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 struct smc_cdc_tx_pend { 25 struct smc_connection *conn; /* socket connection */ 26 union smc_host_cursor cursor; /* tx sndbuf cursor sent */ 27 union smc_host_cursor p_cursor; /* rx RMBE cursor produced */ 28 u16 ctrl_seq; /* conn. tx sequence # */ 29 }; 30 31 /* handler for send/transmission completion of a CDC msg */ 32 static void smc_cdc_tx_handler(struct smc_wr_tx_pend_priv *pnd_snd, 33 struct smc_link *link, 34 enum ib_wc_status wc_status) 35 { 36 struct smc_cdc_tx_pend *cdcpend = (struct smc_cdc_tx_pend *)pnd_snd; 37 struct smc_connection *conn = cdcpend->conn; 38 struct smc_sock *smc; 39 int diff; 40 41 if (!conn) 42 /* already dismissed */ 43 return; 44 45 smc = container_of(conn, struct smc_sock, conn); 46 bh_lock_sock(&smc->sk); 47 if (!wc_status) { 48 diff = smc_curs_diff(cdcpend->conn->sndbuf_desc->len, 49 &cdcpend->conn->tx_curs_fin, 50 &cdcpend->cursor); 51 /* sndbuf_space is decreased in smc_sendmsg */ 52 smp_mb__before_atomic(); 53 atomic_add(diff, &cdcpend->conn->sndbuf_space); 54 /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */ 55 smp_mb__after_atomic(); 56 smc_curs_copy(&conn->tx_curs_fin, &cdcpend->cursor, conn); 57 } 58 smc_tx_sndbuf_nonfull(smc); 59 bh_unlock_sock(&smc->sk); 60 } 61 62 int smc_cdc_get_free_slot(struct smc_connection *conn, 63 struct smc_wr_buf **wr_buf, 64 struct smc_cdc_tx_pend **pend) 65 { 66 struct smc_link *link = &conn->lgr->lnk[SMC_SINGLE_LINK]; 67 int rc; 68 69 rc = smc_wr_tx_get_free_slot(link, smc_cdc_tx_handler, wr_buf, 70 (struct smc_wr_tx_pend_priv **)pend); 71 if (!conn->alert_token_local) 72 /* abnormal termination */ 73 rc = -EPIPE; 74 return rc; 75 } 76 77 static inline void smc_cdc_add_pending_send(struct smc_connection *conn, 78 struct smc_cdc_tx_pend *pend) 79 { 80 BUILD_BUG_ON_MSG( 81 sizeof(struct smc_cdc_msg) > SMC_WR_BUF_SIZE, 82 "must increase SMC_WR_BUF_SIZE to at least sizeof(struct smc_cdc_msg)"); 83 BUILD_BUG_ON_MSG( 84 offsetofend(struct smc_cdc_msg, reserved) > SMC_WR_TX_SIZE, 85 "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()"); 86 BUILD_BUG_ON_MSG( 87 sizeof(struct smc_cdc_tx_pend) > SMC_WR_TX_PEND_PRIV_SIZE, 88 "must increase SMC_WR_TX_PEND_PRIV_SIZE to at least sizeof(struct smc_cdc_tx_pend)"); 89 pend->conn = conn; 90 pend->cursor = conn->tx_curs_sent; 91 pend->p_cursor = conn->local_tx_ctrl.prod; 92 pend->ctrl_seq = conn->tx_cdc_seq; 93 } 94 95 int smc_cdc_msg_send(struct smc_connection *conn, 96 struct smc_wr_buf *wr_buf, 97 struct smc_cdc_tx_pend *pend) 98 { 99 struct smc_link *link; 100 int rc; 101 102 link = &conn->lgr->lnk[SMC_SINGLE_LINK]; 103 104 smc_cdc_add_pending_send(conn, pend); 105 106 conn->tx_cdc_seq++; 107 conn->local_tx_ctrl.seqno = conn->tx_cdc_seq; 108 smc_host_msg_to_cdc((struct smc_cdc_msg *)wr_buf, 109 &conn->local_tx_ctrl, conn); 110 rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend); 111 if (!rc) 112 smc_curs_copy(&conn->rx_curs_confirmed, 113 &conn->local_tx_ctrl.cons, conn); 114 115 return rc; 116 } 117 118 static int smcr_cdc_get_slot_and_msg_send(struct smc_connection *conn) 119 { 120 struct smc_cdc_tx_pend *pend; 121 struct smc_wr_buf *wr_buf; 122 int rc; 123 124 rc = smc_cdc_get_free_slot(conn, &wr_buf, &pend); 125 if (rc) 126 return rc; 127 128 return smc_cdc_msg_send(conn, wr_buf, pend); 129 } 130 131 int smc_cdc_get_slot_and_msg_send(struct smc_connection *conn) 132 { 133 int rc; 134 135 if (conn->lgr->is_smcd) { 136 spin_lock_bh(&conn->send_lock); 137 rc = smcd_cdc_msg_send(conn); 138 spin_unlock_bh(&conn->send_lock); 139 } else { 140 rc = smcr_cdc_get_slot_and_msg_send(conn); 141 } 142 143 return rc; 144 } 145 146 static bool smc_cdc_tx_filter(struct smc_wr_tx_pend_priv *tx_pend, 147 unsigned long data) 148 { 149 struct smc_connection *conn = (struct smc_connection *)data; 150 struct smc_cdc_tx_pend *cdc_pend = 151 (struct smc_cdc_tx_pend *)tx_pend; 152 153 return cdc_pend->conn == conn; 154 } 155 156 static void smc_cdc_tx_dismisser(struct smc_wr_tx_pend_priv *tx_pend) 157 { 158 struct smc_cdc_tx_pend *cdc_pend = 159 (struct smc_cdc_tx_pend *)tx_pend; 160 161 cdc_pend->conn = NULL; 162 } 163 164 void smc_cdc_tx_dismiss_slots(struct smc_connection *conn) 165 { 166 struct smc_link *link = &conn->lgr->lnk[SMC_SINGLE_LINK]; 167 168 smc_wr_tx_dismiss_slots(link, SMC_CDC_MSG_TYPE, 169 smc_cdc_tx_filter, smc_cdc_tx_dismisser, 170 (unsigned long)conn); 171 } 172 173 /* Send a SMC-D CDC header. 174 * This increments the free space available in our send buffer. 175 * Also update the confirmed receive buffer with what was sent to the peer. 176 */ 177 int smcd_cdc_msg_send(struct smc_connection *conn) 178 { 179 struct smc_sock *smc = container_of(conn, struct smc_sock, conn); 180 union smc_host_cursor curs; 181 struct smcd_cdc_msg cdc; 182 int rc, diff; 183 184 memset(&cdc, 0, sizeof(cdc)); 185 cdc.common.type = SMC_CDC_MSG_TYPE; 186 curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.prod.acurs); 187 cdc.prod.wrap = curs.wrap; 188 cdc.prod.count = curs.count; 189 curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.cons.acurs); 190 cdc.cons.wrap = curs.wrap; 191 cdc.cons.count = curs.count; 192 cdc.cons.prod_flags = conn->local_tx_ctrl.prod_flags; 193 cdc.cons.conn_state_flags = conn->local_tx_ctrl.conn_state_flags; 194 rc = smcd_tx_ism_write(conn, &cdc, sizeof(cdc), 0, 1); 195 if (rc) 196 return rc; 197 smc_curs_copy(&conn->rx_curs_confirmed, &curs, conn); 198 /* Calculate transmitted data and increment free send buffer space */ 199 diff = smc_curs_diff(conn->sndbuf_desc->len, &conn->tx_curs_fin, 200 &conn->tx_curs_sent); 201 /* increased by confirmed number of bytes */ 202 smp_mb__before_atomic(); 203 atomic_add(diff, &conn->sndbuf_space); 204 /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */ 205 smp_mb__after_atomic(); 206 smc_curs_copy(&conn->tx_curs_fin, &conn->tx_curs_sent, conn); 207 208 smc_tx_sndbuf_nonfull(smc); 209 return rc; 210 } 211 212 /********************************* receive ***********************************/ 213 214 static inline bool smc_cdc_before(u16 seq1, u16 seq2) 215 { 216 return (s16)(seq1 - seq2) < 0; 217 } 218 219 static void smc_cdc_handle_urg_data_arrival(struct smc_sock *smc, 220 int *diff_prod) 221 { 222 struct smc_connection *conn = &smc->conn; 223 char *base; 224 225 /* new data included urgent business */ 226 smc_curs_copy(&conn->urg_curs, &conn->local_rx_ctrl.prod, conn); 227 conn->urg_state = SMC_URG_VALID; 228 if (!sock_flag(&smc->sk, SOCK_URGINLINE)) 229 /* we'll skip the urgent byte, so don't account for it */ 230 (*diff_prod)--; 231 base = (char *)conn->rmb_desc->cpu_addr + conn->rx_off; 232 if (conn->urg_curs.count) 233 conn->urg_rx_byte = *(base + conn->urg_curs.count - 1); 234 else 235 conn->urg_rx_byte = *(base + conn->rmb_desc->len - 1); 236 sk_send_sigurg(&smc->sk); 237 } 238 239 static void smc_cdc_msg_recv_action(struct smc_sock *smc, 240 struct smc_cdc_msg *cdc) 241 { 242 union smc_host_cursor cons_old, prod_old; 243 struct smc_connection *conn = &smc->conn; 244 int diff_cons, diff_prod; 245 246 smc_curs_copy(&prod_old, &conn->local_rx_ctrl.prod, conn); 247 smc_curs_copy(&cons_old, &conn->local_rx_ctrl.cons, conn); 248 smc_cdc_msg_to_host(&conn->local_rx_ctrl, cdc, conn); 249 250 diff_cons = smc_curs_diff(conn->peer_rmbe_size, &cons_old, 251 &conn->local_rx_ctrl.cons); 252 if (diff_cons) { 253 /* peer_rmbe_space is decreased during data transfer with RDMA 254 * write 255 */ 256 smp_mb__before_atomic(); 257 atomic_add(diff_cons, &conn->peer_rmbe_space); 258 /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */ 259 smp_mb__after_atomic(); 260 } 261 262 diff_prod = smc_curs_diff(conn->rmb_desc->len, &prod_old, 263 &conn->local_rx_ctrl.prod); 264 if (diff_prod) { 265 if (conn->local_rx_ctrl.prod_flags.urg_data_present) 266 smc_cdc_handle_urg_data_arrival(smc, &diff_prod); 267 /* bytes_to_rcv is decreased in smc_recvmsg */ 268 smp_mb__before_atomic(); 269 atomic_add(diff_prod, &conn->bytes_to_rcv); 270 /* guarantee 0 <= bytes_to_rcv <= rmb_desc->len */ 271 smp_mb__after_atomic(); 272 smc->sk.sk_data_ready(&smc->sk); 273 } else { 274 if (conn->local_rx_ctrl.prod_flags.write_blocked || 275 conn->local_rx_ctrl.prod_flags.cons_curs_upd_req || 276 conn->local_rx_ctrl.prod_flags.urg_data_pending) { 277 if (conn->local_rx_ctrl.prod_flags.urg_data_pending) 278 conn->urg_state = SMC_URG_NOTYET; 279 /* force immediate tx of current consumer cursor, but 280 * under send_lock to guarantee arrival in seqno-order 281 */ 282 if (smc->sk.sk_state != SMC_INIT) 283 smc_tx_sndbuf_nonempty(conn); 284 } 285 } 286 287 /* piggy backed tx info */ 288 /* trigger sndbuf consumer: RDMA write into peer RMBE and CDC */ 289 if (diff_cons && smc_tx_prepared_sends(conn)) { 290 smc_tx_sndbuf_nonempty(conn); 291 /* trigger socket release if connection closed */ 292 smc_close_wake_tx_prepared(smc); 293 } 294 if (diff_cons && conn->urg_tx_pend && 295 atomic_read(&conn->peer_rmbe_space) == conn->peer_rmbe_size) { 296 /* urg data confirmed by peer, indicate we're ready for more */ 297 conn->urg_tx_pend = false; 298 smc->sk.sk_write_space(&smc->sk); 299 } 300 301 if (conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) { 302 smc->sk.sk_err = ECONNRESET; 303 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1; 304 } 305 if (smc_cdc_rxed_any_close_or_senddone(conn)) { 306 smc->sk.sk_shutdown |= RCV_SHUTDOWN; 307 if (smc->clcsock && smc->clcsock->sk) 308 smc->clcsock->sk->sk_shutdown |= RCV_SHUTDOWN; 309 sock_set_flag(&smc->sk, SOCK_DONE); 310 sock_hold(&smc->sk); /* sock_put in close_work */ 311 if (!schedule_work(&conn->close_work)) 312 sock_put(&smc->sk); 313 } 314 } 315 316 /* called under tasklet context */ 317 static void smc_cdc_msg_recv(struct smc_sock *smc, struct smc_cdc_msg *cdc) 318 { 319 sock_hold(&smc->sk); 320 bh_lock_sock(&smc->sk); 321 smc_cdc_msg_recv_action(smc, cdc); 322 bh_unlock_sock(&smc->sk); 323 sock_put(&smc->sk); /* no free sk in softirq-context */ 324 } 325 326 /* Schedule a tasklet for this connection. Triggered from the ISM device IRQ 327 * handler to indicate update in the DMBE. 328 * 329 * Context: 330 * - tasklet context 331 */ 332 static void smcd_cdc_rx_tsklet(unsigned long data) 333 { 334 struct smc_connection *conn = (struct smc_connection *)data; 335 struct smcd_cdc_msg *data_cdc; 336 struct smcd_cdc_msg cdc; 337 struct smc_sock *smc; 338 339 if (!conn) 340 return; 341 342 data_cdc = (struct smcd_cdc_msg *)conn->rmb_desc->cpu_addr; 343 smcd_curs_copy(&cdc.prod, &data_cdc->prod, conn); 344 smcd_curs_copy(&cdc.cons, &data_cdc->cons, conn); 345 smc = container_of(conn, struct smc_sock, conn); 346 smc_cdc_msg_recv(smc, (struct smc_cdc_msg *)&cdc); 347 } 348 349 /* Initialize receive tasklet. Called from ISM device IRQ handler to start 350 * receiver side. 351 */ 352 void smcd_cdc_rx_init(struct smc_connection *conn) 353 { 354 tasklet_init(&conn->rx_tsklet, smcd_cdc_rx_tsklet, (unsigned long)conn); 355 } 356 357 /***************************** init, exit, misc ******************************/ 358 359 static void smc_cdc_rx_handler(struct ib_wc *wc, void *buf) 360 { 361 struct smc_link *link = (struct smc_link *)wc->qp->qp_context; 362 struct smc_cdc_msg *cdc = buf; 363 struct smc_connection *conn; 364 struct smc_link_group *lgr; 365 struct smc_sock *smc; 366 367 if (wc->byte_len < offsetof(struct smc_cdc_msg, reserved)) 368 return; /* short message */ 369 if (cdc->len != SMC_WR_TX_SIZE) 370 return; /* invalid message */ 371 372 /* lookup connection */ 373 lgr = smc_get_lgr(link); 374 read_lock_bh(&lgr->conns_lock); 375 conn = smc_lgr_find_conn(ntohl(cdc->token), lgr); 376 read_unlock_bh(&lgr->conns_lock); 377 if (!conn) 378 return; 379 smc = container_of(conn, struct smc_sock, conn); 380 381 if (!cdc->prod_flags.failover_validation) { 382 if (smc_cdc_before(ntohs(cdc->seqno), 383 conn->local_rx_ctrl.seqno)) 384 /* received seqno is old */ 385 return; 386 } 387 smc_cdc_msg_recv(smc, cdc); 388 } 389 390 static struct smc_wr_rx_handler smc_cdc_rx_handlers[] = { 391 { 392 .handler = smc_cdc_rx_handler, 393 .type = SMC_CDC_MSG_TYPE 394 }, 395 { 396 .handler = NULL, 397 } 398 }; 399 400 int __init smc_cdc_init(void) 401 { 402 struct smc_wr_rx_handler *handler; 403 int rc = 0; 404 405 for (handler = smc_cdc_rx_handlers; handler->handler; handler++) { 406 INIT_HLIST_NODE(&handler->list); 407 rc = smc_wr_rx_register_handler(handler); 408 if (rc) 409 break; 410 } 411 return rc; 412 } 413