On Wed, Nov 5, 2014 at 12:27 AM, Christoph Hellwig <hch@xxxxxxxxxxxxx> wrote: > On Mon, Nov 03, 2014 at 11:04:27AM -0800, Linus Torvalds wrote: >> Oh, so you don't actually need any file contents at all? >> >> If that is actually a real usage, then maybe we should just say that >> "O_TMPFILE|O_RDONLY" is fine, and remove the check that it has to be >> writable. > > Wasn't this disallowed to prevent problems on old kernels that don't use > O_TMPFILE? In that case we'd ignore the flag and would just get a file > handle for the directory instead. Yes, that was the idea at first, as discussed at http://article.gmane.org/gmane.linux.file-systems/76273, in commit bb458c644. But soon after that, O_WRONLY was allowed (ba57ea64cb1), and O_RDWR was removed from O_TMPFILE. Now only remain the explicit checks in build_open_flags() that Linus mentioned. If we allow fd = open("/tmp", O_TMPFILE|O_RDONLY, 0600) it would be seen by an old kernel as fd = open("/tmp", O_DIRECTORY|O_RDONLY, 0600) which will succeed. But unlike the other cases the creative definition of O_TMPFILE was meant to prevent, this does not create a security risk for anyone implementing a secure tmpfile, as they would be asking for a writable fd. To implement an atomic open() with O_TMPFILE+flink, if neither O_WRONLY nor O_RDWR is in flags, you would have to manually check with fstat that fd is indeed a regular file and not a directory. At least if you need to run on old kernels. If such a changes goes in, the man page for open(2) should talk about what happens on old kernels (it already has an explanation for the writable case). -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
