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>
27 #include "smc_close.h"
29 #define SMC_LGR_NUM_INCR 256
30 #define SMC_LGR_FREE_DELAY_SERV (600 * HZ)
31 #define SMC_LGR_FREE_DELAY_CLNT (SMC_LGR_FREE_DELAY_SERV + 10)
33 static u32 smc_lgr_num; /* unique link group number */
35 /* Register connection's alert token in our lookup structure.
36 * To use rbtrees we have to implement our own insert core.
37 * Requires @conns_lock
38 * @smc connection to register
39 * Returns 0 on success, != otherwise.
41 static void smc_lgr_add_alert_token(struct smc_connection *conn)
43 struct rb_node **link, *parent = NULL;
44 u32 token = conn->alert_token_local;
46 link = &conn->lgr->conns_all.rb_node;
48 struct smc_connection *cur = rb_entry(*link,
49 struct smc_connection, alert_node);
52 if (cur->alert_token_local > token)
53 link = &parent->rb_left;
55 link = &parent->rb_right;
57 /* Put the new node there */
58 rb_link_node(&conn->alert_node, parent, link);
59 rb_insert_color(&conn->alert_node, &conn->lgr->conns_all);
62 /* Register connection in link group by assigning an alert token
63 * registered in a search tree.
64 * Requires @conns_lock
65 * Note that '0' is a reserved value and not assigned.
67 static void smc_lgr_register_conn(struct smc_connection *conn)
69 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
70 static atomic_t nexttoken = ATOMIC_INIT(0);
72 /* find a new alert_token_local value not yet used by some connection
75 sock_hold(&smc->sk); /* sock_put in smc_lgr_unregister_conn() */
76 while (!conn->alert_token_local) {
77 conn->alert_token_local = atomic_inc_return(&nexttoken);
78 if (smc_lgr_find_conn(conn->alert_token_local, conn->lgr))
79 conn->alert_token_local = 0;
81 smc_lgr_add_alert_token(conn);
82 conn->lgr->conns_num++;
85 /* Unregister connection and reset the alert token of the given connection<
87 static void __smc_lgr_unregister_conn(struct smc_connection *conn)
89 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
90 struct smc_link_group *lgr = conn->lgr;
92 rb_erase(&conn->alert_node, &lgr->conns_all);
94 conn->alert_token_local = 0;
96 sock_put(&smc->sk); /* sock_hold in smc_lgr_register_conn() */
99 /* Unregister connection and trigger lgr freeing if applicable
101 static void smc_lgr_unregister_conn(struct smc_connection *conn)
103 struct smc_link_group *lgr = conn->lgr;
108 write_lock_bh(&lgr->conns_lock);
109 if (conn->alert_token_local) {
111 __smc_lgr_unregister_conn(conn);
113 write_unlock_bh(&lgr->conns_lock);
114 if (!reduced || lgr->conns_num)
116 /* client link group creation always follows the server link group
117 * creation. For client use a somewhat higher removal delay time,
118 * otherwise there is a risk of out-of-sync link groups.
120 mod_delayed_work(system_wq, &lgr->free_work,
121 lgr->role == SMC_CLNT ? SMC_LGR_FREE_DELAY_CLNT :
122 SMC_LGR_FREE_DELAY_SERV);
125 static void smc_lgr_free_work(struct work_struct *work)
127 struct smc_link_group *lgr = container_of(to_delayed_work(work),
128 struct smc_link_group,
132 spin_lock_bh(&smc_lgr_list.lock);
133 read_lock_bh(&lgr->conns_lock);
134 conns = RB_EMPTY_ROOT(&lgr->conns_all);
135 read_unlock_bh(&lgr->conns_lock);
136 if (!conns) { /* number of lgr connections is no longer zero */
137 spin_unlock_bh(&smc_lgr_list.lock);
140 list_del_init(&lgr->list); /* remove from smc_lgr_list */
141 spin_unlock_bh(&smc_lgr_list.lock);
145 /* create a new SMC link group */
146 static int smc_lgr_create(struct smc_sock *smc, __be32 peer_in_addr,
147 struct smc_ib_device *smcibdev, u8 ibport,
148 char *peer_systemid, unsigned short vlan_id)
150 struct smc_link_group *lgr;
151 struct smc_link *lnk;
156 lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
161 lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
162 lgr->sync_err = false;
163 lgr->daddr = peer_in_addr;
164 memcpy(lgr->peer_systemid, peer_systemid, SMC_SYSTEMID_LEN);
165 lgr->vlan_id = vlan_id;
166 rwlock_init(&lgr->sndbufs_lock);
167 rwlock_init(&lgr->rmbs_lock);
168 for (i = 0; i < SMC_RMBE_SIZES; i++) {
169 INIT_LIST_HEAD(&lgr->sndbufs[i]);
170 INIT_LIST_HEAD(&lgr->rmbs[i]);
172 smc_lgr_num += SMC_LGR_NUM_INCR;
173 memcpy(&lgr->id, (u8 *)&smc_lgr_num, SMC_LGR_ID_SIZE);
174 INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
175 lgr->conns_all = RB_ROOT;
177 lnk = &lgr->lnk[SMC_SINGLE_LINK];
178 /* initialize link */
179 lnk->link_id = SMC_SINGLE_LINK;
180 lnk->smcibdev = smcibdev;
181 lnk->ibport = ibport;
182 lnk->path_mtu = smcibdev->pattr[ibport - 1].active_mtu;
183 if (!smcibdev->initialized)
184 smc_ib_setup_per_ibdev(smcibdev);
185 get_random_bytes(rndvec, sizeof(rndvec));
186 lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) + (rndvec[2] << 16);
187 rc = smc_wr_alloc_link_mem(lnk);
190 rc = smc_ib_create_protection_domain(lnk);
193 rc = smc_ib_create_queue_pair(lnk);
196 rc = smc_wr_create_link(lnk);
199 init_completion(&lnk->llc_confirm);
200 init_completion(&lnk->llc_confirm_resp);
203 rwlock_init(&lgr->conns_lock);
204 spin_lock_bh(&smc_lgr_list.lock);
205 list_add(&lgr->list, &smc_lgr_list.list);
206 spin_unlock_bh(&smc_lgr_list.lock);
210 smc_ib_destroy_queue_pair(lnk);
212 smc_ib_dealloc_protection_domain(lnk);
214 smc_wr_free_link_mem(lnk);
221 static void smc_buf_unuse(struct smc_connection *conn)
223 if (conn->sndbuf_desc) {
224 conn->sndbuf_desc->used = 0;
225 conn->sndbuf_size = 0;
227 if (conn->rmb_desc) {
228 conn->rmb_desc->reused = true;
229 conn->rmb_desc->used = 0;
234 /* remove a finished connection from its link group */
235 void smc_conn_free(struct smc_connection *conn)
237 struct smc_link_group *lgr = conn->lgr;
241 smc_cdc_tx_dismiss_slots(conn);
243 smc_lgr_unregister_conn(conn);
246 static void smc_link_clear(struct smc_link *lnk)
249 smc_ib_modify_qp_reset(lnk);
250 smc_wr_free_link(lnk);
251 smc_ib_destroy_queue_pair(lnk);
252 smc_ib_dealloc_protection_domain(lnk);
253 smc_wr_free_link_mem(lnk);
256 static void smc_buf_free(struct smc_buf_desc *buf_desc, struct smc_link *lnk,
260 if (buf_desc->mr_rx[SMC_SINGLE_LINK])
261 smc_ib_put_memory_region(
262 buf_desc->mr_rx[SMC_SINGLE_LINK]);
263 smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
266 smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
269 sg_free_table(&buf_desc->sgt[SMC_SINGLE_LINK]);
270 if (buf_desc->cpu_addr)
271 free_pages((unsigned long)buf_desc->cpu_addr, buf_desc->order);
275 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
277 struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
278 struct smc_buf_desc *buf_desc, *bf_desc;
279 struct list_head *buf_list;
282 for (i = 0; i < SMC_RMBE_SIZES; i++) {
284 buf_list = &lgr->rmbs[i];
286 buf_list = &lgr->sndbufs[i];
287 list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
289 list_del(&buf_desc->list);
290 smc_buf_free(buf_desc, lnk, is_rmb);
295 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
297 /* free send buffers */
298 __smc_lgr_free_bufs(lgr, false);
300 __smc_lgr_free_bufs(lgr, true);
303 /* remove a link group */
304 void smc_lgr_free(struct smc_link_group *lgr)
306 smc_lgr_free_bufs(lgr);
307 smc_link_clear(&lgr->lnk[SMC_SINGLE_LINK]);
311 /* terminate linkgroup abnormally */
312 void smc_lgr_terminate(struct smc_link_group *lgr)
314 struct smc_connection *conn;
315 struct smc_sock *smc;
316 struct rb_node *node;
318 spin_lock_bh(&smc_lgr_list.lock);
319 if (list_empty(&lgr->list)) {
320 /* termination already triggered */
321 spin_unlock_bh(&smc_lgr_list.lock);
324 /* do not use this link group for new connections */
325 list_del_init(&lgr->list);
326 spin_unlock_bh(&smc_lgr_list.lock);
328 write_lock_bh(&lgr->conns_lock);
329 node = rb_first(&lgr->conns_all);
331 conn = rb_entry(node, struct smc_connection, alert_node);
332 smc = container_of(conn, struct smc_sock, conn);
334 __smc_lgr_unregister_conn(conn);
335 schedule_work(&conn->close_work);
337 node = rb_first(&lgr->conns_all);
339 write_unlock_bh(&lgr->conns_lock);
342 /* Determine vlan of internal TCP socket.
343 * @vlan_id: address to store the determined vlan id into
345 static int smc_vlan_by_tcpsk(struct socket *clcsock, unsigned short *vlan_id)
347 struct dst_entry *dst = sk_dst_get(clcsock->sk);
360 if (is_vlan_dev(dst->dev))
361 *vlan_id = vlan_dev_vlan_id(dst->dev);
369 /* determine the link gid matching the vlan id of the link group */
370 static int smc_link_determine_gid(struct smc_link_group *lgr)
372 struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
373 struct ib_gid_attr gattr;
378 lnk->gid = lnk->smcibdev->gid[lnk->ibport - 1];
382 for (i = 0; i < lnk->smcibdev->pattr[lnk->ibport - 1].gid_tbl_len;
384 if (ib_query_gid(lnk->smcibdev->ibdev, lnk->ibport, i, &gid,
388 (vlan_dev_vlan_id(gattr.ndev) == lgr->vlan_id)) {
396 /* create a new SMC connection (and a new link group if necessary) */
397 int smc_conn_create(struct smc_sock *smc, __be32 peer_in_addr,
398 struct smc_ib_device *smcibdev, u8 ibport,
399 struct smc_clc_msg_local *lcl, int srv_first_contact)
401 struct smc_connection *conn = &smc->conn;
402 struct smc_link_group *lgr;
403 unsigned short vlan_id;
404 enum smc_lgr_role role;
405 int local_contact = SMC_FIRST_CONTACT;
408 role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
409 rc = smc_vlan_by_tcpsk(smc->clcsock, &vlan_id);
413 if ((role == SMC_CLNT) && srv_first_contact)
414 /* create new link group as well */
417 /* determine if an existing link group can be reused */
418 spin_lock_bh(&smc_lgr_list.lock);
419 list_for_each_entry(lgr, &smc_lgr_list.list, list) {
420 write_lock_bh(&lgr->conns_lock);
421 if (!memcmp(lgr->peer_systemid, lcl->id_for_peer,
423 !memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_gid, &lcl->gid,
425 !memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_mac, lcl->mac,
428 (lgr->role == role) &&
429 (lgr->vlan_id == vlan_id) &&
430 ((role == SMC_CLNT) ||
431 (lgr->conns_num < SMC_RMBS_PER_LGR_MAX &&
432 !bitmap_full(lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX)))) {
433 /* link group found */
434 local_contact = SMC_REUSE_CONTACT;
436 smc_lgr_register_conn(conn); /* add smc conn to lgr */
437 if (delayed_work_pending(&lgr->free_work))
438 cancel_delayed_work(&lgr->free_work);
439 write_unlock_bh(&lgr->conns_lock);
442 write_unlock_bh(&lgr->conns_lock);
444 spin_unlock_bh(&smc_lgr_list.lock);
446 if (role == SMC_CLNT && !srv_first_contact &&
447 (local_contact == SMC_FIRST_CONTACT)) {
448 /* Server reuses a link group, but Client wants to start
450 * send out_of_sync decline, reason synchr. error
456 if (local_contact == SMC_FIRST_CONTACT) {
457 rc = smc_lgr_create(smc, peer_in_addr, smcibdev, ibport,
458 lcl->id_for_peer, vlan_id);
461 smc_lgr_register_conn(conn); /* add smc conn to lgr */
462 rc = smc_link_determine_gid(conn->lgr);
464 conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
465 conn->local_tx_ctrl.len = sizeof(struct smc_cdc_msg);
466 #ifndef KERNEL_HAS_ATOMIC64
467 spin_lock_init(&conn->acurs_lock);
471 return rc ? rc : local_contact;
474 /* try to reuse a sndbuf or rmb description slot for a certain
475 * buffer size; if not available, return NULL
478 struct smc_buf_desc *smc_buf_get_slot(struct smc_link_group *lgr,
479 int compressed_bufsize,
481 struct list_head *buf_list)
483 struct smc_buf_desc *buf_slot;
486 list_for_each_entry(buf_slot, buf_list, list) {
487 if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
488 read_unlock_bh(lock);
492 read_unlock_bh(lock);
496 /* one of the conditions for announcing a receiver's current window size is
497 * that it "results in a minimum increase in the window size of 10% of the
498 * receive buffer space" [RFC7609]
500 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
502 return max_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
505 static struct smc_buf_desc *smc_new_buf_create(struct smc_link_group *lgr,
506 bool is_rmb, int bufsize)
508 struct smc_buf_desc *buf_desc;
509 struct smc_link *lnk;
512 /* try to alloc a new buffer */
513 buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
515 return ERR_PTR(-ENOMEM);
518 (void *)__get_free_pages(GFP_KERNEL | __GFP_NOWARN |
520 __GFP_NORETRY | __GFP_ZERO,
522 if (!buf_desc->cpu_addr) {
524 return ERR_PTR(-EAGAIN);
526 buf_desc->order = get_order(bufsize);
528 /* build the sg table from the pages */
529 lnk = &lgr->lnk[SMC_SINGLE_LINK];
530 rc = sg_alloc_table(&buf_desc->sgt[SMC_SINGLE_LINK], 1,
533 smc_buf_free(buf_desc, lnk, is_rmb);
536 sg_set_buf(buf_desc->sgt[SMC_SINGLE_LINK].sgl,
537 buf_desc->cpu_addr, bufsize);
539 /* map sg table to DMA address */
540 rc = smc_ib_buf_map_sg(lnk->smcibdev, buf_desc,
541 is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
542 /* SMC protocol depends on mapping to one DMA address only */
544 smc_buf_free(buf_desc, lnk, is_rmb);
545 return ERR_PTR(-EAGAIN);
548 /* create a new memory region for the RMB */
550 rc = smc_ib_get_memory_region(lnk->roce_pd,
551 IB_ACCESS_REMOTE_WRITE |
552 IB_ACCESS_LOCAL_WRITE,
555 smc_buf_free(buf_desc, lnk, is_rmb);
563 static int __smc_buf_create(struct smc_sock *smc, bool is_rmb)
565 struct smc_connection *conn = &smc->conn;
566 struct smc_link_group *lgr = conn->lgr;
567 struct smc_buf_desc *buf_desc = NULL;
568 struct list_head *buf_list;
569 int bufsize, bufsize_short;
574 /* use socket recv buffer size (w/o overhead) as start value */
575 sk_buf_size = smc->sk.sk_rcvbuf / 2;
577 /* use socket send buffer size (w/o overhead) as start value */
578 sk_buf_size = smc->sk.sk_sndbuf / 2;
580 for (bufsize_short = smc_compress_bufsize(sk_buf_size);
581 bufsize_short >= 0; bufsize_short--) {
584 lock = &lgr->rmbs_lock;
585 buf_list = &lgr->rmbs[bufsize_short];
587 lock = &lgr->sndbufs_lock;
588 buf_list = &lgr->sndbufs[bufsize_short];
590 bufsize = smc_uncompress_bufsize(bufsize_short);
591 if ((1 << get_order(bufsize)) > SG_MAX_SINGLE_ALLOC)
594 /* check for reusable slot in the link group */
595 buf_desc = smc_buf_get_slot(lgr, bufsize_short, lock, buf_list);
597 memset(buf_desc->cpu_addr, 0, bufsize);
598 break; /* found reusable slot */
601 buf_desc = smc_new_buf_create(lgr, is_rmb, bufsize);
602 if (PTR_ERR(buf_desc) == -ENOMEM)
604 if (IS_ERR(buf_desc))
609 list_add(&buf_desc->list, buf_list);
610 write_unlock_bh(lock);
614 if (IS_ERR(buf_desc))
618 conn->rmb_desc = buf_desc;
619 conn->rmbe_size = bufsize;
620 conn->rmbe_size_short = bufsize_short;
621 smc->sk.sk_rcvbuf = bufsize * 2;
622 atomic_set(&conn->bytes_to_rcv, 0);
623 conn->rmbe_update_limit = smc_rmb_wnd_update_limit(bufsize);
625 conn->sndbuf_desc = buf_desc;
626 conn->sndbuf_size = bufsize;
627 smc->sk.sk_sndbuf = bufsize * 2;
628 atomic_set(&conn->sndbuf_space, bufsize);
633 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn)
635 struct smc_link_group *lgr = conn->lgr;
637 smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
638 conn->sndbuf_desc, DMA_TO_DEVICE);
641 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
643 struct smc_link_group *lgr = conn->lgr;
645 smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
646 conn->sndbuf_desc, DMA_TO_DEVICE);
649 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
651 struct smc_link_group *lgr = conn->lgr;
653 smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
654 conn->rmb_desc, DMA_FROM_DEVICE);
657 void smc_rmb_sync_sg_for_device(struct smc_connection *conn)
659 struct smc_link_group *lgr = conn->lgr;
661 smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
662 conn->rmb_desc, DMA_FROM_DEVICE);
665 /* create the send and receive buffer for an SMC socket;
666 * receive buffers are called RMBs;
667 * (even though the SMC protocol allows more than one RMB-element per RMB,
668 * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
669 * extra RMB for every connection in a link group
671 int smc_buf_create(struct smc_sock *smc)
675 /* create send buffer */
676 rc = __smc_buf_create(smc, false);
680 rc = __smc_buf_create(smc, true);
682 smc_buf_free(smc->conn.sndbuf_desc,
683 &smc->conn.lgr->lnk[SMC_SINGLE_LINK], false);
687 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
691 for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
692 if (!test_and_set_bit(i, lgr->rtokens_used_mask))
698 /* save rkey and dma_addr received from peer during clc handshake */
699 int smc_rmb_rtoken_handling(struct smc_connection *conn,
700 struct smc_clc_msg_accept_confirm *clc)
702 u64 dma_addr = be64_to_cpu(clc->rmb_dma_addr);
703 struct smc_link_group *lgr = conn->lgr;
704 u32 rkey = ntohl(clc->rmb_rkey);
707 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
708 if ((lgr->rtokens[i][SMC_SINGLE_LINK].rkey == rkey) &&
709 (lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr == dma_addr) &&
710 test_bit(i, lgr->rtokens_used_mask)) {
711 conn->rtoken_idx = i;
715 conn->rtoken_idx = smc_rmb_reserve_rtoken_idx(lgr);
716 if (conn->rtoken_idx < 0)
717 return conn->rtoken_idx;
718 lgr->rtokens[conn->rtoken_idx][SMC_SINGLE_LINK].rkey = rkey;
719 lgr->rtokens[conn->rtoken_idx][SMC_SINGLE_LINK].dma_addr = dma_addr;