fix: count every PHP-FPM stage, stop panicking on a bad URI, report scrape failures

[dimoschi]

Why this PR exists

You asked how so many bugs went unnoticed. Rather than keep surfacing them one at a time, this is the result of a deliberate sweep of phpfpm/, phpfpm/fcgi/, cmd/, internal/server/ and main.go.

Six defects. Five are inherited from upstream and untouched since the fork, and two have open upstream issues that have gone unexplained for years.

Fixes

1. Process accounting — hipages#322, hipages#419

CountProcessState put Finishing, Ending and Getting request information in empty switch cases. Go has no implicit fallthrough, so those processes counted as neither active nor idle and vanished from the total:

6 processes in: active=2 idle=1 total=3   (3 unaccounted for)

PHP-FPM treats a child as idle only while it is accepting and everything else as active. README.md already claimed this flag "will emulate PHP-FPMs implementation including the accumulation of multiple states" — the code simply did not do what its own docs said.

This is consistent with upstream hipages#322, "Why phpfpm_total_processes is a dynamically changing value", open since 2023 with no explanation: under load, processes cycle through the dropped stages and the total sags toward the active count. I can't confirm that reporter used --phpfpm.fix-process-count, so treat it as a strong match rather than a proven one.

The Creating stage was missing entirely — it is a documented stage in php-src sapi/fpm/fpm/fpm_request.c (verified against the requests_stages array), and it is the Unknown process state 'Creating' log spam in upstream hipages#419.

An unrecognised stage is still logged, but now counts as active, so a future PHP release cannot silently shrink the total again.

2. Panic on a malformed scrape URI

url.Parse returns a nil *url.URL alongside its error. parseURL dereferenced it in the error branch:

uri, err := url.Parse(rawurl)
if err != nil {
    return uri.Scheme, uri.Host, uri.Path, err   // uri is nil
}

A single bad --phpfpm.scrape-uri killed the process, and because it runs inside the scrape goroutine there was nothing to recover it. Introduced 2018-03-18 in upstream #23. Not reported upstream.

3. get exited 0 when every scrape failed

PoolManager.Update unconditionally returned nil, so get's log.Fatal("Could not update pool") was unreachable. It now returns the joined per-pool errors; get prints whatever it collected, then exits non-zero. Pool.error already logs each failure, so the redundant second log in the update goroutine is gone.

4. An invalid --log.level killed the process

log.Fatalf with a dead lvl = logrus.InfoLevel assignment directly above it, showing the intent was to fall back. An unusable log level is no reason to take a running exporter's metrics away.

5. Nil logger panic

The package logger was nil until SetLogger ran, so the first unrecognised process state panicked for library callers. It now defaults to a discarding logger, and SetLogger ignores nil.

6. initialPodEnlisting returned an always-nil error

Removed. This is the last outstanding unparam finding.

Behaviour changes

• phpfpm_{active,idle,total}_processes values change under --phpfpm.fix-process-count. Totals go up, because processes that were being dropped are now counted. This is the point of the fix, but it will move graphs.
• get now exits 1 if any target failed. Scripts that ignored the exit code are unaffected; scripts that checked it now get the truth.

Tests

Every fix has a test written failing first. Notably TestCountProcessStateCountsEveryReportedProcess asserts active + idle == total == len(processes) across all eight stages, which is the invariant that was broken.

Verification

$ mise exec -- go test ./... -race
ok  	github.com/hackthebox/php-fpm_exporter/internal/server
ok  	github.com/hackthebox/php-fpm_exporter/phpfpm
ok  	github.com/hackthebox/php-fpm_exporter/phpfpm/fcgi

Confirmed against the built binary: the malformed URI now logs an error instead of panicking, an invalid log level falls back, and get returns exit 1 on scrape failure and 0 otherwise.

golangci-lint (v2.12.2, as CI pins): 5 issues on master → 4 here, none new.

Known gap

The 8 remaining mutation survivors are all in cmd/ — the exit-code branch and the log-level fallback message — and are accepted on the same basis as the earlier batch: cmd is wiring and is untested by design. Moving get's body into internal/ is tracked in #25.