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_sock *smc; 38 int diff; 39 40 if (!cdcpend->conn) 41 /* already dismissed */ 42 return; 43 44 smc = container_of(cdcpend->conn, struct smc_sock, conn); 45 bh_lock_sock(&smc->sk); 46 if (!wc_status) { 47 diff = smc_curs_diff(cdcpend->conn->sndbuf_desc->len, 48 &cdcpend->conn->tx_curs_fin, 49 &cdcpend->cursor); 50 /* sndbuf_space is decreased in smc_sendmsg */ 51 smp_mb__before_atomic(); 52 atomic_add(diff, &cdcpend->conn->sndbuf_space); 53 /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */ 54 smp_mb__after_atomic(); 55 smc_curs_write(&cdcpend->conn->tx_curs_fin, 56 smc_curs_read(&cdcpend->cursor, cdcpend->conn), 57 cdcpend->conn); 58 } 59 smc_tx_sndbuf_nonfull(smc); 60 bh_unlock_sock(&smc->sk); 61 } 62 63 int smc_cdc_get_free_slot(struct smc_connection *conn, 64 struct smc_wr_buf **wr_buf, 65 struct smc_cdc_tx_pend **pend) 66 { 67 struct smc_link *link = &conn->lgr->lnk[SMC_SINGLE_LINK]; 68 int rc; 69 70 rc = smc_wr_tx_get_free_slot(link, smc_cdc_tx_handler, wr_buf, 71 (struct smc_wr_tx_pend_priv **)pend); 72 if (!conn->alert_token_local) 73 /* abnormal termination */ 74 rc = -EPIPE; 75 return rc; 76 } 77 78 static inline void smc_cdc_add_pending_send(struct smc_connection *conn, 79 struct smc_cdc_tx_pend *pend) 80 { 81 BUILD_BUG_ON_MSG( 82 sizeof(struct smc_cdc_msg) > SMC_WR_BUF_SIZE, 83 "must increase SMC_WR_BUF_SIZE to at least sizeof(struct smc_cdc_msg)"); 84 BUILD_BUG_ON_MSG( 85 sizeof(struct smc_cdc_msg) != SMC_WR_TX_SIZE, 86 "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()"); 87 BUILD_BUG_ON_MSG( 88 sizeof(struct smc_cdc_tx_pend) > SMC_WR_TX_PEND_PRIV_SIZE, 89 "must increase SMC_WR_TX_PEND_PRIV_SIZE to at least sizeof(struct smc_cdc_tx_pend)"); 90 pend->conn = conn; 91 pend->cursor = conn->tx_curs_sent; 92 pend->p_cursor = conn->local_tx_ctrl.prod; 93 pend->ctrl_seq = conn->tx_cdc_seq; 94 } 95 96 int smc_cdc_msg_send(struct smc_connection *conn, 97 struct smc_wr_buf *wr_buf, 98 struct smc_cdc_tx_pend *pend) 99 { 100 struct smc_link *link; 101 int rc; 102 103 link = &conn->lgr->lnk[SMC_SINGLE_LINK]; 104 105 smc_cdc_add_pending_send(conn, pend); 106 107 conn->tx_cdc_seq++; 108 conn->local_tx_ctrl.seqno = conn->tx_cdc_seq; 109 smc_host_msg_to_cdc((struct smc_cdc_msg *)wr_buf, 110 &conn->local_tx_ctrl, conn); 111 rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend); 112 if (!rc) 113 smc_curs_write(&conn->rx_curs_confirmed, 114 smc_curs_read(&conn->local_tx_ctrl.cons, conn), 115 conn); 116 117 return rc; 118 } 119 120 static int smcr_cdc_get_slot_and_msg_send(struct smc_connection *conn) 121 { 122 struct smc_cdc_tx_pend *pend; 123 struct smc_wr_buf *wr_buf; 124 int rc; 125 126 rc = smc_cdc_get_free_slot(conn, &wr_buf, &pend); 127 if (rc) 128 return rc; 129 130 return smc_cdc_msg_send(conn, wr_buf, pend); 131 } 132 133 int smc_cdc_get_slot_and_msg_send(struct smc_connection *conn) 134 { 135 int rc; 136 137 if (conn->lgr->is_smcd) { 138 spin_lock_bh(&conn->send_lock); 139 rc = smcd_cdc_msg_send(conn); 140 spin_unlock_bh(&conn->send_lock); 141 } else { 142 rc = smcr_cdc_get_slot_and_msg_send(conn); 143 } 144 145 return rc; 146 } 147 148 static bool smc_cdc_tx_filter(struct smc_wr_tx_pend_priv *tx_pend, 149 unsigned long data) 150 { 151 struct smc_connection *conn = (struct smc_connection *)data; 152 struct smc_cdc_tx_pend *cdc_pend = 153 (struct smc_cdc_tx_pend *)tx_pend; 154 155 return cdc_pend->conn == conn; 156 } 157 158 static void smc_cdc_tx_dismisser(struct smc_wr_tx_pend_priv *tx_pend) 159 { 160 struct smc_cdc_tx_pend *cdc_pend = 161 (struct smc_cdc_tx_pend *)tx_pend; 162 163 cdc_pend->conn = NULL; 164 } 165 166 void smc_cdc_tx_dismiss_slots(struct smc_connection *conn) 167 { 168 struct smc_link *link = &conn->lgr->lnk[SMC_SINGLE_LINK]; 169 170 smc_wr_tx_dismiss_slots(link, SMC_CDC_MSG_TYPE, 171 smc_cdc_tx_filter, smc_cdc_tx_dismisser, 172 (unsigned long)conn); 173 } 174 175 /* Send a SMC-D CDC header. 176 * This increments the free space available in our send buffer. 177 * Also update the confirmed receive buffer with what was sent to the peer. 178 */ 179 int smcd_cdc_msg_send(struct smc_connection *conn) 180 { 181 struct smc_sock *smc = container_of(conn, struct smc_sock, conn); 182 struct smcd_cdc_msg cdc; 183 int rc, diff; 184 185 memset(&cdc, 0, sizeof(cdc)); 186 cdc.common.type = SMC_CDC_MSG_TYPE; 187 cdc.prod_wrap = conn->local_tx_ctrl.prod.wrap; 188 cdc.prod_count = conn->local_tx_ctrl.prod.count; 189 190 cdc.cons_wrap = conn->local_tx_ctrl.cons.wrap; 191 cdc.cons_count = conn->local_tx_ctrl.cons.count; 192 cdc.prod_flags = conn->local_tx_ctrl.prod_flags; 193 cdc.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_write(&conn->rx_curs_confirmed, 198 smc_curs_read(&conn->local_tx_ctrl.cons, conn), conn); 199 /* Calculate transmitted data and increment free send buffer space */ 200 diff = smc_curs_diff(conn->sndbuf_desc->len, &conn->tx_curs_fin, 201 &conn->tx_curs_sent); 202 /* increased by confirmed number of bytes */ 203 smp_mb__before_atomic(); 204 atomic_add(diff, &conn->sndbuf_space); 205 /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */ 206 smp_mb__after_atomic(); 207 smc_curs_write(&conn->tx_curs_fin, 208 smc_curs_read(&conn->tx_curs_sent, conn), conn); 209 210 smc_tx_sndbuf_nonfull(smc); 211 return rc; 212 } 213 214 /********************************* receive ***********************************/ 215 216 static inline bool smc_cdc_before(u16 seq1, u16 seq2) 217 { 218 return (s16)(seq1 - seq2) < 0; 219 } 220 221 static void smc_cdc_handle_urg_data_arrival(struct smc_sock *smc, 222 int *diff_prod) 223 { 224 struct smc_connection *conn = &smc->conn; 225 char *base; 226 227 /* new data included urgent business */ 228 smc_curs_write(&conn->urg_curs, 229 smc_curs_read(&conn->local_rx_ctrl.prod, conn), 230 conn); 231 conn->urg_state = SMC_URG_VALID; 232 if (!sock_flag(&smc->sk, SOCK_URGINLINE)) 233 /* we'll skip the urgent byte, so don't account for it */ 234 (*diff_prod)--; 235 base = (char *)conn->rmb_desc->cpu_addr + conn->rx_off; 236 if (conn->urg_curs.count) 237 conn->urg_rx_byte = *(base + conn->urg_curs.count - 1); 238 else 239 conn->urg_rx_byte = *(base + conn->rmb_desc->len - 1); 240 sk_send_sigurg(&smc->sk); 241 } 242 243 static void smc_cdc_msg_recv_action(struct smc_sock *smc, 244 struct smc_cdc_msg *cdc) 245 { 246 union smc_host_cursor cons_old, prod_old; 247 struct smc_connection *conn = &smc->conn; 248 int diff_cons, diff_prod; 249 250 smc_curs_write(&prod_old, 251 smc_curs_read(&conn->local_rx_ctrl.prod, conn), 252 conn); 253 smc_curs_write(&cons_old, 254 smc_curs_read(&conn->local_rx_ctrl.cons, conn), 255 conn); 256 smc_cdc_msg_to_host(&conn->local_rx_ctrl, cdc, conn); 257 258 diff_cons = smc_curs_diff(conn->peer_rmbe_size, &cons_old, 259 &conn->local_rx_ctrl.cons); 260 if (diff_cons) { 261 /* peer_rmbe_space is decreased during data transfer with RDMA 262 * write 263 */ 264 smp_mb__before_atomic(); 265 atomic_add(diff_cons, &conn->peer_rmbe_space); 266 /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */ 267 smp_mb__after_atomic(); 268 } 269 270 diff_prod = smc_curs_diff(conn->rmb_desc->len, &prod_old, 271 &conn->local_rx_ctrl.prod); 272 if (diff_prod) { 273 if (conn->local_rx_ctrl.prod_flags.urg_data_present) 274 smc_cdc_handle_urg_data_arrival(smc, &diff_prod); 275 /* bytes_to_rcv is decreased in smc_recvmsg */ 276 smp_mb__before_atomic(); 277 atomic_add(diff_prod, &conn->bytes_to_rcv); 278 /* guarantee 0 <= bytes_to_rcv <= rmb_desc->len */ 279 smp_mb__after_atomic(); 280 smc->sk.sk_data_ready(&smc->sk); 281 } else { 282 if (conn->local_rx_ctrl.prod_flags.write_blocked || 283 conn->local_rx_ctrl.prod_flags.cons_curs_upd_req || 284 conn->local_rx_ctrl.prod_flags.urg_data_pending) { 285 if (conn->local_rx_ctrl.prod_flags.urg_data_pending) 286 conn->urg_state = SMC_URG_NOTYET; 287 /* force immediate tx of current consumer cursor, but 288 * under send_lock to guarantee arrival in seqno-order 289 */ 290 smc_tx_sndbuf_nonempty(conn); 291 } 292 } 293 294 /* piggy backed tx info */ 295 /* trigger sndbuf consumer: RDMA write into peer RMBE and CDC */ 296 if (diff_cons && smc_tx_prepared_sends(conn)) { 297 smc_tx_sndbuf_nonempty(conn); 298 /* trigger socket release if connection closed */ 299 smc_close_wake_tx_prepared(smc); 300 } 301 if (diff_cons && conn->urg_tx_pend && 302 atomic_read(&conn->peer_rmbe_space) == conn->peer_rmbe_size) { 303 /* urg data confirmed by peer, indicate we're ready for more */ 304 conn->urg_tx_pend = false; 305 smc->sk.sk_write_space(&smc->sk); 306 } 307 308 if (conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) { 309 smc->sk.sk_err = ECONNRESET; 310 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1; 311 } 312 if (smc_cdc_rxed_any_close_or_senddone(conn)) { 313 smc->sk.sk_shutdown |= RCV_SHUTDOWN; 314 if (smc->clcsock && smc->clcsock->sk) 315 smc->clcsock->sk->sk_shutdown |= RCV_SHUTDOWN; 316 sock_set_flag(&smc->sk, SOCK_DONE); 317 sock_hold(&smc->sk); /* sock_put in close_work */ 318 if (!schedule_work(&conn->close_work)) 319 sock_put(&smc->sk); 320 } 321 } 322 323 /* called under tasklet context */ 324 static void smc_cdc_msg_recv(struct smc_sock *smc, struct smc_cdc_msg *cdc) 325 { 326 sock_hold(&smc->sk); 327 bh_lock_sock(&smc->sk); 328 smc_cdc_msg_recv_action(smc, cdc); 329 bh_unlock_sock(&smc->sk); 330 sock_put(&smc->sk); /* no free sk in softirq-context */ 331 } 332 333 /* Schedule a tasklet for this connection. Triggered from the ISM device IRQ 334 * handler to indicate update in the DMBE. 335 * 336 * Context: 337 * - tasklet context 338 */ 339 static void smcd_cdc_rx_tsklet(unsigned long data) 340 { 341 struct smc_connection *conn = (struct smc_connection *)data; 342 struct smcd_cdc_msg cdc; 343 struct smc_sock *smc; 344 345 if (!conn) 346 return; 347 348 memcpy(&cdc, conn->rmb_desc->cpu_addr, sizeof(cdc)); 349 smc = container_of(conn, struct smc_sock, conn); 350 smc_cdc_msg_recv(smc, (struct smc_cdc_msg *)&cdc); 351 } 352 353 /* Initialize receive tasklet. Called from ISM device IRQ handler to start 354 * receiver side. 355 */ 356 void smcd_cdc_rx_init(struct smc_connection *conn) 357 { 358 tasklet_init(&conn->rx_tsklet, smcd_cdc_rx_tsklet, (unsigned long)conn); 359 } 360 361 /***************************** init, exit, misc ******************************/ 362 363 static void smc_cdc_rx_handler(struct ib_wc *wc, void *buf) 364 { 365 struct smc_link *link = (struct smc_link *)wc->qp->qp_context; 366 struct smc_cdc_msg *cdc = buf; 367 struct smc_connection *conn; 368 struct smc_link_group *lgr; 369 struct smc_sock *smc; 370 371 if (wc->byte_len < offsetof(struct smc_cdc_msg, reserved)) 372 return; /* short message */ 373 if (cdc->len != SMC_WR_TX_SIZE) 374 return; /* invalid message */ 375 376 /* lookup connection */ 377 lgr = container_of(link, struct smc_link_group, lnk[SMC_SINGLE_LINK]); 378 read_lock_bh(&lgr->conns_lock); 379 conn = smc_lgr_find_conn(ntohl(cdc->token), lgr); 380 read_unlock_bh(&lgr->conns_lock); 381 if (!conn) 382 return; 383 smc = container_of(conn, struct smc_sock, conn); 384 385 if (!cdc->prod_flags.failover_validation) { 386 if (smc_cdc_before(ntohs(cdc->seqno), 387 conn->local_rx_ctrl.seqno)) 388 /* received seqno is old */ 389 return; 390 } 391 smc_cdc_msg_recv(smc, cdc); 392 } 393 394 static struct smc_wr_rx_handler smc_cdc_rx_handlers[] = { 395 { 396 .handler = smc_cdc_rx_handler, 397 .type = SMC_CDC_MSG_TYPE 398 }, 399 { 400 .handler = NULL, 401 } 402 }; 403 404 int __init smc_cdc_init(void) 405 { 406 struct smc_wr_rx_handler *handler; 407 int rc = 0; 408 409 for (handler = smc_cdc_rx_handlers; handler->handler; handler++) { 410 INIT_HLIST_NODE(&handler->list); 411 rc = smc_wr_rx_register_handler(handler); 412 if (rc) 413 break; 414 } 415 return rc; 416 } 417