Re: kdevtmpfs oops since yesterdays vfs merge

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

On Sun, Jul 24, 2011 at 09:56:12PM -0400, Dave Jones wrote:

> [    7.760774] dracut: luksOpen /dev/sda2 luks-b5a1fb36-5672-4191-a260-e3f389eb0bb6
> [   14.787158] nodename: dm-0
> [   15.082391] nodename: dm-0
> 
> 
> when it triggers the bug_on(), it's that second nodename that is garbage.

Interesting...  The next experiment would be to stick BUG_ON(!req.dev)
into devtmpfs_create_node() right after the assigment to that field.

We couldn't be hit by the lack of barriers here, could we?  Store to
req.dev happens before spin_unlock(&req_lock), so by the time when
that request is seen by loop in devtmpfsd() and passed to handle() it
should be seen - we have grabbed req_lock, found a pointer to req, dropped
req_lock and called handle().  Should've been enough...

Might be interesting to print &req from devtmpfs_create_node(), both on
entry and on exit, and print req right before the call of handle()...

Incidentally, that disassembly shows one really ugly thing - offset of
->devt in struct device is 0x3c0.  IOW, each of those suckers eats a
kilobyte... ;-/