Internet exposure checks and response rate limiting

Jarl works out which views are reachable from the internet (match-clients
ending in a catch-all, not private-looking addresses: an internal view of a
global IPv6 prefix is still yours; a server without views: open allow-query
and a non-private listen address), with a per-view override (schema 9), and
checks public ones: rate limiting, open resolver, minimal-any, cache
snooping, open NOTIFY and transfers, version and host name, management
interfaces. Fixes are staged like any change, including rate limiting from
ISC's example in log-only mode with the secondaries exempt, then enforcing.
named logs what rate limiting limits (or would, in log-only mode, where its
counters don't move) as query-errors; Jarl routes those lines to its log
and counts them (rrl breakdown, ratelimited top lists, Prometheus
jarl_ratelimit_logged_total), shown on the new Exposure page.

A change that newly opens a public view up must be confirmed (accept_risk,
--accept-risk, a checkbox on the Changes page; 409 confirm_risk
otherwise). Findings can be acknowledged with a reason and stay visible as
info. GET /exposure, POST /exposure/fix|ack|unack, jarl exposure,
docs/exposure.md.

Also fixes query statistics skipping TSIG-signed queries ("IP#port/key
NAME" in named's log).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

internal/engine/engine.go

@@ -15,6 +15,7 @@ import (
 	"io/fs"
 	"log/slog"
 	"net/http"
+	"net/netip"
 	"os"
 	"path/filepath"
 	"slices"
@@ -56,6 +57,10 @@ type Config struct {
 	NoReload bool
 	// SkipCheck disables named-checkconf (only for tests without BIND).
 	SkipCheck bool
+	// LocalAddrs returns this host's addresses, for working out whether a
+	// server without views is reachable from the internet
+	// (docs/exposure.md). nil: none known.
+	LocalAddrs func() []netip.Addr
 }
 
 // Engine applies changesets.
@@ -110,17 +115,20 @@ type Result struct {
 	DryRun      bool               `json:"dry_run,omitempty"`
 	Valid       bool               `json:"valid"`
 	Problems    []validate.Problem `json:"problems"` // problems this change introduces
-	Existing    []validate.Problem `json:"existing_problems,omitempty"`
-	Check       []CheckMessage     `json:"check,omitempty"`
-	CheckOK     bool               `json:"check_ok"`
-	Diff        string             `json:"diff"`
-	Files       []string           `json:"files_changed"`
-	Serials     []SerialChange     `json:"serials,omitempty"`
-	Reloaded    []string           `json:"reloaded,omitempty"`
-	Warnings    []string           `json:"warnings,omitempty"`
-	Drift       []DriftEntry       `json:"drift,omitempty"`
-	Ops         []model.Op         `json:"ops,omitempty"`
-	Error       *Error             `json:"error,omitempty"`
+	// Risks are the introduced problems that open the server up; applying
+	// needs AcceptRisk (docs/exposure.md).
+	Risks    []validate.Problem `json:"risks,omitempty"`
+	Existing []validate.Problem `json:"existing_problems,omitempty"`
+	Check    []CheckMessage     `json:"check,omitempty"`
+	CheckOK  bool               `json:"check_ok"`
+	Diff     string             `json:"diff"`
+	Files    []string           `json:"files_changed"`
+	Serials  []SerialChange     `json:"serials,omitempty"`
+	Reloaded []string           `json:"reloaded,omitempty"`
+	Warnings []string           `json:"warnings,omitempty"`
+	Drift    []DriftEntry       `json:"drift,omitempty"`
+	Ops      []model.Op         `json:"ops,omitempty"`
+	Error    *Error             `json:"error,omitempty"`
 }
 
 // stored is what is kept in a changeset's result column.
@@ -138,6 +146,9 @@ type ApplyOptions struct {
 	Force bool
 	// NoReload writes the files without reloading named.
 	NoReload bool
+	// AcceptRisk confirms a change that opens the server up, such as an
+	// open resolver on a public view (validate.RiskCodes).
+	AcceptRisk bool
 	// wipeZone lets WipeDNSSEC switch signing off for this zone at once.
 	wipeZone int64
 }
@@ -180,7 +191,46 @@ func ValidateOptions(st *model.State) validate.Options {
 }
 
 func (e *Engine) validate(st *model.State) []validate.Problem {
-	return validate.State(st, ValidateOptions(st))
+	return validate.State(st, e.validateOptions(st))
+}
+
+func (e *Engine) validateOptions(st *model.State) validate.Options {
+	opts := ValidateOptions(st)
+	if e.cfg.LocalAddrs != nil {
+		opts.LocalAddrs = e.cfg.LocalAddrs()
+	}
+	return opts
+}
+
+// Validate checks a state as applying it would.
+func (e *Engine) Validate(st *model.State) []validate.Problem { return e.validate(st) }
+
+// Exposures says which views (or the server, without views) are
+// reachable from the internet, and why (docs/exposure.md).
+func (e *Engine) Exposures(st *model.State) []validate.Exposure {
+	return validate.Exposures(st, e.validateOptions(st))
+}
+
+// risks returns the introduced problems that open a public view (or a
+// public server) up: risky warnings that weren't there before at all. One
+// that was there already, acknowledged (taking the acknowledgement back)
+// or not, opens nothing new.
+func risks(before, introduced []validate.Problem, public map[string]bool) []validate.Problem {
+	type key struct {
+		code, objType, view string
+		id                  int64
+	}
+	had := map[key]bool{}
+	for _, p := range before {
+		had[key{p.Code, p.ObjectType, p.View, p.ObjectID}] = true
+	}
+	var out []validate.Problem
+	for _, p := range introduced {
+		if p.Severity == validate.Warning && validate.RiskCodes[p.Code] && public[p.View] && !had[key{p.Code, p.ObjectType, p.View, p.ObjectID}] {
+			out = append(out, p)
+		}
+	}
+	return out
 }
 
 // ---- changeset management
@@ -389,6 +439,8 @@ type plan struct {
 	base, work    *model.State
 	introduced    []validate.Problem
 	existing      []validate.Problem
+	before        []validate.Problem // the problems before the change
+	public        map[string]bool    // public views by name ("": the server), after the change
 	serials       []SerialChange
 	out           *render.Output
 	changed       []*render.File
@@ -419,6 +471,13 @@ func (pl *plan) files() []string {
 func (e *Engine) makePlan(ctx context.Context, base, work *model.State, user string, opts ApplyOptions) (*plan, error) {
 	pl := &plan{base: base, work: work, verify: map[int64]uint32{}}
 	before := e.validate(base)
+	pl.before = before
+	pl.public = map[string]bool{}
+	for _, x := range e.Exposures(work) {
+		pl.public[x.View] = x.Public
+		// Server-wide findings (no view) count when anything is public.
+		pl.public[""] = pl.public[""] || x.Public
+	}
 	after := e.validate(work)
 	pl.introduced = append(validate.Introduced(before, after), e.dnssecChanges(base, work, opts)...)
 	introducedKeys := map[string]bool{}
@@ -652,6 +711,7 @@ func (pl *plan) result() *Result {
 		Files:    pl.files(),
 		Serials:  pl.serials,
 		Drift:    pl.drift,
+		Risks:    risks(pl.before, pl.introduced, pl.public),
 	}
 	r.Valid = r.Valid && r.CheckOK
 	if r.Files == nil {
@@ -781,6 +841,13 @@ func (e *Engine) applyOps(ctx context.Context, cs *model.Changeset, reqs []Reque
 	if opts.DryRun {
 		return res, nil
 	}
+	if rs := risks(pl.before, pl.introduced, pl.public); len(rs) > 0 && !opts.AcceptRisk {
+		msg := rs[0].Message
+		if len(rs) > 1 {
+			msg += fmt.Sprintf(" (and %d more)", len(rs)-1)
+		}
+		return res, &Error{Kind: KindRisk, Message: "this change opens the server up: " + msg + "; apply it again with accept_risk to confirm (jarl: --accept-risk), or acknowledge the finding for good", Problems: rs}
+	}
 	if len(pl.drift) > 0 && !opts.Force {
 		var paths []string
 		for _, d := range pl.drift {