[PATCH 0/1] net: avoid a kernel panic during sk_busy_loop

Hi Dave,

I stumbled across a reproducible kernel panic while playing around with
busy_poll on a Linux 4.9.86 kernel. There's an unfortunate interaction
between init_dummy_netdev, which doesn't bother to fill in netdev_ops, and
sk_busy_loop, which assumes netdev_ops is a valid pointer.

To reproduce on the device under test (DUT), I did:

  $ ip addr show dev wlan0
  8: wlan0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc mq [...]
      inet 172.16.122.6/23 brd 172.16.123.255 scope global wlan0
  $ sysctl -w net.core.busy_read=50
  $ nc -l 172.16.122.6 5001

Then transmitted some data to this socket from a second host:

  $ echo "foo" | nc 172.16.122.6 5001

The DUT immediately hits a kernel panic.

I've attached a patch that applies cleanly to the 4.9.87 stable release.
This fix isn't necessary for net/net-next (ndo_busy_poll was removed in
linux-4.11), but a further backport of this commit is likely required for
any stable releases older than linux-4.5.

I hope this is the right way to raise something like this. I couldn't find
a clear answer from the -stable and netdev howtos for bugs against features
that no longer exist in mainline.

Thanks,
Josh

init_dummy_netdev() leaves its netdev_ops pointer zeroed. This leads
to a NULL pointer dereference when sk_busy_loop fires against an iwlwifi
wireless adapter and checks napi->dev->netdev_ops->ndo_busy_poll.

Avoid this by ensuring that napi->dev is not a dummy device before
dereferencing napi dev's netdev_ops, preventing the following panic:

  BUG: unable to handle kernel NULL pointer dereference at 00000000000000c8
  IP: [<ffffffff817b4b72>] sk_busy_loop+0x92/0x2f0
  Call Trace:
   [<ffffffff815a3134>] ? uart_write_room+0x74/0xf0
   [<ffffffff817964a9>] sock_poll+0x99/0xa0
   [<ffffffff81223142>] do_sys_poll+0x2e2/0x520
   [<ffffffff8118d3fc>] ? get_page_from_freelist+0x3bc/0xa30
   [<ffffffff810ada22>] ? update_curr+0x62/0x140
   [<ffffffff811ea671>] ? __slab_free+0xa1/0x2a0
   [<ffffffff811ea671>] ? __slab_free+0xa1/0x2a0
   [<ffffffff8179dbb1>] ? skb_free_head+0x21/0x30
   [<ffffffff81221bd0>] ? poll_initwait+0x50/0x50
   [<ffffffff811eaa36>] ? kmem_cache_free+0x1c6/0x1e0
   [<ffffffff815a4884>] ? uart_write+0x124/0x1d0
   [<ffffffff810bd1cd>] ? remove_wait_queue+0x4d/0x60
   [<ffffffff810bd224>] ? __wake_up+0x44/0x50
   [<ffffffff81582731>] ? tty_write_unlock+0x31/0x40
   [<ffffffff8158c5c6>] ? tty_ldisc_deref+0x16/0x20
   [<ffffffff81584820>] ? tty_write+0x1e0/0x2f0
   [<ffffffff81587e50>] ? process_echoes+0x80/0x80
   [<ffffffff8120c17b>] ? __vfs_write+0x2b/0x130
   [<ffffffff8120d09a>] ? vfs_write+0x15a/0x1a0
   [<ffffffff81223455>] SyS_poll+0x75/0x100
   [<ffffffff819a6524>] entry_SYSCALL_64_fastpath+0x24/0xcf

Commit 79e7fff47b7b ("net: remove support for per driver ndo_busy_poll()")
indirectly fixed this upstream in linux-4.11 by removing the offending
pointer usage. No other users of napi->dev touch its netdev_ops.

Fixes: 060212928670 ("net: add low latency socket poll")
Fixes: ce6aea93f751 ("net: network drivers no longer need to implement ndo_busy_poll()") - 4.9.y
Signed-off-by: Josh Elsasser <redacted>
---
 net/core/dev.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/net/core/dev.c b/net/core/dev.c
index 8898618bf341..d0f67d544587 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c

@@ -5042,7 +5042,10 @@ bool sk_busy_loop(struct sock *sk, int nonblock)
 		goto out;
 
 	/* Note: ndo_busy_poll method is optional in linux-4.5 */
-	busy_poll = napi->dev->netdev_ops->ndo_busy_poll;
+	if (napi->dev->reg_state != NETREG_DUMMY)
+		busy_poll = napi->dev->netdev_ops->ndo_busy_poll;
+	else
+		busy_poll = NULL;
 
 	do {
 		rc = 0;

-- 
2.11.0

On Sun, Mar 11, 2018 at 12:22 PM, Josh Elsasser [off-list ref] wrote:

init_dummy_netdev() leaves its netdev_ops pointer zeroed. This leads
to a NULL pointer dereference when sk_busy_loop fires against an iwlwifi
wireless adapter and checks napi->dev->netdev_ops->ndo_busy_poll.

Avoid this by ensuring that napi->dev is not a dummy device before
dereferencing napi dev's netdev_ops, preventing the following panic:

Hmm, how about just checking ->netdev_ops? Checking reg_state looks
odd, although works.

On Mon, Mar 12, 2018 at 4:17 PM, Cong Wang [off-list ref] wrote:

On Sun, Mar 11, 2018 at 12:22 PM, Josh Elsasser [off-list ref] wrote:

quoted

init_dummy_netdev() leaves its netdev_ops pointer zeroed. This leads
to a NULL pointer dereference when sk_busy_loop fires against an iwlwifi
wireless adapter and checks napi->dev->netdev_ops->ndo_busy_poll.

Avoid this by ensuring that napi->dev is not a dummy device before
dereferencing napi dev's netdev_ops, preventing the following panic:

Hmm, how about just checking ->netdev_ops? Checking reg_state looks
odd, although works.

Fair point. I was trying to differentiate between an unexpected
NULL pointer and a dummy netdev, but I guess it was clever
at the expense of readability.

I'll push up a v2 that just does the obvious.