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 smcd_dev *smcd = lgr->smcd;
505 struct nlattr *attrs;
508 nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
509 &smc_gen_nl_family, NLM_F_MULTI,
510 SMC_NETLINK_GET_LGR_SMCD);
514 attrs = nla_nest_start(skb, SMC_GEN_LGR_SMCD);
518 if (nla_put_u32(skb, SMC_NLA_LGR_D_ID, *((u32 *)&lgr->id)))
520 if (nla_put_u64_64bit(skb, SMC_NLA_LGR_D_GID,
521 smcd->ops->get_local_gid(smcd),
524 if (nla_put_u64_64bit(skb, SMC_NLA_LGR_D_PEER_GID, lgr->peer_gid,
527 if (nla_put_u8(skb, SMC_NLA_LGR_D_VLAN_ID, lgr->vlan_id))
529 if (nla_put_u32(skb, SMC_NLA_LGR_D_CONNS_NUM, lgr->conns_num))
531 if (nla_put_u32(skb, SMC_NLA_LGR_D_CHID, smc_ism_get_chid(lgr->smcd)))
533 memcpy(smc_pnet, lgr->smcd->pnetid, SMC_MAX_PNETID_LEN);
534 smc_pnet[SMC_MAX_PNETID_LEN] = 0;
535 if (nla_put_string(skb, SMC_NLA_LGR_D_PNETID, smc_pnet))
537 if (lgr->smc_version > SMC_V1) {
538 struct nlattr *v2_attrs;
540 v2_attrs = nla_nest_start(skb, SMC_NLA_LGR_D_V2_COMMON);
543 if (smc_nl_fill_lgr_v2_common(lgr, skb, cb, v2_attrs))
546 nla_nest_end(skb, attrs);
547 genlmsg_end(skb, nlh);
551 nla_nest_cancel(skb, attrs);
553 genlmsg_cancel(skb, nlh);
558 static int smc_nl_handle_smcd_lgr(struct smcd_dev *dev,
560 struct netlink_callback *cb)
562 struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
563 struct smc_link_group *lgr;
564 int snum = cb_ctx->pos[1];
567 spin_lock_bh(&dev->lgr_lock);
568 list_for_each_entry(lgr, &dev->lgr_list, list) {
573 rc = smc_nl_fill_smcd_lgr(lgr, skb, cb);
580 spin_unlock_bh(&dev->lgr_lock);
581 cb_ctx->pos[1] = num;
585 static int smc_nl_fill_smcd_dev(struct smcd_dev_list *dev_list,
587 struct netlink_callback *cb)
589 struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
590 struct smcd_dev *smcd_dev;
591 int snum = cb_ctx->pos[0];
594 mutex_lock(&dev_list->mutex);
595 list_for_each_entry(smcd_dev, &dev_list->list, list) {
596 if (list_empty(&smcd_dev->lgr_list))
600 rc = smc_nl_handle_smcd_lgr(smcd_dev, skb, cb);
607 mutex_unlock(&dev_list->mutex);
608 cb_ctx->pos[0] = num;
612 int smcr_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb)
614 bool list_links = false;
616 smc_nl_fill_lgr_list(&smc_lgr_list, skb, cb, list_links);
620 int smcr_nl_get_link(struct sk_buff *skb, struct netlink_callback *cb)
622 bool list_links = true;
624 smc_nl_fill_lgr_list(&smc_lgr_list, skb, cb, list_links);
628 int smcd_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb)
630 smc_nl_fill_smcd_dev(&smcd_dev_list, skb, cb);
634 void smc_lgr_cleanup_early(struct smc_link_group *lgr)
636 spinlock_t *lgr_lock;
641 smc_lgr_list_head(lgr, &lgr_lock);
642 spin_lock_bh(lgr_lock);
643 /* do not use this link group for new connections */
644 if (!list_empty(&lgr->list))
645 list_del_init(&lgr->list);
646 spin_unlock_bh(lgr_lock);
647 __smc_lgr_terminate(lgr, true);
650 static void smcr_lgr_link_deactivate_all(struct smc_link_group *lgr)
654 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
655 struct smc_link *lnk = &lgr->lnk[i];
657 if (smc_link_sendable(lnk))
658 lnk->state = SMC_LNK_INACTIVE;
660 wake_up_all(&lgr->llc_msg_waiter);
661 wake_up_all(&lgr->llc_flow_waiter);
664 static void smc_lgr_free(struct smc_link_group *lgr);
666 static void smc_lgr_free_work(struct work_struct *work)
668 struct smc_link_group *lgr = container_of(to_delayed_work(work),
669 struct smc_link_group,
671 spinlock_t *lgr_lock;
674 smc_lgr_list_head(lgr, &lgr_lock);
675 spin_lock_bh(lgr_lock);
677 spin_unlock_bh(lgr_lock);
680 read_lock_bh(&lgr->conns_lock);
681 conns = RB_EMPTY_ROOT(&lgr->conns_all);
682 read_unlock_bh(&lgr->conns_lock);
683 if (!conns) { /* number of lgr connections is no longer zero */
684 spin_unlock_bh(lgr_lock);
687 list_del_init(&lgr->list); /* remove from smc_lgr_list */
688 lgr->freeing = 1; /* this instance does the freeing, no new schedule */
689 spin_unlock_bh(lgr_lock);
690 cancel_delayed_work(&lgr->free_work);
692 if (!lgr->is_smcd && !lgr->terminating)
693 smc_llc_send_link_delete_all(lgr, true,
694 SMC_LLC_DEL_PROG_INIT_TERM);
695 if (lgr->is_smcd && !lgr->terminating)
696 smc_ism_signal_shutdown(lgr);
698 smcr_lgr_link_deactivate_all(lgr);
702 static void smc_lgr_terminate_work(struct work_struct *work)
704 struct smc_link_group *lgr = container_of(work, struct smc_link_group,
707 __smc_lgr_terminate(lgr, true);
710 /* return next unique link id for the lgr */
711 static u8 smcr_next_link_id(struct smc_link_group *lgr)
718 link_id = ++lgr->next_link_id;
719 if (!link_id) /* skip zero as link_id */
720 link_id = ++lgr->next_link_id;
721 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
722 if (smc_link_usable(&lgr->lnk[i]) &&
723 lgr->lnk[i].link_id == link_id)
731 static void smcr_copy_dev_info_to_link(struct smc_link *link)
733 struct smc_ib_device *smcibdev = link->smcibdev;
735 snprintf(link->ibname, sizeof(link->ibname), "%s",
736 smcibdev->ibdev->name);
737 link->ndev_ifidx = smcibdev->ndev_ifidx[link->ibport - 1];
740 int smcr_link_init(struct smc_link_group *lgr, struct smc_link *lnk,
741 u8 link_idx, struct smc_init_info *ini)
743 struct smc_ib_device *smcibdev;
747 if (lgr->smc_version == SMC_V2) {
748 lnk->smcibdev = ini->smcrv2.ib_dev_v2;
749 lnk->ibport = ini->smcrv2.ib_port_v2;
751 lnk->smcibdev = ini->ib_dev;
752 lnk->ibport = ini->ib_port;
754 get_device(&lnk->smcibdev->ibdev->dev);
755 atomic_inc(&lnk->smcibdev->lnk_cnt);
756 refcount_set(&lnk->refcnt, 1); /* link refcnt is set to 1 */
758 lnk->path_mtu = lnk->smcibdev->pattr[lnk->ibport - 1].active_mtu;
759 lnk->link_id = smcr_next_link_id(lgr);
761 smc_lgr_hold(lgr); /* lgr_put in smcr_link_clear() */
762 lnk->link_idx = link_idx;
763 lnk->wr_rx_id_compl = 0;
764 smc_ibdev_cnt_inc(lnk);
765 smcr_copy_dev_info_to_link(lnk);
766 atomic_set(&lnk->conn_cnt, 0);
767 smc_llc_link_set_uid(lnk);
768 INIT_WORK(&lnk->link_down_wrk, smc_link_down_work);
769 if (!lnk->smcibdev->initialized) {
770 rc = (int)smc_ib_setup_per_ibdev(lnk->smcibdev);
774 get_random_bytes(rndvec, sizeof(rndvec));
775 lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) +
777 rc = smc_ib_determine_gid(lnk->smcibdev, lnk->ibport,
778 ini->vlan_id, lnk->gid, &lnk->sgid_index,
779 lgr->smc_version == SMC_V2 ?
780 &ini->smcrv2 : NULL);
783 rc = smc_llc_link_init(lnk);
786 rc = smc_wr_alloc_link_mem(lnk);
789 rc = smc_ib_create_protection_domain(lnk);
792 rc = smc_ib_create_queue_pair(lnk);
795 rc = smc_wr_create_link(lnk);
798 lnk->state = SMC_LNK_ACTIVATING;
802 smc_ib_destroy_queue_pair(lnk);
804 smc_ib_dealloc_protection_domain(lnk);
806 smc_wr_free_link_mem(lnk);
808 smc_llc_link_clear(lnk, false);
810 smc_ibdev_cnt_dec(lnk);
811 put_device(&lnk->smcibdev->ibdev->dev);
812 smcibdev = lnk->smcibdev;
813 memset(lnk, 0, sizeof(struct smc_link));
814 lnk->state = SMC_LNK_UNUSED;
815 if (!atomic_dec_return(&smcibdev->lnk_cnt))
816 wake_up(&smcibdev->lnks_deleted);
817 smc_lgr_put(lgr); /* lgr_hold above */
821 /* create a new SMC link group */
822 static int smc_lgr_create(struct smc_sock *smc, struct smc_init_info *ini)
824 struct smc_link_group *lgr;
825 struct list_head *lgr_list;
826 struct smcd_dev *smcd;
827 struct smc_link *lnk;
828 spinlock_t *lgr_lock;
833 if (ini->is_smcd && ini->vlan_id) {
834 if (smc_ism_get_vlan(ini->ism_dev[ini->ism_selected],
836 rc = SMC_CLC_DECL_ISMVLANERR;
841 lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
843 rc = SMC_CLC_DECL_MEM;
846 lgr->tx_wq = alloc_workqueue("smc_tx_wq-%*phN", 0, 0,
847 SMC_LGR_ID_SIZE, &lgr->id);
852 lgr->is_smcd = ini->is_smcd;
854 lgr->terminating = 0;
856 lgr->vlan_id = ini->vlan_id;
857 refcount_set(&lgr->refcnt, 1); /* set lgr refcnt to 1 */
858 init_rwsem(&lgr->sndbufs_lock);
859 init_rwsem(&lgr->rmbs_lock);
860 rwlock_init(&lgr->conns_lock);
861 for (i = 0; i < SMC_RMBE_SIZES; i++) {
862 INIT_LIST_HEAD(&lgr->sndbufs[i]);
863 INIT_LIST_HEAD(&lgr->rmbs[i]);
865 lgr->next_link_id = 0;
866 smc_lgr_list.num += SMC_LGR_NUM_INCR;
867 memcpy(&lgr->id, (u8 *)&smc_lgr_list.num, SMC_LGR_ID_SIZE);
868 INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
869 INIT_WORK(&lgr->terminate_work, smc_lgr_terminate_work);
870 lgr->conns_all = RB_ROOT;
872 /* SMC-D specific settings */
873 smcd = ini->ism_dev[ini->ism_selected];
874 get_device(smcd->ops->get_dev(smcd));
875 lgr->peer_gid = ini->ism_peer_gid[ini->ism_selected];
876 lgr->smcd = ini->ism_dev[ini->ism_selected];
877 lgr_list = &ini->ism_dev[ini->ism_selected]->lgr_list;
878 lgr_lock = &lgr->smcd->lgr_lock;
879 lgr->smc_version = ini->smcd_version;
880 lgr->peer_shutdown = 0;
881 atomic_inc(&ini->ism_dev[ini->ism_selected]->lgr_cnt);
883 /* SMC-R specific settings */
884 struct smc_ib_device *ibdev;
887 lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
888 lgr->smc_version = ini->smcr_version;
889 memcpy(lgr->peer_systemid, ini->peer_systemid,
891 if (lgr->smc_version == SMC_V2) {
892 ibdev = ini->smcrv2.ib_dev_v2;
893 ibport = ini->smcrv2.ib_port_v2;
894 lgr->saddr = ini->smcrv2.saddr;
895 lgr->uses_gateway = ini->smcrv2.uses_gateway;
896 memcpy(lgr->nexthop_mac, ini->smcrv2.nexthop_mac,
900 ibport = ini->ib_port;
902 memcpy(lgr->pnet_id, ibdev->pnetid[ibport - 1],
904 rc = smc_wr_alloc_lgr_mem(lgr);
907 smc_llc_lgr_init(lgr, smc);
909 link_idx = SMC_SINGLE_LINK;
910 lnk = &lgr->lnk[link_idx];
911 rc = smcr_link_init(lgr, lnk, link_idx, ini);
913 smc_wr_free_lgr_mem(lgr);
916 lgr->net = smc_ib_net(lnk->smcibdev);
917 lgr_list = &smc_lgr_list.list;
918 lgr_lock = &smc_lgr_list.lock;
919 lgr->buf_type = lgr->net->smc.sysctl_smcr_buf_type;
920 atomic_inc(&lgr_cnt);
923 spin_lock_bh(lgr_lock);
924 list_add_tail(&lgr->list, lgr_list);
925 spin_unlock_bh(lgr_lock);
929 destroy_workqueue(lgr->tx_wq);
933 if (ini->is_smcd && ini->vlan_id)
934 smc_ism_put_vlan(ini->ism_dev[ini->ism_selected], ini->vlan_id);
938 rc = SMC_CLC_DECL_MEM;
940 rc = SMC_CLC_DECL_INTERR;
945 static int smc_write_space(struct smc_connection *conn)
947 int buffer_len = conn->peer_rmbe_size;
948 union smc_host_cursor prod;
949 union smc_host_cursor cons;
952 smc_curs_copy(&prod, &conn->local_tx_ctrl.prod, conn);
953 smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
954 /* determine rx_buf space */
955 space = buffer_len - smc_curs_diff(buffer_len, &cons, &prod);
959 static int smc_switch_cursor(struct smc_sock *smc, struct smc_cdc_tx_pend *pend,
960 struct smc_wr_buf *wr_buf)
962 struct smc_connection *conn = &smc->conn;
963 union smc_host_cursor cons, fin;
967 smc_curs_copy(&conn->tx_curs_sent, &conn->tx_curs_fin, conn);
968 smc_curs_copy(&fin, &conn->local_tx_ctrl_fin, conn);
969 /* set prod cursor to old state, enforce tx_rdma_writes() */
970 smc_curs_copy(&conn->local_tx_ctrl.prod, &fin, conn);
971 smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
973 if (smc_curs_comp(conn->peer_rmbe_size, &cons, &fin) < 0) {
974 /* cons cursor advanced more than fin, and prod was set
975 * fin above, so now prod is smaller than cons. Fix that.
977 diff = smc_curs_diff(conn->peer_rmbe_size, &fin, &cons);
978 smc_curs_add(conn->sndbuf_desc->len,
979 &conn->tx_curs_sent, diff);
980 smc_curs_add(conn->sndbuf_desc->len,
981 &conn->tx_curs_fin, diff);
983 smp_mb__before_atomic();
984 atomic_add(diff, &conn->sndbuf_space);
985 smp_mb__after_atomic();
987 smc_curs_add(conn->peer_rmbe_size,
988 &conn->local_tx_ctrl.prod, diff);
989 smc_curs_add(conn->peer_rmbe_size,
990 &conn->local_tx_ctrl_fin, diff);
992 /* recalculate, value is used by tx_rdma_writes() */
993 atomic_set(&smc->conn.peer_rmbe_space, smc_write_space(conn));
995 if (smc->sk.sk_state != SMC_INIT &&
996 smc->sk.sk_state != SMC_CLOSED) {
997 rc = smcr_cdc_msg_send_validation(conn, pend, wr_buf);
999 queue_delayed_work(conn->lgr->tx_wq, &conn->tx_work, 0);
1000 smc->sk.sk_data_ready(&smc->sk);
1003 smc_wr_tx_put_slot(conn->lnk,
1004 (struct smc_wr_tx_pend_priv *)pend);
1009 void smc_switch_link_and_count(struct smc_connection *conn,
1010 struct smc_link *to_lnk)
1012 atomic_dec(&conn->lnk->conn_cnt);
1013 /* link_hold in smc_conn_create() */
1014 smcr_link_put(conn->lnk);
1016 atomic_inc(&conn->lnk->conn_cnt);
1017 /* link_put in smc_conn_free() */
1018 smcr_link_hold(conn->lnk);
1021 struct smc_link *smc_switch_conns(struct smc_link_group *lgr,
1022 struct smc_link *from_lnk, bool is_dev_err)
1024 struct smc_link *to_lnk = NULL;
1025 struct smc_cdc_tx_pend *pend;
1026 struct smc_connection *conn;
1027 struct smc_wr_buf *wr_buf;
1028 struct smc_sock *smc;
1029 struct rb_node *node;
1032 /* link is inactive, wake up tx waiters */
1033 smc_wr_wakeup_tx_wait(from_lnk);
1035 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1036 if (!smc_link_active(&lgr->lnk[i]) || i == from_lnk->link_idx)
1038 if (is_dev_err && from_lnk->smcibdev == lgr->lnk[i].smcibdev &&
1039 from_lnk->ibport == lgr->lnk[i].ibport) {
1042 to_lnk = &lgr->lnk[i];
1045 if (!to_lnk || !smc_wr_tx_link_hold(to_lnk)) {
1046 smc_lgr_terminate_sched(lgr);
1050 read_lock_bh(&lgr->conns_lock);
1051 for (node = rb_first(&lgr->conns_all); node; node = rb_next(node)) {
1052 conn = rb_entry(node, struct smc_connection, alert_node);
1053 if (conn->lnk != from_lnk)
1055 smc = container_of(conn, struct smc_sock, conn);
1056 /* conn->lnk not yet set in SMC_INIT state */
1057 if (smc->sk.sk_state == SMC_INIT)
1059 if (smc->sk.sk_state == SMC_CLOSED ||
1060 smc->sk.sk_state == SMC_PEERCLOSEWAIT1 ||
1061 smc->sk.sk_state == SMC_PEERCLOSEWAIT2 ||
1062 smc->sk.sk_state == SMC_APPFINCLOSEWAIT ||
1063 smc->sk.sk_state == SMC_APPCLOSEWAIT1 ||
1064 smc->sk.sk_state == SMC_APPCLOSEWAIT2 ||
1065 smc->sk.sk_state == SMC_PEERFINCLOSEWAIT ||
1066 smc->sk.sk_state == SMC_PEERABORTWAIT ||
1067 smc->sk.sk_state == SMC_PROCESSABORT) {
1068 spin_lock_bh(&conn->send_lock);
1069 smc_switch_link_and_count(conn, to_lnk);
1070 spin_unlock_bh(&conn->send_lock);
1073 sock_hold(&smc->sk);
1074 read_unlock_bh(&lgr->conns_lock);
1075 /* pre-fetch buffer outside of send_lock, might sleep */
1076 rc = smc_cdc_get_free_slot(conn, to_lnk, &wr_buf, NULL, &pend);
1079 /* avoid race with smcr_tx_sndbuf_nonempty() */
1080 spin_lock_bh(&conn->send_lock);
1081 smc_switch_link_and_count(conn, to_lnk);
1082 rc = smc_switch_cursor(smc, pend, wr_buf);
1083 spin_unlock_bh(&conn->send_lock);
1089 read_unlock_bh(&lgr->conns_lock);
1090 smc_wr_tx_link_put(to_lnk);
1094 smcr_link_down_cond_sched(to_lnk);
1095 smc_wr_tx_link_put(to_lnk);
1099 static void smcr_buf_unuse(struct smc_buf_desc *buf_desc, bool is_rmb,
1100 struct smc_link_group *lgr)
1102 struct rw_semaphore *lock; /* lock buffer list */
1105 if (is_rmb && buf_desc->is_conf_rkey && !list_empty(&lgr->list)) {
1106 /* unregister rmb with peer */
1107 rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY);
1109 /* protect against smc_llc_cli_rkey_exchange() */
1110 down_read(&lgr->llc_conf_mutex);
1111 smc_llc_do_delete_rkey(lgr, buf_desc);
1112 buf_desc->is_conf_rkey = false;
1113 up_read(&lgr->llc_conf_mutex);
1114 smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl);
1118 if (buf_desc->is_reg_err) {
1119 /* buf registration failed, reuse not possible */
1120 lock = is_rmb ? &lgr->rmbs_lock :
1123 list_del(&buf_desc->list);
1126 smc_buf_free(lgr, is_rmb, buf_desc);
1128 /* memzero_explicit provides potential memory barrier semantics */
1129 memzero_explicit(buf_desc->cpu_addr, buf_desc->len);
1130 WRITE_ONCE(buf_desc->used, 0);
1134 static void smc_buf_unuse(struct smc_connection *conn,
1135 struct smc_link_group *lgr)
1137 if (conn->sndbuf_desc) {
1138 if (!lgr->is_smcd && conn->sndbuf_desc->is_vm) {
1139 smcr_buf_unuse(conn->sndbuf_desc, false, lgr);
1141 memzero_explicit(conn->sndbuf_desc->cpu_addr, conn->sndbuf_desc->len);
1142 WRITE_ONCE(conn->sndbuf_desc->used, 0);
1145 if (conn->rmb_desc) {
1146 if (!lgr->is_smcd) {
1147 smcr_buf_unuse(conn->rmb_desc, true, lgr);
1149 memzero_explicit(conn->rmb_desc->cpu_addr,
1150 conn->rmb_desc->len + sizeof(struct smcd_cdc_msg));
1151 WRITE_ONCE(conn->rmb_desc->used, 0);
1156 /* remove a finished connection from its link group */
1157 void smc_conn_free(struct smc_connection *conn)
1159 struct smc_link_group *lgr = conn->lgr;
1161 if (!lgr || conn->freed)
1162 /* Connection has never been registered in a
1163 * link group, or has already been freed.
1168 if (!smc_conn_lgr_valid(conn))
1169 /* Connection has already unregistered from
1175 if (!list_empty(&lgr->list))
1176 smc_ism_unset_conn(conn);
1177 tasklet_kill(&conn->rx_tsklet);
1179 smc_cdc_wait_pend_tx_wr(conn);
1180 if (current_work() != &conn->abort_work)
1181 cancel_work_sync(&conn->abort_work);
1183 if (!list_empty(&lgr->list)) {
1184 smc_buf_unuse(conn, lgr); /* allow buffer reuse */
1185 smc_lgr_unregister_conn(conn);
1188 if (!lgr->conns_num)
1189 smc_lgr_schedule_free_work(lgr);
1192 smcr_link_put(conn->lnk); /* link_hold in smc_conn_create() */
1193 smc_lgr_put(lgr); /* lgr_hold in smc_conn_create() */
1196 /* unregister a link from a buf_desc */
1197 static void smcr_buf_unmap_link(struct smc_buf_desc *buf_desc, bool is_rmb,
1198 struct smc_link *lnk)
1200 if (is_rmb || buf_desc->is_vm)
1201 buf_desc->is_reg_mr[lnk->link_idx] = false;
1202 if (!buf_desc->is_map_ib[lnk->link_idx])
1205 if ((is_rmb || buf_desc->is_vm) &&
1206 buf_desc->mr[lnk->link_idx]) {
1207 smc_ib_put_memory_region(buf_desc->mr[lnk->link_idx]);
1208 buf_desc->mr[lnk->link_idx] = NULL;
1211 smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_FROM_DEVICE);
1213 smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_TO_DEVICE);
1215 sg_free_table(&buf_desc->sgt[lnk->link_idx]);
1216 buf_desc->is_map_ib[lnk->link_idx] = false;
1219 /* unmap all buffers of lgr for a deleted link */
1220 static void smcr_buf_unmap_lgr(struct smc_link *lnk)
1222 struct smc_link_group *lgr = lnk->lgr;
1223 struct smc_buf_desc *buf_desc, *bf;
1226 for (i = 0; i < SMC_RMBE_SIZES; i++) {
1227 down_write(&lgr->rmbs_lock);
1228 list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list)
1229 smcr_buf_unmap_link(buf_desc, true, lnk);
1230 up_write(&lgr->rmbs_lock);
1232 down_write(&lgr->sndbufs_lock);
1233 list_for_each_entry_safe(buf_desc, bf, &lgr->sndbufs[i],
1235 smcr_buf_unmap_link(buf_desc, false, lnk);
1236 up_write(&lgr->sndbufs_lock);
1240 static void smcr_rtoken_clear_link(struct smc_link *lnk)
1242 struct smc_link_group *lgr = lnk->lgr;
1245 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1246 lgr->rtokens[i][lnk->link_idx].rkey = 0;
1247 lgr->rtokens[i][lnk->link_idx].dma_addr = 0;
1251 static void __smcr_link_clear(struct smc_link *lnk)
1253 struct smc_link_group *lgr = lnk->lgr;
1254 struct smc_ib_device *smcibdev;
1256 smc_wr_free_link_mem(lnk);
1257 smc_ibdev_cnt_dec(lnk);
1258 put_device(&lnk->smcibdev->ibdev->dev);
1259 smcibdev = lnk->smcibdev;
1260 memset(lnk, 0, sizeof(struct smc_link));
1261 lnk->state = SMC_LNK_UNUSED;
1262 if (!atomic_dec_return(&smcibdev->lnk_cnt))
1263 wake_up(&smcibdev->lnks_deleted);
1264 smc_lgr_put(lgr); /* lgr_hold in smcr_link_init() */
1267 /* must be called under lgr->llc_conf_mutex lock */
1268 void smcr_link_clear(struct smc_link *lnk, bool log)
1270 if (!lnk->lgr || lnk->clearing ||
1271 lnk->state == SMC_LNK_UNUSED)
1275 smc_llc_link_clear(lnk, log);
1276 smcr_buf_unmap_lgr(lnk);
1277 smcr_rtoken_clear_link(lnk);
1278 smc_ib_modify_qp_error(lnk);
1279 smc_wr_free_link(lnk);
1280 smc_ib_destroy_queue_pair(lnk);
1281 smc_ib_dealloc_protection_domain(lnk);
1282 smcr_link_put(lnk); /* theoretically last link_put */
1285 void smcr_link_hold(struct smc_link *lnk)
1287 refcount_inc(&lnk->refcnt);
1290 void smcr_link_put(struct smc_link *lnk)
1292 if (refcount_dec_and_test(&lnk->refcnt))
1293 __smcr_link_clear(lnk);
1296 static void smcr_buf_free(struct smc_link_group *lgr, bool is_rmb,
1297 struct smc_buf_desc *buf_desc)
1301 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
1302 smcr_buf_unmap_link(buf_desc, is_rmb, &lgr->lnk[i]);
1304 if (!buf_desc->is_vm && buf_desc->pages)
1305 __free_pages(buf_desc->pages, buf_desc->order);
1306 else if (buf_desc->is_vm && buf_desc->cpu_addr)
1307 vfree(buf_desc->cpu_addr);
1311 static void smcd_buf_free(struct smc_link_group *lgr, bool is_dmb,
1312 struct smc_buf_desc *buf_desc)
1315 /* restore original buf len */
1316 buf_desc->len += sizeof(struct smcd_cdc_msg);
1317 smc_ism_unregister_dmb(lgr->smcd, buf_desc);
1319 kfree(buf_desc->cpu_addr);
1324 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
1325 struct smc_buf_desc *buf_desc)
1328 smcd_buf_free(lgr, is_rmb, buf_desc);
1330 smcr_buf_free(lgr, is_rmb, buf_desc);
1333 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
1335 struct smc_buf_desc *buf_desc, *bf_desc;
1336 struct list_head *buf_list;
1339 for (i = 0; i < SMC_RMBE_SIZES; i++) {
1341 buf_list = &lgr->rmbs[i];
1343 buf_list = &lgr->sndbufs[i];
1344 list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
1346 list_del(&buf_desc->list);
1347 smc_buf_free(lgr, is_rmb, buf_desc);
1352 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
1354 /* free send buffers */
1355 __smc_lgr_free_bufs(lgr, false);
1357 __smc_lgr_free_bufs(lgr, true);
1360 /* won't be freed until no one accesses to lgr anymore */
1361 static void __smc_lgr_free(struct smc_link_group *lgr)
1363 smc_lgr_free_bufs(lgr);
1365 if (!atomic_dec_return(&lgr->smcd->lgr_cnt))
1366 wake_up(&lgr->smcd->lgrs_deleted);
1368 smc_wr_free_lgr_mem(lgr);
1369 if (!atomic_dec_return(&lgr_cnt))
1370 wake_up(&lgrs_deleted);
1375 /* remove a link group */
1376 static void smc_lgr_free(struct smc_link_group *lgr)
1380 if (!lgr->is_smcd) {
1381 down_write(&lgr->llc_conf_mutex);
1382 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1383 if (lgr->lnk[i].state != SMC_LNK_UNUSED)
1384 smcr_link_clear(&lgr->lnk[i], false);
1386 up_write(&lgr->llc_conf_mutex);
1387 smc_llc_lgr_clear(lgr);
1390 destroy_workqueue(lgr->tx_wq);
1392 smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
1393 put_device(lgr->smcd->ops->get_dev(lgr->smcd));
1395 smc_lgr_put(lgr); /* theoretically last lgr_put */
1398 void smc_lgr_hold(struct smc_link_group *lgr)
1400 refcount_inc(&lgr->refcnt);
1403 void smc_lgr_put(struct smc_link_group *lgr)
1405 if (refcount_dec_and_test(&lgr->refcnt))
1406 __smc_lgr_free(lgr);
1409 static void smc_sk_wake_ups(struct smc_sock *smc)
1411 smc->sk.sk_write_space(&smc->sk);
1412 smc->sk.sk_data_ready(&smc->sk);
1413 smc->sk.sk_state_change(&smc->sk);
1416 /* kill a connection */
1417 static void smc_conn_kill(struct smc_connection *conn, bool soft)
1419 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1421 if (conn->lgr->is_smcd && conn->lgr->peer_shutdown)
1422 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
1424 smc_close_abort(conn);
1426 smc->sk.sk_err = ECONNABORTED;
1427 smc_sk_wake_ups(smc);
1428 if (conn->lgr->is_smcd) {
1429 smc_ism_unset_conn(conn);
1431 tasklet_kill(&conn->rx_tsklet);
1433 tasklet_unlock_wait(&conn->rx_tsklet);
1435 smc_cdc_wait_pend_tx_wr(conn);
1437 smc_lgr_unregister_conn(conn);
1438 smc_close_active_abort(smc);
1441 static void smc_lgr_cleanup(struct smc_link_group *lgr)
1444 smc_ism_signal_shutdown(lgr);
1446 u32 rsn = lgr->llc_termination_rsn;
1449 rsn = SMC_LLC_DEL_PROG_INIT_TERM;
1450 smc_llc_send_link_delete_all(lgr, false, rsn);
1451 smcr_lgr_link_deactivate_all(lgr);
1455 /* terminate link group
1456 * @soft: true if link group shutdown can take its time
1457 * false if immediate link group shutdown is required
1459 static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft)
1461 struct smc_connection *conn;
1462 struct smc_sock *smc;
1463 struct rb_node *node;
1465 if (lgr->terminating)
1466 return; /* lgr already terminating */
1467 /* cancel free_work sync, will terminate when lgr->freeing is set */
1468 cancel_delayed_work(&lgr->free_work);
1469 lgr->terminating = 1;
1471 /* kill remaining link group connections */
1472 read_lock_bh(&lgr->conns_lock);
1473 node = rb_first(&lgr->conns_all);
1475 read_unlock_bh(&lgr->conns_lock);
1476 conn = rb_entry(node, struct smc_connection, alert_node);
1477 smc = container_of(conn, struct smc_sock, conn);
1478 sock_hold(&smc->sk); /* sock_put below */
1479 lock_sock(&smc->sk);
1480 smc_conn_kill(conn, soft);
1481 release_sock(&smc->sk);
1482 sock_put(&smc->sk); /* sock_hold above */
1483 read_lock_bh(&lgr->conns_lock);
1484 node = rb_first(&lgr->conns_all);
1486 read_unlock_bh(&lgr->conns_lock);
1487 smc_lgr_cleanup(lgr);
1491 /* unlink link group and schedule termination */
1492 void smc_lgr_terminate_sched(struct smc_link_group *lgr)
1494 spinlock_t *lgr_lock;
1496 smc_lgr_list_head(lgr, &lgr_lock);
1497 spin_lock_bh(lgr_lock);
1498 if (list_empty(&lgr->list) || lgr->terminating || lgr->freeing) {
1499 spin_unlock_bh(lgr_lock);
1500 return; /* lgr already terminating */
1502 list_del_init(&lgr->list);
1504 spin_unlock_bh(lgr_lock);
1505 schedule_work(&lgr->terminate_work);
1508 /* Called when peer lgr shutdown (regularly or abnormally) is received */
1509 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid, unsigned short vlan)
1511 struct smc_link_group *lgr, *l;
1512 LIST_HEAD(lgr_free_list);
1514 /* run common cleanup function and build free list */
1515 spin_lock_bh(&dev->lgr_lock);
1516 list_for_each_entry_safe(lgr, l, &dev->lgr_list, list) {
1517 if ((!peer_gid || lgr->peer_gid == peer_gid) &&
1518 (vlan == VLAN_VID_MASK || lgr->vlan_id == vlan)) {
1519 if (peer_gid) /* peer triggered termination */
1520 lgr->peer_shutdown = 1;
1521 list_move(&lgr->list, &lgr_free_list);
1525 spin_unlock_bh(&dev->lgr_lock);
1527 /* cancel the regular free workers and actually free lgrs */
1528 list_for_each_entry_safe(lgr, l, &lgr_free_list, list) {
1529 list_del_init(&lgr->list);
1530 schedule_work(&lgr->terminate_work);
1534 /* Called when an SMCD device is removed or the smc module is unloaded */
1535 void smc_smcd_terminate_all(struct smcd_dev *smcd)
1537 struct smc_link_group *lgr, *lg;
1538 LIST_HEAD(lgr_free_list);
1540 spin_lock_bh(&smcd->lgr_lock);
1541 list_splice_init(&smcd->lgr_list, &lgr_free_list);
1542 list_for_each_entry(lgr, &lgr_free_list, list)
1544 spin_unlock_bh(&smcd->lgr_lock);
1546 list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
1547 list_del_init(&lgr->list);
1548 __smc_lgr_terminate(lgr, false);
1551 if (atomic_read(&smcd->lgr_cnt))
1552 wait_event(smcd->lgrs_deleted, !atomic_read(&smcd->lgr_cnt));
1555 /* Called when an SMCR device is removed or the smc module is unloaded.
1556 * If smcibdev is given, all SMCR link groups using this device are terminated.
1557 * If smcibdev is NULL, all SMCR link groups are terminated.
1559 void smc_smcr_terminate_all(struct smc_ib_device *smcibdev)
1561 struct smc_link_group *lgr, *lg;
1562 LIST_HEAD(lgr_free_list);
1565 spin_lock_bh(&smc_lgr_list.lock);
1567 list_splice_init(&smc_lgr_list.list, &lgr_free_list);
1568 list_for_each_entry(lgr, &lgr_free_list, list)
1571 list_for_each_entry_safe(lgr, lg, &smc_lgr_list.list, list) {
1572 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1573 if (lgr->lnk[i].smcibdev == smcibdev)
1574 smcr_link_down_cond_sched(&lgr->lnk[i]);
1578 spin_unlock_bh(&smc_lgr_list.lock);
1580 list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
1581 list_del_init(&lgr->list);
1582 smc_llc_set_termination_rsn(lgr, SMC_LLC_DEL_OP_INIT_TERM);
1583 __smc_lgr_terminate(lgr, false);
1587 if (atomic_read(&smcibdev->lnk_cnt))
1588 wait_event(smcibdev->lnks_deleted,
1589 !atomic_read(&smcibdev->lnk_cnt));
1591 if (atomic_read(&lgr_cnt))
1592 wait_event(lgrs_deleted, !atomic_read(&lgr_cnt));
1596 /* set new lgr type and clear all asymmetric link tagging */
1597 void smcr_lgr_set_type(struct smc_link_group *lgr, enum smc_lgr_type new_type)
1599 char *lgr_type = "";
1602 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
1603 if (smc_link_usable(&lgr->lnk[i]))
1604 lgr->lnk[i].link_is_asym = false;
1605 if (lgr->type == new_type)
1607 lgr->type = new_type;
1609 switch (lgr->type) {
1613 case SMC_LGR_SINGLE:
1614 lgr_type = "SINGLE";
1616 case SMC_LGR_SYMMETRIC:
1617 lgr_type = "SYMMETRIC";
1619 case SMC_LGR_ASYMMETRIC_PEER:
1620 lgr_type = "ASYMMETRIC_PEER";
1622 case SMC_LGR_ASYMMETRIC_LOCAL:
1623 lgr_type = "ASYMMETRIC_LOCAL";
1626 pr_warn_ratelimited("smc: SMC-R lg %*phN net %llu state changed: "
1627 "%s, pnetid %.16s\n", SMC_LGR_ID_SIZE, &lgr->id,
1628 lgr->net->net_cookie, lgr_type, lgr->pnet_id);
1631 /* set new lgr type and tag a link as asymmetric */
1632 void smcr_lgr_set_type_asym(struct smc_link_group *lgr,
1633 enum smc_lgr_type new_type, int asym_lnk_idx)
1635 smcr_lgr_set_type(lgr, new_type);
1636 lgr->lnk[asym_lnk_idx].link_is_asym = true;
1639 /* abort connection, abort_work scheduled from tasklet context */
1640 static void smc_conn_abort_work(struct work_struct *work)
1642 struct smc_connection *conn = container_of(work,
1643 struct smc_connection,
1645 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1647 lock_sock(&smc->sk);
1648 smc_conn_kill(conn, true);
1649 release_sock(&smc->sk);
1650 sock_put(&smc->sk); /* sock_hold done by schedulers of abort_work */
1653 void smcr_port_add(struct smc_ib_device *smcibdev, u8 ibport)
1655 struct smc_link_group *lgr, *n;
1657 spin_lock_bh(&smc_lgr_list.lock);
1658 list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1659 struct smc_link *link;
1661 if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1662 SMC_MAX_PNETID_LEN) ||
1663 lgr->type == SMC_LGR_SYMMETRIC ||
1664 lgr->type == SMC_LGR_ASYMMETRIC_PEER ||
1665 !rdma_dev_access_netns(smcibdev->ibdev, lgr->net))
1668 /* trigger local add link processing */
1669 link = smc_llc_usable_link(lgr);
1671 smc_llc_add_link_local(link);
1673 spin_unlock_bh(&smc_lgr_list.lock);
1676 /* link is down - switch connections to alternate link,
1677 * must be called under lgr->llc_conf_mutex lock
1679 static void smcr_link_down(struct smc_link *lnk)
1681 struct smc_link_group *lgr = lnk->lgr;
1682 struct smc_link *to_lnk;
1685 if (!lgr || lnk->state == SMC_LNK_UNUSED || list_empty(&lgr->list))
1688 to_lnk = smc_switch_conns(lgr, lnk, true);
1689 if (!to_lnk) { /* no backup link available */
1690 smcr_link_clear(lnk, true);
1693 smcr_lgr_set_type(lgr, SMC_LGR_SINGLE);
1694 del_link_id = lnk->link_id;
1696 if (lgr->role == SMC_SERV) {
1697 /* trigger local delete link processing */
1698 smc_llc_srv_delete_link_local(to_lnk, del_link_id);
1700 if (lgr->llc_flow_lcl.type != SMC_LLC_FLOW_NONE) {
1701 /* another llc task is ongoing */
1702 up_write(&lgr->llc_conf_mutex);
1703 wait_event_timeout(lgr->llc_flow_waiter,
1704 (list_empty(&lgr->list) ||
1705 lgr->llc_flow_lcl.type == SMC_LLC_FLOW_NONE),
1707 down_write(&lgr->llc_conf_mutex);
1709 if (!list_empty(&lgr->list)) {
1710 smc_llc_send_delete_link(to_lnk, del_link_id,
1712 SMC_LLC_DEL_LOST_PATH);
1713 smcr_link_clear(lnk, true);
1715 wake_up(&lgr->llc_flow_waiter); /* wake up next waiter */
1719 /* must be called under lgr->llc_conf_mutex lock */
1720 void smcr_link_down_cond(struct smc_link *lnk)
1722 if (smc_link_downing(&lnk->state)) {
1723 trace_smcr_link_down(lnk, __builtin_return_address(0));
1724 smcr_link_down(lnk);
1728 /* will get the lgr->llc_conf_mutex lock */
1729 void smcr_link_down_cond_sched(struct smc_link *lnk)
1731 if (smc_link_downing(&lnk->state)) {
1732 trace_smcr_link_down(lnk, __builtin_return_address(0));
1733 schedule_work(&lnk->link_down_wrk);
1737 void smcr_port_err(struct smc_ib_device *smcibdev, u8 ibport)
1739 struct smc_link_group *lgr, *n;
1742 list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1743 if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1744 SMC_MAX_PNETID_LEN))
1745 continue; /* lgr is not affected */
1746 if (list_empty(&lgr->list))
1748 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1749 struct smc_link *lnk = &lgr->lnk[i];
1751 if (smc_link_usable(lnk) &&
1752 lnk->smcibdev == smcibdev && lnk->ibport == ibport)
1753 smcr_link_down_cond_sched(lnk);
1758 static void smc_link_down_work(struct work_struct *work)
1760 struct smc_link *link = container_of(work, struct smc_link,
1762 struct smc_link_group *lgr = link->lgr;
1764 if (list_empty(&lgr->list))
1766 wake_up_all(&lgr->llc_msg_waiter);
1767 down_write(&lgr->llc_conf_mutex);
1768 smcr_link_down(link);
1769 up_write(&lgr->llc_conf_mutex);
1772 static int smc_vlan_by_tcpsk_walk(struct net_device *lower_dev,
1773 struct netdev_nested_priv *priv)
1775 unsigned short *vlan_id = (unsigned short *)priv->data;
1777 if (is_vlan_dev(lower_dev)) {
1778 *vlan_id = vlan_dev_vlan_id(lower_dev);
1785 /* Determine vlan of internal TCP socket. */
1786 int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini)
1788 struct dst_entry *dst = sk_dst_get(clcsock->sk);
1789 struct netdev_nested_priv priv;
1790 struct net_device *ndev;
1804 if (is_vlan_dev(ndev)) {
1805 ini->vlan_id = vlan_dev_vlan_id(ndev);
1809 priv.data = (void *)&ini->vlan_id;
1811 netdev_walk_all_lower_dev(ndev, smc_vlan_by_tcpsk_walk, &priv);
1820 static bool smcr_lgr_match(struct smc_link_group *lgr, u8 smcr_version,
1824 enum smc_lgr_role role, u32 clcqpn,
1827 struct smc_link *lnk;
1830 if (memcmp(lgr->peer_systemid, peer_systemid, SMC_SYSTEMID_LEN) ||
1834 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1837 if (!smc_link_active(lnk))
1839 /* use verbs API to check netns, instead of lgr->net */
1840 if (!rdma_dev_access_netns(lnk->smcibdev->ibdev, net))
1842 if ((lgr->role == SMC_SERV || lnk->peer_qpn == clcqpn) &&
1843 !memcmp(lnk->peer_gid, peer_gid, SMC_GID_SIZE) &&
1844 (smcr_version == SMC_V2 ||
1845 !memcmp(lnk->peer_mac, peer_mac_v1, ETH_ALEN)))
1851 static bool smcd_lgr_match(struct smc_link_group *lgr,
1852 struct smcd_dev *smcismdev, u64 peer_gid)
1854 return lgr->peer_gid == peer_gid && lgr->smcd == smcismdev;
1857 /* create a new SMC connection (and a new link group if necessary) */
1858 int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini)
1860 struct smc_connection *conn = &smc->conn;
1861 struct net *net = sock_net(&smc->sk);
1862 struct list_head *lgr_list;
1863 struct smc_link_group *lgr;
1864 enum smc_lgr_role role;
1865 spinlock_t *lgr_lock;
1868 lgr_list = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_list :
1870 lgr_lock = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_lock :
1872 ini->first_contact_local = 1;
1873 role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
1874 if (role == SMC_CLNT && ini->first_contact_peer)
1875 /* create new link group as well */
1878 /* determine if an existing link group can be reused */
1879 spin_lock_bh(lgr_lock);
1880 list_for_each_entry(lgr, lgr_list, list) {
1881 write_lock_bh(&lgr->conns_lock);
1883 smcd_lgr_match(lgr, ini->ism_dev[ini->ism_selected],
1884 ini->ism_peer_gid[ini->ism_selected]) :
1885 smcr_lgr_match(lgr, ini->smcr_version,
1887 ini->peer_gid, ini->peer_mac, role,
1888 ini->ib_clcqpn, net)) &&
1890 (ini->smcd_version == SMC_V2 ||
1891 lgr->vlan_id == ini->vlan_id) &&
1892 (role == SMC_CLNT || ini->is_smcd ||
1893 (lgr->conns_num < SMC_RMBS_PER_LGR_MAX &&
1894 !bitmap_full(lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX)))) {
1895 /* link group found */
1896 ini->first_contact_local = 0;
1898 rc = smc_lgr_register_conn(conn, false);
1899 write_unlock_bh(&lgr->conns_lock);
1900 if (!rc && delayed_work_pending(&lgr->free_work))
1901 cancel_delayed_work(&lgr->free_work);
1904 write_unlock_bh(&lgr->conns_lock);
1906 spin_unlock_bh(lgr_lock);
1910 if (role == SMC_CLNT && !ini->first_contact_peer &&
1911 ini->first_contact_local) {
1912 /* Server reuses a link group, but Client wants to start
1914 * send out_of_sync decline, reason synchr. error
1916 return SMC_CLC_DECL_SYNCERR;
1920 if (ini->first_contact_local) {
1921 rc = smc_lgr_create(smc, ini);
1925 write_lock_bh(&lgr->conns_lock);
1926 rc = smc_lgr_register_conn(conn, true);
1927 write_unlock_bh(&lgr->conns_lock);
1929 smc_lgr_cleanup_early(lgr);
1933 smc_lgr_hold(conn->lgr); /* lgr_put in smc_conn_free() */
1934 if (!conn->lgr->is_smcd)
1935 smcr_link_hold(conn->lnk); /* link_put in smc_conn_free() */
1937 conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
1938 conn->local_tx_ctrl.len = SMC_WR_TX_SIZE;
1939 conn->urg_state = SMC_URG_READ;
1940 init_waitqueue_head(&conn->cdc_pend_tx_wq);
1941 INIT_WORK(&smc->conn.abort_work, smc_conn_abort_work);
1943 conn->rx_off = sizeof(struct smcd_cdc_msg);
1944 smcd_cdc_rx_init(conn); /* init tasklet for this conn */
1948 #ifndef KERNEL_HAS_ATOMIC64
1949 spin_lock_init(&conn->acurs_lock);
1956 #define SMCD_DMBE_SIZES 6 /* 0 -> 16KB, 1 -> 32KB, .. 6 -> 1MB */
1957 #define SMCR_RMBE_SIZES 5 /* 0 -> 16KB, 1 -> 32KB, .. 5 -> 512KB */
1959 /* convert the RMB size into the compressed notation (minimum 16K, see
1960 * SMCD/R_DMBE_SIZES.
1961 * In contrast to plain ilog2, this rounds towards the next power of 2,
1962 * so the socket application gets at least its desired sndbuf / rcvbuf size.
1964 static u8 smc_compress_bufsize(int size, bool is_smcd, bool is_rmb)
1966 const unsigned int max_scat = SG_MAX_SINGLE_ALLOC * PAGE_SIZE;
1969 if (size <= SMC_BUF_MIN_SIZE)
1972 size = (size - 1) >> 14; /* convert to 16K multiple */
1973 compressed = min_t(u8, ilog2(size) + 1,
1974 is_smcd ? SMCD_DMBE_SIZES : SMCR_RMBE_SIZES);
1976 if (!is_smcd && is_rmb)
1977 /* RMBs are backed by & limited to max size of scatterlists */
1978 compressed = min_t(u8, compressed, ilog2(max_scat >> 14));
1983 /* convert the RMB size from compressed notation into integer */
1984 int smc_uncompress_bufsize(u8 compressed)
1988 size = 0x00000001 << (((int)compressed) + 14);
1992 /* try to reuse a sndbuf or rmb description slot for a certain
1993 * buffer size; if not available, return NULL
1995 static struct smc_buf_desc *smc_buf_get_slot(int compressed_bufsize,
1996 struct rw_semaphore *lock,
1997 struct list_head *buf_list)
1999 struct smc_buf_desc *buf_slot;
2002 list_for_each_entry(buf_slot, buf_list, list) {
2003 if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
2012 /* one of the conditions for announcing a receiver's current window size is
2013 * that it "results in a minimum increase in the window size of 10% of the
2014 * receive buffer space" [RFC7609]
2016 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
2018 return max_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
2021 /* map an buf to a link */
2022 static int smcr_buf_map_link(struct smc_buf_desc *buf_desc, bool is_rmb,
2023 struct smc_link *lnk)
2025 int rc, i, nents, offset, buf_size, size, access_flags;
2026 struct scatterlist *sg;
2029 if (buf_desc->is_map_ib[lnk->link_idx])
2032 if (buf_desc->is_vm) {
2033 buf = buf_desc->cpu_addr;
2034 buf_size = buf_desc->len;
2035 offset = offset_in_page(buf_desc->cpu_addr);
2036 nents = PAGE_ALIGN(buf_size + offset) / PAGE_SIZE;
2041 rc = sg_alloc_table(&buf_desc->sgt[lnk->link_idx], nents, GFP_KERNEL);
2045 if (buf_desc->is_vm) {
2046 /* virtually contiguous buffer */
2047 for_each_sg(buf_desc->sgt[lnk->link_idx].sgl, sg, nents, i) {
2048 size = min_t(int, PAGE_SIZE - offset, buf_size);
2049 sg_set_page(sg, vmalloc_to_page(buf), size, offset);
2050 buf += size / sizeof(*buf);
2055 /* physically contiguous buffer */
2056 sg_set_buf(buf_desc->sgt[lnk->link_idx].sgl,
2057 buf_desc->cpu_addr, buf_desc->len);
2060 /* map sg table to DMA address */
2061 rc = smc_ib_buf_map_sg(lnk, buf_desc,
2062 is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
2063 /* SMC protocol depends on mapping to one DMA address only */
2069 buf_desc->is_dma_need_sync |=
2070 smc_ib_is_sg_need_sync(lnk, buf_desc) << lnk->link_idx;
2072 if (is_rmb || buf_desc->is_vm) {
2073 /* create a new memory region for the RMB or vzalloced sndbuf */
2074 access_flags = is_rmb ?
2075 IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE :
2076 IB_ACCESS_LOCAL_WRITE;
2078 rc = smc_ib_get_memory_region(lnk->roce_pd, access_flags,
2079 buf_desc, lnk->link_idx);
2082 smc_ib_sync_sg_for_device(lnk, buf_desc,
2083 is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
2085 buf_desc->is_map_ib[lnk->link_idx] = true;
2089 smc_ib_buf_unmap_sg(lnk, buf_desc,
2090 is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
2092 sg_free_table(&buf_desc->sgt[lnk->link_idx]);
2096 /* register a new buf on IB device, rmb or vzalloced sndbuf
2097 * must be called under lgr->llc_conf_mutex lock
2099 int smcr_link_reg_buf(struct smc_link *link, struct smc_buf_desc *buf_desc)
2101 if (list_empty(&link->lgr->list))
2103 if (!buf_desc->is_reg_mr[link->link_idx]) {
2104 /* register memory region for new buf */
2105 if (buf_desc->is_vm)
2106 buf_desc->mr[link->link_idx]->iova =
2107 (uintptr_t)buf_desc->cpu_addr;
2108 if (smc_wr_reg_send(link, buf_desc->mr[link->link_idx])) {
2109 buf_desc->is_reg_err = true;
2112 buf_desc->is_reg_mr[link->link_idx] = true;
2117 static int _smcr_buf_map_lgr(struct smc_link *lnk, struct rw_semaphore *lock,
2118 struct list_head *lst, bool is_rmb)
2120 struct smc_buf_desc *buf_desc, *bf;
2124 list_for_each_entry_safe(buf_desc, bf, lst, list) {
2125 if (!buf_desc->used)
2127 rc = smcr_buf_map_link(buf_desc, is_rmb, lnk);
2136 /* map all used buffers of lgr for a new link */
2137 int smcr_buf_map_lgr(struct smc_link *lnk)
2139 struct smc_link_group *lgr = lnk->lgr;
2142 for (i = 0; i < SMC_RMBE_SIZES; i++) {
2143 rc = _smcr_buf_map_lgr(lnk, &lgr->rmbs_lock,
2144 &lgr->rmbs[i], true);
2147 rc = _smcr_buf_map_lgr(lnk, &lgr->sndbufs_lock,
2148 &lgr->sndbufs[i], false);
2155 /* register all used buffers of lgr for a new link,
2156 * must be called under lgr->llc_conf_mutex lock
2158 int smcr_buf_reg_lgr(struct smc_link *lnk)
2160 struct smc_link_group *lgr = lnk->lgr;
2161 struct smc_buf_desc *buf_desc, *bf;
2164 /* reg all RMBs for a new link */
2165 down_write(&lgr->rmbs_lock);
2166 for (i = 0; i < SMC_RMBE_SIZES; i++) {
2167 list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list) {
2168 if (!buf_desc->used)
2170 rc = smcr_link_reg_buf(lnk, buf_desc);
2172 up_write(&lgr->rmbs_lock);
2177 up_write(&lgr->rmbs_lock);
2179 if (lgr->buf_type == SMCR_PHYS_CONT_BUFS)
2182 /* reg all vzalloced sndbufs for a new link */
2183 down_write(&lgr->sndbufs_lock);
2184 for (i = 0; i < SMC_RMBE_SIZES; i++) {
2185 list_for_each_entry_safe(buf_desc, bf, &lgr->sndbufs[i], list) {
2186 if (!buf_desc->used || !buf_desc->is_vm)
2188 rc = smcr_link_reg_buf(lnk, buf_desc);
2190 up_write(&lgr->sndbufs_lock);
2195 up_write(&lgr->sndbufs_lock);
2199 static struct smc_buf_desc *smcr_new_buf_create(struct smc_link_group *lgr,
2200 bool is_rmb, int bufsize)
2202 struct smc_buf_desc *buf_desc;
2204 /* try to alloc a new buffer */
2205 buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
2207 return ERR_PTR(-ENOMEM);
2209 switch (lgr->buf_type) {
2210 case SMCR_PHYS_CONT_BUFS:
2211 case SMCR_MIXED_BUFS:
2212 buf_desc->order = get_order(bufsize);
2213 buf_desc->pages = alloc_pages(GFP_KERNEL | __GFP_NOWARN |
2214 __GFP_NOMEMALLOC | __GFP_COMP |
2215 __GFP_NORETRY | __GFP_ZERO,
2217 if (buf_desc->pages) {
2218 buf_desc->cpu_addr =
2219 (void *)page_address(buf_desc->pages);
2220 buf_desc->len = bufsize;
2221 buf_desc->is_vm = false;
2224 if (lgr->buf_type == SMCR_PHYS_CONT_BUFS)
2226 fallthrough; // try virtually continguous buf
2227 case SMCR_VIRT_CONT_BUFS:
2228 buf_desc->order = get_order(bufsize);
2229 buf_desc->cpu_addr = vzalloc(PAGE_SIZE << buf_desc->order);
2230 if (!buf_desc->cpu_addr)
2232 buf_desc->pages = NULL;
2233 buf_desc->len = bufsize;
2234 buf_desc->is_vm = true;
2241 return ERR_PTR(-EAGAIN);
2244 /* map buf_desc on all usable links,
2245 * unused buffers stay mapped as long as the link is up
2247 static int smcr_buf_map_usable_links(struct smc_link_group *lgr,
2248 struct smc_buf_desc *buf_desc, bool is_rmb)
2250 int i, rc = 0, cnt = 0;
2252 /* protect against parallel link reconfiguration */
2253 down_read(&lgr->llc_conf_mutex);
2254 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
2255 struct smc_link *lnk = &lgr->lnk[i];
2257 if (!smc_link_usable(lnk))
2259 if (smcr_buf_map_link(buf_desc, is_rmb, lnk)) {
2266 up_read(&lgr->llc_conf_mutex);
2272 static struct smc_buf_desc *smcd_new_buf_create(struct smc_link_group *lgr,
2273 bool is_dmb, int bufsize)
2275 struct smc_buf_desc *buf_desc;
2278 /* try to alloc a new DMB */
2279 buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
2281 return ERR_PTR(-ENOMEM);
2283 rc = smc_ism_register_dmb(lgr, bufsize, buf_desc);
2287 return ERR_PTR(-EAGAIN);
2289 return ERR_PTR(-ENOSPC);
2290 return ERR_PTR(-EIO);
2292 buf_desc->pages = virt_to_page(buf_desc->cpu_addr);
2293 /* CDC header stored in buf. So, pretend it was smaller */
2294 buf_desc->len = bufsize - sizeof(struct smcd_cdc_msg);
2296 buf_desc->cpu_addr = kzalloc(bufsize, GFP_KERNEL |
2297 __GFP_NOWARN | __GFP_NORETRY |
2299 if (!buf_desc->cpu_addr) {
2301 return ERR_PTR(-EAGAIN);
2303 buf_desc->len = bufsize;
2308 static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
2310 struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
2311 struct smc_connection *conn = &smc->conn;
2312 struct smc_link_group *lgr = conn->lgr;
2313 struct list_head *buf_list;
2314 int bufsize, bufsize_comp;
2315 struct rw_semaphore *lock; /* lock buffer list */
2316 bool is_dgraded = false;
2319 /* use socket recv buffer size (w/o overhead) as start value */
2320 bufsize = smc->sk.sk_rcvbuf / 2;
2322 /* use socket send buffer size (w/o overhead) as start value */
2323 bufsize = smc->sk.sk_sndbuf / 2;
2325 for (bufsize_comp = smc_compress_bufsize(bufsize, is_smcd, is_rmb);
2326 bufsize_comp >= 0; bufsize_comp--) {
2328 lock = &lgr->rmbs_lock;
2329 buf_list = &lgr->rmbs[bufsize_comp];
2331 lock = &lgr->sndbufs_lock;
2332 buf_list = &lgr->sndbufs[bufsize_comp];
2334 bufsize = smc_uncompress_bufsize(bufsize_comp);
2336 /* check for reusable slot in the link group */
2337 buf_desc = smc_buf_get_slot(bufsize_comp, lock, buf_list);
2339 buf_desc->is_dma_need_sync = 0;
2340 SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, bufsize);
2341 SMC_STAT_BUF_REUSE(smc, is_smcd, is_rmb);
2342 break; /* found reusable slot */
2346 buf_desc = smcd_new_buf_create(lgr, is_rmb, bufsize);
2348 buf_desc = smcr_new_buf_create(lgr, is_rmb, bufsize);
2350 if (PTR_ERR(buf_desc) == -ENOMEM)
2352 if (IS_ERR(buf_desc)) {
2355 SMC_STAT_RMB_DOWNGRADED(smc, is_smcd, is_rmb);
2360 SMC_STAT_RMB_ALLOC(smc, is_smcd, is_rmb);
2361 SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, bufsize);
2364 list_add(&buf_desc->list, buf_list);
2369 if (IS_ERR(buf_desc))
2370 return PTR_ERR(buf_desc);
2373 if (smcr_buf_map_usable_links(lgr, buf_desc, is_rmb)) {
2374 smcr_buf_unuse(buf_desc, is_rmb, lgr);
2380 conn->rmb_desc = buf_desc;
2381 conn->rmbe_size_comp = bufsize_comp;
2382 smc->sk.sk_rcvbuf = bufsize * 2;
2383 atomic_set(&conn->bytes_to_rcv, 0);
2384 conn->rmbe_update_limit =
2385 smc_rmb_wnd_update_limit(buf_desc->len);
2387 smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */
2389 conn->sndbuf_desc = buf_desc;
2390 smc->sk.sk_sndbuf = bufsize * 2;
2391 atomic_set(&conn->sndbuf_space, bufsize);
2396 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
2398 if (!conn->sndbuf_desc->is_dma_need_sync)
2400 if (!smc_conn_lgr_valid(conn) || conn->lgr->is_smcd ||
2401 !smc_link_active(conn->lnk))
2403 smc_ib_sync_sg_for_device(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE);
2406 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
2410 if (!conn->rmb_desc->is_dma_need_sync)
2412 if (!smc_conn_lgr_valid(conn) || conn->lgr->is_smcd)
2414 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
2415 if (!smc_link_active(&conn->lgr->lnk[i]))
2417 smc_ib_sync_sg_for_cpu(&conn->lgr->lnk[i], conn->rmb_desc,
2422 /* create the send and receive buffer for an SMC socket;
2423 * receive buffers are called RMBs;
2424 * (even though the SMC protocol allows more than one RMB-element per RMB,
2425 * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
2426 * extra RMB for every connection in a link group
2428 int smc_buf_create(struct smc_sock *smc, bool is_smcd)
2432 /* create send buffer */
2433 rc = __smc_buf_create(smc, is_smcd, false);
2437 rc = __smc_buf_create(smc, is_smcd, true);
2439 down_write(&smc->conn.lgr->sndbufs_lock);
2440 list_del(&smc->conn.sndbuf_desc->list);
2441 up_write(&smc->conn.lgr->sndbufs_lock);
2442 smc_buf_free(smc->conn.lgr, false, smc->conn.sndbuf_desc);
2443 smc->conn.sndbuf_desc = NULL;
2448 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
2452 for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
2453 if (!test_and_set_bit(i, lgr->rtokens_used_mask))
2459 static int smc_rtoken_find_by_link(struct smc_link_group *lgr, int lnk_idx,
2464 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2465 if (test_bit(i, lgr->rtokens_used_mask) &&
2466 lgr->rtokens[i][lnk_idx].rkey == rkey)
2472 /* set rtoken for a new link to an existing rmb */
2473 void smc_rtoken_set(struct smc_link_group *lgr, int link_idx, int link_idx_new,
2474 __be32 nw_rkey_known, __be64 nw_vaddr, __be32 nw_rkey)
2478 rtok_idx = smc_rtoken_find_by_link(lgr, link_idx, ntohl(nw_rkey_known));
2479 if (rtok_idx == -ENOENT)
2481 lgr->rtokens[rtok_idx][link_idx_new].rkey = ntohl(nw_rkey);
2482 lgr->rtokens[rtok_idx][link_idx_new].dma_addr = be64_to_cpu(nw_vaddr);
2485 /* set rtoken for a new link whose link_id is given */
2486 void smc_rtoken_set2(struct smc_link_group *lgr, int rtok_idx, int link_id,
2487 __be64 nw_vaddr, __be32 nw_rkey)
2489 u64 dma_addr = be64_to_cpu(nw_vaddr);
2490 u32 rkey = ntohl(nw_rkey);
2494 for (link_idx = 0; link_idx < SMC_LINKS_PER_LGR_MAX; link_idx++) {
2495 if (lgr->lnk[link_idx].link_id == link_id) {
2502 lgr->rtokens[rtok_idx][link_idx].rkey = rkey;
2503 lgr->rtokens[rtok_idx][link_idx].dma_addr = dma_addr;
2506 /* add a new rtoken from peer */
2507 int smc_rtoken_add(struct smc_link *lnk, __be64 nw_vaddr, __be32 nw_rkey)
2509 struct smc_link_group *lgr = smc_get_lgr(lnk);
2510 u64 dma_addr = be64_to_cpu(nw_vaddr);
2511 u32 rkey = ntohl(nw_rkey);
2514 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2515 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
2516 lgr->rtokens[i][lnk->link_idx].dma_addr == dma_addr &&
2517 test_bit(i, lgr->rtokens_used_mask)) {
2518 /* already in list */
2522 i = smc_rmb_reserve_rtoken_idx(lgr);
2525 lgr->rtokens[i][lnk->link_idx].rkey = rkey;
2526 lgr->rtokens[i][lnk->link_idx].dma_addr = dma_addr;
2530 /* delete an rtoken from all links */
2531 int smc_rtoken_delete(struct smc_link *lnk, __be32 nw_rkey)
2533 struct smc_link_group *lgr = smc_get_lgr(lnk);
2534 u32 rkey = ntohl(nw_rkey);
2537 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2538 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
2539 test_bit(i, lgr->rtokens_used_mask)) {
2540 for (j = 0; j < SMC_LINKS_PER_LGR_MAX; j++) {
2541 lgr->rtokens[i][j].rkey = 0;
2542 lgr->rtokens[i][j].dma_addr = 0;
2544 clear_bit(i, lgr->rtokens_used_mask);
2551 /* save rkey and dma_addr received from peer during clc handshake */
2552 int smc_rmb_rtoken_handling(struct smc_connection *conn,
2553 struct smc_link *lnk,
2554 struct smc_clc_msg_accept_confirm *clc)
2556 conn->rtoken_idx = smc_rtoken_add(lnk, clc->r0.rmb_dma_addr,
2558 if (conn->rtoken_idx < 0)
2559 return conn->rtoken_idx;
2563 static void smc_core_going_away(void)
2565 struct smc_ib_device *smcibdev;
2566 struct smcd_dev *smcd;
2568 mutex_lock(&smc_ib_devices.mutex);
2569 list_for_each_entry(smcibdev, &smc_ib_devices.list, list) {
2572 for (i = 0; i < SMC_MAX_PORTS; i++)
2573 set_bit(i, smcibdev->ports_going_away);
2575 mutex_unlock(&smc_ib_devices.mutex);
2577 mutex_lock(&smcd_dev_list.mutex);
2578 list_for_each_entry(smcd, &smcd_dev_list.list, list) {
2579 smcd->going_away = 1;
2581 mutex_unlock(&smcd_dev_list.mutex);
2584 /* Clean up all SMC link groups */
2585 static void smc_lgrs_shutdown(void)
2587 struct smcd_dev *smcd;
2589 smc_core_going_away();
2591 smc_smcr_terminate_all(NULL);
2593 mutex_lock(&smcd_dev_list.mutex);
2594 list_for_each_entry(smcd, &smcd_dev_list.list, list)
2595 smc_smcd_terminate_all(smcd);
2596 mutex_unlock(&smcd_dev_list.mutex);
2599 static int smc_core_reboot_event(struct notifier_block *this,
2600 unsigned long event, void *ptr)
2602 smc_lgrs_shutdown();
2603 smc_ib_unregister_client();
2608 static struct notifier_block smc_reboot_notifier = {
2609 .notifier_call = smc_core_reboot_event,
2612 int __init smc_core_init(void)
2614 return register_reboot_notifier(&smc_reboot_notifier);
2617 /* Called (from smc_exit) when module is removed */
2618 void smc_core_exit(void)
2620 unregister_reboot_notifier(&smc_reboot_notifier);
2621 smc_lgrs_shutdown();