net.ipv6.ioam6_id_wide stores a u64 (netns_sysctl_ipv6.ioam6_id_wide), but its table entry uses proc_doulongvec_minmax() with .maxlen = sizeof(u64) and .extra2 pointing at a u64. On 32-bit, unsigned long is 4 bytes, so the handler treats the field as a vector of two 32-bit values and reads the u64 max through an unsigned long pointer: - reading prints two numbers ("4294967295 16777215" by default on little-endian); - writing one value only replaces one half of the u64: on little-endian "echo 5 >" leaves 0x00ffffff00000005; on big-endian it sets the upper half, and the range check uses the upper word of ioam6_id_wide_max (0x00ffffff). The wide namespace ID emitted in IOAM trace data is then not the configured value. Same class of issue as commit 162cf592b577 ("net/smc: fix limit_smc_hs sysctl on big-endian architecture"). There is no u64 variant of proc_doulongvec_minmax(), so use a small local handler that goes through proc_dostring() and kstrtou64(). A generic u64 helper in kernel/sysctl.c would be a larger change for net and stable; it can follow in net-next. For normal input the handler behaves as the 64-bit one does today: base 0 parsing (decimal, hex or octal), surrounding white space is accepted, values above IOAM6_DEFAULT_ID_WIDE return -EINVAL, and a write of only "\n" returns -EINVAL. A continued write at a non-zero offset in strict mode is ignored, as before. Input too long for the 24-byte buffer is rejected rather than truncated by proc_dostring(). One change on 64-bit: a write of two numbers such as "5 6" is now rejected with -EINVAL, instead of being a partial write that sets the value to 5. The bug was discovered by an AI coding agent (Claude Code) reading the code. It was confirmed and the fix tested on i386 and x86_64 kernels under QEMU, with a small static test program as init: before the fix, i386 reads "4294967295 16777215", and after "echo 5" it reads "5 16777215". With the fix, i386 reads and writes a single 64-bit value and gives the same results as x86_64, and on x86_64 reads, writes and range checks behave as before. No kernel change was needed to reproduce it. Fixes: 9ee11f0fff20 ("ipv6: ioam: Data plane support for Pre-allocated Trace") Signed-off-by: Shashank Mohan Jain --- v2: - Keep -EINVAL for a write of only "\n", as before; a continued write at a non-zero offset in strict mode is still ignored (Justin Iurman). - Changelog: say that on 64-bit "5 6" is now rejected with -EINVAL instead of being a partial write that sets 5 (Justin Iurman). - Changelog: describe how the bug was found and how the fix was tested in plain English, instead of an Assisted-by tag. - Cc the sysctl maintainers. v1: https://lore.kernel.org/netdev/20261118042958.83243-1-jain.sm@gmail.com/ Prepared with Claude Code (Anthropic), model Claude Opus 5.5 (claude-opus-5-5). Tested on net.git main d5b89d55a322: x86_64 and i386 defconfig builds, plus a build-only 32-bit big-endian powerpc (pmac32_defconfig). W=1 shows no warnings for sysctl_net_ipv6.o before or after. Under QEMU (TCG) with a small static test program as init, i386 now gives the same results as x86_64. On x86_64 vs unpatched: a lone "\n" still returns -EINVAL, a continued write at a non-zero offset with sysctl_writes_strict=1 is still ignored and succeeds, "5 6" now returns -EINVAL instead of setting 5, and 21 to 23 digit input is now accepted. Other changes seen: a lone "\n" at a non-zero offset with sysctl_writes_strict=0 is ignored instead of returning -EINVAL, and a read in small chunks returns the whole value. The big-endian kernel was built only, not run. I did not add Cc: stable; please add it if you think it fits. net/ipv6/sysctl_net_ipv6.c | 45 ++++++++++++++++++++++++++++++++++++-- 1 file changed, 43 insertions(+), 2 deletions(-) diff --git a/net/ipv6/sysctl_net_ipv6.c b/net/ipv6/sysctl_net_ipv6.c index 1a0a36dcda..61c28df64d 100644 --- a/net/ipv6/sysctl_net_ipv6.c +++ b/net/ipv6/sysctl_net_ipv6.c @@ -30,6 +30,48 @@ static u32 rt6_multipath_hash_fields_all_mask = static u32 ioam6_id_max = IOAM6_DEFAULT_ID; static u64 ioam6_id_wide_max = IOAM6_DEFAULT_ID_WIDE; +/* unsigned long is 32-bit on some arches, so handle the u64 here */ +static int proc_ioam6_id_wide(const struct ctl_table *table, int write, + void *buffer, size_t *lenp, loff_t *ppos) +{ + u64 *valp = table->data; + char buf[24] = ""; + const struct ctl_table tmp = { + .data = buf, + .maxlen = sizeof(buf), + }; + loff_t pos = *ppos; + u64 val; + int ret; + + if (!write) { + snprintf(buf, sizeof(buf), "%llu", READ_ONCE(*valp)); + return proc_dostring(&tmp, write, buffer, lenp, ppos); + } + + /* an empty write changes nothing, as with proc_doulongvec_minmax() */ + if (!*lenp) + return 0; + + /* proc_dostring() would silently truncate a longer input */ + if (*lenp >= sizeof(buf)) + return -EINVAL; + + ret = proc_dostring(&tmp, write, buffer, lenp, ppos); + if (ret) + return ret; + + /* a lone "\n" is invalid, a continued write in strict mode is ignored */ + if (!buf[0]) + return pos ? 0 : -EINVAL; + + if (kstrtou64(strim(buf), 0, &val) || val > ioam6_id_wide_max) + return -EINVAL; + + WRITE_ONCE(*valp, val); + return 0; +} + static int proc_rt6_multipath_hash_policy(const struct ctl_table *table, int write, void *buffer, size_t *lenp, loff_t *ppos) { @@ -210,8 +252,7 @@ static const struct ctl_table ipv6_table_template[] = { .data = &init_net.ipv6.sysctl.ioam6_id_wide, .maxlen = sizeof(u64), .mode = 0644, - .proc_handler = proc_doulongvec_minmax, - .extra2 = &ioam6_id_wide_max, + .proc_handler = proc_ioam6_id_wide, }, }; -- 2.54.0 (Apple Git-157)