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         sock_put(&smc->sk); /* sock_hold in smc_lgr_register_conn() */
 122 }
 123
 124 /* Unregister connection from lgr
 125  */
 126 static void smc_lgr_unregister_conn(struct smc_connection *conn)
 127 {
 128         struct smc_link_group *lgr = conn->lgr;
 129
 130         if (!lgr)
 131                 return;
 132         write_lock_bh(&lgr->conns_lock);
 133         if (conn->alert_token_local) {
 134                 __smc_lgr_unregister_conn(conn);
 135         }
 136         write_unlock_bh(&lgr->conns_lock);
 137 }
 138
 139 /* Send delete link, either as client to request the initiation
 140  * of the DELETE LINK sequence from server; or as server to
 141  * initiate the delete processing. See smc_llc_rx_delete_link().
 142  */
 143 static int smc_link_send_delete(struct smc_link *lnk)
 144 {
 145         if (lnk->state == SMC_LNK_ACTIVE &&
 146             !smc_llc_send_delete_link(lnk, SMC_LLC_REQ, true)) {
 147                 smc_llc_link_deleting(lnk);
 148                 return 0;
 149         }
 150         return -ENOTCONN;
 151 }
 152
 153 static void smc_lgr_free(struct smc_link_group *lgr);
 154
 155 static void smc_lgr_free_work(struct work_struct *work)
 156 {
 157         struct smc_link_group *lgr = container_of(to_delayed_work(work),
 158                                                   struct smc_link_group,
 159                                                   free_work);
 160         bool conns;
 161
 162         spin_lock_bh(&smc_lgr_list.lock);
 163         read_lock_bh(&lgr->conns_lock);
 164         conns = RB_EMPTY_ROOT(&lgr->conns_all);
 165         read_unlock_bh(&lgr->conns_lock);
 166         if (!conns) { /* number of lgr connections is no longer zero */
 167                 spin_unlock_bh(&smc_lgr_list.lock);
 168                 return;
 169         }
 170         if (!list_empty(&lgr->list))
 171                 list_del_init(&lgr->list); /* remove from smc_lgr_list */
 172         spin_unlock_bh(&smc_lgr_list.lock);
 173
 174         if (!lgr->is_smcd && !lgr->terminating) {
 175                 struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
 176
 177                 /* try to send del link msg, on error free lgr immediately */
 178                 if (lnk->state == SMC_LNK_ACTIVE &&
 179                     !smc_link_send_delete(lnk)) {
 180                         /* reschedule in case we never receive a response */
 181                         smc_lgr_schedule_free_work(lgr);
 182                         return;
 183                 }
 184         }
 185
 186         if (!delayed_work_pending(&lgr->free_work)) {
 187                 struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
 188
 189                 if (!lgr->is_smcd && lnk->state != SMC_LNK_INACTIVE)
 190                         smc_llc_link_inactive(lnk);
 191                 if (lgr->is_smcd)
 192                         smc_ism_signal_shutdown(lgr);
 193                 smc_lgr_free(lgr);
 194         }
 195 }
 196
 197 /* create a new SMC link group */
 198 static int smc_lgr_create(struct smc_sock *smc, struct smc_init_info *ini)
 199 {
 200         struct smc_link_group *lgr;
 201         struct smc_link *lnk;
 202         u8 rndvec[3];
 203         int rc = 0;
 204         int i;
 205
 206         if (ini->is_smcd && ini->vlan_id) {
 207                 if (smc_ism_get_vlan(ini->ism_dev, ini->vlan_id)) {
 208                         rc = SMC_CLC_DECL_ISMVLANERR;
 209                         goto out;
 210                 }
 211         }
 212
 213         lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
 214         if (!lgr) {
 215                 rc = SMC_CLC_DECL_MEM;
 216                 goto out;
 217         }
 218         lgr->is_smcd = ini->is_smcd;
 219         lgr->sync_err = 0;
 220         lgr->vlan_id = ini->vlan_id;
 221         rwlock_init(&lgr->sndbufs_lock);
 222         rwlock_init(&lgr->rmbs_lock);
 223         rwlock_init(&lgr->conns_lock);
 224         for (i = 0; i < SMC_RMBE_SIZES; i++) {
 225                 INIT_LIST_HEAD(&lgr->sndbufs[i]);
 226                 INIT_LIST_HEAD(&lgr->rmbs[i]);
 227         }
 228         smc_lgr_list.num += SMC_LGR_NUM_INCR;
 229         memcpy(&lgr->id, (u8 *)&smc_lgr_list.num, SMC_LGR_ID_SIZE);
 230         INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
 231         lgr->conns_all = RB_ROOT;
 232         if (ini->is_smcd) {
 233                 /* SMC-D specific settings */
 234                 lgr->peer_gid = ini->ism_gid;
 235                 lgr->smcd = ini->ism_dev;
 236         } else {
 237                 /* SMC-R specific settings */
 238                 lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
 239                 memcpy(lgr->peer_systemid, ini->ib_lcl->id_for_peer,
 240                        SMC_SYSTEMID_LEN);
 241
 242                 lnk = &lgr->lnk[SMC_SINGLE_LINK];
 243                 /* initialize link */
 244                 lnk->state = SMC_LNK_ACTIVATING;
 245                 lnk->link_id = SMC_SINGLE_LINK;
 246                 lnk->smcibdev = ini->ib_dev;
 247                 lnk->ibport = ini->ib_port;
 248                 lnk->path_mtu =
 249                         ini->ib_dev->pattr[ini->ib_port - 1].active_mtu;
 250                 if (!ini->ib_dev->initialized)
 251                         smc_ib_setup_per_ibdev(ini->ib_dev);
 252                 get_random_bytes(rndvec, sizeof(rndvec));
 253                 lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) +
 254                         (rndvec[2] << 16);
 255                 rc = smc_ib_determine_gid(lnk->smcibdev, lnk->ibport,
 256                                           ini->vlan_id, lnk->gid,
 257                                           &lnk->sgid_index);
 258                 if (rc)
 259                         goto free_lgr;
 260                 rc = smc_llc_link_init(lnk);
 261                 if (rc)
 262                         goto free_lgr;
 263                 rc = smc_wr_alloc_link_mem(lnk);
 264                 if (rc)
 265                         goto clear_llc_lnk;
 266                 rc = smc_ib_create_protection_domain(lnk);
 267                 if (rc)
 268                         goto free_link_mem;
 269                 rc = smc_ib_create_queue_pair(lnk);
 270                 if (rc)
 271                         goto dealloc_pd;
 272                 rc = smc_wr_create_link(lnk);
 273                 if (rc)
 274                         goto destroy_qp;
 275         }
 276         smc->conn.lgr = lgr;
 277         spin_lock_bh(&smc_lgr_list.lock);
 278         list_add(&lgr->list, &smc_lgr_list.list);
 279         spin_unlock_bh(&smc_lgr_list.lock);
 280         return 0;
 281
 282 destroy_qp:
 283         smc_ib_destroy_queue_pair(lnk);
 284 dealloc_pd:
 285         smc_ib_dealloc_protection_domain(lnk);
 286 free_link_mem:
 287         smc_wr_free_link_mem(lnk);
 288 clear_llc_lnk:
 289         smc_llc_link_clear(lnk);
 290 free_lgr:
 291         kfree(lgr);
 292 out:
 293         if (rc < 0) {
 294                 if (rc == -ENOMEM)
 295                         rc = SMC_CLC_DECL_MEM;
 296                 else
 297                         rc = SMC_CLC_DECL_INTERR;
 298         }
 299         return rc;
 300 }
 301
 302 static void smc_buf_unuse(struct smc_connection *conn,
 303                           struct smc_link_group *lgr)
 304 {
 305         if (conn->sndbuf_desc)
 306                 conn->sndbuf_desc->used = 0;
 307         if (conn->rmb_desc) {
 308                 if (!conn->rmb_desc->regerr) {
 309                         if (!lgr->is_smcd) {
 310                                 /* unregister rmb with peer */
 311                                 smc_llc_do_delete_rkey(
 312                                                 &lgr->lnk[SMC_SINGLE_LINK],
 313                                                 conn->rmb_desc);
 314                         }
 315                         conn->rmb_desc->used = 0;
 316                 } else {
 317                         /* buf registration failed, reuse not possible */
 318                         write_lock_bh(&lgr->rmbs_lock);
 319                         list_del(&conn->rmb_desc->list);
 320                         write_unlock_bh(&lgr->rmbs_lock);
 321
 322                         smc_buf_free(lgr, true, conn->rmb_desc);
 323                 }
 324         }
 325 }
 326
 327 /* remove a finished connection from its link group */
 328 void smc_conn_free(struct smc_connection *conn)
 329 {
 330         struct smc_link_group *lgr = conn->lgr;
 331
 332         if (!lgr)
 333                 return;
 334         if (lgr->is_smcd) {
 335                 smc_ism_unset_conn(conn);
 336                 tasklet_kill(&conn->rx_tsklet);
 337         } else {
 338                 smc_cdc_tx_dismiss_slots(conn);
 339         }
 340         smc_lgr_unregister_conn(conn);
 341         smc_buf_unuse(conn, lgr);               /* allow buffer reuse */
 342         conn->lgr = NULL;
 343
 344         if (!lgr->conns_num)
 345                 smc_lgr_schedule_free_work(lgr);
 346 }
 347
 348 static void smc_link_clear(struct smc_link *lnk)
 349 {
 350         lnk->peer_qpn = 0;
 351         smc_llc_link_clear(lnk);
 352         smc_ib_modify_qp_reset(lnk);
 353         smc_wr_free_link(lnk);
 354         smc_ib_destroy_queue_pair(lnk);
 355         smc_ib_dealloc_protection_domain(lnk);
 356         smc_wr_free_link_mem(lnk);
 357 }
 358
 359 static void smcr_buf_free(struct smc_link_group *lgr, bool is_rmb,
 360                           struct smc_buf_desc *buf_desc)
 361 {
 362         struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
 363
 364         if (is_rmb) {
 365                 if (buf_desc->mr_rx[SMC_SINGLE_LINK])
 366                         smc_ib_put_memory_region(
 367                                         buf_desc->mr_rx[SMC_SINGLE_LINK]);
 368                 smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
 369                                     DMA_FROM_DEVICE);
 370         } else {
 371                 smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
 372                                     DMA_TO_DEVICE);
 373         }
 374         sg_free_table(&buf_desc->sgt[SMC_SINGLE_LINK]);
 375         if (buf_desc->pages)
 376                 __free_pages(buf_desc->pages, buf_desc->order);
 377         kfree(buf_desc);
 378 }
 379
 380 static void smcd_buf_free(struct smc_link_group *lgr, bool is_dmb,
 381                           struct smc_buf_desc *buf_desc)
 382 {
 383         if (is_dmb) {
 384                 /* restore original buf len */
 385                 buf_desc->len += sizeof(struct smcd_cdc_msg);
 386                 smc_ism_unregister_dmb(lgr->smcd, buf_desc);
 387         } else {
 388                 kfree(buf_desc->cpu_addr);
 389         }
 390         kfree(buf_desc);
 391 }
 392
 393 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
 394                          struct smc_buf_desc *buf_desc)
 395 {
 396         if (lgr->is_smcd)
 397                 smcd_buf_free(lgr, is_rmb, buf_desc);
 398         else
 399                 smcr_buf_free(lgr, is_rmb, buf_desc);
 400 }
 401
 402 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
 403 {
 404         struct smc_buf_desc *buf_desc, *bf_desc;
 405         struct list_head *buf_list;
 406         int i;
 407
 408         for (i = 0; i < SMC_RMBE_SIZES; i++) {
 409                 if (is_rmb)
 410                         buf_list = &lgr->rmbs[i];
 411                 else
 412                         buf_list = &lgr->sndbufs[i];
 413                 list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
 414                                          list) {
 415                         list_del(&buf_desc->list);
 416                         smc_buf_free(lgr, is_rmb, buf_desc);
 417                 }
 418         }
 419 }
 420
 421 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
 422 {
 423         /* free send buffers */
 424         __smc_lgr_free_bufs(lgr, false);
 425         /* free rmbs */
 426         __smc_lgr_free_bufs(lgr, true);
 427 }
 428
 429 /* remove a link group */
 430 static void smc_lgr_free(struct smc_link_group *lgr)
 431 {
 432         smc_lgr_free_bufs(lgr);
 433         if (lgr->is_smcd)
 434                 smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
 435         else
 436                 smc_link_clear(&lgr->lnk[SMC_SINGLE_LINK]);
 437         kfree(lgr);
 438 }
 439
 440 void smc_lgr_forget(struct smc_link_group *lgr)
 441 {
 442         spin_lock_bh(&smc_lgr_list.lock);
 443         /* do not use this link group for new connections */
 444         if (!list_empty(&lgr->list))
 445                 list_del_init(&lgr->list);
 446         spin_unlock_bh(&smc_lgr_list.lock);
 447 }
 448
 449 /* terminate linkgroup abnormally */
 450 static void __smc_lgr_terminate(struct smc_link_group *lgr)
 451 {
 452         struct smc_connection *conn;
 453         struct smc_sock *smc;
 454         struct rb_node *node;
 455
 456         if (lgr->terminating)
 457                 return; /* lgr already terminating */
 458         lgr->terminating = 1;
 459         if (!list_empty(&lgr->list)) /* forget lgr */
 460                 list_del_init(&lgr->list);
 461         if (!lgr->is_smcd)
 462                 smc_llc_link_inactive(&lgr->lnk[SMC_SINGLE_LINK]);
 463
 464         write_lock_bh(&lgr->conns_lock);
 465         node = rb_first(&lgr->conns_all);
 466         while (node) {
 467                 conn = rb_entry(node, struct smc_connection, alert_node);
 468                 smc = container_of(conn, struct smc_sock, conn);
 469                 sock_hold(&smc->sk); /* sock_put in close work */
 470                 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
 471                 __smc_lgr_unregister_conn(conn);
 472                 conn->lgr = NULL;
 473                 write_unlock_bh(&lgr->conns_lock);
 474                 if (!schedule_work(&conn->close_work))
 475                         sock_put(&smc->sk);
 476                 write_lock_bh(&lgr->conns_lock);
 477                 node = rb_first(&lgr->conns_all);
 478         }
 479         write_unlock_bh(&lgr->conns_lock);
 480         if (!lgr->is_smcd)
 481                 wake_up(&lgr->lnk[SMC_SINGLE_LINK].wr_reg_wait);
 482         smc_lgr_schedule_free_work(lgr);
 483 }
 484
 485 void smc_lgr_terminate(struct smc_link_group *lgr)
 486 {
 487         spin_lock_bh(&smc_lgr_list.lock);
 488         __smc_lgr_terminate(lgr);
 489         spin_unlock_bh(&smc_lgr_list.lock);
 490 }
 491
 492 /* Called when IB port is terminated */
 493 void smc_port_terminate(struct smc_ib_device *smcibdev, u8 ibport)
 494 {
 495         struct smc_link_group *lgr, *l;
 496
 497         spin_lock_bh(&smc_lgr_list.lock);
 498         list_for_each_entry_safe(lgr, l, &smc_lgr_list.list, list) {
 499                 if (!lgr->is_smcd &&
 500                     lgr->lnk[SMC_SINGLE_LINK].smcibdev == smcibdev &&
 501                     lgr->lnk[SMC_SINGLE_LINK].ibport == ibport)
 502                         __smc_lgr_terminate(lgr);
 503         }
 504         spin_unlock_bh(&smc_lgr_list.lock);
 505 }
 506
 507 /* Called when SMC-D device is terminated or peer is lost */
 508 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid, unsigned short vlan)
 509 {
 510         struct smc_link_group *lgr, *l;
 511         LIST_HEAD(lgr_free_list);
 512
 513         /* run common cleanup function and build free list */
 514         spin_lock_bh(&smc_lgr_list.lock);
 515         list_for_each_entry_safe(lgr, l, &smc_lgr_list.list, list) {
 516                 if (lgr->is_smcd && lgr->smcd == dev &&
 517                     (!peer_gid || lgr->peer_gid == peer_gid) &&
 518                     (vlan == VLAN_VID_MASK || lgr->vlan_id == vlan)) {
 519                         __smc_lgr_terminate(lgr);
 520                         list_move(&lgr->list, &lgr_free_list);
 521                 }
 522         }
 523         spin_unlock_bh(&smc_lgr_list.lock);
 524
 525         /* cancel the regular free workers and actually free lgrs */
 526         list_for_each_entry_safe(lgr, l, &lgr_free_list, list) {
 527                 list_del_init(&lgr->list);
 528                 cancel_delayed_work_sync(&lgr->free_work);
 529                 if (!peer_gid && vlan == VLAN_VID_MASK) /* dev terminated? */
 530                         smc_ism_signal_shutdown(lgr);
 531                 smc_lgr_free(lgr);
 532         }
 533 }
 534
 535 /* Determine vlan of internal TCP socket.
 536  * @vlan_id: address to store the determined vlan id into
 537  */
 538 int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini)
 539 {
 540         struct dst_entry *dst = sk_dst_get(clcsock->sk);
 541         struct net_device *ndev;
 542         int i, nest_lvl, rc = 0;
 543
 544         ini->vlan_id = 0;
 545         if (!dst) {
 546                 rc = -ENOTCONN;
 547                 goto out;
 548         }
 549         if (!dst->dev) {
 550                 rc = -ENODEV;
 551                 goto out_rel;
 552         }
 553
 554         ndev = dst->dev;
 555         if (is_vlan_dev(ndev)) {
 556                 ini->vlan_id = vlan_dev_vlan_id(ndev);
 557                 goto out_rel;
 558         }
 559
 560         rtnl_lock();
 561         nest_lvl = dev_get_nest_level(ndev);
 562         for (i = 0; i < nest_lvl; i++) {
 563                 struct list_head *lower = &ndev->adj_list.lower;
 564
 565                 if (list_empty(lower))
 566                         break;
 567                 lower = lower->next;
 568                 ndev = (struct net_device *)netdev_lower_get_next(ndev, &lower);
 569                 if (is_vlan_dev(ndev)) {
 570                         ini->vlan_id = vlan_dev_vlan_id(ndev);
 571                         break;
 572                 }
 573         }
 574         rtnl_unlock();
 575
 576 out_rel:
 577         dst_release(dst);
 578 out:
 579         return rc;
 580 }
 581
 582 static bool smcr_lgr_match(struct smc_link_group *lgr,
 583                            struct smc_clc_msg_local *lcl,
 584                            enum smc_lgr_role role, u32 clcqpn)
 585 {
 586         return !memcmp(lgr->peer_systemid, lcl->id_for_peer,
 587                        SMC_SYSTEMID_LEN) &&
 588                 !memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_gid, &lcl->gid,
 589                         SMC_GID_SIZE) &&
 590                 !memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_mac, lcl->mac,
 591                         sizeof(lcl->mac)) &&
 592                 lgr->role == role &&
 593                 (lgr->role == SMC_SERV ||
 594                  lgr->lnk[SMC_SINGLE_LINK].peer_qpn == clcqpn);
 595 }
 596
 597 static bool smcd_lgr_match(struct smc_link_group *lgr,
 598                            struct smcd_dev *smcismdev, u64 peer_gid)
 599 {
 600         return lgr->peer_gid == peer_gid && lgr->smcd == smcismdev;
 601 }
 602
 603 /* create a new SMC connection (and a new link group if necessary) */
 604 int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini)
 605 {
 606         struct smc_connection *conn = &smc->conn;
 607         struct smc_link_group *lgr;
 608         enum smc_lgr_role role;
 609         int rc = 0;
 610
 611         ini->cln_first_contact = SMC_FIRST_CONTACT;
 612         role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
 613         if (role == SMC_CLNT && ini->srv_first_contact)
 614                 /* create new link group as well */
 615                 goto create;
 616
 617         /* determine if an existing link group can be reused */
 618         spin_lock_bh(&smc_lgr_list.lock);
 619         list_for_each_entry(lgr, &smc_lgr_list.list, list) {
 620                 write_lock_bh(&lgr->conns_lock);
 621                 if ((ini->is_smcd ?
 622                      smcd_lgr_match(lgr, ini->ism_dev, ini->ism_gid) :
 623                      smcr_lgr_match(lgr, ini->ib_lcl, role, ini->ib_clcqpn)) &&
 624                     !lgr->sync_err &&
 625                     lgr->vlan_id == ini->vlan_id &&
 626                     (role == SMC_CLNT ||
 627                      lgr->conns_num < SMC_RMBS_PER_LGR_MAX)) {
 628                         /* link group found */
 629                         ini->cln_first_contact = SMC_REUSE_CONTACT;
 630                         conn->lgr = lgr;
 631                         smc_lgr_register_conn(conn); /* add smc conn to lgr */
 632                         if (delayed_work_pending(&lgr->free_work))
 633                                 cancel_delayed_work(&lgr->free_work);
 634                         write_unlock_bh(&lgr->conns_lock);
 635                         break;
 636                 }
 637                 write_unlock_bh(&lgr->conns_lock);
 638         }
 639         spin_unlock_bh(&smc_lgr_list.lock);
 640
 641         if (role == SMC_CLNT && !ini->srv_first_contact &&
 642             ini->cln_first_contact == SMC_FIRST_CONTACT) {
 643                 /* Server reuses a link group, but Client wants to start
 644                  * a new one
 645                  * send out_of_sync decline, reason synchr. error
 646                  */
 647                 return SMC_CLC_DECL_SYNCERR;
 648         }
 649
 650 create:
 651         if (ini->cln_first_contact == SMC_FIRST_CONTACT) {
 652                 rc = smc_lgr_create(smc, ini);
 653                 if (rc)
 654                         goto out;
 655                 smc_lgr_register_conn(conn); /* add smc conn to lgr */
 656         }
 657         conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
 658         conn->local_tx_ctrl.len = SMC_WR_TX_SIZE;
 659         conn->urg_state = SMC_URG_READ;
 660         if (ini->is_smcd) {
 661                 conn->rx_off = sizeof(struct smcd_cdc_msg);
 662                 smcd_cdc_rx_init(conn); /* init tasklet for this conn */
 663         }
 664 #ifndef KERNEL_HAS_ATOMIC64
 665         spin_lock_init(&conn->acurs_lock);
 666 #endif
 667
 668 out:
 669         return rc;
 670 }
 671
 672 /* convert the RMB size into the compressed notation - minimum 16K.
 673  * In contrast to plain ilog2, this rounds towards the next power of 2,
 674  * so the socket application gets at least its desired sndbuf / rcvbuf size.
 675  */
 676 static u8 smc_compress_bufsize(int size)
 677 {
 678         u8 compressed;
 679
 680         if (size <= SMC_BUF_MIN_SIZE)
 681                 return 0;
 682
 683         size = (size - 1) >> 14;
 684         compressed = ilog2(size) + 1;
 685         if (compressed >= SMC_RMBE_SIZES)
 686                 compressed = SMC_RMBE_SIZES - 1;
 687         return compressed;
 688 }
 689
 690 /* convert the RMB size from compressed notation into integer */
 691 int smc_uncompress_bufsize(u8 compressed)
 692 {
 693         u32 size;
 694
 695         size = 0x00000001 << (((int)compressed) + 14);
 696         return (int)size;
 697 }
 698
 699 /* try to reuse a sndbuf or rmb description slot for a certain
 700  * buffer size; if not available, return NULL
 701  */
 702 static struct smc_buf_desc *smc_buf_get_slot(int compressed_bufsize,
 703                                              rwlock_t *lock,
 704                                              struct list_head *buf_list)
 705 {
 706         struct smc_buf_desc *buf_slot;
 707
 708         read_lock_bh(lock);
 709         list_for_each_entry(buf_slot, buf_list, list) {
 710                 if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
 711                         read_unlock_bh(lock);
 712                         return buf_slot;
 713                 }
 714         }
 715         read_unlock_bh(lock);
 716         return NULL;
 717 }
 718
 719 /* one of the conditions for announcing a receiver's current window size is
 720  * that it "results in a minimum increase in the window size of 10% of the
 721  * receive buffer space" [RFC7609]
 722  */
 723 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
 724 {
 725         return min_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
 726 }
 727
 728 static struct smc_buf_desc *smcr_new_buf_create(struct smc_link_group *lgr,
 729                                                 bool is_rmb, int bufsize)
 730 {
 731         struct smc_buf_desc *buf_desc;
 732         struct smc_link *lnk;
 733         int rc;
 734
 735         /* try to alloc a new buffer */
 736         buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
 737         if (!buf_desc)
 738                 return ERR_PTR(-ENOMEM);
 739
 740         buf_desc->order = get_order(bufsize);
 741         buf_desc->pages = alloc_pages(GFP_KERNEL | __GFP_NOWARN |
 742                                       __GFP_NOMEMALLOC | __GFP_COMP |
 743                                       __GFP_NORETRY | __GFP_ZERO,
 744                                       buf_desc->order);
 745         if (!buf_desc->pages) {
 746                 kfree(buf_desc);
 747                 return ERR_PTR(-EAGAIN);
 748         }
 749         buf_desc->cpu_addr = (void *)page_address(buf_desc->pages);
 750
 751         /* build the sg table from the pages */
 752         lnk = &lgr->lnk[SMC_SINGLE_LINK];
 753         rc = sg_alloc_table(&buf_desc->sgt[SMC_SINGLE_LINK], 1,
 754                             GFP_KERNEL);
 755         if (rc) {
 756                 smc_buf_free(lgr, is_rmb, buf_desc);
 757                 return ERR_PTR(rc);
 758         }
 759         sg_set_buf(buf_desc->sgt[SMC_SINGLE_LINK].sgl,
 760                    buf_desc->cpu_addr, bufsize);
 761
 762         /* map sg table to DMA address */
 763         rc = smc_ib_buf_map_sg(lnk->smcibdev, buf_desc,
 764                                is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
 765         /* SMC protocol depends on mapping to one DMA address only */
 766         if (rc != 1)  {
 767                 smc_buf_free(lgr, is_rmb, buf_desc);
 768                 return ERR_PTR(-EAGAIN);
 769         }
 770
 771         /* create a new memory region for the RMB */
 772         if (is_rmb) {
 773                 rc = smc_ib_get_memory_region(lnk->roce_pd,
 774                                               IB_ACCESS_REMOTE_WRITE |
 775                                               IB_ACCESS_LOCAL_WRITE,
 776                                               buf_desc);
 777                 if (rc) {
 778                         smc_buf_free(lgr, is_rmb, buf_desc);
 779                         return ERR_PTR(rc);
 780                 }
 781         }
 782
 783         buf_desc->len = bufsize;
 784         return buf_desc;
 785 }
 786
 787 #define SMCD_DMBE_SIZES         7 /* 0 -> 16KB, 1 -> 32KB, .. 6 -> 1MB */
 788
 789 static struct smc_buf_desc *smcd_new_buf_create(struct smc_link_group *lgr,
 790                                                 bool is_dmb, int bufsize)
 791 {
 792         struct smc_buf_desc *buf_desc;
 793         int rc;
 794
 795         if (smc_compress_bufsize(bufsize) > SMCD_DMBE_SIZES)
 796                 return ERR_PTR(-EAGAIN);
 797
 798         /* try to alloc a new DMB */
 799         buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
 800         if (!buf_desc)
 801                 return ERR_PTR(-ENOMEM);
 802         if (is_dmb) {
 803                 rc = smc_ism_register_dmb(lgr, bufsize, buf_desc);
 804                 if (rc) {
 805                         kfree(buf_desc);
 806                         return ERR_PTR(-EAGAIN);
 807                 }
 808                 buf_desc->pages = virt_to_page(buf_desc->cpu_addr);
 809                 /* CDC header stored in buf. So, pretend it was smaller */
 810                 buf_desc->len = bufsize - sizeof(struct smcd_cdc_msg);
 811         } else {
 812                 buf_desc->cpu_addr = kzalloc(bufsize, GFP_KERNEL |
 813                                              __GFP_NOWARN | __GFP_NORETRY |
 814                                              __GFP_NOMEMALLOC);
 815                 if (!buf_desc->cpu_addr) {
 816                         kfree(buf_desc);
 817                         return ERR_PTR(-EAGAIN);
 818                 }
 819                 buf_desc->len = bufsize;
 820         }
 821         return buf_desc;
 822 }
 823
 824 static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
 825 {
 826         struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
 827         struct smc_connection *conn = &smc->conn;
 828         struct smc_link_group *lgr = conn->lgr;
 829         struct list_head *buf_list;
 830         int bufsize, bufsize_short;
 831         int sk_buf_size;
 832         rwlock_t *lock;
 833
 834         if (is_rmb)
 835                 /* use socket recv buffer size (w/o overhead) as start value */
 836                 sk_buf_size = smc->sk.sk_rcvbuf / 2;
 837         else
 838                 /* use socket send buffer size (w/o overhead) as start value */
 839                 sk_buf_size = smc->sk.sk_sndbuf / 2;
 840
 841         for (bufsize_short = smc_compress_bufsize(sk_buf_size);
 842              bufsize_short >= 0; bufsize_short--) {
 843
 844                 if (is_rmb) {
 845                         lock = &lgr->rmbs_lock;
 846                         buf_list = &lgr->rmbs[bufsize_short];
 847                 } else {
 848                         lock = &lgr->sndbufs_lock;
 849                         buf_list = &lgr->sndbufs[bufsize_short];
 850                 }
 851                 bufsize = smc_uncompress_bufsize(bufsize_short);
 852                 if ((1 << get_order(bufsize)) > SG_MAX_SINGLE_ALLOC)
 853                         continue;
 854
 855                 /* check for reusable slot in the link group */
 856                 buf_desc = smc_buf_get_slot(bufsize_short, lock, buf_list);
 857                 if (buf_desc) {
 858                         memset(buf_desc->cpu_addr, 0, bufsize);
 859                         break; /* found reusable slot */
 860                 }
 861
 862                 if (is_smcd)
 863                         buf_desc = smcd_new_buf_create(lgr, is_rmb, bufsize);
 864                 else
 865                         buf_desc = smcr_new_buf_create(lgr, is_rmb, bufsize);
 866
 867                 if (PTR_ERR(buf_desc) == -ENOMEM)
 868                         break;
 869                 if (IS_ERR(buf_desc))
 870                         continue;
 871
 872                 buf_desc->used = 1;
 873                 write_lock_bh(lock);
 874                 list_add(&buf_desc->list, buf_list);
 875                 write_unlock_bh(lock);
 876                 break; /* found */
 877         }
 878
 879         if (IS_ERR(buf_desc))
 880                 return -ENOMEM;
 881
 882         if (is_rmb) {
 883                 conn->rmb_desc = buf_desc;
 884                 conn->rmbe_size_short = bufsize_short;
 885                 smc->sk.sk_rcvbuf = bufsize * 2;
 886                 atomic_set(&conn->bytes_to_rcv, 0);
 887                 conn->rmbe_update_limit =
 888                         smc_rmb_wnd_update_limit(buf_desc->len);
 889                 if (is_smcd)
 890                         smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */
 891         } else {
 892                 conn->sndbuf_desc = buf_desc;
 893                 smc->sk.sk_sndbuf = bufsize * 2;
 894                 atomic_set(&conn->sndbuf_space, bufsize);
 895         }
 896         return 0;
 897 }
 898
 899 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn)
 900 {
 901         struct smc_link_group *lgr = conn->lgr;
 902
 903         if (!conn->lgr || conn->lgr->is_smcd)
 904                 return;
 905         smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 906                                conn->sndbuf_desc, DMA_TO_DEVICE);
 907 }
 908
 909 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
 910 {
 911         struct smc_link_group *lgr = conn->lgr;
 912
 913         if (!conn->lgr || conn->lgr->is_smcd)
 914                 return;
 915         smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 916                                   conn->sndbuf_desc, DMA_TO_DEVICE);
 917 }
 918
 919 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
 920 {
 921         struct smc_link_group *lgr = conn->lgr;
 922
 923         if (!conn->lgr || conn->lgr->is_smcd)
 924                 return;
 925         smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 926                                conn->rmb_desc, DMA_FROM_DEVICE);
 927 }
 928
 929 void smc_rmb_sync_sg_for_device(struct smc_connection *conn)
 930 {
 931         struct smc_link_group *lgr = conn->lgr;
 932
 933         if (!conn->lgr || conn->lgr->is_smcd)
 934                 return;
 935         smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 936                                   conn->rmb_desc, DMA_FROM_DEVICE);
 937 }
 938
 939 /* create the send and receive buffer for an SMC socket;
 940  * receive buffers are called RMBs;
 941  * (even though the SMC protocol allows more than one RMB-element per RMB,
 942  * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
 943  * extra RMB for every connection in a link group
 944  */
 945 int smc_buf_create(struct smc_sock *smc, bool is_smcd)
 946 {
 947         int rc;
 948
 949         /* create send buffer */
 950         rc = __smc_buf_create(smc, is_smcd, false);
 951         if (rc)
 952                 return rc;
 953         /* create rmb */
 954         rc = __smc_buf_create(smc, is_smcd, true);
 955         if (rc)
 956                 smc_buf_free(smc->conn.lgr, false, smc->conn.sndbuf_desc);
 957         return rc;
 958 }
 959
 960 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
 961 {
 962         int i;
 963
 964         for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
 965                 if (!test_and_set_bit(i, lgr->rtokens_used_mask))
 966                         return i;
 967         }
 968         return -ENOSPC;
 969 }
 970
 971 /* add a new rtoken from peer */
 972 int smc_rtoken_add(struct smc_link_group *lgr, __be64 nw_vaddr, __be32 nw_rkey)
 973 {
 974         u64 dma_addr = be64_to_cpu(nw_vaddr);
 975         u32 rkey = ntohl(nw_rkey);
 976         int i;
 977
 978         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
 979                 if ((lgr->rtokens[i][SMC_SINGLE_LINK].rkey == rkey) &&
 980                     (lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr == dma_addr) &&
 981                     test_bit(i, lgr->rtokens_used_mask)) {
 982                         /* already in list */
 983                         return i;
 984                 }
 985         }
 986         i = smc_rmb_reserve_rtoken_idx(lgr);
 987         if (i < 0)
 988                 return i;
 989         lgr->rtokens[i][SMC_SINGLE_LINK].rkey = rkey;
 990         lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr = dma_addr;
 991         return i;
 992 }
 993
 994 /* delete an rtoken */
 995 int smc_rtoken_delete(struct smc_link_group *lgr, __be32 nw_rkey)
 996 {
 997         u32 rkey = ntohl(nw_rkey);
 998         int i;
 999
1000         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
1001                 if (lgr->rtokens[i][SMC_SINGLE_LINK].rkey == rkey &&
1002                     test_bit(i, lgr->rtokens_used_mask)) {
1003                         lgr->rtokens[i][SMC_SINGLE_LINK].rkey = 0;
1004                         lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr = 0;
1005
1006                         clear_bit(i, lgr->rtokens_used_mask);
1007                         return 0;
1008                 }
1009         }
1010         return -ENOENT;
1011 }
1012
1013 /* save rkey and dma_addr received from peer during clc handshake */
1014 int smc_rmb_rtoken_handling(struct smc_connection *conn,
1015                             struct smc_clc_msg_accept_confirm *clc)
1016 {
1017         conn->rtoken_idx = smc_rtoken_add(conn->lgr, clc->rmb_dma_addr,
1018                                           clc->rmb_rkey);
1019         if (conn->rtoken_idx < 0)
1020                 return conn->rtoken_idx;
1021         return 0;
1022 }
1023
1024 /* Called (from smc_exit) when module is removed */
1025 void smc_core_exit(void)
1026 {
1027         struct smc_link_group *lgr, *lg;
1028         LIST_HEAD(lgr_freeing_list);
1029
1030         spin_lock_bh(&smc_lgr_list.lock);
1031         if (!list_empty(&smc_lgr_list.list))
1032                 list_splice_init(&smc_lgr_list.list, &lgr_freeing_list);
1033         spin_unlock_bh(&smc_lgr_list.lock);
1034         list_for_each_entry_safe(lgr, lg, &lgr_freeing_list, list) {
1035                 list_del_init(&lgr->list);
1036                 if (!lgr->is_smcd) {
1037                         struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
1038
1039                         if (lnk->state == SMC_LNK_ACTIVE)
1040                                 smc_llc_send_delete_link(lnk, SMC_LLC_REQ,
1041                                                          false);
1042                         smc_llc_link_inactive(lnk);
1043                 }
1044                 cancel_delayed_work_sync(&lgr->free_work);
1045                 if (lgr->is_smcd)
1046                         smc_ism_signal_shutdown(lgr);
1047                 smc_lgr_free(lgr); /* free link group */
1048         }
1049 }