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 sock_release(smc->clcsock);
143 if (smc->use_fallback) {
144 schedule_delayed_work(&smc->sock_put_work, TCP_TIMEWAIT_LEN);
145 } else if (sk->sk_state == SMC_CLOSED) {
146 smc_conn_free(&smc->conn);
147 schedule_delayed_work(&smc->sock_put_work,
148 SMC_CLOSE_SOCK_PUT_DELAY);
157 static void smc_destruct(struct sock *sk)
159 if (sk->sk_state != SMC_CLOSED)
161 if (!sock_flag(sk, SOCK_DEAD))
164 sk_refcnt_debug_dec(sk);
167 static struct sock *smc_sock_alloc(struct net *net, struct socket *sock)
169 struct smc_sock *smc;
172 sk = sk_alloc(net, PF_SMC, GFP_KERNEL, &smc_proto, 0);
176 sock_init_data(sock, sk); /* sets sk_refcnt to 1 */
177 sk->sk_state = SMC_INIT;
178 sk->sk_destruct = smc_destruct;
179 sk->sk_protocol = SMCPROTO_SMC;
181 INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
182 INIT_LIST_HEAD(&smc->accept_q);
183 spin_lock_init(&smc->accept_q_lock);
184 INIT_DELAYED_WORK(&smc->sock_put_work, smc_close_sock_put_work);
185 sk->sk_prot->hash(sk);
186 sk_refcnt_debug_inc(sk);
191 static int smc_bind(struct socket *sock, struct sockaddr *uaddr,
194 struct sockaddr_in *addr = (struct sockaddr_in *)uaddr;
195 struct sock *sk = sock->sk;
196 struct smc_sock *smc;
201 /* replicate tests from inet_bind(), to be safe wrt. future changes */
203 if (addr_len < sizeof(struct sockaddr_in))
207 /* accept AF_UNSPEC (mapped to AF_INET) only if s_addr is INADDR_ANY */
208 if ((addr->sin_family != AF_INET) &&
209 ((addr->sin_family != AF_UNSPEC) ||
210 (addr->sin_addr.s_addr != htonl(INADDR_ANY))))
215 /* Check if socket is already active */
217 if (sk->sk_state != SMC_INIT)
220 smc->clcsock->sk->sk_reuse = sk->sk_reuse;
221 rc = kernel_bind(smc->clcsock, uaddr, addr_len);
229 static void smc_copy_sock_settings(struct sock *nsk, struct sock *osk,
232 /* options we don't get control via setsockopt for */
233 nsk->sk_type = osk->sk_type;
234 nsk->sk_sndbuf = osk->sk_sndbuf;
235 nsk->sk_rcvbuf = osk->sk_rcvbuf;
236 nsk->sk_sndtimeo = osk->sk_sndtimeo;
237 nsk->sk_rcvtimeo = osk->sk_rcvtimeo;
238 nsk->sk_mark = osk->sk_mark;
239 nsk->sk_priority = osk->sk_priority;
240 nsk->sk_rcvlowat = osk->sk_rcvlowat;
241 nsk->sk_bound_dev_if = osk->sk_bound_dev_if;
242 nsk->sk_err = osk->sk_err;
244 nsk->sk_flags &= ~mask;
245 nsk->sk_flags |= osk->sk_flags & mask;
248 #define SK_FLAGS_SMC_TO_CLC ((1UL << SOCK_URGINLINE) | \
249 (1UL << SOCK_KEEPOPEN) | \
250 (1UL << SOCK_LINGER) | \
251 (1UL << SOCK_BROADCAST) | \
252 (1UL << SOCK_TIMESTAMP) | \
253 (1UL << SOCK_DBG) | \
254 (1UL << SOCK_RCVTSTAMP) | \
255 (1UL << SOCK_RCVTSTAMPNS) | \
256 (1UL << SOCK_LOCALROUTE) | \
257 (1UL << SOCK_TIMESTAMPING_RX_SOFTWARE) | \
258 (1UL << SOCK_RXQ_OVFL) | \
259 (1UL << SOCK_WIFI_STATUS) | \
260 (1UL << SOCK_NOFCS) | \
261 (1UL << SOCK_FILTER_LOCKED))
262 /* copy only relevant settings and flags of SOL_SOCKET level from smc to
263 * clc socket (since smc is not called for these options from net/core)
265 static void smc_copy_sock_settings_to_clc(struct smc_sock *smc)
267 smc_copy_sock_settings(smc->clcsock->sk, &smc->sk, SK_FLAGS_SMC_TO_CLC);
270 #define SK_FLAGS_CLC_TO_SMC ((1UL << SOCK_URGINLINE) | \
271 (1UL << SOCK_KEEPOPEN) | \
272 (1UL << SOCK_LINGER) | \
274 /* copy only settings and flags relevant for smc from clc to smc socket */
275 static void smc_copy_sock_settings_to_smc(struct smc_sock *smc)
277 smc_copy_sock_settings(&smc->sk, smc->clcsock->sk, SK_FLAGS_CLC_TO_SMC);
280 /* determine subnet and mask of internal TCP socket */
281 int smc_netinfo_by_tcpsk(struct socket *clcsock,
282 __be32 *subnet, u8 *prefix_len)
284 struct dst_entry *dst = sk_dst_get(clcsock->sk);
285 struct sockaddr_in addr;
298 /* get address to which the internal TCP socket is bound */
299 kernel_getsockname(clcsock, (struct sockaddr *)&addr, &len);
300 /* analyze IPv4 specific data of net_device belonging to TCP socket */
301 for_ifa(dst->dev->ip_ptr) {
302 if (ifa->ifa_address != addr.sin_addr.s_addr)
304 *prefix_len = inet_mask_len(ifa->ifa_mask);
305 *subnet = ifa->ifa_address & ifa->ifa_mask;
308 } endfor_ifa(dst->dev->ip_ptr);
316 static int smc_clnt_conf_first_link(struct smc_sock *smc, union ib_gid *gid)
318 struct smc_link_group *lgr = smc->conn.lgr;
319 struct smc_link *link;
323 link = &lgr->lnk[SMC_SINGLE_LINK];
324 /* receive CONFIRM LINK request from server over RoCE fabric */
325 rest = wait_for_completion_interruptible_timeout(
327 SMC_LLC_WAIT_FIRST_TIME);
329 struct smc_clc_msg_decline dclc;
331 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
336 rc = smc_ib_modify_qp_rts(link);
338 return SMC_CLC_DECL_INTERR;
340 smc_wr_remember_qp_attr(link);
342 rc = smc_wr_reg_send(link,
343 smc->conn.rmb_desc->mr_rx[SMC_SINGLE_LINK]);
345 return SMC_CLC_DECL_INTERR;
347 /* send CONFIRM LINK response over RoCE fabric */
348 rc = smc_llc_send_confirm_link(link,
349 link->smcibdev->mac[link->ibport - 1],
352 return SMC_CLC_DECL_TCL;
357 static void smc_conn_save_peer_info(struct smc_sock *smc,
358 struct smc_clc_msg_accept_confirm *clc)
360 smc->conn.peer_conn_idx = clc->conn_idx;
361 smc->conn.local_tx_ctrl.token = ntohl(clc->rmbe_alert_token);
362 smc->conn.peer_rmbe_size = smc_uncompress_bufsize(clc->rmbe_size);
363 atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
366 static void smc_link_save_peer_info(struct smc_link *link,
367 struct smc_clc_msg_accept_confirm *clc)
369 link->peer_qpn = ntoh24(clc->qpn);
370 memcpy(link->peer_gid, clc->lcl.gid, SMC_GID_SIZE);
371 memcpy(link->peer_mac, clc->lcl.mac, sizeof(link->peer_mac));
372 link->peer_psn = ntoh24(clc->psn);
373 link->peer_mtu = clc->qp_mtu;
376 /* setup for RDMA connection of client */
377 static int smc_connect_rdma(struct smc_sock *smc)
379 struct sockaddr_in *inaddr = (struct sockaddr_in *)smc->addr;
380 struct smc_clc_msg_accept_confirm aclc;
381 int local_contact = SMC_FIRST_CONTACT;
382 struct smc_ib_device *smcibdev;
383 struct smc_link *link;
384 u8 srv_first_contact;
389 /* IPSec connections opt out of SMC-R optimizations */
390 if (using_ipsec(smc)) {
391 reason_code = SMC_CLC_DECL_IPSEC;
395 /* PNET table look up: search active ib_device and port
396 * within same PNETID that also contains the ethernet device
397 * used for the internal TCP socket
399 smc_pnet_find_roce_resource(smc->clcsock->sk, &smcibdev, &ibport);
401 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
405 /* do inband token exchange */
406 reason_code = smc_clc_send_proposal(smc, smcibdev, ibport);
407 if (reason_code < 0) {
411 if (reason_code > 0) /* configuration error */
413 /* receive SMC Accept CLC message */
414 reason_code = smc_clc_wait_msg(smc, &aclc, sizeof(aclc),
416 if (reason_code < 0) {
423 srv_first_contact = aclc.hdr.flag;
424 mutex_lock(&smc_create_lgr_pending);
425 local_contact = smc_conn_create(smc, inaddr->sin_addr.s_addr, smcibdev,
426 ibport, &aclc.lcl, srv_first_contact);
427 if (local_contact < 0) {
430 reason_code = SMC_CLC_DECL_MEM;/* insufficient memory*/
431 else if (rc == -ENOLINK)
432 reason_code = SMC_CLC_DECL_SYNCERR; /* synchr. error */
433 goto decline_rdma_unlock;
435 link = &smc->conn.lgr->lnk[SMC_SINGLE_LINK];
437 smc_conn_save_peer_info(smc, &aclc);
439 /* create send buffer and rmb */
440 rc = smc_buf_create(smc);
442 reason_code = SMC_CLC_DECL_MEM;
443 goto decline_rdma_unlock;
446 if (local_contact == SMC_FIRST_CONTACT)
447 smc_link_save_peer_info(link, &aclc);
449 rc = smc_rmb_rtoken_handling(&smc->conn, &aclc);
451 reason_code = SMC_CLC_DECL_INTERR;
452 goto decline_rdma_unlock;
458 if (local_contact == SMC_FIRST_CONTACT) {
459 rc = smc_ib_ready_link(link);
461 reason_code = SMC_CLC_DECL_INTERR;
462 goto decline_rdma_unlock;
465 struct smc_buf_desc *buf_desc = smc->conn.rmb_desc;
467 if (!buf_desc->reused) {
468 /* register memory region for new rmb */
469 rc = smc_wr_reg_send(link,
470 buf_desc->mr_rx[SMC_SINGLE_LINK]);
472 reason_code = SMC_CLC_DECL_INTERR;
473 goto decline_rdma_unlock;
477 smc_rmb_sync_sg_for_device(&smc->conn);
479 rc = smc_clc_send_confirm(smc);
483 if (local_contact == SMC_FIRST_CONTACT) {
484 /* QP confirmation over RoCE fabric */
485 reason_code = smc_clnt_conf_first_link(
486 smc, &smcibdev->gid[ibport - 1]);
487 if (reason_code < 0) {
492 goto decline_rdma_unlock;
495 mutex_unlock(&smc_create_lgr_pending);
499 smc_copy_sock_settings_to_clc(smc);
500 if (smc->sk.sk_state == SMC_INIT)
501 smc->sk.sk_state = SMC_ACTIVE;
503 return rc ? rc : local_contact;
506 mutex_unlock(&smc_create_lgr_pending);
507 smc_conn_free(&smc->conn);
509 /* RDMA setup failed, switch back to TCP */
510 smc->use_fallback = true;
511 if (reason_code && (reason_code != SMC_CLC_DECL_REPLY)) {
512 rc = smc_clc_send_decline(smc, reason_code, 0);
513 if (rc < sizeof(struct smc_clc_msg_decline))
519 mutex_unlock(&smc_create_lgr_pending);
520 smc_conn_free(&smc->conn);
525 static int smc_connect(struct socket *sock, struct sockaddr *addr,
528 struct sock *sk = sock->sk;
529 struct smc_sock *smc;
534 /* separate smc parameter checking to be safe */
535 if (alen < sizeof(addr->sa_family))
537 if (addr->sa_family != AF_INET)
539 smc->addr = addr; /* needed for nonblocking connect */
542 switch (sk->sk_state) {
553 smc_copy_sock_settings_to_clc(smc);
554 rc = kernel_connect(smc->clcsock, addr, alen, flags);
558 /* setup RDMA connection */
559 rc = smc_connect_rdma(smc);
563 rc = 0; /* success cases including fallback */
571 static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
573 struct sock *sk = &lsmc->sk;
574 struct socket *new_clcsock;
578 release_sock(&lsmc->sk);
579 new_sk = smc_sock_alloc(sock_net(sk), NULL);
582 lsmc->sk.sk_err = ENOMEM;
584 lock_sock(&lsmc->sk);
587 *new_smc = smc_sk(new_sk);
589 rc = kernel_accept(lsmc->clcsock, &new_clcsock, 0);
590 lock_sock(&lsmc->sk);
592 lsmc->sk.sk_err = -rc;
593 new_sk->sk_state = SMC_CLOSED;
594 sock_set_flag(new_sk, SOCK_DEAD);
595 sk->sk_prot->unhash(new_sk);
600 if (lsmc->sk.sk_state == SMC_CLOSED) {
602 sock_release(new_clcsock);
603 new_sk->sk_state = SMC_CLOSED;
604 sock_set_flag(new_sk, SOCK_DEAD);
605 sk->sk_prot->unhash(new_sk);
611 (*new_smc)->clcsock = new_clcsock;
616 /* add a just created sock to the accept queue of the listen sock as
617 * candidate for a following socket accept call from user space
619 static void smc_accept_enqueue(struct sock *parent, struct sock *sk)
621 struct smc_sock *par = smc_sk(parent);
624 spin_lock(&par->accept_q_lock);
625 list_add_tail(&smc_sk(sk)->accept_q, &par->accept_q);
626 spin_unlock(&par->accept_q_lock);
627 sk_acceptq_added(parent);
630 /* remove a socket from the accept queue of its parental listening socket */
631 static void smc_accept_unlink(struct sock *sk)
633 struct smc_sock *par = smc_sk(sk)->listen_smc;
635 spin_lock(&par->accept_q_lock);
636 list_del_init(&smc_sk(sk)->accept_q);
637 spin_unlock(&par->accept_q_lock);
638 sk_acceptq_removed(&smc_sk(sk)->listen_smc->sk);
642 /* remove a sock from the accept queue to bind it to a new socket created
643 * for a socket accept call from user space
645 struct sock *smc_accept_dequeue(struct sock *parent,
646 struct socket *new_sock)
648 struct smc_sock *isk, *n;
651 list_for_each_entry_safe(isk, n, &smc_sk(parent)->accept_q, accept_q) {
652 new_sk = (struct sock *)isk;
654 smc_accept_unlink(new_sk);
655 if (new_sk->sk_state == SMC_CLOSED) {
656 new_sk->sk_prot->unhash(new_sk);
661 sock_graft(new_sk, new_sock);
667 /* clean up for a created but never accepted sock */
668 void smc_close_non_accepted(struct sock *sk)
670 struct smc_sock *smc = smc_sk(sk);
674 if (!sk->sk_lingertime)
675 /* wait for peer closing */
676 sk->sk_lingertime = SMC_MAX_STREAM_WAIT_TIMEOUT;
677 if (smc->use_fallback) {
678 sk->sk_state = SMC_CLOSED;
680 smc_close_active(smc);
681 sock_set_flag(sk, SOCK_DEAD);
682 sk->sk_shutdown |= SHUTDOWN_MASK;
691 if (smc->use_fallback) {
692 schedule_delayed_work(&smc->sock_put_work, TCP_TIMEWAIT_LEN);
693 } else if (sk->sk_state == SMC_CLOSED) {
694 smc_conn_free(&smc->conn);
695 schedule_delayed_work(&smc->sock_put_work,
696 SMC_CLOSE_SOCK_PUT_DELAY);
702 static int smc_serv_conf_first_link(struct smc_sock *smc)
704 struct smc_link_group *lgr = smc->conn.lgr;
705 struct smc_link *link;
709 link = &lgr->lnk[SMC_SINGLE_LINK];
711 rc = smc_wr_reg_send(link,
712 smc->conn.rmb_desc->mr_rx[SMC_SINGLE_LINK]);
714 return SMC_CLC_DECL_INTERR;
716 /* send CONFIRM LINK request to client over the RoCE fabric */
717 rc = smc_llc_send_confirm_link(link,
718 link->smcibdev->mac[link->ibport - 1],
719 &link->smcibdev->gid[link->ibport - 1],
722 return SMC_CLC_DECL_TCL;
724 /* receive CONFIRM LINK response from client over the RoCE fabric */
725 rest = wait_for_completion_interruptible_timeout(
726 &link->llc_confirm_resp,
727 SMC_LLC_WAIT_FIRST_TIME);
729 struct smc_clc_msg_decline dclc;
731 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
738 /* setup for RDMA connection of server */
739 static void smc_listen_work(struct work_struct *work)
741 struct smc_sock *new_smc = container_of(work, struct smc_sock,
743 struct socket *newclcsock = new_smc->clcsock;
744 struct smc_sock *lsmc = new_smc->listen_smc;
745 struct smc_clc_msg_accept_confirm cclc;
746 int local_contact = SMC_REUSE_CONTACT;
747 struct sock *newsmcsk = &new_smc->sk;
748 struct smc_clc_msg_proposal pclc;
749 struct smc_ib_device *smcibdev;
750 struct sockaddr_in peeraddr;
751 struct smc_link *link;
758 /* do inband token exchange -
759 *wait for and receive SMC Proposal CLC message
761 reason_code = smc_clc_wait_msg(new_smc, &pclc, sizeof(pclc),
768 /* IPSec connections opt out of SMC-R optimizations */
769 if (using_ipsec(new_smc)) {
770 reason_code = SMC_CLC_DECL_IPSEC;
774 /* PNET table look up: search active ib_device and port
775 * within same PNETID that also contains the ethernet device
776 * used for the internal TCP socket
778 smc_pnet_find_roce_resource(newclcsock->sk, &smcibdev, &ibport);
780 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
784 /* determine subnet and mask from internal TCP socket */
785 rc = smc_netinfo_by_tcpsk(newclcsock, &subnet, &prefix_len);
787 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
790 if ((pclc.outgoing_subnet != subnet) ||
791 (pclc.prefix_len != prefix_len)) {
792 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
796 /* get address of the peer connected to the internal TCP socket */
797 kernel_getpeername(newclcsock, (struct sockaddr *)&peeraddr, &len);
799 /* allocate connection / link group */
800 mutex_lock(&smc_create_lgr_pending);
801 local_contact = smc_conn_create(new_smc, peeraddr.sin_addr.s_addr,
802 smcibdev, ibport, &pclc.lcl, 0);
803 if (local_contact < 0) {
806 reason_code = SMC_CLC_DECL_MEM;/* insufficient memory*/
807 else if (rc == -ENOLINK)
808 reason_code = SMC_CLC_DECL_SYNCERR; /* synchr. error */
811 link = &new_smc->conn.lgr->lnk[SMC_SINGLE_LINK];
813 /* create send buffer and rmb */
814 rc = smc_buf_create(new_smc);
816 reason_code = SMC_CLC_DECL_MEM;
820 smc_close_init(new_smc);
821 smc_rx_init(new_smc);
823 if (local_contact != SMC_FIRST_CONTACT) {
824 struct smc_buf_desc *buf_desc = new_smc->conn.rmb_desc;
826 if (!buf_desc->reused) {
827 /* register memory region for new rmb */
828 rc = smc_wr_reg_send(link,
829 buf_desc->mr_rx[SMC_SINGLE_LINK]);
831 reason_code = SMC_CLC_DECL_INTERR;
836 smc_rmb_sync_sg_for_device(&new_smc->conn);
838 rc = smc_clc_send_accept(new_smc, local_contact);
842 /* receive SMC Confirm CLC message */
843 reason_code = smc_clc_wait_msg(new_smc, &cclc, sizeof(cclc),
849 smc_conn_save_peer_info(new_smc, &cclc);
850 if (local_contact == SMC_FIRST_CONTACT)
851 smc_link_save_peer_info(link, &cclc);
853 rc = smc_rmb_rtoken_handling(&new_smc->conn, &cclc);
855 reason_code = SMC_CLC_DECL_INTERR;
859 if (local_contact == SMC_FIRST_CONTACT) {
860 rc = smc_ib_ready_link(link);
862 reason_code = SMC_CLC_DECL_INTERR;
865 /* QP confirmation over RoCE fabric */
866 reason_code = smc_serv_conf_first_link(new_smc);
867 if (reason_code < 0) {
868 /* peer is not aware of a problem */
876 smc_tx_init(new_smc);
879 sk_refcnt_debug_inc(newsmcsk);
880 if (newsmcsk->sk_state == SMC_INIT)
881 newsmcsk->sk_state = SMC_ACTIVE;
883 mutex_unlock(&smc_create_lgr_pending);
884 lock_sock_nested(&lsmc->sk, SINGLE_DEPTH_NESTING);
885 if (lsmc->sk.sk_state == SMC_LISTEN) {
886 smc_accept_enqueue(&lsmc->sk, newsmcsk);
887 } else { /* no longer listening */
888 smc_close_non_accepted(newsmcsk);
890 release_sock(&lsmc->sk);
893 lsmc->sk.sk_data_ready(&lsmc->sk);
894 sock_put(&lsmc->sk); /* sock_hold in smc_tcp_listen_work */
898 /* RDMA setup failed, switch back to TCP */
899 smc_conn_free(&new_smc->conn);
900 new_smc->use_fallback = true;
901 if (reason_code && (reason_code != SMC_CLC_DECL_REPLY)) {
902 rc = smc_clc_send_decline(new_smc, reason_code, 0);
903 if (rc < sizeof(struct smc_clc_msg_decline))
909 newsmcsk->sk_state = SMC_CLOSED;
910 smc_conn_free(&new_smc->conn);
911 goto enqueue; /* queue new sock with sk_err set */
914 static void smc_tcp_listen_work(struct work_struct *work)
916 struct smc_sock *lsmc = container_of(work, struct smc_sock,
918 struct smc_sock *new_smc;
921 lock_sock(&lsmc->sk);
922 while (lsmc->sk.sk_state == SMC_LISTEN) {
923 rc = smc_clcsock_accept(lsmc, &new_smc);
929 new_smc->listen_smc = lsmc;
930 new_smc->use_fallback = false; /* assume rdma capability first*/
931 sock_hold(&lsmc->sk); /* sock_put in smc_listen_work */
932 INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
933 smc_copy_sock_settings_to_smc(new_smc);
934 schedule_work(&new_smc->smc_listen_work);
938 release_sock(&lsmc->sk);
939 lsmc->sk.sk_data_ready(&lsmc->sk); /* no more listening, wake accept */
942 static int smc_listen(struct socket *sock, int backlog)
944 struct sock *sk = sock->sk;
945 struct smc_sock *smc;
952 if ((sk->sk_state != SMC_INIT) && (sk->sk_state != SMC_LISTEN))
956 if (sk->sk_state == SMC_LISTEN) {
957 sk->sk_max_ack_backlog = backlog;
960 /* some socket options are handled in core, so we could not apply
961 * them to the clc socket -- copy smc socket options to clc socket
963 smc_copy_sock_settings_to_clc(smc);
965 rc = kernel_listen(smc->clcsock, backlog);
968 sk->sk_max_ack_backlog = backlog;
969 sk->sk_ack_backlog = 0;
970 sk->sk_state = SMC_LISTEN;
971 INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
972 schedule_work(&smc->tcp_listen_work);
979 static int smc_accept(struct socket *sock, struct socket *new_sock,
980 int flags, bool kern)
982 struct sock *sk = sock->sk, *nsk;
983 DECLARE_WAITQUEUE(wait, current);
984 struct smc_sock *lsmc;
991 if (lsmc->sk.sk_state != SMC_LISTEN) {
996 /* Wait for an incoming connection */
997 timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
998 add_wait_queue_exclusive(sk_sleep(sk), &wait);
999 while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
1000 set_current_state(TASK_INTERRUPTIBLE);
1006 timeo = schedule_timeout(timeo);
1007 /* wakeup by sk_data_ready in smc_listen_work() */
1008 sched_annotate_sleep();
1010 if (signal_pending(current)) {
1011 rc = sock_intr_errno(timeo);
1015 set_current_state(TASK_RUNNING);
1016 remove_wait_queue(sk_sleep(sk), &wait);
1019 rc = sock_error(nsk);
1026 static int smc_getname(struct socket *sock, struct sockaddr *addr,
1029 struct smc_sock *smc;
1031 if (peer && (sock->sk->sk_state != SMC_ACTIVE) &&
1032 (sock->sk->sk_state != SMC_APPCLOSEWAIT1))
1035 smc = smc_sk(sock->sk);
1037 return smc->clcsock->ops->getname(smc->clcsock, addr, len, peer);
1040 static int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
1042 struct sock *sk = sock->sk;
1043 struct smc_sock *smc;
1048 if ((sk->sk_state != SMC_ACTIVE) &&
1049 (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1050 (sk->sk_state != SMC_INIT))
1052 if (smc->use_fallback)
1053 rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
1055 rc = smc_tx_sendmsg(smc, msg, len);
1061 static int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
1064 struct sock *sk = sock->sk;
1065 struct smc_sock *smc;
1070 if ((sk->sk_state == SMC_INIT) ||
1071 (sk->sk_state == SMC_LISTEN) ||
1072 (sk->sk_state == SMC_CLOSED))
1075 if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
1080 if (smc->use_fallback)
1081 rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
1083 rc = smc_rx_recvmsg(smc, msg, len, flags);
1090 static unsigned int smc_accept_poll(struct sock *parent)
1092 struct smc_sock *isk;
1096 list_for_each_entry(isk, &smc_sk(parent)->accept_q, accept_q) {
1097 sk = (struct sock *)isk;
1099 if (sk->sk_state == SMC_ACTIVE) {
1100 release_sock(parent);
1101 return POLLIN | POLLRDNORM;
1104 release_sock(parent);
1109 static unsigned int smc_poll(struct file *file, struct socket *sock,
1112 struct sock *sk = sock->sk;
1113 unsigned int mask = 0;
1114 struct smc_sock *smc;
1117 smc = smc_sk(sock->sk);
1118 if ((sk->sk_state == SMC_INIT) || smc->use_fallback) {
1119 /* delegate to CLC child sock */
1120 mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
1121 /* if non-blocking connect finished ... */
1123 if ((sk->sk_state == SMC_INIT) && (mask & POLLOUT)) {
1124 sk->sk_err = smc->clcsock->sk->sk_err;
1128 rc = smc_connect_rdma(smc);
1132 /* success cases including fallback */
1133 mask |= POLLOUT | POLLWRNORM;
1138 sock_poll_wait(file, sk_sleep(sk), wait);
1139 if (sk->sk_state == SMC_LISTEN)
1140 /* woken up by sk_data_ready in smc_listen_work() */
1141 mask |= smc_accept_poll(sk);
1144 if (atomic_read(&smc->conn.sndbuf_space) ||
1145 (sk->sk_shutdown & SEND_SHUTDOWN)) {
1146 mask |= POLLOUT | POLLWRNORM;
1148 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1149 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1151 if (atomic_read(&smc->conn.bytes_to_rcv))
1152 mask |= POLLIN | POLLRDNORM;
1153 if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
1154 (sk->sk_state == SMC_CLOSED))
1156 if (sk->sk_shutdown & RCV_SHUTDOWN)
1157 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
1158 if (sk->sk_state == SMC_APPCLOSEWAIT1)
1166 static int smc_shutdown(struct socket *sock, int how)
1168 struct sock *sk = sock->sk;
1169 struct smc_sock *smc;
1175 if ((how < SHUT_RD) || (how > SHUT_RDWR))
1181 if ((sk->sk_state != SMC_LISTEN) &&
1182 (sk->sk_state != SMC_ACTIVE) &&
1183 (sk->sk_state != SMC_PEERCLOSEWAIT1) &&
1184 (sk->sk_state != SMC_PEERCLOSEWAIT2) &&
1185 (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1186 (sk->sk_state != SMC_APPCLOSEWAIT2) &&
1187 (sk->sk_state != SMC_APPFINCLOSEWAIT))
1189 if (smc->use_fallback) {
1190 rc = kernel_sock_shutdown(smc->clcsock, how);
1191 sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
1192 if (sk->sk_shutdown == SHUTDOWN_MASK)
1193 sk->sk_state = SMC_CLOSED;
1197 case SHUT_RDWR: /* shutdown in both directions */
1198 rc = smc_close_active(smc);
1201 rc = smc_close_shutdown_write(smc);
1204 if (sk->sk_state == SMC_LISTEN)
1205 rc = smc_close_active(smc);
1208 /* nothing more to do because peer is not involved */
1211 rc1 = kernel_sock_shutdown(smc->clcsock, how);
1212 /* map sock_shutdown_cmd constants to sk_shutdown value range */
1213 sk->sk_shutdown |= how + 1;
1217 return rc ? rc : rc1;
1220 static int smc_setsockopt(struct socket *sock, int level, int optname,
1221 char __user *optval, unsigned int optlen)
1223 struct sock *sk = sock->sk;
1224 struct smc_sock *smc;
1228 /* generic setsockopts reaching us here always apply to the
1231 return smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
1235 static int smc_getsockopt(struct socket *sock, int level, int optname,
1236 char __user *optval, int __user *optlen)
1238 struct smc_sock *smc;
1240 smc = smc_sk(sock->sk);
1241 /* socket options apply to the CLC socket */
1242 return smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
1246 static int smc_ioctl(struct socket *sock, unsigned int cmd,
1249 struct smc_sock *smc;
1251 smc = smc_sk(sock->sk);
1252 if (smc->use_fallback)
1253 return smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
1255 return sock_no_ioctl(sock, cmd, arg);
1258 static ssize_t smc_sendpage(struct socket *sock, struct page *page,
1259 int offset, size_t size, int flags)
1261 struct sock *sk = sock->sk;
1262 struct smc_sock *smc;
1267 if (sk->sk_state != SMC_ACTIVE)
1269 if (smc->use_fallback)
1270 rc = kernel_sendpage(smc->clcsock, page, offset,
1273 rc = sock_no_sendpage(sock, page, offset, size, flags);
1280 static ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
1281 struct pipe_inode_info *pipe, size_t len,
1284 struct sock *sk = sock->sk;
1285 struct smc_sock *smc;
1290 if ((sk->sk_state != SMC_ACTIVE) && (sk->sk_state != SMC_CLOSED))
1292 if (smc->use_fallback) {
1293 rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
1303 /* must look like tcp */
1304 static const struct proto_ops smc_sock_ops = {
1306 .owner = THIS_MODULE,
1307 .release = smc_release,
1309 .connect = smc_connect,
1310 .socketpair = sock_no_socketpair,
1311 .accept = smc_accept,
1312 .getname = smc_getname,
1315 .listen = smc_listen,
1316 .shutdown = smc_shutdown,
1317 .setsockopt = smc_setsockopt,
1318 .getsockopt = smc_getsockopt,
1319 .sendmsg = smc_sendmsg,
1320 .recvmsg = smc_recvmsg,
1321 .mmap = sock_no_mmap,
1322 .sendpage = smc_sendpage,
1323 .splice_read = smc_splice_read,
1326 static int smc_create(struct net *net, struct socket *sock, int protocol,
1329 struct smc_sock *smc;
1333 rc = -ESOCKTNOSUPPORT;
1334 if (sock->type != SOCK_STREAM)
1337 rc = -EPROTONOSUPPORT;
1338 if ((protocol != IPPROTO_IP) && (protocol != IPPROTO_TCP))
1342 sock->ops = &smc_sock_ops;
1343 sk = smc_sock_alloc(net, sock);
1347 /* create internal TCP socket for CLC handshake and fallback */
1349 smc->use_fallback = false; /* assume rdma capability first */
1350 rc = sock_create_kern(net, PF_INET, SOCK_STREAM,
1351 IPPROTO_TCP, &smc->clcsock);
1353 sk_common_release(sk);
1354 smc->sk.sk_sndbuf = max(smc->clcsock->sk->sk_sndbuf, SMC_BUF_MIN_SIZE);
1355 smc->sk.sk_rcvbuf = max(smc->clcsock->sk->sk_rcvbuf, SMC_BUF_MIN_SIZE);
1361 static const struct net_proto_family smc_sock_family_ops = {
1363 .owner = THIS_MODULE,
1364 .create = smc_create,
1367 static int __init smc_init(void)
1371 rc = smc_pnet_init();
1375 rc = smc_llc_init();
1377 pr_err("%s: smc_llc_init fails with %d\n", __func__, rc);
1381 rc = smc_cdc_init();
1383 pr_err("%s: smc_cdc_init fails with %d\n", __func__, rc);
1387 rc = proto_register(&smc_proto, 1);
1389 pr_err("%s: proto_register fails with %d\n", __func__, rc);
1393 rc = sock_register(&smc_sock_family_ops);
1395 pr_err("%s: sock_register fails with %d\n", __func__, rc);
1398 INIT_HLIST_HEAD(&smc_v4_hashinfo.ht);
1400 rc = smc_ib_register_client();
1402 pr_err("%s: ib_register fails with %d\n", __func__, rc);
1409 sock_unregister(PF_SMC);
1411 proto_unregister(&smc_proto);
1417 static void __exit smc_exit(void)
1419 struct smc_link_group *lgr, *lg;
1420 LIST_HEAD(lgr_freeing_list);
1422 spin_lock_bh(&smc_lgr_list.lock);
1423 if (!list_empty(&smc_lgr_list.list))
1424 list_splice_init(&smc_lgr_list.list, &lgr_freeing_list);
1425 spin_unlock_bh(&smc_lgr_list.lock);
1426 list_for_each_entry_safe(lgr, lg, &lgr_freeing_list, list) {
1427 list_del_init(&lgr->list);
1428 smc_lgr_free(lgr); /* free link group */
1430 smc_ib_unregister_client();
1431 sock_unregister(PF_SMC);
1432 proto_unregister(&smc_proto);
1436 module_init(smc_init);
1437 module_exit(smc_exit);
1439 MODULE_AUTHOR("Ursula Braun <ubraun@linux.vnet.ibm.com>");
1440 MODULE_DESCRIPTION("smc socket address family");
1441 MODULE_LICENSE("GPL");
1442 MODULE_ALIAS_NETPROTO(PF_SMC);