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