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 if (!conn) 35 /* already dismissed */ 36 return; 37 38 smc = container_of(conn, struct smc_sock, conn); 39 bh_lock_sock(&smc->sk); 40 if (!wc_status) { 41 diff = smc_curs_diff(cdcpend->conn->sndbuf_desc->len, 42 &cdcpend->conn->tx_curs_fin, 43 &cdcpend->cursor); 44 /* sndbuf_space is decreased in smc_sendmsg */ 45 smp_mb__before_atomic(); 46 atomic_add(diff, &cdcpend->conn->sndbuf_space); 47 /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */ 48 smp_mb__after_atomic(); 49 smc_curs_copy(&conn->tx_curs_fin, &cdcpend->cursor, conn); 50 } 51 smc_tx_sndbuf_nonfull(smc); 52 bh_unlock_sock(&smc->sk); 53 } 54 55 int smc_cdc_get_free_slot(struct smc_connection *conn, 56 struct smc_wr_buf **wr_buf, 57 struct smc_rdma_wr **wr_rdma_buf, 58 struct smc_cdc_tx_pend **pend) 59 { 60 struct smc_link *link = &conn->lgr->lnk[SMC_SINGLE_LINK]; 61 int rc; 62 63 rc = smc_wr_tx_get_free_slot(link, smc_cdc_tx_handler, wr_buf, 64 wr_rdma_buf, 65 (struct smc_wr_tx_pend_priv **)pend); 66 if (!conn->alert_token_local) 67 /* abnormal termination */ 68 rc = -EPIPE; 69 return rc; 70 } 71 72 static inline void smc_cdc_add_pending_send(struct smc_connection *conn, 73 struct smc_cdc_tx_pend *pend) 74 { 75 BUILD_BUG_ON_MSG( 76 sizeof(struct smc_cdc_msg) > SMC_WR_BUF_SIZE, 77 "must increase SMC_WR_BUF_SIZE to at least sizeof(struct smc_cdc_msg)"); 78 BUILD_BUG_ON_MSG( 79 offsetofend(struct smc_cdc_msg, reserved) > SMC_WR_TX_SIZE, 80 "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()"); 81 BUILD_BUG_ON_MSG( 82 sizeof(struct smc_cdc_tx_pend) > SMC_WR_TX_PEND_PRIV_SIZE, 83 "must increase SMC_WR_TX_PEND_PRIV_SIZE to at least sizeof(struct smc_cdc_tx_pend)"); 84 pend->conn = conn; 85 pend->cursor = conn->tx_curs_sent; 86 pend->p_cursor = conn->local_tx_ctrl.prod; 87 pend->ctrl_seq = conn->tx_cdc_seq; 88 } 89 90 int smc_cdc_msg_send(struct smc_connection *conn, 91 struct smc_wr_buf *wr_buf, 92 struct smc_cdc_tx_pend *pend) 93 { 94 union smc_host_cursor cfed; 95 struct smc_link *link; 96 int rc; 97 98 link = &conn->lgr->lnk[SMC_SINGLE_LINK]; 99 100 smc_cdc_add_pending_send(conn, pend); 101 102 conn->tx_cdc_seq++; 103 conn->local_tx_ctrl.seqno = conn->tx_cdc_seq; 104 smc_host_msg_to_cdc((struct smc_cdc_msg *)wr_buf, 105 &conn->local_tx_ctrl, conn); 106 smc_curs_copy(&cfed, &((struct smc_host_cdc_msg *)wr_buf)->cons, conn); 107 rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend); 108 if (!rc) 109 smc_curs_copy(&conn->rx_curs_confirmed, &cfed, conn); 110 111 return rc; 112 } 113 114 static int smcr_cdc_get_slot_and_msg_send(struct smc_connection *conn) 115 { 116 struct smc_cdc_tx_pend *pend; 117 struct smc_wr_buf *wr_buf; 118 int rc; 119 120 rc = smc_cdc_get_free_slot(conn, &wr_buf, NULL, &pend); 121 if (rc) 122 return rc; 123 124 spin_lock_bh(&conn->send_lock); 125 rc = smc_cdc_msg_send(conn, wr_buf, pend); 126 spin_unlock_bh(&conn->send_lock); 127 return rc; 128 } 129 130 int smc_cdc_get_slot_and_msg_send(struct smc_connection *conn) 131 { 132 int rc; 133 134 if (conn->lgr->is_smcd) { 135 spin_lock_bh(&conn->send_lock); 136 rc = smcd_cdc_msg_send(conn); 137 spin_unlock_bh(&conn->send_lock); 138 } else { 139 rc = smcr_cdc_get_slot_and_msg_send(conn); 140 } 141 142 return rc; 143 } 144 145 static bool smc_cdc_tx_filter(struct smc_wr_tx_pend_priv *tx_pend, 146 unsigned long data) 147 { 148 struct smc_connection *conn = (struct smc_connection *)data; 149 struct smc_cdc_tx_pend *cdc_pend = 150 (struct smc_cdc_tx_pend *)tx_pend; 151 152 return cdc_pend->conn == conn; 153 } 154 155 static void smc_cdc_tx_dismisser(struct smc_wr_tx_pend_priv *tx_pend) 156 { 157 struct smc_cdc_tx_pend *cdc_pend = 158 (struct smc_cdc_tx_pend *)tx_pend; 159 160 cdc_pend->conn = NULL; 161 } 162 163 void smc_cdc_tx_dismiss_slots(struct smc_connection *conn) 164 { 165 struct smc_link *link = &conn->lgr->lnk[SMC_SINGLE_LINK]; 166 167 smc_wr_tx_dismiss_slots(link, SMC_CDC_MSG_TYPE, 168 smc_cdc_tx_filter, smc_cdc_tx_dismisser, 169 (unsigned long)conn); 170 } 171 172 /* Send a SMC-D CDC header. 173 * This increments the free space available in our send buffer. 174 * Also update the confirmed receive buffer with what was sent to the peer. 175 */ 176 int smcd_cdc_msg_send(struct smc_connection *conn) 177 { 178 struct smc_sock *smc = container_of(conn, struct smc_sock, conn); 179 union smc_host_cursor curs; 180 struct smcd_cdc_msg cdc; 181 int rc, diff; 182 183 memset(&cdc, 0, sizeof(cdc)); 184 cdc.common.type = SMC_CDC_MSG_TYPE; 185 curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.prod.acurs); 186 cdc.prod.wrap = curs.wrap; 187 cdc.prod.count = curs.count; 188 curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.cons.acurs); 189 cdc.cons.wrap = curs.wrap; 190 cdc.cons.count = curs.count; 191 cdc.cons.prod_flags = conn->local_tx_ctrl.prod_flags; 192 cdc.cons.conn_state_flags = conn->local_tx_ctrl.conn_state_flags; 193 rc = smcd_tx_ism_write(conn, &cdc, sizeof(cdc), 0, 1); 194 if (rc) 195 return rc; 196 smc_curs_copy(&conn->rx_curs_confirmed, &curs, conn); 197 /* Calculate transmitted data and increment free send buffer space */ 198 diff = smc_curs_diff(conn->sndbuf_desc->len, &conn->tx_curs_fin, 199 &conn->tx_curs_sent); 200 /* increased by confirmed number of bytes */ 201 smp_mb__before_atomic(); 202 atomic_add(diff, &conn->sndbuf_space); 203 /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */ 204 smp_mb__after_atomic(); 205 smc_curs_copy(&conn->tx_curs_fin, &conn->tx_curs_sent, conn); 206 207 smc_tx_sndbuf_nonfull(smc); 208 return rc; 209 } 210 211 /********************************* receive ***********************************/ 212 213 static inline bool smc_cdc_before(u16 seq1, u16 seq2) 214 { 215 return (s16)(seq1 - seq2) < 0; 216 } 217 218 static void smc_cdc_handle_urg_data_arrival(struct smc_sock *smc, 219 int *diff_prod) 220 { 221 struct smc_connection *conn = &smc->conn; 222 char *base; 223 224 /* new data included urgent business */ 225 smc_curs_copy(&conn->urg_curs, &conn->local_rx_ctrl.prod, conn); 226 conn->urg_state = SMC_URG_VALID; 227 if (!sock_flag(&smc->sk, SOCK_URGINLINE)) 228 /* we'll skip the urgent byte, so don't account for it */ 229 (*diff_prod)--; 230 base = (char *)conn->rmb_desc->cpu_addr + conn->rx_off; 231 if (conn->urg_curs.count) 232 conn->urg_rx_byte = *(base + conn->urg_curs.count - 1); 233 else 234 conn->urg_rx_byte = *(base + conn->rmb_desc->len - 1); 235 sk_send_sigurg(&smc->sk); 236 } 237 238 static void smc_cdc_msg_recv_action(struct smc_sock *smc, 239 struct smc_cdc_msg *cdc) 240 { 241 union smc_host_cursor cons_old, prod_old; 242 struct smc_connection *conn = &smc->conn; 243 int diff_cons, diff_prod; 244 245 smc_curs_copy(&prod_old, &conn->local_rx_ctrl.prod, conn); 246 smc_curs_copy(&cons_old, &conn->local_rx_ctrl.cons, conn); 247 smc_cdc_msg_to_host(&conn->local_rx_ctrl, cdc, conn); 248 249 diff_cons = smc_curs_diff(conn->peer_rmbe_size, &cons_old, 250 &conn->local_rx_ctrl.cons); 251 if (diff_cons) { 252 /* peer_rmbe_space is decreased during data transfer with RDMA 253 * write 254 */ 255 smp_mb__before_atomic(); 256 atomic_add(diff_cons, &conn->peer_rmbe_space); 257 /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */ 258 smp_mb__after_atomic(); 259 } 260 261 diff_prod = smc_curs_diff(conn->rmb_desc->len, &prod_old, 262 &conn->local_rx_ctrl.prod); 263 if (diff_prod) { 264 if (conn->local_rx_ctrl.prod_flags.urg_data_present) 265 smc_cdc_handle_urg_data_arrival(smc, &diff_prod); 266 /* bytes_to_rcv is decreased in smc_recvmsg */ 267 smp_mb__before_atomic(); 268 atomic_add(diff_prod, &conn->bytes_to_rcv); 269 /* guarantee 0 <= bytes_to_rcv <= rmb_desc->len */ 270 smp_mb__after_atomic(); 271 smc->sk.sk_data_ready(&smc->sk); 272 } else { 273 if (conn->local_rx_ctrl.prod_flags.write_blocked || 274 conn->local_rx_ctrl.prod_flags.cons_curs_upd_req || 275 conn->local_rx_ctrl.prod_flags.urg_data_pending) { 276 if (conn->local_rx_ctrl.prod_flags.urg_data_pending) 277 conn->urg_state = SMC_URG_NOTYET; 278 /* force immediate tx of current consumer cursor, but 279 * under send_lock to guarantee arrival in seqno-order 280 */ 281 if (smc->sk.sk_state != SMC_INIT) 282 smc_tx_sndbuf_nonempty(conn); 283 } 284 } 285 286 /* piggy backed tx info */ 287 /* trigger sndbuf consumer: RDMA write into peer RMBE and CDC */ 288 if (diff_cons && smc_tx_prepared_sends(conn)) { 289 smc_tx_sndbuf_nonempty(conn); 290 /* trigger socket release if connection closed */ 291 smc_close_wake_tx_prepared(smc); 292 } 293 if (diff_cons && conn->urg_tx_pend && 294 atomic_read(&conn->peer_rmbe_space) == conn->peer_rmbe_size) { 295 /* urg data confirmed by peer, indicate we're ready for more */ 296 conn->urg_tx_pend = false; 297 smc->sk.sk_write_space(&smc->sk); 298 } 299 300 if (conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) { 301 smc->sk.sk_err = ECONNRESET; 302 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1; 303 } 304 if (smc_cdc_rxed_any_close_or_senddone(conn)) { 305 smc->sk.sk_shutdown |= RCV_SHUTDOWN; 306 if (smc->clcsock && smc->clcsock->sk) 307 smc->clcsock->sk->sk_shutdown |= RCV_SHUTDOWN; 308 sock_set_flag(&smc->sk, SOCK_DONE); 309 sock_hold(&smc->sk); /* sock_put in close_work */ 310 if (!schedule_work(&conn->close_work)) 311 sock_put(&smc->sk); 312 } 313 } 314 315 /* called under tasklet context */ 316 static void smc_cdc_msg_recv(struct smc_sock *smc, struct smc_cdc_msg *cdc) 317 { 318 sock_hold(&smc->sk); 319 bh_lock_sock(&smc->sk); 320 smc_cdc_msg_recv_action(smc, cdc); 321 bh_unlock_sock(&smc->sk); 322 sock_put(&smc->sk); /* no free sk in softirq-context */ 323 } 324 325 /* Schedule a tasklet for this connection. Triggered from the ISM device IRQ 326 * handler to indicate update in the DMBE. 327 * 328 * Context: 329 * - tasklet context 330 */ 331 static void smcd_cdc_rx_tsklet(unsigned long data) 332 { 333 struct smc_connection *conn = (struct smc_connection *)data; 334 struct smcd_cdc_msg *data_cdc; 335 struct smcd_cdc_msg cdc; 336 struct smc_sock *smc; 337 338 if (!conn) 339 return; 340 341 data_cdc = (struct smcd_cdc_msg *)conn->rmb_desc->cpu_addr; 342 smcd_curs_copy(&cdc.prod, &data_cdc->prod, conn); 343 smcd_curs_copy(&cdc.cons, &data_cdc->cons, conn); 344 smc = container_of(conn, struct smc_sock, conn); 345 smc_cdc_msg_recv(smc, (struct smc_cdc_msg *)&cdc); 346 } 347 348 /* Initialize receive tasklet. Called from ISM device IRQ handler to start 349 * receiver side. 350 */ 351 void smcd_cdc_rx_init(struct smc_connection *conn) 352 { 353 tasklet_init(&conn->rx_tsklet, smcd_cdc_rx_tsklet, (unsigned long)conn); 354 } 355 356 /***************************** init, exit, misc ******************************/ 357 358 static void smc_cdc_rx_handler(struct ib_wc *wc, void *buf) 359 { 360 struct smc_link *link = (struct smc_link *)wc->qp->qp_context; 361 struct smc_cdc_msg *cdc = buf; 362 struct smc_connection *conn; 363 struct smc_link_group *lgr; 364 struct smc_sock *smc; 365 366 if (wc->byte_len < offsetof(struct smc_cdc_msg, reserved)) 367 return; /* short message */ 368 if (cdc->len != SMC_WR_TX_SIZE) 369 return; /* invalid message */ 370 371 /* lookup connection */ 372 lgr = smc_get_lgr(link); 373 read_lock_bh(&lgr->conns_lock); 374 conn = smc_lgr_find_conn(ntohl(cdc->token), lgr); 375 read_unlock_bh(&lgr->conns_lock); 376 if (!conn) 377 return; 378 smc = container_of(conn, struct smc_sock, conn); 379 380 if (!cdc->prod_flags.failover_validation) { 381 if (smc_cdc_before(ntohs(cdc->seqno), 382 conn->local_rx_ctrl.seqno)) 383 /* received seqno is old */ 384 return; 385 } 386 smc_cdc_msg_recv(smc, cdc); 387 } 388 389 static struct smc_wr_rx_handler smc_cdc_rx_handlers[] = { 390 { 391 .handler = smc_cdc_rx_handler, 392 .type = SMC_CDC_MSG_TYPE 393 }, 394 { 395 .handler = NULL, 396 } 397 }; 398 399 int __init smc_cdc_init(void) 400 { 401 struct smc_wr_rx_handler *handler; 402 int rc = 0; 403 404 for (handler = smc_cdc_rx_handlers; handler->handler; handler++) { 405 INIT_HLIST_NODE(&handler->list); 406 rc = smc_wr_rx_register_handler(handler); 407 if (rc) 408 break; 409 } 410 return rc; 411 } 412