xref: /openbmc/linux/net/smc/smc_core.c (revision b8a94bfb)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  *  Shared Memory Communications over RDMA (SMC-R) and RoCE
4  *
5  *  Basic Transport Functions exploiting Infiniband API
6  *
7  *  Copyright IBM Corp. 2016
8  *
9  *  Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
10  */
11 
12 #include <linux/socket.h>
13 #include <linux/if_vlan.h>
14 #include <linux/random.h>
15 #include <linux/workqueue.h>
16 #include <linux/wait.h>
17 #include <linux/reboot.h>
18 #include <linux/mutex.h>
19 #include <linux/list.h>
20 #include <linux/smc.h>
21 #include <net/tcp.h>
22 #include <net/sock.h>
23 #include <rdma/ib_verbs.h>
24 #include <rdma/ib_cache.h>
25 
26 #include "smc.h"
27 #include "smc_clc.h"
28 #include "smc_core.h"
29 #include "smc_ib.h"
30 #include "smc_wr.h"
31 #include "smc_llc.h"
32 #include "smc_cdc.h"
33 #include "smc_close.h"
34 #include "smc_ism.h"
35 #include "smc_netlink.h"
36 #include "smc_stats.h"
37 #include "smc_tracepoint.h"
38 
39 #define SMC_LGR_NUM_INCR		256
40 #define SMC_LGR_FREE_DELAY_SERV		(600 * HZ)
41 #define SMC_LGR_FREE_DELAY_CLNT		(SMC_LGR_FREE_DELAY_SERV + 10 * HZ)
42 
43 struct smc_lgr_list smc_lgr_list = {	/* established link groups */
44 	.lock = __SPIN_LOCK_UNLOCKED(smc_lgr_list.lock),
45 	.list = LIST_HEAD_INIT(smc_lgr_list.list),
46 	.num = 0,
47 };
48 
49 static atomic_t lgr_cnt = ATOMIC_INIT(0); /* number of existing link groups */
50 static DECLARE_WAIT_QUEUE_HEAD(lgrs_deleted);
51 
52 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
53 			 struct smc_buf_desc *buf_desc);
54 static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft);
55 
56 static void smc_link_down_work(struct work_struct *work);
57 
58 /* return head of link group list and its lock for a given link group */
59 static inline struct list_head *smc_lgr_list_head(struct smc_link_group *lgr,
60 						  spinlock_t **lgr_lock)
61 {
62 	if (lgr->is_smcd) {
63 		*lgr_lock = &lgr->smcd->lgr_lock;
64 		return &lgr->smcd->lgr_list;
65 	}
66 
67 	*lgr_lock = &smc_lgr_list.lock;
68 	return &smc_lgr_list.list;
69 }
70 
71 static void smc_ibdev_cnt_inc(struct smc_link *lnk)
72 {
73 	atomic_inc(&lnk->smcibdev->lnk_cnt_by_port[lnk->ibport - 1]);
74 }
75 
76 static void smc_ibdev_cnt_dec(struct smc_link *lnk)
77 {
78 	atomic_dec(&lnk->smcibdev->lnk_cnt_by_port[lnk->ibport - 1]);
79 }
80 
81 static void smc_lgr_schedule_free_work(struct smc_link_group *lgr)
82 {
83 	/* client link group creation always follows the server link group
84 	 * creation. For client use a somewhat higher removal delay time,
85 	 * otherwise there is a risk of out-of-sync link groups.
86 	 */
87 	if (!lgr->freeing) {
88 		mod_delayed_work(system_wq, &lgr->free_work,
89 				 (!lgr->is_smcd && lgr->role == SMC_CLNT) ?
90 						SMC_LGR_FREE_DELAY_CLNT :
91 						SMC_LGR_FREE_DELAY_SERV);
92 	}
93 }
94 
95 /* Register connection's alert token in our lookup structure.
96  * To use rbtrees we have to implement our own insert core.
97  * Requires @conns_lock
98  * @smc		connection to register
99  * Returns 0 on success, != otherwise.
100  */
101 static void smc_lgr_add_alert_token(struct smc_connection *conn)
102 {
103 	struct rb_node **link, *parent = NULL;
104 	u32 token = conn->alert_token_local;
105 
106 	link = &conn->lgr->conns_all.rb_node;
107 	while (*link) {
108 		struct smc_connection *cur = rb_entry(*link,
109 					struct smc_connection, alert_node);
110 
111 		parent = *link;
112 		if (cur->alert_token_local > token)
113 			link = &parent->rb_left;
114 		else
115 			link = &parent->rb_right;
116 	}
117 	/* Put the new node there */
118 	rb_link_node(&conn->alert_node, parent, link);
119 	rb_insert_color(&conn->alert_node, &conn->lgr->conns_all);
120 }
121 
122 /* assign an SMC-R link to the connection */
123 static int smcr_lgr_conn_assign_link(struct smc_connection *conn, bool first)
124 {
125 	enum smc_link_state expected = first ? SMC_LNK_ACTIVATING :
126 				       SMC_LNK_ACTIVE;
127 	int i, j;
128 
129 	/* do link balancing */
130 	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
131 		struct smc_link *lnk = &conn->lgr->lnk[i];
132 
133 		if (lnk->state != expected || lnk->link_is_asym)
134 			continue;
135 		if (conn->lgr->role == SMC_CLNT) {
136 			conn->lnk = lnk; /* temporary, SMC server assigns link*/
137 			break;
138 		}
139 		if (conn->lgr->conns_num % 2) {
140 			for (j = i + 1; j < SMC_LINKS_PER_LGR_MAX; j++) {
141 				struct smc_link *lnk2;
142 
143 				lnk2 = &conn->lgr->lnk[j];
144 				if (lnk2->state == expected &&
145 				    !lnk2->link_is_asym) {
146 					conn->lnk = lnk2;
147 					break;
148 				}
149 			}
150 		}
151 		if (!conn->lnk)
152 			conn->lnk = lnk;
153 		break;
154 	}
155 	if (!conn->lnk)
156 		return SMC_CLC_DECL_NOACTLINK;
157 	atomic_inc(&conn->lnk->conn_cnt);
158 	return 0;
159 }
160 
161 /* Register connection in link group by assigning an alert token
162  * registered in a search tree.
163  * Requires @conns_lock
164  * Note that '0' is a reserved value and not assigned.
165  */
166 static int smc_lgr_register_conn(struct smc_connection *conn, bool first)
167 {
168 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
169 	static atomic_t nexttoken = ATOMIC_INIT(0);
170 	int rc;
171 
172 	if (!conn->lgr->is_smcd) {
173 		rc = smcr_lgr_conn_assign_link(conn, first);
174 		if (rc) {
175 			conn->lgr = NULL;
176 			return rc;
177 		}
178 	}
179 	/* find a new alert_token_local value not yet used by some connection
180 	 * in this link group
181 	 */
182 	sock_hold(&smc->sk); /* sock_put in smc_lgr_unregister_conn() */
183 	while (!conn->alert_token_local) {
184 		conn->alert_token_local = atomic_inc_return(&nexttoken);
185 		if (smc_lgr_find_conn(conn->alert_token_local, conn->lgr))
186 			conn->alert_token_local = 0;
187 	}
188 	smc_lgr_add_alert_token(conn);
189 	conn->lgr->conns_num++;
190 	return 0;
191 }
192 
193 /* Unregister connection and reset the alert token of the given connection<
194  */
195 static void __smc_lgr_unregister_conn(struct smc_connection *conn)
196 {
197 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
198 	struct smc_link_group *lgr = conn->lgr;
199 
200 	rb_erase(&conn->alert_node, &lgr->conns_all);
201 	if (conn->lnk)
202 		atomic_dec(&conn->lnk->conn_cnt);
203 	lgr->conns_num--;
204 	conn->alert_token_local = 0;
205 	sock_put(&smc->sk); /* sock_hold in smc_lgr_register_conn() */
206 }
207 
208 /* Unregister connection from lgr
209  */
210 static void smc_lgr_unregister_conn(struct smc_connection *conn)
211 {
212 	struct smc_link_group *lgr = conn->lgr;
213 
214 	if (!smc_conn_lgr_valid(conn))
215 		return;
216 	write_lock_bh(&lgr->conns_lock);
217 	if (conn->alert_token_local) {
218 		__smc_lgr_unregister_conn(conn);
219 	}
220 	write_unlock_bh(&lgr->conns_lock);
221 }
222 
223 int smc_nl_get_sys_info(struct sk_buff *skb, struct netlink_callback *cb)
224 {
225 	struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
226 	char hostname[SMC_MAX_HOSTNAME_LEN + 1];
227 	char smc_seid[SMC_MAX_EID_LEN + 1];
228 	struct nlattr *attrs;
229 	u8 *seid = NULL;
230 	u8 *host = NULL;
231 	void *nlh;
232 
233 	nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
234 			  &smc_gen_nl_family, NLM_F_MULTI,
235 			  SMC_NETLINK_GET_SYS_INFO);
236 	if (!nlh)
237 		goto errmsg;
238 	if (cb_ctx->pos[0])
239 		goto errout;
240 	attrs = nla_nest_start(skb, SMC_GEN_SYS_INFO);
241 	if (!attrs)
242 		goto errout;
243 	if (nla_put_u8(skb, SMC_NLA_SYS_VER, SMC_V2))
244 		goto errattr;
245 	if (nla_put_u8(skb, SMC_NLA_SYS_REL, SMC_RELEASE))
246 		goto errattr;
247 	if (nla_put_u8(skb, SMC_NLA_SYS_IS_ISM_V2, smc_ism_is_v2_capable()))
248 		goto errattr;
249 	if (nla_put_u8(skb, SMC_NLA_SYS_IS_SMCR_V2, true))
250 		goto errattr;
251 	smc_clc_get_hostname(&host);
252 	if (host) {
253 		memcpy(hostname, host, SMC_MAX_HOSTNAME_LEN);
254 		hostname[SMC_MAX_HOSTNAME_LEN] = 0;
255 		if (nla_put_string(skb, SMC_NLA_SYS_LOCAL_HOST, hostname))
256 			goto errattr;
257 	}
258 	if (smc_ism_is_v2_capable()) {
259 		smc_ism_get_system_eid(&seid);
260 		memcpy(smc_seid, seid, SMC_MAX_EID_LEN);
261 		smc_seid[SMC_MAX_EID_LEN] = 0;
262 		if (nla_put_string(skb, SMC_NLA_SYS_SEID, smc_seid))
263 			goto errattr;
264 	}
265 	nla_nest_end(skb, attrs);
266 	genlmsg_end(skb, nlh);
267 	cb_ctx->pos[0] = 1;
268 	return skb->len;
269 
270 errattr:
271 	nla_nest_cancel(skb, attrs);
272 errout:
273 	genlmsg_cancel(skb, nlh);
274 errmsg:
275 	return skb->len;
276 }
277 
278 /* Fill SMC_NLA_LGR_D_V2_COMMON/SMC_NLA_LGR_R_V2_COMMON nested attributes */
279 static int smc_nl_fill_lgr_v2_common(struct smc_link_group *lgr,
280 				     struct sk_buff *skb,
281 				     struct netlink_callback *cb,
282 				     struct nlattr *v2_attrs)
283 {
284 	char smc_host[SMC_MAX_HOSTNAME_LEN + 1];
285 	char smc_eid[SMC_MAX_EID_LEN + 1];
286 
287 	if (nla_put_u8(skb, SMC_NLA_LGR_V2_VER, lgr->smc_version))
288 		goto errv2attr;
289 	if (nla_put_u8(skb, SMC_NLA_LGR_V2_REL, lgr->peer_smc_release))
290 		goto errv2attr;
291 	if (nla_put_u8(skb, SMC_NLA_LGR_V2_OS, lgr->peer_os))
292 		goto errv2attr;
293 	memcpy(smc_host, lgr->peer_hostname, SMC_MAX_HOSTNAME_LEN);
294 	smc_host[SMC_MAX_HOSTNAME_LEN] = 0;
295 	if (nla_put_string(skb, SMC_NLA_LGR_V2_PEER_HOST, smc_host))
296 		goto errv2attr;
297 	memcpy(smc_eid, lgr->negotiated_eid, SMC_MAX_EID_LEN);
298 	smc_eid[SMC_MAX_EID_LEN] = 0;
299 	if (nla_put_string(skb, SMC_NLA_LGR_V2_NEG_EID, smc_eid))
300 		goto errv2attr;
301 
302 	nla_nest_end(skb, v2_attrs);
303 	return 0;
304 
305 errv2attr:
306 	nla_nest_cancel(skb, v2_attrs);
307 	return -EMSGSIZE;
308 }
309 
310 static int smc_nl_fill_smcr_lgr_v2(struct smc_link_group *lgr,
311 				   struct sk_buff *skb,
312 				   struct netlink_callback *cb)
313 {
314 	struct nlattr *v2_attrs;
315 
316 	v2_attrs = nla_nest_start(skb, SMC_NLA_LGR_R_V2);
317 	if (!v2_attrs)
318 		goto errattr;
319 	if (nla_put_u8(skb, SMC_NLA_LGR_R_V2_DIRECT, !lgr->uses_gateway))
320 		goto errv2attr;
321 
322 	nla_nest_end(skb, v2_attrs);
323 	return 0;
324 
325 errv2attr:
326 	nla_nest_cancel(skb, v2_attrs);
327 errattr:
328 	return -EMSGSIZE;
329 }
330 
331 static int smc_nl_fill_lgr(struct smc_link_group *lgr,
332 			   struct sk_buff *skb,
333 			   struct netlink_callback *cb)
334 {
335 	char smc_target[SMC_MAX_PNETID_LEN + 1];
336 	struct nlattr *attrs, *v2_attrs;
337 
338 	attrs = nla_nest_start(skb, SMC_GEN_LGR_SMCR);
339 	if (!attrs)
340 		goto errout;
341 
342 	if (nla_put_u32(skb, SMC_NLA_LGR_R_ID, *((u32 *)&lgr->id)))
343 		goto errattr;
344 	if (nla_put_u32(skb, SMC_NLA_LGR_R_CONNS_NUM, lgr->conns_num))
345 		goto errattr;
346 	if (nla_put_u8(skb, SMC_NLA_LGR_R_ROLE, lgr->role))
347 		goto errattr;
348 	if (nla_put_u8(skb, SMC_NLA_LGR_R_TYPE, lgr->type))
349 		goto errattr;
350 	if (nla_put_u8(skb, SMC_NLA_LGR_R_BUF_TYPE, lgr->buf_type))
351 		goto errattr;
352 	if (nla_put_u8(skb, SMC_NLA_LGR_R_VLAN_ID, lgr->vlan_id))
353 		goto errattr;
354 	if (nla_put_u64_64bit(skb, SMC_NLA_LGR_R_NET_COOKIE,
355 			      lgr->net->net_cookie, SMC_NLA_LGR_R_PAD))
356 		goto errattr;
357 	memcpy(smc_target, lgr->pnet_id, SMC_MAX_PNETID_LEN);
358 	smc_target[SMC_MAX_PNETID_LEN] = 0;
359 	if (nla_put_string(skb, SMC_NLA_LGR_R_PNETID, smc_target))
360 		goto errattr;
361 	if (lgr->smc_version > SMC_V1) {
362 		v2_attrs = nla_nest_start(skb, SMC_NLA_LGR_R_V2_COMMON);
363 		if (!v2_attrs)
364 			goto errattr;
365 		if (smc_nl_fill_lgr_v2_common(lgr, skb, cb, v2_attrs))
366 			goto errattr;
367 		if (smc_nl_fill_smcr_lgr_v2(lgr, skb, cb))
368 			goto errattr;
369 	}
370 
371 	nla_nest_end(skb, attrs);
372 	return 0;
373 errattr:
374 	nla_nest_cancel(skb, attrs);
375 errout:
376 	return -EMSGSIZE;
377 }
378 
379 static int smc_nl_fill_lgr_link(struct smc_link_group *lgr,
380 				struct smc_link *link,
381 				struct sk_buff *skb,
382 				struct netlink_callback *cb)
383 {
384 	char smc_ibname[IB_DEVICE_NAME_MAX];
385 	u8 smc_gid_target[41];
386 	struct nlattr *attrs;
387 	u32 link_uid = 0;
388 	void *nlh;
389 
390 	nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
391 			  &smc_gen_nl_family, NLM_F_MULTI,
392 			  SMC_NETLINK_GET_LINK_SMCR);
393 	if (!nlh)
394 		goto errmsg;
395 
396 	attrs = nla_nest_start(skb, SMC_GEN_LINK_SMCR);
397 	if (!attrs)
398 		goto errout;
399 
400 	if (nla_put_u8(skb, SMC_NLA_LINK_ID, link->link_id))
401 		goto errattr;
402 	if (nla_put_u32(skb, SMC_NLA_LINK_STATE, link->state))
403 		goto errattr;
404 	if (nla_put_u32(skb, SMC_NLA_LINK_CONN_CNT,
405 			atomic_read(&link->conn_cnt)))
406 		goto errattr;
407 	if (nla_put_u8(skb, SMC_NLA_LINK_IB_PORT, link->ibport))
408 		goto errattr;
409 	if (nla_put_u32(skb, SMC_NLA_LINK_NET_DEV, link->ndev_ifidx))
410 		goto errattr;
411 	snprintf(smc_ibname, sizeof(smc_ibname), "%s", link->ibname);
412 	if (nla_put_string(skb, SMC_NLA_LINK_IB_DEV, smc_ibname))
413 		goto errattr;
414 	memcpy(&link_uid, link->link_uid, sizeof(link_uid));
415 	if (nla_put_u32(skb, SMC_NLA_LINK_UID, link_uid))
416 		goto errattr;
417 	memcpy(&link_uid, link->peer_link_uid, sizeof(link_uid));
418 	if (nla_put_u32(skb, SMC_NLA_LINK_PEER_UID, link_uid))
419 		goto errattr;
420 	memset(smc_gid_target, 0, sizeof(smc_gid_target));
421 	smc_gid_be16_convert(smc_gid_target, link->gid);
422 	if (nla_put_string(skb, SMC_NLA_LINK_GID, smc_gid_target))
423 		goto errattr;
424 	memset(smc_gid_target, 0, sizeof(smc_gid_target));
425 	smc_gid_be16_convert(smc_gid_target, link->peer_gid);
426 	if (nla_put_string(skb, SMC_NLA_LINK_PEER_GID, smc_gid_target))
427 		goto errattr;
428 
429 	nla_nest_end(skb, attrs);
430 	genlmsg_end(skb, nlh);
431 	return 0;
432 errattr:
433 	nla_nest_cancel(skb, attrs);
434 errout:
435 	genlmsg_cancel(skb, nlh);
436 errmsg:
437 	return -EMSGSIZE;
438 }
439 
440 static int smc_nl_handle_lgr(struct smc_link_group *lgr,
441 			     struct sk_buff *skb,
442 			     struct netlink_callback *cb,
443 			     bool list_links)
444 {
445 	void *nlh;
446 	int i;
447 
448 	nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
449 			  &smc_gen_nl_family, NLM_F_MULTI,
450 			  SMC_NETLINK_GET_LGR_SMCR);
451 	if (!nlh)
452 		goto errmsg;
453 	if (smc_nl_fill_lgr(lgr, skb, cb))
454 		goto errout;
455 
456 	genlmsg_end(skb, nlh);
457 	if (!list_links)
458 		goto out;
459 	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
460 		if (!smc_link_usable(&lgr->lnk[i]))
461 			continue;
462 		if (smc_nl_fill_lgr_link(lgr, &lgr->lnk[i], skb, cb))
463 			goto errout;
464 	}
465 out:
466 	return 0;
467 
468 errout:
469 	genlmsg_cancel(skb, nlh);
470 errmsg:
471 	return -EMSGSIZE;
472 }
473 
474 static void smc_nl_fill_lgr_list(struct smc_lgr_list *smc_lgr,
475 				 struct sk_buff *skb,
476 				 struct netlink_callback *cb,
477 				 bool list_links)
478 {
479 	struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
480 	struct smc_link_group *lgr;
481 	int snum = cb_ctx->pos[0];
482 	int num = 0;
483 
484 	spin_lock_bh(&smc_lgr->lock);
485 	list_for_each_entry(lgr, &smc_lgr->list, list) {
486 		if (num < snum)
487 			goto next;
488 		if (smc_nl_handle_lgr(lgr, skb, cb, list_links))
489 			goto errout;
490 next:
491 		num++;
492 	}
493 errout:
494 	spin_unlock_bh(&smc_lgr->lock);
495 	cb_ctx->pos[0] = num;
496 }
497 
498 static int smc_nl_fill_smcd_lgr(struct smc_link_group *lgr,
499 				struct sk_buff *skb,
500 				struct netlink_callback *cb)
501 {
502 	char smc_pnet[SMC_MAX_PNETID_LEN + 1];
503 	struct nlattr *attrs;
504 	void *nlh;
505 
506 	nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
507 			  &smc_gen_nl_family, NLM_F_MULTI,
508 			  SMC_NETLINK_GET_LGR_SMCD);
509 	if (!nlh)
510 		goto errmsg;
511 
512 	attrs = nla_nest_start(skb, SMC_GEN_LGR_SMCD);
513 	if (!attrs)
514 		goto errout;
515 
516 	if (nla_put_u32(skb, SMC_NLA_LGR_D_ID, *((u32 *)&lgr->id)))
517 		goto errattr;
518 	if (nla_put_u64_64bit(skb, SMC_NLA_LGR_D_GID, lgr->smcd->local_gid,
519 			      SMC_NLA_LGR_D_PAD))
520 		goto errattr;
521 	if (nla_put_u64_64bit(skb, SMC_NLA_LGR_D_PEER_GID, lgr->peer_gid,
522 			      SMC_NLA_LGR_D_PAD))
523 		goto errattr;
524 	if (nla_put_u8(skb, SMC_NLA_LGR_D_VLAN_ID, lgr->vlan_id))
525 		goto errattr;
526 	if (nla_put_u32(skb, SMC_NLA_LGR_D_CONNS_NUM, lgr->conns_num))
527 		goto errattr;
528 	if (nla_put_u32(skb, SMC_NLA_LGR_D_CHID, smc_ism_get_chid(lgr->smcd)))
529 		goto errattr;
530 	memcpy(smc_pnet, lgr->smcd->pnetid, SMC_MAX_PNETID_LEN);
531 	smc_pnet[SMC_MAX_PNETID_LEN] = 0;
532 	if (nla_put_string(skb, SMC_NLA_LGR_D_PNETID, smc_pnet))
533 		goto errattr;
534 	if (lgr->smc_version > SMC_V1) {
535 		struct nlattr *v2_attrs;
536 
537 		v2_attrs = nla_nest_start(skb, SMC_NLA_LGR_D_V2_COMMON);
538 		if (!v2_attrs)
539 			goto errattr;
540 		if (smc_nl_fill_lgr_v2_common(lgr, skb, cb, v2_attrs))
541 			goto errattr;
542 	}
543 	nla_nest_end(skb, attrs);
544 	genlmsg_end(skb, nlh);
545 	return 0;
546 
547 errattr:
548 	nla_nest_cancel(skb, attrs);
549 errout:
550 	genlmsg_cancel(skb, nlh);
551 errmsg:
552 	return -EMSGSIZE;
553 }
554 
555 static int smc_nl_handle_smcd_lgr(struct smcd_dev *dev,
556 				  struct sk_buff *skb,
557 				  struct netlink_callback *cb)
558 {
559 	struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
560 	struct smc_link_group *lgr;
561 	int snum = cb_ctx->pos[1];
562 	int rc = 0, num = 0;
563 
564 	spin_lock_bh(&dev->lgr_lock);
565 	list_for_each_entry(lgr, &dev->lgr_list, list) {
566 		if (!lgr->is_smcd)
567 			continue;
568 		if (num < snum)
569 			goto next;
570 		rc = smc_nl_fill_smcd_lgr(lgr, skb, cb);
571 		if (rc)
572 			goto errout;
573 next:
574 		num++;
575 	}
576 errout:
577 	spin_unlock_bh(&dev->lgr_lock);
578 	cb_ctx->pos[1] = num;
579 	return rc;
580 }
581 
582 static int smc_nl_fill_smcd_dev(struct smcd_dev_list *dev_list,
583 				struct sk_buff *skb,
584 				struct netlink_callback *cb)
585 {
586 	struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
587 	struct smcd_dev *smcd_dev;
588 	int snum = cb_ctx->pos[0];
589 	int rc = 0, num = 0;
590 
591 	mutex_lock(&dev_list->mutex);
592 	list_for_each_entry(smcd_dev, &dev_list->list, list) {
593 		if (list_empty(&smcd_dev->lgr_list))
594 			continue;
595 		if (num < snum)
596 			goto next;
597 		rc = smc_nl_handle_smcd_lgr(smcd_dev, skb, cb);
598 		if (rc)
599 			goto errout;
600 next:
601 		num++;
602 	}
603 errout:
604 	mutex_unlock(&dev_list->mutex);
605 	cb_ctx->pos[0] = num;
606 	return rc;
607 }
608 
609 int smcr_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb)
610 {
611 	bool list_links = false;
612 
613 	smc_nl_fill_lgr_list(&smc_lgr_list, skb, cb, list_links);
614 	return skb->len;
615 }
616 
617 int smcr_nl_get_link(struct sk_buff *skb, struct netlink_callback *cb)
618 {
619 	bool list_links = true;
620 
621 	smc_nl_fill_lgr_list(&smc_lgr_list, skb, cb, list_links);
622 	return skb->len;
623 }
624 
625 int smcd_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb)
626 {
627 	smc_nl_fill_smcd_dev(&smcd_dev_list, skb, cb);
628 	return skb->len;
629 }
630 
631 void smc_lgr_cleanup_early(struct smc_link_group *lgr)
632 {
633 	spinlock_t *lgr_lock;
634 
635 	if (!lgr)
636 		return;
637 
638 	smc_lgr_list_head(lgr, &lgr_lock);
639 	spin_lock_bh(lgr_lock);
640 	/* do not use this link group for new connections */
641 	if (!list_empty(&lgr->list))
642 		list_del_init(&lgr->list);
643 	spin_unlock_bh(lgr_lock);
644 	__smc_lgr_terminate(lgr, true);
645 }
646 
647 static void smcr_lgr_link_deactivate_all(struct smc_link_group *lgr)
648 {
649 	int i;
650 
651 	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
652 		struct smc_link *lnk = &lgr->lnk[i];
653 
654 		if (smc_link_sendable(lnk))
655 			lnk->state = SMC_LNK_INACTIVE;
656 	}
657 	wake_up_all(&lgr->llc_msg_waiter);
658 	wake_up_all(&lgr->llc_flow_waiter);
659 }
660 
661 static void smc_lgr_free(struct smc_link_group *lgr);
662 
663 static void smc_lgr_free_work(struct work_struct *work)
664 {
665 	struct smc_link_group *lgr = container_of(to_delayed_work(work),
666 						  struct smc_link_group,
667 						  free_work);
668 	spinlock_t *lgr_lock;
669 	bool conns;
670 
671 	smc_lgr_list_head(lgr, &lgr_lock);
672 	spin_lock_bh(lgr_lock);
673 	if (lgr->freeing) {
674 		spin_unlock_bh(lgr_lock);
675 		return;
676 	}
677 	read_lock_bh(&lgr->conns_lock);
678 	conns = RB_EMPTY_ROOT(&lgr->conns_all);
679 	read_unlock_bh(&lgr->conns_lock);
680 	if (!conns) { /* number of lgr connections is no longer zero */
681 		spin_unlock_bh(lgr_lock);
682 		return;
683 	}
684 	list_del_init(&lgr->list); /* remove from smc_lgr_list */
685 	lgr->freeing = 1; /* this instance does the freeing, no new schedule */
686 	spin_unlock_bh(lgr_lock);
687 	cancel_delayed_work(&lgr->free_work);
688 
689 	if (!lgr->is_smcd && !lgr->terminating)
690 		smc_llc_send_link_delete_all(lgr, true,
691 					     SMC_LLC_DEL_PROG_INIT_TERM);
692 	if (lgr->is_smcd && !lgr->terminating)
693 		smc_ism_signal_shutdown(lgr);
694 	if (!lgr->is_smcd)
695 		smcr_lgr_link_deactivate_all(lgr);
696 	smc_lgr_free(lgr);
697 }
698 
699 static void smc_lgr_terminate_work(struct work_struct *work)
700 {
701 	struct smc_link_group *lgr = container_of(work, struct smc_link_group,
702 						  terminate_work);
703 
704 	__smc_lgr_terminate(lgr, true);
705 }
706 
707 /* return next unique link id for the lgr */
708 static u8 smcr_next_link_id(struct smc_link_group *lgr)
709 {
710 	u8 link_id;
711 	int i;
712 
713 	while (1) {
714 again:
715 		link_id = ++lgr->next_link_id;
716 		if (!link_id)	/* skip zero as link_id */
717 			link_id = ++lgr->next_link_id;
718 		for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
719 			if (smc_link_usable(&lgr->lnk[i]) &&
720 			    lgr->lnk[i].link_id == link_id)
721 				goto again;
722 		}
723 		break;
724 	}
725 	return link_id;
726 }
727 
728 static void smcr_copy_dev_info_to_link(struct smc_link *link)
729 {
730 	struct smc_ib_device *smcibdev = link->smcibdev;
731 
732 	snprintf(link->ibname, sizeof(link->ibname), "%s",
733 		 smcibdev->ibdev->name);
734 	link->ndev_ifidx = smcibdev->ndev_ifidx[link->ibport - 1];
735 }
736 
737 int smcr_link_init(struct smc_link_group *lgr, struct smc_link *lnk,
738 		   u8 link_idx, struct smc_init_info *ini)
739 {
740 	struct smc_ib_device *smcibdev;
741 	u8 rndvec[3];
742 	int rc;
743 
744 	if (lgr->smc_version == SMC_V2) {
745 		lnk->smcibdev = ini->smcrv2.ib_dev_v2;
746 		lnk->ibport = ini->smcrv2.ib_port_v2;
747 	} else {
748 		lnk->smcibdev = ini->ib_dev;
749 		lnk->ibport = ini->ib_port;
750 	}
751 	get_device(&lnk->smcibdev->ibdev->dev);
752 	atomic_inc(&lnk->smcibdev->lnk_cnt);
753 	refcount_set(&lnk->refcnt, 1); /* link refcnt is set to 1 */
754 	lnk->clearing = 0;
755 	lnk->path_mtu = lnk->smcibdev->pattr[lnk->ibport - 1].active_mtu;
756 	lnk->link_id = smcr_next_link_id(lgr);
757 	lnk->lgr = lgr;
758 	smc_lgr_hold(lgr); /* lgr_put in smcr_link_clear() */
759 	lnk->link_idx = link_idx;
760 	lnk->wr_rx_id_compl = 0;
761 	smc_ibdev_cnt_inc(lnk);
762 	smcr_copy_dev_info_to_link(lnk);
763 	atomic_set(&lnk->conn_cnt, 0);
764 	smc_llc_link_set_uid(lnk);
765 	INIT_WORK(&lnk->link_down_wrk, smc_link_down_work);
766 	if (!lnk->smcibdev->initialized) {
767 		rc = (int)smc_ib_setup_per_ibdev(lnk->smcibdev);
768 		if (rc)
769 			goto out;
770 	}
771 	get_random_bytes(rndvec, sizeof(rndvec));
772 	lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) +
773 		(rndvec[2] << 16);
774 	rc = smc_ib_determine_gid(lnk->smcibdev, lnk->ibport,
775 				  ini->vlan_id, lnk->gid, &lnk->sgid_index,
776 				  lgr->smc_version == SMC_V2 ?
777 						  &ini->smcrv2 : NULL);
778 	if (rc)
779 		goto out;
780 	rc = smc_llc_link_init(lnk);
781 	if (rc)
782 		goto out;
783 	rc = smc_wr_alloc_link_mem(lnk);
784 	if (rc)
785 		goto clear_llc_lnk;
786 	rc = smc_ib_create_protection_domain(lnk);
787 	if (rc)
788 		goto free_link_mem;
789 	rc = smc_ib_create_queue_pair(lnk);
790 	if (rc)
791 		goto dealloc_pd;
792 	rc = smc_wr_create_link(lnk);
793 	if (rc)
794 		goto destroy_qp;
795 	lnk->state = SMC_LNK_ACTIVATING;
796 	return 0;
797 
798 destroy_qp:
799 	smc_ib_destroy_queue_pair(lnk);
800 dealloc_pd:
801 	smc_ib_dealloc_protection_domain(lnk);
802 free_link_mem:
803 	smc_wr_free_link_mem(lnk);
804 clear_llc_lnk:
805 	smc_llc_link_clear(lnk, false);
806 out:
807 	smc_ibdev_cnt_dec(lnk);
808 	put_device(&lnk->smcibdev->ibdev->dev);
809 	smcibdev = lnk->smcibdev;
810 	memset(lnk, 0, sizeof(struct smc_link));
811 	lnk->state = SMC_LNK_UNUSED;
812 	if (!atomic_dec_return(&smcibdev->lnk_cnt))
813 		wake_up(&smcibdev->lnks_deleted);
814 	smc_lgr_put(lgr); /* lgr_hold above */
815 	return rc;
816 }
817 
818 /* create a new SMC link group */
819 static int smc_lgr_create(struct smc_sock *smc, struct smc_init_info *ini)
820 {
821 	struct smc_link_group *lgr;
822 	struct list_head *lgr_list;
823 	struct smc_link *lnk;
824 	spinlock_t *lgr_lock;
825 	u8 link_idx;
826 	int rc = 0;
827 	int i;
828 
829 	if (ini->is_smcd && ini->vlan_id) {
830 		if (smc_ism_get_vlan(ini->ism_dev[ini->ism_selected],
831 				     ini->vlan_id)) {
832 			rc = SMC_CLC_DECL_ISMVLANERR;
833 			goto out;
834 		}
835 	}
836 
837 	lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
838 	if (!lgr) {
839 		rc = SMC_CLC_DECL_MEM;
840 		goto ism_put_vlan;
841 	}
842 	lgr->tx_wq = alloc_workqueue("smc_tx_wq-%*phN", 0, 0,
843 				     SMC_LGR_ID_SIZE, &lgr->id);
844 	if (!lgr->tx_wq) {
845 		rc = -ENOMEM;
846 		goto free_lgr;
847 	}
848 	lgr->is_smcd = ini->is_smcd;
849 	lgr->sync_err = 0;
850 	lgr->terminating = 0;
851 	lgr->freeing = 0;
852 	lgr->vlan_id = ini->vlan_id;
853 	refcount_set(&lgr->refcnt, 1); /* set lgr refcnt to 1 */
854 	mutex_init(&lgr->sndbufs_lock);
855 	mutex_init(&lgr->rmbs_lock);
856 	rwlock_init(&lgr->conns_lock);
857 	for (i = 0; i < SMC_RMBE_SIZES; i++) {
858 		INIT_LIST_HEAD(&lgr->sndbufs[i]);
859 		INIT_LIST_HEAD(&lgr->rmbs[i]);
860 	}
861 	lgr->next_link_id = 0;
862 	smc_lgr_list.num += SMC_LGR_NUM_INCR;
863 	memcpy(&lgr->id, (u8 *)&smc_lgr_list.num, SMC_LGR_ID_SIZE);
864 	INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
865 	INIT_WORK(&lgr->terminate_work, smc_lgr_terminate_work);
866 	lgr->conns_all = RB_ROOT;
867 	if (ini->is_smcd) {
868 		/* SMC-D specific settings */
869 		get_device(&ini->ism_dev[ini->ism_selected]->dev);
870 		lgr->peer_gid = ini->ism_peer_gid[ini->ism_selected];
871 		lgr->smcd = ini->ism_dev[ini->ism_selected];
872 		lgr_list = &ini->ism_dev[ini->ism_selected]->lgr_list;
873 		lgr_lock = &lgr->smcd->lgr_lock;
874 		lgr->smc_version = ini->smcd_version;
875 		lgr->peer_shutdown = 0;
876 		atomic_inc(&ini->ism_dev[ini->ism_selected]->lgr_cnt);
877 	} else {
878 		/* SMC-R specific settings */
879 		struct smc_ib_device *ibdev;
880 		int ibport;
881 
882 		lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
883 		lgr->smc_version = ini->smcr_version;
884 		memcpy(lgr->peer_systemid, ini->peer_systemid,
885 		       SMC_SYSTEMID_LEN);
886 		if (lgr->smc_version == SMC_V2) {
887 			ibdev = ini->smcrv2.ib_dev_v2;
888 			ibport = ini->smcrv2.ib_port_v2;
889 			lgr->saddr = ini->smcrv2.saddr;
890 			lgr->uses_gateway = ini->smcrv2.uses_gateway;
891 			memcpy(lgr->nexthop_mac, ini->smcrv2.nexthop_mac,
892 			       ETH_ALEN);
893 		} else {
894 			ibdev = ini->ib_dev;
895 			ibport = ini->ib_port;
896 		}
897 		memcpy(lgr->pnet_id, ibdev->pnetid[ibport - 1],
898 		       SMC_MAX_PNETID_LEN);
899 		if (smc_wr_alloc_lgr_mem(lgr))
900 			goto free_wq;
901 		smc_llc_lgr_init(lgr, smc);
902 
903 		link_idx = SMC_SINGLE_LINK;
904 		lnk = &lgr->lnk[link_idx];
905 		rc = smcr_link_init(lgr, lnk, link_idx, ini);
906 		if (rc) {
907 			smc_wr_free_lgr_mem(lgr);
908 			goto free_wq;
909 		}
910 		lgr->net = smc_ib_net(lnk->smcibdev);
911 		lgr_list = &smc_lgr_list.list;
912 		lgr_lock = &smc_lgr_list.lock;
913 		lgr->buf_type = lgr->net->smc.sysctl_smcr_buf_type;
914 		atomic_inc(&lgr_cnt);
915 	}
916 	smc->conn.lgr = lgr;
917 	spin_lock_bh(lgr_lock);
918 	list_add_tail(&lgr->list, lgr_list);
919 	spin_unlock_bh(lgr_lock);
920 	return 0;
921 
922 free_wq:
923 	destroy_workqueue(lgr->tx_wq);
924 free_lgr:
925 	kfree(lgr);
926 ism_put_vlan:
927 	if (ini->is_smcd && ini->vlan_id)
928 		smc_ism_put_vlan(ini->ism_dev[ini->ism_selected], ini->vlan_id);
929 out:
930 	if (rc < 0) {
931 		if (rc == -ENOMEM)
932 			rc = SMC_CLC_DECL_MEM;
933 		else
934 			rc = SMC_CLC_DECL_INTERR;
935 	}
936 	return rc;
937 }
938 
939 static int smc_write_space(struct smc_connection *conn)
940 {
941 	int buffer_len = conn->peer_rmbe_size;
942 	union smc_host_cursor prod;
943 	union smc_host_cursor cons;
944 	int space;
945 
946 	smc_curs_copy(&prod, &conn->local_tx_ctrl.prod, conn);
947 	smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
948 	/* determine rx_buf space */
949 	space = buffer_len - smc_curs_diff(buffer_len, &cons, &prod);
950 	return space;
951 }
952 
953 static int smc_switch_cursor(struct smc_sock *smc, struct smc_cdc_tx_pend *pend,
954 			     struct smc_wr_buf *wr_buf)
955 {
956 	struct smc_connection *conn = &smc->conn;
957 	union smc_host_cursor cons, fin;
958 	int rc = 0;
959 	int diff;
960 
961 	smc_curs_copy(&conn->tx_curs_sent, &conn->tx_curs_fin, conn);
962 	smc_curs_copy(&fin, &conn->local_tx_ctrl_fin, conn);
963 	/* set prod cursor to old state, enforce tx_rdma_writes() */
964 	smc_curs_copy(&conn->local_tx_ctrl.prod, &fin, conn);
965 	smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
966 
967 	if (smc_curs_comp(conn->peer_rmbe_size, &cons, &fin) < 0) {
968 		/* cons cursor advanced more than fin, and prod was set
969 		 * fin above, so now prod is smaller than cons. Fix that.
970 		 */
971 		diff = smc_curs_diff(conn->peer_rmbe_size, &fin, &cons);
972 		smc_curs_add(conn->sndbuf_desc->len,
973 			     &conn->tx_curs_sent, diff);
974 		smc_curs_add(conn->sndbuf_desc->len,
975 			     &conn->tx_curs_fin, diff);
976 
977 		smp_mb__before_atomic();
978 		atomic_add(diff, &conn->sndbuf_space);
979 		smp_mb__after_atomic();
980 
981 		smc_curs_add(conn->peer_rmbe_size,
982 			     &conn->local_tx_ctrl.prod, diff);
983 		smc_curs_add(conn->peer_rmbe_size,
984 			     &conn->local_tx_ctrl_fin, diff);
985 	}
986 	/* recalculate, value is used by tx_rdma_writes() */
987 	atomic_set(&smc->conn.peer_rmbe_space, smc_write_space(conn));
988 
989 	if (smc->sk.sk_state != SMC_INIT &&
990 	    smc->sk.sk_state != SMC_CLOSED) {
991 		rc = smcr_cdc_msg_send_validation(conn, pend, wr_buf);
992 		if (!rc) {
993 			queue_delayed_work(conn->lgr->tx_wq, &conn->tx_work, 0);
994 			smc->sk.sk_data_ready(&smc->sk);
995 		}
996 	} else {
997 		smc_wr_tx_put_slot(conn->lnk,
998 				   (struct smc_wr_tx_pend_priv *)pend);
999 	}
1000 	return rc;
1001 }
1002 
1003 void smc_switch_link_and_count(struct smc_connection *conn,
1004 			       struct smc_link *to_lnk)
1005 {
1006 	atomic_dec(&conn->lnk->conn_cnt);
1007 	/* link_hold in smc_conn_create() */
1008 	smcr_link_put(conn->lnk);
1009 	conn->lnk = to_lnk;
1010 	atomic_inc(&conn->lnk->conn_cnt);
1011 	/* link_put in smc_conn_free() */
1012 	smcr_link_hold(conn->lnk);
1013 }
1014 
1015 struct smc_link *smc_switch_conns(struct smc_link_group *lgr,
1016 				  struct smc_link *from_lnk, bool is_dev_err)
1017 {
1018 	struct smc_link *to_lnk = NULL;
1019 	struct smc_cdc_tx_pend *pend;
1020 	struct smc_connection *conn;
1021 	struct smc_wr_buf *wr_buf;
1022 	struct smc_sock *smc;
1023 	struct rb_node *node;
1024 	int i, rc = 0;
1025 
1026 	/* link is inactive, wake up tx waiters */
1027 	smc_wr_wakeup_tx_wait(from_lnk);
1028 
1029 	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1030 		if (!smc_link_active(&lgr->lnk[i]) || i == from_lnk->link_idx)
1031 			continue;
1032 		if (is_dev_err && from_lnk->smcibdev == lgr->lnk[i].smcibdev &&
1033 		    from_lnk->ibport == lgr->lnk[i].ibport) {
1034 			continue;
1035 		}
1036 		to_lnk = &lgr->lnk[i];
1037 		break;
1038 	}
1039 	if (!to_lnk || !smc_wr_tx_link_hold(to_lnk)) {
1040 		smc_lgr_terminate_sched(lgr);
1041 		return NULL;
1042 	}
1043 again:
1044 	read_lock_bh(&lgr->conns_lock);
1045 	for (node = rb_first(&lgr->conns_all); node; node = rb_next(node)) {
1046 		conn = rb_entry(node, struct smc_connection, alert_node);
1047 		if (conn->lnk != from_lnk)
1048 			continue;
1049 		smc = container_of(conn, struct smc_sock, conn);
1050 		/* conn->lnk not yet set in SMC_INIT state */
1051 		if (smc->sk.sk_state == SMC_INIT)
1052 			continue;
1053 		if (smc->sk.sk_state == SMC_CLOSED ||
1054 		    smc->sk.sk_state == SMC_PEERCLOSEWAIT1 ||
1055 		    smc->sk.sk_state == SMC_PEERCLOSEWAIT2 ||
1056 		    smc->sk.sk_state == SMC_APPFINCLOSEWAIT ||
1057 		    smc->sk.sk_state == SMC_APPCLOSEWAIT1 ||
1058 		    smc->sk.sk_state == SMC_APPCLOSEWAIT2 ||
1059 		    smc->sk.sk_state == SMC_PEERFINCLOSEWAIT ||
1060 		    smc->sk.sk_state == SMC_PEERABORTWAIT ||
1061 		    smc->sk.sk_state == SMC_PROCESSABORT) {
1062 			spin_lock_bh(&conn->send_lock);
1063 			smc_switch_link_and_count(conn, to_lnk);
1064 			spin_unlock_bh(&conn->send_lock);
1065 			continue;
1066 		}
1067 		sock_hold(&smc->sk);
1068 		read_unlock_bh(&lgr->conns_lock);
1069 		/* pre-fetch buffer outside of send_lock, might sleep */
1070 		rc = smc_cdc_get_free_slot(conn, to_lnk, &wr_buf, NULL, &pend);
1071 		if (rc)
1072 			goto err_out;
1073 		/* avoid race with smcr_tx_sndbuf_nonempty() */
1074 		spin_lock_bh(&conn->send_lock);
1075 		smc_switch_link_and_count(conn, to_lnk);
1076 		rc = smc_switch_cursor(smc, pend, wr_buf);
1077 		spin_unlock_bh(&conn->send_lock);
1078 		sock_put(&smc->sk);
1079 		if (rc)
1080 			goto err_out;
1081 		goto again;
1082 	}
1083 	read_unlock_bh(&lgr->conns_lock);
1084 	smc_wr_tx_link_put(to_lnk);
1085 	return to_lnk;
1086 
1087 err_out:
1088 	smcr_link_down_cond_sched(to_lnk);
1089 	smc_wr_tx_link_put(to_lnk);
1090 	return NULL;
1091 }
1092 
1093 static void smcr_buf_unuse(struct smc_buf_desc *buf_desc, bool is_rmb,
1094 			   struct smc_link_group *lgr)
1095 {
1096 	struct mutex *lock;	/* lock buffer list */
1097 	int rc;
1098 
1099 	if (is_rmb && buf_desc->is_conf_rkey && !list_empty(&lgr->list)) {
1100 		/* unregister rmb with peer */
1101 		rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY);
1102 		if (!rc) {
1103 			/* protect against smc_llc_cli_rkey_exchange() */
1104 			mutex_lock(&lgr->llc_conf_mutex);
1105 			smc_llc_do_delete_rkey(lgr, buf_desc);
1106 			buf_desc->is_conf_rkey = false;
1107 			mutex_unlock(&lgr->llc_conf_mutex);
1108 			smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl);
1109 		}
1110 	}
1111 
1112 	if (buf_desc->is_reg_err) {
1113 		/* buf registration failed, reuse not possible */
1114 		lock = is_rmb ? &lgr->rmbs_lock :
1115 				&lgr->sndbufs_lock;
1116 		mutex_lock(lock);
1117 		list_del(&buf_desc->list);
1118 		mutex_unlock(lock);
1119 
1120 		smc_buf_free(lgr, is_rmb, buf_desc);
1121 	} else {
1122 		buf_desc->used = 0;
1123 		memset(buf_desc->cpu_addr, 0, buf_desc->len);
1124 	}
1125 }
1126 
1127 static void smc_buf_unuse(struct smc_connection *conn,
1128 			  struct smc_link_group *lgr)
1129 {
1130 	if (conn->sndbuf_desc) {
1131 		if (!lgr->is_smcd && conn->sndbuf_desc->is_vm) {
1132 			smcr_buf_unuse(conn->sndbuf_desc, false, lgr);
1133 		} else {
1134 			conn->sndbuf_desc->used = 0;
1135 			memset(conn->sndbuf_desc->cpu_addr, 0,
1136 			       conn->sndbuf_desc->len);
1137 		}
1138 	}
1139 	if (conn->rmb_desc) {
1140 		if (!lgr->is_smcd) {
1141 			smcr_buf_unuse(conn->rmb_desc, true, lgr);
1142 		} else {
1143 			conn->rmb_desc->used = 0;
1144 			memset(conn->rmb_desc->cpu_addr, 0,
1145 			       conn->rmb_desc->len +
1146 			       sizeof(struct smcd_cdc_msg));
1147 		}
1148 	}
1149 }
1150 
1151 /* remove a finished connection from its link group */
1152 void smc_conn_free(struct smc_connection *conn)
1153 {
1154 	struct smc_link_group *lgr = conn->lgr;
1155 
1156 	if (!lgr || conn->freed)
1157 		/* Connection has never been registered in a
1158 		 * link group, or has already been freed.
1159 		 */
1160 		return;
1161 
1162 	conn->freed = 1;
1163 	if (!smc_conn_lgr_valid(conn))
1164 		/* Connection has already unregistered from
1165 		 * link group.
1166 		 */
1167 		goto lgr_put;
1168 
1169 	if (lgr->is_smcd) {
1170 		if (!list_empty(&lgr->list))
1171 			smc_ism_unset_conn(conn);
1172 		tasklet_kill(&conn->rx_tsklet);
1173 	} else {
1174 		smc_cdc_wait_pend_tx_wr(conn);
1175 		if (current_work() != &conn->abort_work)
1176 			cancel_work_sync(&conn->abort_work);
1177 	}
1178 	if (!list_empty(&lgr->list)) {
1179 		smc_buf_unuse(conn, lgr); /* allow buffer reuse */
1180 		smc_lgr_unregister_conn(conn);
1181 	}
1182 
1183 	if (!lgr->conns_num)
1184 		smc_lgr_schedule_free_work(lgr);
1185 lgr_put:
1186 	if (!lgr->is_smcd)
1187 		smcr_link_put(conn->lnk); /* link_hold in smc_conn_create() */
1188 	smc_lgr_put(lgr); /* lgr_hold in smc_conn_create() */
1189 }
1190 
1191 /* unregister a link from a buf_desc */
1192 static void smcr_buf_unmap_link(struct smc_buf_desc *buf_desc, bool is_rmb,
1193 				struct smc_link *lnk)
1194 {
1195 	if (is_rmb || buf_desc->is_vm)
1196 		buf_desc->is_reg_mr[lnk->link_idx] = false;
1197 	if (!buf_desc->is_map_ib[lnk->link_idx])
1198 		return;
1199 
1200 	if ((is_rmb || buf_desc->is_vm) &&
1201 	    buf_desc->mr[lnk->link_idx]) {
1202 		smc_ib_put_memory_region(buf_desc->mr[lnk->link_idx]);
1203 		buf_desc->mr[lnk->link_idx] = NULL;
1204 	}
1205 	if (is_rmb)
1206 		smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_FROM_DEVICE);
1207 	else
1208 		smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_TO_DEVICE);
1209 
1210 	sg_free_table(&buf_desc->sgt[lnk->link_idx]);
1211 	buf_desc->is_map_ib[lnk->link_idx] = false;
1212 }
1213 
1214 /* unmap all buffers of lgr for a deleted link */
1215 static void smcr_buf_unmap_lgr(struct smc_link *lnk)
1216 {
1217 	struct smc_link_group *lgr = lnk->lgr;
1218 	struct smc_buf_desc *buf_desc, *bf;
1219 	int i;
1220 
1221 	for (i = 0; i < SMC_RMBE_SIZES; i++) {
1222 		mutex_lock(&lgr->rmbs_lock);
1223 		list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list)
1224 			smcr_buf_unmap_link(buf_desc, true, lnk);
1225 		mutex_unlock(&lgr->rmbs_lock);
1226 		mutex_lock(&lgr->sndbufs_lock);
1227 		list_for_each_entry_safe(buf_desc, bf, &lgr->sndbufs[i],
1228 					 list)
1229 			smcr_buf_unmap_link(buf_desc, false, lnk);
1230 		mutex_unlock(&lgr->sndbufs_lock);
1231 	}
1232 }
1233 
1234 static void smcr_rtoken_clear_link(struct smc_link *lnk)
1235 {
1236 	struct smc_link_group *lgr = lnk->lgr;
1237 	int i;
1238 
1239 	for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1240 		lgr->rtokens[i][lnk->link_idx].rkey = 0;
1241 		lgr->rtokens[i][lnk->link_idx].dma_addr = 0;
1242 	}
1243 }
1244 
1245 static void __smcr_link_clear(struct smc_link *lnk)
1246 {
1247 	struct smc_link_group *lgr = lnk->lgr;
1248 	struct smc_ib_device *smcibdev;
1249 
1250 	smc_wr_free_link_mem(lnk);
1251 	smc_ibdev_cnt_dec(lnk);
1252 	put_device(&lnk->smcibdev->ibdev->dev);
1253 	smcibdev = lnk->smcibdev;
1254 	memset(lnk, 0, sizeof(struct smc_link));
1255 	lnk->state = SMC_LNK_UNUSED;
1256 	if (!atomic_dec_return(&smcibdev->lnk_cnt))
1257 		wake_up(&smcibdev->lnks_deleted);
1258 	smc_lgr_put(lgr); /* lgr_hold in smcr_link_init() */
1259 }
1260 
1261 /* must be called under lgr->llc_conf_mutex lock */
1262 void smcr_link_clear(struct smc_link *lnk, bool log)
1263 {
1264 	if (!lnk->lgr || lnk->clearing ||
1265 	    lnk->state == SMC_LNK_UNUSED)
1266 		return;
1267 	lnk->clearing = 1;
1268 	lnk->peer_qpn = 0;
1269 	smc_llc_link_clear(lnk, log);
1270 	smcr_buf_unmap_lgr(lnk);
1271 	smcr_rtoken_clear_link(lnk);
1272 	smc_ib_modify_qp_error(lnk);
1273 	smc_wr_free_link(lnk);
1274 	smc_ib_destroy_queue_pair(lnk);
1275 	smc_ib_dealloc_protection_domain(lnk);
1276 	smcr_link_put(lnk); /* theoretically last link_put */
1277 }
1278 
1279 void smcr_link_hold(struct smc_link *lnk)
1280 {
1281 	refcount_inc(&lnk->refcnt);
1282 }
1283 
1284 void smcr_link_put(struct smc_link *lnk)
1285 {
1286 	if (refcount_dec_and_test(&lnk->refcnt))
1287 		__smcr_link_clear(lnk);
1288 }
1289 
1290 static void smcr_buf_free(struct smc_link_group *lgr, bool is_rmb,
1291 			  struct smc_buf_desc *buf_desc)
1292 {
1293 	int i;
1294 
1295 	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
1296 		smcr_buf_unmap_link(buf_desc, is_rmb, &lgr->lnk[i]);
1297 
1298 	if (!buf_desc->is_vm && buf_desc->pages)
1299 		__free_pages(buf_desc->pages, buf_desc->order);
1300 	else if (buf_desc->is_vm && buf_desc->cpu_addr)
1301 		vfree(buf_desc->cpu_addr);
1302 	kfree(buf_desc);
1303 }
1304 
1305 static void smcd_buf_free(struct smc_link_group *lgr, bool is_dmb,
1306 			  struct smc_buf_desc *buf_desc)
1307 {
1308 	if (is_dmb) {
1309 		/* restore original buf len */
1310 		buf_desc->len += sizeof(struct smcd_cdc_msg);
1311 		smc_ism_unregister_dmb(lgr->smcd, buf_desc);
1312 	} else {
1313 		kfree(buf_desc->cpu_addr);
1314 	}
1315 	kfree(buf_desc);
1316 }
1317 
1318 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
1319 			 struct smc_buf_desc *buf_desc)
1320 {
1321 	if (lgr->is_smcd)
1322 		smcd_buf_free(lgr, is_rmb, buf_desc);
1323 	else
1324 		smcr_buf_free(lgr, is_rmb, buf_desc);
1325 }
1326 
1327 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
1328 {
1329 	struct smc_buf_desc *buf_desc, *bf_desc;
1330 	struct list_head *buf_list;
1331 	int i;
1332 
1333 	for (i = 0; i < SMC_RMBE_SIZES; i++) {
1334 		if (is_rmb)
1335 			buf_list = &lgr->rmbs[i];
1336 		else
1337 			buf_list = &lgr->sndbufs[i];
1338 		list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
1339 					 list) {
1340 			list_del(&buf_desc->list);
1341 			smc_buf_free(lgr, is_rmb, buf_desc);
1342 		}
1343 	}
1344 }
1345 
1346 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
1347 {
1348 	/* free send buffers */
1349 	__smc_lgr_free_bufs(lgr, false);
1350 	/* free rmbs */
1351 	__smc_lgr_free_bufs(lgr, true);
1352 }
1353 
1354 /* won't be freed until no one accesses to lgr anymore */
1355 static void __smc_lgr_free(struct smc_link_group *lgr)
1356 {
1357 	smc_lgr_free_bufs(lgr);
1358 	if (lgr->is_smcd) {
1359 		if (!atomic_dec_return(&lgr->smcd->lgr_cnt))
1360 			wake_up(&lgr->smcd->lgrs_deleted);
1361 	} else {
1362 		smc_wr_free_lgr_mem(lgr);
1363 		if (!atomic_dec_return(&lgr_cnt))
1364 			wake_up(&lgrs_deleted);
1365 	}
1366 	kfree(lgr);
1367 }
1368 
1369 /* remove a link group */
1370 static void smc_lgr_free(struct smc_link_group *lgr)
1371 {
1372 	int i;
1373 
1374 	if (!lgr->is_smcd) {
1375 		mutex_lock(&lgr->llc_conf_mutex);
1376 		for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1377 			if (lgr->lnk[i].state != SMC_LNK_UNUSED)
1378 				smcr_link_clear(&lgr->lnk[i], false);
1379 		}
1380 		mutex_unlock(&lgr->llc_conf_mutex);
1381 		smc_llc_lgr_clear(lgr);
1382 	}
1383 
1384 	destroy_workqueue(lgr->tx_wq);
1385 	if (lgr->is_smcd) {
1386 		smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
1387 		put_device(&lgr->smcd->dev);
1388 	}
1389 	smc_lgr_put(lgr); /* theoretically last lgr_put */
1390 }
1391 
1392 void smc_lgr_hold(struct smc_link_group *lgr)
1393 {
1394 	refcount_inc(&lgr->refcnt);
1395 }
1396 
1397 void smc_lgr_put(struct smc_link_group *lgr)
1398 {
1399 	if (refcount_dec_and_test(&lgr->refcnt))
1400 		__smc_lgr_free(lgr);
1401 }
1402 
1403 static void smc_sk_wake_ups(struct smc_sock *smc)
1404 {
1405 	smc->sk.sk_write_space(&smc->sk);
1406 	smc->sk.sk_data_ready(&smc->sk);
1407 	smc->sk.sk_state_change(&smc->sk);
1408 }
1409 
1410 /* kill a connection */
1411 static void smc_conn_kill(struct smc_connection *conn, bool soft)
1412 {
1413 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1414 
1415 	if (conn->lgr->is_smcd && conn->lgr->peer_shutdown)
1416 		conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
1417 	else
1418 		smc_close_abort(conn);
1419 	conn->killed = 1;
1420 	smc->sk.sk_err = ECONNABORTED;
1421 	smc_sk_wake_ups(smc);
1422 	if (conn->lgr->is_smcd) {
1423 		smc_ism_unset_conn(conn);
1424 		if (soft)
1425 			tasklet_kill(&conn->rx_tsklet);
1426 		else
1427 			tasklet_unlock_wait(&conn->rx_tsklet);
1428 	} else {
1429 		smc_cdc_wait_pend_tx_wr(conn);
1430 	}
1431 	smc_lgr_unregister_conn(conn);
1432 	smc_close_active_abort(smc);
1433 }
1434 
1435 static void smc_lgr_cleanup(struct smc_link_group *lgr)
1436 {
1437 	if (lgr->is_smcd) {
1438 		smc_ism_signal_shutdown(lgr);
1439 	} else {
1440 		u32 rsn = lgr->llc_termination_rsn;
1441 
1442 		if (!rsn)
1443 			rsn = SMC_LLC_DEL_PROG_INIT_TERM;
1444 		smc_llc_send_link_delete_all(lgr, false, rsn);
1445 		smcr_lgr_link_deactivate_all(lgr);
1446 	}
1447 }
1448 
1449 /* terminate link group
1450  * @soft: true if link group shutdown can take its time
1451  *	  false if immediate link group shutdown is required
1452  */
1453 static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft)
1454 {
1455 	struct smc_connection *conn;
1456 	struct smc_sock *smc;
1457 	struct rb_node *node;
1458 
1459 	if (lgr->terminating)
1460 		return;	/* lgr already terminating */
1461 	/* cancel free_work sync, will terminate when lgr->freeing is set */
1462 	cancel_delayed_work_sync(&lgr->free_work);
1463 	lgr->terminating = 1;
1464 
1465 	/* kill remaining link group connections */
1466 	read_lock_bh(&lgr->conns_lock);
1467 	node = rb_first(&lgr->conns_all);
1468 	while (node) {
1469 		read_unlock_bh(&lgr->conns_lock);
1470 		conn = rb_entry(node, struct smc_connection, alert_node);
1471 		smc = container_of(conn, struct smc_sock, conn);
1472 		sock_hold(&smc->sk); /* sock_put below */
1473 		lock_sock(&smc->sk);
1474 		smc_conn_kill(conn, soft);
1475 		release_sock(&smc->sk);
1476 		sock_put(&smc->sk); /* sock_hold above */
1477 		read_lock_bh(&lgr->conns_lock);
1478 		node = rb_first(&lgr->conns_all);
1479 	}
1480 	read_unlock_bh(&lgr->conns_lock);
1481 	smc_lgr_cleanup(lgr);
1482 	smc_lgr_free(lgr);
1483 }
1484 
1485 /* unlink link group and schedule termination */
1486 void smc_lgr_terminate_sched(struct smc_link_group *lgr)
1487 {
1488 	spinlock_t *lgr_lock;
1489 
1490 	smc_lgr_list_head(lgr, &lgr_lock);
1491 	spin_lock_bh(lgr_lock);
1492 	if (list_empty(&lgr->list) || lgr->terminating || lgr->freeing) {
1493 		spin_unlock_bh(lgr_lock);
1494 		return;	/* lgr already terminating */
1495 	}
1496 	list_del_init(&lgr->list);
1497 	lgr->freeing = 1;
1498 	spin_unlock_bh(lgr_lock);
1499 	schedule_work(&lgr->terminate_work);
1500 }
1501 
1502 /* Called when peer lgr shutdown (regularly or abnormally) is received */
1503 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid, unsigned short vlan)
1504 {
1505 	struct smc_link_group *lgr, *l;
1506 	LIST_HEAD(lgr_free_list);
1507 
1508 	/* run common cleanup function and build free list */
1509 	spin_lock_bh(&dev->lgr_lock);
1510 	list_for_each_entry_safe(lgr, l, &dev->lgr_list, list) {
1511 		if ((!peer_gid || lgr->peer_gid == peer_gid) &&
1512 		    (vlan == VLAN_VID_MASK || lgr->vlan_id == vlan)) {
1513 			if (peer_gid) /* peer triggered termination */
1514 				lgr->peer_shutdown = 1;
1515 			list_move(&lgr->list, &lgr_free_list);
1516 			lgr->freeing = 1;
1517 		}
1518 	}
1519 	spin_unlock_bh(&dev->lgr_lock);
1520 
1521 	/* cancel the regular free workers and actually free lgrs */
1522 	list_for_each_entry_safe(lgr, l, &lgr_free_list, list) {
1523 		list_del_init(&lgr->list);
1524 		schedule_work(&lgr->terminate_work);
1525 	}
1526 }
1527 
1528 /* Called when an SMCD device is removed or the smc module is unloaded */
1529 void smc_smcd_terminate_all(struct smcd_dev *smcd)
1530 {
1531 	struct smc_link_group *lgr, *lg;
1532 	LIST_HEAD(lgr_free_list);
1533 
1534 	spin_lock_bh(&smcd->lgr_lock);
1535 	list_splice_init(&smcd->lgr_list, &lgr_free_list);
1536 	list_for_each_entry(lgr, &lgr_free_list, list)
1537 		lgr->freeing = 1;
1538 	spin_unlock_bh(&smcd->lgr_lock);
1539 
1540 	list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
1541 		list_del_init(&lgr->list);
1542 		__smc_lgr_terminate(lgr, false);
1543 	}
1544 
1545 	if (atomic_read(&smcd->lgr_cnt))
1546 		wait_event(smcd->lgrs_deleted, !atomic_read(&smcd->lgr_cnt));
1547 }
1548 
1549 /* Called when an SMCR device is removed or the smc module is unloaded.
1550  * If smcibdev is given, all SMCR link groups using this device are terminated.
1551  * If smcibdev is NULL, all SMCR link groups are terminated.
1552  */
1553 void smc_smcr_terminate_all(struct smc_ib_device *smcibdev)
1554 {
1555 	struct smc_link_group *lgr, *lg;
1556 	LIST_HEAD(lgr_free_list);
1557 	int i;
1558 
1559 	spin_lock_bh(&smc_lgr_list.lock);
1560 	if (!smcibdev) {
1561 		list_splice_init(&smc_lgr_list.list, &lgr_free_list);
1562 		list_for_each_entry(lgr, &lgr_free_list, list)
1563 			lgr->freeing = 1;
1564 	} else {
1565 		list_for_each_entry_safe(lgr, lg, &smc_lgr_list.list, list) {
1566 			for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1567 				if (lgr->lnk[i].smcibdev == smcibdev)
1568 					smcr_link_down_cond_sched(&lgr->lnk[i]);
1569 			}
1570 		}
1571 	}
1572 	spin_unlock_bh(&smc_lgr_list.lock);
1573 
1574 	list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
1575 		list_del_init(&lgr->list);
1576 		smc_llc_set_termination_rsn(lgr, SMC_LLC_DEL_OP_INIT_TERM);
1577 		__smc_lgr_terminate(lgr, false);
1578 	}
1579 
1580 	if (smcibdev) {
1581 		if (atomic_read(&smcibdev->lnk_cnt))
1582 			wait_event(smcibdev->lnks_deleted,
1583 				   !atomic_read(&smcibdev->lnk_cnt));
1584 	} else {
1585 		if (atomic_read(&lgr_cnt))
1586 			wait_event(lgrs_deleted, !atomic_read(&lgr_cnt));
1587 	}
1588 }
1589 
1590 /* set new lgr type and clear all asymmetric link tagging */
1591 void smcr_lgr_set_type(struct smc_link_group *lgr, enum smc_lgr_type new_type)
1592 {
1593 	char *lgr_type = "";
1594 	int i;
1595 
1596 	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
1597 		if (smc_link_usable(&lgr->lnk[i]))
1598 			lgr->lnk[i].link_is_asym = false;
1599 	if (lgr->type == new_type)
1600 		return;
1601 	lgr->type = new_type;
1602 
1603 	switch (lgr->type) {
1604 	case SMC_LGR_NONE:
1605 		lgr_type = "NONE";
1606 		break;
1607 	case SMC_LGR_SINGLE:
1608 		lgr_type = "SINGLE";
1609 		break;
1610 	case SMC_LGR_SYMMETRIC:
1611 		lgr_type = "SYMMETRIC";
1612 		break;
1613 	case SMC_LGR_ASYMMETRIC_PEER:
1614 		lgr_type = "ASYMMETRIC_PEER";
1615 		break;
1616 	case SMC_LGR_ASYMMETRIC_LOCAL:
1617 		lgr_type = "ASYMMETRIC_LOCAL";
1618 		break;
1619 	}
1620 	pr_warn_ratelimited("smc: SMC-R lg %*phN net %llu state changed: "
1621 			    "%s, pnetid %.16s\n", SMC_LGR_ID_SIZE, &lgr->id,
1622 			    lgr->net->net_cookie, lgr_type, lgr->pnet_id);
1623 }
1624 
1625 /* set new lgr type and tag a link as asymmetric */
1626 void smcr_lgr_set_type_asym(struct smc_link_group *lgr,
1627 			    enum smc_lgr_type new_type, int asym_lnk_idx)
1628 {
1629 	smcr_lgr_set_type(lgr, new_type);
1630 	lgr->lnk[asym_lnk_idx].link_is_asym = true;
1631 }
1632 
1633 /* abort connection, abort_work scheduled from tasklet context */
1634 static void smc_conn_abort_work(struct work_struct *work)
1635 {
1636 	struct smc_connection *conn = container_of(work,
1637 						   struct smc_connection,
1638 						   abort_work);
1639 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1640 
1641 	lock_sock(&smc->sk);
1642 	smc_conn_kill(conn, true);
1643 	release_sock(&smc->sk);
1644 	sock_put(&smc->sk); /* sock_hold done by schedulers of abort_work */
1645 }
1646 
1647 void smcr_port_add(struct smc_ib_device *smcibdev, u8 ibport)
1648 {
1649 	struct smc_link_group *lgr, *n;
1650 
1651 	list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1652 		struct smc_link *link;
1653 
1654 		if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1655 			    SMC_MAX_PNETID_LEN) ||
1656 		    lgr->type == SMC_LGR_SYMMETRIC ||
1657 		    lgr->type == SMC_LGR_ASYMMETRIC_PEER ||
1658 		    !rdma_dev_access_netns(smcibdev->ibdev, lgr->net))
1659 			continue;
1660 
1661 		/* trigger local add link processing */
1662 		link = smc_llc_usable_link(lgr);
1663 		if (link)
1664 			smc_llc_add_link_local(link);
1665 	}
1666 }
1667 
1668 /* link is down - switch connections to alternate link,
1669  * must be called under lgr->llc_conf_mutex lock
1670  */
1671 static void smcr_link_down(struct smc_link *lnk)
1672 {
1673 	struct smc_link_group *lgr = lnk->lgr;
1674 	struct smc_link *to_lnk;
1675 	int del_link_id;
1676 
1677 	if (!lgr || lnk->state == SMC_LNK_UNUSED || list_empty(&lgr->list))
1678 		return;
1679 
1680 	to_lnk = smc_switch_conns(lgr, lnk, true);
1681 	if (!to_lnk) { /* no backup link available */
1682 		smcr_link_clear(lnk, true);
1683 		return;
1684 	}
1685 	smcr_lgr_set_type(lgr, SMC_LGR_SINGLE);
1686 	del_link_id = lnk->link_id;
1687 
1688 	if (lgr->role == SMC_SERV) {
1689 		/* trigger local delete link processing */
1690 		smc_llc_srv_delete_link_local(to_lnk, del_link_id);
1691 	} else {
1692 		if (lgr->llc_flow_lcl.type != SMC_LLC_FLOW_NONE) {
1693 			/* another llc task is ongoing */
1694 			mutex_unlock(&lgr->llc_conf_mutex);
1695 			wait_event_timeout(lgr->llc_flow_waiter,
1696 				(list_empty(&lgr->list) ||
1697 				 lgr->llc_flow_lcl.type == SMC_LLC_FLOW_NONE),
1698 				SMC_LLC_WAIT_TIME);
1699 			mutex_lock(&lgr->llc_conf_mutex);
1700 		}
1701 		if (!list_empty(&lgr->list)) {
1702 			smc_llc_send_delete_link(to_lnk, del_link_id,
1703 						 SMC_LLC_REQ, true,
1704 						 SMC_LLC_DEL_LOST_PATH);
1705 			smcr_link_clear(lnk, true);
1706 		}
1707 		wake_up(&lgr->llc_flow_waiter);	/* wake up next waiter */
1708 	}
1709 }
1710 
1711 /* must be called under lgr->llc_conf_mutex lock */
1712 void smcr_link_down_cond(struct smc_link *lnk)
1713 {
1714 	if (smc_link_downing(&lnk->state)) {
1715 		trace_smcr_link_down(lnk, __builtin_return_address(0));
1716 		smcr_link_down(lnk);
1717 	}
1718 }
1719 
1720 /* will get the lgr->llc_conf_mutex lock */
1721 void smcr_link_down_cond_sched(struct smc_link *lnk)
1722 {
1723 	if (smc_link_downing(&lnk->state)) {
1724 		trace_smcr_link_down(lnk, __builtin_return_address(0));
1725 		schedule_work(&lnk->link_down_wrk);
1726 	}
1727 }
1728 
1729 void smcr_port_err(struct smc_ib_device *smcibdev, u8 ibport)
1730 {
1731 	struct smc_link_group *lgr, *n;
1732 	int i;
1733 
1734 	list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1735 		if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1736 			    SMC_MAX_PNETID_LEN))
1737 			continue; /* lgr is not affected */
1738 		if (list_empty(&lgr->list))
1739 			continue;
1740 		for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1741 			struct smc_link *lnk = &lgr->lnk[i];
1742 
1743 			if (smc_link_usable(lnk) &&
1744 			    lnk->smcibdev == smcibdev && lnk->ibport == ibport)
1745 				smcr_link_down_cond_sched(lnk);
1746 		}
1747 	}
1748 }
1749 
1750 static void smc_link_down_work(struct work_struct *work)
1751 {
1752 	struct smc_link *link = container_of(work, struct smc_link,
1753 					     link_down_wrk);
1754 	struct smc_link_group *lgr = link->lgr;
1755 
1756 	if (list_empty(&lgr->list))
1757 		return;
1758 	wake_up_all(&lgr->llc_msg_waiter);
1759 	mutex_lock(&lgr->llc_conf_mutex);
1760 	smcr_link_down(link);
1761 	mutex_unlock(&lgr->llc_conf_mutex);
1762 }
1763 
1764 static int smc_vlan_by_tcpsk_walk(struct net_device *lower_dev,
1765 				  struct netdev_nested_priv *priv)
1766 {
1767 	unsigned short *vlan_id = (unsigned short *)priv->data;
1768 
1769 	if (is_vlan_dev(lower_dev)) {
1770 		*vlan_id = vlan_dev_vlan_id(lower_dev);
1771 		return 1;
1772 	}
1773 
1774 	return 0;
1775 }
1776 
1777 /* Determine vlan of internal TCP socket. */
1778 int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini)
1779 {
1780 	struct dst_entry *dst = sk_dst_get(clcsock->sk);
1781 	struct netdev_nested_priv priv;
1782 	struct net_device *ndev;
1783 	int rc = 0;
1784 
1785 	ini->vlan_id = 0;
1786 	if (!dst) {
1787 		rc = -ENOTCONN;
1788 		goto out;
1789 	}
1790 	if (!dst->dev) {
1791 		rc = -ENODEV;
1792 		goto out_rel;
1793 	}
1794 
1795 	ndev = dst->dev;
1796 	if (is_vlan_dev(ndev)) {
1797 		ini->vlan_id = vlan_dev_vlan_id(ndev);
1798 		goto out_rel;
1799 	}
1800 
1801 	priv.data = (void *)&ini->vlan_id;
1802 	rtnl_lock();
1803 	netdev_walk_all_lower_dev(ndev, smc_vlan_by_tcpsk_walk, &priv);
1804 	rtnl_unlock();
1805 
1806 out_rel:
1807 	dst_release(dst);
1808 out:
1809 	return rc;
1810 }
1811 
1812 static bool smcr_lgr_match(struct smc_link_group *lgr, u8 smcr_version,
1813 			   u8 peer_systemid[],
1814 			   u8 peer_gid[],
1815 			   u8 peer_mac_v1[],
1816 			   enum smc_lgr_role role, u32 clcqpn,
1817 			   struct net *net)
1818 {
1819 	struct smc_link *lnk;
1820 	int i;
1821 
1822 	if (memcmp(lgr->peer_systemid, peer_systemid, SMC_SYSTEMID_LEN) ||
1823 	    lgr->role != role)
1824 		return false;
1825 
1826 	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1827 		lnk = &lgr->lnk[i];
1828 
1829 		if (!smc_link_active(lnk))
1830 			continue;
1831 		/* use verbs API to check netns, instead of lgr->net */
1832 		if (!rdma_dev_access_netns(lnk->smcibdev->ibdev, net))
1833 			return false;
1834 		if ((lgr->role == SMC_SERV || lnk->peer_qpn == clcqpn) &&
1835 		    !memcmp(lnk->peer_gid, peer_gid, SMC_GID_SIZE) &&
1836 		    (smcr_version == SMC_V2 ||
1837 		     !memcmp(lnk->peer_mac, peer_mac_v1, ETH_ALEN)))
1838 			return true;
1839 	}
1840 	return false;
1841 }
1842 
1843 static bool smcd_lgr_match(struct smc_link_group *lgr,
1844 			   struct smcd_dev *smcismdev, u64 peer_gid)
1845 {
1846 	return lgr->peer_gid == peer_gid && lgr->smcd == smcismdev;
1847 }
1848 
1849 /* create a new SMC connection (and a new link group if necessary) */
1850 int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini)
1851 {
1852 	struct smc_connection *conn = &smc->conn;
1853 	struct net *net = sock_net(&smc->sk);
1854 	struct list_head *lgr_list;
1855 	struct smc_link_group *lgr;
1856 	enum smc_lgr_role role;
1857 	spinlock_t *lgr_lock;
1858 	int rc = 0;
1859 
1860 	lgr_list = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_list :
1861 				  &smc_lgr_list.list;
1862 	lgr_lock = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_lock :
1863 				  &smc_lgr_list.lock;
1864 	ini->first_contact_local = 1;
1865 	role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
1866 	if (role == SMC_CLNT && ini->first_contact_peer)
1867 		/* create new link group as well */
1868 		goto create;
1869 
1870 	/* determine if an existing link group can be reused */
1871 	spin_lock_bh(lgr_lock);
1872 	list_for_each_entry(lgr, lgr_list, list) {
1873 		write_lock_bh(&lgr->conns_lock);
1874 		if ((ini->is_smcd ?
1875 		     smcd_lgr_match(lgr, ini->ism_dev[ini->ism_selected],
1876 				    ini->ism_peer_gid[ini->ism_selected]) :
1877 		     smcr_lgr_match(lgr, ini->smcr_version,
1878 				    ini->peer_systemid,
1879 				    ini->peer_gid, ini->peer_mac, role,
1880 				    ini->ib_clcqpn, net)) &&
1881 		    !lgr->sync_err &&
1882 		    (ini->smcd_version == SMC_V2 ||
1883 		     lgr->vlan_id == ini->vlan_id) &&
1884 		    (role == SMC_CLNT || ini->is_smcd ||
1885 		    (lgr->conns_num < SMC_RMBS_PER_LGR_MAX &&
1886 		      !bitmap_full(lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX)))) {
1887 			/* link group found */
1888 			ini->first_contact_local = 0;
1889 			conn->lgr = lgr;
1890 			rc = smc_lgr_register_conn(conn, false);
1891 			write_unlock_bh(&lgr->conns_lock);
1892 			if (!rc && delayed_work_pending(&lgr->free_work))
1893 				cancel_delayed_work(&lgr->free_work);
1894 			break;
1895 		}
1896 		write_unlock_bh(&lgr->conns_lock);
1897 	}
1898 	spin_unlock_bh(lgr_lock);
1899 	if (rc)
1900 		return rc;
1901 
1902 	if (role == SMC_CLNT && !ini->first_contact_peer &&
1903 	    ini->first_contact_local) {
1904 		/* Server reuses a link group, but Client wants to start
1905 		 * a new one
1906 		 * send out_of_sync decline, reason synchr. error
1907 		 */
1908 		return SMC_CLC_DECL_SYNCERR;
1909 	}
1910 
1911 create:
1912 	if (ini->first_contact_local) {
1913 		rc = smc_lgr_create(smc, ini);
1914 		if (rc)
1915 			goto out;
1916 		lgr = conn->lgr;
1917 		write_lock_bh(&lgr->conns_lock);
1918 		rc = smc_lgr_register_conn(conn, true);
1919 		write_unlock_bh(&lgr->conns_lock);
1920 		if (rc) {
1921 			smc_lgr_cleanup_early(lgr);
1922 			goto out;
1923 		}
1924 	}
1925 	smc_lgr_hold(conn->lgr); /* lgr_put in smc_conn_free() */
1926 	if (!conn->lgr->is_smcd)
1927 		smcr_link_hold(conn->lnk); /* link_put in smc_conn_free() */
1928 	conn->freed = 0;
1929 	conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
1930 	conn->local_tx_ctrl.len = SMC_WR_TX_SIZE;
1931 	conn->urg_state = SMC_URG_READ;
1932 	init_waitqueue_head(&conn->cdc_pend_tx_wq);
1933 	INIT_WORK(&smc->conn.abort_work, smc_conn_abort_work);
1934 	if (ini->is_smcd) {
1935 		conn->rx_off = sizeof(struct smcd_cdc_msg);
1936 		smcd_cdc_rx_init(conn); /* init tasklet for this conn */
1937 	} else {
1938 		conn->rx_off = 0;
1939 	}
1940 #ifndef KERNEL_HAS_ATOMIC64
1941 	spin_lock_init(&conn->acurs_lock);
1942 #endif
1943 
1944 out:
1945 	return rc;
1946 }
1947 
1948 #define SMCD_DMBE_SIZES		6 /* 0 -> 16KB, 1 -> 32KB, .. 6 -> 1MB */
1949 #define SMCR_RMBE_SIZES		5 /* 0 -> 16KB, 1 -> 32KB, .. 5 -> 512KB */
1950 
1951 /* convert the RMB size into the compressed notation (minimum 16K, see
1952  * SMCD/R_DMBE_SIZES.
1953  * In contrast to plain ilog2, this rounds towards the next power of 2,
1954  * so the socket application gets at least its desired sndbuf / rcvbuf size.
1955  */
1956 static u8 smc_compress_bufsize(int size, bool is_smcd, bool is_rmb)
1957 {
1958 	const unsigned int max_scat = SG_MAX_SINGLE_ALLOC * PAGE_SIZE;
1959 	u8 compressed;
1960 
1961 	if (size <= SMC_BUF_MIN_SIZE)
1962 		return 0;
1963 
1964 	size = (size - 1) >> 14;  /* convert to 16K multiple */
1965 	compressed = min_t(u8, ilog2(size) + 1,
1966 			   is_smcd ? SMCD_DMBE_SIZES : SMCR_RMBE_SIZES);
1967 
1968 	if (!is_smcd && is_rmb)
1969 		/* RMBs are backed by & limited to max size of scatterlists */
1970 		compressed = min_t(u8, compressed, ilog2(max_scat >> 14));
1971 
1972 	return compressed;
1973 }
1974 
1975 /* convert the RMB size from compressed notation into integer */
1976 int smc_uncompress_bufsize(u8 compressed)
1977 {
1978 	u32 size;
1979 
1980 	size = 0x00000001 << (((int)compressed) + 14);
1981 	return (int)size;
1982 }
1983 
1984 /* try to reuse a sndbuf or rmb description slot for a certain
1985  * buffer size; if not available, return NULL
1986  */
1987 static struct smc_buf_desc *smc_buf_get_slot(int compressed_bufsize,
1988 					     struct mutex *lock,
1989 					     struct list_head *buf_list)
1990 {
1991 	struct smc_buf_desc *buf_slot;
1992 
1993 	mutex_lock(lock);
1994 	list_for_each_entry(buf_slot, buf_list, list) {
1995 		if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
1996 			mutex_unlock(lock);
1997 			return buf_slot;
1998 		}
1999 	}
2000 	mutex_unlock(lock);
2001 	return NULL;
2002 }
2003 
2004 /* one of the conditions for announcing a receiver's current window size is
2005  * that it "results in a minimum increase in the window size of 10% of the
2006  * receive buffer space" [RFC7609]
2007  */
2008 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
2009 {
2010 	return max_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
2011 }
2012 
2013 /* map an buf to a link */
2014 static int smcr_buf_map_link(struct smc_buf_desc *buf_desc, bool is_rmb,
2015 			     struct smc_link *lnk)
2016 {
2017 	int rc, i, nents, offset, buf_size, size, access_flags;
2018 	struct scatterlist *sg;
2019 	void *buf;
2020 
2021 	if (buf_desc->is_map_ib[lnk->link_idx])
2022 		return 0;
2023 
2024 	if (buf_desc->is_vm) {
2025 		buf = buf_desc->cpu_addr;
2026 		buf_size = buf_desc->len;
2027 		offset = offset_in_page(buf_desc->cpu_addr);
2028 		nents = PAGE_ALIGN(buf_size + offset) / PAGE_SIZE;
2029 	} else {
2030 		nents = 1;
2031 	}
2032 
2033 	rc = sg_alloc_table(&buf_desc->sgt[lnk->link_idx], nents, GFP_KERNEL);
2034 	if (rc)
2035 		return rc;
2036 
2037 	if (buf_desc->is_vm) {
2038 		/* virtually contiguous buffer */
2039 		for_each_sg(buf_desc->sgt[lnk->link_idx].sgl, sg, nents, i) {
2040 			size = min_t(int, PAGE_SIZE - offset, buf_size);
2041 			sg_set_page(sg, vmalloc_to_page(buf), size, offset);
2042 			buf += size / sizeof(*buf);
2043 			buf_size -= size;
2044 			offset = 0;
2045 		}
2046 	} else {
2047 		/* physically contiguous buffer */
2048 		sg_set_buf(buf_desc->sgt[lnk->link_idx].sgl,
2049 			   buf_desc->cpu_addr, buf_desc->len);
2050 	}
2051 
2052 	/* map sg table to DMA address */
2053 	rc = smc_ib_buf_map_sg(lnk, buf_desc,
2054 			       is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
2055 	/* SMC protocol depends on mapping to one DMA address only */
2056 	if (rc != nents) {
2057 		rc = -EAGAIN;
2058 		goto free_table;
2059 	}
2060 
2061 	buf_desc->is_dma_need_sync |=
2062 		smc_ib_is_sg_need_sync(lnk, buf_desc) << lnk->link_idx;
2063 
2064 	if (is_rmb || buf_desc->is_vm) {
2065 		/* create a new memory region for the RMB or vzalloced sndbuf */
2066 		access_flags = is_rmb ?
2067 			       IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE :
2068 			       IB_ACCESS_LOCAL_WRITE;
2069 
2070 		rc = smc_ib_get_memory_region(lnk->roce_pd, access_flags,
2071 					      buf_desc, lnk->link_idx);
2072 		if (rc)
2073 			goto buf_unmap;
2074 		smc_ib_sync_sg_for_device(lnk, buf_desc,
2075 					  is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
2076 	}
2077 	buf_desc->is_map_ib[lnk->link_idx] = true;
2078 	return 0;
2079 
2080 buf_unmap:
2081 	smc_ib_buf_unmap_sg(lnk, buf_desc,
2082 			    is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
2083 free_table:
2084 	sg_free_table(&buf_desc->sgt[lnk->link_idx]);
2085 	return rc;
2086 }
2087 
2088 /* register a new buf on IB device, rmb or vzalloced sndbuf
2089  * must be called under lgr->llc_conf_mutex lock
2090  */
2091 int smcr_link_reg_buf(struct smc_link *link, struct smc_buf_desc *buf_desc)
2092 {
2093 	if (list_empty(&link->lgr->list))
2094 		return -ENOLINK;
2095 	if (!buf_desc->is_reg_mr[link->link_idx]) {
2096 		/* register memory region for new buf */
2097 		if (buf_desc->is_vm)
2098 			buf_desc->mr[link->link_idx]->iova =
2099 				(uintptr_t)buf_desc->cpu_addr;
2100 		if (smc_wr_reg_send(link, buf_desc->mr[link->link_idx])) {
2101 			buf_desc->is_reg_err = true;
2102 			return -EFAULT;
2103 		}
2104 		buf_desc->is_reg_mr[link->link_idx] = true;
2105 	}
2106 	return 0;
2107 }
2108 
2109 static int _smcr_buf_map_lgr(struct smc_link *lnk, struct mutex *lock,
2110 			     struct list_head *lst, bool is_rmb)
2111 {
2112 	struct smc_buf_desc *buf_desc, *bf;
2113 	int rc = 0;
2114 
2115 	mutex_lock(lock);
2116 	list_for_each_entry_safe(buf_desc, bf, lst, list) {
2117 		if (!buf_desc->used)
2118 			continue;
2119 		rc = smcr_buf_map_link(buf_desc, is_rmb, lnk);
2120 		if (rc)
2121 			goto out;
2122 	}
2123 out:
2124 	mutex_unlock(lock);
2125 	return rc;
2126 }
2127 
2128 /* map all used buffers of lgr for a new link */
2129 int smcr_buf_map_lgr(struct smc_link *lnk)
2130 {
2131 	struct smc_link_group *lgr = lnk->lgr;
2132 	int i, rc = 0;
2133 
2134 	for (i = 0; i < SMC_RMBE_SIZES; i++) {
2135 		rc = _smcr_buf_map_lgr(lnk, &lgr->rmbs_lock,
2136 				       &lgr->rmbs[i], true);
2137 		if (rc)
2138 			return rc;
2139 		rc = _smcr_buf_map_lgr(lnk, &lgr->sndbufs_lock,
2140 				       &lgr->sndbufs[i], false);
2141 		if (rc)
2142 			return rc;
2143 	}
2144 	return 0;
2145 }
2146 
2147 /* register all used buffers of lgr for a new link,
2148  * must be called under lgr->llc_conf_mutex lock
2149  */
2150 int smcr_buf_reg_lgr(struct smc_link *lnk)
2151 {
2152 	struct smc_link_group *lgr = lnk->lgr;
2153 	struct smc_buf_desc *buf_desc, *bf;
2154 	int i, rc = 0;
2155 
2156 	/* reg all RMBs for a new link */
2157 	mutex_lock(&lgr->rmbs_lock);
2158 	for (i = 0; i < SMC_RMBE_SIZES; i++) {
2159 		list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list) {
2160 			if (!buf_desc->used)
2161 				continue;
2162 			rc = smcr_link_reg_buf(lnk, buf_desc);
2163 			if (rc) {
2164 				mutex_unlock(&lgr->rmbs_lock);
2165 				return rc;
2166 			}
2167 		}
2168 	}
2169 	mutex_unlock(&lgr->rmbs_lock);
2170 
2171 	if (lgr->buf_type == SMCR_PHYS_CONT_BUFS)
2172 		return rc;
2173 
2174 	/* reg all vzalloced sndbufs for a new link */
2175 	mutex_lock(&lgr->sndbufs_lock);
2176 	for (i = 0; i < SMC_RMBE_SIZES; i++) {
2177 		list_for_each_entry_safe(buf_desc, bf, &lgr->sndbufs[i], list) {
2178 			if (!buf_desc->used || !buf_desc->is_vm)
2179 				continue;
2180 			rc = smcr_link_reg_buf(lnk, buf_desc);
2181 			if (rc) {
2182 				mutex_unlock(&lgr->sndbufs_lock);
2183 				return rc;
2184 			}
2185 		}
2186 	}
2187 	mutex_unlock(&lgr->sndbufs_lock);
2188 	return rc;
2189 }
2190 
2191 static struct smc_buf_desc *smcr_new_buf_create(struct smc_link_group *lgr,
2192 						bool is_rmb, int bufsize)
2193 {
2194 	struct smc_buf_desc *buf_desc;
2195 
2196 	/* try to alloc a new buffer */
2197 	buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
2198 	if (!buf_desc)
2199 		return ERR_PTR(-ENOMEM);
2200 
2201 	switch (lgr->buf_type) {
2202 	case SMCR_PHYS_CONT_BUFS:
2203 	case SMCR_MIXED_BUFS:
2204 		buf_desc->order = get_order(bufsize);
2205 		buf_desc->pages = alloc_pages(GFP_KERNEL | __GFP_NOWARN |
2206 					      __GFP_NOMEMALLOC | __GFP_COMP |
2207 					      __GFP_NORETRY | __GFP_ZERO,
2208 					      buf_desc->order);
2209 		if (buf_desc->pages) {
2210 			buf_desc->cpu_addr =
2211 				(void *)page_address(buf_desc->pages);
2212 			buf_desc->len = bufsize;
2213 			buf_desc->is_vm = false;
2214 			break;
2215 		}
2216 		if (lgr->buf_type == SMCR_PHYS_CONT_BUFS)
2217 			goto out;
2218 		fallthrough;	// try virtually continguous buf
2219 	case SMCR_VIRT_CONT_BUFS:
2220 		buf_desc->order = get_order(bufsize);
2221 		buf_desc->cpu_addr = vzalloc(PAGE_SIZE << buf_desc->order);
2222 		if (!buf_desc->cpu_addr)
2223 			goto out;
2224 		buf_desc->pages = NULL;
2225 		buf_desc->len = bufsize;
2226 		buf_desc->is_vm = true;
2227 		break;
2228 	}
2229 	return buf_desc;
2230 
2231 out:
2232 	kfree(buf_desc);
2233 	return ERR_PTR(-EAGAIN);
2234 }
2235 
2236 /* map buf_desc on all usable links,
2237  * unused buffers stay mapped as long as the link is up
2238  */
2239 static int smcr_buf_map_usable_links(struct smc_link_group *lgr,
2240 				     struct smc_buf_desc *buf_desc, bool is_rmb)
2241 {
2242 	int i, rc = 0, cnt = 0;
2243 
2244 	/* protect against parallel link reconfiguration */
2245 	mutex_lock(&lgr->llc_conf_mutex);
2246 	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
2247 		struct smc_link *lnk = &lgr->lnk[i];
2248 
2249 		if (!smc_link_usable(lnk))
2250 			continue;
2251 		if (smcr_buf_map_link(buf_desc, is_rmb, lnk)) {
2252 			rc = -ENOMEM;
2253 			goto out;
2254 		}
2255 		cnt++;
2256 	}
2257 out:
2258 	mutex_unlock(&lgr->llc_conf_mutex);
2259 	if (!rc && !cnt)
2260 		rc = -EINVAL;
2261 	return rc;
2262 }
2263 
2264 static struct smc_buf_desc *smcd_new_buf_create(struct smc_link_group *lgr,
2265 						bool is_dmb, int bufsize)
2266 {
2267 	struct smc_buf_desc *buf_desc;
2268 	int rc;
2269 
2270 	/* try to alloc a new DMB */
2271 	buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
2272 	if (!buf_desc)
2273 		return ERR_PTR(-ENOMEM);
2274 	if (is_dmb) {
2275 		rc = smc_ism_register_dmb(lgr, bufsize, buf_desc);
2276 		if (rc) {
2277 			kfree(buf_desc);
2278 			if (rc == -ENOMEM)
2279 				return ERR_PTR(-EAGAIN);
2280 			if (rc == -ENOSPC)
2281 				return ERR_PTR(-ENOSPC);
2282 			return ERR_PTR(-EIO);
2283 		}
2284 		buf_desc->pages = virt_to_page(buf_desc->cpu_addr);
2285 		/* CDC header stored in buf. So, pretend it was smaller */
2286 		buf_desc->len = bufsize - sizeof(struct smcd_cdc_msg);
2287 	} else {
2288 		buf_desc->cpu_addr = kzalloc(bufsize, GFP_KERNEL |
2289 					     __GFP_NOWARN | __GFP_NORETRY |
2290 					     __GFP_NOMEMALLOC);
2291 		if (!buf_desc->cpu_addr) {
2292 			kfree(buf_desc);
2293 			return ERR_PTR(-EAGAIN);
2294 		}
2295 		buf_desc->len = bufsize;
2296 	}
2297 	return buf_desc;
2298 }
2299 
2300 static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
2301 {
2302 	struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
2303 	struct smc_connection *conn = &smc->conn;
2304 	struct smc_link_group *lgr = conn->lgr;
2305 	struct list_head *buf_list;
2306 	int bufsize, bufsize_short;
2307 	bool is_dgraded = false;
2308 	struct mutex *lock;	/* lock buffer list */
2309 	int sk_buf_size;
2310 
2311 	if (is_rmb)
2312 		/* use socket recv buffer size (w/o overhead) as start value */
2313 		sk_buf_size = smc->sk.sk_rcvbuf / 2;
2314 	else
2315 		/* use socket send buffer size (w/o overhead) as start value */
2316 		sk_buf_size = smc->sk.sk_sndbuf / 2;
2317 
2318 	for (bufsize_short = smc_compress_bufsize(sk_buf_size, is_smcd, is_rmb);
2319 	     bufsize_short >= 0; bufsize_short--) {
2320 		if (is_rmb) {
2321 			lock = &lgr->rmbs_lock;
2322 			buf_list = &lgr->rmbs[bufsize_short];
2323 		} else {
2324 			lock = &lgr->sndbufs_lock;
2325 			buf_list = &lgr->sndbufs[bufsize_short];
2326 		}
2327 		bufsize = smc_uncompress_bufsize(bufsize_short);
2328 
2329 		/* check for reusable slot in the link group */
2330 		buf_desc = smc_buf_get_slot(bufsize_short, lock, buf_list);
2331 		if (buf_desc) {
2332 			buf_desc->is_dma_need_sync = 0;
2333 			SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, bufsize);
2334 			SMC_STAT_BUF_REUSE(smc, is_smcd, is_rmb);
2335 			break; /* found reusable slot */
2336 		}
2337 
2338 		if (is_smcd)
2339 			buf_desc = smcd_new_buf_create(lgr, is_rmb, bufsize);
2340 		else
2341 			buf_desc = smcr_new_buf_create(lgr, is_rmb, bufsize);
2342 
2343 		if (PTR_ERR(buf_desc) == -ENOMEM)
2344 			break;
2345 		if (IS_ERR(buf_desc)) {
2346 			if (!is_dgraded) {
2347 				is_dgraded = true;
2348 				SMC_STAT_RMB_DOWNGRADED(smc, is_smcd, is_rmb);
2349 			}
2350 			continue;
2351 		}
2352 
2353 		SMC_STAT_RMB_ALLOC(smc, is_smcd, is_rmb);
2354 		SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, bufsize);
2355 		buf_desc->used = 1;
2356 		mutex_lock(lock);
2357 		list_add(&buf_desc->list, buf_list);
2358 		mutex_unlock(lock);
2359 		break; /* found */
2360 	}
2361 
2362 	if (IS_ERR(buf_desc))
2363 		return PTR_ERR(buf_desc);
2364 
2365 	if (!is_smcd) {
2366 		if (smcr_buf_map_usable_links(lgr, buf_desc, is_rmb)) {
2367 			smcr_buf_unuse(buf_desc, is_rmb, lgr);
2368 			return -ENOMEM;
2369 		}
2370 	}
2371 
2372 	if (is_rmb) {
2373 		conn->rmb_desc = buf_desc;
2374 		conn->rmbe_size_short = bufsize_short;
2375 		smc->sk.sk_rcvbuf = bufsize * 2;
2376 		atomic_set(&conn->bytes_to_rcv, 0);
2377 		conn->rmbe_update_limit =
2378 			smc_rmb_wnd_update_limit(buf_desc->len);
2379 		if (is_smcd)
2380 			smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */
2381 	} else {
2382 		conn->sndbuf_desc = buf_desc;
2383 		smc->sk.sk_sndbuf = bufsize * 2;
2384 		atomic_set(&conn->sndbuf_space, bufsize);
2385 	}
2386 	return 0;
2387 }
2388 
2389 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
2390 {
2391 	if (!conn->sndbuf_desc->is_dma_need_sync)
2392 		return;
2393 	if (!smc_conn_lgr_valid(conn) || conn->lgr->is_smcd ||
2394 	    !smc_link_active(conn->lnk))
2395 		return;
2396 	smc_ib_sync_sg_for_device(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE);
2397 }
2398 
2399 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
2400 {
2401 	int i;
2402 
2403 	if (!conn->rmb_desc->is_dma_need_sync)
2404 		return;
2405 	if (!smc_conn_lgr_valid(conn) || conn->lgr->is_smcd)
2406 		return;
2407 	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
2408 		if (!smc_link_active(&conn->lgr->lnk[i]))
2409 			continue;
2410 		smc_ib_sync_sg_for_cpu(&conn->lgr->lnk[i], conn->rmb_desc,
2411 				       DMA_FROM_DEVICE);
2412 	}
2413 }
2414 
2415 /* create the send and receive buffer for an SMC socket;
2416  * receive buffers are called RMBs;
2417  * (even though the SMC protocol allows more than one RMB-element per RMB,
2418  * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
2419  * extra RMB for every connection in a link group
2420  */
2421 int smc_buf_create(struct smc_sock *smc, bool is_smcd)
2422 {
2423 	int rc;
2424 
2425 	/* create send buffer */
2426 	rc = __smc_buf_create(smc, is_smcd, false);
2427 	if (rc)
2428 		return rc;
2429 	/* create rmb */
2430 	rc = __smc_buf_create(smc, is_smcd, true);
2431 	if (rc) {
2432 		mutex_lock(&smc->conn.lgr->sndbufs_lock);
2433 		list_del(&smc->conn.sndbuf_desc->list);
2434 		mutex_unlock(&smc->conn.lgr->sndbufs_lock);
2435 		smc_buf_free(smc->conn.lgr, false, smc->conn.sndbuf_desc);
2436 		smc->conn.sndbuf_desc = NULL;
2437 	}
2438 	return rc;
2439 }
2440 
2441 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
2442 {
2443 	int i;
2444 
2445 	for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
2446 		if (!test_and_set_bit(i, lgr->rtokens_used_mask))
2447 			return i;
2448 	}
2449 	return -ENOSPC;
2450 }
2451 
2452 static int smc_rtoken_find_by_link(struct smc_link_group *lgr, int lnk_idx,
2453 				   u32 rkey)
2454 {
2455 	int i;
2456 
2457 	for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2458 		if (test_bit(i, lgr->rtokens_used_mask) &&
2459 		    lgr->rtokens[i][lnk_idx].rkey == rkey)
2460 			return i;
2461 	}
2462 	return -ENOENT;
2463 }
2464 
2465 /* set rtoken for a new link to an existing rmb */
2466 void smc_rtoken_set(struct smc_link_group *lgr, int link_idx, int link_idx_new,
2467 		    __be32 nw_rkey_known, __be64 nw_vaddr, __be32 nw_rkey)
2468 {
2469 	int rtok_idx;
2470 
2471 	rtok_idx = smc_rtoken_find_by_link(lgr, link_idx, ntohl(nw_rkey_known));
2472 	if (rtok_idx == -ENOENT)
2473 		return;
2474 	lgr->rtokens[rtok_idx][link_idx_new].rkey = ntohl(nw_rkey);
2475 	lgr->rtokens[rtok_idx][link_idx_new].dma_addr = be64_to_cpu(nw_vaddr);
2476 }
2477 
2478 /* set rtoken for a new link whose link_id is given */
2479 void smc_rtoken_set2(struct smc_link_group *lgr, int rtok_idx, int link_id,
2480 		     __be64 nw_vaddr, __be32 nw_rkey)
2481 {
2482 	u64 dma_addr = be64_to_cpu(nw_vaddr);
2483 	u32 rkey = ntohl(nw_rkey);
2484 	bool found = false;
2485 	int link_idx;
2486 
2487 	for (link_idx = 0; link_idx < SMC_LINKS_PER_LGR_MAX; link_idx++) {
2488 		if (lgr->lnk[link_idx].link_id == link_id) {
2489 			found = true;
2490 			break;
2491 		}
2492 	}
2493 	if (!found)
2494 		return;
2495 	lgr->rtokens[rtok_idx][link_idx].rkey = rkey;
2496 	lgr->rtokens[rtok_idx][link_idx].dma_addr = dma_addr;
2497 }
2498 
2499 /* add a new rtoken from peer */
2500 int smc_rtoken_add(struct smc_link *lnk, __be64 nw_vaddr, __be32 nw_rkey)
2501 {
2502 	struct smc_link_group *lgr = smc_get_lgr(lnk);
2503 	u64 dma_addr = be64_to_cpu(nw_vaddr);
2504 	u32 rkey = ntohl(nw_rkey);
2505 	int i;
2506 
2507 	for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2508 		if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
2509 		    lgr->rtokens[i][lnk->link_idx].dma_addr == dma_addr &&
2510 		    test_bit(i, lgr->rtokens_used_mask)) {
2511 			/* already in list */
2512 			return i;
2513 		}
2514 	}
2515 	i = smc_rmb_reserve_rtoken_idx(lgr);
2516 	if (i < 0)
2517 		return i;
2518 	lgr->rtokens[i][lnk->link_idx].rkey = rkey;
2519 	lgr->rtokens[i][lnk->link_idx].dma_addr = dma_addr;
2520 	return i;
2521 }
2522 
2523 /* delete an rtoken from all links */
2524 int smc_rtoken_delete(struct smc_link *lnk, __be32 nw_rkey)
2525 {
2526 	struct smc_link_group *lgr = smc_get_lgr(lnk);
2527 	u32 rkey = ntohl(nw_rkey);
2528 	int i, j;
2529 
2530 	for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2531 		if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
2532 		    test_bit(i, lgr->rtokens_used_mask)) {
2533 			for (j = 0; j < SMC_LINKS_PER_LGR_MAX; j++) {
2534 				lgr->rtokens[i][j].rkey = 0;
2535 				lgr->rtokens[i][j].dma_addr = 0;
2536 			}
2537 			clear_bit(i, lgr->rtokens_used_mask);
2538 			return 0;
2539 		}
2540 	}
2541 	return -ENOENT;
2542 }
2543 
2544 /* save rkey and dma_addr received from peer during clc handshake */
2545 int smc_rmb_rtoken_handling(struct smc_connection *conn,
2546 			    struct smc_link *lnk,
2547 			    struct smc_clc_msg_accept_confirm *clc)
2548 {
2549 	conn->rtoken_idx = smc_rtoken_add(lnk, clc->r0.rmb_dma_addr,
2550 					  clc->r0.rmb_rkey);
2551 	if (conn->rtoken_idx < 0)
2552 		return conn->rtoken_idx;
2553 	return 0;
2554 }
2555 
2556 static void smc_core_going_away(void)
2557 {
2558 	struct smc_ib_device *smcibdev;
2559 	struct smcd_dev *smcd;
2560 
2561 	mutex_lock(&smc_ib_devices.mutex);
2562 	list_for_each_entry(smcibdev, &smc_ib_devices.list, list) {
2563 		int i;
2564 
2565 		for (i = 0; i < SMC_MAX_PORTS; i++)
2566 			set_bit(i, smcibdev->ports_going_away);
2567 	}
2568 	mutex_unlock(&smc_ib_devices.mutex);
2569 
2570 	mutex_lock(&smcd_dev_list.mutex);
2571 	list_for_each_entry(smcd, &smcd_dev_list.list, list) {
2572 		smcd->going_away = 1;
2573 	}
2574 	mutex_unlock(&smcd_dev_list.mutex);
2575 }
2576 
2577 /* Clean up all SMC link groups */
2578 static void smc_lgrs_shutdown(void)
2579 {
2580 	struct smcd_dev *smcd;
2581 
2582 	smc_core_going_away();
2583 
2584 	smc_smcr_terminate_all(NULL);
2585 
2586 	mutex_lock(&smcd_dev_list.mutex);
2587 	list_for_each_entry(smcd, &smcd_dev_list.list, list)
2588 		smc_smcd_terminate_all(smcd);
2589 	mutex_unlock(&smcd_dev_list.mutex);
2590 }
2591 
2592 static int smc_core_reboot_event(struct notifier_block *this,
2593 				 unsigned long event, void *ptr)
2594 {
2595 	smc_lgrs_shutdown();
2596 	smc_ib_unregister_client();
2597 	return 0;
2598 }
2599 
2600 static struct notifier_block smc_reboot_notifier = {
2601 	.notifier_call = smc_core_reboot_event,
2602 };
2603 
2604 int __init smc_core_init(void)
2605 {
2606 	return register_reboot_notifier(&smc_reboot_notifier);
2607 }
2608 
2609 /* Called (from smc_exit) when module is removed */
2610 void smc_core_exit(void)
2611 {
2612 	unregister_reboot_notifier(&smc_reboot_notifier);
2613 	smc_lgrs_shutdown();
2614 }
2615