xref: /openbmc/linux/net/smc/smc_core.c (revision eb7ebfa3)
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 	conn->lgr = NULL;
122 	sock_put(&smc->sk); /* sock_hold in smc_lgr_register_conn() */
123 }
124 
125 /* Unregister connection and trigger lgr freeing if applicable
126  */
127 static void smc_lgr_unregister_conn(struct smc_connection *conn)
128 {
129 	struct smc_link_group *lgr = conn->lgr;
130 	int reduced = 0;
131 
132 	write_lock_bh(&lgr->conns_lock);
133 	if (conn->alert_token_local) {
134 		reduced = 1;
135 		__smc_lgr_unregister_conn(conn);
136 	}
137 	write_unlock_bh(&lgr->conns_lock);
138 	if (!reduced || lgr->conns_num)
139 		return;
140 	smc_lgr_schedule_free_work(lgr);
141 }
142 
143 /* Send delete link, either as client to request the initiation
144  * of the DELETE LINK sequence from server; or as server to
145  * initiate the delete processing. See smc_llc_rx_delete_link().
146  */
147 static int smc_link_send_delete(struct smc_link *lnk)
148 {
149 	if (lnk->state == SMC_LNK_ACTIVE &&
150 	    !smc_llc_send_delete_link(lnk, SMC_LLC_REQ, true)) {
151 		smc_llc_link_deleting(lnk);
152 		return 0;
153 	}
154 	return -ENOTCONN;
155 }
156 
157 static void smc_lgr_free_work(struct work_struct *work)
158 {
159 	struct smc_link_group *lgr = container_of(to_delayed_work(work),
160 						  struct smc_link_group,
161 						  free_work);
162 	bool conns;
163 
164 	spin_lock_bh(&smc_lgr_list.lock);
165 	if (list_empty(&lgr->list))
166 		goto free;
167 	read_lock_bh(&lgr->conns_lock);
168 	conns = RB_EMPTY_ROOT(&lgr->conns_all);
169 	read_unlock_bh(&lgr->conns_lock);
170 	if (!conns) { /* number of lgr connections is no longer zero */
171 		spin_unlock_bh(&smc_lgr_list.lock);
172 		return;
173 	}
174 	list_del_init(&lgr->list); /* remove from smc_lgr_list */
175 free:
176 	spin_unlock_bh(&smc_lgr_list.lock);
177 
178 	if (!lgr->is_smcd && !lgr->terminating)	{
179 		/* try to send del link msg, on error free lgr immediately */
180 		if (!smc_link_send_delete(&lgr->lnk[SMC_SINGLE_LINK])) {
181 			/* reschedule in case we never receive a response */
182 			smc_lgr_schedule_free_work(lgr);
183 			return;
184 		}
185 	}
186 
187 	if (!delayed_work_pending(&lgr->free_work)) {
188 		struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
189 
190 		if (!lgr->is_smcd && lnk->state != SMC_LNK_INACTIVE)
191 			smc_llc_link_inactive(lnk);
192 		smc_lgr_free(lgr);
193 	}
194 }
195 
196 /* create a new SMC link group */
197 static int smc_lgr_create(struct smc_sock *smc, bool is_smcd,
198 			  struct smc_ib_device *smcibdev, u8 ibport,
199 			  char *peer_systemid, unsigned short vlan_id,
200 			  struct smcd_dev *smcismdev, u64 peer_gid)
201 {
202 	struct smc_link_group *lgr;
203 	struct smc_link *lnk;
204 	u8 rndvec[3];
205 	int rc = 0;
206 	int i;
207 
208 	if (is_smcd && vlan_id) {
209 		rc = smc_ism_get_vlan(smcismdev, vlan_id);
210 		if (rc)
211 			goto out;
212 	}
213 
214 	lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
215 	if (!lgr) {
216 		rc = -ENOMEM;
217 		goto out;
218 	}
219 	lgr->is_smcd = is_smcd;
220 	lgr->sync_err = 0;
221 	lgr->vlan_id = vlan_id;
222 	rwlock_init(&lgr->sndbufs_lock);
223 	rwlock_init(&lgr->rmbs_lock);
224 	rwlock_init(&lgr->conns_lock);
225 	for (i = 0; i < SMC_RMBE_SIZES; i++) {
226 		INIT_LIST_HEAD(&lgr->sndbufs[i]);
227 		INIT_LIST_HEAD(&lgr->rmbs[i]);
228 	}
229 	smc_lgr_list.num += SMC_LGR_NUM_INCR;
230 	memcpy(&lgr->id, (u8 *)&smc_lgr_list.num, SMC_LGR_ID_SIZE);
231 	INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
232 	lgr->conns_all = RB_ROOT;
233 	if (is_smcd) {
234 		/* SMC-D specific settings */
235 		lgr->peer_gid = peer_gid;
236 		lgr->smcd = smcismdev;
237 	} else {
238 		/* SMC-R specific settings */
239 		lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
240 		memcpy(lgr->peer_systemid, peer_systemid, 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 = smcibdev;
247 		lnk->ibport = ibport;
248 		lnk->path_mtu = smcibdev->pattr[ibport - 1].active_mtu;
249 		if (!smcibdev->initialized)
250 			smc_ib_setup_per_ibdev(smcibdev);
251 		get_random_bytes(rndvec, sizeof(rndvec));
252 		lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) +
253 			(rndvec[2] << 16);
254 		rc = smc_ib_determine_gid(lnk->smcibdev, lnk->ibport,
255 					  vlan_id, lnk->gid, &lnk->sgid_index);
256 		if (rc)
257 			goto free_lgr;
258 		rc = smc_llc_link_init(lnk);
259 		if (rc)
260 			goto free_lgr;
261 		rc = smc_wr_alloc_link_mem(lnk);
262 		if (rc)
263 			goto clear_llc_lnk;
264 		rc = smc_ib_create_protection_domain(lnk);
265 		if (rc)
266 			goto free_link_mem;
267 		rc = smc_ib_create_queue_pair(lnk);
268 		if (rc)
269 			goto dealloc_pd;
270 		rc = smc_wr_create_link(lnk);
271 		if (rc)
272 			goto destroy_qp;
273 	}
274 	smc->conn.lgr = lgr;
275 	spin_lock_bh(&smc_lgr_list.lock);
276 	list_add(&lgr->list, &smc_lgr_list.list);
277 	spin_unlock_bh(&smc_lgr_list.lock);
278 	return 0;
279 
280 destroy_qp:
281 	smc_ib_destroy_queue_pair(lnk);
282 dealloc_pd:
283 	smc_ib_dealloc_protection_domain(lnk);
284 free_link_mem:
285 	smc_wr_free_link_mem(lnk);
286 clear_llc_lnk:
287 	smc_llc_link_clear(lnk);
288 free_lgr:
289 	kfree(lgr);
290 out:
291 	return rc;
292 }
293 
294 static void smc_buf_unuse(struct smc_connection *conn)
295 {
296 	if (conn->sndbuf_desc)
297 		conn->sndbuf_desc->used = 0;
298 	if (conn->rmb_desc) {
299 		if (!conn->rmb_desc->regerr) {
300 			conn->rmb_desc->reused = 1;
301 			conn->rmb_desc->used = 0;
302 		} else {
303 			/* buf registration failed, reuse not possible */
304 			struct smc_link_group *lgr = conn->lgr;
305 
306 			write_lock_bh(&lgr->rmbs_lock);
307 			list_del(&conn->rmb_desc->list);
308 			write_unlock_bh(&lgr->rmbs_lock);
309 
310 			smc_buf_free(lgr, true, conn->rmb_desc);
311 		}
312 	}
313 }
314 
315 /* remove a finished connection from its link group */
316 void smc_conn_free(struct smc_connection *conn)
317 {
318 	if (!conn->lgr)
319 		return;
320 	if (conn->lgr->is_smcd) {
321 		smc_ism_unset_conn(conn);
322 		tasklet_kill(&conn->rx_tsklet);
323 	} else {
324 		smc_cdc_tx_dismiss_slots(conn);
325 	}
326 	smc_lgr_unregister_conn(conn);
327 	smc_buf_unuse(conn);
328 }
329 
330 static void smc_link_clear(struct smc_link *lnk)
331 {
332 	lnk->peer_qpn = 0;
333 	smc_llc_link_clear(lnk);
334 	smc_ib_modify_qp_reset(lnk);
335 	smc_wr_free_link(lnk);
336 	smc_ib_destroy_queue_pair(lnk);
337 	smc_ib_dealloc_protection_domain(lnk);
338 	smc_wr_free_link_mem(lnk);
339 }
340 
341 static void smcr_buf_free(struct smc_link_group *lgr, bool is_rmb,
342 			  struct smc_buf_desc *buf_desc)
343 {
344 	struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
345 
346 	if (is_rmb) {
347 		if (buf_desc->mr_rx[SMC_SINGLE_LINK])
348 			smc_ib_put_memory_region(
349 					buf_desc->mr_rx[SMC_SINGLE_LINK]);
350 		smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
351 				    DMA_FROM_DEVICE);
352 	} else {
353 		smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
354 				    DMA_TO_DEVICE);
355 	}
356 	sg_free_table(&buf_desc->sgt[SMC_SINGLE_LINK]);
357 	if (buf_desc->pages)
358 		__free_pages(buf_desc->pages, buf_desc->order);
359 	kfree(buf_desc);
360 }
361 
362 static void smcd_buf_free(struct smc_link_group *lgr, bool is_dmb,
363 			  struct smc_buf_desc *buf_desc)
364 {
365 	if (is_dmb) {
366 		/* restore original buf len */
367 		buf_desc->len += sizeof(struct smcd_cdc_msg);
368 		smc_ism_unregister_dmb(lgr->smcd, buf_desc);
369 	} else {
370 		kfree(buf_desc->cpu_addr);
371 	}
372 	kfree(buf_desc);
373 }
374 
375 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
376 			 struct smc_buf_desc *buf_desc)
377 {
378 	if (lgr->is_smcd)
379 		smcd_buf_free(lgr, is_rmb, buf_desc);
380 	else
381 		smcr_buf_free(lgr, is_rmb, buf_desc);
382 }
383 
384 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
385 {
386 	struct smc_buf_desc *buf_desc, *bf_desc;
387 	struct list_head *buf_list;
388 	int i;
389 
390 	for (i = 0; i < SMC_RMBE_SIZES; i++) {
391 		if (is_rmb)
392 			buf_list = &lgr->rmbs[i];
393 		else
394 			buf_list = &lgr->sndbufs[i];
395 		list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
396 					 list) {
397 			list_del(&buf_desc->list);
398 			smc_buf_free(lgr, is_rmb, buf_desc);
399 		}
400 	}
401 }
402 
403 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
404 {
405 	/* free send buffers */
406 	__smc_lgr_free_bufs(lgr, false);
407 	/* free rmbs */
408 	__smc_lgr_free_bufs(lgr, true);
409 }
410 
411 /* remove a link group */
412 void smc_lgr_free(struct smc_link_group *lgr)
413 {
414 	smc_lgr_free_bufs(lgr);
415 	if (lgr->is_smcd)
416 		smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
417 	else
418 		smc_link_clear(&lgr->lnk[SMC_SINGLE_LINK]);
419 	kfree(lgr);
420 }
421 
422 void smc_lgr_forget(struct smc_link_group *lgr)
423 {
424 	spin_lock_bh(&smc_lgr_list.lock);
425 	/* do not use this link group for new connections */
426 	if (!list_empty(&lgr->list))
427 		list_del_init(&lgr->list);
428 	spin_unlock_bh(&smc_lgr_list.lock);
429 }
430 
431 /* terminate linkgroup abnormally */
432 static void __smc_lgr_terminate(struct smc_link_group *lgr)
433 {
434 	struct smc_connection *conn;
435 	struct smc_sock *smc;
436 	struct rb_node *node;
437 
438 	if (lgr->terminating)
439 		return;	/* lgr already terminating */
440 	lgr->terminating = 1;
441 	if (!list_empty(&lgr->list)) /* forget lgr */
442 		list_del_init(&lgr->list);
443 	if (!lgr->is_smcd)
444 		smc_llc_link_inactive(&lgr->lnk[SMC_SINGLE_LINK]);
445 
446 	write_lock_bh(&lgr->conns_lock);
447 	node = rb_first(&lgr->conns_all);
448 	while (node) {
449 		conn = rb_entry(node, struct smc_connection, alert_node);
450 		smc = container_of(conn, struct smc_sock, conn);
451 		sock_hold(&smc->sk); /* sock_put in close work */
452 		conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
453 		__smc_lgr_unregister_conn(conn);
454 		write_unlock_bh(&lgr->conns_lock);
455 		if (!schedule_work(&conn->close_work))
456 			sock_put(&smc->sk);
457 		write_lock_bh(&lgr->conns_lock);
458 		node = rb_first(&lgr->conns_all);
459 	}
460 	write_unlock_bh(&lgr->conns_lock);
461 	if (!lgr->is_smcd)
462 		wake_up(&lgr->lnk[SMC_SINGLE_LINK].wr_reg_wait);
463 	smc_lgr_schedule_free_work(lgr);
464 }
465 
466 void smc_lgr_terminate(struct smc_link_group *lgr)
467 {
468 	spin_lock_bh(&smc_lgr_list.lock);
469 	__smc_lgr_terminate(lgr);
470 	spin_unlock_bh(&smc_lgr_list.lock);
471 }
472 
473 /* Called when IB port is terminated */
474 void smc_port_terminate(struct smc_ib_device *smcibdev, u8 ibport)
475 {
476 	struct smc_link_group *lgr, *l;
477 
478 	spin_lock_bh(&smc_lgr_list.lock);
479 	list_for_each_entry_safe(lgr, l, &smc_lgr_list.list, list) {
480 		if (!lgr->is_smcd &&
481 		    lgr->lnk[SMC_SINGLE_LINK].smcibdev == smcibdev &&
482 		    lgr->lnk[SMC_SINGLE_LINK].ibport == ibport)
483 			__smc_lgr_terminate(lgr);
484 	}
485 	spin_unlock_bh(&smc_lgr_list.lock);
486 }
487 
488 /* Called when SMC-D device is terminated or peer is lost */
489 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid)
490 {
491 	struct smc_link_group *lgr, *l;
492 	LIST_HEAD(lgr_free_list);
493 
494 	/* run common cleanup function and build free list */
495 	spin_lock_bh(&smc_lgr_list.lock);
496 	list_for_each_entry_safe(lgr, l, &smc_lgr_list.list, list) {
497 		if (lgr->is_smcd && lgr->smcd == dev &&
498 		    (!peer_gid || lgr->peer_gid == peer_gid) &&
499 		    !list_empty(&lgr->list)) {
500 			__smc_lgr_terminate(lgr);
501 			list_move(&lgr->list, &lgr_free_list);
502 		}
503 	}
504 	spin_unlock_bh(&smc_lgr_list.lock);
505 
506 	/* cancel the regular free workers and actually free lgrs */
507 	list_for_each_entry_safe(lgr, l, &lgr_free_list, list) {
508 		list_del_init(&lgr->list);
509 		cancel_delayed_work_sync(&lgr->free_work);
510 		smc_lgr_free(lgr);
511 	}
512 }
513 
514 /* Determine vlan of internal TCP socket.
515  * @vlan_id: address to store the determined vlan id into
516  */
517 int smc_vlan_by_tcpsk(struct socket *clcsock, unsigned short *vlan_id)
518 {
519 	struct dst_entry *dst = sk_dst_get(clcsock->sk);
520 	struct net_device *ndev;
521 	int i, nest_lvl, rc = 0;
522 
523 	*vlan_id = 0;
524 	if (!dst) {
525 		rc = -ENOTCONN;
526 		goto out;
527 	}
528 	if (!dst->dev) {
529 		rc = -ENODEV;
530 		goto out_rel;
531 	}
532 
533 	ndev = dst->dev;
534 	if (is_vlan_dev(ndev)) {
535 		*vlan_id = vlan_dev_vlan_id(ndev);
536 		goto out_rel;
537 	}
538 
539 	rtnl_lock();
540 	nest_lvl = dev_get_nest_level(ndev);
541 	for (i = 0; i < nest_lvl; i++) {
542 		struct list_head *lower = &ndev->adj_list.lower;
543 
544 		if (list_empty(lower))
545 			break;
546 		lower = lower->next;
547 		ndev = (struct net_device *)netdev_lower_get_next(ndev, &lower);
548 		if (is_vlan_dev(ndev)) {
549 			*vlan_id = vlan_dev_vlan_id(ndev);
550 			break;
551 		}
552 	}
553 	rtnl_unlock();
554 
555 out_rel:
556 	dst_release(dst);
557 out:
558 	return rc;
559 }
560 
561 static bool smcr_lgr_match(struct smc_link_group *lgr,
562 			   struct smc_clc_msg_local *lcl,
563 			   enum smc_lgr_role role)
564 {
565 	return !memcmp(lgr->peer_systemid, lcl->id_for_peer,
566 		       SMC_SYSTEMID_LEN) &&
567 		!memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_gid, &lcl->gid,
568 			SMC_GID_SIZE) &&
569 		!memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_mac, lcl->mac,
570 			sizeof(lcl->mac)) &&
571 		lgr->role == role;
572 }
573 
574 static bool smcd_lgr_match(struct smc_link_group *lgr,
575 			   struct smcd_dev *smcismdev, u64 peer_gid)
576 {
577 	return lgr->peer_gid == peer_gid && lgr->smcd == smcismdev;
578 }
579 
580 /* create a new SMC connection (and a new link group if necessary) */
581 int smc_conn_create(struct smc_sock *smc, bool is_smcd, int srv_first_contact,
582 		    struct smc_ib_device *smcibdev, u8 ibport,
583 		    struct smc_clc_msg_local *lcl, struct smcd_dev *smcd,
584 		    u64 peer_gid)
585 {
586 	struct smc_connection *conn = &smc->conn;
587 	int local_contact = SMC_FIRST_CONTACT;
588 	struct smc_link_group *lgr;
589 	unsigned short vlan_id;
590 	enum smc_lgr_role role;
591 	int rc = 0;
592 
593 	role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
594 	rc = smc_vlan_by_tcpsk(smc->clcsock, &vlan_id);
595 	if (rc)
596 		return rc;
597 
598 	if ((role == SMC_CLNT) && srv_first_contact)
599 		/* create new link group as well */
600 		goto create;
601 
602 	/* determine if an existing link group can be reused */
603 	spin_lock_bh(&smc_lgr_list.lock);
604 	list_for_each_entry(lgr, &smc_lgr_list.list, list) {
605 		write_lock_bh(&lgr->conns_lock);
606 		if ((is_smcd ? smcd_lgr_match(lgr, smcd, peer_gid) :
607 		     smcr_lgr_match(lgr, lcl, role)) &&
608 		    !lgr->sync_err &&
609 		    lgr->vlan_id == vlan_id &&
610 		    (role == SMC_CLNT ||
611 		     lgr->conns_num < SMC_RMBS_PER_LGR_MAX)) {
612 			/* link group found */
613 			local_contact = SMC_REUSE_CONTACT;
614 			conn->lgr = lgr;
615 			smc_lgr_register_conn(conn); /* add smc conn to lgr */
616 			write_unlock_bh(&lgr->conns_lock);
617 			break;
618 		}
619 		write_unlock_bh(&lgr->conns_lock);
620 	}
621 	spin_unlock_bh(&smc_lgr_list.lock);
622 
623 	if (role == SMC_CLNT && !srv_first_contact &&
624 	    (local_contact == SMC_FIRST_CONTACT)) {
625 		/* Server reuses a link group, but Client wants to start
626 		 * a new one
627 		 * send out_of_sync decline, reason synchr. error
628 		 */
629 		return -ENOLINK;
630 	}
631 
632 create:
633 	if (local_contact == SMC_FIRST_CONTACT) {
634 		rc = smc_lgr_create(smc, is_smcd, smcibdev, ibport,
635 				    lcl->id_for_peer, vlan_id, smcd, peer_gid);
636 		if (rc)
637 			goto out;
638 		smc_lgr_register_conn(conn); /* add smc conn to lgr */
639 	}
640 	conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
641 	conn->local_tx_ctrl.len = SMC_WR_TX_SIZE;
642 	conn->urg_state = SMC_URG_READ;
643 	if (is_smcd) {
644 		conn->rx_off = sizeof(struct smcd_cdc_msg);
645 		smcd_cdc_rx_init(conn); /* init tasklet for this conn */
646 	}
647 #ifndef KERNEL_HAS_ATOMIC64
648 	spin_lock_init(&conn->acurs_lock);
649 #endif
650 
651 out:
652 	return rc ? rc : local_contact;
653 }
654 
655 /* convert the RMB size into the compressed notation - minimum 16K.
656  * In contrast to plain ilog2, this rounds towards the next power of 2,
657  * so the socket application gets at least its desired sndbuf / rcvbuf size.
658  */
659 static u8 smc_compress_bufsize(int size)
660 {
661 	u8 compressed;
662 
663 	if (size <= SMC_BUF_MIN_SIZE)
664 		return 0;
665 
666 	size = (size - 1) >> 14;
667 	compressed = ilog2(size) + 1;
668 	if (compressed >= SMC_RMBE_SIZES)
669 		compressed = SMC_RMBE_SIZES - 1;
670 	return compressed;
671 }
672 
673 /* convert the RMB size from compressed notation into integer */
674 int smc_uncompress_bufsize(u8 compressed)
675 {
676 	u32 size;
677 
678 	size = 0x00000001 << (((int)compressed) + 14);
679 	return (int)size;
680 }
681 
682 /* try to reuse a sndbuf or rmb description slot for a certain
683  * buffer size; if not available, return NULL
684  */
685 static struct smc_buf_desc *smc_buf_get_slot(int compressed_bufsize,
686 					     rwlock_t *lock,
687 					     struct list_head *buf_list)
688 {
689 	struct smc_buf_desc *buf_slot;
690 
691 	read_lock_bh(lock);
692 	list_for_each_entry(buf_slot, buf_list, list) {
693 		if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
694 			read_unlock_bh(lock);
695 			return buf_slot;
696 		}
697 	}
698 	read_unlock_bh(lock);
699 	return NULL;
700 }
701 
702 /* one of the conditions for announcing a receiver's current window size is
703  * that it "results in a minimum increase in the window size of 10% of the
704  * receive buffer space" [RFC7609]
705  */
706 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
707 {
708 	return min_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
709 }
710 
711 static struct smc_buf_desc *smcr_new_buf_create(struct smc_link_group *lgr,
712 						bool is_rmb, int bufsize)
713 {
714 	struct smc_buf_desc *buf_desc;
715 	struct smc_link *lnk;
716 	int rc;
717 
718 	/* try to alloc a new buffer */
719 	buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
720 	if (!buf_desc)
721 		return ERR_PTR(-ENOMEM);
722 
723 	buf_desc->order = get_order(bufsize);
724 	buf_desc->pages = alloc_pages(GFP_KERNEL | __GFP_NOWARN |
725 				      __GFP_NOMEMALLOC | __GFP_COMP |
726 				      __GFP_NORETRY | __GFP_ZERO,
727 				      buf_desc->order);
728 	if (!buf_desc->pages) {
729 		kfree(buf_desc);
730 		return ERR_PTR(-EAGAIN);
731 	}
732 	buf_desc->cpu_addr = (void *)page_address(buf_desc->pages);
733 
734 	/* build the sg table from the pages */
735 	lnk = &lgr->lnk[SMC_SINGLE_LINK];
736 	rc = sg_alloc_table(&buf_desc->sgt[SMC_SINGLE_LINK], 1,
737 			    GFP_KERNEL);
738 	if (rc) {
739 		smc_buf_free(lgr, is_rmb, buf_desc);
740 		return ERR_PTR(rc);
741 	}
742 	sg_set_buf(buf_desc->sgt[SMC_SINGLE_LINK].sgl,
743 		   buf_desc->cpu_addr, bufsize);
744 
745 	/* map sg table to DMA address */
746 	rc = smc_ib_buf_map_sg(lnk->smcibdev, buf_desc,
747 			       is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
748 	/* SMC protocol depends on mapping to one DMA address only */
749 	if (rc != 1)  {
750 		smc_buf_free(lgr, is_rmb, buf_desc);
751 		return ERR_PTR(-EAGAIN);
752 	}
753 
754 	/* create a new memory region for the RMB */
755 	if (is_rmb) {
756 		rc = smc_ib_get_memory_region(lnk->roce_pd,
757 					      IB_ACCESS_REMOTE_WRITE |
758 					      IB_ACCESS_LOCAL_WRITE,
759 					      buf_desc);
760 		if (rc) {
761 			smc_buf_free(lgr, is_rmb, buf_desc);
762 			return ERR_PTR(rc);
763 		}
764 	}
765 
766 	buf_desc->len = bufsize;
767 	return buf_desc;
768 }
769 
770 #define SMCD_DMBE_SIZES		7 /* 0 -> 16KB, 1 -> 32KB, .. 6 -> 1MB */
771 
772 static struct smc_buf_desc *smcd_new_buf_create(struct smc_link_group *lgr,
773 						bool is_dmb, int bufsize)
774 {
775 	struct smc_buf_desc *buf_desc;
776 	int rc;
777 
778 	if (smc_compress_bufsize(bufsize) > SMCD_DMBE_SIZES)
779 		return ERR_PTR(-EAGAIN);
780 
781 	/* try to alloc a new DMB */
782 	buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
783 	if (!buf_desc)
784 		return ERR_PTR(-ENOMEM);
785 	if (is_dmb) {
786 		rc = smc_ism_register_dmb(lgr, bufsize, buf_desc);
787 		if (rc) {
788 			kfree(buf_desc);
789 			return ERR_PTR(-EAGAIN);
790 		}
791 		buf_desc->pages = virt_to_page(buf_desc->cpu_addr);
792 		/* CDC header stored in buf. So, pretend it was smaller */
793 		buf_desc->len = bufsize - sizeof(struct smcd_cdc_msg);
794 	} else {
795 		buf_desc->cpu_addr = kzalloc(bufsize, GFP_KERNEL |
796 					     __GFP_NOWARN | __GFP_NORETRY |
797 					     __GFP_NOMEMALLOC);
798 		if (!buf_desc->cpu_addr) {
799 			kfree(buf_desc);
800 			return ERR_PTR(-EAGAIN);
801 		}
802 		buf_desc->len = bufsize;
803 	}
804 	return buf_desc;
805 }
806 
807 static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
808 {
809 	struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
810 	struct smc_connection *conn = &smc->conn;
811 	struct smc_link_group *lgr = conn->lgr;
812 	struct list_head *buf_list;
813 	int bufsize, bufsize_short;
814 	int sk_buf_size;
815 	rwlock_t *lock;
816 
817 	if (is_rmb)
818 		/* use socket recv buffer size (w/o overhead) as start value */
819 		sk_buf_size = smc->sk.sk_rcvbuf / 2;
820 	else
821 		/* use socket send buffer size (w/o overhead) as start value */
822 		sk_buf_size = smc->sk.sk_sndbuf / 2;
823 
824 	for (bufsize_short = smc_compress_bufsize(sk_buf_size);
825 	     bufsize_short >= 0; bufsize_short--) {
826 
827 		if (is_rmb) {
828 			lock = &lgr->rmbs_lock;
829 			buf_list = &lgr->rmbs[bufsize_short];
830 		} else {
831 			lock = &lgr->sndbufs_lock;
832 			buf_list = &lgr->sndbufs[bufsize_short];
833 		}
834 		bufsize = smc_uncompress_bufsize(bufsize_short);
835 		if ((1 << get_order(bufsize)) > SG_MAX_SINGLE_ALLOC)
836 			continue;
837 
838 		/* check for reusable slot in the link group */
839 		buf_desc = smc_buf_get_slot(bufsize_short, lock, buf_list);
840 		if (buf_desc) {
841 			memset(buf_desc->cpu_addr, 0, bufsize);
842 			break; /* found reusable slot */
843 		}
844 
845 		if (is_smcd)
846 			buf_desc = smcd_new_buf_create(lgr, is_rmb, bufsize);
847 		else
848 			buf_desc = smcr_new_buf_create(lgr, is_rmb, bufsize);
849 
850 		if (PTR_ERR(buf_desc) == -ENOMEM)
851 			break;
852 		if (IS_ERR(buf_desc))
853 			continue;
854 
855 		buf_desc->used = 1;
856 		write_lock_bh(lock);
857 		list_add(&buf_desc->list, buf_list);
858 		write_unlock_bh(lock);
859 		break; /* found */
860 	}
861 
862 	if (IS_ERR(buf_desc))
863 		return -ENOMEM;
864 
865 	if (is_rmb) {
866 		conn->rmb_desc = buf_desc;
867 		conn->rmbe_size_short = bufsize_short;
868 		smc->sk.sk_rcvbuf = bufsize * 2;
869 		atomic_set(&conn->bytes_to_rcv, 0);
870 		conn->rmbe_update_limit =
871 			smc_rmb_wnd_update_limit(buf_desc->len);
872 		if (is_smcd)
873 			smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */
874 	} else {
875 		conn->sndbuf_desc = buf_desc;
876 		smc->sk.sk_sndbuf = bufsize * 2;
877 		atomic_set(&conn->sndbuf_space, bufsize);
878 	}
879 	return 0;
880 }
881 
882 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn)
883 {
884 	struct smc_link_group *lgr = conn->lgr;
885 
886 	if (!conn->lgr || conn->lgr->is_smcd)
887 		return;
888 	smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
889 			       conn->sndbuf_desc, DMA_TO_DEVICE);
890 }
891 
892 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
893 {
894 	struct smc_link_group *lgr = conn->lgr;
895 
896 	if (!conn->lgr || conn->lgr->is_smcd)
897 		return;
898 	smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
899 				  conn->sndbuf_desc, DMA_TO_DEVICE);
900 }
901 
902 void smc_rmb_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->rmb_desc, DMA_FROM_DEVICE);
910 }
911 
912 void smc_rmb_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->rmb_desc, DMA_FROM_DEVICE);
920 }
921 
922 /* create the send and receive buffer for an SMC socket;
923  * receive buffers are called RMBs;
924  * (even though the SMC protocol allows more than one RMB-element per RMB,
925  * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
926  * extra RMB for every connection in a link group
927  */
928 int smc_buf_create(struct smc_sock *smc, bool is_smcd)
929 {
930 	int rc;
931 
932 	/* create send buffer */
933 	rc = __smc_buf_create(smc, is_smcd, false);
934 	if (rc)
935 		return rc;
936 	/* create rmb */
937 	rc = __smc_buf_create(smc, is_smcd, true);
938 	if (rc)
939 		smc_buf_free(smc->conn.lgr, false, smc->conn.sndbuf_desc);
940 	return rc;
941 }
942 
943 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
944 {
945 	int i;
946 
947 	for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
948 		if (!test_and_set_bit(i, lgr->rtokens_used_mask))
949 			return i;
950 	}
951 	return -ENOSPC;
952 }
953 
954 /* add a new rtoken from peer */
955 int smc_rtoken_add(struct smc_link_group *lgr, __be64 nw_vaddr, __be32 nw_rkey)
956 {
957 	u64 dma_addr = be64_to_cpu(nw_vaddr);
958 	u32 rkey = ntohl(nw_rkey);
959 	int i;
960 
961 	for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
962 		if ((lgr->rtokens[i][SMC_SINGLE_LINK].rkey == rkey) &&
963 		    (lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr == dma_addr) &&
964 		    test_bit(i, lgr->rtokens_used_mask)) {
965 			/* already in list */
966 			return i;
967 		}
968 	}
969 	i = smc_rmb_reserve_rtoken_idx(lgr);
970 	if (i < 0)
971 		return i;
972 	lgr->rtokens[i][SMC_SINGLE_LINK].rkey = rkey;
973 	lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr = dma_addr;
974 	return i;
975 }
976 
977 /* delete an rtoken */
978 int smc_rtoken_delete(struct smc_link_group *lgr, __be32 nw_rkey)
979 {
980 	u32 rkey = ntohl(nw_rkey);
981 	int i;
982 
983 	for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
984 		if (lgr->rtokens[i][SMC_SINGLE_LINK].rkey == rkey &&
985 		    test_bit(i, lgr->rtokens_used_mask)) {
986 			lgr->rtokens[i][SMC_SINGLE_LINK].rkey = 0;
987 			lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr = 0;
988 
989 			clear_bit(i, lgr->rtokens_used_mask);
990 			return 0;
991 		}
992 	}
993 	return -ENOENT;
994 }
995 
996 /* save rkey and dma_addr received from peer during clc handshake */
997 int smc_rmb_rtoken_handling(struct smc_connection *conn,
998 			    struct smc_clc_msg_accept_confirm *clc)
999 {
1000 	conn->rtoken_idx = smc_rtoken_add(conn->lgr, clc->rmb_dma_addr,
1001 					  clc->rmb_rkey);
1002 	if (conn->rtoken_idx < 0)
1003 		return conn->rtoken_idx;
1004 	return 0;
1005 }
1006 
1007 /* Called (from smc_exit) when module is removed */
1008 void smc_core_exit(void)
1009 {
1010 	struct smc_link_group *lgr, *lg;
1011 	LIST_HEAD(lgr_freeing_list);
1012 
1013 	spin_lock_bh(&smc_lgr_list.lock);
1014 	if (!list_empty(&smc_lgr_list.list))
1015 		list_splice_init(&smc_lgr_list.list, &lgr_freeing_list);
1016 	spin_unlock_bh(&smc_lgr_list.lock);
1017 	list_for_each_entry_safe(lgr, lg, &lgr_freeing_list, list) {
1018 		list_del_init(&lgr->list);
1019 		if (!lgr->is_smcd) {
1020 			struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
1021 
1022 			if (lnk->state == SMC_LNK_ACTIVE)
1023 				smc_llc_send_delete_link(lnk, SMC_LLC_REQ,
1024 							 false);
1025 			smc_llc_link_inactive(lnk);
1026 		}
1027 		cancel_delayed_work_sync(&lgr->free_work);
1028 		smc_lgr_free(lgr); /* free link group */
1029 	}
1030 }
1031