1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Shared Memory Communications over RDMA (SMC-R) and RoCE 4 * 5 * Connection Data Control (CDC) 6 * handles flow control 7 * 8 * Copyright IBM Corp. 2016 9 * 10 * Author(s): Ursula Braun <ubraun@linux.vnet.ibm.com> 11 */ 12 13 #include <linux/spinlock.h> 14 15 #include "smc.h" 16 #include "smc_wr.h" 17 #include "smc_cdc.h" 18 #include "smc_tx.h" 19 #include "smc_rx.h" 20 #include "smc_close.h" 21 #include "smc_ism.h" 22 23 /********************************** send *************************************/ 24 25 /* handler for send/transmission completion of a CDC msg */ 26 static void smc_cdc_tx_handler(struct smc_wr_tx_pend_priv *pnd_snd, 27 struct smc_link *link, 28 enum ib_wc_status wc_status) 29 { 30 struct smc_cdc_tx_pend *cdcpend = (struct smc_cdc_tx_pend *)pnd_snd; 31 struct smc_connection *conn = cdcpend->conn; 32 struct smc_buf_desc *sndbuf_desc; 33 struct smc_sock *smc; 34 int diff; 35 36 sndbuf_desc = conn->sndbuf_desc; 37 smc = container_of(conn, struct smc_sock, conn); 38 bh_lock_sock(&smc->sk); 39 if (!wc_status && sndbuf_desc) { 40 diff = smc_curs_diff(sndbuf_desc->len, 41 &cdcpend->conn->tx_curs_fin, 42 &cdcpend->cursor); 43 /* sndbuf_space is decreased in smc_sendmsg */ 44 smp_mb__before_atomic(); 45 atomic_add(diff, &cdcpend->conn->sndbuf_space); 46 /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */ 47 smp_mb__after_atomic(); 48 smc_curs_copy(&conn->tx_curs_fin, &cdcpend->cursor, conn); 49 smc_curs_copy(&conn->local_tx_ctrl_fin, &cdcpend->p_cursor, 50 conn); 51 conn->tx_cdc_seq_fin = cdcpend->ctrl_seq; 52 } 53 54 if (atomic_dec_and_test(&conn->cdc_pend_tx_wr)) { 55 /* If user owns the sock_lock, mark the connection need sending. 56 * User context will later try to send when it release sock_lock 57 * in smc_release_cb() 58 */ 59 if (sock_owned_by_user(&smc->sk)) 60 conn->tx_in_release_sock = true; 61 else 62 smc_tx_pending(conn); 63 64 if (unlikely(wq_has_sleeper(&conn->cdc_pend_tx_wq))) 65 wake_up(&conn->cdc_pend_tx_wq); 66 } 67 WARN_ON(atomic_read(&conn->cdc_pend_tx_wr) < 0); 68 69 smc_tx_sndbuf_nonfull(smc); 70 bh_unlock_sock(&smc->sk); 71 } 72 73 int smc_cdc_get_free_slot(struct smc_connection *conn, 74 struct smc_link *link, 75 struct smc_wr_buf **wr_buf, 76 struct smc_rdma_wr **wr_rdma_buf, 77 struct smc_cdc_tx_pend **pend) 78 { 79 int rc; 80 81 rc = smc_wr_tx_get_free_slot(link, smc_cdc_tx_handler, wr_buf, 82 wr_rdma_buf, 83 (struct smc_wr_tx_pend_priv **)pend); 84 if (conn->killed) { 85 /* abnormal termination */ 86 if (!rc) 87 smc_wr_tx_put_slot(link, 88 (struct smc_wr_tx_pend_priv *)(*pend)); 89 rc = -EPIPE; 90 } 91 return rc; 92 } 93 94 static inline void smc_cdc_add_pending_send(struct smc_connection *conn, 95 struct smc_cdc_tx_pend *pend) 96 { 97 BUILD_BUG_ON_MSG( 98 sizeof(struct smc_cdc_msg) > SMC_WR_BUF_SIZE, 99 "must increase SMC_WR_BUF_SIZE to at least sizeof(struct smc_cdc_msg)"); 100 BUILD_BUG_ON_MSG( 101 offsetofend(struct smc_cdc_msg, reserved) > SMC_WR_TX_SIZE, 102 "must adapt SMC_WR_TX_SIZE to sizeof(struct smc_cdc_msg); if not all smc_wr upper layer protocols use the same message size any more, must start to set link->wr_tx_sges[i].length on each individual smc_wr_tx_send()"); 103 BUILD_BUG_ON_MSG( 104 sizeof(struct smc_cdc_tx_pend) > SMC_WR_TX_PEND_PRIV_SIZE, 105 "must increase SMC_WR_TX_PEND_PRIV_SIZE to at least sizeof(struct smc_cdc_tx_pend)"); 106 pend->conn = conn; 107 pend->cursor = conn->tx_curs_sent; 108 pend->p_cursor = conn->local_tx_ctrl.prod; 109 pend->ctrl_seq = conn->tx_cdc_seq; 110 } 111 112 int smc_cdc_msg_send(struct smc_connection *conn, 113 struct smc_wr_buf *wr_buf, 114 struct smc_cdc_tx_pend *pend) 115 { 116 struct smc_link *link = conn->lnk; 117 union smc_host_cursor cfed; 118 int rc; 119 120 smc_cdc_add_pending_send(conn, pend); 121 122 conn->tx_cdc_seq++; 123 conn->local_tx_ctrl.seqno = conn->tx_cdc_seq; 124 smc_host_msg_to_cdc((struct smc_cdc_msg *)wr_buf, conn, &cfed); 125 126 atomic_inc(&conn->cdc_pend_tx_wr); 127 smp_mb__after_atomic(); /* Make sure cdc_pend_tx_wr added before post */ 128 129 rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend); 130 if (!rc) { 131 smc_curs_copy(&conn->rx_curs_confirmed, &cfed, conn); 132 conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0; 133 } else { 134 conn->tx_cdc_seq--; 135 conn->local_tx_ctrl.seqno = conn->tx_cdc_seq; 136 atomic_dec(&conn->cdc_pend_tx_wr); 137 } 138 139 return rc; 140 } 141 142 /* send a validation msg indicating the move of a conn to an other QP link */ 143 int smcr_cdc_msg_send_validation(struct smc_connection *conn, 144 struct smc_cdc_tx_pend *pend, 145 struct smc_wr_buf *wr_buf) 146 { 147 struct smc_host_cdc_msg *local = &conn->local_tx_ctrl; 148 struct smc_link *link = conn->lnk; 149 struct smc_cdc_msg *peer; 150 int rc; 151 152 peer = (struct smc_cdc_msg *)wr_buf; 153 peer->common.type = local->common.type; 154 peer->len = local->len; 155 peer->seqno = htons(conn->tx_cdc_seq_fin); /* seqno last compl. tx */ 156 peer->token = htonl(local->token); 157 peer->prod_flags.failover_validation = 1; 158 159 /* We need to set pend->conn here to make sure smc_cdc_tx_handler() 160 * can handle properly 161 */ 162 smc_cdc_add_pending_send(conn, pend); 163 164 atomic_inc(&conn->cdc_pend_tx_wr); 165 smp_mb__after_atomic(); /* Make sure cdc_pend_tx_wr added before post */ 166 167 rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend); 168 if (unlikely(rc)) 169 atomic_dec(&conn->cdc_pend_tx_wr); 170 171 return rc; 172 } 173 174 static int smcr_cdc_get_slot_and_msg_send(struct smc_connection *conn) 175 { 176 struct smc_cdc_tx_pend *pend; 177 struct smc_wr_buf *wr_buf; 178 struct smc_link *link; 179 bool again = false; 180 int rc; 181 182 again: 183 link = conn->lnk; 184 if (!smc_wr_tx_link_hold(link)) 185 return -ENOLINK; 186 rc = smc_cdc_get_free_slot(conn, link, &wr_buf, NULL, &pend); 187 if (rc) 188 goto put_out; 189 190 spin_lock_bh(&conn->send_lock); 191 if (link != conn->lnk) { 192 /* link of connection changed, try again one time*/ 193 spin_unlock_bh(&conn->send_lock); 194 smc_wr_tx_put_slot(link, 195 (struct smc_wr_tx_pend_priv *)pend); 196 smc_wr_tx_link_put(link); 197 if (again) 198 return -ENOLINK; 199 again = true; 200 goto again; 201 } 202 rc = smc_cdc_msg_send(conn, wr_buf, pend); 203 spin_unlock_bh(&conn->send_lock); 204 put_out: 205 smc_wr_tx_link_put(link); 206 return rc; 207 } 208 209 int smc_cdc_get_slot_and_msg_send(struct smc_connection *conn) 210 { 211 int rc; 212 213 if (!smc_conn_lgr_valid(conn) || 214 (conn->lgr->is_smcd && conn->lgr->peer_shutdown)) 215 return -EPIPE; 216 217 if (conn->lgr->is_smcd) { 218 spin_lock_bh(&conn->send_lock); 219 rc = smcd_cdc_msg_send(conn); 220 spin_unlock_bh(&conn->send_lock); 221 } else { 222 rc = smcr_cdc_get_slot_and_msg_send(conn); 223 } 224 225 return rc; 226 } 227 228 void smc_cdc_wait_pend_tx_wr(struct smc_connection *conn) 229 { 230 wait_event(conn->cdc_pend_tx_wq, !atomic_read(&conn->cdc_pend_tx_wr)); 231 } 232 233 /* Send a SMC-D CDC header. 234 * This increments the free space available in our send buffer. 235 * Also update the confirmed receive buffer with what was sent to the peer. 236 */ 237 int smcd_cdc_msg_send(struct smc_connection *conn) 238 { 239 struct smc_sock *smc = container_of(conn, struct smc_sock, conn); 240 union smc_host_cursor curs; 241 struct smcd_cdc_msg cdc; 242 int rc, diff; 243 244 memset(&cdc, 0, sizeof(cdc)); 245 cdc.common.type = SMC_CDC_MSG_TYPE; 246 curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.prod.acurs); 247 cdc.prod.wrap = curs.wrap; 248 cdc.prod.count = curs.count; 249 curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.cons.acurs); 250 cdc.cons.wrap = curs.wrap; 251 cdc.cons.count = curs.count; 252 cdc.cons.prod_flags = conn->local_tx_ctrl.prod_flags; 253 cdc.cons.conn_state_flags = conn->local_tx_ctrl.conn_state_flags; 254 rc = smcd_tx_ism_write(conn, &cdc, sizeof(cdc), 0, 1); 255 if (rc) 256 return rc; 257 smc_curs_copy(&conn->rx_curs_confirmed, &curs, conn); 258 conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0; 259 260 if (smc_ism_support_dmb_nocopy(conn->lgr->smcd)) 261 /* if local sndbuf shares the same memory region with 262 * peer DMB, then don't update the tx_curs_fin 263 * and sndbuf_space until peer has consumed the data. 264 */ 265 return 0; 266 267 /* Calculate transmitted data and increment free send buffer space */ 268 diff = smc_curs_diff(conn->sndbuf_desc->len, &conn->tx_curs_fin, 269 &conn->tx_curs_sent); 270 /* increased by confirmed number of bytes */ 271 smp_mb__before_atomic(); 272 atomic_add(diff, &conn->sndbuf_space); 273 /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */ 274 smp_mb__after_atomic(); 275 smc_curs_copy(&conn->tx_curs_fin, &conn->tx_curs_sent, conn); 276 277 smc_tx_sndbuf_nonfull(smc); 278 return 0; 279 } 280 281 /********************************* receive ***********************************/ 282 283 static inline bool smc_cdc_before(u16 seq1, u16 seq2) 284 { 285 return (s16)(seq1 - seq2) < 0; 286 } 287 288 static void smc_cdc_handle_urg_data_arrival(struct smc_sock *smc, 289 int *diff_prod) 290 { 291 struct smc_connection *conn = &smc->conn; 292 char *base; 293 294 /* new data included urgent business */ 295 smc_curs_copy(&conn->urg_curs, &conn->local_rx_ctrl.prod, conn); 296 conn->urg_state = SMC_URG_VALID; 297 if (!sock_flag(&smc->sk, SOCK_URGINLINE)) 298 /* we'll skip the urgent byte, so don't account for it */ 299 (*diff_prod)--; 300 base = (char *)conn->rmb_desc->cpu_addr + conn->rx_off; 301 if (conn->urg_curs.count) 302 conn->urg_rx_byte = *(base + conn->urg_curs.count - 1); 303 else 304 conn->urg_rx_byte = *(base + conn->rmb_desc->len - 1); 305 sk_send_sigurg(&smc->sk); 306 } 307 308 static void smc_cdc_msg_validate(struct smc_sock *smc, struct smc_cdc_msg *cdc, 309 struct smc_link *link) 310 { 311 struct smc_connection *conn = &smc->conn; 312 u16 recv_seq = ntohs(cdc->seqno); 313 s16 diff; 314 315 /* check that seqnum was seen before */ 316 diff = conn->local_rx_ctrl.seqno - recv_seq; 317 if (diff < 0) { /* diff larger than 0x7fff */ 318 /* drop connection */ 319 conn->out_of_sync = 1; /* prevent any further receives */ 320 spin_lock_bh(&conn->send_lock); 321 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1; 322 conn->lnk = link; 323 spin_unlock_bh(&conn->send_lock); 324 sock_hold(&smc->sk); /* sock_put in abort_work */ 325 if (!queue_work(smc_close_wq, &conn->abort_work)) 326 sock_put(&smc->sk); 327 } 328 } 329 330 static void smc_cdc_msg_recv_action(struct smc_sock *smc, 331 struct smc_cdc_msg *cdc) 332 { 333 union smc_host_cursor cons_old, prod_old; 334 struct smc_connection *conn = &smc->conn; 335 int diff_cons, diff_prod, diff_tx; 336 337 smc_curs_copy(&prod_old, &conn->local_rx_ctrl.prod, conn); 338 smc_curs_copy(&cons_old, &conn->local_rx_ctrl.cons, conn); 339 smc_cdc_msg_to_host(&conn->local_rx_ctrl, cdc, conn); 340 341 diff_cons = smc_curs_diff(conn->peer_rmbe_size, &cons_old, 342 &conn->local_rx_ctrl.cons); 343 if (diff_cons) { 344 /* peer_rmbe_space is decreased during data transfer with RDMA 345 * write 346 */ 347 smp_mb__before_atomic(); 348 atomic_add(diff_cons, &conn->peer_rmbe_space); 349 /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */ 350 smp_mb__after_atomic(); 351 352 /* if local sndbuf shares the same memory region with 353 * peer RMB, then update tx_curs_fin and sndbuf_space 354 * here since peer has already consumed the data. 355 */ 356 if (conn->lgr->is_smcd && 357 smc_ism_support_dmb_nocopy(conn->lgr->smcd)) { 358 /* Calculate consumed data and 359 * increment free send buffer space. 360 */ 361 diff_tx = smc_curs_diff(conn->sndbuf_desc->len, 362 &conn->tx_curs_fin, 363 &conn->local_rx_ctrl.cons); 364 /* increase local sndbuf space and fin_curs */ 365 smp_mb__before_atomic(); 366 atomic_add(diff_tx, &conn->sndbuf_space); 367 /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */ 368 smp_mb__after_atomic(); 369 smc_curs_copy(&conn->tx_curs_fin, 370 &conn->local_rx_ctrl.cons, conn); 371 372 smc_tx_sndbuf_nonfull(smc); 373 } 374 } 375 376 diff_prod = smc_curs_diff(conn->rmb_desc->len, &prod_old, 377 &conn->local_rx_ctrl.prod); 378 if (diff_prod) { 379 if (conn->local_rx_ctrl.prod_flags.urg_data_present) 380 smc_cdc_handle_urg_data_arrival(smc, &diff_prod); 381 /* bytes_to_rcv is decreased in smc_recvmsg */ 382 smp_mb__before_atomic(); 383 atomic_add(diff_prod, &conn->bytes_to_rcv); 384 /* guarantee 0 <= bytes_to_rcv <= rmb_desc->len */ 385 smp_mb__after_atomic(); 386 smc->sk.sk_data_ready(&smc->sk); 387 } else { 388 if (conn->local_rx_ctrl.prod_flags.write_blocked) 389 smc->sk.sk_data_ready(&smc->sk); 390 if (conn->local_rx_ctrl.prod_flags.urg_data_pending) 391 conn->urg_state = SMC_URG_NOTYET; 392 } 393 394 /* trigger sndbuf consumer: RDMA write into peer RMBE and CDC */ 395 if ((diff_cons && smc_tx_prepared_sends(conn)) || 396 conn->local_rx_ctrl.prod_flags.cons_curs_upd_req || 397 conn->local_rx_ctrl.prod_flags.urg_data_pending) { 398 if (!sock_owned_by_user(&smc->sk)) 399 smc_tx_pending(conn); 400 else 401 conn->tx_in_release_sock = true; 402 } 403 404 if (diff_cons && conn->urg_tx_pend && 405 atomic_read(&conn->peer_rmbe_space) == conn->peer_rmbe_size) { 406 /* urg data confirmed by peer, indicate we're ready for more */ 407 conn->urg_tx_pend = false; 408 smc->sk.sk_write_space(&smc->sk); 409 } 410 411 if (conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) { 412 smc->sk.sk_err = ECONNRESET; 413 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1; 414 } 415 if (smc_cdc_rxed_any_close_or_senddone(conn)) { 416 smc->sk.sk_shutdown |= RCV_SHUTDOWN; 417 if (smc->clcsock && smc->clcsock->sk) 418 smc->clcsock->sk->sk_shutdown |= RCV_SHUTDOWN; 419 smc_sock_set_flag(&smc->sk, SOCK_DONE); 420 sock_hold(&smc->sk); /* sock_put in close_work */ 421 if (!queue_work(smc_close_wq, &conn->close_work)) 422 sock_put(&smc->sk); 423 } 424 } 425 426 /* called under tasklet context */ 427 static void smc_cdc_msg_recv(struct smc_sock *smc, struct smc_cdc_msg *cdc) 428 { 429 sock_hold(&smc->sk); 430 bh_lock_sock(&smc->sk); 431 smc_cdc_msg_recv_action(smc, cdc); 432 bh_unlock_sock(&smc->sk); 433 sock_put(&smc->sk); /* no free sk in softirq-context */ 434 } 435 436 /* Schedule a tasklet for this connection. Triggered from the ISM device IRQ 437 * handler to indicate update in the DMBE. 438 * 439 * Context: 440 * - tasklet context 441 */ 442 static void smcd_cdc_rx_tsklet(struct tasklet_struct *t) 443 { 444 struct smc_connection *conn = from_tasklet(conn, t, rx_tsklet); 445 struct smcd_cdc_msg *data_cdc; 446 struct smcd_cdc_msg cdc; 447 struct smc_sock *smc; 448 449 if (!conn || conn->killed) 450 return; 451 452 data_cdc = (struct smcd_cdc_msg *)conn->rmb_desc->cpu_addr; 453 smcd_curs_copy(&cdc.prod, &data_cdc->prod, conn); 454 smcd_curs_copy(&cdc.cons, &data_cdc->cons, conn); 455 smc = container_of(conn, struct smc_sock, conn); 456 smc_cdc_msg_recv(smc, (struct smc_cdc_msg *)&cdc); 457 } 458 459 /* Initialize receive tasklet. Called from ISM device IRQ handler to start 460 * receiver side. 461 */ 462 void smcd_cdc_rx_init(struct smc_connection *conn) 463 { 464 tasklet_setup(&conn->rx_tsklet, smcd_cdc_rx_tsklet); 465 } 466 467 /***************************** init, exit, misc ******************************/ 468 469 static void smc_cdc_rx_handler(struct ib_wc *wc, void *buf) 470 { 471 struct smc_link *link = (struct smc_link *)wc->qp->qp_context; 472 struct smc_cdc_msg *cdc = buf; 473 struct smc_connection *conn; 474 struct smc_link_group *lgr; 475 struct smc_sock *smc; 476 477 if (wc->byte_len < offsetof(struct smc_cdc_msg, reserved)) 478 return; /* short message */ 479 if (cdc->len != SMC_WR_TX_SIZE) 480 return; /* invalid message */ 481 482 /* lookup connection */ 483 lgr = smc_get_lgr(link); 484 read_lock_bh(&lgr->conns_lock); 485 conn = smc_lgr_find_conn(ntohl(cdc->token), lgr); 486 read_unlock_bh(&lgr->conns_lock); 487 if (!conn || conn->out_of_sync) 488 return; 489 smc = container_of(conn, struct smc_sock, conn); 490 491 if (cdc->prod_flags.failover_validation) { 492 smc_cdc_msg_validate(smc, cdc, link); 493 return; 494 } 495 if (smc_cdc_before(ntohs(cdc->seqno), 496 conn->local_rx_ctrl.seqno)) 497 /* received seqno is old */ 498 return; 499 500 smc_cdc_msg_recv(smc, cdc); 501 } 502 503 static struct smc_wr_rx_handler smc_cdc_rx_handlers[] = { 504 { 505 .handler = smc_cdc_rx_handler, 506 .type = SMC_CDC_MSG_TYPE 507 }, 508 { 509 .handler = NULL, 510 } 511 }; 512 513 int __init smc_cdc_init(void) 514 { 515 struct smc_wr_rx_handler *handler; 516 int rc = 0; 517 518 for (handler = smc_cdc_rx_handlers; handler->handler; handler++) { 519 INIT_HLIST_NODE(&handler->list); 520 rc = smc_wr_rx_register_handler(handler); 521 if (rc) 522 break; 523 } 524 return rc; 525 } 526