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