xref: /linux/net/smc/smc_cdc.c (revision e999db587312e5b798421d803495f41d1283d7d7)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Shared Memory Communications over RDMA (SMC-R) and RoCE
4  *
5  * Connection Data Control (CDC)
6  * handles flow control
7  *
8  * Copyright IBM Corp. 2016
9  *
10  * Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
11  */
12 
13 #include <linux/spinlock.h>
14 
15 #include "smc.h"
16 #include "smc_wr.h"
17 #include "smc_cdc.h"
18 #include "smc_tx.h"
19 #include "smc_rx.h"
20 #include "smc_close.h"
21 
22 /********************************** send *************************************/
23 
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)
28 {
29 	struct smc_cdc_tx_pend *cdcpend = (struct smc_cdc_tx_pend *)pnd_snd;
30 	struct smc_connection *conn = cdcpend->conn;
31 	struct smc_sock *smc;
32 	int diff;
33 
34 	if (!conn)
35 		/* already dismissed */
36 		return;
37 
38 	smc = container_of(conn, struct smc_sock, conn);
39 	bh_lock_sock(&smc->sk);
40 	if (!wc_status) {
41 		diff = smc_curs_diff(cdcpend->conn->sndbuf_desc->len,
42 				     &cdcpend->conn->tx_curs_fin,
43 				     &cdcpend->cursor);
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);
50 		smc_curs_copy(&conn->local_tx_ctrl_fin, &cdcpend->p_cursor,
51 			      conn);
52 		conn->tx_cdc_seq_fin = cdcpend->ctrl_seq;
53 	}
54 	smc_tx_sndbuf_nonfull(smc);
55 	bh_unlock_sock(&smc->sk);
56 }
57 
58 int smc_cdc_get_free_slot(struct smc_connection *conn,
59 			  struct smc_link *link,
60 			  struct smc_wr_buf **wr_buf,
61 			  struct smc_rdma_wr **wr_rdma_buf,
62 			  struct smc_cdc_tx_pend **pend)
63 {
64 	int rc;
65 
66 	rc = smc_wr_tx_get_free_slot(link, smc_cdc_tx_handler, wr_buf,
67 				     wr_rdma_buf,
68 				     (struct smc_wr_tx_pend_priv **)pend);
69 	if (conn->killed) {
70 		/* abnormal termination */
71 		if (!rc)
72 			smc_wr_tx_put_slot(link,
73 					   (struct smc_wr_tx_pend_priv *)pend);
74 		rc = -EPIPE;
75 	}
76 	return rc;
77 }
78 
79 static inline void smc_cdc_add_pending_send(struct smc_connection *conn,
80 					    struct smc_cdc_tx_pend *pend)
81 {
82 	BUILD_BUG_ON_MSG(
83 		sizeof(struct smc_cdc_msg) > SMC_WR_BUF_SIZE,
84 		"must increase SMC_WR_BUF_SIZE to at least sizeof(struct smc_cdc_msg)");
85 	BUILD_BUG_ON_MSG(
86 		offsetofend(struct smc_cdc_msg, reserved) > SMC_WR_TX_SIZE,
87 		"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()");
88 	BUILD_BUG_ON_MSG(
89 		sizeof(struct smc_cdc_tx_pend) > SMC_WR_TX_PEND_PRIV_SIZE,
90 		"must increase SMC_WR_TX_PEND_PRIV_SIZE to at least sizeof(struct smc_cdc_tx_pend)");
91 	pend->conn = conn;
92 	pend->cursor = conn->tx_curs_sent;
93 	pend->p_cursor = conn->local_tx_ctrl.prod;
94 	pend->ctrl_seq = conn->tx_cdc_seq;
95 }
96 
97 int smc_cdc_msg_send(struct smc_connection *conn,
98 		     struct smc_wr_buf *wr_buf,
99 		     struct smc_cdc_tx_pend *pend)
100 {
101 	struct smc_link *link = conn->lnk;
102 	union smc_host_cursor cfed;
103 	int rc;
104 
105 	smc_cdc_add_pending_send(conn, pend);
106 
107 	conn->tx_cdc_seq++;
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);
111 	if (!rc) {
112 		smc_curs_copy(&conn->rx_curs_confirmed, &cfed, conn);
113 		conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0;
114 	} else {
115 		conn->tx_cdc_seq--;
116 		conn->local_tx_ctrl.seqno = conn->tx_cdc_seq;
117 	}
118 
119 	return rc;
120 }
121 
122 /* send a validation msg indicating the move of a conn to an other QP link */
123 int smcr_cdc_msg_send_validation(struct smc_connection *conn,
124 				 struct smc_cdc_tx_pend *pend,
125 				 struct smc_wr_buf *wr_buf)
126 {
127 	struct smc_host_cdc_msg *local = &conn->local_tx_ctrl;
128 	struct smc_link *link = conn->lnk;
129 	struct smc_cdc_msg *peer;
130 	int rc;
131 
132 	peer = (struct smc_cdc_msg *)wr_buf;
133 	peer->common.type = local->common.type;
134 	peer->len = local->len;
135 	peer->seqno = htons(conn->tx_cdc_seq_fin); /* seqno last compl. tx */
136 	peer->token = htonl(local->token);
137 	peer->prod_flags.failover_validation = 1;
138 
139 	rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend);
140 	return rc;
141 }
142 
143 static int smcr_cdc_get_slot_and_msg_send(struct smc_connection *conn)
144 {
145 	struct smc_cdc_tx_pend *pend;
146 	struct smc_wr_buf *wr_buf;
147 	struct smc_link *link;
148 	bool again = false;
149 	int rc;
150 
151 again:
152 	link = conn->lnk;
153 	rc = smc_cdc_get_free_slot(conn, link, &wr_buf, NULL, &pend);
154 	if (rc)
155 		return rc;
156 
157 	spin_lock_bh(&conn->send_lock);
158 	if (link != conn->lnk) {
159 		/* link of connection changed, try again one time*/
160 		spin_unlock_bh(&conn->send_lock);
161 		smc_wr_tx_put_slot(link,
162 				   (struct smc_wr_tx_pend_priv *)pend);
163 		if (again)
164 			return -ENOLINK;
165 		again = true;
166 		goto again;
167 	}
168 	rc = smc_cdc_msg_send(conn, wr_buf, pend);
169 	spin_unlock_bh(&conn->send_lock);
170 	return rc;
171 }
172 
173 int smc_cdc_get_slot_and_msg_send(struct smc_connection *conn)
174 {
175 	int rc;
176 
177 	if (!conn->lgr || (conn->lgr->is_smcd && conn->lgr->peer_shutdown))
178 		return -EPIPE;
179 
180 	if (conn->lgr->is_smcd) {
181 		spin_lock_bh(&conn->send_lock);
182 		rc = smcd_cdc_msg_send(conn);
183 		spin_unlock_bh(&conn->send_lock);
184 	} else {
185 		rc = smcr_cdc_get_slot_and_msg_send(conn);
186 	}
187 
188 	return rc;
189 }
190 
191 static bool smc_cdc_tx_filter(struct smc_wr_tx_pend_priv *tx_pend,
192 			      unsigned long data)
193 {
194 	struct smc_connection *conn = (struct smc_connection *)data;
195 	struct smc_cdc_tx_pend *cdc_pend =
196 		(struct smc_cdc_tx_pend *)tx_pend;
197 
198 	return cdc_pend->conn == conn;
199 }
200 
201 static void smc_cdc_tx_dismisser(struct smc_wr_tx_pend_priv *tx_pend)
202 {
203 	struct smc_cdc_tx_pend *cdc_pend =
204 		(struct smc_cdc_tx_pend *)tx_pend;
205 
206 	cdc_pend->conn = NULL;
207 }
208 
209 void smc_cdc_tx_dismiss_slots(struct smc_connection *conn)
210 {
211 	struct smc_link *link = conn->lnk;
212 
213 	smc_wr_tx_dismiss_slots(link, SMC_CDC_MSG_TYPE,
214 				smc_cdc_tx_filter, smc_cdc_tx_dismisser,
215 				(unsigned long)conn);
216 }
217 
218 /* Send a SMC-D CDC header.
219  * This increments the free space available in our send buffer.
220  * Also update the confirmed receive buffer with what was sent to the peer.
221  */
222 int smcd_cdc_msg_send(struct smc_connection *conn)
223 {
224 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
225 	union smc_host_cursor curs;
226 	struct smcd_cdc_msg cdc;
227 	int rc, diff;
228 
229 	memset(&cdc, 0, sizeof(cdc));
230 	cdc.common.type = SMC_CDC_MSG_TYPE;
231 	curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.prod.acurs);
232 	cdc.prod.wrap = curs.wrap;
233 	cdc.prod.count = curs.count;
234 	curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.cons.acurs);
235 	cdc.cons.wrap = curs.wrap;
236 	cdc.cons.count = curs.count;
237 	cdc.cons.prod_flags = conn->local_tx_ctrl.prod_flags;
238 	cdc.cons.conn_state_flags = conn->local_tx_ctrl.conn_state_flags;
239 	rc = smcd_tx_ism_write(conn, &cdc, sizeof(cdc), 0, 1);
240 	if (rc)
241 		return rc;
242 	smc_curs_copy(&conn->rx_curs_confirmed, &curs, conn);
243 	conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0;
244 	/* Calculate transmitted data and increment free send buffer space */
245 	diff = smc_curs_diff(conn->sndbuf_desc->len, &conn->tx_curs_fin,
246 			     &conn->tx_curs_sent);
247 	/* increased by confirmed number of bytes */
248 	smp_mb__before_atomic();
249 	atomic_add(diff, &conn->sndbuf_space);
250 	/* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */
251 	smp_mb__after_atomic();
252 	smc_curs_copy(&conn->tx_curs_fin, &conn->tx_curs_sent, conn);
253 
254 	smc_tx_sndbuf_nonfull(smc);
255 	return rc;
256 }
257 
258 /********************************* receive ***********************************/
259 
260 static inline bool smc_cdc_before(u16 seq1, u16 seq2)
261 {
262 	return (s16)(seq1 - seq2) < 0;
263 }
264 
265 static void smc_cdc_handle_urg_data_arrival(struct smc_sock *smc,
266 					    int *diff_prod)
267 {
268 	struct smc_connection *conn = &smc->conn;
269 	char *base;
270 
271 	/* new data included urgent business */
272 	smc_curs_copy(&conn->urg_curs, &conn->local_rx_ctrl.prod, conn);
273 	conn->urg_state = SMC_URG_VALID;
274 	if (!sock_flag(&smc->sk, SOCK_URGINLINE))
275 		/* we'll skip the urgent byte, so don't account for it */
276 		(*diff_prod)--;
277 	base = (char *)conn->rmb_desc->cpu_addr + conn->rx_off;
278 	if (conn->urg_curs.count)
279 		conn->urg_rx_byte = *(base + conn->urg_curs.count - 1);
280 	else
281 		conn->urg_rx_byte = *(base + conn->rmb_desc->len - 1);
282 	sk_send_sigurg(&smc->sk);
283 }
284 
285 static void smc_cdc_msg_validate(struct smc_sock *smc, struct smc_cdc_msg *cdc,
286 				 struct smc_link *link)
287 {
288 	struct smc_connection *conn = &smc->conn;
289 	u16 recv_seq = ntohs(cdc->seqno);
290 	s16 diff;
291 
292 	/* check that seqnum was seen before */
293 	diff = conn->local_rx_ctrl.seqno - recv_seq;
294 	if (diff < 0) { /* diff larger than 0x7fff */
295 		/* drop connection */
296 		conn->out_of_sync = 1;	/* prevent any further receives */
297 		spin_lock_bh(&conn->send_lock);
298 		conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
299 		conn->lnk = link;
300 		spin_unlock_bh(&conn->send_lock);
301 		sock_hold(&smc->sk); /* sock_put in abort_work */
302 		if (!queue_work(smc_close_wq, &conn->abort_work))
303 			sock_put(&smc->sk);
304 	}
305 }
306 
307 static void smc_cdc_msg_recv_action(struct smc_sock *smc,
308 				    struct smc_cdc_msg *cdc)
309 {
310 	union smc_host_cursor cons_old, prod_old;
311 	struct smc_connection *conn = &smc->conn;
312 	int diff_cons, diff_prod;
313 
314 	smc_curs_copy(&prod_old, &conn->local_rx_ctrl.prod, conn);
315 	smc_curs_copy(&cons_old, &conn->local_rx_ctrl.cons, conn);
316 	smc_cdc_msg_to_host(&conn->local_rx_ctrl, cdc, conn);
317 
318 	diff_cons = smc_curs_diff(conn->peer_rmbe_size, &cons_old,
319 				  &conn->local_rx_ctrl.cons);
320 	if (diff_cons) {
321 		/* peer_rmbe_space is decreased during data transfer with RDMA
322 		 * write
323 		 */
324 		smp_mb__before_atomic();
325 		atomic_add(diff_cons, &conn->peer_rmbe_space);
326 		/* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */
327 		smp_mb__after_atomic();
328 	}
329 
330 	diff_prod = smc_curs_diff(conn->rmb_desc->len, &prod_old,
331 				  &conn->local_rx_ctrl.prod);
332 	if (diff_prod) {
333 		if (conn->local_rx_ctrl.prod_flags.urg_data_present)
334 			smc_cdc_handle_urg_data_arrival(smc, &diff_prod);
335 		/* bytes_to_rcv is decreased in smc_recvmsg */
336 		smp_mb__before_atomic();
337 		atomic_add(diff_prod, &conn->bytes_to_rcv);
338 		/* guarantee 0 <= bytes_to_rcv <= rmb_desc->len */
339 		smp_mb__after_atomic();
340 		smc->sk.sk_data_ready(&smc->sk);
341 	} else {
342 		if (conn->local_rx_ctrl.prod_flags.write_blocked)
343 			smc->sk.sk_data_ready(&smc->sk);
344 		if (conn->local_rx_ctrl.prod_flags.urg_data_pending)
345 			conn->urg_state = SMC_URG_NOTYET;
346 	}
347 
348 	/* trigger sndbuf consumer: RDMA write into peer RMBE and CDC */
349 	if ((diff_cons && smc_tx_prepared_sends(conn)) ||
350 	    conn->local_rx_ctrl.prod_flags.cons_curs_upd_req ||
351 	    conn->local_rx_ctrl.prod_flags.urg_data_pending)
352 		smc_tx_sndbuf_nonempty(conn);
353 
354 	if (diff_cons && conn->urg_tx_pend &&
355 	    atomic_read(&conn->peer_rmbe_space) == conn->peer_rmbe_size) {
356 		/* urg data confirmed by peer, indicate we're ready for more */
357 		conn->urg_tx_pend = false;
358 		smc->sk.sk_write_space(&smc->sk);
359 	}
360 
361 	if (conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) {
362 		smc->sk.sk_err = ECONNRESET;
363 		conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
364 	}
365 	if (smc_cdc_rxed_any_close_or_senddone(conn)) {
366 		smc->sk.sk_shutdown |= RCV_SHUTDOWN;
367 		if (smc->clcsock && smc->clcsock->sk)
368 			smc->clcsock->sk->sk_shutdown |= RCV_SHUTDOWN;
369 		sock_set_flag(&smc->sk, SOCK_DONE);
370 		sock_hold(&smc->sk); /* sock_put in close_work */
371 		if (!queue_work(smc_close_wq, &conn->close_work))
372 			sock_put(&smc->sk);
373 	}
374 }
375 
376 /* called under tasklet context */
377 static void smc_cdc_msg_recv(struct smc_sock *smc, struct smc_cdc_msg *cdc)
378 {
379 	sock_hold(&smc->sk);
380 	bh_lock_sock(&smc->sk);
381 	smc_cdc_msg_recv_action(smc, cdc);
382 	bh_unlock_sock(&smc->sk);
383 	sock_put(&smc->sk); /* no free sk in softirq-context */
384 }
385 
386 /* Schedule a tasklet for this connection. Triggered from the ISM device IRQ
387  * handler to indicate update in the DMBE.
388  *
389  * Context:
390  * - tasklet context
391  */
392 static void smcd_cdc_rx_tsklet(struct tasklet_struct *t)
393 {
394 	struct smc_connection *conn = from_tasklet(conn, t, rx_tsklet);
395 	struct smcd_cdc_msg *data_cdc;
396 	struct smcd_cdc_msg cdc;
397 	struct smc_sock *smc;
398 
399 	if (!conn || conn->killed)
400 		return;
401 
402 	data_cdc = (struct smcd_cdc_msg *)conn->rmb_desc->cpu_addr;
403 	smcd_curs_copy(&cdc.prod, &data_cdc->prod, conn);
404 	smcd_curs_copy(&cdc.cons, &data_cdc->cons, conn);
405 	smc = container_of(conn, struct smc_sock, conn);
406 	smc_cdc_msg_recv(smc, (struct smc_cdc_msg *)&cdc);
407 }
408 
409 /* Initialize receive tasklet. Called from ISM device IRQ handler to start
410  * receiver side.
411  */
412 void smcd_cdc_rx_init(struct smc_connection *conn)
413 {
414 	tasklet_setup(&conn->rx_tsklet, smcd_cdc_rx_tsklet);
415 }
416 
417 /***************************** init, exit, misc ******************************/
418 
419 static void smc_cdc_rx_handler(struct ib_wc *wc, void *buf)
420 {
421 	struct smc_link *link = (struct smc_link *)wc->qp->qp_context;
422 	struct smc_cdc_msg *cdc = buf;
423 	struct smc_connection *conn;
424 	struct smc_link_group *lgr;
425 	struct smc_sock *smc;
426 
427 	if (wc->byte_len < offsetof(struct smc_cdc_msg, reserved))
428 		return; /* short message */
429 	if (cdc->len != SMC_WR_TX_SIZE)
430 		return; /* invalid message */
431 
432 	/* lookup connection */
433 	lgr = smc_get_lgr(link);
434 	read_lock_bh(&lgr->conns_lock);
435 	conn = smc_lgr_find_conn(ntohl(cdc->token), lgr);
436 	read_unlock_bh(&lgr->conns_lock);
437 	if (!conn || conn->out_of_sync)
438 		return;
439 	smc = container_of(conn, struct smc_sock, conn);
440 
441 	if (cdc->prod_flags.failover_validation) {
442 		smc_cdc_msg_validate(smc, cdc, link);
443 		return;
444 	}
445 	if (smc_cdc_before(ntohs(cdc->seqno),
446 			   conn->local_rx_ctrl.seqno))
447 		/* received seqno is old */
448 		return;
449 
450 	smc_cdc_msg_recv(smc, cdc);
451 }
452 
453 static struct smc_wr_rx_handler smc_cdc_rx_handlers[] = {
454 	{
455 		.handler	= smc_cdc_rx_handler,
456 		.type		= SMC_CDC_MSG_TYPE
457 	},
458 	{
459 		.handler	= NULL,
460 	}
461 };
462 
463 int __init smc_cdc_init(void)
464 {
465 	struct smc_wr_rx_handler *handler;
466 	int rc = 0;
467 
468 	for (handler = smc_cdc_rx_handlers; handler->handler; handler++) {
469 		INIT_HLIST_NODE(&handler->list);
470 		rc = smc_wr_rx_register_handler(handler);
471 		if (rc)
472 			break;
473 	}
474 	return rc;
475 }
476