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;
37 struct smc_connection *conn = cdcpend->conn;
42 /* already dismissed */
45 smc = container_of(conn, struct smc_sock, conn);
46 bh_lock_sock(&smc->sk);
48 diff = smc_curs_diff(cdcpend->conn->sndbuf_desc->len,
49 &cdcpend->conn->tx_curs_fin,
51 /* sndbuf_space is decreased in smc_sendmsg */
52 smp_mb__before_atomic();
53 atomic_add(diff, &cdcpend->conn->sndbuf_space);
54 /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */
55 smp_mb__after_atomic();
56 smc_curs_copy(&conn->tx_curs_fin, &cdcpend->cursor, conn);
58 smc_tx_sndbuf_nonfull(smc);
59 bh_unlock_sock(&smc->sk);
62 int smc_cdc_get_free_slot(struct smc_connection *conn,
63 struct smc_wr_buf **wr_buf,
64 struct smc_cdc_tx_pend **pend)
66 struct smc_link *link = &conn->lgr->lnk[SMC_SINGLE_LINK];
69 rc = smc_wr_tx_get_free_slot(link, smc_cdc_tx_handler, wr_buf,
70 (struct smc_wr_tx_pend_priv **)pend);
71 if (!conn->alert_token_local)
72 /* abnormal termination */
77 static inline void smc_cdc_add_pending_send(struct smc_connection *conn,
78 struct smc_cdc_tx_pend *pend)
81 sizeof(struct smc_cdc_msg) > SMC_WR_BUF_SIZE,
82 "must increase SMC_WR_BUF_SIZE to at least sizeof(struct smc_cdc_msg)");
84 sizeof(struct smc_cdc_msg) != SMC_WR_TX_SIZE,
85 "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()");
87 sizeof(struct smc_cdc_tx_pend) > SMC_WR_TX_PEND_PRIV_SIZE,
88 "must increase SMC_WR_TX_PEND_PRIV_SIZE to at least sizeof(struct smc_cdc_tx_pend)");
90 pend->cursor = conn->tx_curs_sent;
91 pend->p_cursor = conn->local_tx_ctrl.prod;
92 pend->ctrl_seq = conn->tx_cdc_seq;
95 int smc_cdc_msg_send(struct smc_connection *conn,
96 struct smc_wr_buf *wr_buf,
97 struct smc_cdc_tx_pend *pend)
99 union smc_host_cursor cfed;
100 struct smc_link *link;
103 link = &conn->lgr->lnk[SMC_SINGLE_LINK];
105 smc_cdc_add_pending_send(conn, pend);
108 conn->local_tx_ctrl.seqno = conn->tx_cdc_seq;
109 smc_host_msg_to_cdc((struct smc_cdc_msg *)wr_buf, conn, &cfed);
110 rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend);
112 smc_curs_copy(&conn->rx_curs_confirmed, &cfed, conn);
117 static int smcr_cdc_get_slot_and_msg_send(struct smc_connection *conn)
119 struct smc_cdc_tx_pend *pend;
120 struct smc_wr_buf *wr_buf;
123 rc = smc_cdc_get_free_slot(conn, &wr_buf, &pend);
127 return smc_cdc_msg_send(conn, wr_buf, pend);
130 int smc_cdc_get_slot_and_msg_send(struct smc_connection *conn)
134 if (conn->lgr->is_smcd) {
135 spin_lock_bh(&conn->send_lock);
136 rc = smcd_cdc_msg_send(conn);
137 spin_unlock_bh(&conn->send_lock);
139 rc = smcr_cdc_get_slot_and_msg_send(conn);
145 static bool smc_cdc_tx_filter(struct smc_wr_tx_pend_priv *tx_pend,
148 struct smc_connection *conn = (struct smc_connection *)data;
149 struct smc_cdc_tx_pend *cdc_pend =
150 (struct smc_cdc_tx_pend *)tx_pend;
152 return cdc_pend->conn == conn;
155 static void smc_cdc_tx_dismisser(struct smc_wr_tx_pend_priv *tx_pend)
157 struct smc_cdc_tx_pend *cdc_pend =
158 (struct smc_cdc_tx_pend *)tx_pend;
160 cdc_pend->conn = NULL;
163 void smc_cdc_tx_dismiss_slots(struct smc_connection *conn)
165 struct smc_link *link = &conn->lgr->lnk[SMC_SINGLE_LINK];
167 smc_wr_tx_dismiss_slots(link, SMC_CDC_MSG_TYPE,
168 smc_cdc_tx_filter, smc_cdc_tx_dismisser,
169 (unsigned long)conn);
172 /* Send a SMC-D CDC header.
173 * This increments the free space available in our send buffer.
174 * Also update the confirmed receive buffer with what was sent to the peer.
176 int smcd_cdc_msg_send(struct smc_connection *conn)
178 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
179 struct smcd_cdc_msg cdc;
182 memset(&cdc, 0, sizeof(cdc));
183 cdc.common.type = SMC_CDC_MSG_TYPE;
184 cdc.prod_wrap = conn->local_tx_ctrl.prod.wrap;
185 cdc.prod_count = conn->local_tx_ctrl.prod.count;
187 cdc.cons_wrap = conn->local_tx_ctrl.cons.wrap;
188 cdc.cons_count = conn->local_tx_ctrl.cons.count;
189 cdc.prod_flags = conn->local_tx_ctrl.prod_flags;
190 cdc.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, &conn->local_tx_ctrl.cons,
196 /* Calculate transmitted data and increment free send buffer space */
197 diff = smc_curs_diff(conn->sndbuf_desc->len, &conn->tx_curs_fin,
198 &conn->tx_curs_sent);
199 /* increased by confirmed number of bytes */
200 smp_mb__before_atomic();
201 atomic_add(diff, &conn->sndbuf_space);
202 /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */
203 smp_mb__after_atomic();
204 smc_curs_copy(&conn->tx_curs_fin, &conn->tx_curs_sent, conn);
206 smc_tx_sndbuf_nonfull(smc);
210 /********************************* receive ***********************************/
212 static inline bool smc_cdc_before(u16 seq1, u16 seq2)
214 return (s16)(seq1 - seq2) < 0;
217 static void smc_cdc_handle_urg_data_arrival(struct smc_sock *smc,
220 struct smc_connection *conn = &smc->conn;
223 /* new data included urgent business */
224 smc_curs_copy(&conn->urg_curs, &conn->local_rx_ctrl.prod, conn);
225 conn->urg_state = SMC_URG_VALID;
226 if (!sock_flag(&smc->sk, SOCK_URGINLINE))
227 /* we'll skip the urgent byte, so don't account for it */
229 base = (char *)conn->rmb_desc->cpu_addr + conn->rx_off;
230 if (conn->urg_curs.count)
231 conn->urg_rx_byte = *(base + conn->urg_curs.count - 1);
233 conn->urg_rx_byte = *(base + conn->rmb_desc->len - 1);
234 sk_send_sigurg(&smc->sk);
237 static void smc_cdc_msg_recv_action(struct smc_sock *smc,
238 struct smc_cdc_msg *cdc)
240 union smc_host_cursor cons_old, prod_old;
241 struct smc_connection *conn = &smc->conn;
242 int diff_cons, diff_prod;
244 smc_curs_copy(&prod_old, &conn->local_rx_ctrl.prod, conn);
245 smc_curs_copy(&cons_old, &conn->local_rx_ctrl.cons, conn);
246 smc_cdc_msg_to_host(&conn->local_rx_ctrl, cdc, conn);
248 diff_cons = smc_curs_diff(conn->peer_rmbe_size, &cons_old,
249 &conn->local_rx_ctrl.cons);
251 /* peer_rmbe_space is decreased during data transfer with RDMA
254 smp_mb__before_atomic();
255 atomic_add(diff_cons, &conn->peer_rmbe_space);
256 /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */
257 smp_mb__after_atomic();
260 diff_prod = smc_curs_diff(conn->rmb_desc->len, &prod_old,
261 &conn->local_rx_ctrl.prod);
263 if (conn->local_rx_ctrl.prod_flags.urg_data_present)
264 smc_cdc_handle_urg_data_arrival(smc, &diff_prod);
265 /* bytes_to_rcv is decreased in smc_recvmsg */
266 smp_mb__before_atomic();
267 atomic_add(diff_prod, &conn->bytes_to_rcv);
268 /* guarantee 0 <= bytes_to_rcv <= rmb_desc->len */
269 smp_mb__after_atomic();
270 smc->sk.sk_data_ready(&smc->sk);
272 if (conn->local_rx_ctrl.prod_flags.write_blocked ||
273 conn->local_rx_ctrl.prod_flags.cons_curs_upd_req ||
274 conn->local_rx_ctrl.prod_flags.urg_data_pending) {
275 if (conn->local_rx_ctrl.prod_flags.urg_data_pending)
276 conn->urg_state = SMC_URG_NOTYET;
277 /* force immediate tx of current consumer cursor, but
278 * under send_lock to guarantee arrival in seqno-order
280 if (smc->sk.sk_state != SMC_INIT)
281 smc_tx_sndbuf_nonempty(conn);
285 /* piggy backed tx info */
286 /* trigger sndbuf consumer: RDMA write into peer RMBE and CDC */
287 if (diff_cons && smc_tx_prepared_sends(conn)) {
288 smc_tx_sndbuf_nonempty(conn);
289 /* trigger socket release if connection closed */
290 smc_close_wake_tx_prepared(smc);
292 if (diff_cons && conn->urg_tx_pend &&
293 atomic_read(&conn->peer_rmbe_space) == conn->peer_rmbe_size) {
294 /* urg data confirmed by peer, indicate we're ready for more */
295 conn->urg_tx_pend = false;
296 smc->sk.sk_write_space(&smc->sk);
299 if (conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) {
300 smc->sk.sk_err = ECONNRESET;
301 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
303 if (smc_cdc_rxed_any_close_or_senddone(conn)) {
304 smc->sk.sk_shutdown |= RCV_SHUTDOWN;
305 if (smc->clcsock && smc->clcsock->sk)
306 smc->clcsock->sk->sk_shutdown |= RCV_SHUTDOWN;
307 sock_set_flag(&smc->sk, SOCK_DONE);
308 sock_hold(&smc->sk); /* sock_put in close_work */
309 if (!schedule_work(&conn->close_work))
314 /* called under tasklet context */
315 static void smc_cdc_msg_recv(struct smc_sock *smc, struct smc_cdc_msg *cdc)
318 bh_lock_sock(&smc->sk);
319 smc_cdc_msg_recv_action(smc, cdc);
320 bh_unlock_sock(&smc->sk);
321 sock_put(&smc->sk); /* no free sk in softirq-context */
324 /* Schedule a tasklet for this connection. Triggered from the ISM device IRQ
325 * handler to indicate update in the DMBE.
330 static void smcd_cdc_rx_tsklet(unsigned long data)
332 struct smc_connection *conn = (struct smc_connection *)data;
333 struct smcd_cdc_msg cdc;
334 struct smc_sock *smc;
339 memcpy(&cdc, conn->rmb_desc->cpu_addr, sizeof(cdc));
340 smc = container_of(conn, struct smc_sock, conn);
341 smc_cdc_msg_recv(smc, (struct smc_cdc_msg *)&cdc);
344 /* Initialize receive tasklet. Called from ISM device IRQ handler to start
347 void smcd_cdc_rx_init(struct smc_connection *conn)
349 tasklet_init(&conn->rx_tsklet, smcd_cdc_rx_tsklet, (unsigned long)conn);
352 /***************************** init, exit, misc ******************************/
354 static void smc_cdc_rx_handler(struct ib_wc *wc, void *buf)
356 struct smc_link *link = (struct smc_link *)wc->qp->qp_context;
357 struct smc_cdc_msg *cdc = buf;
358 struct smc_connection *conn;
359 struct smc_link_group *lgr;
360 struct smc_sock *smc;
362 if (wc->byte_len < offsetof(struct smc_cdc_msg, reserved))
363 return; /* short message */
364 if (cdc->len != SMC_WR_TX_SIZE)
365 return; /* invalid message */
367 /* lookup connection */
368 lgr = smc_get_lgr(link);
369 read_lock_bh(&lgr->conns_lock);
370 conn = smc_lgr_find_conn(ntohl(cdc->token), lgr);
371 read_unlock_bh(&lgr->conns_lock);
374 smc = container_of(conn, struct smc_sock, conn);
376 if (!cdc->prod_flags.failover_validation) {
377 if (smc_cdc_before(ntohs(cdc->seqno),
378 conn->local_rx_ctrl.seqno))
379 /* received seqno is old */
382 smc_cdc_msg_recv(smc, cdc);
385 static struct smc_wr_rx_handler smc_cdc_rx_handlers[] = {
387 .handler = smc_cdc_rx_handler,
388 .type = SMC_CDC_MSG_TYPE
395 int __init smc_cdc_init(void)
397 struct smc_wr_rx_handler *handler;
400 for (handler = smc_cdc_rx_handlers; handler->handler; handler++) {
401 INIT_HLIST_NODE(&handler->list);
402 rc = smc_wr_rx_register_handler(handler);