diff options
author | Florin Coras <fcoras@cisco.com> | 2018-05-04 15:46:57 -0700 |
---|---|---|
committer | Dave Barach <openvpp@barachs.net> | 2018-05-09 11:34:06 +0000 |
commit | 8e43d04ca4f4496aaefc4f5e2b6e1c0951624099 (patch) | |
tree | 3addc5766002d5224dde2c7fa4efe018480830e4 | |
parent | ee7f0bd9e7ce4106d3b9511b0efede4326bded51 (diff) |
session: cleanup session tx function
- rework the function to declutter and avoid building more than one tx
frame
- add dual loop although benefits in my tests seem to be minimal
- improve tcp/udp echo external apps. They have slightly better
throughput than internal echo apps.
- udp bugfixes
Change-Id: Iea4a245b1b1bb407a7f403dedcce2664a49f774b
Signed-off-by: Florin Coras <fcoras@cisco.com>
-rw-r--r-- | src/svm/svm_fifo.c | 6 | ||||
-rw-r--r-- | src/tests/vnet/session/tcp_echo.c | 59 | ||||
-rw-r--r-- | src/tests/vnet/session/udp_echo.c | 69 | ||||
-rw-r--r-- | src/vnet/session-apps/echo_client.c | 58 | ||||
-rw-r--r-- | src/vnet/session-apps/echo_client.h | 2 | ||||
-rw-r--r-- | src/vnet/session/application_interface.h | 6 | ||||
-rw-r--r-- | src/vnet/session/session.c | 1 | ||||
-rw-r--r-- | src/vnet/session/session.h | 76 | ||||
-rw-r--r-- | src/vnet/session/session_node.c | 631 |
9 files changed, 532 insertions, 376 deletions
diff --git a/src/svm/svm_fifo.c b/src/svm/svm_fifo.c index 3552192a768..dbdb813a7df 100644 --- a/src/svm/svm_fifo.c +++ b/src/svm/svm_fifo.c @@ -622,14 +622,14 @@ void svm_fifo_overwrite_head (svm_fifo_t * f, u8 * data, u32 len) { u32 first_chunk; + first_chunk = f->nitems - f->head; ASSERT (len <= f->nitems); - if (len < f->nitems - f->head) + if (len <= first_chunk) clib_memcpy (&f->data[f->head], data, len); else { - first_chunk = len - (f->nitems - f->head); clib_memcpy (&f->data[f->head], data, first_chunk); - clib_memcpy (f->data, data + first_chunk, len - first_chunk); + clib_memcpy (&f->data[0], data + first_chunk, len - first_chunk); } } diff --git a/src/tests/vnet/session/tcp_echo.c b/src/tests/vnet/session/tcp_echo.c index 0108b16ac7b..d3d4c0ad0a6 100644 --- a/src/tests/vnet/session/tcp_echo.c +++ b/src/tests/vnet/session/tcp_echo.c @@ -85,7 +85,7 @@ typedef struct int i_am_master; /* drop all packets */ - int drop_packets; + int no_return; /* Our event queue */ svm_queue_t *our_event_queue; @@ -117,6 +117,7 @@ typedef struct u32 client_bytes_received; u8 test_return_packets; u64 bytes_to_send; + u32 fifo_size; /** Flag that decides if socket, instead of svm, api is used to connect to * vpp. If sock api is used, shm binary api is subsequently bootstrapped @@ -250,7 +251,6 @@ application_send_attach (echo_main_t * em) vl_api_application_tls_cert_add_t *cert_mp; vl_api_application_tls_key_add_t *key_mp; - u32 fifo_size = 4 << 20; bmp = vl_msg_api_alloc (sizeof (*bmp)); memset (bmp, 0, sizeof (*bmp)); @@ -260,8 +260,8 @@ application_send_attach (echo_main_t * em) bmp->options[APP_OPTIONS_FLAGS] = APP_OPTIONS_FLAGS_ACCEPT_REDIRECT | APP_OPTIONS_FLAGS_ADD_SEGMENT; bmp->options[APP_OPTIONS_PREALLOC_FIFO_PAIRS] = 16; - bmp->options[APP_OPTIONS_RX_FIFO_SIZE] = fifo_size; - bmp->options[APP_OPTIONS_TX_FIFO_SIZE] = fifo_size; + bmp->options[APP_OPTIONS_RX_FIFO_SIZE] = em->fifo_size; + bmp->options[APP_OPTIONS_TX_FIFO_SIZE] = em->fifo_size; bmp->options[APP_OPTIONS_ADD_SEGMENT_SIZE] = 128 << 20; bmp->options[APP_OPTIONS_SEGMENT_SIZE] = 256 << 20; vl_msg_api_send_shmem (em->vl_input_queue, (u8 *) & bmp); @@ -757,11 +757,10 @@ vl_api_connect_session_reply_t_handler (vl_api_connect_session_reply_t * mp) static void send_test_chunk (echo_main_t * em, svm_fifo_t * tx_fifo, int mypid, u32 bytes) { + u32 bytes_to_snd, enq_space, min_chunk = 16 << 10; u8 *test_data = em->connect_test_data; u64 bytes_sent = 0; int test_buf_offset = 0; - u32 bytes_to_snd; - u32 queue_max_chunk = 128 << 10, actual_write; session_fifo_event_t evt; int rv; @@ -771,11 +770,12 @@ send_test_chunk (echo_main_t * em, svm_fifo_t * tx_fifo, int mypid, u32 bytes) while (bytes_to_snd > 0 && !em->time_to_stop) { - actual_write = (bytes_to_snd > queue_max_chunk) ? - queue_max_chunk : bytes_to_snd; - rv = svm_fifo_enqueue_nowait (tx_fifo, actual_write, + enq_space = svm_fifo_max_enqueue (tx_fifo); + if (enq_space < clib_min (bytes_to_snd, min_chunk)) + continue; + rv = svm_fifo_enqueue_nowait (tx_fifo, + clib_min (bytes_to_snd, enq_space), test_data + test_buf_offset); - if (rv > 0) { bytes_to_snd -= rv; @@ -787,9 +787,8 @@ send_test_chunk (echo_main_t * em, svm_fifo_t * tx_fifo, int mypid, u32 bytes) /* Fabricate TX event, send to vpp */ evt.fifo = tx_fifo; evt.event_type = FIFO_EVENT_APP_TX; - - svm_queue_add (em->vpp_event_queue, - (u8 *) & evt, 0 /* do wait for mutex */ ); + svm_queue_add (em->vpp_event_queue, (u8 *) & evt, + 0 /* do wait for mutex */ ); } } } @@ -824,7 +823,7 @@ client_send_data (echo_main_t * em) if (leftover) send_test_chunk (em, tx_fifo, mypid, leftover); - if (!em->drop_packets) + if (!em->no_return) { f64 timeout = clib_time_now (&em->clib_time) + 10; @@ -839,6 +838,7 @@ client_send_data (echo_main_t * em) } } } + em->time_to_stop = 1; } @@ -911,7 +911,7 @@ client_run (echo_main_t * em) } /* Init test data */ - vec_validate (em->connect_test_data, 128 * 1024 - 1); + vec_validate (em->connect_test_data, 1024 * 1024 - 1); for (i = 0; i < vec_len (em->connect_test_data); i++) em->connect_test_data[i] = i & 0xff; @@ -1130,7 +1130,7 @@ server_handle_fifo_event_rx (echo_main_t * em, session_fifo_event_t * e) } /* Reflect if a non-drop session */ - if (!em->drop_packets && n_read > 0) + if (!em->no_return && n_read > 0) { offset = 0; do @@ -1342,13 +1342,13 @@ tcp_echo_api_hookup (echo_main_t * em) int main (int argc, char **argv) { - int i_am_master = 1, drop_packets = 0, test_return_packets = 0; + int i_am_server = 1, test_return_packets = 0; echo_main_t *em = &echo_main; unformat_input_t _argv, *a = &_argv; u8 *chroot_prefix; u8 *heap, *uri = 0; u8 *bind_uri = (u8 *) "tcp://0.0.0.0/1234"; - u8 *connect_uri = (u8 *) "tcp://6.0.1.2/1234"; + u8 *connect_uri = (u8 *) "tcp://6.0.1.1/1234"; u64 bytes_to_send = 64 << 10, mbytes; char *app_name; u32 tmp; @@ -1365,11 +1365,11 @@ main (int argc, char **argv) vec_validate (em->rx_buf, 128 << 10); em->session_index_by_vpp_handles = hash_create (0, sizeof (uword)); - em->my_pid = getpid (); em->configured_segment_size = 1 << 20; em->socket_name = 0; em->use_sock_api = 1; + em->fifo_size = 64 << 10; clib_time_init (&em->clib_time); init_error_string_table (em); @@ -1388,12 +1388,12 @@ main (int argc, char **argv) em->configured_segment_size = tmp << 20; else if (unformat (a, "segment-size %dG", &tmp)) em->configured_segment_size = tmp << 30; - else if (unformat (a, "master")) - i_am_master = 1; - else if (unformat (a, "slave")) - i_am_master = 0; - else if (unformat (a, "drop")) - drop_packets = 1; + else if (unformat (a, "server")) + i_am_server = 1; + else if (unformat (a, "client")) + i_am_server = 0; + else if (unformat (a, "no-return")) + em->no_return = 1; else if (unformat (a, "test")) test_return_packets = 1; else if (unformat (a, "mbytes %lld", &mbytes)) @@ -1408,6 +1408,8 @@ main (int argc, char **argv) ; else if (unformat (a, "use-svm-api")) em->use_sock_api = 0; + else if (unformat (a, "fifo-size %d", &tmp)) + em->fifo_size = tmp << 10; else { fformat (stderr, "%s: usage [master|slave]\n"); @@ -1429,9 +1431,8 @@ main (int argc, char **argv) em->connect_uri = format (0, "%s%c", connect_uri, 0); } - em->i_am_master = i_am_master; + em->i_am_master = i_am_server; em->segment_main = &svm_fifo_segment_main; - em->drop_packets = drop_packets; em->test_return_packets = test_return_packets; em->bytes_to_send = bytes_to_send; em->time_to_stop = 0; @@ -1439,7 +1440,7 @@ main (int argc, char **argv) setup_signal_handlers (); tcp_echo_api_hookup (em); - app_name = i_am_master ? "tcp_echo_server" : "tcp_echo_client"; + app_name = i_am_server ? "tcp_echo_server" : "tcp_echo_client"; if (connect_to_vpp (app_name) < 0) { svm_region_exit (); @@ -1447,7 +1448,7 @@ main (int argc, char **argv) exit (1); } - if (i_am_master == 0) + if (i_am_server == 0) client_run (em); else server_run (em); diff --git a/src/tests/vnet/session/udp_echo.c b/src/tests/vnet/session/udp_echo.c index fe5461d2c59..54e00b181bd 100644 --- a/src/tests/vnet/session/udp_echo.c +++ b/src/tests/vnet/session/udp_echo.c @@ -125,6 +125,8 @@ typedef struct uword *segments_table; u8 do_echo; + u8 have_return; + u64 bytes_to_send; } udp_echo_main_t; #if CLIB_DEBUG > 0 @@ -147,7 +149,6 @@ static void stats_signal (int signum) { udp_echo_main_t *um = &udp_echo_main; - um->time_to_print_stats = 1; } @@ -593,15 +594,20 @@ send_test_chunk (udp_echo_main_t * utm, app_session_t * s, u32 bytes) u8 *test_data = utm->connect_test_data; int test_buf_offset = 0; u64 bytes_sent = 0; - u32 bytes_to_snd; + u32 bytes_to_snd, enq_space, min_chunk; int rv; + min_chunk = clib_min (65536, s->tx_fifo->nitems); bytes_to_snd = (bytes == 0) ? vec_len (test_data) : bytes; if (bytes_to_snd > vec_len (test_data)) bytes_to_snd = vec_len (test_data); while (bytes_to_snd > 0 && !utm->time_to_stop) { + enq_space = svm_fifo_max_enqueue (s->tx_fifo); + if (enq_space < clib_min (bytes_to_snd, min_chunk)) + continue; + rv = app_send (s, test_data + test_buf_offset, bytes_to_snd, 0); if (rv > 0) { @@ -621,12 +627,14 @@ recv_test_chunk (udp_echo_main_t * utm, app_session_t * s) void client_send_data (udp_echo_main_t * utm) { - u8 *test_data; + f64 start_time, end_time, delta; app_session_t *session; + char *transfer_type; u32 n_iterations; + u8 *test_data; int i; - vec_validate (utm->connect_test_data, 64 * 1024 - 1); + vec_validate (utm->connect_test_data, 1024 * 1024 - 1); for (i = 0; i < vec_len (utm->connect_test_data); i++) utm->connect_test_data[i] = i & 0xff; @@ -635,22 +643,38 @@ client_send_data (udp_echo_main_t * utm) ASSERT (vec_len (test_data) > 0); vec_validate (utm->rx_buf, vec_len (test_data) - 1); - n_iterations = NITER; + n_iterations = utm->bytes_to_send / vec_len (test_data); + if (!n_iterations) + n_iterations = 1; + start_time = clib_time_now (&utm->clib_time); for (i = 0; i < n_iterations; i++) { send_test_chunk (utm, session, 0); - recv_test_chunk (utm, session); + if (utm->have_return) + recv_test_chunk (utm, session); if (utm->time_to_stop) break; } - f64 timeout = clib_time_now (&utm->clib_time) + 5; - while (clib_time_now (&utm->clib_time) < timeout) + if (utm->have_return) { - recv_test_chunk (utm, session); + f64 timeout = clib_time_now (&utm->clib_time) + 5; + while (clib_time_now (&utm->clib_time) < timeout) + recv_test_chunk (utm, session); } + end_time = clib_time_now (&utm->clib_time); + delta = end_time - start_time; + transfer_type = utm->have_return ? "full-duplex" : "half-duplex"; + clib_warning ("%lld bytes (%lld mbytes, %lld gbytes) in %.2f seconds", + utm->bytes_to_send, utm->bytes_to_send / (1ULL << 20), + utm->bytes_to_send / (1ULL << 30), delta); + clib_warning ("%.2f bytes/second %s", ((f64) utm->bytes_to_send) / (delta), + transfer_type); + clib_warning ("%.4f gbit/second %s", + (((f64) utm->bytes_to_send * 8.0) / delta / 1e9), + transfer_type); } static void @@ -1205,7 +1229,7 @@ int main (int argc, char **argv) { udp_echo_main_t *utm = &udp_echo_main; - u8 *uri = (u8 *) "udp://0.0.0.0/1234"; + u8 *uri = (u8 *) "udp://6.0.1.1/1234"; unformat_input_t _argv, *a = &_argv; int i_am_server = 1; app_session_t *session; @@ -1217,27 +1241,27 @@ main (int argc, char **argv) int i; clib_mem_init (0, 256 << 20); - heap = clib_mem_get_per_cpu_heap (); h = mheap_header (heap); - /* make the main heap thread-safe */ h->flags |= MHEAP_FLAG_THREAD_SAFE; + svm_fifo_segment_main_init (0x200000000ULL, 20); vec_validate (utm->rx_buf, 8192); - utm->session_index_by_vpp_handles = hash_create (0, sizeof (uword)); utm->my_pid = getpid (); utm->configured_segment_size = 1 << 20; utm->segments_table = hash_create_vec (0, sizeof (u8), sizeof (u64)); - + utm->have_return = 1; + utm->bytes_to_send = 1024; + utm->fifo_size = 128 << 10; + utm->segment_main = &svm_fifo_segment_main; + utm->cut_through_session_index = ~0; clib_time_init (&utm->clib_time); + init_error_string_table (utm); - svm_fifo_segment_main_init (0x200000000ULL, 20); unformat_init_command_line (a, argv); - utm->fifo_size = 128 << 10; - while (unformat_check_input (a) != UNFORMAT_END_OF_INPUT) { if (unformat (a, "chroot prefix %s", &chroot_prefix)) @@ -1254,6 +1278,12 @@ main (int argc, char **argv) i_am_server = 1; else if (unformat (a, "client")) i_am_server = 0; + else if (unformat (a, "no-return")) + utm->have_return = 0; + else if (unformat (a, "mbytes %d", &tmp)) + utm->bytes_to_send = (u64) tmp << 20; + else if (unformat (a, "fifo-size %d", &tmp)) + utm->fifo_size = tmp << 10; else { fformat (stderr, "%s: usage [server|client]\n"); @@ -1261,9 +1291,7 @@ main (int argc, char **argv) } } - utm->cut_through_session_index = ~0; utm->i_am_server = i_am_server; - utm->segment_main = &svm_fifo_segment_main; setup_signal_handlers (); tcp_echo_api_hookup (utm); @@ -1290,7 +1318,7 @@ main (int argc, char **argv) if (i_am_server == 0) { client_test (utm); - exit (0); + goto done; } /* $$$$ hack preallocation */ @@ -1304,6 +1332,7 @@ main (int argc, char **argv) udp_server_test (utm); +done: vl_client_disconnect_from_vlib (); exit (0); } diff --git a/src/vnet/session-apps/echo_client.c b/src/vnet/session-apps/echo_client.c index d47b0d97337..3c1904cb55e 100644 --- a/src/vnet/session-apps/echo_client.c +++ b/src/vnet/session-apps/echo_client.c @@ -56,11 +56,60 @@ send_data_chunk (echo_client_main_t * ecm, session_t * s) s->bytes_to_send); if (!ecm->is_dgram) - rv = app_send_stream (&s->data, test_data + test_buf_offset, - bytes_this_chunk, 0); + { + if (ecm->no_copy) + { + svm_fifo_t *f = s->data.tx_fifo; + rv = clib_min (svm_fifo_max_enqueue (f), bytes_this_chunk); + svm_fifo_enqueue_nocopy (f, rv); + if (svm_fifo_set_event (f)) + { + session_fifo_event_t evt; + evt.fifo = f; + evt.event_type = FIFO_EVENT_APP_TX; + svm_queue_add (s->data.vpp_evt_q, (u8 *) & evt, 0); + } + } + else + rv = app_send_stream (&s->data, test_data + test_buf_offset, + bytes_this_chunk, 0); + } else - rv = app_send_dgram (&s->data, test_data + test_buf_offset, - bytes_this_chunk, 0); + { + if (ecm->no_copy) + { + session_dgram_hdr_t hdr; + svm_fifo_t *f = s->data.tx_fifo; + app_session_transport_t *at = &s->data.transport; + u32 max_enqueue = svm_fifo_max_enqueue (f); + + if (max_enqueue <= sizeof (session_dgram_hdr_t)) + return; + + max_enqueue -= sizeof (session_dgram_hdr_t); + rv = clib_min (max_enqueue, bytes_this_chunk); + + hdr.data_length = rv; + hdr.data_offset = 0; + clib_memcpy (&hdr.rmt_ip, &at->rmt_ip, sizeof (ip46_address_t)); + hdr.is_ip4 = at->is_ip4; + hdr.rmt_port = at->rmt_port; + clib_memcpy (&hdr.lcl_ip, &at->lcl_ip, sizeof (ip46_address_t)); + hdr.lcl_port = at->lcl_port; + svm_fifo_enqueue_nowait (f, sizeof (hdr), (u8 *) & hdr); + svm_fifo_enqueue_nocopy (f, rv); + if (svm_fifo_set_event (f)) + { + session_fifo_event_t evt; + evt.fifo = f; + evt.event_type = FIFO_EVENT_APP_TX; + svm_queue_add (s->data.vpp_evt_q, (u8 *) & evt, 0); + } + } + else + rv = app_send_dgram (&s->data, test_data + test_buf_offset, + bytes_this_chunk, 0); + } /* If we managed to enqueue data... */ if (rv > 0) @@ -591,6 +640,7 @@ echo_clients_command_fn (vlib_main_t * vm, ecm->test_failed = 0; ecm->vlib_main = vm; ecm->tls_engine = TLS_ENGINE_OPENSSL; + ecm->no_copy = 0; if (thread_main->n_vlib_mains > 1) clib_spinlock_init (&ecm->sessions_lock); diff --git a/src/vnet/session-apps/echo_client.h b/src/vnet/session-apps/echo_client.h index 344e43865b5..8c07b10fbba 100644 --- a/src/vnet/session-apps/echo_client.h +++ b/src/vnet/session-apps/echo_client.h @@ -64,6 +64,8 @@ typedef struct u32 private_segment_size; /**< size of private fifo segs */ u32 tls_engine; /**< TLS engine mbedtls/openssl */ u8 is_dgram; + u32 no_copy; /**< Don't memcpy data to tx fifo */ + /* * Test state variables */ diff --git a/src/vnet/session/application_interface.h b/src/vnet/session/application_interface.h index 5fd218533dc..49c2862ac45 100644 --- a/src/vnet/session/application_interface.h +++ b/src/vnet/session/application_interface.h @@ -204,7 +204,7 @@ app_send_dgram_raw (svm_fifo_t * f, app_session_transport_t * at, int rv; max_enqueue = svm_fifo_max_enqueue (f); - if (svm_fifo_max_enqueue (f) <= sizeof (session_dgram_hdr_t)) + if (max_enqueue <= sizeof (session_dgram_hdr_t)) return 0; max_enqueue -= sizeof (session_dgram_hdr_t); @@ -217,9 +217,6 @@ app_send_dgram_raw (svm_fifo_t * f, app_session_transport_t * at, clib_memcpy (&hdr.lcl_ip, &at->lcl_ip, sizeof (ip46_address_t)); hdr.lcl_port = at->lcl_port; rv = svm_fifo_enqueue_nowait (f, sizeof (hdr), (u8 *) & hdr); - if (rv <= 0) - return 0; - ASSERT (rv == sizeof (hdr)); if ((rv = svm_fifo_enqueue_nowait (f, actual_write, data)) > 0) @@ -231,6 +228,7 @@ app_send_dgram_raw (svm_fifo_t * f, app_session_transport_t * at, svm_queue_add (vpp_evt_q, (u8 *) & evt, noblock); } } + ASSERT (rv); return rv; } diff --git a/src/vnet/session/session.c b/src/vnet/session/session.c index 568065ebc0d..a8a9c66ac81 100644 --- a/src/vnet/session/session.c +++ b/src/vnet/session/session.c @@ -1272,6 +1272,7 @@ session_manager_main_enable (vlib_main_t * vm) vec_validate (smm->free_event_vector, num_threads - 1); vec_validate (smm->vpp_event_queues, num_threads - 1); vec_validate (smm->peekers_rw_locks, num_threads - 1); + vec_validate_aligned (smm->ctx, num_threads - 1, CLIB_CACHE_LINE_BYTES); for (i = 0; i < TRANSPORT_N_PROTO; i++) { diff --git a/src/vnet/session/session.h b/src/vnet/session/session.h index b54f46189f3..c5779b4ffd1 100644 --- a/src/vnet/session/session.h +++ b/src/vnet/session/session.h @@ -103,14 +103,58 @@ typedef CLIB_PACKED (struct { }) session_fifo_event_t; /* *INDENT-ON* */ +typedef struct session_dgram_pre_hdr_ +{ + u32 data_length; + u32 data_offset; +} session_dgram_pre_hdr_t; + +/* *INDENT-OFF* */ +typedef CLIB_PACKED (struct session_dgram_header_ +{ + u32 data_length; + u32 data_offset; + ip46_address_t rmt_ip; + ip46_address_t lcl_ip; + u16 rmt_port; + u16 lcl_port; + u8 is_ip4; +}) session_dgram_hdr_t; +/* *INDENT-ON* */ + +#define SESSION_CONN_ID_LEN 37 +#define SESSION_CONN_HDR_LEN 45 + +STATIC_ASSERT (sizeof (session_dgram_hdr_t) == (SESSION_CONN_ID_LEN + 8), + "session conn id wrong length"); + +typedef struct session_tx_context_ +{ + CLIB_CACHE_LINE_ALIGN_MARK (cacheline0); + stream_session_t *s; + transport_proto_vft_t *transport_vft; + transport_connection_t *tc; + vlib_buffer_t *b; + u32 max_dequeue; + u32 snd_space; + u32 left_to_snd; + u32 tx_offset; + u32 max_len_to_snd; + u16 deq_per_first_buf; + u16 deq_per_buf; + u16 snd_mss; + u8 n_bufs_per_seg; + CLIB_CACHE_LINE_ALIGN_MARK (cacheline1); + session_dgram_hdr_t hdr; +} session_tx_context_t; + /* Forward definition */ typedef struct _session_manager_main session_manager_main_t; typedef int (session_fifo_rx_fn) (vlib_main_t * vm, vlib_node_runtime_t * node, - session_manager_main_t * smm, session_fifo_event_t * e0, stream_session_t * s0, - u32 thread_index, int *n_tx_pkts); + int *n_tx_pkts); extern session_fifo_rx_fn session_tx_fifo_peek_and_snd; extern session_fifo_rx_fn session_tx_fifo_dequeue_and_snd; @@ -144,6 +188,9 @@ struct _session_manager_main /** per-worker postponed disconnects */ session_fifo_event_t **pending_disconnects; + /** per-worker session context */ + session_tx_context_t *ctx; + /** vpp fifo event queue */ svm_queue_t **vpp_event_queues; @@ -204,31 +251,6 @@ struct _session_manager_main }; -typedef struct session_dgram_pre_hdr_ -{ - u32 data_length; - u32 data_offset; -} session_dgram_pre_hdr_t; - -/* *INDENT-OFF* */ -typedef CLIB_PACKED (struct session_dgram_header_ -{ - u32 data_length; - u32 data_offset; - ip46_address_t rmt_ip; - ip46_address_t lcl_ip; - u16 rmt_port; - u16 lcl_port; - u8 is_ip4; -}) session_dgram_hdr_t; -/* *INDENT-ON* */ - -#define SESSION_CONN_ID_LEN 37 -#define SESSION_CONN_HDR_LEN 45 - -STATIC_ASSERT (sizeof (session_dgram_hdr_t) == (SESSION_CONN_ID_LEN + 8), - "session conn id wrong length"); - extern session_manager_main_t session_manager_main; extern vlib_node_registration_t session_queue_node; diff --git a/src/vnet/session/session_node.c b/src/vnet/session/session_node.c index 14716965547..269e2fb591e 100644 --- a/src/vnet/session/session_node.c +++ b/src/vnet/session/session_node.c @@ -18,6 +18,7 @@ #include <vnet/vnet.h> #include <vppinfra/elog.h> #include <vnet/session/transport.h> +#include <vnet/session/session.h> #include <vnet/session/application.h> #include <vnet/session/session_debug.h> #include <svm/queue.h> @@ -65,420 +66,473 @@ static char *session_queue_error_strings[] = { }; always_inline void -session_tx_fifo_chain_tail (session_manager_main_t * smm, vlib_main_t * vm, - u8 thread_index, svm_fifo_t * fifo, - vlib_buffer_t * b0, u32 bi0, u8 n_bufs_per_seg, - u32 left_from_seg, u32 * left_to_snd0, - u16 * n_bufs, u32 * tx_offset, u16 deq_per_buf, - u8 peek_data, transport_tx_fn_type_t tx_type) +session_tx_trace_buffer (vlib_main_t * vm, vlib_node_runtime_t * node, + u32 next_index, vlib_buffer_t * b, + stream_session_t * s, u32 * n_trace) { - vlib_buffer_t *chain_b0, *prev_b0; - u32 chain_bi0, to_deq; - u16 len_to_deq0, n_bytes_read; - u8 *data0, j; - - b0->flags |= VLIB_BUFFER_TOTAL_LENGTH_VALID; - b0->total_length_not_including_first_buffer = 0; + session_queue_trace_t *t; + vlib_trace_buffer (vm, node, next_index, b, 1 /* follow_chain */ ); + vlib_set_trace_count (vm, node, --*n_trace); + t = vlib_add_trace (vm, node, b, sizeof (*t)); + t->session_index = s->session_index; + t->server_thread_index = s->thread_index; +} - chain_bi0 = bi0; - chain_b0 = b0; +always_inline void +session_tx_fifo_chain_tail (vlib_main_t * vm, session_tx_context_t * ctx, + vlib_buffer_t * b, u16 * n_bufs, u8 peek_data) +{ + session_manager_main_t *smm = &session_manager_main; + vlib_buffer_t *chain_b, *prev_b; + u32 chain_bi0, to_deq, left_from_seg; + u16 len_to_deq, n_bytes_read; + u8 *data, j; + + b->flags |= VLIB_BUFFER_TOTAL_LENGTH_VALID; + b->total_length_not_including_first_buffer = 0; + + chain_b = b; + left_from_seg = clib_min (ctx->snd_mss - b->current_length, + ctx->left_to_snd); to_deq = left_from_seg; - for (j = 1; j < n_bufs_per_seg; j++) + for (j = 1; j < ctx->n_bufs_per_seg; j++) { - prev_b0 = chain_b0; - len_to_deq0 = clib_min (to_deq, deq_per_buf); + prev_b = chain_b; + len_to_deq = clib_min (to_deq, ctx->deq_per_buf); *n_bufs -= 1; - chain_bi0 = smm->tx_buffers[thread_index][*n_bufs]; - _vec_len (smm->tx_buffers[thread_index]) = *n_bufs; + chain_bi0 = smm->tx_buffers[ctx->s->thread_index][*n_bufs]; + _vec_len (smm->tx_buffers[ctx->s->thread_index]) = *n_bufs; - chain_b0 = vlib_get_buffer (vm, chain_bi0); - chain_b0->current_data = 0; - data0 = vlib_buffer_get_current (chain_b0); + chain_b = vlib_get_buffer (vm, chain_bi0); + chain_b->current_data = 0; + data = vlib_buffer_get_current (chain_b); if (peek_data) { - n_bytes_read = svm_fifo_peek (fifo, *tx_offset, len_to_deq0, data0); - *tx_offset += n_bytes_read; + n_bytes_read = svm_fifo_peek (ctx->s->server_tx_fifo, + ctx->tx_offset, len_to_deq, data); + ctx->tx_offset += n_bytes_read; } else { - if (tx_type == TRANSPORT_TX_DGRAM) + if (ctx->transport_vft->tx_type == TRANSPORT_TX_DGRAM) { - session_dgram_hdr_t *hdr; + svm_fifo_t *f = ctx->s->server_tx_fifo; + session_dgram_hdr_t *hdr = &ctx->hdr; u16 deq_now; - hdr = (session_dgram_hdr_t *) svm_fifo_head (fifo); deq_now = clib_min (hdr->data_length - hdr->data_offset, - len_to_deq0); - n_bytes_read = svm_fifo_peek (fifo, hdr->data_offset, deq_now, - data0); + len_to_deq); + n_bytes_read = svm_fifo_peek (f, hdr->data_offset, deq_now, + data); ASSERT (n_bytes_read > 0); hdr->data_offset += n_bytes_read; if (hdr->data_offset == hdr->data_length) - svm_fifo_dequeue_drop (fifo, hdr->data_length); + svm_fifo_dequeue_drop (f, hdr->data_length); } else - n_bytes_read = svm_fifo_dequeue_nowait (fifo, len_to_deq0, data0); + n_bytes_read = svm_fifo_dequeue_nowait (ctx->s->server_tx_fifo, + len_to_deq, data); } - ASSERT (n_bytes_read == len_to_deq0); - chain_b0->current_length = n_bytes_read; - b0->total_length_not_including_first_buffer += chain_b0->current_length; + ASSERT (n_bytes_read == len_to_deq); + chain_b->current_length = n_bytes_read; + b->total_length_not_including_first_buffer += chain_b->current_length; /* update previous buffer */ - prev_b0->next_buffer = chain_bi0; - prev_b0->flags |= VLIB_BUFFER_NEXT_PRESENT; + prev_b->next_buffer = chain_bi0; + prev_b->flags |= VLIB_BUFFER_NEXT_PRESENT; /* update current buffer */ - chain_b0->next_buffer = 0; + chain_b->next_buffer = 0; to_deq -= n_bytes_read; if (to_deq == 0) break; } ASSERT (to_deq == 0 - && b0->total_length_not_including_first_buffer == left_from_seg); - *left_to_snd0 -= left_from_seg; + && b->total_length_not_including_first_buffer == left_from_seg); + ctx->left_to_snd -= left_from_seg; } always_inline int -session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node, +session_output_try_get_buffers (vlib_main_t * vm, session_manager_main_t * smm, - session_fifo_event_t * e0, - stream_session_t * s0, u32 thread_index, - int *n_tx_packets, u8 peek_data) + u32 thread_index, u16 * n_bufs, u32 wanted) { - u32 n_trace = vlib_get_trace_count (vm, node); - u32 left_to_snd0, max_len_to_snd0, len_to_deq0, snd_space0; - u32 n_bufs_per_evt, n_frames_per_evt, n_bufs_per_frame; - transport_connection_t *tc0; - transport_proto_vft_t *transport_vft; - transport_proto_t tp; - u32 next_index, next0, *to_next, n_left_to_next, bi0; - vlib_buffer_t *b0; - u32 tx_offset = 0, max_dequeue0, n_bytes_per_seg, left_for_seg; - u16 snd_mss0, n_bufs_per_seg, n_bufs; + u32 bufs_alloc = 0, bufs_now; + vec_validate_aligned (smm->tx_buffers[thread_index], *n_bufs + wanted - 1, + CLIB_CACHE_LINE_BYTES); + do + { + bufs_now = + vlib_buffer_alloc (vm, + &smm->tx_buffers[thread_index][*n_bufs + + bufs_alloc], + wanted - bufs_alloc); + bufs_alloc += bufs_now; + } + while (bufs_now > 0 && ((bufs_alloc + *n_bufs < wanted))); + + *n_bufs += bufs_alloc; + _vec_len (smm->tx_buffers[thread_index]) = *n_bufs; + return bufs_alloc; +} + +always_inline void +session_tx_fill_buffer (vlib_main_t * vm, session_tx_context_t * ctx, + vlib_buffer_t * b, u16 * n_bufs, u8 peek_data) +{ + u32 len_to_deq; u8 *data0; - int i, n_bytes_read; - u32 n_bytes_per_buf, deq_per_buf, deq_per_first_buf; - u32 bufs_alloc, bufs_now; - session_dgram_hdr_t hdr; - - next_index = next0 = smm->session_type_to_next[s0->session_type]; - tp = session_get_transport_proto (s0); - transport_vft = transport_protocol_get_vft (tp); + int n_bytes_read; + + /* + * Start with the first buffer in chain + */ + b->error = 0; + b->flags = VNET_BUFFER_F_LOCALLY_ORIGINATED; + b->current_data = 0; + b->total_length_not_including_first_buffer = 0; + + data0 = vlib_buffer_make_headroom (b, MAX_HDRS_LEN); + len_to_deq = clib_min (ctx->left_to_snd, ctx->deq_per_first_buf); + if (peek_data) { - if (PREDICT_FALSE (s0->session_state < SESSION_STATE_READY)) - { - /* Can retransmit for closed sessions but can't send new data if - * session is not ready or closed */ - vec_add1 (smm->pending_event_vector[thread_index], *e0); - return 0; - } - tc0 = - transport_vft->get_connection (s0->connection_index, thread_index); + n_bytes_read = svm_fifo_peek (ctx->s->server_tx_fifo, ctx->tx_offset, + len_to_deq, data0); + ASSERT (n_bytes_read > 0); + /* Keep track of progress locally, transport is also supposed to + * increment it independently when pushing the header */ + ctx->tx_offset += n_bytes_read; } else { - if (s0->session_state == SESSION_STATE_LISTENING) - tc0 = transport_vft->get_listener (s0->connection_index); + if (ctx->transport_vft->tx_type == TRANSPORT_TX_DGRAM) + { + session_dgram_hdr_t *hdr = &ctx->hdr; + svm_fifo_t *f = ctx->s->server_tx_fifo; + u16 deq_now; + u32 offset; + + ASSERT (hdr->data_length > hdr->data_offset); + deq_now = clib_min (hdr->data_length - hdr->data_offset, + len_to_deq); + offset = hdr->data_offset + SESSION_CONN_HDR_LEN; + n_bytes_read = svm_fifo_peek (f, offset, deq_now, data0); + ASSERT (n_bytes_read > 0); + + if (ctx->s->session_state == SESSION_STATE_LISTENING) + { + ip_copy (&ctx->tc->rmt_ip, &hdr->rmt_ip, ctx->tc->is_ip4); + ctx->tc->rmt_port = hdr->rmt_port; + } + hdr->data_offset += n_bytes_read; + if (hdr->data_offset == hdr->data_length) + { + offset = hdr->data_length + SESSION_CONN_HDR_LEN; + svm_fifo_dequeue_drop (f, offset); + } + } else { - if (PREDICT_FALSE (s0->session_state == SESSION_STATE_CLOSED)) - return 0; - tc0 = transport_vft->get_connection (s0->connection_index, - thread_index); + n_bytes_read = svm_fifo_dequeue_nowait (ctx->s->server_tx_fifo, + len_to_deq, data0); + ASSERT (n_bytes_read > 0); } } + b->current_length = n_bytes_read; + ctx->left_to_snd -= n_bytes_read; - /* Make sure we have space to send and there's something to dequeue */ - snd_mss0 = transport_vft->send_mss (tc0); - snd_space0 = transport_vft->send_space (tc0); + /* + * Fill in the remaining buffers in the chain, if any + */ + if (PREDICT_FALSE (ctx->n_bufs_per_seg > 1 && ctx->left_to_snd)) + session_tx_fifo_chain_tail (vm, ctx, b, n_bufs, peek_data); + + /* *INDENT-OFF* */ + SESSION_EVT_DBG(SESSION_EVT_DEQ, s, ({ + ed->data[0] = e->event_type; + ed->data[1] = max_dequeue; + ed->data[2] = len_to_deq; + ed->data[3] = left_to_snd; + })); + /* *INDENT-ON* */ +} - /* Can't make any progress */ - if (snd_space0 == 0 || snd_mss0 == 0) +always_inline u8 +session_tx_not_ready (stream_session_t * s, u8 peek_data) +{ + if (peek_data) { - vec_add1 (smm->pending_event_vector[thread_index], *e0); - return 0; + /* Can retransmit for closed sessions but can't send new data if + * session is not ready or closed */ + if (s->session_state < SESSION_STATE_READY) + return 1; } + return 0; +} - /* Allow enqueuing of a new event */ - svm_fifo_unset_event (s0->server_tx_fifo); +always_inline transport_connection_t * +session_tx_get_transport (session_tx_context_t * ctx, u8 peek_data) +{ + if (peek_data) + { + return ctx->transport_vft->get_connection (ctx->s->connection_index, + ctx->s->thread_index); + } + else + { + if (ctx->s->session_state == SESSION_STATE_LISTENING) + return ctx->transport_vft->get_listener (ctx->s->connection_index); + else + { + return ctx->transport_vft->get_connection (ctx->s->connection_index, + ctx->s->thread_index); + } + } +} - /* Check how much we can pull. */ - max_dequeue0 = svm_fifo_max_dequeue (s0->server_tx_fifo); +always_inline void +session_tx_set_dequeue_params (vlib_main_t * vm, session_tx_context_t * ctx, + u8 peek_data) +{ + u32 n_bytes_per_buf, n_bytes_per_seg; + ctx->max_dequeue = svm_fifo_max_dequeue (ctx->s->server_tx_fifo); if (peek_data) { /* Offset in rx fifo from where to peek data */ - tx_offset = transport_vft->tx_fifo_offset (tc0); - if (PREDICT_FALSE (tx_offset >= max_dequeue0)) - return 0; - max_dequeue0 -= tx_offset; + ctx->tx_offset = ctx->transport_vft->tx_fifo_offset (ctx->tc); + if (PREDICT_FALSE (ctx->tx_offset >= ctx->max_dequeue)) + { + ctx->max_len_to_snd = 0; + return; + } + ctx->max_dequeue -= ctx->tx_offset; } else { - if (transport_vft->tx_type == TRANSPORT_TX_DGRAM) + if (ctx->transport_vft->tx_type == TRANSPORT_TX_DGRAM) { - if (max_dequeue0 < sizeof (hdr)) - return 0; - svm_fifo_peek (s0->server_tx_fifo, 0, sizeof (hdr), (u8 *) & hdr); - ASSERT (hdr.data_length > hdr.data_offset); - max_dequeue0 = hdr.data_length - hdr.data_offset; + if (ctx->max_dequeue <= sizeof (ctx->hdr)) + { + ctx->max_len_to_snd = 0; + return; + } + svm_fifo_peek (ctx->s->server_tx_fifo, 0, sizeof (ctx->hdr), + (u8 *) & ctx->hdr); + ASSERT (ctx->hdr.data_length > ctx->hdr.data_offset); + ctx->max_dequeue = ctx->hdr.data_length - ctx->hdr.data_offset; } } - ASSERT (max_dequeue0 > 0); + ASSERT (ctx->max_dequeue > 0); /* Ensure we're not writing more than transport window allows */ - if (max_dequeue0 < snd_space0) + if (ctx->max_dequeue < ctx->snd_space) { /* Constrained by tx queue. Try to send only fully formed segments */ - max_len_to_snd0 = (max_dequeue0 > snd_mss0) ? - max_dequeue0 - max_dequeue0 % snd_mss0 : max_dequeue0; + ctx->max_len_to_snd = + (ctx->max_dequeue > ctx->snd_mss) ? + ctx->max_dequeue - ctx->max_dequeue % ctx->snd_mss : ctx->max_dequeue; /* TODO Nagle ? */ } else { /* Expectation is that snd_space0 is already a multiple of snd_mss */ - max_len_to_snd0 = snd_space0; + ctx->max_len_to_snd = ctx->snd_space; } - n_bytes_per_buf = vlib_buffer_free_list_buffer_size - (vm, VLIB_BUFFER_DEFAULT_FREE_LIST_INDEX); + n_bytes_per_buf = vlib_buffer_free_list_buffer_size (vm, + VLIB_BUFFER_DEFAULT_FREE_LIST_INDEX); ASSERT (n_bytes_per_buf > MAX_HDRS_LEN); - n_bytes_per_seg = MAX_HDRS_LEN + snd_mss0; - n_bufs_per_seg = ceil ((double) n_bytes_per_seg / n_bytes_per_buf); - n_bufs_per_evt = ceil ((double) max_len_to_snd0 / n_bytes_per_seg); - n_frames_per_evt = ceil ((double) n_bufs_per_evt / VLIB_FRAME_SIZE); - n_bufs_per_frame = n_bufs_per_seg * VLIB_FRAME_SIZE; + n_bytes_per_seg = MAX_HDRS_LEN + ctx->snd_mss; + ctx->n_bufs_per_seg = ceil ((double) n_bytes_per_seg / n_bytes_per_buf); + ctx->deq_per_buf = clib_min (ctx->snd_mss, n_bytes_per_buf); + ctx->deq_per_first_buf = clib_min (ctx->snd_mss, + n_bytes_per_buf - MAX_HDRS_LEN); +} + +always_inline int +session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node, + session_fifo_event_t * e, + stream_session_t * s, int *n_tx_packets, + u8 peek_data) +{ + u32 next_index, next0, next1, next2, next3, *to_next, n_left_to_next; + u32 n_trace = vlib_get_trace_count (vm, node), n_packets = 0, pbi; + u32 n_bufs_per_frame, thread_index = s->thread_index; + session_manager_main_t *smm = &session_manager_main; + session_tx_context_t *ctx = &smm->ctx[thread_index]; + transport_proto_t tp; + vlib_buffer_t *pb; + u16 n_bufs; - deq_per_buf = clib_min (snd_mss0, n_bytes_per_buf); - deq_per_first_buf = clib_min (snd_mss0, n_bytes_per_buf - MAX_HDRS_LEN); + if (PREDICT_FALSE (session_tx_not_ready (s, peek_data))) + { + vec_add1 (smm->pending_event_vector[thread_index], *e); + return 0; + } + + next_index = smm->session_type_to_next[s->session_type]; + next0 = next1 = next2 = next3 = next_index; + + tp = session_get_transport_proto (s); + ctx->s = s; + ctx->transport_vft = transport_protocol_get_vft (tp); + ctx->tc = session_tx_get_transport (ctx, peek_data); + ctx->snd_mss = ctx->transport_vft->send_mss (ctx->tc); + ctx->snd_space = ctx->transport_vft->send_space (ctx->tc); + if (ctx->snd_space == 0 || ctx->snd_mss == 0) + { + vec_add1 (smm->pending_event_vector[thread_index], *e); + return 0; + } + + /* Allow enqueuing of a new event */ + svm_fifo_unset_event (s->server_tx_fifo); + + /* Check how much we can pull. */ + session_tx_set_dequeue_params (vm, ctx, peek_data); + if (PREDICT_FALSE (!ctx->max_len_to_snd)) + return 0; n_bufs = vec_len (smm->tx_buffers[thread_index]); - left_to_snd0 = max_len_to_snd0; - for (i = 0; i < n_frames_per_evt; i++) + ctx->left_to_snd = ctx->max_len_to_snd; + + /* + * Make sure we have at least one full frame of buffers ready + */ + n_bufs_per_frame = ctx->n_bufs_per_seg * VLIB_FRAME_SIZE; + if (n_bufs < n_bufs_per_frame) { - /* Make sure we have at least one full frame of buffers ready */ + session_output_try_get_buffers (vm, smm, thread_index, &n_bufs, + n_bufs_per_frame); if (PREDICT_FALSE (n_bufs < n_bufs_per_frame)) { - vec_validate (smm->tx_buffers[thread_index], - n_bufs + n_bufs_per_frame - 1); - bufs_alloc = 0; - do - { - bufs_now = - vlib_buffer_alloc (vm, - &smm->tx_buffers[thread_index][n_bufs + - bufs_alloc], - n_bufs_per_frame - bufs_alloc); - bufs_alloc += bufs_now; - } - while (bufs_now > 0 && ((bufs_alloc + n_bufs < n_bufs_per_frame))); - - n_bufs += bufs_alloc; - _vec_len (smm->tx_buffers[thread_index]) = n_bufs; - - if (PREDICT_FALSE (n_bufs < n_bufs_per_frame)) - { - vec_add1 (smm->pending_event_vector[thread_index], *e0); - return -1; - } - ASSERT (n_bufs >= n_bufs_per_frame); + vec_add1 (smm->pending_event_vector[thread_index], *e); + return -1; } + } - vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next); - while (left_to_snd0 && n_left_to_next) + /* + * Write until we fill up a frame + */ + vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next); + while (ctx->left_to_snd && n_left_to_next) + { + while (ctx->left_to_snd > 3 * ctx->snd_mss && n_left_to_next >= 4) { - /* - * Handle first buffer in chain separately - */ + vlib_buffer_t *b0, *b1; + u32 bi0, bi1; - len_to_deq0 = clib_min (left_to_snd0, deq_per_first_buf); - if (left_to_snd0 > len_to_deq0 && n_left_to_next > 1) - { - vlib_buffer_t *pb; - u32 pbi = smm->tx_buffers[thread_index][n_bufs - 2]; - pb = vlib_get_buffer (vm, pbi); - vlib_prefetch_buffer_header (pb, LOAD); - } + pbi = smm->tx_buffers[thread_index][n_bufs - 3]; + pb = vlib_get_buffer (vm, pbi); + vlib_prefetch_buffer_header (pb, STORE); + pbi = smm->tx_buffers[thread_index][n_bufs - 4]; + pb = vlib_get_buffer (vm, pbi); + vlib_prefetch_buffer_header (pb, STORE); - /* Get free buffer */ - ASSERT (n_bufs >= 1); - bi0 = smm->tx_buffers[thread_index][--n_bufs]; - _vec_len (smm->tx_buffers[thread_index]) = n_bufs; - - /* usual speculation, or the enqueue_x1 macro will barf */ - to_next[0] = bi0; - to_next += 1; - n_left_to_next -= 1; + to_next[0] = bi0 = smm->tx_buffers[thread_index][--n_bufs]; + to_next[1] = bi1 = smm->tx_buffers[thread_index][--n_bufs]; b0 = vlib_get_buffer (vm, bi0); - b0->error = 0; - b0->flags = VNET_BUFFER_F_LOCALLY_ORIGINATED; - b0->current_data = 0; - b0->total_length_not_including_first_buffer = 0; + b1 = vlib_get_buffer (vm, bi1); - data0 = vlib_buffer_make_headroom (b0, MAX_HDRS_LEN); - if (peek_data) - { - n_bytes_read = svm_fifo_peek (s0->server_tx_fifo, tx_offset, - len_to_deq0, data0); - if (n_bytes_read <= 0) - goto dequeue_fail; - /* Keep track of progress locally, transport is also supposed to - * increment it independently when pushing the header */ - tx_offset += n_bytes_read; - } - else - { - if (transport_vft->tx_type == TRANSPORT_TX_DGRAM) - { - svm_fifo_t *f = s0->server_tx_fifo; - u16 deq_now; - u32 offset; - - ASSERT (hdr.data_length > hdr.data_offset); - deq_now = clib_min (hdr.data_length - hdr.data_offset, - len_to_deq0); - offset = hdr.data_offset + SESSION_CONN_HDR_LEN; - n_bytes_read = svm_fifo_peek (f, offset, deq_now, data0); - if (PREDICT_FALSE (n_bytes_read <= 0)) - goto dequeue_fail; - - if (s0->session_state == SESSION_STATE_LISTENING) - { - ip_copy (&tc0->rmt_ip, &hdr.rmt_ip, tc0->is_ip4); - tc0->rmt_port = hdr.rmt_port; - } - hdr.data_offset += n_bytes_read; - if (hdr.data_offset == hdr.data_length) - { - offset = hdr.data_length + SESSION_CONN_HDR_LEN; - svm_fifo_dequeue_drop (f, offset); - } - } - else - { - n_bytes_read = svm_fifo_dequeue_nowait (s0->server_tx_fifo, - len_to_deq0, data0); - if (n_bytes_read <= 0) - goto dequeue_fail; - } - } + session_tx_fill_buffer (vm, ctx, b0, &n_bufs, peek_data); + session_tx_fill_buffer (vm, ctx, b1, &n_bufs, peek_data); - b0->current_length = n_bytes_read; - left_to_snd0 -= n_bytes_read; - *n_tx_packets = *n_tx_packets + 1; + ctx->transport_vft->push_header (ctx->tc, b0); + ctx->transport_vft->push_header (ctx->tc, b1); - /* - * Fill in the remaining buffers in the chain, if any - */ - if (PREDICT_FALSE (n_bufs_per_seg > 1 && left_to_snd0)) + to_next += 2; + n_left_to_next -= 2; + n_packets += 2; + + VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b0); + VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b1); + + if (PREDICT_FALSE (n_trace > 0)) { - left_for_seg = clib_min (snd_mss0 - n_bytes_read, left_to_snd0); - session_tx_fifo_chain_tail (smm, vm, thread_index, - s0->server_tx_fifo, b0, bi0, - n_bufs_per_seg, left_for_seg, - &left_to_snd0, &n_bufs, &tx_offset, - deq_per_buf, peek_data, - transport_vft->tx_type); + session_tx_trace_buffer (vm, node, next_index, b0, s, &n_trace); + if (n_trace) + session_tx_trace_buffer (vm, node, next_index, b1, s, + &n_trace); } + vlib_validate_buffer_enqueue_x2 (vm, node, next_index, to_next, + n_left_to_next, bi0, bi1, + next0, next1); + } + while (ctx->left_to_snd && n_left_to_next) + { + vlib_buffer_t *b0; + u32 bi0; + + ASSERT (n_bufs >= 1); + to_next[0] = bi0 = smm->tx_buffers[thread_index][--n_bufs]; + b0 = vlib_get_buffer (vm, bi0); + session_tx_fill_buffer (vm, ctx, b0, &n_bufs, peek_data); + /* Ask transport to push header after current_length and * total_length_not_including_first_buffer are updated */ - transport_vft->push_header (tc0, b0); + ctx->transport_vft->push_header (ctx->tc, b0); - /* *INDENT-OFF* */ - SESSION_EVT_DBG(SESSION_EVT_DEQ, s0, ({ - ed->data[0] = e0->event_type; - ed->data[1] = max_dequeue0; - ed->data[2] = len_to_deq0; - ed->data[3] = left_to_snd0; - })); - /* *INDENT-ON* */ + to_next += 1; + n_left_to_next -= 1; + n_packets += 1; VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b0); - if (PREDICT_FALSE (n_trace > 0)) - { - session_queue_trace_t *t0; - vlib_trace_buffer (vm, node, next_index, b0, - 1 /* follow_chain */ ); - vlib_set_trace_count (vm, node, --n_trace); - t0 = vlib_add_trace (vm, node, b0, sizeof (*t0)); - t0->session_index = s0->session_index; - t0->server_thread_index = s0->thread_index; - } + session_tx_trace_buffer (vm, node, next_index, b0, s, &n_trace); - vlib_validate_buffer_enqueue_x1 (vm, node, next_index, - to_next, n_left_to_next, - bi0, next0); + vlib_validate_buffer_enqueue_x1 (vm, node, next_index, to_next, + n_left_to_next, bi0, next0); } - vlib_put_next_frame (vm, node, next_index, n_left_to_next); } + _vec_len (smm->tx_buffers[thread_index]) = n_bufs; + *n_tx_packets += n_packets; + vlib_put_next_frame (vm, node, next_index, n_left_to_next); /* If we couldn't dequeue all bytes mark as partially read */ - if (max_len_to_snd0 < max_dequeue0) - if (svm_fifo_set_event (s0->server_tx_fifo)) - vec_add1 (smm->pending_event_vector[thread_index], *e0); + if (ctx->max_len_to_snd < ctx->max_dequeue) + if (svm_fifo_set_event (s->server_tx_fifo)) + vec_add1 (smm->pending_event_vector[thread_index], *e); - if (!peek_data && transport_vft->tx_type == TRANSPORT_TX_DGRAM) + if (!peek_data && ctx->transport_vft->tx_type == TRANSPORT_TX_DGRAM) { /* Fix dgram pre header */ - if (max_len_to_snd0 < max_dequeue0) - svm_fifo_overwrite_head (s0->server_tx_fifo, (u8 *) & hdr, + if (ctx->max_len_to_snd < ctx->max_dequeue) + svm_fifo_overwrite_head (s->server_tx_fifo, (u8 *) & ctx->hdr, sizeof (session_dgram_pre_hdr_t)); /* More data needs to be read */ - else if (svm_fifo_max_dequeue (s0->server_tx_fifo) > 0) - vec_add1 (smm->pending_event_vector[thread_index], *e0); + else if (svm_fifo_max_dequeue (s->server_tx_fifo) > 0) + if (svm_fifo_set_event (s->server_tx_fifo)) + vec_add1 (smm->pending_event_vector[thread_index], *e); } return 0; - -dequeue_fail: - /* - * 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"); - if (svm_fifo_set_event (s0->server_tx_fifo)) - { - vec_add1 (smm->pending_event_vector[thread_index], *e0); - } - vlib_put_next_frame (vm, node, next_index, n_left_to_next + 1); - _vec_len (smm->tx_buffers[thread_index]) += 1; - - return 0; } int session_tx_fifo_peek_and_snd (vlib_main_t * vm, vlib_node_runtime_t * node, - session_manager_main_t * smm, session_fifo_event_t * e0, - stream_session_t * s0, u32 thread_index, - int *n_tx_pkts) + stream_session_t * s0, int *n_tx_pkts) { - return session_tx_fifo_read_and_snd_i (vm, node, smm, e0, s0, thread_index, - n_tx_pkts, 1); + return session_tx_fifo_read_and_snd_i (vm, node, e0, s0, n_tx_pkts, 1); } int session_tx_fifo_dequeue_and_snd (vlib_main_t * vm, vlib_node_runtime_t * node, - session_manager_main_t * smm, session_fifo_event_t * e0, - stream_session_t * s0, u32 thread_index, - int *n_tx_pkts) + stream_session_t * s0, int *n_tx_pkts) { - return session_tx_fifo_read_and_snd_i (vm, node, smm, e0, s0, thread_index, - n_tx_pkts, 0); + return session_tx_fifo_read_and_snd_i (vm, node, e0, s0, n_tx_pkts, 0); } int session_tx_fifo_dequeue_internal (vlib_main_t * vm, vlib_node_runtime_t * node, - session_manager_main_t * smm, session_fifo_event_t * e0, - stream_session_t * s0, u32 thread_index, - int *n_tx_pkts) + stream_session_t * s0, int *n_tx_pkts) { application_t *app; app = application_get (s0->opaque); @@ -719,8 +773,7 @@ skip_dequeue: } /* Spray packets in per session type frames, since they go to * different nodes */ - rv = (smm->session_tx_fns[s0->session_type]) (vm, node, smm, e0, s0, - my_thread_index, + rv = (smm->session_tx_fns[s0->session_type]) (vm, node, e0, s0, &n_tx_packets); /* Out of buffers */ if (PREDICT_FALSE (rv < 0)) @@ -732,13 +785,13 @@ skip_dequeue: break; case FIFO_EVENT_DISCONNECT: /* Make sure disconnects run after the pending list is drained */ - if (!e0->postponed) + s0 = session_get_from_handle (e0->session_handle); + if (!e0->postponed || svm_fifo_max_dequeue (s0->server_tx_fifo)) { e0->postponed = 1; vec_add1 (smm->pending_disconnects[my_thread_index], *e0); continue; } - s0 = session_get_from_handle (e0->session_handle); stream_session_disconnect_transport (s0); break; case FIFO_EVENT_BUILTIN_RX: |