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_VLAN_ID, lgr->vlan_id))
352 if (nla_put_u64_64bit(skb, SMC_NLA_LGR_R_NET_COOKIE,
353 lgr->net->net_cookie, SMC_NLA_LGR_R_PAD))
355 memcpy(smc_target, lgr->pnet_id, SMC_MAX_PNETID_LEN);
356 smc_target[SMC_MAX_PNETID_LEN] = 0;
357 if (nla_put_string(skb, SMC_NLA_LGR_R_PNETID, smc_target))
359 if (lgr->smc_version > SMC_V1) {
360 v2_attrs = nla_nest_start(skb, SMC_NLA_LGR_R_V2_COMMON);
363 if (smc_nl_fill_lgr_v2_common(lgr, skb, cb, v2_attrs))
365 if (smc_nl_fill_smcr_lgr_v2(lgr, skb, cb))
369 nla_nest_end(skb, attrs);
372 nla_nest_cancel(skb, attrs);
377 static int smc_nl_fill_lgr_link(struct smc_link_group *lgr,
378 struct smc_link *link,
380 struct netlink_callback *cb)
382 char smc_ibname[IB_DEVICE_NAME_MAX];
383 u8 smc_gid_target[41];
384 struct nlattr *attrs;
388 nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
389 &smc_gen_nl_family, NLM_F_MULTI,
390 SMC_NETLINK_GET_LINK_SMCR);
394 attrs = nla_nest_start(skb, SMC_GEN_LINK_SMCR);
398 if (nla_put_u8(skb, SMC_NLA_LINK_ID, link->link_id))
400 if (nla_put_u32(skb, SMC_NLA_LINK_STATE, link->state))
402 if (nla_put_u32(skb, SMC_NLA_LINK_CONN_CNT,
403 atomic_read(&link->conn_cnt)))
405 if (nla_put_u8(skb, SMC_NLA_LINK_IB_PORT, link->ibport))
407 if (nla_put_u32(skb, SMC_NLA_LINK_NET_DEV, link->ndev_ifidx))
409 snprintf(smc_ibname, sizeof(smc_ibname), "%s", link->ibname);
410 if (nla_put_string(skb, SMC_NLA_LINK_IB_DEV, smc_ibname))
412 memcpy(&link_uid, link->link_uid, sizeof(link_uid));
413 if (nla_put_u32(skb, SMC_NLA_LINK_UID, link_uid))
415 memcpy(&link_uid, link->peer_link_uid, sizeof(link_uid));
416 if (nla_put_u32(skb, SMC_NLA_LINK_PEER_UID, link_uid))
418 memset(smc_gid_target, 0, sizeof(smc_gid_target));
419 smc_gid_be16_convert(smc_gid_target, link->gid);
420 if (nla_put_string(skb, SMC_NLA_LINK_GID, smc_gid_target))
422 memset(smc_gid_target, 0, sizeof(smc_gid_target));
423 smc_gid_be16_convert(smc_gid_target, link->peer_gid);
424 if (nla_put_string(skb, SMC_NLA_LINK_PEER_GID, smc_gid_target))
427 nla_nest_end(skb, attrs);
428 genlmsg_end(skb, nlh);
431 nla_nest_cancel(skb, attrs);
433 genlmsg_cancel(skb, nlh);
438 static int smc_nl_handle_lgr(struct smc_link_group *lgr,
440 struct netlink_callback *cb,
446 nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
447 &smc_gen_nl_family, NLM_F_MULTI,
448 SMC_NETLINK_GET_LGR_SMCR);
451 if (smc_nl_fill_lgr(lgr, skb, cb))
454 genlmsg_end(skb, nlh);
457 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
458 if (!smc_link_usable(&lgr->lnk[i]))
460 if (smc_nl_fill_lgr_link(lgr, &lgr->lnk[i], skb, cb))
467 genlmsg_cancel(skb, nlh);
472 static void smc_nl_fill_lgr_list(struct smc_lgr_list *smc_lgr,
474 struct netlink_callback *cb,
477 struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
478 struct smc_link_group *lgr;
479 int snum = cb_ctx->pos[0];
482 spin_lock_bh(&smc_lgr->lock);
483 list_for_each_entry(lgr, &smc_lgr->list, list) {
486 if (smc_nl_handle_lgr(lgr, skb, cb, list_links))
492 spin_unlock_bh(&smc_lgr->lock);
493 cb_ctx->pos[0] = num;
496 static int smc_nl_fill_smcd_lgr(struct smc_link_group *lgr,
498 struct netlink_callback *cb)
500 char smc_pnet[SMC_MAX_PNETID_LEN + 1];
501 struct nlattr *attrs;
504 nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
505 &smc_gen_nl_family, NLM_F_MULTI,
506 SMC_NETLINK_GET_LGR_SMCD);
510 attrs = nla_nest_start(skb, SMC_GEN_LGR_SMCD);
514 if (nla_put_u32(skb, SMC_NLA_LGR_D_ID, *((u32 *)&lgr->id)))
516 if (nla_put_u64_64bit(skb, SMC_NLA_LGR_D_GID, lgr->smcd->local_gid,
519 if (nla_put_u64_64bit(skb, SMC_NLA_LGR_D_PEER_GID, lgr->peer_gid,
522 if (nla_put_u8(skb, SMC_NLA_LGR_D_VLAN_ID, lgr->vlan_id))
524 if (nla_put_u32(skb, SMC_NLA_LGR_D_CONNS_NUM, lgr->conns_num))
526 if (nla_put_u32(skb, SMC_NLA_LGR_D_CHID, smc_ism_get_chid(lgr->smcd)))
528 memcpy(smc_pnet, lgr->smcd->pnetid, SMC_MAX_PNETID_LEN);
529 smc_pnet[SMC_MAX_PNETID_LEN] = 0;
530 if (nla_put_string(skb, SMC_NLA_LGR_D_PNETID, smc_pnet))
532 if (lgr->smc_version > SMC_V1) {
533 struct nlattr *v2_attrs;
535 v2_attrs = nla_nest_start(skb, SMC_NLA_LGR_D_V2_COMMON);
538 if (smc_nl_fill_lgr_v2_common(lgr, skb, cb, v2_attrs))
541 nla_nest_end(skb, attrs);
542 genlmsg_end(skb, nlh);
546 nla_nest_cancel(skb, attrs);
548 genlmsg_cancel(skb, nlh);
553 static int smc_nl_handle_smcd_lgr(struct smcd_dev *dev,
555 struct netlink_callback *cb)
557 struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
558 struct smc_link_group *lgr;
559 int snum = cb_ctx->pos[1];
562 spin_lock_bh(&dev->lgr_lock);
563 list_for_each_entry(lgr, &dev->lgr_list, list) {
568 rc = smc_nl_fill_smcd_lgr(lgr, skb, cb);
575 spin_unlock_bh(&dev->lgr_lock);
576 cb_ctx->pos[1] = num;
580 static int smc_nl_fill_smcd_dev(struct smcd_dev_list *dev_list,
582 struct netlink_callback *cb)
584 struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
585 struct smcd_dev *smcd_dev;
586 int snum = cb_ctx->pos[0];
589 mutex_lock(&dev_list->mutex);
590 list_for_each_entry(smcd_dev, &dev_list->list, list) {
591 if (list_empty(&smcd_dev->lgr_list))
595 rc = smc_nl_handle_smcd_lgr(smcd_dev, skb, cb);
602 mutex_unlock(&dev_list->mutex);
603 cb_ctx->pos[0] = num;
607 int smcr_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb)
609 bool list_links = false;
611 smc_nl_fill_lgr_list(&smc_lgr_list, skb, cb, list_links);
615 int smcr_nl_get_link(struct sk_buff *skb, struct netlink_callback *cb)
617 bool list_links = true;
619 smc_nl_fill_lgr_list(&smc_lgr_list, skb, cb, list_links);
623 int smcd_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb)
625 smc_nl_fill_smcd_dev(&smcd_dev_list, skb, cb);
629 void smc_lgr_cleanup_early(struct smc_link_group *lgr)
631 spinlock_t *lgr_lock;
636 smc_lgr_list_head(lgr, &lgr_lock);
637 spin_lock_bh(lgr_lock);
638 /* do not use this link group for new connections */
639 if (!list_empty(&lgr->list))
640 list_del_init(&lgr->list);
641 spin_unlock_bh(lgr_lock);
642 __smc_lgr_terminate(lgr, true);
645 static void smcr_lgr_link_deactivate_all(struct smc_link_group *lgr)
649 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
650 struct smc_link *lnk = &lgr->lnk[i];
652 if (smc_link_sendable(lnk))
653 lnk->state = SMC_LNK_INACTIVE;
655 wake_up_all(&lgr->llc_msg_waiter);
656 wake_up_all(&lgr->llc_flow_waiter);
659 static void smc_lgr_free(struct smc_link_group *lgr);
661 static void smc_lgr_free_work(struct work_struct *work)
663 struct smc_link_group *lgr = container_of(to_delayed_work(work),
664 struct smc_link_group,
666 spinlock_t *lgr_lock;
669 smc_lgr_list_head(lgr, &lgr_lock);
670 spin_lock_bh(lgr_lock);
672 spin_unlock_bh(lgr_lock);
675 read_lock_bh(&lgr->conns_lock);
676 conns = RB_EMPTY_ROOT(&lgr->conns_all);
677 read_unlock_bh(&lgr->conns_lock);
678 if (!conns) { /* number of lgr connections is no longer zero */
679 spin_unlock_bh(lgr_lock);
682 list_del_init(&lgr->list); /* remove from smc_lgr_list */
683 lgr->freeing = 1; /* this instance does the freeing, no new schedule */
684 spin_unlock_bh(lgr_lock);
685 cancel_delayed_work(&lgr->free_work);
687 if (!lgr->is_smcd && !lgr->terminating)
688 smc_llc_send_link_delete_all(lgr, true,
689 SMC_LLC_DEL_PROG_INIT_TERM);
690 if (lgr->is_smcd && !lgr->terminating)
691 smc_ism_signal_shutdown(lgr);
693 smcr_lgr_link_deactivate_all(lgr);
697 static void smc_lgr_terminate_work(struct work_struct *work)
699 struct smc_link_group *lgr = container_of(work, struct smc_link_group,
702 __smc_lgr_terminate(lgr, true);
705 /* return next unique link id for the lgr */
706 static u8 smcr_next_link_id(struct smc_link_group *lgr)
713 link_id = ++lgr->next_link_id;
714 if (!link_id) /* skip zero as link_id */
715 link_id = ++lgr->next_link_id;
716 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
717 if (smc_link_usable(&lgr->lnk[i]) &&
718 lgr->lnk[i].link_id == link_id)
726 static void smcr_copy_dev_info_to_link(struct smc_link *link)
728 struct smc_ib_device *smcibdev = link->smcibdev;
730 snprintf(link->ibname, sizeof(link->ibname), "%s",
731 smcibdev->ibdev->name);
732 link->ndev_ifidx = smcibdev->ndev_ifidx[link->ibport - 1];
735 int smcr_link_init(struct smc_link_group *lgr, struct smc_link *lnk,
736 u8 link_idx, struct smc_init_info *ini)
738 struct smc_ib_device *smcibdev;
742 if (lgr->smc_version == SMC_V2) {
743 lnk->smcibdev = ini->smcrv2.ib_dev_v2;
744 lnk->ibport = ini->smcrv2.ib_port_v2;
746 lnk->smcibdev = ini->ib_dev;
747 lnk->ibport = ini->ib_port;
749 get_device(&lnk->smcibdev->ibdev->dev);
750 atomic_inc(&lnk->smcibdev->lnk_cnt);
751 refcount_set(&lnk->refcnt, 1); /* link refcnt is set to 1 */
753 lnk->path_mtu = lnk->smcibdev->pattr[lnk->ibport - 1].active_mtu;
754 lnk->link_id = smcr_next_link_id(lgr);
756 smc_lgr_hold(lgr); /* lgr_put in smcr_link_clear() */
757 lnk->link_idx = link_idx;
758 lnk->wr_rx_id_compl = 0;
759 smc_ibdev_cnt_inc(lnk);
760 smcr_copy_dev_info_to_link(lnk);
761 atomic_set(&lnk->conn_cnt, 0);
762 smc_llc_link_set_uid(lnk);
763 INIT_WORK(&lnk->link_down_wrk, smc_link_down_work);
764 if (!lnk->smcibdev->initialized) {
765 rc = (int)smc_ib_setup_per_ibdev(lnk->smcibdev);
769 get_random_bytes(rndvec, sizeof(rndvec));
770 lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) +
772 rc = smc_ib_determine_gid(lnk->smcibdev, lnk->ibport,
773 ini->vlan_id, lnk->gid, &lnk->sgid_index,
774 lgr->smc_version == SMC_V2 ?
775 &ini->smcrv2 : NULL);
778 rc = smc_llc_link_init(lnk);
781 rc = smc_wr_alloc_link_mem(lnk);
784 rc = smc_ib_create_protection_domain(lnk);
787 rc = smc_ib_create_queue_pair(lnk);
790 rc = smc_wr_create_link(lnk);
793 lnk->state = SMC_LNK_ACTIVATING;
797 smc_ib_destroy_queue_pair(lnk);
799 smc_ib_dealloc_protection_domain(lnk);
801 smc_wr_free_link_mem(lnk);
803 smc_llc_link_clear(lnk, false);
805 smc_ibdev_cnt_dec(lnk);
806 put_device(&lnk->smcibdev->ibdev->dev);
807 smcibdev = lnk->smcibdev;
808 memset(lnk, 0, sizeof(struct smc_link));
809 lnk->state = SMC_LNK_UNUSED;
810 if (!atomic_dec_return(&smcibdev->lnk_cnt))
811 wake_up(&smcibdev->lnks_deleted);
812 smc_lgr_put(lgr); /* lgr_hold above */
816 /* create a new SMC link group */
817 static int smc_lgr_create(struct smc_sock *smc, struct smc_init_info *ini)
819 struct smc_link_group *lgr;
820 struct list_head *lgr_list;
821 struct smc_link *lnk;
822 spinlock_t *lgr_lock;
827 if (ini->is_smcd && ini->vlan_id) {
828 if (smc_ism_get_vlan(ini->ism_dev[ini->ism_selected],
830 rc = SMC_CLC_DECL_ISMVLANERR;
835 lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
837 rc = SMC_CLC_DECL_MEM;
840 lgr->tx_wq = alloc_workqueue("smc_tx_wq-%*phN", 0, 0,
841 SMC_LGR_ID_SIZE, &lgr->id);
846 lgr->is_smcd = ini->is_smcd;
848 lgr->terminating = 0;
850 lgr->vlan_id = ini->vlan_id;
851 refcount_set(&lgr->refcnt, 1); /* set lgr refcnt to 1 */
852 mutex_init(&lgr->sndbufs_lock);
853 mutex_init(&lgr->rmbs_lock);
854 rwlock_init(&lgr->conns_lock);
855 for (i = 0; i < SMC_RMBE_SIZES; i++) {
856 INIT_LIST_HEAD(&lgr->sndbufs[i]);
857 INIT_LIST_HEAD(&lgr->rmbs[i]);
859 lgr->next_link_id = 0;
860 smc_lgr_list.num += SMC_LGR_NUM_INCR;
861 memcpy(&lgr->id, (u8 *)&smc_lgr_list.num, SMC_LGR_ID_SIZE);
862 INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
863 INIT_WORK(&lgr->terminate_work, smc_lgr_terminate_work);
864 lgr->conns_all = RB_ROOT;
866 /* SMC-D specific settings */
867 get_device(&ini->ism_dev[ini->ism_selected]->dev);
868 lgr->peer_gid = ini->ism_peer_gid[ini->ism_selected];
869 lgr->smcd = ini->ism_dev[ini->ism_selected];
870 lgr_list = &ini->ism_dev[ini->ism_selected]->lgr_list;
871 lgr_lock = &lgr->smcd->lgr_lock;
872 lgr->smc_version = ini->smcd_version;
873 lgr->peer_shutdown = 0;
874 atomic_inc(&ini->ism_dev[ini->ism_selected]->lgr_cnt);
876 /* SMC-R specific settings */
877 struct smc_ib_device *ibdev;
880 lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
881 lgr->smc_version = ini->smcr_version;
882 memcpy(lgr->peer_systemid, ini->peer_systemid,
884 if (lgr->smc_version == SMC_V2) {
885 ibdev = ini->smcrv2.ib_dev_v2;
886 ibport = ini->smcrv2.ib_port_v2;
887 lgr->saddr = ini->smcrv2.saddr;
888 lgr->uses_gateway = ini->smcrv2.uses_gateway;
889 memcpy(lgr->nexthop_mac, ini->smcrv2.nexthop_mac,
893 ibport = ini->ib_port;
895 memcpy(lgr->pnet_id, ibdev->pnetid[ibport - 1],
897 if (smc_wr_alloc_lgr_mem(lgr))
899 smc_llc_lgr_init(lgr, smc);
901 link_idx = SMC_SINGLE_LINK;
902 lnk = &lgr->lnk[link_idx];
903 rc = smcr_link_init(lgr, lnk, link_idx, ini);
905 smc_wr_free_lgr_mem(lgr);
908 lgr->net = smc_ib_net(lnk->smcibdev);
909 lgr_list = &smc_lgr_list.list;
910 lgr_lock = &smc_lgr_list.lock;
911 atomic_inc(&lgr_cnt);
914 spin_lock_bh(lgr_lock);
915 list_add_tail(&lgr->list, lgr_list);
916 spin_unlock_bh(lgr_lock);
920 destroy_workqueue(lgr->tx_wq);
924 if (ini->is_smcd && ini->vlan_id)
925 smc_ism_put_vlan(ini->ism_dev[ini->ism_selected], ini->vlan_id);
929 rc = SMC_CLC_DECL_MEM;
931 rc = SMC_CLC_DECL_INTERR;
936 static int smc_write_space(struct smc_connection *conn)
938 int buffer_len = conn->peer_rmbe_size;
939 union smc_host_cursor prod;
940 union smc_host_cursor cons;
943 smc_curs_copy(&prod, &conn->local_tx_ctrl.prod, conn);
944 smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
945 /* determine rx_buf space */
946 space = buffer_len - smc_curs_diff(buffer_len, &cons, &prod);
950 static int smc_switch_cursor(struct smc_sock *smc, struct smc_cdc_tx_pend *pend,
951 struct smc_wr_buf *wr_buf)
953 struct smc_connection *conn = &smc->conn;
954 union smc_host_cursor cons, fin;
958 smc_curs_copy(&conn->tx_curs_sent, &conn->tx_curs_fin, conn);
959 smc_curs_copy(&fin, &conn->local_tx_ctrl_fin, conn);
960 /* set prod cursor to old state, enforce tx_rdma_writes() */
961 smc_curs_copy(&conn->local_tx_ctrl.prod, &fin, conn);
962 smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
964 if (smc_curs_comp(conn->peer_rmbe_size, &cons, &fin) < 0) {
965 /* cons cursor advanced more than fin, and prod was set
966 * fin above, so now prod is smaller than cons. Fix that.
968 diff = smc_curs_diff(conn->peer_rmbe_size, &fin, &cons);
969 smc_curs_add(conn->sndbuf_desc->len,
970 &conn->tx_curs_sent, diff);
971 smc_curs_add(conn->sndbuf_desc->len,
972 &conn->tx_curs_fin, diff);
974 smp_mb__before_atomic();
975 atomic_add(diff, &conn->sndbuf_space);
976 smp_mb__after_atomic();
978 smc_curs_add(conn->peer_rmbe_size,
979 &conn->local_tx_ctrl.prod, diff);
980 smc_curs_add(conn->peer_rmbe_size,
981 &conn->local_tx_ctrl_fin, diff);
983 /* recalculate, value is used by tx_rdma_writes() */
984 atomic_set(&smc->conn.peer_rmbe_space, smc_write_space(conn));
986 if (smc->sk.sk_state != SMC_INIT &&
987 smc->sk.sk_state != SMC_CLOSED) {
988 rc = smcr_cdc_msg_send_validation(conn, pend, wr_buf);
990 queue_delayed_work(conn->lgr->tx_wq, &conn->tx_work, 0);
991 smc->sk.sk_data_ready(&smc->sk);
994 smc_wr_tx_put_slot(conn->lnk,
995 (struct smc_wr_tx_pend_priv *)pend);
1000 void smc_switch_link_and_count(struct smc_connection *conn,
1001 struct smc_link *to_lnk)
1003 atomic_dec(&conn->lnk->conn_cnt);
1004 /* link_hold in smc_conn_create() */
1005 smcr_link_put(conn->lnk);
1007 atomic_inc(&conn->lnk->conn_cnt);
1008 /* link_put in smc_conn_free() */
1009 smcr_link_hold(conn->lnk);
1012 struct smc_link *smc_switch_conns(struct smc_link_group *lgr,
1013 struct smc_link *from_lnk, bool is_dev_err)
1015 struct smc_link *to_lnk = NULL;
1016 struct smc_cdc_tx_pend *pend;
1017 struct smc_connection *conn;
1018 struct smc_wr_buf *wr_buf;
1019 struct smc_sock *smc;
1020 struct rb_node *node;
1023 /* link is inactive, wake up tx waiters */
1024 smc_wr_wakeup_tx_wait(from_lnk);
1026 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1027 if (!smc_link_active(&lgr->lnk[i]) || i == from_lnk->link_idx)
1029 if (is_dev_err && from_lnk->smcibdev == lgr->lnk[i].smcibdev &&
1030 from_lnk->ibport == lgr->lnk[i].ibport) {
1033 to_lnk = &lgr->lnk[i];
1036 if (!to_lnk || !smc_wr_tx_link_hold(to_lnk)) {
1037 smc_lgr_terminate_sched(lgr);
1041 read_lock_bh(&lgr->conns_lock);
1042 for (node = rb_first(&lgr->conns_all); node; node = rb_next(node)) {
1043 conn = rb_entry(node, struct smc_connection, alert_node);
1044 if (conn->lnk != from_lnk)
1046 smc = container_of(conn, struct smc_sock, conn);
1047 /* conn->lnk not yet set in SMC_INIT state */
1048 if (smc->sk.sk_state == SMC_INIT)
1050 if (smc->sk.sk_state == SMC_CLOSED ||
1051 smc->sk.sk_state == SMC_PEERCLOSEWAIT1 ||
1052 smc->sk.sk_state == SMC_PEERCLOSEWAIT2 ||
1053 smc->sk.sk_state == SMC_APPFINCLOSEWAIT ||
1054 smc->sk.sk_state == SMC_APPCLOSEWAIT1 ||
1055 smc->sk.sk_state == SMC_APPCLOSEWAIT2 ||
1056 smc->sk.sk_state == SMC_PEERFINCLOSEWAIT ||
1057 smc->sk.sk_state == SMC_PEERABORTWAIT ||
1058 smc->sk.sk_state == SMC_PROCESSABORT) {
1059 spin_lock_bh(&conn->send_lock);
1060 smc_switch_link_and_count(conn, to_lnk);
1061 spin_unlock_bh(&conn->send_lock);
1064 sock_hold(&smc->sk);
1065 read_unlock_bh(&lgr->conns_lock);
1066 /* pre-fetch buffer outside of send_lock, might sleep */
1067 rc = smc_cdc_get_free_slot(conn, to_lnk, &wr_buf, NULL, &pend);
1070 /* avoid race with smcr_tx_sndbuf_nonempty() */
1071 spin_lock_bh(&conn->send_lock);
1072 smc_switch_link_and_count(conn, to_lnk);
1073 rc = smc_switch_cursor(smc, pend, wr_buf);
1074 spin_unlock_bh(&conn->send_lock);
1080 read_unlock_bh(&lgr->conns_lock);
1081 smc_wr_tx_link_put(to_lnk);
1085 smcr_link_down_cond_sched(to_lnk);
1086 smc_wr_tx_link_put(to_lnk);
1090 static void smcr_buf_unuse(struct smc_buf_desc *rmb_desc,
1091 struct smc_link_group *lgr)
1095 if (rmb_desc->is_conf_rkey && !list_empty(&lgr->list)) {
1096 /* unregister rmb with peer */
1097 rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY);
1099 /* protect against smc_llc_cli_rkey_exchange() */
1100 mutex_lock(&lgr->llc_conf_mutex);
1101 smc_llc_do_delete_rkey(lgr, rmb_desc);
1102 rmb_desc->is_conf_rkey = false;
1103 mutex_unlock(&lgr->llc_conf_mutex);
1104 smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl);
1108 if (rmb_desc->is_reg_err) {
1109 /* buf registration failed, reuse not possible */
1110 mutex_lock(&lgr->rmbs_lock);
1111 list_del(&rmb_desc->list);
1112 mutex_unlock(&lgr->rmbs_lock);
1114 smc_buf_free(lgr, true, rmb_desc);
1117 memset(rmb_desc->cpu_addr, 0, rmb_desc->len);
1121 static void smc_buf_unuse(struct smc_connection *conn,
1122 struct smc_link_group *lgr)
1124 if (conn->sndbuf_desc) {
1125 conn->sndbuf_desc->used = 0;
1126 memset(conn->sndbuf_desc->cpu_addr, 0, conn->sndbuf_desc->len);
1128 if (conn->rmb_desc && lgr->is_smcd) {
1129 conn->rmb_desc->used = 0;
1130 memset(conn->rmb_desc->cpu_addr, 0, conn->rmb_desc->len +
1131 sizeof(struct smcd_cdc_msg));
1132 } else if (conn->rmb_desc) {
1133 smcr_buf_unuse(conn->rmb_desc, lgr);
1137 /* remove a finished connection from its link group */
1138 void smc_conn_free(struct smc_connection *conn)
1140 struct smc_link_group *lgr = conn->lgr;
1142 if (!lgr || conn->freed)
1143 /* Connection has never been registered in a
1144 * link group, or has already been freed.
1149 if (!smc_conn_lgr_valid(conn))
1150 /* Connection has already unregistered from
1156 if (!list_empty(&lgr->list))
1157 smc_ism_unset_conn(conn);
1158 tasklet_kill(&conn->rx_tsklet);
1160 smc_cdc_wait_pend_tx_wr(conn);
1161 if (current_work() != &conn->abort_work)
1162 cancel_work_sync(&conn->abort_work);
1164 if (!list_empty(&lgr->list)) {
1165 smc_buf_unuse(conn, lgr); /* allow buffer reuse */
1166 smc_lgr_unregister_conn(conn);
1169 if (!lgr->conns_num)
1170 smc_lgr_schedule_free_work(lgr);
1173 smcr_link_put(conn->lnk); /* link_hold in smc_conn_create() */
1174 smc_lgr_put(lgr); /* lgr_hold in smc_conn_create() */
1177 /* unregister a link from a buf_desc */
1178 static void smcr_buf_unmap_link(struct smc_buf_desc *buf_desc, bool is_rmb,
1179 struct smc_link *lnk)
1182 buf_desc->is_reg_mr[lnk->link_idx] = false;
1183 if (!buf_desc->is_map_ib[lnk->link_idx])
1186 if (buf_desc->mr_rx[lnk->link_idx]) {
1187 smc_ib_put_memory_region(
1188 buf_desc->mr_rx[lnk->link_idx]);
1189 buf_desc->mr_rx[lnk->link_idx] = NULL;
1191 smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_FROM_DEVICE);
1193 smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_TO_DEVICE);
1195 sg_free_table(&buf_desc->sgt[lnk->link_idx]);
1196 buf_desc->is_map_ib[lnk->link_idx] = false;
1199 /* unmap all buffers of lgr for a deleted link */
1200 static void smcr_buf_unmap_lgr(struct smc_link *lnk)
1202 struct smc_link_group *lgr = lnk->lgr;
1203 struct smc_buf_desc *buf_desc, *bf;
1206 for (i = 0; i < SMC_RMBE_SIZES; i++) {
1207 mutex_lock(&lgr->rmbs_lock);
1208 list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list)
1209 smcr_buf_unmap_link(buf_desc, true, lnk);
1210 mutex_unlock(&lgr->rmbs_lock);
1211 mutex_lock(&lgr->sndbufs_lock);
1212 list_for_each_entry_safe(buf_desc, bf, &lgr->sndbufs[i],
1214 smcr_buf_unmap_link(buf_desc, false, lnk);
1215 mutex_unlock(&lgr->sndbufs_lock);
1219 static void smcr_rtoken_clear_link(struct smc_link *lnk)
1221 struct smc_link_group *lgr = lnk->lgr;
1224 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1225 lgr->rtokens[i][lnk->link_idx].rkey = 0;
1226 lgr->rtokens[i][lnk->link_idx].dma_addr = 0;
1230 static void __smcr_link_clear(struct smc_link *lnk)
1232 struct smc_link_group *lgr = lnk->lgr;
1233 struct smc_ib_device *smcibdev;
1235 smc_wr_free_link_mem(lnk);
1236 smc_ibdev_cnt_dec(lnk);
1237 put_device(&lnk->smcibdev->ibdev->dev);
1238 smcibdev = lnk->smcibdev;
1239 memset(lnk, 0, sizeof(struct smc_link));
1240 lnk->state = SMC_LNK_UNUSED;
1241 if (!atomic_dec_return(&smcibdev->lnk_cnt))
1242 wake_up(&smcibdev->lnks_deleted);
1243 smc_lgr_put(lgr); /* lgr_hold in smcr_link_init() */
1246 /* must be called under lgr->llc_conf_mutex lock */
1247 void smcr_link_clear(struct smc_link *lnk, bool log)
1249 if (!lnk->lgr || lnk->clearing ||
1250 lnk->state == SMC_LNK_UNUSED)
1254 smc_llc_link_clear(lnk, log);
1255 smcr_buf_unmap_lgr(lnk);
1256 smcr_rtoken_clear_link(lnk);
1257 smc_ib_modify_qp_error(lnk);
1258 smc_wr_free_link(lnk);
1259 smc_ib_destroy_queue_pair(lnk);
1260 smc_ib_dealloc_protection_domain(lnk);
1261 smcr_link_put(lnk); /* theoretically last link_put */
1264 void smcr_link_hold(struct smc_link *lnk)
1266 refcount_inc(&lnk->refcnt);
1269 void smcr_link_put(struct smc_link *lnk)
1271 if (refcount_dec_and_test(&lnk->refcnt))
1272 __smcr_link_clear(lnk);
1275 static void smcr_buf_free(struct smc_link_group *lgr, bool is_rmb,
1276 struct smc_buf_desc *buf_desc)
1280 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
1281 smcr_buf_unmap_link(buf_desc, is_rmb, &lgr->lnk[i]);
1283 if (buf_desc->pages)
1284 __free_pages(buf_desc->pages, buf_desc->order);
1288 static void smcd_buf_free(struct smc_link_group *lgr, bool is_dmb,
1289 struct smc_buf_desc *buf_desc)
1292 /* restore original buf len */
1293 buf_desc->len += sizeof(struct smcd_cdc_msg);
1294 smc_ism_unregister_dmb(lgr->smcd, buf_desc);
1296 kfree(buf_desc->cpu_addr);
1301 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
1302 struct smc_buf_desc *buf_desc)
1305 smcd_buf_free(lgr, is_rmb, buf_desc);
1307 smcr_buf_free(lgr, is_rmb, buf_desc);
1310 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
1312 struct smc_buf_desc *buf_desc, *bf_desc;
1313 struct list_head *buf_list;
1316 for (i = 0; i < SMC_RMBE_SIZES; i++) {
1318 buf_list = &lgr->rmbs[i];
1320 buf_list = &lgr->sndbufs[i];
1321 list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
1323 list_del(&buf_desc->list);
1324 smc_buf_free(lgr, is_rmb, buf_desc);
1329 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
1331 /* free send buffers */
1332 __smc_lgr_free_bufs(lgr, false);
1334 __smc_lgr_free_bufs(lgr, true);
1337 /* won't be freed until no one accesses to lgr anymore */
1338 static void __smc_lgr_free(struct smc_link_group *lgr)
1340 smc_lgr_free_bufs(lgr);
1342 if (!atomic_dec_return(&lgr->smcd->lgr_cnt))
1343 wake_up(&lgr->smcd->lgrs_deleted);
1345 smc_wr_free_lgr_mem(lgr);
1346 if (!atomic_dec_return(&lgr_cnt))
1347 wake_up(&lgrs_deleted);
1352 /* remove a link group */
1353 static void smc_lgr_free(struct smc_link_group *lgr)
1357 if (!lgr->is_smcd) {
1358 mutex_lock(&lgr->llc_conf_mutex);
1359 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1360 if (lgr->lnk[i].state != SMC_LNK_UNUSED)
1361 smcr_link_clear(&lgr->lnk[i], false);
1363 mutex_unlock(&lgr->llc_conf_mutex);
1364 smc_llc_lgr_clear(lgr);
1367 destroy_workqueue(lgr->tx_wq);
1369 smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
1370 put_device(&lgr->smcd->dev);
1372 smc_lgr_put(lgr); /* theoretically last lgr_put */
1375 void smc_lgr_hold(struct smc_link_group *lgr)
1377 refcount_inc(&lgr->refcnt);
1380 void smc_lgr_put(struct smc_link_group *lgr)
1382 if (refcount_dec_and_test(&lgr->refcnt))
1383 __smc_lgr_free(lgr);
1386 static void smc_sk_wake_ups(struct smc_sock *smc)
1388 smc->sk.sk_write_space(&smc->sk);
1389 smc->sk.sk_data_ready(&smc->sk);
1390 smc->sk.sk_state_change(&smc->sk);
1393 /* kill a connection */
1394 static void smc_conn_kill(struct smc_connection *conn, bool soft)
1396 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1398 if (conn->lgr->is_smcd && conn->lgr->peer_shutdown)
1399 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
1401 smc_close_abort(conn);
1403 smc->sk.sk_err = ECONNABORTED;
1404 smc_sk_wake_ups(smc);
1405 if (conn->lgr->is_smcd) {
1406 smc_ism_unset_conn(conn);
1408 tasklet_kill(&conn->rx_tsklet);
1410 tasklet_unlock_wait(&conn->rx_tsklet);
1412 smc_cdc_wait_pend_tx_wr(conn);
1414 smc_lgr_unregister_conn(conn);
1415 smc_close_active_abort(smc);
1418 static void smc_lgr_cleanup(struct smc_link_group *lgr)
1421 smc_ism_signal_shutdown(lgr);
1423 u32 rsn = lgr->llc_termination_rsn;
1426 rsn = SMC_LLC_DEL_PROG_INIT_TERM;
1427 smc_llc_send_link_delete_all(lgr, false, rsn);
1428 smcr_lgr_link_deactivate_all(lgr);
1432 /* terminate link group
1433 * @soft: true if link group shutdown can take its time
1434 * false if immediate link group shutdown is required
1436 static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft)
1438 struct smc_connection *conn;
1439 struct smc_sock *smc;
1440 struct rb_node *node;
1442 if (lgr->terminating)
1443 return; /* lgr already terminating */
1444 /* cancel free_work sync, will terminate when lgr->freeing is set */
1445 cancel_delayed_work_sync(&lgr->free_work);
1446 lgr->terminating = 1;
1448 /* kill remaining link group connections */
1449 read_lock_bh(&lgr->conns_lock);
1450 node = rb_first(&lgr->conns_all);
1452 read_unlock_bh(&lgr->conns_lock);
1453 conn = rb_entry(node, struct smc_connection, alert_node);
1454 smc = container_of(conn, struct smc_sock, conn);
1455 sock_hold(&smc->sk); /* sock_put below */
1456 lock_sock(&smc->sk);
1457 smc_conn_kill(conn, soft);
1458 release_sock(&smc->sk);
1459 sock_put(&smc->sk); /* sock_hold above */
1460 read_lock_bh(&lgr->conns_lock);
1461 node = rb_first(&lgr->conns_all);
1463 read_unlock_bh(&lgr->conns_lock);
1464 smc_lgr_cleanup(lgr);
1468 /* unlink link group and schedule termination */
1469 void smc_lgr_terminate_sched(struct smc_link_group *lgr)
1471 spinlock_t *lgr_lock;
1473 smc_lgr_list_head(lgr, &lgr_lock);
1474 spin_lock_bh(lgr_lock);
1475 if (list_empty(&lgr->list) || lgr->terminating || lgr->freeing) {
1476 spin_unlock_bh(lgr_lock);
1477 return; /* lgr already terminating */
1479 list_del_init(&lgr->list);
1481 spin_unlock_bh(lgr_lock);
1482 schedule_work(&lgr->terminate_work);
1485 /* Called when peer lgr shutdown (regularly or abnormally) is received */
1486 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid, unsigned short vlan)
1488 struct smc_link_group *lgr, *l;
1489 LIST_HEAD(lgr_free_list);
1491 /* run common cleanup function and build free list */
1492 spin_lock_bh(&dev->lgr_lock);
1493 list_for_each_entry_safe(lgr, l, &dev->lgr_list, list) {
1494 if ((!peer_gid || lgr->peer_gid == peer_gid) &&
1495 (vlan == VLAN_VID_MASK || lgr->vlan_id == vlan)) {
1496 if (peer_gid) /* peer triggered termination */
1497 lgr->peer_shutdown = 1;
1498 list_move(&lgr->list, &lgr_free_list);
1502 spin_unlock_bh(&dev->lgr_lock);
1504 /* cancel the regular free workers and actually free lgrs */
1505 list_for_each_entry_safe(lgr, l, &lgr_free_list, list) {
1506 list_del_init(&lgr->list);
1507 schedule_work(&lgr->terminate_work);
1511 /* Called when an SMCD device is removed or the smc module is unloaded */
1512 void smc_smcd_terminate_all(struct smcd_dev *smcd)
1514 struct smc_link_group *lgr, *lg;
1515 LIST_HEAD(lgr_free_list);
1517 spin_lock_bh(&smcd->lgr_lock);
1518 list_splice_init(&smcd->lgr_list, &lgr_free_list);
1519 list_for_each_entry(lgr, &lgr_free_list, list)
1521 spin_unlock_bh(&smcd->lgr_lock);
1523 list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
1524 list_del_init(&lgr->list);
1525 __smc_lgr_terminate(lgr, false);
1528 if (atomic_read(&smcd->lgr_cnt))
1529 wait_event(smcd->lgrs_deleted, !atomic_read(&smcd->lgr_cnt));
1532 /* Called when an SMCR device is removed or the smc module is unloaded.
1533 * If smcibdev is given, all SMCR link groups using this device are terminated.
1534 * If smcibdev is NULL, all SMCR link groups are terminated.
1536 void smc_smcr_terminate_all(struct smc_ib_device *smcibdev)
1538 struct smc_link_group *lgr, *lg;
1539 LIST_HEAD(lgr_free_list);
1542 spin_lock_bh(&smc_lgr_list.lock);
1544 list_splice_init(&smc_lgr_list.list, &lgr_free_list);
1545 list_for_each_entry(lgr, &lgr_free_list, list)
1548 list_for_each_entry_safe(lgr, lg, &smc_lgr_list.list, list) {
1549 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1550 if (lgr->lnk[i].smcibdev == smcibdev)
1551 smcr_link_down_cond_sched(&lgr->lnk[i]);
1555 spin_unlock_bh(&smc_lgr_list.lock);
1557 list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
1558 list_del_init(&lgr->list);
1559 smc_llc_set_termination_rsn(lgr, SMC_LLC_DEL_OP_INIT_TERM);
1560 __smc_lgr_terminate(lgr, false);
1564 if (atomic_read(&smcibdev->lnk_cnt))
1565 wait_event(smcibdev->lnks_deleted,
1566 !atomic_read(&smcibdev->lnk_cnt));
1568 if (atomic_read(&lgr_cnt))
1569 wait_event(lgrs_deleted, !atomic_read(&lgr_cnt));
1573 /* set new lgr type and clear all asymmetric link tagging */
1574 void smcr_lgr_set_type(struct smc_link_group *lgr, enum smc_lgr_type new_type)
1576 char *lgr_type = "";
1579 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
1580 if (smc_link_usable(&lgr->lnk[i]))
1581 lgr->lnk[i].link_is_asym = false;
1582 if (lgr->type == new_type)
1584 lgr->type = new_type;
1586 switch (lgr->type) {
1590 case SMC_LGR_SINGLE:
1591 lgr_type = "SINGLE";
1593 case SMC_LGR_SYMMETRIC:
1594 lgr_type = "SYMMETRIC";
1596 case SMC_LGR_ASYMMETRIC_PEER:
1597 lgr_type = "ASYMMETRIC_PEER";
1599 case SMC_LGR_ASYMMETRIC_LOCAL:
1600 lgr_type = "ASYMMETRIC_LOCAL";
1603 pr_warn_ratelimited("smc: SMC-R lg %*phN net %llu state changed: "
1604 "%s, pnetid %.16s\n", SMC_LGR_ID_SIZE, &lgr->id,
1605 lgr->net->net_cookie, lgr_type, lgr->pnet_id);
1608 /* set new lgr type and tag a link as asymmetric */
1609 void smcr_lgr_set_type_asym(struct smc_link_group *lgr,
1610 enum smc_lgr_type new_type, int asym_lnk_idx)
1612 smcr_lgr_set_type(lgr, new_type);
1613 lgr->lnk[asym_lnk_idx].link_is_asym = true;
1616 /* abort connection, abort_work scheduled from tasklet context */
1617 static void smc_conn_abort_work(struct work_struct *work)
1619 struct smc_connection *conn = container_of(work,
1620 struct smc_connection,
1622 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1624 lock_sock(&smc->sk);
1625 smc_conn_kill(conn, true);
1626 release_sock(&smc->sk);
1627 sock_put(&smc->sk); /* sock_hold done by schedulers of abort_work */
1630 void smcr_port_add(struct smc_ib_device *smcibdev, u8 ibport)
1632 struct smc_link_group *lgr, *n;
1634 list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1635 struct smc_link *link;
1637 if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1638 SMC_MAX_PNETID_LEN) ||
1639 lgr->type == SMC_LGR_SYMMETRIC ||
1640 lgr->type == SMC_LGR_ASYMMETRIC_PEER ||
1641 !rdma_dev_access_netns(smcibdev->ibdev, lgr->net))
1644 /* trigger local add link processing */
1645 link = smc_llc_usable_link(lgr);
1647 smc_llc_add_link_local(link);
1651 /* link is down - switch connections to alternate link,
1652 * must be called under lgr->llc_conf_mutex lock
1654 static void smcr_link_down(struct smc_link *lnk)
1656 struct smc_link_group *lgr = lnk->lgr;
1657 struct smc_link *to_lnk;
1660 if (!lgr || lnk->state == SMC_LNK_UNUSED || list_empty(&lgr->list))
1663 to_lnk = smc_switch_conns(lgr, lnk, true);
1664 if (!to_lnk) { /* no backup link available */
1665 smcr_link_clear(lnk, true);
1668 smcr_lgr_set_type(lgr, SMC_LGR_SINGLE);
1669 del_link_id = lnk->link_id;
1671 if (lgr->role == SMC_SERV) {
1672 /* trigger local delete link processing */
1673 smc_llc_srv_delete_link_local(to_lnk, del_link_id);
1675 if (lgr->llc_flow_lcl.type != SMC_LLC_FLOW_NONE) {
1676 /* another llc task is ongoing */
1677 mutex_unlock(&lgr->llc_conf_mutex);
1678 wait_event_timeout(lgr->llc_flow_waiter,
1679 (list_empty(&lgr->list) ||
1680 lgr->llc_flow_lcl.type == SMC_LLC_FLOW_NONE),
1682 mutex_lock(&lgr->llc_conf_mutex);
1684 if (!list_empty(&lgr->list)) {
1685 smc_llc_send_delete_link(to_lnk, del_link_id,
1687 SMC_LLC_DEL_LOST_PATH);
1688 smcr_link_clear(lnk, true);
1690 wake_up(&lgr->llc_flow_waiter); /* wake up next waiter */
1694 /* must be called under lgr->llc_conf_mutex lock */
1695 void smcr_link_down_cond(struct smc_link *lnk)
1697 if (smc_link_downing(&lnk->state)) {
1698 trace_smcr_link_down(lnk, __builtin_return_address(0));
1699 smcr_link_down(lnk);
1703 /* will get the lgr->llc_conf_mutex lock */
1704 void smcr_link_down_cond_sched(struct smc_link *lnk)
1706 if (smc_link_downing(&lnk->state)) {
1707 trace_smcr_link_down(lnk, __builtin_return_address(0));
1708 schedule_work(&lnk->link_down_wrk);
1712 void smcr_port_err(struct smc_ib_device *smcibdev, u8 ibport)
1714 struct smc_link_group *lgr, *n;
1717 list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1718 if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1719 SMC_MAX_PNETID_LEN))
1720 continue; /* lgr is not affected */
1721 if (list_empty(&lgr->list))
1723 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1724 struct smc_link *lnk = &lgr->lnk[i];
1726 if (smc_link_usable(lnk) &&
1727 lnk->smcibdev == smcibdev && lnk->ibport == ibport)
1728 smcr_link_down_cond_sched(lnk);
1733 static void smc_link_down_work(struct work_struct *work)
1735 struct smc_link *link = container_of(work, struct smc_link,
1737 struct smc_link_group *lgr = link->lgr;
1739 if (list_empty(&lgr->list))
1741 wake_up_all(&lgr->llc_msg_waiter);
1742 mutex_lock(&lgr->llc_conf_mutex);
1743 smcr_link_down(link);
1744 mutex_unlock(&lgr->llc_conf_mutex);
1747 static int smc_vlan_by_tcpsk_walk(struct net_device *lower_dev,
1748 struct netdev_nested_priv *priv)
1750 unsigned short *vlan_id = (unsigned short *)priv->data;
1752 if (is_vlan_dev(lower_dev)) {
1753 *vlan_id = vlan_dev_vlan_id(lower_dev);
1760 /* Determine vlan of internal TCP socket. */
1761 int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini)
1763 struct dst_entry *dst = sk_dst_get(clcsock->sk);
1764 struct netdev_nested_priv priv;
1765 struct net_device *ndev;
1779 if (is_vlan_dev(ndev)) {
1780 ini->vlan_id = vlan_dev_vlan_id(ndev);
1784 priv.data = (void *)&ini->vlan_id;
1786 netdev_walk_all_lower_dev(ndev, smc_vlan_by_tcpsk_walk, &priv);
1795 static bool smcr_lgr_match(struct smc_link_group *lgr, u8 smcr_version,
1799 enum smc_lgr_role role, u32 clcqpn,
1802 struct smc_link *lnk;
1805 if (memcmp(lgr->peer_systemid, peer_systemid, SMC_SYSTEMID_LEN) ||
1809 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1812 if (!smc_link_active(lnk))
1814 /* use verbs API to check netns, instead of lgr->net */
1815 if (!rdma_dev_access_netns(lnk->smcibdev->ibdev, net))
1817 if ((lgr->role == SMC_SERV || lnk->peer_qpn == clcqpn) &&
1818 !memcmp(lnk->peer_gid, peer_gid, SMC_GID_SIZE) &&
1819 (smcr_version == SMC_V2 ||
1820 !memcmp(lnk->peer_mac, peer_mac_v1, ETH_ALEN)))
1826 static bool smcd_lgr_match(struct smc_link_group *lgr,
1827 struct smcd_dev *smcismdev, u64 peer_gid)
1829 return lgr->peer_gid == peer_gid && lgr->smcd == smcismdev;
1832 /* create a new SMC connection (and a new link group if necessary) */
1833 int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini)
1835 struct smc_connection *conn = &smc->conn;
1836 struct net *net = sock_net(&smc->sk);
1837 struct list_head *lgr_list;
1838 struct smc_link_group *lgr;
1839 enum smc_lgr_role role;
1840 spinlock_t *lgr_lock;
1843 lgr_list = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_list :
1845 lgr_lock = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_lock :
1847 ini->first_contact_local = 1;
1848 role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
1849 if (role == SMC_CLNT && ini->first_contact_peer)
1850 /* create new link group as well */
1853 /* determine if an existing link group can be reused */
1854 spin_lock_bh(lgr_lock);
1855 list_for_each_entry(lgr, lgr_list, list) {
1856 write_lock_bh(&lgr->conns_lock);
1858 smcd_lgr_match(lgr, ini->ism_dev[ini->ism_selected],
1859 ini->ism_peer_gid[ini->ism_selected]) :
1860 smcr_lgr_match(lgr, ini->smcr_version,
1862 ini->peer_gid, ini->peer_mac, role,
1863 ini->ib_clcqpn, net)) &&
1865 (ini->smcd_version == SMC_V2 ||
1866 lgr->vlan_id == ini->vlan_id) &&
1867 (role == SMC_CLNT || ini->is_smcd ||
1868 (lgr->conns_num < SMC_RMBS_PER_LGR_MAX &&
1869 !bitmap_full(lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX)))) {
1870 /* link group found */
1871 ini->first_contact_local = 0;
1873 rc = smc_lgr_register_conn(conn, false);
1874 write_unlock_bh(&lgr->conns_lock);
1875 if (!rc && delayed_work_pending(&lgr->free_work))
1876 cancel_delayed_work(&lgr->free_work);
1879 write_unlock_bh(&lgr->conns_lock);
1881 spin_unlock_bh(lgr_lock);
1885 if (role == SMC_CLNT && !ini->first_contact_peer &&
1886 ini->first_contact_local) {
1887 /* Server reuses a link group, but Client wants to start
1889 * send out_of_sync decline, reason synchr. error
1891 return SMC_CLC_DECL_SYNCERR;
1895 if (ini->first_contact_local) {
1896 rc = smc_lgr_create(smc, ini);
1900 write_lock_bh(&lgr->conns_lock);
1901 rc = smc_lgr_register_conn(conn, true);
1902 write_unlock_bh(&lgr->conns_lock);
1904 smc_lgr_cleanup_early(lgr);
1908 smc_lgr_hold(conn->lgr); /* lgr_put in smc_conn_free() */
1909 if (!conn->lgr->is_smcd)
1910 smcr_link_hold(conn->lnk); /* link_put in smc_conn_free() */
1912 conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
1913 conn->local_tx_ctrl.len = SMC_WR_TX_SIZE;
1914 conn->urg_state = SMC_URG_READ;
1915 init_waitqueue_head(&conn->cdc_pend_tx_wq);
1916 INIT_WORK(&smc->conn.abort_work, smc_conn_abort_work);
1918 conn->rx_off = sizeof(struct smcd_cdc_msg);
1919 smcd_cdc_rx_init(conn); /* init tasklet for this conn */
1923 #ifndef KERNEL_HAS_ATOMIC64
1924 spin_lock_init(&conn->acurs_lock);
1931 #define SMCD_DMBE_SIZES 6 /* 0 -> 16KB, 1 -> 32KB, .. 6 -> 1MB */
1932 #define SMCR_RMBE_SIZES 5 /* 0 -> 16KB, 1 -> 32KB, .. 5 -> 512KB */
1934 /* convert the RMB size into the compressed notation (minimum 16K, see
1935 * SMCD/R_DMBE_SIZES.
1936 * In contrast to plain ilog2, this rounds towards the next power of 2,
1937 * so the socket application gets at least its desired sndbuf / rcvbuf size.
1939 static u8 smc_compress_bufsize(int size, bool is_smcd, bool is_rmb)
1941 const unsigned int max_scat = SG_MAX_SINGLE_ALLOC * PAGE_SIZE;
1944 if (size <= SMC_BUF_MIN_SIZE)
1947 size = (size - 1) >> 14; /* convert to 16K multiple */
1948 compressed = min_t(u8, ilog2(size) + 1,
1949 is_smcd ? SMCD_DMBE_SIZES : SMCR_RMBE_SIZES);
1951 if (!is_smcd && is_rmb)
1952 /* RMBs are backed by & limited to max size of scatterlists */
1953 compressed = min_t(u8, compressed, ilog2(max_scat >> 14));
1958 /* convert the RMB size from compressed notation into integer */
1959 int smc_uncompress_bufsize(u8 compressed)
1963 size = 0x00000001 << (((int)compressed) + 14);
1967 /* try to reuse a sndbuf or rmb description slot for a certain
1968 * buffer size; if not available, return NULL
1970 static struct smc_buf_desc *smc_buf_get_slot(int compressed_bufsize,
1972 struct list_head *buf_list)
1974 struct smc_buf_desc *buf_slot;
1977 list_for_each_entry(buf_slot, buf_list, list) {
1978 if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
1987 /* one of the conditions for announcing a receiver's current window size is
1988 * that it "results in a minimum increase in the window size of 10% of the
1989 * receive buffer space" [RFC7609]
1991 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
1993 return max_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
1996 /* map an rmb buf to a link */
1997 static int smcr_buf_map_link(struct smc_buf_desc *buf_desc, bool is_rmb,
1998 struct smc_link *lnk)
2002 if (buf_desc->is_map_ib[lnk->link_idx])
2005 rc = sg_alloc_table(&buf_desc->sgt[lnk->link_idx], 1, GFP_KERNEL);
2008 sg_set_buf(buf_desc->sgt[lnk->link_idx].sgl,
2009 buf_desc->cpu_addr, buf_desc->len);
2011 /* map sg table to DMA address */
2012 rc = smc_ib_buf_map_sg(lnk, buf_desc,
2013 is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
2014 /* SMC protocol depends on mapping to one DMA address only */
2020 /* create a new memory region for the RMB */
2022 rc = smc_ib_get_memory_region(lnk->roce_pd,
2023 IB_ACCESS_REMOTE_WRITE |
2024 IB_ACCESS_LOCAL_WRITE,
2025 buf_desc, lnk->link_idx);
2028 smc_ib_sync_sg_for_device(lnk, buf_desc, DMA_FROM_DEVICE);
2030 buf_desc->is_map_ib[lnk->link_idx] = true;
2034 smc_ib_buf_unmap_sg(lnk, buf_desc,
2035 is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
2037 sg_free_table(&buf_desc->sgt[lnk->link_idx]);
2041 /* register a new rmb on IB device,
2042 * must be called under lgr->llc_conf_mutex lock
2044 int smcr_link_reg_rmb(struct smc_link *link, struct smc_buf_desc *rmb_desc)
2046 if (list_empty(&link->lgr->list))
2048 if (!rmb_desc->is_reg_mr[link->link_idx]) {
2049 /* register memory region for new rmb */
2050 if (smc_wr_reg_send(link, rmb_desc->mr_rx[link->link_idx])) {
2051 rmb_desc->is_reg_err = true;
2054 rmb_desc->is_reg_mr[link->link_idx] = true;
2059 static int _smcr_buf_map_lgr(struct smc_link *lnk, struct mutex *lock,
2060 struct list_head *lst, bool is_rmb)
2062 struct smc_buf_desc *buf_desc, *bf;
2066 list_for_each_entry_safe(buf_desc, bf, lst, list) {
2067 if (!buf_desc->used)
2069 rc = smcr_buf_map_link(buf_desc, is_rmb, lnk);
2078 /* map all used buffers of lgr for a new link */
2079 int smcr_buf_map_lgr(struct smc_link *lnk)
2081 struct smc_link_group *lgr = lnk->lgr;
2084 for (i = 0; i < SMC_RMBE_SIZES; i++) {
2085 rc = _smcr_buf_map_lgr(lnk, &lgr->rmbs_lock,
2086 &lgr->rmbs[i], true);
2089 rc = _smcr_buf_map_lgr(lnk, &lgr->sndbufs_lock,
2090 &lgr->sndbufs[i], false);
2097 /* register all used buffers of lgr for a new link,
2098 * must be called under lgr->llc_conf_mutex lock
2100 int smcr_buf_reg_lgr(struct smc_link *lnk)
2102 struct smc_link_group *lgr = lnk->lgr;
2103 struct smc_buf_desc *buf_desc, *bf;
2106 mutex_lock(&lgr->rmbs_lock);
2107 for (i = 0; i < SMC_RMBE_SIZES; i++) {
2108 list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list) {
2109 if (!buf_desc->used)
2111 rc = smcr_link_reg_rmb(lnk, buf_desc);
2117 mutex_unlock(&lgr->rmbs_lock);
2121 static struct smc_buf_desc *smcr_new_buf_create(struct smc_link_group *lgr,
2122 bool is_rmb, int bufsize)
2124 struct smc_buf_desc *buf_desc;
2126 /* try to alloc a new buffer */
2127 buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
2129 return ERR_PTR(-ENOMEM);
2131 buf_desc->order = get_order(bufsize);
2132 buf_desc->pages = alloc_pages(GFP_KERNEL | __GFP_NOWARN |
2133 __GFP_NOMEMALLOC | __GFP_COMP |
2134 __GFP_NORETRY | __GFP_ZERO,
2136 if (!buf_desc->pages) {
2138 return ERR_PTR(-EAGAIN);
2140 buf_desc->cpu_addr = (void *)page_address(buf_desc->pages);
2141 buf_desc->len = bufsize;
2145 /* map buf_desc on all usable links,
2146 * unused buffers stay mapped as long as the link is up
2148 static int smcr_buf_map_usable_links(struct smc_link_group *lgr,
2149 struct smc_buf_desc *buf_desc, bool is_rmb)
2153 /* protect against parallel link reconfiguration */
2154 mutex_lock(&lgr->llc_conf_mutex);
2155 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
2156 struct smc_link *lnk = &lgr->lnk[i];
2158 if (!smc_link_usable(lnk))
2160 if (smcr_buf_map_link(buf_desc, is_rmb, lnk)) {
2166 mutex_unlock(&lgr->llc_conf_mutex);
2170 static struct smc_buf_desc *smcd_new_buf_create(struct smc_link_group *lgr,
2171 bool is_dmb, int bufsize)
2173 struct smc_buf_desc *buf_desc;
2176 /* try to alloc a new DMB */
2177 buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
2179 return ERR_PTR(-ENOMEM);
2181 rc = smc_ism_register_dmb(lgr, bufsize, buf_desc);
2185 return ERR_PTR(-EAGAIN);
2187 return ERR_PTR(-ENOSPC);
2188 return ERR_PTR(-EIO);
2190 buf_desc->pages = virt_to_page(buf_desc->cpu_addr);
2191 /* CDC header stored in buf. So, pretend it was smaller */
2192 buf_desc->len = bufsize - sizeof(struct smcd_cdc_msg);
2194 buf_desc->cpu_addr = kzalloc(bufsize, GFP_KERNEL |
2195 __GFP_NOWARN | __GFP_NORETRY |
2197 if (!buf_desc->cpu_addr) {
2199 return ERR_PTR(-EAGAIN);
2201 buf_desc->len = bufsize;
2206 static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
2208 struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
2209 struct smc_connection *conn = &smc->conn;
2210 struct smc_link_group *lgr = conn->lgr;
2211 struct list_head *buf_list;
2212 int bufsize, bufsize_short;
2213 bool is_dgraded = false;
2214 struct mutex *lock; /* lock buffer list */
2218 /* use socket recv buffer size (w/o overhead) as start value */
2219 sk_buf_size = smc->sk.sk_rcvbuf / 2;
2221 /* use socket send buffer size (w/o overhead) as start value */
2222 sk_buf_size = smc->sk.sk_sndbuf / 2;
2224 for (bufsize_short = smc_compress_bufsize(sk_buf_size, is_smcd, is_rmb);
2225 bufsize_short >= 0; bufsize_short--) {
2227 lock = &lgr->rmbs_lock;
2228 buf_list = &lgr->rmbs[bufsize_short];
2230 lock = &lgr->sndbufs_lock;
2231 buf_list = &lgr->sndbufs[bufsize_short];
2233 bufsize = smc_uncompress_bufsize(bufsize_short);
2235 /* check for reusable slot in the link group */
2236 buf_desc = smc_buf_get_slot(bufsize_short, lock, buf_list);
2238 SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, bufsize);
2239 SMC_STAT_BUF_REUSE(smc, is_smcd, is_rmb);
2240 break; /* found reusable slot */
2244 buf_desc = smcd_new_buf_create(lgr, is_rmb, bufsize);
2246 buf_desc = smcr_new_buf_create(lgr, is_rmb, bufsize);
2248 if (PTR_ERR(buf_desc) == -ENOMEM)
2250 if (IS_ERR(buf_desc)) {
2253 SMC_STAT_RMB_DOWNGRADED(smc, is_smcd, is_rmb);
2258 SMC_STAT_RMB_ALLOC(smc, is_smcd, is_rmb);
2259 SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, bufsize);
2262 list_add(&buf_desc->list, buf_list);
2267 if (IS_ERR(buf_desc))
2268 return PTR_ERR(buf_desc);
2271 if (smcr_buf_map_usable_links(lgr, buf_desc, is_rmb)) {
2272 smcr_buf_unuse(buf_desc, lgr);
2278 conn->rmb_desc = buf_desc;
2279 conn->rmbe_size_short = bufsize_short;
2280 smc->sk.sk_rcvbuf = bufsize * 2;
2281 atomic_set(&conn->bytes_to_rcv, 0);
2282 conn->rmbe_update_limit =
2283 smc_rmb_wnd_update_limit(buf_desc->len);
2285 smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */
2287 conn->sndbuf_desc = buf_desc;
2288 smc->sk.sk_sndbuf = bufsize * 2;
2289 atomic_set(&conn->sndbuf_space, bufsize);
2294 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn)
2296 if (!smc_conn_lgr_valid(conn) || conn->lgr->is_smcd ||
2297 !smc_link_active(conn->lnk))
2299 smc_ib_sync_sg_for_cpu(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE);
2302 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
2304 if (!smc_conn_lgr_valid(conn) || conn->lgr->is_smcd ||
2305 !smc_link_active(conn->lnk))
2307 smc_ib_sync_sg_for_device(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE);
2310 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
2314 if (!smc_conn_lgr_valid(conn) || conn->lgr->is_smcd)
2316 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
2317 if (!smc_link_active(&conn->lgr->lnk[i]))
2319 smc_ib_sync_sg_for_cpu(&conn->lgr->lnk[i], conn->rmb_desc,
2324 void smc_rmb_sync_sg_for_device(struct smc_connection *conn)
2328 if (!smc_conn_lgr_valid(conn) || conn->lgr->is_smcd)
2330 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
2331 if (!smc_link_active(&conn->lgr->lnk[i]))
2333 smc_ib_sync_sg_for_device(&conn->lgr->lnk[i], conn->rmb_desc,
2338 /* create the send and receive buffer for an SMC socket;
2339 * receive buffers are called RMBs;
2340 * (even though the SMC protocol allows more than one RMB-element per RMB,
2341 * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
2342 * extra RMB for every connection in a link group
2344 int smc_buf_create(struct smc_sock *smc, bool is_smcd)
2348 /* create send buffer */
2349 rc = __smc_buf_create(smc, is_smcd, false);
2353 rc = __smc_buf_create(smc, is_smcd, true);
2355 mutex_lock(&smc->conn.lgr->sndbufs_lock);
2356 list_del(&smc->conn.sndbuf_desc->list);
2357 mutex_unlock(&smc->conn.lgr->sndbufs_lock);
2358 smc_buf_free(smc->conn.lgr, false, smc->conn.sndbuf_desc);
2359 smc->conn.sndbuf_desc = NULL;
2364 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
2368 for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
2369 if (!test_and_set_bit(i, lgr->rtokens_used_mask))
2375 static int smc_rtoken_find_by_link(struct smc_link_group *lgr, int lnk_idx,
2380 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2381 if (test_bit(i, lgr->rtokens_used_mask) &&
2382 lgr->rtokens[i][lnk_idx].rkey == rkey)
2388 /* set rtoken for a new link to an existing rmb */
2389 void smc_rtoken_set(struct smc_link_group *lgr, int link_idx, int link_idx_new,
2390 __be32 nw_rkey_known, __be64 nw_vaddr, __be32 nw_rkey)
2394 rtok_idx = smc_rtoken_find_by_link(lgr, link_idx, ntohl(nw_rkey_known));
2395 if (rtok_idx == -ENOENT)
2397 lgr->rtokens[rtok_idx][link_idx_new].rkey = ntohl(nw_rkey);
2398 lgr->rtokens[rtok_idx][link_idx_new].dma_addr = be64_to_cpu(nw_vaddr);
2401 /* set rtoken for a new link whose link_id is given */
2402 void smc_rtoken_set2(struct smc_link_group *lgr, int rtok_idx, int link_id,
2403 __be64 nw_vaddr, __be32 nw_rkey)
2405 u64 dma_addr = be64_to_cpu(nw_vaddr);
2406 u32 rkey = ntohl(nw_rkey);
2410 for (link_idx = 0; link_idx < SMC_LINKS_PER_LGR_MAX; link_idx++) {
2411 if (lgr->lnk[link_idx].link_id == link_id) {
2418 lgr->rtokens[rtok_idx][link_idx].rkey = rkey;
2419 lgr->rtokens[rtok_idx][link_idx].dma_addr = dma_addr;
2422 /* add a new rtoken from peer */
2423 int smc_rtoken_add(struct smc_link *lnk, __be64 nw_vaddr, __be32 nw_rkey)
2425 struct smc_link_group *lgr = smc_get_lgr(lnk);
2426 u64 dma_addr = be64_to_cpu(nw_vaddr);
2427 u32 rkey = ntohl(nw_rkey);
2430 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2431 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
2432 lgr->rtokens[i][lnk->link_idx].dma_addr == dma_addr &&
2433 test_bit(i, lgr->rtokens_used_mask)) {
2434 /* already in list */
2438 i = smc_rmb_reserve_rtoken_idx(lgr);
2441 lgr->rtokens[i][lnk->link_idx].rkey = rkey;
2442 lgr->rtokens[i][lnk->link_idx].dma_addr = dma_addr;
2446 /* delete an rtoken from all links */
2447 int smc_rtoken_delete(struct smc_link *lnk, __be32 nw_rkey)
2449 struct smc_link_group *lgr = smc_get_lgr(lnk);
2450 u32 rkey = ntohl(nw_rkey);
2453 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2454 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
2455 test_bit(i, lgr->rtokens_used_mask)) {
2456 for (j = 0; j < SMC_LINKS_PER_LGR_MAX; j++) {
2457 lgr->rtokens[i][j].rkey = 0;
2458 lgr->rtokens[i][j].dma_addr = 0;
2460 clear_bit(i, lgr->rtokens_used_mask);
2467 /* save rkey and dma_addr received from peer during clc handshake */
2468 int smc_rmb_rtoken_handling(struct smc_connection *conn,
2469 struct smc_link *lnk,
2470 struct smc_clc_msg_accept_confirm *clc)
2472 conn->rtoken_idx = smc_rtoken_add(lnk, clc->r0.rmb_dma_addr,
2474 if (conn->rtoken_idx < 0)
2475 return conn->rtoken_idx;
2479 static void smc_core_going_away(void)
2481 struct smc_ib_device *smcibdev;
2482 struct smcd_dev *smcd;
2484 mutex_lock(&smc_ib_devices.mutex);
2485 list_for_each_entry(smcibdev, &smc_ib_devices.list, list) {
2488 for (i = 0; i < SMC_MAX_PORTS; i++)
2489 set_bit(i, smcibdev->ports_going_away);
2491 mutex_unlock(&smc_ib_devices.mutex);
2493 mutex_lock(&smcd_dev_list.mutex);
2494 list_for_each_entry(smcd, &smcd_dev_list.list, list) {
2495 smcd->going_away = 1;
2497 mutex_unlock(&smcd_dev_list.mutex);
2500 /* Clean up all SMC link groups */
2501 static void smc_lgrs_shutdown(void)
2503 struct smcd_dev *smcd;
2505 smc_core_going_away();
2507 smc_smcr_terminate_all(NULL);
2509 mutex_lock(&smcd_dev_list.mutex);
2510 list_for_each_entry(smcd, &smcd_dev_list.list, list)
2511 smc_smcd_terminate_all(smcd);
2512 mutex_unlock(&smcd_dev_list.mutex);
2515 static int smc_core_reboot_event(struct notifier_block *this,
2516 unsigned long event, void *ptr)
2518 smc_lgrs_shutdown();
2519 smc_ib_unregister_client();
2523 static struct notifier_block smc_reboot_notifier = {
2524 .notifier_call = smc_core_reboot_event,
2527 int __init smc_core_init(void)
2529 return register_reboot_notifier(&smc_reboot_notifier);
2532 /* Called (from smc_exit) when module is removed */
2533 void smc_core_exit(void)
2535 unregister_reboot_notifier(&smc_reboot_notifier);
2536 smc_lgrs_shutdown();