linux/drivers/staging/lustre/lnet/lnet/peer.c
<<
>>
Prefs
   1/*
   2 * GPL HEADER START
   3 *
   4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5 *
   6 * This program is free software; you can redistribute it and/or modify
   7 * it under the terms of the GNU General Public License version 2 only,
   8 * as published by the Free Software Foundation.
   9 *
  10 * This program is distributed in the hope that it will be useful, but
  11 * WITHOUT ANY WARRANTY; without even the implied warranty of
  12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13 * General Public License version 2 for more details (a copy is included
  14 * in the LICENSE file that accompanied this code).
  15 *
  16 * You should have received a copy of the GNU General Public License
  17 * version 2 along with this program; If not, see
  18 * http://www.gnu.org/licenses/gpl-2.0.html
  19 *
  20 * GPL HEADER END
  21 */
  22/*
  23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24 * Use is subject to license terms.
  25 *
  26 * Copyright (c) 2012, Intel Corporation.
  27 */
  28/*
  29 * This file is part of Lustre, http://www.lustre.org/
  30 * Lustre is a trademark of Sun Microsystems, Inc.
  31 *
  32 * lnet/lnet/peer.c
  33 */
  34
  35#define DEBUG_SUBSYSTEM S_LNET
  36
  37#include "../../include/linux/lnet/lib-lnet.h"
  38#include "../../include/linux/lnet/lib-dlc.h"
  39
  40int
  41lnet_peer_tables_create(void)
  42{
  43        struct lnet_peer_table *ptable;
  44        struct list_head *hash;
  45        int i;
  46        int j;
  47
  48        the_lnet.ln_peer_tables = cfs_percpt_alloc(lnet_cpt_table(),
  49                                                   sizeof(*ptable));
  50        if (!the_lnet.ln_peer_tables) {
  51                CERROR("Failed to allocate cpu-partition peer tables\n");
  52                return -ENOMEM;
  53        }
  54
  55        cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
  56                INIT_LIST_HEAD(&ptable->pt_deathrow);
  57
  58                LIBCFS_CPT_ALLOC(hash, lnet_cpt_table(), i,
  59                                 LNET_PEER_HASH_SIZE * sizeof(*hash));
  60                if (!hash) {
  61                        CERROR("Failed to create peer hash table\n");
  62                        lnet_peer_tables_destroy();
  63                        return -ENOMEM;
  64                }
  65
  66                for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
  67                        INIT_LIST_HEAD(&hash[j]);
  68                ptable->pt_hash = hash; /* sign of initialization */
  69        }
  70
  71        return 0;
  72}
  73
  74void
  75lnet_peer_tables_destroy(void)
  76{
  77        struct lnet_peer_table *ptable;
  78        struct list_head *hash;
  79        int i;
  80        int j;
  81
  82        if (!the_lnet.ln_peer_tables)
  83                return;
  84
  85        cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
  86                hash = ptable->pt_hash;
  87                if (!hash) /* not initialized */
  88                        break;
  89
  90                LASSERT(list_empty(&ptable->pt_deathrow));
  91
  92                ptable->pt_hash = NULL;
  93                for (j = 0; j < LNET_PEER_HASH_SIZE; j++)
  94                        LASSERT(list_empty(&hash[j]));
  95
  96                LIBCFS_FREE(hash, LNET_PEER_HASH_SIZE * sizeof(*hash));
  97        }
  98
  99        cfs_percpt_free(the_lnet.ln_peer_tables);
 100        the_lnet.ln_peer_tables = NULL;
 101}
 102
 103static void
 104lnet_peer_table_cleanup_locked(lnet_ni_t *ni, struct lnet_peer_table *ptable)
 105{
 106        int i;
 107        lnet_peer_t *lp;
 108        lnet_peer_t *tmp;
 109
 110        for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
 111                list_for_each_entry_safe(lp, tmp, &ptable->pt_hash[i],
 112                                         lp_hashlist) {
 113                        if (ni && ni != lp->lp_ni)
 114                                continue;
 115                        list_del_init(&lp->lp_hashlist);
 116                        /* Lose hash table's ref */
 117                        ptable->pt_zombies++;
 118                        lnet_peer_decref_locked(lp);
 119                }
 120        }
 121}
 122
 123static void
 124lnet_peer_table_deathrow_wait_locked(struct lnet_peer_table *ptable,
 125                                     int cpt_locked)
 126{
 127        int i;
 128
 129        for (i = 3; ptable->pt_zombies; i++) {
 130                lnet_net_unlock(cpt_locked);
 131
 132                if (is_power_of_2(i)) {
 133                        CDEBUG(D_WARNING,
 134                               "Waiting for %d zombies on peer table\n",
 135                               ptable->pt_zombies);
 136                }
 137                set_current_state(TASK_UNINTERRUPTIBLE);
 138                schedule_timeout(cfs_time_seconds(1) >> 1);
 139                lnet_net_lock(cpt_locked);
 140        }
 141}
 142
 143static void
 144lnet_peer_table_del_rtrs_locked(lnet_ni_t *ni, struct lnet_peer_table *ptable,
 145                                int cpt_locked)
 146{
 147        lnet_peer_t *lp;
 148        lnet_peer_t *tmp;
 149        lnet_nid_t lp_nid;
 150        int i;
 151
 152        for (i = 0; i < LNET_PEER_HASH_SIZE; i++) {
 153                list_for_each_entry_safe(lp, tmp, &ptable->pt_hash[i],
 154                                         lp_hashlist) {
 155                        if (ni != lp->lp_ni)
 156                                continue;
 157
 158                        if (!lp->lp_rtr_refcount)
 159                                continue;
 160
 161                        lp_nid = lp->lp_nid;
 162
 163                        lnet_net_unlock(cpt_locked);
 164                        lnet_del_route(LNET_NIDNET(LNET_NID_ANY), lp_nid);
 165                        lnet_net_lock(cpt_locked);
 166                }
 167        }
 168}
 169
 170void
 171lnet_peer_tables_cleanup(lnet_ni_t *ni)
 172{
 173        struct lnet_peer_table *ptable;
 174        struct list_head deathrow;
 175        lnet_peer_t *lp;
 176        lnet_peer_t *temp;
 177        int i;
 178
 179        INIT_LIST_HEAD(&deathrow);
 180
 181        LASSERT(the_lnet.ln_shutdown || ni);
 182        /*
 183         * If just deleting the peers for a NI, get rid of any routes these
 184         * peers are gateways for.
 185         */
 186        cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 187                lnet_net_lock(i);
 188                lnet_peer_table_del_rtrs_locked(ni, ptable, i);
 189                lnet_net_unlock(i);
 190        }
 191
 192        /*
 193         * Start the process of moving the applicable peers to
 194         * deathrow.
 195         */
 196        cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 197                lnet_net_lock(i);
 198                lnet_peer_table_cleanup_locked(ni, ptable);
 199                lnet_net_unlock(i);
 200        }
 201
 202        /* Cleanup all entries on deathrow. */
 203        cfs_percpt_for_each(ptable, i, the_lnet.ln_peer_tables) {
 204                lnet_net_lock(i);
 205                lnet_peer_table_deathrow_wait_locked(ptable, i);
 206                list_splice_init(&ptable->pt_deathrow, &deathrow);
 207                lnet_net_unlock(i);
 208        }
 209
 210        list_for_each_entry_safe(lp, temp, &deathrow, lp_hashlist) {
 211                list_del(&lp->lp_hashlist);
 212                LIBCFS_FREE(lp, sizeof(*lp));
 213        }
 214}
 215
 216void
 217lnet_destroy_peer_locked(lnet_peer_t *lp)
 218{
 219        struct lnet_peer_table *ptable;
 220
 221        LASSERT(!lp->lp_refcount);
 222        LASSERT(!lp->lp_rtr_refcount);
 223        LASSERT(list_empty(&lp->lp_txq));
 224        LASSERT(list_empty(&lp->lp_hashlist));
 225        LASSERT(!lp->lp_txqnob);
 226
 227        ptable = the_lnet.ln_peer_tables[lp->lp_cpt];
 228        LASSERT(ptable->pt_number > 0);
 229        ptable->pt_number--;
 230
 231        lnet_ni_decref_locked(lp->lp_ni, lp->lp_cpt);
 232        lp->lp_ni = NULL;
 233
 234        list_add(&lp->lp_hashlist, &ptable->pt_deathrow);
 235        LASSERT(ptable->pt_zombies > 0);
 236        ptable->pt_zombies--;
 237}
 238
 239lnet_peer_t *
 240lnet_find_peer_locked(struct lnet_peer_table *ptable, lnet_nid_t nid)
 241{
 242        struct list_head *peers;
 243        lnet_peer_t *lp;
 244
 245        LASSERT(!the_lnet.ln_shutdown);
 246
 247        peers = &ptable->pt_hash[lnet_nid2peerhash(nid)];
 248        list_for_each_entry(lp, peers, lp_hashlist) {
 249                if (lp->lp_nid == nid) {
 250                        lnet_peer_addref_locked(lp);
 251                        return lp;
 252                }
 253        }
 254
 255        return NULL;
 256}
 257
 258int
 259lnet_nid2peer_locked(lnet_peer_t **lpp, lnet_nid_t nid, int cpt)
 260{
 261        struct lnet_peer_table *ptable;
 262        lnet_peer_t *lp = NULL;
 263        lnet_peer_t *lp2;
 264        int cpt2;
 265        int rc = 0;
 266
 267        *lpp = NULL;
 268        if (the_lnet.ln_shutdown) /* it's shutting down */
 269                return -ESHUTDOWN;
 270
 271        /* cpt can be LNET_LOCK_EX if it's called from router functions */
 272        cpt2 = cpt != LNET_LOCK_EX ? cpt : lnet_cpt_of_nid_locked(nid);
 273
 274        ptable = the_lnet.ln_peer_tables[cpt2];
 275        lp = lnet_find_peer_locked(ptable, nid);
 276        if (lp) {
 277                *lpp = lp;
 278                return 0;
 279        }
 280
 281        if (!list_empty(&ptable->pt_deathrow)) {
 282                lp = list_entry(ptable->pt_deathrow.next,
 283                                lnet_peer_t, lp_hashlist);
 284                list_del(&lp->lp_hashlist);
 285        }
 286
 287        /*
 288         * take extra refcount in case another thread has shutdown LNet
 289         * and destroyed locks and peer-table before I finish the allocation
 290         */
 291        ptable->pt_number++;
 292        lnet_net_unlock(cpt);
 293
 294        if (lp)
 295                memset(lp, 0, sizeof(*lp));
 296        else
 297                LIBCFS_CPT_ALLOC(lp, lnet_cpt_table(), cpt2, sizeof(*lp));
 298
 299        if (!lp) {
 300                rc = -ENOMEM;
 301                lnet_net_lock(cpt);
 302                goto out;
 303        }
 304
 305        INIT_LIST_HEAD(&lp->lp_txq);
 306        INIT_LIST_HEAD(&lp->lp_rtrq);
 307        INIT_LIST_HEAD(&lp->lp_routes);
 308
 309        lp->lp_notify = 0;
 310        lp->lp_notifylnd = 0;
 311        lp->lp_notifying = 0;
 312        lp->lp_alive_count = 0;
 313        lp->lp_timestamp = 0;
 314        lp->lp_alive = !lnet_peers_start_down(); /* 1 bit!! */
 315        lp->lp_last_alive = cfs_time_current(); /* assumes alive */
 316        lp->lp_last_query = 0; /* haven't asked NI yet */
 317        lp->lp_ping_timestamp = 0;
 318        lp->lp_ping_feats = LNET_PING_FEAT_INVAL;
 319        lp->lp_nid = nid;
 320        lp->lp_cpt = cpt2;
 321        lp->lp_refcount = 2;    /* 1 for caller; 1 for hash */
 322        lp->lp_rtr_refcount = 0;
 323
 324        lnet_net_lock(cpt);
 325
 326        if (the_lnet.ln_shutdown) {
 327                rc = -ESHUTDOWN;
 328                goto out;
 329        }
 330
 331        lp2 = lnet_find_peer_locked(ptable, nid);
 332        if (lp2) {
 333                *lpp = lp2;
 334                goto out;
 335        }
 336
 337        lp->lp_ni = lnet_net2ni_locked(LNET_NIDNET(nid), cpt2);
 338        if (!lp->lp_ni) {
 339                rc = -EHOSTUNREACH;
 340                goto out;
 341        }
 342
 343        lp->lp_txcredits = lp->lp_ni->ni_peertxcredits;
 344        lp->lp_mintxcredits = lp->lp_ni->ni_peertxcredits;
 345        lp->lp_rtrcredits = lnet_peer_buffer_credits(lp->lp_ni);
 346        lp->lp_minrtrcredits = lnet_peer_buffer_credits(lp->lp_ni);
 347
 348        list_add_tail(&lp->lp_hashlist,
 349                      &ptable->pt_hash[lnet_nid2peerhash(nid)]);
 350        ptable->pt_version++;
 351        *lpp = lp;
 352
 353        return 0;
 354out:
 355        if (lp)
 356                list_add(&lp->lp_hashlist, &ptable->pt_deathrow);
 357        ptable->pt_number--;
 358        return rc;
 359}
 360
 361void
 362lnet_debug_peer(lnet_nid_t nid)
 363{
 364        char *aliveness = "NA";
 365        lnet_peer_t *lp;
 366        int rc;
 367        int cpt;
 368
 369        cpt = lnet_cpt_of_nid(nid);
 370        lnet_net_lock(cpt);
 371
 372        rc = lnet_nid2peer_locked(&lp, nid, cpt);
 373        if (rc) {
 374                lnet_net_unlock(cpt);
 375                CDEBUG(D_WARNING, "No peer %s\n", libcfs_nid2str(nid));
 376                return;
 377        }
 378
 379        if (lnet_isrouter(lp) || lnet_peer_aliveness_enabled(lp))
 380                aliveness = lp->lp_alive ? "up" : "down";
 381
 382        CDEBUG(D_WARNING, "%-24s %4d %5s %5d %5d %5d %5d %5d %ld\n",
 383               libcfs_nid2str(lp->lp_nid), lp->lp_refcount,
 384               aliveness, lp->lp_ni->ni_peertxcredits,
 385               lp->lp_rtrcredits, lp->lp_minrtrcredits,
 386               lp->lp_txcredits, lp->lp_mintxcredits, lp->lp_txqnob);
 387
 388        lnet_peer_decref_locked(lp);
 389
 390        lnet_net_unlock(cpt);
 391}
 392
 393int
 394lnet_get_peer_info(__u32 peer_index, __u64 *nid,
 395                   char aliveness[LNET_MAX_STR_LEN],
 396                   __u32 *cpt_iter, __u32 *refcount,
 397                   __u32 *ni_peer_tx_credits, __u32 *peer_tx_credits,
 398                   __u32 *peer_rtr_credits, __u32 *peer_min_rtr_credits,
 399                   __u32 *peer_tx_qnob)
 400{
 401        struct lnet_peer_table *peer_table;
 402        lnet_peer_t *lp;
 403        bool found = false;
 404        int lncpt, j;
 405
 406        /* get the number of CPTs */
 407        lncpt = cfs_percpt_number(the_lnet.ln_peer_tables);
 408
 409        /*
 410         * if the cpt number to be examined is >= the number of cpts in
 411         * the system then indicate that there are no more cpts to examin
 412         */
 413        if (*cpt_iter >= lncpt)
 414                return -ENOENT;
 415
 416        /* get the current table */
 417        peer_table = the_lnet.ln_peer_tables[*cpt_iter];
 418        /* if the ptable is NULL then there are no more cpts to examine */
 419        if (!peer_table)
 420                return -ENOENT;
 421
 422        lnet_net_lock(*cpt_iter);
 423
 424        for (j = 0; j < LNET_PEER_HASH_SIZE && !found; j++) {
 425                struct list_head *peers = &peer_table->pt_hash[j];
 426
 427                list_for_each_entry(lp, peers, lp_hashlist) {
 428                        if (peer_index-- > 0)
 429                                continue;
 430
 431                        snprintf(aliveness, LNET_MAX_STR_LEN, "NA");
 432                        if (lnet_isrouter(lp) ||
 433                            lnet_peer_aliveness_enabled(lp))
 434                                snprintf(aliveness, LNET_MAX_STR_LEN,
 435                                         lp->lp_alive ? "up" : "down");
 436
 437                        *nid = lp->lp_nid;
 438                        *refcount = lp->lp_refcount;
 439                        *ni_peer_tx_credits = lp->lp_ni->ni_peertxcredits;
 440                        *peer_tx_credits = lp->lp_txcredits;
 441                        *peer_rtr_credits = lp->lp_rtrcredits;
 442                        *peer_min_rtr_credits = lp->lp_mintxcredits;
 443                        *peer_tx_qnob = lp->lp_txqnob;
 444
 445                        found = true;
 446                }
 447        }
 448        lnet_net_unlock(*cpt_iter);
 449
 450        *cpt_iter = lncpt;
 451
 452        return found ? 0 : -ENOENT;
 453}
 454