mox/webmail/webmail.go

// Package webmail implements a webmail client, serving html/js and providing an API for message actions and SSE endpoint for receiving real-time updates.
package webmail

// todo: should we be serving the messages/parts on a separate (sub)domain for user-content? to limit damage if the csp rules aren't enough.

import (
	"archive/zip"
	"bytes"
	"compress/gzip"
	"context"
	"encoding/base64"
	"encoding/json"
	"errors"
	"fmt"
	"io"
	"mime"
	"net/http"
	"os"
	"path/filepath"
	"regexp"
	"runtime/debug"
	"strconv"
	"strings"
	"sync"
	"time"

	_ "embed"

	"golang.org/x/exp/slog"
	"golang.org/x/net/html"

	"github.com/prometheus/client_golang/prometheus"
	"github.com/prometheus/client_golang/prometheus/promauto"

	"github.com/mjl-/bstore"
	"github.com/mjl-/sherpa"

	"github.com/mjl-/mox/message"
	"github.com/mjl-/mox/metrics"
	"github.com/mjl-/mox/mlog"
	"github.com/mjl-/mox/mox-"
	"github.com/mjl-/mox/moxio"
	"github.com/mjl-/mox/moxvar"
	"github.com/mjl-/mox/store"
	"github.com/mjl-/mox/webaccount"
)

func init() {
	mox.LimitersInit()
}

var pkglog = mlog.New("webmail", nil)

// We pass the request to the sherpa handler so the TLS info can be used for
// the Received header in submitted messages. Most API calls need just the
// account name.
type ctxKey string

var requestInfoCtxKey ctxKey = "requestInfo"

type requestInfo struct {
	LoginAddress string
	AccountName  string
	Request      *http.Request // For Proto and TLS connection state during message submit.
}

//go:embed webmail.html
var webmailHTML []byte

//go:embed webmail.js
var webmailJS []byte

//go:embed msg.html
var webmailmsgHTML []byte

//go:embed msg.js
var webmailmsgJS []byte

//go:embed text.html
var webmailtextHTML []byte

//go:embed text.js
var webmailtextJS []byte

var (
	// Similar between ../webmail/webmail.go:/metricSubmission and ../smtpserver/server.go:/metricSubmission
	metricSubmission = promauto.NewCounterVec(
		prometheus.CounterOpts{
			Name: "mox_webmail_submission_total",
			Help: "Webmail message submission results, known values (those ending with error are server errors): ok, badfrom, messagelimiterror, recipientlimiterror, queueerror, storesenterror.",
		},
		[]string{
			"result",
		},
	)
	metricServerErrors = promauto.NewCounterVec(
		prometheus.CounterOpts{
			Name: "mox_webmail_errors_total",
			Help: "Webmail server errors, known values: dkimsign, submit.",
		},
		[]string{
			"error",
		},
	)
	metricSSEConnections = promauto.NewGauge(
		prometheus.GaugeOpts{
			Name: "mox_webmail_sse_connections",
			Help: "Number of active webmail SSE connections.",
		},
	)
)

func xcheckf(ctx context.Context, err error, format string, args ...any) {
	if err == nil {
		return
	}
	msg := fmt.Sprintf(format, args...)
	errmsg := fmt.Sprintf("%s: %s", msg, err)
	pkglog.WithContext(ctx).Errorx(msg, err)
	code := "server:error"
	if errors.Is(err, context.Canceled) || errors.Is(err, context.DeadlineExceeded) {
		code = "user:error"
	}
	panic(&sherpa.Error{Code: code, Message: errmsg})
}

func xcheckuserf(ctx context.Context, err error, format string, args ...any) {
	if err == nil {
		return
	}
	msg := fmt.Sprintf(format, args...)
	errmsg := fmt.Sprintf("%s: %s", msg, err)
	pkglog.WithContext(ctx).Errorx(msg, err)
	panic(&sherpa.Error{Code: "user:error", Message: errmsg})
}

func xdbwrite(ctx context.Context, acc *store.Account, fn func(tx *bstore.Tx)) {
	err := acc.DB.Write(ctx, func(tx *bstore.Tx) error {
		fn(tx)
		return nil
	})
	xcheckf(ctx, err, "transaction")
}

func xdbread(ctx context.Context, acc *store.Account, fn func(tx *bstore.Tx)) {
	err := acc.DB.Read(ctx, func(tx *bstore.Tx) error {
		fn(tx)
		return nil
	})
	xcheckf(ctx, err, "transaction")
}

// We merge the js into the html at first load, cache a gzipped version that is
// generated on first need, and respond with a Last-Modified header. For quickly
// serving a single, compressed, cacheable file.
type merged struct {
	sync.Mutex
	combined                 []byte
	combinedGzip             []byte
	mtime                    time.Time // For Last-Modified and conditional request.
	fallbackHTML, fallbackJS []byte    // The embedded html/js files.
	htmlPath, jsPath         string    // Paths used during development.
}

var webmail = &merged{
	fallbackHTML: webmailHTML,
	fallbackJS:   webmailJS,
	htmlPath:     filepath.FromSlash("webmail/webmail.html"),
	jsPath:       filepath.FromSlash("webmail/webmail.js"),
}

// fallbackMtime returns a time to use for the Last-Modified header in case we
// cannot find a file, e.g. when used in production.
func fallbackMtime(log mlog.Log) time.Time {
	p, err := os.Executable()
	log.Check(err, "finding executable for mtime")
	if err == nil {
		st, err := os.Stat(p)
		log.Check(err, "stat on executable for mtime")
		if err == nil {
			return st.ModTime()
		}
	}
	log.Info("cannot find executable for webmail mtime, using current time")
	return time.Now()
}

func (m *merged) serve(ctx context.Context, log mlog.Log, w http.ResponseWriter, r *http.Request) {
	// We typically return the embedded file, but during development it's handy
	// to load from disk.
	fhtml, _ := os.Open(m.htmlPath)
	if fhtml != nil {
		defer fhtml.Close()
	}
	fjs, _ := os.Open(m.jsPath)
	if fjs != nil {
		defer fjs.Close()
	}

	html := m.fallbackHTML
	js := m.fallbackJS

	var diskmtime time.Time
	var refreshdisk bool
	if fhtml != nil && fjs != nil {
		sth, err := fhtml.Stat()
		xcheckf(ctx, err, "stat html")
		stj, err := fjs.Stat()
		xcheckf(ctx, err, "stat js")

		maxmtime := sth.ModTime()
		if stj.ModTime().After(maxmtime) {
			maxmtime = stj.ModTime()
		}

		m.Lock()
		refreshdisk = maxmtime.After(m.mtime) || m.combined == nil
		m.Unlock()

		if refreshdisk {
			html, err = io.ReadAll(fhtml)
			xcheckf(ctx, err, "reading html")
			js, err = io.ReadAll(fjs)
			xcheckf(ctx, err, "reading js")
			diskmtime = maxmtime
		}
	}

	gz := acceptsGzip(r)
	var out []byte
	var mtime time.Time
	var origSize int64

	func() {
		m.Lock()
		defer m.Unlock()

		if refreshdisk || m.combined == nil {
			script := []byte(`<script>/* placeholder */</script>`)
			index := bytes.Index(html, script)
			if index < 0 {
				xcheckf(ctx, errors.New("script not found"), "generating combined html")
			}
			var b bytes.Buffer
			b.Write(html[:index])
			fmt.Fprintf(&b, "<script>\n// Javascript is generated from typescript, don't modify the javascript because changes will be lost.\nconst moxversion = \"%s\";\n", moxvar.Version)
			b.Write(js)
			b.WriteString("\t\t</script>")
			b.Write(html[index+len(script):])
			out = b.Bytes()
			m.combined = out
			if refreshdisk {
				m.mtime = diskmtime
			} else {
				m.mtime = fallbackMtime(log)
			}
			m.combinedGzip = nil
		} else {
			out = m.combined
		}
		if gz {
			if m.combinedGzip == nil {
				var b bytes.Buffer
				gzw, err := gzip.NewWriterLevel(&b, gzip.BestCompression)
				if err == nil {
					_, err = gzw.Write(out)
				}
				if err == nil {
					err = gzw.Close()
				}
				xcheckf(ctx, err, "gzipping combined html")
				m.combinedGzip = b.Bytes()
			}
			origSize = int64(len(out))
			out = m.combinedGzip
		}
		mtime = m.mtime
	}()

	w.Header().Set("Content-Type", "text/html; charset=utf-8")
	http.ServeContent(gzipInjector{w, gz, origSize}, r, "", mtime, bytes.NewReader(out))
}

// gzipInjector is a http.ResponseWriter that optionally injects a
// Content-Encoding: gzip header, only in case of status 200 OK. Used with
// http.ServeContent to serve gzipped content if the client supports it. We cannot
// just unconditionally add the content-encoding header, because we don't know
// enough if we will be sending data: http.ServeContent may be sending a "not
// modified" response, and possibly others.
type gzipInjector struct {
	http.ResponseWriter // Keep most methods.
	gz                  bool
	origSize            int64
}

// WriteHeader adds a Content-Encoding: gzip header before actually writing the
// headers and status.
func (w gzipInjector) WriteHeader(statusCode int) {
	if w.gz && statusCode == http.StatusOK {
		w.ResponseWriter.Header().Set("Content-Encoding", "gzip")
		if lw, ok := w.ResponseWriter.(interface{ SetUncompressedSize(int64) }); ok {
			lw.SetUncompressedSize(w.origSize)
		}
	}
	w.ResponseWriter.WriteHeader(statusCode)
}

// Serve content, either from a file, or return the fallback data. Caller
// should already have set the content-type. We use this to return a file from
// the local file system (during development), or embedded in the binary (when
// deployed).
func serveContentFallback(log mlog.Log, w http.ResponseWriter, r *http.Request, path string, fallback []byte) {
	f, err := os.Open(path)
	if err == nil {
		defer f.Close()
		st, err := f.Stat()
		if err == nil {
			http.ServeContent(w, r, "", st.ModTime(), f)
			return
		}
	}
	http.ServeContent(w, r, "", fallbackMtime(log), bytes.NewReader(fallback))
}

// Handler returns a handler for the webmail endpoints, customized for the max
// message size coming from the listener.
func Handler(maxMessageSize int64) func(w http.ResponseWriter, r *http.Request) {
	sh, err := makeSherpaHandler(maxMessageSize)
	return func(w http.ResponseWriter, r *http.Request) {
		if err != nil {
			http.Error(w, "500 - internal server error - cannot handle requests", http.StatusInternalServerError)
			return
		}
		handle(sh, w, r)
	}
}

func handle(apiHandler http.Handler, w http.ResponseWriter, r *http.Request) {
	ctx := r.Context()
	log := pkglog.WithContext(ctx).With(slog.String("userauth", ""))

	// Server-sent event connection, for all initial data (list of mailboxes), list of
	// messages, and all events afterwards. Authenticated through a token in the query
	// string, which it got from a Token API call.
	if r.URL.Path == "/events" {
		serveEvents(ctx, log, w, r)
		return
	}

	// HTTP Basic authentication for all requests.
	loginAddress, accName := webaccount.CheckAuth(ctx, log, "webmail", w, r)
	if accName == "" {
		// Error response already sent.
		return
	}

	if lw, ok := w.(interface{ AddAttr(a slog.Attr) }); ok {
		lw.AddAttr(slog.String("authaccount", accName))
	}

	defer func() {
		x := recover()
		if x == nil {
			return
		}
		err, ok := x.(*sherpa.Error)
		if !ok {
			log.WithContext(ctx).Error("handle panic", slog.Any("err", x))
			debug.PrintStack()
			metrics.PanicInc(metrics.Webmailhandle)
			panic(x)
		}
		if strings.HasPrefix(err.Code, "user:") {
			log.Debugx("webmail user error", err)
			http.Error(w, "400 - bad request - "+err.Message, http.StatusBadRequest)
		} else {
			log.Errorx("webmail server error", err)
			http.Error(w, "500 - internal server error - "+err.Message, http.StatusInternalServerError)
		}
	}()

	switch r.URL.Path {
	case "/":
		switch r.Method {
		default:
			http.Error(w, "405 - method not allowed - use get", http.StatusMethodNotAllowed)
			return
		case "GET", "HEAD":
		}

		webmail.serve(ctx, log, w, r)
		return

	case "/msg.js", "/text.js":
		switch r.Method {
		default:
			http.Error(w, "405 - method not allowed - use get", http.StatusMethodNotAllowed)
			return
		case "GET", "HEAD":
		}

		path := filepath.Join("webmail", r.URL.Path[1:])
		var fallback = webmailmsgJS
		if r.URL.Path == "/text.js" {
			fallback = webmailtextJS
		}

		w.Header().Set("Content-Type", "application/javascript; charset=utf-8")
		serveContentFallback(log, w, r, path, fallback)
		return
	}

	// API calls.
	if strings.HasPrefix(r.URL.Path, "/api/") {
		reqInfo := requestInfo{loginAddress, accName, r}
		ctx = context.WithValue(ctx, requestInfoCtxKey, reqInfo)
		apiHandler.ServeHTTP(w, r.WithContext(ctx))
		return
	}

	// We are now expecting the following URLs:
	// .../msg/<msgid>/{attachments.zip,parsedmessage.js,raw}
	// .../msg/<msgid>/{,msg}{text,html,htmlexternal}
	// .../msg/<msgid>/{view,viewtext,download}/<partid>

	if !strings.HasPrefix(r.URL.Path, "/msg/") {
		http.NotFound(w, r)
		return
	}

	t := strings.Split(r.URL.Path[len("/msg/"):], "/")
	if len(t) < 2 {
		http.NotFound(w, r)
		return
	}

	id, err := strconv.ParseInt(t[0], 10, 64)
	if err != nil || id == 0 {
		http.NotFound(w, r)
		return
	}

	// Many of the requests need either a message or a parsed part. Make it easy to
	// fetch/prepare and cleanup. We only do all the work when the request seems legit
	// (valid HTTP route and method).
	xprepare := func() (acc *store.Account, m store.Message, msgr *store.MsgReader, p message.Part, cleanup func(), ok bool) {
		if r.Method != "GET" {
			http.Error(w, "405 - method not allowed - post required", http.StatusMethodNotAllowed)
			return
		}

		defer func() {
			if ok {
				return
			}
			if msgr != nil {
				err := msgr.Close()
				log.Check(err, "closing message reader")
				msgr = nil
			}
			if acc != nil {
				err := acc.Close()
				log.Check(err, "closing account")
				acc = nil
			}
		}()

		var err error

		acc, err = store.OpenAccount(log, accName)
		xcheckf(ctx, err, "open account")

		m = store.Message{ID: id}
		err = acc.DB.Get(ctx, &m)
		if err == bstore.ErrAbsent || err == nil && m.Expunged {
			http.NotFound(w, r)
			return
		}
		xcheckf(ctx, err, "get message")

		msgr = acc.MessageReader(m)

		p, err = m.LoadPart(msgr)
		xcheckf(ctx, err, "load parsed message")

		cleanup = func() {
			err := msgr.Close()
			log.Check(err, "closing message reader")
			err = acc.Close()
			log.Check(err, "closing account")
		}
		ok = true
		return
	}

	h := w.Header()

	// We set a Content-Security-Policy header that is as strict as possible, depending
	// on the type of message/part/html/js. We have to be careful because we are
	// returning data that is coming in from external places. E.g. HTML could contain
	// javascripts that we don't want to execute, especially not on our domain. We load
	// resources in an iframe. The CSP policy starts out  with default-src 'none' to
	// disallow loading anything, then start allowing what is safe, such as inlined
	// datauri images and inline styles. Data can only be loaded when the request is
	// coming from the same origin (so other sites cannot include resources
	// (messages/parts)).
	//
	// We want to load resources in sandbox-mode, causing the page to be loaded as from
	// a different origin. If sameOrigin is set, we have a looser CSP policy:
	// allow-same-origin is set so resources are loaded as coming from this same
	// origin. This is needed for the msg* endpoints that render a message, where we
	// load the message body in a separate iframe again (with stricter CSP again),
	// which we need to access for its inner height. If allowSelfScript is also set
	// (for "msgtext"), the CSP leaves out the sandbox entirely.
	//
	// If allowExternal is set, we allow loading image, media (audio/video), styles and
	// fronts from external URLs as well as inline URI's. By default we don't allow any
	// loading of content, except inlined images (we do that ourselves for images
	// embedded in the email), and we allow inline styles (which are safely constrained
	// to an iframe).
	//
	// If allowSelfScript is set, inline scripts and scripts from our origin are
	// allowed. Used to display a message including header. The header is rendered with
	// javascript, the content is rendered in a separate iframe with a CSP that doesn't
	// have allowSelfScript.
	headers := func(sameOrigin, allowExternal, allowSelfScript bool) {
		// allow-popups is needed to make opening links in new tabs work.
		sb := "sandbox allow-popups allow-popups-to-escape-sandbox; "
		if sameOrigin && allowSelfScript {
			// Sandbox with both allow-same-origin and allow-script would not provide security,
			// and would give warning in console about that.
			sb = ""
		} else if sameOrigin {
			sb = "sandbox allow-popups allow-popups-to-escape-sandbox allow-same-origin; "
		}
		script := ""
		if allowSelfScript {
			script = "; script-src 'unsafe-inline' 'self'; frame-src 'self'; connect-src 'self'"
		}
		var csp string
		if allowExternal {
			csp = sb + "frame-ancestors 'self'; default-src 'none'; img-src data: http: https: 'unsafe-inline'; style-src 'unsafe-inline' data: http: https:; font-src data: http: https: 'unsafe-inline'; media-src 'unsafe-inline' data: http: https:" + script
		} else {
			csp = sb + "frame-ancestors 'self'; default-src 'none'; img-src data:; style-src 'unsafe-inline'" + script
		}
		h.Set("Content-Security-Policy", csp)
		h.Set("X-Frame-Options", "sameorigin") // Duplicate with CSP, but better too much than too little.
		h.Set("X-Content-Type-Options", "nosniff")
		h.Set("Referrer-Policy", "no-referrer")
	}

	switch {
	case len(t) == 2 && t[1] == "attachments.zip":
		acc, m, msgr, p, cleanup, ok := xprepare()
		if !ok {
			return
		}
		defer cleanup()
		state := msgState{acc: acc, m: m, msgr: msgr, part: &p}
		// note: state is cleared by cleanup

		mi, err := messageItem(log, m, &state)
		xcheckf(ctx, err, "parsing message")

		headers(false, false, false)
		h.Set("Content-Type", "application/zip")
		h.Set("Cache-Control", "no-cache, max-age=0")
		var subjectSlug string
		if p.Envelope != nil {
			s := p.Envelope.Subject
			s = strings.ToLower(s)
			s = regexp.MustCompile("[^a-z0-9_.-]").ReplaceAllString(s, "-")
			s = regexp.MustCompile("--*").ReplaceAllString(s, "-")
			s = strings.TrimLeft(s, "-")
			s = strings.TrimRight(s, "-")
			if s != "" {
				s = "-" + s
			}
			subjectSlug = s
		}
		filename := fmt.Sprintf("email-%d-attachments-%s%s.zip", m.ID, m.Received.Format("20060102-150405"), subjectSlug)
		cd := mime.FormatMediaType("attachment", map[string]string{"filename": filename})
		h.Set("Content-Disposition", cd)

		zw := zip.NewWriter(w)
		names := map[string]bool{}
		for _, a := range mi.Attachments {
			ap := a.Part
			name := tryDecodeParam(log, ap.ContentTypeParams["name"])
			if name == "" {
				// We don't check errors, this is all best-effort.
				h, _ := ap.Header()
				disposition := h.Get("Content-Disposition")
				_, params, _ := mime.ParseMediaType(disposition)
				name = tryDecodeParam(log, params["filename"])
			}
			if name != "" {
				name = filepath.Base(name)
			}
			mt := strings.ToLower(ap.MediaType + "/" + ap.MediaSubType)
			if name == "" || names[name] {
				ext := filepath.Ext(name)
				if ext == "" {
					// Handle just a few basic types.
					extensions := map[string]string{
						"text/plain":      ".txt",
						"text/html":       ".html",
						"image/jpeg":      ".jpg",
						"image/png":       ".png",
						"image/gif":       ".gif",
						"application/zip": ".zip",
					}
					ext = extensions[mt]
					if ext == "" {
						ext = ".bin"
					}
				}
				var stem string
				if name != "" && strings.HasSuffix(name, ext) {
					stem = strings.TrimSuffix(name, ext)
				} else {
					stem = "attachment"
					for _, index := range a.Path {
						stem += fmt.Sprintf("-%d", index)
					}
				}
				name = stem + ext
				seq := 0
				for names[name] {
					seq++
					name = stem + fmt.Sprintf("-%d", seq) + ext
				}
			}
			names[name] = true

			fh := zip.FileHeader{
				Name:     name,
				Modified: m.Received,
			}
			nodeflate := map[string]bool{
				"application/x-bzip2":          true,
				"application/zip":              true,
				"application/x-zip-compressed": true,
				"application/gzip":             true,
				"application/x-gzip":           true,
				"application/vnd.rar":          true,
				"application/x-rar-compressed": true,
				"application/x-7z-compressed":  true,
			}
			// Sniff content-type as well for compressed data.
			buf := make([]byte, 512)
			n, _ := io.ReadFull(ap.Reader(), buf)
			var sniffmt string
			if n > 0 {
				sniffmt = strings.ToLower(http.DetectContentType(buf[:n]))
			}
			deflate := ap.MediaType != "VIDEO" && ap.MediaType != "AUDIO" && (ap.MediaType != "IMAGE" || ap.MediaSubType == "BMP") && !nodeflate[mt] && !nodeflate[sniffmt]
			if deflate {
				fh.Method = zip.Deflate
			}
			// We cannot return errors anymore: we have already sent an application/zip header.
			if zf, err := zw.CreateHeader(&fh); err != nil {
				log.Check(err, "adding to zip file")
				return
			} else if _, err := io.Copy(zf, ap.Reader()); err != nil {
				log.Check(err, "writing to zip file")
				return
			}
		}
		err = zw.Close()
		log.Check(err, "final write to zip file")

	// Raw display of a message, as text/plain.
	case len(t) == 2 && t[1] == "raw":
		_, _, msgr, p, cleanup, ok := xprepare()
		if !ok {
			return
		}
		defer cleanup()

		// We intentially use text/plain. We certainly don't want to return a format that
		// browsers or users would think of executing. We do set the charset if available
		// on the outer part. If present, we assume it may be relevant for other parts. If
		// not, there is not much we could do better...
		headers(false, false, false)
		ct := "text/plain"
		params := map[string]string{}
		if charset := p.ContentTypeParams["charset"]; charset != "" {
			params["charset"] = charset
		}
		h.Set("Content-Type", mime.FormatMediaType(ct, params))
		h.Set("Cache-Control", "no-cache, max-age=0")

		_, err := io.Copy(w, &moxio.AtReader{R: msgr})
		log.Check(err, "writing raw")

	case len(t) == 2 && (t[1] == "msgtext" || t[1] == "msghtml" || t[1] == "msghtmlexternal"):
		// msg.html has a javascript tag with message data, and javascript to render the
		// message header like the regular webmail.html and to load the message body in a
		// separate iframe with a separate request with stronger CSP.
		acc, m, msgr, p, cleanup, ok := xprepare()
		if !ok {
			return
		}
		defer cleanup()

		state := msgState{acc: acc, m: m, msgr: msgr, part: &p}
		// note: state is cleared by cleanup

		pm, err := parsedMessage(log, m, &state, true, true)
		xcheckf(ctx, err, "getting parsed message")
		if t[1] == "msgtext" && len(pm.Texts) == 0 || t[1] != "msgtext" && !pm.HasHTML {
			http.Error(w, "400 - bad request - no such part", http.StatusBadRequest)
			return
		}

		sameorigin := true
		loadExternal := t[1] == "msghtmlexternal"
		allowSelfScript := true
		headers(sameorigin, loadExternal, allowSelfScript)
		h.Set("Content-Type", "text/html; charset=utf-8")
		h.Set("Cache-Control", "no-cache, max-age=0")

		path := filepath.FromSlash("webmail/msg.html")
		fallback := webmailmsgHTML
		serveContentFallback(log, w, r, path, fallback)

	case len(t) == 2 && t[1] == "parsedmessage.js":
		// Used by msg.html, for the msg* endpoints, for the data needed to show all data
		// except the message body.
		// This is js with data inside instead so we can load it synchronously, which we do
		// to get a "loaded" event after the page was actually loaded.

		acc, m, msgr, p, cleanup, ok := xprepare()
		if !ok {
			return
		}
		defer cleanup()
		state := msgState{acc: acc, m: m, msgr: msgr, part: &p}
		// note: state is cleared by cleanup

		pm, err := parsedMessage(log, m, &state, true, true)
		xcheckf(ctx, err, "parsing parsedmessage")
		pmjson, err := json.Marshal(pm)
		xcheckf(ctx, err, "marshal parsedmessage")

		m.MsgPrefix = nil
		m.ParsedBuf = nil
		mi := MessageItem{m, pm.envelope, pm.attachments, pm.isSigned, pm.isEncrypted, pm.firstLine, false}
		mijson, err := json.Marshal(mi)
		xcheckf(ctx, err, "marshal messageitem")

		headers(false, false, false)
		h.Set("Content-Type", "application/javascript; charset=utf-8")
		h.Set("Cache-Control", "no-cache, max-age=0")

		_, err = fmt.Fprintf(w, "window.messageItem = %s;\nwindow.parsedMessage = %s;\n", mijson, pmjson)
		log.Check(err, "writing parsedmessage.js")

	case len(t) == 2 && t[1] == "text":
		// Returns text.html whichs loads the message data with a javascript tag and
		// renders just the text content with the same code as webmail.html. Used by the
		// iframe in the msgtext endpoint. Not used by the regular webmail viewer, it
		// renders the text itself, with the same shared js code.
		acc, m, msgr, p, cleanup, ok := xprepare()
		if !ok {
			return
		}
		defer cleanup()

		state := msgState{acc: acc, m: m, msgr: msgr, part: &p}
		// note: state is cleared by cleanup

		pm, err := parsedMessage(log, m, &state, true, true)
		xcheckf(ctx, err, "parsing parsedmessage")

		if len(pm.Texts) == 0 {
			http.Error(w, "400 - bad request - no text part in message", http.StatusBadRequest)
			return
		}

		// Needed for inner document height for outer iframe height in separate message view.
		sameorigin := true
		allowSelfScript := true
		headers(sameorigin, false, allowSelfScript)
		h.Set("Content-Type", "text/html; charset=utf-8")
		h.Set("Cache-Control", "no-cache, max-age=0")

		// We typically return the embedded file, but during development it's handy to load
		// from disk.
		path := filepath.FromSlash("webmail/text.html")
		fallback := webmailtextHTML
		serveContentFallback(log, w, r, path, fallback)

	case len(t) == 2 && (t[1] == "html" || t[1] == "htmlexternal"):
		// Returns the first HTML part, with "cid:" URIs replaced with an inlined datauri
		// if the referenced Content-ID attachment can be found.
		_, _, _, p, cleanup, ok := xprepare()
		if !ok {
			return
		}
		defer cleanup()

		setHeaders := func() {
			// Needed for inner document height for outer iframe height in separate message
			// view. We only need that when displaying as a separate message on the msghtml*
			// endpoints. When displaying in the regular webmail, we don't need to know the
			// inner height so we load it as different origin, which should be safer.
			sameorigin := r.URL.Query().Get("sameorigin") == "true"
			allowExternal := strings.HasSuffix(t[1], "external")
			headers(sameorigin, allowExternal, false)

			h.Set("Content-Type", "text/html; charset=utf-8")
			h.Set("Cache-Control", "no-cache, max-age=0")
		}

		// todo: skip certain html parts? e.g. with content-disposition: attachment?
		var done bool
		var usePart func(p *message.Part, parents []*message.Part)
		usePart = func(p *message.Part, parents []*message.Part) {
			if done {
				return
			}
			mt := p.MediaType + "/" + p.MediaSubType
			switch mt {
			case "TEXT/HTML":
				done = true
				err := inlineSanitizeHTML(log, setHeaders, w, p, parents)
				if err != nil {
					http.Error(w, "400 - bad request - "+err.Error(), http.StatusBadRequest)
				}
				return
			}
			parents = append(parents, p)
			for _, sp := range p.Parts {
				usePart(&sp, parents)
			}
		}
		usePart(&p, nil)

		if !done {
			http.Error(w, "400 - bad request - no html part in message", http.StatusBadRequest)
		}

	case len(t) == 3 && (t[1] == "view" || t[1] == "viewtext" || t[1] == "download"):
		// View any part, as referenced in the last element path. "0" is the whole message,
		// 0.0 is the first subpart, etc. "view" returns it with the content-type from the
		// message (could be dangerous, but we set strict CSP headers), "viewtext" returns
		// data with a text/plain content-type so the browser will attempt to display it,
		// and "download" adds a content-disposition header causing the browser the
		// download the file.
		_, _, _, p, cleanup, ok := xprepare()
		if !ok {
			return
		}
		defer cleanup()

		paths := strings.Split(t[2], ".")
		if len(paths) == 0 || paths[0] != "0" {
			http.NotFound(w, r)
			return
		}
		ap := p
		for _, e := range paths[1:] {
			index, err := strconv.ParseInt(e, 10, 32)
			if err != nil || index < 0 || int(index) >= len(ap.Parts) {
				http.NotFound(w, r)
				return
			}
			ap = ap.Parts[int(index)]
		}

		headers(false, false, false)
		var ct string
		if t[1] == "viewtext" {
			ct = "text/plain"
		} else {
			ct = strings.ToLower(ap.MediaType + "/" + ap.MediaSubType)
		}
		h.Set("Content-Type", ct)
		h.Set("Cache-Control", "no-cache, max-age=0")
		if t[1] == "download" {
			name := tryDecodeParam(log, ap.ContentTypeParams["name"])
			if name == "" {
				// We don't check errors, this is all best-effort.
				h, _ := ap.Header()
				disposition := h.Get("Content-Disposition")
				_, params, _ := mime.ParseMediaType(disposition)
				name = tryDecodeParam(log, params["filename"])
			}
			if name == "" {
				name = "attachment.bin"
			}
			cd := mime.FormatMediaType("attachment", map[string]string{"filename": name})
			h.Set("Content-Disposition", cd)
		}

		_, err := io.Copy(w, ap.Reader())
		if err != nil && !moxio.IsClosed(err) {
			log.Errorx("copying attachment", err)
		}
	default:
		http.NotFound(w, r)
	}
}

func acceptsGzip(r *http.Request) bool {
	s := r.Header.Get("Accept-Encoding")
	t := strings.Split(s, ",")
	for _, e := range t {
		e = strings.TrimSpace(e)
		tt := strings.Split(e, ";")
		if len(tt) > 1 && t[1] == "q=0" {
			continue
		}
		if tt[0] == "gzip" {
			return true
		}
	}
	return false
}

// inlineSanitizeHTML writes the part as HTML, with "cid:" URIs for html "src"
// attributes inlined and with potentially dangerous tags removed (javascript). The
// sanitizing is just a first layer of defense, CSP headers block execution of
// scripts. If the HTML becomes too large, an error is returned. Before writing
// HTML, setHeaders is called to write the required headers for content-type and
// CSP. On error, setHeader is not called, no output is written and the caller
// should write an error response.
func inlineSanitizeHTML(log mlog.Log, setHeaders func(), w io.Writer, p *message.Part, parents []*message.Part) error {
	// Prepare cids if there is a chance we will use them.
	cids := map[string]*message.Part{}
	for _, parent := range parents {
		if parent.MediaType+"/"+parent.MediaSubType == "MULTIPART/RELATED" && p.DecodedSize < 2*1024*1024 {
			for i, rp := range parent.Parts {
				if rp.ContentID != "" {
					cids[strings.ToLower(rp.ContentID)] = &parent.Parts[i]
				}
			}
		}
	}

	node, err := html.Parse(p.ReaderUTF8OrBinary())
	if err != nil {
		return fmt.Errorf("parsing html: %v", err)
	}

	// We track size, if it becomes too much, we abort and still copy as regular html.
	var totalSize int64
	if err := inlineNode(node, cids, &totalSize); err != nil {
		return fmt.Errorf("inline cid uris in html nodes: %w", err)
	}
	sanitizeNode(node)
	setHeaders()
	err = html.Render(w, node)
	log.Check(err, "writing html")
	return nil
}

// We inline cid: URIs into data: URIs. If a cid is missing in the
// multipart/related, we ignore the error and continue with other HTML nodes. It
// will probably just result in a "broken image". We limit the max size we
// generate. We only replace "src" attributes that start with "cid:". A cid URI
// could theoretically occur in many more places, like link href, and css url().
// That's probably not common though. Let's wait for someone to need it.
func inlineNode(node *html.Node, cids map[string]*message.Part, totalSize *int64) error {
	for i, a := range node.Attr {
		if a.Key != "src" || !caselessPrefix(a.Val, "cid:") || a.Namespace != "" {
			continue
		}
		cid := a.Val[4:]
		ap := cids["<"+strings.ToLower(cid)+">"]
		if ap == nil {
			// Missing cid, can happen with email, no need to stop returning data.
			continue
		}
		*totalSize += ap.DecodedSize
		if *totalSize >= 10*1024*1024 {
			return fmt.Errorf("html too large")
		}
		var sb strings.Builder
		if _, err := fmt.Fprintf(&sb, "data:%s;base64,", strings.ToLower(ap.MediaType+"/"+ap.MediaSubType)); err != nil {
			return fmt.Errorf("writing datauri: %v", err)
		}
		w := base64.NewEncoder(base64.StdEncoding, &sb)
		if _, err := io.Copy(w, ap.Reader()); err != nil {
			return fmt.Errorf("writing base64 datauri: %v", err)
		}
		node.Attr[i].Val = sb.String()
	}
	for node = node.FirstChild; node != nil; node = node.NextSibling {
		if err := inlineNode(node, cids, totalSize); err != nil {
			return err
		}
	}
	return nil
}

func caselessPrefix(k, pre string) bool {
	return len(k) >= len(pre) && strings.EqualFold(k[:len(pre)], pre)
}

var targetable = map[string]bool{
	"a":    true,
	"area": true,
	"form": true,
	"base": true,
}

// sanitizeNode removes script elements, on* attributes, javascript: href
// attributes, adds target="_blank" to all links and to a base tag.
func sanitizeNode(node *html.Node) {
	i := 0
	var haveTarget, haveRel bool
	for i < len(node.Attr) {
		a := node.Attr[i]
		// Remove dangerous attributes.
		if strings.HasPrefix(a.Key, "on") || a.Key == "href" && caselessPrefix(a.Val, "javascript:") || a.Key == "src" && caselessPrefix(a.Val, "data:text/html") {
			copy(node.Attr[i:], node.Attr[i+1:])
			node.Attr = node.Attr[:len(node.Attr)-1]
			continue
		}
		if a.Key == "target" {
			node.Attr[i].Val = "_blank"
			haveTarget = true
		}
		if a.Key == "rel" && targetable[node.Data] {
			node.Attr[i].Val = "noopener noreferrer"
			haveRel = true
		}
		i++
	}
	// Ensure target attribute is set for elements that can have it.
	if !haveTarget && node.Type == html.ElementNode && targetable[node.Data] {
		node.Attr = append(node.Attr, html.Attribute{Key: "target", Val: "_blank"})
		haveTarget = true
	}
	if haveTarget && !haveRel {
		node.Attr = append(node.Attr, html.Attribute{Key: "rel", Val: "noopener noreferrer"})
	}

	parent := node
	node = node.FirstChild
	var haveBase bool
	for node != nil {
		// Set next now, we may remove cur, which clears its NextSibling.
		cur := node
		node = node.NextSibling

		// Remove script elements.
		if cur.Type == html.ElementNode && cur.Data == "script" {
			parent.RemoveChild(cur)
			continue
		}
		sanitizeNode(cur)
	}
	if parent.Type == html.ElementNode && parent.Data == "head" && !haveBase {
		n := html.Node{Type: html.ElementNode, Data: "base", Attr: []html.Attribute{{Key: "target", Val: "_blank"}, {Key: "rel", Val: "noopener noreferrer"}}}
		parent.AppendChild(&n)
	}
}