Re: lock i_mutex for fallocate?

[Christoph Hellwig <hch@infradead.org>]

On Wed, Aug 31, 2011 at 05:33:25PM -0700, Allison Henderson wrote:
> Hi All,
> 
> In ext4 punch hole, we realized that the punch hole operation needs
> to be done under i_mutex just like truncate.  i_mutex for truncate
> is held in the vfs layer, so we dont need to lock it at the file
> system layer, but vfs does not lock i_mutex for fallocate.  We can
> lock i_mutex for fallocate at the fs layer, but question was raised
> then: should i_mutex for fallocate be held in the vfs layer instead?
> I do not know if other file systems need i_mutex to be locked for
> fallocate, or if they might be locking it already, so I am doing
> some investigating on this idea, and also the appropriate use of
> i_mutex in general.  Can someone provide some insight this topic?

Don't do it.

i_mutex is already overloaded, and this does not fit into any
of the somewhat reasonable uses cases for it, which are:

 a) for directories the VFS uses it to protect the tree topology
 b) for regular files all generic I/O code currently uses it to
    serialize writers.
 c) the VFS uses it around truncate, and setxattr updates
 d) filesystems abuse it for internal metadata in various places

As you can see right now we do not hold it over any file operation,
and I'm absolutely against adding that.  I'd rather untange the
current uses, specificly:

 - push synchronization of setattr into the filesystems
 - push synchronization of xattr write operations into the filesystems
 - move the read/write synchronization to a separate shared/exclusive
   lock like it's already done in XFS, and like Lukas proposed for
   ext4.  This fixes the Posix compliance corner cases about reads
   beeing atomic vs writes, simplifies direct I/O locking a lot,
   and allows for more parallel direct I/O support like XFS supports.
 - try to get rid of the abuses inside filesystems as much as possible.