Re: [git pull] vfs.git part 2

[Rasmus Villemoes <linux@rasmusvillemoes.dk>]

Al Viro <viro@ZenIV.linux.org.uk> writes:

> On Fri, Jul 12, 2013 at 12:02:45PM +0000, Rasmus Villemoes wrote:
>
>> But isn't the problem the case where dirname does not exist? I.e., the
>> application has to make sure that "/some/where" exists and is a directory
>> before open("/some/where", O_CREAT | O_TMPFILE | O_RDWR, 0666) can be
>> relied upon to fail on kernels not recognizing O_TMPFILE, instead of
>> just creating "where" in "/some".
>> 
>> Just thinking out loud, and please tell me to shut up if it doesn't make
>> sense: The documentation for O_DIRECTORY seems to imply that one could
>> require O_DIRECTORY to be given when using O_TMPFILE. The "If pathname
>> is not a directory, cause the open to fail" certainly seems to make
>> sense when O_TMPFILE is used, and older kernels should complain when
>> seeing the O_CREAT|O_DIRECTORY combination. It is a hack, though.
>
> They should, but they won't ;-/

I see; I should test before I post, but...

> It's the same problem - we do *not* validate the flags argument.
> We'll get to do_last(), hit lookup_open(), which will create the
> sucker and go to finish_open_created.  Which is past the logics
> checking for LOOKUP_DIRECTORY trying to return a non-directory and it
> would've been too late to fail anyway - the file has already been
> created.  IOW, O_DIRECTORY is ignored when O_CREAT is present *and*
> file didn't exist already.  In that case we almost certainly can treat
> that as a bug (i.e. start failing open() on O_CREAT | O_DIRECTORY in
> all cases - I'd be _very_ surprised if somebody called open() with
> such combination of flags), but that doesn't help with older
> kernels...

... it seems that if one then omits O_CREAT, things work out ok, as long
as one uses O_RDWR (which is the only sane thing to do with O_TMPFILE, I
guess):

open("/tmp/test/dir", O_DIRECTORY | O_RDWR, 0666) -> -1; Is a directory
open("/tmp/test/dir", O_DIRECTORY | O_RDONLY, 0666) -> 3; Success
open("/tmp/test/file", O_DIRECTORY | O_RDWR, 0666) -> -1; Not a directory
open("/tmp/test/link_to_file", O_DIRECTORY | O_RDWR, 0666) -> -1; Not a directory
open("/tmp/test/link_to_nowhere", O_DIRECTORY | O_RDWR, 0666) -> -1; No such file or directory
open("/tmp/test/link_to_dir", O_DIRECTORY | O_RDWR, 0666) -> -1; Is a directory
open("/tmp/test/link_to_dir", O_DIRECTORY | O_RDONLY, 0666) -> 3; Success
open("/tmp/test/link_to_dir", O_NOFOLLOW | O_DIRECTORY | O_RDWR, 0666) -> -1; Too many levels of symbolic links
open("/tmp/test/link_to_dir", O_NOFOLLOW | O_DIRECTORY | O_RDONLY, 0666) -> -1; Too many levels of symbolic links

(The above flags are what an old kernel would effectively see with or
without O_TMPFILE present, I suppose.)

How about simply making O_TMPFILE == O_DIRECTORY | O_RDWR |
O_TMPFILE_INTERNAL, and letting the correct use be 

open("/some/dir", O_TMPFILE) [with or without a mode argument]

Using O_DIRECTORY when we don't want to open a directory, and omitting
O_CREAT when we do want to create something new, is somewhat
counter-intuitive, but I think this would solve the problem with old
kernels.

Rasmus