git » debian:dnss » commit c98400d

httpresolver: React faster to network errors

author Alberto Bertogli
2020-08-22 22:49:43 UTC
committer Alberto Bertogli
2020-08-22 22:49:43 UTC
parent da7ed40160b1ce27621a955708562cfd9fb7a1f5

httpresolver: React faster to network errors

When the TCP connection to the upstream server dies unexpectedly, dnss
can return errors for a very long time (hours) instead of reconnecting.
This is because the HTTP/2 transport will insist on using the existing
(dead) connection, as it doesn't give up on it.

This can easily happen when the computer dnss changes wifi networks, for
example.

Unfortunately, there is no practical way to tell the transport to drop
the connections.

So this patch introduces a workaround, which is to switch to a new
client when we've seen errors consistently for 10s. That's enough time
for 2 queries back to back to timeout (each query has a 4s timeout).

internal/httpresolver/resolver.go +106 -16

diff --git a/internal/httpresolver/resolver.go b/internal/httpresolver/resolver.go
index 665fa92..1e0426f 100644
--- a/internal/httpresolver/resolver.go
+++ b/internal/httpresolver/resolver.go
@@ -8,8 +8,10 @@ import (
 	"io"
 	"io/ioutil"
 	"mime"
+	"net"
 	"net/http"
 	"net/url"
+	"sync"
 	"time"
 
 	"blitiri.com.ar/go/dnss/internal/dnsserver"
@@ -22,9 +24,15 @@ import (
 // httpsResolver implements the dnsserver.Resolver interface by querying a
 // server via DNS over HTTPS (DoH, RFC 8484).
 type httpsResolver struct {
-	Upstream *url.URL
-	CAFile   string
+	Upstream  *url.URL
+	CAFile    string
+	tlsConfig *tls.Config
+
+	mu       sync.Mutex
 	client   *http.Client
+	firstErr time.Time
+
+	ev trace.EventLog
 }
 
 func loadCertPool(caFile string) (*x509.CertPool, error) {
@@ -51,12 +59,56 @@ func NewDoH(upstream *url.URL, caFile string) *httpsResolver {
 }
 
 func (r *httpsResolver) Init() error {
+	// If CAFile is empty, we're ok with the defaults (use the system default
+	// CA database).
+	if r.CAFile != "" {
+		pool, err := loadCertPool(r.CAFile)
+		if err != nil {
+			return err
+		}
+
+		r.tlsConfig = &tls.Config{
+			RootCAs: pool,
+		}
+	}
+
+	client, err := r.newClient()
+
+	r.mu.Lock()
+	r.client = client
+	r.mu.Unlock()
+
+	r.ev = trace.NewEventLog("httpresolver", r.Upstream.String())
+	r.ev.Printf("Init complete, client: %p", r.client)
+
+	return err
+}
+
+func (r *httpsResolver) newClient() (*http.Client, error) {
 	transport := &http.Transport{
+		TLSClientConfig: r.tlsConfig,
+
 		// Take the semi-standard proxy settings from the environment.
 		Proxy: http.ProxyFromEnvironment,
+
+		// Drop connections after 30s idle.
+		// This helps prevent connection pile-up on frequent client rotations,
+		// which can happen with intermittent network issues.
+		IdleConnTimeout: 30 * time.Second,
+
+		// Reasonable defaults, based on http.DefaultTransport.
+		DialContext: (&net.Dialer{
+			Timeout:   10 * time.Second,
+			KeepAlive: 1 * time.Second,
+			DualStack: true,
+		}).DialContext,
+		ForceAttemptHTTP2:     true,
+		MaxIdleConns:          10,
+		TLSHandshakeTimeout:   4 * time.Second,
+		ExpectContinueTimeout: 1 * time.Second,
 	}
 
-	r.client = &http.Client{
+	client := &http.Client{
 		// Give our HTTP requests 4 second timeouts: DNS usually doesn't wait
 		// that long anyway, but this helps with slow connections.
 		Timeout: 4 * time.Second,
@@ -64,23 +116,56 @@ func (r *httpsResolver) Init() error {
 		Transport: transport,
 	}
 
-	// If CAFile is empty, we're ok with the defaults (use the system default
-	// CA database).
-	if r.CAFile != "" {
-		pool, err := loadCertPool(r.CAFile)
-		if err != nil {
-			return err
-		}
+	return client, nil
+}
 
-		transport.TLSClientConfig = &tls.Config{
-			RootCAs: pool,
-		}
-	}
+func (r *httpsResolver) setClientError(err error) {
+	r.mu.Lock()
+	defer r.mu.Unlock()
 
-	return nil
+	if err == nil {
+		r.firstErr = time.Time{}
+	} else if r.firstErr.IsZero() {
+		r.firstErr = time.Now()
+	}
 }
 
 func (r *httpsResolver) Maintain() {
+	for range time.Tick(2 * time.Second) {
+		r.maybeRotateClient()
+	}
+}
+
+func (r *httpsResolver) maybeRotateClient() {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+
+	if r.firstErr.IsZero() {
+		return
+	}
+
+	// If we've seen errors for the last 10s, rotate the client.
+	// This is unfortunately needed because the Go HTTP/2 transport will
+	// insist on using a dead connection for a long time, and cannot be told
+	// to close it. This causes problems when the computer changes connections
+	// (e.g. switch wifi network) or is having intermittent network issues.
+	// This workaround works because a new client will initiate a new
+	// connection, and the old one will die in the background.
+	// The time chosen here combines with the transport timeouts set above, so
+	// we never have too many in-flight connections.
+	if time.Since(r.firstErr) > 10*time.Second {
+		r.ev.Printf("Rotating client after %s of errors: %p",
+			time.Since(r.firstErr), r.client)
+		client, err := r.newClient()
+		if err != nil {
+			r.ev.Errorf("Error creating new client: %v", err)
+			return
+		}
+
+		r.client = client
+		r.firstErr = time.Time{}
+		r.ev.Printf("Rotated client: %p", r.client)
+	}
 }
 
 func (r *httpsResolver) Query(req *dns.Msg, tr trace.Trace) (*dns.Msg, error) {
@@ -95,10 +180,15 @@ func (r *httpsResolver) Query(req *dns.Msg, tr trace.Trace) (*dns.Msg, error) {
 
 	// TODO: Accept header.
 
-	hr, err := r.client.Post(
+	r.mu.Lock()
+	client := r.client
+	r.mu.Unlock()
+
+	hr, err := client.Post(
 		r.Upstream.String(),
 		"application/dns-message",
 		bytes.NewReader(packed))
+	r.setClientError(err)
 	if err != nil {
 		return nil, fmt.Errorf("POST failed: %v", err)
 	}