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,
 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         memcpy(lgr->peer_systemid, peer_systemid, SMC_SYSTEMID_LEN);
 165         lgr->vlan_id = vlan_id;
 166         rwlock_init(&lgr->sndbufs_lock);
 167         rwlock_init(&lgr->rmbs_lock);
 168         for (i = 0; i < SMC_RMBE_SIZES; i++) {
 169                 INIT_LIST_HEAD(&lgr->sndbufs[i]);
 170                 INIT_LIST_HEAD(&lgr->rmbs[i]);
 171         }
 172         smc_lgr_num += SMC_LGR_NUM_INCR;
 173         memcpy(&lgr->id, (u8 *)&smc_lgr_num, SMC_LGR_ID_SIZE);
 174         INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
 175         lgr->conns_all = RB_ROOT;
 176
 177         lnk = &lgr->lnk[SMC_SINGLE_LINK];
 178         /* initialize link */
 179         lnk->smcibdev = smcibdev;
 180         lnk->ibport = ibport;
 181         lnk->path_mtu = smcibdev->pattr[ibport - 1].active_mtu;
 182         if (!smcibdev->initialized)
 183                 smc_ib_setup_per_ibdev(smcibdev);
 184         get_random_bytes(rndvec, sizeof(rndvec));
 185         lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) + (rndvec[2] << 16);
 186         rc = smc_wr_alloc_link_mem(lnk);
 187         if (rc)
 188                 goto free_lgr;
 189         rc = smc_ib_create_protection_domain(lnk);
 190         if (rc)
 191                 goto free_link_mem;
 192         rc = smc_ib_create_queue_pair(lnk);
 193         if (rc)
 194                 goto dealloc_pd;
 195         rc = smc_wr_create_link(lnk);
 196         if (rc)
 197                 goto destroy_qp;
 198         init_completion(&lnk->llc_confirm);
 199         init_completion(&lnk->llc_confirm_resp);
 200
 201         smc->conn.lgr = lgr;
 202         rwlock_init(&lgr->conns_lock);
 203         spin_lock_bh(&smc_lgr_list.lock);
 204         list_add(&lgr->list, &smc_lgr_list.list);
 205         spin_unlock_bh(&smc_lgr_list.lock);
 206         return 0;
 207
 208 destroy_qp:
 209         smc_ib_destroy_queue_pair(lnk);
 210 dealloc_pd:
 211         smc_ib_dealloc_protection_domain(lnk);
 212 free_link_mem:
 213         smc_wr_free_link_mem(lnk);
 214 free_lgr:
 215         kfree(lgr);
 216 out:
 217         return rc;
 218 }
 219
 220 static void smc_buf_unuse(struct smc_connection *conn)
 221 {
 222         if (conn->sndbuf_desc) {
 223                 conn->sndbuf_desc->used = 0;
 224                 conn->sndbuf_size = 0;
 225         }
 226         if (conn->rmb_desc) {
 227                 conn->rmb_desc->reused = true;
 228                 conn->rmb_desc->used = 0;
 229                 conn->rmbe_size = 0;
 230         }
 231 }
 232
 233 /* remove a finished connection from its link group */
 234 void smc_conn_free(struct smc_connection *conn)
 235 {
 236         if (!conn->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); /* sock_put in close work */
 331                 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
 332                 __smc_lgr_unregister_conn(conn);
 333                 write_unlock_bh(&lgr->conns_lock);
 334                 if (!schedule_work(&conn->close_work))
 335                         sock_put(&smc->sk);
 336                 write_lock_bh(&lgr->conns_lock);
 337                 node = rb_first(&lgr->conns_all);
 338         }
 339         write_unlock_bh(&lgr->conns_lock);
 340         wake_up(&lgr->lnk[SMC_SINGLE_LINK].wr_reg_wait);
 341 }
 342
 343 /* Determine vlan of internal TCP socket.
 344  * @vlan_id: address to store the determined vlan id into
 345  */
 346 static int smc_vlan_by_tcpsk(struct socket *clcsock, unsigned short *vlan_id)
 347 {
 348         struct dst_entry *dst = sk_dst_get(clcsock->sk);
 349         int rc = 0;
 350
 351         *vlan_id = 0;
 352         if (!dst) {
 353                 rc = -ENOTCONN;
 354                 goto out;
 355         }
 356         if (!dst->dev) {
 357                 rc = -ENODEV;
 358                 goto out_rel;
 359         }
 360
 361         if (is_vlan_dev(dst->dev))
 362                 *vlan_id = vlan_dev_vlan_id(dst->dev);
 363
 364 out_rel:
 365         dst_release(dst);
 366 out:
 367         return rc;
 368 }
 369
 370 /* determine the link gid matching the vlan id of the link group */
 371 static int smc_link_determine_gid(struct smc_link_group *lgr)
 372 {
 373         struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
 374         struct ib_gid_attr gattr;
 375         union ib_gid gid;
 376         int i;
 377
 378         if (!lgr->vlan_id) {
 379                 lnk->gid = lnk->smcibdev->gid[lnk->ibport - 1];
 380                 return 0;
 381         }
 382
 383         for (i = 0; i < lnk->smcibdev->pattr[lnk->ibport - 1].gid_tbl_len;
 384              i++) {
 385                 if (ib_query_gid(lnk->smcibdev->ibdev, lnk->ibport, i, &gid,
 386                                  &gattr))
 387                         continue;
 388                 if (gattr.ndev) {
 389                         if (is_vlan_dev(gattr.ndev) &&
 390                             vlan_dev_vlan_id(gattr.ndev) == lgr->vlan_id) {
 391                                 lnk->gid = gid;
 392                                 dev_put(gattr.ndev);
 393                                 return 0;
 394                         }
 395                         dev_put(gattr.ndev);
 396                 }
 397         }
 398         return -ENODEV;
 399 }
 400
 401 /* create a new SMC connection (and a new link group if necessary) */
 402 int smc_conn_create(struct smc_sock *smc,
 403                     struct smc_ib_device *smcibdev, u8 ibport,
 404                     struct smc_clc_msg_local *lcl, int srv_first_contact)
 405 {
 406         struct smc_connection *conn = &smc->conn;
 407         struct smc_link_group *lgr;
 408         unsigned short vlan_id;
 409         enum smc_lgr_role role;
 410         int local_contact = SMC_FIRST_CONTACT;
 411         int rc = 0;
 412
 413         role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
 414         rc = smc_vlan_by_tcpsk(smc->clcsock, &vlan_id);
 415         if (rc)
 416                 return rc;
 417
 418         if ((role == SMC_CLNT) && srv_first_contact)
 419                 /* create new link group as well */
 420                 goto create;
 421
 422         /* determine if an existing link group can be reused */
 423         spin_lock_bh(&smc_lgr_list.lock);
 424         list_for_each_entry(lgr, &smc_lgr_list.list, list) {
 425                 write_lock_bh(&lgr->conns_lock);
 426                 if (!memcmp(lgr->peer_systemid, lcl->id_for_peer,
 427                             SMC_SYSTEMID_LEN) &&
 428                     !memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_gid, &lcl->gid,
 429                             SMC_GID_SIZE) &&
 430                     !memcmp(lgr->lnk[SMC_SINGLE_LINK].peer_mac, lcl->mac,
 431                             sizeof(lcl->mac)) &&
 432                     !lgr->sync_err &&
 433                     (lgr->role == role) &&
 434                     (lgr->vlan_id == vlan_id) &&
 435                     ((role == SMC_CLNT) ||
 436                      (lgr->conns_num < SMC_RMBS_PER_LGR_MAX))) {
 437                         /* link group found */
 438                         local_contact = SMC_REUSE_CONTACT;
 439                         conn->lgr = lgr;
 440                         smc_lgr_register_conn(conn); /* add smc conn to lgr */
 441                         write_unlock_bh(&lgr->conns_lock);
 442                         break;
 443                 }
 444                 write_unlock_bh(&lgr->conns_lock);
 445         }
 446         spin_unlock_bh(&smc_lgr_list.lock);
 447
 448         if (role == SMC_CLNT && !srv_first_contact &&
 449             (local_contact == SMC_FIRST_CONTACT)) {
 450                 /* Server reuses a link group, but Client wants to start
 451                  * a new one
 452                  * send out_of_sync decline, reason synchr. error
 453                  */
 454                 return -ENOLINK;
 455         }
 456
 457 create:
 458         if (local_contact == SMC_FIRST_CONTACT) {
 459                 rc = smc_lgr_create(smc, smcibdev, ibport,
 460                                     lcl->id_for_peer, vlan_id);
 461                 if (rc)
 462                         goto out;
 463                 smc_lgr_register_conn(conn); /* add smc conn to lgr */
 464                 rc = smc_link_determine_gid(conn->lgr);
 465         }
 466         conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
 467         conn->local_tx_ctrl.len = sizeof(struct smc_cdc_msg);
 468 #ifndef KERNEL_HAS_ATOMIC64
 469         spin_lock_init(&conn->acurs_lock);
 470 #endif
 471
 472 out:
 473         return rc ? rc : local_contact;
 474 }
 475
 476 /* try to reuse a sndbuf or rmb description slot for a certain
 477  * buffer size; if not available, return NULL
 478  */
 479 static inline
 480 struct smc_buf_desc *smc_buf_get_slot(struct smc_link_group *lgr,
 481                                       int compressed_bufsize,
 482                                       rwlock_t *lock,
 483                                       struct list_head *buf_list)
 484 {
 485         struct smc_buf_desc *buf_slot;
 486
 487         read_lock_bh(lock);
 488         list_for_each_entry(buf_slot, buf_list, list) {
 489                 if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
 490                         read_unlock_bh(lock);
 491                         return buf_slot;
 492                 }
 493         }
 494         read_unlock_bh(lock);
 495         return NULL;
 496 }
 497
 498 /* one of the conditions for announcing a receiver's current window size is
 499  * that it "results in a minimum increase in the window size of 10% of the
 500  * receive buffer space" [RFC7609]
 501  */
 502 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
 503 {
 504         return min_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
 505 }
 506
 507 static struct smc_buf_desc *smc_new_buf_create(struct smc_link_group *lgr,
 508                                                bool is_rmb, int bufsize)
 509 {
 510         struct smc_buf_desc *buf_desc;
 511         struct smc_link *lnk;
 512         int rc;
 513
 514         /* try to alloc a new buffer */
 515         buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
 516         if (!buf_desc)
 517                 return ERR_PTR(-ENOMEM);
 518
 519         buf_desc->cpu_addr =
 520                 (void *)__get_free_pages(GFP_KERNEL | __GFP_NOWARN |
 521                                          __GFP_NOMEMALLOC |
 522                                          __GFP_NORETRY | __GFP_ZERO,
 523                                          get_order(bufsize));
 524         if (!buf_desc->cpu_addr) {
 525                 kfree(buf_desc);
 526                 return ERR_PTR(-EAGAIN);
 527         }
 528         buf_desc->order = get_order(bufsize);
 529
 530         /* build the sg table from the pages */
 531         lnk = &lgr->lnk[SMC_SINGLE_LINK];
 532         rc = sg_alloc_table(&buf_desc->sgt[SMC_SINGLE_LINK], 1,
 533                             GFP_KERNEL);
 534         if (rc) {
 535                 smc_buf_free(buf_desc, lnk, is_rmb);
 536                 return ERR_PTR(rc);
 537         }
 538         sg_set_buf(buf_desc->sgt[SMC_SINGLE_LINK].sgl,
 539                    buf_desc->cpu_addr, bufsize);
 540
 541         /* map sg table to DMA address */
 542         rc = smc_ib_buf_map_sg(lnk->smcibdev, buf_desc,
 543                                is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
 544         /* SMC protocol depends on mapping to one DMA address only */
 545         if (rc != 1)  {
 546                 smc_buf_free(buf_desc, lnk, is_rmb);
 547                 return ERR_PTR(-EAGAIN);
 548         }
 549
 550         /* create a new memory region for the RMB */
 551         if (is_rmb) {
 552                 rc = smc_ib_get_memory_region(lnk->roce_pd,
 553                                               IB_ACCESS_REMOTE_WRITE |
 554                                               IB_ACCESS_LOCAL_WRITE,
 555                                               buf_desc);
 556                 if (rc) {
 557                         smc_buf_free(buf_desc, lnk, is_rmb);
 558                         return ERR_PTR(rc);
 559                 }
 560         }
 561
 562         return buf_desc;
 563 }
 564
 565 static int __smc_buf_create(struct smc_sock *smc, bool is_rmb)
 566 {
 567         struct smc_connection *conn = &smc->conn;
 568         struct smc_link_group *lgr = conn->lgr;
 569         struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
 570         struct list_head *buf_list;
 571         int bufsize, bufsize_short;
 572         int sk_buf_size;
 573         rwlock_t *lock;
 574
 575         if (is_rmb)
 576                 /* use socket recv buffer size (w/o overhead) as start value */
 577                 sk_buf_size = smc->sk.sk_rcvbuf / 2;
 578         else
 579                 /* use socket send buffer size (w/o overhead) as start value */
 580                 sk_buf_size = smc->sk.sk_sndbuf / 2;
 581
 582         for (bufsize_short = smc_compress_bufsize(sk_buf_size);
 583              bufsize_short >= 0; bufsize_short--) {
 584
 585                 if (is_rmb) {
 586                         lock = &lgr->rmbs_lock;
 587                         buf_list = &lgr->rmbs[bufsize_short];
 588                 } else {
 589                         lock = &lgr->sndbufs_lock;
 590                         buf_list = &lgr->sndbufs[bufsize_short];
 591                 }
 592                 bufsize = smc_uncompress_bufsize(bufsize_short);
 593                 if ((1 << get_order(bufsize)) > SG_MAX_SINGLE_ALLOC)
 594                         continue;
 595
 596                 /* check for reusable slot in the link group */
 597                 buf_desc = smc_buf_get_slot(lgr, bufsize_short, lock, buf_list);
 598                 if (buf_desc) {
 599                         memset(buf_desc->cpu_addr, 0, bufsize);
 600                         break; /* found reusable slot */
 601                 }
 602
 603                 buf_desc = smc_new_buf_create(lgr, is_rmb, bufsize);
 604                 if (PTR_ERR(buf_desc) == -ENOMEM)
 605                         break;
 606                 if (IS_ERR(buf_desc))
 607                         continue;
 608
 609                 buf_desc->used = 1;
 610                 write_lock_bh(lock);
 611                 list_add(&buf_desc->list, buf_list);
 612                 write_unlock_bh(lock);
 613                 break; /* found */
 614         }
 615
 616         if (IS_ERR(buf_desc))
 617                 return -ENOMEM;
 618
 619         if (is_rmb) {
 620                 conn->rmb_desc = buf_desc;
 621                 conn->rmbe_size = bufsize;
 622                 conn->rmbe_size_short = bufsize_short;
 623                 smc->sk.sk_rcvbuf = bufsize * 2;
 624                 atomic_set(&conn->bytes_to_rcv, 0);
 625                 conn->rmbe_update_limit = smc_rmb_wnd_update_limit(bufsize);
 626         } else {
 627                 conn->sndbuf_desc = buf_desc;
 628                 conn->sndbuf_size = bufsize;
 629                 smc->sk.sk_sndbuf = bufsize * 2;
 630                 atomic_set(&conn->sndbuf_space, bufsize);
 631         }
 632         return 0;
 633 }
 634
 635 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn)
 636 {
 637         struct smc_link_group *lgr = conn->lgr;
 638
 639         smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 640                                conn->sndbuf_desc, DMA_TO_DEVICE);
 641 }
 642
 643 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
 644 {
 645         struct smc_link_group *lgr = conn->lgr;
 646
 647         smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 648                                   conn->sndbuf_desc, DMA_TO_DEVICE);
 649 }
 650
 651 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
 652 {
 653         struct smc_link_group *lgr = conn->lgr;
 654
 655         smc_ib_sync_sg_for_cpu(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 656                                conn->rmb_desc, DMA_FROM_DEVICE);
 657 }
 658
 659 void smc_rmb_sync_sg_for_device(struct smc_connection *conn)
 660 {
 661         struct smc_link_group *lgr = conn->lgr;
 662
 663         smc_ib_sync_sg_for_device(lgr->lnk[SMC_SINGLE_LINK].smcibdev,
 664                                   conn->rmb_desc, DMA_FROM_DEVICE);
 665 }
 666
 667 /* create the send and receive buffer for an SMC socket;
 668  * receive buffers are called RMBs;
 669  * (even though the SMC protocol allows more than one RMB-element per RMB,
 670  * the Linux implementation uses just one RMB-element per RMB, i.e. uses an
 671  * extra RMB for every connection in a link group
 672  */
 673 int smc_buf_create(struct smc_sock *smc)
 674 {
 675         int rc;
 676
 677         /* create send buffer */
 678         rc = __smc_buf_create(smc, false);
 679         if (rc)
 680                 return rc;
 681         /* create rmb */
 682         rc = __smc_buf_create(smc, true);
 683         if (rc)
 684                 smc_buf_free(smc->conn.sndbuf_desc,
 685                              &smc->conn.lgr->lnk[SMC_SINGLE_LINK], false);
 686         return rc;
 687 }
 688
 689 static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
 690 {
 691         int i;
 692
 693         for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
 694                 if (!test_and_set_bit(i, lgr->rtokens_used_mask))
 695                         return i;
 696         }
 697         return -ENOSPC;
 698 }
 699
 700 /* save rkey and dma_addr received from peer during clc handshake */
 701 int smc_rmb_rtoken_handling(struct smc_connection *conn,
 702                             struct smc_clc_msg_accept_confirm *clc)
 703 {
 704         u64 dma_addr = be64_to_cpu(clc->rmb_dma_addr);
 705         struct smc_link_group *lgr = conn->lgr;
 706         u32 rkey = ntohl(clc->rmb_rkey);
 707         int i;
 708
 709         for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
 710                 if ((lgr->rtokens[i][SMC_SINGLE_LINK].rkey == rkey) &&
 711                     (lgr->rtokens[i][SMC_SINGLE_LINK].dma_addr == dma_addr) &&
 712                     test_bit(i, lgr->rtokens_used_mask)) {
 713                         conn->rtoken_idx = i;
 714                         return 0;
 715                 }
 716         }
 717         conn->rtoken_idx = smc_rmb_reserve_rtoken_idx(lgr);
 718         if (conn->rtoken_idx < 0)
 719                 return conn->rtoken_idx;
 720         lgr->rtokens[conn->rtoken_idx][SMC_SINGLE_LINK].rkey = rkey;
 721         lgr->rtokens[conn->rtoken_idx][SMC_SINGLE_LINK].dma_addr = dma_addr;
 722         return 0;
 723 }