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 * - non-blocking connect postponed
11 * - IPv6 support postponed
12 * - support for alternate links postponed
13 * - partial support for non-blocking sockets only
14 * - support for urgent data postponed
16 * Copyright IBM Corp. 2016
18 * Author(s): Ursula Braun <ubraun@linux.vnet.ibm.com>
19 * based on prototype from Frank Blaschka
22 #define KMSG_COMPONENT "smc"
23 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
25 #include <linux/module.h>
26 #include <linux/socket.h>
27 #include <linux/inetdevice.h>
28 #include <linux/workqueue.h>
30 #include <linux/sched/signal.h>
45 #include "smc_close.h"
47 static DEFINE_MUTEX(smc_create_lgr_pending); /* serialize link group
51 struct smc_lgr_list smc_lgr_list = { /* established link groups */
52 .lock = __SPIN_LOCK_UNLOCKED(smc_lgr_list.lock),
53 .list = LIST_HEAD_INIT(smc_lgr_list.list),
56 static void smc_tcp_listen_work(struct work_struct *);
58 static void smc_set_keepalive(struct sock *sk, int val)
60 struct smc_sock *smc = smc_sk(sk);
62 smc->clcsock->sk->sk_prot->keepalive(smc->clcsock->sk, val);
65 static struct smc_hashinfo smc_v4_hashinfo = {
66 .lock = __RW_LOCK_UNLOCKED(smc_v4_hashinfo.lock),
69 int smc_hash_sk(struct sock *sk)
71 struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
72 struct hlist_head *head;
76 write_lock_bh(&h->lock);
77 sk_add_node(sk, head);
78 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
79 write_unlock_bh(&h->lock);
83 EXPORT_SYMBOL_GPL(smc_hash_sk);
85 void smc_unhash_sk(struct sock *sk)
87 struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
89 write_lock_bh(&h->lock);
90 if (sk_del_node_init(sk))
91 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
92 write_unlock_bh(&h->lock);
94 EXPORT_SYMBOL_GPL(smc_unhash_sk);
96 struct proto smc_proto = {
99 .keepalive = smc_set_keepalive,
101 .unhash = smc_unhash_sk,
102 .obj_size = sizeof(struct smc_sock),
103 .h.smc_hash = &smc_v4_hashinfo,
104 .slab_flags = SLAB_TYPESAFE_BY_RCU,
106 EXPORT_SYMBOL_GPL(smc_proto);
108 static int smc_release(struct socket *sock)
110 struct sock *sk = sock->sk;
111 struct smc_sock *smc;
119 if (sk->sk_state == SMC_LISTEN)
120 /* smc_close_non_accepted() is called and acquires
121 * sock lock for child sockets again
123 lock_sock_nested(sk, SINGLE_DEPTH_NESTING);
127 if (smc->use_fallback) {
128 sk->sk_state = SMC_CLOSED;
129 sk->sk_state_change(sk);
131 rc = smc_close_active(smc);
132 sock_set_flag(sk, SOCK_DEAD);
133 sk->sk_shutdown |= SHUTDOWN_MASK;
136 if (smc->use_fallback && sk->sk_state == SMC_LISTEN) {
137 /* wake up clcsock accept */
138 rc = kernel_sock_shutdown(smc->clcsock, SHUT_RDWR);
140 mutex_lock(&smc->clcsock_release_lock);
141 sock_release(smc->clcsock);
143 mutex_unlock(&smc->clcsock_release_lock);
149 if (smc->use_fallback) {
150 schedule_delayed_work(&smc->sock_put_work, TCP_TIMEWAIT_LEN);
151 } else if (sk->sk_state == SMC_CLOSED) {
152 smc_conn_free(&smc->conn);
153 schedule_delayed_work(&smc->sock_put_work,
154 SMC_CLOSE_SOCK_PUT_DELAY);
163 static void smc_destruct(struct sock *sk)
165 if (sk->sk_state != SMC_CLOSED)
167 if (!sock_flag(sk, SOCK_DEAD))
170 sk_refcnt_debug_dec(sk);
173 static struct sock *smc_sock_alloc(struct net *net, struct socket *sock)
175 struct smc_sock *smc;
178 sk = sk_alloc(net, PF_SMC, GFP_KERNEL, &smc_proto, 0);
182 sock_init_data(sock, sk); /* sets sk_refcnt to 1 */
183 sk->sk_state = SMC_INIT;
184 sk->sk_destruct = smc_destruct;
185 sk->sk_protocol = SMCPROTO_SMC;
187 INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
188 INIT_LIST_HEAD(&smc->accept_q);
189 spin_lock_init(&smc->accept_q_lock);
190 INIT_DELAYED_WORK(&smc->sock_put_work, smc_close_sock_put_work);
191 sk->sk_prot->hash(sk);
192 sk_refcnt_debug_inc(sk);
193 mutex_init(&smc->clcsock_release_lock);
198 static int smc_bind(struct socket *sock, struct sockaddr *uaddr,
201 struct sockaddr_in *addr = (struct sockaddr_in *)uaddr;
202 struct sock *sk = sock->sk;
203 struct smc_sock *smc;
208 /* replicate tests from inet_bind(), to be safe wrt. future changes */
210 if (addr_len < sizeof(struct sockaddr_in))
214 /* accept AF_UNSPEC (mapped to AF_INET) only if s_addr is INADDR_ANY */
215 if ((addr->sin_family != AF_INET) &&
216 ((addr->sin_family != AF_UNSPEC) ||
217 (addr->sin_addr.s_addr != htonl(INADDR_ANY))))
222 /* Check if socket is already active */
224 if (sk->sk_state != SMC_INIT)
227 smc->clcsock->sk->sk_reuse = sk->sk_reuse;
228 rc = kernel_bind(smc->clcsock, uaddr, addr_len);
236 static void smc_copy_sock_settings(struct sock *nsk, struct sock *osk,
239 /* options we don't get control via setsockopt for */
240 nsk->sk_type = osk->sk_type;
241 nsk->sk_sndbuf = osk->sk_sndbuf;
242 nsk->sk_rcvbuf = osk->sk_rcvbuf;
243 nsk->sk_sndtimeo = osk->sk_sndtimeo;
244 nsk->sk_rcvtimeo = osk->sk_rcvtimeo;
245 nsk->sk_mark = osk->sk_mark;
246 nsk->sk_priority = osk->sk_priority;
247 nsk->sk_rcvlowat = osk->sk_rcvlowat;
248 nsk->sk_bound_dev_if = osk->sk_bound_dev_if;
249 nsk->sk_err = osk->sk_err;
251 nsk->sk_flags &= ~mask;
252 nsk->sk_flags |= osk->sk_flags & mask;
255 #define SK_FLAGS_SMC_TO_CLC ((1UL << SOCK_URGINLINE) | \
256 (1UL << SOCK_KEEPOPEN) | \
257 (1UL << SOCK_LINGER) | \
258 (1UL << SOCK_BROADCAST) | \
259 (1UL << SOCK_TIMESTAMP) | \
260 (1UL << SOCK_DBG) | \
261 (1UL << SOCK_RCVTSTAMP) | \
262 (1UL << SOCK_RCVTSTAMPNS) | \
263 (1UL << SOCK_LOCALROUTE) | \
264 (1UL << SOCK_TIMESTAMPING_RX_SOFTWARE) | \
265 (1UL << SOCK_RXQ_OVFL) | \
266 (1UL << SOCK_WIFI_STATUS) | \
267 (1UL << SOCK_NOFCS) | \
268 (1UL << SOCK_FILTER_LOCKED))
269 /* copy only relevant settings and flags of SOL_SOCKET level from smc to
270 * clc socket (since smc is not called for these options from net/core)
272 static void smc_copy_sock_settings_to_clc(struct smc_sock *smc)
274 smc_copy_sock_settings(smc->clcsock->sk, &smc->sk, SK_FLAGS_SMC_TO_CLC);
277 #define SK_FLAGS_CLC_TO_SMC ((1UL << SOCK_URGINLINE) | \
278 (1UL << SOCK_KEEPOPEN) | \
279 (1UL << SOCK_LINGER) | \
281 /* copy only settings and flags relevant for smc from clc to smc socket */
282 static void smc_copy_sock_settings_to_smc(struct smc_sock *smc)
284 smc_copy_sock_settings(&smc->sk, smc->clcsock->sk, SK_FLAGS_CLC_TO_SMC);
287 /* determine subnet and mask of internal TCP socket */
288 int smc_netinfo_by_tcpsk(struct socket *clcsock,
289 __be32 *subnet, u8 *prefix_len)
291 struct dst_entry *dst = sk_dst_get(clcsock->sk);
292 struct in_device *in_dev;
293 struct sockaddr_in addr;
306 /* get address to which the internal TCP socket is bound */
307 kernel_getsockname(clcsock, (struct sockaddr *)&addr, &len);
308 /* analyze IPv4 specific data of net_device belonging to TCP socket */
310 in_dev = __in_dev_get_rcu(dst->dev);
312 if (!inet_ifa_match(addr.sin_addr.s_addr, ifa))
314 *prefix_len = inet_mask_len(ifa->ifa_mask);
315 *subnet = ifa->ifa_address & ifa->ifa_mask;
318 } endfor_ifa(in_dev);
327 static int smc_clnt_conf_first_link(struct smc_sock *smc, union ib_gid *gid)
329 struct smc_link_group *lgr = smc->conn.lgr;
330 struct smc_link *link;
334 link = &lgr->lnk[SMC_SINGLE_LINK];
335 /* receive CONFIRM LINK request from server over RoCE fabric */
336 rest = wait_for_completion_interruptible_timeout(
338 SMC_LLC_WAIT_FIRST_TIME);
340 struct smc_clc_msg_decline dclc;
342 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
347 rc = smc_ib_modify_qp_rts(link);
349 return SMC_CLC_DECL_INTERR;
351 smc_wr_remember_qp_attr(link);
353 rc = smc_wr_reg_send(link,
354 smc->conn.rmb_desc->mr_rx[SMC_SINGLE_LINK]);
356 return SMC_CLC_DECL_INTERR;
358 /* send CONFIRM LINK response over RoCE fabric */
359 rc = smc_llc_send_confirm_link(link,
360 link->smcibdev->mac[link->ibport - 1],
363 return SMC_CLC_DECL_TCL;
368 static void smc_conn_save_peer_info(struct smc_sock *smc,
369 struct smc_clc_msg_accept_confirm *clc)
371 smc->conn.peer_conn_idx = clc->conn_idx;
372 smc->conn.local_tx_ctrl.token = ntohl(clc->rmbe_alert_token);
373 smc->conn.peer_rmbe_size = smc_uncompress_bufsize(clc->rmbe_size);
374 atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
377 static void smc_link_save_peer_info(struct smc_link *link,
378 struct smc_clc_msg_accept_confirm *clc)
380 link->peer_qpn = ntoh24(clc->qpn);
381 memcpy(link->peer_gid, clc->lcl.gid, SMC_GID_SIZE);
382 memcpy(link->peer_mac, clc->lcl.mac, sizeof(link->peer_mac));
383 link->peer_psn = ntoh24(clc->psn);
384 link->peer_mtu = clc->qp_mtu;
387 /* setup for RDMA connection of client */
388 static int smc_connect_rdma(struct smc_sock *smc)
390 struct sockaddr_in *inaddr = (struct sockaddr_in *)smc->addr;
391 struct smc_clc_msg_accept_confirm aclc;
392 int local_contact = SMC_FIRST_CONTACT;
393 struct smc_ib_device *smcibdev;
394 struct smc_link *link;
395 u8 srv_first_contact;
400 /* IPSec connections opt out of SMC-R optimizations */
401 if (using_ipsec(smc)) {
402 reason_code = SMC_CLC_DECL_IPSEC;
406 /* PNET table look up: search active ib_device and port
407 * within same PNETID that also contains the ethernet device
408 * used for the internal TCP socket
410 smc_pnet_find_roce_resource(smc->clcsock->sk, &smcibdev, &ibport);
412 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
416 /* do inband token exchange */
417 reason_code = smc_clc_send_proposal(smc, smcibdev, ibport);
418 if (reason_code < 0) {
422 if (reason_code > 0) /* configuration error */
424 /* receive SMC Accept CLC message */
425 reason_code = smc_clc_wait_msg(smc, &aclc, sizeof(aclc),
427 if (reason_code < 0) {
434 srv_first_contact = aclc.hdr.flag;
435 mutex_lock(&smc_create_lgr_pending);
436 local_contact = smc_conn_create(smc, inaddr->sin_addr.s_addr, smcibdev,
437 ibport, &aclc.lcl, srv_first_contact);
438 if (local_contact < 0) {
441 reason_code = SMC_CLC_DECL_MEM;/* insufficient memory*/
442 else if (rc == -ENOLINK)
443 reason_code = SMC_CLC_DECL_SYNCERR; /* synchr. error */
444 goto decline_rdma_unlock;
446 link = &smc->conn.lgr->lnk[SMC_SINGLE_LINK];
448 smc_conn_save_peer_info(smc, &aclc);
450 /* create send buffer and rmb */
451 rc = smc_buf_create(smc);
453 reason_code = SMC_CLC_DECL_MEM;
454 goto decline_rdma_unlock;
457 if (local_contact == SMC_FIRST_CONTACT)
458 smc_link_save_peer_info(link, &aclc);
460 rc = smc_rmb_rtoken_handling(&smc->conn, &aclc);
462 reason_code = SMC_CLC_DECL_INTERR;
463 goto decline_rdma_unlock;
469 if (local_contact == SMC_FIRST_CONTACT) {
470 rc = smc_ib_ready_link(link);
472 reason_code = SMC_CLC_DECL_INTERR;
473 goto decline_rdma_unlock;
476 struct smc_buf_desc *buf_desc = smc->conn.rmb_desc;
478 if (!buf_desc->reused) {
479 /* register memory region for new rmb */
480 rc = smc_wr_reg_send(link,
481 buf_desc->mr_rx[SMC_SINGLE_LINK]);
483 reason_code = SMC_CLC_DECL_INTERR;
484 goto decline_rdma_unlock;
488 smc_rmb_sync_sg_for_device(&smc->conn);
490 rc = smc_clc_send_confirm(smc);
494 if (local_contact == SMC_FIRST_CONTACT) {
495 /* QP confirmation over RoCE fabric */
496 reason_code = smc_clnt_conf_first_link(
497 smc, &smcibdev->gid[ibport - 1]);
498 if (reason_code < 0) {
503 goto decline_rdma_unlock;
506 mutex_unlock(&smc_create_lgr_pending);
510 smc_copy_sock_settings_to_clc(smc);
511 if (smc->sk.sk_state == SMC_INIT)
512 smc->sk.sk_state = SMC_ACTIVE;
514 return rc ? rc : local_contact;
517 mutex_unlock(&smc_create_lgr_pending);
518 smc_conn_free(&smc->conn);
520 /* RDMA setup failed, switch back to TCP */
521 smc->use_fallback = true;
522 if (reason_code && (reason_code != SMC_CLC_DECL_REPLY)) {
523 rc = smc_clc_send_decline(smc, reason_code);
524 if (rc < sizeof(struct smc_clc_msg_decline))
530 mutex_unlock(&smc_create_lgr_pending);
531 smc_conn_free(&smc->conn);
536 static int smc_connect(struct socket *sock, struct sockaddr *addr,
539 struct sock *sk = sock->sk;
540 struct smc_sock *smc;
545 /* separate smc parameter checking to be safe */
546 if (alen < sizeof(addr->sa_family))
548 if (addr->sa_family != AF_INET)
550 smc->addr = addr; /* needed for nonblocking connect */
553 switch (sk->sk_state) {
564 smc_copy_sock_settings_to_clc(smc);
565 rc = kernel_connect(smc->clcsock, addr, alen, flags);
569 /* setup RDMA connection */
570 rc = smc_connect_rdma(smc);
574 rc = 0; /* success cases including fallback */
582 static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
584 struct sock *sk = &lsmc->sk;
585 struct socket *new_clcsock;
589 release_sock(&lsmc->sk);
590 new_sk = smc_sock_alloc(sock_net(sk), NULL);
593 lsmc->sk.sk_err = ENOMEM;
595 lock_sock(&lsmc->sk);
598 *new_smc = smc_sk(new_sk);
600 mutex_lock(&lsmc->clcsock_release_lock);
602 rc = kernel_accept(lsmc->clcsock, &new_clcsock, 0);
603 mutex_unlock(&lsmc->clcsock_release_lock);
604 lock_sock(&lsmc->sk);
606 lsmc->sk.sk_err = -rc;
607 new_sk->sk_state = SMC_CLOSED;
608 sock_set_flag(new_sk, SOCK_DEAD);
609 sk->sk_prot->unhash(new_sk);
614 if (lsmc->sk.sk_state == SMC_CLOSED) {
616 sock_release(new_clcsock);
617 new_sk->sk_state = SMC_CLOSED;
618 sock_set_flag(new_sk, SOCK_DEAD);
619 sk->sk_prot->unhash(new_sk);
625 (*new_smc)->clcsock = new_clcsock;
630 /* add a just created sock to the accept queue of the listen sock as
631 * candidate for a following socket accept call from user space
633 static void smc_accept_enqueue(struct sock *parent, struct sock *sk)
635 struct smc_sock *par = smc_sk(parent);
638 spin_lock(&par->accept_q_lock);
639 list_add_tail(&smc_sk(sk)->accept_q, &par->accept_q);
640 spin_unlock(&par->accept_q_lock);
641 sk_acceptq_added(parent);
644 /* remove a socket from the accept queue of its parental listening socket */
645 static void smc_accept_unlink(struct sock *sk)
647 struct smc_sock *par = smc_sk(sk)->listen_smc;
649 spin_lock(&par->accept_q_lock);
650 list_del_init(&smc_sk(sk)->accept_q);
651 spin_unlock(&par->accept_q_lock);
652 sk_acceptq_removed(&smc_sk(sk)->listen_smc->sk);
656 /* remove a sock from the accept queue to bind it to a new socket created
657 * for a socket accept call from user space
659 struct sock *smc_accept_dequeue(struct sock *parent,
660 struct socket *new_sock)
662 struct smc_sock *isk, *n;
665 list_for_each_entry_safe(isk, n, &smc_sk(parent)->accept_q, accept_q) {
666 new_sk = (struct sock *)isk;
668 smc_accept_unlink(new_sk);
669 if (new_sk->sk_state == SMC_CLOSED) {
670 new_sk->sk_prot->unhash(new_sk);
675 sock_graft(new_sk, new_sock);
681 /* clean up for a created but never accepted sock */
682 void smc_close_non_accepted(struct sock *sk)
684 struct smc_sock *smc = smc_sk(sk);
688 if (!sk->sk_lingertime)
689 /* wait for peer closing */
690 sk->sk_lingertime = SMC_MAX_STREAM_WAIT_TIMEOUT;
691 if (smc->use_fallback) {
692 sk->sk_state = SMC_CLOSED;
694 smc_close_active(smc);
695 sock_set_flag(sk, SOCK_DEAD);
696 sk->sk_shutdown |= SHUTDOWN_MASK;
705 if (smc->use_fallback) {
706 schedule_delayed_work(&smc->sock_put_work, TCP_TIMEWAIT_LEN);
707 } else if (sk->sk_state == SMC_CLOSED) {
708 smc_conn_free(&smc->conn);
709 schedule_delayed_work(&smc->sock_put_work,
710 SMC_CLOSE_SOCK_PUT_DELAY);
716 static int smc_serv_conf_first_link(struct smc_sock *smc)
718 struct smc_link_group *lgr = smc->conn.lgr;
719 struct smc_link *link;
723 link = &lgr->lnk[SMC_SINGLE_LINK];
725 rc = smc_wr_reg_send(link,
726 smc->conn.rmb_desc->mr_rx[SMC_SINGLE_LINK]);
728 return SMC_CLC_DECL_INTERR;
730 /* send CONFIRM LINK request to client over the RoCE fabric */
731 rc = smc_llc_send_confirm_link(link,
732 link->smcibdev->mac[link->ibport - 1],
733 &link->smcibdev->gid[link->ibport - 1],
736 return SMC_CLC_DECL_TCL;
738 /* receive CONFIRM LINK response from client over the RoCE fabric */
739 rest = wait_for_completion_interruptible_timeout(
740 &link->llc_confirm_resp,
741 SMC_LLC_WAIT_FIRST_TIME);
743 struct smc_clc_msg_decline dclc;
745 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
752 /* setup for RDMA connection of server */
753 static void smc_listen_work(struct work_struct *work)
755 struct smc_sock *new_smc = container_of(work, struct smc_sock,
757 struct socket *newclcsock = new_smc->clcsock;
758 struct smc_sock *lsmc = new_smc->listen_smc;
759 struct smc_clc_msg_accept_confirm cclc;
760 int local_contact = SMC_REUSE_CONTACT;
761 struct sock *newsmcsk = &new_smc->sk;
762 struct smc_clc_msg_proposal pclc;
763 struct smc_ib_device *smcibdev;
764 struct sockaddr_in peeraddr;
765 struct smc_link *link;
772 /* do inband token exchange -
773 *wait for and receive SMC Proposal CLC message
775 reason_code = smc_clc_wait_msg(new_smc, &pclc, sizeof(pclc),
782 /* IPSec connections opt out of SMC-R optimizations */
783 if (using_ipsec(new_smc)) {
784 reason_code = SMC_CLC_DECL_IPSEC;
788 /* PNET table look up: search active ib_device and port
789 * within same PNETID that also contains the ethernet device
790 * used for the internal TCP socket
792 smc_pnet_find_roce_resource(newclcsock->sk, &smcibdev, &ibport);
794 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
798 /* determine subnet and mask from internal TCP socket */
799 rc = smc_netinfo_by_tcpsk(newclcsock, &subnet, &prefix_len);
801 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
804 if ((pclc.outgoing_subnet != subnet) ||
805 (pclc.prefix_len != prefix_len)) {
806 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
810 /* get address of the peer connected to the internal TCP socket */
811 kernel_getpeername(newclcsock, (struct sockaddr *)&peeraddr, &len);
813 /* allocate connection / link group */
814 mutex_lock(&smc_create_lgr_pending);
815 local_contact = smc_conn_create(new_smc, peeraddr.sin_addr.s_addr,
816 smcibdev, ibport, &pclc.lcl, 0);
817 if (local_contact < 0) {
820 reason_code = SMC_CLC_DECL_MEM;/* insufficient memory*/
823 link = &new_smc->conn.lgr->lnk[SMC_SINGLE_LINK];
825 /* create send buffer and rmb */
826 rc = smc_buf_create(new_smc);
828 reason_code = SMC_CLC_DECL_MEM;
832 smc_close_init(new_smc);
833 smc_rx_init(new_smc);
835 if (local_contact != SMC_FIRST_CONTACT) {
836 struct smc_buf_desc *buf_desc = new_smc->conn.rmb_desc;
838 if (!buf_desc->reused) {
839 /* register memory region for new rmb */
840 rc = smc_wr_reg_send(link,
841 buf_desc->mr_rx[SMC_SINGLE_LINK]);
843 reason_code = SMC_CLC_DECL_INTERR;
848 smc_rmb_sync_sg_for_device(&new_smc->conn);
850 rc = smc_clc_send_accept(new_smc, local_contact);
854 /* receive SMC Confirm CLC message */
855 reason_code = smc_clc_wait_msg(new_smc, &cclc, sizeof(cclc),
861 smc_conn_save_peer_info(new_smc, &cclc);
862 if (local_contact == SMC_FIRST_CONTACT)
863 smc_link_save_peer_info(link, &cclc);
865 rc = smc_rmb_rtoken_handling(&new_smc->conn, &cclc);
867 reason_code = SMC_CLC_DECL_INTERR;
871 if (local_contact == SMC_FIRST_CONTACT) {
872 rc = smc_ib_ready_link(link);
874 reason_code = SMC_CLC_DECL_INTERR;
877 /* QP confirmation over RoCE fabric */
878 reason_code = smc_serv_conf_first_link(new_smc);
879 if (reason_code < 0) {
880 /* peer is not aware of a problem */
888 smc_tx_init(new_smc);
891 sk_refcnt_debug_inc(newsmcsk);
892 if (newsmcsk->sk_state == SMC_INIT)
893 newsmcsk->sk_state = SMC_ACTIVE;
895 mutex_unlock(&smc_create_lgr_pending);
896 lock_sock_nested(&lsmc->sk, SINGLE_DEPTH_NESTING);
897 if (lsmc->sk.sk_state == SMC_LISTEN) {
898 smc_accept_enqueue(&lsmc->sk, newsmcsk);
899 } else { /* no longer listening */
900 smc_close_non_accepted(newsmcsk);
902 release_sock(&lsmc->sk);
905 lsmc->sk.sk_data_ready(&lsmc->sk);
906 sock_put(&lsmc->sk); /* sock_hold in smc_tcp_listen_work */
910 /* RDMA setup failed, switch back to TCP */
911 smc_conn_free(&new_smc->conn);
912 new_smc->use_fallback = true;
913 if (reason_code && (reason_code != SMC_CLC_DECL_REPLY)) {
914 rc = smc_clc_send_decline(new_smc, reason_code);
915 if (rc < sizeof(struct smc_clc_msg_decline))
921 newsmcsk->sk_state = SMC_CLOSED;
922 smc_conn_free(&new_smc->conn);
923 goto enqueue; /* queue new sock with sk_err set */
926 static void smc_tcp_listen_work(struct work_struct *work)
928 struct smc_sock *lsmc = container_of(work, struct smc_sock,
930 struct smc_sock *new_smc;
933 lock_sock(&lsmc->sk);
934 while (lsmc->sk.sk_state == SMC_LISTEN) {
935 rc = smc_clcsock_accept(lsmc, &new_smc);
941 new_smc->listen_smc = lsmc;
942 new_smc->use_fallback = false; /* assume rdma capability first*/
943 sock_hold(&lsmc->sk); /* sock_put in smc_listen_work */
944 INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
945 smc_copy_sock_settings_to_smc(new_smc);
946 schedule_work(&new_smc->smc_listen_work);
950 release_sock(&lsmc->sk);
951 lsmc->sk.sk_data_ready(&lsmc->sk); /* no more listening, wake accept */
954 static int smc_listen(struct socket *sock, int backlog)
956 struct sock *sk = sock->sk;
957 struct smc_sock *smc;
964 if ((sk->sk_state != SMC_INIT) && (sk->sk_state != SMC_LISTEN))
968 if (sk->sk_state == SMC_LISTEN) {
969 sk->sk_max_ack_backlog = backlog;
972 /* some socket options are handled in core, so we could not apply
973 * them to the clc socket -- copy smc socket options to clc socket
975 smc_copy_sock_settings_to_clc(smc);
977 rc = kernel_listen(smc->clcsock, backlog);
980 sk->sk_max_ack_backlog = backlog;
981 sk->sk_ack_backlog = 0;
982 sk->sk_state = SMC_LISTEN;
983 INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
984 schedule_work(&smc->tcp_listen_work);
991 static int smc_accept(struct socket *sock, struct socket *new_sock,
992 int flags, bool kern)
994 struct sock *sk = sock->sk, *nsk;
995 DECLARE_WAITQUEUE(wait, current);
996 struct smc_sock *lsmc;
1003 if (lsmc->sk.sk_state != SMC_LISTEN) {
1008 /* Wait for an incoming connection */
1009 timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
1010 add_wait_queue_exclusive(sk_sleep(sk), &wait);
1011 while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
1012 set_current_state(TASK_INTERRUPTIBLE);
1018 timeo = schedule_timeout(timeo);
1019 /* wakeup by sk_data_ready in smc_listen_work() */
1020 sched_annotate_sleep();
1022 if (signal_pending(current)) {
1023 rc = sock_intr_errno(timeo);
1027 set_current_state(TASK_RUNNING);
1028 remove_wait_queue(sk_sleep(sk), &wait);
1031 rc = sock_error(nsk);
1038 static int smc_getname(struct socket *sock, struct sockaddr *addr,
1041 struct smc_sock *smc;
1043 if (peer && (sock->sk->sk_state != SMC_ACTIVE) &&
1044 (sock->sk->sk_state != SMC_APPCLOSEWAIT1))
1047 smc = smc_sk(sock->sk);
1049 return smc->clcsock->ops->getname(smc->clcsock, addr, len, peer);
1052 static int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
1054 struct sock *sk = sock->sk;
1055 struct smc_sock *smc;
1060 if ((sk->sk_state != SMC_ACTIVE) &&
1061 (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1062 (sk->sk_state != SMC_INIT))
1064 if (smc->use_fallback)
1065 rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
1067 rc = smc_tx_sendmsg(smc, msg, len);
1073 static int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
1076 struct sock *sk = sock->sk;
1077 struct smc_sock *smc;
1082 if ((sk->sk_state == SMC_INIT) ||
1083 (sk->sk_state == SMC_LISTEN) ||
1084 (sk->sk_state == SMC_CLOSED))
1087 if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
1092 if (smc->use_fallback)
1093 rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
1095 rc = smc_rx_recvmsg(smc, msg, len, flags);
1102 static unsigned int smc_accept_poll(struct sock *parent)
1104 struct smc_sock *isk;
1108 list_for_each_entry(isk, &smc_sk(parent)->accept_q, accept_q) {
1109 sk = (struct sock *)isk;
1111 if (sk->sk_state == SMC_ACTIVE) {
1112 release_sock(parent);
1113 return POLLIN | POLLRDNORM;
1116 release_sock(parent);
1121 static unsigned int smc_poll(struct file *file, struct socket *sock,
1124 struct sock *sk = sock->sk;
1125 unsigned int mask = 0;
1126 struct smc_sock *smc;
1129 smc = smc_sk(sock->sk);
1130 if ((sk->sk_state == SMC_INIT) || smc->use_fallback) {
1131 /* delegate to CLC child sock */
1132 mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
1133 /* if non-blocking connect finished ... */
1135 if ((sk->sk_state == SMC_INIT) && (mask & POLLOUT)) {
1136 sk->sk_err = smc->clcsock->sk->sk_err;
1140 rc = smc_connect_rdma(smc);
1144 /* success cases including fallback */
1145 mask |= POLLOUT | POLLWRNORM;
1150 sock_poll_wait(file, sk_sleep(sk), wait);
1151 if (sk->sk_state == SMC_LISTEN)
1152 /* woken up by sk_data_ready in smc_listen_work() */
1153 mask |= smc_accept_poll(sk);
1156 if (atomic_read(&smc->conn.sndbuf_space) ||
1157 (sk->sk_shutdown & SEND_SHUTDOWN)) {
1158 mask |= POLLOUT | POLLWRNORM;
1160 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1161 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1163 if (atomic_read(&smc->conn.bytes_to_rcv))
1164 mask |= POLLIN | POLLRDNORM;
1165 if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
1166 (sk->sk_state == SMC_CLOSED))
1168 if (sk->sk_shutdown & RCV_SHUTDOWN)
1169 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
1170 if (sk->sk_state == SMC_APPCLOSEWAIT1)
1178 static int smc_shutdown(struct socket *sock, int how)
1180 struct sock *sk = sock->sk;
1181 bool do_shutdown = true;
1182 struct smc_sock *smc;
1189 if ((how < SHUT_RD) || (how > SHUT_RDWR))
1195 if ((sk->sk_state != SMC_ACTIVE) &&
1196 (sk->sk_state != SMC_PEERCLOSEWAIT1) &&
1197 (sk->sk_state != SMC_PEERCLOSEWAIT2) &&
1198 (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1199 (sk->sk_state != SMC_APPCLOSEWAIT2) &&
1200 (sk->sk_state != SMC_APPFINCLOSEWAIT))
1202 if (smc->use_fallback) {
1203 rc = kernel_sock_shutdown(smc->clcsock, how);
1204 sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
1205 if (sk->sk_shutdown == SHUTDOWN_MASK)
1206 sk->sk_state = SMC_CLOSED;
1210 case SHUT_RDWR: /* shutdown in both directions */
1211 old_state = sk->sk_state;
1212 rc = smc_close_active(smc);
1213 if (old_state == SMC_ACTIVE &&
1214 sk->sk_state == SMC_PEERCLOSEWAIT1)
1215 do_shutdown = false;
1218 rc = smc_close_shutdown_write(smc);
1222 /* nothing more to do because peer is not involved */
1225 if (do_shutdown && smc->clcsock)
1226 rc1 = kernel_sock_shutdown(smc->clcsock, how);
1227 /* map sock_shutdown_cmd constants to sk_shutdown value range */
1228 sk->sk_shutdown |= how + 1;
1232 return rc ? rc : rc1;
1235 static int smc_setsockopt(struct socket *sock, int level, int optname,
1236 char __user *optval, unsigned int optlen)
1238 struct sock *sk = sock->sk;
1239 struct smc_sock *smc;
1243 /* generic setsockopts reaching us here always apply to the
1246 return smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
1250 static int smc_getsockopt(struct socket *sock, int level, int optname,
1251 char __user *optval, int __user *optlen)
1253 struct smc_sock *smc;
1255 smc = smc_sk(sock->sk);
1256 /* socket options apply to the CLC socket */
1257 return smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
1261 static int smc_ioctl(struct socket *sock, unsigned int cmd,
1264 struct smc_sock *smc;
1266 smc = smc_sk(sock->sk);
1267 if (smc->use_fallback)
1268 return smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
1270 return sock_no_ioctl(sock, cmd, arg);
1273 static ssize_t smc_sendpage(struct socket *sock, struct page *page,
1274 int offset, size_t size, int flags)
1276 struct sock *sk = sock->sk;
1277 struct smc_sock *smc;
1282 if (sk->sk_state != SMC_ACTIVE) {
1287 if (smc->use_fallback)
1288 rc = kernel_sendpage(smc->clcsock, page, offset,
1291 rc = sock_no_sendpage(sock, page, offset, size, flags);
1297 static ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
1298 struct pipe_inode_info *pipe, size_t len,
1301 struct sock *sk = sock->sk;
1302 struct smc_sock *smc;
1307 if ((sk->sk_state != SMC_ACTIVE) && (sk->sk_state != SMC_CLOSED))
1309 if (smc->use_fallback) {
1310 rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
1320 /* must look like tcp */
1321 static const struct proto_ops smc_sock_ops = {
1323 .owner = THIS_MODULE,
1324 .release = smc_release,
1326 .connect = smc_connect,
1327 .socketpair = sock_no_socketpair,
1328 .accept = smc_accept,
1329 .getname = smc_getname,
1332 .listen = smc_listen,
1333 .shutdown = smc_shutdown,
1334 .setsockopt = smc_setsockopt,
1335 .getsockopt = smc_getsockopt,
1336 .sendmsg = smc_sendmsg,
1337 .recvmsg = smc_recvmsg,
1338 .mmap = sock_no_mmap,
1339 .sendpage = smc_sendpage,
1340 .splice_read = smc_splice_read,
1343 static int smc_create(struct net *net, struct socket *sock, int protocol,
1346 struct smc_sock *smc;
1350 rc = -ESOCKTNOSUPPORT;
1351 if (sock->type != SOCK_STREAM)
1354 rc = -EPROTONOSUPPORT;
1355 if ((protocol != IPPROTO_IP) && (protocol != IPPROTO_TCP))
1359 sock->ops = &smc_sock_ops;
1360 sk = smc_sock_alloc(net, sock);
1364 /* create internal TCP socket for CLC handshake and fallback */
1366 smc->use_fallback = false; /* assume rdma capability first */
1367 rc = sock_create_kern(net, PF_INET, SOCK_STREAM,
1368 IPPROTO_TCP, &smc->clcsock);
1370 sk_common_release(sk);
1373 smc->sk.sk_sndbuf = max(smc->clcsock->sk->sk_sndbuf, SMC_BUF_MIN_SIZE);
1374 smc->sk.sk_rcvbuf = max(smc->clcsock->sk->sk_rcvbuf, SMC_BUF_MIN_SIZE);
1380 static const struct net_proto_family smc_sock_family_ops = {
1382 .owner = THIS_MODULE,
1383 .create = smc_create,
1386 static int __init smc_init(void)
1390 rc = smc_pnet_init();
1394 rc = smc_llc_init();
1396 pr_err("%s: smc_llc_init fails with %d\n", __func__, rc);
1400 rc = smc_cdc_init();
1402 pr_err("%s: smc_cdc_init fails with %d\n", __func__, rc);
1406 rc = proto_register(&smc_proto, 1);
1408 pr_err("%s: proto_register fails with %d\n", __func__, rc);
1412 rc = sock_register(&smc_sock_family_ops);
1414 pr_err("%s: sock_register fails with %d\n", __func__, rc);
1417 INIT_HLIST_HEAD(&smc_v4_hashinfo.ht);
1419 rc = smc_ib_register_client();
1421 pr_err("%s: ib_register fails with %d\n", __func__, rc);
1428 sock_unregister(PF_SMC);
1430 proto_unregister(&smc_proto);
1436 static void __exit smc_exit(void)
1438 struct smc_link_group *lgr, *lg;
1439 LIST_HEAD(lgr_freeing_list);
1441 spin_lock_bh(&smc_lgr_list.lock);
1442 if (!list_empty(&smc_lgr_list.list))
1443 list_splice_init(&smc_lgr_list.list, &lgr_freeing_list);
1444 spin_unlock_bh(&smc_lgr_list.lock);
1445 list_for_each_entry_safe(lgr, lg, &lgr_freeing_list, list) {
1446 list_del_init(&lgr->list);
1447 smc_lgr_free(lgr); /* free link group */
1449 smc_ib_unregister_client();
1450 sock_unregister(PF_SMC);
1451 proto_unregister(&smc_proto);
1455 module_init(smc_init);
1456 module_exit(smc_exit);
1458 MODULE_AUTHOR("Ursula Braun <ubraun@linux.vnet.ibm.com>");
1459 MODULE_DESCRIPTION("smc socket address family");
1460 MODULE_LICENSE("GPL");
1461 MODULE_ALIAS_NETPROTO(PF_SMC);