pollout: handle request for pollout during pollout service
authorAndy Green <andy@warmcat.com>
Wed, 5 Apr 2017 00:30:55 +0000 (08:30 +0800)
committerAndy Green <andy@warmcat.com>
Wed, 5 Apr 2017 00:30:55 +0000 (08:30 +0800)
lib/pollfd.c
lib/private-libwebsockets.h
lib/service.c

index 56cca1d..841019f 100644 (file)
@@ -33,6 +33,26 @@ _lws_change_pollfd(struct lws *wsi, int _and, int _or, struct lws_pollargs *pa)
        if (!wsi || wsi->position_in_fds_table < 0)
                return 0;
 
+       if (wsi->handling_pollout && !_and && _or == LWS_POLLOUT) {
+               /*
+                * Happening alongside service thread handling POLLOUT.
+                * The danger is when he is finished, he will disable POLLOUT,
+                * countermanding what we changed here.
+                *
+                * Instead of changing the fds, inform the service thread
+                * what happened, and ask it to leave POLLOUT active on exit
+                */
+               wsi->leave_pollout_active = 1;
+               /*
+                * by definition service thread is not in poll wait, so no need
+                * to cancel service
+                */
+
+               lwsl_debug("%s: using leave_pollout_active\n", __func__);
+
+               return 0;
+       }
+
        context = wsi->context;
        pt = &context->pt[(int)wsi->tsi];
        assert(wsi->position_in_fds_table >= 0 &&
index 64cfcf6..abe4671 100644 (file)
@@ -1502,7 +1502,7 @@ struct lws {
        unsigned int sending_chunked:1;
        unsigned int already_did_cce:1;
        unsigned int told_user_closed:1;
-       unsigned int :1;
+
 #if defined(LWS_WITH_ESP8266)
        unsigned int pending_send_completion:3;
        unsigned int close_is_pending_send_completion:1;
@@ -1532,6 +1532,10 @@ struct lws {
        unsigned int redirect_to_https:1;
 #endif
 
+       /* volatile to make sure code is aware other thread can change */
+       volatile unsigned int handling_pollout:1;
+       volatile unsigned int leave_pollout_active:1;
+
 #ifndef LWS_NO_CLIENT
        unsigned short c_port;
 #endif
index 6a42dd2..159dd8c 100644 (file)
@@ -64,6 +64,14 @@ lws_handle_POLLOUT_event(struct lws *wsi, struct lws_pollfd *pollfd)
 
        //lwsl_err("%s: %p\n", __func__, wsi);
 
+       wsi->leave_pollout_active = 0;
+       wsi->handling_pollout = 1;
+       /*
+        * if another thread wants POLLOUT on us, from here on while
+        * handling_pollout is set, he will only set leave_pollout_active.
+        * If we are going to disable POLLOUT, we will check that first.
+        */
+
        /*
         * user callback is lowest priority to get these notifications
         * actually, since other pending things cannot be disordered
@@ -77,13 +85,13 @@ lws_handle_POLLOUT_event(struct lws *wsi, struct lws_pollfd *pollfd)
                if (lws_issue_raw(wsi, wsi->trunc_alloc + wsi->trunc_offset,
                                  wsi->trunc_len) < 0) {
                        lwsl_info("%s signalling to close\n", __func__);
-                       return -1;
+                       goto bail_die;
                }
                /* leave POLLOUT active either way */
-               return 0;
+               goto bail_ok;
        } else
                if (wsi->state == LWSS_FLUSHING_STORED_SEND_BEFORE_CLOSE)
-                       return -1; /* retry closing now */
+                       goto bail_die; /* retry closing now */
 
        if (wsi->mode == LWSCM_WSCL_ISSUE_HTTP_BODY)
                goto user_service;
@@ -105,7 +113,7 @@ lws_handle_POLLOUT_event(struct lws *wsi, struct lws_pollfd *pollfd)
                wsi->pps = LWS_PPS_NONE;
                lws_rx_flow_control(wsi, 1);
 
-               return 0; /* leave POLLOUT active */
+               goto bail_ok; /* leave POLLOUT active */
        }
 #endif
 
@@ -134,16 +142,16 @@ lws_handle_POLLOUT_event(struct lws *wsi, struct lws_pollfd *pollfd)
                n = lws_write(wsi, &wsi->u.ws.ping_payload_buf[LWS_PRE],
                              wsi->u.ws.ping_payload_len, write_type);
                if (n < 0)
-                       return -1;
+                       goto bail_die;
 
                /* well he is sent, mark him done */
                wsi->u.ws.ping_pending_flag = 0;
                if (wsi->u.ws.payload_is_close)
                        /* oh... a close frame was it... then we are done */
-                       return -1;
+                       goto bail_die;
 
                /* otherwise for PING, leave POLLOUT active either way */
-               return 0;
+               goto bail_ok;
        }
 
        if (wsi->state == LWSS_ESTABLISHED &&
@@ -155,7 +163,7 @@ lws_handle_POLLOUT_event(struct lws *wsi, struct lws_pollfd *pollfd)
                n = lws_write(wsi, &wsi->u.ws.ping_payload_buf[LWS_PRE],
                              0, LWS_WRITE_PING);
                if (n < 0)
-                       return -1;
+                       goto bail_die;
 
                /*
                 * we apparently were able to send the PING in a reasonable time
@@ -166,7 +174,7 @@ lws_handle_POLLOUT_event(struct lws *wsi, struct lws_pollfd *pollfd)
                lws_set_timeout(wsi, PENDING_TIMEOUT_WS_PONG_CHECK_GET_PONG,
                                wsi->context->timeout_secs);
 
-               return 0;
+               goto bail_ok;
        }
 
        /* Priority 4: if we are closing, not allowed to send more data frags
@@ -185,16 +193,16 @@ lws_handle_POLLOUT_event(struct lws *wsi, struct lws_pollfd *pollfd)
        if (wsi->state == LWSS_ESTABLISHED && wsi->u.ws.tx_draining_ext) {
                lwsl_ext("SERVICING TX EXT DRAINING\n");
                if (lws_write(wsi, NULL, 0, LWS_WRITE_CONTINUATION) < 0)
-                       return -1;
+                       goto bail_die;
                /* leave POLLOUT active */
-               return 0;
+               goto bail_ok;
        }
 
        /* Priority 6: user can get the callback
         */
        m = lws_ext_cb_active(wsi, LWS_EXT_CB_IS_WRITEABLE, NULL, 0);
        if (m)
-               return -1;
+               goto bail_die;
 #ifndef LWS_NO_EXTENSIONS
        if (!wsi->extension_data_pending)
                goto user_service;
@@ -225,7 +233,7 @@ lws_handle_POLLOUT_event(struct lws *wsi, struct lws_pollfd *pollfd)
                                               &eff_buf, 0);
                if (m < 0) {
                        lwsl_err("ext reports fatal error\n");
-                       return -1;
+                       goto bail_die;
                }
                if (m)
                        /*
@@ -241,7 +249,7 @@ lws_handle_POLLOUT_event(struct lws *wsi, struct lws_pollfd *pollfd)
                                          eff_buf.token_len);
                        if (n < 0) {
                                lwsl_info("closing from POLLOUT spill\n");
-                               return -1;
+                               goto bail_die;
                        }
                        /*
                         * Keep amount spilled small to minimize chance of this
@@ -249,7 +257,7 @@ lws_handle_POLLOUT_event(struct lws *wsi, struct lws_pollfd *pollfd)
                        if (n != eff_buf.token_len) {
                                lwsl_err("Unable to spill ext %d vs %d\n",
                                                          eff_buf.token_len, n);
-                               return -1;
+                               goto bail_die;
                        }
                } else
                        continue;
@@ -277,7 +285,7 @@ lws_handle_POLLOUT_event(struct lws *wsi, struct lws_pollfd *pollfd)
                 * when we come back here and there's nothing more to spill.
                 */
 
-               return 0;
+               goto bail_ok;
        }
 #ifndef LWS_NO_EXTENSIONS
        wsi->extension_data_pending = 0;
@@ -285,15 +293,31 @@ lws_handle_POLLOUT_event(struct lws *wsi, struct lws_pollfd *pollfd)
 user_service:
        /* one shot */
 
-       if (pollfd)
-               if (lws_change_pollfd(wsi, LWS_POLLOUT, 0)) {
-                       lwsl_info("failed at set pollfd\n");
-                       return 1;
-               }
+       if (pollfd) {
+               int eff = wsi->leave_pollout_active;
+
+               if (!eff)
+                       if (lws_change_pollfd(wsi, LWS_POLLOUT, 0)) {
+                               lwsl_info("failed at set pollfd\n");
+                               goto bail_die;
+                       }
+
+               wsi->handling_pollout = 0;
+
+               /* cannot get leave_pollout_active set after the above */
+
+               if (!eff && wsi->leave_pollout_active)
+                       /* got set inbetween sampling eff and clearing
+                        * handling_pollout, force POLLOUT on */
+                       lws_calllback_as_writeable(wsi);
+
+               wsi->leave_pollout_active = 0;
+       }
 
        if (wsi->mode != LWSCM_WSCL_ISSUE_HTTP_BODY &&
            !wsi->hdr_parsing_completed)
-               return 0;
+               goto bail_ok;
+
 
 #ifdef LWS_WITH_CGI
 user_service_go_again:
@@ -321,7 +345,7 @@ user_service_go_again:
        wsi->u.http2.requested_POLLOUT = 0;
        if (!wsi->u.http2.initialized) {
                lwsl_info("pollout on uninitialized http2 conn\n");
-               return 0;
+               goto bail_ok;
        }
 
        lwsl_info("%s: doing children\n", __func__);
@@ -344,10 +368,30 @@ user_service_go_again:
 
        lwsl_info("%s: completed\n", __func__);
 
-       return 0;
+       goto bail_ok;
 notify:
 #endif
+       wsi->handling_pollout = 0;
+       wsi->leave_pollout_active = 0;
+
        return lws_calllback_as_writeable(wsi);
+
+       /*
+        * since these don't disable the POLLOUT, they are always doing the
+        * right thing for leave_pollout_active whether it was set or not.
+        */
+
+bail_ok:
+       wsi->handling_pollout = 0;
+       wsi->leave_pollout_active = 0;
+
+       return 0;
+
+bail_die:
+       wsi->handling_pollout = 0;
+       wsi->leave_pollout_active = 0;
+
+       return -1;
 }
 
 int