1 // SPDX-License-Identifier: GPL-2.0
3 * Shared Memory Communications over RDMA (SMC-R) and RoCE
5 * Basic Transport Functions exploiting Infiniband API
7 * Copyright IBM Corp. 2016
9 * Author(s): Ursula Braun <ubraun@linux.vnet.ibm.com>
12 #include <linux/socket.h>
13 #include <linux/if_vlan.h>
14 #include <linux/random.h>
15 #include <linux/workqueue.h>
16 #include <linux/wait.h>
17 #include <linux/reboot.h>
18 #include <linux/mutex.h>
19 #include <linux/list.h>
20 #include <linux/smc.h>
23 #include <rdma/ib_verbs.h>
24 #include <rdma/ib_cache.h>
33 #include "smc_close.h"
35 #include "smc_netlink.h"
36 #include "smc_stats.h"
37 #include "smc_tracepoint.h"
39 #define SMC_LGR_NUM_INCR 256
40 #define SMC_LGR_FREE_DELAY_SERV (600 * HZ)
41 #define SMC_LGR_FREE_DELAY_CLNT (SMC_LGR_FREE_DELAY_SERV + 10 * HZ)
43 struct smc_lgr_list smc_lgr_list = { /* established link groups */
44 .lock = __SPIN_LOCK_UNLOCKED(smc_lgr_list.lock),
45 .list = LIST_HEAD_INIT(smc_lgr_list.list),
49 static atomic_t lgr_cnt = ATOMIC_INIT(0); /* number of existing link groups */
50 static DECLARE_WAIT_QUEUE_HEAD(lgrs_deleted);
52 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
53 struct smc_buf_desc *buf_desc);
54 static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft);
56 static void smc_link_down_work(struct work_struct *work);
58 /* return head of link group list and its lock for a given link group */
59 static inline struct list_head *smc_lgr_list_head(struct smc_link_group *lgr,
60 spinlock_t **lgr_lock)
63 *lgr_lock = &lgr->smcd->lgr_lock;
64 return &lgr->smcd->lgr_list;
67 *lgr_lock = &smc_lgr_list.lock;
68 return &smc_lgr_list.list;
71 static void smc_ibdev_cnt_inc(struct smc_link *lnk)
73 atomic_inc(&lnk->smcibdev->lnk_cnt_by_port[lnk->ibport - 1]);
76 static void smc_ibdev_cnt_dec(struct smc_link *lnk)
78 atomic_dec(&lnk->smcibdev->lnk_cnt_by_port[lnk->ibport - 1]);
81 static void smc_lgr_schedule_free_work(struct smc_link_group *lgr)
83 /* client link group creation always follows the server link group
84 * creation. For client use a somewhat higher removal delay time,
85 * otherwise there is a risk of out-of-sync link groups.
88 mod_delayed_work(system_wq, &lgr->free_work,
89 (!lgr->is_smcd && lgr->role == SMC_CLNT) ?
90 SMC_LGR_FREE_DELAY_CLNT :
91 SMC_LGR_FREE_DELAY_SERV);
95 /* Register connection's alert token in our lookup structure.
96 * To use rbtrees we have to implement our own insert core.
97 * Requires @conns_lock
98 * @smc connection to register
99 * Returns 0 on success, != otherwise.
101 static void smc_lgr_add_alert_token(struct smc_connection *conn)
103 struct rb_node **link, *parent = NULL;
104 u32 token = conn->alert_token_local;
106 link = &conn->lgr->conns_all.rb_node;
108 struct smc_connection *cur = rb_entry(*link,
109 struct smc_connection, alert_node);
112 if (cur->alert_token_local > token)
113 link = &parent->rb_left;
115 link = &parent->rb_right;
117 /* Put the new node there */
118 rb_link_node(&conn->alert_node, parent, link);
119 rb_insert_color(&conn->alert_node, &conn->lgr->conns_all);
122 /* assign an SMC-R link to the connection */
123 static int smcr_lgr_conn_assign_link(struct smc_connection *conn, bool first)
125 enum smc_link_state expected = first ? SMC_LNK_ACTIVATING :
129 /* do link balancing */
130 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
131 struct smc_link *lnk = &conn->lgr->lnk[i];
133 if (lnk->state != expected || lnk->link_is_asym)
135 if (conn->lgr->role == SMC_CLNT) {
136 conn->lnk = lnk; /* temporary, SMC server assigns link*/
139 if (conn->lgr->conns_num % 2) {
140 for (j = i + 1; j < SMC_LINKS_PER_LGR_MAX; j++) {
141 struct smc_link *lnk2;
143 lnk2 = &conn->lgr->lnk[j];
144 if (lnk2->state == expected &&
145 !lnk2->link_is_asym) {
156 return SMC_CLC_DECL_NOACTLINK;
157 atomic_inc(&conn->lnk->conn_cnt);
161 /* Register connection in link group by assigning an alert token
162 * registered in a search tree.
163 * Requires @conns_lock
164 * Note that '0' is a reserved value and not assigned.
166 static int smc_lgr_register_conn(struct smc_connection *conn, bool first)
168 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
169 static atomic_t nexttoken = ATOMIC_INIT(0);
172 if (!conn->lgr->is_smcd) {
173 rc = smcr_lgr_conn_assign_link(conn, first);
179 /* find a new alert_token_local value not yet used by some connection
182 sock_hold(&smc->sk); /* sock_put in smc_lgr_unregister_conn() */
183 while (!conn->alert_token_local) {
184 conn->alert_token_local = atomic_inc_return(&nexttoken);
185 if (smc_lgr_find_conn(conn->alert_token_local, conn->lgr))
186 conn->alert_token_local = 0;
188 smc_lgr_add_alert_token(conn);
189 conn->lgr->conns_num++;
193 /* Unregister connection and reset the alert token of the given connection<
195 static void __smc_lgr_unregister_conn(struct smc_connection *conn)
197 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
198 struct smc_link_group *lgr = conn->lgr;
200 rb_erase(&conn->alert_node, &lgr->conns_all);
202 atomic_dec(&conn->lnk->conn_cnt);
204 conn->alert_token_local = 0;
205 sock_put(&smc->sk); /* sock_hold in smc_lgr_register_conn() */
208 /* Unregister connection from lgr
210 static void smc_lgr_unregister_conn(struct smc_connection *conn)
212 struct smc_link_group *lgr = conn->lgr;
214 if (!smc_conn_lgr_valid(conn))
216 write_lock_bh(&lgr->conns_lock);
217 if (conn->alert_token_local) {
218 __smc_lgr_unregister_conn(conn);
220 write_unlock_bh(&lgr->conns_lock);
223 int smc_nl_get_sys_info(struct sk_buff *skb, struct netlink_callback *cb)
225 struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
226 char hostname[SMC_MAX_HOSTNAME_LEN + 1];
227 char smc_seid[SMC_MAX_EID_LEN + 1];
228 struct nlattr *attrs;
233 nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
234 &smc_gen_nl_family, NLM_F_MULTI,
235 SMC_NETLINK_GET_SYS_INFO);
240 attrs = nla_nest_start(skb, SMC_GEN_SYS_INFO);
243 if (nla_put_u8(skb, SMC_NLA_SYS_VER, SMC_V2))
245 if (nla_put_u8(skb, SMC_NLA_SYS_REL, SMC_RELEASE))
247 if (nla_put_u8(skb, SMC_NLA_SYS_IS_ISM_V2, smc_ism_is_v2_capable()))
249 if (nla_put_u8(skb, SMC_NLA_SYS_IS_SMCR_V2, true))
251 smc_clc_get_hostname(&host);
253 memcpy(hostname, host, SMC_MAX_HOSTNAME_LEN);
254 hostname[SMC_MAX_HOSTNAME_LEN] = 0;
255 if (nla_put_string(skb, SMC_NLA_SYS_LOCAL_HOST, hostname))
258 if (smc_ism_is_v2_capable()) {
259 smc_ism_get_system_eid(&seid);
260 memcpy(smc_seid, seid, SMC_MAX_EID_LEN);
261 smc_seid[SMC_MAX_EID_LEN] = 0;
262 if (nla_put_string(skb, SMC_NLA_SYS_SEID, smc_seid))
265 nla_nest_end(skb, attrs);
266 genlmsg_end(skb, nlh);
271 nla_nest_cancel(skb, attrs);
273 genlmsg_cancel(skb, nlh);
278 /* Fill SMC_NLA_LGR_D_V2_COMMON/SMC_NLA_LGR_R_V2_COMMON nested attributes */
279 static int smc_nl_fill_lgr_v2_common(struct smc_link_group *lgr,
281 struct netlink_callback *cb,
282 struct nlattr *v2_attrs)
284 char smc_host[SMC_MAX_HOSTNAME_LEN + 1];
285 char smc_eid[SMC_MAX_EID_LEN + 1];
287 if (nla_put_u8(skb, SMC_NLA_LGR_V2_VER, lgr->smc_version))
289 if (nla_put_u8(skb, SMC_NLA_LGR_V2_REL, lgr->peer_smc_release))
291 if (nla_put_u8(skb, SMC_NLA_LGR_V2_OS, lgr->peer_os))
293 memcpy(smc_host, lgr->peer_hostname, SMC_MAX_HOSTNAME_LEN);
294 smc_host[SMC_MAX_HOSTNAME_LEN] = 0;
295 if (nla_put_string(skb, SMC_NLA_LGR_V2_PEER_HOST, smc_host))
297 memcpy(smc_eid, lgr->negotiated_eid, SMC_MAX_EID_LEN);
298 smc_eid[SMC_MAX_EID_LEN] = 0;
299 if (nla_put_string(skb, SMC_NLA_LGR_V2_NEG_EID, smc_eid))
302 nla_nest_end(skb, v2_attrs);
306 nla_nest_cancel(skb, v2_attrs);
310 static int smc_nl_fill_smcr_lgr_v2(struct smc_link_group *lgr,
312 struct netlink_callback *cb)
314 struct nlattr *v2_attrs;
316 v2_attrs = nla_nest_start(skb, SMC_NLA_LGR_R_V2);
319 if (nla_put_u8(skb, SMC_NLA_LGR_R_V2_DIRECT, !lgr->uses_gateway))
322 nla_nest_end(skb, v2_attrs);
326 nla_nest_cancel(skb, v2_attrs);
331 static int smc_nl_fill_lgr(struct smc_link_group *lgr,
333 struct netlink_callback *cb)
335 char smc_target[SMC_MAX_PNETID_LEN + 1];
336 struct nlattr *attrs, *v2_attrs;
338 attrs = nla_nest_start(skb, SMC_GEN_LGR_SMCR);
342 if (nla_put_u32(skb, SMC_NLA_LGR_R_ID, *((u32 *)&lgr->id)))
344 if (nla_put_u32(skb, SMC_NLA_LGR_R_CONNS_NUM, lgr->conns_num))
346 if (nla_put_u8(skb, SMC_NLA_LGR_R_ROLE, lgr->role))
348 if (nla_put_u8(skb, SMC_NLA_LGR_R_TYPE, lgr->type))
350 if (nla_put_u8(skb, SMC_NLA_LGR_R_BUF_TYPE, lgr->buf_type))
352 if (nla_put_u8(skb, SMC_NLA_LGR_R_VLAN_ID, lgr->vlan_id))
354 if (nla_put_u64_64bit(skb, SMC_NLA_LGR_R_NET_COOKIE,
355 lgr->net->net_cookie, SMC_NLA_LGR_R_PAD))
357 memcpy(smc_target, lgr->pnet_id, SMC_MAX_PNETID_LEN);
358 smc_target[SMC_MAX_PNETID_LEN] = 0;
359 if (nla_put_string(skb, SMC_NLA_LGR_R_PNETID, smc_target))
361 if (lgr->smc_version > SMC_V1) {
362 v2_attrs = nla_nest_start(skb, SMC_NLA_LGR_R_V2_COMMON);
365 if (smc_nl_fill_lgr_v2_common(lgr, skb, cb, v2_attrs))
367 if (smc_nl_fill_smcr_lgr_v2(lgr, skb, cb))
371 nla_nest_end(skb, attrs);
374 nla_nest_cancel(skb, attrs);
379 static int smc_nl_fill_lgr_link(struct smc_link_group *lgr,
380 struct smc_link *link,
382 struct netlink_callback *cb)
384 char smc_ibname[IB_DEVICE_NAME_MAX];
385 u8 smc_gid_target[41];
386 struct nlattr *attrs;
390 nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
391 &smc_gen_nl_family, NLM_F_MULTI,
392 SMC_NETLINK_GET_LINK_SMCR);
396 attrs = nla_nest_start(skb, SMC_GEN_LINK_SMCR);
400 if (nla_put_u8(skb, SMC_NLA_LINK_ID, link->link_id))
402 if (nla_put_u32(skb, SMC_NLA_LINK_STATE, link->state))
404 if (nla_put_u32(skb, SMC_NLA_LINK_CONN_CNT,
405 atomic_read(&link->conn_cnt)))
407 if (nla_put_u8(skb, SMC_NLA_LINK_IB_PORT, link->ibport))
409 if (nla_put_u32(skb, SMC_NLA_LINK_NET_DEV, link->ndev_ifidx))
411 snprintf(smc_ibname, sizeof(smc_ibname), "%s", link->ibname);
412 if (nla_put_string(skb, SMC_NLA_LINK_IB_DEV, smc_ibname))
414 memcpy(&link_uid, link->link_uid, sizeof(link_uid));
415 if (nla_put_u32(skb, SMC_NLA_LINK_UID, link_uid))
417 memcpy(&link_uid, link->peer_link_uid, sizeof(link_uid));
418 if (nla_put_u32(skb, SMC_NLA_LINK_PEER_UID, link_uid))
420 memset(smc_gid_target, 0, sizeof(smc_gid_target));
421 smc_gid_be16_convert(smc_gid_target, link->gid);
422 if (nla_put_string(skb, SMC_NLA_LINK_GID, smc_gid_target))
424 memset(smc_gid_target, 0, sizeof(smc_gid_target));
425 smc_gid_be16_convert(smc_gid_target, link->peer_gid);
426 if (nla_put_string(skb, SMC_NLA_LINK_PEER_GID, smc_gid_target))
429 nla_nest_end(skb, attrs);
430 genlmsg_end(skb, nlh);
433 nla_nest_cancel(skb, attrs);
435 genlmsg_cancel(skb, nlh);
440 static int smc_nl_handle_lgr(struct smc_link_group *lgr,
442 struct netlink_callback *cb,
448 nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
449 &smc_gen_nl_family, NLM_F_MULTI,
450 SMC_NETLINK_GET_LGR_SMCR);
453 if (smc_nl_fill_lgr(lgr, skb, cb))
456 genlmsg_end(skb, nlh);
459 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
460 if (!smc_link_usable(&lgr->lnk[i]))
462 if (smc_nl_fill_lgr_link(lgr, &lgr->lnk[i], skb, cb))
469 genlmsg_cancel(skb, nlh);
474 static void smc_nl_fill_lgr_list(struct smc_lgr_list *smc_lgr,
476 struct netlink_callback *cb,
479 struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
480 struct smc_link_group *lgr;
481 int snum = cb_ctx->pos[0];
484 spin_lock_bh(&smc_lgr->lock);
485 list_for_each_entry(lgr, &smc_lgr->list, list) {
488 if (smc_nl_handle_lgr(lgr, skb, cb, list_links))
494 spin_unlock_bh(&smc_lgr->lock);
495 cb_ctx->pos[0] = num;
498 static int smc_nl_fill_smcd_lgr(struct smc_link_group *lgr,
500 struct netlink_callback *cb)
502 char smc_pnet[SMC_MAX_PNETID_LEN + 1];
503 struct nlattr *attrs;
506 nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
507 &smc_gen_nl_family, NLM_F_MULTI,
508 SMC_NETLINK_GET_LGR_SMCD);
512 attrs = nla_nest_start(skb, SMC_GEN_LGR_SMCD);
516 if (nla_put_u32(skb, SMC_NLA_LGR_D_ID, *((u32 *)&lgr->id)))
518 if (nla_put_u64_64bit(skb, SMC_NLA_LGR_D_GID, lgr->smcd->local_gid,
521 if (nla_put_u64_64bit(skb, SMC_NLA_LGR_D_PEER_GID, lgr->peer_gid,
524 if (nla_put_u8(skb, SMC_NLA_LGR_D_VLAN_ID, lgr->vlan_id))
526 if (nla_put_u32(skb, SMC_NLA_LGR_D_CONNS_NUM, lgr->conns_num))
528 if (nla_put_u32(skb, SMC_NLA_LGR_D_CHID, smc_ism_get_chid(lgr->smcd)))
530 memcpy(smc_pnet, lgr->smcd->pnetid, SMC_MAX_PNETID_LEN);
531 smc_pnet[SMC_MAX_PNETID_LEN] = 0;
532 if (nla_put_string(skb, SMC_NLA_LGR_D_PNETID, smc_pnet))
534 if (lgr->smc_version > SMC_V1) {
535 struct nlattr *v2_attrs;
537 v2_attrs = nla_nest_start(skb, SMC_NLA_LGR_D_V2_COMMON);
540 if (smc_nl_fill_lgr_v2_common(lgr, skb, cb, v2_attrs))
543 nla_nest_end(skb, attrs);
544 genlmsg_end(skb, nlh);
548 nla_nest_cancel(skb, attrs);
550 genlmsg_cancel(skb, nlh);
555 static int smc_nl_handle_smcd_lgr(struct smcd_dev *dev,
557 struct netlink_callback *cb)
559 struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
560 struct smc_link_group *lgr;
561 int snum = cb_ctx->pos[1];
564 spin_lock_bh(&dev->lgr_lock);
565 list_for_each_entry(lgr, &dev->lgr_list, list) {
570 rc = smc_nl_fill_smcd_lgr(lgr, skb, cb);
577 spin_unlock_bh(&dev->lgr_lock);
578 cb_ctx->pos[1] = num;
582 static int smc_nl_fill_smcd_dev(struct smcd_dev_list *dev_list,
584 struct netlink_callback *cb)
586 struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
587 struct smcd_dev *smcd_dev;
588 int snum = cb_ctx->pos[0];
591 mutex_lock(&dev_list->mutex);
592 list_for_each_entry(smcd_dev, &dev_list->list, list) {
593 if (list_empty(&smcd_dev->lgr_list))
597 rc = smc_nl_handle_smcd_lgr(smcd_dev, skb, cb);
604 mutex_unlock(&dev_list->mutex);
605 cb_ctx->pos[0] = num;
609 int smcr_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb)
611 bool list_links = false;
613 smc_nl_fill_lgr_list(&smc_lgr_list, skb, cb, list_links);
617 int smcr_nl_get_link(struct sk_buff *skb, struct netlink_callback *cb)
619 bool list_links = true;
621 smc_nl_fill_lgr_list(&smc_lgr_list, skb, cb, list_links);
625 int smcd_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb)
627 smc_nl_fill_smcd_dev(&smcd_dev_list, skb, cb);
631 void smc_lgr_cleanup_early(struct smc_link_group *lgr)
633 spinlock_t *lgr_lock;
638 smc_lgr_list_head(lgr, &lgr_lock);
639 spin_lock_bh(lgr_lock);
640 /* do not use this link group for new connections */
641 if (!list_empty(&lgr->list))
642 list_del_init(&lgr->list);
643 spin_unlock_bh(lgr_lock);
644 __smc_lgr_terminate(lgr, true);
647 static void smcr_lgr_link_deactivate_all(struct smc_link_group *lgr)
651 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
652 struct smc_link *lnk = &lgr->lnk[i];
654 if (smc_link_sendable(lnk))
655 lnk->state = SMC_LNK_INACTIVE;
657 wake_up_all(&lgr->llc_msg_waiter);
658 wake_up_all(&lgr->llc_flow_waiter);
661 static void smc_lgr_free(struct smc_link_group *lgr);
663 static void smc_lgr_free_work(struct work_struct *work)
665 struct smc_link_group *lgr = container_of(to_delayed_work(work),
666 struct smc_link_group,
668 spinlock_t *lgr_lock;
671 smc_lgr_list_head(lgr, &lgr_lock);
672 spin_lock_bh(lgr_lock);
674 spin_unlock_bh(lgr_lock);
677 read_lock_bh(&lgr->conns_lock);
678 conns = RB_EMPTY_ROOT(&lgr->conns_all);
679 read_unlock_bh(&lgr->conns_lock);
680 if (!conns) { /* number of lgr connections is no longer zero */
681 spin_unlock_bh(lgr_lock);
684 list_del_init(&lgr->list); /* remove from smc_lgr_list */
685 lgr->freeing = 1; /* this instance does the freeing, no new schedule */
686 spin_unlock_bh(lgr_lock);
687 cancel_delayed_work(&lgr->free_work);
689 if (!lgr->is_smcd && !lgr->terminating)
690 smc_llc_send_link_delete_all(lgr, true,
691 SMC_LLC_DEL_PROG_INIT_TERM);
692 if (lgr->is_smcd && !lgr->terminating)
693 smc_ism_signal_shutdown(lgr);
695 smcr_lgr_link_deactivate_all(lgr);
699 static void smc_lgr_terminate_work(struct work_struct *work)
701 struct smc_link_group *lgr = container_of(work, struct smc_link_group,
704 __smc_lgr_terminate(lgr, true);
707 /* return next unique link id for the lgr */
708 static u8 smcr_next_link_id(struct smc_link_group *lgr)
715 link_id = ++lgr->next_link_id;
716 if (!link_id) /* skip zero as link_id */
717 link_id = ++lgr->next_link_id;
718 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
719 if (smc_link_usable(&lgr->lnk[i]) &&
720 lgr->lnk[i].link_id == link_id)
728 static void smcr_copy_dev_info_to_link(struct smc_link *link)
730 struct smc_ib_device *smcibdev = link->smcibdev;
732 snprintf(link->ibname, sizeof(link->ibname), "%s",
733 smcibdev->ibdev->name);
734 link->ndev_ifidx = smcibdev->ndev_ifidx[link->ibport - 1];
737 int smcr_link_init(struct smc_link_group *lgr, struct smc_link *lnk,
738 u8 link_idx, struct smc_init_info *ini)
740 struct smc_ib_device *smcibdev;
744 if (lgr->smc_version == SMC_V2) {
745 lnk->smcibdev = ini->smcrv2.ib_dev_v2;
746 lnk->ibport = ini->smcrv2.ib_port_v2;
748 lnk->smcibdev = ini->ib_dev;
749 lnk->ibport = ini->ib_port;
751 get_device(&lnk->smcibdev->ibdev->dev);
752 atomic_inc(&lnk->smcibdev->lnk_cnt);
753 refcount_set(&lnk->refcnt, 1); /* link refcnt is set to 1 */
755 lnk->path_mtu = lnk->smcibdev->pattr[lnk->ibport - 1].active_mtu;
756 lnk->link_id = smcr_next_link_id(lgr);
758 smc_lgr_hold(lgr); /* lgr_put in smcr_link_clear() */
759 lnk->link_idx = link_idx;
760 smc_ibdev_cnt_inc(lnk);
761 smcr_copy_dev_info_to_link(lnk);
762 atomic_set(&lnk->conn_cnt, 0);
763 smc_llc_link_set_uid(lnk);
764 INIT_WORK(&lnk->link_down_wrk, smc_link_down_work);
765 if (!lnk->smcibdev->initialized) {
766 rc = (int)smc_ib_setup_per_ibdev(lnk->smcibdev);
770 get_random_bytes(rndvec, sizeof(rndvec));
771 lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) +
773 rc = smc_ib_determine_gid(lnk->smcibdev, lnk->ibport,
774 ini->vlan_id, lnk->gid, &lnk->sgid_index,
775 lgr->smc_version == SMC_V2 ?
776 &ini->smcrv2 : NULL);
779 rc = smc_llc_link_init(lnk);
782 rc = smc_wr_alloc_link_mem(lnk);
785 rc = smc_ib_create_protection_domain(lnk);
788 rc = smc_ib_create_queue_pair(lnk);
791 rc = smc_wr_create_link(lnk);
794 lnk->state = SMC_LNK_ACTIVATING;
798 smc_ib_destroy_queue_pair(lnk);
800 smc_ib_dealloc_protection_domain(lnk);
802 smc_wr_free_link_mem(lnk);
804 smc_llc_link_clear(lnk, false);
806 smc_ibdev_cnt_dec(lnk);
807 put_device(&lnk->smcibdev->ibdev->dev);
808 smcibdev = lnk->smcibdev;
809 memset(lnk, 0, sizeof(struct smc_link));
810 lnk->state = SMC_LNK_UNUSED;
811 if (!atomic_dec_return(&smcibdev->lnk_cnt))
812 wake_up(&smcibdev->lnks_deleted);
813 smc_lgr_put(lgr); /* lgr_hold above */
817 /* create a new SMC link group */
818 static int smc_lgr_create(struct smc_sock *smc, struct smc_init_info *ini)
820 struct smc_link_group *lgr;
821 struct list_head *lgr_list;
822 struct smc_link *lnk;
823 spinlock_t *lgr_lock;
828 if (ini->is_smcd && ini->vlan_id) {
829 if (smc_ism_get_vlan(ini->ism_dev[ini->ism_selected],
831 rc = SMC_CLC_DECL_ISMVLANERR;
836 lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
838 rc = SMC_CLC_DECL_MEM;
841 lgr->tx_wq = alloc_workqueue("smc_tx_wq-%*phN", 0, 0,
842 SMC_LGR_ID_SIZE, &lgr->id);
847 lgr->is_smcd = ini->is_smcd;
849 lgr->terminating = 0;
851 lgr->vlan_id = ini->vlan_id;
852 refcount_set(&lgr->refcnt, 1); /* set lgr refcnt to 1 */
853 mutex_init(&lgr->sndbufs_lock);
854 mutex_init(&lgr->rmbs_lock);
855 rwlock_init(&lgr->conns_lock);
856 for (i = 0; i < SMC_RMBE_SIZES; i++) {
857 INIT_LIST_HEAD(&lgr->sndbufs[i]);
858 INIT_LIST_HEAD(&lgr->rmbs[i]);
860 lgr->next_link_id = 0;
861 smc_lgr_list.num += SMC_LGR_NUM_INCR;
862 memcpy(&lgr->id, (u8 *)&smc_lgr_list.num, SMC_LGR_ID_SIZE);
863 INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
864 INIT_WORK(&lgr->terminate_work, smc_lgr_terminate_work);
865 lgr->conns_all = RB_ROOT;
867 /* SMC-D specific settings */
868 get_device(&ini->ism_dev[ini->ism_selected]->dev);
869 lgr->peer_gid = ini->ism_peer_gid[ini->ism_selected];
870 lgr->smcd = ini->ism_dev[ini->ism_selected];
871 lgr_list = &ini->ism_dev[ini->ism_selected]->lgr_list;
872 lgr_lock = &lgr->smcd->lgr_lock;
873 lgr->smc_version = ini->smcd_version;
874 lgr->peer_shutdown = 0;
875 atomic_inc(&ini->ism_dev[ini->ism_selected]->lgr_cnt);
877 /* SMC-R specific settings */
878 struct smc_ib_device *ibdev;
881 lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
882 lgr->smc_version = ini->smcr_version;
883 memcpy(lgr->peer_systemid, ini->peer_systemid,
885 if (lgr->smc_version == SMC_V2) {
886 ibdev = ini->smcrv2.ib_dev_v2;
887 ibport = ini->smcrv2.ib_port_v2;
888 lgr->saddr = ini->smcrv2.saddr;
889 lgr->uses_gateway = ini->smcrv2.uses_gateway;
890 memcpy(lgr->nexthop_mac, ini->smcrv2.nexthop_mac,
894 ibport = ini->ib_port;
896 memcpy(lgr->pnet_id, ibdev->pnetid[ibport - 1],
898 if (smc_wr_alloc_lgr_mem(lgr))
900 smc_llc_lgr_init(lgr, smc);
902 link_idx = SMC_SINGLE_LINK;
903 lnk = &lgr->lnk[link_idx];
904 rc = smcr_link_init(lgr, lnk, link_idx, ini);
906 smc_wr_free_lgr_mem(lgr);
909 lgr->net = smc_ib_net(lnk->smcibdev);
910 lgr_list = &smc_lgr_list.list;
911 lgr_lock = &smc_lgr_list.lock;
912 lgr->buf_type = lgr->net->smc.sysctl_smcr_buf_type;
913 atomic_inc(&lgr_cnt);
916 spin_lock_bh(lgr_lock);
917 list_add_tail(&lgr->list, lgr_list);
918 spin_unlock_bh(lgr_lock);
922 destroy_workqueue(lgr->tx_wq);
926 if (ini->is_smcd && ini->vlan_id)
927 smc_ism_put_vlan(ini->ism_dev[ini->ism_selected], ini->vlan_id);
931 rc = SMC_CLC_DECL_MEM;
933 rc = SMC_CLC_DECL_INTERR;
938 static int smc_write_space(struct smc_connection *conn)
940 int buffer_len = conn->peer_rmbe_size;
941 union smc_host_cursor prod;
942 union smc_host_cursor cons;
945 smc_curs_copy(&prod, &conn->local_tx_ctrl.prod, conn);
946 smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
947 /* determine rx_buf space */
948 space = buffer_len - smc_curs_diff(buffer_len, &cons, &prod);
952 static int smc_switch_cursor(struct smc_sock *smc, struct smc_cdc_tx_pend *pend,
953 struct smc_wr_buf *wr_buf)
955 struct smc_connection *conn = &smc->conn;
956 union smc_host_cursor cons, fin;
960 smc_curs_copy(&conn->tx_curs_sent, &conn->tx_curs_fin, conn);
961 smc_curs_copy(&fin, &conn->local_tx_ctrl_fin, conn);
962 /* set prod cursor to old state, enforce tx_rdma_writes() */
963 smc_curs_copy(&conn->local_tx_ctrl.prod, &fin, conn);
964 smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
966 if (smc_curs_comp(conn->peer_rmbe_size, &cons, &fin) < 0) {
967 /* cons cursor advanced more than fin, and prod was set
968 * fin above, so now prod is smaller than cons. Fix that.
970 diff = smc_curs_diff(conn->peer_rmbe_size, &fin, &cons);
971 smc_curs_add(conn->sndbuf_desc->len,
972 &conn->tx_curs_sent, diff);
973 smc_curs_add(conn->sndbuf_desc->len,
974 &conn->tx_curs_fin, diff);
976 smp_mb__before_atomic();
977 atomic_add(diff, &conn->sndbuf_space);
978 smp_mb__after_atomic();
980 smc_curs_add(conn->peer_rmbe_size,
981 &conn->local_tx_ctrl.prod, diff);
982 smc_curs_add(conn->peer_rmbe_size,
983 &conn->local_tx_ctrl_fin, diff);
985 /* recalculate, value is used by tx_rdma_writes() */
986 atomic_set(&smc->conn.peer_rmbe_space, smc_write_space(conn));
988 if (smc->sk.sk_state != SMC_INIT &&
989 smc->sk.sk_state != SMC_CLOSED) {
990 rc = smcr_cdc_msg_send_validation(conn, pend, wr_buf);
992 queue_delayed_work(conn->lgr->tx_wq, &conn->tx_work, 0);
993 smc->sk.sk_data_ready(&smc->sk);
996 smc_wr_tx_put_slot(conn->lnk,
997 (struct smc_wr_tx_pend_priv *)pend);
1002 void smc_switch_link_and_count(struct smc_connection *conn,
1003 struct smc_link *to_lnk)
1005 atomic_dec(&conn->lnk->conn_cnt);
1006 /* link_hold in smc_conn_create() */
1007 smcr_link_put(conn->lnk);
1009 atomic_inc(&conn->lnk->conn_cnt);
1010 /* link_put in smc_conn_free() */
1011 smcr_link_hold(conn->lnk);
1014 struct smc_link *smc_switch_conns(struct smc_link_group *lgr,
1015 struct smc_link *from_lnk, bool is_dev_err)
1017 struct smc_link *to_lnk = NULL;
1018 struct smc_cdc_tx_pend *pend;
1019 struct smc_connection *conn;
1020 struct smc_wr_buf *wr_buf;
1021 struct smc_sock *smc;
1022 struct rb_node *node;
1025 /* link is inactive, wake up tx waiters */
1026 smc_wr_wakeup_tx_wait(from_lnk);
1028 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1029 if (!smc_link_active(&lgr->lnk[i]) || i == from_lnk->link_idx)
1031 if (is_dev_err && from_lnk->smcibdev == lgr->lnk[i].smcibdev &&
1032 from_lnk->ibport == lgr->lnk[i].ibport) {
1035 to_lnk = &lgr->lnk[i];
1038 if (!to_lnk || !smc_wr_tx_link_hold(to_lnk)) {
1039 smc_lgr_terminate_sched(lgr);
1043 read_lock_bh(&lgr->conns_lock);
1044 for (node = rb_first(&lgr->conns_all); node; node = rb_next(node)) {
1045 conn = rb_entry(node, struct smc_connection, alert_node);
1046 if (conn->lnk != from_lnk)
1048 smc = container_of(conn, struct smc_sock, conn);
1049 /* conn->lnk not yet set in SMC_INIT state */
1050 if (smc->sk.sk_state == SMC_INIT)
1052 if (smc->sk.sk_state == SMC_CLOSED ||
1053 smc->sk.sk_state == SMC_PEERCLOSEWAIT1 ||
1054 smc->sk.sk_state == SMC_PEERCLOSEWAIT2 ||
1055 smc->sk.sk_state == SMC_APPFINCLOSEWAIT ||
1056 smc->sk.sk_state == SMC_APPCLOSEWAIT1 ||
1057 smc->sk.sk_state == SMC_APPCLOSEWAIT2 ||
1058 smc->sk.sk_state == SMC_PEERFINCLOSEWAIT ||
1059 smc->sk.sk_state == SMC_PEERABORTWAIT ||
1060 smc->sk.sk_state == SMC_PROCESSABORT) {
1061 spin_lock_bh(&conn->send_lock);
1062 smc_switch_link_and_count(conn, to_lnk);
1063 spin_unlock_bh(&conn->send_lock);
1066 sock_hold(&smc->sk);
1067 read_unlock_bh(&lgr->conns_lock);
1068 /* pre-fetch buffer outside of send_lock, might sleep */
1069 rc = smc_cdc_get_free_slot(conn, to_lnk, &wr_buf, NULL, &pend);
1072 /* avoid race with smcr_tx_sndbuf_nonempty() */
1073 spin_lock_bh(&conn->send_lock);
1074 smc_switch_link_and_count(conn, to_lnk);
1075 rc = smc_switch_cursor(smc, pend, wr_buf);
1076 spin_unlock_bh(&conn->send_lock);
1082 read_unlock_bh(&lgr->conns_lock);
1083 smc_wr_tx_link_put(to_lnk);
1087 smcr_link_down_cond_sched(to_lnk);
1088 smc_wr_tx_link_put(to_lnk);
1092 static void smcr_buf_unuse(struct smc_buf_desc *buf_desc, bool is_rmb,
1093 struct smc_link_group *lgr)
1095 struct mutex *lock; /* lock buffer list */
1098 if (is_rmb && buf_desc->is_conf_rkey && !list_empty(&lgr->list)) {
1099 /* unregister rmb with peer */
1100 rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY);
1102 /* protect against smc_llc_cli_rkey_exchange() */
1103 mutex_lock(&lgr->llc_conf_mutex);
1104 smc_llc_do_delete_rkey(lgr, buf_desc);
1105 buf_desc->is_conf_rkey = false;
1106 mutex_unlock(&lgr->llc_conf_mutex);
1107 smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl);
1111 if (buf_desc->is_reg_err) {
1112 /* buf registration failed, reuse not possible */
1113 lock = is_rmb ? &lgr->rmbs_lock :
1116 list_del(&buf_desc->list);
1119 smc_buf_free(lgr, is_rmb, buf_desc);
1122 memset(buf_desc->cpu_addr, 0, buf_desc->len);
1126 static void smc_buf_unuse(struct smc_connection *conn,
1127 struct smc_link_group *lgr)
1129 if (conn->sndbuf_desc) {
1130 if (!lgr->is_smcd && conn->sndbuf_desc->is_vm) {
1131 smcr_buf_unuse(conn->sndbuf_desc, false, lgr);
1133 conn->sndbuf_desc->used = 0;
1134 memset(conn->sndbuf_desc->cpu_addr, 0,
1135 conn->sndbuf_desc->len);
1138 if (conn->rmb_desc) {
1139 if (!lgr->is_smcd) {
1140 smcr_buf_unuse(conn->rmb_desc, true, lgr);
1142 conn->rmb_desc->used = 0;
1143 memset(conn->rmb_desc->cpu_addr, 0,
1144 conn->rmb_desc->len +
1145 sizeof(struct smcd_cdc_msg));
1150 /* remove a finished connection from its link group */
1151 void smc_conn_free(struct smc_connection *conn)
1153 struct smc_link_group *lgr = conn->lgr;
1155 if (!lgr || conn->freed)
1156 /* Connection has never been registered in a
1157 * link group, or has already been freed.
1162 if (!smc_conn_lgr_valid(conn))
1163 /* Connection has already unregistered from
1169 if (!list_empty(&lgr->list))
1170 smc_ism_unset_conn(conn);
1171 tasklet_kill(&conn->rx_tsklet);
1173 smc_cdc_wait_pend_tx_wr(conn);
1174 if (current_work() != &conn->abort_work)
1175 cancel_work_sync(&conn->abort_work);
1177 if (!list_empty(&lgr->list)) {
1178 smc_buf_unuse(conn, lgr); /* allow buffer reuse */
1179 smc_lgr_unregister_conn(conn);
1182 if (!lgr->conns_num)
1183 smc_lgr_schedule_free_work(lgr);
1186 smcr_link_put(conn->lnk); /* link_hold in smc_conn_create() */
1187 smc_lgr_put(lgr); /* lgr_hold in smc_conn_create() */
1190 /* unregister a link from a buf_desc */
1191 static void smcr_buf_unmap_link(struct smc_buf_desc *buf_desc, bool is_rmb,
1192 struct smc_link *lnk)
1194 if (is_rmb || buf_desc->is_vm)
1195 buf_desc->is_reg_mr[lnk->link_idx] = false;
1196 if (!buf_desc->is_map_ib[lnk->link_idx])
1199 if ((is_rmb || buf_desc->is_vm) &&
1200 buf_desc->mr[lnk->link_idx]) {
1201 smc_ib_put_memory_region(buf_desc->mr[lnk->link_idx]);
1202 buf_desc->mr[lnk->link_idx] = NULL;
1205 smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_FROM_DEVICE);
1207 smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_TO_DEVICE);
1209 sg_free_table(&buf_desc->sgt[lnk->link_idx]);
1210 buf_desc->is_map_ib[lnk->link_idx] = false;
1213 /* unmap all buffers of lgr for a deleted link */
1214 static void smcr_buf_unmap_lgr(struct smc_link *lnk)
1216 struct smc_link_group *lgr = lnk->lgr;
1217 struct smc_buf_desc *buf_desc, *bf;
1220 for (i = 0; i < SMC_RMBE_SIZES; i++) {
1221 mutex_lock(&lgr->rmbs_lock);
1222 list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list)
1223 smcr_buf_unmap_link(buf_desc, true, lnk);
1224 mutex_unlock(&lgr->rmbs_lock);
1225 mutex_lock(&lgr->sndbufs_lock);
1226 list_for_each_entry_safe(buf_desc, bf, &lgr->sndbufs[i],
1228 smcr_buf_unmap_link(buf_desc, false, lnk);
1229 mutex_unlock(&lgr->sndbufs_lock);
1233 static void smcr_rtoken_clear_link(struct smc_link *lnk)
1235 struct smc_link_group *lgr = lnk->lgr;
1238 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1239 lgr->rtokens[i][lnk->link_idx].rkey = 0;
1240 lgr->rtokens[i][lnk->link_idx].dma_addr = 0;
1244 static void __smcr_link_clear(struct smc_link *lnk)
1246 struct smc_link_group *lgr = lnk->lgr;
1247 struct smc_ib_device *smcibdev;
1249 smc_wr_free_link_mem(lnk);
1250 smc_ibdev_cnt_dec(lnk);
1251 put_device(&lnk->smcibdev->ibdev->dev);
1252 smcibdev = lnk->smcibdev;
1253 memset(lnk, 0, sizeof(struct smc_link));
1254 lnk->state = SMC_LNK_UNUSED;
1255 if (!atomic_dec_return(&smcibdev->lnk_cnt))
1256 wake_up(&smcibdev->lnks_deleted);
1257 smc_lgr_put(lgr); /* lgr_hold in smcr_link_init() */
1260 /* must be called under lgr->llc_conf_mutex lock */
1261 void smcr_link_clear(struct smc_link *lnk, bool log)
1263 if (!lnk->lgr || lnk->clearing ||
1264 lnk->state == SMC_LNK_UNUSED)
1268 smc_llc_link_clear(lnk, log);
1269 smcr_buf_unmap_lgr(lnk);
1270 smcr_rtoken_clear_link(lnk);
1271 smc_ib_modify_qp_error(lnk);
1272 smc_wr_free_link(lnk);
1273 smc_ib_destroy_queue_pair(lnk);
1274 smc_ib_dealloc_protection_domain(lnk);
1275 smcr_link_put(lnk); /* theoretically last link_put */
1278 void smcr_link_hold(struct smc_link *lnk)
1280 refcount_inc(&lnk->refcnt);
1283 void smcr_link_put(struct smc_link *lnk)
1285 if (refcount_dec_and_test(&lnk->refcnt))
1286 __smcr_link_clear(lnk);
1289 static void smcr_buf_free(struct smc_link_group *lgr, bool is_rmb,
1290 struct smc_buf_desc *buf_desc)
1294 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
1295 smcr_buf_unmap_link(buf_desc, is_rmb, &lgr->lnk[i]);
1297 if (!buf_desc->is_vm && buf_desc->pages)
1298 __free_pages(buf_desc->pages, buf_desc->order);
1299 else if (buf_desc->is_vm && buf_desc->cpu_addr)
1300 vfree(buf_desc->cpu_addr);
1304 static void smcd_buf_free(struct smc_link_group *lgr, bool is_dmb,
1305 struct smc_buf_desc *buf_desc)
1308 /* restore original buf len */
1309 buf_desc->len += sizeof(struct smcd_cdc_msg);
1310 smc_ism_unregister_dmb(lgr->smcd, buf_desc);
1312 kfree(buf_desc->cpu_addr);
1317 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
1318 struct smc_buf_desc *buf_desc)
1321 smcd_buf_free(lgr, is_rmb, buf_desc);
1323 smcr_buf_free(lgr, is_rmb, buf_desc);
1326 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
1328 struct smc_buf_desc *buf_desc, *bf_desc;
1329 struct list_head *buf_list;
1332 for (i = 0; i < SMC_RMBE_SIZES; i++) {
1334 buf_list = &lgr->rmbs[i];
1336 buf_list = &lgr->sndbufs[i];
1337 list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
1339 list_del(&buf_desc->list);
1340 smc_buf_free(lgr, is_rmb, buf_desc);
1345 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
1347 /* free send buffers */
1348 __smc_lgr_free_bufs(lgr, false);
1350 __smc_lgr_free_bufs(lgr, true);
1353 /* won't be freed until no one accesses to lgr anymore */
1354 static void __smc_lgr_free(struct smc_link_group *lgr)
1356 smc_lgr_free_bufs(lgr);
1358 if (!atomic_dec_return(&lgr->smcd->lgr_cnt))
1359 wake_up(&lgr->smcd->lgrs_deleted);
1361 smc_wr_free_lgr_mem(lgr);
1362 if (!atomic_dec_return(&lgr_cnt))
1363 wake_up(&lgrs_deleted);
1368 /* remove a link group */
1369 static void smc_lgr_free(struct smc_link_group *lgr)
1373 if (!lgr->is_smcd) {
1374 mutex_lock(&lgr->llc_conf_mutex);
1375 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1376 if (lgr->lnk[i].state != SMC_LNK_UNUSED)
1377 smcr_link_clear(&lgr->lnk[i], false);
1379 mutex_unlock(&lgr->llc_conf_mutex);
1380 smc_llc_lgr_clear(lgr);
1383 destroy_workqueue(lgr->tx_wq);
1385 smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
1386 put_device(&lgr->smcd->dev);
1388 smc_lgr_put(lgr); /* theoretically last lgr_put */
1391 void smc_lgr_hold(struct smc_link_group *lgr)
1393 refcount_inc(&lgr->refcnt);
1396 void smc_lgr_put(struct smc_link_group *lgr)
1398 if (refcount_dec_and_test(&lgr->refcnt))
1399 __smc_lgr_free(lgr);
1402 static void smc_sk_wake_ups(struct smc_sock *smc)
1404 smc->sk.sk_write_space(&smc->sk);
1405 smc->sk.sk_data_ready(&smc->sk);
1406 smc->sk.sk_state_change(&smc->sk);
1409 /* kill a connection */
1410 static void smc_conn_kill(struct smc_connection *conn, bool soft)
1412 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1414 if (conn->lgr->is_smcd && conn->lgr->peer_shutdown)
1415 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
1417 smc_close_abort(conn);
1419 smc->sk.sk_err = ECONNABORTED;
1420 smc_sk_wake_ups(smc);
1421 if (conn->lgr->is_smcd) {
1422 smc_ism_unset_conn(conn);
1424 tasklet_kill(&conn->rx_tsklet);
1426 tasklet_unlock_wait(&conn->rx_tsklet);
1428 smc_cdc_wait_pend_tx_wr(conn);
1430 smc_lgr_unregister_conn(conn);
1431 smc_close_active_abort(smc);
1434 static void smc_lgr_cleanup(struct smc_link_group *lgr)
1437 smc_ism_signal_shutdown(lgr);
1439 u32 rsn = lgr->llc_termination_rsn;
1442 rsn = SMC_LLC_DEL_PROG_INIT_TERM;
1443 smc_llc_send_link_delete_all(lgr, false, rsn);
1444 smcr_lgr_link_deactivate_all(lgr);
1448 /* terminate link group
1449 * @soft: true if link group shutdown can take its time
1450 * false if immediate link group shutdown is required
1452 static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft)
1454 struct smc_connection *conn;
1455 struct smc_sock *smc;
1456 struct rb_node *node;
1458 if (lgr->terminating)
1459 return; /* lgr already terminating */
1460 /* cancel free_work sync, will terminate when lgr->freeing is set */
1461 cancel_delayed_work_sync(&lgr->free_work);
1462 lgr->terminating = 1;
1464 /* kill remaining link group connections */
1465 read_lock_bh(&lgr->conns_lock);
1466 node = rb_first(&lgr->conns_all);
1468 read_unlock_bh(&lgr->conns_lock);
1469 conn = rb_entry(node, struct smc_connection, alert_node);
1470 smc = container_of(conn, struct smc_sock, conn);
1471 sock_hold(&smc->sk); /* sock_put below */
1472 lock_sock(&smc->sk);
1473 smc_conn_kill(conn, soft);
1474 release_sock(&smc->sk);
1475 sock_put(&smc->sk); /* sock_hold above */
1476 read_lock_bh(&lgr->conns_lock);
1477 node = rb_first(&lgr->conns_all);
1479 read_unlock_bh(&lgr->conns_lock);
1480 smc_lgr_cleanup(lgr);
1484 /* unlink link group and schedule termination */
1485 void smc_lgr_terminate_sched(struct smc_link_group *lgr)
1487 spinlock_t *lgr_lock;
1489 smc_lgr_list_head(lgr, &lgr_lock);
1490 spin_lock_bh(lgr_lock);
1491 if (list_empty(&lgr->list) || lgr->terminating || lgr->freeing) {
1492 spin_unlock_bh(lgr_lock);
1493 return; /* lgr already terminating */
1495 list_del_init(&lgr->list);
1497 spin_unlock_bh(lgr_lock);
1498 schedule_work(&lgr->terminate_work);
1501 /* Called when peer lgr shutdown (regularly or abnormally) is received */
1502 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid, unsigned short vlan)
1504 struct smc_link_group *lgr, *l;
1505 LIST_HEAD(lgr_free_list);
1507 /* run common cleanup function and build free list */
1508 spin_lock_bh(&dev->lgr_lock);
1509 list_for_each_entry_safe(lgr, l, &dev->lgr_list, list) {
1510 if ((!peer_gid || lgr->peer_gid == peer_gid) &&
1511 (vlan == VLAN_VID_MASK || lgr->vlan_id == vlan)) {
1512 if (peer_gid) /* peer triggered termination */
1513 lgr->peer_shutdown = 1;
1514 list_move(&lgr->list, &lgr_free_list);
1518 spin_unlock_bh(&dev->lgr_lock);
1520 /* cancel the regular free workers and actually free lgrs */
1521 list_for_each_entry_safe(lgr, l, &lgr_free_list, list) {
1522 list_del_init(&lgr->list);
1523 schedule_work(&lgr->terminate_work);
1527 /* Called when an SMCD device is removed or the smc module is unloaded */
1528 void smc_smcd_terminate_all(struct smcd_dev *smcd)
1530 struct smc_link_group *lgr, *lg;
1531 LIST_HEAD(lgr_free_list);
1533 spin_lock_bh(&smcd->lgr_lock);
1534 list_splice_init(&smcd->lgr_list, &lgr_free_list);
1535 list_for_each_entry(lgr, &lgr_free_list, list)
1537 spin_unlock_bh(&smcd->lgr_lock);
1539 list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
1540 list_del_init(&lgr->list);
1541 __smc_lgr_terminate(lgr, false);
1544 if (atomic_read(&smcd->lgr_cnt))
1545 wait_event(smcd->lgrs_deleted, !atomic_read(&smcd->lgr_cnt));
1548 /* Called when an SMCR device is removed or the smc module is unloaded.
1549 * If smcibdev is given, all SMCR link groups using this device are terminated.
1550 * If smcibdev is NULL, all SMCR link groups are terminated.
1552 void smc_smcr_terminate_all(struct smc_ib_device *smcibdev)
1554 struct smc_link_group *lgr, *lg;
1555 LIST_HEAD(lgr_free_list);
1558 spin_lock_bh(&smc_lgr_list.lock);
1560 list_splice_init(&smc_lgr_list.list, &lgr_free_list);
1561 list_for_each_entry(lgr, &lgr_free_list, list)
1564 list_for_each_entry_safe(lgr, lg, &smc_lgr_list.list, list) {
1565 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1566 if (lgr->lnk[i].smcibdev == smcibdev)
1567 smcr_link_down_cond_sched(&lgr->lnk[i]);
1571 spin_unlock_bh(&smc_lgr_list.lock);
1573 list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
1574 list_del_init(&lgr->list);
1575 smc_llc_set_termination_rsn(lgr, SMC_LLC_DEL_OP_INIT_TERM);
1576 __smc_lgr_terminate(lgr, false);
1580 if (atomic_read(&smcibdev->lnk_cnt))
1581 wait_event(smcibdev->lnks_deleted,
1582 !atomic_read(&smcibdev->lnk_cnt));
1584 if (atomic_read(&lgr_cnt))
1585 wait_event(lgrs_deleted, !atomic_read(&lgr_cnt));
1589 /* set new lgr type and clear all asymmetric link tagging */
1590 void smcr_lgr_set_type(struct smc_link_group *lgr, enum smc_lgr_type new_type)
1592 char *lgr_type = "";
1595 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
1596 if (smc_link_usable(&lgr->lnk[i]))
1597 lgr->lnk[i].link_is_asym = false;
1598 if (lgr->type == new_type)
1600 lgr->type = new_type;
1602 switch (lgr->type) {
1606 case SMC_LGR_SINGLE:
1607 lgr_type = "SINGLE";
1609 case SMC_LGR_SYMMETRIC:
1610 lgr_type = "SYMMETRIC";
1612 case SMC_LGR_ASYMMETRIC_PEER:
1613 lgr_type = "ASYMMETRIC_PEER";
1615 case SMC_LGR_ASYMMETRIC_LOCAL:
1616 lgr_type = "ASYMMETRIC_LOCAL";
1619 pr_warn_ratelimited("smc: SMC-R lg %*phN net %llu state changed: "
1620 "%s, pnetid %.16s\n", SMC_LGR_ID_SIZE, &lgr->id,
1621 lgr->net->net_cookie, lgr_type, lgr->pnet_id);
1624 /* set new lgr type and tag a link as asymmetric */
1625 void smcr_lgr_set_type_asym(struct smc_link_group *lgr,
1626 enum smc_lgr_type new_type, int asym_lnk_idx)
1628 smcr_lgr_set_type(lgr, new_type);
1629 lgr->lnk[asym_lnk_idx].link_is_asym = true;
1632 /* abort connection, abort_work scheduled from tasklet context */
1633 static void smc_conn_abort_work(struct work_struct *work)
1635 struct smc_connection *conn = container_of(work,
1636 struct smc_connection,
1638 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1640 lock_sock(&smc->sk);
1641 smc_conn_kill(conn, true);
1642 release_sock(&smc->sk);
1643 sock_put(&smc->sk); /* sock_hold done by schedulers of abort_work */
1646 void smcr_port_add(struct smc_ib_device *smcibdev, u8 ibport)
1648 struct smc_link_group *lgr, *n;
1650 list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1651 struct smc_link *link;
1653 if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1654 SMC_MAX_PNETID_LEN) ||
1655 lgr->type == SMC_LGR_SYMMETRIC ||
1656 lgr->type == SMC_LGR_ASYMMETRIC_PEER ||
1657 !rdma_dev_access_netns(smcibdev->ibdev, lgr->net))
1660 /* trigger local add link processing */
1661 link = smc_llc_usable_link(lgr);
1663 smc_llc_add_link_local(link);
1667 /* link is down - switch connections to alternate link,
1668 * must be called under lgr->llc_conf_mutex lock
1670 static void smcr_link_down(struct smc_link *lnk)
1672 struct smc_link_group *lgr = lnk->lgr;
1673 struct smc_link *to_lnk;
1676 if (!lgr || lnk->state == SMC_LNK_UNUSED || list_empty(&lgr->list))
1679 to_lnk = smc_switch_conns(lgr, lnk, true);
1680 if (!to_lnk) { /* no backup link available */
1681 smcr_link_clear(lnk, true);
1684 smcr_lgr_set_type(lgr, SMC_LGR_SINGLE);
1685 del_link_id = lnk->link_id;
1687 if (lgr->role == SMC_SERV) {
1688 /* trigger local delete link processing */
1689 smc_llc_srv_delete_link_local(to_lnk, del_link_id);
1691 if (lgr->llc_flow_lcl.type != SMC_LLC_FLOW_NONE) {
1692 /* another llc task is ongoing */
1693 mutex_unlock(&lgr->llc_conf_mutex);
1694 wait_event_timeout(lgr->llc_flow_waiter,
1695 (list_empty(&lgr->list) ||
1696 lgr->llc_flow_lcl.type == SMC_LLC_FLOW_NONE),
1698 mutex_lock(&lgr->llc_conf_mutex);
1700 if (!list_empty(&lgr->list)) {
1701 smc_llc_send_delete_link(to_lnk, del_link_id,
1703 SMC_LLC_DEL_LOST_PATH);
1704 smcr_link_clear(lnk, true);
1706 wake_up(&lgr->llc_flow_waiter); /* wake up next waiter */
1710 /* must be called under lgr->llc_conf_mutex lock */
1711 void smcr_link_down_cond(struct smc_link *lnk)
1713 if (smc_link_downing(&lnk->state)) {
1714 trace_smcr_link_down(lnk, __builtin_return_address(0));
1715 smcr_link_down(lnk);
1719 /* will get the lgr->llc_conf_mutex lock */
1720 void smcr_link_down_cond_sched(struct smc_link *lnk)
1722 if (smc_link_downing(&lnk->state)) {
1723 trace_smcr_link_down(lnk, __builtin_return_address(0));
1724 schedule_work(&lnk->link_down_wrk);
1728 void smcr_port_err(struct smc_ib_device *smcibdev, u8 ibport)
1730 struct smc_link_group *lgr, *n;
1733 list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1734 if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1735 SMC_MAX_PNETID_LEN))
1736 continue; /* lgr is not affected */
1737 if (list_empty(&lgr->list))
1739 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1740 struct smc_link *lnk = &lgr->lnk[i];
1742 if (smc_link_usable(lnk) &&
1743 lnk->smcibdev == smcibdev && lnk->ibport == ibport)
1744 smcr_link_down_cond_sched(lnk);
1749 static void smc_link_down_work(struct work_struct *work)
1751 struct smc_link *link = container_of(work, struct smc_link,
1753 struct smc_link_group *lgr = link->lgr;
1755 if (list_empty(&lgr->list))
1757 wake_up_all(&lgr->llc_msg_waiter);
1758 mutex_lock(&lgr->llc_conf_mutex);
1759 smcr_link_down(link);
1760 mutex_unlock(&lgr->llc_conf_mutex);
1763 static int smc_vlan_by_tcpsk_walk(struct net_device *lower_dev,
1764 struct netdev_nested_priv *priv)
1766 unsigned short *vlan_id = (unsigned short *)priv->data;
1768 if (is_vlan_dev(lower_dev)) {
1769 *vlan_id = vlan_dev_vlan_id(lower_dev);
1776 /* Determine vlan of internal TCP socket. */
1777 int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini)
1779 struct dst_entry *dst = sk_dst_get(clcsock->sk);
1780 struct netdev_nested_priv priv;
1781 struct net_device *ndev;
1795 if (is_vlan_dev(ndev)) {
1796 ini->vlan_id = vlan_dev_vlan_id(ndev);
1800 priv.data = (void *)&ini->vlan_id;
1802 netdev_walk_all_lower_dev(ndev, smc_vlan_by_tcpsk_walk, &priv);
1811 static bool smcr_lgr_match(struct smc_link_group *lgr, u8 smcr_version,
1815 enum smc_lgr_role role, u32 clcqpn,
1818 struct smc_link *lnk;
1821 if (memcmp(lgr->peer_systemid, peer_systemid, SMC_SYSTEMID_LEN) ||
1825 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1828 if (!smc_link_active(lnk))
1830 /* use verbs API to check netns, instead of lgr->net */
1831 if (!rdma_dev_access_netns(lnk->smcibdev->ibdev, net))
1833 if ((lgr->role == SMC_SERV || lnk->peer_qpn == clcqpn) &&
1834 !memcmp(lnk->peer_gid, peer_gid, SMC_GID_SIZE) &&
1835 (smcr_version == SMC_V2 ||
1836 !memcmp(lnk->peer_mac, peer_mac_v1, ETH_ALEN)))
1842 static bool smcd_lgr_match(struct smc_link_group *lgr,
1843 struct smcd_dev *smcismdev, u64 peer_gid)
1845 return lgr->peer_gid == peer_gid && lgr->smcd == smcismdev;
1848 /* create a new SMC connection (and a new link group if necessary) */
1849 int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini)
1851 struct smc_connection *conn = &smc->conn;
1852 struct net *net = sock_net(&smc->sk);
1853 struct list_head *lgr_list;
1854 struct smc_link_group *lgr;
1855 enum smc_lgr_role role;
1856 spinlock_t *lgr_lock;
1859 lgr_list = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_list :
1861 lgr_lock = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_lock :
1863 ini->first_contact_local = 1;
1864 role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
1865 if (role == SMC_CLNT && ini->first_contact_peer)
1866 /* create new link group as well */
1869 /* determine if an existing link group can be reused */
1870 spin_lock_bh(lgr_lock);
1871 list_for_each_entry(lgr, lgr_list, list) {
1872 write_lock_bh(&lgr->conns_lock);
1874 smcd_lgr_match(lgr, ini->ism_dev[ini->ism_selected],
1875 ini->ism_peer_gid[ini->ism_selected]) :
1876 smcr_lgr_match(lgr, ini->smcr_version,
1878 ini->peer_gid, ini->peer_mac, role,
1879 ini->ib_clcqpn, net)) &&
1881 (ini->smcd_version == SMC_V2 ||
1882 lgr->vlan_id == ini->vlan_id) &&
1883 (role == SMC_CLNT || ini->is_smcd ||
1884 (lgr->conns_num < SMC_RMBS_PER_LGR_MAX &&
1885 !bitmap_full(lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX)))) {
1886 /* link group found */
1887 ini->first_contact_local = 0;
1889 rc = smc_lgr_register_conn(conn, false);
1890 write_unlock_bh(&lgr->conns_lock);
1891 if (!rc && delayed_work_pending(&lgr->free_work))
1892 cancel_delayed_work(&lgr->free_work);
1895 write_unlock_bh(&lgr->conns_lock);
1897 spin_unlock_bh(lgr_lock);
1901 if (role == SMC_CLNT && !ini->first_contact_peer &&
1902 ini->first_contact_local) {
1903 /* Server reuses a link group, but Client wants to start
1905 * send out_of_sync decline, reason synchr. error
1907 return SMC_CLC_DECL_SYNCERR;
1911 if (ini->first_contact_local) {
1912 rc = smc_lgr_create(smc, ini);
1916 write_lock_bh(&lgr->conns_lock);
1917 rc = smc_lgr_register_conn(conn, true);
1918 write_unlock_bh(&lgr->conns_lock);
1920 smc_lgr_cleanup_early(lgr);
1924 smc_lgr_hold(conn->lgr); /* lgr_put in smc_conn_free() */
1925 if (!conn->lgr->is_smcd)
1926 smcr_link_hold(conn->lnk); /* link_put in smc_conn_free() */
1928 conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
1929 conn->local_tx_ctrl.len = SMC_WR_TX_SIZE;
1930 conn->urg_state = SMC_URG_READ;
1931 init_waitqueue_head(&conn->cdc_pend_tx_wq);
1932 INIT_WORK(&smc->conn.abort_work, smc_conn_abort_work);
1934 conn->rx_off = sizeof(struct smcd_cdc_msg);
1935 smcd_cdc_rx_init(conn); /* init tasklet for this conn */
1939 #ifndef KERNEL_HAS_ATOMIC64
1940 spin_lock_init(&conn->acurs_lock);
1947 #define SMCD_DMBE_SIZES 6 /* 0 -> 16KB, 1 -> 32KB, .. 6 -> 1MB */
1948 #define SMCR_RMBE_SIZES 5 /* 0 -> 16KB, 1 -> 32KB, .. 5 -> 512KB */
1950 /* convert the RMB size into the compressed notation (minimum 16K, see
1951 * SMCD/R_DMBE_SIZES.
1952 * In contrast to plain ilog2, this rounds towards the next power of 2,
1953 * so the socket application gets at least its desired sndbuf / rcvbuf size.
1955 static u8 smc_compress_bufsize(int size, bool is_smcd, bool is_rmb)
1957 const unsigned int max_scat = SG_MAX_SINGLE_ALLOC * PAGE_SIZE;
1960 if (size <= SMC_BUF_MIN_SIZE)
1963 size = (size - 1) >> 14; /* convert to 16K multiple */
1964 compressed = min_t(u8, ilog2(size) + 1,
1965 is_smcd ? SMCD_DMBE_SIZES : SMCR_RMBE_SIZES);
1967 if (!is_smcd && is_rmb)
1968 /* RMBs are backed by & limited to max size of scatterlists */
1969 compressed = min_t(u8, compressed, ilog2(max_scat >> 14));
1974 /* convert the RMB size from compressed notation into integer */
1975 int smc_uncompress_bufsize(u8 compressed)
1979 size = 0x00000001 << (((int)compressed) + 14);
1983 /* try to reuse a sndbuf or rmb description slot for a certain
1984 * buffer size; if not available, return NULL
1986 static struct smc_buf_desc *smc_buf_get_slot(int compressed_bufsize,
1988 struct list_head *buf_list)
1990 struct smc_buf_desc *buf_slot;
1993 list_for_each_entry(buf_slot, buf_list, list) {
1994 if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
2003 /* one of the conditions for announcing a receiver's current window size is
2004 * that it "results in a minimum increase in the window size of 10% of the
2005 * receive buffer space" [RFC7609]
2007 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
2009 return max_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
2012 /* map an buf to a link */
2013 static int smcr_buf_map_link(struct smc_buf_desc *buf_desc, bool is_rmb,
2014 struct smc_link *lnk)
2016 int rc, i, nents, offset, buf_size, size, access_flags;
2017 struct scatterlist *sg;
2020 if (buf_desc->is_map_ib[lnk->link_idx])
2023 if (buf_desc->is_vm) {
2024 buf = buf_desc->cpu_addr;
2025 buf_size = buf_desc->len;
2026 offset = offset_in_page(buf_desc->cpu_addr);
2027 nents = PAGE_ALIGN(buf_size + offset) / PAGE_SIZE;
2032 rc = sg_alloc_table(&buf_desc->sgt[lnk->link_idx], nents, GFP_KERNEL);
2036 if (buf_desc->is_vm) {
2037 /* virtually contiguous buffer */
2038 for_each_sg(buf_desc->sgt[lnk->link_idx].sgl, sg, nents, i) {
2039 size = min_t(int, PAGE_SIZE - offset, buf_size);
2040 sg_set_page(sg, vmalloc_to_page(buf), size, offset);
2041 buf += size / sizeof(*buf);
2046 /* physically contiguous buffer */
2047 sg_set_buf(buf_desc->sgt[lnk->link_idx].sgl,
2048 buf_desc->cpu_addr, buf_desc->len);
2051 /* map sg table to DMA address */
2052 rc = smc_ib_buf_map_sg(lnk, buf_desc,
2053 is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
2054 /* SMC protocol depends on mapping to one DMA address only */
2060 buf_desc->is_dma_need_sync |=
2061 smc_ib_is_sg_need_sync(lnk, buf_desc) << lnk->link_idx;
2063 if (is_rmb || buf_desc->is_vm) {
2064 /* create a new memory region for the RMB or vzalloced sndbuf */
2065 access_flags = is_rmb ?
2066 IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE :
2067 IB_ACCESS_LOCAL_WRITE;
2069 rc = smc_ib_get_memory_region(lnk->roce_pd, access_flags,
2070 buf_desc, lnk->link_idx);
2073 smc_ib_sync_sg_for_device(lnk, buf_desc,
2074 is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
2076 buf_desc->is_map_ib[lnk->link_idx] = true;
2080 smc_ib_buf_unmap_sg(lnk, buf_desc,
2081 is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
2083 sg_free_table(&buf_desc->sgt[lnk->link_idx]);
2087 /* register a new buf on IB device, rmb or vzalloced sndbuf
2088 * must be called under lgr->llc_conf_mutex lock
2090 int smcr_link_reg_buf(struct smc_link *link, struct smc_buf_desc *buf_desc)
2092 if (list_empty(&link->lgr->list))
2094 if (!buf_desc->is_reg_mr[link->link_idx]) {
2095 /* register memory region for new buf */
2096 if (buf_desc->is_vm)
2097 buf_desc->mr[link->link_idx]->iova =
2098 (uintptr_t)buf_desc->cpu_addr;
2099 if (smc_wr_reg_send(link, buf_desc->mr[link->link_idx])) {
2100 buf_desc->is_reg_err = true;
2103 buf_desc->is_reg_mr[link->link_idx] = true;
2108 static int _smcr_buf_map_lgr(struct smc_link *lnk, struct mutex *lock,
2109 struct list_head *lst, bool is_rmb)
2111 struct smc_buf_desc *buf_desc, *bf;
2115 list_for_each_entry_safe(buf_desc, bf, lst, list) {
2116 if (!buf_desc->used)
2118 rc = smcr_buf_map_link(buf_desc, is_rmb, lnk);
2127 /* map all used buffers of lgr for a new link */
2128 int smcr_buf_map_lgr(struct smc_link *lnk)
2130 struct smc_link_group *lgr = lnk->lgr;
2133 for (i = 0; i < SMC_RMBE_SIZES; i++) {
2134 rc = _smcr_buf_map_lgr(lnk, &lgr->rmbs_lock,
2135 &lgr->rmbs[i], true);
2138 rc = _smcr_buf_map_lgr(lnk, &lgr->sndbufs_lock,
2139 &lgr->sndbufs[i], false);
2146 /* register all used buffers of lgr for a new link,
2147 * must be called under lgr->llc_conf_mutex lock
2149 int smcr_buf_reg_lgr(struct smc_link *lnk)
2151 struct smc_link_group *lgr = lnk->lgr;
2152 struct smc_buf_desc *buf_desc, *bf;
2155 /* reg all RMBs for a new link */
2156 mutex_lock(&lgr->rmbs_lock);
2157 for (i = 0; i < SMC_RMBE_SIZES; i++) {
2158 list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list) {
2159 if (!buf_desc->used)
2161 rc = smcr_link_reg_buf(lnk, buf_desc);
2163 mutex_unlock(&lgr->rmbs_lock);
2168 mutex_unlock(&lgr->rmbs_lock);
2170 if (lgr->buf_type == SMCR_PHYS_CONT_BUFS)
2173 /* reg all vzalloced sndbufs for a new link */
2174 mutex_lock(&lgr->sndbufs_lock);
2175 for (i = 0; i < SMC_RMBE_SIZES; i++) {
2176 list_for_each_entry_safe(buf_desc, bf, &lgr->sndbufs[i], list) {
2177 if (!buf_desc->used || !buf_desc->is_vm)
2179 rc = smcr_link_reg_buf(lnk, buf_desc);
2181 mutex_unlock(&lgr->sndbufs_lock);
2186 mutex_unlock(&lgr->sndbufs_lock);
2190 static struct smc_buf_desc *smcr_new_buf_create(struct smc_link_group *lgr,
2191 bool is_rmb, int bufsize)
2193 struct smc_buf_desc *buf_desc;
2195 /* try to alloc a new buffer */
2196 buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
2198 return ERR_PTR(-ENOMEM);
2200 switch (lgr->buf_type) {
2201 case SMCR_PHYS_CONT_BUFS:
2202 case SMCR_MIXED_BUFS:
2203 buf_desc->order = get_order(bufsize);
2204 buf_desc->pages = alloc_pages(GFP_KERNEL | __GFP_NOWARN |
2205 __GFP_NOMEMALLOC | __GFP_COMP |
2206 __GFP_NORETRY | __GFP_ZERO,
2208 if (buf_desc->pages) {
2209 buf_desc->cpu_addr =
2210 (void *)page_address(buf_desc->pages);
2211 buf_desc->len = bufsize;
2212 buf_desc->is_vm = false;
2215 if (lgr->buf_type == SMCR_PHYS_CONT_BUFS)
2217 fallthrough; // try virtually continguous buf
2218 case SMCR_VIRT_CONT_BUFS:
2219 buf_desc->order = get_order(bufsize);
2220 buf_desc->cpu_addr = vzalloc(PAGE_SIZE << buf_desc->order);
2221 if (!buf_desc->cpu_addr)
2223 buf_desc->pages = NULL;
2224 buf_desc->len = bufsize;
2225 buf_desc->is_vm = true;
2232 return ERR_PTR(-EAGAIN);
2235 /* map buf_desc on all usable links,
2236 * unused buffers stay mapped as long as the link is up
2238 static int smcr_buf_map_usable_links(struct smc_link_group *lgr,
2239 struct smc_buf_desc *buf_desc, bool is_rmb)
2243 /* protect against parallel link reconfiguration */
2244 mutex_lock(&lgr->llc_conf_mutex);
2245 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
2246 struct smc_link *lnk = &lgr->lnk[i];
2248 if (!smc_link_usable(lnk))
2250 if (smcr_buf_map_link(buf_desc, is_rmb, lnk)) {
2256 mutex_unlock(&lgr->llc_conf_mutex);
2260 static struct smc_buf_desc *smcd_new_buf_create(struct smc_link_group *lgr,
2261 bool is_dmb, int bufsize)
2263 struct smc_buf_desc *buf_desc;
2266 /* try to alloc a new DMB */
2267 buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
2269 return ERR_PTR(-ENOMEM);
2271 rc = smc_ism_register_dmb(lgr, bufsize, buf_desc);
2275 return ERR_PTR(-EAGAIN);
2277 return ERR_PTR(-ENOSPC);
2278 return ERR_PTR(-EIO);
2280 buf_desc->pages = virt_to_page(buf_desc->cpu_addr);
2281 /* CDC header stored in buf. So, pretend it was smaller */
2282 buf_desc->len = bufsize - sizeof(struct smcd_cdc_msg);
2284 buf_desc->cpu_addr = kzalloc(bufsize, GFP_KERNEL |
2285 __GFP_NOWARN | __GFP_NORETRY |
2287 if (!buf_desc->cpu_addr) {
2289 return ERR_PTR(-EAGAIN);
2291 buf_desc->len = bufsize;
2296 static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
2298 struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
2299 struct smc_connection *conn = &smc->conn;
2300 struct smc_link_group *lgr = conn->lgr;
2301 struct list_head *buf_list;
2302 int bufsize, bufsize_short;
2303 bool is_dgraded = false;
2304 struct mutex *lock; /* lock buffer list */
2308 /* use socket recv buffer size (w/o overhead) as start value */
2309 sk_buf_size = smc->sk.sk_rcvbuf / 2;
2311 /* use socket send buffer size (w/o overhead) as start value */
2312 sk_buf_size = smc->sk.sk_sndbuf / 2;
2314 for (bufsize_short = smc_compress_bufsize(sk_buf_size, is_smcd, is_rmb);
2315 bufsize_short >= 0; bufsize_short--) {
2317 lock = &lgr->rmbs_lock;
2318 buf_list = &lgr->rmbs[bufsize_short];
2320 lock = &lgr->sndbufs_lock;
2321 buf_list = &lgr->sndbufs[bufsize_short];
2323 bufsize = smc_uncompress_bufsize(bufsize_short);
2325 /* check for reusable slot in the link group */
2326 buf_desc = smc_buf_get_slot(bufsize_short, lock, buf_list);
2328 buf_desc->is_dma_need_sync = 0;
2329 SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, bufsize);
2330 SMC_STAT_BUF_REUSE(smc, is_smcd, is_rmb);
2331 break; /* found reusable slot */
2335 buf_desc = smcd_new_buf_create(lgr, is_rmb, bufsize);
2337 buf_desc = smcr_new_buf_create(lgr, is_rmb, bufsize);
2339 if (PTR_ERR(buf_desc) == -ENOMEM)
2341 if (IS_ERR(buf_desc)) {
2344 SMC_STAT_RMB_DOWNGRADED(smc, is_smcd, is_rmb);
2349 SMC_STAT_RMB_ALLOC(smc, is_smcd, is_rmb);
2350 SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, bufsize);
2353 list_add(&buf_desc->list, buf_list);
2358 if (IS_ERR(buf_desc))
2359 return PTR_ERR(buf_desc);
2362 if (smcr_buf_map_usable_links(lgr, buf_desc, is_rmb)) {
2363 smcr_buf_unuse(buf_desc, is_rmb, lgr);
2369 conn->rmb_desc = buf_desc;
2370 conn->rmbe_size_short = bufsize_short;
2371 smc->sk.sk_rcvbuf = bufsize * 2;
2372 atomic_set(&conn->bytes_to_rcv, 0);
2373 conn->rmbe_update_limit =
2374 smc_rmb_wnd_update_limit(buf_desc->len);
2376 smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */
2378 conn->sndbuf_desc = buf_desc;
2379 smc->sk.sk_sndbuf = bufsize * 2;
2380 atomic_set(&conn->sndbuf_space, bufsize);
2385 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
2387 if (!conn->sndbuf_desc->is_dma_need_sync)
2389 if (!smc_conn_lgr_valid(conn) || conn->lgr->is_smcd ||
2390 !smc_link_active(conn->lnk))
2392 smc_ib_sync_sg_for_device(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE);
2395 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
2399 if (!conn->rmb_desc->is_dma_need_sync)
2401 if (!smc_conn_lgr_valid(conn) || conn->lgr->is_smcd)
2403 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
2404 if (!smc_link_active(&conn->lgr->lnk[i]))
2406 smc_ib_sync_sg_for_cpu(&conn->lgr->lnk[i], conn->rmb_desc,
2411 /* create the send and receive buffer for an SMC socket;
2412 * receive buffers are called RMBs;
2413 * (even though the SMC protocol allows more than one RMB-element per RMB,
2414 * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
2415 * extra RMB for every connection in a link group
2417 int smc_buf_create(struct smc_sock *smc, bool is_smcd)
2421 /* create send buffer */
2422 rc = __smc_buf_create(smc, is_smcd, false);
2426 rc = __smc_buf_create(smc, is_smcd, true);
2428 mutex_lock(&smc->conn.lgr->sndbufs_lock);
2429 list_del(&smc->conn.sndbuf_desc->list);
2430 mutex_unlock(&smc->conn.lgr->sndbufs_lock);
2431 smc_buf_free(smc->conn.lgr, false, smc->conn.sndbuf_desc);
2432 smc->conn.sndbuf_desc = NULL;
2437 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
2441 for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
2442 if (!test_and_set_bit(i, lgr->rtokens_used_mask))
2448 static int smc_rtoken_find_by_link(struct smc_link_group *lgr, int lnk_idx,
2453 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2454 if (test_bit(i, lgr->rtokens_used_mask) &&
2455 lgr->rtokens[i][lnk_idx].rkey == rkey)
2461 /* set rtoken for a new link to an existing rmb */
2462 void smc_rtoken_set(struct smc_link_group *lgr, int link_idx, int link_idx_new,
2463 __be32 nw_rkey_known, __be64 nw_vaddr, __be32 nw_rkey)
2467 rtok_idx = smc_rtoken_find_by_link(lgr, link_idx, ntohl(nw_rkey_known));
2468 if (rtok_idx == -ENOENT)
2470 lgr->rtokens[rtok_idx][link_idx_new].rkey = ntohl(nw_rkey);
2471 lgr->rtokens[rtok_idx][link_idx_new].dma_addr = be64_to_cpu(nw_vaddr);
2474 /* set rtoken for a new link whose link_id is given */
2475 void smc_rtoken_set2(struct smc_link_group *lgr, int rtok_idx, int link_id,
2476 __be64 nw_vaddr, __be32 nw_rkey)
2478 u64 dma_addr = be64_to_cpu(nw_vaddr);
2479 u32 rkey = ntohl(nw_rkey);
2483 for (link_idx = 0; link_idx < SMC_LINKS_PER_LGR_MAX; link_idx++) {
2484 if (lgr->lnk[link_idx].link_id == link_id) {
2491 lgr->rtokens[rtok_idx][link_idx].rkey = rkey;
2492 lgr->rtokens[rtok_idx][link_idx].dma_addr = dma_addr;
2495 /* add a new rtoken from peer */
2496 int smc_rtoken_add(struct smc_link *lnk, __be64 nw_vaddr, __be32 nw_rkey)
2498 struct smc_link_group *lgr = smc_get_lgr(lnk);
2499 u64 dma_addr = be64_to_cpu(nw_vaddr);
2500 u32 rkey = ntohl(nw_rkey);
2503 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2504 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
2505 lgr->rtokens[i][lnk->link_idx].dma_addr == dma_addr &&
2506 test_bit(i, lgr->rtokens_used_mask)) {
2507 /* already in list */
2511 i = smc_rmb_reserve_rtoken_idx(lgr);
2514 lgr->rtokens[i][lnk->link_idx].rkey = rkey;
2515 lgr->rtokens[i][lnk->link_idx].dma_addr = dma_addr;
2519 /* delete an rtoken from all links */
2520 int smc_rtoken_delete(struct smc_link *lnk, __be32 nw_rkey)
2522 struct smc_link_group *lgr = smc_get_lgr(lnk);
2523 u32 rkey = ntohl(nw_rkey);
2526 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2527 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
2528 test_bit(i, lgr->rtokens_used_mask)) {
2529 for (j = 0; j < SMC_LINKS_PER_LGR_MAX; j++) {
2530 lgr->rtokens[i][j].rkey = 0;
2531 lgr->rtokens[i][j].dma_addr = 0;
2533 clear_bit(i, lgr->rtokens_used_mask);
2540 /* save rkey and dma_addr received from peer during clc handshake */
2541 int smc_rmb_rtoken_handling(struct smc_connection *conn,
2542 struct smc_link *lnk,
2543 struct smc_clc_msg_accept_confirm *clc)
2545 conn->rtoken_idx = smc_rtoken_add(lnk, clc->r0.rmb_dma_addr,
2547 if (conn->rtoken_idx < 0)
2548 return conn->rtoken_idx;
2552 static void smc_core_going_away(void)
2554 struct smc_ib_device *smcibdev;
2555 struct smcd_dev *smcd;
2557 mutex_lock(&smc_ib_devices.mutex);
2558 list_for_each_entry(smcibdev, &smc_ib_devices.list, list) {
2561 for (i = 0; i < SMC_MAX_PORTS; i++)
2562 set_bit(i, smcibdev->ports_going_away);
2564 mutex_unlock(&smc_ib_devices.mutex);
2566 mutex_lock(&smcd_dev_list.mutex);
2567 list_for_each_entry(smcd, &smcd_dev_list.list, list) {
2568 smcd->going_away = 1;
2570 mutex_unlock(&smcd_dev_list.mutex);
2573 /* Clean up all SMC link groups */
2574 static void smc_lgrs_shutdown(void)
2576 struct smcd_dev *smcd;
2578 smc_core_going_away();
2580 smc_smcr_terminate_all(NULL);
2582 mutex_lock(&smcd_dev_list.mutex);
2583 list_for_each_entry(smcd, &smcd_dev_list.list, list)
2584 smc_smcd_terminate_all(smcd);
2585 mutex_unlock(&smcd_dev_list.mutex);
2588 static int smc_core_reboot_event(struct notifier_block *this,
2589 unsigned long event, void *ptr)
2591 smc_lgrs_shutdown();
2592 smc_ib_unregister_client();
2596 static struct notifier_block smc_reboot_notifier = {
2597 .notifier_call = smc_core_reboot_event,
2600 int __init smc_core_init(void)
2602 return register_reboot_notifier(&smc_reboot_notifier);
2605 /* Called (from smc_exit) when module is removed */
2606 void smc_core_exit(void)
2608 unregister_reboot_notifier(&smc_reboot_notifier);
2609 smc_lgrs_shutdown();