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_size, 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_size */ 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 if (smc->sk.sk_state != SMC_ACTIVE) 61 /* wake up smc_close_wait_tx_pends() */ 62 smc->sk.sk_state_change(&smc->sk); 63 bh_unlock_sock(&smc->sk); 64 } 65 66 int smc_cdc_get_free_slot(struct smc_connection *conn, 67 struct smc_wr_buf **wr_buf, 68 struct smc_cdc_tx_pend **pend) 69 { 70 struct smc_link *link = &conn->lgr->lnk[SMC_SINGLE_LINK]; 71 72 return smc_wr_tx_get_free_slot(link, smc_cdc_tx_handler, wr_buf, 73 (struct smc_wr_tx_pend_priv **)pend); 74 } 75 76 static inline void smc_cdc_add_pending_send(struct smc_connection *conn, 77 struct smc_cdc_tx_pend *pend) 78 { 79 BUILD_BUG_ON_MSG( 80 sizeof(struct smc_cdc_msg) > SMC_WR_BUF_SIZE, 81 "must increase SMC_WR_BUF_SIZE to at least sizeof(struct smc_cdc_msg)"); 82 BUILD_BUG_ON_MSG( 83 offsetof(struct smc_cdc_msg, reserved) > SMC_WR_TX_SIZE, 84 "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()"); 85 BUILD_BUG_ON_MSG( 86 sizeof(struct smc_cdc_tx_pend) > SMC_WR_TX_PEND_PRIV_SIZE, 87 "must increase SMC_WR_TX_PEND_PRIV_SIZE to at least sizeof(struct smc_cdc_tx_pend)"); 88 pend->conn = conn; 89 pend->cursor = conn->tx_curs_sent; 90 pend->p_cursor = conn->local_tx_ctrl.prod; 91 pend->ctrl_seq = conn->tx_cdc_seq; 92 } 93 94 int smc_cdc_msg_send(struct smc_connection *conn, 95 struct smc_wr_buf *wr_buf, 96 struct smc_cdc_tx_pend *pend) 97 { 98 struct smc_link *link; 99 int rc; 100 101 link = &conn->lgr->lnk[SMC_SINGLE_LINK]; 102 103 smc_cdc_add_pending_send(conn, pend); 104 105 conn->tx_cdc_seq++; 106 conn->local_tx_ctrl.seqno = conn->tx_cdc_seq; 107 smc_host_msg_to_cdc((struct smc_cdc_msg *)wr_buf, 108 &conn->local_tx_ctrl, conn); 109 rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend); 110 if (!rc) 111 smc_curs_write(&conn->rx_curs_confirmed, 112 smc_curs_read(&conn->local_tx_ctrl.cons, conn), 113 conn); 114 115 return rc; 116 } 117 118 int smc_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 static bool smc_cdc_tx_filter(struct smc_wr_tx_pend_priv *tx_pend, 132 unsigned long data) 133 { 134 struct smc_connection *conn = (struct smc_connection *)data; 135 struct smc_cdc_tx_pend *cdc_pend = 136 (struct smc_cdc_tx_pend *)tx_pend; 137 138 return cdc_pend->conn == conn; 139 } 140 141 static void smc_cdc_tx_dismisser(struct smc_wr_tx_pend_priv *tx_pend) 142 { 143 struct smc_cdc_tx_pend *cdc_pend = 144 (struct smc_cdc_tx_pend *)tx_pend; 145 146 cdc_pend->conn = NULL; 147 } 148 149 void smc_cdc_tx_dismiss_slots(struct smc_connection *conn) 150 { 151 struct smc_link *link = &conn->lgr->lnk[SMC_SINGLE_LINK]; 152 153 smc_wr_tx_dismiss_slots(link, SMC_CDC_MSG_TYPE, 154 smc_cdc_tx_filter, smc_cdc_tx_dismisser, 155 (unsigned long)conn); 156 } 157 158 bool smc_cdc_tx_has_pending(struct smc_connection *conn) 159 { 160 struct smc_link *link = &conn->lgr->lnk[SMC_SINGLE_LINK]; 161 162 return smc_wr_tx_has_pending(link, SMC_CDC_MSG_TYPE, 163 smc_cdc_tx_filter, (unsigned long)conn); 164 } 165 166 /********************************* receive ***********************************/ 167 168 static inline bool smc_cdc_before(u16 seq1, u16 seq2) 169 { 170 return (s16)(seq1 - seq2) < 0; 171 } 172 173 static void smc_cdc_msg_recv_action(struct smc_sock *smc, 174 struct smc_link *link, 175 struct smc_cdc_msg *cdc) 176 { 177 union smc_host_cursor cons_old, prod_old; 178 struct smc_connection *conn = &smc->conn; 179 int diff_cons, diff_prod; 180 181 if (!cdc->prod_flags.failover_validation) { 182 if (smc_cdc_before(ntohs(cdc->seqno), 183 conn->local_rx_ctrl.seqno)) 184 /* received seqno is old */ 185 return; 186 } 187 smc_curs_write(&prod_old, 188 smc_curs_read(&conn->local_rx_ctrl.prod, conn), 189 conn); 190 smc_curs_write(&cons_old, 191 smc_curs_read(&conn->local_rx_ctrl.cons, conn), 192 conn); 193 smc_cdc_msg_to_host(&conn->local_rx_ctrl, cdc, conn); 194 195 diff_cons = smc_curs_diff(conn->peer_rmbe_size, &cons_old, 196 &conn->local_rx_ctrl.cons); 197 if (diff_cons) { 198 /* peer_rmbe_space is decreased during data transfer with RDMA 199 * write 200 */ 201 smp_mb__before_atomic(); 202 atomic_add(diff_cons, &conn->peer_rmbe_space); 203 /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */ 204 smp_mb__after_atomic(); 205 } 206 207 diff_prod = smc_curs_diff(conn->rmbe_size, &prod_old, 208 &conn->local_rx_ctrl.prod); 209 if (diff_prod) { 210 /* bytes_to_rcv is decreased in smc_recvmsg */ 211 smp_mb__before_atomic(); 212 atomic_add(diff_prod, &conn->bytes_to_rcv); 213 /* guarantee 0 <= bytes_to_rcv <= rmbe_size */ 214 smp_mb__after_atomic(); 215 smc->sk.sk_data_ready(&smc->sk); 216 } 217 218 if (conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) { 219 smc->sk.sk_err = ECONNRESET; 220 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1; 221 } 222 if (smc_cdc_rxed_any_close_or_senddone(conn)) { 223 smc->sk.sk_shutdown |= RCV_SHUTDOWN; 224 if (smc->clcsock && smc->clcsock->sk) 225 smc->clcsock->sk->sk_shutdown |= RCV_SHUTDOWN; 226 sock_set_flag(&smc->sk, SOCK_DONE); 227 schedule_work(&conn->close_work); 228 } 229 230 /* piggy backed tx info */ 231 /* trigger sndbuf consumer: RDMA write into peer RMBE and CDC */ 232 if (diff_cons && smc_tx_prepared_sends(conn)) { 233 smc_tx_sndbuf_nonempty(conn); 234 /* trigger socket release if connection closed */ 235 smc_close_wake_tx_prepared(smc); 236 } 237 238 /* socket connected but not accepted */ 239 if (!smc->sk.sk_socket) 240 return; 241 242 /* data available */ 243 if ((conn->local_rx_ctrl.prod_flags.write_blocked) || 244 (conn->local_rx_ctrl.prod_flags.cons_curs_upd_req)) 245 smc_tx_consumer_update(conn); 246 } 247 248 /* called under tasklet context */ 249 static inline void smc_cdc_msg_recv(struct smc_cdc_msg *cdc, 250 struct smc_link *link, u64 wr_id) 251 { 252 struct smc_link_group *lgr = container_of(link, struct smc_link_group, 253 lnk[SMC_SINGLE_LINK]); 254 struct smc_connection *connection; 255 struct smc_sock *smc; 256 257 /* lookup connection */ 258 read_lock_bh(&lgr->conns_lock); 259 connection = smc_lgr_find_conn(ntohl(cdc->token), lgr); 260 if (!connection) { 261 read_unlock_bh(&lgr->conns_lock); 262 return; 263 } 264 smc = container_of(connection, struct smc_sock, conn); 265 sock_hold(&smc->sk); 266 read_unlock_bh(&lgr->conns_lock); 267 bh_lock_sock(&smc->sk); 268 smc_cdc_msg_recv_action(smc, link, cdc); 269 bh_unlock_sock(&smc->sk); 270 sock_put(&smc->sk); /* no free sk in softirq-context */ 271 } 272 273 /***************************** init, exit, misc ******************************/ 274 275 static void smc_cdc_rx_handler(struct ib_wc *wc, void *buf) 276 { 277 struct smc_link *link = (struct smc_link *)wc->qp->qp_context; 278 struct smc_cdc_msg *cdc = buf; 279 280 if (wc->byte_len < offsetof(struct smc_cdc_msg, reserved)) 281 return; /* short message */ 282 if (cdc->len != sizeof(*cdc)) 283 return; /* invalid message */ 284 smc_cdc_msg_recv(cdc, link, wc->wr_id); 285 } 286 287 static struct smc_wr_rx_handler smc_cdc_rx_handlers[] = { 288 { 289 .handler = smc_cdc_rx_handler, 290 .type = SMC_CDC_MSG_TYPE 291 }, 292 { 293 .handler = NULL, 294 } 295 }; 296 297 int __init smc_cdc_init(void) 298 { 299 struct smc_wr_rx_handler *handler; 300 int rc = 0; 301 302 for (handler = smc_cdc_rx_handlers; handler->handler; handler++) { 303 INIT_HLIST_NODE(&handler->list); 304 rc = smc_wr_rx_register_handler(handler); 305 if (rc) 306 break; 307 } 308 return rc; 309 } 310