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                         conn->lgr = NULL;
 175                         return rc;
 176                 }
 177         }
 178         /* find a new alert_token_local value not yet used by some connection
 179          * in this link group
 180          */
 181         sock_hold(&smc->sk); /* sock_put in smc_lgr_unregister_conn() */
 182         while (!conn->alert_token_local) {
 183                 conn->alert_token_local = atomic_inc_return(&nexttoken);
 184                 if (smc_lgr_find_conn(conn->alert_token_local, conn->lgr))
 185                         conn->alert_token_local = 0;
 186         }
 187         smc_lgr_add_alert_token(conn);
 188         conn->lgr->conns_num++;
 189         return 0;
 190 }
 191
 192 /* Unregister connection and reset the alert token of the given connection<
 193  */
 194 static void __smc_lgr_unregister_conn(struct smc_connection *conn)
 195 {
 196         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
 197         struct smc_link_group *lgr = conn->lgr;
 198
 199         rb_erase(&conn->alert_node, &lgr->conns_all);
 200         if (conn->lnk)
 201                 atomic_dec(&conn->lnk->conn_cnt);
 202         lgr->conns_num--;
 203         conn->alert_token_local = 0;
 204         sock_put(&smc->sk); /* sock_hold in smc_lgr_register_conn() */
 205 }
 206
 207 /* Unregister connection from lgr
 208  */
 209 static void smc_lgr_unregister_conn(struct smc_connection *conn)
 210 {
 211         struct smc_link_group *lgr = conn->lgr;
 212
 213         if (!lgr)
 214                 return;
 215         write_lock_bh(&lgr->conns_lock);
 216         if (conn->alert_token_local) {
 217                 __smc_lgr_unregister_conn(conn);
 218         }
 219         write_unlock_bh(&lgr->conns_lock);
 220         conn->lgr = NULL;
 221 }
 222
 223 int smc_nl_get_sys_info(struct sk_buff *skb, struct netlink_callback *cb)
 224 {
 225         struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
 226         char hostname[SMC_MAX_HOSTNAME_LEN + 1];
 227         char smc_seid[SMC_MAX_EID_LEN + 1];
 228         struct smcd_dev *smcd_dev;
 229         struct nlattr *attrs;
 230         u8 *seid = NULL;
 231         u8 *host = NULL;
 232         void *nlh;
 233
 234         nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
 235                           &smc_gen_nl_family, NLM_F_MULTI,
 236                           SMC_NETLINK_GET_SYS_INFO);
 237         if (!nlh)
 238                 goto errmsg;
 239         if (cb_ctx->pos[0])
 240                 goto errout;
 241         attrs = nla_nest_start(skb, SMC_GEN_SYS_INFO);
 242         if (!attrs)
 243                 goto errout;
 244         if (nla_put_u8(skb, SMC_NLA_SYS_VER, SMC_V2))
 245                 goto errattr;
 246         if (nla_put_u8(skb, SMC_NLA_SYS_REL, SMC_RELEASE))
 247                 goto errattr;
 248         if (nla_put_u8(skb, SMC_NLA_SYS_IS_ISM_V2, smc_ism_is_v2_capable()))
 249                 goto errattr;
 250         smc_clc_get_hostname(&host);
 251         if (host) {
 252                 memcpy(hostname, host, SMC_MAX_HOSTNAME_LEN);
 253                 hostname[SMC_MAX_HOSTNAME_LEN] = 0;
 254                 if (nla_put_string(skb, SMC_NLA_SYS_LOCAL_HOST, hostname))
 255                         goto errattr;
 256         }
 257         mutex_lock(&smcd_dev_list.mutex);
 258         smcd_dev = list_first_entry_or_null(&smcd_dev_list.list,
 259                                             struct smcd_dev, list);
 260         if (smcd_dev)
 261                 smc_ism_get_system_eid(smcd_dev, &seid);
 262         mutex_unlock(&smcd_dev_list.mutex);
 263         if (seid && smc_ism_is_v2_capable()) {
 264                 memcpy(smc_seid, seid, SMC_MAX_EID_LEN);
 265                 smc_seid[SMC_MAX_EID_LEN] = 0;
 266                 if (nla_put_string(skb, SMC_NLA_SYS_SEID, smc_seid))
 267                         goto errattr;
 268         }
 269         nla_nest_end(skb, attrs);
 270         genlmsg_end(skb, nlh);
 271         cb_ctx->pos[0] = 1;
 272         return skb->len;
 273
 274 errattr:
 275         nla_nest_cancel(skb, attrs);
 276 errout:
 277         genlmsg_cancel(skb, nlh);
 278 errmsg:
 279         return skb->len;
 280 }
 281
 282 static int smc_nl_fill_lgr(struct smc_link_group *lgr,
 283                            struct sk_buff *skb,
 284                            struct netlink_callback *cb)
 285 {
 286         char smc_target[SMC_MAX_PNETID_LEN + 1];
 287         struct nlattr *attrs;
 288
 289         attrs = nla_nest_start(skb, SMC_GEN_LGR_SMCR);
 290         if (!attrs)
 291                 goto errout;
 292
 293         if (nla_put_u32(skb, SMC_NLA_LGR_R_ID, *((u32 *)&lgr->id)))
 294                 goto errattr;
 295         if (nla_put_u32(skb, SMC_NLA_LGR_R_CONNS_NUM, lgr->conns_num))
 296                 goto errattr;
 297         if (nla_put_u8(skb, SMC_NLA_LGR_R_ROLE, lgr->role))
 298                 goto errattr;
 299         if (nla_put_u8(skb, SMC_NLA_LGR_R_TYPE, lgr->type))
 300                 goto errattr;
 301         if (nla_put_u8(skb, SMC_NLA_LGR_R_VLAN_ID, lgr->vlan_id))
 302                 goto errattr;
 303         memcpy(smc_target, lgr->pnet_id, SMC_MAX_PNETID_LEN);
 304         smc_target[SMC_MAX_PNETID_LEN] = 0;
 305         if (nla_put_string(skb, SMC_NLA_LGR_R_PNETID, smc_target))
 306                 goto errattr;
 307
 308         nla_nest_end(skb, attrs);
 309         return 0;
 310 errattr:
 311         nla_nest_cancel(skb, attrs);
 312 errout:
 313         return -EMSGSIZE;
 314 }
 315
 316 static int smc_nl_fill_lgr_link(struct smc_link_group *lgr,
 317                                 struct smc_link *link,
 318                                 struct sk_buff *skb,
 319                                 struct netlink_callback *cb)
 320 {
 321         char smc_ibname[IB_DEVICE_NAME_MAX];
 322         u8 smc_gid_target[41];
 323         struct nlattr *attrs;
 324         u32 link_uid = 0;
 325         void *nlh;
 326
 327         nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
 328                           &smc_gen_nl_family, NLM_F_MULTI,
 329                           SMC_NETLINK_GET_LINK_SMCR);
 330         if (!nlh)
 331                 goto errmsg;
 332
 333         attrs = nla_nest_start(skb, SMC_GEN_LINK_SMCR);
 334         if (!attrs)
 335                 goto errout;
 336
 337         if (nla_put_u8(skb, SMC_NLA_LINK_ID, link->link_id))
 338                 goto errattr;
 339         if (nla_put_u32(skb, SMC_NLA_LINK_STATE, link->state))
 340                 goto errattr;
 341         if (nla_put_u32(skb, SMC_NLA_LINK_CONN_CNT,
 342                         atomic_read(&link->conn_cnt)))
 343                 goto errattr;
 344         if (nla_put_u8(skb, SMC_NLA_LINK_IB_PORT, link->ibport))
 345                 goto errattr;
 346         if (nla_put_u32(skb, SMC_NLA_LINK_NET_DEV, link->ndev_ifidx))
 347                 goto errattr;
 348         snprintf(smc_ibname, sizeof(smc_ibname), "%s", link->ibname);
 349         if (nla_put_string(skb, SMC_NLA_LINK_IB_DEV, smc_ibname))
 350                 goto errattr;
 351         memcpy(&link_uid, link->link_uid, sizeof(link_uid));
 352         if (nla_put_u32(skb, SMC_NLA_LINK_UID, link_uid))
 353                 goto errattr;
 354         memcpy(&link_uid, link->peer_link_uid, sizeof(link_uid));
 355         if (nla_put_u32(skb, SMC_NLA_LINK_PEER_UID, link_uid))
 356                 goto errattr;
 357         memset(smc_gid_target, 0, sizeof(smc_gid_target));
 358         smc_gid_be16_convert(smc_gid_target, link->gid);
 359         if (nla_put_string(skb, SMC_NLA_LINK_GID, smc_gid_target))
 360                 goto errattr;
 361         memset(smc_gid_target, 0, sizeof(smc_gid_target));
 362         smc_gid_be16_convert(smc_gid_target, link->peer_gid);
 363         if (nla_put_string(skb, SMC_NLA_LINK_PEER_GID, smc_gid_target))
 364                 goto errattr;
 365
 366         nla_nest_end(skb, attrs);
 367         genlmsg_end(skb, nlh);
 368         return 0;
 369 errattr:
 370         nla_nest_cancel(skb, attrs);
 371 errout:
 372         genlmsg_cancel(skb, nlh);
 373 errmsg:
 374         return -EMSGSIZE;
 375 }
 376
 377 static int smc_nl_handle_lgr(struct smc_link_group *lgr,
 378                              struct sk_buff *skb,
 379                              struct netlink_callback *cb,
 380                              bool list_links)
 381 {
 382         void *nlh;
 383         int i;
 384
 385         nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
 386                           &smc_gen_nl_family, NLM_F_MULTI,
 387                           SMC_NETLINK_GET_LGR_SMCR);
 388         if (!nlh)
 389                 goto errmsg;
 390         if (smc_nl_fill_lgr(lgr, skb, cb))
 391                 goto errout;
 392
 393         genlmsg_end(skb, nlh);
 394         if (!list_links)
 395                 goto out;
 396         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 397                 if (!smc_link_usable(&lgr->lnk[i]))
 398                         continue;
 399                 if (smc_nl_fill_lgr_link(lgr, &lgr->lnk[i], skb, cb))
 400                         goto errout;
 401         }
 402 out:
 403         return 0;
 404
 405 errout:
 406         genlmsg_cancel(skb, nlh);
 407 errmsg:
 408         return -EMSGSIZE;
 409 }
 410
 411 static void smc_nl_fill_lgr_list(struct smc_lgr_list *smc_lgr,
 412                                  struct sk_buff *skb,
 413                                  struct netlink_callback *cb,
 414                                  bool list_links)
 415 {
 416         struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
 417         struct smc_link_group *lgr;
 418         int snum = cb_ctx->pos[0];
 419         int num = 0;
 420
 421         spin_lock_bh(&smc_lgr->lock);
 422         list_for_each_entry(lgr, &smc_lgr->list, list) {
 423                 if (num < snum)
 424                         goto next;
 425                 if (smc_nl_handle_lgr(lgr, skb, cb, list_links))
 426                         goto errout;
 427 next:
 428                 num++;
 429         }
 430 errout:
 431         spin_unlock_bh(&smc_lgr->lock);
 432         cb_ctx->pos[0] = num;
 433 }
 434
 435 static int smc_nl_fill_smcd_lgr(struct smc_link_group *lgr,
 436                                 struct sk_buff *skb,
 437                                 struct netlink_callback *cb)
 438 {
 439         char smc_host[SMC_MAX_HOSTNAME_LEN + 1];
 440         char smc_pnet[SMC_MAX_PNETID_LEN + 1];
 441         char smc_eid[SMC_MAX_EID_LEN + 1];
 442         struct nlattr *v2_attrs;
 443         struct nlattr *attrs;
 444         void *nlh;
 445
 446         nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
 447                           &smc_gen_nl_family, NLM_F_MULTI,
 448                           SMC_NETLINK_GET_LGR_SMCD);
 449         if (!nlh)
 450                 goto errmsg;
 451
 452         attrs = nla_nest_start(skb, SMC_GEN_LGR_SMCD);
 453         if (!attrs)
 454                 goto errout;
 455
 456         if (nla_put_u32(skb, SMC_NLA_LGR_D_ID, *((u32 *)&lgr->id)))
 457                 goto errattr;
 458         if (nla_put_u64_64bit(skb, SMC_NLA_LGR_D_GID, lgr->smcd->local_gid,
 459                               SMC_NLA_LGR_D_PAD))
 460                 goto errattr;
 461         if (nla_put_u64_64bit(skb, SMC_NLA_LGR_D_PEER_GID, lgr->peer_gid,
 462                               SMC_NLA_LGR_D_PAD))
 463                 goto errattr;
 464         if (nla_put_u8(skb, SMC_NLA_LGR_D_VLAN_ID, lgr->vlan_id))
 465                 goto errattr;
 466         if (nla_put_u32(skb, SMC_NLA_LGR_D_CONNS_NUM, lgr->conns_num))
 467                 goto errattr;
 468         if (nla_put_u32(skb, SMC_NLA_LGR_D_CHID, smc_ism_get_chid(lgr->smcd)))
 469                 goto errattr;
 470         memcpy(smc_pnet, lgr->smcd->pnetid, SMC_MAX_PNETID_LEN);
 471         smc_pnet[SMC_MAX_PNETID_LEN] = 0;
 472         if (nla_put_string(skb, SMC_NLA_LGR_D_PNETID, smc_pnet))
 473                 goto errattr;
 474
 475         v2_attrs = nla_nest_start(skb, SMC_NLA_LGR_V2);
 476         if (!v2_attrs)
 477                 goto errattr;
 478         if (nla_put_u8(skb, SMC_NLA_LGR_V2_VER, lgr->smc_version))
 479                 goto errv2attr;
 480         if (nla_put_u8(skb, SMC_NLA_LGR_V2_REL, lgr->peer_smc_release))
 481                 goto errv2attr;
 482         if (nla_put_u8(skb, SMC_NLA_LGR_V2_OS, lgr->peer_os))
 483                 goto errv2attr;
 484         memcpy(smc_host, lgr->peer_hostname, SMC_MAX_HOSTNAME_LEN);
 485         smc_host[SMC_MAX_HOSTNAME_LEN] = 0;
 486         if (nla_put_string(skb, SMC_NLA_LGR_V2_PEER_HOST, smc_host))
 487                 goto errv2attr;
 488         memcpy(smc_eid, lgr->negotiated_eid, SMC_MAX_EID_LEN);
 489         smc_eid[SMC_MAX_EID_LEN] = 0;
 490         if (nla_put_string(skb, SMC_NLA_LGR_V2_NEG_EID, smc_eid))
 491                 goto errv2attr;
 492
 493         nla_nest_end(skb, v2_attrs);
 494         nla_nest_end(skb, attrs);
 495         genlmsg_end(skb, nlh);
 496         return 0;
 497
 498 errv2attr:
 499         nla_nest_cancel(skb, v2_attrs);
 500 errattr:
 501         nla_nest_cancel(skb, attrs);
 502 errout:
 503         genlmsg_cancel(skb, nlh);
 504 errmsg:
 505         return -EMSGSIZE;
 506 }
 507
 508 static int smc_nl_handle_smcd_lgr(struct smcd_dev *dev,
 509                                   struct sk_buff *skb,
 510                                   struct netlink_callback *cb)
 511 {
 512         struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
 513         struct smc_link_group *lgr;
 514         int snum = cb_ctx->pos[1];
 515         int rc = 0, num = 0;
 516
 517         spin_lock_bh(&dev->lgr_lock);
 518         list_for_each_entry(lgr, &dev->lgr_list, list) {
 519                 if (!lgr->is_smcd)
 520                         continue;
 521                 if (num < snum)
 522                         goto next;
 523                 rc = smc_nl_fill_smcd_lgr(lgr, skb, cb);
 524                 if (rc)
 525                         goto errout;
 526 next:
 527                 num++;
 528         }
 529 errout:
 530         spin_unlock_bh(&dev->lgr_lock);
 531         cb_ctx->pos[1] = num;
 532         return rc;
 533 }
 534
 535 static int smc_nl_fill_smcd_dev(struct smcd_dev_list *dev_list,
 536                                 struct sk_buff *skb,
 537                                 struct netlink_callback *cb)
 538 {
 539         struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
 540         struct smcd_dev *smcd_dev;
 541         int snum = cb_ctx->pos[0];
 542         int rc = 0, num = 0;
 543
 544         mutex_lock(&dev_list->mutex);
 545         list_for_each_entry(smcd_dev, &dev_list->list, list) {
 546                 if (list_empty(&smcd_dev->lgr_list))
 547                         continue;
 548                 if (num < snum)
 549                         goto next;
 550                 rc = smc_nl_handle_smcd_lgr(smcd_dev, skb, cb);
 551                 if (rc)
 552                         goto errout;
 553 next:
 554                 num++;
 555         }
 556 errout:
 557         mutex_unlock(&dev_list->mutex);
 558         cb_ctx->pos[0] = num;
 559         return rc;
 560 }
 561
 562 int smcr_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb)
 563 {
 564         bool list_links = false;
 565
 566         smc_nl_fill_lgr_list(&smc_lgr_list, skb, cb, list_links);
 567         return skb->len;
 568 }
 569
 570 int smcr_nl_get_link(struct sk_buff *skb, struct netlink_callback *cb)
 571 {
 572         bool list_links = true;
 573
 574         smc_nl_fill_lgr_list(&smc_lgr_list, skb, cb, list_links);
 575         return skb->len;
 576 }
 577
 578 int smcd_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb)
 579 {
 580         smc_nl_fill_smcd_dev(&smcd_dev_list, skb, cb);
 581         return skb->len;
 582 }
 583
 584 void smc_lgr_cleanup_early(struct smc_link_group *lgr)
 585 {
 586         spinlock_t *lgr_lock;
 587
 588         if (!lgr)
 589                 return;
 590
 591         smc_lgr_list_head(lgr, &lgr_lock);
 592         spin_lock_bh(lgr_lock);
 593         /* do not use this link group for new connections */
 594         if (!list_empty(&lgr->list))
 595                 list_del_init(&lgr->list);
 596         spin_unlock_bh(lgr_lock);
 597         __smc_lgr_terminate(lgr, true);
 598 }
 599
 600 static void smcr_lgr_link_deactivate_all(struct smc_link_group *lgr)
 601 {
 602         int i;
 603
 604         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 605                 struct smc_link *lnk = &lgr->lnk[i];
 606
 607                 if (smc_link_sendable(lnk))
 608                         lnk->state = SMC_LNK_INACTIVE;
 609         }
 610         wake_up_all(&lgr->llc_msg_waiter);
 611         wake_up_all(&lgr->llc_flow_waiter);
 612 }
 613
 614 static void smc_lgr_free(struct smc_link_group *lgr);
 615
 616 static void smc_lgr_free_work(struct work_struct *work)
 617 {
 618         struct smc_link_group *lgr = container_of(to_delayed_work(work),
 619                                                   struct smc_link_group,
 620                                                   free_work);
 621         spinlock_t *lgr_lock;
 622         bool conns;
 623
 624         smc_lgr_list_head(lgr, &lgr_lock);
 625         spin_lock_bh(lgr_lock);
 626         if (lgr->freeing) {
 627                 spin_unlock_bh(lgr_lock);
 628                 return;
 629         }
 630         read_lock_bh(&lgr->conns_lock);
 631         conns = RB_EMPTY_ROOT(&lgr->conns_all);
 632         read_unlock_bh(&lgr->conns_lock);
 633         if (!conns) { /* number of lgr connections is no longer zero */
 634                 spin_unlock_bh(lgr_lock);
 635                 return;
 636         }
 637         list_del_init(&lgr->list); /* remove from smc_lgr_list */
 638         lgr->freeing = 1; /* this instance does the freeing, no new schedule */
 639         spin_unlock_bh(lgr_lock);
 640         cancel_delayed_work(&lgr->free_work);
 641
 642         if (!lgr->is_smcd && !lgr->terminating)
 643                 smc_llc_send_link_delete_all(lgr, true,
 644                                              SMC_LLC_DEL_PROG_INIT_TERM);
 645         if (lgr->is_smcd && !lgr->terminating)
 646                 smc_ism_signal_shutdown(lgr);
 647         if (!lgr->is_smcd)
 648                 smcr_lgr_link_deactivate_all(lgr);
 649         smc_lgr_free(lgr);
 650 }
 651
 652 static void smc_lgr_terminate_work(struct work_struct *work)
 653 {
 654         struct smc_link_group *lgr = container_of(work, struct smc_link_group,
 655                                                   terminate_work);
 656
 657         __smc_lgr_terminate(lgr, true);
 658 }
 659
 660 /* return next unique link id for the lgr */
 661 static u8 smcr_next_link_id(struct smc_link_group *lgr)
 662 {
 663         u8 link_id;
 664         int i;
 665
 666         while (1) {
 667 again:
 668                 link_id = ++lgr->next_link_id;
 669                 if (!link_id)   /* skip zero as link_id */
 670                         link_id = ++lgr->next_link_id;
 671                 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 672                         if (smc_link_usable(&lgr->lnk[i]) &&
 673                             lgr->lnk[i].link_id == link_id)
 674                                 goto again;
 675                 }
 676                 break;
 677         }
 678         return link_id;
 679 }
 680
 681 static void smcr_copy_dev_info_to_link(struct smc_link *link)
 682 {
 683         struct smc_ib_device *smcibdev = link->smcibdev;
 684
 685         snprintf(link->ibname, sizeof(link->ibname), "%s",
 686                  smcibdev->ibdev->name);
 687         link->ndev_ifidx = smcibdev->ndev_ifidx[link->ibport - 1];
 688 }
 689
 690 int smcr_link_init(struct smc_link_group *lgr, struct smc_link *lnk,
 691                    u8 link_idx, struct smc_init_info *ini)
 692 {
 693         u8 rndvec[3];
 694         int rc;
 695
 696         get_device(&ini->ib_dev->ibdev->dev);
 697         atomic_inc(&ini->ib_dev->lnk_cnt);
 698         lnk->link_id = smcr_next_link_id(lgr);
 699         lnk->lgr = lgr;
 700         lnk->link_idx = link_idx;
 701         lnk->smcibdev = ini->ib_dev;
 702         lnk->ibport = ini->ib_port;
 703         smc_ibdev_cnt_inc(lnk);
 704         smcr_copy_dev_info_to_link(lnk);
 705         lnk->path_mtu = ini->ib_dev->pattr[ini->ib_port - 1].active_mtu;
 706         atomic_set(&lnk->conn_cnt, 0);
 707         smc_llc_link_set_uid(lnk);
 708         INIT_WORK(&lnk->link_down_wrk, smc_link_down_work);
 709         if (!ini->ib_dev->initialized) {
 710                 rc = (int)smc_ib_setup_per_ibdev(ini->ib_dev);
 711                 if (rc)
 712                         goto out;
 713         }
 714         get_random_bytes(rndvec, sizeof(rndvec));
 715         lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) +
 716                 (rndvec[2] << 16);
 717         rc = smc_ib_determine_gid(lnk->smcibdev, lnk->ibport,
 718                                   ini->vlan_id, lnk->gid, &lnk->sgid_index);
 719         if (rc)
 720                 goto out;
 721         rc = smc_llc_link_init(lnk);
 722         if (rc)
 723                 goto out;
 724         rc = smc_wr_alloc_link_mem(lnk);
 725         if (rc)
 726                 goto clear_llc_lnk;
 727         rc = smc_ib_create_protection_domain(lnk);
 728         if (rc)
 729                 goto free_link_mem;
 730         rc = smc_ib_create_queue_pair(lnk);
 731         if (rc)
 732                 goto dealloc_pd;
 733         rc = smc_wr_create_link(lnk);
 734         if (rc)
 735                 goto destroy_qp;
 736         lnk->state = SMC_LNK_ACTIVATING;
 737         return 0;
 738
 739 destroy_qp:
 740         smc_ib_destroy_queue_pair(lnk);
 741 dealloc_pd:
 742         smc_ib_dealloc_protection_domain(lnk);
 743 free_link_mem:
 744         smc_wr_free_link_mem(lnk);
 745 clear_llc_lnk:
 746         smc_llc_link_clear(lnk, false);
 747 out:
 748         smc_ibdev_cnt_dec(lnk);
 749         put_device(&ini->ib_dev->ibdev->dev);
 750         memset(lnk, 0, sizeof(struct smc_link));
 751         lnk->state = SMC_LNK_UNUSED;
 752         if (!atomic_dec_return(&ini->ib_dev->lnk_cnt))
 753                 wake_up(&ini->ib_dev->lnks_deleted);
 754         return rc;
 755 }
 756
 757 /* create a new SMC link group */
 758 static int smc_lgr_create(struct smc_sock *smc, struct smc_init_info *ini)
 759 {
 760         struct smc_link_group *lgr;
 761         struct list_head *lgr_list;
 762         struct smc_link *lnk;
 763         spinlock_t *lgr_lock;
 764         u8 link_idx;
 765         int rc = 0;
 766         int i;
 767
 768         if (ini->is_smcd && ini->vlan_id) {
 769                 if (smc_ism_get_vlan(ini->ism_dev[ini->ism_selected],
 770                                      ini->vlan_id)) {
 771                         rc = SMC_CLC_DECL_ISMVLANERR;
 772                         goto out;
 773                 }
 774         }
 775
 776         lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
 777         if (!lgr) {
 778                 rc = SMC_CLC_DECL_MEM;
 779                 goto ism_put_vlan;
 780         }
 781         lgr->tx_wq = alloc_workqueue("smc_tx_wq-%*phN", 0, 0,
 782                                      SMC_LGR_ID_SIZE, &lgr->id);
 783         if (!lgr->tx_wq) {
 784                 rc = -ENOMEM;
 785                 goto free_lgr;
 786         }
 787         lgr->is_smcd = ini->is_smcd;
 788         lgr->sync_err = 0;
 789         lgr->terminating = 0;
 790         lgr->freeing = 0;
 791         lgr->vlan_id = ini->vlan_id;
 792         mutex_init(&lgr->sndbufs_lock);
 793         mutex_init(&lgr->rmbs_lock);
 794         rwlock_init(&lgr->conns_lock);
 795         for (i = 0; i < SMC_RMBE_SIZES; i++) {
 796                 INIT_LIST_HEAD(&lgr->sndbufs[i]);
 797                 INIT_LIST_HEAD(&lgr->rmbs[i]);
 798         }
 799         lgr->next_link_id = 0;
 800         smc_lgr_list.num += SMC_LGR_NUM_INCR;
 801         memcpy(&lgr->id, (u8 *)&smc_lgr_list.num, SMC_LGR_ID_SIZE);
 802         INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
 803         INIT_WORK(&lgr->terminate_work, smc_lgr_terminate_work);
 804         lgr->conns_all = RB_ROOT;
 805         if (ini->is_smcd) {
 806                 /* SMC-D specific settings */
 807                 get_device(&ini->ism_dev[ini->ism_selected]->dev);
 808                 lgr->peer_gid = ini->ism_peer_gid[ini->ism_selected];
 809                 lgr->smcd = ini->ism_dev[ini->ism_selected];
 810                 lgr_list = &ini->ism_dev[ini->ism_selected]->lgr_list;
 811                 lgr_lock = &lgr->smcd->lgr_lock;
 812                 lgr->smc_version = ini->smcd_version;
 813                 lgr->peer_shutdown = 0;
 814                 atomic_inc(&ini->ism_dev[ini->ism_selected]->lgr_cnt);
 815         } else {
 816                 /* SMC-R specific settings */
 817                 lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
 818                 memcpy(lgr->peer_systemid, ini->ib_lcl->id_for_peer,
 819                        SMC_SYSTEMID_LEN);
 820                 memcpy(lgr->pnet_id, ini->ib_dev->pnetid[ini->ib_port - 1],
 821                        SMC_MAX_PNETID_LEN);
 822                 smc_llc_lgr_init(lgr, smc);
 823
 824                 link_idx = SMC_SINGLE_LINK;
 825                 lnk = &lgr->lnk[link_idx];
 826                 rc = smcr_link_init(lgr, lnk, link_idx, ini);
 827                 if (rc)
 828                         goto free_wq;
 829                 lgr_list = &smc_lgr_list.list;
 830                 lgr_lock = &smc_lgr_list.lock;
 831                 atomic_inc(&lgr_cnt);
 832         }
 833         smc->conn.lgr = lgr;
 834         spin_lock_bh(lgr_lock);
 835         list_add_tail(&lgr->list, lgr_list);
 836         spin_unlock_bh(lgr_lock);
 837         return 0;
 838
 839 free_wq:
 840         destroy_workqueue(lgr->tx_wq);
 841 free_lgr:
 842         kfree(lgr);
 843 ism_put_vlan:
 844         if (ini->is_smcd && ini->vlan_id)
 845                 smc_ism_put_vlan(ini->ism_dev[ini->ism_selected], ini->vlan_id);
 846 out:
 847         if (rc < 0) {
 848                 if (rc == -ENOMEM)
 849                         rc = SMC_CLC_DECL_MEM;
 850                 else
 851                         rc = SMC_CLC_DECL_INTERR;
 852         }
 853         return rc;
 854 }
 855
 856 static int smc_write_space(struct smc_connection *conn)
 857 {
 858         int buffer_len = conn->peer_rmbe_size;
 859         union smc_host_cursor prod;
 860         union smc_host_cursor cons;
 861         int space;
 862
 863         smc_curs_copy(&prod, &conn->local_tx_ctrl.prod, conn);
 864         smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
 865         /* determine rx_buf space */
 866         space = buffer_len - smc_curs_diff(buffer_len, &cons, &prod);
 867         return space;
 868 }
 869
 870 static int smc_switch_cursor(struct smc_sock *smc, struct smc_cdc_tx_pend *pend,
 871                              struct smc_wr_buf *wr_buf)
 872 {
 873         struct smc_connection *conn = &smc->conn;
 874         union smc_host_cursor cons, fin;
 875         int rc = 0;
 876         int diff;
 877
 878         smc_curs_copy(&conn->tx_curs_sent, &conn->tx_curs_fin, conn);
 879         smc_curs_copy(&fin, &conn->local_tx_ctrl_fin, conn);
 880         /* set prod cursor to old state, enforce tx_rdma_writes() */
 881         smc_curs_copy(&conn->local_tx_ctrl.prod, &fin, conn);
 882         smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
 883
 884         if (smc_curs_comp(conn->peer_rmbe_size, &cons, &fin) < 0) {
 885                 /* cons cursor advanced more than fin, and prod was set
 886                  * fin above, so now prod is smaller than cons. Fix that.
 887                  */
 888                 diff = smc_curs_diff(conn->peer_rmbe_size, &fin, &cons);
 889                 smc_curs_add(conn->sndbuf_desc->len,
 890                              &conn->tx_curs_sent, diff);
 891                 smc_curs_add(conn->sndbuf_desc->len,
 892                              &conn->tx_curs_fin, diff);
 893
 894                 smp_mb__before_atomic();
 895                 atomic_add(diff, &conn->sndbuf_space);
 896                 smp_mb__after_atomic();
 897
 898                 smc_curs_add(conn->peer_rmbe_size,
 899                              &conn->local_tx_ctrl.prod, diff);
 900                 smc_curs_add(conn->peer_rmbe_size,
 901                              &conn->local_tx_ctrl_fin, diff);
 902         }
 903         /* recalculate, value is used by tx_rdma_writes() */
 904         atomic_set(&smc->conn.peer_rmbe_space, smc_write_space(conn));
 905
 906         if (smc->sk.sk_state != SMC_INIT &&
 907             smc->sk.sk_state != SMC_CLOSED) {
 908                 rc = smcr_cdc_msg_send_validation(conn, pend, wr_buf);
 909                 if (!rc) {
 910                         queue_delayed_work(conn->lgr->tx_wq, &conn->tx_work, 0);
 911                         smc->sk.sk_data_ready(&smc->sk);
 912                 }
 913         } else {
 914                 smc_wr_tx_put_slot(conn->lnk,
 915                                    (struct smc_wr_tx_pend_priv *)pend);
 916         }
 917         return rc;
 918 }
 919
 920 void smc_switch_link_and_count(struct smc_connection *conn,
 921                                struct smc_link *to_lnk)
 922 {
 923         atomic_dec(&conn->lnk->conn_cnt);
 924         conn->lnk = to_lnk;
 925         atomic_inc(&conn->lnk->conn_cnt);
 926 }
 927
 928 struct smc_link *smc_switch_conns(struct smc_link_group *lgr,
 929                                   struct smc_link *from_lnk, bool is_dev_err)
 930 {
 931         struct smc_link *to_lnk = NULL;
 932         struct smc_cdc_tx_pend *pend;
 933         struct smc_connection *conn;
 934         struct smc_wr_buf *wr_buf;
 935         struct smc_sock *smc;
 936         struct rb_node *node;
 937         int i, rc = 0;
 938
 939         /* link is inactive, wake up tx waiters */
 940         smc_wr_wakeup_tx_wait(from_lnk);
 941
 942         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 943                 if (!smc_link_active(&lgr->lnk[i]) || i == from_lnk->link_idx)
 944                         continue;
 945                 if (is_dev_err && from_lnk->smcibdev == lgr->lnk[i].smcibdev &&
 946                     from_lnk->ibport == lgr->lnk[i].ibport) {
 947                         continue;
 948                 }
 949                 to_lnk = &lgr->lnk[i];
 950                 break;
 951         }
 952         if (!to_lnk || !smc_wr_tx_link_hold(to_lnk)) {
 953                 smc_lgr_terminate_sched(lgr);
 954                 return NULL;
 955         }
 956 again:
 957         read_lock_bh(&lgr->conns_lock);
 958         for (node = rb_first(&lgr->conns_all); node; node = rb_next(node)) {
 959                 conn = rb_entry(node, struct smc_connection, alert_node);
 960                 if (conn->lnk != from_lnk)
 961                         continue;
 962                 smc = container_of(conn, struct smc_sock, conn);
 963                 /* conn->lnk not yet set in SMC_INIT state */
 964                 if (smc->sk.sk_state == SMC_INIT)
 965                         continue;
 966                 if (smc->sk.sk_state == SMC_CLOSED ||
 967                     smc->sk.sk_state == SMC_PEERCLOSEWAIT1 ||
 968                     smc->sk.sk_state == SMC_PEERCLOSEWAIT2 ||
 969                     smc->sk.sk_state == SMC_APPFINCLOSEWAIT ||
 970                     smc->sk.sk_state == SMC_APPCLOSEWAIT1 ||
 971                     smc->sk.sk_state == SMC_APPCLOSEWAIT2 ||
 972                     smc->sk.sk_state == SMC_PEERFINCLOSEWAIT ||
 973                     smc->sk.sk_state == SMC_PEERABORTWAIT ||
 974                     smc->sk.sk_state == SMC_PROCESSABORT) {
 975                         spin_lock_bh(&conn->send_lock);
 976                         smc_switch_link_and_count(conn, to_lnk);
 977                         spin_unlock_bh(&conn->send_lock);
 978                         continue;
 979                 }
 980                 sock_hold(&smc->sk);
 981                 read_unlock_bh(&lgr->conns_lock);
 982                 /* pre-fetch buffer outside of send_lock, might sleep */
 983                 rc = smc_cdc_get_free_slot(conn, to_lnk, &wr_buf, NULL, &pend);
 984                 if (rc)
 985                         goto err_out;
 986                 /* avoid race with smcr_tx_sndbuf_nonempty() */
 987                 spin_lock_bh(&conn->send_lock);
 988                 smc_switch_link_and_count(conn, to_lnk);
 989                 rc = smc_switch_cursor(smc, pend, wr_buf);
 990                 spin_unlock_bh(&conn->send_lock);
 991                 sock_put(&smc->sk);
 992                 if (rc)
 993                         goto err_out;
 994                 goto again;
 995         }
 996         read_unlock_bh(&lgr->conns_lock);
 997         smc_wr_tx_link_put(to_lnk);
 998         return to_lnk;
 999
1000 err_out:
1001         smcr_link_down_cond_sched(to_lnk);
1002         smc_wr_tx_link_put(to_lnk);
1003         return NULL;
1004 }
1005
1006 static void smcr_buf_unuse(struct smc_buf_desc *rmb_desc,
1007                            struct smc_link_group *lgr)
1008 {
1009         int rc;
1010
1011         if (rmb_desc->is_conf_rkey && !list_empty(&lgr->list)) {
1012                 /* unregister rmb with peer */
1013                 rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY);
1014                 if (!rc) {
1015                         /* protect against smc_llc_cli_rkey_exchange() */
1016                         mutex_lock(&lgr->llc_conf_mutex);
1017                         smc_llc_do_delete_rkey(lgr, rmb_desc);
1018                         rmb_desc->is_conf_rkey = false;
1019                         mutex_unlock(&lgr->llc_conf_mutex);
1020                         smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl);
1021                 }
1022         }
1023
1024         if (rmb_desc->is_reg_err) {
1025                 /* buf registration failed, reuse not possible */
1026                 mutex_lock(&lgr->rmbs_lock);
1027                 list_del(&rmb_desc->list);
1028                 mutex_unlock(&lgr->rmbs_lock);
1029
1030                 smc_buf_free(lgr, true, rmb_desc);
1031         } else {
1032                 rmb_desc->used = 0;
1033         }
1034 }
1035
1036 static void smc_buf_unuse(struct smc_connection *conn,
1037                           struct smc_link_group *lgr)
1038 {
1039         if (conn->sndbuf_desc)
1040                 conn->sndbuf_desc->used = 0;
1041         if (conn->rmb_desc && lgr->is_smcd)
1042                 conn->rmb_desc->used = 0;
1043         else if (conn->rmb_desc)
1044                 smcr_buf_unuse(conn->rmb_desc, lgr);
1045 }
1046
1047 /* remove a finished connection from its link group */
1048 void smc_conn_free(struct smc_connection *conn)
1049 {
1050         struct smc_link_group *lgr = conn->lgr;
1051
1052         if (!lgr)
1053                 return;
1054         if (lgr->is_smcd) {
1055                 if (!list_empty(&lgr->list))
1056                         smc_ism_unset_conn(conn);
1057                 tasklet_kill(&conn->rx_tsklet);
1058         } else {
1059                 smc_cdc_wait_pend_tx_wr(conn);
1060                 if (current_work() != &conn->abort_work)
1061                         cancel_work_sync(&conn->abort_work);
1062         }
1063         if (!list_empty(&lgr->list)) {
1064                 smc_buf_unuse(conn, lgr); /* allow buffer reuse */
1065                 smc_lgr_unregister_conn(conn);
1066         }
1067
1068         if (!lgr->conns_num)
1069                 smc_lgr_schedule_free_work(lgr);
1070 }
1071
1072 /* unregister a link from a buf_desc */
1073 static void smcr_buf_unmap_link(struct smc_buf_desc *buf_desc, bool is_rmb,
1074                                 struct smc_link *lnk)
1075 {
1076         if (is_rmb)
1077                 buf_desc->is_reg_mr[lnk->link_idx] = false;
1078         if (!buf_desc->is_map_ib[lnk->link_idx])
1079                 return;
1080         if (is_rmb) {
1081                 if (buf_desc->mr_rx[lnk->link_idx]) {
1082                         smc_ib_put_memory_region(
1083                                         buf_desc->mr_rx[lnk->link_idx]);
1084                         buf_desc->mr_rx[lnk->link_idx] = NULL;
1085                 }
1086                 smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_FROM_DEVICE);
1087         } else {
1088                 smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_TO_DEVICE);
1089         }
1090         sg_free_table(&buf_desc->sgt[lnk->link_idx]);
1091         buf_desc->is_map_ib[lnk->link_idx] = false;
1092 }
1093
1094 /* unmap all buffers of lgr for a deleted link */
1095 static void smcr_buf_unmap_lgr(struct smc_link *lnk)
1096 {
1097         struct smc_link_group *lgr = lnk->lgr;
1098         struct smc_buf_desc *buf_desc, *bf;
1099         int i;
1100
1101         for (i = 0; i < SMC_RMBE_SIZES; i++) {
1102                 mutex_lock(&lgr->rmbs_lock);
1103                 list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list)
1104                         smcr_buf_unmap_link(buf_desc, true, lnk);
1105                 mutex_unlock(&lgr->rmbs_lock);
1106                 mutex_lock(&lgr->sndbufs_lock);
1107                 list_for_each_entry_safe(buf_desc, bf, &lgr->sndbufs[i],
1108                                          list)
1109                         smcr_buf_unmap_link(buf_desc, false, lnk);
1110                 mutex_unlock(&lgr->sndbufs_lock);
1111         }
1112 }
1113
1114 static void smcr_rtoken_clear_link(struct smc_link *lnk)
1115 {
1116         struct smc_link_group *lgr = lnk->lgr;
1117         int i;
1118
1119         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1120                 lgr->rtokens[i][lnk->link_idx].rkey = 0;
1121                 lgr->rtokens[i][lnk->link_idx].dma_addr = 0;
1122         }
1123 }
1124
1125 /* must be called under lgr->llc_conf_mutex lock */
1126 void smcr_link_clear(struct smc_link *lnk, bool log)
1127 {
1128         struct smc_ib_device *smcibdev;
1129
1130         if (!lnk->lgr || lnk->state == SMC_LNK_UNUSED)
1131                 return;
1132         lnk->peer_qpn = 0;
1133         smc_llc_link_clear(lnk, log);
1134         smcr_buf_unmap_lgr(lnk);
1135         smcr_rtoken_clear_link(lnk);
1136         smc_ib_modify_qp_error(lnk);
1137         smc_wr_free_link(lnk);
1138         smc_ib_destroy_queue_pair(lnk);
1139         smc_ib_dealloc_protection_domain(lnk);
1140         smc_wr_free_link_mem(lnk);
1141         smc_ibdev_cnt_dec(lnk);
1142         put_device(&lnk->smcibdev->ibdev->dev);
1143         smcibdev = lnk->smcibdev;
1144         memset(lnk, 0, sizeof(struct smc_link));
1145         lnk->state = SMC_LNK_UNUSED;
1146         if (!atomic_dec_return(&smcibdev->lnk_cnt))
1147                 wake_up(&smcibdev->lnks_deleted);
1148 }
1149
1150 static void smcr_buf_free(struct smc_link_group *lgr, bool is_rmb,
1151                           struct smc_buf_desc *buf_desc)
1152 {
1153         int i;
1154
1155         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
1156                 smcr_buf_unmap_link(buf_desc, is_rmb, &lgr->lnk[i]);
1157
1158         if (buf_desc->pages)
1159                 __free_pages(buf_desc->pages, buf_desc->order);
1160         kfree(buf_desc);
1161 }
1162
1163 static void smcd_buf_free(struct smc_link_group *lgr, bool is_dmb,
1164                           struct smc_buf_desc *buf_desc)
1165 {
1166         if (is_dmb) {
1167                 /* restore original buf len */
1168                 buf_desc->len += sizeof(struct smcd_cdc_msg);
1169                 smc_ism_unregister_dmb(lgr->smcd, buf_desc);
1170         } else {
1171                 kfree(buf_desc->cpu_addr);
1172         }
1173         kfree(buf_desc);
1174 }
1175
1176 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
1177                          struct smc_buf_desc *buf_desc)
1178 {
1179         if (lgr->is_smcd)
1180                 smcd_buf_free(lgr, is_rmb, buf_desc);
1181         else
1182                 smcr_buf_free(lgr, is_rmb, buf_desc);
1183 }
1184
1185 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
1186 {
1187         struct smc_buf_desc *buf_desc, *bf_desc;
1188         struct list_head *buf_list;
1189         int i;
1190
1191         for (i = 0; i < SMC_RMBE_SIZES; i++) {
1192                 if (is_rmb)
1193                         buf_list = &lgr->rmbs[i];
1194                 else
1195                         buf_list = &lgr->sndbufs[i];
1196                 list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
1197                                          list) {
1198                         list_del(&buf_desc->list);
1199                         smc_buf_free(lgr, is_rmb, buf_desc);
1200                 }
1201         }
1202 }
1203
1204 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
1205 {
1206         /* free send buffers */
1207         __smc_lgr_free_bufs(lgr, false);
1208         /* free rmbs */
1209         __smc_lgr_free_bufs(lgr, true);
1210 }
1211
1212 /* remove a link group */
1213 static void smc_lgr_free(struct smc_link_group *lgr)
1214 {
1215         int i;
1216
1217         if (!lgr->is_smcd) {
1218                 mutex_lock(&lgr->llc_conf_mutex);
1219                 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1220                         if (lgr->lnk[i].state != SMC_LNK_UNUSED)
1221                                 smcr_link_clear(&lgr->lnk[i], false);
1222                 }
1223                 mutex_unlock(&lgr->llc_conf_mutex);
1224                 smc_llc_lgr_clear(lgr);
1225         }
1226
1227         smc_lgr_free_bufs(lgr);
1228         destroy_workqueue(lgr->tx_wq);
1229         if (lgr->is_smcd) {
1230                 smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
1231                 put_device(&lgr->smcd->dev);
1232                 if (!atomic_dec_return(&lgr->smcd->lgr_cnt))
1233                         wake_up(&lgr->smcd->lgrs_deleted);
1234         } else {
1235                 if (!atomic_dec_return(&lgr_cnt))
1236                         wake_up(&lgrs_deleted);
1237         }
1238         kfree(lgr);
1239 }
1240
1241 static void smc_sk_wake_ups(struct smc_sock *smc)
1242 {
1243         smc->sk.sk_write_space(&smc->sk);
1244         smc->sk.sk_data_ready(&smc->sk);
1245         smc->sk.sk_state_change(&smc->sk);
1246 }
1247
1248 /* kill a connection */
1249 static void smc_conn_kill(struct smc_connection *conn, bool soft)
1250 {
1251         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1252
1253         if (conn->lgr->is_smcd && conn->lgr->peer_shutdown)
1254                 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
1255         else
1256                 smc_close_abort(conn);
1257         conn->killed = 1;
1258         smc->sk.sk_err = ECONNABORTED;
1259         smc_sk_wake_ups(smc);
1260         if (conn->lgr->is_smcd) {
1261                 smc_ism_unset_conn(conn);
1262                 if (soft)
1263                         tasklet_kill(&conn->rx_tsklet);
1264                 else
1265                         tasklet_unlock_wait(&conn->rx_tsklet);
1266         } else {
1267                 smc_cdc_wait_pend_tx_wr(conn);
1268         }
1269         smc_lgr_unregister_conn(conn);
1270         smc_close_active_abort(smc);
1271 }
1272
1273 static void smc_lgr_cleanup(struct smc_link_group *lgr)
1274 {
1275         if (lgr->is_smcd) {
1276                 smc_ism_signal_shutdown(lgr);
1277         } else {
1278                 u32 rsn = lgr->llc_termination_rsn;
1279
1280                 if (!rsn)
1281                         rsn = SMC_LLC_DEL_PROG_INIT_TERM;
1282                 smc_llc_send_link_delete_all(lgr, false, rsn);
1283                 smcr_lgr_link_deactivate_all(lgr);
1284         }
1285 }
1286
1287 /* terminate link group
1288  * @soft: true if link group shutdown can take its time
1289  *        false if immediate link group shutdown is required
1290  */
1291 static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft)
1292 {
1293         struct smc_connection *conn;
1294         struct smc_sock *smc;
1295         struct rb_node *node;
1296
1297         if (lgr->terminating)
1298                 return; /* lgr already terminating */
1299         /* cancel free_work sync, will terminate when lgr->freeing is set */
1300         cancel_delayed_work_sync(&lgr->free_work);
1301         lgr->terminating = 1;
1302
1303         /* kill remaining link group connections */
1304         read_lock_bh(&lgr->conns_lock);
1305         node = rb_first(&lgr->conns_all);
1306         while (node) {
1307                 read_unlock_bh(&lgr->conns_lock);
1308                 conn = rb_entry(node, struct smc_connection, alert_node);
1309                 smc = container_of(conn, struct smc_sock, conn);
1310                 sock_hold(&smc->sk); /* sock_put below */
1311                 lock_sock(&smc->sk);
1312                 smc_conn_kill(conn, soft);
1313                 release_sock(&smc->sk);
1314                 sock_put(&smc->sk); /* sock_hold above */
1315                 read_lock_bh(&lgr->conns_lock);
1316                 node = rb_first(&lgr->conns_all);
1317         }
1318         read_unlock_bh(&lgr->conns_lock);
1319         smc_lgr_cleanup(lgr);
1320         smc_lgr_free(lgr);
1321 }
1322
1323 /* unlink link group and schedule termination */
1324 void smc_lgr_terminate_sched(struct smc_link_group *lgr)
1325 {
1326         spinlock_t *lgr_lock;
1327
1328         smc_lgr_list_head(lgr, &lgr_lock);
1329         spin_lock_bh(lgr_lock);
1330         if (list_empty(&lgr->list) || lgr->terminating || lgr->freeing) {
1331                 spin_unlock_bh(lgr_lock);
1332                 return; /* lgr already terminating */
1333         }
1334         list_del_init(&lgr->list);
1335         lgr->freeing = 1;
1336         spin_unlock_bh(lgr_lock);
1337         schedule_work(&lgr->terminate_work);
1338 }
1339
1340 /* Called when peer lgr shutdown (regularly or abnormally) is received */
1341 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid, unsigned short vlan)
1342 {
1343         struct smc_link_group *lgr, *l;
1344         LIST_HEAD(lgr_free_list);
1345
1346         /* run common cleanup function and build free list */
1347         spin_lock_bh(&dev->lgr_lock);
1348         list_for_each_entry_safe(lgr, l, &dev->lgr_list, list) {
1349                 if ((!peer_gid || lgr->peer_gid == peer_gid) &&
1350                     (vlan == VLAN_VID_MASK || lgr->vlan_id == vlan)) {
1351                         if (peer_gid) /* peer triggered termination */
1352                                 lgr->peer_shutdown = 1;
1353                         list_move(&lgr->list, &lgr_free_list);
1354                         lgr->freeing = 1;
1355                 }
1356         }
1357         spin_unlock_bh(&dev->lgr_lock);
1358
1359         /* cancel the regular free workers and actually free lgrs */
1360         list_for_each_entry_safe(lgr, l, &lgr_free_list, list) {
1361                 list_del_init(&lgr->list);
1362                 schedule_work(&lgr->terminate_work);
1363         }
1364 }
1365
1366 /* Called when an SMCD device is removed or the smc module is unloaded */
1367 void smc_smcd_terminate_all(struct smcd_dev *smcd)
1368 {
1369         struct smc_link_group *lgr, *lg;
1370         LIST_HEAD(lgr_free_list);
1371
1372         spin_lock_bh(&smcd->lgr_lock);
1373         list_splice_init(&smcd->lgr_list, &lgr_free_list);
1374         list_for_each_entry(lgr, &lgr_free_list, list)
1375                 lgr->freeing = 1;
1376         spin_unlock_bh(&smcd->lgr_lock);
1377
1378         list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
1379                 list_del_init(&lgr->list);
1380                 __smc_lgr_terminate(lgr, false);
1381         }
1382
1383         if (atomic_read(&smcd->lgr_cnt))
1384                 wait_event(smcd->lgrs_deleted, !atomic_read(&smcd->lgr_cnt));
1385 }
1386
1387 /* Called when an SMCR device is removed or the smc module is unloaded.
1388  * If smcibdev is given, all SMCR link groups using this device are terminated.
1389  * If smcibdev is NULL, all SMCR link groups are terminated.
1390  */
1391 void smc_smcr_terminate_all(struct smc_ib_device *smcibdev)
1392 {
1393         struct smc_link_group *lgr, *lg;
1394         LIST_HEAD(lgr_free_list);
1395         int i;
1396
1397         spin_lock_bh(&smc_lgr_list.lock);
1398         if (!smcibdev) {
1399                 list_splice_init(&smc_lgr_list.list, &lgr_free_list);
1400                 list_for_each_entry(lgr, &lgr_free_list, list)
1401                         lgr->freeing = 1;
1402         } else {
1403                 list_for_each_entry_safe(lgr, lg, &smc_lgr_list.list, list) {
1404                         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1405                                 if (lgr->lnk[i].smcibdev == smcibdev)
1406                                         smcr_link_down_cond_sched(&lgr->lnk[i]);
1407                         }
1408                 }
1409         }
1410         spin_unlock_bh(&smc_lgr_list.lock);
1411
1412         list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
1413                 list_del_init(&lgr->list);
1414                 smc_llc_set_termination_rsn(lgr, SMC_LLC_DEL_OP_INIT_TERM);
1415                 __smc_lgr_terminate(lgr, false);
1416         }
1417
1418         if (smcibdev) {
1419                 if (atomic_read(&smcibdev->lnk_cnt))
1420                         wait_event(smcibdev->lnks_deleted,
1421                                    !atomic_read(&smcibdev->lnk_cnt));
1422         } else {
1423                 if (atomic_read(&lgr_cnt))
1424                         wait_event(lgrs_deleted, !atomic_read(&lgr_cnt));
1425         }
1426 }
1427
1428 /* set new lgr type and clear all asymmetric link tagging */
1429 void smcr_lgr_set_type(struct smc_link_group *lgr, enum smc_lgr_type new_type)
1430 {
1431         char *lgr_type = "";
1432         int i;
1433
1434         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
1435                 if (smc_link_usable(&lgr->lnk[i]))
1436                         lgr->lnk[i].link_is_asym = false;
1437         if (lgr->type == new_type)
1438                 return;
1439         lgr->type = new_type;
1440
1441         switch (lgr->type) {
1442         case SMC_LGR_NONE:
1443                 lgr_type = "NONE";
1444                 break;
1445         case SMC_LGR_SINGLE:
1446                 lgr_type = "SINGLE";
1447                 break;
1448         case SMC_LGR_SYMMETRIC:
1449                 lgr_type = "SYMMETRIC";
1450                 break;
1451         case SMC_LGR_ASYMMETRIC_PEER:
1452                 lgr_type = "ASYMMETRIC_PEER";
1453                 break;
1454         case SMC_LGR_ASYMMETRIC_LOCAL:
1455                 lgr_type = "ASYMMETRIC_LOCAL";
1456                 break;
1457         }
1458         pr_warn_ratelimited("smc: SMC-R lg %*phN state changed: "
1459                             "%s, pnetid %.16s\n", SMC_LGR_ID_SIZE, &lgr->id,
1460                             lgr_type, lgr->pnet_id);
1461 }
1462
1463 /* set new lgr type and tag a link as asymmetric */
1464 void smcr_lgr_set_type_asym(struct smc_link_group *lgr,
1465                             enum smc_lgr_type new_type, int asym_lnk_idx)
1466 {
1467         smcr_lgr_set_type(lgr, new_type);
1468         lgr->lnk[asym_lnk_idx].link_is_asym = true;
1469 }
1470
1471 /* abort connection, abort_work scheduled from tasklet context */
1472 static void smc_conn_abort_work(struct work_struct *work)
1473 {
1474         struct smc_connection *conn = container_of(work,
1475                                                    struct smc_connection,
1476                                                    abort_work);
1477         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1478
1479         lock_sock(&smc->sk);
1480         smc_conn_kill(conn, true);
1481         release_sock(&smc->sk);
1482         sock_put(&smc->sk); /* sock_hold done by schedulers of abort_work */
1483 }
1484
1485 void smcr_port_add(struct smc_ib_device *smcibdev, u8 ibport)
1486 {
1487         struct smc_link_group *lgr, *n;
1488
1489         list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1490                 struct smc_link *link;
1491
1492                 if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1493                             SMC_MAX_PNETID_LEN) ||
1494                     lgr->type == SMC_LGR_SYMMETRIC ||
1495                     lgr->type == SMC_LGR_ASYMMETRIC_PEER)
1496                         continue;
1497
1498                 /* trigger local add link processing */
1499                 link = smc_llc_usable_link(lgr);
1500                 if (link)
1501                         smc_llc_add_link_local(link);
1502         }
1503 }
1504
1505 /* link is down - switch connections to alternate link,
1506  * must be called under lgr->llc_conf_mutex lock
1507  */
1508 static void smcr_link_down(struct smc_link *lnk)
1509 {
1510         struct smc_link_group *lgr = lnk->lgr;
1511         struct smc_link *to_lnk;
1512         int del_link_id;
1513
1514         if (!lgr || lnk->state == SMC_LNK_UNUSED || list_empty(&lgr->list))
1515                 return;
1516
1517         to_lnk = smc_switch_conns(lgr, lnk, true);
1518         if (!to_lnk) { /* no backup link available */
1519                 smcr_link_clear(lnk, true);
1520                 return;
1521         }
1522         smcr_lgr_set_type(lgr, SMC_LGR_SINGLE);
1523         del_link_id = lnk->link_id;
1524
1525         if (lgr->role == SMC_SERV) {
1526                 /* trigger local delete link processing */
1527                 smc_llc_srv_delete_link_local(to_lnk, del_link_id);
1528         } else {
1529                 if (lgr->llc_flow_lcl.type != SMC_LLC_FLOW_NONE) {
1530                         /* another llc task is ongoing */
1531                         mutex_unlock(&lgr->llc_conf_mutex);
1532                         wait_event_timeout(lgr->llc_flow_waiter,
1533                                 (list_empty(&lgr->list) ||
1534                                  lgr->llc_flow_lcl.type == SMC_LLC_FLOW_NONE),
1535                                 SMC_LLC_WAIT_TIME);
1536                         mutex_lock(&lgr->llc_conf_mutex);
1537                 }
1538                 if (!list_empty(&lgr->list)) {
1539                         smc_llc_send_delete_link(to_lnk, del_link_id,
1540                                                  SMC_LLC_REQ, true,
1541                                                  SMC_LLC_DEL_LOST_PATH);
1542                         smcr_link_clear(lnk, true);
1543                 }
1544                 wake_up(&lgr->llc_flow_waiter); /* wake up next waiter */
1545         }
1546 }
1547
1548 /* must be called under lgr->llc_conf_mutex lock */
1549 void smcr_link_down_cond(struct smc_link *lnk)
1550 {
1551         if (smc_link_downing(&lnk->state))
1552                 smcr_link_down(lnk);
1553 }
1554
1555 /* will get the lgr->llc_conf_mutex lock */
1556 void smcr_link_down_cond_sched(struct smc_link *lnk)
1557 {
1558         if (smc_link_downing(&lnk->state))
1559                 schedule_work(&lnk->link_down_wrk);
1560 }
1561
1562 void smcr_port_err(struct smc_ib_device *smcibdev, u8 ibport)
1563 {
1564         struct smc_link_group *lgr, *n;
1565         int i;
1566
1567         list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1568                 if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1569                             SMC_MAX_PNETID_LEN))
1570                         continue; /* lgr is not affected */
1571                 if (list_empty(&lgr->list))
1572                         continue;
1573                 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1574                         struct smc_link *lnk = &lgr->lnk[i];
1575
1576                         if (smc_link_usable(lnk) &&
1577                             lnk->smcibdev == smcibdev && lnk->ibport == ibport)
1578                                 smcr_link_down_cond_sched(lnk);
1579                 }
1580         }
1581 }
1582
1583 static void smc_link_down_work(struct work_struct *work)
1584 {
1585         struct smc_link *link = container_of(work, struct smc_link,
1586                                              link_down_wrk);
1587         struct smc_link_group *lgr = link->lgr;
1588
1589         if (list_empty(&lgr->list))
1590                 return;
1591         wake_up_all(&lgr->llc_msg_waiter);
1592         mutex_lock(&lgr->llc_conf_mutex);
1593         smcr_link_down(link);
1594         mutex_unlock(&lgr->llc_conf_mutex);
1595 }
1596
1597 static int smc_vlan_by_tcpsk_walk(struct net_device *lower_dev,
1598                                   struct netdev_nested_priv *priv)
1599 {
1600         unsigned short *vlan_id = (unsigned short *)priv->data;
1601
1602         if (is_vlan_dev(lower_dev)) {
1603                 *vlan_id = vlan_dev_vlan_id(lower_dev);
1604                 return 1;
1605         }
1606
1607         return 0;
1608 }
1609
1610 /* Determine vlan of internal TCP socket. */
1611 int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini)
1612 {
1613         struct dst_entry *dst = sk_dst_get(clcsock->sk);
1614         struct netdev_nested_priv priv;
1615         struct net_device *ndev;
1616         int rc = 0;
1617
1618         ini->vlan_id = 0;
1619         if (!dst) {
1620                 rc = -ENOTCONN;
1621                 goto out;
1622         }
1623         if (!dst->dev) {
1624                 rc = -ENODEV;
1625                 goto out_rel;
1626         }
1627
1628         ndev = dst->dev;
1629         if (is_vlan_dev(ndev)) {
1630                 ini->vlan_id = vlan_dev_vlan_id(ndev);
1631                 goto out_rel;
1632         }
1633
1634         priv.data = (void *)&ini->vlan_id;
1635         rtnl_lock();
1636         netdev_walk_all_lower_dev(ndev, smc_vlan_by_tcpsk_walk, &priv);
1637         rtnl_unlock();
1638
1639 out_rel:
1640         dst_release(dst);
1641 out:
1642         return rc;
1643 }
1644
1645 static bool smcr_lgr_match(struct smc_link_group *lgr,
1646                            struct smc_clc_msg_local *lcl,
1647                            enum smc_lgr_role role, u32 clcqpn)
1648 {
1649         int i;
1650
1651         if (memcmp(lgr->peer_systemid, lcl->id_for_peer, SMC_SYSTEMID_LEN) ||
1652             lgr->role != role)
1653                 return false;
1654
1655         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1656                 if (!smc_link_active(&lgr->lnk[i]))
1657                         continue;
1658                 if ((lgr->role == SMC_SERV || lgr->lnk[i].peer_qpn == clcqpn) &&
1659                     !memcmp(lgr->lnk[i].peer_gid, &lcl->gid, SMC_GID_SIZE) &&
1660                     !memcmp(lgr->lnk[i].peer_mac, lcl->mac, sizeof(lcl->mac)))
1661                         return true;
1662         }
1663         return false;
1664 }
1665
1666 static bool smcd_lgr_match(struct smc_link_group *lgr,
1667                            struct smcd_dev *smcismdev, u64 peer_gid)
1668 {
1669         return lgr->peer_gid == peer_gid && lgr->smcd == smcismdev;
1670 }
1671
1672 /* create a new SMC connection (and a new link group if necessary) */
1673 int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini)
1674 {
1675         struct smc_connection *conn = &smc->conn;
1676         struct list_head *lgr_list;
1677         struct smc_link_group *lgr;
1678         enum smc_lgr_role role;
1679         spinlock_t *lgr_lock;
1680         int rc = 0;
1681
1682         lgr_list = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_list :
1683                                   &smc_lgr_list.list;
1684         lgr_lock = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_lock :
1685                                   &smc_lgr_list.lock;
1686         ini->first_contact_local = 1;
1687         role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
1688         if (role == SMC_CLNT && ini->first_contact_peer)
1689                 /* create new link group as well */
1690                 goto create;
1691
1692         /* determine if an existing link group can be reused */
1693         spin_lock_bh(lgr_lock);
1694         list_for_each_entry(lgr, lgr_list, list) {
1695                 write_lock_bh(&lgr->conns_lock);
1696                 if ((ini->is_smcd ?
1697                      smcd_lgr_match(lgr, ini->ism_dev[ini->ism_selected],
1698                                     ini->ism_peer_gid[ini->ism_selected]) :
1699                      smcr_lgr_match(lgr, ini->ib_lcl, role, ini->ib_clcqpn)) &&
1700                     !lgr->sync_err &&
1701                     (ini->smcd_version == SMC_V2 ||
1702                      lgr->vlan_id == ini->vlan_id) &&
1703                     (role == SMC_CLNT || ini->is_smcd ||
1704                     (lgr->conns_num < SMC_RMBS_PER_LGR_MAX &&
1705                       !bitmap_full(lgr->rtokens_used_mask, 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                         smc_lgr_cleanup_early(lgr);
1741                         goto out;
1742                 }
1743         }
1744         conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
1745         conn->local_tx_ctrl.len = SMC_WR_TX_SIZE;
1746         conn->urg_state = SMC_URG_READ;
1747         init_waitqueue_head(&conn->cdc_pend_tx_wq);
1748         INIT_WORK(&smc->conn.abort_work, smc_conn_abort_work);
1749         if (ini->is_smcd) {
1750                 conn->rx_off = sizeof(struct smcd_cdc_msg);
1751                 smcd_cdc_rx_init(conn); /* init tasklet for this conn */
1752         } else {
1753                 conn->rx_off = 0;
1754         }
1755 #ifndef KERNEL_HAS_ATOMIC64
1756         spin_lock_init(&conn->acurs_lock);
1757 #endif
1758
1759 out:
1760         return rc;
1761 }
1762
1763 #define SMCD_DMBE_SIZES         6 /* 0 -> 16KB, 1 -> 32KB, .. 6 -> 1MB */
1764 #define SMCR_RMBE_SIZES         5 /* 0 -> 16KB, 1 -> 32KB, .. 5 -> 512KB */
1765
1766 /* convert the RMB size into the compressed notation (minimum 16K, see
1767  * SMCD/R_DMBE_SIZES.
1768  * In contrast to plain ilog2, this rounds towards the next power of 2,
1769  * so the socket application gets at least its desired sndbuf / rcvbuf size.
1770  */
1771 static u8 smc_compress_bufsize(int size, bool is_smcd, bool is_rmb)
1772 {
1773         const unsigned int max_scat = SG_MAX_SINGLE_ALLOC * PAGE_SIZE;
1774         u8 compressed;
1775
1776         if (size <= SMC_BUF_MIN_SIZE)
1777                 return 0;
1778
1779         size = (size - 1) >> 14;  /* convert to 16K multiple */
1780         compressed = min_t(u8, ilog2(size) + 1,
1781                            is_smcd ? SMCD_DMBE_SIZES : SMCR_RMBE_SIZES);
1782
1783         if (!is_smcd && is_rmb)
1784                 /* RMBs are backed by & limited to max size of scatterlists */
1785                 compressed = min_t(u8, compressed, ilog2(max_scat >> 14));
1786
1787         return compressed;
1788 }
1789
1790 /* convert the RMB size from compressed notation into integer */
1791 int smc_uncompress_bufsize(u8 compressed)
1792 {
1793         u32 size;
1794
1795         size = 0x00000001 << (((int)compressed) + 14);
1796         return (int)size;
1797 }
1798
1799 /* try to reuse a sndbuf or rmb description slot for a certain
1800  * buffer size; if not available, return NULL
1801  */
1802 static struct smc_buf_desc *smc_buf_get_slot(int compressed_bufsize,
1803                                              struct mutex *lock,
1804                                              struct list_head *buf_list)
1805 {
1806         struct smc_buf_desc *buf_slot;
1807
1808         mutex_lock(lock);
1809         list_for_each_entry(buf_slot, buf_list, list) {
1810                 if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
1811                         mutex_unlock(lock);
1812                         return buf_slot;
1813                 }
1814         }
1815         mutex_unlock(lock);
1816         return NULL;
1817 }
1818
1819 /* one of the conditions for announcing a receiver's current window size is
1820  * that it "results in a minimum increase in the window size of 10% of the
1821  * receive buffer space" [RFC7609]
1822  */
1823 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
1824 {
1825         return max_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
1826 }
1827
1828 /* map an rmb buf to a link */
1829 static int smcr_buf_map_link(struct smc_buf_desc *buf_desc, bool is_rmb,
1830                              struct smc_link *lnk)
1831 {
1832         int rc;
1833
1834         if (buf_desc->is_map_ib[lnk->link_idx])
1835                 return 0;
1836
1837         rc = sg_alloc_table(&buf_desc->sgt[lnk->link_idx], 1, GFP_KERNEL);
1838         if (rc)
1839                 return rc;
1840         sg_set_buf(buf_desc->sgt[lnk->link_idx].sgl,
1841                    buf_desc->cpu_addr, buf_desc->len);
1842
1843         /* map sg table to DMA address */
1844         rc = smc_ib_buf_map_sg(lnk, buf_desc,
1845                                is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
1846         /* SMC protocol depends on mapping to one DMA address only */
1847         if (rc != 1) {
1848                 rc = -EAGAIN;
1849                 goto free_table;
1850         }
1851
1852         /* create a new memory region for the RMB */
1853         if (is_rmb) {
1854                 rc = smc_ib_get_memory_region(lnk->roce_pd,
1855                                               IB_ACCESS_REMOTE_WRITE |
1856                                               IB_ACCESS_LOCAL_WRITE,
1857                                               buf_desc, lnk->link_idx);
1858                 if (rc)
1859                         goto buf_unmap;
1860                 smc_ib_sync_sg_for_device(lnk, buf_desc, DMA_FROM_DEVICE);
1861         }
1862         buf_desc->is_map_ib[lnk->link_idx] = true;
1863         return 0;
1864
1865 buf_unmap:
1866         smc_ib_buf_unmap_sg(lnk, buf_desc,
1867                             is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
1868 free_table:
1869         sg_free_table(&buf_desc->sgt[lnk->link_idx]);
1870         return rc;
1871 }
1872
1873 /* register a new rmb on IB device,
1874  * must be called under lgr->llc_conf_mutex lock
1875  */
1876 int smcr_link_reg_rmb(struct smc_link *link, struct smc_buf_desc *rmb_desc)
1877 {
1878         if (list_empty(&link->lgr->list))
1879                 return -ENOLINK;
1880         if (!rmb_desc->is_reg_mr[link->link_idx]) {
1881                 /* register memory region for new rmb */
1882                 if (smc_wr_reg_send(link, rmb_desc->mr_rx[link->link_idx])) {
1883                         rmb_desc->is_reg_err = true;
1884                         return -EFAULT;
1885                 }
1886                 rmb_desc->is_reg_mr[link->link_idx] = true;
1887         }
1888         return 0;
1889 }
1890
1891 static int _smcr_buf_map_lgr(struct smc_link *lnk, struct mutex *lock,
1892                              struct list_head *lst, bool is_rmb)
1893 {
1894         struct smc_buf_desc *buf_desc, *bf;
1895         int rc = 0;
1896
1897         mutex_lock(lock);
1898         list_for_each_entry_safe(buf_desc, bf, lst, list) {
1899                 if (!buf_desc->used)
1900                         continue;
1901                 rc = smcr_buf_map_link(buf_desc, is_rmb, lnk);
1902                 if (rc)
1903                         goto out;
1904         }
1905 out:
1906         mutex_unlock(lock);
1907         return rc;
1908 }
1909
1910 /* map all used buffers of lgr for a new link */
1911 int smcr_buf_map_lgr(struct smc_link *lnk)
1912 {
1913         struct smc_link_group *lgr = lnk->lgr;
1914         int i, rc = 0;
1915
1916         for (i = 0; i < SMC_RMBE_SIZES; i++) {
1917                 rc = _smcr_buf_map_lgr(lnk, &lgr->rmbs_lock,
1918                                        &lgr->rmbs[i], true);
1919                 if (rc)
1920                         return rc;
1921                 rc = _smcr_buf_map_lgr(lnk, &lgr->sndbufs_lock,
1922                                        &lgr->sndbufs[i], false);
1923                 if (rc)
1924                         return rc;
1925         }
1926         return 0;
1927 }
1928
1929 /* register all used buffers of lgr for a new link,
1930  * must be called under lgr->llc_conf_mutex lock
1931  */
1932 int smcr_buf_reg_lgr(struct smc_link *lnk)
1933 {
1934         struct smc_link_group *lgr = lnk->lgr;
1935         struct smc_buf_desc *buf_desc, *bf;
1936         int i, rc = 0;
1937
1938         mutex_lock(&lgr->rmbs_lock);
1939         for (i = 0; i < SMC_RMBE_SIZES; i++) {
1940                 list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list) {
1941                         if (!buf_desc->used)
1942                                 continue;
1943                         rc = smcr_link_reg_rmb(lnk, buf_desc);
1944                         if (rc)
1945                                 goto out;
1946                 }
1947         }
1948 out:
1949         mutex_unlock(&lgr->rmbs_lock);
1950         return rc;
1951 }
1952
1953 static struct smc_buf_desc *smcr_new_buf_create(struct smc_link_group *lgr,
1954                                                 bool is_rmb, int bufsize)
1955 {
1956         struct smc_buf_desc *buf_desc;
1957
1958         /* try to alloc a new buffer */
1959         buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
1960         if (!buf_desc)
1961                 return ERR_PTR(-ENOMEM);
1962
1963         buf_desc->order = get_order(bufsize);
1964         buf_desc->pages = alloc_pages(GFP_KERNEL | __GFP_NOWARN |
1965                                       __GFP_NOMEMALLOC | __GFP_COMP |
1966                                       __GFP_NORETRY | __GFP_ZERO,
1967                                       buf_desc->order);
1968         if (!buf_desc->pages) {
1969                 kfree(buf_desc);
1970                 return ERR_PTR(-EAGAIN);
1971         }
1972         buf_desc->cpu_addr = (void *)page_address(buf_desc->pages);
1973         buf_desc->len = bufsize;
1974         return buf_desc;
1975 }
1976
1977 /* map buf_desc on all usable links,
1978  * unused buffers stay mapped as long as the link is up
1979  */
1980 static int smcr_buf_map_usable_links(struct smc_link_group *lgr,
1981                                      struct smc_buf_desc *buf_desc, bool is_rmb)
1982 {
1983         int i, rc = 0;
1984
1985         /* protect against parallel link reconfiguration */
1986         mutex_lock(&lgr->llc_conf_mutex);
1987         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1988                 struct smc_link *lnk = &lgr->lnk[i];
1989
1990                 if (!smc_link_usable(lnk))
1991                         continue;
1992                 if (smcr_buf_map_link(buf_desc, is_rmb, lnk)) {
1993                         rc = -ENOMEM;
1994                         goto out;
1995                 }
1996         }
1997 out:
1998         mutex_unlock(&lgr->llc_conf_mutex);
1999         return rc;
2000 }
2001
2002 static struct smc_buf_desc *smcd_new_buf_create(struct smc_link_group *lgr,
2003                                                 bool is_dmb, int bufsize)
2004 {
2005         struct smc_buf_desc *buf_desc;
2006         int rc;
2007
2008         /* try to alloc a new DMB */
2009         buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
2010         if (!buf_desc)
2011                 return ERR_PTR(-ENOMEM);
2012         if (is_dmb) {
2013                 rc = smc_ism_register_dmb(lgr, bufsize, buf_desc);
2014                 if (rc) {
2015                         kfree(buf_desc);
2016                         if (rc == -ENOMEM)
2017                                 return ERR_PTR(-EAGAIN);
2018                         if (rc == -ENOSPC)
2019                                 return ERR_PTR(-ENOSPC);
2020                         return ERR_PTR(-EIO);
2021                 }
2022                 buf_desc->pages = virt_to_page(buf_desc->cpu_addr);
2023                 /* CDC header stored in buf. So, pretend it was smaller */
2024                 buf_desc->len = bufsize - sizeof(struct smcd_cdc_msg);
2025         } else {
2026                 buf_desc->cpu_addr = kzalloc(bufsize, GFP_KERNEL |
2027                                              __GFP_NOWARN | __GFP_NORETRY |
2028                                              __GFP_NOMEMALLOC);
2029                 if (!buf_desc->cpu_addr) {
2030                         kfree(buf_desc);
2031                         return ERR_PTR(-EAGAIN);
2032                 }
2033                 buf_desc->len = bufsize;
2034         }
2035         return buf_desc;
2036 }
2037
2038 static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
2039 {
2040         struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
2041         struct smc_connection *conn = &smc->conn;
2042         struct smc_link_group *lgr = conn->lgr;
2043         struct list_head *buf_list;
2044         int bufsize, bufsize_short;
2045         bool is_dgraded = false;
2046         struct mutex *lock;     /* lock buffer list */
2047         int sk_buf_size;
2048
2049         if (is_rmb)
2050                 /* use socket recv buffer size (w/o overhead) as start value */
2051                 sk_buf_size = smc->sk.sk_rcvbuf / 2;
2052         else
2053                 /* use socket send buffer size (w/o overhead) as start value */
2054                 sk_buf_size = smc->sk.sk_sndbuf / 2;
2055
2056         for (bufsize_short = smc_compress_bufsize(sk_buf_size, is_smcd, is_rmb);
2057              bufsize_short >= 0; bufsize_short--) {
2058                 if (is_rmb) {
2059                         lock = &lgr->rmbs_lock;
2060                         buf_list = &lgr->rmbs[bufsize_short];
2061                 } else {
2062                         lock = &lgr->sndbufs_lock;
2063                         buf_list = &lgr->sndbufs[bufsize_short];
2064                 }
2065                 bufsize = smc_uncompress_bufsize(bufsize_short);
2066
2067                 /* check for reusable slot in the link group */
2068                 buf_desc = smc_buf_get_slot(bufsize_short, lock, buf_list);
2069                 if (buf_desc) {
2070                         SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, bufsize);
2071                         SMC_STAT_BUF_REUSE(smc, is_smcd, is_rmb);
2072                         memset(buf_desc->cpu_addr, 0, bufsize);
2073                         break; /* found reusable slot */
2074                 }
2075
2076                 if (is_smcd)
2077                         buf_desc = smcd_new_buf_create(lgr, is_rmb, bufsize);
2078                 else
2079                         buf_desc = smcr_new_buf_create(lgr, is_rmb, bufsize);
2080
2081                 if (PTR_ERR(buf_desc) == -ENOMEM)
2082                         break;
2083                 if (IS_ERR(buf_desc)) {
2084                         if (!is_dgraded) {
2085                                 is_dgraded = true;
2086                                 SMC_STAT_RMB_DOWNGRADED(smc, is_smcd, is_rmb);
2087                         }
2088                         continue;
2089                 }
2090
2091                 SMC_STAT_RMB_ALLOC(smc, is_smcd, is_rmb);
2092                 SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, bufsize);
2093                 buf_desc->used = 1;
2094                 mutex_lock(lock);
2095                 list_add(&buf_desc->list, buf_list);
2096                 mutex_unlock(lock);
2097                 break; /* found */
2098         }
2099
2100         if (IS_ERR(buf_desc))
2101                 return PTR_ERR(buf_desc);
2102
2103         if (!is_smcd) {
2104                 if (smcr_buf_map_usable_links(lgr, buf_desc, is_rmb)) {
2105                         smcr_buf_unuse(buf_desc, lgr);
2106                         return -ENOMEM;
2107                 }
2108         }
2109
2110         if (is_rmb) {
2111                 conn->rmb_desc = buf_desc;
2112                 conn->rmbe_size_short = bufsize_short;
2113                 smc->sk.sk_rcvbuf = bufsize * 2;
2114                 atomic_set(&conn->bytes_to_rcv, 0);
2115                 conn->rmbe_update_limit =
2116                         smc_rmb_wnd_update_limit(buf_desc->len);
2117                 if (is_smcd)
2118                         smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */
2119         } else {
2120                 conn->sndbuf_desc = buf_desc;
2121                 smc->sk.sk_sndbuf = bufsize * 2;
2122                 atomic_set(&conn->sndbuf_space, bufsize);
2123         }
2124         return 0;
2125 }
2126
2127 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn)
2128 {
2129         if (!conn->lgr || conn->lgr->is_smcd || !smc_link_active(conn->lnk))
2130                 return;
2131         smc_ib_sync_sg_for_cpu(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE);
2132 }
2133
2134 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
2135 {
2136         if (!conn->lgr || conn->lgr->is_smcd || !smc_link_active(conn->lnk))
2137                 return;
2138         smc_ib_sync_sg_for_device(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE);
2139 }
2140
2141 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
2142 {
2143         int i;
2144
2145         if (!conn->lgr || conn->lgr->is_smcd)
2146                 return;
2147         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
2148                 if (!smc_link_active(&conn->lgr->lnk[i]))
2149                         continue;
2150                 smc_ib_sync_sg_for_cpu(&conn->lgr->lnk[i], conn->rmb_desc,
2151                                        DMA_FROM_DEVICE);
2152         }
2153 }
2154
2155 void smc_rmb_sync_sg_for_device(struct smc_connection *conn)
2156 {
2157         int i;
2158
2159         if (!conn->lgr || conn->lgr->is_smcd)
2160                 return;
2161         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
2162                 if (!smc_link_active(&conn->lgr->lnk[i]))
2163                         continue;
2164                 smc_ib_sync_sg_for_device(&conn->lgr->lnk[i], conn->rmb_desc,
2165                                           DMA_FROM_DEVICE);
2166         }
2167 }
2168
2169 /* create the send and receive buffer for an SMC socket;
2170  * receive buffers are called RMBs;
2171  * (even though the SMC protocol allows more than one RMB-element per RMB,
2172  * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
2173  * extra RMB for every connection in a link group
2174  */
2175 int smc_buf_create(struct smc_sock *smc, bool is_smcd)
2176 {
2177         int rc;
2178
2179         /* create send buffer */
2180         rc = __smc_buf_create(smc, is_smcd, false);
2181         if (rc)
2182                 return rc;
2183         /* create rmb */
2184         rc = __smc_buf_create(smc, is_smcd, true);
2185         if (rc) {
2186                 mutex_lock(&smc->conn.lgr->sndbufs_lock);
2187                 list_del(&smc->conn.sndbuf_desc->list);
2188                 mutex_unlock(&smc->conn.lgr->sndbufs_lock);
2189                 smc_buf_free(smc->conn.lgr, false, smc->conn.sndbuf_desc);
2190                 smc->conn.sndbuf_desc = NULL;
2191         }
2192         return rc;
2193 }
2194
2195 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
2196 {
2197         int i;
2198
2199         for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
2200                 if (!test_and_set_bit(i, lgr->rtokens_used_mask))
2201                         return i;
2202         }
2203         return -ENOSPC;
2204 }
2205
2206 static int smc_rtoken_find_by_link(struct smc_link_group *lgr, int lnk_idx,
2207                                    u32 rkey)
2208 {
2209         int i;
2210
2211         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2212                 if (test_bit(i, lgr->rtokens_used_mask) &&
2213                     lgr->rtokens[i][lnk_idx].rkey == rkey)
2214                         return i;
2215         }
2216         return -ENOENT;
2217 }
2218
2219 /* set rtoken for a new link to an existing rmb */
2220 void smc_rtoken_set(struct smc_link_group *lgr, int link_idx, int link_idx_new,
2221                     __be32 nw_rkey_known, __be64 nw_vaddr, __be32 nw_rkey)
2222 {
2223         int rtok_idx;
2224
2225         rtok_idx = smc_rtoken_find_by_link(lgr, link_idx, ntohl(nw_rkey_known));
2226         if (rtok_idx == -ENOENT)
2227                 return;
2228         lgr->rtokens[rtok_idx][link_idx_new].rkey = ntohl(nw_rkey);
2229         lgr->rtokens[rtok_idx][link_idx_new].dma_addr = be64_to_cpu(nw_vaddr);
2230 }
2231
2232 /* set rtoken for a new link whose link_id is given */
2233 void smc_rtoken_set2(struct smc_link_group *lgr, int rtok_idx, int link_id,
2234                      __be64 nw_vaddr, __be32 nw_rkey)
2235 {
2236         u64 dma_addr = be64_to_cpu(nw_vaddr);
2237         u32 rkey = ntohl(nw_rkey);
2238         bool found = false;
2239         int link_idx;
2240
2241         for (link_idx = 0; link_idx < SMC_LINKS_PER_LGR_MAX; link_idx++) {
2242                 if (lgr->lnk[link_idx].link_id == link_id) {
2243                         found = true;
2244                         break;
2245                 }
2246         }
2247         if (!found)
2248                 return;
2249         lgr->rtokens[rtok_idx][link_idx].rkey = rkey;
2250         lgr->rtokens[rtok_idx][link_idx].dma_addr = dma_addr;
2251 }
2252
2253 /* add a new rtoken from peer */
2254 int smc_rtoken_add(struct smc_link *lnk, __be64 nw_vaddr, __be32 nw_rkey)
2255 {
2256         struct smc_link_group *lgr = smc_get_lgr(lnk);
2257         u64 dma_addr = be64_to_cpu(nw_vaddr);
2258         u32 rkey = ntohl(nw_rkey);
2259         int i;
2260
2261         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2262                 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
2263                     lgr->rtokens[i][lnk->link_idx].dma_addr == dma_addr &&
2264                     test_bit(i, lgr->rtokens_used_mask)) {
2265                         /* already in list */
2266                         return i;
2267                 }
2268         }
2269         i = smc_rmb_reserve_rtoken_idx(lgr);
2270         if (i < 0)
2271                 return i;
2272         lgr->rtokens[i][lnk->link_idx].rkey = rkey;
2273         lgr->rtokens[i][lnk->link_idx].dma_addr = dma_addr;
2274         return i;
2275 }
2276
2277 /* delete an rtoken from all links */
2278 int smc_rtoken_delete(struct smc_link *lnk, __be32 nw_rkey)
2279 {
2280         struct smc_link_group *lgr = smc_get_lgr(lnk);
2281         u32 rkey = ntohl(nw_rkey);
2282         int i, j;
2283
2284         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
2285                 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
2286                     test_bit(i, lgr->rtokens_used_mask)) {
2287                         for (j = 0; j < SMC_LINKS_PER_LGR_MAX; j++) {
2288                                 lgr->rtokens[i][j].rkey = 0;
2289                                 lgr->rtokens[i][j].dma_addr = 0;
2290                         }
2291                         clear_bit(i, lgr->rtokens_used_mask);
2292                         return 0;
2293                 }
2294         }
2295         return -ENOENT;
2296 }
2297
2298 /* save rkey and dma_addr received from peer during clc handshake */
2299 int smc_rmb_rtoken_handling(struct smc_connection *conn,
2300                             struct smc_link *lnk,
2301                             struct smc_clc_msg_accept_confirm *clc)
2302 {
2303         conn->rtoken_idx = smc_rtoken_add(lnk, clc->r0.rmb_dma_addr,
2304                                           clc->r0.rmb_rkey);
2305         if (conn->rtoken_idx < 0)
2306                 return conn->rtoken_idx;
2307         return 0;
2308 }
2309
2310 static void smc_core_going_away(void)
2311 {
2312         struct smc_ib_device *smcibdev;
2313         struct smcd_dev *smcd;
2314
2315         mutex_lock(&smc_ib_devices.mutex);
2316         list_for_each_entry(smcibdev, &smc_ib_devices.list, list) {
2317                 int i;
2318
2319                 for (i = 0; i < SMC_MAX_PORTS; i++)
2320                         set_bit(i, smcibdev->ports_going_away);
2321         }
2322         mutex_unlock(&smc_ib_devices.mutex);
2323
2324         mutex_lock(&smcd_dev_list.mutex);
2325         list_for_each_entry(smcd, &smcd_dev_list.list, list) {
2326                 smcd->going_away = 1;
2327         }
2328         mutex_unlock(&smcd_dev_list.mutex);
2329 }
2330
2331 /* Clean up all SMC link groups */
2332 static void smc_lgrs_shutdown(void)
2333 {
2334         struct smcd_dev *smcd;
2335
2336         smc_core_going_away();
2337
2338         smc_smcr_terminate_all(NULL);
2339
2340         mutex_lock(&smcd_dev_list.mutex);
2341         list_for_each_entry(smcd, &smcd_dev_list.list, list)
2342                 smc_smcd_terminate_all(smcd);
2343         mutex_unlock(&smcd_dev_list.mutex);
2344 }
2345
2346 static int smc_core_reboot_event(struct notifier_block *this,
2347                                  unsigned long event, void *ptr)
2348 {
2349         smc_lgrs_shutdown();
2350         smc_ib_unregister_client();
2351         return 0;
2352 }
2353
2354 static struct notifier_block smc_reboot_notifier = {
2355         .notifier_call = smc_core_reboot_event,
2356 };
2357
2358 int __init smc_core_init(void)
2359 {
2360         return register_reboot_notifier(&smc_reboot_notifier);
2361 }
2362
2363 /* Called (from smc_exit) when module is removed */
2364 void smc_core_exit(void)
2365 {
2366         unregister_reboot_notifier(&smc_reboot_notifier);
2367         smc_lgrs_shutdown();
2368 }