1 /* 2 * Shared Memory Communications over RDMA (SMC-R) and RoCE 3 * 4 * Connection Data Control (CDC) 5 * handles flow control 6 * 7 * Copyright IBM Corp. 2016 8 * 9 * Author(s): Ursula Braun <ubraun@linux.vnet.ibm.com> 10 */ 11 12 #include <linux/spinlock.h> 13 14 #include "smc.h" 15 #include "smc_wr.h" 16 #include "smc_cdc.h" 17 #include "smc_tx.h" 18 #include "smc_rx.h" 19 #include "smc_close.h" 20 21 /********************************** send *************************************/ 22 23 struct smc_cdc_tx_pend { 24 struct smc_connection *conn; /* socket connection */ 25 union smc_host_cursor cursor; /* tx sndbuf cursor sent */ 26 union smc_host_cursor p_cursor; /* rx RMBE cursor produced */ 27 u16 ctrl_seq; /* conn. tx sequence # */ 28 }; 29 30 /* handler for send/transmission completion of a CDC msg */ 31 static void smc_cdc_tx_handler(struct smc_wr_tx_pend_priv *pnd_snd, 32 struct smc_link *link, 33 enum ib_wc_status wc_status) 34 { 35 struct smc_cdc_tx_pend *cdcpend = (struct smc_cdc_tx_pend *)pnd_snd; 36 struct smc_sock *smc; 37 int diff; 38 39 if (!cdcpend->conn) 40 /* already dismissed */ 41 return; 42 43 smc = container_of(cdcpend->conn, struct smc_sock, conn); 44 bh_lock_sock(&smc->sk); 45 if (!wc_status) { 46 diff = smc_curs_diff(cdcpend->conn->sndbuf_size, 47 &cdcpend->conn->tx_curs_fin, 48 &cdcpend->cursor); 49 /* sndbuf_space is decreased in smc_sendmsg */ 50 smp_mb__before_atomic(); 51 atomic_add(diff, &cdcpend->conn->sndbuf_space); 52 /* guarantee 0 <= sndbuf_space <= sndbuf_size */ 53 smp_mb__after_atomic(); 54 smc_curs_write(&cdcpend->conn->tx_curs_fin, 55 smc_curs_read(&cdcpend->cursor, cdcpend->conn), 56 cdcpend->conn); 57 } 58 smc_tx_sndbuf_nonfull(smc); 59 if (smc->sk.sk_state != SMC_ACTIVE) 60 /* wake up smc_close_wait_tx_pends() */ 61 smc->sk.sk_state_change(&smc->sk); 62 bh_unlock_sock(&smc->sk); 63 } 64 65 int smc_cdc_get_free_slot(struct smc_connection *conn, 66 struct smc_wr_buf **wr_buf, 67 struct smc_cdc_tx_pend **pend) 68 { 69 struct smc_link *link = &conn->lgr->lnk[SMC_SINGLE_LINK]; 70 71 return smc_wr_tx_get_free_slot(link, smc_cdc_tx_handler, wr_buf, 72 (struct smc_wr_tx_pend_priv **)pend); 73 } 74 75 static inline void smc_cdc_add_pending_send(struct smc_connection *conn, 76 struct smc_cdc_tx_pend *pend) 77 { 78 BUILD_BUG_ON_MSG( 79 sizeof(struct smc_cdc_msg) > SMC_WR_BUF_SIZE, 80 "must increase SMC_WR_BUF_SIZE to at least sizeof(struct smc_cdc_msg)"); 81 BUILD_BUG_ON_MSG( 82 offsetof(struct smc_cdc_msg, reserved) > SMC_WR_TX_SIZE, 83 "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()"); 84 BUILD_BUG_ON_MSG( 85 sizeof(struct smc_cdc_tx_pend) > SMC_WR_TX_PEND_PRIV_SIZE, 86 "must increase SMC_WR_TX_PEND_PRIV_SIZE to at least sizeof(struct smc_cdc_tx_pend)"); 87 pend->conn = conn; 88 pend->cursor = conn->tx_curs_sent; 89 pend->p_cursor = conn->local_tx_ctrl.prod; 90 pend->ctrl_seq = conn->tx_cdc_seq; 91 } 92 93 int smc_cdc_msg_send(struct smc_connection *conn, 94 struct smc_wr_buf *wr_buf, 95 struct smc_cdc_tx_pend *pend) 96 { 97 struct smc_link *link; 98 int rc; 99 100 link = &conn->lgr->lnk[SMC_SINGLE_LINK]; 101 102 smc_cdc_add_pending_send(conn, pend); 103 104 conn->tx_cdc_seq++; 105 conn->local_tx_ctrl.seqno = conn->tx_cdc_seq; 106 smc_host_msg_to_cdc((struct smc_cdc_msg *)wr_buf, 107 &conn->local_tx_ctrl, conn); 108 rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend); 109 if (!rc) 110 smc_curs_write(&conn->rx_curs_confirmed, 111 smc_curs_read(&conn->local_tx_ctrl.cons, conn), 112 conn); 113 114 return rc; 115 } 116 117 int smc_cdc_get_slot_and_msg_send(struct smc_connection *conn) 118 { 119 struct smc_cdc_tx_pend *pend; 120 struct smc_wr_buf *wr_buf; 121 int rc; 122 123 rc = smc_cdc_get_free_slot(conn, &wr_buf, &pend); 124 if (rc) 125 return rc; 126 127 return smc_cdc_msg_send(conn, wr_buf, pend); 128 } 129 130 static bool smc_cdc_tx_filter(struct smc_wr_tx_pend_priv *tx_pend, 131 unsigned long data) 132 { 133 struct smc_connection *conn = (struct smc_connection *)data; 134 struct smc_cdc_tx_pend *cdc_pend = 135 (struct smc_cdc_tx_pend *)tx_pend; 136 137 return cdc_pend->conn == conn; 138 } 139 140 static void smc_cdc_tx_dismisser(struct smc_wr_tx_pend_priv *tx_pend) 141 { 142 struct smc_cdc_tx_pend *cdc_pend = 143 (struct smc_cdc_tx_pend *)tx_pend; 144 145 cdc_pend->conn = NULL; 146 } 147 148 void smc_cdc_tx_dismiss_slots(struct smc_connection *conn) 149 { 150 struct smc_link *link = &conn->lgr->lnk[SMC_SINGLE_LINK]; 151 152 smc_wr_tx_dismiss_slots(link, SMC_CDC_MSG_TYPE, 153 smc_cdc_tx_filter, smc_cdc_tx_dismisser, 154 (unsigned long)conn); 155 } 156 157 bool smc_cdc_tx_has_pending(struct smc_connection *conn) 158 { 159 struct smc_link *link = &conn->lgr->lnk[SMC_SINGLE_LINK]; 160 161 return smc_wr_tx_has_pending(link, SMC_CDC_MSG_TYPE, 162 smc_cdc_tx_filter, (unsigned long)conn); 163 } 164 165 /********************************* receive ***********************************/ 166 167 static inline bool smc_cdc_before(u16 seq1, u16 seq2) 168 { 169 return (s16)(seq1 - seq2) < 0; 170 } 171 172 static void smc_cdc_msg_recv_action(struct smc_sock *smc, 173 struct smc_link *link, 174 struct smc_cdc_msg *cdc) 175 { 176 union smc_host_cursor cons_old, prod_old; 177 struct smc_connection *conn = &smc->conn; 178 int diff_cons, diff_prod; 179 180 if (!cdc->prod_flags.failover_validation) { 181 if (smc_cdc_before(ntohs(cdc->seqno), 182 conn->local_rx_ctrl.seqno)) 183 /* received seqno is old */ 184 return; 185 } 186 smc_curs_write(&prod_old, 187 smc_curs_read(&conn->local_rx_ctrl.prod, conn), 188 conn); 189 smc_curs_write(&cons_old, 190 smc_curs_read(&conn->local_rx_ctrl.cons, conn), 191 conn); 192 smc_cdc_msg_to_host(&conn->local_rx_ctrl, cdc, conn); 193 194 diff_cons = smc_curs_diff(conn->peer_rmbe_size, &cons_old, 195 &conn->local_rx_ctrl.cons); 196 if (diff_cons) { 197 /* peer_rmbe_space is decreased during data transfer with RDMA 198 * write 199 */ 200 smp_mb__before_atomic(); 201 atomic_add(diff_cons, &conn->peer_rmbe_space); 202 /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */ 203 smp_mb__after_atomic(); 204 } 205 206 diff_prod = smc_curs_diff(conn->rmbe_size, &prod_old, 207 &conn->local_rx_ctrl.prod); 208 if (diff_prod) { 209 /* bytes_to_rcv is decreased in smc_recvmsg */ 210 smp_mb__before_atomic(); 211 atomic_add(diff_prod, &conn->bytes_to_rcv); 212 /* guarantee 0 <= bytes_to_rcv <= rmbe_size */ 213 smp_mb__after_atomic(); 214 smc->sk.sk_data_ready(&smc->sk); 215 } 216 217 if (conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) { 218 smc->sk.sk_err = ECONNRESET; 219 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1; 220 } 221 if (smc_cdc_rxed_any_close_or_senddone(conn)) { 222 smc->sk.sk_shutdown |= RCV_SHUTDOWN; 223 if (smc->clcsock && smc->clcsock->sk) 224 smc->clcsock->sk->sk_shutdown |= RCV_SHUTDOWN; 225 sock_set_flag(&smc->sk, SOCK_DONE); 226 schedule_work(&conn->close_work); 227 } 228 229 /* piggy backed tx info */ 230 /* trigger sndbuf consumer: RDMA write into peer RMBE and CDC */ 231 if (diff_cons && smc_tx_prepared_sends(conn)) { 232 smc_tx_sndbuf_nonempty(conn); 233 /* trigger socket release if connection closed */ 234 smc_close_wake_tx_prepared(smc); 235 } 236 237 /* socket connected but not accepted */ 238 if (!smc->sk.sk_socket) 239 return; 240 241 /* data available */ 242 if ((conn->local_rx_ctrl.prod_flags.write_blocked) || 243 (conn->local_rx_ctrl.prod_flags.cons_curs_upd_req)) 244 smc_tx_consumer_update(conn); 245 } 246 247 /* called under tasklet context */ 248 static inline void smc_cdc_msg_recv(struct smc_cdc_msg *cdc, 249 struct smc_link *link, u64 wr_id) 250 { 251 struct smc_link_group *lgr = container_of(link, struct smc_link_group, 252 lnk[SMC_SINGLE_LINK]); 253 struct smc_connection *connection; 254 struct smc_sock *smc; 255 256 /* lookup connection */ 257 read_lock_bh(&lgr->conns_lock); 258 connection = smc_lgr_find_conn(ntohl(cdc->token), lgr); 259 if (!connection) { 260 read_unlock_bh(&lgr->conns_lock); 261 return; 262 } 263 smc = container_of(connection, struct smc_sock, conn); 264 sock_hold(&smc->sk); 265 read_unlock_bh(&lgr->conns_lock); 266 bh_lock_sock(&smc->sk); 267 smc_cdc_msg_recv_action(smc, link, cdc); 268 bh_unlock_sock(&smc->sk); 269 sock_put(&smc->sk); /* no free sk in softirq-context */ 270 } 271 272 /***************************** init, exit, misc ******************************/ 273 274 static void smc_cdc_rx_handler(struct ib_wc *wc, void *buf) 275 { 276 struct smc_link *link = (struct smc_link *)wc->qp->qp_context; 277 struct smc_cdc_msg *cdc = buf; 278 279 if (wc->byte_len < offsetof(struct smc_cdc_msg, reserved)) 280 return; /* short message */ 281 if (cdc->len != sizeof(*cdc)) 282 return; /* invalid message */ 283 smc_cdc_msg_recv(cdc, link, wc->wr_id); 284 } 285 286 static struct smc_wr_rx_handler smc_cdc_rx_handlers[] = { 287 { 288 .handler = smc_cdc_rx_handler, 289 .type = SMC_CDC_MSG_TYPE 290 }, 291 { 292 .handler = NULL, 293 } 294 }; 295 296 int __init smc_cdc_init(void) 297 { 298 struct smc_wr_rx_handler *handler; 299 int rc = 0; 300 301 for (handler = smc_cdc_rx_handlers; handler->handler; handler++) { 302 INIT_HLIST_NODE(&handler->list); 303 rc = smc_wr_rx_register_handler(handler); 304 if (rc) 305 break; 306 } 307 return rc; 308 } 309