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_link *link, 66 struct smc_wr_buf **wr_buf, 67 struct smc_cdc_tx_pend **pend) 68 { 69 return smc_wr_tx_get_free_slot(link, smc_cdc_tx_handler, wr_buf, 70 (struct smc_wr_tx_pend_priv **)pend); 71 } 72 73 static inline void smc_cdc_add_pending_send(struct smc_connection *conn, 74 struct smc_cdc_tx_pend *pend) 75 { 76 BUILD_BUG_ON_MSG( 77 sizeof(struct smc_cdc_msg) > SMC_WR_BUF_SIZE, 78 "must increase SMC_WR_BUF_SIZE to at least sizeof(struct smc_cdc_msg)"); 79 BUILD_BUG_ON_MSG( 80 offsetof(struct smc_cdc_msg, reserved) > SMC_WR_TX_SIZE, 81 "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()"); 82 BUILD_BUG_ON_MSG( 83 sizeof(struct smc_cdc_tx_pend) > SMC_WR_TX_PEND_PRIV_SIZE, 84 "must increase SMC_WR_TX_PEND_PRIV_SIZE to at least sizeof(struct smc_cdc_tx_pend)"); 85 pend->conn = conn; 86 pend->cursor = conn->tx_curs_sent; 87 pend->p_cursor = conn->local_tx_ctrl.prod; 88 pend->ctrl_seq = conn->tx_cdc_seq; 89 } 90 91 int smc_cdc_msg_send(struct smc_connection *conn, 92 struct smc_wr_buf *wr_buf, 93 struct smc_cdc_tx_pend *pend) 94 { 95 struct smc_link *link; 96 int rc; 97 98 link = &conn->lgr->lnk[SMC_SINGLE_LINK]; 99 100 smc_cdc_add_pending_send(conn, pend); 101 102 conn->tx_cdc_seq++; 103 conn->local_tx_ctrl.seqno = conn->tx_cdc_seq; 104 smc_host_msg_to_cdc((struct smc_cdc_msg *)wr_buf, 105 &conn->local_tx_ctrl, conn); 106 rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend); 107 if (!rc) 108 smc_curs_write(&conn->rx_curs_confirmed, 109 smc_curs_read(&conn->local_tx_ctrl.cons, conn), 110 conn); 111 112 return rc; 113 } 114 115 int smc_cdc_get_slot_and_msg_send(struct smc_connection *conn) 116 { 117 struct smc_cdc_tx_pend *pend; 118 struct smc_wr_buf *wr_buf; 119 int rc; 120 121 rc = smc_cdc_get_free_slot(&conn->lgr->lnk[SMC_SINGLE_LINK], &wr_buf, 122 &pend); 123 if (rc) 124 return rc; 125 126 return smc_cdc_msg_send(conn, wr_buf, pend); 127 } 128 129 static bool smc_cdc_tx_filter(struct smc_wr_tx_pend_priv *tx_pend, 130 unsigned long data) 131 { 132 struct smc_connection *conn = (struct smc_connection *)data; 133 struct smc_cdc_tx_pend *cdc_pend = 134 (struct smc_cdc_tx_pend *)tx_pend; 135 136 return cdc_pend->conn == conn; 137 } 138 139 static void smc_cdc_tx_dismisser(struct smc_wr_tx_pend_priv *tx_pend) 140 { 141 struct smc_cdc_tx_pend *cdc_pend = 142 (struct smc_cdc_tx_pend *)tx_pend; 143 144 cdc_pend->conn = NULL; 145 } 146 147 void smc_cdc_tx_dismiss_slots(struct smc_connection *conn) 148 { 149 struct smc_link *link = &conn->lgr->lnk[SMC_SINGLE_LINK]; 150 151 smc_wr_tx_dismiss_slots(link, SMC_CDC_MSG_TYPE, 152 smc_cdc_tx_filter, smc_cdc_tx_dismisser, 153 (unsigned long)conn); 154 } 155 156 bool smc_cdc_tx_has_pending(struct smc_connection *conn) 157 { 158 struct smc_link *link = &conn->lgr->lnk[SMC_SINGLE_LINK]; 159 160 return smc_wr_tx_has_pending(link, SMC_CDC_MSG_TYPE, 161 smc_cdc_tx_filter, (unsigned long)conn); 162 } 163 164 /********************************* receive ***********************************/ 165 166 static inline bool smc_cdc_before(u16 seq1, u16 seq2) 167 { 168 return (s16)(seq1 - seq2) < 0; 169 } 170 171 static void smc_cdc_msg_recv_action(struct smc_sock *smc, 172 struct smc_link *link, 173 struct smc_cdc_msg *cdc) 174 { 175 union smc_host_cursor cons_old, prod_old; 176 struct smc_connection *conn = &smc->conn; 177 int diff_cons, diff_prod; 178 179 if (!cdc->prod_flags.failover_validation) { 180 if (smc_cdc_before(ntohs(cdc->seqno), 181 conn->local_rx_ctrl.seqno)) 182 /* received seqno is old */ 183 return; 184 } 185 smc_curs_write(&prod_old, 186 smc_curs_read(&conn->local_rx_ctrl.prod, conn), 187 conn); 188 smc_curs_write(&cons_old, 189 smc_curs_read(&conn->local_rx_ctrl.cons, conn), 190 conn); 191 smc_cdc_msg_to_host(&conn->local_rx_ctrl, cdc, conn); 192 193 diff_cons = smc_curs_diff(conn->peer_rmbe_size, &cons_old, 194 &conn->local_rx_ctrl.cons); 195 if (diff_cons) { 196 /* peer_rmbe_space is decreased during data transfer with RDMA 197 * write 198 */ 199 smp_mb__before_atomic(); 200 atomic_add(diff_cons, &conn->peer_rmbe_space); 201 /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */ 202 smp_mb__after_atomic(); 203 } 204 205 diff_prod = smc_curs_diff(conn->rmbe_size, &prod_old, 206 &conn->local_rx_ctrl.prod); 207 if (diff_prod) { 208 /* bytes_to_rcv is decreased in smc_recvmsg */ 209 smp_mb__before_atomic(); 210 atomic_add(diff_prod, &conn->bytes_to_rcv); 211 /* guarantee 0 <= bytes_to_rcv <= rmbe_size */ 212 smp_mb__after_atomic(); 213 smc->sk.sk_data_ready(&smc->sk); 214 } 215 216 if (conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) { 217 smc->sk.sk_err = ECONNRESET; 218 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1; 219 } 220 if (smc_cdc_rxed_any_close_or_senddone(conn)) { 221 smc->sk.sk_shutdown |= RCV_SHUTDOWN; 222 if (smc->clcsock && smc->clcsock->sk) 223 smc->clcsock->sk->sk_shutdown |= RCV_SHUTDOWN; 224 sock_set_flag(&smc->sk, SOCK_DONE); 225 schedule_work(&conn->close_work); 226 } 227 228 /* piggy backed tx info */ 229 /* trigger sndbuf consumer: RDMA write into peer RMBE and CDC */ 230 if (diff_cons && smc_tx_prepared_sends(conn)) { 231 smc_tx_sndbuf_nonempty(conn); 232 /* trigger socket release if connection closed */ 233 smc_close_wake_tx_prepared(smc); 234 } 235 236 /* socket connected but not accepted */ 237 if (!smc->sk.sk_socket) 238 return; 239 240 /* data available */ 241 if ((conn->local_rx_ctrl.prod_flags.write_blocked) || 242 (conn->local_rx_ctrl.prod_flags.cons_curs_upd_req)) 243 smc_tx_consumer_update(conn); 244 } 245 246 /* called under tasklet context */ 247 static inline void smc_cdc_msg_recv(struct smc_cdc_msg *cdc, 248 struct smc_link *link, u64 wr_id) 249 { 250 struct smc_link_group *lgr = container_of(link, struct smc_link_group, 251 lnk[SMC_SINGLE_LINK]); 252 struct smc_connection *connection; 253 struct smc_sock *smc; 254 255 /* lookup connection */ 256 read_lock_bh(&lgr->conns_lock); 257 connection = smc_lgr_find_conn(ntohl(cdc->token), lgr); 258 if (!connection) { 259 read_unlock_bh(&lgr->conns_lock); 260 return; 261 } 262 smc = container_of(connection, struct smc_sock, conn); 263 sock_hold(&smc->sk); 264 read_unlock_bh(&lgr->conns_lock); 265 bh_lock_sock(&smc->sk); 266 smc_cdc_msg_recv_action(smc, link, cdc); 267 bh_unlock_sock(&smc->sk); 268 sock_put(&smc->sk); /* no free sk in softirq-context */ 269 } 270 271 /***************************** init, exit, misc ******************************/ 272 273 static void smc_cdc_rx_handler(struct ib_wc *wc, void *buf) 274 { 275 struct smc_link *link = (struct smc_link *)wc->qp->qp_context; 276 struct smc_cdc_msg *cdc = buf; 277 278 if (wc->byte_len < offsetof(struct smc_cdc_msg, reserved)) 279 return; /* short message */ 280 if (cdc->len != sizeof(*cdc)) 281 return; /* invalid message */ 282 smc_cdc_msg_recv(cdc, link, wc->wr_id); 283 } 284 285 static struct smc_wr_rx_handler smc_cdc_rx_handlers[] = { 286 { 287 .handler = smc_cdc_rx_handler, 288 .type = SMC_CDC_MSG_TYPE 289 }, 290 { 291 .handler = NULL, 292 } 293 }; 294 295 int __init smc_cdc_init(void) 296 { 297 struct smc_wr_rx_handler *handler; 298 int rc = 0; 299 300 for (handler = smc_cdc_rx_handlers; handler->handler; handler++) { 301 INIT_HLIST_NODE(&handler->list); 302 rc = smc_wr_rx_register_handler(handler); 303 if (rc) 304 break; 305 } 306 return rc; 307 } 308