Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. |
| 3 | * |
| 4 | * Copyright (c) 2011, 2012, Intel Corporation. |
| 5 | * |
| 6 | * Author: Zach Brown <zab@zabbo.net> |
| 7 | * Author: Peter J. Braam <braam@clusterfs.com> |
| 8 | * Author: Phil Schwan <phil@clusterfs.com> |
| 9 | * Author: Eric Barton <eric@bartonsoftware.com> |
| 10 | * |
| 11 | * This file is part of Portals, http://www.sf.net/projects/sandiaportals/ |
| 12 | * |
| 13 | * Portals is free software; you can redistribute it and/or |
| 14 | * modify it under the terms of version 2 of the GNU General Public |
| 15 | * License as published by the Free Software Foundation. |
| 16 | * |
| 17 | * Portals is distributed in the hope that it will be useful, |
| 18 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 19 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 20 | * GNU General Public License for more details. |
| 21 | * |
| 22 | * You should have received a copy of the GNU General Public License |
| 23 | * along with Portals; if not, write to the Free Software |
| 24 | * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. |
| 25 | */ |
| 26 | |
| 27 | #include "socklnd.h" |
| 28 | |
| 29 | ksock_tx_t * |
| 30 | ksocknal_alloc_tx(int type, int size) |
| 31 | { |
| 32 | ksock_tx_t *tx = NULL; |
| 33 | |
| 34 | if (type == KSOCK_MSG_NOOP) { |
| 35 | LASSERT(size == KSOCK_NOOP_TX_SIZE); |
| 36 | |
| 37 | /* searching for a noop tx in free list */ |
| 38 | spin_lock(&ksocknal_data.ksnd_tx_lock); |
| 39 | |
| 40 | if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) { |
| 41 | tx = list_entry(ksocknal_data.ksnd_idle_noop_txs. \ |
| 42 | next, ksock_tx_t, tx_list); |
| 43 | LASSERT(tx->tx_desc_size == size); |
| 44 | list_del(&tx->tx_list); |
| 45 | } |
| 46 | |
| 47 | spin_unlock(&ksocknal_data.ksnd_tx_lock); |
| 48 | } |
| 49 | |
| 50 | if (tx == NULL) |
| 51 | LIBCFS_ALLOC(tx, size); |
| 52 | |
| 53 | if (tx == NULL) |
| 54 | return NULL; |
| 55 | |
| 56 | atomic_set(&tx->tx_refcount, 1); |
| 57 | tx->tx_zc_aborted = 0; |
| 58 | tx->tx_zc_capable = 0; |
| 59 | tx->tx_zc_checked = 0; |
| 60 | tx->tx_desc_size = size; |
| 61 | |
| 62 | atomic_inc(&ksocknal_data.ksnd_nactive_txs); |
| 63 | |
| 64 | return tx; |
| 65 | } |
| 66 | |
| 67 | ksock_tx_t * |
| 68 | ksocknal_alloc_tx_noop(__u64 cookie, int nonblk) |
| 69 | { |
| 70 | ksock_tx_t *tx; |
| 71 | |
| 72 | tx = ksocknal_alloc_tx(KSOCK_MSG_NOOP, KSOCK_NOOP_TX_SIZE); |
| 73 | if (tx == NULL) { |
| 74 | CERROR("Can't allocate noop tx desc\n"); |
| 75 | return NULL; |
| 76 | } |
| 77 | |
| 78 | tx->tx_conn = NULL; |
| 79 | tx->tx_lnetmsg = NULL; |
| 80 | tx->tx_kiov = NULL; |
| 81 | tx->tx_nkiov = 0; |
| 82 | tx->tx_iov = tx->tx_frags.virt.iov; |
| 83 | tx->tx_niov = 1; |
| 84 | tx->tx_nonblk = nonblk; |
| 85 | |
| 86 | socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_NOOP); |
| 87 | tx->tx_msg.ksm_zc_cookies[1] = cookie; |
| 88 | |
| 89 | return tx; |
| 90 | } |
| 91 | |
| 92 | |
| 93 | void |
| 94 | ksocknal_free_tx (ksock_tx_t *tx) |
| 95 | { |
| 96 | atomic_dec(&ksocknal_data.ksnd_nactive_txs); |
| 97 | |
| 98 | if (tx->tx_lnetmsg == NULL && tx->tx_desc_size == KSOCK_NOOP_TX_SIZE) { |
| 99 | /* it's a noop tx */ |
| 100 | spin_lock(&ksocknal_data.ksnd_tx_lock); |
| 101 | |
| 102 | list_add(&tx->tx_list, &ksocknal_data.ksnd_idle_noop_txs); |
| 103 | |
| 104 | spin_unlock(&ksocknal_data.ksnd_tx_lock); |
| 105 | } else { |
| 106 | LIBCFS_FREE(tx, tx->tx_desc_size); |
| 107 | } |
| 108 | } |
| 109 | |
| 110 | int |
| 111 | ksocknal_send_iov (ksock_conn_t *conn, ksock_tx_t *tx) |
| 112 | { |
| 113 | struct iovec *iov = tx->tx_iov; |
| 114 | int nob; |
| 115 | int rc; |
| 116 | |
| 117 | LASSERT (tx->tx_niov > 0); |
| 118 | |
| 119 | /* Never touch tx->tx_iov inside ksocknal_lib_send_iov() */ |
| 120 | rc = ksocknal_lib_send_iov(conn, tx); |
| 121 | |
| 122 | if (rc <= 0) /* sent nothing? */ |
| 123 | return (rc); |
| 124 | |
| 125 | nob = rc; |
| 126 | LASSERT (nob <= tx->tx_resid); |
| 127 | tx->tx_resid -= nob; |
| 128 | |
| 129 | /* "consume" iov */ |
| 130 | do { |
| 131 | LASSERT (tx->tx_niov > 0); |
| 132 | |
| 133 | if (nob < (int) iov->iov_len) { |
| 134 | iov->iov_base = (void *)((char *)iov->iov_base + nob); |
| 135 | iov->iov_len -= nob; |
| 136 | return (rc); |
| 137 | } |
| 138 | |
| 139 | nob -= iov->iov_len; |
| 140 | tx->tx_iov = ++iov; |
| 141 | tx->tx_niov--; |
| 142 | } while (nob != 0); |
| 143 | |
| 144 | return (rc); |
| 145 | } |
| 146 | |
| 147 | int |
| 148 | ksocknal_send_kiov (ksock_conn_t *conn, ksock_tx_t *tx) |
| 149 | { |
| 150 | lnet_kiov_t *kiov = tx->tx_kiov; |
| 151 | int nob; |
| 152 | int rc; |
| 153 | |
| 154 | LASSERT (tx->tx_niov == 0); |
| 155 | LASSERT (tx->tx_nkiov > 0); |
| 156 | |
| 157 | /* Never touch tx->tx_kiov inside ksocknal_lib_send_kiov() */ |
| 158 | rc = ksocknal_lib_send_kiov(conn, tx); |
| 159 | |
| 160 | if (rc <= 0) /* sent nothing? */ |
| 161 | return (rc); |
| 162 | |
| 163 | nob = rc; |
| 164 | LASSERT (nob <= tx->tx_resid); |
| 165 | tx->tx_resid -= nob; |
| 166 | |
| 167 | /* "consume" kiov */ |
| 168 | do { |
| 169 | LASSERT(tx->tx_nkiov > 0); |
| 170 | |
| 171 | if (nob < (int)kiov->kiov_len) { |
| 172 | kiov->kiov_offset += nob; |
| 173 | kiov->kiov_len -= nob; |
| 174 | return rc; |
| 175 | } |
| 176 | |
| 177 | nob -= (int)kiov->kiov_len; |
| 178 | tx->tx_kiov = ++kiov; |
| 179 | tx->tx_nkiov--; |
| 180 | } while (nob != 0); |
| 181 | |
| 182 | return (rc); |
| 183 | } |
| 184 | |
| 185 | int |
| 186 | ksocknal_transmit (ksock_conn_t *conn, ksock_tx_t *tx) |
| 187 | { |
| 188 | int rc; |
| 189 | int bufnob; |
| 190 | |
| 191 | if (ksocknal_data.ksnd_stall_tx != 0) { |
| 192 | cfs_pause(cfs_time_seconds(ksocknal_data.ksnd_stall_tx)); |
| 193 | } |
| 194 | |
| 195 | LASSERT (tx->tx_resid != 0); |
| 196 | |
| 197 | rc = ksocknal_connsock_addref(conn); |
| 198 | if (rc != 0) { |
| 199 | LASSERT (conn->ksnc_closing); |
| 200 | return (-ESHUTDOWN); |
| 201 | } |
| 202 | |
| 203 | do { |
| 204 | if (ksocknal_data.ksnd_enomem_tx > 0) { |
| 205 | /* testing... */ |
| 206 | ksocknal_data.ksnd_enomem_tx--; |
| 207 | rc = -EAGAIN; |
| 208 | } else if (tx->tx_niov != 0) { |
| 209 | rc = ksocknal_send_iov (conn, tx); |
| 210 | } else { |
| 211 | rc = ksocknal_send_kiov (conn, tx); |
| 212 | } |
| 213 | |
| 214 | bufnob = cfs_sock_wmem_queued(conn->ksnc_sock); |
| 215 | if (rc > 0) /* sent something? */ |
| 216 | conn->ksnc_tx_bufnob += rc; /* account it */ |
| 217 | |
| 218 | if (bufnob < conn->ksnc_tx_bufnob) { |
| 219 | /* allocated send buffer bytes < computed; infer |
| 220 | * something got ACKed */ |
| 221 | conn->ksnc_tx_deadline = |
| 222 | cfs_time_shift(*ksocknal_tunables.ksnd_timeout); |
| 223 | conn->ksnc_peer->ksnp_last_alive = cfs_time_current(); |
| 224 | conn->ksnc_tx_bufnob = bufnob; |
| 225 | mb(); |
| 226 | } |
| 227 | |
| 228 | if (rc <= 0) { /* Didn't write anything? */ |
| 229 | |
| 230 | if (rc == 0) /* some stacks return 0 instead of -EAGAIN */ |
| 231 | rc = -EAGAIN; |
| 232 | |
| 233 | /* Check if EAGAIN is due to memory pressure */ |
| 234 | if(rc == -EAGAIN && ksocknal_lib_memory_pressure(conn)) |
| 235 | rc = -ENOMEM; |
| 236 | |
| 237 | break; |
| 238 | } |
| 239 | |
| 240 | /* socket's wmem_queued now includes 'rc' bytes */ |
| 241 | atomic_sub (rc, &conn->ksnc_tx_nob); |
| 242 | rc = 0; |
| 243 | |
| 244 | } while (tx->tx_resid != 0); |
| 245 | |
| 246 | ksocknal_connsock_decref(conn); |
| 247 | return (rc); |
| 248 | } |
| 249 | |
| 250 | int |
| 251 | ksocknal_recv_iov (ksock_conn_t *conn) |
| 252 | { |
| 253 | struct iovec *iov = conn->ksnc_rx_iov; |
| 254 | int nob; |
| 255 | int rc; |
| 256 | |
| 257 | LASSERT (conn->ksnc_rx_niov > 0); |
| 258 | |
| 259 | /* Never touch conn->ksnc_rx_iov or change connection |
| 260 | * status inside ksocknal_lib_recv_iov */ |
| 261 | rc = ksocknal_lib_recv_iov(conn); |
| 262 | |
| 263 | if (rc <= 0) |
| 264 | return (rc); |
| 265 | |
| 266 | /* received something... */ |
| 267 | nob = rc; |
| 268 | |
| 269 | conn->ksnc_peer->ksnp_last_alive = cfs_time_current(); |
| 270 | conn->ksnc_rx_deadline = |
| 271 | cfs_time_shift(*ksocknal_tunables.ksnd_timeout); |
| 272 | mb(); /* order with setting rx_started */ |
| 273 | conn->ksnc_rx_started = 1; |
| 274 | |
| 275 | conn->ksnc_rx_nob_wanted -= nob; |
| 276 | conn->ksnc_rx_nob_left -= nob; |
| 277 | |
| 278 | do { |
| 279 | LASSERT (conn->ksnc_rx_niov > 0); |
| 280 | |
| 281 | if (nob < (int)iov->iov_len) { |
| 282 | iov->iov_len -= nob; |
| 283 | iov->iov_base = (void *)((char *)iov->iov_base + nob); |
| 284 | return (-EAGAIN); |
| 285 | } |
| 286 | |
| 287 | nob -= iov->iov_len; |
| 288 | conn->ksnc_rx_iov = ++iov; |
| 289 | conn->ksnc_rx_niov--; |
| 290 | } while (nob != 0); |
| 291 | |
| 292 | return (rc); |
| 293 | } |
| 294 | |
| 295 | int |
| 296 | ksocknal_recv_kiov (ksock_conn_t *conn) |
| 297 | { |
| 298 | lnet_kiov_t *kiov = conn->ksnc_rx_kiov; |
| 299 | int nob; |
| 300 | int rc; |
| 301 | LASSERT (conn->ksnc_rx_nkiov > 0); |
| 302 | |
| 303 | /* Never touch conn->ksnc_rx_kiov or change connection |
| 304 | * status inside ksocknal_lib_recv_iov */ |
| 305 | rc = ksocknal_lib_recv_kiov(conn); |
| 306 | |
| 307 | if (rc <= 0) |
| 308 | return (rc); |
| 309 | |
| 310 | /* received something... */ |
| 311 | nob = rc; |
| 312 | |
| 313 | conn->ksnc_peer->ksnp_last_alive = cfs_time_current(); |
| 314 | conn->ksnc_rx_deadline = |
| 315 | cfs_time_shift(*ksocknal_tunables.ksnd_timeout); |
| 316 | mb(); /* order with setting rx_started */ |
| 317 | conn->ksnc_rx_started = 1; |
| 318 | |
| 319 | conn->ksnc_rx_nob_wanted -= nob; |
| 320 | conn->ksnc_rx_nob_left -= nob; |
| 321 | |
| 322 | do { |
| 323 | LASSERT (conn->ksnc_rx_nkiov > 0); |
| 324 | |
| 325 | if (nob < (int) kiov->kiov_len) { |
| 326 | kiov->kiov_offset += nob; |
| 327 | kiov->kiov_len -= nob; |
| 328 | return -EAGAIN; |
| 329 | } |
| 330 | |
| 331 | nob -= kiov->kiov_len; |
| 332 | conn->ksnc_rx_kiov = ++kiov; |
| 333 | conn->ksnc_rx_nkiov--; |
| 334 | } while (nob != 0); |
| 335 | |
| 336 | return 1; |
| 337 | } |
| 338 | |
| 339 | int |
| 340 | ksocknal_receive (ksock_conn_t *conn) |
| 341 | { |
| 342 | /* Return 1 on success, 0 on EOF, < 0 on error. |
| 343 | * Caller checks ksnc_rx_nob_wanted to determine |
| 344 | * progress/completion. */ |
| 345 | int rc; |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 346 | |
| 347 | if (ksocknal_data.ksnd_stall_rx != 0) { |
| 348 | cfs_pause(cfs_time_seconds (ksocknal_data.ksnd_stall_rx)); |
| 349 | } |
| 350 | |
| 351 | rc = ksocknal_connsock_addref(conn); |
| 352 | if (rc != 0) { |
| 353 | LASSERT (conn->ksnc_closing); |
| 354 | return (-ESHUTDOWN); |
| 355 | } |
| 356 | |
| 357 | for (;;) { |
| 358 | if (conn->ksnc_rx_niov != 0) |
| 359 | rc = ksocknal_recv_iov (conn); |
| 360 | else |
| 361 | rc = ksocknal_recv_kiov (conn); |
| 362 | |
| 363 | if (rc <= 0) { |
| 364 | /* error/EOF or partial receive */ |
| 365 | if (rc == -EAGAIN) { |
| 366 | rc = 1; |
| 367 | } else if (rc == 0 && conn->ksnc_rx_started) { |
| 368 | /* EOF in the middle of a message */ |
| 369 | rc = -EPROTO; |
| 370 | } |
| 371 | break; |
| 372 | } |
| 373 | |
| 374 | /* Completed a fragment */ |
| 375 | |
| 376 | if (conn->ksnc_rx_nob_wanted == 0) { |
| 377 | rc = 1; |
| 378 | break; |
| 379 | } |
| 380 | } |
| 381 | |
| 382 | ksocknal_connsock_decref(conn); |
Greg Kroah-Hartman | 0a3bdb0 | 2013-08-03 10:35:28 +0800 | [diff] [blame] | 383 | return rc; |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 384 | } |
| 385 | |
| 386 | void |
| 387 | ksocknal_tx_done (lnet_ni_t *ni, ksock_tx_t *tx) |
| 388 | { |
| 389 | lnet_msg_t *lnetmsg = tx->tx_lnetmsg; |
| 390 | int rc = (tx->tx_resid == 0 && !tx->tx_zc_aborted) ? 0 : -EIO; |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 391 | |
| 392 | LASSERT(ni != NULL || tx->tx_conn != NULL); |
| 393 | |
| 394 | if (tx->tx_conn != NULL) |
| 395 | ksocknal_conn_decref(tx->tx_conn); |
| 396 | |
| 397 | if (ni == NULL && tx->tx_conn != NULL) |
| 398 | ni = tx->tx_conn->ksnc_peer->ksnp_ni; |
| 399 | |
| 400 | ksocknal_free_tx (tx); |
| 401 | if (lnetmsg != NULL) /* KSOCK_MSG_NOOP go without lnetmsg */ |
| 402 | lnet_finalize (ni, lnetmsg, rc); |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 403 | } |
| 404 | |
| 405 | void |
| 406 | ksocknal_txlist_done (lnet_ni_t *ni, struct list_head *txlist, int error) |
| 407 | { |
| 408 | ksock_tx_t *tx; |
| 409 | |
| 410 | while (!list_empty (txlist)) { |
| 411 | tx = list_entry (txlist->next, ksock_tx_t, tx_list); |
| 412 | |
| 413 | if (error && tx->tx_lnetmsg != NULL) { |
| 414 | CNETERR("Deleting packet type %d len %d %s->%s\n", |
| 415 | le32_to_cpu (tx->tx_lnetmsg->msg_hdr.type), |
| 416 | le32_to_cpu (tx->tx_lnetmsg->msg_hdr.payload_length), |
| 417 | libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.src_nid)), |
| 418 | libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.dest_nid))); |
| 419 | } else if (error) { |
| 420 | CNETERR("Deleting noop packet\n"); |
| 421 | } |
| 422 | |
| 423 | list_del (&tx->tx_list); |
| 424 | |
| 425 | LASSERT (atomic_read(&tx->tx_refcount) == 1); |
| 426 | ksocknal_tx_done (ni, tx); |
| 427 | } |
| 428 | } |
| 429 | |
| 430 | static void |
| 431 | ksocknal_check_zc_req(ksock_tx_t *tx) |
| 432 | { |
| 433 | ksock_conn_t *conn = tx->tx_conn; |
| 434 | ksock_peer_t *peer = conn->ksnc_peer; |
| 435 | |
| 436 | /* Set tx_msg.ksm_zc_cookies[0] to a unique non-zero cookie and add tx |
| 437 | * to ksnp_zc_req_list if some fragment of this message should be sent |
| 438 | * zero-copy. Our peer will send an ACK containing this cookie when |
| 439 | * she has received this message to tell us we can signal completion. |
| 440 | * tx_msg.ksm_zc_cookies[0] remains non-zero while tx is on |
| 441 | * ksnp_zc_req_list. */ |
| 442 | LASSERT (tx->tx_msg.ksm_type != KSOCK_MSG_NOOP); |
| 443 | LASSERT (tx->tx_zc_capable); |
| 444 | |
| 445 | tx->tx_zc_checked = 1; |
| 446 | |
| 447 | if (conn->ksnc_proto == &ksocknal_protocol_v1x || |
| 448 | !conn->ksnc_zc_capable) |
| 449 | return; |
| 450 | |
| 451 | /* assign cookie and queue tx to pending list, it will be released when |
| 452 | * a matching ack is received. See ksocknal_handle_zcack() */ |
| 453 | |
| 454 | ksocknal_tx_addref(tx); |
| 455 | |
| 456 | spin_lock(&peer->ksnp_lock); |
| 457 | |
| 458 | /* ZC_REQ is going to be pinned to the peer */ |
| 459 | tx->tx_deadline = |
| 460 | cfs_time_shift(*ksocknal_tunables.ksnd_timeout); |
| 461 | |
| 462 | LASSERT (tx->tx_msg.ksm_zc_cookies[0] == 0); |
| 463 | |
| 464 | tx->tx_msg.ksm_zc_cookies[0] = peer->ksnp_zc_next_cookie++; |
| 465 | |
| 466 | if (peer->ksnp_zc_next_cookie == 0) |
| 467 | peer->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1; |
| 468 | |
| 469 | list_add_tail(&tx->tx_zc_list, &peer->ksnp_zc_req_list); |
| 470 | |
| 471 | spin_unlock(&peer->ksnp_lock); |
| 472 | } |
| 473 | |
| 474 | static void |
| 475 | ksocknal_uncheck_zc_req(ksock_tx_t *tx) |
| 476 | { |
| 477 | ksock_peer_t *peer = tx->tx_conn->ksnc_peer; |
| 478 | |
| 479 | LASSERT(tx->tx_msg.ksm_type != KSOCK_MSG_NOOP); |
| 480 | LASSERT(tx->tx_zc_capable); |
| 481 | |
| 482 | tx->tx_zc_checked = 0; |
| 483 | |
| 484 | spin_lock(&peer->ksnp_lock); |
| 485 | |
| 486 | if (tx->tx_msg.ksm_zc_cookies[0] == 0) { |
| 487 | /* Not waiting for an ACK */ |
| 488 | spin_unlock(&peer->ksnp_lock); |
| 489 | return; |
| 490 | } |
| 491 | |
| 492 | tx->tx_msg.ksm_zc_cookies[0] = 0; |
| 493 | list_del(&tx->tx_zc_list); |
| 494 | |
| 495 | spin_unlock(&peer->ksnp_lock); |
| 496 | |
| 497 | ksocknal_tx_decref(tx); |
| 498 | } |
| 499 | |
| 500 | int |
| 501 | ksocknal_process_transmit (ksock_conn_t *conn, ksock_tx_t *tx) |
| 502 | { |
| 503 | int rc; |
| 504 | |
| 505 | if (tx->tx_zc_capable && !tx->tx_zc_checked) |
| 506 | ksocknal_check_zc_req(tx); |
| 507 | |
| 508 | rc = ksocknal_transmit (conn, tx); |
| 509 | |
| 510 | CDEBUG (D_NET, "send(%d) %d\n", tx->tx_resid, rc); |
| 511 | |
| 512 | if (tx->tx_resid == 0) { |
| 513 | /* Sent everything OK */ |
| 514 | LASSERT (rc == 0); |
| 515 | |
| 516 | return (0); |
| 517 | } |
| 518 | |
| 519 | if (rc == -EAGAIN) |
| 520 | return (rc); |
| 521 | |
| 522 | if (rc == -ENOMEM) { |
| 523 | static int counter; |
| 524 | |
| 525 | counter++; /* exponential backoff warnings */ |
| 526 | if ((counter & (-counter)) == counter) |
| 527 | CWARN("%u ENOMEM tx %p (%u allocated)\n", |
| 528 | counter, conn, atomic_read(&libcfs_kmemory)); |
| 529 | |
| 530 | /* Queue on ksnd_enomem_conns for retry after a timeout */ |
| 531 | spin_lock_bh(&ksocknal_data.ksnd_reaper_lock); |
| 532 | |
| 533 | /* enomem list takes over scheduler's ref... */ |
| 534 | LASSERT (conn->ksnc_tx_scheduled); |
| 535 | list_add_tail(&conn->ksnc_tx_list, |
| 536 | &ksocknal_data.ksnd_enomem_conns); |
| 537 | if (!cfs_time_aftereq(cfs_time_add(cfs_time_current(), |
| 538 | SOCKNAL_ENOMEM_RETRY), |
| 539 | ksocknal_data.ksnd_reaper_waketime)) |
| 540 | wake_up (&ksocknal_data.ksnd_reaper_waitq); |
| 541 | |
| 542 | spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock); |
| 543 | return (rc); |
| 544 | } |
| 545 | |
| 546 | /* Actual error */ |
| 547 | LASSERT (rc < 0); |
| 548 | |
| 549 | if (!conn->ksnc_closing) { |
| 550 | switch (rc) { |
| 551 | case -ECONNRESET: |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 552 | LCONSOLE_WARN("Host %pI4h reset our connection " |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 553 | "while we were sending data; it may have " |
| 554 | "rebooted.\n", |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 555 | &conn->ksnc_ipaddr); |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 556 | break; |
| 557 | default: |
| 558 | LCONSOLE_WARN("There was an unexpected network error " |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 559 | "while writing to %pI4h: %d.\n", |
| 560 | &conn->ksnc_ipaddr, rc); |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 561 | break; |
| 562 | } |
| 563 | CDEBUG(D_NET, "[%p] Error %d on write to %s" |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 564 | " ip %pI4h:%d\n", conn, rc, |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 565 | libcfs_id2str(conn->ksnc_peer->ksnp_id), |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 566 | &conn->ksnc_ipaddr, |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 567 | conn->ksnc_port); |
| 568 | } |
| 569 | |
| 570 | if (tx->tx_zc_checked) |
| 571 | ksocknal_uncheck_zc_req(tx); |
| 572 | |
| 573 | /* it's not an error if conn is being closed */ |
| 574 | ksocknal_close_conn_and_siblings (conn, |
| 575 | (conn->ksnc_closing) ? 0 : rc); |
| 576 | |
| 577 | return (rc); |
| 578 | } |
| 579 | |
| 580 | void |
| 581 | ksocknal_launch_connection_locked (ksock_route_t *route) |
| 582 | { |
| 583 | |
| 584 | /* called holding write lock on ksnd_global_lock */ |
| 585 | |
| 586 | LASSERT (!route->ksnr_scheduled); |
| 587 | LASSERT (!route->ksnr_connecting); |
| 588 | LASSERT ((ksocknal_route_mask() & ~route->ksnr_connected) != 0); |
| 589 | |
| 590 | route->ksnr_scheduled = 1; /* scheduling conn for connd */ |
| 591 | ksocknal_route_addref(route); /* extra ref for connd */ |
| 592 | |
| 593 | spin_lock_bh(&ksocknal_data.ksnd_connd_lock); |
| 594 | |
| 595 | list_add_tail(&route->ksnr_connd_list, |
| 596 | &ksocknal_data.ksnd_connd_routes); |
| 597 | wake_up(&ksocknal_data.ksnd_connd_waitq); |
| 598 | |
| 599 | spin_unlock_bh(&ksocknal_data.ksnd_connd_lock); |
| 600 | } |
| 601 | |
| 602 | void |
| 603 | ksocknal_launch_all_connections_locked (ksock_peer_t *peer) |
| 604 | { |
| 605 | ksock_route_t *route; |
| 606 | |
| 607 | /* called holding write lock on ksnd_global_lock */ |
| 608 | for (;;) { |
| 609 | /* launch any/all connections that need it */ |
| 610 | route = ksocknal_find_connectable_route_locked(peer); |
| 611 | if (route == NULL) |
| 612 | return; |
| 613 | |
| 614 | ksocknal_launch_connection_locked(route); |
| 615 | } |
| 616 | } |
| 617 | |
| 618 | ksock_conn_t * |
| 619 | ksocknal_find_conn_locked(ksock_peer_t *peer, ksock_tx_t *tx, int nonblk) |
| 620 | { |
| 621 | struct list_head *tmp; |
| 622 | ksock_conn_t *conn; |
| 623 | ksock_conn_t *typed = NULL; |
| 624 | ksock_conn_t *fallback = NULL; |
| 625 | int tnob = 0; |
| 626 | int fnob = 0; |
| 627 | |
| 628 | list_for_each (tmp, &peer->ksnp_conns) { |
| 629 | ksock_conn_t *c = list_entry(tmp, ksock_conn_t, ksnc_list); |
| 630 | int nob = atomic_read(&c->ksnc_tx_nob) + |
| 631 | cfs_sock_wmem_queued(c->ksnc_sock); |
| 632 | int rc; |
| 633 | |
| 634 | LASSERT (!c->ksnc_closing); |
| 635 | LASSERT (c->ksnc_proto != NULL && |
| 636 | c->ksnc_proto->pro_match_tx != NULL); |
| 637 | |
| 638 | rc = c->ksnc_proto->pro_match_tx(c, tx, nonblk); |
| 639 | |
| 640 | switch (rc) { |
| 641 | default: |
| 642 | LBUG(); |
| 643 | case SOCKNAL_MATCH_NO: /* protocol rejected the tx */ |
| 644 | continue; |
| 645 | |
| 646 | case SOCKNAL_MATCH_YES: /* typed connection */ |
| 647 | if (typed == NULL || tnob > nob || |
| 648 | (tnob == nob && *ksocknal_tunables.ksnd_round_robin && |
| 649 | cfs_time_after(typed->ksnc_tx_last_post, c->ksnc_tx_last_post))) { |
| 650 | typed = c; |
| 651 | tnob = nob; |
| 652 | } |
| 653 | break; |
| 654 | |
| 655 | case SOCKNAL_MATCH_MAY: /* fallback connection */ |
| 656 | if (fallback == NULL || fnob > nob || |
| 657 | (fnob == nob && *ksocknal_tunables.ksnd_round_robin && |
| 658 | cfs_time_after(fallback->ksnc_tx_last_post, c->ksnc_tx_last_post))) { |
| 659 | fallback = c; |
| 660 | fnob = nob; |
| 661 | } |
| 662 | break; |
| 663 | } |
| 664 | } |
| 665 | |
| 666 | /* prefer the typed selection */ |
| 667 | conn = (typed != NULL) ? typed : fallback; |
| 668 | |
| 669 | if (conn != NULL) |
| 670 | conn->ksnc_tx_last_post = cfs_time_current(); |
| 671 | |
| 672 | return conn; |
| 673 | } |
| 674 | |
| 675 | void |
| 676 | ksocknal_tx_prep(ksock_conn_t *conn, ksock_tx_t *tx) |
| 677 | { |
| 678 | conn->ksnc_proto->pro_pack(tx); |
| 679 | |
| 680 | atomic_add (tx->tx_nob, &conn->ksnc_tx_nob); |
| 681 | ksocknal_conn_addref(conn); /* +1 ref for tx */ |
| 682 | tx->tx_conn = conn; |
| 683 | } |
| 684 | |
| 685 | void |
| 686 | ksocknal_queue_tx_locked (ksock_tx_t *tx, ksock_conn_t *conn) |
| 687 | { |
| 688 | ksock_sched_t *sched = conn->ksnc_scheduler; |
| 689 | ksock_msg_t *msg = &tx->tx_msg; |
| 690 | ksock_tx_t *ztx = NULL; |
| 691 | int bufnob = 0; |
| 692 | |
| 693 | /* called holding global lock (read or irq-write) and caller may |
| 694 | * not have dropped this lock between finding conn and calling me, |
| 695 | * so we don't need the {get,put}connsock dance to deref |
| 696 | * ksnc_sock... */ |
| 697 | LASSERT(!conn->ksnc_closing); |
| 698 | |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 699 | CDEBUG(D_NET, "Sending to %s ip %pI4h:%d\n", |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 700 | libcfs_id2str(conn->ksnc_peer->ksnp_id), |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 701 | &conn->ksnc_ipaddr, |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 702 | conn->ksnc_port); |
| 703 | |
| 704 | ksocknal_tx_prep(conn, tx); |
| 705 | |
| 706 | /* Ensure the frags we've been given EXACTLY match the number of |
| 707 | * bytes we want to send. Many TCP/IP stacks disregard any total |
| 708 | * size parameters passed to them and just look at the frags. |
| 709 | * |
| 710 | * We always expect at least 1 mapped fragment containing the |
| 711 | * complete ksocknal message header. */ |
| 712 | LASSERT (lnet_iov_nob (tx->tx_niov, tx->tx_iov) + |
| 713 | lnet_kiov_nob(tx->tx_nkiov, tx->tx_kiov) == |
| 714 | (unsigned int)tx->tx_nob); |
| 715 | LASSERT (tx->tx_niov >= 1); |
| 716 | LASSERT (tx->tx_resid == tx->tx_nob); |
| 717 | |
| 718 | CDEBUG (D_NET, "Packet %p type %d, nob %d niov %d nkiov %d\n", |
| 719 | tx, (tx->tx_lnetmsg != NULL) ? tx->tx_lnetmsg->msg_hdr.type: |
| 720 | KSOCK_MSG_NOOP, |
| 721 | tx->tx_nob, tx->tx_niov, tx->tx_nkiov); |
| 722 | |
| 723 | /* |
| 724 | * FIXME: SOCK_WMEM_QUEUED and SOCK_ERROR could block in __DARWIN8__ |
| 725 | * but they're used inside spinlocks a lot. |
| 726 | */ |
| 727 | bufnob = cfs_sock_wmem_queued(conn->ksnc_sock); |
| 728 | spin_lock_bh(&sched->kss_lock); |
| 729 | |
| 730 | if (list_empty(&conn->ksnc_tx_queue) && bufnob == 0) { |
| 731 | /* First packet starts the timeout */ |
| 732 | conn->ksnc_tx_deadline = |
| 733 | cfs_time_shift(*ksocknal_tunables.ksnd_timeout); |
| 734 | if (conn->ksnc_tx_bufnob > 0) /* something got ACKed */ |
| 735 | conn->ksnc_peer->ksnp_last_alive = cfs_time_current(); |
| 736 | conn->ksnc_tx_bufnob = 0; |
| 737 | mb(); /* order with adding to tx_queue */ |
| 738 | } |
| 739 | |
| 740 | if (msg->ksm_type == KSOCK_MSG_NOOP) { |
| 741 | /* The packet is noop ZC ACK, try to piggyback the ack_cookie |
| 742 | * on a normal packet so I don't need to send it */ |
| 743 | LASSERT (msg->ksm_zc_cookies[1] != 0); |
| 744 | LASSERT (conn->ksnc_proto->pro_queue_tx_zcack != NULL); |
| 745 | |
| 746 | if (conn->ksnc_proto->pro_queue_tx_zcack(conn, tx, 0)) |
| 747 | ztx = tx; /* ZC ACK piggybacked on ztx release tx later */ |
| 748 | |
| 749 | } else { |
| 750 | /* It's a normal packet - can it piggback a noop zc-ack that |
| 751 | * has been queued already? */ |
| 752 | LASSERT (msg->ksm_zc_cookies[1] == 0); |
| 753 | LASSERT (conn->ksnc_proto->pro_queue_tx_msg != NULL); |
| 754 | |
| 755 | ztx = conn->ksnc_proto->pro_queue_tx_msg(conn, tx); |
| 756 | /* ztx will be released later */ |
| 757 | } |
| 758 | |
| 759 | if (ztx != NULL) { |
| 760 | atomic_sub (ztx->tx_nob, &conn->ksnc_tx_nob); |
| 761 | list_add_tail(&ztx->tx_list, &sched->kss_zombie_noop_txs); |
| 762 | } |
| 763 | |
| 764 | if (conn->ksnc_tx_ready && /* able to send */ |
| 765 | !conn->ksnc_tx_scheduled) { /* not scheduled to send */ |
| 766 | /* +1 ref for scheduler */ |
| 767 | ksocknal_conn_addref(conn); |
| 768 | list_add_tail (&conn->ksnc_tx_list, |
| 769 | &sched->kss_tx_conns); |
| 770 | conn->ksnc_tx_scheduled = 1; |
| 771 | wake_up (&sched->kss_waitq); |
| 772 | } |
| 773 | |
| 774 | spin_unlock_bh(&sched->kss_lock); |
| 775 | } |
| 776 | |
| 777 | |
| 778 | ksock_route_t * |
| 779 | ksocknal_find_connectable_route_locked (ksock_peer_t *peer) |
| 780 | { |
| 781 | cfs_time_t now = cfs_time_current(); |
| 782 | struct list_head *tmp; |
| 783 | ksock_route_t *route; |
| 784 | |
| 785 | list_for_each (tmp, &peer->ksnp_routes) { |
| 786 | route = list_entry (tmp, ksock_route_t, ksnr_list); |
| 787 | |
| 788 | LASSERT (!route->ksnr_connecting || route->ksnr_scheduled); |
| 789 | |
| 790 | if (route->ksnr_scheduled) /* connections being established */ |
| 791 | continue; |
| 792 | |
| 793 | /* all route types connected ? */ |
| 794 | if ((ksocknal_route_mask() & ~route->ksnr_connected) == 0) |
| 795 | continue; |
| 796 | |
| 797 | if (!(route->ksnr_retry_interval == 0 || /* first attempt */ |
| 798 | cfs_time_aftereq(now, route->ksnr_timeout))) { |
| 799 | CDEBUG(D_NET, |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 800 | "Too soon to retry route %pI4h " |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 801 | "(cnted %d, interval %ld, %ld secs later)\n", |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 802 | &route->ksnr_ipaddr, |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 803 | route->ksnr_connected, |
| 804 | route->ksnr_retry_interval, |
| 805 | cfs_duration_sec(route->ksnr_timeout - now)); |
| 806 | continue; |
| 807 | } |
| 808 | |
| 809 | return (route); |
| 810 | } |
| 811 | |
| 812 | return (NULL); |
| 813 | } |
| 814 | |
| 815 | ksock_route_t * |
| 816 | ksocknal_find_connecting_route_locked (ksock_peer_t *peer) |
| 817 | { |
| 818 | struct list_head *tmp; |
| 819 | ksock_route_t *route; |
| 820 | |
| 821 | list_for_each (tmp, &peer->ksnp_routes) { |
| 822 | route = list_entry (tmp, ksock_route_t, ksnr_list); |
| 823 | |
| 824 | LASSERT (!route->ksnr_connecting || route->ksnr_scheduled); |
| 825 | |
| 826 | if (route->ksnr_scheduled) |
| 827 | return (route); |
| 828 | } |
| 829 | |
| 830 | return (NULL); |
| 831 | } |
| 832 | |
| 833 | int |
| 834 | ksocknal_launch_packet (lnet_ni_t *ni, ksock_tx_t *tx, lnet_process_id_t id) |
| 835 | { |
| 836 | ksock_peer_t *peer; |
| 837 | ksock_conn_t *conn; |
| 838 | rwlock_t *g_lock; |
| 839 | int retry; |
| 840 | int rc; |
| 841 | |
| 842 | LASSERT (tx->tx_conn == NULL); |
| 843 | |
| 844 | g_lock = &ksocknal_data.ksnd_global_lock; |
| 845 | |
| 846 | for (retry = 0;; retry = 1) { |
| 847 | read_lock(g_lock); |
| 848 | peer = ksocknal_find_peer_locked(ni, id); |
| 849 | if (peer != NULL) { |
| 850 | if (ksocknal_find_connectable_route_locked(peer) == NULL) { |
| 851 | conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk); |
| 852 | if (conn != NULL) { |
| 853 | /* I've got no routes that need to be |
| 854 | * connecting and I do have an actual |
| 855 | * connection... */ |
| 856 | ksocknal_queue_tx_locked (tx, conn); |
| 857 | read_unlock(g_lock); |
| 858 | return (0); |
| 859 | } |
| 860 | } |
| 861 | } |
| 862 | |
| 863 | /* I'll need a write lock... */ |
| 864 | read_unlock(g_lock); |
| 865 | |
| 866 | write_lock_bh(g_lock); |
| 867 | |
| 868 | peer = ksocknal_find_peer_locked(ni, id); |
| 869 | if (peer != NULL) |
| 870 | break; |
| 871 | |
| 872 | write_unlock_bh(g_lock); |
| 873 | |
| 874 | if ((id.pid & LNET_PID_USERFLAG) != 0) { |
| 875 | CERROR("Refusing to create a connection to " |
| 876 | "userspace process %s\n", libcfs_id2str(id)); |
| 877 | return -EHOSTUNREACH; |
| 878 | } |
| 879 | |
| 880 | if (retry) { |
| 881 | CERROR("Can't find peer %s\n", libcfs_id2str(id)); |
| 882 | return -EHOSTUNREACH; |
| 883 | } |
| 884 | |
| 885 | rc = ksocknal_add_peer(ni, id, |
| 886 | LNET_NIDADDR(id.nid), |
| 887 | lnet_acceptor_port()); |
| 888 | if (rc != 0) { |
| 889 | CERROR("Can't add peer %s: %d\n", |
| 890 | libcfs_id2str(id), rc); |
| 891 | return rc; |
| 892 | } |
| 893 | } |
| 894 | |
| 895 | ksocknal_launch_all_connections_locked(peer); |
| 896 | |
| 897 | conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk); |
| 898 | if (conn != NULL) { |
| 899 | /* Connection exists; queue message on it */ |
| 900 | ksocknal_queue_tx_locked (tx, conn); |
| 901 | write_unlock_bh(g_lock); |
| 902 | return (0); |
| 903 | } |
| 904 | |
| 905 | if (peer->ksnp_accepting > 0 || |
| 906 | ksocknal_find_connecting_route_locked (peer) != NULL) { |
| 907 | /* the message is going to be pinned to the peer */ |
| 908 | tx->tx_deadline = |
| 909 | cfs_time_shift(*ksocknal_tunables.ksnd_timeout); |
| 910 | |
| 911 | /* Queue the message until a connection is established */ |
| 912 | list_add_tail (&tx->tx_list, &peer->ksnp_tx_queue); |
| 913 | write_unlock_bh(g_lock); |
| 914 | return 0; |
| 915 | } |
| 916 | |
| 917 | write_unlock_bh(g_lock); |
| 918 | |
| 919 | /* NB Routes may be ignored if connections to them failed recently */ |
| 920 | CNETERR("No usable routes to %s\n", libcfs_id2str(id)); |
| 921 | return (-EHOSTUNREACH); |
| 922 | } |
| 923 | |
| 924 | int |
| 925 | ksocknal_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg) |
| 926 | { |
Oleg Drokin | aadbacc | 2014-01-23 23:45:05 -0500 | [diff] [blame] | 927 | int mpflag = 1; |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 928 | int type = lntmsg->msg_type; |
| 929 | lnet_process_id_t target = lntmsg->msg_target; |
| 930 | unsigned int payload_niov = lntmsg->msg_niov; |
| 931 | struct iovec *payload_iov = lntmsg->msg_iov; |
| 932 | lnet_kiov_t *payload_kiov = lntmsg->msg_kiov; |
| 933 | unsigned int payload_offset = lntmsg->msg_offset; |
| 934 | unsigned int payload_nob = lntmsg->msg_len; |
| 935 | ksock_tx_t *tx; |
| 936 | int desc_size; |
| 937 | int rc; |
| 938 | |
| 939 | /* NB 'private' is different depending on what we're sending. |
| 940 | * Just ignore it... */ |
| 941 | |
| 942 | CDEBUG(D_NET, "sending %u bytes in %d frags to %s\n", |
| 943 | payload_nob, payload_niov, libcfs_id2str(target)); |
| 944 | |
| 945 | LASSERT (payload_nob == 0 || payload_niov > 0); |
| 946 | LASSERT (payload_niov <= LNET_MAX_IOV); |
| 947 | /* payload is either all vaddrs or all pages */ |
| 948 | LASSERT (!(payload_kiov != NULL && payload_iov != NULL)); |
| 949 | LASSERT (!in_interrupt ()); |
| 950 | |
| 951 | if (payload_iov != NULL) |
| 952 | desc_size = offsetof(ksock_tx_t, |
| 953 | tx_frags.virt.iov[1 + payload_niov]); |
| 954 | else |
| 955 | desc_size = offsetof(ksock_tx_t, |
| 956 | tx_frags.paged.kiov[payload_niov]); |
| 957 | |
| 958 | if (lntmsg->msg_vmflush) |
| 959 | mpflag = cfs_memory_pressure_get_and_set(); |
| 960 | tx = ksocknal_alloc_tx(KSOCK_MSG_LNET, desc_size); |
| 961 | if (tx == NULL) { |
| 962 | CERROR("Can't allocate tx desc type %d size %d\n", |
| 963 | type, desc_size); |
| 964 | if (lntmsg->msg_vmflush) |
| 965 | cfs_memory_pressure_restore(mpflag); |
| 966 | return (-ENOMEM); |
| 967 | } |
| 968 | |
| 969 | tx->tx_conn = NULL; /* set when assigned a conn */ |
| 970 | tx->tx_lnetmsg = lntmsg; |
| 971 | |
| 972 | if (payload_iov != NULL) { |
| 973 | tx->tx_kiov = NULL; |
| 974 | tx->tx_nkiov = 0; |
| 975 | tx->tx_iov = tx->tx_frags.virt.iov; |
| 976 | tx->tx_niov = 1 + |
| 977 | lnet_extract_iov(payload_niov, &tx->tx_iov[1], |
| 978 | payload_niov, payload_iov, |
| 979 | payload_offset, payload_nob); |
| 980 | } else { |
| 981 | tx->tx_niov = 1; |
| 982 | tx->tx_iov = &tx->tx_frags.paged.iov; |
| 983 | tx->tx_kiov = tx->tx_frags.paged.kiov; |
| 984 | tx->tx_nkiov = lnet_extract_kiov(payload_niov, tx->tx_kiov, |
| 985 | payload_niov, payload_kiov, |
| 986 | payload_offset, payload_nob); |
| 987 | |
| 988 | if (payload_nob >= *ksocknal_tunables.ksnd_zc_min_payload) |
| 989 | tx->tx_zc_capable = 1; |
| 990 | } |
| 991 | |
| 992 | socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_LNET); |
| 993 | |
| 994 | /* The first fragment will be set later in pro_pack */ |
| 995 | rc = ksocknal_launch_packet(ni, tx, target); |
Oleg Drokin | aadbacc | 2014-01-23 23:45:05 -0500 | [diff] [blame] | 996 | if (!mpflag) |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 997 | cfs_memory_pressure_restore(mpflag); |
Oleg Drokin | aadbacc | 2014-01-23 23:45:05 -0500 | [diff] [blame] | 998 | |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 999 | if (rc == 0) |
| 1000 | return (0); |
| 1001 | |
| 1002 | ksocknal_free_tx(tx); |
| 1003 | return (-EIO); |
| 1004 | } |
| 1005 | |
| 1006 | int |
| 1007 | ksocknal_thread_start(int (*fn)(void *arg), void *arg, char *name) |
| 1008 | { |
Kees Cook | 9edf0f6 | 2013-09-10 21:37:19 -0700 | [diff] [blame] | 1009 | struct task_struct *task = kthread_run(fn, arg, "%s", name); |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 1010 | |
| 1011 | if (IS_ERR(task)) |
| 1012 | return PTR_ERR(task); |
| 1013 | |
| 1014 | write_lock_bh(&ksocknal_data.ksnd_global_lock); |
| 1015 | ksocknal_data.ksnd_nthreads++; |
| 1016 | write_unlock_bh(&ksocknal_data.ksnd_global_lock); |
| 1017 | return 0; |
| 1018 | } |
| 1019 | |
| 1020 | void |
| 1021 | ksocknal_thread_fini (void) |
| 1022 | { |
| 1023 | write_lock_bh(&ksocknal_data.ksnd_global_lock); |
| 1024 | ksocknal_data.ksnd_nthreads--; |
| 1025 | write_unlock_bh(&ksocknal_data.ksnd_global_lock); |
| 1026 | } |
| 1027 | |
| 1028 | int |
| 1029 | ksocknal_new_packet (ksock_conn_t *conn, int nob_to_skip) |
| 1030 | { |
| 1031 | static char ksocknal_slop_buffer[4096]; |
| 1032 | |
| 1033 | int nob; |
| 1034 | unsigned int niov; |
| 1035 | int skipped; |
| 1036 | |
| 1037 | LASSERT(conn->ksnc_proto != NULL); |
| 1038 | |
| 1039 | if ((*ksocknal_tunables.ksnd_eager_ack & conn->ksnc_type) != 0) { |
| 1040 | /* Remind the socket to ack eagerly... */ |
| 1041 | ksocknal_lib_eager_ack(conn); |
| 1042 | } |
| 1043 | |
| 1044 | if (nob_to_skip == 0) { /* right at next packet boundary now */ |
| 1045 | conn->ksnc_rx_started = 0; |
| 1046 | mb(); /* racing with timeout thread */ |
| 1047 | |
| 1048 | switch (conn->ksnc_proto->pro_version) { |
| 1049 | case KSOCK_PROTO_V2: |
| 1050 | case KSOCK_PROTO_V3: |
| 1051 | conn->ksnc_rx_state = SOCKNAL_RX_KSM_HEADER; |
| 1052 | conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space; |
| 1053 | conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg; |
| 1054 | |
| 1055 | conn->ksnc_rx_nob_wanted = offsetof(ksock_msg_t, ksm_u); |
| 1056 | conn->ksnc_rx_nob_left = offsetof(ksock_msg_t, ksm_u); |
| 1057 | conn->ksnc_rx_iov[0].iov_len = offsetof(ksock_msg_t, ksm_u); |
| 1058 | break; |
| 1059 | |
| 1060 | case KSOCK_PROTO_V1: |
| 1061 | /* Receiving bare lnet_hdr_t */ |
| 1062 | conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER; |
| 1063 | conn->ksnc_rx_nob_wanted = sizeof(lnet_hdr_t); |
| 1064 | conn->ksnc_rx_nob_left = sizeof(lnet_hdr_t); |
| 1065 | |
| 1066 | conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space; |
| 1067 | conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg; |
| 1068 | conn->ksnc_rx_iov[0].iov_len = sizeof (lnet_hdr_t); |
| 1069 | break; |
| 1070 | |
| 1071 | default: |
| 1072 | LBUG (); |
| 1073 | } |
| 1074 | conn->ksnc_rx_niov = 1; |
| 1075 | |
| 1076 | conn->ksnc_rx_kiov = NULL; |
| 1077 | conn->ksnc_rx_nkiov = 0; |
| 1078 | conn->ksnc_rx_csum = ~0; |
| 1079 | return (1); |
| 1080 | } |
| 1081 | |
| 1082 | /* Set up to skip as much as possible now. If there's more left |
| 1083 | * (ran out of iov entries) we'll get called again */ |
| 1084 | |
| 1085 | conn->ksnc_rx_state = SOCKNAL_RX_SLOP; |
| 1086 | conn->ksnc_rx_nob_left = nob_to_skip; |
| 1087 | conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space; |
| 1088 | skipped = 0; |
| 1089 | niov = 0; |
| 1090 | |
| 1091 | do { |
| 1092 | nob = MIN (nob_to_skip, sizeof (ksocknal_slop_buffer)); |
| 1093 | |
| 1094 | conn->ksnc_rx_iov[niov].iov_base = ksocknal_slop_buffer; |
| 1095 | conn->ksnc_rx_iov[niov].iov_len = nob; |
| 1096 | niov++; |
| 1097 | skipped += nob; |
| 1098 | nob_to_skip -=nob; |
| 1099 | |
| 1100 | } while (nob_to_skip != 0 && /* mustn't overflow conn's rx iov */ |
| 1101 | niov < sizeof(conn->ksnc_rx_iov_space) / sizeof (struct iovec)); |
| 1102 | |
| 1103 | conn->ksnc_rx_niov = niov; |
| 1104 | conn->ksnc_rx_kiov = NULL; |
| 1105 | conn->ksnc_rx_nkiov = 0; |
| 1106 | conn->ksnc_rx_nob_wanted = skipped; |
| 1107 | return (0); |
| 1108 | } |
| 1109 | |
| 1110 | int |
| 1111 | ksocknal_process_receive (ksock_conn_t *conn) |
| 1112 | { |
| 1113 | lnet_hdr_t *lhdr; |
| 1114 | lnet_process_id_t *id; |
| 1115 | int rc; |
| 1116 | |
| 1117 | LASSERT (atomic_read(&conn->ksnc_conn_refcount) > 0); |
| 1118 | |
| 1119 | /* NB: sched lock NOT held */ |
Masanari Iida | 2b28432 | 2013-08-25 10:10:14 +0900 | [diff] [blame] | 1120 | /* SOCKNAL_RX_LNET_HEADER is here for backward compatibility */ |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 1121 | LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_KSM_HEADER || |
| 1122 | conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD || |
| 1123 | conn->ksnc_rx_state == SOCKNAL_RX_LNET_HEADER || |
| 1124 | conn->ksnc_rx_state == SOCKNAL_RX_SLOP); |
| 1125 | again: |
| 1126 | if (conn->ksnc_rx_nob_wanted != 0) { |
| 1127 | rc = ksocknal_receive(conn); |
| 1128 | |
| 1129 | if (rc <= 0) { |
| 1130 | LASSERT (rc != -EAGAIN); |
| 1131 | |
| 1132 | if (rc == 0) |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 1133 | CDEBUG(D_NET, "[%p] EOF from %s" |
| 1134 | " ip %pI4h:%d\n", conn, |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 1135 | libcfs_id2str(conn->ksnc_peer->ksnp_id), |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 1136 | &conn->ksnc_ipaddr, |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 1137 | conn->ksnc_port); |
| 1138 | else if (!conn->ksnc_closing) |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 1139 | CERROR("[%p] Error %d on read from %s" |
| 1140 | " ip %pI4h:%d\n", |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 1141 | conn, rc, |
| 1142 | libcfs_id2str(conn->ksnc_peer->ksnp_id), |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 1143 | &conn->ksnc_ipaddr, |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 1144 | conn->ksnc_port); |
| 1145 | |
| 1146 | /* it's not an error if conn is being closed */ |
| 1147 | ksocknal_close_conn_and_siblings (conn, |
| 1148 | (conn->ksnc_closing) ? 0 : rc); |
| 1149 | return (rc == 0 ? -ESHUTDOWN : rc); |
| 1150 | } |
| 1151 | |
| 1152 | if (conn->ksnc_rx_nob_wanted != 0) { |
| 1153 | /* short read */ |
| 1154 | return (-EAGAIN); |
| 1155 | } |
| 1156 | } |
| 1157 | switch (conn->ksnc_rx_state) { |
| 1158 | case SOCKNAL_RX_KSM_HEADER: |
| 1159 | if (conn->ksnc_flip) { |
| 1160 | __swab32s(&conn->ksnc_msg.ksm_type); |
| 1161 | __swab32s(&conn->ksnc_msg.ksm_csum); |
| 1162 | __swab64s(&conn->ksnc_msg.ksm_zc_cookies[0]); |
| 1163 | __swab64s(&conn->ksnc_msg.ksm_zc_cookies[1]); |
| 1164 | } |
| 1165 | |
| 1166 | if (conn->ksnc_msg.ksm_type != KSOCK_MSG_NOOP && |
| 1167 | conn->ksnc_msg.ksm_type != KSOCK_MSG_LNET) { |
| 1168 | CERROR("%s: Unknown message type: %x\n", |
| 1169 | libcfs_id2str(conn->ksnc_peer->ksnp_id), |
| 1170 | conn->ksnc_msg.ksm_type); |
| 1171 | ksocknal_new_packet(conn, 0); |
| 1172 | ksocknal_close_conn_and_siblings(conn, -EPROTO); |
| 1173 | return (-EPROTO); |
| 1174 | } |
| 1175 | |
| 1176 | if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP && |
| 1177 | conn->ksnc_msg.ksm_csum != 0 && /* has checksum */ |
| 1178 | conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) { |
| 1179 | /* NOOP Checksum error */ |
| 1180 | CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n", |
| 1181 | libcfs_id2str(conn->ksnc_peer->ksnp_id), |
| 1182 | conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum); |
| 1183 | ksocknal_new_packet(conn, 0); |
| 1184 | ksocknal_close_conn_and_siblings(conn, -EPROTO); |
| 1185 | return (-EIO); |
| 1186 | } |
| 1187 | |
| 1188 | if (conn->ksnc_msg.ksm_zc_cookies[1] != 0) { |
| 1189 | __u64 cookie = 0; |
| 1190 | |
| 1191 | LASSERT (conn->ksnc_proto != &ksocknal_protocol_v1x); |
| 1192 | |
| 1193 | if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP) |
| 1194 | cookie = conn->ksnc_msg.ksm_zc_cookies[0]; |
| 1195 | |
| 1196 | rc = conn->ksnc_proto->pro_handle_zcack(conn, cookie, |
| 1197 | conn->ksnc_msg.ksm_zc_cookies[1]); |
| 1198 | |
| 1199 | if (rc != 0) { |
| 1200 | CERROR("%s: Unknown ZC-ACK cookie: "LPU64", "LPU64"\n", |
| 1201 | libcfs_id2str(conn->ksnc_peer->ksnp_id), |
| 1202 | cookie, conn->ksnc_msg.ksm_zc_cookies[1]); |
| 1203 | ksocknal_new_packet(conn, 0); |
| 1204 | ksocknal_close_conn_and_siblings(conn, -EPROTO); |
| 1205 | return (rc); |
| 1206 | } |
| 1207 | } |
| 1208 | |
| 1209 | if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP) { |
| 1210 | ksocknal_new_packet (conn, 0); |
| 1211 | return 0; /* NOOP is done and just return */ |
| 1212 | } |
| 1213 | |
| 1214 | conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER; |
| 1215 | conn->ksnc_rx_nob_wanted = sizeof(ksock_lnet_msg_t); |
| 1216 | conn->ksnc_rx_nob_left = sizeof(ksock_lnet_msg_t); |
| 1217 | |
| 1218 | conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space; |
| 1219 | conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg; |
| 1220 | conn->ksnc_rx_iov[0].iov_len = sizeof(ksock_lnet_msg_t); |
| 1221 | |
| 1222 | conn->ksnc_rx_niov = 1; |
| 1223 | conn->ksnc_rx_kiov = NULL; |
| 1224 | conn->ksnc_rx_nkiov = 0; |
| 1225 | |
| 1226 | goto again; /* read lnet header now */ |
| 1227 | |
| 1228 | case SOCKNAL_RX_LNET_HEADER: |
| 1229 | /* unpack message header */ |
| 1230 | conn->ksnc_proto->pro_unpack(&conn->ksnc_msg); |
| 1231 | |
| 1232 | if ((conn->ksnc_peer->ksnp_id.pid & LNET_PID_USERFLAG) != 0) { |
| 1233 | /* Userspace peer */ |
| 1234 | lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr; |
| 1235 | id = &conn->ksnc_peer->ksnp_id; |
| 1236 | |
| 1237 | /* Substitute process ID assigned at connection time */ |
| 1238 | lhdr->src_pid = cpu_to_le32(id->pid); |
| 1239 | lhdr->src_nid = cpu_to_le64(id->nid); |
| 1240 | } |
| 1241 | |
| 1242 | conn->ksnc_rx_state = SOCKNAL_RX_PARSE; |
| 1243 | ksocknal_conn_addref(conn); /* ++ref while parsing */ |
| 1244 | |
| 1245 | rc = lnet_parse(conn->ksnc_peer->ksnp_ni, |
| 1246 | &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr, |
| 1247 | conn->ksnc_peer->ksnp_id.nid, conn, 0); |
| 1248 | if (rc < 0) { |
| 1249 | /* I just received garbage: give up on this conn */ |
| 1250 | ksocknal_new_packet(conn, 0); |
| 1251 | ksocknal_close_conn_and_siblings (conn, rc); |
| 1252 | ksocknal_conn_decref(conn); |
| 1253 | return (-EPROTO); |
| 1254 | } |
| 1255 | |
| 1256 | /* I'm racing with ksocknal_recv() */ |
| 1257 | LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_PARSE || |
| 1258 | conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD); |
| 1259 | |
| 1260 | if (conn->ksnc_rx_state != SOCKNAL_RX_LNET_PAYLOAD) |
| 1261 | return 0; |
| 1262 | |
| 1263 | /* ksocknal_recv() got called */ |
| 1264 | goto again; |
| 1265 | |
| 1266 | case SOCKNAL_RX_LNET_PAYLOAD: |
| 1267 | /* payload all received */ |
| 1268 | rc = 0; |
| 1269 | |
| 1270 | if (conn->ksnc_rx_nob_left == 0 && /* not truncating */ |
| 1271 | conn->ksnc_msg.ksm_csum != 0 && /* has checksum */ |
| 1272 | conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) { |
| 1273 | CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n", |
| 1274 | libcfs_id2str(conn->ksnc_peer->ksnp_id), |
| 1275 | conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum); |
| 1276 | rc = -EIO; |
| 1277 | } |
| 1278 | |
| 1279 | if (rc == 0 && conn->ksnc_msg.ksm_zc_cookies[0] != 0) { |
| 1280 | LASSERT(conn->ksnc_proto != &ksocknal_protocol_v1x); |
| 1281 | |
| 1282 | lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr; |
| 1283 | id = &conn->ksnc_peer->ksnp_id; |
| 1284 | |
| 1285 | rc = conn->ksnc_proto->pro_handle_zcreq(conn, |
| 1286 | conn->ksnc_msg.ksm_zc_cookies[0], |
| 1287 | *ksocknal_tunables.ksnd_nonblk_zcack || |
| 1288 | le64_to_cpu(lhdr->src_nid) != id->nid); |
| 1289 | } |
| 1290 | |
| 1291 | lnet_finalize(conn->ksnc_peer->ksnp_ni, conn->ksnc_cookie, rc); |
| 1292 | |
| 1293 | if (rc != 0) { |
| 1294 | ksocknal_new_packet(conn, 0); |
| 1295 | ksocknal_close_conn_and_siblings (conn, rc); |
| 1296 | return (-EPROTO); |
| 1297 | } |
| 1298 | /* Fall through */ |
| 1299 | |
| 1300 | case SOCKNAL_RX_SLOP: |
| 1301 | /* starting new packet? */ |
| 1302 | if (ksocknal_new_packet (conn, conn->ksnc_rx_nob_left)) |
| 1303 | return 0; /* come back later */ |
| 1304 | goto again; /* try to finish reading slop now */ |
| 1305 | |
| 1306 | default: |
| 1307 | break; |
| 1308 | } |
| 1309 | |
| 1310 | /* Not Reached */ |
| 1311 | LBUG (); |
| 1312 | return (-EINVAL); /* keep gcc happy */ |
| 1313 | } |
| 1314 | |
| 1315 | int |
| 1316 | ksocknal_recv (lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed, |
| 1317 | unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov, |
| 1318 | unsigned int offset, unsigned int mlen, unsigned int rlen) |
| 1319 | { |
| 1320 | ksock_conn_t *conn = (ksock_conn_t *)private; |
| 1321 | ksock_sched_t *sched = conn->ksnc_scheduler; |
| 1322 | |
| 1323 | LASSERT (mlen <= rlen); |
| 1324 | LASSERT (niov <= LNET_MAX_IOV); |
| 1325 | |
| 1326 | conn->ksnc_cookie = msg; |
| 1327 | conn->ksnc_rx_nob_wanted = mlen; |
| 1328 | conn->ksnc_rx_nob_left = rlen; |
| 1329 | |
| 1330 | if (mlen == 0 || iov != NULL) { |
| 1331 | conn->ksnc_rx_nkiov = 0; |
| 1332 | conn->ksnc_rx_kiov = NULL; |
| 1333 | conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov; |
| 1334 | conn->ksnc_rx_niov = |
| 1335 | lnet_extract_iov(LNET_MAX_IOV, conn->ksnc_rx_iov, |
| 1336 | niov, iov, offset, mlen); |
| 1337 | } else { |
| 1338 | conn->ksnc_rx_niov = 0; |
| 1339 | conn->ksnc_rx_iov = NULL; |
| 1340 | conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov; |
| 1341 | conn->ksnc_rx_nkiov = |
| 1342 | lnet_extract_kiov(LNET_MAX_IOV, conn->ksnc_rx_kiov, |
| 1343 | niov, kiov, offset, mlen); |
| 1344 | } |
| 1345 | |
| 1346 | LASSERT (mlen == |
| 1347 | lnet_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) + |
| 1348 | lnet_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov)); |
| 1349 | |
| 1350 | LASSERT (conn->ksnc_rx_scheduled); |
| 1351 | |
| 1352 | spin_lock_bh(&sched->kss_lock); |
| 1353 | |
| 1354 | switch (conn->ksnc_rx_state) { |
| 1355 | case SOCKNAL_RX_PARSE_WAIT: |
| 1356 | list_add_tail(&conn->ksnc_rx_list, &sched->kss_rx_conns); |
| 1357 | wake_up (&sched->kss_waitq); |
| 1358 | LASSERT (conn->ksnc_rx_ready); |
| 1359 | break; |
| 1360 | |
| 1361 | case SOCKNAL_RX_PARSE: |
| 1362 | /* scheduler hasn't noticed I'm parsing yet */ |
| 1363 | break; |
| 1364 | } |
| 1365 | |
| 1366 | conn->ksnc_rx_state = SOCKNAL_RX_LNET_PAYLOAD; |
| 1367 | |
| 1368 | spin_unlock_bh(&sched->kss_lock); |
| 1369 | ksocknal_conn_decref(conn); |
| 1370 | return 0; |
| 1371 | } |
| 1372 | |
| 1373 | static inline int |
| 1374 | ksocknal_sched_cansleep(ksock_sched_t *sched) |
| 1375 | { |
| 1376 | int rc; |
| 1377 | |
| 1378 | spin_lock_bh(&sched->kss_lock); |
| 1379 | |
| 1380 | rc = (!ksocknal_data.ksnd_shuttingdown && |
| 1381 | list_empty(&sched->kss_rx_conns) && |
| 1382 | list_empty(&sched->kss_tx_conns)); |
| 1383 | |
| 1384 | spin_unlock_bh(&sched->kss_lock); |
| 1385 | return rc; |
| 1386 | } |
| 1387 | |
| 1388 | int ksocknal_scheduler(void *arg) |
| 1389 | { |
| 1390 | struct ksock_sched_info *info; |
| 1391 | ksock_sched_t *sched; |
| 1392 | ksock_conn_t *conn; |
| 1393 | ksock_tx_t *tx; |
| 1394 | int rc; |
| 1395 | int nloops = 0; |
| 1396 | long id = (long)arg; |
| 1397 | |
| 1398 | info = ksocknal_data.ksnd_sched_info[KSOCK_THREAD_CPT(id)]; |
| 1399 | sched = &info->ksi_scheds[KSOCK_THREAD_SID(id)]; |
| 1400 | |
| 1401 | cfs_block_allsigs(); |
| 1402 | |
| 1403 | rc = cfs_cpt_bind(lnet_cpt_table(), info->ksi_cpt); |
| 1404 | if (rc != 0) { |
| 1405 | CERROR("Can't set CPT affinity to %d: %d\n", |
| 1406 | info->ksi_cpt, rc); |
| 1407 | } |
| 1408 | |
| 1409 | spin_lock_bh(&sched->kss_lock); |
| 1410 | |
| 1411 | while (!ksocknal_data.ksnd_shuttingdown) { |
| 1412 | int did_something = 0; |
| 1413 | |
| 1414 | /* Ensure I progress everything semi-fairly */ |
| 1415 | |
| 1416 | if (!list_empty (&sched->kss_rx_conns)) { |
| 1417 | conn = list_entry(sched->kss_rx_conns.next, |
| 1418 | ksock_conn_t, ksnc_rx_list); |
| 1419 | list_del(&conn->ksnc_rx_list); |
| 1420 | |
| 1421 | LASSERT(conn->ksnc_rx_scheduled); |
| 1422 | LASSERT(conn->ksnc_rx_ready); |
| 1423 | |
| 1424 | /* clear rx_ready in case receive isn't complete. |
| 1425 | * Do it BEFORE we call process_recv, since |
| 1426 | * data_ready can set it any time after we release |
| 1427 | * kss_lock. */ |
| 1428 | conn->ksnc_rx_ready = 0; |
| 1429 | spin_unlock_bh(&sched->kss_lock); |
| 1430 | |
| 1431 | rc = ksocknal_process_receive(conn); |
| 1432 | |
| 1433 | spin_lock_bh(&sched->kss_lock); |
| 1434 | |
| 1435 | /* I'm the only one that can clear this flag */ |
| 1436 | LASSERT(conn->ksnc_rx_scheduled); |
| 1437 | |
| 1438 | /* Did process_receive get everything it wanted? */ |
| 1439 | if (rc == 0) |
| 1440 | conn->ksnc_rx_ready = 1; |
| 1441 | |
| 1442 | if (conn->ksnc_rx_state == SOCKNAL_RX_PARSE) { |
| 1443 | /* Conn blocked waiting for ksocknal_recv() |
| 1444 | * I change its state (under lock) to signal |
| 1445 | * it can be rescheduled */ |
| 1446 | conn->ksnc_rx_state = SOCKNAL_RX_PARSE_WAIT; |
| 1447 | } else if (conn->ksnc_rx_ready) { |
| 1448 | /* reschedule for rx */ |
| 1449 | list_add_tail (&conn->ksnc_rx_list, |
| 1450 | &sched->kss_rx_conns); |
| 1451 | } else { |
| 1452 | conn->ksnc_rx_scheduled = 0; |
| 1453 | /* drop my ref */ |
| 1454 | ksocknal_conn_decref(conn); |
| 1455 | } |
| 1456 | |
| 1457 | did_something = 1; |
| 1458 | } |
| 1459 | |
| 1460 | if (!list_empty (&sched->kss_tx_conns)) { |
| 1461 | LIST_HEAD (zlist); |
| 1462 | |
| 1463 | if (!list_empty(&sched->kss_zombie_noop_txs)) { |
| 1464 | list_add(&zlist, |
| 1465 | &sched->kss_zombie_noop_txs); |
| 1466 | list_del_init(&sched->kss_zombie_noop_txs); |
| 1467 | } |
| 1468 | |
| 1469 | conn = list_entry(sched->kss_tx_conns.next, |
| 1470 | ksock_conn_t, ksnc_tx_list); |
| 1471 | list_del (&conn->ksnc_tx_list); |
| 1472 | |
| 1473 | LASSERT(conn->ksnc_tx_scheduled); |
| 1474 | LASSERT(conn->ksnc_tx_ready); |
| 1475 | LASSERT(!list_empty(&conn->ksnc_tx_queue)); |
| 1476 | |
| 1477 | tx = list_entry(conn->ksnc_tx_queue.next, |
| 1478 | ksock_tx_t, tx_list); |
| 1479 | |
| 1480 | if (conn->ksnc_tx_carrier == tx) |
| 1481 | ksocknal_next_tx_carrier(conn); |
| 1482 | |
| 1483 | /* dequeue now so empty list => more to send */ |
| 1484 | list_del(&tx->tx_list); |
| 1485 | |
| 1486 | /* Clear tx_ready in case send isn't complete. Do |
| 1487 | * it BEFORE we call process_transmit, since |
| 1488 | * write_space can set it any time after we release |
| 1489 | * kss_lock. */ |
| 1490 | conn->ksnc_tx_ready = 0; |
| 1491 | spin_unlock_bh(&sched->kss_lock); |
| 1492 | |
| 1493 | if (!list_empty(&zlist)) { |
| 1494 | /* free zombie noop txs, it's fast because |
| 1495 | * noop txs are just put in freelist */ |
| 1496 | ksocknal_txlist_done(NULL, &zlist, 0); |
| 1497 | } |
| 1498 | |
| 1499 | rc = ksocknal_process_transmit(conn, tx); |
| 1500 | |
| 1501 | if (rc == -ENOMEM || rc == -EAGAIN) { |
| 1502 | /* Incomplete send: replace tx on HEAD of tx_queue */ |
| 1503 | spin_lock_bh(&sched->kss_lock); |
| 1504 | list_add(&tx->tx_list, |
| 1505 | &conn->ksnc_tx_queue); |
| 1506 | } else { |
| 1507 | /* Complete send; tx -ref */ |
| 1508 | ksocknal_tx_decref(tx); |
| 1509 | |
| 1510 | spin_lock_bh(&sched->kss_lock); |
| 1511 | /* assume space for more */ |
| 1512 | conn->ksnc_tx_ready = 1; |
| 1513 | } |
| 1514 | |
| 1515 | if (rc == -ENOMEM) { |
| 1516 | /* Do nothing; after a short timeout, this |
| 1517 | * conn will be reposted on kss_tx_conns. */ |
| 1518 | } else if (conn->ksnc_tx_ready && |
| 1519 | !list_empty (&conn->ksnc_tx_queue)) { |
| 1520 | /* reschedule for tx */ |
| 1521 | list_add_tail (&conn->ksnc_tx_list, |
| 1522 | &sched->kss_tx_conns); |
| 1523 | } else { |
| 1524 | conn->ksnc_tx_scheduled = 0; |
| 1525 | /* drop my ref */ |
| 1526 | ksocknal_conn_decref(conn); |
| 1527 | } |
| 1528 | |
| 1529 | did_something = 1; |
| 1530 | } |
| 1531 | if (!did_something || /* nothing to do */ |
| 1532 | ++nloops == SOCKNAL_RESCHED) { /* hogging CPU? */ |
| 1533 | spin_unlock_bh(&sched->kss_lock); |
| 1534 | |
| 1535 | nloops = 0; |
| 1536 | |
| 1537 | if (!did_something) { /* wait for something to do */ |
| 1538 | cfs_wait_event_interruptible_exclusive( |
| 1539 | sched->kss_waitq, |
| 1540 | !ksocknal_sched_cansleep(sched), rc); |
| 1541 | LASSERT (rc == 0); |
| 1542 | } else { |
| 1543 | cond_resched(); |
| 1544 | } |
| 1545 | |
| 1546 | spin_lock_bh(&sched->kss_lock); |
| 1547 | } |
| 1548 | } |
| 1549 | |
| 1550 | spin_unlock_bh(&sched->kss_lock); |
| 1551 | ksocknal_thread_fini(); |
| 1552 | return 0; |
| 1553 | } |
| 1554 | |
| 1555 | /* |
| 1556 | * Add connection to kss_rx_conns of scheduler |
| 1557 | * and wakeup the scheduler. |
| 1558 | */ |
| 1559 | void ksocknal_read_callback (ksock_conn_t *conn) |
| 1560 | { |
| 1561 | ksock_sched_t *sched; |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 1562 | |
| 1563 | sched = conn->ksnc_scheduler; |
| 1564 | |
| 1565 | spin_lock_bh(&sched->kss_lock); |
| 1566 | |
| 1567 | conn->ksnc_rx_ready = 1; |
| 1568 | |
| 1569 | if (!conn->ksnc_rx_scheduled) { /* not being progressed */ |
| 1570 | list_add_tail(&conn->ksnc_rx_list, |
| 1571 | &sched->kss_rx_conns); |
| 1572 | conn->ksnc_rx_scheduled = 1; |
| 1573 | /* extra ref for scheduler */ |
| 1574 | ksocknal_conn_addref(conn); |
| 1575 | |
| 1576 | wake_up (&sched->kss_waitq); |
| 1577 | } |
| 1578 | spin_unlock_bh(&sched->kss_lock); |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 1579 | } |
| 1580 | |
| 1581 | /* |
| 1582 | * Add connection to kss_tx_conns of scheduler |
| 1583 | * and wakeup the scheduler. |
| 1584 | */ |
| 1585 | void ksocknal_write_callback (ksock_conn_t *conn) |
| 1586 | { |
| 1587 | ksock_sched_t *sched; |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 1588 | |
| 1589 | sched = conn->ksnc_scheduler; |
| 1590 | |
| 1591 | spin_lock_bh(&sched->kss_lock); |
| 1592 | |
| 1593 | conn->ksnc_tx_ready = 1; |
| 1594 | |
| 1595 | if (!conn->ksnc_tx_scheduled && // not being progressed |
| 1596 | !list_empty(&conn->ksnc_tx_queue)){//packets to send |
| 1597 | list_add_tail (&conn->ksnc_tx_list, |
| 1598 | &sched->kss_tx_conns); |
| 1599 | conn->ksnc_tx_scheduled = 1; |
| 1600 | /* extra ref for scheduler */ |
| 1601 | ksocknal_conn_addref(conn); |
| 1602 | |
| 1603 | wake_up (&sched->kss_waitq); |
| 1604 | } |
| 1605 | |
| 1606 | spin_unlock_bh(&sched->kss_lock); |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 1607 | } |
| 1608 | |
| 1609 | ksock_proto_t * |
| 1610 | ksocknal_parse_proto_version (ksock_hello_msg_t *hello) |
| 1611 | { |
| 1612 | __u32 version = 0; |
| 1613 | |
| 1614 | if (hello->kshm_magic == LNET_PROTO_MAGIC) |
| 1615 | version = hello->kshm_version; |
| 1616 | else if (hello->kshm_magic == __swab32(LNET_PROTO_MAGIC)) |
| 1617 | version = __swab32(hello->kshm_version); |
| 1618 | |
| 1619 | if (version != 0) { |
| 1620 | #if SOCKNAL_VERSION_DEBUG |
| 1621 | if (*ksocknal_tunables.ksnd_protocol == 1) |
| 1622 | return NULL; |
| 1623 | |
| 1624 | if (*ksocknal_tunables.ksnd_protocol == 2 && |
| 1625 | version == KSOCK_PROTO_V3) |
| 1626 | return NULL; |
| 1627 | #endif |
| 1628 | if (version == KSOCK_PROTO_V2) |
| 1629 | return &ksocknal_protocol_v2x; |
| 1630 | |
| 1631 | if (version == KSOCK_PROTO_V3) |
| 1632 | return &ksocknal_protocol_v3x; |
| 1633 | |
| 1634 | return NULL; |
| 1635 | } |
| 1636 | |
| 1637 | if (hello->kshm_magic == le32_to_cpu(LNET_PROTO_TCP_MAGIC)) { |
| 1638 | lnet_magicversion_t *hmv = (lnet_magicversion_t *)hello; |
| 1639 | |
| 1640 | CLASSERT (sizeof (lnet_magicversion_t) == |
| 1641 | offsetof (ksock_hello_msg_t, kshm_src_nid)); |
| 1642 | |
| 1643 | if (hmv->version_major == cpu_to_le16 (KSOCK_PROTO_V1_MAJOR) && |
| 1644 | hmv->version_minor == cpu_to_le16 (KSOCK_PROTO_V1_MINOR)) |
| 1645 | return &ksocknal_protocol_v1x; |
| 1646 | } |
| 1647 | |
| 1648 | return NULL; |
| 1649 | } |
| 1650 | |
| 1651 | int |
| 1652 | ksocknal_send_hello (lnet_ni_t *ni, ksock_conn_t *conn, |
| 1653 | lnet_nid_t peer_nid, ksock_hello_msg_t *hello) |
| 1654 | { |
| 1655 | /* CAVEAT EMPTOR: this byte flips 'ipaddrs' */ |
| 1656 | ksock_net_t *net = (ksock_net_t *)ni->ni_data; |
| 1657 | |
| 1658 | LASSERT (hello->kshm_nips <= LNET_MAX_INTERFACES); |
| 1659 | |
| 1660 | /* rely on caller to hold a ref on socket so it wouldn't disappear */ |
| 1661 | LASSERT (conn->ksnc_proto != NULL); |
| 1662 | |
| 1663 | hello->kshm_src_nid = ni->ni_nid; |
| 1664 | hello->kshm_dst_nid = peer_nid; |
| 1665 | hello->kshm_src_pid = the_lnet.ln_pid; |
| 1666 | |
| 1667 | hello->kshm_src_incarnation = net->ksnn_incarnation; |
| 1668 | hello->kshm_ctype = conn->ksnc_type; |
| 1669 | |
| 1670 | return conn->ksnc_proto->pro_send_hello(conn, hello); |
| 1671 | } |
| 1672 | |
| 1673 | int |
| 1674 | ksocknal_invert_type(int type) |
| 1675 | { |
| 1676 | switch (type) |
| 1677 | { |
| 1678 | case SOCKLND_CONN_ANY: |
| 1679 | case SOCKLND_CONN_CONTROL: |
| 1680 | return (type); |
| 1681 | case SOCKLND_CONN_BULK_IN: |
| 1682 | return SOCKLND_CONN_BULK_OUT; |
| 1683 | case SOCKLND_CONN_BULK_OUT: |
| 1684 | return SOCKLND_CONN_BULK_IN; |
| 1685 | default: |
| 1686 | return (SOCKLND_CONN_NONE); |
| 1687 | } |
| 1688 | } |
| 1689 | |
| 1690 | int |
| 1691 | ksocknal_recv_hello (lnet_ni_t *ni, ksock_conn_t *conn, |
| 1692 | ksock_hello_msg_t *hello, lnet_process_id_t *peerid, |
| 1693 | __u64 *incarnation) |
| 1694 | { |
| 1695 | /* Return < 0 fatal error |
| 1696 | * 0 success |
| 1697 | * EALREADY lost connection race |
| 1698 | * EPROTO protocol version mismatch |
| 1699 | */ |
| 1700 | socket_t *sock = conn->ksnc_sock; |
| 1701 | int active = (conn->ksnc_proto != NULL); |
| 1702 | int timeout; |
| 1703 | int proto_match; |
| 1704 | int rc; |
| 1705 | ksock_proto_t *proto; |
| 1706 | lnet_process_id_t recv_id; |
| 1707 | |
| 1708 | /* socket type set on active connections - not set on passive */ |
| 1709 | LASSERT (!active == !(conn->ksnc_type != SOCKLND_CONN_NONE)); |
| 1710 | |
| 1711 | timeout = active ? *ksocknal_tunables.ksnd_timeout : |
| 1712 | lnet_acceptor_timeout(); |
| 1713 | |
| 1714 | rc = libcfs_sock_read(sock, &hello->kshm_magic, sizeof (hello->kshm_magic), timeout); |
| 1715 | if (rc != 0) { |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 1716 | CERROR("Error %d reading HELLO from %pI4h\n", |
| 1717 | rc, &conn->ksnc_ipaddr); |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 1718 | LASSERT (rc < 0); |
| 1719 | return rc; |
| 1720 | } |
| 1721 | |
| 1722 | if (hello->kshm_magic != LNET_PROTO_MAGIC && |
| 1723 | hello->kshm_magic != __swab32(LNET_PROTO_MAGIC) && |
| 1724 | hello->kshm_magic != le32_to_cpu (LNET_PROTO_TCP_MAGIC)) { |
| 1725 | /* Unexpected magic! */ |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 1726 | CERROR("Bad magic(1) %#08x (%#08x expected) from " |
| 1727 | "%pI4h\n", __cpu_to_le32 (hello->kshm_magic), |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 1728 | LNET_PROTO_TCP_MAGIC, |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 1729 | &conn->ksnc_ipaddr); |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 1730 | return -EPROTO; |
| 1731 | } |
| 1732 | |
| 1733 | rc = libcfs_sock_read(sock, &hello->kshm_version, |
| 1734 | sizeof(hello->kshm_version), timeout); |
| 1735 | if (rc != 0) { |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 1736 | CERROR("Error %d reading HELLO from %pI4h\n", |
| 1737 | rc, &conn->ksnc_ipaddr); |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 1738 | LASSERT (rc < 0); |
| 1739 | return rc; |
| 1740 | } |
| 1741 | |
| 1742 | proto = ksocknal_parse_proto_version(hello); |
| 1743 | if (proto == NULL) { |
| 1744 | if (!active) { |
| 1745 | /* unknown protocol from peer, tell peer my protocol */ |
| 1746 | conn->ksnc_proto = &ksocknal_protocol_v3x; |
| 1747 | #if SOCKNAL_VERSION_DEBUG |
| 1748 | if (*ksocknal_tunables.ksnd_protocol == 2) |
| 1749 | conn->ksnc_proto = &ksocknal_protocol_v2x; |
| 1750 | else if (*ksocknal_tunables.ksnd_protocol == 1) |
| 1751 | conn->ksnc_proto = &ksocknal_protocol_v1x; |
| 1752 | #endif |
| 1753 | hello->kshm_nips = 0; |
| 1754 | ksocknal_send_hello(ni, conn, ni->ni_nid, hello); |
| 1755 | } |
| 1756 | |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 1757 | CERROR("Unknown protocol version (%d.x expected)" |
| 1758 | " from %pI4h\n", |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 1759 | conn->ksnc_proto->pro_version, |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 1760 | &conn->ksnc_ipaddr); |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 1761 | |
| 1762 | return -EPROTO; |
| 1763 | } |
| 1764 | |
| 1765 | proto_match = (conn->ksnc_proto == proto); |
| 1766 | conn->ksnc_proto = proto; |
| 1767 | |
| 1768 | /* receive the rest of hello message anyway */ |
| 1769 | rc = conn->ksnc_proto->pro_recv_hello(conn, hello, timeout); |
| 1770 | if (rc != 0) { |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 1771 | CERROR("Error %d reading or checking hello from from %pI4h\n", |
| 1772 | rc, &conn->ksnc_ipaddr); |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 1773 | LASSERT (rc < 0); |
| 1774 | return rc; |
| 1775 | } |
| 1776 | |
| 1777 | *incarnation = hello->kshm_src_incarnation; |
| 1778 | |
| 1779 | if (hello->kshm_src_nid == LNET_NID_ANY) { |
| 1780 | CERROR("Expecting a HELLO hdr with a NID, but got LNET_NID_ANY" |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 1781 | "from %pI4h\n", &conn->ksnc_ipaddr); |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 1782 | return -EPROTO; |
| 1783 | } |
| 1784 | |
| 1785 | if (!active && |
| 1786 | conn->ksnc_port > LNET_ACCEPTOR_MAX_RESERVED_PORT) { |
| 1787 | /* Userspace NAL assigns peer process ID from socket */ |
| 1788 | recv_id.pid = conn->ksnc_port | LNET_PID_USERFLAG; |
| 1789 | recv_id.nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), conn->ksnc_ipaddr); |
| 1790 | } else { |
| 1791 | recv_id.nid = hello->kshm_src_nid; |
| 1792 | recv_id.pid = hello->kshm_src_pid; |
| 1793 | } |
| 1794 | |
| 1795 | if (!active) { |
| 1796 | *peerid = recv_id; |
| 1797 | |
| 1798 | /* peer determines type */ |
| 1799 | conn->ksnc_type = ksocknal_invert_type(hello->kshm_ctype); |
| 1800 | if (conn->ksnc_type == SOCKLND_CONN_NONE) { |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 1801 | CERROR("Unexpected type %d from %s ip %pI4h\n", |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 1802 | hello->kshm_ctype, libcfs_id2str(*peerid), |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 1803 | &conn->ksnc_ipaddr); |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 1804 | return -EPROTO; |
| 1805 | } |
| 1806 | |
| 1807 | return 0; |
| 1808 | } |
| 1809 | |
| 1810 | if (peerid->pid != recv_id.pid || |
| 1811 | peerid->nid != recv_id.nid) { |
| 1812 | LCONSOLE_ERROR_MSG(0x130, "Connected successfully to %s on host" |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 1813 | " %pI4h, but they claimed they were " |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 1814 | "%s; please check your Lustre " |
| 1815 | "configuration.\n", |
| 1816 | libcfs_id2str(*peerid), |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 1817 | &conn->ksnc_ipaddr, |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 1818 | libcfs_id2str(recv_id)); |
| 1819 | return -EPROTO; |
| 1820 | } |
| 1821 | |
| 1822 | if (hello->kshm_ctype == SOCKLND_CONN_NONE) { |
| 1823 | /* Possible protocol mismatch or I lost the connection race */ |
| 1824 | return proto_match ? EALREADY : EPROTO; |
| 1825 | } |
| 1826 | |
| 1827 | if (ksocknal_invert_type(hello->kshm_ctype) != conn->ksnc_type) { |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 1828 | CERROR("Mismatched types: me %d, %s ip %pI4h %d\n", |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 1829 | conn->ksnc_type, libcfs_id2str(*peerid), |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 1830 | &conn->ksnc_ipaddr, |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 1831 | hello->kshm_ctype); |
| 1832 | return -EPROTO; |
| 1833 | } |
| 1834 | |
| 1835 | return 0; |
| 1836 | } |
| 1837 | |
| 1838 | int |
| 1839 | ksocknal_connect (ksock_route_t *route) |
| 1840 | { |
| 1841 | LIST_HEAD (zombies); |
| 1842 | ksock_peer_t *peer = route->ksnr_peer; |
| 1843 | int type; |
| 1844 | int wanted; |
| 1845 | socket_t *sock; |
| 1846 | cfs_time_t deadline; |
| 1847 | int retry_later = 0; |
| 1848 | int rc = 0; |
| 1849 | |
| 1850 | deadline = cfs_time_add(cfs_time_current(), |
| 1851 | cfs_time_seconds(*ksocknal_tunables.ksnd_timeout)); |
| 1852 | |
| 1853 | write_lock_bh(&ksocknal_data.ksnd_global_lock); |
| 1854 | |
| 1855 | LASSERT (route->ksnr_scheduled); |
| 1856 | LASSERT (!route->ksnr_connecting); |
| 1857 | |
| 1858 | route->ksnr_connecting = 1; |
| 1859 | |
| 1860 | for (;;) { |
| 1861 | wanted = ksocknal_route_mask() & ~route->ksnr_connected; |
| 1862 | |
| 1863 | /* stop connecting if peer/route got closed under me, or |
| 1864 | * route got connected while queued */ |
| 1865 | if (peer->ksnp_closing || route->ksnr_deleted || |
| 1866 | wanted == 0) { |
| 1867 | retry_later = 0; |
| 1868 | break; |
| 1869 | } |
| 1870 | |
| 1871 | /* reschedule if peer is connecting to me */ |
| 1872 | if (peer->ksnp_accepting > 0) { |
| 1873 | CDEBUG(D_NET, |
| 1874 | "peer %s(%d) already connecting to me, retry later.\n", |
| 1875 | libcfs_nid2str(peer->ksnp_id.nid), peer->ksnp_accepting); |
| 1876 | retry_later = 1; |
| 1877 | } |
| 1878 | |
| 1879 | if (retry_later) /* needs reschedule */ |
| 1880 | break; |
| 1881 | |
| 1882 | if ((wanted & (1 << SOCKLND_CONN_ANY)) != 0) { |
| 1883 | type = SOCKLND_CONN_ANY; |
| 1884 | } else if ((wanted & (1 << SOCKLND_CONN_CONTROL)) != 0) { |
| 1885 | type = SOCKLND_CONN_CONTROL; |
| 1886 | } else if ((wanted & (1 << SOCKLND_CONN_BULK_IN)) != 0) { |
| 1887 | type = SOCKLND_CONN_BULK_IN; |
| 1888 | } else { |
| 1889 | LASSERT ((wanted & (1 << SOCKLND_CONN_BULK_OUT)) != 0); |
| 1890 | type = SOCKLND_CONN_BULK_OUT; |
| 1891 | } |
| 1892 | |
| 1893 | write_unlock_bh(&ksocknal_data.ksnd_global_lock); |
| 1894 | |
| 1895 | if (cfs_time_aftereq(cfs_time_current(), deadline)) { |
| 1896 | rc = -ETIMEDOUT; |
| 1897 | lnet_connect_console_error(rc, peer->ksnp_id.nid, |
| 1898 | route->ksnr_ipaddr, |
| 1899 | route->ksnr_port); |
| 1900 | goto failed; |
| 1901 | } |
| 1902 | |
| 1903 | rc = lnet_connect(&sock, peer->ksnp_id.nid, |
| 1904 | route->ksnr_myipaddr, |
| 1905 | route->ksnr_ipaddr, route->ksnr_port); |
| 1906 | if (rc != 0) |
| 1907 | goto failed; |
| 1908 | |
| 1909 | rc = ksocknal_create_conn(peer->ksnp_ni, route, sock, type); |
| 1910 | if (rc < 0) { |
| 1911 | lnet_connect_console_error(rc, peer->ksnp_id.nid, |
| 1912 | route->ksnr_ipaddr, |
| 1913 | route->ksnr_port); |
| 1914 | goto failed; |
| 1915 | } |
| 1916 | |
| 1917 | /* A +ve RC means I have to retry because I lost the connection |
| 1918 | * race or I have to renegotiate protocol version */ |
| 1919 | retry_later = (rc != 0); |
| 1920 | if (retry_later) |
| 1921 | CDEBUG(D_NET, "peer %s: conn race, retry later.\n", |
| 1922 | libcfs_nid2str(peer->ksnp_id.nid)); |
| 1923 | |
| 1924 | write_lock_bh(&ksocknal_data.ksnd_global_lock); |
| 1925 | } |
| 1926 | |
| 1927 | route->ksnr_scheduled = 0; |
| 1928 | route->ksnr_connecting = 0; |
| 1929 | |
| 1930 | if (retry_later) { |
| 1931 | /* re-queue for attention; this frees me up to handle |
| 1932 | * the peer's incoming connection request */ |
| 1933 | |
| 1934 | if (rc == EALREADY || |
| 1935 | (rc == 0 && peer->ksnp_accepting > 0)) { |
| 1936 | /* We want to introduce a delay before next |
| 1937 | * attempt to connect if we lost conn race, |
| 1938 | * but the race is resolved quickly usually, |
| 1939 | * so min_reconnectms should be good heuristic */ |
| 1940 | route->ksnr_retry_interval = |
| 1941 | cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms)/1000; |
| 1942 | route->ksnr_timeout = cfs_time_add(cfs_time_current(), |
| 1943 | route->ksnr_retry_interval); |
| 1944 | } |
| 1945 | |
| 1946 | ksocknal_launch_connection_locked(route); |
| 1947 | } |
| 1948 | |
| 1949 | write_unlock_bh(&ksocknal_data.ksnd_global_lock); |
| 1950 | return retry_later; |
| 1951 | |
| 1952 | failed: |
| 1953 | write_lock_bh(&ksocknal_data.ksnd_global_lock); |
| 1954 | |
| 1955 | route->ksnr_scheduled = 0; |
| 1956 | route->ksnr_connecting = 0; |
| 1957 | |
| 1958 | /* This is a retry rather than a new connection */ |
| 1959 | route->ksnr_retry_interval *= 2; |
| 1960 | route->ksnr_retry_interval = |
| 1961 | MAX(route->ksnr_retry_interval, |
| 1962 | cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms)/1000); |
| 1963 | route->ksnr_retry_interval = |
| 1964 | MIN(route->ksnr_retry_interval, |
| 1965 | cfs_time_seconds(*ksocknal_tunables.ksnd_max_reconnectms)/1000); |
| 1966 | |
| 1967 | LASSERT (route->ksnr_retry_interval != 0); |
| 1968 | route->ksnr_timeout = cfs_time_add(cfs_time_current(), |
| 1969 | route->ksnr_retry_interval); |
| 1970 | |
| 1971 | if (!list_empty(&peer->ksnp_tx_queue) && |
| 1972 | peer->ksnp_accepting == 0 && |
| 1973 | ksocknal_find_connecting_route_locked(peer) == NULL) { |
| 1974 | ksock_conn_t *conn; |
| 1975 | |
| 1976 | /* ksnp_tx_queue is queued on a conn on successful |
| 1977 | * connection for V1.x and V2.x */ |
| 1978 | if (!list_empty (&peer->ksnp_conns)) { |
| 1979 | conn = list_entry(peer->ksnp_conns.next, |
| 1980 | ksock_conn_t, ksnc_list); |
| 1981 | LASSERT (conn->ksnc_proto == &ksocknal_protocol_v3x); |
| 1982 | } |
| 1983 | |
| 1984 | /* take all the blocked packets while I've got the lock and |
| 1985 | * complete below... */ |
| 1986 | list_splice_init(&peer->ksnp_tx_queue, &zombies); |
| 1987 | } |
| 1988 | |
Masanari Iida | 2b28432 | 2013-08-25 10:10:14 +0900 | [diff] [blame] | 1989 | #if 0 /* irrelevant with only eager routes */ |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 1990 | if (!route->ksnr_deleted) { |
| 1991 | /* make this route least-favourite for re-selection */ |
| 1992 | list_del(&route->ksnr_list); |
| 1993 | list_add_tail(&route->ksnr_list, &peer->ksnp_routes); |
| 1994 | } |
| 1995 | #endif |
| 1996 | write_unlock_bh(&ksocknal_data.ksnd_global_lock); |
| 1997 | |
| 1998 | ksocknal_peer_failed(peer); |
| 1999 | ksocknal_txlist_done(peer->ksnp_ni, &zombies, 1); |
| 2000 | return 0; |
| 2001 | } |
| 2002 | |
| 2003 | /* |
| 2004 | * check whether we need to create more connds. |
| 2005 | * It will try to create new thread if it's necessary, @timeout can |
| 2006 | * be updated if failed to create, so caller wouldn't keep try while |
| 2007 | * running out of resource. |
| 2008 | */ |
| 2009 | static int |
| 2010 | ksocknal_connd_check_start(long sec, long *timeout) |
| 2011 | { |
| 2012 | char name[16]; |
| 2013 | int rc; |
| 2014 | int total = ksocknal_data.ksnd_connd_starting + |
| 2015 | ksocknal_data.ksnd_connd_running; |
| 2016 | |
| 2017 | if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) { |
| 2018 | /* still in initializing */ |
| 2019 | return 0; |
| 2020 | } |
| 2021 | |
| 2022 | if (total >= *ksocknal_tunables.ksnd_nconnds_max || |
| 2023 | total > ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV) { |
| 2024 | /* can't create more connd, or still have enough |
| 2025 | * threads to handle more connecting */ |
| 2026 | return 0; |
| 2027 | } |
| 2028 | |
| 2029 | if (list_empty(&ksocknal_data.ksnd_connd_routes)) { |
| 2030 | /* no pending connecting request */ |
| 2031 | return 0; |
| 2032 | } |
| 2033 | |
| 2034 | if (sec - ksocknal_data.ksnd_connd_failed_stamp <= 1) { |
| 2035 | /* may run out of resource, retry later */ |
| 2036 | *timeout = cfs_time_seconds(1); |
| 2037 | return 0; |
| 2038 | } |
| 2039 | |
| 2040 | if (ksocknal_data.ksnd_connd_starting > 0) { |
| 2041 | /* serialize starting to avoid flood */ |
| 2042 | return 0; |
| 2043 | } |
| 2044 | |
| 2045 | ksocknal_data.ksnd_connd_starting_stamp = sec; |
| 2046 | ksocknal_data.ksnd_connd_starting++; |
| 2047 | spin_unlock_bh(&ksocknal_data.ksnd_connd_lock); |
| 2048 | |
| 2049 | /* NB: total is the next id */ |
| 2050 | snprintf(name, sizeof(name), "socknal_cd%02d", total); |
| 2051 | rc = ksocknal_thread_start(ksocknal_connd, NULL, name); |
| 2052 | |
| 2053 | spin_lock_bh(&ksocknal_data.ksnd_connd_lock); |
| 2054 | if (rc == 0) |
| 2055 | return 1; |
| 2056 | |
| 2057 | /* we tried ... */ |
| 2058 | LASSERT(ksocknal_data.ksnd_connd_starting > 0); |
| 2059 | ksocknal_data.ksnd_connd_starting--; |
| 2060 | ksocknal_data.ksnd_connd_failed_stamp = cfs_time_current_sec(); |
| 2061 | |
| 2062 | return 1; |
| 2063 | } |
| 2064 | |
| 2065 | /* |
| 2066 | * check whether current thread can exit, it will return 1 if there are too |
| 2067 | * many threads and no creating in past 120 seconds. |
| 2068 | * Also, this function may update @timeout to make caller come back |
| 2069 | * again to recheck these conditions. |
| 2070 | */ |
| 2071 | static int |
| 2072 | ksocknal_connd_check_stop(long sec, long *timeout) |
| 2073 | { |
| 2074 | int val; |
| 2075 | |
| 2076 | if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) { |
| 2077 | /* still in initializing */ |
| 2078 | return 0; |
| 2079 | } |
| 2080 | |
| 2081 | if (ksocknal_data.ksnd_connd_starting > 0) { |
| 2082 | /* in progress of starting new thread */ |
| 2083 | return 0; |
| 2084 | } |
| 2085 | |
| 2086 | if (ksocknal_data.ksnd_connd_running <= |
| 2087 | *ksocknal_tunables.ksnd_nconnds) { /* can't shrink */ |
| 2088 | return 0; |
| 2089 | } |
| 2090 | |
| 2091 | /* created thread in past 120 seconds? */ |
| 2092 | val = (int)(ksocknal_data.ksnd_connd_starting_stamp + |
| 2093 | SOCKNAL_CONND_TIMEOUT - sec); |
| 2094 | |
| 2095 | *timeout = (val > 0) ? cfs_time_seconds(val) : |
| 2096 | cfs_time_seconds(SOCKNAL_CONND_TIMEOUT); |
| 2097 | if (val > 0) |
| 2098 | return 0; |
| 2099 | |
| 2100 | /* no creating in past 120 seconds */ |
| 2101 | |
| 2102 | return ksocknal_data.ksnd_connd_running > |
| 2103 | ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV; |
| 2104 | } |
| 2105 | |
| 2106 | /* Go through connd_routes queue looking for a route that we can process |
| 2107 | * right now, @timeout_p can be updated if we need to come back later */ |
| 2108 | static ksock_route_t * |
| 2109 | ksocknal_connd_get_route_locked(signed long *timeout_p) |
| 2110 | { |
| 2111 | ksock_route_t *route; |
| 2112 | cfs_time_t now; |
| 2113 | |
| 2114 | now = cfs_time_current(); |
| 2115 | |
| 2116 | /* connd_routes can contain both pending and ordinary routes */ |
| 2117 | list_for_each_entry (route, &ksocknal_data.ksnd_connd_routes, |
| 2118 | ksnr_connd_list) { |
| 2119 | |
| 2120 | if (route->ksnr_retry_interval == 0 || |
| 2121 | cfs_time_aftereq(now, route->ksnr_timeout)) |
| 2122 | return route; |
| 2123 | |
| 2124 | if (*timeout_p == MAX_SCHEDULE_TIMEOUT || |
| 2125 | (int)*timeout_p > (int)(route->ksnr_timeout - now)) |
| 2126 | *timeout_p = (int)(route->ksnr_timeout - now); |
| 2127 | } |
| 2128 | |
| 2129 | return NULL; |
| 2130 | } |
| 2131 | |
| 2132 | int |
| 2133 | ksocknal_connd (void *arg) |
| 2134 | { |
| 2135 | spinlock_t *connd_lock = &ksocknal_data.ksnd_connd_lock; |
| 2136 | ksock_connreq_t *cr; |
| 2137 | wait_queue_t wait; |
| 2138 | int nloops = 0; |
| 2139 | int cons_retry = 0; |
| 2140 | |
| 2141 | cfs_block_allsigs (); |
| 2142 | |
| 2143 | init_waitqueue_entry_current (&wait); |
| 2144 | |
| 2145 | spin_lock_bh(connd_lock); |
| 2146 | |
| 2147 | LASSERT(ksocknal_data.ksnd_connd_starting > 0); |
| 2148 | ksocknal_data.ksnd_connd_starting--; |
| 2149 | ksocknal_data.ksnd_connd_running++; |
| 2150 | |
| 2151 | while (!ksocknal_data.ksnd_shuttingdown) { |
| 2152 | ksock_route_t *route = NULL; |
| 2153 | long sec = cfs_time_current_sec(); |
| 2154 | long timeout = MAX_SCHEDULE_TIMEOUT; |
| 2155 | int dropped_lock = 0; |
| 2156 | |
| 2157 | if (ksocknal_connd_check_stop(sec, &timeout)) { |
| 2158 | /* wakeup another one to check stop */ |
| 2159 | wake_up(&ksocknal_data.ksnd_connd_waitq); |
| 2160 | break; |
| 2161 | } |
| 2162 | |
| 2163 | if (ksocknal_connd_check_start(sec, &timeout)) { |
| 2164 | /* created new thread */ |
| 2165 | dropped_lock = 1; |
| 2166 | } |
| 2167 | |
| 2168 | if (!list_empty(&ksocknal_data.ksnd_connd_connreqs)) { |
| 2169 | /* Connection accepted by the listener */ |
| 2170 | cr = list_entry(ksocknal_data.ksnd_connd_connreqs. \ |
| 2171 | next, ksock_connreq_t, ksncr_list); |
| 2172 | |
| 2173 | list_del(&cr->ksncr_list); |
| 2174 | spin_unlock_bh(connd_lock); |
| 2175 | dropped_lock = 1; |
| 2176 | |
| 2177 | ksocknal_create_conn(cr->ksncr_ni, NULL, |
| 2178 | cr->ksncr_sock, SOCKLND_CONN_NONE); |
| 2179 | lnet_ni_decref(cr->ksncr_ni); |
| 2180 | LIBCFS_FREE(cr, sizeof(*cr)); |
| 2181 | |
| 2182 | spin_lock_bh(connd_lock); |
| 2183 | } |
| 2184 | |
| 2185 | /* Only handle an outgoing connection request if there |
| 2186 | * is a thread left to handle incoming connections and |
| 2187 | * create new connd */ |
| 2188 | if (ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV < |
| 2189 | ksocknal_data.ksnd_connd_running) { |
| 2190 | route = ksocknal_connd_get_route_locked(&timeout); |
| 2191 | } |
| 2192 | if (route != NULL) { |
| 2193 | list_del (&route->ksnr_connd_list); |
| 2194 | ksocknal_data.ksnd_connd_connecting++; |
| 2195 | spin_unlock_bh(connd_lock); |
| 2196 | dropped_lock = 1; |
| 2197 | |
| 2198 | if (ksocknal_connect(route)) { |
| 2199 | /* consecutive retry */ |
| 2200 | if (cons_retry++ > SOCKNAL_INSANITY_RECONN) { |
| 2201 | CWARN("massive consecutive " |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 2202 | "re-connecting to %pI4h\n", |
| 2203 | &route->ksnr_ipaddr); |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 2204 | cons_retry = 0; |
| 2205 | } |
| 2206 | } else { |
| 2207 | cons_retry = 0; |
| 2208 | } |
| 2209 | |
| 2210 | ksocknal_route_decref(route); |
| 2211 | |
| 2212 | spin_lock_bh(connd_lock); |
| 2213 | ksocknal_data.ksnd_connd_connecting--; |
| 2214 | } |
| 2215 | |
| 2216 | if (dropped_lock) { |
| 2217 | if (++nloops < SOCKNAL_RESCHED) |
| 2218 | continue; |
| 2219 | spin_unlock_bh(connd_lock); |
| 2220 | nloops = 0; |
| 2221 | cond_resched(); |
| 2222 | spin_lock_bh(connd_lock); |
| 2223 | continue; |
| 2224 | } |
| 2225 | |
| 2226 | /* Nothing to do for 'timeout' */ |
| 2227 | set_current_state(TASK_INTERRUPTIBLE); |
| 2228 | add_wait_queue_exclusive(&ksocknal_data.ksnd_connd_waitq, &wait); |
| 2229 | spin_unlock_bh(connd_lock); |
| 2230 | |
| 2231 | nloops = 0; |
| 2232 | waitq_timedwait(&wait, TASK_INTERRUPTIBLE, timeout); |
| 2233 | |
| 2234 | set_current_state(TASK_RUNNING); |
| 2235 | remove_wait_queue(&ksocknal_data.ksnd_connd_waitq, &wait); |
| 2236 | spin_lock_bh(connd_lock); |
| 2237 | } |
| 2238 | ksocknal_data.ksnd_connd_running--; |
| 2239 | spin_unlock_bh(connd_lock); |
| 2240 | |
| 2241 | ksocknal_thread_fini(); |
| 2242 | return 0; |
| 2243 | } |
| 2244 | |
| 2245 | ksock_conn_t * |
| 2246 | ksocknal_find_timed_out_conn (ksock_peer_t *peer) |
| 2247 | { |
| 2248 | /* We're called with a shared lock on ksnd_global_lock */ |
| 2249 | ksock_conn_t *conn; |
| 2250 | struct list_head *ctmp; |
| 2251 | |
| 2252 | list_for_each (ctmp, &peer->ksnp_conns) { |
| 2253 | int error; |
| 2254 | conn = list_entry (ctmp, ksock_conn_t, ksnc_list); |
| 2255 | |
| 2256 | /* Don't need the {get,put}connsock dance to deref ksnc_sock */ |
| 2257 | LASSERT (!conn->ksnc_closing); |
| 2258 | |
| 2259 | /* SOCK_ERROR will reset error code of socket in |
| 2260 | * some platform (like Darwin8.x) */ |
| 2261 | error = cfs_sock_error(conn->ksnc_sock); |
| 2262 | if (error != 0) { |
| 2263 | ksocknal_conn_addref(conn); |
| 2264 | |
| 2265 | switch (error) { |
| 2266 | case ECONNRESET: |
| 2267 | CNETERR("A connection with %s " |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 2268 | "(%pI4h:%d) was reset; " |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 2269 | "it may have rebooted.\n", |
| 2270 | libcfs_id2str(peer->ksnp_id), |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 2271 | &conn->ksnc_ipaddr, |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 2272 | conn->ksnc_port); |
| 2273 | break; |
| 2274 | case ETIMEDOUT: |
| 2275 | CNETERR("A connection with %s " |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 2276 | "(%pI4h:%d) timed out; the " |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 2277 | "network or node may be down.\n", |
| 2278 | libcfs_id2str(peer->ksnp_id), |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 2279 | &conn->ksnc_ipaddr, |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 2280 | conn->ksnc_port); |
| 2281 | break; |
| 2282 | default: |
| 2283 | CNETERR("An unexpected network error %d " |
| 2284 | "occurred with %s " |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 2285 | "(%pI4h:%d\n", error, |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 2286 | libcfs_id2str(peer->ksnp_id), |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 2287 | &conn->ksnc_ipaddr, |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 2288 | conn->ksnc_port); |
| 2289 | break; |
| 2290 | } |
| 2291 | |
| 2292 | return (conn); |
| 2293 | } |
| 2294 | |
| 2295 | if (conn->ksnc_rx_started && |
| 2296 | cfs_time_aftereq(cfs_time_current(), |
| 2297 | conn->ksnc_rx_deadline)) { |
| 2298 | /* Timed out incomplete incoming message */ |
| 2299 | ksocknal_conn_addref(conn); |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 2300 | CNETERR("Timeout receiving from %s (%pI4h:%d), " |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 2301 | "state %d wanted %d left %d\n", |
| 2302 | libcfs_id2str(peer->ksnp_id), |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 2303 | &conn->ksnc_ipaddr, |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 2304 | conn->ksnc_port, |
| 2305 | conn->ksnc_rx_state, |
| 2306 | conn->ksnc_rx_nob_wanted, |
| 2307 | conn->ksnc_rx_nob_left); |
| 2308 | return (conn); |
| 2309 | } |
| 2310 | |
| 2311 | if ((!list_empty(&conn->ksnc_tx_queue) || |
| 2312 | cfs_sock_wmem_queued(conn->ksnc_sock) != 0) && |
| 2313 | cfs_time_aftereq(cfs_time_current(), |
| 2314 | conn->ksnc_tx_deadline)) { |
| 2315 | /* Timed out messages queued for sending or |
| 2316 | * buffered in the socket's send buffer */ |
| 2317 | ksocknal_conn_addref(conn); |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 2318 | CNETERR("Timeout sending data to %s (%pI4h:%d) " |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 2319 | "the network or that node may be down.\n", |
| 2320 | libcfs_id2str(peer->ksnp_id), |
Peng Tao | 5e8f692 | 2013-07-15 22:27:09 +0800 | [diff] [blame] | 2321 | &conn->ksnc_ipaddr, |
Peng Tao | d7e09d0 | 2013-05-02 16:46:55 +0800 | [diff] [blame] | 2322 | conn->ksnc_port); |
| 2323 | return (conn); |
| 2324 | } |
| 2325 | } |
| 2326 | |
| 2327 | return (NULL); |
| 2328 | } |
| 2329 | |
| 2330 | static inline void |
| 2331 | ksocknal_flush_stale_txs(ksock_peer_t *peer) |
| 2332 | { |
| 2333 | ksock_tx_t *tx; |
| 2334 | LIST_HEAD (stale_txs); |
| 2335 | |
| 2336 | write_lock_bh(&ksocknal_data.ksnd_global_lock); |
| 2337 | |
| 2338 | while (!list_empty (&peer->ksnp_tx_queue)) { |
| 2339 | tx = list_entry (peer->ksnp_tx_queue.next, |
| 2340 | ksock_tx_t, tx_list); |
| 2341 | |
| 2342 | if (!cfs_time_aftereq(cfs_time_current(), |
| 2343 | tx->tx_deadline)) |
| 2344 | break; |
| 2345 | |
| 2346 | list_del (&tx->tx_list); |
| 2347 | list_add_tail (&tx->tx_list, &stale_txs); |
| 2348 | } |
| 2349 | |
| 2350 | write_unlock_bh(&ksocknal_data.ksnd_global_lock); |
| 2351 | |
| 2352 | ksocknal_txlist_done(peer->ksnp_ni, &stale_txs, 1); |
| 2353 | } |
| 2354 | |
| 2355 | int |
| 2356 | ksocknal_send_keepalive_locked(ksock_peer_t *peer) |
| 2357 | { |
| 2358 | ksock_sched_t *sched; |
| 2359 | ksock_conn_t *conn; |
| 2360 | ksock_tx_t *tx; |
| 2361 | |
| 2362 | if (list_empty(&peer->ksnp_conns)) /* last_alive will be updated by create_conn */ |
| 2363 | return 0; |
| 2364 | |
| 2365 | if (peer->ksnp_proto != &ksocknal_protocol_v3x) |
| 2366 | return 0; |
| 2367 | |
| 2368 | if (*ksocknal_tunables.ksnd_keepalive <= 0 || |
| 2369 | cfs_time_before(cfs_time_current(), |
| 2370 | cfs_time_add(peer->ksnp_last_alive, |
| 2371 | cfs_time_seconds(*ksocknal_tunables.ksnd_keepalive)))) |
| 2372 | return 0; |
| 2373 | |
| 2374 | if (cfs_time_before(cfs_time_current(), |
| 2375 | peer->ksnp_send_keepalive)) |
| 2376 | return 0; |
| 2377 | |
| 2378 | /* retry 10 secs later, so we wouldn't put pressure |
| 2379 | * on this peer if we failed to send keepalive this time */ |
| 2380 | peer->ksnp_send_keepalive = cfs_time_shift(10); |
| 2381 | |
| 2382 | conn = ksocknal_find_conn_locked(peer, NULL, 1); |
| 2383 | if (conn != NULL) { |
| 2384 | sched = conn->ksnc_scheduler; |
| 2385 | |
| 2386 | spin_lock_bh(&sched->kss_lock); |
| 2387 | if (!list_empty(&conn->ksnc_tx_queue)) { |
| 2388 | spin_unlock_bh(&sched->kss_lock); |
| 2389 | /* there is an queued ACK, don't need keepalive */ |
| 2390 | return 0; |
| 2391 | } |
| 2392 | |
| 2393 | spin_unlock_bh(&sched->kss_lock); |
| 2394 | } |
| 2395 | |
| 2396 | read_unlock(&ksocknal_data.ksnd_global_lock); |
| 2397 | |
| 2398 | /* cookie = 1 is reserved for keepalive PING */ |
| 2399 | tx = ksocknal_alloc_tx_noop(1, 1); |
| 2400 | if (tx == NULL) { |
| 2401 | read_lock(&ksocknal_data.ksnd_global_lock); |
| 2402 | return -ENOMEM; |
| 2403 | } |
| 2404 | |
| 2405 | if (ksocknal_launch_packet(peer->ksnp_ni, tx, peer->ksnp_id) == 0) { |
| 2406 | read_lock(&ksocknal_data.ksnd_global_lock); |
| 2407 | return 1; |
| 2408 | } |
| 2409 | |
| 2410 | ksocknal_free_tx(tx); |
| 2411 | read_lock(&ksocknal_data.ksnd_global_lock); |
| 2412 | |
| 2413 | return -EIO; |
| 2414 | } |
| 2415 | |
| 2416 | |
| 2417 | void |
| 2418 | ksocknal_check_peer_timeouts (int idx) |
| 2419 | { |
| 2420 | struct list_head *peers = &ksocknal_data.ksnd_peers[idx]; |
| 2421 | ksock_peer_t *peer; |
| 2422 | ksock_conn_t *conn; |
| 2423 | ksock_tx_t *tx; |
| 2424 | |
| 2425 | again: |
| 2426 | /* NB. We expect to have a look at all the peers and not find any |
| 2427 | * connections to time out, so we just use a shared lock while we |
| 2428 | * take a look... */ |
| 2429 | read_lock(&ksocknal_data.ksnd_global_lock); |
| 2430 | |
| 2431 | list_for_each_entry(peer, peers, ksnp_list) { |
| 2432 | cfs_time_t deadline = 0; |
| 2433 | int resid = 0; |
| 2434 | int n = 0; |
| 2435 | |
| 2436 | if (ksocknal_send_keepalive_locked(peer) != 0) { |
| 2437 | read_unlock(&ksocknal_data.ksnd_global_lock); |
| 2438 | goto again; |
| 2439 | } |
| 2440 | |
| 2441 | conn = ksocknal_find_timed_out_conn (peer); |
| 2442 | |
| 2443 | if (conn != NULL) { |
| 2444 | read_unlock(&ksocknal_data.ksnd_global_lock); |
| 2445 | |
| 2446 | ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT); |
| 2447 | |
| 2448 | /* NB we won't find this one again, but we can't |
| 2449 | * just proceed with the next peer, since we dropped |
| 2450 | * ksnd_global_lock and it might be dead already! */ |
| 2451 | ksocknal_conn_decref(conn); |
| 2452 | goto again; |
| 2453 | } |
| 2454 | |
| 2455 | /* we can't process stale txs right here because we're |
| 2456 | * holding only shared lock */ |
| 2457 | if (!list_empty (&peer->ksnp_tx_queue)) { |
| 2458 | ksock_tx_t *tx = |
| 2459 | list_entry (peer->ksnp_tx_queue.next, |
| 2460 | ksock_tx_t, tx_list); |
| 2461 | |
| 2462 | if (cfs_time_aftereq(cfs_time_current(), |
| 2463 | tx->tx_deadline)) { |
| 2464 | |
| 2465 | ksocknal_peer_addref(peer); |
| 2466 | read_unlock(&ksocknal_data.ksnd_global_lock); |
| 2467 | |
| 2468 | ksocknal_flush_stale_txs(peer); |
| 2469 | |
| 2470 | ksocknal_peer_decref(peer); |
| 2471 | goto again; |
| 2472 | } |
| 2473 | } |
| 2474 | |
| 2475 | if (list_empty(&peer->ksnp_zc_req_list)) |
| 2476 | continue; |
| 2477 | |
| 2478 | spin_lock(&peer->ksnp_lock); |
| 2479 | list_for_each_entry(tx, &peer->ksnp_zc_req_list, tx_zc_list) { |
| 2480 | if (!cfs_time_aftereq(cfs_time_current(), |
| 2481 | tx->tx_deadline)) |
| 2482 | break; |
| 2483 | /* ignore the TX if connection is being closed */ |
| 2484 | if (tx->tx_conn->ksnc_closing) |
| 2485 | continue; |
| 2486 | n++; |
| 2487 | } |
| 2488 | |
| 2489 | if (n == 0) { |
| 2490 | spin_unlock(&peer->ksnp_lock); |
| 2491 | continue; |
| 2492 | } |
| 2493 | |
| 2494 | tx = list_entry(peer->ksnp_zc_req_list.next, |
| 2495 | ksock_tx_t, tx_zc_list); |
| 2496 | deadline = tx->tx_deadline; |
| 2497 | resid = tx->tx_resid; |
| 2498 | conn = tx->tx_conn; |
| 2499 | ksocknal_conn_addref(conn); |
| 2500 | |
| 2501 | spin_unlock(&peer->ksnp_lock); |
| 2502 | read_unlock(&ksocknal_data.ksnd_global_lock); |
| 2503 | |
| 2504 | CERROR("Total %d stale ZC_REQs for peer %s detected; the " |
| 2505 | "oldest(%p) timed out %ld secs ago, " |
| 2506 | "resid: %d, wmem: %d\n", |
| 2507 | n, libcfs_nid2str(peer->ksnp_id.nid), tx, |
| 2508 | cfs_duration_sec(cfs_time_current() - deadline), |
| 2509 | resid, cfs_sock_wmem_queued(conn->ksnc_sock)); |
| 2510 | |
| 2511 | ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT); |
| 2512 | ksocknal_conn_decref(conn); |
| 2513 | goto again; |
| 2514 | } |
| 2515 | |
| 2516 | read_unlock(&ksocknal_data.ksnd_global_lock); |
| 2517 | } |
| 2518 | |
| 2519 | int |
| 2520 | ksocknal_reaper (void *arg) |
| 2521 | { |
| 2522 | wait_queue_t wait; |
| 2523 | ksock_conn_t *conn; |
| 2524 | ksock_sched_t *sched; |
| 2525 | struct list_head enomem_conns; |
| 2526 | int nenomem_conns; |
| 2527 | cfs_duration_t timeout; |
| 2528 | int i; |
| 2529 | int peer_index = 0; |
| 2530 | cfs_time_t deadline = cfs_time_current(); |
| 2531 | |
| 2532 | cfs_block_allsigs (); |
| 2533 | |
| 2534 | INIT_LIST_HEAD(&enomem_conns); |
| 2535 | init_waitqueue_entry_current (&wait); |
| 2536 | |
| 2537 | spin_lock_bh(&ksocknal_data.ksnd_reaper_lock); |
| 2538 | |
| 2539 | while (!ksocknal_data.ksnd_shuttingdown) { |
| 2540 | |
| 2541 | if (!list_empty (&ksocknal_data.ksnd_deathrow_conns)) { |
| 2542 | conn = list_entry (ksocknal_data. \ |
| 2543 | ksnd_deathrow_conns.next, |
| 2544 | ksock_conn_t, ksnc_list); |
| 2545 | list_del (&conn->ksnc_list); |
| 2546 | |
| 2547 | spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock); |
| 2548 | |
| 2549 | ksocknal_terminate_conn(conn); |
| 2550 | ksocknal_conn_decref(conn); |
| 2551 | |
| 2552 | spin_lock_bh(&ksocknal_data.ksnd_reaper_lock); |
| 2553 | continue; |
| 2554 | } |
| 2555 | |
| 2556 | if (!list_empty (&ksocknal_data.ksnd_zombie_conns)) { |
| 2557 | conn = list_entry (ksocknal_data.ksnd_zombie_conns.\ |
| 2558 | next, ksock_conn_t, ksnc_list); |
| 2559 | list_del (&conn->ksnc_list); |
| 2560 | |
| 2561 | spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock); |
| 2562 | |
| 2563 | ksocknal_destroy_conn(conn); |
| 2564 | |
| 2565 | spin_lock_bh(&ksocknal_data.ksnd_reaper_lock); |
| 2566 | continue; |
| 2567 | } |
| 2568 | |
| 2569 | if (!list_empty (&ksocknal_data.ksnd_enomem_conns)) { |
| 2570 | list_add(&enomem_conns, |
| 2571 | &ksocknal_data.ksnd_enomem_conns); |
| 2572 | list_del_init(&ksocknal_data.ksnd_enomem_conns); |
| 2573 | } |
| 2574 | |
| 2575 | spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock); |
| 2576 | |
| 2577 | /* reschedule all the connections that stalled with ENOMEM... */ |
| 2578 | nenomem_conns = 0; |
| 2579 | while (!list_empty (&enomem_conns)) { |
| 2580 | conn = list_entry (enomem_conns.next, |
| 2581 | ksock_conn_t, ksnc_tx_list); |
| 2582 | list_del (&conn->ksnc_tx_list); |
| 2583 | |
| 2584 | sched = conn->ksnc_scheduler; |
| 2585 | |
| 2586 | spin_lock_bh(&sched->kss_lock); |
| 2587 | |
| 2588 | LASSERT(conn->ksnc_tx_scheduled); |
| 2589 | conn->ksnc_tx_ready = 1; |
| 2590 | list_add_tail(&conn->ksnc_tx_list, |
| 2591 | &sched->kss_tx_conns); |
| 2592 | wake_up(&sched->kss_waitq); |
| 2593 | |
| 2594 | spin_unlock_bh(&sched->kss_lock); |
| 2595 | nenomem_conns++; |
| 2596 | } |
| 2597 | |
| 2598 | /* careful with the jiffy wrap... */ |
| 2599 | while ((timeout = cfs_time_sub(deadline, |
| 2600 | cfs_time_current())) <= 0) { |
| 2601 | const int n = 4; |
| 2602 | const int p = 1; |
| 2603 | int chunk = ksocknal_data.ksnd_peer_hash_size; |
| 2604 | |
| 2605 | /* Time to check for timeouts on a few more peers: I do |
| 2606 | * checks every 'p' seconds on a proportion of the peer |
| 2607 | * table and I need to check every connection 'n' times |
| 2608 | * within a timeout interval, to ensure I detect a |
| 2609 | * timeout on any connection within (n+1)/n times the |
| 2610 | * timeout interval. */ |
| 2611 | |
| 2612 | if (*ksocknal_tunables.ksnd_timeout > n * p) |
| 2613 | chunk = (chunk * n * p) / |
| 2614 | *ksocknal_tunables.ksnd_timeout; |
| 2615 | if (chunk == 0) |
| 2616 | chunk = 1; |
| 2617 | |
| 2618 | for (i = 0; i < chunk; i++) { |
| 2619 | ksocknal_check_peer_timeouts (peer_index); |
| 2620 | peer_index = (peer_index + 1) % |
| 2621 | ksocknal_data.ksnd_peer_hash_size; |
| 2622 | } |
| 2623 | |
| 2624 | deadline = cfs_time_add(deadline, cfs_time_seconds(p)); |
| 2625 | } |
| 2626 | |
| 2627 | if (nenomem_conns != 0) { |
| 2628 | /* Reduce my timeout if I rescheduled ENOMEM conns. |
| 2629 | * This also prevents me getting woken immediately |
| 2630 | * if any go back on my enomem list. */ |
| 2631 | timeout = SOCKNAL_ENOMEM_RETRY; |
| 2632 | } |
| 2633 | ksocknal_data.ksnd_reaper_waketime = |
| 2634 | cfs_time_add(cfs_time_current(), timeout); |
| 2635 | |
| 2636 | set_current_state (TASK_INTERRUPTIBLE); |
| 2637 | add_wait_queue (&ksocknal_data.ksnd_reaper_waitq, &wait); |
| 2638 | |
| 2639 | if (!ksocknal_data.ksnd_shuttingdown && |
| 2640 | list_empty (&ksocknal_data.ksnd_deathrow_conns) && |
| 2641 | list_empty (&ksocknal_data.ksnd_zombie_conns)) |
| 2642 | waitq_timedwait (&wait, TASK_INTERRUPTIBLE, |
| 2643 | timeout); |
| 2644 | |
| 2645 | set_current_state (TASK_RUNNING); |
| 2646 | remove_wait_queue (&ksocknal_data.ksnd_reaper_waitq, &wait); |
| 2647 | |
| 2648 | spin_lock_bh(&ksocknal_data.ksnd_reaper_lock); |
| 2649 | } |
| 2650 | |
| 2651 | spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock); |
| 2652 | |
| 2653 | ksocknal_thread_fini(); |
| 2654 | return 0; |
| 2655 | } |