2 * Shared Memory Communications over RDMA (SMC-R) and RoCE
4 * AF_SMC protocol family socket handler keeping the AF_INET sock address type
5 * applies to SOCK_STREAM sockets only
6 * offers an alternative communication option for TCP-protocol sockets
7 * applicable with RoCE-cards only
9 * Initial restrictions:
10 * - support for alternate links postponed
12 * Copyright IBM Corp. 2016, 2018
14 * Author(s): Ursula Braun <ubraun@linux.vnet.ibm.com>
15 * based on prototype from Frank Blaschka
18 #define KMSG_COMPONENT "smc"
19 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
21 #include <linux/module.h>
22 #include <linux/socket.h>
23 #include <linux/workqueue.h>
25 #include <linux/sched/signal.h>
26 #include <linux/if_vlan.h>
31 #include <asm/ioctls.h>
43 #include "smc_close.h"
45 static DEFINE_MUTEX(smc_create_lgr_pending); /* serialize link group
49 static void smc_tcp_listen_work(struct work_struct *);
50 static void smc_connect_work(struct work_struct *);
52 static void smc_set_keepalive(struct sock *sk, int val)
54 struct smc_sock *smc = smc_sk(sk);
56 smc->clcsock->sk->sk_prot->keepalive(smc->clcsock->sk, val);
59 static struct smc_hashinfo smc_v4_hashinfo = {
60 .lock = __RW_LOCK_UNLOCKED(smc_v4_hashinfo.lock),
63 static struct smc_hashinfo smc_v6_hashinfo = {
64 .lock = __RW_LOCK_UNLOCKED(smc_v6_hashinfo.lock),
67 int smc_hash_sk(struct sock *sk)
69 struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
70 struct hlist_head *head;
74 write_lock_bh(&h->lock);
75 sk_add_node(sk, head);
76 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
77 write_unlock_bh(&h->lock);
81 EXPORT_SYMBOL_GPL(smc_hash_sk);
83 void smc_unhash_sk(struct sock *sk)
85 struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
87 write_lock_bh(&h->lock);
88 if (sk_del_node_init(sk))
89 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
90 write_unlock_bh(&h->lock);
92 EXPORT_SYMBOL_GPL(smc_unhash_sk);
94 struct proto smc_proto = {
97 .keepalive = smc_set_keepalive,
99 .unhash = smc_unhash_sk,
100 .obj_size = sizeof(struct smc_sock),
101 .h.smc_hash = &smc_v4_hashinfo,
102 .slab_flags = SLAB_TYPESAFE_BY_RCU,
104 EXPORT_SYMBOL_GPL(smc_proto);
106 struct proto smc_proto6 = {
108 .owner = THIS_MODULE,
109 .keepalive = smc_set_keepalive,
111 .unhash = smc_unhash_sk,
112 .obj_size = sizeof(struct smc_sock),
113 .h.smc_hash = &smc_v6_hashinfo,
114 .slab_flags = SLAB_TYPESAFE_BY_RCU,
116 EXPORT_SYMBOL_GPL(smc_proto6);
118 static int smc_release(struct socket *sock)
120 struct sock *sk = sock->sk;
121 struct smc_sock *smc;
129 /* cleanup for a dangling non-blocking connect */
130 flush_work(&smc->connect_work);
131 kfree(smc->connect_info);
132 smc->connect_info = NULL;
134 if (sk->sk_state == SMC_LISTEN)
135 /* smc_close_non_accepted() is called and acquires
136 * sock lock for child sockets again
138 lock_sock_nested(sk, SINGLE_DEPTH_NESTING);
142 if (!smc->use_fallback) {
143 rc = smc_close_active(smc);
144 sock_set_flag(sk, SOCK_DEAD);
145 sk->sk_shutdown |= SHUTDOWN_MASK;
148 sk->sk_prot->unhash(sk);
151 if (smc->use_fallback && sk->sk_state == SMC_LISTEN) {
152 /* wake up clcsock accept */
153 rc = kernel_sock_shutdown(smc->clcsock, SHUT_RDWR);
155 mutex_lock(&smc->clcsock_release_lock);
156 sock_release(smc->clcsock);
158 mutex_unlock(&smc->clcsock_release_lock);
160 if (smc->use_fallback) {
161 if (sk->sk_state != SMC_LISTEN && sk->sk_state != SMC_INIT)
162 sock_put(sk); /* passive closing */
163 sk->sk_state = SMC_CLOSED;
164 sk->sk_state_change(sk);
170 if (!smc->use_fallback && sk->sk_state == SMC_CLOSED)
171 smc_conn_free(&smc->conn);
174 sock_put(sk); /* final sock_put */
179 static void smc_destruct(struct sock *sk)
181 if (sk->sk_state != SMC_CLOSED)
183 if (!sock_flag(sk, SOCK_DEAD))
186 sk_refcnt_debug_dec(sk);
189 static struct sock *smc_sock_alloc(struct net *net, struct socket *sock,
192 struct smc_sock *smc;
196 prot = (protocol == SMCPROTO_SMC6) ? &smc_proto6 : &smc_proto;
197 sk = sk_alloc(net, PF_SMC, GFP_KERNEL, prot, 0);
201 sock_init_data(sock, sk); /* sets sk_refcnt to 1 */
202 sk->sk_state = SMC_INIT;
203 sk->sk_destruct = smc_destruct;
204 sk->sk_protocol = protocol;
206 INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
207 INIT_WORK(&smc->connect_work, smc_connect_work);
208 INIT_DELAYED_WORK(&smc->conn.tx_work, smc_tx_work);
209 INIT_LIST_HEAD(&smc->accept_q);
210 spin_lock_init(&smc->accept_q_lock);
211 spin_lock_init(&smc->conn.send_lock);
212 sk->sk_prot->hash(sk);
213 sk_refcnt_debug_inc(sk);
214 mutex_init(&smc->clcsock_release_lock);
219 static int smc_bind(struct socket *sock, struct sockaddr *uaddr,
222 struct sockaddr_in *addr = (struct sockaddr_in *)uaddr;
223 struct sock *sk = sock->sk;
224 struct smc_sock *smc;
229 /* replicate tests from inet_bind(), to be safe wrt. future changes */
231 if (addr_len < sizeof(struct sockaddr_in))
235 if (addr->sin_family != AF_INET &&
236 addr->sin_family != AF_INET6 &&
237 addr->sin_family != AF_UNSPEC)
239 /* accept AF_UNSPEC (mapped to AF_INET) only if s_addr is INADDR_ANY */
240 if (addr->sin_family == AF_UNSPEC &&
241 addr->sin_addr.s_addr != htonl(INADDR_ANY))
246 /* Check if socket is already active */
248 if (sk->sk_state != SMC_INIT)
251 smc->clcsock->sk->sk_reuse = sk->sk_reuse;
252 rc = kernel_bind(smc->clcsock, uaddr, addr_len);
260 static void smc_copy_sock_settings(struct sock *nsk, struct sock *osk,
263 /* options we don't get control via setsockopt for */
264 nsk->sk_type = osk->sk_type;
265 nsk->sk_sndbuf = osk->sk_sndbuf;
266 nsk->sk_rcvbuf = osk->sk_rcvbuf;
267 nsk->sk_sndtimeo = osk->sk_sndtimeo;
268 nsk->sk_rcvtimeo = osk->sk_rcvtimeo;
269 nsk->sk_mark = osk->sk_mark;
270 nsk->sk_priority = osk->sk_priority;
271 nsk->sk_rcvlowat = osk->sk_rcvlowat;
272 nsk->sk_bound_dev_if = osk->sk_bound_dev_if;
273 nsk->sk_err = osk->sk_err;
275 nsk->sk_flags &= ~mask;
276 nsk->sk_flags |= osk->sk_flags & mask;
279 #define SK_FLAGS_SMC_TO_CLC ((1UL << SOCK_URGINLINE) | \
280 (1UL << SOCK_KEEPOPEN) | \
281 (1UL << SOCK_LINGER) | \
282 (1UL << SOCK_BROADCAST) | \
283 (1UL << SOCK_TIMESTAMP) | \
284 (1UL << SOCK_DBG) | \
285 (1UL << SOCK_RCVTSTAMP) | \
286 (1UL << SOCK_RCVTSTAMPNS) | \
287 (1UL << SOCK_LOCALROUTE) | \
288 (1UL << SOCK_TIMESTAMPING_RX_SOFTWARE) | \
289 (1UL << SOCK_RXQ_OVFL) | \
290 (1UL << SOCK_WIFI_STATUS) | \
291 (1UL << SOCK_NOFCS) | \
292 (1UL << SOCK_FILTER_LOCKED))
293 /* copy only relevant settings and flags of SOL_SOCKET level from smc to
294 * clc socket (since smc is not called for these options from net/core)
296 static void smc_copy_sock_settings_to_clc(struct smc_sock *smc)
298 smc_copy_sock_settings(smc->clcsock->sk, &smc->sk, SK_FLAGS_SMC_TO_CLC);
301 #define SK_FLAGS_CLC_TO_SMC ((1UL << SOCK_URGINLINE) | \
302 (1UL << SOCK_KEEPOPEN) | \
303 (1UL << SOCK_LINGER) | \
305 /* copy only settings and flags relevant for smc from clc to smc socket */
306 static void smc_copy_sock_settings_to_smc(struct smc_sock *smc)
308 smc_copy_sock_settings(&smc->sk, smc->clcsock->sk, SK_FLAGS_CLC_TO_SMC);
311 /* register a new rmb, optionally send confirm_rkey msg to register with peer */
312 static int smc_reg_rmb(struct smc_link *link, struct smc_buf_desc *rmb_desc,
315 /* register memory region for new rmb */
316 if (smc_wr_reg_send(link, rmb_desc->mr_rx[SMC_SINGLE_LINK])) {
317 rmb_desc->regerr = 1;
322 /* exchange confirm_rkey msg with peer */
323 if (smc_llc_do_confirm_rkey(link, rmb_desc)) {
324 rmb_desc->regerr = 1;
330 static int smc_clnt_conf_first_link(struct smc_sock *smc)
332 struct net *net = sock_net(smc->clcsock->sk);
333 struct smc_link_group *lgr = smc->conn.lgr;
334 struct smc_link *link;
338 link = &lgr->lnk[SMC_SINGLE_LINK];
339 /* receive CONFIRM LINK request from server over RoCE fabric */
340 rest = wait_for_completion_interruptible_timeout(
342 SMC_LLC_WAIT_FIRST_TIME);
344 struct smc_clc_msg_decline dclc;
346 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
351 if (link->llc_confirm_rc)
352 return SMC_CLC_DECL_RMBE_EC;
354 rc = smc_ib_modify_qp_rts(link);
356 return SMC_CLC_DECL_ERR_RDYLNK;
358 smc_wr_remember_qp_attr(link);
360 if (smc_reg_rmb(link, smc->conn.rmb_desc, false))
361 return SMC_CLC_DECL_ERR_REGRMB;
363 /* send CONFIRM LINK response over RoCE fabric */
364 rc = smc_llc_send_confirm_link(link, SMC_LLC_RESP);
366 return SMC_CLC_DECL_TIMEOUT_CL;
368 /* receive ADD LINK request from server over RoCE fabric */
369 rest = wait_for_completion_interruptible_timeout(&link->llc_add,
372 struct smc_clc_msg_decline dclc;
374 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
379 /* send add link reject message, only one link supported for now */
380 rc = smc_llc_send_add_link(link,
381 link->smcibdev->mac[link->ibport - 1],
382 link->gid, SMC_LLC_RESP);
384 return SMC_CLC_DECL_TIMEOUT_AL;
386 smc_llc_link_active(link, net->ipv4.sysctl_tcp_keepalive_time);
391 static void smcr_conn_save_peer_info(struct smc_sock *smc,
392 struct smc_clc_msg_accept_confirm *clc)
394 int bufsize = smc_uncompress_bufsize(clc->rmbe_size);
396 smc->conn.peer_rmbe_idx = clc->rmbe_idx;
397 smc->conn.local_tx_ctrl.token = ntohl(clc->rmbe_alert_token);
398 smc->conn.peer_rmbe_size = bufsize;
399 atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
400 smc->conn.tx_off = bufsize * (smc->conn.peer_rmbe_idx - 1);
403 static void smcd_conn_save_peer_info(struct smc_sock *smc,
404 struct smc_clc_msg_accept_confirm *clc)
406 int bufsize = smc_uncompress_bufsize(clc->dmbe_size);
408 smc->conn.peer_rmbe_idx = clc->dmbe_idx;
409 smc->conn.peer_token = clc->token;
410 /* msg header takes up space in the buffer */
411 smc->conn.peer_rmbe_size = bufsize - sizeof(struct smcd_cdc_msg);
412 atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
413 smc->conn.tx_off = bufsize * smc->conn.peer_rmbe_idx;
416 static void smc_conn_save_peer_info(struct smc_sock *smc,
417 struct smc_clc_msg_accept_confirm *clc)
419 if (smc->conn.lgr->is_smcd)
420 smcd_conn_save_peer_info(smc, clc);
422 smcr_conn_save_peer_info(smc, clc);
425 static void smc_link_save_peer_info(struct smc_link *link,
426 struct smc_clc_msg_accept_confirm *clc)
428 link->peer_qpn = ntoh24(clc->qpn);
429 memcpy(link->peer_gid, clc->lcl.gid, SMC_GID_SIZE);
430 memcpy(link->peer_mac, clc->lcl.mac, sizeof(link->peer_mac));
431 link->peer_psn = ntoh24(clc->psn);
432 link->peer_mtu = clc->qp_mtu;
435 /* fall back during connect */
436 static int smc_connect_fallback(struct smc_sock *smc, int reason_code)
438 smc->use_fallback = true;
439 smc->fallback_rsn = reason_code;
440 smc_copy_sock_settings_to_clc(smc);
441 if (smc->sk.sk_state == SMC_INIT)
442 smc->sk.sk_state = SMC_ACTIVE;
446 /* decline and fall back during connect */
447 static int smc_connect_decline_fallback(struct smc_sock *smc, int reason_code)
451 if (reason_code < 0) { /* error, fallback is not possible */
452 if (smc->sk.sk_state == SMC_INIT)
453 sock_put(&smc->sk); /* passive closing */
456 if (reason_code != SMC_CLC_DECL_PEERDECL) {
457 rc = smc_clc_send_decline(smc, reason_code);
459 if (smc->sk.sk_state == SMC_INIT)
460 sock_put(&smc->sk); /* passive closing */
464 return smc_connect_fallback(smc, reason_code);
467 /* abort connecting */
468 static int smc_connect_abort(struct smc_sock *smc, int reason_code,
471 if (local_contact == SMC_FIRST_CONTACT)
472 smc_lgr_forget(smc->conn.lgr);
473 mutex_unlock(&smc_create_lgr_pending);
474 smc_conn_free(&smc->conn);
478 /* check if there is a rdma device available for this connection. */
479 /* called for connect and listen */
480 static int smc_check_rdma(struct smc_sock *smc, struct smc_ib_device **ibdev,
481 u8 *ibport, unsigned short vlan_id, u8 gid[])
485 /* PNET table look up: search active ib_device and port
486 * within same PNETID that also contains the ethernet device
487 * used for the internal TCP socket
489 smc_pnet_find_roce_resource(smc->clcsock->sk, ibdev, ibport, vlan_id,
492 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
497 /* check if there is an ISM device available for this connection. */
498 /* called for connect and listen */
499 static int smc_check_ism(struct smc_sock *smc, struct smcd_dev **ismdev)
501 /* Find ISM device with same PNETID as connecting interface */
502 smc_pnet_find_ism_resource(smc->clcsock->sk, ismdev);
504 return SMC_CLC_DECL_CNFERR; /* configuration error */
508 /* Check for VLAN ID and register it on ISM device just for CLC handshake */
509 static int smc_connect_ism_vlan_setup(struct smc_sock *smc,
510 struct smcd_dev *ismdev,
511 unsigned short vlan_id)
513 if (vlan_id && smc_ism_get_vlan(ismdev, vlan_id))
514 return SMC_CLC_DECL_CNFERR;
518 /* cleanup temporary VLAN ID registration used for CLC handshake. If ISM is
519 * used, the VLAN ID will be registered again during the connection setup.
521 static int smc_connect_ism_vlan_cleanup(struct smc_sock *smc, bool is_smcd,
522 struct smcd_dev *ismdev,
523 unsigned short vlan_id)
527 if (vlan_id && smc_ism_put_vlan(ismdev, vlan_id))
528 return SMC_CLC_DECL_CNFERR;
532 /* CLC handshake during connect */
533 static int smc_connect_clc(struct smc_sock *smc, int smc_type,
534 struct smc_clc_msg_accept_confirm *aclc,
535 struct smc_ib_device *ibdev, u8 ibport,
536 u8 gid[], struct smcd_dev *ismdev)
540 /* do inband token exchange */
541 rc = smc_clc_send_proposal(smc, smc_type, ibdev, ibport, gid, ismdev);
544 /* receive SMC Accept CLC message */
545 return smc_clc_wait_msg(smc, aclc, sizeof(*aclc), SMC_CLC_ACCEPT);
548 /* setup for RDMA connection of client */
549 static int smc_connect_rdma(struct smc_sock *smc,
550 struct smc_clc_msg_accept_confirm *aclc,
551 struct smc_ib_device *ibdev, u8 ibport)
553 int local_contact = SMC_FIRST_CONTACT;
554 struct smc_link *link;
557 mutex_lock(&smc_create_lgr_pending);
558 local_contact = smc_conn_create(smc, false, aclc->hdr.flag, ibdev,
559 ibport, &aclc->lcl, NULL, 0);
560 if (local_contact < 0) {
561 if (local_contact == -ENOMEM)
562 reason_code = SMC_CLC_DECL_MEM;/* insufficient memory*/
563 else if (local_contact == -ENOLINK)
564 reason_code = SMC_CLC_DECL_SYNCERR; /* synchr. error */
566 reason_code = SMC_CLC_DECL_INTERR; /* other error */
567 return smc_connect_abort(smc, reason_code, 0);
569 link = &smc->conn.lgr->lnk[SMC_SINGLE_LINK];
571 smc_conn_save_peer_info(smc, aclc);
573 /* create send buffer and rmb */
574 if (smc_buf_create(smc, false))
575 return smc_connect_abort(smc, SMC_CLC_DECL_MEM, local_contact);
577 if (local_contact == SMC_FIRST_CONTACT)
578 smc_link_save_peer_info(link, aclc);
580 if (smc_rmb_rtoken_handling(&smc->conn, aclc))
581 return smc_connect_abort(smc, SMC_CLC_DECL_ERR_RTOK,
587 if (local_contact == SMC_FIRST_CONTACT) {
588 if (smc_ib_ready_link(link))
589 return smc_connect_abort(smc, SMC_CLC_DECL_ERR_RDYLNK,
592 if (!smc->conn.rmb_desc->reused &&
593 smc_reg_rmb(link, smc->conn.rmb_desc, true))
594 return smc_connect_abort(smc, SMC_CLC_DECL_ERR_REGRMB,
597 smc_rmb_sync_sg_for_device(&smc->conn);
599 reason_code = smc_clc_send_confirm(smc);
601 return smc_connect_abort(smc, reason_code, local_contact);
605 if (local_contact == SMC_FIRST_CONTACT) {
606 /* QP confirmation over RoCE fabric */
607 reason_code = smc_clnt_conf_first_link(smc);
609 return smc_connect_abort(smc, reason_code,
612 mutex_unlock(&smc_create_lgr_pending);
614 smc_copy_sock_settings_to_clc(smc);
615 if (smc->sk.sk_state == SMC_INIT)
616 smc->sk.sk_state = SMC_ACTIVE;
621 /* setup for ISM connection of client */
622 static int smc_connect_ism(struct smc_sock *smc,
623 struct smc_clc_msg_accept_confirm *aclc,
624 struct smcd_dev *ismdev)
626 int local_contact = SMC_FIRST_CONTACT;
629 mutex_lock(&smc_create_lgr_pending);
630 local_contact = smc_conn_create(smc, true, aclc->hdr.flag, NULL, 0,
631 NULL, ismdev, aclc->gid);
632 if (local_contact < 0)
633 return smc_connect_abort(smc, SMC_CLC_DECL_MEM, 0);
635 /* Create send and receive buffers */
636 if (smc_buf_create(smc, true))
637 return smc_connect_abort(smc, SMC_CLC_DECL_MEM, local_contact);
639 smc_conn_save_peer_info(smc, aclc);
644 rc = smc_clc_send_confirm(smc);
646 return smc_connect_abort(smc, rc, local_contact);
647 mutex_unlock(&smc_create_lgr_pending);
649 smc_copy_sock_settings_to_clc(smc);
650 if (smc->sk.sk_state == SMC_INIT)
651 smc->sk.sk_state = SMC_ACTIVE;
656 /* perform steps before actually connecting */
657 static int __smc_connect(struct smc_sock *smc)
659 bool ism_supported = false, rdma_supported = false;
660 struct smc_clc_msg_accept_confirm aclc;
661 struct smc_ib_device *ibdev;
662 struct smcd_dev *ismdev;
663 u8 gid[SMC_GID_SIZE];
669 sock_hold(&smc->sk); /* sock put in passive closing */
671 if (smc->use_fallback)
672 return smc_connect_fallback(smc, smc->fallback_rsn);
674 /* if peer has not signalled SMC-capability, fall back */
675 if (!tcp_sk(smc->clcsock->sk)->syn_smc)
676 return smc_connect_fallback(smc, SMC_CLC_DECL_PEERNOSMC);
678 /* IPSec connections opt out of SMC-R optimizations */
679 if (using_ipsec(smc))
680 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_IPSEC);
682 /* check for VLAN ID */
683 if (smc_vlan_by_tcpsk(smc->clcsock, &vlan))
684 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_CNFERR);
686 /* check if there is an ism device available */
687 if (!smc_check_ism(smc, &ismdev) &&
688 !smc_connect_ism_vlan_setup(smc, ismdev, vlan)) {
689 /* ISM is supported for this connection */
690 ism_supported = true;
691 smc_type = SMC_TYPE_D;
694 /* check if there is a rdma device available */
695 if (!smc_check_rdma(smc, &ibdev, &ibport, vlan, gid)) {
696 /* RDMA is supported for this connection */
697 rdma_supported = true;
699 smc_type = SMC_TYPE_B; /* both */
701 smc_type = SMC_TYPE_R; /* only RDMA */
704 /* if neither ISM nor RDMA are supported, fallback */
705 if (!rdma_supported && !ism_supported)
706 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_NOSMCDEV);
708 /* perform CLC handshake */
709 rc = smc_connect_clc(smc, smc_type, &aclc, ibdev, ibport, gid, ismdev);
711 smc_connect_ism_vlan_cleanup(smc, ism_supported, ismdev, vlan);
712 return smc_connect_decline_fallback(smc, rc);
715 /* depending on previous steps, connect using rdma or ism */
716 if (rdma_supported && aclc.hdr.path == SMC_TYPE_R)
717 rc = smc_connect_rdma(smc, &aclc, ibdev, ibport);
718 else if (ism_supported && aclc.hdr.path == SMC_TYPE_D)
719 rc = smc_connect_ism(smc, &aclc, ismdev);
721 rc = SMC_CLC_DECL_MODEUNSUPP;
723 smc_connect_ism_vlan_cleanup(smc, ism_supported, ismdev, vlan);
724 return smc_connect_decline_fallback(smc, rc);
727 smc_connect_ism_vlan_cleanup(smc, ism_supported, ismdev, vlan);
731 static void smc_connect_work(struct work_struct *work)
733 struct smc_sock *smc = container_of(work, struct smc_sock,
738 rc = kernel_connect(smc->clcsock, &smc->connect_info->addr,
739 smc->connect_info->alen, smc->connect_info->flags);
740 if (smc->clcsock->sk->sk_err) {
741 smc->sk.sk_err = smc->clcsock->sk->sk_err;
745 smc->sk.sk_err = -rc;
749 rc = __smc_connect(smc);
751 smc->sk.sk_err = -rc;
755 smc->sk.sk_state_change(&smc->sk);
757 smc->sk.sk_write_space(&smc->sk);
758 kfree(smc->connect_info);
759 smc->connect_info = NULL;
760 release_sock(&smc->sk);
763 static int smc_connect(struct socket *sock, struct sockaddr *addr,
766 struct sock *sk = sock->sk;
767 struct smc_sock *smc;
772 /* separate smc parameter checking to be safe */
773 if (alen < sizeof(addr->sa_family))
775 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
779 switch (sk->sk_state) {
790 smc_copy_sock_settings_to_clc(smc);
791 tcp_sk(smc->clcsock->sk)->syn_smc = 1;
792 if (flags & O_NONBLOCK) {
793 if (smc->connect_info) {
797 smc->connect_info = kzalloc(alen + 2 * sizeof(int), GFP_KERNEL);
798 if (!smc->connect_info) {
802 smc->connect_info->alen = alen;
803 smc->connect_info->flags = flags ^ O_NONBLOCK;
804 memcpy(&smc->connect_info->addr, addr, alen);
805 schedule_work(&smc->connect_work);
808 rc = kernel_connect(smc->clcsock, addr, alen, flags);
812 rc = __smc_connect(smc);
816 rc = 0; /* success cases including fallback */
825 static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
827 struct socket *new_clcsock = NULL;
828 struct sock *lsk = &lsmc->sk;
833 new_sk = smc_sock_alloc(sock_net(lsk), NULL, lsk->sk_protocol);
836 lsk->sk_err = ENOMEM;
841 *new_smc = smc_sk(new_sk);
843 mutex_lock(&lsmc->clcsock_release_lock);
845 rc = kernel_accept(lsmc->clcsock, &new_clcsock, 0);
846 mutex_unlock(&lsmc->clcsock_release_lock);
850 if (rc < 0 || lsk->sk_state == SMC_CLOSED) {
851 new_sk->sk_prot->unhash(new_sk);
853 sock_release(new_clcsock);
854 new_sk->sk_state = SMC_CLOSED;
855 sock_set_flag(new_sk, SOCK_DEAD);
856 sock_put(new_sk); /* final */
861 (*new_smc)->clcsock = new_clcsock;
866 /* add a just created sock to the accept queue of the listen sock as
867 * candidate for a following socket accept call from user space
869 static void smc_accept_enqueue(struct sock *parent, struct sock *sk)
871 struct smc_sock *par = smc_sk(parent);
873 sock_hold(sk); /* sock_put in smc_accept_unlink () */
874 spin_lock(&par->accept_q_lock);
875 list_add_tail(&smc_sk(sk)->accept_q, &par->accept_q);
876 spin_unlock(&par->accept_q_lock);
877 sk_acceptq_added(parent);
880 /* remove a socket from the accept queue of its parental listening socket */
881 static void smc_accept_unlink(struct sock *sk)
883 struct smc_sock *par = smc_sk(sk)->listen_smc;
885 spin_lock(&par->accept_q_lock);
886 list_del_init(&smc_sk(sk)->accept_q);
887 spin_unlock(&par->accept_q_lock);
888 sk_acceptq_removed(&smc_sk(sk)->listen_smc->sk);
889 sock_put(sk); /* sock_hold in smc_accept_enqueue */
892 /* remove a sock from the accept queue to bind it to a new socket created
893 * for a socket accept call from user space
895 struct sock *smc_accept_dequeue(struct sock *parent,
896 struct socket *new_sock)
898 struct smc_sock *isk, *n;
901 list_for_each_entry_safe(isk, n, &smc_sk(parent)->accept_q, accept_q) {
902 new_sk = (struct sock *)isk;
904 smc_accept_unlink(new_sk);
905 if (new_sk->sk_state == SMC_CLOSED) {
906 new_sk->sk_prot->unhash(new_sk);
908 sock_release(isk->clcsock);
911 sock_put(new_sk); /* final */
915 sock_graft(new_sk, new_sock);
921 /* clean up for a created but never accepted sock */
922 void smc_close_non_accepted(struct sock *sk)
924 struct smc_sock *smc = smc_sk(sk);
927 if (!sk->sk_lingertime)
928 /* wait for peer closing */
929 sk->sk_lingertime = SMC_MAX_STREAM_WAIT_TIMEOUT;
930 if (!smc->use_fallback) {
931 smc_close_active(smc);
932 sock_set_flag(sk, SOCK_DEAD);
933 sk->sk_shutdown |= SHUTDOWN_MASK;
935 sk->sk_prot->unhash(sk);
943 if (smc->use_fallback) {
944 sock_put(sk); /* passive closing */
945 sk->sk_state = SMC_CLOSED;
947 if (sk->sk_state == SMC_CLOSED)
948 smc_conn_free(&smc->conn);
951 sock_put(sk); /* final sock_put */
954 static int smc_serv_conf_first_link(struct smc_sock *smc)
956 struct net *net = sock_net(smc->clcsock->sk);
957 struct smc_link_group *lgr = smc->conn.lgr;
958 struct smc_link *link;
962 link = &lgr->lnk[SMC_SINGLE_LINK];
964 if (smc_reg_rmb(link, smc->conn.rmb_desc, false))
965 return SMC_CLC_DECL_ERR_REGRMB;
967 /* send CONFIRM LINK request to client over the RoCE fabric */
968 rc = smc_llc_send_confirm_link(link, SMC_LLC_REQ);
970 return SMC_CLC_DECL_TIMEOUT_CL;
972 /* receive CONFIRM LINK response from client over the RoCE fabric */
973 rest = wait_for_completion_interruptible_timeout(
974 &link->llc_confirm_resp,
975 SMC_LLC_WAIT_FIRST_TIME);
977 struct smc_clc_msg_decline dclc;
979 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
984 if (link->llc_confirm_resp_rc)
985 return SMC_CLC_DECL_RMBE_EC;
987 /* send ADD LINK request to client over the RoCE fabric */
988 rc = smc_llc_send_add_link(link,
989 link->smcibdev->mac[link->ibport - 1],
990 link->gid, SMC_LLC_REQ);
992 return SMC_CLC_DECL_TIMEOUT_AL;
994 /* receive ADD LINK response from client over the RoCE fabric */
995 rest = wait_for_completion_interruptible_timeout(&link->llc_add_resp,
998 struct smc_clc_msg_decline dclc;
1000 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
1005 smc_llc_link_active(link, net->ipv4.sysctl_tcp_keepalive_time);
1010 /* listen worker: finish */
1011 static void smc_listen_out(struct smc_sock *new_smc)
1013 struct smc_sock *lsmc = new_smc->listen_smc;
1014 struct sock *newsmcsk = &new_smc->sk;
1016 lock_sock_nested(&lsmc->sk, SINGLE_DEPTH_NESTING);
1017 if (lsmc->sk.sk_state == SMC_LISTEN) {
1018 smc_accept_enqueue(&lsmc->sk, newsmcsk);
1019 } else { /* no longer listening */
1020 smc_close_non_accepted(newsmcsk);
1022 release_sock(&lsmc->sk);
1024 /* Wake up accept */
1025 lsmc->sk.sk_data_ready(&lsmc->sk);
1026 sock_put(&lsmc->sk); /* sock_hold in smc_tcp_listen_work */
1029 /* listen worker: finish in state connected */
1030 static void smc_listen_out_connected(struct smc_sock *new_smc)
1032 struct sock *newsmcsk = &new_smc->sk;
1034 sk_refcnt_debug_inc(newsmcsk);
1035 if (newsmcsk->sk_state == SMC_INIT)
1036 newsmcsk->sk_state = SMC_ACTIVE;
1038 smc_listen_out(new_smc);
1041 /* listen worker: finish in error state */
1042 static void smc_listen_out_err(struct smc_sock *new_smc)
1044 struct sock *newsmcsk = &new_smc->sk;
1046 if (newsmcsk->sk_state == SMC_INIT)
1047 sock_put(&new_smc->sk); /* passive closing */
1048 newsmcsk->sk_state = SMC_CLOSED;
1049 smc_conn_free(&new_smc->conn);
1051 smc_listen_out(new_smc);
1054 /* listen worker: decline and fall back if possible */
1055 static void smc_listen_decline(struct smc_sock *new_smc, int reason_code,
1058 /* RDMA setup failed, switch back to TCP */
1059 if (local_contact == SMC_FIRST_CONTACT)
1060 smc_lgr_forget(new_smc->conn.lgr);
1061 if (reason_code < 0) { /* error, no fallback possible */
1062 smc_listen_out_err(new_smc);
1065 smc_conn_free(&new_smc->conn);
1066 new_smc->use_fallback = true;
1067 new_smc->fallback_rsn = reason_code;
1068 if (reason_code && reason_code != SMC_CLC_DECL_PEERDECL) {
1069 if (smc_clc_send_decline(new_smc, reason_code) < 0) {
1070 smc_listen_out_err(new_smc);
1074 smc_listen_out_connected(new_smc);
1077 /* listen worker: check prefixes */
1078 static int smc_listen_rdma_check(struct smc_sock *new_smc,
1079 struct smc_clc_msg_proposal *pclc)
1081 struct smc_clc_msg_proposal_prefix *pclc_prfx;
1082 struct socket *newclcsock = new_smc->clcsock;
1084 pclc_prfx = smc_clc_proposal_get_prefix(pclc);
1085 if (smc_clc_prfx_match(newclcsock, pclc_prfx))
1086 return SMC_CLC_DECL_CNFERR;
1091 /* listen worker: initialize connection and buffers */
1092 static int smc_listen_rdma_init(struct smc_sock *new_smc,
1093 struct smc_clc_msg_proposal *pclc,
1094 struct smc_ib_device *ibdev, u8 ibport,
1097 /* allocate connection / link group */
1098 *local_contact = smc_conn_create(new_smc, false, 0, ibdev, ibport,
1099 &pclc->lcl, NULL, 0);
1100 if (*local_contact < 0) {
1101 if (*local_contact == -ENOMEM)
1102 return SMC_CLC_DECL_MEM;/* insufficient memory*/
1103 return SMC_CLC_DECL_INTERR; /* other error */
1106 /* create send buffer and rmb */
1107 if (smc_buf_create(new_smc, false))
1108 return SMC_CLC_DECL_MEM;
1113 /* listen worker: initialize connection and buffers for SMC-D */
1114 static int smc_listen_ism_init(struct smc_sock *new_smc,
1115 struct smc_clc_msg_proposal *pclc,
1116 struct smcd_dev *ismdev,
1119 struct smc_clc_msg_smcd *pclc_smcd;
1121 pclc_smcd = smc_get_clc_msg_smcd(pclc);
1122 *local_contact = smc_conn_create(new_smc, true, 0, NULL, 0, NULL,
1123 ismdev, pclc_smcd->gid);
1124 if (*local_contact < 0) {
1125 if (*local_contact == -ENOMEM)
1126 return SMC_CLC_DECL_MEM;/* insufficient memory*/
1127 return SMC_CLC_DECL_INTERR; /* other error */
1130 /* Check if peer can be reached via ISM device */
1131 if (smc_ism_cantalk(new_smc->conn.lgr->peer_gid,
1132 new_smc->conn.lgr->vlan_id,
1133 new_smc->conn.lgr->smcd)) {
1134 if (*local_contact == SMC_FIRST_CONTACT)
1135 smc_lgr_forget(new_smc->conn.lgr);
1136 smc_conn_free(&new_smc->conn);
1137 return SMC_CLC_DECL_CNFERR;
1140 /* Create send and receive buffers */
1141 if (smc_buf_create(new_smc, true)) {
1142 if (*local_contact == SMC_FIRST_CONTACT)
1143 smc_lgr_forget(new_smc->conn.lgr);
1144 smc_conn_free(&new_smc->conn);
1145 return SMC_CLC_DECL_MEM;
1151 /* listen worker: register buffers */
1152 static int smc_listen_rdma_reg(struct smc_sock *new_smc, int local_contact)
1154 struct smc_link *link = &new_smc->conn.lgr->lnk[SMC_SINGLE_LINK];
1156 if (local_contact != SMC_FIRST_CONTACT) {
1157 if (!new_smc->conn.rmb_desc->reused) {
1158 if (smc_reg_rmb(link, new_smc->conn.rmb_desc, true))
1159 return SMC_CLC_DECL_ERR_REGRMB;
1162 smc_rmb_sync_sg_for_device(&new_smc->conn);
1167 /* listen worker: finish RDMA setup */
1168 static int smc_listen_rdma_finish(struct smc_sock *new_smc,
1169 struct smc_clc_msg_accept_confirm *cclc,
1172 struct smc_link *link = &new_smc->conn.lgr->lnk[SMC_SINGLE_LINK];
1173 int reason_code = 0;
1175 if (local_contact == SMC_FIRST_CONTACT)
1176 smc_link_save_peer_info(link, cclc);
1178 if (smc_rmb_rtoken_handling(&new_smc->conn, cclc)) {
1179 reason_code = SMC_CLC_DECL_ERR_RTOK;
1183 if (local_contact == SMC_FIRST_CONTACT) {
1184 if (smc_ib_ready_link(link)) {
1185 reason_code = SMC_CLC_DECL_ERR_RDYLNK;
1188 /* QP confirmation over RoCE fabric */
1189 reason_code = smc_serv_conf_first_link(new_smc);
1196 mutex_unlock(&smc_create_lgr_pending);
1197 smc_listen_decline(new_smc, reason_code, local_contact);
1201 /* setup for RDMA connection of server */
1202 static void smc_listen_work(struct work_struct *work)
1204 struct smc_sock *new_smc = container_of(work, struct smc_sock,
1206 struct socket *newclcsock = new_smc->clcsock;
1207 struct smc_clc_msg_accept_confirm cclc;
1208 struct smc_clc_msg_proposal *pclc;
1209 struct smc_ib_device *ibdev;
1210 bool ism_supported = false;
1211 struct smcd_dev *ismdev;
1212 u8 buf[SMC_CLC_MAX_LEN];
1213 int local_contact = 0;
1214 unsigned short vlan;
1215 int reason_code = 0;
1219 if (new_smc->use_fallback) {
1220 smc_listen_out_connected(new_smc);
1224 /* check if peer is smc capable */
1225 if (!tcp_sk(newclcsock->sk)->syn_smc) {
1226 new_smc->use_fallback = true;
1227 new_smc->fallback_rsn = SMC_CLC_DECL_PEERNOSMC;
1228 smc_listen_out_connected(new_smc);
1232 /* do inband token exchange -
1233 * wait for and receive SMC Proposal CLC message
1235 pclc = (struct smc_clc_msg_proposal *)&buf;
1236 reason_code = smc_clc_wait_msg(new_smc, pclc, SMC_CLC_MAX_LEN,
1239 smc_listen_decline(new_smc, reason_code, 0);
1243 /* IPSec connections opt out of SMC-R optimizations */
1244 if (using_ipsec(new_smc)) {
1245 smc_listen_decline(new_smc, SMC_CLC_DECL_IPSEC, 0);
1249 mutex_lock(&smc_create_lgr_pending);
1250 smc_close_init(new_smc);
1251 smc_rx_init(new_smc);
1252 smc_tx_init(new_smc);
1254 /* check if ISM is available */
1255 if ((pclc->hdr.path == SMC_TYPE_D || pclc->hdr.path == SMC_TYPE_B) &&
1256 !smc_check_ism(new_smc, &ismdev) &&
1257 !smc_listen_ism_init(new_smc, pclc, ismdev, &local_contact)) {
1258 ism_supported = true;
1261 /* check if RDMA is available */
1262 if (!ism_supported &&
1263 ((pclc->hdr.path != SMC_TYPE_R && pclc->hdr.path != SMC_TYPE_B) ||
1264 smc_vlan_by_tcpsk(new_smc->clcsock, &vlan) ||
1265 smc_check_rdma(new_smc, &ibdev, &ibport, vlan, NULL) ||
1266 smc_listen_rdma_check(new_smc, pclc) ||
1267 smc_listen_rdma_init(new_smc, pclc, ibdev, ibport,
1269 smc_listen_rdma_reg(new_smc, local_contact))) {
1270 /* SMC not supported, decline */
1271 mutex_unlock(&smc_create_lgr_pending);
1272 smc_listen_decline(new_smc, SMC_CLC_DECL_MODEUNSUPP,
1277 /* send SMC Accept CLC message */
1278 rc = smc_clc_send_accept(new_smc, local_contact);
1280 mutex_unlock(&smc_create_lgr_pending);
1281 smc_listen_decline(new_smc, rc, local_contact);
1285 /* receive SMC Confirm CLC message */
1286 reason_code = smc_clc_wait_msg(new_smc, &cclc, sizeof(cclc),
1289 mutex_unlock(&smc_create_lgr_pending);
1290 smc_listen_decline(new_smc, reason_code, local_contact);
1295 if (!ism_supported) {
1296 if (smc_listen_rdma_finish(new_smc, &cclc, local_contact))
1299 smc_conn_save_peer_info(new_smc, &cclc);
1300 mutex_unlock(&smc_create_lgr_pending);
1301 smc_listen_out_connected(new_smc);
1304 static void smc_tcp_listen_work(struct work_struct *work)
1306 struct smc_sock *lsmc = container_of(work, struct smc_sock,
1308 struct sock *lsk = &lsmc->sk;
1309 struct smc_sock *new_smc;
1313 while (lsk->sk_state == SMC_LISTEN) {
1314 rc = smc_clcsock_accept(lsmc, &new_smc);
1320 new_smc->listen_smc = lsmc;
1321 new_smc->use_fallback = lsmc->use_fallback;
1322 new_smc->fallback_rsn = lsmc->fallback_rsn;
1323 sock_hold(lsk); /* sock_put in smc_listen_work */
1324 INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
1325 smc_copy_sock_settings_to_smc(new_smc);
1326 new_smc->sk.sk_sndbuf = lsmc->sk.sk_sndbuf;
1327 new_smc->sk.sk_rcvbuf = lsmc->sk.sk_rcvbuf;
1328 sock_hold(&new_smc->sk); /* sock_put in passive closing */
1329 if (!schedule_work(&new_smc->smc_listen_work))
1330 sock_put(&new_smc->sk);
1335 sock_put(&lsmc->sk); /* sock_hold in smc_listen */
1338 static int smc_listen(struct socket *sock, int backlog)
1340 struct sock *sk = sock->sk;
1341 struct smc_sock *smc;
1348 if ((sk->sk_state != SMC_INIT) && (sk->sk_state != SMC_LISTEN))
1352 if (sk->sk_state == SMC_LISTEN) {
1353 sk->sk_max_ack_backlog = backlog;
1356 /* some socket options are handled in core, so we could not apply
1357 * them to the clc socket -- copy smc socket options to clc socket
1359 smc_copy_sock_settings_to_clc(smc);
1360 if (!smc->use_fallback)
1361 tcp_sk(smc->clcsock->sk)->syn_smc = 1;
1363 rc = kernel_listen(smc->clcsock, backlog);
1366 sk->sk_max_ack_backlog = backlog;
1367 sk->sk_ack_backlog = 0;
1368 sk->sk_state = SMC_LISTEN;
1369 INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
1370 sock_hold(sk); /* sock_hold in tcp_listen_worker */
1371 if (!schedule_work(&smc->tcp_listen_work))
1379 static int smc_accept(struct socket *sock, struct socket *new_sock,
1380 int flags, bool kern)
1382 struct sock *sk = sock->sk, *nsk;
1383 DECLARE_WAITQUEUE(wait, current);
1384 struct smc_sock *lsmc;
1389 sock_hold(sk); /* sock_put below */
1392 if (lsmc->sk.sk_state != SMC_LISTEN) {
1398 /* Wait for an incoming connection */
1399 timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
1400 add_wait_queue_exclusive(sk_sleep(sk), &wait);
1401 while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
1402 set_current_state(TASK_INTERRUPTIBLE);
1408 timeo = schedule_timeout(timeo);
1409 /* wakeup by sk_data_ready in smc_listen_work() */
1410 sched_annotate_sleep();
1412 if (signal_pending(current)) {
1413 rc = sock_intr_errno(timeo);
1417 set_current_state(TASK_RUNNING);
1418 remove_wait_queue(sk_sleep(sk), &wait);
1421 rc = sock_error(nsk);
1426 if (lsmc->sockopt_defer_accept && !(flags & O_NONBLOCK)) {
1427 /* wait till data arrives on the socket */
1428 timeo = msecs_to_jiffies(lsmc->sockopt_defer_accept *
1430 if (smc_sk(nsk)->use_fallback) {
1431 struct sock *clcsk = smc_sk(nsk)->clcsock->sk;
1434 if (skb_queue_empty(&clcsk->sk_receive_queue))
1435 sk_wait_data(clcsk, &timeo, NULL);
1436 release_sock(clcsk);
1437 } else if (!atomic_read(&smc_sk(nsk)->conn.bytes_to_rcv)) {
1439 smc_rx_wait(smc_sk(nsk), &timeo, smc_rx_data_available);
1445 sock_put(sk); /* sock_hold above */
1449 static int smc_getname(struct socket *sock, struct sockaddr *addr,
1452 struct smc_sock *smc;
1454 if (peer && (sock->sk->sk_state != SMC_ACTIVE) &&
1455 (sock->sk->sk_state != SMC_APPCLOSEWAIT1))
1458 smc = smc_sk(sock->sk);
1460 return smc->clcsock->ops->getname(smc->clcsock, addr, peer);
1463 static int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
1465 struct sock *sk = sock->sk;
1466 struct smc_sock *smc;
1471 if ((sk->sk_state != SMC_ACTIVE) &&
1472 (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1473 (sk->sk_state != SMC_INIT))
1476 if (msg->msg_flags & MSG_FASTOPEN) {
1477 if (sk->sk_state == SMC_INIT) {
1478 smc->use_fallback = true;
1479 smc->fallback_rsn = SMC_CLC_DECL_OPTUNSUPP;
1486 if (smc->use_fallback)
1487 rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
1489 rc = smc_tx_sendmsg(smc, msg, len);
1495 static int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
1498 struct sock *sk = sock->sk;
1499 struct smc_sock *smc;
1504 if ((sk->sk_state == SMC_INIT) ||
1505 (sk->sk_state == SMC_LISTEN) ||
1506 (sk->sk_state == SMC_CLOSED))
1509 if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
1514 if (smc->use_fallback) {
1515 rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
1517 msg->msg_namelen = 0;
1518 rc = smc_rx_recvmsg(smc, msg, NULL, len, flags);
1526 static __poll_t smc_accept_poll(struct sock *parent)
1528 struct smc_sock *isk = smc_sk(parent);
1531 spin_lock(&isk->accept_q_lock);
1532 if (!list_empty(&isk->accept_q))
1533 mask = EPOLLIN | EPOLLRDNORM;
1534 spin_unlock(&isk->accept_q_lock);
1539 static __poll_t smc_poll(struct file *file, struct socket *sock,
1542 struct sock *sk = sock->sk;
1544 struct smc_sock *smc;
1549 smc = smc_sk(sock->sk);
1550 if (smc->use_fallback) {
1551 /* delegate to CLC child sock */
1552 mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
1553 sk->sk_err = smc->clcsock->sk->sk_err;
1557 if (sk->sk_state != SMC_CLOSED)
1558 sock_poll_wait(file, sock, wait);
1561 if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
1562 (sk->sk_state == SMC_CLOSED))
1564 if (sk->sk_state == SMC_LISTEN) {
1565 /* woken up by sk_data_ready in smc_listen_work() */
1566 mask = smc_accept_poll(sk);
1568 if (atomic_read(&smc->conn.sndbuf_space) ||
1569 sk->sk_shutdown & SEND_SHUTDOWN) {
1570 mask |= EPOLLOUT | EPOLLWRNORM;
1572 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1573 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1575 if (atomic_read(&smc->conn.bytes_to_rcv))
1576 mask |= EPOLLIN | EPOLLRDNORM;
1577 if (sk->sk_shutdown & RCV_SHUTDOWN)
1578 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
1579 if (sk->sk_state == SMC_APPCLOSEWAIT1)
1581 if (smc->conn.urg_state == SMC_URG_VALID)
1589 static int smc_shutdown(struct socket *sock, int how)
1591 struct sock *sk = sock->sk;
1592 bool do_shutdown = true;
1593 struct smc_sock *smc;
1600 if ((how < SHUT_RD) || (how > SHUT_RDWR))
1606 if ((sk->sk_state != SMC_ACTIVE) &&
1607 (sk->sk_state != SMC_PEERCLOSEWAIT1) &&
1608 (sk->sk_state != SMC_PEERCLOSEWAIT2) &&
1609 (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1610 (sk->sk_state != SMC_APPCLOSEWAIT2) &&
1611 (sk->sk_state != SMC_APPFINCLOSEWAIT))
1613 if (smc->use_fallback) {
1614 rc = kernel_sock_shutdown(smc->clcsock, how);
1615 sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
1616 if (sk->sk_shutdown == SHUTDOWN_MASK)
1617 sk->sk_state = SMC_CLOSED;
1621 case SHUT_RDWR: /* shutdown in both directions */
1622 old_state = sk->sk_state;
1623 rc = smc_close_active(smc);
1624 if (old_state == SMC_ACTIVE &&
1625 sk->sk_state == SMC_PEERCLOSEWAIT1)
1626 do_shutdown = false;
1629 rc = smc_close_shutdown_write(smc);
1633 /* nothing more to do because peer is not involved */
1636 if (do_shutdown && smc->clcsock)
1637 rc1 = kernel_sock_shutdown(smc->clcsock, how);
1638 /* map sock_shutdown_cmd constants to sk_shutdown value range */
1639 sk->sk_shutdown |= how + 1;
1643 return rc ? rc : rc1;
1646 static int smc_setsockopt(struct socket *sock, int level, int optname,
1647 char __user *optval, unsigned int optlen)
1649 struct sock *sk = sock->sk;
1650 struct smc_sock *smc;
1653 if (level == SOL_TCP && optname == TCP_ULP)
1658 /* generic setsockopts reaching us here always apply to the
1661 rc = smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
1663 if (smc->clcsock->sk->sk_err) {
1664 sk->sk_err = smc->clcsock->sk->sk_err;
1665 sk->sk_error_report(sk);
1670 if (optlen < sizeof(int))
1672 if (get_user(val, (int __user *)optval))
1678 case TCP_FASTOPEN_CONNECT:
1679 case TCP_FASTOPEN_KEY:
1680 case TCP_FASTOPEN_NO_COOKIE:
1681 /* option not supported by SMC */
1682 if (sk->sk_state == SMC_INIT) {
1683 smc->use_fallback = true;
1684 smc->fallback_rsn = SMC_CLC_DECL_OPTUNSUPP;
1686 if (!smc->use_fallback)
1691 if (sk->sk_state != SMC_INIT &&
1692 sk->sk_state != SMC_LISTEN &&
1693 sk->sk_state != SMC_CLOSED) {
1694 if (val && !smc->use_fallback)
1695 mod_delayed_work(system_wq, &smc->conn.tx_work,
1700 if (sk->sk_state != SMC_INIT &&
1701 sk->sk_state != SMC_LISTEN &&
1702 sk->sk_state != SMC_CLOSED) {
1703 if (!val && !smc->use_fallback)
1704 mod_delayed_work(system_wq, &smc->conn.tx_work,
1708 case TCP_DEFER_ACCEPT:
1709 smc->sockopt_defer_accept = val;
1719 static int smc_getsockopt(struct socket *sock, int level, int optname,
1720 char __user *optval, int __user *optlen)
1722 struct smc_sock *smc;
1724 smc = smc_sk(sock->sk);
1725 /* socket options apply to the CLC socket */
1726 return smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
1730 static int smc_ioctl(struct socket *sock, unsigned int cmd,
1733 union smc_host_cursor cons, urg;
1734 struct smc_connection *conn;
1735 struct smc_sock *smc;
1738 smc = smc_sk(sock->sk);
1740 lock_sock(&smc->sk);
1741 if (smc->use_fallback) {
1742 if (!smc->clcsock) {
1743 release_sock(&smc->sk);
1746 answ = smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
1747 release_sock(&smc->sk);
1751 case SIOCINQ: /* same as FIONREAD */
1752 if (smc->sk.sk_state == SMC_LISTEN) {
1753 release_sock(&smc->sk);
1756 if (smc->sk.sk_state == SMC_INIT ||
1757 smc->sk.sk_state == SMC_CLOSED)
1760 answ = atomic_read(&smc->conn.bytes_to_rcv);
1763 /* output queue size (not send + not acked) */
1764 if (smc->sk.sk_state == SMC_LISTEN) {
1765 release_sock(&smc->sk);
1768 if (smc->sk.sk_state == SMC_INIT ||
1769 smc->sk.sk_state == SMC_CLOSED)
1772 answ = smc->conn.sndbuf_desc->len -
1773 atomic_read(&smc->conn.sndbuf_space);
1776 /* output queue size (not send only) */
1777 if (smc->sk.sk_state == SMC_LISTEN) {
1778 release_sock(&smc->sk);
1781 if (smc->sk.sk_state == SMC_INIT ||
1782 smc->sk.sk_state == SMC_CLOSED)
1785 answ = smc_tx_prepared_sends(&smc->conn);
1788 if (smc->sk.sk_state == SMC_LISTEN) {
1789 release_sock(&smc->sk);
1792 if (smc->sk.sk_state == SMC_INIT ||
1793 smc->sk.sk_state == SMC_CLOSED) {
1796 smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
1797 smc_curs_copy(&urg, &conn->urg_curs, conn);
1798 answ = smc_curs_diff(conn->rmb_desc->len,
1803 release_sock(&smc->sk);
1804 return -ENOIOCTLCMD;
1806 release_sock(&smc->sk);
1808 return put_user(answ, (int __user *)arg);
1811 static ssize_t smc_sendpage(struct socket *sock, struct page *page,
1812 int offset, size_t size, int flags)
1814 struct sock *sk = sock->sk;
1815 struct smc_sock *smc;
1820 if (sk->sk_state != SMC_ACTIVE) {
1825 if (smc->use_fallback)
1826 rc = kernel_sendpage(smc->clcsock, page, offset,
1829 rc = sock_no_sendpage(sock, page, offset, size, flags);
1835 /* Map the affected portions of the rmbe into an spd, note the number of bytes
1836 * to splice in conn->splice_pending, and press 'go'. Delays consumer cursor
1837 * updates till whenever a respective page has been fully processed.
1838 * Note that subsequent recv() calls have to wait till all splice() processing
1841 static ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
1842 struct pipe_inode_info *pipe, size_t len,
1845 struct sock *sk = sock->sk;
1846 struct smc_sock *smc;
1852 if (sk->sk_state == SMC_INIT ||
1853 sk->sk_state == SMC_LISTEN ||
1854 sk->sk_state == SMC_CLOSED)
1857 if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
1862 if (smc->use_fallback) {
1863 rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
1870 if (flags & SPLICE_F_NONBLOCK)
1871 flags = MSG_DONTWAIT;
1874 rc = smc_rx_recvmsg(smc, NULL, pipe, len, flags);
1882 /* must look like tcp */
1883 static const struct proto_ops smc_sock_ops = {
1885 .owner = THIS_MODULE,
1886 .release = smc_release,
1888 .connect = smc_connect,
1889 .socketpair = sock_no_socketpair,
1890 .accept = smc_accept,
1891 .getname = smc_getname,
1894 .listen = smc_listen,
1895 .shutdown = smc_shutdown,
1896 .setsockopt = smc_setsockopt,
1897 .getsockopt = smc_getsockopt,
1898 .sendmsg = smc_sendmsg,
1899 .recvmsg = smc_recvmsg,
1900 .mmap = sock_no_mmap,
1901 .sendpage = smc_sendpage,
1902 .splice_read = smc_splice_read,
1905 static int smc_create(struct net *net, struct socket *sock, int protocol,
1908 int family = (protocol == SMCPROTO_SMC6) ? PF_INET6 : PF_INET;
1909 struct smc_sock *smc;
1913 rc = -ESOCKTNOSUPPORT;
1914 if (sock->type != SOCK_STREAM)
1917 rc = -EPROTONOSUPPORT;
1918 if (protocol != SMCPROTO_SMC && protocol != SMCPROTO_SMC6)
1922 sock->ops = &smc_sock_ops;
1923 sk = smc_sock_alloc(net, sock, protocol);
1927 /* create internal TCP socket for CLC handshake and fallback */
1929 smc->use_fallback = false; /* assume rdma capability first */
1930 smc->fallback_rsn = 0;
1931 rc = sock_create_kern(net, family, SOCK_STREAM, IPPROTO_TCP,
1934 sk_common_release(sk);
1937 smc->sk.sk_sndbuf = max(smc->clcsock->sk->sk_sndbuf, SMC_BUF_MIN_SIZE);
1938 smc->sk.sk_rcvbuf = max(smc->clcsock->sk->sk_rcvbuf, SMC_BUF_MIN_SIZE);
1944 static const struct net_proto_family smc_sock_family_ops = {
1946 .owner = THIS_MODULE,
1947 .create = smc_create,
1950 static int __init smc_init(void)
1954 rc = smc_pnet_init();
1958 rc = smc_llc_init();
1960 pr_err("%s: smc_llc_init fails with %d\n", __func__, rc);
1964 rc = smc_cdc_init();
1966 pr_err("%s: smc_cdc_init fails with %d\n", __func__, rc);
1970 rc = proto_register(&smc_proto, 1);
1972 pr_err("%s: proto_register(v4) fails with %d\n", __func__, rc);
1976 rc = proto_register(&smc_proto6, 1);
1978 pr_err("%s: proto_register(v6) fails with %d\n", __func__, rc);
1982 rc = sock_register(&smc_sock_family_ops);
1984 pr_err("%s: sock_register fails with %d\n", __func__, rc);
1987 INIT_HLIST_HEAD(&smc_v4_hashinfo.ht);
1988 INIT_HLIST_HEAD(&smc_v6_hashinfo.ht);
1990 rc = smc_ib_register_client();
1992 pr_err("%s: ib_register fails with %d\n", __func__, rc);
1996 static_branch_enable(&tcp_have_smc);
2000 sock_unregister(PF_SMC);
2002 proto_unregister(&smc_proto6);
2004 proto_unregister(&smc_proto);
2010 static void __exit smc_exit(void)
2013 static_branch_disable(&tcp_have_smc);
2014 smc_ib_unregister_client();
2015 sock_unregister(PF_SMC);
2016 proto_unregister(&smc_proto6);
2017 proto_unregister(&smc_proto);
2021 module_init(smc_init);
2022 module_exit(smc_exit);
2024 MODULE_AUTHOR("Ursula Braun <ubraun@linux.vnet.ibm.com>");
2025 MODULE_DESCRIPTION("smc socket address family");
2026 MODULE_LICENSE("GPL");
2027 MODULE_ALIAS_NETPROTO(PF_SMC);