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         write_lock_bh(&lgr->conns_lock);
 132         if (conn->alert_token_local) {
 133                 __smc_lgr_unregister_conn(conn);
 134         }
 135         write_unlock_bh(&lgr->conns_lock);
 136 }
 137
 138 /* Send delete link, either as client to request the initiation
 139  * of the DELETE LINK sequence from server; or as server to
 140  * initiate the delete processing. See smc_llc_rx_delete_link().
 141  */
 142 static int smc_link_send_delete(struct smc_link *lnk)
 143 {
 144         if (lnk->state == SMC_LNK_ACTIVE &&
 145             !smc_llc_send_delete_link(lnk, SMC_LLC_REQ, true)) {
 146                 smc_llc_link_deleting(lnk);
 147                 return 0;
 148         }
 149         return -ENOTCONN;
 150 }
 151
 152 static void smc_lgr_free_work(struct work_struct *work)
 153 {
 154         struct smc_link_group *lgr = container_of(to_delayed_work(work),
 155                                                   struct smc_link_group,
 156                                                   free_work);
 157         bool conns;
 158
 159         spin_lock_bh(&smc_lgr_list.lock);
 160         if (list_empty(&lgr->list))
 161                 goto free;
 162         read_lock_bh(&lgr->conns_lock);
 163         conns = RB_EMPTY_ROOT(&lgr->conns_all);
 164         read_unlock_bh(&lgr->conns_lock);
 165         if (!conns) { /* number of lgr connections is no longer zero */
 166                 spin_unlock_bh(&smc_lgr_list.lock);
 167                 return;
 168         }
 169         list_del_init(&lgr->list); /* remove from smc_lgr_list */
 170 free:
 171         spin_unlock_bh(&smc_lgr_list.lock);
 172
 173         if (!lgr->is_smcd && !lgr->terminating) {
 174                 /* try to send del link msg, on error free lgr immediately */
 175                 if (!smc_link_send_delete(&lgr->lnk[SMC_SINGLE_LINK])) {
 176                         /* reschedule in case we never receive a response */
 177                         smc_lgr_schedule_free_work(lgr);
 178                         return;
 179                 }
 180         }
 181
 182         if (!delayed_work_pending(&lgr->free_work)) {
 183                 struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
 184
 185                 if (!lgr->is_smcd && lnk->state != SMC_LNK_INACTIVE)
 186                         smc_llc_link_inactive(lnk);
 187                 smc_lgr_free(lgr);
 188         }
 189 }
 190
 191 /* create a new SMC link group */
 192 static int smc_lgr_create(struct smc_sock *smc, bool is_smcd,
 193                           struct smc_ib_device *smcibdev, u8 ibport,
 194                           char *peer_systemid, unsigned short vlan_id,
 195                           struct smcd_dev *smcismdev, u64 peer_gid)
 196 {
 197         struct smc_link_group *lgr;
 198         struct smc_link *lnk;
 199         u8 rndvec[3];
 200         int rc = 0;
 201         int i;
 202
 203         if (is_smcd && vlan_id) {
 204                 rc = smc_ism_get_vlan(smcismdev, vlan_id);
 205                 if (rc)
 206                         goto out;
 207         }
 208
 209         lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
 210         if (!lgr) {
 211                 rc = -ENOMEM;
 212                 goto out;
 213         }
 214         lgr->is_smcd = is_smcd;
 215         lgr->sync_err = 0;
 216         lgr->vlan_id = vlan_id;
 217         rwlock_init(&lgr->sndbufs_lock);
 218         rwlock_init(&lgr->rmbs_lock);
 219         rwlock_init(&lgr->conns_lock);
 220         for (i = 0; i < SMC_RMBE_SIZES; i++) {
 221                 INIT_LIST_HEAD(&lgr->sndbufs[i]);
 222                 INIT_LIST_HEAD(&lgr->rmbs[i]);
 223         }
 224         smc_lgr_list.num += SMC_LGR_NUM_INCR;
 225         memcpy(&lgr->id, (u8 *)&smc_lgr_list.num, SMC_LGR_ID_SIZE);
 226         INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
 227         lgr->conns_all = RB_ROOT;
 228         if (is_smcd) {
 229                 /* SMC-D specific settings */
 230                 lgr->peer_gid = peer_gid;
 231                 lgr->smcd = smcismdev;
 232         } else {
 233                 /* SMC-R specific settings */
 234                 lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
 235                 memcpy(lgr->peer_systemid, peer_systemid, SMC_SYSTEMID_LEN);
 236
 237                 lnk = &lgr->lnk[SMC_SINGLE_LINK];
 238                 /* initialize link */
 239                 lnk->state = SMC_LNK_ACTIVATING;
 240                 lnk->link_id = SMC_SINGLE_LINK;
 241                 lnk->smcibdev = smcibdev;
 242                 lnk->ibport = ibport;
 243                 lnk->path_mtu = smcibdev->pattr[ibport - 1].active_mtu;
 244                 if (!smcibdev->initialized)
 245                         smc_ib_setup_per_ibdev(smcibdev);
 246                 get_random_bytes(rndvec, sizeof(rndvec));
 247                 lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) +
 248                         (rndvec[2] << 16);
 249                 rc = smc_ib_determine_gid(lnk->smcibdev, lnk->ibport,
 250                                           vlan_id, lnk->gid, &lnk->sgid_index);
 251                 if (rc)
 252                         goto free_lgr;
 253                 rc = smc_llc_link_init(lnk);
 254                 if (rc)
 255                         goto free_lgr;
 256                 rc = smc_wr_alloc_link_mem(lnk);
 257                 if (rc)
 258                         goto clear_llc_lnk;
 259                 rc = smc_ib_create_protection_domain(lnk);
 260                 if (rc)
 261                         goto free_link_mem;
 262                 rc = smc_ib_create_queue_pair(lnk);
 263                 if (rc)
 264                         goto dealloc_pd;
 265                 rc = smc_wr_create_link(lnk);
 266                 if (rc)
 267                         goto destroy_qp;
 268         }
 269         smc->conn.lgr = lgr;
 270         spin_lock_bh(&smc_lgr_list.lock);
 271         list_add(&lgr->list, &smc_lgr_list.list);
 272         spin_unlock_bh(&smc_lgr_list.lock);
 273         return 0;
 274
 275 destroy_qp:
 276         smc_ib_destroy_queue_pair(lnk);
 277 dealloc_pd:
 278         smc_ib_dealloc_protection_domain(lnk);
 279 free_link_mem:
 280         smc_wr_free_link_mem(lnk);
 281 clear_llc_lnk:
 282         smc_llc_link_clear(lnk);
 283 free_lgr:
 284         kfree(lgr);
 285 out:
 286         return rc;
 287 }
 288
 289 static void smc_buf_unuse(struct smc_connection *conn,
 290                           struct smc_link_group *lgr)
 291 {
 292         if (conn->sndbuf_desc)
 293                 conn->sndbuf_desc->used = 0;
 294         if (conn->rmb_desc) {
 295                 if (!conn->rmb_desc->regerr) {
 296                         conn->rmb_desc->reused = 1;
 297                         conn->rmb_desc->used = 0;
 298                 } else {
 299                         /* buf registration failed, reuse not possible */
 300                         write_lock_bh(&lgr->rmbs_lock);
 301                         list_del(&conn->rmb_desc->list);
 302                         write_unlock_bh(&lgr->rmbs_lock);
 303
 304                         smc_buf_free(lgr, true, conn->rmb_desc);
 305                 }
 306         }
 307 }
 308
 309 /* remove a finished connection from its link group */
 310 void smc_conn_free(struct smc_connection *conn)
 311 {
 312         struct smc_link_group *lgr = conn->lgr;
 313
 314         if (!lgr)
 315                 return;
 316         if (lgr->is_smcd) {
 317                 smc_ism_unset_conn(conn);
 318                 tasklet_kill(&conn->rx_tsklet);
 319         } else {
 320                 smc_cdc_tx_dismiss_slots(conn);
 321         }
 322         smc_lgr_unregister_conn(conn);          /* unsets conn->lgr */
 323         smc_buf_unuse(conn, lgr);               /* allow buffer reuse */
 324
 325         if (!lgr->conns_num)
 326                 smc_lgr_schedule_free_work(lgr);
 327 }
 328
 329 static void smc_link_clear(struct smc_link *lnk)
 330 {
 331         lnk->peer_qpn = 0;
 332         smc_llc_link_clear(lnk);
 333         smc_ib_modify_qp_reset(lnk);
 334         smc_wr_free_link(lnk);
 335         smc_ib_destroy_queue_pair(lnk);
 336         smc_ib_dealloc_protection_domain(lnk);
 337         smc_wr_free_link_mem(lnk);
 338 }
 339
 340 static void smcr_buf_free(struct smc_link_group *lgr, bool is_rmb,
 341                           struct smc_buf_desc *buf_desc)
 342 {
 343         struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
 344
 345         if (is_rmb) {
 346                 if (buf_desc->mr_rx[SMC_SINGLE_LINK])
 347                         smc_ib_put_memory_region(
 348                                         buf_desc->mr_rx[SMC_SINGLE_LINK]);
 349                 smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
 350                                     DMA_FROM_DEVICE);
 351         } else {
 352                 smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
 353                                     DMA_TO_DEVICE);
 354         }
 355         sg_free_table(&buf_desc->sgt[SMC_SINGLE_LINK]);
 356         if (buf_desc->pages)
 357                 __free_pages(buf_desc->pages, buf_desc->order);
 358         kfree(buf_desc);
 359 }
 360
 361 static void smcd_buf_free(struct smc_link_group *lgr, bool is_dmb,
 362                           struct smc_buf_desc *buf_desc)
 363 {
 364         if (is_dmb) {
 365                 /* restore original buf len */
 366                 buf_desc->len += sizeof(struct smcd_cdc_msg);
 367                 smc_ism_unregister_dmb(lgr->smcd, buf_desc);
 368         } else {
 369                 kfree(buf_desc->cpu_addr);
 370         }
 371         kfree(buf_desc);
 372 }
 373
 374 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
 375                          struct smc_buf_desc *buf_desc)
 376 {
 377         if (lgr->is_smcd)
 378                 smcd_buf_free(lgr, is_rmb, buf_desc);
 379         else
 380                 smcr_buf_free(lgr, is_rmb, buf_desc);
 381 }
 382
 383 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
 384 {
 385         struct smc_buf_desc *buf_desc, *bf_desc;
 386         struct list_head *buf_list;
 387         int i;
 388
 389         for (i = 0; i < SMC_RMBE_SIZES; i++) {
 390                 if (is_rmb)
 391                         buf_list = &lgr->rmbs[i];
 392                 else
 393                         buf_list = &lgr->sndbufs[i];
 394                 list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
 395                                          list) {
 396                         list_del(&buf_desc->list);
 397                         smc_buf_free(lgr, is_rmb, buf_desc);
 398                 }
 399         }
 400 }
 401
 402 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
 403 {
 404         /* free send buffers */
 405         __smc_lgr_free_bufs(lgr, false);
 406         /* free rmbs */
 407         __smc_lgr_free_bufs(lgr, true);
 408 }
 409
 410 /* remove a link group */
 411 void smc_lgr_free(struct smc_link_group *lgr)
 412 {
 413         smc_lgr_free_bufs(lgr);
 414         if (lgr->is_smcd)
 415                 smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
 416         else
 417                 smc_link_clear(&lgr->lnk[SMC_SINGLE_LINK]);
 418         kfree(lgr);
 419 }
 420
 421 void smc_lgr_forget(struct smc_link_group *lgr)
 422 {
 423         spin_lock_bh(&smc_lgr_list.lock);
 424         /* do not use this link group for new connections */
 425         if (!list_empty(&lgr->list))
 426                 list_del_init(&lgr->list);
 427         spin_unlock_bh(&smc_lgr_list.lock);
 428 }
 429
 430 /* terminate linkgroup abnormally */
 431 static void __smc_lgr_terminate(struct smc_link_group *lgr)
 432 {
 433         struct smc_connection *conn;
 434         struct smc_sock *smc;
 435         struct rb_node *node;
 436
 437         if (lgr->terminating)
 438                 return; /* lgr already terminating */
 439         lgr->terminating = 1;
 440         if (!list_empty(&lgr->list)) /* forget lgr */
 441                 list_del_init(&lgr->list);
 442         if (!lgr->is_smcd)
 443                 smc_llc_link_inactive(&lgr->lnk[SMC_SINGLE_LINK]);
 444
 445         write_lock_bh(&lgr->conns_lock);
 446         node = rb_first(&lgr->conns_all);
 447         while (node) {
 448                 conn = rb_entry(node, struct smc_connection, alert_node);
 449                 smc = container_of(conn, struct smc_sock, conn);
 450                 sock_hold(&smc->sk); /* sock_put in close work */
 451                 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
 452                 __smc_lgr_unregister_conn(conn);
 453                 write_unlock_bh(&lgr->conns_lock);
 454                 if (!schedule_work(&conn->close_work))
 455                         sock_put(&smc->sk);
 456                 write_lock_bh(&lgr->conns_lock);
 457                 node = rb_first(&lgr->conns_all);
 458         }
 459         write_unlock_bh(&lgr->conns_lock);
 460         if (!lgr->is_smcd)
 461                 wake_up(&lgr->lnk[SMC_SINGLE_LINK].wr_reg_wait);
 462         smc_lgr_schedule_free_work(lgr);
 463 }
 464
 465 void smc_lgr_terminate(struct smc_link_group *lgr)
 466 {
 467         spin_lock_bh(&smc_lgr_list.lock);
 468         __smc_lgr_terminate(lgr);
 469         spin_unlock_bh(&smc_lgr_list.lock);
 470 }
 471
 472 /* Called when IB port is terminated */
 473 void smc_port_terminate(struct smc_ib_device *smcibdev, u8 ibport)
 474 {
 475         struct smc_link_group *lgr, *l;
 476
 477         spin_lock_bh(&smc_lgr_list.lock);
 478         list_for_each_entry_safe(lgr, l, &smc_lgr_list.list, list) {
 479                 if (!lgr->is_smcd &&
 480                     lgr->lnk[SMC_SINGLE_LINK].smcibdev == smcibdev &&
 481                     lgr->lnk[SMC_SINGLE_LINK].ibport == ibport)
 482                         __smc_lgr_terminate(lgr);
 483         }
 484         spin_unlock_bh(&smc_lgr_list.lock);
 485 }
 486
 487 /* Called when SMC-D device is terminated or peer is lost */
 488 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid)
 489 {
 490         struct smc_link_group *lgr, *l;
 491         LIST_HEAD(lgr_free_list);
 492
 493         /* run common cleanup function and build free list */
 494         spin_lock_bh(&smc_lgr_list.lock);
 495         list_for_each_entry_safe(lgr, l, &smc_lgr_list.list, list) {
 496                 if (lgr->is_smcd && lgr->smcd == dev &&
 497                     (!peer_gid || lgr->peer_gid == peer_gid) &&
 498                     !list_empty(&lgr->list)) {
 499                         __smc_lgr_terminate(lgr);
 500                         list_move(&lgr->list, &lgr_free_list);
 501                 }
 502         }
 503         spin_unlock_bh(&smc_lgr_list.lock);
 504
 505         /* cancel the regular free workers and actually free lgrs */
 506         list_for_each_entry_safe(lgr, l, &lgr_free_list, list) {
 507                 list_del_init(&lgr->list);
 508                 cancel_delayed_work_sync(&lgr->free_work);
 509                 smc_lgr_free(lgr);
 510         }
 511 }
 512
 513 /* Determine vlan of internal TCP socket.
 514  * @vlan_id: address to store the determined vlan id into
 515  */
 516 int smc_vlan_by_tcpsk(struct socket *clcsock, unsigned short *vlan_id)
 517 {
 518         struct dst_entry *dst = sk_dst_get(clcsock->sk);
 519         struct net_device *ndev;
 520         int i, nest_lvl, rc = 0;
 521
 522         *vlan_id = 0;
 523         if (!dst) {
 524                 rc = -ENOTCONN;
 525                 goto out;
 526         }
 527         if (!dst->dev) {
 528                 rc = -ENODEV;
 529                 goto out_rel;
 530         }
 531
 532         ndev = dst->dev;
 533         if (is_vlan_dev(ndev)) {
 534                 *vlan_id = vlan_dev_vlan_id(ndev);
 535                 goto out_rel;
 536         }
 537
 538         rtnl_lock();
 539         nest_lvl = dev_get_nest_level(ndev);
 540         for (i = 0; i < nest_lvl; i++) {
 541                 struct list_head *lower = &ndev->adj_list.lower;
 542
 543                 if (list_empty(lower))
 544                         break;
 545                 lower = lower->next;
 546                 ndev = (struct net_device *)netdev_lower_get_next(ndev, &lower);
 547                 if (is_vlan_dev(ndev)) {
 548                         *vlan_id = vlan_dev_vlan_id(ndev);
 549                         break;
 550                 }
 551         }
 552         rtnl_unlock();
 553
 554 out_rel:
 555         dst_release(dst);
 556 out:
 557         return rc;
 558 }
 559
 560 static bool smcr_lgr_match(struct smc_link_group *lgr,
 561                            struct smc_clc_msg_local *lcl,
 562                            enum smc_lgr_role role)
 563 {
 564         return !memcmp(lgr->peer_systemid, lcl->id_for_peer,
 565                        SMC_SYSTEMID_LEN) &&
 566                 !memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_gid, &lcl->gid,
 567                         SMC_GID_SIZE) &&
 568                 !memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_mac, lcl->mac,
 569                         sizeof(lcl->mac)) &&
 570                 lgr->role == role;
 571 }
 572
 573 static bool smcd_lgr_match(struct smc_link_group *lgr,
 574                            struct smcd_dev *smcismdev, u64 peer_gid)
 575 {
 576         return lgr->peer_gid == peer_gid && lgr->smcd == smcismdev;
 577 }
 578
 579 /* create a new SMC connection (and a new link group if necessary) */
 580 int smc_conn_create(struct smc_sock *smc, bool is_smcd, int srv_first_contact,
 581                     struct smc_ib_device *smcibdev, u8 ibport,
 582                     struct smc_clc_msg_local *lcl, struct smcd_dev *smcd,
 583                     u64 peer_gid)
 584 {
 585         struct smc_connection *conn = &smc->conn;
 586         int local_contact = SMC_FIRST_CONTACT;
 587         struct smc_link_group *lgr;
 588         unsigned short vlan_id;
 589         enum smc_lgr_role role;
 590         int rc = 0;
 591
 592         role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
 593         rc = smc_vlan_by_tcpsk(smc->clcsock, &vlan_id);
 594         if (rc)
 595                 return rc;
 596
 597         if ((role == SMC_CLNT) && srv_first_contact)
 598                 /* create new link group as well */
 599                 goto create;
 600
 601         /* determine if an existing link group can be reused */
 602         spin_lock_bh(&smc_lgr_list.lock);
 603         list_for_each_entry(lgr, &smc_lgr_list.list, list) {
 604                 write_lock_bh(&lgr->conns_lock);
 605                 if ((is_smcd ? smcd_lgr_match(lgr, smcd, peer_gid) :
 606                      smcr_lgr_match(lgr, lcl, role)) &&
 607                     !lgr->sync_err &&
 608                     lgr->vlan_id == vlan_id &&
 609                     (role == SMC_CLNT ||
 610                      lgr->conns_num < SMC_RMBS_PER_LGR_MAX)) {
 611                         /* link group found */
 612                         local_contact = SMC_REUSE_CONTACT;
 613                         conn->lgr = lgr;
 614                         smc_lgr_register_conn(conn); /* add smc conn to lgr */
 615                         write_unlock_bh(&lgr->conns_lock);
 616                         break;
 617                 }
 618                 write_unlock_bh(&lgr->conns_lock);
 619         }
 620         spin_unlock_bh(&smc_lgr_list.lock);
 621
 622         if (role == SMC_CLNT && !srv_first_contact &&
 623             (local_contact == SMC_FIRST_CONTACT)) {
 624                 /* Server reuses a link group, but Client wants to start
 625                  * a new one
 626                  * send out_of_sync decline, reason synchr. error
 627                  */
 628                 return -ENOLINK;
 629         }
 630
 631 create:
 632         if (local_contact == SMC_FIRST_CONTACT) {
 633                 rc = smc_lgr_create(smc, is_smcd, smcibdev, ibport,
 634                                     lcl->id_for_peer, vlan_id, smcd, peer_gid);
 635                 if (rc)
 636                         goto out;
 637                 smc_lgr_register_conn(conn); /* add smc conn to lgr */
 638         }
 639         conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
 640         conn->local_tx_ctrl.len = SMC_WR_TX_SIZE;
 641         conn->urg_state = SMC_URG_READ;
 642         if (is_smcd) {
 643                 conn->rx_off = sizeof(struct smcd_cdc_msg);
 644                 smcd_cdc_rx_init(conn); /* init tasklet for this conn */
 645         }
 646 #ifndef KERNEL_HAS_ATOMIC64
 647         spin_lock_init(&conn->acurs_lock);
 648 #endif
 649
 650 out:
 651         return rc ? rc : local_contact;
 652 }
 653
 654 /* convert the RMB size into the compressed notation - minimum 16K.
 655  * In contrast to plain ilog2, this rounds towards the next power of 2,
 656  * so the socket application gets at least its desired sndbuf / rcvbuf size.
 657  */
 658 static u8 smc_compress_bufsize(int size)
 659 {
 660         u8 compressed;
 661
 662         if (size <= SMC_BUF_MIN_SIZE)
 663                 return 0;
 664
 665         size = (size - 1) >> 14;
 666         compressed = ilog2(size) + 1;
 667         if (compressed >= SMC_RMBE_SIZES)
 668                 compressed = SMC_RMBE_SIZES - 1;
 669         return compressed;
 670 }
 671
 672 /* convert the RMB size from compressed notation into integer */
 673 int smc_uncompress_bufsize(u8 compressed)
 674 {
 675         u32 size;
 676
 677         size = 0x00000001 << (((int)compressed) + 14);
 678         return (int)size;
 679 }
 680
 681 /* try to reuse a sndbuf or rmb description slot for a certain
 682  * buffer size; if not available, return NULL
 683  */
 684 static struct smc_buf_desc *smc_buf_get_slot(int compressed_bufsize,
 685                                              rwlock_t *lock,
 686                                              struct list_head *buf_list)
 687 {
 688         struct smc_buf_desc *buf_slot;
 689
 690         read_lock_bh(lock);
 691         list_for_each_entry(buf_slot, buf_list, list) {
 692                 if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
 693                         read_unlock_bh(lock);
 694                         return buf_slot;
 695                 }
 696         }
 697         read_unlock_bh(lock);
 698         return NULL;
 699 }
 700
 701 /* one of the conditions for announcing a receiver's current window size is
 702  * that it "results in a minimum increase in the window size of 10% of the
 703  * receive buffer space" [RFC7609]
 704  */
 705 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
 706 {
 707         return min_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
 708 }
 709
 710 static struct smc_buf_desc *smcr_new_buf_create(struct smc_link_group *lgr,
 711                                                 bool is_rmb, int bufsize)
 712 {
 713         struct smc_buf_desc *buf_desc;
 714         struct smc_link *lnk;
 715         int rc;
 716
 717         /* try to alloc a new buffer */
 718         buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
 719         if (!buf_desc)
 720                 return ERR_PTR(-ENOMEM);
 721
 722         buf_desc->order = get_order(bufsize);
 723         buf_desc->pages = alloc_pages(GFP_KERNEL | __GFP_NOWARN |
 724                                       __GFP_NOMEMALLOC | __GFP_COMP |
 725                                       __GFP_NORETRY | __GFP_ZERO,
 726                                       buf_desc->order);
 727         if (!buf_desc->pages) {
 728                 kfree(buf_desc);
 729                 return ERR_PTR(-EAGAIN);
 730         }
 731         buf_desc->cpu_addr = (void *)page_address(buf_desc->pages);
 732
 733         /* build the sg table from the pages */
 734         lnk = &lgr->lnk[SMC_SINGLE_LINK];
 735         rc = sg_alloc_table(&buf_desc->sgt[SMC_SINGLE_LINK], 1,
 736                             GFP_KERNEL);
 737         if (rc) {
 738                 smc_buf_free(lgr, is_rmb, buf_desc);
 739                 return ERR_PTR(rc);
 740         }
 741         sg_set_buf(buf_desc->sgt[SMC_SINGLE_LINK].sgl,
 742                    buf_desc->cpu_addr, bufsize);
 743
 744         /* map sg table to DMA address */
 745         rc = smc_ib_buf_map_sg(lnk->smcibdev, buf_desc,
 746                                is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
 747         /* SMC protocol depends on mapping to one DMA address only */
 748         if (rc != 1)  {
 749                 smc_buf_free(lgr, is_rmb, buf_desc);
 750                 return ERR_PTR(-EAGAIN);
 751         }
 752
 753         /* create a new memory region for the RMB */
 754         if (is_rmb) {
 755                 rc = smc_ib_get_memory_region(lnk->roce_pd,
 756                                               IB_ACCESS_REMOTE_WRITE |
 757                                               IB_ACCESS_LOCAL_WRITE,
 758                                               buf_desc);
 759                 if (rc) {
 760                         smc_buf_free(lgr, is_rmb, buf_desc);
 761                         return ERR_PTR(rc);
 762                 }
 763         }
 764
 765         buf_desc->len = bufsize;
 766         return buf_desc;
 767 }
 768
 769 #define SMCD_DMBE_SIZES         7 /* 0 -> 16KB, 1 -> 32KB, .. 6 -> 1MB */
 770
 771 static struct smc_buf_desc *smcd_new_buf_create(struct smc_link_group *lgr,
 772                                                 bool is_dmb, int bufsize)
 773 {
 774         struct smc_buf_desc *buf_desc;
 775         int rc;
 776
 777         if (smc_compress_bufsize(bufsize) > SMCD_DMBE_SIZES)
 778                 return ERR_PTR(-EAGAIN);
 779
 780         /* try to alloc a new DMB */
 781         buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
 782         if (!buf_desc)
 783                 return ERR_PTR(-ENOMEM);
 784         if (is_dmb) {
 785                 rc = smc_ism_register_dmb(lgr, bufsize, buf_desc);
 786                 if (rc) {
 787                         kfree(buf_desc);
 788                         return ERR_PTR(-EAGAIN);
 789                 }
 790                 buf_desc->pages = virt_to_page(buf_desc->cpu_addr);
 791                 /* CDC header stored in buf. So, pretend it was smaller */
 792                 buf_desc->len = bufsize - sizeof(struct smcd_cdc_msg);
 793         } else {
 794                 buf_desc->cpu_addr = kzalloc(bufsize, GFP_KERNEL |
 795                                              __GFP_NOWARN | __GFP_NORETRY |
 796                                              __GFP_NOMEMALLOC);
 797                 if (!buf_desc->cpu_addr) {
 798                         kfree(buf_desc);
 799                         return ERR_PTR(-EAGAIN);
 800                 }
 801                 buf_desc->len = bufsize;
 802         }
 803         return buf_desc;
 804 }
 805
 806 static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
 807 {
 808         struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
 809         struct smc_connection *conn = &smc->conn;
 810         struct smc_link_group *lgr = conn->lgr;
 811         struct list_head *buf_list;
 812         int bufsize, bufsize_short;
 813         int sk_buf_size;
 814         rwlock_t *lock;
 815
 816         if (is_rmb)
 817                 /* use socket recv buffer size (w/o overhead) as start value */
 818                 sk_buf_size = smc->sk.sk_rcvbuf / 2;
 819         else
 820                 /* use socket send buffer size (w/o overhead) as start value */
 821                 sk_buf_size = smc->sk.sk_sndbuf / 2;
 822
 823         for (bufsize_short = smc_compress_bufsize(sk_buf_size);
 824              bufsize_short >= 0; bufsize_short--) {
 825
 826                 if (is_rmb) {
 827                         lock = &lgr->rmbs_lock;
 828                         buf_list = &lgr->rmbs[bufsize_short];
 829                 } else {
 830                         lock = &lgr->sndbufs_lock;
 831                         buf_list = &lgr->sndbufs[bufsize_short];
 832                 }
 833                 bufsize = smc_uncompress_bufsize(bufsize_short);
 834                 if ((1 << get_order(bufsize)) > SG_MAX_SINGLE_ALLOC)
 835                         continue;
 836
 837                 /* check for reusable slot in the link group */
 838                 buf_desc = smc_buf_get_slot(bufsize_short, lock, buf_list);
 839                 if (buf_desc) {
 840                         memset(buf_desc->cpu_addr, 0, bufsize);
 841                         break; /* found reusable slot */
 842                 }
 843
 844                 if (is_smcd)
 845                         buf_desc = smcd_new_buf_create(lgr, is_rmb, bufsize);
 846                 else
 847                         buf_desc = smcr_new_buf_create(lgr, is_rmb, bufsize);
 848
 849                 if (PTR_ERR(buf_desc) == -ENOMEM)
 850                         break;
 851                 if (IS_ERR(buf_desc))
 852                         continue;
 853
 854                 buf_desc->used = 1;
 855                 write_lock_bh(lock);
 856                 list_add(&buf_desc->list, buf_list);
 857                 write_unlock_bh(lock);
 858                 break; /* found */
 859         }
 860
 861         if (IS_ERR(buf_desc))
 862                 return -ENOMEM;
 863
 864         if (is_rmb) {
 865                 conn->rmb_desc = buf_desc;
 866                 conn->rmbe_size_short = bufsize_short;
 867                 smc->sk.sk_rcvbuf = bufsize * 2;
 868                 atomic_set(&conn->bytes_to_rcv, 0);
 869                 conn->rmbe_update_limit =
 870                         smc_rmb_wnd_update_limit(buf_desc->len);
 871                 if (is_smcd)
 872                         smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */
 873         } else {
 874                 conn->sndbuf_desc = buf_desc;
 875                 smc->sk.sk_sndbuf = bufsize * 2;
 876                 atomic_set(&conn->sndbuf_space, bufsize);
 877         }
 878         return 0;
 879 }
 880
 881 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn)
 882 {
 883         struct smc_link_group *lgr = conn->lgr;
 884
 885         if (!conn->lgr || conn->lgr->is_smcd)
 886                 return;
 887         smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 888                                conn->sndbuf_desc, DMA_TO_DEVICE);
 889 }
 890
 891 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
 892 {
 893         struct smc_link_group *lgr = conn->lgr;
 894
 895         if (!conn->lgr || conn->lgr->is_smcd)
 896                 return;
 897         smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 898                                   conn->sndbuf_desc, DMA_TO_DEVICE);
 899 }
 900
 901 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
 902 {
 903         struct smc_link_group *lgr = conn->lgr;
 904
 905         if (!conn->lgr || conn->lgr->is_smcd)
 906                 return;
 907         smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 908                                conn->rmb_desc, DMA_FROM_DEVICE);
 909 }
 910
 911 void smc_rmb_sync_sg_for_device(struct smc_connection *conn)
 912 {
 913         struct smc_link_group *lgr = conn->lgr;
 914
 915         if (!conn->lgr || conn->lgr->is_smcd)
 916                 return;
 917         smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 918                                   conn->rmb_desc, DMA_FROM_DEVICE);
 919 }
 920
 921 /* create the send and receive buffer for an SMC socket;
 922  * receive buffers are called RMBs;
 923  * (even though the SMC protocol allows more than one RMB-element per RMB,
 924  * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
 925  * extra RMB for every connection in a link group
 926  */
 927 int smc_buf_create(struct smc_sock *smc, bool is_smcd)
 928 {
 929         int rc;
 930
 931         /* create send buffer */
 932         rc = __smc_buf_create(smc, is_smcd, false);
 933         if (rc)
 934                 return rc;
 935         /* create rmb */
 936         rc = __smc_buf_create(smc, is_smcd, true);
 937         if (rc)
 938                 smc_buf_free(smc->conn.lgr, false, smc->conn.sndbuf_desc);
 939         return rc;
 940 }
 941
 942 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
 943 {
 944         int i;
 945
 946         for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
 947                 if (!test_and_set_bit(i, lgr->rtokens_used_mask))
 948                         return i;
 949         }
 950         return -ENOSPC;
 951 }
 952
 953 /* add a new rtoken from peer */
 954 int smc_rtoken_add(struct smc_link_group *lgr, __be64 nw_vaddr, __be32 nw_rkey)
 955 {
 956         u64 dma_addr = be64_to_cpu(nw_vaddr);
 957         u32 rkey = ntohl(nw_rkey);
 958         int i;
 959
 960         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
 961                 if ((lgr->rtokens[i][SMC_SINGLE_LINK].rkey == rkey) &&
 962                     (lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr == dma_addr) &&
 963                     test_bit(i, lgr->rtokens_used_mask)) {
 964                         /* already in list */
 965                         return i;
 966                 }
 967         }
 968         i = smc_rmb_reserve_rtoken_idx(lgr);
 969         if (i < 0)
 970                 return i;
 971         lgr->rtokens[i][SMC_SINGLE_LINK].rkey = rkey;
 972         lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr = dma_addr;
 973         return i;
 974 }
 975
 976 /* delete an rtoken */
 977 int smc_rtoken_delete(struct smc_link_group *lgr, __be32 nw_rkey)
 978 {
 979         u32 rkey = ntohl(nw_rkey);
 980         int i;
 981
 982         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
 983                 if (lgr->rtokens[i][SMC_SINGLE_LINK].rkey == rkey &&
 984                     test_bit(i, lgr->rtokens_used_mask)) {
 985                         lgr->rtokens[i][SMC_SINGLE_LINK].rkey = 0;
 986                         lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr = 0;
 987
 988                         clear_bit(i, lgr->rtokens_used_mask);
 989                         return 0;
 990                 }
 991         }
 992         return -ENOENT;
 993 }
 994
 995 /* save rkey and dma_addr received from peer during clc handshake */
 996 int smc_rmb_rtoken_handling(struct smc_connection *conn,
 997                             struct smc_clc_msg_accept_confirm *clc)
 998 {
 999         conn->rtoken_idx = smc_rtoken_add(conn->lgr, clc->rmb_dma_addr,
1000                                           clc->rmb_rkey);
1001         if (conn->rtoken_idx < 0)
1002                 return conn->rtoken_idx;
1003         return 0;
1004 }
1005
1006 /* Called (from smc_exit) when module is removed */
1007 void smc_core_exit(void)
1008 {
1009         struct smc_link_group *lgr, *lg;
1010         LIST_HEAD(lgr_freeing_list);
1011
1012         spin_lock_bh(&smc_lgr_list.lock);
1013         if (!list_empty(&smc_lgr_list.list))
1014                 list_splice_init(&smc_lgr_list.list, &lgr_freeing_list);
1015         spin_unlock_bh(&smc_lgr_list.lock);
1016         list_for_each_entry_safe(lgr, lg, &lgr_freeing_list, list) {
1017                 list_del_init(&lgr->list);
1018                 if (!lgr->is_smcd) {
1019                         struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
1020
1021                         if (lnk->state == SMC_LNK_ACTIVE)
1022                                 smc_llc_send_delete_link(lnk, SMC_LLC_REQ,
1023                                                          false);
1024                         smc_llc_link_inactive(lnk);
1025                 }
1026                 cancel_delayed_work_sync(&lgr->free_work);
1027                 smc_lgr_free(lgr); /* free link group */
1028         }
1029 }