Re: [Bluez-devel] Oops involving RFCOMM and sysfs

[Al Viro <viro@ZenIV.linux.org.uk>]

On Wed, Jan 02, 2008 at 04:16:42PM +0100, Gabor Gombas wrote:
> Heh, it seems talking about a bug makes it trigger:
> 
> Jan  2 16:05:45 twister kernel: Unable to handle kernel NULL pointer dereference at 00000000000000b8 RIP: 
> Jan  2 16:05:45 twister kernel:  [<ffffffff804720a5>] mutex_lock+0x10/0x1d
 
> So the patch referenced above does not help. But I've found a very easy
> way to trigger the bug:
> 
> - do a "cat /dev/zero > /dev/rfcomm0"
> - switch the phone off
> - switch the phone on, and the kernel oopses

sysfs_get_dentry(),
                mutex_lock(&parent->d_inode->i_mutex);
hitting parent->d_inode either NULL or very close to it, depending on your
.config; most likely NULL, if offset of i_mutex is 0xb8 in your build.
That's plausible - 0xb8 is what you'd get on UP build without spinlock
debugging, lockdep, etc.

Assuming that this is what we get, everything looks explainable - we
have sysfs_rename_dir() calling sysfs_get_dentry() while the parent
gets evicted.  We don't have any exclusion, so while we are playing
silly buggers with lookups in sysfs_get_dentry() we have parent become
negative; the rest is obvious...

AFAICS, the locking here is quite broken and frankly, sysfs_get_dentry()
and the way it plays with fs/namei.c are ucking fugly.

Could you stick
	if (!parent->d_inode)
		printk(KERN_WARNING "sysfs locking blows: %s",
			parent->d_name.name);
right before
                mutex_lock(&parent->d_inode->i_mutex);
                dentry = lookup_one_noperm(cur->s_name, parent);
                mutex_unlock(&parent->d_inode->i_mutex);
in sysfs_get_dentry() (fs/sysfs/dir.c) and verify that it does, indeed,
trigger?