1 // SPDX-License-Identifier: GPL-2.0-only
3 * Shared Memory Communications over RDMA (SMC-R) and RoCE
5 * AF_SMC protocol family socket handler keeping the AF_INET sock address type
6 * applies to SOCK_STREAM sockets only
7 * offers an alternative communication option for TCP-protocol sockets
8 * applicable with RoCE-cards only
10 * Initial restrictions:
11 * - support for alternate links postponed
13 * Copyright IBM Corp. 2016, 2018
15 * Author(s): Ursula Braun <ubraun@linux.vnet.ibm.com>
16 * based on prototype from Frank Blaschka
19 #define KMSG_COMPONENT "smc"
20 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
22 #include <linux/module.h>
23 #include <linux/socket.h>
24 #include <linux/workqueue.h>
26 #include <linux/sched/signal.h>
27 #include <linux/if_vlan.h>
28 #include <linux/rcupdate_wait.h>
29 #include <linux/ctype.h>
30 #include <linux/splice.h>
35 #include <asm/ioctls.h>
37 #include <net/net_namespace.h>
38 #include <net/netns/generic.h>
39 #include "smc_netns.h"
49 #include "smc_netlink.h"
52 #include "smc_close.h"
53 #include "smc_stats.h"
54 #include "smc_tracepoint.h"
55 #include "smc_sysctl.h"
57 static DEFINE_MUTEX(smc_server_lgr_pending); /* serialize link group
60 static DEFINE_MUTEX(smc_client_lgr_pending); /* serialize link group
64 static struct workqueue_struct *smc_tcp_ls_wq; /* wq for tcp listen work */
65 struct workqueue_struct *smc_hs_wq; /* wq for handshake work */
66 struct workqueue_struct *smc_close_wq; /* wq for close work */
68 static void smc_tcp_listen_work(struct work_struct *);
69 static void smc_connect_work(struct work_struct *);
71 int smc_nl_dump_hs_limitation(struct sk_buff *skb, struct netlink_callback *cb)
73 struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
79 hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
80 &smc_gen_nl_family, NLM_F_MULTI,
81 SMC_NETLINK_DUMP_HS_LIMITATION);
85 if (nla_put_u8(skb, SMC_NLA_HS_LIMITATION_ENABLED,
86 sock_net(skb->sk)->smc.limit_smc_hs))
89 genlmsg_end(skb, hdr);
94 genlmsg_cancel(skb, hdr);
98 int smc_nl_enable_hs_limitation(struct sk_buff *skb, struct genl_info *info)
100 sock_net(skb->sk)->smc.limit_smc_hs = true;
104 int smc_nl_disable_hs_limitation(struct sk_buff *skb, struct genl_info *info)
106 sock_net(skb->sk)->smc.limit_smc_hs = false;
110 static void smc_set_keepalive(struct sock *sk, int val)
112 struct smc_sock *smc = smc_sk(sk);
114 smc->clcsock->sk->sk_prot->keepalive(smc->clcsock->sk, val);
117 static struct sock *smc_tcp_syn_recv_sock(const struct sock *sk,
119 struct request_sock *req,
120 struct dst_entry *dst,
121 struct request_sock *req_unhash,
124 struct smc_sock *smc;
127 smc = smc_clcsock_user_data(sk);
129 if (READ_ONCE(sk->sk_ack_backlog) + atomic_read(&smc->queued_smc_hs) >
130 sk->sk_max_ack_backlog)
133 if (sk_acceptq_is_full(&smc->sk)) {
134 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
138 /* passthrough to original syn recv sock fct */
139 child = smc->ori_af_ops->syn_recv_sock(sk, skb, req, dst, req_unhash,
141 /* child must not inherit smc or its ops */
143 rcu_assign_sk_user_data(child, NULL);
145 /* v4-mapped sockets don't inherit parent ops. Don't restore. */
146 if (inet_csk(child)->icsk_af_ops == inet_csk(sk)->icsk_af_ops)
147 inet_csk(child)->icsk_af_ops = smc->ori_af_ops;
157 static bool smc_hs_congested(const struct sock *sk)
159 const struct smc_sock *smc;
161 smc = smc_clcsock_user_data(sk);
166 if (workqueue_congested(WORK_CPU_UNBOUND, smc_hs_wq))
172 static struct smc_hashinfo smc_v4_hashinfo = {
173 .lock = __RW_LOCK_UNLOCKED(smc_v4_hashinfo.lock),
176 static struct smc_hashinfo smc_v6_hashinfo = {
177 .lock = __RW_LOCK_UNLOCKED(smc_v6_hashinfo.lock),
180 int smc_hash_sk(struct sock *sk)
182 struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
183 struct hlist_head *head;
187 write_lock_bh(&h->lock);
188 sk_add_node(sk, head);
189 write_unlock_bh(&h->lock);
190 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
194 EXPORT_SYMBOL_GPL(smc_hash_sk);
196 void smc_unhash_sk(struct sock *sk)
198 struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
200 write_lock_bh(&h->lock);
201 if (sk_del_node_init(sk))
202 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
203 write_unlock_bh(&h->lock);
205 EXPORT_SYMBOL_GPL(smc_unhash_sk);
207 /* This will be called before user really release sock_lock. So do the
208 * work which we didn't do because of user hold the sock_lock in the
211 static void smc_release_cb(struct sock *sk)
213 struct smc_sock *smc = smc_sk(sk);
215 if (smc->conn.tx_in_release_sock) {
216 smc_tx_pending(&smc->conn);
217 smc->conn.tx_in_release_sock = false;
221 struct proto smc_proto = {
223 .owner = THIS_MODULE,
224 .keepalive = smc_set_keepalive,
226 .unhash = smc_unhash_sk,
227 .release_cb = smc_release_cb,
228 .obj_size = sizeof(struct smc_sock),
229 .h.smc_hash = &smc_v4_hashinfo,
230 .slab_flags = SLAB_TYPESAFE_BY_RCU,
232 EXPORT_SYMBOL_GPL(smc_proto);
234 struct proto smc_proto6 = {
236 .owner = THIS_MODULE,
237 .keepalive = smc_set_keepalive,
239 .unhash = smc_unhash_sk,
240 .release_cb = smc_release_cb,
241 .obj_size = sizeof(struct smc_sock),
242 .h.smc_hash = &smc_v6_hashinfo,
243 .slab_flags = SLAB_TYPESAFE_BY_RCU,
245 EXPORT_SYMBOL_GPL(smc_proto6);
247 static void smc_fback_restore_callbacks(struct smc_sock *smc)
249 struct sock *clcsk = smc->clcsock->sk;
251 write_lock_bh(&clcsk->sk_callback_lock);
252 clcsk->sk_user_data = NULL;
254 smc_clcsock_restore_cb(&clcsk->sk_state_change, &smc->clcsk_state_change);
255 smc_clcsock_restore_cb(&clcsk->sk_data_ready, &smc->clcsk_data_ready);
256 smc_clcsock_restore_cb(&clcsk->sk_write_space, &smc->clcsk_write_space);
257 smc_clcsock_restore_cb(&clcsk->sk_error_report, &smc->clcsk_error_report);
259 write_unlock_bh(&clcsk->sk_callback_lock);
262 static void smc_restore_fallback_changes(struct smc_sock *smc)
264 if (smc->clcsock->file) { /* non-accepted sockets have no file yet */
265 smc->clcsock->file->private_data = smc->sk.sk_socket;
266 smc->clcsock->file = NULL;
267 smc_fback_restore_callbacks(smc);
271 static int __smc_release(struct smc_sock *smc)
273 struct sock *sk = &smc->sk;
276 if (!smc->use_fallback) {
277 rc = smc_close_active(smc);
278 smc_sock_set_flag(sk, SOCK_DEAD);
279 sk->sk_shutdown |= SHUTDOWN_MASK;
281 if (sk->sk_state != SMC_CLOSED) {
282 if (sk->sk_state != SMC_LISTEN &&
283 sk->sk_state != SMC_INIT)
284 sock_put(sk); /* passive closing */
285 if (sk->sk_state == SMC_LISTEN) {
286 /* wake up clcsock accept */
287 rc = kernel_sock_shutdown(smc->clcsock,
290 sk->sk_state = SMC_CLOSED;
291 sk->sk_state_change(sk);
293 smc_restore_fallback_changes(smc);
296 sk->sk_prot->unhash(sk);
298 if (sk->sk_state == SMC_CLOSED) {
301 smc_clcsock_release(smc);
304 if (!smc->use_fallback)
305 smc_conn_free(&smc->conn);
311 static int smc_release(struct socket *sock)
313 struct sock *sk = sock->sk;
314 struct smc_sock *smc;
315 int old_state, rc = 0;
320 sock_hold(sk); /* sock_put below */
323 old_state = sk->sk_state;
325 /* cleanup for a dangling non-blocking connect */
326 if (smc->connect_nonblock && old_state == SMC_INIT)
327 tcp_abort(smc->clcsock->sk, ECONNABORTED);
329 if (cancel_work_sync(&smc->connect_work))
330 sock_put(&smc->sk); /* sock_hold in smc_connect for passive closing */
332 if (sk->sk_state == SMC_LISTEN)
333 /* smc_close_non_accepted() is called and acquires
334 * sock lock for child sockets again
336 lock_sock_nested(sk, SINGLE_DEPTH_NESTING);
340 if (old_state == SMC_INIT && sk->sk_state == SMC_ACTIVE &&
342 smc_close_active_abort(smc);
344 rc = __smc_release(smc);
351 sock_put(sk); /* sock_hold above */
352 sock_put(sk); /* final sock_put */
357 static void smc_destruct(struct sock *sk)
359 if (sk->sk_state != SMC_CLOSED)
361 if (!sock_flag(sk, SOCK_DEAD))
365 static struct sock *smc_sock_alloc(struct net *net, struct socket *sock,
368 struct smc_sock *smc;
372 prot = (protocol == SMCPROTO_SMC6) ? &smc_proto6 : &smc_proto;
373 sk = sk_alloc(net, PF_SMC, GFP_KERNEL, prot, 0);
377 sock_init_data(sock, sk); /* sets sk_refcnt to 1 */
378 sk->sk_state = SMC_INIT;
379 sk->sk_destruct = smc_destruct;
380 sk->sk_protocol = protocol;
381 WRITE_ONCE(sk->sk_sndbuf, 2 * READ_ONCE(net->smc.sysctl_wmem));
382 WRITE_ONCE(sk->sk_rcvbuf, 2 * READ_ONCE(net->smc.sysctl_rmem));
384 INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
385 INIT_WORK(&smc->connect_work, smc_connect_work);
386 INIT_DELAYED_WORK(&smc->conn.tx_work, smc_tx_work);
387 INIT_LIST_HEAD(&smc->accept_q);
388 spin_lock_init(&smc->accept_q_lock);
389 spin_lock_init(&smc->conn.send_lock);
390 sk->sk_prot->hash(sk);
391 mutex_init(&smc->clcsock_release_lock);
392 smc_init_saved_callbacks(smc);
397 static int smc_bind(struct socket *sock, struct sockaddr *uaddr,
400 struct sockaddr_in *addr = (struct sockaddr_in *)uaddr;
401 struct sock *sk = sock->sk;
402 struct smc_sock *smc;
407 /* replicate tests from inet_bind(), to be safe wrt. future changes */
409 if (addr_len < sizeof(struct sockaddr_in))
413 if (addr->sin_family != AF_INET &&
414 addr->sin_family != AF_INET6 &&
415 addr->sin_family != AF_UNSPEC)
417 /* accept AF_UNSPEC (mapped to AF_INET) only if s_addr is INADDR_ANY */
418 if (addr->sin_family == AF_UNSPEC &&
419 addr->sin_addr.s_addr != htonl(INADDR_ANY))
424 /* Check if socket is already active */
426 if (sk->sk_state != SMC_INIT || smc->connect_nonblock)
429 smc->clcsock->sk->sk_reuse = sk->sk_reuse;
430 smc->clcsock->sk->sk_reuseport = sk->sk_reuseport;
431 rc = kernel_bind(smc->clcsock, uaddr, addr_len);
439 /* copy only relevant settings and flags of SOL_SOCKET level from smc to
440 * clc socket (since smc is not called for these options from net/core)
443 #define SK_FLAGS_SMC_TO_CLC ((1UL << SOCK_URGINLINE) | \
444 (1UL << SOCK_KEEPOPEN) | \
445 (1UL << SOCK_LINGER) | \
446 (1UL << SOCK_BROADCAST) | \
447 (1UL << SOCK_TIMESTAMP) | \
448 (1UL << SOCK_DBG) | \
449 (1UL << SOCK_RCVTSTAMP) | \
450 (1UL << SOCK_RCVTSTAMPNS) | \
451 (1UL << SOCK_LOCALROUTE) | \
452 (1UL << SOCK_TIMESTAMPING_RX_SOFTWARE) | \
453 (1UL << SOCK_RXQ_OVFL) | \
454 (1UL << SOCK_WIFI_STATUS) | \
455 (1UL << SOCK_NOFCS) | \
456 (1UL << SOCK_FILTER_LOCKED) | \
457 (1UL << SOCK_TSTAMP_NEW))
459 /* if set, use value set by setsockopt() - else use IPv4 or SMC sysctl value */
460 static void smc_adjust_sock_bufsizes(struct sock *nsk, struct sock *osk,
463 struct net *nnet = sock_net(nsk);
465 nsk->sk_userlocks = osk->sk_userlocks;
466 if (osk->sk_userlocks & SOCK_SNDBUF_LOCK) {
467 nsk->sk_sndbuf = osk->sk_sndbuf;
469 if (mask == SK_FLAGS_SMC_TO_CLC)
470 WRITE_ONCE(nsk->sk_sndbuf,
471 READ_ONCE(nnet->ipv4.sysctl_tcp_wmem[1]));
473 WRITE_ONCE(nsk->sk_sndbuf,
474 2 * READ_ONCE(nnet->smc.sysctl_wmem));
476 if (osk->sk_userlocks & SOCK_RCVBUF_LOCK) {
477 nsk->sk_rcvbuf = osk->sk_rcvbuf;
479 if (mask == SK_FLAGS_SMC_TO_CLC)
480 WRITE_ONCE(nsk->sk_rcvbuf,
481 READ_ONCE(nnet->ipv4.sysctl_tcp_rmem[1]));
483 WRITE_ONCE(nsk->sk_rcvbuf,
484 2 * READ_ONCE(nnet->smc.sysctl_rmem));
488 static void smc_copy_sock_settings(struct sock *nsk, struct sock *osk,
491 /* options we don't get control via setsockopt for */
492 nsk->sk_type = osk->sk_type;
493 nsk->sk_sndtimeo = osk->sk_sndtimeo;
494 nsk->sk_rcvtimeo = osk->sk_rcvtimeo;
495 nsk->sk_mark = READ_ONCE(osk->sk_mark);
496 nsk->sk_priority = READ_ONCE(osk->sk_priority);
497 nsk->sk_rcvlowat = osk->sk_rcvlowat;
498 nsk->sk_bound_dev_if = osk->sk_bound_dev_if;
499 nsk->sk_err = osk->sk_err;
501 nsk->sk_flags &= ~mask;
502 nsk->sk_flags |= osk->sk_flags & mask;
504 smc_adjust_sock_bufsizes(nsk, osk, mask);
507 static void smc_copy_sock_settings_to_clc(struct smc_sock *smc)
509 smc_copy_sock_settings(smc->clcsock->sk, &smc->sk, SK_FLAGS_SMC_TO_CLC);
512 #define SK_FLAGS_CLC_TO_SMC ((1UL << SOCK_URGINLINE) | \
513 (1UL << SOCK_KEEPOPEN) | \
514 (1UL << SOCK_LINGER) | \
516 /* copy only settings and flags relevant for smc from clc to smc socket */
517 static void smc_copy_sock_settings_to_smc(struct smc_sock *smc)
519 smc_copy_sock_settings(&smc->sk, smc->clcsock->sk, SK_FLAGS_CLC_TO_SMC);
522 /* register the new vzalloced sndbuf on all links */
523 static int smcr_lgr_reg_sndbufs(struct smc_link *link,
524 struct smc_buf_desc *snd_desc)
526 struct smc_link_group *lgr = link->lgr;
529 if (!snd_desc->is_vm)
532 /* protect against parallel smcr_link_reg_buf() */
533 down_write(&lgr->llc_conf_mutex);
534 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
535 if (!smc_link_active(&lgr->lnk[i]))
537 rc = smcr_link_reg_buf(&lgr->lnk[i], snd_desc);
541 up_write(&lgr->llc_conf_mutex);
545 /* register the new rmb on all links */
546 static int smcr_lgr_reg_rmbs(struct smc_link *link,
547 struct smc_buf_desc *rmb_desc)
549 struct smc_link_group *lgr = link->lgr;
550 bool do_slow = false;
553 rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY);
557 down_read(&lgr->llc_conf_mutex);
558 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
559 if (!smc_link_active(&lgr->lnk[i]))
561 if (!rmb_desc->is_reg_mr[link->link_idx]) {
562 up_read(&lgr->llc_conf_mutex);
566 /* mr register already */
570 /* protect against parallel smc_llc_cli_rkey_exchange() and
571 * parallel smcr_link_reg_buf()
573 down_write(&lgr->llc_conf_mutex);
574 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
575 if (!smc_link_active(&lgr->lnk[i]))
577 rc = smcr_link_reg_buf(&lgr->lnk[i], rmb_desc);
582 /* exchange confirm_rkey msg with peer */
583 rc = smc_llc_do_confirm_rkey(link, rmb_desc);
588 rmb_desc->is_conf_rkey = true;
590 do_slow ? up_write(&lgr->llc_conf_mutex) : up_read(&lgr->llc_conf_mutex);
591 smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl);
595 static int smcr_clnt_conf_first_link(struct smc_sock *smc)
597 struct smc_link *link = smc->conn.lnk;
598 struct smc_llc_qentry *qentry;
601 /* Receive CONFIRM LINK request from server over RoCE fabric.
602 * Increasing the client's timeout by twice as much as the server's
603 * timeout by default can temporarily avoid decline messages of
604 * both sides crossing or colliding
606 qentry = smc_llc_wait(link->lgr, NULL, 2 * SMC_LLC_WAIT_TIME,
607 SMC_LLC_CONFIRM_LINK);
609 struct smc_clc_msg_decline dclc;
611 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
612 SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
613 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
615 smc_llc_save_peer_uid(qentry);
616 rc = smc_llc_eval_conf_link(qentry, SMC_LLC_REQ);
617 smc_llc_flow_qentry_del(&link->lgr->llc_flow_lcl);
619 return SMC_CLC_DECL_RMBE_EC;
621 rc = smc_ib_modify_qp_rts(link);
623 return SMC_CLC_DECL_ERR_RDYLNK;
625 smc_wr_remember_qp_attr(link);
627 /* reg the sndbuf if it was vzalloced */
628 if (smc->conn.sndbuf_desc->is_vm) {
629 if (smcr_link_reg_buf(link, smc->conn.sndbuf_desc))
630 return SMC_CLC_DECL_ERR_REGBUF;
634 if (smcr_link_reg_buf(link, smc->conn.rmb_desc))
635 return SMC_CLC_DECL_ERR_REGBUF;
637 /* confirm_rkey is implicit on 1st contact */
638 smc->conn.rmb_desc->is_conf_rkey = true;
640 /* send CONFIRM LINK response over RoCE fabric */
641 rc = smc_llc_send_confirm_link(link, SMC_LLC_RESP);
643 return SMC_CLC_DECL_TIMEOUT_CL;
645 smc_llc_link_active(link);
646 smcr_lgr_set_type(link->lgr, SMC_LGR_SINGLE);
648 if (link->lgr->max_links > 1) {
649 /* optional 2nd link, receive ADD LINK request from server */
650 qentry = smc_llc_wait(link->lgr, NULL, SMC_LLC_WAIT_TIME,
653 struct smc_clc_msg_decline dclc;
655 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
656 SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
658 rc = 0; /* no DECLINE received, go with one link */
661 smc_llc_flow_qentry_clr(&link->lgr->llc_flow_lcl);
662 smc_llc_cli_add_link(link, qentry);
667 static bool smc_isascii(char *hostname)
671 for (i = 0; i < SMC_MAX_HOSTNAME_LEN; i++)
672 if (!isascii(hostname[i]))
677 static void smc_conn_save_peer_info_fce(struct smc_sock *smc,
678 struct smc_clc_msg_accept_confirm *clc)
680 struct smc_clc_first_contact_ext *fce;
683 if (clc->hdr.version == SMC_V1 ||
684 !(clc->hdr.typev2 & SMC_FIRST_CONTACT_MASK))
687 if (smc->conn.lgr->is_smcd) {
688 memcpy(smc->conn.lgr->negotiated_eid, clc->d1.eid,
690 clc_v2_len = offsetofend(struct smc_clc_msg_accept_confirm, d1);
692 memcpy(smc->conn.lgr->negotiated_eid, clc->r1.eid,
694 clc_v2_len = offsetofend(struct smc_clc_msg_accept_confirm, r1);
696 fce = (struct smc_clc_first_contact_ext *)(((u8 *)clc) + clc_v2_len);
697 smc->conn.lgr->peer_os = fce->os_type;
698 smc->conn.lgr->peer_smc_release = fce->release;
699 if (smc_isascii(fce->hostname))
700 memcpy(smc->conn.lgr->peer_hostname, fce->hostname,
701 SMC_MAX_HOSTNAME_LEN);
704 static void smcr_conn_save_peer_info(struct smc_sock *smc,
705 struct smc_clc_msg_accept_confirm *clc)
707 int bufsize = smc_uncompress_bufsize(clc->r0.rmbe_size);
709 smc->conn.peer_rmbe_idx = clc->r0.rmbe_idx;
710 smc->conn.local_tx_ctrl.token = ntohl(clc->r0.rmbe_alert_token);
711 smc->conn.peer_rmbe_size = bufsize;
712 atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
713 smc->conn.tx_off = bufsize * (smc->conn.peer_rmbe_idx - 1);
716 static void smcd_conn_save_peer_info(struct smc_sock *smc,
717 struct smc_clc_msg_accept_confirm *clc)
719 int bufsize = smc_uncompress_bufsize(clc->d0.dmbe_size);
721 smc->conn.peer_rmbe_idx = clc->d0.dmbe_idx;
722 smc->conn.peer_token = ntohll(clc->d0.token);
723 /* msg header takes up space in the buffer */
724 smc->conn.peer_rmbe_size = bufsize - sizeof(struct smcd_cdc_msg);
725 atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
726 smc->conn.tx_off = bufsize * smc->conn.peer_rmbe_idx;
729 static void smc_conn_save_peer_info(struct smc_sock *smc,
730 struct smc_clc_msg_accept_confirm *clc)
732 if (smc->conn.lgr->is_smcd)
733 smcd_conn_save_peer_info(smc, clc);
735 smcr_conn_save_peer_info(smc, clc);
736 smc_conn_save_peer_info_fce(smc, clc);
739 static void smc_link_save_peer_info(struct smc_link *link,
740 struct smc_clc_msg_accept_confirm *clc,
741 struct smc_init_info *ini)
743 link->peer_qpn = ntoh24(clc->r0.qpn);
744 memcpy(link->peer_gid, ini->peer_gid, SMC_GID_SIZE);
745 memcpy(link->peer_mac, ini->peer_mac, sizeof(link->peer_mac));
746 link->peer_psn = ntoh24(clc->r0.psn);
747 link->peer_mtu = clc->r0.qp_mtu;
750 static void smc_stat_inc_fback_rsn_cnt(struct smc_sock *smc,
751 struct smc_stats_fback *fback_arr)
755 for (cnt = 0; cnt < SMC_MAX_FBACK_RSN_CNT; cnt++) {
756 if (fback_arr[cnt].fback_code == smc->fallback_rsn) {
757 fback_arr[cnt].count++;
760 if (!fback_arr[cnt].fback_code) {
761 fback_arr[cnt].fback_code = smc->fallback_rsn;
762 fback_arr[cnt].count++;
768 static void smc_stat_fallback(struct smc_sock *smc)
770 struct net *net = sock_net(&smc->sk);
772 mutex_lock(&net->smc.mutex_fback_rsn);
773 if (smc->listen_smc) {
774 smc_stat_inc_fback_rsn_cnt(smc, net->smc.fback_rsn->srv);
775 net->smc.fback_rsn->srv_fback_cnt++;
777 smc_stat_inc_fback_rsn_cnt(smc, net->smc.fback_rsn->clnt);
778 net->smc.fback_rsn->clnt_fback_cnt++;
780 mutex_unlock(&net->smc.mutex_fback_rsn);
783 /* must be called under rcu read lock */
784 static void smc_fback_wakeup_waitqueue(struct smc_sock *smc, void *key)
786 struct socket_wq *wq;
789 wq = rcu_dereference(smc->sk.sk_wq);
790 if (!skwq_has_sleeper(wq))
793 /* wake up smc sk->sk_wq */
795 /* sk_state_change */
796 wake_up_interruptible_all(&wq->wait);
798 flags = key_to_poll(key);
799 if (flags & (EPOLLIN | EPOLLOUT))
800 /* sk_data_ready or sk_write_space */
801 wake_up_interruptible_sync_poll(&wq->wait, flags);
802 else if (flags & EPOLLERR)
803 /* sk_error_report */
804 wake_up_interruptible_poll(&wq->wait, flags);
808 static int smc_fback_mark_woken(wait_queue_entry_t *wait,
809 unsigned int mode, int sync, void *key)
811 struct smc_mark_woken *mark =
812 container_of(wait, struct smc_mark_woken, wait_entry);
819 static void smc_fback_forward_wakeup(struct smc_sock *smc, struct sock *clcsk,
820 void (*clcsock_callback)(struct sock *sk))
822 struct smc_mark_woken mark = { .woken = false };
823 struct socket_wq *wq;
825 init_waitqueue_func_entry(&mark.wait_entry,
826 smc_fback_mark_woken);
828 wq = rcu_dereference(clcsk->sk_wq);
831 add_wait_queue(sk_sleep(clcsk), &mark.wait_entry);
832 clcsock_callback(clcsk);
833 remove_wait_queue(sk_sleep(clcsk), &mark.wait_entry);
836 smc_fback_wakeup_waitqueue(smc, mark.key);
841 static void smc_fback_state_change(struct sock *clcsk)
843 struct smc_sock *smc;
845 read_lock_bh(&clcsk->sk_callback_lock);
846 smc = smc_clcsock_user_data(clcsk);
848 smc_fback_forward_wakeup(smc, clcsk,
849 smc->clcsk_state_change);
850 read_unlock_bh(&clcsk->sk_callback_lock);
853 static void smc_fback_data_ready(struct sock *clcsk)
855 struct smc_sock *smc;
857 read_lock_bh(&clcsk->sk_callback_lock);
858 smc = smc_clcsock_user_data(clcsk);
860 smc_fback_forward_wakeup(smc, clcsk,
861 smc->clcsk_data_ready);
862 read_unlock_bh(&clcsk->sk_callback_lock);
865 static void smc_fback_write_space(struct sock *clcsk)
867 struct smc_sock *smc;
869 read_lock_bh(&clcsk->sk_callback_lock);
870 smc = smc_clcsock_user_data(clcsk);
872 smc_fback_forward_wakeup(smc, clcsk,
873 smc->clcsk_write_space);
874 read_unlock_bh(&clcsk->sk_callback_lock);
877 static void smc_fback_error_report(struct sock *clcsk)
879 struct smc_sock *smc;
881 read_lock_bh(&clcsk->sk_callback_lock);
882 smc = smc_clcsock_user_data(clcsk);
884 smc_fback_forward_wakeup(smc, clcsk,
885 smc->clcsk_error_report);
886 read_unlock_bh(&clcsk->sk_callback_lock);
889 static void smc_fback_replace_callbacks(struct smc_sock *smc)
891 struct sock *clcsk = smc->clcsock->sk;
893 write_lock_bh(&clcsk->sk_callback_lock);
894 clcsk->sk_user_data = (void *)((uintptr_t)smc | SK_USER_DATA_NOCOPY);
896 smc_clcsock_replace_cb(&clcsk->sk_state_change, smc_fback_state_change,
897 &smc->clcsk_state_change);
898 smc_clcsock_replace_cb(&clcsk->sk_data_ready, smc_fback_data_ready,
899 &smc->clcsk_data_ready);
900 smc_clcsock_replace_cb(&clcsk->sk_write_space, smc_fback_write_space,
901 &smc->clcsk_write_space);
902 smc_clcsock_replace_cb(&clcsk->sk_error_report, smc_fback_error_report,
903 &smc->clcsk_error_report);
905 write_unlock_bh(&clcsk->sk_callback_lock);
908 static int smc_switch_to_fallback(struct smc_sock *smc, int reason_code)
912 mutex_lock(&smc->clcsock_release_lock);
918 smc->use_fallback = true;
919 smc->fallback_rsn = reason_code;
920 smc_stat_fallback(smc);
921 trace_smc_switch_to_fallback(smc, reason_code);
922 if (smc->sk.sk_socket && smc->sk.sk_socket->file) {
923 smc->clcsock->file = smc->sk.sk_socket->file;
924 smc->clcsock->file->private_data = smc->clcsock;
925 smc->clcsock->wq.fasync_list =
926 smc->sk.sk_socket->wq.fasync_list;
927 smc->sk.sk_socket->wq.fasync_list = NULL;
929 /* There might be some wait entries remaining
930 * in smc sk->sk_wq and they should be woken up
931 * as clcsock's wait queue is woken up.
933 smc_fback_replace_callbacks(smc);
936 mutex_unlock(&smc->clcsock_release_lock);
940 /* fall back during connect */
941 static int smc_connect_fallback(struct smc_sock *smc, int reason_code)
943 struct net *net = sock_net(&smc->sk);
946 rc = smc_switch_to_fallback(smc, reason_code);
947 if (rc) { /* fallback fails */
948 this_cpu_inc(net->smc.smc_stats->clnt_hshake_err_cnt);
949 if (smc->sk.sk_state == SMC_INIT)
950 sock_put(&smc->sk); /* passive closing */
953 smc_copy_sock_settings_to_clc(smc);
954 smc->connect_nonblock = 0;
955 if (smc->sk.sk_state == SMC_INIT)
956 smc->sk.sk_state = SMC_ACTIVE;
960 /* decline and fall back during connect */
961 static int smc_connect_decline_fallback(struct smc_sock *smc, int reason_code,
964 struct net *net = sock_net(&smc->sk);
967 if (reason_code < 0) { /* error, fallback is not possible */
968 this_cpu_inc(net->smc.smc_stats->clnt_hshake_err_cnt);
969 if (smc->sk.sk_state == SMC_INIT)
970 sock_put(&smc->sk); /* passive closing */
973 if (reason_code != SMC_CLC_DECL_PEERDECL) {
974 rc = smc_clc_send_decline(smc, reason_code, version);
976 this_cpu_inc(net->smc.smc_stats->clnt_hshake_err_cnt);
977 if (smc->sk.sk_state == SMC_INIT)
978 sock_put(&smc->sk); /* passive closing */
982 return smc_connect_fallback(smc, reason_code);
985 static void smc_conn_abort(struct smc_sock *smc, int local_first)
987 struct smc_connection *conn = &smc->conn;
988 struct smc_link_group *lgr = conn->lgr;
989 bool lgr_valid = false;
991 if (smc_conn_lgr_valid(conn))
995 if (local_first && lgr_valid)
996 smc_lgr_cleanup_early(lgr);
999 /* check if there is a rdma device available for this connection. */
1000 /* called for connect and listen */
1001 static int smc_find_rdma_device(struct smc_sock *smc, struct smc_init_info *ini)
1003 /* PNET table look up: search active ib_device and port
1004 * within same PNETID that also contains the ethernet device
1005 * used for the internal TCP socket
1007 smc_pnet_find_roce_resource(smc->clcsock->sk, ini);
1008 if (!ini->check_smcrv2 && !ini->ib_dev)
1009 return SMC_CLC_DECL_NOSMCRDEV;
1010 if (ini->check_smcrv2 && !ini->smcrv2.ib_dev_v2)
1011 return SMC_CLC_DECL_NOSMCRDEV;
1015 /* check if there is an ISM device available for this connection. */
1016 /* called for connect and listen */
1017 static int smc_find_ism_device(struct smc_sock *smc, struct smc_init_info *ini)
1019 /* Find ISM device with same PNETID as connecting interface */
1020 smc_pnet_find_ism_resource(smc->clcsock->sk, ini);
1021 if (!ini->ism_dev[0])
1022 return SMC_CLC_DECL_NOSMCDDEV;
1024 ini->ism_chid[0] = smc_ism_get_chid(ini->ism_dev[0]);
1028 /* is chid unique for the ism devices that are already determined? */
1029 static bool smc_find_ism_v2_is_unique_chid(u16 chid, struct smc_init_info *ini,
1032 int i = (!ini->ism_dev[0]) ? 1 : 0;
1034 for (; i < cnt; i++)
1035 if (ini->ism_chid[i] == chid)
1040 /* determine possible V2 ISM devices (either without PNETID or with PNETID plus
1041 * PNETID matching net_device)
1043 static int smc_find_ism_v2_device_clnt(struct smc_sock *smc,
1044 struct smc_init_info *ini)
1046 int rc = SMC_CLC_DECL_NOSMCDDEV;
1047 struct smcd_dev *smcd;
1048 int i = 1, entry = 1;
1052 if (smcd_indicated(ini->smc_type_v1))
1053 rc = 0; /* already initialized for V1 */
1054 mutex_lock(&smcd_dev_list.mutex);
1055 list_for_each_entry(smcd, &smcd_dev_list.list, list) {
1056 if (smcd->going_away || smcd == ini->ism_dev[0])
1058 chid = smc_ism_get_chid(smcd);
1059 if (!smc_find_ism_v2_is_unique_chid(chid, ini, i))
1061 is_virtual = __smc_ism_is_virtual(chid);
1062 if (!smc_pnet_is_pnetid_set(smcd->pnetid) ||
1063 smc_pnet_is_ndev_pnetid(sock_net(&smc->sk), smcd->pnetid)) {
1064 if (is_virtual && entry == SMCD_CLC_MAX_V2_GID_ENTRIES)
1065 /* It's the last GID-CHID entry left in CLC
1066 * Proposal SMC-Dv2 extension, but a virtual
1067 * ISM device will take two entries. So give
1068 * up it and try the next potential ISM device.
1071 ini->ism_dev[i] = smcd;
1072 ini->ism_chid[i] = chid;
1073 ini->is_smcd = true;
1076 entry = is_virtual ? entry + 2 : entry + 1;
1077 if (entry > SMCD_CLC_MAX_V2_GID_ENTRIES)
1081 mutex_unlock(&smcd_dev_list.mutex);
1082 ini->ism_offered_cnt = i - 1;
1083 if (!ini->ism_dev[0] && !ini->ism_dev[1])
1084 ini->smcd_version = 0;
1089 /* Check for VLAN ID and register it on ISM device just for CLC handshake */
1090 static int smc_connect_ism_vlan_setup(struct smc_sock *smc,
1091 struct smc_init_info *ini)
1093 if (ini->vlan_id && smc_ism_get_vlan(ini->ism_dev[0], ini->vlan_id))
1094 return SMC_CLC_DECL_ISMVLANERR;
1098 static int smc_find_proposal_devices(struct smc_sock *smc,
1099 struct smc_init_info *ini)
1103 /* check if there is an ism device available */
1104 if (!(ini->smcd_version & SMC_V1) ||
1105 smc_find_ism_device(smc, ini) ||
1106 smc_connect_ism_vlan_setup(smc, ini))
1107 ini->smcd_version &= ~SMC_V1;
1108 /* else ISM V1 is supported for this connection */
1110 /* check if there is an rdma device available */
1111 if (!(ini->smcr_version & SMC_V1) ||
1112 smc_find_rdma_device(smc, ini))
1113 ini->smcr_version &= ~SMC_V1;
1114 /* else RDMA is supported for this connection */
1116 ini->smc_type_v1 = smc_indicated_type(ini->smcd_version & SMC_V1,
1117 ini->smcr_version & SMC_V1);
1119 /* check if there is an ism v2 device available */
1120 if (!(ini->smcd_version & SMC_V2) ||
1121 !smc_ism_is_v2_capable() ||
1122 smc_find_ism_v2_device_clnt(smc, ini))
1123 ini->smcd_version &= ~SMC_V2;
1125 /* check if there is an rdma v2 device available */
1126 ini->check_smcrv2 = true;
1127 ini->smcrv2.saddr = smc->clcsock->sk->sk_rcv_saddr;
1128 if (!(ini->smcr_version & SMC_V2) ||
1129 smc->clcsock->sk->sk_family != AF_INET ||
1130 !smc_clc_ueid_count() ||
1131 smc_find_rdma_device(smc, ini))
1132 ini->smcr_version &= ~SMC_V2;
1133 ini->check_smcrv2 = false;
1135 ini->smc_type_v2 = smc_indicated_type(ini->smcd_version & SMC_V2,
1136 ini->smcr_version & SMC_V2);
1138 /* if neither ISM nor RDMA are supported, fallback */
1139 if (ini->smc_type_v1 == SMC_TYPE_N && ini->smc_type_v2 == SMC_TYPE_N)
1140 rc = SMC_CLC_DECL_NOSMCDEV;
1145 /* cleanup temporary VLAN ID registration used for CLC handshake. If ISM is
1146 * used, the VLAN ID will be registered again during the connection setup.
1148 static int smc_connect_ism_vlan_cleanup(struct smc_sock *smc,
1149 struct smc_init_info *ini)
1151 if (!smcd_indicated(ini->smc_type_v1))
1153 if (ini->vlan_id && smc_ism_put_vlan(ini->ism_dev[0], ini->vlan_id))
1154 return SMC_CLC_DECL_CNFERR;
1158 #define SMC_CLC_MAX_ACCEPT_LEN \
1159 (sizeof(struct smc_clc_msg_accept_confirm) + \
1160 sizeof(struct smc_clc_first_contact_ext_v2x) + \
1161 sizeof(struct smc_clc_msg_trail))
1163 /* CLC handshake during connect */
1164 static int smc_connect_clc(struct smc_sock *smc,
1165 struct smc_clc_msg_accept_confirm *aclc,
1166 struct smc_init_info *ini)
1170 /* do inband token exchange */
1171 rc = smc_clc_send_proposal(smc, ini);
1174 /* receive SMC Accept CLC message */
1175 return smc_clc_wait_msg(smc, aclc, SMC_CLC_MAX_ACCEPT_LEN,
1176 SMC_CLC_ACCEPT, CLC_WAIT_TIME);
1179 void smc_fill_gid_list(struct smc_link_group *lgr,
1180 struct smc_gidlist *gidlist,
1181 struct smc_ib_device *known_dev, u8 *known_gid)
1183 struct smc_init_info *alt_ini = NULL;
1185 memset(gidlist, 0, sizeof(*gidlist));
1186 memcpy(gidlist->list[gidlist->len++], known_gid, SMC_GID_SIZE);
1188 alt_ini = kzalloc(sizeof(*alt_ini), GFP_KERNEL);
1192 alt_ini->vlan_id = lgr->vlan_id;
1193 alt_ini->check_smcrv2 = true;
1194 alt_ini->smcrv2.saddr = lgr->saddr;
1195 smc_pnet_find_alt_roce(lgr, alt_ini, known_dev);
1197 if (!alt_ini->smcrv2.ib_dev_v2)
1200 memcpy(gidlist->list[gidlist->len++], alt_ini->smcrv2.ib_gid_v2,
1207 static int smc_connect_rdma_v2_prepare(struct smc_sock *smc,
1208 struct smc_clc_msg_accept_confirm *aclc,
1209 struct smc_init_info *ini)
1211 struct smc_clc_first_contact_ext *fce =
1212 smc_get_clc_first_contact_ext(aclc, false);
1213 struct net *net = sock_net(&smc->sk);
1216 if (!ini->first_contact_peer || aclc->hdr.version == SMC_V1)
1219 if (fce->v2_direct) {
1220 memcpy(ini->smcrv2.nexthop_mac, &aclc->r0.lcl.mac, ETH_ALEN);
1221 ini->smcrv2.uses_gateway = false;
1223 if (smc_ib_find_route(net, smc->clcsock->sk->sk_rcv_saddr,
1224 smc_ib_gid_to_ipv4(aclc->r0.lcl.gid),
1225 ini->smcrv2.nexthop_mac,
1226 &ini->smcrv2.uses_gateway))
1227 return SMC_CLC_DECL_NOROUTE;
1228 if (!ini->smcrv2.uses_gateway) {
1229 /* mismatch: peer claims indirect, but its direct */
1230 return SMC_CLC_DECL_NOINDIRECT;
1234 ini->release_nr = fce->release;
1235 rc = smc_clc_clnt_v2x_features_validate(fce, ini);
1242 /* setup for RDMA connection of client */
1243 static int smc_connect_rdma(struct smc_sock *smc,
1244 struct smc_clc_msg_accept_confirm *aclc,
1245 struct smc_init_info *ini)
1247 int i, reason_code = 0;
1248 struct smc_link *link;
1251 ini->is_smcd = false;
1252 ini->ib_clcqpn = ntoh24(aclc->r0.qpn);
1253 ini->first_contact_peer = aclc->hdr.typev2 & SMC_FIRST_CONTACT_MASK;
1254 memcpy(ini->peer_systemid, aclc->r0.lcl.id_for_peer, SMC_SYSTEMID_LEN);
1255 memcpy(ini->peer_gid, aclc->r0.lcl.gid, SMC_GID_SIZE);
1256 memcpy(ini->peer_mac, aclc->r0.lcl.mac, ETH_ALEN);
1257 ini->max_conns = SMC_CONN_PER_LGR_MAX;
1258 ini->max_links = SMC_LINKS_ADD_LNK_MAX;
1260 reason_code = smc_connect_rdma_v2_prepare(smc, aclc, ini);
1264 mutex_lock(&smc_client_lgr_pending);
1265 reason_code = smc_conn_create(smc, ini);
1267 mutex_unlock(&smc_client_lgr_pending);
1271 smc_conn_save_peer_info(smc, aclc);
1273 if (ini->first_contact_local) {
1274 link = smc->conn.lnk;
1276 /* set link that was assigned by server */
1278 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1279 struct smc_link *l = &smc->conn.lgr->lnk[i];
1281 if (l->peer_qpn == ntoh24(aclc->r0.qpn) &&
1282 !memcmp(l->peer_gid, &aclc->r0.lcl.gid,
1284 (aclc->hdr.version > SMC_V1 ||
1285 !memcmp(l->peer_mac, &aclc->r0.lcl.mac,
1286 sizeof(l->peer_mac)))) {
1292 reason_code = SMC_CLC_DECL_NOSRVLINK;
1295 smc_switch_link_and_count(&smc->conn, link);
1298 /* create send buffer and rmb */
1299 if (smc_buf_create(smc, false)) {
1300 reason_code = SMC_CLC_DECL_MEM;
1304 if (ini->first_contact_local)
1305 smc_link_save_peer_info(link, aclc, ini);
1307 if (smc_rmb_rtoken_handling(&smc->conn, link, aclc)) {
1308 reason_code = SMC_CLC_DECL_ERR_RTOK;
1312 smc_close_init(smc);
1315 if (ini->first_contact_local) {
1316 if (smc_ib_ready_link(link)) {
1317 reason_code = SMC_CLC_DECL_ERR_RDYLNK;
1321 /* reg sendbufs if they were vzalloced */
1322 if (smc->conn.sndbuf_desc->is_vm) {
1323 if (smcr_lgr_reg_sndbufs(link, smc->conn.sndbuf_desc)) {
1324 reason_code = SMC_CLC_DECL_ERR_REGBUF;
1328 if (smcr_lgr_reg_rmbs(link, smc->conn.rmb_desc)) {
1329 reason_code = SMC_CLC_DECL_ERR_REGBUF;
1334 if (aclc->hdr.version > SMC_V1) {
1336 if (ini->first_contact_local)
1337 smc_fill_gid_list(link->lgr, &ini->smcrv2.gidlist,
1338 link->smcibdev, link->gid);
1341 reason_code = smc_clc_send_confirm(smc, ini->first_contact_local,
1342 aclc->hdr.version, eid, ini);
1348 if (ini->first_contact_local) {
1349 /* QP confirmation over RoCE fabric */
1350 smc_llc_flow_initiate(link->lgr, SMC_LLC_FLOW_ADD_LINK);
1351 reason_code = smcr_clnt_conf_first_link(smc);
1352 smc_llc_flow_stop(link->lgr, &link->lgr->llc_flow_lcl);
1356 mutex_unlock(&smc_client_lgr_pending);
1358 smc_copy_sock_settings_to_clc(smc);
1359 smc->connect_nonblock = 0;
1360 if (smc->sk.sk_state == SMC_INIT)
1361 smc->sk.sk_state = SMC_ACTIVE;
1365 smc_conn_abort(smc, ini->first_contact_local);
1366 mutex_unlock(&smc_client_lgr_pending);
1367 smc->connect_nonblock = 0;
1372 /* The server has chosen one of the proposed ISM devices for the communication.
1373 * Determine from the CHID of the received CLC ACCEPT the ISM device chosen.
1376 smc_v2_determine_accepted_chid(struct smc_clc_msg_accept_confirm *aclc,
1377 struct smc_init_info *ini)
1381 for (i = 0; i < ini->ism_offered_cnt + 1; i++) {
1382 if (ini->ism_chid[i] == ntohs(aclc->d1.chid)) {
1383 ini->ism_selected = i;
1391 /* setup for ISM connection of client */
1392 static int smc_connect_ism(struct smc_sock *smc,
1393 struct smc_clc_msg_accept_confirm *aclc,
1394 struct smc_init_info *ini)
1399 ini->is_smcd = true;
1400 ini->first_contact_peer = aclc->hdr.typev2 & SMC_FIRST_CONTACT_MASK;
1402 if (aclc->hdr.version == SMC_V2) {
1403 if (ini->first_contact_peer) {
1404 struct smc_clc_first_contact_ext *fce =
1405 smc_get_clc_first_contact_ext(aclc, true);
1407 ini->release_nr = fce->release;
1408 rc = smc_clc_clnt_v2x_features_validate(fce, ini);
1413 rc = smc_v2_determine_accepted_chid(aclc, ini);
1417 if (__smc_ism_is_virtual(ini->ism_chid[ini->ism_selected]))
1418 ini->ism_peer_gid[ini->ism_selected].gid_ext =
1419 ntohll(aclc->d1.gid_ext);
1420 /* for non-virtual ISM devices, peer gid_ext remains 0. */
1422 ini->ism_peer_gid[ini->ism_selected].gid = ntohll(aclc->d0.gid);
1424 /* there is only one lgr role for SMC-D; use server lock */
1425 mutex_lock(&smc_server_lgr_pending);
1426 rc = smc_conn_create(smc, ini);
1428 mutex_unlock(&smc_server_lgr_pending);
1432 /* Create send and receive buffers */
1433 rc = smc_buf_create(smc, true);
1435 rc = (rc == -ENOSPC) ? SMC_CLC_DECL_MAX_DMB : SMC_CLC_DECL_MEM;
1439 smc_conn_save_peer_info(smc, aclc);
1440 smc_close_init(smc);
1444 if (aclc->hdr.version > SMC_V1)
1447 rc = smc_clc_send_confirm(smc, ini->first_contact_local,
1448 aclc->hdr.version, eid, ini);
1451 mutex_unlock(&smc_server_lgr_pending);
1453 smc_copy_sock_settings_to_clc(smc);
1454 smc->connect_nonblock = 0;
1455 if (smc->sk.sk_state == SMC_INIT)
1456 smc->sk.sk_state = SMC_ACTIVE;
1460 smc_conn_abort(smc, ini->first_contact_local);
1461 mutex_unlock(&smc_server_lgr_pending);
1462 smc->connect_nonblock = 0;
1467 /* check if received accept type and version matches a proposed one */
1468 static int smc_connect_check_aclc(struct smc_init_info *ini,
1469 struct smc_clc_msg_accept_confirm *aclc)
1471 if (aclc->hdr.typev1 != SMC_TYPE_R &&
1472 aclc->hdr.typev1 != SMC_TYPE_D)
1473 return SMC_CLC_DECL_MODEUNSUPP;
1475 if (aclc->hdr.version >= SMC_V2) {
1476 if ((aclc->hdr.typev1 == SMC_TYPE_R &&
1477 !smcr_indicated(ini->smc_type_v2)) ||
1478 (aclc->hdr.typev1 == SMC_TYPE_D &&
1479 !smcd_indicated(ini->smc_type_v2)))
1480 return SMC_CLC_DECL_MODEUNSUPP;
1482 if ((aclc->hdr.typev1 == SMC_TYPE_R &&
1483 !smcr_indicated(ini->smc_type_v1)) ||
1484 (aclc->hdr.typev1 == SMC_TYPE_D &&
1485 !smcd_indicated(ini->smc_type_v1)))
1486 return SMC_CLC_DECL_MODEUNSUPP;
1492 /* perform steps before actually connecting */
1493 static int __smc_connect(struct smc_sock *smc)
1495 u8 version = smc_ism_is_v2_capable() ? SMC_V2 : SMC_V1;
1496 struct smc_clc_msg_accept_confirm *aclc;
1497 struct smc_init_info *ini = NULL;
1501 if (smc->use_fallback)
1502 return smc_connect_fallback(smc, smc->fallback_rsn);
1504 /* if peer has not signalled SMC-capability, fall back */
1505 if (!tcp_sk(smc->clcsock->sk)->syn_smc)
1506 return smc_connect_fallback(smc, SMC_CLC_DECL_PEERNOSMC);
1508 /* IPSec connections opt out of SMC optimizations */
1509 if (using_ipsec(smc))
1510 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_IPSEC,
1513 ini = kzalloc(sizeof(*ini), GFP_KERNEL);
1515 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_MEM,
1518 ini->smcd_version = SMC_V1 | SMC_V2;
1519 ini->smcr_version = SMC_V1 | SMC_V2;
1520 ini->smc_type_v1 = SMC_TYPE_B;
1521 ini->smc_type_v2 = SMC_TYPE_B;
1523 /* get vlan id from IP device */
1524 if (smc_vlan_by_tcpsk(smc->clcsock, ini)) {
1525 ini->smcd_version &= ~SMC_V1;
1526 ini->smcr_version = 0;
1527 ini->smc_type_v1 = SMC_TYPE_N;
1528 if (!ini->smcd_version) {
1529 rc = SMC_CLC_DECL_GETVLANERR;
1534 rc = smc_find_proposal_devices(smc, ini);
1538 buf = kzalloc(SMC_CLC_MAX_ACCEPT_LEN, GFP_KERNEL);
1540 rc = SMC_CLC_DECL_MEM;
1543 aclc = (struct smc_clc_msg_accept_confirm *)buf;
1545 /* perform CLC handshake */
1546 rc = smc_connect_clc(smc, aclc, ini);
1548 /* -EAGAIN on timeout, see tcp_recvmsg() */
1549 if (rc == -EAGAIN) {
1551 smc->sk.sk_err = ETIMEDOUT;
1556 /* check if smc modes and versions of CLC proposal and accept match */
1557 rc = smc_connect_check_aclc(ini, aclc);
1558 version = aclc->hdr.version == SMC_V1 ? SMC_V1 : SMC_V2;
1562 /* depending on previous steps, connect using rdma or ism */
1563 if (aclc->hdr.typev1 == SMC_TYPE_R) {
1564 ini->smcr_version = version;
1565 rc = smc_connect_rdma(smc, aclc, ini);
1566 } else if (aclc->hdr.typev1 == SMC_TYPE_D) {
1567 ini->smcd_version = version;
1568 rc = smc_connect_ism(smc, aclc, ini);
1573 SMC_STAT_CLNT_SUCC_INC(sock_net(smc->clcsock->sk), aclc);
1574 smc_connect_ism_vlan_cleanup(smc, ini);
1580 smc_connect_ism_vlan_cleanup(smc, ini);
1584 return smc_connect_decline_fallback(smc, rc, version);
1587 static void smc_connect_work(struct work_struct *work)
1589 struct smc_sock *smc = container_of(work, struct smc_sock,
1591 long timeo = smc->sk.sk_sndtimeo;
1595 timeo = MAX_SCHEDULE_TIMEOUT;
1596 lock_sock(smc->clcsock->sk);
1597 if (smc->clcsock->sk->sk_err) {
1598 smc->sk.sk_err = smc->clcsock->sk->sk_err;
1599 } else if ((1 << smc->clcsock->sk->sk_state) &
1600 (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
1601 rc = sk_stream_wait_connect(smc->clcsock->sk, &timeo);
1602 if ((rc == -EPIPE) &&
1603 ((1 << smc->clcsock->sk->sk_state) &
1604 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)))
1607 release_sock(smc->clcsock->sk);
1608 lock_sock(&smc->sk);
1609 if (rc != 0 || smc->sk.sk_err) {
1610 smc->sk.sk_state = SMC_CLOSED;
1611 if (rc == -EPIPE || rc == -EAGAIN)
1612 smc->sk.sk_err = EPIPE;
1613 else if (rc == -ECONNREFUSED)
1614 smc->sk.sk_err = ECONNREFUSED;
1615 else if (signal_pending(current))
1616 smc->sk.sk_err = -sock_intr_errno(timeo);
1617 sock_put(&smc->sk); /* passive closing */
1621 rc = __smc_connect(smc);
1623 smc->sk.sk_err = -rc;
1626 if (!sock_flag(&smc->sk, SOCK_DEAD)) {
1627 if (smc->sk.sk_err) {
1628 smc->sk.sk_state_change(&smc->sk);
1629 } else { /* allow polling before and after fallback decision */
1630 smc->clcsock->sk->sk_write_space(smc->clcsock->sk);
1631 smc->sk.sk_write_space(&smc->sk);
1634 release_sock(&smc->sk);
1637 static int smc_connect(struct socket *sock, struct sockaddr *addr,
1638 int alen, int flags)
1640 struct sock *sk = sock->sk;
1641 struct smc_sock *smc;
1646 /* separate smc parameter checking to be safe */
1647 if (alen < sizeof(addr->sa_family))
1649 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
1653 switch (sock->state) {
1658 rc = sk->sk_state == SMC_ACTIVE ? -EISCONN : -EINVAL;
1661 if (sk->sk_state == SMC_ACTIVE)
1664 case SS_UNCONNECTED:
1665 sock->state = SS_CONNECTING;
1669 switch (sk->sk_state) {
1673 rc = sock_error(sk) ? : -ECONNABORTED;
1674 sock->state = SS_UNCONNECTED;
1683 smc_copy_sock_settings_to_clc(smc);
1684 tcp_sk(smc->clcsock->sk)->syn_smc = 1;
1685 if (smc->connect_nonblock) {
1689 rc = kernel_connect(smc->clcsock, addr, alen, flags);
1690 if (rc && rc != -EINPROGRESS)
1693 if (smc->use_fallback) {
1694 sock->state = rc ? SS_CONNECTING : SS_CONNECTED;
1697 sock_hold(&smc->sk); /* sock put in passive closing */
1698 if (flags & O_NONBLOCK) {
1699 if (queue_work(smc_hs_wq, &smc->connect_work))
1700 smc->connect_nonblock = 1;
1704 rc = __smc_connect(smc);
1711 sock->state = SS_CONNECTED;
1718 static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
1720 struct socket *new_clcsock = NULL;
1721 struct sock *lsk = &lsmc->sk;
1722 struct sock *new_sk;
1726 new_sk = smc_sock_alloc(sock_net(lsk), NULL, lsk->sk_protocol);
1729 lsk->sk_err = ENOMEM;
1734 *new_smc = smc_sk(new_sk);
1736 mutex_lock(&lsmc->clcsock_release_lock);
1738 rc = kernel_accept(lsmc->clcsock, &new_clcsock, SOCK_NONBLOCK);
1739 mutex_unlock(&lsmc->clcsock_release_lock);
1741 if (rc < 0 && rc != -EAGAIN)
1743 if (rc < 0 || lsk->sk_state == SMC_CLOSED) {
1744 new_sk->sk_prot->unhash(new_sk);
1746 sock_release(new_clcsock);
1747 new_sk->sk_state = SMC_CLOSED;
1748 smc_sock_set_flag(new_sk, SOCK_DEAD);
1749 sock_put(new_sk); /* final */
1754 /* new clcsock has inherited the smc listen-specific sk_data_ready
1755 * function; switch it back to the original sk_data_ready function
1757 new_clcsock->sk->sk_data_ready = lsmc->clcsk_data_ready;
1759 /* if new clcsock has also inherited the fallback-specific callback
1760 * functions, switch them back to the original ones.
1762 if (lsmc->use_fallback) {
1763 if (lsmc->clcsk_state_change)
1764 new_clcsock->sk->sk_state_change = lsmc->clcsk_state_change;
1765 if (lsmc->clcsk_write_space)
1766 new_clcsock->sk->sk_write_space = lsmc->clcsk_write_space;
1767 if (lsmc->clcsk_error_report)
1768 new_clcsock->sk->sk_error_report = lsmc->clcsk_error_report;
1771 (*new_smc)->clcsock = new_clcsock;
1776 /* add a just created sock to the accept queue of the listen sock as
1777 * candidate for a following socket accept call from user space
1779 static void smc_accept_enqueue(struct sock *parent, struct sock *sk)
1781 struct smc_sock *par = smc_sk(parent);
1783 sock_hold(sk); /* sock_put in smc_accept_unlink () */
1784 spin_lock(&par->accept_q_lock);
1785 list_add_tail(&smc_sk(sk)->accept_q, &par->accept_q);
1786 spin_unlock(&par->accept_q_lock);
1787 sk_acceptq_added(parent);
1790 /* remove a socket from the accept queue of its parental listening socket */
1791 static void smc_accept_unlink(struct sock *sk)
1793 struct smc_sock *par = smc_sk(sk)->listen_smc;
1795 spin_lock(&par->accept_q_lock);
1796 list_del_init(&smc_sk(sk)->accept_q);
1797 spin_unlock(&par->accept_q_lock);
1798 sk_acceptq_removed(&smc_sk(sk)->listen_smc->sk);
1799 sock_put(sk); /* sock_hold in smc_accept_enqueue */
1802 /* remove a sock from the accept queue to bind it to a new socket created
1803 * for a socket accept call from user space
1805 struct sock *smc_accept_dequeue(struct sock *parent,
1806 struct socket *new_sock)
1808 struct smc_sock *isk, *n;
1809 struct sock *new_sk;
1811 list_for_each_entry_safe(isk, n, &smc_sk(parent)->accept_q, accept_q) {
1812 new_sk = (struct sock *)isk;
1814 smc_accept_unlink(new_sk);
1815 if (new_sk->sk_state == SMC_CLOSED) {
1816 new_sk->sk_prot->unhash(new_sk);
1818 sock_release(isk->clcsock);
1819 isk->clcsock = NULL;
1821 sock_put(new_sk); /* final */
1825 sock_graft(new_sk, new_sock);
1826 new_sock->state = SS_CONNECTED;
1827 if (isk->use_fallback) {
1828 smc_sk(new_sk)->clcsock->file = new_sock->file;
1829 isk->clcsock->file->private_data = isk->clcsock;
1837 /* clean up for a created but never accepted sock */
1838 void smc_close_non_accepted(struct sock *sk)
1840 struct smc_sock *smc = smc_sk(sk);
1842 sock_hold(sk); /* sock_put below */
1844 if (!sk->sk_lingertime)
1845 /* wait for peer closing */
1846 WRITE_ONCE(sk->sk_lingertime, SMC_MAX_STREAM_WAIT_TIMEOUT);
1849 sock_put(sk); /* sock_hold above */
1850 sock_put(sk); /* final sock_put */
1853 static int smcr_serv_conf_first_link(struct smc_sock *smc)
1855 struct smc_link *link = smc->conn.lnk;
1856 struct smc_llc_qentry *qentry;
1859 /* reg the sndbuf if it was vzalloced*/
1860 if (smc->conn.sndbuf_desc->is_vm) {
1861 if (smcr_link_reg_buf(link, smc->conn.sndbuf_desc))
1862 return SMC_CLC_DECL_ERR_REGBUF;
1866 if (smcr_link_reg_buf(link, smc->conn.rmb_desc))
1867 return SMC_CLC_DECL_ERR_REGBUF;
1869 /* send CONFIRM LINK request to client over the RoCE fabric */
1870 rc = smc_llc_send_confirm_link(link, SMC_LLC_REQ);
1872 return SMC_CLC_DECL_TIMEOUT_CL;
1874 /* receive CONFIRM LINK response from client over the RoCE fabric */
1875 qentry = smc_llc_wait(link->lgr, link, SMC_LLC_WAIT_TIME,
1876 SMC_LLC_CONFIRM_LINK);
1878 struct smc_clc_msg_decline dclc;
1880 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
1881 SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
1882 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
1884 smc_llc_save_peer_uid(qentry);
1885 rc = smc_llc_eval_conf_link(qentry, SMC_LLC_RESP);
1886 smc_llc_flow_qentry_del(&link->lgr->llc_flow_lcl);
1888 return SMC_CLC_DECL_RMBE_EC;
1890 /* confirm_rkey is implicit on 1st contact */
1891 smc->conn.rmb_desc->is_conf_rkey = true;
1893 smc_llc_link_active(link);
1894 smcr_lgr_set_type(link->lgr, SMC_LGR_SINGLE);
1896 if (link->lgr->max_links > 1) {
1897 down_write(&link->lgr->llc_conf_mutex);
1898 /* initial contact - try to establish second link */
1899 smc_llc_srv_add_link(link, NULL);
1900 up_write(&link->lgr->llc_conf_mutex);
1905 /* listen worker: finish */
1906 static void smc_listen_out(struct smc_sock *new_smc)
1908 struct smc_sock *lsmc = new_smc->listen_smc;
1909 struct sock *newsmcsk = &new_smc->sk;
1911 if (tcp_sk(new_smc->clcsock->sk)->syn_smc)
1912 atomic_dec(&lsmc->queued_smc_hs);
1914 if (lsmc->sk.sk_state == SMC_LISTEN) {
1915 lock_sock_nested(&lsmc->sk, SINGLE_DEPTH_NESTING);
1916 smc_accept_enqueue(&lsmc->sk, newsmcsk);
1917 release_sock(&lsmc->sk);
1918 } else { /* no longer listening */
1919 smc_close_non_accepted(newsmcsk);
1922 /* Wake up accept */
1923 lsmc->sk.sk_data_ready(&lsmc->sk);
1924 sock_put(&lsmc->sk); /* sock_hold in smc_tcp_listen_work */
1927 /* listen worker: finish in state connected */
1928 static void smc_listen_out_connected(struct smc_sock *new_smc)
1930 struct sock *newsmcsk = &new_smc->sk;
1932 if (newsmcsk->sk_state == SMC_INIT)
1933 newsmcsk->sk_state = SMC_ACTIVE;
1935 smc_listen_out(new_smc);
1938 /* listen worker: finish in error state */
1939 static void smc_listen_out_err(struct smc_sock *new_smc)
1941 struct sock *newsmcsk = &new_smc->sk;
1942 struct net *net = sock_net(newsmcsk);
1944 this_cpu_inc(net->smc.smc_stats->srv_hshake_err_cnt);
1945 if (newsmcsk->sk_state == SMC_INIT)
1946 sock_put(&new_smc->sk); /* passive closing */
1947 newsmcsk->sk_state = SMC_CLOSED;
1949 smc_listen_out(new_smc);
1952 /* listen worker: decline and fall back if possible */
1953 static void smc_listen_decline(struct smc_sock *new_smc, int reason_code,
1954 int local_first, u8 version)
1956 /* RDMA setup failed, switch back to TCP */
1957 smc_conn_abort(new_smc, local_first);
1958 if (reason_code < 0 ||
1959 smc_switch_to_fallback(new_smc, reason_code)) {
1960 /* error, no fallback possible */
1961 smc_listen_out_err(new_smc);
1964 if (reason_code && reason_code != SMC_CLC_DECL_PEERDECL) {
1965 if (smc_clc_send_decline(new_smc, reason_code, version) < 0) {
1966 smc_listen_out_err(new_smc);
1970 smc_listen_out_connected(new_smc);
1973 /* listen worker: version checking */
1974 static int smc_listen_v2_check(struct smc_sock *new_smc,
1975 struct smc_clc_msg_proposal *pclc,
1976 struct smc_init_info *ini)
1978 struct smc_clc_smcd_v2_extension *pclc_smcd_v2_ext;
1979 struct smc_clc_v2_extension *pclc_v2_ext;
1980 int rc = SMC_CLC_DECL_PEERNOSMC;
1982 ini->smc_type_v1 = pclc->hdr.typev1;
1983 ini->smc_type_v2 = pclc->hdr.typev2;
1984 ini->smcd_version = smcd_indicated(ini->smc_type_v1) ? SMC_V1 : 0;
1985 ini->smcr_version = smcr_indicated(ini->smc_type_v1) ? SMC_V1 : 0;
1986 if (pclc->hdr.version > SMC_V1) {
1987 if (smcd_indicated(ini->smc_type_v2))
1988 ini->smcd_version |= SMC_V2;
1989 if (smcr_indicated(ini->smc_type_v2))
1990 ini->smcr_version |= SMC_V2;
1992 if (!(ini->smcd_version & SMC_V2) && !(ini->smcr_version & SMC_V2)) {
1993 rc = SMC_CLC_DECL_PEERNOSMC;
1996 pclc_v2_ext = smc_get_clc_v2_ext(pclc);
1998 ini->smcd_version &= ~SMC_V2;
1999 ini->smcr_version &= ~SMC_V2;
2000 rc = SMC_CLC_DECL_NOV2EXT;
2003 pclc_smcd_v2_ext = smc_get_clc_smcd_v2_ext(pclc_v2_ext);
2004 if (ini->smcd_version & SMC_V2) {
2005 if (!smc_ism_is_v2_capable()) {
2006 ini->smcd_version &= ~SMC_V2;
2007 rc = SMC_CLC_DECL_NOISM2SUPP;
2008 } else if (!pclc_smcd_v2_ext) {
2009 ini->smcd_version &= ~SMC_V2;
2010 rc = SMC_CLC_DECL_NOV2DEXT;
2011 } else if (!pclc_v2_ext->hdr.eid_cnt &&
2012 !pclc_v2_ext->hdr.flag.seid) {
2013 ini->smcd_version &= ~SMC_V2;
2014 rc = SMC_CLC_DECL_NOUEID;
2017 if (ini->smcr_version & SMC_V2) {
2018 if (!pclc_v2_ext->hdr.eid_cnt) {
2019 ini->smcr_version &= ~SMC_V2;
2020 rc = SMC_CLC_DECL_NOUEID;
2024 ini->release_nr = pclc_v2_ext->hdr.flag.release;
2025 if (pclc_v2_ext->hdr.flag.release > SMC_RELEASE)
2026 ini->release_nr = SMC_RELEASE;
2029 if (!ini->smcd_version && !ini->smcr_version)
2035 /* listen worker: check prefixes */
2036 static int smc_listen_prfx_check(struct smc_sock *new_smc,
2037 struct smc_clc_msg_proposal *pclc)
2039 struct smc_clc_msg_proposal_prefix *pclc_prfx;
2040 struct socket *newclcsock = new_smc->clcsock;
2042 if (pclc->hdr.typev1 == SMC_TYPE_N)
2044 pclc_prfx = smc_clc_proposal_get_prefix(pclc);
2045 if (smc_clc_prfx_match(newclcsock, pclc_prfx))
2046 return SMC_CLC_DECL_DIFFPREFIX;
2051 /* listen worker: initialize connection and buffers */
2052 static int smc_listen_rdma_init(struct smc_sock *new_smc,
2053 struct smc_init_info *ini)
2057 /* allocate connection / link group */
2058 rc = smc_conn_create(new_smc, ini);
2062 /* create send buffer and rmb */
2063 if (smc_buf_create(new_smc, false)) {
2064 smc_conn_abort(new_smc, ini->first_contact_local);
2065 return SMC_CLC_DECL_MEM;
2071 /* listen worker: initialize connection and buffers for SMC-D */
2072 static int smc_listen_ism_init(struct smc_sock *new_smc,
2073 struct smc_init_info *ini)
2077 rc = smc_conn_create(new_smc, ini);
2081 /* Create send and receive buffers */
2082 rc = smc_buf_create(new_smc, true);
2084 smc_conn_abort(new_smc, ini->first_contact_local);
2085 return (rc == -ENOSPC) ? SMC_CLC_DECL_MAX_DMB :
2092 static bool smc_is_already_selected(struct smcd_dev *smcd,
2093 struct smc_init_info *ini,
2098 for (i = 0; i < matches; i++)
2099 if (smcd == ini->ism_dev[i])
2105 /* check for ISM devices matching proposed ISM devices */
2106 static void smc_check_ism_v2_match(struct smc_init_info *ini,
2108 struct smcd_gid *proposed_gid,
2109 unsigned int *matches)
2111 struct smcd_dev *smcd;
2113 list_for_each_entry(smcd, &smcd_dev_list.list, list) {
2114 if (smcd->going_away)
2116 if (smc_is_already_selected(smcd, ini, *matches))
2118 if (smc_ism_get_chid(smcd) == proposed_chid &&
2119 !smc_ism_cantalk(proposed_gid, ISM_RESERVED_VLANID, smcd)) {
2120 ini->ism_peer_gid[*matches].gid = proposed_gid->gid;
2121 if (__smc_ism_is_virtual(proposed_chid))
2122 ini->ism_peer_gid[*matches].gid_ext =
2123 proposed_gid->gid_ext;
2124 /* non-virtual ISM's peer gid_ext remains 0. */
2125 ini->ism_dev[*matches] = smcd;
2132 static void smc_find_ism_store_rc(u32 rc, struct smc_init_info *ini)
2138 static void smc_find_ism_v2_device_serv(struct smc_sock *new_smc,
2139 struct smc_clc_msg_proposal *pclc,
2140 struct smc_init_info *ini)
2142 struct smc_clc_smcd_v2_extension *smcd_v2_ext;
2143 struct smc_clc_v2_extension *smc_v2_ext;
2144 struct smc_clc_msg_smcd *pclc_smcd;
2145 unsigned int matches = 0;
2146 struct smcd_gid smcd_gid;
2152 if (!(ini->smcd_version & SMC_V2) || !smcd_indicated(ini->smc_type_v2))
2155 pclc_smcd = smc_get_clc_msg_smcd(pclc);
2156 smc_v2_ext = smc_get_clc_v2_ext(pclc);
2157 smcd_v2_ext = smc_get_clc_smcd_v2_ext(smc_v2_ext);
2159 mutex_lock(&smcd_dev_list.mutex);
2160 if (pclc_smcd->ism.chid) {
2161 /* check for ISM device matching proposed native ISM device */
2162 smcd_gid.gid = ntohll(pclc_smcd->ism.gid);
2163 smcd_gid.gid_ext = 0;
2164 smc_check_ism_v2_match(ini, ntohs(pclc_smcd->ism.chid),
2165 &smcd_gid, &matches);
2167 for (i = 0; i < smc_v2_ext->hdr.ism_gid_cnt; i++) {
2168 /* check for ISM devices matching proposed non-native ISM
2171 smcd_gid.gid = ntohll(smcd_v2_ext->gidchid[i].gid);
2172 smcd_gid.gid_ext = 0;
2173 chid = ntohs(smcd_v2_ext->gidchid[i].chid);
2174 if (__smc_ism_is_virtual(chid)) {
2175 if ((i + 1) == smc_v2_ext->hdr.ism_gid_cnt ||
2176 chid != ntohs(smcd_v2_ext->gidchid[i + 1].chid))
2177 /* each virtual ISM device takes two GID-CHID
2178 * entries and CHID of the second entry repeats
2179 * that of the first entry.
2181 * So check if the next GID-CHID entry exists
2182 * and both two entries' CHIDs are the same.
2186 ntohll(smcd_v2_ext->gidchid[++i].gid);
2188 smc_check_ism_v2_match(ini, chid, &smcd_gid, &matches);
2190 mutex_unlock(&smcd_dev_list.mutex);
2192 if (!ini->ism_dev[0]) {
2193 smc_find_ism_store_rc(SMC_CLC_DECL_NOSMCD2DEV, ini);
2197 smc_ism_get_system_eid(&eid);
2198 if (!smc_clc_match_eid(ini->negotiated_eid, smc_v2_ext,
2199 smcd_v2_ext->system_eid, eid))
2202 /* separate - outside the smcd_dev_list.lock */
2203 smcd_version = ini->smcd_version;
2204 for (i = 0; i < matches; i++) {
2205 ini->smcd_version = SMC_V2;
2206 ini->is_smcd = true;
2207 ini->ism_selected = i;
2208 rc = smc_listen_ism_init(new_smc, ini);
2210 smc_find_ism_store_rc(rc, ini);
2211 /* try next active ISM device */
2214 return; /* matching and usable V2 ISM device found */
2216 /* no V2 ISM device could be initialized */
2217 ini->smcd_version = smcd_version; /* restore original value */
2218 ini->negotiated_eid[0] = 0;
2221 ini->smcd_version &= ~SMC_V2;
2222 ini->ism_dev[0] = NULL;
2223 ini->is_smcd = false;
2226 static void smc_find_ism_v1_device_serv(struct smc_sock *new_smc,
2227 struct smc_clc_msg_proposal *pclc,
2228 struct smc_init_info *ini)
2230 struct smc_clc_msg_smcd *pclc_smcd = smc_get_clc_msg_smcd(pclc);
2233 /* check if ISM V1 is available */
2234 if (!(ini->smcd_version & SMC_V1) || !smcd_indicated(ini->smc_type_v1))
2236 ini->is_smcd = true; /* prepare ISM check */
2237 ini->ism_peer_gid[0].gid = ntohll(pclc_smcd->ism.gid);
2238 ini->ism_peer_gid[0].gid_ext = 0;
2239 rc = smc_find_ism_device(new_smc, ini);
2242 ini->ism_selected = 0;
2243 rc = smc_listen_ism_init(new_smc, ini);
2245 return; /* V1 ISM device found */
2248 smc_find_ism_store_rc(rc, ini);
2249 ini->smcd_version &= ~SMC_V1;
2250 ini->ism_dev[0] = NULL;
2251 ini->is_smcd = false;
2254 /* listen worker: register buffers */
2255 static int smc_listen_rdma_reg(struct smc_sock *new_smc, bool local_first)
2257 struct smc_connection *conn = &new_smc->conn;
2260 /* reg sendbufs if they were vzalloced */
2261 if (conn->sndbuf_desc->is_vm) {
2262 if (smcr_lgr_reg_sndbufs(conn->lnk,
2264 return SMC_CLC_DECL_ERR_REGBUF;
2266 if (smcr_lgr_reg_rmbs(conn->lnk, conn->rmb_desc))
2267 return SMC_CLC_DECL_ERR_REGBUF;
2273 static void smc_find_rdma_v2_device_serv(struct smc_sock *new_smc,
2274 struct smc_clc_msg_proposal *pclc,
2275 struct smc_init_info *ini)
2277 struct smc_clc_v2_extension *smc_v2_ext;
2281 if (!(ini->smcr_version & SMC_V2) || !smcr_indicated(ini->smc_type_v2))
2284 smc_v2_ext = smc_get_clc_v2_ext(pclc);
2285 if (!smc_clc_match_eid(ini->negotiated_eid, smc_v2_ext, NULL, NULL))
2288 /* prepare RDMA check */
2289 memcpy(ini->peer_systemid, pclc->lcl.id_for_peer, SMC_SYSTEMID_LEN);
2290 memcpy(ini->peer_gid, smc_v2_ext->roce, SMC_GID_SIZE);
2291 memcpy(ini->peer_mac, pclc->lcl.mac, ETH_ALEN);
2292 ini->check_smcrv2 = true;
2293 ini->smcrv2.clc_sk = new_smc->clcsock->sk;
2294 ini->smcrv2.saddr = new_smc->clcsock->sk->sk_rcv_saddr;
2295 ini->smcrv2.daddr = smc_ib_gid_to_ipv4(smc_v2_ext->roce);
2296 rc = smc_find_rdma_device(new_smc, ini);
2298 smc_find_ism_store_rc(rc, ini);
2301 if (!ini->smcrv2.uses_gateway)
2302 memcpy(ini->smcrv2.nexthop_mac, pclc->lcl.mac, ETH_ALEN);
2304 smcr_version = ini->smcr_version;
2305 ini->smcr_version = SMC_V2;
2306 rc = smc_listen_rdma_init(new_smc, ini);
2308 rc = smc_listen_rdma_reg(new_smc, ini->first_contact_local);
2310 smc_conn_abort(new_smc, ini->first_contact_local);
2314 ini->smcr_version = smcr_version;
2315 smc_find_ism_store_rc(rc, ini);
2318 ini->smcr_version &= ~SMC_V2;
2319 ini->smcrv2.ib_dev_v2 = NULL;
2320 ini->check_smcrv2 = false;
2323 static int smc_find_rdma_v1_device_serv(struct smc_sock *new_smc,
2324 struct smc_clc_msg_proposal *pclc,
2325 struct smc_init_info *ini)
2329 if (!(ini->smcr_version & SMC_V1) || !smcr_indicated(ini->smc_type_v1))
2330 return SMC_CLC_DECL_NOSMCDEV;
2332 /* prepare RDMA check */
2333 memcpy(ini->peer_systemid, pclc->lcl.id_for_peer, SMC_SYSTEMID_LEN);
2334 memcpy(ini->peer_gid, pclc->lcl.gid, SMC_GID_SIZE);
2335 memcpy(ini->peer_mac, pclc->lcl.mac, ETH_ALEN);
2336 rc = smc_find_rdma_device(new_smc, ini);
2338 /* no RDMA device found */
2339 return SMC_CLC_DECL_NOSMCDEV;
2341 rc = smc_listen_rdma_init(new_smc, ini);
2344 return smc_listen_rdma_reg(new_smc, ini->first_contact_local);
2347 /* determine the local device matching to proposal */
2348 static int smc_listen_find_device(struct smc_sock *new_smc,
2349 struct smc_clc_msg_proposal *pclc,
2350 struct smc_init_info *ini)
2354 /* check for ISM device matching V2 proposed device */
2355 smc_find_ism_v2_device_serv(new_smc, pclc, ini);
2356 if (ini->ism_dev[0])
2359 /* check for matching IP prefix and subnet length (V1) */
2360 prfx_rc = smc_listen_prfx_check(new_smc, pclc);
2362 smc_find_ism_store_rc(prfx_rc, ini);
2364 /* get vlan id from IP device */
2365 if (smc_vlan_by_tcpsk(new_smc->clcsock, ini))
2366 return ini->rc ?: SMC_CLC_DECL_GETVLANERR;
2368 /* check for ISM device matching V1 proposed device */
2370 smc_find_ism_v1_device_serv(new_smc, pclc, ini);
2371 if (ini->ism_dev[0])
2374 if (!smcr_indicated(pclc->hdr.typev1) &&
2375 !smcr_indicated(pclc->hdr.typev2))
2376 /* skip RDMA and decline */
2377 return ini->rc ?: SMC_CLC_DECL_NOSMCDDEV;
2379 /* check if RDMA V2 is available */
2380 smc_find_rdma_v2_device_serv(new_smc, pclc, ini);
2381 if (ini->smcrv2.ib_dev_v2)
2384 /* check if RDMA V1 is available */
2388 rc = smc_find_rdma_v1_device_serv(new_smc, pclc, ini);
2389 smc_find_ism_store_rc(rc, ini);
2390 return (!rc) ? 0 : ini->rc;
2395 /* listen worker: finish RDMA setup */
2396 static int smc_listen_rdma_finish(struct smc_sock *new_smc,
2397 struct smc_clc_msg_accept_confirm *cclc,
2399 struct smc_init_info *ini)
2401 struct smc_link *link = new_smc->conn.lnk;
2402 int reason_code = 0;
2405 smc_link_save_peer_info(link, cclc, ini);
2407 if (smc_rmb_rtoken_handling(&new_smc->conn, link, cclc))
2408 return SMC_CLC_DECL_ERR_RTOK;
2411 if (smc_ib_ready_link(link))
2412 return SMC_CLC_DECL_ERR_RDYLNK;
2413 /* QP confirmation over RoCE fabric */
2414 smc_llc_flow_initiate(link->lgr, SMC_LLC_FLOW_ADD_LINK);
2415 reason_code = smcr_serv_conf_first_link(new_smc);
2416 smc_llc_flow_stop(link->lgr, &link->lgr->llc_flow_lcl);
2421 /* setup for connection of server */
2422 static void smc_listen_work(struct work_struct *work)
2424 struct smc_sock *new_smc = container_of(work, struct smc_sock,
2426 struct socket *newclcsock = new_smc->clcsock;
2427 struct smc_clc_msg_accept_confirm *cclc;
2428 struct smc_clc_msg_proposal_area *buf;
2429 struct smc_clc_msg_proposal *pclc;
2430 struct smc_init_info *ini = NULL;
2431 u8 proposal_version = SMC_V1;
2435 if (new_smc->listen_smc->sk.sk_state != SMC_LISTEN)
2436 return smc_listen_out_err(new_smc);
2438 if (new_smc->use_fallback) {
2439 smc_listen_out_connected(new_smc);
2443 /* check if peer is smc capable */
2444 if (!tcp_sk(newclcsock->sk)->syn_smc) {
2445 rc = smc_switch_to_fallback(new_smc, SMC_CLC_DECL_PEERNOSMC);
2447 smc_listen_out_err(new_smc);
2449 smc_listen_out_connected(new_smc);
2453 /* do inband token exchange -
2454 * wait for and receive SMC Proposal CLC message
2456 buf = kzalloc(sizeof(*buf), GFP_KERNEL);
2458 rc = SMC_CLC_DECL_MEM;
2461 pclc = (struct smc_clc_msg_proposal *)buf;
2462 rc = smc_clc_wait_msg(new_smc, pclc, sizeof(*buf),
2463 SMC_CLC_PROPOSAL, CLC_WAIT_TIME);
2467 if (pclc->hdr.version > SMC_V1)
2468 proposal_version = SMC_V2;
2470 /* IPSec connections opt out of SMC optimizations */
2471 if (using_ipsec(new_smc)) {
2472 rc = SMC_CLC_DECL_IPSEC;
2476 ini = kzalloc(sizeof(*ini), GFP_KERNEL);
2478 rc = SMC_CLC_DECL_MEM;
2482 /* initial version checking */
2483 rc = smc_listen_v2_check(new_smc, pclc, ini);
2487 rc = smc_clc_srv_v2x_features_validate(new_smc, pclc, ini);
2491 mutex_lock(&smc_server_lgr_pending);
2492 smc_close_init(new_smc);
2493 smc_rx_init(new_smc);
2494 smc_tx_init(new_smc);
2496 /* determine ISM or RoCE device used for connection */
2497 rc = smc_listen_find_device(new_smc, pclc, ini);
2501 /* send SMC Accept CLC message */
2502 accept_version = ini->is_smcd ? ini->smcd_version : ini->smcr_version;
2503 rc = smc_clc_send_accept(new_smc, ini->first_contact_local,
2504 accept_version, ini->negotiated_eid, ini);
2508 /* SMC-D does not need this lock any more */
2510 mutex_unlock(&smc_server_lgr_pending);
2512 /* receive SMC Confirm CLC message */
2513 memset(buf, 0, sizeof(*buf));
2514 cclc = (struct smc_clc_msg_accept_confirm *)buf;
2515 rc = smc_clc_wait_msg(new_smc, cclc, sizeof(*buf),
2516 SMC_CLC_CONFIRM, CLC_WAIT_TIME);
2523 rc = smc_clc_v2x_features_confirm_check(cclc, ini);
2530 /* fce smc release version is needed in smc_listen_rdma_finish,
2531 * so save fce info here.
2533 smc_conn_save_peer_info_fce(new_smc, cclc);
2536 if (!ini->is_smcd) {
2537 rc = smc_listen_rdma_finish(new_smc, cclc,
2538 ini->first_contact_local, ini);
2541 mutex_unlock(&smc_server_lgr_pending);
2543 smc_conn_save_peer_info(new_smc, cclc);
2544 smc_listen_out_connected(new_smc);
2545 SMC_STAT_SERV_SUCC_INC(sock_net(newclcsock->sk), ini);
2549 mutex_unlock(&smc_server_lgr_pending);
2551 smc_listen_decline(new_smc, rc, ini ? ini->first_contact_local : 0,
2558 static void smc_tcp_listen_work(struct work_struct *work)
2560 struct smc_sock *lsmc = container_of(work, struct smc_sock,
2562 struct sock *lsk = &lsmc->sk;
2563 struct smc_sock *new_smc;
2567 while (lsk->sk_state == SMC_LISTEN) {
2568 rc = smc_clcsock_accept(lsmc, &new_smc);
2569 if (rc) /* clcsock accept queue empty or error */
2574 if (tcp_sk(new_smc->clcsock->sk)->syn_smc)
2575 atomic_inc(&lsmc->queued_smc_hs);
2577 new_smc->listen_smc = lsmc;
2578 new_smc->use_fallback = lsmc->use_fallback;
2579 new_smc->fallback_rsn = lsmc->fallback_rsn;
2580 sock_hold(lsk); /* sock_put in smc_listen_work */
2581 INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
2582 smc_copy_sock_settings_to_smc(new_smc);
2583 sock_hold(&new_smc->sk); /* sock_put in passive closing */
2584 if (!queue_work(smc_hs_wq, &new_smc->smc_listen_work))
2585 sock_put(&new_smc->sk);
2590 sock_put(&lsmc->sk); /* sock_hold in smc_clcsock_data_ready() */
2593 static void smc_clcsock_data_ready(struct sock *listen_clcsock)
2595 struct smc_sock *lsmc;
2597 read_lock_bh(&listen_clcsock->sk_callback_lock);
2598 lsmc = smc_clcsock_user_data(listen_clcsock);
2601 lsmc->clcsk_data_ready(listen_clcsock);
2602 if (lsmc->sk.sk_state == SMC_LISTEN) {
2603 sock_hold(&lsmc->sk); /* sock_put in smc_tcp_listen_work() */
2604 if (!queue_work(smc_tcp_ls_wq, &lsmc->tcp_listen_work))
2605 sock_put(&lsmc->sk);
2608 read_unlock_bh(&listen_clcsock->sk_callback_lock);
2611 static int smc_listen(struct socket *sock, int backlog)
2613 struct sock *sk = sock->sk;
2614 struct smc_sock *smc;
2621 if ((sk->sk_state != SMC_INIT && sk->sk_state != SMC_LISTEN) ||
2622 smc->connect_nonblock || sock->state != SS_UNCONNECTED)
2626 if (sk->sk_state == SMC_LISTEN) {
2627 sk->sk_max_ack_backlog = backlog;
2630 /* some socket options are handled in core, so we could not apply
2631 * them to the clc socket -- copy smc socket options to clc socket
2633 smc_copy_sock_settings_to_clc(smc);
2634 if (!smc->use_fallback)
2635 tcp_sk(smc->clcsock->sk)->syn_smc = 1;
2637 /* save original sk_data_ready function and establish
2638 * smc-specific sk_data_ready function
2640 write_lock_bh(&smc->clcsock->sk->sk_callback_lock);
2641 smc->clcsock->sk->sk_user_data =
2642 (void *)((uintptr_t)smc | SK_USER_DATA_NOCOPY);
2643 smc_clcsock_replace_cb(&smc->clcsock->sk->sk_data_ready,
2644 smc_clcsock_data_ready, &smc->clcsk_data_ready);
2645 write_unlock_bh(&smc->clcsock->sk->sk_callback_lock);
2647 /* save original ops */
2648 smc->ori_af_ops = inet_csk(smc->clcsock->sk)->icsk_af_ops;
2650 smc->af_ops = *smc->ori_af_ops;
2651 smc->af_ops.syn_recv_sock = smc_tcp_syn_recv_sock;
2653 inet_csk(smc->clcsock->sk)->icsk_af_ops = &smc->af_ops;
2655 if (smc->limit_smc_hs)
2656 tcp_sk(smc->clcsock->sk)->smc_hs_congested = smc_hs_congested;
2658 rc = kernel_listen(smc->clcsock, backlog);
2660 write_lock_bh(&smc->clcsock->sk->sk_callback_lock);
2661 smc_clcsock_restore_cb(&smc->clcsock->sk->sk_data_ready,
2662 &smc->clcsk_data_ready);
2663 smc->clcsock->sk->sk_user_data = NULL;
2664 write_unlock_bh(&smc->clcsock->sk->sk_callback_lock);
2667 sk->sk_max_ack_backlog = backlog;
2668 sk->sk_ack_backlog = 0;
2669 sk->sk_state = SMC_LISTEN;
2676 static int smc_accept(struct socket *sock, struct socket *new_sock,
2677 int flags, bool kern)
2679 struct sock *sk = sock->sk, *nsk;
2680 DECLARE_WAITQUEUE(wait, current);
2681 struct smc_sock *lsmc;
2686 sock_hold(sk); /* sock_put below */
2689 if (lsmc->sk.sk_state != SMC_LISTEN) {
2695 /* Wait for an incoming connection */
2696 timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
2697 add_wait_queue_exclusive(sk_sleep(sk), &wait);
2698 while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
2699 set_current_state(TASK_INTERRUPTIBLE);
2705 timeo = schedule_timeout(timeo);
2706 /* wakeup by sk_data_ready in smc_listen_work() */
2707 sched_annotate_sleep();
2709 if (signal_pending(current)) {
2710 rc = sock_intr_errno(timeo);
2714 set_current_state(TASK_RUNNING);
2715 remove_wait_queue(sk_sleep(sk), &wait);
2718 rc = sock_error(nsk);
2723 if (lsmc->sockopt_defer_accept && !(flags & O_NONBLOCK)) {
2724 /* wait till data arrives on the socket */
2725 timeo = msecs_to_jiffies(lsmc->sockopt_defer_accept *
2727 if (smc_sk(nsk)->use_fallback) {
2728 struct sock *clcsk = smc_sk(nsk)->clcsock->sk;
2731 if (skb_queue_empty(&clcsk->sk_receive_queue))
2732 sk_wait_data(clcsk, &timeo, NULL);
2733 release_sock(clcsk);
2734 } else if (!atomic_read(&smc_sk(nsk)->conn.bytes_to_rcv)) {
2736 smc_rx_wait(smc_sk(nsk), &timeo, smc_rx_data_available);
2742 sock_put(sk); /* sock_hold above */
2746 static int smc_getname(struct socket *sock, struct sockaddr *addr,
2749 struct smc_sock *smc;
2751 if (peer && (sock->sk->sk_state != SMC_ACTIVE) &&
2752 (sock->sk->sk_state != SMC_APPCLOSEWAIT1))
2755 smc = smc_sk(sock->sk);
2757 return smc->clcsock->ops->getname(smc->clcsock, addr, peer);
2760 static int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
2762 struct sock *sk = sock->sk;
2763 struct smc_sock *smc;
2769 /* SMC does not support connect with fastopen */
2770 if (msg->msg_flags & MSG_FASTOPEN) {
2771 /* not connected yet, fallback */
2772 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
2773 rc = smc_switch_to_fallback(smc, SMC_CLC_DECL_OPTUNSUPP);
2780 } else if ((sk->sk_state != SMC_ACTIVE) &&
2781 (sk->sk_state != SMC_APPCLOSEWAIT1) &&
2782 (sk->sk_state != SMC_INIT)) {
2787 if (smc->use_fallback) {
2788 rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
2790 rc = smc_tx_sendmsg(smc, msg, len);
2791 SMC_STAT_TX_PAYLOAD(smc, len, rc);
2798 static int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
2801 struct sock *sk = sock->sk;
2802 struct smc_sock *smc;
2807 if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
2808 /* socket was connected before, no more data to read */
2812 if ((sk->sk_state == SMC_INIT) ||
2813 (sk->sk_state == SMC_LISTEN) ||
2814 (sk->sk_state == SMC_CLOSED))
2817 if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
2822 if (smc->use_fallback) {
2823 rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
2825 msg->msg_namelen = 0;
2826 rc = smc_rx_recvmsg(smc, msg, NULL, len, flags);
2827 SMC_STAT_RX_PAYLOAD(smc, rc, rc);
2835 static __poll_t smc_accept_poll(struct sock *parent)
2837 struct smc_sock *isk = smc_sk(parent);
2840 spin_lock(&isk->accept_q_lock);
2841 if (!list_empty(&isk->accept_q))
2842 mask = EPOLLIN | EPOLLRDNORM;
2843 spin_unlock(&isk->accept_q_lock);
2848 static __poll_t smc_poll(struct file *file, struct socket *sock,
2851 struct sock *sk = sock->sk;
2852 struct smc_sock *smc;
2858 smc = smc_sk(sock->sk);
2859 if (smc->use_fallback) {
2860 /* delegate to CLC child sock */
2861 mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
2862 sk->sk_err = smc->clcsock->sk->sk_err;
2864 if (sk->sk_state != SMC_CLOSED)
2865 sock_poll_wait(file, sock, wait);
2868 if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
2869 (sk->sk_state == SMC_CLOSED))
2871 if (sk->sk_state == SMC_LISTEN) {
2872 /* woken up by sk_data_ready in smc_listen_work() */
2873 mask |= smc_accept_poll(sk);
2874 } else if (smc->use_fallback) { /* as result of connect_work()*/
2875 mask |= smc->clcsock->ops->poll(file, smc->clcsock,
2877 sk->sk_err = smc->clcsock->sk->sk_err;
2879 if ((sk->sk_state != SMC_INIT &&
2880 atomic_read(&smc->conn.sndbuf_space)) ||
2881 sk->sk_shutdown & SEND_SHUTDOWN) {
2882 mask |= EPOLLOUT | EPOLLWRNORM;
2884 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
2885 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
2887 if (atomic_read(&smc->conn.bytes_to_rcv))
2888 mask |= EPOLLIN | EPOLLRDNORM;
2889 if (sk->sk_shutdown & RCV_SHUTDOWN)
2890 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
2891 if (sk->sk_state == SMC_APPCLOSEWAIT1)
2893 if (smc->conn.urg_state == SMC_URG_VALID)
2901 static int smc_shutdown(struct socket *sock, int how)
2903 struct sock *sk = sock->sk;
2904 bool do_shutdown = true;
2905 struct smc_sock *smc;
2912 if ((how < SHUT_RD) || (how > SHUT_RDWR))
2917 if (sock->state == SS_CONNECTING) {
2918 if (sk->sk_state == SMC_ACTIVE)
2919 sock->state = SS_CONNECTED;
2920 else if (sk->sk_state == SMC_PEERCLOSEWAIT1 ||
2921 sk->sk_state == SMC_PEERCLOSEWAIT2 ||
2922 sk->sk_state == SMC_APPCLOSEWAIT1 ||
2923 sk->sk_state == SMC_APPCLOSEWAIT2 ||
2924 sk->sk_state == SMC_APPFINCLOSEWAIT)
2925 sock->state = SS_DISCONNECTING;
2929 if ((sk->sk_state != SMC_ACTIVE) &&
2930 (sk->sk_state != SMC_PEERCLOSEWAIT1) &&
2931 (sk->sk_state != SMC_PEERCLOSEWAIT2) &&
2932 (sk->sk_state != SMC_APPCLOSEWAIT1) &&
2933 (sk->sk_state != SMC_APPCLOSEWAIT2) &&
2934 (sk->sk_state != SMC_APPFINCLOSEWAIT))
2936 if (smc->use_fallback) {
2937 rc = kernel_sock_shutdown(smc->clcsock, how);
2938 sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
2939 if (sk->sk_shutdown == SHUTDOWN_MASK) {
2940 sk->sk_state = SMC_CLOSED;
2941 sk->sk_socket->state = SS_UNCONNECTED;
2947 case SHUT_RDWR: /* shutdown in both directions */
2948 old_state = sk->sk_state;
2949 rc = smc_close_active(smc);
2950 if (old_state == SMC_ACTIVE &&
2951 sk->sk_state == SMC_PEERCLOSEWAIT1)
2952 do_shutdown = false;
2955 rc = smc_close_shutdown_write(smc);
2959 /* nothing more to do because peer is not involved */
2962 if (do_shutdown && smc->clcsock)
2963 rc1 = kernel_sock_shutdown(smc->clcsock, how);
2964 /* map sock_shutdown_cmd constants to sk_shutdown value range */
2965 sk->sk_shutdown |= how + 1;
2967 if (sk->sk_state == SMC_CLOSED)
2968 sock->state = SS_UNCONNECTED;
2970 sock->state = SS_DISCONNECTING;
2973 return rc ? rc : rc1;
2976 static int __smc_getsockopt(struct socket *sock, int level, int optname,
2977 char __user *optval, int __user *optlen)
2979 struct smc_sock *smc;
2982 smc = smc_sk(sock->sk);
2984 if (get_user(len, optlen))
2987 len = min_t(int, len, sizeof(int));
2994 val = smc->limit_smc_hs;
3000 if (put_user(len, optlen))
3002 if (copy_to_user(optval, &val, len))
3008 static int __smc_setsockopt(struct socket *sock, int level, int optname,
3009 sockptr_t optval, unsigned int optlen)
3011 struct sock *sk = sock->sk;
3012 struct smc_sock *smc;
3020 if (optlen < sizeof(int)) {
3024 if (copy_from_sockptr(&val, optval, sizeof(int))) {
3029 smc->limit_smc_hs = !!val;
3041 static int smc_setsockopt(struct socket *sock, int level, int optname,
3042 sockptr_t optval, unsigned int optlen)
3044 struct sock *sk = sock->sk;
3045 struct smc_sock *smc;
3048 if (level == SOL_TCP && optname == TCP_ULP)
3050 else if (level == SOL_SMC)
3051 return __smc_setsockopt(sock, level, optname, optval, optlen);
3055 /* generic setsockopts reaching us here always apply to the
3058 mutex_lock(&smc->clcsock_release_lock);
3059 if (!smc->clcsock) {
3060 mutex_unlock(&smc->clcsock_release_lock);
3063 if (unlikely(!smc->clcsock->ops->setsockopt))
3066 rc = smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
3068 if (smc->clcsock->sk->sk_err) {
3069 sk->sk_err = smc->clcsock->sk->sk_err;
3070 sk_error_report(sk);
3072 mutex_unlock(&smc->clcsock_release_lock);
3074 if (optlen < sizeof(int))
3076 if (copy_from_sockptr(&val, optval, sizeof(int)))
3080 if (rc || smc->use_fallback)
3084 case TCP_FASTOPEN_CONNECT:
3085 case TCP_FASTOPEN_KEY:
3086 case TCP_FASTOPEN_NO_COOKIE:
3087 /* option not supported by SMC */
3088 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
3089 rc = smc_switch_to_fallback(smc, SMC_CLC_DECL_OPTUNSUPP);
3095 if (sk->sk_state != SMC_INIT &&
3096 sk->sk_state != SMC_LISTEN &&
3097 sk->sk_state != SMC_CLOSED) {
3099 SMC_STAT_INC(smc, ndly_cnt);
3100 smc_tx_pending(&smc->conn);
3101 cancel_delayed_work(&smc->conn.tx_work);
3106 if (sk->sk_state != SMC_INIT &&
3107 sk->sk_state != SMC_LISTEN &&
3108 sk->sk_state != SMC_CLOSED) {
3110 SMC_STAT_INC(smc, cork_cnt);
3111 smc_tx_pending(&smc->conn);
3112 cancel_delayed_work(&smc->conn.tx_work);
3116 case TCP_DEFER_ACCEPT:
3117 smc->sockopt_defer_accept = val;
3128 static int smc_getsockopt(struct socket *sock, int level, int optname,
3129 char __user *optval, int __user *optlen)
3131 struct smc_sock *smc;
3134 if (level == SOL_SMC)
3135 return __smc_getsockopt(sock, level, optname, optval, optlen);
3137 smc = smc_sk(sock->sk);
3138 mutex_lock(&smc->clcsock_release_lock);
3139 if (!smc->clcsock) {
3140 mutex_unlock(&smc->clcsock_release_lock);
3143 /* socket options apply to the CLC socket */
3144 if (unlikely(!smc->clcsock->ops->getsockopt)) {
3145 mutex_unlock(&smc->clcsock_release_lock);
3148 rc = smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
3150 mutex_unlock(&smc->clcsock_release_lock);
3154 static int smc_ioctl(struct socket *sock, unsigned int cmd,
3157 union smc_host_cursor cons, urg;
3158 struct smc_connection *conn;
3159 struct smc_sock *smc;
3162 smc = smc_sk(sock->sk);
3164 lock_sock(&smc->sk);
3165 if (smc->use_fallback) {
3166 if (!smc->clcsock) {
3167 release_sock(&smc->sk);
3170 answ = smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
3171 release_sock(&smc->sk);
3175 case SIOCINQ: /* same as FIONREAD */
3176 if (smc->sk.sk_state == SMC_LISTEN) {
3177 release_sock(&smc->sk);
3180 if (smc->sk.sk_state == SMC_INIT ||
3181 smc->sk.sk_state == SMC_CLOSED)
3184 answ = atomic_read(&smc->conn.bytes_to_rcv);
3187 /* output queue size (not send + not acked) */
3188 if (smc->sk.sk_state == SMC_LISTEN) {
3189 release_sock(&smc->sk);
3192 if (smc->sk.sk_state == SMC_INIT ||
3193 smc->sk.sk_state == SMC_CLOSED)
3196 answ = smc->conn.sndbuf_desc->len -
3197 atomic_read(&smc->conn.sndbuf_space);
3200 /* output queue size (not send only) */
3201 if (smc->sk.sk_state == SMC_LISTEN) {
3202 release_sock(&smc->sk);
3205 if (smc->sk.sk_state == SMC_INIT ||
3206 smc->sk.sk_state == SMC_CLOSED)
3209 answ = smc_tx_prepared_sends(&smc->conn);
3212 if (smc->sk.sk_state == SMC_LISTEN) {
3213 release_sock(&smc->sk);
3216 if (smc->sk.sk_state == SMC_INIT ||
3217 smc->sk.sk_state == SMC_CLOSED) {
3220 smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
3221 smc_curs_copy(&urg, &conn->urg_curs, conn);
3222 answ = smc_curs_diff(conn->rmb_desc->len,
3227 release_sock(&smc->sk);
3228 return -ENOIOCTLCMD;
3230 release_sock(&smc->sk);
3232 return put_user(answ, (int __user *)arg);
3235 /* Map the affected portions of the rmbe into an spd, note the number of bytes
3236 * to splice in conn->splice_pending, and press 'go'. Delays consumer cursor
3237 * updates till whenever a respective page has been fully processed.
3238 * Note that subsequent recv() calls have to wait till all splice() processing
3241 static ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
3242 struct pipe_inode_info *pipe, size_t len,
3245 struct sock *sk = sock->sk;
3246 struct smc_sock *smc;
3251 if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
3252 /* socket was connected before, no more data to read */
3256 if (sk->sk_state == SMC_INIT ||
3257 sk->sk_state == SMC_LISTEN ||
3258 sk->sk_state == SMC_CLOSED)
3261 if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
3266 if (smc->use_fallback) {
3267 rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
3274 if (flags & SPLICE_F_NONBLOCK)
3275 flags = MSG_DONTWAIT;
3278 SMC_STAT_INC(smc, splice_cnt);
3279 rc = smc_rx_recvmsg(smc, NULL, pipe, len, flags);
3287 /* must look like tcp */
3288 static const struct proto_ops smc_sock_ops = {
3290 .owner = THIS_MODULE,
3291 .release = smc_release,
3293 .connect = smc_connect,
3294 .socketpair = sock_no_socketpair,
3295 .accept = smc_accept,
3296 .getname = smc_getname,
3299 .listen = smc_listen,
3300 .shutdown = smc_shutdown,
3301 .setsockopt = smc_setsockopt,
3302 .getsockopt = smc_getsockopt,
3303 .sendmsg = smc_sendmsg,
3304 .recvmsg = smc_recvmsg,
3305 .mmap = sock_no_mmap,
3306 .splice_read = smc_splice_read,
3309 static int __smc_create(struct net *net, struct socket *sock, int protocol,
3310 int kern, struct socket *clcsock)
3312 int family = (protocol == SMCPROTO_SMC6) ? PF_INET6 : PF_INET;
3313 struct smc_sock *smc;
3317 rc = -ESOCKTNOSUPPORT;
3318 if (sock->type != SOCK_STREAM)
3321 rc = -EPROTONOSUPPORT;
3322 if (protocol != SMCPROTO_SMC && protocol != SMCPROTO_SMC6)
3326 sock->ops = &smc_sock_ops;
3327 sock->state = SS_UNCONNECTED;
3328 sk = smc_sock_alloc(net, sock, protocol);
3332 /* create internal TCP socket for CLC handshake and fallback */
3334 smc->use_fallback = false; /* assume rdma capability first */
3335 smc->fallback_rsn = 0;
3337 /* default behavior from limit_smc_hs in every net namespace */
3338 smc->limit_smc_hs = net->smc.limit_smc_hs;
3342 rc = sock_create_kern(net, family, SOCK_STREAM, IPPROTO_TCP,
3345 sk_common_release(sk);
3349 /* smc_clcsock_release() does not wait smc->clcsock->sk's
3350 * destruction; its sk_state might not be TCP_CLOSE after
3351 * smc->sk is close()d, and TCP timers can be fired later,
3352 * which need net ref.
3354 sk = smc->clcsock->sk;
3355 __netns_tracker_free(net, &sk->ns_tracker, false);
3356 sk->sk_net_refcnt = 1;
3357 get_net_track(net, &sk->ns_tracker, GFP_KERNEL);
3358 sock_inuse_add(net, 1);
3360 smc->clcsock = clcsock;
3367 static int smc_create(struct net *net, struct socket *sock, int protocol,
3370 return __smc_create(net, sock, protocol, kern, NULL);
3373 static const struct net_proto_family smc_sock_family_ops = {
3375 .owner = THIS_MODULE,
3376 .create = smc_create,
3379 static int smc_ulp_init(struct sock *sk)
3381 struct socket *tcp = sk->sk_socket;
3382 struct net *net = sock_net(sk);
3383 struct socket *smcsock;
3386 /* only TCP can be replaced */
3387 if (tcp->type != SOCK_STREAM || sk->sk_protocol != IPPROTO_TCP ||
3388 (sk->sk_family != AF_INET && sk->sk_family != AF_INET6))
3389 return -ESOCKTNOSUPPORT;
3390 /* don't handle wq now */
3391 if (tcp->state != SS_UNCONNECTED || !tcp->file || tcp->wq.fasync_list)
3394 if (sk->sk_family == AF_INET)
3395 protocol = SMCPROTO_SMC;
3397 protocol = SMCPROTO_SMC6;
3399 smcsock = sock_alloc();
3403 smcsock->type = SOCK_STREAM;
3404 __module_get(THIS_MODULE); /* tried in __tcp_ulp_find_autoload */
3405 ret = __smc_create(net, smcsock, protocol, 1, tcp);
3407 sock_release(smcsock); /* module_put() which ops won't be NULL */
3411 /* replace tcp socket to smc */
3412 smcsock->file = tcp->file;
3413 smcsock->file->private_data = smcsock;
3414 smcsock->file->f_inode = SOCK_INODE(smcsock); /* replace inode when sock_close */
3415 smcsock->file->f_path.dentry->d_inode = SOCK_INODE(smcsock); /* dput() in __fput */
3421 static void smc_ulp_clone(const struct request_sock *req, struct sock *newsk,
3422 const gfp_t priority)
3424 struct inet_connection_sock *icsk = inet_csk(newsk);
3426 /* don't inherit ulp ops to child when listen */
3427 icsk->icsk_ulp_ops = NULL;
3430 static struct tcp_ulp_ops smc_ulp_ops __read_mostly = {
3432 .owner = THIS_MODULE,
3433 .init = smc_ulp_init,
3434 .clone = smc_ulp_clone,
3437 unsigned int smc_net_id;
3439 static __net_init int smc_net_init(struct net *net)
3443 rc = smc_sysctl_net_init(net);
3446 return smc_pnet_net_init(net);
3449 static void __net_exit smc_net_exit(struct net *net)
3451 smc_sysctl_net_exit(net);
3452 smc_pnet_net_exit(net);
3455 static __net_init int smc_net_stat_init(struct net *net)
3457 return smc_stats_init(net);
3460 static void __net_exit smc_net_stat_exit(struct net *net)
3462 smc_stats_exit(net);
3465 static struct pernet_operations smc_net_ops = {
3466 .init = smc_net_init,
3467 .exit = smc_net_exit,
3469 .size = sizeof(struct smc_net),
3472 static struct pernet_operations smc_net_stat_ops = {
3473 .init = smc_net_stat_init,
3474 .exit = smc_net_stat_exit,
3477 static int __init smc_init(void)
3481 rc = register_pernet_subsys(&smc_net_ops);
3485 rc = register_pernet_subsys(&smc_net_stat_ops);
3487 goto out_pernet_subsys;
3489 rc = smc_ism_init();
3491 goto out_pernet_subsys_stat;
3498 rc = smc_pnet_init();
3504 smc_tcp_ls_wq = alloc_workqueue("smc_tcp_ls_wq", 0, 0);
3508 smc_hs_wq = alloc_workqueue("smc_hs_wq", 0, 0);
3510 goto out_alloc_tcp_ls_wq;
3512 smc_close_wq = alloc_workqueue("smc_close_wq", 0, 0);
3514 goto out_alloc_hs_wq;
3516 rc = smc_core_init();
3518 pr_err("%s: smc_core_init fails with %d\n", __func__, rc);
3522 rc = smc_llc_init();
3524 pr_err("%s: smc_llc_init fails with %d\n", __func__, rc);
3528 rc = smc_cdc_init();
3530 pr_err("%s: smc_cdc_init fails with %d\n", __func__, rc);
3534 rc = proto_register(&smc_proto, 1);
3536 pr_err("%s: proto_register(v4) fails with %d\n", __func__, rc);
3540 rc = proto_register(&smc_proto6, 1);
3542 pr_err("%s: proto_register(v6) fails with %d\n", __func__, rc);
3546 rc = sock_register(&smc_sock_family_ops);
3548 pr_err("%s: sock_register fails with %d\n", __func__, rc);
3551 INIT_HLIST_HEAD(&smc_v4_hashinfo.ht);
3552 INIT_HLIST_HEAD(&smc_v6_hashinfo.ht);
3554 rc = smc_ib_register_client();
3556 pr_err("%s: ib_register fails with %d\n", __func__, rc);
3560 rc = tcp_register_ulp(&smc_ulp_ops);
3562 pr_err("%s: tcp_ulp_register fails with %d\n", __func__, rc);
3566 static_branch_enable(&tcp_have_smc);
3570 smc_ib_unregister_client();
3572 sock_unregister(PF_SMC);
3574 proto_unregister(&smc_proto6);
3576 proto_unregister(&smc_proto);
3580 destroy_workqueue(smc_close_wq);
3582 destroy_workqueue(smc_hs_wq);
3583 out_alloc_tcp_ls_wq:
3584 destroy_workqueue(smc_tcp_ls_wq);
3592 out_pernet_subsys_stat:
3593 unregister_pernet_subsys(&smc_net_stat_ops);
3595 unregister_pernet_subsys(&smc_net_ops);
3600 static void __exit smc_exit(void)
3602 static_branch_disable(&tcp_have_smc);
3603 tcp_unregister_ulp(&smc_ulp_ops);
3604 sock_unregister(PF_SMC);
3606 smc_ib_unregister_client();
3608 destroy_workqueue(smc_close_wq);
3609 destroy_workqueue(smc_tcp_ls_wq);
3610 destroy_workqueue(smc_hs_wq);
3611 proto_unregister(&smc_proto6);
3612 proto_unregister(&smc_proto);
3616 unregister_pernet_subsys(&smc_net_stat_ops);
3617 unregister_pernet_subsys(&smc_net_ops);
3621 module_init(smc_init);
3622 module_exit(smc_exit);
3624 MODULE_AUTHOR("Ursula Braun <ubraun@linux.vnet.ibm.com>");
3625 MODULE_DESCRIPTION("smc socket address family");
3626 MODULE_LICENSE("GPL");
3627 MODULE_ALIAS_NETPROTO(PF_SMC);
3628 MODULE_ALIAS_TCP_ULP("smc");
3629 MODULE_ALIAS_GENL_FAMILY(SMC_GENL_FAMILY_NAME);