Re: [RFC PATCH] fs code: Push the BKL down into the file system ioctl handlers

[Arnd Bergmann]

On Friday 23 May 2008, Alan Cox wrote:
> --- a/drivers/char/raw.c
> +++ b/drivers/char/raw.c
> @@ -116,13 +116,15 @@ static int raw_release(struct inode *inode, struct file *filp)
>  /*
>   * Forward ioctls to the underlying block device.
>   */
> -static int
> -raw_ioctl(struct inode *inode, struct file *filp,
> -                 unsigned int command, unsigned long arg)
> +static long raw_ioctl(struct file *filp, unsigned int command,
> +                                               unsigned long arg)
>  {
> +       long ret;
>         struct block_device *bdev = filp->private_data;
> -
> -       return blkdev_ioctl(bdev->bd_inode, NULL, command, arg);
> +       lock_kernel();
> +       ret = blkdev_ioctl(bdev->bd_inode, NULL, command, arg);
> +       unlock_kernel();
> +       return ret;
>  }
>  
>  static void bind_device(struct raw_config_request *rq)

blkdev_ioctl and compat_blkdev_ioctl are converted already (they take
the BKL in all places of doubt), so you can call these directly without
calling lock_kernel() again.

However, reading through that code again, I noticed two nastybits:
* there is no compat_raw_ioctl function, so any user space program
trying 32 bit block ioctl calls on a raw device has been broken ever
since my patch that moved the block compat_ioctl handlers into
the block/compat_ioctl.c file. The same problem seems to exist
on pktcdvd.c.

* If any block driver had implemented unlocked_ioctl, it would be
even worse: The raw driver passes filp->private_data->bd_inode as
the inode and NULL as the file. blkdev_ioctl drops the inode
argument when calling into a driver and only passes filp, so 
if a driver then tries to access filp->f_mapping->host,
it gets a NULL pointer dereference.
Fortunately, no block driver to date implements unlocked_ioctl, and
compat_ioctl was protected from this bug by the missing
compat_raw_ioctl.

I suppose the right fix for this will be to make the blkdev
unlocked_ioctl and compat_ioctl take a bdev argument instead
of file and inode.

	Arnd <><
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/