1 // SPDX-License-Identifier: GPL-2.0
3 * Shared Memory Communications over RDMA (SMC-R) and RoCE
5 * Connection Data Control (CDC)
8 * Copyright IBM Corp. 2016
10 * Author(s): Ursula Braun <ubraun@linux.vnet.ibm.com>
13 #include <linux/spinlock.h>
20 #include "smc_close.h"
22 /********************************** send *************************************/
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 # */
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)
36 struct smc_cdc_tx_pend *cdcpend = (struct smc_cdc_tx_pend *)pnd_snd;
41 /* already dismissed */
44 smc = container_of(cdcpend->conn, struct smc_sock, conn);
45 bh_lock_sock(&smc->sk);
47 diff = smc_curs_diff(cdcpend->conn->sndbuf_size,
48 &cdcpend->conn->tx_curs_fin,
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),
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);
66 int smc_cdc_get_free_slot(struct smc_link *link,
67 struct smc_wr_buf **wr_buf,
68 struct smc_cdc_tx_pend **pend)
70 return smc_wr_tx_get_free_slot(link, smc_cdc_tx_handler, wr_buf,
71 (struct smc_wr_tx_pend_priv **)pend);
74 static inline void smc_cdc_add_pending_send(struct smc_connection *conn,
75 struct smc_cdc_tx_pend *pend)
78 sizeof(struct smc_cdc_msg) > SMC_WR_BUF_SIZE,
79 "must increase SMC_WR_BUF_SIZE to at least sizeof(struct smc_cdc_msg)");
81 offsetof(struct smc_cdc_msg, reserved) > SMC_WR_TX_SIZE,
82 "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 sizeof(struct smc_cdc_tx_pend) > SMC_WR_TX_PEND_PRIV_SIZE,
85 "must increase SMC_WR_TX_PEND_PRIV_SIZE to at least sizeof(struct smc_cdc_tx_pend)");
87 pend->cursor = conn->tx_curs_sent;
88 pend->p_cursor = conn->local_tx_ctrl.prod;
89 pend->ctrl_seq = conn->tx_cdc_seq;
92 int smc_cdc_msg_send(struct smc_connection *conn,
93 struct smc_wr_buf *wr_buf,
94 struct smc_cdc_tx_pend *pend)
96 struct smc_link *link;
99 link = &conn->lgr->lnk[SMC_SINGLE_LINK];
101 smc_cdc_add_pending_send(conn, pend);
104 conn->local_tx_ctrl.seqno = conn->tx_cdc_seq;
105 smc_host_msg_to_cdc((struct smc_cdc_msg *)wr_buf,
106 &conn->local_tx_ctrl, conn);
107 rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend);
109 smc_curs_write(&conn->rx_curs_confirmed,
110 smc_curs_read(&conn->local_tx_ctrl.cons, conn),
116 int smc_cdc_get_slot_and_msg_send(struct smc_connection *conn)
118 struct smc_cdc_tx_pend *pend;
119 struct smc_wr_buf *wr_buf;
122 rc = smc_cdc_get_free_slot(&conn->lgr->lnk[SMC_SINGLE_LINK], &wr_buf,
127 return smc_cdc_msg_send(conn, wr_buf, pend);
130 static bool smc_cdc_tx_filter(struct smc_wr_tx_pend_priv *tx_pend,
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;
137 return cdc_pend->conn == conn;
140 static void smc_cdc_tx_dismisser(struct smc_wr_tx_pend_priv *tx_pend)
142 struct smc_cdc_tx_pend *cdc_pend =
143 (struct smc_cdc_tx_pend *)tx_pend;
145 cdc_pend->conn = NULL;
148 void smc_cdc_tx_dismiss_slots(struct smc_connection *conn)
150 struct smc_link *link = &conn->lgr->lnk[SMC_SINGLE_LINK];
152 smc_wr_tx_dismiss_slots(link, SMC_CDC_MSG_TYPE,
153 smc_cdc_tx_filter, smc_cdc_tx_dismisser,
154 (unsigned long)conn);
157 bool smc_cdc_tx_has_pending(struct smc_connection *conn)
159 struct smc_link *link = &conn->lgr->lnk[SMC_SINGLE_LINK];
161 return smc_wr_tx_has_pending(link, SMC_CDC_MSG_TYPE,
162 smc_cdc_tx_filter, (unsigned long)conn);
165 /********************************* receive ***********************************/
167 static inline bool smc_cdc_before(u16 seq1, u16 seq2)
169 return (s16)(seq1 - seq2) < 0;
172 static void smc_cdc_msg_recv_action(struct smc_sock *smc,
173 struct smc_link *link,
174 struct smc_cdc_msg *cdc)
176 union smc_host_cursor cons_old, prod_old;
177 struct smc_connection *conn = &smc->conn;
178 int diff_cons, diff_prod;
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 */
186 smc_curs_write(&prod_old,
187 smc_curs_read(&conn->local_rx_ctrl.prod, conn),
189 smc_curs_write(&cons_old,
190 smc_curs_read(&conn->local_rx_ctrl.cons, conn),
192 smc_cdc_msg_to_host(&conn->local_rx_ctrl, cdc, conn);
194 diff_cons = smc_curs_diff(conn->peer_rmbe_size, &cons_old,
195 &conn->local_rx_ctrl.cons);
197 /* peer_rmbe_space is decreased during data transfer with RDMA
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();
206 diff_prod = smc_curs_diff(conn->rmbe_size, &prod_old,
207 &conn->local_rx_ctrl.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);
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;
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);
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);
237 /* socket connected but not accepted */
238 if (!smc->sk.sk_socket)
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);
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)
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;
256 /* lookup connection */
257 read_lock_bh(&lgr->conns_lock);
258 connection = smc_lgr_find_conn(ntohl(cdc->token), lgr);
260 read_unlock_bh(&lgr->conns_lock);
263 smc = container_of(connection, struct smc_sock, conn);
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 */
272 /***************************** init, exit, misc ******************************/
274 static void smc_cdc_rx_handler(struct ib_wc *wc, void *buf)
276 struct smc_link *link = (struct smc_link *)wc->qp->qp_context;
277 struct smc_cdc_msg *cdc = buf;
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);
286 static struct smc_wr_rx_handler smc_cdc_rx_handlers[] = {
288 .handler = smc_cdc_rx_handler,
289 .type = SMC_CDC_MSG_TYPE
296 int __init smc_cdc_init(void)
298 struct smc_wr_rx_handler *handler;
301 for (handler = smc_cdc_rx_handlers; handler->handler; handler++) {
302 INIT_HLIST_NODE(&handler->list);
303 rc = smc_wr_rx_register_handler(handler);