Mercurial > hg > ngx_http_upstream_keepalive

diff ngx_http_upstream_keepalive_module.c @ 11:15530a464dba

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Keepalive: don't cache invalid connections. 1. Remember failed status, since peer.free() may be called more than once. It's called twice if peer fails - once from ngx_http_upstream_next() with NGX_PEER_FAILED set, and then from ngx_http_upstream_finalize_request() without. 2. We shouldn't cache connection unless we aren't expecting anything from upstream. For memcached this means either 404 response or 200 response with all body read (body may not be read e.g. when serving HEAD request).
author Maxim Dounin <mdounin@mdounin.ru>
date Thu, 13 Nov 2008 19:36:23 +0300
parents 06bd0e50e696
children 28af4b0b32c1
line wrap: on
line diff
--- a/ngx_http_upstream_keepalive_module.c
+++ b/ngx_http_upstream_keepalive_module.c
@@ -25,11 +25,15 @@ typedef struct {
 typedef struct {
     ngx_http_upstream_keepalive_srv_conf_t  *conf;
 
+    ngx_http_upstream_t               *upstream;
+
     void                              *data;
 
     ngx_event_get_peer_pt              original_get_peer;
     ngx_event_free_peer_pt             original_free_peer;
 
+    ngx_uint_t                         failed;       /* unsigned:1 */
+
 } ngx_http_upstream_keepalive_peer_data_t;
 
 
@@ -169,6 +173,7 @@ ngx_http_upstream_init_keepalive_peer(ng
     }
 
     kp->conf = kcf;
+    kp->upstream = r->upstream;
     kp->data = r->upstream->peer.data;
     kp->original_get_peer = r->upstream->peer.get;
     kp->original_free_peer = r->upstream->peer.free;
@@ -194,6 +199,8 @@ ngx_http_upstream_get_keepalive_peer(ngx
     ngx_log_debug0(NGX_LOG_DEBUG_HTTP, pc->log, 0,
                    "get keepalive peer");
 
+    kp->failed = 0;
+
     /* single pool of cached connections */
 
     if (kp->conf->single && !ngx_queue_empty(&kp->conf->cache)) {
@@ -264,14 +271,50 @@ ngx_http_upstream_free_keepalive_peer(ng
     ngx_http_upstream_keepalive_peer_data_t  *kp = data;
     ngx_http_upstream_keepalive_cache_t      *item;
 
-    ngx_queue_t       *q;
-    ngx_connection_t  *c;
+    ngx_uint_t            status;
+    ngx_queue_t          *q;
+    ngx_connection_t     *c;
+    ngx_http_upstream_t  *u;
 
     ngx_log_debug0(NGX_LOG_DEBUG_HTTP, pc->log, 0,
                    "free keepalive peer");
 
-    if (!(state & NGX_PEER_FAILED)
-        && pc->connection != NULL)
+    /* remember failed state - peer.free() may be called more than once */
+
+    if (state & NGX_PEER_FAILED) {
+        kp->failed = 1;
+    }
+
+    /*
+     * cache valid connections
+     *
+     * For memcached this means status either 404 or 200.  For status 200 we
+     * should also check if all response body was read (u->length == 0) and
+     * make sure that u->length is valid (we use u->header_sent flag to test
+     * this).  Memcached is the only supported protocol for now.
+     *
+     * Some notes on other possibilities (incomplete):
+     *
+     * fastcgi: u->pipe->upstream_done should be sufficient
+     *
+     * proxy buffered: u->pipe->upstream_done, 304 replies, replies to head
+     * requests (see RFC 2616, 4.4 Message Length)
+     *
+     * proxy unbuffered: 200 as for memcached (with u->length == 0 and
+     * header_sent), 304, replies to head requests
+     *
+     * subrequest_in_memory: won't work as of now
+     *
+     * TODO: move this logic to protocol modules (NGX_PEER_KEEPALIVE?)
+     */
+
+    u = kp->upstream;
+    status = u->headers_in.status_n;
+
+    if (!kp->failed
+        && pc->connection != NULL
+        && (status == NGX_HTTP_NOT_FOUND
+            || (status == NGX_HTTP_OK && u->header_sent && u->length == 0)))
     {
         c = pc->connection;