Loading...
Note: File does not exist in v3.5.6.
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
38#define SMC_LGR_NUM_INCR 256
39#define SMC_LGR_FREE_DELAY_SERV (600 * HZ)
40#define SMC_LGR_FREE_DELAY_CLNT (SMC_LGR_FREE_DELAY_SERV + 10 * HZ)
41
42struct smc_lgr_list smc_lgr_list = { /* established link groups */
43 .lock = __SPIN_LOCK_UNLOCKED(smc_lgr_list.lock),
44 .list = LIST_HEAD_INIT(smc_lgr_list.list),
45 .num = 0,
46};
47
48static atomic_t lgr_cnt = ATOMIC_INIT(0); /* number of existing link groups */
49static DECLARE_WAIT_QUEUE_HEAD(lgrs_deleted);
50
51static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
52 struct smc_buf_desc *buf_desc);
53static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft);
54
55static void smc_link_down_work(struct work_struct *work);
56
57/* return head of link group list and its lock for a given link group */
58static inline struct list_head *smc_lgr_list_head(struct smc_link_group *lgr,
59 spinlock_t **lgr_lock)
60{
61 if (lgr->is_smcd) {
62 *lgr_lock = &lgr->smcd->lgr_lock;
63 return &lgr->smcd->lgr_list;
64 }
65
66 *lgr_lock = &smc_lgr_list.lock;
67 return &smc_lgr_list.list;
68}
69
70static void smc_ibdev_cnt_inc(struct smc_link *lnk)
71{
72 atomic_inc(&lnk->smcibdev->lnk_cnt_by_port[lnk->ibport - 1]);
73}
74
75static void smc_ibdev_cnt_dec(struct smc_link *lnk)
76{
77 atomic_dec(&lnk->smcibdev->lnk_cnt_by_port[lnk->ibport - 1]);
78}
79
80static void smc_lgr_schedule_free_work(struct smc_link_group *lgr)
81{
82 /* client link group creation always follows the server link group
83 * creation. For client use a somewhat higher removal delay time,
84 * otherwise there is a risk of out-of-sync link groups.
85 */
86 if (!lgr->freeing) {
87 mod_delayed_work(system_wq, &lgr->free_work,
88 (!lgr->is_smcd && lgr->role == SMC_CLNT) ?
89 SMC_LGR_FREE_DELAY_CLNT :
90 SMC_LGR_FREE_DELAY_SERV);
91 }
92}
93
94/* Register connection's alert token in our lookup structure.
95 * To use rbtrees we have to implement our own insert core.
96 * Requires @conns_lock
97 * @smc connection to register
98 * Returns 0 on success, != otherwise.
99 */
100static void smc_lgr_add_alert_token(struct smc_connection *conn)
101{
102 struct rb_node **link, *parent = NULL;
103 u32 token = conn->alert_token_local;
104
105 link = &conn->lgr->conns_all.rb_node;
106 while (*link) {
107 struct smc_connection *cur = rb_entry(*link,
108 struct smc_connection, alert_node);
109
110 parent = *link;
111 if (cur->alert_token_local > token)
112 link = &parent->rb_left;
113 else
114 link = &parent->rb_right;
115 }
116 /* Put the new node there */
117 rb_link_node(&conn->alert_node, parent, link);
118 rb_insert_color(&conn->alert_node, &conn->lgr->conns_all);
119}
120
121/* assign an SMC-R link to the connection */
122static int smcr_lgr_conn_assign_link(struct smc_connection *conn, bool first)
123{
124 enum smc_link_state expected = first ? SMC_LNK_ACTIVATING :
125 SMC_LNK_ACTIVE;
126 int i, j;
127
128 /* do link balancing */
129 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
130 struct smc_link *lnk = &conn->lgr->lnk[i];
131
132 if (lnk->state != expected || lnk->link_is_asym)
133 continue;
134 if (conn->lgr->role == SMC_CLNT) {
135 conn->lnk = lnk; /* temporary, SMC server assigns link*/
136 break;
137 }
138 if (conn->lgr->conns_num % 2) {
139 for (j = i + 1; j < SMC_LINKS_PER_LGR_MAX; j++) {
140 struct smc_link *lnk2;
141
142 lnk2 = &conn->lgr->lnk[j];
143 if (lnk2->state == expected &&
144 !lnk2->link_is_asym) {
145 conn->lnk = lnk2;
146 break;
147 }
148 }
149 }
150 if (!conn->lnk)
151 conn->lnk = lnk;
152 break;
153 }
154 if (!conn->lnk)
155 return SMC_CLC_DECL_NOACTLINK;
156 atomic_inc(&conn->lnk->conn_cnt);
157 return 0;
158}
159
160/* Register connection in link group by assigning an alert token
161 * registered in a search tree.
162 * Requires @conns_lock
163 * Note that '0' is a reserved value and not assigned.
164 */
165static int smc_lgr_register_conn(struct smc_connection *conn, bool first)
166{
167 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
168 static atomic_t nexttoken = ATOMIC_INIT(0);
169 int rc;
170
171 if (!conn->lgr->is_smcd) {
172 rc = smcr_lgr_conn_assign_link(conn, first);
173 if (rc)
174 return rc;
175 }
176 /* find a new alert_token_local value not yet used by some connection
177 * in this link group
178 */
179 sock_hold(&smc->sk); /* sock_put in smc_lgr_unregister_conn() */
180 while (!conn->alert_token_local) {
181 conn->alert_token_local = atomic_inc_return(&nexttoken);
182 if (smc_lgr_find_conn(conn->alert_token_local, conn->lgr))
183 conn->alert_token_local = 0;
184 }
185 smc_lgr_add_alert_token(conn);
186 conn->lgr->conns_num++;
187 return 0;
188}
189
190/* Unregister connection and reset the alert token of the given connection<
191 */
192static void __smc_lgr_unregister_conn(struct smc_connection *conn)
193{
194 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
195 struct smc_link_group *lgr = conn->lgr;
196
197 rb_erase(&conn->alert_node, &lgr->conns_all);
198 if (conn->lnk)
199 atomic_dec(&conn->lnk->conn_cnt);
200 lgr->conns_num--;
201 conn->alert_token_local = 0;
202 sock_put(&smc->sk); /* sock_hold in smc_lgr_register_conn() */
203}
204
205/* Unregister connection from lgr
206 */
207static void smc_lgr_unregister_conn(struct smc_connection *conn)
208{
209 struct smc_link_group *lgr = conn->lgr;
210
211 if (!lgr)
212 return;
213 write_lock_bh(&lgr->conns_lock);
214 if (conn->alert_token_local) {
215 __smc_lgr_unregister_conn(conn);
216 }
217 write_unlock_bh(&lgr->conns_lock);
218 conn->lgr = NULL;
219}
220
221int smc_nl_get_sys_info(struct sk_buff *skb, struct netlink_callback *cb)
222{
223 struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
224 char hostname[SMC_MAX_HOSTNAME_LEN + 1];
225 char smc_seid[SMC_MAX_EID_LEN + 1];
226 struct smcd_dev *smcd_dev;
227 struct nlattr *attrs;
228 u8 *seid = NULL;
229 u8 *host = NULL;
230 void *nlh;
231
232 nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
233 &smc_gen_nl_family, NLM_F_MULTI,
234 SMC_NETLINK_GET_SYS_INFO);
235 if (!nlh)
236 goto errmsg;
237 if (cb_ctx->pos[0])
238 goto errout;
239 attrs = nla_nest_start(skb, SMC_GEN_SYS_INFO);
240 if (!attrs)
241 goto errout;
242 if (nla_put_u8(skb, SMC_NLA_SYS_VER, SMC_V2))
243 goto errattr;
244 if (nla_put_u8(skb, SMC_NLA_SYS_REL, SMC_RELEASE))
245 goto errattr;
246 if (nla_put_u8(skb, SMC_NLA_SYS_IS_ISM_V2, smc_ism_is_v2_capable()))
247 goto errattr;
248 smc_clc_get_hostname(&host);
249 if (host) {
250 memcpy(hostname, host, SMC_MAX_HOSTNAME_LEN);
251 hostname[SMC_MAX_HOSTNAME_LEN] = 0;
252 if (nla_put_string(skb, SMC_NLA_SYS_LOCAL_HOST, hostname))
253 goto errattr;
254 }
255 mutex_lock(&smcd_dev_list.mutex);
256 smcd_dev = list_first_entry_or_null(&smcd_dev_list.list,
257 struct smcd_dev, list);
258 if (smcd_dev)
259 smc_ism_get_system_eid(smcd_dev, &seid);
260 mutex_unlock(&smcd_dev_list.mutex);
261 if (seid && smc_ism_is_v2_capable()) {
262 memcpy(smc_seid, seid, SMC_MAX_EID_LEN);
263 smc_seid[SMC_MAX_EID_LEN] = 0;
264 if (nla_put_string(skb, SMC_NLA_SYS_SEID, smc_seid))
265 goto errattr;
266 }
267 nla_nest_end(skb, attrs);
268 genlmsg_end(skb, nlh);
269 cb_ctx->pos[0] = 1;
270 return skb->len;
271
272errattr:
273 nla_nest_cancel(skb, attrs);
274errout:
275 genlmsg_cancel(skb, nlh);
276errmsg:
277 return skb->len;
278}
279
280static int smc_nl_fill_lgr(struct smc_link_group *lgr,
281 struct sk_buff *skb,
282 struct netlink_callback *cb)
283{
284 char smc_target[SMC_MAX_PNETID_LEN + 1];
285 struct nlattr *attrs;
286
287 attrs = nla_nest_start(skb, SMC_GEN_LGR_SMCR);
288 if (!attrs)
289 goto errout;
290
291 if (nla_put_u32(skb, SMC_NLA_LGR_R_ID, *((u32 *)&lgr->id)))
292 goto errattr;
293 if (nla_put_u32(skb, SMC_NLA_LGR_R_CONNS_NUM, lgr->conns_num))
294 goto errattr;
295 if (nla_put_u8(skb, SMC_NLA_LGR_R_ROLE, lgr->role))
296 goto errattr;
297 if (nla_put_u8(skb, SMC_NLA_LGR_R_TYPE, lgr->type))
298 goto errattr;
299 if (nla_put_u8(skb, SMC_NLA_LGR_R_VLAN_ID, lgr->vlan_id))
300 goto errattr;
301 memcpy(smc_target, lgr->pnet_id, SMC_MAX_PNETID_LEN);
302 smc_target[SMC_MAX_PNETID_LEN] = 0;
303 if (nla_put_string(skb, SMC_NLA_LGR_R_PNETID, smc_target))
304 goto errattr;
305
306 nla_nest_end(skb, attrs);
307 return 0;
308errattr:
309 nla_nest_cancel(skb, attrs);
310errout:
311 return -EMSGSIZE;
312}
313
314static int smc_nl_fill_lgr_link(struct smc_link_group *lgr,
315 struct smc_link *link,
316 struct sk_buff *skb,
317 struct netlink_callback *cb)
318{
319 char smc_ibname[IB_DEVICE_NAME_MAX];
320 u8 smc_gid_target[41];
321 struct nlattr *attrs;
322 u32 link_uid = 0;
323 void *nlh;
324
325 nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
326 &smc_gen_nl_family, NLM_F_MULTI,
327 SMC_NETLINK_GET_LINK_SMCR);
328 if (!nlh)
329 goto errmsg;
330
331 attrs = nla_nest_start(skb, SMC_GEN_LINK_SMCR);
332 if (!attrs)
333 goto errout;
334
335 if (nla_put_u8(skb, SMC_NLA_LINK_ID, link->link_id))
336 goto errattr;
337 if (nla_put_u32(skb, SMC_NLA_LINK_STATE, link->state))
338 goto errattr;
339 if (nla_put_u32(skb, SMC_NLA_LINK_CONN_CNT,
340 atomic_read(&link->conn_cnt)))
341 goto errattr;
342 if (nla_put_u8(skb, SMC_NLA_LINK_IB_PORT, link->ibport))
343 goto errattr;
344 if (nla_put_u32(skb, SMC_NLA_LINK_NET_DEV, link->ndev_ifidx))
345 goto errattr;
346 snprintf(smc_ibname, sizeof(smc_ibname), "%s", link->ibname);
347 if (nla_put_string(skb, SMC_NLA_LINK_IB_DEV, smc_ibname))
348 goto errattr;
349 memcpy(&link_uid, link->link_uid, sizeof(link_uid));
350 if (nla_put_u32(skb, SMC_NLA_LINK_UID, link_uid))
351 goto errattr;
352 memcpy(&link_uid, link->peer_link_uid, sizeof(link_uid));
353 if (nla_put_u32(skb, SMC_NLA_LINK_PEER_UID, link_uid))
354 goto errattr;
355 memset(smc_gid_target, 0, sizeof(smc_gid_target));
356 smc_gid_be16_convert(smc_gid_target, link->gid);
357 if (nla_put_string(skb, SMC_NLA_LINK_GID, smc_gid_target))
358 goto errattr;
359 memset(smc_gid_target, 0, sizeof(smc_gid_target));
360 smc_gid_be16_convert(smc_gid_target, link->peer_gid);
361 if (nla_put_string(skb, SMC_NLA_LINK_PEER_GID, smc_gid_target))
362 goto errattr;
363
364 nla_nest_end(skb, attrs);
365 genlmsg_end(skb, nlh);
366 return 0;
367errattr:
368 nla_nest_cancel(skb, attrs);
369errout:
370 genlmsg_cancel(skb, nlh);
371errmsg:
372 return -EMSGSIZE;
373}
374
375static int smc_nl_handle_lgr(struct smc_link_group *lgr,
376 struct sk_buff *skb,
377 struct netlink_callback *cb,
378 bool list_links)
379{
380 void *nlh;
381 int i;
382
383 nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
384 &smc_gen_nl_family, NLM_F_MULTI,
385 SMC_NETLINK_GET_LGR_SMCR);
386 if (!nlh)
387 goto errmsg;
388 if (smc_nl_fill_lgr(lgr, skb, cb))
389 goto errout;
390
391 genlmsg_end(skb, nlh);
392 if (!list_links)
393 goto out;
394 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
395 if (!smc_link_usable(&lgr->lnk[i]))
396 continue;
397 if (smc_nl_fill_lgr_link(lgr, &lgr->lnk[i], skb, cb))
398 goto errout;
399 }
400out:
401 return 0;
402
403errout:
404 genlmsg_cancel(skb, nlh);
405errmsg:
406 return -EMSGSIZE;
407}
408
409static void smc_nl_fill_lgr_list(struct smc_lgr_list *smc_lgr,
410 struct sk_buff *skb,
411 struct netlink_callback *cb,
412 bool list_links)
413{
414 struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
415 struct smc_link_group *lgr;
416 int snum = cb_ctx->pos[0];
417 int num = 0;
418
419 spin_lock_bh(&smc_lgr->lock);
420 list_for_each_entry(lgr, &smc_lgr->list, list) {
421 if (num < snum)
422 goto next;
423 if (smc_nl_handle_lgr(lgr, skb, cb, list_links))
424 goto errout;
425next:
426 num++;
427 }
428errout:
429 spin_unlock_bh(&smc_lgr->lock);
430 cb_ctx->pos[0] = num;
431}
432
433static int smc_nl_fill_smcd_lgr(struct smc_link_group *lgr,
434 struct sk_buff *skb,
435 struct netlink_callback *cb)
436{
437 char smc_host[SMC_MAX_HOSTNAME_LEN + 1];
438 char smc_pnet[SMC_MAX_PNETID_LEN + 1];
439 char smc_eid[SMC_MAX_EID_LEN + 1];
440 struct nlattr *v2_attrs;
441 struct nlattr *attrs;
442 void *nlh;
443
444 nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
445 &smc_gen_nl_family, NLM_F_MULTI,
446 SMC_NETLINK_GET_LGR_SMCD);
447 if (!nlh)
448 goto errmsg;
449
450 attrs = nla_nest_start(skb, SMC_GEN_LGR_SMCD);
451 if (!attrs)
452 goto errout;
453
454 if (nla_put_u32(skb, SMC_NLA_LGR_D_ID, *((u32 *)&lgr->id)))
455 goto errattr;
456 if (nla_put_u64_64bit(skb, SMC_NLA_LGR_D_GID, lgr->smcd->local_gid,
457 SMC_NLA_LGR_D_PAD))
458 goto errattr;
459 if (nla_put_u64_64bit(skb, SMC_NLA_LGR_D_PEER_GID, lgr->peer_gid,
460 SMC_NLA_LGR_D_PAD))
461 goto errattr;
462 if (nla_put_u8(skb, SMC_NLA_LGR_D_VLAN_ID, lgr->vlan_id))
463 goto errattr;
464 if (nla_put_u32(skb, SMC_NLA_LGR_D_CONNS_NUM, lgr->conns_num))
465 goto errattr;
466 if (nla_put_u32(skb, SMC_NLA_LGR_D_CHID, smc_ism_get_chid(lgr->smcd)))
467 goto errattr;
468 memcpy(smc_pnet, lgr->smcd->pnetid, SMC_MAX_PNETID_LEN);
469 smc_pnet[SMC_MAX_PNETID_LEN] = 0;
470 if (nla_put_string(skb, SMC_NLA_LGR_D_PNETID, smc_pnet))
471 goto errattr;
472
473 v2_attrs = nla_nest_start(skb, SMC_NLA_LGR_V2);
474 if (!v2_attrs)
475 goto errattr;
476 if (nla_put_u8(skb, SMC_NLA_LGR_V2_VER, lgr->smc_version))
477 goto errv2attr;
478 if (nla_put_u8(skb, SMC_NLA_LGR_V2_REL, lgr->peer_smc_release))
479 goto errv2attr;
480 if (nla_put_u8(skb, SMC_NLA_LGR_V2_OS, lgr->peer_os))
481 goto errv2attr;
482 memcpy(smc_host, lgr->peer_hostname, SMC_MAX_HOSTNAME_LEN);
483 smc_host[SMC_MAX_HOSTNAME_LEN] = 0;
484 if (nla_put_string(skb, SMC_NLA_LGR_V2_PEER_HOST, smc_host))
485 goto errv2attr;
486 memcpy(smc_eid, lgr->negotiated_eid, SMC_MAX_EID_LEN);
487 smc_eid[SMC_MAX_EID_LEN] = 0;
488 if (nla_put_string(skb, SMC_NLA_LGR_V2_NEG_EID, smc_eid))
489 goto errv2attr;
490
491 nla_nest_end(skb, v2_attrs);
492 nla_nest_end(skb, attrs);
493 genlmsg_end(skb, nlh);
494 return 0;
495
496errv2attr:
497 nla_nest_cancel(skb, v2_attrs);
498errattr:
499 nla_nest_cancel(skb, attrs);
500errout:
501 genlmsg_cancel(skb, nlh);
502errmsg:
503 return -EMSGSIZE;
504}
505
506static int smc_nl_handle_smcd_lgr(struct smcd_dev *dev,
507 struct sk_buff *skb,
508 struct netlink_callback *cb)
509{
510 struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
511 struct smc_link_group *lgr;
512 int snum = cb_ctx->pos[1];
513 int rc = 0, num = 0;
514
515 spin_lock_bh(&dev->lgr_lock);
516 list_for_each_entry(lgr, &dev->lgr_list, list) {
517 if (!lgr->is_smcd)
518 continue;
519 if (num < snum)
520 goto next;
521 rc = smc_nl_fill_smcd_lgr(lgr, skb, cb);
522 if (rc)
523 goto errout;
524next:
525 num++;
526 }
527errout:
528 spin_unlock_bh(&dev->lgr_lock);
529 cb_ctx->pos[1] = num;
530 return rc;
531}
532
533static int smc_nl_fill_smcd_dev(struct smcd_dev_list *dev_list,
534 struct sk_buff *skb,
535 struct netlink_callback *cb)
536{
537 struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
538 struct smcd_dev *smcd_dev;
539 int snum = cb_ctx->pos[0];
540 int rc = 0, num = 0;
541
542 mutex_lock(&dev_list->mutex);
543 list_for_each_entry(smcd_dev, &dev_list->list, list) {
544 if (list_empty(&smcd_dev->lgr_list))
545 continue;
546 if (num < snum)
547 goto next;
548 rc = smc_nl_handle_smcd_lgr(smcd_dev, skb, cb);
549 if (rc)
550 goto errout;
551next:
552 num++;
553 }
554errout:
555 mutex_unlock(&dev_list->mutex);
556 cb_ctx->pos[0] = num;
557 return rc;
558}
559
560int smcr_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb)
561{
562 bool list_links = false;
563
564 smc_nl_fill_lgr_list(&smc_lgr_list, skb, cb, list_links);
565 return skb->len;
566}
567
568int smcr_nl_get_link(struct sk_buff *skb, struct netlink_callback *cb)
569{
570 bool list_links = true;
571
572 smc_nl_fill_lgr_list(&smc_lgr_list, skb, cb, list_links);
573 return skb->len;
574}
575
576int smcd_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb)
577{
578 smc_nl_fill_smcd_dev(&smcd_dev_list, skb, cb);
579 return skb->len;
580}
581
582void smc_lgr_cleanup_early(struct smc_connection *conn)
583{
584 struct smc_link_group *lgr = conn->lgr;
585 struct list_head *lgr_list;
586 spinlock_t *lgr_lock;
587
588 if (!lgr)
589 return;
590
591 smc_conn_free(conn);
592 lgr_list = smc_lgr_list_head(lgr, &lgr_lock);
593 spin_lock_bh(lgr_lock);
594 /* do not use this link group for new connections */
595 if (!list_empty(lgr_list))
596 list_del_init(lgr_list);
597 spin_unlock_bh(lgr_lock);
598 __smc_lgr_terminate(lgr, true);
599}
600
601static void smcr_lgr_link_deactivate_all(struct smc_link_group *lgr)
602{
603 int i;
604
605 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
606 struct smc_link *lnk = &lgr->lnk[i];
607
608 if (smc_link_usable(lnk))
609 lnk->state = SMC_LNK_INACTIVE;
610 }
611 wake_up_all(&lgr->llc_msg_waiter);
612 wake_up_all(&lgr->llc_flow_waiter);
613}
614
615static void smc_lgr_free(struct smc_link_group *lgr);
616
617static void smc_lgr_free_work(struct work_struct *work)
618{
619 struct smc_link_group *lgr = container_of(to_delayed_work(work),
620 struct smc_link_group,
621 free_work);
622 spinlock_t *lgr_lock;
623 bool conns;
624
625 smc_lgr_list_head(lgr, &lgr_lock);
626 spin_lock_bh(lgr_lock);
627 if (lgr->freeing) {
628 spin_unlock_bh(lgr_lock);
629 return;
630 }
631 read_lock_bh(&lgr->conns_lock);
632 conns = RB_EMPTY_ROOT(&lgr->conns_all);
633 read_unlock_bh(&lgr->conns_lock);
634 if (!conns) { /* number of lgr connections is no longer zero */
635 spin_unlock_bh(lgr_lock);
636 return;
637 }
638 list_del_init(&lgr->list); /* remove from smc_lgr_list */
639 lgr->freeing = 1; /* this instance does the freeing, no new schedule */
640 spin_unlock_bh(lgr_lock);
641 cancel_delayed_work(&lgr->free_work);
642
643 if (!lgr->is_smcd && !lgr->terminating)
644 smc_llc_send_link_delete_all(lgr, true,
645 SMC_LLC_DEL_PROG_INIT_TERM);
646 if (lgr->is_smcd && !lgr->terminating)
647 smc_ism_signal_shutdown(lgr);
648 if (!lgr->is_smcd)
649 smcr_lgr_link_deactivate_all(lgr);
650 smc_lgr_free(lgr);
651}
652
653static void smc_lgr_terminate_work(struct work_struct *work)
654{
655 struct smc_link_group *lgr = container_of(work, struct smc_link_group,
656 terminate_work);
657
658 __smc_lgr_terminate(lgr, true);
659}
660
661/* return next unique link id for the lgr */
662static u8 smcr_next_link_id(struct smc_link_group *lgr)
663{
664 u8 link_id;
665 int i;
666
667 while (1) {
668 link_id = ++lgr->next_link_id;
669 if (!link_id) /* skip zero as link_id */
670 link_id = ++lgr->next_link_id;
671 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
672 if (smc_link_usable(&lgr->lnk[i]) &&
673 lgr->lnk[i].link_id == link_id)
674 continue;
675 }
676 break;
677 }
678 return link_id;
679}
680
681static void smcr_copy_dev_info_to_link(struct smc_link *link)
682{
683 struct smc_ib_device *smcibdev = link->smcibdev;
684
685 snprintf(link->ibname, sizeof(link->ibname), "%s",
686 smcibdev->ibdev->name);
687 link->ndev_ifidx = smcibdev->ndev_ifidx[link->ibport - 1];
688}
689
690int smcr_link_init(struct smc_link_group *lgr, struct smc_link *lnk,
691 u8 link_idx, struct smc_init_info *ini)
692{
693 u8 rndvec[3];
694 int rc;
695
696 get_device(&ini->ib_dev->ibdev->dev);
697 atomic_inc(&ini->ib_dev->lnk_cnt);
698 lnk->link_id = smcr_next_link_id(lgr);
699 lnk->lgr = lgr;
700 lnk->link_idx = link_idx;
701 lnk->smcibdev = ini->ib_dev;
702 lnk->ibport = ini->ib_port;
703 smc_ibdev_cnt_inc(lnk);
704 smcr_copy_dev_info_to_link(lnk);
705 lnk->path_mtu = ini->ib_dev->pattr[ini->ib_port - 1].active_mtu;
706 atomic_set(&lnk->conn_cnt, 0);
707 smc_llc_link_set_uid(lnk);
708 INIT_WORK(&lnk->link_down_wrk, smc_link_down_work);
709 if (!ini->ib_dev->initialized) {
710 rc = (int)smc_ib_setup_per_ibdev(ini->ib_dev);
711 if (rc)
712 goto out;
713 }
714 get_random_bytes(rndvec, sizeof(rndvec));
715 lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) +
716 (rndvec[2] << 16);
717 rc = smc_ib_determine_gid(lnk->smcibdev, lnk->ibport,
718 ini->vlan_id, lnk->gid, &lnk->sgid_index);
719 if (rc)
720 goto out;
721 rc = smc_llc_link_init(lnk);
722 if (rc)
723 goto out;
724 rc = smc_wr_alloc_link_mem(lnk);
725 if (rc)
726 goto clear_llc_lnk;
727 rc = smc_ib_create_protection_domain(lnk);
728 if (rc)
729 goto free_link_mem;
730 rc = smc_ib_create_queue_pair(lnk);
731 if (rc)
732 goto dealloc_pd;
733 rc = smc_wr_create_link(lnk);
734 if (rc)
735 goto destroy_qp;
736 lnk->state = SMC_LNK_ACTIVATING;
737 return 0;
738
739destroy_qp:
740 smc_ib_destroy_queue_pair(lnk);
741dealloc_pd:
742 smc_ib_dealloc_protection_domain(lnk);
743free_link_mem:
744 smc_wr_free_link_mem(lnk);
745clear_llc_lnk:
746 smc_llc_link_clear(lnk, false);
747out:
748 smc_ibdev_cnt_dec(lnk);
749 put_device(&ini->ib_dev->ibdev->dev);
750 memset(lnk, 0, sizeof(struct smc_link));
751 lnk->state = SMC_LNK_UNUSED;
752 if (!atomic_dec_return(&ini->ib_dev->lnk_cnt))
753 wake_up(&ini->ib_dev->lnks_deleted);
754 return rc;
755}
756
757/* create a new SMC link group */
758static int smc_lgr_create(struct smc_sock *smc, struct smc_init_info *ini)
759{
760 struct smc_link_group *lgr;
761 struct list_head *lgr_list;
762 struct smc_link *lnk;
763 spinlock_t *lgr_lock;
764 u8 link_idx;
765 int rc = 0;
766 int i;
767
768 if (ini->is_smcd && ini->vlan_id) {
769 if (smc_ism_get_vlan(ini->ism_dev[ini->ism_selected],
770 ini->vlan_id)) {
771 rc = SMC_CLC_DECL_ISMVLANERR;
772 goto out;
773 }
774 }
775
776 lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
777 if (!lgr) {
778 rc = SMC_CLC_DECL_MEM;
779 goto ism_put_vlan;
780 }
781 lgr->tx_wq = alloc_workqueue("smc_tx_wq-%*phN", 0, 0,
782 SMC_LGR_ID_SIZE, &lgr->id);
783 if (!lgr->tx_wq) {
784 rc = -ENOMEM;
785 goto free_lgr;
786 }
787 lgr->is_smcd = ini->is_smcd;
788 lgr->sync_err = 0;
789 lgr->terminating = 0;
790 lgr->freeing = 0;
791 lgr->vlan_id = ini->vlan_id;
792 mutex_init(&lgr->sndbufs_lock);
793 mutex_init(&lgr->rmbs_lock);
794 rwlock_init(&lgr->conns_lock);
795 for (i = 0; i < SMC_RMBE_SIZES; i++) {
796 INIT_LIST_HEAD(&lgr->sndbufs[i]);
797 INIT_LIST_HEAD(&lgr->rmbs[i]);
798 }
799 lgr->next_link_id = 0;
800 smc_lgr_list.num += SMC_LGR_NUM_INCR;
801 memcpy(&lgr->id, (u8 *)&smc_lgr_list.num, SMC_LGR_ID_SIZE);
802 INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
803 INIT_WORK(&lgr->terminate_work, smc_lgr_terminate_work);
804 lgr->conns_all = RB_ROOT;
805 if (ini->is_smcd) {
806 /* SMC-D specific settings */
807 get_device(&ini->ism_dev[ini->ism_selected]->dev);
808 lgr->peer_gid = ini->ism_peer_gid[ini->ism_selected];
809 lgr->smcd = ini->ism_dev[ini->ism_selected];
810 lgr_list = &ini->ism_dev[ini->ism_selected]->lgr_list;
811 lgr_lock = &lgr->smcd->lgr_lock;
812 lgr->smc_version = ini->smcd_version;
813 lgr->peer_shutdown = 0;
814 atomic_inc(&ini->ism_dev[ini->ism_selected]->lgr_cnt);
815 } else {
816 /* SMC-R specific settings */
817 lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
818 memcpy(lgr->peer_systemid, ini->ib_lcl->id_for_peer,
819 SMC_SYSTEMID_LEN);
820 memcpy(lgr->pnet_id, ini->ib_dev->pnetid[ini->ib_port - 1],
821 SMC_MAX_PNETID_LEN);
822 smc_llc_lgr_init(lgr, smc);
823
824 link_idx = SMC_SINGLE_LINK;
825 lnk = &lgr->lnk[link_idx];
826 rc = smcr_link_init(lgr, lnk, link_idx, ini);
827 if (rc)
828 goto free_wq;
829 lgr_list = &smc_lgr_list.list;
830 lgr_lock = &smc_lgr_list.lock;
831 atomic_inc(&lgr_cnt);
832 }
833 smc->conn.lgr = lgr;
834 spin_lock_bh(lgr_lock);
835 list_add_tail(&lgr->list, lgr_list);
836 spin_unlock_bh(lgr_lock);
837 return 0;
838
839free_wq:
840 destroy_workqueue(lgr->tx_wq);
841free_lgr:
842 kfree(lgr);
843ism_put_vlan:
844 if (ini->is_smcd && ini->vlan_id)
845 smc_ism_put_vlan(ini->ism_dev[ini->ism_selected], ini->vlan_id);
846out:
847 if (rc < 0) {
848 if (rc == -ENOMEM)
849 rc = SMC_CLC_DECL_MEM;
850 else
851 rc = SMC_CLC_DECL_INTERR;
852 }
853 return rc;
854}
855
856static int smc_write_space(struct smc_connection *conn)
857{
858 int buffer_len = conn->peer_rmbe_size;
859 union smc_host_cursor prod;
860 union smc_host_cursor cons;
861 int space;
862
863 smc_curs_copy(&prod, &conn->local_tx_ctrl.prod, conn);
864 smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
865 /* determine rx_buf space */
866 space = buffer_len - smc_curs_diff(buffer_len, &cons, &prod);
867 return space;
868}
869
870static int smc_switch_cursor(struct smc_sock *smc, struct smc_cdc_tx_pend *pend,
871 struct smc_wr_buf *wr_buf)
872{
873 struct smc_connection *conn = &smc->conn;
874 union smc_host_cursor cons, fin;
875 int rc = 0;
876 int diff;
877
878 smc_curs_copy(&conn->tx_curs_sent, &conn->tx_curs_fin, conn);
879 smc_curs_copy(&fin, &conn->local_tx_ctrl_fin, conn);
880 /* set prod cursor to old state, enforce tx_rdma_writes() */
881 smc_curs_copy(&conn->local_tx_ctrl.prod, &fin, conn);
882 smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
883
884 if (smc_curs_comp(conn->peer_rmbe_size, &cons, &fin) < 0) {
885 /* cons cursor advanced more than fin, and prod was set
886 * fin above, so now prod is smaller than cons. Fix that.
887 */
888 diff = smc_curs_diff(conn->peer_rmbe_size, &fin, &cons);
889 smc_curs_add(conn->sndbuf_desc->len,
890 &conn->tx_curs_sent, diff);
891 smc_curs_add(conn->sndbuf_desc->len,
892 &conn->tx_curs_fin, diff);
893
894 smp_mb__before_atomic();
895 atomic_add(diff, &conn->sndbuf_space);
896 smp_mb__after_atomic();
897
898 smc_curs_add(conn->peer_rmbe_size,
899 &conn->local_tx_ctrl.prod, diff);
900 smc_curs_add(conn->peer_rmbe_size,
901 &conn->local_tx_ctrl_fin, diff);
902 }
903 /* recalculate, value is used by tx_rdma_writes() */
904 atomic_set(&smc->conn.peer_rmbe_space, smc_write_space(conn));
905
906 if (smc->sk.sk_state != SMC_INIT &&
907 smc->sk.sk_state != SMC_CLOSED) {
908 rc = smcr_cdc_msg_send_validation(conn, pend, wr_buf);
909 if (!rc) {
910 queue_delayed_work(conn->lgr->tx_wq, &conn->tx_work, 0);
911 smc->sk.sk_data_ready(&smc->sk);
912 }
913 } else {
914 smc_wr_tx_put_slot(conn->lnk,
915 (struct smc_wr_tx_pend_priv *)pend);
916 }
917 return rc;
918}
919
920void smc_switch_link_and_count(struct smc_connection *conn,
921 struct smc_link *to_lnk)
922{
923 atomic_dec(&conn->lnk->conn_cnt);
924 conn->lnk = to_lnk;
925 atomic_inc(&conn->lnk->conn_cnt);
926}
927
928struct smc_link *smc_switch_conns(struct smc_link_group *lgr,
929 struct smc_link *from_lnk, bool is_dev_err)
930{
931 struct smc_link *to_lnk = NULL;
932 struct smc_cdc_tx_pend *pend;
933 struct smc_connection *conn;
934 struct smc_wr_buf *wr_buf;
935 struct smc_sock *smc;
936 struct rb_node *node;
937 int i, rc = 0;
938
939 /* link is inactive, wake up tx waiters */
940 smc_wr_wakeup_tx_wait(from_lnk);
941
942 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
943 if (!smc_link_active(&lgr->lnk[i]) || i == from_lnk->link_idx)
944 continue;
945 if (is_dev_err && from_lnk->smcibdev == lgr->lnk[i].smcibdev &&
946 from_lnk->ibport == lgr->lnk[i].ibport) {
947 continue;
948 }
949 to_lnk = &lgr->lnk[i];
950 break;
951 }
952 if (!to_lnk || !smc_wr_tx_link_hold(to_lnk)) {
953 smc_lgr_terminate_sched(lgr);
954 return NULL;
955 }
956again:
957 read_lock_bh(&lgr->conns_lock);
958 for (node = rb_first(&lgr->conns_all); node; node = rb_next(node)) {
959 conn = rb_entry(node, struct smc_connection, alert_node);
960 if (conn->lnk != from_lnk)
961 continue;
962 smc = container_of(conn, struct smc_sock, conn);
963 /* conn->lnk not yet set in SMC_INIT state */
964 if (smc->sk.sk_state == SMC_INIT)
965 continue;
966 if (smc->sk.sk_state == SMC_CLOSED ||
967 smc->sk.sk_state == SMC_PEERCLOSEWAIT1 ||
968 smc->sk.sk_state == SMC_PEERCLOSEWAIT2 ||
969 smc->sk.sk_state == SMC_APPFINCLOSEWAIT ||
970 smc->sk.sk_state == SMC_APPCLOSEWAIT1 ||
971 smc->sk.sk_state == SMC_APPCLOSEWAIT2 ||
972 smc->sk.sk_state == SMC_PEERFINCLOSEWAIT ||
973 smc->sk.sk_state == SMC_PEERABORTWAIT ||
974 smc->sk.sk_state == SMC_PROCESSABORT) {
975 spin_lock_bh(&conn->send_lock);
976 smc_switch_link_and_count(conn, to_lnk);
977 spin_unlock_bh(&conn->send_lock);
978 continue;
979 }
980 sock_hold(&smc->sk);
981 read_unlock_bh(&lgr->conns_lock);
982 /* pre-fetch buffer outside of send_lock, might sleep */
983 rc = smc_cdc_get_free_slot(conn, to_lnk, &wr_buf, NULL, &pend);
984 if (rc)
985 goto err_out;
986 /* avoid race with smcr_tx_sndbuf_nonempty() */
987 spin_lock_bh(&conn->send_lock);
988 smc_switch_link_and_count(conn, to_lnk);
989 rc = smc_switch_cursor(smc, pend, wr_buf);
990 spin_unlock_bh(&conn->send_lock);
991 sock_put(&smc->sk);
992 if (rc)
993 goto err_out;
994 goto again;
995 }
996 read_unlock_bh(&lgr->conns_lock);
997 smc_wr_tx_link_put(to_lnk);
998 return to_lnk;
999
1000err_out:
1001 smcr_link_down_cond_sched(to_lnk);
1002 smc_wr_tx_link_put(to_lnk);
1003 return NULL;
1004}
1005
1006static void smcr_buf_unuse(struct smc_buf_desc *rmb_desc,
1007 struct smc_link_group *lgr)
1008{
1009 int rc;
1010
1011 if (rmb_desc->is_conf_rkey && !list_empty(&lgr->list)) {
1012 /* unregister rmb with peer */
1013 rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY);
1014 if (!rc) {
1015 /* protect against smc_llc_cli_rkey_exchange() */
1016 mutex_lock(&lgr->llc_conf_mutex);
1017 smc_llc_do_delete_rkey(lgr, rmb_desc);
1018 rmb_desc->is_conf_rkey = false;
1019 mutex_unlock(&lgr->llc_conf_mutex);
1020 smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl);
1021 }
1022 }
1023
1024 if (rmb_desc->is_reg_err) {
1025 /* buf registration failed, reuse not possible */
1026 mutex_lock(&lgr->rmbs_lock);
1027 list_del(&rmb_desc->list);
1028 mutex_unlock(&lgr->rmbs_lock);
1029
1030 smc_buf_free(lgr, true, rmb_desc);
1031 } else {
1032 rmb_desc->used = 0;
1033 }
1034}
1035
1036static void smc_buf_unuse(struct smc_connection *conn,
1037 struct smc_link_group *lgr)
1038{
1039 if (conn->sndbuf_desc)
1040 conn->sndbuf_desc->used = 0;
1041 if (conn->rmb_desc && lgr->is_smcd)
1042 conn->rmb_desc->used = 0;
1043 else if (conn->rmb_desc)
1044 smcr_buf_unuse(conn->rmb_desc, lgr);
1045}
1046
1047/* remove a finished connection from its link group */
1048void smc_conn_free(struct smc_connection *conn)
1049{
1050 struct smc_link_group *lgr = conn->lgr;
1051
1052 if (!lgr)
1053 return;
1054 if (lgr->is_smcd) {
1055 if (!list_empty(&lgr->list))
1056 smc_ism_unset_conn(conn);
1057 tasklet_kill(&conn->rx_tsklet);
1058 } else {
1059 smc_cdc_tx_dismiss_slots(conn);
1060 if (current_work() != &conn->abort_work)
1061 cancel_work_sync(&conn->abort_work);
1062 }
1063 if (!list_empty(&lgr->list)) {
1064 smc_lgr_unregister_conn(conn);
1065 smc_buf_unuse(conn, lgr); /* allow buffer reuse */
1066 }
1067
1068 if (!lgr->conns_num)
1069 smc_lgr_schedule_free_work(lgr);
1070}
1071
1072/* unregister a link from a buf_desc */
1073static void smcr_buf_unmap_link(struct smc_buf_desc *buf_desc, bool is_rmb,
1074 struct smc_link *lnk)
1075{
1076 if (is_rmb)
1077 buf_desc->is_reg_mr[lnk->link_idx] = false;
1078 if (!buf_desc->is_map_ib[lnk->link_idx])
1079 return;
1080 if (is_rmb) {
1081 if (buf_desc->mr_rx[lnk->link_idx]) {
1082 smc_ib_put_memory_region(
1083 buf_desc->mr_rx[lnk->link_idx]);
1084 buf_desc->mr_rx[lnk->link_idx] = NULL;
1085 }
1086 smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_FROM_DEVICE);
1087 } else {
1088 smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_TO_DEVICE);
1089 }
1090 sg_free_table(&buf_desc->sgt[lnk->link_idx]);
1091 buf_desc->is_map_ib[lnk->link_idx] = false;
1092}
1093
1094/* unmap all buffers of lgr for a deleted link */
1095static void smcr_buf_unmap_lgr(struct smc_link *lnk)
1096{
1097 struct smc_link_group *lgr = lnk->lgr;
1098 struct smc_buf_desc *buf_desc, *bf;
1099 int i;
1100
1101 for (i = 0; i < SMC_RMBE_SIZES; i++) {
1102 mutex_lock(&lgr->rmbs_lock);
1103 list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list)
1104 smcr_buf_unmap_link(buf_desc, true, lnk);
1105 mutex_unlock(&lgr->rmbs_lock);
1106 mutex_lock(&lgr->sndbufs_lock);
1107 list_for_each_entry_safe(buf_desc, bf, &lgr->sndbufs[i],
1108 list)
1109 smcr_buf_unmap_link(buf_desc, false, lnk);
1110 mutex_unlock(&lgr->sndbufs_lock);
1111 }
1112}
1113
1114static void smcr_rtoken_clear_link(struct smc_link *lnk)
1115{
1116 struct smc_link_group *lgr = lnk->lgr;
1117 int i;
1118
1119 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1120 lgr->rtokens[i][lnk->link_idx].rkey = 0;
1121 lgr->rtokens[i][lnk->link_idx].dma_addr = 0;
1122 }
1123}
1124
1125/* must be called under lgr->llc_conf_mutex lock */
1126void smcr_link_clear(struct smc_link *lnk, bool log)
1127{
1128 struct smc_ib_device *smcibdev;
1129
1130 if (!lnk->lgr || lnk->state == SMC_LNK_UNUSED)
1131 return;
1132 lnk->peer_qpn = 0;
1133 smc_llc_link_clear(lnk, log);
1134 smcr_buf_unmap_lgr(lnk);
1135 smcr_rtoken_clear_link(lnk);
1136 smc_ib_modify_qp_reset(lnk);
1137 smc_wr_free_link(lnk);
1138 smc_ib_destroy_queue_pair(lnk);
1139 smc_ib_dealloc_protection_domain(lnk);
1140 smc_wr_free_link_mem(lnk);
1141 smc_ibdev_cnt_dec(lnk);
1142 put_device(&lnk->smcibdev->ibdev->dev);
1143 smcibdev = lnk->smcibdev;
1144 memset(lnk, 0, sizeof(struct smc_link));
1145 lnk->state = SMC_LNK_UNUSED;
1146 if (!atomic_dec_return(&smcibdev->lnk_cnt))
1147 wake_up(&smcibdev->lnks_deleted);
1148}
1149
1150static void smcr_buf_free(struct smc_link_group *lgr, bool is_rmb,
1151 struct smc_buf_desc *buf_desc)
1152{
1153 int i;
1154
1155 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
1156 smcr_buf_unmap_link(buf_desc, is_rmb, &lgr->lnk[i]);
1157
1158 if (buf_desc->pages)
1159 __free_pages(buf_desc->pages, buf_desc->order);
1160 kfree(buf_desc);
1161}
1162
1163static void smcd_buf_free(struct smc_link_group *lgr, bool is_dmb,
1164 struct smc_buf_desc *buf_desc)
1165{
1166 if (is_dmb) {
1167 /* restore original buf len */
1168 buf_desc->len += sizeof(struct smcd_cdc_msg);
1169 smc_ism_unregister_dmb(lgr->smcd, buf_desc);
1170 } else {
1171 kfree(buf_desc->cpu_addr);
1172 }
1173 kfree(buf_desc);
1174}
1175
1176static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
1177 struct smc_buf_desc *buf_desc)
1178{
1179 if (lgr->is_smcd)
1180 smcd_buf_free(lgr, is_rmb, buf_desc);
1181 else
1182 smcr_buf_free(lgr, is_rmb, buf_desc);
1183}
1184
1185static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
1186{
1187 struct smc_buf_desc *buf_desc, *bf_desc;
1188 struct list_head *buf_list;
1189 int i;
1190
1191 for (i = 0; i < SMC_RMBE_SIZES; i++) {
1192 if (is_rmb)
1193 buf_list = &lgr->rmbs[i];
1194 else
1195 buf_list = &lgr->sndbufs[i];
1196 list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
1197 list) {
1198 list_del(&buf_desc->list);
1199 smc_buf_free(lgr, is_rmb, buf_desc);
1200 }
1201 }
1202}
1203
1204static void smc_lgr_free_bufs(struct smc_link_group *lgr)
1205{
1206 /* free send buffers */
1207 __smc_lgr_free_bufs(lgr, false);
1208 /* free rmbs */
1209 __smc_lgr_free_bufs(lgr, true);
1210}
1211
1212/* remove a link group */
1213static void smc_lgr_free(struct smc_link_group *lgr)
1214{
1215 int i;
1216
1217 if (!lgr->is_smcd) {
1218 mutex_lock(&lgr->llc_conf_mutex);
1219 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1220 if (lgr->lnk[i].state != SMC_LNK_UNUSED)
1221 smcr_link_clear(&lgr->lnk[i], false);
1222 }
1223 mutex_unlock(&lgr->llc_conf_mutex);
1224 smc_llc_lgr_clear(lgr);
1225 }
1226
1227 smc_lgr_free_bufs(lgr);
1228 destroy_workqueue(lgr->tx_wq);
1229 if (lgr->is_smcd) {
1230 smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
1231 put_device(&lgr->smcd->dev);
1232 if (!atomic_dec_return(&lgr->smcd->lgr_cnt))
1233 wake_up(&lgr->smcd->lgrs_deleted);
1234 } else {
1235 if (!atomic_dec_return(&lgr_cnt))
1236 wake_up(&lgrs_deleted);
1237 }
1238 kfree(lgr);
1239}
1240
1241static void smc_sk_wake_ups(struct smc_sock *smc)
1242{
1243 smc->sk.sk_write_space(&smc->sk);
1244 smc->sk.sk_data_ready(&smc->sk);
1245 smc->sk.sk_state_change(&smc->sk);
1246}
1247
1248/* kill a connection */
1249static void smc_conn_kill(struct smc_connection *conn, bool soft)
1250{
1251 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1252
1253 if (conn->lgr->is_smcd && conn->lgr->peer_shutdown)
1254 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
1255 else
1256 smc_close_abort(conn);
1257 conn->killed = 1;
1258 smc->sk.sk_err = ECONNABORTED;
1259 smc_sk_wake_ups(smc);
1260 if (conn->lgr->is_smcd) {
1261 smc_ism_unset_conn(conn);
1262 if (soft)
1263 tasklet_kill(&conn->rx_tsklet);
1264 else
1265 tasklet_unlock_wait(&conn->rx_tsklet);
1266 } else {
1267 smc_cdc_tx_dismiss_slots(conn);
1268 }
1269 smc_lgr_unregister_conn(conn);
1270 smc_close_active_abort(smc);
1271}
1272
1273static void smc_lgr_cleanup(struct smc_link_group *lgr)
1274{
1275 if (lgr->is_smcd) {
1276 smc_ism_signal_shutdown(lgr);
1277 } else {
1278 u32 rsn = lgr->llc_termination_rsn;
1279
1280 if (!rsn)
1281 rsn = SMC_LLC_DEL_PROG_INIT_TERM;
1282 smc_llc_send_link_delete_all(lgr, false, rsn);
1283 smcr_lgr_link_deactivate_all(lgr);
1284 }
1285}
1286
1287/* terminate link group
1288 * @soft: true if link group shutdown can take its time
1289 * false if immediate link group shutdown is required
1290 */
1291static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft)
1292{
1293 struct smc_connection *conn;
1294 struct smc_sock *smc;
1295 struct rb_node *node;
1296
1297 if (lgr->terminating)
1298 return; /* lgr already terminating */
1299 /* cancel free_work sync, will terminate when lgr->freeing is set */
1300 cancel_delayed_work_sync(&lgr->free_work);
1301 lgr->terminating = 1;
1302
1303 /* kill remaining link group connections */
1304 read_lock_bh(&lgr->conns_lock);
1305 node = rb_first(&lgr->conns_all);
1306 while (node) {
1307 read_unlock_bh(&lgr->conns_lock);
1308 conn = rb_entry(node, struct smc_connection, alert_node);
1309 smc = container_of(conn, struct smc_sock, conn);
1310 sock_hold(&smc->sk); /* sock_put below */
1311 lock_sock(&smc->sk);
1312 smc_conn_kill(conn, soft);
1313 release_sock(&smc->sk);
1314 sock_put(&smc->sk); /* sock_hold above */
1315 read_lock_bh(&lgr->conns_lock);
1316 node = rb_first(&lgr->conns_all);
1317 }
1318 read_unlock_bh(&lgr->conns_lock);
1319 smc_lgr_cleanup(lgr);
1320 smc_lgr_free(lgr);
1321}
1322
1323/* unlink link group and schedule termination */
1324void smc_lgr_terminate_sched(struct smc_link_group *lgr)
1325{
1326 spinlock_t *lgr_lock;
1327
1328 smc_lgr_list_head(lgr, &lgr_lock);
1329 spin_lock_bh(lgr_lock);
1330 if (list_empty(&lgr->list) || lgr->terminating || lgr->freeing) {
1331 spin_unlock_bh(lgr_lock);
1332 return; /* lgr already terminating */
1333 }
1334 list_del_init(&lgr->list);
1335 lgr->freeing = 1;
1336 spin_unlock_bh(lgr_lock);
1337 schedule_work(&lgr->terminate_work);
1338}
1339
1340/* Called when peer lgr shutdown (regularly or abnormally) is received */
1341void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid, unsigned short vlan)
1342{
1343 struct smc_link_group *lgr, *l;
1344 LIST_HEAD(lgr_free_list);
1345
1346 /* run common cleanup function and build free list */
1347 spin_lock_bh(&dev->lgr_lock);
1348 list_for_each_entry_safe(lgr, l, &dev->lgr_list, list) {
1349 if ((!peer_gid || lgr->peer_gid == peer_gid) &&
1350 (vlan == VLAN_VID_MASK || lgr->vlan_id == vlan)) {
1351 if (peer_gid) /* peer triggered termination */
1352 lgr->peer_shutdown = 1;
1353 list_move(&lgr->list, &lgr_free_list);
1354 lgr->freeing = 1;
1355 }
1356 }
1357 spin_unlock_bh(&dev->lgr_lock);
1358
1359 /* cancel the regular free workers and actually free lgrs */
1360 list_for_each_entry_safe(lgr, l, &lgr_free_list, list) {
1361 list_del_init(&lgr->list);
1362 schedule_work(&lgr->terminate_work);
1363 }
1364}
1365
1366/* Called when an SMCD device is removed or the smc module is unloaded */
1367void smc_smcd_terminate_all(struct smcd_dev *smcd)
1368{
1369 struct smc_link_group *lgr, *lg;
1370 LIST_HEAD(lgr_free_list);
1371
1372 spin_lock_bh(&smcd->lgr_lock);
1373 list_splice_init(&smcd->lgr_list, &lgr_free_list);
1374 list_for_each_entry(lgr, &lgr_free_list, list)
1375 lgr->freeing = 1;
1376 spin_unlock_bh(&smcd->lgr_lock);
1377
1378 list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
1379 list_del_init(&lgr->list);
1380 __smc_lgr_terminate(lgr, false);
1381 }
1382
1383 if (atomic_read(&smcd->lgr_cnt))
1384 wait_event(smcd->lgrs_deleted, !atomic_read(&smcd->lgr_cnt));
1385}
1386
1387/* Called when an SMCR device is removed or the smc module is unloaded.
1388 * If smcibdev is given, all SMCR link groups using this device are terminated.
1389 * If smcibdev is NULL, all SMCR link groups are terminated.
1390 */
1391void smc_smcr_terminate_all(struct smc_ib_device *smcibdev)
1392{
1393 struct smc_link_group *lgr, *lg;
1394 LIST_HEAD(lgr_free_list);
1395 int i;
1396
1397 spin_lock_bh(&smc_lgr_list.lock);
1398 if (!smcibdev) {
1399 list_splice_init(&smc_lgr_list.list, &lgr_free_list);
1400 list_for_each_entry(lgr, &lgr_free_list, list)
1401 lgr->freeing = 1;
1402 } else {
1403 list_for_each_entry_safe(lgr, lg, &smc_lgr_list.list, list) {
1404 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1405 if (lgr->lnk[i].smcibdev == smcibdev)
1406 smcr_link_down_cond_sched(&lgr->lnk[i]);
1407 }
1408 }
1409 }
1410 spin_unlock_bh(&smc_lgr_list.lock);
1411
1412 list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
1413 list_del_init(&lgr->list);
1414 smc_llc_set_termination_rsn(lgr, SMC_LLC_DEL_OP_INIT_TERM);
1415 __smc_lgr_terminate(lgr, false);
1416 }
1417
1418 if (smcibdev) {
1419 if (atomic_read(&smcibdev->lnk_cnt))
1420 wait_event(smcibdev->lnks_deleted,
1421 !atomic_read(&smcibdev->lnk_cnt));
1422 } else {
1423 if (atomic_read(&lgr_cnt))
1424 wait_event(lgrs_deleted, !atomic_read(&lgr_cnt));
1425 }
1426}
1427
1428/* set new lgr type and clear all asymmetric link tagging */
1429void smcr_lgr_set_type(struct smc_link_group *lgr, enum smc_lgr_type new_type)
1430{
1431 char *lgr_type = "";
1432 int i;
1433
1434 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
1435 if (smc_link_usable(&lgr->lnk[i]))
1436 lgr->lnk[i].link_is_asym = false;
1437 if (lgr->type == new_type)
1438 return;
1439 lgr->type = new_type;
1440
1441 switch (lgr->type) {
1442 case SMC_LGR_NONE:
1443 lgr_type = "NONE";
1444 break;
1445 case SMC_LGR_SINGLE:
1446 lgr_type = "SINGLE";
1447 break;
1448 case SMC_LGR_SYMMETRIC:
1449 lgr_type = "SYMMETRIC";
1450 break;
1451 case SMC_LGR_ASYMMETRIC_PEER:
1452 lgr_type = "ASYMMETRIC_PEER";
1453 break;
1454 case SMC_LGR_ASYMMETRIC_LOCAL:
1455 lgr_type = "ASYMMETRIC_LOCAL";
1456 break;
1457 }
1458 pr_warn_ratelimited("smc: SMC-R lg %*phN state changed: "
1459 "%s, pnetid %.16s\n", SMC_LGR_ID_SIZE, &lgr->id,
1460 lgr_type, lgr->pnet_id);
1461}
1462
1463/* set new lgr type and tag a link as asymmetric */
1464void smcr_lgr_set_type_asym(struct smc_link_group *lgr,
1465 enum smc_lgr_type new_type, int asym_lnk_idx)
1466{
1467 smcr_lgr_set_type(lgr, new_type);
1468 lgr->lnk[asym_lnk_idx].link_is_asym = true;
1469}
1470
1471/* abort connection, abort_work scheduled from tasklet context */
1472static void smc_conn_abort_work(struct work_struct *work)
1473{
1474 struct smc_connection *conn = container_of(work,
1475 struct smc_connection,
1476 abort_work);
1477 struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1478
1479 lock_sock(&smc->sk);
1480 smc_conn_kill(conn, true);
1481 release_sock(&smc->sk);
1482 sock_put(&smc->sk); /* sock_hold done by schedulers of abort_work */
1483}
1484
1485void smcr_port_add(struct smc_ib_device *smcibdev, u8 ibport)
1486{
1487 struct smc_link_group *lgr, *n;
1488
1489 list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1490 struct smc_link *link;
1491
1492 if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1493 SMC_MAX_PNETID_LEN) ||
1494 lgr->type == SMC_LGR_SYMMETRIC ||
1495 lgr->type == SMC_LGR_ASYMMETRIC_PEER)
1496 continue;
1497
1498 /* trigger local add link processing */
1499 link = smc_llc_usable_link(lgr);
1500 if (link)
1501 smc_llc_add_link_local(link);
1502 }
1503}
1504
1505/* link is down - switch connections to alternate link,
1506 * must be called under lgr->llc_conf_mutex lock
1507 */
1508static void smcr_link_down(struct smc_link *lnk)
1509{
1510 struct smc_link_group *lgr = lnk->lgr;
1511 struct smc_link *to_lnk;
1512 int del_link_id;
1513
1514 if (!lgr || lnk->state == SMC_LNK_UNUSED || list_empty(&lgr->list))
1515 return;
1516
1517 smc_ib_modify_qp_reset(lnk);
1518 to_lnk = smc_switch_conns(lgr, lnk, true);
1519 if (!to_lnk) { /* no backup link available */
1520 smcr_link_clear(lnk, true);
1521 return;
1522 }
1523 smcr_lgr_set_type(lgr, SMC_LGR_SINGLE);
1524 del_link_id = lnk->link_id;
1525
1526 if (lgr->role == SMC_SERV) {
1527 /* trigger local delete link processing */
1528 smc_llc_srv_delete_link_local(to_lnk, del_link_id);
1529 } else {
1530 if (lgr->llc_flow_lcl.type != SMC_LLC_FLOW_NONE) {
1531 /* another llc task is ongoing */
1532 mutex_unlock(&lgr->llc_conf_mutex);
1533 wait_event_timeout(lgr->llc_flow_waiter,
1534 (list_empty(&lgr->list) ||
1535 lgr->llc_flow_lcl.type == SMC_LLC_FLOW_NONE),
1536 SMC_LLC_WAIT_TIME);
1537 mutex_lock(&lgr->llc_conf_mutex);
1538 }
1539 if (!list_empty(&lgr->list)) {
1540 smc_llc_send_delete_link(to_lnk, del_link_id,
1541 SMC_LLC_REQ, true,
1542 SMC_LLC_DEL_LOST_PATH);
1543 smcr_link_clear(lnk, true);
1544 }
1545 wake_up(&lgr->llc_flow_waiter); /* wake up next waiter */
1546 }
1547}
1548
1549/* must be called under lgr->llc_conf_mutex lock */
1550void smcr_link_down_cond(struct smc_link *lnk)
1551{
1552 if (smc_link_downing(&lnk->state))
1553 smcr_link_down(lnk);
1554}
1555
1556/* will get the lgr->llc_conf_mutex lock */
1557void smcr_link_down_cond_sched(struct smc_link *lnk)
1558{
1559 if (smc_link_downing(&lnk->state))
1560 schedule_work(&lnk->link_down_wrk);
1561}
1562
1563void smcr_port_err(struct smc_ib_device *smcibdev, u8 ibport)
1564{
1565 struct smc_link_group *lgr, *n;
1566 int i;
1567
1568 list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1569 if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1570 SMC_MAX_PNETID_LEN))
1571 continue; /* lgr is not affected */
1572 if (list_empty(&lgr->list))
1573 continue;
1574 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1575 struct smc_link *lnk = &lgr->lnk[i];
1576
1577 if (smc_link_usable(lnk) &&
1578 lnk->smcibdev == smcibdev && lnk->ibport == ibport)
1579 smcr_link_down_cond_sched(lnk);
1580 }
1581 }
1582}
1583
1584static void smc_link_down_work(struct work_struct *work)
1585{
1586 struct smc_link *link = container_of(work, struct smc_link,
1587 link_down_wrk);
1588 struct smc_link_group *lgr = link->lgr;
1589
1590 if (list_empty(&lgr->list))
1591 return;
1592 wake_up_all(&lgr->llc_msg_waiter);
1593 mutex_lock(&lgr->llc_conf_mutex);
1594 smcr_link_down(link);
1595 mutex_unlock(&lgr->llc_conf_mutex);
1596}
1597
1598/* Determine vlan of internal TCP socket.
1599 * @vlan_id: address to store the determined vlan id into
1600 */
1601int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini)
1602{
1603 struct dst_entry *dst = sk_dst_get(clcsock->sk);
1604 struct net_device *ndev;
1605 int i, nest_lvl, rc = 0;
1606
1607 ini->vlan_id = 0;
1608 if (!dst) {
1609 rc = -ENOTCONN;
1610 goto out;
1611 }
1612 if (!dst->dev) {
1613 rc = -ENODEV;
1614 goto out_rel;
1615 }
1616
1617 ndev = dst->dev;
1618 if (is_vlan_dev(ndev)) {
1619 ini->vlan_id = vlan_dev_vlan_id(ndev);
1620 goto out_rel;
1621 }
1622
1623 rtnl_lock();
1624 nest_lvl = ndev->lower_level;
1625 for (i = 0; i < nest_lvl; i++) {
1626 struct list_head *lower = &ndev->adj_list.lower;
1627
1628 if (list_empty(lower))
1629 break;
1630 lower = lower->next;
1631 ndev = (struct net_device *)netdev_lower_get_next(ndev, &lower);
1632 if (is_vlan_dev(ndev)) {
1633 ini->vlan_id = vlan_dev_vlan_id(ndev);
1634 break;
1635 }
1636 }
1637 rtnl_unlock();
1638
1639out_rel:
1640 dst_release(dst);
1641out:
1642 return rc;
1643}
1644
1645static bool smcr_lgr_match(struct smc_link_group *lgr,
1646 struct smc_clc_msg_local *lcl,
1647 enum smc_lgr_role role, u32 clcqpn)
1648{
1649 int i;
1650
1651 if (memcmp(lgr->peer_systemid, lcl->id_for_peer, SMC_SYSTEMID_LEN) ||
1652 lgr->role != role)
1653 return false;
1654
1655 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1656 if (!smc_link_active(&lgr->lnk[i]))
1657 continue;
1658 if ((lgr->role == SMC_SERV || lgr->lnk[i].peer_qpn == clcqpn) &&
1659 !memcmp(lgr->lnk[i].peer_gid, &lcl->gid, SMC_GID_SIZE) &&
1660 !memcmp(lgr->lnk[i].peer_mac, lcl->mac, sizeof(lcl->mac)))
1661 return true;
1662 }
1663 return false;
1664}
1665
1666static bool smcd_lgr_match(struct smc_link_group *lgr,
1667 struct smcd_dev *smcismdev, u64 peer_gid)
1668{
1669 return lgr->peer_gid == peer_gid && lgr->smcd == smcismdev;
1670}
1671
1672/* create a new SMC connection (and a new link group if necessary) */
1673int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini)
1674{
1675 struct smc_connection *conn = &smc->conn;
1676 struct list_head *lgr_list;
1677 struct smc_link_group *lgr;
1678 enum smc_lgr_role role;
1679 spinlock_t *lgr_lock;
1680 int rc = 0;
1681
1682 lgr_list = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_list :
1683 &smc_lgr_list.list;
1684 lgr_lock = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_lock :
1685 &smc_lgr_list.lock;
1686 ini->first_contact_local = 1;
1687 role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
1688 if (role == SMC_CLNT && ini->first_contact_peer)
1689 /* create new link group as well */
1690 goto create;
1691
1692 /* determine if an existing link group can be reused */
1693 spin_lock_bh(lgr_lock);
1694 list_for_each_entry(lgr, lgr_list, list) {
1695 write_lock_bh(&lgr->conns_lock);
1696 if ((ini->is_smcd ?
1697 smcd_lgr_match(lgr, ini->ism_dev[ini->ism_selected],
1698 ini->ism_peer_gid[ini->ism_selected]) :
1699 smcr_lgr_match(lgr, ini->ib_lcl, role, ini->ib_clcqpn)) &&
1700 !lgr->sync_err &&
1701 (ini->smcd_version == SMC_V2 ||
1702 lgr->vlan_id == ini->vlan_id) &&
1703 (role == SMC_CLNT || ini->is_smcd ||
1704 lgr->conns_num < SMC_RMBS_PER_LGR_MAX)) {
1705 /* link group found */
1706 ini->first_contact_local = 0;
1707 conn->lgr = lgr;
1708 rc = smc_lgr_register_conn(conn, false);
1709 write_unlock_bh(&lgr->conns_lock);
1710 if (!rc && delayed_work_pending(&lgr->free_work))
1711 cancel_delayed_work(&lgr->free_work);
1712 break;
1713 }
1714 write_unlock_bh(&lgr->conns_lock);
1715 }
1716 spin_unlock_bh(lgr_lock);
1717 if (rc)
1718 return rc;
1719
1720 if (role == SMC_CLNT && !ini->first_contact_peer &&
1721 ini->first_contact_local) {
1722 /* Server reuses a link group, but Client wants to start
1723 * a new one
1724 * send out_of_sync decline, reason synchr. error
1725 */
1726 return SMC_CLC_DECL_SYNCERR;
1727 }
1728
1729create:
1730 if (ini->first_contact_local) {
1731 rc = smc_lgr_create(smc, ini);
1732 if (rc)
1733 goto out;
1734 lgr = conn->lgr;
1735 write_lock_bh(&lgr->conns_lock);
1736 rc = smc_lgr_register_conn(conn, true);
1737 write_unlock_bh(&lgr->conns_lock);
1738 if (rc)
1739 goto out;
1740 }
1741 conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
1742 conn->local_tx_ctrl.len = SMC_WR_TX_SIZE;
1743 conn->urg_state = SMC_URG_READ;
1744 INIT_WORK(&smc->conn.abort_work, smc_conn_abort_work);
1745 if (ini->is_smcd) {
1746 conn->rx_off = sizeof(struct smcd_cdc_msg);
1747 smcd_cdc_rx_init(conn); /* init tasklet for this conn */
1748 } else {
1749 conn->rx_off = 0;
1750 }
1751#ifndef KERNEL_HAS_ATOMIC64
1752 spin_lock_init(&conn->acurs_lock);
1753#endif
1754
1755out:
1756 return rc;
1757}
1758
1759/* convert the RMB size into the compressed notation - minimum 16K.
1760 * In contrast to plain ilog2, this rounds towards the next power of 2,
1761 * so the socket application gets at least its desired sndbuf / rcvbuf size.
1762 */
1763static u8 smc_compress_bufsize(int size)
1764{
1765 u8 compressed;
1766
1767 if (size <= SMC_BUF_MIN_SIZE)
1768 return 0;
1769
1770 size = (size - 1) >> 14;
1771 compressed = ilog2(size) + 1;
1772 if (compressed >= SMC_RMBE_SIZES)
1773 compressed = SMC_RMBE_SIZES - 1;
1774 return compressed;
1775}
1776
1777/* convert the RMB size from compressed notation into integer */
1778int smc_uncompress_bufsize(u8 compressed)
1779{
1780 u32 size;
1781
1782 size = 0x00000001 << (((int)compressed) + 14);
1783 return (int)size;
1784}
1785
1786/* try to reuse a sndbuf or rmb description slot for a certain
1787 * buffer size; if not available, return NULL
1788 */
1789static struct smc_buf_desc *smc_buf_get_slot(int compressed_bufsize,
1790 struct mutex *lock,
1791 struct list_head *buf_list)
1792{
1793 struct smc_buf_desc *buf_slot;
1794
1795 mutex_lock(lock);
1796 list_for_each_entry(buf_slot, buf_list, list) {
1797 if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
1798 mutex_unlock(lock);
1799 return buf_slot;
1800 }
1801 }
1802 mutex_unlock(lock);
1803 return NULL;
1804}
1805
1806/* one of the conditions for announcing a receiver's current window size is
1807 * that it "results in a minimum increase in the window size of 10% of the
1808 * receive buffer space" [RFC7609]
1809 */
1810static inline int smc_rmb_wnd_update_limit(int rmbe_size)
1811{
1812 return min_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
1813}
1814
1815/* map an rmb buf to a link */
1816static int smcr_buf_map_link(struct smc_buf_desc *buf_desc, bool is_rmb,
1817 struct smc_link *lnk)
1818{
1819 int rc;
1820
1821 if (buf_desc->is_map_ib[lnk->link_idx])
1822 return 0;
1823
1824 rc = sg_alloc_table(&buf_desc->sgt[lnk->link_idx], 1, GFP_KERNEL);
1825 if (rc)
1826 return rc;
1827 sg_set_buf(buf_desc->sgt[lnk->link_idx].sgl,
1828 buf_desc->cpu_addr, buf_desc->len);
1829
1830 /* map sg table to DMA address */
1831 rc = smc_ib_buf_map_sg(lnk, buf_desc,
1832 is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
1833 /* SMC protocol depends on mapping to one DMA address only */
1834 if (rc != 1) {
1835 rc = -EAGAIN;
1836 goto free_table;
1837 }
1838
1839 /* create a new memory region for the RMB */
1840 if (is_rmb) {
1841 rc = smc_ib_get_memory_region(lnk->roce_pd,
1842 IB_ACCESS_REMOTE_WRITE |
1843 IB_ACCESS_LOCAL_WRITE,
1844 buf_desc, lnk->link_idx);
1845 if (rc)
1846 goto buf_unmap;
1847 smc_ib_sync_sg_for_device(lnk, buf_desc, DMA_FROM_DEVICE);
1848 }
1849 buf_desc->is_map_ib[lnk->link_idx] = true;
1850 return 0;
1851
1852buf_unmap:
1853 smc_ib_buf_unmap_sg(lnk, buf_desc,
1854 is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
1855free_table:
1856 sg_free_table(&buf_desc->sgt[lnk->link_idx]);
1857 return rc;
1858}
1859
1860/* register a new rmb on IB device,
1861 * must be called under lgr->llc_conf_mutex lock
1862 */
1863int smcr_link_reg_rmb(struct smc_link *link, struct smc_buf_desc *rmb_desc)
1864{
1865 if (list_empty(&link->lgr->list))
1866 return -ENOLINK;
1867 if (!rmb_desc->is_reg_mr[link->link_idx]) {
1868 /* register memory region for new rmb */
1869 if (smc_wr_reg_send(link, rmb_desc->mr_rx[link->link_idx])) {
1870 rmb_desc->is_reg_err = true;
1871 return -EFAULT;
1872 }
1873 rmb_desc->is_reg_mr[link->link_idx] = true;
1874 }
1875 return 0;
1876}
1877
1878static int _smcr_buf_map_lgr(struct smc_link *lnk, struct mutex *lock,
1879 struct list_head *lst, bool is_rmb)
1880{
1881 struct smc_buf_desc *buf_desc, *bf;
1882 int rc = 0;
1883
1884 mutex_lock(lock);
1885 list_for_each_entry_safe(buf_desc, bf, lst, list) {
1886 if (!buf_desc->used)
1887 continue;
1888 rc = smcr_buf_map_link(buf_desc, is_rmb, lnk);
1889 if (rc)
1890 goto out;
1891 }
1892out:
1893 mutex_unlock(lock);
1894 return rc;
1895}
1896
1897/* map all used buffers of lgr for a new link */
1898int smcr_buf_map_lgr(struct smc_link *lnk)
1899{
1900 struct smc_link_group *lgr = lnk->lgr;
1901 int i, rc = 0;
1902
1903 for (i = 0; i < SMC_RMBE_SIZES; i++) {
1904 rc = _smcr_buf_map_lgr(lnk, &lgr->rmbs_lock,
1905 &lgr->rmbs[i], true);
1906 if (rc)
1907 return rc;
1908 rc = _smcr_buf_map_lgr(lnk, &lgr->sndbufs_lock,
1909 &lgr->sndbufs[i], false);
1910 if (rc)
1911 return rc;
1912 }
1913 return 0;
1914}
1915
1916/* register all used buffers of lgr for a new link,
1917 * must be called under lgr->llc_conf_mutex lock
1918 */
1919int smcr_buf_reg_lgr(struct smc_link *lnk)
1920{
1921 struct smc_link_group *lgr = lnk->lgr;
1922 struct smc_buf_desc *buf_desc, *bf;
1923 int i, rc = 0;
1924
1925 mutex_lock(&lgr->rmbs_lock);
1926 for (i = 0; i < SMC_RMBE_SIZES; i++) {
1927 list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list) {
1928 if (!buf_desc->used)
1929 continue;
1930 rc = smcr_link_reg_rmb(lnk, buf_desc);
1931 if (rc)
1932 goto out;
1933 }
1934 }
1935out:
1936 mutex_unlock(&lgr->rmbs_lock);
1937 return rc;
1938}
1939
1940static struct smc_buf_desc *smcr_new_buf_create(struct smc_link_group *lgr,
1941 bool is_rmb, int bufsize)
1942{
1943 struct smc_buf_desc *buf_desc;
1944
1945 /* try to alloc a new buffer */
1946 buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
1947 if (!buf_desc)
1948 return ERR_PTR(-ENOMEM);
1949
1950 buf_desc->order = get_order(bufsize);
1951 buf_desc->pages = alloc_pages(GFP_KERNEL | __GFP_NOWARN |
1952 __GFP_NOMEMALLOC | __GFP_COMP |
1953 __GFP_NORETRY | __GFP_ZERO,
1954 buf_desc->order);
1955 if (!buf_desc->pages) {
1956 kfree(buf_desc);
1957 return ERR_PTR(-EAGAIN);
1958 }
1959 buf_desc->cpu_addr = (void *)page_address(buf_desc->pages);
1960 buf_desc->len = bufsize;
1961 return buf_desc;
1962}
1963
1964/* map buf_desc on all usable links,
1965 * unused buffers stay mapped as long as the link is up
1966 */
1967static int smcr_buf_map_usable_links(struct smc_link_group *lgr,
1968 struct smc_buf_desc *buf_desc, bool is_rmb)
1969{
1970 int i, rc = 0;
1971
1972 /* protect against parallel link reconfiguration */
1973 mutex_lock(&lgr->llc_conf_mutex);
1974 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1975 struct smc_link *lnk = &lgr->lnk[i];
1976
1977 if (!smc_link_usable(lnk))
1978 continue;
1979 if (smcr_buf_map_link(buf_desc, is_rmb, lnk)) {
1980 rc = -ENOMEM;
1981 goto out;
1982 }
1983 }
1984out:
1985 mutex_unlock(&lgr->llc_conf_mutex);
1986 return rc;
1987}
1988
1989#define SMCD_DMBE_SIZES 6 /* 0 -> 16KB, 1 -> 32KB, .. 6 -> 1MB */
1990
1991static struct smc_buf_desc *smcd_new_buf_create(struct smc_link_group *lgr,
1992 bool is_dmb, int bufsize)
1993{
1994 struct smc_buf_desc *buf_desc;
1995 int rc;
1996
1997 if (smc_compress_bufsize(bufsize) > SMCD_DMBE_SIZES)
1998 return ERR_PTR(-EAGAIN);
1999
2000 /* try to alloc a new DMB */
2001 buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
2002 if (!buf_desc)
2003 return ERR_PTR(-ENOMEM);
2004 if (is_dmb) {
2005 rc = smc_ism_register_dmb(lgr, bufsize, buf_desc);
2006 if (rc) {
2007 kfree(buf_desc);
2008 if (rc == -ENOMEM)
2009 return ERR_PTR(-EAGAIN);
2010 if (rc == -ENOSPC)
2011 return ERR_PTR(-ENOSPC);
2012 return ERR_PTR(-EIO);
2013 }
2014 buf_desc->pages = virt_to_page(buf_desc->cpu_addr);
2015 /* CDC header stored in buf. So, pretend it was smaller */
2016 buf_desc->len = bufsize - sizeof(struct smcd_cdc_msg);
2017 } else {
2018 buf_desc->cpu_addr = kzalloc(bufsize, GFP_KERNEL |
2019 __GFP_NOWARN | __GFP_NORETRY |
2020 __GFP_NOMEMALLOC);
2021 if (!buf_desc->cpu_addr) {
2022 kfree(buf_desc);
2023 return ERR_PTR(-EAGAIN);
2024 }
2025 buf_desc->len = bufsize;
2026 }
2027 return buf_desc;
2028}
2029
2030static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
2031{
2032 struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
2033 struct smc_connection *conn = &smc->conn;
2034 struct smc_link_group *lgr = conn->lgr;
2035 struct list_head *buf_list;
2036 int bufsize, bufsize_short;
2037 bool is_dgraded = false;
2038 struct mutex *lock; /* lock buffer list */
2039 int sk_buf_size;
2040
2041 if (is_rmb)
2042 /* use socket recv buffer size (w/o overhead) as start value */
2043 sk_buf_size = smc->sk.sk_rcvbuf / 2;
2044 else
2045 /* use socket send buffer size (w/o overhead) as start value */
2046 sk_buf_size = smc->sk.sk_sndbuf / 2;
2047
2048 for (bufsize_short = smc_compress_bufsize(sk_buf_size);
2049 bufsize_short >= 0; bufsize_short--) {
2050
2051 if (is_rmb) {
2052 lock = &lgr->rmbs_lock;
2053 buf_list = &lgr->rmbs[bufsize_short];
2054 } else {
2055 lock = &lgr->sndbufs_lock;
2056 buf_list = &lgr->sndbufs[bufsize_short];
2057 }
2058 bufsize = smc_uncompress_bufsize(bufsize_short);
2059 if ((1 << get_order(bufsize)) > SG_MAX_SINGLE_ALLOC)
2060 continue;
2061
2062 /* check for reusable slot in the link group */
2063 buf_desc = smc_buf_get_slot(bufsize_short, lock, buf_list);
2064 if (buf_desc) {
2065 SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, bufsize);
2066 SMC_STAT_BUF_REUSE(smc, is_smcd, is_rmb);
2067 memset(buf_desc->cpu_addr, 0, bufsize);
2068 break; /* found reusable slot */
2069 }
2070
2071 if (is_smcd)
2072 buf_desc = smcd_new_buf_create(lgr, is_rmb, bufsize);
2073 else
2074 buf_desc = smcr_new_buf_create(lgr, is_rmb, bufsize);
2075
2076 if (PTR_ERR(buf_desc) == -ENOMEM)
2077 break;
2078 if (IS_ERR(buf_desc)) {
2079 if (!is_dgraded) {
2080 is_dgraded = true;
2081 SMC_STAT_RMB_DOWNGRADED(smc, is_smcd, is_rmb);
2082 }
2083 continue;
2084 }
2085
2086 SMC_STAT_RMB_ALLOC(smc, is_smcd, is_rmb);
2087 SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, bufsize);
2088 buf_desc->used = 1;
2089 mutex_lock(lock);
2090 list_add(&buf_desc->list, buf_list);
2091 mutex_unlock(lock);
2092 break; /* found */
2093 }
2094
2095 if (IS_ERR(buf_desc))
2096 return PTR_ERR(buf_desc);
2097
2098 if (!is_smcd) {
2099 if (smcr_buf_map_usable_links(lgr, buf_desc, is_rmb)) {
2100 smcr_buf_unuse(buf_desc, lgr);
2101 return -ENOMEM;
2102 }
2103 }
2104
2105 if (is_rmb) {
2106 conn->rmb_desc = buf_desc;
2107 conn->rmbe_size_short = bufsize_short;
2108 smc->sk.sk_rcvbuf = bufsize * 2;
2109 atomic_set(&conn->bytes_to_rcv, 0);
2110 conn->rmbe_update_limit =
2111 smc_rmb_wnd_update_limit(buf_desc->len);
2112 if (is_smcd)
2113 smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */
2114 } else {
2115 conn->sndbuf_desc = buf_desc;
2116 smc->sk.sk_sndbuf = bufsize * 2;
2117 atomic_set(&conn->sndbuf_space, bufsize);
2118 }
2119 return 0;
2120}
2121
2122void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn)
2123{
2124 if (!conn->lgr || conn->lgr->is_smcd || !smc_link_active(conn->lnk))
2125 return;
2126 smc_ib_sync_sg_for_cpu(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE);
2127}
2128
2129void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
2130{
2131 if (!conn->lgr || conn->lgr->is_smcd || !smc_link_active(conn->lnk))
2132 return;
2133 smc_ib_sync_sg_for_device(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE);
2134}
2135
2136void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
2137{
2138 int i;
2139
2140 if (!conn->lgr || conn->lgr->is_smcd)
2141 return;
2142 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
2143 if (!smc_link_active(&conn->lgr->lnk[i]))
2144 continue;
2145 smc_ib_sync_sg_for_cpu(&conn->lgr->lnk[i], conn->rmb_desc,
2146 DMA_FROM_DEVICE);
2147 }
2148}
2149
2150void smc_rmb_sync_sg_for_device(struct smc_connection *conn)
2151{
2152 int i;
2153
2154 if (!conn->lgr || conn->lgr->is_smcd)
2155 return;
2156 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
2157 if (!smc_link_active(&conn->lgr->lnk[i]))
2158 continue;
2159 smc_ib_sync_sg_for_device(&conn->lgr->lnk[i], conn->rmb_desc,
2160 DMA_FROM_DEVICE);
2161 }
2162}
2163
2164/* create the send and receive buffer for an SMC socket;
2165 * receive buffers are called RMBs;
2166 * (even though the SMC protocol allows more than one RMB-element per RMB,
2167 * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
2168 * extra RMB for every connection in a link group
2169 */
2170int smc_buf_create(struct smc_sock *smc, bool is_smcd)
2171{
2172 int rc;
2173
2174 /* create send buffer */
2175 rc = __smc_buf_create(smc, is_smcd, false);
2176 if (rc)
2177 return rc;
2178 /* create rmb */
2179 rc = __smc_buf_create(smc, is_smcd, true);
2180 if (rc) {
2181 mutex_lock(&smc->conn.lgr->sndbufs_lock);
2182 list_del(&smc->conn.sndbuf_desc->list);
2183 mutex_unlock(&smc->conn.lgr->sndbufs_lock);
2184 smc_buf_free(smc->conn.lgr, false, smc->conn.sndbuf_desc);
2185 smc->conn.sndbuf_desc = NULL;
2186 }
2187 return rc;
2188}
2189
2190static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
2191{
2192 int i;
2193
2194 for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
2195 if (!test_and_set_bit(i, lgr->rtokens_used_mask))
2196 return i;
2197 }
2198 return -ENOSPC;
2199}
2200
2201static int smc_rtoken_find_by_link(struct smc_link_group *lgr, int lnk_idx,
2202 u32 rkey)
2203{
2204 int i;
2205
2206 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2207 if (test_bit(i, lgr->rtokens_used_mask) &&
2208 lgr->rtokens[i][lnk_idx].rkey == rkey)
2209 return i;
2210 }
2211 return -ENOENT;
2212}
2213
2214/* set rtoken for a new link to an existing rmb */
2215void smc_rtoken_set(struct smc_link_group *lgr, int link_idx, int link_idx_new,
2216 __be32 nw_rkey_known, __be64 nw_vaddr, __be32 nw_rkey)
2217{
2218 int rtok_idx;
2219
2220 rtok_idx = smc_rtoken_find_by_link(lgr, link_idx, ntohl(nw_rkey_known));
2221 if (rtok_idx == -ENOENT)
2222 return;
2223 lgr->rtokens[rtok_idx][link_idx_new].rkey = ntohl(nw_rkey);
2224 lgr->rtokens[rtok_idx][link_idx_new].dma_addr = be64_to_cpu(nw_vaddr);
2225}
2226
2227/* set rtoken for a new link whose link_id is given */
2228void smc_rtoken_set2(struct smc_link_group *lgr, int rtok_idx, int link_id,
2229 __be64 nw_vaddr, __be32 nw_rkey)
2230{
2231 u64 dma_addr = be64_to_cpu(nw_vaddr);
2232 u32 rkey = ntohl(nw_rkey);
2233 bool found = false;
2234 int link_idx;
2235
2236 for (link_idx = 0; link_idx < SMC_LINKS_PER_LGR_MAX; link_idx++) {
2237 if (lgr->lnk[link_idx].link_id == link_id) {
2238 found = true;
2239 break;
2240 }
2241 }
2242 if (!found)
2243 return;
2244 lgr->rtokens[rtok_idx][link_idx].rkey = rkey;
2245 lgr->rtokens[rtok_idx][link_idx].dma_addr = dma_addr;
2246}
2247
2248/* add a new rtoken from peer */
2249int smc_rtoken_add(struct smc_link *lnk, __be64 nw_vaddr, __be32 nw_rkey)
2250{
2251 struct smc_link_group *lgr = smc_get_lgr(lnk);
2252 u64 dma_addr = be64_to_cpu(nw_vaddr);
2253 u32 rkey = ntohl(nw_rkey);
2254 int i;
2255
2256 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2257 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
2258 lgr->rtokens[i][lnk->link_idx].dma_addr == dma_addr &&
2259 test_bit(i, lgr->rtokens_used_mask)) {
2260 /* already in list */
2261 return i;
2262 }
2263 }
2264 i = smc_rmb_reserve_rtoken_idx(lgr);
2265 if (i < 0)
2266 return i;
2267 lgr->rtokens[i][lnk->link_idx].rkey = rkey;
2268 lgr->rtokens[i][lnk->link_idx].dma_addr = dma_addr;
2269 return i;
2270}
2271
2272/* delete an rtoken from all links */
2273int smc_rtoken_delete(struct smc_link *lnk, __be32 nw_rkey)
2274{
2275 struct smc_link_group *lgr = smc_get_lgr(lnk);
2276 u32 rkey = ntohl(nw_rkey);
2277 int i, j;
2278
2279 for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2280 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
2281 test_bit(i, lgr->rtokens_used_mask)) {
2282 for (j = 0; j < SMC_LINKS_PER_LGR_MAX; j++) {
2283 lgr->rtokens[i][j].rkey = 0;
2284 lgr->rtokens[i][j].dma_addr = 0;
2285 }
2286 clear_bit(i, lgr->rtokens_used_mask);
2287 return 0;
2288 }
2289 }
2290 return -ENOENT;
2291}
2292
2293/* save rkey and dma_addr received from peer during clc handshake */
2294int smc_rmb_rtoken_handling(struct smc_connection *conn,
2295 struct smc_link *lnk,
2296 struct smc_clc_msg_accept_confirm *clc)
2297{
2298 conn->rtoken_idx = smc_rtoken_add(lnk, clc->r0.rmb_dma_addr,
2299 clc->r0.rmb_rkey);
2300 if (conn->rtoken_idx < 0)
2301 return conn->rtoken_idx;
2302 return 0;
2303}
2304
2305static void smc_core_going_away(void)
2306{
2307 struct smc_ib_device *smcibdev;
2308 struct smcd_dev *smcd;
2309
2310 mutex_lock(&smc_ib_devices.mutex);
2311 list_for_each_entry(smcibdev, &smc_ib_devices.list, list) {
2312 int i;
2313
2314 for (i = 0; i < SMC_MAX_PORTS; i++)
2315 set_bit(i, smcibdev->ports_going_away);
2316 }
2317 mutex_unlock(&smc_ib_devices.mutex);
2318
2319 mutex_lock(&smcd_dev_list.mutex);
2320 list_for_each_entry(smcd, &smcd_dev_list.list, list) {
2321 smcd->going_away = 1;
2322 }
2323 mutex_unlock(&smcd_dev_list.mutex);
2324}
2325
2326/* Clean up all SMC link groups */
2327static void smc_lgrs_shutdown(void)
2328{
2329 struct smcd_dev *smcd;
2330
2331 smc_core_going_away();
2332
2333 smc_smcr_terminate_all(NULL);
2334
2335 mutex_lock(&smcd_dev_list.mutex);
2336 list_for_each_entry(smcd, &smcd_dev_list.list, list)
2337 smc_smcd_terminate_all(smcd);
2338 mutex_unlock(&smcd_dev_list.mutex);
2339}
2340
2341static int smc_core_reboot_event(struct notifier_block *this,
2342 unsigned long event, void *ptr)
2343{
2344 smc_lgrs_shutdown();
2345 smc_ib_unregister_client();
2346 return 0;
2347}
2348
2349static struct notifier_block smc_reboot_notifier = {
2350 .notifier_call = smc_core_reboot_event,
2351};
2352
2353int __init smc_core_init(void)
2354{
2355 return register_reboot_notifier(&smc_reboot_notifier);
2356}
2357
2358/* Called (from smc_exit) when module is removed */
2359void smc_core_exit(void)
2360{
2361 unregister_reboot_notifier(&smc_reboot_notifier);
2362 smc_lgrs_shutdown();
2363}