xref: /openbmc/linux/net/smc/smc_cdc.c (revision fe7498ef)
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 	if (!smc_wr_tx_link_hold(link))
154 		return -ENOLINK;
155 	rc = smc_cdc_get_free_slot(conn, link, &wr_buf, NULL, &pend);
156 	if (rc)
157 		goto put_out;
158 
159 	spin_lock_bh(&conn->send_lock);
160 	if (link != conn->lnk) {
161 		/* link of connection changed, try again one time*/
162 		spin_unlock_bh(&conn->send_lock);
163 		smc_wr_tx_put_slot(link,
164 				   (struct smc_wr_tx_pend_priv *)pend);
165 		smc_wr_tx_link_put(link);
166 		if (again)
167 			return -ENOLINK;
168 		again = true;
169 		goto again;
170 	}
171 	rc = smc_cdc_msg_send(conn, wr_buf, pend);
172 	spin_unlock_bh(&conn->send_lock);
173 put_out:
174 	smc_wr_tx_link_put(link);
175 	return rc;
176 }
177 
178 int smc_cdc_get_slot_and_msg_send(struct smc_connection *conn)
179 {
180 	int rc;
181 
182 	if (!conn->lgr || (conn->lgr->is_smcd && conn->lgr->peer_shutdown))
183 		return -EPIPE;
184 
185 	if (conn->lgr->is_smcd) {
186 		spin_lock_bh(&conn->send_lock);
187 		rc = smcd_cdc_msg_send(conn);
188 		spin_unlock_bh(&conn->send_lock);
189 	} else {
190 		rc = smcr_cdc_get_slot_and_msg_send(conn);
191 	}
192 
193 	return rc;
194 }
195 
196 static bool smc_cdc_tx_filter(struct smc_wr_tx_pend_priv *tx_pend,
197 			      unsigned long data)
198 {
199 	struct smc_connection *conn = (struct smc_connection *)data;
200 	struct smc_cdc_tx_pend *cdc_pend =
201 		(struct smc_cdc_tx_pend *)tx_pend;
202 
203 	return cdc_pend->conn == conn;
204 }
205 
206 static void smc_cdc_tx_dismisser(struct smc_wr_tx_pend_priv *tx_pend)
207 {
208 	struct smc_cdc_tx_pend *cdc_pend =
209 		(struct smc_cdc_tx_pend *)tx_pend;
210 
211 	cdc_pend->conn = NULL;
212 }
213 
214 void smc_cdc_tx_dismiss_slots(struct smc_connection *conn)
215 {
216 	struct smc_link *link = conn->lnk;
217 
218 	smc_wr_tx_dismiss_slots(link, SMC_CDC_MSG_TYPE,
219 				smc_cdc_tx_filter, smc_cdc_tx_dismisser,
220 				(unsigned long)conn);
221 }
222 
223 /* Send a SMC-D CDC header.
224  * This increments the free space available in our send buffer.
225  * Also update the confirmed receive buffer with what was sent to the peer.
226  */
227 int smcd_cdc_msg_send(struct smc_connection *conn)
228 {
229 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
230 	union smc_host_cursor curs;
231 	struct smcd_cdc_msg cdc;
232 	int rc, diff;
233 
234 	memset(&cdc, 0, sizeof(cdc));
235 	cdc.common.type = SMC_CDC_MSG_TYPE;
236 	curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.prod.acurs);
237 	cdc.prod.wrap = curs.wrap;
238 	cdc.prod.count = curs.count;
239 	curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.cons.acurs);
240 	cdc.cons.wrap = curs.wrap;
241 	cdc.cons.count = curs.count;
242 	cdc.cons.prod_flags = conn->local_tx_ctrl.prod_flags;
243 	cdc.cons.conn_state_flags = conn->local_tx_ctrl.conn_state_flags;
244 	rc = smcd_tx_ism_write(conn, &cdc, sizeof(cdc), 0, 1);
245 	if (rc)
246 		return rc;
247 	smc_curs_copy(&conn->rx_curs_confirmed, &curs, conn);
248 	conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0;
249 	/* Calculate transmitted data and increment free send buffer space */
250 	diff = smc_curs_diff(conn->sndbuf_desc->len, &conn->tx_curs_fin,
251 			     &conn->tx_curs_sent);
252 	/* increased by confirmed number of bytes */
253 	smp_mb__before_atomic();
254 	atomic_add(diff, &conn->sndbuf_space);
255 	/* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */
256 	smp_mb__after_atomic();
257 	smc_curs_copy(&conn->tx_curs_fin, &conn->tx_curs_sent, conn);
258 
259 	smc_tx_sndbuf_nonfull(smc);
260 	return rc;
261 }
262 
263 /********************************* receive ***********************************/
264 
265 static inline bool smc_cdc_before(u16 seq1, u16 seq2)
266 {
267 	return (s16)(seq1 - seq2) < 0;
268 }
269 
270 static void smc_cdc_handle_urg_data_arrival(struct smc_sock *smc,
271 					    int *diff_prod)
272 {
273 	struct smc_connection *conn = &smc->conn;
274 	char *base;
275 
276 	/* new data included urgent business */
277 	smc_curs_copy(&conn->urg_curs, &conn->local_rx_ctrl.prod, conn);
278 	conn->urg_state = SMC_URG_VALID;
279 	if (!sock_flag(&smc->sk, SOCK_URGINLINE))
280 		/* we'll skip the urgent byte, so don't account for it */
281 		(*diff_prod)--;
282 	base = (char *)conn->rmb_desc->cpu_addr + conn->rx_off;
283 	if (conn->urg_curs.count)
284 		conn->urg_rx_byte = *(base + conn->urg_curs.count - 1);
285 	else
286 		conn->urg_rx_byte = *(base + conn->rmb_desc->len - 1);
287 	sk_send_sigurg(&smc->sk);
288 }
289 
290 static void smc_cdc_msg_validate(struct smc_sock *smc, struct smc_cdc_msg *cdc,
291 				 struct smc_link *link)
292 {
293 	struct smc_connection *conn = &smc->conn;
294 	u16 recv_seq = ntohs(cdc->seqno);
295 	s16 diff;
296 
297 	/* check that seqnum was seen before */
298 	diff = conn->local_rx_ctrl.seqno - recv_seq;
299 	if (diff < 0) { /* diff larger than 0x7fff */
300 		/* drop connection */
301 		conn->out_of_sync = 1;	/* prevent any further receives */
302 		spin_lock_bh(&conn->send_lock);
303 		conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
304 		conn->lnk = link;
305 		spin_unlock_bh(&conn->send_lock);
306 		sock_hold(&smc->sk); /* sock_put in abort_work */
307 		if (!queue_work(smc_close_wq, &conn->abort_work))
308 			sock_put(&smc->sk);
309 	}
310 }
311 
312 static void smc_cdc_msg_recv_action(struct smc_sock *smc,
313 				    struct smc_cdc_msg *cdc)
314 {
315 	union smc_host_cursor cons_old, prod_old;
316 	struct smc_connection *conn = &smc->conn;
317 	int diff_cons, diff_prod;
318 
319 	smc_curs_copy(&prod_old, &conn->local_rx_ctrl.prod, conn);
320 	smc_curs_copy(&cons_old, &conn->local_rx_ctrl.cons, conn);
321 	smc_cdc_msg_to_host(&conn->local_rx_ctrl, cdc, conn);
322 
323 	diff_cons = smc_curs_diff(conn->peer_rmbe_size, &cons_old,
324 				  &conn->local_rx_ctrl.cons);
325 	if (diff_cons) {
326 		/* peer_rmbe_space is decreased during data transfer with RDMA
327 		 * write
328 		 */
329 		smp_mb__before_atomic();
330 		atomic_add(diff_cons, &conn->peer_rmbe_space);
331 		/* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */
332 		smp_mb__after_atomic();
333 	}
334 
335 	diff_prod = smc_curs_diff(conn->rmb_desc->len, &prod_old,
336 				  &conn->local_rx_ctrl.prod);
337 	if (diff_prod) {
338 		if (conn->local_rx_ctrl.prod_flags.urg_data_present)
339 			smc_cdc_handle_urg_data_arrival(smc, &diff_prod);
340 		/* bytes_to_rcv is decreased in smc_recvmsg */
341 		smp_mb__before_atomic();
342 		atomic_add(diff_prod, &conn->bytes_to_rcv);
343 		/* guarantee 0 <= bytes_to_rcv <= rmb_desc->len */
344 		smp_mb__after_atomic();
345 		smc->sk.sk_data_ready(&smc->sk);
346 	} else {
347 		if (conn->local_rx_ctrl.prod_flags.write_blocked)
348 			smc->sk.sk_data_ready(&smc->sk);
349 		if (conn->local_rx_ctrl.prod_flags.urg_data_pending)
350 			conn->urg_state = SMC_URG_NOTYET;
351 	}
352 
353 	/* trigger sndbuf consumer: RDMA write into peer RMBE and CDC */
354 	if ((diff_cons && smc_tx_prepared_sends(conn)) ||
355 	    conn->local_rx_ctrl.prod_flags.cons_curs_upd_req ||
356 	    conn->local_rx_ctrl.prod_flags.urg_data_pending)
357 		smc_tx_sndbuf_nonempty(conn);
358 
359 	if (diff_cons && conn->urg_tx_pend &&
360 	    atomic_read(&conn->peer_rmbe_space) == conn->peer_rmbe_size) {
361 		/* urg data confirmed by peer, indicate we're ready for more */
362 		conn->urg_tx_pend = false;
363 		smc->sk.sk_write_space(&smc->sk);
364 	}
365 
366 	if (conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) {
367 		smc->sk.sk_err = ECONNRESET;
368 		conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
369 	}
370 	if (smc_cdc_rxed_any_close_or_senddone(conn)) {
371 		smc->sk.sk_shutdown |= RCV_SHUTDOWN;
372 		if (smc->clcsock && smc->clcsock->sk)
373 			smc->clcsock->sk->sk_shutdown |= RCV_SHUTDOWN;
374 		sock_set_flag(&smc->sk, SOCK_DONE);
375 		sock_hold(&smc->sk); /* sock_put in close_work */
376 		if (!queue_work(smc_close_wq, &conn->close_work))
377 			sock_put(&smc->sk);
378 	}
379 }
380 
381 /* called under tasklet context */
382 static void smc_cdc_msg_recv(struct smc_sock *smc, struct smc_cdc_msg *cdc)
383 {
384 	sock_hold(&smc->sk);
385 	bh_lock_sock(&smc->sk);
386 	smc_cdc_msg_recv_action(smc, cdc);
387 	bh_unlock_sock(&smc->sk);
388 	sock_put(&smc->sk); /* no free sk in softirq-context */
389 }
390 
391 /* Schedule a tasklet for this connection. Triggered from the ISM device IRQ
392  * handler to indicate update in the DMBE.
393  *
394  * Context:
395  * - tasklet context
396  */
397 static void smcd_cdc_rx_tsklet(struct tasklet_struct *t)
398 {
399 	struct smc_connection *conn = from_tasklet(conn, t, rx_tsklet);
400 	struct smcd_cdc_msg *data_cdc;
401 	struct smcd_cdc_msg cdc;
402 	struct smc_sock *smc;
403 
404 	if (!conn || conn->killed)
405 		return;
406 
407 	data_cdc = (struct smcd_cdc_msg *)conn->rmb_desc->cpu_addr;
408 	smcd_curs_copy(&cdc.prod, &data_cdc->prod, conn);
409 	smcd_curs_copy(&cdc.cons, &data_cdc->cons, conn);
410 	smc = container_of(conn, struct smc_sock, conn);
411 	smc_cdc_msg_recv(smc, (struct smc_cdc_msg *)&cdc);
412 }
413 
414 /* Initialize receive tasklet. Called from ISM device IRQ handler to start
415  * receiver side.
416  */
417 void smcd_cdc_rx_init(struct smc_connection *conn)
418 {
419 	tasklet_setup(&conn->rx_tsklet, smcd_cdc_rx_tsklet);
420 }
421 
422 /***************************** init, exit, misc ******************************/
423 
424 static void smc_cdc_rx_handler(struct ib_wc *wc, void *buf)
425 {
426 	struct smc_link *link = (struct smc_link *)wc->qp->qp_context;
427 	struct smc_cdc_msg *cdc = buf;
428 	struct smc_connection *conn;
429 	struct smc_link_group *lgr;
430 	struct smc_sock *smc;
431 
432 	if (wc->byte_len < offsetof(struct smc_cdc_msg, reserved))
433 		return; /* short message */
434 	if (cdc->len != SMC_WR_TX_SIZE)
435 		return; /* invalid message */
436 
437 	/* lookup connection */
438 	lgr = smc_get_lgr(link);
439 	read_lock_bh(&lgr->conns_lock);
440 	conn = smc_lgr_find_conn(ntohl(cdc->token), lgr);
441 	read_unlock_bh(&lgr->conns_lock);
442 	if (!conn || conn->out_of_sync)
443 		return;
444 	smc = container_of(conn, struct smc_sock, conn);
445 
446 	if (cdc->prod_flags.failover_validation) {
447 		smc_cdc_msg_validate(smc, cdc, link);
448 		return;
449 	}
450 	if (smc_cdc_before(ntohs(cdc->seqno),
451 			   conn->local_rx_ctrl.seqno))
452 		/* received seqno is old */
453 		return;
454 
455 	smc_cdc_msg_recv(smc, cdc);
456 }
457 
458 static struct smc_wr_rx_handler smc_cdc_rx_handlers[] = {
459 	{
460 		.handler	= smc_cdc_rx_handler,
461 		.type		= SMC_CDC_MSG_TYPE
462 	},
463 	{
464 		.handler	= NULL,
465 	}
466 };
467 
468 int __init smc_cdc_init(void)
469 {
470 	struct smc_wr_rx_handler *handler;
471 	int rc = 0;
472 
473 	for (handler = smc_cdc_rx_handlers; handler->handler; handler++) {
474 		INIT_HLIST_NODE(&handler->list);
475 		rc = smc_wr_rx_register_handler(handler);
476 		if (rc)
477 			break;
478 	}
479 	return rc;
480 }
481