]> git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blame - net/smc/af_smc.c
projects / mirror_ubuntu-jammy-kernel.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
net/smc: introduce CLC first contact extension
[mirror_ubuntu-jammy-kernel.git] / net / smc / af_smc.c
CommitLineData
09c434b8 1// SPDX-License-Identifier: GPL-2.0-only
ac713874
UB		 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 *
a046d57d 10 * Initial restrictions:
a046d57d 11 * - support for alternate links postponed
a046d57d 12 *
aaa4d33f 13 * Copyright IBM Corp. 2016, 2018
ac713874
UB		 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>
a046d57d 24#include <linux/workqueue.h>
5f08318f 25#include <linux/in.h>
c3edc401 26#include <linux/sched/signal.h>
41349844 27#include <linux/if_vlan.h>
4ead9c96 28#include <linux/rcupdate_wait.h>
b81a5eb7 29#include <linux/ctype.h>
c3edc401 30
ac713874 31#include <net/sock.h>
a046d57d 32#include <net/tcp.h>
f16a7dd5 33#include <net/smc.h>
9b67e26f 34#include <asm/ioctls.h>
ac713874 35
64e28b52
HW		 36#include <net/net_namespace.h>
37#include <net/netns/generic.h>
38#include "smc_netns.h"
39
ac713874 40#include "smc.h"
a046d57d 41#include "smc_clc.h"
9bf9abea 42#include "smc_llc.h"
5f08318f 43#include "smc_cdc.h"
0cfdd8f9 44#include "smc_core.h"
a4cf0443 45#include "smc_ib.h"
41349844 46#include "smc_ism.h"
6812baab 47#include "smc_pnet.h"
e6727f39 48#include "smc_tx.h"
952310cc 49#include "smc_rx.h"
b38d7324 50#include "smc_close.h"
ac713874 51
72a36a8a
HW		 52static DEFINE_MUTEX(smc_server_lgr_pending); /* serialize link group
53 * creation on server
54 */
55static DEFINE_MUTEX(smc_client_lgr_pending); /* serialize link group
56 * creation on client
0cfdd8f9
UB		 57 */
58
22ef473d
KG		 59struct workqueue_struct *smc_hs_wq; /* wq for handshake work */
60struct workqueue_struct *smc_close_wq; /* wq for close work */
61
a046d57d 62static void smc_tcp_listen_work(struct work_struct *);
24ac3a08 63static void smc_connect_work(struct work_struct *);
a046d57d 64
ac713874
UB		 65static 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
f16a7dd5
UB		 72static struct smc_hashinfo smc_v4_hashinfo = {
73 .lock = __RW_LOCK_UNLOCKED(smc_v4_hashinfo.lock),
74};
75
aaa4d33f
KG		 76static struct smc_hashinfo smc_v6_hashinfo = {
77 .lock = __RW_LOCK_UNLOCKED(smc_v6_hashinfo.lock),
78};
79
f16a7dd5
UB		 80int 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}
94EXPORT_SYMBOL_GPL(smc_hash_sk);
95
96void 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}
105EXPORT_SYMBOL_GPL(smc_unhash_sk);
106
107struct proto smc_proto = {
ac713874
UB		 108 .name = "SMC",
109 .owner = THIS_MODULE,
110 .keepalive = smc_set_keepalive,
f16a7dd5
UB		 111 .hash = smc_hash_sk,
112 .unhash = smc_unhash_sk,
ac713874 113 .obj_size = sizeof(struct smc_sock),
f16a7dd5 114 .h.smc_hash = &smc_v4_hashinfo,
5f0d5a3a 115 .slab_flags = SLAB_TYPESAFE_BY_RCU,
ac713874 116};
f16a7dd5 117EXPORT_SYMBOL_GPL(smc_proto);
ac713874 118
aaa4d33f
KG		 119struct 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};
129EXPORT_SYMBOL_GPL(smc_proto6);
130
f536dffc
UB		 131static void smc_restore_fallback_changes(struct smc_sock *smc)
132{
1ad24058
KG		 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 }
f536dffc
UB		 137}
138
39f41f36 139static int __smc_release(struct smc_sock *smc)
ac713874 140{
39f41f36 141 struct sock *sk = &smc->sk;
b38d7324 142 int rc = 0;
ac713874 143
51f1de79 144 if (!smc->use_fallback) {
b38d7324
UB		 145 rc = smc_close_active(smc);
146 sock_set_flag(sk, SOCK_DEAD);
147 sk->sk_shutdown |= SHUTDOWN_MASK;
b03faa1f
UB		 148 } else {
149 if (sk->sk_state != SMC_LISTEN && sk->sk_state != SMC_INIT)
150 sock_put(sk); /* passive closing */
151 if (sk->sk_state == SMC_LISTEN) {
78abe3d0
MJ		 152 /* wake up clcsock accept */
153 rc = kernel_sock_shutdown(smc->clcsock, SHUT_RDWR);
154 }
51f1de79
UB		 155 sk->sk_state = SMC_CLOSED;
156 sk->sk_state_change(sk);
f536dffc 157 smc_restore_fallback_changes(smc);
51f1de79 158 }
ac713874 159
b03faa1f
UB		 160 sk->sk_prot->unhash(sk);
161
162 if (sk->sk_state == SMC_CLOSED) {
163 if (smc->clcsock) {
fd57770d
KG		 164 release_sock(sk);
165 smc_clcsock_release(smc);
166 lock_sock(sk);
b03faa1f
UB		 167 }
168 if (!smc->use_fallback)
169 smc_conn_free(&smc->conn);
170 }
171
39f41f36
UB		 172 return rc;
173}
174
175static int smc_release(struct socket *sock)
176{
177 struct sock *sk = sock->sk;
178 struct smc_sock *smc;
179 int rc = 0;
180
181 if (!sk)
182 goto out;
183
81cf4f47 184 sock_hold(sk); /* sock_put below */
39f41f36
UB		 185 smc = smc_sk(sk);
186
187 /* cleanup for a dangling non-blocking connect */
188 if (smc->connect_nonblock && sk->sk_state == SMC_INIT)
189 tcp_abort(smc->clcsock->sk, ECONNABORTED);
190 flush_work(&smc->connect_work);
191
192 if (sk->sk_state == SMC_LISTEN)
193 /* smc_close_non_accepted() is called and acquires
194 * sock lock for child sockets again
195 */
196 lock_sock_nested(sk, SINGLE_DEPTH_NESTING);
197 else
198 lock_sock(sk);
199
200 rc = __smc_release(smc);
201
ac713874
UB		 202 /* detach socket */
203 sock_orphan(sk);
204 sock->sk = NULL;
205 release_sock(sk);
206
81cf4f47 207 sock_put(sk); /* sock_hold above */
51f1de79 208 sock_put(sk); /* final sock_put */
ac713874 209out:
b38d7324 210 return rc;
ac713874
UB		 211}
212
213static void smc_destruct(struct sock *sk)
214{
215 if (sk->sk_state != SMC_CLOSED)
216 return;
217 if (!sock_flag(sk, SOCK_DEAD))
218 return;
219
220 sk_refcnt_debug_dec(sk);
221}
222
aaa4d33f
KG		 223static struct sock *smc_sock_alloc(struct net *net, struct socket *sock,
224 int protocol)
ac713874
UB		 225{
226 struct smc_sock *smc;
aaa4d33f 227 struct proto *prot;
ac713874
UB		 228 struct sock *sk;
229
aaa4d33f
KG		 230 prot = (protocol == SMCPROTO_SMC6) ? &smc_proto6 : &smc_proto;
231 sk = sk_alloc(net, PF_SMC, GFP_KERNEL, prot, 0);
ac713874
UB		 232 if (!sk)
233 return NULL;
234
235 sock_init_data(sock, sk); /* sets sk_refcnt to 1 */
236 sk->sk_state = SMC_INIT;
237 sk->sk_destruct = smc_destruct;
aaa4d33f 238 sk->sk_protocol = protocol;
ac713874 239 smc = smc_sk(sk);
a046d57d 240 INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
24ac3a08 241 INIT_WORK(&smc->connect_work, smc_connect_work);
be7f3e59 242 INIT_DELAYED_WORK(&smc->conn.tx_work, smc_tx_work);
a046d57d
UB		 243 INIT_LIST_HEAD(&smc->accept_q);
244 spin_lock_init(&smc->accept_q_lock);
be7f3e59 245 spin_lock_init(&smc->conn.send_lock);
f16a7dd5 246 sk->sk_prot->hash(sk);
a046d57d 247 sk_refcnt_debug_inc(sk);
78abe3d0 248 mutex_init(&smc->clcsock_release_lock);
ac713874
UB		 249
250 return sk;
251}
252
253static int smc_bind(struct socket *sock, struct sockaddr *uaddr,
254 int addr_len)
255{
256 struct sockaddr_in *addr = (struct sockaddr_in *)uaddr;
257 struct sock *sk = sock->sk;
258 struct smc_sock *smc;
259 int rc;
260
261 smc = smc_sk(sk);
262
263 /* replicate tests from inet_bind(), to be safe wrt. future changes */
264 rc = -EINVAL;
265 if (addr_len < sizeof(struct sockaddr_in))
266 goto out;
267
268 rc = -EAFNOSUPPORT;
aaa4d33f
KG		 269 if (addr->sin_family != AF_INET &&
270 addr->sin_family != AF_INET6 &&
271 addr->sin_family != AF_UNSPEC)
272 goto out;
ac713874 273 /* accept AF_UNSPEC (mapped to AF_INET) only if s_addr is INADDR_ANY */
aaa4d33f
KG		 274 if (addr->sin_family == AF_UNSPEC &&
275 addr->sin_addr.s_addr != htonl(INADDR_ANY))
ac713874
UB		 276 goto out;
277
278 lock_sock(sk);
279
280 /* Check if socket is already active */
281 rc = -EINVAL;
cd206360 282 if (sk->sk_state != SMC_INIT || smc->connect_nonblock)
ac713874
UB		 283 goto out_rel;
284
285 smc->clcsock->sk->sk_reuse = sk->sk_reuse;
286 rc = kernel_bind(smc->clcsock, uaddr, addr_len);
287
288out_rel:
289 release_sock(sk);
290out:
291 return rc;
292}
293
294static void smc_copy_sock_settings(struct sock *nsk, struct sock *osk,
295 unsigned long mask)
296{
297 /* options we don't get control via setsockopt for */
298 nsk->sk_type = osk->sk_type;
299 nsk->sk_sndbuf = osk->sk_sndbuf;
300 nsk->sk_rcvbuf = osk->sk_rcvbuf;
301 nsk->sk_sndtimeo = osk->sk_sndtimeo;
302 nsk->sk_rcvtimeo = osk->sk_rcvtimeo;
303 nsk->sk_mark = osk->sk_mark;
304 nsk->sk_priority = osk->sk_priority;
305 nsk->sk_rcvlowat = osk->sk_rcvlowat;
306 nsk->sk_bound_dev_if = osk->sk_bound_dev_if;
307 nsk->sk_err = osk->sk_err;
308
309 nsk->sk_flags &= ~mask;
310 nsk->sk_flags |= osk->sk_flags & mask;
311}
312
313#define SK_FLAGS_SMC_TO_CLC ((1UL << SOCK_URGINLINE) | \
314 (1UL << SOCK_KEEPOPEN) | \
315 (1UL << SOCK_LINGER) | \
316 (1UL << SOCK_BROADCAST) | \
317 (1UL << SOCK_TIMESTAMP) | \
318 (1UL << SOCK_DBG) | \
319 (1UL << SOCK_RCVTSTAMP) | \
320 (1UL << SOCK_RCVTSTAMPNS) | \
321 (1UL << SOCK_LOCALROUTE) | \
322 (1UL << SOCK_TIMESTAMPING_RX_SOFTWARE) | \
323 (1UL << SOCK_RXQ_OVFL) | \
324 (1UL << SOCK_WIFI_STATUS) | \
325 (1UL << SOCK_NOFCS) | \
9718475e
DD		 326 (1UL << SOCK_FILTER_LOCKED) | \
327 (1UL << SOCK_TSTAMP_NEW))
ac713874
UB		 328/* copy only relevant settings and flags of SOL_SOCKET level from smc to
329 * clc socket (since smc is not called for these options from net/core)
330 */
331static void smc_copy_sock_settings_to_clc(struct smc_sock *smc)
332{
333 smc_copy_sock_settings(smc->clcsock->sk, &smc->sk, SK_FLAGS_SMC_TO_CLC);
334}
335
336#define SK_FLAGS_CLC_TO_SMC ((1UL << SOCK_URGINLINE) | \
337 (1UL << SOCK_KEEPOPEN) | \
338 (1UL << SOCK_LINGER) | \
339 (1UL << SOCK_DBG))
340/* copy only settings and flags relevant for smc from clc to smc socket */
341static void smc_copy_sock_settings_to_smc(struct smc_sock *smc)
342{
343 smc_copy_sock_settings(&smc->sk, smc->clcsock->sk, SK_FLAGS_CLC_TO_SMC);
344}
345
b9247544 346/* register the new rmb on all links */
7562a13d 347static int smcr_lgr_reg_rmbs(struct smc_link *link,
b9247544
KG		 348 struct smc_buf_desc *rmb_desc)
349{
7562a13d
KG		 350 struct smc_link_group *lgr = link->lgr;
351 int i, rc = 0;
b9247544 352
d5500667
KG		 353 rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY);
354 if (rc)
355 return rc;
356 /* protect against parallel smc_llc_cli_rkey_exchange() and
357 * parallel smcr_link_reg_rmb()
358 */
359 mutex_lock(&lgr->llc_conf_mutex);
b9247544 360 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
741a49a4 361 if (!smc_link_active(&lgr->lnk[i]))
b9247544 362 continue;
7562a13d 363 rc = smcr_link_reg_rmb(&lgr->lnk[i], rmb_desc);
b9247544 364 if (rc)
7562a13d 365 goto out;
44aa81ce 366 }
7562a13d
KG		 367
368 /* exchange confirm_rkey msg with peer */
369 rc = smc_llc_do_confirm_rkey(link, rmb_desc);
370 if (rc) {
371 rc = -EFAULT;
372 goto out;
373 }
374 rmb_desc->is_conf_rkey = true;
375out:
d5500667
KG		 376 mutex_unlock(&lgr->llc_conf_mutex);
377 smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl);
7562a13d 378 return rc;
e63a5f8c
KG		 379}
380
b9247544 381static int smcr_clnt_conf_first_link(struct smc_sock *smc)
9bf9abea 382{
387707fd 383 struct smc_link *link = smc->conn.lnk;
0fb0b02b 384 struct smc_llc_qentry *qentry;
9bf9abea
UB		 385 int rc;
386
9bf9abea 387 /* receive CONFIRM LINK request from server over RoCE fabric */
0fb0b02b
KG		 388 qentry = smc_llc_wait(link->lgr, NULL, SMC_LLC_WAIT_TIME,
389 SMC_LLC_CONFIRM_LINK);
390 if (!qentry) {
9bf9abea
UB		 391 struct smc_clc_msg_decline dclc;
392
393 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
2b59f58e 394 SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
9ed28556 395 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
9bf9abea 396 }
649758ff 397 smc_llc_save_peer_uid(qentry);
0fb0b02b
KG		 398 rc = smc_llc_eval_conf_link(qentry, SMC_LLC_REQ);
399 smc_llc_flow_qentry_del(&link->lgr->llc_flow_lcl);
400 if (rc)
75d320d6
KG		 401 return SMC_CLC_DECL_RMBE_EC;
402
9bf9abea
UB		 403 rc = smc_ib_modify_qp_rts(link);
404 if (rc)
603cc149 405 return SMC_CLC_DECL_ERR_RDYLNK;
9bf9abea
UB		 406
407 smc_wr_remember_qp_attr(link);
652a1e41 408
7562a13d 409 if (smcr_link_reg_rmb(link, smc->conn.rmb_desc))
603cc149 410 return SMC_CLC_DECL_ERR_REGRMB;
652a1e41 411
0fb0b02b
KG		 412 /* confirm_rkey is implicit on 1st contact */
413 smc->conn.rmb_desc->is_conf_rkey = true;
414
9bf9abea 415 /* send CONFIRM LINK response over RoCE fabric */
947541f3 416 rc = smc_llc_send_confirm_link(link, SMC_LLC_RESP);
9bf9abea 417 if (rc < 0)
603cc149 418 return SMC_CLC_DECL_TIMEOUT_CL;
9bf9abea 419
0fb0b02b 420 smc_llc_link_active(link);
0a99be43 421 smcr_lgr_set_type(link->lgr, SMC_LGR_SINGLE);
0fb0b02b
KG		 422
423 /* optional 2nd link, receive ADD LINK request from server */
424 qentry = smc_llc_wait(link->lgr, NULL, SMC_LLC_WAIT_TIME,
425 SMC_LLC_ADD_LINK);
426 if (!qentry) {
52bedf37
KG		 427 struct smc_clc_msg_decline dclc;
428
429 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
2b59f58e 430 SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
0fb0b02b
KG		 431 if (rc == -EAGAIN)
432 rc = 0; /* no DECLINE received, go with one link */
433 return rc;
52bedf37 434 }
0fb0b02b 435 smc_llc_flow_qentry_clr(&link->lgr->llc_flow_lcl);
b1570a87 436 smc_llc_cli_add_link(link, qentry);
75d320d6 437 return 0;
9bf9abea
UB		 438}
439
41349844
HW		 440static void smcr_conn_save_peer_info(struct smc_sock *smc,
441 struct smc_clc_msg_accept_confirm *clc)
0cfdd8f9 442{
3d9725a6 443 int bufsize = smc_uncompress_bufsize(clc->r0.rmbe_size);
95d8d263 444
3d9725a6
UB		 445 smc->conn.peer_rmbe_idx = clc->r0.rmbe_idx;
446 smc->conn.local_tx_ctrl.token = ntohl(clc->r0.rmbe_alert_token);
95d8d263 447 smc->conn.peer_rmbe_size = bufsize;
cd6851f3 448 atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
95d8d263 449 smc->conn.tx_off = bufsize * (smc->conn.peer_rmbe_idx - 1);
0cfdd8f9
UB		 450}
451
b81a5eb7
UB		 452static bool smc_isascii(char *hostname)
453{
454 int i;
455
456 for (i = 0; i < SMC_MAX_HOSTNAME_LEN; i++)
457 if (!isascii(hostname[i]))
458 return false;
459 return true;
460}
461
41349844
HW		 462static void smcd_conn_save_peer_info(struct smc_sock *smc,
463 struct smc_clc_msg_accept_confirm *clc)
464{
3d9725a6 465 int bufsize = smc_uncompress_bufsize(clc->d0.dmbe_size);
41349844 466
3d9725a6
UB		 467 smc->conn.peer_rmbe_idx = clc->d0.dmbe_idx;
468 smc->conn.peer_token = clc->d0.token;
41349844
HW		 469 /* msg header takes up space in the buffer */
470 smc->conn.peer_rmbe_size = bufsize - sizeof(struct smcd_cdc_msg);
471 atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
472 smc->conn.tx_off = bufsize * smc->conn.peer_rmbe_idx;
b81a5eb7
UB		 473 if (clc->hdr.version > SMC_V1 &&
474 (clc->hdr.typev2 & SMC_FIRST_CONTACT_MASK)) {
475 struct smc_clc_msg_accept_confirm_v2 *clc_v2 =
476 (struct smc_clc_msg_accept_confirm_v2 *)clc;
477 struct smc_clc_first_contact_ext *fce =
478 (struct smc_clc_first_contact_ext *)
479 (((u8 *)clc_v2) + sizeof(*clc_v2));
480
481 memcpy(smc->conn.lgr->negotiated_eid, clc_v2->eid,
482 SMC_MAX_EID_LEN);
483 smc->conn.lgr->peer_os = fce->os_type;
484 smc->conn.lgr->peer_smc_release = fce->release;
485 if (smc_isascii(fce->hostname))
486 memcpy(smc->conn.lgr->peer_hostname, fce->hostname,
487 SMC_MAX_HOSTNAME_LEN);
488 }
41349844
HW		 489}
490
491static void smc_conn_save_peer_info(struct smc_sock *smc,
492 struct smc_clc_msg_accept_confirm *clc)
493{
494 if (smc->conn.lgr->is_smcd)
495 smcd_conn_save_peer_info(smc, clc);
496 else
497 smcr_conn_save_peer_info(smc, clc);
498}
499
0cfdd8f9
UB		 500static void smc_link_save_peer_info(struct smc_link *link,
501 struct smc_clc_msg_accept_confirm *clc)
502{
3d9725a6
UB		 503 link->peer_qpn = ntoh24(clc->r0.qpn);
504 memcpy(link->peer_gid, clc->r0.lcl.gid, SMC_GID_SIZE);
505 memcpy(link->peer_mac, clc->r0.lcl.mac, sizeof(link->peer_mac));
506 link->peer_psn = ntoh24(clc->r0.psn);
507 link->peer_mtu = clc->r0.qp_mtu;
0cfdd8f9
UB		 508}
509
07603b23
UB		 510static void smc_switch_to_fallback(struct smc_sock *smc)
511{
512 smc->use_fallback = true;
513 if (smc->sk.sk_socket && smc->sk.sk_socket->file) {
514 smc->clcsock->file = smc->sk.sk_socket->file;
515 smc->clcsock->file->private_data = smc->clcsock;
67f562e3
UB		 516 smc->clcsock->wq.fasync_list =
517 smc->sk.sk_socket->wq.fasync_list;
07603b23
UB		 518 }
519}
520
3b2dec26 521/* fall back during connect */
603cc149 522static int smc_connect_fallback(struct smc_sock *smc, int reason_code)
a046d57d 523{
07603b23 524 smc_switch_to_fallback(smc);
603cc149 525 smc->fallback_rsn = reason_code;
3b2dec26 526 smc_copy_sock_settings_to_clc(smc);
50717a37 527 smc->connect_nonblock = 0;
3b2dec26
HW		 528 if (smc->sk.sk_state == SMC_INIT)
529 smc->sk.sk_state = SMC_ACTIVE;
530 return 0;
531}
51f1de79 532
3b2dec26
HW		 533/* decline and fall back during connect */
534static int smc_connect_decline_fallback(struct smc_sock *smc, int reason_code)
535{
536 int rc;
ee9dfbef 537
e1bbdd57
UB		 538 if (reason_code < 0) { /* error, fallback is not possible */
539 if (smc->sk.sk_state == SMC_INIT)
540 sock_put(&smc->sk); /* passive closing */
3b2dec26 541 return reason_code;
e1bbdd57 542 }
603cc149 543 if (reason_code != SMC_CLC_DECL_PEERDECL) {
3b2dec26 544 rc = smc_clc_send_decline(smc, reason_code);
e1bbdd57
UB		 545 if (rc < 0) {
546 if (smc->sk.sk_state == SMC_INIT)
547 sock_put(&smc->sk); /* passive closing */
3b2dec26 548 return rc;
e1bbdd57 549 }
c5c1cc9c 550 }
603cc149 551 return smc_connect_fallback(smc, reason_code);
3b2dec26 552}
c5c1cc9c 553
3b2dec26
HW		 554/* abort connecting */
555static int smc_connect_abort(struct smc_sock *smc, int reason_code,
5ac54d87 556 int local_first)
3b2dec26 557{
51e3dfa8
UB		 558 bool is_smcd = smc->conn.lgr->is_smcd;
559
5ac54d87 560 if (local_first)
51e3dfa8
UB		 561 smc_lgr_cleanup_early(&smc->conn);
562 else
563 smc_conn_free(&smc->conn);
564 if (is_smcd)
72a36a8a
HW		 565 /* there is only one lgr role for SMC-D; use server lock */
566 mutex_unlock(&smc_server_lgr_pending);
567 else
568 mutex_unlock(&smc_client_lgr_pending);
569
50717a37 570 smc->connect_nonblock = 0;
3b2dec26
HW		 571 return reason_code;
572}
573
574/* check if there is a rdma device available for this connection. */
575/* called for connect and listen */
228bae05 576static int smc_find_rdma_device(struct smc_sock *smc, struct smc_init_info *ini)
3b2dec26 577{
a046d57d
UB		 578 /* PNET table look up: search active ib_device and port
579 * within same PNETID that also contains the ethernet device
580 * used for the internal TCP socket
581 */
bc36d2fc 582 smc_pnet_find_roce_resource(smc->clcsock->sk, ini);
9aa68d29
KG		 583 if (!ini->ib_dev)
584 return SMC_CLC_DECL_NOSMCRDEV;
bc36d2fc 585 return 0;
3b2dec26
HW		 586}
587
41349844
HW		 588/* check if there is an ISM device available for this connection. */
589/* called for connect and listen */
228bae05 590static int smc_find_ism_device(struct smc_sock *smc, struct smc_init_info *ini)
41349844
HW		 591{
592 /* Find ISM device with same PNETID as connecting interface */
bc36d2fc 593 smc_pnet_find_ism_resource(smc->clcsock->sk, ini);
3fc64937 594 if (!ini->ism_dev[0])
9aa68d29 595 return SMC_CLC_DECL_NOSMCDDEV;
8caaccf5
UB		 596 else
597 ini->ism_chid[0] = smc_ism_get_chid(ini->ism_dev[0]);
41349844
HW		 598 return 0;
599}
600
d70bf4f7
UB		 601/* determine possible V2 ISM devices (either without PNETID or with PNETID plus
602 * PNETID matching net_device)
603 */
604static int smc_find_ism_v2_device_clnt(struct smc_sock *smc,
605 struct smc_init_info *ini)
606{
607 int rc = SMC_CLC_DECL_NOSMCDDEV;
608 struct smcd_dev *smcd;
609 int i = 1;
610
611 if (smcd_indicated(ini->smc_type_v1))
612 rc = 0; /* already initialized for V1 */
613 mutex_lock(&smcd_dev_list.mutex);
614 list_for_each_entry(smcd, &smcd_dev_list.list, list) {
615 if (smcd->going_away || smcd == ini->ism_dev[0])
616 continue;
617 if (!smc_pnet_is_pnetid_set(smcd->pnetid) ||
618 smc_pnet_is_ndev_pnetid(sock_net(&smc->sk), smcd->pnetid)) {
619 ini->ism_dev[i] = smcd;
620 ini->ism_chid[i] = smc_ism_get_chid(ini->ism_dev[i]);
621 ini->is_smcd = true;
622 rc = 0;
623 i++;
624 if (i > SMC_MAX_ISM_DEVS)
625 break;
626 }
627 }
628 mutex_unlock(&smcd_dev_list.mutex);
629 ini->ism_offered_cnt = i - 1;
630 if (!ini->ism_dev[0] && !ini->ism_dev[1])
631 ini->smcd_version = 0;
632
633 return rc;
634}
635
41349844
HW		 636/* Check for VLAN ID and register it on ISM device just for CLC handshake */
637static int smc_connect_ism_vlan_setup(struct smc_sock *smc,
bc36d2fc 638 struct smc_init_info *ini)
41349844 639{
3fc64937 640 if (ini->vlan_id && smc_ism_get_vlan(ini->ism_dev[0], ini->vlan_id))
7a62725a 641 return SMC_CLC_DECL_ISMVLANERR;
41349844
HW		 642 return 0;
643}
644
d70bf4f7
UB		 645static int smc_find_proposal_devices(struct smc_sock *smc,
646 struct smc_init_info *ini)
647{
648 int rc = 0;
649
650 /* check if there is an ism device available */
651 if (ini->smcd_version & SMC_V1) {
652 if (smc_find_ism_device(smc, ini) ||
653 smc_connect_ism_vlan_setup(smc, ini)) {
654 if (ini->smc_type_v1 == SMC_TYPE_B)
655 ini->smc_type_v1 = SMC_TYPE_R;
656 else
657 ini->smc_type_v1 = SMC_TYPE_N;
658 } /* else ISM V1 is supported for this connection */
659 if (smc_find_rdma_device(smc, ini)) {
660 if (ini->smc_type_v1 == SMC_TYPE_B)
661 ini->smc_type_v1 = SMC_TYPE_D;
662 else
663 ini->smc_type_v1 = SMC_TYPE_N;
664 } /* else RDMA is supported for this connection */
665 }
666 if (smc_ism_v2_capable && smc_find_ism_v2_device_clnt(smc, ini))
667 ini->smc_type_v2 = SMC_TYPE_N;
668
669 /* if neither ISM nor RDMA are supported, fallback */
670 if (!smcr_indicated(ini->smc_type_v1) &&
671 ini->smc_type_v1 == SMC_TYPE_N && ini->smc_type_v2 == SMC_TYPE_N)
672 rc = SMC_CLC_DECL_NOSMCDEV;
673
674 return rc;
675}
676
41349844
HW		 677/* cleanup temporary VLAN ID registration used for CLC handshake. If ISM is
678 * used, the VLAN ID will be registered again during the connection setup.
679 */
d70bf4f7 680static int smc_connect_ism_vlan_cleanup(struct smc_sock *smc,
bc36d2fc 681 struct smc_init_info *ini)
41349844 682{
d70bf4f7 683 if (!smcd_indicated(ini->smc_type_v1))
41349844 684 return 0;
3fc64937 685 if (ini->vlan_id && smc_ism_put_vlan(ini->ism_dev[0], ini->vlan_id))
41349844
HW		 686 return SMC_CLC_DECL_CNFERR;
687 return 0;
688}
689
a7c9c5f4
UB		 690#define SMC_CLC_MAX_ACCEPT_LEN \
691 (sizeof(struct smc_clc_msg_accept_confirm_v2) + \
b81a5eb7 692 sizeof(struct smc_clc_first_contact_ext) + \
a7c9c5f4
UB		 693 sizeof(struct smc_clc_msg_trail))
694
3b2dec26 695/* CLC handshake during connect */
d70bf4f7 696static int smc_connect_clc(struct smc_sock *smc,
a7c9c5f4 697 struct smc_clc_msg_accept_confirm_v2 *aclc2,
bc36d2fc 698 struct smc_init_info *ini)
3b2dec26
HW		 699{
700 int rc = 0;
a046d57d
UB		 701
702 /* do inband token exchange */
d70bf4f7 703 rc = smc_clc_send_proposal(smc, ini);
3b2dec26
HW		 704 if (rc)
705 return rc;
a046d57d 706 /* receive SMC Accept CLC message */
a7c9c5f4
UB		 707 return smc_clc_wait_msg(smc, aclc2, SMC_CLC_MAX_ACCEPT_LEN,
708 SMC_CLC_ACCEPT, CLC_WAIT_TIME);
3b2dec26
HW		 709}
710
711/* setup for RDMA connection of client */
712static int smc_connect_rdma(struct smc_sock *smc,
713 struct smc_clc_msg_accept_confirm *aclc,
bc36d2fc 714 struct smc_init_info *ini)
3b2dec26 715{
0fb0b02b 716 int i, reason_code = 0;
3b2dec26 717 struct smc_link *link;
a046d57d 718
bc36d2fc 719 ini->is_smcd = false;
3d9725a6
UB		 720 ini->ib_lcl = &aclc->r0.lcl;
721 ini->ib_clcqpn = ntoh24(aclc->r0.qpn);
f1eb02f9 722 ini->first_contact_peer = aclc->hdr.typev2 & SMC_FIRST_CONTACT_MASK;
bc36d2fc 723
72a36a8a 724 mutex_lock(&smc_client_lgr_pending);
7a62725a
KG		 725 reason_code = smc_conn_create(smc, ini);
726 if (reason_code) {
72a36a8a
HW		 727 mutex_unlock(&smc_client_lgr_pending);
728 return reason_code;
0cfdd8f9 729 }
a046d57d 730
3b2dec26 731 smc_conn_save_peer_info(smc, aclc);
cd6851f3 732
5ac54d87 733 if (ini->first_contact_local) {
0fb0b02b
KG		 734 link = smc->conn.lnk;
735 } else {
736 /* set link that was assigned by server */
737 link = NULL;
738 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
739 struct smc_link *l = &smc->conn.lgr->lnk[i];
740
3d9725a6
UB		 741 if (l->peer_qpn == ntoh24(aclc->r0.qpn) &&
742 !memcmp(l->peer_gid, &aclc->r0.lcl.gid,
743 SMC_GID_SIZE) &&
744 !memcmp(l->peer_mac, &aclc->r0.lcl.mac,
745 sizeof(l->peer_mac))) {
0fb0b02b
KG		 746 link = l;
747 break;
748 }
749 }
750 if (!link)
751 return smc_connect_abort(smc, SMC_CLC_DECL_NOSRVLINK,
5ac54d87 752 ini->first_contact_local);
0fb0b02b
KG		 753 smc->conn.lnk = link;
754 }
755
3e034725 756 /* create send buffer and rmb */
c6ba7c9b 757 if (smc_buf_create(smc, false))
7a62725a 758 return smc_connect_abort(smc, SMC_CLC_DECL_MEM,
5ac54d87 759 ini->first_contact_local);
cd6851f3 760
5ac54d87 761 if (ini->first_contact_local)
3b2dec26 762 smc_link_save_peer_info(link, aclc);
bd4ad577 763
e07d31dc 764 if (smc_rmb_rtoken_handling(&smc->conn, link, aclc))
603cc149 765 return smc_connect_abort(smc, SMC_CLC_DECL_ERR_RTOK,
5ac54d87 766 ini->first_contact_local);
bd4ad577 767
46c28dbd
UB		 768 smc_close_init(smc);
769 smc_rx_init(smc);
770
5ac54d87 771 if (ini->first_contact_local) {
3b2dec26 772 if (smc_ib_ready_link(link))
603cc149 773 return smc_connect_abort(smc, SMC_CLC_DECL_ERR_RDYLNK,
5ac54d87 774 ini->first_contact_local);
652a1e41 775 } else {
7562a13d 776 if (smcr_lgr_reg_rmbs(link, smc->conn.rmb_desc))
603cc149 777 return smc_connect_abort(smc, SMC_CLC_DECL_ERR_REGRMB,
5ac54d87 778 ini->first_contact_local);
bd4ad577 779 }
10428dd8 780 smc_rmb_sync_sg_for_device(&smc->conn);
a046d57d 781
a7c9c5f4
UB		 782 reason_code = smc_clc_send_confirm(smc, ini->first_contact_local,
783 SMC_V1);
3b2dec26 784 if (reason_code)
7a62725a 785 return smc_connect_abort(smc, reason_code,
5ac54d87 786 ini->first_contact_local);
3b2dec26
HW		 787
788 smc_tx_init(smc);
a046d57d 789
5ac54d87 790 if (ini->first_contact_local) {
9bf9abea 791 /* QP confirmation over RoCE fabric */
0fb0b02b 792 smc_llc_flow_initiate(link->lgr, SMC_LLC_FLOW_ADD_LINK);
b9247544 793 reason_code = smcr_clnt_conf_first_link(smc);
0fb0b02b 794 smc_llc_flow_stop(link->lgr, &link->lgr->llc_flow_lcl);
3b2dec26
HW		 795 if (reason_code)
796 return smc_connect_abort(smc, reason_code,
5ac54d87 797 ini->first_contact_local);
9bf9abea 798 }
72a36a8a 799 mutex_unlock(&smc_client_lgr_pending);
e6727f39 800
a046d57d 801 smc_copy_sock_settings_to_clc(smc);
50717a37 802 smc->connect_nonblock = 0;
b38d7324
UB		 803 if (smc->sk.sk_state == SMC_INIT)
804 smc->sk.sk_state = SMC_ACTIVE;
a046d57d 805
3b2dec26
HW		 806 return 0;
807}
a046d57d 808
a7c9c5f4
UB		 809/* The server has chosen one of the proposed ISM devices for the communication.
810 * Determine from the CHID of the received CLC ACCEPT the ISM device chosen.
811 */
812static int
813smc_v2_determine_accepted_chid(struct smc_clc_msg_accept_confirm_v2 *aclc,
814 struct smc_init_info *ini)
815{
816 int i;
817
818 for (i = 0; i < ini->ism_offered_cnt + 1; i++) {
819 if (ini->ism_chid[i] == ntohs(aclc->chid)) {
820 ini->ism_selected = i;
821 return 0;
822 }
823 }
824
825 return -EPROTO;
826}
827
41349844
HW		 828/* setup for ISM connection of client */
829static int smc_connect_ism(struct smc_sock *smc,
830 struct smc_clc_msg_accept_confirm *aclc,
bc36d2fc 831 struct smc_init_info *ini)
41349844 832{
41349844
HW		 833 int rc = 0;
834
bc36d2fc 835 ini->is_smcd = true;
f1eb02f9 836 ini->first_contact_peer = aclc->hdr.typev2 & SMC_FIRST_CONTACT_MASK;
bc36d2fc 837
a7c9c5f4
UB		 838 if (aclc->hdr.version == SMC_V2) {
839 struct smc_clc_msg_accept_confirm_v2 *aclc_v2 =
840 (struct smc_clc_msg_accept_confirm_v2 *)aclc;
841
842 rc = smc_v2_determine_accepted_chid(aclc_v2, ini);
843 if (rc)
844 return rc;
845 }
846 ini->ism_peer_gid[ini->ism_selected] = aclc->d0.gid;
847
72a36a8a
HW		 848 /* there is only one lgr role for SMC-D; use server lock */
849 mutex_lock(&smc_server_lgr_pending);
7a62725a
KG		 850 rc = smc_conn_create(smc, ini);
851 if (rc) {
72a36a8a 852 mutex_unlock(&smc_server_lgr_pending);
7a62725a 853 return rc;
72a36a8a 854 }
41349844
HW		 855
856 /* Create send and receive buffers */
72b7f6c4
KG		 857 rc = smc_buf_create(smc, true);
858 if (rc)
859 return smc_connect_abort(smc, (rc == -ENOSPC) ?
860 SMC_CLC_DECL_MAX_DMB :
861 SMC_CLC_DECL_MEM,
5ac54d87 862 ini->first_contact_local);
41349844
HW		 863
864 smc_conn_save_peer_info(smc, aclc);
865 smc_close_init(smc);
866 smc_rx_init(smc);
867 smc_tx_init(smc);
868
a7c9c5f4
UB		 869 rc = smc_clc_send_confirm(smc, ini->first_contact_local,
870 aclc->hdr.version);
41349844 871 if (rc)
5ac54d87 872 return smc_connect_abort(smc, rc, ini->first_contact_local);
72a36a8a 873 mutex_unlock(&smc_server_lgr_pending);
41349844
HW		 874
875 smc_copy_sock_settings_to_clc(smc);
50717a37 876 smc->connect_nonblock = 0;
41349844
HW		 877 if (smc->sk.sk_state == SMC_INIT)
878 smc->sk.sk_state = SMC_ACTIVE;
879
880 return 0;
881}
882
d70bf4f7
UB		 883/* check if received accept type and version matches a proposed one */
884static int smc_connect_check_aclc(struct smc_init_info *ini,
885 struct smc_clc_msg_accept_confirm *aclc)
886{
887 if ((aclc->hdr.typev1 == SMC_TYPE_R &&
888 !smcr_indicated(ini->smc_type_v1)) ||
889 (aclc->hdr.typev1 == SMC_TYPE_D &&
a7c9c5f4
UB		 890 ((!smcd_indicated(ini->smc_type_v1) &&
891 !smcd_indicated(ini->smc_type_v2)) ||
892 (aclc->hdr.version == SMC_V1 &&
893 !smcd_indicated(ini->smc_type_v1)) ||
894 (aclc->hdr.version == SMC_V2 &&
895 !smcd_indicated(ini->smc_type_v2)))))
d70bf4f7
UB		 896 return SMC_CLC_DECL_MODEUNSUPP;
897
898 return 0;
899}
900
3b2dec26
HW		 901/* perform steps before actually connecting */
902static int __smc_connect(struct smc_sock *smc)
903{
a7c9c5f4
UB		 904 struct smc_clc_msg_accept_confirm_v2 *aclc2;
905 struct smc_clc_msg_accept_confirm *aclc;
3fc64937 906 struct smc_init_info *ini = NULL;
a7c9c5f4 907 u8 *buf = NULL;
3b2dec26 908 int rc = 0;
a046d57d 909
3b2dec26 910 if (smc->use_fallback)
603cc149 911 return smc_connect_fallback(smc, smc->fallback_rsn);
3b2dec26
HW		 912
913 /* if peer has not signalled SMC-capability, fall back */
914 if (!tcp_sk(smc->clcsock->sk)->syn_smc)
603cc149 915 return smc_connect_fallback(smc, SMC_CLC_DECL_PEERNOSMC);
3b2dec26
HW		 916
917 /* IPSec connections opt out of SMC-R optimizations */
918 if (using_ipsec(smc))
919 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_IPSEC);
920
3fc64937
UB		 921 ini = kzalloc(sizeof(*ini), GFP_KERNEL);
922 if (!ini)
923 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_MEM);
924
d70bf4f7
UB		 925 ini->smcd_version = SMC_V1;
926 ini->smcd_version |= smc_ism_v2_capable ? SMC_V2 : 0;
927 ini->smc_type_v1 = SMC_TYPE_B;
928 ini->smc_type_v2 = smc_ism_v2_capable ? SMC_TYPE_D : SMC_TYPE_N;
929
fba7e8ef 930 /* get vlan id from IP device */
3fc64937 931 if (smc_vlan_by_tcpsk(smc->clcsock, ini)) {
d70bf4f7
UB		 932 ini->smcd_version &= ~SMC_V1;
933 ini->smc_type_v1 = SMC_TYPE_N;
934 if (!ini->smcd_version) {
935 rc = SMC_CLC_DECL_GETVLANERR;
936 goto fallback;
937 }
41349844
HW		 938 }
939
d70bf4f7
UB		 940 rc = smc_find_proposal_devices(smc, ini);
941 if (rc)
942 goto fallback;
3b2dec26 943
a7c9c5f4
UB		 944 buf = kzalloc(SMC_CLC_MAX_ACCEPT_LEN, GFP_KERNEL);
945 if (!buf) {
946 rc = SMC_CLC_DECL_MEM;
947 goto fallback;
948 }
949 aclc2 = (struct smc_clc_msg_accept_confirm_v2 *)buf;
950 aclc = (struct smc_clc_msg_accept_confirm *)aclc2;
951
3b2dec26 952 /* perform CLC handshake */
a7c9c5f4 953 rc = smc_connect_clc(smc, aclc2, ini);
d70bf4f7
UB		 954 if (rc)
955 goto vlan_cleanup;
956
957 /* check if smc modes and versions of CLC proposal and accept match */
a7c9c5f4 958 rc = smc_connect_check_aclc(ini, aclc);
d70bf4f7
UB		 959 if (rc)
960 goto vlan_cleanup;
3b2dec26 961
41349844 962 /* depending on previous steps, connect using rdma or ism */
a7c9c5f4
UB		 963 if (aclc->hdr.typev1 == SMC_TYPE_R)
964 rc = smc_connect_rdma(smc, aclc, ini);
965 else if (aclc->hdr.typev1 == SMC_TYPE_D)
966 rc = smc_connect_ism(smc, aclc, ini);
d70bf4f7
UB		 967 if (rc)
968 goto vlan_cleanup;
3b2dec26 969
d70bf4f7 970 smc_connect_ism_vlan_cleanup(smc, ini);
a7c9c5f4 971 kfree(buf);
3fc64937 972 kfree(ini);
3b2dec26 973 return 0;
d70bf4f7
UB		 974
975vlan_cleanup:
976 smc_connect_ism_vlan_cleanup(smc, ini);
a7c9c5f4 977 kfree(buf);
d70bf4f7
UB		 978fallback:
979 kfree(ini);
980 return smc_connect_decline_fallback(smc, rc);
a046d57d
UB		 981}
982
24ac3a08
UB		 983static void smc_connect_work(struct work_struct *work)
984{
985 struct smc_sock *smc = container_of(work, struct smc_sock,
986 connect_work);
50717a37
UB		 987 long timeo = smc->sk.sk_sndtimeo;
988 int rc = 0;
24ac3a08 989
50717a37
UB		 990 if (!timeo)
991 timeo = MAX_SCHEDULE_TIMEOUT;
992 lock_sock(smc->clcsock->sk);
24ac3a08
UB		 993 if (smc->clcsock->sk->sk_err) {
994 smc->sk.sk_err = smc->clcsock->sk->sk_err;
50717a37
UB		 995 } else if ((1 << smc->clcsock->sk->sk_state) &
996 (TCPF_SYN_SENT | TCP_SYN_RECV)) {
997 rc = sk_stream_wait_connect(smc->clcsock->sk, &timeo);
998 if ((rc == -EPIPE) &&
999 ((1 << smc->clcsock->sk->sk_state) &
1000 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)))
1001 rc = 0;
24ac3a08 1002 }
50717a37
UB		 1003 release_sock(smc->clcsock->sk);
1004 lock_sock(&smc->sk);
1005 if (rc != 0 || smc->sk.sk_err) {
1006 smc->sk.sk_state = SMC_CLOSED;
1007 if (rc == -EPIPE || rc == -EAGAIN)
1008 smc->sk.sk_err = EPIPE;
1009 else if (signal_pending(current))
1010 smc->sk.sk_err = -sock_intr_errno(timeo);
6d6dd528 1011 sock_put(&smc->sk); /* passive closing */
24ac3a08
UB		 1012 goto out;
1013 }
1014
1015 rc = __smc_connect(smc);
1016 if (rc < 0)
1017 smc->sk.sk_err = -rc;
1018
1019out:
07603b23
UB		 1020 if (!sock_flag(&smc->sk, SOCK_DEAD)) {
1021 if (smc->sk.sk_err) {
1022 smc->sk.sk_state_change(&smc->sk);
1023 } else { /* allow polling before and after fallback decision */
1024 smc->clcsock->sk->sk_write_space(smc->clcsock->sk);
1025 smc->sk.sk_write_space(&smc->sk);
1026 }
1027 }
24ac3a08
UB		 1028 release_sock(&smc->sk);
1029}
1030
ac713874
UB		 1031static int smc_connect(struct socket *sock, struct sockaddr *addr,
1032 int alen, int flags)
1033{
1034 struct sock *sk = sock->sk;
1035 struct smc_sock *smc;
1036 int rc = -EINVAL;
1037
1038 smc = smc_sk(sk);
1039
1040 /* separate smc parameter checking to be safe */
1041 if (alen < sizeof(addr->sa_family))
1042 goto out_err;
aaa4d33f 1043 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
ac713874
UB		 1044 goto out_err;
1045
1046 lock_sock(sk);
1047 switch (sk->sk_state) {
1048 default:
1049 goto out;
1050 case SMC_ACTIVE:
1051 rc = -EISCONN;
1052 goto out;
1053 case SMC_INIT:
1054 rc = 0;
1055 break;
1056 }
1057
1058 smc_copy_sock_settings_to_clc(smc);
c5c1cc9c 1059 tcp_sk(smc->clcsock->sk)->syn_smc = 1;
50717a37
UB		 1060 if (smc->connect_nonblock) {
1061 rc = -EALREADY;
1062 goto out;
1063 }
1064 rc = kernel_connect(smc->clcsock, addr, alen, flags);
1065 if (rc && rc != -EINPROGRESS)
1066 goto out;
301428ea
UB		 1067
1068 sock_hold(&smc->sk); /* sock put in passive closing */
86434744
UB		 1069 if (smc->use_fallback)
1070 goto out;
24ac3a08 1071 if (flags & O_NONBLOCK) {
22ef473d 1072 if (queue_work(smc_hs_wq, &smc->connect_work))
50717a37 1073 smc->connect_nonblock = 1;
24ac3a08
UB		 1074 rc = -EINPROGRESS;
1075 } else {
24ac3a08
UB		 1076 rc = __smc_connect(smc);
1077 if (rc < 0)
1078 goto out;
1079 else
1080 rc = 0; /* success cases including fallback */
1081 }
ac713874
UB		 1082
1083out:
1084 release_sock(sk);
1085out_err:
1086 return rc;
1087}
1088
1089static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
1090{
3163c507
UB		 1091 struct socket *new_clcsock = NULL;
1092 struct sock *lsk = &lsmc->sk;
ac713874 1093 struct sock *new_sk;
78abe3d0 1094 int rc = -EINVAL;
ac713874 1095
3163c507 1096 release_sock(lsk);
aaa4d33f 1097 new_sk = smc_sock_alloc(sock_net(lsk), NULL, lsk->sk_protocol);
ac713874
UB		 1098 if (!new_sk) {
1099 rc = -ENOMEM;
3163c507 1100 lsk->sk_err = ENOMEM;
ac713874 1101 *new_smc = NULL;
3163c507 1102 lock_sock(lsk);
ac713874
UB		 1103 goto out;
1104 }
1105 *new_smc = smc_sk(new_sk);
1106
78abe3d0
MJ		 1107 mutex_lock(&lsmc->clcsock_release_lock);
1108 if (lsmc->clcsock)
a60a2b1e 1109 rc = kernel_accept(lsmc->clcsock, &new_clcsock, SOCK_NONBLOCK);
78abe3d0 1110 mutex_unlock(&lsmc->clcsock_release_lock);
3163c507 1111 lock_sock(lsk);
a60a2b1e 1112 if (rc < 0 && rc != -EAGAIN)
3163c507 1113 lsk->sk_err = -rc;
35a6b178 1114 if (rc < 0 || lsk->sk_state == SMC_CLOSED) {
f61bca58 1115 new_sk->sk_prot->unhash(new_sk);
a046d57d
UB		 1116 if (new_clcsock)
1117 sock_release(new_clcsock);
1118 new_sk->sk_state = SMC_CLOSED;
1119 sock_set_flag(new_sk, SOCK_DEAD);
51f1de79 1120 sock_put(new_sk); /* final */
ac713874
UB		 1121 *new_smc = NULL;
1122 goto out;
1123 }
1124
a60a2b1e
UB		 1125 /* new clcsock has inherited the smc listen-specific sk_data_ready
1126 * function; switch it back to the original sk_data_ready function
1127 */
1128 new_clcsock->sk->sk_data_ready = lsmc->clcsk_data_ready;
ac713874
UB		 1129 (*new_smc)->clcsock = new_clcsock;
1130out:
1131 return rc;
1132}
1133
a046d57d
UB		 1134/* add a just created sock to the accept queue of the listen sock as
1135 * candidate for a following socket accept call from user space
1136 */
1137static void smc_accept_enqueue(struct sock *parent, struct sock *sk)
1138{
1139 struct smc_sock *par = smc_sk(parent);
1140
51f1de79 1141 sock_hold(sk); /* sock_put in smc_accept_unlink () */
a046d57d
UB		 1142 spin_lock(&par->accept_q_lock);
1143 list_add_tail(&smc_sk(sk)->accept_q, &par->accept_q);
1144 spin_unlock(&par->accept_q_lock);
1145 sk_acceptq_added(parent);
1146}
1147
1148/* remove a socket from the accept queue of its parental listening socket */
1149static void smc_accept_unlink(struct sock *sk)
1150{
1151 struct smc_sock *par = smc_sk(sk)->listen_smc;
1152
1153 spin_lock(&par->accept_q_lock);
1154 list_del_init(&smc_sk(sk)->accept_q);
1155 spin_unlock(&par->accept_q_lock);
1156 sk_acceptq_removed(&smc_sk(sk)->listen_smc->sk);
51f1de79 1157 sock_put(sk); /* sock_hold in smc_accept_enqueue */
a046d57d
UB		 1158}
1159
1160/* remove a sock from the accept queue to bind it to a new socket created
1161 * for a socket accept call from user space
1162 */
b38d7324
UB		 1163struct sock *smc_accept_dequeue(struct sock *parent,
1164 struct socket *new_sock)
a046d57d
UB		 1165{
1166 struct smc_sock *isk, *n;
1167 struct sock *new_sk;
1168
1169 list_for_each_entry_safe(isk, n, &smc_sk(parent)->accept_q, accept_q) {
1170 new_sk = (struct sock *)isk;
1171
1172 smc_accept_unlink(new_sk);
1173 if (new_sk->sk_state == SMC_CLOSED) {
f61bca58 1174 new_sk->sk_prot->unhash(new_sk);
127f4970
UB		 1175 if (isk->clcsock) {
1176 sock_release(isk->clcsock);
1177 isk->clcsock = NULL;
1178 }
51f1de79 1179 sock_put(new_sk); /* final */
a046d57d
UB		 1180 continue;
1181 }
07603b23 1182 if (new_sock) {
a046d57d 1183 sock_graft(new_sk, new_sock);
07603b23
UB		 1184 if (isk->use_fallback) {
1185 smc_sk(new_sk)->clcsock->file = new_sock->file;
1186 isk->clcsock->file->private_data = isk->clcsock;
1187 }
1188 }
a046d57d
UB		 1189 return new_sk;
1190 }
1191 return NULL;
1192}
1193
1194/* clean up for a created but never accepted sock */
b38d7324 1195void smc_close_non_accepted(struct sock *sk)
a046d57d
UB		 1196{
1197 struct smc_sock *smc = smc_sk(sk);
1198
81cf4f47 1199 sock_hold(sk); /* sock_put below */
b38d7324
UB		 1200 lock_sock(sk);
1201 if (!sk->sk_lingertime)
1202 /* wait for peer closing */
1203 sk->sk_lingertime = SMC_MAX_STREAM_WAIT_TIMEOUT;
39f41f36 1204 __smc_release(smc);
b38d7324 1205 release_sock(sk);
81cf4f47 1206 sock_put(sk); /* sock_hold above */
51f1de79 1207 sock_put(sk); /* final sock_put */
a046d57d
UB		 1208}
1209
b9247544 1210static int smcr_serv_conf_first_link(struct smc_sock *smc)
9bf9abea 1211{
387707fd 1212 struct smc_link *link = smc->conn.lnk;
4667bb4a 1213 struct smc_llc_qentry *qentry;
9bf9abea
UB		 1214 int rc;
1215
7562a13d 1216 if (smcr_link_reg_rmb(link, smc->conn.rmb_desc))
603cc149 1217 return SMC_CLC_DECL_ERR_REGRMB;
652a1e41 1218
9bf9abea 1219 /* send CONFIRM LINK request to client over the RoCE fabric */
947541f3 1220 rc = smc_llc_send_confirm_link(link, SMC_LLC_REQ);
9bf9abea 1221 if (rc < 0)
603cc149 1222 return SMC_CLC_DECL_TIMEOUT_CL;
9bf9abea
UB		 1223
1224 /* receive CONFIRM LINK response from client over the RoCE fabric */
4667bb4a
KG		 1225 qentry = smc_llc_wait(link->lgr, link, SMC_LLC_WAIT_TIME,
1226 SMC_LLC_CONFIRM_LINK);
1227 if (!qentry) {
9bf9abea
UB		 1228 struct smc_clc_msg_decline dclc;
1229
1230 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
2b59f58e 1231 SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
9ed28556 1232 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
9bf9abea 1233 }
649758ff 1234 smc_llc_save_peer_uid(qentry);
4667bb4a
KG		 1235 rc = smc_llc_eval_conf_link(qentry, SMC_LLC_RESP);
1236 smc_llc_flow_qentry_del(&link->lgr->llc_flow_lcl);
1237 if (rc)
75d320d6
KG		 1238 return SMC_CLC_DECL_RMBE_EC;
1239
4667bb4a
KG		 1240 /* confirm_rkey is implicit on 1st contact */
1241 smc->conn.rmb_desc->is_conf_rkey = true;
52bedf37 1242
00a049cf 1243 smc_llc_link_active(link);
0a99be43 1244 smcr_lgr_set_type(link->lgr, SMC_LGR_SINGLE);
52bedf37 1245
4667bb4a 1246 /* initial contact - try to establish second link */
2d2209f2 1247 smc_llc_srv_add_link(link);
75d320d6 1248 return 0;
9bf9abea
UB		 1249}
1250
3b2dec26
HW		 1251/* listen worker: finish */
1252static void smc_listen_out(struct smc_sock *new_smc)
a046d57d 1253{
a046d57d 1254 struct smc_sock *lsmc = new_smc->listen_smc;
a046d57d 1255 struct sock *newsmcsk = &new_smc->sk;
a046d57d 1256
3b2dec26 1257 if (lsmc->sk.sk_state == SMC_LISTEN) {
fd57770d 1258 lock_sock_nested(&lsmc->sk, SINGLE_DEPTH_NESTING);
3b2dec26 1259 smc_accept_enqueue(&lsmc->sk, newsmcsk);
fd57770d 1260 release_sock(&lsmc->sk);
3b2dec26
HW		 1261 } else { /* no longer listening */
1262 smc_close_non_accepted(newsmcsk);
c5c1cc9c
UB		 1263 }
1264
3b2dec26
HW		 1265 /* Wake up accept */
1266 lsmc->sk.sk_data_ready(&lsmc->sk);
1267 sock_put(&lsmc->sk); /* sock_hold in smc_tcp_listen_work */
1268}
a046d57d 1269
3b2dec26
HW		 1270/* listen worker: finish in state connected */
1271static void smc_listen_out_connected(struct smc_sock *new_smc)
1272{
1273 struct sock *newsmcsk = &new_smc->sk;
a046d57d 1274
3b2dec26
HW		 1275 sk_refcnt_debug_inc(newsmcsk);
1276 if (newsmcsk->sk_state == SMC_INIT)
1277 newsmcsk->sk_state = SMC_ACTIVE;
1278
1279 smc_listen_out(new_smc);
1280}
1281
1282/* listen worker: finish in error state */
1283static void smc_listen_out_err(struct smc_sock *new_smc)
1284{
1285 struct sock *newsmcsk = &new_smc->sk;
1286
1287 if (newsmcsk->sk_state == SMC_INIT)
1288 sock_put(&new_smc->sk); /* passive closing */
1289 newsmcsk->sk_state = SMC_CLOSED;
3b2dec26
HW		 1290
1291 smc_listen_out(new_smc);
1292}
1293
1294/* listen worker: decline and fall back if possible */
1295static void smc_listen_decline(struct smc_sock *new_smc, int reason_code,
a7c9c5f4 1296 struct smc_init_info *ini)
3b2dec26
HW		 1297{
1298 /* RDMA setup failed, switch back to TCP */
a7c9c5f4 1299 if (ini->first_contact_local)
51e3dfa8
UB		 1300 smc_lgr_cleanup_early(&new_smc->conn);
1301 else
1302 smc_conn_free(&new_smc->conn);
3b2dec26
HW		 1303 if (reason_code < 0) { /* error, no fallback possible */
1304 smc_listen_out_err(new_smc);
1305 return;
1306 }
07603b23 1307 smc_switch_to_fallback(new_smc);
603cc149
KG		 1308 new_smc->fallback_rsn = reason_code;
1309 if (reason_code && reason_code != SMC_CLC_DECL_PEERDECL) {
3b2dec26
HW		 1310 if (smc_clc_send_decline(new_smc, reason_code) < 0) {
1311 smc_listen_out_err(new_smc);
1312 return;
1313 }
a046d57d 1314 }
3b2dec26
HW		 1315 smc_listen_out_connected(new_smc);
1316}
1317
5c21c4cc
UB		 1318/* listen worker: version checking */
1319static int smc_listen_v2_check(struct smc_sock *new_smc,
1320 struct smc_clc_msg_proposal *pclc,
1321 struct smc_init_info *ini)
1322{
1323 struct smc_clc_smcd_v2_extension *pclc_smcd_v2_ext;
1324 struct smc_clc_v2_extension *pclc_v2_ext;
1325
1326 ini->smc_type_v1 = pclc->hdr.typev1;
1327 ini->smc_type_v2 = pclc->hdr.typev2;
1328 ini->smcd_version = ini->smc_type_v1 != SMC_TYPE_N ? SMC_V1 : 0;
1329 if (pclc->hdr.version > SMC_V1)
1330 ini->smcd_version |=
1331 ini->smc_type_v2 != SMC_TYPE_N ? SMC_V2 : 0;
1332 if (!smc_ism_v2_capable) {
1333 ini->smcd_version &= ~SMC_V2;
1334 goto out;
1335 }
1336 pclc_v2_ext = smc_get_clc_v2_ext(pclc);
1337 if (!pclc_v2_ext) {
1338 ini->smcd_version &= ~SMC_V2;
1339 goto out;
1340 }
1341 pclc_smcd_v2_ext = smc_get_clc_smcd_v2_ext(pclc_v2_ext);
1342 if (!pclc_smcd_v2_ext)
1343 ini->smcd_version &= ~SMC_V2;
1344
1345out:
1346 if (!ini->smcd_version) {
1347 if (pclc->hdr.typev1 == SMC_TYPE_B ||
1348 pclc->hdr.typev2 == SMC_TYPE_B)
1349 return SMC_CLC_DECL_NOSMCDEV;
1350 if (pclc->hdr.typev1 == SMC_TYPE_D ||
1351 pclc->hdr.typev2 == SMC_TYPE_D)
1352 return SMC_CLC_DECL_NOSMCDDEV;
1353 return SMC_CLC_DECL_NOSMCRDEV;
1354 }
1355
1356 return 0;
1357}
1358
3b2dec26 1359/* listen worker: check prefixes */
59886697 1360static int smc_listen_prfx_check(struct smc_sock *new_smc,
3b2dec26
HW		 1361 struct smc_clc_msg_proposal *pclc)
1362{
1363 struct smc_clc_msg_proposal_prefix *pclc_prfx;
1364 struct socket *newclcsock = new_smc->clcsock;
a046d57d 1365
5c21c4cc
UB		 1366 if (pclc->hdr.typev1 == SMC_TYPE_N)
1367 return 0;
e7b7a64a 1368 pclc_prfx = smc_clc_proposal_get_prefix(pclc);
3b2dec26 1369 if (smc_clc_prfx_match(newclcsock, pclc_prfx))
59886697 1370 return SMC_CLC_DECL_DIFFPREFIX;
c246d942 1371
3b2dec26
HW		 1372 return 0;
1373}
a046d57d 1374
3b2dec26
HW		 1375/* listen worker: initialize connection and buffers */
1376static int smc_listen_rdma_init(struct smc_sock *new_smc,
7a62725a 1377 struct smc_init_info *ini)
3b2dec26 1378{
7a62725a
KG		 1379 int rc;
1380
0cfdd8f9 1381 /* allocate connection / link group */
7a62725a
KG		 1382 rc = smc_conn_create(new_smc, ini);
1383 if (rc)
1384 return rc;
a046d57d 1385
3e034725 1386 /* create send buffer and rmb */
c6ba7c9b 1387 if (smc_buf_create(new_smc, false))
3b2dec26 1388 return SMC_CLC_DECL_MEM;
a046d57d 1389
3b2dec26
HW		 1390 return 0;
1391}
1392
41349844
HW		 1393/* listen worker: initialize connection and buffers for SMC-D */
1394static int smc_listen_ism_init(struct smc_sock *new_smc,
7a62725a 1395 struct smc_init_info *ini)
41349844 1396{
7a62725a 1397 int rc;
41349844 1398
7a62725a
KG		 1399 rc = smc_conn_create(new_smc, ini);
1400 if (rc)
1401 return rc;
41349844 1402
41349844 1403 /* Create send and receive buffers */
72b7f6c4
KG		 1404 rc = smc_buf_create(new_smc, true);
1405 if (rc) {
5ac54d87 1406 if (ini->first_contact_local)
51e3dfa8
UB		 1407 smc_lgr_cleanup_early(&new_smc->conn);
1408 else
1409 smc_conn_free(&new_smc->conn);
72b7f6c4
KG		 1410 return (rc == -ENOSPC) ? SMC_CLC_DECL_MAX_DMB :
1411 SMC_CLC_DECL_MEM;
41349844
HW		 1412 }
1413
1414 return 0;
1415}
1416
5c21c4cc
UB		 1417static bool smc_is_already_selected(struct smcd_dev *smcd,
1418 struct smc_init_info *ini,
1419 int matches)
1420{
1421 int i;
1422
1423 for (i = 0; i < matches; i++)
1424 if (smcd == ini->ism_dev[i])
1425 return true;
1426
1427 return false;
1428}
1429
1430/* check for ISM devices matching proposed ISM devices */
1431static void smc_check_ism_v2_match(struct smc_init_info *ini,
1432 u16 proposed_chid, u64 proposed_gid,
1433 unsigned int *matches)
1434{
1435 struct smcd_dev *smcd;
1436
1437 list_for_each_entry(smcd, &smcd_dev_list.list, list) {
1438 if (smcd->going_away)
1439 continue;
1440 if (smc_is_already_selected(smcd, ini, *matches))
1441 continue;
1442 if (smc_ism_get_chid(smcd) == proposed_chid &&
1443 !smc_ism_cantalk(proposed_gid, ISM_RESERVED_VLANID, smcd)) {
1444 ini->ism_peer_gid[*matches] = proposed_gid;
1445 ini->ism_dev[*matches] = smcd;
1446 (*matches)++;
1447 break;
1448 }
1449 }
1450}
1451
1452static void smc_find_ism_v2_device_serv(struct smc_sock *new_smc,
1453 struct smc_clc_msg_proposal *pclc,
1454 struct smc_init_info *ini)
1455{
1456 struct smc_clc_smcd_v2_extension *smcd_v2_ext;
1457 struct smc_clc_v2_extension *smc_v2_ext;
1458 struct smc_clc_msg_smcd *pclc_smcd;
1459 unsigned int matches = 0;
1460 u8 *eid = NULL;
1461 int i;
1462
1463 if (!(ini->smcd_version & SMC_V2) || !smcd_indicated(ini->smc_type_v2))
1464 return;
1465
1466 pclc_smcd = smc_get_clc_msg_smcd(pclc);
1467 smc_v2_ext = smc_get_clc_v2_ext(pclc);
1468 smcd_v2_ext = smc_get_clc_smcd_v2_ext(smc_v2_ext);
1469 if (!smcd_v2_ext ||
1470 !smc_v2_ext->hdr.flag.seid) /* no system EID support for SMCD */
1471 goto not_found;
1472
1473 mutex_lock(&smcd_dev_list.mutex);
1474 if (pclc_smcd->ism.chid)
1475 /* check for ISM device matching proposed native ISM device */
1476 smc_check_ism_v2_match(ini, ntohs(pclc_smcd->ism.chid),
1477 ntohll(pclc_smcd->ism.gid), &matches);
1478 for (i = 1; i <= smc_v2_ext->hdr.ism_gid_cnt; i++) {
1479 /* check for ISM devices matching proposed non-native ISM
1480 * devices
1481 */
1482 smc_check_ism_v2_match(ini,
1483 ntohs(smcd_v2_ext->gidchid[i - 1].chid),
1484 ntohll(smcd_v2_ext->gidchid[i - 1].gid),
1485 &matches);
1486 }
1487 mutex_unlock(&smcd_dev_list.mutex);
1488
1489 if (ini->ism_dev[0]) {
1490 smc_ism_get_system_eid(ini->ism_dev[0], &eid);
1491 if (memcmp(eid, smcd_v2_ext->system_eid, SMC_MAX_EID_LEN))
1492 goto not_found;
1493 } else {
1494 goto not_found;
1495 }
1496
1497 /* separate - outside the smcd_dev_list.lock */
1498 for (i = 0; i < matches; i++) {
1499 ini->smcd_version = SMC_V2;
1500 ini->is_smcd = true;
1501 ini->ism_selected = i;
1502 if (smc_listen_ism_init(new_smc, ini))
1503 /* try next active ISM device */
1504 continue;
1505 return; /* matching and usable V2 ISM device found */
1506 }
1507
1508not_found:
1509 ini->smcd_version &= ~SMC_V2;
1510 ini->ism_dev[0] = NULL;
1511 ini->is_smcd = false;
1512}
1513
1514static void smc_find_ism_v1_device_serv(struct smc_sock *new_smc,
1515 struct smc_clc_msg_proposal *pclc,
1516 struct smc_init_info *ini)
7affc809
UB		 1517{
1518 struct smc_clc_msg_smcd *pclc_smcd = smc_get_clc_msg_smcd(pclc);
1519
5c21c4cc
UB		 1520 /* check if ISM V1 is available */
1521 if (!(ini->smcd_version & SMC_V1) || !smcd_indicated(ini->smc_type_v1))
7affc809
UB		 1522 goto not_found;
1523 ini->is_smcd = true; /* prepare ISM check */
8c3dca34 1524 ini->ism_peer_gid[0] = ntohll(pclc_smcd->ism.gid);
7affc809
UB		 1525 if (smc_find_ism_device(new_smc, ini))
1526 goto not_found;
5c21c4cc 1527 ini->ism_selected = 0;
7affc809 1528 if (!smc_listen_ism_init(new_smc, ini))
5c21c4cc 1529 return; /* V1 ISM device found */
7affc809
UB		 1530
1531not_found:
3fc64937 1532 ini->ism_dev[0] = NULL;
7affc809
UB		 1533 ini->is_smcd = false;
1534}
1535
3b2dec26 1536/* listen worker: register buffers */
5ac54d87 1537static int smc_listen_rdma_reg(struct smc_sock *new_smc, bool local_first)
3b2dec26 1538{
b9247544 1539 struct smc_connection *conn = &new_smc->conn;
46c28dbd 1540
5ac54d87 1541 if (!local_first) {
7562a13d 1542 if (smcr_lgr_reg_rmbs(conn->lnk, conn->rmb_desc))
c7674c00 1543 return SMC_CLC_DECL_ERR_REGRMB;
652a1e41 1544 }
10428dd8 1545 smc_rmb_sync_sg_for_device(&new_smc->conn);
652a1e41 1546
3b2dec26
HW		 1547 return 0;
1548}
1549
5c21c4cc
UB		 1550static int smc_find_rdma_v1_device_serv(struct smc_sock *new_smc,
1551 struct smc_clc_msg_proposal *pclc,
1552 struct smc_init_info *ini)
7affc809
UB		 1553{
1554 int rc;
1555
5c21c4cc 1556 if (!smcr_indicated(ini->smc_type_v1))
7affc809
UB		 1557 return SMC_CLC_DECL_NOSMCDEV;
1558
1559 /* prepare RDMA check */
1560 ini->ib_lcl = &pclc->lcl;
1561 rc = smc_find_rdma_device(new_smc, ini);
1562 if (rc) {
1563 /* no RDMA device found */
5c21c4cc 1564 if (ini->smc_type_v1 == SMC_TYPE_B)
7affc809
UB		 1565 /* neither ISM nor RDMA device found */
1566 rc = SMC_CLC_DECL_NOSMCDEV;
1567 return rc;
1568 }
1569 rc = smc_listen_rdma_init(new_smc, ini);
1570 if (rc)
1571 return rc;
1572 return smc_listen_rdma_reg(new_smc, ini->first_contact_local);
1573}
1574
1575/* determine the local device matching to proposal */
1576static int smc_listen_find_device(struct smc_sock *new_smc,
1577 struct smc_clc_msg_proposal *pclc,
1578 struct smc_init_info *ini)
1579{
5c21c4cc
UB		 1580 int rc;
1581
1582 /* check for ISM device matching V2 proposed device */
1583 smc_find_ism_v2_device_serv(new_smc, pclc, ini);
1584 if (ini->ism_dev[0])
7affc809 1585 return 0;
5c21c4cc
UB		 1586
1587 if (!(ini->smcd_version & SMC_V1))
1588 return SMC_CLC_DECL_NOSMCDEV;
1589
1590 /* check for matching IP prefix and subnet length */
1591 rc = smc_listen_prfx_check(new_smc, pclc);
1592 if (rc)
1593 return rc;
1594
1595 /* get vlan id from IP device */
1596 if (smc_vlan_by_tcpsk(new_smc->clcsock, ini))
1597 return SMC_CLC_DECL_GETVLANERR;
1598
1599 /* check for ISM device matching V1 proposed device */
1600 smc_find_ism_v1_device_serv(new_smc, pclc, ini);
1601 if (ini->ism_dev[0])
1602 return 0;
1603
7affc809
UB		 1604 if (pclc->hdr.typev1 == SMC_TYPE_D)
1605 return SMC_CLC_DECL_NOSMCDDEV; /* skip RDMA and decline */
1606
1607 /* check if RDMA is available */
5c21c4cc 1608 return smc_find_rdma_v1_device_serv(new_smc, pclc, ini);
7affc809
UB		 1609}
1610
3b2dec26 1611/* listen worker: finish RDMA setup */
1ca52fcf
UB		 1612static int smc_listen_rdma_finish(struct smc_sock *new_smc,
1613 struct smc_clc_msg_accept_confirm *cclc,
5ac54d87 1614 bool local_first)
3b2dec26 1615{
387707fd 1616 struct smc_link *link = new_smc->conn.lnk;
3b2dec26 1617 int reason_code = 0;
a046d57d 1618
5ac54d87 1619 if (local_first)
3b2dec26 1620 smc_link_save_peer_info(link, cclc);
a046d57d 1621
0c881ada
UB		 1622 if (smc_rmb_rtoken_handling(&new_smc->conn, link, cclc))
1623 return SMC_CLC_DECL_ERR_RTOK;
bd4ad577 1624
5ac54d87 1625 if (local_first) {
0c881ada
UB		 1626 if (smc_ib_ready_link(link))
1627 return SMC_CLC_DECL_ERR_RDYLNK;
9bf9abea 1628 /* QP confirmation over RoCE fabric */
4667bb4a 1629 smc_llc_flow_initiate(link->lgr, SMC_LLC_FLOW_ADD_LINK);
b9247544 1630 reason_code = smcr_serv_conf_first_link(new_smc);
4667bb4a 1631 smc_llc_flow_stop(link->lgr, &link->lgr->llc_flow_lcl);
bd4ad577 1632 }
1ca52fcf 1633 return reason_code;
3b2dec26 1634}
e6727f39 1635
7affc809 1636/* setup for connection of server */
3b2dec26
HW		 1637static void smc_listen_work(struct work_struct *work)
1638{
1639 struct smc_sock *new_smc = container_of(work, struct smc_sock,
1640 smc_listen_work);
1641 struct socket *newclcsock = new_smc->clcsock;
a7c9c5f4
UB		 1642 struct smc_clc_msg_accept_confirm_v2 *cclc2;
1643 struct smc_clc_msg_accept_confirm *cclc;
6bb14e48 1644 struct smc_clc_msg_proposal_area *buf;
3b2dec26 1645 struct smc_clc_msg_proposal *pclc;
3fc64937 1646 struct smc_init_info *ini = NULL;
3b2dec26 1647 int rc = 0;
3b2dec26 1648
fd57770d
KG		 1649 if (new_smc->listen_smc->sk.sk_state != SMC_LISTEN)
1650 return smc_listen_out_err(new_smc);
1651
3b2dec26
HW		 1652 if (new_smc->use_fallback) {
1653 smc_listen_out_connected(new_smc);
1654 return;
a046d57d 1655 }
a046d57d 1656
3b2dec26
HW		 1657 /* check if peer is smc capable */
1658 if (!tcp_sk(newclcsock->sk)->syn_smc) {
07603b23 1659 smc_switch_to_fallback(new_smc);
603cc149 1660 new_smc->fallback_rsn = SMC_CLC_DECL_PEERNOSMC;
3b2dec26
HW		 1661 smc_listen_out_connected(new_smc);
1662 return;
1663 }
a046d57d 1664
3b2dec26
HW		 1665 /* do inband token exchange -
1666 * wait for and receive SMC Proposal CLC message
1667 */
6bb14e48
UB		 1668 buf = kzalloc(sizeof(*buf), GFP_KERNEL);
1669 if (!buf) {
1670 rc = SMC_CLC_DECL_MEM;
1671 goto out_decl;
1672 }
1673 pclc = (struct smc_clc_msg_proposal *)buf;
1674 rc = smc_clc_wait_msg(new_smc, pclc, sizeof(*buf),
228bae05 1675 SMC_CLC_PROPOSAL, CLC_WAIT_TIME);
9aa68d29
KG		 1676 if (rc)
1677 goto out_decl;
a046d57d 1678
3b2dec26
HW		 1679 /* IPSec connections opt out of SMC-R optimizations */
1680 if (using_ipsec(new_smc)) {
9aa68d29
KG		 1681 rc = SMC_CLC_DECL_IPSEC;
1682 goto out_decl;
3b2dec26
HW		 1683 }
1684
3fc64937
UB		 1685 ini = kzalloc(sizeof(*ini), GFP_KERNEL);
1686 if (!ini) {
1687 rc = SMC_CLC_DECL_MEM;
1688 goto out_decl;
1689 }
1690
5c21c4cc
UB		 1691 /* initial version checking */
1692 rc = smc_listen_v2_check(new_smc, pclc, ini);
1693 if (rc)
9aa68d29 1694 goto out_decl;
fba7e8ef 1695
72a36a8a 1696 mutex_lock(&smc_server_lgr_pending);
3b2dec26
HW		 1697 smc_close_init(new_smc);
1698 smc_rx_init(new_smc);
1699 smc_tx_init(new_smc);
1700
7affc809 1701 /* determine ISM or RoCE device used for connection */
3fc64937 1702 rc = smc_listen_find_device(new_smc, pclc, ini);
7affc809
UB		 1703 if (rc)
1704 goto out_unlock;
3b2dec26
HW		 1705
1706 /* send SMC Accept CLC message */
a7c9c5f4
UB		 1707 rc = smc_clc_send_accept(new_smc, ini->first_contact_local,
1708 ini->smcd_version == SMC_V2 ? SMC_V2 : SMC_V1);
9aa68d29
KG		 1709 if (rc)
1710 goto out_unlock;
3b2dec26 1711
62c7139f 1712 /* SMC-D does not need this lock any more */
3fc64937 1713 if (ini->is_smcd)
72a36a8a 1714 mutex_unlock(&smc_server_lgr_pending);
62c7139f 1715
3b2dec26 1716 /* receive SMC Confirm CLC message */
a7c9c5f4
UB		 1717 cclc2 = (struct smc_clc_msg_accept_confirm_v2 *)buf;
1718 cclc = (struct smc_clc_msg_accept_confirm *)cclc2;
1719 memset(buf, 0, sizeof(struct smc_clc_msg_proposal_area));
1720 rc = smc_clc_wait_msg(new_smc, cclc2,
1721 sizeof(struct smc_clc_msg_proposal_area),
228bae05
KG		 1722 SMC_CLC_CONFIRM, CLC_WAIT_TIME);
1723 if (rc) {
3fc64937 1724 if (!ini->is_smcd)
9aa68d29
KG		 1725 goto out_unlock;
1726 goto out_decl;
3b2dec26
HW		 1727 }
1728
1729 /* finish worker */
3fc64937 1730 if (!ini->is_smcd) {
a7c9c5f4 1731 rc = smc_listen_rdma_finish(new_smc, cclc,
3fc64937 1732 ini->first_contact_local);
62c7139f 1733 if (rc)
0c881ada
UB		 1734 goto out_unlock;
1735 mutex_unlock(&smc_server_lgr_pending);
1ca52fcf 1736 }
a7c9c5f4 1737 smc_conn_save_peer_info(new_smc, cclc);
3b2dec26 1738 smc_listen_out_connected(new_smc);
ac679364 1739 goto out_free;
9aa68d29
KG		 1740
1741out_unlock:
1742 mutex_unlock(&smc_server_lgr_pending);
1743out_decl:
a7c9c5f4 1744 smc_listen_decline(new_smc, rc, ini);
ac679364 1745out_free:
3fc64937 1746 kfree(ini);
6bb14e48 1747 kfree(buf);
a046d57d
UB		 1748}
1749
1750static void smc_tcp_listen_work(struct work_struct *work)
1751{
1752 struct smc_sock *lsmc = container_of(work, struct smc_sock,
1753 tcp_listen_work);
3163c507 1754 struct sock *lsk = &lsmc->sk;
a046d57d
UB		 1755 struct smc_sock *new_smc;
1756 int rc = 0;
1757
3163c507
UB		 1758 lock_sock(lsk);
1759 while (lsk->sk_state == SMC_LISTEN) {
a046d57d 1760 rc = smc_clcsock_accept(lsmc, &new_smc);
a60a2b1e 1761 if (rc) /* clcsock accept queue empty or error */
a046d57d
UB		 1762 goto out;
1763 if (!new_smc)
1764 continue;
1765
1766 new_smc->listen_smc = lsmc;
ee9dfbef 1767 new_smc->use_fallback = lsmc->use_fallback;
603cc149 1768 new_smc->fallback_rsn = lsmc->fallback_rsn;
3163c507 1769 sock_hold(lsk); /* sock_put in smc_listen_work */
a046d57d
UB		 1770 INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
1771 smc_copy_sock_settings_to_smc(new_smc);
bd58c7e0
UB		 1772 new_smc->sk.sk_sndbuf = lsmc->sk.sk_sndbuf;
1773 new_smc->sk.sk_rcvbuf = lsmc->sk.sk_rcvbuf;
51f1de79 1774 sock_hold(&new_smc->sk); /* sock_put in passive closing */
22ef473d 1775 if (!queue_work(smc_hs_wq, &new_smc->smc_listen_work))
51f1de79 1776 sock_put(&new_smc->sk);
a046d57d
UB		 1777 }
1778
1779out:
3163c507 1780 release_sock(lsk);
a60a2b1e
UB		 1781 sock_put(&lsmc->sk); /* sock_hold in smc_clcsock_data_ready() */
1782}
1783
1784static void smc_clcsock_data_ready(struct sock *listen_clcsock)
1785{
1786 struct smc_sock *lsmc;
1787
1788 lsmc = (struct smc_sock *)
1789 ((uintptr_t)listen_clcsock->sk_user_data & ~SK_USER_DATA_NOCOPY);
1790 if (!lsmc)
1791 return;
1792 lsmc->clcsk_data_ready(listen_clcsock);
1793 if (lsmc->sk.sk_state == SMC_LISTEN) {
1794 sock_hold(&lsmc->sk); /* sock_put in smc_tcp_listen_work() */
22ef473d 1795 if (!queue_work(smc_hs_wq, &lsmc->tcp_listen_work))
a60a2b1e
UB		 1796 sock_put(&lsmc->sk);
1797 }
a046d57d
UB		 1798}
1799
ac713874
UB		 1800static int smc_listen(struct socket *sock, int backlog)
1801{
1802 struct sock *sk = sock->sk;
1803 struct smc_sock *smc;
1804 int rc;
1805
1806 smc = smc_sk(sk);
1807 lock_sock(sk);
1808
1809 rc = -EINVAL;
cd206360
UB		 1810 if ((sk->sk_state != SMC_INIT && sk->sk_state != SMC_LISTEN) ||
1811 smc->connect_nonblock)
ac713874
UB		 1812 goto out;
1813
1814 rc = 0;
1815 if (sk->sk_state == SMC_LISTEN) {
1816 sk->sk_max_ack_backlog = backlog;
1817 goto out;
1818 }
1819 /* some socket options are handled in core, so we could not apply
1820 * them to the clc socket -- copy smc socket options to clc socket
1821 */
1822 smc_copy_sock_settings_to_clc(smc);
ee9dfbef
UB		 1823 if (!smc->use_fallback)
1824 tcp_sk(smc->clcsock->sk)->syn_smc = 1;
ac713874 1825
a60a2b1e
UB		 1826 /* save original sk_data_ready function and establish
1827 * smc-specific sk_data_ready function
1828 */
1829 smc->clcsk_data_ready = smc->clcsock->sk->sk_data_ready;
1830 smc->clcsock->sk->sk_data_ready = smc_clcsock_data_ready;
1831 smc->clcsock->sk->sk_user_data =
1832 (void *)((uintptr_t)smc | SK_USER_DATA_NOCOPY);
ac713874
UB		 1833 rc = kernel_listen(smc->clcsock, backlog);
1834 if (rc)
1835 goto out;
1836 sk->sk_max_ack_backlog = backlog;
1837 sk->sk_ack_backlog = 0;
1838 sk->sk_state = SMC_LISTEN;
1839
1840out:
1841 release_sock(sk);
1842 return rc;
1843}
1844
1845static int smc_accept(struct socket *sock, struct socket *new_sock,
cdfbabfb 1846 int flags, bool kern)
ac713874 1847{
a046d57d
UB		 1848 struct sock *sk = sock->sk, *nsk;
1849 DECLARE_WAITQUEUE(wait, current);
ac713874 1850 struct smc_sock *lsmc;
a046d57d
UB		 1851 long timeo;
1852 int rc = 0;
ac713874
UB		 1853
1854 lsmc = smc_sk(sk);
51f1de79 1855 sock_hold(sk); /* sock_put below */
ac713874
UB		 1856 lock_sock(sk);
1857
1858 if (lsmc->sk.sk_state != SMC_LISTEN) {
1859 rc = -EINVAL;
abb190f1 1860 release_sock(sk);
ac713874
UB		 1861 goto out;
1862 }
1863
a046d57d
UB		 1864 /* Wait for an incoming connection */
1865 timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
1866 add_wait_queue_exclusive(sk_sleep(sk), &wait);
1867 while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
1868 set_current_state(TASK_INTERRUPTIBLE);
1869 if (!timeo) {
1870 rc = -EAGAIN;
1871 break;
1872 }
1873 release_sock(sk);
1874 timeo = schedule_timeout(timeo);
1875 /* wakeup by sk_data_ready in smc_listen_work() */
1876 sched_annotate_sleep();
1877 lock_sock(sk);
1878 if (signal_pending(current)) {
1879 rc = sock_intr_errno(timeo);
1880 break;
1881 }
1882 }
1883 set_current_state(TASK_RUNNING);
1884 remove_wait_queue(sk_sleep(sk), &wait);
ac713874 1885
a046d57d
UB		 1886 if (!rc)
1887 rc = sock_error(nsk);
abb190f1
UB		 1888 release_sock(sk);
1889 if (rc)
1890 goto out;
1891
1892 if (lsmc->sockopt_defer_accept && !(flags & O_NONBLOCK)) {
1893 /* wait till data arrives on the socket */
1894 timeo = msecs_to_jiffies(lsmc->sockopt_defer_accept *
1895 MSEC_PER_SEC);
1896 if (smc_sk(nsk)->use_fallback) {
1897 struct sock *clcsk = smc_sk(nsk)->clcsock->sk;
1898
1899 lock_sock(clcsk);
1900 if (skb_queue_empty(&clcsk->sk_receive_queue))
1901 sk_wait_data(clcsk, &timeo, NULL);
1902 release_sock(clcsk);
1903 } else if (!atomic_read(&smc_sk(nsk)->conn.bytes_to_rcv)) {
1904 lock_sock(nsk);
b51fa1b1 1905 smc_rx_wait(smc_sk(nsk), &timeo, smc_rx_data_available);
abb190f1
UB		 1906 release_sock(nsk);
1907 }
1908 }
ac713874
UB		 1909
1910out:
51f1de79 1911 sock_put(sk); /* sock_hold above */
ac713874
UB		 1912 return rc;
1913}
1914
1915static int smc_getname(struct socket *sock, struct sockaddr *addr,
9b2c45d4 1916 int peer)
ac713874
UB		 1917{
1918 struct smc_sock *smc;
1919
b38d7324
UB		 1920 if (peer && (sock->sk->sk_state != SMC_ACTIVE) &&
1921 (sock->sk->sk_state != SMC_APPCLOSEWAIT1))
ac713874
UB		 1922 return -ENOTCONN;
1923
1924 smc = smc_sk(sock->sk);
1925
9b2c45d4 1926 return smc->clcsock->ops->getname(smc->clcsock, addr, peer);
ac713874
UB		 1927}
1928
1929static int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
1930{
1931 struct sock *sk = sock->sk;
1932 struct smc_sock *smc;
1933 int rc = -EPIPE;
1934
1935 smc = smc_sk(sk);
1936 lock_sock(sk);
b38d7324
UB		 1937 if ((sk->sk_state != SMC_ACTIVE) &&
1938 (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1939 (sk->sk_state != SMC_INIT))
ac713874 1940 goto out;
ee9dfbef
UB		 1941
1942 if (msg->msg_flags & MSG_FASTOPEN) {
cd206360 1943 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
07603b23 1944 smc_switch_to_fallback(smc);
603cc149 1945 smc->fallback_rsn = SMC_CLC_DECL_OPTUNSUPP;
ee9dfbef
UB		 1946 } else {
1947 rc = -EINVAL;
1948 goto out;
1949 }
1950 }
1951
ac713874
UB		 1952 if (smc->use_fallback)
1953 rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
1954 else
e6727f39 1955 rc = smc_tx_sendmsg(smc, msg, len);
ac713874
UB		 1956out:
1957 release_sock(sk);
1958 return rc;
1959}
1960
1961static int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
1962 int flags)
1963{
1964 struct sock *sk = sock->sk;
1965 struct smc_sock *smc;
1966 int rc = -ENOTCONN;
1967
1968 smc = smc_sk(sk);
1969 lock_sock(sk);
51c5aba3
KG		 1970 if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
1971 /* socket was connected before, no more data to read */
1972 rc = 0;
1973 goto out;
1974 }
b38d7324
UB		 1975 if ((sk->sk_state == SMC_INIT) ||
1976 (sk->sk_state == SMC_LISTEN) ||
1977 (sk->sk_state == SMC_CLOSED))
ac713874
UB		 1978 goto out;
1979
b38d7324
UB		 1980 if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
1981 rc = 0;
1982 goto out;
1983 }
1984
9014db20 1985 if (smc->use_fallback) {
ac713874 1986 rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
9014db20
SR		 1987 } else {
1988 msg->msg_namelen = 0;
1989 rc = smc_rx_recvmsg(smc, msg, NULL, len, flags);
1990 }
b38d7324 1991
ac713874
UB		 1992out:
1993 release_sock(sk);
1994 return rc;
1995}
1996
ade994f4 1997static __poll_t smc_accept_poll(struct sock *parent)
a046d57d 1998{
8dce2786 1999 struct smc_sock *isk = smc_sk(parent);
63e2480c 2000 __poll_t mask = 0;
a046d57d 2001
8dce2786
UB		 2002 spin_lock(&isk->accept_q_lock);
2003 if (!list_empty(&isk->accept_q))
a9a08845 2004 mask = EPOLLIN | EPOLLRDNORM;
8dce2786 2005 spin_unlock(&isk->accept_q_lock);
a046d57d 2006
8dce2786 2007 return mask;
a046d57d
UB		 2008}
2009
a11e1d43
LT		 2010static __poll_t smc_poll(struct file *file, struct socket *sock,
2011 poll_table *wait)
ac713874
UB		 2012{
2013 struct sock *sk = sock->sk;
ac713874 2014 struct smc_sock *smc;
50717a37 2015 __poll_t mask = 0;
ac713874 2016
8dce2786 2017 if (!sk)
a9a08845 2018 return EPOLLNVAL;
8dce2786 2019
ac713874 2020 smc = smc_sk(sock->sk);
648a5a7a 2021 if (smc->use_fallback) {
a046d57d 2022 /* delegate to CLC child sock */
a11e1d43 2023 mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
784813ae 2024 sk->sk_err = smc->clcsock->sk->sk_err;
ac713874 2025 } else {
410da1e1 2026 if (sk->sk_state != SMC_CLOSED)
89ab066d 2027 sock_poll_wait(file, sock, wait);
a046d57d 2028 if (sk->sk_err)
a9a08845 2029 mask |= EPOLLERR;
b38d7324
UB		 2030 if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
2031 (sk->sk_state == SMC_CLOSED))
a9a08845 2032 mask |= EPOLLHUP;
8dce2786
UB		 2033 if (sk->sk_state == SMC_LISTEN) {
2034 /* woken up by sk_data_ready in smc_listen_work() */
50717a37
UB		 2035 mask |= smc_accept_poll(sk);
2036 } else if (smc->use_fallback) { /* as result of connect_work()*/
2037 mask |= smc->clcsock->ops->poll(file, smc->clcsock,
2038 wait);
2039 sk->sk_err = smc->clcsock->sk->sk_err;
8dce2786 2040 } else {
50717a37
UB		 2041 if ((sk->sk_state != SMC_INIT &&
2042 atomic_read(&smc->conn.sndbuf_space)) ||
8dce2786 2043 sk->sk_shutdown & SEND_SHUTDOWN) {
a9a08845 2044 mask |= EPOLLOUT | EPOLLWRNORM;
8dce2786
UB		 2045 } else {
2046 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
2047 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
2048 }
2049 if (atomic_read(&smc->conn.bytes_to_rcv))
a9a08845 2050 mask |= EPOLLIN | EPOLLRDNORM;
8dce2786 2051 if (sk->sk_shutdown & RCV_SHUTDOWN)
a9a08845 2052 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
8dce2786 2053 if (sk->sk_state == SMC_APPCLOSEWAIT1)
a9a08845 2054 mask |= EPOLLIN;
71d117f5
KG		 2055 if (smc->conn.urg_state == SMC_URG_VALID)
2056 mask |= EPOLLPRI;
8dce2786 2057 }
ac713874
UB		 2058 }
2059
2060 return mask;
2061}
2062
2063static int smc_shutdown(struct socket *sock, int how)
2064{
2065 struct sock *sk = sock->sk;
2066 struct smc_sock *smc;
2067 int rc = -EINVAL;
b38d7324 2068 int rc1 = 0;
ac713874
UB		 2069
2070 smc = smc_sk(sk);
2071
2072 if ((how < SHUT_RD) || (how > SHUT_RDWR))
b38d7324 2073 return rc;
ac713874
UB		 2074
2075 lock_sock(sk);
2076
2077 rc = -ENOTCONN;
caa21e19 2078 if ((sk->sk_state != SMC_ACTIVE) &&
b38d7324
UB		 2079 (sk->sk_state != SMC_PEERCLOSEWAIT1) &&
2080 (sk->sk_state != SMC_PEERCLOSEWAIT2) &&
2081 (sk->sk_state != SMC_APPCLOSEWAIT1) &&
2082 (sk->sk_state != SMC_APPCLOSEWAIT2) &&
2083 (sk->sk_state != SMC_APPFINCLOSEWAIT))
ac713874
UB		 2084 goto out;
2085 if (smc->use_fallback) {
2086 rc = kernel_sock_shutdown(smc->clcsock, how);
2087 sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
2088 if (sk->sk_shutdown == SHUTDOWN_MASK)
2089 sk->sk_state = SMC_CLOSED;
b38d7324 2090 goto out;
ac713874 2091 }
b38d7324
UB		 2092 switch (how) {
2093 case SHUT_RDWR: /* shutdown in both directions */
2094 rc = smc_close_active(smc);
2095 break;
2096 case SHUT_WR:
2097 rc = smc_close_shutdown_write(smc);
2098 break;
2099 case SHUT_RD:
1255fcb2
UB		 2100 rc = 0;
2101 /* nothing more to do because peer is not involved */
b38d7324
UB		 2102 break;
2103 }
1255fcb2
UB		 2104 if (smc->clcsock)
2105 rc1 = kernel_sock_shutdown(smc->clcsock, how);
b38d7324
UB		 2106 /* map sock_shutdown_cmd constants to sk_shutdown value range */
2107 sk->sk_shutdown |= how + 1;
ac713874
UB		 2108
2109out:
2110 release_sock(sk);
b38d7324 2111 return rc ? rc : rc1;
ac713874
UB		 2112}
2113
2114static int smc_setsockopt(struct socket *sock, int level, int optname,
a7b75c5a 2115 sockptr_t optval, unsigned int optlen)
ac713874
UB		 2116{
2117 struct sock *sk = sock->sk;
2118 struct smc_sock *smc;
01d2f7e2 2119 int val, rc;
ac713874
UB		 2120
2121 smc = smc_sk(sk);
2122
2123 /* generic setsockopts reaching us here always apply to the
2124 * CLC socket
2125 */
a44d9e72
CH		 2126 if (unlikely(!smc->clcsock->ops->setsockopt))
2127 rc = -EOPNOTSUPP;
2128 else
2129 rc = smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
2130 optval, optlen);
ee9dfbef
UB		 2131 if (smc->clcsock->sk->sk_err) {
2132 sk->sk_err = smc->clcsock->sk->sk_err;
2133 sk->sk_error_report(sk);
2134 }
ee9dfbef 2135
01d2f7e2 2136 if (optlen < sizeof(int))
3dc9f558 2137 return -EINVAL;
a7b75c5a 2138 if (copy_from_sockptr(&val, optval, sizeof(int)))
ac0107ed 2139 return -EFAULT;
01d2f7e2 2140
ee9dfbef 2141 lock_sock(sk);
86434744
UB		 2142 if (rc || smc->use_fallback)
2143 goto out;
ee9dfbef
UB		 2144 switch (optname) {
2145 case TCP_ULP:
2146 case TCP_FASTOPEN:
2147 case TCP_FASTOPEN_CONNECT:
2148 case TCP_FASTOPEN_KEY:
2149 case TCP_FASTOPEN_NO_COOKIE:
2150 /* option not supported by SMC */
8204df72 2151 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
07603b23 2152 smc_switch_to_fallback(smc);
603cc149 2153 smc->fallback_rsn = SMC_CLC_DECL_OPTUNSUPP;
ee9dfbef 2154 } else {
86434744 2155 rc = -EINVAL;
ee9dfbef
UB		 2156 }
2157 break;
01d2f7e2 2158 case TCP_NODELAY:
f9cedf1a
UB		 2159 if (sk->sk_state != SMC_INIT &&
2160 sk->sk_state != SMC_LISTEN &&
2161 sk->sk_state != SMC_CLOSED) {
86434744 2162 if (val)
22ef473d
KG		 2163 mod_delayed_work(smc->conn.lgr->tx_wq,
2164 &smc->conn.tx_work, 0);
01d2f7e2
UB		 2165 }
2166 break;
2167 case TCP_CORK:
f9cedf1a
UB		 2168 if (sk->sk_state != SMC_INIT &&
2169 sk->sk_state != SMC_LISTEN &&
2170 sk->sk_state != SMC_CLOSED) {
86434744 2171 if (!val)
22ef473d
KG		 2172 mod_delayed_work(smc->conn.lgr->tx_wq,
2173 &smc->conn.tx_work, 0);
01d2f7e2
UB		 2174 }
2175 break;
abb190f1
UB		 2176 case TCP_DEFER_ACCEPT:
2177 smc->sockopt_defer_accept = val;
2178 break;
ee9dfbef
UB		 2179 default:
2180 break;
2181 }
86434744 2182out:
ee9dfbef
UB		 2183 release_sock(sk);
2184
2185 return rc;
ac713874
UB		 2186}
2187
2188static int smc_getsockopt(struct socket *sock, int level, int optname,
2189 char __user *optval, int __user *optlen)
2190{
2191 struct smc_sock *smc;
2192
2193 smc = smc_sk(sock->sk);
2194 /* socket options apply to the CLC socket */
a44d9e72
CH		 2195 if (unlikely(!smc->clcsock->ops->getsockopt))
2196 return -EOPNOTSUPP;
ac713874
UB		 2197 return smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
2198 optval, optlen);
2199}
2200
2201static int smc_ioctl(struct socket *sock, unsigned int cmd,
2202 unsigned long arg)
2203{
de8474eb
SR		 2204 union smc_host_cursor cons, urg;
2205 struct smc_connection *conn;
ac713874 2206 struct smc_sock *smc;
9b67e26f 2207 int answ;
ac713874
UB		 2208
2209 smc = smc_sk(sock->sk);
de8474eb 2210 conn = &smc->conn;
7311d665 2211 lock_sock(&smc->sk);
9b67e26f 2212 if (smc->use_fallback) {
7311d665
UB		 2213 if (!smc->clcsock) {
2214 release_sock(&smc->sk);
9b67e26f 2215 return -EBADF;
7311d665
UB		 2216 }
2217 answ = smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
2218 release_sock(&smc->sk);
2219 return answ;
9b67e26f
UB		 2220 }
2221 switch (cmd) {
2222 case SIOCINQ: /* same as FIONREAD */
1992d998
UB		 2223 if (smc->sk.sk_state == SMC_LISTEN) {
2224 release_sock(&smc->sk);
9b67e26f 2225 return -EINVAL;
1992d998 2226 }
2351abe6
UB		 2227 if (smc->sk.sk_state == SMC_INIT ||
2228 smc->sk.sk_state == SMC_CLOSED)
2229 answ = 0;
2230 else
2231 answ = atomic_read(&smc->conn.bytes_to_rcv);
9b67e26f
UB		 2232 break;
2233 case SIOCOUTQ:
2234 /* output queue size (not send + not acked) */
1992d998
UB		 2235 if (smc->sk.sk_state == SMC_LISTEN) {
2236 release_sock(&smc->sk);
9b67e26f 2237 return -EINVAL;
1992d998 2238 }
2351abe6
UB		 2239 if (smc->sk.sk_state == SMC_INIT ||
2240 smc->sk.sk_state == SMC_CLOSED)
2241 answ = 0;
2242 else
2243 answ = smc->conn.sndbuf_desc->len -
9b67e26f
UB		 2244 atomic_read(&smc->conn.sndbuf_space);
2245 break;
2246 case SIOCOUTQNSD:
2247 /* output queue size (not send only) */
1992d998
UB		 2248 if (smc->sk.sk_state == SMC_LISTEN) {
2249 release_sock(&smc->sk);
9b67e26f 2250 return -EINVAL;
1992d998 2251 }
2351abe6
UB		 2252 if (smc->sk.sk_state == SMC_INIT ||
2253 smc->sk.sk_state == SMC_CLOSED)
2254 answ = 0;
2255 else
2256 answ = smc_tx_prepared_sends(&smc->conn);
9b67e26f 2257 break;
de8474eb 2258 case SIOCATMARK:
1992d998
UB		 2259 if (smc->sk.sk_state == SMC_LISTEN) {
2260 release_sock(&smc->sk);
de8474eb 2261 return -EINVAL;
1992d998 2262 }
de8474eb
SR		 2263 if (smc->sk.sk_state == SMC_INIT ||
2264 smc->sk.sk_state == SMC_CLOSED) {
2265 answ = 0;
2266 } else {
bac6de7b
SR		 2267 smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
2268 smc_curs_copy(&urg, &conn->urg_curs, conn);
de8474eb
SR		 2269 answ = smc_curs_diff(conn->rmb_desc->len,
2270 &cons, &urg) == 1;
2271 }
2272 break;
9b67e26f 2273 default:
1992d998 2274 release_sock(&smc->sk);
9b67e26f
UB		 2275 return -ENOIOCTLCMD;
2276 }
1992d998 2277 release_sock(&smc->sk);
9b67e26f
UB		 2278
2279 return put_user(answ, (int __user *)arg);
ac713874
UB		 2280}
2281
2282static ssize_t smc_sendpage(struct socket *sock, struct page *page,
2283 int offset, size_t size, int flags)
2284{
2285 struct sock *sk = sock->sk;
2286 struct smc_sock *smc;
2287 int rc = -EPIPE;
2288
2289 smc = smc_sk(sk);
2290 lock_sock(sk);
bda27ff5
SR		 2291 if (sk->sk_state != SMC_ACTIVE) {
2292 release_sock(sk);
ac713874 2293 goto out;
bda27ff5
SR		 2294 }
2295 release_sock(sk);
ac713874
UB		 2296 if (smc->use_fallback)
2297 rc = kernel_sendpage(smc->clcsock, page, offset,
2298 size, flags);
2299 else
2300 rc = sock_no_sendpage(sock, page, offset, size, flags);
2301
2302out:
ac713874
UB		 2303 return rc;
2304}
2305
9014db20
SR		 2306/* Map the affected portions of the rmbe into an spd, note the number of bytes
2307 * to splice in conn->splice_pending, and press 'go'. Delays consumer cursor
2308 * updates till whenever a respective page has been fully processed.
2309 * Note that subsequent recv() calls have to wait till all splice() processing
2310 * completed.
2311 */
ac713874
UB		 2312static ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
2313 struct pipe_inode_info *pipe, size_t len,
9014db20 2314 unsigned int flags)
ac713874
UB		 2315{
2316 struct sock *sk = sock->sk;
2317 struct smc_sock *smc;
2318 int rc = -ENOTCONN;
2319
2320 smc = smc_sk(sk);
2321 lock_sock(sk);
51c5aba3
KG		 2322 if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
2323 /* socket was connected before, no more data to read */
2324 rc = 0;
2325 goto out;
2326 }
9014db20
SR		 2327 if (sk->sk_state == SMC_INIT ||
2328 sk->sk_state == SMC_LISTEN ||
2329 sk->sk_state == SMC_CLOSED)
2330 goto out;
2331
2332 if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
2333 rc = 0;
ac713874 2334 goto out;
9014db20
SR		 2335 }
2336
ac713874
UB		 2337 if (smc->use_fallback) {
2338 rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
2339 pipe, len, flags);
2340 } else {
9014db20
SR		 2341 if (*ppos) {
2342 rc = -ESPIPE;
2343 goto out;
2344 }
2345 if (flags & SPLICE_F_NONBLOCK)
2346 flags = MSG_DONTWAIT;
2347 else
2348 flags = 0;
2349 rc = smc_rx_recvmsg(smc, NULL, pipe, len, flags);
ac713874
UB		 2350 }
2351out:
2352 release_sock(sk);
9014db20 2353
ac713874
UB		 2354 return rc;
2355}
2356
2357/* must look like tcp */
2358static const struct proto_ops smc_sock_ops = {
2359 .family = PF_SMC,
2360 .owner = THIS_MODULE,
2361 .release = smc_release,
2362 .bind = smc_bind,
2363 .connect = smc_connect,
2364 .socketpair = sock_no_socketpair,
2365 .accept = smc_accept,
2366 .getname = smc_getname,
a11e1d43 2367 .poll = smc_poll,
ac713874
UB		 2368 .ioctl = smc_ioctl,
2369 .listen = smc_listen,
2370 .shutdown = smc_shutdown,
2371 .setsockopt = smc_setsockopt,
2372 .getsockopt = smc_getsockopt,
2373 .sendmsg = smc_sendmsg,
2374 .recvmsg = smc_recvmsg,
2375 .mmap = sock_no_mmap,
2376 .sendpage = smc_sendpage,
2377 .splice_read = smc_splice_read,
2378};
2379
2380static int smc_create(struct net *net, struct socket *sock, int protocol,
2381 int kern)
2382{
aaa4d33f 2383 int family = (protocol == SMCPROTO_SMC6) ? PF_INET6 : PF_INET;
ac713874
UB		 2384 struct smc_sock *smc;
2385 struct sock *sk;
2386 int rc;
2387
2388 rc = -ESOCKTNOSUPPORT;
2389 if (sock->type != SOCK_STREAM)
2390 goto out;
2391
2392 rc = -EPROTONOSUPPORT;
aaa4d33f 2393 if (protocol != SMCPROTO_SMC && protocol != SMCPROTO_SMC6)
ac713874
UB		 2394 goto out;
2395
2396 rc = -ENOBUFS;
2397 sock->ops = &smc_sock_ops;
aaa4d33f 2398 sk = smc_sock_alloc(net, sock, protocol);
ac713874
UB		 2399 if (!sk)
2400 goto out;
2401
2402 /* create internal TCP socket for CLC handshake and fallback */
2403 smc = smc_sk(sk);
a046d57d 2404 smc->use_fallback = false; /* assume rdma capability first */
603cc149 2405 smc->fallback_rsn = 0;
aaa4d33f
KG		 2406 rc = sock_create_kern(net, family, SOCK_STREAM, IPPROTO_TCP,
2407 &smc->clcsock);
a5dcb73b 2408 if (rc) {
ac713874 2409 sk_common_release(sk);
a5dcb73b
DC		 2410 goto out;
2411 }
cd6851f3
UB		 2412 smc->sk.sk_sndbuf = max(smc->clcsock->sk->sk_sndbuf, SMC_BUF_MIN_SIZE);
2413 smc->sk.sk_rcvbuf = max(smc->clcsock->sk->sk_rcvbuf, SMC_BUF_MIN_SIZE);
ac713874
UB		 2414
2415out:
2416 return rc;
2417}
2418
2419static const struct net_proto_family smc_sock_family_ops = {
2420 .family = PF_SMC,
2421 .owner = THIS_MODULE,
2422 .create = smc_create,
2423};
2424
64e28b52
HW		 2425unsigned int smc_net_id;
2426
2427static __net_init int smc_net_init(struct net *net)
2428{
2429 return smc_pnet_net_init(net);
2430}
2431
2432static void __net_exit smc_net_exit(struct net *net)
2433{
2434 smc_pnet_net_exit(net);
2435}
2436
2437static struct pernet_operations smc_net_ops = {
2438 .init = smc_net_init,
2439 .exit = smc_net_exit,
2440 .id = &smc_net_id,
2441 .size = sizeof(struct smc_net),
2442};
2443
ac713874
UB		 2444static int __init smc_init(void)
2445{
2446 int rc;
2447
64e28b52
HW		 2448 rc = register_pernet_subsys(&smc_net_ops);
2449 if (rc)
2450 return rc;
2451
201091eb 2452 smc_ism_init();
b81a5eb7 2453 smc_clc_init();
201091eb 2454
6812baab
TR		 2455 rc = smc_pnet_init();
2456 if (rc)
8c33bf1b 2457 goto out_pernet_subsys;
6812baab 2458
22ef473d
KG		 2459 rc = -ENOMEM;
2460 smc_hs_wq = alloc_workqueue("smc_hs_wq", 0, 0);
2461 if (!smc_hs_wq)
2462 goto out_pnet;
2463
2464 smc_close_wq = alloc_workqueue("smc_close_wq", 0, 0);
2465 if (!smc_close_wq)
2466 goto out_alloc_hs_wq;
2467
6dabd405
UB		 2468 rc = smc_core_init();
2469 if (rc) {
2470 pr_err("%s: smc_core_init fails with %d\n", __func__, rc);
22ef473d 2471 goto out_alloc_wqs;
6dabd405
UB		 2472 }
2473
9bf9abea
UB		 2474 rc = smc_llc_init();
2475 if (rc) {
2476 pr_err("%s: smc_llc_init fails with %d\n", __func__, rc);
6dabd405 2477 goto out_core;
9bf9abea
UB		 2478 }
2479
5f08318f
UB		 2480 rc = smc_cdc_init();
2481 if (rc) {
2482 pr_err("%s: smc_cdc_init fails with %d\n", __func__, rc);
6dabd405 2483 goto out_core;
5f08318f
UB		 2484 }
2485
ac713874
UB		 2486 rc = proto_register(&smc_proto, 1);
2487 if (rc) {
aaa4d33f 2488 pr_err("%s: proto_register(v4) fails with %d\n", __func__, rc);
6dabd405 2489 goto out_core;
ac713874
UB		 2490 }
2491
aaa4d33f
KG		 2492 rc = proto_register(&smc_proto6, 1);
2493 if (rc) {
2494 pr_err("%s: proto_register(v6) fails with %d\n", __func__, rc);
2495 goto out_proto;
2496 }
2497
ac713874
UB		 2498 rc = sock_register(&smc_sock_family_ops);
2499 if (rc) {
2500 pr_err("%s: sock_register fails with %d\n", __func__, rc);
aaa4d33f 2501 goto out_proto6;
ac713874 2502 }
f16a7dd5 2503 INIT_HLIST_HEAD(&smc_v4_hashinfo.ht);
aaa4d33f 2504 INIT_HLIST_HEAD(&smc_v6_hashinfo.ht);
ac713874 2505
a4cf0443
UB		 2506 rc = smc_ib_register_client();
2507 if (rc) {
2508 pr_err("%s: ib_register fails with %d\n", __func__, rc);
2509 goto out_sock;
2510 }
2511
c5c1cc9c 2512 static_branch_enable(&tcp_have_smc);
ac713874
UB		 2513 return 0;
2514
a4cf0443
UB		 2515out_sock:
2516 sock_unregister(PF_SMC);
aaa4d33f
KG		 2517out_proto6:
2518 proto_unregister(&smc_proto6);
ac713874
UB		 2519out_proto:
2520 proto_unregister(&smc_proto);
6dabd405
UB		 2521out_core:
2522 smc_core_exit();
22ef473d
KG		 2523out_alloc_wqs:
2524 destroy_workqueue(smc_close_wq);
2525out_alloc_hs_wq:
2526 destroy_workqueue(smc_hs_wq);
6812baab
TR		 2527out_pnet:
2528 smc_pnet_exit();
8c33bf1b
Y		 2529out_pernet_subsys:
2530 unregister_pernet_subsys(&smc_net_ops);
2531
ac713874
UB		 2532 return rc;
2533}
2534
2535static void __exit smc_exit(void)
2536{
c5c1cc9c 2537 static_branch_disable(&tcp_have_smc);
ac713874 2538 sock_unregister(PF_SMC);
6dabd405
UB		 2539 smc_core_exit();
2540 smc_ib_unregister_client();
22ef473d
KG		 2541 destroy_workqueue(smc_close_wq);
2542 destroy_workqueue(smc_hs_wq);
aaa4d33f 2543 proto_unregister(&smc_proto6);
ac713874 2544 proto_unregister(&smc_proto);
6812baab 2545 smc_pnet_exit();
64e28b52 2546 unregister_pernet_subsys(&smc_net_ops);
4ead9c96 2547 rcu_barrier();
ac713874
UB		 2548}
2549
2550module_init(smc_init);
2551module_exit(smc_exit);
2552
2553MODULE_AUTHOR("Ursula Braun <ubraun@linux.vnet.ibm.com>");
2554MODULE_DESCRIPTION("smc socket address family");
2555MODULE_LICENSE("GPL");
2556MODULE_ALIAS_NETPROTO(PF_SMC);
Ubuntu Jammy Linux kernel mirror