Re: RFC: Fix f_flags races without the BKL

[Oleg Nesterov <oleg@redhat.com>]

On 12/29, Jonathan Corbet wrote:
>
> After pondering for a while, I couldn't come up with anything better than a
> global file->f_flags mutex.  There's no point in bloating struct file with
> a mutex just for this purpose; it's hard to imagine that there will be any
> real contention for this lock.

Yes, this patch is simple and straightforward, but now we can't change
->f_flags in non-preempible context. And the global lock is not very
nice anyway.

Once again, can't we use O_LOCK_FLAGS bit? I agree, it is a bit ugly,
and I won't insist if you don't like is.

	static inline int try_lock_f_flags(struct file *file)
	{
		return !test_and_set_bit(O_LOCK_FLAGS, file->f_flags);
	}

	static inline set_f_flags(struct file *file, unsigned int flags)
	{
		file->f_flags = flags & ~O_LOCK_FLAGS;
	}

Now, nobody should change ->f_flags directly (except create/open
pathes. For example, ioctl_fionbio() should be changed:

		if (try_lock_f_flags(filp)) {
			if (on)
				set_f_flags(filp, filp->f_flags | flag);
			else
				set_f_flags(filp, filp->f_flags & ~flag);
		}

If try_lock_f_flags() fails we do nothing, as if the current owner of
O_LOCK_FLAGS changes ->f_flags after us.

What do you think?


> @@ -1116,6 +1116,7 @@ static int blkdev_open(struct inode * inode, struct file * filp)
>  	 * binary needs it. We might want to drop this workaround
>  	 * during an unstable branch.
>  	 */
> +	lock_file_flags();
>  	filp->f_flags |= O_LARGEFILE;
>
>  	if (filp->f_flags & O_NDELAY)
> @@ -1124,6 +1125,7 @@ static int blkdev_open(struct inode * inode, struct file * filp)
>  		filp->f_mode |= FMODE_EXCL;
>  	if ((filp->f_flags & O_ACCMODE) == 3)
>  		filp->f_mode |= FMODE_WRITE_IOCTL;
> +	unlock_file_flags();

do we really need lock_file_flags() here?

> diff --git a/fs/pipe.c b/fs/pipe.c
> index 7aea8b8..23ae227 100644
> --- a/fs/pipe.c
> +++ b/fs/pipe.c
> @@ -945,7 +945,9 @@ struct file *create_write_pipe(int flags)
>  		goto err_dentry;
>  	f->f_mapping = inode->i_mapping;
>
> +	lock_file_flags();
>  	f->f_flags = O_WRONLY | (flags & O_NONBLOCK);
> +	unlock_file_flags();
>  	f->f_version = 0;
>
>  	return f;
> @@ -981,7 +983,9 @@ struct file *create_read_pipe(struct file *wrf, int flags)
>  	f->f_mapping = wrf->f_path.dentry->d_inode->i_mapping;
>
>  	f->f_pos = 0;
> +	lock_file_flags();
>  	f->f_flags = O_RDONLY | (flags & O_NONBLOCK);
> +	unlock_file_flags();

Ditto. Nobody can see this file yet, we can change ->f_flags lockless.

But please correct me if I am wrong, I know nothing about fs/.

Oleg.