xref: /openbmc/linux/net/smc/smc_cdc.c (revision e733f3e9)
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_buf_desc *sndbuf_desc;
32 	struct smc_sock *smc;
33 	int diff;
34 
35 	sndbuf_desc = conn->sndbuf_desc;
36 	smc = container_of(conn, struct smc_sock, conn);
37 	bh_lock_sock(&smc->sk);
38 	if (!wc_status && sndbuf_desc) {
39 		diff = smc_curs_diff(sndbuf_desc->len,
40 				     &cdcpend->conn->tx_curs_fin,
41 				     &cdcpend->cursor);
42 		/* sndbuf_space is decreased in smc_sendmsg */
43 		smp_mb__before_atomic();
44 		atomic_add(diff, &cdcpend->conn->sndbuf_space);
45 		/* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */
46 		smp_mb__after_atomic();
47 		smc_curs_copy(&conn->tx_curs_fin, &cdcpend->cursor, conn);
48 		smc_curs_copy(&conn->local_tx_ctrl_fin, &cdcpend->p_cursor,
49 			      conn);
50 		conn->tx_cdc_seq_fin = cdcpend->ctrl_seq;
51 	}
52 
53 	if (atomic_dec_and_test(&conn->cdc_pend_tx_wr)) {
54 		/* If user owns the sock_lock, mark the connection need sending.
55 		 * User context will later try to send when it release sock_lock
56 		 * in smc_release_cb()
57 		 */
58 		if (sock_owned_by_user(&smc->sk))
59 			conn->tx_in_release_sock = true;
60 		else
61 			smc_tx_pending(conn);
62 
63 		if (unlikely(wq_has_sleeper(&conn->cdc_pend_tx_wq)))
64 			wake_up(&conn->cdc_pend_tx_wq);
65 	}
66 	WARN_ON(atomic_read(&conn->cdc_pend_tx_wr) < 0);
67 
68 	smc_tx_sndbuf_nonfull(smc);
69 	bh_unlock_sock(&smc->sk);
70 }
71 
72 int smc_cdc_get_free_slot(struct smc_connection *conn,
73 			  struct smc_link *link,
74 			  struct smc_wr_buf **wr_buf,
75 			  struct smc_rdma_wr **wr_rdma_buf,
76 			  struct smc_cdc_tx_pend **pend)
77 {
78 	int rc;
79 
80 	rc = smc_wr_tx_get_free_slot(link, smc_cdc_tx_handler, wr_buf,
81 				     wr_rdma_buf,
82 				     (struct smc_wr_tx_pend_priv **)pend);
83 	if (conn->killed) {
84 		/* abnormal termination */
85 		if (!rc)
86 			smc_wr_tx_put_slot(link,
87 					   (struct smc_wr_tx_pend_priv *)(*pend));
88 		rc = -EPIPE;
89 	}
90 	return rc;
91 }
92 
93 static inline void smc_cdc_add_pending_send(struct smc_connection *conn,
94 					    struct smc_cdc_tx_pend *pend)
95 {
96 	BUILD_BUG_ON_MSG(
97 		sizeof(struct smc_cdc_msg) > SMC_WR_BUF_SIZE,
98 		"must increase SMC_WR_BUF_SIZE to at least sizeof(struct smc_cdc_msg)");
99 	BUILD_BUG_ON_MSG(
100 		offsetofend(struct smc_cdc_msg, reserved) > SMC_WR_TX_SIZE,
101 		"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()");
102 	BUILD_BUG_ON_MSG(
103 		sizeof(struct smc_cdc_tx_pend) > SMC_WR_TX_PEND_PRIV_SIZE,
104 		"must increase SMC_WR_TX_PEND_PRIV_SIZE to at least sizeof(struct smc_cdc_tx_pend)");
105 	pend->conn = conn;
106 	pend->cursor = conn->tx_curs_sent;
107 	pend->p_cursor = conn->local_tx_ctrl.prod;
108 	pend->ctrl_seq = conn->tx_cdc_seq;
109 }
110 
111 int smc_cdc_msg_send(struct smc_connection *conn,
112 		     struct smc_wr_buf *wr_buf,
113 		     struct smc_cdc_tx_pend *pend)
114 {
115 	struct smc_link *link = conn->lnk;
116 	union smc_host_cursor cfed;
117 	int rc;
118 
119 	smc_cdc_add_pending_send(conn, pend);
120 
121 	conn->tx_cdc_seq++;
122 	conn->local_tx_ctrl.seqno = conn->tx_cdc_seq;
123 	smc_host_msg_to_cdc((struct smc_cdc_msg *)wr_buf, conn, &cfed);
124 
125 	atomic_inc(&conn->cdc_pend_tx_wr);
126 	smp_mb__after_atomic(); /* Make sure cdc_pend_tx_wr added before post */
127 
128 	rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend);
129 	if (!rc) {
130 		smc_curs_copy(&conn->rx_curs_confirmed, &cfed, conn);
131 		conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0;
132 	} else {
133 		conn->tx_cdc_seq--;
134 		conn->local_tx_ctrl.seqno = conn->tx_cdc_seq;
135 		atomic_dec(&conn->cdc_pend_tx_wr);
136 	}
137 
138 	return rc;
139 }
140 
141 /* send a validation msg indicating the move of a conn to an other QP link */
142 int smcr_cdc_msg_send_validation(struct smc_connection *conn,
143 				 struct smc_cdc_tx_pend *pend,
144 				 struct smc_wr_buf *wr_buf)
145 {
146 	struct smc_host_cdc_msg *local = &conn->local_tx_ctrl;
147 	struct smc_link *link = conn->lnk;
148 	struct smc_cdc_msg *peer;
149 	int rc;
150 
151 	peer = (struct smc_cdc_msg *)wr_buf;
152 	peer->common.type = local->common.type;
153 	peer->len = local->len;
154 	peer->seqno = htons(conn->tx_cdc_seq_fin); /* seqno last compl. tx */
155 	peer->token = htonl(local->token);
156 	peer->prod_flags.failover_validation = 1;
157 
158 	/* We need to set pend->conn here to make sure smc_cdc_tx_handler()
159 	 * can handle properly
160 	 */
161 	smc_cdc_add_pending_send(conn, pend);
162 
163 	atomic_inc(&conn->cdc_pend_tx_wr);
164 	smp_mb__after_atomic(); /* Make sure cdc_pend_tx_wr added before post */
165 
166 	rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend);
167 	if (unlikely(rc))
168 		atomic_dec(&conn->cdc_pend_tx_wr);
169 
170 	return rc;
171 }
172 
173 static int smcr_cdc_get_slot_and_msg_send(struct smc_connection *conn)
174 {
175 	struct smc_cdc_tx_pend *pend;
176 	struct smc_wr_buf *wr_buf;
177 	struct smc_link *link;
178 	bool again = false;
179 	int rc;
180 
181 again:
182 	link = conn->lnk;
183 	if (!smc_wr_tx_link_hold(link))
184 		return -ENOLINK;
185 	rc = smc_cdc_get_free_slot(conn, link, &wr_buf, NULL, &pend);
186 	if (rc)
187 		goto put_out;
188 
189 	spin_lock_bh(&conn->send_lock);
190 	if (link != conn->lnk) {
191 		/* link of connection changed, try again one time*/
192 		spin_unlock_bh(&conn->send_lock);
193 		smc_wr_tx_put_slot(link,
194 				   (struct smc_wr_tx_pend_priv *)pend);
195 		smc_wr_tx_link_put(link);
196 		if (again)
197 			return -ENOLINK;
198 		again = true;
199 		goto again;
200 	}
201 	rc = smc_cdc_msg_send(conn, wr_buf, pend);
202 	spin_unlock_bh(&conn->send_lock);
203 put_out:
204 	smc_wr_tx_link_put(link);
205 	return rc;
206 }
207 
208 int smc_cdc_get_slot_and_msg_send(struct smc_connection *conn)
209 {
210 	int rc;
211 
212 	if (!smc_conn_lgr_valid(conn) ||
213 	    (conn->lgr->is_smcd && conn->lgr->peer_shutdown))
214 		return -EPIPE;
215 
216 	if (conn->lgr->is_smcd) {
217 		spin_lock_bh(&conn->send_lock);
218 		rc = smcd_cdc_msg_send(conn);
219 		spin_unlock_bh(&conn->send_lock);
220 	} else {
221 		rc = smcr_cdc_get_slot_and_msg_send(conn);
222 	}
223 
224 	return rc;
225 }
226 
227 void smc_cdc_wait_pend_tx_wr(struct smc_connection *conn)
228 {
229 	wait_event(conn->cdc_pend_tx_wq, !atomic_read(&conn->cdc_pend_tx_wr));
230 }
231 
232 /* Send a SMC-D CDC header.
233  * This increments the free space available in our send buffer.
234  * Also update the confirmed receive buffer with what was sent to the peer.
235  */
236 int smcd_cdc_msg_send(struct smc_connection *conn)
237 {
238 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
239 	union smc_host_cursor curs;
240 	struct smcd_cdc_msg cdc;
241 	int rc, diff;
242 
243 	memset(&cdc, 0, sizeof(cdc));
244 	cdc.common.type = SMC_CDC_MSG_TYPE;
245 	curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.prod.acurs);
246 	cdc.prod.wrap = curs.wrap;
247 	cdc.prod.count = curs.count;
248 	curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.cons.acurs);
249 	cdc.cons.wrap = curs.wrap;
250 	cdc.cons.count = curs.count;
251 	cdc.cons.prod_flags = conn->local_tx_ctrl.prod_flags;
252 	cdc.cons.conn_state_flags = conn->local_tx_ctrl.conn_state_flags;
253 	rc = smcd_tx_ism_write(conn, &cdc, sizeof(cdc), 0, 1);
254 	if (rc)
255 		return rc;
256 	smc_curs_copy(&conn->rx_curs_confirmed, &curs, conn);
257 	conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0;
258 	/* Calculate transmitted data and increment free send buffer space */
259 	diff = smc_curs_diff(conn->sndbuf_desc->len, &conn->tx_curs_fin,
260 			     &conn->tx_curs_sent);
261 	/* increased by confirmed number of bytes */
262 	smp_mb__before_atomic();
263 	atomic_add(diff, &conn->sndbuf_space);
264 	/* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */
265 	smp_mb__after_atomic();
266 	smc_curs_copy(&conn->tx_curs_fin, &conn->tx_curs_sent, conn);
267 
268 	smc_tx_sndbuf_nonfull(smc);
269 	return rc;
270 }
271 
272 /********************************* receive ***********************************/
273 
274 static inline bool smc_cdc_before(u16 seq1, u16 seq2)
275 {
276 	return (s16)(seq1 - seq2) < 0;
277 }
278 
279 static void smc_cdc_handle_urg_data_arrival(struct smc_sock *smc,
280 					    int *diff_prod)
281 {
282 	struct smc_connection *conn = &smc->conn;
283 	char *base;
284 
285 	/* new data included urgent business */
286 	smc_curs_copy(&conn->urg_curs, &conn->local_rx_ctrl.prod, conn);
287 	conn->urg_state = SMC_URG_VALID;
288 	if (!sock_flag(&smc->sk, SOCK_URGINLINE))
289 		/* we'll skip the urgent byte, so don't account for it */
290 		(*diff_prod)--;
291 	base = (char *)conn->rmb_desc->cpu_addr + conn->rx_off;
292 	if (conn->urg_curs.count)
293 		conn->urg_rx_byte = *(base + conn->urg_curs.count - 1);
294 	else
295 		conn->urg_rx_byte = *(base + conn->rmb_desc->len - 1);
296 	sk_send_sigurg(&smc->sk);
297 }
298 
299 static void smc_cdc_msg_validate(struct smc_sock *smc, struct smc_cdc_msg *cdc,
300 				 struct smc_link *link)
301 {
302 	struct smc_connection *conn = &smc->conn;
303 	u16 recv_seq = ntohs(cdc->seqno);
304 	s16 diff;
305 
306 	/* check that seqnum was seen before */
307 	diff = conn->local_rx_ctrl.seqno - recv_seq;
308 	if (diff < 0) { /* diff larger than 0x7fff */
309 		/* drop connection */
310 		conn->out_of_sync = 1;	/* prevent any further receives */
311 		spin_lock_bh(&conn->send_lock);
312 		conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
313 		conn->lnk = link;
314 		spin_unlock_bh(&conn->send_lock);
315 		sock_hold(&smc->sk); /* sock_put in abort_work */
316 		if (!queue_work(smc_close_wq, &conn->abort_work))
317 			sock_put(&smc->sk);
318 	}
319 }
320 
321 static void smc_cdc_msg_recv_action(struct smc_sock *smc,
322 				    struct smc_cdc_msg *cdc)
323 {
324 	union smc_host_cursor cons_old, prod_old;
325 	struct smc_connection *conn = &smc->conn;
326 	int diff_cons, diff_prod;
327 
328 	smc_curs_copy(&prod_old, &conn->local_rx_ctrl.prod, conn);
329 	smc_curs_copy(&cons_old, &conn->local_rx_ctrl.cons, conn);
330 	smc_cdc_msg_to_host(&conn->local_rx_ctrl, cdc, conn);
331 
332 	diff_cons = smc_curs_diff(conn->peer_rmbe_size, &cons_old,
333 				  &conn->local_rx_ctrl.cons);
334 	if (diff_cons) {
335 		/* peer_rmbe_space is decreased during data transfer with RDMA
336 		 * write
337 		 */
338 		smp_mb__before_atomic();
339 		atomic_add(diff_cons, &conn->peer_rmbe_space);
340 		/* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */
341 		smp_mb__after_atomic();
342 	}
343 
344 	diff_prod = smc_curs_diff(conn->rmb_desc->len, &prod_old,
345 				  &conn->local_rx_ctrl.prod);
346 	if (diff_prod) {
347 		if (conn->local_rx_ctrl.prod_flags.urg_data_present)
348 			smc_cdc_handle_urg_data_arrival(smc, &diff_prod);
349 		/* bytes_to_rcv is decreased in smc_recvmsg */
350 		smp_mb__before_atomic();
351 		atomic_add(diff_prod, &conn->bytes_to_rcv);
352 		/* guarantee 0 <= bytes_to_rcv <= rmb_desc->len */
353 		smp_mb__after_atomic();
354 		smc->sk.sk_data_ready(&smc->sk);
355 	} else {
356 		if (conn->local_rx_ctrl.prod_flags.write_blocked)
357 			smc->sk.sk_data_ready(&smc->sk);
358 		if (conn->local_rx_ctrl.prod_flags.urg_data_pending)
359 			conn->urg_state = SMC_URG_NOTYET;
360 	}
361 
362 	/* trigger sndbuf consumer: RDMA write into peer RMBE and CDC */
363 	if ((diff_cons && smc_tx_prepared_sends(conn)) ||
364 	    conn->local_rx_ctrl.prod_flags.cons_curs_upd_req ||
365 	    conn->local_rx_ctrl.prod_flags.urg_data_pending) {
366 		if (!sock_owned_by_user(&smc->sk))
367 			smc_tx_pending(conn);
368 		else
369 			conn->tx_in_release_sock = true;
370 	}
371 
372 	if (diff_cons && conn->urg_tx_pend &&
373 	    atomic_read(&conn->peer_rmbe_space) == conn->peer_rmbe_size) {
374 		/* urg data confirmed by peer, indicate we're ready for more */
375 		conn->urg_tx_pend = false;
376 		smc->sk.sk_write_space(&smc->sk);
377 	}
378 
379 	if (conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) {
380 		smc->sk.sk_err = ECONNRESET;
381 		conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
382 	}
383 	if (smc_cdc_rxed_any_close_or_senddone(conn)) {
384 		smc->sk.sk_shutdown |= RCV_SHUTDOWN;
385 		if (smc->clcsock && smc->clcsock->sk)
386 			smc->clcsock->sk->sk_shutdown |= RCV_SHUTDOWN;
387 		smc_sock_set_flag(&smc->sk, SOCK_DONE);
388 		sock_hold(&smc->sk); /* sock_put in close_work */
389 		if (!queue_work(smc_close_wq, &conn->close_work))
390 			sock_put(&smc->sk);
391 	}
392 }
393 
394 /* called under tasklet context */
395 static void smc_cdc_msg_recv(struct smc_sock *smc, struct smc_cdc_msg *cdc)
396 {
397 	sock_hold(&smc->sk);
398 	bh_lock_sock(&smc->sk);
399 	smc_cdc_msg_recv_action(smc, cdc);
400 	bh_unlock_sock(&smc->sk);
401 	sock_put(&smc->sk); /* no free sk in softirq-context */
402 }
403 
404 /* Schedule a tasklet for this connection. Triggered from the ISM device IRQ
405  * handler to indicate update in the DMBE.
406  *
407  * Context:
408  * - tasklet context
409  */
410 static void smcd_cdc_rx_tsklet(struct tasklet_struct *t)
411 {
412 	struct smc_connection *conn = from_tasklet(conn, t, rx_tsklet);
413 	struct smcd_cdc_msg *data_cdc;
414 	struct smcd_cdc_msg cdc;
415 	struct smc_sock *smc;
416 
417 	if (!conn || conn->killed)
418 		return;
419 
420 	data_cdc = (struct smcd_cdc_msg *)conn->rmb_desc->cpu_addr;
421 	smcd_curs_copy(&cdc.prod, &data_cdc->prod, conn);
422 	smcd_curs_copy(&cdc.cons, &data_cdc->cons, conn);
423 	smc = container_of(conn, struct smc_sock, conn);
424 	smc_cdc_msg_recv(smc, (struct smc_cdc_msg *)&cdc);
425 }
426 
427 /* Initialize receive tasklet. Called from ISM device IRQ handler to start
428  * receiver side.
429  */
430 void smcd_cdc_rx_init(struct smc_connection *conn)
431 {
432 	tasklet_setup(&conn->rx_tsklet, smcd_cdc_rx_tsklet);
433 }
434 
435 /***************************** init, exit, misc ******************************/
436 
437 static void smc_cdc_rx_handler(struct ib_wc *wc, void *buf)
438 {
439 	struct smc_link *link = (struct smc_link *)wc->qp->qp_context;
440 	struct smc_cdc_msg *cdc = buf;
441 	struct smc_connection *conn;
442 	struct smc_link_group *lgr;
443 	struct smc_sock *smc;
444 
445 	if (wc->byte_len < offsetof(struct smc_cdc_msg, reserved))
446 		return; /* short message */
447 	if (cdc->len != SMC_WR_TX_SIZE)
448 		return; /* invalid message */
449 
450 	/* lookup connection */
451 	lgr = smc_get_lgr(link);
452 	read_lock_bh(&lgr->conns_lock);
453 	conn = smc_lgr_find_conn(ntohl(cdc->token), lgr);
454 	read_unlock_bh(&lgr->conns_lock);
455 	if (!conn || conn->out_of_sync)
456 		return;
457 	smc = container_of(conn, struct smc_sock, conn);
458 
459 	if (cdc->prod_flags.failover_validation) {
460 		smc_cdc_msg_validate(smc, cdc, link);
461 		return;
462 	}
463 	if (smc_cdc_before(ntohs(cdc->seqno),
464 			   conn->local_rx_ctrl.seqno))
465 		/* received seqno is old */
466 		return;
467 
468 	smc_cdc_msg_recv(smc, cdc);
469 }
470 
471 static struct smc_wr_rx_handler smc_cdc_rx_handlers[] = {
472 	{
473 		.handler	= smc_cdc_rx_handler,
474 		.type		= SMC_CDC_MSG_TYPE
475 	},
476 	{
477 		.handler	= NULL,
478 	}
479 };
480 
481 int __init smc_cdc_init(void)
482 {
483 	struct smc_wr_rx_handler *handler;
484 	int rc = 0;
485 
486 	for (handler = smc_cdc_rx_handlers; handler->handler; handler++) {
487 		INIT_HLIST_NODE(&handler->list);
488 		rc = smc_wr_rx_register_handler(handler);
489 		if (rc)
490 			break;
491 	}
492 	return rc;
493 }
494