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
 24struct smc_cdc_tx_pend {
 25	struct smc_connection	*conn;		/* socket connection */
 26	union smc_host_cursor	cursor;	/* tx sndbuf cursor sent */
 27	union smc_host_cursor	p_cursor;	/* rx RMBE cursor produced */
 28	u16			ctrl_seq;	/* conn. tx sequence # */
 29};
 30
 31/* handler for send/transmission completion of a CDC msg */
 32static void smc_cdc_tx_handler(struct smc_wr_tx_pend_priv *pnd_snd,
 33			       struct smc_link *link,
 34			       enum ib_wc_status wc_status)
 35{
 36	struct smc_cdc_tx_pend *cdcpend = (struct smc_cdc_tx_pend *)pnd_snd;
 37	struct smc_sock *smc;
 38	int diff;
 39
 40	if (!cdcpend->conn)
 41		/* already dismissed */
 42		return;
 43
 44	smc = container_of(cdcpend->conn, struct smc_sock, conn);
 45	bh_lock_sock(&smc->sk);
 46	if (!wc_status) {
 47		diff = smc_curs_diff(cdcpend->conn->sndbuf_size,
 48				     &cdcpend->conn->tx_curs_fin,
 49				     &cdcpend->cursor);
 50		/* sndbuf_space is decreased in smc_sendmsg */
 51		smp_mb__before_atomic();
 52		atomic_add(diff, &cdcpend->conn->sndbuf_space);
 53		/* guarantee 0 <= sndbuf_space <= sndbuf_size */
 54		smp_mb__after_atomic();
 55		smc_curs_write(&cdcpend->conn->tx_curs_fin,
 56			       smc_curs_read(&cdcpend->cursor, cdcpend->conn),
 57			       cdcpend->conn);
 58	}
 59	smc_tx_sndbuf_nonfull(smc);
 60	bh_unlock_sock(&smc->sk);
 61}
 62
 63int smc_cdc_get_free_slot(struct smc_connection *conn,
 64			  struct smc_wr_buf **wr_buf,
 65			  struct smc_cdc_tx_pend **pend)
 66{
 67	struct smc_link *link = &conn->lgr->lnk[SMC_SINGLE_LINK];
 68	int rc;
 69
 70	rc = smc_wr_tx_get_free_slot(link, smc_cdc_tx_handler, wr_buf,
 71				     (struct smc_wr_tx_pend_priv **)pend);
 72	if (!conn->alert_token_local)
 73		/* abnormal termination */
 74		rc = -EPIPE;
 75	return rc;
 76}
 77
 78static inline void smc_cdc_add_pending_send(struct smc_connection *conn,
 79					    struct smc_cdc_tx_pend *pend)
 80{
 81	BUILD_BUG_ON_MSG(
 82		sizeof(struct smc_cdc_msg) > SMC_WR_BUF_SIZE,
 83		"must increase SMC_WR_BUF_SIZE to at least sizeof(struct smc_cdc_msg)");
 84	BUILD_BUG_ON_MSG(
 85		offsetof(struct smc_cdc_msg, reserved) > SMC_WR_TX_SIZE,
 86		"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()");
 87	BUILD_BUG_ON_MSG(
 88		sizeof(struct smc_cdc_tx_pend) > SMC_WR_TX_PEND_PRIV_SIZE,
 89		"must increase SMC_WR_TX_PEND_PRIV_SIZE to at least sizeof(struct smc_cdc_tx_pend)");
 90	pend->conn = conn;
 91	pend->cursor = conn->tx_curs_sent;
 92	pend->p_cursor = conn->local_tx_ctrl.prod;
 93	pend->ctrl_seq = conn->tx_cdc_seq;
 94}
 95
 96int smc_cdc_msg_send(struct smc_connection *conn,
 97		     struct smc_wr_buf *wr_buf,
 98		     struct smc_cdc_tx_pend *pend)
 99{
100	struct smc_link *link;
101	int rc;
102
103	link = &conn->lgr->lnk[SMC_SINGLE_LINK];
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,
110			    &conn->local_tx_ctrl, conn);
111	rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend);
112	if (!rc)
113		smc_curs_write(&conn->rx_curs_confirmed,
114			       smc_curs_read(&conn->local_tx_ctrl.cons, conn),
115			       conn);
116
117	return rc;
118}
119
120int smc_cdc_get_slot_and_msg_send(struct smc_connection *conn)
121{
122	struct smc_cdc_tx_pend *pend;
123	struct smc_wr_buf *wr_buf;
124	int rc;
125
126	rc = smc_cdc_get_free_slot(conn, &wr_buf, &pend);
127	if (rc)
128		return rc;
129
130	return smc_cdc_msg_send(conn, wr_buf, pend);
131}
132
133static bool smc_cdc_tx_filter(struct smc_wr_tx_pend_priv *tx_pend,
134			      unsigned long data)
135{
136	struct smc_connection *conn = (struct smc_connection *)data;
137	struct smc_cdc_tx_pend *cdc_pend =
138		(struct smc_cdc_tx_pend *)tx_pend;
139
140	return cdc_pend->conn == conn;
141}
142
143static void smc_cdc_tx_dismisser(struct smc_wr_tx_pend_priv *tx_pend)
144{
145	struct smc_cdc_tx_pend *cdc_pend =
146		(struct smc_cdc_tx_pend *)tx_pend;
147
148	cdc_pend->conn = NULL;
149}
150
151void smc_cdc_tx_dismiss_slots(struct smc_connection *conn)
152{
153	struct smc_link *link = &conn->lgr->lnk[SMC_SINGLE_LINK];
154
155	smc_wr_tx_dismiss_slots(link, SMC_CDC_MSG_TYPE,
156				smc_cdc_tx_filter, smc_cdc_tx_dismisser,
157				(unsigned long)conn);
158}
159
160/********************************* receive ***********************************/
161
162static inline bool smc_cdc_before(u16 seq1, u16 seq2)
163{
164	return (s16)(seq1 - seq2) < 0;
165}
166
167static void smc_cdc_msg_recv_action(struct smc_sock *smc,
168				    struct smc_link *link,
169				    struct smc_cdc_msg *cdc)
170{
171	union smc_host_cursor cons_old, prod_old;
172	struct smc_connection *conn = &smc->conn;
173	int diff_cons, diff_prod;
174
175	if (!cdc->prod_flags.failover_validation) {
176		if (smc_cdc_before(ntohs(cdc->seqno),
177				   conn->local_rx_ctrl.seqno))
178			/* received seqno is old */
179			return;
180	}
181	smc_curs_write(&prod_old,
182		       smc_curs_read(&conn->local_rx_ctrl.prod, conn),
183		       conn);
184	smc_curs_write(&cons_old,
185		       smc_curs_read(&conn->local_rx_ctrl.cons, conn),
186		       conn);
187	smc_cdc_msg_to_host(&conn->local_rx_ctrl, cdc, conn);
188
189	diff_cons = smc_curs_diff(conn->peer_rmbe_size, &cons_old,
190				  &conn->local_rx_ctrl.cons);
191	if (diff_cons) {
192		/* peer_rmbe_space is decreased during data transfer with RDMA
193		 * write
194		 */
195		smp_mb__before_atomic();
196		atomic_add(diff_cons, &conn->peer_rmbe_space);
197		/* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */
198		smp_mb__after_atomic();
199	}
200
201	diff_prod = smc_curs_diff(conn->rmbe_size, &prod_old,
202				  &conn->local_rx_ctrl.prod);
203	if (diff_prod) {
204		/* bytes_to_rcv is decreased in smc_recvmsg */
205		smp_mb__before_atomic();
206		atomic_add(diff_prod, &conn->bytes_to_rcv);
207		/* guarantee 0 <= bytes_to_rcv <= rmbe_size */
208		smp_mb__after_atomic();
209		smc->sk.sk_data_ready(&smc->sk);
210	} else if ((conn->local_rx_ctrl.prod_flags.write_blocked) ||
211		   (conn->local_rx_ctrl.prod_flags.cons_curs_upd_req)) {
212		smc->sk.sk_data_ready(&smc->sk);
213	}
214
215	/* piggy backed tx info */
216	/* trigger sndbuf consumer: RDMA write into peer RMBE and CDC */
217	if (diff_cons && smc_tx_prepared_sends(conn)) {
218		smc_tx_sndbuf_nonempty(conn);
219		/* trigger socket release if connection closed */
220		smc_close_wake_tx_prepared(smc);
221	}
222
223	if (conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) {
224		smc->sk.sk_err = ECONNRESET;
225		conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
226	}
227	if (smc_cdc_rxed_any_close_or_senddone(conn)) {
228		smc->sk.sk_shutdown |= RCV_SHUTDOWN;
229		if (smc->clcsock && smc->clcsock->sk)
230			smc->clcsock->sk->sk_shutdown |= RCV_SHUTDOWN;
231		sock_set_flag(&smc->sk, SOCK_DONE);
232		sock_hold(&smc->sk); /* sock_put in close_work */
233		if (!schedule_work(&conn->close_work))
234			sock_put(&smc->sk);
235	}
236}
237
238/* called under tasklet context */
239static inline void smc_cdc_msg_recv(struct smc_cdc_msg *cdc,
240				    struct smc_link *link, u64 wr_id)
241{
242	struct smc_link_group *lgr = container_of(link, struct smc_link_group,
243						  lnk[SMC_SINGLE_LINK]);
244	struct smc_connection *connection;
245	struct smc_sock *smc;
246
247	/* lookup connection */
248	read_lock_bh(&lgr->conns_lock);
249	connection = smc_lgr_find_conn(ntohl(cdc->token), lgr);
250	if (!connection) {
251		read_unlock_bh(&lgr->conns_lock);
252		return;
253	}
254	smc = container_of(connection, struct smc_sock, conn);
255	sock_hold(&smc->sk);
256	read_unlock_bh(&lgr->conns_lock);
257	bh_lock_sock(&smc->sk);
258	smc_cdc_msg_recv_action(smc, link, cdc);
259	bh_unlock_sock(&smc->sk);
260	sock_put(&smc->sk); /* no free sk in softirq-context */
261}
262
263/***************************** init, exit, misc ******************************/
264
265static void smc_cdc_rx_handler(struct ib_wc *wc, void *buf)
266{
267	struct smc_link *link = (struct smc_link *)wc->qp->qp_context;
268	struct smc_cdc_msg *cdc = buf;
269
270	if (wc->byte_len < offsetof(struct smc_cdc_msg, reserved))
271		return; /* short message */
272	if (cdc->len != SMC_WR_TX_SIZE)
273		return; /* invalid message */
274	smc_cdc_msg_recv(cdc, link, wc->wr_id);
275}
276
277static struct smc_wr_rx_handler smc_cdc_rx_handlers[] = {
278	{
279		.handler	= smc_cdc_rx_handler,
280		.type		= SMC_CDC_MSG_TYPE
281	},
282	{
283		.handler	= NULL,
284	}
285};
286
287int __init smc_cdc_init(void)
288{
289	struct smc_wr_rx_handler *handler;
290	int rc = 0;
291
292	for (handler = smc_cdc_rx_handlers; handler->handler; handler++) {
293		INIT_HLIST_NODE(&handler->list);
294		rc = smc_wr_rx_register_handler(handler);
295		if (rc)
296			break;
297	}
298	return rc;
299}