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 conn->lnk = NULL; /* reset conn->lnk first */
131 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
132 struct smc_link *lnk = &conn->lgr->lnk[i];
134 if (lnk->state != expected || lnk->link_is_asym)
136 if (conn->lgr->role == SMC_CLNT) {
137 conn->lnk = lnk; /* temporary, SMC server assigns link*/
140 if (conn->lgr->conns_num % 2) {
141 for (j = i + 1; j < SMC_LINKS_PER_LGR_MAX; j++) {
142 struct smc_link *lnk2;
144 lnk2 = &conn->lgr->lnk[j];
145 if (lnk2->state == expected &&
146 !lnk2->link_is_asym) {
157 return SMC_CLC_DECL_NOACTLINK;
158 atomic_inc(&conn->lnk->conn_cnt);
162 /* Register connection in link group by assigning an alert token
163 * registered in a search tree.
164 * Requires @conns_lock
165 * Note that '0' is a reserved value and not assigned.
167 static int smc_lgr_register_conn(struct smc_connection *conn, bool first)
169 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
170 static atomic_t nexttoken = ATOMIC_INIT(0);
173 if (!conn->lgr->is_smcd) {
174 rc = smcr_lgr_conn_assign_link(conn, first);
180 /* find a new alert_token_local value not yet used by some connection
183 sock_hold(&smc->sk); /* sock_put in smc_lgr_unregister_conn() */
184 while (!conn->alert_token_local) {
185 conn->alert_token_local = atomic_inc_return(&nexttoken);
186 if (smc_lgr_find_conn(conn->alert_token_local, conn->lgr))
187 conn->alert_token_local = 0;
189 smc_lgr_add_alert_token(conn);
190 conn->lgr->conns_num++;
194 /* Unregister connection and reset the alert token of the given connection<
196 static void __smc_lgr_unregister_conn(struct smc_connection *conn)
198 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
199 struct smc_link_group *lgr = conn->lgr;
201 rb_erase(&conn->alert_node, &lgr->conns_all);
203 atomic_dec(&conn->lnk->conn_cnt);
205 conn->alert_token_local = 0;
206 sock_put(&smc->sk); /* sock_hold in smc_lgr_register_conn() */
209 /* Unregister connection from lgr
211 static void smc_lgr_unregister_conn(struct smc_connection *conn)
213 struct smc_link_group *lgr = conn->lgr;
215 if (!smc_conn_lgr_valid(conn))
217 write_lock_bh(&lgr->conns_lock);
218 if (conn->alert_token_local) {
219 __smc_lgr_unregister_conn(conn);
221 write_unlock_bh(&lgr->conns_lock);
224 int smc_nl_get_sys_info(struct sk_buff *skb, struct netlink_callback *cb)
226 struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
227 char hostname[SMC_MAX_HOSTNAME_LEN + 1];
228 char smc_seid[SMC_MAX_EID_LEN + 1];
229 struct nlattr *attrs;
234 nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
235 &smc_gen_nl_family, NLM_F_MULTI,
236 SMC_NETLINK_GET_SYS_INFO);
241 attrs = nla_nest_start(skb, SMC_GEN_SYS_INFO);
244 if (nla_put_u8(skb, SMC_NLA_SYS_VER, SMC_V2))
246 if (nla_put_u8(skb, SMC_NLA_SYS_REL, SMC_RELEASE))
248 if (nla_put_u8(skb, SMC_NLA_SYS_IS_ISM_V2, smc_ism_is_v2_capable()))
250 if (nla_put_u8(skb, SMC_NLA_SYS_IS_SMCR_V2, true))
252 smc_clc_get_hostname(&host);
254 memcpy(hostname, host, SMC_MAX_HOSTNAME_LEN);
255 hostname[SMC_MAX_HOSTNAME_LEN] = 0;
256 if (nla_put_string(skb, SMC_NLA_SYS_LOCAL_HOST, hostname))
259 if (smc_ism_is_v2_capable()) {
260 smc_ism_get_system_eid(&seid);
261 memcpy(smc_seid, seid, SMC_MAX_EID_LEN);
262 smc_seid[SMC_MAX_EID_LEN] = 0;
263 if (nla_put_string(skb, SMC_NLA_SYS_SEID, smc_seid))
266 nla_nest_end(skb, attrs);
267 genlmsg_end(skb, nlh);
272 nla_nest_cancel(skb, attrs);
274 genlmsg_cancel(skb, nlh);
279 /* Fill SMC_NLA_LGR_D_V2_COMMON/SMC_NLA_LGR_R_V2_COMMON nested attributes */
280 static int smc_nl_fill_lgr_v2_common(struct smc_link_group *lgr,
282 struct netlink_callback *cb,
283 struct nlattr *v2_attrs)
285 char smc_host[SMC_MAX_HOSTNAME_LEN + 1];
286 char smc_eid[SMC_MAX_EID_LEN + 1];
288 if (nla_put_u8(skb, SMC_NLA_LGR_V2_VER, lgr->smc_version))
290 if (nla_put_u8(skb, SMC_NLA_LGR_V2_REL, lgr->peer_smc_release))
292 if (nla_put_u8(skb, SMC_NLA_LGR_V2_OS, lgr->peer_os))
294 memcpy(smc_host, lgr->peer_hostname, SMC_MAX_HOSTNAME_LEN);
295 smc_host[SMC_MAX_HOSTNAME_LEN] = 0;
296 if (nla_put_string(skb, SMC_NLA_LGR_V2_PEER_HOST, smc_host))
298 memcpy(smc_eid, lgr->negotiated_eid, SMC_MAX_EID_LEN);
299 smc_eid[SMC_MAX_EID_LEN] = 0;
300 if (nla_put_string(skb, SMC_NLA_LGR_V2_NEG_EID, smc_eid))
303 nla_nest_end(skb, v2_attrs);
307 nla_nest_cancel(skb, v2_attrs);
311 static int smc_nl_fill_smcr_lgr_v2(struct smc_link_group *lgr,
313 struct netlink_callback *cb)
315 struct nlattr *v2_attrs;
317 v2_attrs = nla_nest_start(skb, SMC_NLA_LGR_R_V2);
320 if (nla_put_u8(skb, SMC_NLA_LGR_R_V2_DIRECT, !lgr->uses_gateway))
323 nla_nest_end(skb, v2_attrs);
327 nla_nest_cancel(skb, v2_attrs);
332 static int smc_nl_fill_lgr(struct smc_link_group *lgr,
334 struct netlink_callback *cb)
336 char smc_target[SMC_MAX_PNETID_LEN + 1];
337 struct nlattr *attrs, *v2_attrs;
339 attrs = nla_nest_start(skb, SMC_GEN_LGR_SMCR);
343 if (nla_put_u32(skb, SMC_NLA_LGR_R_ID, *((u32 *)&lgr->id)))
345 if (nla_put_u32(skb, SMC_NLA_LGR_R_CONNS_NUM, lgr->conns_num))
347 if (nla_put_u8(skb, SMC_NLA_LGR_R_ROLE, lgr->role))
349 if (nla_put_u8(skb, SMC_NLA_LGR_R_TYPE, lgr->type))
351 if (nla_put_u8(skb, SMC_NLA_LGR_R_BUF_TYPE, lgr->buf_type))
353 if (nla_put_u8(skb, SMC_NLA_LGR_R_VLAN_ID, lgr->vlan_id))
355 if (nla_put_u64_64bit(skb, SMC_NLA_LGR_R_NET_COOKIE,
356 lgr->net->net_cookie, SMC_NLA_LGR_R_PAD))
358 memcpy(smc_target, lgr->pnet_id, SMC_MAX_PNETID_LEN);
359 smc_target[SMC_MAX_PNETID_LEN] = 0;
360 if (nla_put_string(skb, SMC_NLA_LGR_R_PNETID, smc_target))
362 if (lgr->smc_version > SMC_V1) {
363 v2_attrs = nla_nest_start(skb, SMC_NLA_LGR_R_V2_COMMON);
366 if (smc_nl_fill_lgr_v2_common(lgr, skb, cb, v2_attrs))
368 if (smc_nl_fill_smcr_lgr_v2(lgr, skb, cb))
372 nla_nest_end(skb, attrs);
375 nla_nest_cancel(skb, attrs);
380 static int smc_nl_fill_lgr_link(struct smc_link_group *lgr,
381 struct smc_link *link,
383 struct netlink_callback *cb)
385 char smc_ibname[IB_DEVICE_NAME_MAX];
386 u8 smc_gid_target[41];
387 struct nlattr *attrs;
391 nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
392 &smc_gen_nl_family, NLM_F_MULTI,
393 SMC_NETLINK_GET_LINK_SMCR);
397 attrs = nla_nest_start(skb, SMC_GEN_LINK_SMCR);
401 if (nla_put_u8(skb, SMC_NLA_LINK_ID, link->link_id))
403 if (nla_put_u32(skb, SMC_NLA_LINK_STATE, link->state))
405 if (nla_put_u32(skb, SMC_NLA_LINK_CONN_CNT,
406 atomic_read(&link->conn_cnt)))
408 if (nla_put_u8(skb, SMC_NLA_LINK_IB_PORT, link->ibport))
410 if (nla_put_u32(skb, SMC_NLA_LINK_NET_DEV, link->ndev_ifidx))
412 snprintf(smc_ibname, sizeof(smc_ibname), "%s", link->ibname);
413 if (nla_put_string(skb, SMC_NLA_LINK_IB_DEV, smc_ibname))
415 memcpy(&link_uid, link->link_uid, sizeof(link_uid));
416 if (nla_put_u32(skb, SMC_NLA_LINK_UID, link_uid))
418 memcpy(&link_uid, link->peer_link_uid, sizeof(link_uid));
419 if (nla_put_u32(skb, SMC_NLA_LINK_PEER_UID, link_uid))
421 memset(smc_gid_target, 0, sizeof(smc_gid_target));
422 smc_gid_be16_convert(smc_gid_target, link->gid);
423 if (nla_put_string(skb, SMC_NLA_LINK_GID, smc_gid_target))
425 memset(smc_gid_target, 0, sizeof(smc_gid_target));
426 smc_gid_be16_convert(smc_gid_target, link->peer_gid);
427 if (nla_put_string(skb, SMC_NLA_LINK_PEER_GID, smc_gid_target))
430 nla_nest_end(skb, attrs);
431 genlmsg_end(skb, nlh);
434 nla_nest_cancel(skb, attrs);
436 genlmsg_cancel(skb, nlh);
441 static int smc_nl_handle_lgr(struct smc_link_group *lgr,
443 struct netlink_callback *cb,
449 nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
450 &smc_gen_nl_family, NLM_F_MULTI,
451 SMC_NETLINK_GET_LGR_SMCR);
454 if (smc_nl_fill_lgr(lgr, skb, cb))
457 genlmsg_end(skb, nlh);
460 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
461 if (!smc_link_usable(&lgr->lnk[i]))
463 if (smc_nl_fill_lgr_link(lgr, &lgr->lnk[i], skb, cb))
470 genlmsg_cancel(skb, nlh);
475 static void smc_nl_fill_lgr_list(struct smc_lgr_list *smc_lgr,
477 struct netlink_callback *cb,
480 struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
481 struct smc_link_group *lgr;
482 int snum = cb_ctx->pos[0];
485 spin_lock_bh(&smc_lgr->lock);
486 list_for_each_entry(lgr, &smc_lgr->list, list) {
489 if (smc_nl_handle_lgr(lgr, skb, cb, list_links))
495 spin_unlock_bh(&smc_lgr->lock);
496 cb_ctx->pos[0] = num;
499 static int smc_nl_fill_smcd_lgr(struct smc_link_group *lgr,
501 struct netlink_callback *cb)
503 char smc_pnet[SMC_MAX_PNETID_LEN + 1];
504 struct nlattr *attrs;
507 nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
508 &smc_gen_nl_family, NLM_F_MULTI,
509 SMC_NETLINK_GET_LGR_SMCD);
513 attrs = nla_nest_start(skb, SMC_GEN_LGR_SMCD);
517 if (nla_put_u32(skb, SMC_NLA_LGR_D_ID, *((u32 *)&lgr->id)))
519 if (nla_put_u64_64bit(skb, SMC_NLA_LGR_D_GID, lgr->smcd->local_gid,
522 if (nla_put_u64_64bit(skb, SMC_NLA_LGR_D_PEER_GID, lgr->peer_gid,
525 if (nla_put_u8(skb, SMC_NLA_LGR_D_VLAN_ID, lgr->vlan_id))
527 if (nla_put_u32(skb, SMC_NLA_LGR_D_CONNS_NUM, lgr->conns_num))
529 if (nla_put_u32(skb, SMC_NLA_LGR_D_CHID, smc_ism_get_chid(lgr->smcd)))
531 memcpy(smc_pnet, lgr->smcd->pnetid, SMC_MAX_PNETID_LEN);
532 smc_pnet[SMC_MAX_PNETID_LEN] = 0;
533 if (nla_put_string(skb, SMC_NLA_LGR_D_PNETID, smc_pnet))
535 if (lgr->smc_version > SMC_V1) {
536 struct nlattr *v2_attrs;
538 v2_attrs = nla_nest_start(skb, SMC_NLA_LGR_D_V2_COMMON);
541 if (smc_nl_fill_lgr_v2_common(lgr, skb, cb, v2_attrs))
544 nla_nest_end(skb, attrs);
545 genlmsg_end(skb, nlh);
549 nla_nest_cancel(skb, attrs);
551 genlmsg_cancel(skb, nlh);
556 static int smc_nl_handle_smcd_lgr(struct smcd_dev *dev,
558 struct netlink_callback *cb)
560 struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
561 struct smc_link_group *lgr;
562 int snum = cb_ctx->pos[1];
565 spin_lock_bh(&dev->lgr_lock);
566 list_for_each_entry(lgr, &dev->lgr_list, list) {
571 rc = smc_nl_fill_smcd_lgr(lgr, skb, cb);
578 spin_unlock_bh(&dev->lgr_lock);
579 cb_ctx->pos[1] = num;
583 static int smc_nl_fill_smcd_dev(struct smcd_dev_list *dev_list,
585 struct netlink_callback *cb)
587 struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
588 struct smcd_dev *smcd_dev;
589 int snum = cb_ctx->pos[0];
592 mutex_lock(&dev_list->mutex);
593 list_for_each_entry(smcd_dev, &dev_list->list, list) {
594 if (list_empty(&smcd_dev->lgr_list))
598 rc = smc_nl_handle_smcd_lgr(smcd_dev, skb, cb);
605 mutex_unlock(&dev_list->mutex);
606 cb_ctx->pos[0] = num;
610 int smcr_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb)
612 bool list_links = false;
614 smc_nl_fill_lgr_list(&smc_lgr_list, skb, cb, list_links);
618 int smcr_nl_get_link(struct sk_buff *skb, struct netlink_callback *cb)
620 bool list_links = true;
622 smc_nl_fill_lgr_list(&smc_lgr_list, skb, cb, list_links);
626 int smcd_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb)
628 smc_nl_fill_smcd_dev(&smcd_dev_list, skb, cb);
632 void smc_lgr_cleanup_early(struct smc_link_group *lgr)
634 spinlock_t *lgr_lock;
639 smc_lgr_list_head(lgr, &lgr_lock);
640 spin_lock_bh(lgr_lock);
641 /* do not use this link group for new connections */
642 if (!list_empty(&lgr->list))
643 list_del_init(&lgr->list);
644 spin_unlock_bh(lgr_lock);
645 __smc_lgr_terminate(lgr, true);
648 static void smcr_lgr_link_deactivate_all(struct smc_link_group *lgr)
652 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
653 struct smc_link *lnk = &lgr->lnk[i];
655 if (smc_link_sendable(lnk))
656 lnk->state = SMC_LNK_INACTIVE;
658 wake_up_all(&lgr->llc_msg_waiter);
659 wake_up_all(&lgr->llc_flow_waiter);
662 static void smc_lgr_free(struct smc_link_group *lgr);
664 static void smc_lgr_free_work(struct work_struct *work)
666 struct smc_link_group *lgr = container_of(to_delayed_work(work),
667 struct smc_link_group,
669 spinlock_t *lgr_lock;
672 smc_lgr_list_head(lgr, &lgr_lock);
673 spin_lock_bh(lgr_lock);
675 spin_unlock_bh(lgr_lock);
678 read_lock_bh(&lgr->conns_lock);
679 conns = RB_EMPTY_ROOT(&lgr->conns_all);
680 read_unlock_bh(&lgr->conns_lock);
681 if (!conns) { /* number of lgr connections is no longer zero */
682 spin_unlock_bh(lgr_lock);
685 list_del_init(&lgr->list); /* remove from smc_lgr_list */
686 lgr->freeing = 1; /* this instance does the freeing, no new schedule */
687 spin_unlock_bh(lgr_lock);
688 cancel_delayed_work(&lgr->free_work);
690 if (!lgr->is_smcd && !lgr->terminating)
691 smc_llc_send_link_delete_all(lgr, true,
692 SMC_LLC_DEL_PROG_INIT_TERM);
693 if (lgr->is_smcd && !lgr->terminating)
694 smc_ism_signal_shutdown(lgr);
696 smcr_lgr_link_deactivate_all(lgr);
700 static void smc_lgr_terminate_work(struct work_struct *work)
702 struct smc_link_group *lgr = container_of(work, struct smc_link_group,
705 __smc_lgr_terminate(lgr, true);
708 /* return next unique link id for the lgr */
709 static u8 smcr_next_link_id(struct smc_link_group *lgr)
716 link_id = ++lgr->next_link_id;
717 if (!link_id) /* skip zero as link_id */
718 link_id = ++lgr->next_link_id;
719 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
720 if (smc_link_usable(&lgr->lnk[i]) &&
721 lgr->lnk[i].link_id == link_id)
729 static void smcr_copy_dev_info_to_link(struct smc_link *link)
731 struct smc_ib_device *smcibdev = link->smcibdev;
733 snprintf(link->ibname, sizeof(link->ibname), "%s",
734 smcibdev->ibdev->name);
735 link->ndev_ifidx = smcibdev->ndev_ifidx[link->ibport - 1];
738 int smcr_link_init(struct smc_link_group *lgr, struct smc_link *lnk,
739 u8 link_idx, struct smc_init_info *ini)
741 struct smc_ib_device *smcibdev;
745 if (lgr->smc_version == SMC_V2) {
746 lnk->smcibdev = ini->smcrv2.ib_dev_v2;
747 lnk->ibport = ini->smcrv2.ib_port_v2;
749 lnk->smcibdev = ini->ib_dev;
750 lnk->ibport = ini->ib_port;
752 get_device(&lnk->smcibdev->ibdev->dev);
753 atomic_inc(&lnk->smcibdev->lnk_cnt);
754 refcount_set(&lnk->refcnt, 1); /* link refcnt is set to 1 */
756 lnk->path_mtu = lnk->smcibdev->pattr[lnk->ibport - 1].active_mtu;
757 lnk->link_id = smcr_next_link_id(lgr);
759 smc_lgr_hold(lgr); /* lgr_put in smcr_link_clear() */
760 lnk->link_idx = link_idx;
761 lnk->wr_rx_id_compl = 0;
762 smc_ibdev_cnt_inc(lnk);
763 smcr_copy_dev_info_to_link(lnk);
764 atomic_set(&lnk->conn_cnt, 0);
765 smc_llc_link_set_uid(lnk);
766 INIT_WORK(&lnk->link_down_wrk, smc_link_down_work);
767 if (!lnk->smcibdev->initialized) {
768 rc = (int)smc_ib_setup_per_ibdev(lnk->smcibdev);
772 get_random_bytes(rndvec, sizeof(rndvec));
773 lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) +
775 rc = smc_ib_determine_gid(lnk->smcibdev, lnk->ibport,
776 ini->vlan_id, lnk->gid, &lnk->sgid_index,
777 lgr->smc_version == SMC_V2 ?
778 &ini->smcrv2 : NULL);
781 rc = smc_llc_link_init(lnk);
784 rc = smc_wr_alloc_link_mem(lnk);
787 rc = smc_ib_create_protection_domain(lnk);
790 rc = smc_ib_create_queue_pair(lnk);
793 rc = smc_wr_create_link(lnk);
796 lnk->state = SMC_LNK_ACTIVATING;
800 smc_ib_destroy_queue_pair(lnk);
802 smc_ib_dealloc_protection_domain(lnk);
804 smc_wr_free_link_mem(lnk);
806 smc_llc_link_clear(lnk, false);
808 smc_ibdev_cnt_dec(lnk);
809 put_device(&lnk->smcibdev->ibdev->dev);
810 smcibdev = lnk->smcibdev;
811 memset(lnk, 0, sizeof(struct smc_link));
812 lnk->state = SMC_LNK_UNUSED;
813 if (!atomic_dec_return(&smcibdev->lnk_cnt))
814 wake_up(&smcibdev->lnks_deleted);
815 smc_lgr_put(lgr); /* lgr_hold above */
819 /* create a new SMC link group */
820 static int smc_lgr_create(struct smc_sock *smc, struct smc_init_info *ini)
822 struct smc_link_group *lgr;
823 struct list_head *lgr_list;
824 struct smc_link *lnk;
825 spinlock_t *lgr_lock;
830 if (ini->is_smcd && ini->vlan_id) {
831 if (smc_ism_get_vlan(ini->ism_dev[ini->ism_selected],
833 rc = SMC_CLC_DECL_ISMVLANERR;
838 lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
840 rc = SMC_CLC_DECL_MEM;
843 lgr->tx_wq = alloc_workqueue("smc_tx_wq-%*phN", 0, 0,
844 SMC_LGR_ID_SIZE, &lgr->id);
849 lgr->is_smcd = ini->is_smcd;
851 lgr->terminating = 0;
853 lgr->vlan_id = ini->vlan_id;
854 refcount_set(&lgr->refcnt, 1); /* set lgr refcnt to 1 */
855 mutex_init(&lgr->sndbufs_lock);
856 mutex_init(&lgr->rmbs_lock);
857 rwlock_init(&lgr->conns_lock);
858 for (i = 0; i < SMC_RMBE_SIZES; i++) {
859 INIT_LIST_HEAD(&lgr->sndbufs[i]);
860 INIT_LIST_HEAD(&lgr->rmbs[i]);
862 lgr->next_link_id = 0;
863 smc_lgr_list.num += SMC_LGR_NUM_INCR;
864 memcpy(&lgr->id, (u8 *)&smc_lgr_list.num, SMC_LGR_ID_SIZE);
865 INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
866 INIT_WORK(&lgr->terminate_work, smc_lgr_terminate_work);
867 lgr->conns_all = RB_ROOT;
869 /* SMC-D specific settings */
870 get_device(&ini->ism_dev[ini->ism_selected]->dev);
871 lgr->peer_gid = ini->ism_peer_gid[ini->ism_selected];
872 lgr->smcd = ini->ism_dev[ini->ism_selected];
873 lgr_list = &ini->ism_dev[ini->ism_selected]->lgr_list;
874 lgr_lock = &lgr->smcd->lgr_lock;
875 lgr->smc_version = ini->smcd_version;
876 lgr->peer_shutdown = 0;
877 atomic_inc(&ini->ism_dev[ini->ism_selected]->lgr_cnt);
879 /* SMC-R specific settings */
880 struct smc_ib_device *ibdev;
883 lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
884 lgr->smc_version = ini->smcr_version;
885 memcpy(lgr->peer_systemid, ini->peer_systemid,
887 if (lgr->smc_version == SMC_V2) {
888 ibdev = ini->smcrv2.ib_dev_v2;
889 ibport = ini->smcrv2.ib_port_v2;
890 lgr->saddr = ini->smcrv2.saddr;
891 lgr->uses_gateway = ini->smcrv2.uses_gateway;
892 memcpy(lgr->nexthop_mac, ini->smcrv2.nexthop_mac,
896 ibport = ini->ib_port;
898 memcpy(lgr->pnet_id, ibdev->pnetid[ibport - 1],
900 rc = smc_wr_alloc_lgr_mem(lgr);
903 smc_llc_lgr_init(lgr, smc);
905 link_idx = SMC_SINGLE_LINK;
906 lnk = &lgr->lnk[link_idx];
907 rc = smcr_link_init(lgr, lnk, link_idx, ini);
909 smc_wr_free_lgr_mem(lgr);
912 lgr->net = smc_ib_net(lnk->smcibdev);
913 lgr_list = &smc_lgr_list.list;
914 lgr_lock = &smc_lgr_list.lock;
915 lgr->buf_type = lgr->net->smc.sysctl_smcr_buf_type;
916 atomic_inc(&lgr_cnt);
919 spin_lock_bh(lgr_lock);
920 list_add_tail(&lgr->list, lgr_list);
921 spin_unlock_bh(lgr_lock);
925 destroy_workqueue(lgr->tx_wq);
929 if (ini->is_smcd && ini->vlan_id)
930 smc_ism_put_vlan(ini->ism_dev[ini->ism_selected], ini->vlan_id);
934 rc = SMC_CLC_DECL_MEM;
936 rc = SMC_CLC_DECL_INTERR;
941 static int smc_write_space(struct smc_connection *conn)
943 int buffer_len = conn->peer_rmbe_size;
944 union smc_host_cursor prod;
945 union smc_host_cursor cons;
948 smc_curs_copy(&prod, &conn->local_tx_ctrl.prod, conn);
949 smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
950 /* determine rx_buf space */
951 space = buffer_len - smc_curs_diff(buffer_len, &cons, &prod);
955 static int smc_switch_cursor(struct smc_sock *smc, struct smc_cdc_tx_pend *pend,
956 struct smc_wr_buf *wr_buf)
958 struct smc_connection *conn = &smc->conn;
959 union smc_host_cursor cons, fin;
963 smc_curs_copy(&conn->tx_curs_sent, &conn->tx_curs_fin, conn);
964 smc_curs_copy(&fin, &conn->local_tx_ctrl_fin, conn);
965 /* set prod cursor to old state, enforce tx_rdma_writes() */
966 smc_curs_copy(&conn->local_tx_ctrl.prod, &fin, conn);
967 smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
969 if (smc_curs_comp(conn->peer_rmbe_size, &cons, &fin) < 0) {
970 /* cons cursor advanced more than fin, and prod was set
971 * fin above, so now prod is smaller than cons. Fix that.
973 diff = smc_curs_diff(conn->peer_rmbe_size, &fin, &cons);
974 smc_curs_add(conn->sndbuf_desc->len,
975 &conn->tx_curs_sent, diff);
976 smc_curs_add(conn->sndbuf_desc->len,
977 &conn->tx_curs_fin, diff);
979 smp_mb__before_atomic();
980 atomic_add(diff, &conn->sndbuf_space);
981 smp_mb__after_atomic();
983 smc_curs_add(conn->peer_rmbe_size,
984 &conn->local_tx_ctrl.prod, diff);
985 smc_curs_add(conn->peer_rmbe_size,
986 &conn->local_tx_ctrl_fin, diff);
988 /* recalculate, value is used by tx_rdma_writes() */
989 atomic_set(&smc->conn.peer_rmbe_space, smc_write_space(conn));
991 if (smc->sk.sk_state != SMC_INIT &&
992 smc->sk.sk_state != SMC_CLOSED) {
993 rc = smcr_cdc_msg_send_validation(conn, pend, wr_buf);
995 queue_delayed_work(conn->lgr->tx_wq, &conn->tx_work, 0);
996 smc->sk.sk_data_ready(&smc->sk);
999 smc_wr_tx_put_slot(conn->lnk,
1000 (struct smc_wr_tx_pend_priv *)pend);
1005 void smc_switch_link_and_count(struct smc_connection *conn,
1006 struct smc_link *to_lnk)
1008 atomic_dec(&conn->lnk->conn_cnt);
1009 /* link_hold in smc_conn_create() */
1010 smcr_link_put(conn->lnk);
1012 atomic_inc(&conn->lnk->conn_cnt);
1013 /* link_put in smc_conn_free() */
1014 smcr_link_hold(conn->lnk);
1017 struct smc_link *smc_switch_conns(struct smc_link_group *lgr,
1018 struct smc_link *from_lnk, bool is_dev_err)
1020 struct smc_link *to_lnk = NULL;
1021 struct smc_cdc_tx_pend *pend;
1022 struct smc_connection *conn;
1023 struct smc_wr_buf *wr_buf;
1024 struct smc_sock *smc;
1025 struct rb_node *node;
1028 /* link is inactive, wake up tx waiters */
1029 smc_wr_wakeup_tx_wait(from_lnk);
1031 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1032 if (!smc_link_active(&lgr->lnk[i]) || i == from_lnk->link_idx)
1034 if (is_dev_err && from_lnk->smcibdev == lgr->lnk[i].smcibdev &&
1035 from_lnk->ibport == lgr->lnk[i].ibport) {
1038 to_lnk = &lgr->lnk[i];
1041 if (!to_lnk || !smc_wr_tx_link_hold(to_lnk)) {
1042 smc_lgr_terminate_sched(lgr);
1046 read_lock_bh(&lgr->conns_lock);
1047 for (node = rb_first(&lgr->conns_all); node; node = rb_next(node)) {
1048 conn = rb_entry(node, struct smc_connection, alert_node);
1049 if (conn->lnk != from_lnk)
1051 smc = container_of(conn, struct smc_sock, conn);
1052 /* conn->lnk not yet set in SMC_INIT state */
1053 if (smc->sk.sk_state == SMC_INIT)
1055 if (smc->sk.sk_state == SMC_CLOSED ||
1056 smc->sk.sk_state == SMC_PEERCLOSEWAIT1 ||
1057 smc->sk.sk_state == SMC_PEERCLOSEWAIT2 ||
1058 smc->sk.sk_state == SMC_APPFINCLOSEWAIT ||
1059 smc->sk.sk_state == SMC_APPCLOSEWAIT1 ||
1060 smc->sk.sk_state == SMC_APPCLOSEWAIT2 ||
1061 smc->sk.sk_state == SMC_PEERFINCLOSEWAIT ||
1062 smc->sk.sk_state == SMC_PEERABORTWAIT ||
1063 smc->sk.sk_state == SMC_PROCESSABORT) {
1064 spin_lock_bh(&conn->send_lock);
1065 smc_switch_link_and_count(conn, to_lnk);
1066 spin_unlock_bh(&conn->send_lock);
1069 sock_hold(&smc->sk);
1070 read_unlock_bh(&lgr->conns_lock);
1071 /* pre-fetch buffer outside of send_lock, might sleep */
1072 rc = smc_cdc_get_free_slot(conn, to_lnk, &wr_buf, NULL, &pend);
1075 /* avoid race with smcr_tx_sndbuf_nonempty() */
1076 spin_lock_bh(&conn->send_lock);
1077 smc_switch_link_and_count(conn, to_lnk);
1078 rc = smc_switch_cursor(smc, pend, wr_buf);
1079 spin_unlock_bh(&conn->send_lock);
1085 read_unlock_bh(&lgr->conns_lock);
1086 smc_wr_tx_link_put(to_lnk);
1090 smcr_link_down_cond_sched(to_lnk);
1091 smc_wr_tx_link_put(to_lnk);
1095 static void smcr_buf_unuse(struct smc_buf_desc *buf_desc, bool is_rmb,
1096 struct smc_link_group *lgr)
1098 struct mutex *lock; /* lock buffer list */
1101 if (is_rmb && buf_desc->is_conf_rkey && !list_empty(&lgr->list)) {
1102 /* unregister rmb with peer */
1103 rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY);
1105 /* protect against smc_llc_cli_rkey_exchange() */
1106 mutex_lock(&lgr->llc_conf_mutex);
1107 smc_llc_do_delete_rkey(lgr, buf_desc);
1108 buf_desc->is_conf_rkey = false;
1109 mutex_unlock(&lgr->llc_conf_mutex);
1110 smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl);
1114 if (buf_desc->is_reg_err) {
1115 /* buf registration failed, reuse not possible */
1116 lock = is_rmb ? &lgr->rmbs_lock :
1119 list_del(&buf_desc->list);
1122 smc_buf_free(lgr, is_rmb, buf_desc);
1124 /* memzero_explicit provides potential memory barrier semantics */
1125 memzero_explicit(buf_desc->cpu_addr, buf_desc->len);
1126 WRITE_ONCE(buf_desc->used, 0);
1130 static void smc_buf_unuse(struct smc_connection *conn,
1131 struct smc_link_group *lgr)
1133 if (conn->sndbuf_desc) {
1134 if (!lgr->is_smcd && conn->sndbuf_desc->is_vm) {
1135 smcr_buf_unuse(conn->sndbuf_desc, false, lgr);
1137 memzero_explicit(conn->sndbuf_desc->cpu_addr, conn->sndbuf_desc->len);
1138 WRITE_ONCE(conn->sndbuf_desc->used, 0);
1141 if (conn->rmb_desc) {
1142 if (!lgr->is_smcd) {
1143 smcr_buf_unuse(conn->rmb_desc, true, lgr);
1145 memzero_explicit(conn->rmb_desc->cpu_addr,
1146 conn->rmb_desc->len + sizeof(struct smcd_cdc_msg));
1147 WRITE_ONCE(conn->rmb_desc->used, 0);
1152 /* remove a finished connection from its link group */
1153 void smc_conn_free(struct smc_connection *conn)
1155 struct smc_link_group *lgr = conn->lgr;
1157 if (!lgr || conn->freed)
1158 /* Connection has never been registered in a
1159 * link group, or has already been freed.
1164 if (!smc_conn_lgr_valid(conn))
1165 /* Connection has already unregistered from
1171 if (!list_empty(&lgr->list))
1172 smc_ism_unset_conn(conn);
1173 tasklet_kill(&conn->rx_tsklet);
1175 smc_cdc_wait_pend_tx_wr(conn);
1176 if (current_work() != &conn->abort_work)
1177 cancel_work_sync(&conn->abort_work);
1179 if (!list_empty(&lgr->list)) {
1180 smc_buf_unuse(conn, lgr); /* allow buffer reuse */
1181 smc_lgr_unregister_conn(conn);
1184 if (!lgr->conns_num)
1185 smc_lgr_schedule_free_work(lgr);
1188 smcr_link_put(conn->lnk); /* link_hold in smc_conn_create() */
1189 smc_lgr_put(lgr); /* lgr_hold in smc_conn_create() */
1192 /* unregister a link from a buf_desc */
1193 static void smcr_buf_unmap_link(struct smc_buf_desc *buf_desc, bool is_rmb,
1194 struct smc_link *lnk)
1196 if (is_rmb || buf_desc->is_vm)
1197 buf_desc->is_reg_mr[lnk->link_idx] = false;
1198 if (!buf_desc->is_map_ib[lnk->link_idx])
1201 if ((is_rmb || buf_desc->is_vm) &&
1202 buf_desc->mr[lnk->link_idx]) {
1203 smc_ib_put_memory_region(buf_desc->mr[lnk->link_idx]);
1204 buf_desc->mr[lnk->link_idx] = NULL;
1207 smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_FROM_DEVICE);
1209 smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_TO_DEVICE);
1211 sg_free_table(&buf_desc->sgt[lnk->link_idx]);
1212 buf_desc->is_map_ib[lnk->link_idx] = false;
1215 /* unmap all buffers of lgr for a deleted link */
1216 static void smcr_buf_unmap_lgr(struct smc_link *lnk)
1218 struct smc_link_group *lgr = lnk->lgr;
1219 struct smc_buf_desc *buf_desc, *bf;
1222 for (i = 0; i < SMC_RMBE_SIZES; i++) {
1223 mutex_lock(&lgr->rmbs_lock);
1224 list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list)
1225 smcr_buf_unmap_link(buf_desc, true, lnk);
1226 mutex_unlock(&lgr->rmbs_lock);
1227 mutex_lock(&lgr->sndbufs_lock);
1228 list_for_each_entry_safe(buf_desc, bf, &lgr->sndbufs[i],
1230 smcr_buf_unmap_link(buf_desc, false, lnk);
1231 mutex_unlock(&lgr->sndbufs_lock);
1235 static void smcr_rtoken_clear_link(struct smc_link *lnk)
1237 struct smc_link_group *lgr = lnk->lgr;
1240 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1241 lgr->rtokens[i][lnk->link_idx].rkey = 0;
1242 lgr->rtokens[i][lnk->link_idx].dma_addr = 0;
1246 static void __smcr_link_clear(struct smc_link *lnk)
1248 struct smc_link_group *lgr = lnk->lgr;
1249 struct smc_ib_device *smcibdev;
1251 smc_wr_free_link_mem(lnk);
1252 smc_ibdev_cnt_dec(lnk);
1253 put_device(&lnk->smcibdev->ibdev->dev);
1254 smcibdev = lnk->smcibdev;
1255 memset(lnk, 0, sizeof(struct smc_link));
1256 lnk->state = SMC_LNK_UNUSED;
1257 if (!atomic_dec_return(&smcibdev->lnk_cnt))
1258 wake_up(&smcibdev->lnks_deleted);
1259 smc_lgr_put(lgr); /* lgr_hold in smcr_link_init() */
1262 /* must be called under lgr->llc_conf_mutex lock */
1263 void smcr_link_clear(struct smc_link *lnk, bool log)
1265 if (!lnk->lgr || lnk->clearing ||
1266 lnk->state == SMC_LNK_UNUSED)
1270 smc_llc_link_clear(lnk, log);
1271 smcr_buf_unmap_lgr(lnk);
1272 smcr_rtoken_clear_link(lnk);
1273 smc_ib_modify_qp_error(lnk);
1274 smc_wr_free_link(lnk);
1275 smc_ib_destroy_queue_pair(lnk);
1276 smc_ib_dealloc_protection_domain(lnk);
1277 smcr_link_put(lnk); /* theoretically last link_put */
1280 void smcr_link_hold(struct smc_link *lnk)
1282 refcount_inc(&lnk->refcnt);
1285 void smcr_link_put(struct smc_link *lnk)
1287 if (refcount_dec_and_test(&lnk->refcnt))
1288 __smcr_link_clear(lnk);
1291 static void smcr_buf_free(struct smc_link_group *lgr, bool is_rmb,
1292 struct smc_buf_desc *buf_desc)
1296 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
1297 smcr_buf_unmap_link(buf_desc, is_rmb, &lgr->lnk[i]);
1299 if (!buf_desc->is_vm && buf_desc->pages)
1300 __free_pages(buf_desc->pages, buf_desc->order);
1301 else if (buf_desc->is_vm && buf_desc->cpu_addr)
1302 vfree(buf_desc->cpu_addr);
1306 static void smcd_buf_free(struct smc_link_group *lgr, bool is_dmb,
1307 struct smc_buf_desc *buf_desc)
1310 /* restore original buf len */
1311 buf_desc->len += sizeof(struct smcd_cdc_msg);
1312 smc_ism_unregister_dmb(lgr->smcd, buf_desc);
1314 kfree(buf_desc->cpu_addr);
1319 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
1320 struct smc_buf_desc *buf_desc)
1323 smcd_buf_free(lgr, is_rmb, buf_desc);
1325 smcr_buf_free(lgr, is_rmb, buf_desc);
1328 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
1330 struct smc_buf_desc *buf_desc, *bf_desc;
1331 struct list_head *buf_list;
1334 for (i = 0; i < SMC_RMBE_SIZES; i++) {
1336 buf_list = &lgr->rmbs[i];
1338 buf_list = &lgr->sndbufs[i];
1339 list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
1341 list_del(&buf_desc->list);
1342 smc_buf_free(lgr, is_rmb, buf_desc);
1347 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
1349 /* free send buffers */
1350 __smc_lgr_free_bufs(lgr, false);
1352 __smc_lgr_free_bufs(lgr, true);
1355 /* won't be freed until no one accesses to lgr anymore */
1356 static void __smc_lgr_free(struct smc_link_group *lgr)
1358 smc_lgr_free_bufs(lgr);
1360 if (!atomic_dec_return(&lgr->smcd->lgr_cnt))
1361 wake_up(&lgr->smcd->lgrs_deleted);
1363 smc_wr_free_lgr_mem(lgr);
1364 if (!atomic_dec_return(&lgr_cnt))
1365 wake_up(&lgrs_deleted);
1370 /* remove a link group */
1371 static void smc_lgr_free(struct smc_link_group *lgr)
1375 if (!lgr->is_smcd) {
1376 mutex_lock(&lgr->llc_conf_mutex);
1377 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1378 if (lgr->lnk[i].state != SMC_LNK_UNUSED)
1379 smcr_link_clear(&lgr->lnk[i], false);
1381 mutex_unlock(&lgr->llc_conf_mutex);
1382 smc_llc_lgr_clear(lgr);
1385 destroy_workqueue(lgr->tx_wq);
1387 smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
1388 put_device(&lgr->smcd->dev);
1390 smc_lgr_put(lgr); /* theoretically last lgr_put */
1393 void smc_lgr_hold(struct smc_link_group *lgr)
1395 refcount_inc(&lgr->refcnt);
1398 void smc_lgr_put(struct smc_link_group *lgr)
1400 if (refcount_dec_and_test(&lgr->refcnt))
1401 __smc_lgr_free(lgr);
1404 static void smc_sk_wake_ups(struct smc_sock *smc)
1406 smc->sk.sk_write_space(&smc->sk);
1407 smc->sk.sk_data_ready(&smc->sk);
1408 smc->sk.sk_state_change(&smc->sk);
1411 /* kill a connection */
1412 static void smc_conn_kill(struct smc_connection *conn, bool soft)
1414 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1416 if (conn->lgr->is_smcd && conn->lgr->peer_shutdown)
1417 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
1419 smc_close_abort(conn);
1421 smc->sk.sk_err = ECONNABORTED;
1422 smc_sk_wake_ups(smc);
1423 if (conn->lgr->is_smcd) {
1424 smc_ism_unset_conn(conn);
1426 tasklet_kill(&conn->rx_tsklet);
1428 tasklet_unlock_wait(&conn->rx_tsklet);
1430 smc_cdc_wait_pend_tx_wr(conn);
1432 smc_lgr_unregister_conn(conn);
1433 smc_close_active_abort(smc);
1436 static void smc_lgr_cleanup(struct smc_link_group *lgr)
1439 smc_ism_signal_shutdown(lgr);
1441 u32 rsn = lgr->llc_termination_rsn;
1444 rsn = SMC_LLC_DEL_PROG_INIT_TERM;
1445 smc_llc_send_link_delete_all(lgr, false, rsn);
1446 smcr_lgr_link_deactivate_all(lgr);
1450 /* terminate link group
1451 * @soft: true if link group shutdown can take its time
1452 * false if immediate link group shutdown is required
1454 static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft)
1456 struct smc_connection *conn;
1457 struct smc_sock *smc;
1458 struct rb_node *node;
1460 if (lgr->terminating)
1461 return; /* lgr already terminating */
1462 /* cancel free_work sync, will terminate when lgr->freeing is set */
1463 cancel_delayed_work(&lgr->free_work);
1464 lgr->terminating = 1;
1466 /* kill remaining link group connections */
1467 read_lock_bh(&lgr->conns_lock);
1468 node = rb_first(&lgr->conns_all);
1470 read_unlock_bh(&lgr->conns_lock);
1471 conn = rb_entry(node, struct smc_connection, alert_node);
1472 smc = container_of(conn, struct smc_sock, conn);
1473 sock_hold(&smc->sk); /* sock_put below */
1474 lock_sock(&smc->sk);
1475 smc_conn_kill(conn, soft);
1476 release_sock(&smc->sk);
1477 sock_put(&smc->sk); /* sock_hold above */
1478 read_lock_bh(&lgr->conns_lock);
1479 node = rb_first(&lgr->conns_all);
1481 read_unlock_bh(&lgr->conns_lock);
1482 smc_lgr_cleanup(lgr);
1486 /* unlink link group and schedule termination */
1487 void smc_lgr_terminate_sched(struct smc_link_group *lgr)
1489 spinlock_t *lgr_lock;
1491 smc_lgr_list_head(lgr, &lgr_lock);
1492 spin_lock_bh(lgr_lock);
1493 if (list_empty(&lgr->list) || lgr->terminating || lgr->freeing) {
1494 spin_unlock_bh(lgr_lock);
1495 return; /* lgr already terminating */
1497 list_del_init(&lgr->list);
1499 spin_unlock_bh(lgr_lock);
1500 schedule_work(&lgr->terminate_work);
1503 /* Called when peer lgr shutdown (regularly or abnormally) is received */
1504 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid, unsigned short vlan)
1506 struct smc_link_group *lgr, *l;
1507 LIST_HEAD(lgr_free_list);
1509 /* run common cleanup function and build free list */
1510 spin_lock_bh(&dev->lgr_lock);
1511 list_for_each_entry_safe(lgr, l, &dev->lgr_list, list) {
1512 if ((!peer_gid || lgr->peer_gid == peer_gid) &&
1513 (vlan == VLAN_VID_MASK || lgr->vlan_id == vlan)) {
1514 if (peer_gid) /* peer triggered termination */
1515 lgr->peer_shutdown = 1;
1516 list_move(&lgr->list, &lgr_free_list);
1520 spin_unlock_bh(&dev->lgr_lock);
1522 /* cancel the regular free workers and actually free lgrs */
1523 list_for_each_entry_safe(lgr, l, &lgr_free_list, list) {
1524 list_del_init(&lgr->list);
1525 schedule_work(&lgr->terminate_work);
1529 /* Called when an SMCD device is removed or the smc module is unloaded */
1530 void smc_smcd_terminate_all(struct smcd_dev *smcd)
1532 struct smc_link_group *lgr, *lg;
1533 LIST_HEAD(lgr_free_list);
1535 spin_lock_bh(&smcd->lgr_lock);
1536 list_splice_init(&smcd->lgr_list, &lgr_free_list);
1537 list_for_each_entry(lgr, &lgr_free_list, list)
1539 spin_unlock_bh(&smcd->lgr_lock);
1541 list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
1542 list_del_init(&lgr->list);
1543 __smc_lgr_terminate(lgr, false);
1546 if (atomic_read(&smcd->lgr_cnt))
1547 wait_event(smcd->lgrs_deleted, !atomic_read(&smcd->lgr_cnt));
1550 /* Called when an SMCR device is removed or the smc module is unloaded.
1551 * If smcibdev is given, all SMCR link groups using this device are terminated.
1552 * If smcibdev is NULL, all SMCR link groups are terminated.
1554 void smc_smcr_terminate_all(struct smc_ib_device *smcibdev)
1556 struct smc_link_group *lgr, *lg;
1557 LIST_HEAD(lgr_free_list);
1560 spin_lock_bh(&smc_lgr_list.lock);
1562 list_splice_init(&smc_lgr_list.list, &lgr_free_list);
1563 list_for_each_entry(lgr, &lgr_free_list, list)
1566 list_for_each_entry_safe(lgr, lg, &smc_lgr_list.list, list) {
1567 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1568 if (lgr->lnk[i].smcibdev == smcibdev)
1569 smcr_link_down_cond_sched(&lgr->lnk[i]);
1573 spin_unlock_bh(&smc_lgr_list.lock);
1575 list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
1576 list_del_init(&lgr->list);
1577 smc_llc_set_termination_rsn(lgr, SMC_LLC_DEL_OP_INIT_TERM);
1578 __smc_lgr_terminate(lgr, false);
1582 if (atomic_read(&smcibdev->lnk_cnt))
1583 wait_event(smcibdev->lnks_deleted,
1584 !atomic_read(&smcibdev->lnk_cnt));
1586 if (atomic_read(&lgr_cnt))
1587 wait_event(lgrs_deleted, !atomic_read(&lgr_cnt));
1591 /* set new lgr type and clear all asymmetric link tagging */
1592 void smcr_lgr_set_type(struct smc_link_group *lgr, enum smc_lgr_type new_type)
1594 char *lgr_type = "";
1597 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
1598 if (smc_link_usable(&lgr->lnk[i]))
1599 lgr->lnk[i].link_is_asym = false;
1600 if (lgr->type == new_type)
1602 lgr->type = new_type;
1604 switch (lgr->type) {
1608 case SMC_LGR_SINGLE:
1609 lgr_type = "SINGLE";
1611 case SMC_LGR_SYMMETRIC:
1612 lgr_type = "SYMMETRIC";
1614 case SMC_LGR_ASYMMETRIC_PEER:
1615 lgr_type = "ASYMMETRIC_PEER";
1617 case SMC_LGR_ASYMMETRIC_LOCAL:
1618 lgr_type = "ASYMMETRIC_LOCAL";
1621 pr_warn_ratelimited("smc: SMC-R lg %*phN net %llu state changed: "
1622 "%s, pnetid %.16s\n", SMC_LGR_ID_SIZE, &lgr->id,
1623 lgr->net->net_cookie, lgr_type, lgr->pnet_id);
1626 /* set new lgr type and tag a link as asymmetric */
1627 void smcr_lgr_set_type_asym(struct smc_link_group *lgr,
1628 enum smc_lgr_type new_type, int asym_lnk_idx)
1630 smcr_lgr_set_type(lgr, new_type);
1631 lgr->lnk[asym_lnk_idx].link_is_asym = true;
1634 /* abort connection, abort_work scheduled from tasklet context */
1635 static void smc_conn_abort_work(struct work_struct *work)
1637 struct smc_connection *conn = container_of(work,
1638 struct smc_connection,
1640 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1642 lock_sock(&smc->sk);
1643 smc_conn_kill(conn, true);
1644 release_sock(&smc->sk);
1645 sock_put(&smc->sk); /* sock_hold done by schedulers of abort_work */
1648 void smcr_port_add(struct smc_ib_device *smcibdev, u8 ibport)
1650 struct smc_link_group *lgr, *n;
1652 list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1653 struct smc_link *link;
1655 if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1656 SMC_MAX_PNETID_LEN) ||
1657 lgr->type == SMC_LGR_SYMMETRIC ||
1658 lgr->type == SMC_LGR_ASYMMETRIC_PEER ||
1659 !rdma_dev_access_netns(smcibdev->ibdev, lgr->net))
1662 /* trigger local add link processing */
1663 link = smc_llc_usable_link(lgr);
1665 smc_llc_add_link_local(link);
1669 /* link is down - switch connections to alternate link,
1670 * must be called under lgr->llc_conf_mutex lock
1672 static void smcr_link_down(struct smc_link *lnk)
1674 struct smc_link_group *lgr = lnk->lgr;
1675 struct smc_link *to_lnk;
1678 if (!lgr || lnk->state == SMC_LNK_UNUSED || list_empty(&lgr->list))
1681 to_lnk = smc_switch_conns(lgr, lnk, true);
1682 if (!to_lnk) { /* no backup link available */
1683 smcr_link_clear(lnk, true);
1686 smcr_lgr_set_type(lgr, SMC_LGR_SINGLE);
1687 del_link_id = lnk->link_id;
1689 if (lgr->role == SMC_SERV) {
1690 /* trigger local delete link processing */
1691 smc_llc_srv_delete_link_local(to_lnk, del_link_id);
1693 if (lgr->llc_flow_lcl.type != SMC_LLC_FLOW_NONE) {
1694 /* another llc task is ongoing */
1695 mutex_unlock(&lgr->llc_conf_mutex);
1696 wait_event_timeout(lgr->llc_flow_waiter,
1697 (list_empty(&lgr->list) ||
1698 lgr->llc_flow_lcl.type == SMC_LLC_FLOW_NONE),
1700 mutex_lock(&lgr->llc_conf_mutex);
1702 if (!list_empty(&lgr->list)) {
1703 smc_llc_send_delete_link(to_lnk, del_link_id,
1705 SMC_LLC_DEL_LOST_PATH);
1706 smcr_link_clear(lnk, true);
1708 wake_up(&lgr->llc_flow_waiter); /* wake up next waiter */
1712 /* must be called under lgr->llc_conf_mutex lock */
1713 void smcr_link_down_cond(struct smc_link *lnk)
1715 if (smc_link_downing(&lnk->state)) {
1716 trace_smcr_link_down(lnk, __builtin_return_address(0));
1717 smcr_link_down(lnk);
1721 /* will get the lgr->llc_conf_mutex lock */
1722 void smcr_link_down_cond_sched(struct smc_link *lnk)
1724 if (smc_link_downing(&lnk->state)) {
1725 trace_smcr_link_down(lnk, __builtin_return_address(0));
1726 schedule_work(&lnk->link_down_wrk);
1730 void smcr_port_err(struct smc_ib_device *smcibdev, u8 ibport)
1732 struct smc_link_group *lgr, *n;
1735 list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1736 if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1737 SMC_MAX_PNETID_LEN))
1738 continue; /* lgr is not affected */
1739 if (list_empty(&lgr->list))
1741 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1742 struct smc_link *lnk = &lgr->lnk[i];
1744 if (smc_link_usable(lnk) &&
1745 lnk->smcibdev == smcibdev && lnk->ibport == ibport)
1746 smcr_link_down_cond_sched(lnk);
1751 static void smc_link_down_work(struct work_struct *work)
1753 struct smc_link *link = container_of(work, struct smc_link,
1755 struct smc_link_group *lgr = link->lgr;
1757 if (list_empty(&lgr->list))
1759 wake_up_all(&lgr->llc_msg_waiter);
1760 mutex_lock(&lgr->llc_conf_mutex);
1761 smcr_link_down(link);
1762 mutex_unlock(&lgr->llc_conf_mutex);
1765 static int smc_vlan_by_tcpsk_walk(struct net_device *lower_dev,
1766 struct netdev_nested_priv *priv)
1768 unsigned short *vlan_id = (unsigned short *)priv->data;
1770 if (is_vlan_dev(lower_dev)) {
1771 *vlan_id = vlan_dev_vlan_id(lower_dev);
1778 /* Determine vlan of internal TCP socket. */
1779 int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini)
1781 struct dst_entry *dst = sk_dst_get(clcsock->sk);
1782 struct netdev_nested_priv priv;
1783 struct net_device *ndev;
1797 if (is_vlan_dev(ndev)) {
1798 ini->vlan_id = vlan_dev_vlan_id(ndev);
1802 priv.data = (void *)&ini->vlan_id;
1804 netdev_walk_all_lower_dev(ndev, smc_vlan_by_tcpsk_walk, &priv);
1813 static bool smcr_lgr_match(struct smc_link_group *lgr, u8 smcr_version,
1817 enum smc_lgr_role role, u32 clcqpn,
1820 struct smc_link *lnk;
1823 if (memcmp(lgr->peer_systemid, peer_systemid, SMC_SYSTEMID_LEN) ||
1827 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1830 if (!smc_link_active(lnk))
1832 /* use verbs API to check netns, instead of lgr->net */
1833 if (!rdma_dev_access_netns(lnk->smcibdev->ibdev, net))
1835 if ((lgr->role == SMC_SERV || lnk->peer_qpn == clcqpn) &&
1836 !memcmp(lnk->peer_gid, peer_gid, SMC_GID_SIZE) &&
1837 (smcr_version == SMC_V2 ||
1838 !memcmp(lnk->peer_mac, peer_mac_v1, ETH_ALEN)))
1844 static bool smcd_lgr_match(struct smc_link_group *lgr,
1845 struct smcd_dev *smcismdev, u64 peer_gid)
1847 return lgr->peer_gid == peer_gid && lgr->smcd == smcismdev;
1850 /* create a new SMC connection (and a new link group if necessary) */
1851 int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini)
1853 struct smc_connection *conn = &smc->conn;
1854 struct net *net = sock_net(&smc->sk);
1855 struct list_head *lgr_list;
1856 struct smc_link_group *lgr;
1857 enum smc_lgr_role role;
1858 spinlock_t *lgr_lock;
1861 lgr_list = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_list :
1863 lgr_lock = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_lock :
1865 ini->first_contact_local = 1;
1866 role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
1867 if (role == SMC_CLNT && ini->first_contact_peer)
1868 /* create new link group as well */
1871 /* determine if an existing link group can be reused */
1872 spin_lock_bh(lgr_lock);
1873 list_for_each_entry(lgr, lgr_list, list) {
1874 write_lock_bh(&lgr->conns_lock);
1876 smcd_lgr_match(lgr, ini->ism_dev[ini->ism_selected],
1877 ini->ism_peer_gid[ini->ism_selected]) :
1878 smcr_lgr_match(lgr, ini->smcr_version,
1880 ini->peer_gid, ini->peer_mac, role,
1881 ini->ib_clcqpn, net)) &&
1883 (ini->smcd_version == SMC_V2 ||
1884 lgr->vlan_id == ini->vlan_id) &&
1885 (role == SMC_CLNT || ini->is_smcd ||
1886 (lgr->conns_num < SMC_RMBS_PER_LGR_MAX &&
1887 !bitmap_full(lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX)))) {
1888 /* link group found */
1889 ini->first_contact_local = 0;
1891 rc = smc_lgr_register_conn(conn, false);
1892 write_unlock_bh(&lgr->conns_lock);
1893 if (!rc && delayed_work_pending(&lgr->free_work))
1894 cancel_delayed_work(&lgr->free_work);
1897 write_unlock_bh(&lgr->conns_lock);
1899 spin_unlock_bh(lgr_lock);
1903 if (role == SMC_CLNT && !ini->first_contact_peer &&
1904 ini->first_contact_local) {
1905 /* Server reuses a link group, but Client wants to start
1907 * send out_of_sync decline, reason synchr. error
1909 return SMC_CLC_DECL_SYNCERR;
1913 if (ini->first_contact_local) {
1914 rc = smc_lgr_create(smc, ini);
1918 write_lock_bh(&lgr->conns_lock);
1919 rc = smc_lgr_register_conn(conn, true);
1920 write_unlock_bh(&lgr->conns_lock);
1922 smc_lgr_cleanup_early(lgr);
1926 smc_lgr_hold(conn->lgr); /* lgr_put in smc_conn_free() */
1927 if (!conn->lgr->is_smcd)
1928 smcr_link_hold(conn->lnk); /* link_put in smc_conn_free() */
1930 conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
1931 conn->local_tx_ctrl.len = SMC_WR_TX_SIZE;
1932 conn->urg_state = SMC_URG_READ;
1933 init_waitqueue_head(&conn->cdc_pend_tx_wq);
1934 INIT_WORK(&smc->conn.abort_work, smc_conn_abort_work);
1936 conn->rx_off = sizeof(struct smcd_cdc_msg);
1937 smcd_cdc_rx_init(conn); /* init tasklet for this conn */
1941 #ifndef KERNEL_HAS_ATOMIC64
1942 spin_lock_init(&conn->acurs_lock);
1949 #define SMCD_DMBE_SIZES 6 /* 0 -> 16KB, 1 -> 32KB, .. 6 -> 1MB */
1950 #define SMCR_RMBE_SIZES 5 /* 0 -> 16KB, 1 -> 32KB, .. 5 -> 512KB */
1952 /* convert the RMB size into the compressed notation (minimum 16K, see
1953 * SMCD/R_DMBE_SIZES.
1954 * In contrast to plain ilog2, this rounds towards the next power of 2,
1955 * so the socket application gets at least its desired sndbuf / rcvbuf size.
1957 static u8 smc_compress_bufsize(int size, bool is_smcd, bool is_rmb)
1959 const unsigned int max_scat = SG_MAX_SINGLE_ALLOC * PAGE_SIZE;
1962 if (size <= SMC_BUF_MIN_SIZE)
1965 size = (size - 1) >> 14; /* convert to 16K multiple */
1966 compressed = min_t(u8, ilog2(size) + 1,
1967 is_smcd ? SMCD_DMBE_SIZES : SMCR_RMBE_SIZES);
1969 if (!is_smcd && is_rmb)
1970 /* RMBs are backed by & limited to max size of scatterlists */
1971 compressed = min_t(u8, compressed, ilog2(max_scat >> 14));
1976 /* convert the RMB size from compressed notation into integer */
1977 int smc_uncompress_bufsize(u8 compressed)
1981 size = 0x00000001 << (((int)compressed) + 14);
1985 /* try to reuse a sndbuf or rmb description slot for a certain
1986 * buffer size; if not available, return NULL
1988 static struct smc_buf_desc *smc_buf_get_slot(int compressed_bufsize,
1990 struct list_head *buf_list)
1992 struct smc_buf_desc *buf_slot;
1995 list_for_each_entry(buf_slot, buf_list, list) {
1996 if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
2005 /* one of the conditions for announcing a receiver's current window size is
2006 * that it "results in a minimum increase in the window size of 10% of the
2007 * receive buffer space" [RFC7609]
2009 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
2011 return max_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
2014 /* map an buf to a link */
2015 static int smcr_buf_map_link(struct smc_buf_desc *buf_desc, bool is_rmb,
2016 struct smc_link *lnk)
2018 int rc, i, nents, offset, buf_size, size, access_flags;
2019 struct scatterlist *sg;
2022 if (buf_desc->is_map_ib[lnk->link_idx])
2025 if (buf_desc->is_vm) {
2026 buf = buf_desc->cpu_addr;
2027 buf_size = buf_desc->len;
2028 offset = offset_in_page(buf_desc->cpu_addr);
2029 nents = PAGE_ALIGN(buf_size + offset) / PAGE_SIZE;
2034 rc = sg_alloc_table(&buf_desc->sgt[lnk->link_idx], nents, GFP_KERNEL);
2038 if (buf_desc->is_vm) {
2039 /* virtually contiguous buffer */
2040 for_each_sg(buf_desc->sgt[lnk->link_idx].sgl, sg, nents, i) {
2041 size = min_t(int, PAGE_SIZE - offset, buf_size);
2042 sg_set_page(sg, vmalloc_to_page(buf), size, offset);
2043 buf += size / sizeof(*buf);
2048 /* physically contiguous buffer */
2049 sg_set_buf(buf_desc->sgt[lnk->link_idx].sgl,
2050 buf_desc->cpu_addr, buf_desc->len);
2053 /* map sg table to DMA address */
2054 rc = smc_ib_buf_map_sg(lnk, buf_desc,
2055 is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
2056 /* SMC protocol depends on mapping to one DMA address only */
2062 buf_desc->is_dma_need_sync |=
2063 smc_ib_is_sg_need_sync(lnk, buf_desc) << lnk->link_idx;
2065 if (is_rmb || buf_desc->is_vm) {
2066 /* create a new memory region for the RMB or vzalloced sndbuf */
2067 access_flags = is_rmb ?
2068 IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE :
2069 IB_ACCESS_LOCAL_WRITE;
2071 rc = smc_ib_get_memory_region(lnk->roce_pd, access_flags,
2072 buf_desc, lnk->link_idx);
2075 smc_ib_sync_sg_for_device(lnk, buf_desc,
2076 is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
2078 buf_desc->is_map_ib[lnk->link_idx] = true;
2082 smc_ib_buf_unmap_sg(lnk, buf_desc,
2083 is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
2085 sg_free_table(&buf_desc->sgt[lnk->link_idx]);
2089 /* register a new buf on IB device, rmb or vzalloced sndbuf
2090 * must be called under lgr->llc_conf_mutex lock
2092 int smcr_link_reg_buf(struct smc_link *link, struct smc_buf_desc *buf_desc)
2094 if (list_empty(&link->lgr->list))
2096 if (!buf_desc->is_reg_mr[link->link_idx]) {
2097 /* register memory region for new buf */
2098 if (buf_desc->is_vm)
2099 buf_desc->mr[link->link_idx]->iova =
2100 (uintptr_t)buf_desc->cpu_addr;
2101 if (smc_wr_reg_send(link, buf_desc->mr[link->link_idx])) {
2102 buf_desc->is_reg_err = true;
2105 buf_desc->is_reg_mr[link->link_idx] = true;
2110 static int _smcr_buf_map_lgr(struct smc_link *lnk, struct mutex *lock,
2111 struct list_head *lst, bool is_rmb)
2113 struct smc_buf_desc *buf_desc, *bf;
2117 list_for_each_entry_safe(buf_desc, bf, lst, list) {
2118 if (!buf_desc->used)
2120 rc = smcr_buf_map_link(buf_desc, is_rmb, lnk);
2129 /* map all used buffers of lgr for a new link */
2130 int smcr_buf_map_lgr(struct smc_link *lnk)
2132 struct smc_link_group *lgr = lnk->lgr;
2135 for (i = 0; i < SMC_RMBE_SIZES; i++) {
2136 rc = _smcr_buf_map_lgr(lnk, &lgr->rmbs_lock,
2137 &lgr->rmbs[i], true);
2140 rc = _smcr_buf_map_lgr(lnk, &lgr->sndbufs_lock,
2141 &lgr->sndbufs[i], false);
2148 /* register all used buffers of lgr for a new link,
2149 * must be called under lgr->llc_conf_mutex lock
2151 int smcr_buf_reg_lgr(struct smc_link *lnk)
2153 struct smc_link_group *lgr = lnk->lgr;
2154 struct smc_buf_desc *buf_desc, *bf;
2157 /* reg all RMBs for a new link */
2158 mutex_lock(&lgr->rmbs_lock);
2159 for (i = 0; i < SMC_RMBE_SIZES; i++) {
2160 list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list) {
2161 if (!buf_desc->used)
2163 rc = smcr_link_reg_buf(lnk, buf_desc);
2165 mutex_unlock(&lgr->rmbs_lock);
2170 mutex_unlock(&lgr->rmbs_lock);
2172 if (lgr->buf_type == SMCR_PHYS_CONT_BUFS)
2175 /* reg all vzalloced sndbufs for a new link */
2176 mutex_lock(&lgr->sndbufs_lock);
2177 for (i = 0; i < SMC_RMBE_SIZES; i++) {
2178 list_for_each_entry_safe(buf_desc, bf, &lgr->sndbufs[i], list) {
2179 if (!buf_desc->used || !buf_desc->is_vm)
2181 rc = smcr_link_reg_buf(lnk, buf_desc);
2183 mutex_unlock(&lgr->sndbufs_lock);
2188 mutex_unlock(&lgr->sndbufs_lock);
2192 static struct smc_buf_desc *smcr_new_buf_create(struct smc_link_group *lgr,
2193 bool is_rmb, int bufsize)
2195 struct smc_buf_desc *buf_desc;
2197 /* try to alloc a new buffer */
2198 buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
2200 return ERR_PTR(-ENOMEM);
2202 switch (lgr->buf_type) {
2203 case SMCR_PHYS_CONT_BUFS:
2204 case SMCR_MIXED_BUFS:
2205 buf_desc->order = get_order(bufsize);
2206 buf_desc->pages = alloc_pages(GFP_KERNEL | __GFP_NOWARN |
2207 __GFP_NOMEMALLOC | __GFP_COMP |
2208 __GFP_NORETRY | __GFP_ZERO,
2210 if (buf_desc->pages) {
2211 buf_desc->cpu_addr =
2212 (void *)page_address(buf_desc->pages);
2213 buf_desc->len = bufsize;
2214 buf_desc->is_vm = false;
2217 if (lgr->buf_type == SMCR_PHYS_CONT_BUFS)
2219 fallthrough; // try virtually continguous buf
2220 case SMCR_VIRT_CONT_BUFS:
2221 buf_desc->order = get_order(bufsize);
2222 buf_desc->cpu_addr = vzalloc(PAGE_SIZE << buf_desc->order);
2223 if (!buf_desc->cpu_addr)
2225 buf_desc->pages = NULL;
2226 buf_desc->len = bufsize;
2227 buf_desc->is_vm = true;
2234 return ERR_PTR(-EAGAIN);
2237 /* map buf_desc on all usable links,
2238 * unused buffers stay mapped as long as the link is up
2240 static int smcr_buf_map_usable_links(struct smc_link_group *lgr,
2241 struct smc_buf_desc *buf_desc, bool is_rmb)
2243 int i, rc = 0, cnt = 0;
2245 /* protect against parallel link reconfiguration */
2246 mutex_lock(&lgr->llc_conf_mutex);
2247 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
2248 struct smc_link *lnk = &lgr->lnk[i];
2250 if (!smc_link_usable(lnk))
2252 if (smcr_buf_map_link(buf_desc, is_rmb, lnk)) {
2259 mutex_unlock(&lgr->llc_conf_mutex);
2265 static struct smc_buf_desc *smcd_new_buf_create(struct smc_link_group *lgr,
2266 bool is_dmb, int bufsize)
2268 struct smc_buf_desc *buf_desc;
2271 /* try to alloc a new DMB */
2272 buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
2274 return ERR_PTR(-ENOMEM);
2276 rc = smc_ism_register_dmb(lgr, bufsize, buf_desc);
2280 return ERR_PTR(-EAGAIN);
2282 return ERR_PTR(-ENOSPC);
2283 return ERR_PTR(-EIO);
2285 buf_desc->pages = virt_to_page(buf_desc->cpu_addr);
2286 /* CDC header stored in buf. So, pretend it was smaller */
2287 buf_desc->len = bufsize - sizeof(struct smcd_cdc_msg);
2289 buf_desc->cpu_addr = kzalloc(bufsize, GFP_KERNEL |
2290 __GFP_NOWARN | __GFP_NORETRY |
2292 if (!buf_desc->cpu_addr) {
2294 return ERR_PTR(-EAGAIN);
2296 buf_desc->len = bufsize;
2301 static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
2303 struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
2304 struct smc_connection *conn = &smc->conn;
2305 struct smc_link_group *lgr = conn->lgr;
2306 struct list_head *buf_list;
2307 int bufsize, bufsize_short;
2308 bool is_dgraded = false;
2309 struct mutex *lock; /* lock buffer list */
2313 /* use socket recv buffer size (w/o overhead) as start value */
2314 sk_buf_size = smc->sk.sk_rcvbuf;
2316 /* use socket send buffer size (w/o overhead) as start value */
2317 sk_buf_size = smc->sk.sk_sndbuf;
2319 for (bufsize_short = smc_compress_bufsize(sk_buf_size, is_smcd, is_rmb);
2320 bufsize_short >= 0; bufsize_short--) {
2322 lock = &lgr->rmbs_lock;
2323 buf_list = &lgr->rmbs[bufsize_short];
2325 lock = &lgr->sndbufs_lock;
2326 buf_list = &lgr->sndbufs[bufsize_short];
2328 bufsize = smc_uncompress_bufsize(bufsize_short);
2330 /* check for reusable slot in the link group */
2331 buf_desc = smc_buf_get_slot(bufsize_short, lock, buf_list);
2333 buf_desc->is_dma_need_sync = 0;
2334 SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, bufsize);
2335 SMC_STAT_BUF_REUSE(smc, is_smcd, is_rmb);
2336 break; /* found reusable slot */
2340 buf_desc = smcd_new_buf_create(lgr, is_rmb, bufsize);
2342 buf_desc = smcr_new_buf_create(lgr, is_rmb, bufsize);
2344 if (PTR_ERR(buf_desc) == -ENOMEM)
2346 if (IS_ERR(buf_desc)) {
2349 SMC_STAT_RMB_DOWNGRADED(smc, is_smcd, is_rmb);
2354 SMC_STAT_RMB_ALLOC(smc, is_smcd, is_rmb);
2355 SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, bufsize);
2358 list_add(&buf_desc->list, buf_list);
2363 if (IS_ERR(buf_desc))
2364 return PTR_ERR(buf_desc);
2367 if (smcr_buf_map_usable_links(lgr, buf_desc, is_rmb)) {
2368 smcr_buf_unuse(buf_desc, is_rmb, lgr);
2374 conn->rmb_desc = buf_desc;
2375 conn->rmbe_size_short = bufsize_short;
2376 smc->sk.sk_rcvbuf = bufsize;
2377 atomic_set(&conn->bytes_to_rcv, 0);
2378 conn->rmbe_update_limit =
2379 smc_rmb_wnd_update_limit(buf_desc->len);
2381 smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */
2383 conn->sndbuf_desc = buf_desc;
2384 smc->sk.sk_sndbuf = bufsize;
2385 atomic_set(&conn->sndbuf_space, bufsize);
2390 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
2392 if (!conn->sndbuf_desc->is_dma_need_sync)
2394 if (!smc_conn_lgr_valid(conn) || conn->lgr->is_smcd ||
2395 !smc_link_active(conn->lnk))
2397 smc_ib_sync_sg_for_device(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE);
2400 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
2404 if (!conn->rmb_desc->is_dma_need_sync)
2406 if (!smc_conn_lgr_valid(conn) || conn->lgr->is_smcd)
2408 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
2409 if (!smc_link_active(&conn->lgr->lnk[i]))
2411 smc_ib_sync_sg_for_cpu(&conn->lgr->lnk[i], conn->rmb_desc,
2416 /* create the send and receive buffer for an SMC socket;
2417 * receive buffers are called RMBs;
2418 * (even though the SMC protocol allows more than one RMB-element per RMB,
2419 * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
2420 * extra RMB for every connection in a link group
2422 int smc_buf_create(struct smc_sock *smc, bool is_smcd)
2426 /* create send buffer */
2427 rc = __smc_buf_create(smc, is_smcd, false);
2431 rc = __smc_buf_create(smc, is_smcd, true);
2433 mutex_lock(&smc->conn.lgr->sndbufs_lock);
2434 list_del(&smc->conn.sndbuf_desc->list);
2435 mutex_unlock(&smc->conn.lgr->sndbufs_lock);
2436 smc_buf_free(smc->conn.lgr, false, smc->conn.sndbuf_desc);
2437 smc->conn.sndbuf_desc = NULL;
2442 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
2446 for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
2447 if (!test_and_set_bit(i, lgr->rtokens_used_mask))
2453 static int smc_rtoken_find_by_link(struct smc_link_group *lgr, int lnk_idx,
2458 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2459 if (test_bit(i, lgr->rtokens_used_mask) &&
2460 lgr->rtokens[i][lnk_idx].rkey == rkey)
2466 /* set rtoken for a new link to an existing rmb */
2467 void smc_rtoken_set(struct smc_link_group *lgr, int link_idx, int link_idx_new,
2468 __be32 nw_rkey_known, __be64 nw_vaddr, __be32 nw_rkey)
2472 rtok_idx = smc_rtoken_find_by_link(lgr, link_idx, ntohl(nw_rkey_known));
2473 if (rtok_idx == -ENOENT)
2475 lgr->rtokens[rtok_idx][link_idx_new].rkey = ntohl(nw_rkey);
2476 lgr->rtokens[rtok_idx][link_idx_new].dma_addr = be64_to_cpu(nw_vaddr);
2479 /* set rtoken for a new link whose link_id is given */
2480 void smc_rtoken_set2(struct smc_link_group *lgr, int rtok_idx, int link_id,
2481 __be64 nw_vaddr, __be32 nw_rkey)
2483 u64 dma_addr = be64_to_cpu(nw_vaddr);
2484 u32 rkey = ntohl(nw_rkey);
2488 for (link_idx = 0; link_idx < SMC_LINKS_PER_LGR_MAX; link_idx++) {
2489 if (lgr->lnk[link_idx].link_id == link_id) {
2496 lgr->rtokens[rtok_idx][link_idx].rkey = rkey;
2497 lgr->rtokens[rtok_idx][link_idx].dma_addr = dma_addr;
2500 /* add a new rtoken from peer */
2501 int smc_rtoken_add(struct smc_link *lnk, __be64 nw_vaddr, __be32 nw_rkey)
2503 struct smc_link_group *lgr = smc_get_lgr(lnk);
2504 u64 dma_addr = be64_to_cpu(nw_vaddr);
2505 u32 rkey = ntohl(nw_rkey);
2508 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2509 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
2510 lgr->rtokens[i][lnk->link_idx].dma_addr == dma_addr &&
2511 test_bit(i, lgr->rtokens_used_mask)) {
2512 /* already in list */
2516 i = smc_rmb_reserve_rtoken_idx(lgr);
2519 lgr->rtokens[i][lnk->link_idx].rkey = rkey;
2520 lgr->rtokens[i][lnk->link_idx].dma_addr = dma_addr;
2524 /* delete an rtoken from all links */
2525 int smc_rtoken_delete(struct smc_link *lnk, __be32 nw_rkey)
2527 struct smc_link_group *lgr = smc_get_lgr(lnk);
2528 u32 rkey = ntohl(nw_rkey);
2531 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2532 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
2533 test_bit(i, lgr->rtokens_used_mask)) {
2534 for (j = 0; j < SMC_LINKS_PER_LGR_MAX; j++) {
2535 lgr->rtokens[i][j].rkey = 0;
2536 lgr->rtokens[i][j].dma_addr = 0;
2538 clear_bit(i, lgr->rtokens_used_mask);
2545 /* save rkey and dma_addr received from peer during clc handshake */
2546 int smc_rmb_rtoken_handling(struct smc_connection *conn,
2547 struct smc_link *lnk,
2548 struct smc_clc_msg_accept_confirm *clc)
2550 conn->rtoken_idx = smc_rtoken_add(lnk, clc->r0.rmb_dma_addr,
2552 if (conn->rtoken_idx < 0)
2553 return conn->rtoken_idx;
2557 static void smc_core_going_away(void)
2559 struct smc_ib_device *smcibdev;
2560 struct smcd_dev *smcd;
2562 mutex_lock(&smc_ib_devices.mutex);
2563 list_for_each_entry(smcibdev, &smc_ib_devices.list, list) {
2566 for (i = 0; i < SMC_MAX_PORTS; i++)
2567 set_bit(i, smcibdev->ports_going_away);
2569 mutex_unlock(&smc_ib_devices.mutex);
2571 mutex_lock(&smcd_dev_list.mutex);
2572 list_for_each_entry(smcd, &smcd_dev_list.list, list) {
2573 smcd->going_away = 1;
2575 mutex_unlock(&smcd_dev_list.mutex);
2578 /* Clean up all SMC link groups */
2579 static void smc_lgrs_shutdown(void)
2581 struct smcd_dev *smcd;
2583 smc_core_going_away();
2585 smc_smcr_terminate_all(NULL);
2587 mutex_lock(&smcd_dev_list.mutex);
2588 list_for_each_entry(smcd, &smcd_dev_list.list, list)
2589 smc_smcd_terminate_all(smcd);
2590 mutex_unlock(&smcd_dev_list.mutex);
2593 static int smc_core_reboot_event(struct notifier_block *this,
2594 unsigned long event, void *ptr)
2596 smc_lgrs_shutdown();
2597 smc_ib_unregister_client();
2601 static struct notifier_block smc_reboot_notifier = {
2602 .notifier_call = smc_core_reboot_event,
2605 int __init smc_core_init(void)
2607 return register_reboot_notifier(&smc_reboot_notifier);
2610 /* Called (from smc_exit) when module is removed */
2611 void smc_core_exit(void)
2613 unregister_reboot_notifier(&smc_reboot_notifier);
2614 smc_lgrs_shutdown();