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;
106 write_lock_bh(&lgr->conns_lock);
107 if (conn->alert_token_local) {
109 __smc_lgr_unregister_conn(conn);
111 write_unlock_bh(&lgr->conns_lock);
112 if (!reduced || lgr->conns_num)
114 /* client link group creation always follows the server link group
115 * creation. For client use a somewhat higher removal delay time,
116 * otherwise there is a risk of out-of-sync link groups.
118 mod_delayed_work(system_wq, &lgr->free_work,
119 lgr->role == SMC_CLNT ? SMC_LGR_FREE_DELAY_CLNT :
120 SMC_LGR_FREE_DELAY_SERV);
123 static void smc_lgr_free_work(struct work_struct *work)
125 struct smc_link_group *lgr = container_of(to_delayed_work(work),
126 struct smc_link_group,
130 spin_lock_bh(&smc_lgr_list.lock);
131 read_lock_bh(&lgr->conns_lock);
132 conns = RB_EMPTY_ROOT(&lgr->conns_all);
133 read_unlock_bh(&lgr->conns_lock);
134 if (!conns) { /* number of lgr connections is no longer zero */
135 spin_unlock_bh(&smc_lgr_list.lock);
138 list_del_init(&lgr->list); /* remove from smc_lgr_list */
139 spin_unlock_bh(&smc_lgr_list.lock);
143 /* create a new SMC link group */
144 static int smc_lgr_create(struct smc_sock *smc, __be32 peer_in_addr,
145 struct smc_ib_device *smcibdev, u8 ibport,
146 char *peer_systemid, unsigned short vlan_id)
148 struct smc_link_group *lgr;
149 struct smc_link *lnk;
154 lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
159 lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
160 lgr->sync_err = false;
161 lgr->daddr = peer_in_addr;
162 memcpy(lgr->peer_systemid, peer_systemid, SMC_SYSTEMID_LEN);
163 lgr->vlan_id = vlan_id;
164 rwlock_init(&lgr->sndbufs_lock);
165 rwlock_init(&lgr->rmbs_lock);
166 for (i = 0; i < SMC_RMBE_SIZES; i++) {
167 INIT_LIST_HEAD(&lgr->sndbufs[i]);
168 INIT_LIST_HEAD(&lgr->rmbs[i]);
170 smc_lgr_num += SMC_LGR_NUM_INCR;
171 memcpy(&lgr->id, (u8 *)&smc_lgr_num, SMC_LGR_ID_SIZE);
172 INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
173 lgr->conns_all = RB_ROOT;
175 lnk = &lgr->lnk[SMC_SINGLE_LINK];
176 /* initialize link */
177 lnk->smcibdev = smcibdev;
178 lnk->ibport = ibport;
179 lnk->path_mtu = smcibdev->pattr[ibport - 1].active_mtu;
180 if (!smcibdev->initialized)
181 smc_ib_setup_per_ibdev(smcibdev);
182 get_random_bytes(rndvec, sizeof(rndvec));
183 lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) + (rndvec[2] << 16);
184 rc = smc_wr_alloc_link_mem(lnk);
187 rc = smc_ib_create_protection_domain(lnk);
190 rc = smc_ib_create_queue_pair(lnk);
193 rc = smc_wr_create_link(lnk);
196 init_completion(&lnk->llc_confirm);
197 init_completion(&lnk->llc_confirm_resp);
200 rwlock_init(&lgr->conns_lock);
201 spin_lock_bh(&smc_lgr_list.lock);
202 list_add(&lgr->list, &smc_lgr_list.list);
203 spin_unlock_bh(&smc_lgr_list.lock);
207 smc_ib_destroy_queue_pair(lnk);
209 smc_ib_dealloc_protection_domain(lnk);
211 smc_wr_free_link_mem(lnk);
218 static void smc_buf_unuse(struct smc_connection *conn)
220 if (conn->sndbuf_desc) {
221 conn->sndbuf_desc->used = 0;
222 conn->sndbuf_size = 0;
224 if (conn->rmb_desc) {
225 conn->rmb_desc->reused = true;
226 conn->rmb_desc->used = 0;
231 /* remove a finished connection from its link group */
232 void smc_conn_free(struct smc_connection *conn)
234 struct smc_link_group *lgr = conn->lgr;
238 smc_cdc_tx_dismiss_slots(conn);
239 smc_lgr_unregister_conn(conn);
243 static void smc_link_clear(struct smc_link *lnk)
246 smc_ib_modify_qp_reset(lnk);
247 smc_wr_free_link(lnk);
248 smc_ib_destroy_queue_pair(lnk);
249 smc_ib_dealloc_protection_domain(lnk);
250 smc_wr_free_link_mem(lnk);
253 static void smc_buf_free(struct smc_buf_desc *buf_desc, struct smc_link *lnk,
257 if (buf_desc->mr_rx[SMC_SINGLE_LINK])
258 smc_ib_put_memory_region(
259 buf_desc->mr_rx[SMC_SINGLE_LINK]);
260 smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
263 smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
266 sg_free_table(&buf_desc->sgt[SMC_SINGLE_LINK]);
267 if (buf_desc->cpu_addr)
268 free_pages((unsigned long)buf_desc->cpu_addr, buf_desc->order);
272 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
274 struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
275 struct smc_buf_desc *buf_desc, *bf_desc;
276 struct list_head *buf_list;
279 for (i = 0; i < SMC_RMBE_SIZES; i++) {
281 buf_list = &lgr->rmbs[i];
283 buf_list = &lgr->sndbufs[i];
284 list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
286 list_del(&buf_desc->list);
287 smc_buf_free(buf_desc, lnk, is_rmb);
292 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
294 /* free send buffers */
295 __smc_lgr_free_bufs(lgr, false);
297 __smc_lgr_free_bufs(lgr, true);
300 /* remove a link group */
301 void smc_lgr_free(struct smc_link_group *lgr)
303 smc_lgr_free_bufs(lgr);
304 smc_link_clear(&lgr->lnk[SMC_SINGLE_LINK]);
308 /* terminate linkgroup abnormally */
309 void smc_lgr_terminate(struct smc_link_group *lgr)
311 struct smc_connection *conn;
312 struct smc_sock *smc;
313 struct rb_node *node;
315 spin_lock_bh(&smc_lgr_list.lock);
316 if (list_empty(&lgr->list)) {
317 /* termination already triggered */
318 spin_unlock_bh(&smc_lgr_list.lock);
321 /* do not use this link group for new connections */
322 list_del_init(&lgr->list);
323 spin_unlock_bh(&smc_lgr_list.lock);
325 write_lock_bh(&lgr->conns_lock);
326 node = rb_first(&lgr->conns_all);
328 conn = rb_entry(node, struct smc_connection, alert_node);
329 smc = container_of(conn, struct smc_sock, conn);
331 __smc_lgr_unregister_conn(conn);
332 schedule_work(&conn->close_work);
334 node = rb_first(&lgr->conns_all);
336 write_unlock_bh(&lgr->conns_lock);
339 /* Determine vlan of internal TCP socket.
340 * @vlan_id: address to store the determined vlan id into
342 static int smc_vlan_by_tcpsk(struct socket *clcsock, unsigned short *vlan_id)
344 struct dst_entry *dst = sk_dst_get(clcsock->sk);
357 if (is_vlan_dev(dst->dev))
358 *vlan_id = vlan_dev_vlan_id(dst->dev);
366 /* determine the link gid matching the vlan id of the link group */
367 static int smc_link_determine_gid(struct smc_link_group *lgr)
369 struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
370 struct ib_gid_attr gattr;
375 lnk->gid = lnk->smcibdev->gid[lnk->ibport - 1];
379 for (i = 0; i < lnk->smcibdev->pattr[lnk->ibport - 1].gid_tbl_len;
381 if (ib_query_gid(lnk->smcibdev->ibdev, lnk->ibport, i, &gid,
385 if (is_vlan_dev(gattr.ndev) &&
386 vlan_dev_vlan_id(gattr.ndev) == lgr->vlan_id) {
397 /* create a new SMC connection (and a new link group if necessary) */
398 int smc_conn_create(struct smc_sock *smc, __be32 peer_in_addr,
399 struct smc_ib_device *smcibdev, u8 ibport,
400 struct smc_clc_msg_local *lcl, int srv_first_contact)
402 struct smc_connection *conn = &smc->conn;
403 struct smc_link_group *lgr;
404 unsigned short vlan_id;
405 enum smc_lgr_role role;
406 int local_contact = SMC_FIRST_CONTACT;
409 role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
410 rc = smc_vlan_by_tcpsk(smc->clcsock, &vlan_id);
414 if ((role == SMC_CLNT) && srv_first_contact)
415 /* create new link group as well */
418 /* determine if an existing link group can be reused */
419 spin_lock_bh(&smc_lgr_list.lock);
420 list_for_each_entry(lgr, &smc_lgr_list.list, list) {
421 write_lock_bh(&lgr->conns_lock);
422 if (!memcmp(lgr->peer_systemid, lcl->id_for_peer,
424 !memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_gid, &lcl->gid,
426 !memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_mac, lcl->mac,
429 (lgr->role == role) &&
430 (lgr->vlan_id == vlan_id) &&
431 ((role == SMC_CLNT) ||
432 (lgr->conns_num < 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 write_unlock_bh(&lgr->conns_lock);
440 write_unlock_bh(&lgr->conns_lock);
442 spin_unlock_bh(&smc_lgr_list.lock);
444 if (role == SMC_CLNT && !srv_first_contact &&
445 (local_contact == SMC_FIRST_CONTACT)) {
446 /* Server reuses a link group, but Client wants to start
448 * send out_of_sync decline, reason synchr. error
454 if (local_contact == SMC_FIRST_CONTACT) {
455 rc = smc_lgr_create(smc, peer_in_addr, smcibdev, ibport,
456 lcl->id_for_peer, vlan_id);
459 smc_lgr_register_conn(conn); /* add smc conn to lgr */
460 rc = smc_link_determine_gid(conn->lgr);
462 conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
463 conn->local_tx_ctrl.len = sizeof(struct smc_cdc_msg);
464 #ifndef KERNEL_HAS_ATOMIC64
465 spin_lock_init(&conn->acurs_lock);
469 return rc ? rc : local_contact;
472 /* try to reuse a sndbuf or rmb description slot for a certain
473 * buffer size; if not available, return NULL
476 struct smc_buf_desc *smc_buf_get_slot(struct smc_link_group *lgr,
477 int compressed_bufsize,
479 struct list_head *buf_list)
481 struct smc_buf_desc *buf_slot;
484 list_for_each_entry(buf_slot, buf_list, list) {
485 if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
486 read_unlock_bh(lock);
490 read_unlock_bh(lock);
494 /* one of the conditions for announcing a receiver's current window size is
495 * that it "results in a minimum increase in the window size of 10% of the
496 * receive buffer space" [RFC7609]
498 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
500 return min_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
503 static struct smc_buf_desc *smc_new_buf_create(struct smc_link_group *lgr,
504 bool is_rmb, int bufsize)
506 struct smc_buf_desc *buf_desc;
507 struct smc_link *lnk;
510 /* try to alloc a new buffer */
511 buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
513 return ERR_PTR(-ENOMEM);
516 (void *)__get_free_pages(GFP_KERNEL | __GFP_NOWARN |
518 __GFP_NORETRY | __GFP_ZERO,
520 if (!buf_desc->cpu_addr) {
522 return ERR_PTR(-EAGAIN);
524 buf_desc->order = get_order(bufsize);
526 /* build the sg table from the pages */
527 lnk = &lgr->lnk[SMC_SINGLE_LINK];
528 rc = sg_alloc_table(&buf_desc->sgt[SMC_SINGLE_LINK], 1,
531 smc_buf_free(buf_desc, lnk, is_rmb);
534 sg_set_buf(buf_desc->sgt[SMC_SINGLE_LINK].sgl,
535 buf_desc->cpu_addr, bufsize);
537 /* map sg table to DMA address */
538 rc = smc_ib_buf_map_sg(lnk->smcibdev, buf_desc,
539 is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
540 /* SMC protocol depends on mapping to one DMA address only */
542 smc_buf_free(buf_desc, lnk, is_rmb);
543 return ERR_PTR(-EAGAIN);
546 /* create a new memory region for the RMB */
548 rc = smc_ib_get_memory_region(lnk->roce_pd,
549 IB_ACCESS_REMOTE_WRITE |
550 IB_ACCESS_LOCAL_WRITE,
553 smc_buf_free(buf_desc, lnk, is_rmb);
561 static int __smc_buf_create(struct smc_sock *smc, bool is_rmb)
563 struct smc_connection *conn = &smc->conn;
564 struct smc_link_group *lgr = conn->lgr;
565 struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
566 struct list_head *buf_list;
567 int bufsize, bufsize_short;
572 /* use socket recv buffer size (w/o overhead) as start value */
573 sk_buf_size = smc->sk.sk_rcvbuf / 2;
575 /* use socket send buffer size (w/o overhead) as start value */
576 sk_buf_size = smc->sk.sk_sndbuf / 2;
578 for (bufsize_short = smc_compress_bufsize(sk_buf_size);
579 bufsize_short >= 0; bufsize_short--) {
582 lock = &lgr->rmbs_lock;
583 buf_list = &lgr->rmbs[bufsize_short];
585 lock = &lgr->sndbufs_lock;
586 buf_list = &lgr->sndbufs[bufsize_short];
588 bufsize = smc_uncompress_bufsize(bufsize_short);
589 if ((1 << get_order(bufsize)) > SG_MAX_SINGLE_ALLOC)
592 /* check for reusable slot in the link group */
593 buf_desc = smc_buf_get_slot(lgr, bufsize_short, lock, buf_list);
595 memset(buf_desc->cpu_addr, 0, bufsize);
596 break; /* found reusable slot */
599 buf_desc = smc_new_buf_create(lgr, is_rmb, bufsize);
600 if (PTR_ERR(buf_desc) == -ENOMEM)
602 if (IS_ERR(buf_desc))
607 list_add(&buf_desc->list, buf_list);
608 write_unlock_bh(lock);
612 if (IS_ERR(buf_desc))
616 conn->rmb_desc = buf_desc;
617 conn->rmbe_size = bufsize;
618 conn->rmbe_size_short = bufsize_short;
619 smc->sk.sk_rcvbuf = bufsize * 2;
620 atomic_set(&conn->bytes_to_rcv, 0);
621 conn->rmbe_update_limit = smc_rmb_wnd_update_limit(bufsize);
623 conn->sndbuf_desc = buf_desc;
624 conn->sndbuf_size = bufsize;
625 smc->sk.sk_sndbuf = bufsize * 2;
626 atomic_set(&conn->sndbuf_space, bufsize);
631 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn)
633 struct smc_link_group *lgr = conn->lgr;
635 smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
636 conn->sndbuf_desc, DMA_TO_DEVICE);
639 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
641 struct smc_link_group *lgr = conn->lgr;
643 smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
644 conn->sndbuf_desc, DMA_TO_DEVICE);
647 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
649 struct smc_link_group *lgr = conn->lgr;
651 smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
652 conn->rmb_desc, DMA_FROM_DEVICE);
655 void smc_rmb_sync_sg_for_device(struct smc_connection *conn)
657 struct smc_link_group *lgr = conn->lgr;
659 smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
660 conn->rmb_desc, DMA_FROM_DEVICE);
663 /* create the send and receive buffer for an SMC socket;
664 * receive buffers are called RMBs;
665 * (even though the SMC protocol allows more than one RMB-element per RMB,
666 * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
667 * extra RMB for every connection in a link group
669 int smc_buf_create(struct smc_sock *smc)
673 /* create send buffer */
674 rc = __smc_buf_create(smc, false);
678 rc = __smc_buf_create(smc, true);
680 smc_buf_free(smc->conn.sndbuf_desc,
681 &smc->conn.lgr->lnk[SMC_SINGLE_LINK], false);
685 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
689 for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
690 if (!test_and_set_bit(i, lgr->rtokens_used_mask))
696 /* save rkey and dma_addr received from peer during clc handshake */
697 int smc_rmb_rtoken_handling(struct smc_connection *conn,
698 struct smc_clc_msg_accept_confirm *clc)
700 u64 dma_addr = be64_to_cpu(clc->rmb_dma_addr);
701 struct smc_link_group *lgr = conn->lgr;
702 u32 rkey = ntohl(clc->rmb_rkey);
705 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
706 if ((lgr->rtokens[i][SMC_SINGLE_LINK].rkey == rkey) &&
707 (lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr == dma_addr) &&
708 test_bit(i, lgr->rtokens_used_mask)) {
709 conn->rtoken_idx = i;
713 conn->rtoken_idx = smc_rmb_reserve_rtoken_idx(lgr);
714 if (conn->rtoken_idx < 0)
715 return conn->rtoken_idx;
716 lgr->rtokens[conn->rtoken_idx][SMC_SINGLE_LINK].rkey = rkey;
717 lgr->rtokens[conn->rtoken_idx][SMC_SINGLE_LINK].dma_addr = dma_addr;