]> git.ipfire.org Git - thirdparty/apache/httpd.git/commitdiff
mod_proxy_http: handle async tunneling of Upgrade(d) protocols.
authorYann Ylavic <ylavic@apache.org>
Thu, 2 Jul 2020 00:14:26 +0000 (00:14 +0000)
committerYann Ylavic <ylavic@apache.org>
Thu, 2 Jul 2020 00:14:26 +0000 (00:14 +0000)
When supported by the MPM (i.e. "event"), provide async callbacks and let
them be scheduled by ap_mpm_register_poll_callback_timeout(), while the
handler returns SUSPENDED.

The new ProxyAsyncDelay directive (if positive) enables async handling,
while ProxyAsyncIdleTimeout determines the timeout applied on both ends
while tunneling.

Github: closes #126

git-svn-id: https://svn.apache.org/repos/asf/httpd/httpd/trunk@1879419 13f79535-47bb-0310-9956-ffa450edef68

modules/proxy/mod_proxy.c
modules/proxy/mod_proxy.h
modules/proxy/mod_proxy_http.c

index d730a6d0477bbd6a623f8b0fc0b2a6d95bbfa06b..33060b08f240222a5b9db4d06d3597c052c1f3c4 100644 (file)
@@ -1842,6 +1842,7 @@ static void *create_proxy_dir_config(apr_pool_t *p, char *dummy)
     new->add_forwarded_headers_set = 0;
     new->forward_100_continue = 1;
     new->forward_100_continue_set = 0;
+    new->async_delay = -1;
 
     return (void *) new;
 }
@@ -1889,17 +1890,30 @@ static void *merge_proxy_dir_config(apr_pool_t *p, void *basev, void *addv)
     new->error_override_set = add->error_override_set || base->error_override_set;
     new->alias = (add->alias_set == 0) ? base->alias : add->alias;
     new->alias_set = add->alias_set || base->alias_set;
+
     new->add_forwarded_headers =
         (add->add_forwarded_headers_set == 0) ? base->add_forwarded_headers
         : add->add_forwarded_headers;
     new->add_forwarded_headers_set = add->add_forwarded_headers_set
         || base->add_forwarded_headers_set;
+
     new->forward_100_continue =
         (add->forward_100_continue_set == 0) ? base->forward_100_continue
                                              : add->forward_100_continue;
     new->forward_100_continue_set = add->forward_100_continue_set
                                     || base->forward_100_continue_set;
 
+    new->async_delay =
+        (add->async_delay_set == 0) ? base->async_delay
+                                    : add->async_delay;
+    new->async_delay_set = add->async_delay_set
+                           || base->async_delay_set;
+    new->async_idle_timeout =
+        (add->async_idle_timeout_set == 0) ? base->async_idle_timeout
+                                           : add->async_idle_timeout;
+    new->async_idle_timeout_set = add->async_idle_timeout_set
+                                  || base->async_idle_timeout_set;
+
     return new;
 }
 
@@ -2479,6 +2493,33 @@ static const char *
    return NULL;
 }
 
+static const char *
+    set_proxy_async_delay(cmd_parms *parms, void *dconf, const char *arg)
+{
+    proxy_dir_conf *conf = dconf;
+    if (strcmp(arg, "-1") == 0) {
+        conf->async_delay = -1;
+    }
+    else if (ap_timeout_parameter_parse(arg, &conf->async_delay, "s")
+                || conf->async_delay < 0) {
+        return "ProxyAsyncDelay has wrong format";
+    }
+    conf->async_delay_set = 1;
+    return NULL;
+}
+
+static const char *
+    set_proxy_async_idle(cmd_parms *parms, void *dconf, const char *arg)
+{
+    proxy_dir_conf *conf = dconf;
+    if (ap_timeout_parameter_parse(arg, &conf->async_idle_timeout, "s")
+            || conf->async_idle_timeout < 0) {
+        return "ProxyAsyncIdleTimeout has wrong format";
+    }
+    conf->async_idle_timeout_set = 1;
+    return NULL;
+}
+
 static const char *
     set_recv_buffer_size(cmd_parms *parms, void *dummy, const char *arg)
 {
@@ -3068,6 +3109,10 @@ static const command_rec proxy_cmds[] =
     AP_INIT_FLAG("Proxy100Continue", forward_100_continue, NULL, RSRC_CONF|ACCESS_CONF,
      "on if 100-Continue should be forwarded to the origin server, off if the "
      "proxy should handle it by itself"),
+    AP_INIT_TAKE1("ProxyAsyncDelay", set_proxy_async_delay, NULL, RSRC_CONF|ACCESS_CONF,
+     "Amount of time to poll before going asynchronous"),
+    AP_INIT_TAKE1("ProxyAsyncIdleTimeout", set_proxy_async_idle, NULL, RSRC_CONF|ACCESS_CONF,
+     "Timeout for asynchronous inactivity, ProxyTimeout by default"),
     {NULL}
 };
 
index b013fcf0897aca2b74f1f398580960b3c3d33062..2250febcb0ec01805d5d17294e8e2182ad920874 100644 (file)
@@ -247,6 +247,11 @@ typedef struct {
     unsigned int forward_100_continue_set:1;
 
     apr_array_header_t *error_override_codes;
+
+    apr_interval_time_t async_delay;
+    apr_interval_time_t async_idle_timeout;
+    unsigned int async_delay_set:1;
+    unsigned int async_idle_timeout_set:1;
 } proxy_dir_conf;
 
 /* if we interpolate env vars per-request, we'll need a per-request
index 01ba1ce5347e58f5ae1cab6c2cdd1560f0836b9e..e34ab31fca4a45f6a8cdd3678bc10777c49e951e 100644 (file)
 
 #include "mod_proxy.h"
 #include "ap_regex.h"
+#include "ap_mpm.h"
 
 module AP_MODULE_DECLARE_DATA proxy_http_module;
 
 static int (*ap_proxy_clear_connection_fn)(request_rec *r, apr_table_t *headers) =
         NULL;
 
-static apr_status_t ap_proxy_http_cleanup(const char *scheme,
-                                          request_rec *r,
-                                          proxy_conn_rec *backend);
-
 static apr_status_t ap_proxygetline(apr_bucket_brigade *bb, char *s, int n,
                                     request_rec *r, int flags, int *read);
 
+
 /*
  * Canonicalise http-like URLs.
  *  scheme is the scheme for the URL
@@ -219,6 +217,12 @@ static void add_cl(apr_pool_t *p,
 
 #define MAX_MEM_SPOOL 16384
 
+typedef enum {
+    PROXY_HTTP_REQ_HAVE_HEADER = 0,
+
+    PROXY_HTTP_TUNNELING
+} proxy_http_state;
+
 typedef enum {
     RB_INIT = 0,
     RB_STREAM_CL,
@@ -229,29 +233,129 @@ typedef enum {
 typedef struct {
     apr_pool_t *p;
     request_rec *r;
+    const char *proto;
     proxy_worker *worker;
+    proxy_dir_conf *dconf;
     proxy_server_conf *sconf;
-
     char server_portstr[32];
+
     proxy_conn_rec *backend;
     conn_rec *origin;
 
     apr_bucket_alloc_t *bucket_alloc;
     apr_bucket_brigade *header_brigade;
     apr_bucket_brigade *input_brigade;
+
     char *old_cl_val, *old_te_val;
     apr_off_t cl_val;
 
+    proxy_http_state state;
     rb_methods rb_method;
 
-    int force10;
     const char *upgrade;
-
-    int expecting_100;
-    unsigned int do_100_continue:1,
-                 prefetch_nonblocking:1;
+    proxy_tunnel_rec *tunnel;
+    apr_array_header_t *pfds;
+    apr_interval_time_t idle_timeout;
+
+    unsigned int can_go_async           :1,
+                 expecting_100          :1,
+                 do_100_continue        :1,
+                 prefetch_nonblocking   :1,
+                 force10                :1;
 } proxy_http_req_t;
 
+static void proxy_http_async_finish(proxy_http_req_t *req)
+{ 
+    conn_rec *c = req->r->connection;
+
+    ap_log_rerror(APLOG_MARK, APLOG_TRACE1, 0, req->r,
+                  "proxy %s: finish async", req->proto);
+
+    proxy_run_detach_backend(req->r, req->backend);
+    ap_proxy_release_connection(req->proto, req->backend, req->r->server);
+
+    ap_finalize_request_protocol(req->r);
+    ap_process_request_after_handler(req->r);
+    /* don't touch req or req->r from here */
+
+    c->cs->state = CONN_STATE_LINGER;
+    ap_mpm_resume_suspended(c);
+}
+
+/* If neither socket becomes readable in the specified timeout,
+ * this callback will kill the request.
+ * We do not have to worry about having a cancel and a IO both queued.
+ */
+static void proxy_http_async_cancel_cb(void *baton)
+{ 
+    proxy_http_req_t *req = (proxy_http_req_t *)baton;
+
+    ap_log_rerror(APLOG_MARK, APLOG_TRACE1, 0, req->r,
+                  "proxy %s: cancel async", req->proto);
+
+    req->r->connection->keepalive = AP_CONN_CLOSE;
+    req->backend->close = 1;
+    proxy_http_async_finish(req);
+}
+
+/* Invoked by the event loop when data is ready on either end. 
+ * We don't need the invoke_mtx, since we never put multiple callback events
+ * in the queue.
+ */
+static void proxy_http_async_cb(void *baton)
+{ 
+    proxy_http_req_t *req = (proxy_http_req_t *)baton;
+    int status;
+
+    if (req->pfds) {
+        apr_pool_clear(req->pfds->pool);
+    }
+
+    switch (req->state) {
+    case PROXY_HTTP_TUNNELING:
+        /* Pump both ends until they'd block and then start over again */
+        status = ap_proxy_tunnel_run(req->tunnel);
+        if (status == HTTP_GATEWAY_TIME_OUT) {
+            if (req->pfds) {
+                apr_pollfd_t *async_pfds = (void *)req->pfds->elts;
+                apr_pollfd_t *tunnel_pfds = (void *)req->tunnel->pfds->elts;
+                async_pfds[0].reqevents = tunnel_pfds[0].reqevents;
+                async_pfds[1].reqevents = tunnel_pfds[1].reqevents;
+            }
+            else {
+                req->pfds = apr_array_copy(req->p, req->tunnel->pfds);
+                apr_pool_create(&req->pfds->pool, req->p);
+            }
+            status = SUSPENDED;
+        }
+        break;
+
+    default:
+        ap_log_rerror(APLOG_MARK, APLOG_ERR, 0, req->r,
+                      "proxy %s: unexpected async state (%i)",
+                      req->proto, (int)req->state);
+        status = HTTP_INTERNAL_SERVER_ERROR;
+        break;
+    }
+
+    if (status == SUSPENDED) {
+        ap_log_rerror(APLOG_MARK, APLOG_TRACE1, 0, req->r,
+                      "proxy %s: suspended, going async",
+                      req->proto);
+
+        ap_mpm_register_poll_callback_timeout(req->pfds,
+                                              proxy_http_async_cb, 
+                                              proxy_http_async_cancel_cb, 
+                                              req, req->idle_timeout);
+    }
+    else if (status != OK) {
+        proxy_http_async_cancel_cb(req);
+    }
+    else {
+        proxy_http_async_finish(req);
+    }
+}
+
 /* Read what's in the client pipe. If nonblocking is set and read is EAGAIN,
  * pass a FLUSH bucket to the backend and read again in blocking mode.
  */
@@ -1200,13 +1304,11 @@ int ap_proxy_http_process_response(proxy_http_req_t *req)
     int i;
     const char *te = NULL;
     int original_status = r->status;
-    int proxy_status = OK;
     const char *original_status_line = r->status_line;
     const char *proxy_status_line = NULL;
     apr_interval_time_t old_timeout = 0;
-    proxy_dir_conf *dconf;
-
-    dconf = ap_get_module_config(r->per_dir_config, &proxy_module);
+    proxy_dir_conf *dconf = req->dconf;
+    int proxy_status = OK;
 
     bb = apr_brigade_create(p, c->bucket_alloc);
     pass_bb = apr_brigade_create(p, c->bucket_alloc);
@@ -1634,9 +1736,6 @@ int ap_proxy_http_process_response(proxy_http_req_t *req)
 
         if (proxy_status == HTTP_SWITCHING_PROTOCOLS) {
             apr_status_t rv;
-            proxy_tunnel_rec *tunnel;
-            apr_interval_time_t client_timeout = -1,
-                                backend_timeout = -1;
 
             /* If we didn't send the full body yet, do it now */
             if (do_100_continue) {
@@ -1650,41 +1749,35 @@ int ap_proxy_http_process_response(proxy_http_req_t *req)
             ap_log_rerror(APLOG_MARK, APLOG_DEBUG, 0, r, APLOGNO(10239)
                           "HTTP: tunneling protocol %s", upgrade);
 
-            rv = ap_proxy_tunnel_create(&tunnel, r, origin, "HTTP");
+            rv = ap_proxy_tunnel_create(&req->tunnel, r, origin, upgrade);
             if (rv != APR_SUCCESS) {
                 ap_log_rerror(APLOG_MARK, APLOG_ERR, rv, r, APLOGNO(10240)
                               "can't create tunnel for %s", upgrade);
                 return HTTP_INTERNAL_SERVER_ERROR;
             }
 
-            /* Set timeout to the lowest configured for client or backend */
-            apr_socket_timeout_get(backend->sock, &backend_timeout);
-            apr_socket_timeout_get(ap_get_conn_socket(c), &client_timeout);
-            if (backend_timeout >= 0 && backend_timeout < client_timeout) {
-                tunnel->timeout = backend_timeout;
-            }
-            else {
-                tunnel->timeout = client_timeout;
-            }
+            req->proto = upgrade;
 
-            /* Let proxy tunnel forward everything */
-            status = ap_proxy_tunnel_run(tunnel);
-            if (ap_is_HTTP_ERROR(status)) {
-                /* Tunnel always return HTTP_GATEWAY_TIME_OUT on timeout,
-                 * but we can differentiate between client and backend here.
-                 */
-                if (status == HTTP_GATEWAY_TIME_OUT
-                        && tunnel->timeout == client_timeout) {
-                    status = HTTP_REQUEST_TIME_OUT;
-                }
+            if (req->can_go_async) {
+                /* Let the MPM schedule the work when idle */
+                req->state = PROXY_HTTP_TUNNELING;
+                req->tunnel->timeout = dconf->async_delay;
+                proxy_http_async_cb(req);
+                return SUSPENDED;
             }
-            else {
+
+            /* Let proxy tunnel forward everything within this thread */
+            req->tunnel->timeout = req->idle_timeout;
+            status = ap_proxy_tunnel_run(req->tunnel);
+            if (!ap_is_HTTP_ERROR(status)) {
                 /* Update r->status for custom log */
                 status = HTTP_SWITCHING_PROTOCOLS;
             }
             r->status = status;
 
             /* We are done with both connections */
+            r->connection->keepalive = AP_CONN_CLOSE;
+            backend->close = 1;
             return DONE;
         }
 
@@ -2000,14 +2093,6 @@ int ap_proxy_http_process_response(proxy_http_req_t *req)
     return OK;
 }
 
-static
-apr_status_t ap_proxy_http_cleanup(const char *scheme, request_rec *r,
-                                   proxy_conn_rec *backend)
-{
-    ap_proxy_release_connection(scheme, backend, r->server);
-    return OK;
-}
-
 /*
  * This handles http:// URLs, and other URLs using a remote proxy over http
  * If proxyhost is NULL, then contact the server directly, otherwise
@@ -2029,6 +2114,7 @@ static int proxy_http_handler(request_rec *r, proxy_worker *worker,
     proxy_http_req_t *req = NULL;
     proxy_conn_rec *backend = NULL;
     apr_bucket_brigade *input_brigade = NULL;
+    int mpm_can_poll = 0;
     int is_ssl = 0;
     conn_rec *c = r->connection;
     proxy_dir_conf *dconf;
@@ -2080,20 +2166,26 @@ static int proxy_http_handler(request_rec *r, proxy_worker *worker,
                                               worker, r->server)) != OK) {
         return status;
     }
-
     backend->is_ssl = is_ssl;
 
+    dconf = ap_get_module_config(r->per_dir_config, &proxy_module);
+    ap_mpm_query(AP_MPMQ_CAN_POLL, &mpm_can_poll);
+
     req = apr_pcalloc(p, sizeof(*req));
     req->p = p;
     req->r = r;
     req->sconf = conf;
+    req->dconf = dconf;
     req->worker = worker;
     req->backend = backend;
+    req->proto = proxy_function;
     req->bucket_alloc = c->bucket_alloc;
+    req->can_go_async = (mpm_can_poll &&
+                         dconf->async_delay_set &&
+                         dconf->async_delay >= 0);
+    req->state = PROXY_HTTP_REQ_HAVE_HEADER;
     req->rb_method = RB_INIT;
 
-    dconf = ap_get_module_config(r->per_dir_config, &proxy_module);
-
     if (apr_table_get(r->subprocess_env, "force-proxy-request-1.0")) {
         req->force10 = 1;
     }
@@ -2105,6 +2197,22 @@ static int proxy_http_handler(request_rec *r, proxy_worker *worker,
         }
     }
 
+    if (req->can_go_async || req->upgrade) {
+        /* If ProxyAsyncIdleTimeout is not set, use backend timeout */
+        if (req->can_go_async && dconf->async_idle_timeout_set) {
+            req->idle_timeout = dconf->async_idle_timeout;
+        }
+        else if (worker->s->timeout_set) {
+            req->idle_timeout = worker->s->timeout;
+        }
+        else if (conf->timeout_set) {
+            req->idle_timeout = conf->timeout;
+        }
+        else {
+            req->idle_timeout = r->server->timeout;
+        }
+    }
+
     /* We possibly reuse input data prefetched in previous call(s), e.g. for a
      * balancer fallback scenario, and in this case the 100 continue settings
      * should be consistent between balancer members. If not, we need to ignore
@@ -2128,15 +2236,19 @@ static int proxy_http_handler(request_rec *r, proxy_worker *worker,
          * req->expecting_100 (i.e. cleared only if mod_proxy_http sent the
          * "100 Continue" according to its policy).
          */
-        req->do_100_continue = req->prefetch_nonblocking = 1;
-        req->expecting_100 = r->expecting_100;
+        req->do_100_continue = 1;
+        req->expecting_100 = (r->expecting_100 != 0);
         r->expecting_100 = 0;
     }
+
     /* Should we block while prefetching the body or try nonblocking and flush
      * data to the backend ASAP?
      */
-    else if (input_brigade || apr_table_get(r->subprocess_env,
-                                            "proxy-prefetch-nonblocking")) {
+    if (input_brigade
+             || req->can_go_async
+             || req->do_100_continue
+             || apr_table_get(r->subprocess_env,
+                              "proxy-prefetch-nonblocking")) {
         req->prefetch_nonblocking = 1;
     }
 
@@ -2255,6 +2367,9 @@ static int proxy_http_handler(request_rec *r, proxy_worker *worker,
 
         /* Step Five: Receive the Response... Fall thru to cleanup */
         status = ap_proxy_http_process_response(req);
+        if (status == SUSPENDED) {
+            return SUSPENDED;
+        }
         if (req->backend) {
             proxy_run_detach_backend(r, req->backend);
         }
@@ -2267,7 +2382,8 @@ cleanup:
     if (req->backend) {
         if (status != OK)
             req->backend->close = 1;
-        ap_proxy_http_cleanup(proxy_function, r, req->backend);
+        ap_proxy_release_connection(proxy_function, req->backend,
+                                    r->server);
     }
     if (req->expecting_100) {
         /* Restore r->expecting_100 if we didn't touch it */