BUG triggered with GRO SKB list_head changes

Starting with the following net-next commit, I see a BUG when starting a LXD container inside of a KVM guest using virtio-net:

d4546c2509b1 net: Convert GRO SKB handling to list_head.

Here's what the kernel spits out:

kernel BUG at /var/scm/kernel/linux/include/linux/skbuff.h:2080! invalid opcode: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC PTI

  CPU: 0 PID: 1362 Comm: libvirtd Not tainted 4.18.0-rc2+ #69
  Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Ubuntu-1.8.2-1ubuntu1 04/01/2014
  RIP: 0010:skb_pull+0x36/0x40
  Code: c6 77 24 29 f0 3b 87 84 00 00 00 89 87 80 00 00 00 72 17 89 f6 48 89 f0 48 03 87 d8 00 00 00 48 89 87 d8 00 00 00 c3 31 c0 c3 <0f> 0b 0f 1f 84 00 00 00

00 00 0f 1f 44 00 00 39 b7 80 00 00 00 76
RSP: 0000:ffff96737f6039f0 EFLAGS: 00010297
RAX: 000000009c66e2f2 RBX: 0000000000000000 RCX: 0000000000000501
RDX: 0000000000000001 RSI: 000000000000000e RDI: ffff96737f7e3938
RBP: ffff967379f40020 R08: 0000000000000000 R09: 0000000000000000
R10: ffff96737f603988 R11: ffffffffc0461335 R12: ffff967379f409e0
R13: ffff96737f7e3938 R14: 0000000000000000 R15: ffff967379e96ac0
FS: 00007fc96087e640(0000) GS:ffff96737f600000(0000) knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 00007fc913608aa0 CR3: 000000005dacc001 CR4: 00000000001606f0

  Call Trace:
   <IRQ>
   br_dev_xmit+0xe1/0x3d0 [bridge]
   dev_hard_start_xmit+0xbc/0x3b0
   __dev_queue_xmit+0xb98/0xc30
   ip_finish_output2+0x3e5/0x670
   ? ip_output+0x7f/0x250
   ip_output+0x7f/0x250
   ? ip_fragment.constprop.5+0x80/0x80
   ip_forward+0x3e2/0x650
   ? ipv4_frags_init_net+0x130/0x130
   ip_rcv+0x2be/0x500
   ? ip_local_deliver_finish+0x3b0/0x3b0
   __netif_receive_skb_core+0x6a8/0xb30
   ? lock_acquire+0xab/0x200
   ? netif_receive_skb_internal+0x2a/0x380
   netif_receive_skb_internal+0x73/0x380
   ? napi_gro_complete+0xcf/0x1b0
   dev_gro_receive+0x374/0x730
   napi_gro_receive+0x4f/0x1d0
   receive_buf+0x4b6/0x1930 [virtio_net]
   ? detach_buf+0x69/0x120 [virtio_ring]
   virtnet_poll+0x122/0x2e0 [virtio_net]
   net_rx_action+0x207/0x450
   __do_softirq+0x149/0x4ea
   irq_exit+0xbf/0xd0
   do_IRQ+0x6c/0x130
   common_interrupt+0xf/0xf
   </IRQ>
  RIP: 0010:__radix_tree_lookup+0x28/0xe0
  Code: 00 00 53 49 89 ca 41 bb 40 00 00 00 4c 8b 47 50 4c 89 c0 83 e0 03 48 83 f8 01 0f 85 a8 00 00 00 4c 89 c0 48 83 e0 fe 0f b6 08 <4c> 89 d8 48 d3 e0 48 83

e8 01 48 39 c6 76 11 e9 9f 00 00 00 4c 89
RSP: 0000:ffffae150048fcc0 EFLAGS: 00000282 ORIG_RAX: ffffffffffffffd9
RAX: ffff96735d2ef908 RBX: 000000000000001f RCX: 0000000000000006
RDX: 0000000000000000 RSI: 00000000000002e2 RDI: ffff96735d10b788
RBP: 00000000000002e2 R08: ffff96735d2ef909 R09: 0000000000000000
R10: 0000000000000000 R11: 0000000000000040 R12: 000000000000001f
R13: ffffec01c15f3a80 R14: 000000000000001f R15: ffffae150048fd18

   __do_page_cache_readahead+0x11f/0x2e0
   filemap_fault+0x408/0x660
   ext4_filemap_fault+0x2f/0x40
   __do_fault+0x1f/0xd0
   __handle_mm_fault+0x915/0xfa0
   handle_mm_fault+0x1c2/0x390
   __do_page_fault+0x2f6/0x580
   ? async_page_fault+0x5/0x20
   async_page_fault+0x1b/0x20
  RIP: 0033:0x7fc913608aa0
  Code: Bad RIP value.
  RSP: 002b:00007ffcfa9c7f08 EFLAGS: 00010206
  RAX: 0000000000000000 RBX: 0000000000000003 RCX: 0000000000000080
  RDX: 0000000000000006 RSI: 00007fc913a74bf8 RDI: 00007fc913df9720
  RBP: 0000000000000001 R08: 000055df45795700 R09: 0000000000000000
  R10: 000055df4574c010 R11: 0000000000000001 R12: 00007ffcfa9c8c38
  R13: 00007ffcfa9c8c48 R14: 00007fc913dc3d70 R15: 000055df4578ab30
  Modules linked in: veth ebtable_filter ebtables ipt_MASQUERADE xt_CHECKSUM xt_comment xt_tcpudp iptable_nat nf_conntrack_ipv4 nf_defrag_ipv4 nf_nat_ipv4 nf_nat nf_conntrack libcrc32c iptable_mangle iptable_filter bpfilter bridge stp llc fuse kvm_intel kvm irqbypass 9pnet_virtio 9pnet virtio_balloon ib_iser rdma_cm configfs iw_cm ib_cm ib_core iscsi_tcp libiscsi_tcp libiscsi scsi_transport_iscsi ip_tables x_tables virtio_net net_failover virtio_blk failover crc32_pclmul crc32c_intel pcbc aesni_intel aes_x86_64 crypto_simd cryptd glue_helper virtio_pci psmouse virtio_ring virtio

I'm not very familiar with the GRO or IP fragmentation code but I was able to identify that this change "fixes" the issue:

include/linux/skbuff.h +1 -1

index 7ccc601b55d9..a5cea572a7f1 100644
@@ -666,6 +666,7 @@ struct sk_buff {
  			/* These two members must be first. */
  			struct sk_buff		*next;
  			struct sk_buff		*prev;
+			struct list_head	list;
 
  			union {
  				struct net_device	*dev;

@@ -678,7 +679,6 @@ struct sk_buff {
  			};
  		};
  		struct rb_node		rbnode; /* used in netem & tcp stack */
-		struct list_head	list;
  	};
  	struct sock		*sk;
 

That's not the correct fix, as we wouldn't want to waste space with two list implementations always being around, but I think it shows that perhaps there is something in the call stack attempting to use both the list_head list and the ip_defrag_offset at the same time and unintentionally trouncing over the other member in the union.

I wish I had a proper fix but I suspect that someone more familiar with this code will spot the issue quickly. I didn't see anything incorrect in the list manipulations in the offending commit so some deeper knowledge of the network stack is needed.

Tyler