xref: /linux/net/smc/smc_core.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
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 	conn->lnk = NULL;	/* reset conn->lnk first */
131 	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
132 		struct smc_link *lnk = &conn->lgr->lnk[i];
133 
134 		if (lnk->state != expected || lnk->link_is_asym)
135 			continue;
136 		if (conn->lgr->role == SMC_CLNT) {
137 			conn->lnk = lnk; /* temporary, SMC server assigns link*/
138 			break;
139 		}
140 		if (conn->lgr->conns_num % 2) {
141 			for (j = i + 1; j < SMC_LINKS_PER_LGR_MAX; j++) {
142 				struct smc_link *lnk2;
143 
144 				lnk2 = &conn->lgr->lnk[j];
145 				if (lnk2->state == expected &&
146 				    !lnk2->link_is_asym) {
147 					conn->lnk = lnk2;
148 					break;
149 				}
150 			}
151 		}
152 		if (!conn->lnk)
153 			conn->lnk = lnk;
154 		break;
155 	}
156 	if (!conn->lnk)
157 		return SMC_CLC_DECL_NOACTLINK;
158 	atomic_inc(&conn->lnk->conn_cnt);
159 	return 0;
160 }
161 
162 /* Register connection in link group by assigning an alert token
163  * registered in a search tree.
164  * Requires @conns_lock
165  * Note that '0' is a reserved value and not assigned.
166  */
167 static int smc_lgr_register_conn(struct smc_connection *conn, bool first)
168 {
169 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
170 	static atomic_t nexttoken = ATOMIC_INIT(0);
171 	int rc;
172 
173 	if (!conn->lgr->is_smcd) {
174 		rc = smcr_lgr_conn_assign_link(conn, first);
175 		if (rc) {
176 			conn->lgr = NULL;
177 			return rc;
178 		}
179 	}
180 	/* find a new alert_token_local value not yet used by some connection
181 	 * in this link group
182 	 */
183 	sock_hold(&smc->sk); /* sock_put in smc_lgr_unregister_conn() */
184 	while (!conn->alert_token_local) {
185 		conn->alert_token_local = atomic_inc_return(&nexttoken);
186 		if (smc_lgr_find_conn(conn->alert_token_local, conn->lgr))
187 			conn->alert_token_local = 0;
188 	}
189 	smc_lgr_add_alert_token(conn);
190 	conn->lgr->conns_num++;
191 	return 0;
192 }
193 
194 /* Unregister connection and reset the alert token of the given connection<
195  */
196 static void __smc_lgr_unregister_conn(struct smc_connection *conn)
197 {
198 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
199 	struct smc_link_group *lgr = conn->lgr;
200 
201 	rb_erase(&conn->alert_node, &lgr->conns_all);
202 	if (conn->lnk)
203 		atomic_dec(&conn->lnk->conn_cnt);
204 	lgr->conns_num--;
205 	conn->alert_token_local = 0;
206 	sock_put(&smc->sk); /* sock_hold in smc_lgr_register_conn() */
207 }
208 
209 /* Unregister connection from lgr
210  */
211 static void smc_lgr_unregister_conn(struct smc_connection *conn)
212 {
213 	struct smc_link_group *lgr = conn->lgr;
214 
215 	if (!smc_conn_lgr_valid(conn))
216 		return;
217 	write_lock_bh(&lgr->conns_lock);
218 	if (conn->alert_token_local) {
219 		__smc_lgr_unregister_conn(conn);
220 	}
221 	write_unlock_bh(&lgr->conns_lock);
222 }
223 
224 static void smc_lgr_buf_list_add(struct smc_link_group *lgr,
225 				 bool is_rmb,
226 				 struct list_head *buf_list,
227 				 struct smc_buf_desc *buf_desc)
228 {
229 	list_add(&buf_desc->list, buf_list);
230 	if (is_rmb) {
231 		lgr->alloc_rmbs += buf_desc->len;
232 		lgr->alloc_rmbs +=
233 			lgr->is_smcd ? sizeof(struct smcd_cdc_msg) : 0;
234 	} else {
235 		lgr->alloc_sndbufs += buf_desc->len;
236 	}
237 }
238 
239 static void smc_lgr_buf_list_del(struct smc_link_group *lgr,
240 				 bool is_rmb,
241 				 struct smc_buf_desc *buf_desc)
242 {
243 	list_del(&buf_desc->list);
244 	if (is_rmb) {
245 		lgr->alloc_rmbs -= buf_desc->len;
246 		lgr->alloc_rmbs -=
247 			lgr->is_smcd ? sizeof(struct smcd_cdc_msg) : 0;
248 	} else {
249 		lgr->alloc_sndbufs -= buf_desc->len;
250 	}
251 }
252 
253 int smc_nl_get_sys_info(struct sk_buff *skb, struct netlink_callback *cb)
254 {
255 	struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
256 	char hostname[SMC_MAX_HOSTNAME_LEN + 1];
257 	char smc_seid[SMC_MAX_EID_LEN + 1];
258 	struct nlattr *attrs;
259 	u8 *seid = NULL;
260 	u8 *host = NULL;
261 	void *nlh;
262 
263 	nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
264 			  &smc_gen_nl_family, NLM_F_MULTI,
265 			  SMC_NETLINK_GET_SYS_INFO);
266 	if (!nlh)
267 		goto errmsg;
268 	if (cb_ctx->pos[0])
269 		goto errout;
270 	attrs = nla_nest_start(skb, SMC_GEN_SYS_INFO);
271 	if (!attrs)
272 		goto errout;
273 	if (nla_put_u8(skb, SMC_NLA_SYS_VER, SMC_V2))
274 		goto errattr;
275 	if (nla_put_u8(skb, SMC_NLA_SYS_REL, SMC_RELEASE))
276 		goto errattr;
277 	if (nla_put_u8(skb, SMC_NLA_SYS_IS_ISM_V2, smc_ism_is_v2_capable()))
278 		goto errattr;
279 	if (nla_put_u8(skb, SMC_NLA_SYS_IS_SMCR_V2, true))
280 		goto errattr;
281 	smc_clc_get_hostname(&host);
282 	if (host) {
283 		memcpy(hostname, host, SMC_MAX_HOSTNAME_LEN);
284 		hostname[SMC_MAX_HOSTNAME_LEN] = 0;
285 		if (nla_put_string(skb, SMC_NLA_SYS_LOCAL_HOST, hostname))
286 			goto errattr;
287 	}
288 	if (smc_ism_is_v2_capable()) {
289 		smc_ism_get_system_eid(&seid);
290 		memcpy(smc_seid, seid, SMC_MAX_EID_LEN);
291 		smc_seid[SMC_MAX_EID_LEN] = 0;
292 		if (nla_put_string(skb, SMC_NLA_SYS_SEID, smc_seid))
293 			goto errattr;
294 	}
295 	nla_nest_end(skb, attrs);
296 	genlmsg_end(skb, nlh);
297 	cb_ctx->pos[0] = 1;
298 	return skb->len;
299 
300 errattr:
301 	nla_nest_cancel(skb, attrs);
302 errout:
303 	genlmsg_cancel(skb, nlh);
304 errmsg:
305 	return skb->len;
306 }
307 
308 /* Fill SMC_NLA_LGR_D_V2_COMMON/SMC_NLA_LGR_R_V2_COMMON nested attributes */
309 static int smc_nl_fill_lgr_v2_common(struct smc_link_group *lgr,
310 				     struct sk_buff *skb,
311 				     struct netlink_callback *cb,
312 				     struct nlattr *v2_attrs)
313 {
314 	char smc_host[SMC_MAX_HOSTNAME_LEN + 1];
315 	char smc_eid[SMC_MAX_EID_LEN + 1];
316 
317 	if (nla_put_u8(skb, SMC_NLA_LGR_V2_VER, lgr->smc_version))
318 		goto errv2attr;
319 	if (nla_put_u8(skb, SMC_NLA_LGR_V2_REL, lgr->peer_smc_release))
320 		goto errv2attr;
321 	if (nla_put_u8(skb, SMC_NLA_LGR_V2_OS, lgr->peer_os))
322 		goto errv2attr;
323 	memcpy(smc_host, lgr->peer_hostname, SMC_MAX_HOSTNAME_LEN);
324 	smc_host[SMC_MAX_HOSTNAME_LEN] = 0;
325 	if (nla_put_string(skb, SMC_NLA_LGR_V2_PEER_HOST, smc_host))
326 		goto errv2attr;
327 	memcpy(smc_eid, lgr->negotiated_eid, SMC_MAX_EID_LEN);
328 	smc_eid[SMC_MAX_EID_LEN] = 0;
329 	if (nla_put_string(skb, SMC_NLA_LGR_V2_NEG_EID, smc_eid))
330 		goto errv2attr;
331 
332 	nla_nest_end(skb, v2_attrs);
333 	return 0;
334 
335 errv2attr:
336 	nla_nest_cancel(skb, v2_attrs);
337 	return -EMSGSIZE;
338 }
339 
340 static int smc_nl_fill_smcr_lgr_v2(struct smc_link_group *lgr,
341 				   struct sk_buff *skb,
342 				   struct netlink_callback *cb)
343 {
344 	struct nlattr *v2_attrs;
345 
346 	v2_attrs = nla_nest_start(skb, SMC_NLA_LGR_R_V2);
347 	if (!v2_attrs)
348 		goto errattr;
349 	if (nla_put_u8(skb, SMC_NLA_LGR_R_V2_DIRECT, !lgr->uses_gateway))
350 		goto errv2attr;
351 	if (nla_put_u8(skb, SMC_NLA_LGR_R_V2_MAX_CONNS, lgr->max_conns))
352 		goto errv2attr;
353 	if (nla_put_u8(skb, SMC_NLA_LGR_R_V2_MAX_LINKS, lgr->max_links))
354 		goto errv2attr;
355 
356 	nla_nest_end(skb, v2_attrs);
357 	return 0;
358 
359 errv2attr:
360 	nla_nest_cancel(skb, v2_attrs);
361 errattr:
362 	return -EMSGSIZE;
363 }
364 
365 static int smc_nl_fill_lgr(struct smc_link_group *lgr,
366 			   struct sk_buff *skb,
367 			   struct netlink_callback *cb)
368 {
369 	char smc_target[SMC_MAX_PNETID_LEN + 1];
370 	struct nlattr *attrs, *v2_attrs;
371 
372 	attrs = nla_nest_start(skb, SMC_GEN_LGR_SMCR);
373 	if (!attrs)
374 		goto errout;
375 
376 	if (nla_put_u32(skb, SMC_NLA_LGR_R_ID, *((u32 *)&lgr->id)))
377 		goto errattr;
378 	if (nla_put_u32(skb, SMC_NLA_LGR_R_CONNS_NUM, lgr->conns_num))
379 		goto errattr;
380 	if (nla_put_u8(skb, SMC_NLA_LGR_R_ROLE, lgr->role))
381 		goto errattr;
382 	if (nla_put_u8(skb, SMC_NLA_LGR_R_TYPE, lgr->type))
383 		goto errattr;
384 	if (nla_put_u8(skb, SMC_NLA_LGR_R_BUF_TYPE, lgr->buf_type))
385 		goto errattr;
386 	if (nla_put_u8(skb, SMC_NLA_LGR_R_VLAN_ID, lgr->vlan_id))
387 		goto errattr;
388 	if (nla_put_u64_64bit(skb, SMC_NLA_LGR_R_NET_COOKIE,
389 			      lgr->net->net_cookie, SMC_NLA_LGR_R_PAD))
390 		goto errattr;
391 	memcpy(smc_target, lgr->pnet_id, SMC_MAX_PNETID_LEN);
392 	smc_target[SMC_MAX_PNETID_LEN] = 0;
393 	if (nla_put_string(skb, SMC_NLA_LGR_R_PNETID, smc_target))
394 		goto errattr;
395 	if (nla_put_uint(skb, SMC_NLA_LGR_R_SNDBUF_ALLOC, lgr->alloc_sndbufs))
396 		goto errattr;
397 	if (nla_put_uint(skb, SMC_NLA_LGR_R_RMB_ALLOC, lgr->alloc_rmbs))
398 		goto errattr;
399 	if (lgr->smc_version > SMC_V1) {
400 		v2_attrs = nla_nest_start(skb, SMC_NLA_LGR_R_V2_COMMON);
401 		if (!v2_attrs)
402 			goto errattr;
403 		if (smc_nl_fill_lgr_v2_common(lgr, skb, cb, v2_attrs))
404 			goto errattr;
405 		if (smc_nl_fill_smcr_lgr_v2(lgr, skb, cb))
406 			goto errattr;
407 	}
408 
409 	nla_nest_end(skb, attrs);
410 	return 0;
411 errattr:
412 	nla_nest_cancel(skb, attrs);
413 errout:
414 	return -EMSGSIZE;
415 }
416 
417 static int smc_nl_fill_lgr_link(struct smc_link_group *lgr,
418 				struct smc_link *link,
419 				struct sk_buff *skb,
420 				struct netlink_callback *cb)
421 {
422 	char smc_ibname[IB_DEVICE_NAME_MAX];
423 	u8 smc_gid_target[41];
424 	struct nlattr *attrs;
425 	u32 link_uid = 0;
426 	void *nlh;
427 
428 	nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
429 			  &smc_gen_nl_family, NLM_F_MULTI,
430 			  SMC_NETLINK_GET_LINK_SMCR);
431 	if (!nlh)
432 		goto errmsg;
433 
434 	attrs = nla_nest_start(skb, SMC_GEN_LINK_SMCR);
435 	if (!attrs)
436 		goto errout;
437 
438 	if (nla_put_u8(skb, SMC_NLA_LINK_ID, link->link_id))
439 		goto errattr;
440 	if (nla_put_u32(skb, SMC_NLA_LINK_STATE, link->state))
441 		goto errattr;
442 	if (nla_put_u32(skb, SMC_NLA_LINK_CONN_CNT,
443 			atomic_read(&link->conn_cnt)))
444 		goto errattr;
445 	if (nla_put_u8(skb, SMC_NLA_LINK_IB_PORT, link->ibport))
446 		goto errattr;
447 	if (nla_put_u32(skb, SMC_NLA_LINK_NET_DEV, link->ndev_ifidx))
448 		goto errattr;
449 	snprintf(smc_ibname, sizeof(smc_ibname), "%s", link->ibname);
450 	if (nla_put_string(skb, SMC_NLA_LINK_IB_DEV, smc_ibname))
451 		goto errattr;
452 	memcpy(&link_uid, link->link_uid, sizeof(link_uid));
453 	if (nla_put_u32(skb, SMC_NLA_LINK_UID, link_uid))
454 		goto errattr;
455 	memcpy(&link_uid, link->peer_link_uid, sizeof(link_uid));
456 	if (nla_put_u32(skb, SMC_NLA_LINK_PEER_UID, link_uid))
457 		goto errattr;
458 	memset(smc_gid_target, 0, sizeof(smc_gid_target));
459 	smc_gid_be16_convert(smc_gid_target, link->gid);
460 	if (nla_put_string(skb, SMC_NLA_LINK_GID, smc_gid_target))
461 		goto errattr;
462 	memset(smc_gid_target, 0, sizeof(smc_gid_target));
463 	smc_gid_be16_convert(smc_gid_target, link->peer_gid);
464 	if (nla_put_string(skb, SMC_NLA_LINK_PEER_GID, smc_gid_target))
465 		goto errattr;
466 
467 	nla_nest_end(skb, attrs);
468 	genlmsg_end(skb, nlh);
469 	return 0;
470 errattr:
471 	nla_nest_cancel(skb, attrs);
472 errout:
473 	genlmsg_cancel(skb, nlh);
474 errmsg:
475 	return -EMSGSIZE;
476 }
477 
478 static int smc_nl_handle_lgr(struct smc_link_group *lgr,
479 			     struct sk_buff *skb,
480 			     struct netlink_callback *cb,
481 			     bool list_links)
482 {
483 	void *nlh;
484 	int i;
485 
486 	nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
487 			  &smc_gen_nl_family, NLM_F_MULTI,
488 			  SMC_NETLINK_GET_LGR_SMCR);
489 	if (!nlh)
490 		goto errmsg;
491 	if (smc_nl_fill_lgr(lgr, skb, cb))
492 		goto errout;
493 
494 	genlmsg_end(skb, nlh);
495 	if (!list_links)
496 		goto out;
497 	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
498 		if (!smc_link_usable(&lgr->lnk[i]))
499 			continue;
500 		if (smc_nl_fill_lgr_link(lgr, &lgr->lnk[i], skb, cb))
501 			goto errout;
502 	}
503 out:
504 	return 0;
505 
506 errout:
507 	genlmsg_cancel(skb, nlh);
508 errmsg:
509 	return -EMSGSIZE;
510 }
511 
512 static void smc_nl_fill_lgr_list(struct smc_lgr_list *smc_lgr,
513 				 struct sk_buff *skb,
514 				 struct netlink_callback *cb,
515 				 bool list_links)
516 {
517 	struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
518 	struct smc_link_group *lgr;
519 	int snum = cb_ctx->pos[0];
520 	int num = 0;
521 
522 	spin_lock_bh(&smc_lgr->lock);
523 	list_for_each_entry(lgr, &smc_lgr->list, list) {
524 		if (num < snum)
525 			goto next;
526 		if (smc_nl_handle_lgr(lgr, skb, cb, list_links))
527 			goto errout;
528 next:
529 		num++;
530 	}
531 errout:
532 	spin_unlock_bh(&smc_lgr->lock);
533 	cb_ctx->pos[0] = num;
534 }
535 
536 static int smc_nl_fill_smcd_lgr(struct smc_link_group *lgr,
537 				struct sk_buff *skb,
538 				struct netlink_callback *cb)
539 {
540 	char smc_pnet[SMC_MAX_PNETID_LEN + 1];
541 	struct smcd_dev *smcd = lgr->smcd;
542 	struct smcd_gid smcd_gid;
543 	struct nlattr *attrs;
544 	void *nlh;
545 
546 	nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
547 			  &smc_gen_nl_family, NLM_F_MULTI,
548 			  SMC_NETLINK_GET_LGR_SMCD);
549 	if (!nlh)
550 		goto errmsg;
551 
552 	attrs = nla_nest_start(skb, SMC_GEN_LGR_SMCD);
553 	if (!attrs)
554 		goto errout;
555 
556 	if (nla_put_u32(skb, SMC_NLA_LGR_D_ID, *((u32 *)&lgr->id)))
557 		goto errattr;
558 	smcd->ops->get_local_gid(smcd, &smcd_gid);
559 	if (nla_put_u64_64bit(skb, SMC_NLA_LGR_D_GID,
560 			      smcd_gid.gid, SMC_NLA_LGR_D_PAD))
561 		goto errattr;
562 	if (nla_put_u64_64bit(skb, SMC_NLA_LGR_D_EXT_GID,
563 			      smcd_gid.gid_ext, SMC_NLA_LGR_D_PAD))
564 		goto errattr;
565 	if (nla_put_u64_64bit(skb, SMC_NLA_LGR_D_PEER_GID, lgr->peer_gid.gid,
566 			      SMC_NLA_LGR_D_PAD))
567 		goto errattr;
568 	if (nla_put_u64_64bit(skb, SMC_NLA_LGR_D_PEER_EXT_GID,
569 			      lgr->peer_gid.gid_ext, SMC_NLA_LGR_D_PAD))
570 		goto errattr;
571 	if (nla_put_u8(skb, SMC_NLA_LGR_D_VLAN_ID, lgr->vlan_id))
572 		goto errattr;
573 	if (nla_put_u32(skb, SMC_NLA_LGR_D_CONNS_NUM, lgr->conns_num))
574 		goto errattr;
575 	if (nla_put_u32(skb, SMC_NLA_LGR_D_CHID, smc_ism_get_chid(lgr->smcd)))
576 		goto errattr;
577 	if (nla_put_uint(skb, SMC_NLA_LGR_D_SNDBUF_ALLOC, lgr->alloc_sndbufs))
578 		goto errattr;
579 	if (nla_put_uint(skb, SMC_NLA_LGR_D_DMB_ALLOC, lgr->alloc_rmbs))
580 		goto errattr;
581 	memcpy(smc_pnet, lgr->smcd->pnetid, SMC_MAX_PNETID_LEN);
582 	smc_pnet[SMC_MAX_PNETID_LEN] = 0;
583 	if (nla_put_string(skb, SMC_NLA_LGR_D_PNETID, smc_pnet))
584 		goto errattr;
585 	if (lgr->smc_version > SMC_V1) {
586 		struct nlattr *v2_attrs;
587 
588 		v2_attrs = nla_nest_start(skb, SMC_NLA_LGR_D_V2_COMMON);
589 		if (!v2_attrs)
590 			goto errattr;
591 		if (smc_nl_fill_lgr_v2_common(lgr, skb, cb, v2_attrs))
592 			goto errattr;
593 	}
594 	nla_nest_end(skb, attrs);
595 	genlmsg_end(skb, nlh);
596 	return 0;
597 
598 errattr:
599 	nla_nest_cancel(skb, attrs);
600 errout:
601 	genlmsg_cancel(skb, nlh);
602 errmsg:
603 	return -EMSGSIZE;
604 }
605 
606 static int smc_nl_handle_smcd_lgr(struct smcd_dev *dev,
607 				  struct sk_buff *skb,
608 				  struct netlink_callback *cb)
609 {
610 	struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
611 	struct smc_link_group *lgr;
612 	int snum = cb_ctx->pos[1];
613 	int rc = 0, num = 0;
614 
615 	spin_lock_bh(&dev->lgr_lock);
616 	list_for_each_entry(lgr, &dev->lgr_list, list) {
617 		if (!lgr->is_smcd)
618 			continue;
619 		if (num < snum)
620 			goto next;
621 		rc = smc_nl_fill_smcd_lgr(lgr, skb, cb);
622 		if (rc)
623 			goto errout;
624 next:
625 		num++;
626 	}
627 errout:
628 	spin_unlock_bh(&dev->lgr_lock);
629 	cb_ctx->pos[1] = num;
630 	return rc;
631 }
632 
633 static int smc_nl_fill_smcd_dev(struct smcd_dev_list *dev_list,
634 				struct sk_buff *skb,
635 				struct netlink_callback *cb)
636 {
637 	struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
638 	struct smcd_dev *smcd_dev;
639 	int snum = cb_ctx->pos[0];
640 	int rc = 0, num = 0;
641 
642 	mutex_lock(&dev_list->mutex);
643 	list_for_each_entry(smcd_dev, &dev_list->list, list) {
644 		if (list_empty(&smcd_dev->lgr_list))
645 			continue;
646 		if (num < snum)
647 			goto next;
648 		rc = smc_nl_handle_smcd_lgr(smcd_dev, skb, cb);
649 		if (rc)
650 			goto errout;
651 next:
652 		num++;
653 	}
654 errout:
655 	mutex_unlock(&dev_list->mutex);
656 	cb_ctx->pos[0] = num;
657 	return rc;
658 }
659 
660 int smcr_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb)
661 {
662 	bool list_links = false;
663 
664 	smc_nl_fill_lgr_list(&smc_lgr_list, skb, cb, list_links);
665 	return skb->len;
666 }
667 
668 int smcr_nl_get_link(struct sk_buff *skb, struct netlink_callback *cb)
669 {
670 	bool list_links = true;
671 
672 	smc_nl_fill_lgr_list(&smc_lgr_list, skb, cb, list_links);
673 	return skb->len;
674 }
675 
676 int smcd_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb)
677 {
678 	smc_nl_fill_smcd_dev(&smcd_dev_list, skb, cb);
679 	return skb->len;
680 }
681 
682 void smc_lgr_cleanup_early(struct smc_link_group *lgr)
683 {
684 	spinlock_t *lgr_lock;
685 
686 	if (!lgr)
687 		return;
688 
689 	smc_lgr_list_head(lgr, &lgr_lock);
690 	spin_lock_bh(lgr_lock);
691 	/* do not use this link group for new connections */
692 	if (!list_empty(&lgr->list))
693 		list_del_init(&lgr->list);
694 	spin_unlock_bh(lgr_lock);
695 	__smc_lgr_terminate(lgr, true);
696 }
697 
698 static void smcr_lgr_link_deactivate_all(struct smc_link_group *lgr)
699 {
700 	int i;
701 
702 	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
703 		struct smc_link *lnk = &lgr->lnk[i];
704 
705 		if (smc_link_sendable(lnk))
706 			lnk->state = SMC_LNK_INACTIVE;
707 	}
708 	wake_up_all(&lgr->llc_msg_waiter);
709 	wake_up_all(&lgr->llc_flow_waiter);
710 }
711 
712 static void smc_lgr_free(struct smc_link_group *lgr);
713 
714 static void smc_lgr_free_work(struct work_struct *work)
715 {
716 	struct smc_link_group *lgr = container_of(to_delayed_work(work),
717 						  struct smc_link_group,
718 						  free_work);
719 	spinlock_t *lgr_lock;
720 	bool conns;
721 
722 	smc_lgr_list_head(lgr, &lgr_lock);
723 	spin_lock_bh(lgr_lock);
724 	if (lgr->freeing) {
725 		spin_unlock_bh(lgr_lock);
726 		return;
727 	}
728 	read_lock_bh(&lgr->conns_lock);
729 	conns = RB_EMPTY_ROOT(&lgr->conns_all);
730 	read_unlock_bh(&lgr->conns_lock);
731 	if (!conns) { /* number of lgr connections is no longer zero */
732 		spin_unlock_bh(lgr_lock);
733 		return;
734 	}
735 	list_del_init(&lgr->list); /* remove from smc_lgr_list */
736 	lgr->freeing = 1; /* this instance does the freeing, no new schedule */
737 	spin_unlock_bh(lgr_lock);
738 	cancel_delayed_work(&lgr->free_work);
739 
740 	if (!lgr->is_smcd && !lgr->terminating)
741 		smc_llc_send_link_delete_all(lgr, true,
742 					     SMC_LLC_DEL_PROG_INIT_TERM);
743 	if (lgr->is_smcd && !lgr->terminating)
744 		smc_ism_signal_shutdown(lgr);
745 	if (!lgr->is_smcd)
746 		smcr_lgr_link_deactivate_all(lgr);
747 	smc_lgr_free(lgr);
748 }
749 
750 static void smc_lgr_terminate_work(struct work_struct *work)
751 {
752 	struct smc_link_group *lgr = container_of(work, struct smc_link_group,
753 						  terminate_work);
754 
755 	__smc_lgr_terminate(lgr, true);
756 }
757 
758 /* return next unique link id for the lgr */
759 static u8 smcr_next_link_id(struct smc_link_group *lgr)
760 {
761 	u8 link_id;
762 	int i;
763 
764 	while (1) {
765 again:
766 		link_id = ++lgr->next_link_id;
767 		if (!link_id)	/* skip zero as link_id */
768 			link_id = ++lgr->next_link_id;
769 		for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
770 			if (smc_link_usable(&lgr->lnk[i]) &&
771 			    lgr->lnk[i].link_id == link_id)
772 				goto again;
773 		}
774 		break;
775 	}
776 	return link_id;
777 }
778 
779 static void smcr_copy_dev_info_to_link(struct smc_link *link)
780 {
781 	struct smc_ib_device *smcibdev = link->smcibdev;
782 
783 	snprintf(link->ibname, sizeof(link->ibname), "%s",
784 		 smcibdev->ibdev->name);
785 	link->ndev_ifidx = smcibdev->ndev_ifidx[link->ibport - 1];
786 }
787 
788 int smcr_link_init(struct smc_link_group *lgr, struct smc_link *lnk,
789 		   u8 link_idx, struct smc_init_info *ini)
790 {
791 	struct smc_ib_device *smcibdev;
792 	u8 rndvec[3];
793 	int rc;
794 
795 	if (lgr->smc_version == SMC_V2) {
796 		lnk->smcibdev = ini->smcrv2.ib_dev_v2;
797 		lnk->ibport = ini->smcrv2.ib_port_v2;
798 	} else {
799 		lnk->smcibdev = ini->ib_dev;
800 		lnk->ibport = ini->ib_port;
801 	}
802 	get_device(&lnk->smcibdev->ibdev->dev);
803 	atomic_inc(&lnk->smcibdev->lnk_cnt);
804 	refcount_set(&lnk->refcnt, 1); /* link refcnt is set to 1 */
805 	lnk->clearing = 0;
806 	lnk->path_mtu = lnk->smcibdev->pattr[lnk->ibport - 1].active_mtu;
807 	lnk->link_id = smcr_next_link_id(lgr);
808 	lnk->lgr = lgr;
809 	smc_lgr_hold(lgr); /* lgr_put in smcr_link_clear() */
810 	lnk->link_idx = link_idx;
811 	lnk->wr_rx_id_compl = 0;
812 	smc_ibdev_cnt_inc(lnk);
813 	smcr_copy_dev_info_to_link(lnk);
814 	atomic_set(&lnk->conn_cnt, 0);
815 	smc_llc_link_set_uid(lnk);
816 	INIT_WORK(&lnk->link_down_wrk, smc_link_down_work);
817 	if (!lnk->smcibdev->initialized) {
818 		rc = (int)smc_ib_setup_per_ibdev(lnk->smcibdev);
819 		if (rc)
820 			goto out;
821 	}
822 	get_random_bytes(rndvec, sizeof(rndvec));
823 	lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) +
824 		(rndvec[2] << 16);
825 	rc = smc_ib_determine_gid(lnk->smcibdev, lnk->ibport,
826 				  ini->vlan_id, lnk->gid, &lnk->sgid_index,
827 				  lgr->smc_version == SMC_V2 ?
828 						  &ini->smcrv2 : NULL);
829 	if (rc)
830 		goto out;
831 	rc = smc_llc_link_init(lnk);
832 	if (rc)
833 		goto out;
834 	rc = smc_wr_alloc_link_mem(lnk);
835 	if (rc)
836 		goto clear_llc_lnk;
837 	rc = smc_ib_create_protection_domain(lnk);
838 	if (rc)
839 		goto free_link_mem;
840 	rc = smc_ib_create_queue_pair(lnk);
841 	if (rc)
842 		goto dealloc_pd;
843 	rc = smc_wr_create_link(lnk);
844 	if (rc)
845 		goto destroy_qp;
846 	lnk->state = SMC_LNK_ACTIVATING;
847 	return 0;
848 
849 destroy_qp:
850 	smc_ib_destroy_queue_pair(lnk);
851 dealloc_pd:
852 	smc_ib_dealloc_protection_domain(lnk);
853 free_link_mem:
854 	smc_wr_free_link_mem(lnk);
855 clear_llc_lnk:
856 	smc_llc_link_clear(lnk, false);
857 out:
858 	smc_ibdev_cnt_dec(lnk);
859 	put_device(&lnk->smcibdev->ibdev->dev);
860 	smcibdev = lnk->smcibdev;
861 	memset(lnk, 0, sizeof(struct smc_link));
862 	lnk->state = SMC_LNK_UNUSED;
863 	if (!atomic_dec_return(&smcibdev->lnk_cnt))
864 		wake_up(&smcibdev->lnks_deleted);
865 	smc_lgr_put(lgr); /* lgr_hold above */
866 	return rc;
867 }
868 
869 /* create a new SMC link group */
870 static int smc_lgr_create(struct smc_sock *smc, struct smc_init_info *ini)
871 {
872 	struct smc_link_group *lgr;
873 	struct list_head *lgr_list;
874 	struct smcd_dev *smcd;
875 	struct smc_link *lnk;
876 	spinlock_t *lgr_lock;
877 	u8 link_idx;
878 	int rc = 0;
879 	int i;
880 
881 	if (ini->is_smcd && ini->vlan_id) {
882 		if (smc_ism_get_vlan(ini->ism_dev[ini->ism_selected],
883 				     ini->vlan_id)) {
884 			rc = SMC_CLC_DECL_ISMVLANERR;
885 			goto out;
886 		}
887 	}
888 
889 	lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
890 	if (!lgr) {
891 		rc = SMC_CLC_DECL_MEM;
892 		goto ism_put_vlan;
893 	}
894 	lgr->tx_wq = alloc_workqueue("smc_tx_wq-%*phN", 0, 0,
895 				     SMC_LGR_ID_SIZE, &lgr->id);
896 	if (!lgr->tx_wq) {
897 		rc = -ENOMEM;
898 		goto free_lgr;
899 	}
900 	lgr->is_smcd = ini->is_smcd;
901 	lgr->sync_err = 0;
902 	lgr->terminating = 0;
903 	lgr->freeing = 0;
904 	lgr->vlan_id = ini->vlan_id;
905 	refcount_set(&lgr->refcnt, 1); /* set lgr refcnt to 1 */
906 	init_rwsem(&lgr->sndbufs_lock);
907 	init_rwsem(&lgr->rmbs_lock);
908 	rwlock_init(&lgr->conns_lock);
909 	for (i = 0; i < SMC_RMBE_SIZES; i++) {
910 		INIT_LIST_HEAD(&lgr->sndbufs[i]);
911 		INIT_LIST_HEAD(&lgr->rmbs[i]);
912 	}
913 	lgr->next_link_id = 0;
914 	smc_lgr_list.num += SMC_LGR_NUM_INCR;
915 	memcpy(&lgr->id, (u8 *)&smc_lgr_list.num, SMC_LGR_ID_SIZE);
916 	INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
917 	INIT_WORK(&lgr->terminate_work, smc_lgr_terminate_work);
918 	lgr->conns_all = RB_ROOT;
919 	if (ini->is_smcd) {
920 		/* SMC-D specific settings */
921 		smcd = ini->ism_dev[ini->ism_selected];
922 		get_device(smcd->ops->get_dev(smcd));
923 		lgr->peer_gid.gid =
924 			ini->ism_peer_gid[ini->ism_selected].gid;
925 		lgr->peer_gid.gid_ext =
926 			ini->ism_peer_gid[ini->ism_selected].gid_ext;
927 		lgr->smcd = ini->ism_dev[ini->ism_selected];
928 		lgr_list = &ini->ism_dev[ini->ism_selected]->lgr_list;
929 		lgr_lock = &lgr->smcd->lgr_lock;
930 		lgr->smc_version = ini->smcd_version;
931 		lgr->peer_shutdown = 0;
932 		atomic_inc(&ini->ism_dev[ini->ism_selected]->lgr_cnt);
933 	} else {
934 		/* SMC-R specific settings */
935 		struct smc_ib_device *ibdev;
936 		int ibport;
937 
938 		lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
939 		lgr->smc_version = ini->smcr_version;
940 		memcpy(lgr->peer_systemid, ini->peer_systemid,
941 		       SMC_SYSTEMID_LEN);
942 		if (lgr->smc_version == SMC_V2) {
943 			ibdev = ini->smcrv2.ib_dev_v2;
944 			ibport = ini->smcrv2.ib_port_v2;
945 			lgr->saddr = ini->smcrv2.saddr;
946 			lgr->uses_gateway = ini->smcrv2.uses_gateway;
947 			memcpy(lgr->nexthop_mac, ini->smcrv2.nexthop_mac,
948 			       ETH_ALEN);
949 			lgr->max_conns = ini->max_conns;
950 			lgr->max_links = ini->max_links;
951 		} else {
952 			ibdev = ini->ib_dev;
953 			ibport = ini->ib_port;
954 			lgr->max_conns = SMC_CONN_PER_LGR_MAX;
955 			lgr->max_links = SMC_LINKS_ADD_LNK_MAX;
956 		}
957 		memcpy(lgr->pnet_id, ibdev->pnetid[ibport - 1],
958 		       SMC_MAX_PNETID_LEN);
959 		rc = smc_wr_alloc_lgr_mem(lgr);
960 		if (rc)
961 			goto free_wq;
962 		smc_llc_lgr_init(lgr, smc);
963 
964 		link_idx = SMC_SINGLE_LINK;
965 		lnk = &lgr->lnk[link_idx];
966 		rc = smcr_link_init(lgr, lnk, link_idx, ini);
967 		if (rc) {
968 			smc_wr_free_lgr_mem(lgr);
969 			goto free_wq;
970 		}
971 		lgr->net = smc_ib_net(lnk->smcibdev);
972 		lgr_list = &smc_lgr_list.list;
973 		lgr_lock = &smc_lgr_list.lock;
974 		lgr->buf_type = lgr->net->smc.sysctl_smcr_buf_type;
975 		atomic_inc(&lgr_cnt);
976 	}
977 	smc->conn.lgr = lgr;
978 	spin_lock_bh(lgr_lock);
979 	list_add_tail(&lgr->list, lgr_list);
980 	spin_unlock_bh(lgr_lock);
981 	return 0;
982 
983 free_wq:
984 	destroy_workqueue(lgr->tx_wq);
985 free_lgr:
986 	kfree(lgr);
987 ism_put_vlan:
988 	if (ini->is_smcd && ini->vlan_id)
989 		smc_ism_put_vlan(ini->ism_dev[ini->ism_selected], ini->vlan_id);
990 out:
991 	if (rc < 0) {
992 		if (rc == -ENOMEM)
993 			rc = SMC_CLC_DECL_MEM;
994 		else
995 			rc = SMC_CLC_DECL_INTERR;
996 	}
997 	return rc;
998 }
999 
1000 static int smc_write_space(struct smc_connection *conn)
1001 {
1002 	int buffer_len = conn->peer_rmbe_size;
1003 	union smc_host_cursor prod;
1004 	union smc_host_cursor cons;
1005 	int space;
1006 
1007 	smc_curs_copy(&prod, &conn->local_tx_ctrl.prod, conn);
1008 	smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
1009 	/* determine rx_buf space */
1010 	space = buffer_len - smc_curs_diff(buffer_len, &cons, &prod);
1011 	return space;
1012 }
1013 
1014 static int smc_switch_cursor(struct smc_sock *smc, struct smc_cdc_tx_pend *pend,
1015 			     struct smc_wr_buf *wr_buf)
1016 {
1017 	struct smc_connection *conn = &smc->conn;
1018 	union smc_host_cursor cons, fin;
1019 	int rc = 0;
1020 	int diff;
1021 
1022 	smc_curs_copy(&conn->tx_curs_sent, &conn->tx_curs_fin, conn);
1023 	smc_curs_copy(&fin, &conn->local_tx_ctrl_fin, conn);
1024 	/* set prod cursor to old state, enforce tx_rdma_writes() */
1025 	smc_curs_copy(&conn->local_tx_ctrl.prod, &fin, conn);
1026 	smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
1027 
1028 	if (smc_curs_comp(conn->peer_rmbe_size, &cons, &fin) < 0) {
1029 		/* cons cursor advanced more than fin, and prod was set
1030 		 * fin above, so now prod is smaller than cons. Fix that.
1031 		 */
1032 		diff = smc_curs_diff(conn->peer_rmbe_size, &fin, &cons);
1033 		smc_curs_add(conn->sndbuf_desc->len,
1034 			     &conn->tx_curs_sent, diff);
1035 		smc_curs_add(conn->sndbuf_desc->len,
1036 			     &conn->tx_curs_fin, diff);
1037 
1038 		smp_mb__before_atomic();
1039 		atomic_add(diff, &conn->sndbuf_space);
1040 		smp_mb__after_atomic();
1041 
1042 		smc_curs_add(conn->peer_rmbe_size,
1043 			     &conn->local_tx_ctrl.prod, diff);
1044 		smc_curs_add(conn->peer_rmbe_size,
1045 			     &conn->local_tx_ctrl_fin, diff);
1046 	}
1047 	/* recalculate, value is used by tx_rdma_writes() */
1048 	atomic_set(&smc->conn.peer_rmbe_space, smc_write_space(conn));
1049 
1050 	if (smc->sk.sk_state != SMC_INIT &&
1051 	    smc->sk.sk_state != SMC_CLOSED) {
1052 		rc = smcr_cdc_msg_send_validation(conn, pend, wr_buf);
1053 		if (!rc) {
1054 			queue_delayed_work(conn->lgr->tx_wq, &conn->tx_work, 0);
1055 			smc->sk.sk_data_ready(&smc->sk);
1056 		}
1057 	} else {
1058 		smc_wr_tx_put_slot(conn->lnk,
1059 				   (struct smc_wr_tx_pend_priv *)pend);
1060 	}
1061 	return rc;
1062 }
1063 
1064 void smc_switch_link_and_count(struct smc_connection *conn,
1065 			       struct smc_link *to_lnk)
1066 {
1067 	atomic_dec(&conn->lnk->conn_cnt);
1068 	/* link_hold in smc_conn_create() */
1069 	smcr_link_put(conn->lnk);
1070 	conn->lnk = to_lnk;
1071 	atomic_inc(&conn->lnk->conn_cnt);
1072 	/* link_put in smc_conn_free() */
1073 	smcr_link_hold(conn->lnk);
1074 }
1075 
1076 struct smc_link *smc_switch_conns(struct smc_link_group *lgr,
1077 				  struct smc_link *from_lnk, bool is_dev_err)
1078 {
1079 	struct smc_link *to_lnk = NULL;
1080 	struct smc_cdc_tx_pend *pend;
1081 	struct smc_connection *conn;
1082 	struct smc_wr_buf *wr_buf;
1083 	struct smc_sock *smc;
1084 	struct rb_node *node;
1085 	int i, rc = 0;
1086 
1087 	/* link is inactive, wake up tx waiters */
1088 	smc_wr_wakeup_tx_wait(from_lnk);
1089 
1090 	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1091 		if (!smc_link_active(&lgr->lnk[i]) || i == from_lnk->link_idx)
1092 			continue;
1093 		if (is_dev_err && from_lnk->smcibdev == lgr->lnk[i].smcibdev &&
1094 		    from_lnk->ibport == lgr->lnk[i].ibport) {
1095 			continue;
1096 		}
1097 		to_lnk = &lgr->lnk[i];
1098 		break;
1099 	}
1100 	if (!to_lnk || !smc_wr_tx_link_hold(to_lnk)) {
1101 		smc_lgr_terminate_sched(lgr);
1102 		return NULL;
1103 	}
1104 again:
1105 	read_lock_bh(&lgr->conns_lock);
1106 	for (node = rb_first(&lgr->conns_all); node; node = rb_next(node)) {
1107 		conn = rb_entry(node, struct smc_connection, alert_node);
1108 		if (conn->lnk != from_lnk)
1109 			continue;
1110 		smc = container_of(conn, struct smc_sock, conn);
1111 		/* conn->lnk not yet set in SMC_INIT state */
1112 		if (smc->sk.sk_state == SMC_INIT)
1113 			continue;
1114 		if (smc->sk.sk_state == SMC_CLOSED ||
1115 		    smc->sk.sk_state == SMC_PEERCLOSEWAIT1 ||
1116 		    smc->sk.sk_state == SMC_PEERCLOSEWAIT2 ||
1117 		    smc->sk.sk_state == SMC_APPFINCLOSEWAIT ||
1118 		    smc->sk.sk_state == SMC_APPCLOSEWAIT1 ||
1119 		    smc->sk.sk_state == SMC_APPCLOSEWAIT2 ||
1120 		    smc->sk.sk_state == SMC_PEERFINCLOSEWAIT ||
1121 		    smc->sk.sk_state == SMC_PEERABORTWAIT ||
1122 		    smc->sk.sk_state == SMC_PROCESSABORT) {
1123 			spin_lock_bh(&conn->send_lock);
1124 			smc_switch_link_and_count(conn, to_lnk);
1125 			spin_unlock_bh(&conn->send_lock);
1126 			continue;
1127 		}
1128 		sock_hold(&smc->sk);
1129 		read_unlock_bh(&lgr->conns_lock);
1130 		/* pre-fetch buffer outside of send_lock, might sleep */
1131 		rc = smc_cdc_get_free_slot(conn, to_lnk, &wr_buf, NULL, &pend);
1132 		if (rc)
1133 			goto err_out;
1134 		/* avoid race with smcr_tx_sndbuf_nonempty() */
1135 		spin_lock_bh(&conn->send_lock);
1136 		smc_switch_link_and_count(conn, to_lnk);
1137 		rc = smc_switch_cursor(smc, pend, wr_buf);
1138 		spin_unlock_bh(&conn->send_lock);
1139 		sock_put(&smc->sk);
1140 		if (rc)
1141 			goto err_out;
1142 		goto again;
1143 	}
1144 	read_unlock_bh(&lgr->conns_lock);
1145 	smc_wr_tx_link_put(to_lnk);
1146 	return to_lnk;
1147 
1148 err_out:
1149 	smcr_link_down_cond_sched(to_lnk);
1150 	smc_wr_tx_link_put(to_lnk);
1151 	return NULL;
1152 }
1153 
1154 static void smcr_buf_unuse(struct smc_buf_desc *buf_desc, bool is_rmb,
1155 			   struct smc_link_group *lgr)
1156 {
1157 	struct rw_semaphore *lock;	/* lock buffer list */
1158 	int rc;
1159 
1160 	if (is_rmb && buf_desc->is_conf_rkey && !list_empty(&lgr->list)) {
1161 		/* unregister rmb with peer */
1162 		rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY);
1163 		if (!rc) {
1164 			/* protect against smc_llc_cli_rkey_exchange() */
1165 			down_read(&lgr->llc_conf_mutex);
1166 			smc_llc_do_delete_rkey(lgr, buf_desc);
1167 			buf_desc->is_conf_rkey = false;
1168 			up_read(&lgr->llc_conf_mutex);
1169 			smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl);
1170 		}
1171 	}
1172 
1173 	if (buf_desc->is_reg_err) {
1174 		/* buf registration failed, reuse not possible */
1175 		lock = is_rmb ? &lgr->rmbs_lock :
1176 				&lgr->sndbufs_lock;
1177 		down_write(lock);
1178 		smc_lgr_buf_list_del(lgr, is_rmb, buf_desc);
1179 		up_write(lock);
1180 
1181 		smc_buf_free(lgr, is_rmb, buf_desc);
1182 	} else {
1183 		/* memzero_explicit provides potential memory barrier semantics */
1184 		memzero_explicit(buf_desc->cpu_addr, buf_desc->len);
1185 		WRITE_ONCE(buf_desc->used, 0);
1186 	}
1187 }
1188 
1189 static void smcd_buf_detach(struct smc_connection *conn)
1190 {
1191 	struct smcd_dev *smcd = conn->lgr->smcd;
1192 	u64 peer_token = conn->peer_token;
1193 
1194 	if (!conn->sndbuf_desc)
1195 		return;
1196 
1197 	smc_ism_detach_dmb(smcd, peer_token);
1198 
1199 	kfree(conn->sndbuf_desc);
1200 	conn->sndbuf_desc = NULL;
1201 }
1202 
1203 static void smc_buf_unuse(struct smc_connection *conn,
1204 			  struct smc_link_group *lgr)
1205 {
1206 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1207 	bool is_smcd = lgr->is_smcd;
1208 	int bufsize;
1209 
1210 	if (conn->sndbuf_desc) {
1211 		bufsize = conn->sndbuf_desc->len;
1212 		if (!is_smcd && conn->sndbuf_desc->is_vm) {
1213 			smcr_buf_unuse(conn->sndbuf_desc, false, lgr);
1214 		} else {
1215 			memzero_explicit(conn->sndbuf_desc->cpu_addr, bufsize);
1216 			WRITE_ONCE(conn->sndbuf_desc->used, 0);
1217 		}
1218 		SMC_STAT_RMB_SIZE(smc, is_smcd, false, false, bufsize);
1219 	}
1220 	if (conn->rmb_desc) {
1221 		bufsize = conn->rmb_desc->len;
1222 		if (!is_smcd) {
1223 			smcr_buf_unuse(conn->rmb_desc, true, lgr);
1224 		} else {
1225 			bufsize += sizeof(struct smcd_cdc_msg);
1226 			memzero_explicit(conn->rmb_desc->cpu_addr, bufsize);
1227 			WRITE_ONCE(conn->rmb_desc->used, 0);
1228 		}
1229 		SMC_STAT_RMB_SIZE(smc, is_smcd, true, false, bufsize);
1230 	}
1231 }
1232 
1233 /* remove a finished connection from its link group */
1234 void smc_conn_free(struct smc_connection *conn)
1235 {
1236 	struct smc_link_group *lgr = conn->lgr;
1237 
1238 	if (!lgr || conn->freed)
1239 		/* Connection has never been registered in a
1240 		 * link group, or has already been freed.
1241 		 */
1242 		return;
1243 
1244 	conn->freed = 1;
1245 	if (!smc_conn_lgr_valid(conn))
1246 		/* Connection has already unregistered from
1247 		 * link group.
1248 		 */
1249 		goto lgr_put;
1250 
1251 	if (lgr->is_smcd) {
1252 		if (!list_empty(&lgr->list))
1253 			smc_ism_unset_conn(conn);
1254 		if (smc_ism_support_dmb_nocopy(lgr->smcd))
1255 			smcd_buf_detach(conn);
1256 		tasklet_kill(&conn->rx_tsklet);
1257 	} else {
1258 		smc_cdc_wait_pend_tx_wr(conn);
1259 		if (current_work() != &conn->abort_work)
1260 			cancel_work_sync(&conn->abort_work);
1261 	}
1262 	if (!list_empty(&lgr->list)) {
1263 		smc_buf_unuse(conn, lgr); /* allow buffer reuse */
1264 		smc_lgr_unregister_conn(conn);
1265 	}
1266 
1267 	if (!lgr->conns_num)
1268 		smc_lgr_schedule_free_work(lgr);
1269 lgr_put:
1270 	if (!lgr->is_smcd)
1271 		smcr_link_put(conn->lnk); /* link_hold in smc_conn_create() */
1272 	smc_lgr_put(lgr); /* lgr_hold in smc_conn_create() */
1273 }
1274 
1275 /* unregister a link from a buf_desc */
1276 static void smcr_buf_unmap_link(struct smc_buf_desc *buf_desc, bool is_rmb,
1277 				struct smc_link *lnk)
1278 {
1279 	if (is_rmb || buf_desc->is_vm)
1280 		buf_desc->is_reg_mr[lnk->link_idx] = false;
1281 	if (!buf_desc->is_map_ib[lnk->link_idx])
1282 		return;
1283 
1284 	if ((is_rmb || buf_desc->is_vm) &&
1285 	    buf_desc->mr[lnk->link_idx]) {
1286 		smc_ib_put_memory_region(buf_desc->mr[lnk->link_idx]);
1287 		buf_desc->mr[lnk->link_idx] = NULL;
1288 	}
1289 	if (is_rmb)
1290 		smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_FROM_DEVICE);
1291 	else
1292 		smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_TO_DEVICE);
1293 
1294 	sg_free_table(&buf_desc->sgt[lnk->link_idx]);
1295 	buf_desc->is_map_ib[lnk->link_idx] = false;
1296 }
1297 
1298 /* unmap all buffers of lgr for a deleted link */
1299 static void smcr_buf_unmap_lgr(struct smc_link *lnk)
1300 {
1301 	struct smc_link_group *lgr = lnk->lgr;
1302 	struct smc_buf_desc *buf_desc, *bf;
1303 	int i;
1304 
1305 	for (i = 0; i < SMC_RMBE_SIZES; i++) {
1306 		down_write(&lgr->rmbs_lock);
1307 		list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list)
1308 			smcr_buf_unmap_link(buf_desc, true, lnk);
1309 		up_write(&lgr->rmbs_lock);
1310 
1311 		down_write(&lgr->sndbufs_lock);
1312 		list_for_each_entry_safe(buf_desc, bf, &lgr->sndbufs[i],
1313 					 list)
1314 			smcr_buf_unmap_link(buf_desc, false, lnk);
1315 		up_write(&lgr->sndbufs_lock);
1316 	}
1317 }
1318 
1319 static void smcr_rtoken_clear_link(struct smc_link *lnk)
1320 {
1321 	struct smc_link_group *lgr = lnk->lgr;
1322 	int i;
1323 
1324 	for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1325 		lgr->rtokens[i][lnk->link_idx].rkey = 0;
1326 		lgr->rtokens[i][lnk->link_idx].dma_addr = 0;
1327 	}
1328 }
1329 
1330 static void __smcr_link_clear(struct smc_link *lnk)
1331 {
1332 	struct smc_link_group *lgr = lnk->lgr;
1333 	struct smc_ib_device *smcibdev;
1334 
1335 	smc_wr_free_link_mem(lnk);
1336 	smc_ibdev_cnt_dec(lnk);
1337 	put_device(&lnk->smcibdev->ibdev->dev);
1338 	smcibdev = lnk->smcibdev;
1339 	memset(lnk, 0, sizeof(struct smc_link));
1340 	lnk->state = SMC_LNK_UNUSED;
1341 	if (!atomic_dec_return(&smcibdev->lnk_cnt))
1342 		wake_up(&smcibdev->lnks_deleted);
1343 	smc_lgr_put(lgr); /* lgr_hold in smcr_link_init() */
1344 }
1345 
1346 /* must be called under lgr->llc_conf_mutex lock */
1347 void smcr_link_clear(struct smc_link *lnk, bool log)
1348 {
1349 	if (!lnk->lgr || lnk->clearing ||
1350 	    lnk->state == SMC_LNK_UNUSED)
1351 		return;
1352 	lnk->clearing = 1;
1353 	lnk->peer_qpn = 0;
1354 	smc_llc_link_clear(lnk, log);
1355 	smcr_buf_unmap_lgr(lnk);
1356 	smcr_rtoken_clear_link(lnk);
1357 	smc_ib_modify_qp_error(lnk);
1358 	smc_wr_free_link(lnk);
1359 	smc_ib_destroy_queue_pair(lnk);
1360 	smc_ib_dealloc_protection_domain(lnk);
1361 	smcr_link_put(lnk); /* theoretically last link_put */
1362 }
1363 
1364 void smcr_link_hold(struct smc_link *lnk)
1365 {
1366 	refcount_inc(&lnk->refcnt);
1367 }
1368 
1369 void smcr_link_put(struct smc_link *lnk)
1370 {
1371 	if (refcount_dec_and_test(&lnk->refcnt))
1372 		__smcr_link_clear(lnk);
1373 }
1374 
1375 static void smcr_buf_free(struct smc_link_group *lgr, bool is_rmb,
1376 			  struct smc_buf_desc *buf_desc)
1377 {
1378 	int i;
1379 
1380 	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
1381 		smcr_buf_unmap_link(buf_desc, is_rmb, &lgr->lnk[i]);
1382 
1383 	if (!buf_desc->is_vm && buf_desc->pages)
1384 		__free_pages(buf_desc->pages, buf_desc->order);
1385 	else if (buf_desc->is_vm && buf_desc->cpu_addr)
1386 		vfree(buf_desc->cpu_addr);
1387 	kfree(buf_desc);
1388 }
1389 
1390 static void smcd_buf_free(struct smc_link_group *lgr, bool is_dmb,
1391 			  struct smc_buf_desc *buf_desc)
1392 {
1393 	if (is_dmb) {
1394 		/* restore original buf len */
1395 		buf_desc->len += sizeof(struct smcd_cdc_msg);
1396 		smc_ism_unregister_dmb(lgr->smcd, buf_desc);
1397 	} else {
1398 		kfree(buf_desc->cpu_addr);
1399 	}
1400 	kfree(buf_desc);
1401 }
1402 
1403 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
1404 			 struct smc_buf_desc *buf_desc)
1405 {
1406 	if (lgr->is_smcd)
1407 		smcd_buf_free(lgr, is_rmb, buf_desc);
1408 	else
1409 		smcr_buf_free(lgr, is_rmb, buf_desc);
1410 }
1411 
1412 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
1413 {
1414 	struct smc_buf_desc *buf_desc, *bf_desc;
1415 	struct list_head *buf_list;
1416 	int i;
1417 
1418 	for (i = 0; i < SMC_RMBE_SIZES; i++) {
1419 		if (is_rmb)
1420 			buf_list = &lgr->rmbs[i];
1421 		else
1422 			buf_list = &lgr->sndbufs[i];
1423 		list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
1424 					 list) {
1425 			smc_lgr_buf_list_del(lgr, is_rmb, buf_desc);
1426 			smc_buf_free(lgr, is_rmb, buf_desc);
1427 		}
1428 	}
1429 }
1430 
1431 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
1432 {
1433 	/* free send buffers */
1434 	__smc_lgr_free_bufs(lgr, false);
1435 	/* free rmbs */
1436 	__smc_lgr_free_bufs(lgr, true);
1437 }
1438 
1439 /* won't be freed until no one accesses to lgr anymore */
1440 static void __smc_lgr_free(struct smc_link_group *lgr)
1441 {
1442 	smc_lgr_free_bufs(lgr);
1443 	if (lgr->is_smcd) {
1444 		if (!atomic_dec_return(&lgr->smcd->lgr_cnt))
1445 			wake_up(&lgr->smcd->lgrs_deleted);
1446 	} else {
1447 		smc_wr_free_lgr_mem(lgr);
1448 		if (!atomic_dec_return(&lgr_cnt))
1449 			wake_up(&lgrs_deleted);
1450 	}
1451 	kfree(lgr);
1452 }
1453 
1454 /* remove a link group */
1455 static void smc_lgr_free(struct smc_link_group *lgr)
1456 {
1457 	int i;
1458 
1459 	if (!lgr->is_smcd) {
1460 		down_write(&lgr->llc_conf_mutex);
1461 		for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1462 			if (lgr->lnk[i].state != SMC_LNK_UNUSED)
1463 				smcr_link_clear(&lgr->lnk[i], false);
1464 		}
1465 		up_write(&lgr->llc_conf_mutex);
1466 		smc_llc_lgr_clear(lgr);
1467 	}
1468 
1469 	destroy_workqueue(lgr->tx_wq);
1470 	if (lgr->is_smcd) {
1471 		smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
1472 		put_device(lgr->smcd->ops->get_dev(lgr->smcd));
1473 	}
1474 	smc_lgr_put(lgr); /* theoretically last lgr_put */
1475 }
1476 
1477 void smc_lgr_hold(struct smc_link_group *lgr)
1478 {
1479 	refcount_inc(&lgr->refcnt);
1480 }
1481 
1482 void smc_lgr_put(struct smc_link_group *lgr)
1483 {
1484 	if (refcount_dec_and_test(&lgr->refcnt))
1485 		__smc_lgr_free(lgr);
1486 }
1487 
1488 static void smc_sk_wake_ups(struct smc_sock *smc)
1489 {
1490 	smc->sk.sk_write_space(&smc->sk);
1491 	smc->sk.sk_data_ready(&smc->sk);
1492 	smc->sk.sk_state_change(&smc->sk);
1493 }
1494 
1495 /* kill a connection */
1496 static void smc_conn_kill(struct smc_connection *conn, bool soft)
1497 {
1498 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1499 
1500 	if (conn->lgr->is_smcd && conn->lgr->peer_shutdown)
1501 		conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
1502 	else
1503 		smc_close_abort(conn);
1504 	conn->killed = 1;
1505 	smc->sk.sk_err = ECONNABORTED;
1506 	smc_sk_wake_ups(smc);
1507 	if (conn->lgr->is_smcd) {
1508 		smc_ism_unset_conn(conn);
1509 		if (smc_ism_support_dmb_nocopy(conn->lgr->smcd))
1510 			smcd_buf_detach(conn);
1511 		if (soft)
1512 			tasklet_kill(&conn->rx_tsklet);
1513 		else
1514 			tasklet_unlock_wait(&conn->rx_tsklet);
1515 	} else {
1516 		smc_cdc_wait_pend_tx_wr(conn);
1517 	}
1518 	smc_lgr_unregister_conn(conn);
1519 	smc_close_active_abort(smc);
1520 }
1521 
1522 static void smc_lgr_cleanup(struct smc_link_group *lgr)
1523 {
1524 	if (lgr->is_smcd) {
1525 		smc_ism_signal_shutdown(lgr);
1526 	} else {
1527 		u32 rsn = lgr->llc_termination_rsn;
1528 
1529 		if (!rsn)
1530 			rsn = SMC_LLC_DEL_PROG_INIT_TERM;
1531 		smc_llc_send_link_delete_all(lgr, false, rsn);
1532 		smcr_lgr_link_deactivate_all(lgr);
1533 	}
1534 }
1535 
1536 /* terminate link group
1537  * @soft: true if link group shutdown can take its time
1538  *	  false if immediate link group shutdown is required
1539  */
1540 static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft)
1541 {
1542 	struct smc_connection *conn;
1543 	struct smc_sock *smc;
1544 	struct rb_node *node;
1545 
1546 	if (lgr->terminating)
1547 		return;	/* lgr already terminating */
1548 	/* cancel free_work sync, will terminate when lgr->freeing is set */
1549 	cancel_delayed_work(&lgr->free_work);
1550 	lgr->terminating = 1;
1551 
1552 	/* kill remaining link group connections */
1553 	read_lock_bh(&lgr->conns_lock);
1554 	node = rb_first(&lgr->conns_all);
1555 	while (node) {
1556 		read_unlock_bh(&lgr->conns_lock);
1557 		conn = rb_entry(node, struct smc_connection, alert_node);
1558 		smc = container_of(conn, struct smc_sock, conn);
1559 		sock_hold(&smc->sk); /* sock_put below */
1560 		lock_sock(&smc->sk);
1561 		smc_conn_kill(conn, soft);
1562 		release_sock(&smc->sk);
1563 		sock_put(&smc->sk); /* sock_hold above */
1564 		read_lock_bh(&lgr->conns_lock);
1565 		node = rb_first(&lgr->conns_all);
1566 	}
1567 	read_unlock_bh(&lgr->conns_lock);
1568 	smc_lgr_cleanup(lgr);
1569 	smc_lgr_free(lgr);
1570 }
1571 
1572 /* unlink link group and schedule termination */
1573 void smc_lgr_terminate_sched(struct smc_link_group *lgr)
1574 {
1575 	spinlock_t *lgr_lock;
1576 
1577 	smc_lgr_list_head(lgr, &lgr_lock);
1578 	spin_lock_bh(lgr_lock);
1579 	if (list_empty(&lgr->list) || lgr->terminating || lgr->freeing) {
1580 		spin_unlock_bh(lgr_lock);
1581 		return;	/* lgr already terminating */
1582 	}
1583 	list_del_init(&lgr->list);
1584 	lgr->freeing = 1;
1585 	spin_unlock_bh(lgr_lock);
1586 	schedule_work(&lgr->terminate_work);
1587 }
1588 
1589 /* Called when peer lgr shutdown (regularly or abnormally) is received */
1590 void smc_smcd_terminate(struct smcd_dev *dev, struct smcd_gid *peer_gid,
1591 			unsigned short vlan)
1592 {
1593 	struct smc_link_group *lgr, *l;
1594 	LIST_HEAD(lgr_free_list);
1595 
1596 	/* run common cleanup function and build free list */
1597 	spin_lock_bh(&dev->lgr_lock);
1598 	list_for_each_entry_safe(lgr, l, &dev->lgr_list, list) {
1599 		if ((!peer_gid->gid ||
1600 		     (lgr->peer_gid.gid == peer_gid->gid &&
1601 		      !smc_ism_is_emulated(dev) ? 1 :
1602 		      lgr->peer_gid.gid_ext == peer_gid->gid_ext)) &&
1603 		    (vlan == VLAN_VID_MASK || lgr->vlan_id == vlan)) {
1604 			if (peer_gid->gid) /* peer triggered termination */
1605 				lgr->peer_shutdown = 1;
1606 			list_move(&lgr->list, &lgr_free_list);
1607 			lgr->freeing = 1;
1608 		}
1609 	}
1610 	spin_unlock_bh(&dev->lgr_lock);
1611 
1612 	/* cancel the regular free workers and actually free lgrs */
1613 	list_for_each_entry_safe(lgr, l, &lgr_free_list, list) {
1614 		list_del_init(&lgr->list);
1615 		schedule_work(&lgr->terminate_work);
1616 	}
1617 }
1618 
1619 /* Called when an SMCD device is removed or the smc module is unloaded */
1620 void smc_smcd_terminate_all(struct smcd_dev *smcd)
1621 {
1622 	struct smc_link_group *lgr, *lg;
1623 	LIST_HEAD(lgr_free_list);
1624 
1625 	spin_lock_bh(&smcd->lgr_lock);
1626 	list_splice_init(&smcd->lgr_list, &lgr_free_list);
1627 	list_for_each_entry(lgr, &lgr_free_list, list)
1628 		lgr->freeing = 1;
1629 	spin_unlock_bh(&smcd->lgr_lock);
1630 
1631 	list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
1632 		list_del_init(&lgr->list);
1633 		__smc_lgr_terminate(lgr, false);
1634 	}
1635 
1636 	if (atomic_read(&smcd->lgr_cnt))
1637 		wait_event(smcd->lgrs_deleted, !atomic_read(&smcd->lgr_cnt));
1638 }
1639 
1640 /* Called when an SMCR device is removed or the smc module is unloaded.
1641  * If smcibdev is given, all SMCR link groups using this device are terminated.
1642  * If smcibdev is NULL, all SMCR link groups are terminated.
1643  */
1644 void smc_smcr_terminate_all(struct smc_ib_device *smcibdev)
1645 {
1646 	struct smc_link_group *lgr, *lg;
1647 	LIST_HEAD(lgr_free_list);
1648 	int i;
1649 
1650 	spin_lock_bh(&smc_lgr_list.lock);
1651 	if (!smcibdev) {
1652 		list_splice_init(&smc_lgr_list.list, &lgr_free_list);
1653 		list_for_each_entry(lgr, &lgr_free_list, list)
1654 			lgr->freeing = 1;
1655 	} else {
1656 		list_for_each_entry_safe(lgr, lg, &smc_lgr_list.list, list) {
1657 			for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1658 				if (lgr->lnk[i].smcibdev == smcibdev)
1659 					smcr_link_down_cond_sched(&lgr->lnk[i]);
1660 			}
1661 		}
1662 	}
1663 	spin_unlock_bh(&smc_lgr_list.lock);
1664 
1665 	list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
1666 		list_del_init(&lgr->list);
1667 		smc_llc_set_termination_rsn(lgr, SMC_LLC_DEL_OP_INIT_TERM);
1668 		__smc_lgr_terminate(lgr, false);
1669 	}
1670 
1671 	if (smcibdev) {
1672 		if (atomic_read(&smcibdev->lnk_cnt))
1673 			wait_event(smcibdev->lnks_deleted,
1674 				   !atomic_read(&smcibdev->lnk_cnt));
1675 	} else {
1676 		if (atomic_read(&lgr_cnt))
1677 			wait_event(lgrs_deleted, !atomic_read(&lgr_cnt));
1678 	}
1679 }
1680 
1681 /* set new lgr type and clear all asymmetric link tagging */
1682 void smcr_lgr_set_type(struct smc_link_group *lgr, enum smc_lgr_type new_type)
1683 {
1684 	char *lgr_type = "";
1685 	int i;
1686 
1687 	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
1688 		if (smc_link_usable(&lgr->lnk[i]))
1689 			lgr->lnk[i].link_is_asym = false;
1690 	if (lgr->type == new_type)
1691 		return;
1692 	lgr->type = new_type;
1693 
1694 	switch (lgr->type) {
1695 	case SMC_LGR_NONE:
1696 		lgr_type = "NONE";
1697 		break;
1698 	case SMC_LGR_SINGLE:
1699 		lgr_type = "SINGLE";
1700 		break;
1701 	case SMC_LGR_SYMMETRIC:
1702 		lgr_type = "SYMMETRIC";
1703 		break;
1704 	case SMC_LGR_ASYMMETRIC_PEER:
1705 		lgr_type = "ASYMMETRIC_PEER";
1706 		break;
1707 	case SMC_LGR_ASYMMETRIC_LOCAL:
1708 		lgr_type = "ASYMMETRIC_LOCAL";
1709 		break;
1710 	}
1711 	pr_warn_ratelimited("smc: SMC-R lg %*phN net %llu state changed: "
1712 			    "%s, pnetid %.16s\n", SMC_LGR_ID_SIZE, &lgr->id,
1713 			    lgr->net->net_cookie, lgr_type, lgr->pnet_id);
1714 }
1715 
1716 /* set new lgr type and tag a link as asymmetric */
1717 void smcr_lgr_set_type_asym(struct smc_link_group *lgr,
1718 			    enum smc_lgr_type new_type, int asym_lnk_idx)
1719 {
1720 	smcr_lgr_set_type(lgr, new_type);
1721 	lgr->lnk[asym_lnk_idx].link_is_asym = true;
1722 }
1723 
1724 /* abort connection, abort_work scheduled from tasklet context */
1725 static void smc_conn_abort_work(struct work_struct *work)
1726 {
1727 	struct smc_connection *conn = container_of(work,
1728 						   struct smc_connection,
1729 						   abort_work);
1730 	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1731 
1732 	lock_sock(&smc->sk);
1733 	smc_conn_kill(conn, true);
1734 	release_sock(&smc->sk);
1735 	sock_put(&smc->sk); /* sock_hold done by schedulers of abort_work */
1736 }
1737 
1738 void smcr_port_add(struct smc_ib_device *smcibdev, u8 ibport)
1739 {
1740 	struct smc_link_group *lgr, *n;
1741 
1742 	spin_lock_bh(&smc_lgr_list.lock);
1743 	list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1744 		struct smc_link *link;
1745 
1746 		if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1747 			    SMC_MAX_PNETID_LEN) ||
1748 		    lgr->type == SMC_LGR_SYMMETRIC ||
1749 		    lgr->type == SMC_LGR_ASYMMETRIC_PEER ||
1750 		    !rdma_dev_access_netns(smcibdev->ibdev, lgr->net))
1751 			continue;
1752 
1753 		if (lgr->type == SMC_LGR_SINGLE && lgr->max_links <= 1)
1754 			continue;
1755 
1756 		/* trigger local add link processing */
1757 		link = smc_llc_usable_link(lgr);
1758 		if (link)
1759 			smc_llc_add_link_local(link);
1760 	}
1761 	spin_unlock_bh(&smc_lgr_list.lock);
1762 }
1763 
1764 /* link is down - switch connections to alternate link,
1765  * must be called under lgr->llc_conf_mutex lock
1766  */
1767 static void smcr_link_down(struct smc_link *lnk)
1768 {
1769 	struct smc_link_group *lgr = lnk->lgr;
1770 	struct smc_link *to_lnk;
1771 	int del_link_id;
1772 
1773 	if (!lgr || lnk->state == SMC_LNK_UNUSED || list_empty(&lgr->list))
1774 		return;
1775 
1776 	to_lnk = smc_switch_conns(lgr, lnk, true);
1777 	if (!to_lnk) { /* no backup link available */
1778 		smcr_link_clear(lnk, true);
1779 		return;
1780 	}
1781 	smcr_lgr_set_type(lgr, SMC_LGR_SINGLE);
1782 	del_link_id = lnk->link_id;
1783 
1784 	if (lgr->role == SMC_SERV) {
1785 		/* trigger local delete link processing */
1786 		smc_llc_srv_delete_link_local(to_lnk, del_link_id);
1787 	} else {
1788 		if (lgr->llc_flow_lcl.type != SMC_LLC_FLOW_NONE) {
1789 			/* another llc task is ongoing */
1790 			up_write(&lgr->llc_conf_mutex);
1791 			wait_event_timeout(lgr->llc_flow_waiter,
1792 				(list_empty(&lgr->list) ||
1793 				 lgr->llc_flow_lcl.type == SMC_LLC_FLOW_NONE),
1794 				SMC_LLC_WAIT_TIME);
1795 			down_write(&lgr->llc_conf_mutex);
1796 		}
1797 		if (!list_empty(&lgr->list)) {
1798 			smc_llc_send_delete_link(to_lnk, del_link_id,
1799 						 SMC_LLC_REQ, true,
1800 						 SMC_LLC_DEL_LOST_PATH);
1801 			smcr_link_clear(lnk, true);
1802 		}
1803 		wake_up(&lgr->llc_flow_waiter);	/* wake up next waiter */
1804 	}
1805 }
1806 
1807 /* must be called under lgr->llc_conf_mutex lock */
1808 void smcr_link_down_cond(struct smc_link *lnk)
1809 {
1810 	if (smc_link_downing(&lnk->state)) {
1811 		trace_smcr_link_down(lnk, __builtin_return_address(0));
1812 		smcr_link_down(lnk);
1813 	}
1814 }
1815 
1816 /* will get the lgr->llc_conf_mutex lock */
1817 void smcr_link_down_cond_sched(struct smc_link *lnk)
1818 {
1819 	if (smc_link_downing(&lnk->state)) {
1820 		trace_smcr_link_down(lnk, __builtin_return_address(0));
1821 		schedule_work(&lnk->link_down_wrk);
1822 	}
1823 }
1824 
1825 void smcr_port_err(struct smc_ib_device *smcibdev, u8 ibport)
1826 {
1827 	struct smc_link_group *lgr, *n;
1828 	int i;
1829 
1830 	list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1831 		if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1832 			    SMC_MAX_PNETID_LEN))
1833 			continue; /* lgr is not affected */
1834 		if (list_empty(&lgr->list))
1835 			continue;
1836 		for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1837 			struct smc_link *lnk = &lgr->lnk[i];
1838 
1839 			if (smc_link_usable(lnk) &&
1840 			    lnk->smcibdev == smcibdev && lnk->ibport == ibport)
1841 				smcr_link_down_cond_sched(lnk);
1842 		}
1843 	}
1844 }
1845 
1846 static void smc_link_down_work(struct work_struct *work)
1847 {
1848 	struct smc_link *link = container_of(work, struct smc_link,
1849 					     link_down_wrk);
1850 	struct smc_link_group *lgr = link->lgr;
1851 
1852 	if (list_empty(&lgr->list))
1853 		return;
1854 	wake_up_all(&lgr->llc_msg_waiter);
1855 	down_write(&lgr->llc_conf_mutex);
1856 	smcr_link_down(link);
1857 	up_write(&lgr->llc_conf_mutex);
1858 }
1859 
1860 static int smc_vlan_by_tcpsk_walk(struct net_device *lower_dev,
1861 				  struct netdev_nested_priv *priv)
1862 {
1863 	unsigned short *vlan_id = (unsigned short *)priv->data;
1864 
1865 	if (is_vlan_dev(lower_dev)) {
1866 		*vlan_id = vlan_dev_vlan_id(lower_dev);
1867 		return 1;
1868 	}
1869 
1870 	return 0;
1871 }
1872 
1873 /* Determine vlan of internal TCP socket. */
1874 int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini)
1875 {
1876 	struct dst_entry *dst = sk_dst_get(clcsock->sk);
1877 	struct netdev_nested_priv priv;
1878 	struct net_device *ndev;
1879 	int rc = 0;
1880 
1881 	ini->vlan_id = 0;
1882 	if (!dst) {
1883 		rc = -ENOTCONN;
1884 		goto out;
1885 	}
1886 	if (!dst->dev) {
1887 		rc = -ENODEV;
1888 		goto out_rel;
1889 	}
1890 
1891 	ndev = dst->dev;
1892 	if (is_vlan_dev(ndev)) {
1893 		ini->vlan_id = vlan_dev_vlan_id(ndev);
1894 		goto out_rel;
1895 	}
1896 
1897 	priv.data = (void *)&ini->vlan_id;
1898 	rtnl_lock();
1899 	netdev_walk_all_lower_dev(ndev, smc_vlan_by_tcpsk_walk, &priv);
1900 	rtnl_unlock();
1901 
1902 out_rel:
1903 	dst_release(dst);
1904 out:
1905 	return rc;
1906 }
1907 
1908 static bool smcr_lgr_match(struct smc_link_group *lgr, u8 smcr_version,
1909 			   u8 peer_systemid[],
1910 			   u8 peer_gid[],
1911 			   u8 peer_mac_v1[],
1912 			   enum smc_lgr_role role, u32 clcqpn,
1913 			   struct net *net)
1914 {
1915 	struct smc_link *lnk;
1916 	int i;
1917 
1918 	if (memcmp(lgr->peer_systemid, peer_systemid, SMC_SYSTEMID_LEN) ||
1919 	    lgr->role != role)
1920 		return false;
1921 
1922 	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1923 		lnk = &lgr->lnk[i];
1924 
1925 		if (!smc_link_active(lnk))
1926 			continue;
1927 		/* use verbs API to check netns, instead of lgr->net */
1928 		if (!rdma_dev_access_netns(lnk->smcibdev->ibdev, net))
1929 			return false;
1930 		if ((lgr->role == SMC_SERV || lnk->peer_qpn == clcqpn) &&
1931 		    !memcmp(lnk->peer_gid, peer_gid, SMC_GID_SIZE) &&
1932 		    (smcr_version == SMC_V2 ||
1933 		     !memcmp(lnk->peer_mac, peer_mac_v1, ETH_ALEN)))
1934 			return true;
1935 	}
1936 	return false;
1937 }
1938 
1939 static bool smcd_lgr_match(struct smc_link_group *lgr,
1940 			   struct smcd_dev *smcismdev,
1941 			   struct smcd_gid *peer_gid)
1942 {
1943 	if (lgr->peer_gid.gid != peer_gid->gid ||
1944 	    lgr->smcd != smcismdev)
1945 		return false;
1946 
1947 	if (smc_ism_is_emulated(smcismdev) &&
1948 	    lgr->peer_gid.gid_ext != peer_gid->gid_ext)
1949 		return false;
1950 
1951 	return true;
1952 }
1953 
1954 /* create a new SMC connection (and a new link group if necessary) */
1955 int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini)
1956 {
1957 	struct smc_connection *conn = &smc->conn;
1958 	struct net *net = sock_net(&smc->sk);
1959 	struct list_head *lgr_list;
1960 	struct smc_link_group *lgr;
1961 	enum smc_lgr_role role;
1962 	spinlock_t *lgr_lock;
1963 	int rc = 0;
1964 
1965 	lgr_list = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_list :
1966 				  &smc_lgr_list.list;
1967 	lgr_lock = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_lock :
1968 				  &smc_lgr_list.lock;
1969 	ini->first_contact_local = 1;
1970 	role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
1971 	if (role == SMC_CLNT && ini->first_contact_peer)
1972 		/* create new link group as well */
1973 		goto create;
1974 
1975 	/* determine if an existing link group can be reused */
1976 	spin_lock_bh(lgr_lock);
1977 	list_for_each_entry(lgr, lgr_list, list) {
1978 		write_lock_bh(&lgr->conns_lock);
1979 		if ((ini->is_smcd ?
1980 		     smcd_lgr_match(lgr, ini->ism_dev[ini->ism_selected],
1981 				    &ini->ism_peer_gid[ini->ism_selected]) :
1982 		     smcr_lgr_match(lgr, ini->smcr_version,
1983 				    ini->peer_systemid,
1984 				    ini->peer_gid, ini->peer_mac, role,
1985 				    ini->ib_clcqpn, net)) &&
1986 		    !lgr->sync_err &&
1987 		    (ini->smcd_version == SMC_V2 ||
1988 		     lgr->vlan_id == ini->vlan_id) &&
1989 		    (role == SMC_CLNT || ini->is_smcd ||
1990 		    (lgr->conns_num < lgr->max_conns &&
1991 		      !bitmap_full(lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX)))) {
1992 			/* link group found */
1993 			ini->first_contact_local = 0;
1994 			conn->lgr = lgr;
1995 			rc = smc_lgr_register_conn(conn, false);
1996 			write_unlock_bh(&lgr->conns_lock);
1997 			if (!rc && delayed_work_pending(&lgr->free_work))
1998 				cancel_delayed_work(&lgr->free_work);
1999 			break;
2000 		}
2001 		write_unlock_bh(&lgr->conns_lock);
2002 	}
2003 	spin_unlock_bh(lgr_lock);
2004 	if (rc)
2005 		return rc;
2006 
2007 	if (role == SMC_CLNT && !ini->first_contact_peer &&
2008 	    ini->first_contact_local) {
2009 		/* Server reuses a link group, but Client wants to start
2010 		 * a new one
2011 		 * send out_of_sync decline, reason synchr. error
2012 		 */
2013 		return SMC_CLC_DECL_SYNCERR;
2014 	}
2015 
2016 create:
2017 	if (ini->first_contact_local) {
2018 		rc = smc_lgr_create(smc, ini);
2019 		if (rc)
2020 			goto out;
2021 		lgr = conn->lgr;
2022 		write_lock_bh(&lgr->conns_lock);
2023 		rc = smc_lgr_register_conn(conn, true);
2024 		write_unlock_bh(&lgr->conns_lock);
2025 		if (rc) {
2026 			smc_lgr_cleanup_early(lgr);
2027 			goto out;
2028 		}
2029 	}
2030 	smc_lgr_hold(conn->lgr); /* lgr_put in smc_conn_free() */
2031 	if (!conn->lgr->is_smcd)
2032 		smcr_link_hold(conn->lnk); /* link_put in smc_conn_free() */
2033 	conn->freed = 0;
2034 	conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
2035 	conn->local_tx_ctrl.len = SMC_WR_TX_SIZE;
2036 	conn->urg_state = SMC_URG_READ;
2037 	init_waitqueue_head(&conn->cdc_pend_tx_wq);
2038 	INIT_WORK(&smc->conn.abort_work, smc_conn_abort_work);
2039 	if (ini->is_smcd) {
2040 		conn->rx_off = sizeof(struct smcd_cdc_msg);
2041 		smcd_cdc_rx_init(conn); /* init tasklet for this conn */
2042 	} else {
2043 		conn->rx_off = 0;
2044 	}
2045 #ifndef KERNEL_HAS_ATOMIC64
2046 	spin_lock_init(&conn->acurs_lock);
2047 #endif
2048 
2049 out:
2050 	return rc;
2051 }
2052 
2053 #define SMCD_DMBE_SIZES		6 /* 0 -> 16KB, 1 -> 32KB, .. 6 -> 1MB */
2054 #define SMCR_RMBE_SIZES		15 /* 0 -> 16KB, 1 -> 32KB, .. 15 -> 512MB */
2055 
2056 /* convert the RMB size into the compressed notation (minimum 16K, see
2057  * SMCD/R_DMBE_SIZES.
2058  * In contrast to plain ilog2, this rounds towards the next power of 2,
2059  * so the socket application gets at least its desired sndbuf / rcvbuf size.
2060  */
2061 static u8 smc_compress_bufsize(int size, bool is_smcd, bool is_rmb)
2062 {
2063 	u8 compressed;
2064 
2065 	if (size <= SMC_BUF_MIN_SIZE)
2066 		return 0;
2067 
2068 	size = (size - 1) >> 14;  /* convert to 16K multiple */
2069 	compressed = min_t(u8, ilog2(size) + 1,
2070 			   is_smcd ? SMCD_DMBE_SIZES : SMCR_RMBE_SIZES);
2071 
2072 #ifdef CONFIG_ARCH_NO_SG_CHAIN
2073 	if (!is_smcd && is_rmb)
2074 		/* RMBs are backed by & limited to max size of scatterlists */
2075 		compressed = min_t(u8, compressed, ilog2((SG_MAX_SINGLE_ALLOC * PAGE_SIZE) >> 14));
2076 #endif
2077 
2078 	return compressed;
2079 }
2080 
2081 /* convert the RMB size from compressed notation into integer */
2082 int smc_uncompress_bufsize(u8 compressed)
2083 {
2084 	u32 size;
2085 
2086 	size = 0x00000001 << (((int)compressed) + 14);
2087 	return (int)size;
2088 }
2089 
2090 /* try to reuse a sndbuf or rmb description slot for a certain
2091  * buffer size; if not available, return NULL
2092  */
2093 static struct smc_buf_desc *smc_buf_get_slot(int compressed_bufsize,
2094 					     struct rw_semaphore *lock,
2095 					     struct list_head *buf_list)
2096 {
2097 	struct smc_buf_desc *buf_slot;
2098 
2099 	down_read(lock);
2100 	list_for_each_entry(buf_slot, buf_list, list) {
2101 		if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
2102 			up_read(lock);
2103 			return buf_slot;
2104 		}
2105 	}
2106 	up_read(lock);
2107 	return NULL;
2108 }
2109 
2110 /* one of the conditions for announcing a receiver's current window size is
2111  * that it "results in a minimum increase in the window size of 10% of the
2112  * receive buffer space" [RFC7609]
2113  */
2114 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
2115 {
2116 	return max_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
2117 }
2118 
2119 /* map an buf to a link */
2120 static int smcr_buf_map_link(struct smc_buf_desc *buf_desc, bool is_rmb,
2121 			     struct smc_link *lnk)
2122 {
2123 	int rc, i, nents, offset, buf_size, size, access_flags;
2124 	struct scatterlist *sg;
2125 	void *buf;
2126 
2127 	if (buf_desc->is_map_ib[lnk->link_idx])
2128 		return 0;
2129 
2130 	if (buf_desc->is_vm) {
2131 		buf = buf_desc->cpu_addr;
2132 		buf_size = buf_desc->len;
2133 		offset = offset_in_page(buf_desc->cpu_addr);
2134 		nents = PAGE_ALIGN(buf_size + offset) / PAGE_SIZE;
2135 	} else {
2136 		nents = 1;
2137 	}
2138 
2139 	rc = sg_alloc_table(&buf_desc->sgt[lnk->link_idx], nents, GFP_KERNEL);
2140 	if (rc)
2141 		return rc;
2142 
2143 	if (buf_desc->is_vm) {
2144 		/* virtually contiguous buffer */
2145 		for_each_sg(buf_desc->sgt[lnk->link_idx].sgl, sg, nents, i) {
2146 			size = min_t(int, PAGE_SIZE - offset, buf_size);
2147 			sg_set_page(sg, vmalloc_to_page(buf), size, offset);
2148 			buf += size / sizeof(*buf);
2149 			buf_size -= size;
2150 			offset = 0;
2151 		}
2152 	} else {
2153 		/* physically contiguous buffer */
2154 		sg_set_buf(buf_desc->sgt[lnk->link_idx].sgl,
2155 			   buf_desc->cpu_addr, buf_desc->len);
2156 	}
2157 
2158 	/* map sg table to DMA address */
2159 	rc = smc_ib_buf_map_sg(lnk, buf_desc,
2160 			       is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
2161 	/* SMC protocol depends on mapping to one DMA address only */
2162 	if (rc != nents) {
2163 		rc = -EAGAIN;
2164 		goto free_table;
2165 	}
2166 
2167 	buf_desc->is_dma_need_sync |=
2168 		smc_ib_is_sg_need_sync(lnk, buf_desc) << lnk->link_idx;
2169 
2170 	if (is_rmb || buf_desc->is_vm) {
2171 		/* create a new memory region for the RMB or vzalloced sndbuf */
2172 		access_flags = is_rmb ?
2173 			       IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE :
2174 			       IB_ACCESS_LOCAL_WRITE;
2175 
2176 		rc = smc_ib_get_memory_region(lnk->roce_pd, access_flags,
2177 					      buf_desc, lnk->link_idx);
2178 		if (rc)
2179 			goto buf_unmap;
2180 		smc_ib_sync_sg_for_device(lnk, buf_desc,
2181 					  is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
2182 	}
2183 	buf_desc->is_map_ib[lnk->link_idx] = true;
2184 	return 0;
2185 
2186 buf_unmap:
2187 	smc_ib_buf_unmap_sg(lnk, buf_desc,
2188 			    is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
2189 free_table:
2190 	sg_free_table(&buf_desc->sgt[lnk->link_idx]);
2191 	return rc;
2192 }
2193 
2194 /* register a new buf on IB device, rmb or vzalloced sndbuf
2195  * must be called under lgr->llc_conf_mutex lock
2196  */
2197 int smcr_link_reg_buf(struct smc_link *link, struct smc_buf_desc *buf_desc)
2198 {
2199 	if (list_empty(&link->lgr->list))
2200 		return -ENOLINK;
2201 	if (!buf_desc->is_reg_mr[link->link_idx]) {
2202 		/* register memory region for new buf */
2203 		if (buf_desc->is_vm)
2204 			buf_desc->mr[link->link_idx]->iova =
2205 				(uintptr_t)buf_desc->cpu_addr;
2206 		if (smc_wr_reg_send(link, buf_desc->mr[link->link_idx])) {
2207 			buf_desc->is_reg_err = true;
2208 			return -EFAULT;
2209 		}
2210 		buf_desc->is_reg_mr[link->link_idx] = true;
2211 	}
2212 	return 0;
2213 }
2214 
2215 static int _smcr_buf_map_lgr(struct smc_link *lnk, struct rw_semaphore *lock,
2216 			     struct list_head *lst, bool is_rmb)
2217 {
2218 	struct smc_buf_desc *buf_desc, *bf;
2219 	int rc = 0;
2220 
2221 	down_write(lock);
2222 	list_for_each_entry_safe(buf_desc, bf, lst, list) {
2223 		if (!buf_desc->used)
2224 			continue;
2225 		rc = smcr_buf_map_link(buf_desc, is_rmb, lnk);
2226 		if (rc)
2227 			goto out;
2228 	}
2229 out:
2230 	up_write(lock);
2231 	return rc;
2232 }
2233 
2234 /* map all used buffers of lgr for a new link */
2235 int smcr_buf_map_lgr(struct smc_link *lnk)
2236 {
2237 	struct smc_link_group *lgr = lnk->lgr;
2238 	int i, rc = 0;
2239 
2240 	for (i = 0; i < SMC_RMBE_SIZES; i++) {
2241 		rc = _smcr_buf_map_lgr(lnk, &lgr->rmbs_lock,
2242 				       &lgr->rmbs[i], true);
2243 		if (rc)
2244 			return rc;
2245 		rc = _smcr_buf_map_lgr(lnk, &lgr->sndbufs_lock,
2246 				       &lgr->sndbufs[i], false);
2247 		if (rc)
2248 			return rc;
2249 	}
2250 	return 0;
2251 }
2252 
2253 /* register all used buffers of lgr for a new link,
2254  * must be called under lgr->llc_conf_mutex lock
2255  */
2256 int smcr_buf_reg_lgr(struct smc_link *lnk)
2257 {
2258 	struct smc_link_group *lgr = lnk->lgr;
2259 	struct smc_buf_desc *buf_desc, *bf;
2260 	int i, rc = 0;
2261 
2262 	/* reg all RMBs for a new link */
2263 	down_write(&lgr->rmbs_lock);
2264 	for (i = 0; i < SMC_RMBE_SIZES; i++) {
2265 		list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list) {
2266 			if (!buf_desc->used)
2267 				continue;
2268 			rc = smcr_link_reg_buf(lnk, buf_desc);
2269 			if (rc) {
2270 				up_write(&lgr->rmbs_lock);
2271 				return rc;
2272 			}
2273 		}
2274 	}
2275 	up_write(&lgr->rmbs_lock);
2276 
2277 	if (lgr->buf_type == SMCR_PHYS_CONT_BUFS)
2278 		return rc;
2279 
2280 	/* reg all vzalloced sndbufs for a new link */
2281 	down_write(&lgr->sndbufs_lock);
2282 	for (i = 0; i < SMC_RMBE_SIZES; i++) {
2283 		list_for_each_entry_safe(buf_desc, bf, &lgr->sndbufs[i], list) {
2284 			if (!buf_desc->used || !buf_desc->is_vm)
2285 				continue;
2286 			rc = smcr_link_reg_buf(lnk, buf_desc);
2287 			if (rc) {
2288 				up_write(&lgr->sndbufs_lock);
2289 				return rc;
2290 			}
2291 		}
2292 	}
2293 	up_write(&lgr->sndbufs_lock);
2294 	return rc;
2295 }
2296 
2297 static struct smc_buf_desc *smcr_new_buf_create(struct smc_link_group *lgr,
2298 						int bufsize)
2299 {
2300 	struct smc_buf_desc *buf_desc;
2301 
2302 	/* try to alloc a new buffer */
2303 	buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
2304 	if (!buf_desc)
2305 		return ERR_PTR(-ENOMEM);
2306 
2307 	switch (lgr->buf_type) {
2308 	case SMCR_PHYS_CONT_BUFS:
2309 	case SMCR_MIXED_BUFS:
2310 		buf_desc->order = get_order(bufsize);
2311 		buf_desc->pages = alloc_pages(GFP_KERNEL | __GFP_NOWARN |
2312 					      __GFP_NOMEMALLOC | __GFP_COMP |
2313 					      __GFP_NORETRY | __GFP_ZERO,
2314 					      buf_desc->order);
2315 		if (buf_desc->pages) {
2316 			buf_desc->cpu_addr =
2317 				(void *)page_address(buf_desc->pages);
2318 			buf_desc->len = bufsize;
2319 			buf_desc->is_vm = false;
2320 			break;
2321 		}
2322 		if (lgr->buf_type == SMCR_PHYS_CONT_BUFS)
2323 			goto out;
2324 		fallthrough;	// try virtually continguous buf
2325 	case SMCR_VIRT_CONT_BUFS:
2326 		buf_desc->order = get_order(bufsize);
2327 		buf_desc->cpu_addr = vzalloc(PAGE_SIZE << buf_desc->order);
2328 		if (!buf_desc->cpu_addr)
2329 			goto out;
2330 		buf_desc->pages = NULL;
2331 		buf_desc->len = bufsize;
2332 		buf_desc->is_vm = true;
2333 		break;
2334 	}
2335 	return buf_desc;
2336 
2337 out:
2338 	kfree(buf_desc);
2339 	return ERR_PTR(-EAGAIN);
2340 }
2341 
2342 /* map buf_desc on all usable links,
2343  * unused buffers stay mapped as long as the link is up
2344  */
2345 static int smcr_buf_map_usable_links(struct smc_link_group *lgr,
2346 				     struct smc_buf_desc *buf_desc, bool is_rmb)
2347 {
2348 	int i, rc = 0, cnt = 0;
2349 
2350 	/* protect against parallel link reconfiguration */
2351 	down_read(&lgr->llc_conf_mutex);
2352 	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
2353 		struct smc_link *lnk = &lgr->lnk[i];
2354 
2355 		if (!smc_link_usable(lnk))
2356 			continue;
2357 		if (smcr_buf_map_link(buf_desc, is_rmb, lnk)) {
2358 			rc = -ENOMEM;
2359 			goto out;
2360 		}
2361 		cnt++;
2362 	}
2363 out:
2364 	up_read(&lgr->llc_conf_mutex);
2365 	if (!rc && !cnt)
2366 		rc = -EINVAL;
2367 	return rc;
2368 }
2369 
2370 static struct smc_buf_desc *smcd_new_buf_create(struct smc_link_group *lgr,
2371 						bool is_dmb, int bufsize)
2372 {
2373 	struct smc_buf_desc *buf_desc;
2374 	int rc;
2375 
2376 	/* try to alloc a new DMB */
2377 	buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
2378 	if (!buf_desc)
2379 		return ERR_PTR(-ENOMEM);
2380 	if (is_dmb) {
2381 		rc = smc_ism_register_dmb(lgr, bufsize, buf_desc);
2382 		if (rc) {
2383 			kfree(buf_desc);
2384 			if (rc == -ENOMEM)
2385 				return ERR_PTR(-EAGAIN);
2386 			if (rc == -ENOSPC)
2387 				return ERR_PTR(-ENOSPC);
2388 			return ERR_PTR(-EIO);
2389 		}
2390 		buf_desc->pages = virt_to_page(buf_desc->cpu_addr);
2391 		/* CDC header stored in buf. So, pretend it was smaller */
2392 		buf_desc->len = bufsize - sizeof(struct smcd_cdc_msg);
2393 	} else {
2394 		buf_desc->cpu_addr = kzalloc(bufsize, GFP_KERNEL |
2395 					     __GFP_NOWARN | __GFP_NORETRY |
2396 					     __GFP_NOMEMALLOC);
2397 		if (!buf_desc->cpu_addr) {
2398 			kfree(buf_desc);
2399 			return ERR_PTR(-EAGAIN);
2400 		}
2401 		buf_desc->len = bufsize;
2402 	}
2403 	return buf_desc;
2404 }
2405 
2406 static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
2407 {
2408 	struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
2409 	struct smc_connection *conn = &smc->conn;
2410 	struct smc_link_group *lgr = conn->lgr;
2411 	struct list_head *buf_list;
2412 	int bufsize, bufsize_comp;
2413 	struct rw_semaphore *lock;	/* lock buffer list */
2414 	bool is_dgraded = false;
2415 
2416 	if (is_rmb)
2417 		/* use socket recv buffer size (w/o overhead) as start value */
2418 		bufsize = smc->sk.sk_rcvbuf / 2;
2419 	else
2420 		/* use socket send buffer size (w/o overhead) as start value */
2421 		bufsize = smc->sk.sk_sndbuf / 2;
2422 
2423 	for (bufsize_comp = smc_compress_bufsize(bufsize, is_smcd, is_rmb);
2424 	     bufsize_comp >= 0; bufsize_comp--) {
2425 		if (is_rmb) {
2426 			lock = &lgr->rmbs_lock;
2427 			buf_list = &lgr->rmbs[bufsize_comp];
2428 		} else {
2429 			lock = &lgr->sndbufs_lock;
2430 			buf_list = &lgr->sndbufs[bufsize_comp];
2431 		}
2432 		bufsize = smc_uncompress_bufsize(bufsize_comp);
2433 
2434 		/* check for reusable slot in the link group */
2435 		buf_desc = smc_buf_get_slot(bufsize_comp, lock, buf_list);
2436 		if (buf_desc) {
2437 			buf_desc->is_dma_need_sync = 0;
2438 			SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, true, bufsize);
2439 			SMC_STAT_BUF_REUSE(smc, is_smcd, is_rmb);
2440 			break; /* found reusable slot */
2441 		}
2442 
2443 		if (is_smcd)
2444 			buf_desc = smcd_new_buf_create(lgr, is_rmb, bufsize);
2445 		else
2446 			buf_desc = smcr_new_buf_create(lgr, bufsize);
2447 
2448 		if (PTR_ERR(buf_desc) == -ENOMEM)
2449 			break;
2450 		if (IS_ERR(buf_desc)) {
2451 			if (!is_dgraded) {
2452 				is_dgraded = true;
2453 				SMC_STAT_RMB_DOWNGRADED(smc, is_smcd, is_rmb);
2454 			}
2455 			continue;
2456 		}
2457 
2458 		SMC_STAT_RMB_ALLOC(smc, is_smcd, is_rmb);
2459 		SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, true, bufsize);
2460 		buf_desc->used = 1;
2461 		down_write(lock);
2462 		smc_lgr_buf_list_add(lgr, is_rmb, buf_list, buf_desc);
2463 		up_write(lock);
2464 		break; /* found */
2465 	}
2466 
2467 	if (IS_ERR(buf_desc))
2468 		return PTR_ERR(buf_desc);
2469 
2470 	if (!is_smcd) {
2471 		if (smcr_buf_map_usable_links(lgr, buf_desc, is_rmb)) {
2472 			smcr_buf_unuse(buf_desc, is_rmb, lgr);
2473 			return -ENOMEM;
2474 		}
2475 	}
2476 
2477 	if (is_rmb) {
2478 		conn->rmb_desc = buf_desc;
2479 		conn->rmbe_size_comp = bufsize_comp;
2480 		smc->sk.sk_rcvbuf = bufsize * 2;
2481 		atomic_set(&conn->bytes_to_rcv, 0);
2482 		conn->rmbe_update_limit =
2483 			smc_rmb_wnd_update_limit(buf_desc->len);
2484 		if (is_smcd)
2485 			smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */
2486 	} else {
2487 		conn->sndbuf_desc = buf_desc;
2488 		smc->sk.sk_sndbuf = bufsize * 2;
2489 		atomic_set(&conn->sndbuf_space, bufsize);
2490 	}
2491 	return 0;
2492 }
2493 
2494 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
2495 {
2496 	if (!conn->sndbuf_desc->is_dma_need_sync)
2497 		return;
2498 	if (!smc_conn_lgr_valid(conn) || conn->lgr->is_smcd ||
2499 	    !smc_link_active(conn->lnk))
2500 		return;
2501 	smc_ib_sync_sg_for_device(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE);
2502 }
2503 
2504 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
2505 {
2506 	int i;
2507 
2508 	if (!conn->rmb_desc->is_dma_need_sync)
2509 		return;
2510 	if (!smc_conn_lgr_valid(conn) || conn->lgr->is_smcd)
2511 		return;
2512 	for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
2513 		if (!smc_link_active(&conn->lgr->lnk[i]))
2514 			continue;
2515 		smc_ib_sync_sg_for_cpu(&conn->lgr->lnk[i], conn->rmb_desc,
2516 				       DMA_FROM_DEVICE);
2517 	}
2518 }
2519 
2520 /* create the send and receive buffer for an SMC socket;
2521  * receive buffers are called RMBs;
2522  * (even though the SMC protocol allows more than one RMB-element per RMB,
2523  * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
2524  * extra RMB for every connection in a link group
2525  */
2526 int smc_buf_create(struct smc_sock *smc, bool is_smcd)
2527 {
2528 	int rc;
2529 
2530 	/* create send buffer */
2531 	if (is_smcd &&
2532 	    smc_ism_support_dmb_nocopy(smc->conn.lgr->smcd))
2533 		goto create_rmb;
2534 
2535 	rc = __smc_buf_create(smc, is_smcd, false);
2536 	if (rc)
2537 		return rc;
2538 
2539 create_rmb:
2540 	/* create rmb */
2541 	rc = __smc_buf_create(smc, is_smcd, true);
2542 	if (rc && smc->conn.sndbuf_desc) {
2543 		down_write(&smc->conn.lgr->sndbufs_lock);
2544 		smc_lgr_buf_list_del(smc->conn.lgr, false,
2545 				     smc->conn.sndbuf_desc);
2546 		up_write(&smc->conn.lgr->sndbufs_lock);
2547 		smc_buf_free(smc->conn.lgr, false, smc->conn.sndbuf_desc);
2548 		smc->conn.sndbuf_desc = NULL;
2549 	}
2550 	return rc;
2551 }
2552 
2553 int smcd_buf_attach(struct smc_sock *smc)
2554 {
2555 	struct smc_connection *conn = &smc->conn;
2556 	struct smcd_dev *smcd = conn->lgr->smcd;
2557 	u64 peer_token = conn->peer_token;
2558 	struct smc_buf_desc *buf_desc;
2559 	int rc;
2560 
2561 	buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
2562 	if (!buf_desc)
2563 		return -ENOMEM;
2564 
2565 	/* The ghost sndbuf_desc describes the same memory region as
2566 	 * peer RMB. Its lifecycle is consistent with the connection's
2567 	 * and it will be freed with the connections instead of the
2568 	 * link group.
2569 	 */
2570 	rc = smc_ism_attach_dmb(smcd, peer_token, buf_desc);
2571 	if (rc)
2572 		goto free;
2573 
2574 	smc->sk.sk_sndbuf = buf_desc->len;
2575 	buf_desc->cpu_addr =
2576 		(u8 *)buf_desc->cpu_addr + sizeof(struct smcd_cdc_msg);
2577 	buf_desc->len -= sizeof(struct smcd_cdc_msg);
2578 	conn->sndbuf_desc = buf_desc;
2579 	conn->sndbuf_desc->used = 1;
2580 	atomic_set(&conn->sndbuf_space, conn->sndbuf_desc->len);
2581 	return 0;
2582 
2583 free:
2584 	kfree(buf_desc);
2585 	return rc;
2586 }
2587 
2588 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
2589 {
2590 	int i;
2591 
2592 	for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
2593 		if (!test_and_set_bit(i, lgr->rtokens_used_mask))
2594 			return i;
2595 	}
2596 	return -ENOSPC;
2597 }
2598 
2599 static int smc_rtoken_find_by_link(struct smc_link_group *lgr, int lnk_idx,
2600 				   u32 rkey)
2601 {
2602 	int i;
2603 
2604 	for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2605 		if (test_bit(i, lgr->rtokens_used_mask) &&
2606 		    lgr->rtokens[i][lnk_idx].rkey == rkey)
2607 			return i;
2608 	}
2609 	return -ENOENT;
2610 }
2611 
2612 /* set rtoken for a new link to an existing rmb */
2613 void smc_rtoken_set(struct smc_link_group *lgr, int link_idx, int link_idx_new,
2614 		    __be32 nw_rkey_known, __be64 nw_vaddr, __be32 nw_rkey)
2615 {
2616 	int rtok_idx;
2617 
2618 	rtok_idx = smc_rtoken_find_by_link(lgr, link_idx, ntohl(nw_rkey_known));
2619 	if (rtok_idx == -ENOENT)
2620 		return;
2621 	lgr->rtokens[rtok_idx][link_idx_new].rkey = ntohl(nw_rkey);
2622 	lgr->rtokens[rtok_idx][link_idx_new].dma_addr = be64_to_cpu(nw_vaddr);
2623 }
2624 
2625 /* set rtoken for a new link whose link_id is given */
2626 void smc_rtoken_set2(struct smc_link_group *lgr, int rtok_idx, int link_id,
2627 		     __be64 nw_vaddr, __be32 nw_rkey)
2628 {
2629 	u64 dma_addr = be64_to_cpu(nw_vaddr);
2630 	u32 rkey = ntohl(nw_rkey);
2631 	bool found = false;
2632 	int link_idx;
2633 
2634 	for (link_idx = 0; link_idx < SMC_LINKS_PER_LGR_MAX; link_idx++) {
2635 		if (lgr->lnk[link_idx].link_id == link_id) {
2636 			found = true;
2637 			break;
2638 		}
2639 	}
2640 	if (!found)
2641 		return;
2642 	lgr->rtokens[rtok_idx][link_idx].rkey = rkey;
2643 	lgr->rtokens[rtok_idx][link_idx].dma_addr = dma_addr;
2644 }
2645 
2646 /* add a new rtoken from peer */
2647 int smc_rtoken_add(struct smc_link *lnk, __be64 nw_vaddr, __be32 nw_rkey)
2648 {
2649 	struct smc_link_group *lgr = smc_get_lgr(lnk);
2650 	u64 dma_addr = be64_to_cpu(nw_vaddr);
2651 	u32 rkey = ntohl(nw_rkey);
2652 	int i;
2653 
2654 	for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2655 		if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
2656 		    lgr->rtokens[i][lnk->link_idx].dma_addr == dma_addr &&
2657 		    test_bit(i, lgr->rtokens_used_mask)) {
2658 			/* already in list */
2659 			return i;
2660 		}
2661 	}
2662 	i = smc_rmb_reserve_rtoken_idx(lgr);
2663 	if (i < 0)
2664 		return i;
2665 	lgr->rtokens[i][lnk->link_idx].rkey = rkey;
2666 	lgr->rtokens[i][lnk->link_idx].dma_addr = dma_addr;
2667 	return i;
2668 }
2669 
2670 /* delete an rtoken from all links */
2671 int smc_rtoken_delete(struct smc_link *lnk, __be32 nw_rkey)
2672 {
2673 	struct smc_link_group *lgr = smc_get_lgr(lnk);
2674 	u32 rkey = ntohl(nw_rkey);
2675 	int i, j;
2676 
2677 	for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2678 		if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
2679 		    test_bit(i, lgr->rtokens_used_mask)) {
2680 			for (j = 0; j < SMC_LINKS_PER_LGR_MAX; j++) {
2681 				lgr->rtokens[i][j].rkey = 0;
2682 				lgr->rtokens[i][j].dma_addr = 0;
2683 			}
2684 			clear_bit(i, lgr->rtokens_used_mask);
2685 			return 0;
2686 		}
2687 	}
2688 	return -ENOENT;
2689 }
2690 
2691 /* save rkey and dma_addr received from peer during clc handshake */
2692 int smc_rmb_rtoken_handling(struct smc_connection *conn,
2693 			    struct smc_link *lnk,
2694 			    struct smc_clc_msg_accept_confirm *clc)
2695 {
2696 	conn->rtoken_idx = smc_rtoken_add(lnk, clc->r0.rmb_dma_addr,
2697 					  clc->r0.rmb_rkey);
2698 	if (conn->rtoken_idx < 0)
2699 		return conn->rtoken_idx;
2700 	return 0;
2701 }
2702 
2703 static void smc_core_going_away(void)
2704 {
2705 	struct smc_ib_device *smcibdev;
2706 	struct smcd_dev *smcd;
2707 
2708 	mutex_lock(&smc_ib_devices.mutex);
2709 	list_for_each_entry(smcibdev, &smc_ib_devices.list, list) {
2710 		int i;
2711 
2712 		for (i = 0; i < SMC_MAX_PORTS; i++)
2713 			set_bit(i, smcibdev->ports_going_away);
2714 	}
2715 	mutex_unlock(&smc_ib_devices.mutex);
2716 
2717 	mutex_lock(&smcd_dev_list.mutex);
2718 	list_for_each_entry(smcd, &smcd_dev_list.list, list) {
2719 		smcd->going_away = 1;
2720 	}
2721 	mutex_unlock(&smcd_dev_list.mutex);
2722 }
2723 
2724 /* Clean up all SMC link groups */
2725 static void smc_lgrs_shutdown(void)
2726 {
2727 	struct smcd_dev *smcd;
2728 
2729 	smc_core_going_away();
2730 
2731 	smc_smcr_terminate_all(NULL);
2732 
2733 	mutex_lock(&smcd_dev_list.mutex);
2734 	list_for_each_entry(smcd, &smcd_dev_list.list, list)
2735 		smc_smcd_terminate_all(smcd);
2736 	mutex_unlock(&smcd_dev_list.mutex);
2737 }
2738 
2739 static int smc_core_reboot_event(struct notifier_block *this,
2740 				 unsigned long event, void *ptr)
2741 {
2742 	smc_lgrs_shutdown();
2743 	smc_ib_unregister_client();
2744 	smc_ism_exit();
2745 	return 0;
2746 }
2747 
2748 static struct notifier_block smc_reboot_notifier = {
2749 	.notifier_call = smc_core_reboot_event,
2750 };
2751 
2752 int __init smc_core_init(void)
2753 {
2754 	return register_reboot_notifier(&smc_reboot_notifier);
2755 }
2756 
2757 /* Called (from smc_exit) when module is removed */
2758 void smc_core_exit(void)
2759 {
2760 	unregister_reboot_notifier(&smc_reboot_notifier);
2761 	smc_lgrs_shutdown();
2762 }
2763