net: rcu lock and preempt disable missing around generic xdp
authorJohn Fastabend <john.fastabend@gmail.com>
Fri, 8 Sep 2017 21:00:30 +0000 (14:00 -0700)
committerDavid S. Miller <davem@davemloft.net>
Sat, 9 Sep 2017 04:11:00 +0000 (21:11 -0700)
do_xdp_generic must be called inside rcu critical section with preempt
disabled to ensure BPF programs are valid and per-cpu variables used
for redirect operations are consistent. This patch ensures this is true
and fixes the splat below.

The netif_receive_skb_internal() code path is now broken into two rcu
critical sections. I decided it was better to limit the preempt_enable/disable
block to just the xdp static key portion and the fallout is more
rcu_read_lock/unlock calls. Seems like the best option to me.

[  607.596901] =============================
[  607.596906] WARNING: suspicious RCU usage
[  607.596912] 4.13.0-rc4+ #570 Not tainted
[  607.596917] -----------------------------
[  607.596923] net/core/dev.c:3948 suspicious rcu_dereference_check() usage!
[  607.596927]
[  607.596927] other info that might help us debug this:
[  607.596927]
[  607.596933]
[  607.596933] rcu_scheduler_active = 2, debug_locks = 1
[  607.596938] 2 locks held by pool/14624:
[  607.596943]  #0:  (rcu_read_lock_bh){......}, at: [<ffffffff95445ffd>] ip_finish_output2+0x14d/0x890
[  607.596973]  #1:  (rcu_read_lock_bh){......}, at: [<ffffffff953c8e3a>] __dev_queue_xmit+0x14a/0xfd0
[  607.597000]
[  607.597000] stack backtrace:
[  607.597006] CPU: 5 PID: 14624 Comm: pool Not tainted 4.13.0-rc4+ #570
[  607.597011] Hardware name: Dell Inc. Precision Tower 5810/0HHV7N, BIOS A17 03/01/2017
[  607.597016] Call Trace:
[  607.597027]  dump_stack+0x67/0x92
[  607.597040]  lockdep_rcu_suspicious+0xdd/0x110
[  607.597054]  do_xdp_generic+0x313/0xa50
[  607.597068]  ? time_hardirqs_on+0x5b/0x150
[  607.597076]  ? mark_held_locks+0x6b/0xc0
[  607.597088]  ? netdev_pick_tx+0x150/0x150
[  607.597117]  netif_rx_internal+0x205/0x3f0
[  607.597127]  ? do_xdp_generic+0xa50/0xa50
[  607.597144]  ? lock_downgrade+0x2b0/0x2b0
[  607.597158]  ? __lock_is_held+0x93/0x100
[  607.597187]  netif_rx+0x119/0x190
[  607.597202]  loopback_xmit+0xfd/0x1b0
[  607.597214]  dev_hard_start_xmit+0x127/0x4e0

Fixes: d445516966dc ("net: xdp: support xdp generic on virtual devices")
Fixes: b5cdae3291f7 ("net: Generic XDP")
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: John Fastabend <john.fastabend@gmail.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/core/dev.c

index 6f845e4fec175f333b568bfc44ed85a6a400ff6f..fb766d906148e7d1e85a654eaf6048c1cdf49364 100644 (file)
@@ -3981,8 +3981,13 @@ static int netif_rx_internal(struct sk_buff *skb)
        trace_netif_rx(skb);
 
        if (static_key_false(&generic_xdp_needed)) {
-               int ret = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog),
-                                        skb);
+               int ret;
+
+               preempt_disable();
+               rcu_read_lock();
+               ret = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), skb);
+               rcu_read_unlock();
+               preempt_enable();
 
                /* Consider XDP consuming the packet a success from
                 * the netdev point of view we do not want to count
@@ -4500,18 +4505,20 @@ static int netif_receive_skb_internal(struct sk_buff *skb)
        if (skb_defer_rx_timestamp(skb))
                return NET_RX_SUCCESS;
 
-       rcu_read_lock();
-
        if (static_key_false(&generic_xdp_needed)) {
-               int ret = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog),
-                                        skb);
+               int ret;
 
-               if (ret != XDP_PASS) {
-                       rcu_read_unlock();
+               preempt_disable();
+               rcu_read_lock();
+               ret = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), skb);
+               rcu_read_unlock();
+               preempt_enable();
+
+               if (ret != XDP_PASS)
                        return NET_RX_DROP;
-               }
        }
 
+       rcu_read_lock();
 #ifdef CONFIG_RPS
        if (static_key_false(&rps_needed)) {
                struct rps_dev_flow voidflow, *rflow = &voidflow;