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, 2015, 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 #include <libcfs/libcfs.h>
  46 #include <linux/module.h>
  47 #include <linux/math64.h>
  48
  49 #include <obd.h>
  50 #include <obd_class.h>
  51 #include <obd_support.h>
  52 #include <lprocfs_status.h>
  53 #include <lustre_req_layout.h>
  54 #include <lustre_fld.h>
  55 #include <lustre_mdc.h>
  56 #include "fld_internal.h"
  57
  58 static int fld_rrb_hash(struct lu_client_fld *fld, u64 seq)
  59 {
  60         LASSERT(fld->lcf_count > 0);
  61         return do_div(seq, fld->lcf_count);
  62 }
  63
  64 static struct lu_fld_target *
  65 fld_rrb_scan(struct lu_client_fld *fld, u64 seq)
  66 {
  67         struct lu_fld_target *target;
  68         int hash;
  69         ENTRY;
  70
  71         /* Because almost all of special sequence located in MDT0,
  72          * it should go to index 0 directly, instead of calculating
  73          * hash again, and also if other MDTs is not being connected,
  74          * the fld lookup requests(for seq on MDT0) should not be
  75          * blocked because of other MDTs */
  76         if (fid_seq_is_norm(seq))
  77                 hash = fld_rrb_hash(fld, seq);
  78         else
  79                 hash = 0;
  80
  81 again:
  82         list_for_each_entry(target, &fld->lcf_targets, ft_chain) {
  83                 if (target->ft_idx == hash)
  84                         RETURN(target);
  85         }
  86
  87         if (hash != 0) {
  88                 /* It is possible the remote target(MDT) are not connected to
  89                  * with client yet, so we will refer this to MDT0, which should
  90                  * be connected during mount */
  91                 hash = 0;
  92                 goto again;
  93         }
  94
  95         CERROR("%s: Can't find target by hash %d (seq "LPX64"). "
  96                "Targets (%d):\n", fld->lcf_name, hash, seq,
  97                fld->lcf_count);
  98
  99         list_for_each_entry(target, &fld->lcf_targets, ft_chain) {
 100                 const char *srv_name = target->ft_srv != NULL  ?
 101                         target->ft_srv->lsf_name : "<null>";
 102                 const char *exp_name = target->ft_exp != NULL ?
 103                         (char *)target->ft_exp->exp_obd->obd_uuid.uuid :
 104                         "<null>";
 105
 106                 CERROR("  exp: 0x%p (%s), srv: 0x%p (%s), idx: "LPU64"\n",
 107                        target->ft_exp, exp_name, target->ft_srv,
 108                        srv_name, target->ft_idx);
 109         }
 110
 111         /*
 112          * If target is not found, there is logical error anyway, so here is
 113          * LBUG() to catch this situation.
 114          */
 115         LBUG();
 116         RETURN(NULL);
 117 }
 118
 119 struct lu_fld_hash fld_hash[] = {
 120         {
 121                 .fh_name = "RRB",
 122                 .fh_hash_func = fld_rrb_hash,
 123                 .fh_scan_func = fld_rrb_scan
 124         },
 125         {
 126                 NULL,
 127         }
 128 };
 129
 130 static struct lu_fld_target *
 131 fld_client_get_target(struct lu_client_fld *fld, u64 seq)
 132 {
 133         struct lu_fld_target *target;
 134         ENTRY;
 135
 136         LASSERT(fld->lcf_hash != NULL);
 137
 138         spin_lock(&fld->lcf_lock);
 139         target = fld->lcf_hash->fh_scan_func(fld, seq);
 140         spin_unlock(&fld->lcf_lock);
 141
 142         if (target != NULL) {
 143                 CDEBUG(D_INFO, "%s: Found target (idx "LPU64
 144                        ") by seq "LPX64"\n", fld->lcf_name,
 145                        target->ft_idx, seq);
 146         }
 147
 148         RETURN(target);
 149 }
 150
 151 /*
 152  * Add export to FLD. This is usually done by CMM and LMV as they are main users
 153  * of FLD module.
 154  */
 155 int fld_client_add_target(struct lu_client_fld *fld,
 156                           struct lu_fld_target *tar)
 157 {
 158         const char *name;
 159         struct lu_fld_target *target, *tmp;
 160         ENTRY;
 161
 162         LASSERT(tar != NULL);
 163         name = fld_target_name(tar);
 164         LASSERT(name != NULL);
 165         LASSERT(tar->ft_srv != NULL || tar->ft_exp != NULL);
 166
 167         CDEBUG(D_INFO, "%s: Adding target %s (idx "LPU64")\n", fld->lcf_name,
 168                name, tar->ft_idx);
 169
 170         OBD_ALLOC_PTR(target);
 171         if (target == NULL)
 172                 RETURN(-ENOMEM);
 173
 174         spin_lock(&fld->lcf_lock);
 175         list_for_each_entry(tmp, &fld->lcf_targets, ft_chain) {
 176                 if (tmp->ft_idx == tar->ft_idx) {
 177                         spin_unlock(&fld->lcf_lock);
 178                         OBD_FREE_PTR(target);
 179                         CERROR("Target %s exists in FLD and known as %s:#"LPU64"\n",
 180                                name, fld_target_name(tmp), tmp->ft_idx);
 181                         RETURN(-EEXIST);
 182                 }
 183         }
 184
 185         target->ft_exp = tar->ft_exp;
 186         if (target->ft_exp != NULL)
 187                 class_export_get(target->ft_exp);
 188         target->ft_srv = tar->ft_srv;
 189         target->ft_idx = tar->ft_idx;
 190
 191         list_add_tail(&target->ft_chain, &fld->lcf_targets);
 192
 193         fld->lcf_count++;
 194         spin_unlock(&fld->lcf_lock);
 195
 196         RETURN(0);
 197 }
 198 EXPORT_SYMBOL(fld_client_add_target);
 199
 200 /* Remove export from FLD */
 201 int fld_client_del_target(struct lu_client_fld *fld, __u64 idx)
 202 {
 203         struct lu_fld_target *target, *tmp;
 204         ENTRY;
 205
 206         spin_lock(&fld->lcf_lock);
 207         list_for_each_entry_safe(target, tmp, &fld->lcf_targets, ft_chain) {
 208                 if (target->ft_idx == idx) {
 209                         fld->lcf_count--;
 210                         list_del(&target->ft_chain);
 211                         spin_unlock(&fld->lcf_lock);
 212
 213                         if (target->ft_exp != NULL)
 214                                 class_export_put(target->ft_exp);
 215
 216                         OBD_FREE_PTR(target);
 217                         RETURN(0);
 218                 }
 219         }
 220         spin_unlock(&fld->lcf_lock);
 221         RETURN(-ENOENT);
 222 }
 223
 224 #ifdef CONFIG_PROC_FS
 225 static int fld_client_proc_init(struct lu_client_fld *fld)
 226 {
 227         int rc;
 228         ENTRY;
 229
 230         fld->lcf_proc_dir = lprocfs_register(fld->lcf_name, fld_type_proc_dir,
 231                                              NULL, NULL);
 232         if (IS_ERR(fld->lcf_proc_dir)) {
 233                 CERROR("%s: LProcFS failed in fld-init\n",
 234                        fld->lcf_name);
 235                 rc = PTR_ERR(fld->lcf_proc_dir);
 236                 RETURN(rc);
 237         }
 238
 239         rc = lprocfs_add_vars(fld->lcf_proc_dir, fld_client_proc_list, fld);
 240         if (rc) {
 241                 CERROR("%s: Can't init FLD proc, rc %d\n",
 242                        fld->lcf_name, rc);
 243                 GOTO(out_cleanup, rc);
 244         }
 245
 246         RETURN(0);
 247
 248 out_cleanup:
 249         fld_client_proc_fini(fld);
 250         return rc;
 251 }
 252
 253 void fld_client_proc_fini(struct lu_client_fld *fld)
 254 {
 255         ENTRY;
 256         if (fld->lcf_proc_dir) {
 257                 if (!IS_ERR(fld->lcf_proc_dir))
 258                         lprocfs_remove(&fld->lcf_proc_dir);
 259                 fld->lcf_proc_dir = NULL;
 260         }
 261         EXIT;
 262 }
 263 #else /* !CONFIG_PROC_FS */
 264 static int fld_client_proc_init(struct lu_client_fld *fld)
 265 {
 266         return 0;
 267 }
 268
 269 void fld_client_proc_fini(struct lu_client_fld *fld)
 270 {
 271         return;
 272 }
 273 #endif /* CONFIG_PROC_FS */
 274
 275 EXPORT_SYMBOL(fld_client_proc_fini);
 276
 277 static inline int hash_is_sane(int hash)
 278 {
 279         return (hash >= 0 && hash < ARRAY_SIZE(fld_hash));
 280 }
 281
 282 int fld_client_init(struct lu_client_fld *fld,
 283                     const char *prefix, int hash)
 284 {
 285         int cache_size, cache_threshold;
 286         int rc;
 287         ENTRY;
 288
 289         LASSERT(fld != NULL);
 290
 291         snprintf(fld->lcf_name, sizeof(fld->lcf_name),
 292                  "cli-%s", prefix);
 293
 294         if (!hash_is_sane(hash)) {
 295                 CERROR("%s: Wrong hash function %#x\n",
 296                        fld->lcf_name, hash);
 297                 RETURN(-EINVAL);
 298         }
 299
 300         fld->lcf_count = 0;
 301         spin_lock_init(&fld->lcf_lock);
 302         fld->lcf_hash = &fld_hash[hash];
 303         INIT_LIST_HEAD(&fld->lcf_targets);
 304
 305         cache_size = FLD_CLIENT_CACHE_SIZE /
 306                 sizeof(struct fld_cache_entry);
 307
 308         cache_threshold = cache_size *
 309                 FLD_CLIENT_CACHE_THRESHOLD / 100;
 310
 311         fld->lcf_cache = fld_cache_init(fld->lcf_name,
 312                                         cache_size, cache_threshold);
 313         if (IS_ERR(fld->lcf_cache)) {
 314                 rc = PTR_ERR(fld->lcf_cache);
 315                 fld->lcf_cache = NULL;
 316                 GOTO(out, rc);
 317         }
 318
 319         rc = fld_client_proc_init(fld);
 320         if (rc)
 321                 GOTO(out, rc);
 322         EXIT;
 323 out:
 324         if (rc)
 325                 fld_client_fini(fld);
 326         else
 327                 CDEBUG(D_INFO, "%s: Using \"%s\" hash\n",
 328                        fld->lcf_name, fld->lcf_hash->fh_name);
 329         return rc;
 330 }
 331 EXPORT_SYMBOL(fld_client_init);
 332
 333 void fld_client_fini(struct lu_client_fld *fld)
 334 {
 335         struct lu_fld_target *target, *tmp;
 336         ENTRY;
 337
 338         spin_lock(&fld->lcf_lock);
 339         list_for_each_entry_safe(target, tmp, &fld->lcf_targets, ft_chain) {
 340                 fld->lcf_count--;
 341                 list_del(&target->ft_chain);
 342                 if (target->ft_exp != NULL)
 343                         class_export_put(target->ft_exp);
 344                 OBD_FREE_PTR(target);
 345         }
 346         spin_unlock(&fld->lcf_lock);
 347
 348         if (fld->lcf_cache != NULL) {
 349                 if (!IS_ERR(fld->lcf_cache))
 350                         fld_cache_fini(fld->lcf_cache);
 351                 fld->lcf_cache = NULL;
 352         }
 353
 354         EXIT;
 355 }
 356 EXPORT_SYMBOL(fld_client_fini);
 357
 358 int fld_client_rpc(struct obd_export *exp,
 359                    struct lu_seq_range *range, __u32 fld_op,
 360                    struct ptlrpc_request **reqp)
 361 {
 362         struct ptlrpc_request *req = NULL;
 363         struct lu_seq_range   *prange;
 364         __u32                 *op;
 365         int                    rc = 0;
 366         struct obd_import     *imp;
 367         ENTRY;
 368
 369         LASSERT(exp != NULL);
 370
 371 again:
 372         imp = class_exp2cliimp(exp);
 373         switch (fld_op) {
 374         case FLD_QUERY:
 375                 req = ptlrpc_request_alloc_pack(imp, &RQF_FLD_QUERY,
 376                                                 LUSTRE_MDS_VERSION, FLD_QUERY);
 377                 if (req == NULL)
 378                         RETURN(-ENOMEM);
 379
 380                 /* XXX: only needed when talking to old server(< 2.6), it should
 381                  * be removed when < 2.6 server is not supported */
 382                 op = req_capsule_client_get(&req->rq_pill, &RMF_FLD_OPC);
 383                 *op = FLD_LOOKUP;
 384
 385                 /* For MDS_MDS seq lookup, it will always use LWP connection,
 386                  * but LWP will be evicted after restart, so cause the error.
 387                  * so we will set no_delay for seq lookup request, once the
 388                  * request fails because of the eviction. always retry here */
 389                 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
 390                         req->rq_allow_replay = 1;
 391                         req->rq_no_delay = 1;
 392                 }
 393                 break;
 394         case FLD_READ:
 395                 req = ptlrpc_request_alloc_pack(imp, &RQF_FLD_READ,
 396                                                 LUSTRE_MDS_VERSION, FLD_READ);
 397                 if (req == NULL)
 398                         RETURN(-ENOMEM);
 399
 400                 req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA,
 401                                      RCL_SERVER, PAGE_CACHE_SIZE);
 402                 break;
 403         default:
 404                 rc = -EINVAL;
 405                 break;
 406         }
 407
 408         if (rc != 0)
 409                 RETURN(rc);
 410
 411         prange = req_capsule_client_get(&req->rq_pill, &RMF_FLD_MDFLD);
 412         *prange = *range;
 413         ptlrpc_request_set_replen(req);
 414         req->rq_request_portal = FLD_REQUEST_PORTAL;
 415         req->rq_reply_portal = MDC_REPLY_PORTAL;
 416         ptlrpc_at_set_req_timeout(req);
 417
 418         obd_get_request_slot(&exp->exp_obd->u.cli);
 419         rc = ptlrpc_queue_wait(req);
 420         obd_put_request_slot(&exp->exp_obd->u.cli);
 421
 422         if (rc == -ENOENT) {
 423                 /* Don't loop forever on non-existing FID sequences. */
 424                 GOTO(out_req, rc);
 425         }
 426
 427         if (rc != 0) {
 428                 if (imp->imp_state != LUSTRE_IMP_CLOSED && !imp->imp_deactive) {
 429                         /* Since LWP is not replayable, so it will keep
 430                          * trying unless umount happens, otherwise it would
 431                          * cause unecessary failure of the application. */
 432                         ptlrpc_req_finished(req);
 433                         rc = 0;
 434                         goto again;
 435                 }
 436                 GOTO(out_req, rc);
 437         }
 438
 439         if (fld_op == FLD_QUERY) {
 440                 prange = req_capsule_server_get(&req->rq_pill,
 441                                                 &RMF_FLD_MDFLD);
 442                 if (prange == NULL)
 443                         GOTO(out_req, rc = -EFAULT);
 444                 *range = *prange;
 445         }
 446
 447         EXIT;
 448 out_req:
 449         if (rc != 0 || reqp == NULL) {
 450                 ptlrpc_req_finished(req);
 451                 req = NULL;
 452         }
 453
 454         if (reqp != NULL)
 455                 *reqp = req;
 456
 457         return rc;
 458 }
 459
 460 int fld_client_lookup(struct lu_client_fld *fld, u64 seq, u32 *mds,
 461                       __u32 flags, const struct lu_env *env)
 462 {
 463         struct lu_seq_range res = { 0 };
 464         struct lu_fld_target *target;
 465         struct lu_fld_target *origin;
 466         int rc;
 467         ENTRY;
 468
 469         rc = fld_cache_lookup(fld->lcf_cache, seq, &res);
 470         if (rc == 0) {
 471                 *mds = res.lsr_index;
 472                 RETURN(0);
 473         }
 474
 475         /* Can not find it in the cache */
 476         target = fld_client_get_target(fld, seq);
 477         LASSERT(target != NULL);
 478         origin = target;
 479 again:
 480         CDEBUG(D_INFO, "%s: Lookup fld entry (seq: "LPX64") on "
 481                "target %s (idx "LPU64")\n", fld->lcf_name, seq,
 482                fld_target_name(target), target->ft_idx);
 483
 484         res.lsr_start = seq;
 485         fld_range_set_type(&res, flags);
 486
 487 #ifdef HAVE_SERVER_SUPPORT
 488         if (target->ft_srv != NULL) {
 489                 LASSERT(env != NULL);
 490                 rc = fld_server_lookup(env, target->ft_srv, seq, &res);
 491         } else
 492 #endif /* HAVE_SERVER_SUPPORT */
 493         {
 494                 rc = fld_client_rpc(target->ft_exp, &res, FLD_QUERY, NULL);
 495         }
 496
 497         if (rc == -ESHUTDOWN) {
 498                 /* If fld lookup failed because the target has been shutdown,
 499                  * then try next target in the list, until trying all targets
 500                  * or fld lookup succeeds */
 501                 spin_lock(&fld->lcf_lock);
 502
 503                 /* If the next entry in the list is the head of the list,
 504                  * move to the next entry after the head and retrieve
 505                  * the target. Else retreive the next target entry. */
 506
 507                 if (target->ft_chain.next == &fld->lcf_targets)
 508                         target = list_entry(target->ft_chain.next->next,
 509                                             struct lu_fld_target, ft_chain);
 510                 else
 511                         target = list_entry(target->ft_chain.next,
 512                                                  struct lu_fld_target,
 513                                                  ft_chain);
 514                 spin_unlock(&fld->lcf_lock);
 515                 if (target != origin)
 516                         goto again;
 517         }
 518         if (rc == 0) {
 519                 *mds = res.lsr_index;
 520                 fld_cache_insert(fld->lcf_cache, &res);
 521         }
 522
 523         RETURN(rc);
 524 }
 525 EXPORT_SYMBOL(fld_client_lookup);
 526
 527 void fld_client_flush(struct lu_client_fld *fld)
 528 {
 529         fld_cache_flush(fld->lcf_cache);
 530 }
 531
 532
 533 struct proc_dir_entry *fld_type_proc_dir;
 534
 535 static int __init fld_init(void)
 536 {
 537         fld_type_proc_dir = lprocfs_register(LUSTRE_FLD_NAME,
 538                                              proc_lustre_root,
 539                                              NULL, NULL);
 540         if (IS_ERR(fld_type_proc_dir))
 541                 return PTR_ERR(fld_type_proc_dir);
 542
 543 #ifdef HAVE_SERVER_SUPPORT
 544         fld_server_mod_init();
 545 #endif /* HAVE_SERVER_SUPPORT */
 546
 547         return 0;
 548 }
 549
 550 static void __exit fld_exit(void)
 551 {
 552 #ifdef HAVE_SERVER_SUPPORT
 553         fld_server_mod_exit();
 554 #endif /* HAVE_SERVER_SUPPORT */
 555
 556         if (fld_type_proc_dir != NULL && !IS_ERR(fld_type_proc_dir)) {
 557                 lprocfs_remove(&fld_type_proc_dir);
 558                 fld_type_proc_dir = NULL;
 559         }
 560 }
 561
 562 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
 563 MODULE_DESCRIPTION("Lustre FID Location Database");
 564 MODULE_VERSION(LUSTRE_VERSION_STRING);
 565 MODULE_LICENSE("GPL");
 566
 567 module_init(fld_init);
 568 module_exit(fld_exit);