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 <net/tcp.h>
  19 #include <net/sock.h>
  20 #include <rdma/ib_verbs.h>
  21 #include <rdma/ib_cache.h>
  22
  23 #include "smc.h"
  24 #include "smc_clc.h"
  25 #include "smc_core.h"
  26 #include "smc_ib.h"
  27 #include "smc_wr.h"
  28 #include "smc_llc.h"
  29 #include "smc_cdc.h"
  30 #include "smc_close.h"
  31 #include "smc_ism.h"
  32
  33 #define SMC_LGR_NUM_INCR                256
  34 #define SMC_LGR_FREE_DELAY_SERV         (600 * HZ)
  35 #define SMC_LGR_FREE_DELAY_CLNT         (SMC_LGR_FREE_DELAY_SERV + 10 * HZ)
  36 #define SMC_LGR_FREE_DELAY_FAST         (8 * 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 /* return head of link group list and its lock for a given link group */
  52 static inline struct list_head *smc_lgr_list_head(struct smc_link_group *lgr,
  53                                                   spinlock_t **lgr_lock)
  54 {
  55         if (lgr->is_smcd) {
  56                 *lgr_lock = &lgr->smcd->lgr_lock;
  57                 return &lgr->smcd->lgr_list;
  58         }
  59
  60         *lgr_lock = &smc_lgr_list.lock;
  61         return &smc_lgr_list.list;
  62 }
  63
  64 static void smc_lgr_schedule_free_work(struct smc_link_group *lgr)
  65 {
  66         /* client link group creation always follows the server link group
  67          * creation. For client use a somewhat higher removal delay time,
  68          * otherwise there is a risk of out-of-sync link groups.
  69          */
  70         if (!lgr->freeing && !lgr->freefast) {
  71                 mod_delayed_work(system_wq, &lgr->free_work,
  72                                  (!lgr->is_smcd && lgr->role == SMC_CLNT) ?
  73                                                 SMC_LGR_FREE_DELAY_CLNT :
  74                                                 SMC_LGR_FREE_DELAY_SERV);
  75         }
  76 }
  77
  78 void smc_lgr_schedule_free_work_fast(struct smc_link_group *lgr)
  79 {
  80         if (!lgr->freeing && !lgr->freefast) {
  81                 lgr->freefast = 1;
  82                 mod_delayed_work(system_wq, &lgr->free_work,
  83                                  SMC_LGR_FREE_DELAY_FAST);
  84         }
  85 }
  86
  87 /* Register connection's alert token in our lookup structure.
  88  * To use rbtrees we have to implement our own insert core.
  89  * Requires @conns_lock
  90  * @smc         connection to register
  91  * Returns 0 on success, != otherwise.
  92  */
  93 static void smc_lgr_add_alert_token(struct smc_connection *conn)
  94 {
  95         struct rb_node **link, *parent = NULL;
  96         u32 token = conn->alert_token_local;
  97
  98         link = &conn->lgr->conns_all.rb_node;
  99         while (*link) {
 100                 struct smc_connection *cur = rb_entry(*link,
 101                                         struct smc_connection, alert_node);
 102
 103                 parent = *link;
 104                 if (cur->alert_token_local > token)
 105                         link = &parent->rb_left;
 106                 else
 107                         link = &parent->rb_right;
 108         }
 109         /* Put the new node there */
 110         rb_link_node(&conn->alert_node, parent, link);
 111         rb_insert_color(&conn->alert_node, &conn->lgr->conns_all);
 112 }
 113
 114 /* Register connection in link group by assigning an alert token
 115  * registered in a search tree.
 116  * Requires @conns_lock
 117  * Note that '0' is a reserved value and not assigned.
 118  */
 119 static int smc_lgr_register_conn(struct smc_connection *conn)
 120 {
 121         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
 122         static atomic_t nexttoken = ATOMIC_INIT(0);
 123
 124         /* find a new alert_token_local value not yet used by some connection
 125          * in this link group
 126          */
 127         sock_hold(&smc->sk); /* sock_put in smc_lgr_unregister_conn() */
 128         while (!conn->alert_token_local) {
 129                 conn->alert_token_local = atomic_inc_return(&nexttoken);
 130                 if (smc_lgr_find_conn(conn->alert_token_local, conn->lgr))
 131                         conn->alert_token_local = 0;
 132         }
 133         smc_lgr_add_alert_token(conn);
 134
 135         /* assign the new connection to a link */
 136         if (!conn->lgr->is_smcd) {
 137                 struct smc_link *lnk;
 138                 int i;
 139
 140                 /* tbd - link balancing */
 141                 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 142                         lnk = &conn->lgr->lnk[i];
 143                         if (lnk->state == SMC_LNK_ACTIVATING ||
 144                             lnk->state == SMC_LNK_ACTIVE)
 145                                 conn->lnk = lnk;
 146                 }
 147                 if (!conn->lnk)
 148                         return SMC_CLC_DECL_NOACTLINK;
 149         }
 150         conn->lgr->conns_num++;
 151         return 0;
 152 }
 153
 154 /* Unregister connection and reset the alert token of the given connection<
 155  */
 156 static void __smc_lgr_unregister_conn(struct smc_connection *conn)
 157 {
 158         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
 159         struct smc_link_group *lgr = conn->lgr;
 160
 161         rb_erase(&conn->alert_node, &lgr->conns_all);
 162         lgr->conns_num--;
 163         conn->alert_token_local = 0;
 164         sock_put(&smc->sk); /* sock_hold in smc_lgr_register_conn() */
 165 }
 166
 167 /* Unregister connection from lgr
 168  */
 169 static void smc_lgr_unregister_conn(struct smc_connection *conn)
 170 {
 171         struct smc_link_group *lgr = conn->lgr;
 172
 173         if (!lgr)
 174                 return;
 175         write_lock_bh(&lgr->conns_lock);
 176         if (conn->alert_token_local) {
 177                 __smc_lgr_unregister_conn(conn);
 178         }
 179         write_unlock_bh(&lgr->conns_lock);
 180         conn->lgr = NULL;
 181 }
 182
 183 void smc_lgr_cleanup_early(struct smc_connection *conn)
 184 {
 185         struct smc_link_group *lgr = conn->lgr;
 186
 187         if (!lgr)
 188                 return;
 189
 190         smc_conn_free(conn);
 191         smc_lgr_forget(lgr);
 192         smc_lgr_schedule_free_work_fast(lgr);
 193 }
 194
 195 /* Send delete link, either as client to request the initiation
 196  * of the DELETE LINK sequence from server; or as server to
 197  * initiate the delete processing. See smc_llc_rx_delete_link().
 198  */
 199 static int smcr_link_send_delete(struct smc_link *lnk, bool orderly)
 200 {
 201         if (lnk->state == SMC_LNK_ACTIVE &&
 202             !smc_llc_send_delete_link(lnk, SMC_LLC_REQ, orderly)) {
 203                 smc_llc_link_deleting(lnk);
 204                 return 0;
 205         }
 206         return -ENOTCONN;
 207 }
 208
 209 static void smc_lgr_free(struct smc_link_group *lgr);
 210
 211 static void smc_lgr_free_work(struct work_struct *work)
 212 {
 213         struct smc_link_group *lgr = container_of(to_delayed_work(work),
 214                                                   struct smc_link_group,
 215                                                   free_work);
 216         spinlock_t *lgr_lock;
 217         bool conns;
 218         int i;
 219
 220         smc_lgr_list_head(lgr, &lgr_lock);
 221         spin_lock_bh(lgr_lock);
 222         if (lgr->freeing) {
 223                 spin_unlock_bh(lgr_lock);
 224                 return;
 225         }
 226         read_lock_bh(&lgr->conns_lock);
 227         conns = RB_EMPTY_ROOT(&lgr->conns_all);
 228         read_unlock_bh(&lgr->conns_lock);
 229         if (!conns) { /* number of lgr connections is no longer zero */
 230                 spin_unlock_bh(lgr_lock);
 231                 return;
 232         }
 233         list_del_init(&lgr->list); /* remove from smc_lgr_list */
 234
 235         if (!lgr->is_smcd && !lgr->terminating) {
 236                 bool do_wait = false;
 237
 238                 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 239                         struct smc_link *lnk = &lgr->lnk[i];
 240                         /* try to send del link msg, on err free immediately */
 241                         if (lnk->state == SMC_LNK_ACTIVE &&
 242                             !smcr_link_send_delete(lnk, true)) {
 243                                 /* reschedule in case we never receive a resp */
 244                                 smc_lgr_schedule_free_work(lgr);
 245                                 do_wait = true;
 246                         }
 247                 }
 248                 if (do_wait) {
 249                         spin_unlock_bh(lgr_lock);
 250                         return; /* wait for resp, see smc_llc_rx_delete_link */
 251                 }
 252         }
 253         lgr->freeing = 1; /* this instance does the freeing, no new schedule */
 254         spin_unlock_bh(lgr_lock);
 255         cancel_delayed_work(&lgr->free_work);
 256
 257         if (lgr->is_smcd && !lgr->terminating)
 258                 smc_ism_signal_shutdown(lgr);
 259         if (!lgr->is_smcd) {
 260                 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 261                         struct smc_link *lnk = &lgr->lnk[i];
 262
 263                         if (smc_link_usable(lnk))
 264                                 lnk->state = SMC_LNK_INACTIVE;
 265                 }
 266         }
 267         smc_lgr_free(lgr);
 268 }
 269
 270 static void smc_lgr_terminate_work(struct work_struct *work)
 271 {
 272         struct smc_link_group *lgr = container_of(work, struct smc_link_group,
 273                                                   terminate_work);
 274
 275         __smc_lgr_terminate(lgr, true);
 276 }
 277
 278 /* return next unique link id for the lgr */
 279 static u8 smcr_next_link_id(struct smc_link_group *lgr)
 280 {
 281         u8 link_id;
 282         int i;
 283
 284         while (1) {
 285                 link_id = ++lgr->next_link_id;
 286                 if (!link_id)   /* skip zero as link_id */
 287                         link_id = ++lgr->next_link_id;
 288                 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 289                         if (smc_link_usable(&lgr->lnk[i]) &&
 290                             lgr->lnk[i].link_id == link_id)
 291                                 continue;
 292                 }
 293                 break;
 294         }
 295         return link_id;
 296 }
 297
 298 static int smcr_link_init(struct smc_link_group *lgr, struct smc_link *lnk,
 299                           u8 link_idx, struct smc_init_info *ini)
 300 {
 301         u8 rndvec[3];
 302         int rc;
 303
 304         get_device(&ini->ib_dev->ibdev->dev);
 305         atomic_inc(&ini->ib_dev->lnk_cnt);
 306         lnk->state = SMC_LNK_ACTIVATING;
 307         lnk->link_id = smcr_next_link_id(lgr);
 308         lnk->lgr = lgr;
 309         lnk->link_idx = link_idx;
 310         lnk->smcibdev = ini->ib_dev;
 311         lnk->ibport = ini->ib_port;
 312         lnk->path_mtu = ini->ib_dev->pattr[ini->ib_port - 1].active_mtu;
 313         if (!ini->ib_dev->initialized) {
 314                 rc = (int)smc_ib_setup_per_ibdev(ini->ib_dev);
 315                 if (rc)
 316                         goto out;
 317         }
 318         get_random_bytes(rndvec, sizeof(rndvec));
 319         lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) +
 320                 (rndvec[2] << 16);
 321         rc = smc_ib_determine_gid(lnk->smcibdev, lnk->ibport,
 322                                   ini->vlan_id, lnk->gid, &lnk->sgid_index);
 323         if (rc)
 324                 goto out;
 325         rc = smc_llc_link_init(lnk);
 326         if (rc)
 327                 goto out;
 328         rc = smc_wr_alloc_link_mem(lnk);
 329         if (rc)
 330                 goto clear_llc_lnk;
 331         rc = smc_ib_create_protection_domain(lnk);
 332         if (rc)
 333                 goto free_link_mem;
 334         rc = smc_ib_create_queue_pair(lnk);
 335         if (rc)
 336                 goto dealloc_pd;
 337         rc = smc_wr_create_link(lnk);
 338         if (rc)
 339                 goto destroy_qp;
 340         return 0;
 341
 342 destroy_qp:
 343         smc_ib_destroy_queue_pair(lnk);
 344 dealloc_pd:
 345         smc_ib_dealloc_protection_domain(lnk);
 346 free_link_mem:
 347         smc_wr_free_link_mem(lnk);
 348 clear_llc_lnk:
 349         smc_llc_link_clear(lnk);
 350 out:
 351         put_device(&ini->ib_dev->ibdev->dev);
 352         memset(lnk, 0, sizeof(struct smc_link));
 353         lnk->state = SMC_LNK_UNUSED;
 354         if (!atomic_dec_return(&ini->ib_dev->lnk_cnt))
 355                 wake_up(&ini->ib_dev->lnks_deleted);
 356         return rc;
 357 }
 358
 359 /* create a new SMC link group */
 360 static int smc_lgr_create(struct smc_sock *smc, struct smc_init_info *ini)
 361 {
 362         struct smc_link_group *lgr;
 363         struct list_head *lgr_list;
 364         struct smc_link *lnk;
 365         spinlock_t *lgr_lock;
 366         u8 link_idx;
 367         int rc = 0;
 368         int i;
 369
 370         if (ini->is_smcd && ini->vlan_id) {
 371                 if (smc_ism_get_vlan(ini->ism_dev, ini->vlan_id)) {
 372                         rc = SMC_CLC_DECL_ISMVLANERR;
 373                         goto out;
 374                 }
 375         }
 376
 377         lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
 378         if (!lgr) {
 379                 rc = SMC_CLC_DECL_MEM;
 380                 goto ism_put_vlan;
 381         }
 382         lgr->is_smcd = ini->is_smcd;
 383         lgr->sync_err = 0;
 384         lgr->terminating = 0;
 385         lgr->freefast = 0;
 386         lgr->freeing = 0;
 387         lgr->vlan_id = ini->vlan_id;
 388         mutex_init(&lgr->sndbufs_lock);
 389         mutex_init(&lgr->rmbs_lock);
 390         rwlock_init(&lgr->conns_lock);
 391         for (i = 0; i < SMC_RMBE_SIZES; i++) {
 392                 INIT_LIST_HEAD(&lgr->sndbufs[i]);
 393                 INIT_LIST_HEAD(&lgr->rmbs[i]);
 394         }
 395         lgr->next_link_id = 0;
 396         smc_lgr_list.num += SMC_LGR_NUM_INCR;
 397         memcpy(&lgr->id, (u8 *)&smc_lgr_list.num, SMC_LGR_ID_SIZE);
 398         INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
 399         INIT_WORK(&lgr->terminate_work, smc_lgr_terminate_work);
 400         lgr->conns_all = RB_ROOT;
 401         if (ini->is_smcd) {
 402                 /* SMC-D specific settings */
 403                 get_device(&ini->ism_dev->dev);
 404                 lgr->peer_gid = ini->ism_gid;
 405                 lgr->smcd = ini->ism_dev;
 406                 lgr_list = &ini->ism_dev->lgr_list;
 407                 lgr_lock = &lgr->smcd->lgr_lock;
 408                 lgr->peer_shutdown = 0;
 409                 atomic_inc(&ini->ism_dev->lgr_cnt);
 410         } else {
 411                 /* SMC-R specific settings */
 412                 lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
 413                 memcpy(lgr->peer_systemid, ini->ib_lcl->id_for_peer,
 414                        SMC_SYSTEMID_LEN);
 415                 smc_llc_lgr_init(lgr, smc);
 416
 417                 link_idx = SMC_SINGLE_LINK;
 418                 lnk = &lgr->lnk[link_idx];
 419                 rc = smcr_link_init(lgr, lnk, link_idx, ini);
 420                 if (rc)
 421                         goto free_lgr;
 422                 lgr_list = &smc_lgr_list.list;
 423                 lgr_lock = &smc_lgr_list.lock;
 424                 atomic_inc(&lgr_cnt);
 425         }
 426         smc->conn.lgr = lgr;
 427         spin_lock_bh(lgr_lock);
 428         list_add(&lgr->list, lgr_list);
 429         spin_unlock_bh(lgr_lock);
 430         return 0;
 431
 432 free_lgr:
 433         kfree(lgr);
 434 ism_put_vlan:
 435         if (ini->is_smcd && ini->vlan_id)
 436                 smc_ism_put_vlan(ini->ism_dev, ini->vlan_id);
 437 out:
 438         if (rc < 0) {
 439                 if (rc == -ENOMEM)
 440                         rc = SMC_CLC_DECL_MEM;
 441                 else
 442                         rc = SMC_CLC_DECL_INTERR;
 443         }
 444         return rc;
 445 }
 446
 447 static void smcr_buf_unuse(struct smc_buf_desc *rmb_desc,
 448                            struct smc_link *lnk)
 449 {
 450         struct smc_link_group *lgr = lnk->lgr;
 451
 452         if (rmb_desc->is_conf_rkey && !list_empty(&lgr->list)) {
 453                 /* unregister rmb with peer */
 454                 smc_llc_do_delete_rkey(lnk, rmb_desc);
 455                 rmb_desc->is_conf_rkey = false;
 456         }
 457         if (rmb_desc->is_reg_err) {
 458                 /* buf registration failed, reuse not possible */
 459                 mutex_lock(&lgr->rmbs_lock);
 460                 list_del(&rmb_desc->list);
 461                 mutex_unlock(&lgr->rmbs_lock);
 462
 463                 smc_buf_free(lgr, true, rmb_desc);
 464         } else {
 465                 rmb_desc->used = 0;
 466         }
 467 }
 468
 469 static void smc_buf_unuse(struct smc_connection *conn,
 470                           struct smc_link_group *lgr)
 471 {
 472         if (conn->sndbuf_desc)
 473                 conn->sndbuf_desc->used = 0;
 474         if (conn->rmb_desc && lgr->is_smcd)
 475                 conn->rmb_desc->used = 0;
 476         else if (conn->rmb_desc)
 477                 smcr_buf_unuse(conn->rmb_desc, conn->lnk);
 478 }
 479
 480 /* remove a finished connection from its link group */
 481 void smc_conn_free(struct smc_connection *conn)
 482 {
 483         struct smc_link_group *lgr = conn->lgr;
 484
 485         if (!lgr)
 486                 return;
 487         if (lgr->is_smcd) {
 488                 if (!list_empty(&lgr->list))
 489                         smc_ism_unset_conn(conn);
 490                 tasklet_kill(&conn->rx_tsklet);
 491         } else {
 492                 smc_cdc_tx_dismiss_slots(conn);
 493         }
 494         if (!list_empty(&lgr->list)) {
 495                 smc_lgr_unregister_conn(conn);
 496                 smc_buf_unuse(conn, lgr); /* allow buffer reuse */
 497         }
 498
 499         if (!lgr->conns_num)
 500                 smc_lgr_schedule_free_work(lgr);
 501 }
 502
 503 static void smcr_link_clear(struct smc_link *lnk)
 504 {
 505         struct smc_ib_device *smcibdev;
 506
 507         if (lnk->peer_qpn == 0)
 508                 return;
 509         lnk->peer_qpn = 0;
 510         smc_llc_link_clear(lnk);
 511         smc_ib_modify_qp_reset(lnk);
 512         smc_wr_free_link(lnk);
 513         smc_ib_destroy_queue_pair(lnk);
 514         smc_ib_dealloc_protection_domain(lnk);
 515         smc_wr_free_link_mem(lnk);
 516         put_device(&lnk->smcibdev->ibdev->dev);
 517         smcibdev = lnk->smcibdev;
 518         memset(lnk, 0, sizeof(struct smc_link));
 519         lnk->state = SMC_LNK_UNUSED;
 520         if (!atomic_dec_return(&smcibdev->lnk_cnt))
 521                 wake_up(&smcibdev->lnks_deleted);
 522 }
 523
 524 static void smcr_buf_free(struct smc_link_group *lgr, bool is_rmb,
 525                           struct smc_buf_desc *buf_desc)
 526 {
 527         struct smc_link *lnk;
 528         int i;
 529
 530         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 531                 lnk = &lgr->lnk[i];
 532                 if (!buf_desc->is_map_ib[lnk->link_idx])
 533                         continue;
 534                 if (is_rmb) {
 535                         if (buf_desc->mr_rx[lnk->link_idx])
 536                                 smc_ib_put_memory_region(
 537                                                 buf_desc->mr_rx[lnk->link_idx]);
 538                         smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_FROM_DEVICE);
 539                 } else {
 540                         smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_TO_DEVICE);
 541                 }
 542                 sg_free_table(&buf_desc->sgt[lnk->link_idx]);
 543         }
 544
 545         if (buf_desc->pages)
 546                 __free_pages(buf_desc->pages, buf_desc->order);
 547         kfree(buf_desc);
 548 }
 549
 550 static void smcd_buf_free(struct smc_link_group *lgr, bool is_dmb,
 551                           struct smc_buf_desc *buf_desc)
 552 {
 553         if (is_dmb) {
 554                 /* restore original buf len */
 555                 buf_desc->len += sizeof(struct smcd_cdc_msg);
 556                 smc_ism_unregister_dmb(lgr->smcd, buf_desc);
 557         } else {
 558                 kfree(buf_desc->cpu_addr);
 559         }
 560         kfree(buf_desc);
 561 }
 562
 563 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
 564                          struct smc_buf_desc *buf_desc)
 565 {
 566         if (lgr->is_smcd)
 567                 smcd_buf_free(lgr, is_rmb, buf_desc);
 568         else
 569                 smcr_buf_free(lgr, is_rmb, buf_desc);
 570 }
 571
 572 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
 573 {
 574         struct smc_buf_desc *buf_desc, *bf_desc;
 575         struct list_head *buf_list;
 576         int i;
 577
 578         for (i = 0; i < SMC_RMBE_SIZES; i++) {
 579                 if (is_rmb)
 580                         buf_list = &lgr->rmbs[i];
 581                 else
 582                         buf_list = &lgr->sndbufs[i];
 583                 list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
 584                                          list) {
 585                         list_del(&buf_desc->list);
 586                         smc_buf_free(lgr, is_rmb, buf_desc);
 587                 }
 588         }
 589 }
 590
 591 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
 592 {
 593         /* free send buffers */
 594         __smc_lgr_free_bufs(lgr, false);
 595         /* free rmbs */
 596         __smc_lgr_free_bufs(lgr, true);
 597 }
 598
 599 /* remove a link group */
 600 static void smc_lgr_free(struct smc_link_group *lgr)
 601 {
 602         int i;
 603
 604         smc_lgr_free_bufs(lgr);
 605         if (lgr->is_smcd) {
 606                 if (!lgr->terminating) {
 607                         smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
 608                         put_device(&lgr->smcd->dev);
 609                 }
 610                 if (!atomic_dec_return(&lgr->smcd->lgr_cnt))
 611                         wake_up(&lgr->smcd->lgrs_deleted);
 612         } else {
 613                 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 614                         if (lgr->lnk[i].state != SMC_LNK_UNUSED)
 615                                 smcr_link_clear(&lgr->lnk[i]);
 616                 }
 617                 smc_llc_lgr_clear(lgr);
 618                 if (!atomic_dec_return(&lgr_cnt))
 619                         wake_up(&lgrs_deleted);
 620         }
 621         kfree(lgr);
 622 }
 623
 624 void smc_lgr_forget(struct smc_link_group *lgr)
 625 {
 626         struct list_head *lgr_list;
 627         spinlock_t *lgr_lock;
 628
 629         lgr_list = smc_lgr_list_head(lgr, &lgr_lock);
 630         spin_lock_bh(lgr_lock);
 631         /* do not use this link group for new connections */
 632         if (!list_empty(lgr_list))
 633                 list_del_init(lgr_list);
 634         spin_unlock_bh(lgr_lock);
 635 }
 636
 637 static void smcd_unregister_all_dmbs(struct smc_link_group *lgr)
 638 {
 639         int i;
 640
 641         for (i = 0; i < SMC_RMBE_SIZES; i++) {
 642                 struct smc_buf_desc *buf_desc;
 643
 644                 list_for_each_entry(buf_desc, &lgr->rmbs[i], list) {
 645                         buf_desc->len += sizeof(struct smcd_cdc_msg);
 646                         smc_ism_unregister_dmb(lgr->smcd, buf_desc);
 647                 }
 648         }
 649 }
 650
 651 static void smc_sk_wake_ups(struct smc_sock *smc)
 652 {
 653         smc->sk.sk_write_space(&smc->sk);
 654         smc->sk.sk_data_ready(&smc->sk);
 655         smc->sk.sk_state_change(&smc->sk);
 656 }
 657
 658 /* kill a connection */
 659 static void smc_conn_kill(struct smc_connection *conn, bool soft)
 660 {
 661         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
 662
 663         if (conn->lgr->is_smcd && conn->lgr->peer_shutdown)
 664                 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
 665         else
 666                 smc_close_abort(conn);
 667         conn->killed = 1;
 668         smc->sk.sk_err = ECONNABORTED;
 669         smc_sk_wake_ups(smc);
 670         if (conn->lgr->is_smcd) {
 671                 smc_ism_unset_conn(conn);
 672                 if (soft)
 673                         tasklet_kill(&conn->rx_tsklet);
 674                 else
 675                         tasklet_unlock_wait(&conn->rx_tsklet);
 676         } else {
 677                 smc_cdc_tx_dismiss_slots(conn);
 678         }
 679         smc_lgr_unregister_conn(conn);
 680         smc_close_active_abort(smc);
 681 }
 682
 683 static void smc_lgr_cleanup(struct smc_link_group *lgr)
 684 {
 685         int i;
 686
 687         if (lgr->is_smcd) {
 688                 smc_ism_signal_shutdown(lgr);
 689                 smcd_unregister_all_dmbs(lgr);
 690                 smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
 691                 put_device(&lgr->smcd->dev);
 692         } else {
 693                 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 694                         struct smc_link *lnk = &lgr->lnk[i];
 695
 696                         if (smc_link_usable(lnk))
 697                                 lnk->state = SMC_LNK_INACTIVE;
 698                 }
 699         }
 700 }
 701
 702 /* terminate link group
 703  * @soft: true if link group shutdown can take its time
 704  *        false if immediate link group shutdown is required
 705  */
 706 static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft)
 707 {
 708         struct smc_connection *conn;
 709         struct smc_sock *smc;
 710         struct rb_node *node;
 711
 712         if (lgr->terminating)
 713                 return; /* lgr already terminating */
 714         if (!soft)
 715                 cancel_delayed_work_sync(&lgr->free_work);
 716         lgr->terminating = 1;
 717
 718         /* kill remaining link group connections */
 719         read_lock_bh(&lgr->conns_lock);
 720         node = rb_first(&lgr->conns_all);
 721         while (node) {
 722                 read_unlock_bh(&lgr->conns_lock);
 723                 conn = rb_entry(node, struct smc_connection, alert_node);
 724                 smc = container_of(conn, struct smc_sock, conn);
 725                 sock_hold(&smc->sk); /* sock_put below */
 726                 lock_sock(&smc->sk);
 727                 smc_conn_kill(conn, soft);
 728                 release_sock(&smc->sk);
 729                 sock_put(&smc->sk); /* sock_hold above */
 730                 read_lock_bh(&lgr->conns_lock);
 731                 node = rb_first(&lgr->conns_all);
 732         }
 733         read_unlock_bh(&lgr->conns_lock);
 734         smc_lgr_cleanup(lgr);
 735         if (soft)
 736                 smc_lgr_schedule_free_work_fast(lgr);
 737         else
 738                 smc_lgr_free(lgr);
 739 }
 740
 741 /* unlink link group and schedule termination */
 742 void smc_lgr_terminate_sched(struct smc_link_group *lgr)
 743 {
 744         spinlock_t *lgr_lock;
 745
 746         smc_lgr_list_head(lgr, &lgr_lock);
 747         spin_lock_bh(lgr_lock);
 748         if (list_empty(&lgr->list) || lgr->terminating || lgr->freeing) {
 749                 spin_unlock_bh(lgr_lock);
 750                 return; /* lgr already terminating */
 751         }
 752         list_del_init(&lgr->list);
 753         spin_unlock_bh(lgr_lock);
 754         schedule_work(&lgr->terminate_work);
 755 }
 756
 757 /* Called when IB port is terminated */
 758 void smc_port_terminate(struct smc_ib_device *smcibdev, u8 ibport)
 759 {
 760         struct smc_link_group *lgr, *l;
 761         LIST_HEAD(lgr_free_list);
 762         int i;
 763
 764         spin_lock_bh(&smc_lgr_list.lock);
 765         list_for_each_entry_safe(lgr, l, &smc_lgr_list.list, list) {
 766                 if (lgr->is_smcd)
 767                         continue;
 768                 /* tbd - terminate only when no more links are active */
 769                 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 770                         if (!smc_link_usable(&lgr->lnk[i]) ||
 771                             lgr->lnk[i].state == SMC_LNK_DELETING)
 772                                 continue;
 773                         if (lgr->lnk[i].smcibdev == smcibdev &&
 774                             lgr->lnk[i].ibport == ibport) {
 775                                 list_move(&lgr->list, &lgr_free_list);
 776                                 lgr->freeing = 1;
 777                         }
 778                 }
 779         }
 780         spin_unlock_bh(&smc_lgr_list.lock);
 781
 782         list_for_each_entry_safe(lgr, l, &lgr_free_list, list) {
 783                 list_del_init(&lgr->list);
 784                 __smc_lgr_terminate(lgr, false);
 785         }
 786 }
 787
 788 /* Called when peer lgr shutdown (regularly or abnormally) is received */
 789 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid, unsigned short vlan)
 790 {
 791         struct smc_link_group *lgr, *l;
 792         LIST_HEAD(lgr_free_list);
 793
 794         /* run common cleanup function and build free list */
 795         spin_lock_bh(&dev->lgr_lock);
 796         list_for_each_entry_safe(lgr, l, &dev->lgr_list, list) {
 797                 if ((!peer_gid || lgr->peer_gid == peer_gid) &&
 798                     (vlan == VLAN_VID_MASK || lgr->vlan_id == vlan)) {
 799                         if (peer_gid) /* peer triggered termination */
 800                                 lgr->peer_shutdown = 1;
 801                         list_move(&lgr->list, &lgr_free_list);
 802                 }
 803         }
 804         spin_unlock_bh(&dev->lgr_lock);
 805
 806         /* cancel the regular free workers and actually free lgrs */
 807         list_for_each_entry_safe(lgr, l, &lgr_free_list, list) {
 808                 list_del_init(&lgr->list);
 809                 schedule_work(&lgr->terminate_work);
 810         }
 811 }
 812
 813 /* Called when an SMCD device is removed or the smc module is unloaded */
 814 void smc_smcd_terminate_all(struct smcd_dev *smcd)
 815 {
 816         struct smc_link_group *lgr, *lg;
 817         LIST_HEAD(lgr_free_list);
 818
 819         spin_lock_bh(&smcd->lgr_lock);
 820         list_splice_init(&smcd->lgr_list, &lgr_free_list);
 821         list_for_each_entry(lgr, &lgr_free_list, list)
 822                 lgr->freeing = 1;
 823         spin_unlock_bh(&smcd->lgr_lock);
 824
 825         list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
 826                 list_del_init(&lgr->list);
 827                 __smc_lgr_terminate(lgr, false);
 828         }
 829
 830         if (atomic_read(&smcd->lgr_cnt))
 831                 wait_event(smcd->lgrs_deleted, !atomic_read(&smcd->lgr_cnt));
 832 }
 833
 834 /* Called when an SMCR device is removed or the smc module is unloaded.
 835  * If smcibdev is given, all SMCR link groups using this device are terminated.
 836  * If smcibdev is NULL, all SMCR link groups are terminated.
 837  */
 838 void smc_smcr_terminate_all(struct smc_ib_device *smcibdev)
 839 {
 840         struct smc_link_group *lgr, *lg;
 841         LIST_HEAD(lgr_free_list);
 842         int i;
 843
 844         spin_lock_bh(&smc_lgr_list.lock);
 845         if (!smcibdev) {
 846                 list_splice_init(&smc_lgr_list.list, &lgr_free_list);
 847                 list_for_each_entry(lgr, &lgr_free_list, list)
 848                         lgr->freeing = 1;
 849         } else {
 850                 list_for_each_entry_safe(lgr, lg, &smc_lgr_list.list, list) {
 851                         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 852                                 if (lgr->lnk[i].smcibdev == smcibdev) {
 853                                         list_move(&lgr->list, &lgr_free_list);
 854                                         lgr->freeing = 1;
 855                                         break;
 856                                 }
 857                         }
 858                 }
 859         }
 860         spin_unlock_bh(&smc_lgr_list.lock);
 861
 862         list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
 863                 list_del_init(&lgr->list);
 864                 __smc_lgr_terminate(lgr, false);
 865         }
 866
 867         if (smcibdev) {
 868                 if (atomic_read(&smcibdev->lnk_cnt))
 869                         wait_event(smcibdev->lnks_deleted,
 870                                    !atomic_read(&smcibdev->lnk_cnt));
 871         } else {
 872                 if (atomic_read(&lgr_cnt))
 873                         wait_event(lgrs_deleted, !atomic_read(&lgr_cnt));
 874         }
 875 }
 876
 877 /* Determine vlan of internal TCP socket.
 878  * @vlan_id: address to store the determined vlan id into
 879  */
 880 int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini)
 881 {
 882         struct dst_entry *dst = sk_dst_get(clcsock->sk);
 883         struct net_device *ndev;
 884         int i, nest_lvl, rc = 0;
 885
 886         ini->vlan_id = 0;
 887         if (!dst) {
 888                 rc = -ENOTCONN;
 889                 goto out;
 890         }
 891         if (!dst->dev) {
 892                 rc = -ENODEV;
 893                 goto out_rel;
 894         }
 895
 896         ndev = dst->dev;
 897         if (is_vlan_dev(ndev)) {
 898                 ini->vlan_id = vlan_dev_vlan_id(ndev);
 899                 goto out_rel;
 900         }
 901
 902         rtnl_lock();
 903         nest_lvl = ndev->lower_level;
 904         for (i = 0; i < nest_lvl; i++) {
 905                 struct list_head *lower = &ndev->adj_list.lower;
 906
 907                 if (list_empty(lower))
 908                         break;
 909                 lower = lower->next;
 910                 ndev = (struct net_device *)netdev_lower_get_next(ndev, &lower);
 911                 if (is_vlan_dev(ndev)) {
 912                         ini->vlan_id = vlan_dev_vlan_id(ndev);
 913                         break;
 914                 }
 915         }
 916         rtnl_unlock();
 917
 918 out_rel:
 919         dst_release(dst);
 920 out:
 921         return rc;
 922 }
 923
 924 static bool smcr_lgr_match(struct smc_link_group *lgr,
 925                            struct smc_clc_msg_local *lcl,
 926                            enum smc_lgr_role role, u32 clcqpn)
 927 {
 928         int i;
 929
 930         if (memcmp(lgr->peer_systemid, lcl->id_for_peer, SMC_SYSTEMID_LEN) ||
 931             lgr->role != role)
 932                 return false;
 933
 934         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 935                 if (lgr->lnk[i].state != SMC_LNK_ACTIVE)
 936                         continue;
 937                 if ((lgr->role == SMC_SERV || lgr->lnk[i].peer_qpn == clcqpn) &&
 938                     !memcmp(lgr->lnk[i].peer_gid, &lcl->gid, SMC_GID_SIZE) &&
 939                     !memcmp(lgr->lnk[i].peer_mac, lcl->mac, sizeof(lcl->mac)))
 940                         return true;
 941         }
 942         return false;
 943 }
 944
 945 static bool smcd_lgr_match(struct smc_link_group *lgr,
 946                            struct smcd_dev *smcismdev, u64 peer_gid)
 947 {
 948         return lgr->peer_gid == peer_gid && lgr->smcd == smcismdev;
 949 }
 950
 951 /* create a new SMC connection (and a new link group if necessary) */
 952 int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini)
 953 {
 954         struct smc_connection *conn = &smc->conn;
 955         struct list_head *lgr_list;
 956         struct smc_link_group *lgr;
 957         enum smc_lgr_role role;
 958         spinlock_t *lgr_lock;
 959         int rc = 0;
 960
 961         lgr_list = ini->is_smcd ? &ini->ism_dev->lgr_list : &smc_lgr_list.list;
 962         lgr_lock = ini->is_smcd ? &ini->ism_dev->lgr_lock : &smc_lgr_list.lock;
 963         ini->cln_first_contact = SMC_FIRST_CONTACT;
 964         role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
 965         if (role == SMC_CLNT && ini->srv_first_contact)
 966                 /* create new link group as well */
 967                 goto create;
 968
 969         /* determine if an existing link group can be reused */
 970         spin_lock_bh(lgr_lock);
 971         list_for_each_entry(lgr, lgr_list, list) {
 972                 write_lock_bh(&lgr->conns_lock);
 973                 if ((ini->is_smcd ?
 974                      smcd_lgr_match(lgr, ini->ism_dev, ini->ism_gid) :
 975                      smcr_lgr_match(lgr, ini->ib_lcl, role, ini->ib_clcqpn)) &&
 976                     !lgr->sync_err &&
 977                     lgr->vlan_id == ini->vlan_id &&
 978                     (role == SMC_CLNT ||
 979                      lgr->conns_num < SMC_RMBS_PER_LGR_MAX)) {
 980                         /* link group found */
 981                         ini->cln_first_contact = SMC_REUSE_CONTACT;
 982                         conn->lgr = lgr;
 983                         rc = smc_lgr_register_conn(conn); /* add conn to lgr */
 984                         write_unlock_bh(&lgr->conns_lock);
 985                         if (!rc && delayed_work_pending(&lgr->free_work))
 986                                 cancel_delayed_work(&lgr->free_work);
 987                         break;
 988                 }
 989                 write_unlock_bh(&lgr->conns_lock);
 990         }
 991         spin_unlock_bh(lgr_lock);
 992         if (rc)
 993                 return rc;
 994
 995         if (role == SMC_CLNT && !ini->srv_first_contact &&
 996             ini->cln_first_contact == SMC_FIRST_CONTACT) {
 997                 /* Server reuses a link group, but Client wants to start
 998                  * a new one
 999                  * send out_of_sync decline, reason synchr. error
1000                  */
1001                 return SMC_CLC_DECL_SYNCERR;
1002         }
1003
1004 create:
1005         if (ini->cln_first_contact == SMC_FIRST_CONTACT) {
1006                 rc = smc_lgr_create(smc, ini);
1007                 if (rc)
1008                         goto out;
1009                 lgr = conn->lgr;
1010                 write_lock_bh(&lgr->conns_lock);
1011                 rc = smc_lgr_register_conn(conn); /* add smc conn to lgr */
1012                 write_unlock_bh(&lgr->conns_lock);
1013                 if (rc)
1014                         goto out;
1015         }
1016         conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
1017         conn->local_tx_ctrl.len = SMC_WR_TX_SIZE;
1018         conn->urg_state = SMC_URG_READ;
1019         if (ini->is_smcd) {
1020                 conn->rx_off = sizeof(struct smcd_cdc_msg);
1021                 smcd_cdc_rx_init(conn); /* init tasklet for this conn */
1022         }
1023 #ifndef KERNEL_HAS_ATOMIC64
1024         spin_lock_init(&conn->acurs_lock);
1025 #endif
1026
1027 out:
1028         return rc;
1029 }
1030
1031 /* convert the RMB size into the compressed notation - minimum 16K.
1032  * In contrast to plain ilog2, this rounds towards the next power of 2,
1033  * so the socket application gets at least its desired sndbuf / rcvbuf size.
1034  */
1035 static u8 smc_compress_bufsize(int size)
1036 {
1037         u8 compressed;
1038
1039         if (size <= SMC_BUF_MIN_SIZE)
1040                 return 0;
1041
1042         size = (size - 1) >> 14;
1043         compressed = ilog2(size) + 1;
1044         if (compressed >= SMC_RMBE_SIZES)
1045                 compressed = SMC_RMBE_SIZES - 1;
1046         return compressed;
1047 }
1048
1049 /* convert the RMB size from compressed notation into integer */
1050 int smc_uncompress_bufsize(u8 compressed)
1051 {
1052         u32 size;
1053
1054         size = 0x00000001 << (((int)compressed) + 14);
1055         return (int)size;
1056 }
1057
1058 /* try to reuse a sndbuf or rmb description slot for a certain
1059  * buffer size; if not available, return NULL
1060  */
1061 static struct smc_buf_desc *smc_buf_get_slot(int compressed_bufsize,
1062                                              struct mutex *lock,
1063                                              struct list_head *buf_list)
1064 {
1065         struct smc_buf_desc *buf_slot;
1066
1067         mutex_lock(lock);
1068         list_for_each_entry(buf_slot, buf_list, list) {
1069                 if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
1070                         mutex_unlock(lock);
1071                         return buf_slot;
1072                 }
1073         }
1074         mutex_unlock(lock);
1075         return NULL;
1076 }
1077
1078 /* one of the conditions for announcing a receiver's current window size is
1079  * that it "results in a minimum increase in the window size of 10% of the
1080  * receive buffer space" [RFC7609]
1081  */
1082 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
1083 {
1084         return min_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
1085 }
1086
1087 /* map an rmb buf to a link */
1088 static int smcr_buf_map_link(struct smc_buf_desc *buf_desc, bool is_rmb,
1089                              struct smc_link *lnk)
1090 {
1091         int rc;
1092
1093         if (buf_desc->is_map_ib[lnk->link_idx])
1094                 return 0;
1095
1096         rc = sg_alloc_table(&buf_desc->sgt[lnk->link_idx], 1, GFP_KERNEL);
1097         if (rc)
1098                 return rc;
1099         sg_set_buf(buf_desc->sgt[lnk->link_idx].sgl,
1100                    buf_desc->cpu_addr, buf_desc->len);
1101
1102         /* map sg table to DMA address */
1103         rc = smc_ib_buf_map_sg(lnk, buf_desc,
1104                                is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
1105         /* SMC protocol depends on mapping to one DMA address only */
1106         if (rc != 1) {
1107                 rc = -EAGAIN;
1108                 goto free_table;
1109         }
1110
1111         /* create a new memory region for the RMB */
1112         if (is_rmb) {
1113                 rc = smc_ib_get_memory_region(lnk->roce_pd,
1114                                               IB_ACCESS_REMOTE_WRITE |
1115                                               IB_ACCESS_LOCAL_WRITE,
1116                                               buf_desc, lnk->link_idx);
1117                 if (rc)
1118                         goto buf_unmap;
1119                 smc_ib_sync_sg_for_device(lnk, buf_desc, DMA_FROM_DEVICE);
1120         }
1121         buf_desc->is_map_ib[lnk->link_idx] = true;
1122         return 0;
1123
1124 buf_unmap:
1125         smc_ib_buf_unmap_sg(lnk, buf_desc,
1126                             is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
1127 free_table:
1128         sg_free_table(&buf_desc->sgt[lnk->link_idx]);
1129         return rc;
1130 }
1131
1132 static struct smc_buf_desc *smcr_new_buf_create(struct smc_link_group *lgr,
1133                                                 bool is_rmb, int bufsize)
1134 {
1135         struct smc_buf_desc *buf_desc;
1136
1137         /* try to alloc a new buffer */
1138         buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
1139         if (!buf_desc)
1140                 return ERR_PTR(-ENOMEM);
1141
1142         buf_desc->order = get_order(bufsize);
1143         buf_desc->pages = alloc_pages(GFP_KERNEL | __GFP_NOWARN |
1144                                       __GFP_NOMEMALLOC | __GFP_COMP |
1145                                       __GFP_NORETRY | __GFP_ZERO,
1146                                       buf_desc->order);
1147         if (!buf_desc->pages) {
1148                 kfree(buf_desc);
1149                 return ERR_PTR(-EAGAIN);
1150         }
1151         buf_desc->cpu_addr = (void *)page_address(buf_desc->pages);
1152         buf_desc->len = bufsize;
1153         return buf_desc;
1154 }
1155
1156 /* map buf_desc on all usable links,
1157  * unused buffers stay mapped as long as the link is up
1158  */
1159 static int smcr_buf_map_usable_links(struct smc_link_group *lgr,
1160                                      struct smc_buf_desc *buf_desc, bool is_rmb)
1161 {
1162         int i, rc = 0;
1163
1164         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1165                 struct smc_link *lnk = &lgr->lnk[i];
1166
1167                 if (!smc_link_usable(lnk))
1168                         continue;
1169                 if (smcr_buf_map_link(buf_desc, is_rmb, lnk)) {
1170                         smcr_buf_unuse(buf_desc, lnk);
1171                         rc = -ENOMEM;
1172                         goto out;
1173                 }
1174         }
1175 out:
1176         return rc;
1177 }
1178
1179 #define SMCD_DMBE_SIZES         7 /* 0 -> 16KB, 1 -> 32KB, .. 6 -> 1MB */
1180
1181 static struct smc_buf_desc *smcd_new_buf_create(struct smc_link_group *lgr,
1182                                                 bool is_dmb, int bufsize)
1183 {
1184         struct smc_buf_desc *buf_desc;
1185         int rc;
1186
1187         if (smc_compress_bufsize(bufsize) > SMCD_DMBE_SIZES)
1188                 return ERR_PTR(-EAGAIN);
1189
1190         /* try to alloc a new DMB */
1191         buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
1192         if (!buf_desc)
1193                 return ERR_PTR(-ENOMEM);
1194         if (is_dmb) {
1195                 rc = smc_ism_register_dmb(lgr, bufsize, buf_desc);
1196                 if (rc) {
1197                         kfree(buf_desc);
1198                         return ERR_PTR(-EAGAIN);
1199                 }
1200                 buf_desc->pages = virt_to_page(buf_desc->cpu_addr);
1201                 /* CDC header stored in buf. So, pretend it was smaller */
1202                 buf_desc->len = bufsize - sizeof(struct smcd_cdc_msg);
1203         } else {
1204                 buf_desc->cpu_addr = kzalloc(bufsize, GFP_KERNEL |
1205                                              __GFP_NOWARN | __GFP_NORETRY |
1206                                              __GFP_NOMEMALLOC);
1207                 if (!buf_desc->cpu_addr) {
1208                         kfree(buf_desc);
1209                         return ERR_PTR(-EAGAIN);
1210                 }
1211                 buf_desc->len = bufsize;
1212         }
1213         return buf_desc;
1214 }
1215
1216 static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
1217 {
1218         struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
1219         struct smc_connection *conn = &smc->conn;
1220         struct smc_link_group *lgr = conn->lgr;
1221         struct list_head *buf_list;
1222         int bufsize, bufsize_short;
1223         struct mutex *lock;     /* lock buffer list */
1224         int sk_buf_size;
1225
1226         if (is_rmb)
1227                 /* use socket recv buffer size (w/o overhead) as start value */
1228                 sk_buf_size = smc->sk.sk_rcvbuf / 2;
1229         else
1230                 /* use socket send buffer size (w/o overhead) as start value */
1231                 sk_buf_size = smc->sk.sk_sndbuf / 2;
1232
1233         for (bufsize_short = smc_compress_bufsize(sk_buf_size);
1234              bufsize_short >= 0; bufsize_short--) {
1235
1236                 if (is_rmb) {
1237                         lock = &lgr->rmbs_lock;
1238                         buf_list = &lgr->rmbs[bufsize_short];
1239                 } else {
1240                         lock = &lgr->sndbufs_lock;
1241                         buf_list = &lgr->sndbufs[bufsize_short];
1242                 }
1243                 bufsize = smc_uncompress_bufsize(bufsize_short);
1244                 if ((1 << get_order(bufsize)) > SG_MAX_SINGLE_ALLOC)
1245                         continue;
1246
1247                 /* check for reusable slot in the link group */
1248                 buf_desc = smc_buf_get_slot(bufsize_short, lock, buf_list);
1249                 if (buf_desc) {
1250                         memset(buf_desc->cpu_addr, 0, bufsize);
1251                         break; /* found reusable slot */
1252                 }
1253
1254                 if (is_smcd)
1255                         buf_desc = smcd_new_buf_create(lgr, is_rmb, bufsize);
1256                 else
1257                         buf_desc = smcr_new_buf_create(lgr, is_rmb, bufsize);
1258
1259                 if (PTR_ERR(buf_desc) == -ENOMEM)
1260                         break;
1261                 if (IS_ERR(buf_desc))
1262                         continue;
1263
1264                 buf_desc->used = 1;
1265                 mutex_lock(lock);
1266                 list_add(&buf_desc->list, buf_list);
1267                 mutex_unlock(lock);
1268                 break; /* found */
1269         }
1270
1271         if (IS_ERR(buf_desc))
1272                 return -ENOMEM;
1273
1274         if (!is_smcd) {
1275                 if (smcr_buf_map_usable_links(lgr, buf_desc, is_rmb)) {
1276                         return -ENOMEM;
1277                 }
1278         }
1279
1280         if (is_rmb) {
1281                 conn->rmb_desc = buf_desc;
1282                 conn->rmbe_size_short = bufsize_short;
1283                 smc->sk.sk_rcvbuf = bufsize * 2;
1284                 atomic_set(&conn->bytes_to_rcv, 0);
1285                 conn->rmbe_update_limit =
1286                         smc_rmb_wnd_update_limit(buf_desc->len);
1287                 if (is_smcd)
1288                         smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */
1289         } else {
1290                 conn->sndbuf_desc = buf_desc;
1291                 smc->sk.sk_sndbuf = bufsize * 2;
1292                 atomic_set(&conn->sndbuf_space, bufsize);
1293         }
1294         return 0;
1295 }
1296
1297 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn)
1298 {
1299         if (!conn->lgr || conn->lgr->is_smcd || !smc_link_usable(conn->lnk))
1300                 return;
1301         smc_ib_sync_sg_for_cpu(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE);
1302 }
1303
1304 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
1305 {
1306         if (!conn->lgr || conn->lgr->is_smcd || !smc_link_usable(conn->lnk))
1307                 return;
1308         smc_ib_sync_sg_for_device(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE);
1309 }
1310
1311 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
1312 {
1313         int i;
1314
1315         if (!conn->lgr || conn->lgr->is_smcd)
1316                 return;
1317         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1318                 if (!smc_link_usable(&conn->lgr->lnk[i]))
1319                         continue;
1320                 smc_ib_sync_sg_for_cpu(&conn->lgr->lnk[i], conn->rmb_desc,
1321                                        DMA_FROM_DEVICE);
1322         }
1323 }
1324
1325 void smc_rmb_sync_sg_for_device(struct smc_connection *conn)
1326 {
1327         int i;
1328
1329         if (!conn->lgr || conn->lgr->is_smcd)
1330                 return;
1331         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1332                 if (!smc_link_usable(&conn->lgr->lnk[i]))
1333                         continue;
1334                 smc_ib_sync_sg_for_device(&conn->lgr->lnk[i], conn->rmb_desc,
1335                                           DMA_FROM_DEVICE);
1336         }
1337 }
1338
1339 /* create the send and receive buffer for an SMC socket;
1340  * receive buffers are called RMBs;
1341  * (even though the SMC protocol allows more than one RMB-element per RMB,
1342  * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
1343  * extra RMB for every connection in a link group
1344  */
1345 int smc_buf_create(struct smc_sock *smc, bool is_smcd)
1346 {
1347         int rc;
1348
1349         /* create send buffer */
1350         rc = __smc_buf_create(smc, is_smcd, false);
1351         if (rc)
1352                 return rc;
1353         /* create rmb */
1354         rc = __smc_buf_create(smc, is_smcd, true);
1355         if (rc)
1356                 smc_buf_free(smc->conn.lgr, false, smc->conn.sndbuf_desc);
1357         return rc;
1358 }
1359
1360 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
1361 {
1362         int i;
1363
1364         for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
1365                 if (!test_and_set_bit(i, lgr->rtokens_used_mask))
1366                         return i;
1367         }
1368         return -ENOSPC;
1369 }
1370
1371 /* add a new rtoken from peer */
1372 int smc_rtoken_add(struct smc_link *lnk, __be64 nw_vaddr, __be32 nw_rkey)
1373 {
1374         struct smc_link_group *lgr = smc_get_lgr(lnk);
1375         u64 dma_addr = be64_to_cpu(nw_vaddr);
1376         u32 rkey = ntohl(nw_rkey);
1377         int i;
1378
1379         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1380                 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
1381                     lgr->rtokens[i][lnk->link_idx].dma_addr == dma_addr &&
1382                     test_bit(i, lgr->rtokens_used_mask)) {
1383                         /* already in list */
1384                         return i;
1385                 }
1386         }
1387         i = smc_rmb_reserve_rtoken_idx(lgr);
1388         if (i < 0)
1389                 return i;
1390         lgr->rtokens[i][lnk->link_idx].rkey = rkey;
1391         lgr->rtokens[i][lnk->link_idx].dma_addr = dma_addr;
1392         return i;
1393 }
1394
1395 /* delete an rtoken from all links */
1396 int smc_rtoken_delete(struct smc_link *lnk, __be32 nw_rkey)
1397 {
1398         struct smc_link_group *lgr = smc_get_lgr(lnk);
1399         u32 rkey = ntohl(nw_rkey);
1400         int i, j;
1401
1402         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1403                 if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
1404                     test_bit(i, lgr->rtokens_used_mask)) {
1405                         for (j = 0; j < SMC_LINKS_PER_LGR_MAX; j++) {
1406                                 lgr->rtokens[i][j].rkey = 0;
1407                                 lgr->rtokens[i][j].dma_addr = 0;
1408                         }
1409                         clear_bit(i, lgr->rtokens_used_mask);
1410                         return 0;
1411                 }
1412         }
1413         return -ENOENT;
1414 }
1415
1416 /* save rkey and dma_addr received from peer during clc handshake */
1417 int smc_rmb_rtoken_handling(struct smc_connection *conn,
1418                             struct smc_link *lnk,
1419                             struct smc_clc_msg_accept_confirm *clc)
1420 {
1421         conn->rtoken_idx = smc_rtoken_add(lnk, clc->rmb_dma_addr,
1422                                           clc->rmb_rkey);
1423         if (conn->rtoken_idx < 0)
1424                 return conn->rtoken_idx;
1425         return 0;
1426 }
1427
1428 static void smc_core_going_away(void)
1429 {
1430         struct smc_ib_device *smcibdev;
1431         struct smcd_dev *smcd;
1432
1433         spin_lock(&smc_ib_devices.lock);
1434         list_for_each_entry(smcibdev, &smc_ib_devices.list, list) {
1435                 int i;
1436
1437                 for (i = 0; i < SMC_MAX_PORTS; i++)
1438                         set_bit(i, smcibdev->ports_going_away);
1439         }
1440         spin_unlock(&smc_ib_devices.lock);
1441
1442         spin_lock(&smcd_dev_list.lock);
1443         list_for_each_entry(smcd, &smcd_dev_list.list, list) {
1444                 smcd->going_away = 1;
1445         }
1446         spin_unlock(&smcd_dev_list.lock);
1447 }
1448
1449 /* Clean up all SMC link groups */
1450 static void smc_lgrs_shutdown(void)
1451 {
1452         struct smcd_dev *smcd;
1453
1454         smc_core_going_away();
1455
1456         smc_smcr_terminate_all(NULL);
1457
1458         spin_lock(&smcd_dev_list.lock);
1459         list_for_each_entry(smcd, &smcd_dev_list.list, list)
1460                 smc_smcd_terminate_all(smcd);
1461         spin_unlock(&smcd_dev_list.lock);
1462 }
1463
1464 static int smc_core_reboot_event(struct notifier_block *this,
1465                                  unsigned long event, void *ptr)
1466 {
1467         smc_lgrs_shutdown();
1468         smc_ib_unregister_client();
1469         return 0;
1470 }
1471
1472 static struct notifier_block smc_reboot_notifier = {
1473         .notifier_call = smc_core_reboot_event,
1474 };
1475
1476 int __init smc_core_init(void)
1477 {
1478         return register_reboot_notifier(&smc_reboot_notifier);
1479 }
1480
1481 /* Called (from smc_exit) when module is removed */
1482 void smc_core_exit(void)
1483 {
1484         unregister_reboot_notifier(&smc_reboot_notifier);
1485         smc_lgrs_shutdown();
1486 }