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 22 /********************************** send *************************************/ 23 24 /* handler for send/transmission completion of a CDC msg */ 25 static void smc_cdc_tx_handler(struct smc_wr_tx_pend_priv *pnd_snd, 26 struct smc_link *link, 27 enum ib_wc_status wc_status) 28 { 29 struct smc_cdc_tx_pend *cdcpend = (struct smc_cdc_tx_pend *)pnd_snd; 30 struct smc_connection *conn = cdcpend->conn; 31 struct smc_sock *smc; 32 int diff; 33 34 smc = container_of(conn, struct smc_sock, conn); 35 bh_lock_sock(&smc->sk); 36 if (!wc_status) { 37 diff = smc_curs_diff(cdcpend->conn->sndbuf_desc->len, 38 &cdcpend->conn->tx_curs_fin, 39 &cdcpend->cursor); 40 /* sndbuf_space is decreased in smc_sendmsg */ 41 smp_mb__before_atomic(); 42 atomic_add(diff, &cdcpend->conn->sndbuf_space); 43 /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */ 44 smp_mb__after_atomic(); 45 smc_curs_copy(&conn->tx_curs_fin, &cdcpend->cursor, conn); 46 smc_curs_copy(&conn->local_tx_ctrl_fin, &cdcpend->p_cursor, 47 conn); 48 conn->tx_cdc_seq_fin = cdcpend->ctrl_seq; 49 } 50 51 if (atomic_dec_and_test(&conn->cdc_pend_tx_wr)) { 52 /* If user owns the sock_lock, mark the connection need sending. 53 * User context will later try to send when it release sock_lock 54 * in smc_release_cb() 55 */ 56 if (sock_owned_by_user(&smc->sk)) 57 conn->tx_in_release_sock = true; 58 else 59 smc_tx_pending(conn); 60 61 if (unlikely(wq_has_sleeper(&conn->cdc_pend_tx_wq))) 62 wake_up(&conn->cdc_pend_tx_wq); 63 } 64 WARN_ON(atomic_read(&conn->cdc_pend_tx_wr) < 0); 65 66 smc_tx_sndbuf_nonfull(smc); 67 bh_unlock_sock(&smc->sk); 68 } 69 70 int smc_cdc_get_free_slot(struct smc_connection *conn, 71 struct smc_link *link, 72 struct smc_wr_buf **wr_buf, 73 struct smc_rdma_wr **wr_rdma_buf, 74 struct smc_cdc_tx_pend **pend) 75 { 76 int rc; 77 78 rc = smc_wr_tx_get_free_slot(link, smc_cdc_tx_handler, wr_buf, 79 wr_rdma_buf, 80 (struct smc_wr_tx_pend_priv **)pend); 81 if (conn->killed) { 82 /* abnormal termination */ 83 if (!rc) 84 smc_wr_tx_put_slot(link, 85 (struct smc_wr_tx_pend_priv *)(*pend)); 86 rc = -EPIPE; 87 } 88 return rc; 89 } 90 91 static inline void smc_cdc_add_pending_send(struct smc_connection *conn, 92 struct smc_cdc_tx_pend *pend) 93 { 94 BUILD_BUG_ON_MSG( 95 sizeof(struct smc_cdc_msg) > SMC_WR_BUF_SIZE, 96 "must increase SMC_WR_BUF_SIZE to at least sizeof(struct smc_cdc_msg)"); 97 BUILD_BUG_ON_MSG( 98 offsetofend(struct smc_cdc_msg, reserved) > SMC_WR_TX_SIZE, 99 "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()"); 100 BUILD_BUG_ON_MSG( 101 sizeof(struct smc_cdc_tx_pend) > SMC_WR_TX_PEND_PRIV_SIZE, 102 "must increase SMC_WR_TX_PEND_PRIV_SIZE to at least sizeof(struct smc_cdc_tx_pend)"); 103 pend->conn = conn; 104 pend->cursor = conn->tx_curs_sent; 105 pend->p_cursor = conn->local_tx_ctrl.prod; 106 pend->ctrl_seq = conn->tx_cdc_seq; 107 } 108 109 int smc_cdc_msg_send(struct smc_connection *conn, 110 struct smc_wr_buf *wr_buf, 111 struct smc_cdc_tx_pend *pend) 112 { 113 struct smc_link *link = conn->lnk; 114 union smc_host_cursor cfed; 115 int rc; 116 117 if (unlikely(!READ_ONCE(conn->sndbuf_desc))) 118 return -ENOBUFS; 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 /* Calculate transmitted data and increment free send buffer space */ 260 diff = smc_curs_diff(conn->sndbuf_desc->len, &conn->tx_curs_fin, 261 &conn->tx_curs_sent); 262 /* increased by confirmed number of bytes */ 263 smp_mb__before_atomic(); 264 atomic_add(diff, &conn->sndbuf_space); 265 /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */ 266 smp_mb__after_atomic(); 267 smc_curs_copy(&conn->tx_curs_fin, &conn->tx_curs_sent, conn); 268 269 smc_tx_sndbuf_nonfull(smc); 270 return rc; 271 } 272 273 /********************************* receive ***********************************/ 274 275 static inline bool smc_cdc_before(u16 seq1, u16 seq2) 276 { 277 return (s16)(seq1 - seq2) < 0; 278 } 279 280 static void smc_cdc_handle_urg_data_arrival(struct smc_sock *smc, 281 int *diff_prod) 282 { 283 struct smc_connection *conn = &smc->conn; 284 char *base; 285 286 /* new data included urgent business */ 287 smc_curs_copy(&conn->urg_curs, &conn->local_rx_ctrl.prod, conn); 288 conn->urg_state = SMC_URG_VALID; 289 if (!sock_flag(&smc->sk, SOCK_URGINLINE)) 290 /* we'll skip the urgent byte, so don't account for it */ 291 (*diff_prod)--; 292 base = (char *)conn->rmb_desc->cpu_addr + conn->rx_off; 293 if (conn->urg_curs.count) 294 conn->urg_rx_byte = *(base + conn->urg_curs.count - 1); 295 else 296 conn->urg_rx_byte = *(base + conn->rmb_desc->len - 1); 297 sk_send_sigurg(&smc->sk); 298 } 299 300 static void smc_cdc_msg_validate(struct smc_sock *smc, struct smc_cdc_msg *cdc, 301 struct smc_link *link) 302 { 303 struct smc_connection *conn = &smc->conn; 304 u16 recv_seq = ntohs(cdc->seqno); 305 s16 diff; 306 307 /* check that seqnum was seen before */ 308 diff = conn->local_rx_ctrl.seqno - recv_seq; 309 if (diff < 0) { /* diff larger than 0x7fff */ 310 /* drop connection */ 311 conn->out_of_sync = 1; /* prevent any further receives */ 312 spin_lock_bh(&conn->send_lock); 313 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1; 314 conn->lnk = link; 315 spin_unlock_bh(&conn->send_lock); 316 sock_hold(&smc->sk); /* sock_put in abort_work */ 317 if (!queue_work(smc_close_wq, &conn->abort_work)) 318 sock_put(&smc->sk); 319 } 320 } 321 322 static void smc_cdc_msg_recv_action(struct smc_sock *smc, 323 struct smc_cdc_msg *cdc) 324 { 325 union smc_host_cursor cons_old, prod_old; 326 struct smc_connection *conn = &smc->conn; 327 int diff_cons, diff_prod; 328 329 smc_curs_copy(&prod_old, &conn->local_rx_ctrl.prod, conn); 330 smc_curs_copy(&cons_old, &conn->local_rx_ctrl.cons, conn); 331 smc_cdc_msg_to_host(&conn->local_rx_ctrl, cdc, conn); 332 333 diff_cons = smc_curs_diff(conn->peer_rmbe_size, &cons_old, 334 &conn->local_rx_ctrl.cons); 335 if (diff_cons) { 336 /* peer_rmbe_space is decreased during data transfer with RDMA 337 * write 338 */ 339 smp_mb__before_atomic(); 340 atomic_add(diff_cons, &conn->peer_rmbe_space); 341 /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */ 342 smp_mb__after_atomic(); 343 } 344 345 diff_prod = smc_curs_diff(conn->rmb_desc->len, &prod_old, 346 &conn->local_rx_ctrl.prod); 347 if (diff_prod) { 348 if (conn->local_rx_ctrl.prod_flags.urg_data_present) 349 smc_cdc_handle_urg_data_arrival(smc, &diff_prod); 350 /* bytes_to_rcv is decreased in smc_recvmsg */ 351 smp_mb__before_atomic(); 352 atomic_add(diff_prod, &conn->bytes_to_rcv); 353 /* guarantee 0 <= bytes_to_rcv <= rmb_desc->len */ 354 smp_mb__after_atomic(); 355 smc->sk.sk_data_ready(&smc->sk); 356 } else { 357 if (conn->local_rx_ctrl.prod_flags.write_blocked) 358 smc->sk.sk_data_ready(&smc->sk); 359 if (conn->local_rx_ctrl.prod_flags.urg_data_pending) 360 conn->urg_state = SMC_URG_NOTYET; 361 } 362 363 /* trigger sndbuf consumer: RDMA write into peer RMBE and CDC */ 364 if ((diff_cons && smc_tx_prepared_sends(conn)) || 365 conn->local_rx_ctrl.prod_flags.cons_curs_upd_req || 366 conn->local_rx_ctrl.prod_flags.urg_data_pending) { 367 if (!sock_owned_by_user(&smc->sk)) 368 smc_tx_pending(conn); 369 else 370 conn->tx_in_release_sock = true; 371 } 372 373 if (diff_cons && conn->urg_tx_pend && 374 atomic_read(&conn->peer_rmbe_space) == conn->peer_rmbe_size) { 375 /* urg data confirmed by peer, indicate we're ready for more */ 376 conn->urg_tx_pend = false; 377 smc->sk.sk_write_space(&smc->sk); 378 } 379 380 if (conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) { 381 smc->sk.sk_err = ECONNRESET; 382 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1; 383 } 384 if (smc_cdc_rxed_any_close_or_senddone(conn)) { 385 smc->sk.sk_shutdown |= RCV_SHUTDOWN; 386 if (smc->clcsock && smc->clcsock->sk) 387 smc->clcsock->sk->sk_shutdown |= RCV_SHUTDOWN; 388 sock_set_flag(&smc->sk, SOCK_DONE); 389 sock_hold(&smc->sk); /* sock_put in close_work */ 390 if (!queue_work(smc_close_wq, &conn->close_work)) 391 sock_put(&smc->sk); 392 } 393 } 394 395 /* called under tasklet context */ 396 static void smc_cdc_msg_recv(struct smc_sock *smc, struct smc_cdc_msg *cdc) 397 { 398 sock_hold(&smc->sk); 399 bh_lock_sock(&smc->sk); 400 smc_cdc_msg_recv_action(smc, cdc); 401 bh_unlock_sock(&smc->sk); 402 sock_put(&smc->sk); /* no free sk in softirq-context */ 403 } 404 405 /* Schedule a tasklet for this connection. Triggered from the ISM device IRQ 406 * handler to indicate update in the DMBE. 407 * 408 * Context: 409 * - tasklet context 410 */ 411 static void smcd_cdc_rx_tsklet(struct tasklet_struct *t) 412 { 413 struct smc_connection *conn = from_tasklet(conn, t, rx_tsklet); 414 struct smcd_cdc_msg *data_cdc; 415 struct smcd_cdc_msg cdc; 416 struct smc_sock *smc; 417 418 if (!conn || conn->killed) 419 return; 420 421 data_cdc = (struct smcd_cdc_msg *)conn->rmb_desc->cpu_addr; 422 smcd_curs_copy(&cdc.prod, &data_cdc->prod, conn); 423 smcd_curs_copy(&cdc.cons, &data_cdc->cons, conn); 424 smc = container_of(conn, struct smc_sock, conn); 425 smc_cdc_msg_recv(smc, (struct smc_cdc_msg *)&cdc); 426 } 427 428 /* Initialize receive tasklet. Called from ISM device IRQ handler to start 429 * receiver side. 430 */ 431 void smcd_cdc_rx_init(struct smc_connection *conn) 432 { 433 tasklet_setup(&conn->rx_tsklet, smcd_cdc_rx_tsklet); 434 } 435 436 /***************************** init, exit, misc ******************************/ 437 438 static void smc_cdc_rx_handler(struct ib_wc *wc, void *buf) 439 { 440 struct smc_link *link = (struct smc_link *)wc->qp->qp_context; 441 struct smc_cdc_msg *cdc = buf; 442 struct smc_connection *conn; 443 struct smc_link_group *lgr; 444 struct smc_sock *smc; 445 446 if (wc->byte_len < offsetof(struct smc_cdc_msg, reserved)) 447 return; /* short message */ 448 if (cdc->len != SMC_WR_TX_SIZE) 449 return; /* invalid message */ 450 451 /* lookup connection */ 452 lgr = smc_get_lgr(link); 453 read_lock_bh(&lgr->conns_lock); 454 conn = smc_lgr_find_conn(ntohl(cdc->token), lgr); 455 read_unlock_bh(&lgr->conns_lock); 456 if (!conn || conn->out_of_sync) 457 return; 458 smc = container_of(conn, struct smc_sock, conn); 459 460 if (cdc->prod_flags.failover_validation) { 461 smc_cdc_msg_validate(smc, cdc, link); 462 return; 463 } 464 if (smc_cdc_before(ntohs(cdc->seqno), 465 conn->local_rx_ctrl.seqno)) 466 /* received seqno is old */ 467 return; 468 469 smc_cdc_msg_recv(smc, cdc); 470 } 471 472 static struct smc_wr_rx_handler smc_cdc_rx_handlers[] = { 473 { 474 .handler = smc_cdc_rx_handler, 475 .type = SMC_CDC_MSG_TYPE 476 }, 477 { 478 .handler = NULL, 479 } 480 }; 481 482 int __init smc_cdc_init(void) 483 { 484 struct smc_wr_rx_handler *handler; 485 int rc = 0; 486 487 for (handler = smc_cdc_rx_handlers; handler->handler; handler++) { 488 INIT_HLIST_NODE(&handler->list); 489 rc = smc_wr_rx_register_handler(handler); 490 if (rc) 491 break; 492 } 493 return rc; 494 } 495