xref: /linux/net/smc/smc_core.c (revision fd639726bf15fca8ee1a00dce8e0096d0ad9bd18)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  *  Shared Memory Communications over RDMA (SMC-R) and RoCE
4  *
5  *  Basic Transport Functions exploiting Infiniband API
6  *
7  *  Copyright IBM Corp. 2016
8  *
9  *  Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
10  */
11 
12 #include <linux/socket.h>
13 #include <linux/if_vlan.h>
14 #include <linux/random.h>
15 #include <linux/workqueue.h>
16 #include <net/tcp.h>
17 #include <net/sock.h>
18 #include <rdma/ib_verbs.h>
19 
20 #include "smc.h"
21 #include "smc_clc.h"
22 #include "smc_core.h"
23 #include "smc_ib.h"
24 #include "smc_wr.h"
25 #include "smc_llc.h"
26 #include "smc_cdc.h"
27 #include "smc_close.h"
28 
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)
32 
33 static u32 smc_lgr_num;			/* unique link group number */
34 
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.
40  */
41 static void smc_lgr_add_alert_token(struct smc_connection *conn)
42 {
43 	struct rb_node **link, *parent = NULL;
44 	u32 token = conn->alert_token_local;
45 
46 	link = &conn->lgr->conns_all.rb_node;
47 	while (*link) {
48 		struct smc_connection *cur = rb_entry(*link,
49 					struct smc_connection, alert_node);
50 
51 		parent = *link;
52 		if (cur->alert_token_local > token)
53 			link = &parent->rb_left;
54 		else
55 			link = &parent->rb_right;
56 	}
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);
60 }
61 
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.
66  */
67 static void smc_lgr_register_conn(struct smc_connection *conn)
68 {
69 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
70 	static atomic_t nexttoken = ATOMIC_INIT(0);
71 
72 	/* find a new alert_token_local value not yet used by some connection
73 	 * in this link group
74 	 */
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;
80 	}
81 	smc_lgr_add_alert_token(conn);
82 	conn->lgr->conns_num++;
83 }
84 
85 /* Unregister connection and reset the alert token of the given connection<
86  */
87 static void __smc_lgr_unregister_conn(struct smc_connection *conn)
88 {
89 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
90 	struct smc_link_group *lgr = conn->lgr;
91 
92 	rb_erase(&conn->alert_node, &lgr->conns_all);
93 	lgr->conns_num--;
94 	conn->alert_token_local = 0;
95 	conn->lgr = NULL;
96 	sock_put(&smc->sk); /* sock_hold in smc_lgr_register_conn() */
97 }
98 
99 /* Unregister connection and trigger lgr freeing if applicable
100  */
101 static void smc_lgr_unregister_conn(struct smc_connection *conn)
102 {
103 	struct smc_link_group *lgr = conn->lgr;
104 	int reduced = 0;
105 
106 	write_lock_bh(&lgr->conns_lock);
107 	if (conn->alert_token_local) {
108 		reduced = 1;
109 		__smc_lgr_unregister_conn(conn);
110 	}
111 	write_unlock_bh(&lgr->conns_lock);
112 	if (!reduced || lgr->conns_num)
113 		return;
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.
117 	 */
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);
121 }
122 
123 static void smc_lgr_free_work(struct work_struct *work)
124 {
125 	struct smc_link_group *lgr = container_of(to_delayed_work(work),
126 						  struct smc_link_group,
127 						  free_work);
128 	bool conns;
129 
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);
136 		return;
137 	}
138 	list_del_init(&lgr->list); /* remove from smc_lgr_list */
139 	spin_unlock_bh(&smc_lgr_list.lock);
140 	smc_lgr_free(lgr);
141 }
142 
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)
147 {
148 	struct smc_link_group *lgr;
149 	struct smc_link *lnk;
150 	u8 rndvec[3];
151 	int rc = 0;
152 	int i;
153 
154 	lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
155 	if (!lgr) {
156 		rc = -ENOMEM;
157 		goto out;
158 	}
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]);
169 	}
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;
174 
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);
185 	if (rc)
186 		goto free_lgr;
187 	rc = smc_ib_create_protection_domain(lnk);
188 	if (rc)
189 		goto free_link_mem;
190 	rc = smc_ib_create_queue_pair(lnk);
191 	if (rc)
192 		goto dealloc_pd;
193 	rc = smc_wr_create_link(lnk);
194 	if (rc)
195 		goto destroy_qp;
196 	init_completion(&lnk->llc_confirm);
197 	init_completion(&lnk->llc_confirm_resp);
198 
199 	smc->conn.lgr = lgr;
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);
204 	return 0;
205 
206 destroy_qp:
207 	smc_ib_destroy_queue_pair(lnk);
208 dealloc_pd:
209 	smc_ib_dealloc_protection_domain(lnk);
210 free_link_mem:
211 	smc_wr_free_link_mem(lnk);
212 free_lgr:
213 	kfree(lgr);
214 out:
215 	return rc;
216 }
217 
218 static void smc_buf_unuse(struct smc_connection *conn)
219 {
220 	if (conn->sndbuf_desc) {
221 		conn->sndbuf_desc->used = 0;
222 		conn->sndbuf_size = 0;
223 	}
224 	if (conn->rmb_desc) {
225 		conn->rmb_desc->reused = true;
226 		conn->rmb_desc->used = 0;
227 		conn->rmbe_size = 0;
228 	}
229 }
230 
231 /* remove a finished connection from its link group */
232 void smc_conn_free(struct smc_connection *conn)
233 {
234 	struct smc_link_group *lgr = conn->lgr;
235 
236 	if (!lgr)
237 		return;
238 	smc_cdc_tx_dismiss_slots(conn);
239 	smc_lgr_unregister_conn(conn);
240 	smc_buf_unuse(conn);
241 }
242 
243 static void smc_link_clear(struct smc_link *lnk)
244 {
245 	lnk->peer_qpn = 0;
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);
251 }
252 
253 static void smc_buf_free(struct smc_buf_desc *buf_desc, struct smc_link *lnk,
254 			 bool is_rmb)
255 {
256 	if (is_rmb) {
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,
261 				    DMA_FROM_DEVICE);
262 	} else {
263 		smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
264 				    DMA_TO_DEVICE);
265 	}
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);
269 	kfree(buf_desc);
270 }
271 
272 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
273 {
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;
277 	int i;
278 
279 	for (i = 0; i < SMC_RMBE_SIZES; i++) {
280 		if (is_rmb)
281 			buf_list = &lgr->rmbs[i];
282 		else
283 			buf_list = &lgr->sndbufs[i];
284 		list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
285 					 list) {
286 			list_del(&buf_desc->list);
287 			smc_buf_free(buf_desc, lnk, is_rmb);
288 		}
289 	}
290 }
291 
292 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
293 {
294 	/* free send buffers */
295 	__smc_lgr_free_bufs(lgr, false);
296 	/* free rmbs */
297 	__smc_lgr_free_bufs(lgr, true);
298 }
299 
300 /* remove a link group */
301 void smc_lgr_free(struct smc_link_group *lgr)
302 {
303 	smc_lgr_free_bufs(lgr);
304 	smc_link_clear(&lgr->lnk[SMC_SINGLE_LINK]);
305 	kfree(lgr);
306 }
307 
308 /* terminate linkgroup abnormally */
309 void smc_lgr_terminate(struct smc_link_group *lgr)
310 {
311 	struct smc_connection *conn;
312 	struct smc_sock *smc;
313 	struct rb_node *node;
314 
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);
319 		return;
320 	}
321 	/* do not use this link group for new connections */
322 	list_del_init(&lgr->list);
323 	spin_unlock_bh(&smc_lgr_list.lock);
324 
325 	write_lock_bh(&lgr->conns_lock);
326 	node = rb_first(&lgr->conns_all);
327 	while (node) {
328 		conn = rb_entry(node, struct smc_connection, alert_node);
329 		smc = container_of(conn, struct smc_sock, conn);
330 		sock_hold(&smc->sk);
331 		__smc_lgr_unregister_conn(conn);
332 		schedule_work(&conn->close_work);
333 		sock_put(&smc->sk);
334 		node = rb_first(&lgr->conns_all);
335 	}
336 	write_unlock_bh(&lgr->conns_lock);
337 }
338 
339 /* Determine vlan of internal TCP socket.
340  * @vlan_id: address to store the determined vlan id into
341  */
342 static int smc_vlan_by_tcpsk(struct socket *clcsock, unsigned short *vlan_id)
343 {
344 	struct dst_entry *dst = sk_dst_get(clcsock->sk);
345 	int rc = 0;
346 
347 	*vlan_id = 0;
348 	if (!dst) {
349 		rc = -ENOTCONN;
350 		goto out;
351 	}
352 	if (!dst->dev) {
353 		rc = -ENODEV;
354 		goto out_rel;
355 	}
356 
357 	if (is_vlan_dev(dst->dev))
358 		*vlan_id = vlan_dev_vlan_id(dst->dev);
359 
360 out_rel:
361 	dst_release(dst);
362 out:
363 	return rc;
364 }
365 
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)
368 {
369 	struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
370 	struct ib_gid_attr gattr;
371 	union ib_gid gid;
372 	int i;
373 
374 	if (!lgr->vlan_id) {
375 		lnk->gid = lnk->smcibdev->gid[lnk->ibport - 1];
376 		return 0;
377 	}
378 
379 	for (i = 0; i < lnk->smcibdev->pattr[lnk->ibport - 1].gid_tbl_len;
380 	     i++) {
381 		if (ib_query_gid(lnk->smcibdev->ibdev, lnk->ibport, i, &gid,
382 				 &gattr))
383 			continue;
384 		if (gattr.ndev) {
385 			if (is_vlan_dev(gattr.ndev) &&
386 			    vlan_dev_vlan_id(gattr.ndev) == lgr->vlan_id) {
387 				lnk->gid = gid;
388 				dev_put(gattr.ndev);
389 				return 0;
390 			}
391 			dev_put(gattr.ndev);
392 		}
393 	}
394 	return -ENODEV;
395 }
396 
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)
401 {
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;
407 	int rc = 0;
408 
409 	role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
410 	rc = smc_vlan_by_tcpsk(smc->clcsock, &vlan_id);
411 	if (rc)
412 		return rc;
413 
414 	if ((role == SMC_CLNT) && srv_first_contact)
415 		/* create new link group as well */
416 		goto create;
417 
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,
423 			    SMC_SYSTEMID_LEN) &&
424 		    !memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_gid, &lcl->gid,
425 			    SMC_GID_SIZE) &&
426 		    !memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_mac, lcl->mac,
427 			    sizeof(lcl->mac)) &&
428 		    !lgr->sync_err &&
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;
435 			conn->lgr = lgr;
436 			smc_lgr_register_conn(conn); /* add smc conn to lgr */
437 			write_unlock_bh(&lgr->conns_lock);
438 			break;
439 		}
440 		write_unlock_bh(&lgr->conns_lock);
441 	}
442 	spin_unlock_bh(&smc_lgr_list.lock);
443 
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
447 		 * a new one
448 		 * send out_of_sync decline, reason synchr. error
449 		 */
450 		return -ENOLINK;
451 	}
452 
453 create:
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);
457 		if (rc)
458 			goto out;
459 		smc_lgr_register_conn(conn); /* add smc conn to lgr */
460 		rc = smc_link_determine_gid(conn->lgr);
461 	}
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);
466 #endif
467 
468 out:
469 	return rc ? rc : local_contact;
470 }
471 
472 /* try to reuse a sndbuf or rmb description slot for a certain
473  * buffer size; if not available, return NULL
474  */
475 static inline
476 struct smc_buf_desc *smc_buf_get_slot(struct smc_link_group *lgr,
477 				      int compressed_bufsize,
478 				      rwlock_t *lock,
479 				      struct list_head *buf_list)
480 {
481 	struct smc_buf_desc *buf_slot;
482 
483 	read_lock_bh(lock);
484 	list_for_each_entry(buf_slot, buf_list, list) {
485 		if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
486 			read_unlock_bh(lock);
487 			return buf_slot;
488 		}
489 	}
490 	read_unlock_bh(lock);
491 	return NULL;
492 }
493 
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]
497  */
498 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
499 {
500 	return min_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
501 }
502 
503 static struct smc_buf_desc *smc_new_buf_create(struct smc_link_group *lgr,
504 					       bool is_rmb, int bufsize)
505 {
506 	struct smc_buf_desc *buf_desc;
507 	struct smc_link *lnk;
508 	int rc;
509 
510 	/* try to alloc a new buffer */
511 	buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
512 	if (!buf_desc)
513 		return ERR_PTR(-ENOMEM);
514 
515 	buf_desc->cpu_addr =
516 		(void *)__get_free_pages(GFP_KERNEL | __GFP_NOWARN |
517 					 __GFP_NOMEMALLOC |
518 					 __GFP_NORETRY | __GFP_ZERO,
519 					 get_order(bufsize));
520 	if (!buf_desc->cpu_addr) {
521 		kfree(buf_desc);
522 		return ERR_PTR(-EAGAIN);
523 	}
524 	buf_desc->order = get_order(bufsize);
525 
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,
529 			    GFP_KERNEL);
530 	if (rc) {
531 		smc_buf_free(buf_desc, lnk, is_rmb);
532 		return ERR_PTR(rc);
533 	}
534 	sg_set_buf(buf_desc->sgt[SMC_SINGLE_LINK].sgl,
535 		   buf_desc->cpu_addr, bufsize);
536 
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 */
541 	if (rc != 1)  {
542 		smc_buf_free(buf_desc, lnk, is_rmb);
543 		return ERR_PTR(-EAGAIN);
544 	}
545 
546 	/* create a new memory region for the RMB */
547 	if (is_rmb) {
548 		rc = smc_ib_get_memory_region(lnk->roce_pd,
549 					      IB_ACCESS_REMOTE_WRITE |
550 					      IB_ACCESS_LOCAL_WRITE,
551 					      buf_desc);
552 		if (rc) {
553 			smc_buf_free(buf_desc, lnk, is_rmb);
554 			return ERR_PTR(rc);
555 		}
556 	}
557 
558 	return buf_desc;
559 }
560 
561 static int __smc_buf_create(struct smc_sock *smc, bool is_rmb)
562 {
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;
568 	int sk_buf_size;
569 	rwlock_t *lock;
570 
571 	if (is_rmb)
572 		/* use socket recv buffer size (w/o overhead) as start value */
573 		sk_buf_size = smc->sk.sk_rcvbuf / 2;
574 	else
575 		/* use socket send buffer size (w/o overhead) as start value */
576 		sk_buf_size = smc->sk.sk_sndbuf / 2;
577 
578 	for (bufsize_short = smc_compress_bufsize(sk_buf_size);
579 	     bufsize_short >= 0; bufsize_short--) {
580 
581 		if (is_rmb) {
582 			lock = &lgr->rmbs_lock;
583 			buf_list = &lgr->rmbs[bufsize_short];
584 		} else {
585 			lock = &lgr->sndbufs_lock;
586 			buf_list = &lgr->sndbufs[bufsize_short];
587 		}
588 		bufsize = smc_uncompress_bufsize(bufsize_short);
589 		if ((1 << get_order(bufsize)) > SG_MAX_SINGLE_ALLOC)
590 			continue;
591 
592 		/* check for reusable slot in the link group */
593 		buf_desc = smc_buf_get_slot(lgr, bufsize_short, lock, buf_list);
594 		if (buf_desc) {
595 			memset(buf_desc->cpu_addr, 0, bufsize);
596 			break; /* found reusable slot */
597 		}
598 
599 		buf_desc = smc_new_buf_create(lgr, is_rmb, bufsize);
600 		if (PTR_ERR(buf_desc) == -ENOMEM)
601 			break;
602 		if (IS_ERR(buf_desc))
603 			continue;
604 
605 		buf_desc->used = 1;
606 		write_lock_bh(lock);
607 		list_add(&buf_desc->list, buf_list);
608 		write_unlock_bh(lock);
609 		break; /* found */
610 	}
611 
612 	if (IS_ERR(buf_desc))
613 		return -ENOMEM;
614 
615 	if (is_rmb) {
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);
622 	} else {
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);
627 	}
628 	return 0;
629 }
630 
631 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn)
632 {
633 	struct smc_link_group *lgr = conn->lgr;
634 
635 	smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
636 			       conn->sndbuf_desc, DMA_TO_DEVICE);
637 }
638 
639 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
640 {
641 	struct smc_link_group *lgr = conn->lgr;
642 
643 	smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
644 				  conn->sndbuf_desc, DMA_TO_DEVICE);
645 }
646 
647 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
648 {
649 	struct smc_link_group *lgr = conn->lgr;
650 
651 	smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
652 			       conn->rmb_desc, DMA_FROM_DEVICE);
653 }
654 
655 void smc_rmb_sync_sg_for_device(struct smc_connection *conn)
656 {
657 	struct smc_link_group *lgr = conn->lgr;
658 
659 	smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
660 				  conn->rmb_desc, DMA_FROM_DEVICE);
661 }
662 
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
668  */
669 int smc_buf_create(struct smc_sock *smc)
670 {
671 	int rc;
672 
673 	/* create send buffer */
674 	rc = __smc_buf_create(smc, false);
675 	if (rc)
676 		return rc;
677 	/* create rmb */
678 	rc = __smc_buf_create(smc, true);
679 	if (rc)
680 		smc_buf_free(smc->conn.sndbuf_desc,
681 			     &smc->conn.lgr->lnk[SMC_SINGLE_LINK], false);
682 	return rc;
683 }
684 
685 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
686 {
687 	int i;
688 
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))
691 			return i;
692 	}
693 	return -ENOSPC;
694 }
695 
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)
699 {
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);
703 	int i;
704 
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;
710 			return 0;
711 		}
712 	}
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;
718 	return 0;
719 }
720