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>
18 #include <rdma/ib_verbs.h>
19 #include <rdma/ib_cache.h>
28 #include "smc_close.h"
31 #define SMC_LGR_NUM_INCR 256
32 #define SMC_LGR_FREE_DELAY_SERV (600 * HZ)
33 #define SMC_LGR_FREE_DELAY_CLNT (SMC_LGR_FREE_DELAY_SERV + 10 * HZ)
34 #define SMC_LGR_FREE_DELAY_FAST (8 * HZ)
36 static struct smc_lgr_list smc_lgr_list = { /* established link groups */
37 .lock = __SPIN_LOCK_UNLOCKED(smc_lgr_list.lock),
38 .list = LIST_HEAD_INIT(smc_lgr_list.list),
42 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
43 struct smc_buf_desc *buf_desc);
45 static void smc_lgr_schedule_free_work(struct smc_link_group *lgr)
47 /* client link group creation always follows the server link group
48 * creation. For client use a somewhat higher removal delay time,
49 * otherwise there is a risk of out-of-sync link groups.
51 mod_delayed_work(system_wq, &lgr->free_work,
52 (!lgr->is_smcd && lgr->role == SMC_CLNT) ?
53 SMC_LGR_FREE_DELAY_CLNT : SMC_LGR_FREE_DELAY_SERV);
56 void smc_lgr_schedule_free_work_fast(struct smc_link_group *lgr)
58 mod_delayed_work(system_wq, &lgr->free_work, SMC_LGR_FREE_DELAY_FAST);
61 /* Register connection's alert token in our lookup structure.
62 * To use rbtrees we have to implement our own insert core.
63 * Requires @conns_lock
64 * @smc connection to register
65 * Returns 0 on success, != otherwise.
67 static void smc_lgr_add_alert_token(struct smc_connection *conn)
69 struct rb_node **link, *parent = NULL;
70 u32 token = conn->alert_token_local;
72 link = &conn->lgr->conns_all.rb_node;
74 struct smc_connection *cur = rb_entry(*link,
75 struct smc_connection, alert_node);
78 if (cur->alert_token_local > token)
79 link = &parent->rb_left;
81 link = &parent->rb_right;
83 /* Put the new node there */
84 rb_link_node(&conn->alert_node, parent, link);
85 rb_insert_color(&conn->alert_node, &conn->lgr->conns_all);
88 /* Register connection in link group by assigning an alert token
89 * registered in a search tree.
90 * Requires @conns_lock
91 * Note that '0' is a reserved value and not assigned.
93 static void smc_lgr_register_conn(struct smc_connection *conn)
95 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
96 static atomic_t nexttoken = ATOMIC_INIT(0);
98 /* find a new alert_token_local value not yet used by some connection
101 sock_hold(&smc->sk); /* sock_put in smc_lgr_unregister_conn() */
102 while (!conn->alert_token_local) {
103 conn->alert_token_local = atomic_inc_return(&nexttoken);
104 if (smc_lgr_find_conn(conn->alert_token_local, conn->lgr))
105 conn->alert_token_local = 0;
107 smc_lgr_add_alert_token(conn);
108 conn->lgr->conns_num++;
111 /* Unregister connection and reset the alert token of the given connection<
113 static void __smc_lgr_unregister_conn(struct smc_connection *conn)
115 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
116 struct smc_link_group *lgr = conn->lgr;
118 rb_erase(&conn->alert_node, &lgr->conns_all);
120 conn->alert_token_local = 0;
122 sock_put(&smc->sk); /* sock_hold in smc_lgr_register_conn() */
125 /* Unregister connection from lgr
127 static void smc_lgr_unregister_conn(struct smc_connection *conn)
129 struct smc_link_group *lgr = conn->lgr;
133 write_lock_bh(&lgr->conns_lock);
134 if (conn->alert_token_local) {
135 __smc_lgr_unregister_conn(conn);
137 write_unlock_bh(&lgr->conns_lock);
140 /* Send delete link, either as client to request the initiation
141 * of the DELETE LINK sequence from server; or as server to
142 * initiate the delete processing. See smc_llc_rx_delete_link().
144 static int smc_link_send_delete(struct smc_link *lnk)
146 if (lnk->state == SMC_LNK_ACTIVE &&
147 !smc_llc_send_delete_link(lnk, SMC_LLC_REQ, true)) {
148 smc_llc_link_deleting(lnk);
154 static void smc_lgr_free_work(struct work_struct *work)
156 struct smc_link_group *lgr = container_of(to_delayed_work(work),
157 struct smc_link_group,
161 spin_lock_bh(&smc_lgr_list.lock);
162 if (list_empty(&lgr->list))
164 read_lock_bh(&lgr->conns_lock);
165 conns = RB_EMPTY_ROOT(&lgr->conns_all);
166 read_unlock_bh(&lgr->conns_lock);
167 if (!conns) { /* number of lgr connections is no longer zero */
168 spin_unlock_bh(&smc_lgr_list.lock);
171 list_del_init(&lgr->list); /* remove from smc_lgr_list */
173 spin_unlock_bh(&smc_lgr_list.lock);
175 if (!lgr->is_smcd && !lgr->terminating) {
176 /* try to send del link msg, on error free lgr immediately */
177 if (!smc_link_send_delete(&lgr->lnk[SMC_SINGLE_LINK])) {
178 /* reschedule in case we never receive a response */
179 smc_lgr_schedule_free_work(lgr);
184 if (!delayed_work_pending(&lgr->free_work)) {
185 struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
187 if (!lgr->is_smcd && lnk->state != SMC_LNK_INACTIVE)
188 smc_llc_link_inactive(lnk);
193 /* create a new SMC link group */
194 static int smc_lgr_create(struct smc_sock *smc, bool is_smcd,
195 struct smc_ib_device *smcibdev, u8 ibport,
196 char *peer_systemid, unsigned short vlan_id,
197 struct smcd_dev *smcismdev, u64 peer_gid)
199 struct smc_link_group *lgr;
200 struct smc_link *lnk;
205 if (is_smcd && vlan_id) {
206 rc = smc_ism_get_vlan(smcismdev, vlan_id);
211 lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
216 lgr->is_smcd = is_smcd;
218 lgr->vlan_id = vlan_id;
219 rwlock_init(&lgr->sndbufs_lock);
220 rwlock_init(&lgr->rmbs_lock);
221 rwlock_init(&lgr->conns_lock);
222 for (i = 0; i < SMC_RMBE_SIZES; i++) {
223 INIT_LIST_HEAD(&lgr->sndbufs[i]);
224 INIT_LIST_HEAD(&lgr->rmbs[i]);
226 smc_lgr_list.num += SMC_LGR_NUM_INCR;
227 memcpy(&lgr->id, (u8 *)&smc_lgr_list.num, SMC_LGR_ID_SIZE);
228 INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
229 lgr->conns_all = RB_ROOT;
231 /* SMC-D specific settings */
232 lgr->peer_gid = peer_gid;
233 lgr->smcd = smcismdev;
235 /* SMC-R specific settings */
236 lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
237 memcpy(lgr->peer_systemid, peer_systemid, SMC_SYSTEMID_LEN);
239 lnk = &lgr->lnk[SMC_SINGLE_LINK];
240 /* initialize link */
241 lnk->state = SMC_LNK_ACTIVATING;
242 lnk->link_id = SMC_SINGLE_LINK;
243 lnk->smcibdev = smcibdev;
244 lnk->ibport = ibport;
245 lnk->path_mtu = smcibdev->pattr[ibport - 1].active_mtu;
246 if (!smcibdev->initialized)
247 smc_ib_setup_per_ibdev(smcibdev);
248 get_random_bytes(rndvec, sizeof(rndvec));
249 lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) +
251 rc = smc_ib_determine_gid(lnk->smcibdev, lnk->ibport,
252 vlan_id, lnk->gid, &lnk->sgid_index);
255 rc = smc_llc_link_init(lnk);
258 rc = smc_wr_alloc_link_mem(lnk);
261 rc = smc_ib_create_protection_domain(lnk);
264 rc = smc_ib_create_queue_pair(lnk);
267 rc = smc_wr_create_link(lnk);
272 spin_lock_bh(&smc_lgr_list.lock);
273 list_add(&lgr->list, &smc_lgr_list.list);
274 spin_unlock_bh(&smc_lgr_list.lock);
278 smc_ib_destroy_queue_pair(lnk);
280 smc_ib_dealloc_protection_domain(lnk);
282 smc_wr_free_link_mem(lnk);
284 smc_llc_link_clear(lnk);
291 static void smc_buf_unuse(struct smc_connection *conn,
292 struct smc_link_group *lgr)
294 if (conn->sndbuf_desc)
295 conn->sndbuf_desc->used = 0;
296 if (conn->rmb_desc) {
297 if (!conn->rmb_desc->regerr) {
298 conn->rmb_desc->reused = 1;
299 conn->rmb_desc->used = 0;
301 /* buf registration failed, reuse not possible */
302 write_lock_bh(&lgr->rmbs_lock);
303 list_del(&conn->rmb_desc->list);
304 write_unlock_bh(&lgr->rmbs_lock);
306 smc_buf_free(lgr, true, conn->rmb_desc);
311 /* remove a finished connection from its link group */
312 void smc_conn_free(struct smc_connection *conn)
314 struct smc_link_group *lgr = conn->lgr;
319 smc_ism_unset_conn(conn);
320 tasklet_kill(&conn->rx_tsklet);
322 smc_cdc_tx_dismiss_slots(conn);
324 smc_buf_unuse(conn, lgr); /* allow buffer reuse */
325 smc_lgr_unregister_conn(conn); /* unsets conn->lgr */
328 smc_lgr_schedule_free_work(lgr);
331 static void smc_link_clear(struct smc_link *lnk)
334 smc_llc_link_clear(lnk);
335 smc_ib_modify_qp_reset(lnk);
336 smc_wr_free_link(lnk);
337 smc_ib_destroy_queue_pair(lnk);
338 smc_ib_dealloc_protection_domain(lnk);
339 smc_wr_free_link_mem(lnk);
342 static void smcr_buf_free(struct smc_link_group *lgr, bool is_rmb,
343 struct smc_buf_desc *buf_desc)
345 struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
348 if (buf_desc->mr_rx[SMC_SINGLE_LINK])
349 smc_ib_put_memory_region(
350 buf_desc->mr_rx[SMC_SINGLE_LINK]);
351 smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
354 smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
357 sg_free_table(&buf_desc->sgt[SMC_SINGLE_LINK]);
359 __free_pages(buf_desc->pages, buf_desc->order);
363 static void smcd_buf_free(struct smc_link_group *lgr, bool is_dmb,
364 struct smc_buf_desc *buf_desc)
367 /* restore original buf len */
368 buf_desc->len += sizeof(struct smcd_cdc_msg);
369 smc_ism_unregister_dmb(lgr->smcd, buf_desc);
371 kfree(buf_desc->cpu_addr);
376 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
377 struct smc_buf_desc *buf_desc)
380 smcd_buf_free(lgr, is_rmb, buf_desc);
382 smcr_buf_free(lgr, is_rmb, buf_desc);
385 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
387 struct smc_buf_desc *buf_desc, *bf_desc;
388 struct list_head *buf_list;
391 for (i = 0; i < SMC_RMBE_SIZES; i++) {
393 buf_list = &lgr->rmbs[i];
395 buf_list = &lgr->sndbufs[i];
396 list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
398 list_del(&buf_desc->list);
399 smc_buf_free(lgr, is_rmb, buf_desc);
404 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
406 /* free send buffers */
407 __smc_lgr_free_bufs(lgr, false);
409 __smc_lgr_free_bufs(lgr, true);
412 /* remove a link group */
413 void smc_lgr_free(struct smc_link_group *lgr)
415 smc_lgr_free_bufs(lgr);
417 smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
419 smc_link_clear(&lgr->lnk[SMC_SINGLE_LINK]);
423 void smc_lgr_forget(struct smc_link_group *lgr)
425 spin_lock_bh(&smc_lgr_list.lock);
426 /* do not use this link group for new connections */
427 if (!list_empty(&lgr->list))
428 list_del_init(&lgr->list);
429 spin_unlock_bh(&smc_lgr_list.lock);
432 /* terminate linkgroup abnormally */
433 static void __smc_lgr_terminate(struct smc_link_group *lgr)
435 struct smc_connection *conn;
436 struct smc_sock *smc;
437 struct rb_node *node;
439 if (lgr->terminating)
440 return; /* lgr already terminating */
441 lgr->terminating = 1;
442 if (!list_empty(&lgr->list)) /* forget lgr */
443 list_del_init(&lgr->list);
445 smc_llc_link_inactive(&lgr->lnk[SMC_SINGLE_LINK]);
447 write_lock_bh(&lgr->conns_lock);
448 node = rb_first(&lgr->conns_all);
450 conn = rb_entry(node, struct smc_connection, alert_node);
451 smc = container_of(conn, struct smc_sock, conn);
452 sock_hold(&smc->sk); /* sock_put in close work */
453 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
454 __smc_lgr_unregister_conn(conn);
455 write_unlock_bh(&lgr->conns_lock);
456 if (!schedule_work(&conn->close_work))
458 write_lock_bh(&lgr->conns_lock);
459 node = rb_first(&lgr->conns_all);
461 write_unlock_bh(&lgr->conns_lock);
463 wake_up(&lgr->lnk[SMC_SINGLE_LINK].wr_reg_wait);
464 smc_lgr_schedule_free_work(lgr);
467 void smc_lgr_terminate(struct smc_link_group *lgr)
469 spin_lock_bh(&smc_lgr_list.lock);
470 __smc_lgr_terminate(lgr);
471 spin_unlock_bh(&smc_lgr_list.lock);
474 /* Called when IB port is terminated */
475 void smc_port_terminate(struct smc_ib_device *smcibdev, u8 ibport)
477 struct smc_link_group *lgr, *l;
479 spin_lock_bh(&smc_lgr_list.lock);
480 list_for_each_entry_safe(lgr, l, &smc_lgr_list.list, list) {
482 lgr->lnk[SMC_SINGLE_LINK].smcibdev == smcibdev &&
483 lgr->lnk[SMC_SINGLE_LINK].ibport == ibport)
484 __smc_lgr_terminate(lgr);
486 spin_unlock_bh(&smc_lgr_list.lock);
489 /* Called when SMC-D device is terminated or peer is lost */
490 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid)
492 struct smc_link_group *lgr, *l;
493 LIST_HEAD(lgr_free_list);
495 /* run common cleanup function and build free list */
496 spin_lock_bh(&smc_lgr_list.lock);
497 list_for_each_entry_safe(lgr, l, &smc_lgr_list.list, list) {
498 if (lgr->is_smcd && lgr->smcd == dev &&
499 (!peer_gid || lgr->peer_gid == peer_gid) &&
500 !list_empty(&lgr->list)) {
501 __smc_lgr_terminate(lgr);
502 list_move(&lgr->list, &lgr_free_list);
505 spin_unlock_bh(&smc_lgr_list.lock);
507 /* cancel the regular free workers and actually free lgrs */
508 list_for_each_entry_safe(lgr, l, &lgr_free_list, list) {
509 list_del_init(&lgr->list);
510 cancel_delayed_work_sync(&lgr->free_work);
515 /* Determine vlan of internal TCP socket.
516 * @vlan_id: address to store the determined vlan id into
518 int smc_vlan_by_tcpsk(struct socket *clcsock, unsigned short *vlan_id)
520 struct dst_entry *dst = sk_dst_get(clcsock->sk);
521 struct net_device *ndev;
522 int i, nest_lvl, rc = 0;
535 if (is_vlan_dev(ndev)) {
536 *vlan_id = vlan_dev_vlan_id(ndev);
541 nest_lvl = dev_get_nest_level(ndev);
542 for (i = 0; i < nest_lvl; i++) {
543 struct list_head *lower = &ndev->adj_list.lower;
545 if (list_empty(lower))
548 ndev = (struct net_device *)netdev_lower_get_next(ndev, &lower);
549 if (is_vlan_dev(ndev)) {
550 *vlan_id = vlan_dev_vlan_id(ndev);
562 static bool smcr_lgr_match(struct smc_link_group *lgr,
563 struct smc_clc_msg_local *lcl,
564 enum smc_lgr_role role)
566 return !memcmp(lgr->peer_systemid, lcl->id_for_peer,
568 !memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_gid, &lcl->gid,
570 !memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_mac, lcl->mac,
575 static bool smcd_lgr_match(struct smc_link_group *lgr,
576 struct smcd_dev *smcismdev, u64 peer_gid)
578 return lgr->peer_gid == peer_gid && lgr->smcd == smcismdev;
581 /* create a new SMC connection (and a new link group if necessary) */
582 int smc_conn_create(struct smc_sock *smc, bool is_smcd, int srv_first_contact,
583 struct smc_ib_device *smcibdev, u8 ibport,
584 struct smc_clc_msg_local *lcl, struct smcd_dev *smcd,
587 struct smc_connection *conn = &smc->conn;
588 int local_contact = SMC_FIRST_CONTACT;
589 struct smc_link_group *lgr;
590 unsigned short vlan_id;
591 enum smc_lgr_role role;
594 role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
595 rc = smc_vlan_by_tcpsk(smc->clcsock, &vlan_id);
599 if ((role == SMC_CLNT) && srv_first_contact)
600 /* create new link group as well */
603 /* determine if an existing link group can be reused */
604 spin_lock_bh(&smc_lgr_list.lock);
605 list_for_each_entry(lgr, &smc_lgr_list.list, list) {
606 write_lock_bh(&lgr->conns_lock);
607 if ((is_smcd ? smcd_lgr_match(lgr, smcd, peer_gid) :
608 smcr_lgr_match(lgr, lcl, role)) &&
610 lgr->vlan_id == vlan_id &&
612 (lgr->conns_num < SMC_RMBS_PER_LGR_MAX &&
613 !bitmap_full(lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX)))) {
614 /* link group found */
615 local_contact = SMC_REUSE_CONTACT;
617 smc_lgr_register_conn(conn); /* add smc conn to lgr */
618 if (delayed_work_pending(&lgr->free_work))
619 cancel_delayed_work(&lgr->free_work);
620 write_unlock_bh(&lgr->conns_lock);
623 write_unlock_bh(&lgr->conns_lock);
625 spin_unlock_bh(&smc_lgr_list.lock);
627 if (role == SMC_CLNT && !srv_first_contact &&
628 (local_contact == SMC_FIRST_CONTACT)) {
629 /* Server reuses a link group, but Client wants to start
631 * send out_of_sync decline, reason synchr. error
637 if (local_contact == SMC_FIRST_CONTACT) {
638 rc = smc_lgr_create(smc, is_smcd, smcibdev, ibport,
639 lcl->id_for_peer, vlan_id, smcd, peer_gid);
642 smc_lgr_register_conn(conn); /* add smc conn to lgr */
644 conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
645 conn->local_tx_ctrl.len = SMC_WR_TX_SIZE;
646 conn->urg_state = SMC_URG_READ;
648 conn->rx_off = sizeof(struct smcd_cdc_msg);
649 smcd_cdc_rx_init(conn); /* init tasklet for this conn */
651 #ifndef KERNEL_HAS_ATOMIC64
652 spin_lock_init(&conn->acurs_lock);
656 return rc ? rc : local_contact;
659 /* convert the RMB size into the compressed notation - minimum 16K.
660 * In contrast to plain ilog2, this rounds towards the next power of 2,
661 * so the socket application gets at least its desired sndbuf / rcvbuf size.
663 static u8 smc_compress_bufsize(int size)
667 if (size <= SMC_BUF_MIN_SIZE)
670 size = (size - 1) >> 14;
671 compressed = ilog2(size) + 1;
672 if (compressed >= SMC_RMBE_SIZES)
673 compressed = SMC_RMBE_SIZES - 1;
677 /* convert the RMB size from compressed notation into integer */
678 int smc_uncompress_bufsize(u8 compressed)
682 size = 0x00000001 << (((int)compressed) + 14);
686 /* try to reuse a sndbuf or rmb description slot for a certain
687 * buffer size; if not available, return NULL
689 static struct smc_buf_desc *smc_buf_get_slot(int compressed_bufsize,
691 struct list_head *buf_list)
693 struct smc_buf_desc *buf_slot;
696 list_for_each_entry(buf_slot, buf_list, list) {
697 if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
698 read_unlock_bh(lock);
702 read_unlock_bh(lock);
706 /* one of the conditions for announcing a receiver's current window size is
707 * that it "results in a minimum increase in the window size of 10% of the
708 * receive buffer space" [RFC7609]
710 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
712 return max_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
715 static struct smc_buf_desc *smcr_new_buf_create(struct smc_link_group *lgr,
716 bool is_rmb, int bufsize)
718 struct smc_buf_desc *buf_desc;
719 struct smc_link *lnk;
722 /* try to alloc a new buffer */
723 buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
725 return ERR_PTR(-ENOMEM);
727 buf_desc->order = get_order(bufsize);
728 buf_desc->pages = alloc_pages(GFP_KERNEL | __GFP_NOWARN |
729 __GFP_NOMEMALLOC | __GFP_COMP |
730 __GFP_NORETRY | __GFP_ZERO,
732 if (!buf_desc->pages) {
734 return ERR_PTR(-EAGAIN);
736 buf_desc->cpu_addr = (void *)page_address(buf_desc->pages);
738 /* build the sg table from the pages */
739 lnk = &lgr->lnk[SMC_SINGLE_LINK];
740 rc = sg_alloc_table(&buf_desc->sgt[SMC_SINGLE_LINK], 1,
743 smc_buf_free(lgr, is_rmb, buf_desc);
746 sg_set_buf(buf_desc->sgt[SMC_SINGLE_LINK].sgl,
747 buf_desc->cpu_addr, bufsize);
749 /* map sg table to DMA address */
750 rc = smc_ib_buf_map_sg(lnk->smcibdev, buf_desc,
751 is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
752 /* SMC protocol depends on mapping to one DMA address only */
754 smc_buf_free(lgr, is_rmb, buf_desc);
755 return ERR_PTR(-EAGAIN);
758 /* create a new memory region for the RMB */
760 rc = smc_ib_get_memory_region(lnk->roce_pd,
761 IB_ACCESS_REMOTE_WRITE |
762 IB_ACCESS_LOCAL_WRITE,
765 smc_buf_free(lgr, is_rmb, buf_desc);
770 buf_desc->len = bufsize;
774 #define SMCD_DMBE_SIZES 6 /* 0 -> 16KB, 1 -> 32KB, .. 6 -> 1MB */
776 static struct smc_buf_desc *smcd_new_buf_create(struct smc_link_group *lgr,
777 bool is_dmb, int bufsize)
779 struct smc_buf_desc *buf_desc;
782 if (smc_compress_bufsize(bufsize) > SMCD_DMBE_SIZES)
783 return ERR_PTR(-EAGAIN);
785 /* try to alloc a new DMB */
786 buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
788 return ERR_PTR(-ENOMEM);
790 rc = smc_ism_register_dmb(lgr, bufsize, buf_desc);
793 return ERR_PTR(-EAGAIN);
795 buf_desc->pages = virt_to_page(buf_desc->cpu_addr);
796 /* CDC header stored in buf. So, pretend it was smaller */
797 buf_desc->len = bufsize - sizeof(struct smcd_cdc_msg);
799 buf_desc->cpu_addr = kzalloc(bufsize, GFP_KERNEL |
800 __GFP_NOWARN | __GFP_NORETRY |
802 if (!buf_desc->cpu_addr) {
804 return ERR_PTR(-EAGAIN);
806 buf_desc->len = bufsize;
811 static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
813 struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
814 struct smc_connection *conn = &smc->conn;
815 struct smc_link_group *lgr = conn->lgr;
816 struct list_head *buf_list;
817 int bufsize, bufsize_short;
822 /* use socket recv buffer size (w/o overhead) as start value */
823 sk_buf_size = smc->sk.sk_rcvbuf / 2;
825 /* use socket send buffer size (w/o overhead) as start value */
826 sk_buf_size = smc->sk.sk_sndbuf / 2;
828 for (bufsize_short = smc_compress_bufsize(sk_buf_size);
829 bufsize_short >= 0; bufsize_short--) {
832 lock = &lgr->rmbs_lock;
833 buf_list = &lgr->rmbs[bufsize_short];
835 lock = &lgr->sndbufs_lock;
836 buf_list = &lgr->sndbufs[bufsize_short];
838 bufsize = smc_uncompress_bufsize(bufsize_short);
839 if ((1 << get_order(bufsize)) > SG_MAX_SINGLE_ALLOC)
842 /* check for reusable slot in the link group */
843 buf_desc = smc_buf_get_slot(bufsize_short, lock, buf_list);
845 memset(buf_desc->cpu_addr, 0, bufsize);
846 break; /* found reusable slot */
850 buf_desc = smcd_new_buf_create(lgr, is_rmb, bufsize);
852 buf_desc = smcr_new_buf_create(lgr, is_rmb, bufsize);
854 if (PTR_ERR(buf_desc) == -ENOMEM)
856 if (IS_ERR(buf_desc))
861 list_add(&buf_desc->list, buf_list);
862 write_unlock_bh(lock);
866 if (IS_ERR(buf_desc))
870 conn->rmb_desc = buf_desc;
871 conn->rmbe_size_short = bufsize_short;
872 smc->sk.sk_rcvbuf = bufsize * 2;
873 atomic_set(&conn->bytes_to_rcv, 0);
874 conn->rmbe_update_limit =
875 smc_rmb_wnd_update_limit(buf_desc->len);
877 smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */
879 conn->sndbuf_desc = buf_desc;
880 smc->sk.sk_sndbuf = bufsize * 2;
881 atomic_set(&conn->sndbuf_space, bufsize);
886 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn)
888 struct smc_link_group *lgr = conn->lgr;
890 if (!conn->lgr || conn->lgr->is_smcd)
892 smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
893 conn->sndbuf_desc, DMA_TO_DEVICE);
896 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
898 struct smc_link_group *lgr = conn->lgr;
900 if (!conn->lgr || conn->lgr->is_smcd)
902 smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
903 conn->sndbuf_desc, DMA_TO_DEVICE);
906 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
908 struct smc_link_group *lgr = conn->lgr;
910 if (!conn->lgr || conn->lgr->is_smcd)
912 smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
913 conn->rmb_desc, DMA_FROM_DEVICE);
916 void smc_rmb_sync_sg_for_device(struct smc_connection *conn)
918 struct smc_link_group *lgr = conn->lgr;
920 if (!conn->lgr || conn->lgr->is_smcd)
922 smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
923 conn->rmb_desc, DMA_FROM_DEVICE);
926 /* create the send and receive buffer for an SMC socket;
927 * receive buffers are called RMBs;
928 * (even though the SMC protocol allows more than one RMB-element per RMB,
929 * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
930 * extra RMB for every connection in a link group
932 int smc_buf_create(struct smc_sock *smc, bool is_smcd)
936 /* create send buffer */
937 rc = __smc_buf_create(smc, is_smcd, false);
941 rc = __smc_buf_create(smc, is_smcd, true);
943 smc_buf_free(smc->conn.lgr, false, smc->conn.sndbuf_desc);
947 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
951 for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
952 if (!test_and_set_bit(i, lgr->rtokens_used_mask))
958 /* add a new rtoken from peer */
959 int smc_rtoken_add(struct smc_link_group *lgr, __be64 nw_vaddr, __be32 nw_rkey)
961 u64 dma_addr = be64_to_cpu(nw_vaddr);
962 u32 rkey = ntohl(nw_rkey);
965 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
966 if ((lgr->rtokens[i][SMC_SINGLE_LINK].rkey == rkey) &&
967 (lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr == dma_addr) &&
968 test_bit(i, lgr->rtokens_used_mask)) {
969 /* already in list */
973 i = smc_rmb_reserve_rtoken_idx(lgr);
976 lgr->rtokens[i][SMC_SINGLE_LINK].rkey = rkey;
977 lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr = dma_addr;
981 /* delete an rtoken */
982 int smc_rtoken_delete(struct smc_link_group *lgr, __be32 nw_rkey)
984 u32 rkey = ntohl(nw_rkey);
987 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
988 if (lgr->rtokens[i][SMC_SINGLE_LINK].rkey == rkey &&
989 test_bit(i, lgr->rtokens_used_mask)) {
990 lgr->rtokens[i][SMC_SINGLE_LINK].rkey = 0;
991 lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr = 0;
993 clear_bit(i, lgr->rtokens_used_mask);
1000 /* save rkey and dma_addr received from peer during clc handshake */
1001 int smc_rmb_rtoken_handling(struct smc_connection *conn,
1002 struct smc_clc_msg_accept_confirm *clc)
1004 conn->rtoken_idx = smc_rtoken_add(conn->lgr, clc->rmb_dma_addr,
1006 if (conn->rtoken_idx < 0)
1007 return conn->rtoken_idx;
1011 /* Called (from smc_exit) when module is removed */
1012 void smc_core_exit(void)
1014 struct smc_link_group *lgr, *lg;
1015 LIST_HEAD(lgr_freeing_list);
1017 spin_lock_bh(&smc_lgr_list.lock);
1018 if (!list_empty(&smc_lgr_list.list))
1019 list_splice_init(&smc_lgr_list.list, &lgr_freeing_list);
1020 spin_unlock_bh(&smc_lgr_list.lock);
1021 list_for_each_entry_safe(lgr, lg, &lgr_freeing_list, list) {
1022 list_del_init(&lgr->list);
1023 if (!lgr->is_smcd) {
1024 struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
1026 if (lnk->state == SMC_LNK_ACTIVE)
1027 smc_llc_send_delete_link(lnk, SMC_LLC_REQ,
1029 smc_llc_link_inactive(lnk);
1031 cancel_delayed_work_sync(&lgr->free_work);
1032 smc_lgr_free(lgr); /* free link group */