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;
34 smc = container_of(conn, struct smc_sock, conn);
35 bh_lock_sock(&smc->sk);
37 diff = smc_curs_diff(cdcpend->conn->sndbuf_desc->len,
38 &cdcpend->conn->tx_curs_fin,
40 /* sndbuf_space is decreased in smc_sendmsg */
41 smp_mb__before_atomic();
42 atomic_add(diff, &cdcpend->conn->sndbuf_space);
43 /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */
44 smp_mb__after_atomic();
45 smc_curs_copy(&conn->tx_curs_fin, &cdcpend->cursor, conn);
46 smc_curs_copy(&conn->local_tx_ctrl_fin, &cdcpend->p_cursor,
48 conn->tx_cdc_seq_fin = cdcpend->ctrl_seq;
51 if (atomic_dec_and_test(&conn->cdc_pend_tx_wr) &&
52 unlikely(wq_has_sleeper(&conn->cdc_pend_tx_wq)))
53 wake_up(&conn->cdc_pend_tx_wq);
54 WARN_ON(atomic_read(&conn->cdc_pend_tx_wr) < 0);
56 smc_tx_sndbuf_nonfull(smc);
57 bh_unlock_sock(&smc->sk);
60 int smc_cdc_get_free_slot(struct smc_connection *conn,
61 struct smc_link *link,
62 struct smc_wr_buf **wr_buf,
63 struct smc_rdma_wr **wr_rdma_buf,
64 struct smc_cdc_tx_pend **pend)
68 rc = smc_wr_tx_get_free_slot(link, smc_cdc_tx_handler, wr_buf,
70 (struct smc_wr_tx_pend_priv **)pend);
72 /* abnormal termination */
74 smc_wr_tx_put_slot(link,
75 (struct smc_wr_tx_pend_priv *)(*pend));
81 static inline void smc_cdc_add_pending_send(struct smc_connection *conn,
82 struct smc_cdc_tx_pend *pend)
85 sizeof(struct smc_cdc_msg) > SMC_WR_BUF_SIZE,
86 "must increase SMC_WR_BUF_SIZE to at least sizeof(struct smc_cdc_msg)");
88 offsetofend(struct smc_cdc_msg, reserved) > SMC_WR_TX_SIZE,
89 "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()");
91 sizeof(struct smc_cdc_tx_pend) > SMC_WR_TX_PEND_PRIV_SIZE,
92 "must increase SMC_WR_TX_PEND_PRIV_SIZE to at least sizeof(struct smc_cdc_tx_pend)");
94 pend->cursor = conn->tx_curs_sent;
95 pend->p_cursor = conn->local_tx_ctrl.prod;
96 pend->ctrl_seq = conn->tx_cdc_seq;
99 int smc_cdc_msg_send(struct smc_connection *conn,
100 struct smc_wr_buf *wr_buf,
101 struct smc_cdc_tx_pend *pend)
103 struct smc_link *link = conn->lnk;
104 union smc_host_cursor cfed;
107 if (unlikely(!READ_ONCE(conn->sndbuf_desc)))
110 smc_cdc_add_pending_send(conn, pend);
113 conn->local_tx_ctrl.seqno = conn->tx_cdc_seq;
114 smc_host_msg_to_cdc((struct smc_cdc_msg *)wr_buf, conn, &cfed);
116 atomic_inc(&conn->cdc_pend_tx_wr);
117 smp_mb__after_atomic(); /* Make sure cdc_pend_tx_wr added before post */
119 rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend);
121 smc_curs_copy(&conn->rx_curs_confirmed, &cfed, conn);
122 conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0;
125 conn->local_tx_ctrl.seqno = conn->tx_cdc_seq;
126 atomic_dec(&conn->cdc_pend_tx_wr);
132 /* send a validation msg indicating the move of a conn to an other QP link */
133 int smcr_cdc_msg_send_validation(struct smc_connection *conn,
134 struct smc_cdc_tx_pend *pend,
135 struct smc_wr_buf *wr_buf)
137 struct smc_host_cdc_msg *local = &conn->local_tx_ctrl;
138 struct smc_link *link = conn->lnk;
139 struct smc_cdc_msg *peer;
142 peer = (struct smc_cdc_msg *)wr_buf;
143 peer->common.type = local->common.type;
144 peer->len = local->len;
145 peer->seqno = htons(conn->tx_cdc_seq_fin); /* seqno last compl. tx */
146 peer->token = htonl(local->token);
147 peer->prod_flags.failover_validation = 1;
149 /* We need to set pend->conn here to make sure smc_cdc_tx_handler()
150 * can handle properly
152 smc_cdc_add_pending_send(conn, pend);
154 atomic_inc(&conn->cdc_pend_tx_wr);
155 smp_mb__after_atomic(); /* Make sure cdc_pend_tx_wr added before post */
157 rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend);
159 atomic_dec(&conn->cdc_pend_tx_wr);
164 static int smcr_cdc_get_slot_and_msg_send(struct smc_connection *conn)
166 struct smc_cdc_tx_pend *pend;
167 struct smc_wr_buf *wr_buf;
168 struct smc_link *link;
174 if (!smc_wr_tx_link_hold(link))
176 rc = smc_cdc_get_free_slot(conn, link, &wr_buf, NULL, &pend);
180 spin_lock_bh(&conn->send_lock);
181 if (link != conn->lnk) {
182 /* link of connection changed, try again one time*/
183 spin_unlock_bh(&conn->send_lock);
184 smc_wr_tx_put_slot(link,
185 (struct smc_wr_tx_pend_priv *)pend);
186 smc_wr_tx_link_put(link);
192 rc = smc_cdc_msg_send(conn, wr_buf, pend);
193 spin_unlock_bh(&conn->send_lock);
195 smc_wr_tx_link_put(link);
199 int smc_cdc_get_slot_and_msg_send(struct smc_connection *conn)
203 if (!conn->lgr || (conn->lgr->is_smcd && conn->lgr->peer_shutdown))
206 if (conn->lgr->is_smcd) {
207 spin_lock_bh(&conn->send_lock);
208 rc = smcd_cdc_msg_send(conn);
209 spin_unlock_bh(&conn->send_lock);
211 rc = smcr_cdc_get_slot_and_msg_send(conn);
217 void smc_cdc_wait_pend_tx_wr(struct smc_connection *conn)
219 wait_event(conn->cdc_pend_tx_wq, !atomic_read(&conn->cdc_pend_tx_wr));
222 /* Send a SMC-D CDC header.
223 * This increments the free space available in our send buffer.
224 * Also update the confirmed receive buffer with what was sent to the peer.
226 int smcd_cdc_msg_send(struct smc_connection *conn)
228 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
229 union smc_host_cursor curs;
230 struct smcd_cdc_msg cdc;
233 memset(&cdc, 0, sizeof(cdc));
234 cdc.common.type = SMC_CDC_MSG_TYPE;
235 curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.prod.acurs);
236 cdc.prod.wrap = curs.wrap;
237 cdc.prod.count = curs.count;
238 curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.cons.acurs);
239 cdc.cons.wrap = curs.wrap;
240 cdc.cons.count = curs.count;
241 cdc.cons.prod_flags = conn->local_tx_ctrl.prod_flags;
242 cdc.cons.conn_state_flags = conn->local_tx_ctrl.conn_state_flags;
243 rc = smcd_tx_ism_write(conn, &cdc, sizeof(cdc), 0, 1);
246 smc_curs_copy(&conn->rx_curs_confirmed, &curs, conn);
247 conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0;
248 /* Calculate transmitted data and increment free send buffer space */
249 diff = smc_curs_diff(conn->sndbuf_desc->len, &conn->tx_curs_fin,
250 &conn->tx_curs_sent);
251 /* increased by confirmed number of bytes */
252 smp_mb__before_atomic();
253 atomic_add(diff, &conn->sndbuf_space);
254 /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */
255 smp_mb__after_atomic();
256 smc_curs_copy(&conn->tx_curs_fin, &conn->tx_curs_sent, conn);
258 smc_tx_sndbuf_nonfull(smc);
262 /********************************* receive ***********************************/
264 static inline bool smc_cdc_before(u16 seq1, u16 seq2)
266 return (s16)(seq1 - seq2) < 0;
269 static void smc_cdc_handle_urg_data_arrival(struct smc_sock *smc,
272 struct smc_connection *conn = &smc->conn;
275 /* new data included urgent business */
276 smc_curs_copy(&conn->urg_curs, &conn->local_rx_ctrl.prod, conn);
277 conn->urg_state = SMC_URG_VALID;
278 if (!sock_flag(&smc->sk, SOCK_URGINLINE))
279 /* we'll skip the urgent byte, so don't account for it */
281 base = (char *)conn->rmb_desc->cpu_addr + conn->rx_off;
282 if (conn->urg_curs.count)
283 conn->urg_rx_byte = *(base + conn->urg_curs.count - 1);
285 conn->urg_rx_byte = *(base + conn->rmb_desc->len - 1);
286 sk_send_sigurg(&smc->sk);
289 static void smc_cdc_msg_validate(struct smc_sock *smc, struct smc_cdc_msg *cdc,
290 struct smc_link *link)
292 struct smc_connection *conn = &smc->conn;
293 u16 recv_seq = ntohs(cdc->seqno);
296 /* check that seqnum was seen before */
297 diff = conn->local_rx_ctrl.seqno - recv_seq;
298 if (diff < 0) { /* diff larger than 0x7fff */
299 /* drop connection */
300 conn->out_of_sync = 1; /* prevent any further receives */
301 spin_lock_bh(&conn->send_lock);
302 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
304 spin_unlock_bh(&conn->send_lock);
305 sock_hold(&smc->sk); /* sock_put in abort_work */
306 if (!queue_work(smc_close_wq, &conn->abort_work))
311 static void smc_cdc_msg_recv_action(struct smc_sock *smc,
312 struct smc_cdc_msg *cdc)
314 union smc_host_cursor cons_old, prod_old;
315 struct smc_connection *conn = &smc->conn;
316 int diff_cons, diff_prod;
318 smc_curs_copy(&prod_old, &conn->local_rx_ctrl.prod, conn);
319 smc_curs_copy(&cons_old, &conn->local_rx_ctrl.cons, conn);
320 smc_cdc_msg_to_host(&conn->local_rx_ctrl, cdc, conn);
322 diff_cons = smc_curs_diff(conn->peer_rmbe_size, &cons_old,
323 &conn->local_rx_ctrl.cons);
325 /* peer_rmbe_space is decreased during data transfer with RDMA
328 smp_mb__before_atomic();
329 atomic_add(diff_cons, &conn->peer_rmbe_space);
330 /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */
331 smp_mb__after_atomic();
334 diff_prod = smc_curs_diff(conn->rmb_desc->len, &prod_old,
335 &conn->local_rx_ctrl.prod);
337 if (conn->local_rx_ctrl.prod_flags.urg_data_present)
338 smc_cdc_handle_urg_data_arrival(smc, &diff_prod);
339 /* bytes_to_rcv is decreased in smc_recvmsg */
340 smp_mb__before_atomic();
341 atomic_add(diff_prod, &conn->bytes_to_rcv);
342 /* guarantee 0 <= bytes_to_rcv <= rmb_desc->len */
343 smp_mb__after_atomic();
344 smc->sk.sk_data_ready(&smc->sk);
346 if (conn->local_rx_ctrl.prod_flags.write_blocked)
347 smc->sk.sk_data_ready(&smc->sk);
348 if (conn->local_rx_ctrl.prod_flags.urg_data_pending)
349 conn->urg_state = SMC_URG_NOTYET;
352 /* trigger sndbuf consumer: RDMA write into peer RMBE and CDC */
353 if ((diff_cons && smc_tx_prepared_sends(conn)) ||
354 conn->local_rx_ctrl.prod_flags.cons_curs_upd_req ||
355 conn->local_rx_ctrl.prod_flags.urg_data_pending)
356 smc_tx_sndbuf_nonempty(conn);
358 if (diff_cons && conn->urg_tx_pend &&
359 atomic_read(&conn->peer_rmbe_space) == conn->peer_rmbe_size) {
360 /* urg data confirmed by peer, indicate we're ready for more */
361 conn->urg_tx_pend = false;
362 smc->sk.sk_write_space(&smc->sk);
365 if (conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) {
366 smc->sk.sk_err = ECONNRESET;
367 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
369 if (smc_cdc_rxed_any_close_or_senddone(conn)) {
370 smc->sk.sk_shutdown |= RCV_SHUTDOWN;
371 if (smc->clcsock && smc->clcsock->sk)
372 smc->clcsock->sk->sk_shutdown |= RCV_SHUTDOWN;
373 sock_set_flag(&smc->sk, SOCK_DONE);
374 sock_hold(&smc->sk); /* sock_put in close_work */
375 if (!queue_work(smc_close_wq, &conn->close_work))
380 /* called under tasklet context */
381 static void smc_cdc_msg_recv(struct smc_sock *smc, struct smc_cdc_msg *cdc)
384 bh_lock_sock(&smc->sk);
385 smc_cdc_msg_recv_action(smc, cdc);
386 bh_unlock_sock(&smc->sk);
387 sock_put(&smc->sk); /* no free sk in softirq-context */
390 /* Schedule a tasklet for this connection. Triggered from the ISM device IRQ
391 * handler to indicate update in the DMBE.
396 static void smcd_cdc_rx_tsklet(struct tasklet_struct *t)
398 struct smc_connection *conn = from_tasklet(conn, t, rx_tsklet);
399 struct smcd_cdc_msg *data_cdc;
400 struct smcd_cdc_msg cdc;
401 struct smc_sock *smc;
403 if (!conn || conn->killed)
406 data_cdc = (struct smcd_cdc_msg *)conn->rmb_desc->cpu_addr;
407 smcd_curs_copy(&cdc.prod, &data_cdc->prod, conn);
408 smcd_curs_copy(&cdc.cons, &data_cdc->cons, conn);
409 smc = container_of(conn, struct smc_sock, conn);
410 smc_cdc_msg_recv(smc, (struct smc_cdc_msg *)&cdc);
413 /* Initialize receive tasklet. Called from ISM device IRQ handler to start
416 void smcd_cdc_rx_init(struct smc_connection *conn)
418 tasklet_setup(&conn->rx_tsklet, smcd_cdc_rx_tsklet);
421 /***************************** init, exit, misc ******************************/
423 static void smc_cdc_rx_handler(struct ib_wc *wc, void *buf)
425 struct smc_link *link = (struct smc_link *)wc->qp->qp_context;
426 struct smc_cdc_msg *cdc = buf;
427 struct smc_connection *conn;
428 struct smc_link_group *lgr;
429 struct smc_sock *smc;
431 if (wc->byte_len < offsetof(struct smc_cdc_msg, reserved))
432 return; /* short message */
433 if (cdc->len != SMC_WR_TX_SIZE)
434 return; /* invalid message */
436 /* lookup connection */
437 lgr = smc_get_lgr(link);
438 read_lock_bh(&lgr->conns_lock);
439 conn = smc_lgr_find_conn(ntohl(cdc->token), lgr);
440 read_unlock_bh(&lgr->conns_lock);
441 if (!conn || conn->out_of_sync)
443 smc = container_of(conn, struct smc_sock, conn);
445 if (cdc->prod_flags.failover_validation) {
446 smc_cdc_msg_validate(smc, cdc, link);
449 if (smc_cdc_before(ntohs(cdc->seqno),
450 conn->local_rx_ctrl.seqno))
451 /* received seqno is old */
454 smc_cdc_msg_recv(smc, cdc);
457 static struct smc_wr_rx_handler smc_cdc_rx_handlers[] = {
459 .handler = smc_cdc_rx_handler,
460 .type = SMC_CDC_MSG_TYPE
467 int __init smc_cdc_init(void)
469 struct smc_wr_rx_handler *handler;
472 for (handler = smc_cdc_rx_handlers; handler->handler; handler++) {
473 INIT_HLIST_NODE(&handler->list);
474 rc = smc_wr_rx_register_handler(handler);