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 /* handler for send/transmission completion of a CDC msg */
25 static void smc_cdc_tx_handler(struct smc_wr_tx_pend_priv *pnd_snd,
26 struct smc_link *link,
27 enum ib_wc_status wc_status)
29 struct smc_cdc_tx_pend *cdcpend = (struct smc_cdc_tx_pend *)pnd_snd;
30 struct smc_connection *conn = cdcpend->conn;
35 /* already dismissed */
38 smc = container_of(conn, struct smc_sock, conn);
39 bh_lock_sock(&smc->sk);
41 diff = smc_curs_diff(cdcpend->conn->sndbuf_desc->len,
42 &cdcpend->conn->tx_curs_fin,
44 /* sndbuf_space is decreased in smc_sendmsg */
45 smp_mb__before_atomic();
46 atomic_add(diff, &cdcpend->conn->sndbuf_space);
47 /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */
48 smp_mb__after_atomic();
49 smc_curs_copy(&conn->tx_curs_fin, &cdcpend->cursor, conn);
51 smc_tx_sndbuf_nonfull(smc);
52 bh_unlock_sock(&smc->sk);
55 int smc_cdc_get_free_slot(struct smc_connection *conn,
56 struct smc_wr_buf **wr_buf,
57 struct smc_rdma_wr **wr_rdma_buf,
58 struct smc_cdc_tx_pend **pend)
60 struct smc_link *link = &conn->lgr->lnk[SMC_SINGLE_LINK];
63 rc = smc_wr_tx_get_free_slot(link, smc_cdc_tx_handler, wr_buf,
65 (struct smc_wr_tx_pend_priv **)pend);
66 if (!conn->alert_token_local)
67 /* abnormal termination */
72 static inline void smc_cdc_add_pending_send(struct smc_connection *conn,
73 struct smc_cdc_tx_pend *pend)
76 sizeof(struct smc_cdc_msg) > SMC_WR_BUF_SIZE,
77 "must increase SMC_WR_BUF_SIZE to at least sizeof(struct smc_cdc_msg)");
79 offsetofend(struct smc_cdc_msg, reserved) > SMC_WR_TX_SIZE,
80 "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 sizeof(struct smc_cdc_tx_pend) > SMC_WR_TX_PEND_PRIV_SIZE,
83 "must increase SMC_WR_TX_PEND_PRIV_SIZE to at least sizeof(struct smc_cdc_tx_pend)");
85 pend->cursor = conn->tx_curs_sent;
86 pend->p_cursor = conn->local_tx_ctrl.prod;
87 pend->ctrl_seq = conn->tx_cdc_seq;
90 int smc_cdc_msg_send(struct smc_connection *conn,
91 struct smc_wr_buf *wr_buf,
92 struct smc_cdc_tx_pend *pend)
94 union smc_host_cursor cfed;
95 struct smc_link *link;
98 link = &conn->lgr->lnk[SMC_SINGLE_LINK];
100 smc_cdc_add_pending_send(conn, pend);
103 conn->local_tx_ctrl.seqno = conn->tx_cdc_seq;
104 smc_host_msg_to_cdc((struct smc_cdc_msg *)wr_buf, conn, &cfed);
105 rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend);
107 smc_curs_copy(&conn->rx_curs_confirmed, &cfed, conn);
112 static int smcr_cdc_get_slot_and_msg_send(struct smc_connection *conn)
114 struct smc_cdc_tx_pend *pend;
115 struct smc_wr_buf *wr_buf;
118 rc = smc_cdc_get_free_slot(conn, &wr_buf, NULL, &pend);
122 spin_lock_bh(&conn->send_lock);
123 rc = smc_cdc_msg_send(conn, wr_buf, pend);
124 spin_unlock_bh(&conn->send_lock);
128 int smc_cdc_get_slot_and_msg_send(struct smc_connection *conn)
132 if (conn->lgr->is_smcd) {
133 spin_lock_bh(&conn->send_lock);
134 rc = smcd_cdc_msg_send(conn);
135 spin_unlock_bh(&conn->send_lock);
137 rc = smcr_cdc_get_slot_and_msg_send(conn);
143 static bool smc_cdc_tx_filter(struct smc_wr_tx_pend_priv *tx_pend,
146 struct smc_connection *conn = (struct smc_connection *)data;
147 struct smc_cdc_tx_pend *cdc_pend =
148 (struct smc_cdc_tx_pend *)tx_pend;
150 return cdc_pend->conn == conn;
153 static void smc_cdc_tx_dismisser(struct smc_wr_tx_pend_priv *tx_pend)
155 struct smc_cdc_tx_pend *cdc_pend =
156 (struct smc_cdc_tx_pend *)tx_pend;
158 cdc_pend->conn = NULL;
161 void smc_cdc_tx_dismiss_slots(struct smc_connection *conn)
163 struct smc_link *link = &conn->lgr->lnk[SMC_SINGLE_LINK];
165 smc_wr_tx_dismiss_slots(link, SMC_CDC_MSG_TYPE,
166 smc_cdc_tx_filter, smc_cdc_tx_dismisser,
167 (unsigned long)conn);
170 /* Send a SMC-D CDC header.
171 * This increments the free space available in our send buffer.
172 * Also update the confirmed receive buffer with what was sent to the peer.
174 int smcd_cdc_msg_send(struct smc_connection *conn)
176 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
177 union smc_host_cursor curs;
178 struct smcd_cdc_msg cdc;
181 memset(&cdc, 0, sizeof(cdc));
182 cdc.common.type = SMC_CDC_MSG_TYPE;
183 curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.prod.acurs);
184 cdc.prod.wrap = curs.wrap;
185 cdc.prod.count = curs.count;
186 curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.cons.acurs);
187 cdc.cons.wrap = curs.wrap;
188 cdc.cons.count = curs.count;
189 cdc.cons.prod_flags = conn->local_tx_ctrl.prod_flags;
190 cdc.cons.conn_state_flags = conn->local_tx_ctrl.conn_state_flags;
191 rc = smcd_tx_ism_write(conn, &cdc, sizeof(cdc), 0, 1);
194 smc_curs_copy(&conn->rx_curs_confirmed, &curs, conn);
195 /* Calculate transmitted data and increment free send buffer space */
196 diff = smc_curs_diff(conn->sndbuf_desc->len, &conn->tx_curs_fin,
197 &conn->tx_curs_sent);
198 /* increased by confirmed number of bytes */
199 smp_mb__before_atomic();
200 atomic_add(diff, &conn->sndbuf_space);
201 /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */
202 smp_mb__after_atomic();
203 smc_curs_copy(&conn->tx_curs_fin, &conn->tx_curs_sent, conn);
205 smc_tx_sndbuf_nonfull(smc);
209 /********************************* receive ***********************************/
211 static inline bool smc_cdc_before(u16 seq1, u16 seq2)
213 return (s16)(seq1 - seq2) < 0;
216 static void smc_cdc_handle_urg_data_arrival(struct smc_sock *smc,
219 struct smc_connection *conn = &smc->conn;
222 /* new data included urgent business */
223 smc_curs_copy(&conn->urg_curs, &conn->local_rx_ctrl.prod, conn);
224 conn->urg_state = SMC_URG_VALID;
225 if (!sock_flag(&smc->sk, SOCK_URGINLINE))
226 /* we'll skip the urgent byte, so don't account for it */
228 base = (char *)conn->rmb_desc->cpu_addr + conn->rx_off;
229 if (conn->urg_curs.count)
230 conn->urg_rx_byte = *(base + conn->urg_curs.count - 1);
232 conn->urg_rx_byte = *(base + conn->rmb_desc->len - 1);
233 sk_send_sigurg(&smc->sk);
236 static void smc_cdc_msg_recv_action(struct smc_sock *smc,
237 struct smc_cdc_msg *cdc)
239 union smc_host_cursor cons_old, prod_old;
240 struct smc_connection *conn = &smc->conn;
241 int diff_cons, diff_prod;
243 smc_curs_copy(&prod_old, &conn->local_rx_ctrl.prod, conn);
244 smc_curs_copy(&cons_old, &conn->local_rx_ctrl.cons, conn);
245 smc_cdc_msg_to_host(&conn->local_rx_ctrl, cdc, conn);
247 diff_cons = smc_curs_diff(conn->peer_rmbe_size, &cons_old,
248 &conn->local_rx_ctrl.cons);
250 /* peer_rmbe_space is decreased during data transfer with RDMA
253 smp_mb__before_atomic();
254 atomic_add(diff_cons, &conn->peer_rmbe_space);
255 /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */
256 smp_mb__after_atomic();
259 diff_prod = smc_curs_diff(conn->rmb_desc->len, &prod_old,
260 &conn->local_rx_ctrl.prod);
262 if (conn->local_rx_ctrl.prod_flags.urg_data_present)
263 smc_cdc_handle_urg_data_arrival(smc, &diff_prod);
264 /* bytes_to_rcv is decreased in smc_recvmsg */
265 smp_mb__before_atomic();
266 atomic_add(diff_prod, &conn->bytes_to_rcv);
267 /* guarantee 0 <= bytes_to_rcv <= rmb_desc->len */
268 smp_mb__after_atomic();
269 smc->sk.sk_data_ready(&smc->sk);
271 if (conn->local_rx_ctrl.prod_flags.write_blocked ||
272 conn->local_rx_ctrl.prod_flags.cons_curs_upd_req ||
273 conn->local_rx_ctrl.prod_flags.urg_data_pending) {
274 if (conn->local_rx_ctrl.prod_flags.urg_data_pending)
275 conn->urg_state = SMC_URG_NOTYET;
276 /* force immediate tx of current consumer cursor, but
277 * under send_lock to guarantee arrival in seqno-order
279 if (smc->sk.sk_state != SMC_INIT)
280 smc_tx_sndbuf_nonempty(conn);
284 /* piggy backed tx info */
285 /* trigger sndbuf consumer: RDMA write into peer RMBE and CDC */
286 if (diff_cons && smc_tx_prepared_sends(conn)) {
287 smc_tx_sndbuf_nonempty(conn);
288 /* trigger socket release if connection closed */
289 smc_close_wake_tx_prepared(smc);
291 if (diff_cons && conn->urg_tx_pend &&
292 atomic_read(&conn->peer_rmbe_space) == conn->peer_rmbe_size) {
293 /* urg data confirmed by peer, indicate we're ready for more */
294 conn->urg_tx_pend = false;
295 smc->sk.sk_write_space(&smc->sk);
298 if (conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) {
299 smc->sk.sk_err = ECONNRESET;
300 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
302 if (smc_cdc_rxed_any_close_or_senddone(conn)) {
303 smc->sk.sk_shutdown |= RCV_SHUTDOWN;
304 if (smc->clcsock && smc->clcsock->sk)
305 smc->clcsock->sk->sk_shutdown |= RCV_SHUTDOWN;
306 sock_set_flag(&smc->sk, SOCK_DONE);
307 sock_hold(&smc->sk); /* sock_put in close_work */
308 if (!schedule_work(&conn->close_work))
313 /* called under tasklet context */
314 static void smc_cdc_msg_recv(struct smc_sock *smc, struct smc_cdc_msg *cdc)
317 bh_lock_sock(&smc->sk);
318 smc_cdc_msg_recv_action(smc, cdc);
319 bh_unlock_sock(&smc->sk);
320 sock_put(&smc->sk); /* no free sk in softirq-context */
323 /* Schedule a tasklet for this connection. Triggered from the ISM device IRQ
324 * handler to indicate update in the DMBE.
329 static void smcd_cdc_rx_tsklet(unsigned long data)
331 struct smc_connection *conn = (struct smc_connection *)data;
332 struct smcd_cdc_msg *data_cdc;
333 struct smcd_cdc_msg cdc;
334 struct smc_sock *smc;
339 data_cdc = (struct smcd_cdc_msg *)conn->rmb_desc->cpu_addr;
340 smcd_curs_copy(&cdc.prod, &data_cdc->prod, conn);
341 smcd_curs_copy(&cdc.cons, &data_cdc->cons, conn);
342 smc = container_of(conn, struct smc_sock, conn);
343 smc_cdc_msg_recv(smc, (struct smc_cdc_msg *)&cdc);
346 /* Initialize receive tasklet. Called from ISM device IRQ handler to start
349 void smcd_cdc_rx_init(struct smc_connection *conn)
351 tasklet_init(&conn->rx_tsklet, smcd_cdc_rx_tsklet, (unsigned long)conn);
354 /***************************** init, exit, misc ******************************/
356 static void smc_cdc_rx_handler(struct ib_wc *wc, void *buf)
358 struct smc_link *link = (struct smc_link *)wc->qp->qp_context;
359 struct smc_cdc_msg *cdc = buf;
360 struct smc_connection *conn;
361 struct smc_link_group *lgr;
362 struct smc_sock *smc;
364 if (wc->byte_len < offsetof(struct smc_cdc_msg, reserved))
365 return; /* short message */
366 if (cdc->len != SMC_WR_TX_SIZE)
367 return; /* invalid message */
369 /* lookup connection */
370 lgr = smc_get_lgr(link);
371 read_lock_bh(&lgr->conns_lock);
372 conn = smc_lgr_find_conn(ntohl(cdc->token), lgr);
373 read_unlock_bh(&lgr->conns_lock);
376 smc = container_of(conn, struct smc_sock, conn);
378 if (!cdc->prod_flags.failover_validation) {
379 if (smc_cdc_before(ntohs(cdc->seqno),
380 conn->local_rx_ctrl.seqno))
381 /* received seqno is old */
384 smc_cdc_msg_recv(smc, cdc);
387 static struct smc_wr_rx_handler smc_cdc_rx_handlers[] = {
389 .handler = smc_cdc_rx_handler,
390 .type = SMC_CDC_MSG_TYPE
397 int __init smc_cdc_init(void)
399 struct smc_wr_rx_handler *handler;
402 for (handler = smc_cdc_rx_handlers; handler->handler; handler++) {
403 INIT_HLIST_NODE(&handler->list);
404 rc = smc_wr_rx_register_handler(handler);