xref: /linux/net/smc/smc_cdc.c (revision 24168c5e6dfbdd5b414f048f47f75d64533296ca)
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 #include "smc_ism.h"
22 
23 /********************************** send *************************************/
24 
25 /* handler for send/transmission completion of a CDC msg */
26 static void smc_cdc_tx_handler(struct smc_wr_tx_pend_priv *pnd_snd,
27 			       struct smc_link *link,
28 			       enum ib_wc_status wc_status)
29 {
30 	struct smc_cdc_tx_pend *cdcpend = (struct smc_cdc_tx_pend *)pnd_snd;
31 	struct smc_connection *conn = cdcpend->conn;
32 	struct smc_buf_desc *sndbuf_desc;
33 	struct smc_sock *smc;
34 	int diff;
35 
36 	sndbuf_desc = conn->sndbuf_desc;
37 	smc = container_of(conn, struct smc_sock, conn);
38 	bh_lock_sock(&smc->sk);
39 	if (!wc_status && sndbuf_desc) {
40 		diff = smc_curs_diff(sndbuf_desc->len,
41 				     &cdcpend->conn->tx_curs_fin,
42 				     &cdcpend->cursor);
43 		/* sndbuf_space is decreased in smc_sendmsg */
44 		smp_mb__before_atomic();
45 		atomic_add(diff, &cdcpend->conn->sndbuf_space);
46 		/* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */
47 		smp_mb__after_atomic();
48 		smc_curs_copy(&conn->tx_curs_fin, &cdcpend->cursor, conn);
49 		smc_curs_copy(&conn->local_tx_ctrl_fin, &cdcpend->p_cursor,
50 			      conn);
51 		conn->tx_cdc_seq_fin = cdcpend->ctrl_seq;
52 	}
53 
54 	if (atomic_dec_and_test(&conn->cdc_pend_tx_wr)) {
55 		/* If user owns the sock_lock, mark the connection need sending.
56 		 * User context will later try to send when it release sock_lock
57 		 * in smc_release_cb()
58 		 */
59 		if (sock_owned_by_user(&smc->sk))
60 			conn->tx_in_release_sock = true;
61 		else
62 			smc_tx_pending(conn);
63 
64 		if (unlikely(wq_has_sleeper(&conn->cdc_pend_tx_wq)))
65 			wake_up(&conn->cdc_pend_tx_wq);
66 	}
67 	WARN_ON(atomic_read(&conn->cdc_pend_tx_wr) < 0);
68 
69 	smc_tx_sndbuf_nonfull(smc);
70 	bh_unlock_sock(&smc->sk);
71 }
72 
73 int smc_cdc_get_free_slot(struct smc_connection *conn,
74 			  struct smc_link *link,
75 			  struct smc_wr_buf **wr_buf,
76 			  struct smc_rdma_wr **wr_rdma_buf,
77 			  struct smc_cdc_tx_pend **pend)
78 {
79 	int rc;
80 
81 	rc = smc_wr_tx_get_free_slot(link, smc_cdc_tx_handler, wr_buf,
82 				     wr_rdma_buf,
83 				     (struct smc_wr_tx_pend_priv **)pend);
84 	if (conn->killed) {
85 		/* abnormal termination */
86 		if (!rc)
87 			smc_wr_tx_put_slot(link,
88 					   (struct smc_wr_tx_pend_priv *)(*pend));
89 		rc = -EPIPE;
90 	}
91 	return rc;
92 }
93 
94 static inline void smc_cdc_add_pending_send(struct smc_connection *conn,
95 					    struct smc_cdc_tx_pend *pend)
96 {
97 	BUILD_BUG_ON_MSG(
98 		sizeof(struct smc_cdc_msg) > SMC_WR_BUF_SIZE,
99 		"must increase SMC_WR_BUF_SIZE to at least sizeof(struct smc_cdc_msg)");
100 	BUILD_BUG_ON_MSG(
101 		offsetofend(struct smc_cdc_msg, reserved) > SMC_WR_TX_SIZE,
102 		"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()");
103 	BUILD_BUG_ON_MSG(
104 		sizeof(struct smc_cdc_tx_pend) > SMC_WR_TX_PEND_PRIV_SIZE,
105 		"must increase SMC_WR_TX_PEND_PRIV_SIZE to at least sizeof(struct smc_cdc_tx_pend)");
106 	pend->conn = conn;
107 	pend->cursor = conn->tx_curs_sent;
108 	pend->p_cursor = conn->local_tx_ctrl.prod;
109 	pend->ctrl_seq = conn->tx_cdc_seq;
110 }
111 
112 int smc_cdc_msg_send(struct smc_connection *conn,
113 		     struct smc_wr_buf *wr_buf,
114 		     struct smc_cdc_tx_pend *pend)
115 {
116 	struct smc_link *link = conn->lnk;
117 	union smc_host_cursor cfed;
118 	int rc;
119 
120 	smc_cdc_add_pending_send(conn, pend);
121 
122 	conn->tx_cdc_seq++;
123 	conn->local_tx_ctrl.seqno = conn->tx_cdc_seq;
124 	smc_host_msg_to_cdc((struct smc_cdc_msg *)wr_buf, conn, &cfed);
125 
126 	atomic_inc(&conn->cdc_pend_tx_wr);
127 	smp_mb__after_atomic(); /* Make sure cdc_pend_tx_wr added before post */
128 
129 	rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend);
130 	if (!rc) {
131 		smc_curs_copy(&conn->rx_curs_confirmed, &cfed, conn);
132 		conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0;
133 	} else {
134 		conn->tx_cdc_seq--;
135 		conn->local_tx_ctrl.seqno = conn->tx_cdc_seq;
136 		atomic_dec(&conn->cdc_pend_tx_wr);
137 	}
138 
139 	return rc;
140 }
141 
142 /* send a validation msg indicating the move of a conn to an other QP link */
143 int smcr_cdc_msg_send_validation(struct smc_connection *conn,
144 				 struct smc_cdc_tx_pend *pend,
145 				 struct smc_wr_buf *wr_buf)
146 {
147 	struct smc_host_cdc_msg *local = &conn->local_tx_ctrl;
148 	struct smc_link *link = conn->lnk;
149 	struct smc_cdc_msg *peer;
150 	int rc;
151 
152 	peer = (struct smc_cdc_msg *)wr_buf;
153 	peer->common.type = local->common.type;
154 	peer->len = local->len;
155 	peer->seqno = htons(conn->tx_cdc_seq_fin); /* seqno last compl. tx */
156 	peer->token = htonl(local->token);
157 	peer->prod_flags.failover_validation = 1;
158 
159 	/* We need to set pend->conn here to make sure smc_cdc_tx_handler()
160 	 * can handle properly
161 	 */
162 	smc_cdc_add_pending_send(conn, pend);
163 
164 	atomic_inc(&conn->cdc_pend_tx_wr);
165 	smp_mb__after_atomic(); /* Make sure cdc_pend_tx_wr added before post */
166 
167 	rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend);
168 	if (unlikely(rc))
169 		atomic_dec(&conn->cdc_pend_tx_wr);
170 
171 	return rc;
172 }
173 
174 static int smcr_cdc_get_slot_and_msg_send(struct smc_connection *conn)
175 {
176 	struct smc_cdc_tx_pend *pend;
177 	struct smc_wr_buf *wr_buf;
178 	struct smc_link *link;
179 	bool again = false;
180 	int rc;
181 
182 again:
183 	link = conn->lnk;
184 	if (!smc_wr_tx_link_hold(link))
185 		return -ENOLINK;
186 	rc = smc_cdc_get_free_slot(conn, link, &wr_buf, NULL, &pend);
187 	if (rc)
188 		goto put_out;
189 
190 	spin_lock_bh(&conn->send_lock);
191 	if (link != conn->lnk) {
192 		/* link of connection changed, try again one time*/
193 		spin_unlock_bh(&conn->send_lock);
194 		smc_wr_tx_put_slot(link,
195 				   (struct smc_wr_tx_pend_priv *)pend);
196 		smc_wr_tx_link_put(link);
197 		if (again)
198 			return -ENOLINK;
199 		again = true;
200 		goto again;
201 	}
202 	rc = smc_cdc_msg_send(conn, wr_buf, pend);
203 	spin_unlock_bh(&conn->send_lock);
204 put_out:
205 	smc_wr_tx_link_put(link);
206 	return rc;
207 }
208 
209 int smc_cdc_get_slot_and_msg_send(struct smc_connection *conn)
210 {
211 	int rc;
212 
213 	if (!smc_conn_lgr_valid(conn) ||
214 	    (conn->lgr->is_smcd && conn->lgr->peer_shutdown))
215 		return -EPIPE;
216 
217 	if (conn->lgr->is_smcd) {
218 		spin_lock_bh(&conn->send_lock);
219 		rc = smcd_cdc_msg_send(conn);
220 		spin_unlock_bh(&conn->send_lock);
221 	} else {
222 		rc = smcr_cdc_get_slot_and_msg_send(conn);
223 	}
224 
225 	return rc;
226 }
227 
228 void smc_cdc_wait_pend_tx_wr(struct smc_connection *conn)
229 {
230 	wait_event(conn->cdc_pend_tx_wq, !atomic_read(&conn->cdc_pend_tx_wr));
231 }
232 
233 /* Send a SMC-D CDC header.
234  * This increments the free space available in our send buffer.
235  * Also update the confirmed receive buffer with what was sent to the peer.
236  */
237 int smcd_cdc_msg_send(struct smc_connection *conn)
238 {
239 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
240 	union smc_host_cursor curs;
241 	struct smcd_cdc_msg cdc;
242 	int rc, diff;
243 
244 	memset(&cdc, 0, sizeof(cdc));
245 	cdc.common.type = SMC_CDC_MSG_TYPE;
246 	curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.prod.acurs);
247 	cdc.prod.wrap = curs.wrap;
248 	cdc.prod.count = curs.count;
249 	curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.cons.acurs);
250 	cdc.cons.wrap = curs.wrap;
251 	cdc.cons.count = curs.count;
252 	cdc.cons.prod_flags = conn->local_tx_ctrl.prod_flags;
253 	cdc.cons.conn_state_flags = conn->local_tx_ctrl.conn_state_flags;
254 	rc = smcd_tx_ism_write(conn, &cdc, sizeof(cdc), 0, 1);
255 	if (rc)
256 		return rc;
257 	smc_curs_copy(&conn->rx_curs_confirmed, &curs, conn);
258 	conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0;
259 
260 	if (smc_ism_support_dmb_nocopy(conn->lgr->smcd))
261 		/* if local sndbuf shares the same memory region with
262 		 * peer DMB, then don't update the tx_curs_fin
263 		 * and sndbuf_space until peer has consumed the data.
264 		 */
265 		return 0;
266 
267 	/* Calculate transmitted data and increment free send buffer space */
268 	diff = smc_curs_diff(conn->sndbuf_desc->len, &conn->tx_curs_fin,
269 			     &conn->tx_curs_sent);
270 	/* increased by confirmed number of bytes */
271 	smp_mb__before_atomic();
272 	atomic_add(diff, &conn->sndbuf_space);
273 	/* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */
274 	smp_mb__after_atomic();
275 	smc_curs_copy(&conn->tx_curs_fin, &conn->tx_curs_sent, conn);
276 
277 	smc_tx_sndbuf_nonfull(smc);
278 	return 0;
279 }
280 
281 /********************************* receive ***********************************/
282 
283 static inline bool smc_cdc_before(u16 seq1, u16 seq2)
284 {
285 	return (s16)(seq1 - seq2) < 0;
286 }
287 
288 static void smc_cdc_handle_urg_data_arrival(struct smc_sock *smc,
289 					    int *diff_prod)
290 {
291 	struct smc_connection *conn = &smc->conn;
292 	char *base;
293 
294 	/* new data included urgent business */
295 	smc_curs_copy(&conn->urg_curs, &conn->local_rx_ctrl.prod, conn);
296 	conn->urg_state = SMC_URG_VALID;
297 	if (!sock_flag(&smc->sk, SOCK_URGINLINE))
298 		/* we'll skip the urgent byte, so don't account for it */
299 		(*diff_prod)--;
300 	base = (char *)conn->rmb_desc->cpu_addr + conn->rx_off;
301 	if (conn->urg_curs.count)
302 		conn->urg_rx_byte = *(base + conn->urg_curs.count - 1);
303 	else
304 		conn->urg_rx_byte = *(base + conn->rmb_desc->len - 1);
305 	sk_send_sigurg(&smc->sk);
306 }
307 
308 static void smc_cdc_msg_validate(struct smc_sock *smc, struct smc_cdc_msg *cdc,
309 				 struct smc_link *link)
310 {
311 	struct smc_connection *conn = &smc->conn;
312 	u16 recv_seq = ntohs(cdc->seqno);
313 	s16 diff;
314 
315 	/* check that seqnum was seen before */
316 	diff = conn->local_rx_ctrl.seqno - recv_seq;
317 	if (diff < 0) { /* diff larger than 0x7fff */
318 		/* drop connection */
319 		conn->out_of_sync = 1;	/* prevent any further receives */
320 		spin_lock_bh(&conn->send_lock);
321 		conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
322 		conn->lnk = link;
323 		spin_unlock_bh(&conn->send_lock);
324 		sock_hold(&smc->sk); /* sock_put in abort_work */
325 		if (!queue_work(smc_close_wq, &conn->abort_work))
326 			sock_put(&smc->sk);
327 	}
328 }
329 
330 static void smc_cdc_msg_recv_action(struct smc_sock *smc,
331 				    struct smc_cdc_msg *cdc)
332 {
333 	union smc_host_cursor cons_old, prod_old;
334 	struct smc_connection *conn = &smc->conn;
335 	int diff_cons, diff_prod, diff_tx;
336 
337 	smc_curs_copy(&prod_old, &conn->local_rx_ctrl.prod, conn);
338 	smc_curs_copy(&cons_old, &conn->local_rx_ctrl.cons, conn);
339 	smc_cdc_msg_to_host(&conn->local_rx_ctrl, cdc, conn);
340 
341 	diff_cons = smc_curs_diff(conn->peer_rmbe_size, &cons_old,
342 				  &conn->local_rx_ctrl.cons);
343 	if (diff_cons) {
344 		/* peer_rmbe_space is decreased during data transfer with RDMA
345 		 * write
346 		 */
347 		smp_mb__before_atomic();
348 		atomic_add(diff_cons, &conn->peer_rmbe_space);
349 		/* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */
350 		smp_mb__after_atomic();
351 
352 		/* if local sndbuf shares the same memory region with
353 		 * peer RMB, then update tx_curs_fin and sndbuf_space
354 		 * here since peer has already consumed the data.
355 		 */
356 		if (conn->lgr->is_smcd &&
357 		    smc_ism_support_dmb_nocopy(conn->lgr->smcd)) {
358 			/* Calculate consumed data and
359 			 * increment free send buffer space.
360 			 */
361 			diff_tx = smc_curs_diff(conn->sndbuf_desc->len,
362 						&conn->tx_curs_fin,
363 						&conn->local_rx_ctrl.cons);
364 			/* increase local sndbuf space and fin_curs */
365 			smp_mb__before_atomic();
366 			atomic_add(diff_tx, &conn->sndbuf_space);
367 			/* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */
368 			smp_mb__after_atomic();
369 			smc_curs_copy(&conn->tx_curs_fin,
370 				      &conn->local_rx_ctrl.cons, conn);
371 
372 			smc_tx_sndbuf_nonfull(smc);
373 		}
374 	}
375 
376 	diff_prod = smc_curs_diff(conn->rmb_desc->len, &prod_old,
377 				  &conn->local_rx_ctrl.prod);
378 	if (diff_prod) {
379 		if (conn->local_rx_ctrl.prod_flags.urg_data_present)
380 			smc_cdc_handle_urg_data_arrival(smc, &diff_prod);
381 		/* bytes_to_rcv is decreased in smc_recvmsg */
382 		smp_mb__before_atomic();
383 		atomic_add(diff_prod, &conn->bytes_to_rcv);
384 		/* guarantee 0 <= bytes_to_rcv <= rmb_desc->len */
385 		smp_mb__after_atomic();
386 		smc->sk.sk_data_ready(&smc->sk);
387 	} else {
388 		if (conn->local_rx_ctrl.prod_flags.write_blocked)
389 			smc->sk.sk_data_ready(&smc->sk);
390 		if (conn->local_rx_ctrl.prod_flags.urg_data_pending)
391 			conn->urg_state = SMC_URG_NOTYET;
392 	}
393 
394 	/* trigger sndbuf consumer: RDMA write into peer RMBE and CDC */
395 	if ((diff_cons && smc_tx_prepared_sends(conn)) ||
396 	    conn->local_rx_ctrl.prod_flags.cons_curs_upd_req ||
397 	    conn->local_rx_ctrl.prod_flags.urg_data_pending) {
398 		if (!sock_owned_by_user(&smc->sk))
399 			smc_tx_pending(conn);
400 		else
401 			conn->tx_in_release_sock = true;
402 	}
403 
404 	if (diff_cons && conn->urg_tx_pend &&
405 	    atomic_read(&conn->peer_rmbe_space) == conn->peer_rmbe_size) {
406 		/* urg data confirmed by peer, indicate we're ready for more */
407 		conn->urg_tx_pend = false;
408 		smc->sk.sk_write_space(&smc->sk);
409 	}
410 
411 	if (conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) {
412 		smc->sk.sk_err = ECONNRESET;
413 		conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
414 	}
415 	if (smc_cdc_rxed_any_close_or_senddone(conn)) {
416 		smc->sk.sk_shutdown |= RCV_SHUTDOWN;
417 		if (smc->clcsock && smc->clcsock->sk)
418 			smc->clcsock->sk->sk_shutdown |= RCV_SHUTDOWN;
419 		smc_sock_set_flag(&smc->sk, SOCK_DONE);
420 		sock_hold(&smc->sk); /* sock_put in close_work */
421 		if (!queue_work(smc_close_wq, &conn->close_work))
422 			sock_put(&smc->sk);
423 	}
424 }
425 
426 /* called under tasklet context */
427 static void smc_cdc_msg_recv(struct smc_sock *smc, struct smc_cdc_msg *cdc)
428 {
429 	sock_hold(&smc->sk);
430 	bh_lock_sock(&smc->sk);
431 	smc_cdc_msg_recv_action(smc, cdc);
432 	bh_unlock_sock(&smc->sk);
433 	sock_put(&smc->sk); /* no free sk in softirq-context */
434 }
435 
436 /* Schedule a tasklet for this connection. Triggered from the ISM device IRQ
437  * handler to indicate update in the DMBE.
438  *
439  * Context:
440  * - tasklet context
441  */
442 static void smcd_cdc_rx_tsklet(struct tasklet_struct *t)
443 {
444 	struct smc_connection *conn = from_tasklet(conn, t, rx_tsklet);
445 	struct smcd_cdc_msg *data_cdc;
446 	struct smcd_cdc_msg cdc;
447 	struct smc_sock *smc;
448 
449 	if (!conn || conn->killed)
450 		return;
451 
452 	data_cdc = (struct smcd_cdc_msg *)conn->rmb_desc->cpu_addr;
453 	smcd_curs_copy(&cdc.prod, &data_cdc->prod, conn);
454 	smcd_curs_copy(&cdc.cons, &data_cdc->cons, conn);
455 	smc = container_of(conn, struct smc_sock, conn);
456 	smc_cdc_msg_recv(smc, (struct smc_cdc_msg *)&cdc);
457 }
458 
459 /* Initialize receive tasklet. Called from ISM device IRQ handler to start
460  * receiver side.
461  */
462 void smcd_cdc_rx_init(struct smc_connection *conn)
463 {
464 	tasklet_setup(&conn->rx_tsklet, smcd_cdc_rx_tsklet);
465 }
466 
467 /***************************** init, exit, misc ******************************/
468 
469 static void smc_cdc_rx_handler(struct ib_wc *wc, void *buf)
470 {
471 	struct smc_link *link = (struct smc_link *)wc->qp->qp_context;
472 	struct smc_cdc_msg *cdc = buf;
473 	struct smc_connection *conn;
474 	struct smc_link_group *lgr;
475 	struct smc_sock *smc;
476 
477 	if (wc->byte_len < offsetof(struct smc_cdc_msg, reserved))
478 		return; /* short message */
479 	if (cdc->len != SMC_WR_TX_SIZE)
480 		return; /* invalid message */
481 
482 	/* lookup connection */
483 	lgr = smc_get_lgr(link);
484 	read_lock_bh(&lgr->conns_lock);
485 	conn = smc_lgr_find_conn(ntohl(cdc->token), lgr);
486 	read_unlock_bh(&lgr->conns_lock);
487 	if (!conn || conn->out_of_sync)
488 		return;
489 	smc = container_of(conn, struct smc_sock, conn);
490 
491 	if (cdc->prod_flags.failover_validation) {
492 		smc_cdc_msg_validate(smc, cdc, link);
493 		return;
494 	}
495 	if (smc_cdc_before(ntohs(cdc->seqno),
496 			   conn->local_rx_ctrl.seqno))
497 		/* received seqno is old */
498 		return;
499 
500 	smc_cdc_msg_recv(smc, cdc);
501 }
502 
503 static struct smc_wr_rx_handler smc_cdc_rx_handlers[] = {
504 	{
505 		.handler	= smc_cdc_rx_handler,
506 		.type		= SMC_CDC_MSG_TYPE
507 	},
508 	{
509 		.handler	= NULL,
510 	}
511 };
512 
513 int __init smc_cdc_init(void)
514 {
515 	struct smc_wr_rx_handler *handler;
516 	int rc = 0;
517 
518 	for (handler = smc_cdc_rx_handlers; handler->handler; handler++) {
519 		INIT_HLIST_NODE(&handler->list);
520 		rc = smc_wr_rx_register_handler(handler);
521 		if (rc)
522 			break;
523 	}
524 	return rc;
525 }
526