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
  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->smcibdev = smcibdev;
 181         lnk->ibport = ibport;
 182         lnk->path_mtu = smcibdev->pattr[ibport - 1].active_mtu;
 183         if (!smcibdev->initialized)
 184                 smc_ib_setup_per_ibdev(smcibdev);
 185         get_random_bytes(rndvec, sizeof(rndvec));
 186         lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) + (rndvec[2] << 16);
 187         rc = smc_wr_alloc_link_mem(lnk);
 188         if (rc)
 189                 goto free_lgr;
 190         rc = smc_ib_create_protection_domain(lnk);
 191         if (rc)
 192                 goto free_link_mem;
 193         rc = smc_ib_create_queue_pair(lnk);
 194         if (rc)
 195                 goto dealloc_pd;
 196         rc = smc_wr_create_link(lnk);
 197         if (rc)
 198                 goto destroy_qp;
 199         init_completion(&lnk->llc_confirm);
 200         init_completion(&lnk->llc_confirm_resp);
 201
 202         smc->conn.lgr = lgr;
 203         rwlock_init(&lgr->conns_lock);
 204         spin_lock_bh(&smc_lgr_list.lock);
 205         list_add(&lgr->list, &smc_lgr_list.list);
 206         spin_unlock_bh(&smc_lgr_list.lock);
 207         return 0;
 208
 209 destroy_qp:
 210         smc_ib_destroy_queue_pair(lnk);
 211 dealloc_pd:
 212         smc_ib_dealloc_protection_domain(lnk);
 213 free_link_mem:
 214         smc_wr_free_link_mem(lnk);
 215 free_lgr:
 216         kfree(lgr);
 217 out:
 218         return rc;
 219 }
 220
 221 static void smc_buf_unuse(struct smc_connection *conn)
 222 {
 223         if (conn->sndbuf_desc) {
 224                 conn->sndbuf_desc->used = 0;
 225                 conn->sndbuf_size = 0;
 226         }
 227         if (conn->rmb_desc) {
 228                 conn->rmb_desc->reused = true;
 229                 conn->rmb_desc->used = 0;
 230                 conn->rmbe_size = 0;
 231         }
 232 }
 233
 234 /* remove a finished connection from its link group */
 235 void smc_conn_free(struct smc_connection *conn)
 236 {
 237         if (!conn->lgr)
 238                 return;
 239         smc_cdc_tx_dismiss_slots(conn);
 240         smc_lgr_unregister_conn(conn);
 241         smc_buf_unuse(conn);
 242 }
 243
 244 static void smc_link_clear(struct smc_link *lnk)
 245 {
 246         lnk->peer_qpn = 0;
 247         smc_ib_modify_qp_reset(lnk);
 248         smc_wr_free_link(lnk);
 249         smc_ib_destroy_queue_pair(lnk);
 250         smc_ib_dealloc_protection_domain(lnk);
 251         smc_wr_free_link_mem(lnk);
 252 }
 253
 254 static void smc_buf_free(struct smc_buf_desc *buf_desc, struct smc_link *lnk,
 255                          bool is_rmb)
 256 {
 257         if (is_rmb) {
 258                 if (buf_desc->mr_rx[SMC_SINGLE_LINK])
 259                         smc_ib_put_memory_region(
 260                                         buf_desc->mr_rx[SMC_SINGLE_LINK]);
 261                 smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
 262                                     DMA_FROM_DEVICE);
 263         } else {
 264                 smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
 265                                     DMA_TO_DEVICE);
 266         }
 267         sg_free_table(&buf_desc->sgt[SMC_SINGLE_LINK]);
 268         if (buf_desc->cpu_addr)
 269                 free_pages((unsigned long)buf_desc->cpu_addr, buf_desc->order);
 270         kfree(buf_desc);
 271 }
 272
 273 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
 274 {
 275         struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
 276         struct smc_buf_desc *buf_desc, *bf_desc;
 277         struct list_head *buf_list;
 278         int i;
 279
 280         for (i = 0; i < SMC_RMBE_SIZES; i++) {
 281                 if (is_rmb)
 282                         buf_list = &lgr->rmbs[i];
 283                 else
 284                         buf_list = &lgr->sndbufs[i];
 285                 list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
 286                                          list) {
 287                         list_del(&buf_desc->list);
 288                         smc_buf_free(buf_desc, lnk, is_rmb);
 289                 }
 290         }
 291 }
 292
 293 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
 294 {
 295         /* free send buffers */
 296         __smc_lgr_free_bufs(lgr, false);
 297         /* free rmbs */
 298         __smc_lgr_free_bufs(lgr, true);
 299 }
 300
 301 /* remove a link group */
 302 void smc_lgr_free(struct smc_link_group *lgr)
 303 {
 304         smc_lgr_free_bufs(lgr);
 305         smc_link_clear(&lgr->lnk[SMC_SINGLE_LINK]);
 306         kfree(lgr);
 307 }
 308
 309 /* terminate linkgroup abnormally */
 310 void smc_lgr_terminate(struct smc_link_group *lgr)
 311 {
 312         struct smc_connection *conn;
 313         struct smc_sock *smc;
 314         struct rb_node *node;
 315
 316         spin_lock_bh(&smc_lgr_list.lock);
 317         if (list_empty(&lgr->list)) {
 318                 /* termination already triggered */
 319                 spin_unlock_bh(&smc_lgr_list.lock);
 320                 return;
 321         }
 322         /* do not use this link group for new connections */
 323         list_del_init(&lgr->list);
 324         spin_unlock_bh(&smc_lgr_list.lock);
 325
 326         write_lock_bh(&lgr->conns_lock);
 327         node = rb_first(&lgr->conns_all);
 328         while (node) {
 329                 conn = rb_entry(node, struct smc_connection, alert_node);
 330                 smc = container_of(conn, struct smc_sock, conn);
 331                 sock_hold(&smc->sk); /* sock_put in close work */
 332                 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
 333                 __smc_lgr_unregister_conn(conn);
 334                 write_unlock_bh(&lgr->conns_lock);
 335                 if (!schedule_work(&conn->close_work))
 336                         sock_put(&smc->sk);
 337                 write_lock_bh(&lgr->conns_lock);
 338                 node = rb_first(&lgr->conns_all);
 339         }
 340         write_unlock_bh(&lgr->conns_lock);
 341         wake_up(&lgr->lnk[SMC_SINGLE_LINK].wr_reg_wait);
 342 }
 343
 344 /* Determine vlan of internal TCP socket.
 345  * @vlan_id: address to store the determined vlan id into
 346  */
 347 static int smc_vlan_by_tcpsk(struct socket *clcsock, unsigned short *vlan_id)
 348 {
 349         struct dst_entry *dst = sk_dst_get(clcsock->sk);
 350         int rc = 0;
 351
 352         *vlan_id = 0;
 353         if (!dst) {
 354                 rc = -ENOTCONN;
 355                 goto out;
 356         }
 357         if (!dst->dev) {
 358                 rc = -ENODEV;
 359                 goto out_rel;
 360         }
 361
 362         if (is_vlan_dev(dst->dev))
 363                 *vlan_id = vlan_dev_vlan_id(dst->dev);
 364
 365 out_rel:
 366         dst_release(dst);
 367 out:
 368         return rc;
 369 }
 370
 371 /* determine the link gid matching the vlan id of the link group */
 372 static int smc_link_determine_gid(struct smc_link_group *lgr)
 373 {
 374         struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
 375         struct ib_gid_attr gattr;
 376         union ib_gid gid;
 377         int i;
 378
 379         if (!lgr->vlan_id) {
 380                 lnk->gid = lnk->smcibdev->gid[lnk->ibport - 1];
 381                 return 0;
 382         }
 383
 384         for (i = 0; i < lnk->smcibdev->pattr[lnk->ibport - 1].gid_tbl_len;
 385              i++) {
 386                 if (ib_query_gid(lnk->smcibdev->ibdev, lnk->ibport, i, &gid,
 387                                  &gattr))
 388                         continue;
 389                 if (gattr.ndev) {
 390                         if (is_vlan_dev(gattr.ndev) &&
 391                             vlan_dev_vlan_id(gattr.ndev) == lgr->vlan_id) {
 392                                 lnk->gid = gid;
 393                                 dev_put(gattr.ndev);
 394                                 return 0;
 395                         }
 396                         dev_put(gattr.ndev);
 397                 }
 398         }
 399         return -ENODEV;
 400 }
 401
 402 /* create a new SMC connection (and a new link group if necessary) */
 403 int smc_conn_create(struct smc_sock *smc, __be32 peer_in_addr,
 404                     struct smc_ib_device *smcibdev, u8 ibport,
 405                     struct smc_clc_msg_local *lcl, int srv_first_contact)
 406 {
 407         struct smc_connection *conn = &smc->conn;
 408         struct smc_link_group *lgr;
 409         unsigned short vlan_id;
 410         enum smc_lgr_role role;
 411         int local_contact = SMC_FIRST_CONTACT;
 412         int rc = 0;
 413
 414         role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
 415         rc = smc_vlan_by_tcpsk(smc->clcsock, &vlan_id);
 416         if (rc)
 417                 return rc;
 418
 419         if ((role == SMC_CLNT) && srv_first_contact)
 420                 /* create new link group as well */
 421                 goto create;
 422
 423         /* determine if an existing link group can be reused */
 424         spin_lock_bh(&smc_lgr_list.lock);
 425         list_for_each_entry(lgr, &smc_lgr_list.list, list) {
 426                 write_lock_bh(&lgr->conns_lock);
 427                 if (!memcmp(lgr->peer_systemid, lcl->id_for_peer,
 428                             SMC_SYSTEMID_LEN) &&
 429                     !memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_gid, &lcl->gid,
 430                             SMC_GID_SIZE) &&
 431                     !memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_mac, lcl->mac,
 432                             sizeof(lcl->mac)) &&
 433                     !lgr->sync_err &&
 434                     (lgr->role == role) &&
 435                     (lgr->vlan_id == vlan_id) &&
 436                     ((role == SMC_CLNT) ||
 437                      (lgr->conns_num < SMC_RMBS_PER_LGR_MAX))) {
 438                         /* link group found */
 439                         local_contact = SMC_REUSE_CONTACT;
 440                         conn->lgr = lgr;
 441                         smc_lgr_register_conn(conn); /* add smc conn to lgr */
 442                         write_unlock_bh(&lgr->conns_lock);
 443                         break;
 444                 }
 445                 write_unlock_bh(&lgr->conns_lock);
 446         }
 447         spin_unlock_bh(&smc_lgr_list.lock);
 448
 449         if (role == SMC_CLNT && !srv_first_contact &&
 450             (local_contact == SMC_FIRST_CONTACT)) {
 451                 /* Server reuses a link group, but Client wants to start
 452                  * a new one
 453                  * send out_of_sync decline, reason synchr. error
 454                  */
 455                 return -ENOLINK;
 456         }
 457
 458 create:
 459         if (local_contact == SMC_FIRST_CONTACT) {
 460                 rc = smc_lgr_create(smc, peer_in_addr, smcibdev, ibport,
 461                                     lcl->id_for_peer, vlan_id);
 462                 if (rc)
 463                         goto out;
 464                 smc_lgr_register_conn(conn); /* add smc conn to lgr */
 465                 rc = smc_link_determine_gid(conn->lgr);
 466         }
 467         conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
 468         conn->local_tx_ctrl.len = sizeof(struct smc_cdc_msg);
 469 #ifndef KERNEL_HAS_ATOMIC64
 470         spin_lock_init(&conn->acurs_lock);
 471 #endif
 472
 473 out:
 474         return rc ? rc : local_contact;
 475 }
 476
 477 /* try to reuse a sndbuf or rmb description slot for a certain
 478  * buffer size; if not available, return NULL
 479  */
 480 static inline
 481 struct smc_buf_desc *smc_buf_get_slot(struct smc_link_group *lgr,
 482                                       int compressed_bufsize,
 483                                       rwlock_t *lock,
 484                                       struct list_head *buf_list)
 485 {
 486         struct smc_buf_desc *buf_slot;
 487
 488         read_lock_bh(lock);
 489         list_for_each_entry(buf_slot, buf_list, list) {
 490                 if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
 491                         read_unlock_bh(lock);
 492                         return buf_slot;
 493                 }
 494         }
 495         read_unlock_bh(lock);
 496         return NULL;
 497 }
 498
 499 /* one of the conditions for announcing a receiver's current window size is
 500  * that it "results in a minimum increase in the window size of 10% of the
 501  * receive buffer space" [RFC7609]
 502  */
 503 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
 504 {
 505         return min_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
 506 }
 507
 508 static struct smc_buf_desc *smc_new_buf_create(struct smc_link_group *lgr,
 509                                                bool is_rmb, int bufsize)
 510 {
 511         struct smc_buf_desc *buf_desc;
 512         struct smc_link *lnk;
 513         int rc;
 514
 515         /* try to alloc a new buffer */
 516         buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
 517         if (!buf_desc)
 518                 return ERR_PTR(-ENOMEM);
 519
 520         buf_desc->cpu_addr =
 521                 (void *)__get_free_pages(GFP_KERNEL | __GFP_NOWARN |
 522                                          __GFP_NOMEMALLOC |
 523                                          __GFP_NORETRY | __GFP_ZERO,
 524                                          get_order(bufsize));
 525         if (!buf_desc->cpu_addr) {
 526                 kfree(buf_desc);
 527                 return ERR_PTR(-EAGAIN);
 528         }
 529         buf_desc->order = get_order(bufsize);
 530
 531         /* build the sg table from the pages */
 532         lnk = &lgr->lnk[SMC_SINGLE_LINK];
 533         rc = sg_alloc_table(&buf_desc->sgt[SMC_SINGLE_LINK], 1,
 534                             GFP_KERNEL);
 535         if (rc) {
 536                 smc_buf_free(buf_desc, lnk, is_rmb);
 537                 return ERR_PTR(rc);
 538         }
 539         sg_set_buf(buf_desc->sgt[SMC_SINGLE_LINK].sgl,
 540                    buf_desc->cpu_addr, bufsize);
 541
 542         /* map sg table to DMA address */
 543         rc = smc_ib_buf_map_sg(lnk->smcibdev, buf_desc,
 544                                is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
 545         /* SMC protocol depends on mapping to one DMA address only */
 546         if (rc != 1)  {
 547                 smc_buf_free(buf_desc, lnk, is_rmb);
 548                 return ERR_PTR(-EAGAIN);
 549         }
 550
 551         /* create a new memory region for the RMB */
 552         if (is_rmb) {
 553                 rc = smc_ib_get_memory_region(lnk->roce_pd,
 554                                               IB_ACCESS_REMOTE_WRITE |
 555                                               IB_ACCESS_LOCAL_WRITE,
 556                                               buf_desc);
 557                 if (rc) {
 558                         smc_buf_free(buf_desc, lnk, is_rmb);
 559                         return ERR_PTR(rc);
 560                 }
 561         }
 562
 563         return buf_desc;
 564 }
 565
 566 static int __smc_buf_create(struct smc_sock *smc, bool is_rmb)
 567 {
 568         struct smc_connection *conn = &smc->conn;
 569         struct smc_link_group *lgr = conn->lgr;
 570         struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
 571         struct list_head *buf_list;
 572         int bufsize, bufsize_short;
 573         int sk_buf_size;
 574         rwlock_t *lock;
 575
 576         if (is_rmb)
 577                 /* use socket recv buffer size (w/o overhead) as start value */
 578                 sk_buf_size = smc->sk.sk_rcvbuf / 2;
 579         else
 580                 /* use socket send buffer size (w/o overhead) as start value */
 581                 sk_buf_size = smc->sk.sk_sndbuf / 2;
 582
 583         for (bufsize_short = smc_compress_bufsize(sk_buf_size);
 584              bufsize_short >= 0; bufsize_short--) {
 585
 586                 if (is_rmb) {
 587                         lock = &lgr->rmbs_lock;
 588                         buf_list = &lgr->rmbs[bufsize_short];
 589                 } else {
 590                         lock = &lgr->sndbufs_lock;
 591                         buf_list = &lgr->sndbufs[bufsize_short];
 592                 }
 593                 bufsize = smc_uncompress_bufsize(bufsize_short);
 594                 if ((1 << get_order(bufsize)) > SG_MAX_SINGLE_ALLOC)
 595                         continue;
 596
 597                 /* check for reusable slot in the link group */
 598                 buf_desc = smc_buf_get_slot(lgr, bufsize_short, lock, buf_list);
 599                 if (buf_desc) {
 600                         memset(buf_desc->cpu_addr, 0, bufsize);
 601                         break; /* found reusable slot */
 602                 }
 603
 604                 buf_desc = smc_new_buf_create(lgr, is_rmb, bufsize);
 605                 if (PTR_ERR(buf_desc) == -ENOMEM)
 606                         break;
 607                 if (IS_ERR(buf_desc))
 608                         continue;
 609
 610                 buf_desc->used = 1;
 611                 write_lock_bh(lock);
 612                 list_add(&buf_desc->list, buf_list);
 613                 write_unlock_bh(lock);
 614                 break; /* found */
 615         }
 616
 617         if (IS_ERR(buf_desc))
 618                 return -ENOMEM;
 619
 620         if (is_rmb) {
 621                 conn->rmb_desc = buf_desc;
 622                 conn->rmbe_size = bufsize;
 623                 conn->rmbe_size_short = bufsize_short;
 624                 smc->sk.sk_rcvbuf = bufsize * 2;
 625                 atomic_set(&conn->bytes_to_rcv, 0);
 626                 conn->rmbe_update_limit = smc_rmb_wnd_update_limit(bufsize);
 627         } else {
 628                 conn->sndbuf_desc = buf_desc;
 629                 conn->sndbuf_size = bufsize;
 630                 smc->sk.sk_sndbuf = bufsize * 2;
 631                 atomic_set(&conn->sndbuf_space, bufsize);
 632         }
 633         return 0;
 634 }
 635
 636 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn)
 637 {
 638         struct smc_link_group *lgr = conn->lgr;
 639
 640         smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 641                                conn->sndbuf_desc, DMA_TO_DEVICE);
 642 }
 643
 644 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
 645 {
 646         struct smc_link_group *lgr = conn->lgr;
 647
 648         smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 649                                   conn->sndbuf_desc, DMA_TO_DEVICE);
 650 }
 651
 652 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
 653 {
 654         struct smc_link_group *lgr = conn->lgr;
 655
 656         smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 657                                conn->rmb_desc, DMA_FROM_DEVICE);
 658 }
 659
 660 void smc_rmb_sync_sg_for_device(struct smc_connection *conn)
 661 {
 662         struct smc_link_group *lgr = conn->lgr;
 663
 664         smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 665                                   conn->rmb_desc, DMA_FROM_DEVICE);
 666 }
 667
 668 /* create the send and receive buffer for an SMC socket;
 669  * receive buffers are called RMBs;
 670  * (even though the SMC protocol allows more than one RMB-element per RMB,
 671  * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
 672  * extra RMB for every connection in a link group
 673  */
 674 int smc_buf_create(struct smc_sock *smc)
 675 {
 676         int rc;
 677
 678         /* create send buffer */
 679         rc = __smc_buf_create(smc, false);
 680         if (rc)
 681                 return rc;
 682         /* create rmb */
 683         rc = __smc_buf_create(smc, true);
 684         if (rc)
 685                 smc_buf_free(smc->conn.sndbuf_desc,
 686                              &smc->conn.lgr->lnk[SMC_SINGLE_LINK], false);
 687         return rc;
 688 }
 689
 690 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
 691 {
 692         int i;
 693
 694         for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
 695                 if (!test_and_set_bit(i, lgr->rtokens_used_mask))
 696                         return i;
 697         }
 698         return -ENOSPC;
 699 }
 700
 701 /* save rkey and dma_addr received from peer during clc handshake */
 702 int smc_rmb_rtoken_handling(struct smc_connection *conn,
 703                             struct smc_clc_msg_accept_confirm *clc)
 704 {
 705         u64 dma_addr = be64_to_cpu(clc->rmb_dma_addr);
 706         struct smc_link_group *lgr = conn->lgr;
 707         u32 rkey = ntohl(clc->rmb_rkey);
 708         int i;
 709
 710         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
 711                 if ((lgr->rtokens[i][SMC_SINGLE_LINK].rkey == rkey) &&
 712                     (lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr == dma_addr) &&
 713                     test_bit(i, lgr->rtokens_used_mask)) {
 714                         conn->rtoken_idx = i;
 715                         return 0;
 716                 }
 717         }
 718         conn->rtoken_idx = smc_rmb_reserve_rtoken_idx(lgr);
 719         if (conn->rtoken_idx < 0)
 720                 return conn->rtoken_idx;
 721         lgr->rtokens[conn->rtoken_idx][SMC_SINGLE_LINK].rkey = rkey;
 722         lgr->rtokens[conn->rtoken_idx][SMC_SINGLE_LINK].dma_addr = dma_addr;
 723         return 0;
 724 }