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