xref: /openbmc/linux/net/smc/smc_cdc.c (revision f8e17c17)
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->lgr->lnk[SMC_SINGLE_LINK];
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 	union smc_host_cursor cfed;
95 	struct smc_link *link;
96 	int rc;
97 
98 	link = &conn->lgr->lnk[SMC_SINGLE_LINK];
99 
100 	smc_cdc_add_pending_send(conn, pend);
101 
102 	conn->tx_cdc_seq++;
103 	conn->local_tx_ctrl.seqno = conn->tx_cdc_seq;
104 	smc_host_msg_to_cdc((struct smc_cdc_msg *)wr_buf, conn, &cfed);
105 	rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend);
106 	if (!rc) {
107 		smc_curs_copy(&conn->rx_curs_confirmed, &cfed, conn);
108 		conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0;
109 	}
110 
111 	return rc;
112 }
113 
114 static int smcr_cdc_get_slot_and_msg_send(struct smc_connection *conn)
115 {
116 	struct smc_cdc_tx_pend *pend;
117 	struct smc_wr_buf *wr_buf;
118 	int rc;
119 
120 	rc = smc_cdc_get_free_slot(conn, &wr_buf, NULL, &pend);
121 	if (rc)
122 		return rc;
123 
124 	spin_lock_bh(&conn->send_lock);
125 	rc = smc_cdc_msg_send(conn, wr_buf, pend);
126 	spin_unlock_bh(&conn->send_lock);
127 	return rc;
128 }
129 
130 int smc_cdc_get_slot_and_msg_send(struct smc_connection *conn)
131 {
132 	int rc;
133 
134 	if (!conn->lgr || (conn->lgr->is_smcd && conn->lgr->peer_shutdown))
135 		return -EPIPE;
136 
137 	if (conn->lgr->is_smcd) {
138 		spin_lock_bh(&conn->send_lock);
139 		rc = smcd_cdc_msg_send(conn);
140 		spin_unlock_bh(&conn->send_lock);
141 	} else {
142 		rc = smcr_cdc_get_slot_and_msg_send(conn);
143 	}
144 
145 	return rc;
146 }
147 
148 static bool smc_cdc_tx_filter(struct smc_wr_tx_pend_priv *tx_pend,
149 			      unsigned long data)
150 {
151 	struct smc_connection *conn = (struct smc_connection *)data;
152 	struct smc_cdc_tx_pend *cdc_pend =
153 		(struct smc_cdc_tx_pend *)tx_pend;
154 
155 	return cdc_pend->conn == conn;
156 }
157 
158 static void smc_cdc_tx_dismisser(struct smc_wr_tx_pend_priv *tx_pend)
159 {
160 	struct smc_cdc_tx_pend *cdc_pend =
161 		(struct smc_cdc_tx_pend *)tx_pend;
162 
163 	cdc_pend->conn = NULL;
164 }
165 
166 void smc_cdc_tx_dismiss_slots(struct smc_connection *conn)
167 {
168 	struct smc_link *link = &conn->lgr->lnk[SMC_SINGLE_LINK];
169 
170 	smc_wr_tx_dismiss_slots(link, SMC_CDC_MSG_TYPE,
171 				smc_cdc_tx_filter, smc_cdc_tx_dismisser,
172 				(unsigned long)conn);
173 }
174 
175 /* Send a SMC-D CDC header.
176  * This increments the free space available in our send buffer.
177  * Also update the confirmed receive buffer with what was sent to the peer.
178  */
179 int smcd_cdc_msg_send(struct smc_connection *conn)
180 {
181 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
182 	union smc_host_cursor curs;
183 	struct smcd_cdc_msg cdc;
184 	int rc, diff;
185 
186 	memset(&cdc, 0, sizeof(cdc));
187 	cdc.common.type = SMC_CDC_MSG_TYPE;
188 	curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.prod.acurs);
189 	cdc.prod.wrap = curs.wrap;
190 	cdc.prod.count = curs.count;
191 	curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.cons.acurs);
192 	cdc.cons.wrap = curs.wrap;
193 	cdc.cons.count = curs.count;
194 	cdc.cons.prod_flags = conn->local_tx_ctrl.prod_flags;
195 	cdc.cons.conn_state_flags = conn->local_tx_ctrl.conn_state_flags;
196 	rc = smcd_tx_ism_write(conn, &cdc, sizeof(cdc), 0, 1);
197 	if (rc)
198 		return rc;
199 	smc_curs_copy(&conn->rx_curs_confirmed, &curs, conn);
200 	conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0;
201 	/* Calculate transmitted data and increment free send buffer space */
202 	diff = smc_curs_diff(conn->sndbuf_desc->len, &conn->tx_curs_fin,
203 			     &conn->tx_curs_sent);
204 	/* increased by confirmed number of bytes */
205 	smp_mb__before_atomic();
206 	atomic_add(diff, &conn->sndbuf_space);
207 	/* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */
208 	smp_mb__after_atomic();
209 	smc_curs_copy(&conn->tx_curs_fin, &conn->tx_curs_sent, conn);
210 
211 	smc_tx_sndbuf_nonfull(smc);
212 	return rc;
213 }
214 
215 /********************************* receive ***********************************/
216 
217 static inline bool smc_cdc_before(u16 seq1, u16 seq2)
218 {
219 	return (s16)(seq1 - seq2) < 0;
220 }
221 
222 static void smc_cdc_handle_urg_data_arrival(struct smc_sock *smc,
223 					    int *diff_prod)
224 {
225 	struct smc_connection *conn = &smc->conn;
226 	char *base;
227 
228 	/* new data included urgent business */
229 	smc_curs_copy(&conn->urg_curs, &conn->local_rx_ctrl.prod, conn);
230 	conn->urg_state = SMC_URG_VALID;
231 	if (!sock_flag(&smc->sk, SOCK_URGINLINE))
232 		/* we'll skip the urgent byte, so don't account for it */
233 		(*diff_prod)--;
234 	base = (char *)conn->rmb_desc->cpu_addr + conn->rx_off;
235 	if (conn->urg_curs.count)
236 		conn->urg_rx_byte = *(base + conn->urg_curs.count - 1);
237 	else
238 		conn->urg_rx_byte = *(base + conn->rmb_desc->len - 1);
239 	sk_send_sigurg(&smc->sk);
240 }
241 
242 static void smc_cdc_msg_recv_action(struct smc_sock *smc,
243 				    struct smc_cdc_msg *cdc)
244 {
245 	union smc_host_cursor cons_old, prod_old;
246 	struct smc_connection *conn = &smc->conn;
247 	int diff_cons, diff_prod;
248 
249 	smc_curs_copy(&prod_old, &conn->local_rx_ctrl.prod, conn);
250 	smc_curs_copy(&cons_old, &conn->local_rx_ctrl.cons, conn);
251 	smc_cdc_msg_to_host(&conn->local_rx_ctrl, cdc, conn);
252 
253 	diff_cons = smc_curs_diff(conn->peer_rmbe_size, &cons_old,
254 				  &conn->local_rx_ctrl.cons);
255 	if (diff_cons) {
256 		/* peer_rmbe_space is decreased during data transfer with RDMA
257 		 * write
258 		 */
259 		smp_mb__before_atomic();
260 		atomic_add(diff_cons, &conn->peer_rmbe_space);
261 		/* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */
262 		smp_mb__after_atomic();
263 	}
264 
265 	diff_prod = smc_curs_diff(conn->rmb_desc->len, &prod_old,
266 				  &conn->local_rx_ctrl.prod);
267 	if (diff_prod) {
268 		if (conn->local_rx_ctrl.prod_flags.urg_data_present)
269 			smc_cdc_handle_urg_data_arrival(smc, &diff_prod);
270 		/* bytes_to_rcv is decreased in smc_recvmsg */
271 		smp_mb__before_atomic();
272 		atomic_add(diff_prod, &conn->bytes_to_rcv);
273 		/* guarantee 0 <= bytes_to_rcv <= rmb_desc->len */
274 		smp_mb__after_atomic();
275 		smc->sk.sk_data_ready(&smc->sk);
276 	} else {
277 		if (conn->local_rx_ctrl.prod_flags.write_blocked)
278 			smc->sk.sk_data_ready(&smc->sk);
279 		if (conn->local_rx_ctrl.prod_flags.urg_data_pending)
280 			conn->urg_state = SMC_URG_NOTYET;
281 	}
282 
283 	/* trigger sndbuf consumer: RDMA write into peer RMBE and CDC */
284 	if ((diff_cons && smc_tx_prepared_sends(conn)) ||
285 	    conn->local_rx_ctrl.prod_flags.cons_curs_upd_req ||
286 	    conn->local_rx_ctrl.prod_flags.urg_data_pending)
287 		smc_tx_sndbuf_nonempty(conn);
288 
289 	if (diff_cons && conn->urg_tx_pend &&
290 	    atomic_read(&conn->peer_rmbe_space) == conn->peer_rmbe_size) {
291 		/* urg data confirmed by peer, indicate we're ready for more */
292 		conn->urg_tx_pend = false;
293 		smc->sk.sk_write_space(&smc->sk);
294 	}
295 
296 	if (conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) {
297 		smc->sk.sk_err = ECONNRESET;
298 		conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
299 	}
300 	if (smc_cdc_rxed_any_close_or_senddone(conn)) {
301 		smc->sk.sk_shutdown |= RCV_SHUTDOWN;
302 		if (smc->clcsock && smc->clcsock->sk)
303 			smc->clcsock->sk->sk_shutdown |= RCV_SHUTDOWN;
304 		sock_set_flag(&smc->sk, SOCK_DONE);
305 		sock_hold(&smc->sk); /* sock_put in close_work */
306 		if (!schedule_work(&conn->close_work))
307 			sock_put(&smc->sk);
308 	}
309 }
310 
311 /* called under tasklet context */
312 static void smc_cdc_msg_recv(struct smc_sock *smc, struct smc_cdc_msg *cdc)
313 {
314 	sock_hold(&smc->sk);
315 	bh_lock_sock(&smc->sk);
316 	smc_cdc_msg_recv_action(smc, cdc);
317 	bh_unlock_sock(&smc->sk);
318 	sock_put(&smc->sk); /* no free sk in softirq-context */
319 }
320 
321 /* Schedule a tasklet for this connection. Triggered from the ISM device IRQ
322  * handler to indicate update in the DMBE.
323  *
324  * Context:
325  * - tasklet context
326  */
327 static void smcd_cdc_rx_tsklet(unsigned long data)
328 {
329 	struct smc_connection *conn = (struct smc_connection *)data;
330 	struct smcd_cdc_msg *data_cdc;
331 	struct smcd_cdc_msg cdc;
332 	struct smc_sock *smc;
333 
334 	if (!conn || conn->killed)
335 		return;
336 
337 	data_cdc = (struct smcd_cdc_msg *)conn->rmb_desc->cpu_addr;
338 	smcd_curs_copy(&cdc.prod, &data_cdc->prod, conn);
339 	smcd_curs_copy(&cdc.cons, &data_cdc->cons, conn);
340 	smc = container_of(conn, struct smc_sock, conn);
341 	smc_cdc_msg_recv(smc, (struct smc_cdc_msg *)&cdc);
342 }
343 
344 /* Initialize receive tasklet. Called from ISM device IRQ handler to start
345  * receiver side.
346  */
347 void smcd_cdc_rx_init(struct smc_connection *conn)
348 {
349 	tasklet_init(&conn->rx_tsklet, smcd_cdc_rx_tsklet, (unsigned long)conn);
350 }
351 
352 /***************************** init, exit, misc ******************************/
353 
354 static void smc_cdc_rx_handler(struct ib_wc *wc, void *buf)
355 {
356 	struct smc_link *link = (struct smc_link *)wc->qp->qp_context;
357 	struct smc_cdc_msg *cdc = buf;
358 	struct smc_connection *conn;
359 	struct smc_link_group *lgr;
360 	struct smc_sock *smc;
361 
362 	if (wc->byte_len < offsetof(struct smc_cdc_msg, reserved))
363 		return; /* short message */
364 	if (cdc->len != SMC_WR_TX_SIZE)
365 		return; /* invalid message */
366 
367 	/* lookup connection */
368 	lgr = smc_get_lgr(link);
369 	read_lock_bh(&lgr->conns_lock);
370 	conn = smc_lgr_find_conn(ntohl(cdc->token), lgr);
371 	read_unlock_bh(&lgr->conns_lock);
372 	if (!conn)
373 		return;
374 	smc = container_of(conn, struct smc_sock, conn);
375 
376 	if (!cdc->prod_flags.failover_validation) {
377 		if (smc_cdc_before(ntohs(cdc->seqno),
378 				   conn->local_rx_ctrl.seqno))
379 			/* received seqno is old */
380 			return;
381 	}
382 	smc_cdc_msg_recv(smc, cdc);
383 }
384 
385 static struct smc_wr_rx_handler smc_cdc_rx_handlers[] = {
386 	{
387 		.handler	= smc_cdc_rx_handler,
388 		.type		= SMC_CDC_MSG_TYPE
389 	},
390 	{
391 		.handler	= NULL,
392 	}
393 };
394 
395 int __init smc_cdc_init(void)
396 {
397 	struct smc_wr_rx_handler *handler;
398 	int rc = 0;
399 
400 	for (handler = smc_cdc_rx_handlers; handler->handler; handler++) {
401 		INIT_HLIST_NODE(&handler->list);
402 		rc = smc_wr_rx_register_handler(handler);
403 		if (rc)
404 			break;
405 	}
406 	return rc;
407 }
408