net/smc/smc_core.c

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