2019-06-30 17:07:58 -05:00
|
|
|
// Copyright 2015 Matthew Holt and The Caddy Authors
|
|
|
|
//
|
|
|
|
// Licensed under the Apache License, Version 2.0 (the "License");
|
|
|
|
// you may not use this file except in compliance with the License.
|
|
|
|
// You may obtain a copy of the License at
|
|
|
|
//
|
|
|
|
// http://www.apache.org/licenses/LICENSE-2.0
|
|
|
|
//
|
|
|
|
// Unless required by applicable law or agreed to in writing, software
|
|
|
|
// distributed under the License is distributed on an "AS IS" BASIS,
|
|
|
|
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
|
|
// See the License for the specific language governing permissions and
|
|
|
|
// limitations under the License.
|
|
|
|
|
2019-05-04 11:49:50 -05:00
|
|
|
package reverseproxy
|
|
|
|
|
|
|
|
import (
|
2019-11-15 19:15:33 -05:00
|
|
|
"bytes"
|
2019-05-04 11:49:50 -05:00
|
|
|
"context"
|
2019-09-02 23:01:02 -05:00
|
|
|
"encoding/json"
|
2020-11-25 12:54:23 -05:00
|
|
|
"errors"
|
2019-05-04 11:49:50 -05:00
|
|
|
"fmt"
|
2019-11-15 19:15:33 -05:00
|
|
|
"io"
|
2019-05-04 11:49:50 -05:00
|
|
|
"net"
|
|
|
|
"net/http"
|
2022-08-09 11:53:24 -05:00
|
|
|
"net/http/httptrace"
|
2022-08-17 17:10:57 -05:00
|
|
|
"net/netip"
|
2021-08-12 11:48:24 -05:00
|
|
|
"net/textproto"
|
2021-03-29 19:36:40 -05:00
|
|
|
"net/url"
|
2019-09-03 13:10:11 -05:00
|
|
|
"regexp"
|
2022-08-09 11:53:24 -05:00
|
|
|
"runtime"
|
2020-06-04 13:06:38 -05:00
|
|
|
"strconv"
|
2019-05-04 11:49:50 -05:00
|
|
|
"strings"
|
2019-11-15 19:15:33 -05:00
|
|
|
"sync"
|
2019-05-04 11:49:50 -05:00
|
|
|
"time"
|
|
|
|
|
2019-09-02 23:01:02 -05:00
|
|
|
"github.com/caddyserver/caddy/v2"
|
2021-05-02 13:39:06 -05:00
|
|
|
"github.com/caddyserver/caddy/v2/caddyconfig/caddyfile"
|
2019-09-02 23:01:02 -05:00
|
|
|
"github.com/caddyserver/caddy/v2/modules/caddyhttp"
|
2019-09-14 14:25:26 -05:00
|
|
|
"github.com/caddyserver/caddy/v2/modules/caddyhttp/headers"
|
2022-05-06 09:50:26 -05:00
|
|
|
"github.com/caddyserver/caddy/v2/modules/caddyhttp/rewrite"
|
2019-10-29 17:02:40 -05:00
|
|
|
"go.uber.org/zap"
|
2019-05-04 11:49:50 -05:00
|
|
|
"golang.org/x/net/http/httpguts"
|
|
|
|
)
|
|
|
|
|
2022-08-09 11:53:24 -05:00
|
|
|
var supports1xx bool
|
|
|
|
|
2019-09-02 23:01:02 -05:00
|
|
|
func init() {
|
2022-08-29 13:31:53 -05:00
|
|
|
// Caddy requires at least Go 1.18, but Early Hints requires Go 1.19; thus we can simply check for 1.18 in version string
|
|
|
|
// TODO: remove this once our minimum Go version is 1.19
|
|
|
|
supports1xx = !strings.Contains(runtime.Version(), "go1.18")
|
2022-08-09 11:53:24 -05:00
|
|
|
|
2019-09-02 23:01:02 -05:00
|
|
|
caddy.RegisterModule(Handler{})
|
2019-05-04 11:49:50 -05:00
|
|
|
}
|
|
|
|
|
2019-09-03 17:56:09 -05:00
|
|
|
// Handler implements a highly configurable and production-ready reverse proxy.
|
2019-12-29 15:16:34 -05:00
|
|
|
//
|
2019-12-23 14:45:35 -05:00
|
|
|
// Upon proxying, this module sets the following placeholders (which can be used
|
2021-02-22 13:57:21 -05:00
|
|
|
// both within and after this handler; for example, in response headers):
|
2019-12-23 14:45:35 -05:00
|
|
|
//
|
2019-12-29 15:16:34 -05:00
|
|
|
// Placeholder | Description
|
|
|
|
// ------------|-------------
|
|
|
|
// `{http.reverse_proxy.upstream.address}` | The full address to the upstream as given in the config
|
|
|
|
// `{http.reverse_proxy.upstream.hostport}` | The host:port of the upstream
|
|
|
|
// `{http.reverse_proxy.upstream.host}` | The host of the upstream
|
|
|
|
// `{http.reverse_proxy.upstream.port}` | The port of the upstream
|
|
|
|
// `{http.reverse_proxy.upstream.requests}` | The approximate current number of requests to the upstream
|
|
|
|
// `{http.reverse_proxy.upstream.max_requests}` | The maximum approximate number of requests allowed to the upstream
|
|
|
|
// `{http.reverse_proxy.upstream.fails}` | The number of recent failed requests to the upstream
|
2021-02-22 13:57:21 -05:00
|
|
|
// `{http.reverse_proxy.upstream.latency}` | How long it took the proxy upstream to write the response header.
|
2022-04-11 14:04:05 -05:00
|
|
|
// `{http.reverse_proxy.upstream.latency_ms}` | Same as 'latency', but in milliseconds.
|
2021-02-22 13:57:21 -05:00
|
|
|
// `{http.reverse_proxy.upstream.duration}` | Time spent proxying to the upstream, including writing response body to client.
|
2022-04-11 14:04:05 -05:00
|
|
|
// `{http.reverse_proxy.upstream.duration_ms}` | Same as 'upstream.duration', but in milliseconds.
|
2021-02-22 13:57:21 -05:00
|
|
|
// `{http.reverse_proxy.duration}` | Total time spent proxying, including selecting an upstream, retries, and writing response.
|
2022-04-11 14:04:05 -05:00
|
|
|
// `{http.reverse_proxy.duration_ms}` | Same as 'duration', but in milliseconds.
|
2019-09-02 23:01:02 -05:00
|
|
|
type Handler struct {
|
2019-12-23 14:45:35 -05:00
|
|
|
// Configures the method of transport for the proxy. A transport
|
|
|
|
// is what performs the actual "round trip" to the backend.
|
|
|
|
// The default transport is plaintext HTTP.
|
|
|
|
TransportRaw json.RawMessage `json:"transport,omitempty" caddy:"namespace=http.reverse_proxy.transport inline_key=protocol"`
|
|
|
|
|
|
|
|
// A circuit breaker may be used to relieve pressure on a backend
|
|
|
|
// that is beginning to exhibit symptoms of stress or latency.
|
|
|
|
// By default, there is no circuit breaker.
|
|
|
|
CBRaw json.RawMessage `json:"circuit_breaker,omitempty" caddy:"namespace=http.reverse_proxy.circuit_breakers inline_key=type"`
|
|
|
|
|
|
|
|
// Load balancing distributes load/requests between backends.
|
|
|
|
LoadBalancing *LoadBalancing `json:"load_balancing,omitempty"`
|
|
|
|
|
|
|
|
// Health checks update the status of backends, whether they are
|
|
|
|
// up or down. Down backends will not be proxied to.
|
|
|
|
HealthChecks *HealthChecks `json:"health_checks,omitempty"`
|
|
|
|
|
2022-03-06 19:43:39 -05:00
|
|
|
// Upstreams is the static list of backends to proxy to.
|
2019-12-23 14:45:35 -05:00
|
|
|
Upstreams UpstreamPool `json:"upstreams,omitempty"`
|
|
|
|
|
2022-03-06 19:43:39 -05:00
|
|
|
// A module for retrieving the list of upstreams dynamically. Dynamic
|
|
|
|
// upstreams are retrieved at every iteration of the proxy loop for
|
|
|
|
// each request (i.e. before every proxy attempt within every request).
|
|
|
|
// Active health checks do not work on dynamic upstreams, and passive
|
|
|
|
// health checks are only effective on dynamic upstreams if the proxy
|
|
|
|
// server is busy enough that concurrent requests to the same backends
|
|
|
|
// are continuous. Instead of health checks for dynamic upstreams, it
|
|
|
|
// is recommended that the dynamic upstream module only return available
|
|
|
|
// backends in the first place.
|
|
|
|
DynamicUpstreamsRaw json.RawMessage `json:"dynamic_upstreams,omitempty" caddy:"namespace=http.reverse_proxy.upstreams inline_key=source"`
|
|
|
|
|
2021-08-12 11:48:24 -05:00
|
|
|
// Adjusts how often to flush the response buffer. By default,
|
|
|
|
// no periodic flushing is done. A negative value disables
|
|
|
|
// response buffering, and flushes immediately after each
|
|
|
|
// write to the client. This option is ignored when the upstream's
|
|
|
|
// response is recognized as a streaming response, or if its
|
|
|
|
// content length is -1; for such responses, writes are flushed
|
|
|
|
// to the client immediately.
|
2022-08-12 14:15:41 -05:00
|
|
|
//
|
|
|
|
// Normally, a request will be canceled if the client disconnects
|
|
|
|
// before the response is received from the backend. If explicitly
|
|
|
|
// set to -1, client disconnection will be ignored and the request
|
|
|
|
// will be completed to help facilitate low-latency streaming.
|
2019-12-23 14:45:35 -05:00
|
|
|
FlushInterval caddy.Duration `json:"flush_interval,omitempty"`
|
|
|
|
|
2022-03-06 18:51:55 -05:00
|
|
|
// A list of IP ranges (supports CIDR notation) from which
|
|
|
|
// X-Forwarded-* header values should be trusted. By default,
|
|
|
|
// no proxies are trusted, so existing values will be ignored
|
|
|
|
// when setting these headers. If the proxy is trusted, then
|
|
|
|
// existing values will be used when constructing the final
|
|
|
|
// header values.
|
|
|
|
TrustedProxies []string `json:"trusted_proxies,omitempty"`
|
|
|
|
|
2019-12-23 14:45:35 -05:00
|
|
|
// Headers manipulates headers between Caddy and the backend.
|
2020-04-17 10:53:06 -05:00
|
|
|
// By default, all headers are passed-thru without changes,
|
|
|
|
// with the exceptions of special hop-by-hop headers.
|
|
|
|
//
|
2022-03-06 18:51:55 -05:00
|
|
|
// X-Forwarded-For, X-Forwarded-Proto and X-Forwarded-Host
|
|
|
|
// are also set implicitly.
|
2019-12-23 14:45:35 -05:00
|
|
|
Headers *headers.Handler `json:"headers,omitempty"`
|
|
|
|
|
|
|
|
// If true, the entire request body will be read and buffered
|
|
|
|
// in memory before being proxied to the backend. This should
|
2021-02-09 16:15:04 -05:00
|
|
|
// be avoided if at all possible for performance reasons, but
|
|
|
|
// could be useful if the backend is intolerant of read latency.
|
2019-12-23 14:45:35 -05:00
|
|
|
BufferRequests bool `json:"buffer_requests,omitempty"`
|
2019-09-02 23:01:02 -05:00
|
|
|
|
2021-02-09 16:15:04 -05:00
|
|
|
// If true, the entire response body will be read and buffered
|
|
|
|
// in memory before being proxied to the client. This should
|
|
|
|
// be avoided if at all possible for performance reasons, but
|
|
|
|
// could be useful if the backend has tighter memory constraints.
|
|
|
|
BufferResponses bool `json:"buffer_responses,omitempty"`
|
|
|
|
|
|
|
|
// If body buffering is enabled, the maximum size of the buffers
|
|
|
|
// used for the requests and responses (in bytes).
|
|
|
|
MaxBufferSize int64 `json:"max_buffer_size,omitempty"`
|
|
|
|
|
2022-05-06 09:50:26 -05:00
|
|
|
// If configured, rewrites the copy of the upstream request.
|
|
|
|
// Allows changing the request method and URI (path and query).
|
|
|
|
// Since the rewrite is applied to the copy, it does not persist
|
|
|
|
// past the reverse proxy handler.
|
|
|
|
// If the method is changed to `GET` or `HEAD`, the request body
|
|
|
|
// will not be copied to the backend. This allows a later request
|
|
|
|
// handler -- either in a `handle_response` route, or after -- to
|
|
|
|
// read the body.
|
|
|
|
// By default, no rewrite is performed, and the method and URI
|
|
|
|
// from the incoming request is used as-is for proxying.
|
|
|
|
Rewrite *rewrite.Rewrite `json:"rewrite,omitempty"`
|
|
|
|
|
2020-05-27 11:15:20 -05:00
|
|
|
// List of handlers and their associated matchers to evaluate
|
|
|
|
// after successful roundtrips. The first handler that matches
|
|
|
|
// the response from a backend will be invoked. The response
|
|
|
|
// body from the backend will not be written to the client;
|
|
|
|
// it is up to the handler to finish handling the response.
|
|
|
|
// If passive health checks are enabled, any errors from the
|
|
|
|
// handler chain will not affect the health status of the
|
|
|
|
// backend.
|
|
|
|
//
|
2021-05-12 15:19:08 -05:00
|
|
|
// Three new placeholders are available in this handler chain:
|
|
|
|
// - `{http.reverse_proxy.status_code}` The status code from the response
|
|
|
|
// - `{http.reverse_proxy.status_text}` The status text from the response
|
|
|
|
// - `{http.reverse_proxy.header.*}` The headers from the response
|
2020-05-27 11:15:20 -05:00
|
|
|
HandleResponse []caddyhttp.ResponseHandler `json:"handle_response,omitempty"`
|
|
|
|
|
2022-03-06 19:43:39 -05:00
|
|
|
Transport http.RoundTripper `json:"-"`
|
|
|
|
CB CircuitBreaker `json:"-"`
|
|
|
|
DynamicUpstreams UpstreamSource `json:"-"`
|
2019-10-29 17:02:40 -05:00
|
|
|
|
2022-03-06 18:51:55 -05:00
|
|
|
// Holds the parsed CIDR ranges from TrustedProxies
|
2022-08-17 17:10:57 -05:00
|
|
|
trustedProxies []netip.Prefix
|
2022-03-06 18:51:55 -05:00
|
|
|
|
2021-05-02 13:39:06 -05:00
|
|
|
// Holds the named response matchers from the Caddyfile while adapting
|
|
|
|
responseMatchers map[string]caddyhttp.ResponseMatcher
|
|
|
|
|
|
|
|
// Holds the handle_response Caddyfile tokens while adapting
|
|
|
|
handleResponseSegments []*caddyfile.Dispenser
|
|
|
|
|
2020-08-07 19:02:24 -05:00
|
|
|
ctx caddy.Context
|
2019-10-29 17:02:40 -05:00
|
|
|
logger *zap.Logger
|
2019-09-02 23:01:02 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
// CaddyModule returns the Caddy module information.
|
|
|
|
func (Handler) CaddyModule() caddy.ModuleInfo {
|
|
|
|
return caddy.ModuleInfo{
|
2019-12-10 15:36:46 -05:00
|
|
|
ID: "http.handlers.reverse_proxy",
|
|
|
|
New: func() caddy.Module { return new(Handler) },
|
2019-05-04 11:49:50 -05:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2019-09-03 17:56:09 -05:00
|
|
|
// Provision ensures that h is set up properly before use.
|
2019-09-02 23:01:02 -05:00
|
|
|
func (h *Handler) Provision(ctx caddy.Context) error {
|
2020-08-07 19:02:24 -05:00
|
|
|
h.ctx = ctx
|
2019-10-29 17:02:40 -05:00
|
|
|
h.logger = ctx.Logger(h)
|
|
|
|
|
2022-03-06 19:43:39 -05:00
|
|
|
// verify SRV compatibility - TODO: LookupSRV deprecated; will be removed
|
2020-10-01 15:05:39 -05:00
|
|
|
for i, v := range h.Upstreams {
|
|
|
|
if v.LookupSRV == "" {
|
|
|
|
continue
|
|
|
|
}
|
|
|
|
if h.HealthChecks != nil && h.HealthChecks.Active != nil {
|
|
|
|
return fmt.Errorf(`upstream: lookup_srv is incompatible with active health checks: %d: {"dial": %q, "lookup_srv": %q}`, i, v.Dial, v.LookupSRV)
|
|
|
|
}
|
|
|
|
if v.Dial != "" {
|
|
|
|
return fmt.Errorf(`upstream: specifying dial address is incompatible with lookup_srv: %d: {"dial": %q, "lookup_srv": %q}`, i, v.Dial, v.LookupSRV)
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2019-09-03 20:06:54 -05:00
|
|
|
// start by loading modules
|
2019-09-02 23:01:02 -05:00
|
|
|
if h.TransportRaw != nil {
|
2019-12-10 15:36:46 -05:00
|
|
|
mod, err := ctx.LoadModule(h, "TransportRaw")
|
2019-09-02 23:01:02 -05:00
|
|
|
if err != nil {
|
2019-12-10 15:36:46 -05:00
|
|
|
return fmt.Errorf("loading transport: %v", err)
|
2019-05-04 11:49:50 -05:00
|
|
|
}
|
2019-12-10 15:36:46 -05:00
|
|
|
h.Transport = mod.(http.RoundTripper)
|
2019-09-02 23:01:02 -05:00
|
|
|
}
|
|
|
|
if h.LoadBalancing != nil && h.LoadBalancing.SelectionPolicyRaw != nil {
|
2019-12-10 15:36:46 -05:00
|
|
|
mod, err := ctx.LoadModule(h.LoadBalancing, "SelectionPolicyRaw")
|
2019-09-02 23:01:02 -05:00
|
|
|
if err != nil {
|
2019-12-10 15:36:46 -05:00
|
|
|
return fmt.Errorf("loading load balancing selection policy: %s", err)
|
2019-05-04 11:49:50 -05:00
|
|
|
}
|
2019-12-10 15:36:46 -05:00
|
|
|
h.LoadBalancing.SelectionPolicy = mod.(Selector)
|
2019-05-04 11:49:50 -05:00
|
|
|
}
|
2019-09-03 20:06:54 -05:00
|
|
|
if h.CBRaw != nil {
|
2019-12-10 15:36:46 -05:00
|
|
|
mod, err := ctx.LoadModule(h, "CBRaw")
|
2019-09-03 20:06:54 -05:00
|
|
|
if err != nil {
|
2019-12-10 15:36:46 -05:00
|
|
|
return fmt.Errorf("loading circuit breaker: %s", err)
|
2019-09-03 20:06:54 -05:00
|
|
|
}
|
2019-12-10 15:36:46 -05:00
|
|
|
h.CB = mod.(CircuitBreaker)
|
2019-09-03 20:06:54 -05:00
|
|
|
}
|
2022-03-06 19:43:39 -05:00
|
|
|
if h.DynamicUpstreamsRaw != nil {
|
|
|
|
mod, err := ctx.LoadModule(h, "DynamicUpstreamsRaw")
|
|
|
|
if err != nil {
|
|
|
|
return fmt.Errorf("loading upstream source module: %v", err)
|
|
|
|
}
|
|
|
|
h.DynamicUpstreams = mod.(UpstreamSource)
|
|
|
|
}
|
2019-05-04 11:49:50 -05:00
|
|
|
|
2022-03-06 18:51:55 -05:00
|
|
|
// parse trusted proxy CIDRs ahead of time
|
|
|
|
for _, str := range h.TrustedProxies {
|
|
|
|
if strings.Contains(str, "/") {
|
2022-08-17 17:10:57 -05:00
|
|
|
ipNet, err := netip.ParsePrefix(str)
|
2022-03-06 18:51:55 -05:00
|
|
|
if err != nil {
|
2022-08-17 17:10:57 -05:00
|
|
|
return fmt.Errorf("parsing CIDR expression: '%s': %v", str, err)
|
2022-03-06 18:51:55 -05:00
|
|
|
}
|
|
|
|
h.trustedProxies = append(h.trustedProxies, ipNet)
|
|
|
|
} else {
|
2022-08-17 17:10:57 -05:00
|
|
|
ipAddr, err := netip.ParseAddr(str)
|
|
|
|
if err != nil {
|
|
|
|
return fmt.Errorf("invalid IP address: '%s': %v", str, err)
|
2022-03-06 18:51:55 -05:00
|
|
|
}
|
2022-08-17 17:10:57 -05:00
|
|
|
ipNew := netip.PrefixFrom(ipAddr, ipAddr.BitLen())
|
|
|
|
h.trustedProxies = append(h.trustedProxies, ipNew)
|
2022-03-06 18:51:55 -05:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2020-04-22 20:57:06 -05:00
|
|
|
// ensure any embedded headers handler module gets provisioned
|
|
|
|
// (see https://caddy.community/t/set-cookie-manipulation-in-reverse-proxy/7666?u=matt
|
|
|
|
// for what happens if we forget to provision it)
|
|
|
|
if h.Headers != nil {
|
|
|
|
err := h.Headers.Provision(ctx)
|
|
|
|
if err != nil {
|
|
|
|
return fmt.Errorf("provisioning embedded headers handler: %v", err)
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2022-05-06 09:50:26 -05:00
|
|
|
if h.Rewrite != nil {
|
|
|
|
err := h.Rewrite.Provision(ctx)
|
|
|
|
if err != nil {
|
|
|
|
return fmt.Errorf("provisioning rewrite: %v", err)
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2019-10-11 15:25:39 -05:00
|
|
|
// set up transport
|
2019-09-02 23:01:02 -05:00
|
|
|
if h.Transport == nil {
|
2021-11-24 01:32:25 -05:00
|
|
|
t := &HTTPTransport{}
|
2019-09-05 14:14:39 -05:00
|
|
|
err := t.Provision(ctx)
|
|
|
|
if err != nil {
|
|
|
|
return fmt.Errorf("provisioning default transport: %v", err)
|
|
|
|
}
|
|
|
|
h.Transport = t
|
2019-05-04 11:49:50 -05:00
|
|
|
}
|
|
|
|
|
2019-10-11 15:25:39 -05:00
|
|
|
// set up load balancing
|
2019-09-02 23:01:02 -05:00
|
|
|
if h.LoadBalancing == nil {
|
|
|
|
h.LoadBalancing = new(LoadBalancing)
|
|
|
|
}
|
|
|
|
if h.LoadBalancing.SelectionPolicy == nil {
|
|
|
|
h.LoadBalancing.SelectionPolicy = RandomSelection{}
|
|
|
|
}
|
|
|
|
if h.LoadBalancing.TryDuration > 0 && h.LoadBalancing.TryInterval == 0 {
|
|
|
|
// a non-zero try_duration with a zero try_interval
|
|
|
|
// will always spin the CPU for try_duration if the
|
2019-09-03 13:10:11 -05:00
|
|
|
// upstream is local or low-latency; avoid that by
|
|
|
|
// defaulting to a sane wait period between attempts
|
2019-09-02 23:01:02 -05:00
|
|
|
h.LoadBalancing.TryInterval = caddy.Duration(250 * time.Millisecond)
|
2019-05-04 11:49:50 -05:00
|
|
|
}
|
2019-12-10 15:36:46 -05:00
|
|
|
lbMatcherSets, err := ctx.LoadModule(h.LoadBalancing, "RetryMatchRaw")
|
|
|
|
if err != nil {
|
|
|
|
return err
|
|
|
|
}
|
|
|
|
err = h.LoadBalancing.RetryMatch.FromInterface(lbMatcherSets)
|
2019-10-05 17:22:05 -05:00
|
|
|
if err != nil {
|
|
|
|
return err
|
|
|
|
}
|
2019-05-04 11:49:50 -05:00
|
|
|
|
2019-10-11 15:25:39 -05:00
|
|
|
// set up upstreams
|
2022-03-06 19:43:39 -05:00
|
|
|
for _, u := range h.Upstreams {
|
|
|
|
h.provisionUpstream(u)
|
2019-05-04 11:49:50 -05:00
|
|
|
}
|
2019-09-02 23:01:02 -05:00
|
|
|
|
2020-08-07 19:02:24 -05:00
|
|
|
if h.HealthChecks != nil {
|
|
|
|
// set defaults on passive health checks, if necessary
|
|
|
|
if h.HealthChecks.Passive != nil {
|
|
|
|
if h.HealthChecks.Passive.FailDuration > 0 && h.HealthChecks.Passive.MaxFails == 0 {
|
|
|
|
h.HealthChecks.Passive.MaxFails = 1
|
|
|
|
}
|
2020-02-23 16:30:52 -05:00
|
|
|
}
|
|
|
|
|
2020-08-07 19:02:24 -05:00
|
|
|
// if active health checks are enabled, configure them and start a worker
|
2021-03-29 19:36:40 -05:00
|
|
|
if h.HealthChecks.Active != nil && (h.HealthChecks.Active.Path != "" ||
|
|
|
|
h.HealthChecks.Active.URI != "" ||
|
|
|
|
h.HealthChecks.Active.Port != 0) {
|
|
|
|
|
2020-08-07 19:02:24 -05:00
|
|
|
h.HealthChecks.Active.logger = h.logger.Named("health_checker.active")
|
2020-02-23 16:30:52 -05:00
|
|
|
|
2020-08-07 19:02:24 -05:00
|
|
|
timeout := time.Duration(h.HealthChecks.Active.Timeout)
|
|
|
|
if timeout == 0 {
|
|
|
|
timeout = 5 * time.Second
|
|
|
|
}
|
2020-02-23 16:30:52 -05:00
|
|
|
|
2021-03-29 19:36:40 -05:00
|
|
|
if h.HealthChecks.Active.Path != "" {
|
|
|
|
h.HealthChecks.Active.logger.Warn("the 'path' option is deprecated, please use 'uri' instead!")
|
|
|
|
}
|
|
|
|
|
|
|
|
// parse the URI string (supports path and query)
|
|
|
|
if h.HealthChecks.Active.URI != "" {
|
|
|
|
parsedURI, err := url.Parse(h.HealthChecks.Active.URI)
|
|
|
|
if err != nil {
|
|
|
|
return err
|
|
|
|
}
|
|
|
|
h.HealthChecks.Active.uri = parsedURI
|
|
|
|
}
|
|
|
|
|
2020-08-07 19:02:24 -05:00
|
|
|
h.HealthChecks.Active.httpClient = &http.Client{
|
|
|
|
Timeout: timeout,
|
|
|
|
Transport: h.Transport,
|
|
|
|
}
|
|
|
|
|
2020-09-17 11:25:34 -05:00
|
|
|
for _, upstream := range h.Upstreams {
|
|
|
|
// if there's an alternative port for health-check provided in the config,
|
|
|
|
// then use it, otherwise use the port of upstream.
|
|
|
|
if h.HealthChecks.Active.Port != 0 {
|
|
|
|
upstream.activeHealthCheckPort = h.HealthChecks.Active.Port
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2020-08-07 19:02:24 -05:00
|
|
|
if h.HealthChecks.Active.Interval == 0 {
|
|
|
|
h.HealthChecks.Active.Interval = caddy.Duration(30 * time.Second)
|
|
|
|
}
|
|
|
|
|
|
|
|
if h.HealthChecks.Active.ExpectBody != "" {
|
|
|
|
var err error
|
|
|
|
h.HealthChecks.Active.bodyRegexp, err = regexp.Compile(h.HealthChecks.Active.ExpectBody)
|
|
|
|
if err != nil {
|
|
|
|
return fmt.Errorf("expect_body: compiling regular expression: %v", err)
|
|
|
|
}
|
2020-02-23 16:30:52 -05:00
|
|
|
}
|
|
|
|
|
2020-08-07 19:02:24 -05:00
|
|
|
go h.activeHealthChecker()
|
|
|
|
}
|
2020-02-23 16:30:52 -05:00
|
|
|
}
|
|
|
|
|
2020-05-27 11:15:20 -05:00
|
|
|
// set up any response routes
|
|
|
|
for i, rh := range h.HandleResponse {
|
|
|
|
err := rh.Provision(ctx)
|
|
|
|
if err != nil {
|
|
|
|
return fmt.Errorf("provisioning response handler %d: %v", i, err)
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2022-08-27 13:30:23 -05:00
|
|
|
upstreamHealthyUpdater := newMetricsUpstreamsHealthyUpdater(h)
|
|
|
|
upstreamHealthyUpdater.Init()
|
|
|
|
|
2019-09-02 23:01:02 -05:00
|
|
|
return nil
|
2019-05-04 11:49:50 -05:00
|
|
|
}
|
|
|
|
|
2019-09-03 13:10:11 -05:00
|
|
|
// Cleanup cleans up the resources made by h during provisioning.
|
2019-09-02 23:01:02 -05:00
|
|
|
func (h *Handler) Cleanup() error {
|
2019-10-10 16:38:30 -05:00
|
|
|
// TODO: Close keepalive connections on reload? https://github.com/caddyserver/caddy/pull/2507/files#diff-70219fd88fe3f36834f474ce6537ed26R762
|
|
|
|
|
2019-09-03 13:10:11 -05:00
|
|
|
// remove hosts from our config from the pool
|
2019-09-02 23:01:02 -05:00
|
|
|
for _, upstream := range h.Upstreams {
|
2020-11-22 16:50:29 -05:00
|
|
|
_, _ = hosts.Delete(upstream.String())
|
2019-05-04 11:49:50 -05:00
|
|
|
}
|
2019-09-03 13:10:11 -05:00
|
|
|
|
2019-09-02 23:01:02 -05:00
|
|
|
return nil
|
2019-05-04 11:49:50 -05:00
|
|
|
}
|
|
|
|
|
2019-09-02 23:01:02 -05:00
|
|
|
func (h *Handler) ServeHTTP(w http.ResponseWriter, r *http.Request, next caddyhttp.Handler) error {
|
2019-12-29 15:12:52 -05:00
|
|
|
repl := r.Context().Value(caddy.ReplacerCtxKey).(*caddy.Replacer)
|
2019-09-14 14:25:26 -05:00
|
|
|
|
2019-09-02 23:01:02 -05:00
|
|
|
// prepare the request for proxying; this is needed only once
|
2022-05-06 09:50:26 -05:00
|
|
|
clonedReq, err := h.prepareRequest(r, repl)
|
2019-09-02 23:01:02 -05:00
|
|
|
if err != nil {
|
|
|
|
return caddyhttp.Error(http.StatusInternalServerError,
|
|
|
|
fmt.Errorf("preparing request for upstream round-trip: %v", err))
|
2019-05-04 11:49:50 -05:00
|
|
|
}
|
|
|
|
|
2020-07-17 18:54:58 -05:00
|
|
|
// we will need the original headers and Host value if
|
2022-03-03 11:54:45 -05:00
|
|
|
// header operations are configured; this is so that each
|
|
|
|
// retry can apply the modifications, because placeholders
|
|
|
|
// may be used which depend on the selected upstream for
|
|
|
|
// their values
|
|
|
|
reqHost := clonedReq.Host
|
|
|
|
reqHeader := clonedReq.Header
|
2019-09-14 14:25:26 -05:00
|
|
|
|
2019-09-02 23:01:02 -05:00
|
|
|
start := time.Now()
|
2021-02-22 13:57:21 -05:00
|
|
|
defer func() {
|
|
|
|
// total proxying duration, including time spent on LB and retries
|
|
|
|
repl.Set("http.reverse_proxy.duration", time.Since(start))
|
2022-04-11 14:04:05 -05:00
|
|
|
repl.Set("http.reverse_proxy.duration_ms", time.Since(start).Seconds()*1e3) // multiply seconds to preserve decimal (see #4666)
|
2021-02-22 13:57:21 -05:00
|
|
|
}()
|
2019-09-02 23:01:02 -05:00
|
|
|
|
2022-03-06 19:43:39 -05:00
|
|
|
// in the proxy loop, each iteration is an attempt to proxy the request,
|
|
|
|
// and because we may retry some number of times, carry over the error
|
|
|
|
// from previous tries because of the nuances of load balancing & retries
|
2019-09-02 23:01:02 -05:00
|
|
|
var proxyErr error
|
2022-07-13 15:15:00 -05:00
|
|
|
var retries int
|
2019-09-02 23:01:02 -05:00
|
|
|
for {
|
2022-03-06 19:43:39 -05:00
|
|
|
var done bool
|
2022-07-13 15:15:00 -05:00
|
|
|
done, proxyErr = h.proxyLoopIteration(clonedReq, r, w, proxyErr, start, retries, repl, reqHeader, reqHost, next)
|
2022-03-06 19:43:39 -05:00
|
|
|
if done {
|
|
|
|
break
|
2019-09-02 23:01:02 -05:00
|
|
|
}
|
2022-07-13 15:15:00 -05:00
|
|
|
retries++
|
2022-03-06 19:43:39 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
if proxyErr != nil {
|
|
|
|
return statusError(proxyErr)
|
|
|
|
}
|
2019-09-02 23:01:02 -05:00
|
|
|
|
2022-03-06 19:43:39 -05:00
|
|
|
return nil
|
|
|
|
}
|
|
|
|
|
|
|
|
// proxyLoopIteration implements an iteration of the proxy loop. Despite the enormous amount of local state
|
|
|
|
// that has to be passed in, we brought this into its own method so that we could run defer more easily.
|
|
|
|
// It returns true when the loop is done and should break; false otherwise. The error value returned should
|
|
|
|
// be assigned to the proxyErr value for the next iteration of the loop (or the error handled after break).
|
2022-07-13 15:15:00 -05:00
|
|
|
func (h *Handler) proxyLoopIteration(r *http.Request, origReq *http.Request, w http.ResponseWriter, proxyErr error, start time.Time, retries int,
|
2022-03-06 19:43:39 -05:00
|
|
|
repl *caddy.Replacer, reqHeader http.Header, reqHost string, next caddyhttp.Handler) (bool, error) {
|
|
|
|
// get the updated list of upstreams
|
|
|
|
upstreams := h.Upstreams
|
|
|
|
if h.DynamicUpstreams != nil {
|
|
|
|
dUpstreams, err := h.DynamicUpstreams.GetUpstreams(r)
|
2019-10-11 15:25:39 -05:00
|
|
|
if err != nil {
|
2022-03-06 19:43:39 -05:00
|
|
|
h.logger.Error("failed getting dynamic upstreams; falling back to static upstreams", zap.Error(err))
|
|
|
|
} else {
|
|
|
|
upstreams = dUpstreams
|
|
|
|
for _, dUp := range dUpstreams {
|
|
|
|
h.provisionUpstream(dUp)
|
|
|
|
}
|
|
|
|
h.logger.Debug("provisioned dynamic upstreams", zap.Int("count", len(dUpstreams)))
|
|
|
|
defer func() {
|
|
|
|
// these upstreams are dynamic, so they are only used for this iteration
|
|
|
|
// of the proxy loop; be sure to let them go away when we're done with them
|
|
|
|
for _, upstream := range dUpstreams {
|
|
|
|
_, _ = hosts.Delete(upstream.String())
|
|
|
|
}
|
|
|
|
}()
|
2019-10-11 15:25:39 -05:00
|
|
|
}
|
2022-03-06 19:43:39 -05:00
|
|
|
}
|
2019-10-11 15:25:39 -05:00
|
|
|
|
2022-03-06 19:43:39 -05:00
|
|
|
// choose an available upstream
|
|
|
|
upstream := h.LoadBalancing.SelectionPolicy.Select(upstreams, r, w)
|
|
|
|
if upstream == nil {
|
|
|
|
if proxyErr == nil {
|
2022-07-08 14:01:02 -05:00
|
|
|
proxyErr = caddyhttp.Error(http.StatusServiceUnavailable, fmt.Errorf("no upstreams available"))
|
2019-09-14 14:25:26 -05:00
|
|
|
}
|
2022-07-13 15:15:00 -05:00
|
|
|
if !h.LoadBalancing.tryAgain(h.ctx, start, retries, proxyErr, r) {
|
2022-03-06 19:43:39 -05:00
|
|
|
return true, proxyErr
|
2019-09-02 23:01:02 -05:00
|
|
|
}
|
2022-03-06 19:43:39 -05:00
|
|
|
return false, proxyErr
|
|
|
|
}
|
2019-09-02 23:01:02 -05:00
|
|
|
|
2022-03-06 19:43:39 -05:00
|
|
|
// the dial address may vary per-request if placeholders are
|
|
|
|
// used, so perform those replacements here; the resulting
|
|
|
|
// DialInfo struct should have valid network address syntax
|
|
|
|
dialInfo, err := upstream.fillDialInfo(r)
|
|
|
|
if err != nil {
|
|
|
|
return true, fmt.Errorf("making dial info: %v", err)
|
|
|
|
}
|
2020-05-27 11:15:20 -05:00
|
|
|
|
2022-03-06 19:43:39 -05:00
|
|
|
h.logger.Debug("selected upstream",
|
|
|
|
zap.String("dial", dialInfo.Address),
|
|
|
|
zap.Int("total_upstreams", len(upstreams)))
|
2019-09-02 23:01:02 -05:00
|
|
|
|
2022-03-06 19:43:39 -05:00
|
|
|
// attach to the request information about how to dial the upstream;
|
|
|
|
// this is necessary because the information cannot be sufficiently
|
|
|
|
// or satisfactorily represented in a URL
|
|
|
|
caddyhttp.SetVar(r.Context(), dialInfoVarKey, dialInfo)
|
|
|
|
|
|
|
|
// set placeholders with information about this upstream
|
|
|
|
repl.Set("http.reverse_proxy.upstream.address", dialInfo.String())
|
|
|
|
repl.Set("http.reverse_proxy.upstream.hostport", dialInfo.Address)
|
|
|
|
repl.Set("http.reverse_proxy.upstream.host", dialInfo.Host)
|
|
|
|
repl.Set("http.reverse_proxy.upstream.port", dialInfo.Port)
|
|
|
|
repl.Set("http.reverse_proxy.upstream.requests", upstream.Host.NumRequests())
|
|
|
|
repl.Set("http.reverse_proxy.upstream.max_requests", upstream.MaxRequests)
|
|
|
|
repl.Set("http.reverse_proxy.upstream.fails", upstream.Host.Fails())
|
|
|
|
|
|
|
|
// mutate request headers according to this upstream;
|
|
|
|
// because we're in a retry loop, we have to copy
|
|
|
|
// headers (and the r.Host value) from the original
|
|
|
|
// so that each retry is identical to the first
|
|
|
|
if h.Headers != nil && h.Headers.Request != nil {
|
|
|
|
r.Header = make(http.Header)
|
|
|
|
copyHeader(r.Header, reqHeader)
|
|
|
|
r.Host = reqHost
|
|
|
|
h.Headers.Request.ApplyToRequest(r)
|
2019-05-04 11:49:50 -05:00
|
|
|
}
|
|
|
|
|
2022-03-06 19:43:39 -05:00
|
|
|
// proxy the request to that upstream
|
2022-05-06 09:50:26 -05:00
|
|
|
proxyErr = h.reverseProxy(w, r, origReq, repl, dialInfo, next)
|
2022-06-03 15:13:47 -05:00
|
|
|
if proxyErr == nil || errors.Is(proxyErr, context.Canceled) {
|
2022-03-06 19:43:39 -05:00
|
|
|
// context.Canceled happens when the downstream client
|
|
|
|
// cancels the request, which is not our failure
|
|
|
|
return true, nil
|
|
|
|
}
|
|
|
|
|
|
|
|
// if the roundtrip was successful, don't retry the request or
|
|
|
|
// ding the health status of the upstream (an error can still
|
|
|
|
// occur after the roundtrip if, for example, a response handler
|
|
|
|
// after the roundtrip returns an error)
|
|
|
|
if succ, ok := proxyErr.(roundtripSucceeded); ok {
|
|
|
|
return true, succ.error
|
|
|
|
}
|
|
|
|
|
|
|
|
// remember this failure (if enabled)
|
|
|
|
h.countFailure(upstream)
|
|
|
|
|
|
|
|
// if we've tried long enough, break
|
2022-07-13 15:15:00 -05:00
|
|
|
if !h.LoadBalancing.tryAgain(h.ctx, start, retries, proxyErr, r) {
|
2022-03-06 19:43:39 -05:00
|
|
|
return true, proxyErr
|
|
|
|
}
|
|
|
|
|
|
|
|
return false, proxyErr
|
2019-09-02 23:01:02 -05:00
|
|
|
}
|
|
|
|
|
2022-03-03 11:54:45 -05:00
|
|
|
// prepareRequest clones req so that it can be safely modified without
|
|
|
|
// changing the original request or introducing data races. It then
|
|
|
|
// modifies it so that it is ready to be proxied, except for directing
|
|
|
|
// to a specific upstream. This method adjusts headers and other relevant
|
|
|
|
// properties of the cloned request and should be done just once (before
|
|
|
|
// proxying) regardless of proxy retries. This assumes that no mutations
|
|
|
|
// of the cloned request are performed by h during or after proxying.
|
2022-05-06 09:50:26 -05:00
|
|
|
func (h Handler) prepareRequest(req *http.Request, repl *caddy.Replacer) (*http.Request, error) {
|
2022-03-03 11:54:45 -05:00
|
|
|
req = cloneRequest(req)
|
|
|
|
|
2022-05-06 09:50:26 -05:00
|
|
|
// if enabled, perform rewrites on the cloned request; if
|
|
|
|
// the method is GET or HEAD, prevent the request body
|
|
|
|
// from being copied to the upstream
|
|
|
|
if h.Rewrite != nil {
|
|
|
|
changed := h.Rewrite.Rewrite(req, repl)
|
|
|
|
if changed && (h.Rewrite.Method == "GET" || h.Rewrite.Method == "HEAD") {
|
|
|
|
req.ContentLength = 0
|
|
|
|
req.Body = nil
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2022-03-03 11:54:45 -05:00
|
|
|
// if enabled, buffer client request; this should only be
|
|
|
|
// enabled if the upstream requires it and does not work
|
|
|
|
// with "slow clients" (gunicorn, etc.) - this obviously
|
|
|
|
// has a perf overhead and makes the proxy at risk of
|
|
|
|
// exhausting memory and more susceptible to slowloris
|
|
|
|
// attacks, so it is strongly recommended to only use this
|
|
|
|
// feature if absolutely required, if read timeouts are
|
|
|
|
// set, and if body size is limited
|
2022-05-06 09:50:26 -05:00
|
|
|
if h.BufferRequests && req.Body != nil {
|
2022-03-03 11:54:45 -05:00
|
|
|
req.Body = h.bufferedBody(req.Body)
|
|
|
|
}
|
2019-09-05 14:14:39 -05:00
|
|
|
|
2019-05-04 11:49:50 -05:00
|
|
|
if req.ContentLength == 0 {
|
2019-09-02 23:01:02 -05:00
|
|
|
req.Body = nil // Issue golang/go#16036: nil Body for http.Transport retries
|
2019-05-04 11:49:50 -05:00
|
|
|
}
|
|
|
|
|
2019-09-02 23:01:02 -05:00
|
|
|
req.Close = false
|
2019-05-04 11:49:50 -05:00
|
|
|
|
2019-09-02 23:01:02 -05:00
|
|
|
// if User-Agent is not set by client, then explicitly
|
|
|
|
// disable it so it's not set to default value by std lib
|
|
|
|
if _, ok := req.Header["User-Agent"]; !ok {
|
|
|
|
req.Header.Set("User-Agent", "")
|
|
|
|
}
|
|
|
|
|
|
|
|
reqUpType := upgradeType(req.Header)
|
|
|
|
removeConnectionHeaders(req.Header)
|
2019-05-04 11:49:50 -05:00
|
|
|
|
|
|
|
// Remove hop-by-hop headers to the backend. Especially
|
|
|
|
// important is "Connection" because we want a persistent
|
|
|
|
// connection, regardless of what the client sent to us.
|
2021-06-04 16:21:02 -05:00
|
|
|
// Issue golang/go#46313: don't skip if field is empty.
|
2019-05-04 11:49:50 -05:00
|
|
|
for _, h := range hopHeaders {
|
2021-06-04 16:21:02 -05:00
|
|
|
// Issue golang/go#21096: tell backend applications that care about trailer support
|
|
|
|
// that we support trailers. (We do, but we don't go out of our way to
|
|
|
|
// advertise that unless the incoming client request thought it was worth
|
|
|
|
// mentioning.)
|
|
|
|
if h == "Te" && httpguts.HeaderValuesContainsToken(req.Header["Te"], "trailers") {
|
|
|
|
req.Header.Set("Te", "trailers")
|
2019-05-04 11:49:50 -05:00
|
|
|
continue
|
|
|
|
}
|
2019-09-02 23:01:02 -05:00
|
|
|
req.Header.Del(h)
|
2019-05-04 11:49:50 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
// After stripping all the hop-by-hop connection headers above, add back any
|
|
|
|
// necessary for protocol upgrades, such as for websockets.
|
|
|
|
if reqUpType != "" {
|
2019-09-02 23:01:02 -05:00
|
|
|
req.Header.Set("Connection", "Upgrade")
|
|
|
|
req.Header.Set("Upgrade", reqUpType)
|
2019-05-04 11:49:50 -05:00
|
|
|
}
|
|
|
|
|
2022-03-06 18:51:55 -05:00
|
|
|
// Add the supported X-Forwarded-* headers
|
|
|
|
err := h.addForwardedHeaders(req)
|
|
|
|
if err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
|
|
|
|
|
|
|
return req, nil
|
|
|
|
}
|
|
|
|
|
|
|
|
// addForwardedHeaders adds the de-facto standard X-Forwarded-*
|
|
|
|
// headers to the request before it is sent upstream.
|
|
|
|
//
|
|
|
|
// These headers are security sensitive, so care is taken to only
|
|
|
|
// use existing values for these headers from the incoming request
|
|
|
|
// if the client IP is trusted (i.e. coming from a trusted proxy
|
|
|
|
// sitting in front of this server). If the request didn't have
|
|
|
|
// the headers at all, then they will be added with the values
|
|
|
|
// that we can glean from the request.
|
|
|
|
func (h Handler) addForwardedHeaders(req *http.Request) error {
|
|
|
|
// Parse the remote IP, ignore the error as non-fatal,
|
|
|
|
// but the remote IP is required to continue, so we
|
|
|
|
// just return early. This should probably never happen
|
|
|
|
// though, unless some other module manipulated the request's
|
|
|
|
// remote address and used an invalid value.
|
|
|
|
clientIP, _, err := net.SplitHostPort(req.RemoteAddr)
|
|
|
|
if err != nil {
|
|
|
|
// Remove the `X-Forwarded-*` headers to avoid upstreams
|
|
|
|
// potentially trusting a header that came from the client
|
|
|
|
req.Header.Del("X-Forwarded-For")
|
|
|
|
req.Header.Del("X-Forwarded-Proto")
|
|
|
|
req.Header.Del("X-Forwarded-Host")
|
|
|
|
return nil
|
|
|
|
}
|
|
|
|
|
|
|
|
// Client IP may contain a zone if IPv6, so we need
|
|
|
|
// to pull that out before parsing the IP
|
2022-08-06 23:03:37 -05:00
|
|
|
if before, _, found := strings.Cut(clientIP, "%"); found {
|
|
|
|
clientIP = before
|
2022-03-06 18:51:55 -05:00
|
|
|
}
|
2022-08-17 17:10:57 -05:00
|
|
|
ipAddr, err := netip.ParseAddr(clientIP)
|
|
|
|
if err != nil {
|
|
|
|
return fmt.Errorf("invalid IP address: '%s': %v", clientIP, err)
|
2019-05-04 11:49:50 -05:00
|
|
|
}
|
|
|
|
|
2022-03-06 18:51:55 -05:00
|
|
|
// Check if the client is a trusted proxy
|
|
|
|
trusted := false
|
|
|
|
for _, ipRange := range h.trustedProxies {
|
2022-08-17 17:10:57 -05:00
|
|
|
if ipRange.Contains(ipAddr) {
|
2022-03-06 18:51:55 -05:00
|
|
|
trusted = true
|
|
|
|
break
|
2020-05-20 12:33:17 -05:00
|
|
|
}
|
2022-03-06 18:51:55 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
// If we aren't the first proxy, and the proxy is trusted,
|
|
|
|
// retain prior X-Forwarded-For information as a comma+space
|
|
|
|
// separated list and fold multiple headers into one.
|
|
|
|
clientXFF := clientIP
|
|
|
|
prior, ok, omit := allHeaderValues(req.Header, "X-Forwarded-For")
|
|
|
|
if trusted && ok && prior != "" {
|
|
|
|
clientXFF = prior + ", " + clientXFF
|
|
|
|
}
|
|
|
|
if !omit {
|
|
|
|
req.Header.Set("X-Forwarded-For", clientXFF)
|
|
|
|
}
|
|
|
|
|
|
|
|
// Set X-Forwarded-Proto; many backend apps expect this,
|
|
|
|
// so that they can properly craft URLs with the right
|
|
|
|
// scheme to match the original request
|
|
|
|
proto := "https"
|
|
|
|
if req.TLS == nil {
|
|
|
|
proto = "http"
|
|
|
|
}
|
|
|
|
prior, ok, omit = lastHeaderValue(req.Header, "X-Forwarded-Proto")
|
|
|
|
if trusted && ok && prior != "" {
|
|
|
|
proto = prior
|
|
|
|
}
|
|
|
|
if !omit {
|
2020-05-20 12:33:17 -05:00
|
|
|
req.Header.Set("X-Forwarded-Proto", proto)
|
2020-04-17 10:53:06 -05:00
|
|
|
}
|
|
|
|
|
2022-03-06 18:51:55 -05:00
|
|
|
// Set X-Forwarded-Host; often this is redundant because
|
|
|
|
// we pass through the request Host as-is, but in situations
|
|
|
|
// where we proxy over HTTPS, the user may need to override
|
|
|
|
// Host themselves, so it's helpful to send the original too.
|
2022-05-06 09:50:26 -05:00
|
|
|
host := req.Host
|
2022-03-06 18:51:55 -05:00
|
|
|
prior, ok, omit = lastHeaderValue(req.Header, "X-Forwarded-Host")
|
|
|
|
if trusted && ok && prior != "" {
|
|
|
|
host = prior
|
|
|
|
}
|
|
|
|
if !omit {
|
|
|
|
req.Header.Set("X-Forwarded-Host", host)
|
|
|
|
}
|
|
|
|
|
|
|
|
return nil
|
2019-09-02 23:01:02 -05:00
|
|
|
}
|
|
|
|
|
2019-09-03 17:56:09 -05:00
|
|
|
// reverseProxy performs a round-trip to the given backend and processes the response with the client.
|
|
|
|
// (This method is mostly the beginning of what was borrowed from the net/http/httputil package in the
|
|
|
|
// Go standard library which was used as the foundation.)
|
2022-05-06 09:50:26 -05:00
|
|
|
func (h *Handler) reverseProxy(rw http.ResponseWriter, req *http.Request, origReq *http.Request, repl *caddy.Replacer, di DialInfo, next caddyhttp.Handler) error {
|
2022-03-06 19:43:39 -05:00
|
|
|
_ = di.Upstream.Host.countRequest(1)
|
2020-11-22 16:50:29 -05:00
|
|
|
//nolint:errcheck
|
2022-03-06 19:43:39 -05:00
|
|
|
defer di.Upstream.Host.countRequest(-1)
|
2019-09-02 23:01:02 -05:00
|
|
|
|
|
|
|
// point the request to this upstream
|
2019-10-11 15:25:39 -05:00
|
|
|
h.directRequest(req, di)
|
2019-09-02 23:01:02 -05:00
|
|
|
|
2021-12-02 15:26:24 -05:00
|
|
|
server := req.Context().Value(caddyhttp.ServerCtxKey).(*caddyhttp.Server)
|
|
|
|
shouldLogCredentials := server.Logs != nil && server.Logs.ShouldLogCredentials
|
|
|
|
|
2022-08-09 11:53:24 -05:00
|
|
|
if supports1xx {
|
|
|
|
// Forward 1xx status codes, backported from https://github.com/golang/go/pull/53164
|
|
|
|
trace := &httptrace.ClientTrace{
|
|
|
|
Got1xxResponse: func(code int, header textproto.MIMEHeader) error {
|
|
|
|
h := rw.Header()
|
|
|
|
copyHeader(h, http.Header(header))
|
|
|
|
rw.WriteHeader(code)
|
|
|
|
|
|
|
|
// Clear headers, it's not automatically done by ResponseWriter.WriteHeader() for 1xx responses
|
|
|
|
for k := range h {
|
|
|
|
delete(h, k)
|
|
|
|
}
|
|
|
|
|
|
|
|
return nil
|
|
|
|
},
|
|
|
|
}
|
|
|
|
req = req.WithContext(httptrace.WithClientTrace(req.Context(), trace))
|
|
|
|
}
|
|
|
|
|
2022-08-12 14:15:41 -05:00
|
|
|
// if FlushInterval is explicitly configured to -1 (i.e. flush continuously to achieve
|
|
|
|
// low-latency streaming), don't let the transport cancel the request if the client
|
|
|
|
// disconnects: user probably wants us to finish sending the data to the upstream
|
|
|
|
// regardless, and we should expect client disconnection in low-latency streaming
|
|
|
|
// scenarios (see issue #4922)
|
|
|
|
if h.FlushInterval == -1 {
|
|
|
|
req = req.WithContext(ignoreClientGoneContext{req.Context(), h.ctx.Done()})
|
|
|
|
}
|
|
|
|
|
2020-05-18 15:08:11 -05:00
|
|
|
// do the round-trip; emit debug log with values we know are
|
|
|
|
// safe, or if there is no error, emit fuller log entry
|
2019-09-02 23:01:02 -05:00
|
|
|
start := time.Now()
|
|
|
|
res, err := h.Transport.RoundTrip(req)
|
2019-12-10 15:36:46 -05:00
|
|
|
duration := time.Since(start)
|
2020-05-18 15:08:11 -05:00
|
|
|
logger := h.logger.With(
|
2020-03-27 15:29:01 -05:00
|
|
|
zap.String("upstream", di.Upstream.String()),
|
2021-09-24 19:29:23 -05:00
|
|
|
zap.Duration("duration", duration),
|
2021-12-02 15:26:24 -05:00
|
|
|
zap.Object("request", caddyhttp.LoggableHTTPRequest{
|
|
|
|
Request: req,
|
|
|
|
ShouldLogCredentials: shouldLogCredentials,
|
|
|
|
}),
|
2021-05-05 15:52:18 -05:00
|
|
|
)
|
2020-05-18 14:50:40 -05:00
|
|
|
if err != nil {
|
2021-09-27 13:06:06 -05:00
|
|
|
logger.Debug("upstream roundtrip", zap.Error(err))
|
2020-05-18 14:50:40 -05:00
|
|
|
return err
|
|
|
|
}
|
2020-05-18 15:08:11 -05:00
|
|
|
logger.Debug("upstream roundtrip",
|
2021-12-02 15:26:24 -05:00
|
|
|
zap.Object("headers", caddyhttp.LoggableHTTPHeader{
|
|
|
|
Header: res.Header,
|
|
|
|
ShouldLogCredentials: shouldLogCredentials,
|
|
|
|
}),
|
2020-05-27 11:15:20 -05:00
|
|
|
zap.Int("status", res.StatusCode))
|
2019-11-05 18:27:51 -05:00
|
|
|
|
2021-02-22 13:57:21 -05:00
|
|
|
// duration until upstream wrote response headers (roundtrip duration)
|
|
|
|
repl.Set("http.reverse_proxy.upstream.latency", duration)
|
2022-04-11 14:04:05 -05:00
|
|
|
repl.Set("http.reverse_proxy.upstream.latency_ms", duration.Seconds()*1e3) // multiply seconds to preserve decimal (see #4666)
|
2021-02-22 13:57:21 -05:00
|
|
|
|
2019-09-03 20:06:54 -05:00
|
|
|
// update circuit breaker on current conditions
|
2019-10-11 15:25:39 -05:00
|
|
|
if di.Upstream.cb != nil {
|
2019-12-10 15:36:46 -05:00
|
|
|
di.Upstream.cb.RecordMetric(res.StatusCode, duration)
|
2019-09-03 20:06:54 -05:00
|
|
|
}
|
|
|
|
|
2019-09-02 23:01:02 -05:00
|
|
|
// perform passive health checks (if enabled)
|
|
|
|
if h.HealthChecks != nil && h.HealthChecks.Passive != nil {
|
|
|
|
// strike if the status code matches one that is "bad"
|
|
|
|
for _, badStatus := range h.HealthChecks.Passive.UnhealthyStatus {
|
|
|
|
if caddyhttp.StatusCodeMatches(res.StatusCode, badStatus) {
|
2019-10-11 15:25:39 -05:00
|
|
|
h.countFailure(di.Upstream)
|
2019-09-02 23:01:02 -05:00
|
|
|
}
|
2019-05-04 11:49:50 -05:00
|
|
|
}
|
|
|
|
|
2019-09-02 23:01:02 -05:00
|
|
|
// strike if the roundtrip took too long
|
|
|
|
if h.HealthChecks.Passive.UnhealthyLatency > 0 &&
|
2019-12-10 15:36:46 -05:00
|
|
|
duration >= time.Duration(h.HealthChecks.Passive.UnhealthyLatency) {
|
2019-10-11 15:25:39 -05:00
|
|
|
h.countFailure(di.Upstream)
|
2019-09-02 23:01:02 -05:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2021-02-09 16:15:04 -05:00
|
|
|
// if enabled, buffer the response body
|
|
|
|
if h.BufferResponses {
|
|
|
|
res.Body = h.bufferedBody(res.Body)
|
|
|
|
}
|
|
|
|
|
2020-06-04 13:06:38 -05:00
|
|
|
// see if any response handler is configured for this response from the backend
|
2020-05-27 11:15:20 -05:00
|
|
|
for i, rh := range h.HandleResponse {
|
2020-06-04 13:06:38 -05:00
|
|
|
if rh.Match != nil && !rh.Match.Match(res.StatusCode, res.Header) {
|
2020-05-27 11:15:20 -05:00
|
|
|
continue
|
|
|
|
}
|
2020-06-04 13:06:38 -05:00
|
|
|
|
2022-06-22 14:10:14 -05:00
|
|
|
// if configured to only change the status code,
|
|
|
|
// do that then continue regular proxy response
|
2020-06-04 13:06:38 -05:00
|
|
|
if statusCodeStr := rh.StatusCode.String(); statusCodeStr != "" {
|
|
|
|
statusCode, err := strconv.Atoi(repl.ReplaceAll(statusCodeStr, ""))
|
|
|
|
if err != nil {
|
|
|
|
return caddyhttp.Error(http.StatusInternalServerError, err)
|
|
|
|
}
|
|
|
|
if statusCode != 0 {
|
|
|
|
res.StatusCode = statusCode
|
|
|
|
}
|
|
|
|
break
|
|
|
|
}
|
|
|
|
|
|
|
|
// otherwise, if there are any routes configured, execute those as the
|
|
|
|
// actual response instead of what we got from the proxy backend
|
|
|
|
if len(rh.Routes) == 0 {
|
2020-05-27 11:15:20 -05:00
|
|
|
continue
|
|
|
|
}
|
2021-05-12 15:19:08 -05:00
|
|
|
|
|
|
|
// set up the replacer so that parts of the original response can be
|
|
|
|
// used for routing decisions
|
|
|
|
for field, value := range res.Header {
|
|
|
|
repl.Set("http.reverse_proxy.header."+field, strings.Join(value, ","))
|
|
|
|
}
|
2020-05-27 11:15:20 -05:00
|
|
|
repl.Set("http.reverse_proxy.status_code", res.StatusCode)
|
|
|
|
repl.Set("http.reverse_proxy.status_text", res.Status)
|
2021-05-12 15:19:08 -05:00
|
|
|
|
2020-05-27 11:15:20 -05:00
|
|
|
h.logger.Debug("handling response", zap.Int("handler", i))
|
2021-10-18 13:00:43 -05:00
|
|
|
|
reverseproxy: copy_response and copy_response_headers for handle_response routes (#4391)
* reverseproxy: New `copy_response` handler for `handle_response` routes
Followup to #4298 and #4388.
This adds a new `copy_response` handler which may only be used in `reverse_proxy`'s `handle_response` routes, which can be used to actually copy the proxy response downstream.
Previously, if `handle_response` was used (with routes, not the status code mode), it was impossible to use the upstream's response body at all, because we would always close the body, expecting the routes to write a new body from scratch.
To implement this, I had to refactor `h.reverseProxy()` to move all the code that came after the `HandleResponse` loop into a new function. This new function `h.finalizeResponse()` takes care of preparing the response by removing extra headers, dealing with trailers, then copying the headers and body downstream.
Since basically what we want `copy_response` to do is invoke `h.finalizeResponse()` at a configurable point in time, we need to pass down the proxy handler, the response, and some other state via a new `req.WithContext(ctx)`. Wrapping a new context is pretty much the only way we have to jump a few layers in the HTTP middleware chain and let a handler pick up this information. Feels a bit dirty, but it works.
Also fixed a bug with the `http.reverse_proxy.upstream.duration` placeholder, it always had the same duration as `http.reverse_proxy.upstream.latency`, but the former was meant to be the time taken for the roundtrip _plus_ copying/writing the response.
* Delete the "Content-Length" header if we aren't copying
Fixes a bug where the Content-Length will mismatch the actual bytes written if we skipped copying the response, so we get a message like this when using curl:
```
curl: (18) transfer closed with 18 bytes remaining to read
```
To replicate:
```
{
admin off
debug
}
:8881 {
reverse_proxy 127.0.0.1:8882 {
@200 status 200
handle_response @200 {
header Foo bar
}
}
}
:8882 {
header Content-Type application/json
respond `{"hello": "world"}` 200
}
```
* Implement `copy_response_headers`, with include/exclude list support
* Apply suggestions from code review
Co-authored-by: Matt Holt <mholt@users.noreply.github.com>
2022-03-09 13:00:51 -05:00
|
|
|
// we make some data available via request context to child routes
|
|
|
|
// so that they may inherit some options and functions from the
|
|
|
|
// handler, and be able to copy the response.
|
2022-05-06 09:50:26 -05:00
|
|
|
// we use the original request here, so that any routes from 'next'
|
|
|
|
// see the original request rather than the proxy cloned request.
|
reverseproxy: copy_response and copy_response_headers for handle_response routes (#4391)
* reverseproxy: New `copy_response` handler for `handle_response` routes
Followup to #4298 and #4388.
This adds a new `copy_response` handler which may only be used in `reverse_proxy`'s `handle_response` routes, which can be used to actually copy the proxy response downstream.
Previously, if `handle_response` was used (with routes, not the status code mode), it was impossible to use the upstream's response body at all, because we would always close the body, expecting the routes to write a new body from scratch.
To implement this, I had to refactor `h.reverseProxy()` to move all the code that came after the `HandleResponse` loop into a new function. This new function `h.finalizeResponse()` takes care of preparing the response by removing extra headers, dealing with trailers, then copying the headers and body downstream.
Since basically what we want `copy_response` to do is invoke `h.finalizeResponse()` at a configurable point in time, we need to pass down the proxy handler, the response, and some other state via a new `req.WithContext(ctx)`. Wrapping a new context is pretty much the only way we have to jump a few layers in the HTTP middleware chain and let a handler pick up this information. Feels a bit dirty, but it works.
Also fixed a bug with the `http.reverse_proxy.upstream.duration` placeholder, it always had the same duration as `http.reverse_proxy.upstream.latency`, but the former was meant to be the time taken for the roundtrip _plus_ copying/writing the response.
* Delete the "Content-Length" header if we aren't copying
Fixes a bug where the Content-Length will mismatch the actual bytes written if we skipped copying the response, so we get a message like this when using curl:
```
curl: (18) transfer closed with 18 bytes remaining to read
```
To replicate:
```
{
admin off
debug
}
:8881 {
reverse_proxy 127.0.0.1:8882 {
@200 status 200
handle_response @200 {
header Foo bar
}
}
}
:8882 {
header Content-Type application/json
respond `{"hello": "world"}` 200
}
```
* Implement `copy_response_headers`, with include/exclude list support
* Apply suggestions from code review
Co-authored-by: Matt Holt <mholt@users.noreply.github.com>
2022-03-09 13:00:51 -05:00
|
|
|
hrc := &handleResponseContext{
|
|
|
|
handler: h,
|
|
|
|
response: res,
|
|
|
|
start: start,
|
|
|
|
logger: logger,
|
|
|
|
}
|
2022-05-06 09:50:26 -05:00
|
|
|
ctx := origReq.Context()
|
reverseproxy: copy_response and copy_response_headers for handle_response routes (#4391)
* reverseproxy: New `copy_response` handler for `handle_response` routes
Followup to #4298 and #4388.
This adds a new `copy_response` handler which may only be used in `reverse_proxy`'s `handle_response` routes, which can be used to actually copy the proxy response downstream.
Previously, if `handle_response` was used (with routes, not the status code mode), it was impossible to use the upstream's response body at all, because we would always close the body, expecting the routes to write a new body from scratch.
To implement this, I had to refactor `h.reverseProxy()` to move all the code that came after the `HandleResponse` loop into a new function. This new function `h.finalizeResponse()` takes care of preparing the response by removing extra headers, dealing with trailers, then copying the headers and body downstream.
Since basically what we want `copy_response` to do is invoke `h.finalizeResponse()` at a configurable point in time, we need to pass down the proxy handler, the response, and some other state via a new `req.WithContext(ctx)`. Wrapping a new context is pretty much the only way we have to jump a few layers in the HTTP middleware chain and let a handler pick up this information. Feels a bit dirty, but it works.
Also fixed a bug with the `http.reverse_proxy.upstream.duration` placeholder, it always had the same duration as `http.reverse_proxy.upstream.latency`, but the former was meant to be the time taken for the roundtrip _plus_ copying/writing the response.
* Delete the "Content-Length" header if we aren't copying
Fixes a bug where the Content-Length will mismatch the actual bytes written if we skipped copying the response, so we get a message like this when using curl:
```
curl: (18) transfer closed with 18 bytes remaining to read
```
To replicate:
```
{
admin off
debug
}
:8881 {
reverse_proxy 127.0.0.1:8882 {
@200 status 200
handle_response @200 {
header Foo bar
}
}
}
:8882 {
header Content-Type application/json
respond `{"hello": "world"}` 200
}
```
* Implement `copy_response_headers`, with include/exclude list support
* Apply suggestions from code review
Co-authored-by: Matt Holt <mholt@users.noreply.github.com>
2022-03-09 13:00:51 -05:00
|
|
|
ctx = context.WithValue(ctx, proxyHandleResponseContextCtxKey, hrc)
|
|
|
|
|
2021-10-18 13:00:43 -05:00
|
|
|
// pass the request through the response handler routes
|
2022-05-06 09:50:26 -05:00
|
|
|
routeErr := rh.Routes.Compile(next).ServeHTTP(rw, origReq.WithContext(ctx))
|
reverseproxy: copy_response and copy_response_headers for handle_response routes (#4391)
* reverseproxy: New `copy_response` handler for `handle_response` routes
Followup to #4298 and #4388.
This adds a new `copy_response` handler which may only be used in `reverse_proxy`'s `handle_response` routes, which can be used to actually copy the proxy response downstream.
Previously, if `handle_response` was used (with routes, not the status code mode), it was impossible to use the upstream's response body at all, because we would always close the body, expecting the routes to write a new body from scratch.
To implement this, I had to refactor `h.reverseProxy()` to move all the code that came after the `HandleResponse` loop into a new function. This new function `h.finalizeResponse()` takes care of preparing the response by removing extra headers, dealing with trailers, then copying the headers and body downstream.
Since basically what we want `copy_response` to do is invoke `h.finalizeResponse()` at a configurable point in time, we need to pass down the proxy handler, the response, and some other state via a new `req.WithContext(ctx)`. Wrapping a new context is pretty much the only way we have to jump a few layers in the HTTP middleware chain and let a handler pick up this information. Feels a bit dirty, but it works.
Also fixed a bug with the `http.reverse_proxy.upstream.duration` placeholder, it always had the same duration as `http.reverse_proxy.upstream.latency`, but the former was meant to be the time taken for the roundtrip _plus_ copying/writing the response.
* Delete the "Content-Length" header if we aren't copying
Fixes a bug where the Content-Length will mismatch the actual bytes written if we skipped copying the response, so we get a message like this when using curl:
```
curl: (18) transfer closed with 18 bytes remaining to read
```
To replicate:
```
{
admin off
debug
}
:8881 {
reverse_proxy 127.0.0.1:8882 {
@200 status 200
handle_response @200 {
header Foo bar
}
}
}
:8882 {
header Content-Type application/json
respond `{"hello": "world"}` 200
}
```
* Implement `copy_response_headers`, with include/exclude list support
* Apply suggestions from code review
Co-authored-by: Matt Holt <mholt@users.noreply.github.com>
2022-03-09 13:00:51 -05:00
|
|
|
|
2022-06-22 14:10:14 -05:00
|
|
|
// close the response body afterwards, since we don't need it anymore;
|
|
|
|
// either a route had 'copy_response' which already consumed the body,
|
|
|
|
// or some other terminal handler ran which doesn't need the response
|
|
|
|
// body after that point (e.g. 'file_server' for X-Accel-Redirect flow),
|
|
|
|
// or we fell through to subsequent handlers past this proxy
|
|
|
|
// (e.g. forward auth's 2xx response flow).
|
|
|
|
if !hrc.isFinalized {
|
|
|
|
res.Body.Close()
|
reverseproxy: copy_response and copy_response_headers for handle_response routes (#4391)
* reverseproxy: New `copy_response` handler for `handle_response` routes
Followup to #4298 and #4388.
This adds a new `copy_response` handler which may only be used in `reverse_proxy`'s `handle_response` routes, which can be used to actually copy the proxy response downstream.
Previously, if `handle_response` was used (with routes, not the status code mode), it was impossible to use the upstream's response body at all, because we would always close the body, expecting the routes to write a new body from scratch.
To implement this, I had to refactor `h.reverseProxy()` to move all the code that came after the `HandleResponse` loop into a new function. This new function `h.finalizeResponse()` takes care of preparing the response by removing extra headers, dealing with trailers, then copying the headers and body downstream.
Since basically what we want `copy_response` to do is invoke `h.finalizeResponse()` at a configurable point in time, we need to pass down the proxy handler, the response, and some other state via a new `req.WithContext(ctx)`. Wrapping a new context is pretty much the only way we have to jump a few layers in the HTTP middleware chain and let a handler pick up this information. Feels a bit dirty, but it works.
Also fixed a bug with the `http.reverse_proxy.upstream.duration` placeholder, it always had the same duration as `http.reverse_proxy.upstream.latency`, but the former was meant to be the time taken for the roundtrip _plus_ copying/writing the response.
* Delete the "Content-Length" header if we aren't copying
Fixes a bug where the Content-Length will mismatch the actual bytes written if we skipped copying the response, so we get a message like this when using curl:
```
curl: (18) transfer closed with 18 bytes remaining to read
```
To replicate:
```
{
admin off
debug
}
:8881 {
reverse_proxy 127.0.0.1:8882 {
@200 status 200
handle_response @200 {
header Foo bar
}
}
}
:8882 {
header Content-Type application/json
respond `{"hello": "world"}` 200
}
```
* Implement `copy_response_headers`, with include/exclude list support
* Apply suggestions from code review
Co-authored-by: Matt Holt <mholt@users.noreply.github.com>
2022-03-09 13:00:51 -05:00
|
|
|
}
|
2021-10-18 13:00:43 -05:00
|
|
|
|
2022-06-22 14:10:14 -05:00
|
|
|
// wrap any route error in roundtripSucceeded so caller knows that
|
|
|
|
// the roundtrip was successful and to not retry
|
2021-10-18 13:00:43 -05:00
|
|
|
if routeErr != nil {
|
2020-05-27 11:15:20 -05:00
|
|
|
return roundtripSucceeded{routeErr}
|
|
|
|
}
|
reverseproxy: copy_response and copy_response_headers for handle_response routes (#4391)
* reverseproxy: New `copy_response` handler for `handle_response` routes
Followup to #4298 and #4388.
This adds a new `copy_response` handler which may only be used in `reverse_proxy`'s `handle_response` routes, which can be used to actually copy the proxy response downstream.
Previously, if `handle_response` was used (with routes, not the status code mode), it was impossible to use the upstream's response body at all, because we would always close the body, expecting the routes to write a new body from scratch.
To implement this, I had to refactor `h.reverseProxy()` to move all the code that came after the `HandleResponse` loop into a new function. This new function `h.finalizeResponse()` takes care of preparing the response by removing extra headers, dealing with trailers, then copying the headers and body downstream.
Since basically what we want `copy_response` to do is invoke `h.finalizeResponse()` at a configurable point in time, we need to pass down the proxy handler, the response, and some other state via a new `req.WithContext(ctx)`. Wrapping a new context is pretty much the only way we have to jump a few layers in the HTTP middleware chain and let a handler pick up this information. Feels a bit dirty, but it works.
Also fixed a bug with the `http.reverse_proxy.upstream.duration` placeholder, it always had the same duration as `http.reverse_proxy.upstream.latency`, but the former was meant to be the time taken for the roundtrip _plus_ copying/writing the response.
* Delete the "Content-Length" header if we aren't copying
Fixes a bug where the Content-Length will mismatch the actual bytes written if we skipped copying the response, so we get a message like this when using curl:
```
curl: (18) transfer closed with 18 bytes remaining to read
```
To replicate:
```
{
admin off
debug
}
:8881 {
reverse_proxy 127.0.0.1:8882 {
@200 status 200
handle_response @200 {
header Foo bar
}
}
}
:8882 {
header Content-Type application/json
respond `{"hello": "world"}` 200
}
```
* Implement `copy_response_headers`, with include/exclude list support
* Apply suggestions from code review
Co-authored-by: Matt Holt <mholt@users.noreply.github.com>
2022-03-09 13:00:51 -05:00
|
|
|
|
2022-06-22 14:10:14 -05:00
|
|
|
// we're done handling the response, and we don't want to
|
|
|
|
// fall through to the default finalize/copy behaviour
|
|
|
|
return nil
|
2020-05-27 11:15:20 -05:00
|
|
|
}
|
|
|
|
|
2022-06-22 14:10:14 -05:00
|
|
|
// copy the response body and headers back to the upstream client
|
|
|
|
return h.finalizeResponse(rw, req, res, repl, start, logger)
|
reverseproxy: copy_response and copy_response_headers for handle_response routes (#4391)
* reverseproxy: New `copy_response` handler for `handle_response` routes
Followup to #4298 and #4388.
This adds a new `copy_response` handler which may only be used in `reverse_proxy`'s `handle_response` routes, which can be used to actually copy the proxy response downstream.
Previously, if `handle_response` was used (with routes, not the status code mode), it was impossible to use the upstream's response body at all, because we would always close the body, expecting the routes to write a new body from scratch.
To implement this, I had to refactor `h.reverseProxy()` to move all the code that came after the `HandleResponse` loop into a new function. This new function `h.finalizeResponse()` takes care of preparing the response by removing extra headers, dealing with trailers, then copying the headers and body downstream.
Since basically what we want `copy_response` to do is invoke `h.finalizeResponse()` at a configurable point in time, we need to pass down the proxy handler, the response, and some other state via a new `req.WithContext(ctx)`. Wrapping a new context is pretty much the only way we have to jump a few layers in the HTTP middleware chain and let a handler pick up this information. Feels a bit dirty, but it works.
Also fixed a bug with the `http.reverse_proxy.upstream.duration` placeholder, it always had the same duration as `http.reverse_proxy.upstream.latency`, but the former was meant to be the time taken for the roundtrip _plus_ copying/writing the response.
* Delete the "Content-Length" header if we aren't copying
Fixes a bug where the Content-Length will mismatch the actual bytes written if we skipped copying the response, so we get a message like this when using curl:
```
curl: (18) transfer closed with 18 bytes remaining to read
```
To replicate:
```
{
admin off
debug
}
:8881 {
reverse_proxy 127.0.0.1:8882 {
@200 status 200
handle_response @200 {
header Foo bar
}
}
}
:8882 {
header Content-Type application/json
respond `{"hello": "world"}` 200
}
```
* Implement `copy_response_headers`, with include/exclude list support
* Apply suggestions from code review
Co-authored-by: Matt Holt <mholt@users.noreply.github.com>
2022-03-09 13:00:51 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
// finalizeResponse prepares and copies the response.
|
|
|
|
func (h Handler) finalizeResponse(
|
|
|
|
rw http.ResponseWriter,
|
|
|
|
req *http.Request,
|
|
|
|
res *http.Response,
|
|
|
|
repl *caddy.Replacer,
|
|
|
|
start time.Time,
|
|
|
|
logger *zap.Logger,
|
|
|
|
) error {
|
2021-02-09 16:15:04 -05:00
|
|
|
// deal with 101 Switching Protocols responses: (WebSocket, h2c, etc)
|
2019-09-02 23:01:02 -05:00
|
|
|
if res.StatusCode == http.StatusSwitchingProtocols {
|
2021-05-05 15:52:18 -05:00
|
|
|
h.handleUpgradeResponse(logger, rw, req, res)
|
2019-09-02 23:01:02 -05:00
|
|
|
return nil
|
2019-05-04 11:49:50 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
removeConnectionHeaders(res.Header)
|
|
|
|
|
|
|
|
for _, h := range hopHeaders {
|
|
|
|
res.Header.Del(h)
|
|
|
|
}
|
|
|
|
|
2020-05-13 10:52:20 -05:00
|
|
|
// apply any response header operations
|
|
|
|
if h.Headers != nil && h.Headers.Response != nil {
|
|
|
|
if h.Headers.Response.Require == nil ||
|
|
|
|
h.Headers.Response.Require.Match(res.StatusCode, res.Header) {
|
|
|
|
h.Headers.Response.ApplyTo(res.Header, repl)
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2019-05-04 11:49:50 -05:00
|
|
|
copyHeader(rw.Header(), res.Header)
|
|
|
|
|
|
|
|
// The "Trailer" header isn't included in the Transport's response,
|
|
|
|
// at least for *http.Transport. Build it up from Trailer.
|
|
|
|
announcedTrailers := len(res.Trailer)
|
|
|
|
if announcedTrailers > 0 {
|
|
|
|
trailerKeys := make([]string, 0, len(res.Trailer))
|
|
|
|
for k := range res.Trailer {
|
|
|
|
trailerKeys = append(trailerKeys, k)
|
|
|
|
}
|
|
|
|
rw.Header().Add("Trailer", strings.Join(trailerKeys, ", "))
|
|
|
|
}
|
|
|
|
|
|
|
|
rw.WriteHeader(res.StatusCode)
|
2022-06-22 14:10:14 -05:00
|
|
|
|
|
|
|
err := h.copyResponse(rw, res.Body, h.flushInterval(req, res))
|
|
|
|
res.Body.Close() // close now, instead of defer, to populate res.Trailer
|
|
|
|
if err != nil {
|
|
|
|
// we're streaming the response and we've already written headers, so
|
|
|
|
// there's nothing an error handler can do to recover at this point;
|
|
|
|
// the standard lib's proxy panics at this point, but we'll just log
|
|
|
|
// the error and abort the stream here
|
|
|
|
h.logger.Error("aborting with incomplete response", zap.Error(err))
|
|
|
|
return nil
|
2019-05-04 11:49:50 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
if len(res.Trailer) > 0 {
|
|
|
|
// Force chunking if we saw a response trailer.
|
|
|
|
// This prevents net/http from calculating the length for short
|
|
|
|
// bodies and adding a Content-Length.
|
|
|
|
if fl, ok := rw.(http.Flusher); ok {
|
|
|
|
fl.Flush()
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2021-02-22 13:57:21 -05:00
|
|
|
// total duration spent proxying, including writing response body
|
reverseproxy: copy_response and copy_response_headers for handle_response routes (#4391)
* reverseproxy: New `copy_response` handler for `handle_response` routes
Followup to #4298 and #4388.
This adds a new `copy_response` handler which may only be used in `reverse_proxy`'s `handle_response` routes, which can be used to actually copy the proxy response downstream.
Previously, if `handle_response` was used (with routes, not the status code mode), it was impossible to use the upstream's response body at all, because we would always close the body, expecting the routes to write a new body from scratch.
To implement this, I had to refactor `h.reverseProxy()` to move all the code that came after the `HandleResponse` loop into a new function. This new function `h.finalizeResponse()` takes care of preparing the response by removing extra headers, dealing with trailers, then copying the headers and body downstream.
Since basically what we want `copy_response` to do is invoke `h.finalizeResponse()` at a configurable point in time, we need to pass down the proxy handler, the response, and some other state via a new `req.WithContext(ctx)`. Wrapping a new context is pretty much the only way we have to jump a few layers in the HTTP middleware chain and let a handler pick up this information. Feels a bit dirty, but it works.
Also fixed a bug with the `http.reverse_proxy.upstream.duration` placeholder, it always had the same duration as `http.reverse_proxy.upstream.latency`, but the former was meant to be the time taken for the roundtrip _plus_ copying/writing the response.
* Delete the "Content-Length" header if we aren't copying
Fixes a bug where the Content-Length will mismatch the actual bytes written if we skipped copying the response, so we get a message like this when using curl:
```
curl: (18) transfer closed with 18 bytes remaining to read
```
To replicate:
```
{
admin off
debug
}
:8881 {
reverse_proxy 127.0.0.1:8882 {
@200 status 200
handle_response @200 {
header Foo bar
}
}
}
:8882 {
header Content-Type application/json
respond `{"hello": "world"}` 200
}
```
* Implement `copy_response_headers`, with include/exclude list support
* Apply suggestions from code review
Co-authored-by: Matt Holt <mholt@users.noreply.github.com>
2022-03-09 13:00:51 -05:00
|
|
|
repl.Set("http.reverse_proxy.upstream.duration", time.Since(start))
|
2022-04-11 14:04:05 -05:00
|
|
|
repl.Set("http.reverse_proxy.upstream.duration_ms", time.Since(start).Seconds()*1e3)
|
2021-02-22 13:57:21 -05:00
|
|
|
|
2019-05-04 11:49:50 -05:00
|
|
|
if len(res.Trailer) == announcedTrailers {
|
|
|
|
copyHeader(rw.Header(), res.Trailer)
|
2019-09-02 23:01:02 -05:00
|
|
|
return nil
|
2019-05-04 11:49:50 -05:00
|
|
|
}
|
|
|
|
|
|
|
|
for k, vv := range res.Trailer {
|
|
|
|
k = http.TrailerPrefix + k
|
|
|
|
for _, v := range vv {
|
|
|
|
rw.Header().Add(k, v)
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2019-09-02 23:01:02 -05:00
|
|
|
return nil
|
2019-05-04 11:49:50 -05:00
|
|
|
}
|
|
|
|
|
2022-07-13 15:15:00 -05:00
|
|
|
// tryAgain takes the time that the handler was initially invoked,
|
|
|
|
// the amount of retries already performed, as well as any error
|
|
|
|
// currently obtained, and the request being tried, and returns
|
|
|
|
// true if another attempt should be made at proxying the request.
|
|
|
|
// If true is returned, it has already blocked long enough before
|
|
|
|
// the next retry (i.e. no more sleeping is needed). If false is
|
|
|
|
// returned, the handler should stop trying to proxy the request.
|
|
|
|
func (lb LoadBalancing) tryAgain(ctx caddy.Context, start time.Time, retries int, proxyErr error, req *http.Request) bool {
|
|
|
|
// no retries are configured
|
|
|
|
if lb.TryDuration == 0 && lb.Retries == 0 {
|
|
|
|
return false
|
|
|
|
}
|
|
|
|
|
2019-09-02 23:01:02 -05:00
|
|
|
// if we've tried long enough, break
|
2022-07-13 15:15:00 -05:00
|
|
|
if lb.TryDuration > 0 && time.Since(start) >= time.Duration(lb.TryDuration) {
|
|
|
|
return false
|
|
|
|
}
|
|
|
|
|
|
|
|
// if we've reached the retry limit, break
|
|
|
|
if lb.Retries > 0 && retries >= lb.Retries {
|
2019-09-02 23:01:02 -05:00
|
|
|
return false
|
2019-05-04 11:49:50 -05:00
|
|
|
}
|
2019-10-05 17:22:05 -05:00
|
|
|
|
|
|
|
// if the error occurred while dialing (i.e. a connection
|
|
|
|
// could not even be established to the upstream), then it
|
|
|
|
// should be safe to retry, since without a connection, no
|
|
|
|
// HTTP request can be transmitted; but if the error is not
|
|
|
|
// specifically a dialer error, we need to be careful
|
|
|
|
if _, ok := proxyErr.(DialError); proxyErr != nil && !ok {
|
|
|
|
// if the error occurred after a connection was established,
|
|
|
|
// we have to assume the upstream received the request, and
|
|
|
|
// retries need to be carefully decided, because some requests
|
|
|
|
// are not idempotent
|
|
|
|
if lb.RetryMatch == nil && req.Method != "GET" {
|
|
|
|
// by default, don't retry requests if they aren't GET
|
|
|
|
return false
|
|
|
|
}
|
|
|
|
if !lb.RetryMatch.AnyMatch(req) {
|
|
|
|
return false
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2022-07-13 15:15:00 -05:00
|
|
|
// fast path; if the interval is zero, we don't need to wait
|
|
|
|
if lb.TryInterval == 0 {
|
|
|
|
return true
|
|
|
|
}
|
|
|
|
|
2019-09-02 23:01:02 -05:00
|
|
|
// otherwise, wait and try the next available host
|
2022-01-04 14:14:18 -05:00
|
|
|
timer := time.NewTimer(time.Duration(lb.TryInterval))
|
2020-08-07 19:02:24 -05:00
|
|
|
select {
|
2022-01-04 14:14:18 -05:00
|
|
|
case <-timer.C:
|
2020-08-07 19:02:24 -05:00
|
|
|
return true
|
|
|
|
case <-ctx.Done():
|
2022-01-04 14:14:18 -05:00
|
|
|
if !timer.Stop() {
|
|
|
|
// if the timer has been stopped then read from the channel
|
|
|
|
<-timer.C
|
|
|
|
}
|
2020-08-07 19:02:24 -05:00
|
|
|
return false
|
|
|
|
}
|
2019-05-04 11:49:50 -05:00
|
|
|
}
|
|
|
|
|
2019-10-11 15:25:39 -05:00
|
|
|
// directRequest modifies only req.URL so that it points to the upstream
|
|
|
|
// in the given DialInfo. It must modify ONLY the request URL.
|
2022-03-03 11:54:45 -05:00
|
|
|
func (Handler) directRequest(req *http.Request, di DialInfo) {
|
2020-04-21 21:34:00 -05:00
|
|
|
// we need a host, so set the upstream's host address
|
|
|
|
reqHost := di.Address
|
2019-09-11 19:53:44 -05:00
|
|
|
|
2020-04-21 21:34:00 -05:00
|
|
|
// if the port equates to the scheme, strip the port because
|
|
|
|
// it's weird to make a request like http://example.com:80/.
|
|
|
|
if (req.URL.Scheme == "http" && di.Port == "80") ||
|
|
|
|
(req.URL.Scheme == "https" && di.Port == "443") {
|
|
|
|
reqHost = di.Host
|
2019-05-04 11:49:50 -05:00
|
|
|
}
|
2020-04-21 21:34:00 -05:00
|
|
|
|
|
|
|
req.URL.Host = reqHost
|
2019-05-04 11:49:50 -05:00
|
|
|
}
|
|
|
|
|
2022-03-06 19:43:39 -05:00
|
|
|
func (h Handler) provisionUpstream(upstream *Upstream) {
|
|
|
|
// create or get the host representation for this upstream
|
|
|
|
upstream.fillHost()
|
|
|
|
|
|
|
|
// give it the circuit breaker, if any
|
|
|
|
upstream.cb = h.CB
|
|
|
|
|
|
|
|
// if the passive health checker has a non-zero UnhealthyRequestCount
|
|
|
|
// but the upstream has no MaxRequests set (they are the same thing,
|
|
|
|
// but the passive health checker is a default value for for upstreams
|
|
|
|
// without MaxRequests), copy the value into this upstream, since the
|
|
|
|
// value in the upstream (MaxRequests) is what is used during
|
|
|
|
// availability checks
|
|
|
|
if h.HealthChecks != nil && h.HealthChecks.Passive != nil {
|
|
|
|
h.HealthChecks.Passive.logger = h.logger.Named("health_checker.passive")
|
|
|
|
if h.HealthChecks.Passive.UnhealthyRequestCount > 0 &&
|
|
|
|
upstream.MaxRequests == 0 {
|
|
|
|
upstream.MaxRequests = h.HealthChecks.Passive.UnhealthyRequestCount
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// upstreams need independent access to the passive
|
|
|
|
// health check policy because passive health checks
|
|
|
|
// run without access to h.
|
|
|
|
if h.HealthChecks != nil {
|
|
|
|
upstream.healthCheckPolicy = h.HealthChecks.Passive
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2021-02-09 16:15:04 -05:00
|
|
|
// bufferedBody reads originalBody into a buffer, then returns a reader for the buffer.
|
|
|
|
// Always close the return value when done with it, just like if it was the original body!
|
|
|
|
func (h Handler) bufferedBody(originalBody io.ReadCloser) io.ReadCloser {
|
|
|
|
buf := bufPool.Get().(*bytes.Buffer)
|
|
|
|
buf.Reset()
|
|
|
|
if h.MaxBufferSize > 0 {
|
|
|
|
n, err := io.CopyN(buf, originalBody, h.MaxBufferSize)
|
|
|
|
if err != nil || n == h.MaxBufferSize {
|
|
|
|
return bodyReadCloser{
|
|
|
|
Reader: io.MultiReader(buf, originalBody),
|
|
|
|
buf: buf,
|
|
|
|
body: originalBody,
|
|
|
|
}
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
_, _ = io.Copy(buf, originalBody)
|
|
|
|
}
|
|
|
|
originalBody.Close() // no point in keeping it open
|
|
|
|
return bodyReadCloser{
|
|
|
|
Reader: buf,
|
|
|
|
buf: buf,
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2022-03-03 11:54:45 -05:00
|
|
|
// cloneRequest makes a semi-deep clone of origReq.
|
|
|
|
//
|
|
|
|
// Most of this code is borrowed from the Go stdlib reverse proxy,
|
|
|
|
// but we make a shallow-ish clone the request (deep clone only
|
|
|
|
// the headers and URL) so we can avoid manipulating the original
|
|
|
|
// request when using it to proxy upstream. This prevents request
|
|
|
|
// corruption and data races.
|
|
|
|
func cloneRequest(origReq *http.Request) *http.Request {
|
|
|
|
req := new(http.Request)
|
|
|
|
*req = *origReq
|
|
|
|
if origReq.URL != nil {
|
|
|
|
newURL := new(url.URL)
|
|
|
|
*newURL = *origReq.URL
|
|
|
|
if origReq.URL.User != nil {
|
|
|
|
newURL.User = new(url.Userinfo)
|
|
|
|
*newURL.User = *origReq.URL.User
|
|
|
|
}
|
|
|
|
// sanitize the request URL; we expect it to not contain the
|
|
|
|
// scheme and host since those should be determined by r.TLS
|
|
|
|
// and r.Host respectively, but some clients may include it
|
|
|
|
// in the request-line, which is technically valid in HTTP,
|
|
|
|
// but breaks reverseproxy behaviour, overriding how the
|
|
|
|
// dialer will behave. See #4237 for context.
|
|
|
|
newURL.Scheme = ""
|
|
|
|
newURL.Host = ""
|
|
|
|
req.URL = newURL
|
|
|
|
}
|
|
|
|
if origReq.Header != nil {
|
|
|
|
req.Header = origReq.Header.Clone()
|
|
|
|
}
|
|
|
|
if origReq.Trailer != nil {
|
|
|
|
req.Trailer = origReq.Trailer.Clone()
|
|
|
|
}
|
|
|
|
return req
|
|
|
|
}
|
|
|
|
|
2019-09-02 23:01:02 -05:00
|
|
|
func copyHeader(dst, src http.Header) {
|
|
|
|
for k, vv := range src {
|
|
|
|
for _, v := range vv {
|
|
|
|
dst.Add(k, v)
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2022-03-06 18:51:55 -05:00
|
|
|
// allHeaderValues gets all values for a given header field,
|
|
|
|
// joined by a comma and space if more than one is set. If the
|
|
|
|
// header field is nil, then the omit is true, meaning some
|
|
|
|
// earlier logic in the server wanted to prevent this header from
|
|
|
|
// getting written at all. If the header is empty, then ok is
|
|
|
|
// false. Callers should still check that the value is not empty
|
|
|
|
// (the header field may be set but have an empty value).
|
|
|
|
func allHeaderValues(h http.Header, field string) (value string, ok bool, omit bool) {
|
|
|
|
values, ok := h[http.CanonicalHeaderKey(field)]
|
|
|
|
if ok && values == nil {
|
|
|
|
return "", true, true
|
|
|
|
}
|
|
|
|
if len(values) == 0 {
|
|
|
|
return "", false, false
|
|
|
|
}
|
|
|
|
return strings.Join(values, ", "), true, false
|
|
|
|
}
|
|
|
|
|
|
|
|
// lastHeaderValue gets the last value for a given header field
|
|
|
|
// if more than one is set. If the header field is nil, then
|
|
|
|
// the omit is true, meaning some earlier logic in the server
|
|
|
|
// wanted to prevent this header from getting written at all.
|
|
|
|
// If the header is empty, then ok is false. Callers should
|
|
|
|
// still check that the value is not empty (the header field
|
|
|
|
// may be set but have an empty value).
|
|
|
|
func lastHeaderValue(h http.Header, field string) (value string, ok bool, omit bool) {
|
|
|
|
values, ok := h[http.CanonicalHeaderKey(field)]
|
|
|
|
if ok && values == nil {
|
|
|
|
return "", true, true
|
|
|
|
}
|
|
|
|
if len(values) == 0 {
|
|
|
|
return "", false, false
|
|
|
|
}
|
|
|
|
return values[len(values)-1], true, false
|
|
|
|
}
|
|
|
|
|
2019-05-04 11:49:50 -05:00
|
|
|
func upgradeType(h http.Header) string {
|
|
|
|
if !httpguts.HeaderValuesContainsToken(h["Connection"], "Upgrade") {
|
|
|
|
return ""
|
|
|
|
}
|
|
|
|
return strings.ToLower(h.Get("Upgrade"))
|
|
|
|
}
|
|
|
|
|
2019-09-02 23:01:02 -05:00
|
|
|
// removeConnectionHeaders removes hop-by-hop headers listed in the "Connection" header of h.
|
|
|
|
// See RFC 7230, section 6.1
|
|
|
|
func removeConnectionHeaders(h http.Header) {
|
2021-08-12 11:48:24 -05:00
|
|
|
for _, f := range h["Connection"] {
|
|
|
|
for _, sf := range strings.Split(f, ",") {
|
|
|
|
if sf = textproto.TrimString(sf); sf != "" {
|
|
|
|
h.Del(sf)
|
2019-09-02 23:01:02 -05:00
|
|
|
}
|
|
|
|
}
|
2019-05-04 11:49:50 -05:00
|
|
|
}
|
2019-09-02 23:01:02 -05:00
|
|
|
}
|
|
|
|
|
2020-11-25 12:54:23 -05:00
|
|
|
// statusError returns an error value that has a status code.
|
|
|
|
func statusError(err error) error {
|
|
|
|
// errors proxying usually mean there is a problem with the upstream(s)
|
|
|
|
statusCode := http.StatusBadGateway
|
|
|
|
|
2022-06-03 15:13:47 -05:00
|
|
|
// timeout errors have a standard status code (see issue #4823)
|
|
|
|
if err, ok := err.(net.Error); ok && err.Timeout() {
|
|
|
|
statusCode = http.StatusGatewayTimeout
|
|
|
|
}
|
|
|
|
|
2020-11-25 12:54:23 -05:00
|
|
|
// if the client canceled the request (usually this means they closed
|
|
|
|
// the connection, so they won't see any response), we can report it
|
|
|
|
// as a client error (4xx) and not a server error (5xx); unfortunately
|
|
|
|
// the Go standard library, at least at time of writing in late 2020,
|
|
|
|
// obnoxiously wraps the exported, standard context.Canceled error with
|
|
|
|
// an unexported garbage value that we have to do a substring check for:
|
|
|
|
// https://github.com/golang/go/blob/6965b01ea248cabb70c3749fd218b36089a21efb/src/net/net.go#L416-L430
|
|
|
|
if errors.Is(err, context.Canceled) || strings.Contains(err.Error(), "operation was canceled") {
|
|
|
|
// regrettably, there is no standard error code for "client closed connection", but
|
|
|
|
// for historical reasons we can use a code that a lot of people are already using;
|
|
|
|
// using 5xx is problematic for users; see #3748
|
|
|
|
statusCode = 499
|
|
|
|
}
|
|
|
|
return caddyhttp.Error(statusCode, err)
|
|
|
|
}
|
|
|
|
|
2019-09-03 17:56:09 -05:00
|
|
|
// LoadBalancing has parameters related to load balancing.
|
2019-09-02 23:01:02 -05:00
|
|
|
type LoadBalancing struct {
|
2019-12-23 14:45:35 -05:00
|
|
|
// A selection policy is how to choose an available backend.
|
|
|
|
// The default policy is random selection.
|
|
|
|
SelectionPolicyRaw json.RawMessage `json:"selection_policy,omitempty" caddy:"namespace=http.reverse_proxy.selection_policies inline_key=policy"`
|
|
|
|
|
2022-07-13 15:15:00 -05:00
|
|
|
// How many times to retry selecting available backends for each
|
|
|
|
// request if the next available host is down. If try_duration is
|
|
|
|
// also configured, then retries may stop early if the duration
|
|
|
|
// is reached. By default, retries are disabled (zero).
|
|
|
|
Retries int `json:"retries,omitempty"`
|
|
|
|
|
2019-12-23 14:45:35 -05:00
|
|
|
// How long to try selecting available backends for each request
|
2022-07-13 15:15:00 -05:00
|
|
|
// if the next available host is down. Clients will wait for up
|
|
|
|
// to this long while the load balancer tries to find an available
|
|
|
|
// upstream host. If retries is also configured, tries may stop
|
|
|
|
// early if the maximum retries is reached. By default, retries
|
|
|
|
// are disabled (zero duration).
|
2019-12-23 14:45:35 -05:00
|
|
|
TryDuration caddy.Duration `json:"try_duration,omitempty"`
|
|
|
|
|
2022-07-13 15:15:00 -05:00
|
|
|
// How long to wait between selecting the next host from the pool.
|
|
|
|
// Default is 250ms if try_duration is enabled, otherwise zero. Only
|
|
|
|
// relevant when a request to an upstream host fails. Be aware that
|
|
|
|
// setting this to 0 with a non-zero try_duration can cause the CPU
|
|
|
|
// to spin if all backends are down and latency is very low.
|
2019-12-23 14:45:35 -05:00
|
|
|
TryInterval caddy.Duration `json:"try_interval,omitempty"`
|
|
|
|
|
|
|
|
// A list of matcher sets that restricts with which requests retries are
|
|
|
|
// allowed. A request must match any of the given matcher sets in order
|
|
|
|
// to be retried if the connection to the upstream succeeded but the
|
|
|
|
// subsequent round-trip failed. If the connection to the upstream failed,
|
|
|
|
// a retry is always allowed. If unspecified, only GET requests will be
|
|
|
|
// allowed to be retried. Note that a retry is done with the next available
|
|
|
|
// host according to the load balancing policy.
|
|
|
|
RetryMatchRaw caddyhttp.RawMatcherSets `json:"retry_match,omitempty" caddy:"namespace=http.matchers"`
|
2019-09-02 23:01:02 -05:00
|
|
|
|
2019-10-05 17:22:05 -05:00
|
|
|
SelectionPolicy Selector `json:"-"`
|
|
|
|
RetryMatch caddyhttp.MatcherSets `json:"-"`
|
2019-09-02 23:01:02 -05:00
|
|
|
}
|
|
|
|
|
2019-09-03 17:56:09 -05:00
|
|
|
// Selector selects an available upstream from the pool.
|
2019-09-02 23:01:02 -05:00
|
|
|
type Selector interface {
|
2020-11-20 14:39:26 -05:00
|
|
|
Select(UpstreamPool, *http.Request, http.ResponseWriter) *Upstream
|
2019-09-02 23:01:02 -05:00
|
|
|
}
|
|
|
|
|
2022-03-06 19:43:39 -05:00
|
|
|
// UpstreamSource gets the list of upstreams that can be used when
|
|
|
|
// proxying a request. Returned upstreams will be load balanced and
|
|
|
|
// health-checked. This should be a very fast function -- instant
|
|
|
|
// if possible -- and the return value must be as stable as possible.
|
|
|
|
// In other words, the list of upstreams should ideally not change much
|
|
|
|
// across successive calls. If the list of upstreams changes or the
|
|
|
|
// ordering is not stable, load balancing will suffer. This function
|
|
|
|
// may be called during each retry, multiple times per request, and as
|
|
|
|
// such, needs to be instantaneous. The returned slice will not be
|
|
|
|
// modified.
|
|
|
|
type UpstreamSource interface {
|
|
|
|
GetUpstreams(*http.Request) ([]*Upstream, error)
|
|
|
|
}
|
|
|
|
|
2019-09-02 23:01:02 -05:00
|
|
|
// Hop-by-hop headers. These are removed when sent to the backend.
|
|
|
|
// As of RFC 7230, hop-by-hop headers are required to appear in the
|
|
|
|
// Connection header field. These are the headers defined by the
|
|
|
|
// obsoleted RFC 2616 (section 13.5.1) and are used for backward
|
|
|
|
// compatibility.
|
|
|
|
var hopHeaders = []string{
|
2020-03-20 07:54:28 -05:00
|
|
|
"Alt-Svc",
|
2019-09-02 23:01:02 -05:00
|
|
|
"Connection",
|
|
|
|
"Proxy-Connection", // non-standard but still sent by libcurl and rejected by e.g. google
|
|
|
|
"Keep-Alive",
|
|
|
|
"Proxy-Authenticate",
|
|
|
|
"Proxy-Authorization",
|
|
|
|
"Te", // canonicalized version of "TE"
|
|
|
|
"Trailer", // not Trailers per URL above; https://www.rfc-editor.org/errata_search.php?eid=4522
|
|
|
|
"Transfer-Encoding",
|
|
|
|
"Upgrade",
|
|
|
|
}
|
|
|
|
|
2019-10-05 17:22:05 -05:00
|
|
|
// DialError is an error that specifically occurs
|
|
|
|
// in a call to Dial or DialContext.
|
2020-04-07 09:31:52 -05:00
|
|
|
type DialError struct{ error }
|
|
|
|
|
|
|
|
// TLSTransport is implemented by transports
|
|
|
|
// that are capable of using TLS.
|
|
|
|
type TLSTransport interface {
|
|
|
|
// TLSEnabled returns true if the transport
|
|
|
|
// has TLS enabled, false otherwise.
|
|
|
|
TLSEnabled() bool
|
|
|
|
|
|
|
|
// EnableTLS enables TLS within the transport
|
|
|
|
// if it is not already, using the provided
|
|
|
|
// value as a basis for the TLS config.
|
|
|
|
EnableTLS(base *TLSConfig) error
|
2019-10-05 17:22:05 -05:00
|
|
|
}
|
|
|
|
|
2020-05-27 11:15:20 -05:00
|
|
|
// roundtripSucceeded is an error type that is returned if the
|
|
|
|
// roundtrip succeeded, but an error occurred after-the-fact.
|
|
|
|
type roundtripSucceeded struct{ error }
|
|
|
|
|
2021-02-09 16:15:04 -05:00
|
|
|
// bodyReadCloser is a reader that, upon closing, will return
|
|
|
|
// its buffer to the pool and close the underlying body reader.
|
|
|
|
type bodyReadCloser struct {
|
|
|
|
io.Reader
|
|
|
|
buf *bytes.Buffer
|
|
|
|
body io.ReadCloser
|
|
|
|
}
|
|
|
|
|
|
|
|
func (brc bodyReadCloser) Close() error {
|
|
|
|
bufPool.Put(brc.buf)
|
|
|
|
if brc.body != nil {
|
|
|
|
return brc.body.Close()
|
|
|
|
}
|
|
|
|
return nil
|
|
|
|
}
|
|
|
|
|
|
|
|
// bufPool is used for buffering requests and responses.
|
2019-11-15 19:15:33 -05:00
|
|
|
var bufPool = sync.Pool{
|
2022-08-02 15:39:09 -05:00
|
|
|
New: func() any {
|
2019-11-15 19:15:33 -05:00
|
|
|
return new(bytes.Buffer)
|
|
|
|
},
|
|
|
|
}
|
2019-09-03 13:10:11 -05:00
|
|
|
|
reverseproxy: copy_response and copy_response_headers for handle_response routes (#4391)
* reverseproxy: New `copy_response` handler for `handle_response` routes
Followup to #4298 and #4388.
This adds a new `copy_response` handler which may only be used in `reverse_proxy`'s `handle_response` routes, which can be used to actually copy the proxy response downstream.
Previously, if `handle_response` was used (with routes, not the status code mode), it was impossible to use the upstream's response body at all, because we would always close the body, expecting the routes to write a new body from scratch.
To implement this, I had to refactor `h.reverseProxy()` to move all the code that came after the `HandleResponse` loop into a new function. This new function `h.finalizeResponse()` takes care of preparing the response by removing extra headers, dealing with trailers, then copying the headers and body downstream.
Since basically what we want `copy_response` to do is invoke `h.finalizeResponse()` at a configurable point in time, we need to pass down the proxy handler, the response, and some other state via a new `req.WithContext(ctx)`. Wrapping a new context is pretty much the only way we have to jump a few layers in the HTTP middleware chain and let a handler pick up this information. Feels a bit dirty, but it works.
Also fixed a bug with the `http.reverse_proxy.upstream.duration` placeholder, it always had the same duration as `http.reverse_proxy.upstream.latency`, but the former was meant to be the time taken for the roundtrip _plus_ copying/writing the response.
* Delete the "Content-Length" header if we aren't copying
Fixes a bug where the Content-Length will mismatch the actual bytes written if we skipped copying the response, so we get a message like this when using curl:
```
curl: (18) transfer closed with 18 bytes remaining to read
```
To replicate:
```
{
admin off
debug
}
:8881 {
reverse_proxy 127.0.0.1:8882 {
@200 status 200
handle_response @200 {
header Foo bar
}
}
}
:8882 {
header Content-Type application/json
respond `{"hello": "world"}` 200
}
```
* Implement `copy_response_headers`, with include/exclude list support
* Apply suggestions from code review
Co-authored-by: Matt Holt <mholt@users.noreply.github.com>
2022-03-09 13:00:51 -05:00
|
|
|
// handleResponseContext carries some contextual information about the
|
2022-08-27 15:39:26 -05:00
|
|
|
// current proxy handling.
|
reverseproxy: copy_response and copy_response_headers for handle_response routes (#4391)
* reverseproxy: New `copy_response` handler for `handle_response` routes
Followup to #4298 and #4388.
This adds a new `copy_response` handler which may only be used in `reverse_proxy`'s `handle_response` routes, which can be used to actually copy the proxy response downstream.
Previously, if `handle_response` was used (with routes, not the status code mode), it was impossible to use the upstream's response body at all, because we would always close the body, expecting the routes to write a new body from scratch.
To implement this, I had to refactor `h.reverseProxy()` to move all the code that came after the `HandleResponse` loop into a new function. This new function `h.finalizeResponse()` takes care of preparing the response by removing extra headers, dealing with trailers, then copying the headers and body downstream.
Since basically what we want `copy_response` to do is invoke `h.finalizeResponse()` at a configurable point in time, we need to pass down the proxy handler, the response, and some other state via a new `req.WithContext(ctx)`. Wrapping a new context is pretty much the only way we have to jump a few layers in the HTTP middleware chain and let a handler pick up this information. Feels a bit dirty, but it works.
Also fixed a bug with the `http.reverse_proxy.upstream.duration` placeholder, it always had the same duration as `http.reverse_proxy.upstream.latency`, but the former was meant to be the time taken for the roundtrip _plus_ copying/writing the response.
* Delete the "Content-Length" header if we aren't copying
Fixes a bug where the Content-Length will mismatch the actual bytes written if we skipped copying the response, so we get a message like this when using curl:
```
curl: (18) transfer closed with 18 bytes remaining to read
```
To replicate:
```
{
admin off
debug
}
:8881 {
reverse_proxy 127.0.0.1:8882 {
@200 status 200
handle_response @200 {
header Foo bar
}
}
}
:8882 {
header Content-Type application/json
respond `{"hello": "world"}` 200
}
```
* Implement `copy_response_headers`, with include/exclude list support
* Apply suggestions from code review
Co-authored-by: Matt Holt <mholt@users.noreply.github.com>
2022-03-09 13:00:51 -05:00
|
|
|
type handleResponseContext struct {
|
|
|
|
// handler is the active proxy handler instance, so that
|
|
|
|
// routes like copy_response may inherit some config
|
|
|
|
// options and have access to handler methods.
|
|
|
|
handler *Handler
|
|
|
|
|
|
|
|
// response is the actual response received from the proxy
|
|
|
|
// roundtrip, to potentially be copied if a copy_response
|
|
|
|
// handler is in the handle_response routes.
|
|
|
|
response *http.Response
|
|
|
|
|
|
|
|
// start is the time just before the proxy roundtrip was
|
|
|
|
// performed, used for logging.
|
|
|
|
start time.Time
|
|
|
|
|
|
|
|
// logger is the prepared logger which is used to write logs
|
|
|
|
// with the request, duration, and selected upstream attached.
|
|
|
|
logger *zap.Logger
|
|
|
|
|
|
|
|
// isFinalized is whether the response has been finalized,
|
|
|
|
// i.e. copied and closed, to make sure that it doesn't
|
|
|
|
// happen twice.
|
|
|
|
isFinalized bool
|
|
|
|
}
|
|
|
|
|
2022-08-12 14:15:41 -05:00
|
|
|
// ignoreClientGoneContext is a special context.Context type
|
|
|
|
// intended for use when doing a RoundTrip where you don't
|
|
|
|
// want a client disconnection to cancel the request during
|
|
|
|
// the roundtrip. Set its done field to a Done() channel
|
|
|
|
// of a context that doesn't get canceled when the client
|
|
|
|
// disconnects, such as caddy.Context.Done() instead.
|
|
|
|
type ignoreClientGoneContext struct {
|
|
|
|
context.Context
|
|
|
|
done <-chan struct{}
|
|
|
|
}
|
|
|
|
|
|
|
|
func (c ignoreClientGoneContext) Done() <-chan struct{} { return c.done }
|
|
|
|
|
reverseproxy: copy_response and copy_response_headers for handle_response routes (#4391)
* reverseproxy: New `copy_response` handler for `handle_response` routes
Followup to #4298 and #4388.
This adds a new `copy_response` handler which may only be used in `reverse_proxy`'s `handle_response` routes, which can be used to actually copy the proxy response downstream.
Previously, if `handle_response` was used (with routes, not the status code mode), it was impossible to use the upstream's response body at all, because we would always close the body, expecting the routes to write a new body from scratch.
To implement this, I had to refactor `h.reverseProxy()` to move all the code that came after the `HandleResponse` loop into a new function. This new function `h.finalizeResponse()` takes care of preparing the response by removing extra headers, dealing with trailers, then copying the headers and body downstream.
Since basically what we want `copy_response` to do is invoke `h.finalizeResponse()` at a configurable point in time, we need to pass down the proxy handler, the response, and some other state via a new `req.WithContext(ctx)`. Wrapping a new context is pretty much the only way we have to jump a few layers in the HTTP middleware chain and let a handler pick up this information. Feels a bit dirty, but it works.
Also fixed a bug with the `http.reverse_proxy.upstream.duration` placeholder, it always had the same duration as `http.reverse_proxy.upstream.latency`, but the former was meant to be the time taken for the roundtrip _plus_ copying/writing the response.
* Delete the "Content-Length" header if we aren't copying
Fixes a bug where the Content-Length will mismatch the actual bytes written if we skipped copying the response, so we get a message like this when using curl:
```
curl: (18) transfer closed with 18 bytes remaining to read
```
To replicate:
```
{
admin off
debug
}
:8881 {
reverse_proxy 127.0.0.1:8882 {
@200 status 200
handle_response @200 {
header Foo bar
}
}
}
:8882 {
header Content-Type application/json
respond `{"hello": "world"}` 200
}
```
* Implement `copy_response_headers`, with include/exclude list support
* Apply suggestions from code review
Co-authored-by: Matt Holt <mholt@users.noreply.github.com>
2022-03-09 13:00:51 -05:00
|
|
|
// proxyHandleResponseContextCtxKey is the context key for the active proxy handler
|
|
|
|
// so that handle_response routes can inherit some config options
|
|
|
|
// from the proxy handler.
|
|
|
|
const proxyHandleResponseContextCtxKey caddy.CtxKey = "reverse_proxy_handle_response_context"
|
|
|
|
|
2019-09-02 23:01:02 -05:00
|
|
|
// Interface guards
|
|
|
|
var (
|
|
|
|
_ caddy.Provisioner = (*Handler)(nil)
|
|
|
|
_ caddy.CleanerUpper = (*Handler)(nil)
|
2019-09-03 17:56:09 -05:00
|
|
|
_ caddyhttp.MiddlewareHandler = (*Handler)(nil)
|
2019-09-02 23:01:02 -05:00
|
|
|
)
|