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 #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                 0,
 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         if (fld->lcf_flags != LUSTRE_FLD_INIT) {
 168                 CERROR("%s: Attempt to add target %s (idx "LPU64") "
 169                        "on fly - skip it\n", fld->lcf_name, name,
 170                        tar->ft_idx);
 171                 RETURN(0);
 172         } else {
 173                 CDEBUG(D_INFO, "%s: Adding target %s (idx "
 174                        LPU64")\n", fld->lcf_name, name, tar->ft_idx);
 175         }
 176
 177         OBD_ALLOC_PTR(target);
 178         if (target == NULL)
 179                 RETURN(-ENOMEM);
 180
 181         spin_lock(&fld->lcf_lock);
 182         list_for_each_entry(tmp, &fld->lcf_targets, ft_chain) {
 183                 if (tmp->ft_idx == tar->ft_idx) {
 184                         spin_unlock(&fld->lcf_lock);
 185                         OBD_FREE_PTR(target);
 186                         CERROR("Target %s exists in FLD and known as %s:#"LPU64"\n",
 187                                name, fld_target_name(tmp), tmp->ft_idx);
 188                         RETURN(-EEXIST);
 189                 }
 190         }
 191
 192         target->ft_exp = tar->ft_exp;
 193         if (target->ft_exp != NULL)
 194                 class_export_get(target->ft_exp);
 195         target->ft_srv = tar->ft_srv;
 196         target->ft_idx = tar->ft_idx;
 197
 198         list_add_tail(&target->ft_chain, &fld->lcf_targets);
 199
 200         fld->lcf_count++;
 201         spin_unlock(&fld->lcf_lock);
 202
 203         RETURN(0);
 204 }
 205 EXPORT_SYMBOL(fld_client_add_target);
 206
 207 /* Remove export from FLD */
 208 int fld_client_del_target(struct lu_client_fld *fld, __u64 idx)
 209 {
 210         struct lu_fld_target *target, *tmp;
 211         ENTRY;
 212
 213         spin_lock(&fld->lcf_lock);
 214         list_for_each_entry_safe(target, tmp, &fld->lcf_targets, ft_chain) {
 215                 if (target->ft_idx == idx) {
 216                         fld->lcf_count--;
 217                         list_del(&target->ft_chain);
 218                         spin_unlock(&fld->lcf_lock);
 219
 220                         if (target->ft_exp != NULL)
 221                                 class_export_put(target->ft_exp);
 222
 223                         OBD_FREE_PTR(target);
 224                         RETURN(0);
 225                 }
 226         }
 227         spin_unlock(&fld->lcf_lock);
 228         RETURN(-ENOENT);
 229 }
 230 EXPORT_SYMBOL(fld_client_del_target);
 231
 232 #ifdef LPROCFS
 233 static int fld_client_proc_init(struct lu_client_fld *fld)
 234 {
 235         int rc;
 236         ENTRY;
 237
 238         fld->lcf_proc_dir = lprocfs_seq_register(fld->lcf_name,
 239                                                  fld_type_proc_dir,
 240                                                  NULL, NULL);
 241         if (IS_ERR(fld->lcf_proc_dir)) {
 242                 CERROR("%s: LProcFS failed in fld-init\n",
 243                        fld->lcf_name);
 244                 rc = PTR_ERR(fld->lcf_proc_dir);
 245                 RETURN(rc);
 246         }
 247
 248         rc = lprocfs_seq_add_vars(fld->lcf_proc_dir,
 249                                   fld_client_proc_list, fld);
 250         if (rc) {
 251                 CERROR("%s: Can't init FLD proc, rc %d\n",
 252                        fld->lcf_name, rc);
 253                 GOTO(out_cleanup, rc);
 254         }
 255
 256         RETURN(0);
 257
 258 out_cleanup:
 259         fld_client_proc_fini(fld);
 260         return rc;
 261 }
 262
 263 void fld_client_proc_fini(struct lu_client_fld *fld)
 264 {
 265         ENTRY;
 266         if (fld->lcf_proc_dir) {
 267                 if (!IS_ERR(fld->lcf_proc_dir))
 268                         lprocfs_remove(&fld->lcf_proc_dir);
 269                 fld->lcf_proc_dir = NULL;
 270         }
 271         EXIT;
 272 }
 273 #else /* LPROCFS */
 274 static int fld_client_proc_init(struct lu_client_fld *fld)
 275 {
 276         return 0;
 277 }
 278
 279 void fld_client_proc_fini(struct lu_client_fld *fld)
 280 {
 281         return;
 282 }
 283 #endif /* !LPROCFS */
 284
 285 EXPORT_SYMBOL(fld_client_proc_fini);
 286
 287 static inline int hash_is_sane(int hash)
 288 {
 289         return (hash >= 0 && hash < ARRAY_SIZE(fld_hash));
 290 }
 291
 292 int fld_client_init(struct lu_client_fld *fld,
 293                     const char *prefix, int hash)
 294 {
 295         int cache_size, cache_threshold;
 296         int rc;
 297         ENTRY;
 298
 299         LASSERT(fld != NULL);
 300
 301         snprintf(fld->lcf_name, sizeof(fld->lcf_name),
 302                  "cli-%s", prefix);
 303
 304         if (!hash_is_sane(hash)) {
 305                 CERROR("%s: Wrong hash function %#x\n",
 306                        fld->lcf_name, hash);
 307                 RETURN(-EINVAL);
 308         }
 309
 310         fld->lcf_count = 0;
 311         spin_lock_init(&fld->lcf_lock);
 312         fld->lcf_hash = &fld_hash[hash];
 313         fld->lcf_flags = LUSTRE_FLD_INIT;
 314         INIT_LIST_HEAD(&fld->lcf_targets);
 315
 316         cache_size = FLD_CLIENT_CACHE_SIZE /
 317                 sizeof(struct fld_cache_entry);
 318
 319         cache_threshold = cache_size *
 320                 FLD_CLIENT_CACHE_THRESHOLD / 100;
 321
 322         fld->lcf_cache = fld_cache_init(fld->lcf_name,
 323                                         cache_size, cache_threshold);
 324         if (IS_ERR(fld->lcf_cache)) {
 325                 rc = PTR_ERR(fld->lcf_cache);
 326                 fld->lcf_cache = NULL;
 327                 GOTO(out, rc);
 328         }
 329
 330         rc = fld_client_proc_init(fld);
 331         if (rc)
 332                 GOTO(out, rc);
 333         EXIT;
 334 out:
 335         if (rc)
 336                 fld_client_fini(fld);
 337         else
 338                 CDEBUG(D_INFO, "%s: Using \"%s\" hash\n",
 339                        fld->lcf_name, fld->lcf_hash->fh_name);
 340         return rc;
 341 }
 342 EXPORT_SYMBOL(fld_client_init);
 343
 344 void fld_client_fini(struct lu_client_fld *fld)
 345 {
 346         struct lu_fld_target *target, *tmp;
 347         ENTRY;
 348
 349         spin_lock(&fld->lcf_lock);
 350         list_for_each_entry_safe(target, tmp, &fld->lcf_targets, ft_chain) {
 351                 fld->lcf_count--;
 352                 list_del(&target->ft_chain);
 353                 if (target->ft_exp != NULL)
 354                         class_export_put(target->ft_exp);
 355                 OBD_FREE_PTR(target);
 356         }
 357         spin_unlock(&fld->lcf_lock);
 358
 359         if (fld->lcf_cache != NULL) {
 360                 if (!IS_ERR(fld->lcf_cache))
 361                         fld_cache_fini(fld->lcf_cache);
 362                 fld->lcf_cache = NULL;
 363         }
 364
 365         EXIT;
 366 }
 367 EXPORT_SYMBOL(fld_client_fini);
 368
 369 int fld_client_rpc(struct obd_export *exp,
 370                    struct lu_seq_range *range, __u32 fld_op,
 371                    struct ptlrpc_request **reqp)
 372 {
 373         struct ptlrpc_request *req = NULL;
 374         struct lu_seq_range   *prange;
 375         __u32                 *op;
 376         int                    rc = 0;
 377         struct obd_import     *imp;
 378         ENTRY;
 379
 380         LASSERT(exp != NULL);
 381
 382 again:
 383         imp = class_exp2cliimp(exp);
 384         switch (fld_op) {
 385         case FLD_QUERY:
 386                 req = ptlrpc_request_alloc_pack(imp, &RQF_FLD_QUERY,
 387                                                 LUSTRE_MDS_VERSION, FLD_QUERY);
 388                 if (req == NULL)
 389                         RETURN(-ENOMEM);
 390
 391                 /* XXX: only needed when talking to old server(< 2.6), it should
 392                  * be removed when < 2.6 server is not supported */
 393                 op = req_capsule_client_get(&req->rq_pill, &RMF_FLD_OPC);
 394                 *op = FLD_LOOKUP;
 395
 396                 /* For MDS_MDS seq lookup, it will always use LWP connection,
 397                  * but LWP will be evicted after restart, so cause the error.
 398                  * so we will set no_delay for seq lookup request, once the
 399                  * request fails because of the eviction. always retry here */
 400                 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
 401                         req->rq_allow_replay = 1;
 402                         req->rq_no_delay = 1;
 403                 }
 404                 break;
 405         case FLD_READ:
 406                 req = ptlrpc_request_alloc_pack(imp, &RQF_FLD_READ,
 407                                                 LUSTRE_MDS_VERSION, FLD_READ);
 408                 if (req == NULL)
 409                         RETURN(-ENOMEM);
 410
 411                 req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA,
 412                                      RCL_SERVER, PAGE_CACHE_SIZE);
 413                 break;
 414         default:
 415                 rc = -EINVAL;
 416                 break;
 417         }
 418
 419         if (rc != 0)
 420                 RETURN(rc);
 421
 422         prange = req_capsule_client_get(&req->rq_pill, &RMF_FLD_MDFLD);
 423         *prange = *range;
 424         ptlrpc_request_set_replen(req);
 425         req->rq_request_portal = FLD_REQUEST_PORTAL;
 426         req->rq_reply_portal = MDC_REPLY_PORTAL;
 427         ptlrpc_at_set_req_timeout(req);
 428
 429         obd_get_request_slot(&exp->exp_obd->u.cli);
 430         rc = ptlrpc_queue_wait(req);
 431         obd_put_request_slot(&exp->exp_obd->u.cli);
 432         if (rc != 0) {
 433                 if (imp->imp_state != LUSTRE_IMP_CLOSED) {
 434                         /* Since LWP is not replayable, so it will keep
 435                          * trying unless umount happens, otherwise it would
 436                          * cause unecessary failure of the application. */
 437                         ptlrpc_req_finished(req);
 438                         rc = 0;
 439                         goto again;
 440                 }
 441                 GOTO(out_req, rc);
 442         }
 443
 444         if (fld_op == FLD_QUERY) {
 445                 prange = req_capsule_server_get(&req->rq_pill,
 446                                                 &RMF_FLD_MDFLD);
 447                 if (prange == NULL)
 448                         GOTO(out_req, rc = -EFAULT);
 449                 *range = *prange;
 450         }
 451
 452         EXIT;
 453 out_req:
 454         if (rc != 0 || reqp == NULL) {
 455                 ptlrpc_req_finished(req);
 456                 req = NULL;
 457         }
 458
 459         if (reqp != NULL)
 460                 *reqp = req;
 461
 462         return rc;
 463 }
 464
 465 int fld_client_lookup(struct lu_client_fld *fld, u64 seq, u32 *mds,
 466                       __u32 flags, const struct lu_env *env)
 467 {
 468         struct lu_seq_range res = { 0 };
 469         struct lu_fld_target *target;
 470         int rc;
 471         ENTRY;
 472
 473         fld->lcf_flags |= LUSTRE_FLD_RUN;
 474
 475         rc = fld_cache_lookup(fld->lcf_cache, seq, &res);
 476         if (rc == 0) {
 477                 *mds = res.lsr_index;
 478                 RETURN(0);
 479         }
 480
 481         /* Can not find it in the cache */
 482         target = fld_client_get_target(fld, seq);
 483         LASSERT(target != NULL);
 484
 485         CDEBUG(D_INFO, "%s: Lookup fld entry (seq: "LPX64") on "
 486                "target %s (idx "LPU64")\n", fld->lcf_name, seq,
 487                fld_target_name(target), target->ft_idx);
 488
 489         res.lsr_start = seq;
 490         fld_range_set_type(&res, flags);
 491
 492 #ifdef HAVE_SERVER_SUPPORT
 493         if (target->ft_srv != NULL) {
 494                 LASSERT(env != NULL);
 495                 rc = fld_server_lookup(env, target->ft_srv, seq, &res);
 496         } else
 497 #endif /* HAVE_SERVER_SUPPORT */
 498         {
 499                 rc = fld_client_rpc(target->ft_exp, &res, FLD_QUERY, NULL);
 500         }
 501
 502         if (rc == 0) {
 503                 *mds = res.lsr_index;
 504                 fld_cache_insert(fld->lcf_cache, &res);
 505         }
 506
 507         RETURN(rc);
 508 }
 509 EXPORT_SYMBOL(fld_client_lookup);
 510
 511 void fld_client_flush(struct lu_client_fld *fld)
 512 {
 513         fld_cache_flush(fld->lcf_cache);
 514 }
 515 EXPORT_SYMBOL(fld_client_flush);
 516
 517
 518 struct proc_dir_entry *fld_type_proc_dir;
 519
 520 static int __init fld_mod_init(void)
 521 {
 522         fld_type_proc_dir = lprocfs_seq_register(LUSTRE_FLD_NAME,
 523                                                  proc_lustre_root,
 524                                                  NULL, NULL);
 525         if (IS_ERR(fld_type_proc_dir))
 526                 return PTR_ERR(fld_type_proc_dir);
 527
 528 #ifdef HAVE_SERVER_SUPPORT
 529         fld_server_mod_init();
 530 #endif /* HAVE_SERVER_SUPPORT */
 531
 532         return 0;
 533 }
 534
 535 static void __exit fld_mod_exit(void)
 536 {
 537 #ifdef HAVE_SERVER_SUPPORT
 538         fld_server_mod_exit();
 539 #endif /* HAVE_SERVER_SUPPORT */
 540
 541         if (fld_type_proc_dir != NULL && !IS_ERR(fld_type_proc_dir)) {
 542                 lprocfs_remove(&fld_type_proc_dir);
 543                 fld_type_proc_dir = NULL;
 544         }
 545 }
 546
 547 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
 548 MODULE_DESCRIPTION("Lustre FLD");
 549 MODULE_LICENSE("GPL");
 550
 551 cfs_module(mdd, LUSTRE_VERSION_STRING, fld_mod_init, fld_mod_exit);