Bug 2129713

From a capture shared offline: router advertisements are sent with an incorrect checksum, whose value is 58 (decimal, same as ICMPv6 protocol number) plus the expected checksum. It's as if in some builds, this snippet from ndp.c:

	ip6hr->hop_limit = IPPROTO_ICMPV6;
	ihr->icmp6_cksum = 0;
	ihr->icmp6_cksum = csum_unaligned(ip6hr, sizeof(*ip6hr) +
						 sizeof(*ihr) + len, 0);

	ip6hr->version = 6;
	ip6hr->nexthdr = IPPROTO_ICMPV6;
	ip6hr->hop_limit = 255;

where, for convenience, hop_limit is first set to IPPROTO_ICMPV6 to match the IPv6 pseudo-header for ICMPv6 checksum, and later set to its intended value, happened to be equivalent to:

	ihr->icmp6_cksum = 0;
	ihr->icmp6_cksum = csum_unaligned(ip6hr, sizeof(*ip6hr) +
						 sizeof(*ihr) + len, 0);

	ip6hr->version = 6;
	ip6hr->nexthdr = IPPROTO_ICMPV6;
	ip6hr->hop_limit = 255;

At a first glance I don't see any justification why the compiler would be allowed to elide the initial assignment of hop_limit, though.

From a second (still quick) look:

- the gcc version used in passt-0.git.2022_06_08.d7d467f-0.el9.x86_64 is 11.2.1-9.4.el9 (https://download.copr.fedorainfracloud.org/results/sbrivio/passt/epel-9-x86_64/04776284-passt/builder-live.log.gz) -- this is the same as the one used for the most recent EPEL 9 build, 0^20220924.g8978f65-1.el9.x86_64

- comparing that part of the NDP implementation between the most recent EPEL 9 build and the most recent Fedora 36 build (using gcc 12.2.1-2.fc36.x86_64), it looks like there are some notable differences -- something makes me think that the hop_limit store is actually missing in the EPEL 9 build, but I couldn't grasp enough of it, yet

- if that store is really missing, this would be similar in nature to the issue described at https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101354, which however applies only to functions marked as "naked", not the case here

It might be worth to check for differences in intermediate files (passing -save-temps in CFLAGS) between the two gcc versions.

Weird:

$ gcc --version
gcc (GCC) 11.2.1 20220127 (Red Hat 11.2.1-9)
Copyright (C) 2021 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

$ (make clean; CFLAGS='-g' make) >/dev/null
$ objdump -DSdslrx passt|grep -A15 "ndp.c\:192" -m1
/root/passt_gcc_debug/passt/ndp.c:192
	ihr->icmp6_cksum = 0;
    b8bc:	31 c9                	xor    %ecx,%ecx
/root/passt_gcc_debug/passt/ndp.c:193
	ihr->icmp6_cksum = csum_unaligned(ip6hr, sizeof(*ip6hr) +
    b8be:	31 d2                	xor    %edx,%edx
/root/passt_gcc_debug/passt/ndp.c:191
	ip6hr->hop_limit = IPPROTO_ICMPV6;
    b8c0:	c6 44 24 35 3a       	movb   $0x3a,0x35(%rsp)
__bswap_16():
/usr/include/bits/byteswap.h:37
  return __builtin_bswap16 (__bsx);
    b8c5:	66 c1 c0 08          	rol    $0x8,%ax
ndp():
/root/passt_gcc_debug/passt/ndp.c:193
	ihr->icmp6_cksum = csum_unaligned(ip6hr, sizeof(*ip6hr) +

Here, 0x3a (IPPROTO_ICMPV6) is stored before the call to csum_unaligned(). But not if I build with -flto=auto (that's a default flag for at least EPEL 9 packages):

$ (make clean; CFLAGS='-g -flto=auto' make) >/dev/null
$ objdump -DSdslrx passt|grep -A15 "ndp.c\:192" -m1
/root/passt_gcc_debug/passt/ndp.c:192
	ihr->icmp6_cksum = 0;
    bb0c:	45 31 c0             	xor    %r8d,%r8d
__bswap_16():
/usr/include/bits/byteswap.h:37
  return __builtin_bswap16 (__bsx);
    bb0f:	66 c1 c0 08          	rol    $0x8,%ax
ndp():
/root/passt_gcc_debug/passt/ndp.c:192
    bb13:	66 44 89 44 24 58    	mov    %r8w,0x58(%rsp)
/root/passt_gcc_debug/passt/ndp.c:190
	ip6hr->payload_len = htons(sizeof(*ihr) + len);
    bb19:	66 89 44 24 32       	mov    %ax,0x32(%rsp)
/root/passt_gcc_debug/passt/ndp.c:193
	ihr->icmp6_cksum = csum_unaligned(ip6hr, sizeof(*ip6hr) +
    bb1e:	48 8d 46 28          	lea    0x28(%rsi),%rax

I also tested on a recent gcc 12.2.0, same behaviour.

I introduced a workaround, that is, declaring csum_unaligned() as "noipa" for the affected gcc versions, depending on CFLAGS:
  https://passt.top/passt/commit/?id=06aa26fcf398f5d19ab46e42996190d7f95e837a

and it's now available in the 0^20220929.g06aa26f-1 EPEL 9 build (Copr repository only at the moment).

Verified with passt-0^20221104.ge308018-1.el9.x86_64, it's passed and can get IPV6 gateway on the VM

on host: 
[root@dell-per440-18 ~]# rpm -qa |grep passt
passt-0^20221104.ge308018-1.el9.x86_64
[root@dell-per440-18 ~]# gcc --version
gcc (GCC) 11.2.1 20220127 (Red Hat 11.2.1-9)

on VM:
[root@dell-per440-18 ~]# dhclient -6 eth0
grep: /etc/sysconfig/network-scripts/ifcfg-*: No such file or directory
grep: /etc/sysconfig/network-scripts/ifcfg-*: No such file or directory
grep: /etc/sysconfig/network-scripts/ifcfg-*: No such file or directory
grep: /etc/sysconfig/network-scripts/ifcfg-*: No such file or directory
grep: /etc/sysconfig/network-scripts/ifcfg-*: No such file or directory
grep: /etc/sysconfig/network-scripts/ifcfg-*: No such file or directory
[root@dell-per440-18 ~]# ip -j -6 ro sh|jq -rM '.[] | select(.dst == "default").gateway'
fe80::cee1:9402:8b35:be41

Set it to Veirified.