On 7/16/21 10:43 PM, Andrii Nakryiko wrote:
On Fri, Jul 16, 2021 at 3:05 AM Lorenz Bauer <lmb@xxxxxxxxxxxxxx> wrote:
We got the following UBSAN report on one of our testing machines:
================================================================================
UBSAN: array-index-out-of-bounds in kernel/bpf/syscall.c:2389:24
index 6 is out of range for type 'char *[6]'
CPU: 43 PID: 930921 Comm: systemd-coredum Tainted: G O 5.10.48-cloudflare-kasan-2021.7.0 #1
Hardware name: <snip>
Call Trace:
dump_stack+0x7d/0xa3
ubsan_epilogue+0x5/0x40
__ubsan_handle_out_of_bounds.cold+0x43/0x48
? seq_printf+0x17d/0x250
bpf_link_show_fdinfo+0x329/0x380
? bpf_map_value_size+0xe0/0xe0
? put_files_struct+0x20/0x2d0
? __kasan_kmalloc.constprop.0+0xc2/0xd0
seq_show+0x3f7/0x540
seq_read_iter+0x3f8/0x1040
seq_read+0x329/0x500
? seq_read_iter+0x1040/0x1040
? __fsnotify_parent+0x80/0x820
? __fsnotify_update_child_dentry_flags+0x380/0x380
vfs_read+0x123/0x460
ksys_read+0xed/0x1c0
? __x64_sys_pwrite64+0x1f0/0x1f0
do_syscall_64+0x33/0x40
entry_SYSCALL_64_after_hwframe+0x44/0xa9
<snip>
================================================================================
================================================================================
UBSAN: object-size-mismatch in kernel/bpf/syscall.c:2384:2
From the report, we can infer that some array access in bpf_link_show_fdinfo at index 6
is out of bounds. The obvious candidate is bpf_link_type_strs[BPF_LINK_TYPE_XDP] with
BPF_LINK_TYPE_XDP == 6. It turns out that BPF_LINK_TYPE_XDP is missing from bpf_types.h
and therefore doesn't have an entry in bpf_link_type_strs:
pos: 0
flags: 02000000
mnt_id: 13
link_type: (null)
link_id: 4
prog_tag: bcf7977d3b93787c
prog_id: 4
ifindex: 1
Fixes: aa8d3a716b59 ("bpf, xdp: Add bpf_link-based XDP attachment API")
Signed-off-by: Lorenz Bauer <lmb@xxxxxxxxxxxxxx>
---
Well, oops. Thanks for the fix!
Acked-by: Andrii Nakryiko <andrii@xxxxxxxxxx>
It would be great to have a compilation error for something like this.
I wonder if we can do something to detect this going forward?
include/linux/bpf_types.h | 1 +
1 file changed, 1 insertion(+)
diff --git a/include/linux/bpf_types.h b/include/linux/bpf_types.h
index a9db1eae6796..be95f2722ad9 100644
--- a/include/linux/bpf_types.h
+++ b/include/linux/bpf_types.h
@@ -135,3 +135,4 @@ BPF_LINK_TYPE(BPF_LINK_TYPE_ITER, iter)
#ifdef CONFIG_NET
BPF_LINK_TYPE(BPF_LINK_TYPE_NETNS, netns)
#endif
+BPF_LINK_TYPE(BPF_LINK_TYPE_XDP, xdp)
Lorenz, does this compile when you don't have CONFIG_NET configured? I would assume
this needs to go right below the netns one depending on CONFIG_NET.. at least the
bpf_xdp_link_lops are in net/core/dev.c which is only built under CONFIG_NET.
Thanks,
Daniel
