xref: /linux/net/smc/smc_core.c (revision 2dbc0838bcf24ca59cabc3130cf3b1d6809cdcd4)
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 #include <rdma/ib_cache.h>
20 
21 #include "smc.h"
22 #include "smc_clc.h"
23 #include "smc_core.h"
24 #include "smc_ib.h"
25 #include "smc_wr.h"
26 #include "smc_llc.h"
27 #include "smc_cdc.h"
28 #include "smc_close.h"
29 #include "smc_ism.h"
30 
31 #define SMC_LGR_NUM_INCR		256
32 #define SMC_LGR_FREE_DELAY_SERV		(600 * HZ)
33 #define SMC_LGR_FREE_DELAY_CLNT		(SMC_LGR_FREE_DELAY_SERV + 10 * HZ)
34 #define SMC_LGR_FREE_DELAY_FAST		(8 * HZ)
35 
36 static struct smc_lgr_list smc_lgr_list = {	/* established link groups */
37 	.lock = __SPIN_LOCK_UNLOCKED(smc_lgr_list.lock),
38 	.list = LIST_HEAD_INIT(smc_lgr_list.list),
39 	.num = 0,
40 };
41 
42 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
43 			 struct smc_buf_desc *buf_desc);
44 
45 static void smc_lgr_schedule_free_work(struct smc_link_group *lgr)
46 {
47 	/* client link group creation always follows the server link group
48 	 * creation. For client use a somewhat higher removal delay time,
49 	 * otherwise there is a risk of out-of-sync link groups.
50 	 */
51 	mod_delayed_work(system_wq, &lgr->free_work,
52 			 (!lgr->is_smcd && lgr->role == SMC_CLNT) ?
53 			 SMC_LGR_FREE_DELAY_CLNT : SMC_LGR_FREE_DELAY_SERV);
54 }
55 
56 void smc_lgr_schedule_free_work_fast(struct smc_link_group *lgr)
57 {
58 	mod_delayed_work(system_wq, &lgr->free_work, SMC_LGR_FREE_DELAY_FAST);
59 }
60 
61 /* Register connection's alert token in our lookup structure.
62  * To use rbtrees we have to implement our own insert core.
63  * Requires @conns_lock
64  * @smc		connection to register
65  * Returns 0 on success, != otherwise.
66  */
67 static void smc_lgr_add_alert_token(struct smc_connection *conn)
68 {
69 	struct rb_node **link, *parent = NULL;
70 	u32 token = conn->alert_token_local;
71 
72 	link = &conn->lgr->conns_all.rb_node;
73 	while (*link) {
74 		struct smc_connection *cur = rb_entry(*link,
75 					struct smc_connection, alert_node);
76 
77 		parent = *link;
78 		if (cur->alert_token_local > token)
79 			link = &parent->rb_left;
80 		else
81 			link = &parent->rb_right;
82 	}
83 	/* Put the new node there */
84 	rb_link_node(&conn->alert_node, parent, link);
85 	rb_insert_color(&conn->alert_node, &conn->lgr->conns_all);
86 }
87 
88 /* Register connection in link group by assigning an alert token
89  * registered in a search tree.
90  * Requires @conns_lock
91  * Note that '0' is a reserved value and not assigned.
92  */
93 static void smc_lgr_register_conn(struct smc_connection *conn)
94 {
95 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
96 	static atomic_t nexttoken = ATOMIC_INIT(0);
97 
98 	/* find a new alert_token_local value not yet used by some connection
99 	 * in this link group
100 	 */
101 	sock_hold(&smc->sk); /* sock_put in smc_lgr_unregister_conn() */
102 	while (!conn->alert_token_local) {
103 		conn->alert_token_local = atomic_inc_return(&nexttoken);
104 		if (smc_lgr_find_conn(conn->alert_token_local, conn->lgr))
105 			conn->alert_token_local = 0;
106 	}
107 	smc_lgr_add_alert_token(conn);
108 	conn->lgr->conns_num++;
109 }
110 
111 /* Unregister connection and reset the alert token of the given connection<
112  */
113 static void __smc_lgr_unregister_conn(struct smc_connection *conn)
114 {
115 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
116 	struct smc_link_group *lgr = conn->lgr;
117 
118 	rb_erase(&conn->alert_node, &lgr->conns_all);
119 	lgr->conns_num--;
120 	conn->alert_token_local = 0;
121 	sock_put(&smc->sk); /* sock_hold in smc_lgr_register_conn() */
122 }
123 
124 /* Unregister connection from lgr
125  */
126 static void smc_lgr_unregister_conn(struct smc_connection *conn)
127 {
128 	struct smc_link_group *lgr = conn->lgr;
129 
130 	if (!lgr)
131 		return;
132 	write_lock_bh(&lgr->conns_lock);
133 	if (conn->alert_token_local) {
134 		__smc_lgr_unregister_conn(conn);
135 	}
136 	write_unlock_bh(&lgr->conns_lock);
137 }
138 
139 /* Send delete link, either as client to request the initiation
140  * of the DELETE LINK sequence from server; or as server to
141  * initiate the delete processing. See smc_llc_rx_delete_link().
142  */
143 static int smc_link_send_delete(struct smc_link *lnk)
144 {
145 	if (lnk->state == SMC_LNK_ACTIVE &&
146 	    !smc_llc_send_delete_link(lnk, SMC_LLC_REQ, true)) {
147 		smc_llc_link_deleting(lnk);
148 		return 0;
149 	}
150 	return -ENOTCONN;
151 }
152 
153 static void smc_lgr_free(struct smc_link_group *lgr);
154 
155 static void smc_lgr_free_work(struct work_struct *work)
156 {
157 	struct smc_link_group *lgr = container_of(to_delayed_work(work),
158 						  struct smc_link_group,
159 						  free_work);
160 	bool conns;
161 
162 	spin_lock_bh(&smc_lgr_list.lock);
163 	read_lock_bh(&lgr->conns_lock);
164 	conns = RB_EMPTY_ROOT(&lgr->conns_all);
165 	read_unlock_bh(&lgr->conns_lock);
166 	if (!conns) { /* number of lgr connections is no longer zero */
167 		spin_unlock_bh(&smc_lgr_list.lock);
168 		return;
169 	}
170 	if (!list_empty(&lgr->list))
171 		list_del_init(&lgr->list); /* remove from smc_lgr_list */
172 	spin_unlock_bh(&smc_lgr_list.lock);
173 
174 	if (!lgr->is_smcd && !lgr->terminating)	{
175 		struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
176 
177 		/* try to send del link msg, on error free lgr immediately */
178 		if (lnk->state == SMC_LNK_ACTIVE &&
179 		    !smc_link_send_delete(lnk)) {
180 			/* reschedule in case we never receive a response */
181 			smc_lgr_schedule_free_work(lgr);
182 			return;
183 		}
184 	}
185 
186 	if (!delayed_work_pending(&lgr->free_work)) {
187 		struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
188 
189 		if (!lgr->is_smcd && lnk->state != SMC_LNK_INACTIVE)
190 			smc_llc_link_inactive(lnk);
191 		if (lgr->is_smcd)
192 			smc_ism_signal_shutdown(lgr);
193 		smc_lgr_free(lgr);
194 	}
195 }
196 
197 /* create a new SMC link group */
198 static int smc_lgr_create(struct smc_sock *smc, struct smc_init_info *ini)
199 {
200 	struct smc_link_group *lgr;
201 	struct smc_link *lnk;
202 	u8 rndvec[3];
203 	int rc = 0;
204 	int i;
205 
206 	if (ini->is_smcd && ini->vlan_id) {
207 		if (smc_ism_get_vlan(ini->ism_dev, ini->vlan_id)) {
208 			rc = SMC_CLC_DECL_ISMVLANERR;
209 			goto out;
210 		}
211 	}
212 
213 	lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
214 	if (!lgr) {
215 		rc = SMC_CLC_DECL_MEM;
216 		goto out;
217 	}
218 	lgr->is_smcd = ini->is_smcd;
219 	lgr->sync_err = 0;
220 	lgr->vlan_id = ini->vlan_id;
221 	rwlock_init(&lgr->sndbufs_lock);
222 	rwlock_init(&lgr->rmbs_lock);
223 	rwlock_init(&lgr->conns_lock);
224 	for (i = 0; i < SMC_RMBE_SIZES; i++) {
225 		INIT_LIST_HEAD(&lgr->sndbufs[i]);
226 		INIT_LIST_HEAD(&lgr->rmbs[i]);
227 	}
228 	smc_lgr_list.num += SMC_LGR_NUM_INCR;
229 	memcpy(&lgr->id, (u8 *)&smc_lgr_list.num, SMC_LGR_ID_SIZE);
230 	INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
231 	lgr->conns_all = RB_ROOT;
232 	if (ini->is_smcd) {
233 		/* SMC-D specific settings */
234 		lgr->peer_gid = ini->ism_gid;
235 		lgr->smcd = ini->ism_dev;
236 	} else {
237 		/* SMC-R specific settings */
238 		lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
239 		memcpy(lgr->peer_systemid, ini->ib_lcl->id_for_peer,
240 		       SMC_SYSTEMID_LEN);
241 
242 		lnk = &lgr->lnk[SMC_SINGLE_LINK];
243 		/* initialize link */
244 		lnk->state = SMC_LNK_ACTIVATING;
245 		lnk->link_id = SMC_SINGLE_LINK;
246 		lnk->smcibdev = ini->ib_dev;
247 		lnk->ibport = ini->ib_port;
248 		lnk->path_mtu =
249 			ini->ib_dev->pattr[ini->ib_port - 1].active_mtu;
250 		if (!ini->ib_dev->initialized)
251 			smc_ib_setup_per_ibdev(ini->ib_dev);
252 		get_random_bytes(rndvec, sizeof(rndvec));
253 		lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) +
254 			(rndvec[2] << 16);
255 		rc = smc_ib_determine_gid(lnk->smcibdev, lnk->ibport,
256 					  ini->vlan_id, lnk->gid,
257 					  &lnk->sgid_index);
258 		if (rc)
259 			goto free_lgr;
260 		rc = smc_llc_link_init(lnk);
261 		if (rc)
262 			goto free_lgr;
263 		rc = smc_wr_alloc_link_mem(lnk);
264 		if (rc)
265 			goto clear_llc_lnk;
266 		rc = smc_ib_create_protection_domain(lnk);
267 		if (rc)
268 			goto free_link_mem;
269 		rc = smc_ib_create_queue_pair(lnk);
270 		if (rc)
271 			goto dealloc_pd;
272 		rc = smc_wr_create_link(lnk);
273 		if (rc)
274 			goto destroy_qp;
275 	}
276 	smc->conn.lgr = lgr;
277 	spin_lock_bh(&smc_lgr_list.lock);
278 	list_add(&lgr->list, &smc_lgr_list.list);
279 	spin_unlock_bh(&smc_lgr_list.lock);
280 	return 0;
281 
282 destroy_qp:
283 	smc_ib_destroy_queue_pair(lnk);
284 dealloc_pd:
285 	smc_ib_dealloc_protection_domain(lnk);
286 free_link_mem:
287 	smc_wr_free_link_mem(lnk);
288 clear_llc_lnk:
289 	smc_llc_link_clear(lnk);
290 free_lgr:
291 	kfree(lgr);
292 out:
293 	if (rc < 0) {
294 		if (rc == -ENOMEM)
295 			rc = SMC_CLC_DECL_MEM;
296 		else
297 			rc = SMC_CLC_DECL_INTERR;
298 	}
299 	return rc;
300 }
301 
302 static void smc_buf_unuse(struct smc_connection *conn,
303 			  struct smc_link_group *lgr)
304 {
305 	if (conn->sndbuf_desc)
306 		conn->sndbuf_desc->used = 0;
307 	if (conn->rmb_desc) {
308 		if (!conn->rmb_desc->regerr) {
309 			if (!lgr->is_smcd) {
310 				/* unregister rmb with peer */
311 				smc_llc_do_delete_rkey(
312 						&lgr->lnk[SMC_SINGLE_LINK],
313 						conn->rmb_desc);
314 			}
315 			conn->rmb_desc->used = 0;
316 		} else {
317 			/* buf registration failed, reuse not possible */
318 			write_lock_bh(&lgr->rmbs_lock);
319 			list_del(&conn->rmb_desc->list);
320 			write_unlock_bh(&lgr->rmbs_lock);
321 
322 			smc_buf_free(lgr, true, conn->rmb_desc);
323 		}
324 	}
325 }
326 
327 /* remove a finished connection from its link group */
328 void smc_conn_free(struct smc_connection *conn)
329 {
330 	struct smc_link_group *lgr = conn->lgr;
331 
332 	if (!lgr)
333 		return;
334 	if (lgr->is_smcd) {
335 		smc_ism_unset_conn(conn);
336 		tasklet_kill(&conn->rx_tsklet);
337 	} else {
338 		smc_cdc_tx_dismiss_slots(conn);
339 	}
340 	smc_lgr_unregister_conn(conn);
341 	smc_buf_unuse(conn, lgr);		/* allow buffer reuse */
342 	conn->lgr = NULL;
343 
344 	if (!lgr->conns_num)
345 		smc_lgr_schedule_free_work(lgr);
346 }
347 
348 static void smc_link_clear(struct smc_link *lnk)
349 {
350 	lnk->peer_qpn = 0;
351 	smc_llc_link_clear(lnk);
352 	smc_ib_modify_qp_reset(lnk);
353 	smc_wr_free_link(lnk);
354 	smc_ib_destroy_queue_pair(lnk);
355 	smc_ib_dealloc_protection_domain(lnk);
356 	smc_wr_free_link_mem(lnk);
357 }
358 
359 static void smcr_buf_free(struct smc_link_group *lgr, bool is_rmb,
360 			  struct smc_buf_desc *buf_desc)
361 {
362 	struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
363 
364 	if (is_rmb) {
365 		if (buf_desc->mr_rx[SMC_SINGLE_LINK])
366 			smc_ib_put_memory_region(
367 					buf_desc->mr_rx[SMC_SINGLE_LINK]);
368 		smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
369 				    DMA_FROM_DEVICE);
370 	} else {
371 		smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
372 				    DMA_TO_DEVICE);
373 	}
374 	sg_free_table(&buf_desc->sgt[SMC_SINGLE_LINK]);
375 	if (buf_desc->pages)
376 		__free_pages(buf_desc->pages, buf_desc->order);
377 	kfree(buf_desc);
378 }
379 
380 static void smcd_buf_free(struct smc_link_group *lgr, bool is_dmb,
381 			  struct smc_buf_desc *buf_desc)
382 {
383 	if (is_dmb) {
384 		/* restore original buf len */
385 		buf_desc->len += sizeof(struct smcd_cdc_msg);
386 		smc_ism_unregister_dmb(lgr->smcd, buf_desc);
387 	} else {
388 		kfree(buf_desc->cpu_addr);
389 	}
390 	kfree(buf_desc);
391 }
392 
393 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
394 			 struct smc_buf_desc *buf_desc)
395 {
396 	if (lgr->is_smcd)
397 		smcd_buf_free(lgr, is_rmb, buf_desc);
398 	else
399 		smcr_buf_free(lgr, is_rmb, buf_desc);
400 }
401 
402 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
403 {
404 	struct smc_buf_desc *buf_desc, *bf_desc;
405 	struct list_head *buf_list;
406 	int i;
407 
408 	for (i = 0; i < SMC_RMBE_SIZES; i++) {
409 		if (is_rmb)
410 			buf_list = &lgr->rmbs[i];
411 		else
412 			buf_list = &lgr->sndbufs[i];
413 		list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
414 					 list) {
415 			list_del(&buf_desc->list);
416 			smc_buf_free(lgr, is_rmb, buf_desc);
417 		}
418 	}
419 }
420 
421 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
422 {
423 	/* free send buffers */
424 	__smc_lgr_free_bufs(lgr, false);
425 	/* free rmbs */
426 	__smc_lgr_free_bufs(lgr, true);
427 }
428 
429 /* remove a link group */
430 static void smc_lgr_free(struct smc_link_group *lgr)
431 {
432 	smc_lgr_free_bufs(lgr);
433 	if (lgr->is_smcd)
434 		smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
435 	else
436 		smc_link_clear(&lgr->lnk[SMC_SINGLE_LINK]);
437 	kfree(lgr);
438 }
439 
440 void smc_lgr_forget(struct smc_link_group *lgr)
441 {
442 	spin_lock_bh(&smc_lgr_list.lock);
443 	/* do not use this link group for new connections */
444 	if (!list_empty(&lgr->list))
445 		list_del_init(&lgr->list);
446 	spin_unlock_bh(&smc_lgr_list.lock);
447 }
448 
449 /* terminate linkgroup abnormally */
450 static void __smc_lgr_terminate(struct smc_link_group *lgr)
451 {
452 	struct smc_connection *conn;
453 	struct smc_sock *smc;
454 	struct rb_node *node;
455 
456 	if (lgr->terminating)
457 		return;	/* lgr already terminating */
458 	lgr->terminating = 1;
459 	if (!list_empty(&lgr->list)) /* forget lgr */
460 		list_del_init(&lgr->list);
461 	if (!lgr->is_smcd)
462 		smc_llc_link_inactive(&lgr->lnk[SMC_SINGLE_LINK]);
463 
464 	write_lock_bh(&lgr->conns_lock);
465 	node = rb_first(&lgr->conns_all);
466 	while (node) {
467 		conn = rb_entry(node, struct smc_connection, alert_node);
468 		smc = container_of(conn, struct smc_sock, conn);
469 		sock_hold(&smc->sk); /* sock_put in close work */
470 		conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
471 		__smc_lgr_unregister_conn(conn);
472 		conn->lgr = NULL;
473 		write_unlock_bh(&lgr->conns_lock);
474 		if (!schedule_work(&conn->close_work))
475 			sock_put(&smc->sk);
476 		write_lock_bh(&lgr->conns_lock);
477 		node = rb_first(&lgr->conns_all);
478 	}
479 	write_unlock_bh(&lgr->conns_lock);
480 	if (!lgr->is_smcd)
481 		wake_up(&lgr->lnk[SMC_SINGLE_LINK].wr_reg_wait);
482 	smc_lgr_schedule_free_work(lgr);
483 }
484 
485 void smc_lgr_terminate(struct smc_link_group *lgr)
486 {
487 	spin_lock_bh(&smc_lgr_list.lock);
488 	__smc_lgr_terminate(lgr);
489 	spin_unlock_bh(&smc_lgr_list.lock);
490 }
491 
492 /* Called when IB port is terminated */
493 void smc_port_terminate(struct smc_ib_device *smcibdev, u8 ibport)
494 {
495 	struct smc_link_group *lgr, *l;
496 
497 	spin_lock_bh(&smc_lgr_list.lock);
498 	list_for_each_entry_safe(lgr, l, &smc_lgr_list.list, list) {
499 		if (!lgr->is_smcd &&
500 		    lgr->lnk[SMC_SINGLE_LINK].smcibdev == smcibdev &&
501 		    lgr->lnk[SMC_SINGLE_LINK].ibport == ibport)
502 			__smc_lgr_terminate(lgr);
503 	}
504 	spin_unlock_bh(&smc_lgr_list.lock);
505 }
506 
507 /* Called when SMC-D device is terminated or peer is lost */
508 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid, unsigned short vlan)
509 {
510 	struct smc_link_group *lgr, *l;
511 	LIST_HEAD(lgr_free_list);
512 
513 	/* run common cleanup function and build free list */
514 	spin_lock_bh(&smc_lgr_list.lock);
515 	list_for_each_entry_safe(lgr, l, &smc_lgr_list.list, list) {
516 		if (lgr->is_smcd && lgr->smcd == dev &&
517 		    (!peer_gid || lgr->peer_gid == peer_gid) &&
518 		    (vlan == VLAN_VID_MASK || lgr->vlan_id == vlan)) {
519 			__smc_lgr_terminate(lgr);
520 			list_move(&lgr->list, &lgr_free_list);
521 		}
522 	}
523 	spin_unlock_bh(&smc_lgr_list.lock);
524 
525 	/* cancel the regular free workers and actually free lgrs */
526 	list_for_each_entry_safe(lgr, l, &lgr_free_list, list) {
527 		list_del_init(&lgr->list);
528 		cancel_delayed_work_sync(&lgr->free_work);
529 		if (!peer_gid && vlan == VLAN_VID_MASK) /* dev terminated? */
530 			smc_ism_signal_shutdown(lgr);
531 		smc_lgr_free(lgr);
532 	}
533 }
534 
535 /* Determine vlan of internal TCP socket.
536  * @vlan_id: address to store the determined vlan id into
537  */
538 int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini)
539 {
540 	struct dst_entry *dst = sk_dst_get(clcsock->sk);
541 	struct net_device *ndev;
542 	int i, nest_lvl, rc = 0;
543 
544 	ini->vlan_id = 0;
545 	if (!dst) {
546 		rc = -ENOTCONN;
547 		goto out;
548 	}
549 	if (!dst->dev) {
550 		rc = -ENODEV;
551 		goto out_rel;
552 	}
553 
554 	ndev = dst->dev;
555 	if (is_vlan_dev(ndev)) {
556 		ini->vlan_id = vlan_dev_vlan_id(ndev);
557 		goto out_rel;
558 	}
559 
560 	rtnl_lock();
561 	nest_lvl = dev_get_nest_level(ndev);
562 	for (i = 0; i < nest_lvl; i++) {
563 		struct list_head *lower = &ndev->adj_list.lower;
564 
565 		if (list_empty(lower))
566 			break;
567 		lower = lower->next;
568 		ndev = (struct net_device *)netdev_lower_get_next(ndev, &lower);
569 		if (is_vlan_dev(ndev)) {
570 			ini->vlan_id = vlan_dev_vlan_id(ndev);
571 			break;
572 		}
573 	}
574 	rtnl_unlock();
575 
576 out_rel:
577 	dst_release(dst);
578 out:
579 	return rc;
580 }
581 
582 static bool smcr_lgr_match(struct smc_link_group *lgr,
583 			   struct smc_clc_msg_local *lcl,
584 			   enum smc_lgr_role role, u32 clcqpn)
585 {
586 	return !memcmp(lgr->peer_systemid, lcl->id_for_peer,
587 		       SMC_SYSTEMID_LEN) &&
588 		!memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_gid, &lcl->gid,
589 			SMC_GID_SIZE) &&
590 		!memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_mac, lcl->mac,
591 			sizeof(lcl->mac)) &&
592 		lgr->role == role &&
593 		(lgr->role == SMC_SERV ||
594 		 lgr->lnk[SMC_SINGLE_LINK].peer_qpn == clcqpn);
595 }
596 
597 static bool smcd_lgr_match(struct smc_link_group *lgr,
598 			   struct smcd_dev *smcismdev, u64 peer_gid)
599 {
600 	return lgr->peer_gid == peer_gid && lgr->smcd == smcismdev;
601 }
602 
603 /* create a new SMC connection (and a new link group if necessary) */
604 int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini)
605 {
606 	struct smc_connection *conn = &smc->conn;
607 	struct smc_link_group *lgr;
608 	enum smc_lgr_role role;
609 	int rc = 0;
610 
611 	ini->cln_first_contact = SMC_FIRST_CONTACT;
612 	role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
613 	if (role == SMC_CLNT && ini->srv_first_contact)
614 		/* create new link group as well */
615 		goto create;
616 
617 	/* determine if an existing link group can be reused */
618 	spin_lock_bh(&smc_lgr_list.lock);
619 	list_for_each_entry(lgr, &smc_lgr_list.list, list) {
620 		write_lock_bh(&lgr->conns_lock);
621 		if ((ini->is_smcd ?
622 		     smcd_lgr_match(lgr, ini->ism_dev, ini->ism_gid) :
623 		     smcr_lgr_match(lgr, ini->ib_lcl, role, ini->ib_clcqpn)) &&
624 		    !lgr->sync_err &&
625 		    lgr->vlan_id == ini->vlan_id &&
626 		    (role == SMC_CLNT ||
627 		     lgr->conns_num < SMC_RMBS_PER_LGR_MAX)) {
628 			/* link group found */
629 			ini->cln_first_contact = SMC_REUSE_CONTACT;
630 			conn->lgr = lgr;
631 			smc_lgr_register_conn(conn); /* add smc conn to lgr */
632 			if (delayed_work_pending(&lgr->free_work))
633 				cancel_delayed_work(&lgr->free_work);
634 			write_unlock_bh(&lgr->conns_lock);
635 			break;
636 		}
637 		write_unlock_bh(&lgr->conns_lock);
638 	}
639 	spin_unlock_bh(&smc_lgr_list.lock);
640 
641 	if (role == SMC_CLNT && !ini->srv_first_contact &&
642 	    ini->cln_first_contact == SMC_FIRST_CONTACT) {
643 		/* Server reuses a link group, but Client wants to start
644 		 * a new one
645 		 * send out_of_sync decline, reason synchr. error
646 		 */
647 		return SMC_CLC_DECL_SYNCERR;
648 	}
649 
650 create:
651 	if (ini->cln_first_contact == SMC_FIRST_CONTACT) {
652 		rc = smc_lgr_create(smc, ini);
653 		if (rc)
654 			goto out;
655 		lgr = conn->lgr;
656 		write_lock_bh(&lgr->conns_lock);
657 		smc_lgr_register_conn(conn); /* add smc conn to lgr */
658 		write_unlock_bh(&lgr->conns_lock);
659 	}
660 	conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
661 	conn->local_tx_ctrl.len = SMC_WR_TX_SIZE;
662 	conn->urg_state = SMC_URG_READ;
663 	if (ini->is_smcd) {
664 		conn->rx_off = sizeof(struct smcd_cdc_msg);
665 		smcd_cdc_rx_init(conn); /* init tasklet for this conn */
666 	}
667 #ifndef KERNEL_HAS_ATOMIC64
668 	spin_lock_init(&conn->acurs_lock);
669 #endif
670 
671 out:
672 	return rc;
673 }
674 
675 /* convert the RMB size into the compressed notation - minimum 16K.
676  * In contrast to plain ilog2, this rounds towards the next power of 2,
677  * so the socket application gets at least its desired sndbuf / rcvbuf size.
678  */
679 static u8 smc_compress_bufsize(int size)
680 {
681 	u8 compressed;
682 
683 	if (size <= SMC_BUF_MIN_SIZE)
684 		return 0;
685 
686 	size = (size - 1) >> 14;
687 	compressed = ilog2(size) + 1;
688 	if (compressed >= SMC_RMBE_SIZES)
689 		compressed = SMC_RMBE_SIZES - 1;
690 	return compressed;
691 }
692 
693 /* convert the RMB size from compressed notation into integer */
694 int smc_uncompress_bufsize(u8 compressed)
695 {
696 	u32 size;
697 
698 	size = 0x00000001 << (((int)compressed) + 14);
699 	return (int)size;
700 }
701 
702 /* try to reuse a sndbuf or rmb description slot for a certain
703  * buffer size; if not available, return NULL
704  */
705 static struct smc_buf_desc *smc_buf_get_slot(int compressed_bufsize,
706 					     rwlock_t *lock,
707 					     struct list_head *buf_list)
708 {
709 	struct smc_buf_desc *buf_slot;
710 
711 	read_lock_bh(lock);
712 	list_for_each_entry(buf_slot, buf_list, list) {
713 		if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
714 			read_unlock_bh(lock);
715 			return buf_slot;
716 		}
717 	}
718 	read_unlock_bh(lock);
719 	return NULL;
720 }
721 
722 /* one of the conditions for announcing a receiver's current window size is
723  * that it "results in a minimum increase in the window size of 10% of the
724  * receive buffer space" [RFC7609]
725  */
726 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
727 {
728 	return min_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
729 }
730 
731 static struct smc_buf_desc *smcr_new_buf_create(struct smc_link_group *lgr,
732 						bool is_rmb, int bufsize)
733 {
734 	struct smc_buf_desc *buf_desc;
735 	struct smc_link *lnk;
736 	int rc;
737 
738 	/* try to alloc a new buffer */
739 	buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
740 	if (!buf_desc)
741 		return ERR_PTR(-ENOMEM);
742 
743 	buf_desc->order = get_order(bufsize);
744 	buf_desc->pages = alloc_pages(GFP_KERNEL | __GFP_NOWARN |
745 				      __GFP_NOMEMALLOC | __GFP_COMP |
746 				      __GFP_NORETRY | __GFP_ZERO,
747 				      buf_desc->order);
748 	if (!buf_desc->pages) {
749 		kfree(buf_desc);
750 		return ERR_PTR(-EAGAIN);
751 	}
752 	buf_desc->cpu_addr = (void *)page_address(buf_desc->pages);
753 
754 	/* build the sg table from the pages */
755 	lnk = &lgr->lnk[SMC_SINGLE_LINK];
756 	rc = sg_alloc_table(&buf_desc->sgt[SMC_SINGLE_LINK], 1,
757 			    GFP_KERNEL);
758 	if (rc) {
759 		smc_buf_free(lgr, is_rmb, buf_desc);
760 		return ERR_PTR(rc);
761 	}
762 	sg_set_buf(buf_desc->sgt[SMC_SINGLE_LINK].sgl,
763 		   buf_desc->cpu_addr, bufsize);
764 
765 	/* map sg table to DMA address */
766 	rc = smc_ib_buf_map_sg(lnk->smcibdev, buf_desc,
767 			       is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
768 	/* SMC protocol depends on mapping to one DMA address only */
769 	if (rc != 1)  {
770 		smc_buf_free(lgr, is_rmb, buf_desc);
771 		return ERR_PTR(-EAGAIN);
772 	}
773 
774 	/* create a new memory region for the RMB */
775 	if (is_rmb) {
776 		rc = smc_ib_get_memory_region(lnk->roce_pd,
777 					      IB_ACCESS_REMOTE_WRITE |
778 					      IB_ACCESS_LOCAL_WRITE,
779 					      buf_desc);
780 		if (rc) {
781 			smc_buf_free(lgr, is_rmb, buf_desc);
782 			return ERR_PTR(rc);
783 		}
784 	}
785 
786 	buf_desc->len = bufsize;
787 	return buf_desc;
788 }
789 
790 #define SMCD_DMBE_SIZES		7 /* 0 -> 16KB, 1 -> 32KB, .. 6 -> 1MB */
791 
792 static struct smc_buf_desc *smcd_new_buf_create(struct smc_link_group *lgr,
793 						bool is_dmb, int bufsize)
794 {
795 	struct smc_buf_desc *buf_desc;
796 	int rc;
797 
798 	if (smc_compress_bufsize(bufsize) > SMCD_DMBE_SIZES)
799 		return ERR_PTR(-EAGAIN);
800 
801 	/* try to alloc a new DMB */
802 	buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
803 	if (!buf_desc)
804 		return ERR_PTR(-ENOMEM);
805 	if (is_dmb) {
806 		rc = smc_ism_register_dmb(lgr, bufsize, buf_desc);
807 		if (rc) {
808 			kfree(buf_desc);
809 			return ERR_PTR(-EAGAIN);
810 		}
811 		buf_desc->pages = virt_to_page(buf_desc->cpu_addr);
812 		/* CDC header stored in buf. So, pretend it was smaller */
813 		buf_desc->len = bufsize - sizeof(struct smcd_cdc_msg);
814 	} else {
815 		buf_desc->cpu_addr = kzalloc(bufsize, GFP_KERNEL |
816 					     __GFP_NOWARN | __GFP_NORETRY |
817 					     __GFP_NOMEMALLOC);
818 		if (!buf_desc->cpu_addr) {
819 			kfree(buf_desc);
820 			return ERR_PTR(-EAGAIN);
821 		}
822 		buf_desc->len = bufsize;
823 	}
824 	return buf_desc;
825 }
826 
827 static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
828 {
829 	struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
830 	struct smc_connection *conn = &smc->conn;
831 	struct smc_link_group *lgr = conn->lgr;
832 	struct list_head *buf_list;
833 	int bufsize, bufsize_short;
834 	int sk_buf_size;
835 	rwlock_t *lock;
836 
837 	if (is_rmb)
838 		/* use socket recv buffer size (w/o overhead) as start value */
839 		sk_buf_size = smc->sk.sk_rcvbuf / 2;
840 	else
841 		/* use socket send buffer size (w/o overhead) as start value */
842 		sk_buf_size = smc->sk.sk_sndbuf / 2;
843 
844 	for (bufsize_short = smc_compress_bufsize(sk_buf_size);
845 	     bufsize_short >= 0; bufsize_short--) {
846 
847 		if (is_rmb) {
848 			lock = &lgr->rmbs_lock;
849 			buf_list = &lgr->rmbs[bufsize_short];
850 		} else {
851 			lock = &lgr->sndbufs_lock;
852 			buf_list = &lgr->sndbufs[bufsize_short];
853 		}
854 		bufsize = smc_uncompress_bufsize(bufsize_short);
855 		if ((1 << get_order(bufsize)) > SG_MAX_SINGLE_ALLOC)
856 			continue;
857 
858 		/* check for reusable slot in the link group */
859 		buf_desc = smc_buf_get_slot(bufsize_short, lock, buf_list);
860 		if (buf_desc) {
861 			memset(buf_desc->cpu_addr, 0, bufsize);
862 			break; /* found reusable slot */
863 		}
864 
865 		if (is_smcd)
866 			buf_desc = smcd_new_buf_create(lgr, is_rmb, bufsize);
867 		else
868 			buf_desc = smcr_new_buf_create(lgr, is_rmb, bufsize);
869 
870 		if (PTR_ERR(buf_desc) == -ENOMEM)
871 			break;
872 		if (IS_ERR(buf_desc))
873 			continue;
874 
875 		buf_desc->used = 1;
876 		write_lock_bh(lock);
877 		list_add(&buf_desc->list, buf_list);
878 		write_unlock_bh(lock);
879 		break; /* found */
880 	}
881 
882 	if (IS_ERR(buf_desc))
883 		return -ENOMEM;
884 
885 	if (is_rmb) {
886 		conn->rmb_desc = buf_desc;
887 		conn->rmbe_size_short = bufsize_short;
888 		smc->sk.sk_rcvbuf = bufsize * 2;
889 		atomic_set(&conn->bytes_to_rcv, 0);
890 		conn->rmbe_update_limit =
891 			smc_rmb_wnd_update_limit(buf_desc->len);
892 		if (is_smcd)
893 			smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */
894 	} else {
895 		conn->sndbuf_desc = buf_desc;
896 		smc->sk.sk_sndbuf = bufsize * 2;
897 		atomic_set(&conn->sndbuf_space, bufsize);
898 	}
899 	return 0;
900 }
901 
902 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn)
903 {
904 	struct smc_link_group *lgr = conn->lgr;
905 
906 	if (!conn->lgr || conn->lgr->is_smcd)
907 		return;
908 	smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
909 			       conn->sndbuf_desc, DMA_TO_DEVICE);
910 }
911 
912 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
913 {
914 	struct smc_link_group *lgr = conn->lgr;
915 
916 	if (!conn->lgr || conn->lgr->is_smcd)
917 		return;
918 	smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
919 				  conn->sndbuf_desc, DMA_TO_DEVICE);
920 }
921 
922 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
923 {
924 	struct smc_link_group *lgr = conn->lgr;
925 
926 	if (!conn->lgr || conn->lgr->is_smcd)
927 		return;
928 	smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
929 			       conn->rmb_desc, DMA_FROM_DEVICE);
930 }
931 
932 void smc_rmb_sync_sg_for_device(struct smc_connection *conn)
933 {
934 	struct smc_link_group *lgr = conn->lgr;
935 
936 	if (!conn->lgr || conn->lgr->is_smcd)
937 		return;
938 	smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
939 				  conn->rmb_desc, DMA_FROM_DEVICE);
940 }
941 
942 /* create the send and receive buffer for an SMC socket;
943  * receive buffers are called RMBs;
944  * (even though the SMC protocol allows more than one RMB-element per RMB,
945  * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
946  * extra RMB for every connection in a link group
947  */
948 int smc_buf_create(struct smc_sock *smc, bool is_smcd)
949 {
950 	int rc;
951 
952 	/* create send buffer */
953 	rc = __smc_buf_create(smc, is_smcd, false);
954 	if (rc)
955 		return rc;
956 	/* create rmb */
957 	rc = __smc_buf_create(smc, is_smcd, true);
958 	if (rc)
959 		smc_buf_free(smc->conn.lgr, false, smc->conn.sndbuf_desc);
960 	return rc;
961 }
962 
963 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
964 {
965 	int i;
966 
967 	for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
968 		if (!test_and_set_bit(i, lgr->rtokens_used_mask))
969 			return i;
970 	}
971 	return -ENOSPC;
972 }
973 
974 /* add a new rtoken from peer */
975 int smc_rtoken_add(struct smc_link_group *lgr, __be64 nw_vaddr, __be32 nw_rkey)
976 {
977 	u64 dma_addr = be64_to_cpu(nw_vaddr);
978 	u32 rkey = ntohl(nw_rkey);
979 	int i;
980 
981 	for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
982 		if ((lgr->rtokens[i][SMC_SINGLE_LINK].rkey == rkey) &&
983 		    (lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr == dma_addr) &&
984 		    test_bit(i, lgr->rtokens_used_mask)) {
985 			/* already in list */
986 			return i;
987 		}
988 	}
989 	i = smc_rmb_reserve_rtoken_idx(lgr);
990 	if (i < 0)
991 		return i;
992 	lgr->rtokens[i][SMC_SINGLE_LINK].rkey = rkey;
993 	lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr = dma_addr;
994 	return i;
995 }
996 
997 /* delete an rtoken */
998 int smc_rtoken_delete(struct smc_link_group *lgr, __be32 nw_rkey)
999 {
1000 	u32 rkey = ntohl(nw_rkey);
1001 	int i;
1002 
1003 	for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1004 		if (lgr->rtokens[i][SMC_SINGLE_LINK].rkey == rkey &&
1005 		    test_bit(i, lgr->rtokens_used_mask)) {
1006 			lgr->rtokens[i][SMC_SINGLE_LINK].rkey = 0;
1007 			lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr = 0;
1008 
1009 			clear_bit(i, lgr->rtokens_used_mask);
1010 			return 0;
1011 		}
1012 	}
1013 	return -ENOENT;
1014 }
1015 
1016 /* save rkey and dma_addr received from peer during clc handshake */
1017 int smc_rmb_rtoken_handling(struct smc_connection *conn,
1018 			    struct smc_clc_msg_accept_confirm *clc)
1019 {
1020 	conn->rtoken_idx = smc_rtoken_add(conn->lgr, clc->rmb_dma_addr,
1021 					  clc->rmb_rkey);
1022 	if (conn->rtoken_idx < 0)
1023 		return conn->rtoken_idx;
1024 	return 0;
1025 }
1026 
1027 /* Called (from smc_exit) when module is removed */
1028 void smc_core_exit(void)
1029 {
1030 	struct smc_link_group *lgr, *lg;
1031 	LIST_HEAD(lgr_freeing_list);
1032 
1033 	spin_lock_bh(&smc_lgr_list.lock);
1034 	if (!list_empty(&smc_lgr_list.list))
1035 		list_splice_init(&smc_lgr_list.list, &lgr_freeing_list);
1036 	spin_unlock_bh(&smc_lgr_list.lock);
1037 	list_for_each_entry_safe(lgr, lg, &lgr_freeing_list, list) {
1038 		list_del_init(&lgr->list);
1039 		if (!lgr->is_smcd) {
1040 			struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
1041 
1042 			if (lnk->state == SMC_LNK_ACTIVE)
1043 				smc_llc_send_delete_link(lnk, SMC_LLC_REQ,
1044 							 false);
1045 			smc_llc_link_inactive(lnk);
1046 		}
1047 		cancel_delayed_work_sync(&lgr->free_work);
1048 		if (lgr->is_smcd)
1049 			smc_ism_signal_shutdown(lgr);
1050 		smc_lgr_free(lgr); /* free link group */
1051 	}
1052 }
1053