinking is to allow the user to specify a comma separated
list to mke2fs -T, with subsequent type parameters (as defined in
/etc/mke2fs.conf) overriding previous ones, so you could do something
like "mke2fs -T installer,ext2", and have it do something sane. But
more thought is required here...
- Ted
--
To UNSUBSCRIBE, email to [EMAIL PROTECTED]
with a subject of "unsubscribe". Trouble? Contact [EMAIL PROTECTED]
