Files
oikos/internal/httpapi/ratelimit.go
dtoro fa79c1ea25
Some checks failed
ci / build-test (push) Has been cancelled
ci / docker-build (push) Has been cancelled
ci / web (push) Has been cancelled
Desktop App / Build Linux (amd64) (push) Has been cancelled
Desktop App / Attach to Release (push) Has been cancelled
0.28.0 — operational hardening (plan D1–D5): CI deploy gate, versioned images, rate limiting, resource limits, health probes
D1: deploy.sh CI gate — read-only SHA via git ls-remote, Gitea commit-status
    poll, portable mkdir deploy lock (macOS, no flock), TOCTOU guard, token
    passed via curl --config - (not argv), graceful misconfig tolerance.
D2: version-tagged images — OIKOS_VERSION=v$VERSION, keep-last-3 prune derived
    from 'docker compose config --images'; VERSION read after pull.
D3: per-IP rate limiting — new internal/httpapi/ratelimit.go (x/time/rate),
    rightmost-XFF, /healthz exempt, ctx-driven sweep; disabled by default.
D4: mem_limit/cpus on all 10 compose services.
D5: staleness-aware health probes — new internal/health package wired into
    scheduler (:8093) and notifier (:8094); nomos already had :8092.

Two /review passes hardened the deploy lock, TOCTOU guard, token hygiene,
and XFF handling.
2026-08-08 21:31:16 +02:00

149 lines
4.5 KiB
Go

package httpapi
import (
"context"
"log/slog"
"net"
"net/http"
"strings"
"sync"
"time"
"golang.org/x/time/rate"
)
// rateLimiter is a per-client (IP) token-bucket limiter registry. Each unique
// client gets its own *rate.Limiter; idle entries are swept periodically so a
// flood of distinct IPs can't grow the map unbounded. A rate of zero (rps==0)
// disables limiting entirely — the returned middleware is a no-op.
//
// Client identity is the source IP. Behind Caddy the real client is in
// X-Forwarded-For: Caddy appends the immediate client as the LAST hop, while
// earlier hops are client-supplied and spoofable. clientIP therefore takes the
// rightmost XFF entry (the proxy's contribution) rather than the first.
type rateLimiter struct {
mu sync.Mutex
limiters map[string]*entry
rps rate.Limit
burst int
}
type entry struct {
limiter *rate.Limiter
lastSeen time.Time
}
// newRateLimiter builds the registry and starts the idle-entry sweeper tied to
// ctx, so the ticker is stopped when the server shuts down.
func newRateLimiter(ctx context.Context, rps, burst int) *rateLimiter {
rl := &rateLimiter{
limiters: make(map[string]*entry),
rps: rate.Limit(rps),
burst: burst,
}
if rps > 0 {
go rl.sweep(ctx)
}
return rl
}
// sweep drops entries untouched since the last sweep so the registry doesn't
// grow without bound under a rotating-IP attack or long-lived process. Exits
// (and stops its ticker) when ctx is cancelled.
func (rl *rateLimiter) sweep(ctx context.Context) {
ticker := time.NewTicker(5 * time.Minute)
defer ticker.Stop()
for {
select {
case <-ctx.Done():
return
case <-ticker.C:
rl.mu.Lock()
for ip, e := range rl.limiters {
if time.Since(e.lastSeen) > 10*time.Minute {
delete(rl.limiters, ip)
}
}
rl.mu.Unlock()
}
}
}
func (rl *rateLimiter) get(ip string) *rate.Limiter {
rl.mu.Lock()
defer rl.mu.Unlock()
if e, ok := rl.limiters[ip]; ok {
e.lastSeen = time.Now()
return e.limiter
}
l := rate.NewLimiter(rl.rps, rl.burst)
rl.limiters[ip] = &entry{limiter: l, lastSeen: time.Now()}
return l
}
// middleware returns a chi-style middleware that enforces the per-IP limit.
// Call with rps==0 to get a pass-through no-op.
func (rl *rateLimiter) middleware(next http.Handler) http.Handler {
if rl.rps <= 0 {
return next
}
return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
// Exempt infra liveness probes so Caddy/compose healthchecks can't be
// throttled into marking the service unhealthy.
if r.URL.Path == "/healthz" {
next.ServeHTTP(w, r)
return
}
if !rl.get(clientIP(r)).Allow() {
w.Header().Set("Retry-After", "1")
writeProblem(w, r, http.StatusTooManyRequests, "rate limit exceeded", "")
return
}
next.ServeHTTP(w, r)
})
}
// clientIP extracts the originating client address. It takes the rightmost
// X-Forwarded-For hop — the one the reverse proxy (Caddy) appends for the
// immediate client — because earlier hops are attacker-controlled and could
// be spoofed to dodge the limit or exhaust another client's bucket. Falls
// back to r.RemoteAddr when no XFF header is present.
//
// Known limitation: this is only trustworthy when the request actually
// traverses Caddy. A client connecting directly to the published :8090 (not
// behind the proxy) can set a single-hop XFF and have it trusted. That only
// evades rate limiting (auth is still required), and rate limiting is off by
// default, so the blast radius is narrow. Fully closing it requires either
// Caddy trusted_proxies (so it overwrites XFF / sets a non-spoofable
// X-Real-Ip) or keying the limiter on the auth token instead of IP.
func clientIP(r *http.Request) string {
if xff := r.Header.Get("X-Forwarded-For"); xff != "" {
if idx := strings.LastIndex(xff, ","); idx >= 0 {
xff = xff[idx+1:]
}
if ip := strings.TrimSpace(xff); ip != "" {
return ip
}
}
host, _, err := net.SplitHostPort(r.RemoteAddr)
if err != nil {
return r.RemoteAddr
}
return host
}
// newRateLimiterFromConfig builds the limiter from API config, logging the
// chosen policy once at startup. rps<=0 means "disabled" (returns a no-op
// middleware) so dev/single-user setups aren't throttled by default.
func newRateLimiterFromConfig(ctx context.Context, rps, burst int) *rateLimiter {
if rps <= 0 {
slog.Info("api rate limiting disabled (OIKOS_API_RATE_LIMIT unset)")
return &rateLimiter{rps: 0}
}
if burst <= 0 {
burst = rps * 2
}
slog.Info("api rate limiting enabled", "rps", rps, "burst", burst)
return newRateLimiter(ctx, rps, burst)
}