From 6792ec059696a358b6c98d8d86e9740b34c01e24 Mon Sep 17 00:00:00 2001 From: Florin Coras Date: Mon, 13 Mar 2017 03:49:51 -0700 Subject: TCP/session improvements - Added svm fifo flag for tracking fifo dequeue events (replaces event length). Updated all code to switch to the new scheme. - More session debugging - Fix peek index wrap - Add a trivial socket test client - Fast retransmit/cc fixes - tx and rx SACK fixes and unit testing - SRTT computation fix - remove dupack/ack burst filters - improve ack rx - improved segment rx - builtin client test code Change-Id: Ic4eb2d5ca446eb2260ccd3ccbcdaa73c64e7f4e1 Signed-off-by: Florin Coras Signed-off-by: Dave Barach --- src/vnet/session/application.h | 3 +- src/vnet/session/node.c | 127 +++++++++++++++++++++++---------------- src/vnet/session/session.c | 63 +++++++++++++------ src/vnet/session/session.h | 19 +++--- src/vnet/session/session_cli.c | 2 +- src/vnet/session/session_debug.h | 38 ++++++++++-- src/vnet/session/transport.h | 2 +- 7 files changed, 165 insertions(+), 89 deletions(-) (limited to 'src/vnet/session') diff --git a/src/vnet/session/application.h b/src/vnet/session/application.h index a60a8b8b5f9..480828f7b34 100644 --- a/src/vnet/session/application.h +++ b/src/vnet/session/application.h @@ -45,8 +45,7 @@ typedef struct _stream_session_cb_vft void (*session_reset_callback) (stream_session_t * s); /* Direct RX callback, for built-in servers */ - int (*builtin_server_rx_callback) (stream_session_t * session, - session_fifo_event_t * ep); + int (*builtin_server_rx_callback) (stream_session_t * session); /* Redirect connection to local server */ int (*redirect_connect_callback) (u32 api_client_index, void *mp); diff --git a/src/vnet/session/node.c b/src/vnet/session/node.c index 822afebde0c..8681105c284 100644 --- a/src/vnet/session/node.c +++ b/src/vnet/session/node.c @@ -13,21 +13,14 @@ * limitations under the License. */ +#include #include #include -#include -#include - #include - -#include -#include #include -#include - -#include -#include +#include #include +#include vlib_node_registration_t session_queue_node; @@ -52,8 +45,8 @@ format_session_queue_trace (u8 * s, va_list * args) vlib_node_registration_t session_queue_node; -#define foreach_session_queue_error \ -_(TX, "Packets transmitted") \ +#define foreach_session_queue_error \ +_(TX, "Packets transmitted") \ _(TIMER, "Timer events") typedef enum @@ -91,10 +84,10 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node, transport_proto_vft_t *transport_vft; u32 next_index, next0, *to_next, n_left_to_next, bi0; vlib_buffer_t *b0; - u32 rx_offset; + u32 rx_offset = 0, max_dequeue0; u16 snd_mss0; u8 *data0; - int i; + int i, n_bytes_read; next_index = next0 = session_type_to_next[s0->session_type]; @@ -106,24 +99,33 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node, snd_mss0 = transport_vft->send_mss (tc0); /* Can't make any progress */ - if (snd_space0 == 0 || svm_fifo_max_dequeue (s0->server_tx_fifo) == 0 - || snd_mss0 == 0) + if (snd_space0 == 0 || snd_mss0 == 0) { vec_add1 (smm->evts_partially_read[thread_index], *e0); return 0; } - ASSERT (e0->enqueue_length > 0); - - /* Ensure we're not writing more than transport window allows */ - max_len_to_snd0 = clib_min (e0->enqueue_length, snd_space0); - if (peek_data) { /* Offset in rx fifo from where to peek data */ rx_offset = transport_vft->tx_fifo_offset (tc0); } + /* Check how much we can pull. If buffering, subtract the offset */ + max_dequeue0 = svm_fifo_max_dequeue (s0->server_tx_fifo) - rx_offset; + + /* Allow enqueuing of a new event */ + svm_fifo_unset_event (s0->server_tx_fifo); + + /* Nothing to read return */ + if (max_dequeue0 == 0) + { + return 0; + } + + /* Ensure we're not writing more than transport window allows */ + max_len_to_snd0 = clib_min (max_dequeue0, snd_space0); + /* TODO check if transport is willing to send len_to_snd0 * bytes (Nagle) */ @@ -147,13 +149,10 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node, * XXX 0.9 because when debugging we might not get a full frame */ if (PREDICT_FALSE (n_bufs < 0.9 * VLIB_FRAME_SIZE)) { - /* Keep track of how much we've dequeued and exit */ - if (left_to_snd0 != max_len_to_snd0) + if (svm_fifo_set_event (s0->server_tx_fifo)) { - e0->enqueue_length -= max_len_to_snd0 - left_to_snd0; vec_add1 (smm->evts_partially_read[thread_index], *e0); } - return -1; } @@ -198,9 +197,9 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node, len_to_deq0 = (left_to_snd0 < snd_mss0) ? left_to_snd0 : snd_mss0; /* *INDENT-OFF* */ - SESSION_EVT_DBG(s0, SESSION_EVT_DEQ, ({ + SESSION_EVT_DBG(SESSION_EVT_DEQ, s0, ({ ed->data[0] = e0->event_id; - ed->data[1] = e0->enqueue_length; + ed->data[1] = max_dequeue0; ed->data[2] = len_to_deq0; ed->data[3] = left_to_snd0; })); @@ -214,29 +213,30 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node, * 2) buffer chains */ if (peek_data) { - int n_bytes_read; n_bytes_read = svm_fifo_peek (s0->server_tx_fifo, s0->pid, rx_offset, len_to_deq0, data0); - if (n_bytes_read < 0) + if (n_bytes_read <= 0) goto dequeue_fail; /* Keep track of progress locally, transport is also supposed to - * increment it independently when pushing header */ + * increment it independently when pushing the header */ rx_offset += n_bytes_read; } else { - if (svm_fifo_dequeue_nowait (s0->server_tx_fifo, s0->pid, - len_to_deq0, data0) < 0) + n_bytes_read = svm_fifo_dequeue_nowait (s0->server_tx_fifo, + s0->pid, len_to_deq0, + data0); + if (n_bytes_read <= 0) goto dequeue_fail; } - b0->current_length = len_to_deq0; + b0->current_length = n_bytes_read; /* Ask transport to push header */ transport_vft->push_header (tc0, b0); - left_to_snd0 -= len_to_deq0; + left_to_snd0 -= n_bytes_read; *n_tx_packets = *n_tx_packets + 1; vlib_validate_buffer_enqueue_x1 (vm, node, next_index, @@ -246,25 +246,31 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node, vlib_put_next_frame (vm, node, next_index, n_left_to_next); } - /* If we couldn't dequeue all bytes store progress */ - if (max_len_to_snd0 < e0->enqueue_length) + /* If we couldn't dequeue all bytes mark as partially read */ + if (max_len_to_snd0 < max_dequeue0) { - e0->enqueue_length -= max_len_to_snd0; - vec_add1 (smm->evts_partially_read[thread_index], *e0); + /* If we don't already have new event */ + if (svm_fifo_set_event (s0->server_tx_fifo)) + { + vec_add1 (smm->evts_partially_read[thread_index], *e0); + } } return 0; dequeue_fail: - /* Can't read from fifo. Store event rx progress, save as partially read, - * return buff to free list and return */ - e0->enqueue_length -= max_len_to_snd0 - left_to_snd0; - vec_add1 (smm->evts_partially_read[thread_index], *e0); + /* + * Can't read from fifo. If we don't already have an event, save as partially + * read, return buff to free list and return + */ + clib_warning ("dequeue fail"); - to_next -= 1; - n_left_to_next += 1; + if (svm_fifo_set_event (s0->server_tx_fifo)) + { + vec_add1 (smm->evts_partially_read[thread_index], *e0); + } + vlib_put_next_frame (vm, node, next_index, n_left_to_next + 1); _vec_len (smm->tx_buffers[thread_index]) += 1; - clib_warning ("dequeue fail"); return 0; } @@ -298,6 +304,7 @@ session_queue_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node, session_fifo_event_t *my_fifo_events, *e; u32 n_to_dequeue, n_events; unix_shared_memory_queue_t *q; + application_t *app; int n_tx_packets = 0; u32 my_thread_index = vm->cpu_index; int i, rv; @@ -321,13 +328,18 @@ session_queue_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node, if (n_to_dequeue == 0 && vec_len (my_fifo_events) == 0) return 0; + SESSION_EVT_DBG (SESSION_EVT_DEQ_NODE, 0); + /* * If we didn't manage to process previous events try going * over them again without dequeuing new ones. */ /* XXX: Block senders to sessions that can't keep up */ if (vec_len (my_fifo_events) >= 100) - goto skip_dequeue; + { + clib_warning ("too many fifo events unsolved"); + goto skip_dequeue; + } /* See you in the next life, don't be late */ if (pthread_mutex_trylock (&q->mutex)) @@ -352,19 +364,17 @@ skip_dequeue: { svm_fifo_t *f0; /* $$$ prefetch 1 ahead maybe */ stream_session_t *s0; - u32 server_session_index0, server_thread_index0; + u32 session_index0; session_fifo_event_t *e0; e0 = &my_fifo_events[i]; f0 = e0->fifo; - server_session_index0 = f0->server_session_index; - server_thread_index0 = f0->server_thread_index; + session_index0 = f0->server_session_index; /* $$$ add multiple event queues, per vpp worker thread */ - ASSERT (server_thread_index0 == my_thread_index); + ASSERT (f0->server_thread_index == my_thread_index); - s0 = stream_session_get_if_valid (server_session_index0, - my_thread_index); + s0 = stream_session_get_if_valid (session_index0, my_thread_index); if (CLIB_DEBUG && !s0) { @@ -385,11 +395,20 @@ skip_dequeue: rv = (smm->session_tx_fns[s0->session_type]) (vm, node, smm, e0, s0, my_thread_index, &n_tx_packets); + /* Out of buffers */ if (rv < 0) goto done; break; - + case FIFO_EVENT_SERVER_EXIT: + stream_session_disconnect (s0); + break; + case FIFO_EVENT_BUILTIN_RX: + svm_fifo_unset_event (s0->server_rx_fifo); + /* Get session's server */ + app = application_get (s0->app_index); + app->cb_fns.builtin_server_rx_callback (s0); + break; default: clib_warning ("unhandled event type %d", e0->event_type); } @@ -418,6 +437,8 @@ done: vlib_node_increment_counter (vm, session_queue_node.index, SESSION_QUEUE_ERROR_TX, n_tx_packets); + SESSION_EVT_DBG (SESSION_EVT_DEQ_NODE, 1); + return n_tx_packets; } diff --git a/src/vnet/session/session.c b/src/vnet/session/session.c index 06e2a09af31..f10918aa3c4 100644 --- a/src/vnet/session/session.c +++ b/src/vnet/session/session.c @@ -804,30 +804,36 @@ stream_session_enqueue_notify (stream_session_t * s, u8 block) /* Get session's server */ app = application_get (s->app_index); - /* Fabricate event */ - evt.fifo = s->server_rx_fifo; - evt.event_type = FIFO_EVENT_SERVER_RX; - evt.event_id = serial_number++; - evt.enqueue_length = svm_fifo_max_dequeue (s->server_rx_fifo); - /* Built-in server? Hand event to the callback... */ if (app->cb_fns.builtin_server_rx_callback) - return app->cb_fns.builtin_server_rx_callback (s, &evt); - - /* Add event to server's event queue */ - q = app->event_queue; + return app->cb_fns.builtin_server_rx_callback (s); - /* Based on request block (or not) for lack of space */ - if (block || PREDICT_TRUE (q->cursize < q->maxsize)) - unix_shared_memory_queue_add (app->event_queue, (u8 *) & evt, - 0 /* do wait for mutex */ ); - else - return -1; + /* If no event, send one */ + if (svm_fifo_set_event (s->server_rx_fifo)) + { + /* Fabricate event */ + evt.fifo = s->server_rx_fifo; + evt.event_type = FIFO_EVENT_SERVER_RX; + evt.event_id = serial_number++; + + /* Add event to server's event queue */ + q = app->event_queue; + + /* Based on request block (or not) for lack of space */ + if (block || PREDICT_TRUE (q->cursize < q->maxsize)) + unix_shared_memory_queue_add (app->event_queue, (u8 *) & evt, + 0 /* do wait for mutex */ ); + else + { + clib_warning ("fifo full"); + return -1; + } + } /* *INDENT-OFF* */ - SESSION_EVT_DBG(s, SESSION_EVT_ENQ, ({ + SESSION_EVT_DBG(SESSION_EVT_ENQ, s, ({ ed->data[0] = evt.event_id; - ed->data[1] = evt.enqueue_length; + ed->data[1] = svm_fifo_max_dequeue (s->server_rx_fifo); })); /* *INDENT-ON* */ @@ -1192,8 +1198,29 @@ stream_session_open (u8 sst, ip46_address_t * addr, u16 port_host_byte_order, void stream_session_disconnect (stream_session_t * s) { +// session_fifo_event_t evt; + s->session_state = SESSION_STATE_CLOSED; + /* RPC to vpp evt queue in the right thread */ + tp_vfts[s->session_type].close (s->connection_index, s->thread_index); + +// { +// /* Fabricate event */ +// evt.fifo = s->server_rx_fifo; +// evt.event_type = FIFO_EVENT_SERVER_RX; +// evt.event_id = serial_number++; +// +// /* Based on request block (or not) for lack of space */ +// if (PREDICT_TRUE(q->cursize < q->maxsize)) +// unix_shared_memory_queue_add (app->event_queue, (u8 *) &evt, +// 0 /* do wait for mutex */); +// else +// { +// clib_warning("fifo full"); +// return -1; +// } +// } } /** diff --git a/src/vnet/session/session.h b/src/vnet/session/session.h index 96c00d87766..a39bc06ffca 100644 --- a/src/vnet/session/session.h +++ b/src/vnet/session/session.h @@ -33,6 +33,7 @@ typedef enum FIFO_EVENT_SERVER_TX, FIFO_EVENT_TIMEOUT, FIFO_EVENT_SERVER_EXIT, + FIFO_EVENT_BUILTIN_RX } fifo_event_type_t; #define foreach_session_input_error \ @@ -91,14 +92,13 @@ typedef enum SESSION_STATE_N_STATES, } stream_session_state_t; -typedef CLIB_PACKED (struct - { - svm_fifo_t * fifo; - u8 event_type; - /* $$$$ for event logging */ - u16 event_id; - u32 enqueue_length; - }) session_fifo_event_t; +/* *INDENT-OFF* */ +typedef CLIB_PACKED (struct { + svm_fifo_t * fifo; + u8 event_type; + u16 event_id; +}) session_fifo_event_t; +/* *INDENT-ON* */ typedef struct _stream_session_t { @@ -333,7 +333,7 @@ stream_session_get_index (stream_session_t * s) } always_inline u32 -stream_session_max_enqueue (transport_connection_t * tc) +stream_session_max_rx_enqueue (transport_connection_t * tc) { stream_session_t *s = stream_session_get (tc->s_index, tc->thread_index); return svm_fifo_max_enqueue (s->server_rx_fifo); @@ -346,7 +346,6 @@ stream_session_fifo_size (transport_connection_t * tc) return s->server_rx_fifo->nitems; } - int stream_session_enqueue_data (transport_connection_t * tc, u8 * data, u16 len, u8 queue_event); diff --git a/src/vnet/session/session_cli.c b/src/vnet/session/session_cli.c index b029ee6549a..38762afc8bb 100644 --- a/src/vnet/session/session_cli.c +++ b/src/vnet/session/session_cli.c @@ -107,7 +107,7 @@ show_session_command_fn (vlib_main_t * vm, unformat_input_t * input, { if (once_per_pool) { - str = format (str, "%-40s%-20s%-20s%-15s", + str = format (str, "%-50s%-20s%-20s%-15s", "Connection", "Rx fifo", "Tx fifo", "Session Index"); vlib_cli_output (vm, "%v", str); diff --git a/src/vnet/session/session_debug.h b/src/vnet/session/session_debug.h index 858f12e0287..80a97cd5f09 100644 --- a/src/vnet/session/session_debug.h +++ b/src/vnet/session/session_debug.h @@ -21,7 +21,8 @@ #define foreach_session_dbg_evt \ _(ENQ, "enqueue") \ - _(DEQ, "dequeue") + _(DEQ, "dequeue") \ + _(DEQ_NODE, "dequeue") typedef enum _session_evt_dbg { @@ -30,7 +31,10 @@ typedef enum _session_evt_dbg #undef _ } session_evt_dbg_e; -#if TRANSPORT_DEBUG +#define SESSION_DBG (0) +#define SESSION_DEQ_NODE_EVTS (0) + +#if TRANSPORT_DEBUG && SESSION_DBG #define DEC_SESSION_ETD(_s, _e, _size) \ struct \ @@ -44,6 +48,12 @@ typedef enum _session_evt_dbg ed = ELOG_TRACK_DATA (&vlib_global_main.elog_main, \ _e, _tc->elog_track) +#define DEC_SESSION_ED(_e, _size) \ + struct \ + { \ + u32 data[_size]; \ + } * ed; \ + ed = ELOG_DATA (&vlib_global_main.elog_main, _e) #define SESSION_EVT_DEQ_HANDLER(_s, _body) \ { \ @@ -67,13 +77,33 @@ typedef enum _session_evt_dbg do { _body; } while (0); \ } +#if SESSION_DEQ_NODE_EVTS +#define SESSION_EVT_DEQ_NODE_HANDLER(_node_evt) \ +{ \ + ELOG_TYPE_DECLARE (_e) = \ + { \ + .format = "deq-node: %s", \ + .format_args = "t4", \ + .n_enum_strings = 2, \ + .enum_strings = { \ + "start", \ + "end", \ + }, \ + }; \ + DEC_SESSION_ED(_e, 1); \ + ed->data[0] = _node_evt; \ +} +#else +#define SESSION_EVT_DEQ_NODE_HANDLER(_node_evt) +#endif + #define CONCAT_HELPER(_a, _b) _a##_b #define CC(_a, _b) CONCAT_HELPER(_a, _b) -#define SESSION_EVT_DBG(_s, _evt, _body) CC(_evt, _HANDLER)(_s, _body) +#define SESSION_EVT_DBG(_evt, _args...) CC(_evt, _HANDLER)(_args) #else -#define SESSION_EVT_DBG(_s, _evt, _body) +#define SESSION_EVT_DBG(_evt, _args...) #endif #endif /* SRC_VNET_SESSION_SESSION_DEBUG_H_ */ diff --git a/src/vnet/session/transport.h b/src/vnet/session/transport.h index 421121d2fa6..2f912cbc987 100644 --- a/src/vnet/session/transport.h +++ b/src/vnet/session/transport.h @@ -38,7 +38,7 @@ typedef struct _transport_connection u32 thread_index; /**< Worker-thread index */ #if TRANSPORT_DEBUG - elog_track_t elog_track; /**< Debug purposes */ + elog_track_t elog_track; /**< Event logging */ #endif /** Macros for 'derived classes' where base is named "connection" */ -- cgit 1.2.3-korg