When using SUSPENDED connections to serve HTTP request the connection is never closed. For example, as discussed at http://thread.gmane.org/gmane.comp.apache.user/104996, this test case https://gist.github.com/ArtemGr/9870554 will leak connections and will fail an "ab" test. Please introduce an API allowing one to return a SUSPENDED connection back to the mpm_event queue, so that the mpm_event will properly flush and close the connection (or even reuse it with KeepAlive). A rough example of such an API was hacked together in the following patch: https://gist.github.com/ArtemGr/9887564 (for httpd 2.4.6).
FYI: In httpd 2.4.9 my hack no longer works for some reason.
Created attachment 31519 [details] Patch for httpd-2.4.9 implementing `ap_resume_suspended`. A (hopefully improved) version of the patch.
Created attachment 31520 [details] Test case working with httpd-2.4.9. Nevermind my previous comment about patch no longer working with httpd-2.4.9. It works (if ap_mpm_register_timed_callback isn't used). Attached is the test case. Without the patch it hangs on "ab" and with the patch it works NP.
Comment on attachment 31520 [details] Test case working with httpd-2.4.9. >// apxs -i -a -c mod_sustest.c >// SetHandler sustest >// ab -n 9999 -c 9 http://localhost/sustest > >#include <httpd.h> >#include <http_protocol.h> >#include <http_config.h> >#include <http_request.h> >#include <http_log.h> >#include <http_connection.h> >#include <ap_mpm.h> > >#include <unistd.h> > >module sustest_module; >APLOG_USE_MODULE (sustest); > >static void render (request_rec* r) { > r->content_type = "text/html; charset=utf-8"; > if (!r->header_only) { > ap_rputs ("<!doctype html><html lang=\"ru\"><head><meta charset=\"UTF-8\"><title>test</title></head><body>\n", r); > if (r->filename) ap_rprintf (r, "FS path: %s<br/>\n", r->filename); > ap_rputs ("v. 02<br/>", r); > ap_rputs ("</body></html>\n", r); > } >} > >static void setupSuspended (request_rec* r) { >} > >void ap_resume_suspended (conn_rec* c); > >static void finishSuspended (request_rec* r) { > conn_rec* c = r->connection; > apr_thread_mutex_lock (r->invoke_mtx); > > if (r->connection->aborted) { > ap_die (HTTP_INTERNAL_SERVER_ERROR, r); > } else if (r->status == HTTP_OK) { > apr_bucket_brigade* bb = apr_brigade_create (r->pool, r->connection->bucket_alloc); > //APR_BRIGADE_INSERT_TAIL (bb, apr_bucket_flush_create (r->connection->bucket_alloc)); > APR_BRIGADE_INSERT_TAIL (bb, apr_bucket_eos_create (r->connection->bucket_alloc)); > ap_pass_brigade (r->output_filters, bb); > apr_brigade_destroy (bb); > ap_finalize_request_protocol (r); > } > > apr_thread_mutex_unlock (r->invoke_mtx); > ap_process_request_after_handler (r); > ap_log_rerror (APLOG_MARK, LOG_INFO, r->status, r, "Calling resume_suspended..."); > ap_resume_suspended (c); >} > >static void callback (void* vr) { > request_rec* r = (request_rec*) vr; > ap_log_rerror (APLOG_MARK, LOG_INFO, r->status, r, "Callback."); > > if (!r->connection->aborted) render (r); > > finishSuspended (r); >} > >static void* thread_func (void* vr) { > usleep (100); > callback (vr); > return NULL; >} > >static int sustest_handler (request_rec* r) { > if (!r->handler || strcmp (r->handler, "sustest")) return DECLINED; > if (ap_meets_conditions (r) != OK) {ap_log_rerror (APLOG_MARK, APLOG_NOTICE, 0, r, "!ap_meets_conditions"); return DECLINED;} > > ap_log_rerror (APLOG_MARK, LOG_INFO, r->status, r, "sustest_handler"); > //render (r); return OK; > setupSuspended (r); > > pthread_t thread; > if (pthread_create (&thread, NULL, thread_func, r) != 0) return DECLINED; > pthread_detach (thread); > //ap_mpm_register_timed_callback (apr_time_from_msec (90), callback, r); > return SUSPENDED; >} > >static void sustest_hooks (apr_pool_t *pool) { > ap_hook_handler (sustest_handler, NULL, NULL, APR_HOOK_FIRST); >} > >module sustest_module = { > STANDARD20_MODULE_STUFF, > NULL, > NULL, > NULL, > NULL, > NULL, > sustest_hooks >};
Created attachment 31521 [details] Test case working with httpd-2.4.9. (Keep-alive works fine with the patch).
Created attachment 31548 [details] Add support for connection suspension in Event MPM Cleaned up the code; I changed the globally exposed function to a hook, modeled after the timed_callback which is very similar. The code within the hook is yours, though. In doing so, I had to introduce a new ap_mpm_query() query, to make it worker/prefork-safe. Any code calling the new function should check if the MPM supports suspension first.
Created attachment 31549 [details] Testcase for new patch Forgot to mention that the new patch is against trunk, not against 2.4.9. Here's the testcase, changed a little for the new patch.
Created attachment 31551 [details] Add support for resuming a suspended connection in Event MPM Fixed the error messages in the Edward Lu's patch to match the new function name.
Edward, thanks for looking at this!
http://svn.apache.org/r1593860