vpp-maglev/cmd/tester/main.go

// SPDX-License-Identifier: Apache-2.0

// maglevt is a tiny TUI that reads maglev.yaml, enumerates every VIP
// and hits it on a tight cadence (default 100ms) from outside the load
// balancer. HTTP/HTTPS VIPs get a HEAD request with per-VIP rolling
// latency stats, success/failure ratios, and a running tally of a
// configurable response header (default: X-IPng-Frontend) so pool-
// failover events show up as visible reshuffles in the tally. Non-HTTP
// VIPs get a plain TCP-connect probe for liveness. See maglevt --help
// for the flag surface.
package main

import (
	"context"
	"flag"
	"fmt"
	"net"
	"os"
	"regexp"
	"sort"
	"strings"
	"time"

	tea "github.com/charmbracelet/bubbletea"

	buildinfo "git.ipng.ch/ipng/vpp-maglev/cmd"
	"git.ipng.ch/ipng/vpp-maglev/internal/config"
)

func main() {
	if err := run(); err != nil {
		fmt.Fprintf(os.Stderr, "maglevt: %v\n", err)
		os.Exit(1)
	}
}

func run() error {
	var cfgPaths multiFlag
	flag.Var(&cfgPaths, "config", "path to maglev.yaml (repeatable; also accepts a comma-separated list). Frontends are unioned across files, deduplicated by (address, protocol, port).")
	interval := flag.Duration("interval", 100*time.Millisecond, "probe interval per VIP (±10% jitter)")
	timeout := flag.Duration("timeout", 2*time.Second, "per-request timeout")
	host := flag.String("host", "", "Host header override (default: VIP address literal)")
	// Default probe URI: a small, deliberate health-check path that
	// typically returns 204 No Content and doesn't hit the backend
	// app logs. /.well-known/ipng/healthz is the convention for
	// IPng deployments; override with --uri for anything else.
	// --path is registered as a synonym for backward compatibility
	// with the pre-1.0 flag name — both set the same variable, so
	// whichever the operator types last on the command line wins.
	const defaultURI = "/.well-known/ipng/healthz"
	path := flag.String("uri", defaultURI, "HTTP request path (URI) used in the GET request")
	flag.StringVar(path, "path", defaultURI, "alias for --uri")
	header := flag.String("header", "X-IPng-Frontend", "response header to extract and tally")
	insecure := flag.Bool("insecure", true, "skip TLS verification for HTTPS")
	keepalive := flag.Bool("keepalive", false, "enable HTTP keep-alives (disabled by default so each probe opens a fresh connection — required for failover visibility)")
	flag.BoolVar(keepalive, "k", false, "shorthand for --keepalive")
	filter := flag.String("filter", "", "only probe frontends whose name matches this regex")
	printVersion := flag.Bool("version", false, "print version and exit")
	flag.Parse()

	if *printVersion {
		fmt.Printf("maglevt %s (commit %s, built %s)\n",
			buildinfo.Version(), buildinfo.Commit(), buildinfo.Date())
		return nil
	}

	if len(cfgPaths) == 0 {
		cfgPaths = multiFlag{"/etc/vpp-maglev/maglev.yaml"}
	}

	// Load every requested config. A parse/semantic error on any of
	// them is fatal — we want the user to see it rather than silently
	// probing a reduced set of VIPs because one file was broken.
	configs := make([]*config.Config, 0, len(cfgPaths))
	for _, p := range cfgPaths {
		cfg, res := config.Check(p)
		if !res.OK() {
			if res.ParseError != "" {
				return fmt.Errorf("config parse %s: %s", p, res.ParseError)
			}
			return fmt.Errorf("config semantic %s: %s", p, res.SemanticError)
		}
		configs = append(configs, cfg)
	}

	var filterRe *regexp.Regexp
	if *filter != "" {
		var err error
		filterRe, err = regexp.Compile(*filter)
		if err != nil {
			return fmt.Errorf("invalid --filter regex: %w", err)
		}
	}

	opts := probeOpts{
		Interval:  *interval,
		Timeout:   *timeout,
		Host:      *host,
		Path:      *path,
		Header:    *header,
		Insecure:  *insecure,
		KeepAlive: *keepalive,
	}

	vips := buildVIPsUnion(configs, cfgPaths, filterRe, opts)
	if len(vips) == 0 {
		return fmt.Errorf("no matching frontends in %s", strings.Join(cfgPaths, ", "))
	}

	m := Model{
		host:    fqdn(),
		vips:    vips,
		opts:    opts,
		startAt: time.Now(),
		showDNS: true,
	}

	ctx, cancel := context.WithCancel(context.Background())
	defer cancel()

	prog := tea.NewProgram(m, tea.WithAltScreen())

	// Spawn one probe goroutine per VIP. Each sends probeResultMsg
	// into the tea.Program via prog.Send, which is thread-safe.
	// Alongside the prober we kick off a one-shot reverse-DNS
	// lookup so the 'd' toggle has a hostname to display; the
	// lookup is best-effort and simply drops on timeout or NXDOMAIN.
	for _, v := range vips {
		go runProbeLoop(ctx, v.info, opts, prog.Send)
		go runDNSLookup(ctx, v.info, prog.Send)
	}

	_, err := prog.Run()
	cancel()
	// Give the workers a beat to observe ctx and exit. This isn't
	// strictly required — the process is exiting anyway — but a clean
	// shutdown avoids the "unexpected EOF writing to a closed
	// transport" spam that HTTP clients sometimes emit on ctrl-C.
	time.Sleep(50 * time.Millisecond)
	return err
}

// buildVIPsUnion flattens frontends from multiple configs into a
// single deduplicated probe list, keyed by the (scheme, address,
// port) tuple. The typical use case is a pair of maglevds fronting
// the same two VIPs (vip0 / vip1, IPv4 + IPv6, × port 80 + 443 = 8
// probers) — the operator passes both yaml files and maglevt unions
// them so the probe grid doesn't grow duplicates from mirrored
// configs. The symbolic frontend name from yaml is intentionally
// dropped: when two files use the same name for different tuples
// (common in cross-deployment comparisons) the name would be
// ambiguous, and the tuple is the only stable identity. Only the
// --filter regex still uses the name, as a pre-dedup match.
//
// Dedup key uses net.IP.String() which canonicalises IPv6 zero-
// compression, so 2001:db8::1 and 2001:db8:0:0:0:0:0:1 collapse
// onto one entry. Iteration order across files is stable for
// deterministic TUI layout: within a file, frontends are visited
// in name-sorted order; across files, the first occurrence of each
// tuple wins and fixes its slot in the output.
// fqdn returns the system's fully-qualified hostname: gethostname(2)
// via os.Hostname() for the short name, then a CNAME lookup to reach
// the canonical form the resolver would hand back — the same two-step
// dance `hostname -f` performs. Falls back to the short name when the
// resolver has nothing to add, so the TUI header always renders.
func fqdn() string {
	h, err := os.Hostname()
	if err != nil {
		return "unknown"
	}
	if cname, err := net.LookupCNAME(h); err == nil {
		if s := strings.TrimSuffix(cname, "."); s != "" {
			return s
		}
	}
	return h
}

func buildVIPsUnion(cfgs []*config.Config, cfgPaths []string, filterRe *regexp.Regexp, opts probeOpts) []*vipState {
	_ = cfgPaths // reserved for future diagnostics (e.g. which file this tuple came from)
	type key struct {
		ip     string
		scheme string
		port   uint16
	}
	seen := map[key]*vipState{}
	var order []key

	for _, cfg := range cfgs {
		names := make([]string, 0, len(cfg.Frontends))
		for name := range cfg.Frontends {
			names = append(names, name)
		}
		sortStringsInPlace(names)

		for _, name := range names {
			fe := cfg.Frontends[name]
			if filterRe != nil && !filterRe.MatchString(name) {
				continue
			}
			if strings.ToLower(fe.Protocol) != "tcp" || fe.Port == 0 {
				continue
			}
			scheme := schemeForPort(fe.Port)
			k := key{ip: fe.Address.String(), scheme: scheme, port: fe.Port}
			if _, ok := seen[k]; ok {
				continue // already claimed by an earlier file
			}
			info := &vipInfo{
				idx:    len(order),
				scheme: scheme,
				ip:     fe.Address,
				port:   fe.Port,
			}
			if scheme == "http" || scheme == "https" {
				info.url = buildURL(scheme, fe.Address, fe.Port, opts.Path)
				info.client = newHTTPClient(opts)
			}
			v := &vipState{
				info:     info,
				rolling:  newRolling(),
				tally:    map[string]int{},
				tallyOld: map[string]int{},
				tallyNew: map[string]int{},
			}
			seen[k] = v
			order = append(order, k)
		}
	}

	out := make([]*vipState, len(order))
	for i, k := range order {
		out[i] = seen[k]
	}
	// Display order: IPv6 before IPv4, higher ports before lower
	// within each address family, then address string as a final
	// tiebreaker for determinism across runs. HTTPS :443 sitting
	// above HTTP :80 matches the "secure first" reading order most
	// operators expect, and clustering all the IPv6 rows at the top
	// keeps a mixed-family deployment visually coherent as the
	// operator scans down the table.
	sort.SliceStable(out, func(i, j int) bool {
		vi, vj := out[i].info, out[j].info
		iIs6 := vi.ip.To4() == nil
		jIs6 := vj.ip.To4() == nil
		if iIs6 != jIs6 {
			return iIs6
		}
		if vi.port != vj.port {
			return vi.port > vj.port
		}
		return vi.ip.String() < vj.ip.String()
	})
	// Re-index after the sort so info.idx matches the slot each VIP
	// now occupies in out — probeResultMsg.VIPIdx is looked up via
	// this index in Model.Update, so they must agree.
	for i, v := range out {
		v.info.idx = i
	}
	return out
}

// multiFlag is a flag.Value implementation that accumulates repeated
// --config occurrences into a slice, and also accepts comma-separated
// values on a single flag instance so `--config a.yaml,b.yaml` and
// `--config a.yaml --config b.yaml` produce the same result.
type multiFlag []string

func (m *multiFlag) String() string {
	return strings.Join(*m, ",")
}

func (m *multiFlag) Set(v string) error {
	for _, p := range strings.Split(v, ",") {
		p = strings.TrimSpace(p)
		if p != "" {
			*m = append(*m, p)
		}
	}
	return nil
}

// schemeForPort infers HTTP vs HTTPS from the VIP's TCP port, falling
// back to "tcp" (connect-only probe) for anything that isn't
// unambiguously web. Intentionally narrow — we'd rather under-classify
// than send HEAD / at an IMAPS VIP and spew protocol errors into the
// logs. Adding more here is fine later (e.g. 8080/8443) but defaults
// should stay conservative.
func schemeForPort(port uint16) string {
	switch port {
	case 80:
		return "http"
	case 443:
		return "https"
	}
	return "tcp"
}

// buildURL constructs the probe URL for an HTTP/HTTPS VIP. IPv6
// literals are bracketed per RFC 3986 §3.2.2 so the colon in the
// address isn't confused with the port separator.
func buildURL(scheme string, ip net.IP, port uint16, path string) string {
	host := ip.String()
	if ip.To4() == nil {
		host = "[" + host + "]"
	}
	if path == "" {
		path = "/"
	}
	return fmt.Sprintf("%s://%s:%d%s", scheme, host, port, path)
}

// sortStringsInPlace is a tiny shim so we don't import "sort" just
// for a single call from buildVIPs. The sorted-names slice is at
// most a few dozen elements so an insertion sort is fine and avoids
// the import churn.
func sortStringsInPlace(s []string) {
	for i := 1; i < len(s); i++ {
		for j := i; j > 0 && s[j-1] > s[j]; j-- {
			s[j-1], s[j] = s[j], s[j-1]
		}
	}
}