vfs: add i_op->atomic_open()

Add a new inode operation which is called on the last component of an open.
Using this the filesystem can look up, possibly create and open the file in one
atomic operation.  If it cannot perform this (e.g. the file type turned out to
be wrong) it may signal this by returning NULL instead of an open struct file
pointer.

i_op->atomic_open() is only called if the last component is negative or needs
lookup.  Handling cached positive dentries here doesn't add much value: these
can be opened using f_op->open().  If the cached file turns out to be invalid,
the open can be retried, this time using ->atomic_open() with a fresh dentry.

For now leave the old way of using open intents in lookup and revalidate in
place.  This will be removed once all the users are converted.

David Howells noticed that if ->atomic_open() opens the file but does not create
it, handle_truncate() will be called on it even if it is not a regular file.
Fix this by checking the file type in this case too.

Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>

Documentation/filesystems/Locking

@@ -62,6 +62,9 @@ ata *);
 	int (*removexattr) (struct dentry *, const char *);
 	int (*fiemap)(struct inode *, struct fiemap_extent_info *, u64 start, u64 len);
 	void (*update_time)(struct inode *, struct timespec *, int);
+	struct file * (*atomic_open)(struct inode *, struct dentry *,
+				struct opendata *, unsigned open_flag,
+				umode_t create_mode, bool *created);
 
 locking rules:
 	all may block
@@ -89,6 +92,7 @@ listxattr:	no
 removexattr:	yes
 fiemap:		no
 update_time:	no
+atomic_open:	yes
 
 	Additionally, ->rmdir(), ->unlink() and ->rename() have ->i_mutex on
 victim.