xref: /linux/net/smc/smc_core.c (revision a44e4f3ab16bc808590763a543a93b6fbf3abcc4)
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 ism_put_vlan;
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 ism_put_vlan:
293 	if (ini->is_smcd && ini->vlan_id)
294 		smc_ism_put_vlan(ini->ism_dev, ini->vlan_id);
295 out:
296 	if (rc < 0) {
297 		if (rc == -ENOMEM)
298 			rc = SMC_CLC_DECL_MEM;
299 		else
300 			rc = SMC_CLC_DECL_INTERR;
301 	}
302 	return rc;
303 }
304 
305 static void smc_buf_unuse(struct smc_connection *conn,
306 			  struct smc_link_group *lgr)
307 {
308 	if (conn->sndbuf_desc)
309 		conn->sndbuf_desc->used = 0;
310 	if (conn->rmb_desc) {
311 		if (!conn->rmb_desc->regerr) {
312 			if (!lgr->is_smcd) {
313 				/* unregister rmb with peer */
314 				smc_llc_do_delete_rkey(
315 						&lgr->lnk[SMC_SINGLE_LINK],
316 						conn->rmb_desc);
317 			}
318 			conn->rmb_desc->used = 0;
319 		} else {
320 			/* buf registration failed, reuse not possible */
321 			write_lock_bh(&lgr->rmbs_lock);
322 			list_del(&conn->rmb_desc->list);
323 			write_unlock_bh(&lgr->rmbs_lock);
324 
325 			smc_buf_free(lgr, true, conn->rmb_desc);
326 		}
327 	}
328 }
329 
330 /* remove a finished connection from its link group */
331 void smc_conn_free(struct smc_connection *conn)
332 {
333 	struct smc_link_group *lgr = conn->lgr;
334 
335 	if (!lgr)
336 		return;
337 	if (lgr->is_smcd) {
338 		smc_ism_unset_conn(conn);
339 		tasklet_kill(&conn->rx_tsklet);
340 	} else {
341 		smc_cdc_tx_dismiss_slots(conn);
342 	}
343 	smc_lgr_unregister_conn(conn);
344 	smc_buf_unuse(conn, lgr);		/* allow buffer reuse */
345 	conn->lgr = NULL;
346 
347 	if (!lgr->conns_num)
348 		smc_lgr_schedule_free_work(lgr);
349 }
350 
351 static void smc_link_clear(struct smc_link *lnk)
352 {
353 	lnk->peer_qpn = 0;
354 	smc_llc_link_clear(lnk);
355 	smc_ib_modify_qp_reset(lnk);
356 	smc_wr_free_link(lnk);
357 	smc_ib_destroy_queue_pair(lnk);
358 	smc_ib_dealloc_protection_domain(lnk);
359 	smc_wr_free_link_mem(lnk);
360 }
361 
362 static void smcr_buf_free(struct smc_link_group *lgr, bool is_rmb,
363 			  struct smc_buf_desc *buf_desc)
364 {
365 	struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
366 
367 	if (is_rmb) {
368 		if (buf_desc->mr_rx[SMC_SINGLE_LINK])
369 			smc_ib_put_memory_region(
370 					buf_desc->mr_rx[SMC_SINGLE_LINK]);
371 		smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
372 				    DMA_FROM_DEVICE);
373 	} else {
374 		smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
375 				    DMA_TO_DEVICE);
376 	}
377 	sg_free_table(&buf_desc->sgt[SMC_SINGLE_LINK]);
378 	if (buf_desc->pages)
379 		__free_pages(buf_desc->pages, buf_desc->order);
380 	kfree(buf_desc);
381 }
382 
383 static void smcd_buf_free(struct smc_link_group *lgr, bool is_dmb,
384 			  struct smc_buf_desc *buf_desc)
385 {
386 	if (is_dmb) {
387 		/* restore original buf len */
388 		buf_desc->len += sizeof(struct smcd_cdc_msg);
389 		smc_ism_unregister_dmb(lgr->smcd, buf_desc);
390 	} else {
391 		kfree(buf_desc->cpu_addr);
392 	}
393 	kfree(buf_desc);
394 }
395 
396 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
397 			 struct smc_buf_desc *buf_desc)
398 {
399 	if (lgr->is_smcd)
400 		smcd_buf_free(lgr, is_rmb, buf_desc);
401 	else
402 		smcr_buf_free(lgr, is_rmb, buf_desc);
403 }
404 
405 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
406 {
407 	struct smc_buf_desc *buf_desc, *bf_desc;
408 	struct list_head *buf_list;
409 	int i;
410 
411 	for (i = 0; i < SMC_RMBE_SIZES; i++) {
412 		if (is_rmb)
413 			buf_list = &lgr->rmbs[i];
414 		else
415 			buf_list = &lgr->sndbufs[i];
416 		list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
417 					 list) {
418 			list_del(&buf_desc->list);
419 			smc_buf_free(lgr, is_rmb, buf_desc);
420 		}
421 	}
422 }
423 
424 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
425 {
426 	/* free send buffers */
427 	__smc_lgr_free_bufs(lgr, false);
428 	/* free rmbs */
429 	__smc_lgr_free_bufs(lgr, true);
430 }
431 
432 /* remove a link group */
433 static void smc_lgr_free(struct smc_link_group *lgr)
434 {
435 	smc_lgr_free_bufs(lgr);
436 	if (lgr->is_smcd)
437 		smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
438 	else
439 		smc_link_clear(&lgr->lnk[SMC_SINGLE_LINK]);
440 	kfree(lgr);
441 }
442 
443 void smc_lgr_forget(struct smc_link_group *lgr)
444 {
445 	spin_lock_bh(&smc_lgr_list.lock);
446 	/* do not use this link group for new connections */
447 	if (!list_empty(&lgr->list))
448 		list_del_init(&lgr->list);
449 	spin_unlock_bh(&smc_lgr_list.lock);
450 }
451 
452 /* terminate linkgroup abnormally */
453 static void __smc_lgr_terminate(struct smc_link_group *lgr)
454 {
455 	struct smc_connection *conn;
456 	struct smc_sock *smc;
457 	struct rb_node *node;
458 
459 	if (lgr->terminating)
460 		return;	/* lgr already terminating */
461 	lgr->terminating = 1;
462 	if (!list_empty(&lgr->list)) /* forget lgr */
463 		list_del_init(&lgr->list);
464 	if (!lgr->is_smcd)
465 		smc_llc_link_inactive(&lgr->lnk[SMC_SINGLE_LINK]);
466 
467 	write_lock_bh(&lgr->conns_lock);
468 	node = rb_first(&lgr->conns_all);
469 	while (node) {
470 		conn = rb_entry(node, struct smc_connection, alert_node);
471 		smc = container_of(conn, struct smc_sock, conn);
472 		sock_hold(&smc->sk); /* sock_put in close work */
473 		conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
474 		__smc_lgr_unregister_conn(conn);
475 		conn->lgr = NULL;
476 		write_unlock_bh(&lgr->conns_lock);
477 		if (!schedule_work(&conn->close_work))
478 			sock_put(&smc->sk);
479 		write_lock_bh(&lgr->conns_lock);
480 		node = rb_first(&lgr->conns_all);
481 	}
482 	write_unlock_bh(&lgr->conns_lock);
483 	if (!lgr->is_smcd)
484 		wake_up(&lgr->lnk[SMC_SINGLE_LINK].wr_reg_wait);
485 	smc_lgr_schedule_free_work(lgr);
486 }
487 
488 void smc_lgr_terminate(struct smc_link_group *lgr)
489 {
490 	spin_lock_bh(&smc_lgr_list.lock);
491 	__smc_lgr_terminate(lgr);
492 	spin_unlock_bh(&smc_lgr_list.lock);
493 }
494 
495 /* Called when IB port is terminated */
496 void smc_port_terminate(struct smc_ib_device *smcibdev, u8 ibport)
497 {
498 	struct smc_link_group *lgr, *l;
499 
500 	spin_lock_bh(&smc_lgr_list.lock);
501 	list_for_each_entry_safe(lgr, l, &smc_lgr_list.list, list) {
502 		if (!lgr->is_smcd &&
503 		    lgr->lnk[SMC_SINGLE_LINK].smcibdev == smcibdev &&
504 		    lgr->lnk[SMC_SINGLE_LINK].ibport == ibport)
505 			__smc_lgr_terminate(lgr);
506 	}
507 	spin_unlock_bh(&smc_lgr_list.lock);
508 }
509 
510 /* Called when SMC-D device is terminated or peer is lost */
511 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid, unsigned short vlan)
512 {
513 	struct smc_link_group *lgr, *l;
514 	LIST_HEAD(lgr_free_list);
515 
516 	/* run common cleanup function and build free list */
517 	spin_lock_bh(&smc_lgr_list.lock);
518 	list_for_each_entry_safe(lgr, l, &smc_lgr_list.list, list) {
519 		if (lgr->is_smcd && lgr->smcd == dev &&
520 		    (!peer_gid || lgr->peer_gid == peer_gid) &&
521 		    (vlan == VLAN_VID_MASK || lgr->vlan_id == vlan)) {
522 			__smc_lgr_terminate(lgr);
523 			list_move(&lgr->list, &lgr_free_list);
524 		}
525 	}
526 	spin_unlock_bh(&smc_lgr_list.lock);
527 
528 	/* cancel the regular free workers and actually free lgrs */
529 	list_for_each_entry_safe(lgr, l, &lgr_free_list, list) {
530 		list_del_init(&lgr->list);
531 		cancel_delayed_work_sync(&lgr->free_work);
532 		if (!peer_gid && vlan == VLAN_VID_MASK) /* dev terminated? */
533 			smc_ism_signal_shutdown(lgr);
534 		smc_lgr_free(lgr);
535 	}
536 }
537 
538 /* Determine vlan of internal TCP socket.
539  * @vlan_id: address to store the determined vlan id into
540  */
541 int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini)
542 {
543 	struct dst_entry *dst = sk_dst_get(clcsock->sk);
544 	struct net_device *ndev;
545 	int i, nest_lvl, rc = 0;
546 
547 	ini->vlan_id = 0;
548 	if (!dst) {
549 		rc = -ENOTCONN;
550 		goto out;
551 	}
552 	if (!dst->dev) {
553 		rc = -ENODEV;
554 		goto out_rel;
555 	}
556 
557 	ndev = dst->dev;
558 	if (is_vlan_dev(ndev)) {
559 		ini->vlan_id = vlan_dev_vlan_id(ndev);
560 		goto out_rel;
561 	}
562 
563 	rtnl_lock();
564 	nest_lvl = ndev->lower_level;
565 	for (i = 0; i < nest_lvl; i++) {
566 		struct list_head *lower = &ndev->adj_list.lower;
567 
568 		if (list_empty(lower))
569 			break;
570 		lower = lower->next;
571 		ndev = (struct net_device *)netdev_lower_get_next(ndev, &lower);
572 		if (is_vlan_dev(ndev)) {
573 			ini->vlan_id = vlan_dev_vlan_id(ndev);
574 			break;
575 		}
576 	}
577 	rtnl_unlock();
578 
579 out_rel:
580 	dst_release(dst);
581 out:
582 	return rc;
583 }
584 
585 static bool smcr_lgr_match(struct smc_link_group *lgr,
586 			   struct smc_clc_msg_local *lcl,
587 			   enum smc_lgr_role role, u32 clcqpn)
588 {
589 	return !memcmp(lgr->peer_systemid, lcl->id_for_peer,
590 		       SMC_SYSTEMID_LEN) &&
591 		!memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_gid, &lcl->gid,
592 			SMC_GID_SIZE) &&
593 		!memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_mac, lcl->mac,
594 			sizeof(lcl->mac)) &&
595 		lgr->role == role &&
596 		(lgr->role == SMC_SERV ||
597 		 lgr->lnk[SMC_SINGLE_LINK].peer_qpn == clcqpn);
598 }
599 
600 static bool smcd_lgr_match(struct smc_link_group *lgr,
601 			   struct smcd_dev *smcismdev, u64 peer_gid)
602 {
603 	return lgr->peer_gid == peer_gid && lgr->smcd == smcismdev;
604 }
605 
606 /* create a new SMC connection (and a new link group if necessary) */
607 int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini)
608 {
609 	struct smc_connection *conn = &smc->conn;
610 	struct smc_link_group *lgr;
611 	enum smc_lgr_role role;
612 	int rc = 0;
613 
614 	ini->cln_first_contact = SMC_FIRST_CONTACT;
615 	role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
616 	if (role == SMC_CLNT && ini->srv_first_contact)
617 		/* create new link group as well */
618 		goto create;
619 
620 	/* determine if an existing link group can be reused */
621 	spin_lock_bh(&smc_lgr_list.lock);
622 	list_for_each_entry(lgr, &smc_lgr_list.list, list) {
623 		write_lock_bh(&lgr->conns_lock);
624 		if ((ini->is_smcd ?
625 		     smcd_lgr_match(lgr, ini->ism_dev, ini->ism_gid) :
626 		     smcr_lgr_match(lgr, ini->ib_lcl, role, ini->ib_clcqpn)) &&
627 		    !lgr->sync_err &&
628 		    lgr->vlan_id == ini->vlan_id &&
629 		    (role == SMC_CLNT ||
630 		     lgr->conns_num < SMC_RMBS_PER_LGR_MAX)) {
631 			/* link group found */
632 			ini->cln_first_contact = SMC_REUSE_CONTACT;
633 			conn->lgr = lgr;
634 			smc_lgr_register_conn(conn); /* add smc conn to lgr */
635 			if (delayed_work_pending(&lgr->free_work))
636 				cancel_delayed_work(&lgr->free_work);
637 			write_unlock_bh(&lgr->conns_lock);
638 			break;
639 		}
640 		write_unlock_bh(&lgr->conns_lock);
641 	}
642 	spin_unlock_bh(&smc_lgr_list.lock);
643 
644 	if (role == SMC_CLNT && !ini->srv_first_contact &&
645 	    ini->cln_first_contact == SMC_FIRST_CONTACT) {
646 		/* Server reuses a link group, but Client wants to start
647 		 * a new one
648 		 * send out_of_sync decline, reason synchr. error
649 		 */
650 		return SMC_CLC_DECL_SYNCERR;
651 	}
652 
653 create:
654 	if (ini->cln_first_contact == SMC_FIRST_CONTACT) {
655 		rc = smc_lgr_create(smc, ini);
656 		if (rc)
657 			goto out;
658 		lgr = conn->lgr;
659 		write_lock_bh(&lgr->conns_lock);
660 		smc_lgr_register_conn(conn); /* add smc conn to lgr */
661 		write_unlock_bh(&lgr->conns_lock);
662 	}
663 	conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
664 	conn->local_tx_ctrl.len = SMC_WR_TX_SIZE;
665 	conn->urg_state = SMC_URG_READ;
666 	if (ini->is_smcd) {
667 		conn->rx_off = sizeof(struct smcd_cdc_msg);
668 		smcd_cdc_rx_init(conn); /* init tasklet for this conn */
669 	}
670 #ifndef KERNEL_HAS_ATOMIC64
671 	spin_lock_init(&conn->acurs_lock);
672 #endif
673 
674 out:
675 	return rc;
676 }
677 
678 /* convert the RMB size into the compressed notation - minimum 16K.
679  * In contrast to plain ilog2, this rounds towards the next power of 2,
680  * so the socket application gets at least its desired sndbuf / rcvbuf size.
681  */
682 static u8 smc_compress_bufsize(int size)
683 {
684 	u8 compressed;
685 
686 	if (size <= SMC_BUF_MIN_SIZE)
687 		return 0;
688 
689 	size = (size - 1) >> 14;
690 	compressed = ilog2(size) + 1;
691 	if (compressed >= SMC_RMBE_SIZES)
692 		compressed = SMC_RMBE_SIZES - 1;
693 	return compressed;
694 }
695 
696 /* convert the RMB size from compressed notation into integer */
697 int smc_uncompress_bufsize(u8 compressed)
698 {
699 	u32 size;
700 
701 	size = 0x00000001 << (((int)compressed) + 14);
702 	return (int)size;
703 }
704 
705 /* try to reuse a sndbuf or rmb description slot for a certain
706  * buffer size; if not available, return NULL
707  */
708 static struct smc_buf_desc *smc_buf_get_slot(int compressed_bufsize,
709 					     rwlock_t *lock,
710 					     struct list_head *buf_list)
711 {
712 	struct smc_buf_desc *buf_slot;
713 
714 	read_lock_bh(lock);
715 	list_for_each_entry(buf_slot, buf_list, list) {
716 		if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
717 			read_unlock_bh(lock);
718 			return buf_slot;
719 		}
720 	}
721 	read_unlock_bh(lock);
722 	return NULL;
723 }
724 
725 /* one of the conditions for announcing a receiver's current window size is
726  * that it "results in a minimum increase in the window size of 10% of the
727  * receive buffer space" [RFC7609]
728  */
729 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
730 {
731 	return min_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
732 }
733 
734 static struct smc_buf_desc *smcr_new_buf_create(struct smc_link_group *lgr,
735 						bool is_rmb, int bufsize)
736 {
737 	struct smc_buf_desc *buf_desc;
738 	struct smc_link *lnk;
739 	int rc;
740 
741 	/* try to alloc a new buffer */
742 	buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
743 	if (!buf_desc)
744 		return ERR_PTR(-ENOMEM);
745 
746 	buf_desc->order = get_order(bufsize);
747 	buf_desc->pages = alloc_pages(GFP_KERNEL | __GFP_NOWARN |
748 				      __GFP_NOMEMALLOC | __GFP_COMP |
749 				      __GFP_NORETRY | __GFP_ZERO,
750 				      buf_desc->order);
751 	if (!buf_desc->pages) {
752 		kfree(buf_desc);
753 		return ERR_PTR(-EAGAIN);
754 	}
755 	buf_desc->cpu_addr = (void *)page_address(buf_desc->pages);
756 
757 	/* build the sg table from the pages */
758 	lnk = &lgr->lnk[SMC_SINGLE_LINK];
759 	rc = sg_alloc_table(&buf_desc->sgt[SMC_SINGLE_LINK], 1,
760 			    GFP_KERNEL);
761 	if (rc) {
762 		smc_buf_free(lgr, is_rmb, buf_desc);
763 		return ERR_PTR(rc);
764 	}
765 	sg_set_buf(buf_desc->sgt[SMC_SINGLE_LINK].sgl,
766 		   buf_desc->cpu_addr, bufsize);
767 
768 	/* map sg table to DMA address */
769 	rc = smc_ib_buf_map_sg(lnk->smcibdev, buf_desc,
770 			       is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
771 	/* SMC protocol depends on mapping to one DMA address only */
772 	if (rc != 1)  {
773 		smc_buf_free(lgr, is_rmb, buf_desc);
774 		return ERR_PTR(-EAGAIN);
775 	}
776 
777 	/* create a new memory region for the RMB */
778 	if (is_rmb) {
779 		rc = smc_ib_get_memory_region(lnk->roce_pd,
780 					      IB_ACCESS_REMOTE_WRITE |
781 					      IB_ACCESS_LOCAL_WRITE,
782 					      buf_desc);
783 		if (rc) {
784 			smc_buf_free(lgr, is_rmb, buf_desc);
785 			return ERR_PTR(rc);
786 		}
787 	}
788 
789 	buf_desc->len = bufsize;
790 	return buf_desc;
791 }
792 
793 #define SMCD_DMBE_SIZES		7 /* 0 -> 16KB, 1 -> 32KB, .. 6 -> 1MB */
794 
795 static struct smc_buf_desc *smcd_new_buf_create(struct smc_link_group *lgr,
796 						bool is_dmb, int bufsize)
797 {
798 	struct smc_buf_desc *buf_desc;
799 	int rc;
800 
801 	if (smc_compress_bufsize(bufsize) > SMCD_DMBE_SIZES)
802 		return ERR_PTR(-EAGAIN);
803 
804 	/* try to alloc a new DMB */
805 	buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
806 	if (!buf_desc)
807 		return ERR_PTR(-ENOMEM);
808 	if (is_dmb) {
809 		rc = smc_ism_register_dmb(lgr, bufsize, buf_desc);
810 		if (rc) {
811 			kfree(buf_desc);
812 			return ERR_PTR(-EAGAIN);
813 		}
814 		buf_desc->pages = virt_to_page(buf_desc->cpu_addr);
815 		/* CDC header stored in buf. So, pretend it was smaller */
816 		buf_desc->len = bufsize - sizeof(struct smcd_cdc_msg);
817 	} else {
818 		buf_desc->cpu_addr = kzalloc(bufsize, GFP_KERNEL |
819 					     __GFP_NOWARN | __GFP_NORETRY |
820 					     __GFP_NOMEMALLOC);
821 		if (!buf_desc->cpu_addr) {
822 			kfree(buf_desc);
823 			return ERR_PTR(-EAGAIN);
824 		}
825 		buf_desc->len = bufsize;
826 	}
827 	return buf_desc;
828 }
829 
830 static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
831 {
832 	struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
833 	struct smc_connection *conn = &smc->conn;
834 	struct smc_link_group *lgr = conn->lgr;
835 	struct list_head *buf_list;
836 	int bufsize, bufsize_short;
837 	int sk_buf_size;
838 	rwlock_t *lock;
839 
840 	if (is_rmb)
841 		/* use socket recv buffer size (w/o overhead) as start value */
842 		sk_buf_size = smc->sk.sk_rcvbuf / 2;
843 	else
844 		/* use socket send buffer size (w/o overhead) as start value */
845 		sk_buf_size = smc->sk.sk_sndbuf / 2;
846 
847 	for (bufsize_short = smc_compress_bufsize(sk_buf_size);
848 	     bufsize_short >= 0; bufsize_short--) {
849 
850 		if (is_rmb) {
851 			lock = &lgr->rmbs_lock;
852 			buf_list = &lgr->rmbs[bufsize_short];
853 		} else {
854 			lock = &lgr->sndbufs_lock;
855 			buf_list = &lgr->sndbufs[bufsize_short];
856 		}
857 		bufsize = smc_uncompress_bufsize(bufsize_short);
858 		if ((1 << get_order(bufsize)) > SG_MAX_SINGLE_ALLOC)
859 			continue;
860 
861 		/* check for reusable slot in the link group */
862 		buf_desc = smc_buf_get_slot(bufsize_short, lock, buf_list);
863 		if (buf_desc) {
864 			memset(buf_desc->cpu_addr, 0, bufsize);
865 			break; /* found reusable slot */
866 		}
867 
868 		if (is_smcd)
869 			buf_desc = smcd_new_buf_create(lgr, is_rmb, bufsize);
870 		else
871 			buf_desc = smcr_new_buf_create(lgr, is_rmb, bufsize);
872 
873 		if (PTR_ERR(buf_desc) == -ENOMEM)
874 			break;
875 		if (IS_ERR(buf_desc))
876 			continue;
877 
878 		buf_desc->used = 1;
879 		write_lock_bh(lock);
880 		list_add(&buf_desc->list, buf_list);
881 		write_unlock_bh(lock);
882 		break; /* found */
883 	}
884 
885 	if (IS_ERR(buf_desc))
886 		return -ENOMEM;
887 
888 	if (is_rmb) {
889 		conn->rmb_desc = buf_desc;
890 		conn->rmbe_size_short = bufsize_short;
891 		smc->sk.sk_rcvbuf = bufsize * 2;
892 		atomic_set(&conn->bytes_to_rcv, 0);
893 		conn->rmbe_update_limit =
894 			smc_rmb_wnd_update_limit(buf_desc->len);
895 		if (is_smcd)
896 			smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */
897 	} else {
898 		conn->sndbuf_desc = buf_desc;
899 		smc->sk.sk_sndbuf = bufsize * 2;
900 		atomic_set(&conn->sndbuf_space, bufsize);
901 	}
902 	return 0;
903 }
904 
905 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn)
906 {
907 	struct smc_link_group *lgr = conn->lgr;
908 
909 	if (!conn->lgr || conn->lgr->is_smcd)
910 		return;
911 	smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
912 			       conn->sndbuf_desc, DMA_TO_DEVICE);
913 }
914 
915 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
916 {
917 	struct smc_link_group *lgr = conn->lgr;
918 
919 	if (!conn->lgr || conn->lgr->is_smcd)
920 		return;
921 	smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
922 				  conn->sndbuf_desc, DMA_TO_DEVICE);
923 }
924 
925 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
926 {
927 	struct smc_link_group *lgr = conn->lgr;
928 
929 	if (!conn->lgr || conn->lgr->is_smcd)
930 		return;
931 	smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
932 			       conn->rmb_desc, DMA_FROM_DEVICE);
933 }
934 
935 void smc_rmb_sync_sg_for_device(struct smc_connection *conn)
936 {
937 	struct smc_link_group *lgr = conn->lgr;
938 
939 	if (!conn->lgr || conn->lgr->is_smcd)
940 		return;
941 	smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
942 				  conn->rmb_desc, DMA_FROM_DEVICE);
943 }
944 
945 /* create the send and receive buffer for an SMC socket;
946  * receive buffers are called RMBs;
947  * (even though the SMC protocol allows more than one RMB-element per RMB,
948  * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
949  * extra RMB for every connection in a link group
950  */
951 int smc_buf_create(struct smc_sock *smc, bool is_smcd)
952 {
953 	int rc;
954 
955 	/* create send buffer */
956 	rc = __smc_buf_create(smc, is_smcd, false);
957 	if (rc)
958 		return rc;
959 	/* create rmb */
960 	rc = __smc_buf_create(smc, is_smcd, true);
961 	if (rc)
962 		smc_buf_free(smc->conn.lgr, false, smc->conn.sndbuf_desc);
963 	return rc;
964 }
965 
966 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
967 {
968 	int i;
969 
970 	for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
971 		if (!test_and_set_bit(i, lgr->rtokens_used_mask))
972 			return i;
973 	}
974 	return -ENOSPC;
975 }
976 
977 /* add a new rtoken from peer */
978 int smc_rtoken_add(struct smc_link_group *lgr, __be64 nw_vaddr, __be32 nw_rkey)
979 {
980 	u64 dma_addr = be64_to_cpu(nw_vaddr);
981 	u32 rkey = ntohl(nw_rkey);
982 	int i;
983 
984 	for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
985 		if ((lgr->rtokens[i][SMC_SINGLE_LINK].rkey == rkey) &&
986 		    (lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr == dma_addr) &&
987 		    test_bit(i, lgr->rtokens_used_mask)) {
988 			/* already in list */
989 			return i;
990 		}
991 	}
992 	i = smc_rmb_reserve_rtoken_idx(lgr);
993 	if (i < 0)
994 		return i;
995 	lgr->rtokens[i][SMC_SINGLE_LINK].rkey = rkey;
996 	lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr = dma_addr;
997 	return i;
998 }
999 
1000 /* delete an rtoken */
1001 int smc_rtoken_delete(struct smc_link_group *lgr, __be32 nw_rkey)
1002 {
1003 	u32 rkey = ntohl(nw_rkey);
1004 	int i;
1005 
1006 	for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1007 		if (lgr->rtokens[i][SMC_SINGLE_LINK].rkey == rkey &&
1008 		    test_bit(i, lgr->rtokens_used_mask)) {
1009 			lgr->rtokens[i][SMC_SINGLE_LINK].rkey = 0;
1010 			lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr = 0;
1011 
1012 			clear_bit(i, lgr->rtokens_used_mask);
1013 			return 0;
1014 		}
1015 	}
1016 	return -ENOENT;
1017 }
1018 
1019 /* save rkey and dma_addr received from peer during clc handshake */
1020 int smc_rmb_rtoken_handling(struct smc_connection *conn,
1021 			    struct smc_clc_msg_accept_confirm *clc)
1022 {
1023 	conn->rtoken_idx = smc_rtoken_add(conn->lgr, clc->rmb_dma_addr,
1024 					  clc->rmb_rkey);
1025 	if (conn->rtoken_idx < 0)
1026 		return conn->rtoken_idx;
1027 	return 0;
1028 }
1029 
1030 /* Called (from smc_exit) when module is removed */
1031 void smc_core_exit(void)
1032 {
1033 	struct smc_link_group *lgr, *lg;
1034 	LIST_HEAD(lgr_freeing_list);
1035 
1036 	spin_lock_bh(&smc_lgr_list.lock);
1037 	if (!list_empty(&smc_lgr_list.list))
1038 		list_splice_init(&smc_lgr_list.list, &lgr_freeing_list);
1039 	spin_unlock_bh(&smc_lgr_list.lock);
1040 	list_for_each_entry_safe(lgr, lg, &lgr_freeing_list, list) {
1041 		list_del_init(&lgr->list);
1042 		if (!lgr->is_smcd) {
1043 			struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
1044 
1045 			if (lnk->state == SMC_LNK_ACTIVE)
1046 				smc_llc_send_delete_link(lnk, SMC_LLC_REQ,
1047 							 false);
1048 			smc_llc_link_inactive(lnk);
1049 		}
1050 		cancel_delayed_work_sync(&lgr->free_work);
1051 		if (lgr->is_smcd)
1052 			smc_ism_signal_shutdown(lgr);
1053 		smc_lgr_free(lgr); /* free link group */
1054 	}
1055 }
1056