xref: /linux/net/smc/smc_core.c (revision a8b70ccf10e38775785d9cb12ead916474549f99)
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 static void smc_lgr_schedule_free_work(struct smc_link_group *lgr)
36 {
37 	/* client link group creation always follows the server link group
38 	 * creation. For client use a somewhat higher removal delay time,
39 	 * otherwise there is a risk of out-of-sync link groups.
40 	 */
41 	mod_delayed_work(system_wq, &lgr->free_work,
42 			 lgr->role == SMC_CLNT ? SMC_LGR_FREE_DELAY_CLNT :
43 						 SMC_LGR_FREE_DELAY_SERV);
44 }
45 
46 /* Register connection's alert token in our lookup structure.
47  * To use rbtrees we have to implement our own insert core.
48  * Requires @conns_lock
49  * @smc		connection to register
50  * Returns 0 on success, != otherwise.
51  */
52 static void smc_lgr_add_alert_token(struct smc_connection *conn)
53 {
54 	struct rb_node **link, *parent = NULL;
55 	u32 token = conn->alert_token_local;
56 
57 	link = &conn->lgr->conns_all.rb_node;
58 	while (*link) {
59 		struct smc_connection *cur = rb_entry(*link,
60 					struct smc_connection, alert_node);
61 
62 		parent = *link;
63 		if (cur->alert_token_local > token)
64 			link = &parent->rb_left;
65 		else
66 			link = &parent->rb_right;
67 	}
68 	/* Put the new node there */
69 	rb_link_node(&conn->alert_node, parent, link);
70 	rb_insert_color(&conn->alert_node, &conn->lgr->conns_all);
71 }
72 
73 /* Register connection in link group by assigning an alert token
74  * registered in a search tree.
75  * Requires @conns_lock
76  * Note that '0' is a reserved value and not assigned.
77  */
78 static void smc_lgr_register_conn(struct smc_connection *conn)
79 {
80 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
81 	static atomic_t nexttoken = ATOMIC_INIT(0);
82 
83 	/* find a new alert_token_local value not yet used by some connection
84 	 * in this link group
85 	 */
86 	sock_hold(&smc->sk); /* sock_put in smc_lgr_unregister_conn() */
87 	while (!conn->alert_token_local) {
88 		conn->alert_token_local = atomic_inc_return(&nexttoken);
89 		if (smc_lgr_find_conn(conn->alert_token_local, conn->lgr))
90 			conn->alert_token_local = 0;
91 	}
92 	smc_lgr_add_alert_token(conn);
93 	conn->lgr->conns_num++;
94 }
95 
96 /* Unregister connection and reset the alert token of the given connection<
97  */
98 static void __smc_lgr_unregister_conn(struct smc_connection *conn)
99 {
100 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
101 	struct smc_link_group *lgr = conn->lgr;
102 
103 	rb_erase(&conn->alert_node, &lgr->conns_all);
104 	lgr->conns_num--;
105 	conn->alert_token_local = 0;
106 	conn->lgr = NULL;
107 	sock_put(&smc->sk); /* sock_hold in smc_lgr_register_conn() */
108 }
109 
110 /* Unregister connection and trigger lgr freeing if applicable
111  */
112 static void smc_lgr_unregister_conn(struct smc_connection *conn)
113 {
114 	struct smc_link_group *lgr = conn->lgr;
115 	int reduced = 0;
116 
117 	write_lock_bh(&lgr->conns_lock);
118 	if (conn->alert_token_local) {
119 		reduced = 1;
120 		__smc_lgr_unregister_conn(conn);
121 	}
122 	write_unlock_bh(&lgr->conns_lock);
123 	if (!reduced || lgr->conns_num)
124 		return;
125 	smc_lgr_schedule_free_work(lgr);
126 }
127 
128 static void smc_lgr_free_work(struct work_struct *work)
129 {
130 	struct smc_link_group *lgr = container_of(to_delayed_work(work),
131 						  struct smc_link_group,
132 						  free_work);
133 	bool conns;
134 
135 	spin_lock_bh(&smc_lgr_list.lock);
136 	if (list_empty(&lgr->list))
137 		goto free;
138 	read_lock_bh(&lgr->conns_lock);
139 	conns = RB_EMPTY_ROOT(&lgr->conns_all);
140 	read_unlock_bh(&lgr->conns_lock);
141 	if (!conns) { /* number of lgr connections is no longer zero */
142 		spin_unlock_bh(&smc_lgr_list.lock);
143 		return;
144 	}
145 	list_del_init(&lgr->list); /* remove from smc_lgr_list */
146 free:
147 	spin_unlock_bh(&smc_lgr_list.lock);
148 	if (!delayed_work_pending(&lgr->free_work))
149 		smc_lgr_free(lgr);
150 }
151 
152 /* create a new SMC link group */
153 static int smc_lgr_create(struct smc_sock *smc,
154 			  struct smc_ib_device *smcibdev, u8 ibport,
155 			  char *peer_systemid, unsigned short vlan_id)
156 {
157 	struct smc_link_group *lgr;
158 	struct smc_link *lnk;
159 	u8 rndvec[3];
160 	int rc = 0;
161 	int i;
162 
163 	lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
164 	if (!lgr) {
165 		rc = -ENOMEM;
166 		goto out;
167 	}
168 	lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
169 	lgr->sync_err = false;
170 	memcpy(lgr->peer_systemid, peer_systemid, SMC_SYSTEMID_LEN);
171 	lgr->vlan_id = vlan_id;
172 	rwlock_init(&lgr->sndbufs_lock);
173 	rwlock_init(&lgr->rmbs_lock);
174 	for (i = 0; i < SMC_RMBE_SIZES; i++) {
175 		INIT_LIST_HEAD(&lgr->sndbufs[i]);
176 		INIT_LIST_HEAD(&lgr->rmbs[i]);
177 	}
178 	smc_lgr_num += SMC_LGR_NUM_INCR;
179 	memcpy(&lgr->id, (u8 *)&smc_lgr_num, SMC_LGR_ID_SIZE);
180 	INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
181 	lgr->conns_all = RB_ROOT;
182 
183 	lnk = &lgr->lnk[SMC_SINGLE_LINK];
184 	/* initialize link */
185 	lnk->state = SMC_LNK_ACTIVATING;
186 	lnk->link_id = SMC_SINGLE_LINK;
187 	lnk->smcibdev = smcibdev;
188 	lnk->ibport = ibport;
189 	lnk->path_mtu = smcibdev->pattr[ibport - 1].active_mtu;
190 	if (!smcibdev->initialized)
191 		smc_ib_setup_per_ibdev(smcibdev);
192 	get_random_bytes(rndvec, sizeof(rndvec));
193 	lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) + (rndvec[2] << 16);
194 	rc = smc_wr_alloc_link_mem(lnk);
195 	if (rc)
196 		goto free_lgr;
197 	rc = smc_ib_create_protection_domain(lnk);
198 	if (rc)
199 		goto free_link_mem;
200 	rc = smc_ib_create_queue_pair(lnk);
201 	if (rc)
202 		goto dealloc_pd;
203 	rc = smc_wr_create_link(lnk);
204 	if (rc)
205 		goto destroy_qp;
206 	init_completion(&lnk->llc_confirm);
207 	init_completion(&lnk->llc_confirm_resp);
208 	init_completion(&lnk->llc_add);
209 	init_completion(&lnk->llc_add_resp);
210 
211 	smc->conn.lgr = lgr;
212 	rwlock_init(&lgr->conns_lock);
213 	spin_lock_bh(&smc_lgr_list.lock);
214 	list_add(&lgr->list, &smc_lgr_list.list);
215 	spin_unlock_bh(&smc_lgr_list.lock);
216 	return 0;
217 
218 destroy_qp:
219 	smc_ib_destroy_queue_pair(lnk);
220 dealloc_pd:
221 	smc_ib_dealloc_protection_domain(lnk);
222 free_link_mem:
223 	smc_wr_free_link_mem(lnk);
224 free_lgr:
225 	kfree(lgr);
226 out:
227 	return rc;
228 }
229 
230 static void smc_buf_unuse(struct smc_connection *conn)
231 {
232 	if (conn->sndbuf_desc) {
233 		conn->sndbuf_desc->used = 0;
234 		conn->sndbuf_size = 0;
235 	}
236 	if (conn->rmb_desc) {
237 		conn->rmb_desc->reused = true;
238 		conn->rmb_desc->used = 0;
239 		conn->rmbe_size = 0;
240 	}
241 }
242 
243 /* remove a finished connection from its link group */
244 void smc_conn_free(struct smc_connection *conn)
245 {
246 	if (!conn->lgr)
247 		return;
248 	smc_cdc_tx_dismiss_slots(conn);
249 	smc_lgr_unregister_conn(conn);
250 	smc_buf_unuse(conn);
251 }
252 
253 static void smc_link_clear(struct smc_link *lnk)
254 {
255 	lnk->peer_qpn = 0;
256 	smc_ib_modify_qp_reset(lnk);
257 	smc_wr_free_link(lnk);
258 	smc_ib_destroy_queue_pair(lnk);
259 	smc_ib_dealloc_protection_domain(lnk);
260 	smc_wr_free_link_mem(lnk);
261 }
262 
263 static void smc_buf_free(struct smc_buf_desc *buf_desc, struct smc_link *lnk,
264 			 bool is_rmb)
265 {
266 	if (is_rmb) {
267 		if (buf_desc->mr_rx[SMC_SINGLE_LINK])
268 			smc_ib_put_memory_region(
269 					buf_desc->mr_rx[SMC_SINGLE_LINK]);
270 		smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
271 				    DMA_FROM_DEVICE);
272 	} else {
273 		smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
274 				    DMA_TO_DEVICE);
275 	}
276 	sg_free_table(&buf_desc->sgt[SMC_SINGLE_LINK]);
277 	if (buf_desc->cpu_addr)
278 		free_pages((unsigned long)buf_desc->cpu_addr, buf_desc->order);
279 	kfree(buf_desc);
280 }
281 
282 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
283 {
284 	struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
285 	struct smc_buf_desc *buf_desc, *bf_desc;
286 	struct list_head *buf_list;
287 	int i;
288 
289 	for (i = 0; i < SMC_RMBE_SIZES; i++) {
290 		if (is_rmb)
291 			buf_list = &lgr->rmbs[i];
292 		else
293 			buf_list = &lgr->sndbufs[i];
294 		list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
295 					 list) {
296 			list_del(&buf_desc->list);
297 			smc_buf_free(buf_desc, lnk, is_rmb);
298 		}
299 	}
300 }
301 
302 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
303 {
304 	/* free send buffers */
305 	__smc_lgr_free_bufs(lgr, false);
306 	/* free rmbs */
307 	__smc_lgr_free_bufs(lgr, true);
308 }
309 
310 /* remove a link group */
311 void smc_lgr_free(struct smc_link_group *lgr)
312 {
313 	smc_lgr_free_bufs(lgr);
314 	smc_link_clear(&lgr->lnk[SMC_SINGLE_LINK]);
315 	kfree(lgr);
316 }
317 
318 void smc_lgr_forget(struct smc_link_group *lgr)
319 {
320 	spin_lock_bh(&smc_lgr_list.lock);
321 	/* do not use this link group for new connections */
322 	if (!list_empty(&lgr->list))
323 		list_del_init(&lgr->list);
324 	spin_unlock_bh(&smc_lgr_list.lock);
325 }
326 
327 /* terminate linkgroup abnormally */
328 void smc_lgr_terminate(struct smc_link_group *lgr)
329 {
330 	struct smc_connection *conn;
331 	struct smc_sock *smc;
332 	struct rb_node *node;
333 
334 	smc_lgr_forget(lgr);
335 
336 	write_lock_bh(&lgr->conns_lock);
337 	node = rb_first(&lgr->conns_all);
338 	while (node) {
339 		conn = rb_entry(node, struct smc_connection, alert_node);
340 		smc = container_of(conn, struct smc_sock, conn);
341 		sock_hold(&smc->sk); /* sock_put in close work */
342 		conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
343 		__smc_lgr_unregister_conn(conn);
344 		write_unlock_bh(&lgr->conns_lock);
345 		if (!schedule_work(&conn->close_work))
346 			sock_put(&smc->sk);
347 		write_lock_bh(&lgr->conns_lock);
348 		node = rb_first(&lgr->conns_all);
349 	}
350 	write_unlock_bh(&lgr->conns_lock);
351 	wake_up(&lgr->lnk[SMC_SINGLE_LINK].wr_reg_wait);
352 	smc_lgr_schedule_free_work(lgr);
353 }
354 
355 /* Determine vlan of internal TCP socket.
356  * @vlan_id: address to store the determined vlan id into
357  */
358 static int smc_vlan_by_tcpsk(struct socket *clcsock, unsigned short *vlan_id)
359 {
360 	struct dst_entry *dst = sk_dst_get(clcsock->sk);
361 	int rc = 0;
362 
363 	*vlan_id = 0;
364 	if (!dst) {
365 		rc = -ENOTCONN;
366 		goto out;
367 	}
368 	if (!dst->dev) {
369 		rc = -ENODEV;
370 		goto out_rel;
371 	}
372 
373 	if (is_vlan_dev(dst->dev))
374 		*vlan_id = vlan_dev_vlan_id(dst->dev);
375 
376 out_rel:
377 	dst_release(dst);
378 out:
379 	return rc;
380 }
381 
382 /* determine the link gid matching the vlan id of the link group */
383 static int smc_link_determine_gid(struct smc_link_group *lgr)
384 {
385 	struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
386 	struct ib_gid_attr gattr;
387 	union ib_gid gid;
388 	int i;
389 
390 	if (!lgr->vlan_id) {
391 		lnk->gid = lnk->smcibdev->gid[lnk->ibport - 1];
392 		return 0;
393 	}
394 
395 	for (i = 0; i < lnk->smcibdev->pattr[lnk->ibport - 1].gid_tbl_len;
396 	     i++) {
397 		if (ib_query_gid(lnk->smcibdev->ibdev, lnk->ibport, i, &gid,
398 				 &gattr))
399 			continue;
400 		if (gattr.ndev) {
401 			if (is_vlan_dev(gattr.ndev) &&
402 			    vlan_dev_vlan_id(gattr.ndev) == lgr->vlan_id) {
403 				lnk->gid = gid;
404 				dev_put(gattr.ndev);
405 				return 0;
406 			}
407 			dev_put(gattr.ndev);
408 		}
409 	}
410 	return -ENODEV;
411 }
412 
413 /* create a new SMC connection (and a new link group if necessary) */
414 int smc_conn_create(struct smc_sock *smc,
415 		    struct smc_ib_device *smcibdev, u8 ibport,
416 		    struct smc_clc_msg_local *lcl, int srv_first_contact)
417 {
418 	struct smc_connection *conn = &smc->conn;
419 	struct smc_link_group *lgr;
420 	unsigned short vlan_id;
421 	enum smc_lgr_role role;
422 	int local_contact = SMC_FIRST_CONTACT;
423 	int rc = 0;
424 
425 	role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
426 	rc = smc_vlan_by_tcpsk(smc->clcsock, &vlan_id);
427 	if (rc)
428 		return rc;
429 
430 	if ((role == SMC_CLNT) && srv_first_contact)
431 		/* create new link group as well */
432 		goto create;
433 
434 	/* determine if an existing link group can be reused */
435 	spin_lock_bh(&smc_lgr_list.lock);
436 	list_for_each_entry(lgr, &smc_lgr_list.list, list) {
437 		write_lock_bh(&lgr->conns_lock);
438 		if (!memcmp(lgr->peer_systemid, lcl->id_for_peer,
439 			    SMC_SYSTEMID_LEN) &&
440 		    !memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_gid, &lcl->gid,
441 			    SMC_GID_SIZE) &&
442 		    !memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_mac, lcl->mac,
443 			    sizeof(lcl->mac)) &&
444 		    !lgr->sync_err &&
445 		    (lgr->role == role) &&
446 		    (lgr->vlan_id == vlan_id) &&
447 		    ((role == SMC_CLNT) ||
448 		     (lgr->conns_num < SMC_RMBS_PER_LGR_MAX))) {
449 			/* link group found */
450 			local_contact = SMC_REUSE_CONTACT;
451 			conn->lgr = lgr;
452 			smc_lgr_register_conn(conn); /* add smc conn to lgr */
453 			write_unlock_bh(&lgr->conns_lock);
454 			break;
455 		}
456 		write_unlock_bh(&lgr->conns_lock);
457 	}
458 	spin_unlock_bh(&smc_lgr_list.lock);
459 
460 	if (role == SMC_CLNT && !srv_first_contact &&
461 	    (local_contact == SMC_FIRST_CONTACT)) {
462 		/* Server reuses a link group, but Client wants to start
463 		 * a new one
464 		 * send out_of_sync decline, reason synchr. error
465 		 */
466 		return -ENOLINK;
467 	}
468 
469 create:
470 	if (local_contact == SMC_FIRST_CONTACT) {
471 		rc = smc_lgr_create(smc, smcibdev, ibport,
472 				    lcl->id_for_peer, vlan_id);
473 		if (rc)
474 			goto out;
475 		smc_lgr_register_conn(conn); /* add smc conn to lgr */
476 		rc = smc_link_determine_gid(conn->lgr);
477 	}
478 	conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
479 	conn->local_tx_ctrl.len = SMC_WR_TX_SIZE;
480 #ifndef KERNEL_HAS_ATOMIC64
481 	spin_lock_init(&conn->acurs_lock);
482 #endif
483 
484 out:
485 	return rc ? rc : local_contact;
486 }
487 
488 /* try to reuse a sndbuf or rmb description slot for a certain
489  * buffer size; if not available, return NULL
490  */
491 static inline
492 struct smc_buf_desc *smc_buf_get_slot(struct smc_link_group *lgr,
493 				      int compressed_bufsize,
494 				      rwlock_t *lock,
495 				      struct list_head *buf_list)
496 {
497 	struct smc_buf_desc *buf_slot;
498 
499 	read_lock_bh(lock);
500 	list_for_each_entry(buf_slot, buf_list, list) {
501 		if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
502 			read_unlock_bh(lock);
503 			return buf_slot;
504 		}
505 	}
506 	read_unlock_bh(lock);
507 	return NULL;
508 }
509 
510 /* one of the conditions for announcing a receiver's current window size is
511  * that it "results in a minimum increase in the window size of 10% of the
512  * receive buffer space" [RFC7609]
513  */
514 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
515 {
516 	return min_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
517 }
518 
519 static struct smc_buf_desc *smc_new_buf_create(struct smc_link_group *lgr,
520 					       bool is_rmb, int bufsize)
521 {
522 	struct smc_buf_desc *buf_desc;
523 	struct smc_link *lnk;
524 	int rc;
525 
526 	/* try to alloc a new buffer */
527 	buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
528 	if (!buf_desc)
529 		return ERR_PTR(-ENOMEM);
530 
531 	buf_desc->cpu_addr =
532 		(void *)__get_free_pages(GFP_KERNEL | __GFP_NOWARN |
533 					 __GFP_NOMEMALLOC |
534 					 __GFP_NORETRY | __GFP_ZERO,
535 					 get_order(bufsize));
536 	if (!buf_desc->cpu_addr) {
537 		kfree(buf_desc);
538 		return ERR_PTR(-EAGAIN);
539 	}
540 	buf_desc->order = get_order(bufsize);
541 
542 	/* build the sg table from the pages */
543 	lnk = &lgr->lnk[SMC_SINGLE_LINK];
544 	rc = sg_alloc_table(&buf_desc->sgt[SMC_SINGLE_LINK], 1,
545 			    GFP_KERNEL);
546 	if (rc) {
547 		smc_buf_free(buf_desc, lnk, is_rmb);
548 		return ERR_PTR(rc);
549 	}
550 	sg_set_buf(buf_desc->sgt[SMC_SINGLE_LINK].sgl,
551 		   buf_desc->cpu_addr, bufsize);
552 
553 	/* map sg table to DMA address */
554 	rc = smc_ib_buf_map_sg(lnk->smcibdev, buf_desc,
555 			       is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
556 	/* SMC protocol depends on mapping to one DMA address only */
557 	if (rc != 1)  {
558 		smc_buf_free(buf_desc, lnk, is_rmb);
559 		return ERR_PTR(-EAGAIN);
560 	}
561 
562 	/* create a new memory region for the RMB */
563 	if (is_rmb) {
564 		rc = smc_ib_get_memory_region(lnk->roce_pd,
565 					      IB_ACCESS_REMOTE_WRITE |
566 					      IB_ACCESS_LOCAL_WRITE,
567 					      buf_desc);
568 		if (rc) {
569 			smc_buf_free(buf_desc, lnk, is_rmb);
570 			return ERR_PTR(rc);
571 		}
572 	}
573 
574 	return buf_desc;
575 }
576 
577 static int __smc_buf_create(struct smc_sock *smc, bool is_rmb)
578 {
579 	struct smc_connection *conn = &smc->conn;
580 	struct smc_link_group *lgr = conn->lgr;
581 	struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
582 	struct list_head *buf_list;
583 	int bufsize, bufsize_short;
584 	int sk_buf_size;
585 	rwlock_t *lock;
586 
587 	if (is_rmb)
588 		/* use socket recv buffer size (w/o overhead) as start value */
589 		sk_buf_size = smc->sk.sk_rcvbuf / 2;
590 	else
591 		/* use socket send buffer size (w/o overhead) as start value */
592 		sk_buf_size = smc->sk.sk_sndbuf / 2;
593 
594 	for (bufsize_short = smc_compress_bufsize(sk_buf_size);
595 	     bufsize_short >= 0; bufsize_short--) {
596 
597 		if (is_rmb) {
598 			lock = &lgr->rmbs_lock;
599 			buf_list = &lgr->rmbs[bufsize_short];
600 		} else {
601 			lock = &lgr->sndbufs_lock;
602 			buf_list = &lgr->sndbufs[bufsize_short];
603 		}
604 		bufsize = smc_uncompress_bufsize(bufsize_short);
605 		if ((1 << get_order(bufsize)) > SG_MAX_SINGLE_ALLOC)
606 			continue;
607 
608 		/* check for reusable slot in the link group */
609 		buf_desc = smc_buf_get_slot(lgr, bufsize_short, lock, buf_list);
610 		if (buf_desc) {
611 			memset(buf_desc->cpu_addr, 0, bufsize);
612 			break; /* found reusable slot */
613 		}
614 
615 		buf_desc = smc_new_buf_create(lgr, is_rmb, bufsize);
616 		if (PTR_ERR(buf_desc) == -ENOMEM)
617 			break;
618 		if (IS_ERR(buf_desc))
619 			continue;
620 
621 		buf_desc->used = 1;
622 		write_lock_bh(lock);
623 		list_add(&buf_desc->list, buf_list);
624 		write_unlock_bh(lock);
625 		break; /* found */
626 	}
627 
628 	if (IS_ERR(buf_desc))
629 		return -ENOMEM;
630 
631 	if (is_rmb) {
632 		conn->rmb_desc = buf_desc;
633 		conn->rmbe_size = bufsize;
634 		conn->rmbe_size_short = bufsize_short;
635 		smc->sk.sk_rcvbuf = bufsize * 2;
636 		atomic_set(&conn->bytes_to_rcv, 0);
637 		conn->rmbe_update_limit = smc_rmb_wnd_update_limit(bufsize);
638 	} else {
639 		conn->sndbuf_desc = buf_desc;
640 		conn->sndbuf_size = bufsize;
641 		smc->sk.sk_sndbuf = bufsize * 2;
642 		atomic_set(&conn->sndbuf_space, bufsize);
643 	}
644 	return 0;
645 }
646 
647 void smc_sndbuf_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->sndbuf_desc, DMA_TO_DEVICE);
653 }
654 
655 void smc_sndbuf_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->sndbuf_desc, DMA_TO_DEVICE);
661 }
662 
663 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
664 {
665 	struct smc_link_group *lgr = conn->lgr;
666 
667 	smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
668 			       conn->rmb_desc, DMA_FROM_DEVICE);
669 }
670 
671 void smc_rmb_sync_sg_for_device(struct smc_connection *conn)
672 {
673 	struct smc_link_group *lgr = conn->lgr;
674 
675 	smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
676 				  conn->rmb_desc, DMA_FROM_DEVICE);
677 }
678 
679 /* create the send and receive buffer for an SMC socket;
680  * receive buffers are called RMBs;
681  * (even though the SMC protocol allows more than one RMB-element per RMB,
682  * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
683  * extra RMB for every connection in a link group
684  */
685 int smc_buf_create(struct smc_sock *smc)
686 {
687 	int rc;
688 
689 	/* create send buffer */
690 	rc = __smc_buf_create(smc, false);
691 	if (rc)
692 		return rc;
693 	/* create rmb */
694 	rc = __smc_buf_create(smc, true);
695 	if (rc)
696 		smc_buf_free(smc->conn.sndbuf_desc,
697 			     &smc->conn.lgr->lnk[SMC_SINGLE_LINK], false);
698 	return rc;
699 }
700 
701 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
702 {
703 	int i;
704 
705 	for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
706 		if (!test_and_set_bit(i, lgr->rtokens_used_mask))
707 			return i;
708 	}
709 	return -ENOSPC;
710 }
711 
712 /* add a new rtoken from peer */
713 int smc_rtoken_add(struct smc_link_group *lgr, __be64 nw_vaddr, __be32 nw_rkey)
714 {
715 	u64 dma_addr = be64_to_cpu(nw_vaddr);
716 	u32 rkey = ntohl(nw_rkey);
717 	int i;
718 
719 	for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
720 		if ((lgr->rtokens[i][SMC_SINGLE_LINK].rkey == rkey) &&
721 		    (lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr == dma_addr) &&
722 		    test_bit(i, lgr->rtokens_used_mask)) {
723 			/* already in list */
724 			return i;
725 		}
726 	}
727 	i = smc_rmb_reserve_rtoken_idx(lgr);
728 	if (i < 0)
729 		return i;
730 	lgr->rtokens[i][SMC_SINGLE_LINK].rkey = rkey;
731 	lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr = dma_addr;
732 	return i;
733 }
734 
735 /* delete an rtoken */
736 int smc_rtoken_delete(struct smc_link_group *lgr, __be32 nw_rkey)
737 {
738 	u32 rkey = ntohl(nw_rkey);
739 	int i;
740 
741 	for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
742 		if (lgr->rtokens[i][SMC_SINGLE_LINK].rkey == rkey &&
743 		    test_bit(i, lgr->rtokens_used_mask)) {
744 			lgr->rtokens[i][SMC_SINGLE_LINK].rkey = 0;
745 			lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr = 0;
746 
747 			clear_bit(i, lgr->rtokens_used_mask);
748 			return 0;
749 		}
750 	}
751 	return -ENOENT;
752 }
753 
754 /* save rkey and dma_addr received from peer during clc handshake */
755 int smc_rmb_rtoken_handling(struct smc_connection *conn,
756 			    struct smc_clc_msg_accept_confirm *clc)
757 {
758 	conn->rtoken_idx = smc_rtoken_add(conn->lgr, clc->rmb_dma_addr,
759 					  clc->rmb_rkey);
760 	if (conn->rtoken_idx < 0)
761 		return conn->rtoken_idx;
762 	return 0;
763 }
764