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 <[email protected]>
  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         struct list_head *lgr_list;
 208         spinlock_t *lgr_lock;
 209
 210         if (!lgr)
 211                 return;
 212
 213         smc_conn_free(conn);
 214         lgr_list = smc_lgr_list_head(lgr, &lgr_lock);
 215         spin_lock_bh(lgr_lock);
 216         /* do not use this link group for new connections */
 217         if (!list_empty(lgr_list))
 218                 list_del_init(lgr_list);
 219         spin_unlock_bh(lgr_lock);
 220         __smc_lgr_terminate(lgr, true);
 221 }
 222
 223 static void smcr_lgr_link_deactivate_all(struct smc_link_group *lgr)
 224 {
 225         int i;
 226
 227         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 228                 struct smc_link *lnk = &lgr->lnk[i];
 229
 230                 if (smc_link_usable(lnk))
 231                         lnk->state = SMC_LNK_INACTIVE;
 232         }
 233         wake_up_all(&lgr->llc_msg_waiter);
 234         wake_up_all(&lgr->llc_flow_waiter);
 235 }
 236
 237 static void smc_lgr_free(struct smc_link_group *lgr);
 238
 239 static void smc_lgr_free_work(struct work_struct *work)
 240 {
 241         struct smc_link_group *lgr = container_of(to_delayed_work(work),
 242                                                   struct smc_link_group,
 243                                                   free_work);
 244         spinlock_t *lgr_lock;
 245         bool conns;
 246
 247         smc_lgr_list_head(lgr, &lgr_lock);
 248         spin_lock_bh(lgr_lock);
 249         if (lgr->freeing) {
 250                 spin_unlock_bh(lgr_lock);
 251                 return;
 252         }
 253         read_lock_bh(&lgr->conns_lock);
 254         conns = RB_EMPTY_ROOT(&lgr->conns_all);
 255         read_unlock_bh(&lgr->conns_lock);
 256         if (!conns) { /* number of lgr connections is no longer zero */
 257                 spin_unlock_bh(lgr_lock);
 258                 return;
 259         }
 260         list_del_init(&lgr->list); /* remove from smc_lgr_list */
 261         lgr->freeing = 1; /* this instance does the freeing, no new schedule */
 262         spin_unlock_bh(lgr_lock);
 263         cancel_delayed_work(&lgr->free_work);
 264
 265         if (!lgr->is_smcd && !lgr->terminating)
 266                 smc_llc_send_link_delete_all(lgr, true,
 267                                              SMC_LLC_DEL_PROG_INIT_TERM);
 268         if (lgr->is_smcd && !lgr->terminating)
 269                 smc_ism_signal_shutdown(lgr);
 270         if (!lgr->is_smcd)
 271                 smcr_lgr_link_deactivate_all(lgr);
 272         smc_lgr_free(lgr);
 273 }
 274
 275 static void smc_lgr_terminate_work(struct work_struct *work)
 276 {
 277         struct smc_link_group *lgr = container_of(work, struct smc_link_group,
 278                                                   terminate_work);
 279
 280         __smc_lgr_terminate(lgr, true);
 281 }
 282
 283 /* return next unique link id for the lgr */
 284 static u8 smcr_next_link_id(struct smc_link_group *lgr)
 285 {
 286         u8 link_id;
 287         int i;
 288
 289         while (1) {
 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                                 continue;
 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) {
 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                         smcr_link_down_cond_sched(to_lnk);
 587                         return NULL;
 588                 }
 589                 /* avoid race with smcr_tx_sndbuf_nonempty() */
 590                 spin_lock_bh(&conn->send_lock);
 591                 conn->lnk = to_lnk;
 592                 rc = smc_switch_cursor(smc, pend, wr_buf);
 593                 spin_unlock_bh(&conn->send_lock);
 594                 sock_put(&smc->sk);
 595                 if (rc) {
 596                         smcr_link_down_cond_sched(to_lnk);
 597                         return NULL;
 598                 }
 599                 goto again;
 600         }
 601         read_unlock_bh(&lgr->conns_lock);
 602         return to_lnk;
 603 }
 604
 605 static void smcr_buf_unuse(struct smc_buf_desc *rmb_desc,
 606                            struct smc_link_group *lgr)
 607 {
 608         int rc;
 609
 610         if (rmb_desc->is_conf_rkey && !list_empty(&lgr->list)) {
 611                 /* unregister rmb with peer */
 612                 rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY);
 613                 if (!rc) {
 614                         /* protect against smc_llc_cli_rkey_exchange() */
 615                         mutex_lock(&lgr->llc_conf_mutex);
 616                         smc_llc_do_delete_rkey(lgr, rmb_desc);
 617                         rmb_desc->is_conf_rkey = false;
 618                         mutex_unlock(&lgr->llc_conf_mutex);
 619                         smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl);
 620                 }
 621         }
 622
 623         if (rmb_desc->is_reg_err) {
 624                 /* buf registration failed, reuse not possible */
 625                 mutex_lock(&lgr->rmbs_lock);
 626                 list_del(&rmb_desc->list);
 627                 mutex_unlock(&lgr->rmbs_lock);
 628
 629                 smc_buf_free(lgr, true, rmb_desc);
 630         } else {
 631                 rmb_desc->used = 0;
 632         }
 633 }
 634
 635 static void smc_buf_unuse(struct smc_connection *conn,
 636                           struct smc_link_group *lgr)
 637 {
 638         if (conn->sndbuf_desc)
 639                 conn->sndbuf_desc->used = 0;
 640         if (conn->rmb_desc && lgr->is_smcd)
 641                 conn->rmb_desc->used = 0;
 642         else if (conn->rmb_desc)
 643                 smcr_buf_unuse(conn->rmb_desc, lgr);
 644 }
 645
 646 /* remove a finished connection from its link group */
 647 void smc_conn_free(struct smc_connection *conn)
 648 {
 649         struct smc_link_group *lgr = conn->lgr;
 650
 651         if (!lgr)
 652                 return;
 653         if (lgr->is_smcd) {
 654                 if (!list_empty(&lgr->list))
 655                         smc_ism_unset_conn(conn);
 656                 tasklet_kill(&conn->rx_tsklet);
 657         } else {
 658                 smc_cdc_tx_dismiss_slots(conn);
 659                 if (current_work() != &conn->abort_work)
 660                         cancel_work_sync(&conn->abort_work);
 661         }
 662         if (!list_empty(&lgr->list)) {
 663                 smc_lgr_unregister_conn(conn);
 664                 smc_buf_unuse(conn, lgr); /* allow buffer reuse */
 665         }
 666
 667         if (!lgr->conns_num)
 668                 smc_lgr_schedule_free_work(lgr);
 669 }
 670
 671 /* unregister a link from a buf_desc */
 672 static void smcr_buf_unmap_link(struct smc_buf_desc *buf_desc, bool is_rmb,
 673                                 struct smc_link *lnk)
 674 {
 675         if (is_rmb)
 676                 buf_desc->is_reg_mr[lnk->link_idx] = false;
 677         if (!buf_desc->is_map_ib[lnk->link_idx])
 678                 return;
 679         if (is_rmb) {
 680                 if (buf_desc->mr_rx[lnk->link_idx]) {
 681                         smc_ib_put_memory_region(
 682                                         buf_desc->mr_rx[lnk->link_idx]);
 683                         buf_desc->mr_rx[lnk->link_idx] = NULL;
 684                 }
 685                 smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_FROM_DEVICE);
 686         } else {
 687                 smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_TO_DEVICE);
 688         }
 689         sg_free_table(&buf_desc->sgt[lnk->link_idx]);
 690         buf_desc->is_map_ib[lnk->link_idx] = false;
 691 }
 692
 693 /* unmap all buffers of lgr for a deleted link */
 694 static void smcr_buf_unmap_lgr(struct smc_link *lnk)
 695 {
 696         struct smc_link_group *lgr = lnk->lgr;
 697         struct smc_buf_desc *buf_desc, *bf;
 698         int i;
 699
 700         for (i = 0; i < SMC_RMBE_SIZES; i++) {
 701                 mutex_lock(&lgr->rmbs_lock);
 702                 list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list)
 703                         smcr_buf_unmap_link(buf_desc, true, lnk);
 704                 mutex_unlock(&lgr->rmbs_lock);
 705                 mutex_lock(&lgr->sndbufs_lock);
 706                 list_for_each_entry_safe(buf_desc, bf, &lgr->sndbufs[i],
 707                                          list)
 708                         smcr_buf_unmap_link(buf_desc, false, lnk);
 709                 mutex_unlock(&lgr->sndbufs_lock);
 710         }
 711 }
 712
 713 static void smcr_rtoken_clear_link(struct smc_link *lnk)
 714 {
 715         struct smc_link_group *lgr = lnk->lgr;
 716         int i;
 717
 718         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
 719                 lgr->rtokens[i][lnk->link_idx].rkey = 0;
 720                 lgr->rtokens[i][lnk->link_idx].dma_addr = 0;
 721         }
 722 }
 723
 724 /* must be called under lgr->llc_conf_mutex lock */
 725 void smcr_link_clear(struct smc_link *lnk, bool log)
 726 {
 727         struct smc_ib_device *smcibdev;
 728
 729         if (!lnk->lgr || lnk->state == SMC_LNK_UNUSED)
 730                 return;
 731         lnk->peer_qpn = 0;
 732         smc_llc_link_clear(lnk, log);
 733         smcr_buf_unmap_lgr(lnk);
 734         smcr_rtoken_clear_link(lnk);
 735         smc_ib_modify_qp_reset(lnk);
 736         smc_wr_free_link(lnk);
 737         smc_ib_destroy_queue_pair(lnk);
 738         smc_ib_dealloc_protection_domain(lnk);
 739         smc_wr_free_link_mem(lnk);
 740         put_device(&lnk->smcibdev->ibdev->dev);
 741         smcibdev = lnk->smcibdev;
 742         memset(lnk, 0, sizeof(struct smc_link));
 743         lnk->state = SMC_LNK_UNUSED;
 744         if (!atomic_dec_return(&smcibdev->lnk_cnt))
 745                 wake_up(&smcibdev->lnks_deleted);
 746 }
 747
 748 static void smcr_buf_free(struct smc_link_group *lgr, bool is_rmb,
 749                           struct smc_buf_desc *buf_desc)
 750 {
 751         int i;
 752
 753         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
 754                 smcr_buf_unmap_link(buf_desc, is_rmb, &lgr->lnk[i]);
 755
 756         if (buf_desc->pages)
 757                 __free_pages(buf_desc->pages, buf_desc->order);
 758         kfree(buf_desc);
 759 }
 760
 761 static void smcd_buf_free(struct smc_link_group *lgr, bool is_dmb,
 762                           struct smc_buf_desc *buf_desc)
 763 {
 764         if (is_dmb) {
 765                 /* restore original buf len */
 766                 buf_desc->len += sizeof(struct smcd_cdc_msg);
 767                 smc_ism_unregister_dmb(lgr->smcd, buf_desc);
 768         } else {
 769                 kfree(buf_desc->cpu_addr);
 770         }
 771         kfree(buf_desc);
 772 }
 773
 774 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
 775                          struct smc_buf_desc *buf_desc)
 776 {
 777         if (lgr->is_smcd)
 778                 smcd_buf_free(lgr, is_rmb, buf_desc);
 779         else
 780                 smcr_buf_free(lgr, is_rmb, buf_desc);
 781 }
 782
 783 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
 784 {
 785         struct smc_buf_desc *buf_desc, *bf_desc;
 786         struct list_head *buf_list;
 787         int i;
 788
 789         for (i = 0; i < SMC_RMBE_SIZES; i++) {
 790                 if (is_rmb)
 791                         buf_list = &lgr->rmbs[i];
 792                 else
 793                         buf_list = &lgr->sndbufs[i];
 794                 list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
 795                                          list) {
 796                         list_del(&buf_desc->list);
 797                         smc_buf_free(lgr, is_rmb, buf_desc);
 798                 }
 799         }
 800 }
 801
 802 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
 803 {
 804         /* free send buffers */
 805         __smc_lgr_free_bufs(lgr, false);
 806         /* free rmbs */
 807         __smc_lgr_free_bufs(lgr, true);
 808 }
 809
 810 /* remove a link group */
 811 static void smc_lgr_free(struct smc_link_group *lgr)
 812 {
 813         int i;
 814
 815         if (!lgr->is_smcd) {
 816                 mutex_lock(&lgr->llc_conf_mutex);
 817                 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 818                         if (lgr->lnk[i].state != SMC_LNK_UNUSED)
 819                                 smcr_link_clear(&lgr->lnk[i], false);
 820                 }
 821                 mutex_unlock(&lgr->llc_conf_mutex);
 822                 smc_llc_lgr_clear(lgr);
 823         }
 824
 825         smc_lgr_free_bufs(lgr);
 826         destroy_workqueue(lgr->tx_wq);
 827         if (lgr->is_smcd) {
 828                 smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
 829                 put_device(&lgr->smcd->dev);
 830                 if (!atomic_dec_return(&lgr->smcd->lgr_cnt))
 831                         wake_up(&lgr->smcd->lgrs_deleted);
 832         } else {
 833                 if (!atomic_dec_return(&lgr_cnt))
 834                         wake_up(&lgrs_deleted);
 835         }
 836         kfree(lgr);
 837 }
 838
 839 static void smcd_unregister_all_dmbs(struct smc_link_group *lgr)
 840 {
 841         int i;
 842
 843         for (i = 0; i < SMC_RMBE_SIZES; i++) {
 844                 struct smc_buf_desc *buf_desc;
 845
 846                 list_for_each_entry(buf_desc, &lgr->rmbs[i], list) {
 847                         buf_desc->len += sizeof(struct smcd_cdc_msg);
 848                         smc_ism_unregister_dmb(lgr->smcd, buf_desc);
 849                 }
 850         }
 851 }
 852
 853 static void smc_sk_wake_ups(struct smc_sock *smc)
 854 {
 855         smc->sk.sk_write_space(&smc->sk);
 856         smc->sk.sk_data_ready(&smc->sk);
 857         smc->sk.sk_state_change(&smc->sk);
 858 }
 859
 860 /* kill a connection */
 861 static void smc_conn_kill(struct smc_connection *conn, bool soft)
 862 {
 863         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
 864
 865         if (conn->lgr->is_smcd && conn->lgr->peer_shutdown)
 866                 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
 867         else
 868                 smc_close_abort(conn);
 869         conn->killed = 1;
 870         smc->sk.sk_err = ECONNABORTED;
 871         smc_sk_wake_ups(smc);
 872         if (conn->lgr->is_smcd) {
 873                 smc_ism_unset_conn(conn);
 874                 if (soft)
 875                         tasklet_kill(&conn->rx_tsklet);
 876                 else
 877                         tasklet_unlock_wait(&conn->rx_tsklet);
 878         } else {
 879                 smc_cdc_tx_dismiss_slots(conn);
 880         }
 881         smc_lgr_unregister_conn(conn);
 882         smc_close_active_abort(smc);
 883 }
 884
 885 static void smc_lgr_cleanup(struct smc_link_group *lgr)
 886 {
 887         if (lgr->is_smcd) {
 888                 smc_ism_signal_shutdown(lgr);
 889                 smcd_unregister_all_dmbs(lgr);
 890         } else {
 891                 u32 rsn = lgr->llc_termination_rsn;
 892
 893                 if (!rsn)
 894                         rsn = SMC_LLC_DEL_PROG_INIT_TERM;
 895                 smc_llc_send_link_delete_all(lgr, false, rsn);
 896                 smcr_lgr_link_deactivate_all(lgr);
 897         }
 898 }
 899
 900 /* terminate link group
 901  * @soft: true if link group shutdown can take its time
 902  *        false if immediate link group shutdown is required
 903  */
 904 static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft)
 905 {
 906         struct smc_connection *conn;
 907         struct smc_sock *smc;
 908         struct rb_node *node;
 909
 910         if (lgr->terminating)
 911                 return; /* lgr already terminating */
 912         /* cancel free_work sync, will terminate when lgr->freeing is set */
 913         cancel_delayed_work_sync(&lgr->free_work);
 914         lgr->terminating = 1;
 915
 916         /* kill remaining link group connections */
 917         read_lock_bh(&lgr->conns_lock);
 918         node = rb_first(&lgr->conns_all);
 919         while (node) {
 920                 read_unlock_bh(&lgr->conns_lock);
 921                 conn = rb_entry(node, struct smc_connection, alert_node);
 922                 smc = container_of(conn, struct smc_sock, conn);
 923                 sock_hold(&smc->sk); /* sock_put below */
 924                 lock_sock(&smc->sk);
 925                 smc_conn_kill(conn, soft);
 926                 release_sock(&smc->sk);
 927                 sock_put(&smc->sk); /* sock_hold above */
 928                 read_lock_bh(&lgr->conns_lock);
 929                 node = rb_first(&lgr->conns_all);
 930         }
 931         read_unlock_bh(&lgr->conns_lock);
 932         smc_lgr_cleanup(lgr);
 933         smc_lgr_free(lgr);
 934 }
 935
 936 /* unlink link group and schedule termination */
 937 void smc_lgr_terminate_sched(struct smc_link_group *lgr)
 938 {
 939         spinlock_t *lgr_lock;
 940
 941         smc_lgr_list_head(lgr, &lgr_lock);
 942         spin_lock_bh(lgr_lock);
 943         if (list_empty(&lgr->list) || lgr->terminating || lgr->freeing) {
 944                 spin_unlock_bh(lgr_lock);
 945                 return; /* lgr already terminating */
 946         }
 947         list_del_init(&lgr->list);
 948         lgr->freeing = 1;
 949         spin_unlock_bh(lgr_lock);
 950         schedule_work(&lgr->terminate_work);
 951 }
 952
 953 /* Called when peer lgr shutdown (regularly or abnormally) is received */
 954 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid, unsigned short vlan)
 955 {
 956         struct smc_link_group *lgr, *l;
 957         LIST_HEAD(lgr_free_list);
 958
 959         /* run common cleanup function and build free list */
 960         spin_lock_bh(&dev->lgr_lock);
 961         list_for_each_entry_safe(lgr, l, &dev->lgr_list, list) {
 962                 if ((!peer_gid || lgr->peer_gid == peer_gid) &&
 963                     (vlan == VLAN_VID_MASK || lgr->vlan_id == vlan)) {
 964                         if (peer_gid) /* peer triggered termination */
 965                                 lgr->peer_shutdown = 1;
 966                         list_move(&lgr->list, &lgr_free_list);
 967                         lgr->freeing = 1;
 968                 }
 969         }
 970         spin_unlock_bh(&dev->lgr_lock);
 971
 972         /* cancel the regular free workers and actually free lgrs */
 973         list_for_each_entry_safe(lgr, l, &lgr_free_list, list) {
 974                 list_del_init(&lgr->list);
 975                 schedule_work(&lgr->terminate_work);
 976         }
 977 }
 978
 979 /* Called when an SMCD device is removed or the smc module is unloaded */
 980 void smc_smcd_terminate_all(struct smcd_dev *smcd)
 981 {
 982         struct smc_link_group *lgr, *lg;
 983         LIST_HEAD(lgr_free_list);
 984
 985         spin_lock_bh(&smcd->lgr_lock);
 986         list_splice_init(&smcd->lgr_list, &lgr_free_list);
 987         list_for_each_entry(lgr, &lgr_free_list, list)
 988                 lgr->freeing = 1;
 989         spin_unlock_bh(&smcd->lgr_lock);
 990
 991         list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
 992                 list_del_init(&lgr->list);
 993                 __smc_lgr_terminate(lgr, false);
 994         }
 995
 996         if (atomic_read(&smcd->lgr_cnt))
 997                 wait_event(smcd->lgrs_deleted, !atomic_read(&smcd->lgr_cnt));
 998 }
 999
1000 /* Called when an SMCR device is removed or the smc module is unloaded.
1001  * If smcibdev is given, all SMCR link groups using this device are terminated.
1002  * If smcibdev is NULL, all SMCR link groups are terminated.
1003  */
1004 void smc_smcr_terminate_all(struct smc_ib_device *smcibdev)
1005 {
1006         struct smc_link_group *lgr, *lg;
1007         LIST_HEAD(lgr_free_list);
1008         int i;
1009
1010         spin_lock_bh(&smc_lgr_list.lock);
1011         if (!smcibdev) {
1012                 list_splice_init(&smc_lgr_list.list, &lgr_free_list);
1013                 list_for_each_entry(lgr, &lgr_free_list, list)
1014                         lgr->freeing = 1;
1015         } else {
1016                 list_for_each_entry_safe(lgr, lg, &smc_lgr_list.list, list) {
1017                         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1018                                 if (lgr->lnk[i].smcibdev == smcibdev)
1019                                         smcr_link_down_cond_sched(&lgr->lnk[i]);
1020                         }
1021                 }
1022         }
1023         spin_unlock_bh(&smc_lgr_list.lock);
1024
1025         list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
1026                 list_del_init(&lgr->list);
1027                 smc_llc_set_termination_rsn(lgr, SMC_LLC_DEL_OP_INIT_TERM);
1028                 __smc_lgr_terminate(lgr, false);
1029         }
1030
1031         if (smcibdev) {
1032                 if (atomic_read(&smcibdev->lnk_cnt))
1033                         wait_event(smcibdev->lnks_deleted,
1034                                    !atomic_read(&smcibdev->lnk_cnt));
1035         } else {
1036                 if (atomic_read(&lgr_cnt))
1037                         wait_event(lgrs_deleted, !atomic_read(&lgr_cnt));
1038         }
1039 }
1040
1041 /* set new lgr type and clear all asymmetric link tagging */
1042 void smcr_lgr_set_type(struct smc_link_group *lgr, enum smc_lgr_type new_type)
1043 {
1044         char *lgr_type = "";
1045         int i;
1046
1047         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
1048                 if (smc_link_usable(&lgr->lnk[i]))
1049                         lgr->lnk[i].link_is_asym = false;
1050         if (lgr->type == new_type)
1051                 return;
1052         lgr->type = new_type;
1053
1054         switch (lgr->type) {
1055         case SMC_LGR_NONE:
1056                 lgr_type = "NONE";
1057                 break;
1058         case SMC_LGR_SINGLE:
1059                 lgr_type = "SINGLE";
1060                 break;
1061         case SMC_LGR_SYMMETRIC:
1062                 lgr_type = "SYMMETRIC";
1063                 break;
1064         case SMC_LGR_ASYMMETRIC_PEER:
1065                 lgr_type = "ASYMMETRIC_PEER";
1066                 break;
1067         case SMC_LGR_ASYMMETRIC_LOCAL:
1068                 lgr_type = "ASYMMETRIC_LOCAL";
1069                 break;
1070         }
1071         pr_warn_ratelimited("smc: SMC-R lg %*phN state changed: "
1072                             "%s, pnetid %.16s\n", SMC_LGR_ID_SIZE, &lgr->id,
1073                             lgr_type, lgr->pnet_id);
1074 }
1075
1076 /* set new lgr type and tag a link as asymmetric */
1077 void smcr_lgr_set_type_asym(struct smc_link_group *lgr,
1078                             enum smc_lgr_type new_type, int asym_lnk_idx)
1079 {
1080         smcr_lgr_set_type(lgr, new_type);
1081         lgr->lnk[asym_lnk_idx].link_is_asym = true;
1082 }
1083
1084 /* abort connection, abort_work scheduled from tasklet context */
1085 static void smc_conn_abort_work(struct work_struct *work)
1086 {
1087         struct smc_connection *conn = container_of(work,
1088                                                    struct smc_connection,
1089                                                    abort_work);
1090         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
1091
1092         smc_conn_kill(conn, true);
1093         sock_put(&smc->sk); /* sock_hold done by schedulers of abort_work */
1094 }
1095
1096 void smcr_port_add(struct smc_ib_device *smcibdev, u8 ibport)
1097 {
1098         struct smc_link_group *lgr, *n;
1099
1100         list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1101                 struct smc_link *link;
1102
1103                 if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1104                             SMC_MAX_PNETID_LEN) ||
1105                     lgr->type == SMC_LGR_SYMMETRIC ||
1106                     lgr->type == SMC_LGR_ASYMMETRIC_PEER)
1107                         continue;
1108
1109                 /* trigger local add link processing */
1110                 link = smc_llc_usable_link(lgr);
1111                 if (link)
1112                         smc_llc_add_link_local(link);
1113         }
1114 }
1115
1116 /* link is down - switch connections to alternate link,
1117  * must be called under lgr->llc_conf_mutex lock
1118  */
1119 static void smcr_link_down(struct smc_link *lnk)
1120 {
1121         struct smc_link_group *lgr = lnk->lgr;
1122         struct smc_link *to_lnk;
1123         int del_link_id;
1124
1125         if (!lgr || lnk->state == SMC_LNK_UNUSED || list_empty(&lgr->list))
1126                 return;
1127
1128         smc_ib_modify_qp_reset(lnk);
1129         to_lnk = smc_switch_conns(lgr, lnk, true);
1130         if (!to_lnk) { /* no backup link available */
1131                 smcr_link_clear(lnk, true);
1132                 return;
1133         }
1134         smcr_lgr_set_type(lgr, SMC_LGR_SINGLE);
1135         del_link_id = lnk->link_id;
1136
1137         if (lgr->role == SMC_SERV) {
1138                 /* trigger local delete link processing */
1139                 smc_llc_srv_delete_link_local(to_lnk, del_link_id);
1140         } else {
1141                 if (lgr->llc_flow_lcl.type != SMC_LLC_FLOW_NONE) {
1142                         /* another llc task is ongoing */
1143                         mutex_unlock(&lgr->llc_conf_mutex);
1144                         wait_event_timeout(lgr->llc_flow_waiter,
1145                                 (list_empty(&lgr->list) ||
1146                                  lgr->llc_flow_lcl.type == SMC_LLC_FLOW_NONE),
1147                                 SMC_LLC_WAIT_TIME);
1148                         mutex_lock(&lgr->llc_conf_mutex);
1149                 }
1150                 if (!list_empty(&lgr->list)) {
1151                         smc_llc_send_delete_link(to_lnk, del_link_id,
1152                                                  SMC_LLC_REQ, true,
1153                                                  SMC_LLC_DEL_LOST_PATH);
1154                         smcr_link_clear(lnk, true);
1155                 }
1156                 wake_up(&lgr->llc_flow_waiter); /* wake up next waiter */
1157         }
1158 }
1159
1160 /* must be called under lgr->llc_conf_mutex lock */
1161 void smcr_link_down_cond(struct smc_link *lnk)
1162 {
1163         if (smc_link_downing(&lnk->state))
1164                 smcr_link_down(lnk);
1165 }
1166
1167 /* will get the lgr->llc_conf_mutex lock */
1168 void smcr_link_down_cond_sched(struct smc_link *lnk)
1169 {
1170         if (smc_link_downing(&lnk->state))
1171                 schedule_work(&lnk->link_down_wrk);
1172 }
1173
1174 void smcr_port_err(struct smc_ib_device *smcibdev, u8 ibport)
1175 {
1176         struct smc_link_group *lgr, *n;
1177         int i;
1178
1179         list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
1180                 if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
1181                             SMC_MAX_PNETID_LEN))
1182                         continue; /* lgr is not affected */
1183                 if (list_empty(&lgr->list))
1184                         continue;
1185                 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1186                         struct smc_link *lnk = &lgr->lnk[i];
1187
1188                         if (smc_link_usable(lnk) &&
1189                             lnk->smcibdev == smcibdev && lnk->ibport == ibport)
1190                                 smcr_link_down_cond_sched(lnk);
1191                 }
1192         }
1193 }
1194
1195 static void smc_link_down_work(struct work_struct *work)
1196 {
1197         struct smc_link *link = container_of(work, struct smc_link,
1198                                              link_down_wrk);
1199         struct smc_link_group *lgr = link->lgr;
1200
1201         if (list_empty(&lgr->list))
1202                 return;
1203         wake_up_all(&lgr->llc_msg_waiter);
1204         mutex_lock(&lgr->llc_conf_mutex);
1205         smcr_link_down(link);
1206         mutex_unlock(&lgr->llc_conf_mutex);
1207 }
1208
1209 /* Determine vlan of internal TCP socket.
1210  * @vlan_id: address to store the determined vlan id into
1211  */
1212 int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini)
1213 {
1214         struct dst_entry *dst = sk_dst_get(clcsock->sk);
1215         struct net_device *ndev;
1216         int i, nest_lvl, rc = 0;
1217
1218         ini->vlan_id = 0;
1219         if (!dst) {
1220                 rc = -ENOTCONN;
1221                 goto out;
1222         }
1223         if (!dst->dev) {
1224                 rc = -ENODEV;
1225                 goto out_rel;
1226         }
1227
1228         ndev = dst->dev;
1229         if (is_vlan_dev(ndev)) {
1230                 ini->vlan_id = vlan_dev_vlan_id(ndev);
1231                 goto out_rel;
1232         }
1233
1234         rtnl_lock();
1235         nest_lvl = ndev->lower_level;
1236         for (i = 0; i < nest_lvl; i++) {
1237                 struct list_head *lower = &ndev->adj_list.lower;
1238
1239                 if (list_empty(lower))
1240                         break;
1241                 lower = lower->next;
1242                 ndev = (struct net_device *)netdev_lower_get_next(ndev, &lower);
1243                 if (is_vlan_dev(ndev)) {
1244                         ini->vlan_id = vlan_dev_vlan_id(ndev);
1245                         break;
1246                 }
1247         }
1248         rtnl_unlock();
1249
1250 out_rel:
1251         dst_release(dst);
1252 out:
1253         return rc;
1254 }
1255
1256 static bool smcr_lgr_match(struct smc_link_group *lgr,
1257                            struct smc_clc_msg_local *lcl,
1258                            enum smc_lgr_role role, u32 clcqpn)
1259 {
1260         int i;
1261
1262         if (memcmp(lgr->peer_systemid, lcl->id_for_peer, SMC_SYSTEMID_LEN) ||
1263             lgr->role != role)
1264                 return false;
1265
1266         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1267                 if (!smc_link_active(&lgr->lnk[i]))
1268                         continue;
1269                 if ((lgr->role == SMC_SERV || lgr->lnk[i].peer_qpn == clcqpn) &&
1270                     !memcmp(lgr->lnk[i].peer_gid, &lcl->gid, SMC_GID_SIZE) &&
1271                     !memcmp(lgr->lnk[i].peer_mac, lcl->mac, sizeof(lcl->mac)))
1272                         return true;
1273         }
1274         return false;
1275 }
1276
1277 static bool smcd_lgr_match(struct smc_link_group *lgr,
1278                            struct smcd_dev *smcismdev, u64 peer_gid)
1279 {
1280         return lgr->peer_gid == peer_gid && lgr->smcd == smcismdev;
1281 }
1282
1283 /* create a new SMC connection (and a new link group if necessary) */
1284 int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini)
1285 {
1286         struct smc_connection *conn = &smc->conn;
1287         struct list_head *lgr_list;
1288         struct smc_link_group *lgr;
1289         enum smc_lgr_role role;
1290         spinlock_t *lgr_lock;
1291         int rc = 0;
1292
1293         lgr_list = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_list :
1294                                   &smc_lgr_list.list;
1295         lgr_lock = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_lock :
1296                                   &smc_lgr_list.lock;
1297         ini->first_contact_local = 1;
1298         role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
1299         if (role == SMC_CLNT && ini->first_contact_peer)
1300                 /* create new link group as well */
1301                 goto create;
1302
1303         /* determine if an existing link group can be reused */
1304         spin_lock_bh(lgr_lock);
1305         list_for_each_entry(lgr, lgr_list, list) {
1306                 write_lock_bh(&lgr->conns_lock);
1307                 if ((ini->is_smcd ?
1308                      smcd_lgr_match(lgr, ini->ism_dev[ini->ism_selected],
1309                                     ini->ism_peer_gid[ini->ism_selected]) :
1310                      smcr_lgr_match(lgr, ini->ib_lcl, role, ini->ib_clcqpn)) &&
1311                     !lgr->sync_err &&
1312                     (ini->smcd_version == SMC_V2 ||
1313                      lgr->vlan_id == ini->vlan_id) &&
1314                     (role == SMC_CLNT || ini->is_smcd ||
1315                      lgr->conns_num < SMC_RMBS_PER_LGR_MAX)) {
1316                         /* link group found */
1317                         ini->first_contact_local = 0;
1318                         conn->lgr = lgr;
1319                         rc = smc_lgr_register_conn(conn, false);
1320                         write_unlock_bh(&lgr->conns_lock);
1321                         if (!rc && delayed_work_pending(&lgr->free_work))
1322                                 cancel_delayed_work(&lgr->free_work);
1323                         break;
1324                 }
1325                 write_unlock_bh(&lgr->conns_lock);
1326         }
1327         spin_unlock_bh(lgr_lock);
1328         if (rc)
1329                 return rc;
1330
1331         if (role == SMC_CLNT && !ini->first_contact_peer &&
1332             ini->first_contact_local) {
1333                 /* Server reuses a link group, but Client wants to start
1334                  * a new one
1335                  * send out_of_sync decline, reason synchr. error
1336                  */
1337                 return SMC_CLC_DECL_SYNCERR;
1338         }
1339
1340 create:
1341         if (ini->first_contact_local) {
1342                 rc = smc_lgr_create(smc, ini);
1343                 if (rc)
1344                         goto out;
1345                 lgr = conn->lgr;
1346                 write_lock_bh(&lgr->conns_lock);
1347                 rc = smc_lgr_register_conn(conn, true);
1348                 write_unlock_bh(&lgr->conns_lock);
1349                 if (rc)
1350                         goto out;
1351         }
1352         conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
1353         conn->local_tx_ctrl.len = SMC_WR_TX_SIZE;
1354         conn->urg_state = SMC_URG_READ;
1355         INIT_WORK(&smc->conn.abort_work, smc_conn_abort_work);
1356         if (ini->is_smcd) {
1357                 conn->rx_off = sizeof(struct smcd_cdc_msg);
1358                 smcd_cdc_rx_init(conn); /* init tasklet for this conn */
1359         } else {
1360                 conn->rx_off = 0;
1361         }
1362 #ifndef KERNEL_HAS_ATOMIC64
1363         spin_lock_init(&conn->acurs_lock);
1364 #endif
1365
1366 out:
1367         return rc;
1368 }
1369
1370 /* convert the RMB size into the compressed notation - minimum 16K.
1371  * In contrast to plain ilog2, this rounds towards the next power of 2,
1372  * so the socket application gets at least its desired sndbuf / rcvbuf size.
1373  */
1374 static u8 smc_compress_bufsize(int size)
1375 {
1376         u8 compressed;
1377
1378         if (size <= SMC_BUF_MIN_SIZE)
1379                 return 0;
1380
1381         size = (size - 1) >> 14;
1382         compressed = ilog2(size) + 1;
1383         if (compressed >= SMC_RMBE_SIZES)
1384                 compressed = SMC_RMBE_SIZES - 1;
1385         return compressed;
1386 }
1387
1388 /* convert the RMB size from compressed notation into integer */
1389 int smc_uncompress_bufsize(u8 compressed)
1390 {
1391         u32 size;
1392
1393         size = 0x00000001 << (((int)compressed) + 14);
1394         return (int)size;
1395 }
1396
1397 /* try to reuse a sndbuf or rmb description slot for a certain
1398  * buffer size; if not available, return NULL
1399  */
1400 static struct smc_buf_desc *smc_buf_get_slot(int compressed_bufsize,
1401                                              struct mutex *lock,
1402                                              struct list_head *buf_list)
1403 {
1404         struct smc_buf_desc *buf_slot;
1405
1406         mutex_lock(lock);
1407         list_for_each_entry(buf_slot, buf_list, list) {
1408                 if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
1409                         mutex_unlock(lock);
1410                         return buf_slot;
1411                 }
1412         }
1413         mutex_unlock(lock);
1414         return NULL;
1415 }
1416
1417 /* one of the conditions for announcing a receiver's current window size is
1418  * that it "results in a minimum increase in the window size of 10% of the
1419  * receive buffer space" [RFC7609]
1420  */
1421 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
1422 {
1423         return min_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
1424 }
1425
1426 /* map an rmb buf to a link */
1427 static int smcr_buf_map_link(struct smc_buf_desc *buf_desc, bool is_rmb,
1428                              struct smc_link *lnk)
1429 {
1430         int rc;
1431
1432         if (buf_desc->is_map_ib[lnk->link_idx])
1433                 return 0;
1434
1435         rc = sg_alloc_table(&buf_desc->sgt[lnk->link_idx], 1, GFP_KERNEL);
1436         if (rc)
1437                 return rc;
1438         sg_set_buf(buf_desc->sgt[lnk->link_idx].sgl,
1439                    buf_desc->cpu_addr, buf_desc->len);
1440
1441         /* map sg table to DMA address */
1442         rc = smc_ib_buf_map_sg(lnk, buf_desc,
1443                                is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
1444         /* SMC protocol depends on mapping to one DMA address only */
1445         if (rc != 1) {
1446                 rc = -EAGAIN;
1447                 goto free_table;
1448         }
1449
1450         /* create a new memory region for the RMB */
1451         if (is_rmb) {
1452                 rc = smc_ib_get_memory_region(lnk->roce_pd,
1453                                               IB_ACCESS_REMOTE_WRITE |
1454                                               IB_ACCESS_LOCAL_WRITE,
1455                                               buf_desc, lnk->link_idx);
1456                 if (rc)
1457                         goto buf_unmap;
1458                 smc_ib_sync_sg_for_device(lnk, buf_desc, DMA_FROM_DEVICE);
1459         }
1460         buf_desc->is_map_ib[lnk->link_idx] = true;
1461         return 0;
1462
1463 buf_unmap:
1464         smc_ib_buf_unmap_sg(lnk, buf_desc,
1465                             is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
1466 free_table:
1467         sg_free_table(&buf_desc->sgt[lnk->link_idx]);
1468         return rc;
1469 }
1470
1471 /* register a new rmb on IB device,
1472  * must be called under lgr->llc_conf_mutex lock
1473  */
1474 int smcr_link_reg_rmb(struct smc_link *link, struct smc_buf_desc *rmb_desc)
1475 {
1476         if (list_empty(&link->lgr->list))
1477                 return -ENOLINK;
1478         if (!rmb_desc->is_reg_mr[link->link_idx]) {
1479                 /* register memory region for new rmb */
1480                 if (smc_wr_reg_send(link, rmb_desc->mr_rx[link->link_idx])) {
1481                         rmb_desc->is_reg_err = true;
1482                         return -EFAULT;
1483                 }
1484                 rmb_desc->is_reg_mr[link->link_idx] = true;
1485         }
1486         return 0;
1487 }
1488
1489 static int _smcr_buf_map_lgr(struct smc_link *lnk, struct mutex *lock,
1490                              struct list_head *lst, bool is_rmb)
1491 {
1492         struct smc_buf_desc *buf_desc, *bf;
1493         int rc = 0;
1494
1495         mutex_lock(lock);
1496         list_for_each_entry_safe(buf_desc, bf, lst, list) {
1497                 if (!buf_desc->used)
1498                         continue;
1499                 rc = smcr_buf_map_link(buf_desc, is_rmb, lnk);
1500                 if (rc)
1501                         goto out;
1502         }
1503 out:
1504         mutex_unlock(lock);
1505         return rc;
1506 }
1507
1508 /* map all used buffers of lgr for a new link */
1509 int smcr_buf_map_lgr(struct smc_link *lnk)
1510 {
1511         struct smc_link_group *lgr = lnk->lgr;
1512         int i, rc = 0;
1513
1514         for (i = 0; i < SMC_RMBE_SIZES; i++) {
1515                 rc = _smcr_buf_map_lgr(lnk, &lgr->rmbs_lock,
1516                                        &lgr->rmbs[i], true);
1517                 if (rc)
1518                         return rc;
1519                 rc = _smcr_buf_map_lgr(lnk, &lgr->sndbufs_lock,
1520                                        &lgr->sndbufs[i], false);
1521                 if (rc)
1522                         return rc;
1523         }
1524         return 0;
1525 }
1526
1527 /* register all used buffers of lgr for a new link,
1528  * must be called under lgr->llc_conf_mutex lock
1529  */
1530 int smcr_buf_reg_lgr(struct smc_link *lnk)
1531 {
1532         struct smc_link_group *lgr = lnk->lgr;
1533         struct smc_buf_desc *buf_desc, *bf;
1534         int i, rc = 0;
1535
1536         mutex_lock(&lgr->rmbs_lock);
1537         for (i = 0; i < SMC_RMBE_SIZES; i++) {
1538                 list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list) {
1539                         if (!buf_desc->used)
1540                                 continue;
1541                         rc = smcr_link_reg_rmb(lnk, buf_desc);
1542                         if (rc)
1543                                 goto out;
1544                 }
1545         }
1546 out:
1547         mutex_unlock(&lgr->rmbs_lock);
1548         return rc;
1549 }
1550
1551 static struct smc_buf_desc *smcr_new_buf_create(struct smc_link_group *lgr,
1552                                                 bool is_rmb, int bufsize)
1553 {
1554         struct smc_buf_desc *buf_desc;
1555
1556         /* try to alloc a new buffer */
1557         buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
1558         if (!buf_desc)
1559                 return ERR_PTR(-ENOMEM);
1560
1561         buf_desc->order = get_order(bufsize);
1562         buf_desc->pages = alloc_pages(GFP_KERNEL | __GFP_NOWARN |
1563                                       __GFP_NOMEMALLOC | __GFP_COMP |
1564                                       __GFP_NORETRY | __GFP_ZERO,
1565                                       buf_desc->order);
1566         if (!buf_desc->pages) {
1567                 kfree(buf_desc);
1568                 return ERR_PTR(-EAGAIN);
1569         }
1570         buf_desc->cpu_addr = (void *)page_address(buf_desc->pages);
1571         buf_desc->len = bufsize;
1572         return buf_desc;
1573 }
1574
1575 /* map buf_desc on all usable links,
1576  * unused buffers stay mapped as long as the link is up
1577  */
1578 static int smcr_buf_map_usable_links(struct smc_link_group *lgr,
1579                                      struct smc_buf_desc *buf_desc, bool is_rmb)
1580 {
1581         int i, rc = 0;
1582
1583         /* protect against parallel link reconfiguration */
1584         mutex_lock(&lgr->llc_conf_mutex);
1585         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1586                 struct smc_link *lnk = &lgr->lnk[i];
1587
1588                 if (!smc_link_usable(lnk))
1589                         continue;
1590                 if (smcr_buf_map_link(buf_desc, is_rmb, lnk)) {
1591                         rc = -ENOMEM;
1592                         goto out;
1593                 }
1594         }
1595 out:
1596         mutex_unlock(&lgr->llc_conf_mutex);
1597         return rc;
1598 }
1599
1600 #define SMCD_DMBE_SIZES         6 /* 0 -> 16KB, 1 -> 32KB, .. 6 -> 1MB */
1601
1602 static struct smc_buf_desc *smcd_new_buf_create(struct smc_link_group *lgr,
1603                                                 bool is_dmb, int bufsize)
1604 {
1605         struct smc_buf_desc *buf_desc;
1606         int rc;
1607
1608         if (smc_compress_bufsize(bufsize) > SMCD_DMBE_SIZES)
1609                 return ERR_PTR(-EAGAIN);
1610
1611         /* try to alloc a new DMB */
1612         buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
1613         if (!buf_desc)
1614                 return ERR_PTR(-ENOMEM);
1615         if (is_dmb) {
1616                 rc = smc_ism_register_dmb(lgr, bufsize, buf_desc);
1617                 if (rc) {
1618                         kfree(buf_desc);
1619                         if (rc == -ENOMEM)
1620                                 return ERR_PTR(-EAGAIN);
1621                         if (rc == -ENOSPC)
1622                                 return ERR_PTR(-ENOSPC);
1623                         return ERR_PTR(-EIO);
1624                 }
1625                 buf_desc->pages = virt_to_page(buf_desc->cpu_addr);
1626                 /* CDC header stored in buf. So, pretend it was smaller */
1627                 buf_desc->len = bufsize - sizeof(struct smcd_cdc_msg);
1628         } else {
1629                 buf_desc->cpu_addr = kzalloc(bufsize, GFP_KERNEL |
1630                                              __GFP_NOWARN | __GFP_NORETRY |
1631                                              __GFP_NOMEMALLOC);
1632                 if (!buf_desc->cpu_addr) {
1633                         kfree(buf_desc);
1634                         return ERR_PTR(-EAGAIN);
1635                 }
1636                 buf_desc->len = bufsize;
1637         }
1638         return buf_desc;
1639 }
1640
1641 static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
1642 {
1643         struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
1644         struct smc_connection *conn = &smc->conn;
1645         struct smc_link_group *lgr = conn->lgr;
1646         struct list_head *buf_list;
1647         int bufsize, bufsize_short;
1648         struct mutex *lock;     /* lock buffer list */
1649         int sk_buf_size;
1650
1651         if (is_rmb)
1652                 /* use socket recv buffer size (w/o overhead) as start value */
1653                 sk_buf_size = smc->sk.sk_rcvbuf / 2;
1654         else
1655                 /* use socket send buffer size (w/o overhead) as start value */
1656                 sk_buf_size = smc->sk.sk_sndbuf / 2;
1657
1658         for (bufsize_short = smc_compress_bufsize(sk_buf_size);
1659              bufsize_short >= 0; bufsize_short--) {
1660
1661                 if (is_rmb) {
1662                         lock = &lgr->rmbs_lock;
1663                         buf_list = &lgr->rmbs[bufsize_short];
1664                 } else {
1665                         lock = &lgr->sndbufs_lock;
1666                         buf_list = &lgr->sndbufs[bufsize_short];
1667                 }
1668                 bufsize = smc_uncompress_bufsize(bufsize_short);
1669                 if ((1 << get_order(bufsize)) > SG_MAX_SINGLE_ALLOC)
1670                         continue;
1671
1672                 /* check for reusable slot in the link group */
1673                 buf_desc = smc_buf_get_slot(bufsize_short, lock, buf_list);
1674                 if (buf_desc) {
1675                         memset(buf_desc->cpu_addr, 0, bufsize);
1676                         break; /* found reusable slot */
1677                 }
1678
1679                 if (is_smcd)
1680                         buf_desc = smcd_new_buf_create(lgr, is_rmb, bufsize);
1681                 else
1682                         buf_desc = smcr_new_buf_create(lgr, is_rmb, bufsize);
1683
1684                 if (PTR_ERR(buf_desc) == -ENOMEM)
1685                         break;
1686                 if (IS_ERR(buf_desc))
1687                         continue;
1688
1689                 buf_desc->used = 1;
1690                 mutex_lock(lock);
1691                 list_add(&buf_desc->list, buf_list);
1692                 mutex_unlock(lock);
1693                 break; /* found */
1694         }
1695
1696         if (IS_ERR(buf_desc))
1697                 return PTR_ERR(buf_desc);
1698
1699         if (!is_smcd) {
1700                 if (smcr_buf_map_usable_links(lgr, buf_desc, is_rmb)) {
1701                         smcr_buf_unuse(buf_desc, lgr);
1702                         return -ENOMEM;
1703                 }
1704         }
1705
1706         if (is_rmb) {
1707                 conn->rmb_desc = buf_desc;
1708                 conn->rmbe_size_short = bufsize_short;
1709                 smc->sk.sk_rcvbuf = bufsize * 2;
1710                 atomic_set(&conn->bytes_to_rcv, 0);
1711                 conn->rmbe_update_limit =
1712                         smc_rmb_wnd_update_limit(buf_desc->len);
1713                 if (is_smcd)
1714                         smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */
1715         } else {
1716                 conn->sndbuf_desc = buf_desc;
1717                 smc->sk.sk_sndbuf = bufsize * 2;
1718                 atomic_set(&conn->sndbuf_space, bufsize);
1719         }
1720         return 0;
1721 }
1722
1723 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn)
1724 {
1725         if (!conn->lgr || conn->lgr->is_smcd || !smc_link_active(conn->lnk))
1726                 return;
1727         smc_ib_sync_sg_for_cpu(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE);
1728 }
1729
1730 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
1731 {
1732         if (!conn->lgr || conn->lgr->is_smcd || !smc_link_active(conn->lnk))
1733                 return;
1734         smc_ib_sync_sg_for_device(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE);
1735 }
1736
1737 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
1738 {
1739         int i;
1740
1741         if (!conn->lgr || conn->lgr->is_smcd)
1742                 return;
1743         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1744                 if (!smc_link_active(&conn->lgr->lnk[i]))
1745                         continue;
1746                 smc_ib_sync_sg_for_cpu(&conn->lgr->lnk[i], conn->rmb_desc,
1747                                        DMA_FROM_DEVICE);
1748         }
1749 }
1750
1751 void smc_rmb_sync_sg_for_device(struct smc_connection *conn)
1752 {
1753         int i;
1754
1755         if (!conn->lgr || conn->lgr->is_smcd)
1756                 return;
1757         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1758                 if (!smc_link_active(&conn->lgr->lnk[i]))
1759                         continue;
1760                 smc_ib_sync_sg_for_device(&conn->lgr->lnk[i], conn->rmb_desc,
1761                                           DMA_FROM_DEVICE);
1762         }
1763 }
1764
1765 /* create the send and receive buffer for an SMC socket;
1766  * receive buffers are called RMBs;
1767  * (even though the SMC protocol allows more than one RMB-element per RMB,
1768  * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
1769  * extra RMB for every connection in a link group
1770  */
1771 int smc_buf_create(struct smc_sock *smc, bool is_smcd)
1772 {
1773         int rc;
1774
1775         /* create send buffer */
1776         rc = __smc_buf_create(smc, is_smcd, false);
1777         if (rc)
1778                 return rc;
1779         /* create rmb */
1780         rc = __smc_buf_create(smc, is_smcd, true);
1781         if (rc) {
1782                 mutex_lock(&smc->conn.lgr->sndbufs_lock);
1783                 list_del(&smc->conn.sndbuf_desc->list);
1784                 mutex_unlock(&smc->conn.lgr->sndbufs_lock);
1785                 smc_buf_free(smc->conn.lgr, false, smc->conn.sndbuf_desc);
1786                 smc->conn.sndbuf_desc = NULL;
1787         }
1788         return rc;
1789 }
1790
1791 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
1792 {
1793         int i;
1794
1795         for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
1796                 if (!test_and_set_bit(i, lgr->rtokens_used_mask))
1797                         return i;
1798         }
1799         return -ENOSPC;
1800 }
1801
1802 static int smc_rtoken_find_by_link(struct smc_link_group *lgr, int lnk_idx,
1803                                    u32 rkey)
1804 {
1805         int i;
1806
1807         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1808                 if (test_bit(i, lgr->rtokens_used_mask) &&
1809                     lgr->rtokens[i][lnk_idx].rkey == rkey)
1810                         return i;
1811         }
1812         return -ENOENT;
1813 }
1814
1815 /* set rtoken for a new link to an existing rmb */
1816 void smc_rtoken_set(struct smc_link_group *lgr, int link_idx, int link_idx_new,
1817                     __be32 nw_rkey_known, __be64 nw_vaddr, __be32 nw_rkey)
1818 {
1819         int rtok_idx;
1820
1821         rtok_idx = smc_rtoken_find_by_link(lgr, link_idx, ntohl(nw_rkey_known));
1822         if (rtok_idx == -ENOENT)
1823                 return;
1824         lgr->rtokens[rtok_idx][link_idx_new].rkey = ntohl(nw_rkey);
1825         lgr->rtokens[rtok_idx][link_idx_new].dma_addr = be64_to_cpu(nw_vaddr);
1826 }
1827
1828 /* set rtoken for a new link whose link_id is given */
1829 void smc_rtoken_set2(struct smc_link_group *lgr, int rtok_idx, int link_id,
1830                      __be64 nw_vaddr, __be32 nw_rkey)
1831 {
1832         u64 dma_addr = be64_to_cpu(nw_vaddr);
1833         u32 rkey = ntohl(nw_rkey);
1834         bool found = false;
1835         int link_idx;
1836
1837         for (link_idx = 0; link_idx < SMC_LINKS_PER_LGR_MAX; link_idx++) {
1838                 if (lgr->lnk[link_idx].link_id == link_id) {
1839                         found = true;
1840                         break;
1841                 }
1842         }
1843         if (!found)
1844                 return;
1845         lgr->rtokens[rtok_idx][link_idx].rkey = rkey;
1846         lgr->rtokens[rtok_idx][link_idx].dma_addr = dma_addr;
1847 }
1848
1849 /* add a new rtoken from peer */
1850 int smc_rtoken_add(struct smc_link *lnk, __be64 nw_vaddr, __be32 nw_rkey)
1851 {
1852         struct smc_link_group *lgr = smc_get_lgr(lnk);
1853         u64 dma_addr = be64_to_cpu(nw_vaddr);
1854         u32 rkey = ntohl(nw_rkey);
1855         int i;
1856
1857         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1858                 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
1859                     lgr->rtokens[i][lnk->link_idx].dma_addr == dma_addr &&
1860                     test_bit(i, lgr->rtokens_used_mask)) {
1861                         /* already in list */
1862                         return i;
1863                 }
1864         }
1865         i = smc_rmb_reserve_rtoken_idx(lgr);
1866         if (i < 0)
1867                 return i;
1868         lgr->rtokens[i][lnk->link_idx].rkey = rkey;
1869         lgr->rtokens[i][lnk->link_idx].dma_addr = dma_addr;
1870         return i;
1871 }
1872
1873 /* delete an rtoken from all links */
1874 int smc_rtoken_delete(struct smc_link *lnk, __be32 nw_rkey)
1875 {
1876         struct smc_link_group *lgr = smc_get_lgr(lnk);
1877         u32 rkey = ntohl(nw_rkey);
1878         int i, j;
1879
1880         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1881                 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
1882                     test_bit(i, lgr->rtokens_used_mask)) {
1883                         for (j = 0; j < SMC_LINKS_PER_LGR_MAX; j++) {
1884                                 lgr->rtokens[i][j].rkey = 0;
1885                                 lgr->rtokens[i][j].dma_addr = 0;
1886                         }
1887                         clear_bit(i, lgr->rtokens_used_mask);
1888                         return 0;
1889                 }
1890         }
1891         return -ENOENT;
1892 }
1893
1894 /* save rkey and dma_addr received from peer during clc handshake */
1895 int smc_rmb_rtoken_handling(struct smc_connection *conn,
1896                             struct smc_link *lnk,
1897                             struct smc_clc_msg_accept_confirm *clc)
1898 {
1899         conn->rtoken_idx = smc_rtoken_add(lnk, clc->r0.rmb_dma_addr,
1900                                           clc->r0.rmb_rkey);
1901         if (conn->rtoken_idx < 0)
1902                 return conn->rtoken_idx;
1903         return 0;
1904 }
1905
1906 static void smc_core_going_away(void)
1907 {
1908         struct smc_ib_device *smcibdev;
1909         struct smcd_dev *smcd;
1910
1911         mutex_lock(&smc_ib_devices.mutex);
1912         list_for_each_entry(smcibdev, &smc_ib_devices.list, list) {
1913                 int i;
1914
1915                 for (i = 0; i < SMC_MAX_PORTS; i++)
1916                         set_bit(i, smcibdev->ports_going_away);
1917         }
1918         mutex_unlock(&smc_ib_devices.mutex);
1919
1920         mutex_lock(&smcd_dev_list.mutex);
1921         list_for_each_entry(smcd, &smcd_dev_list.list, list) {
1922                 smcd->going_away = 1;
1923         }
1924         mutex_unlock(&smcd_dev_list.mutex);
1925 }
1926
1927 /* Clean up all SMC link groups */
1928 static void smc_lgrs_shutdown(void)
1929 {
1930         struct smcd_dev *smcd;
1931
1932         smc_core_going_away();
1933
1934         smc_smcr_terminate_all(NULL);
1935
1936         mutex_lock(&smcd_dev_list.mutex);
1937         list_for_each_entry(smcd, &smcd_dev_list.list, list)
1938                 smc_smcd_terminate_all(smcd);
1939         mutex_unlock(&smcd_dev_list.mutex);
1940 }
1941
1942 static int smc_core_reboot_event(struct notifier_block *this,
1943                                  unsigned long event, void *ptr)
1944 {
1945         smc_lgrs_shutdown();
1946         smc_ib_unregister_client();
1947         return 0;
1948 }
1949
1950 static struct notifier_block smc_reboot_notifier = {
1951         .notifier_call = smc_core_reboot_event,
1952 };
1953
1954 int __init smc_core_init(void)
1955 {
1956         return register_reboot_notifier(&smc_reboot_notifier);
1957 }
1958
1959 /* Called (from smc_exit) when module is removed */
1960 void smc_core_exit(void)
1961 {
1962         unregister_reboot_notifier(&smc_reboot_notifier);
1963         smc_lgrs_shutdown();
1964 }