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
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 if (smc->clcsock->file) { /* non-accepted sockets have no file yet */
130 smc->clcsock->file->private_data = smc->sk.sk_socket;
131 smc->clcsock->file = NULL;
135 static int __smc_release(struct smc_sock *smc)
137 struct sock *sk = &smc->sk;
140 if (!smc->use_fallback) {
141 rc = smc_close_active(smc);
142 sock_set_flag(sk, SOCK_DEAD);
143 sk->sk_shutdown |= SHUTDOWN_MASK;
145 if (sk->sk_state != SMC_LISTEN && sk->sk_state != SMC_INIT)
146 sock_put(sk); /* passive closing */
147 if (sk->sk_state == SMC_LISTEN) {
148 /* wake up clcsock accept */
149 rc = kernel_sock_shutdown(smc->clcsock, SHUT_RDWR);
151 sk->sk_state = SMC_CLOSED;
152 sk->sk_state_change(sk);
153 smc_restore_fallback_changes(smc);
156 sk->sk_prot->unhash(sk);
158 if (sk->sk_state == SMC_CLOSED) {
161 smc_clcsock_release(smc);
164 if (!smc->use_fallback)
165 smc_conn_free(&smc->conn);
171 static int smc_release(struct socket *sock)
173 struct sock *sk = sock->sk;
174 struct smc_sock *smc;
180 sock_hold(sk); /* sock_put below */
183 /* cleanup for a dangling non-blocking connect */
184 if (smc->connect_nonblock && sk->sk_state == SMC_INIT)
185 tcp_abort(smc->clcsock->sk, ECONNABORTED);
186 flush_work(&smc->connect_work);
188 if (sk->sk_state == SMC_LISTEN)
189 /* smc_close_non_accepted() is called and acquires
190 * sock lock for child sockets again
192 lock_sock_nested(sk, SINGLE_DEPTH_NESTING);
196 rc = __smc_release(smc);
203 sock_put(sk); /* sock_hold above */
204 sock_put(sk); /* final sock_put */
209 static void smc_destruct(struct sock *sk)
211 if (sk->sk_state != SMC_CLOSED)
213 if (!sock_flag(sk, SOCK_DEAD))
216 sk_refcnt_debug_dec(sk);
219 static struct sock *smc_sock_alloc(struct net *net, struct socket *sock,
222 struct smc_sock *smc;
226 prot = (protocol == SMCPROTO_SMC6) ? &smc_proto6 : &smc_proto;
227 sk = sk_alloc(net, PF_SMC, GFP_KERNEL, prot, 0);
231 sock_init_data(sock, sk); /* sets sk_refcnt to 1 */
232 sk->sk_state = SMC_INIT;
233 sk->sk_destruct = smc_destruct;
234 sk->sk_protocol = protocol;
236 INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
237 INIT_WORK(&smc->connect_work, smc_connect_work);
238 INIT_DELAYED_WORK(&smc->conn.tx_work, smc_tx_work);
239 INIT_LIST_HEAD(&smc->accept_q);
240 spin_lock_init(&smc->accept_q_lock);
241 spin_lock_init(&smc->conn.send_lock);
242 sk->sk_prot->hash(sk);
243 sk_refcnt_debug_inc(sk);
244 mutex_init(&smc->clcsock_release_lock);
249 static int smc_bind(struct socket *sock, struct sockaddr *uaddr,
252 struct sockaddr_in *addr = (struct sockaddr_in *)uaddr;
253 struct sock *sk = sock->sk;
254 struct smc_sock *smc;
259 /* replicate tests from inet_bind(), to be safe wrt. future changes */
261 if (addr_len < sizeof(struct sockaddr_in))
265 if (addr->sin_family != AF_INET &&
266 addr->sin_family != AF_INET6 &&
267 addr->sin_family != AF_UNSPEC)
269 /* accept AF_UNSPEC (mapped to AF_INET) only if s_addr is INADDR_ANY */
270 if (addr->sin_family == AF_UNSPEC &&
271 addr->sin_addr.s_addr != htonl(INADDR_ANY))
276 /* Check if socket is already active */
278 if (sk->sk_state != SMC_INIT || smc->connect_nonblock)
281 smc->clcsock->sk->sk_reuse = sk->sk_reuse;
282 rc = kernel_bind(smc->clcsock, uaddr, addr_len);
290 static void smc_copy_sock_settings(struct sock *nsk, struct sock *osk,
293 /* options we don't get control via setsockopt for */
294 nsk->sk_type = osk->sk_type;
295 nsk->sk_sndbuf = osk->sk_sndbuf;
296 nsk->sk_rcvbuf = osk->sk_rcvbuf;
297 nsk->sk_sndtimeo = osk->sk_sndtimeo;
298 nsk->sk_rcvtimeo = osk->sk_rcvtimeo;
299 nsk->sk_mark = osk->sk_mark;
300 nsk->sk_priority = osk->sk_priority;
301 nsk->sk_rcvlowat = osk->sk_rcvlowat;
302 nsk->sk_bound_dev_if = osk->sk_bound_dev_if;
303 nsk->sk_err = osk->sk_err;
305 nsk->sk_flags &= ~mask;
306 nsk->sk_flags |= osk->sk_flags & mask;
309 #define SK_FLAGS_SMC_TO_CLC ((1UL << SOCK_URGINLINE) | \
310 (1UL << SOCK_KEEPOPEN) | \
311 (1UL << SOCK_LINGER) | \
312 (1UL << SOCK_BROADCAST) | \
313 (1UL << SOCK_TIMESTAMP) | \
314 (1UL << SOCK_DBG) | \
315 (1UL << SOCK_RCVTSTAMP) | \
316 (1UL << SOCK_RCVTSTAMPNS) | \
317 (1UL << SOCK_LOCALROUTE) | \
318 (1UL << SOCK_TIMESTAMPING_RX_SOFTWARE) | \
319 (1UL << SOCK_RXQ_OVFL) | \
320 (1UL << SOCK_WIFI_STATUS) | \
321 (1UL << SOCK_NOFCS) | \
322 (1UL << SOCK_FILTER_LOCKED) | \
323 (1UL << SOCK_TSTAMP_NEW))
324 /* copy only relevant settings and flags of SOL_SOCKET level from smc to
325 * clc socket (since smc is not called for these options from net/core)
327 static void smc_copy_sock_settings_to_clc(struct smc_sock *smc)
329 smc_copy_sock_settings(smc->clcsock->sk, &smc->sk, SK_FLAGS_SMC_TO_CLC);
332 #define SK_FLAGS_CLC_TO_SMC ((1UL << SOCK_URGINLINE) | \
333 (1UL << SOCK_KEEPOPEN) | \
334 (1UL << SOCK_LINGER) | \
336 /* copy only settings and flags relevant for smc from clc to smc socket */
337 static void smc_copy_sock_settings_to_smc(struct smc_sock *smc)
339 smc_copy_sock_settings(&smc->sk, smc->clcsock->sk, SK_FLAGS_CLC_TO_SMC);
342 /* register the new rmb on all links */
343 static int smcr_lgr_reg_rmbs(struct smc_link *link,
344 struct smc_buf_desc *rmb_desc)
346 struct smc_link_group *lgr = link->lgr;
349 rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY);
352 /* protect against parallel smc_llc_cli_rkey_exchange() and
353 * parallel smcr_link_reg_rmb()
355 mutex_lock(&lgr->llc_conf_mutex);
356 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
357 if (!smc_link_active(&lgr->lnk[i]))
359 rc = smcr_link_reg_rmb(&lgr->lnk[i], rmb_desc);
364 /* exchange confirm_rkey msg with peer */
365 rc = smc_llc_do_confirm_rkey(link, rmb_desc);
370 rmb_desc->is_conf_rkey = true;
372 mutex_unlock(&lgr->llc_conf_mutex);
373 smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl);
377 static int smcr_clnt_conf_first_link(struct smc_sock *smc)
379 struct smc_link *link = smc->conn.lnk;
380 struct smc_llc_qentry *qentry;
383 /* receive CONFIRM LINK request from server over RoCE fabric */
384 qentry = smc_llc_wait(link->lgr, NULL, SMC_LLC_WAIT_TIME,
385 SMC_LLC_CONFIRM_LINK);
387 struct smc_clc_msg_decline dclc;
389 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
390 SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
391 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
393 smc_llc_save_peer_uid(qentry);
394 rc = smc_llc_eval_conf_link(qentry, SMC_LLC_REQ);
395 smc_llc_flow_qentry_del(&link->lgr->llc_flow_lcl);
397 return SMC_CLC_DECL_RMBE_EC;
399 rc = smc_ib_modify_qp_rts(link);
401 return SMC_CLC_DECL_ERR_RDYLNK;
403 smc_wr_remember_qp_attr(link);
405 if (smcr_link_reg_rmb(link, smc->conn.rmb_desc))
406 return SMC_CLC_DECL_ERR_REGRMB;
408 /* confirm_rkey is implicit on 1st contact */
409 smc->conn.rmb_desc->is_conf_rkey = true;
411 /* send CONFIRM LINK response over RoCE fabric */
412 rc = smc_llc_send_confirm_link(link, SMC_LLC_RESP);
414 return SMC_CLC_DECL_TIMEOUT_CL;
416 smc_llc_link_active(link);
417 smcr_lgr_set_type(link->lgr, SMC_LGR_SINGLE);
419 /* optional 2nd link, receive ADD LINK request from server */
420 qentry = smc_llc_wait(link->lgr, NULL, SMC_LLC_WAIT_TIME,
423 struct smc_clc_msg_decline dclc;
425 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
426 SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
428 rc = 0; /* no DECLINE received, go with one link */
431 smc_llc_flow_qentry_clr(&link->lgr->llc_flow_lcl);
432 smc_llc_cli_add_link(link, qentry);
436 static void smcr_conn_save_peer_info(struct smc_sock *smc,
437 struct smc_clc_msg_accept_confirm *clc)
439 int bufsize = smc_uncompress_bufsize(clc->rmbe_size);
441 smc->conn.peer_rmbe_idx = clc->rmbe_idx;
442 smc->conn.local_tx_ctrl.token = ntohl(clc->rmbe_alert_token);
443 smc->conn.peer_rmbe_size = bufsize;
444 atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
445 smc->conn.tx_off = bufsize * (smc->conn.peer_rmbe_idx - 1);
448 static void smcd_conn_save_peer_info(struct smc_sock *smc,
449 struct smc_clc_msg_accept_confirm *clc)
451 int bufsize = smc_uncompress_bufsize(clc->dmbe_size);
453 smc->conn.peer_rmbe_idx = clc->dmbe_idx;
454 smc->conn.peer_token = clc->token;
455 /* msg header takes up space in the buffer */
456 smc->conn.peer_rmbe_size = bufsize - sizeof(struct smcd_cdc_msg);
457 atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
458 smc->conn.tx_off = bufsize * smc->conn.peer_rmbe_idx;
461 static void smc_conn_save_peer_info(struct smc_sock *smc,
462 struct smc_clc_msg_accept_confirm *clc)
464 if (smc->conn.lgr->is_smcd)
465 smcd_conn_save_peer_info(smc, clc);
467 smcr_conn_save_peer_info(smc, clc);
470 static void smc_link_save_peer_info(struct smc_link *link,
471 struct smc_clc_msg_accept_confirm *clc)
473 link->peer_qpn = ntoh24(clc->qpn);
474 memcpy(link->peer_gid, clc->lcl.gid, SMC_GID_SIZE);
475 memcpy(link->peer_mac, clc->lcl.mac, sizeof(link->peer_mac));
476 link->peer_psn = ntoh24(clc->psn);
477 link->peer_mtu = clc->qp_mtu;
480 static void smc_switch_to_fallback(struct smc_sock *smc)
482 smc->use_fallback = true;
483 if (smc->sk.sk_socket && smc->sk.sk_socket->file) {
484 smc->clcsock->file = smc->sk.sk_socket->file;
485 smc->clcsock->file->private_data = smc->clcsock;
486 smc->clcsock->wq.fasync_list =
487 smc->sk.sk_socket->wq.fasync_list;
491 /* fall back during connect */
492 static int smc_connect_fallback(struct smc_sock *smc, int reason_code)
494 smc_switch_to_fallback(smc);
495 smc->fallback_rsn = reason_code;
496 smc_copy_sock_settings_to_clc(smc);
497 smc->connect_nonblock = 0;
498 if (smc->sk.sk_state == SMC_INIT)
499 smc->sk.sk_state = SMC_ACTIVE;
503 /* decline and fall back during connect */
504 static int smc_connect_decline_fallback(struct smc_sock *smc, int reason_code)
508 if (reason_code < 0) { /* error, fallback is not possible */
509 if (smc->sk.sk_state == SMC_INIT)
510 sock_put(&smc->sk); /* passive closing */
513 if (reason_code != SMC_CLC_DECL_PEERDECL) {
514 rc = smc_clc_send_decline(smc, reason_code);
516 if (smc->sk.sk_state == SMC_INIT)
517 sock_put(&smc->sk); /* passive closing */
521 return smc_connect_fallback(smc, reason_code);
524 /* abort connecting */
525 static int smc_connect_abort(struct smc_sock *smc, int reason_code,
528 bool is_smcd = smc->conn.lgr->is_smcd;
530 if (local_contact == SMC_FIRST_CONTACT)
531 smc_lgr_cleanup_early(&smc->conn);
533 smc_conn_free(&smc->conn);
535 /* there is only one lgr role for SMC-D; use server lock */
536 mutex_unlock(&smc_server_lgr_pending);
538 mutex_unlock(&smc_client_lgr_pending);
540 smc->connect_nonblock = 0;
544 /* check if there is a rdma device available for this connection. */
545 /* called for connect and listen */
546 static int smc_find_rdma_device(struct smc_sock *smc, struct smc_init_info *ini)
548 /* PNET table look up: search active ib_device and port
549 * within same PNETID that also contains the ethernet device
550 * used for the internal TCP socket
552 smc_pnet_find_roce_resource(smc->clcsock->sk, ini);
554 return SMC_CLC_DECL_NOSMCRDEV;
558 /* check if there is an ISM device available for this connection. */
559 /* called for connect and listen */
560 static int smc_find_ism_device(struct smc_sock *smc, struct smc_init_info *ini)
562 /* Find ISM device with same PNETID as connecting interface */
563 smc_pnet_find_ism_resource(smc->clcsock->sk, ini);
565 return SMC_CLC_DECL_NOSMCDDEV;
569 /* Check for VLAN ID and register it on ISM device just for CLC handshake */
570 static int smc_connect_ism_vlan_setup(struct smc_sock *smc,
571 struct smc_init_info *ini)
573 if (ini->vlan_id && smc_ism_get_vlan(ini->ism_dev, ini->vlan_id))
574 return SMC_CLC_DECL_ISMVLANERR;
578 /* cleanup temporary VLAN ID registration used for CLC handshake. If ISM is
579 * used, the VLAN ID will be registered again during the connection setup.
581 static int smc_connect_ism_vlan_cleanup(struct smc_sock *smc, bool is_smcd,
582 struct smc_init_info *ini)
586 if (ini->vlan_id && smc_ism_put_vlan(ini->ism_dev, ini->vlan_id))
587 return SMC_CLC_DECL_CNFERR;
591 /* CLC handshake during connect */
592 static int smc_connect_clc(struct smc_sock *smc, int smc_type,
593 struct smc_clc_msg_accept_confirm *aclc,
594 struct smc_init_info *ini)
598 /* do inband token exchange */
599 rc = smc_clc_send_proposal(smc, smc_type, ini);
602 /* receive SMC Accept CLC message */
603 return smc_clc_wait_msg(smc, aclc, sizeof(*aclc), SMC_CLC_ACCEPT,
607 /* setup for RDMA connection of client */
608 static int smc_connect_rdma(struct smc_sock *smc,
609 struct smc_clc_msg_accept_confirm *aclc,
610 struct smc_init_info *ini)
612 int i, reason_code = 0;
613 struct smc_link *link;
615 ini->is_smcd = false;
616 ini->ib_lcl = &aclc->lcl;
617 ini->ib_clcqpn = ntoh24(aclc->qpn);
618 ini->srv_first_contact = aclc->hdr.flag;
620 mutex_lock(&smc_client_lgr_pending);
621 reason_code = smc_conn_create(smc, ini);
623 mutex_unlock(&smc_client_lgr_pending);
627 smc_conn_save_peer_info(smc, aclc);
629 if (ini->cln_first_contact == SMC_FIRST_CONTACT) {
630 link = smc->conn.lnk;
632 /* set link that was assigned by server */
634 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
635 struct smc_link *l = &smc->conn.lgr->lnk[i];
637 if (l->peer_qpn == ntoh24(aclc->qpn) &&
638 !memcmp(l->peer_gid, &aclc->lcl.gid, SMC_GID_SIZE) &&
639 !memcmp(l->peer_mac, &aclc->lcl.mac, sizeof(l->peer_mac))) {
645 return smc_connect_abort(smc, SMC_CLC_DECL_NOSRVLINK,
646 ini->cln_first_contact);
647 smc->conn.lnk = link;
650 /* create send buffer and rmb */
651 if (smc_buf_create(smc, false))
652 return smc_connect_abort(smc, SMC_CLC_DECL_MEM,
653 ini->cln_first_contact);
655 if (ini->cln_first_contact == SMC_FIRST_CONTACT)
656 smc_link_save_peer_info(link, aclc);
658 if (smc_rmb_rtoken_handling(&smc->conn, link, aclc))
659 return smc_connect_abort(smc, SMC_CLC_DECL_ERR_RTOK,
660 ini->cln_first_contact);
665 if (ini->cln_first_contact == SMC_FIRST_CONTACT) {
666 if (smc_ib_ready_link(link))
667 return smc_connect_abort(smc, SMC_CLC_DECL_ERR_RDYLNK,
668 ini->cln_first_contact);
670 if (smcr_lgr_reg_rmbs(link, smc->conn.rmb_desc))
671 return smc_connect_abort(smc, SMC_CLC_DECL_ERR_REGRMB,
672 ini->cln_first_contact);
674 smc_rmb_sync_sg_for_device(&smc->conn);
676 reason_code = smc_clc_send_confirm(smc);
678 return smc_connect_abort(smc, reason_code,
679 ini->cln_first_contact);
683 if (ini->cln_first_contact == SMC_FIRST_CONTACT) {
684 /* QP confirmation over RoCE fabric */
685 smc_llc_flow_initiate(link->lgr, SMC_LLC_FLOW_ADD_LINK);
686 reason_code = smcr_clnt_conf_first_link(smc);
687 smc_llc_flow_stop(link->lgr, &link->lgr->llc_flow_lcl);
689 return smc_connect_abort(smc, reason_code,
690 ini->cln_first_contact);
692 mutex_unlock(&smc_client_lgr_pending);
694 smc_copy_sock_settings_to_clc(smc);
695 smc->connect_nonblock = 0;
696 if (smc->sk.sk_state == SMC_INIT)
697 smc->sk.sk_state = SMC_ACTIVE;
702 /* setup for ISM connection of client */
703 static int smc_connect_ism(struct smc_sock *smc,
704 struct smc_clc_msg_accept_confirm *aclc,
705 struct smc_init_info *ini)
710 ini->ism_gid = aclc->gid;
711 ini->srv_first_contact = aclc->hdr.flag;
713 /* there is only one lgr role for SMC-D; use server lock */
714 mutex_lock(&smc_server_lgr_pending);
715 rc = smc_conn_create(smc, ini);
717 mutex_unlock(&smc_server_lgr_pending);
721 /* Create send and receive buffers */
722 rc = smc_buf_create(smc, true);
724 return smc_connect_abort(smc, (rc == -ENOSPC) ?
725 SMC_CLC_DECL_MAX_DMB :
727 ini->cln_first_contact);
729 smc_conn_save_peer_info(smc, aclc);
734 rc = smc_clc_send_confirm(smc);
736 return smc_connect_abort(smc, rc, ini->cln_first_contact);
737 mutex_unlock(&smc_server_lgr_pending);
739 smc_copy_sock_settings_to_clc(smc);
740 smc->connect_nonblock = 0;
741 if (smc->sk.sk_state == SMC_INIT)
742 smc->sk.sk_state = SMC_ACTIVE;
747 /* perform steps before actually connecting */
748 static int __smc_connect(struct smc_sock *smc)
750 bool ism_supported = false, rdma_supported = false;
751 struct smc_clc_msg_accept_confirm aclc;
752 struct smc_init_info ini = {0};
756 if (smc->use_fallback)
757 return smc_connect_fallback(smc, smc->fallback_rsn);
759 /* if peer has not signalled SMC-capability, fall back */
760 if (!tcp_sk(smc->clcsock->sk)->syn_smc)
761 return smc_connect_fallback(smc, SMC_CLC_DECL_PEERNOSMC);
763 /* IPSec connections opt out of SMC-R optimizations */
764 if (using_ipsec(smc))
765 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_IPSEC);
767 /* get vlan id from IP device */
768 if (smc_vlan_by_tcpsk(smc->clcsock, &ini))
769 return smc_connect_decline_fallback(smc,
770 SMC_CLC_DECL_GETVLANERR);
772 /* check if there is an ism device available */
773 if (!smc_find_ism_device(smc, &ini) &&
774 !smc_connect_ism_vlan_setup(smc, &ini)) {
775 /* ISM is supported for this connection */
776 ism_supported = true;
777 smc_type = SMC_TYPE_D;
780 /* check if there is a rdma device available */
781 if (!smc_find_rdma_device(smc, &ini)) {
782 /* RDMA is supported for this connection */
783 rdma_supported = true;
785 smc_type = SMC_TYPE_B; /* both */
787 smc_type = SMC_TYPE_R; /* only RDMA */
790 /* if neither ISM nor RDMA are supported, fallback */
791 if (!rdma_supported && !ism_supported)
792 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_NOSMCDEV);
794 /* perform CLC handshake */
795 rc = smc_connect_clc(smc, smc_type, &aclc, &ini);
797 smc_connect_ism_vlan_cleanup(smc, ism_supported, &ini);
798 return smc_connect_decline_fallback(smc, rc);
801 /* depending on previous steps, connect using rdma or ism */
802 if (rdma_supported && aclc.hdr.path == SMC_TYPE_R)
803 rc = smc_connect_rdma(smc, &aclc, &ini);
804 else if (ism_supported && aclc.hdr.path == SMC_TYPE_D)
805 rc = smc_connect_ism(smc, &aclc, &ini);
807 rc = SMC_CLC_DECL_MODEUNSUPP;
809 smc_connect_ism_vlan_cleanup(smc, ism_supported, &ini);
810 return smc_connect_decline_fallback(smc, rc);
813 smc_connect_ism_vlan_cleanup(smc, ism_supported, &ini);
817 static void smc_connect_work(struct work_struct *work)
819 struct smc_sock *smc = container_of(work, struct smc_sock,
821 long timeo = smc->sk.sk_sndtimeo;
825 timeo = MAX_SCHEDULE_TIMEOUT;
826 lock_sock(smc->clcsock->sk);
827 if (smc->clcsock->sk->sk_err) {
828 smc->sk.sk_err = smc->clcsock->sk->sk_err;
829 } else if ((1 << smc->clcsock->sk->sk_state) &
830 (TCPF_SYN_SENT | TCP_SYN_RECV)) {
831 rc = sk_stream_wait_connect(smc->clcsock->sk, &timeo);
832 if ((rc == -EPIPE) &&
833 ((1 << smc->clcsock->sk->sk_state) &
834 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)))
837 release_sock(smc->clcsock->sk);
839 if (rc != 0 || smc->sk.sk_err) {
840 smc->sk.sk_state = SMC_CLOSED;
841 if (rc == -EPIPE || rc == -EAGAIN)
842 smc->sk.sk_err = EPIPE;
843 else if (signal_pending(current))
844 smc->sk.sk_err = -sock_intr_errno(timeo);
845 sock_put(&smc->sk); /* passive closing */
849 rc = __smc_connect(smc);
851 smc->sk.sk_err = -rc;
854 if (!sock_flag(&smc->sk, SOCK_DEAD)) {
855 if (smc->sk.sk_err) {
856 smc->sk.sk_state_change(&smc->sk);
857 } else { /* allow polling before and after fallback decision */
858 smc->clcsock->sk->sk_write_space(smc->clcsock->sk);
859 smc->sk.sk_write_space(&smc->sk);
862 release_sock(&smc->sk);
865 static int smc_connect(struct socket *sock, struct sockaddr *addr,
868 struct sock *sk = sock->sk;
869 struct smc_sock *smc;
874 /* separate smc parameter checking to be safe */
875 if (alen < sizeof(addr->sa_family))
877 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
881 switch (sk->sk_state) {
892 smc_copy_sock_settings_to_clc(smc);
893 tcp_sk(smc->clcsock->sk)->syn_smc = 1;
894 if (smc->connect_nonblock) {
898 rc = kernel_connect(smc->clcsock, addr, alen, flags);
899 if (rc && rc != -EINPROGRESS)
902 sock_hold(&smc->sk); /* sock put in passive closing */
903 if (smc->use_fallback)
905 if (flags & O_NONBLOCK) {
906 if (schedule_work(&smc->connect_work))
907 smc->connect_nonblock = 1;
910 rc = __smc_connect(smc);
914 rc = 0; /* success cases including fallback */
923 static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
925 struct socket *new_clcsock = NULL;
926 struct sock *lsk = &lsmc->sk;
931 new_sk = smc_sock_alloc(sock_net(lsk), NULL, lsk->sk_protocol);
934 lsk->sk_err = ENOMEM;
939 *new_smc = smc_sk(new_sk);
941 mutex_lock(&lsmc->clcsock_release_lock);
943 rc = kernel_accept(lsmc->clcsock, &new_clcsock, 0);
944 mutex_unlock(&lsmc->clcsock_release_lock);
948 if (rc < 0 || lsk->sk_state == SMC_CLOSED) {
949 new_sk->sk_prot->unhash(new_sk);
951 sock_release(new_clcsock);
952 new_sk->sk_state = SMC_CLOSED;
953 sock_set_flag(new_sk, SOCK_DEAD);
954 sock_put(new_sk); /* final */
959 (*new_smc)->clcsock = new_clcsock;
964 /* add a just created sock to the accept queue of the listen sock as
965 * candidate for a following socket accept call from user space
967 static void smc_accept_enqueue(struct sock *parent, struct sock *sk)
969 struct smc_sock *par = smc_sk(parent);
971 sock_hold(sk); /* sock_put in smc_accept_unlink () */
972 spin_lock(&par->accept_q_lock);
973 list_add_tail(&smc_sk(sk)->accept_q, &par->accept_q);
974 spin_unlock(&par->accept_q_lock);
975 sk_acceptq_added(parent);
978 /* remove a socket from the accept queue of its parental listening socket */
979 static void smc_accept_unlink(struct sock *sk)
981 struct smc_sock *par = smc_sk(sk)->listen_smc;
983 spin_lock(&par->accept_q_lock);
984 list_del_init(&smc_sk(sk)->accept_q);
985 spin_unlock(&par->accept_q_lock);
986 sk_acceptq_removed(&smc_sk(sk)->listen_smc->sk);
987 sock_put(sk); /* sock_hold in smc_accept_enqueue */
990 /* remove a sock from the accept queue to bind it to a new socket created
991 * for a socket accept call from user space
993 struct sock *smc_accept_dequeue(struct sock *parent,
994 struct socket *new_sock)
996 struct smc_sock *isk, *n;
999 list_for_each_entry_safe(isk, n, &smc_sk(parent)->accept_q, accept_q) {
1000 new_sk = (struct sock *)isk;
1002 smc_accept_unlink(new_sk);
1003 if (new_sk->sk_state == SMC_CLOSED) {
1004 new_sk->sk_prot->unhash(new_sk);
1006 sock_release(isk->clcsock);
1007 isk->clcsock = NULL;
1009 sock_put(new_sk); /* final */
1013 sock_graft(new_sk, new_sock);
1014 if (isk->use_fallback) {
1015 smc_sk(new_sk)->clcsock->file = new_sock->file;
1016 isk->clcsock->file->private_data = isk->clcsock;
1024 /* clean up for a created but never accepted sock */
1025 void smc_close_non_accepted(struct sock *sk)
1027 struct smc_sock *smc = smc_sk(sk);
1029 sock_hold(sk); /* sock_put below */
1031 if (!sk->sk_lingertime)
1032 /* wait for peer closing */
1033 sk->sk_lingertime = SMC_MAX_STREAM_WAIT_TIMEOUT;
1036 sock_put(sk); /* sock_hold above */
1037 sock_put(sk); /* final sock_put */
1040 static int smcr_serv_conf_first_link(struct smc_sock *smc)
1042 struct smc_link *link = smc->conn.lnk;
1043 struct smc_llc_qentry *qentry;
1046 if (smcr_link_reg_rmb(link, smc->conn.rmb_desc))
1047 return SMC_CLC_DECL_ERR_REGRMB;
1049 /* send CONFIRM LINK request to client over the RoCE fabric */
1050 rc = smc_llc_send_confirm_link(link, SMC_LLC_REQ);
1052 return SMC_CLC_DECL_TIMEOUT_CL;
1054 /* receive CONFIRM LINK response from client over the RoCE fabric */
1055 qentry = smc_llc_wait(link->lgr, link, SMC_LLC_WAIT_TIME,
1056 SMC_LLC_CONFIRM_LINK);
1058 struct smc_clc_msg_decline dclc;
1060 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
1061 SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
1062 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
1064 smc_llc_save_peer_uid(qentry);
1065 rc = smc_llc_eval_conf_link(qentry, SMC_LLC_RESP);
1066 smc_llc_flow_qentry_del(&link->lgr->llc_flow_lcl);
1068 return SMC_CLC_DECL_RMBE_EC;
1070 /* confirm_rkey is implicit on 1st contact */
1071 smc->conn.rmb_desc->is_conf_rkey = true;
1073 smc_llc_link_active(link);
1074 smcr_lgr_set_type(link->lgr, SMC_LGR_SINGLE);
1076 /* initial contact - try to establish second link */
1077 smc_llc_srv_add_link(link);
1081 /* listen worker: finish */
1082 static void smc_listen_out(struct smc_sock *new_smc)
1084 struct smc_sock *lsmc = new_smc->listen_smc;
1085 struct sock *newsmcsk = &new_smc->sk;
1087 if (lsmc->sk.sk_state == SMC_LISTEN) {
1088 lock_sock_nested(&lsmc->sk, SINGLE_DEPTH_NESTING);
1089 smc_accept_enqueue(&lsmc->sk, newsmcsk);
1090 release_sock(&lsmc->sk);
1091 } else { /* no longer listening */
1092 smc_close_non_accepted(newsmcsk);
1095 /* Wake up accept */
1096 lsmc->sk.sk_data_ready(&lsmc->sk);
1097 sock_put(&lsmc->sk); /* sock_hold in smc_tcp_listen_work */
1100 /* listen worker: finish in state connected */
1101 static void smc_listen_out_connected(struct smc_sock *new_smc)
1103 struct sock *newsmcsk = &new_smc->sk;
1105 sk_refcnt_debug_inc(newsmcsk);
1106 if (newsmcsk->sk_state == SMC_INIT)
1107 newsmcsk->sk_state = SMC_ACTIVE;
1109 smc_listen_out(new_smc);
1112 /* listen worker: finish in error state */
1113 static void smc_listen_out_err(struct smc_sock *new_smc)
1115 struct sock *newsmcsk = &new_smc->sk;
1117 if (newsmcsk->sk_state == SMC_INIT)
1118 sock_put(&new_smc->sk); /* passive closing */
1119 newsmcsk->sk_state = SMC_CLOSED;
1121 smc_listen_out(new_smc);
1124 /* listen worker: decline and fall back if possible */
1125 static void smc_listen_decline(struct smc_sock *new_smc, int reason_code,
1128 /* RDMA setup failed, switch back to TCP */
1129 if (local_contact == SMC_FIRST_CONTACT)
1130 smc_lgr_cleanup_early(&new_smc->conn);
1132 smc_conn_free(&new_smc->conn);
1133 if (reason_code < 0) { /* error, no fallback possible */
1134 smc_listen_out_err(new_smc);
1137 smc_switch_to_fallback(new_smc);
1138 new_smc->fallback_rsn = reason_code;
1139 if (reason_code && reason_code != SMC_CLC_DECL_PEERDECL) {
1140 if (smc_clc_send_decline(new_smc, reason_code) < 0) {
1141 smc_listen_out_err(new_smc);
1145 smc_listen_out_connected(new_smc);
1148 /* listen worker: check prefixes */
1149 static int smc_listen_prfx_check(struct smc_sock *new_smc,
1150 struct smc_clc_msg_proposal *pclc)
1152 struct smc_clc_msg_proposal_prefix *pclc_prfx;
1153 struct socket *newclcsock = new_smc->clcsock;
1155 pclc_prfx = smc_clc_proposal_get_prefix(pclc);
1156 if (smc_clc_prfx_match(newclcsock, pclc_prfx))
1157 return SMC_CLC_DECL_DIFFPREFIX;
1162 /* listen worker: initialize connection and buffers */
1163 static int smc_listen_rdma_init(struct smc_sock *new_smc,
1164 struct smc_init_info *ini)
1168 /* allocate connection / link group */
1169 rc = smc_conn_create(new_smc, ini);
1173 /* create send buffer and rmb */
1174 if (smc_buf_create(new_smc, false))
1175 return SMC_CLC_DECL_MEM;
1180 /* listen worker: initialize connection and buffers for SMC-D */
1181 static int smc_listen_ism_init(struct smc_sock *new_smc,
1182 struct smc_clc_msg_proposal *pclc,
1183 struct smc_init_info *ini)
1185 struct smc_clc_msg_smcd *pclc_smcd;
1188 pclc_smcd = smc_get_clc_msg_smcd(pclc);
1189 ini->ism_gid = pclc_smcd->gid;
1190 rc = smc_conn_create(new_smc, ini);
1194 /* Check if peer can be reached via ISM device */
1195 if (smc_ism_cantalk(new_smc->conn.lgr->peer_gid,
1196 new_smc->conn.lgr->vlan_id,
1197 new_smc->conn.lgr->smcd)) {
1198 if (ini->cln_first_contact == SMC_FIRST_CONTACT)
1199 smc_lgr_cleanup_early(&new_smc->conn);
1201 smc_conn_free(&new_smc->conn);
1202 return SMC_CLC_DECL_SMCDNOTALK;
1205 /* Create send and receive buffers */
1206 rc = smc_buf_create(new_smc, true);
1208 if (ini->cln_first_contact == SMC_FIRST_CONTACT)
1209 smc_lgr_cleanup_early(&new_smc->conn);
1211 smc_conn_free(&new_smc->conn);
1212 return (rc == -ENOSPC) ? SMC_CLC_DECL_MAX_DMB :
1219 /* listen worker: register buffers */
1220 static int smc_listen_rdma_reg(struct smc_sock *new_smc, int local_contact)
1222 struct smc_connection *conn = &new_smc->conn;
1224 if (local_contact != SMC_FIRST_CONTACT) {
1225 if (smcr_lgr_reg_rmbs(conn->lnk, conn->rmb_desc))
1226 return SMC_CLC_DECL_ERR_REGRMB;
1228 smc_rmb_sync_sg_for_device(&new_smc->conn);
1233 /* listen worker: finish RDMA setup */
1234 static int smc_listen_rdma_finish(struct smc_sock *new_smc,
1235 struct smc_clc_msg_accept_confirm *cclc,
1238 struct smc_link *link = new_smc->conn.lnk;
1239 int reason_code = 0;
1241 if (local_contact == SMC_FIRST_CONTACT)
1242 smc_link_save_peer_info(link, cclc);
1244 if (smc_rmb_rtoken_handling(&new_smc->conn, link, cclc)) {
1245 reason_code = SMC_CLC_DECL_ERR_RTOK;
1249 if (local_contact == SMC_FIRST_CONTACT) {
1250 if (smc_ib_ready_link(link)) {
1251 reason_code = SMC_CLC_DECL_ERR_RDYLNK;
1254 /* QP confirmation over RoCE fabric */
1255 smc_llc_flow_initiate(link->lgr, SMC_LLC_FLOW_ADD_LINK);
1256 reason_code = smcr_serv_conf_first_link(new_smc);
1257 smc_llc_flow_stop(link->lgr, &link->lgr->llc_flow_lcl);
1264 smc_listen_decline(new_smc, reason_code, local_contact);
1268 /* setup for RDMA connection of server */
1269 static void smc_listen_work(struct work_struct *work)
1271 struct smc_sock *new_smc = container_of(work, struct smc_sock,
1273 struct socket *newclcsock = new_smc->clcsock;
1274 struct smc_clc_msg_accept_confirm cclc;
1275 struct smc_clc_msg_proposal *pclc;
1276 struct smc_init_info ini = {0};
1277 bool ism_supported = false;
1278 u8 buf[SMC_CLC_MAX_LEN];
1281 if (new_smc->listen_smc->sk.sk_state != SMC_LISTEN)
1282 return smc_listen_out_err(new_smc);
1284 if (new_smc->use_fallback) {
1285 smc_listen_out_connected(new_smc);
1289 /* check if peer is smc capable */
1290 if (!tcp_sk(newclcsock->sk)->syn_smc) {
1291 smc_switch_to_fallback(new_smc);
1292 new_smc->fallback_rsn = SMC_CLC_DECL_PEERNOSMC;
1293 smc_listen_out_connected(new_smc);
1297 /* do inband token exchange -
1298 * wait for and receive SMC Proposal CLC message
1300 pclc = (struct smc_clc_msg_proposal *)&buf;
1301 rc = smc_clc_wait_msg(new_smc, pclc, SMC_CLC_MAX_LEN,
1302 SMC_CLC_PROPOSAL, CLC_WAIT_TIME);
1306 /* IPSec connections opt out of SMC-R optimizations */
1307 if (using_ipsec(new_smc)) {
1308 rc = SMC_CLC_DECL_IPSEC;
1312 /* check for matching IP prefix and subnet length */
1313 rc = smc_listen_prfx_check(new_smc, pclc);
1317 /* get vlan id from IP device */
1318 if (smc_vlan_by_tcpsk(new_smc->clcsock, &ini)) {
1319 rc = SMC_CLC_DECL_GETVLANERR;
1323 mutex_lock(&smc_server_lgr_pending);
1324 smc_close_init(new_smc);
1325 smc_rx_init(new_smc);
1326 smc_tx_init(new_smc);
1328 /* check if ISM is available */
1329 if (pclc->hdr.path == SMC_TYPE_D || pclc->hdr.path == SMC_TYPE_B) {
1330 ini.is_smcd = true; /* prepare ISM check */
1331 rc = smc_find_ism_device(new_smc, &ini);
1333 rc = smc_listen_ism_init(new_smc, pclc, &ini);
1335 ism_supported = true;
1336 else if (pclc->hdr.path == SMC_TYPE_D)
1337 goto out_unlock; /* skip RDMA and decline */
1340 /* check if RDMA is available */
1341 if (!ism_supported) { /* SMC_TYPE_R or SMC_TYPE_B */
1342 /* prepare RDMA check */
1343 ini.is_smcd = false;
1345 ini.ib_lcl = &pclc->lcl;
1346 rc = smc_find_rdma_device(new_smc, &ini);
1348 /* no RDMA device found */
1349 if (pclc->hdr.path == SMC_TYPE_B)
1350 /* neither ISM nor RDMA device found */
1351 rc = SMC_CLC_DECL_NOSMCDEV;
1354 rc = smc_listen_rdma_init(new_smc, &ini);
1357 rc = smc_listen_rdma_reg(new_smc, ini.cln_first_contact);
1362 /* send SMC Accept CLC message */
1363 rc = smc_clc_send_accept(new_smc, ini.cln_first_contact);
1367 /* SMC-D does not need this lock any more */
1369 mutex_unlock(&smc_server_lgr_pending);
1371 /* receive SMC Confirm CLC message */
1372 rc = smc_clc_wait_msg(new_smc, &cclc, sizeof(cclc),
1373 SMC_CLC_CONFIRM, CLC_WAIT_TIME);
1381 if (!ism_supported) {
1382 rc = smc_listen_rdma_finish(new_smc, &cclc,
1383 ini.cln_first_contact);
1384 mutex_unlock(&smc_server_lgr_pending);
1388 smc_conn_save_peer_info(new_smc, &cclc);
1389 smc_listen_out_connected(new_smc);
1393 mutex_unlock(&smc_server_lgr_pending);
1395 smc_listen_decline(new_smc, rc, ini.cln_first_contact);
1398 static void smc_tcp_listen_work(struct work_struct *work)
1400 struct smc_sock *lsmc = container_of(work, struct smc_sock,
1402 struct sock *lsk = &lsmc->sk;
1403 struct smc_sock *new_smc;
1407 while (lsk->sk_state == SMC_LISTEN) {
1408 rc = smc_clcsock_accept(lsmc, &new_smc);
1414 new_smc->listen_smc = lsmc;
1415 new_smc->use_fallback = lsmc->use_fallback;
1416 new_smc->fallback_rsn = lsmc->fallback_rsn;
1417 sock_hold(lsk); /* sock_put in smc_listen_work */
1418 INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
1419 smc_copy_sock_settings_to_smc(new_smc);
1420 new_smc->sk.sk_sndbuf = lsmc->sk.sk_sndbuf;
1421 new_smc->sk.sk_rcvbuf = lsmc->sk.sk_rcvbuf;
1422 sock_hold(&new_smc->sk); /* sock_put in passive closing */
1423 if (!schedule_work(&new_smc->smc_listen_work))
1424 sock_put(&new_smc->sk);
1429 sock_put(&lsmc->sk); /* sock_hold in smc_listen */
1432 static int smc_listen(struct socket *sock, int backlog)
1434 struct sock *sk = sock->sk;
1435 struct smc_sock *smc;
1442 if ((sk->sk_state != SMC_INIT && sk->sk_state != SMC_LISTEN) ||
1443 smc->connect_nonblock)
1447 if (sk->sk_state == SMC_LISTEN) {
1448 sk->sk_max_ack_backlog = backlog;
1451 /* some socket options are handled in core, so we could not apply
1452 * them to the clc socket -- copy smc socket options to clc socket
1454 smc_copy_sock_settings_to_clc(smc);
1455 if (!smc->use_fallback)
1456 tcp_sk(smc->clcsock->sk)->syn_smc = 1;
1458 rc = kernel_listen(smc->clcsock, backlog);
1461 sk->sk_max_ack_backlog = backlog;
1462 sk->sk_ack_backlog = 0;
1463 sk->sk_state = SMC_LISTEN;
1464 sock_hold(sk); /* sock_hold in tcp_listen_worker */
1465 if (!schedule_work(&smc->tcp_listen_work))
1473 static int smc_accept(struct socket *sock, struct socket *new_sock,
1474 int flags, bool kern)
1476 struct sock *sk = sock->sk, *nsk;
1477 DECLARE_WAITQUEUE(wait, current);
1478 struct smc_sock *lsmc;
1483 sock_hold(sk); /* sock_put below */
1486 if (lsmc->sk.sk_state != SMC_LISTEN) {
1492 /* Wait for an incoming connection */
1493 timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
1494 add_wait_queue_exclusive(sk_sleep(sk), &wait);
1495 while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
1496 set_current_state(TASK_INTERRUPTIBLE);
1502 timeo = schedule_timeout(timeo);
1503 /* wakeup by sk_data_ready in smc_listen_work() */
1504 sched_annotate_sleep();
1506 if (signal_pending(current)) {
1507 rc = sock_intr_errno(timeo);
1511 set_current_state(TASK_RUNNING);
1512 remove_wait_queue(sk_sleep(sk), &wait);
1515 rc = sock_error(nsk);
1520 if (lsmc->sockopt_defer_accept && !(flags & O_NONBLOCK)) {
1521 /* wait till data arrives on the socket */
1522 timeo = msecs_to_jiffies(lsmc->sockopt_defer_accept *
1524 if (smc_sk(nsk)->use_fallback) {
1525 struct sock *clcsk = smc_sk(nsk)->clcsock->sk;
1528 if (skb_queue_empty(&clcsk->sk_receive_queue))
1529 sk_wait_data(clcsk, &timeo, NULL);
1530 release_sock(clcsk);
1531 } else if (!atomic_read(&smc_sk(nsk)->conn.bytes_to_rcv)) {
1533 smc_rx_wait(smc_sk(nsk), &timeo, smc_rx_data_available);
1539 sock_put(sk); /* sock_hold above */
1543 static int smc_getname(struct socket *sock, struct sockaddr *addr,
1546 struct smc_sock *smc;
1548 if (peer && (sock->sk->sk_state != SMC_ACTIVE) &&
1549 (sock->sk->sk_state != SMC_APPCLOSEWAIT1))
1552 smc = smc_sk(sock->sk);
1554 return smc->clcsock->ops->getname(smc->clcsock, addr, peer);
1557 static int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
1559 struct sock *sk = sock->sk;
1560 struct smc_sock *smc;
1565 if ((sk->sk_state != SMC_ACTIVE) &&
1566 (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1567 (sk->sk_state != SMC_INIT))
1570 if (msg->msg_flags & MSG_FASTOPEN) {
1571 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
1572 smc_switch_to_fallback(smc);
1573 smc->fallback_rsn = SMC_CLC_DECL_OPTUNSUPP;
1580 if (smc->use_fallback)
1581 rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
1583 rc = smc_tx_sendmsg(smc, msg, len);
1589 static int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
1592 struct sock *sk = sock->sk;
1593 struct smc_sock *smc;
1598 if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
1599 /* socket was connected before, no more data to read */
1603 if ((sk->sk_state == SMC_INIT) ||
1604 (sk->sk_state == SMC_LISTEN) ||
1605 (sk->sk_state == SMC_CLOSED))
1608 if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
1613 if (smc->use_fallback) {
1614 rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
1616 msg->msg_namelen = 0;
1617 rc = smc_rx_recvmsg(smc, msg, NULL, len, flags);
1625 static __poll_t smc_accept_poll(struct sock *parent)
1627 struct smc_sock *isk = smc_sk(parent);
1630 spin_lock(&isk->accept_q_lock);
1631 if (!list_empty(&isk->accept_q))
1632 mask = EPOLLIN | EPOLLRDNORM;
1633 spin_unlock(&isk->accept_q_lock);
1638 static __poll_t smc_poll(struct file *file, struct socket *sock,
1641 struct sock *sk = sock->sk;
1642 struct smc_sock *smc;
1648 smc = smc_sk(sock->sk);
1649 if (smc->use_fallback) {
1650 /* delegate to CLC child sock */
1651 mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
1652 sk->sk_err = smc->clcsock->sk->sk_err;
1654 if (sk->sk_state != SMC_CLOSED)
1655 sock_poll_wait(file, sock, wait);
1658 if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
1659 (sk->sk_state == SMC_CLOSED))
1661 if (sk->sk_state == SMC_LISTEN) {
1662 /* woken up by sk_data_ready in smc_listen_work() */
1663 mask |= smc_accept_poll(sk);
1664 } else if (smc->use_fallback) { /* as result of connect_work()*/
1665 mask |= smc->clcsock->ops->poll(file, smc->clcsock,
1667 sk->sk_err = smc->clcsock->sk->sk_err;
1669 if ((sk->sk_state != SMC_INIT &&
1670 atomic_read(&smc->conn.sndbuf_space)) ||
1671 sk->sk_shutdown & SEND_SHUTDOWN) {
1672 mask |= EPOLLOUT | EPOLLWRNORM;
1674 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1675 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1677 if (atomic_read(&smc->conn.bytes_to_rcv))
1678 mask |= EPOLLIN | EPOLLRDNORM;
1679 if (sk->sk_shutdown & RCV_SHUTDOWN)
1680 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
1681 if (sk->sk_state == SMC_APPCLOSEWAIT1)
1683 if (smc->conn.urg_state == SMC_URG_VALID)
1691 static int smc_shutdown(struct socket *sock, int how)
1693 struct sock *sk = sock->sk;
1694 struct smc_sock *smc;
1700 if ((how < SHUT_RD) || (how > SHUT_RDWR))
1706 if ((sk->sk_state != SMC_ACTIVE) &&
1707 (sk->sk_state != SMC_PEERCLOSEWAIT1) &&
1708 (sk->sk_state != SMC_PEERCLOSEWAIT2) &&
1709 (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1710 (sk->sk_state != SMC_APPCLOSEWAIT2) &&
1711 (sk->sk_state != SMC_APPFINCLOSEWAIT))
1713 if (smc->use_fallback) {
1714 rc = kernel_sock_shutdown(smc->clcsock, how);
1715 sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
1716 if (sk->sk_shutdown == SHUTDOWN_MASK)
1717 sk->sk_state = SMC_CLOSED;
1721 case SHUT_RDWR: /* shutdown in both directions */
1722 rc = smc_close_active(smc);
1725 rc = smc_close_shutdown_write(smc);
1729 /* nothing more to do because peer is not involved */
1733 rc1 = kernel_sock_shutdown(smc->clcsock, how);
1734 /* map sock_shutdown_cmd constants to sk_shutdown value range */
1735 sk->sk_shutdown |= how + 1;
1739 return rc ? rc : rc1;
1742 static int smc_setsockopt(struct socket *sock, int level, int optname,
1743 sockptr_t optval, unsigned int optlen)
1745 struct sock *sk = sock->sk;
1746 struct smc_sock *smc;
1751 /* generic setsockopts reaching us here always apply to the
1754 if (unlikely(!smc->clcsock->ops->setsockopt))
1757 rc = smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
1759 if (smc->clcsock->sk->sk_err) {
1760 sk->sk_err = smc->clcsock->sk->sk_err;
1761 sk->sk_error_report(sk);
1764 if (optlen < sizeof(int))
1766 if (copy_from_sockptr(&val, optval, sizeof(int)))
1770 if (rc || smc->use_fallback)
1775 case TCP_FASTOPEN_CONNECT:
1776 case TCP_FASTOPEN_KEY:
1777 case TCP_FASTOPEN_NO_COOKIE:
1778 /* option not supported by SMC */
1779 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
1780 smc_switch_to_fallback(smc);
1781 smc->fallback_rsn = SMC_CLC_DECL_OPTUNSUPP;
1787 if (sk->sk_state != SMC_INIT &&
1788 sk->sk_state != SMC_LISTEN &&
1789 sk->sk_state != SMC_CLOSED) {
1791 mod_delayed_work(system_wq, &smc->conn.tx_work,
1796 if (sk->sk_state != SMC_INIT &&
1797 sk->sk_state != SMC_LISTEN &&
1798 sk->sk_state != SMC_CLOSED) {
1800 mod_delayed_work(system_wq, &smc->conn.tx_work,
1804 case TCP_DEFER_ACCEPT:
1805 smc->sockopt_defer_accept = val;
1816 static int smc_getsockopt(struct socket *sock, int level, int optname,
1817 char __user *optval, int __user *optlen)
1819 struct smc_sock *smc;
1821 smc = smc_sk(sock->sk);
1822 /* socket options apply to the CLC socket */
1823 if (unlikely(!smc->clcsock->ops->getsockopt))
1825 return smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
1829 static int smc_ioctl(struct socket *sock, unsigned int cmd,
1832 union smc_host_cursor cons, urg;
1833 struct smc_connection *conn;
1834 struct smc_sock *smc;
1837 smc = smc_sk(sock->sk);
1839 lock_sock(&smc->sk);
1840 if (smc->use_fallback) {
1841 if (!smc->clcsock) {
1842 release_sock(&smc->sk);
1845 answ = smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
1846 release_sock(&smc->sk);
1850 case SIOCINQ: /* same as FIONREAD */
1851 if (smc->sk.sk_state == SMC_LISTEN) {
1852 release_sock(&smc->sk);
1855 if (smc->sk.sk_state == SMC_INIT ||
1856 smc->sk.sk_state == SMC_CLOSED)
1859 answ = atomic_read(&smc->conn.bytes_to_rcv);
1862 /* output queue size (not send + not acked) */
1863 if (smc->sk.sk_state == SMC_LISTEN) {
1864 release_sock(&smc->sk);
1867 if (smc->sk.sk_state == SMC_INIT ||
1868 smc->sk.sk_state == SMC_CLOSED)
1871 answ = smc->conn.sndbuf_desc->len -
1872 atomic_read(&smc->conn.sndbuf_space);
1875 /* output queue size (not send only) */
1876 if (smc->sk.sk_state == SMC_LISTEN) {
1877 release_sock(&smc->sk);
1880 if (smc->sk.sk_state == SMC_INIT ||
1881 smc->sk.sk_state == SMC_CLOSED)
1884 answ = smc_tx_prepared_sends(&smc->conn);
1887 if (smc->sk.sk_state == SMC_LISTEN) {
1888 release_sock(&smc->sk);
1891 if (smc->sk.sk_state == SMC_INIT ||
1892 smc->sk.sk_state == SMC_CLOSED) {
1895 smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
1896 smc_curs_copy(&urg, &conn->urg_curs, conn);
1897 answ = smc_curs_diff(conn->rmb_desc->len,
1902 release_sock(&smc->sk);
1903 return -ENOIOCTLCMD;
1905 release_sock(&smc->sk);
1907 return put_user(answ, (int __user *)arg);
1910 static ssize_t smc_sendpage(struct socket *sock, struct page *page,
1911 int offset, size_t size, int flags)
1913 struct sock *sk = sock->sk;
1914 struct smc_sock *smc;
1919 if (sk->sk_state != SMC_ACTIVE) {
1924 if (smc->use_fallback)
1925 rc = kernel_sendpage(smc->clcsock, page, offset,
1928 rc = sock_no_sendpage(sock, page, offset, size, flags);
1934 /* Map the affected portions of the rmbe into an spd, note the number of bytes
1935 * to splice in conn->splice_pending, and press 'go'. Delays consumer cursor
1936 * updates till whenever a respective page has been fully processed.
1937 * Note that subsequent recv() calls have to wait till all splice() processing
1940 static ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
1941 struct pipe_inode_info *pipe, size_t len,
1944 struct sock *sk = sock->sk;
1945 struct smc_sock *smc;
1950 if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
1951 /* socket was connected before, no more data to read */
1955 if (sk->sk_state == SMC_INIT ||
1956 sk->sk_state == SMC_LISTEN ||
1957 sk->sk_state == SMC_CLOSED)
1960 if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
1965 if (smc->use_fallback) {
1966 rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
1973 if (flags & SPLICE_F_NONBLOCK)
1974 flags = MSG_DONTWAIT;
1977 rc = smc_rx_recvmsg(smc, NULL, pipe, len, flags);
1985 /* must look like tcp */
1986 static const struct proto_ops smc_sock_ops = {
1988 .owner = THIS_MODULE,
1989 .release = smc_release,
1991 .connect = smc_connect,
1992 .socketpair = sock_no_socketpair,
1993 .accept = smc_accept,
1994 .getname = smc_getname,
1997 .listen = smc_listen,
1998 .shutdown = smc_shutdown,
1999 .setsockopt = smc_setsockopt,
2000 .getsockopt = smc_getsockopt,
2001 .sendmsg = smc_sendmsg,
2002 .recvmsg = smc_recvmsg,
2003 .mmap = sock_no_mmap,
2004 .sendpage = smc_sendpage,
2005 .splice_read = smc_splice_read,
2008 static int smc_create(struct net *net, struct socket *sock, int protocol,
2011 int family = (protocol == SMCPROTO_SMC6) ? PF_INET6 : PF_INET;
2012 struct smc_sock *smc;
2016 rc = -ESOCKTNOSUPPORT;
2017 if (sock->type != SOCK_STREAM)
2020 rc = -EPROTONOSUPPORT;
2021 if (protocol != SMCPROTO_SMC && protocol != SMCPROTO_SMC6)
2025 sock->ops = &smc_sock_ops;
2026 sk = smc_sock_alloc(net, sock, protocol);
2030 /* create internal TCP socket for CLC handshake and fallback */
2032 smc->use_fallback = false; /* assume rdma capability first */
2033 smc->fallback_rsn = 0;
2034 rc = sock_create_kern(net, family, SOCK_STREAM, IPPROTO_TCP,
2037 sk_common_release(sk);
2040 smc->sk.sk_sndbuf = max(smc->clcsock->sk->sk_sndbuf, SMC_BUF_MIN_SIZE);
2041 smc->sk.sk_rcvbuf = max(smc->clcsock->sk->sk_rcvbuf, SMC_BUF_MIN_SIZE);
2047 static const struct net_proto_family smc_sock_family_ops = {
2049 .owner = THIS_MODULE,
2050 .create = smc_create,
2053 unsigned int smc_net_id;
2055 static __net_init int smc_net_init(struct net *net)
2057 return smc_pnet_net_init(net);
2060 static void __net_exit smc_net_exit(struct net *net)
2062 smc_pnet_net_exit(net);
2065 static struct pernet_operations smc_net_ops = {
2066 .init = smc_net_init,
2067 .exit = smc_net_exit,
2069 .size = sizeof(struct smc_net),
2072 static int __init smc_init(void)
2076 rc = register_pernet_subsys(&smc_net_ops);
2080 rc = smc_pnet_init();
2082 goto out_pernet_subsys;
2084 rc = smc_core_init();
2086 pr_err("%s: smc_core_init fails with %d\n", __func__, rc);
2090 rc = smc_llc_init();
2092 pr_err("%s: smc_llc_init fails with %d\n", __func__, rc);
2096 rc = smc_cdc_init();
2098 pr_err("%s: smc_cdc_init fails with %d\n", __func__, rc);
2102 rc = proto_register(&smc_proto, 1);
2104 pr_err("%s: proto_register(v4) fails with %d\n", __func__, rc);
2108 rc = proto_register(&smc_proto6, 1);
2110 pr_err("%s: proto_register(v6) fails with %d\n", __func__, rc);
2114 rc = sock_register(&smc_sock_family_ops);
2116 pr_err("%s: sock_register fails with %d\n", __func__, rc);
2119 INIT_HLIST_HEAD(&smc_v4_hashinfo.ht);
2120 INIT_HLIST_HEAD(&smc_v6_hashinfo.ht);
2122 rc = smc_ib_register_client();
2124 pr_err("%s: ib_register fails with %d\n", __func__, rc);
2128 static_branch_enable(&tcp_have_smc);
2132 sock_unregister(PF_SMC);
2134 proto_unregister(&smc_proto6);
2136 proto_unregister(&smc_proto);
2142 unregister_pernet_subsys(&smc_net_ops);
2147 static void __exit smc_exit(void)
2149 static_branch_disable(&tcp_have_smc);
2150 sock_unregister(PF_SMC);
2152 smc_ib_unregister_client();
2153 proto_unregister(&smc_proto6);
2154 proto_unregister(&smc_proto);
2156 unregister_pernet_subsys(&smc_net_ops);
2160 module_init(smc_init);
2161 module_exit(smc_exit);
2164 MODULE_DESCRIPTION("smc socket address family");
2165 MODULE_LICENSE("GPL");
2166 MODULE_ALIAS_NETPROTO(PF_SMC);