net/smc/smc_core.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *  Shared Memory Communications over RDMA (SMC-R) and RoCE
   4  *
   5  *  Basic Transport Functions exploiting Infiniband API
   6  *
   7  *  Copyright IBM Corp. 2016
   8  *
   9  *  Author(s):  Ursula Braun <[email protected]>
  10  */
  11
  12 #include <linux/socket.h>
  13 #include <linux/if_vlan.h>
  14 #include <linux/random.h>
  15 #include <linux/workqueue.h>
  16 #include <net/tcp.h>
  17 #include <net/sock.h>
  18 #include <rdma/ib_verbs.h>
  19
  20 #include "smc.h"
  21 #include "smc_clc.h"
  22 #include "smc_core.h"
  23 #include "smc_ib.h"
  24 #include "smc_wr.h"
  25 #include "smc_llc.h"
  26 #include "smc_cdc.h"
  27 #include "smc_close.h"
  28
  29 #define SMC_LGR_NUM_INCR                256
  30 #define SMC_LGR_FREE_DELAY_SERV         (600 * HZ)
  31 #define SMC_LGR_FREE_DELAY_CLNT         (SMC_LGR_FREE_DELAY_SERV + 10)
  32
  33 static u32 smc_lgr_num;                 /* unique link group number */
  34
  35 /* Register connection's alert token in our lookup structure.
  36  * To use rbtrees we have to implement our own insert core.
  37  * Requires @conns_lock
  38  * @smc         connection to register
  39  * Returns 0 on success, != otherwise.
  40  */
  41 static void smc_lgr_add_alert_token(struct smc_connection *conn)
  42 {
  43         struct rb_node **link, *parent = NULL;
  44         u32 token = conn->alert_token_local;
  45
  46         link = &conn->lgr->conns_all.rb_node;
  47         while (*link) {
  48                 struct smc_connection *cur = rb_entry(*link,
  49                                         struct smc_connection, alert_node);
  50
  51                 parent = *link;
  52                 if (cur->alert_token_local > token)
  53                         link = &parent->rb_left;
  54                 else
  55                         link = &parent->rb_right;
  56         }
  57         /* Put the new node there */
  58         rb_link_node(&conn->alert_node, parent, link);
  59         rb_insert_color(&conn->alert_node, &conn->lgr->conns_all);
  60 }
  61
  62 /* Register connection in link group by assigning an alert token
  63  * registered in a search tree.
  64  * Requires @conns_lock
  65  * Note that '0' is a reserved value and not assigned.
  66  */
  67 static void smc_lgr_register_conn(struct smc_connection *conn)
  68 {
  69         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
  70         static atomic_t nexttoken = ATOMIC_INIT(0);
  71
  72         /* find a new alert_token_local value not yet used by some connection
  73          * in this link group
  74          */
  75         sock_hold(&smc->sk); /* sock_put in smc_lgr_unregister_conn() */
  76         while (!conn->alert_token_local) {
  77                 conn->alert_token_local = atomic_inc_return(&nexttoken);
  78                 if (smc_lgr_find_conn(conn->alert_token_local, conn->lgr))
  79                         conn->alert_token_local = 0;
  80         }
  81         smc_lgr_add_alert_token(conn);
  82         conn->lgr->conns_num++;
  83 }
  84
  85 /* Unregister connection and reset the alert token of the given connection<
  86  */
  87 static void __smc_lgr_unregister_conn(struct smc_connection *conn)
  88 {
  89         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
  90         struct smc_link_group *lgr = conn->lgr;
  91
  92         rb_erase(&conn->alert_node, &lgr->conns_all);
  93         lgr->conns_num--;
  94         conn->alert_token_local = 0;
  95         conn->lgr = NULL;
  96         sock_put(&smc->sk); /* sock_hold in smc_lgr_register_conn() */
  97 }
  98
  99 /* Unregister connection and trigger lgr freeing if applicable
 100  */
 101 static void smc_lgr_unregister_conn(struct smc_connection *conn)
 102 {
 103         struct smc_link_group *lgr = conn->lgr;
 104         int reduced = 0;
 105
 106         write_lock_bh(&lgr->conns_lock);
 107         if (conn->alert_token_local) {
 108                 reduced = 1;
 109                 __smc_lgr_unregister_conn(conn);
 110         }
 111         write_unlock_bh(&lgr->conns_lock);
 112         if (!reduced || lgr->conns_num)
 113                 return;
 114         /* client link group creation always follows the server link group
 115          * creation. For client use a somewhat higher removal delay time,
 116          * otherwise there is a risk of out-of-sync link groups.
 117          */
 118         mod_delayed_work(system_wq, &lgr->free_work,
 119                          lgr->role == SMC_CLNT ? SMC_LGR_FREE_DELAY_CLNT :
 120                                                  SMC_LGR_FREE_DELAY_SERV);
 121 }
 122
 123 static void smc_lgr_free_work(struct work_struct *work)
 124 {
 125         struct smc_link_group *lgr = container_of(to_delayed_work(work),
 126                                                   struct smc_link_group,
 127                                                   free_work);
 128         bool conns;
 129
 130         spin_lock_bh(&smc_lgr_list.lock);
 131         if (list_empty(&lgr->list))
 132                 goto free;
 133         read_lock_bh(&lgr->conns_lock);
 134         conns = RB_EMPTY_ROOT(&lgr->conns_all);
 135         read_unlock_bh(&lgr->conns_lock);
 136         if (!conns) { /* number of lgr connections is no longer zero */
 137                 spin_unlock_bh(&smc_lgr_list.lock);
 138                 return;
 139         }
 140         list_del_init(&lgr->list); /* remove from smc_lgr_list */
 141 free:
 142         spin_unlock_bh(&smc_lgr_list.lock);
 143         smc_lgr_free(lgr);
 144 }
 145
 146 /* create a new SMC link group */
 147 static int smc_lgr_create(struct smc_sock *smc, __be32 peer_in_addr,
 148                           struct smc_ib_device *smcibdev, u8 ibport,
 149                           char *peer_systemid, unsigned short vlan_id)
 150 {
 151         struct smc_link_group *lgr;
 152         struct smc_link *lnk;
 153         u8 rndvec[3];
 154         int rc = 0;
 155         int i;
 156
 157         lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
 158         if (!lgr) {
 159                 rc = -ENOMEM;
 160                 goto out;
 161         }
 162         lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
 163         lgr->sync_err = false;
 164         lgr->daddr = peer_in_addr;
 165         memcpy(lgr->peer_systemid, peer_systemid, SMC_SYSTEMID_LEN);
 166         lgr->vlan_id = vlan_id;
 167         rwlock_init(&lgr->sndbufs_lock);
 168         rwlock_init(&lgr->rmbs_lock);
 169         for (i = 0; i < SMC_RMBE_SIZES; i++) {
 170                 INIT_LIST_HEAD(&lgr->sndbufs[i]);
 171                 INIT_LIST_HEAD(&lgr->rmbs[i]);
 172         }
 173         smc_lgr_num += SMC_LGR_NUM_INCR;
 174         memcpy(&lgr->id, (u8 *)&smc_lgr_num, SMC_LGR_ID_SIZE);
 175         INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
 176         lgr->conns_all = RB_ROOT;
 177
 178         lnk = &lgr->lnk[SMC_SINGLE_LINK];
 179         /* initialize link */
 180         lnk->link_id = SMC_SINGLE_LINK;
 181         lnk->smcibdev = smcibdev;
 182         lnk->ibport = ibport;
 183         lnk->path_mtu = smcibdev->pattr[ibport - 1].active_mtu;
 184         if (!smcibdev->initialized)
 185                 smc_ib_setup_per_ibdev(smcibdev);
 186         get_random_bytes(rndvec, sizeof(rndvec));
 187         lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) + (rndvec[2] << 16);
 188         rc = smc_wr_alloc_link_mem(lnk);
 189         if (rc)
 190                 goto free_lgr;
 191         rc = smc_ib_create_protection_domain(lnk);
 192         if (rc)
 193                 goto free_link_mem;
 194         rc = smc_ib_create_queue_pair(lnk);
 195         if (rc)
 196                 goto dealloc_pd;
 197         rc = smc_wr_create_link(lnk);
 198         if (rc)
 199                 goto destroy_qp;
 200         init_completion(&lnk->llc_confirm);
 201         init_completion(&lnk->llc_confirm_resp);
 202
 203         smc->conn.lgr = lgr;
 204         rwlock_init(&lgr->conns_lock);
 205         spin_lock_bh(&smc_lgr_list.lock);
 206         list_add(&lgr->list, &smc_lgr_list.list);
 207         spin_unlock_bh(&smc_lgr_list.lock);
 208         return 0;
 209
 210 destroy_qp:
 211         smc_ib_destroy_queue_pair(lnk);
 212 dealloc_pd:
 213         smc_ib_dealloc_protection_domain(lnk);
 214 free_link_mem:
 215         smc_wr_free_link_mem(lnk);
 216 free_lgr:
 217         kfree(lgr);
 218 out:
 219         return rc;
 220 }
 221
 222 static void smc_buf_unuse(struct smc_connection *conn)
 223 {
 224         if (conn->sndbuf_desc) {
 225                 conn->sndbuf_desc->used = 0;
 226                 conn->sndbuf_size = 0;
 227         }
 228         if (conn->rmb_desc) {
 229                 conn->rmb_desc->reused = true;
 230                 conn->rmb_desc->used = 0;
 231                 conn->rmbe_size = 0;
 232         }
 233 }
 234
 235 /* remove a finished connection from its link group */
 236 void smc_conn_free(struct smc_connection *conn)
 237 {
 238         if (!conn->lgr)
 239                 return;
 240         smc_cdc_tx_dismiss_slots(conn);
 241         smc_lgr_unregister_conn(conn);
 242         smc_buf_unuse(conn);
 243 }
 244
 245 static void smc_link_clear(struct smc_link *lnk)
 246 {
 247         lnk->peer_qpn = 0;
 248         smc_ib_modify_qp_reset(lnk);
 249         smc_wr_free_link(lnk);
 250         smc_ib_destroy_queue_pair(lnk);
 251         smc_ib_dealloc_protection_domain(lnk);
 252         smc_wr_free_link_mem(lnk);
 253 }
 254
 255 static void smc_buf_free(struct smc_buf_desc *buf_desc, struct smc_link *lnk,
 256                          bool is_rmb)
 257 {
 258         if (is_rmb) {
 259                 if (buf_desc->mr_rx[SMC_SINGLE_LINK])
 260                         smc_ib_put_memory_region(
 261                                         buf_desc->mr_rx[SMC_SINGLE_LINK]);
 262                 smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
 263                                     DMA_FROM_DEVICE);
 264         } else {
 265                 smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
 266                                     DMA_TO_DEVICE);
 267         }
 268         sg_free_table(&buf_desc->sgt[SMC_SINGLE_LINK]);
 269         if (buf_desc->cpu_addr)
 270                 free_pages((unsigned long)buf_desc->cpu_addr, buf_desc->order);
 271         kfree(buf_desc);
 272 }
 273
 274 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
 275 {
 276         struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
 277         struct smc_buf_desc *buf_desc, *bf_desc;
 278         struct list_head *buf_list;
 279         int i;
 280
 281         for (i = 0; i < SMC_RMBE_SIZES; i++) {
 282                 if (is_rmb)
 283                         buf_list = &lgr->rmbs[i];
 284                 else
 285                         buf_list = &lgr->sndbufs[i];
 286                 list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
 287                                          list) {
 288                         list_del(&buf_desc->list);
 289                         smc_buf_free(buf_desc, lnk, is_rmb);
 290                 }
 291         }
 292 }
 293
 294 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
 295 {
 296         /* free send buffers */
 297         __smc_lgr_free_bufs(lgr, false);
 298         /* free rmbs */
 299         __smc_lgr_free_bufs(lgr, true);
 300 }
 301
 302 /* remove a link group */
 303 void smc_lgr_free(struct smc_link_group *lgr)
 304 {
 305         smc_lgr_free_bufs(lgr);
 306         smc_link_clear(&lgr->lnk[SMC_SINGLE_LINK]);
 307         kfree(lgr);
 308 }
 309
 310 /* terminate linkgroup abnormally */
 311 void smc_lgr_terminate(struct smc_link_group *lgr)
 312 {
 313         struct smc_connection *conn;
 314         struct smc_sock *smc;
 315         struct rb_node *node;
 316
 317         spin_lock_bh(&smc_lgr_list.lock);
 318         if (list_empty(&lgr->list)) {
 319                 /* termination already triggered */
 320                 spin_unlock_bh(&smc_lgr_list.lock);
 321                 return;
 322         }
 323         /* do not use this link group for new connections */
 324         list_del_init(&lgr->list);
 325         spin_unlock_bh(&smc_lgr_list.lock);
 326
 327         write_lock_bh(&lgr->conns_lock);
 328         node = rb_first(&lgr->conns_all);
 329         while (node) {
 330                 conn = rb_entry(node, struct smc_connection, alert_node);
 331                 smc = container_of(conn, struct smc_sock, conn);
 332                 sock_hold(&smc->sk); /* sock_put in close work */
 333                 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
 334                 __smc_lgr_unregister_conn(conn);
 335                 write_unlock_bh(&lgr->conns_lock);
 336                 if (!schedule_work(&conn->close_work))
 337                         sock_put(&smc->sk);
 338                 write_lock_bh(&lgr->conns_lock);
 339                 node = rb_first(&lgr->conns_all);
 340         }
 341         write_unlock_bh(&lgr->conns_lock);
 342         wake_up(&lgr->lnk[SMC_SINGLE_LINK].wr_reg_wait);
 343 }
 344
 345 /* Determine vlan of internal TCP socket.
 346  * @vlan_id: address to store the determined vlan id into
 347  */
 348 static int smc_vlan_by_tcpsk(struct socket *clcsock, unsigned short *vlan_id)
 349 {
 350         struct dst_entry *dst = sk_dst_get(clcsock->sk);
 351         int rc = 0;
 352
 353         *vlan_id = 0;
 354         if (!dst) {
 355                 rc = -ENOTCONN;
 356                 goto out;
 357         }
 358         if (!dst->dev) {
 359                 rc = -ENODEV;
 360                 goto out_rel;
 361         }
 362
 363         if (is_vlan_dev(dst->dev))
 364                 *vlan_id = vlan_dev_vlan_id(dst->dev);
 365
 366 out_rel:
 367         dst_release(dst);
 368 out:
 369         return rc;
 370 }
 371
 372 /* determine the link gid matching the vlan id of the link group */
 373 static int smc_link_determine_gid(struct smc_link_group *lgr)
 374 {
 375         struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
 376         struct ib_gid_attr gattr;
 377         union ib_gid gid;
 378         int i;
 379
 380         if (!lgr->vlan_id) {
 381                 lnk->gid = lnk->smcibdev->gid[lnk->ibport - 1];
 382                 return 0;
 383         }
 384
 385         for (i = 0; i < lnk->smcibdev->pattr[lnk->ibport - 1].gid_tbl_len;
 386              i++) {
 387                 if (ib_query_gid(lnk->smcibdev->ibdev, lnk->ibport, i, &gid,
 388                                  &gattr))
 389                         continue;
 390                 if (gattr.ndev) {
 391                         if (is_vlan_dev(gattr.ndev) &&
 392                             vlan_dev_vlan_id(gattr.ndev) == lgr->vlan_id) {
 393                                 lnk->gid = gid;
 394                                 dev_put(gattr.ndev);
 395                                 return 0;
 396                         }
 397                         dev_put(gattr.ndev);
 398                 }
 399         }
 400         return -ENODEV;
 401 }
 402
 403 /* create a new SMC connection (and a new link group if necessary) */
 404 int smc_conn_create(struct smc_sock *smc, __be32 peer_in_addr,
 405                     struct smc_ib_device *smcibdev, u8 ibport,
 406                     struct smc_clc_msg_local *lcl, int srv_first_contact)
 407 {
 408         struct smc_connection *conn = &smc->conn;
 409         struct smc_link_group *lgr;
 410         unsigned short vlan_id;
 411         enum smc_lgr_role role;
 412         int local_contact = SMC_FIRST_CONTACT;
 413         int rc = 0;
 414
 415         role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
 416         rc = smc_vlan_by_tcpsk(smc->clcsock, &vlan_id);
 417         if (rc)
 418                 return rc;
 419
 420         if ((role == SMC_CLNT) && srv_first_contact)
 421                 /* create new link group as well */
 422                 goto create;
 423
 424         /* determine if an existing link group can be reused */
 425         spin_lock_bh(&smc_lgr_list.lock);
 426         list_for_each_entry(lgr, &smc_lgr_list.list, list) {
 427                 write_lock_bh(&lgr->conns_lock);
 428                 if (!memcmp(lgr->peer_systemid, lcl->id_for_peer,
 429                             SMC_SYSTEMID_LEN) &&
 430                     !memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_gid, &lcl->gid,
 431                             SMC_GID_SIZE) &&
 432                     !memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_mac, lcl->mac,
 433                             sizeof(lcl->mac)) &&
 434                     !lgr->sync_err &&
 435                     (lgr->role == role) &&
 436                     (lgr->vlan_id == vlan_id) &&
 437                     ((role == SMC_CLNT) ||
 438                      (lgr->conns_num < SMC_RMBS_PER_LGR_MAX))) {
 439                         /* link group found */
 440                         local_contact = SMC_REUSE_CONTACT;
 441                         conn->lgr = lgr;
 442                         smc_lgr_register_conn(conn); /* add smc conn to lgr */
 443                         write_unlock_bh(&lgr->conns_lock);
 444                         break;
 445                 }
 446                 write_unlock_bh(&lgr->conns_lock);
 447         }
 448         spin_unlock_bh(&smc_lgr_list.lock);
 449
 450         if (role == SMC_CLNT && !srv_first_contact &&
 451             (local_contact == SMC_FIRST_CONTACT)) {
 452                 /* Server reuses a link group, but Client wants to start
 453                  * a new one
 454                  * send out_of_sync decline, reason synchr. error
 455                  */
 456                 return -ENOLINK;
 457         }
 458
 459 create:
 460         if (local_contact == SMC_FIRST_CONTACT) {
 461                 rc = smc_lgr_create(smc, peer_in_addr, smcibdev, ibport,
 462                                     lcl->id_for_peer, vlan_id);
 463                 if (rc)
 464                         goto out;
 465                 smc_lgr_register_conn(conn); /* add smc conn to lgr */
 466                 rc = smc_link_determine_gid(conn->lgr);
 467         }
 468         conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
 469         conn->local_tx_ctrl.len = SMC_WR_TX_SIZE;
 470 #ifndef KERNEL_HAS_ATOMIC64
 471         spin_lock_init(&conn->acurs_lock);
 472 #endif
 473
 474 out:
 475         return rc ? rc : local_contact;
 476 }
 477
 478 /* try to reuse a sndbuf or rmb description slot for a certain
 479  * buffer size; if not available, return NULL
 480  */
 481 static inline
 482 struct smc_buf_desc *smc_buf_get_slot(struct smc_link_group *lgr,
 483                                       int compressed_bufsize,
 484                                       rwlock_t *lock,
 485                                       struct list_head *buf_list)
 486 {
 487         struct smc_buf_desc *buf_slot;
 488
 489         read_lock_bh(lock);
 490         list_for_each_entry(buf_slot, buf_list, list) {
 491                 if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
 492                         read_unlock_bh(lock);
 493                         return buf_slot;
 494                 }
 495         }
 496         read_unlock_bh(lock);
 497         return NULL;
 498 }
 499
 500 /* one of the conditions for announcing a receiver's current window size is
 501  * that it "results in a minimum increase in the window size of 10% of the
 502  * receive buffer space" [RFC7609]
 503  */
 504 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
 505 {
 506         return min_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
 507 }
 508
 509 static struct smc_buf_desc *smc_new_buf_create(struct smc_link_group *lgr,
 510                                                bool is_rmb, int bufsize)
 511 {
 512         struct smc_buf_desc *buf_desc;
 513         struct smc_link *lnk;
 514         int rc;
 515
 516         /* try to alloc a new buffer */
 517         buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
 518         if (!buf_desc)
 519                 return ERR_PTR(-ENOMEM);
 520
 521         buf_desc->cpu_addr =
 522                 (void *)__get_free_pages(GFP_KERNEL | __GFP_NOWARN |
 523                                          __GFP_NOMEMALLOC |
 524                                          __GFP_NORETRY | __GFP_ZERO,
 525                                          get_order(bufsize));
 526         if (!buf_desc->cpu_addr) {
 527                 kfree(buf_desc);
 528                 return ERR_PTR(-EAGAIN);
 529         }
 530         buf_desc->order = get_order(bufsize);
 531
 532         /* build the sg table from the pages */
 533         lnk = &lgr->lnk[SMC_SINGLE_LINK];
 534         rc = sg_alloc_table(&buf_desc->sgt[SMC_SINGLE_LINK], 1,
 535                             GFP_KERNEL);
 536         if (rc) {
 537                 smc_buf_free(buf_desc, lnk, is_rmb);
 538                 return ERR_PTR(rc);
 539         }
 540         sg_set_buf(buf_desc->sgt[SMC_SINGLE_LINK].sgl,
 541                    buf_desc->cpu_addr, bufsize);
 542
 543         /* map sg table to DMA address */
 544         rc = smc_ib_buf_map_sg(lnk->smcibdev, buf_desc,
 545                                is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
 546         /* SMC protocol depends on mapping to one DMA address only */
 547         if (rc != 1)  {
 548                 smc_buf_free(buf_desc, lnk, is_rmb);
 549                 return ERR_PTR(-EAGAIN);
 550         }
 551
 552         /* create a new memory region for the RMB */
 553         if (is_rmb) {
 554                 rc = smc_ib_get_memory_region(lnk->roce_pd,
 555                                               IB_ACCESS_REMOTE_WRITE |
 556                                               IB_ACCESS_LOCAL_WRITE,
 557                                               buf_desc);
 558                 if (rc) {
 559                         smc_buf_free(buf_desc, lnk, is_rmb);
 560                         return ERR_PTR(rc);
 561                 }
 562         }
 563
 564         return buf_desc;
 565 }
 566
 567 static int __smc_buf_create(struct smc_sock *smc, bool is_rmb)
 568 {
 569         struct smc_connection *conn = &smc->conn;
 570         struct smc_link_group *lgr = conn->lgr;
 571         struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
 572         struct list_head *buf_list;
 573         int bufsize, bufsize_short;
 574         int sk_buf_size;
 575         rwlock_t *lock;
 576
 577         if (is_rmb)
 578                 /* use socket recv buffer size (w/o overhead) as start value */
 579                 sk_buf_size = smc->sk.sk_rcvbuf / 2;
 580         else
 581                 /* use socket send buffer size (w/o overhead) as start value */
 582                 sk_buf_size = smc->sk.sk_sndbuf / 2;
 583
 584         for (bufsize_short = smc_compress_bufsize(sk_buf_size);
 585              bufsize_short >= 0; bufsize_short--) {
 586
 587                 if (is_rmb) {
 588                         lock = &lgr->rmbs_lock;
 589                         buf_list = &lgr->rmbs[bufsize_short];
 590                 } else {
 591                         lock = &lgr->sndbufs_lock;
 592                         buf_list = &lgr->sndbufs[bufsize_short];
 593                 }
 594                 bufsize = smc_uncompress_bufsize(bufsize_short);
 595                 if ((1 << get_order(bufsize)) > SG_MAX_SINGLE_ALLOC)
 596                         continue;
 597
 598                 /* check for reusable slot in the link group */
 599                 buf_desc = smc_buf_get_slot(lgr, bufsize_short, lock, buf_list);
 600                 if (buf_desc) {
 601                         memset(buf_desc->cpu_addr, 0, bufsize);
 602                         break; /* found reusable slot */
 603                 }
 604
 605                 buf_desc = smc_new_buf_create(lgr, is_rmb, bufsize);
 606                 if (PTR_ERR(buf_desc) == -ENOMEM)
 607                         break;
 608                 if (IS_ERR(buf_desc))
 609                         continue;
 610
 611                 buf_desc->used = 1;
 612                 write_lock_bh(lock);
 613                 list_add(&buf_desc->list, buf_list);
 614                 write_unlock_bh(lock);
 615                 break; /* found */
 616         }
 617
 618         if (IS_ERR(buf_desc))
 619                 return -ENOMEM;
 620
 621         if (is_rmb) {
 622                 conn->rmb_desc = buf_desc;
 623                 conn->rmbe_size = bufsize;
 624                 conn->rmbe_size_short = bufsize_short;
 625                 smc->sk.sk_rcvbuf = bufsize * 2;
 626                 atomic_set(&conn->bytes_to_rcv, 0);
 627                 conn->rmbe_update_limit = smc_rmb_wnd_update_limit(bufsize);
 628         } else {
 629                 conn->sndbuf_desc = buf_desc;
 630                 conn->sndbuf_size = bufsize;
 631                 smc->sk.sk_sndbuf = bufsize * 2;
 632                 atomic_set(&conn->sndbuf_space, bufsize);
 633         }
 634         return 0;
 635 }
 636
 637 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn)
 638 {
 639         struct smc_link_group *lgr = conn->lgr;
 640
 641         smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 642                                conn->sndbuf_desc, DMA_TO_DEVICE);
 643 }
 644
 645 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
 646 {
 647         struct smc_link_group *lgr = conn->lgr;
 648
 649         smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 650                                   conn->sndbuf_desc, DMA_TO_DEVICE);
 651 }
 652
 653 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
 654 {
 655         struct smc_link_group *lgr = conn->lgr;
 656
 657         smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 658                                conn->rmb_desc, DMA_FROM_DEVICE);
 659 }
 660
 661 void smc_rmb_sync_sg_for_device(struct smc_connection *conn)
 662 {
 663         struct smc_link_group *lgr = conn->lgr;
 664
 665         smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 666                                   conn->rmb_desc, DMA_FROM_DEVICE);
 667 }
 668
 669 /* create the send and receive buffer for an SMC socket;
 670  * receive buffers are called RMBs;
 671  * (even though the SMC protocol allows more than one RMB-element per RMB,
 672  * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
 673  * extra RMB for every connection in a link group
 674  */
 675 int smc_buf_create(struct smc_sock *smc)
 676 {
 677         int rc;
 678
 679         /* create send buffer */
 680         rc = __smc_buf_create(smc, false);
 681         if (rc)
 682                 return rc;
 683         /* create rmb */
 684         rc = __smc_buf_create(smc, true);
 685         if (rc)
 686                 smc_buf_free(smc->conn.sndbuf_desc,
 687                              &smc->conn.lgr->lnk[SMC_SINGLE_LINK], false);
 688         return rc;
 689 }
 690
 691 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
 692 {
 693         int i;
 694
 695         for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
 696                 if (!test_and_set_bit(i, lgr->rtokens_used_mask))
 697                         return i;
 698         }
 699         return -ENOSPC;
 700 }
 701
 702 /* save rkey and dma_addr received from peer during clc handshake */
 703 int smc_rmb_rtoken_handling(struct smc_connection *conn,
 704                             struct smc_clc_msg_accept_confirm *clc)
 705 {
 706         u64 dma_addr = be64_to_cpu(clc->rmb_dma_addr);
 707         struct smc_link_group *lgr = conn->lgr;
 708         u32 rkey = ntohl(clc->rmb_rkey);
 709         int i;
 710
 711         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
 712                 if ((lgr->rtokens[i][SMC_SINGLE_LINK].rkey == rkey) &&
 713                     (lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr == dma_addr) &&
 714                     test_bit(i, lgr->rtokens_used_mask)) {
 715                         conn->rtoken_idx = i;
 716                         return 0;
 717                 }
 718         }
 719         conn->rtoken_idx = smc_rmb_reserve_rtoken_idx(lgr);
 720         if (conn->rtoken_idx < 0)
 721                 return conn->rtoken_idx;
 722         lgr->rtokens[conn->rtoken_idx][SMC_SINGLE_LINK].rkey = rkey;
 723         lgr->rtokens[conn->rtoken_idx][SMC_SINGLE_LINK].dma_addr = dma_addr;
 724         return 0;
 725 }