1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Shared Memory Communications over RDMA (SMC-R) and RoCE
4 *
5 * AF_SMC protocol family socket handler keeping the AF_INET sock address type
6 * applies to SOCK_STREAM sockets only
7 * offers an alternative communication option for TCP-protocol sockets
8 * applicable with RoCE-cards only
9 *
10 * Initial restrictions:
11 * - support for alternate links postponed
12 *
13 * Copyright IBM Corp. 2016, 2018
14 *
15 * Author(s): Ursula Braun <ubraun@linux.vnet.ibm.com>
16 * based on prototype from Frank Blaschka
17 */
18
19 #define KMSG_COMPONENT "smc"
20 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
21
22 #include <linux/module.h>
23 #include <linux/socket.h>
24 #include <linux/workqueue.h>
25 #include <linux/in.h>
26 #include <linux/sched/signal.h>
27 #include <linux/if_vlan.h>
28 #include <linux/rcupdate_wait.h>
29 #include <linux/ctype.h>
30
31 #include <net/sock.h>
32 #include <net/tcp.h>
33 #include <net/smc.h>
34 #include <asm/ioctls.h>
35
36 #include <net/net_namespace.h>
37 #include <net/netns/generic.h>
38 #include "smc_netns.h"
39
40 #include "smc.h"
41 #include "smc_clc.h"
42 #include "smc_llc.h"
43 #include "smc_cdc.h"
44 #include "smc_core.h"
45 #include "smc_ib.h"
46 #include "smc_ism.h"
47 #include "smc_pnet.h"
48 #include "smc_tx.h"
49 #include "smc_rx.h"
50 #include "smc_close.h"
51
52 static DEFINE_MUTEX(smc_server_lgr_pending); /* serialize link group
53 * creation on server
54 */
55 static DEFINE_MUTEX(smc_client_lgr_pending); /* serialize link group
56 * creation on client
57 */
58
59 struct workqueue_struct *smc_hs_wq; /* wq for handshake work */
60 struct workqueue_struct *smc_close_wq; /* wq for close work */
61
62 static void smc_tcp_listen_work(struct work_struct *);
63 static void smc_connect_work(struct work_struct *);
64
smc_set_keepalive(struct sock * sk,int val)65 static void smc_set_keepalive(struct sock *sk, int val)
66 {
67 struct smc_sock *smc = smc_sk(sk);
68
69 smc->clcsock->sk->sk_prot->keepalive(smc->clcsock->sk, val);
70 }
71
72 static struct smc_hashinfo smc_v4_hashinfo = {
73 .lock = __RW_LOCK_UNLOCKED(smc_v4_hashinfo.lock),
74 };
75
76 static struct smc_hashinfo smc_v6_hashinfo = {
77 .lock = __RW_LOCK_UNLOCKED(smc_v6_hashinfo.lock),
78 };
79
smc_hash_sk(struct sock * sk)80 int smc_hash_sk(struct sock *sk)
81 {
82 struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
83 struct hlist_head *head;
84
85 head = &h->ht;
86
87 write_lock_bh(&h->lock);
88 sk_add_node(sk, head);
89 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
90 write_unlock_bh(&h->lock);
91
92 return 0;
93 }
94 EXPORT_SYMBOL_GPL(smc_hash_sk);
95
smc_unhash_sk(struct sock * sk)96 void smc_unhash_sk(struct sock *sk)
97 {
98 struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
99
100 write_lock_bh(&h->lock);
101 if (sk_del_node_init(sk))
102 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
103 write_unlock_bh(&h->lock);
104 }
105 EXPORT_SYMBOL_GPL(smc_unhash_sk);
106
107 struct proto smc_proto = {
108 .name = "SMC",
109 .owner = THIS_MODULE,
110 .keepalive = smc_set_keepalive,
111 .hash = smc_hash_sk,
112 .unhash = smc_unhash_sk,
113 .obj_size = sizeof(struct smc_sock),
114 .h.smc_hash = &smc_v4_hashinfo,
115 .slab_flags = SLAB_TYPESAFE_BY_RCU,
116 };
117 EXPORT_SYMBOL_GPL(smc_proto);
118
119 struct proto smc_proto6 = {
120 .name = "SMC6",
121 .owner = THIS_MODULE,
122 .keepalive = smc_set_keepalive,
123 .hash = smc_hash_sk,
124 .unhash = smc_unhash_sk,
125 .obj_size = sizeof(struct smc_sock),
126 .h.smc_hash = &smc_v6_hashinfo,
127 .slab_flags = SLAB_TYPESAFE_BY_RCU,
128 };
129 EXPORT_SYMBOL_GPL(smc_proto6);
130
smc_restore_fallback_changes(struct smc_sock * smc)131 static void smc_restore_fallback_changes(struct smc_sock *smc)
132 {
133 if (smc->clcsock->file) { /* non-accepted sockets have no file yet */
134 smc->clcsock->file->private_data = smc->sk.sk_socket;
135 smc->clcsock->file = NULL;
136 }
137 }
138
__smc_release(struct smc_sock * smc)139 static int __smc_release(struct smc_sock *smc)
140 {
141 struct sock *sk = &smc->sk;
142 int rc = 0;
143
144 if (!smc->use_fallback) {
145 rc = smc_close_active(smc);
146 sock_set_flag(sk, SOCK_DEAD);
147 sk->sk_shutdown |= SHUTDOWN_MASK;
148 } else {
149 if (sk->sk_state != SMC_CLOSED) {
150 if (sk->sk_state != SMC_LISTEN &&
151 sk->sk_state != SMC_INIT)
152 sock_put(sk); /* passive closing */
153 if (sk->sk_state == SMC_LISTEN) {
154 /* wake up clcsock accept */
155 rc = kernel_sock_shutdown(smc->clcsock,
156 SHUT_RDWR);
157 }
158 sk->sk_state = SMC_CLOSED;
159 sk->sk_state_change(sk);
160 }
161 smc_restore_fallback_changes(smc);
162 }
163
164 sk->sk_prot->unhash(sk);
165
166 if (sk->sk_state == SMC_CLOSED) {
167 if (smc->clcsock) {
168 release_sock(sk);
169 smc_clcsock_release(smc);
170 lock_sock(sk);
171 }
172 if (!smc->use_fallback)
173 smc_conn_free(&smc->conn);
174 }
175
176 return rc;
177 }
178
smc_release(struct socket * sock)179 static int smc_release(struct socket *sock)
180 {
181 struct sock *sk = sock->sk;
182 struct smc_sock *smc;
183 int old_state, rc = 0;
184
185 if (!sk)
186 goto out;
187
188 sock_hold(sk); /* sock_put below */
189 smc = smc_sk(sk);
190
191 old_state = sk->sk_state;
192
193 /* cleanup for a dangling non-blocking connect */
194 if (smc->connect_nonblock && old_state == SMC_INIT)
195 tcp_abort(smc->clcsock->sk, ECONNABORTED);
196
197 if (cancel_work_sync(&smc->connect_work))
198 sock_put(&smc->sk); /* sock_hold in smc_connect for passive closing */
199
200 if (sk->sk_state == SMC_LISTEN)
201 /* smc_close_non_accepted() is called and acquires
202 * sock lock for child sockets again
203 */
204 lock_sock_nested(sk, SINGLE_DEPTH_NESTING);
205 else
206 lock_sock(sk);
207
208 if (old_state == SMC_INIT && sk->sk_state == SMC_ACTIVE &&
209 !smc->use_fallback)
210 smc_close_active_abort(smc);
211
212 rc = __smc_release(smc);
213
214 /* detach socket */
215 sock_orphan(sk);
216 sock->sk = NULL;
217 release_sock(sk);
218
219 sock_put(sk); /* sock_hold above */
220 sock_put(sk); /* final sock_put */
221 out:
222 return rc;
223 }
224
smc_destruct(struct sock * sk)225 static void smc_destruct(struct sock *sk)
226 {
227 if (sk->sk_state != SMC_CLOSED)
228 return;
229 if (!sock_flag(sk, SOCK_DEAD))
230 return;
231
232 sk_refcnt_debug_dec(sk);
233 }
234
smc_sock_alloc(struct net * net,struct socket * sock,int protocol)235 static struct sock *smc_sock_alloc(struct net *net, struct socket *sock,
236 int protocol)
237 {
238 struct smc_sock *smc;
239 struct proto *prot;
240 struct sock *sk;
241
242 prot = (protocol == SMCPROTO_SMC6) ? &smc_proto6 : &smc_proto;
243 sk = sk_alloc(net, PF_SMC, GFP_KERNEL, prot, 0);
244 if (!sk)
245 return NULL;
246
247 sock_init_data(sock, sk); /* sets sk_refcnt to 1 */
248 sk->sk_state = SMC_INIT;
249 sk->sk_destruct = smc_destruct;
250 sk->sk_protocol = protocol;
251 smc = smc_sk(sk);
252 INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
253 INIT_WORK(&smc->connect_work, smc_connect_work);
254 INIT_DELAYED_WORK(&smc->conn.tx_work, smc_tx_work);
255 INIT_LIST_HEAD(&smc->accept_q);
256 spin_lock_init(&smc->accept_q_lock);
257 spin_lock_init(&smc->conn.send_lock);
258 sk->sk_prot->hash(sk);
259 sk_refcnt_debug_inc(sk);
260 mutex_init(&smc->clcsock_release_lock);
261
262 return sk;
263 }
264
smc_bind(struct socket * sock,struct sockaddr * uaddr,int addr_len)265 static int smc_bind(struct socket *sock, struct sockaddr *uaddr,
266 int addr_len)
267 {
268 struct sockaddr_in *addr = (struct sockaddr_in *)uaddr;
269 struct sock *sk = sock->sk;
270 struct smc_sock *smc;
271 int rc;
272
273 smc = smc_sk(sk);
274
275 /* replicate tests from inet_bind(), to be safe wrt. future changes */
276 rc = -EINVAL;
277 if (addr_len < sizeof(struct sockaddr_in))
278 goto out;
279
280 rc = -EAFNOSUPPORT;
281 if (addr->sin_family != AF_INET &&
282 addr->sin_family != AF_INET6 &&
283 addr->sin_family != AF_UNSPEC)
284 goto out;
285 /* accept AF_UNSPEC (mapped to AF_INET) only if s_addr is INADDR_ANY */
286 if (addr->sin_family == AF_UNSPEC &&
287 addr->sin_addr.s_addr != htonl(INADDR_ANY))
288 goto out;
289
290 lock_sock(sk);
291
292 /* Check if socket is already active */
293 rc = -EINVAL;
294 if (sk->sk_state != SMC_INIT || smc->connect_nonblock)
295 goto out_rel;
296
297 smc->clcsock->sk->sk_reuse = sk->sk_reuse;
298 rc = kernel_bind(smc->clcsock, uaddr, addr_len);
299
300 out_rel:
301 release_sock(sk);
302 out:
303 return rc;
304 }
305
smc_copy_sock_settings(struct sock * nsk,struct sock * osk,unsigned long mask)306 static void smc_copy_sock_settings(struct sock *nsk, struct sock *osk,
307 unsigned long mask)
308 {
309 /* options we don't get control via setsockopt for */
310 nsk->sk_type = osk->sk_type;
311 nsk->sk_sndbuf = osk->sk_sndbuf;
312 nsk->sk_rcvbuf = osk->sk_rcvbuf;
313 nsk->sk_sndtimeo = osk->sk_sndtimeo;
314 nsk->sk_rcvtimeo = osk->sk_rcvtimeo;
315 nsk->sk_mark = osk->sk_mark;
316 nsk->sk_priority = osk->sk_priority;
317 nsk->sk_rcvlowat = osk->sk_rcvlowat;
318 nsk->sk_bound_dev_if = osk->sk_bound_dev_if;
319 nsk->sk_err = osk->sk_err;
320
321 nsk->sk_flags &= ~mask;
322 nsk->sk_flags |= osk->sk_flags & mask;
323 }
324
325 #define SK_FLAGS_SMC_TO_CLC ((1UL << SOCK_URGINLINE) | \
326 (1UL << SOCK_KEEPOPEN) | \
327 (1UL << SOCK_LINGER) | \
328 (1UL << SOCK_BROADCAST) | \
329 (1UL << SOCK_TIMESTAMP) | \
330 (1UL << SOCK_DBG) | \
331 (1UL << SOCK_RCVTSTAMP) | \
332 (1UL << SOCK_RCVTSTAMPNS) | \
333 (1UL << SOCK_LOCALROUTE) | \
334 (1UL << SOCK_TIMESTAMPING_RX_SOFTWARE) | \
335 (1UL << SOCK_RXQ_OVFL) | \
336 (1UL << SOCK_WIFI_STATUS) | \
337 (1UL << SOCK_NOFCS) | \
338 (1UL << SOCK_FILTER_LOCKED) | \
339 (1UL << SOCK_TSTAMP_NEW))
340 /* copy only relevant settings and flags of SOL_SOCKET level from smc to
341 * clc socket (since smc is not called for these options from net/core)
342 */
smc_copy_sock_settings_to_clc(struct smc_sock * smc)343 static void smc_copy_sock_settings_to_clc(struct smc_sock *smc)
344 {
345 smc_copy_sock_settings(smc->clcsock->sk, &smc->sk, SK_FLAGS_SMC_TO_CLC);
346 }
347
348 #define SK_FLAGS_CLC_TO_SMC ((1UL << SOCK_URGINLINE) | \
349 (1UL << SOCK_KEEPOPEN) | \
350 (1UL << SOCK_LINGER) | \
351 (1UL << SOCK_DBG))
352 /* copy only settings and flags relevant for smc from clc to smc socket */
smc_copy_sock_settings_to_smc(struct smc_sock * smc)353 static void smc_copy_sock_settings_to_smc(struct smc_sock *smc)
354 {
355 smc_copy_sock_settings(&smc->sk, smc->clcsock->sk, SK_FLAGS_CLC_TO_SMC);
356 }
357
358 /* register the new rmb on all links */
smcr_lgr_reg_rmbs(struct smc_link * link,struct smc_buf_desc * rmb_desc)359 static int smcr_lgr_reg_rmbs(struct smc_link *link,
360 struct smc_buf_desc *rmb_desc)
361 {
362 struct smc_link_group *lgr = link->lgr;
363 int i, rc = 0;
364
365 rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY);
366 if (rc)
367 return rc;
368 /* protect against parallel smc_llc_cli_rkey_exchange() and
369 * parallel smcr_link_reg_rmb()
370 */
371 mutex_lock(&lgr->llc_conf_mutex);
372 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
373 if (!smc_link_active(&lgr->lnk[i]))
374 continue;
375 rc = smcr_link_reg_rmb(&lgr->lnk[i], rmb_desc);
376 if (rc)
377 goto out;
378 }
379
380 /* exchange confirm_rkey msg with peer */
381 rc = smc_llc_do_confirm_rkey(link, rmb_desc);
382 if (rc) {
383 rc = -EFAULT;
384 goto out;
385 }
386 rmb_desc->is_conf_rkey = true;
387 out:
388 mutex_unlock(&lgr->llc_conf_mutex);
389 smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl);
390 return rc;
391 }
392
smcr_clnt_conf_first_link(struct smc_sock * smc)393 static int smcr_clnt_conf_first_link(struct smc_sock *smc)
394 {
395 struct smc_link *link = smc->conn.lnk;
396 struct smc_llc_qentry *qentry;
397 int rc;
398
399 /* receive CONFIRM LINK request from server over RoCE fabric */
400 qentry = smc_llc_wait(link->lgr, NULL, SMC_LLC_WAIT_TIME,
401 SMC_LLC_CONFIRM_LINK);
402 if (!qentry) {
403 struct smc_clc_msg_decline dclc;
404
405 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
406 SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
407 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
408 }
409 smc_llc_save_peer_uid(qentry);
410 rc = smc_llc_eval_conf_link(qentry, SMC_LLC_REQ);
411 smc_llc_flow_qentry_del(&link->lgr->llc_flow_lcl);
412 if (rc)
413 return SMC_CLC_DECL_RMBE_EC;
414
415 rc = smc_ib_modify_qp_rts(link);
416 if (rc)
417 return SMC_CLC_DECL_ERR_RDYLNK;
418
419 smc_wr_remember_qp_attr(link);
420
421 if (smcr_link_reg_rmb(link, smc->conn.rmb_desc))
422 return SMC_CLC_DECL_ERR_REGRMB;
423
424 /* confirm_rkey is implicit on 1st contact */
425 smc->conn.rmb_desc->is_conf_rkey = true;
426
427 /* send CONFIRM LINK response over RoCE fabric */
428 rc = smc_llc_send_confirm_link(link, SMC_LLC_RESP);
429 if (rc < 0)
430 return SMC_CLC_DECL_TIMEOUT_CL;
431
432 smc_llc_link_active(link);
433 smcr_lgr_set_type(link->lgr, SMC_LGR_SINGLE);
434
435 /* optional 2nd link, receive ADD LINK request from server */
436 qentry = smc_llc_wait(link->lgr, NULL, SMC_LLC_WAIT_TIME,
437 SMC_LLC_ADD_LINK);
438 if (!qentry) {
439 struct smc_clc_msg_decline dclc;
440
441 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
442 SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
443 if (rc == -EAGAIN)
444 rc = 0; /* no DECLINE received, go with one link */
445 return rc;
446 }
447 smc_llc_flow_qentry_clr(&link->lgr->llc_flow_lcl);
448 smc_llc_cli_add_link(link, qentry);
449 return 0;
450 }
451
smcr_conn_save_peer_info(struct smc_sock * smc,struct smc_clc_msg_accept_confirm * clc)452 static void smcr_conn_save_peer_info(struct smc_sock *smc,
453 struct smc_clc_msg_accept_confirm *clc)
454 {
455 int bufsize = smc_uncompress_bufsize(clc->r0.rmbe_size);
456
457 smc->conn.peer_rmbe_idx = clc->r0.rmbe_idx;
458 smc->conn.local_tx_ctrl.token = ntohl(clc->r0.rmbe_alert_token);
459 smc->conn.peer_rmbe_size = bufsize;
460 atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
461 smc->conn.tx_off = bufsize * (smc->conn.peer_rmbe_idx - 1);
462 }
463
smc_isascii(char * hostname)464 static bool smc_isascii(char *hostname)
465 {
466 int i;
467
468 for (i = 0; i < SMC_MAX_HOSTNAME_LEN; i++)
469 if (!isascii(hostname[i]))
470 return false;
471 return true;
472 }
473
smcd_conn_save_peer_info(struct smc_sock * smc,struct smc_clc_msg_accept_confirm * clc)474 static void smcd_conn_save_peer_info(struct smc_sock *smc,
475 struct smc_clc_msg_accept_confirm *clc)
476 {
477 int bufsize = smc_uncompress_bufsize(clc->d0.dmbe_size);
478
479 smc->conn.peer_rmbe_idx = clc->d0.dmbe_idx;
480 smc->conn.peer_token = clc->d0.token;
481 /* msg header takes up space in the buffer */
482 smc->conn.peer_rmbe_size = bufsize - sizeof(struct smcd_cdc_msg);
483 atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
484 smc->conn.tx_off = bufsize * smc->conn.peer_rmbe_idx;
485 if (clc->hdr.version > SMC_V1 &&
486 (clc->hdr.typev2 & SMC_FIRST_CONTACT_MASK)) {
487 struct smc_clc_msg_accept_confirm_v2 *clc_v2 =
488 (struct smc_clc_msg_accept_confirm_v2 *)clc;
489 struct smc_clc_first_contact_ext *fce =
490 (struct smc_clc_first_contact_ext *)
491 (((u8 *)clc_v2) + sizeof(*clc_v2));
492
493 memcpy(smc->conn.lgr->negotiated_eid, clc_v2->eid,
494 SMC_MAX_EID_LEN);
495 smc->conn.lgr->peer_os = fce->os_type;
496 smc->conn.lgr->peer_smc_release = fce->release;
497 if (smc_isascii(fce->hostname))
498 memcpy(smc->conn.lgr->peer_hostname, fce->hostname,
499 SMC_MAX_HOSTNAME_LEN);
500 }
501 }
502
smc_conn_save_peer_info(struct smc_sock * smc,struct smc_clc_msg_accept_confirm * clc)503 static void smc_conn_save_peer_info(struct smc_sock *smc,
504 struct smc_clc_msg_accept_confirm *clc)
505 {
506 if (smc->conn.lgr->is_smcd)
507 smcd_conn_save_peer_info(smc, clc);
508 else
509 smcr_conn_save_peer_info(smc, clc);
510 }
511
smc_link_save_peer_info(struct smc_link * link,struct smc_clc_msg_accept_confirm * clc)512 static void smc_link_save_peer_info(struct smc_link *link,
513 struct smc_clc_msg_accept_confirm *clc)
514 {
515 link->peer_qpn = ntoh24(clc->r0.qpn);
516 memcpy(link->peer_gid, clc->r0.lcl.gid, SMC_GID_SIZE);
517 memcpy(link->peer_mac, clc->r0.lcl.mac, sizeof(link->peer_mac));
518 link->peer_psn = ntoh24(clc->r0.psn);
519 link->peer_mtu = clc->r0.qp_mtu;
520 }
521
smc_switch_to_fallback(struct smc_sock * smc)522 static void smc_switch_to_fallback(struct smc_sock *smc)
523 {
524 wait_queue_head_t *smc_wait = sk_sleep(&smc->sk);
525 wait_queue_head_t *clc_wait = sk_sleep(smc->clcsock->sk);
526 unsigned long flags;
527
528 smc->use_fallback = true;
529 if (smc->sk.sk_socket && smc->sk.sk_socket->file) {
530 smc->clcsock->file = smc->sk.sk_socket->file;
531 smc->clcsock->file->private_data = smc->clcsock;
532 smc->clcsock->wq.fasync_list =
533 smc->sk.sk_socket->wq.fasync_list;
534
535 /* There may be some entries remaining in
536 * smc socket->wq, which should be removed
537 * to clcsocket->wq during the fallback.
538 */
539 spin_lock_irqsave(&smc_wait->lock, flags);
540 spin_lock_nested(&clc_wait->lock, SINGLE_DEPTH_NESTING);
541 list_splice_init(&smc_wait->head, &clc_wait->head);
542 spin_unlock(&clc_wait->lock);
543 spin_unlock_irqrestore(&smc_wait->lock, flags);
544 }
545 }
546
547 /* fall back during connect */
smc_connect_fallback(struct smc_sock * smc,int reason_code)548 static int smc_connect_fallback(struct smc_sock *smc, int reason_code)
549 {
550 smc_switch_to_fallback(smc);
551 smc->fallback_rsn = reason_code;
552 smc_copy_sock_settings_to_clc(smc);
553 smc->connect_nonblock = 0;
554 if (smc->sk.sk_state == SMC_INIT)
555 smc->sk.sk_state = SMC_ACTIVE;
556 return 0;
557 }
558
559 /* decline and fall back during connect */
smc_connect_decline_fallback(struct smc_sock * smc,int reason_code,u8 version)560 static int smc_connect_decline_fallback(struct smc_sock *smc, int reason_code,
561 u8 version)
562 {
563 int rc;
564
565 if (reason_code < 0) { /* error, fallback is not possible */
566 if (smc->sk.sk_state == SMC_INIT)
567 sock_put(&smc->sk); /* passive closing */
568 return reason_code;
569 }
570 if (reason_code != SMC_CLC_DECL_PEERDECL) {
571 rc = smc_clc_send_decline(smc, reason_code, version);
572 if (rc < 0) {
573 if (smc->sk.sk_state == SMC_INIT)
574 sock_put(&smc->sk); /* passive closing */
575 return rc;
576 }
577 }
578 return smc_connect_fallback(smc, reason_code);
579 }
580
581 /* abort connecting */
smc_connect_abort(struct smc_sock * smc,int local_first)582 static void smc_connect_abort(struct smc_sock *smc, int local_first)
583 {
584 if (local_first)
585 smc_lgr_cleanup_early(&smc->conn);
586 else
587 smc_conn_free(&smc->conn);
588 }
589
590 /* check if there is a rdma device available for this connection. */
591 /* called for connect and listen */
smc_find_rdma_device(struct smc_sock * smc,struct smc_init_info * ini)592 static int smc_find_rdma_device(struct smc_sock *smc, struct smc_init_info *ini)
593 {
594 /* PNET table look up: search active ib_device and port
595 * within same PNETID that also contains the ethernet device
596 * used for the internal TCP socket
597 */
598 smc_pnet_find_roce_resource(smc->clcsock->sk, ini);
599 if (!ini->ib_dev)
600 return SMC_CLC_DECL_NOSMCRDEV;
601 return 0;
602 }
603
604 /* check if there is an ISM device available for this connection. */
605 /* called for connect and listen */
smc_find_ism_device(struct smc_sock * smc,struct smc_init_info * ini)606 static int smc_find_ism_device(struct smc_sock *smc, struct smc_init_info *ini)
607 {
608 /* Find ISM device with same PNETID as connecting interface */
609 smc_pnet_find_ism_resource(smc->clcsock->sk, ini);
610 if (!ini->ism_dev[0])
611 return SMC_CLC_DECL_NOSMCDDEV;
612 else
613 ini->ism_chid[0] = smc_ism_get_chid(ini->ism_dev[0]);
614 return 0;
615 }
616
617 /* is chid unique for the ism devices that are already determined? */
smc_find_ism_v2_is_unique_chid(u16 chid,struct smc_init_info * ini,int cnt)618 static bool smc_find_ism_v2_is_unique_chid(u16 chid, struct smc_init_info *ini,
619 int cnt)
620 {
621 int i = (!ini->ism_dev[0]) ? 1 : 0;
622
623 for (; i < cnt; i++)
624 if (ini->ism_chid[i] == chid)
625 return false;
626 return true;
627 }
628
629 /* determine possible V2 ISM devices (either without PNETID or with PNETID plus
630 * PNETID matching net_device)
631 */
smc_find_ism_v2_device_clnt(struct smc_sock * smc,struct smc_init_info * ini)632 static int smc_find_ism_v2_device_clnt(struct smc_sock *smc,
633 struct smc_init_info *ini)
634 {
635 int rc = SMC_CLC_DECL_NOSMCDDEV;
636 struct smcd_dev *smcd;
637 int i = 1;
638 u16 chid;
639
640 if (smcd_indicated(ini->smc_type_v1))
641 rc = 0; /* already initialized for V1 */
642 mutex_lock(&smcd_dev_list.mutex);
643 list_for_each_entry(smcd, &smcd_dev_list.list, list) {
644 if (smcd->going_away || smcd == ini->ism_dev[0])
645 continue;
646 chid = smc_ism_get_chid(smcd);
647 if (!smc_find_ism_v2_is_unique_chid(chid, ini, i))
648 continue;
649 if (!smc_pnet_is_pnetid_set(smcd->pnetid) ||
650 smc_pnet_is_ndev_pnetid(sock_net(&smc->sk), smcd->pnetid)) {
651 ini->ism_dev[i] = smcd;
652 ini->ism_chid[i] = chid;
653 ini->is_smcd = true;
654 rc = 0;
655 i++;
656 if (i > SMC_MAX_ISM_DEVS)
657 break;
658 }
659 }
660 mutex_unlock(&smcd_dev_list.mutex);
661 ini->ism_offered_cnt = i - 1;
662 if (!ini->ism_dev[0] && !ini->ism_dev[1])
663 ini->smcd_version = 0;
664
665 return rc;
666 }
667
668 /* Check for VLAN ID and register it on ISM device just for CLC handshake */
smc_connect_ism_vlan_setup(struct smc_sock * smc,struct smc_init_info * ini)669 static int smc_connect_ism_vlan_setup(struct smc_sock *smc,
670 struct smc_init_info *ini)
671 {
672 if (ini->vlan_id && smc_ism_get_vlan(ini->ism_dev[0], ini->vlan_id))
673 return SMC_CLC_DECL_ISMVLANERR;
674 return 0;
675 }
676
smc_find_proposal_devices(struct smc_sock * smc,struct smc_init_info * ini)677 static int smc_find_proposal_devices(struct smc_sock *smc,
678 struct smc_init_info *ini)
679 {
680 int rc = 0;
681
682 /* check if there is an ism device available */
683 if (ini->smcd_version & SMC_V1) {
684 if (smc_find_ism_device(smc, ini) ||
685 smc_connect_ism_vlan_setup(smc, ini)) {
686 if (ini->smc_type_v1 == SMC_TYPE_B)
687 ini->smc_type_v1 = SMC_TYPE_R;
688 else
689 ini->smc_type_v1 = SMC_TYPE_N;
690 } /* else ISM V1 is supported for this connection */
691 if (smc_find_rdma_device(smc, ini)) {
692 if (ini->smc_type_v1 == SMC_TYPE_B)
693 ini->smc_type_v1 = SMC_TYPE_D;
694 else
695 ini->smc_type_v1 = SMC_TYPE_N;
696 } /* else RDMA is supported for this connection */
697 }
698 if (smc_ism_v2_capable && smc_find_ism_v2_device_clnt(smc, ini))
699 ini->smc_type_v2 = SMC_TYPE_N;
700
701 /* if neither ISM nor RDMA are supported, fallback */
702 if (!smcr_indicated(ini->smc_type_v1) &&
703 ini->smc_type_v1 == SMC_TYPE_N && ini->smc_type_v2 == SMC_TYPE_N)
704 rc = SMC_CLC_DECL_NOSMCDEV;
705
706 return rc;
707 }
708
709 /* cleanup temporary VLAN ID registration used for CLC handshake. If ISM is
710 * used, the VLAN ID will be registered again during the connection setup.
711 */
smc_connect_ism_vlan_cleanup(struct smc_sock * smc,struct smc_init_info * ini)712 static int smc_connect_ism_vlan_cleanup(struct smc_sock *smc,
713 struct smc_init_info *ini)
714 {
715 if (!smcd_indicated(ini->smc_type_v1))
716 return 0;
717 if (ini->vlan_id && smc_ism_put_vlan(ini->ism_dev[0], ini->vlan_id))
718 return SMC_CLC_DECL_CNFERR;
719 return 0;
720 }
721
722 #define SMC_CLC_MAX_ACCEPT_LEN \
723 (sizeof(struct smc_clc_msg_accept_confirm_v2) + \
724 sizeof(struct smc_clc_first_contact_ext) + \
725 sizeof(struct smc_clc_msg_trail))
726
727 /* CLC handshake during connect */
smc_connect_clc(struct smc_sock * smc,struct smc_clc_msg_accept_confirm_v2 * aclc2,struct smc_init_info * ini)728 static int smc_connect_clc(struct smc_sock *smc,
729 struct smc_clc_msg_accept_confirm_v2 *aclc2,
730 struct smc_init_info *ini)
731 {
732 int rc = 0;
733
734 /* do inband token exchange */
735 rc = smc_clc_send_proposal(smc, ini);
736 if (rc)
737 return rc;
738 /* receive SMC Accept CLC message */
739 return smc_clc_wait_msg(smc, aclc2, SMC_CLC_MAX_ACCEPT_LEN,
740 SMC_CLC_ACCEPT, CLC_WAIT_TIME);
741 }
742
743 /* setup for RDMA connection of client */
smc_connect_rdma(struct smc_sock * smc,struct smc_clc_msg_accept_confirm * aclc,struct smc_init_info * ini)744 static int smc_connect_rdma(struct smc_sock *smc,
745 struct smc_clc_msg_accept_confirm *aclc,
746 struct smc_init_info *ini)
747 {
748 int i, reason_code = 0;
749 struct smc_link *link;
750
751 ini->is_smcd = false;
752 ini->ib_lcl = &aclc->r0.lcl;
753 ini->ib_clcqpn = ntoh24(aclc->r0.qpn);
754 ini->first_contact_peer = aclc->hdr.typev2 & SMC_FIRST_CONTACT_MASK;
755
756 mutex_lock(&smc_client_lgr_pending);
757 reason_code = smc_conn_create(smc, ini);
758 if (reason_code) {
759 mutex_unlock(&smc_client_lgr_pending);
760 return reason_code;
761 }
762
763 smc_conn_save_peer_info(smc, aclc);
764
765 if (ini->first_contact_local) {
766 link = smc->conn.lnk;
767 } else {
768 /* set link that was assigned by server */
769 link = NULL;
770 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
771 struct smc_link *l = &smc->conn.lgr->lnk[i];
772
773 if (l->peer_qpn == ntoh24(aclc->r0.qpn) &&
774 !memcmp(l->peer_gid, &aclc->r0.lcl.gid,
775 SMC_GID_SIZE) &&
776 !memcmp(l->peer_mac, &aclc->r0.lcl.mac,
777 sizeof(l->peer_mac))) {
778 link = l;
779 break;
780 }
781 }
782 if (!link) {
783 reason_code = SMC_CLC_DECL_NOSRVLINK;
784 goto connect_abort;
785 }
786 smc->conn.lnk = link;
787 }
788
789 /* create send buffer and rmb */
790 if (smc_buf_create(smc, false)) {
791 reason_code = SMC_CLC_DECL_MEM;
792 goto connect_abort;
793 }
794
795 if (ini->first_contact_local)
796 smc_link_save_peer_info(link, aclc);
797
798 if (smc_rmb_rtoken_handling(&smc->conn, link, aclc)) {
799 reason_code = SMC_CLC_DECL_ERR_RTOK;
800 goto connect_abort;
801 }
802
803 smc_close_init(smc);
804 smc_rx_init(smc);
805
806 if (ini->first_contact_local) {
807 if (smc_ib_ready_link(link)) {
808 reason_code = SMC_CLC_DECL_ERR_RDYLNK;
809 goto connect_abort;
810 }
811 } else {
812 if (smcr_lgr_reg_rmbs(link, smc->conn.rmb_desc)) {
813 reason_code = SMC_CLC_DECL_ERR_REGRMB;
814 goto connect_abort;
815 }
816 }
817 smc_rmb_sync_sg_for_device(&smc->conn);
818
819 reason_code = smc_clc_send_confirm(smc, ini->first_contact_local,
820 SMC_V1);
821 if (reason_code)
822 goto connect_abort;
823
824 smc_tx_init(smc);
825
826 if (ini->first_contact_local) {
827 /* QP confirmation over RoCE fabric */
828 smc_llc_flow_initiate(link->lgr, SMC_LLC_FLOW_ADD_LINK);
829 reason_code = smcr_clnt_conf_first_link(smc);
830 smc_llc_flow_stop(link->lgr, &link->lgr->llc_flow_lcl);
831 if (reason_code)
832 goto connect_abort;
833 }
834 mutex_unlock(&smc_client_lgr_pending);
835
836 smc_copy_sock_settings_to_clc(smc);
837 smc->connect_nonblock = 0;
838 if (smc->sk.sk_state == SMC_INIT)
839 smc->sk.sk_state = SMC_ACTIVE;
840
841 return 0;
842 connect_abort:
843 smc_connect_abort(smc, ini->first_contact_local);
844 mutex_unlock(&smc_client_lgr_pending);
845 smc->connect_nonblock = 0;
846
847 return reason_code;
848 }
849
850 /* The server has chosen one of the proposed ISM devices for the communication.
851 * Determine from the CHID of the received CLC ACCEPT the ISM device chosen.
852 */
853 static int
smc_v2_determine_accepted_chid(struct smc_clc_msg_accept_confirm_v2 * aclc,struct smc_init_info * ini)854 smc_v2_determine_accepted_chid(struct smc_clc_msg_accept_confirm_v2 *aclc,
855 struct smc_init_info *ini)
856 {
857 int i;
858
859 for (i = 0; i < ini->ism_offered_cnt + 1; i++) {
860 if (ini->ism_chid[i] == ntohs(aclc->chid)) {
861 ini->ism_selected = i;
862 return 0;
863 }
864 }
865
866 return -EPROTO;
867 }
868
869 /* setup for ISM connection of client */
smc_connect_ism(struct smc_sock * smc,struct smc_clc_msg_accept_confirm * aclc,struct smc_init_info * ini)870 static int smc_connect_ism(struct smc_sock *smc,
871 struct smc_clc_msg_accept_confirm *aclc,
872 struct smc_init_info *ini)
873 {
874 int rc = 0;
875
876 ini->is_smcd = true;
877 ini->first_contact_peer = aclc->hdr.typev2 & SMC_FIRST_CONTACT_MASK;
878
879 if (aclc->hdr.version == SMC_V2) {
880 struct smc_clc_msg_accept_confirm_v2 *aclc_v2 =
881 (struct smc_clc_msg_accept_confirm_v2 *)aclc;
882
883 rc = smc_v2_determine_accepted_chid(aclc_v2, ini);
884 if (rc)
885 return rc;
886 }
887 ini->ism_peer_gid[ini->ism_selected] = aclc->d0.gid;
888
889 /* there is only one lgr role for SMC-D; use server lock */
890 mutex_lock(&smc_server_lgr_pending);
891 rc = smc_conn_create(smc, ini);
892 if (rc) {
893 mutex_unlock(&smc_server_lgr_pending);
894 return rc;
895 }
896
897 /* Create send and receive buffers */
898 rc = smc_buf_create(smc, true);
899 if (rc) {
900 rc = (rc == -ENOSPC) ? SMC_CLC_DECL_MAX_DMB : SMC_CLC_DECL_MEM;
901 goto connect_abort;
902 }
903
904 smc_conn_save_peer_info(smc, aclc);
905 smc_close_init(smc);
906 smc_rx_init(smc);
907 smc_tx_init(smc);
908
909 rc = smc_clc_send_confirm(smc, ini->first_contact_local,
910 aclc->hdr.version);
911 if (rc)
912 goto connect_abort;
913 mutex_unlock(&smc_server_lgr_pending);
914
915 smc_copy_sock_settings_to_clc(smc);
916 smc->connect_nonblock = 0;
917 if (smc->sk.sk_state == SMC_INIT)
918 smc->sk.sk_state = SMC_ACTIVE;
919
920 return 0;
921 connect_abort:
922 smc_connect_abort(smc, ini->first_contact_local);
923 mutex_unlock(&smc_server_lgr_pending);
924 smc->connect_nonblock = 0;
925
926 return rc;
927 }
928
929 /* check if received accept type and version matches a proposed one */
smc_connect_check_aclc(struct smc_init_info * ini,struct smc_clc_msg_accept_confirm * aclc)930 static int smc_connect_check_aclc(struct smc_init_info *ini,
931 struct smc_clc_msg_accept_confirm *aclc)
932 {
933 if ((aclc->hdr.typev1 == SMC_TYPE_R &&
934 !smcr_indicated(ini->smc_type_v1)) ||
935 (aclc->hdr.typev1 == SMC_TYPE_D &&
936 ((!smcd_indicated(ini->smc_type_v1) &&
937 !smcd_indicated(ini->smc_type_v2)) ||
938 (aclc->hdr.version == SMC_V1 &&
939 !smcd_indicated(ini->smc_type_v1)) ||
940 (aclc->hdr.version == SMC_V2 &&
941 !smcd_indicated(ini->smc_type_v2)))))
942 return SMC_CLC_DECL_MODEUNSUPP;
943
944 return 0;
945 }
946
947 /* perform steps before actually connecting */
__smc_connect(struct smc_sock * smc)948 static int __smc_connect(struct smc_sock *smc)
949 {
950 u8 version = smc_ism_v2_capable ? SMC_V2 : SMC_V1;
951 struct smc_clc_msg_accept_confirm_v2 *aclc2;
952 struct smc_clc_msg_accept_confirm *aclc;
953 struct smc_init_info *ini = NULL;
954 u8 *buf = NULL;
955 int rc = 0;
956
957 if (smc->use_fallback)
958 return smc_connect_fallback(smc, smc->fallback_rsn);
959
960 /* if peer has not signalled SMC-capability, fall back */
961 if (!tcp_sk(smc->clcsock->sk)->syn_smc)
962 return smc_connect_fallback(smc, SMC_CLC_DECL_PEERNOSMC);
963
964 /* IPSec connections opt out of SMC optimizations */
965 if (using_ipsec(smc))
966 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_IPSEC,
967 version);
968
969 ini = kzalloc(sizeof(*ini), GFP_KERNEL);
970 if (!ini)
971 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_MEM,
972 version);
973
974 ini->smcd_version = SMC_V1;
975 ini->smcd_version |= smc_ism_v2_capable ? SMC_V2 : 0;
976 ini->smc_type_v1 = SMC_TYPE_B;
977 ini->smc_type_v2 = smc_ism_v2_capable ? SMC_TYPE_D : SMC_TYPE_N;
978
979 /* get vlan id from IP device */
980 if (smc_vlan_by_tcpsk(smc->clcsock, ini)) {
981 ini->smcd_version &= ~SMC_V1;
982 ini->smc_type_v1 = SMC_TYPE_N;
983 if (!ini->smcd_version) {
984 rc = SMC_CLC_DECL_GETVLANERR;
985 goto fallback;
986 }
987 }
988
989 rc = smc_find_proposal_devices(smc, ini);
990 if (rc)
991 goto fallback;
992
993 buf = kzalloc(SMC_CLC_MAX_ACCEPT_LEN, GFP_KERNEL);
994 if (!buf) {
995 rc = SMC_CLC_DECL_MEM;
996 goto fallback;
997 }
998 aclc2 = (struct smc_clc_msg_accept_confirm_v2 *)buf;
999 aclc = (struct smc_clc_msg_accept_confirm *)aclc2;
1000
1001 /* perform CLC handshake */
1002 rc = smc_connect_clc(smc, aclc2, ini);
1003 if (rc)
1004 goto vlan_cleanup;
1005
1006 /* check if smc modes and versions of CLC proposal and accept match */
1007 rc = smc_connect_check_aclc(ini, aclc);
1008 version = aclc->hdr.version == SMC_V1 ? SMC_V1 : SMC_V2;
1009 ini->smcd_version = version;
1010 if (rc)
1011 goto vlan_cleanup;
1012
1013 /* depending on previous steps, connect using rdma or ism */
1014 if (aclc->hdr.typev1 == SMC_TYPE_R)
1015 rc = smc_connect_rdma(smc, aclc, ini);
1016 else if (aclc->hdr.typev1 == SMC_TYPE_D)
1017 rc = smc_connect_ism(smc, aclc, ini);
1018 if (rc)
1019 goto vlan_cleanup;
1020
1021 smc_connect_ism_vlan_cleanup(smc, ini);
1022 kfree(buf);
1023 kfree(ini);
1024 return 0;
1025
1026 vlan_cleanup:
1027 smc_connect_ism_vlan_cleanup(smc, ini);
1028 kfree(buf);
1029 fallback:
1030 kfree(ini);
1031 return smc_connect_decline_fallback(smc, rc, version);
1032 }
1033
smc_connect_work(struct work_struct * work)1034 static void smc_connect_work(struct work_struct *work)
1035 {
1036 struct smc_sock *smc = container_of(work, struct smc_sock,
1037 connect_work);
1038 long timeo = smc->sk.sk_sndtimeo;
1039 int rc = 0;
1040
1041 if (!timeo)
1042 timeo = MAX_SCHEDULE_TIMEOUT;
1043 lock_sock(smc->clcsock->sk);
1044 if (smc->clcsock->sk->sk_err) {
1045 smc->sk.sk_err = smc->clcsock->sk->sk_err;
1046 } else if ((1 << smc->clcsock->sk->sk_state) &
1047 (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
1048 rc = sk_stream_wait_connect(smc->clcsock->sk, &timeo);
1049 if ((rc == -EPIPE) &&
1050 ((1 << smc->clcsock->sk->sk_state) &
1051 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)))
1052 rc = 0;
1053 }
1054 release_sock(smc->clcsock->sk);
1055 lock_sock(&smc->sk);
1056 if (rc != 0 || smc->sk.sk_err) {
1057 smc->sk.sk_state = SMC_CLOSED;
1058 if (rc == -EPIPE || rc == -EAGAIN)
1059 smc->sk.sk_err = EPIPE;
1060 else if (rc == -ECONNREFUSED)
1061 smc->sk.sk_err = ECONNREFUSED;
1062 else if (signal_pending(current))
1063 smc->sk.sk_err = -sock_intr_errno(timeo);
1064 sock_put(&smc->sk); /* passive closing */
1065 goto out;
1066 }
1067
1068 rc = __smc_connect(smc);
1069 if (rc < 0)
1070 smc->sk.sk_err = -rc;
1071
1072 out:
1073 if (!sock_flag(&smc->sk, SOCK_DEAD)) {
1074 if (smc->sk.sk_err) {
1075 smc->sk.sk_state_change(&smc->sk);
1076 } else { /* allow polling before and after fallback decision */
1077 smc->clcsock->sk->sk_write_space(smc->clcsock->sk);
1078 smc->sk.sk_write_space(&smc->sk);
1079 }
1080 }
1081 release_sock(&smc->sk);
1082 }
1083
smc_connect(struct socket * sock,struct sockaddr * addr,int alen,int flags)1084 static int smc_connect(struct socket *sock, struct sockaddr *addr,
1085 int alen, int flags)
1086 {
1087 struct sock *sk = sock->sk;
1088 struct smc_sock *smc;
1089 int rc = -EINVAL;
1090
1091 smc = smc_sk(sk);
1092
1093 /* separate smc parameter checking to be safe */
1094 if (alen < sizeof(addr->sa_family))
1095 goto out_err;
1096 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
1097 goto out_err;
1098
1099 lock_sock(sk);
1100 switch (sk->sk_state) {
1101 default:
1102 goto out;
1103 case SMC_ACTIVE:
1104 rc = -EISCONN;
1105 goto out;
1106 case SMC_INIT:
1107 rc = 0;
1108 break;
1109 }
1110
1111 smc_copy_sock_settings_to_clc(smc);
1112 tcp_sk(smc->clcsock->sk)->syn_smc = 1;
1113 if (smc->connect_nonblock) {
1114 rc = -EALREADY;
1115 goto out;
1116 }
1117 rc = kernel_connect(smc->clcsock, addr, alen, flags);
1118 if (rc && rc != -EINPROGRESS)
1119 goto out;
1120
1121 if (smc->use_fallback)
1122 goto out;
1123 sock_hold(&smc->sk); /* sock put in passive closing */
1124 if (flags & O_NONBLOCK) {
1125 if (queue_work(smc_hs_wq, &smc->connect_work))
1126 smc->connect_nonblock = 1;
1127 rc = -EINPROGRESS;
1128 } else {
1129 rc = __smc_connect(smc);
1130 if (rc < 0)
1131 goto out;
1132 else
1133 rc = 0; /* success cases including fallback */
1134 }
1135
1136 out:
1137 release_sock(sk);
1138 out_err:
1139 return rc;
1140 }
1141
smc_clcsock_accept(struct smc_sock * lsmc,struct smc_sock ** new_smc)1142 static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
1143 {
1144 struct socket *new_clcsock = NULL;
1145 struct sock *lsk = &lsmc->sk;
1146 struct sock *new_sk;
1147 int rc = -EINVAL;
1148
1149 release_sock(lsk);
1150 new_sk = smc_sock_alloc(sock_net(lsk), NULL, lsk->sk_protocol);
1151 if (!new_sk) {
1152 rc = -ENOMEM;
1153 lsk->sk_err = ENOMEM;
1154 *new_smc = NULL;
1155 lock_sock(lsk);
1156 goto out;
1157 }
1158 *new_smc = smc_sk(new_sk);
1159
1160 mutex_lock(&lsmc->clcsock_release_lock);
1161 if (lsmc->clcsock)
1162 rc = kernel_accept(lsmc->clcsock, &new_clcsock, SOCK_NONBLOCK);
1163 mutex_unlock(&lsmc->clcsock_release_lock);
1164 lock_sock(lsk);
1165 if (rc < 0 && rc != -EAGAIN)
1166 lsk->sk_err = -rc;
1167 if (rc < 0 || lsk->sk_state == SMC_CLOSED) {
1168 new_sk->sk_prot->unhash(new_sk);
1169 if (new_clcsock)
1170 sock_release(new_clcsock);
1171 new_sk->sk_state = SMC_CLOSED;
1172 sock_set_flag(new_sk, SOCK_DEAD);
1173 sock_put(new_sk); /* final */
1174 *new_smc = NULL;
1175 goto out;
1176 }
1177
1178 /* new clcsock has inherited the smc listen-specific sk_data_ready
1179 * function; switch it back to the original sk_data_ready function
1180 */
1181 new_clcsock->sk->sk_data_ready = lsmc->clcsk_data_ready;
1182 (*new_smc)->clcsock = new_clcsock;
1183 out:
1184 return rc;
1185 }
1186
1187 /* add a just created sock to the accept queue of the listen sock as
1188 * candidate for a following socket accept call from user space
1189 */
smc_accept_enqueue(struct sock * parent,struct sock * sk)1190 static void smc_accept_enqueue(struct sock *parent, struct sock *sk)
1191 {
1192 struct smc_sock *par = smc_sk(parent);
1193
1194 sock_hold(sk); /* sock_put in smc_accept_unlink () */
1195 spin_lock(&par->accept_q_lock);
1196 list_add_tail(&smc_sk(sk)->accept_q, &par->accept_q);
1197 spin_unlock(&par->accept_q_lock);
1198 sk_acceptq_added(parent);
1199 }
1200
1201 /* remove a socket from the accept queue of its parental listening socket */
smc_accept_unlink(struct sock * sk)1202 static void smc_accept_unlink(struct sock *sk)
1203 {
1204 struct smc_sock *par = smc_sk(sk)->listen_smc;
1205
1206 spin_lock(&par->accept_q_lock);
1207 list_del_init(&smc_sk(sk)->accept_q);
1208 spin_unlock(&par->accept_q_lock);
1209 sk_acceptq_removed(&smc_sk(sk)->listen_smc->sk);
1210 sock_put(sk); /* sock_hold in smc_accept_enqueue */
1211 }
1212
1213 /* remove a sock from the accept queue to bind it to a new socket created
1214 * for a socket accept call from user space
1215 */
smc_accept_dequeue(struct sock * parent,struct socket * new_sock)1216 struct sock *smc_accept_dequeue(struct sock *parent,
1217 struct socket *new_sock)
1218 {
1219 struct smc_sock *isk, *n;
1220 struct sock *new_sk;
1221
1222 list_for_each_entry_safe(isk, n, &smc_sk(parent)->accept_q, accept_q) {
1223 new_sk = (struct sock *)isk;
1224
1225 smc_accept_unlink(new_sk);
1226 if (new_sk->sk_state == SMC_CLOSED) {
1227 new_sk->sk_prot->unhash(new_sk);
1228 if (isk->clcsock) {
1229 sock_release(isk->clcsock);
1230 isk->clcsock = NULL;
1231 }
1232 sock_put(new_sk); /* final */
1233 continue;
1234 }
1235 if (new_sock) {
1236 sock_graft(new_sk, new_sock);
1237 if (isk->use_fallback) {
1238 smc_sk(new_sk)->clcsock->file = new_sock->file;
1239 isk->clcsock->file->private_data = isk->clcsock;
1240 }
1241 }
1242 return new_sk;
1243 }
1244 return NULL;
1245 }
1246
1247 /* clean up for a created but never accepted sock */
smc_close_non_accepted(struct sock * sk)1248 void smc_close_non_accepted(struct sock *sk)
1249 {
1250 struct smc_sock *smc = smc_sk(sk);
1251
1252 sock_hold(sk); /* sock_put below */
1253 lock_sock(sk);
1254 if (!sk->sk_lingertime)
1255 /* wait for peer closing */
1256 sk->sk_lingertime = SMC_MAX_STREAM_WAIT_TIMEOUT;
1257 __smc_release(smc);
1258 release_sock(sk);
1259 sock_put(sk); /* sock_hold above */
1260 sock_put(sk); /* final sock_put */
1261 }
1262
smcr_serv_conf_first_link(struct smc_sock * smc)1263 static int smcr_serv_conf_first_link(struct smc_sock *smc)
1264 {
1265 struct smc_link *link = smc->conn.lnk;
1266 struct smc_llc_qentry *qentry;
1267 int rc;
1268
1269 if (smcr_link_reg_rmb(link, smc->conn.rmb_desc))
1270 return SMC_CLC_DECL_ERR_REGRMB;
1271
1272 /* send CONFIRM LINK request to client over the RoCE fabric */
1273 rc = smc_llc_send_confirm_link(link, SMC_LLC_REQ);
1274 if (rc < 0)
1275 return SMC_CLC_DECL_TIMEOUT_CL;
1276
1277 /* receive CONFIRM LINK response from client over the RoCE fabric */
1278 qentry = smc_llc_wait(link->lgr, link, SMC_LLC_WAIT_TIME,
1279 SMC_LLC_CONFIRM_LINK);
1280 if (!qentry) {
1281 struct smc_clc_msg_decline dclc;
1282
1283 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
1284 SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
1285 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
1286 }
1287 smc_llc_save_peer_uid(qentry);
1288 rc = smc_llc_eval_conf_link(qentry, SMC_LLC_RESP);
1289 smc_llc_flow_qentry_del(&link->lgr->llc_flow_lcl);
1290 if (rc)
1291 return SMC_CLC_DECL_RMBE_EC;
1292
1293 /* confirm_rkey is implicit on 1st contact */
1294 smc->conn.rmb_desc->is_conf_rkey = true;
1295
1296 smc_llc_link_active(link);
1297 smcr_lgr_set_type(link->lgr, SMC_LGR_SINGLE);
1298
1299 /* initial contact - try to establish second link */
1300 smc_llc_srv_add_link(link);
1301 return 0;
1302 }
1303
1304 /* listen worker: finish */
smc_listen_out(struct smc_sock * new_smc)1305 static void smc_listen_out(struct smc_sock *new_smc)
1306 {
1307 struct smc_sock *lsmc = new_smc->listen_smc;
1308 struct sock *newsmcsk = &new_smc->sk;
1309
1310 if (lsmc->sk.sk_state == SMC_LISTEN) {
1311 lock_sock_nested(&lsmc->sk, SINGLE_DEPTH_NESTING);
1312 smc_accept_enqueue(&lsmc->sk, newsmcsk);
1313 release_sock(&lsmc->sk);
1314 } else { /* no longer listening */
1315 smc_close_non_accepted(newsmcsk);
1316 }
1317
1318 /* Wake up accept */
1319 lsmc->sk.sk_data_ready(&lsmc->sk);
1320 sock_put(&lsmc->sk); /* sock_hold in smc_tcp_listen_work */
1321 }
1322
1323 /* listen worker: finish in state connected */
smc_listen_out_connected(struct smc_sock * new_smc)1324 static void smc_listen_out_connected(struct smc_sock *new_smc)
1325 {
1326 struct sock *newsmcsk = &new_smc->sk;
1327
1328 if (newsmcsk->sk_state == SMC_INIT)
1329 newsmcsk->sk_state = SMC_ACTIVE;
1330
1331 smc_listen_out(new_smc);
1332 }
1333
1334 /* listen worker: finish in error state */
smc_listen_out_err(struct smc_sock * new_smc)1335 static void smc_listen_out_err(struct smc_sock *new_smc)
1336 {
1337 struct sock *newsmcsk = &new_smc->sk;
1338
1339 if (newsmcsk->sk_state == SMC_INIT)
1340 sock_put(&new_smc->sk); /* passive closing */
1341 newsmcsk->sk_state = SMC_CLOSED;
1342
1343 smc_listen_out(new_smc);
1344 }
1345
1346 /* listen worker: decline and fall back if possible */
smc_listen_decline(struct smc_sock * new_smc,int reason_code,int local_first,u8 version)1347 static void smc_listen_decline(struct smc_sock *new_smc, int reason_code,
1348 int local_first, u8 version)
1349 {
1350 /* RDMA setup failed, switch back to TCP */
1351 if (local_first)
1352 smc_lgr_cleanup_early(&new_smc->conn);
1353 else
1354 smc_conn_free(&new_smc->conn);
1355 if (reason_code < 0) { /* error, no fallback possible */
1356 smc_listen_out_err(new_smc);
1357 return;
1358 }
1359 smc_switch_to_fallback(new_smc);
1360 new_smc->fallback_rsn = reason_code;
1361 if (reason_code && reason_code != SMC_CLC_DECL_PEERDECL) {
1362 if (smc_clc_send_decline(new_smc, reason_code, version) < 0) {
1363 smc_listen_out_err(new_smc);
1364 return;
1365 }
1366 }
1367 smc_listen_out_connected(new_smc);
1368 }
1369
1370 /* listen worker: version checking */
smc_listen_v2_check(struct smc_sock * new_smc,struct smc_clc_msg_proposal * pclc,struct smc_init_info * ini)1371 static int smc_listen_v2_check(struct smc_sock *new_smc,
1372 struct smc_clc_msg_proposal *pclc,
1373 struct smc_init_info *ini)
1374 {
1375 struct smc_clc_smcd_v2_extension *pclc_smcd_v2_ext;
1376 struct smc_clc_v2_extension *pclc_v2_ext;
1377
1378 ini->smc_type_v1 = pclc->hdr.typev1;
1379 ini->smc_type_v2 = pclc->hdr.typev2;
1380 ini->smcd_version = ini->smc_type_v1 != SMC_TYPE_N ? SMC_V1 : 0;
1381 if (pclc->hdr.version > SMC_V1)
1382 ini->smcd_version |=
1383 ini->smc_type_v2 != SMC_TYPE_N ? SMC_V2 : 0;
1384 if (!smc_ism_v2_capable) {
1385 ini->smcd_version &= ~SMC_V2;
1386 goto out;
1387 }
1388 pclc_v2_ext = smc_get_clc_v2_ext(pclc);
1389 if (!pclc_v2_ext) {
1390 ini->smcd_version &= ~SMC_V2;
1391 goto out;
1392 }
1393 pclc_smcd_v2_ext = smc_get_clc_smcd_v2_ext(pclc_v2_ext);
1394 if (!pclc_smcd_v2_ext)
1395 ini->smcd_version &= ~SMC_V2;
1396
1397 out:
1398 if (!ini->smcd_version) {
1399 if (pclc->hdr.typev1 == SMC_TYPE_B ||
1400 pclc->hdr.typev2 == SMC_TYPE_B)
1401 return SMC_CLC_DECL_NOSMCDEV;
1402 if (pclc->hdr.typev1 == SMC_TYPE_D ||
1403 pclc->hdr.typev2 == SMC_TYPE_D)
1404 return SMC_CLC_DECL_NOSMCDDEV;
1405 return SMC_CLC_DECL_NOSMCRDEV;
1406 }
1407
1408 return 0;
1409 }
1410
1411 /* listen worker: check prefixes */
smc_listen_prfx_check(struct smc_sock * new_smc,struct smc_clc_msg_proposal * pclc)1412 static int smc_listen_prfx_check(struct smc_sock *new_smc,
1413 struct smc_clc_msg_proposal *pclc)
1414 {
1415 struct smc_clc_msg_proposal_prefix *pclc_prfx;
1416 struct socket *newclcsock = new_smc->clcsock;
1417
1418 if (pclc->hdr.typev1 == SMC_TYPE_N)
1419 return 0;
1420 pclc_prfx = smc_clc_proposal_get_prefix(pclc);
1421 if (smc_clc_prfx_match(newclcsock, pclc_prfx))
1422 return SMC_CLC_DECL_DIFFPREFIX;
1423
1424 return 0;
1425 }
1426
1427 /* listen worker: initialize connection and buffers */
smc_listen_rdma_init(struct smc_sock * new_smc,struct smc_init_info * ini)1428 static int smc_listen_rdma_init(struct smc_sock *new_smc,
1429 struct smc_init_info *ini)
1430 {
1431 int rc;
1432
1433 /* allocate connection / link group */
1434 rc = smc_conn_create(new_smc, ini);
1435 if (rc)
1436 return rc;
1437
1438 /* create send buffer and rmb */
1439 if (smc_buf_create(new_smc, false))
1440 return SMC_CLC_DECL_MEM;
1441
1442 return 0;
1443 }
1444
1445 /* listen worker: initialize connection and buffers for SMC-D */
smc_listen_ism_init(struct smc_sock * new_smc,struct smc_init_info * ini)1446 static int smc_listen_ism_init(struct smc_sock *new_smc,
1447 struct smc_init_info *ini)
1448 {
1449 int rc;
1450
1451 rc = smc_conn_create(new_smc, ini);
1452 if (rc)
1453 return rc;
1454
1455 /* Create send and receive buffers */
1456 rc = smc_buf_create(new_smc, true);
1457 if (rc) {
1458 if (ini->first_contact_local)
1459 smc_lgr_cleanup_early(&new_smc->conn);
1460 else
1461 smc_conn_free(&new_smc->conn);
1462 return (rc == -ENOSPC) ? SMC_CLC_DECL_MAX_DMB :
1463 SMC_CLC_DECL_MEM;
1464 }
1465
1466 return 0;
1467 }
1468
smc_is_already_selected(struct smcd_dev * smcd,struct smc_init_info * ini,int matches)1469 static bool smc_is_already_selected(struct smcd_dev *smcd,
1470 struct smc_init_info *ini,
1471 int matches)
1472 {
1473 int i;
1474
1475 for (i = 0; i < matches; i++)
1476 if (smcd == ini->ism_dev[i])
1477 return true;
1478
1479 return false;
1480 }
1481
1482 /* check for ISM devices matching proposed ISM devices */
smc_check_ism_v2_match(struct smc_init_info * ini,u16 proposed_chid,u64 proposed_gid,unsigned int * matches)1483 static void smc_check_ism_v2_match(struct smc_init_info *ini,
1484 u16 proposed_chid, u64 proposed_gid,
1485 unsigned int *matches)
1486 {
1487 struct smcd_dev *smcd;
1488
1489 list_for_each_entry(smcd, &smcd_dev_list.list, list) {
1490 if (smcd->going_away)
1491 continue;
1492 if (smc_is_already_selected(smcd, ini, *matches))
1493 continue;
1494 if (smc_ism_get_chid(smcd) == proposed_chid &&
1495 !smc_ism_cantalk(proposed_gid, ISM_RESERVED_VLANID, smcd)) {
1496 ini->ism_peer_gid[*matches] = proposed_gid;
1497 ini->ism_dev[*matches] = smcd;
1498 (*matches)++;
1499 break;
1500 }
1501 }
1502 }
1503
smc_find_ism_v2_device_serv(struct smc_sock * new_smc,struct smc_clc_msg_proposal * pclc,struct smc_init_info * ini)1504 static void smc_find_ism_v2_device_serv(struct smc_sock *new_smc,
1505 struct smc_clc_msg_proposal *pclc,
1506 struct smc_init_info *ini)
1507 {
1508 struct smc_clc_smcd_v2_extension *smcd_v2_ext;
1509 struct smc_clc_v2_extension *smc_v2_ext;
1510 struct smc_clc_msg_smcd *pclc_smcd;
1511 unsigned int matches = 0;
1512 u8 smcd_version;
1513 u8 *eid = NULL;
1514 int i;
1515
1516 if (!(ini->smcd_version & SMC_V2) || !smcd_indicated(ini->smc_type_v2))
1517 goto not_found;
1518
1519 pclc_smcd = smc_get_clc_msg_smcd(pclc);
1520 smc_v2_ext = smc_get_clc_v2_ext(pclc);
1521 smcd_v2_ext = smc_get_clc_smcd_v2_ext(smc_v2_ext);
1522 if (!smcd_v2_ext ||
1523 !smc_v2_ext->hdr.flag.seid) /* no system EID support for SMCD */
1524 goto not_found;
1525
1526 mutex_lock(&smcd_dev_list.mutex);
1527 if (pclc_smcd->ism.chid)
1528 /* check for ISM device matching proposed native ISM device */
1529 smc_check_ism_v2_match(ini, ntohs(pclc_smcd->ism.chid),
1530 ntohll(pclc_smcd->ism.gid), &matches);
1531 for (i = 1; i <= smc_v2_ext->hdr.ism_gid_cnt; i++) {
1532 /* check for ISM devices matching proposed non-native ISM
1533 * devices
1534 */
1535 smc_check_ism_v2_match(ini,
1536 ntohs(smcd_v2_ext->gidchid[i - 1].chid),
1537 ntohll(smcd_v2_ext->gidchid[i - 1].gid),
1538 &matches);
1539 }
1540 mutex_unlock(&smcd_dev_list.mutex);
1541
1542 if (ini->ism_dev[0]) {
1543 smc_ism_get_system_eid(ini->ism_dev[0], &eid);
1544 if (memcmp(eid, smcd_v2_ext->system_eid, SMC_MAX_EID_LEN))
1545 goto not_found;
1546 } else {
1547 goto not_found;
1548 }
1549
1550 /* separate - outside the smcd_dev_list.lock */
1551 smcd_version = ini->smcd_version;
1552 for (i = 0; i < matches; i++) {
1553 ini->smcd_version = SMC_V2;
1554 ini->is_smcd = true;
1555 ini->ism_selected = i;
1556 if (smc_listen_ism_init(new_smc, ini))
1557 /* try next active ISM device */
1558 continue;
1559 return; /* matching and usable V2 ISM device found */
1560 }
1561 /* no V2 ISM device could be initialized */
1562 ini->smcd_version = smcd_version; /* restore original value */
1563
1564 not_found:
1565 ini->smcd_version &= ~SMC_V2;
1566 ini->ism_dev[0] = NULL;
1567 ini->is_smcd = false;
1568 }
1569
smc_find_ism_v1_device_serv(struct smc_sock * new_smc,struct smc_clc_msg_proposal * pclc,struct smc_init_info * ini)1570 static void smc_find_ism_v1_device_serv(struct smc_sock *new_smc,
1571 struct smc_clc_msg_proposal *pclc,
1572 struct smc_init_info *ini)
1573 {
1574 struct smc_clc_msg_smcd *pclc_smcd = smc_get_clc_msg_smcd(pclc);
1575
1576 /* check if ISM V1 is available */
1577 if (!(ini->smcd_version & SMC_V1) || !smcd_indicated(ini->smc_type_v1))
1578 goto not_found;
1579 ini->is_smcd = true; /* prepare ISM check */
1580 ini->ism_peer_gid[0] = ntohll(pclc_smcd->ism.gid);
1581 if (smc_find_ism_device(new_smc, ini))
1582 goto not_found;
1583 ini->ism_selected = 0;
1584 if (!smc_listen_ism_init(new_smc, ini))
1585 return; /* V1 ISM device found */
1586
1587 not_found:
1588 ini->ism_dev[0] = NULL;
1589 ini->is_smcd = false;
1590 }
1591
1592 /* listen worker: register buffers */
smc_listen_rdma_reg(struct smc_sock * new_smc,bool local_first)1593 static int smc_listen_rdma_reg(struct smc_sock *new_smc, bool local_first)
1594 {
1595 struct smc_connection *conn = &new_smc->conn;
1596
1597 if (!local_first) {
1598 if (smcr_lgr_reg_rmbs(conn->lnk, conn->rmb_desc))
1599 return SMC_CLC_DECL_ERR_REGRMB;
1600 }
1601 smc_rmb_sync_sg_for_device(&new_smc->conn);
1602
1603 return 0;
1604 }
1605
smc_find_rdma_v1_device_serv(struct smc_sock * new_smc,struct smc_clc_msg_proposal * pclc,struct smc_init_info * ini)1606 static int smc_find_rdma_v1_device_serv(struct smc_sock *new_smc,
1607 struct smc_clc_msg_proposal *pclc,
1608 struct smc_init_info *ini)
1609 {
1610 int rc;
1611
1612 if (!smcr_indicated(ini->smc_type_v1))
1613 return SMC_CLC_DECL_NOSMCDEV;
1614
1615 /* prepare RDMA check */
1616 ini->ib_lcl = &pclc->lcl;
1617 rc = smc_find_rdma_device(new_smc, ini);
1618 if (rc) {
1619 /* no RDMA device found */
1620 if (ini->smc_type_v1 == SMC_TYPE_B)
1621 /* neither ISM nor RDMA device found */
1622 rc = SMC_CLC_DECL_NOSMCDEV;
1623 return rc;
1624 }
1625 rc = smc_listen_rdma_init(new_smc, ini);
1626 if (rc)
1627 return rc;
1628 return smc_listen_rdma_reg(new_smc, ini->first_contact_local);
1629 }
1630
1631 /* determine the local device matching to proposal */
smc_listen_find_device(struct smc_sock * new_smc,struct smc_clc_msg_proposal * pclc,struct smc_init_info * ini)1632 static int smc_listen_find_device(struct smc_sock *new_smc,
1633 struct smc_clc_msg_proposal *pclc,
1634 struct smc_init_info *ini)
1635 {
1636 int rc;
1637
1638 /* check for ISM device matching V2 proposed device */
1639 smc_find_ism_v2_device_serv(new_smc, pclc, ini);
1640 if (ini->ism_dev[0])
1641 return 0;
1642
1643 if (!(ini->smcd_version & SMC_V1))
1644 return SMC_CLC_DECL_NOSMCDEV;
1645
1646 /* check for matching IP prefix and subnet length */
1647 rc = smc_listen_prfx_check(new_smc, pclc);
1648 if (rc)
1649 return rc;
1650
1651 /* get vlan id from IP device */
1652 if (smc_vlan_by_tcpsk(new_smc->clcsock, ini))
1653 return SMC_CLC_DECL_GETVLANERR;
1654
1655 /* check for ISM device matching V1 proposed device */
1656 smc_find_ism_v1_device_serv(new_smc, pclc, ini);
1657 if (ini->ism_dev[0])
1658 return 0;
1659
1660 if (pclc->hdr.typev1 == SMC_TYPE_D)
1661 return SMC_CLC_DECL_NOSMCDDEV; /* skip RDMA and decline */
1662
1663 /* check if RDMA is available */
1664 return smc_find_rdma_v1_device_serv(new_smc, pclc, ini);
1665 }
1666
1667 /* listen worker: finish RDMA setup */
smc_listen_rdma_finish(struct smc_sock * new_smc,struct smc_clc_msg_accept_confirm * cclc,bool local_first)1668 static int smc_listen_rdma_finish(struct smc_sock *new_smc,
1669 struct smc_clc_msg_accept_confirm *cclc,
1670 bool local_first)
1671 {
1672 struct smc_link *link = new_smc->conn.lnk;
1673 int reason_code = 0;
1674
1675 if (local_first)
1676 smc_link_save_peer_info(link, cclc);
1677
1678 if (smc_rmb_rtoken_handling(&new_smc->conn, link, cclc))
1679 return SMC_CLC_DECL_ERR_RTOK;
1680
1681 if (local_first) {
1682 if (smc_ib_ready_link(link))
1683 return SMC_CLC_DECL_ERR_RDYLNK;
1684 /* QP confirmation over RoCE fabric */
1685 smc_llc_flow_initiate(link->lgr, SMC_LLC_FLOW_ADD_LINK);
1686 reason_code = smcr_serv_conf_first_link(new_smc);
1687 smc_llc_flow_stop(link->lgr, &link->lgr->llc_flow_lcl);
1688 }
1689 return reason_code;
1690 }
1691
1692 /* setup for connection of server */
smc_listen_work(struct work_struct * work)1693 static void smc_listen_work(struct work_struct *work)
1694 {
1695 struct smc_sock *new_smc = container_of(work, struct smc_sock,
1696 smc_listen_work);
1697 u8 version = smc_ism_v2_capable ? SMC_V2 : SMC_V1;
1698 struct socket *newclcsock = new_smc->clcsock;
1699 struct smc_clc_msg_accept_confirm *cclc;
1700 struct smc_clc_msg_proposal_area *buf;
1701 struct smc_clc_msg_proposal *pclc;
1702 struct smc_init_info *ini = NULL;
1703 int rc = 0;
1704
1705 if (new_smc->listen_smc->sk.sk_state != SMC_LISTEN)
1706 return smc_listen_out_err(new_smc);
1707
1708 if (new_smc->use_fallback) {
1709 smc_listen_out_connected(new_smc);
1710 return;
1711 }
1712
1713 /* check if peer is smc capable */
1714 if (!tcp_sk(newclcsock->sk)->syn_smc) {
1715 smc_switch_to_fallback(new_smc);
1716 new_smc->fallback_rsn = SMC_CLC_DECL_PEERNOSMC;
1717 smc_listen_out_connected(new_smc);
1718 return;
1719 }
1720
1721 /* do inband token exchange -
1722 * wait for and receive SMC Proposal CLC message
1723 */
1724 buf = kzalloc(sizeof(*buf), GFP_KERNEL);
1725 if (!buf) {
1726 rc = SMC_CLC_DECL_MEM;
1727 goto out_decl;
1728 }
1729 pclc = (struct smc_clc_msg_proposal *)buf;
1730 rc = smc_clc_wait_msg(new_smc, pclc, sizeof(*buf),
1731 SMC_CLC_PROPOSAL, CLC_WAIT_TIME);
1732 if (rc)
1733 goto out_decl;
1734 version = pclc->hdr.version == SMC_V1 ? SMC_V1 : version;
1735
1736 /* IPSec connections opt out of SMC optimizations */
1737 if (using_ipsec(new_smc)) {
1738 rc = SMC_CLC_DECL_IPSEC;
1739 goto out_decl;
1740 }
1741
1742 ini = kzalloc(sizeof(*ini), GFP_KERNEL);
1743 if (!ini) {
1744 rc = SMC_CLC_DECL_MEM;
1745 goto out_decl;
1746 }
1747
1748 /* initial version checking */
1749 rc = smc_listen_v2_check(new_smc, pclc, ini);
1750 if (rc)
1751 goto out_decl;
1752
1753 mutex_lock(&smc_server_lgr_pending);
1754 smc_close_init(new_smc);
1755 smc_rx_init(new_smc);
1756 smc_tx_init(new_smc);
1757
1758 /* determine ISM or RoCE device used for connection */
1759 rc = smc_listen_find_device(new_smc, pclc, ini);
1760 if (rc)
1761 goto out_unlock;
1762
1763 /* send SMC Accept CLC message */
1764 rc = smc_clc_send_accept(new_smc, ini->first_contact_local,
1765 ini->smcd_version == SMC_V2 ? SMC_V2 : SMC_V1);
1766 if (rc)
1767 goto out_unlock;
1768
1769 /* SMC-D does not need this lock any more */
1770 if (ini->is_smcd)
1771 mutex_unlock(&smc_server_lgr_pending);
1772
1773 /* receive SMC Confirm CLC message */
1774 memset(buf, 0, sizeof(*buf));
1775 cclc = (struct smc_clc_msg_accept_confirm *)buf;
1776 rc = smc_clc_wait_msg(new_smc, cclc, sizeof(*buf),
1777 SMC_CLC_CONFIRM, CLC_WAIT_TIME);
1778 if (rc) {
1779 if (!ini->is_smcd)
1780 goto out_unlock;
1781 goto out_decl;
1782 }
1783
1784 /* finish worker */
1785 if (!ini->is_smcd) {
1786 rc = smc_listen_rdma_finish(new_smc, cclc,
1787 ini->first_contact_local);
1788 if (rc)
1789 goto out_unlock;
1790 mutex_unlock(&smc_server_lgr_pending);
1791 }
1792 smc_conn_save_peer_info(new_smc, cclc);
1793 smc_listen_out_connected(new_smc);
1794 goto out_free;
1795
1796 out_unlock:
1797 mutex_unlock(&smc_server_lgr_pending);
1798 out_decl:
1799 smc_listen_decline(new_smc, rc, ini ? ini->first_contact_local : 0,
1800 version);
1801 out_free:
1802 kfree(ini);
1803 kfree(buf);
1804 }
1805
smc_tcp_listen_work(struct work_struct * work)1806 static void smc_tcp_listen_work(struct work_struct *work)
1807 {
1808 struct smc_sock *lsmc = container_of(work, struct smc_sock,
1809 tcp_listen_work);
1810 struct sock *lsk = &lsmc->sk;
1811 struct smc_sock *new_smc;
1812 int rc = 0;
1813
1814 lock_sock(lsk);
1815 while (lsk->sk_state == SMC_LISTEN) {
1816 rc = smc_clcsock_accept(lsmc, &new_smc);
1817 if (rc) /* clcsock accept queue empty or error */
1818 goto out;
1819 if (!new_smc)
1820 continue;
1821
1822 new_smc->listen_smc = lsmc;
1823 new_smc->use_fallback = lsmc->use_fallback;
1824 new_smc->fallback_rsn = lsmc->fallback_rsn;
1825 sock_hold(lsk); /* sock_put in smc_listen_work */
1826 INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
1827 smc_copy_sock_settings_to_smc(new_smc);
1828 new_smc->sk.sk_sndbuf = lsmc->sk.sk_sndbuf;
1829 new_smc->sk.sk_rcvbuf = lsmc->sk.sk_rcvbuf;
1830 sock_hold(&new_smc->sk); /* sock_put in passive closing */
1831 if (!queue_work(smc_hs_wq, &new_smc->smc_listen_work))
1832 sock_put(&new_smc->sk);
1833 }
1834
1835 out:
1836 release_sock(lsk);
1837 sock_put(&lsmc->sk); /* sock_hold in smc_clcsock_data_ready() */
1838 }
1839
smc_clcsock_data_ready(struct sock * listen_clcsock)1840 static void smc_clcsock_data_ready(struct sock *listen_clcsock)
1841 {
1842 struct smc_sock *lsmc;
1843
1844 lsmc = (struct smc_sock *)
1845 ((uintptr_t)listen_clcsock->sk_user_data & ~SK_USER_DATA_NOCOPY);
1846 if (!lsmc)
1847 return;
1848 lsmc->clcsk_data_ready(listen_clcsock);
1849 if (lsmc->sk.sk_state == SMC_LISTEN) {
1850 sock_hold(&lsmc->sk); /* sock_put in smc_tcp_listen_work() */
1851 if (!queue_work(smc_hs_wq, &lsmc->tcp_listen_work))
1852 sock_put(&lsmc->sk);
1853 }
1854 }
1855
smc_listen(struct socket * sock,int backlog)1856 static int smc_listen(struct socket *sock, int backlog)
1857 {
1858 struct sock *sk = sock->sk;
1859 struct smc_sock *smc;
1860 int rc;
1861
1862 smc = smc_sk(sk);
1863 lock_sock(sk);
1864
1865 rc = -EINVAL;
1866 if ((sk->sk_state != SMC_INIT && sk->sk_state != SMC_LISTEN) ||
1867 smc->connect_nonblock)
1868 goto out;
1869
1870 rc = 0;
1871 if (sk->sk_state == SMC_LISTEN) {
1872 sk->sk_max_ack_backlog = backlog;
1873 goto out;
1874 }
1875 /* some socket options are handled in core, so we could not apply
1876 * them to the clc socket -- copy smc socket options to clc socket
1877 */
1878 smc_copy_sock_settings_to_clc(smc);
1879 if (!smc->use_fallback)
1880 tcp_sk(smc->clcsock->sk)->syn_smc = 1;
1881
1882 /* save original sk_data_ready function and establish
1883 * smc-specific sk_data_ready function
1884 */
1885 smc->clcsk_data_ready = smc->clcsock->sk->sk_data_ready;
1886 smc->clcsock->sk->sk_data_ready = smc_clcsock_data_ready;
1887 smc->clcsock->sk->sk_user_data =
1888 (void *)((uintptr_t)smc | SK_USER_DATA_NOCOPY);
1889 rc = kernel_listen(smc->clcsock, backlog);
1890 if (rc) {
1891 smc->clcsock->sk->sk_data_ready = smc->clcsk_data_ready;
1892 goto out;
1893 }
1894 sk->sk_max_ack_backlog = backlog;
1895 sk->sk_ack_backlog = 0;
1896 sk->sk_state = SMC_LISTEN;
1897
1898 out:
1899 release_sock(sk);
1900 return rc;
1901 }
1902
smc_accept(struct socket * sock,struct socket * new_sock,int flags,bool kern)1903 static int smc_accept(struct socket *sock, struct socket *new_sock,
1904 int flags, bool kern)
1905 {
1906 struct sock *sk = sock->sk, *nsk;
1907 DECLARE_WAITQUEUE(wait, current);
1908 struct smc_sock *lsmc;
1909 long timeo;
1910 int rc = 0;
1911
1912 lsmc = smc_sk(sk);
1913 sock_hold(sk); /* sock_put below */
1914 lock_sock(sk);
1915
1916 if (lsmc->sk.sk_state != SMC_LISTEN) {
1917 rc = -EINVAL;
1918 release_sock(sk);
1919 goto out;
1920 }
1921
1922 /* Wait for an incoming connection */
1923 timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
1924 add_wait_queue_exclusive(sk_sleep(sk), &wait);
1925 while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
1926 set_current_state(TASK_INTERRUPTIBLE);
1927 if (!timeo) {
1928 rc = -EAGAIN;
1929 break;
1930 }
1931 release_sock(sk);
1932 timeo = schedule_timeout(timeo);
1933 /* wakeup by sk_data_ready in smc_listen_work() */
1934 sched_annotate_sleep();
1935 lock_sock(sk);
1936 if (signal_pending(current)) {
1937 rc = sock_intr_errno(timeo);
1938 break;
1939 }
1940 }
1941 set_current_state(TASK_RUNNING);
1942 remove_wait_queue(sk_sleep(sk), &wait);
1943
1944 if (!rc)
1945 rc = sock_error(nsk);
1946 release_sock(sk);
1947 if (rc)
1948 goto out;
1949
1950 if (lsmc->sockopt_defer_accept && !(flags & O_NONBLOCK)) {
1951 /* wait till data arrives on the socket */
1952 timeo = msecs_to_jiffies(lsmc->sockopt_defer_accept *
1953 MSEC_PER_SEC);
1954 if (smc_sk(nsk)->use_fallback) {
1955 struct sock *clcsk = smc_sk(nsk)->clcsock->sk;
1956
1957 lock_sock(clcsk);
1958 if (skb_queue_empty(&clcsk->sk_receive_queue))
1959 sk_wait_data(clcsk, &timeo, NULL);
1960 release_sock(clcsk);
1961 } else if (!atomic_read(&smc_sk(nsk)->conn.bytes_to_rcv)) {
1962 lock_sock(nsk);
1963 smc_rx_wait(smc_sk(nsk), &timeo, smc_rx_data_available);
1964 release_sock(nsk);
1965 }
1966 }
1967
1968 out:
1969 sock_put(sk); /* sock_hold above */
1970 return rc;
1971 }
1972
smc_getname(struct socket * sock,struct sockaddr * addr,int peer)1973 static int smc_getname(struct socket *sock, struct sockaddr *addr,
1974 int peer)
1975 {
1976 struct smc_sock *smc;
1977
1978 if (peer && (sock->sk->sk_state != SMC_ACTIVE) &&
1979 (sock->sk->sk_state != SMC_APPCLOSEWAIT1))
1980 return -ENOTCONN;
1981
1982 smc = smc_sk(sock->sk);
1983
1984 return smc->clcsock->ops->getname(smc->clcsock, addr, peer);
1985 }
1986
smc_sendmsg(struct socket * sock,struct msghdr * msg,size_t len)1987 static int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
1988 {
1989 struct sock *sk = sock->sk;
1990 struct smc_sock *smc;
1991 int rc = -EPIPE;
1992
1993 smc = smc_sk(sk);
1994 lock_sock(sk);
1995 if ((sk->sk_state != SMC_ACTIVE) &&
1996 (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1997 (sk->sk_state != SMC_INIT))
1998 goto out;
1999
2000 if (msg->msg_flags & MSG_FASTOPEN) {
2001 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
2002 smc_switch_to_fallback(smc);
2003 smc->fallback_rsn = SMC_CLC_DECL_OPTUNSUPP;
2004 } else {
2005 rc = -EINVAL;
2006 goto out;
2007 }
2008 }
2009
2010 if (smc->use_fallback)
2011 rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
2012 else
2013 rc = smc_tx_sendmsg(smc, msg, len);
2014 out:
2015 release_sock(sk);
2016 return rc;
2017 }
2018
smc_recvmsg(struct socket * sock,struct msghdr * msg,size_t len,int flags)2019 static int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
2020 int flags)
2021 {
2022 struct sock *sk = sock->sk;
2023 struct smc_sock *smc;
2024 int rc = -ENOTCONN;
2025
2026 smc = smc_sk(sk);
2027 lock_sock(sk);
2028 if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
2029 /* socket was connected before, no more data to read */
2030 rc = 0;
2031 goto out;
2032 }
2033 if ((sk->sk_state == SMC_INIT) ||
2034 (sk->sk_state == SMC_LISTEN) ||
2035 (sk->sk_state == SMC_CLOSED))
2036 goto out;
2037
2038 if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
2039 rc = 0;
2040 goto out;
2041 }
2042
2043 if (smc->use_fallback) {
2044 rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
2045 } else {
2046 msg->msg_namelen = 0;
2047 rc = smc_rx_recvmsg(smc, msg, NULL, len, flags);
2048 }
2049
2050 out:
2051 release_sock(sk);
2052 return rc;
2053 }
2054
smc_accept_poll(struct sock * parent)2055 static __poll_t smc_accept_poll(struct sock *parent)
2056 {
2057 struct smc_sock *isk = smc_sk(parent);
2058 __poll_t mask = 0;
2059
2060 spin_lock(&isk->accept_q_lock);
2061 if (!list_empty(&isk->accept_q))
2062 mask = EPOLLIN | EPOLLRDNORM;
2063 spin_unlock(&isk->accept_q_lock);
2064
2065 return mask;
2066 }
2067
smc_poll(struct file * file,struct socket * sock,poll_table * wait)2068 static __poll_t smc_poll(struct file *file, struct socket *sock,
2069 poll_table *wait)
2070 {
2071 struct sock *sk = sock->sk;
2072 struct smc_sock *smc;
2073 __poll_t mask = 0;
2074
2075 if (!sk)
2076 return EPOLLNVAL;
2077
2078 smc = smc_sk(sock->sk);
2079 if (smc->use_fallback) {
2080 /* delegate to CLC child sock */
2081 mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
2082 sk->sk_err = smc->clcsock->sk->sk_err;
2083 } else {
2084 if (sk->sk_state != SMC_CLOSED)
2085 sock_poll_wait(file, sock, wait);
2086 if (sk->sk_err)
2087 mask |= EPOLLERR;
2088 if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
2089 (sk->sk_state == SMC_CLOSED))
2090 mask |= EPOLLHUP;
2091 if (sk->sk_state == SMC_LISTEN) {
2092 /* woken up by sk_data_ready in smc_listen_work() */
2093 mask |= smc_accept_poll(sk);
2094 } else if (smc->use_fallback) { /* as result of connect_work()*/
2095 mask |= smc->clcsock->ops->poll(file, smc->clcsock,
2096 wait);
2097 sk->sk_err = smc->clcsock->sk->sk_err;
2098 } else {
2099 if ((sk->sk_state != SMC_INIT &&
2100 atomic_read(&smc->conn.sndbuf_space)) ||
2101 sk->sk_shutdown & SEND_SHUTDOWN) {
2102 mask |= EPOLLOUT | EPOLLWRNORM;
2103 } else {
2104 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
2105 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
2106 }
2107 if (atomic_read(&smc->conn.bytes_to_rcv))
2108 mask |= EPOLLIN | EPOLLRDNORM;
2109 if (sk->sk_shutdown & RCV_SHUTDOWN)
2110 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
2111 if (sk->sk_state == SMC_APPCLOSEWAIT1)
2112 mask |= EPOLLIN;
2113 if (smc->conn.urg_state == SMC_URG_VALID)
2114 mask |= EPOLLPRI;
2115 }
2116 }
2117
2118 return mask;
2119 }
2120
smc_shutdown(struct socket * sock,int how)2121 static int smc_shutdown(struct socket *sock, int how)
2122 {
2123 struct sock *sk = sock->sk;
2124 bool do_shutdown = true;
2125 struct smc_sock *smc;
2126 int rc = -EINVAL;
2127 int old_state;
2128 int rc1 = 0;
2129
2130 smc = smc_sk(sk);
2131
2132 if ((how < SHUT_RD) || (how > SHUT_RDWR))
2133 return rc;
2134
2135 lock_sock(sk);
2136
2137 rc = -ENOTCONN;
2138 if ((sk->sk_state != SMC_ACTIVE) &&
2139 (sk->sk_state != SMC_PEERCLOSEWAIT1) &&
2140 (sk->sk_state != SMC_PEERCLOSEWAIT2) &&
2141 (sk->sk_state != SMC_APPCLOSEWAIT1) &&
2142 (sk->sk_state != SMC_APPCLOSEWAIT2) &&
2143 (sk->sk_state != SMC_APPFINCLOSEWAIT))
2144 goto out;
2145 if (smc->use_fallback) {
2146 rc = kernel_sock_shutdown(smc->clcsock, how);
2147 sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
2148 if (sk->sk_shutdown == SHUTDOWN_MASK) {
2149 sk->sk_state = SMC_CLOSED;
2150 sock_put(sk);
2151 }
2152 goto out;
2153 }
2154 switch (how) {
2155 case SHUT_RDWR: /* shutdown in both directions */
2156 old_state = sk->sk_state;
2157 rc = smc_close_active(smc);
2158 if (old_state == SMC_ACTIVE &&
2159 sk->sk_state == SMC_PEERCLOSEWAIT1)
2160 do_shutdown = false;
2161 break;
2162 case SHUT_WR:
2163 rc = smc_close_shutdown_write(smc);
2164 break;
2165 case SHUT_RD:
2166 rc = 0;
2167 /* nothing more to do because peer is not involved */
2168 break;
2169 }
2170 if (do_shutdown && smc->clcsock)
2171 rc1 = kernel_sock_shutdown(smc->clcsock, how);
2172 /* map sock_shutdown_cmd constants to sk_shutdown value range */
2173 sk->sk_shutdown |= how + 1;
2174
2175 out:
2176 release_sock(sk);
2177 return rc ? rc : rc1;
2178 }
2179
smc_setsockopt(struct socket * sock,int level,int optname,sockptr_t optval,unsigned int optlen)2180 static int smc_setsockopt(struct socket *sock, int level, int optname,
2181 sockptr_t optval, unsigned int optlen)
2182 {
2183 struct sock *sk = sock->sk;
2184 struct smc_sock *smc;
2185 int val, rc;
2186
2187 if (level == SOL_TCP && optname == TCP_ULP)
2188 return -EOPNOTSUPP;
2189
2190 smc = smc_sk(sk);
2191
2192 /* generic setsockopts reaching us here always apply to the
2193 * CLC socket
2194 */
2195 if (unlikely(!smc->clcsock->ops->setsockopt))
2196 rc = -EOPNOTSUPP;
2197 else
2198 rc = smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
2199 optval, optlen);
2200 if (smc->clcsock->sk->sk_err) {
2201 sk->sk_err = smc->clcsock->sk->sk_err;
2202 sk->sk_error_report(sk);
2203 }
2204
2205 if (optlen < sizeof(int))
2206 return -EINVAL;
2207 if (copy_from_sockptr(&val, optval, sizeof(int)))
2208 return -EFAULT;
2209
2210 lock_sock(sk);
2211 if (rc || smc->use_fallback)
2212 goto out;
2213 switch (optname) {
2214 case TCP_FASTOPEN:
2215 case TCP_FASTOPEN_CONNECT:
2216 case TCP_FASTOPEN_KEY:
2217 case TCP_FASTOPEN_NO_COOKIE:
2218 /* option not supported by SMC */
2219 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
2220 smc_switch_to_fallback(smc);
2221 smc->fallback_rsn = SMC_CLC_DECL_OPTUNSUPP;
2222 } else {
2223 rc = -EINVAL;
2224 }
2225 break;
2226 case TCP_NODELAY:
2227 if (sk->sk_state != SMC_INIT &&
2228 sk->sk_state != SMC_LISTEN &&
2229 sk->sk_state != SMC_CLOSED) {
2230 if (val)
2231 mod_delayed_work(smc->conn.lgr->tx_wq,
2232 &smc->conn.tx_work, 0);
2233 }
2234 break;
2235 case TCP_CORK:
2236 if (sk->sk_state != SMC_INIT &&
2237 sk->sk_state != SMC_LISTEN &&
2238 sk->sk_state != SMC_CLOSED) {
2239 if (!val)
2240 mod_delayed_work(smc->conn.lgr->tx_wq,
2241 &smc->conn.tx_work, 0);
2242 }
2243 break;
2244 case TCP_DEFER_ACCEPT:
2245 smc->sockopt_defer_accept = val;
2246 break;
2247 default:
2248 break;
2249 }
2250 out:
2251 release_sock(sk);
2252
2253 return rc;
2254 }
2255
smc_getsockopt(struct socket * sock,int level,int optname,char __user * optval,int __user * optlen)2256 static int smc_getsockopt(struct socket *sock, int level, int optname,
2257 char __user *optval, int __user *optlen)
2258 {
2259 struct smc_sock *smc;
2260
2261 smc = smc_sk(sock->sk);
2262 /* socket options apply to the CLC socket */
2263 if (unlikely(!smc->clcsock->ops->getsockopt))
2264 return -EOPNOTSUPP;
2265 return smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
2266 optval, optlen);
2267 }
2268
smc_ioctl(struct socket * sock,unsigned int cmd,unsigned long arg)2269 static int smc_ioctl(struct socket *sock, unsigned int cmd,
2270 unsigned long arg)
2271 {
2272 union smc_host_cursor cons, urg;
2273 struct smc_connection *conn;
2274 struct smc_sock *smc;
2275 int answ;
2276
2277 smc = smc_sk(sock->sk);
2278 conn = &smc->conn;
2279 lock_sock(&smc->sk);
2280 if (smc->use_fallback) {
2281 if (!smc->clcsock) {
2282 release_sock(&smc->sk);
2283 return -EBADF;
2284 }
2285 answ = smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
2286 release_sock(&smc->sk);
2287 return answ;
2288 }
2289 switch (cmd) {
2290 case SIOCINQ: /* same as FIONREAD */
2291 if (smc->sk.sk_state == SMC_LISTEN) {
2292 release_sock(&smc->sk);
2293 return -EINVAL;
2294 }
2295 if (smc->sk.sk_state == SMC_INIT ||
2296 smc->sk.sk_state == SMC_CLOSED)
2297 answ = 0;
2298 else
2299 answ = atomic_read(&smc->conn.bytes_to_rcv);
2300 break;
2301 case SIOCOUTQ:
2302 /* output queue size (not send + not acked) */
2303 if (smc->sk.sk_state == SMC_LISTEN) {
2304 release_sock(&smc->sk);
2305 return -EINVAL;
2306 }
2307 if (smc->sk.sk_state == SMC_INIT ||
2308 smc->sk.sk_state == SMC_CLOSED)
2309 answ = 0;
2310 else
2311 answ = smc->conn.sndbuf_desc->len -
2312 atomic_read(&smc->conn.sndbuf_space);
2313 break;
2314 case SIOCOUTQNSD:
2315 /* output queue size (not send only) */
2316 if (smc->sk.sk_state == SMC_LISTEN) {
2317 release_sock(&smc->sk);
2318 return -EINVAL;
2319 }
2320 if (smc->sk.sk_state == SMC_INIT ||
2321 smc->sk.sk_state == SMC_CLOSED)
2322 answ = 0;
2323 else
2324 answ = smc_tx_prepared_sends(&smc->conn);
2325 break;
2326 case SIOCATMARK:
2327 if (smc->sk.sk_state == SMC_LISTEN) {
2328 release_sock(&smc->sk);
2329 return -EINVAL;
2330 }
2331 if (smc->sk.sk_state == SMC_INIT ||
2332 smc->sk.sk_state == SMC_CLOSED) {
2333 answ = 0;
2334 } else {
2335 smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
2336 smc_curs_copy(&urg, &conn->urg_curs, conn);
2337 answ = smc_curs_diff(conn->rmb_desc->len,
2338 &cons, &urg) == 1;
2339 }
2340 break;
2341 default:
2342 release_sock(&smc->sk);
2343 return -ENOIOCTLCMD;
2344 }
2345 release_sock(&smc->sk);
2346
2347 return put_user(answ, (int __user *)arg);
2348 }
2349
smc_sendpage(struct socket * sock,struct page * page,int offset,size_t size,int flags)2350 static ssize_t smc_sendpage(struct socket *sock, struct page *page,
2351 int offset, size_t size, int flags)
2352 {
2353 struct sock *sk = sock->sk;
2354 struct smc_sock *smc;
2355 int rc = -EPIPE;
2356
2357 smc = smc_sk(sk);
2358 lock_sock(sk);
2359 if (sk->sk_state != SMC_ACTIVE) {
2360 release_sock(sk);
2361 goto out;
2362 }
2363 release_sock(sk);
2364 if (smc->use_fallback)
2365 rc = kernel_sendpage(smc->clcsock, page, offset,
2366 size, flags);
2367 else
2368 rc = sock_no_sendpage(sock, page, offset, size, flags);
2369
2370 out:
2371 return rc;
2372 }
2373
2374 /* Map the affected portions of the rmbe into an spd, note the number of bytes
2375 * to splice in conn->splice_pending, and press 'go'. Delays consumer cursor
2376 * updates till whenever a respective page has been fully processed.
2377 * Note that subsequent recv() calls have to wait till all splice() processing
2378 * completed.
2379 */
smc_splice_read(struct socket * sock,loff_t * ppos,struct pipe_inode_info * pipe,size_t len,unsigned int flags)2380 static ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
2381 struct pipe_inode_info *pipe, size_t len,
2382 unsigned int flags)
2383 {
2384 struct sock *sk = sock->sk;
2385 struct smc_sock *smc;
2386 int rc = -ENOTCONN;
2387
2388 smc = smc_sk(sk);
2389 lock_sock(sk);
2390 if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
2391 /* socket was connected before, no more data to read */
2392 rc = 0;
2393 goto out;
2394 }
2395 if (sk->sk_state == SMC_INIT ||
2396 sk->sk_state == SMC_LISTEN ||
2397 sk->sk_state == SMC_CLOSED)
2398 goto out;
2399
2400 if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
2401 rc = 0;
2402 goto out;
2403 }
2404
2405 if (smc->use_fallback) {
2406 rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
2407 pipe, len, flags);
2408 } else {
2409 if (*ppos) {
2410 rc = -ESPIPE;
2411 goto out;
2412 }
2413 if (flags & SPLICE_F_NONBLOCK)
2414 flags = MSG_DONTWAIT;
2415 else
2416 flags = 0;
2417 rc = smc_rx_recvmsg(smc, NULL, pipe, len, flags);
2418 }
2419 out:
2420 release_sock(sk);
2421
2422 return rc;
2423 }
2424
2425 /* must look like tcp */
2426 static const struct proto_ops smc_sock_ops = {
2427 .family = PF_SMC,
2428 .owner = THIS_MODULE,
2429 .release = smc_release,
2430 .bind = smc_bind,
2431 .connect = smc_connect,
2432 .socketpair = sock_no_socketpair,
2433 .accept = smc_accept,
2434 .getname = smc_getname,
2435 .poll = smc_poll,
2436 .ioctl = smc_ioctl,
2437 .listen = smc_listen,
2438 .shutdown = smc_shutdown,
2439 .setsockopt = smc_setsockopt,
2440 .getsockopt = smc_getsockopt,
2441 .sendmsg = smc_sendmsg,
2442 .recvmsg = smc_recvmsg,
2443 .mmap = sock_no_mmap,
2444 .sendpage = smc_sendpage,
2445 .splice_read = smc_splice_read,
2446 };
2447
smc_create(struct net * net,struct socket * sock,int protocol,int kern)2448 static int smc_create(struct net *net, struct socket *sock, int protocol,
2449 int kern)
2450 {
2451 int family = (protocol == SMCPROTO_SMC6) ? PF_INET6 : PF_INET;
2452 struct smc_sock *smc;
2453 struct sock *sk;
2454 int rc;
2455
2456 rc = -ESOCKTNOSUPPORT;
2457 if (sock->type != SOCK_STREAM)
2458 goto out;
2459
2460 rc = -EPROTONOSUPPORT;
2461 if (protocol != SMCPROTO_SMC && protocol != SMCPROTO_SMC6)
2462 goto out;
2463
2464 rc = -ENOBUFS;
2465 sock->ops = &smc_sock_ops;
2466 sk = smc_sock_alloc(net, sock, protocol);
2467 if (!sk)
2468 goto out;
2469
2470 /* create internal TCP socket for CLC handshake and fallback */
2471 smc = smc_sk(sk);
2472 smc->use_fallback = false; /* assume rdma capability first */
2473 smc->fallback_rsn = 0;
2474 rc = sock_create_kern(net, family, SOCK_STREAM, IPPROTO_TCP,
2475 &smc->clcsock);
2476 if (rc) {
2477 sk_common_release(sk);
2478 goto out;
2479 }
2480 smc->sk.sk_sndbuf = max(smc->clcsock->sk->sk_sndbuf, SMC_BUF_MIN_SIZE);
2481 smc->sk.sk_rcvbuf = max(smc->clcsock->sk->sk_rcvbuf, SMC_BUF_MIN_SIZE);
2482
2483 out:
2484 return rc;
2485 }
2486
2487 static const struct net_proto_family smc_sock_family_ops = {
2488 .family = PF_SMC,
2489 .owner = THIS_MODULE,
2490 .create = smc_create,
2491 };
2492
2493 unsigned int smc_net_id;
2494
smc_net_init(struct net * net)2495 static __net_init int smc_net_init(struct net *net)
2496 {
2497 return smc_pnet_net_init(net);
2498 }
2499
smc_net_exit(struct net * net)2500 static void __net_exit smc_net_exit(struct net *net)
2501 {
2502 smc_pnet_net_exit(net);
2503 }
2504
2505 static struct pernet_operations smc_net_ops = {
2506 .init = smc_net_init,
2507 .exit = smc_net_exit,
2508 .id = &smc_net_id,
2509 .size = sizeof(struct smc_net),
2510 };
2511
smc_init(void)2512 static int __init smc_init(void)
2513 {
2514 int rc;
2515
2516 rc = register_pernet_subsys(&smc_net_ops);
2517 if (rc)
2518 return rc;
2519
2520 smc_ism_init();
2521 smc_clc_init();
2522
2523 rc = smc_pnet_init();
2524 if (rc)
2525 goto out_pernet_subsys;
2526
2527 rc = -ENOMEM;
2528 smc_hs_wq = alloc_workqueue("smc_hs_wq", 0, 0);
2529 if (!smc_hs_wq)
2530 goto out_pnet;
2531
2532 smc_close_wq = alloc_workqueue("smc_close_wq", 0, 0);
2533 if (!smc_close_wq)
2534 goto out_alloc_hs_wq;
2535
2536 rc = smc_core_init();
2537 if (rc) {
2538 pr_err("%s: smc_core_init fails with %d\n", __func__, rc);
2539 goto out_alloc_wqs;
2540 }
2541
2542 rc = smc_llc_init();
2543 if (rc) {
2544 pr_err("%s: smc_llc_init fails with %d\n", __func__, rc);
2545 goto out_core;
2546 }
2547
2548 rc = smc_cdc_init();
2549 if (rc) {
2550 pr_err("%s: smc_cdc_init fails with %d\n", __func__, rc);
2551 goto out_core;
2552 }
2553
2554 rc = proto_register(&smc_proto, 1);
2555 if (rc) {
2556 pr_err("%s: proto_register(v4) fails with %d\n", __func__, rc);
2557 goto out_core;
2558 }
2559
2560 rc = proto_register(&smc_proto6, 1);
2561 if (rc) {
2562 pr_err("%s: proto_register(v6) fails with %d\n", __func__, rc);
2563 goto out_proto;
2564 }
2565
2566 rc = sock_register(&smc_sock_family_ops);
2567 if (rc) {
2568 pr_err("%s: sock_register fails with %d\n", __func__, rc);
2569 goto out_proto6;
2570 }
2571 INIT_HLIST_HEAD(&smc_v4_hashinfo.ht);
2572 INIT_HLIST_HEAD(&smc_v6_hashinfo.ht);
2573
2574 rc = smc_ib_register_client();
2575 if (rc) {
2576 pr_err("%s: ib_register fails with %d\n", __func__, rc);
2577 goto out_sock;
2578 }
2579
2580 static_branch_enable(&tcp_have_smc);
2581 return 0;
2582
2583 out_sock:
2584 sock_unregister(PF_SMC);
2585 out_proto6:
2586 proto_unregister(&smc_proto6);
2587 out_proto:
2588 proto_unregister(&smc_proto);
2589 out_core:
2590 smc_core_exit();
2591 out_alloc_wqs:
2592 destroy_workqueue(smc_close_wq);
2593 out_alloc_hs_wq:
2594 destroy_workqueue(smc_hs_wq);
2595 out_pnet:
2596 smc_pnet_exit();
2597 out_pernet_subsys:
2598 unregister_pernet_subsys(&smc_net_ops);
2599
2600 return rc;
2601 }
2602
smc_exit(void)2603 static void __exit smc_exit(void)
2604 {
2605 static_branch_disable(&tcp_have_smc);
2606 sock_unregister(PF_SMC);
2607 smc_core_exit();
2608 smc_ib_unregister_client();
2609 destroy_workqueue(smc_close_wq);
2610 destroy_workqueue(smc_hs_wq);
2611 proto_unregister(&smc_proto6);
2612 proto_unregister(&smc_proto);
2613 smc_pnet_exit();
2614 unregister_pernet_subsys(&smc_net_ops);
2615 rcu_barrier();
2616 }
2617
2618 module_init(smc_init);
2619 module_exit(smc_exit);
2620
2621 MODULE_AUTHOR("Ursula Braun <ubraun@linux.vnet.ibm.com>");
2622 MODULE_DESCRIPTION("smc socket address family");
2623 MODULE_LICENSE("GPL");
2624 MODULE_ALIAS_NETPROTO(PF_SMC);
2625