xref: /linux/net/smc/smc_cdc.c (revision 40e79150c1686263e6a031d7702aec63aff31332)
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 	}
51 	smc_tx_sndbuf_nonfull(smc);
52 	bh_unlock_sock(&smc->sk);
53 }
54 
55 int smc_cdc_get_free_slot(struct smc_connection *conn,
56 			  struct smc_wr_buf **wr_buf,
57 			  struct smc_rdma_wr **wr_rdma_buf,
58 			  struct smc_cdc_tx_pend **pend)
59 {
60 	struct smc_link *link = conn->lnk;
61 	int rc;
62 
63 	rc = smc_wr_tx_get_free_slot(link, smc_cdc_tx_handler, wr_buf,
64 				     wr_rdma_buf,
65 				     (struct smc_wr_tx_pend_priv **)pend);
66 	if (conn->killed)
67 		/* abnormal termination */
68 		rc = -EPIPE;
69 	return rc;
70 }
71 
72 static inline void smc_cdc_add_pending_send(struct smc_connection *conn,
73 					    struct smc_cdc_tx_pend *pend)
74 {
75 	BUILD_BUG_ON_MSG(
76 		sizeof(struct smc_cdc_msg) > SMC_WR_BUF_SIZE,
77 		"must increase SMC_WR_BUF_SIZE to at least sizeof(struct smc_cdc_msg)");
78 	BUILD_BUG_ON_MSG(
79 		offsetofend(struct smc_cdc_msg, reserved) > SMC_WR_TX_SIZE,
80 		"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()");
81 	BUILD_BUG_ON_MSG(
82 		sizeof(struct smc_cdc_tx_pend) > SMC_WR_TX_PEND_PRIV_SIZE,
83 		"must increase SMC_WR_TX_PEND_PRIV_SIZE to at least sizeof(struct smc_cdc_tx_pend)");
84 	pend->conn = conn;
85 	pend->cursor = conn->tx_curs_sent;
86 	pend->p_cursor = conn->local_tx_ctrl.prod;
87 	pend->ctrl_seq = conn->tx_cdc_seq;
88 }
89 
90 int smc_cdc_msg_send(struct smc_connection *conn,
91 		     struct smc_wr_buf *wr_buf,
92 		     struct smc_cdc_tx_pend *pend)
93 {
94 	struct smc_link *link = conn->lnk;
95 	union smc_host_cursor cfed;
96 	int rc;
97 
98 	smc_cdc_add_pending_send(conn, pend);
99 
100 	conn->tx_cdc_seq++;
101 	conn->local_tx_ctrl.seqno = conn->tx_cdc_seq;
102 	smc_host_msg_to_cdc((struct smc_cdc_msg *)wr_buf, conn, &cfed);
103 	rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend);
104 	if (!rc) {
105 		smc_curs_copy(&conn->rx_curs_confirmed, &cfed, conn);
106 		conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0;
107 	}
108 
109 	return rc;
110 }
111 
112 static int smcr_cdc_get_slot_and_msg_send(struct smc_connection *conn)
113 {
114 	struct smc_cdc_tx_pend *pend;
115 	struct smc_wr_buf *wr_buf;
116 	int rc;
117 
118 	rc = smc_cdc_get_free_slot(conn, &wr_buf, NULL, &pend);
119 	if (rc)
120 		return rc;
121 
122 	spin_lock_bh(&conn->send_lock);
123 	rc = smc_cdc_msg_send(conn, wr_buf, pend);
124 	spin_unlock_bh(&conn->send_lock);
125 	return rc;
126 }
127 
128 int smc_cdc_get_slot_and_msg_send(struct smc_connection *conn)
129 {
130 	int rc;
131 
132 	if (!conn->lgr || (conn->lgr->is_smcd && conn->lgr->peer_shutdown))
133 		return -EPIPE;
134 
135 	if (conn->lgr->is_smcd) {
136 		spin_lock_bh(&conn->send_lock);
137 		rc = smcd_cdc_msg_send(conn);
138 		spin_unlock_bh(&conn->send_lock);
139 	} else {
140 		rc = smcr_cdc_get_slot_and_msg_send(conn);
141 	}
142 
143 	return rc;
144 }
145 
146 static bool smc_cdc_tx_filter(struct smc_wr_tx_pend_priv *tx_pend,
147 			      unsigned long data)
148 {
149 	struct smc_connection *conn = (struct smc_connection *)data;
150 	struct smc_cdc_tx_pend *cdc_pend =
151 		(struct smc_cdc_tx_pend *)tx_pend;
152 
153 	return cdc_pend->conn == conn;
154 }
155 
156 static void smc_cdc_tx_dismisser(struct smc_wr_tx_pend_priv *tx_pend)
157 {
158 	struct smc_cdc_tx_pend *cdc_pend =
159 		(struct smc_cdc_tx_pend *)tx_pend;
160 
161 	cdc_pend->conn = NULL;
162 }
163 
164 void smc_cdc_tx_dismiss_slots(struct smc_connection *conn)
165 {
166 	struct smc_link *link = conn->lnk;
167 
168 	smc_wr_tx_dismiss_slots(link, SMC_CDC_MSG_TYPE,
169 				smc_cdc_tx_filter, smc_cdc_tx_dismisser,
170 				(unsigned long)conn);
171 }
172 
173 /* Send a SMC-D CDC header.
174  * This increments the free space available in our send buffer.
175  * Also update the confirmed receive buffer with what was sent to the peer.
176  */
177 int smcd_cdc_msg_send(struct smc_connection *conn)
178 {
179 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
180 	union smc_host_cursor curs;
181 	struct smcd_cdc_msg cdc;
182 	int rc, diff;
183 
184 	memset(&cdc, 0, sizeof(cdc));
185 	cdc.common.type = SMC_CDC_MSG_TYPE;
186 	curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.prod.acurs);
187 	cdc.prod.wrap = curs.wrap;
188 	cdc.prod.count = curs.count;
189 	curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.cons.acurs);
190 	cdc.cons.wrap = curs.wrap;
191 	cdc.cons.count = curs.count;
192 	cdc.cons.prod_flags = conn->local_tx_ctrl.prod_flags;
193 	cdc.cons.conn_state_flags = conn->local_tx_ctrl.conn_state_flags;
194 	rc = smcd_tx_ism_write(conn, &cdc, sizeof(cdc), 0, 1);
195 	if (rc)
196 		return rc;
197 	smc_curs_copy(&conn->rx_curs_confirmed, &curs, conn);
198 	conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0;
199 	/* Calculate transmitted data and increment free send buffer space */
200 	diff = smc_curs_diff(conn->sndbuf_desc->len, &conn->tx_curs_fin,
201 			     &conn->tx_curs_sent);
202 	/* increased by confirmed number of bytes */
203 	smp_mb__before_atomic();
204 	atomic_add(diff, &conn->sndbuf_space);
205 	/* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */
206 	smp_mb__after_atomic();
207 	smc_curs_copy(&conn->tx_curs_fin, &conn->tx_curs_sent, conn);
208 
209 	smc_tx_sndbuf_nonfull(smc);
210 	return rc;
211 }
212 
213 /********************************* receive ***********************************/
214 
215 static inline bool smc_cdc_before(u16 seq1, u16 seq2)
216 {
217 	return (s16)(seq1 - seq2) < 0;
218 }
219 
220 static void smc_cdc_handle_urg_data_arrival(struct smc_sock *smc,
221 					    int *diff_prod)
222 {
223 	struct smc_connection *conn = &smc->conn;
224 	char *base;
225 
226 	/* new data included urgent business */
227 	smc_curs_copy(&conn->urg_curs, &conn->local_rx_ctrl.prod, conn);
228 	conn->urg_state = SMC_URG_VALID;
229 	if (!sock_flag(&smc->sk, SOCK_URGINLINE))
230 		/* we'll skip the urgent byte, so don't account for it */
231 		(*diff_prod)--;
232 	base = (char *)conn->rmb_desc->cpu_addr + conn->rx_off;
233 	if (conn->urg_curs.count)
234 		conn->urg_rx_byte = *(base + conn->urg_curs.count - 1);
235 	else
236 		conn->urg_rx_byte = *(base + conn->rmb_desc->len - 1);
237 	sk_send_sigurg(&smc->sk);
238 }
239 
240 static void smc_cdc_msg_recv_action(struct smc_sock *smc,
241 				    struct smc_cdc_msg *cdc)
242 {
243 	union smc_host_cursor cons_old, prod_old;
244 	struct smc_connection *conn = &smc->conn;
245 	int diff_cons, diff_prod;
246 
247 	smc_curs_copy(&prod_old, &conn->local_rx_ctrl.prod, conn);
248 	smc_curs_copy(&cons_old, &conn->local_rx_ctrl.cons, conn);
249 	smc_cdc_msg_to_host(&conn->local_rx_ctrl, cdc, conn);
250 
251 	diff_cons = smc_curs_diff(conn->peer_rmbe_size, &cons_old,
252 				  &conn->local_rx_ctrl.cons);
253 	if (diff_cons) {
254 		/* peer_rmbe_space is decreased during data transfer with RDMA
255 		 * write
256 		 */
257 		smp_mb__before_atomic();
258 		atomic_add(diff_cons, &conn->peer_rmbe_space);
259 		/* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */
260 		smp_mb__after_atomic();
261 	}
262 
263 	diff_prod = smc_curs_diff(conn->rmb_desc->len, &prod_old,
264 				  &conn->local_rx_ctrl.prod);
265 	if (diff_prod) {
266 		if (conn->local_rx_ctrl.prod_flags.urg_data_present)
267 			smc_cdc_handle_urg_data_arrival(smc, &diff_prod);
268 		/* bytes_to_rcv is decreased in smc_recvmsg */
269 		smp_mb__before_atomic();
270 		atomic_add(diff_prod, &conn->bytes_to_rcv);
271 		/* guarantee 0 <= bytes_to_rcv <= rmb_desc->len */
272 		smp_mb__after_atomic();
273 		smc->sk.sk_data_ready(&smc->sk);
274 	} else {
275 		if (conn->local_rx_ctrl.prod_flags.write_blocked)
276 			smc->sk.sk_data_ready(&smc->sk);
277 		if (conn->local_rx_ctrl.prod_flags.urg_data_pending)
278 			conn->urg_state = SMC_URG_NOTYET;
279 	}
280 
281 	/* trigger sndbuf consumer: RDMA write into peer RMBE and CDC */
282 	if ((diff_cons && smc_tx_prepared_sends(conn)) ||
283 	    conn->local_rx_ctrl.prod_flags.cons_curs_upd_req ||
284 	    conn->local_rx_ctrl.prod_flags.urg_data_pending)
285 		smc_tx_sndbuf_nonempty(conn);
286 
287 	if (diff_cons && conn->urg_tx_pend &&
288 	    atomic_read(&conn->peer_rmbe_space) == conn->peer_rmbe_size) {
289 		/* urg data confirmed by peer, indicate we're ready for more */
290 		conn->urg_tx_pend = false;
291 		smc->sk.sk_write_space(&smc->sk);
292 	}
293 
294 	if (conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) {
295 		smc->sk.sk_err = ECONNRESET;
296 		conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
297 	}
298 	if (smc_cdc_rxed_any_close_or_senddone(conn)) {
299 		smc->sk.sk_shutdown |= RCV_SHUTDOWN;
300 		if (smc->clcsock && smc->clcsock->sk)
301 			smc->clcsock->sk->sk_shutdown |= RCV_SHUTDOWN;
302 		sock_set_flag(&smc->sk, SOCK_DONE);
303 		sock_hold(&smc->sk); /* sock_put in close_work */
304 		if (!schedule_work(&conn->close_work))
305 			sock_put(&smc->sk);
306 	}
307 }
308 
309 /* called under tasklet context */
310 static void smc_cdc_msg_recv(struct smc_sock *smc, struct smc_cdc_msg *cdc)
311 {
312 	sock_hold(&smc->sk);
313 	bh_lock_sock(&smc->sk);
314 	smc_cdc_msg_recv_action(smc, cdc);
315 	bh_unlock_sock(&smc->sk);
316 	sock_put(&smc->sk); /* no free sk in softirq-context */
317 }
318 
319 /* Schedule a tasklet for this connection. Triggered from the ISM device IRQ
320  * handler to indicate update in the DMBE.
321  *
322  * Context:
323  * - tasklet context
324  */
325 static void smcd_cdc_rx_tsklet(unsigned long data)
326 {
327 	struct smc_connection *conn = (struct smc_connection *)data;
328 	struct smcd_cdc_msg *data_cdc;
329 	struct smcd_cdc_msg cdc;
330 	struct smc_sock *smc;
331 
332 	if (!conn || conn->killed)
333 		return;
334 
335 	data_cdc = (struct smcd_cdc_msg *)conn->rmb_desc->cpu_addr;
336 	smcd_curs_copy(&cdc.prod, &data_cdc->prod, conn);
337 	smcd_curs_copy(&cdc.cons, &data_cdc->cons, conn);
338 	smc = container_of(conn, struct smc_sock, conn);
339 	smc_cdc_msg_recv(smc, (struct smc_cdc_msg *)&cdc);
340 }
341 
342 /* Initialize receive tasklet. Called from ISM device IRQ handler to start
343  * receiver side.
344  */
345 void smcd_cdc_rx_init(struct smc_connection *conn)
346 {
347 	tasklet_init(&conn->rx_tsklet, smcd_cdc_rx_tsklet, (unsigned long)conn);
348 }
349 
350 /***************************** init, exit, misc ******************************/
351 
352 static void smc_cdc_rx_handler(struct ib_wc *wc, void *buf)
353 {
354 	struct smc_link *link = (struct smc_link *)wc->qp->qp_context;
355 	struct smc_cdc_msg *cdc = buf;
356 	struct smc_connection *conn;
357 	struct smc_link_group *lgr;
358 	struct smc_sock *smc;
359 
360 	if (wc->byte_len < offsetof(struct smc_cdc_msg, reserved))
361 		return; /* short message */
362 	if (cdc->len != SMC_WR_TX_SIZE)
363 		return; /* invalid message */
364 
365 	/* lookup connection */
366 	lgr = smc_get_lgr(link);
367 	read_lock_bh(&lgr->conns_lock);
368 	conn = smc_lgr_find_conn(ntohl(cdc->token), lgr);
369 	read_unlock_bh(&lgr->conns_lock);
370 	if (!conn)
371 		return;
372 	smc = container_of(conn, struct smc_sock, conn);
373 
374 	if (!cdc->prod_flags.failover_validation) {
375 		if (smc_cdc_before(ntohs(cdc->seqno),
376 				   conn->local_rx_ctrl.seqno))
377 			/* received seqno is old */
378 			return;
379 	}
380 	smc_cdc_msg_recv(smc, cdc);
381 }
382 
383 static struct smc_wr_rx_handler smc_cdc_rx_handlers[] = {
384 	{
385 		.handler	= smc_cdc_rx_handler,
386 		.type		= SMC_CDC_MSG_TYPE
387 	},
388 	{
389 		.handler	= NULL,
390 	}
391 };
392 
393 int __init smc_cdc_init(void)
394 {
395 	struct smc_wr_rx_handler *handler;
396 	int rc = 0;
397 
398 	for (handler = smc_cdc_rx_handlers; handler->handler; handler++) {
399 		INIT_HLIST_NODE(&handler->list);
400 		rc = smc_wr_rx_register_handler(handler);
401 		if (rc)
402 			break;
403 	}
404 	return rc;
405 }
406