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         read_lock_bh(&lgr->conns_lock);
 132         conns = RB_EMPTY_ROOT(&lgr->conns_all);
 133         read_unlock_bh(&lgr->conns_lock);
 134         if (!conns) { /* number of lgr connections is no longer zero */
 135                 spin_unlock_bh(&smc_lgr_list.lock);
 136                 return;
 137         }
 138         list_del_init(&lgr->list); /* remove from smc_lgr_list */
 139         spin_unlock_bh(&smc_lgr_list.lock);
 140         smc_lgr_free(lgr);
 141 }
 142
 143 /* create a new SMC link group */
 144 static int smc_lgr_create(struct smc_sock *smc, __be32 peer_in_addr,
 145                           struct smc_ib_device *smcibdev, u8 ibport,
 146                           char *peer_systemid, unsigned short vlan_id)
 147 {
 148         struct smc_link_group *lgr;
 149         struct smc_link *lnk;
 150         u8 rndvec[3];
 151         int rc = 0;
 152         int i;
 153
 154         lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
 155         if (!lgr) {
 156                 rc = -ENOMEM;
 157                 goto out;
 158         }
 159         lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
 160         lgr->sync_err = false;
 161         lgr->daddr = peer_in_addr;
 162         memcpy(lgr->peer_systemid, peer_systemid, SMC_SYSTEMID_LEN);
 163         lgr->vlan_id = vlan_id;
 164         rwlock_init(&lgr->sndbufs_lock);
 165         rwlock_init(&lgr->rmbs_lock);
 166         for (i = 0; i < SMC_RMBE_SIZES; i++) {
 167                 INIT_LIST_HEAD(&lgr->sndbufs[i]);
 168                 INIT_LIST_HEAD(&lgr->rmbs[i]);
 169         }
 170         smc_lgr_num += SMC_LGR_NUM_INCR;
 171         memcpy(&lgr->id, (u8 *)&smc_lgr_num, SMC_LGR_ID_SIZE);
 172         INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
 173         lgr->conns_all = RB_ROOT;
 174
 175         lnk = &lgr->lnk[SMC_SINGLE_LINK];
 176         /* initialize link */
 177         lnk->smcibdev = smcibdev;
 178         lnk->ibport = ibport;
 179         lnk->path_mtu = smcibdev->pattr[ibport - 1].active_mtu;
 180         if (!smcibdev->initialized)
 181                 smc_ib_setup_per_ibdev(smcibdev);
 182         get_random_bytes(rndvec, sizeof(rndvec));
 183         lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) + (rndvec[2] << 16);
 184         rc = smc_wr_alloc_link_mem(lnk);
 185         if (rc)
 186                 goto free_lgr;
 187         rc = smc_ib_create_protection_domain(lnk);
 188         if (rc)
 189                 goto free_link_mem;
 190         rc = smc_ib_create_queue_pair(lnk);
 191         if (rc)
 192                 goto dealloc_pd;
 193         rc = smc_wr_create_link(lnk);
 194         if (rc)
 195                 goto destroy_qp;
 196         init_completion(&lnk->llc_confirm);
 197         init_completion(&lnk->llc_confirm_resp);
 198
 199         smc->conn.lgr = lgr;
 200         rwlock_init(&lgr->conns_lock);
 201         spin_lock_bh(&smc_lgr_list.lock);
 202         list_add(&lgr->list, &smc_lgr_list.list);
 203         spin_unlock_bh(&smc_lgr_list.lock);
 204         return 0;
 205
 206 destroy_qp:
 207         smc_ib_destroy_queue_pair(lnk);
 208 dealloc_pd:
 209         smc_ib_dealloc_protection_domain(lnk);
 210 free_link_mem:
 211         smc_wr_free_link_mem(lnk);
 212 free_lgr:
 213         kfree(lgr);
 214 out:
 215         return rc;
 216 }
 217
 218 static void smc_buf_unuse(struct smc_connection *conn)
 219 {
 220         if (conn->sndbuf_desc) {
 221                 conn->sndbuf_desc->used = 0;
 222                 conn->sndbuf_size = 0;
 223         }
 224         if (conn->rmb_desc) {
 225                 conn->rmb_desc->reused = true;
 226                 conn->rmb_desc->used = 0;
 227                 conn->rmbe_size = 0;
 228         }
 229 }
 230
 231 /* remove a finished connection from its link group */
 232 void smc_conn_free(struct smc_connection *conn)
 233 {
 234         struct smc_link_group *lgr = conn->lgr;
 235
 236         if (!lgr)
 237                 return;
 238         smc_cdc_tx_dismiss_slots(conn);
 239         smc_lgr_unregister_conn(conn);
 240         smc_buf_unuse(conn);
 241 }
 242
 243 static void smc_link_clear(struct smc_link *lnk)
 244 {
 245         lnk->peer_qpn = 0;
 246         smc_ib_modify_qp_reset(lnk);
 247         smc_wr_free_link(lnk);
 248         smc_ib_destroy_queue_pair(lnk);
 249         smc_ib_dealloc_protection_domain(lnk);
 250         smc_wr_free_link_mem(lnk);
 251 }
 252
 253 static void smc_buf_free(struct smc_buf_desc *buf_desc, struct smc_link *lnk,
 254                          bool is_rmb)
 255 {
 256         if (is_rmb) {
 257                 if (buf_desc->mr_rx[SMC_SINGLE_LINK])
 258                         smc_ib_put_memory_region(
 259                                         buf_desc->mr_rx[SMC_SINGLE_LINK]);
 260                 smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
 261                                     DMA_FROM_DEVICE);
 262         } else {
 263                 smc_ib_buf_unmap_sg(lnk->smcibdev, buf_desc,
 264                                     DMA_TO_DEVICE);
 265         }
 266         sg_free_table(&buf_desc->sgt[SMC_SINGLE_LINK]);
 267         if (buf_desc->cpu_addr)
 268                 free_pages((unsigned long)buf_desc->cpu_addr, buf_desc->order);
 269         kfree(buf_desc);
 270 }
 271
 272 static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
 273 {
 274         struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
 275         struct smc_buf_desc *buf_desc, *bf_desc;
 276         struct list_head *buf_list;
 277         int i;
 278
 279         for (i = 0; i < SMC_RMBE_SIZES; i++) {
 280                 if (is_rmb)
 281                         buf_list = &lgr->rmbs[i];
 282                 else
 283                         buf_list = &lgr->sndbufs[i];
 284                 list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
 285                                          list) {
 286                         list_del(&buf_desc->list);
 287                         smc_buf_free(buf_desc, lnk, is_rmb);
 288                 }
 289         }
 290 }
 291
 292 static void smc_lgr_free_bufs(struct smc_link_group *lgr)
 293 {
 294         /* free send buffers */
 295         __smc_lgr_free_bufs(lgr, false);
 296         /* free rmbs */
 297         __smc_lgr_free_bufs(lgr, true);
 298 }
 299
 300 /* remove a link group */
 301 void smc_lgr_free(struct smc_link_group *lgr)
 302 {
 303         smc_lgr_free_bufs(lgr);
 304         smc_link_clear(&lgr->lnk[SMC_SINGLE_LINK]);
 305         kfree(lgr);
 306 }
 307
 308 /* terminate linkgroup abnormally */
 309 void smc_lgr_terminate(struct smc_link_group *lgr)
 310 {
 311         struct smc_connection *conn;
 312         struct smc_sock *smc;
 313         struct rb_node *node;
 314
 315         spin_lock_bh(&smc_lgr_list.lock);
 316         if (list_empty(&lgr->list)) {
 317                 /* termination already triggered */
 318                 spin_unlock_bh(&smc_lgr_list.lock);
 319                 return;
 320         }
 321         /* do not use this link group for new connections */
 322         list_del_init(&lgr->list);
 323         spin_unlock_bh(&smc_lgr_list.lock);
 324
 325         write_lock_bh(&lgr->conns_lock);
 326         node = rb_first(&lgr->conns_all);
 327         while (node) {
 328                 conn = rb_entry(node, struct smc_connection, alert_node);
 329                 smc = container_of(conn, struct smc_sock, conn);
 330                 sock_hold(&smc->sk);
 331                 __smc_lgr_unregister_conn(conn);
 332                 schedule_work(&conn->close_work);
 333                 sock_put(&smc->sk);
 334                 node = rb_first(&lgr->conns_all);
 335         }
 336         write_unlock_bh(&lgr->conns_lock);
 337 }
 338
 339 /* Determine vlan of internal TCP socket.
 340  * @vlan_id: address to store the determined vlan id into
 341  */
 342 static int smc_vlan_by_tcpsk(struct socket *clcsock, unsigned short *vlan_id)
 343 {
 344         struct dst_entry *dst = sk_dst_get(clcsock->sk);
 345         int rc = 0;
 346
 347         *vlan_id = 0;
 348         if (!dst) {
 349                 rc = -ENOTCONN;
 350                 goto out;
 351         }
 352         if (!dst->dev) {
 353                 rc = -ENODEV;
 354                 goto out_rel;
 355         }
 356
 357         if (is_vlan_dev(dst->dev))
 358                 *vlan_id = vlan_dev_vlan_id(dst->dev);
 359
 360 out_rel:
 361         dst_release(dst);
 362 out:
 363         return rc;
 364 }
 365
 366 /* determine the link gid matching the vlan id of the link group */
 367 static int smc_link_determine_gid(struct smc_link_group *lgr)
 368 {
 369         struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
 370         struct ib_gid_attr gattr;
 371         union ib_gid gid;
 372         int i;
 373
 374         if (!lgr->vlan_id) {
 375                 lnk->gid = lnk->smcibdev->gid[lnk->ibport - 1];
 376                 return 0;
 377         }
 378
 379         for (i = 0; i < lnk->smcibdev->pattr[lnk->ibport - 1].gid_tbl_len;
 380              i++) {
 381                 if (ib_query_gid(lnk->smcibdev->ibdev, lnk->ibport, i, &gid,
 382                                  &gattr))
 383                         continue;
 384                 if (gattr.ndev &&
 385                     (vlan_dev_vlan_id(gattr.ndev) == lgr->vlan_id)) {
 386                         lnk->gid = gid;
 387                         return 0;
 388                 }
 389         }
 390         return -ENODEV;
 391 }
 392
 393 /* create a new SMC connection (and a new link group if necessary) */
 394 int smc_conn_create(struct smc_sock *smc, __be32 peer_in_addr,
 395                     struct smc_ib_device *smcibdev, u8 ibport,
 396                     struct smc_clc_msg_local *lcl, int srv_first_contact)
 397 {
 398         struct smc_connection *conn = &smc->conn;
 399         struct smc_link_group *lgr;
 400         unsigned short vlan_id;
 401         enum smc_lgr_role role;
 402         int local_contact = SMC_FIRST_CONTACT;
 403         int rc = 0;
 404
 405         role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
 406         rc = smc_vlan_by_tcpsk(smc->clcsock, &vlan_id);
 407         if (rc)
 408                 return rc;
 409
 410         if ((role == SMC_CLNT) && srv_first_contact)
 411                 /* create new link group as well */
 412                 goto create;
 413
 414         /* determine if an existing link group can be reused */
 415         spin_lock_bh(&smc_lgr_list.lock);
 416         list_for_each_entry(lgr, &smc_lgr_list.list, list) {
 417                 write_lock_bh(&lgr->conns_lock);
 418                 if (!memcmp(lgr->peer_systemid, lcl->id_for_peer,
 419                             SMC_SYSTEMID_LEN) &&
 420                     !memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_gid, &lcl->gid,
 421                             SMC_GID_SIZE) &&
 422                     !memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_mac, lcl->mac,
 423                             sizeof(lcl->mac)) &&
 424                     !lgr->sync_err &&
 425                     (lgr->role == role) &&
 426                     (lgr->vlan_id == vlan_id) &&
 427                     ((role == SMC_CLNT) ||
 428                      (lgr->conns_num < SMC_RMBS_PER_LGR_MAX))) {
 429                         /* link group found */
 430                         local_contact = SMC_REUSE_CONTACT;
 431                         conn->lgr = lgr;
 432                         smc_lgr_register_conn(conn); /* add smc conn to lgr */
 433                         write_unlock_bh(&lgr->conns_lock);
 434                         break;
 435                 }
 436                 write_unlock_bh(&lgr->conns_lock);
 437         }
 438         spin_unlock_bh(&smc_lgr_list.lock);
 439
 440         if (role == SMC_CLNT && !srv_first_contact &&
 441             (local_contact == SMC_FIRST_CONTACT)) {
 442                 /* Server reuses a link group, but Client wants to start
 443                  * a new one
 444                  * send out_of_sync decline, reason synchr. error
 445                  */
 446                 return -ENOLINK;
 447         }
 448
 449 create:
 450         if (local_contact == SMC_FIRST_CONTACT) {
 451                 rc = smc_lgr_create(smc, peer_in_addr, smcibdev, ibport,
 452                                     lcl->id_for_peer, vlan_id);
 453                 if (rc)
 454                         goto out;
 455                 smc_lgr_register_conn(conn); /* add smc conn to lgr */
 456                 rc = smc_link_determine_gid(conn->lgr);
 457         }
 458         conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
 459         conn->local_tx_ctrl.len = sizeof(struct smc_cdc_msg);
 460 #ifndef KERNEL_HAS_ATOMIC64
 461         spin_lock_init(&conn->acurs_lock);
 462 #endif
 463
 464 out:
 465         return rc ? rc : local_contact;
 466 }
 467
 468 /* try to reuse a sndbuf or rmb description slot for a certain
 469  * buffer size; if not available, return NULL
 470  */
 471 static inline
 472 struct smc_buf_desc *smc_buf_get_slot(struct smc_link_group *lgr,
 473                                       int compressed_bufsize,
 474                                       rwlock_t *lock,
 475                                       struct list_head *buf_list)
 476 {
 477         struct smc_buf_desc *buf_slot;
 478
 479         read_lock_bh(lock);
 480         list_for_each_entry(buf_slot, buf_list, list) {
 481                 if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
 482                         read_unlock_bh(lock);
 483                         return buf_slot;
 484                 }
 485         }
 486         read_unlock_bh(lock);
 487         return NULL;
 488 }
 489
 490 /* one of the conditions for announcing a receiver's current window size is
 491  * that it "results in a minimum increase in the window size of 10% of the
 492  * receive buffer space" [RFC7609]
 493  */
 494 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
 495 {
 496         return min_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
 497 }
 498
 499 static struct smc_buf_desc *smc_new_buf_create(struct smc_link_group *lgr,
 500                                                bool is_rmb, int bufsize)
 501 {
 502         struct smc_buf_desc *buf_desc;
 503         struct smc_link *lnk;
 504         int rc;
 505
 506         /* try to alloc a new buffer */
 507         buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
 508         if (!buf_desc)
 509                 return ERR_PTR(-ENOMEM);
 510
 511         buf_desc->cpu_addr =
 512                 (void *)__get_free_pages(GFP_KERNEL | __GFP_NOWARN |
 513                                          __GFP_NOMEMALLOC |
 514                                          __GFP_NORETRY | __GFP_ZERO,
 515                                          get_order(bufsize));
 516         if (!buf_desc->cpu_addr) {
 517                 kfree(buf_desc);
 518                 return ERR_PTR(-EAGAIN);
 519         }
 520         buf_desc->order = get_order(bufsize);
 521
 522         /* build the sg table from the pages */
 523         lnk = &lgr->lnk[SMC_SINGLE_LINK];
 524         rc = sg_alloc_table(&buf_desc->sgt[SMC_SINGLE_LINK], 1,
 525                             GFP_KERNEL);
 526         if (rc) {
 527                 smc_buf_free(buf_desc, lnk, is_rmb);
 528                 return ERR_PTR(rc);
 529         }
 530         sg_set_buf(buf_desc->sgt[SMC_SINGLE_LINK].sgl,
 531                    buf_desc->cpu_addr, bufsize);
 532
 533         /* map sg table to DMA address */
 534         rc = smc_ib_buf_map_sg(lnk->smcibdev, buf_desc,
 535                                is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
 536         /* SMC protocol depends on mapping to one DMA address only */
 537         if (rc != 1)  {
 538                 smc_buf_free(buf_desc, lnk, is_rmb);
 539                 return ERR_PTR(-EAGAIN);
 540         }
 541
 542         /* create a new memory region for the RMB */
 543         if (is_rmb) {
 544                 rc = smc_ib_get_memory_region(lnk->roce_pd,
 545                                               IB_ACCESS_REMOTE_WRITE |
 546                                               IB_ACCESS_LOCAL_WRITE,
 547                                               buf_desc);
 548                 if (rc) {
 549                         smc_buf_free(buf_desc, lnk, is_rmb);
 550                         return ERR_PTR(rc);
 551                 }
 552         }
 553
 554         return buf_desc;
 555 }
 556
 557 static int __smc_buf_create(struct smc_sock *smc, bool is_rmb)
 558 {
 559         struct smc_connection *conn = &smc->conn;
 560         struct smc_link_group *lgr = conn->lgr;
 561         struct smc_buf_desc *buf_desc = NULL;
 562         struct list_head *buf_list;
 563         int bufsize, bufsize_short;
 564         int sk_buf_size;
 565         rwlock_t *lock;
 566
 567         if (is_rmb)
 568                 /* use socket recv buffer size (w/o overhead) as start value */
 569                 sk_buf_size = smc->sk.sk_rcvbuf / 2;
 570         else
 571                 /* use socket send buffer size (w/o overhead) as start value */
 572                 sk_buf_size = smc->sk.sk_sndbuf / 2;
 573
 574         for (bufsize_short = smc_compress_bufsize(smc->sk.sk_sndbuf / 2);
 575              bufsize_short >= 0; bufsize_short--) {
 576
 577                 if (is_rmb) {
 578                         lock = &lgr->rmbs_lock;
 579                         buf_list = &lgr->rmbs[bufsize_short];
 580                 } else {
 581                         lock = &lgr->sndbufs_lock;
 582                         buf_list = &lgr->sndbufs[bufsize_short];
 583                 }
 584                 bufsize = smc_uncompress_bufsize(bufsize_short);
 585                 if ((1 << get_order(bufsize)) > SG_MAX_SINGLE_ALLOC)
 586                         continue;
 587
 588                 /* check for reusable slot in the link group */
 589                 buf_desc = smc_buf_get_slot(lgr, bufsize_short, lock, buf_list);
 590                 if (buf_desc) {
 591                         memset(buf_desc->cpu_addr, 0, bufsize);
 592                         break; /* found reusable slot */
 593                 }
 594
 595                 buf_desc = smc_new_buf_create(lgr, is_rmb, bufsize);
 596                 if (PTR_ERR(buf_desc) == -ENOMEM)
 597                         break;
 598                 if (IS_ERR(buf_desc))
 599                         continue;
 600
 601                 buf_desc->used = 1;
 602                 write_lock_bh(lock);
 603                 list_add(&buf_desc->list, buf_list);
 604                 write_unlock_bh(lock);
 605                 break; /* found */
 606         }
 607
 608         if (IS_ERR(buf_desc))
 609                 return -ENOMEM;
 610
 611         if (is_rmb) {
 612                 conn->rmb_desc = buf_desc;
 613                 conn->rmbe_size = bufsize;
 614                 conn->rmbe_size_short = bufsize_short;
 615                 smc->sk.sk_rcvbuf = bufsize * 2;
 616                 atomic_set(&conn->bytes_to_rcv, 0);
 617                 conn->rmbe_update_limit = smc_rmb_wnd_update_limit(bufsize);
 618         } else {
 619                 conn->sndbuf_desc = buf_desc;
 620                 conn->sndbuf_size = bufsize;
 621                 smc->sk.sk_sndbuf = bufsize * 2;
 622                 atomic_set(&conn->sndbuf_space, bufsize);
 623         }
 624         return 0;
 625 }
 626
 627 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn)
 628 {
 629         struct smc_link_group *lgr = conn->lgr;
 630
 631         smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 632                                conn->sndbuf_desc, DMA_TO_DEVICE);
 633 }
 634
 635 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
 636 {
 637         struct smc_link_group *lgr = conn->lgr;
 638
 639         smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 640                                   conn->sndbuf_desc, DMA_TO_DEVICE);
 641 }
 642
 643 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
 644 {
 645         struct smc_link_group *lgr = conn->lgr;
 646
 647         smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 648                                conn->rmb_desc, DMA_FROM_DEVICE);
 649 }
 650
 651 void smc_rmb_sync_sg_for_device(struct smc_connection *conn)
 652 {
 653         struct smc_link_group *lgr = conn->lgr;
 654
 655         smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 656                                   conn->rmb_desc, DMA_FROM_DEVICE);
 657 }
 658
 659 /* create the send and receive buffer for an SMC socket;
 660  * receive buffers are called RMBs;
 661  * (even though the SMC protocol allows more than one RMB-element per RMB,
 662  * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
 663  * extra RMB for every connection in a link group
 664  */
 665 int smc_buf_create(struct smc_sock *smc)
 666 {
 667         int rc;
 668
 669         /* create send buffer */
 670         rc = __smc_buf_create(smc, false);
 671         if (rc)
 672                 return rc;
 673         /* create rmb */
 674         rc = __smc_buf_create(smc, true);
 675         if (rc)
 676                 smc_buf_free(smc->conn.sndbuf_desc,
 677                              &smc->conn.lgr->lnk[SMC_SINGLE_LINK], false);
 678         return rc;
 679 }
 680
 681 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
 682 {
 683         int i;
 684
 685         for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
 686                 if (!test_and_set_bit(i, lgr->rtokens_used_mask))
 687                         return i;
 688         }
 689         return -ENOSPC;
 690 }
 691
 692 /* save rkey and dma_addr received from peer during clc handshake */
 693 int smc_rmb_rtoken_handling(struct smc_connection *conn,
 694                             struct smc_clc_msg_accept_confirm *clc)
 695 {
 696         u64 dma_addr = be64_to_cpu(clc->rmb_dma_addr);
 697         struct smc_link_group *lgr = conn->lgr;
 698         u32 rkey = ntohl(clc->rmb_rkey);
 699         int i;
 700
 701         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
 702                 if ((lgr->rtokens[i][SMC_SINGLE_LINK].rkey == rkey) &&
 703                     (lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr == dma_addr) &&
 704                     test_bit(i, lgr->rtokens_used_mask)) {
 705                         conn->rtoken_idx = i;
 706                         return 0;
 707                 }
 708         }
 709         conn->rtoken_idx = smc_rmb_reserve_rtoken_idx(lgr);
 710         if (conn->rtoken_idx < 0)
 711                 return conn->rtoken_idx;
 712         lgr->rtokens[conn->rtoken_idx][SMC_SINGLE_LINK].rkey = rkey;
 713         lgr->rtokens[conn->rtoken_idx][SMC_SINGLE_LINK].dma_addr = dma_addr;
 714         return 0;
 715 }