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 <net/tcp.h>
  17 #include <net/sock.h>
  18 #include <rdma/ib_verbs.h>
  19 #include <rdma/ib_cache.h>
  20
  21 #include "smc.h"
  22 #include "smc_clc.h"
  23 #include "smc_core.h"
  24 #include "smc_ib.h"
  25 #include "smc_wr.h"
  26 #include "smc_llc.h"
  27 #include "smc_cdc.h"
  28 #include "smc_close.h"
  29 #include "smc_ism.h"
  30
  31 #define SMC_LGR_NUM_INCR                256
  32 #define SMC_LGR_FREE_DELAY_SERV         (600 * HZ)
  33 #define SMC_LGR_FREE_DELAY_CLNT         (SMC_LGR_FREE_DELAY_SERV + 10 * HZ)
  34 #define SMC_LGR_FREE_DELAY_FAST         (8 * HZ)
  35
  36 static struct smc_lgr_list smc_lgr_list = {     /* established link groups */
  37         .lock = __SPIN_LOCK_UNLOCKED(smc_lgr_list.lock),
  38         .list = LIST_HEAD_INIT(smc_lgr_list.list),
  39         .num = 0,
  40 };
  41
  42 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
  43                          struct smc_buf_desc *buf_desc);
  44
  45 static void smc_lgr_schedule_free_work(struct smc_link_group *lgr)
  46 {
  47         /* client link group creation always follows the server link group
  48          * creation. For client use a somewhat higher removal delay time,
  49          * otherwise there is a risk of out-of-sync link groups.
  50          */
  51         mod_delayed_work(system_wq, &lgr->free_work,
  52                          (!lgr->is_smcd && lgr->role == SMC_CLNT) ?
  53                          SMC_LGR_FREE_DELAY_CLNT : SMC_LGR_FREE_DELAY_SERV);
  54 }
  55
  56 void smc_lgr_schedule_free_work_fast(struct smc_link_group *lgr)
  57 {
  58         mod_delayed_work(system_wq, &lgr->free_work, SMC_LGR_FREE_DELAY_FAST);
  59 }
  60
  61 /* Register connection's alert token in our lookup structure.
  62  * To use rbtrees we have to implement our own insert core.
  63  * Requires @conns_lock
  64  * @smc         connection to register
  65  * Returns 0 on success, != otherwise.
  66  */
  67 static void smc_lgr_add_alert_token(struct smc_connection *conn)
  68 {
  69         struct rb_node **link, *parent = NULL;
  70         u32 token = conn->alert_token_local;
  71
  72         link = &conn->lgr->conns_all.rb_node;
  73         while (*link) {
  74                 struct smc_connection *cur = rb_entry(*link,
  75                                         struct smc_connection, alert_node);
  76
  77                 parent = *link;
  78                 if (cur->alert_token_local > token)
  79                         link = &parent->rb_left;
  80                 else
  81                         link = &parent->rb_right;
  82         }
  83         /* Put the new node there */
  84         rb_link_node(&conn->alert_node, parent, link);
  85         rb_insert_color(&conn->alert_node, &conn->lgr->conns_all);
  86 }
  87
  88 /* Register connection in link group by assigning an alert token
  89  * registered in a search tree.
  90  * Requires @conns_lock
  91  * Note that '0' is a reserved value and not assigned.
  92  */
  93 static void smc_lgr_register_conn(struct smc_connection *conn)
  94 {
  95         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
  96         static atomic_t nexttoken = ATOMIC_INIT(0);
  97
  98         /* find a new alert_token_local value not yet used by some connection
  99          * in this link group
 100          */
 101         sock_hold(&smc->sk); /* sock_put in smc_lgr_unregister_conn() */
 102         while (!conn->alert_token_local) {
 103                 conn->alert_token_local = atomic_inc_return(&nexttoken);
 104                 if (smc_lgr_find_conn(conn->alert_token_local, conn->lgr))
 105                         conn->alert_token_local = 0;
 106         }
 107         smc_lgr_add_alert_token(conn);
 108         conn->lgr->conns_num++;
 109 }
 110
 111 /* Unregister connection and reset the alert token of the given connection<
 112  */
 113 static void __smc_lgr_unregister_conn(struct smc_connection *conn)
 114 {
 115         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
 116         struct smc_link_group *lgr = conn->lgr;
 117
 118         rb_erase(&conn->alert_node, &lgr->conns_all);
 119         lgr->conns_num--;
 120         conn->alert_token_local = 0;
 121         conn->lgr = NULL;
 122         sock_put(&smc->sk); /* sock_hold in smc_lgr_register_conn() */
 123 }
 124
 125 /* Unregister connection from lgr
 126  */
 127 static void smc_lgr_unregister_conn(struct smc_connection *conn)
 128 {
 129         struct smc_link_group *lgr = conn->lgr;
 130
 131         if (!lgr)
 132                 return;
 133         write_lock_bh(&lgr->conns_lock);
 134         if (conn->alert_token_local) {
 135                 __smc_lgr_unregister_conn(conn);
 136         }
 137         write_unlock_bh(&lgr->conns_lock);
 138 }
 139
 140 /* Send delete link, either as client to request the initiation
 141  * of the DELETE LINK sequence from server; or as server to
 142  * initiate the delete processing. See smc_llc_rx_delete_link().
 143  */
 144 static int smc_link_send_delete(struct smc_link *lnk)
 145 {
 146         if (lnk->state == SMC_LNK_ACTIVE &&
 147             !smc_llc_send_delete_link(lnk, SMC_LLC_REQ, true)) {
 148                 smc_llc_link_deleting(lnk);
 149                 return 0;
 150         }
 151         return -ENOTCONN;
 152 }
 153
 154 static void smc_lgr_free_work(struct work_struct *work)
 155 {
 156         struct smc_link_group *lgr = container_of(to_delayed_work(work),
 157                                                   struct smc_link_group,
 158                                                   free_work);
 159         bool conns;
 160
 161         spin_lock_bh(&smc_lgr_list.lock);
 162         if (list_empty(&lgr->list))
 163                 goto free;
 164         read_lock_bh(&lgr->conns_lock);
 165         conns = RB_EMPTY_ROOT(&lgr->conns_all);
 166         read_unlock_bh(&lgr->conns_lock);
 167         if (!conns) { /* number of lgr connections is no longer zero */
 168                 spin_unlock_bh(&smc_lgr_list.lock);
 169                 return;
 170         }
 171         list_del_init(&lgr->list); /* remove from smc_lgr_list */
 172 free:
 173         spin_unlock_bh(&smc_lgr_list.lock);
 174
 175         if (!lgr->is_smcd && !lgr->terminating) {
 176                 /* try to send del link msg, on error free lgr immediately */
 177                 if (!smc_link_send_delete(&lgr->lnk[SMC_SINGLE_LINK])) {
 178                         /* reschedule in case we never receive a response */
 179                         smc_lgr_schedule_free_work(lgr);
 180                         return;
 181                 }
 182         }
 183
 184         if (!delayed_work_pending(&lgr->free_work)) {
 185                 struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
 186
 187                 if (!lgr->is_smcd && lnk->state != SMC_LNK_INACTIVE)
 188                         smc_llc_link_inactive(lnk);
 189                 smc_lgr_free(lgr);
 190         }
 191 }
 192
 193 /* create a new SMC link group */
 194 static int smc_lgr_create(struct smc_sock *smc, bool is_smcd,
 195                           struct smc_ib_device *smcibdev, u8 ibport,
 196                           char *peer_systemid, unsigned short vlan_id,
 197                           struct smcd_dev *smcismdev, u64 peer_gid)
 198 {
 199         struct smc_link_group *lgr;
 200         struct smc_link *lnk;
 201         u8 rndvec[3];
 202         int rc = 0;
 203         int i;
 204
 205         if (is_smcd && vlan_id) {
 206                 rc = smc_ism_get_vlan(smcismdev, vlan_id);
 207                 if (rc)
 208                         goto out;
 209         }
 210
 211         lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
 212         if (!lgr) {
 213                 rc = -ENOMEM;
 214                 goto out;
 215         }
 216         lgr->is_smcd = is_smcd;
 217         lgr->sync_err = 0;
 218         lgr->vlan_id = vlan_id;
 219         rwlock_init(&lgr->sndbufs_lock);
 220         rwlock_init(&lgr->rmbs_lock);
 221         rwlock_init(&lgr->conns_lock);
 222         for (i = 0; i < SMC_RMBE_SIZES; i++) {
 223                 INIT_LIST_HEAD(&lgr->sndbufs[i]);
 224                 INIT_LIST_HEAD(&lgr->rmbs[i]);
 225         }
 226         smc_lgr_list.num += SMC_LGR_NUM_INCR;
 227         memcpy(&lgr->id, (u8 *)&smc_lgr_list.num, SMC_LGR_ID_SIZE);
 228         INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
 229         lgr->conns_all = RB_ROOT;
 230         if (is_smcd) {
 231                 /* SMC-D specific settings */
 232                 lgr->peer_gid = peer_gid;
 233                 lgr->smcd = smcismdev;
 234         } else {
 235                 /* SMC-R specific settings */
 236                 lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
 237                 memcpy(lgr->peer_systemid, peer_systemid, SMC_SYSTEMID_LEN);
 238
 239                 lnk = &lgr->lnk[SMC_SINGLE_LINK];
 240                 /* initialize link */
 241                 lnk->state = SMC_LNK_ACTIVATING;
 242                 lnk->link_id = SMC_SINGLE_LINK;
 243                 lnk->smcibdev = smcibdev;
 244                 lnk->ibport = ibport;
 245                 lnk->path_mtu = smcibdev->pattr[ibport - 1].active_mtu;
 246                 if (!smcibdev->initialized)
 247                         smc_ib_setup_per_ibdev(smcibdev);
 248                 get_random_bytes(rndvec, sizeof(rndvec));
 249                 lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) +
 250                         (rndvec[2] << 16);
 251                 rc = smc_ib_determine_gid(lnk->smcibdev, lnk->ibport,
 252                                           vlan_id, lnk->gid, &lnk->sgid_index);
 253                 if (rc)
 254                         goto free_lgr;
 255                 rc = smc_llc_link_init(lnk);
 256                 if (rc)
 257                         goto free_lgr;
 258                 rc = smc_wr_alloc_link_mem(lnk);
 259                 if (rc)
 260                         goto clear_llc_lnk;
 261                 rc = smc_ib_create_protection_domain(lnk);
 262                 if (rc)
 263                         goto free_link_mem;
 264                 rc = smc_ib_create_queue_pair(lnk);
 265                 if (rc)
 266                         goto dealloc_pd;
 267                 rc = smc_wr_create_link(lnk);
 268                 if (rc)
 269                         goto destroy_qp;
 270         }
 271         smc->conn.lgr = lgr;
 272         spin_lock_bh(&smc_lgr_list.lock);
 273         list_add(&lgr->list, &smc_lgr_list.list);
 274         spin_unlock_bh(&smc_lgr_list.lock);
 275         return 0;
 276
 277 destroy_qp:
 278         smc_ib_destroy_queue_pair(lnk);
 279 dealloc_pd:
 280         smc_ib_dealloc_protection_domain(lnk);
 281 free_link_mem:
 282         smc_wr_free_link_mem(lnk);
 283 clear_llc_lnk:
 284         smc_llc_link_clear(lnk);
 285 free_lgr:
 286         kfree(lgr);
 287 out:
 288         return rc;
 289 }
 290
 291 static void smc_buf_unuse(struct smc_connection *conn,
 292                           struct smc_link_group *lgr)
 293 {
 294         if (conn->sndbuf_desc)
 295                 conn->sndbuf_desc->used = 0;
 296         if (conn->rmb_desc) {
 297                 if (!conn->rmb_desc->regerr) {
 298                         conn->rmb_desc->reused = 1;
 299                         conn->rmb_desc->used = 0;
 300                 } else {
 301                         /* buf registration failed, reuse not possible */
 302                         write_lock_bh(&lgr->rmbs_lock);
 303                         list_del(&conn->rmb_desc->list);
 304                         write_unlock_bh(&lgr->rmbs_lock);
 305
 306                         smc_buf_free(lgr, true, conn->rmb_desc);
 307                 }
 308         }
 309 }
 310
 311 /* remove a finished connection from its link group */
 312 void smc_conn_free(struct smc_connection *conn)
 313 {
 314         struct smc_link_group *lgr = conn->lgr;
 315
 316         if (!lgr)
 317                 return;
 318         if (lgr->is_smcd) {
 319                 smc_ism_unset_conn(conn);
 320                 tasklet_kill(&conn->rx_tsklet);
 321         } else {
 322                 smc_cdc_tx_dismiss_slots(conn);
 323         }
 324         smc_lgr_unregister_conn(conn);          /* unsets conn->lgr */
 325         smc_buf_unuse(conn, lgr);               /* allow buffer reuse */
 326
 327         if (!lgr->conns_num)
 328                 smc_lgr_schedule_free_work(lgr);
 329 }
 330
 331 static void smc_link_clear(struct smc_link *lnk)
 332 {
 333         lnk->peer_qpn = 0;
 334         smc_llc_link_clear(lnk);
 335         smc_ib_modify_qp_reset(lnk);
 336         smc_wr_free_link(lnk);
 337         smc_ib_destroy_queue_pair(lnk);
 338         smc_ib_dealloc_protection_domain(lnk);
 339         smc_wr_free_link_mem(lnk);
 340 }
 341
 342 static void smcr_buf_free(struct smc_link_group *lgr, bool is_rmb,
 343                           struct smc_buf_desc *buf_desc)
 344 {
 345         struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
 346
 347         if (is_rmb) {
 348                 if (buf_desc->mr_rx[SMC_SINGLE_LINK])
 349                         smc_ib_put_memory_region(
 350                                         buf_desc->mr_rx[SMC_SINGLE_LINK]);
 351                 smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
 352                                     DMA_FROM_DEVICE);
 353         } else {
 354                 smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
 355                                     DMA_TO_DEVICE);
 356         }
 357         sg_free_table(&buf_desc->sgt[SMC_SINGLE_LINK]);
 358         if (buf_desc->pages)
 359                 __free_pages(buf_desc->pages, buf_desc->order);
 360         kfree(buf_desc);
 361 }
 362
 363 static void smcd_buf_free(struct smc_link_group *lgr, bool is_dmb,
 364                           struct smc_buf_desc *buf_desc)
 365 {
 366         if (is_dmb) {
 367                 /* restore original buf len */
 368                 buf_desc->len += sizeof(struct smcd_cdc_msg);
 369                 smc_ism_unregister_dmb(lgr->smcd, buf_desc);
 370         } else {
 371                 kfree(buf_desc->cpu_addr);
 372         }
 373         kfree(buf_desc);
 374 }
 375
 376 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
 377                          struct smc_buf_desc *buf_desc)
 378 {
 379         if (lgr->is_smcd)
 380                 smcd_buf_free(lgr, is_rmb, buf_desc);
 381         else
 382                 smcr_buf_free(lgr, is_rmb, buf_desc);
 383 }
 384
 385 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
 386 {
 387         struct smc_buf_desc *buf_desc, *bf_desc;
 388         struct list_head *buf_list;
 389         int i;
 390
 391         for (i = 0; i < SMC_RMBE_SIZES; i++) {
 392                 if (is_rmb)
 393                         buf_list = &lgr->rmbs[i];
 394                 else
 395                         buf_list = &lgr->sndbufs[i];
 396                 list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
 397                                          list) {
 398                         list_del(&buf_desc->list);
 399                         smc_buf_free(lgr, is_rmb, buf_desc);
 400                 }
 401         }
 402 }
 403
 404 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
 405 {
 406         /* free send buffers */
 407         __smc_lgr_free_bufs(lgr, false);
 408         /* free rmbs */
 409         __smc_lgr_free_bufs(lgr, true);
 410 }
 411
 412 /* remove a link group */
 413 void smc_lgr_free(struct smc_link_group *lgr)
 414 {
 415         smc_lgr_free_bufs(lgr);
 416         if (lgr->is_smcd)
 417                 smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
 418         else
 419                 smc_link_clear(&lgr->lnk[SMC_SINGLE_LINK]);
 420         kfree(lgr);
 421 }
 422
 423 void smc_lgr_forget(struct smc_link_group *lgr)
 424 {
 425         spin_lock_bh(&smc_lgr_list.lock);
 426         /* do not use this link group for new connections */
 427         if (!list_empty(&lgr->list))
 428                 list_del_init(&lgr->list);
 429         spin_unlock_bh(&smc_lgr_list.lock);
 430 }
 431
 432 /* terminate linkgroup abnormally */
 433 static void __smc_lgr_terminate(struct smc_link_group *lgr)
 434 {
 435         struct smc_connection *conn;
 436         struct smc_sock *smc;
 437         struct rb_node *node;
 438
 439         if (lgr->terminating)
 440                 return; /* lgr already terminating */
 441         lgr->terminating = 1;
 442         if (!list_empty(&lgr->list)) /* forget lgr */
 443                 list_del_init(&lgr->list);
 444         if (!lgr->is_smcd)
 445                 smc_llc_link_inactive(&lgr->lnk[SMC_SINGLE_LINK]);
 446
 447         write_lock_bh(&lgr->conns_lock);
 448         node = rb_first(&lgr->conns_all);
 449         while (node) {
 450                 conn = rb_entry(node, struct smc_connection, alert_node);
 451                 smc = container_of(conn, struct smc_sock, conn);
 452                 sock_hold(&smc->sk); /* sock_put in close work */
 453                 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
 454                 __smc_lgr_unregister_conn(conn);
 455                 write_unlock_bh(&lgr->conns_lock);
 456                 if (!schedule_work(&conn->close_work))
 457                         sock_put(&smc->sk);
 458                 write_lock_bh(&lgr->conns_lock);
 459                 node = rb_first(&lgr->conns_all);
 460         }
 461         write_unlock_bh(&lgr->conns_lock);
 462         if (!lgr->is_smcd)
 463                 wake_up(&lgr->lnk[SMC_SINGLE_LINK].wr_reg_wait);
 464         smc_lgr_schedule_free_work(lgr);
 465 }
 466
 467 void smc_lgr_terminate(struct smc_link_group *lgr)
 468 {
 469         spin_lock_bh(&smc_lgr_list.lock);
 470         __smc_lgr_terminate(lgr);
 471         spin_unlock_bh(&smc_lgr_list.lock);
 472 }
 473
 474 /* Called when IB port is terminated */
 475 void smc_port_terminate(struct smc_ib_device *smcibdev, u8 ibport)
 476 {
 477         struct smc_link_group *lgr, *l;
 478
 479         spin_lock_bh(&smc_lgr_list.lock);
 480         list_for_each_entry_safe(lgr, l, &smc_lgr_list.list, list) {
 481                 if (!lgr->is_smcd &&
 482                     lgr->lnk[SMC_SINGLE_LINK].smcibdev == smcibdev &&
 483                     lgr->lnk[SMC_SINGLE_LINK].ibport == ibport)
 484                         __smc_lgr_terminate(lgr);
 485         }
 486         spin_unlock_bh(&smc_lgr_list.lock);
 487 }
 488
 489 /* Called when SMC-D device is terminated or peer is lost */
 490 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid)
 491 {
 492         struct smc_link_group *lgr, *l;
 493         LIST_HEAD(lgr_free_list);
 494
 495         /* run common cleanup function and build free list */
 496         spin_lock_bh(&smc_lgr_list.lock);
 497         list_for_each_entry_safe(lgr, l, &smc_lgr_list.list, list) {
 498                 if (lgr->is_smcd && lgr->smcd == dev &&
 499                     (!peer_gid || lgr->peer_gid == peer_gid) &&
 500                     !list_empty(&lgr->list)) {
 501                         __smc_lgr_terminate(lgr);
 502                         list_move(&lgr->list, &lgr_free_list);
 503                 }
 504         }
 505         spin_unlock_bh(&smc_lgr_list.lock);
 506
 507         /* cancel the regular free workers and actually free lgrs */
 508         list_for_each_entry_safe(lgr, l, &lgr_free_list, list) {
 509                 list_del_init(&lgr->list);
 510                 cancel_delayed_work_sync(&lgr->free_work);
 511                 smc_lgr_free(lgr);
 512         }
 513 }
 514
 515 /* Determine vlan of internal TCP socket.
 516  * @vlan_id: address to store the determined vlan id into
 517  */
 518 int smc_vlan_by_tcpsk(struct socket *clcsock, unsigned short *vlan_id)
 519 {
 520         struct dst_entry *dst = sk_dst_get(clcsock->sk);
 521         struct net_device *ndev;
 522         int i, nest_lvl, rc = 0;
 523
 524         *vlan_id = 0;
 525         if (!dst) {
 526                 rc = -ENOTCONN;
 527                 goto out;
 528         }
 529         if (!dst->dev) {
 530                 rc = -ENODEV;
 531                 goto out_rel;
 532         }
 533
 534         ndev = dst->dev;
 535         if (is_vlan_dev(ndev)) {
 536                 *vlan_id = vlan_dev_vlan_id(ndev);
 537                 goto out_rel;
 538         }
 539
 540         rtnl_lock();
 541         nest_lvl = dev_get_nest_level(ndev);
 542         for (i = 0; i < nest_lvl; i++) {
 543                 struct list_head *lower = &ndev->adj_list.lower;
 544
 545                 if (list_empty(lower))
 546                         break;
 547                 lower = lower->next;
 548                 ndev = (struct net_device *)netdev_lower_get_next(ndev, &lower);
 549                 if (is_vlan_dev(ndev)) {
 550                         *vlan_id = vlan_dev_vlan_id(ndev);
 551                         break;
 552                 }
 553         }
 554         rtnl_unlock();
 555
 556 out_rel:
 557         dst_release(dst);
 558 out:
 559         return rc;
 560 }
 561
 562 static bool smcr_lgr_match(struct smc_link_group *lgr,
 563                            struct smc_clc_msg_local *lcl,
 564                            enum smc_lgr_role role)
 565 {
 566         return !memcmp(lgr->peer_systemid, lcl->id_for_peer,
 567                        SMC_SYSTEMID_LEN) &&
 568                 !memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_gid, &lcl->gid,
 569                         SMC_GID_SIZE) &&
 570                 !memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_mac, lcl->mac,
 571                         sizeof(lcl->mac)) &&
 572                 lgr->role == role;
 573 }
 574
 575 static bool smcd_lgr_match(struct smc_link_group *lgr,
 576                            struct smcd_dev *smcismdev, u64 peer_gid)
 577 {
 578         return lgr->peer_gid == peer_gid && lgr->smcd == smcismdev;
 579 }
 580
 581 /* create a new SMC connection (and a new link group if necessary) */
 582 int smc_conn_create(struct smc_sock *smc, bool is_smcd, int srv_first_contact,
 583                     struct smc_ib_device *smcibdev, u8 ibport,
 584                     struct smc_clc_msg_local *lcl, struct smcd_dev *smcd,
 585                     u64 peer_gid)
 586 {
 587         struct smc_connection *conn = &smc->conn;
 588         int local_contact = SMC_FIRST_CONTACT;
 589         struct smc_link_group *lgr;
 590         unsigned short vlan_id;
 591         enum smc_lgr_role role;
 592         int rc = 0;
 593
 594         role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
 595         rc = smc_vlan_by_tcpsk(smc->clcsock, &vlan_id);
 596         if (rc)
 597                 return rc;
 598
 599         if ((role == SMC_CLNT) && srv_first_contact)
 600                 /* create new link group as well */
 601                 goto create;
 602
 603         /* determine if an existing link group can be reused */
 604         spin_lock_bh(&smc_lgr_list.lock);
 605         list_for_each_entry(lgr, &smc_lgr_list.list, list) {
 606                 write_lock_bh(&lgr->conns_lock);
 607                 if ((is_smcd ? smcd_lgr_match(lgr, smcd, peer_gid) :
 608                      smcr_lgr_match(lgr, lcl, role)) &&
 609                     !lgr->sync_err &&
 610                     lgr->vlan_id == vlan_id &&
 611                     (role == SMC_CLNT ||
 612                      lgr->conns_num < SMC_RMBS_PER_LGR_MAX)) {
 613                         /* link group found */
 614                         local_contact = SMC_REUSE_CONTACT;
 615                         conn->lgr = lgr;
 616                         smc_lgr_register_conn(conn); /* add smc conn to lgr */
 617                         if (delayed_work_pending(&lgr->free_work))
 618                                 cancel_delayed_work(&lgr->free_work);
 619                         write_unlock_bh(&lgr->conns_lock);
 620                         break;
 621                 }
 622                 write_unlock_bh(&lgr->conns_lock);
 623         }
 624         spin_unlock_bh(&smc_lgr_list.lock);
 625
 626         if (role == SMC_CLNT && !srv_first_contact &&
 627             (local_contact == SMC_FIRST_CONTACT)) {
 628                 /* Server reuses a link group, but Client wants to start
 629                  * a new one
 630                  * send out_of_sync decline, reason synchr. error
 631                  */
 632                 return -ENOLINK;
 633         }
 634
 635 create:
 636         if (local_contact == SMC_FIRST_CONTACT) {
 637                 rc = smc_lgr_create(smc, is_smcd, smcibdev, ibport,
 638                                     lcl->id_for_peer, vlan_id, smcd, peer_gid);
 639                 if (rc)
 640                         goto out;
 641                 smc_lgr_register_conn(conn); /* add smc conn to lgr */
 642         }
 643         conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
 644         conn->local_tx_ctrl.len = SMC_WR_TX_SIZE;
 645         conn->urg_state = SMC_URG_READ;
 646         if (is_smcd) {
 647                 conn->rx_off = sizeof(struct smcd_cdc_msg);
 648                 smcd_cdc_rx_init(conn); /* init tasklet for this conn */
 649         }
 650 #ifndef KERNEL_HAS_ATOMIC64
 651         spin_lock_init(&conn->acurs_lock);
 652 #endif
 653
 654 out:
 655         return rc ? rc : local_contact;
 656 }
 657
 658 /* convert the RMB size into the compressed notation - minimum 16K.
 659  * In contrast to plain ilog2, this rounds towards the next power of 2,
 660  * so the socket application gets at least its desired sndbuf / rcvbuf size.
 661  */
 662 static u8 smc_compress_bufsize(int size)
 663 {
 664         u8 compressed;
 665
 666         if (size <= SMC_BUF_MIN_SIZE)
 667                 return 0;
 668
 669         size = (size - 1) >> 14;
 670         compressed = ilog2(size) + 1;
 671         if (compressed >= SMC_RMBE_SIZES)
 672                 compressed = SMC_RMBE_SIZES - 1;
 673         return compressed;
 674 }
 675
 676 /* convert the RMB size from compressed notation into integer */
 677 int smc_uncompress_bufsize(u8 compressed)
 678 {
 679         u32 size;
 680
 681         size = 0x00000001 << (((int)compressed) + 14);
 682         return (int)size;
 683 }
 684
 685 /* try to reuse a sndbuf or rmb description slot for a certain
 686  * buffer size; if not available, return NULL
 687  */
 688 static struct smc_buf_desc *smc_buf_get_slot(int compressed_bufsize,
 689                                              rwlock_t *lock,
 690                                              struct list_head *buf_list)
 691 {
 692         struct smc_buf_desc *buf_slot;
 693
 694         read_lock_bh(lock);
 695         list_for_each_entry(buf_slot, buf_list, list) {
 696                 if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
 697                         read_unlock_bh(lock);
 698                         return buf_slot;
 699                 }
 700         }
 701         read_unlock_bh(lock);
 702         return NULL;
 703 }
 704
 705 /* one of the conditions for announcing a receiver's current window size is
 706  * that it "results in a minimum increase in the window size of 10% of the
 707  * receive buffer space" [RFC7609]
 708  */
 709 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
 710 {
 711         return min_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
 712 }
 713
 714 static struct smc_buf_desc *smcr_new_buf_create(struct smc_link_group *lgr,
 715                                                 bool is_rmb, int bufsize)
 716 {
 717         struct smc_buf_desc *buf_desc;
 718         struct smc_link *lnk;
 719         int rc;
 720
 721         /* try to alloc a new buffer */
 722         buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
 723         if (!buf_desc)
 724                 return ERR_PTR(-ENOMEM);
 725
 726         buf_desc->order = get_order(bufsize);
 727         buf_desc->pages = alloc_pages(GFP_KERNEL | __GFP_NOWARN |
 728                                       __GFP_NOMEMALLOC | __GFP_COMP |
 729                                       __GFP_NORETRY | __GFP_ZERO,
 730                                       buf_desc->order);
 731         if (!buf_desc->pages) {
 732                 kfree(buf_desc);
 733                 return ERR_PTR(-EAGAIN);
 734         }
 735         buf_desc->cpu_addr = (void *)page_address(buf_desc->pages);
 736
 737         /* build the sg table from the pages */
 738         lnk = &lgr->lnk[SMC_SINGLE_LINK];
 739         rc = sg_alloc_table(&buf_desc->sgt[SMC_SINGLE_LINK], 1,
 740                             GFP_KERNEL);
 741         if (rc) {
 742                 smc_buf_free(lgr, is_rmb, buf_desc);
 743                 return ERR_PTR(rc);
 744         }
 745         sg_set_buf(buf_desc->sgt[SMC_SINGLE_LINK].sgl,
 746                    buf_desc->cpu_addr, bufsize);
 747
 748         /* map sg table to DMA address */
 749         rc = smc_ib_buf_map_sg(lnk->smcibdev, buf_desc,
 750                                is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
 751         /* SMC protocol depends on mapping to one DMA address only */
 752         if (rc != 1)  {
 753                 smc_buf_free(lgr, is_rmb, buf_desc);
 754                 return ERR_PTR(-EAGAIN);
 755         }
 756
 757         /* create a new memory region for the RMB */
 758         if (is_rmb) {
 759                 rc = smc_ib_get_memory_region(lnk->roce_pd,
 760                                               IB_ACCESS_REMOTE_WRITE |
 761                                               IB_ACCESS_LOCAL_WRITE,
 762                                               buf_desc);
 763                 if (rc) {
 764                         smc_buf_free(lgr, is_rmb, buf_desc);
 765                         return ERR_PTR(rc);
 766                 }
 767         }
 768
 769         buf_desc->len = bufsize;
 770         return buf_desc;
 771 }
 772
 773 #define SMCD_DMBE_SIZES         6 /* 0 -> 16KB, 1 -> 32KB, .. 6 -> 1MB */
 774
 775 static struct smc_buf_desc *smcd_new_buf_create(struct smc_link_group *lgr,
 776                                                 bool is_dmb, int bufsize)
 777 {
 778         struct smc_buf_desc *buf_desc;
 779         int rc;
 780
 781         if (smc_compress_bufsize(bufsize) > SMCD_DMBE_SIZES)
 782                 return ERR_PTR(-EAGAIN);
 783
 784         /* try to alloc a new DMB */
 785         buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
 786         if (!buf_desc)
 787                 return ERR_PTR(-ENOMEM);
 788         if (is_dmb) {
 789                 rc = smc_ism_register_dmb(lgr, bufsize, buf_desc);
 790                 if (rc) {
 791                         kfree(buf_desc);
 792                         return ERR_PTR(-EAGAIN);
 793                 }
 794                 buf_desc->pages = virt_to_page(buf_desc->cpu_addr);
 795                 /* CDC header stored in buf. So, pretend it was smaller */
 796                 buf_desc->len = bufsize - sizeof(struct smcd_cdc_msg);
 797         } else {
 798                 buf_desc->cpu_addr = kzalloc(bufsize, GFP_KERNEL |
 799                                              __GFP_NOWARN | __GFP_NORETRY |
 800                                              __GFP_NOMEMALLOC);
 801                 if (!buf_desc->cpu_addr) {
 802                         kfree(buf_desc);
 803                         return ERR_PTR(-EAGAIN);
 804                 }
 805                 buf_desc->len = bufsize;
 806         }
 807         return buf_desc;
 808 }
 809
 810 static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
 811 {
 812         struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
 813         struct smc_connection *conn = &smc->conn;
 814         struct smc_link_group *lgr = conn->lgr;
 815         struct list_head *buf_list;
 816         int bufsize, bufsize_short;
 817         int sk_buf_size;
 818         rwlock_t *lock;
 819
 820         if (is_rmb)
 821                 /* use socket recv buffer size (w/o overhead) as start value */
 822                 sk_buf_size = smc->sk.sk_rcvbuf / 2;
 823         else
 824                 /* use socket send buffer size (w/o overhead) as start value */
 825                 sk_buf_size = smc->sk.sk_sndbuf / 2;
 826
 827         for (bufsize_short = smc_compress_bufsize(sk_buf_size);
 828              bufsize_short >= 0; bufsize_short--) {
 829
 830                 if (is_rmb) {
 831                         lock = &lgr->rmbs_lock;
 832                         buf_list = &lgr->rmbs[bufsize_short];
 833                 } else {
 834                         lock = &lgr->sndbufs_lock;
 835                         buf_list = &lgr->sndbufs[bufsize_short];
 836                 }
 837                 bufsize = smc_uncompress_bufsize(bufsize_short);
 838                 if ((1 << get_order(bufsize)) > SG_MAX_SINGLE_ALLOC)
 839                         continue;
 840
 841                 /* check for reusable slot in the link group */
 842                 buf_desc = smc_buf_get_slot(bufsize_short, lock, buf_list);
 843                 if (buf_desc) {
 844                         memset(buf_desc->cpu_addr, 0, bufsize);
 845                         break; /* found reusable slot */
 846                 }
 847
 848                 if (is_smcd)
 849                         buf_desc = smcd_new_buf_create(lgr, is_rmb, bufsize);
 850                 else
 851                         buf_desc = smcr_new_buf_create(lgr, is_rmb, bufsize);
 852
 853                 if (PTR_ERR(buf_desc) == -ENOMEM)
 854                         break;
 855                 if (IS_ERR(buf_desc))
 856                         continue;
 857
 858                 buf_desc->used = 1;
 859                 write_lock_bh(lock);
 860                 list_add(&buf_desc->list, buf_list);
 861                 write_unlock_bh(lock);
 862                 break; /* found */
 863         }
 864
 865         if (IS_ERR(buf_desc))
 866                 return -ENOMEM;
 867
 868         if (is_rmb) {
 869                 conn->rmb_desc = buf_desc;
 870                 conn->rmbe_size_short = bufsize_short;
 871                 smc->sk.sk_rcvbuf = bufsize * 2;
 872                 atomic_set(&conn->bytes_to_rcv, 0);
 873                 conn->rmbe_update_limit =
 874                         smc_rmb_wnd_update_limit(buf_desc->len);
 875                 if (is_smcd)
 876                         smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */
 877         } else {
 878                 conn->sndbuf_desc = buf_desc;
 879                 smc->sk.sk_sndbuf = bufsize * 2;
 880                 atomic_set(&conn->sndbuf_space, bufsize);
 881         }
 882         return 0;
 883 }
 884
 885 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn)
 886 {
 887         struct smc_link_group *lgr = conn->lgr;
 888
 889         if (!conn->lgr || conn->lgr->is_smcd)
 890                 return;
 891         smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 892                                conn->sndbuf_desc, DMA_TO_DEVICE);
 893 }
 894
 895 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
 896 {
 897         struct smc_link_group *lgr = conn->lgr;
 898
 899         if (!conn->lgr || conn->lgr->is_smcd)
 900                 return;
 901         smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 902                                   conn->sndbuf_desc, DMA_TO_DEVICE);
 903 }
 904
 905 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
 906 {
 907         struct smc_link_group *lgr = conn->lgr;
 908
 909         if (!conn->lgr || conn->lgr->is_smcd)
 910                 return;
 911         smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 912                                conn->rmb_desc, DMA_FROM_DEVICE);
 913 }
 914
 915 void smc_rmb_sync_sg_for_device(struct smc_connection *conn)
 916 {
 917         struct smc_link_group *lgr = conn->lgr;
 918
 919         if (!conn->lgr || conn->lgr->is_smcd)
 920                 return;
 921         smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 922                                   conn->rmb_desc, DMA_FROM_DEVICE);
 923 }
 924
 925 /* create the send and receive buffer for an SMC socket;
 926  * receive buffers are called RMBs;
 927  * (even though the SMC protocol allows more than one RMB-element per RMB,
 928  * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
 929  * extra RMB for every connection in a link group
 930  */
 931 int smc_buf_create(struct smc_sock *smc, bool is_smcd)
 932 {
 933         int rc;
 934
 935         /* create send buffer */
 936         rc = __smc_buf_create(smc, is_smcd, false);
 937         if (rc)
 938                 return rc;
 939         /* create rmb */
 940         rc = __smc_buf_create(smc, is_smcd, true);
 941         if (rc)
 942                 smc_buf_free(smc->conn.lgr, false, smc->conn.sndbuf_desc);
 943         return rc;
 944 }
 945
 946 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
 947 {
 948         int i;
 949
 950         for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
 951                 if (!test_and_set_bit(i, lgr->rtokens_used_mask))
 952                         return i;
 953         }
 954         return -ENOSPC;
 955 }
 956
 957 /* add a new rtoken from peer */
 958 int smc_rtoken_add(struct smc_link_group *lgr, __be64 nw_vaddr, __be32 nw_rkey)
 959 {
 960         u64 dma_addr = be64_to_cpu(nw_vaddr);
 961         u32 rkey = ntohl(nw_rkey);
 962         int i;
 963
 964         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
 965                 if ((lgr->rtokens[i][SMC_SINGLE_LINK].rkey == rkey) &&
 966                     (lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr == dma_addr) &&
 967                     test_bit(i, lgr->rtokens_used_mask)) {
 968                         /* already in list */
 969                         return i;
 970                 }
 971         }
 972         i = smc_rmb_reserve_rtoken_idx(lgr);
 973         if (i < 0)
 974                 return i;
 975         lgr->rtokens[i][SMC_SINGLE_LINK].rkey = rkey;
 976         lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr = dma_addr;
 977         return i;
 978 }
 979
 980 /* delete an rtoken */
 981 int smc_rtoken_delete(struct smc_link_group *lgr, __be32 nw_rkey)
 982 {
 983         u32 rkey = ntohl(nw_rkey);
 984         int i;
 985
 986         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
 987                 if (lgr->rtokens[i][SMC_SINGLE_LINK].rkey == rkey &&
 988                     test_bit(i, lgr->rtokens_used_mask)) {
 989                         lgr->rtokens[i][SMC_SINGLE_LINK].rkey = 0;
 990                         lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr = 0;
 991
 992                         clear_bit(i, lgr->rtokens_used_mask);
 993                         return 0;
 994                 }
 995         }
 996         return -ENOENT;
 997 }
 998
 999 /* save rkey and dma_addr received from peer during clc handshake */
1000 int smc_rmb_rtoken_handling(struct smc_connection *conn,
1001                             struct smc_clc_msg_accept_confirm *clc)
1002 {
1003         conn->rtoken_idx = smc_rtoken_add(conn->lgr, clc->rmb_dma_addr,
1004                                           clc->rmb_rkey);
1005         if (conn->rtoken_idx < 0)
1006                 return conn->rtoken_idx;
1007         return 0;
1008 }
1009
1010 /* Called (from smc_exit) when module is removed */
1011 void smc_core_exit(void)
1012 {
1013         struct smc_link_group *lgr, *lg;
1014         LIST_HEAD(lgr_freeing_list);
1015
1016         spin_lock_bh(&smc_lgr_list.lock);
1017         if (!list_empty(&smc_lgr_list.list))
1018                 list_splice_init(&smc_lgr_list.list, &lgr_freeing_list);
1019         spin_unlock_bh(&smc_lgr_list.lock);
1020         list_for_each_entry_safe(lgr, lg, &lgr_freeing_list, list) {
1021                 list_del_init(&lgr->list);
1022                 if (!lgr->is_smcd) {
1023                         struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
1024
1025                         if (lnk->state == SMC_LNK_ACTIVE)
1026                                 smc_llc_send_delete_link(lnk, SMC_LLC_REQ,
1027                                                          false);
1028                         smc_llc_link_inactive(lnk);
1029                 }
1030                 cancel_delayed_work_sync(&lgr->free_work);
1031                 smc_lgr_free(lgr); /* free link group */
1032         }
1033 }