Open and own the required ICMP sockets before permanently dropping
setuid root, then parse input without elevated identity. Close-on-exec
socket ownership, shared IPv4/IPv6 dispatch and fault-injected
identity/capability calls are covered. The owned fixture tests check
deny/allow behavior without contacting external devices.
Step 2/14 of the complete C17/native-correctness series. Base: the
preceding series branch; merge/review in order so this PR shows only its
own step.
Earlier validation, before this review follow-up, in isolated copied
snapshots on Linux arm64:
- Clean GCC build and make check: 13 passed, 1 existing prerequisite
skips, 0 failures.
- ASan/UBSan: 12 passed, 2 existing privilege/allocator-interposition
prerequisite skips, 0 failures; no sanitizer findings.
- Production source/profile integrity verification against disposable
MariaDB and loopback SNMP fixtures passed.
- Release archive extracted into a separate source tree and built/tested
out of tree; fuzz checks passed.
Earlier local profiles retain source hashes for the copied validation
tree; GitHub checks validate the published commit. No production
database/device, host installation or live privilege change is part of
validation. Existing prerequisite skips were retained; no test or
finding was suppressed to obtain a pass.
Series navigation: [Series
index](https://github.com/Cacti/spine/pull/638); [previous
step](https://github.com/Cacti/spine/pull/638). [Next
step](https://github.com/Cacti/spine/pull/640).
Review follow-up (2026-10-09): Restacked on the corrected preceding
branches. The SNMP status fix is introduced in #638, the
reindex/completion fixes in #640, and focused regression coverage in
#650.
Restacked tip validation: isolated Linux arm64 build and make check
passed (18 passed, 1 setuid prerequisite skip, no failures).
GitHub checks validate the updated published commits; earlier
profile/archive/sanitizer evidence above belongs to the previous
revisions.
---------
Signed-off-by: Thomas Vincent <thomasvincent@gmail.com>
## Summary
- make subprocess-registry detach atomic so exactly one closer owns each
entry
- make cancellation cleanup release only the detached node it owns
- replace `vfork`/`execve` child setup with `posix_spawn` and explicit
file actions
- mark every pipe and stdio-collision duplicate close-on-exec
- restore the child's default SIGPIPE disposition while preserving the
parent's handling
- bound TERM/KILL reaping and park unreaped children in a fixed-capacity
sweep list
- harden PHP script-server slot ownership, startup, restart, signal,
descriptor, and cleanup paths
- retire and restart failed or oversized reads so stale response tails
cannot poison later data sources
- revalidate every command retry and cap recovery handshakes at two
seconds
This folds the child-process lifecycle portion of #597 into the narrower
ownership PR. It does not absorb unrelated poll-host, SQL, SNMP, ICMP,
or build-infrastructure work.
## Review-driven corrections
- close-on-exec is explicitly restored after `dup()` in both read and
write stdio-collision branches
- the source invariant check now uses portable POSIX character classes
- direct tests cover duplicate flags, both collision branches, failure
after pipe creation, abandoned-child sweeping, and capacity overflow
- collision cleanup preserves current-child redirects even when an older
registry entry occupies fd 0 or 1
- script children receive SIGTERM before SIGKILL and unreaped PHP
children transfer to the bounded sweep registry
- the Docker integration suite now starts a real PHP script server,
verifies its startup protocol and database result, and checks that
shutdown leaves no zombie PHP process
## Validation
- Ubuntu 24.04 container with GCC and GNU ld: `make check`, 6/6 targets
passed with no skips, including `test_php_runtime`
- real PHP script-server Docker integration: 6/6 checks passed; exact
value `42` reached `poller_output` and no zombie remained
- macOS Homebrew dependency build passed
- macOS `make check`: 5/5 available binaries passed; GNU `--wrap`
runtime suite is Linux-only
- PHP fixture syntax checked through mise PHP 8.3
- actionlint, shellcheck, and `git diff --check` passed
Closes#610.
---------
Signed-off-by: Thomas Vincent <thomasvincent@gmail.com>
Co-authored-by: TheWitness <thewitness@cacti.net>
Poller correctness and memory-safety fixes, with the ICMP hardening from
#543 folded in.
Verified in a clean `ubuntu:24.04` container: builds at **38 warnings**
against develop's 39, and the unit tests build and pass on Linux (39
assertions across three binaries).
**Fixes**
- `php_readpipe` off-by-one — the old bound let `bptr` reach
`result_string + 2048` and then wrote the terminator one byte past a
2048-byte allocation.
- `get_namebyhost` used `strtok`, which keeps one process-wide save
pointer, and runs unlocked on every poller thread. Two devices could be
polled at each other's addresses. Now `strtok_r`.
- The log date format was read from `default_dateformat`; Cacti's key is
`default_date_format`, so the lookup never matched and the format was
never loaded.
- `db_query` returns `mysql_store_result()` directly, which is NULL on
store failure. The NULL check ran *after* `mysql_fetch_row`, so the
crash happened one line before the guard.
- `snmp_count` logged a Net-SNMP `errstat` without setting either loop
flag, rebuilding the identical GETNEXT forever and pinning a thread
until the poll interval ended.
- `php_close` skipped the `close()` when the quit write failed with
EPIPE, leaking a descriptor per script-server restart.
- The reap loop's delay was compiled out on `SOLAR_THREAD` builds, so
both phases burned twenty `WNOHANG` polls in nanoseconds and the child
was never reaped.
- ICMP replies are validated for length and matched on id/sequence; on
Cygwin the peeked datagram is now drained, which previously busy-spun
until host timeout.
- The fatal signal handler no longer calls
`malloc`/`localtime_r`/`fprintf` before `_exit` — on a SIGSEGV caused by
heap corruption that deadlocked on the allocator's lock instead of
dying.
- The unit test Makefile was hardcoded to `/opt/homebrew` paths and
`-std=gnu23`, so it could not build on Linux at all. Now pkg-config
driven.
Supersedes #543.
**Added since the first revision**
- `snmp_get_multi()` now checks the varbind type before formatting.
Under v2c an agent reports a per-OID failure as an inline exception
varbind while the PDU `errstat` stays `NOERROR`, so `snprint_value()`
rendered the exception as literal text and stored it as a value.
- Two warning sources removed: a dangling line continuation ending
`sql.h`, and `strncopy()` now using `memcpy` where the byte count is
already exact. Warnings drop from **39 to 7**.
Closes#551Closes#232
- `read_config_options()` issued 23 separate `SELECT ... FROM settings
WHERE name = ?` round trips at startup; they are now one bulk read
served from memory. Call sites are unchanged.
Closes#330
**Test infrastructure**
- The SNMPv3 integration fixture was broken: the image shipped a binary
that could not start (`libmysqlclient.so.21: cannot open shared object
file`). The build context carried `*.o`, `spine`, `Makefile` and
`config.status`, so `make` treated them as current, skipped compiling,
and packaged a binary linked against libraries the runtime image does
not have. `.dockerignore` now excludes build output. The fixture runs
end to end: spine polls the SNMPv3 device, reads `uptime`, writes to
`poller_output` and exits 0.
- Two libFuzzer targets, run per PR and weekly: the host string parser
and the ICMP reply parser. The second drives
`spine_icmp_classify_reply()`, extracted from `ping_icmp` so the bounds
walk over a shared raw socket is reachable by a test. 15M executions
under ASan/UBSan, no crashes.
**Added in this revision**
- `output_regex` was applied only in the in-loop OID drain, not the pass
after the loop that handles whatever did not fill a `MAX_OIDS` batch, so
a device whose OID count missed a boundary stored raw values. A single
data source is always the leftover case.
`tests/integration/test_output_regex.sh` now asserts on the value: it
fails on the parent commit with the raw reading and passes here with the
filtered one. Addresses #210; the feature stays dormant until Cacti
populates `poller_item.output_regex`.
Merge before #555, which needs the `get_namebyhost` fix above to resolve
a device hostname at all.
---------
Signed-off-by: Thomas Vincent <thomasvincent@gmail.com>
QA: Oops removing snmp_init broke snmpv3
Feature: don't complain on wrong snmp types, just roll with it
git-svn-id: svn+ssh://repo.cacti.net/var/data/svnroot/cacti/spine/branches/0.8.7@4551 860744bd-22fc-0310-8c96-e9fe5004b5ca
2) -feature: add Net-SNMP Library vs. Header check prior to execution.
3) More clearly indicated script servers being shut down during debug
4) Optimize some of the latest code based upon what would appear to be common sense.
git-svn-id: svn+ssh://repo.cacti.net/var/data/svnroot/cacti/branches/BRANCH_0_8_6/cactid@3567 860744bd-22fc-0310-8c96-e9fe5004b5ca
Thomas Vincent