Re: [PATCH 2/9] vfs: export do_splice_direct() to modules

[Al Viro <viro@ZenIV.linux.org.uk>]

On Tue, Mar 19, 2013 at 11:29:41AM +0100, Miklos Szeredi wrote:

> Copy up is a once-in-a-lifetime event for an object.  Optimizing it is
> way down in the list of things to do.  I'd drop splice in a jiffy if
> it's in the way.

What makes you think that write is any better?  Same deadlock there - check
generic_file_aio_write(), it calls the same sb_start_write()...  IOW,
switching from splice to write won't help at all.

> Much more interesting question:  what happens if we crash during a
> rename?  Whiteout implemented in the filesystem won't save us.  And
> the results are interesting: old versions of files become visible and
> similar fun.  Far from likely to happen, but ...
> 
> Add a rename-with-whiteout primitive on filesystems?  That one is not
> going to be as simple as plain whiteout.  Or?

Umm...  If/when we start caring about that kind of atomicity (and I agree
that we ought to) overlayfs approach to whiteouts will actually have much
harder time - it doesn't take much to teach a journalling fs how to do that
kind of ->rename() in a single transaction; the only question is how to tell
it that we want to leave a whiteout behind us.  Hell knows; one variant is
to add a flag, of course.  Another might be more interesting - we want some
kind of "directory is opaque" flag, so if we start reshuffling the methods,
we might try to merge unlink/rmdir/whiteout.  Rules:
	* victim is negative => create a whiteout
	* victim is a directory, parent opaque => rmdir
	* victim is a non-directory, parent opaque => unlink
	* victim is positive, parent _not_ opaque => replace with whiteout
	* old_dir in case of ->rename() is opaque => normal rename
	* old_dir in case of ->rename() is not opaque => leave whiteout behind
Non-unioned => opaque, of course (nothing showing through it).

Getting good behaviour on rename interrupted by crash is going to be _very_
tricky with any strategy other than whiteouts-in-fs, AFAICS.

Again, I have no problem whatsoever with changing the set of directory
methods, as long as the replacement is sane.  We'd done that kind of thing
before and it's not a problem.