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 unlikely(wq_has_sleeper(&conn->cdc_pend_tx_wq))) 53 wake_up(&conn->cdc_pend_tx_wq); 54 WARN_ON(atomic_read(&conn->cdc_pend_tx_wr) < 0); 55 56 smc_tx_sndbuf_nonfull(smc); 57 bh_unlock_sock(&smc->sk); 58 } 59 60 int smc_cdc_get_free_slot(struct smc_connection *conn, 61 struct smc_link *link, 62 struct smc_wr_buf **wr_buf, 63 struct smc_rdma_wr **wr_rdma_buf, 64 struct smc_cdc_tx_pend **pend) 65 { 66 int rc; 67 68 rc = smc_wr_tx_get_free_slot(link, smc_cdc_tx_handler, wr_buf, 69 wr_rdma_buf, 70 (struct smc_wr_tx_pend_priv **)pend); 71 if (conn->killed) { 72 /* abnormal termination */ 73 if (!rc) 74 smc_wr_tx_put_slot(link, 75 (struct smc_wr_tx_pend_priv *)pend); 76 rc = -EPIPE; 77 } 78 return rc; 79 } 80 81 static inline void smc_cdc_add_pending_send(struct smc_connection *conn, 82 struct smc_cdc_tx_pend *pend) 83 { 84 BUILD_BUG_ON_MSG( 85 sizeof(struct smc_cdc_msg) > SMC_WR_BUF_SIZE, 86 "must increase SMC_WR_BUF_SIZE to at least sizeof(struct smc_cdc_msg)"); 87 BUILD_BUG_ON_MSG( 88 offsetofend(struct smc_cdc_msg, reserved) > SMC_WR_TX_SIZE, 89 "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()"); 90 BUILD_BUG_ON_MSG( 91 sizeof(struct smc_cdc_tx_pend) > SMC_WR_TX_PEND_PRIV_SIZE, 92 "must increase SMC_WR_TX_PEND_PRIV_SIZE to at least sizeof(struct smc_cdc_tx_pend)"); 93 pend->conn = conn; 94 pend->cursor = conn->tx_curs_sent; 95 pend->p_cursor = conn->local_tx_ctrl.prod; 96 pend->ctrl_seq = conn->tx_cdc_seq; 97 } 98 99 int smc_cdc_msg_send(struct smc_connection *conn, 100 struct smc_wr_buf *wr_buf, 101 struct smc_cdc_tx_pend *pend) 102 { 103 struct smc_link *link = conn->lnk; 104 union smc_host_cursor cfed; 105 int rc; 106 107 smc_cdc_add_pending_send(conn, pend); 108 109 conn->tx_cdc_seq++; 110 conn->local_tx_ctrl.seqno = conn->tx_cdc_seq; 111 smc_host_msg_to_cdc((struct smc_cdc_msg *)wr_buf, conn, &cfed); 112 113 atomic_inc(&conn->cdc_pend_tx_wr); 114 smp_mb__after_atomic(); /* Make sure cdc_pend_tx_wr added before post */ 115 116 rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend); 117 if (!rc) { 118 smc_curs_copy(&conn->rx_curs_confirmed, &cfed, conn); 119 conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0; 120 } else { 121 conn->tx_cdc_seq--; 122 conn->local_tx_ctrl.seqno = conn->tx_cdc_seq; 123 atomic_dec(&conn->cdc_pend_tx_wr); 124 } 125 126 return rc; 127 } 128 129 /* send a validation msg indicating the move of a conn to an other QP link */ 130 int smcr_cdc_msg_send_validation(struct smc_connection *conn, 131 struct smc_cdc_tx_pend *pend, 132 struct smc_wr_buf *wr_buf) 133 { 134 struct smc_host_cdc_msg *local = &conn->local_tx_ctrl; 135 struct smc_link *link = conn->lnk; 136 struct smc_cdc_msg *peer; 137 int rc; 138 139 peer = (struct smc_cdc_msg *)wr_buf; 140 peer->common.type = local->common.type; 141 peer->len = local->len; 142 peer->seqno = htons(conn->tx_cdc_seq_fin); /* seqno last compl. tx */ 143 peer->token = htonl(local->token); 144 peer->prod_flags.failover_validation = 1; 145 146 /* We need to set pend->conn here to make sure smc_cdc_tx_handler() 147 * can handle properly 148 */ 149 smc_cdc_add_pending_send(conn, pend); 150 151 atomic_inc(&conn->cdc_pend_tx_wr); 152 smp_mb__after_atomic(); /* Make sure cdc_pend_tx_wr added before post */ 153 154 rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend); 155 if (unlikely(rc)) 156 atomic_dec(&conn->cdc_pend_tx_wr); 157 158 return rc; 159 } 160 161 static int smcr_cdc_get_slot_and_msg_send(struct smc_connection *conn) 162 { 163 struct smc_cdc_tx_pend *pend; 164 struct smc_wr_buf *wr_buf; 165 struct smc_link *link; 166 bool again = false; 167 int rc; 168 169 again: 170 link = conn->lnk; 171 if (!smc_wr_tx_link_hold(link)) 172 return -ENOLINK; 173 rc = smc_cdc_get_free_slot(conn, link, &wr_buf, NULL, &pend); 174 if (rc) 175 goto put_out; 176 177 spin_lock_bh(&conn->send_lock); 178 if (link != conn->lnk) { 179 /* link of connection changed, try again one time*/ 180 spin_unlock_bh(&conn->send_lock); 181 smc_wr_tx_put_slot(link, 182 (struct smc_wr_tx_pend_priv *)pend); 183 smc_wr_tx_link_put(link); 184 if (again) 185 return -ENOLINK; 186 again = true; 187 goto again; 188 } 189 rc = smc_cdc_msg_send(conn, wr_buf, pend); 190 spin_unlock_bh(&conn->send_lock); 191 put_out: 192 smc_wr_tx_link_put(link); 193 return rc; 194 } 195 196 int smc_cdc_get_slot_and_msg_send(struct smc_connection *conn) 197 { 198 int rc; 199 200 if (!conn->lgr || (conn->lgr->is_smcd && conn->lgr->peer_shutdown)) 201 return -EPIPE; 202 203 if (conn->lgr->is_smcd) { 204 spin_lock_bh(&conn->send_lock); 205 rc = smcd_cdc_msg_send(conn); 206 spin_unlock_bh(&conn->send_lock); 207 } else { 208 rc = smcr_cdc_get_slot_and_msg_send(conn); 209 } 210 211 return rc; 212 } 213 214 void smc_cdc_wait_pend_tx_wr(struct smc_connection *conn) 215 { 216 wait_event(conn->cdc_pend_tx_wq, !atomic_read(&conn->cdc_pend_tx_wr)); 217 } 218 219 /* Send a SMC-D CDC header. 220 * This increments the free space available in our send buffer. 221 * Also update the confirmed receive buffer with what was sent to the peer. 222 */ 223 int smcd_cdc_msg_send(struct smc_connection *conn) 224 { 225 struct smc_sock *smc = container_of(conn, struct smc_sock, conn); 226 union smc_host_cursor curs; 227 struct smcd_cdc_msg cdc; 228 int rc, diff; 229 230 memset(&cdc, 0, sizeof(cdc)); 231 cdc.common.type = SMC_CDC_MSG_TYPE; 232 curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.prod.acurs); 233 cdc.prod.wrap = curs.wrap; 234 cdc.prod.count = curs.count; 235 curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.cons.acurs); 236 cdc.cons.wrap = curs.wrap; 237 cdc.cons.count = curs.count; 238 cdc.cons.prod_flags = conn->local_tx_ctrl.prod_flags; 239 cdc.cons.conn_state_flags = conn->local_tx_ctrl.conn_state_flags; 240 rc = smcd_tx_ism_write(conn, &cdc, sizeof(cdc), 0, 1); 241 if (rc) 242 return rc; 243 smc_curs_copy(&conn->rx_curs_confirmed, &curs, conn); 244 conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0; 245 /* Calculate transmitted data and increment free send buffer space */ 246 diff = smc_curs_diff(conn->sndbuf_desc->len, &conn->tx_curs_fin, 247 &conn->tx_curs_sent); 248 /* increased by confirmed number of bytes */ 249 smp_mb__before_atomic(); 250 atomic_add(diff, &conn->sndbuf_space); 251 /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */ 252 smp_mb__after_atomic(); 253 smc_curs_copy(&conn->tx_curs_fin, &conn->tx_curs_sent, conn); 254 255 smc_tx_sndbuf_nonfull(smc); 256 return rc; 257 } 258 259 /********************************* receive ***********************************/ 260 261 static inline bool smc_cdc_before(u16 seq1, u16 seq2) 262 { 263 return (s16)(seq1 - seq2) < 0; 264 } 265 266 static void smc_cdc_handle_urg_data_arrival(struct smc_sock *smc, 267 int *diff_prod) 268 { 269 struct smc_connection *conn = &smc->conn; 270 char *base; 271 272 /* new data included urgent business */ 273 smc_curs_copy(&conn->urg_curs, &conn->local_rx_ctrl.prod, conn); 274 conn->urg_state = SMC_URG_VALID; 275 if (!sock_flag(&smc->sk, SOCK_URGINLINE)) 276 /* we'll skip the urgent byte, so don't account for it */ 277 (*diff_prod)--; 278 base = (char *)conn->rmb_desc->cpu_addr + conn->rx_off; 279 if (conn->urg_curs.count) 280 conn->urg_rx_byte = *(base + conn->urg_curs.count - 1); 281 else 282 conn->urg_rx_byte = *(base + conn->rmb_desc->len - 1); 283 sk_send_sigurg(&smc->sk); 284 } 285 286 static void smc_cdc_msg_validate(struct smc_sock *smc, struct smc_cdc_msg *cdc, 287 struct smc_link *link) 288 { 289 struct smc_connection *conn = &smc->conn; 290 u16 recv_seq = ntohs(cdc->seqno); 291 s16 diff; 292 293 /* check that seqnum was seen before */ 294 diff = conn->local_rx_ctrl.seqno - recv_seq; 295 if (diff < 0) { /* diff larger than 0x7fff */ 296 /* drop connection */ 297 conn->out_of_sync = 1; /* prevent any further receives */ 298 spin_lock_bh(&conn->send_lock); 299 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1; 300 conn->lnk = link; 301 spin_unlock_bh(&conn->send_lock); 302 sock_hold(&smc->sk); /* sock_put in abort_work */ 303 if (!queue_work(smc_close_wq, &conn->abort_work)) 304 sock_put(&smc->sk); 305 } 306 } 307 308 static void smc_cdc_msg_recv_action(struct smc_sock *smc, 309 struct smc_cdc_msg *cdc) 310 { 311 union smc_host_cursor cons_old, prod_old; 312 struct smc_connection *conn = &smc->conn; 313 int diff_cons, diff_prod; 314 315 smc_curs_copy(&prod_old, &conn->local_rx_ctrl.prod, conn); 316 smc_curs_copy(&cons_old, &conn->local_rx_ctrl.cons, conn); 317 smc_cdc_msg_to_host(&conn->local_rx_ctrl, cdc, conn); 318 319 diff_cons = smc_curs_diff(conn->peer_rmbe_size, &cons_old, 320 &conn->local_rx_ctrl.cons); 321 if (diff_cons) { 322 /* peer_rmbe_space is decreased during data transfer with RDMA 323 * write 324 */ 325 smp_mb__before_atomic(); 326 atomic_add(diff_cons, &conn->peer_rmbe_space); 327 /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */ 328 smp_mb__after_atomic(); 329 } 330 331 diff_prod = smc_curs_diff(conn->rmb_desc->len, &prod_old, 332 &conn->local_rx_ctrl.prod); 333 if (diff_prod) { 334 if (conn->local_rx_ctrl.prod_flags.urg_data_present) 335 smc_cdc_handle_urg_data_arrival(smc, &diff_prod); 336 /* bytes_to_rcv is decreased in smc_recvmsg */ 337 smp_mb__before_atomic(); 338 atomic_add(diff_prod, &conn->bytes_to_rcv); 339 /* guarantee 0 <= bytes_to_rcv <= rmb_desc->len */ 340 smp_mb__after_atomic(); 341 smc->sk.sk_data_ready(&smc->sk); 342 } else { 343 if (conn->local_rx_ctrl.prod_flags.write_blocked) 344 smc->sk.sk_data_ready(&smc->sk); 345 if (conn->local_rx_ctrl.prod_flags.urg_data_pending) 346 conn->urg_state = SMC_URG_NOTYET; 347 } 348 349 /* trigger sndbuf consumer: RDMA write into peer RMBE and CDC */ 350 if ((diff_cons && smc_tx_prepared_sends(conn)) || 351 conn->local_rx_ctrl.prod_flags.cons_curs_upd_req || 352 conn->local_rx_ctrl.prod_flags.urg_data_pending) 353 smc_tx_sndbuf_nonempty(conn); 354 355 if (diff_cons && conn->urg_tx_pend && 356 atomic_read(&conn->peer_rmbe_space) == conn->peer_rmbe_size) { 357 /* urg data confirmed by peer, indicate we're ready for more */ 358 conn->urg_tx_pend = false; 359 smc->sk.sk_write_space(&smc->sk); 360 } 361 362 if (conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) { 363 smc->sk.sk_err = ECONNRESET; 364 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1; 365 } 366 if (smc_cdc_rxed_any_close_or_senddone(conn)) { 367 smc->sk.sk_shutdown |= RCV_SHUTDOWN; 368 if (smc->clcsock && smc->clcsock->sk) 369 smc->clcsock->sk->sk_shutdown |= RCV_SHUTDOWN; 370 sock_set_flag(&smc->sk, SOCK_DONE); 371 sock_hold(&smc->sk); /* sock_put in close_work */ 372 if (!queue_work(smc_close_wq, &conn->close_work)) 373 sock_put(&smc->sk); 374 } 375 } 376 377 /* called under tasklet context */ 378 static void smc_cdc_msg_recv(struct smc_sock *smc, struct smc_cdc_msg *cdc) 379 { 380 sock_hold(&smc->sk); 381 bh_lock_sock(&smc->sk); 382 smc_cdc_msg_recv_action(smc, cdc); 383 bh_unlock_sock(&smc->sk); 384 sock_put(&smc->sk); /* no free sk in softirq-context */ 385 } 386 387 /* Schedule a tasklet for this connection. Triggered from the ISM device IRQ 388 * handler to indicate update in the DMBE. 389 * 390 * Context: 391 * - tasklet context 392 */ 393 static void smcd_cdc_rx_tsklet(struct tasklet_struct *t) 394 { 395 struct smc_connection *conn = from_tasklet(conn, t, rx_tsklet); 396 struct smcd_cdc_msg *data_cdc; 397 struct smcd_cdc_msg cdc; 398 struct smc_sock *smc; 399 400 if (!conn || conn->killed) 401 return; 402 403 data_cdc = (struct smcd_cdc_msg *)conn->rmb_desc->cpu_addr; 404 smcd_curs_copy(&cdc.prod, &data_cdc->prod, conn); 405 smcd_curs_copy(&cdc.cons, &data_cdc->cons, conn); 406 smc = container_of(conn, struct smc_sock, conn); 407 smc_cdc_msg_recv(smc, (struct smc_cdc_msg *)&cdc); 408 } 409 410 /* Initialize receive tasklet. Called from ISM device IRQ handler to start 411 * receiver side. 412 */ 413 void smcd_cdc_rx_init(struct smc_connection *conn) 414 { 415 tasklet_setup(&conn->rx_tsklet, smcd_cdc_rx_tsklet); 416 } 417 418 /***************************** init, exit, misc ******************************/ 419 420 static void smc_cdc_rx_handler(struct ib_wc *wc, void *buf) 421 { 422 struct smc_link *link = (struct smc_link *)wc->qp->qp_context; 423 struct smc_cdc_msg *cdc = buf; 424 struct smc_connection *conn; 425 struct smc_link_group *lgr; 426 struct smc_sock *smc; 427 428 if (wc->byte_len < offsetof(struct smc_cdc_msg, reserved)) 429 return; /* short message */ 430 if (cdc->len != SMC_WR_TX_SIZE) 431 return; /* invalid message */ 432 433 /* lookup connection */ 434 lgr = smc_get_lgr(link); 435 read_lock_bh(&lgr->conns_lock); 436 conn = smc_lgr_find_conn(ntohl(cdc->token), lgr); 437 read_unlock_bh(&lgr->conns_lock); 438 if (!conn || conn->out_of_sync) 439 return; 440 smc = container_of(conn, struct smc_sock, conn); 441 442 if (cdc->prod_flags.failover_validation) { 443 smc_cdc_msg_validate(smc, cdc, link); 444 return; 445 } 446 if (smc_cdc_before(ntohs(cdc->seqno), 447 conn->local_rx_ctrl.seqno)) 448 /* received seqno is old */ 449 return; 450 451 smc_cdc_msg_recv(smc, cdc); 452 } 453 454 static struct smc_wr_rx_handler smc_cdc_rx_handlers[] = { 455 { 456 .handler = smc_cdc_rx_handler, 457 .type = SMC_CDC_MSG_TYPE 458 }, 459 { 460 .handler = NULL, 461 } 462 }; 463 464 int __init smc_cdc_init(void) 465 { 466 struct smc_wr_rx_handler *handler; 467 int rc = 0; 468 469 for (handler = smc_cdc_rx_handlers; handler->handler; handler++) { 470 INIT_HLIST_NODE(&handler->list); 471 rc = smc_wr_rx_register_handler(handler); 472 if (rc) 473 break; 474 } 475 return rc; 476 } 477