Re: [syzbot] [bpf?] KCSAN: data-race in bpf_percpu_array_update / bpf_percpu_array_update (2)

[Marco Elver]

On Tue, 29 Aug 2023 at 22:04, Daniel Borkmann <daniel@xxxxxxxxxxxxx> wrote:
>
> On 8/29/23 8:53 PM, Marco Elver wrote:
> > On Tue, 29 Aug 2023 at 20:30, Yonghong Song <yonghong.song@xxxxxxxxx> wrote:
> >> On 8/29/23 5:39 AM, syzbot wrote:
> >>> Hello,
> >>>
> >>> syzbot found the following issue on:
> >>>
> >>> HEAD commit:    727dbda16b83 Merge tag 'hardening-v6.6-rc1' of git://git.k..
> >>> git tree:       upstream
> >>> console output: https://syzkaller.appspot.com/x/log.txt?x=136f39dfa80000
> >>> kernel config:  https://syzkaller.appspot.com/x/.config?x=dea9c2ce3f646a25
> >>> dashboard link: https://syzkaller.appspot.com/bug?extid=97522333291430dd277f
> >>> compiler:       Debian clang version 15.0.6, GNU ld (GNU Binutils for Debian) 2.40
> >>>
> >>> Unfortunately, I don't have any reproducer for this issue yet.
> >>>
> >>> Downloadable assets:
> >>> disk image: https://storage.googleapis.com/syzbot-assets/9923a023ab11/disk-727dbda1.raw.xz
> >>> vmlinux: https://storage.googleapis.com/syzbot-assets/650dbc695d77/vmlinux-727dbda1.xz
> >>> kernel image: https://storage.googleapis.com/syzbot-assets/361da71276bf/bzImage-727dbda1.xz
> >>>
> >>> IMPORTANT: if you fix the issue, please add the following tag to the commit:
> >>> Reported-by: syzbot+97522333291430dd277f@xxxxxxxxxxxxxxxxxxxxxxxxx
> >>>
> >>> ==================================================================
> >>> BUG: KCSAN: data-race in bpf_percpu_array_update / bpf_percpu_array_update
> >>>
> >>> write to 0xffffe8fffe7425d8 of 8 bytes by task 8257 on cpu 1:
> >>>    bpf_long_memcpy include/linux/bpf.h:428 [inline]
> >>>    bpf_obj_memcpy include/linux/bpf.h:441 [inline]
> >>>    copy_map_value_long include/linux/bpf.h:464 [inline]
> >>>    bpf_percpu_array_update+0x3bb/0x500 kernel/bpf/arraymap.c:380
> >>>    bpf_map_update_value+0x190/0x370 kernel/bpf/syscall.c:175
> >>>    generic_map_update_batch+0x3ae/0x4f0 kernel/bpf/syscall.c:1749
> >>>    bpf_map_do_batch+0x2df/0x3d0 kernel/bpf/syscall.c:4648
> >>>    __sys_bpf+0x28a/0x780
> >>>    __do_sys_bpf kernel/bpf/syscall.c:5241 [inline]
> >>>    __se_sys_bpf kernel/bpf/syscall.c:5239 [inline]
> >>>    __x64_sys_bpf+0x43/0x50 kernel/bpf/syscall.c:5239
> >>>    do_syscall_x64 arch/x86/entry/common.c:50 [inline]
> >>>    do_syscall_64+0x41/0xc0 arch/x86/entry/common.c:80
> >>>    entry_SYSCALL_64_after_hwframe+0x63/0xcd
> >>>
> >>> write to 0xffffe8fffe7425d8 of 8 bytes by task 8268 on cpu 0:
> >>>    bpf_long_memcpy include/linux/bpf.h:428 [inline]
> >>>    bpf_obj_memcpy include/linux/bpf.h:441 [inline]
> >>>    copy_map_value_long include/linux/bpf.h:464 [inline]
> >>>    bpf_percpu_array_update+0x3bb/0x500 kernel/bpf/arraymap.c:380
> >>>    bpf_map_update_value+0x190/0x370 kernel/bpf/syscall.c:175
> >>>    generic_map_update_batch+0x3ae/0x4f0 kernel/bpf/syscall.c:1749
> >>>    bpf_map_do_batch+0x2df/0x3d0 kernel/bpf/syscall.c:4648
> >>>    __sys_bpf+0x28a/0x780
> >>>    __do_sys_bpf kernel/bpf/syscall.c:5241 [inline]
> >>>    __se_sys_bpf kernel/bpf/syscall.c:5239 [inline]
> >>>    __x64_sys_bpf+0x43/0x50 kernel/bpf/syscall.c:5239
> >>>    do_syscall_x64 arch/x86/entry/common.c:50 [inline]
> >>>    do_syscall_64+0x41/0xc0 arch/x86/entry/common.c:80
> >>>    entry_SYSCALL_64_after_hwframe+0x63/0xcd
> >>>
> >>> value changed: 0x0000000000000000 -> 0xfffffff000002788
> >>>
> >>> Reported by Kernel Concurrency Sanitizer on:
> >>> CPU: 0 PID: 8268 Comm: syz-executor.4 Not tainted 6.5.0-syzkaller-00453-g727dbda16b83 #0
> >>> Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 07/26/2023
> >>> ==================================================================
> >>
> >> This case is with two tasks doing bpf_map batch update together for the
> >> same map and key.
> >>     > write to 0xffffe8fffe7425d8 of 8 bytes by task 8257 on cpu 1:
> >>     > write to 0xffffe8fffe7425d8 of 8 bytes by task 8268 on cpu 0:
> >>
> >> So concurrency is introduced by user applications.
> >> In my opinion, this probably not an issue from kernel perspective.
> >
> > Perhaps not, but I recall there being a discussion about making KCSAN
> > aware of memory accesses done by BPF programs (memcpy being a tiny
> > subset of those). Not sure if the above data race qualifies as
> > something we might want to still detect, i.e. a kernel dev testing
> > their kernel might be interested in such a report.
> >
> > Regardless, in this case we should teach syzkaller to ignore KCSAN
> > data races that originate from bpf user operations whatever the
> > origin.
>
> I presume KCSAN could be silenced here via READ_ONCE/WRITE_ONCE conversion?
>
> diff --git a/include/linux/bpf.h b/include/linux/bpf.h
> index f58895830ada..32c4a37045f2 100644
> --- a/include/linux/bpf.h
> +++ b/include/linux/bpf.h
> @@ -424,8 +424,11 @@ static inline void bpf_long_memcpy(void *dst, const void *src, u32 size)
>         long *ldst = dst;
>
>         size /= sizeof(long);
> -       while (size--)
> -               *ldst++ = *lsrc++;
> +       while (size--) {
> +               WRITE_ONCE(*ldst, READ_ONCE(*lsrc));
> +               ldst++;
> +               lsrc++;
> +       }

In this case, maybe data_race(*ldst++ = *lsrc++) would be more
appropriate and efficient. Unlikely that READ_ONCE()/WRITE_ONCE() here
helps make this any safer, i.e. the memcpy is still not atomic and if
it's a bug on the user's side, it'll corrupt data either way.