net/smc/smc_core.c

   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
  42 struct 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
  48 static atomic_t lgr_cnt = ATOMIC_INIT(0); /* number of existing link groups */
  49 static DECLARE_WAIT_QUEUE_HEAD(lgrs_deleted);
  50
  51 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
  52                          struct smc_buf_desc *buf_desc);
  53 static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft);
  54
  55 static 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 */
  58 static 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
  70 static void smc_ibdev_cnt_inc(struct smc_link *lnk)
  71 {
  72         atomic_inc(&lnk->smcibdev->lnk_cnt_by_port[lnk->ibport - 1]);
  73 }
  74
  75 static void smc_ibdev_cnt_dec(struct smc_link *lnk)
  76 {
  77         atomic_dec(&lnk->smcibdev->lnk_cnt_by_port[lnk->ibport - 1]);
  78 }
  79
  80 static 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  */
 100 static 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 */
 122 static 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  */
 165 static 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  */
 192 static 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  */
 207 static 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
 221 int 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
 272 errattr:
 273         nla_nest_cancel(skb, attrs);
 274 errout:
 275         genlmsg_cancel(skb, nlh);
 276 errmsg:
 277         return skb->len;
 278 }
 279
 280 static 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;
 308 errattr:
 309         nla_nest_cancel(skb, attrs);
 310 errout:
 311         return -EMSGSIZE;
 312 }
 313
 314 static 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;
 367 errattr:
 368         nla_nest_cancel(skb, attrs);
 369 errout:
 370         genlmsg_cancel(skb, nlh);
 371 errmsg:
 372         return -EMSGSIZE;
 373 }
 374
 375 static 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         }
 400 out:
 401         return 0;
 402
 403 errout:
 404         genlmsg_cancel(skb, nlh);
 405 errmsg:
 406         return -EMSGSIZE;
 407 }
 408
 409 static 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;
 425 next:
 426                 num++;
 427         }
 428 errout:
 429         spin_unlock_bh(&smc_lgr->lock);
 430         cb_ctx->pos[0] = num;
 431 }
 432
 433 static 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
 496 errv2attr:
 497         nla_nest_cancel(skb, v2_attrs);
 498 errattr:
 499         nla_nest_cancel(skb, attrs);
 500 errout:
 501         genlmsg_cancel(skb, nlh);
 502 errmsg:
 503         return -EMSGSIZE;
 504 }
 505
 506 static 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;
 524 next:
 525                 num++;
 526         }
 527 errout:
 528         spin_unlock_bh(&dev->lgr_lock);
 529         cb_ctx->pos[1] = num;
 530         return rc;
 531 }
 532
 533 static 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;
 551 next:
 552                 num++;
 553         }
 554 errout:
 555         mutex_unlock(&dev_list->mutex);
 556         cb_ctx->pos[0] = num;
 557         return rc;
 558 }
 559
 560 int 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
 568 int 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
 576 int 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
 582 void 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
 601 static 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
 615 static void smc_lgr_free(struct smc_link_group *lgr);
 616
 617 static 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
 653 static 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 */
 662 static u8 smcr_next_link_id(struct smc_link_group *lgr)
 663 {
 664         u8 link_id;
 665         int i;
 666
 667         while (1) {
 668 again:
 669                 link_id = ++lgr->next_link_id;
 670                 if (!link_id)   /* skip zero as link_id */
 671                         link_id = ++lgr->next_link_id;
 672                 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 673                         if (smc_link_usable(&lgr->lnk[i]) &&
 674                             lgr->lnk[i].link_id == link_id)
 675                                 goto again;
 676                 }
 677                 break;
 678         }
 679         return link_id;
 680 }
 681
 682 static void smcr_copy_dev_info_to_link(struct smc_link *link)
 683 {
 684         struct smc_ib_device *smcibdev = link->smcibdev;
 685
 686         snprintf(link->ibname, sizeof(link->ibname), "%s",
 687                  smcibdev->ibdev->name);
 688         link->ndev_ifidx = smcibdev->ndev_ifidx[link->ibport - 1];
 689 }
 690
 691 int smcr_link_init(struct smc_link_group *lgr, struct smc_link *lnk,
 692                    u8 link_idx, struct smc_init_info *ini)
 693 {
 694         u8 rndvec[3];
 695         int rc;
 696
 697         get_device(&ini->ib_dev->ibdev->dev);
 698         atomic_inc(&ini->ib_dev->lnk_cnt);
 699         lnk->link_id = smcr_next_link_id(lgr);
 700         lnk->lgr = lgr;
 701         lnk->link_idx = link_idx;
 702         lnk->smcibdev = ini->ib_dev;
 703         lnk->ibport = ini->ib_port;
 704         smc_ibdev_cnt_inc(lnk);
 705         smcr_copy_dev_info_to_link(lnk);
 706         lnk->path_mtu = ini->ib_dev->pattr[ini->ib_port - 1].active_mtu;
 707         atomic_set(&lnk->conn_cnt, 0);
 708         smc_llc_link_set_uid(lnk);
 709         INIT_WORK(&lnk->link_down_wrk, smc_link_down_work);
 710         if (!ini->ib_dev->initialized) {
 711                 rc = (int)smc_ib_setup_per_ibdev(ini->ib_dev);
 712                 if (rc)
 713                         goto out;
 714         }
 715         get_random_bytes(rndvec, sizeof(rndvec));
 716         lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) +
 717                 (rndvec[2] << 16);
 718         rc = smc_ib_determine_gid(lnk->smcibdev, lnk->ibport,
 719                                   ini->vlan_id, lnk->gid, &lnk->sgid_index);
 720         if (rc)
 721                 goto out;
 722         rc = smc_llc_link_init(lnk);
 723         if (rc)
 724                 goto out;
 725         rc = smc_wr_alloc_link_mem(lnk);
 726         if (rc)
 727                 goto clear_llc_lnk;
 728         rc = smc_ib_create_protection_domain(lnk);
 729         if (rc)
 730                 goto free_link_mem;
 731         rc = smc_ib_create_queue_pair(lnk);
 732         if (rc)
 733                 goto dealloc_pd;
 734         rc = smc_wr_create_link(lnk);
 735         if (rc)
 736                 goto destroy_qp;
 737         lnk->state = SMC_LNK_ACTIVATING;
 738         return 0;
 739
 740 destroy_qp:
 741         smc_ib_destroy_queue_pair(lnk);
 742 dealloc_pd:
 743         smc_ib_dealloc_protection_domain(lnk);
 744 free_link_mem:
 745         smc_wr_free_link_mem(lnk);
 746 clear_llc_lnk:
 747         smc_llc_link_clear(lnk, false);
 748 out:
 749         smc_ibdev_cnt_dec(lnk);
 750         put_device(&ini->ib_dev->ibdev->dev);
 751         memset(lnk, 0, sizeof(struct smc_link));
 752         lnk->state = SMC_LNK_UNUSED;
 753         if (!atomic_dec_return(&ini->ib_dev->lnk_cnt))
 754                 wake_up(&ini->ib_dev->lnks_deleted);
 755         return rc;
 756 }
 757
 758 /* create a new SMC link group */
 759 static int smc_lgr_create(struct smc_sock *smc, struct smc_init_info *ini)
 760 {
 761         struct smc_link_group *lgr;
 762         struct list_head *lgr_list;
 763         struct smc_link *lnk;
 764         spinlock_t *lgr_lock;
 765         u8 link_idx;
 766         int rc = 0;
 767         int i;
 768
 769         if (ini->is_smcd && ini->vlan_id) {
 770                 if (smc_ism_get_vlan(ini->ism_dev[ini->ism_selected],
 771                                      ini->vlan_id)) {
 772                         rc = SMC_CLC_DECL_ISMVLANERR;
 773                         goto out;
 774                 }
 775         }
 776
 777         lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
 778         if (!lgr) {
 779                 rc = SMC_CLC_DECL_MEM;
 780                 goto ism_put_vlan;
 781         }
 782         lgr->tx_wq = alloc_workqueue("smc_tx_wq-%*phN", 0, 0,
 783                                      SMC_LGR_ID_SIZE, &lgr->id);
 784         if (!lgr->tx_wq) {
 785                 rc = -ENOMEM;
 786                 goto free_lgr;
 787         }
 788         lgr->is_smcd = ini->is_smcd;
 789         lgr->sync_err = 0;
 790         lgr->terminating = 0;
 791         lgr->freeing = 0;
 792         lgr->vlan_id = ini->vlan_id;
 793         mutex_init(&lgr->sndbufs_lock);
 794         mutex_init(&lgr->rmbs_lock);
 795         rwlock_init(&lgr->conns_lock);
 796         for (i = 0; i < SMC_RMBE_SIZES; i++) {
 797                 INIT_LIST_HEAD(&lgr->sndbufs[i]);
 798                 INIT_LIST_HEAD(&lgr->rmbs[i]);
 799         }
 800         lgr->next_link_id = 0;
 801         smc_lgr_list.num += SMC_LGR_NUM_INCR;
 802         memcpy(&lgr->id, (u8 *)&smc_lgr_list.num, SMC_LGR_ID_SIZE);
 803         INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
 804         INIT_WORK(&lgr->terminate_work, smc_lgr_terminate_work);
 805         lgr->conns_all = RB_ROOT;
 806         if (ini->is_smcd) {
 807                 /* SMC-D specific settings */
 808                 get_device(&ini->ism_dev[ini->ism_selected]->dev);
 809                 lgr->peer_gid = ini->ism_peer_gid[ini->ism_selected];
 810                 lgr->smcd = ini->ism_dev[ini->ism_selected];
 811                 lgr_list = &ini->ism_dev[ini->ism_selected]->lgr_list;
 812                 lgr_lock = &lgr->smcd->lgr_lock;
 813                 lgr->smc_version = ini->smcd_version;
 814                 lgr->peer_shutdown = 0;
 815                 atomic_inc(&ini->ism_dev[ini->ism_selected]->lgr_cnt);
 816         } else {
 817                 /* SMC-R specific settings */
 818                 lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
 819                 memcpy(lgr->peer_systemid, ini->ib_lcl->id_for_peer,
 820                        SMC_SYSTEMID_LEN);
 821                 memcpy(lgr->pnet_id, ini->ib_dev->pnetid[ini->ib_port - 1],
 822                        SMC_MAX_PNETID_LEN);
 823                 smc_llc_lgr_init(lgr, smc);
 824
 825                 link_idx = SMC_SINGLE_LINK;
 826                 lnk = &lgr->lnk[link_idx];
 827                 rc = smcr_link_init(lgr, lnk, link_idx, ini);
 828                 if (rc)
 829                         goto free_wq;
 830                 lgr_list = &smc_lgr_list.list;
 831                 lgr_lock = &smc_lgr_list.lock;
 832                 atomic_inc(&lgr_cnt);
 833         }
 834         smc->conn.lgr = lgr;
 835         spin_lock_bh(lgr_lock);
 836         list_add_tail(&lgr->list, lgr_list);
 837         spin_unlock_bh(lgr_lock);
 838         return 0;
 839
 840 free_wq:
 841         destroy_workqueue(lgr->tx_wq);
 842 free_lgr:
 843         kfree(lgr);
 844 ism_put_vlan:
 845         if (ini->is_smcd && ini->vlan_id)
 846                 smc_ism_put_vlan(ini->ism_dev[ini->ism_selected], ini->vlan_id);
 847 out:
 848         if (rc < 0) {
 849                 if (rc == -ENOMEM)
 850                         rc = SMC_CLC_DECL_MEM;
 851                 else
 852                         rc = SMC_CLC_DECL_INTERR;
 853         }
 854         return rc;
 855 }
 856
 857 static int smc_write_space(struct smc_connection *conn)
 858 {
 859         int buffer_len = conn->peer_rmbe_size;
 860         union smc_host_cursor prod;
 861         union smc_host_cursor cons;
 862         int space;
 863
 864         smc_curs_copy(&prod, &conn->local_tx_ctrl.prod, conn);
 865         smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
 866         /* determine rx_buf space */
 867         space = buffer_len - smc_curs_diff(buffer_len, &cons, &prod);
 868         return space;
 869 }
 870
 871 static int smc_switch_cursor(struct smc_sock *smc, struct smc_cdc_tx_pend *pend,
 872                              struct smc_wr_buf *wr_buf)
 873 {
 874         struct smc_connection *conn = &smc->conn;
 875         union smc_host_cursor cons, fin;
 876         int rc = 0;
 877         int diff;
 878
 879         smc_curs_copy(&conn->tx_curs_sent, &conn->tx_curs_fin, conn);
 880         smc_curs_copy(&fin, &conn->local_tx_ctrl_fin, conn);
 881         /* set prod cursor to old state, enforce tx_rdma_writes() */
 882         smc_curs_copy(&conn->local_tx_ctrl.prod, &fin, conn);
 883         smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
 884
 885         if (smc_curs_comp(conn->peer_rmbe_size, &cons, &fin) < 0) {
 886                 /* cons cursor advanced more than fin, and prod was set
 887                  * fin above, so now prod is smaller than cons. Fix that.
 888                  */
 889                 diff = smc_curs_diff(conn->peer_rmbe_size, &fin, &cons);
 890                 smc_curs_add(conn->sndbuf_desc->len,
 891                              &conn->tx_curs_sent, diff);
 892                 smc_curs_add(conn->sndbuf_desc->len,
 893                              &conn->tx_curs_fin, diff);
 894
 895                 smp_mb__before_atomic();
 896                 atomic_add(diff, &conn->sndbuf_space);
 897                 smp_mb__after_atomic();
 898
 899                 smc_curs_add(conn->peer_rmbe_size,
 900                              &conn->local_tx_ctrl.prod, diff);
 901                 smc_curs_add(conn->peer_rmbe_size,
 902                              &conn->local_tx_ctrl_fin, diff);
 903         }
 904         /* recalculate, value is used by tx_rdma_writes() */
 905         atomic_set(&smc->conn.peer_rmbe_space, smc_write_space(conn));
 906
 907         if (smc->sk.sk_state != SMC_INIT &&
 908             smc->sk.sk_state != SMC_CLOSED) {
 909                 rc = smcr_cdc_msg_send_validation(conn, pend, wr_buf);
 910                 if (!rc) {
 911                         queue_delayed_work(conn->lgr->tx_wq, &conn->tx_work, 0);
 912                         smc->sk.sk_data_ready(&smc->sk);
 913                 }
 914         } else {
 915                 smc_wr_tx_put_slot(conn->lnk,
 916                                    (struct smc_wr_tx_pend_priv *)pend);
 917         }
 918         return rc;
 919 }
 920
 921 void smc_switch_link_and_count(struct smc_connection *conn,
 922                                struct smc_link *to_lnk)
 923 {
 924         atomic_dec(&conn->lnk->conn_cnt);
 925         conn->lnk = to_lnk;
 926         atomic_inc(&conn->lnk->conn_cnt);
 927 }
 928
 929 struct smc_link *smc_switch_conns(struct smc_link_group *lgr,
 930                                   struct smc_link *from_lnk, bool is_dev_err)
 931 {
 932         struct smc_link *to_lnk = NULL;
 933         struct smc_cdc_tx_pend *pend;
 934         struct smc_connection *conn;
 935         struct smc_wr_buf *wr_buf;
 936         struct smc_sock *smc;
 937         struct rb_node *node;
 938         int i, rc = 0;
 939
 940         /* link is inactive, wake up tx waiters */
 941         smc_wr_wakeup_tx_wait(from_lnk);
 942
 943         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 944                 if (!smc_link_active(&lgr->lnk[i]) || i == from_lnk->link_idx)
 945                         continue;
 946                 if (is_dev_err && from_lnk->smcibdev == lgr->lnk[i].smcibdev &&
 947                     from_lnk->ibport == lgr->lnk[i].ibport) {
 948                         continue;
 949                 }
 950                 to_lnk = &lgr->lnk[i];
 951                 break;
 952         }
 953         if (!to_lnk || !smc_wr_tx_link_hold(to_lnk)) {
 954                 smc_lgr_terminate_sched(lgr);
 955                 return NULL;
 956         }
 957 again:
 958         read_lock_bh(&lgr->conns_lock);
 959         for (node = rb_first(&lgr->conns_all); node; node = rb_next(node)) {
 960                 conn = rb_entry(node, struct smc_connection, alert_node);
 961                 if (conn->lnk != from_lnk)
 962                         continue;
 963                 smc = container_of(conn, struct smc_sock, conn);
 964                 /* conn->lnk not yet set in SMC_INIT state */
 965                 if (smc->sk.sk_state == SMC_INIT)
 966                         continue;
 967                 if (smc->sk.sk_state == SMC_CLOSED ||
 968                     smc->sk.sk_state == SMC_PEERCLOSEWAIT1 ||
 969                     smc->sk.sk_state == SMC_PEERCLOSEWAIT2 ||
 970                     smc->sk.sk_state == SMC_APPFINCLOSEWAIT ||
 971                     smc->sk.sk_state == SMC_APPCLOSEWAIT1 ||
 972                     smc->sk.sk_state == SMC_APPCLOSEWAIT2 ||
 973                     smc->sk.sk_state == SMC_PEERFINCLOSEWAIT ||
 974                     smc->sk.sk_state == SMC_PEERABORTWAIT ||
 975                     smc->sk.sk_state == SMC_PROCESSABORT) {
 976                         spin_lock_bh(&conn->send_lock);
 977                         smc_switch_link_and_count(conn, to_lnk);
 978                         spin_unlock_bh(&conn->send_lock);
 979                         continue;
 980                 }
 981                 sock_hold(&smc->sk);
 982                 read_unlock_bh(&lgr->conns_lock);
 983                 /* pre-fetch buffer outside of send_lock, might sleep */
 984                 rc = smc_cdc_get_free_slot(conn, to_lnk, &wr_buf, NULL, &pend);
 985                 if (rc)
 986                         goto err_out;
 987                 /* avoid race with smcr_tx_sndbuf_nonempty() */
 988                 spin_lock_bh(&conn->send_lock);
 989                 smc_switch_link_and_count(conn, to_lnk);
 990                 rc = smc_switch_cursor(smc, pend, wr_buf);
 991                 spin_unlock_bh(&conn->send_lock);
 992                 sock_put(&smc->sk);
 993                 if (rc)
 994                         goto err_out;
 995                 goto again;
 996         }
 997         read_unlock_bh(&lgr->conns_lock);
 998         smc_wr_tx_link_put(to_lnk);
 999         return to_lnk;
1000
1001 err_out:
1002         smcr_link_down_cond_sched(to_lnk);
1003         smc_wr_tx_link_put(to_lnk);
1004         return NULL;
1005 }
1006
1007 static void smcr_buf_unuse(struct smc_buf_desc *rmb_desc,
1008                            struct smc_link_group *lgr)
1009 {
1010         int rc;
1011
1012         if (rmb_desc->is_conf_rkey && !list_empty(&lgr->list)) {
1013                 /* unregister rmb with peer */
1014                 rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY);
1015                 if (!rc) {
1016                         /* protect against smc_llc_cli_rkey_exchange() */
1017                         mutex_lock(&lgr->llc_conf_mutex);
1018                         smc_llc_do_delete_rkey(lgr, rmb_desc);
1019                         rmb_desc->is_conf_rkey = false;
1020                         mutex_unlock(&lgr->llc_conf_mutex);
1021                         smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl);
1022                 }
1023         }
1024
1025         if (rmb_desc->is_reg_err) {
1026                 /* buf registration failed, reuse not possible */
1027                 mutex_lock(&lgr->rmbs_lock);
1028                 list_del(&rmb_desc->list);
1029                 mutex_unlock(&lgr->rmbs_lock);
1030
1031                 smc_buf_free(lgr, true, rmb_desc);
1032         } else {
1033                 rmb_desc->used = 0;
1034         }
1035 }
1036
1037 static void smc_buf_unuse(struct smc_connection *conn,
1038                           struct smc_link_group *lgr)
1039 {
1040         if (conn->sndbuf_desc)
1041                 conn->sndbuf_desc->used = 0;
1042         if (conn->rmb_desc && lgr->is_smcd)
1043                 conn->rmb_desc->used = 0;
1044         else if (conn->rmb_desc)
1045                 smcr_buf_unuse(conn->rmb_desc, lgr);
1046 }
1047
1048 /* remove a finished connection from its link group */
1049 void smc_conn_free(struct smc_connection *conn)
1050 {
1051         struct smc_link_group *lgr = conn->lgr;
1052
1053         if (!lgr)
1054                 return;
1055         if (lgr->is_smcd) {
1056                 if (!list_empty(&lgr->list))
1057                         smc_ism_unset_conn(conn);
1058                 tasklet_kill(&conn->rx_tsklet);
1059         } else {
1060                 smc_cdc_tx_dismiss_slots(conn);
1061                 if (current_work() != &conn->abort_work)
1062                         cancel_work_sync(&conn->abort_work);
1063         }
1064         if (!list_empty(&lgr->list)) {
1065                 smc_lgr_unregister_conn(conn);
1066                 smc_buf_unuse(conn, lgr); /* allow buffer reuse */
1067         }
1068
1069         if (!lgr->conns_num)
1070                 smc_lgr_schedule_free_work(lgr);
1071 }
1072
1073 /* unregister a link from a buf_desc */
1074 static void smcr_buf_unmap_link(struct smc_buf_desc *buf_desc, bool is_rmb,
1075                                 struct smc_link *lnk)
1076 {
1077         if (is_rmb)
1078                 buf_desc->is_reg_mr[lnk->link_idx] = false;
1079         if (!buf_desc->is_map_ib[lnk->link_idx])
1080                 return;
1081         if (is_rmb) {
1082                 if (buf_desc->mr_rx[lnk->link_idx]) {
1083                         smc_ib_put_memory_region(
1084                                         buf_desc->mr_rx[lnk->link_idx]);
1085                         buf_desc->mr_rx[lnk->link_idx] = NULL;
1086                 }
1087                 smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_FROM_DEVICE);
1088         } else {
1089                 smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_TO_DEVICE);
1090         }
1091         sg_free_table(&buf_desc->sgt[lnk->link_idx]);
1092         buf_desc->is_map_ib[lnk->link_idx] = false;
1093 }
1094
1095 /* unmap all buffers of lgr for a deleted link */
1096 static void smcr_buf_unmap_lgr(struct smc_link *lnk)
1097 {
1098         struct smc_link_group *lgr = lnk->lgr;
1099         struct smc_buf_desc *buf_desc, *bf;
1100         int i;
1101
1102         for (i = 0; i < SMC_RMBE_SIZES; i++) {
1103                 mutex_lock(&lgr->rmbs_lock);
1104                 list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list)
1105                         smcr_buf_unmap_link(buf_desc, true, lnk);
1106                 mutex_unlock(&lgr->rmbs_lock);
1107                 mutex_lock(&lgr->sndbufs_lock);
1108                 list_for_each_entry_safe(buf_desc, bf, &lgr->sndbufs[i],
1109                                          list)
1110                         smcr_buf_unmap_link(buf_desc, false, lnk);
1111                 mutex_unlock(&lgr->sndbufs_lock);
1112         }
1113 }
1114
1115 static void smcr_rtoken_clear_link(struct smc_link *lnk)
1116 {
1117         struct smc_link_group *lgr = lnk->lgr;
1118         int i;
1119
1120         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1121                 lgr->rtokens[i][lnk->link_idx].rkey = 0;
1122                 lgr->rtokens[i][lnk->link_idx].dma_addr = 0;
1123         }
1124 }
1125
1126 /* must be called under lgr->llc_conf_mutex lock */
1127 void smcr_link_clear(struct smc_link *lnk, bool log)
1128 {
1129         struct smc_ib_device *smcibdev;
1130
1131         if (!lnk->lgr || lnk->state == SMC_LNK_UNUSED)
1132                 return;
1133         lnk->peer_qpn = 0;
1134         smc_llc_link_clear(lnk, log);
1135         smcr_buf_unmap_lgr(lnk);
1136         smcr_rtoken_clear_link(lnk);
1137         smc_ib_modify_qp_reset(lnk);
1138         smc_wr_free_link(lnk);
1139         smc_ib_destroy_queue_pair(lnk);
1140         smc_ib_dealloc_protection_domain(lnk);
1141         smc_wr_free_link_mem(lnk);
1142         smc_ibdev_cnt_dec(lnk);
1143         put_device(&lnk->smcibdev->ibdev->dev);
1144         smcibdev = lnk->smcibdev;
1145         memset(lnk, 0, sizeof(struct smc_link));
1146         lnk->state = SMC_LNK_UNUSED;
1147         if (!atomic_dec_return(&smcibdev->lnk_cnt))
1148                 wake_up(&smcibdev->lnks_deleted);
1149 }
1150
1151 static void smcr_buf_free(struct smc_link_group *lgr, bool is_rmb,
1152                           struct smc_buf_desc *buf_desc)
1153 {
1154         int i;
1155
1156         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
1157                 smcr_buf_unmap_link(buf_desc, is_rmb, &lgr->lnk[i]);
1158
1159         if (buf_desc->pages)
1160                 __free_pages(buf_desc->pages, buf_desc->order);
1161         kfree(buf_desc);
1162 }
1163
1164 static void smcd_buf_free(struct smc_link_group *lgr, bool is_dmb,
1165                           struct smc_buf_desc *buf_desc)
1166 {
1167         if (is_dmb) {
1168                 /* restore original buf len */
1169                 buf_desc->len += sizeof(struct smcd_cdc_msg);
1170                 smc_ism_unregister_dmb(lgr->smcd, buf_desc);
1171         } else {
1172                 kfree(buf_desc->cpu_addr);
1173         }
1174         kfree(buf_desc);
1175 }
1176
1177 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
1178                          struct smc_buf_desc *buf_desc)
1179 {
1180         if (lgr->is_smcd)
1181                 smcd_buf_free(lgr, is_rmb, buf_desc);
1182         else
1183                 smcr_buf_free(lgr, is_rmb, buf_desc);
1184 }
1185
1186 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
1187 {
1188         struct smc_buf_desc *buf_desc, *bf_desc;
1189         struct list_head *buf_list;
1190         int i;
1191
1192         for (i = 0; i < SMC_RMBE_SIZES; i++) {
1193                 if (is_rmb)
1194                         buf_list = &lgr->rmbs[i];
1195                 else
1196                         buf_list = &lgr->sndbufs[i];
1197                 list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
1198                                          list) {
1199                         list_del(&buf_desc->list);
1200                         smc_buf_free(lgr, is_rmb, buf_desc);
1201                 }
1202         }
1203 }
1204
1205 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
1206 {
1207         /* free send buffers */
1208         __smc_lgr_free_bufs(lgr, false);
1209         /* free rmbs */
1210         __smc_lgr_free_bufs(lgr, true);
1211 }
1212
1213 /* remove a link group */
1214 static void smc_lgr_free(struct smc_link_group *lgr)
1215 {
1216         int i;
1217
1218         if (!lgr->is_smcd) {
1219                 mutex_lock(&lgr->llc_conf_mutex);
1220                 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1221                         if (lgr->lnk[i].state != SMC_LNK_UNUSED)
1222                                 smcr_link_clear(&lgr->lnk[i], false);
1223                 }
1224                 mutex_unlock(&lgr->llc_conf_mutex);
1225                 smc_llc_lgr_clear(lgr);
1226         }
1227
1228         smc_lgr_free_bufs(lgr);
1229         destroy_workqueue(lgr->tx_wq);
1230         if (lgr->is_smcd) {
1231                 smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
1232                 put_device(&lgr->smcd->dev);
1233                 if (!atomic_dec_return(&lgr->smcd->lgr_cnt))
1234                         wake_up(&lgr->smcd->lgrs_deleted);
1235         } else {
1236                 if (!atomic_dec_return(&lgr_cnt))
1237                         wake_up(&lgrs_deleted);
1238         }
1239         kfree(lgr);
1240 }
1241
1242 static void smc_sk_wake_ups(struct smc_sock *smc)
1243 {
1244         smc->sk.sk_write_space(&smc->sk);
1245         smc->sk.sk_data_ready(&smc->sk);
1246         smc->sk.sk_state_change(&smc->sk);
1247 }
1248
1249 /* kill a connection */
1250 static void smc_conn_kill(struct smc_connection *conn, bool soft)
1251 {
1252         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1253
1254         if (conn->lgr->is_smcd && conn->lgr->peer_shutdown)
1255                 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
1256         else
1257                 smc_close_abort(conn);
1258         conn->killed = 1;
1259         smc->sk.sk_err = ECONNABORTED;
1260         smc_sk_wake_ups(smc);
1261         if (conn->lgr->is_smcd) {
1262                 smc_ism_unset_conn(conn);
1263                 if (soft)
1264                         tasklet_kill(&conn->rx_tsklet);
1265                 else
1266                         tasklet_unlock_wait(&conn->rx_tsklet);
1267         } else {
1268                 smc_cdc_tx_dismiss_slots(conn);
1269         }
1270         smc_lgr_unregister_conn(conn);
1271         smc_close_active_abort(smc);
1272 }
1273
1274 static void smc_lgr_cleanup(struct smc_link_group *lgr)
1275 {
1276         if (lgr->is_smcd) {
1277                 smc_ism_signal_shutdown(lgr);
1278         } else {
1279                 u32 rsn = lgr->llc_termination_rsn;
1280
1281                 if (!rsn)
1282                         rsn = SMC_LLC_DEL_PROG_INIT_TERM;
1283                 smc_llc_send_link_delete_all(lgr, false, rsn);
1284                 smcr_lgr_link_deactivate_all(lgr);
1285         }
1286 }
1287
1288 /* terminate link group
1289  * @soft: true if link group shutdown can take its time
1290  *        false if immediate link group shutdown is required
1291  */
1292 static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft)
1293 {
1294         struct smc_connection *conn;
1295         struct smc_sock *smc;
1296         struct rb_node *node;
1297
1298         if (lgr->terminating)
1299                 return; /* lgr already terminating */
1300         /* cancel free_work sync, will terminate when lgr->freeing is set */
1301         cancel_delayed_work_sync(&lgr->free_work);
1302         lgr->terminating = 1;
1303
1304         /* kill remaining link group connections */
1305         read_lock_bh(&lgr->conns_lock);
1306         node = rb_first(&lgr->conns_all);
1307         while (node) {
1308                 read_unlock_bh(&lgr->conns_lock);
1309                 conn = rb_entry(node, struct smc_connection, alert_node);
1310                 smc = container_of(conn, struct smc_sock, conn);
1311                 sock_hold(&smc->sk); /* sock_put below */
1312                 lock_sock(&smc->sk);
1313                 smc_conn_kill(conn, soft);
1314                 release_sock(&smc->sk);
1315                 sock_put(&smc->sk); /* sock_hold above */
1316                 read_lock_bh(&lgr->conns_lock);
1317                 node = rb_first(&lgr->conns_all);
1318         }
1319         read_unlock_bh(&lgr->conns_lock);
1320         smc_lgr_cleanup(lgr);
1321         smc_lgr_free(lgr);
1322 }
1323
1324 /* unlink link group and schedule termination */
1325 void smc_lgr_terminate_sched(struct smc_link_group *lgr)
1326 {
1327         spinlock_t *lgr_lock;
1328
1329         smc_lgr_list_head(lgr, &lgr_lock);
1330         spin_lock_bh(lgr_lock);
1331         if (list_empty(&lgr->list) || lgr->terminating || lgr->freeing) {
1332                 spin_unlock_bh(lgr_lock);
1333                 return; /* lgr already terminating */
1334         }
1335         list_del_init(&lgr->list);
1336         lgr->freeing = 1;
1337         spin_unlock_bh(lgr_lock);
1338         schedule_work(&lgr->terminate_work);
1339 }
1340
1341 /* Called when peer lgr shutdown (regularly or abnormally) is received */
1342 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid, unsigned short vlan)
1343 {
1344         struct smc_link_group *lgr, *l;
1345         LIST_HEAD(lgr_free_list);
1346
1347         /* run common cleanup function and build free list */
1348         spin_lock_bh(&dev->lgr_lock);
1349         list_for_each_entry_safe(lgr, l, &dev->lgr_list, list) {
1350                 if ((!peer_gid || lgr->peer_gid == peer_gid) &&
1351                     (vlan == VLAN_VID_MASK || lgr->vlan_id == vlan)) {
1352                         if (peer_gid) /* peer triggered termination */
1353                                 lgr->peer_shutdown = 1;
1354                         list_move(&lgr->list, &lgr_free_list);
1355                         lgr->freeing = 1;
1356                 }
1357         }
1358         spin_unlock_bh(&dev->lgr_lock);
1359
1360         /* cancel the regular free workers and actually free lgrs */
1361         list_for_each_entry_safe(lgr, l, &lgr_free_list, list) {
1362                 list_del_init(&lgr->list);
1363                 schedule_work(&lgr->terminate_work);
1364         }
1365 }
1366
1367 /* Called when an SMCD device is removed or the smc module is unloaded */
1368 void smc_smcd_terminate_all(struct smcd_dev *smcd)
1369 {
1370         struct smc_link_group *lgr, *lg;
1371         LIST_HEAD(lgr_free_list);
1372
1373         spin_lock_bh(&smcd->lgr_lock);
1374         list_splice_init(&smcd->lgr_list, &lgr_free_list);
1375         list_for_each_entry(lgr, &lgr_free_list, list)
1376                 lgr->freeing = 1;
1377         spin_unlock_bh(&smcd->lgr_lock);
1378
1379         list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
1380                 list_del_init(&lgr->list);
1381                 __smc_lgr_terminate(lgr, false);
1382         }
1383
1384         if (atomic_read(&smcd->lgr_cnt))
1385                 wait_event(smcd->lgrs_deleted, !atomic_read(&smcd->lgr_cnt));
1386 }
1387
1388 /* Called when an SMCR device is removed or the smc module is unloaded.
1389  * If smcibdev is given, all SMCR link groups using this device are terminated.
1390  * If smcibdev is NULL, all SMCR link groups are terminated.
1391  */
1392 void smc_smcr_terminate_all(struct smc_ib_device *smcibdev)
1393 {
1394         struct smc_link_group *lgr, *lg;
1395         LIST_HEAD(lgr_free_list);
1396         int i;
1397
1398         spin_lock_bh(&smc_lgr_list.lock);
1399         if (!smcibdev) {
1400                 list_splice_init(&smc_lgr_list.list, &lgr_free_list);
1401                 list_for_each_entry(lgr, &lgr_free_list, list)
1402                         lgr->freeing = 1;
1403         } else {
1404                 list_for_each_entry_safe(lgr, lg, &smc_lgr_list.list, list) {
1405                         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1406                                 if (lgr->lnk[i].smcibdev == smcibdev)
1407                                         smcr_link_down_cond_sched(&lgr->lnk[i]);
1408                         }
1409                 }
1410         }
1411         spin_unlock_bh(&smc_lgr_list.lock);
1412
1413         list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
1414                 list_del_init(&lgr->list);
1415                 smc_llc_set_termination_rsn(lgr, SMC_LLC_DEL_OP_INIT_TERM);
1416                 __smc_lgr_terminate(lgr, false);
1417         }
1418
1419         if (smcibdev) {
1420                 if (atomic_read(&smcibdev->lnk_cnt))
1421                         wait_event(smcibdev->lnks_deleted,
1422                                    !atomic_read(&smcibdev->lnk_cnt));
1423         } else {
1424                 if (atomic_read(&lgr_cnt))
1425                         wait_event(lgrs_deleted, !atomic_read(&lgr_cnt));
1426         }
1427 }
1428
1429 /* set new lgr type and clear all asymmetric link tagging */
1430 void smcr_lgr_set_type(struct smc_link_group *lgr, enum smc_lgr_type new_type)
1431 {
1432         char *lgr_type = "";
1433         int i;
1434
1435         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
1436                 if (smc_link_usable(&lgr->lnk[i]))
1437                         lgr->lnk[i].link_is_asym = false;
1438         if (lgr->type == new_type)
1439                 return;
1440         lgr->type = new_type;
1441
1442         switch (lgr->type) {
1443         case SMC_LGR_NONE:
1444                 lgr_type = "NONE";
1445                 break;
1446         case SMC_LGR_SINGLE:
1447                 lgr_type = "SINGLE";
1448                 break;
1449         case SMC_LGR_SYMMETRIC:
1450                 lgr_type = "SYMMETRIC";
1451                 break;
1452         case SMC_LGR_ASYMMETRIC_PEER:
1453                 lgr_type = "ASYMMETRIC_PEER";
1454                 break;
1455         case SMC_LGR_ASYMMETRIC_LOCAL:
1456                 lgr_type = "ASYMMETRIC_LOCAL";
1457                 break;
1458         }
1459         pr_warn_ratelimited("smc: SMC-R lg %*phN state changed: "
1460                             "%s, pnetid %.16s\n", SMC_LGR_ID_SIZE, &lgr->id,
1461                             lgr_type, lgr->pnet_id);
1462 }
1463
1464 /* set new lgr type and tag a link as asymmetric */
1465 void smcr_lgr_set_type_asym(struct smc_link_group *lgr,
1466                             enum smc_lgr_type new_type, int asym_lnk_idx)
1467 {
1468         smcr_lgr_set_type(lgr, new_type);
1469         lgr->lnk[asym_lnk_idx].link_is_asym = true;
1470 }
1471
1472 /* abort connection, abort_work scheduled from tasklet context */
1473 static void smc_conn_abort_work(struct work_struct *work)
1474 {
1475         struct smc_connection *conn = container_of(work,
1476                                                    struct smc_connection,
1477                                                    abort_work);
1478         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1479
1480         lock_sock(&smc->sk);
1481         smc_conn_kill(conn, true);
1482         release_sock(&smc->sk);
1483         sock_put(&smc->sk); /* sock_hold done by schedulers of abort_work */
1484 }
1485
1486 void smcr_port_add(struct smc_ib_device *smcibdev, u8 ibport)
1487 {
1488         struct smc_link_group *lgr, *n;
1489
1490         list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1491                 struct smc_link *link;
1492
1493                 if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1494                             SMC_MAX_PNETID_LEN) ||
1495                     lgr->type == SMC_LGR_SYMMETRIC ||
1496                     lgr->type == SMC_LGR_ASYMMETRIC_PEER)
1497                         continue;
1498
1499                 /* trigger local add link processing */
1500                 link = smc_llc_usable_link(lgr);
1501                 if (link)
1502                         smc_llc_add_link_local(link);
1503         }
1504 }
1505
1506 /* link is down - switch connections to alternate link,
1507  * must be called under lgr->llc_conf_mutex lock
1508  */
1509 static void smcr_link_down(struct smc_link *lnk)
1510 {
1511         struct smc_link_group *lgr = lnk->lgr;
1512         struct smc_link *to_lnk;
1513         int del_link_id;
1514
1515         if (!lgr || lnk->state == SMC_LNK_UNUSED || list_empty(&lgr->list))
1516                 return;
1517
1518         smc_ib_modify_qp_reset(lnk);
1519         to_lnk = smc_switch_conns(lgr, lnk, true);
1520         if (!to_lnk) { /* no backup link available */
1521                 smcr_link_clear(lnk, true);
1522                 return;
1523         }
1524         smcr_lgr_set_type(lgr, SMC_LGR_SINGLE);
1525         del_link_id = lnk->link_id;
1526
1527         if (lgr->role == SMC_SERV) {
1528                 /* trigger local delete link processing */
1529                 smc_llc_srv_delete_link_local(to_lnk, del_link_id);
1530         } else {
1531                 if (lgr->llc_flow_lcl.type != SMC_LLC_FLOW_NONE) {
1532                         /* another llc task is ongoing */
1533                         mutex_unlock(&lgr->llc_conf_mutex);
1534                         wait_event_timeout(lgr->llc_flow_waiter,
1535                                 (list_empty(&lgr->list) ||
1536                                  lgr->llc_flow_lcl.type == SMC_LLC_FLOW_NONE),
1537                                 SMC_LLC_WAIT_TIME);
1538                         mutex_lock(&lgr->llc_conf_mutex);
1539                 }
1540                 if (!list_empty(&lgr->list)) {
1541                         smc_llc_send_delete_link(to_lnk, del_link_id,
1542                                                  SMC_LLC_REQ, true,
1543                                                  SMC_LLC_DEL_LOST_PATH);
1544                         smcr_link_clear(lnk, true);
1545                 }
1546                 wake_up(&lgr->llc_flow_waiter); /* wake up next waiter */
1547         }
1548 }
1549
1550 /* must be called under lgr->llc_conf_mutex lock */
1551 void smcr_link_down_cond(struct smc_link *lnk)
1552 {
1553         if (smc_link_downing(&lnk->state))
1554                 smcr_link_down(lnk);
1555 }
1556
1557 /* will get the lgr->llc_conf_mutex lock */
1558 void smcr_link_down_cond_sched(struct smc_link *lnk)
1559 {
1560         if (smc_link_downing(&lnk->state))
1561                 schedule_work(&lnk->link_down_wrk);
1562 }
1563
1564 void smcr_port_err(struct smc_ib_device *smcibdev, u8 ibport)
1565 {
1566         struct smc_link_group *lgr, *n;
1567         int i;
1568
1569         list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1570                 if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1571                             SMC_MAX_PNETID_LEN))
1572                         continue; /* lgr is not affected */
1573                 if (list_empty(&lgr->list))
1574                         continue;
1575                 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1576                         struct smc_link *lnk = &lgr->lnk[i];
1577
1578                         if (smc_link_usable(lnk) &&
1579                             lnk->smcibdev == smcibdev && lnk->ibport == ibport)
1580                                 smcr_link_down_cond_sched(lnk);
1581                 }
1582         }
1583 }
1584
1585 static void smc_link_down_work(struct work_struct *work)
1586 {
1587         struct smc_link *link = container_of(work, struct smc_link,
1588                                              link_down_wrk);
1589         struct smc_link_group *lgr = link->lgr;
1590
1591         if (list_empty(&lgr->list))
1592                 return;
1593         wake_up_all(&lgr->llc_msg_waiter);
1594         mutex_lock(&lgr->llc_conf_mutex);
1595         smcr_link_down(link);
1596         mutex_unlock(&lgr->llc_conf_mutex);
1597 }
1598
1599 /* Determine vlan of internal TCP socket.
1600  * @vlan_id: address to store the determined vlan id into
1601  */
1602 int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini)
1603 {
1604         struct dst_entry *dst = sk_dst_get(clcsock->sk);
1605         struct net_device *ndev;
1606         int i, nest_lvl, rc = 0;
1607
1608         ini->vlan_id = 0;
1609         if (!dst) {
1610                 rc = -ENOTCONN;
1611                 goto out;
1612         }
1613         if (!dst->dev) {
1614                 rc = -ENODEV;
1615                 goto out_rel;
1616         }
1617
1618         ndev = dst->dev;
1619         if (is_vlan_dev(ndev)) {
1620                 ini->vlan_id = vlan_dev_vlan_id(ndev);
1621                 goto out_rel;
1622         }
1623
1624         rtnl_lock();
1625         nest_lvl = ndev->lower_level;
1626         for (i = 0; i < nest_lvl; i++) {
1627                 struct list_head *lower = &ndev->adj_list.lower;
1628
1629                 if (list_empty(lower))
1630                         break;
1631                 lower = lower->next;
1632                 ndev = (struct net_device *)netdev_lower_get_next(ndev, &lower);
1633                 if (is_vlan_dev(ndev)) {
1634                         ini->vlan_id = vlan_dev_vlan_id(ndev);
1635                         break;
1636                 }
1637         }
1638         rtnl_unlock();
1639
1640 out_rel:
1641         dst_release(dst);
1642 out:
1643         return rc;
1644 }
1645
1646 static bool smcr_lgr_match(struct smc_link_group *lgr,
1647                            struct smc_clc_msg_local *lcl,
1648                            enum smc_lgr_role role, u32 clcqpn)
1649 {
1650         int i;
1651
1652         if (memcmp(lgr->peer_systemid, lcl->id_for_peer, SMC_SYSTEMID_LEN) ||
1653             lgr->role != role)
1654                 return false;
1655
1656         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1657                 if (!smc_link_active(&lgr->lnk[i]))
1658                         continue;
1659                 if ((lgr->role == SMC_SERV || lgr->lnk[i].peer_qpn == clcqpn) &&
1660                     !memcmp(lgr->lnk[i].peer_gid, &lcl->gid, SMC_GID_SIZE) &&
1661                     !memcmp(lgr->lnk[i].peer_mac, lcl->mac, sizeof(lcl->mac)))
1662                         return true;
1663         }
1664         return false;
1665 }
1666
1667 static bool smcd_lgr_match(struct smc_link_group *lgr,
1668                            struct smcd_dev *smcismdev, u64 peer_gid)
1669 {
1670         return lgr->peer_gid == peer_gid && lgr->smcd == smcismdev;
1671 }
1672
1673 /* create a new SMC connection (and a new link group if necessary) */
1674 int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini)
1675 {
1676         struct smc_connection *conn = &smc->conn;
1677         struct list_head *lgr_list;
1678         struct smc_link_group *lgr;
1679         enum smc_lgr_role role;
1680         spinlock_t *lgr_lock;
1681         int rc = 0;
1682
1683         lgr_list = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_list :
1684                                   &smc_lgr_list.list;
1685         lgr_lock = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_lock :
1686                                   &smc_lgr_list.lock;
1687         ini->first_contact_local = 1;
1688         role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
1689         if (role == SMC_CLNT && ini->first_contact_peer)
1690                 /* create new link group as well */
1691                 goto create;
1692
1693         /* determine if an existing link group can be reused */
1694         spin_lock_bh(lgr_lock);
1695         list_for_each_entry(lgr, lgr_list, list) {
1696                 write_lock_bh(&lgr->conns_lock);
1697                 if ((ini->is_smcd ?
1698                      smcd_lgr_match(lgr, ini->ism_dev[ini->ism_selected],
1699                                     ini->ism_peer_gid[ini->ism_selected]) :
1700                      smcr_lgr_match(lgr, ini->ib_lcl, role, ini->ib_clcqpn)) &&
1701                     !lgr->sync_err &&
1702                     (ini->smcd_version == SMC_V2 ||
1703                      lgr->vlan_id == ini->vlan_id) &&
1704                     (role == SMC_CLNT || ini->is_smcd ||
1705                      lgr->conns_num < SMC_RMBS_PER_LGR_MAX)) {
1706                         /* link group found */
1707                         ini->first_contact_local = 0;
1708                         conn->lgr = lgr;
1709                         rc = smc_lgr_register_conn(conn, false);
1710                         write_unlock_bh(&lgr->conns_lock);
1711                         if (!rc && delayed_work_pending(&lgr->free_work))
1712                                 cancel_delayed_work(&lgr->free_work);
1713                         break;
1714                 }
1715                 write_unlock_bh(&lgr->conns_lock);
1716         }
1717         spin_unlock_bh(lgr_lock);
1718         if (rc)
1719                 return rc;
1720
1721         if (role == SMC_CLNT && !ini->first_contact_peer &&
1722             ini->first_contact_local) {
1723                 /* Server reuses a link group, but Client wants to start
1724                  * a new one
1725                  * send out_of_sync decline, reason synchr. error
1726                  */
1727                 return SMC_CLC_DECL_SYNCERR;
1728         }
1729
1730 create:
1731         if (ini->first_contact_local) {
1732                 rc = smc_lgr_create(smc, ini);
1733                 if (rc)
1734                         goto out;
1735                 lgr = conn->lgr;
1736                 write_lock_bh(&lgr->conns_lock);
1737                 rc = smc_lgr_register_conn(conn, true);
1738                 write_unlock_bh(&lgr->conns_lock);
1739                 if (rc)
1740                         goto out;
1741         }
1742         conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
1743         conn->local_tx_ctrl.len = SMC_WR_TX_SIZE;
1744         conn->urg_state = SMC_URG_READ;
1745         INIT_WORK(&smc->conn.abort_work, smc_conn_abort_work);
1746         if (ini->is_smcd) {
1747                 conn->rx_off = sizeof(struct smcd_cdc_msg);
1748                 smcd_cdc_rx_init(conn); /* init tasklet for this conn */
1749         } else {
1750                 conn->rx_off = 0;
1751         }
1752 #ifndef KERNEL_HAS_ATOMIC64
1753         spin_lock_init(&conn->acurs_lock);
1754 #endif
1755
1756 out:
1757         return rc;
1758 }
1759
1760 #define SMCD_DMBE_SIZES         6 /* 0 -> 16KB, 1 -> 32KB, .. 6 -> 1MB */
1761 #define SMCR_RMBE_SIZES         5 /* 0 -> 16KB, 1 -> 32KB, .. 5 -> 512KB */
1762
1763 /* convert the RMB size into the compressed notation (minimum 16K, see
1764  * SMCD/R_DMBE_SIZES.
1765  * In contrast to plain ilog2, this rounds towards the next power of 2,
1766  * so the socket application gets at least its desired sndbuf / rcvbuf size.
1767  */
1768 static u8 smc_compress_bufsize(int size, bool is_smcd, bool is_rmb)
1769 {
1770         const unsigned int max_scat = SG_MAX_SINGLE_ALLOC * PAGE_SIZE;
1771         u8 compressed;
1772
1773         if (size <= SMC_BUF_MIN_SIZE)
1774                 return 0;
1775
1776         size = (size - 1) >> 14;  /* convert to 16K multiple */
1777         compressed = min_t(u8, ilog2(size) + 1,
1778                            is_smcd ? SMCD_DMBE_SIZES : SMCR_RMBE_SIZES);
1779
1780         if (!is_smcd && is_rmb)
1781                 /* RMBs are backed by & limited to max size of scatterlists */
1782                 compressed = min_t(u8, compressed, ilog2(max_scat >> 14));
1783
1784         return compressed;
1785 }
1786
1787 /* convert the RMB size from compressed notation into integer */
1788 int smc_uncompress_bufsize(u8 compressed)
1789 {
1790         u32 size;
1791
1792         size = 0x00000001 << (((int)compressed) + 14);
1793         return (int)size;
1794 }
1795
1796 /* try to reuse a sndbuf or rmb description slot for a certain
1797  * buffer size; if not available, return NULL
1798  */
1799 static struct smc_buf_desc *smc_buf_get_slot(int compressed_bufsize,
1800                                              struct mutex *lock,
1801                                              struct list_head *buf_list)
1802 {
1803         struct smc_buf_desc *buf_slot;
1804
1805         mutex_lock(lock);
1806         list_for_each_entry(buf_slot, buf_list, list) {
1807                 if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
1808                         mutex_unlock(lock);
1809                         return buf_slot;
1810                 }
1811         }
1812         mutex_unlock(lock);
1813         return NULL;
1814 }
1815
1816 /* one of the conditions for announcing a receiver's current window size is
1817  * that it "results in a minimum increase in the window size of 10% of the
1818  * receive buffer space" [RFC7609]
1819  */
1820 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
1821 {
1822         return min_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
1823 }
1824
1825 /* map an rmb buf to a link */
1826 static int smcr_buf_map_link(struct smc_buf_desc *buf_desc, bool is_rmb,
1827                              struct smc_link *lnk)
1828 {
1829         int rc;
1830
1831         if (buf_desc->is_map_ib[lnk->link_idx])
1832                 return 0;
1833
1834         rc = sg_alloc_table(&buf_desc->sgt[lnk->link_idx], 1, GFP_KERNEL);
1835         if (rc)
1836                 return rc;
1837         sg_set_buf(buf_desc->sgt[lnk->link_idx].sgl,
1838                    buf_desc->cpu_addr, buf_desc->len);
1839
1840         /* map sg table to DMA address */
1841         rc = smc_ib_buf_map_sg(lnk, buf_desc,
1842                                is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
1843         /* SMC protocol depends on mapping to one DMA address only */
1844         if (rc != 1) {
1845                 rc = -EAGAIN;
1846                 goto free_table;
1847         }
1848
1849         /* create a new memory region for the RMB */
1850         if (is_rmb) {
1851                 rc = smc_ib_get_memory_region(lnk->roce_pd,
1852                                               IB_ACCESS_REMOTE_WRITE |
1853                                               IB_ACCESS_LOCAL_WRITE,
1854                                               buf_desc, lnk->link_idx);
1855                 if (rc)
1856                         goto buf_unmap;
1857                 smc_ib_sync_sg_for_device(lnk, buf_desc, DMA_FROM_DEVICE);
1858         }
1859         buf_desc->is_map_ib[lnk->link_idx] = true;
1860         return 0;
1861
1862 buf_unmap:
1863         smc_ib_buf_unmap_sg(lnk, buf_desc,
1864                             is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
1865 free_table:
1866         sg_free_table(&buf_desc->sgt[lnk->link_idx]);
1867         return rc;
1868 }
1869
1870 /* register a new rmb on IB device,
1871  * must be called under lgr->llc_conf_mutex lock
1872  */
1873 int smcr_link_reg_rmb(struct smc_link *link, struct smc_buf_desc *rmb_desc)
1874 {
1875         if (list_empty(&link->lgr->list))
1876                 return -ENOLINK;
1877         if (!rmb_desc->is_reg_mr[link->link_idx]) {
1878                 /* register memory region for new rmb */
1879                 if (smc_wr_reg_send(link, rmb_desc->mr_rx[link->link_idx])) {
1880                         rmb_desc->is_reg_err = true;
1881                         return -EFAULT;
1882                 }
1883                 rmb_desc->is_reg_mr[link->link_idx] = true;
1884         }
1885         return 0;
1886 }
1887
1888 static int _smcr_buf_map_lgr(struct smc_link *lnk, struct mutex *lock,
1889                              struct list_head *lst, bool is_rmb)
1890 {
1891         struct smc_buf_desc *buf_desc, *bf;
1892         int rc = 0;
1893
1894         mutex_lock(lock);
1895         list_for_each_entry_safe(buf_desc, bf, lst, list) {
1896                 if (!buf_desc->used)
1897                         continue;
1898                 rc = smcr_buf_map_link(buf_desc, is_rmb, lnk);
1899                 if (rc)
1900                         goto out;
1901         }
1902 out:
1903         mutex_unlock(lock);
1904         return rc;
1905 }
1906
1907 /* map all used buffers of lgr for a new link */
1908 int smcr_buf_map_lgr(struct smc_link *lnk)
1909 {
1910         struct smc_link_group *lgr = lnk->lgr;
1911         int i, rc = 0;
1912
1913         for (i = 0; i < SMC_RMBE_SIZES; i++) {
1914                 rc = _smcr_buf_map_lgr(lnk, &lgr->rmbs_lock,
1915                                        &lgr->rmbs[i], true);
1916                 if (rc)
1917                         return rc;
1918                 rc = _smcr_buf_map_lgr(lnk, &lgr->sndbufs_lock,
1919                                        &lgr->sndbufs[i], false);
1920                 if (rc)
1921                         return rc;
1922         }
1923         return 0;
1924 }
1925
1926 /* register all used buffers of lgr for a new link,
1927  * must be called under lgr->llc_conf_mutex lock
1928  */
1929 int smcr_buf_reg_lgr(struct smc_link *lnk)
1930 {
1931         struct smc_link_group *lgr = lnk->lgr;
1932         struct smc_buf_desc *buf_desc, *bf;
1933         int i, rc = 0;
1934
1935         mutex_lock(&lgr->rmbs_lock);
1936         for (i = 0; i < SMC_RMBE_SIZES; i++) {
1937                 list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list) {
1938                         if (!buf_desc->used)
1939                                 continue;
1940                         rc = smcr_link_reg_rmb(lnk, buf_desc);
1941                         if (rc)
1942                                 goto out;
1943                 }
1944         }
1945 out:
1946         mutex_unlock(&lgr->rmbs_lock);
1947         return rc;
1948 }
1949
1950 static struct smc_buf_desc *smcr_new_buf_create(struct smc_link_group *lgr,
1951                                                 bool is_rmb, int bufsize)
1952 {
1953         struct smc_buf_desc *buf_desc;
1954
1955         /* try to alloc a new buffer */
1956         buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
1957         if (!buf_desc)
1958                 return ERR_PTR(-ENOMEM);
1959
1960         buf_desc->order = get_order(bufsize);
1961         buf_desc->pages = alloc_pages(GFP_KERNEL | __GFP_NOWARN |
1962                                       __GFP_NOMEMALLOC | __GFP_COMP |
1963                                       __GFP_NORETRY | __GFP_ZERO,
1964                                       buf_desc->order);
1965         if (!buf_desc->pages) {
1966                 kfree(buf_desc);
1967                 return ERR_PTR(-EAGAIN);
1968         }
1969         buf_desc->cpu_addr = (void *)page_address(buf_desc->pages);
1970         buf_desc->len = bufsize;
1971         return buf_desc;
1972 }
1973
1974 /* map buf_desc on all usable links,
1975  * unused buffers stay mapped as long as the link is up
1976  */
1977 static int smcr_buf_map_usable_links(struct smc_link_group *lgr,
1978                                      struct smc_buf_desc *buf_desc, bool is_rmb)
1979 {
1980         int i, rc = 0;
1981
1982         /* protect against parallel link reconfiguration */
1983         mutex_lock(&lgr->llc_conf_mutex);
1984         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1985                 struct smc_link *lnk = &lgr->lnk[i];
1986
1987                 if (!smc_link_usable(lnk))
1988                         continue;
1989                 if (smcr_buf_map_link(buf_desc, is_rmb, lnk)) {
1990                         rc = -ENOMEM;
1991                         goto out;
1992                 }
1993         }
1994 out:
1995         mutex_unlock(&lgr->llc_conf_mutex);
1996         return rc;
1997 }
1998
1999 static struct smc_buf_desc *smcd_new_buf_create(struct smc_link_group *lgr,
2000                                                 bool is_dmb, int bufsize)
2001 {
2002         struct smc_buf_desc *buf_desc;
2003         int rc;
2004
2005         /* try to alloc a new DMB */
2006         buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
2007         if (!buf_desc)
2008                 return ERR_PTR(-ENOMEM);
2009         if (is_dmb) {
2010                 rc = smc_ism_register_dmb(lgr, bufsize, buf_desc);
2011                 if (rc) {
2012                         kfree(buf_desc);
2013                         if (rc == -ENOMEM)
2014                                 return ERR_PTR(-EAGAIN);
2015                         if (rc == -ENOSPC)
2016                                 return ERR_PTR(-ENOSPC);
2017                         return ERR_PTR(-EIO);
2018                 }
2019                 buf_desc->pages = virt_to_page(buf_desc->cpu_addr);
2020                 /* CDC header stored in buf. So, pretend it was smaller */
2021                 buf_desc->len = bufsize - sizeof(struct smcd_cdc_msg);
2022         } else {
2023                 buf_desc->cpu_addr = kzalloc(bufsize, GFP_KERNEL |
2024                                              __GFP_NOWARN | __GFP_NORETRY |
2025                                              __GFP_NOMEMALLOC);
2026                 if (!buf_desc->cpu_addr) {
2027                         kfree(buf_desc);
2028                         return ERR_PTR(-EAGAIN);
2029                 }
2030                 buf_desc->len = bufsize;
2031         }
2032         return buf_desc;
2033 }
2034
2035 static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
2036 {
2037         struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
2038         struct smc_connection *conn = &smc->conn;
2039         struct smc_link_group *lgr = conn->lgr;
2040         struct list_head *buf_list;
2041         int bufsize, bufsize_short;
2042         bool is_dgraded = false;
2043         struct mutex *lock;     /* lock buffer list */
2044         int sk_buf_size;
2045
2046         if (is_rmb)
2047                 /* use socket recv buffer size (w/o overhead) as start value */
2048                 sk_buf_size = smc->sk.sk_rcvbuf / 2;
2049         else
2050                 /* use socket send buffer size (w/o overhead) as start value */
2051                 sk_buf_size = smc->sk.sk_sndbuf / 2;
2052
2053         for (bufsize_short = smc_compress_bufsize(sk_buf_size, is_smcd, is_rmb);
2054              bufsize_short >= 0; bufsize_short--) {
2055                 if (is_rmb) {
2056                         lock = &lgr->rmbs_lock;
2057                         buf_list = &lgr->rmbs[bufsize_short];
2058                 } else {
2059                         lock = &lgr->sndbufs_lock;
2060                         buf_list = &lgr->sndbufs[bufsize_short];
2061                 }
2062                 bufsize = smc_uncompress_bufsize(bufsize_short);
2063
2064                 /* check for reusable slot in the link group */
2065                 buf_desc = smc_buf_get_slot(bufsize_short, lock, buf_list);
2066                 if (buf_desc) {
2067                         SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, bufsize);
2068                         SMC_STAT_BUF_REUSE(smc, is_smcd, is_rmb);
2069                         memset(buf_desc->cpu_addr, 0, bufsize);
2070                         break; /* found reusable slot */
2071                 }
2072
2073                 if (is_smcd)
2074                         buf_desc = smcd_new_buf_create(lgr, is_rmb, bufsize);
2075                 else
2076                         buf_desc = smcr_new_buf_create(lgr, is_rmb, bufsize);
2077
2078                 if (PTR_ERR(buf_desc) == -ENOMEM)
2079                         break;
2080                 if (IS_ERR(buf_desc)) {
2081                         if (!is_dgraded) {
2082                                 is_dgraded = true;
2083                                 SMC_STAT_RMB_DOWNGRADED(smc, is_smcd, is_rmb);
2084                         }
2085                         continue;
2086                 }
2087
2088                 SMC_STAT_RMB_ALLOC(smc, is_smcd, is_rmb);
2089                 SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, bufsize);
2090                 buf_desc->used = 1;
2091                 mutex_lock(lock);
2092                 list_add(&buf_desc->list, buf_list);
2093                 mutex_unlock(lock);
2094                 break; /* found */
2095         }
2096
2097         if (IS_ERR(buf_desc))
2098                 return PTR_ERR(buf_desc);
2099
2100         if (!is_smcd) {
2101                 if (smcr_buf_map_usable_links(lgr, buf_desc, is_rmb)) {
2102                         smcr_buf_unuse(buf_desc, lgr);
2103                         return -ENOMEM;
2104                 }
2105         }
2106
2107         if (is_rmb) {
2108                 conn->rmb_desc = buf_desc;
2109                 conn->rmbe_size_short = bufsize_short;
2110                 smc->sk.sk_rcvbuf = bufsize * 2;
2111                 atomic_set(&conn->bytes_to_rcv, 0);
2112                 conn->rmbe_update_limit =
2113                         smc_rmb_wnd_update_limit(buf_desc->len);
2114                 if (is_smcd)
2115                         smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */
2116         } else {
2117                 conn->sndbuf_desc = buf_desc;
2118                 smc->sk.sk_sndbuf = bufsize * 2;
2119                 atomic_set(&conn->sndbuf_space, bufsize);
2120         }
2121         return 0;
2122 }
2123
2124 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn)
2125 {
2126         if (!conn->lgr || conn->lgr->is_smcd || !smc_link_active(conn->lnk))
2127                 return;
2128         smc_ib_sync_sg_for_cpu(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE);
2129 }
2130
2131 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
2132 {
2133         if (!conn->lgr || conn->lgr->is_smcd || !smc_link_active(conn->lnk))
2134                 return;
2135         smc_ib_sync_sg_for_device(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE);
2136 }
2137
2138 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
2139 {
2140         int i;
2141
2142         if (!conn->lgr || conn->lgr->is_smcd)
2143                 return;
2144         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
2145                 if (!smc_link_active(&conn->lgr->lnk[i]))
2146                         continue;
2147                 smc_ib_sync_sg_for_cpu(&conn->lgr->lnk[i], conn->rmb_desc,
2148                                        DMA_FROM_DEVICE);
2149         }
2150 }
2151
2152 void smc_rmb_sync_sg_for_device(struct smc_connection *conn)
2153 {
2154         int i;
2155
2156         if (!conn->lgr || conn->lgr->is_smcd)
2157                 return;
2158         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
2159                 if (!smc_link_active(&conn->lgr->lnk[i]))
2160                         continue;
2161                 smc_ib_sync_sg_for_device(&conn->lgr->lnk[i], conn->rmb_desc,
2162                                           DMA_FROM_DEVICE);
2163         }
2164 }
2165
2166 /* create the send and receive buffer for an SMC socket;
2167  * receive buffers are called RMBs;
2168  * (even though the SMC protocol allows more than one RMB-element per RMB,
2169  * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
2170  * extra RMB for every connection in a link group
2171  */
2172 int smc_buf_create(struct smc_sock *smc, bool is_smcd)
2173 {
2174         int rc;
2175
2176         /* create send buffer */
2177         rc = __smc_buf_create(smc, is_smcd, false);
2178         if (rc)
2179                 return rc;
2180         /* create rmb */
2181         rc = __smc_buf_create(smc, is_smcd, true);
2182         if (rc) {
2183                 mutex_lock(&smc->conn.lgr->sndbufs_lock);
2184                 list_del(&smc->conn.sndbuf_desc->list);
2185                 mutex_unlock(&smc->conn.lgr->sndbufs_lock);
2186                 smc_buf_free(smc->conn.lgr, false, smc->conn.sndbuf_desc);
2187                 smc->conn.sndbuf_desc = NULL;
2188         }
2189         return rc;
2190 }
2191
2192 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
2193 {
2194         int i;
2195
2196         for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
2197                 if (!test_and_set_bit(i, lgr->rtokens_used_mask))
2198                         return i;
2199         }
2200         return -ENOSPC;
2201 }
2202
2203 static int smc_rtoken_find_by_link(struct smc_link_group *lgr, int lnk_idx,
2204                                    u32 rkey)
2205 {
2206         int i;
2207
2208         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2209                 if (test_bit(i, lgr->rtokens_used_mask) &&
2210                     lgr->rtokens[i][lnk_idx].rkey == rkey)
2211                         return i;
2212         }
2213         return -ENOENT;
2214 }
2215
2216 /* set rtoken for a new link to an existing rmb */
2217 void smc_rtoken_set(struct smc_link_group *lgr, int link_idx, int link_idx_new,
2218                     __be32 nw_rkey_known, __be64 nw_vaddr, __be32 nw_rkey)
2219 {
2220         int rtok_idx;
2221
2222         rtok_idx = smc_rtoken_find_by_link(lgr, link_idx, ntohl(nw_rkey_known));
2223         if (rtok_idx == -ENOENT)
2224                 return;
2225         lgr->rtokens[rtok_idx][link_idx_new].rkey = ntohl(nw_rkey);
2226         lgr->rtokens[rtok_idx][link_idx_new].dma_addr = be64_to_cpu(nw_vaddr);
2227 }
2228
2229 /* set rtoken for a new link whose link_id is given */
2230 void smc_rtoken_set2(struct smc_link_group *lgr, int rtok_idx, int link_id,
2231                      __be64 nw_vaddr, __be32 nw_rkey)
2232 {
2233         u64 dma_addr = be64_to_cpu(nw_vaddr);
2234         u32 rkey = ntohl(nw_rkey);
2235         bool found = false;
2236         int link_idx;
2237
2238         for (link_idx = 0; link_idx < SMC_LINKS_PER_LGR_MAX; link_idx++) {
2239                 if (lgr->lnk[link_idx].link_id == link_id) {
2240                         found = true;
2241                         break;
2242                 }
2243         }
2244         if (!found)
2245                 return;
2246         lgr->rtokens[rtok_idx][link_idx].rkey = rkey;
2247         lgr->rtokens[rtok_idx][link_idx].dma_addr = dma_addr;
2248 }
2249
2250 /* add a new rtoken from peer */
2251 int smc_rtoken_add(struct smc_link *lnk, __be64 nw_vaddr, __be32 nw_rkey)
2252 {
2253         struct smc_link_group *lgr = smc_get_lgr(lnk);
2254         u64 dma_addr = be64_to_cpu(nw_vaddr);
2255         u32 rkey = ntohl(nw_rkey);
2256         int i;
2257
2258         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2259                 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
2260                     lgr->rtokens[i][lnk->link_idx].dma_addr == dma_addr &&
2261                     test_bit(i, lgr->rtokens_used_mask)) {
2262                         /* already in list */
2263                         return i;
2264                 }
2265         }
2266         i = smc_rmb_reserve_rtoken_idx(lgr);
2267         if (i < 0)
2268                 return i;
2269         lgr->rtokens[i][lnk->link_idx].rkey = rkey;
2270         lgr->rtokens[i][lnk->link_idx].dma_addr = dma_addr;
2271         return i;
2272 }
2273
2274 /* delete an rtoken from all links */
2275 int smc_rtoken_delete(struct smc_link *lnk, __be32 nw_rkey)
2276 {
2277         struct smc_link_group *lgr = smc_get_lgr(lnk);
2278         u32 rkey = ntohl(nw_rkey);
2279         int i, j;
2280
2281         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2282                 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
2283                     test_bit(i, lgr->rtokens_used_mask)) {
2284                         for (j = 0; j < SMC_LINKS_PER_LGR_MAX; j++) {
2285                                 lgr->rtokens[i][j].rkey = 0;
2286                                 lgr->rtokens[i][j].dma_addr = 0;
2287                         }
2288                         clear_bit(i, lgr->rtokens_used_mask);
2289                         return 0;
2290                 }
2291         }
2292         return -ENOENT;
2293 }
2294
2295 /* save rkey and dma_addr received from peer during clc handshake */
2296 int smc_rmb_rtoken_handling(struct smc_connection *conn,
2297                             struct smc_link *lnk,
2298                             struct smc_clc_msg_accept_confirm *clc)
2299 {
2300         conn->rtoken_idx = smc_rtoken_add(lnk, clc->r0.rmb_dma_addr,
2301                                           clc->r0.rmb_rkey);
2302         if (conn->rtoken_idx < 0)
2303                 return conn->rtoken_idx;
2304         return 0;
2305 }
2306
2307 static void smc_core_going_away(void)
2308 {
2309         struct smc_ib_device *smcibdev;
2310         struct smcd_dev *smcd;
2311
2312         mutex_lock(&smc_ib_devices.mutex);
2313         list_for_each_entry(smcibdev, &smc_ib_devices.list, list) {
2314                 int i;
2315
2316                 for (i = 0; i < SMC_MAX_PORTS; i++)
2317                         set_bit(i, smcibdev->ports_going_away);
2318         }
2319         mutex_unlock(&smc_ib_devices.mutex);
2320
2321         mutex_lock(&smcd_dev_list.mutex);
2322         list_for_each_entry(smcd, &smcd_dev_list.list, list) {
2323                 smcd->going_away = 1;
2324         }
2325         mutex_unlock(&smcd_dev_list.mutex);
2326 }
2327
2328 /* Clean up all SMC link groups */
2329 static void smc_lgrs_shutdown(void)
2330 {
2331         struct smcd_dev *smcd;
2332
2333         smc_core_going_away();
2334
2335         smc_smcr_terminate_all(NULL);
2336
2337         mutex_lock(&smcd_dev_list.mutex);
2338         list_for_each_entry(smcd, &smcd_dev_list.list, list)
2339                 smc_smcd_terminate_all(smcd);
2340         mutex_unlock(&smcd_dev_list.mutex);
2341 }
2342
2343 static int smc_core_reboot_event(struct notifier_block *this,
2344                                  unsigned long event, void *ptr)
2345 {
2346         smc_lgrs_shutdown();
2347         smc_ib_unregister_client();
2348         return 0;
2349 }
2350
2351 static struct notifier_block smc_reboot_notifier = {
2352         .notifier_call = smc_core_reboot_event,
2353 };
2354
2355 int __init smc_core_init(void)
2356 {
2357         return register_reboot_notifier(&smc_reboot_notifier);
2358 }
2359
2360 /* Called (from smc_exit) when module is removed */
2361 void smc_core_exit(void)
2362 {
2363         unregister_reboot_notifier(&smc_reboot_notifier);
2364         smc_lgrs_shutdown();
2365 }