Re: [syzbot] KMSAN: uninit-value in virtio_net_hdr_to_skb

On Mon, May 17, 2021 at 10:57 AM Dmitry Vyukov [off-list ref] wrote:

On Mon, May 17, 2021 at 4:06 PM Willem de Bruijn
[off-list ref] wrote:

quoted

On Mon, May 17, 2021 at 7:27 AM syzbot
[off-list ref] wrote:

quoted

Hello,

syzbot found the following issue on:

HEAD commit:    4ebaab5f kmsan: drop unneeded references to kmsan_context_..
git tree:       https://github.com/google/kmsan.git master
console output: https://syzkaller.appspot.com/x/log.txt?x=17ac508ed00000
kernel config:  https://syzkaller.appspot.com/x/.config?x=ab8076fe8508c0d3
dashboard link: https://syzkaller.appspot.com/bug?extid=106457891e3cf3b273a9
compiler:       Debian clang version 11.0.1-2
syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=138f4972d00000
C reproducer:   https://syzkaller.appspot.com/x/repro.c?x=1624ffced00000

IMPORTANT: if you fix the issue, please add the following tag to the commit:
Reported-by: syzbot+106457891e3cf3b273a9@syzkaller.appspotmail.com

=====================================================
BUG: KMSAN: uninit-value in virtio_net_hdr_to_skb+0x1414/0x14f0 include/linux/virtio_net.h:86

No answer/fix, just initial investigation.

This is an odd location. Line 86 is the inner if statement. Both
protocol and skb->protocol are clearly initialized by then. But, that
is also not the allocation that MSAN reports, see below.

                        if (!skb->protocol) {
                                __be16 protocol =
dev_parse_header_protocol(skb);

                                virtio_net_hdr_set_proto(skb, hdr);
                                if (protocol && protocol != skb->protocol)
                                        return -EINVAL;
                        }

The repro itself seems mostly straightforward:

- create a packet socket
- enable PACKET_VNET_HDR with setsockopt(r3, 0x107, 0xf ..)
- bind to AF_PACKET (0x11)

- create a pipe
- write to pipe[1]
- splice pipe[0] to the packet socket

there are a few other calls that I think are irrelevant and/or would fail.

Perhaps there is some race condition in device refcounting, as bind
operates on that?

quoted

CPU: 0 PID: 8426 Comm: syz-executor777 Not tainted 5.12.0-rc6-syzkaller #0
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011
Call Trace:
 __dump_stack lib/dump_stack.c:79 [inline]
 dump_stack+0x24c/0x2e0 lib/dump_stack.c:120
 kmsan_report+0xfb/0x1e0 mm/kmsan/kmsan_report.c:118
 __msan_warning+0x5c/0xa0 mm/kmsan/kmsan_instr.c:197
 virtio_net_hdr_to_skb+0x1414/0x14f0 include/linux/virtio_net.h:86
 packet_snd net/packet/af_packet.c:2994 [inline]
 packet_sendmsg+0x85b8/0x99d0 net/packet/af_packet.c:3031
 sock_sendmsg_nosec net/socket.c:654 [inline]
 sock_sendmsg net/socket.c:674 [inline]
 kernel_sendmsg+0x22c/0x2f0 net/socket.c:694
 sock_no_sendpage+0x205/0x2b0 net/core/sock.c:2860
 kernel_sendpage+0x47a/0x590 net/socket.c:3631
 sock_sendpage+0x161/0x1a0 net/socket.c:947
 pipe_to_sendpage+0x3e4/0x520 fs/splice.c:364
 splice_from_pipe_feed fs/splice.c:418 [inline]
 __splice_from_pipe+0x5e3/0xff0 fs/splice.c:562
 splice_from_pipe fs/splice.c:597 [inline]
 generic_splice_sendpage+0x1d5/0x2c0 fs/splice.c:746
 do_splice_from fs/splice.c:767 [inline]
 do_splice+0x23c3/0x2c10 fs/splice.c:1079
 __do_splice fs/splice.c:1144 [inline]
 __do_sys_splice fs/splice.c:1350 [inline]
 __se_sys_splice+0x8fa/0xb50 fs/splice.c:1332
 __x64_sys_splice+0x6e/0x90 fs/splice.c:1332
 do_syscall_64+0x9f/0x140 arch/x86/entry/common.c:48
 entry_SYSCALL_64_after_hwframe+0x44/0xae
RIP: 0033:0x449a39
Code: 28 00 00 00 75 05 48 83 c4 28 c3 e8 d1 15 00 00 90 48 89 f8 48 89 f7 48 89 d6 48 89 ca 4d 89 c2 4d 89 c8 4c 8b 4c 24 08 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 c7 c1 b8 ff ff ff f7 d8 64 89 01 48
RSP: 002b:00007f8ed790b2f8 EFLAGS: 00000246 ORIG_RAX: 0000000000000113
RAX: ffffffffffffffda RBX: 00000000004cf518 RCX: 0000000000449a39
RDX: 0000000000000005 RSI: 0000000000000000 RDI: 0000000000000003
RBP: 00000000004cf510 R08: 000000000004ffe0 R09: 0000000000000000
R10: 0000000000000000 R11: 0000000000000246 R12: 00000000004cf51c
R13: 000000000049e46c R14: 6d32cc5e8ead0600 R15: 0000000000022000

Uninit was created at:
 kmsan_save_stack_with_flags+0x3c/0x90 mm/kmsan/kmsan.c:121
 kmsan_alloc_page+0xd0/0x1e0 mm/kmsan/kmsan_shadow.c:274
 __alloc_pages_nodemask+0x827/0xf90 mm/page_alloc.c:5044
 alloc_pages_current+0x7b6/0xb60 mm/mempolicy.c:2277
 alloc_pages include/linux/gfp.h:561 [inline]
 alloc_slab_page mm/slub.c:1653 [inline]
 allocate_slab+0x364/0x1260 mm/slub.c:1793
 new_slab mm/slub.c:1856 [inline]
 new_slab_objects mm/slub.c:2602 [inline]
 ___slab_alloc+0xd42/0x1930 mm/slub.c:2765
 __slab_alloc mm/slub.c:2805 [inline]
 slab_alloc_node mm/slub.c:2886 [inline]
 slab_alloc mm/slub.c:2931 [inline]
 kmem_cache_alloc_trace+0xc53/0x1030 mm/slub.c:2948
 kmalloc include/linux/slab.h:554 [inline]
 kzalloc include/linux/slab.h:684 [inline]
 ____ip_mc_inc_group+0x4d7/0x10b0 net/ipv4/igmp.c:1435

This allocates ip_mc_list, but it uses kzalloc. Can that ever count as
uninitialized?

Yes, kzalloc should never be a source of uninitialized-ness.
But it's not actually this kzalloc, it's underlying page allocation
(that is allocated uninitialized, so can be source of
uninitialized-ness).
If it would be this kzalloc, then stack would be shorter, along the
lines of kzalloc->kmem_cache_alloc_trace->kmsan_save_stack_with_flags.

This smells like a wild access in virtio_net_hdr_to_skb, which just
hit a random uninit somewhere.
Searching for virtio_net_hdr_to_skb I found this:

KASAN: use-after-free Read in eth_header_parse_protocol
https://syzkaller.appspot.com/bug?id=a486048b63065fd224f57b16d5a2fdece2b40eca

Can it be a dup of that bug?

Great find.

That commit is not yet present at kmsan.git at 4ebaab5fb428.

Certainly sounds plausible.