xref: /openbmc/linux/net/smc/smc_core.c (revision 59b4412f)
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 <linux/wait.h>
17 #include <linux/reboot.h>
18 #include <net/tcp.h>
19 #include <net/sock.h>
20 #include <rdma/ib_verbs.h>
21 #include <rdma/ib_cache.h>
22 
23 #include "smc.h"
24 #include "smc_clc.h"
25 #include "smc_core.h"
26 #include "smc_ib.h"
27 #include "smc_wr.h"
28 #include "smc_llc.h"
29 #include "smc_cdc.h"
30 #include "smc_close.h"
31 #include "smc_ism.h"
32 
33 #define SMC_LGR_NUM_INCR		256
34 #define SMC_LGR_FREE_DELAY_SERV		(600 * HZ)
35 #define SMC_LGR_FREE_DELAY_CLNT		(SMC_LGR_FREE_DELAY_SERV + 10 * HZ)
36 #define SMC_LGR_FREE_DELAY_FAST		(8 * HZ)
37 
38 static struct smc_lgr_list smc_lgr_list = {	/* established link groups */
39 	.lock = __SPIN_LOCK_UNLOCKED(smc_lgr_list.lock),
40 	.list = LIST_HEAD_INIT(smc_lgr_list.list),
41 	.num = 0,
42 };
43 
44 static atomic_t lgr_cnt = ATOMIC_INIT(0); /* number of existing link groups */
45 static DECLARE_WAIT_QUEUE_HEAD(lgrs_deleted);
46 
47 struct smc_ib_up_work {
48 	struct work_struct	work;
49 	struct smc_link_group	*lgr;
50 	struct smc_ib_device	*smcibdev;
51 	u8			ibport;
52 };
53 
54 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
55 			 struct smc_buf_desc *buf_desc);
56 static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft);
57 
58 static void smc_link_up_work(struct work_struct *work);
59 static void smc_link_down_work(struct work_struct *work);
60 
61 /* return head of link group list and its lock for a given link group */
62 static inline struct list_head *smc_lgr_list_head(struct smc_link_group *lgr,
63 						  spinlock_t **lgr_lock)
64 {
65 	if (lgr->is_smcd) {
66 		*lgr_lock = &lgr->smcd->lgr_lock;
67 		return &lgr->smcd->lgr_list;
68 	}
69 
70 	*lgr_lock = &smc_lgr_list.lock;
71 	return &smc_lgr_list.list;
72 }
73 
74 static void smc_lgr_schedule_free_work(struct smc_link_group *lgr)
75 {
76 	/* client link group creation always follows the server link group
77 	 * creation. For client use a somewhat higher removal delay time,
78 	 * otherwise there is a risk of out-of-sync link groups.
79 	 */
80 	if (!lgr->freeing && !lgr->freefast) {
81 		mod_delayed_work(system_wq, &lgr->free_work,
82 				 (!lgr->is_smcd && lgr->role == SMC_CLNT) ?
83 						SMC_LGR_FREE_DELAY_CLNT :
84 						SMC_LGR_FREE_DELAY_SERV);
85 	}
86 }
87 
88 void smc_lgr_schedule_free_work_fast(struct smc_link_group *lgr)
89 {
90 	if (!lgr->freeing && !lgr->freefast) {
91 		lgr->freefast = 1;
92 		mod_delayed_work(system_wq, &lgr->free_work,
93 				 SMC_LGR_FREE_DELAY_FAST);
94 	}
95 }
96 
97 /* Register connection's alert token in our lookup structure.
98  * To use rbtrees we have to implement our own insert core.
99  * Requires @conns_lock
100  * @smc		connection to register
101  * Returns 0 on success, != otherwise.
102  */
103 static void smc_lgr_add_alert_token(struct smc_connection *conn)
104 {
105 	struct rb_node **link, *parent = NULL;
106 	u32 token = conn->alert_token_local;
107 
108 	link = &conn->lgr->conns_all.rb_node;
109 	while (*link) {
110 		struct smc_connection *cur = rb_entry(*link,
111 					struct smc_connection, alert_node);
112 
113 		parent = *link;
114 		if (cur->alert_token_local > token)
115 			link = &parent->rb_left;
116 		else
117 			link = &parent->rb_right;
118 	}
119 	/* Put the new node there */
120 	rb_link_node(&conn->alert_node, parent, link);
121 	rb_insert_color(&conn->alert_node, &conn->lgr->conns_all);
122 }
123 
124 /* assign an SMC-R link to the connection */
125 static int smcr_lgr_conn_assign_link(struct smc_connection *conn, bool first)
126 {
127 	enum smc_link_state expected = first ? SMC_LNK_ACTIVATING :
128 				       SMC_LNK_ACTIVE;
129 	int i, j;
130 
131 	/* do link balancing */
132 	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
133 		struct smc_link *lnk = &conn->lgr->lnk[i];
134 
135 		if (lnk->state != expected || lnk->link_is_asym)
136 			continue;
137 		if (conn->lgr->role == SMC_CLNT) {
138 			conn->lnk = lnk; /* temporary, SMC server assigns link*/
139 			break;
140 		}
141 		if (conn->lgr->conns_num % 2) {
142 			for (j = i + 1; j < SMC_LINKS_PER_LGR_MAX; j++) {
143 				struct smc_link *lnk2;
144 
145 				lnk2 = &conn->lgr->lnk[j];
146 				if (lnk2->state == expected &&
147 				    !lnk2->link_is_asym) {
148 					conn->lnk = lnk2;
149 					break;
150 				}
151 			}
152 		}
153 		if (!conn->lnk)
154 			conn->lnk = lnk;
155 		break;
156 	}
157 	if (!conn->lnk)
158 		return SMC_CLC_DECL_NOACTLINK;
159 	return 0;
160 }
161 
162 /* Register connection in link group by assigning an alert token
163  * registered in a search tree.
164  * Requires @conns_lock
165  * Note that '0' is a reserved value and not assigned.
166  */
167 static int smc_lgr_register_conn(struct smc_connection *conn, bool first)
168 {
169 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
170 	static atomic_t nexttoken = ATOMIC_INIT(0);
171 	int rc;
172 
173 	if (!conn->lgr->is_smcd) {
174 		rc = smcr_lgr_conn_assign_link(conn, first);
175 		if (rc)
176 			return rc;
177 	}
178 	/* find a new alert_token_local value not yet used by some connection
179 	 * in this link group
180 	 */
181 	sock_hold(&smc->sk); /* sock_put in smc_lgr_unregister_conn() */
182 	while (!conn->alert_token_local) {
183 		conn->alert_token_local = atomic_inc_return(&nexttoken);
184 		if (smc_lgr_find_conn(conn->alert_token_local, conn->lgr))
185 			conn->alert_token_local = 0;
186 	}
187 	smc_lgr_add_alert_token(conn);
188 	conn->lgr->conns_num++;
189 	return 0;
190 }
191 
192 /* Unregister connection and reset the alert token of the given connection<
193  */
194 static void __smc_lgr_unregister_conn(struct smc_connection *conn)
195 {
196 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
197 	struct smc_link_group *lgr = conn->lgr;
198 
199 	rb_erase(&conn->alert_node, &lgr->conns_all);
200 	lgr->conns_num--;
201 	conn->alert_token_local = 0;
202 	sock_put(&smc->sk); /* sock_hold in smc_lgr_register_conn() */
203 }
204 
205 /* Unregister connection from lgr
206  */
207 static void smc_lgr_unregister_conn(struct smc_connection *conn)
208 {
209 	struct smc_link_group *lgr = conn->lgr;
210 
211 	if (!lgr)
212 		return;
213 	write_lock_bh(&lgr->conns_lock);
214 	if (conn->alert_token_local) {
215 		__smc_lgr_unregister_conn(conn);
216 	}
217 	write_unlock_bh(&lgr->conns_lock);
218 	conn->lgr = NULL;
219 }
220 
221 void smc_lgr_cleanup_early(struct smc_connection *conn)
222 {
223 	struct smc_link_group *lgr = conn->lgr;
224 	struct list_head *lgr_list;
225 	spinlock_t *lgr_lock;
226 
227 	if (!lgr)
228 		return;
229 
230 	smc_conn_free(conn);
231 	lgr_list = smc_lgr_list_head(lgr, &lgr_lock);
232 	spin_lock_bh(lgr_lock);
233 	/* do not use this link group for new connections */
234 	if (!list_empty(lgr_list))
235 		list_del_init(lgr_list);
236 	spin_unlock_bh(lgr_lock);
237 	smc_lgr_schedule_free_work_fast(lgr);
238 }
239 
240 static void smcr_lgr_link_deactivate_all(struct smc_link_group *lgr)
241 {
242 	int i;
243 
244 	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
245 		struct smc_link *lnk = &lgr->lnk[i];
246 
247 		if (smc_link_usable(lnk))
248 			lnk->state = SMC_LNK_INACTIVE;
249 	}
250 	wake_up_interruptible_all(&lgr->llc_waiter);
251 }
252 
253 static void smc_lgr_free(struct smc_link_group *lgr);
254 
255 static void smc_lgr_free_work(struct work_struct *work)
256 {
257 	struct smc_link_group *lgr = container_of(to_delayed_work(work),
258 						  struct smc_link_group,
259 						  free_work);
260 	spinlock_t *lgr_lock;
261 	bool conns;
262 
263 	smc_lgr_list_head(lgr, &lgr_lock);
264 	spin_lock_bh(lgr_lock);
265 	if (lgr->freeing) {
266 		spin_unlock_bh(lgr_lock);
267 		return;
268 	}
269 	read_lock_bh(&lgr->conns_lock);
270 	conns = RB_EMPTY_ROOT(&lgr->conns_all);
271 	read_unlock_bh(&lgr->conns_lock);
272 	if (!conns) { /* number of lgr connections is no longer zero */
273 		spin_unlock_bh(lgr_lock);
274 		return;
275 	}
276 	list_del_init(&lgr->list); /* remove from smc_lgr_list */
277 	lgr->freeing = 1; /* this instance does the freeing, no new schedule */
278 	spin_unlock_bh(lgr_lock);
279 	cancel_delayed_work(&lgr->free_work);
280 
281 	if (!lgr->is_smcd && !lgr->terminating)
282 		smc_llc_send_link_delete_all(lgr, true,
283 					     SMC_LLC_DEL_PROG_INIT_TERM);
284 	if (lgr->is_smcd && !lgr->terminating)
285 		smc_ism_signal_shutdown(lgr);
286 	if (!lgr->is_smcd)
287 		smcr_lgr_link_deactivate_all(lgr);
288 	smc_lgr_free(lgr);
289 }
290 
291 static void smc_lgr_terminate_work(struct work_struct *work)
292 {
293 	struct smc_link_group *lgr = container_of(work, struct smc_link_group,
294 						  terminate_work);
295 
296 	__smc_lgr_terminate(lgr, true);
297 }
298 
299 /* return next unique link id for the lgr */
300 static u8 smcr_next_link_id(struct smc_link_group *lgr)
301 {
302 	u8 link_id;
303 	int i;
304 
305 	while (1) {
306 		link_id = ++lgr->next_link_id;
307 		if (!link_id)	/* skip zero as link_id */
308 			link_id = ++lgr->next_link_id;
309 		for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
310 			if (smc_link_usable(&lgr->lnk[i]) &&
311 			    lgr->lnk[i].link_id == link_id)
312 				continue;
313 		}
314 		break;
315 	}
316 	return link_id;
317 }
318 
319 int smcr_link_init(struct smc_link_group *lgr, struct smc_link *lnk,
320 		   u8 link_idx, struct smc_init_info *ini)
321 {
322 	u8 rndvec[3];
323 	int rc;
324 
325 	get_device(&ini->ib_dev->ibdev->dev);
326 	atomic_inc(&ini->ib_dev->lnk_cnt);
327 	lnk->state = SMC_LNK_ACTIVATING;
328 	lnk->link_id = smcr_next_link_id(lgr);
329 	lnk->lgr = lgr;
330 	lnk->link_idx = link_idx;
331 	lnk->smcibdev = ini->ib_dev;
332 	lnk->ibport = ini->ib_port;
333 	lnk->path_mtu = ini->ib_dev->pattr[ini->ib_port - 1].active_mtu;
334 	smc_llc_link_set_uid(lnk);
335 	INIT_WORK(&lnk->link_down_wrk, smc_link_down_work);
336 	if (!ini->ib_dev->initialized) {
337 		rc = (int)smc_ib_setup_per_ibdev(ini->ib_dev);
338 		if (rc)
339 			goto out;
340 	}
341 	get_random_bytes(rndvec, sizeof(rndvec));
342 	lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) +
343 		(rndvec[2] << 16);
344 	rc = smc_ib_determine_gid(lnk->smcibdev, lnk->ibport,
345 				  ini->vlan_id, lnk->gid, &lnk->sgid_index);
346 	if (rc)
347 		goto out;
348 	rc = smc_llc_link_init(lnk);
349 	if (rc)
350 		goto out;
351 	rc = smc_wr_alloc_link_mem(lnk);
352 	if (rc)
353 		goto clear_llc_lnk;
354 	rc = smc_ib_create_protection_domain(lnk);
355 	if (rc)
356 		goto free_link_mem;
357 	rc = smc_ib_create_queue_pair(lnk);
358 	if (rc)
359 		goto dealloc_pd;
360 	rc = smc_wr_create_link(lnk);
361 	if (rc)
362 		goto destroy_qp;
363 	return 0;
364 
365 destroy_qp:
366 	smc_ib_destroy_queue_pair(lnk);
367 dealloc_pd:
368 	smc_ib_dealloc_protection_domain(lnk);
369 free_link_mem:
370 	smc_wr_free_link_mem(lnk);
371 clear_llc_lnk:
372 	smc_llc_link_clear(lnk, false);
373 out:
374 	put_device(&ini->ib_dev->ibdev->dev);
375 	memset(lnk, 0, sizeof(struct smc_link));
376 	lnk->state = SMC_LNK_UNUSED;
377 	if (!atomic_dec_return(&ini->ib_dev->lnk_cnt))
378 		wake_up(&ini->ib_dev->lnks_deleted);
379 	return rc;
380 }
381 
382 /* create a new SMC link group */
383 static int smc_lgr_create(struct smc_sock *smc, struct smc_init_info *ini)
384 {
385 	struct smc_link_group *lgr;
386 	struct list_head *lgr_list;
387 	struct smc_link *lnk;
388 	spinlock_t *lgr_lock;
389 	u8 link_idx;
390 	int rc = 0;
391 	int i;
392 
393 	if (ini->is_smcd && ini->vlan_id) {
394 		if (smc_ism_get_vlan(ini->ism_dev, ini->vlan_id)) {
395 			rc = SMC_CLC_DECL_ISMVLANERR;
396 			goto out;
397 		}
398 	}
399 
400 	lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
401 	if (!lgr) {
402 		rc = SMC_CLC_DECL_MEM;
403 		goto ism_put_vlan;
404 	}
405 	lgr->is_smcd = ini->is_smcd;
406 	lgr->sync_err = 0;
407 	lgr->terminating = 0;
408 	lgr->freefast = 0;
409 	lgr->freeing = 0;
410 	lgr->vlan_id = ini->vlan_id;
411 	mutex_init(&lgr->sndbufs_lock);
412 	mutex_init(&lgr->rmbs_lock);
413 	rwlock_init(&lgr->conns_lock);
414 	for (i = 0; i < SMC_RMBE_SIZES; i++) {
415 		INIT_LIST_HEAD(&lgr->sndbufs[i]);
416 		INIT_LIST_HEAD(&lgr->rmbs[i]);
417 	}
418 	lgr->next_link_id = 0;
419 	smc_lgr_list.num += SMC_LGR_NUM_INCR;
420 	memcpy(&lgr->id, (u8 *)&smc_lgr_list.num, SMC_LGR_ID_SIZE);
421 	INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
422 	INIT_WORK(&lgr->terminate_work, smc_lgr_terminate_work);
423 	lgr->conns_all = RB_ROOT;
424 	if (ini->is_smcd) {
425 		/* SMC-D specific settings */
426 		get_device(&ini->ism_dev->dev);
427 		lgr->peer_gid = ini->ism_gid;
428 		lgr->smcd = ini->ism_dev;
429 		lgr_list = &ini->ism_dev->lgr_list;
430 		lgr_lock = &lgr->smcd->lgr_lock;
431 		lgr->peer_shutdown = 0;
432 		atomic_inc(&ini->ism_dev->lgr_cnt);
433 	} else {
434 		/* SMC-R specific settings */
435 		lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
436 		memcpy(lgr->peer_systemid, ini->ib_lcl->id_for_peer,
437 		       SMC_SYSTEMID_LEN);
438 		memcpy(lgr->pnet_id, ini->ib_dev->pnetid[ini->ib_port - 1],
439 		       SMC_MAX_PNETID_LEN);
440 		smc_llc_lgr_init(lgr, smc);
441 
442 		link_idx = SMC_SINGLE_LINK;
443 		lnk = &lgr->lnk[link_idx];
444 		rc = smcr_link_init(lgr, lnk, link_idx, ini);
445 		if (rc)
446 			goto free_lgr;
447 		lgr_list = &smc_lgr_list.list;
448 		lgr_lock = &smc_lgr_list.lock;
449 		atomic_inc(&lgr_cnt);
450 	}
451 	smc->conn.lgr = lgr;
452 	spin_lock_bh(lgr_lock);
453 	list_add(&lgr->list, lgr_list);
454 	spin_unlock_bh(lgr_lock);
455 	return 0;
456 
457 free_lgr:
458 	kfree(lgr);
459 ism_put_vlan:
460 	if (ini->is_smcd && ini->vlan_id)
461 		smc_ism_put_vlan(ini->ism_dev, ini->vlan_id);
462 out:
463 	if (rc < 0) {
464 		if (rc == -ENOMEM)
465 			rc = SMC_CLC_DECL_MEM;
466 		else
467 			rc = SMC_CLC_DECL_INTERR;
468 	}
469 	return rc;
470 }
471 
472 static int smc_write_space(struct smc_connection *conn)
473 {
474 	int buffer_len = conn->peer_rmbe_size;
475 	union smc_host_cursor prod;
476 	union smc_host_cursor cons;
477 	int space;
478 
479 	smc_curs_copy(&prod, &conn->local_tx_ctrl.prod, conn);
480 	smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
481 	/* determine rx_buf space */
482 	space = buffer_len - smc_curs_diff(buffer_len, &cons, &prod);
483 	return space;
484 }
485 
486 static int smc_switch_cursor(struct smc_sock *smc, struct smc_cdc_tx_pend *pend,
487 			     struct smc_wr_buf *wr_buf)
488 {
489 	struct smc_connection *conn = &smc->conn;
490 	union smc_host_cursor cons, fin;
491 	int rc = 0;
492 	int diff;
493 
494 	smc_curs_copy(&conn->tx_curs_sent, &conn->tx_curs_fin, conn);
495 	smc_curs_copy(&fin, &conn->local_tx_ctrl_fin, conn);
496 	/* set prod cursor to old state, enforce tx_rdma_writes() */
497 	smc_curs_copy(&conn->local_tx_ctrl.prod, &fin, conn);
498 	smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
499 
500 	if (smc_curs_comp(conn->peer_rmbe_size, &cons, &fin) < 0) {
501 		/* cons cursor advanced more than fin, and prod was set
502 		 * fin above, so now prod is smaller than cons. Fix that.
503 		 */
504 		diff = smc_curs_diff(conn->peer_rmbe_size, &fin, &cons);
505 		smc_curs_add(conn->sndbuf_desc->len,
506 			     &conn->tx_curs_sent, diff);
507 		smc_curs_add(conn->sndbuf_desc->len,
508 			     &conn->tx_curs_fin, diff);
509 
510 		smp_mb__before_atomic();
511 		atomic_add(diff, &conn->sndbuf_space);
512 		smp_mb__after_atomic();
513 
514 		smc_curs_add(conn->peer_rmbe_size,
515 			     &conn->local_tx_ctrl.prod, diff);
516 		smc_curs_add(conn->peer_rmbe_size,
517 			     &conn->local_tx_ctrl_fin, diff);
518 	}
519 	/* recalculate, value is used by tx_rdma_writes() */
520 	atomic_set(&smc->conn.peer_rmbe_space, smc_write_space(conn));
521 
522 	if (smc->sk.sk_state != SMC_INIT &&
523 	    smc->sk.sk_state != SMC_CLOSED) {
524 		rc = smcr_cdc_msg_send_validation(conn, pend, wr_buf);
525 		if (!rc) {
526 			schedule_delayed_work(&conn->tx_work, 0);
527 			smc->sk.sk_data_ready(&smc->sk);
528 		}
529 	} else {
530 		smc_wr_tx_put_slot(conn->lnk,
531 				   (struct smc_wr_tx_pend_priv *)pend);
532 	}
533 	return rc;
534 }
535 
536 struct smc_link *smc_switch_conns(struct smc_link_group *lgr,
537 				  struct smc_link *from_lnk, bool is_dev_err)
538 {
539 	struct smc_link *to_lnk = NULL;
540 	struct smc_cdc_tx_pend *pend;
541 	struct smc_connection *conn;
542 	struct smc_wr_buf *wr_buf;
543 	struct smc_sock *smc;
544 	struct rb_node *node;
545 	int i, rc = 0;
546 
547 	/* link is inactive, wake up tx waiters */
548 	smc_wr_wakeup_tx_wait(from_lnk);
549 
550 	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
551 		if (lgr->lnk[i].state != SMC_LNK_ACTIVE ||
552 		    i == from_lnk->link_idx)
553 			continue;
554 		if (is_dev_err && from_lnk->smcibdev == lgr->lnk[i].smcibdev &&
555 		    from_lnk->ibport == lgr->lnk[i].ibport) {
556 			continue;
557 		}
558 		to_lnk = &lgr->lnk[i];
559 		break;
560 	}
561 	if (!to_lnk) {
562 		smc_lgr_terminate_sched(lgr);
563 		return NULL;
564 	}
565 again:
566 	read_lock_bh(&lgr->conns_lock);
567 	for (node = rb_first(&lgr->conns_all); node; node = rb_next(node)) {
568 		conn = rb_entry(node, struct smc_connection, alert_node);
569 		if (conn->lnk != from_lnk)
570 			continue;
571 		smc = container_of(conn, struct smc_sock, conn);
572 		/* conn->lnk not yet set in SMC_INIT state */
573 		if (smc->sk.sk_state == SMC_INIT)
574 			continue;
575 		if (smc->sk.sk_state == SMC_CLOSED ||
576 		    smc->sk.sk_state == SMC_PEERCLOSEWAIT1 ||
577 		    smc->sk.sk_state == SMC_PEERCLOSEWAIT2 ||
578 		    smc->sk.sk_state == SMC_APPFINCLOSEWAIT ||
579 		    smc->sk.sk_state == SMC_APPCLOSEWAIT1 ||
580 		    smc->sk.sk_state == SMC_APPCLOSEWAIT2 ||
581 		    smc->sk.sk_state == SMC_PEERFINCLOSEWAIT ||
582 		    smc->sk.sk_state == SMC_PEERABORTWAIT ||
583 		    smc->sk.sk_state == SMC_PROCESSABORT) {
584 			spin_lock_bh(&conn->send_lock);
585 			conn->lnk = to_lnk;
586 			spin_unlock_bh(&conn->send_lock);
587 			continue;
588 		}
589 		sock_hold(&smc->sk);
590 		read_unlock_bh(&lgr->conns_lock);
591 		/* pre-fetch buffer outside of send_lock, might sleep */
592 		rc = smc_cdc_get_free_slot(conn, to_lnk, &wr_buf, NULL, &pend);
593 		if (rc) {
594 			smcr_link_down_cond_sched(to_lnk);
595 			return NULL;
596 		}
597 		/* avoid race with smcr_tx_sndbuf_nonempty() */
598 		spin_lock_bh(&conn->send_lock);
599 		conn->lnk = to_lnk;
600 		rc = smc_switch_cursor(smc, pend, wr_buf);
601 		spin_unlock_bh(&conn->send_lock);
602 		sock_put(&smc->sk);
603 		if (rc) {
604 			smcr_link_down_cond_sched(to_lnk);
605 			return NULL;
606 		}
607 		goto again;
608 	}
609 	read_unlock_bh(&lgr->conns_lock);
610 	return to_lnk;
611 }
612 
613 static void smcr_buf_unuse(struct smc_buf_desc *rmb_desc,
614 			   struct smc_link_group *lgr)
615 {
616 	int rc;
617 
618 	if (rmb_desc->is_conf_rkey && !list_empty(&lgr->list)) {
619 		/* unregister rmb with peer */
620 		rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY);
621 		if (!rc) {
622 			/* protect against smc_llc_cli_rkey_exchange() */
623 			mutex_lock(&lgr->llc_conf_mutex);
624 			smc_llc_do_delete_rkey(lgr, rmb_desc);
625 			rmb_desc->is_conf_rkey = false;
626 			mutex_unlock(&lgr->llc_conf_mutex);
627 			smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl);
628 		}
629 	}
630 
631 	if (rmb_desc->is_reg_err) {
632 		/* buf registration failed, reuse not possible */
633 		mutex_lock(&lgr->rmbs_lock);
634 		list_del(&rmb_desc->list);
635 		mutex_unlock(&lgr->rmbs_lock);
636 
637 		smc_buf_free(lgr, true, rmb_desc);
638 	} else {
639 		rmb_desc->used = 0;
640 	}
641 }
642 
643 static void smc_buf_unuse(struct smc_connection *conn,
644 			  struct smc_link_group *lgr)
645 {
646 	if (conn->sndbuf_desc)
647 		conn->sndbuf_desc->used = 0;
648 	if (conn->rmb_desc && lgr->is_smcd)
649 		conn->rmb_desc->used = 0;
650 	else if (conn->rmb_desc)
651 		smcr_buf_unuse(conn->rmb_desc, lgr);
652 }
653 
654 /* remove a finished connection from its link group */
655 void smc_conn_free(struct smc_connection *conn)
656 {
657 	struct smc_link_group *lgr = conn->lgr;
658 
659 	if (!lgr)
660 		return;
661 	if (lgr->is_smcd) {
662 		if (!list_empty(&lgr->list))
663 			smc_ism_unset_conn(conn);
664 		tasklet_kill(&conn->rx_tsklet);
665 	} else {
666 		smc_cdc_tx_dismiss_slots(conn);
667 		if (current_work() != &conn->abort_work)
668 			cancel_work_sync(&conn->abort_work);
669 	}
670 	if (!list_empty(&lgr->list)) {
671 		smc_lgr_unregister_conn(conn);
672 		smc_buf_unuse(conn, lgr); /* allow buffer reuse */
673 	}
674 
675 	if (!lgr->conns_num)
676 		smc_lgr_schedule_free_work(lgr);
677 }
678 
679 /* unregister a link from a buf_desc */
680 static void smcr_buf_unmap_link(struct smc_buf_desc *buf_desc, bool is_rmb,
681 				struct smc_link *lnk)
682 {
683 	if (is_rmb)
684 		buf_desc->is_reg_mr[lnk->link_idx] = false;
685 	if (!buf_desc->is_map_ib[lnk->link_idx])
686 		return;
687 	if (is_rmb) {
688 		if (buf_desc->mr_rx[lnk->link_idx]) {
689 			smc_ib_put_memory_region(
690 					buf_desc->mr_rx[lnk->link_idx]);
691 			buf_desc->mr_rx[lnk->link_idx] = NULL;
692 		}
693 		smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_FROM_DEVICE);
694 	} else {
695 		smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_TO_DEVICE);
696 	}
697 	sg_free_table(&buf_desc->sgt[lnk->link_idx]);
698 	buf_desc->is_map_ib[lnk->link_idx] = false;
699 }
700 
701 /* unmap all buffers of lgr for a deleted link */
702 static void smcr_buf_unmap_lgr(struct smc_link *lnk)
703 {
704 	struct smc_link_group *lgr = lnk->lgr;
705 	struct smc_buf_desc *buf_desc, *bf;
706 	int i;
707 
708 	for (i = 0; i < SMC_RMBE_SIZES; i++) {
709 		mutex_lock(&lgr->rmbs_lock);
710 		list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list)
711 			smcr_buf_unmap_link(buf_desc, true, lnk);
712 		mutex_unlock(&lgr->rmbs_lock);
713 		mutex_lock(&lgr->sndbufs_lock);
714 		list_for_each_entry_safe(buf_desc, bf, &lgr->sndbufs[i],
715 					 list)
716 			smcr_buf_unmap_link(buf_desc, false, lnk);
717 		mutex_unlock(&lgr->sndbufs_lock);
718 	}
719 }
720 
721 static void smcr_rtoken_clear_link(struct smc_link *lnk)
722 {
723 	struct smc_link_group *lgr = lnk->lgr;
724 	int i;
725 
726 	for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
727 		lgr->rtokens[i][lnk->link_idx].rkey = 0;
728 		lgr->rtokens[i][lnk->link_idx].dma_addr = 0;
729 	}
730 }
731 
732 /* must be called under lgr->llc_conf_mutex lock */
733 void smcr_link_clear(struct smc_link *lnk, bool log)
734 {
735 	struct smc_ib_device *smcibdev;
736 
737 	if (!lnk->lgr || lnk->state == SMC_LNK_UNUSED)
738 		return;
739 	lnk->peer_qpn = 0;
740 	smc_llc_link_clear(lnk, log);
741 	smcr_buf_unmap_lgr(lnk);
742 	smcr_rtoken_clear_link(lnk);
743 	smc_ib_modify_qp_reset(lnk);
744 	smc_wr_free_link(lnk);
745 	smc_ib_destroy_queue_pair(lnk);
746 	smc_ib_dealloc_protection_domain(lnk);
747 	smc_wr_free_link_mem(lnk);
748 	put_device(&lnk->smcibdev->ibdev->dev);
749 	smcibdev = lnk->smcibdev;
750 	memset(lnk, 0, sizeof(struct smc_link));
751 	lnk->state = SMC_LNK_UNUSED;
752 	if (!atomic_dec_return(&smcibdev->lnk_cnt))
753 		wake_up(&smcibdev->lnks_deleted);
754 }
755 
756 static void smcr_buf_free(struct smc_link_group *lgr, bool is_rmb,
757 			  struct smc_buf_desc *buf_desc)
758 {
759 	int i;
760 
761 	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
762 		smcr_buf_unmap_link(buf_desc, is_rmb, &lgr->lnk[i]);
763 
764 	if (buf_desc->pages)
765 		__free_pages(buf_desc->pages, buf_desc->order);
766 	kfree(buf_desc);
767 }
768 
769 static void smcd_buf_free(struct smc_link_group *lgr, bool is_dmb,
770 			  struct smc_buf_desc *buf_desc)
771 {
772 	if (is_dmb) {
773 		/* restore original buf len */
774 		buf_desc->len += sizeof(struct smcd_cdc_msg);
775 		smc_ism_unregister_dmb(lgr->smcd, buf_desc);
776 	} else {
777 		kfree(buf_desc->cpu_addr);
778 	}
779 	kfree(buf_desc);
780 }
781 
782 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
783 			 struct smc_buf_desc *buf_desc)
784 {
785 	if (lgr->is_smcd)
786 		smcd_buf_free(lgr, is_rmb, buf_desc);
787 	else
788 		smcr_buf_free(lgr, is_rmb, buf_desc);
789 }
790 
791 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
792 {
793 	struct smc_buf_desc *buf_desc, *bf_desc;
794 	struct list_head *buf_list;
795 	int i;
796 
797 	for (i = 0; i < SMC_RMBE_SIZES; i++) {
798 		if (is_rmb)
799 			buf_list = &lgr->rmbs[i];
800 		else
801 			buf_list = &lgr->sndbufs[i];
802 		list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
803 					 list) {
804 			list_del(&buf_desc->list);
805 			smc_buf_free(lgr, is_rmb, buf_desc);
806 		}
807 	}
808 }
809 
810 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
811 {
812 	/* free send buffers */
813 	__smc_lgr_free_bufs(lgr, false);
814 	/* free rmbs */
815 	__smc_lgr_free_bufs(lgr, true);
816 }
817 
818 /* remove a link group */
819 static void smc_lgr_free(struct smc_link_group *lgr)
820 {
821 	int i;
822 
823 	if (!lgr->is_smcd) {
824 		mutex_lock(&lgr->llc_conf_mutex);
825 		for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
826 			if (lgr->lnk[i].state != SMC_LNK_UNUSED)
827 				smcr_link_clear(&lgr->lnk[i], false);
828 		}
829 		mutex_unlock(&lgr->llc_conf_mutex);
830 		smc_llc_lgr_clear(lgr);
831 	}
832 
833 	smc_lgr_free_bufs(lgr);
834 	if (lgr->is_smcd) {
835 		if (!lgr->terminating) {
836 			smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
837 			put_device(&lgr->smcd->dev);
838 		}
839 		if (!atomic_dec_return(&lgr->smcd->lgr_cnt))
840 			wake_up(&lgr->smcd->lgrs_deleted);
841 	} else {
842 		if (!atomic_dec_return(&lgr_cnt))
843 			wake_up(&lgrs_deleted);
844 	}
845 	kfree(lgr);
846 }
847 
848 static void smcd_unregister_all_dmbs(struct smc_link_group *lgr)
849 {
850 	int i;
851 
852 	for (i = 0; i < SMC_RMBE_SIZES; i++) {
853 		struct smc_buf_desc *buf_desc;
854 
855 		list_for_each_entry(buf_desc, &lgr->rmbs[i], list) {
856 			buf_desc->len += sizeof(struct smcd_cdc_msg);
857 			smc_ism_unregister_dmb(lgr->smcd, buf_desc);
858 		}
859 	}
860 }
861 
862 static void smc_sk_wake_ups(struct smc_sock *smc)
863 {
864 	smc->sk.sk_write_space(&smc->sk);
865 	smc->sk.sk_data_ready(&smc->sk);
866 	smc->sk.sk_state_change(&smc->sk);
867 }
868 
869 /* kill a connection */
870 static void smc_conn_kill(struct smc_connection *conn, bool soft)
871 {
872 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
873 
874 	if (conn->lgr->is_smcd && conn->lgr->peer_shutdown)
875 		conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
876 	else
877 		smc_close_abort(conn);
878 	conn->killed = 1;
879 	smc->sk.sk_err = ECONNABORTED;
880 	smc_sk_wake_ups(smc);
881 	if (conn->lgr->is_smcd) {
882 		smc_ism_unset_conn(conn);
883 		if (soft)
884 			tasklet_kill(&conn->rx_tsklet);
885 		else
886 			tasklet_unlock_wait(&conn->rx_tsklet);
887 	} else {
888 		smc_cdc_tx_dismiss_slots(conn);
889 	}
890 	smc_lgr_unregister_conn(conn);
891 	smc_close_active_abort(smc);
892 }
893 
894 static void smc_lgr_cleanup(struct smc_link_group *lgr)
895 {
896 	if (lgr->is_smcd) {
897 		smc_ism_signal_shutdown(lgr);
898 		smcd_unregister_all_dmbs(lgr);
899 		smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
900 		put_device(&lgr->smcd->dev);
901 	} else {
902 		u32 rsn = lgr->llc_termination_rsn;
903 
904 		if (!rsn)
905 			rsn = SMC_LLC_DEL_PROG_INIT_TERM;
906 		smc_llc_send_link_delete_all(lgr, false, rsn);
907 		smcr_lgr_link_deactivate_all(lgr);
908 	}
909 }
910 
911 /* terminate link group
912  * @soft: true if link group shutdown can take its time
913  *	  false if immediate link group shutdown is required
914  */
915 static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft)
916 {
917 	struct smc_connection *conn;
918 	struct smc_sock *smc;
919 	struct rb_node *node;
920 
921 	if (lgr->terminating)
922 		return;	/* lgr already terminating */
923 	/* cancel free_work sync, will terminate when lgr->freeing is set */
924 	cancel_delayed_work_sync(&lgr->free_work);
925 	lgr->terminating = 1;
926 
927 	/* kill remaining link group connections */
928 	read_lock_bh(&lgr->conns_lock);
929 	node = rb_first(&lgr->conns_all);
930 	while (node) {
931 		read_unlock_bh(&lgr->conns_lock);
932 		conn = rb_entry(node, struct smc_connection, alert_node);
933 		smc = container_of(conn, struct smc_sock, conn);
934 		sock_hold(&smc->sk); /* sock_put below */
935 		lock_sock(&smc->sk);
936 		smc_conn_kill(conn, soft);
937 		release_sock(&smc->sk);
938 		sock_put(&smc->sk); /* sock_hold above */
939 		read_lock_bh(&lgr->conns_lock);
940 		node = rb_first(&lgr->conns_all);
941 	}
942 	read_unlock_bh(&lgr->conns_lock);
943 	smc_lgr_cleanup(lgr);
944 	smc_lgr_free(lgr);
945 }
946 
947 /* unlink link group and schedule termination */
948 void smc_lgr_terminate_sched(struct smc_link_group *lgr)
949 {
950 	spinlock_t *lgr_lock;
951 
952 	smc_lgr_list_head(lgr, &lgr_lock);
953 	spin_lock_bh(lgr_lock);
954 	if (list_empty(&lgr->list) || lgr->terminating || lgr->freeing) {
955 		spin_unlock_bh(lgr_lock);
956 		return;	/* lgr already terminating */
957 	}
958 	list_del_init(&lgr->list);
959 	lgr->freeing = 1;
960 	spin_unlock_bh(lgr_lock);
961 	schedule_work(&lgr->terminate_work);
962 }
963 
964 /* Called when peer lgr shutdown (regularly or abnormally) is received */
965 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid, unsigned short vlan)
966 {
967 	struct smc_link_group *lgr, *l;
968 	LIST_HEAD(lgr_free_list);
969 
970 	/* run common cleanup function and build free list */
971 	spin_lock_bh(&dev->lgr_lock);
972 	list_for_each_entry_safe(lgr, l, &dev->lgr_list, list) {
973 		if ((!peer_gid || lgr->peer_gid == peer_gid) &&
974 		    (vlan == VLAN_VID_MASK || lgr->vlan_id == vlan)) {
975 			if (peer_gid) /* peer triggered termination */
976 				lgr->peer_shutdown = 1;
977 			list_move(&lgr->list, &lgr_free_list);
978 			lgr->freeing = 1;
979 		}
980 	}
981 	spin_unlock_bh(&dev->lgr_lock);
982 
983 	/* cancel the regular free workers and actually free lgrs */
984 	list_for_each_entry_safe(lgr, l, &lgr_free_list, list) {
985 		list_del_init(&lgr->list);
986 		schedule_work(&lgr->terminate_work);
987 	}
988 }
989 
990 /* Called when an SMCD device is removed or the smc module is unloaded */
991 void smc_smcd_terminate_all(struct smcd_dev *smcd)
992 {
993 	struct smc_link_group *lgr, *lg;
994 	LIST_HEAD(lgr_free_list);
995 
996 	spin_lock_bh(&smcd->lgr_lock);
997 	list_splice_init(&smcd->lgr_list, &lgr_free_list);
998 	list_for_each_entry(lgr, &lgr_free_list, list)
999 		lgr->freeing = 1;
1000 	spin_unlock_bh(&smcd->lgr_lock);
1001 
1002 	list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
1003 		list_del_init(&lgr->list);
1004 		__smc_lgr_terminate(lgr, false);
1005 	}
1006 
1007 	if (atomic_read(&smcd->lgr_cnt))
1008 		wait_event(smcd->lgrs_deleted, !atomic_read(&smcd->lgr_cnt));
1009 }
1010 
1011 /* Called when an SMCR device is removed or the smc module is unloaded.
1012  * If smcibdev is given, all SMCR link groups using this device are terminated.
1013  * If smcibdev is NULL, all SMCR link groups are terminated.
1014  */
1015 void smc_smcr_terminate_all(struct smc_ib_device *smcibdev)
1016 {
1017 	struct smc_link_group *lgr, *lg;
1018 	LIST_HEAD(lgr_free_list);
1019 	int i;
1020 
1021 	spin_lock_bh(&smc_lgr_list.lock);
1022 	if (!smcibdev) {
1023 		list_splice_init(&smc_lgr_list.list, &lgr_free_list);
1024 		list_for_each_entry(lgr, &lgr_free_list, list)
1025 			lgr->freeing = 1;
1026 	} else {
1027 		list_for_each_entry_safe(lgr, lg, &smc_lgr_list.list, list) {
1028 			for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1029 				if (lgr->lnk[i].smcibdev == smcibdev)
1030 					smcr_link_down_cond_sched(&lgr->lnk[i]);
1031 			}
1032 		}
1033 	}
1034 	spin_unlock_bh(&smc_lgr_list.lock);
1035 
1036 	list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
1037 		list_del_init(&lgr->list);
1038 		smc_llc_set_termination_rsn(lgr, SMC_LLC_DEL_OP_INIT_TERM);
1039 		__smc_lgr_terminate(lgr, false);
1040 	}
1041 
1042 	if (smcibdev) {
1043 		if (atomic_read(&smcibdev->lnk_cnt))
1044 			wait_event(smcibdev->lnks_deleted,
1045 				   !atomic_read(&smcibdev->lnk_cnt));
1046 	} else {
1047 		if (atomic_read(&lgr_cnt))
1048 			wait_event(lgrs_deleted, !atomic_read(&lgr_cnt));
1049 	}
1050 }
1051 
1052 /* set new lgr type and clear all asymmetric link tagging */
1053 void smcr_lgr_set_type(struct smc_link_group *lgr, enum smc_lgr_type new_type)
1054 {
1055 	char *lgr_type = "";
1056 	int i;
1057 
1058 	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
1059 		if (smc_link_usable(&lgr->lnk[i]))
1060 			lgr->lnk[i].link_is_asym = false;
1061 	if (lgr->type == new_type)
1062 		return;
1063 	lgr->type = new_type;
1064 
1065 	switch (lgr->type) {
1066 	case SMC_LGR_NONE:
1067 		lgr_type = "NONE";
1068 		break;
1069 	case SMC_LGR_SINGLE:
1070 		lgr_type = "SINGLE";
1071 		break;
1072 	case SMC_LGR_SYMMETRIC:
1073 		lgr_type = "SYMMETRIC";
1074 		break;
1075 	case SMC_LGR_ASYMMETRIC_PEER:
1076 		lgr_type = "ASYMMETRIC_PEER";
1077 		break;
1078 	case SMC_LGR_ASYMMETRIC_LOCAL:
1079 		lgr_type = "ASYMMETRIC_LOCAL";
1080 		break;
1081 	}
1082 	pr_warn_ratelimited("smc: SMC-R lg %*phN state changed: "
1083 			    "%s, pnetid %.16s\n", SMC_LGR_ID_SIZE, &lgr->id,
1084 			    lgr_type, lgr->pnet_id);
1085 }
1086 
1087 /* set new lgr type and tag a link as asymmetric */
1088 void smcr_lgr_set_type_asym(struct smc_link_group *lgr,
1089 			    enum smc_lgr_type new_type, int asym_lnk_idx)
1090 {
1091 	smcr_lgr_set_type(lgr, new_type);
1092 	lgr->lnk[asym_lnk_idx].link_is_asym = true;
1093 }
1094 
1095 /* abort connection, abort_work scheduled from tasklet context */
1096 static void smc_conn_abort_work(struct work_struct *work)
1097 {
1098 	struct smc_connection *conn = container_of(work,
1099 						   struct smc_connection,
1100 						   abort_work);
1101 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1102 
1103 	smc_conn_kill(conn, true);
1104 	sock_put(&smc->sk); /* sock_hold done by schedulers of abort_work */
1105 }
1106 
1107 /* link is up - establish alternate link if applicable */
1108 static void smcr_link_up(struct smc_link_group *lgr,
1109 			 struct smc_ib_device *smcibdev, u8 ibport)
1110 {
1111 	struct smc_link *link = NULL;
1112 
1113 	if (list_empty(&lgr->list) ||
1114 	    lgr->type == SMC_LGR_SYMMETRIC ||
1115 	    lgr->type == SMC_LGR_ASYMMETRIC_PEER)
1116 		return;
1117 
1118 	if (lgr->role == SMC_SERV) {
1119 		/* trigger local add link processing */
1120 		link = smc_llc_usable_link(lgr);
1121 		if (!link)
1122 			return;
1123 		smc_llc_srv_add_link_local(link);
1124 	} else {
1125 		/* invite server to start add link processing */
1126 		u8 gid[SMC_GID_SIZE];
1127 
1128 		if (smc_ib_determine_gid(smcibdev, ibport, lgr->vlan_id, gid,
1129 					 NULL))
1130 			return;
1131 		if (lgr->llc_flow_lcl.type != SMC_LLC_FLOW_NONE) {
1132 			/* some other llc task is ongoing */
1133 			wait_event_interruptible_timeout(lgr->llc_waiter,
1134 				(lgr->llc_flow_lcl.type == SMC_LLC_FLOW_NONE),
1135 				SMC_LLC_WAIT_TIME);
1136 		}
1137 		if (list_empty(&lgr->list) ||
1138 		    !smc_ib_port_active(smcibdev, ibport))
1139 			return; /* lgr or device no longer active */
1140 		link = smc_llc_usable_link(lgr);
1141 		if (!link)
1142 			return;
1143 		smc_llc_send_add_link(link, smcibdev->mac[ibport - 1], gid,
1144 				      NULL, SMC_LLC_REQ);
1145 	}
1146 }
1147 
1148 void smcr_port_add(struct smc_ib_device *smcibdev, u8 ibport)
1149 {
1150 	struct smc_ib_up_work *ib_work;
1151 	struct smc_link_group *lgr, *n;
1152 
1153 	list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1154 		if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1155 			    SMC_MAX_PNETID_LEN) ||
1156 		    lgr->type == SMC_LGR_SYMMETRIC ||
1157 		    lgr->type == SMC_LGR_ASYMMETRIC_PEER)
1158 			continue;
1159 		ib_work = kmalloc(sizeof(*ib_work), GFP_KERNEL);
1160 		if (!ib_work)
1161 			continue;
1162 		INIT_WORK(&ib_work->work, smc_link_up_work);
1163 		ib_work->lgr = lgr;
1164 		ib_work->smcibdev = smcibdev;
1165 		ib_work->ibport = ibport;
1166 		schedule_work(&ib_work->work);
1167 	}
1168 }
1169 
1170 /* link is down - switch connections to alternate link,
1171  * must be called under lgr->llc_conf_mutex lock
1172  */
1173 static void smcr_link_down(struct smc_link *lnk)
1174 {
1175 	struct smc_link_group *lgr = lnk->lgr;
1176 	struct smc_link *to_lnk;
1177 	int del_link_id;
1178 
1179 	if (!lgr || lnk->state == SMC_LNK_UNUSED || list_empty(&lgr->list))
1180 		return;
1181 
1182 	smc_ib_modify_qp_reset(lnk);
1183 	to_lnk = smc_switch_conns(lgr, lnk, true);
1184 	if (!to_lnk) { /* no backup link available */
1185 		smcr_link_clear(lnk, true);
1186 		return;
1187 	}
1188 	smcr_lgr_set_type(lgr, SMC_LGR_SINGLE);
1189 	del_link_id = lnk->link_id;
1190 
1191 	if (lgr->role == SMC_SERV) {
1192 		/* trigger local delete link processing */
1193 		smc_llc_srv_delete_link_local(to_lnk, del_link_id);
1194 	} else {
1195 		if (lgr->llc_flow_lcl.type != SMC_LLC_FLOW_NONE) {
1196 			/* another llc task is ongoing */
1197 			mutex_unlock(&lgr->llc_conf_mutex);
1198 			wait_event_interruptible_timeout(lgr->llc_waiter,
1199 				(lgr->llc_flow_lcl.type == SMC_LLC_FLOW_NONE),
1200 				SMC_LLC_WAIT_TIME);
1201 			mutex_lock(&lgr->llc_conf_mutex);
1202 		}
1203 		smc_llc_send_delete_link(to_lnk, del_link_id, SMC_LLC_REQ, true,
1204 					 SMC_LLC_DEL_LOST_PATH);
1205 	}
1206 }
1207 
1208 /* must be called under lgr->llc_conf_mutex lock */
1209 void smcr_link_down_cond(struct smc_link *lnk)
1210 {
1211 	if (smc_link_downing(&lnk->state))
1212 		smcr_link_down(lnk);
1213 }
1214 
1215 /* will get the lgr->llc_conf_mutex lock */
1216 void smcr_link_down_cond_sched(struct smc_link *lnk)
1217 {
1218 	if (smc_link_downing(&lnk->state))
1219 		schedule_work(&lnk->link_down_wrk);
1220 }
1221 
1222 void smcr_port_err(struct smc_ib_device *smcibdev, u8 ibport)
1223 {
1224 	struct smc_link_group *lgr, *n;
1225 	int i;
1226 
1227 	list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1228 		if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1229 			    SMC_MAX_PNETID_LEN))
1230 			continue; /* lgr is not affected */
1231 		if (list_empty(&lgr->list))
1232 			continue;
1233 		for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1234 			struct smc_link *lnk = &lgr->lnk[i];
1235 
1236 			if (smc_link_usable(lnk) &&
1237 			    lnk->smcibdev == smcibdev && lnk->ibport == ibport)
1238 				smcr_link_down_cond_sched(lnk);
1239 		}
1240 	}
1241 }
1242 
1243 static void smc_link_up_work(struct work_struct *work)
1244 {
1245 	struct smc_ib_up_work *ib_work = container_of(work,
1246 						      struct smc_ib_up_work,
1247 						      work);
1248 	struct smc_link_group *lgr = ib_work->lgr;
1249 
1250 	if (list_empty(&lgr->list))
1251 		goto out;
1252 	smcr_link_up(lgr, ib_work->smcibdev, ib_work->ibport);
1253 out:
1254 	kfree(ib_work);
1255 }
1256 
1257 static void smc_link_down_work(struct work_struct *work)
1258 {
1259 	struct smc_link *link = container_of(work, struct smc_link,
1260 					     link_down_wrk);
1261 	struct smc_link_group *lgr = link->lgr;
1262 
1263 	if (list_empty(&lgr->list))
1264 		return;
1265 	wake_up_interruptible_all(&lgr->llc_waiter);
1266 	mutex_lock(&lgr->llc_conf_mutex);
1267 	smcr_link_down(link);
1268 	mutex_unlock(&lgr->llc_conf_mutex);
1269 }
1270 
1271 /* Determine vlan of internal TCP socket.
1272  * @vlan_id: address to store the determined vlan id into
1273  */
1274 int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini)
1275 {
1276 	struct dst_entry *dst = sk_dst_get(clcsock->sk);
1277 	struct net_device *ndev;
1278 	int i, nest_lvl, rc = 0;
1279 
1280 	ini->vlan_id = 0;
1281 	if (!dst) {
1282 		rc = -ENOTCONN;
1283 		goto out;
1284 	}
1285 	if (!dst->dev) {
1286 		rc = -ENODEV;
1287 		goto out_rel;
1288 	}
1289 
1290 	ndev = dst->dev;
1291 	if (is_vlan_dev(ndev)) {
1292 		ini->vlan_id = vlan_dev_vlan_id(ndev);
1293 		goto out_rel;
1294 	}
1295 
1296 	rtnl_lock();
1297 	nest_lvl = ndev->lower_level;
1298 	for (i = 0; i < nest_lvl; i++) {
1299 		struct list_head *lower = &ndev->adj_list.lower;
1300 
1301 		if (list_empty(lower))
1302 			break;
1303 		lower = lower->next;
1304 		ndev = (struct net_device *)netdev_lower_get_next(ndev, &lower);
1305 		if (is_vlan_dev(ndev)) {
1306 			ini->vlan_id = vlan_dev_vlan_id(ndev);
1307 			break;
1308 		}
1309 	}
1310 	rtnl_unlock();
1311 
1312 out_rel:
1313 	dst_release(dst);
1314 out:
1315 	return rc;
1316 }
1317 
1318 static bool smcr_lgr_match(struct smc_link_group *lgr,
1319 			   struct smc_clc_msg_local *lcl,
1320 			   enum smc_lgr_role role, u32 clcqpn)
1321 {
1322 	int i;
1323 
1324 	if (memcmp(lgr->peer_systemid, lcl->id_for_peer, SMC_SYSTEMID_LEN) ||
1325 	    lgr->role != role)
1326 		return false;
1327 
1328 	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1329 		if (lgr->lnk[i].state != SMC_LNK_ACTIVE)
1330 			continue;
1331 		if ((lgr->role == SMC_SERV || lgr->lnk[i].peer_qpn == clcqpn) &&
1332 		    !memcmp(lgr->lnk[i].peer_gid, &lcl->gid, SMC_GID_SIZE) &&
1333 		    !memcmp(lgr->lnk[i].peer_mac, lcl->mac, sizeof(lcl->mac)))
1334 			return true;
1335 	}
1336 	return false;
1337 }
1338 
1339 static bool smcd_lgr_match(struct smc_link_group *lgr,
1340 			   struct smcd_dev *smcismdev, u64 peer_gid)
1341 {
1342 	return lgr->peer_gid == peer_gid && lgr->smcd == smcismdev;
1343 }
1344 
1345 /* create a new SMC connection (and a new link group if necessary) */
1346 int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini)
1347 {
1348 	struct smc_connection *conn = &smc->conn;
1349 	struct list_head *lgr_list;
1350 	struct smc_link_group *lgr;
1351 	enum smc_lgr_role role;
1352 	spinlock_t *lgr_lock;
1353 	int rc = 0;
1354 
1355 	lgr_list = ini->is_smcd ? &ini->ism_dev->lgr_list : &smc_lgr_list.list;
1356 	lgr_lock = ini->is_smcd ? &ini->ism_dev->lgr_lock : &smc_lgr_list.lock;
1357 	ini->cln_first_contact = SMC_FIRST_CONTACT;
1358 	role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
1359 	if (role == SMC_CLNT && ini->srv_first_contact)
1360 		/* create new link group as well */
1361 		goto create;
1362 
1363 	/* determine if an existing link group can be reused */
1364 	spin_lock_bh(lgr_lock);
1365 	list_for_each_entry(lgr, lgr_list, list) {
1366 		write_lock_bh(&lgr->conns_lock);
1367 		if ((ini->is_smcd ?
1368 		     smcd_lgr_match(lgr, ini->ism_dev, ini->ism_gid) :
1369 		     smcr_lgr_match(lgr, ini->ib_lcl, role, ini->ib_clcqpn)) &&
1370 		    !lgr->sync_err &&
1371 		    lgr->vlan_id == ini->vlan_id &&
1372 		    (role == SMC_CLNT ||
1373 		     lgr->conns_num < SMC_RMBS_PER_LGR_MAX)) {
1374 			/* link group found */
1375 			ini->cln_first_contact = SMC_REUSE_CONTACT;
1376 			conn->lgr = lgr;
1377 			rc = smc_lgr_register_conn(conn, false);
1378 			write_unlock_bh(&lgr->conns_lock);
1379 			if (!rc && delayed_work_pending(&lgr->free_work))
1380 				cancel_delayed_work(&lgr->free_work);
1381 			break;
1382 		}
1383 		write_unlock_bh(&lgr->conns_lock);
1384 	}
1385 	spin_unlock_bh(lgr_lock);
1386 	if (rc)
1387 		return rc;
1388 
1389 	if (role == SMC_CLNT && !ini->srv_first_contact &&
1390 	    ini->cln_first_contact == SMC_FIRST_CONTACT) {
1391 		/* Server reuses a link group, but Client wants to start
1392 		 * a new one
1393 		 * send out_of_sync decline, reason synchr. error
1394 		 */
1395 		return SMC_CLC_DECL_SYNCERR;
1396 	}
1397 
1398 create:
1399 	if (ini->cln_first_contact == SMC_FIRST_CONTACT) {
1400 		rc = smc_lgr_create(smc, ini);
1401 		if (rc)
1402 			goto out;
1403 		lgr = conn->lgr;
1404 		write_lock_bh(&lgr->conns_lock);
1405 		rc = smc_lgr_register_conn(conn, true);
1406 		write_unlock_bh(&lgr->conns_lock);
1407 		if (rc)
1408 			goto out;
1409 	}
1410 	conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
1411 	conn->local_tx_ctrl.len = SMC_WR_TX_SIZE;
1412 	conn->urg_state = SMC_URG_READ;
1413 	INIT_WORK(&smc->conn.abort_work, smc_conn_abort_work);
1414 	if (ini->is_smcd) {
1415 		conn->rx_off = sizeof(struct smcd_cdc_msg);
1416 		smcd_cdc_rx_init(conn); /* init tasklet for this conn */
1417 	}
1418 #ifndef KERNEL_HAS_ATOMIC64
1419 	spin_lock_init(&conn->acurs_lock);
1420 #endif
1421 
1422 out:
1423 	return rc;
1424 }
1425 
1426 /* convert the RMB size into the compressed notation - minimum 16K.
1427  * In contrast to plain ilog2, this rounds towards the next power of 2,
1428  * so the socket application gets at least its desired sndbuf / rcvbuf size.
1429  */
1430 static u8 smc_compress_bufsize(int size)
1431 {
1432 	u8 compressed;
1433 
1434 	if (size <= SMC_BUF_MIN_SIZE)
1435 		return 0;
1436 
1437 	size = (size - 1) >> 14;
1438 	compressed = ilog2(size) + 1;
1439 	if (compressed >= SMC_RMBE_SIZES)
1440 		compressed = SMC_RMBE_SIZES - 1;
1441 	return compressed;
1442 }
1443 
1444 /* convert the RMB size from compressed notation into integer */
1445 int smc_uncompress_bufsize(u8 compressed)
1446 {
1447 	u32 size;
1448 
1449 	size = 0x00000001 << (((int)compressed) + 14);
1450 	return (int)size;
1451 }
1452 
1453 /* try to reuse a sndbuf or rmb description slot for a certain
1454  * buffer size; if not available, return NULL
1455  */
1456 static struct smc_buf_desc *smc_buf_get_slot(int compressed_bufsize,
1457 					     struct mutex *lock,
1458 					     struct list_head *buf_list)
1459 {
1460 	struct smc_buf_desc *buf_slot;
1461 
1462 	mutex_lock(lock);
1463 	list_for_each_entry(buf_slot, buf_list, list) {
1464 		if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
1465 			mutex_unlock(lock);
1466 			return buf_slot;
1467 		}
1468 	}
1469 	mutex_unlock(lock);
1470 	return NULL;
1471 }
1472 
1473 /* one of the conditions for announcing a receiver's current window size is
1474  * that it "results in a minimum increase in the window size of 10% of the
1475  * receive buffer space" [RFC7609]
1476  */
1477 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
1478 {
1479 	return min_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
1480 }
1481 
1482 /* map an rmb buf to a link */
1483 static int smcr_buf_map_link(struct smc_buf_desc *buf_desc, bool is_rmb,
1484 			     struct smc_link *lnk)
1485 {
1486 	int rc;
1487 
1488 	if (buf_desc->is_map_ib[lnk->link_idx])
1489 		return 0;
1490 
1491 	rc = sg_alloc_table(&buf_desc->sgt[lnk->link_idx], 1, GFP_KERNEL);
1492 	if (rc)
1493 		return rc;
1494 	sg_set_buf(buf_desc->sgt[lnk->link_idx].sgl,
1495 		   buf_desc->cpu_addr, buf_desc->len);
1496 
1497 	/* map sg table to DMA address */
1498 	rc = smc_ib_buf_map_sg(lnk, buf_desc,
1499 			       is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
1500 	/* SMC protocol depends on mapping to one DMA address only */
1501 	if (rc != 1) {
1502 		rc = -EAGAIN;
1503 		goto free_table;
1504 	}
1505 
1506 	/* create a new memory region for the RMB */
1507 	if (is_rmb) {
1508 		rc = smc_ib_get_memory_region(lnk->roce_pd,
1509 					      IB_ACCESS_REMOTE_WRITE |
1510 					      IB_ACCESS_LOCAL_WRITE,
1511 					      buf_desc, lnk->link_idx);
1512 		if (rc)
1513 			goto buf_unmap;
1514 		smc_ib_sync_sg_for_device(lnk, buf_desc, DMA_FROM_DEVICE);
1515 	}
1516 	buf_desc->is_map_ib[lnk->link_idx] = true;
1517 	return 0;
1518 
1519 buf_unmap:
1520 	smc_ib_buf_unmap_sg(lnk, buf_desc,
1521 			    is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
1522 free_table:
1523 	sg_free_table(&buf_desc->sgt[lnk->link_idx]);
1524 	return rc;
1525 }
1526 
1527 /* register a new rmb on IB device,
1528  * must be called under lgr->llc_conf_mutex lock
1529  */
1530 int smcr_link_reg_rmb(struct smc_link *link, struct smc_buf_desc *rmb_desc)
1531 {
1532 	if (list_empty(&link->lgr->list))
1533 		return -ENOLINK;
1534 	if (!rmb_desc->is_reg_mr[link->link_idx]) {
1535 		/* register memory region for new rmb */
1536 		if (smc_wr_reg_send(link, rmb_desc->mr_rx[link->link_idx])) {
1537 			rmb_desc->is_reg_err = true;
1538 			return -EFAULT;
1539 		}
1540 		rmb_desc->is_reg_mr[link->link_idx] = true;
1541 	}
1542 	return 0;
1543 }
1544 
1545 static int _smcr_buf_map_lgr(struct smc_link *lnk, struct mutex *lock,
1546 			     struct list_head *lst, bool is_rmb)
1547 {
1548 	struct smc_buf_desc *buf_desc, *bf;
1549 	int rc = 0;
1550 
1551 	mutex_lock(lock);
1552 	list_for_each_entry_safe(buf_desc, bf, lst, list) {
1553 		if (!buf_desc->used)
1554 			continue;
1555 		rc = smcr_buf_map_link(buf_desc, is_rmb, lnk);
1556 		if (rc)
1557 			goto out;
1558 	}
1559 out:
1560 	mutex_unlock(lock);
1561 	return rc;
1562 }
1563 
1564 /* map all used buffers of lgr for a new link */
1565 int smcr_buf_map_lgr(struct smc_link *lnk)
1566 {
1567 	struct smc_link_group *lgr = lnk->lgr;
1568 	int i, rc = 0;
1569 
1570 	for (i = 0; i < SMC_RMBE_SIZES; i++) {
1571 		rc = _smcr_buf_map_lgr(lnk, &lgr->rmbs_lock,
1572 				       &lgr->rmbs[i], true);
1573 		if (rc)
1574 			return rc;
1575 		rc = _smcr_buf_map_lgr(lnk, &lgr->sndbufs_lock,
1576 				       &lgr->sndbufs[i], false);
1577 		if (rc)
1578 			return rc;
1579 	}
1580 	return 0;
1581 }
1582 
1583 /* register all used buffers of lgr for a new link,
1584  * must be called under lgr->llc_conf_mutex lock
1585  */
1586 int smcr_buf_reg_lgr(struct smc_link *lnk)
1587 {
1588 	struct smc_link_group *lgr = lnk->lgr;
1589 	struct smc_buf_desc *buf_desc, *bf;
1590 	int i, rc = 0;
1591 
1592 	mutex_lock(&lgr->rmbs_lock);
1593 	for (i = 0; i < SMC_RMBE_SIZES; i++) {
1594 		list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list) {
1595 			if (!buf_desc->used)
1596 				continue;
1597 			rc = smcr_link_reg_rmb(lnk, buf_desc);
1598 			if (rc)
1599 				goto out;
1600 		}
1601 	}
1602 out:
1603 	mutex_unlock(&lgr->rmbs_lock);
1604 	return rc;
1605 }
1606 
1607 static struct smc_buf_desc *smcr_new_buf_create(struct smc_link_group *lgr,
1608 						bool is_rmb, int bufsize)
1609 {
1610 	struct smc_buf_desc *buf_desc;
1611 
1612 	/* try to alloc a new buffer */
1613 	buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
1614 	if (!buf_desc)
1615 		return ERR_PTR(-ENOMEM);
1616 
1617 	buf_desc->order = get_order(bufsize);
1618 	buf_desc->pages = alloc_pages(GFP_KERNEL | __GFP_NOWARN |
1619 				      __GFP_NOMEMALLOC | __GFP_COMP |
1620 				      __GFP_NORETRY | __GFP_ZERO,
1621 				      buf_desc->order);
1622 	if (!buf_desc->pages) {
1623 		kfree(buf_desc);
1624 		return ERR_PTR(-EAGAIN);
1625 	}
1626 	buf_desc->cpu_addr = (void *)page_address(buf_desc->pages);
1627 	buf_desc->len = bufsize;
1628 	return buf_desc;
1629 }
1630 
1631 /* map buf_desc on all usable links,
1632  * unused buffers stay mapped as long as the link is up
1633  */
1634 static int smcr_buf_map_usable_links(struct smc_link_group *lgr,
1635 				     struct smc_buf_desc *buf_desc, bool is_rmb)
1636 {
1637 	int i, rc = 0;
1638 
1639 	/* protect against parallel link reconfiguration */
1640 	mutex_lock(&lgr->llc_conf_mutex);
1641 	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1642 		struct smc_link *lnk = &lgr->lnk[i];
1643 
1644 		if (!smc_link_usable(lnk))
1645 			continue;
1646 		if (smcr_buf_map_link(buf_desc, is_rmb, lnk)) {
1647 			rc = -ENOMEM;
1648 			goto out;
1649 		}
1650 	}
1651 out:
1652 	mutex_unlock(&lgr->llc_conf_mutex);
1653 	return rc;
1654 }
1655 
1656 #define SMCD_DMBE_SIZES		7 /* 0 -> 16KB, 1 -> 32KB, .. 6 -> 1MB */
1657 
1658 static struct smc_buf_desc *smcd_new_buf_create(struct smc_link_group *lgr,
1659 						bool is_dmb, int bufsize)
1660 {
1661 	struct smc_buf_desc *buf_desc;
1662 	int rc;
1663 
1664 	if (smc_compress_bufsize(bufsize) > SMCD_DMBE_SIZES)
1665 		return ERR_PTR(-EAGAIN);
1666 
1667 	/* try to alloc a new DMB */
1668 	buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
1669 	if (!buf_desc)
1670 		return ERR_PTR(-ENOMEM);
1671 	if (is_dmb) {
1672 		rc = smc_ism_register_dmb(lgr, bufsize, buf_desc);
1673 		if (rc) {
1674 			kfree(buf_desc);
1675 			return ERR_PTR(-EAGAIN);
1676 		}
1677 		buf_desc->pages = virt_to_page(buf_desc->cpu_addr);
1678 		/* CDC header stored in buf. So, pretend it was smaller */
1679 		buf_desc->len = bufsize - sizeof(struct smcd_cdc_msg);
1680 	} else {
1681 		buf_desc->cpu_addr = kzalloc(bufsize, GFP_KERNEL |
1682 					     __GFP_NOWARN | __GFP_NORETRY |
1683 					     __GFP_NOMEMALLOC);
1684 		if (!buf_desc->cpu_addr) {
1685 			kfree(buf_desc);
1686 			return ERR_PTR(-EAGAIN);
1687 		}
1688 		buf_desc->len = bufsize;
1689 	}
1690 	return buf_desc;
1691 }
1692 
1693 static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
1694 {
1695 	struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
1696 	struct smc_connection *conn = &smc->conn;
1697 	struct smc_link_group *lgr = conn->lgr;
1698 	struct list_head *buf_list;
1699 	int bufsize, bufsize_short;
1700 	struct mutex *lock;	/* lock buffer list */
1701 	int sk_buf_size;
1702 
1703 	if (is_rmb)
1704 		/* use socket recv buffer size (w/o overhead) as start value */
1705 		sk_buf_size = smc->sk.sk_rcvbuf / 2;
1706 	else
1707 		/* use socket send buffer size (w/o overhead) as start value */
1708 		sk_buf_size = smc->sk.sk_sndbuf / 2;
1709 
1710 	for (bufsize_short = smc_compress_bufsize(sk_buf_size);
1711 	     bufsize_short >= 0; bufsize_short--) {
1712 
1713 		if (is_rmb) {
1714 			lock = &lgr->rmbs_lock;
1715 			buf_list = &lgr->rmbs[bufsize_short];
1716 		} else {
1717 			lock = &lgr->sndbufs_lock;
1718 			buf_list = &lgr->sndbufs[bufsize_short];
1719 		}
1720 		bufsize = smc_uncompress_bufsize(bufsize_short);
1721 		if ((1 << get_order(bufsize)) > SG_MAX_SINGLE_ALLOC)
1722 			continue;
1723 
1724 		/* check for reusable slot in the link group */
1725 		buf_desc = smc_buf_get_slot(bufsize_short, lock, buf_list);
1726 		if (buf_desc) {
1727 			memset(buf_desc->cpu_addr, 0, bufsize);
1728 			break; /* found reusable slot */
1729 		}
1730 
1731 		if (is_smcd)
1732 			buf_desc = smcd_new_buf_create(lgr, is_rmb, bufsize);
1733 		else
1734 			buf_desc = smcr_new_buf_create(lgr, is_rmb, bufsize);
1735 
1736 		if (PTR_ERR(buf_desc) == -ENOMEM)
1737 			break;
1738 		if (IS_ERR(buf_desc))
1739 			continue;
1740 
1741 		buf_desc->used = 1;
1742 		mutex_lock(lock);
1743 		list_add(&buf_desc->list, buf_list);
1744 		mutex_unlock(lock);
1745 		break; /* found */
1746 	}
1747 
1748 	if (IS_ERR(buf_desc))
1749 		return -ENOMEM;
1750 
1751 	if (!is_smcd) {
1752 		if (smcr_buf_map_usable_links(lgr, buf_desc, is_rmb)) {
1753 			smcr_buf_unuse(buf_desc, lgr);
1754 			return -ENOMEM;
1755 		}
1756 	}
1757 
1758 	if (is_rmb) {
1759 		conn->rmb_desc = buf_desc;
1760 		conn->rmbe_size_short = bufsize_short;
1761 		smc->sk.sk_rcvbuf = bufsize * 2;
1762 		atomic_set(&conn->bytes_to_rcv, 0);
1763 		conn->rmbe_update_limit =
1764 			smc_rmb_wnd_update_limit(buf_desc->len);
1765 		if (is_smcd)
1766 			smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */
1767 	} else {
1768 		conn->sndbuf_desc = buf_desc;
1769 		smc->sk.sk_sndbuf = bufsize * 2;
1770 		atomic_set(&conn->sndbuf_space, bufsize);
1771 	}
1772 	return 0;
1773 }
1774 
1775 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn)
1776 {
1777 	if (!conn->lgr || conn->lgr->is_smcd || !smc_link_usable(conn->lnk))
1778 		return;
1779 	smc_ib_sync_sg_for_cpu(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE);
1780 }
1781 
1782 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
1783 {
1784 	if (!conn->lgr || conn->lgr->is_smcd || !smc_link_usable(conn->lnk))
1785 		return;
1786 	smc_ib_sync_sg_for_device(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE);
1787 }
1788 
1789 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
1790 {
1791 	int i;
1792 
1793 	if (!conn->lgr || conn->lgr->is_smcd)
1794 		return;
1795 	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1796 		if (!smc_link_usable(&conn->lgr->lnk[i]))
1797 			continue;
1798 		smc_ib_sync_sg_for_cpu(&conn->lgr->lnk[i], conn->rmb_desc,
1799 				       DMA_FROM_DEVICE);
1800 	}
1801 }
1802 
1803 void smc_rmb_sync_sg_for_device(struct smc_connection *conn)
1804 {
1805 	int i;
1806 
1807 	if (!conn->lgr || conn->lgr->is_smcd)
1808 		return;
1809 	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1810 		if (!smc_link_usable(&conn->lgr->lnk[i]))
1811 			continue;
1812 		smc_ib_sync_sg_for_device(&conn->lgr->lnk[i], conn->rmb_desc,
1813 					  DMA_FROM_DEVICE);
1814 	}
1815 }
1816 
1817 /* create the send and receive buffer for an SMC socket;
1818  * receive buffers are called RMBs;
1819  * (even though the SMC protocol allows more than one RMB-element per RMB,
1820  * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
1821  * extra RMB for every connection in a link group
1822  */
1823 int smc_buf_create(struct smc_sock *smc, bool is_smcd)
1824 {
1825 	int rc;
1826 
1827 	/* create send buffer */
1828 	rc = __smc_buf_create(smc, is_smcd, false);
1829 	if (rc)
1830 		return rc;
1831 	/* create rmb */
1832 	rc = __smc_buf_create(smc, is_smcd, true);
1833 	if (rc)
1834 		smc_buf_free(smc->conn.lgr, false, smc->conn.sndbuf_desc);
1835 	return rc;
1836 }
1837 
1838 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
1839 {
1840 	int i;
1841 
1842 	for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
1843 		if (!test_and_set_bit(i, lgr->rtokens_used_mask))
1844 			return i;
1845 	}
1846 	return -ENOSPC;
1847 }
1848 
1849 static int smc_rtoken_find_by_link(struct smc_link_group *lgr, int lnk_idx,
1850 				   u32 rkey)
1851 {
1852 	int i;
1853 
1854 	for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1855 		if (test_bit(i, lgr->rtokens_used_mask) &&
1856 		    lgr->rtokens[i][lnk_idx].rkey == rkey)
1857 			return i;
1858 	}
1859 	return -ENOENT;
1860 }
1861 
1862 /* set rtoken for a new link to an existing rmb */
1863 void smc_rtoken_set(struct smc_link_group *lgr, int link_idx, int link_idx_new,
1864 		    __be32 nw_rkey_known, __be64 nw_vaddr, __be32 nw_rkey)
1865 {
1866 	int rtok_idx;
1867 
1868 	rtok_idx = smc_rtoken_find_by_link(lgr, link_idx, ntohl(nw_rkey_known));
1869 	if (rtok_idx == -ENOENT)
1870 		return;
1871 	lgr->rtokens[rtok_idx][link_idx_new].rkey = ntohl(nw_rkey);
1872 	lgr->rtokens[rtok_idx][link_idx_new].dma_addr = be64_to_cpu(nw_vaddr);
1873 }
1874 
1875 /* set rtoken for a new link whose link_id is given */
1876 void smc_rtoken_set2(struct smc_link_group *lgr, int rtok_idx, int link_id,
1877 		     __be64 nw_vaddr, __be32 nw_rkey)
1878 {
1879 	u64 dma_addr = be64_to_cpu(nw_vaddr);
1880 	u32 rkey = ntohl(nw_rkey);
1881 	bool found = false;
1882 	int link_idx;
1883 
1884 	for (link_idx = 0; link_idx < SMC_LINKS_PER_LGR_MAX; link_idx++) {
1885 		if (lgr->lnk[link_idx].link_id == link_id) {
1886 			found = true;
1887 			break;
1888 		}
1889 	}
1890 	if (!found)
1891 		return;
1892 	lgr->rtokens[rtok_idx][link_idx].rkey = rkey;
1893 	lgr->rtokens[rtok_idx][link_idx].dma_addr = dma_addr;
1894 }
1895 
1896 /* add a new rtoken from peer */
1897 int smc_rtoken_add(struct smc_link *lnk, __be64 nw_vaddr, __be32 nw_rkey)
1898 {
1899 	struct smc_link_group *lgr = smc_get_lgr(lnk);
1900 	u64 dma_addr = be64_to_cpu(nw_vaddr);
1901 	u32 rkey = ntohl(nw_rkey);
1902 	int i;
1903 
1904 	for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1905 		if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
1906 		    lgr->rtokens[i][lnk->link_idx].dma_addr == dma_addr &&
1907 		    test_bit(i, lgr->rtokens_used_mask)) {
1908 			/* already in list */
1909 			return i;
1910 		}
1911 	}
1912 	i = smc_rmb_reserve_rtoken_idx(lgr);
1913 	if (i < 0)
1914 		return i;
1915 	lgr->rtokens[i][lnk->link_idx].rkey = rkey;
1916 	lgr->rtokens[i][lnk->link_idx].dma_addr = dma_addr;
1917 	return i;
1918 }
1919 
1920 /* delete an rtoken from all links */
1921 int smc_rtoken_delete(struct smc_link *lnk, __be32 nw_rkey)
1922 {
1923 	struct smc_link_group *lgr = smc_get_lgr(lnk);
1924 	u32 rkey = ntohl(nw_rkey);
1925 	int i, j;
1926 
1927 	for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1928 		if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
1929 		    test_bit(i, lgr->rtokens_used_mask)) {
1930 			for (j = 0; j < SMC_LINKS_PER_LGR_MAX; j++) {
1931 				lgr->rtokens[i][j].rkey = 0;
1932 				lgr->rtokens[i][j].dma_addr = 0;
1933 			}
1934 			clear_bit(i, lgr->rtokens_used_mask);
1935 			return 0;
1936 		}
1937 	}
1938 	return -ENOENT;
1939 }
1940 
1941 /* save rkey and dma_addr received from peer during clc handshake */
1942 int smc_rmb_rtoken_handling(struct smc_connection *conn,
1943 			    struct smc_link *lnk,
1944 			    struct smc_clc_msg_accept_confirm *clc)
1945 {
1946 	conn->rtoken_idx = smc_rtoken_add(lnk, clc->rmb_dma_addr,
1947 					  clc->rmb_rkey);
1948 	if (conn->rtoken_idx < 0)
1949 		return conn->rtoken_idx;
1950 	return 0;
1951 }
1952 
1953 static void smc_core_going_away(void)
1954 {
1955 	struct smc_ib_device *smcibdev;
1956 	struct smcd_dev *smcd;
1957 
1958 	spin_lock(&smc_ib_devices.lock);
1959 	list_for_each_entry(smcibdev, &smc_ib_devices.list, list) {
1960 		int i;
1961 
1962 		for (i = 0; i < SMC_MAX_PORTS; i++)
1963 			set_bit(i, smcibdev->ports_going_away);
1964 	}
1965 	spin_unlock(&smc_ib_devices.lock);
1966 
1967 	spin_lock(&smcd_dev_list.lock);
1968 	list_for_each_entry(smcd, &smcd_dev_list.list, list) {
1969 		smcd->going_away = 1;
1970 	}
1971 	spin_unlock(&smcd_dev_list.lock);
1972 }
1973 
1974 /* Clean up all SMC link groups */
1975 static void smc_lgrs_shutdown(void)
1976 {
1977 	struct smcd_dev *smcd;
1978 
1979 	smc_core_going_away();
1980 
1981 	smc_smcr_terminate_all(NULL);
1982 
1983 	spin_lock(&smcd_dev_list.lock);
1984 	list_for_each_entry(smcd, &smcd_dev_list.list, list)
1985 		smc_smcd_terminate_all(smcd);
1986 	spin_unlock(&smcd_dev_list.lock);
1987 }
1988 
1989 static int smc_core_reboot_event(struct notifier_block *this,
1990 				 unsigned long event, void *ptr)
1991 {
1992 	smc_lgrs_shutdown();
1993 	smc_ib_unregister_client();
1994 	return 0;
1995 }
1996 
1997 static struct notifier_block smc_reboot_notifier = {
1998 	.notifier_call = smc_core_reboot_event,
1999 };
2000 
2001 int __init smc_core_init(void)
2002 {
2003 	return register_reboot_notifier(&smc_reboot_notifier);
2004 }
2005 
2006 /* Called (from smc_exit) when module is removed */
2007 void smc_core_exit(void)
2008 {
2009 	unregister_reboot_notifier(&smc_reboot_notifier);
2010 	smc_lgrs_shutdown();
2011 }
2012