diff options
Diffstat (limited to 'src/vnet/session/application_local.c')
| -rw-r--r-- | src/vnet/session/application_local.c | 578 |
1 files changed, 449 insertions, 129 deletions
diff --git a/src/vnet/session/application_local.c b/src/vnet/session/application_local.c index 3c62dade0f5..3cb743d10e0 100644 --- a/src/vnet/session/application_local.c +++ b/src/vnet/session/application_local.c @@ -41,9 +41,25 @@ typedef struct ct_segments_ ct_segment_t *segments; } ct_segments_ctx_t; +typedef struct ct_cleanup_req_ +{ + u32 ct_index; +} ct_cleanup_req_t; + +typedef struct ct_worker_ +{ + ct_connection_t *connections; /**< Per-worker connection pools */ + u32 *pending_connects; /**< Fifo of pending ho indices */ + ct_cleanup_req_t *pending_cleanups; /**< Fifo of pending indices */ + u8 have_connects; /**< Set if connect rpc pending */ + u8 have_cleanups; /**< Set if cleanup rpc pending */ + clib_spinlock_t pending_connects_lock; /**< Lock for pending connects */ + u32 *new_connects; /**< Burst of connects to be done */ +} ct_worker_t; + typedef struct ct_main_ { - ct_connection_t **connections; /**< Per-worker connection pools */ + ct_worker_t *wrk; /**< Per-worker state */ u32 n_workers; /**< Number of vpp workers */ u32 n_sessions; /**< Cumulative sessions counter */ u32 *ho_reusable; /**< Vector of reusable ho indices */ @@ -51,17 +67,28 @@ typedef struct ct_main_ clib_rwlock_t app_segs_lock; /**< RW lock for seg contexts */ uword *app_segs_ctxs_table; /**< App handle to segment pool map */ ct_segments_ctx_t *app_seg_ctxs; /**< Pool of ct segment contexts */ + u32 **fwrk_pending_connects; /**< First wrk pending half-opens */ + u32 fwrk_thread; /**< First worker thread */ + u8 fwrk_have_flush; /**< Flag for connect flush rpc */ } ct_main_t; static ct_main_t ct_main; +static inline ct_worker_t * +ct_worker_get (u32 thread_index) +{ + return &ct_main.wrk[thread_index]; +} + static ct_connection_t * ct_connection_alloc (u32 thread_index) { + ct_worker_t *wrk = ct_worker_get (thread_index); ct_connection_t *ct; - pool_get_zero (ct_main.connections[thread_index], ct); - ct->c_c_index = ct - ct_main.connections[thread_index]; + pool_get_aligned_safe (wrk->connections, ct, CLIB_CACHE_LINE_BYTES); + clib_memset (ct, 0, sizeof (*ct)); + ct->c_c_index = ct - wrk->connections; ct->c_thread_index = thread_index; ct->client_wrk = ~0; ct->server_wrk = ~0; @@ -73,22 +100,25 @@ ct_connection_alloc (u32 thread_index) static ct_connection_t * ct_connection_get (u32 ct_index, u32 thread_index) { - if (pool_is_free_index (ct_main.connections[thread_index], ct_index)) + ct_worker_t *wrk = ct_worker_get (thread_index); + + if (pool_is_free_index (wrk->connections, ct_index)) return 0; - return pool_elt_at_index (ct_main.connections[thread_index], ct_index); + return pool_elt_at_index (wrk->connections, ct_index); } static void ct_connection_free (ct_connection_t * ct) { + ct_worker_t *wrk = ct_worker_get (ct->c_thread_index); + if (CLIB_DEBUG) { - u32 thread_index = ct->c_thread_index; - memset (ct, 0xfc, sizeof (*ct)); - pool_put (ct_main.connections[thread_index], ct); + clib_memset (ct, 0xfc, sizeof (*ct)); + pool_put (wrk->connections, ct); return; } - pool_put (ct_main.connections[ct->c_thread_index], ct); + pool_put (wrk->connections, ct); } static ct_connection_t * @@ -99,11 +129,18 @@ ct_half_open_alloc (void) clib_spinlock_lock (&cm->ho_reuseable_lock); vec_foreach (hip, cm->ho_reusable) - pool_put_index (cm->connections[0], *hip); + pool_put_index (cm->wrk[cm->fwrk_thread].connections, *hip); vec_reset_length (cm->ho_reusable); clib_spinlock_unlock (&cm->ho_reuseable_lock); - return ct_connection_alloc (0); + return ct_connection_alloc (cm->fwrk_thread); +} + +static ct_connection_t * +ct_half_open_get (u32 ho_index) +{ + ct_main_t *cm = &ct_main; + return ct_connection_get (ho_index, cm->fwrk_thread); } void @@ -137,6 +174,33 @@ ct_session_endpoint (session_t * ll, session_endpoint_t * sep) } static void +ct_set_invalid_app_wrk (ct_connection_t *ct, u8 is_client) +{ + ct_connection_t *peer_ct; + + peer_ct = ct_connection_get (ct->peer_index, ct->c_thread_index); + + if (is_client) + { + ct->client_wrk = APP_INVALID_INDEX; + if (peer_ct) + ct->client_wrk = APP_INVALID_INDEX; + } + else + { + ct->server_wrk = APP_INVALID_INDEX; + if (peer_ct) + ct->server_wrk = APP_INVALID_INDEX; + } +} + +static inline u64 +ct_client_seg_handle (u64 server_sh, u32 client_wrk_index) +{ + return (((u64) client_wrk_index << 56) | server_sh); +} + +static void ct_session_dealloc_fifos (ct_connection_t *ct, svm_fifo_t *rx_fifo, svm_fifo_t *tx_fifo) { @@ -146,8 +210,8 @@ ct_session_dealloc_fifos (ct_connection_t *ct, svm_fifo_t *rx_fifo, app_worker_t *app_wrk; ct_segment_t *ct_seg; fifo_segment_t *fs; - u8 del_segment = 0; u32 seg_index; + session_t *s; int cnt; /* @@ -202,77 +266,82 @@ ct_session_dealloc_fifos (ct_connection_t *ct, svm_fifo_t *rx_fifo, if (ct->flags & CT_CONN_F_CLIENT) { cnt = ct_seg->client_n_sessions; - if (!cnt) - ct_seg->flags |= CT_SEGMENT_F_CLIENT_DETACHED; + if (cnt) + goto done; + ct_seg->flags |= CT_SEGMENT_F_CLIENT_DETACHED; + s = session_get (ct->c_s_index, ct->c_thread_index); + if (s->app_wrk_index == APP_INVALID_INDEX) + ct_set_invalid_app_wrk (ct, 1 /* is_client */); } else { cnt = ct_seg->server_n_sessions; - if (!cnt) - ct_seg->flags |= CT_SEGMENT_F_SERVER_DETACHED; + if (cnt) + goto done; + ct_seg->flags |= CT_SEGMENT_F_SERVER_DETACHED; + s = session_get (ct->c_s_index, ct->c_thread_index); + if (s->app_wrk_index == APP_INVALID_INDEX) + ct_set_invalid_app_wrk (ct, 0 /* is_client */); } + if (!(ct_seg->flags & CT_SEGMENT_F_CLIENT_DETACHED) || + !(ct_seg->flags & CT_SEGMENT_F_SERVER_DETACHED)) + goto done; + /* * Remove segment context because both client and server detached */ - if (!cnt && (ct_seg->flags & CT_SEGMENT_F_CLIENT_DETACHED) && - (ct_seg->flags & CT_SEGMENT_F_SERVER_DETACHED)) - { - pool_put_index (seg_ctx->segments, ct->ct_seg_index); + pool_put_index (seg_ctx->segments, ct->ct_seg_index); - /* - * No more segment indices left, remove the segments context - */ - if (!pool_elts (seg_ctx->segments)) - { - u64 table_handle = seg_ctx->client_wrk << 16 | seg_ctx->server_wrk; - table_handle = (u64) seg_ctx->sm_index << 32 | table_handle; - hash_unset (cm->app_segs_ctxs_table, table_handle); - pool_free (seg_ctx->segments); - pool_put_index (cm->app_seg_ctxs, ct->seg_ctx_index); - } - del_segment = 1; + /* + * No more segment indices left, remove the segments context + */ + if (!pool_elts (seg_ctx->segments)) + { + u64 table_handle = seg_ctx->client_wrk << 16 | seg_ctx->server_wrk; + table_handle = (u64) seg_ctx->sm_index << 32 | table_handle; + hash_unset (cm->app_segs_ctxs_table, table_handle); + pool_free (seg_ctx->segments); + pool_put_index (cm->app_seg_ctxs, ct->seg_ctx_index); } - clib_rwlock_writer_unlock (&cm->app_segs_lock); - /* - * Session counter went to zero, notify the app that detached + * Segment to be removed so notify both apps */ - if (cnt) - return; - if (ct->flags & CT_CONN_F_CLIENT) - { - app_wrk = app_worker_get_if_valid (ct->client_wrk); - /* Determine if client app still needs notification, i.e., if it is - * still attached. If client detached and this is the last ct session - * on this segment, then its connects segment manager should also be - * detached, so do not send notification */ - if (app_wrk) - { - segment_manager_t *csm; - csm = app_worker_get_connect_segment_manager (app_wrk); - if (!segment_manager_app_detached (csm)) - app_worker_del_segment_notify (app_wrk, ct->segment_handle); - } - } - else if (!segment_manager_app_detached (sm)) + app_wrk = app_worker_get_if_valid (ct->client_wrk); + /* Determine if client app still needs notification, i.e., if it is + * still attached. If client detached and this is the last ct session + * on this segment, then its connects segment manager should also be + * detached, so do not send notification */ + if (app_wrk) { - app_wrk = app_worker_get (ct->server_wrk); - app_worker_del_segment_notify (app_wrk, ct->segment_handle); + segment_manager_t *csm; + csm = app_worker_get_connect_segment_manager (app_wrk); + if (!segment_manager_app_detached (csm)) + app_worker_del_segment_notify ( + app_wrk, ct_client_seg_handle (ct->segment_handle, ct->client_wrk)); } - if (!del_segment) - return; - + /* Notify server app and free segment */ segment_manager_lock_and_del_segment (sm, seg_index); /* Cleanup segment manager if needed. If server detaches there's a chance * the client's sessions will hold up segment removal */ if (segment_manager_app_detached (sm) && !segment_manager_has_fifos (sm)) segment_manager_free_safe (sm); + +done: + + clib_rwlock_writer_unlock (&cm->app_segs_lock); +} + +static void +ct_session_force_disconnect_server (ct_connection_t *sct) +{ + sct->peer_index = ~0; + session_transport_closing_notify (&sct->connection); } int @@ -294,9 +363,7 @@ ct_session_connect_notify (session_t *ss, session_error_t err) /* Client closed while waiting for reply from server */ if (PREDICT_FALSE (!cct)) { - session_transport_closing_notify (&sct->connection); - session_transport_delete_notify (&sct->connection); - ct_connection_free (sct); + ct_session_force_disconnect_server (sct); return 0; } @@ -307,16 +374,19 @@ ct_session_connect_notify (session_t *ss, session_error_t err) goto connect_error; /* - * Alloc client session + * Alloc client session, server session assumed to be established */ + ASSERT (ss->session_state >= SESSION_STATE_READY); + cs = session_alloc (thread_index); ss = session_get (ss_index, thread_index); cs->session_type = ss->session_type; cs->listener_handle = SESSION_INVALID_HANDLE; - cs->session_state = SESSION_STATE_CONNECTING; + session_set_state (cs, SESSION_STATE_CONNECTING); cs->app_wrk_index = client_wrk->wrk_index; cs->connection_index = cct->c_c_index; + cs->opaque = opaque; cct->c_s_index = cs->session_index; /* This will allocate fifos for the session. They won't be used for @@ -325,23 +395,23 @@ ct_session_connect_notify (session_t *ss, session_error_t err) if ((err = app_worker_init_connected (client_wrk, cs))) { session_free (cs); - session_close (ss); + ct_session_force_disconnect_server (sct); err = SESSION_E_ALLOC; goto connect_error; } - cs->session_state = SESSION_STATE_CONNECTING; + session_set_state (cs, SESSION_STATE_CONNECTING); if (app_worker_connect_notify (client_wrk, cs, 0, opaque)) { segment_manager_dealloc_fifos (cs->rx_fifo, cs->tx_fifo); session_free (cs); - session_close (ss); + ct_session_force_disconnect_server (sct); goto cleanup_client; } cs = session_get (cct->c_s_index, cct->c_thread_index); - cs->session_state = SESSION_STATE_READY; + session_set_state (cs, SESSION_STATE_READY); return 0; @@ -373,9 +443,6 @@ ct_lookup_free_segment (ct_main_t *cm, segment_manager_t *sm, pool_foreach (ct_seg, seg_ctx->segments) { /* Client or server has detached so segment cannot be used */ - if ((ct_seg->flags & CT_SEGMENT_F_SERVER_DETACHED) || - (ct_seg->flags & CT_SEGMENT_F_CLIENT_DETACHED)) - continue; fs = segment_manager_get_segment (sm, ct_seg->segment_index); free_bytes = fifo_segment_available_bytes (fs); max_fifos = fifo_segment_size (fs) / seg_ctx->fifo_pair_bytes; @@ -395,11 +462,11 @@ ct_alloc_segment (ct_main_t *cm, app_worker_t *server_wrk, u64 table_handle, segment_manager_t *sm, u32 client_wrk_index) { u32 seg_ctx_index = ~0, sm_index, pair_bytes; + u64 seg_size, seg_handle, client_seg_handle; segment_manager_props_t *props; const u32 margin = 16 << 10; ct_segments_ctx_t *seg_ctx; app_worker_t *client_wrk; - u64 seg_size, seg_handle; application_t *server; ct_segment_t *ct_seg; uword *spp; @@ -461,7 +528,11 @@ ct_alloc_segment (ct_main_t *cm, app_worker_t *server_wrk, u64 table_handle, goto error; client_wrk = app_worker_get (client_wrk_index); - if (app_worker_add_segment_notify (client_wrk, seg_handle)) + /* Make sure client workers do not have overlapping segment handles. + * Ideally, we should attach fs to client worker segment manager and + * create a new handle but that's not currently possible. */ + client_seg_handle = ct_client_seg_handle (seg_handle, client_wrk_index); + if (app_worker_add_segment_notify (client_wrk, client_seg_handle)) { app_worker_del_segment_notify (server_wrk, seg_handle); goto error; @@ -515,6 +586,8 @@ ct_init_accepted_session (app_worker_t *server_wrk, ct_connection_t *ct, ct->seg_ctx_index = ct_seg->seg_ctx_index; ct->ct_seg_index = ct_seg->ct_seg_index; fs_index = ct_seg->segment_index; + ct_seg->flags &= + ~(CT_SEGMENT_F_SERVER_DETACHED | CT_SEGMENT_F_CLIENT_DETACHED); __atomic_add_fetch (&ct_seg->server_n_sessions, 1, __ATOMIC_RELAXED); __atomic_add_fetch (&ct_seg->client_n_sessions, 1, __ATOMIC_RELAXED); } @@ -573,10 +646,6 @@ ct_init_accepted_session (app_worker_t *server_wrk, ct_connection_t *ct, ls->tx_fifo->shr->master_session_index = ls->session_index; ls->rx_fifo->master_thread_index = ls->thread_index; ls->tx_fifo->master_thread_index = ls->thread_index; - ls->rx_fifo->segment_manager = sm_index; - ls->tx_fifo->segment_manager = sm_index; - ls->rx_fifo->segment_index = fs_index; - ls->tx_fifo->segment_index = fs_index; seg_handle = segment_manager_segment_handle (sm, fs); segment_manager_segment_reader_unlock (sm); @@ -587,23 +656,21 @@ ct_init_accepted_session (app_worker_t *server_wrk, ct_connection_t *ct, } static void -ct_accept_rpc_wrk_handler (void *accept_args) +ct_accept_one (u32 thread_index, u32 ho_index) { - u32 cct_index, ho_index, thread_index, ll_index; ct_connection_t *sct, *cct, *ho; transport_connection_t *ll_ct; app_worker_t *server_wrk; + u32 cct_index, ll_index; session_t *ss, *ll; /* * Alloc client ct and initialize from ho */ - thread_index = vlib_get_thread_index (); cct = ct_connection_alloc (thread_index); cct_index = cct->c_c_index; - ho_index = pointer_to_uword (accept_args); - ho = ct_connection_get (ho_index, 0); + ho = ct_half_open_get (ho_index); /* Unlikely but half-open session and transport could have been freed */ if (PREDICT_FALSE (!ho)) @@ -640,7 +707,7 @@ ct_accept_rpc_wrk_handler (void *accept_args) sct->c_rmt_port = 0; sct->c_lcl_port = ll_ct->lcl_port; sct->c_is_ip4 = cct->c_is_ip4; - clib_memcpy (&sct->c_lcl_ip, &ll_ct->lcl_ip, sizeof (ll_ct->lcl_ip)); + clib_memcpy (&sct->c_lcl_ip, &cct->c_rmt_ip, sizeof (cct->c_rmt_ip)); sct->client_wrk = cct->client_wrk; sct->c_proto = TRANSPORT_PROTO_NONE; sct->client_opaque = cct->client_opaque; @@ -659,7 +726,7 @@ ct_accept_rpc_wrk_handler (void *accept_args) sct->c_is_ip4); ss->connection_index = sct->c_c_index; ss->listener_handle = listen_session_get_handle (ll); - ss->session_state = SESSION_STATE_CREATED; + session_set_state (ss, SESSION_STATE_CREATED); server_wrk = application_listener_select_worker (ll); ss->app_wrk_index = server_wrk->wrk_index; @@ -675,15 +742,17 @@ ct_accept_rpc_wrk_handler (void *accept_args) return; } + cct->server_wrk = sct->server_wrk; cct->seg_ctx_index = sct->seg_ctx_index; cct->ct_seg_index = sct->ct_seg_index; cct->client_rx_fifo = ss->tx_fifo; cct->client_tx_fifo = ss->rx_fifo; cct->client_rx_fifo->refcnt++; cct->client_tx_fifo->refcnt++; - cct->segment_handle = sct->segment_handle; + cct->segment_handle = + ct_client_seg_handle (sct->segment_handle, cct->client_wrk); - ss->session_state = SESSION_STATE_ACCEPTING; + session_set_state (ss, SESSION_STATE_ACCEPTING); if (app_worker_accept_notify (server_wrk, ss)) { ct_session_connect_notify (ss, SESSION_E_REFUSED); @@ -693,13 +762,93 @@ ct_accept_rpc_wrk_handler (void *accept_args) } } -static int -ct_connect (app_worker_t * client_wrk, session_t * ll, - session_endpoint_cfg_t * sep) +static void +ct_accept_rpc_wrk_handler (void *rpc_args) { - u32 thread_index, ho_index; + u32 thread_index, n_connects, i, n_pending; + const u32 max_connects = 32; + ct_worker_t *wrk; + u8 need_rpc = 0; + + thread_index = pointer_to_uword (rpc_args); + wrk = ct_worker_get (thread_index); + + /* Connects could be handled without worker barrier so grab lock */ + clib_spinlock_lock (&wrk->pending_connects_lock); + + n_pending = clib_fifo_elts (wrk->pending_connects); + n_connects = clib_min (n_pending, max_connects); + vec_validate (wrk->new_connects, n_connects); + + for (i = 0; i < n_connects; i++) + clib_fifo_sub1 (wrk->pending_connects, wrk->new_connects[i]); + + if (n_pending == n_connects) + wrk->have_connects = 0; + else + need_rpc = 1; + + clib_spinlock_unlock (&wrk->pending_connects_lock); + + for (i = 0; i < n_connects; i++) + ct_accept_one (thread_index, wrk->new_connects[i]); + + if (need_rpc) + session_send_rpc_evt_to_thread_force ( + thread_index, ct_accept_rpc_wrk_handler, + uword_to_pointer (thread_index, void *)); +} + +static void +ct_fwrk_flush_connects (void *rpc_args) +{ + u32 thread_index, fwrk_index, n_workers; ct_main_t *cm = &ct_main; - ct_connection_t *ho; + ct_worker_t *wrk; + u8 need_rpc; + + fwrk_index = cm->fwrk_thread; + n_workers = vec_len (cm->fwrk_pending_connects); + + for (thread_index = fwrk_index; thread_index < n_workers; thread_index++) + { + if (!vec_len (cm->fwrk_pending_connects[thread_index])) + continue; + + wrk = ct_worker_get (thread_index); + + /* Connects can be done without worker barrier, grab dst worker lock */ + if (thread_index != fwrk_index) + clib_spinlock_lock (&wrk->pending_connects_lock); + + clib_fifo_add (wrk->pending_connects, + cm->fwrk_pending_connects[thread_index], + vec_len (cm->fwrk_pending_connects[thread_index])); + if (!wrk->have_connects) + { + wrk->have_connects = 1; + need_rpc = 1; + } + + if (thread_index != fwrk_index) + clib_spinlock_unlock (&wrk->pending_connects_lock); + + vec_reset_length (cm->fwrk_pending_connects[thread_index]); + + if (need_rpc) + session_send_rpc_evt_to_thread_force ( + thread_index, ct_accept_rpc_wrk_handler, + uword_to_pointer (thread_index, void *)); + } + + cm->fwrk_have_flush = 0; +} + +static void +ct_program_connect_to_wrk (u32 ho_index) +{ + ct_main_t *cm = &ct_main; + u32 thread_index; /* Simple round-robin policy for spreading sessions over workers. We skip * thread index 0, i.e., offset the index by 1, when we have workers as it @@ -708,6 +857,25 @@ ct_connect (app_worker_t * client_wrk, session_t * ll, cm->n_sessions += 1; thread_index = cm->n_workers ? (cm->n_sessions % cm->n_workers) + 1 : 0; + /* Pospone flushing of connect request to dst worker until after session + * layer fully initializes the half-open session. */ + vec_add1 (cm->fwrk_pending_connects[thread_index], ho_index); + if (!cm->fwrk_have_flush) + { + session_send_rpc_evt_to_thread_force ( + cm->fwrk_thread, ct_fwrk_flush_connects, + uword_to_pointer (thread_index, void *)); + cm->fwrk_have_flush = 1; + } +} + +static int +ct_connect (app_worker_t *client_wrk, session_t *ll, + session_endpoint_cfg_t *sep) +{ + ct_connection_t *ho; + u32 ho_index; + /* * Alloc and init client half-open transport */ @@ -725,22 +893,19 @@ ct_connect (app_worker_t * client_wrk, session_t * ll, clib_memcpy (&ho->c_rmt_ip, &sep->ip, sizeof (sep->ip)); ho->flags |= CT_CONN_F_CLIENT; ho->c_s_index = ~0; - ho->actual_tp = sep->transport_proto; + ho->actual_tp = sep->original_tp; /* - * Accept connection on thread selected above. Connected reply comes + * Program connect on a worker, connected reply comes * after server accepts the connection. */ - - session_send_rpc_evt_to_thread_force (thread_index, - ct_accept_rpc_wrk_handler, - uword_to_pointer (ho_index, void *)); + ct_program_connect_to_wrk (ho_index); return ho_index; } static u32 -ct_start_listen (u32 app_listener_index, transport_endpoint_t * tep) +ct_start_listen (u32 app_listener_index, transport_endpoint_cfg_t *tep) { session_endpoint_cfg_t *sep; ct_connection_t *ct; @@ -772,9 +937,9 @@ ct_listener_get (u32 ct_index) } static transport_connection_t * -ct_half_open_get (u32 ct_index) +ct_session_half_open_get (u32 ct_index) { - return (transport_connection_t *) ct_connection_get (ct_index, 0); + return (transport_connection_t *) ct_half_open_get (ct_index); } static void @@ -796,7 +961,10 @@ ct_session_cleanup (u32 conn_index, u32 thread_index) static void ct_cleanup_ho (u32 ho_index) { - ct_connection_free (ct_connection_get (ho_index, 0)); + ct_connection_t *ho; + + ho = ct_half_open_get (ho_index); + ct_connection_free (ho); } static int @@ -827,7 +995,7 @@ ct_session_connect (transport_endpoint_cfg_t * tep) goto global_scope; ll = listen_session_get_from_handle (lh); - al = app_listener_get_w_session (ll); + al = app_listener_get (ll->al_index); /* * Break loop if rule in local table points to connecting app. This @@ -856,58 +1024,189 @@ global_scope: ll = session_lookup_listener_wildcard (table_index, sep); /* Avoid connecting app to own listener */ - if (ll && ll->app_index != app->app_index) - return ct_connect (app_wrk, ll, sep_ext); + if (ll) + { + al = app_listener_get (ll->al_index); + if (al->app_index != app->app_index) + return ct_connect (app_wrk, ll, sep_ext); + } /* Failed to connect but no error */ return SESSION_E_LOCAL_CONNECT; } +static inline int +ct_close_is_reset (ct_connection_t *ct, session_t *s) +{ + if (ct->flags & CT_CONN_F_RESET) + return 1; + if (ct->flags & CT_CONN_F_CLIENT) + return (svm_fifo_max_dequeue (ct->client_rx_fifo) > 0); + else + return (svm_fifo_max_dequeue (s->rx_fifo) > 0); +} + static void -ct_session_close (u32 ct_index, u32 thread_index) +ct_session_cleanup_server_session (session_t *s) { - ct_connection_t *ct, *peer_ct; + ct_connection_t *ct; + + ct = (ct_connection_t *) session_get_transport (s); + ct_session_dealloc_fifos (ct, s->rx_fifo, s->tx_fifo); + session_free (s); + ct_connection_free (ct); +} + +static void +ct_session_postponed_cleanup (ct_connection_t *ct) +{ + ct_connection_t *peer_ct; app_worker_t *app_wrk; session_t *s; - ct = ct_connection_get (ct_index, thread_index); s = session_get (ct->c_s_index, ct->c_thread_index); - peer_ct = ct_connection_get (ct->peer_index, thread_index); + app_wrk = app_worker_get_if_valid (s->app_wrk_index); + + peer_ct = ct_connection_get (ct->peer_index, ct->c_thread_index); if (peer_ct) { - peer_ct->peer_index = ~0; - /* Make sure session was allocated */ - if (peer_ct->flags & CT_CONN_F_HALF_OPEN) - { - ct_session_connect_notify (s, SESSION_E_REFUSED); - } - else if (peer_ct->c_s_index != ~0) - session_transport_closing_notify (&peer_ct->connection); + if (ct_close_is_reset (ct, s)) + session_transport_reset_notify (&peer_ct->connection); else - ct_connection_free (peer_ct); + session_transport_closing_notify (&peer_ct->connection); } + session_transport_closed_notify (&ct->connection); + + /* It would be cleaner to call session_transport_delete_notify + * but then we can't control session cleanup lower */ + session_set_state (s, SESSION_STATE_TRANSPORT_DELETED); + if (app_wrk) + app_worker_cleanup_notify (app_wrk, s, SESSION_CLEANUP_TRANSPORT); if (ct->flags & CT_CONN_F_CLIENT) { /* Normal free for client session as the fifos are allocated through * the connects segment manager in a segment that's not shared with * the server */ - session_free_w_fifos (s); ct_session_dealloc_fifos (ct, ct->client_rx_fifo, ct->client_tx_fifo); + session_program_cleanup (s); + ct_connection_free (ct); } else { /* Manual session and fifo segment cleanup to avoid implicit * segment manager cleanups and notifications */ - app_wrk = app_worker_get_if_valid (s->app_wrk_index); if (app_wrk) - app_worker_cleanup_notify (app_wrk, s, SESSION_CLEANUP_SESSION); + { + /* Remove custom cleanup notify infra when/if switching to normal + * session cleanup. Note that ct is freed in the cb function */ + app_worker_cleanup_notify_custom (app_wrk, s, + SESSION_CLEANUP_SESSION, + ct_session_cleanup_server_session); + } + else + { + ct_connection_free (ct); + } + } +} + +static void +ct_handle_cleanups (void *args) +{ + uword thread_index = pointer_to_uword (args); + const u32 max_cleanups = 100; + ct_cleanup_req_t *req; + ct_connection_t *ct; + u32 n_to_handle = 0; + ct_worker_t *wrk; + session_t *s; + + wrk = ct_worker_get (thread_index); + wrk->have_cleanups = 0; + n_to_handle = clib_fifo_elts (wrk->pending_cleanups); + n_to_handle = clib_min (n_to_handle, max_cleanups); + + while (n_to_handle) + { + clib_fifo_sub2 (wrk->pending_cleanups, req); + ct = ct_connection_get (req->ct_index, thread_index); + s = session_get (ct->c_s_index, ct->c_thread_index); + if (svm_fifo_has_event (s->tx_fifo) || (s->flags & SESSION_F_RX_EVT)) + clib_fifo_add1 (wrk->pending_cleanups, *req); + else + ct_session_postponed_cleanup (ct); + n_to_handle -= 1; + } - ct_session_dealloc_fifos (ct, s->rx_fifo, s->tx_fifo); - session_free (s); + if (clib_fifo_elts (wrk->pending_cleanups)) + { + wrk->have_cleanups = 1; + session_send_rpc_evt_to_thread_force ( + thread_index, ct_handle_cleanups, + uword_to_pointer (thread_index, void *)); } +} - ct_connection_free (ct); +static void +ct_program_cleanup (ct_connection_t *ct) +{ + ct_cleanup_req_t *req; + uword thread_index; + ct_worker_t *wrk; + + thread_index = ct->c_thread_index; + wrk = ct_worker_get (ct->c_thread_index); + + clib_fifo_add2 (wrk->pending_cleanups, req); + req->ct_index = ct->c_c_index; + + if (wrk->have_cleanups) + return; + + wrk->have_cleanups = 1; + session_send_rpc_evt_to_thread_force ( + thread_index, ct_handle_cleanups, uword_to_pointer (thread_index, void *)); +} + +static void +ct_session_close (u32 ct_index, u32 thread_index) +{ + ct_connection_t *ct, *peer_ct; + session_t *s; + + ct = ct_connection_get (ct_index, thread_index); + s = session_get (ct->c_s_index, ct->c_thread_index); + peer_ct = ct_connection_get (ct->peer_index, thread_index); + if (peer_ct) + { + peer_ct->peer_index = ~0; + /* Make sure session was allocated */ + if (peer_ct->flags & CT_CONN_F_HALF_OPEN) + { + ct_session_connect_notify (s, SESSION_E_REFUSED); + ct->peer_index = ~0; + } + else if (peer_ct->c_s_index == ~0) + { + /* should not happen */ + clib_warning ("ct peer without session"); + ct_connection_free (peer_ct); + } + } + + /* Do not send closed notify to make sure pending tx events are + * still delivered and program cleanup */ + ct_program_cleanup (ct); +} + +static void +ct_session_reset (u32 ct_index, u32 thread_index) +{ + ct_connection_t *ct; + ct = ct_connection_get (ct_index, thread_index); + ct->flags |= CT_CONN_F_RESET; + ct_session_close (ct_index, thread_index); } static transport_connection_t * @@ -966,12 +1265,17 @@ static int ct_app_rx_evt (transport_connection_t * tc) { ct_connection_t *ct = (ct_connection_t *) tc, *peer_ct; - session_t *ps; + session_t *ps, *s; + s = session_get (ct->c_s_index, ct->c_thread_index); + if (session_has_transport (s) || s->session_state < SESSION_STATE_READY) + return -1; peer_ct = ct_connection_get (ct->peer_index, tc->thread_index); - if (!peer_ct) + if (!peer_ct || (peer_ct->flags & CT_CONN_F_HALF_OPEN)) return -1; ps = session_get (peer_ct->c_s_index, peer_ct->c_thread_index); + if (ps->session_state >= SESSION_STATE_TRANSPORT_CLOSING) + return -1; return session_dequeue_notify (ps); } @@ -993,7 +1297,7 @@ format_ct_half_open (u8 *s, va_list *args) { u32 ho_index = va_arg (*args, u32); u32 verbose = va_arg (*args, u32); - ct_connection_t *ct = ct_connection_get (ho_index, 0); + ct_connection_t *ct = ct_half_open_get (ho_index); s = format (s, "%-" SESSION_CLI_ID_LEN "U", format_ct_connection_id, ct); if (verbose) s = format (s, "%-" SESSION_CLI_STATE_LEN "s", "HALF-OPEN"); @@ -1042,27 +1346,33 @@ format_ct_session (u8 * s, va_list * args) clib_error_t * ct_enable_disable (vlib_main_t * vm, u8 is_en) { + vlib_thread_main_t *vtm = &vlib_thread_main; ct_main_t *cm = &ct_main; + ct_worker_t *wrk; cm->n_workers = vlib_num_workers (); - vec_validate (cm->connections, cm->n_workers); + cm->fwrk_thread = transport_cl_thread (); + vec_validate (cm->wrk, vtm->n_vlib_mains); + vec_foreach (wrk, cm->wrk) + clib_spinlock_init (&wrk->pending_connects_lock); clib_spinlock_init (&cm->ho_reuseable_lock); clib_rwlock_init (&cm->app_segs_lock); + vec_validate (cm->fwrk_pending_connects, cm->n_workers); return 0; } -/* *INDENT-OFF* */ static const transport_proto_vft_t cut_thru_proto = { .enable = ct_enable_disable, .start_listen = ct_start_listen, .stop_listen = ct_stop_listen, .get_connection = ct_session_get, .get_listener = ct_listener_get, - .get_half_open = ct_half_open_get, + .get_half_open = ct_session_half_open_get, .cleanup = ct_session_cleanup, .cleanup_ho = ct_cleanup_ho, .connect = ct_session_connect, .close = ct_session_close, + .reset = ct_session_reset, .custom_tx = ct_custom_tx, .app_rx_evt = ct_app_rx_evt, .format_listener = format_ct_listener, @@ -1075,7 +1385,14 @@ static const transport_proto_vft_t cut_thru_proto = { .service_type = TRANSPORT_SERVICE_VC, }, }; -/* *INDENT-ON* */ + +static inline int +ct_session_can_tx (session_t *s) +{ + return (s->session_state == SESSION_STATE_READY || + s->session_state == SESSION_STATE_CLOSING || + s->session_state == SESSION_STATE_APP_CLOSED); +} int ct_session_tx (session_t * s) @@ -1083,6 +1400,8 @@ ct_session_tx (session_t * s) ct_connection_t *ct, *peer_ct; session_t *peer_s; + if (!ct_session_can_tx (s)) + return 0; ct = (ct_connection_t *) session_get_transport (s); peer_ct = ct_connection_get (ct->peer_index, ct->c_thread_index); if (!peer_ct) @@ -1090,6 +1409,7 @@ ct_session_tx (session_t * s) peer_s = session_get (peer_ct->c_s_index, peer_ct->c_thread_index); if (peer_s->session_state >= SESSION_STATE_TRANSPORT_CLOSING) return 0; + peer_s->flags |= SESSION_F_RX_EVT; return session_enqueue_notify (peer_s); } |