Re: [PATCH 2/5] ext4: Correctly handle EOFBLOCKS flag in ext4_ext_punch_hole

[Ted Ts'o <tytso@mit.edu>]

On Wed, Mar 21, 2012 at 08:23:55AM +0100, Lukas Czerner wrote:
> +	/*
> +	 * This is fugly, but even though we're going to get rid of the
> +	 * EOFBLOCKS_LF in the future, we have to handle it correctly now
> +	 * because there are still versions of e2fsck out there which
> +	 * would scream otherwise. Once the new e2fsck code ignoring
> +	 * this flag is common enough this can be removed entirely.
> +	 */
> +	if (ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS)) {
> +		struct ext4_ext_path *path;
> +		ext4_lblk_t last_block;
> +
> +		mutex_lock(&inode->i_mutex);
> +		down_read(&EXT4_I(inode)->i_data_sem);

I was looking at this patch, and I was wondering why we weren't taking
i_mutex earlier in ext4_ext_punch_hole().  The primary use of i_mutex
is to protect writes racing with each other and with truncate.  Given
that punch essentially works like truncate, and all of ext4_truncate()
is run with i_mutex down, and currently ext4_ext_punch_hole() (before
applying this patch) doesn't isn't taking i_mutex at all, I'm
wondering if we can run into problems where punch is racing against a
write --- if the pages are already in mapped, then the write might not
even need to take i_data_sem.

Lukas, Allison --- am I missing something here?

						- Ted