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>
21 #include <rdma/ib_verbs.h>
22 #include <rdma/ib_cache.h>
31 #include "smc_close.h"
34 #define SMC_LGR_NUM_INCR 256
35 #define SMC_LGR_FREE_DELAY_SERV (600 * HZ)
36 #define SMC_LGR_FREE_DELAY_CLNT (SMC_LGR_FREE_DELAY_SERV + 10 * HZ)
37 #define SMC_LGR_FREE_DELAY_FAST (8 * HZ)
39 static struct smc_lgr_list smc_lgr_list = { /* established link groups */
40 .lock = __SPIN_LOCK_UNLOCKED(smc_lgr_list.lock),
41 .list = LIST_HEAD_INIT(smc_lgr_list.list),
45 static atomic_t lgr_cnt = ATOMIC_INIT(0); /* number of existing link groups */
46 static DECLARE_WAIT_QUEUE_HEAD(lgrs_deleted);
48 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
49 struct smc_buf_desc *buf_desc);
50 static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft);
52 static void smc_link_down_work(struct work_struct *work);
54 /* return head of link group list and its lock for a given link group */
55 static inline struct list_head *smc_lgr_list_head(struct smc_link_group *lgr,
56 spinlock_t **lgr_lock)
59 *lgr_lock = &lgr->smcd->lgr_lock;
60 return &lgr->smcd->lgr_list;
63 *lgr_lock = &smc_lgr_list.lock;
64 return &smc_lgr_list.list;
67 static void smc_lgr_schedule_free_work(struct smc_link_group *lgr)
69 /* client link group creation always follows the server link group
70 * creation. For client use a somewhat higher removal delay time,
71 * otherwise there is a risk of out-of-sync link groups.
73 if (!lgr->freeing && !lgr->freefast) {
74 mod_delayed_work(system_wq, &lgr->free_work,
75 (!lgr->is_smcd && lgr->role == SMC_CLNT) ?
76 SMC_LGR_FREE_DELAY_CLNT :
77 SMC_LGR_FREE_DELAY_SERV);
81 void smc_lgr_schedule_free_work_fast(struct smc_link_group *lgr)
83 if (!lgr->freeing && !lgr->freefast) {
85 mod_delayed_work(system_wq, &lgr->free_work,
86 SMC_LGR_FREE_DELAY_FAST);
90 /* Register connection's alert token in our lookup structure.
91 * To use rbtrees we have to implement our own insert core.
92 * Requires @conns_lock
93 * @smc connection to register
94 * Returns 0 on success, != otherwise.
96 static void smc_lgr_add_alert_token(struct smc_connection *conn)
98 struct rb_node **link, *parent = NULL;
99 u32 token = conn->alert_token_local;
101 link = &conn->lgr->conns_all.rb_node;
103 struct smc_connection *cur = rb_entry(*link,
104 struct smc_connection, alert_node);
107 if (cur->alert_token_local > token)
108 link = &parent->rb_left;
110 link = &parent->rb_right;
112 /* Put the new node there */
113 rb_link_node(&conn->alert_node, parent, link);
114 rb_insert_color(&conn->alert_node, &conn->lgr->conns_all);
117 /* assign an SMC-R link to the connection */
118 static int smcr_lgr_conn_assign_link(struct smc_connection *conn, bool first)
120 enum smc_link_state expected = first ? SMC_LNK_ACTIVATING :
124 /* do link balancing */
125 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
126 struct smc_link *lnk = &conn->lgr->lnk[i];
128 if (lnk->state != expected || lnk->link_is_asym)
130 if (conn->lgr->role == SMC_CLNT) {
131 conn->lnk = lnk; /* temporary, SMC server assigns link*/
134 if (conn->lgr->conns_num % 2) {
135 for (j = i + 1; j < SMC_LINKS_PER_LGR_MAX; j++) {
136 struct smc_link *lnk2;
138 lnk2 = &conn->lgr->lnk[j];
139 if (lnk2->state == expected &&
140 !lnk2->link_is_asym) {
151 return SMC_CLC_DECL_NOACTLINK;
155 /* Register connection in link group by assigning an alert token
156 * registered in a search tree.
157 * Requires @conns_lock
158 * Note that '0' is a reserved value and not assigned.
160 static int smc_lgr_register_conn(struct smc_connection *conn, bool first)
162 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
163 static atomic_t nexttoken = ATOMIC_INIT(0);
166 if (!conn->lgr->is_smcd) {
167 rc = smcr_lgr_conn_assign_link(conn, first);
171 /* find a new alert_token_local value not yet used by some connection
174 sock_hold(&smc->sk); /* sock_put in smc_lgr_unregister_conn() */
175 while (!conn->alert_token_local) {
176 conn->alert_token_local = atomic_inc_return(&nexttoken);
177 if (smc_lgr_find_conn(conn->alert_token_local, conn->lgr))
178 conn->alert_token_local = 0;
180 smc_lgr_add_alert_token(conn);
181 conn->lgr->conns_num++;
185 /* Unregister connection and reset the alert token of the given connection<
187 static void __smc_lgr_unregister_conn(struct smc_connection *conn)
189 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
190 struct smc_link_group *lgr = conn->lgr;
192 rb_erase(&conn->alert_node, &lgr->conns_all);
194 conn->alert_token_local = 0;
195 sock_put(&smc->sk); /* sock_hold in smc_lgr_register_conn() */
198 /* Unregister connection from lgr
200 static void smc_lgr_unregister_conn(struct smc_connection *conn)
202 struct smc_link_group *lgr = conn->lgr;
206 write_lock_bh(&lgr->conns_lock);
207 if (conn->alert_token_local) {
208 __smc_lgr_unregister_conn(conn);
210 write_unlock_bh(&lgr->conns_lock);
214 void smc_lgr_cleanup_early(struct smc_connection *conn)
216 struct smc_link_group *lgr = conn->lgr;
217 struct list_head *lgr_list;
218 spinlock_t *lgr_lock;
224 lgr_list = smc_lgr_list_head(lgr, &lgr_lock);
225 spin_lock_bh(lgr_lock);
226 /* do not use this link group for new connections */
227 if (!list_empty(lgr_list))
228 list_del_init(lgr_list);
229 spin_unlock_bh(lgr_lock);
230 smc_lgr_schedule_free_work_fast(lgr);
233 static void smcr_lgr_link_deactivate_all(struct smc_link_group *lgr)
237 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
238 struct smc_link *lnk = &lgr->lnk[i];
240 if (smc_link_usable(lnk))
241 lnk->state = SMC_LNK_INACTIVE;
243 wake_up_all(&lgr->llc_msg_waiter);
244 wake_up_all(&lgr->llc_flow_waiter);
247 static void smc_lgr_free(struct smc_link_group *lgr);
249 static void smc_lgr_free_work(struct work_struct *work)
251 struct smc_link_group *lgr = container_of(to_delayed_work(work),
252 struct smc_link_group,
254 spinlock_t *lgr_lock;
257 smc_lgr_list_head(lgr, &lgr_lock);
258 spin_lock_bh(lgr_lock);
260 spin_unlock_bh(lgr_lock);
263 read_lock_bh(&lgr->conns_lock);
264 conns = RB_EMPTY_ROOT(&lgr->conns_all);
265 read_unlock_bh(&lgr->conns_lock);
266 if (!conns) { /* number of lgr connections is no longer zero */
267 spin_unlock_bh(lgr_lock);
270 list_del_init(&lgr->list); /* remove from smc_lgr_list */
271 lgr->freeing = 1; /* this instance does the freeing, no new schedule */
272 spin_unlock_bh(lgr_lock);
273 cancel_delayed_work(&lgr->free_work);
275 if (!lgr->is_smcd && !lgr->terminating)
276 smc_llc_send_link_delete_all(lgr, true,
277 SMC_LLC_DEL_PROG_INIT_TERM);
278 if (lgr->is_smcd && !lgr->terminating)
279 smc_ism_signal_shutdown(lgr);
281 smcr_lgr_link_deactivate_all(lgr);
285 static void smc_lgr_terminate_work(struct work_struct *work)
287 struct smc_link_group *lgr = container_of(work, struct smc_link_group,
290 __smc_lgr_terminate(lgr, true);
293 /* return next unique link id for the lgr */
294 static u8 smcr_next_link_id(struct smc_link_group *lgr)
300 link_id = ++lgr->next_link_id;
301 if (!link_id) /* skip zero as link_id */
302 link_id = ++lgr->next_link_id;
303 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
304 if (smc_link_usable(&lgr->lnk[i]) &&
305 lgr->lnk[i].link_id == link_id)
313 int smcr_link_init(struct smc_link_group *lgr, struct smc_link *lnk,
314 u8 link_idx, struct smc_init_info *ini)
319 get_device(&ini->ib_dev->ibdev->dev);
320 atomic_inc(&ini->ib_dev->lnk_cnt);
321 lnk->link_id = smcr_next_link_id(lgr);
323 lnk->link_idx = link_idx;
324 lnk->smcibdev = ini->ib_dev;
325 lnk->ibport = ini->ib_port;
326 lnk->path_mtu = ini->ib_dev->pattr[ini->ib_port - 1].active_mtu;
327 smc_llc_link_set_uid(lnk);
328 INIT_WORK(&lnk->link_down_wrk, smc_link_down_work);
329 if (!ini->ib_dev->initialized) {
330 rc = (int)smc_ib_setup_per_ibdev(ini->ib_dev);
334 get_random_bytes(rndvec, sizeof(rndvec));
335 lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) +
337 rc = smc_ib_determine_gid(lnk->smcibdev, lnk->ibport,
338 ini->vlan_id, lnk->gid, &lnk->sgid_index);
341 rc = smc_llc_link_init(lnk);
344 rc = smc_wr_alloc_link_mem(lnk);
347 rc = smc_ib_create_protection_domain(lnk);
350 rc = smc_ib_create_queue_pair(lnk);
353 rc = smc_wr_create_link(lnk);
356 lnk->state = SMC_LNK_ACTIVATING;
360 smc_ib_destroy_queue_pair(lnk);
362 smc_ib_dealloc_protection_domain(lnk);
364 smc_wr_free_link_mem(lnk);
366 smc_llc_link_clear(lnk, false);
368 put_device(&ini->ib_dev->ibdev->dev);
369 memset(lnk, 0, sizeof(struct smc_link));
370 lnk->state = SMC_LNK_UNUSED;
371 if (!atomic_dec_return(&ini->ib_dev->lnk_cnt))
372 wake_up(&ini->ib_dev->lnks_deleted);
376 /* create a new SMC link group */
377 static int smc_lgr_create(struct smc_sock *smc, struct smc_init_info *ini)
379 struct smc_link_group *lgr;
380 struct list_head *lgr_list;
381 struct smc_link *lnk;
382 spinlock_t *lgr_lock;
387 if (ini->is_smcd && ini->vlan_id) {
388 if (smc_ism_get_vlan(ini->ism_dev, ini->vlan_id)) {
389 rc = SMC_CLC_DECL_ISMVLANERR;
394 lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
396 rc = SMC_CLC_DECL_MEM;
399 lgr->is_smcd = ini->is_smcd;
401 lgr->terminating = 0;
404 lgr->vlan_id = ini->vlan_id;
405 mutex_init(&lgr->sndbufs_lock);
406 mutex_init(&lgr->rmbs_lock);
407 rwlock_init(&lgr->conns_lock);
408 for (i = 0; i < SMC_RMBE_SIZES; i++) {
409 INIT_LIST_HEAD(&lgr->sndbufs[i]);
410 INIT_LIST_HEAD(&lgr->rmbs[i]);
412 lgr->next_link_id = 0;
413 smc_lgr_list.num += SMC_LGR_NUM_INCR;
414 memcpy(&lgr->id, (u8 *)&smc_lgr_list.num, SMC_LGR_ID_SIZE);
415 INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
416 INIT_WORK(&lgr->terminate_work, smc_lgr_terminate_work);
417 lgr->conns_all = RB_ROOT;
419 /* SMC-D specific settings */
420 get_device(&ini->ism_dev->dev);
421 lgr->peer_gid = ini->ism_gid;
422 lgr->smcd = ini->ism_dev;
423 lgr_list = &ini->ism_dev->lgr_list;
424 lgr_lock = &lgr->smcd->lgr_lock;
425 lgr->peer_shutdown = 0;
426 atomic_inc(&ini->ism_dev->lgr_cnt);
428 /* SMC-R specific settings */
429 lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
430 memcpy(lgr->peer_systemid, ini->ib_lcl->id_for_peer,
432 memcpy(lgr->pnet_id, ini->ib_dev->pnetid[ini->ib_port - 1],
434 smc_llc_lgr_init(lgr, smc);
436 link_idx = SMC_SINGLE_LINK;
437 lnk = &lgr->lnk[link_idx];
438 rc = smcr_link_init(lgr, lnk, link_idx, ini);
441 lgr_list = &smc_lgr_list.list;
442 lgr_lock = &smc_lgr_list.lock;
443 atomic_inc(&lgr_cnt);
446 spin_lock_bh(lgr_lock);
447 list_add_tail(&lgr->list, lgr_list);
448 spin_unlock_bh(lgr_lock);
454 if (ini->is_smcd && ini->vlan_id)
455 smc_ism_put_vlan(ini->ism_dev, ini->vlan_id);
459 rc = SMC_CLC_DECL_MEM;
461 rc = SMC_CLC_DECL_INTERR;
466 static int smc_write_space(struct smc_connection *conn)
468 int buffer_len = conn->peer_rmbe_size;
469 union smc_host_cursor prod;
470 union smc_host_cursor cons;
473 smc_curs_copy(&prod, &conn->local_tx_ctrl.prod, conn);
474 smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
475 /* determine rx_buf space */
476 space = buffer_len - smc_curs_diff(buffer_len, &cons, &prod);
480 static int smc_switch_cursor(struct smc_sock *smc, struct smc_cdc_tx_pend *pend,
481 struct smc_wr_buf *wr_buf)
483 struct smc_connection *conn = &smc->conn;
484 union smc_host_cursor cons, fin;
488 smc_curs_copy(&conn->tx_curs_sent, &conn->tx_curs_fin, conn);
489 smc_curs_copy(&fin, &conn->local_tx_ctrl_fin, conn);
490 /* set prod cursor to old state, enforce tx_rdma_writes() */
491 smc_curs_copy(&conn->local_tx_ctrl.prod, &fin, conn);
492 smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
494 if (smc_curs_comp(conn->peer_rmbe_size, &cons, &fin) < 0) {
495 /* cons cursor advanced more than fin, and prod was set
496 * fin above, so now prod is smaller than cons. Fix that.
498 diff = smc_curs_diff(conn->peer_rmbe_size, &fin, &cons);
499 smc_curs_add(conn->sndbuf_desc->len,
500 &conn->tx_curs_sent, diff);
501 smc_curs_add(conn->sndbuf_desc->len,
502 &conn->tx_curs_fin, diff);
504 smp_mb__before_atomic();
505 atomic_add(diff, &conn->sndbuf_space);
506 smp_mb__after_atomic();
508 smc_curs_add(conn->peer_rmbe_size,
509 &conn->local_tx_ctrl.prod, diff);
510 smc_curs_add(conn->peer_rmbe_size,
511 &conn->local_tx_ctrl_fin, diff);
513 /* recalculate, value is used by tx_rdma_writes() */
514 atomic_set(&smc->conn.peer_rmbe_space, smc_write_space(conn));
516 if (smc->sk.sk_state != SMC_INIT &&
517 smc->sk.sk_state != SMC_CLOSED) {
518 rc = smcr_cdc_msg_send_validation(conn, pend, wr_buf);
520 schedule_delayed_work(&conn->tx_work, 0);
521 smc->sk.sk_data_ready(&smc->sk);
524 smc_wr_tx_put_slot(conn->lnk,
525 (struct smc_wr_tx_pend_priv *)pend);
530 struct smc_link *smc_switch_conns(struct smc_link_group *lgr,
531 struct smc_link *from_lnk, bool is_dev_err)
533 struct smc_link *to_lnk = NULL;
534 struct smc_cdc_tx_pend *pend;
535 struct smc_connection *conn;
536 struct smc_wr_buf *wr_buf;
537 struct smc_sock *smc;
538 struct rb_node *node;
541 /* link is inactive, wake up tx waiters */
542 smc_wr_wakeup_tx_wait(from_lnk);
544 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
545 if (!smc_link_active(&lgr->lnk[i]) || i == from_lnk->link_idx)
547 if (is_dev_err && from_lnk->smcibdev == lgr->lnk[i].smcibdev &&
548 from_lnk->ibport == lgr->lnk[i].ibport) {
551 to_lnk = &lgr->lnk[i];
555 smc_lgr_terminate_sched(lgr);
559 read_lock_bh(&lgr->conns_lock);
560 for (node = rb_first(&lgr->conns_all); node; node = rb_next(node)) {
561 conn = rb_entry(node, struct smc_connection, alert_node);
562 if (conn->lnk != from_lnk)
564 smc = container_of(conn, struct smc_sock, conn);
565 /* conn->lnk not yet set in SMC_INIT state */
566 if (smc->sk.sk_state == SMC_INIT)
568 if (smc->sk.sk_state == SMC_CLOSED ||
569 smc->sk.sk_state == SMC_PEERCLOSEWAIT1 ||
570 smc->sk.sk_state == SMC_PEERCLOSEWAIT2 ||
571 smc->sk.sk_state == SMC_APPFINCLOSEWAIT ||
572 smc->sk.sk_state == SMC_APPCLOSEWAIT1 ||
573 smc->sk.sk_state == SMC_APPCLOSEWAIT2 ||
574 smc->sk.sk_state == SMC_PEERFINCLOSEWAIT ||
575 smc->sk.sk_state == SMC_PEERABORTWAIT ||
576 smc->sk.sk_state == SMC_PROCESSABORT) {
577 spin_lock_bh(&conn->send_lock);
579 spin_unlock_bh(&conn->send_lock);
583 read_unlock_bh(&lgr->conns_lock);
584 /* pre-fetch buffer outside of send_lock, might sleep */
585 rc = smc_cdc_get_free_slot(conn, to_lnk, &wr_buf, NULL, &pend);
587 smcr_link_down_cond_sched(to_lnk);
590 /* avoid race with smcr_tx_sndbuf_nonempty() */
591 spin_lock_bh(&conn->send_lock);
593 rc = smc_switch_cursor(smc, pend, wr_buf);
594 spin_unlock_bh(&conn->send_lock);
597 smcr_link_down_cond_sched(to_lnk);
602 read_unlock_bh(&lgr->conns_lock);
606 static void smcr_buf_unuse(struct smc_buf_desc *rmb_desc,
607 struct smc_link_group *lgr)
611 if (rmb_desc->is_conf_rkey && !list_empty(&lgr->list)) {
612 /* unregister rmb with peer */
613 rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY);
615 /* protect against smc_llc_cli_rkey_exchange() */
616 mutex_lock(&lgr->llc_conf_mutex);
617 smc_llc_do_delete_rkey(lgr, rmb_desc);
618 rmb_desc->is_conf_rkey = false;
619 mutex_unlock(&lgr->llc_conf_mutex);
620 smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl);
624 if (rmb_desc->is_reg_err) {
625 /* buf registration failed, reuse not possible */
626 mutex_lock(&lgr->rmbs_lock);
627 list_del(&rmb_desc->list);
628 mutex_unlock(&lgr->rmbs_lock);
630 smc_buf_free(lgr, true, rmb_desc);
636 static void smc_buf_unuse(struct smc_connection *conn,
637 struct smc_link_group *lgr)
639 if (conn->sndbuf_desc)
640 conn->sndbuf_desc->used = 0;
641 if (conn->rmb_desc && lgr->is_smcd)
642 conn->rmb_desc->used = 0;
643 else if (conn->rmb_desc)
644 smcr_buf_unuse(conn->rmb_desc, lgr);
647 /* remove a finished connection from its link group */
648 void smc_conn_free(struct smc_connection *conn)
650 struct smc_link_group *lgr = conn->lgr;
655 if (!list_empty(&lgr->list))
656 smc_ism_unset_conn(conn);
657 tasklet_kill(&conn->rx_tsklet);
659 smc_cdc_tx_dismiss_slots(conn);
660 if (current_work() != &conn->abort_work)
661 cancel_work_sync(&conn->abort_work);
663 if (!list_empty(&lgr->list)) {
664 smc_lgr_unregister_conn(conn);
665 smc_buf_unuse(conn, lgr); /* allow buffer reuse */
669 smc_lgr_schedule_free_work(lgr);
672 /* unregister a link from a buf_desc */
673 static void smcr_buf_unmap_link(struct smc_buf_desc *buf_desc, bool is_rmb,
674 struct smc_link *lnk)
677 buf_desc->is_reg_mr[lnk->link_idx] = false;
678 if (!buf_desc->is_map_ib[lnk->link_idx])
681 if (buf_desc->mr_rx[lnk->link_idx]) {
682 smc_ib_put_memory_region(
683 buf_desc->mr_rx[lnk->link_idx]);
684 buf_desc->mr_rx[lnk->link_idx] = NULL;
686 smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_FROM_DEVICE);
688 smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_TO_DEVICE);
690 sg_free_table(&buf_desc->sgt[lnk->link_idx]);
691 buf_desc->is_map_ib[lnk->link_idx] = false;
694 /* unmap all buffers of lgr for a deleted link */
695 static void smcr_buf_unmap_lgr(struct smc_link *lnk)
697 struct smc_link_group *lgr = lnk->lgr;
698 struct smc_buf_desc *buf_desc, *bf;
701 for (i = 0; i < SMC_RMBE_SIZES; i++) {
702 mutex_lock(&lgr->rmbs_lock);
703 list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list)
704 smcr_buf_unmap_link(buf_desc, true, lnk);
705 mutex_unlock(&lgr->rmbs_lock);
706 mutex_lock(&lgr->sndbufs_lock);
707 list_for_each_entry_safe(buf_desc, bf, &lgr->sndbufs[i],
709 smcr_buf_unmap_link(buf_desc, false, lnk);
710 mutex_unlock(&lgr->sndbufs_lock);
714 static void smcr_rtoken_clear_link(struct smc_link *lnk)
716 struct smc_link_group *lgr = lnk->lgr;
719 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
720 lgr->rtokens[i][lnk->link_idx].rkey = 0;
721 lgr->rtokens[i][lnk->link_idx].dma_addr = 0;
725 /* must be called under lgr->llc_conf_mutex lock */
726 void smcr_link_clear(struct smc_link *lnk, bool log)
728 struct smc_ib_device *smcibdev;
730 if (!lnk->lgr || lnk->state == SMC_LNK_UNUSED)
733 smc_llc_link_clear(lnk, log);
734 smcr_buf_unmap_lgr(lnk);
735 smcr_rtoken_clear_link(lnk);
736 smc_ib_modify_qp_reset(lnk);
737 smc_wr_free_link(lnk);
738 smc_ib_destroy_queue_pair(lnk);
739 smc_ib_dealloc_protection_domain(lnk);
740 smc_wr_free_link_mem(lnk);
741 put_device(&lnk->smcibdev->ibdev->dev);
742 smcibdev = lnk->smcibdev;
743 memset(lnk, 0, sizeof(struct smc_link));
744 lnk->state = SMC_LNK_UNUSED;
745 if (!atomic_dec_return(&smcibdev->lnk_cnt))
746 wake_up(&smcibdev->lnks_deleted);
749 static void smcr_buf_free(struct smc_link_group *lgr, bool is_rmb,
750 struct smc_buf_desc *buf_desc)
754 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
755 smcr_buf_unmap_link(buf_desc, is_rmb, &lgr->lnk[i]);
758 __free_pages(buf_desc->pages, buf_desc->order);
762 static void smcd_buf_free(struct smc_link_group *lgr, bool is_dmb,
763 struct smc_buf_desc *buf_desc)
766 /* restore original buf len */
767 buf_desc->len += sizeof(struct smcd_cdc_msg);
768 smc_ism_unregister_dmb(lgr->smcd, buf_desc);
770 kfree(buf_desc->cpu_addr);
775 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
776 struct smc_buf_desc *buf_desc)
779 smcd_buf_free(lgr, is_rmb, buf_desc);
781 smcr_buf_free(lgr, is_rmb, buf_desc);
784 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
786 struct smc_buf_desc *buf_desc, *bf_desc;
787 struct list_head *buf_list;
790 for (i = 0; i < SMC_RMBE_SIZES; i++) {
792 buf_list = &lgr->rmbs[i];
794 buf_list = &lgr->sndbufs[i];
795 list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
797 list_del(&buf_desc->list);
798 smc_buf_free(lgr, is_rmb, buf_desc);
803 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
805 /* free send buffers */
806 __smc_lgr_free_bufs(lgr, false);
808 __smc_lgr_free_bufs(lgr, true);
811 /* remove a link group */
812 static void smc_lgr_free(struct smc_link_group *lgr)
817 mutex_lock(&lgr->llc_conf_mutex);
818 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
819 if (lgr->lnk[i].state != SMC_LNK_UNUSED)
820 smcr_link_clear(&lgr->lnk[i], false);
822 mutex_unlock(&lgr->llc_conf_mutex);
823 smc_llc_lgr_clear(lgr);
826 smc_lgr_free_bufs(lgr);
828 if (!lgr->terminating) {
829 smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
830 put_device(&lgr->smcd->dev);
832 if (!atomic_dec_return(&lgr->smcd->lgr_cnt))
833 wake_up(&lgr->smcd->lgrs_deleted);
835 if (!atomic_dec_return(&lgr_cnt))
836 wake_up(&lgrs_deleted);
841 static void smcd_unregister_all_dmbs(struct smc_link_group *lgr)
845 for (i = 0; i < SMC_RMBE_SIZES; i++) {
846 struct smc_buf_desc *buf_desc;
848 list_for_each_entry(buf_desc, &lgr->rmbs[i], list) {
849 buf_desc->len += sizeof(struct smcd_cdc_msg);
850 smc_ism_unregister_dmb(lgr->smcd, buf_desc);
855 static void smc_sk_wake_ups(struct smc_sock *smc)
857 smc->sk.sk_write_space(&smc->sk);
858 smc->sk.sk_data_ready(&smc->sk);
859 smc->sk.sk_state_change(&smc->sk);
862 /* kill a connection */
863 static void smc_conn_kill(struct smc_connection *conn, bool soft)
865 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
867 if (conn->lgr->is_smcd && conn->lgr->peer_shutdown)
868 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
870 smc_close_abort(conn);
872 smc->sk.sk_err = ECONNABORTED;
873 smc_sk_wake_ups(smc);
874 if (conn->lgr->is_smcd) {
875 smc_ism_unset_conn(conn);
877 tasklet_kill(&conn->rx_tsklet);
879 tasklet_unlock_wait(&conn->rx_tsklet);
881 smc_cdc_tx_dismiss_slots(conn);
883 smc_lgr_unregister_conn(conn);
884 smc_close_active_abort(smc);
887 static void smc_lgr_cleanup(struct smc_link_group *lgr)
890 smc_ism_signal_shutdown(lgr);
891 smcd_unregister_all_dmbs(lgr);
892 smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
893 put_device(&lgr->smcd->dev);
895 u32 rsn = lgr->llc_termination_rsn;
898 rsn = SMC_LLC_DEL_PROG_INIT_TERM;
899 smc_llc_send_link_delete_all(lgr, false, rsn);
900 smcr_lgr_link_deactivate_all(lgr);
904 /* terminate link group
905 * @soft: true if link group shutdown can take its time
906 * false if immediate link group shutdown is required
908 static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft)
910 struct smc_connection *conn;
911 struct smc_sock *smc;
912 struct rb_node *node;
914 if (lgr->terminating)
915 return; /* lgr already terminating */
916 /* cancel free_work sync, will terminate when lgr->freeing is set */
917 cancel_delayed_work_sync(&lgr->free_work);
918 lgr->terminating = 1;
920 /* kill remaining link group connections */
921 read_lock_bh(&lgr->conns_lock);
922 node = rb_first(&lgr->conns_all);
924 read_unlock_bh(&lgr->conns_lock);
925 conn = rb_entry(node, struct smc_connection, alert_node);
926 smc = container_of(conn, struct smc_sock, conn);
927 sock_hold(&smc->sk); /* sock_put below */
929 smc_conn_kill(conn, soft);
930 release_sock(&smc->sk);
931 sock_put(&smc->sk); /* sock_hold above */
932 read_lock_bh(&lgr->conns_lock);
933 node = rb_first(&lgr->conns_all);
935 read_unlock_bh(&lgr->conns_lock);
936 smc_lgr_cleanup(lgr);
940 /* unlink link group and schedule termination */
941 void smc_lgr_terminate_sched(struct smc_link_group *lgr)
943 spinlock_t *lgr_lock;
945 smc_lgr_list_head(lgr, &lgr_lock);
946 spin_lock_bh(lgr_lock);
947 if (list_empty(&lgr->list) || lgr->terminating || lgr->freeing) {
948 spin_unlock_bh(lgr_lock);
949 return; /* lgr already terminating */
951 list_del_init(&lgr->list);
953 spin_unlock_bh(lgr_lock);
954 schedule_work(&lgr->terminate_work);
957 /* Called when peer lgr shutdown (regularly or abnormally) is received */
958 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid, unsigned short vlan)
960 struct smc_link_group *lgr, *l;
961 LIST_HEAD(lgr_free_list);
963 /* run common cleanup function and build free list */
964 spin_lock_bh(&dev->lgr_lock);
965 list_for_each_entry_safe(lgr, l, &dev->lgr_list, list) {
966 if ((!peer_gid || lgr->peer_gid == peer_gid) &&
967 (vlan == VLAN_VID_MASK || lgr->vlan_id == vlan)) {
968 if (peer_gid) /* peer triggered termination */
969 lgr->peer_shutdown = 1;
970 list_move(&lgr->list, &lgr_free_list);
974 spin_unlock_bh(&dev->lgr_lock);
976 /* cancel the regular free workers and actually free lgrs */
977 list_for_each_entry_safe(lgr, l, &lgr_free_list, list) {
978 list_del_init(&lgr->list);
979 schedule_work(&lgr->terminate_work);
983 /* Called when an SMCD device is removed or the smc module is unloaded */
984 void smc_smcd_terminate_all(struct smcd_dev *smcd)
986 struct smc_link_group *lgr, *lg;
987 LIST_HEAD(lgr_free_list);
989 spin_lock_bh(&smcd->lgr_lock);
990 list_splice_init(&smcd->lgr_list, &lgr_free_list);
991 list_for_each_entry(lgr, &lgr_free_list, list)
993 spin_unlock_bh(&smcd->lgr_lock);
995 list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
996 list_del_init(&lgr->list);
997 __smc_lgr_terminate(lgr, false);
1000 if (atomic_read(&smcd->lgr_cnt))
1001 wait_event(smcd->lgrs_deleted, !atomic_read(&smcd->lgr_cnt));
1004 /* Called when an SMCR device is removed or the smc module is unloaded.
1005 * If smcibdev is given, all SMCR link groups using this device are terminated.
1006 * If smcibdev is NULL, all SMCR link groups are terminated.
1008 void smc_smcr_terminate_all(struct smc_ib_device *smcibdev)
1010 struct smc_link_group *lgr, *lg;
1011 LIST_HEAD(lgr_free_list);
1014 spin_lock_bh(&smc_lgr_list.lock);
1016 list_splice_init(&smc_lgr_list.list, &lgr_free_list);
1017 list_for_each_entry(lgr, &lgr_free_list, list)
1020 list_for_each_entry_safe(lgr, lg, &smc_lgr_list.list, list) {
1021 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1022 if (lgr->lnk[i].smcibdev == smcibdev)
1023 smcr_link_down_cond_sched(&lgr->lnk[i]);
1027 spin_unlock_bh(&smc_lgr_list.lock);
1029 list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
1030 list_del_init(&lgr->list);
1031 smc_llc_set_termination_rsn(lgr, SMC_LLC_DEL_OP_INIT_TERM);
1032 __smc_lgr_terminate(lgr, false);
1036 if (atomic_read(&smcibdev->lnk_cnt))
1037 wait_event(smcibdev->lnks_deleted,
1038 !atomic_read(&smcibdev->lnk_cnt));
1040 if (atomic_read(&lgr_cnt))
1041 wait_event(lgrs_deleted, !atomic_read(&lgr_cnt));
1045 /* set new lgr type and clear all asymmetric link tagging */
1046 void smcr_lgr_set_type(struct smc_link_group *lgr, enum smc_lgr_type new_type)
1048 char *lgr_type = "";
1051 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
1052 if (smc_link_usable(&lgr->lnk[i]))
1053 lgr->lnk[i].link_is_asym = false;
1054 if (lgr->type == new_type)
1056 lgr->type = new_type;
1058 switch (lgr->type) {
1062 case SMC_LGR_SINGLE:
1063 lgr_type = "SINGLE";
1065 case SMC_LGR_SYMMETRIC:
1066 lgr_type = "SYMMETRIC";
1068 case SMC_LGR_ASYMMETRIC_PEER:
1069 lgr_type = "ASYMMETRIC_PEER";
1071 case SMC_LGR_ASYMMETRIC_LOCAL:
1072 lgr_type = "ASYMMETRIC_LOCAL";
1075 pr_warn_ratelimited("smc: SMC-R lg %*phN state changed: "
1076 "%s, pnetid %.16s\n", SMC_LGR_ID_SIZE, &lgr->id,
1077 lgr_type, lgr->pnet_id);
1080 /* set new lgr type and tag a link as asymmetric */
1081 void smcr_lgr_set_type_asym(struct smc_link_group *lgr,
1082 enum smc_lgr_type new_type, int asym_lnk_idx)
1084 smcr_lgr_set_type(lgr, new_type);
1085 lgr->lnk[asym_lnk_idx].link_is_asym = true;
1088 /* abort connection, abort_work scheduled from tasklet context */
1089 static void smc_conn_abort_work(struct work_struct *work)
1091 struct smc_connection *conn = container_of(work,
1092 struct smc_connection,
1094 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1096 smc_conn_kill(conn, true);
1097 sock_put(&smc->sk); /* sock_hold done by schedulers of abort_work */
1100 void smcr_port_add(struct smc_ib_device *smcibdev, u8 ibport)
1102 struct smc_link_group *lgr, *n;
1104 list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1105 struct smc_link *link;
1107 if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1108 SMC_MAX_PNETID_LEN) ||
1109 lgr->type == SMC_LGR_SYMMETRIC ||
1110 lgr->type == SMC_LGR_ASYMMETRIC_PEER)
1113 /* trigger local add link processing */
1114 link = smc_llc_usable_link(lgr);
1116 smc_llc_add_link_local(link);
1120 /* link is down - switch connections to alternate link,
1121 * must be called under lgr->llc_conf_mutex lock
1123 static void smcr_link_down(struct smc_link *lnk)
1125 struct smc_link_group *lgr = lnk->lgr;
1126 struct smc_link *to_lnk;
1129 if (!lgr || lnk->state == SMC_LNK_UNUSED || list_empty(&lgr->list))
1132 smc_ib_modify_qp_reset(lnk);
1133 to_lnk = smc_switch_conns(lgr, lnk, true);
1134 if (!to_lnk) { /* no backup link available */
1135 smcr_link_clear(lnk, true);
1138 smcr_lgr_set_type(lgr, SMC_LGR_SINGLE);
1139 del_link_id = lnk->link_id;
1141 if (lgr->role == SMC_SERV) {
1142 /* trigger local delete link processing */
1143 smc_llc_srv_delete_link_local(to_lnk, del_link_id);
1145 if (lgr->llc_flow_lcl.type != SMC_LLC_FLOW_NONE) {
1146 /* another llc task is ongoing */
1147 mutex_unlock(&lgr->llc_conf_mutex);
1148 wait_event_timeout(lgr->llc_flow_waiter,
1149 (list_empty(&lgr->list) ||
1150 lgr->llc_flow_lcl.type == SMC_LLC_FLOW_NONE),
1152 mutex_lock(&lgr->llc_conf_mutex);
1154 if (!list_empty(&lgr->list)) {
1155 smc_llc_send_delete_link(to_lnk, del_link_id,
1157 SMC_LLC_DEL_LOST_PATH);
1158 smcr_link_clear(lnk, true);
1160 wake_up(&lgr->llc_flow_waiter); /* wake up next waiter */
1164 /* must be called under lgr->llc_conf_mutex lock */
1165 void smcr_link_down_cond(struct smc_link *lnk)
1167 if (smc_link_downing(&lnk->state))
1168 smcr_link_down(lnk);
1171 /* will get the lgr->llc_conf_mutex lock */
1172 void smcr_link_down_cond_sched(struct smc_link *lnk)
1174 if (smc_link_downing(&lnk->state))
1175 schedule_work(&lnk->link_down_wrk);
1178 void smcr_port_err(struct smc_ib_device *smcibdev, u8 ibport)
1180 struct smc_link_group *lgr, *n;
1183 list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1184 if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1185 SMC_MAX_PNETID_LEN))
1186 continue; /* lgr is not affected */
1187 if (list_empty(&lgr->list))
1189 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1190 struct smc_link *lnk = &lgr->lnk[i];
1192 if (smc_link_usable(lnk) &&
1193 lnk->smcibdev == smcibdev && lnk->ibport == ibport)
1194 smcr_link_down_cond_sched(lnk);
1199 static void smc_link_down_work(struct work_struct *work)
1201 struct smc_link *link = container_of(work, struct smc_link,
1203 struct smc_link_group *lgr = link->lgr;
1205 if (list_empty(&lgr->list))
1207 wake_up_all(&lgr->llc_msg_waiter);
1208 mutex_lock(&lgr->llc_conf_mutex);
1209 smcr_link_down(link);
1210 mutex_unlock(&lgr->llc_conf_mutex);
1213 /* Determine vlan of internal TCP socket.
1214 * @vlan_id: address to store the determined vlan id into
1216 int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini)
1218 struct dst_entry *dst = sk_dst_get(clcsock->sk);
1219 struct net_device *ndev;
1220 int i, nest_lvl, rc = 0;
1233 if (is_vlan_dev(ndev)) {
1234 ini->vlan_id = vlan_dev_vlan_id(ndev);
1239 nest_lvl = ndev->lower_level;
1240 for (i = 0; i < nest_lvl; i++) {
1241 struct list_head *lower = &ndev->adj_list.lower;
1243 if (list_empty(lower))
1245 lower = lower->next;
1246 ndev = (struct net_device *)netdev_lower_get_next(ndev, &lower);
1247 if (is_vlan_dev(ndev)) {
1248 ini->vlan_id = vlan_dev_vlan_id(ndev);
1260 static bool smcr_lgr_match(struct smc_link_group *lgr,
1261 struct smc_clc_msg_local *lcl,
1262 enum smc_lgr_role role, u32 clcqpn)
1266 if (memcmp(lgr->peer_systemid, lcl->id_for_peer, SMC_SYSTEMID_LEN) ||
1270 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1271 if (!smc_link_active(&lgr->lnk[i]))
1273 if ((lgr->role == SMC_SERV || lgr->lnk[i].peer_qpn == clcqpn) &&
1274 !memcmp(lgr->lnk[i].peer_gid, &lcl->gid, SMC_GID_SIZE) &&
1275 !memcmp(lgr->lnk[i].peer_mac, lcl->mac, sizeof(lcl->mac)))
1281 static bool smcd_lgr_match(struct smc_link_group *lgr,
1282 struct smcd_dev *smcismdev, u64 peer_gid)
1284 return lgr->peer_gid == peer_gid && lgr->smcd == smcismdev;
1287 /* create a new SMC connection (and a new link group if necessary) */
1288 int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini)
1290 struct smc_connection *conn = &smc->conn;
1291 struct list_head *lgr_list;
1292 struct smc_link_group *lgr;
1293 enum smc_lgr_role role;
1294 spinlock_t *lgr_lock;
1297 lgr_list = ini->is_smcd ? &ini->ism_dev->lgr_list : &smc_lgr_list.list;
1298 lgr_lock = ini->is_smcd ? &ini->ism_dev->lgr_lock : &smc_lgr_list.lock;
1299 ini->cln_first_contact = SMC_FIRST_CONTACT;
1300 role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
1301 if (role == SMC_CLNT && ini->srv_first_contact)
1302 /* create new link group as well */
1305 /* determine if an existing link group can be reused */
1306 spin_lock_bh(lgr_lock);
1307 list_for_each_entry(lgr, lgr_list, list) {
1308 write_lock_bh(&lgr->conns_lock);
1310 smcd_lgr_match(lgr, ini->ism_dev, ini->ism_gid) :
1311 smcr_lgr_match(lgr, ini->ib_lcl, role, ini->ib_clcqpn)) &&
1313 lgr->vlan_id == ini->vlan_id &&
1314 (role == SMC_CLNT || ini->is_smcd ||
1315 lgr->conns_num < SMC_RMBS_PER_LGR_MAX)) {
1316 /* link group found */
1317 ini->cln_first_contact = SMC_REUSE_CONTACT;
1319 rc = smc_lgr_register_conn(conn, false);
1320 write_unlock_bh(&lgr->conns_lock);
1321 if (!rc && delayed_work_pending(&lgr->free_work))
1322 cancel_delayed_work(&lgr->free_work);
1325 write_unlock_bh(&lgr->conns_lock);
1327 spin_unlock_bh(lgr_lock);
1331 if (role == SMC_CLNT && !ini->srv_first_contact &&
1332 ini->cln_first_contact == SMC_FIRST_CONTACT) {
1333 /* Server reuses a link group, but Client wants to start
1335 * send out_of_sync decline, reason synchr. error
1337 return SMC_CLC_DECL_SYNCERR;
1341 if (ini->cln_first_contact == SMC_FIRST_CONTACT) {
1342 rc = smc_lgr_create(smc, ini);
1346 write_lock_bh(&lgr->conns_lock);
1347 rc = smc_lgr_register_conn(conn, true);
1348 write_unlock_bh(&lgr->conns_lock);
1352 conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
1353 conn->local_tx_ctrl.len = SMC_WR_TX_SIZE;
1354 conn->urg_state = SMC_URG_READ;
1355 INIT_WORK(&smc->conn.abort_work, smc_conn_abort_work);
1357 conn->rx_off = sizeof(struct smcd_cdc_msg);
1358 smcd_cdc_rx_init(conn); /* init tasklet for this conn */
1362 #ifndef KERNEL_HAS_ATOMIC64
1363 spin_lock_init(&conn->acurs_lock);
1370 /* convert the RMB size into the compressed notation - minimum 16K.
1371 * In contrast to plain ilog2, this rounds towards the next power of 2,
1372 * so the socket application gets at least its desired sndbuf / rcvbuf size.
1374 static u8 smc_compress_bufsize(int size)
1378 if (size <= SMC_BUF_MIN_SIZE)
1381 size = (size - 1) >> 14;
1382 compressed = ilog2(size) + 1;
1383 if (compressed >= SMC_RMBE_SIZES)
1384 compressed = SMC_RMBE_SIZES - 1;
1388 /* convert the RMB size from compressed notation into integer */
1389 int smc_uncompress_bufsize(u8 compressed)
1393 size = 0x00000001 << (((int)compressed) + 14);
1397 /* try to reuse a sndbuf or rmb description slot for a certain
1398 * buffer size; if not available, return NULL
1400 static struct smc_buf_desc *smc_buf_get_slot(int compressed_bufsize,
1402 struct list_head *buf_list)
1404 struct smc_buf_desc *buf_slot;
1407 list_for_each_entry(buf_slot, buf_list, list) {
1408 if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
1417 /* one of the conditions for announcing a receiver's current window size is
1418 * that it "results in a minimum increase in the window size of 10% of the
1419 * receive buffer space" [RFC7609]
1421 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
1423 return min_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
1426 /* map an rmb buf to a link */
1427 static int smcr_buf_map_link(struct smc_buf_desc *buf_desc, bool is_rmb,
1428 struct smc_link *lnk)
1432 if (buf_desc->is_map_ib[lnk->link_idx])
1435 rc = sg_alloc_table(&buf_desc->sgt[lnk->link_idx], 1, GFP_KERNEL);
1438 sg_set_buf(buf_desc->sgt[lnk->link_idx].sgl,
1439 buf_desc->cpu_addr, buf_desc->len);
1441 /* map sg table to DMA address */
1442 rc = smc_ib_buf_map_sg(lnk, buf_desc,
1443 is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
1444 /* SMC protocol depends on mapping to one DMA address only */
1450 /* create a new memory region for the RMB */
1452 rc = smc_ib_get_memory_region(lnk->roce_pd,
1453 IB_ACCESS_REMOTE_WRITE |
1454 IB_ACCESS_LOCAL_WRITE,
1455 buf_desc, lnk->link_idx);
1458 smc_ib_sync_sg_for_device(lnk, buf_desc, DMA_FROM_DEVICE);
1460 buf_desc->is_map_ib[lnk->link_idx] = true;
1464 smc_ib_buf_unmap_sg(lnk, buf_desc,
1465 is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
1467 sg_free_table(&buf_desc->sgt[lnk->link_idx]);
1471 /* register a new rmb on IB device,
1472 * must be called under lgr->llc_conf_mutex lock
1474 int smcr_link_reg_rmb(struct smc_link *link, struct smc_buf_desc *rmb_desc)
1476 if (list_empty(&link->lgr->list))
1478 if (!rmb_desc->is_reg_mr[link->link_idx]) {
1479 /* register memory region for new rmb */
1480 if (smc_wr_reg_send(link, rmb_desc->mr_rx[link->link_idx])) {
1481 rmb_desc->is_reg_err = true;
1484 rmb_desc->is_reg_mr[link->link_idx] = true;
1489 static int _smcr_buf_map_lgr(struct smc_link *lnk, struct mutex *lock,
1490 struct list_head *lst, bool is_rmb)
1492 struct smc_buf_desc *buf_desc, *bf;
1496 list_for_each_entry_safe(buf_desc, bf, lst, list) {
1497 if (!buf_desc->used)
1499 rc = smcr_buf_map_link(buf_desc, is_rmb, lnk);
1508 /* map all used buffers of lgr for a new link */
1509 int smcr_buf_map_lgr(struct smc_link *lnk)
1511 struct smc_link_group *lgr = lnk->lgr;
1514 for (i = 0; i < SMC_RMBE_SIZES; i++) {
1515 rc = _smcr_buf_map_lgr(lnk, &lgr->rmbs_lock,
1516 &lgr->rmbs[i], true);
1519 rc = _smcr_buf_map_lgr(lnk, &lgr->sndbufs_lock,
1520 &lgr->sndbufs[i], false);
1527 /* register all used buffers of lgr for a new link,
1528 * must be called under lgr->llc_conf_mutex lock
1530 int smcr_buf_reg_lgr(struct smc_link *lnk)
1532 struct smc_link_group *lgr = lnk->lgr;
1533 struct smc_buf_desc *buf_desc, *bf;
1536 mutex_lock(&lgr->rmbs_lock);
1537 for (i = 0; i < SMC_RMBE_SIZES; i++) {
1538 list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list) {
1539 if (!buf_desc->used)
1541 rc = smcr_link_reg_rmb(lnk, buf_desc);
1547 mutex_unlock(&lgr->rmbs_lock);
1551 static struct smc_buf_desc *smcr_new_buf_create(struct smc_link_group *lgr,
1552 bool is_rmb, int bufsize)
1554 struct smc_buf_desc *buf_desc;
1556 /* try to alloc a new buffer */
1557 buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
1559 return ERR_PTR(-ENOMEM);
1561 buf_desc->order = get_order(bufsize);
1562 buf_desc->pages = alloc_pages(GFP_KERNEL | __GFP_NOWARN |
1563 __GFP_NOMEMALLOC | __GFP_COMP |
1564 __GFP_NORETRY | __GFP_ZERO,
1566 if (!buf_desc->pages) {
1568 return ERR_PTR(-EAGAIN);
1570 buf_desc->cpu_addr = (void *)page_address(buf_desc->pages);
1571 buf_desc->len = bufsize;
1575 /* map buf_desc on all usable links,
1576 * unused buffers stay mapped as long as the link is up
1578 static int smcr_buf_map_usable_links(struct smc_link_group *lgr,
1579 struct smc_buf_desc *buf_desc, bool is_rmb)
1583 /* protect against parallel link reconfiguration */
1584 mutex_lock(&lgr->llc_conf_mutex);
1585 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1586 struct smc_link *lnk = &lgr->lnk[i];
1588 if (!smc_link_usable(lnk))
1590 if (smcr_buf_map_link(buf_desc, is_rmb, lnk)) {
1596 mutex_unlock(&lgr->llc_conf_mutex);
1600 #define SMCD_DMBE_SIZES 7 /* 0 -> 16KB, 1 -> 32KB, .. 6 -> 1MB */
1602 static struct smc_buf_desc *smcd_new_buf_create(struct smc_link_group *lgr,
1603 bool is_dmb, int bufsize)
1605 struct smc_buf_desc *buf_desc;
1608 if (smc_compress_bufsize(bufsize) > SMCD_DMBE_SIZES)
1609 return ERR_PTR(-EAGAIN);
1611 /* try to alloc a new DMB */
1612 buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
1614 return ERR_PTR(-ENOMEM);
1616 rc = smc_ism_register_dmb(lgr, bufsize, buf_desc);
1619 return (rc == -ENOMEM) ? ERR_PTR(-EAGAIN) : ERR_PTR(rc);
1621 buf_desc->pages = virt_to_page(buf_desc->cpu_addr);
1622 /* CDC header stored in buf. So, pretend it was smaller */
1623 buf_desc->len = bufsize - sizeof(struct smcd_cdc_msg);
1625 buf_desc->cpu_addr = kzalloc(bufsize, GFP_KERNEL |
1626 __GFP_NOWARN | __GFP_NORETRY |
1628 if (!buf_desc->cpu_addr) {
1630 return ERR_PTR(-EAGAIN);
1632 buf_desc->len = bufsize;
1637 static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
1639 struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
1640 struct smc_connection *conn = &smc->conn;
1641 struct smc_link_group *lgr = conn->lgr;
1642 struct list_head *buf_list;
1643 int bufsize, bufsize_short;
1644 struct mutex *lock; /* lock buffer list */
1648 /* use socket recv buffer size (w/o overhead) as start value */
1649 sk_buf_size = smc->sk.sk_rcvbuf / 2;
1651 /* use socket send buffer size (w/o overhead) as start value */
1652 sk_buf_size = smc->sk.sk_sndbuf / 2;
1654 for (bufsize_short = smc_compress_bufsize(sk_buf_size);
1655 bufsize_short >= 0; bufsize_short--) {
1658 lock = &lgr->rmbs_lock;
1659 buf_list = &lgr->rmbs[bufsize_short];
1661 lock = &lgr->sndbufs_lock;
1662 buf_list = &lgr->sndbufs[bufsize_short];
1664 bufsize = smc_uncompress_bufsize(bufsize_short);
1665 if ((1 << get_order(bufsize)) > SG_MAX_SINGLE_ALLOC)
1668 /* check for reusable slot in the link group */
1669 buf_desc = smc_buf_get_slot(bufsize_short, lock, buf_list);
1671 memset(buf_desc->cpu_addr, 0, bufsize);
1672 break; /* found reusable slot */
1676 buf_desc = smcd_new_buf_create(lgr, is_rmb, bufsize);
1678 buf_desc = smcr_new_buf_create(lgr, is_rmb, bufsize);
1680 if (PTR_ERR(buf_desc) == -ENOMEM)
1682 if (IS_ERR(buf_desc))
1687 list_add(&buf_desc->list, buf_list);
1692 if (IS_ERR(buf_desc))
1693 return PTR_ERR(buf_desc);
1696 if (smcr_buf_map_usable_links(lgr, buf_desc, is_rmb)) {
1697 smcr_buf_unuse(buf_desc, lgr);
1703 conn->rmb_desc = buf_desc;
1704 conn->rmbe_size_short = bufsize_short;
1705 smc->sk.sk_rcvbuf = bufsize * 2;
1706 atomic_set(&conn->bytes_to_rcv, 0);
1707 conn->rmbe_update_limit =
1708 smc_rmb_wnd_update_limit(buf_desc->len);
1710 smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */
1712 conn->sndbuf_desc = buf_desc;
1713 smc->sk.sk_sndbuf = bufsize * 2;
1714 atomic_set(&conn->sndbuf_space, bufsize);
1719 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn)
1721 if (!conn->lgr || conn->lgr->is_smcd || !smc_link_active(conn->lnk))
1723 smc_ib_sync_sg_for_cpu(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE);
1726 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
1728 if (!conn->lgr || conn->lgr->is_smcd || !smc_link_active(conn->lnk))
1730 smc_ib_sync_sg_for_device(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE);
1733 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
1737 if (!conn->lgr || conn->lgr->is_smcd)
1739 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1740 if (!smc_link_active(&conn->lgr->lnk[i]))
1742 smc_ib_sync_sg_for_cpu(&conn->lgr->lnk[i], conn->rmb_desc,
1747 void smc_rmb_sync_sg_for_device(struct smc_connection *conn)
1751 if (!conn->lgr || conn->lgr->is_smcd)
1753 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1754 if (!smc_link_active(&conn->lgr->lnk[i]))
1756 smc_ib_sync_sg_for_device(&conn->lgr->lnk[i], conn->rmb_desc,
1761 /* create the send and receive buffer for an SMC socket;
1762 * receive buffers are called RMBs;
1763 * (even though the SMC protocol allows more than one RMB-element per RMB,
1764 * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
1765 * extra RMB for every connection in a link group
1767 int smc_buf_create(struct smc_sock *smc, bool is_smcd)
1771 /* create send buffer */
1772 rc = __smc_buf_create(smc, is_smcd, false);
1776 rc = __smc_buf_create(smc, is_smcd, true);
1778 mutex_lock(&smc->conn.lgr->sndbufs_lock);
1779 list_del(&smc->conn.sndbuf_desc->list);
1780 mutex_unlock(&smc->conn.lgr->sndbufs_lock);
1781 smc_buf_free(smc->conn.lgr, false, smc->conn.sndbuf_desc);
1782 smc->conn.sndbuf_desc = NULL;
1787 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
1791 for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
1792 if (!test_and_set_bit(i, lgr->rtokens_used_mask))
1798 static int smc_rtoken_find_by_link(struct smc_link_group *lgr, int lnk_idx,
1803 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1804 if (test_bit(i, lgr->rtokens_used_mask) &&
1805 lgr->rtokens[i][lnk_idx].rkey == rkey)
1811 /* set rtoken for a new link to an existing rmb */
1812 void smc_rtoken_set(struct smc_link_group *lgr, int link_idx, int link_idx_new,
1813 __be32 nw_rkey_known, __be64 nw_vaddr, __be32 nw_rkey)
1817 rtok_idx = smc_rtoken_find_by_link(lgr, link_idx, ntohl(nw_rkey_known));
1818 if (rtok_idx == -ENOENT)
1820 lgr->rtokens[rtok_idx][link_idx_new].rkey = ntohl(nw_rkey);
1821 lgr->rtokens[rtok_idx][link_idx_new].dma_addr = be64_to_cpu(nw_vaddr);
1824 /* set rtoken for a new link whose link_id is given */
1825 void smc_rtoken_set2(struct smc_link_group *lgr, int rtok_idx, int link_id,
1826 __be64 nw_vaddr, __be32 nw_rkey)
1828 u64 dma_addr = be64_to_cpu(nw_vaddr);
1829 u32 rkey = ntohl(nw_rkey);
1833 for (link_idx = 0; link_idx < SMC_LINKS_PER_LGR_MAX; link_idx++) {
1834 if (lgr->lnk[link_idx].link_id == link_id) {
1841 lgr->rtokens[rtok_idx][link_idx].rkey = rkey;
1842 lgr->rtokens[rtok_idx][link_idx].dma_addr = dma_addr;
1845 /* add a new rtoken from peer */
1846 int smc_rtoken_add(struct smc_link *lnk, __be64 nw_vaddr, __be32 nw_rkey)
1848 struct smc_link_group *lgr = smc_get_lgr(lnk);
1849 u64 dma_addr = be64_to_cpu(nw_vaddr);
1850 u32 rkey = ntohl(nw_rkey);
1853 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1854 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
1855 lgr->rtokens[i][lnk->link_idx].dma_addr == dma_addr &&
1856 test_bit(i, lgr->rtokens_used_mask)) {
1857 /* already in list */
1861 i = smc_rmb_reserve_rtoken_idx(lgr);
1864 lgr->rtokens[i][lnk->link_idx].rkey = rkey;
1865 lgr->rtokens[i][lnk->link_idx].dma_addr = dma_addr;
1869 /* delete an rtoken from all links */
1870 int smc_rtoken_delete(struct smc_link *lnk, __be32 nw_rkey)
1872 struct smc_link_group *lgr = smc_get_lgr(lnk);
1873 u32 rkey = ntohl(nw_rkey);
1876 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1877 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
1878 test_bit(i, lgr->rtokens_used_mask)) {
1879 for (j = 0; j < SMC_LINKS_PER_LGR_MAX; j++) {
1880 lgr->rtokens[i][j].rkey = 0;
1881 lgr->rtokens[i][j].dma_addr = 0;
1883 clear_bit(i, lgr->rtokens_used_mask);
1890 /* save rkey and dma_addr received from peer during clc handshake */
1891 int smc_rmb_rtoken_handling(struct smc_connection *conn,
1892 struct smc_link *lnk,
1893 struct smc_clc_msg_accept_confirm *clc)
1895 conn->rtoken_idx = smc_rtoken_add(lnk, clc->rmb_dma_addr,
1897 if (conn->rtoken_idx < 0)
1898 return conn->rtoken_idx;
1902 static void smc_core_going_away(void)
1904 struct smc_ib_device *smcibdev;
1905 struct smcd_dev *smcd;
1907 mutex_lock(&smc_ib_devices.mutex);
1908 list_for_each_entry(smcibdev, &smc_ib_devices.list, list) {
1911 for (i = 0; i < SMC_MAX_PORTS; i++)
1912 set_bit(i, smcibdev->ports_going_away);
1914 mutex_unlock(&smc_ib_devices.mutex);
1916 mutex_lock(&smcd_dev_list.mutex);
1917 list_for_each_entry(smcd, &smcd_dev_list.list, list) {
1918 smcd->going_away = 1;
1920 mutex_unlock(&smcd_dev_list.mutex);
1923 /* Clean up all SMC link groups */
1924 static void smc_lgrs_shutdown(void)
1926 struct smcd_dev *smcd;
1928 smc_core_going_away();
1930 smc_smcr_terminate_all(NULL);
1932 mutex_lock(&smcd_dev_list.mutex);
1933 list_for_each_entry(smcd, &smcd_dev_list.list, list)
1934 smc_smcd_terminate_all(smcd);
1935 mutex_unlock(&smcd_dev_list.mutex);
1938 static int smc_core_reboot_event(struct notifier_block *this,
1939 unsigned long event, void *ptr)
1941 smc_lgrs_shutdown();
1942 smc_ib_unregister_client();
1946 static struct notifier_block smc_reboot_notifier = {
1947 .notifier_call = smc_core_reboot_event,
1950 int __init smc_core_init(void)
1952 return register_reboot_notifier(&smc_reboot_notifier);
1955 /* Called (from smc_exit) when module is removed */
1956 void smc_core_exit(void)
1958 unregister_reboot_notifier(&smc_reboot_notifier);
1959 smc_lgrs_shutdown();