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;
131 write_lock_bh(&lgr->conns_lock);
132 if (conn->alert_token_local) {
133 __smc_lgr_unregister_conn(conn);
135 write_unlock_bh(&lgr->conns_lock);
138 /* Send delete link, either as client to request the initiation
139 * of the DELETE LINK sequence from server; or as server to
140 * initiate the delete processing. See smc_llc_rx_delete_link().
142 static int smc_link_send_delete(struct smc_link *lnk)
144 if (lnk->state == SMC_LNK_ACTIVE &&
145 !smc_llc_send_delete_link(lnk, SMC_LLC_REQ, true)) {
146 smc_llc_link_deleting(lnk);
152 static void smc_lgr_free(struct smc_link_group *lgr);
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 struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
178 /* try to send del link msg, on error free lgr immediately */
179 if (lnk->state == SMC_LNK_ACTIVE &&
180 !smc_link_send_delete(lnk)) {
181 /* reschedule in case we never receive a response */
182 smc_lgr_schedule_free_work(lgr);
187 if (!delayed_work_pending(&lgr->free_work)) {
188 struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
190 if (!lgr->is_smcd && lnk->state != SMC_LNK_INACTIVE)
191 smc_llc_link_inactive(lnk);
193 smc_ism_signal_shutdown(lgr);
198 /* create a new SMC link group */
199 static int smc_lgr_create(struct smc_sock *smc, bool is_smcd,
200 struct smc_ib_device *smcibdev, u8 ibport,
201 char *peer_systemid, unsigned short vlan_id,
202 struct smcd_dev *smcismdev, u64 peer_gid)
204 struct smc_link_group *lgr;
205 struct smc_link *lnk;
210 if (is_smcd && vlan_id) {
211 rc = smc_ism_get_vlan(smcismdev, vlan_id);
216 lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
221 lgr->is_smcd = is_smcd;
223 lgr->vlan_id = vlan_id;
224 rwlock_init(&lgr->sndbufs_lock);
225 rwlock_init(&lgr->rmbs_lock);
226 rwlock_init(&lgr->conns_lock);
227 for (i = 0; i < SMC_RMBE_SIZES; i++) {
228 INIT_LIST_HEAD(&lgr->sndbufs[i]);
229 INIT_LIST_HEAD(&lgr->rmbs[i]);
231 smc_lgr_list.num += SMC_LGR_NUM_INCR;
232 memcpy(&lgr->id, (u8 *)&smc_lgr_list.num, SMC_LGR_ID_SIZE);
233 INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
234 lgr->conns_all = RB_ROOT;
236 /* SMC-D specific settings */
237 lgr->peer_gid = peer_gid;
238 lgr->smcd = smcismdev;
240 /* SMC-R specific settings */
241 lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
242 memcpy(lgr->peer_systemid, peer_systemid, SMC_SYSTEMID_LEN);
244 lnk = &lgr->lnk[SMC_SINGLE_LINK];
245 /* initialize link */
246 lnk->state = SMC_LNK_ACTIVATING;
247 lnk->link_id = SMC_SINGLE_LINK;
248 lnk->smcibdev = smcibdev;
249 lnk->ibport = ibport;
250 lnk->path_mtu = smcibdev->pattr[ibport - 1].active_mtu;
251 if (!smcibdev->initialized)
252 smc_ib_setup_per_ibdev(smcibdev);
253 get_random_bytes(rndvec, sizeof(rndvec));
254 lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) +
256 rc = smc_ib_determine_gid(lnk->smcibdev, lnk->ibport,
257 vlan_id, lnk->gid, &lnk->sgid_index);
260 rc = smc_llc_link_init(lnk);
263 rc = smc_wr_alloc_link_mem(lnk);
266 rc = smc_ib_create_protection_domain(lnk);
269 rc = smc_ib_create_queue_pair(lnk);
272 rc = smc_wr_create_link(lnk);
277 spin_lock_bh(&smc_lgr_list.lock);
278 list_add(&lgr->list, &smc_lgr_list.list);
279 spin_unlock_bh(&smc_lgr_list.lock);
283 smc_ib_destroy_queue_pair(lnk);
285 smc_ib_dealloc_protection_domain(lnk);
287 smc_wr_free_link_mem(lnk);
289 smc_llc_link_clear(lnk);
296 static void smc_buf_unuse(struct smc_connection *conn,
297 struct smc_link_group *lgr)
299 if (conn->sndbuf_desc)
300 conn->sndbuf_desc->used = 0;
301 if (conn->rmb_desc) {
302 if (!conn->rmb_desc->regerr) {
303 conn->rmb_desc->used = 0;
305 /* unregister rmb with peer */
306 smc_llc_do_delete_rkey(
307 &lgr->lnk[SMC_SINGLE_LINK],
311 /* buf registration failed, reuse not possible */
312 write_lock_bh(&lgr->rmbs_lock);
313 list_del(&conn->rmb_desc->list);
314 write_unlock_bh(&lgr->rmbs_lock);
316 smc_buf_free(lgr, true, conn->rmb_desc);
321 /* remove a finished connection from its link group */
322 void smc_conn_free(struct smc_connection *conn)
324 struct smc_link_group *lgr = conn->lgr;
329 smc_ism_unset_conn(conn);
330 tasklet_kill(&conn->rx_tsklet);
332 smc_cdc_tx_dismiss_slots(conn);
334 smc_lgr_unregister_conn(conn); /* unsets conn->lgr */
335 smc_buf_unuse(conn, lgr); /* allow buffer reuse */
338 smc_lgr_schedule_free_work(lgr);
341 static void smc_link_clear(struct smc_link *lnk)
344 smc_llc_link_clear(lnk);
345 smc_ib_modify_qp_reset(lnk);
346 smc_wr_free_link(lnk);
347 smc_ib_destroy_queue_pair(lnk);
348 smc_ib_dealloc_protection_domain(lnk);
349 smc_wr_free_link_mem(lnk);
352 static void smcr_buf_free(struct smc_link_group *lgr, bool is_rmb,
353 struct smc_buf_desc *buf_desc)
355 struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
358 if (buf_desc->mr_rx[SMC_SINGLE_LINK])
359 smc_ib_put_memory_region(
360 buf_desc->mr_rx[SMC_SINGLE_LINK]);
361 smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
364 smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
367 sg_free_table(&buf_desc->sgt[SMC_SINGLE_LINK]);
369 __free_pages(buf_desc->pages, buf_desc->order);
373 static void smcd_buf_free(struct smc_link_group *lgr, bool is_dmb,
374 struct smc_buf_desc *buf_desc)
377 /* restore original buf len */
378 buf_desc->len += sizeof(struct smcd_cdc_msg);
379 smc_ism_unregister_dmb(lgr->smcd, buf_desc);
381 kfree(buf_desc->cpu_addr);
386 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
387 struct smc_buf_desc *buf_desc)
390 smcd_buf_free(lgr, is_rmb, buf_desc);
392 smcr_buf_free(lgr, is_rmb, buf_desc);
395 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
397 struct smc_buf_desc *buf_desc, *bf_desc;
398 struct list_head *buf_list;
401 for (i = 0; i < SMC_RMBE_SIZES; i++) {
403 buf_list = &lgr->rmbs[i];
405 buf_list = &lgr->sndbufs[i];
406 list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
408 list_del(&buf_desc->list);
409 smc_buf_free(lgr, is_rmb, buf_desc);
414 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
416 /* free send buffers */
417 __smc_lgr_free_bufs(lgr, false);
419 __smc_lgr_free_bufs(lgr, true);
422 /* remove a link group */
423 static void smc_lgr_free(struct smc_link_group *lgr)
425 smc_lgr_free_bufs(lgr);
427 smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
429 smc_link_clear(&lgr->lnk[SMC_SINGLE_LINK]);
433 void smc_lgr_forget(struct smc_link_group *lgr)
435 spin_lock_bh(&smc_lgr_list.lock);
436 /* do not use this link group for new connections */
437 if (!list_empty(&lgr->list))
438 list_del_init(&lgr->list);
439 spin_unlock_bh(&smc_lgr_list.lock);
442 /* terminate linkgroup abnormally */
443 static void __smc_lgr_terminate(struct smc_link_group *lgr)
445 struct smc_connection *conn;
446 struct smc_sock *smc;
447 struct rb_node *node;
449 if (lgr->terminating)
450 return; /* lgr already terminating */
451 lgr->terminating = 1;
452 if (!list_empty(&lgr->list)) /* forget lgr */
453 list_del_init(&lgr->list);
455 smc_llc_link_inactive(&lgr->lnk[SMC_SINGLE_LINK]);
457 write_lock_bh(&lgr->conns_lock);
458 node = rb_first(&lgr->conns_all);
460 conn = rb_entry(node, struct smc_connection, alert_node);
461 smc = container_of(conn, struct smc_sock, conn);
462 sock_hold(&smc->sk); /* sock_put in close work */
463 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
464 __smc_lgr_unregister_conn(conn);
465 write_unlock_bh(&lgr->conns_lock);
466 if (!schedule_work(&conn->close_work))
468 write_lock_bh(&lgr->conns_lock);
469 node = rb_first(&lgr->conns_all);
471 write_unlock_bh(&lgr->conns_lock);
473 wake_up(&lgr->lnk[SMC_SINGLE_LINK].wr_reg_wait);
474 smc_lgr_schedule_free_work(lgr);
477 void smc_lgr_terminate(struct smc_link_group *lgr)
479 spin_lock_bh(&smc_lgr_list.lock);
480 __smc_lgr_terminate(lgr);
481 spin_unlock_bh(&smc_lgr_list.lock);
484 /* Called when IB port is terminated */
485 void smc_port_terminate(struct smc_ib_device *smcibdev, u8 ibport)
487 struct smc_link_group *lgr, *l;
489 spin_lock_bh(&smc_lgr_list.lock);
490 list_for_each_entry_safe(lgr, l, &smc_lgr_list.list, list) {
492 lgr->lnk[SMC_SINGLE_LINK].smcibdev == smcibdev &&
493 lgr->lnk[SMC_SINGLE_LINK].ibport == ibport)
494 __smc_lgr_terminate(lgr);
496 spin_unlock_bh(&smc_lgr_list.lock);
499 /* Called when SMC-D device is terminated or peer is lost */
500 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid, unsigned short vlan)
502 struct smc_link_group *lgr, *l;
503 LIST_HEAD(lgr_free_list);
505 /* run common cleanup function and build free list */
506 spin_lock_bh(&smc_lgr_list.lock);
507 list_for_each_entry_safe(lgr, l, &smc_lgr_list.list, list) {
508 if (lgr->is_smcd && lgr->smcd == dev &&
509 (!peer_gid || lgr->peer_gid == peer_gid) &&
510 (vlan == VLAN_VID_MASK || lgr->vlan_id == vlan)) {
511 __smc_lgr_terminate(lgr);
512 list_move(&lgr->list, &lgr_free_list);
515 spin_unlock_bh(&smc_lgr_list.lock);
517 /* cancel the regular free workers and actually free lgrs */
518 list_for_each_entry_safe(lgr, l, &lgr_free_list, list) {
519 list_del_init(&lgr->list);
520 cancel_delayed_work_sync(&lgr->free_work);
521 if (!peer_gid && vlan == VLAN_VID_MASK) /* dev terminated? */
522 smc_ism_signal_shutdown(lgr);
527 /* Determine vlan of internal TCP socket.
528 * @vlan_id: address to store the determined vlan id into
530 int smc_vlan_by_tcpsk(struct socket *clcsock, unsigned short *vlan_id)
532 struct dst_entry *dst = sk_dst_get(clcsock->sk);
533 struct net_device *ndev;
534 int i, nest_lvl, rc = 0;
547 if (is_vlan_dev(ndev)) {
548 *vlan_id = vlan_dev_vlan_id(ndev);
553 nest_lvl = dev_get_nest_level(ndev);
554 for (i = 0; i < nest_lvl; i++) {
555 struct list_head *lower = &ndev->adj_list.lower;
557 if (list_empty(lower))
560 ndev = (struct net_device *)netdev_lower_get_next(ndev, &lower);
561 if (is_vlan_dev(ndev)) {
562 *vlan_id = vlan_dev_vlan_id(ndev);
574 static bool smcr_lgr_match(struct smc_link_group *lgr,
575 struct smc_clc_msg_local *lcl,
576 enum smc_lgr_role role, u32 clcqpn)
578 return !memcmp(lgr->peer_systemid, lcl->id_for_peer,
580 !memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_gid, &lcl->gid,
582 !memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_mac, lcl->mac,
585 (lgr->role == SMC_SERV ||
586 lgr->lnk[SMC_SINGLE_LINK].peer_qpn == clcqpn);
589 static bool smcd_lgr_match(struct smc_link_group *lgr,
590 struct smcd_dev *smcismdev, u64 peer_gid)
592 return lgr->peer_gid == peer_gid && lgr->smcd == smcismdev;
595 /* create a new SMC connection (and a new link group if necessary) */
596 int smc_conn_create(struct smc_sock *smc, bool is_smcd, int srv_first_contact,
597 struct smc_ib_device *smcibdev, u8 ibport, u32 clcqpn,
598 struct smc_clc_msg_local *lcl, struct smcd_dev *smcd,
601 struct smc_connection *conn = &smc->conn;
602 int local_contact = SMC_FIRST_CONTACT;
603 struct smc_link_group *lgr;
604 unsigned short vlan_id;
605 enum smc_lgr_role role;
608 role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
609 rc = smc_vlan_by_tcpsk(smc->clcsock, &vlan_id);
613 if ((role == SMC_CLNT) && srv_first_contact)
614 /* create new link group as well */
617 /* determine if an existing link group can be reused */
618 spin_lock_bh(&smc_lgr_list.lock);
619 list_for_each_entry(lgr, &smc_lgr_list.list, list) {
620 write_lock_bh(&lgr->conns_lock);
621 if ((is_smcd ? smcd_lgr_match(lgr, smcd, peer_gid) :
622 smcr_lgr_match(lgr, lcl, role, clcqpn)) &&
624 lgr->vlan_id == vlan_id &&
626 lgr->conns_num < SMC_RMBS_PER_LGR_MAX)) {
627 /* link group found */
628 local_contact = SMC_REUSE_CONTACT;
630 smc_lgr_register_conn(conn); /* add smc conn to lgr */
631 write_unlock_bh(&lgr->conns_lock);
634 write_unlock_bh(&lgr->conns_lock);
636 spin_unlock_bh(&smc_lgr_list.lock);
638 if (role == SMC_CLNT && !srv_first_contact &&
639 (local_contact == SMC_FIRST_CONTACT)) {
640 /* Server reuses a link group, but Client wants to start
642 * send out_of_sync decline, reason synchr. error
648 if (local_contact == SMC_FIRST_CONTACT) {
649 rc = smc_lgr_create(smc, is_smcd, smcibdev, ibport,
650 lcl->id_for_peer, vlan_id, smcd, peer_gid);
653 smc_lgr_register_conn(conn); /* add smc conn to lgr */
655 conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
656 conn->local_tx_ctrl.len = SMC_WR_TX_SIZE;
657 conn->urg_state = SMC_URG_READ;
659 conn->rx_off = sizeof(struct smcd_cdc_msg);
660 smcd_cdc_rx_init(conn); /* init tasklet for this conn */
662 #ifndef KERNEL_HAS_ATOMIC64
663 spin_lock_init(&conn->acurs_lock);
667 return rc ? rc : local_contact;
670 /* convert the RMB size into the compressed notation - minimum 16K.
671 * In contrast to plain ilog2, this rounds towards the next power of 2,
672 * so the socket application gets at least its desired sndbuf / rcvbuf size.
674 static u8 smc_compress_bufsize(int size)
678 if (size <= SMC_BUF_MIN_SIZE)
681 size = (size - 1) >> 14;
682 compressed = ilog2(size) + 1;
683 if (compressed >= SMC_RMBE_SIZES)
684 compressed = SMC_RMBE_SIZES - 1;
688 /* convert the RMB size from compressed notation into integer */
689 int smc_uncompress_bufsize(u8 compressed)
693 size = 0x00000001 << (((int)compressed) + 14);
697 /* try to reuse a sndbuf or rmb description slot for a certain
698 * buffer size; if not available, return NULL
700 static struct smc_buf_desc *smc_buf_get_slot(int compressed_bufsize,
702 struct list_head *buf_list)
704 struct smc_buf_desc *buf_slot;
707 list_for_each_entry(buf_slot, buf_list, list) {
708 if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
709 read_unlock_bh(lock);
713 read_unlock_bh(lock);
717 /* one of the conditions for announcing a receiver's current window size is
718 * that it "results in a minimum increase in the window size of 10% of the
719 * receive buffer space" [RFC7609]
721 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
723 return min_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
726 static struct smc_buf_desc *smcr_new_buf_create(struct smc_link_group *lgr,
727 bool is_rmb, int bufsize)
729 struct smc_buf_desc *buf_desc;
730 struct smc_link *lnk;
733 /* try to alloc a new buffer */
734 buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
736 return ERR_PTR(-ENOMEM);
738 buf_desc->order = get_order(bufsize);
739 buf_desc->pages = alloc_pages(GFP_KERNEL | __GFP_NOWARN |
740 __GFP_NOMEMALLOC | __GFP_COMP |
741 __GFP_NORETRY | __GFP_ZERO,
743 if (!buf_desc->pages) {
745 return ERR_PTR(-EAGAIN);
747 buf_desc->cpu_addr = (void *)page_address(buf_desc->pages);
749 /* build the sg table from the pages */
750 lnk = &lgr->lnk[SMC_SINGLE_LINK];
751 rc = sg_alloc_table(&buf_desc->sgt[SMC_SINGLE_LINK], 1,
754 smc_buf_free(lgr, is_rmb, buf_desc);
757 sg_set_buf(buf_desc->sgt[SMC_SINGLE_LINK].sgl,
758 buf_desc->cpu_addr, bufsize);
760 /* map sg table to DMA address */
761 rc = smc_ib_buf_map_sg(lnk->smcibdev, buf_desc,
762 is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
763 /* SMC protocol depends on mapping to one DMA address only */
765 smc_buf_free(lgr, is_rmb, buf_desc);
766 return ERR_PTR(-EAGAIN);
769 /* create a new memory region for the RMB */
771 rc = smc_ib_get_memory_region(lnk->roce_pd,
772 IB_ACCESS_REMOTE_WRITE |
773 IB_ACCESS_LOCAL_WRITE,
776 smc_buf_free(lgr, is_rmb, buf_desc);
781 buf_desc->len = bufsize;
785 #define SMCD_DMBE_SIZES 7 /* 0 -> 16KB, 1 -> 32KB, .. 6 -> 1MB */
787 static struct smc_buf_desc *smcd_new_buf_create(struct smc_link_group *lgr,
788 bool is_dmb, int bufsize)
790 struct smc_buf_desc *buf_desc;
793 if (smc_compress_bufsize(bufsize) > SMCD_DMBE_SIZES)
794 return ERR_PTR(-EAGAIN);
796 /* try to alloc a new DMB */
797 buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
799 return ERR_PTR(-ENOMEM);
801 rc = smc_ism_register_dmb(lgr, bufsize, buf_desc);
804 return ERR_PTR(-EAGAIN);
806 buf_desc->pages = virt_to_page(buf_desc->cpu_addr);
807 /* CDC header stored in buf. So, pretend it was smaller */
808 buf_desc->len = bufsize - sizeof(struct smcd_cdc_msg);
810 buf_desc->cpu_addr = kzalloc(bufsize, GFP_KERNEL |
811 __GFP_NOWARN | __GFP_NORETRY |
813 if (!buf_desc->cpu_addr) {
815 return ERR_PTR(-EAGAIN);
817 buf_desc->len = bufsize;
822 static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
824 struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
825 struct smc_connection *conn = &smc->conn;
826 struct smc_link_group *lgr = conn->lgr;
827 struct list_head *buf_list;
828 int bufsize, bufsize_short;
833 /* use socket recv buffer size (w/o overhead) as start value */
834 sk_buf_size = smc->sk.sk_rcvbuf / 2;
836 /* use socket send buffer size (w/o overhead) as start value */
837 sk_buf_size = smc->sk.sk_sndbuf / 2;
839 for (bufsize_short = smc_compress_bufsize(sk_buf_size);
840 bufsize_short >= 0; bufsize_short--) {
843 lock = &lgr->rmbs_lock;
844 buf_list = &lgr->rmbs[bufsize_short];
846 lock = &lgr->sndbufs_lock;
847 buf_list = &lgr->sndbufs[bufsize_short];
849 bufsize = smc_uncompress_bufsize(bufsize_short);
850 if ((1 << get_order(bufsize)) > SG_MAX_SINGLE_ALLOC)
853 /* check for reusable slot in the link group */
854 buf_desc = smc_buf_get_slot(bufsize_short, lock, buf_list);
856 memset(buf_desc->cpu_addr, 0, bufsize);
857 break; /* found reusable slot */
861 buf_desc = smcd_new_buf_create(lgr, is_rmb, bufsize);
863 buf_desc = smcr_new_buf_create(lgr, is_rmb, bufsize);
865 if (PTR_ERR(buf_desc) == -ENOMEM)
867 if (IS_ERR(buf_desc))
872 list_add(&buf_desc->list, buf_list);
873 write_unlock_bh(lock);
877 if (IS_ERR(buf_desc))
881 conn->rmb_desc = buf_desc;
882 conn->rmbe_size_short = bufsize_short;
883 smc->sk.sk_rcvbuf = bufsize * 2;
884 atomic_set(&conn->bytes_to_rcv, 0);
885 conn->rmbe_update_limit =
886 smc_rmb_wnd_update_limit(buf_desc->len);
888 smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */
890 conn->sndbuf_desc = buf_desc;
891 smc->sk.sk_sndbuf = bufsize * 2;
892 atomic_set(&conn->sndbuf_space, bufsize);
897 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn)
899 struct smc_link_group *lgr = conn->lgr;
901 if (!conn->lgr || conn->lgr->is_smcd)
903 smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
904 conn->sndbuf_desc, DMA_TO_DEVICE);
907 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
909 struct smc_link_group *lgr = conn->lgr;
911 if (!conn->lgr || conn->lgr->is_smcd)
913 smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
914 conn->sndbuf_desc, DMA_TO_DEVICE);
917 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
919 struct smc_link_group *lgr = conn->lgr;
921 if (!conn->lgr || conn->lgr->is_smcd)
923 smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
924 conn->rmb_desc, DMA_FROM_DEVICE);
927 void smc_rmb_sync_sg_for_device(struct smc_connection *conn)
929 struct smc_link_group *lgr = conn->lgr;
931 if (!conn->lgr || conn->lgr->is_smcd)
933 smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
934 conn->rmb_desc, DMA_FROM_DEVICE);
937 /* create the send and receive buffer for an SMC socket;
938 * receive buffers are called RMBs;
939 * (even though the SMC protocol allows more than one RMB-element per RMB,
940 * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
941 * extra RMB for every connection in a link group
943 int smc_buf_create(struct smc_sock *smc, bool is_smcd)
947 /* create send buffer */
948 rc = __smc_buf_create(smc, is_smcd, false);
952 rc = __smc_buf_create(smc, is_smcd, true);
954 smc_buf_free(smc->conn.lgr, false, smc->conn.sndbuf_desc);
958 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
962 for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
963 if (!test_and_set_bit(i, lgr->rtokens_used_mask))
969 /* add a new rtoken from peer */
970 int smc_rtoken_add(struct smc_link_group *lgr, __be64 nw_vaddr, __be32 nw_rkey)
972 u64 dma_addr = be64_to_cpu(nw_vaddr);
973 u32 rkey = ntohl(nw_rkey);
976 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
977 if ((lgr->rtokens[i][SMC_SINGLE_LINK].rkey == rkey) &&
978 (lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr == dma_addr) &&
979 test_bit(i, lgr->rtokens_used_mask)) {
980 /* already in list */
984 i = smc_rmb_reserve_rtoken_idx(lgr);
987 lgr->rtokens[i][SMC_SINGLE_LINK].rkey = rkey;
988 lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr = dma_addr;
992 /* delete an rtoken */
993 int smc_rtoken_delete(struct smc_link_group *lgr, __be32 nw_rkey)
995 u32 rkey = ntohl(nw_rkey);
998 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
999 if (lgr->rtokens[i][SMC_SINGLE_LINK].rkey == rkey &&
1000 test_bit(i, lgr->rtokens_used_mask)) {
1001 lgr->rtokens[i][SMC_SINGLE_LINK].rkey = 0;
1002 lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr = 0;
1004 clear_bit(i, lgr->rtokens_used_mask);
1011 /* save rkey and dma_addr received from peer during clc handshake */
1012 int smc_rmb_rtoken_handling(struct smc_connection *conn,
1013 struct smc_clc_msg_accept_confirm *clc)
1015 conn->rtoken_idx = smc_rtoken_add(conn->lgr, clc->rmb_dma_addr,
1017 if (conn->rtoken_idx < 0)
1018 return conn->rtoken_idx;
1022 /* Called (from smc_exit) when module is removed */
1023 void smc_core_exit(void)
1025 struct smc_link_group *lgr, *lg;
1026 LIST_HEAD(lgr_freeing_list);
1028 spin_lock_bh(&smc_lgr_list.lock);
1029 if (!list_empty(&smc_lgr_list.list))
1030 list_splice_init(&smc_lgr_list.list, &lgr_freeing_list);
1031 spin_unlock_bh(&smc_lgr_list.lock);
1032 list_for_each_entry_safe(lgr, lg, &lgr_freeing_list, list) {
1033 list_del_init(&lgr->list);
1034 if (!lgr->is_smcd) {
1035 struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
1037 if (lnk->state == SMC_LNK_ACTIVE)
1038 smc_llc_send_delete_link(lnk, SMC_LLC_REQ,
1040 smc_llc_link_inactive(lnk);
1042 cancel_delayed_work_sync(&lgr->free_work);
1044 smc_ism_signal_shutdown(lgr);
1045 smc_lgr_free(lgr); /* free link group */