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 smc_cdc_add_pending_send(conn, pend); 118 119 conn->tx_cdc_seq++; 120 conn->local_tx_ctrl.seqno = conn->tx_cdc_seq; 121 smc_host_msg_to_cdc((struct smc_cdc_msg *)wr_buf, conn, &cfed); 122 123 atomic_inc(&conn->cdc_pend_tx_wr); 124 smp_mb__after_atomic(); /* Make sure cdc_pend_tx_wr added before post */ 125 126 rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend); 127 if (!rc) { 128 smc_curs_copy(&conn->rx_curs_confirmed, &cfed, conn); 129 conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0; 130 } else { 131 conn->tx_cdc_seq--; 132 conn->local_tx_ctrl.seqno = conn->tx_cdc_seq; 133 atomic_dec(&conn->cdc_pend_tx_wr); 134 } 135 136 return rc; 137 } 138 139 /* send a validation msg indicating the move of a conn to an other QP link */ 140 int smcr_cdc_msg_send_validation(struct smc_connection *conn, 141 struct smc_cdc_tx_pend *pend, 142 struct smc_wr_buf *wr_buf) 143 { 144 struct smc_host_cdc_msg *local = &conn->local_tx_ctrl; 145 struct smc_link *link = conn->lnk; 146 struct smc_cdc_msg *peer; 147 int rc; 148 149 peer = (struct smc_cdc_msg *)wr_buf; 150 peer->common.type = local->common.type; 151 peer->len = local->len; 152 peer->seqno = htons(conn->tx_cdc_seq_fin); /* seqno last compl. tx */ 153 peer->token = htonl(local->token); 154 peer->prod_flags.failover_validation = 1; 155 156 /* We need to set pend->conn here to make sure smc_cdc_tx_handler() 157 * can handle properly 158 */ 159 smc_cdc_add_pending_send(conn, pend); 160 161 atomic_inc(&conn->cdc_pend_tx_wr); 162 smp_mb__after_atomic(); /* Make sure cdc_pend_tx_wr added before post */ 163 164 rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend); 165 if (unlikely(rc)) 166 atomic_dec(&conn->cdc_pend_tx_wr); 167 168 return rc; 169 } 170 171 static int smcr_cdc_get_slot_and_msg_send(struct smc_connection *conn) 172 { 173 struct smc_cdc_tx_pend *pend; 174 struct smc_wr_buf *wr_buf; 175 struct smc_link *link; 176 bool again = false; 177 int rc; 178 179 again: 180 link = conn->lnk; 181 if (!smc_wr_tx_link_hold(link)) 182 return -ENOLINK; 183 rc = smc_cdc_get_free_slot(conn, link, &wr_buf, NULL, &pend); 184 if (rc) 185 goto put_out; 186 187 spin_lock_bh(&conn->send_lock); 188 if (link != conn->lnk) { 189 /* link of connection changed, try again one time*/ 190 spin_unlock_bh(&conn->send_lock); 191 smc_wr_tx_put_slot(link, 192 (struct smc_wr_tx_pend_priv *)pend); 193 smc_wr_tx_link_put(link); 194 if (again) 195 return -ENOLINK; 196 again = true; 197 goto again; 198 } 199 rc = smc_cdc_msg_send(conn, wr_buf, pend); 200 spin_unlock_bh(&conn->send_lock); 201 put_out: 202 smc_wr_tx_link_put(link); 203 return rc; 204 } 205 206 int smc_cdc_get_slot_and_msg_send(struct smc_connection *conn) 207 { 208 int rc; 209 210 if (!smc_conn_lgr_valid(conn) || 211 (conn->lgr->is_smcd && conn->lgr->peer_shutdown)) 212 return -EPIPE; 213 214 if (conn->lgr->is_smcd) { 215 spin_lock_bh(&conn->send_lock); 216 rc = smcd_cdc_msg_send(conn); 217 spin_unlock_bh(&conn->send_lock); 218 } else { 219 rc = smcr_cdc_get_slot_and_msg_send(conn); 220 } 221 222 return rc; 223 } 224 225 void smc_cdc_wait_pend_tx_wr(struct smc_connection *conn) 226 { 227 wait_event(conn->cdc_pend_tx_wq, !atomic_read(&conn->cdc_pend_tx_wr)); 228 } 229 230 /* Send a SMC-D CDC header. 231 * This increments the free space available in our send buffer. 232 * Also update the confirmed receive buffer with what was sent to the peer. 233 */ 234 int smcd_cdc_msg_send(struct smc_connection *conn) 235 { 236 struct smc_sock *smc = container_of(conn, struct smc_sock, conn); 237 union smc_host_cursor curs; 238 struct smcd_cdc_msg cdc; 239 int rc, diff; 240 241 memset(&cdc, 0, sizeof(cdc)); 242 cdc.common.type = SMC_CDC_MSG_TYPE; 243 curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.prod.acurs); 244 cdc.prod.wrap = curs.wrap; 245 cdc.prod.count = curs.count; 246 curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.cons.acurs); 247 cdc.cons.wrap = curs.wrap; 248 cdc.cons.count = curs.count; 249 cdc.cons.prod_flags = conn->local_tx_ctrl.prod_flags; 250 cdc.cons.conn_state_flags = conn->local_tx_ctrl.conn_state_flags; 251 rc = smcd_tx_ism_write(conn, &cdc, sizeof(cdc), 0, 1); 252 if (rc) 253 return rc; 254 smc_curs_copy(&conn->rx_curs_confirmed, &curs, conn); 255 conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0; 256 /* Calculate transmitted data and increment free send buffer space */ 257 diff = smc_curs_diff(conn->sndbuf_desc->len, &conn->tx_curs_fin, 258 &conn->tx_curs_sent); 259 /* increased by confirmed number of bytes */ 260 smp_mb__before_atomic(); 261 atomic_add(diff, &conn->sndbuf_space); 262 /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */ 263 smp_mb__after_atomic(); 264 smc_curs_copy(&conn->tx_curs_fin, &conn->tx_curs_sent, conn); 265 266 smc_tx_sndbuf_nonfull(smc); 267 return rc; 268 } 269 270 /********************************* receive ***********************************/ 271 272 static inline bool smc_cdc_before(u16 seq1, u16 seq2) 273 { 274 return (s16)(seq1 - seq2) < 0; 275 } 276 277 static void smc_cdc_handle_urg_data_arrival(struct smc_sock *smc, 278 int *diff_prod) 279 { 280 struct smc_connection *conn = &smc->conn; 281 char *base; 282 283 /* new data included urgent business */ 284 smc_curs_copy(&conn->urg_curs, &conn->local_rx_ctrl.prod, conn); 285 conn->urg_state = SMC_URG_VALID; 286 if (!sock_flag(&smc->sk, SOCK_URGINLINE)) 287 /* we'll skip the urgent byte, so don't account for it */ 288 (*diff_prod)--; 289 base = (char *)conn->rmb_desc->cpu_addr + conn->rx_off; 290 if (conn->urg_curs.count) 291 conn->urg_rx_byte = *(base + conn->urg_curs.count - 1); 292 else 293 conn->urg_rx_byte = *(base + conn->rmb_desc->len - 1); 294 sk_send_sigurg(&smc->sk); 295 } 296 297 static void smc_cdc_msg_validate(struct smc_sock *smc, struct smc_cdc_msg *cdc, 298 struct smc_link *link) 299 { 300 struct smc_connection *conn = &smc->conn; 301 u16 recv_seq = ntohs(cdc->seqno); 302 s16 diff; 303 304 /* check that seqnum was seen before */ 305 diff = conn->local_rx_ctrl.seqno - recv_seq; 306 if (diff < 0) { /* diff larger than 0x7fff */ 307 /* drop connection */ 308 conn->out_of_sync = 1; /* prevent any further receives */ 309 spin_lock_bh(&conn->send_lock); 310 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1; 311 conn->lnk = link; 312 spin_unlock_bh(&conn->send_lock); 313 sock_hold(&smc->sk); /* sock_put in abort_work */ 314 if (!queue_work(smc_close_wq, &conn->abort_work)) 315 sock_put(&smc->sk); 316 } 317 } 318 319 static void smc_cdc_msg_recv_action(struct smc_sock *smc, 320 struct smc_cdc_msg *cdc) 321 { 322 union smc_host_cursor cons_old, prod_old; 323 struct smc_connection *conn = &smc->conn; 324 int diff_cons, diff_prod; 325 326 smc_curs_copy(&prod_old, &conn->local_rx_ctrl.prod, conn); 327 smc_curs_copy(&cons_old, &conn->local_rx_ctrl.cons, conn); 328 smc_cdc_msg_to_host(&conn->local_rx_ctrl, cdc, conn); 329 330 diff_cons = smc_curs_diff(conn->peer_rmbe_size, &cons_old, 331 &conn->local_rx_ctrl.cons); 332 if (diff_cons) { 333 /* peer_rmbe_space is decreased during data transfer with RDMA 334 * write 335 */ 336 smp_mb__before_atomic(); 337 atomic_add(diff_cons, &conn->peer_rmbe_space); 338 /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */ 339 smp_mb__after_atomic(); 340 } 341 342 diff_prod = smc_curs_diff(conn->rmb_desc->len, &prod_old, 343 &conn->local_rx_ctrl.prod); 344 if (diff_prod) { 345 if (conn->local_rx_ctrl.prod_flags.urg_data_present) 346 smc_cdc_handle_urg_data_arrival(smc, &diff_prod); 347 /* bytes_to_rcv is decreased in smc_recvmsg */ 348 smp_mb__before_atomic(); 349 atomic_add(diff_prod, &conn->bytes_to_rcv); 350 /* guarantee 0 <= bytes_to_rcv <= rmb_desc->len */ 351 smp_mb__after_atomic(); 352 smc->sk.sk_data_ready(&smc->sk); 353 } else { 354 if (conn->local_rx_ctrl.prod_flags.write_blocked) 355 smc->sk.sk_data_ready(&smc->sk); 356 if (conn->local_rx_ctrl.prod_flags.urg_data_pending) 357 conn->urg_state = SMC_URG_NOTYET; 358 } 359 360 /* trigger sndbuf consumer: RDMA write into peer RMBE and CDC */ 361 if ((diff_cons && smc_tx_prepared_sends(conn)) || 362 conn->local_rx_ctrl.prod_flags.cons_curs_upd_req || 363 conn->local_rx_ctrl.prod_flags.urg_data_pending) { 364 if (!sock_owned_by_user(&smc->sk)) 365 smc_tx_pending(conn); 366 else 367 conn->tx_in_release_sock = true; 368 } 369 370 if (diff_cons && conn->urg_tx_pend && 371 atomic_read(&conn->peer_rmbe_space) == conn->peer_rmbe_size) { 372 /* urg data confirmed by peer, indicate we're ready for more */ 373 conn->urg_tx_pend = false; 374 smc->sk.sk_write_space(&smc->sk); 375 } 376 377 if (conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) { 378 smc->sk.sk_err = ECONNRESET; 379 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1; 380 } 381 if (smc_cdc_rxed_any_close_or_senddone(conn)) { 382 smc->sk.sk_shutdown |= RCV_SHUTDOWN; 383 if (smc->clcsock && smc->clcsock->sk) 384 smc->clcsock->sk->sk_shutdown |= RCV_SHUTDOWN; 385 sock_set_flag(&smc->sk, SOCK_DONE); 386 sock_hold(&smc->sk); /* sock_put in close_work */ 387 if (!queue_work(smc_close_wq, &conn->close_work)) 388 sock_put(&smc->sk); 389 } 390 } 391 392 /* called under tasklet context */ 393 static void smc_cdc_msg_recv(struct smc_sock *smc, struct smc_cdc_msg *cdc) 394 { 395 sock_hold(&smc->sk); 396 bh_lock_sock(&smc->sk); 397 smc_cdc_msg_recv_action(smc, cdc); 398 bh_unlock_sock(&smc->sk); 399 sock_put(&smc->sk); /* no free sk in softirq-context */ 400 } 401 402 /* Schedule a tasklet for this connection. Triggered from the ISM device IRQ 403 * handler to indicate update in the DMBE. 404 * 405 * Context: 406 * - tasklet context 407 */ 408 static void smcd_cdc_rx_tsklet(struct tasklet_struct *t) 409 { 410 struct smc_connection *conn = from_tasklet(conn, t, rx_tsklet); 411 struct smcd_cdc_msg *data_cdc; 412 struct smcd_cdc_msg cdc; 413 struct smc_sock *smc; 414 415 if (!conn || conn->killed) 416 return; 417 418 data_cdc = (struct smcd_cdc_msg *)conn->rmb_desc->cpu_addr; 419 smcd_curs_copy(&cdc.prod, &data_cdc->prod, conn); 420 smcd_curs_copy(&cdc.cons, &data_cdc->cons, conn); 421 smc = container_of(conn, struct smc_sock, conn); 422 smc_cdc_msg_recv(smc, (struct smc_cdc_msg *)&cdc); 423 } 424 425 /* Initialize receive tasklet. Called from ISM device IRQ handler to start 426 * receiver side. 427 */ 428 void smcd_cdc_rx_init(struct smc_connection *conn) 429 { 430 tasklet_setup(&conn->rx_tsklet, smcd_cdc_rx_tsklet); 431 } 432 433 /***************************** init, exit, misc ******************************/ 434 435 static void smc_cdc_rx_handler(struct ib_wc *wc, void *buf) 436 { 437 struct smc_link *link = (struct smc_link *)wc->qp->qp_context; 438 struct smc_cdc_msg *cdc = buf; 439 struct smc_connection *conn; 440 struct smc_link_group *lgr; 441 struct smc_sock *smc; 442 443 if (wc->byte_len < offsetof(struct smc_cdc_msg, reserved)) 444 return; /* short message */ 445 if (cdc->len != SMC_WR_TX_SIZE) 446 return; /* invalid message */ 447 448 /* lookup connection */ 449 lgr = smc_get_lgr(link); 450 read_lock_bh(&lgr->conns_lock); 451 conn = smc_lgr_find_conn(ntohl(cdc->token), lgr); 452 read_unlock_bh(&lgr->conns_lock); 453 if (!conn || conn->out_of_sync) 454 return; 455 smc = container_of(conn, struct smc_sock, conn); 456 457 if (cdc->prod_flags.failover_validation) { 458 smc_cdc_msg_validate(smc, cdc, link); 459 return; 460 } 461 if (smc_cdc_before(ntohs(cdc->seqno), 462 conn->local_rx_ctrl.seqno)) 463 /* received seqno is old */ 464 return; 465 466 smc_cdc_msg_recv(smc, cdc); 467 } 468 469 static struct smc_wr_rx_handler smc_cdc_rx_handlers[] = { 470 { 471 .handler = smc_cdc_rx_handler, 472 .type = SMC_CDC_MSG_TYPE 473 }, 474 { 475 .handler = NULL, 476 } 477 }; 478 479 int __init smc_cdc_init(void) 480 { 481 struct smc_wr_rx_handler *handler; 482 int rc = 0; 483 484 for (handler = smc_cdc_rx_handlers; handler->handler; handler++) { 485 INIT_HLIST_NODE(&handler->list); 486 rc = smc_wr_rx_register_handler(handler); 487 if (rc) 488 break; 489 } 490 return rc; 491 } 492