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>
33 #include <asm/ioctls.h>
35 #include <net/net_namespace.h>
36 #include <net/netns/generic.h>
37 #include "smc_netns.h"
49 #include "smc_close.h"
51 static DEFINE_MUTEX(smc_server_lgr_pending); /* serialize link group
54 static DEFINE_MUTEX(smc_client_lgr_pending); /* serialize link group
58 static void smc_tcp_listen_work(struct work_struct *);
59 static void smc_connect_work(struct work_struct *);
61 static void smc_set_keepalive(struct sock *sk, int val)
63 struct smc_sock *smc = smc_sk(sk);
65 smc->clcsock->sk->sk_prot->keepalive(smc->clcsock->sk, val);
68 static struct smc_hashinfo smc_v4_hashinfo = {
69 .lock = __RW_LOCK_UNLOCKED(smc_v4_hashinfo.lock),
72 static struct smc_hashinfo smc_v6_hashinfo = {
73 .lock = __RW_LOCK_UNLOCKED(smc_v6_hashinfo.lock),
76 int smc_hash_sk(struct sock *sk)
78 struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
79 struct hlist_head *head;
83 write_lock_bh(&h->lock);
84 sk_add_node(sk, head);
85 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
86 write_unlock_bh(&h->lock);
90 EXPORT_SYMBOL_GPL(smc_hash_sk);
92 void smc_unhash_sk(struct sock *sk)
94 struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
96 write_lock_bh(&h->lock);
97 if (sk_del_node_init(sk))
98 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
99 write_unlock_bh(&h->lock);
101 EXPORT_SYMBOL_GPL(smc_unhash_sk);
103 struct proto smc_proto = {
105 .owner = THIS_MODULE,
106 .keepalive = smc_set_keepalive,
108 .unhash = smc_unhash_sk,
109 .obj_size = sizeof(struct smc_sock),
110 .h.smc_hash = &smc_v4_hashinfo,
111 .slab_flags = SLAB_TYPESAFE_BY_RCU,
113 EXPORT_SYMBOL_GPL(smc_proto);
115 struct proto smc_proto6 = {
117 .owner = THIS_MODULE,
118 .keepalive = smc_set_keepalive,
120 .unhash = smc_unhash_sk,
121 .obj_size = sizeof(struct smc_sock),
122 .h.smc_hash = &smc_v6_hashinfo,
123 .slab_flags = SLAB_TYPESAFE_BY_RCU,
125 EXPORT_SYMBOL_GPL(smc_proto6);
127 static void smc_restore_fallback_changes(struct smc_sock *smc)
129 smc->clcsock->file->private_data = smc->sk.sk_socket;
130 smc->clcsock->file = NULL;
133 static int __smc_release(struct smc_sock *smc)
135 struct sock *sk = &smc->sk;
138 if (!smc->use_fallback) {
139 rc = smc_close_active(smc);
140 sock_set_flag(sk, SOCK_DEAD);
141 sk->sk_shutdown |= SHUTDOWN_MASK;
143 if (sk->sk_state != SMC_LISTEN && sk->sk_state != SMC_INIT)
144 sock_put(sk); /* passive closing */
145 if (sk->sk_state == SMC_LISTEN) {
146 /* wake up clcsock accept */
147 rc = kernel_sock_shutdown(smc->clcsock, SHUT_RDWR);
149 sk->sk_state = SMC_CLOSED;
150 sk->sk_state_change(sk);
151 smc_restore_fallback_changes(smc);
154 sk->sk_prot->unhash(sk);
156 if (sk->sk_state == SMC_CLOSED) {
159 smc_clcsock_release(smc);
162 if (!smc->use_fallback)
163 smc_conn_free(&smc->conn);
169 static int smc_release(struct socket *sock)
171 struct sock *sk = sock->sk;
172 struct smc_sock *smc;
178 sock_hold(sk); /* sock_put below */
181 /* cleanup for a dangling non-blocking connect */
182 if (smc->connect_nonblock && sk->sk_state == SMC_INIT)
183 tcp_abort(smc->clcsock->sk, ECONNABORTED);
184 flush_work(&smc->connect_work);
186 if (sk->sk_state == SMC_LISTEN)
187 /* smc_close_non_accepted() is called and acquires
188 * sock lock for child sockets again
190 lock_sock_nested(sk, SINGLE_DEPTH_NESTING);
194 rc = __smc_release(smc);
201 sock_put(sk); /* sock_hold above */
202 sock_put(sk); /* final sock_put */
207 static void smc_destruct(struct sock *sk)
209 if (sk->sk_state != SMC_CLOSED)
211 if (!sock_flag(sk, SOCK_DEAD))
214 sk_refcnt_debug_dec(sk);
217 static struct sock *smc_sock_alloc(struct net *net, struct socket *sock,
220 struct smc_sock *smc;
224 prot = (protocol == SMCPROTO_SMC6) ? &smc_proto6 : &smc_proto;
225 sk = sk_alloc(net, PF_SMC, GFP_KERNEL, prot, 0);
229 sock_init_data(sock, sk); /* sets sk_refcnt to 1 */
230 sk->sk_state = SMC_INIT;
231 sk->sk_destruct = smc_destruct;
232 sk->sk_protocol = protocol;
234 INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
235 INIT_WORK(&smc->connect_work, smc_connect_work);
236 INIT_DELAYED_WORK(&smc->conn.tx_work, smc_tx_work);
237 INIT_LIST_HEAD(&smc->accept_q);
238 spin_lock_init(&smc->accept_q_lock);
239 spin_lock_init(&smc->conn.send_lock);
240 sk->sk_prot->hash(sk);
241 sk_refcnt_debug_inc(sk);
242 mutex_init(&smc->clcsock_release_lock);
247 static int smc_bind(struct socket *sock, struct sockaddr *uaddr,
250 struct sockaddr_in *addr = (struct sockaddr_in *)uaddr;
251 struct sock *sk = sock->sk;
252 struct smc_sock *smc;
257 /* replicate tests from inet_bind(), to be safe wrt. future changes */
259 if (addr_len < sizeof(struct sockaddr_in))
263 if (addr->sin_family != AF_INET &&
264 addr->sin_family != AF_INET6 &&
265 addr->sin_family != AF_UNSPEC)
267 /* accept AF_UNSPEC (mapped to AF_INET) only if s_addr is INADDR_ANY */
268 if (addr->sin_family == AF_UNSPEC &&
269 addr->sin_addr.s_addr != htonl(INADDR_ANY))
274 /* Check if socket is already active */
276 if (sk->sk_state != SMC_INIT || smc->connect_nonblock)
279 smc->clcsock->sk->sk_reuse = sk->sk_reuse;
280 rc = kernel_bind(smc->clcsock, uaddr, addr_len);
288 static void smc_copy_sock_settings(struct sock *nsk, struct sock *osk,
291 /* options we don't get control via setsockopt for */
292 nsk->sk_type = osk->sk_type;
293 nsk->sk_sndbuf = osk->sk_sndbuf;
294 nsk->sk_rcvbuf = osk->sk_rcvbuf;
295 nsk->sk_sndtimeo = osk->sk_sndtimeo;
296 nsk->sk_rcvtimeo = osk->sk_rcvtimeo;
297 nsk->sk_mark = osk->sk_mark;
298 nsk->sk_priority = osk->sk_priority;
299 nsk->sk_rcvlowat = osk->sk_rcvlowat;
300 nsk->sk_bound_dev_if = osk->sk_bound_dev_if;
301 nsk->sk_err = osk->sk_err;
303 nsk->sk_flags &= ~mask;
304 nsk->sk_flags |= osk->sk_flags & mask;
307 #define SK_FLAGS_SMC_TO_CLC ((1UL << SOCK_URGINLINE) | \
308 (1UL << SOCK_KEEPOPEN) | \
309 (1UL << SOCK_LINGER) | \
310 (1UL << SOCK_BROADCAST) | \
311 (1UL << SOCK_TIMESTAMP) | \
312 (1UL << SOCK_DBG) | \
313 (1UL << SOCK_RCVTSTAMP) | \
314 (1UL << SOCK_RCVTSTAMPNS) | \
315 (1UL << SOCK_LOCALROUTE) | \
316 (1UL << SOCK_TIMESTAMPING_RX_SOFTWARE) | \
317 (1UL << SOCK_RXQ_OVFL) | \
318 (1UL << SOCK_WIFI_STATUS) | \
319 (1UL << SOCK_NOFCS) | \
320 (1UL << SOCK_FILTER_LOCKED) | \
321 (1UL << SOCK_TSTAMP_NEW))
322 /* copy only relevant settings and flags of SOL_SOCKET level from smc to
323 * clc socket (since smc is not called for these options from net/core)
325 static void smc_copy_sock_settings_to_clc(struct smc_sock *smc)
327 smc_copy_sock_settings(smc->clcsock->sk, &smc->sk, SK_FLAGS_SMC_TO_CLC);
330 #define SK_FLAGS_CLC_TO_SMC ((1UL << SOCK_URGINLINE) | \
331 (1UL << SOCK_KEEPOPEN) | \
332 (1UL << SOCK_LINGER) | \
334 /* copy only settings and flags relevant for smc from clc to smc socket */
335 static void smc_copy_sock_settings_to_smc(struct smc_sock *smc)
337 smc_copy_sock_settings(&smc->sk, smc->clcsock->sk, SK_FLAGS_CLC_TO_SMC);
340 /* register the new rmb on all links */
341 static int smcr_lgr_reg_rmbs(struct smc_link *link,
342 struct smc_buf_desc *rmb_desc)
344 struct smc_link_group *lgr = link->lgr;
347 rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY);
350 /* protect against parallel smc_llc_cli_rkey_exchange() and
351 * parallel smcr_link_reg_rmb()
353 mutex_lock(&lgr->llc_conf_mutex);
354 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
355 if (lgr->lnk[i].state != SMC_LNK_ACTIVE)
357 rc = smcr_link_reg_rmb(&lgr->lnk[i], rmb_desc);
362 /* exchange confirm_rkey msg with peer */
363 rc = smc_llc_do_confirm_rkey(link, rmb_desc);
368 rmb_desc->is_conf_rkey = true;
370 mutex_unlock(&lgr->llc_conf_mutex);
371 smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl);
375 static int smcr_clnt_conf_first_link(struct smc_sock *smc)
377 struct smc_link *link = smc->conn.lnk;
378 struct smc_llc_qentry *qentry;
381 /* receive CONFIRM LINK request from server over RoCE fabric */
382 qentry = smc_llc_wait(link->lgr, NULL, SMC_LLC_WAIT_TIME,
383 SMC_LLC_CONFIRM_LINK);
385 struct smc_clc_msg_decline dclc;
387 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
388 SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
389 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
391 smc_llc_save_peer_uid(qentry);
392 rc = smc_llc_eval_conf_link(qentry, SMC_LLC_REQ);
393 smc_llc_flow_qentry_del(&link->lgr->llc_flow_lcl);
395 return SMC_CLC_DECL_RMBE_EC;
397 rc = smc_ib_modify_qp_rts(link);
399 return SMC_CLC_DECL_ERR_RDYLNK;
401 smc_wr_remember_qp_attr(link);
403 if (smcr_link_reg_rmb(link, smc->conn.rmb_desc))
404 return SMC_CLC_DECL_ERR_REGRMB;
406 /* confirm_rkey is implicit on 1st contact */
407 smc->conn.rmb_desc->is_conf_rkey = true;
409 /* send CONFIRM LINK response over RoCE fabric */
410 rc = smc_llc_send_confirm_link(link, SMC_LLC_RESP);
412 return SMC_CLC_DECL_TIMEOUT_CL;
414 smc_llc_link_active(link);
415 smcr_lgr_set_type(link->lgr, SMC_LGR_SINGLE);
417 /* optional 2nd link, receive ADD LINK request from server */
418 qentry = smc_llc_wait(link->lgr, NULL, SMC_LLC_WAIT_TIME,
421 struct smc_clc_msg_decline dclc;
423 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
424 SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
426 rc = 0; /* no DECLINE received, go with one link */
429 smc_llc_flow_qentry_clr(&link->lgr->llc_flow_lcl);
430 smc_llc_cli_add_link(link, qentry);
434 static void smcr_conn_save_peer_info(struct smc_sock *smc,
435 struct smc_clc_msg_accept_confirm *clc)
437 int bufsize = smc_uncompress_bufsize(clc->rmbe_size);
439 smc->conn.peer_rmbe_idx = clc->rmbe_idx;
440 smc->conn.local_tx_ctrl.token = ntohl(clc->rmbe_alert_token);
441 smc->conn.peer_rmbe_size = bufsize;
442 atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
443 smc->conn.tx_off = bufsize * (smc->conn.peer_rmbe_idx - 1);
446 static void smcd_conn_save_peer_info(struct smc_sock *smc,
447 struct smc_clc_msg_accept_confirm *clc)
449 int bufsize = smc_uncompress_bufsize(clc->dmbe_size);
451 smc->conn.peer_rmbe_idx = clc->dmbe_idx;
452 smc->conn.peer_token = clc->token;
453 /* msg header takes up space in the buffer */
454 smc->conn.peer_rmbe_size = bufsize - sizeof(struct smcd_cdc_msg);
455 atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
456 smc->conn.tx_off = bufsize * smc->conn.peer_rmbe_idx;
459 static void smc_conn_save_peer_info(struct smc_sock *smc,
460 struct smc_clc_msg_accept_confirm *clc)
462 if (smc->conn.lgr->is_smcd)
463 smcd_conn_save_peer_info(smc, clc);
465 smcr_conn_save_peer_info(smc, clc);
468 static void smc_link_save_peer_info(struct smc_link *link,
469 struct smc_clc_msg_accept_confirm *clc)
471 link->peer_qpn = ntoh24(clc->qpn);
472 memcpy(link->peer_gid, clc->lcl.gid, SMC_GID_SIZE);
473 memcpy(link->peer_mac, clc->lcl.mac, sizeof(link->peer_mac));
474 link->peer_psn = ntoh24(clc->psn);
475 link->peer_mtu = clc->qp_mtu;
478 static void smc_switch_to_fallback(struct smc_sock *smc)
480 smc->use_fallback = true;
481 if (smc->sk.sk_socket && smc->sk.sk_socket->file) {
482 smc->clcsock->file = smc->sk.sk_socket->file;
483 smc->clcsock->file->private_data = smc->clcsock;
484 smc->clcsock->wq.fasync_list =
485 smc->sk.sk_socket->wq.fasync_list;
489 /* fall back during connect */
490 static int smc_connect_fallback(struct smc_sock *smc, int reason_code)
492 smc_switch_to_fallback(smc);
493 smc->fallback_rsn = reason_code;
494 smc_copy_sock_settings_to_clc(smc);
495 smc->connect_nonblock = 0;
496 if (smc->sk.sk_state == SMC_INIT)
497 smc->sk.sk_state = SMC_ACTIVE;
501 /* decline and fall back during connect */
502 static int smc_connect_decline_fallback(struct smc_sock *smc, int reason_code)
506 if (reason_code < 0) { /* error, fallback is not possible */
507 if (smc->sk.sk_state == SMC_INIT)
508 sock_put(&smc->sk); /* passive closing */
511 if (reason_code != SMC_CLC_DECL_PEERDECL) {
512 rc = smc_clc_send_decline(smc, reason_code);
514 if (smc->sk.sk_state == SMC_INIT)
515 sock_put(&smc->sk); /* passive closing */
519 return smc_connect_fallback(smc, reason_code);
522 /* abort connecting */
523 static int smc_connect_abort(struct smc_sock *smc, int reason_code,
526 bool is_smcd = smc->conn.lgr->is_smcd;
528 if (local_contact == SMC_FIRST_CONTACT)
529 smc_lgr_cleanup_early(&smc->conn);
531 smc_conn_free(&smc->conn);
533 /* there is only one lgr role for SMC-D; use server lock */
534 mutex_unlock(&smc_server_lgr_pending);
536 mutex_unlock(&smc_client_lgr_pending);
538 smc->connect_nonblock = 0;
542 /* check if there is a rdma device available for this connection. */
543 /* called for connect and listen */
544 static int smc_find_rdma_device(struct smc_sock *smc, struct smc_init_info *ini)
546 /* PNET table look up: search active ib_device and port
547 * within same PNETID that also contains the ethernet device
548 * used for the internal TCP socket
550 smc_pnet_find_roce_resource(smc->clcsock->sk, ini);
552 return SMC_CLC_DECL_NOSMCRDEV;
556 /* check if there is an ISM device available for this connection. */
557 /* called for connect and listen */
558 static int smc_find_ism_device(struct smc_sock *smc, struct smc_init_info *ini)
560 /* Find ISM device with same PNETID as connecting interface */
561 smc_pnet_find_ism_resource(smc->clcsock->sk, ini);
563 return SMC_CLC_DECL_NOSMCDDEV;
567 /* Check for VLAN ID and register it on ISM device just for CLC handshake */
568 static int smc_connect_ism_vlan_setup(struct smc_sock *smc,
569 struct smc_init_info *ini)
571 if (ini->vlan_id && smc_ism_get_vlan(ini->ism_dev, ini->vlan_id))
572 return SMC_CLC_DECL_ISMVLANERR;
576 /* cleanup temporary VLAN ID registration used for CLC handshake. If ISM is
577 * used, the VLAN ID will be registered again during the connection setup.
579 static int smc_connect_ism_vlan_cleanup(struct smc_sock *smc, bool is_smcd,
580 struct smc_init_info *ini)
584 if (ini->vlan_id && smc_ism_put_vlan(ini->ism_dev, ini->vlan_id))
585 return SMC_CLC_DECL_CNFERR;
589 /* CLC handshake during connect */
590 static int smc_connect_clc(struct smc_sock *smc, int smc_type,
591 struct smc_clc_msg_accept_confirm *aclc,
592 struct smc_init_info *ini)
596 /* do inband token exchange */
597 rc = smc_clc_send_proposal(smc, smc_type, ini);
600 /* receive SMC Accept CLC message */
601 return smc_clc_wait_msg(smc, aclc, sizeof(*aclc), SMC_CLC_ACCEPT,
605 /* setup for RDMA connection of client */
606 static int smc_connect_rdma(struct smc_sock *smc,
607 struct smc_clc_msg_accept_confirm *aclc,
608 struct smc_init_info *ini)
610 int i, reason_code = 0;
611 struct smc_link *link;
613 ini->is_smcd = false;
614 ini->ib_lcl = &aclc->lcl;
615 ini->ib_clcqpn = ntoh24(aclc->qpn);
616 ini->srv_first_contact = aclc->hdr.flag;
618 mutex_lock(&smc_client_lgr_pending);
619 reason_code = smc_conn_create(smc, ini);
621 mutex_unlock(&smc_client_lgr_pending);
625 smc_conn_save_peer_info(smc, aclc);
627 if (ini->cln_first_contact == SMC_FIRST_CONTACT) {
628 link = smc->conn.lnk;
630 /* set link that was assigned by server */
632 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
633 struct smc_link *l = &smc->conn.lgr->lnk[i];
635 if (l->peer_qpn == ntoh24(aclc->qpn)) {
641 return smc_connect_abort(smc, SMC_CLC_DECL_NOSRVLINK,
642 ini->cln_first_contact);
643 smc->conn.lnk = link;
646 /* create send buffer and rmb */
647 if (smc_buf_create(smc, false))
648 return smc_connect_abort(smc, SMC_CLC_DECL_MEM,
649 ini->cln_first_contact);
651 if (ini->cln_first_contact == SMC_FIRST_CONTACT)
652 smc_link_save_peer_info(link, aclc);
654 if (smc_rmb_rtoken_handling(&smc->conn, link, aclc))
655 return smc_connect_abort(smc, SMC_CLC_DECL_ERR_RTOK,
656 ini->cln_first_contact);
661 if (ini->cln_first_contact == SMC_FIRST_CONTACT) {
662 if (smc_ib_ready_link(link))
663 return smc_connect_abort(smc, SMC_CLC_DECL_ERR_RDYLNK,
664 ini->cln_first_contact);
666 if (smcr_lgr_reg_rmbs(link, smc->conn.rmb_desc))
667 return smc_connect_abort(smc, SMC_CLC_DECL_ERR_REGRMB,
668 ini->cln_first_contact);
670 smc_rmb_sync_sg_for_device(&smc->conn);
672 reason_code = smc_clc_send_confirm(smc);
674 return smc_connect_abort(smc, reason_code,
675 ini->cln_first_contact);
679 if (ini->cln_first_contact == SMC_FIRST_CONTACT) {
680 /* QP confirmation over RoCE fabric */
681 smc_llc_flow_initiate(link->lgr, SMC_LLC_FLOW_ADD_LINK);
682 reason_code = smcr_clnt_conf_first_link(smc);
683 smc_llc_flow_stop(link->lgr, &link->lgr->llc_flow_lcl);
685 return smc_connect_abort(smc, reason_code,
686 ini->cln_first_contact);
688 mutex_unlock(&smc_client_lgr_pending);
690 smc_copy_sock_settings_to_clc(smc);
691 smc->connect_nonblock = 0;
692 if (smc->sk.sk_state == SMC_INIT)
693 smc->sk.sk_state = SMC_ACTIVE;
698 /* setup for ISM connection of client */
699 static int smc_connect_ism(struct smc_sock *smc,
700 struct smc_clc_msg_accept_confirm *aclc,
701 struct smc_init_info *ini)
706 ini->ism_gid = aclc->gid;
707 ini->srv_first_contact = aclc->hdr.flag;
709 /* there is only one lgr role for SMC-D; use server lock */
710 mutex_lock(&smc_server_lgr_pending);
711 rc = smc_conn_create(smc, ini);
713 mutex_unlock(&smc_server_lgr_pending);
717 /* Create send and receive buffers */
718 if (smc_buf_create(smc, true))
719 return smc_connect_abort(smc, SMC_CLC_DECL_MEM,
720 ini->cln_first_contact);
722 smc_conn_save_peer_info(smc, aclc);
727 rc = smc_clc_send_confirm(smc);
729 return smc_connect_abort(smc, rc, ini->cln_first_contact);
730 mutex_unlock(&smc_server_lgr_pending);
732 smc_copy_sock_settings_to_clc(smc);
733 smc->connect_nonblock = 0;
734 if (smc->sk.sk_state == SMC_INIT)
735 smc->sk.sk_state = SMC_ACTIVE;
740 /* perform steps before actually connecting */
741 static int __smc_connect(struct smc_sock *smc)
743 bool ism_supported = false, rdma_supported = false;
744 struct smc_clc_msg_accept_confirm aclc;
745 struct smc_init_info ini = {0};
749 if (smc->use_fallback)
750 return smc_connect_fallback(smc, smc->fallback_rsn);
752 /* if peer has not signalled SMC-capability, fall back */
753 if (!tcp_sk(smc->clcsock->sk)->syn_smc)
754 return smc_connect_fallback(smc, SMC_CLC_DECL_PEERNOSMC);
756 /* IPSec connections opt out of SMC-R optimizations */
757 if (using_ipsec(smc))
758 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_IPSEC);
760 /* get vlan id from IP device */
761 if (smc_vlan_by_tcpsk(smc->clcsock, &ini))
762 return smc_connect_decline_fallback(smc,
763 SMC_CLC_DECL_GETVLANERR);
765 /* check if there is an ism device available */
766 if (!smc_find_ism_device(smc, &ini) &&
767 !smc_connect_ism_vlan_setup(smc, &ini)) {
768 /* ISM is supported for this connection */
769 ism_supported = true;
770 smc_type = SMC_TYPE_D;
773 /* check if there is a rdma device available */
774 if (!smc_find_rdma_device(smc, &ini)) {
775 /* RDMA is supported for this connection */
776 rdma_supported = true;
778 smc_type = SMC_TYPE_B; /* both */
780 smc_type = SMC_TYPE_R; /* only RDMA */
783 /* if neither ISM nor RDMA are supported, fallback */
784 if (!rdma_supported && !ism_supported)
785 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_NOSMCDEV);
787 /* perform CLC handshake */
788 rc = smc_connect_clc(smc, smc_type, &aclc, &ini);
790 smc_connect_ism_vlan_cleanup(smc, ism_supported, &ini);
791 return smc_connect_decline_fallback(smc, rc);
794 /* depending on previous steps, connect using rdma or ism */
795 if (rdma_supported && aclc.hdr.path == SMC_TYPE_R)
796 rc = smc_connect_rdma(smc, &aclc, &ini);
797 else if (ism_supported && aclc.hdr.path == SMC_TYPE_D)
798 rc = smc_connect_ism(smc, &aclc, &ini);
800 rc = SMC_CLC_DECL_MODEUNSUPP;
802 smc_connect_ism_vlan_cleanup(smc, ism_supported, &ini);
803 return smc_connect_decline_fallback(smc, rc);
806 smc_connect_ism_vlan_cleanup(smc, ism_supported, &ini);
810 static void smc_connect_work(struct work_struct *work)
812 struct smc_sock *smc = container_of(work, struct smc_sock,
814 long timeo = smc->sk.sk_sndtimeo;
818 timeo = MAX_SCHEDULE_TIMEOUT;
819 lock_sock(smc->clcsock->sk);
820 if (smc->clcsock->sk->sk_err) {
821 smc->sk.sk_err = smc->clcsock->sk->sk_err;
822 } else if ((1 << smc->clcsock->sk->sk_state) &
823 (TCPF_SYN_SENT | TCP_SYN_RECV)) {
824 rc = sk_stream_wait_connect(smc->clcsock->sk, &timeo);
825 if ((rc == -EPIPE) &&
826 ((1 << smc->clcsock->sk->sk_state) &
827 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)))
830 release_sock(smc->clcsock->sk);
832 if (rc != 0 || smc->sk.sk_err) {
833 smc->sk.sk_state = SMC_CLOSED;
834 if (rc == -EPIPE || rc == -EAGAIN)
835 smc->sk.sk_err = EPIPE;
836 else if (signal_pending(current))
837 smc->sk.sk_err = -sock_intr_errno(timeo);
838 sock_put(&smc->sk); /* passive closing */
842 rc = __smc_connect(smc);
844 smc->sk.sk_err = -rc;
847 if (!sock_flag(&smc->sk, SOCK_DEAD)) {
848 if (smc->sk.sk_err) {
849 smc->sk.sk_state_change(&smc->sk);
850 } else { /* allow polling before and after fallback decision */
851 smc->clcsock->sk->sk_write_space(smc->clcsock->sk);
852 smc->sk.sk_write_space(&smc->sk);
855 release_sock(&smc->sk);
858 static int smc_connect(struct socket *sock, struct sockaddr *addr,
861 struct sock *sk = sock->sk;
862 struct smc_sock *smc;
867 /* separate smc parameter checking to be safe */
868 if (alen < sizeof(addr->sa_family))
870 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
874 switch (sk->sk_state) {
885 smc_copy_sock_settings_to_clc(smc);
886 tcp_sk(smc->clcsock->sk)->syn_smc = 1;
887 if (smc->connect_nonblock) {
891 rc = kernel_connect(smc->clcsock, addr, alen, flags);
892 if (rc && rc != -EINPROGRESS)
895 sock_hold(&smc->sk); /* sock put in passive closing */
896 if (smc->use_fallback)
898 if (flags & O_NONBLOCK) {
899 if (schedule_work(&smc->connect_work))
900 smc->connect_nonblock = 1;
903 rc = __smc_connect(smc);
907 rc = 0; /* success cases including fallback */
916 static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
918 struct socket *new_clcsock = NULL;
919 struct sock *lsk = &lsmc->sk;
924 new_sk = smc_sock_alloc(sock_net(lsk), NULL, lsk->sk_protocol);
927 lsk->sk_err = ENOMEM;
932 *new_smc = smc_sk(new_sk);
934 mutex_lock(&lsmc->clcsock_release_lock);
936 rc = kernel_accept(lsmc->clcsock, &new_clcsock, 0);
937 mutex_unlock(&lsmc->clcsock_release_lock);
941 if (rc < 0 || lsk->sk_state == SMC_CLOSED) {
942 new_sk->sk_prot->unhash(new_sk);
944 sock_release(new_clcsock);
945 new_sk->sk_state = SMC_CLOSED;
946 sock_set_flag(new_sk, SOCK_DEAD);
947 sock_put(new_sk); /* final */
952 (*new_smc)->clcsock = new_clcsock;
957 /* add a just created sock to the accept queue of the listen sock as
958 * candidate for a following socket accept call from user space
960 static void smc_accept_enqueue(struct sock *parent, struct sock *sk)
962 struct smc_sock *par = smc_sk(parent);
964 sock_hold(sk); /* sock_put in smc_accept_unlink () */
965 spin_lock(&par->accept_q_lock);
966 list_add_tail(&smc_sk(sk)->accept_q, &par->accept_q);
967 spin_unlock(&par->accept_q_lock);
968 sk_acceptq_added(parent);
971 /* remove a socket from the accept queue of its parental listening socket */
972 static void smc_accept_unlink(struct sock *sk)
974 struct smc_sock *par = smc_sk(sk)->listen_smc;
976 spin_lock(&par->accept_q_lock);
977 list_del_init(&smc_sk(sk)->accept_q);
978 spin_unlock(&par->accept_q_lock);
979 sk_acceptq_removed(&smc_sk(sk)->listen_smc->sk);
980 sock_put(sk); /* sock_hold in smc_accept_enqueue */
983 /* remove a sock from the accept queue to bind it to a new socket created
984 * for a socket accept call from user space
986 struct sock *smc_accept_dequeue(struct sock *parent,
987 struct socket *new_sock)
989 struct smc_sock *isk, *n;
992 list_for_each_entry_safe(isk, n, &smc_sk(parent)->accept_q, accept_q) {
993 new_sk = (struct sock *)isk;
995 smc_accept_unlink(new_sk);
996 if (new_sk->sk_state == SMC_CLOSED) {
997 new_sk->sk_prot->unhash(new_sk);
999 sock_release(isk->clcsock);
1000 isk->clcsock = NULL;
1002 sock_put(new_sk); /* final */
1006 sock_graft(new_sk, new_sock);
1007 if (isk->use_fallback) {
1008 smc_sk(new_sk)->clcsock->file = new_sock->file;
1009 isk->clcsock->file->private_data = isk->clcsock;
1017 /* clean up for a created but never accepted sock */
1018 void smc_close_non_accepted(struct sock *sk)
1020 struct smc_sock *smc = smc_sk(sk);
1022 sock_hold(sk); /* sock_put below */
1024 if (!sk->sk_lingertime)
1025 /* wait for peer closing */
1026 sk->sk_lingertime = SMC_MAX_STREAM_WAIT_TIMEOUT;
1029 sock_put(sk); /* sock_hold above */
1030 sock_put(sk); /* final sock_put */
1033 static int smcr_serv_conf_first_link(struct smc_sock *smc)
1035 struct smc_link *link = smc->conn.lnk;
1036 struct smc_llc_qentry *qentry;
1039 if (smcr_link_reg_rmb(link, smc->conn.rmb_desc))
1040 return SMC_CLC_DECL_ERR_REGRMB;
1042 /* send CONFIRM LINK request to client over the RoCE fabric */
1043 rc = smc_llc_send_confirm_link(link, SMC_LLC_REQ);
1045 return SMC_CLC_DECL_TIMEOUT_CL;
1047 /* receive CONFIRM LINK response from client over the RoCE fabric */
1048 qentry = smc_llc_wait(link->lgr, link, SMC_LLC_WAIT_TIME,
1049 SMC_LLC_CONFIRM_LINK);
1051 struct smc_clc_msg_decline dclc;
1053 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
1054 SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
1055 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
1057 smc_llc_save_peer_uid(qentry);
1058 rc = smc_llc_eval_conf_link(qentry, SMC_LLC_RESP);
1059 smc_llc_flow_qentry_del(&link->lgr->llc_flow_lcl);
1061 return SMC_CLC_DECL_RMBE_EC;
1063 /* confirm_rkey is implicit on 1st contact */
1064 smc->conn.rmb_desc->is_conf_rkey = true;
1066 smc_llc_link_active(link);
1067 smcr_lgr_set_type(link->lgr, SMC_LGR_SINGLE);
1069 /* initial contact - try to establish second link */
1070 smc_llc_srv_add_link(link);
1074 /* listen worker: finish */
1075 static void smc_listen_out(struct smc_sock *new_smc)
1077 struct smc_sock *lsmc = new_smc->listen_smc;
1078 struct sock *newsmcsk = &new_smc->sk;
1080 if (lsmc->sk.sk_state == SMC_LISTEN) {
1081 lock_sock_nested(&lsmc->sk, SINGLE_DEPTH_NESTING);
1082 smc_accept_enqueue(&lsmc->sk, newsmcsk);
1083 release_sock(&lsmc->sk);
1084 } else { /* no longer listening */
1085 smc_close_non_accepted(newsmcsk);
1088 /* Wake up accept */
1089 lsmc->sk.sk_data_ready(&lsmc->sk);
1090 sock_put(&lsmc->sk); /* sock_hold in smc_tcp_listen_work */
1093 /* listen worker: finish in state connected */
1094 static void smc_listen_out_connected(struct smc_sock *new_smc)
1096 struct sock *newsmcsk = &new_smc->sk;
1098 sk_refcnt_debug_inc(newsmcsk);
1099 if (newsmcsk->sk_state == SMC_INIT)
1100 newsmcsk->sk_state = SMC_ACTIVE;
1102 smc_listen_out(new_smc);
1105 /* listen worker: finish in error state */
1106 static void smc_listen_out_err(struct smc_sock *new_smc)
1108 struct sock *newsmcsk = &new_smc->sk;
1110 if (newsmcsk->sk_state == SMC_INIT)
1111 sock_put(&new_smc->sk); /* passive closing */
1112 newsmcsk->sk_state = SMC_CLOSED;
1114 smc_listen_out(new_smc);
1117 /* listen worker: decline and fall back if possible */
1118 static void smc_listen_decline(struct smc_sock *new_smc, int reason_code,
1121 /* RDMA setup failed, switch back to TCP */
1122 if (local_contact == SMC_FIRST_CONTACT)
1123 smc_lgr_cleanup_early(&new_smc->conn);
1125 smc_conn_free(&new_smc->conn);
1126 if (reason_code < 0) { /* error, no fallback possible */
1127 smc_listen_out_err(new_smc);
1130 smc_switch_to_fallback(new_smc);
1131 new_smc->fallback_rsn = reason_code;
1132 if (reason_code && reason_code != SMC_CLC_DECL_PEERDECL) {
1133 if (smc_clc_send_decline(new_smc, reason_code) < 0) {
1134 smc_listen_out_err(new_smc);
1138 smc_listen_out_connected(new_smc);
1141 /* listen worker: check prefixes */
1142 static int smc_listen_prfx_check(struct smc_sock *new_smc,
1143 struct smc_clc_msg_proposal *pclc)
1145 struct smc_clc_msg_proposal_prefix *pclc_prfx;
1146 struct socket *newclcsock = new_smc->clcsock;
1148 pclc_prfx = smc_clc_proposal_get_prefix(pclc);
1149 if (smc_clc_prfx_match(newclcsock, pclc_prfx))
1150 return SMC_CLC_DECL_DIFFPREFIX;
1155 /* listen worker: initialize connection and buffers */
1156 static int smc_listen_rdma_init(struct smc_sock *new_smc,
1157 struct smc_init_info *ini)
1161 /* allocate connection / link group */
1162 rc = smc_conn_create(new_smc, ini);
1166 /* create send buffer and rmb */
1167 if (smc_buf_create(new_smc, false))
1168 return SMC_CLC_DECL_MEM;
1173 /* listen worker: initialize connection and buffers for SMC-D */
1174 static int smc_listen_ism_init(struct smc_sock *new_smc,
1175 struct smc_clc_msg_proposal *pclc,
1176 struct smc_init_info *ini)
1178 struct smc_clc_msg_smcd *pclc_smcd;
1181 pclc_smcd = smc_get_clc_msg_smcd(pclc);
1182 ini->ism_gid = pclc_smcd->gid;
1183 rc = smc_conn_create(new_smc, ini);
1187 /* Check if peer can be reached via ISM device */
1188 if (smc_ism_cantalk(new_smc->conn.lgr->peer_gid,
1189 new_smc->conn.lgr->vlan_id,
1190 new_smc->conn.lgr->smcd)) {
1191 if (ini->cln_first_contact == SMC_FIRST_CONTACT)
1192 smc_lgr_cleanup_early(&new_smc->conn);
1194 smc_conn_free(&new_smc->conn);
1195 return SMC_CLC_DECL_SMCDNOTALK;
1198 /* Create send and receive buffers */
1199 if (smc_buf_create(new_smc, true)) {
1200 if (ini->cln_first_contact == SMC_FIRST_CONTACT)
1201 smc_lgr_cleanup_early(&new_smc->conn);
1203 smc_conn_free(&new_smc->conn);
1204 return SMC_CLC_DECL_MEM;
1210 /* listen worker: register buffers */
1211 static int smc_listen_rdma_reg(struct smc_sock *new_smc, int local_contact)
1213 struct smc_connection *conn = &new_smc->conn;
1215 if (local_contact != SMC_FIRST_CONTACT) {
1216 if (smcr_lgr_reg_rmbs(conn->lnk, conn->rmb_desc))
1217 return SMC_CLC_DECL_ERR_REGRMB;
1219 smc_rmb_sync_sg_for_device(&new_smc->conn);
1224 /* listen worker: finish RDMA setup */
1225 static int smc_listen_rdma_finish(struct smc_sock *new_smc,
1226 struct smc_clc_msg_accept_confirm *cclc,
1229 struct smc_link *link = new_smc->conn.lnk;
1230 int reason_code = 0;
1232 if (local_contact == SMC_FIRST_CONTACT)
1233 smc_link_save_peer_info(link, cclc);
1235 if (smc_rmb_rtoken_handling(&new_smc->conn, link, cclc)) {
1236 reason_code = SMC_CLC_DECL_ERR_RTOK;
1240 if (local_contact == SMC_FIRST_CONTACT) {
1241 if (smc_ib_ready_link(link)) {
1242 reason_code = SMC_CLC_DECL_ERR_RDYLNK;
1245 /* QP confirmation over RoCE fabric */
1246 smc_llc_flow_initiate(link->lgr, SMC_LLC_FLOW_ADD_LINK);
1247 reason_code = smcr_serv_conf_first_link(new_smc);
1248 smc_llc_flow_stop(link->lgr, &link->lgr->llc_flow_lcl);
1255 smc_listen_decline(new_smc, reason_code, local_contact);
1259 /* setup for RDMA connection of server */
1260 static void smc_listen_work(struct work_struct *work)
1262 struct smc_sock *new_smc = container_of(work, struct smc_sock,
1264 struct socket *newclcsock = new_smc->clcsock;
1265 struct smc_clc_msg_accept_confirm cclc;
1266 struct smc_clc_msg_proposal *pclc;
1267 struct smc_init_info ini = {0};
1268 bool ism_supported = false;
1269 u8 buf[SMC_CLC_MAX_LEN];
1272 if (new_smc->listen_smc->sk.sk_state != SMC_LISTEN)
1273 return smc_listen_out_err(new_smc);
1275 if (new_smc->use_fallback) {
1276 smc_listen_out_connected(new_smc);
1280 /* check if peer is smc capable */
1281 if (!tcp_sk(newclcsock->sk)->syn_smc) {
1282 smc_switch_to_fallback(new_smc);
1283 new_smc->fallback_rsn = SMC_CLC_DECL_PEERNOSMC;
1284 smc_listen_out_connected(new_smc);
1288 /* do inband token exchange -
1289 * wait for and receive SMC Proposal CLC message
1291 pclc = (struct smc_clc_msg_proposal *)&buf;
1292 rc = smc_clc_wait_msg(new_smc, pclc, SMC_CLC_MAX_LEN,
1293 SMC_CLC_PROPOSAL, CLC_WAIT_TIME);
1297 /* IPSec connections opt out of SMC-R optimizations */
1298 if (using_ipsec(new_smc)) {
1299 rc = SMC_CLC_DECL_IPSEC;
1303 /* check for matching IP prefix and subnet length */
1304 rc = smc_listen_prfx_check(new_smc, pclc);
1308 /* get vlan id from IP device */
1309 if (smc_vlan_by_tcpsk(new_smc->clcsock, &ini)) {
1310 rc = SMC_CLC_DECL_GETVLANERR;
1314 mutex_lock(&smc_server_lgr_pending);
1315 smc_close_init(new_smc);
1316 smc_rx_init(new_smc);
1317 smc_tx_init(new_smc);
1319 /* check if ISM is available */
1320 if (pclc->hdr.path == SMC_TYPE_D || pclc->hdr.path == SMC_TYPE_B) {
1321 ini.is_smcd = true; /* prepare ISM check */
1322 rc = smc_find_ism_device(new_smc, &ini);
1324 rc = smc_listen_ism_init(new_smc, pclc, &ini);
1326 ism_supported = true;
1327 else if (pclc->hdr.path == SMC_TYPE_D)
1328 goto out_unlock; /* skip RDMA and decline */
1331 /* check if RDMA is available */
1332 if (!ism_supported) { /* SMC_TYPE_R or SMC_TYPE_B */
1333 /* prepare RDMA check */
1334 ini.is_smcd = false;
1336 ini.ib_lcl = &pclc->lcl;
1337 rc = smc_find_rdma_device(new_smc, &ini);
1339 /* no RDMA device found */
1340 if (pclc->hdr.path == SMC_TYPE_B)
1341 /* neither ISM nor RDMA device found */
1342 rc = SMC_CLC_DECL_NOSMCDEV;
1345 rc = smc_listen_rdma_init(new_smc, &ini);
1348 rc = smc_listen_rdma_reg(new_smc, ini.cln_first_contact);
1353 /* send SMC Accept CLC message */
1354 rc = smc_clc_send_accept(new_smc, ini.cln_first_contact);
1358 /* SMC-D does not need this lock any more */
1360 mutex_unlock(&smc_server_lgr_pending);
1362 /* receive SMC Confirm CLC message */
1363 rc = smc_clc_wait_msg(new_smc, &cclc, sizeof(cclc),
1364 SMC_CLC_CONFIRM, CLC_WAIT_TIME);
1372 if (!ism_supported) {
1373 rc = smc_listen_rdma_finish(new_smc, &cclc,
1374 ini.cln_first_contact);
1375 mutex_unlock(&smc_server_lgr_pending);
1379 smc_conn_save_peer_info(new_smc, &cclc);
1380 smc_listen_out_connected(new_smc);
1384 mutex_unlock(&smc_server_lgr_pending);
1386 smc_listen_decline(new_smc, rc, ini.cln_first_contact);
1389 static void smc_tcp_listen_work(struct work_struct *work)
1391 struct smc_sock *lsmc = container_of(work, struct smc_sock,
1393 struct sock *lsk = &lsmc->sk;
1394 struct smc_sock *new_smc;
1398 while (lsk->sk_state == SMC_LISTEN) {
1399 rc = smc_clcsock_accept(lsmc, &new_smc);
1405 new_smc->listen_smc = lsmc;
1406 new_smc->use_fallback = lsmc->use_fallback;
1407 new_smc->fallback_rsn = lsmc->fallback_rsn;
1408 sock_hold(lsk); /* sock_put in smc_listen_work */
1409 INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
1410 smc_copy_sock_settings_to_smc(new_smc);
1411 new_smc->sk.sk_sndbuf = lsmc->sk.sk_sndbuf;
1412 new_smc->sk.sk_rcvbuf = lsmc->sk.sk_rcvbuf;
1413 sock_hold(&new_smc->sk); /* sock_put in passive closing */
1414 if (!schedule_work(&new_smc->smc_listen_work))
1415 sock_put(&new_smc->sk);
1420 sock_put(&lsmc->sk); /* sock_hold in smc_listen */
1423 static int smc_listen(struct socket *sock, int backlog)
1425 struct sock *sk = sock->sk;
1426 struct smc_sock *smc;
1433 if ((sk->sk_state != SMC_INIT && sk->sk_state != SMC_LISTEN) ||
1434 smc->connect_nonblock)
1438 if (sk->sk_state == SMC_LISTEN) {
1439 sk->sk_max_ack_backlog = backlog;
1442 /* some socket options are handled in core, so we could not apply
1443 * them to the clc socket -- copy smc socket options to clc socket
1445 smc_copy_sock_settings_to_clc(smc);
1446 if (!smc->use_fallback)
1447 tcp_sk(smc->clcsock->sk)->syn_smc = 1;
1449 rc = kernel_listen(smc->clcsock, backlog);
1452 sk->sk_max_ack_backlog = backlog;
1453 sk->sk_ack_backlog = 0;
1454 sk->sk_state = SMC_LISTEN;
1455 sock_hold(sk); /* sock_hold in tcp_listen_worker */
1456 if (!schedule_work(&smc->tcp_listen_work))
1464 static int smc_accept(struct socket *sock, struct socket *new_sock,
1465 int flags, bool kern)
1467 struct sock *sk = sock->sk, *nsk;
1468 DECLARE_WAITQUEUE(wait, current);
1469 struct smc_sock *lsmc;
1474 sock_hold(sk); /* sock_put below */
1477 if (lsmc->sk.sk_state != SMC_LISTEN) {
1483 /* Wait for an incoming connection */
1484 timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
1485 add_wait_queue_exclusive(sk_sleep(sk), &wait);
1486 while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
1487 set_current_state(TASK_INTERRUPTIBLE);
1493 timeo = schedule_timeout(timeo);
1494 /* wakeup by sk_data_ready in smc_listen_work() */
1495 sched_annotate_sleep();
1497 if (signal_pending(current)) {
1498 rc = sock_intr_errno(timeo);
1502 set_current_state(TASK_RUNNING);
1503 remove_wait_queue(sk_sleep(sk), &wait);
1506 rc = sock_error(nsk);
1511 if (lsmc->sockopt_defer_accept && !(flags & O_NONBLOCK)) {
1512 /* wait till data arrives on the socket */
1513 timeo = msecs_to_jiffies(lsmc->sockopt_defer_accept *
1515 if (smc_sk(nsk)->use_fallback) {
1516 struct sock *clcsk = smc_sk(nsk)->clcsock->sk;
1519 if (skb_queue_empty(&clcsk->sk_receive_queue))
1520 sk_wait_data(clcsk, &timeo, NULL);
1521 release_sock(clcsk);
1522 } else if (!atomic_read(&smc_sk(nsk)->conn.bytes_to_rcv)) {
1524 smc_rx_wait(smc_sk(nsk), &timeo, smc_rx_data_available);
1530 sock_put(sk); /* sock_hold above */
1534 static int smc_getname(struct socket *sock, struct sockaddr *addr,
1537 struct smc_sock *smc;
1539 if (peer && (sock->sk->sk_state != SMC_ACTIVE) &&
1540 (sock->sk->sk_state != SMC_APPCLOSEWAIT1))
1543 smc = smc_sk(sock->sk);
1545 return smc->clcsock->ops->getname(smc->clcsock, addr, peer);
1548 static int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
1550 struct sock *sk = sock->sk;
1551 struct smc_sock *smc;
1556 if ((sk->sk_state != SMC_ACTIVE) &&
1557 (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1558 (sk->sk_state != SMC_INIT))
1561 if (msg->msg_flags & MSG_FASTOPEN) {
1562 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
1563 smc_switch_to_fallback(smc);
1564 smc->fallback_rsn = SMC_CLC_DECL_OPTUNSUPP;
1571 if (smc->use_fallback)
1572 rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
1574 rc = smc_tx_sendmsg(smc, msg, len);
1580 static int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
1583 struct sock *sk = sock->sk;
1584 struct smc_sock *smc;
1589 if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
1590 /* socket was connected before, no more data to read */
1594 if ((sk->sk_state == SMC_INIT) ||
1595 (sk->sk_state == SMC_LISTEN) ||
1596 (sk->sk_state == SMC_CLOSED))
1599 if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
1604 if (smc->use_fallback) {
1605 rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
1607 msg->msg_namelen = 0;
1608 rc = smc_rx_recvmsg(smc, msg, NULL, len, flags);
1616 static __poll_t smc_accept_poll(struct sock *parent)
1618 struct smc_sock *isk = smc_sk(parent);
1621 spin_lock(&isk->accept_q_lock);
1622 if (!list_empty(&isk->accept_q))
1623 mask = EPOLLIN | EPOLLRDNORM;
1624 spin_unlock(&isk->accept_q_lock);
1629 static __poll_t smc_poll(struct file *file, struct socket *sock,
1632 struct sock *sk = sock->sk;
1633 struct smc_sock *smc;
1639 smc = smc_sk(sock->sk);
1640 if (smc->use_fallback) {
1641 /* delegate to CLC child sock */
1642 mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
1643 sk->sk_err = smc->clcsock->sk->sk_err;
1645 if (sk->sk_state != SMC_CLOSED)
1646 sock_poll_wait(file, sock, wait);
1649 if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
1650 (sk->sk_state == SMC_CLOSED))
1652 if (sk->sk_state == SMC_LISTEN) {
1653 /* woken up by sk_data_ready in smc_listen_work() */
1654 mask |= smc_accept_poll(sk);
1655 } else if (smc->use_fallback) { /* as result of connect_work()*/
1656 mask |= smc->clcsock->ops->poll(file, smc->clcsock,
1658 sk->sk_err = smc->clcsock->sk->sk_err;
1660 if ((sk->sk_state != SMC_INIT &&
1661 atomic_read(&smc->conn.sndbuf_space)) ||
1662 sk->sk_shutdown & SEND_SHUTDOWN) {
1663 mask |= EPOLLOUT | EPOLLWRNORM;
1665 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1666 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1668 if (atomic_read(&smc->conn.bytes_to_rcv))
1669 mask |= EPOLLIN | EPOLLRDNORM;
1670 if (sk->sk_shutdown & RCV_SHUTDOWN)
1671 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
1672 if (sk->sk_state == SMC_APPCLOSEWAIT1)
1674 if (smc->conn.urg_state == SMC_URG_VALID)
1682 static int smc_shutdown(struct socket *sock, int how)
1684 struct sock *sk = sock->sk;
1685 struct smc_sock *smc;
1691 if ((how < SHUT_RD) || (how > SHUT_RDWR))
1697 if ((sk->sk_state != SMC_ACTIVE) &&
1698 (sk->sk_state != SMC_PEERCLOSEWAIT1) &&
1699 (sk->sk_state != SMC_PEERCLOSEWAIT2) &&
1700 (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1701 (sk->sk_state != SMC_APPCLOSEWAIT2) &&
1702 (sk->sk_state != SMC_APPFINCLOSEWAIT))
1704 if (smc->use_fallback) {
1705 rc = kernel_sock_shutdown(smc->clcsock, how);
1706 sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
1707 if (sk->sk_shutdown == SHUTDOWN_MASK)
1708 sk->sk_state = SMC_CLOSED;
1712 case SHUT_RDWR: /* shutdown in both directions */
1713 rc = smc_close_active(smc);
1716 rc = smc_close_shutdown_write(smc);
1720 /* nothing more to do because peer is not involved */
1724 rc1 = kernel_sock_shutdown(smc->clcsock, how);
1725 /* map sock_shutdown_cmd constants to sk_shutdown value range */
1726 sk->sk_shutdown |= how + 1;
1730 return rc ? rc : rc1;
1733 static int smc_setsockopt(struct socket *sock, int level, int optname,
1734 char __user *optval, unsigned int optlen)
1736 struct sock *sk = sock->sk;
1737 struct smc_sock *smc;
1742 /* generic setsockopts reaching us here always apply to the
1745 rc = smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
1747 if (smc->clcsock->sk->sk_err) {
1748 sk->sk_err = smc->clcsock->sk->sk_err;
1749 sk->sk_error_report(sk);
1752 if (optlen < sizeof(int))
1754 if (get_user(val, (int __user *)optval))
1758 if (rc || smc->use_fallback)
1763 case TCP_FASTOPEN_CONNECT:
1764 case TCP_FASTOPEN_KEY:
1765 case TCP_FASTOPEN_NO_COOKIE:
1766 /* option not supported by SMC */
1767 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
1768 smc_switch_to_fallback(smc);
1769 smc->fallback_rsn = SMC_CLC_DECL_OPTUNSUPP;
1775 if (sk->sk_state != SMC_INIT &&
1776 sk->sk_state != SMC_LISTEN &&
1777 sk->sk_state != SMC_CLOSED) {
1779 mod_delayed_work(system_wq, &smc->conn.tx_work,
1784 if (sk->sk_state != SMC_INIT &&
1785 sk->sk_state != SMC_LISTEN &&
1786 sk->sk_state != SMC_CLOSED) {
1788 mod_delayed_work(system_wq, &smc->conn.tx_work,
1792 case TCP_DEFER_ACCEPT:
1793 smc->sockopt_defer_accept = val;
1804 static int smc_getsockopt(struct socket *sock, int level, int optname,
1805 char __user *optval, int __user *optlen)
1807 struct smc_sock *smc;
1809 smc = smc_sk(sock->sk);
1810 /* socket options apply to the CLC socket */
1811 return smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
1815 static int smc_ioctl(struct socket *sock, unsigned int cmd,
1818 union smc_host_cursor cons, urg;
1819 struct smc_connection *conn;
1820 struct smc_sock *smc;
1823 smc = smc_sk(sock->sk);
1825 lock_sock(&smc->sk);
1826 if (smc->use_fallback) {
1827 if (!smc->clcsock) {
1828 release_sock(&smc->sk);
1831 answ = smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
1832 release_sock(&smc->sk);
1836 case SIOCINQ: /* same as FIONREAD */
1837 if (smc->sk.sk_state == SMC_LISTEN) {
1838 release_sock(&smc->sk);
1841 if (smc->sk.sk_state == SMC_INIT ||
1842 smc->sk.sk_state == SMC_CLOSED)
1845 answ = atomic_read(&smc->conn.bytes_to_rcv);
1848 /* output queue size (not send + not acked) */
1849 if (smc->sk.sk_state == SMC_LISTEN) {
1850 release_sock(&smc->sk);
1853 if (smc->sk.sk_state == SMC_INIT ||
1854 smc->sk.sk_state == SMC_CLOSED)
1857 answ = smc->conn.sndbuf_desc->len -
1858 atomic_read(&smc->conn.sndbuf_space);
1861 /* output queue size (not send only) */
1862 if (smc->sk.sk_state == SMC_LISTEN) {
1863 release_sock(&smc->sk);
1866 if (smc->sk.sk_state == SMC_INIT ||
1867 smc->sk.sk_state == SMC_CLOSED)
1870 answ = smc_tx_prepared_sends(&smc->conn);
1873 if (smc->sk.sk_state == SMC_LISTEN) {
1874 release_sock(&smc->sk);
1877 if (smc->sk.sk_state == SMC_INIT ||
1878 smc->sk.sk_state == SMC_CLOSED) {
1881 smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
1882 smc_curs_copy(&urg, &conn->urg_curs, conn);
1883 answ = smc_curs_diff(conn->rmb_desc->len,
1888 release_sock(&smc->sk);
1889 return -ENOIOCTLCMD;
1891 release_sock(&smc->sk);
1893 return put_user(answ, (int __user *)arg);
1896 static ssize_t smc_sendpage(struct socket *sock, struct page *page,
1897 int offset, size_t size, int flags)
1899 struct sock *sk = sock->sk;
1900 struct smc_sock *smc;
1905 if (sk->sk_state != SMC_ACTIVE) {
1910 if (smc->use_fallback)
1911 rc = kernel_sendpage(smc->clcsock, page, offset,
1914 rc = sock_no_sendpage(sock, page, offset, size, flags);
1920 /* Map the affected portions of the rmbe into an spd, note the number of bytes
1921 * to splice in conn->splice_pending, and press 'go'. Delays consumer cursor
1922 * updates till whenever a respective page has been fully processed.
1923 * Note that subsequent recv() calls have to wait till all splice() processing
1926 static ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
1927 struct pipe_inode_info *pipe, size_t len,
1930 struct sock *sk = sock->sk;
1931 struct smc_sock *smc;
1936 if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
1937 /* socket was connected before, no more data to read */
1941 if (sk->sk_state == SMC_INIT ||
1942 sk->sk_state == SMC_LISTEN ||
1943 sk->sk_state == SMC_CLOSED)
1946 if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
1951 if (smc->use_fallback) {
1952 rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
1959 if (flags & SPLICE_F_NONBLOCK)
1960 flags = MSG_DONTWAIT;
1963 rc = smc_rx_recvmsg(smc, NULL, pipe, len, flags);
1971 /* must look like tcp */
1972 static const struct proto_ops smc_sock_ops = {
1974 .owner = THIS_MODULE,
1975 .release = smc_release,
1977 .connect = smc_connect,
1978 .socketpair = sock_no_socketpair,
1979 .accept = smc_accept,
1980 .getname = smc_getname,
1983 .listen = smc_listen,
1984 .shutdown = smc_shutdown,
1985 .setsockopt = smc_setsockopt,
1986 .getsockopt = smc_getsockopt,
1987 .sendmsg = smc_sendmsg,
1988 .recvmsg = smc_recvmsg,
1989 .mmap = sock_no_mmap,
1990 .sendpage = smc_sendpage,
1991 .splice_read = smc_splice_read,
1994 static int smc_create(struct net *net, struct socket *sock, int protocol,
1997 int family = (protocol == SMCPROTO_SMC6) ? PF_INET6 : PF_INET;
1998 struct smc_sock *smc;
2002 rc = -ESOCKTNOSUPPORT;
2003 if (sock->type != SOCK_STREAM)
2006 rc = -EPROTONOSUPPORT;
2007 if (protocol != SMCPROTO_SMC && protocol != SMCPROTO_SMC6)
2011 sock->ops = &smc_sock_ops;
2012 sk = smc_sock_alloc(net, sock, protocol);
2016 /* create internal TCP socket for CLC handshake and fallback */
2018 smc->use_fallback = false; /* assume rdma capability first */
2019 smc->fallback_rsn = 0;
2020 rc = sock_create_kern(net, family, SOCK_STREAM, IPPROTO_TCP,
2023 sk_common_release(sk);
2026 smc->sk.sk_sndbuf = max(smc->clcsock->sk->sk_sndbuf, SMC_BUF_MIN_SIZE);
2027 smc->sk.sk_rcvbuf = max(smc->clcsock->sk->sk_rcvbuf, SMC_BUF_MIN_SIZE);
2033 static const struct net_proto_family smc_sock_family_ops = {
2035 .owner = THIS_MODULE,
2036 .create = smc_create,
2039 unsigned int smc_net_id;
2041 static __net_init int smc_net_init(struct net *net)
2043 return smc_pnet_net_init(net);
2046 static void __net_exit smc_net_exit(struct net *net)
2048 smc_pnet_net_exit(net);
2051 static struct pernet_operations smc_net_ops = {
2052 .init = smc_net_init,
2053 .exit = smc_net_exit,
2055 .size = sizeof(struct smc_net),
2058 static int __init smc_init(void)
2062 rc = register_pernet_subsys(&smc_net_ops);
2066 rc = smc_pnet_init();
2068 goto out_pernet_subsys;
2070 rc = smc_core_init();
2072 pr_err("%s: smc_core_init fails with %d\n", __func__, rc);
2076 rc = smc_llc_init();
2078 pr_err("%s: smc_llc_init fails with %d\n", __func__, rc);
2082 rc = smc_cdc_init();
2084 pr_err("%s: smc_cdc_init fails with %d\n", __func__, rc);
2088 rc = proto_register(&smc_proto, 1);
2090 pr_err("%s: proto_register(v4) fails with %d\n", __func__, rc);
2094 rc = proto_register(&smc_proto6, 1);
2096 pr_err("%s: proto_register(v6) fails with %d\n", __func__, rc);
2100 rc = sock_register(&smc_sock_family_ops);
2102 pr_err("%s: sock_register fails with %d\n", __func__, rc);
2105 INIT_HLIST_HEAD(&smc_v4_hashinfo.ht);
2106 INIT_HLIST_HEAD(&smc_v6_hashinfo.ht);
2108 rc = smc_ib_register_client();
2110 pr_err("%s: ib_register fails with %d\n", __func__, rc);
2114 static_branch_enable(&tcp_have_smc);
2118 sock_unregister(PF_SMC);
2120 proto_unregister(&smc_proto6);
2122 proto_unregister(&smc_proto);
2128 unregister_pernet_subsys(&smc_net_ops);
2133 static void __exit smc_exit(void)
2135 static_branch_disable(&tcp_have_smc);
2136 sock_unregister(PF_SMC);
2138 smc_ib_unregister_client();
2139 proto_unregister(&smc_proto6);
2140 proto_unregister(&smc_proto);
2142 unregister_pernet_subsys(&smc_net_ops);
2146 module_init(smc_init);
2147 module_exit(smc_exit);
2149 MODULE_AUTHOR("Ursula Braun <ubraun@linux.vnet.ibm.com>");
2150 MODULE_DESCRIPTION("smc socket address family");
2151 MODULE_LICENSE("GPL");
2152 MODULE_ALIAS_NETPROTO(PF_SMC);