lustre/fld/fld_request.c

   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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/fld/fld_request.c
  37  *
  38  * FLD (Fids Location Database)
  39  *
  40  * Author: Yury Umanets <umka@clusterfs.com>
  41  */
  42
  43 #define DEBUG_SUBSYSTEM S_FLD
  44
  45 #ifdef __KERNEL__
  46 # include <libcfs/libcfs.h>
  47 # include <linux/module.h>
  48 # include <asm/div64.h>
  49 #else /* __KERNEL__ */
  50 # include <liblustre.h>
  51 # include <libcfs/list.h>
  52 #endif
  53
  54 #include <obd.h>
  55 #include <obd_class.h>
  56 #include <obd_support.h>
  57 #include <lprocfs_status.h>
  58 #include <lustre_req_layout.h>
  59 #include <lustre_fld.h>
  60 #include <lustre_mdc.h>
  61 #include "fld_internal.h"
  62
  63 /* TODO: these 3 functions are copies of flow-control code from mdc_lib.c
  64  * It should be common thing. The same about mdc RPC lock */
  65 static int fld_req_avail(struct client_obd *cli, struct mdc_cache_waiter *mcw)
  66 {
  67         int rc;
  68         ENTRY;
  69         client_obd_list_lock(&cli->cl_loi_list_lock);
  70         rc = cfs_list_empty(&mcw->mcw_entry);
  71         client_obd_list_unlock(&cli->cl_loi_list_lock);
  72         RETURN(rc);
  73 };
  74
  75 static void fld_enter_request(struct client_obd *cli)
  76 {
  77         struct mdc_cache_waiter mcw;
  78         struct l_wait_info lwi = { 0 };
  79
  80         client_obd_list_lock(&cli->cl_loi_list_lock);
  81         if (cli->cl_r_in_flight >= cli->cl_max_rpcs_in_flight) {
  82                 cfs_list_add_tail(&mcw.mcw_entry, &cli->cl_cache_waiters);
  83                 init_waitqueue_head(&mcw.mcw_waitq);
  84                 client_obd_list_unlock(&cli->cl_loi_list_lock);
  85                 l_wait_event(mcw.mcw_waitq, fld_req_avail(cli, &mcw), &lwi);
  86         } else {
  87                 cli->cl_r_in_flight++;
  88                 client_obd_list_unlock(&cli->cl_loi_list_lock);
  89         }
  90 }
  91
  92 static void fld_exit_request(struct client_obd *cli)
  93 {
  94         cfs_list_t *l, *tmp;
  95         struct mdc_cache_waiter *mcw;
  96
  97         client_obd_list_lock(&cli->cl_loi_list_lock);
  98         cli->cl_r_in_flight--;
  99         cfs_list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
 100
 101                 if (cli->cl_r_in_flight >= cli->cl_max_rpcs_in_flight) {
 102                         /* No free request slots anymore */
 103                         break;
 104                 }
 105
 106                 mcw = cfs_list_entry(l, struct mdc_cache_waiter, mcw_entry);
 107                 cfs_list_del_init(&mcw->mcw_entry);
 108                 cli->cl_r_in_flight++;
 109                 wake_up(&mcw->mcw_waitq);
 110         }
 111         client_obd_list_unlock(&cli->cl_loi_list_lock);
 112 }
 113
 114 static int fld_rrb_hash(struct lu_client_fld *fld,
 115                         seqno_t seq)
 116 {
 117         LASSERT(fld->lcf_count > 0);
 118         return do_div(seq, fld->lcf_count);
 119 }
 120
 121 static struct lu_fld_target *
 122 fld_rrb_scan(struct lu_client_fld *fld, seqno_t seq)
 123 {
 124         struct lu_fld_target *target;
 125         int hash;
 126         ENTRY;
 127
 128         /* Because almost all of special sequence located in MDT0,
 129          * it should go to index 0 directly, instead of calculating
 130          * hash again, and also if other MDTs is not being connected,
 131          * the fld lookup requests(for seq on MDT0) should not be
 132          * blocked because of other MDTs */
 133         if (fid_seq_is_norm(seq))
 134                 hash = fld_rrb_hash(fld, seq);
 135         else
 136                 hash = 0;
 137
 138         cfs_list_for_each_entry(target, &fld->lcf_targets, ft_chain) {
 139                 if (target->ft_idx == hash)
 140                         RETURN(target);
 141         }
 142
 143         CERROR("%s: Can't find target by hash %d (seq "LPX64"). "
 144                "Targets (%d):\n", fld->lcf_name, hash, seq,
 145                fld->lcf_count);
 146
 147         cfs_list_for_each_entry(target, &fld->lcf_targets, ft_chain) {
 148                 const char *srv_name = target->ft_srv != NULL  ?
 149                         target->ft_srv->lsf_name : "<null>";
 150                 const char *exp_name = target->ft_exp != NULL ?
 151                         (char *)target->ft_exp->exp_obd->obd_uuid.uuid :
 152                         "<null>";
 153
 154                 CERROR("  exp: 0x%p (%s), srv: 0x%p (%s), idx: "LPU64"\n",
 155                        target->ft_exp, exp_name, target->ft_srv,
 156                        srv_name, target->ft_idx);
 157         }
 158
 159         /*
 160          * If target is not found, there is logical error anyway, so here is
 161          * LBUG() to catch this situation.
 162          */
 163         LBUG();
 164         RETURN(NULL);
 165 }
 166
 167 struct lu_fld_hash fld_hash[] = {
 168         {
 169                 .fh_name = "RRB",
 170                 .fh_hash_func = fld_rrb_hash,
 171                 .fh_scan_func = fld_rrb_scan
 172         },
 173         {
 174                 0,
 175         }
 176 };
 177
 178 static struct lu_fld_target *
 179 fld_client_get_target(struct lu_client_fld *fld, seqno_t seq)
 180 {
 181         struct lu_fld_target *target;
 182         ENTRY;
 183
 184         LASSERT(fld->lcf_hash != NULL);
 185
 186         spin_lock(&fld->lcf_lock);
 187         target = fld->lcf_hash->fh_scan_func(fld, seq);
 188         spin_unlock(&fld->lcf_lock);
 189
 190         if (target != NULL) {
 191                 CDEBUG(D_INFO, "%s: Found target (idx "LPU64
 192                        ") by seq "LPX64"\n", fld->lcf_name,
 193                        target->ft_idx, seq);
 194         }
 195
 196         RETURN(target);
 197 }
 198
 199 /*
 200  * Add export to FLD. This is usually done by CMM and LMV as they are main users
 201  * of FLD module.
 202  */
 203 int fld_client_add_target(struct lu_client_fld *fld,
 204                           struct lu_fld_target *tar)
 205 {
 206         const char *name;
 207         struct lu_fld_target *target, *tmp;
 208         ENTRY;
 209
 210         LASSERT(tar != NULL);
 211         name = fld_target_name(tar);
 212         LASSERT(name != NULL);
 213         LASSERT(tar->ft_srv != NULL || tar->ft_exp != NULL);
 214
 215         if (fld->lcf_flags != LUSTRE_FLD_INIT) {
 216                 CERROR("%s: Attempt to add target %s (idx "LPU64") "
 217                        "on fly - skip it\n", fld->lcf_name, name,
 218                        tar->ft_idx);
 219                 RETURN(0);
 220         } else {
 221                 CDEBUG(D_INFO, "%s: Adding target %s (idx "
 222                        LPU64")\n", fld->lcf_name, name, tar->ft_idx);
 223         }
 224
 225         OBD_ALLOC_PTR(target);
 226         if (target == NULL)
 227                 RETURN(-ENOMEM);
 228
 229         spin_lock(&fld->lcf_lock);
 230         cfs_list_for_each_entry(tmp, &fld->lcf_targets, ft_chain) {
 231                 if (tmp->ft_idx == tar->ft_idx) {
 232                         spin_unlock(&fld->lcf_lock);
 233                         OBD_FREE_PTR(target);
 234                         CERROR("Target %s exists in FLD and known as %s:#"LPU64"\n",
 235                                name, fld_target_name(tmp), tmp->ft_idx);
 236                         RETURN(-EEXIST);
 237                 }
 238         }
 239
 240         target->ft_exp = tar->ft_exp;
 241         if (target->ft_exp != NULL)
 242                 class_export_get(target->ft_exp);
 243         target->ft_srv = tar->ft_srv;
 244         target->ft_idx = tar->ft_idx;
 245
 246         cfs_list_add_tail(&target->ft_chain,
 247                           &fld->lcf_targets);
 248
 249         fld->lcf_count++;
 250         spin_unlock(&fld->lcf_lock);
 251
 252         RETURN(0);
 253 }
 254 EXPORT_SYMBOL(fld_client_add_target);
 255
 256 /* Remove export from FLD */
 257 int fld_client_del_target(struct lu_client_fld *fld, __u64 idx)
 258 {
 259         struct lu_fld_target *target, *tmp;
 260         ENTRY;
 261
 262         spin_lock(&fld->lcf_lock);
 263         cfs_list_for_each_entry_safe(target, tmp,
 264                                      &fld->lcf_targets, ft_chain) {
 265                 if (target->ft_idx == idx) {
 266                         fld->lcf_count--;
 267                         cfs_list_del(&target->ft_chain);
 268                         spin_unlock(&fld->lcf_lock);
 269
 270                         if (target->ft_exp != NULL)
 271                                 class_export_put(target->ft_exp);
 272
 273                         OBD_FREE_PTR(target);
 274                         RETURN(0);
 275                 }
 276         }
 277         spin_unlock(&fld->lcf_lock);
 278         RETURN(-ENOENT);
 279 }
 280 EXPORT_SYMBOL(fld_client_del_target);
 281
 282 #ifdef LPROCFS
 283 static int fld_client_proc_init(struct lu_client_fld *fld)
 284 {
 285         int rc;
 286         ENTRY;
 287
 288         fld->lcf_proc_dir = lprocfs_seq_register(fld->lcf_name,
 289                                                  fld_type_proc_dir,
 290                                                  NULL, NULL);
 291         if (IS_ERR(fld->lcf_proc_dir)) {
 292                 CERROR("%s: LProcFS failed in fld-init\n",
 293                        fld->lcf_name);
 294                 rc = PTR_ERR(fld->lcf_proc_dir);
 295                 RETURN(rc);
 296         }
 297
 298         rc = lprocfs_seq_add_vars(fld->lcf_proc_dir,
 299                                   fld_client_proc_list, fld);
 300         if (rc) {
 301                 CERROR("%s: Can't init FLD proc, rc %d\n",
 302                        fld->lcf_name, rc);
 303                 GOTO(out_cleanup, rc);
 304         }
 305
 306         RETURN(0);
 307
 308 out_cleanup:
 309         fld_client_proc_fini(fld);
 310         return rc;
 311 }
 312
 313 void fld_client_proc_fini(struct lu_client_fld *fld)
 314 {
 315         ENTRY;
 316         if (fld->lcf_proc_dir) {
 317                 if (!IS_ERR(fld->lcf_proc_dir))
 318                         lprocfs_remove(&fld->lcf_proc_dir);
 319                 fld->lcf_proc_dir = NULL;
 320         }
 321         EXIT;
 322 }
 323 #else
 324 static int fld_client_proc_init(struct lu_client_fld *fld)
 325 {
 326         return 0;
 327 }
 328
 329 void fld_client_proc_fini(struct lu_client_fld *fld)
 330 {
 331         return;
 332 }
 333 #endif
 334
 335 EXPORT_SYMBOL(fld_client_proc_fini);
 336
 337 static inline int hash_is_sane(int hash)
 338 {
 339         return (hash >= 0 && hash < ARRAY_SIZE(fld_hash));
 340 }
 341
 342 int fld_client_init(struct lu_client_fld *fld,
 343                     const char *prefix, int hash)
 344 {
 345         int cache_size, cache_threshold;
 346         int rc;
 347         ENTRY;
 348
 349         LASSERT(fld != NULL);
 350
 351         snprintf(fld->lcf_name, sizeof(fld->lcf_name),
 352                  "cli-%s", prefix);
 353
 354         if (!hash_is_sane(hash)) {
 355                 CERROR("%s: Wrong hash function %#x\n",
 356                        fld->lcf_name, hash);
 357                 RETURN(-EINVAL);
 358         }
 359
 360         fld->lcf_count = 0;
 361         spin_lock_init(&fld->lcf_lock);
 362         fld->lcf_hash = &fld_hash[hash];
 363         fld->lcf_flags = LUSTRE_FLD_INIT;
 364         CFS_INIT_LIST_HEAD(&fld->lcf_targets);
 365
 366         cache_size = FLD_CLIENT_CACHE_SIZE /
 367                 sizeof(struct fld_cache_entry);
 368
 369         cache_threshold = cache_size *
 370                 FLD_CLIENT_CACHE_THRESHOLD / 100;
 371
 372         fld->lcf_cache = fld_cache_init(fld->lcf_name,
 373                                         cache_size, cache_threshold);
 374         if (IS_ERR(fld->lcf_cache)) {
 375                 rc = PTR_ERR(fld->lcf_cache);
 376                 fld->lcf_cache = NULL;
 377                 GOTO(out, rc);
 378         }
 379
 380         rc = fld_client_proc_init(fld);
 381         if (rc)
 382                 GOTO(out, rc);
 383         EXIT;
 384 out:
 385         if (rc)
 386                 fld_client_fini(fld);
 387         else
 388                 CDEBUG(D_INFO, "%s: Using \"%s\" hash\n",
 389                        fld->lcf_name, fld->lcf_hash->fh_name);
 390         return rc;
 391 }
 392 EXPORT_SYMBOL(fld_client_init);
 393
 394 void fld_client_fini(struct lu_client_fld *fld)
 395 {
 396         struct lu_fld_target *target, *tmp;
 397         ENTRY;
 398
 399         spin_lock(&fld->lcf_lock);
 400         cfs_list_for_each_entry_safe(target, tmp,
 401                                      &fld->lcf_targets, ft_chain) {
 402                 fld->lcf_count--;
 403                 cfs_list_del(&target->ft_chain);
 404                 if (target->ft_exp != NULL)
 405                         class_export_put(target->ft_exp);
 406                 OBD_FREE_PTR(target);
 407         }
 408         spin_unlock(&fld->lcf_lock);
 409
 410         if (fld->lcf_cache != NULL) {
 411                 if (!IS_ERR(fld->lcf_cache))
 412                         fld_cache_fini(fld->lcf_cache);
 413                 fld->lcf_cache = NULL;
 414         }
 415
 416         EXIT;
 417 }
 418 EXPORT_SYMBOL(fld_client_fini);
 419
 420 int fld_client_rpc(struct obd_export *exp,
 421                    struct lu_seq_range *range, __u32 fld_op,
 422                    struct ptlrpc_request **reqp)
 423 {
 424         struct ptlrpc_request *req = NULL;
 425         struct lu_seq_range   *prange;
 426         __u32                 *op;
 427         int                    rc = 0;
 428         struct obd_import     *imp;
 429         ENTRY;
 430
 431         LASSERT(exp != NULL);
 432
 433 again:
 434         imp = class_exp2cliimp(exp);
 435         switch (fld_op) {
 436         case FLD_QUERY:
 437                 req = ptlrpc_request_alloc_pack(imp, &RQF_FLD_QUERY,
 438                                                 LUSTRE_MDS_VERSION, FLD_QUERY);
 439                 if (req == NULL)
 440                         RETURN(-ENOMEM);
 441
 442                 /* XXX: only needed when talking to old server(< 2.6), it should
 443                  * be removed when < 2.6 server is not supported */
 444                 op = req_capsule_client_get(&req->rq_pill, &RMF_FLD_OPC);
 445                 *op = FLD_LOOKUP;
 446
 447                 /* For MDS_MDS seq lookup, it will always use LWP connection,
 448                  * but LWP will be evicted after restart, so cause the error.
 449                  * so we will set no_delay for seq lookup request, once the
 450                  * request fails because of the eviction. always retry here */
 451                 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
 452                         req->rq_allow_replay = 1;
 453                         req->rq_no_delay = 1;
 454                 }
 455                 break;
 456         case FLD_READ:
 457                 req = ptlrpc_request_alloc_pack(imp, &RQF_FLD_READ,
 458                                                 LUSTRE_MDS_VERSION, FLD_READ);
 459                 if (req == NULL)
 460                         RETURN(-ENOMEM);
 461
 462                 req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA,
 463                                      RCL_SERVER, PAGE_CACHE_SIZE);
 464                 break;
 465         default:
 466                 rc = -EINVAL;
 467                 break;
 468         }
 469
 470         if (rc != 0)
 471                 RETURN(rc);
 472
 473         prange = req_capsule_client_get(&req->rq_pill, &RMF_FLD_MDFLD);
 474         *prange = *range;
 475         ptlrpc_request_set_replen(req);
 476         req->rq_request_portal = FLD_REQUEST_PORTAL;
 477         req->rq_reply_portal = MDC_REPLY_PORTAL;
 478         ptlrpc_at_set_req_timeout(req);
 479
 480         fld_enter_request(&exp->exp_obd->u.cli);
 481         rc = ptlrpc_queue_wait(req);
 482         fld_exit_request(&exp->exp_obd->u.cli);
 483         if (rc != 0) {
 484                 if (rc == -EWOULDBLOCK) {
 485                         /* For no_delay req(see above), EWOULDBLOCK means the
 486                          * connection is being evicted, but this seq lookup
 487                          * should not return error, since it would cause
 488                          * unecessary failure of the application, instead
 489                          * it should retry here */
 490                         ptlrpc_req_finished(req);
 491                         goto again;
 492                 }
 493                 GOTO(out_req, rc);
 494         }
 495
 496         if (fld_op == FLD_QUERY) {
 497                 prange = req_capsule_server_get(&req->rq_pill,
 498                                                 &RMF_FLD_MDFLD);
 499                 if (prange == NULL)
 500                         GOTO(out_req, rc = -EFAULT);
 501                 *range = *prange;
 502         }
 503
 504         EXIT;
 505 out_req:
 506         if (rc != 0 || reqp == NULL) {
 507                 ptlrpc_req_finished(req);
 508                 req = NULL;
 509         }
 510
 511         if (reqp != NULL)
 512                 *reqp = req;
 513
 514         return rc;
 515 }
 516
 517 int fld_client_lookup(struct lu_client_fld *fld, seqno_t seq, mdsno_t *mds,
 518                       __u32 flags, const struct lu_env *env)
 519 {
 520         struct lu_seq_range res = { 0 };
 521         struct lu_fld_target *target;
 522         int rc;
 523         ENTRY;
 524
 525         fld->lcf_flags |= LUSTRE_FLD_RUN;
 526
 527         rc = fld_cache_lookup(fld->lcf_cache, seq, &res);
 528         if (rc == 0) {
 529                 *mds = res.lsr_index;
 530                 RETURN(0);
 531         }
 532
 533         /* Can not find it in the cache */
 534         target = fld_client_get_target(fld, seq);
 535         LASSERT(target != NULL);
 536
 537         CDEBUG(D_INFO, "%s: Lookup fld entry (seq: "LPX64") on "
 538                "target %s (idx "LPU64")\n", fld->lcf_name, seq,
 539                fld_target_name(target), target->ft_idx);
 540
 541         res.lsr_start = seq;
 542         fld_range_set_type(&res, flags);
 543
 544 #if defined(__KERNEL__) && defined(HAVE_SERVER_SUPPORT)
 545         if (target->ft_srv != NULL) {
 546                 LASSERT(env != NULL);
 547                 rc = fld_server_lookup(env, target->ft_srv, seq, &res);
 548         } else
 549 #endif
 550         {
 551                 rc = fld_client_rpc(target->ft_exp, &res, FLD_QUERY, NULL);
 552         }
 553
 554         if (rc == 0) {
 555                 *mds = res.lsr_index;
 556                 fld_cache_insert(fld->lcf_cache, &res);
 557         }
 558
 559         RETURN(rc);
 560 }
 561 EXPORT_SYMBOL(fld_client_lookup);
 562
 563 void fld_client_flush(struct lu_client_fld *fld)
 564 {
 565         fld_cache_flush(fld->lcf_cache);
 566 }
 567 EXPORT_SYMBOL(fld_client_flush);
 568
 569 #ifdef __KERNEL__
 570
 571 struct proc_dir_entry *fld_type_proc_dir;
 572
 573 static int __init fld_mod_init(void)
 574 {
 575         fld_type_proc_dir = lprocfs_seq_register(LUSTRE_FLD_NAME,
 576                                                  proc_lustre_root,
 577                                                  NULL, NULL);
 578         if (IS_ERR(fld_type_proc_dir))
 579                 return PTR_ERR(fld_type_proc_dir);
 580
 581 #ifdef HAVE_SERVER_SUPPORT
 582         fld_server_mod_init();
 583 #endif
 584
 585         return 0;
 586 }
 587
 588 static void __exit fld_mod_exit(void)
 589 {
 590 #ifdef HAVE_SERVER_SUPPORT
 591         fld_server_mod_exit();
 592 #endif
 593
 594         if (fld_type_proc_dir != NULL && !IS_ERR(fld_type_proc_dir)) {
 595                 lprocfs_remove(&fld_type_proc_dir);
 596                 fld_type_proc_dir = NULL;
 597         }
 598 }
 599
 600 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
 601 MODULE_DESCRIPTION("Lustre FLD");
 602 MODULE_LICENSE("GPL");
 603
 604 cfs_module(mdd, LUSTRE_VERSION_STRING, fld_mod_init, fld_mod_exit);
 605 #endif /* __KERNEL__ */