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 int smc_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 static bool smc_cdc_tx_filter(struct smc_wr_tx_pend_priv *tx_pend, 134 unsigned long data) 135 { 136 struct smc_connection *conn = (struct smc_connection *)data; 137 struct smc_cdc_tx_pend *cdc_pend = 138 (struct smc_cdc_tx_pend *)tx_pend; 139 140 return cdc_pend->conn == conn; 141 } 142 143 static void smc_cdc_tx_dismisser(struct smc_wr_tx_pend_priv *tx_pend) 144 { 145 struct smc_cdc_tx_pend *cdc_pend = 146 (struct smc_cdc_tx_pend *)tx_pend; 147 148 cdc_pend->conn = NULL; 149 } 150 151 void smc_cdc_tx_dismiss_slots(struct smc_connection *conn) 152 { 153 struct smc_link *link = &conn->lgr->lnk[SMC_SINGLE_LINK]; 154 155 smc_wr_tx_dismiss_slots(link, SMC_CDC_MSG_TYPE, 156 smc_cdc_tx_filter, smc_cdc_tx_dismisser, 157 (unsigned long)conn); 158 } 159 160 /********************************* receive ***********************************/ 161 162 static inline bool smc_cdc_before(u16 seq1, u16 seq2) 163 { 164 return (s16)(seq1 - seq2) < 0; 165 } 166 167 static void smc_cdc_handle_urg_data_arrival(struct smc_sock *smc, 168 int *diff_prod) 169 { 170 struct smc_connection *conn = &smc->conn; 171 char *base; 172 173 /* new data included urgent business */ 174 smc_curs_write(&conn->urg_curs, 175 smc_curs_read(&conn->local_rx_ctrl.prod, conn), 176 conn); 177 conn->urg_state = SMC_URG_VALID; 178 if (!sock_flag(&smc->sk, SOCK_URGINLINE)) 179 /* we'll skip the urgent byte, so don't account for it */ 180 (*diff_prod)--; 181 base = (char *)conn->rmb_desc->cpu_addr; 182 if (conn->urg_curs.count) 183 conn->urg_rx_byte = *(base + conn->urg_curs.count - 1); 184 else 185 conn->urg_rx_byte = *(base + conn->rmb_desc->len - 1); 186 sk_send_sigurg(&smc->sk); 187 } 188 189 static void smc_cdc_msg_recv_action(struct smc_sock *smc, 190 struct smc_cdc_msg *cdc) 191 { 192 union smc_host_cursor cons_old, prod_old; 193 struct smc_connection *conn = &smc->conn; 194 int diff_cons, diff_prod; 195 196 smc_curs_write(&prod_old, 197 smc_curs_read(&conn->local_rx_ctrl.prod, conn), 198 conn); 199 smc_curs_write(&cons_old, 200 smc_curs_read(&conn->local_rx_ctrl.cons, conn), 201 conn); 202 smc_cdc_msg_to_host(&conn->local_rx_ctrl, cdc, conn); 203 204 diff_cons = smc_curs_diff(conn->peer_rmbe_size, &cons_old, 205 &conn->local_rx_ctrl.cons); 206 if (diff_cons) { 207 /* peer_rmbe_space is decreased during data transfer with RDMA 208 * write 209 */ 210 smp_mb__before_atomic(); 211 atomic_add(diff_cons, &conn->peer_rmbe_space); 212 /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */ 213 smp_mb__after_atomic(); 214 } 215 216 diff_prod = smc_curs_diff(conn->rmb_desc->len, &prod_old, 217 &conn->local_rx_ctrl.prod); 218 if (diff_prod) { 219 if (conn->local_rx_ctrl.prod_flags.urg_data_present) 220 smc_cdc_handle_urg_data_arrival(smc, &diff_prod); 221 /* bytes_to_rcv is decreased in smc_recvmsg */ 222 smp_mb__before_atomic(); 223 atomic_add(diff_prod, &conn->bytes_to_rcv); 224 /* guarantee 0 <= bytes_to_rcv <= rmb_desc->len */ 225 smp_mb__after_atomic(); 226 smc->sk.sk_data_ready(&smc->sk); 227 } else { 228 if (conn->local_rx_ctrl.prod_flags.write_blocked || 229 conn->local_rx_ctrl.prod_flags.cons_curs_upd_req || 230 conn->local_rx_ctrl.prod_flags.urg_data_pending) { 231 if (conn->local_rx_ctrl.prod_flags.urg_data_pending) 232 conn->urg_state = SMC_URG_NOTYET; 233 /* force immediate tx of current consumer cursor, but 234 * under send_lock to guarantee arrival in seqno-order 235 */ 236 if (smc->sk.sk_state != SMC_INIT) 237 smc_tx_sndbuf_nonempty(conn); 238 } 239 } 240 241 /* piggy backed tx info */ 242 /* trigger sndbuf consumer: RDMA write into peer RMBE and CDC */ 243 if (diff_cons && smc_tx_prepared_sends(conn)) { 244 smc_tx_sndbuf_nonempty(conn); 245 /* trigger socket release if connection closed */ 246 smc_close_wake_tx_prepared(smc); 247 } 248 if (diff_cons && conn->urg_tx_pend && 249 atomic_read(&conn->peer_rmbe_space) == conn->peer_rmbe_size) { 250 /* urg data confirmed by peer, indicate we're ready for more */ 251 conn->urg_tx_pend = false; 252 smc->sk.sk_write_space(&smc->sk); 253 } 254 255 if (conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) { 256 smc->sk.sk_err = ECONNRESET; 257 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1; 258 } 259 if (smc_cdc_rxed_any_close_or_senddone(conn)) { 260 smc->sk.sk_shutdown |= RCV_SHUTDOWN; 261 if (smc->clcsock && smc->clcsock->sk) 262 smc->clcsock->sk->sk_shutdown |= RCV_SHUTDOWN; 263 sock_set_flag(&smc->sk, SOCK_DONE); 264 sock_hold(&smc->sk); /* sock_put in close_work */ 265 if (!schedule_work(&conn->close_work)) 266 sock_put(&smc->sk); 267 } 268 } 269 270 /* called under tasklet context */ 271 static void smc_cdc_msg_recv(struct smc_sock *smc, struct smc_cdc_msg *cdc) 272 { 273 sock_hold(&smc->sk); 274 bh_lock_sock(&smc->sk); 275 smc_cdc_msg_recv_action(smc, cdc); 276 bh_unlock_sock(&smc->sk); 277 sock_put(&smc->sk); /* no free sk in softirq-context */ 278 } 279 280 /***************************** init, exit, misc ******************************/ 281 282 static void smc_cdc_rx_handler(struct ib_wc *wc, void *buf) 283 { 284 struct smc_link *link = (struct smc_link *)wc->qp->qp_context; 285 struct smc_cdc_msg *cdc = buf; 286 struct smc_connection *conn; 287 struct smc_link_group *lgr; 288 struct smc_sock *smc; 289 290 if (wc->byte_len < offsetof(struct smc_cdc_msg, reserved)) 291 return; /* short message */ 292 if (cdc->len != SMC_WR_TX_SIZE) 293 return; /* invalid message */ 294 295 /* lookup connection */ 296 lgr = container_of(link, struct smc_link_group, lnk[SMC_SINGLE_LINK]); 297 read_lock_bh(&lgr->conns_lock); 298 conn = smc_lgr_find_conn(ntohl(cdc->token), lgr); 299 read_unlock_bh(&lgr->conns_lock); 300 if (!conn) 301 return; 302 smc = container_of(conn, struct smc_sock, conn); 303 304 if (!cdc->prod_flags.failover_validation) { 305 if (smc_cdc_before(ntohs(cdc->seqno), 306 conn->local_rx_ctrl.seqno)) 307 /* received seqno is old */ 308 return; 309 } 310 smc_cdc_msg_recv(smc, cdc); 311 } 312 313 static struct smc_wr_rx_handler smc_cdc_rx_handlers[] = { 314 { 315 .handler = smc_cdc_rx_handler, 316 .type = SMC_CDC_MSG_TYPE 317 }, 318 { 319 .handler = NULL, 320 } 321 }; 322 323 int __init smc_cdc_init(void) 324 { 325 struct smc_wr_rx_handler *handler; 326 int rc = 0; 327 328 for (handler = smc_cdc_rx_handlers; handler->handler; handler++) { 329 INIT_HLIST_NODE(&handler->list); 330 rc = smc_wr_rx_register_handler(handler); 331 if (rc) 332 break; 333 } 334 return rc; 335 } 336