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 <net/tcp.h>
  20 #include <net/sock.h>
  21 #include <rdma/ib_verbs.h>
  22 #include <rdma/ib_cache.h>
  23
  24 #include "smc.h"
  25 #include "smc_clc.h"
  26 #include "smc_core.h"
  27 #include "smc_ib.h"
  28 #include "smc_wr.h"
  29 #include "smc_llc.h"
  30 #include "smc_cdc.h"
  31 #include "smc_close.h"
  32 #include "smc_ism.h"
  33
  34 #define SMC_LGR_NUM_INCR                256
  35 #define SMC_LGR_FREE_DELAY_SERV         (600 * HZ)
  36 #define SMC_LGR_FREE_DELAY_CLNT         (SMC_LGR_FREE_DELAY_SERV + 10 * HZ)
  37
  38 static struct smc_lgr_list smc_lgr_list = {     /* established link groups */
  39         .lock = __SPIN_LOCK_UNLOCKED(smc_lgr_list.lock),
  40         .list = LIST_HEAD_INIT(smc_lgr_list.list),
  41         .num = 0,
  42 };
  43
  44 static atomic_t lgr_cnt = ATOMIC_INIT(0); /* number of existing link groups */
  45 static DECLARE_WAIT_QUEUE_HEAD(lgrs_deleted);
  46
  47 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
  48                          struct smc_buf_desc *buf_desc);
  49 static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft);
  50
  51 static void smc_link_down_work(struct work_struct *work);
  52
  53 /* return head of link group list and its lock for a given link group */
  54 static inline struct list_head *smc_lgr_list_head(struct smc_link_group *lgr,
  55                                                   spinlock_t **lgr_lock)
  56 {
  57         if (lgr->is_smcd) {
  58                 *lgr_lock = &lgr->smcd->lgr_lock;
  59                 return &lgr->smcd->lgr_list;
  60         }
  61
  62         *lgr_lock = &smc_lgr_list.lock;
  63         return &smc_lgr_list.list;
  64 }
  65
  66 static void smc_lgr_schedule_free_work(struct smc_link_group *lgr)
  67 {
  68         /* client link group creation always follows the server link group
  69          * creation. For client use a somewhat higher removal delay time,
  70          * otherwise there is a risk of out-of-sync link groups.
  71          */
  72         if (!lgr->freeing) {
  73                 mod_delayed_work(system_wq, &lgr->free_work,
  74                                  (!lgr->is_smcd && lgr->role == SMC_CLNT) ?
  75                                                 SMC_LGR_FREE_DELAY_CLNT :
  76                                                 SMC_LGR_FREE_DELAY_SERV);
  77         }
  78 }
  79
  80 /* Register connection's alert token in our lookup structure.
  81  * To use rbtrees we have to implement our own insert core.
  82  * Requires @conns_lock
  83  * @smc         connection to register
  84  * Returns 0 on success, != otherwise.
  85  */
  86 static void smc_lgr_add_alert_token(struct smc_connection *conn)
  87 {
  88         struct rb_node **link, *parent = NULL;
  89         u32 token = conn->alert_token_local;
  90
  91         link = &conn->lgr->conns_all.rb_node;
  92         while (*link) {
  93                 struct smc_connection *cur = rb_entry(*link,
  94                                         struct smc_connection, alert_node);
  95
  96                 parent = *link;
  97                 if (cur->alert_token_local > token)
  98                         link = &parent->rb_left;
  99                 else
 100                         link = &parent->rb_right;
 101         }
 102         /* Put the new node there */
 103         rb_link_node(&conn->alert_node, parent, link);
 104         rb_insert_color(&conn->alert_node, &conn->lgr->conns_all);
 105 }
 106
 107 /* assign an SMC-R link to the connection */
 108 static int smcr_lgr_conn_assign_link(struct smc_connection *conn, bool first)
 109 {
 110         enum smc_link_state expected = first ? SMC_LNK_ACTIVATING :
 111                                        SMC_LNK_ACTIVE;
 112         int i, j;
 113
 114         /* do link balancing */
 115         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 116                 struct smc_link *lnk = &conn->lgr->lnk[i];
 117
 118                 if (lnk->state != expected || lnk->link_is_asym)
 119                         continue;
 120                 if (conn->lgr->role == SMC_CLNT) {
 121                         conn->lnk = lnk; /* temporary, SMC server assigns link*/
 122                         break;
 123                 }
 124                 if (conn->lgr->conns_num % 2) {
 125                         for (j = i + 1; j < SMC_LINKS_PER_LGR_MAX; j++) {
 126                                 struct smc_link *lnk2;
 127
 128                                 lnk2 = &conn->lgr->lnk[j];
 129                                 if (lnk2->state == expected &&
 130                                     !lnk2->link_is_asym) {
 131                                         conn->lnk = lnk2;
 132                                         break;
 133                                 }
 134                         }
 135                 }
 136                 if (!conn->lnk)
 137                         conn->lnk = lnk;
 138                 break;
 139         }
 140         if (!conn->lnk)
 141                 return SMC_CLC_DECL_NOACTLINK;
 142         return 0;
 143 }
 144
 145 /* Register connection in link group by assigning an alert token
 146  * registered in a search tree.
 147  * Requires @conns_lock
 148  * Note that '0' is a reserved value and not assigned.
 149  */
 150 static int smc_lgr_register_conn(struct smc_connection *conn, bool first)
 151 {
 152         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
 153         static atomic_t nexttoken = ATOMIC_INIT(0);
 154         int rc;
 155
 156         if (!conn->lgr->is_smcd) {
 157                 rc = smcr_lgr_conn_assign_link(conn, first);
 158                 if (rc)
 159                         return rc;
 160         }
 161         /* find a new alert_token_local value not yet used by some connection
 162          * in this link group
 163          */
 164         sock_hold(&smc->sk); /* sock_put in smc_lgr_unregister_conn() */
 165         while (!conn->alert_token_local) {
 166                 conn->alert_token_local = atomic_inc_return(&nexttoken);
 167                 if (smc_lgr_find_conn(conn->alert_token_local, conn->lgr))
 168                         conn->alert_token_local = 0;
 169         }
 170         smc_lgr_add_alert_token(conn);
 171         conn->lgr->conns_num++;
 172         return 0;
 173 }
 174
 175 /* Unregister connection and reset the alert token of the given connection<
 176  */
 177 static void __smc_lgr_unregister_conn(struct smc_connection *conn)
 178 {
 179         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
 180         struct smc_link_group *lgr = conn->lgr;
 181
 182         rb_erase(&conn->alert_node, &lgr->conns_all);
 183         lgr->conns_num--;
 184         conn->alert_token_local = 0;
 185         sock_put(&smc->sk); /* sock_hold in smc_lgr_register_conn() */
 186 }
 187
 188 /* Unregister connection from lgr
 189  */
 190 static void smc_lgr_unregister_conn(struct smc_connection *conn)
 191 {
 192         struct smc_link_group *lgr = conn->lgr;
 193
 194         if (!lgr)
 195                 return;
 196         write_lock_bh(&lgr->conns_lock);
 197         if (conn->alert_token_local) {
 198                 __smc_lgr_unregister_conn(conn);
 199         }
 200         write_unlock_bh(&lgr->conns_lock);
 201         conn->lgr = NULL;
 202 }
 203
 204 void smc_lgr_cleanup_early(struct smc_connection *conn)
 205 {
 206         struct smc_link_group *lgr = conn->lgr;
 207         spinlock_t *lgr_lock;
 208
 209         if (!lgr)
 210                 return;
 211
 212         smc_conn_free(conn);
 213         smc_lgr_list_head(lgr, &lgr_lock);
 214         spin_lock_bh(lgr_lock);
 215         /* do not use this link group for new connections */
 216         if (!list_empty(&lgr->list))
 217                 list_del_init(&lgr->list);
 218         spin_unlock_bh(lgr_lock);
 219         __smc_lgr_terminate(lgr, true);
 220 }
 221
 222 static void smcr_lgr_link_deactivate_all(struct smc_link_group *lgr)
 223 {
 224         int i;
 225
 226         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 227                 struct smc_link *lnk = &lgr->lnk[i];
 228
 229                 if (smc_link_sendable(lnk))
 230                         lnk->state = SMC_LNK_INACTIVE;
 231         }
 232         wake_up_all(&lgr->llc_msg_waiter);
 233         wake_up_all(&lgr->llc_flow_waiter);
 234 }
 235
 236 static void smc_lgr_free(struct smc_link_group *lgr);
 237
 238 static void smc_lgr_free_work(struct work_struct *work)
 239 {
 240         struct smc_link_group *lgr = container_of(to_delayed_work(work),
 241                                                   struct smc_link_group,
 242                                                   free_work);
 243         spinlock_t *lgr_lock;
 244         bool conns;
 245
 246         smc_lgr_list_head(lgr, &lgr_lock);
 247         spin_lock_bh(lgr_lock);
 248         if (lgr->freeing) {
 249                 spin_unlock_bh(lgr_lock);
 250                 return;
 251         }
 252         read_lock_bh(&lgr->conns_lock);
 253         conns = RB_EMPTY_ROOT(&lgr->conns_all);
 254         read_unlock_bh(&lgr->conns_lock);
 255         if (!conns) { /* number of lgr connections is no longer zero */
 256                 spin_unlock_bh(lgr_lock);
 257                 return;
 258         }
 259         list_del_init(&lgr->list); /* remove from smc_lgr_list */
 260         lgr->freeing = 1; /* this instance does the freeing, no new schedule */
 261         spin_unlock_bh(lgr_lock);
 262         cancel_delayed_work(&lgr->free_work);
 263
 264         if (!lgr->is_smcd && !lgr->terminating)
 265                 smc_llc_send_link_delete_all(lgr, true,
 266                                              SMC_LLC_DEL_PROG_INIT_TERM);
 267         if (lgr->is_smcd && !lgr->terminating)
 268                 smc_ism_signal_shutdown(lgr);
 269         if (!lgr->is_smcd)
 270                 smcr_lgr_link_deactivate_all(lgr);
 271         smc_lgr_free(lgr);
 272 }
 273
 274 static void smc_lgr_terminate_work(struct work_struct *work)
 275 {
 276         struct smc_link_group *lgr = container_of(work, struct smc_link_group,
 277                                                   terminate_work);
 278
 279         __smc_lgr_terminate(lgr, true);
 280 }
 281
 282 /* return next unique link id for the lgr */
 283 static u8 smcr_next_link_id(struct smc_link_group *lgr)
 284 {
 285         u8 link_id;
 286         int i;
 287
 288         while (1) {
 289 again:
 290                 link_id = ++lgr->next_link_id;
 291                 if (!link_id)   /* skip zero as link_id */
 292                         link_id = ++lgr->next_link_id;
 293                 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 294                         if (smc_link_usable(&lgr->lnk[i]) &&
 295                             lgr->lnk[i].link_id == link_id)
 296                                 goto again;
 297                 }
 298                 break;
 299         }
 300         return link_id;
 301 }
 302
 303 int smcr_link_init(struct smc_link_group *lgr, struct smc_link *lnk,
 304                    u8 link_idx, struct smc_init_info *ini)
 305 {
 306         u8 rndvec[3];
 307         int rc;
 308
 309         get_device(&ini->ib_dev->ibdev->dev);
 310         atomic_inc(&ini->ib_dev->lnk_cnt);
 311         lnk->link_id = smcr_next_link_id(lgr);
 312         lnk->lgr = lgr;
 313         lnk->link_idx = link_idx;
 314         lnk->smcibdev = ini->ib_dev;
 315         lnk->ibport = ini->ib_port;
 316         lnk->path_mtu = ini->ib_dev->pattr[ini->ib_port - 1].active_mtu;
 317         smc_llc_link_set_uid(lnk);
 318         INIT_WORK(&lnk->link_down_wrk, smc_link_down_work);
 319         if (!ini->ib_dev->initialized) {
 320                 rc = (int)smc_ib_setup_per_ibdev(ini->ib_dev);
 321                 if (rc)
 322                         goto out;
 323         }
 324         get_random_bytes(rndvec, sizeof(rndvec));
 325         lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) +
 326                 (rndvec[2] << 16);
 327         rc = smc_ib_determine_gid(lnk->smcibdev, lnk->ibport,
 328                                   ini->vlan_id, lnk->gid, &lnk->sgid_index);
 329         if (rc)
 330                 goto out;
 331         rc = smc_llc_link_init(lnk);
 332         if (rc)
 333                 goto out;
 334         rc = smc_wr_alloc_link_mem(lnk);
 335         if (rc)
 336                 goto clear_llc_lnk;
 337         rc = smc_ib_create_protection_domain(lnk);
 338         if (rc)
 339                 goto free_link_mem;
 340         rc = smc_ib_create_queue_pair(lnk);
 341         if (rc)
 342                 goto dealloc_pd;
 343         rc = smc_wr_create_link(lnk);
 344         if (rc)
 345                 goto destroy_qp;
 346         lnk->state = SMC_LNK_ACTIVATING;
 347         return 0;
 348
 349 destroy_qp:
 350         smc_ib_destroy_queue_pair(lnk);
 351 dealloc_pd:
 352         smc_ib_dealloc_protection_domain(lnk);
 353 free_link_mem:
 354         smc_wr_free_link_mem(lnk);
 355 clear_llc_lnk:
 356         smc_llc_link_clear(lnk, false);
 357 out:
 358         put_device(&ini->ib_dev->ibdev->dev);
 359         memset(lnk, 0, sizeof(struct smc_link));
 360         lnk->state = SMC_LNK_UNUSED;
 361         if (!atomic_dec_return(&ini->ib_dev->lnk_cnt))
 362                 wake_up(&ini->ib_dev->lnks_deleted);
 363         return rc;
 364 }
 365
 366 /* create a new SMC link group */
 367 static int smc_lgr_create(struct smc_sock *smc, struct smc_init_info *ini)
 368 {
 369         struct smc_link_group *lgr;
 370         struct list_head *lgr_list;
 371         struct smc_link *lnk;
 372         spinlock_t *lgr_lock;
 373         u8 link_idx;
 374         int rc = 0;
 375         int i;
 376
 377         if (ini->is_smcd && ini->vlan_id) {
 378                 if (smc_ism_get_vlan(ini->ism_dev[ini->ism_selected],
 379                                      ini->vlan_id)) {
 380                         rc = SMC_CLC_DECL_ISMVLANERR;
 381                         goto out;
 382                 }
 383         }
 384
 385         lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
 386         if (!lgr) {
 387                 rc = SMC_CLC_DECL_MEM;
 388                 goto ism_put_vlan;
 389         }
 390         lgr->tx_wq = alloc_workqueue("smc_tx_wq-%*phN", 0, 0,
 391                                      SMC_LGR_ID_SIZE, &lgr->id);
 392         if (!lgr->tx_wq) {
 393                 rc = -ENOMEM;
 394                 goto free_lgr;
 395         }
 396         lgr->is_smcd = ini->is_smcd;
 397         lgr->sync_err = 0;
 398         lgr->terminating = 0;
 399         lgr->freeing = 0;
 400         lgr->vlan_id = ini->vlan_id;
 401         mutex_init(&lgr->sndbufs_lock);
 402         mutex_init(&lgr->rmbs_lock);
 403         rwlock_init(&lgr->conns_lock);
 404         for (i = 0; i < SMC_RMBE_SIZES; i++) {
 405                 INIT_LIST_HEAD(&lgr->sndbufs[i]);
 406                 INIT_LIST_HEAD(&lgr->rmbs[i]);
 407         }
 408         lgr->next_link_id = 0;
 409         smc_lgr_list.num += SMC_LGR_NUM_INCR;
 410         memcpy(&lgr->id, (u8 *)&smc_lgr_list.num, SMC_LGR_ID_SIZE);
 411         INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
 412         INIT_WORK(&lgr->terminate_work, smc_lgr_terminate_work);
 413         lgr->conns_all = RB_ROOT;
 414         if (ini->is_smcd) {
 415                 /* SMC-D specific settings */
 416                 get_device(&ini->ism_dev[ini->ism_selected]->dev);
 417                 lgr->peer_gid = ini->ism_peer_gid[ini->ism_selected];
 418                 lgr->smcd = ini->ism_dev[ini->ism_selected];
 419                 lgr_list = &ini->ism_dev[ini->ism_selected]->lgr_list;
 420                 lgr_lock = &lgr->smcd->lgr_lock;
 421                 lgr->smc_version = ini->smcd_version;
 422                 lgr->peer_shutdown = 0;
 423                 atomic_inc(&ini->ism_dev[ini->ism_selected]->lgr_cnt);
 424         } else {
 425                 /* SMC-R specific settings */
 426                 lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
 427                 memcpy(lgr->peer_systemid, ini->ib_lcl->id_for_peer,
 428                        SMC_SYSTEMID_LEN);
 429                 memcpy(lgr->pnet_id, ini->ib_dev->pnetid[ini->ib_port - 1],
 430                        SMC_MAX_PNETID_LEN);
 431                 smc_llc_lgr_init(lgr, smc);
 432
 433                 link_idx = SMC_SINGLE_LINK;
 434                 lnk = &lgr->lnk[link_idx];
 435                 rc = smcr_link_init(lgr, lnk, link_idx, ini);
 436                 if (rc)
 437                         goto free_wq;
 438                 lgr_list = &smc_lgr_list.list;
 439                 lgr_lock = &smc_lgr_list.lock;
 440                 atomic_inc(&lgr_cnt);
 441         }
 442         smc->conn.lgr = lgr;
 443         spin_lock_bh(lgr_lock);
 444         list_add_tail(&lgr->list, lgr_list);
 445         spin_unlock_bh(lgr_lock);
 446         return 0;
 447
 448 free_wq:
 449         destroy_workqueue(lgr->tx_wq);
 450 free_lgr:
 451         kfree(lgr);
 452 ism_put_vlan:
 453         if (ini->is_smcd && ini->vlan_id)
 454                 smc_ism_put_vlan(ini->ism_dev[ini->ism_selected], ini->vlan_id);
 455 out:
 456         if (rc < 0) {
 457                 if (rc == -ENOMEM)
 458                         rc = SMC_CLC_DECL_MEM;
 459                 else
 460                         rc = SMC_CLC_DECL_INTERR;
 461         }
 462         return rc;
 463 }
 464
 465 static int smc_write_space(struct smc_connection *conn)
 466 {
 467         int buffer_len = conn->peer_rmbe_size;
 468         union smc_host_cursor prod;
 469         union smc_host_cursor cons;
 470         int space;
 471
 472         smc_curs_copy(&prod, &conn->local_tx_ctrl.prod, conn);
 473         smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
 474         /* determine rx_buf space */
 475         space = buffer_len - smc_curs_diff(buffer_len, &cons, &prod);
 476         return space;
 477 }
 478
 479 static int smc_switch_cursor(struct smc_sock *smc, struct smc_cdc_tx_pend *pend,
 480                              struct smc_wr_buf *wr_buf)
 481 {
 482         struct smc_connection *conn = &smc->conn;
 483         union smc_host_cursor cons, fin;
 484         int rc = 0;
 485         int diff;
 486
 487         smc_curs_copy(&conn->tx_curs_sent, &conn->tx_curs_fin, conn);
 488         smc_curs_copy(&fin, &conn->local_tx_ctrl_fin, conn);
 489         /* set prod cursor to old state, enforce tx_rdma_writes() */
 490         smc_curs_copy(&conn->local_tx_ctrl.prod, &fin, conn);
 491         smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
 492
 493         if (smc_curs_comp(conn->peer_rmbe_size, &cons, &fin) < 0) {
 494                 /* cons cursor advanced more than fin, and prod was set
 495                  * fin above, so now prod is smaller than cons. Fix that.
 496                  */
 497                 diff = smc_curs_diff(conn->peer_rmbe_size, &fin, &cons);
 498                 smc_curs_add(conn->sndbuf_desc->len,
 499                              &conn->tx_curs_sent, diff);
 500                 smc_curs_add(conn->sndbuf_desc->len,
 501                              &conn->tx_curs_fin, diff);
 502
 503                 smp_mb__before_atomic();
 504                 atomic_add(diff, &conn->sndbuf_space);
 505                 smp_mb__after_atomic();
 506
 507                 smc_curs_add(conn->peer_rmbe_size,
 508                              &conn->local_tx_ctrl.prod, diff);
 509                 smc_curs_add(conn->peer_rmbe_size,
 510                              &conn->local_tx_ctrl_fin, diff);
 511         }
 512         /* recalculate, value is used by tx_rdma_writes() */
 513         atomic_set(&smc->conn.peer_rmbe_space, smc_write_space(conn));
 514
 515         if (smc->sk.sk_state != SMC_INIT &&
 516             smc->sk.sk_state != SMC_CLOSED) {
 517                 rc = smcr_cdc_msg_send_validation(conn, pend, wr_buf);
 518                 if (!rc) {
 519                         queue_delayed_work(conn->lgr->tx_wq, &conn->tx_work, 0);
 520                         smc->sk.sk_data_ready(&smc->sk);
 521                 }
 522         } else {
 523                 smc_wr_tx_put_slot(conn->lnk,
 524                                    (struct smc_wr_tx_pend_priv *)pend);
 525         }
 526         return rc;
 527 }
 528
 529 struct smc_link *smc_switch_conns(struct smc_link_group *lgr,
 530                                   struct smc_link *from_lnk, bool is_dev_err)
 531 {
 532         struct smc_link *to_lnk = NULL;
 533         struct smc_cdc_tx_pend *pend;
 534         struct smc_connection *conn;
 535         struct smc_wr_buf *wr_buf;
 536         struct smc_sock *smc;
 537         struct rb_node *node;
 538         int i, rc = 0;
 539
 540         /* link is inactive, wake up tx waiters */
 541         smc_wr_wakeup_tx_wait(from_lnk);
 542
 543         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 544                 if (!smc_link_active(&lgr->lnk[i]) || i == from_lnk->link_idx)
 545                         continue;
 546                 if (is_dev_err && from_lnk->smcibdev == lgr->lnk[i].smcibdev &&
 547                     from_lnk->ibport == lgr->lnk[i].ibport) {
 548                         continue;
 549                 }
 550                 to_lnk = &lgr->lnk[i];
 551                 break;
 552         }
 553         if (!to_lnk || !smc_wr_tx_link_hold(to_lnk)) {
 554                 smc_lgr_terminate_sched(lgr);
 555                 return NULL;
 556         }
 557 again:
 558         read_lock_bh(&lgr->conns_lock);
 559         for (node = rb_first(&lgr->conns_all); node; node = rb_next(node)) {
 560                 conn = rb_entry(node, struct smc_connection, alert_node);
 561                 if (conn->lnk != from_lnk)
 562                         continue;
 563                 smc = container_of(conn, struct smc_sock, conn);
 564                 /* conn->lnk not yet set in SMC_INIT state */
 565                 if (smc->sk.sk_state == SMC_INIT)
 566                         continue;
 567                 if (smc->sk.sk_state == SMC_CLOSED ||
 568                     smc->sk.sk_state == SMC_PEERCLOSEWAIT1 ||
 569                     smc->sk.sk_state == SMC_PEERCLOSEWAIT2 ||
 570                     smc->sk.sk_state == SMC_APPFINCLOSEWAIT ||
 571                     smc->sk.sk_state == SMC_APPCLOSEWAIT1 ||
 572                     smc->sk.sk_state == SMC_APPCLOSEWAIT2 ||
 573                     smc->sk.sk_state == SMC_PEERFINCLOSEWAIT ||
 574                     smc->sk.sk_state == SMC_PEERABORTWAIT ||
 575                     smc->sk.sk_state == SMC_PROCESSABORT) {
 576                         spin_lock_bh(&conn->send_lock);
 577                         conn->lnk = to_lnk;
 578                         spin_unlock_bh(&conn->send_lock);
 579                         continue;
 580                 }
 581                 sock_hold(&smc->sk);
 582                 read_unlock_bh(&lgr->conns_lock);
 583                 /* pre-fetch buffer outside of send_lock, might sleep */
 584                 rc = smc_cdc_get_free_slot(conn, to_lnk, &wr_buf, NULL, &pend);
 585                 if (rc)
 586                         goto err_out;
 587                 /* avoid race with smcr_tx_sndbuf_nonempty() */
 588                 spin_lock_bh(&conn->send_lock);
 589                 conn->lnk = to_lnk;
 590                 rc = smc_switch_cursor(smc, pend, wr_buf);
 591                 spin_unlock_bh(&conn->send_lock);
 592                 sock_put(&smc->sk);
 593                 if (rc)
 594                         goto err_out;
 595                 goto again;
 596         }
 597         read_unlock_bh(&lgr->conns_lock);
 598         smc_wr_tx_link_put(to_lnk);
 599         return to_lnk;
 600
 601 err_out:
 602         smcr_link_down_cond_sched(to_lnk);
 603         smc_wr_tx_link_put(to_lnk);
 604         return NULL;
 605 }
 606
 607 static void smcr_buf_unuse(struct smc_buf_desc *rmb_desc,
 608                            struct smc_link_group *lgr)
 609 {
 610         int rc;
 611
 612         if (rmb_desc->is_conf_rkey && !list_empty(&lgr->list)) {
 613                 /* unregister rmb with peer */
 614                 rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY);
 615                 if (!rc) {
 616                         /* protect against smc_llc_cli_rkey_exchange() */
 617                         mutex_lock(&lgr->llc_conf_mutex);
 618                         smc_llc_do_delete_rkey(lgr, rmb_desc);
 619                         rmb_desc->is_conf_rkey = false;
 620                         mutex_unlock(&lgr->llc_conf_mutex);
 621                         smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl);
 622                 }
 623         }
 624
 625         if (rmb_desc->is_reg_err) {
 626                 /* buf registration failed, reuse not possible */
 627                 mutex_lock(&lgr->rmbs_lock);
 628                 list_del(&rmb_desc->list);
 629                 mutex_unlock(&lgr->rmbs_lock);
 630
 631                 smc_buf_free(lgr, true, rmb_desc);
 632         } else {
 633                 rmb_desc->used = 0;
 634         }
 635 }
 636
 637 static void smc_buf_unuse(struct smc_connection *conn,
 638                           struct smc_link_group *lgr)
 639 {
 640         if (conn->sndbuf_desc)
 641                 conn->sndbuf_desc->used = 0;
 642         if (conn->rmb_desc && lgr->is_smcd)
 643                 conn->rmb_desc->used = 0;
 644         else if (conn->rmb_desc)
 645                 smcr_buf_unuse(conn->rmb_desc, lgr);
 646 }
 647
 648 /* remove a finished connection from its link group */
 649 void smc_conn_free(struct smc_connection *conn)
 650 {
 651         struct smc_link_group *lgr = conn->lgr;
 652
 653         if (!lgr)
 654                 return;
 655         if (lgr->is_smcd) {
 656                 if (!list_empty(&lgr->list))
 657                         smc_ism_unset_conn(conn);
 658                 tasklet_kill(&conn->rx_tsklet);
 659         } else {
 660                 smc_cdc_wait_pend_tx_wr(conn);
 661                 if (current_work() != &conn->abort_work)
 662                         cancel_work_sync(&conn->abort_work);
 663         }
 664         if (!list_empty(&lgr->list)) {
 665                 smc_buf_unuse(conn, lgr); /* allow buffer reuse */
 666                 smc_lgr_unregister_conn(conn);
 667         }
 668
 669         if (!lgr->conns_num)
 670                 smc_lgr_schedule_free_work(lgr);
 671 }
 672
 673 /* unregister a link from a buf_desc */
 674 static void smcr_buf_unmap_link(struct smc_buf_desc *buf_desc, bool is_rmb,
 675                                 struct smc_link *lnk)
 676 {
 677         if (is_rmb)
 678                 buf_desc->is_reg_mr[lnk->link_idx] = false;
 679         if (!buf_desc->is_map_ib[lnk->link_idx])
 680                 return;
 681         if (is_rmb) {
 682                 if (buf_desc->mr_rx[lnk->link_idx]) {
 683                         smc_ib_put_memory_region(
 684                                         buf_desc->mr_rx[lnk->link_idx]);
 685                         buf_desc->mr_rx[lnk->link_idx] = NULL;
 686                 }
 687                 smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_FROM_DEVICE);
 688         } else {
 689                 smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_TO_DEVICE);
 690         }
 691         sg_free_table(&buf_desc->sgt[lnk->link_idx]);
 692         buf_desc->is_map_ib[lnk->link_idx] = false;
 693 }
 694
 695 /* unmap all buffers of lgr for a deleted link */
 696 static void smcr_buf_unmap_lgr(struct smc_link *lnk)
 697 {
 698         struct smc_link_group *lgr = lnk->lgr;
 699         struct smc_buf_desc *buf_desc, *bf;
 700         int i;
 701
 702         for (i = 0; i < SMC_RMBE_SIZES; i++) {
 703                 mutex_lock(&lgr->rmbs_lock);
 704                 list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list)
 705                         smcr_buf_unmap_link(buf_desc, true, lnk);
 706                 mutex_unlock(&lgr->rmbs_lock);
 707                 mutex_lock(&lgr->sndbufs_lock);
 708                 list_for_each_entry_safe(buf_desc, bf, &lgr->sndbufs[i],
 709                                          list)
 710                         smcr_buf_unmap_link(buf_desc, false, lnk);
 711                 mutex_unlock(&lgr->sndbufs_lock);
 712         }
 713 }
 714
 715 static void smcr_rtoken_clear_link(struct smc_link *lnk)
 716 {
 717         struct smc_link_group *lgr = lnk->lgr;
 718         int i;
 719
 720         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
 721                 lgr->rtokens[i][lnk->link_idx].rkey = 0;
 722                 lgr->rtokens[i][lnk->link_idx].dma_addr = 0;
 723         }
 724 }
 725
 726 /* must be called under lgr->llc_conf_mutex lock */
 727 void smcr_link_clear(struct smc_link *lnk, bool log)
 728 {
 729         struct smc_ib_device *smcibdev;
 730
 731         if (!lnk->lgr || lnk->state == SMC_LNK_UNUSED)
 732                 return;
 733         lnk->peer_qpn = 0;
 734         smc_llc_link_clear(lnk, log);
 735         smcr_buf_unmap_lgr(lnk);
 736         smcr_rtoken_clear_link(lnk);
 737         smc_ib_modify_qp_error(lnk);
 738         smc_wr_free_link(lnk);
 739         smc_ib_destroy_queue_pair(lnk);
 740         smc_ib_dealloc_protection_domain(lnk);
 741         smc_wr_free_link_mem(lnk);
 742         put_device(&lnk->smcibdev->ibdev->dev);
 743         smcibdev = lnk->smcibdev;
 744         memset(lnk, 0, sizeof(struct smc_link));
 745         lnk->state = SMC_LNK_UNUSED;
 746         if (!atomic_dec_return(&smcibdev->lnk_cnt))
 747                 wake_up(&smcibdev->lnks_deleted);
 748 }
 749
 750 static void smcr_buf_free(struct smc_link_group *lgr, bool is_rmb,
 751                           struct smc_buf_desc *buf_desc)
 752 {
 753         int i;
 754
 755         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
 756                 smcr_buf_unmap_link(buf_desc, is_rmb, &lgr->lnk[i]);
 757
 758         if (buf_desc->pages)
 759                 __free_pages(buf_desc->pages, buf_desc->order);
 760         kfree(buf_desc);
 761 }
 762
 763 static void smcd_buf_free(struct smc_link_group *lgr, bool is_dmb,
 764                           struct smc_buf_desc *buf_desc)
 765 {
 766         if (is_dmb) {
 767                 /* restore original buf len */
 768                 buf_desc->len += sizeof(struct smcd_cdc_msg);
 769                 smc_ism_unregister_dmb(lgr->smcd, buf_desc);
 770         } else {
 771                 kfree(buf_desc->cpu_addr);
 772         }
 773         kfree(buf_desc);
 774 }
 775
 776 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
 777                          struct smc_buf_desc *buf_desc)
 778 {
 779         if (lgr->is_smcd)
 780                 smcd_buf_free(lgr, is_rmb, buf_desc);
 781         else
 782                 smcr_buf_free(lgr, is_rmb, buf_desc);
 783 }
 784
 785 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
 786 {
 787         struct smc_buf_desc *buf_desc, *bf_desc;
 788         struct list_head *buf_list;
 789         int i;
 790
 791         for (i = 0; i < SMC_RMBE_SIZES; i++) {
 792                 if (is_rmb)
 793                         buf_list = &lgr->rmbs[i];
 794                 else
 795                         buf_list = &lgr->sndbufs[i];
 796                 list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
 797                                          list) {
 798                         list_del(&buf_desc->list);
 799                         smc_buf_free(lgr, is_rmb, buf_desc);
 800                 }
 801         }
 802 }
 803
 804 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
 805 {
 806         /* free send buffers */
 807         __smc_lgr_free_bufs(lgr, false);
 808         /* free rmbs */
 809         __smc_lgr_free_bufs(lgr, true);
 810 }
 811
 812 /* remove a link group */
 813 static void smc_lgr_free(struct smc_link_group *lgr)
 814 {
 815         int i;
 816
 817         if (!lgr->is_smcd) {
 818                 mutex_lock(&lgr->llc_conf_mutex);
 819                 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 820                         if (lgr->lnk[i].state != SMC_LNK_UNUSED)
 821                                 smcr_link_clear(&lgr->lnk[i], false);
 822                 }
 823                 mutex_unlock(&lgr->llc_conf_mutex);
 824                 smc_llc_lgr_clear(lgr);
 825         }
 826
 827         smc_lgr_free_bufs(lgr);
 828         destroy_workqueue(lgr->tx_wq);
 829         if (lgr->is_smcd) {
 830                 smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
 831                 put_device(&lgr->smcd->dev);
 832                 if (!atomic_dec_return(&lgr->smcd->lgr_cnt))
 833                         wake_up(&lgr->smcd->lgrs_deleted);
 834         } else {
 835                 if (!atomic_dec_return(&lgr_cnt))
 836                         wake_up(&lgrs_deleted);
 837         }
 838         kfree(lgr);
 839 }
 840
 841 static void smcd_unregister_all_dmbs(struct smc_link_group *lgr)
 842 {
 843         int i;
 844
 845         for (i = 0; i < SMC_RMBE_SIZES; i++) {
 846                 struct smc_buf_desc *buf_desc;
 847
 848                 list_for_each_entry(buf_desc, &lgr->rmbs[i], list) {
 849                         buf_desc->len += sizeof(struct smcd_cdc_msg);
 850                         smc_ism_unregister_dmb(lgr->smcd, buf_desc);
 851                 }
 852         }
 853 }
 854
 855 static void smc_sk_wake_ups(struct smc_sock *smc)
 856 {
 857         smc->sk.sk_write_space(&smc->sk);
 858         smc->sk.sk_data_ready(&smc->sk);
 859         smc->sk.sk_state_change(&smc->sk);
 860 }
 861
 862 /* kill a connection */
 863 static void smc_conn_kill(struct smc_connection *conn, bool soft)
 864 {
 865         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
 866
 867         if (conn->lgr->is_smcd && conn->lgr->peer_shutdown)
 868                 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
 869         else
 870                 smc_close_abort(conn);
 871         conn->killed = 1;
 872         smc->sk.sk_err = ECONNABORTED;
 873         smc_sk_wake_ups(smc);
 874         if (conn->lgr->is_smcd) {
 875                 smc_ism_unset_conn(conn);
 876                 if (soft)
 877                         tasklet_kill(&conn->rx_tsklet);
 878                 else
 879                         tasklet_unlock_wait(&conn->rx_tsklet);
 880         } else {
 881                 smc_cdc_wait_pend_tx_wr(conn);
 882         }
 883         smc_lgr_unregister_conn(conn);
 884         smc_close_active_abort(smc);
 885 }
 886
 887 static void smc_lgr_cleanup(struct smc_link_group *lgr)
 888 {
 889         if (lgr->is_smcd) {
 890                 smc_ism_signal_shutdown(lgr);
 891                 smcd_unregister_all_dmbs(lgr);
 892         } else {
 893                 u32 rsn = lgr->llc_termination_rsn;
 894
 895                 if (!rsn)
 896                         rsn = SMC_LLC_DEL_PROG_INIT_TERM;
 897                 smc_llc_send_link_delete_all(lgr, false, rsn);
 898                 smcr_lgr_link_deactivate_all(lgr);
 899         }
 900 }
 901
 902 /* terminate link group
 903  * @soft: true if link group shutdown can take its time
 904  *        false if immediate link group shutdown is required
 905  */
 906 static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft)
 907 {
 908         struct smc_connection *conn;
 909         struct smc_sock *smc;
 910         struct rb_node *node;
 911
 912         if (lgr->terminating)
 913                 return; /* lgr already terminating */
 914         /* cancel free_work sync, will terminate when lgr->freeing is set */
 915         cancel_delayed_work(&lgr->free_work);
 916         lgr->terminating = 1;
 917
 918         /* kill remaining link group connections */
 919         read_lock_bh(&lgr->conns_lock);
 920         node = rb_first(&lgr->conns_all);
 921         while (node) {
 922                 read_unlock_bh(&lgr->conns_lock);
 923                 conn = rb_entry(node, struct smc_connection, alert_node);
 924                 smc = container_of(conn, struct smc_sock, conn);
 925                 sock_hold(&smc->sk); /* sock_put below */
 926                 lock_sock(&smc->sk);
 927                 smc_conn_kill(conn, soft);
 928                 release_sock(&smc->sk);
 929                 sock_put(&smc->sk); /* sock_hold above */
 930                 read_lock_bh(&lgr->conns_lock);
 931                 node = rb_first(&lgr->conns_all);
 932         }
 933         read_unlock_bh(&lgr->conns_lock);
 934         smc_lgr_cleanup(lgr);
 935         smc_lgr_free(lgr);
 936 }
 937
 938 /* unlink link group and schedule termination */
 939 void smc_lgr_terminate_sched(struct smc_link_group *lgr)
 940 {
 941         spinlock_t *lgr_lock;
 942
 943         smc_lgr_list_head(lgr, &lgr_lock);
 944         spin_lock_bh(lgr_lock);
 945         if (list_empty(&lgr->list) || lgr->terminating || lgr->freeing) {
 946                 spin_unlock_bh(lgr_lock);
 947                 return; /* lgr already terminating */
 948         }
 949         list_del_init(&lgr->list);
 950         lgr->freeing = 1;
 951         spin_unlock_bh(lgr_lock);
 952         schedule_work(&lgr->terminate_work);
 953 }
 954
 955 /* Called when peer lgr shutdown (regularly or abnormally) is received */
 956 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid, unsigned short vlan)
 957 {
 958         struct smc_link_group *lgr, *l;
 959         LIST_HEAD(lgr_free_list);
 960
 961         /* run common cleanup function and build free list */
 962         spin_lock_bh(&dev->lgr_lock);
 963         list_for_each_entry_safe(lgr, l, &dev->lgr_list, list) {
 964                 if ((!peer_gid || lgr->peer_gid == peer_gid) &&
 965                     (vlan == VLAN_VID_MASK || lgr->vlan_id == vlan)) {
 966                         if (peer_gid) /* peer triggered termination */
 967                                 lgr->peer_shutdown = 1;
 968                         list_move(&lgr->list, &lgr_free_list);
 969                         lgr->freeing = 1;
 970                 }
 971         }
 972         spin_unlock_bh(&dev->lgr_lock);
 973
 974         /* cancel the regular free workers and actually free lgrs */
 975         list_for_each_entry_safe(lgr, l, &lgr_free_list, list) {
 976                 list_del_init(&lgr->list);
 977                 schedule_work(&lgr->terminate_work);
 978         }
 979 }
 980
 981 /* Called when an SMCD device is removed or the smc module is unloaded */
 982 void smc_smcd_terminate_all(struct smcd_dev *smcd)
 983 {
 984         struct smc_link_group *lgr, *lg;
 985         LIST_HEAD(lgr_free_list);
 986
 987         spin_lock_bh(&smcd->lgr_lock);
 988         list_splice_init(&smcd->lgr_list, &lgr_free_list);
 989         list_for_each_entry(lgr, &lgr_free_list, list)
 990                 lgr->freeing = 1;
 991         spin_unlock_bh(&smcd->lgr_lock);
 992
 993         list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
 994                 list_del_init(&lgr->list);
 995                 __smc_lgr_terminate(lgr, false);
 996         }
 997
 998         if (atomic_read(&smcd->lgr_cnt))
 999                 wait_event(smcd->lgrs_deleted, !atomic_read(&smcd->lgr_cnt));
1000 }
1001
1002 /* Called when an SMCR device is removed or the smc module is unloaded.
1003  * If smcibdev is given, all SMCR link groups using this device are terminated.
1004  * If smcibdev is NULL, all SMCR link groups are terminated.
1005  */
1006 void smc_smcr_terminate_all(struct smc_ib_device *smcibdev)
1007 {
1008         struct smc_link_group *lgr, *lg;
1009         LIST_HEAD(lgr_free_list);
1010         int i;
1011
1012         spin_lock_bh(&smc_lgr_list.lock);
1013         if (!smcibdev) {
1014                 list_splice_init(&smc_lgr_list.list, &lgr_free_list);
1015                 list_for_each_entry(lgr, &lgr_free_list, list)
1016                         lgr->freeing = 1;
1017         } else {
1018                 list_for_each_entry_safe(lgr, lg, &smc_lgr_list.list, list) {
1019                         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1020                                 if (lgr->lnk[i].smcibdev == smcibdev)
1021                                         smcr_link_down_cond_sched(&lgr->lnk[i]);
1022                         }
1023                 }
1024         }
1025         spin_unlock_bh(&smc_lgr_list.lock);
1026
1027         list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
1028                 list_del_init(&lgr->list);
1029                 smc_llc_set_termination_rsn(lgr, SMC_LLC_DEL_OP_INIT_TERM);
1030                 __smc_lgr_terminate(lgr, false);
1031         }
1032
1033         if (smcibdev) {
1034                 if (atomic_read(&smcibdev->lnk_cnt))
1035                         wait_event(smcibdev->lnks_deleted,
1036                                    !atomic_read(&smcibdev->lnk_cnt));
1037         } else {
1038                 if (atomic_read(&lgr_cnt))
1039                         wait_event(lgrs_deleted, !atomic_read(&lgr_cnt));
1040         }
1041 }
1042
1043 /* set new lgr type and clear all asymmetric link tagging */
1044 void smcr_lgr_set_type(struct smc_link_group *lgr, enum smc_lgr_type new_type)
1045 {
1046         char *lgr_type = "";
1047         int i;
1048
1049         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
1050                 if (smc_link_usable(&lgr->lnk[i]))
1051                         lgr->lnk[i].link_is_asym = false;
1052         if (lgr->type == new_type)
1053                 return;
1054         lgr->type = new_type;
1055
1056         switch (lgr->type) {
1057         case SMC_LGR_NONE:
1058                 lgr_type = "NONE";
1059                 break;
1060         case SMC_LGR_SINGLE:
1061                 lgr_type = "SINGLE";
1062                 break;
1063         case SMC_LGR_SYMMETRIC:
1064                 lgr_type = "SYMMETRIC";
1065                 break;
1066         case SMC_LGR_ASYMMETRIC_PEER:
1067                 lgr_type = "ASYMMETRIC_PEER";
1068                 break;
1069         case SMC_LGR_ASYMMETRIC_LOCAL:
1070                 lgr_type = "ASYMMETRIC_LOCAL";
1071                 break;
1072         }
1073         pr_warn_ratelimited("smc: SMC-R lg %*phN state changed: "
1074                             "%s, pnetid %.16s\n", SMC_LGR_ID_SIZE, &lgr->id,
1075                             lgr_type, lgr->pnet_id);
1076 }
1077
1078 /* set new lgr type and tag a link as asymmetric */
1079 void smcr_lgr_set_type_asym(struct smc_link_group *lgr,
1080                             enum smc_lgr_type new_type, int asym_lnk_idx)
1081 {
1082         smcr_lgr_set_type(lgr, new_type);
1083         lgr->lnk[asym_lnk_idx].link_is_asym = true;
1084 }
1085
1086 /* abort connection, abort_work scheduled from tasklet context */
1087 static void smc_conn_abort_work(struct work_struct *work)
1088 {
1089         struct smc_connection *conn = container_of(work,
1090                                                    struct smc_connection,
1091                                                    abort_work);
1092         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1093
1094         lock_sock(&smc->sk);
1095         smc_conn_kill(conn, true);
1096         release_sock(&smc->sk);
1097         sock_put(&smc->sk); /* sock_hold done by schedulers of abort_work */
1098 }
1099
1100 void smcr_port_add(struct smc_ib_device *smcibdev, u8 ibport)
1101 {
1102         struct smc_link_group *lgr, *n;
1103
1104         spin_lock_bh(&smc_lgr_list.lock);
1105         list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1106                 struct smc_link *link;
1107
1108                 if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1109                             SMC_MAX_PNETID_LEN) ||
1110                     lgr->type == SMC_LGR_SYMMETRIC ||
1111                     lgr->type == SMC_LGR_ASYMMETRIC_PEER)
1112                         continue;
1113
1114                 /* trigger local add link processing */
1115                 link = smc_llc_usable_link(lgr);
1116                 if (link)
1117                         smc_llc_add_link_local(link);
1118         }
1119         spin_unlock_bh(&smc_lgr_list.lock);
1120 }
1121
1122 /* link is down - switch connections to alternate link,
1123  * must be called under lgr->llc_conf_mutex lock
1124  */
1125 static void smcr_link_down(struct smc_link *lnk)
1126 {
1127         struct smc_link_group *lgr = lnk->lgr;
1128         struct smc_link *to_lnk;
1129         int del_link_id;
1130
1131         if (!lgr || lnk->state == SMC_LNK_UNUSED || list_empty(&lgr->list))
1132                 return;
1133
1134         to_lnk = smc_switch_conns(lgr, lnk, true);
1135         if (!to_lnk) { /* no backup link available */
1136                 smcr_link_clear(lnk, true);
1137                 return;
1138         }
1139         smcr_lgr_set_type(lgr, SMC_LGR_SINGLE);
1140         del_link_id = lnk->link_id;
1141
1142         if (lgr->role == SMC_SERV) {
1143                 /* trigger local delete link processing */
1144                 smc_llc_srv_delete_link_local(to_lnk, del_link_id);
1145         } else {
1146                 if (lgr->llc_flow_lcl.type != SMC_LLC_FLOW_NONE) {
1147                         /* another llc task is ongoing */
1148                         mutex_unlock(&lgr->llc_conf_mutex);
1149                         wait_event_timeout(lgr->llc_flow_waiter,
1150                                 (list_empty(&lgr->list) ||
1151                                  lgr->llc_flow_lcl.type == SMC_LLC_FLOW_NONE),
1152                                 SMC_LLC_WAIT_TIME);
1153                         mutex_lock(&lgr->llc_conf_mutex);
1154                 }
1155                 if (!list_empty(&lgr->list)) {
1156                         smc_llc_send_delete_link(to_lnk, del_link_id,
1157                                                  SMC_LLC_REQ, true,
1158                                                  SMC_LLC_DEL_LOST_PATH);
1159                         smcr_link_clear(lnk, true);
1160                 }
1161                 wake_up(&lgr->llc_flow_waiter); /* wake up next waiter */
1162         }
1163 }
1164
1165 /* must be called under lgr->llc_conf_mutex lock */
1166 void smcr_link_down_cond(struct smc_link *lnk)
1167 {
1168         if (smc_link_downing(&lnk->state))
1169                 smcr_link_down(lnk);
1170 }
1171
1172 /* will get the lgr->llc_conf_mutex lock */
1173 void smcr_link_down_cond_sched(struct smc_link *lnk)
1174 {
1175         if (smc_link_downing(&lnk->state))
1176                 schedule_work(&lnk->link_down_wrk);
1177 }
1178
1179 void smcr_port_err(struct smc_ib_device *smcibdev, u8 ibport)
1180 {
1181         struct smc_link_group *lgr, *n;
1182         int i;
1183
1184         list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1185                 if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1186                             SMC_MAX_PNETID_LEN))
1187                         continue; /* lgr is not affected */
1188                 if (list_empty(&lgr->list))
1189                         continue;
1190                 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1191                         struct smc_link *lnk = &lgr->lnk[i];
1192
1193                         if (smc_link_usable(lnk) &&
1194                             lnk->smcibdev == smcibdev && lnk->ibport == ibport)
1195                                 smcr_link_down_cond_sched(lnk);
1196                 }
1197         }
1198 }
1199
1200 static void smc_link_down_work(struct work_struct *work)
1201 {
1202         struct smc_link *link = container_of(work, struct smc_link,
1203                                              link_down_wrk);
1204         struct smc_link_group *lgr = link->lgr;
1205
1206         if (list_empty(&lgr->list))
1207                 return;
1208         wake_up_all(&lgr->llc_msg_waiter);
1209         mutex_lock(&lgr->llc_conf_mutex);
1210         smcr_link_down(link);
1211         mutex_unlock(&lgr->llc_conf_mutex);
1212 }
1213
1214 static int smc_vlan_by_tcpsk_walk(struct net_device *lower_dev,
1215                                   struct netdev_nested_priv *priv)
1216 {
1217         unsigned short *vlan_id = (unsigned short *)priv->data;
1218
1219         if (is_vlan_dev(lower_dev)) {
1220                 *vlan_id = vlan_dev_vlan_id(lower_dev);
1221                 return 1;
1222         }
1223
1224         return 0;
1225 }
1226
1227 /* Determine vlan of internal TCP socket. */
1228 int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini)
1229 {
1230         struct dst_entry *dst = sk_dst_get(clcsock->sk);
1231         struct netdev_nested_priv priv;
1232         struct net_device *ndev;
1233         int rc = 0;
1234
1235         ini->vlan_id = 0;
1236         if (!dst) {
1237                 rc = -ENOTCONN;
1238                 goto out;
1239         }
1240         if (!dst->dev) {
1241                 rc = -ENODEV;
1242                 goto out_rel;
1243         }
1244
1245         ndev = dst->dev;
1246         if (is_vlan_dev(ndev)) {
1247                 ini->vlan_id = vlan_dev_vlan_id(ndev);
1248                 goto out_rel;
1249         }
1250
1251         priv.data = (void *)&ini->vlan_id;
1252         rtnl_lock();
1253         netdev_walk_all_lower_dev(ndev, smc_vlan_by_tcpsk_walk, &priv);
1254         rtnl_unlock();
1255
1256 out_rel:
1257         dst_release(dst);
1258 out:
1259         return rc;
1260 }
1261
1262 static bool smcr_lgr_match(struct smc_link_group *lgr,
1263                            struct smc_clc_msg_local *lcl,
1264                            enum smc_lgr_role role, u32 clcqpn)
1265 {
1266         int i;
1267
1268         if (memcmp(lgr->peer_systemid, lcl->id_for_peer, SMC_SYSTEMID_LEN) ||
1269             lgr->role != role)
1270                 return false;
1271
1272         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1273                 if (!smc_link_active(&lgr->lnk[i]))
1274                         continue;
1275                 if ((lgr->role == SMC_SERV || lgr->lnk[i].peer_qpn == clcqpn) &&
1276                     !memcmp(lgr->lnk[i].peer_gid, &lcl->gid, SMC_GID_SIZE) &&
1277                     !memcmp(lgr->lnk[i].peer_mac, lcl->mac, sizeof(lcl->mac)))
1278                         return true;
1279         }
1280         return false;
1281 }
1282
1283 static bool smcd_lgr_match(struct smc_link_group *lgr,
1284                            struct smcd_dev *smcismdev, u64 peer_gid)
1285 {
1286         return lgr->peer_gid == peer_gid && lgr->smcd == smcismdev;
1287 }
1288
1289 /* create a new SMC connection (and a new link group if necessary) */
1290 int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini)
1291 {
1292         struct smc_connection *conn = &smc->conn;
1293         struct list_head *lgr_list;
1294         struct smc_link_group *lgr;
1295         enum smc_lgr_role role;
1296         spinlock_t *lgr_lock;
1297         int rc = 0;
1298
1299         lgr_list = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_list :
1300                                   &smc_lgr_list.list;
1301         lgr_lock = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_lock :
1302                                   &smc_lgr_list.lock;
1303         ini->first_contact_local = 1;
1304         role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
1305         if (role == SMC_CLNT && ini->first_contact_peer)
1306                 /* create new link group as well */
1307                 goto create;
1308
1309         /* determine if an existing link group can be reused */
1310         spin_lock_bh(lgr_lock);
1311         list_for_each_entry(lgr, lgr_list, list) {
1312                 write_lock_bh(&lgr->conns_lock);
1313                 if ((ini->is_smcd ?
1314                      smcd_lgr_match(lgr, ini->ism_dev[ini->ism_selected],
1315                                     ini->ism_peer_gid[ini->ism_selected]) :
1316                      smcr_lgr_match(lgr, ini->ib_lcl, role, ini->ib_clcqpn)) &&
1317                     !lgr->sync_err &&
1318                     (ini->smcd_version == SMC_V2 ||
1319                      lgr->vlan_id == ini->vlan_id) &&
1320                     (role == SMC_CLNT || ini->is_smcd ||
1321                     (lgr->conns_num < SMC_RMBS_PER_LGR_MAX &&
1322                       !bitmap_full(lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX)))) {
1323                         /* link group found */
1324                         ini->first_contact_local = 0;
1325                         conn->lgr = lgr;
1326                         rc = smc_lgr_register_conn(conn, false);
1327                         write_unlock_bh(&lgr->conns_lock);
1328                         if (!rc && delayed_work_pending(&lgr->free_work))
1329                                 cancel_delayed_work(&lgr->free_work);
1330                         break;
1331                 }
1332                 write_unlock_bh(&lgr->conns_lock);
1333         }
1334         spin_unlock_bh(lgr_lock);
1335         if (rc)
1336                 return rc;
1337
1338         if (role == SMC_CLNT && !ini->first_contact_peer &&
1339             ini->first_contact_local) {
1340                 /* Server reuses a link group, but Client wants to start
1341                  * a new one
1342                  * send out_of_sync decline, reason synchr. error
1343                  */
1344                 return SMC_CLC_DECL_SYNCERR;
1345         }
1346
1347 create:
1348         if (ini->first_contact_local) {
1349                 rc = smc_lgr_create(smc, ini);
1350                 if (rc)
1351                         goto out;
1352                 lgr = conn->lgr;
1353                 write_lock_bh(&lgr->conns_lock);
1354                 rc = smc_lgr_register_conn(conn, true);
1355                 write_unlock_bh(&lgr->conns_lock);
1356                 if (rc)
1357                         goto out;
1358         }
1359         conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
1360         conn->local_tx_ctrl.len = SMC_WR_TX_SIZE;
1361         conn->urg_state = SMC_URG_READ;
1362         init_waitqueue_head(&conn->cdc_pend_tx_wq);
1363         INIT_WORK(&smc->conn.abort_work, smc_conn_abort_work);
1364         if (ini->is_smcd) {
1365                 conn->rx_off = sizeof(struct smcd_cdc_msg);
1366                 smcd_cdc_rx_init(conn); /* init tasklet for this conn */
1367         } else {
1368                 conn->rx_off = 0;
1369         }
1370 #ifndef KERNEL_HAS_ATOMIC64
1371         spin_lock_init(&conn->acurs_lock);
1372 #endif
1373
1374 out:
1375         return rc;
1376 }
1377
1378 /* convert the RMB size into the compressed notation - minimum 16K.
1379  * In contrast to plain ilog2, this rounds towards the next power of 2,
1380  * so the socket application gets at least its desired sndbuf / rcvbuf size.
1381  */
1382 static u8 smc_compress_bufsize(int size)
1383 {
1384         u8 compressed;
1385
1386         if (size <= SMC_BUF_MIN_SIZE)
1387                 return 0;
1388
1389         size = (size - 1) >> 14;
1390         compressed = ilog2(size) + 1;
1391         if (compressed >= SMC_RMBE_SIZES)
1392                 compressed = SMC_RMBE_SIZES - 1;
1393         return compressed;
1394 }
1395
1396 /* convert the RMB size from compressed notation into integer */
1397 int smc_uncompress_bufsize(u8 compressed)
1398 {
1399         u32 size;
1400
1401         size = 0x00000001 << (((int)compressed) + 14);
1402         return (int)size;
1403 }
1404
1405 /* try to reuse a sndbuf or rmb description slot for a certain
1406  * buffer size; if not available, return NULL
1407  */
1408 static struct smc_buf_desc *smc_buf_get_slot(int compressed_bufsize,
1409                                              struct mutex *lock,
1410                                              struct list_head *buf_list)
1411 {
1412         struct smc_buf_desc *buf_slot;
1413
1414         mutex_lock(lock);
1415         list_for_each_entry(buf_slot, buf_list, list) {
1416                 if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
1417                         mutex_unlock(lock);
1418                         return buf_slot;
1419                 }
1420         }
1421         mutex_unlock(lock);
1422         return NULL;
1423 }
1424
1425 /* one of the conditions for announcing a receiver's current window size is
1426  * that it "results in a minimum increase in the window size of 10% of the
1427  * receive buffer space" [RFC7609]
1428  */
1429 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
1430 {
1431         return max_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
1432 }
1433
1434 /* map an rmb buf to a link */
1435 static int smcr_buf_map_link(struct smc_buf_desc *buf_desc, bool is_rmb,
1436                              struct smc_link *lnk)
1437 {
1438         int rc;
1439
1440         if (buf_desc->is_map_ib[lnk->link_idx])
1441                 return 0;
1442
1443         rc = sg_alloc_table(&buf_desc->sgt[lnk->link_idx], 1, GFP_KERNEL);
1444         if (rc)
1445                 return rc;
1446         sg_set_buf(buf_desc->sgt[lnk->link_idx].sgl,
1447                    buf_desc->cpu_addr, buf_desc->len);
1448
1449         /* map sg table to DMA address */
1450         rc = smc_ib_buf_map_sg(lnk, buf_desc,
1451                                is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
1452         /* SMC protocol depends on mapping to one DMA address only */
1453         if (rc != 1) {
1454                 rc = -EAGAIN;
1455                 goto free_table;
1456         }
1457
1458         /* create a new memory region for the RMB */
1459         if (is_rmb) {
1460                 rc = smc_ib_get_memory_region(lnk->roce_pd,
1461                                               IB_ACCESS_REMOTE_WRITE |
1462                                               IB_ACCESS_LOCAL_WRITE,
1463                                               buf_desc, lnk->link_idx);
1464                 if (rc)
1465                         goto buf_unmap;
1466                 smc_ib_sync_sg_for_device(lnk, buf_desc, DMA_FROM_DEVICE);
1467         }
1468         buf_desc->is_map_ib[lnk->link_idx] = true;
1469         return 0;
1470
1471 buf_unmap:
1472         smc_ib_buf_unmap_sg(lnk, buf_desc,
1473                             is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
1474 free_table:
1475         sg_free_table(&buf_desc->sgt[lnk->link_idx]);
1476         return rc;
1477 }
1478
1479 /* register a new rmb on IB device,
1480  * must be called under lgr->llc_conf_mutex lock
1481  */
1482 int smcr_link_reg_rmb(struct smc_link *link, struct smc_buf_desc *rmb_desc)
1483 {
1484         if (list_empty(&link->lgr->list))
1485                 return -ENOLINK;
1486         if (!rmb_desc->is_reg_mr[link->link_idx]) {
1487                 /* register memory region for new rmb */
1488                 if (smc_wr_reg_send(link, rmb_desc->mr_rx[link->link_idx])) {
1489                         rmb_desc->is_reg_err = true;
1490                         return -EFAULT;
1491                 }
1492                 rmb_desc->is_reg_mr[link->link_idx] = true;
1493         }
1494         return 0;
1495 }
1496
1497 static int _smcr_buf_map_lgr(struct smc_link *lnk, struct mutex *lock,
1498                              struct list_head *lst, bool is_rmb)
1499 {
1500         struct smc_buf_desc *buf_desc, *bf;
1501         int rc = 0;
1502
1503         mutex_lock(lock);
1504         list_for_each_entry_safe(buf_desc, bf, lst, list) {
1505                 if (!buf_desc->used)
1506                         continue;
1507                 rc = smcr_buf_map_link(buf_desc, is_rmb, lnk);
1508                 if (rc)
1509                         goto out;
1510         }
1511 out:
1512         mutex_unlock(lock);
1513         return rc;
1514 }
1515
1516 /* map all used buffers of lgr for a new link */
1517 int smcr_buf_map_lgr(struct smc_link *lnk)
1518 {
1519         struct smc_link_group *lgr = lnk->lgr;
1520         int i, rc = 0;
1521
1522         for (i = 0; i < SMC_RMBE_SIZES; i++) {
1523                 rc = _smcr_buf_map_lgr(lnk, &lgr->rmbs_lock,
1524                                        &lgr->rmbs[i], true);
1525                 if (rc)
1526                         return rc;
1527                 rc = _smcr_buf_map_lgr(lnk, &lgr->sndbufs_lock,
1528                                        &lgr->sndbufs[i], false);
1529                 if (rc)
1530                         return rc;
1531         }
1532         return 0;
1533 }
1534
1535 /* register all used buffers of lgr for a new link,
1536  * must be called under lgr->llc_conf_mutex lock
1537  */
1538 int smcr_buf_reg_lgr(struct smc_link *lnk)
1539 {
1540         struct smc_link_group *lgr = lnk->lgr;
1541         struct smc_buf_desc *buf_desc, *bf;
1542         int i, rc = 0;
1543
1544         mutex_lock(&lgr->rmbs_lock);
1545         for (i = 0; i < SMC_RMBE_SIZES; i++) {
1546                 list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list) {
1547                         if (!buf_desc->used)
1548                                 continue;
1549                         rc = smcr_link_reg_rmb(lnk, buf_desc);
1550                         if (rc)
1551                                 goto out;
1552                 }
1553         }
1554 out:
1555         mutex_unlock(&lgr->rmbs_lock);
1556         return rc;
1557 }
1558
1559 static struct smc_buf_desc *smcr_new_buf_create(struct smc_link_group *lgr,
1560                                                 bool is_rmb, int bufsize)
1561 {
1562         struct smc_buf_desc *buf_desc;
1563
1564         /* try to alloc a new buffer */
1565         buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
1566         if (!buf_desc)
1567                 return ERR_PTR(-ENOMEM);
1568
1569         buf_desc->order = get_order(bufsize);
1570         buf_desc->pages = alloc_pages(GFP_KERNEL | __GFP_NOWARN |
1571                                       __GFP_NOMEMALLOC | __GFP_COMP |
1572                                       __GFP_NORETRY | __GFP_ZERO,
1573                                       buf_desc->order);
1574         if (!buf_desc->pages) {
1575                 kfree(buf_desc);
1576                 return ERR_PTR(-EAGAIN);
1577         }
1578         buf_desc->cpu_addr = (void *)page_address(buf_desc->pages);
1579         buf_desc->len = bufsize;
1580         return buf_desc;
1581 }
1582
1583 /* map buf_desc on all usable links,
1584  * unused buffers stay mapped as long as the link is up
1585  */
1586 static int smcr_buf_map_usable_links(struct smc_link_group *lgr,
1587                                      struct smc_buf_desc *buf_desc, bool is_rmb)
1588 {
1589         int i, rc = 0, cnt = 0;
1590
1591         /* protect against parallel link reconfiguration */
1592         mutex_lock(&lgr->llc_conf_mutex);
1593         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1594                 struct smc_link *lnk = &lgr->lnk[i];
1595
1596                 if (!smc_link_usable(lnk))
1597                         continue;
1598                 if (smcr_buf_map_link(buf_desc, is_rmb, lnk)) {
1599                         rc = -ENOMEM;
1600                         goto out;
1601                 }
1602                 cnt++;
1603         }
1604 out:
1605         mutex_unlock(&lgr->llc_conf_mutex);
1606         if (!rc && !cnt)
1607                 rc = -EINVAL;
1608         return rc;
1609 }
1610
1611 #define SMCD_DMBE_SIZES         6 /* 0 -> 16KB, 1 -> 32KB, .. 6 -> 1MB */
1612
1613 static struct smc_buf_desc *smcd_new_buf_create(struct smc_link_group *lgr,
1614                                                 bool is_dmb, int bufsize)
1615 {
1616         struct smc_buf_desc *buf_desc;
1617         int rc;
1618
1619         if (smc_compress_bufsize(bufsize) > SMCD_DMBE_SIZES)
1620                 return ERR_PTR(-EAGAIN);
1621
1622         /* try to alloc a new DMB */
1623         buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
1624         if (!buf_desc)
1625                 return ERR_PTR(-ENOMEM);
1626         if (is_dmb) {
1627                 rc = smc_ism_register_dmb(lgr, bufsize, buf_desc);
1628                 if (rc) {
1629                         kfree(buf_desc);
1630                         if (rc == -ENOMEM)
1631                                 return ERR_PTR(-EAGAIN);
1632                         if (rc == -ENOSPC)
1633                                 return ERR_PTR(-ENOSPC);
1634                         return ERR_PTR(-EIO);
1635                 }
1636                 buf_desc->pages = virt_to_page(buf_desc->cpu_addr);
1637                 /* CDC header stored in buf. So, pretend it was smaller */
1638                 buf_desc->len = bufsize - sizeof(struct smcd_cdc_msg);
1639         } else {
1640                 buf_desc->cpu_addr = kzalloc(bufsize, GFP_KERNEL |
1641                                              __GFP_NOWARN | __GFP_NORETRY |
1642                                              __GFP_NOMEMALLOC);
1643                 if (!buf_desc->cpu_addr) {
1644                         kfree(buf_desc);
1645                         return ERR_PTR(-EAGAIN);
1646                 }
1647                 buf_desc->len = bufsize;
1648         }
1649         return buf_desc;
1650 }
1651
1652 static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
1653 {
1654         struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
1655         struct smc_connection *conn = &smc->conn;
1656         struct smc_link_group *lgr = conn->lgr;
1657         struct list_head *buf_list;
1658         int bufsize, bufsize_short;
1659         struct mutex *lock;     /* lock buffer list */
1660         int sk_buf_size;
1661
1662         if (is_rmb)
1663                 /* use socket recv buffer size (w/o overhead) as start value */
1664                 sk_buf_size = smc->sk.sk_rcvbuf / 2;
1665         else
1666                 /* use socket send buffer size (w/o overhead) as start value */
1667                 sk_buf_size = smc->sk.sk_sndbuf / 2;
1668
1669         for (bufsize_short = smc_compress_bufsize(sk_buf_size);
1670              bufsize_short >= 0; bufsize_short--) {
1671
1672                 if (is_rmb) {
1673                         lock = &lgr->rmbs_lock;
1674                         buf_list = &lgr->rmbs[bufsize_short];
1675                 } else {
1676                         lock = &lgr->sndbufs_lock;
1677                         buf_list = &lgr->sndbufs[bufsize_short];
1678                 }
1679                 bufsize = smc_uncompress_bufsize(bufsize_short);
1680                 if ((1 << get_order(bufsize)) > SG_MAX_SINGLE_ALLOC)
1681                         continue;
1682
1683                 /* check for reusable slot in the link group */
1684                 buf_desc = smc_buf_get_slot(bufsize_short, lock, buf_list);
1685                 if (buf_desc) {
1686                         memset(buf_desc->cpu_addr, 0, bufsize);
1687                         break; /* found reusable slot */
1688                 }
1689
1690                 if (is_smcd)
1691                         buf_desc = smcd_new_buf_create(lgr, is_rmb, bufsize);
1692                 else
1693                         buf_desc = smcr_new_buf_create(lgr, is_rmb, bufsize);
1694
1695                 if (PTR_ERR(buf_desc) == -ENOMEM)
1696                         break;
1697                 if (IS_ERR(buf_desc))
1698                         continue;
1699
1700                 buf_desc->used = 1;
1701                 mutex_lock(lock);
1702                 list_add(&buf_desc->list, buf_list);
1703                 mutex_unlock(lock);
1704                 break; /* found */
1705         }
1706
1707         if (IS_ERR(buf_desc))
1708                 return PTR_ERR(buf_desc);
1709
1710         if (!is_smcd) {
1711                 if (smcr_buf_map_usable_links(lgr, buf_desc, is_rmb)) {
1712                         smcr_buf_unuse(buf_desc, lgr);
1713                         return -ENOMEM;
1714                 }
1715         }
1716
1717         if (is_rmb) {
1718                 conn->rmb_desc = buf_desc;
1719                 conn->rmbe_size_short = bufsize_short;
1720                 smc->sk.sk_rcvbuf = bufsize * 2;
1721                 atomic_set(&conn->bytes_to_rcv, 0);
1722                 conn->rmbe_update_limit =
1723                         smc_rmb_wnd_update_limit(buf_desc->len);
1724                 if (is_smcd)
1725                         smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */
1726         } else {
1727                 conn->sndbuf_desc = buf_desc;
1728                 smc->sk.sk_sndbuf = bufsize * 2;
1729                 atomic_set(&conn->sndbuf_space, bufsize);
1730         }
1731         return 0;
1732 }
1733
1734 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn)
1735 {
1736         if (!conn->lgr || conn->lgr->is_smcd || !smc_link_active(conn->lnk))
1737                 return;
1738         smc_ib_sync_sg_for_cpu(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE);
1739 }
1740
1741 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
1742 {
1743         if (!conn->lgr || conn->lgr->is_smcd || !smc_link_active(conn->lnk))
1744                 return;
1745         smc_ib_sync_sg_for_device(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE);
1746 }
1747
1748 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
1749 {
1750         int i;
1751
1752         if (!conn->lgr || conn->lgr->is_smcd)
1753                 return;
1754         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1755                 if (!smc_link_active(&conn->lgr->lnk[i]))
1756                         continue;
1757                 smc_ib_sync_sg_for_cpu(&conn->lgr->lnk[i], conn->rmb_desc,
1758                                        DMA_FROM_DEVICE);
1759         }
1760 }
1761
1762 void smc_rmb_sync_sg_for_device(struct smc_connection *conn)
1763 {
1764         int i;
1765
1766         if (!conn->lgr || conn->lgr->is_smcd)
1767                 return;
1768         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1769                 if (!smc_link_active(&conn->lgr->lnk[i]))
1770                         continue;
1771                 smc_ib_sync_sg_for_device(&conn->lgr->lnk[i], conn->rmb_desc,
1772                                           DMA_FROM_DEVICE);
1773         }
1774 }
1775
1776 /* create the send and receive buffer for an SMC socket;
1777  * receive buffers are called RMBs;
1778  * (even though the SMC protocol allows more than one RMB-element per RMB,
1779  * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
1780  * extra RMB for every connection in a link group
1781  */
1782 int smc_buf_create(struct smc_sock *smc, bool is_smcd)
1783 {
1784         int rc;
1785
1786         /* create send buffer */
1787         rc = __smc_buf_create(smc, is_smcd, false);
1788         if (rc)
1789                 return rc;
1790         /* create rmb */
1791         rc = __smc_buf_create(smc, is_smcd, true);
1792         if (rc) {
1793                 mutex_lock(&smc->conn.lgr->sndbufs_lock);
1794                 list_del(&smc->conn.sndbuf_desc->list);
1795                 mutex_unlock(&smc->conn.lgr->sndbufs_lock);
1796                 smc_buf_free(smc->conn.lgr, false, smc->conn.sndbuf_desc);
1797                 smc->conn.sndbuf_desc = NULL;
1798         }
1799         return rc;
1800 }
1801
1802 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
1803 {
1804         int i;
1805
1806         for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
1807                 if (!test_and_set_bit(i, lgr->rtokens_used_mask))
1808                         return i;
1809         }
1810         return -ENOSPC;
1811 }
1812
1813 static int smc_rtoken_find_by_link(struct smc_link_group *lgr, int lnk_idx,
1814                                    u32 rkey)
1815 {
1816         int i;
1817
1818         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1819                 if (test_bit(i, lgr->rtokens_used_mask) &&
1820                     lgr->rtokens[i][lnk_idx].rkey == rkey)
1821                         return i;
1822         }
1823         return -ENOENT;
1824 }
1825
1826 /* set rtoken for a new link to an existing rmb */
1827 void smc_rtoken_set(struct smc_link_group *lgr, int link_idx, int link_idx_new,
1828                     __be32 nw_rkey_known, __be64 nw_vaddr, __be32 nw_rkey)
1829 {
1830         int rtok_idx;
1831
1832         rtok_idx = smc_rtoken_find_by_link(lgr, link_idx, ntohl(nw_rkey_known));
1833         if (rtok_idx == -ENOENT)
1834                 return;
1835         lgr->rtokens[rtok_idx][link_idx_new].rkey = ntohl(nw_rkey);
1836         lgr->rtokens[rtok_idx][link_idx_new].dma_addr = be64_to_cpu(nw_vaddr);
1837 }
1838
1839 /* set rtoken for a new link whose link_id is given */
1840 void smc_rtoken_set2(struct smc_link_group *lgr, int rtok_idx, int link_id,
1841                      __be64 nw_vaddr, __be32 nw_rkey)
1842 {
1843         u64 dma_addr = be64_to_cpu(nw_vaddr);
1844         u32 rkey = ntohl(nw_rkey);
1845         bool found = false;
1846         int link_idx;
1847
1848         for (link_idx = 0; link_idx < SMC_LINKS_PER_LGR_MAX; link_idx++) {
1849                 if (lgr->lnk[link_idx].link_id == link_id) {
1850                         found = true;
1851                         break;
1852                 }
1853         }
1854         if (!found)
1855                 return;
1856         lgr->rtokens[rtok_idx][link_idx].rkey = rkey;
1857         lgr->rtokens[rtok_idx][link_idx].dma_addr = dma_addr;
1858 }
1859
1860 /* add a new rtoken from peer */
1861 int smc_rtoken_add(struct smc_link *lnk, __be64 nw_vaddr, __be32 nw_rkey)
1862 {
1863         struct smc_link_group *lgr = smc_get_lgr(lnk);
1864         u64 dma_addr = be64_to_cpu(nw_vaddr);
1865         u32 rkey = ntohl(nw_rkey);
1866         int i;
1867
1868         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1869                 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
1870                     lgr->rtokens[i][lnk->link_idx].dma_addr == dma_addr &&
1871                     test_bit(i, lgr->rtokens_used_mask)) {
1872                         /* already in list */
1873                         return i;
1874                 }
1875         }
1876         i = smc_rmb_reserve_rtoken_idx(lgr);
1877         if (i < 0)
1878                 return i;
1879         lgr->rtokens[i][lnk->link_idx].rkey = rkey;
1880         lgr->rtokens[i][lnk->link_idx].dma_addr = dma_addr;
1881         return i;
1882 }
1883
1884 /* delete an rtoken from all links */
1885 int smc_rtoken_delete(struct smc_link *lnk, __be32 nw_rkey)
1886 {
1887         struct smc_link_group *lgr = smc_get_lgr(lnk);
1888         u32 rkey = ntohl(nw_rkey);
1889         int i, j;
1890
1891         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1892                 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
1893                     test_bit(i, lgr->rtokens_used_mask)) {
1894                         for (j = 0; j < SMC_LINKS_PER_LGR_MAX; j++) {
1895                                 lgr->rtokens[i][j].rkey = 0;
1896                                 lgr->rtokens[i][j].dma_addr = 0;
1897                         }
1898                         clear_bit(i, lgr->rtokens_used_mask);
1899                         return 0;
1900                 }
1901         }
1902         return -ENOENT;
1903 }
1904
1905 /* save rkey and dma_addr received from peer during clc handshake */
1906 int smc_rmb_rtoken_handling(struct smc_connection *conn,
1907                             struct smc_link *lnk,
1908                             struct smc_clc_msg_accept_confirm *clc)
1909 {
1910         conn->rtoken_idx = smc_rtoken_add(lnk, clc->r0.rmb_dma_addr,
1911                                           clc->r0.rmb_rkey);
1912         if (conn->rtoken_idx < 0)
1913                 return conn->rtoken_idx;
1914         return 0;
1915 }
1916
1917 static void smc_core_going_away(void)
1918 {
1919         struct smc_ib_device *smcibdev;
1920         struct smcd_dev *smcd;
1921
1922         mutex_lock(&smc_ib_devices.mutex);
1923         list_for_each_entry(smcibdev, &smc_ib_devices.list, list) {
1924                 int i;
1925
1926                 for (i = 0; i < SMC_MAX_PORTS; i++)
1927                         set_bit(i, smcibdev->ports_going_away);
1928         }
1929         mutex_unlock(&smc_ib_devices.mutex);
1930
1931         mutex_lock(&smcd_dev_list.mutex);
1932         list_for_each_entry(smcd, &smcd_dev_list.list, list) {
1933                 smcd->going_away = 1;
1934         }
1935         mutex_unlock(&smcd_dev_list.mutex);
1936 }
1937
1938 /* Clean up all SMC link groups */
1939 static void smc_lgrs_shutdown(void)
1940 {
1941         struct smcd_dev *smcd;
1942
1943         smc_core_going_away();
1944
1945         smc_smcr_terminate_all(NULL);
1946
1947         mutex_lock(&smcd_dev_list.mutex);
1948         list_for_each_entry(smcd, &smcd_dev_list.list, list)
1949                 smc_smcd_terminate_all(smcd);
1950         mutex_unlock(&smcd_dev_list.mutex);
1951 }
1952
1953 static int smc_core_reboot_event(struct notifier_block *this,
1954                                  unsigned long event, void *ptr)
1955 {
1956         smc_lgrs_shutdown();
1957         smc_ib_unregister_client();
1958         return 0;
1959 }
1960
1961 static struct notifier_block smc_reboot_notifier = {
1962         .notifier_call = smc_core_reboot_event,
1963 };
1964
1965 int __init smc_core_init(void)
1966 {
1967         return register_reboot_notifier(&smc_reboot_notifier);
1968 }
1969
1970 /* Called (from smc_exit) when module is removed */
1971 void smc_core_exit(void)
1972 {
1973         unregister_reboot_notifier(&smc_reboot_notifier);
1974         smc_lgrs_shutdown();
1975 }