Re: [RFC] net: remove erroneous sk null assignment in timestamping

[Johannes Berg <johannes@sipsolutions.net>]

On Fri, 2011-10-07 at 19:40 +0200, Johannes Berg wrote:

> > It looks like skb_clone_tx_timestamp() sets clone->sk without any
> > proper refcounting, so I bet this NULL'ing it out is working
> > around that bug.
> 
> You're right. But this bug can actually trigger another way: The only
> user of this is dp83640_txtstamp() which might do kfree_skb() on this
> skb, in which case that'll likely crash trying to clean up the sk.

Not only that -- I don't even see what causes the socket to not go away
while we're waiting for the ts_work that dp83640_txtstamp() fires off to
run.

And actually, no, this is all wrong. It shouldn't kick off ts_work
anyway, ts_work only handles *RX* timestamps, never *TX*. The *TX*
timestamps are handled by decode_txts() which is called when we receive
a TXTS frame from the device... *that* dequeues the skb from tx_queue
there.

I'm afraid as is this is terminally broken. I don't have a device with a
dp83640 PHY, but I bet you can cause kernel crashes by doing something
like

while (1)
	fd = open()
	enable tx timestamping on fd;
	send(fd, frame)
	close(fd);

I don't even think any of this is privileged.

johannes