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.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) 2011, 2017, 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  * lustre/fld/fld_request.c
  33  *
  34  * FLD (Fids Location Database)
  35  *
  36  * Author: Yury Umanets <umka@clusterfs.com>
  37  */
  38
  39 #define DEBUG_SUBSYSTEM S_FLD
  40
  41 #include <libcfs/libcfs.h>
  42 #include <linux/module.h>
  43 #include <linux/math64.h>
  44
  45 #include <obd.h>
  46 #include <obd_class.h>
  47 #include <obd_support.h>
  48 #include <lprocfs_status.h>
  49 #include <lustre_req_layout.h>
  50 #include <lustre_fld.h>
  51 #include <lustre_mdc.h>
  52 #include "fld_internal.h"
  53
  54 static int fld_rrb_hash(struct lu_client_fld *fld, u64 seq)
  55 {
  56         LASSERT(fld->lcf_count > 0);
  57         return do_div(seq, fld->lcf_count);
  58 }
  59
  60 static struct lu_fld_target *
  61 fld_rrb_scan(struct lu_client_fld *fld, u64 seq)
  62 {
  63         struct lu_fld_target *target;
  64         int hash;
  65         ENTRY;
  66
  67         /* Because almost all of special sequence located in MDT0,
  68          * it should go to index 0 directly, instead of calculating
  69          * hash again, and also if other MDTs is not being connected,
  70          * the fld lookup requests(for seq on MDT0) should not be
  71          * blocked because of other MDTs */
  72         if (fid_seq_is_norm(seq))
  73                 hash = fld_rrb_hash(fld, seq);
  74         else
  75                 hash = 0;
  76
  77 again:
  78         list_for_each_entry(target, &fld->lcf_targets, ft_chain) {
  79                 if (target->ft_idx == hash)
  80                         RETURN(target);
  81         }
  82
  83         if (hash != 0) {
  84                 /* It is possible the remote target(MDT) are not connected to
  85                  * with client yet, so we will refer this to MDT0, which should
  86                  * be connected during mount */
  87                 hash = 0;
  88                 goto again;
  89         }
  90
  91         CERROR("%s: Can't find target by hash %d (seq %#llx). "
  92                "Targets (%d):\n", fld->lcf_name, hash, seq,
  93                fld->lcf_count);
  94
  95         list_for_each_entry(target, &fld->lcf_targets, ft_chain) {
  96                 const char *srv_name = target->ft_srv != NULL  ?
  97                         target->ft_srv->lsf_name : "<null>";
  98                 const char *exp_name = target->ft_exp != NULL ?
  99                         (char *)target->ft_exp->exp_obd->obd_uuid.uuid :
 100                         "<null>";
 101
 102                 CERROR("  exp: 0x%p (%s), srv: 0x%p (%s), idx: %llu\n",
 103                        target->ft_exp, exp_name, target->ft_srv,
 104                        srv_name, target->ft_idx);
 105         }
 106
 107         /*
 108          * If target is not found, there is logical error anyway, so here is
 109          * LBUG() to catch this situation.
 110          */
 111         LBUG();
 112         RETURN(NULL);
 113 }
 114
 115 struct lu_fld_hash fld_hash[] = {
 116         {
 117                 .fh_name = "RRB",
 118                 .fh_hash_func = fld_rrb_hash,
 119                 .fh_scan_func = fld_rrb_scan
 120         },
 121         {
 122                 NULL,
 123         }
 124 };
 125
 126 static struct lu_fld_target *
 127 fld_client_get_target(struct lu_client_fld *fld, u64 seq)
 128 {
 129         struct lu_fld_target *target;
 130         ENTRY;
 131
 132         LASSERT(fld->lcf_hash != NULL);
 133
 134         spin_lock(&fld->lcf_lock);
 135         target = fld->lcf_hash->fh_scan_func(fld, seq);
 136         spin_unlock(&fld->lcf_lock);
 137
 138         if (target != NULL) {
 139                 CDEBUG(D_INFO, "%s: Found target (idx %llu"
 140                        ") by seq %#llx\n", fld->lcf_name,
 141                        target->ft_idx, seq);
 142         }
 143
 144         RETURN(target);
 145 }
 146
 147 /*
 148  * Add export to FLD. This is usually done by CMM and LMV as they are main users
 149  * of FLD module.
 150  */
 151 int fld_client_add_target(struct lu_client_fld *fld,
 152                           struct lu_fld_target *tar)
 153 {
 154         const char *name;
 155         struct lu_fld_target *target, *tmp;
 156         ENTRY;
 157
 158         LASSERT(tar != NULL);
 159         name = fld_target_name(tar);
 160         LASSERT(name != NULL);
 161         LASSERT(tar->ft_srv != NULL || tar->ft_exp != NULL);
 162
 163         CDEBUG(D_INFO, "%s: Adding target %s (idx %llu)\n", fld->lcf_name,
 164                name, tar->ft_idx);
 165
 166         OBD_ALLOC_PTR(target);
 167         if (target == NULL)
 168                 RETURN(-ENOMEM);
 169
 170         spin_lock(&fld->lcf_lock);
 171         list_for_each_entry(tmp, &fld->lcf_targets, ft_chain) {
 172                 if (tmp->ft_idx == tar->ft_idx) {
 173                         spin_unlock(&fld->lcf_lock);
 174                         OBD_FREE_PTR(target);
 175                         CERROR("Target %s exists in FLD and known as %s:#%llu\n",
 176                                name, fld_target_name(tmp), tmp->ft_idx);
 177                         RETURN(-EEXIST);
 178                 }
 179         }
 180
 181         target->ft_exp = tar->ft_exp;
 182         if (target->ft_exp != NULL)
 183                 class_export_get(target->ft_exp);
 184         target->ft_srv = tar->ft_srv;
 185         target->ft_idx = tar->ft_idx;
 186
 187         list_add_tail(&target->ft_chain, &fld->lcf_targets);
 188
 189         fld->lcf_count++;
 190         spin_unlock(&fld->lcf_lock);
 191
 192         RETURN(0);
 193 }
 194 EXPORT_SYMBOL(fld_client_add_target);
 195
 196 /* Remove export from FLD */
 197 int fld_client_del_target(struct lu_client_fld *fld, __u64 idx)
 198 {
 199         struct lu_fld_target *target, *tmp;
 200         ENTRY;
 201
 202         spin_lock(&fld->lcf_lock);
 203         list_for_each_entry_safe(target, tmp, &fld->lcf_targets, ft_chain) {
 204                 if (target->ft_idx == idx) {
 205                         fld->lcf_count--;
 206                         list_del(&target->ft_chain);
 207                         spin_unlock(&fld->lcf_lock);
 208
 209                         if (target->ft_exp != NULL)
 210                                 class_export_put(target->ft_exp);
 211
 212                         OBD_FREE_PTR(target);
 213                         RETURN(0);
 214                 }
 215         }
 216         spin_unlock(&fld->lcf_lock);
 217         RETURN(-ENOENT);
 218 }
 219
 220 struct dentry *fld_debugfs_dir;
 221
 222 static int fld_client_debugfs_init(struct lu_client_fld *fld)
 223 {
 224         int rc;
 225
 226         ENTRY;
 227         fld->lcf_debugfs_entry = ldebugfs_register(fld->lcf_name,
 228                                                    fld_debugfs_dir,
 229                                                    fld_client_debugfs_list,
 230                                                    fld);
 231         if (IS_ERR_OR_NULL(fld->lcf_debugfs_entry)) {
 232                 CERROR("%s: LdebugFS failed in fld-init\n", fld->lcf_name);
 233                 rc = fld->lcf_debugfs_entry ? PTR_ERR(fld->lcf_debugfs_entry)
 234                                             : -ENOMEM;
 235                 fld->lcf_debugfs_entry = NULL;
 236                 RETURN(rc);
 237         }
 238
 239         return 0;
 240 }
 241
 242 void fld_client_debugfs_fini(struct lu_client_fld *fld)
 243 {
 244         if (!IS_ERR_OR_NULL(fld->lcf_debugfs_entry))
 245                 ldebugfs_remove(&fld->lcf_debugfs_entry);
 246 }
 247 EXPORT_SYMBOL(fld_client_debugfs_fini);
 248
 249 static inline int hash_is_sane(int hash)
 250 {
 251         return (hash >= 0 && hash < ARRAY_SIZE(fld_hash));
 252 }
 253
 254 int fld_client_init(struct lu_client_fld *fld,
 255                     const char *prefix, int hash)
 256 {
 257         int cache_size, cache_threshold;
 258         int rc;
 259
 260         ENTRY;
 261         snprintf(fld->lcf_name, sizeof(fld->lcf_name),
 262                  "cli-%s", prefix);
 263
 264         if (!hash_is_sane(hash)) {
 265                 CERROR("%s: Wrong hash function %#x\n",
 266                        fld->lcf_name, hash);
 267                 RETURN(-EINVAL);
 268         }
 269
 270         fld->lcf_count = 0;
 271         spin_lock_init(&fld->lcf_lock);
 272         fld->lcf_hash = &fld_hash[hash];
 273         INIT_LIST_HEAD(&fld->lcf_targets);
 274
 275         cache_size = FLD_CLIENT_CACHE_SIZE /
 276                 sizeof(struct fld_cache_entry);
 277
 278         cache_threshold = cache_size *
 279                 FLD_CLIENT_CACHE_THRESHOLD / 100;
 280
 281         fld->lcf_cache = fld_cache_init(fld->lcf_name,
 282                                         cache_size, cache_threshold);
 283         if (IS_ERR(fld->lcf_cache)) {
 284                 rc = PTR_ERR(fld->lcf_cache);
 285                 fld->lcf_cache = NULL;
 286                 GOTO(out, rc);
 287         }
 288
 289         rc = fld_client_debugfs_init(fld);
 290         if (rc)
 291                 GOTO(out, rc);
 292         EXIT;
 293 out:
 294         if (rc)
 295                 fld_client_fini(fld);
 296         else
 297                 CDEBUG(D_INFO, "%s: Using \"%s\" hash\n",
 298                        fld->lcf_name, fld->lcf_hash->fh_name);
 299         return rc;
 300 }
 301 EXPORT_SYMBOL(fld_client_init);
 302
 303 void fld_client_fini(struct lu_client_fld *fld)
 304 {
 305         struct lu_fld_target *target, *tmp;
 306         ENTRY;
 307
 308         spin_lock(&fld->lcf_lock);
 309         list_for_each_entry_safe(target, tmp, &fld->lcf_targets, ft_chain) {
 310                 fld->lcf_count--;
 311                 list_del(&target->ft_chain);
 312                 if (target->ft_exp != NULL)
 313                         class_export_put(target->ft_exp);
 314                 OBD_FREE_PTR(target);
 315         }
 316         spin_unlock(&fld->lcf_lock);
 317
 318         if (fld->lcf_cache != NULL) {
 319                 if (!IS_ERR(fld->lcf_cache))
 320                         fld_cache_fini(fld->lcf_cache);
 321                 fld->lcf_cache = NULL;
 322         }
 323
 324         EXIT;
 325 }
 326 EXPORT_SYMBOL(fld_client_fini);
 327
 328 int fld_client_rpc(struct obd_export *exp,
 329                    struct lu_seq_range *range, __u32 fld_op,
 330                    struct ptlrpc_request **reqp)
 331 {
 332         struct ptlrpc_request *req = NULL;
 333         struct lu_seq_range   *prange;
 334         __u32                 *op;
 335         int                    rc = 0;
 336         struct obd_import     *imp;
 337         ENTRY;
 338
 339         LASSERT(exp != NULL);
 340
 341 again:
 342         imp = class_exp2cliimp(exp);
 343         switch (fld_op) {
 344         case FLD_QUERY:
 345                 req = ptlrpc_request_alloc_pack(imp, &RQF_FLD_QUERY,
 346                                                 LUSTRE_MDS_VERSION, FLD_QUERY);
 347                 if (req == NULL)
 348                         RETURN(-ENOMEM);
 349
 350                 /* XXX: only needed when talking to old server(< 2.6), it should
 351                  * be removed when < 2.6 server is not supported */
 352                 op = req_capsule_client_get(&req->rq_pill, &RMF_FLD_OPC);
 353                 *op = FLD_LOOKUP;
 354
 355                 /* For MDS_MDS seq lookup, it will always use LWP connection,
 356                  * but LWP will be evicted after restart, so cause the error.
 357                  * so we will set no_delay for seq lookup request, once the
 358                  * request fails because of the eviction. always retry here */
 359                 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
 360                         req->rq_allow_replay = 1;
 361                         req->rq_no_delay = 1;
 362                 }
 363                 break;
 364         case FLD_READ:
 365                 req = ptlrpc_request_alloc_pack(imp, &RQF_FLD_READ,
 366                                                 LUSTRE_MDS_VERSION, FLD_READ);
 367                 if (req == NULL)
 368                         RETURN(-ENOMEM);
 369
 370                 req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA,
 371                                      RCL_SERVER, PAGE_SIZE);
 372                 break;
 373         default:
 374                 rc = -EINVAL;
 375                 break;
 376         }
 377
 378         if (rc != 0)
 379                 RETURN(rc);
 380
 381         prange = req_capsule_client_get(&req->rq_pill, &RMF_FLD_MDFLD);
 382         *prange = *range;
 383         ptlrpc_request_set_replen(req);
 384         req->rq_request_portal = FLD_REQUEST_PORTAL;
 385         req->rq_reply_portal = MDC_REPLY_PORTAL;
 386         ptlrpc_at_set_req_timeout(req);
 387
 388         obd_get_request_slot(&exp->exp_obd->u.cli);
 389         rc = ptlrpc_queue_wait(req);
 390         obd_put_request_slot(&exp->exp_obd->u.cli);
 391
 392         if (rc == -ENOENT) {
 393                 /* Don't loop forever on non-existing FID sequences. */
 394                 GOTO(out_req, rc);
 395         }
 396
 397         if (rc != 0) {
 398                 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 399                     !imp->imp_deactive &&
 400                     imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS &&
 401                     OCD_HAS_FLAG(&imp->imp_connect_data, LIGHTWEIGHT) &&
 402                     rc != -ENOTSUPP) {
 403                         /* Since LWP is not replayable, so it will keep
 404                          * trying unless umount happens or the remote
 405                          * target does not support the operation, otherwise
 406                          * it would cause unecessary failure of the
 407                          * application. */
 408                         ptlrpc_req_finished(req);
 409                         rc = 0;
 410                         goto again;
 411                 }
 412                 GOTO(out_req, rc);
 413         }
 414
 415         if (fld_op == FLD_QUERY) {
 416                 prange = req_capsule_server_get(&req->rq_pill,
 417                                                 &RMF_FLD_MDFLD);
 418                 if (prange == NULL)
 419                         GOTO(out_req, rc = -EFAULT);
 420                 *range = *prange;
 421         }
 422
 423         EXIT;
 424 out_req:
 425         if (rc != 0 || reqp == NULL) {
 426                 ptlrpc_req_finished(req);
 427                 req = NULL;
 428         }
 429
 430         if (reqp != NULL)
 431                 *reqp = req;
 432
 433         return rc;
 434 }
 435
 436 int fld_client_lookup(struct lu_client_fld *fld, u64 seq, u32 *mds,
 437                       __u32 flags, const struct lu_env *env)
 438 {
 439         struct lu_seq_range res = { 0 };
 440         struct lu_fld_target *target;
 441         struct lu_fld_target *origin;
 442         int rc;
 443         ENTRY;
 444
 445         rc = fld_cache_lookup(fld->lcf_cache, seq, &res);
 446         if (rc == 0) {
 447                 *mds = res.lsr_index;
 448                 RETURN(0);
 449         }
 450
 451         /* Can not find it in the cache */
 452         target = fld_client_get_target(fld, seq);
 453         LASSERT(target != NULL);
 454         origin = target;
 455 again:
 456         CDEBUG(D_INFO, "%s: Lookup fld entry (seq: %#llx) on "
 457                "target %s (idx %llu)\n", fld->lcf_name, seq,
 458                fld_target_name(target), target->ft_idx);
 459
 460         res.lsr_start = seq;
 461         fld_range_set_type(&res, flags);
 462
 463 #ifdef HAVE_SERVER_SUPPORT
 464         if (target->ft_srv != NULL) {
 465                 LASSERT(env != NULL);
 466                 rc = fld_server_lookup(env, target->ft_srv, seq, &res);
 467         } else
 468 #endif /* HAVE_SERVER_SUPPORT */
 469         {
 470                 rc = fld_client_rpc(target->ft_exp, &res, FLD_QUERY, NULL);
 471         }
 472
 473         if (rc == -ESHUTDOWN) {
 474                 /* If fld lookup failed because the target has been shutdown,
 475                  * then try next target in the list, until trying all targets
 476                  * or fld lookup succeeds */
 477                 spin_lock(&fld->lcf_lock);
 478
 479                 /* If the next entry in the list is the head of the list,
 480                  * move to the next entry after the head and retrieve
 481                  * the target. Else retreive the next target entry. */
 482
 483                 if (target->ft_chain.next == &fld->lcf_targets)
 484                         target = list_entry(target->ft_chain.next->next,
 485                                             struct lu_fld_target, ft_chain);
 486                 else
 487                         target = list_entry(target->ft_chain.next,
 488                                                  struct lu_fld_target,
 489                                                  ft_chain);
 490                 spin_unlock(&fld->lcf_lock);
 491                 if (target != origin)
 492                         goto again;
 493         }
 494         if (rc == 0) {
 495                 *mds = res.lsr_index;
 496                 fld_cache_insert(fld->lcf_cache, &res);
 497         }
 498
 499         RETURN(rc);
 500 }
 501 EXPORT_SYMBOL(fld_client_lookup);
 502
 503 void fld_client_flush(struct lu_client_fld *fld)
 504 {
 505         fld_cache_flush(fld->lcf_cache);
 506 }
 507
 508 static int __init fld_init(void)
 509 {
 510 #ifdef HAVE_SERVER_SUPPORT
 511         int rc;
 512
 513         rc = fld_server_mod_init();
 514         if (rc)
 515                 return rc;
 516 #endif /* HAVE_SERVER_SUPPORT */
 517
 518         fld_debugfs_dir = ldebugfs_register(LUSTRE_FLD_NAME,
 519                                             debugfs_lustre_root,
 520                                             NULL, NULL);
 521         return PTR_ERR_OR_ZERO(fld_debugfs_dir);
 522 }
 523
 524 static void __exit fld_exit(void)
 525 {
 526 #ifdef HAVE_SERVER_SUPPORT
 527         fld_server_mod_exit();
 528 #endif /* HAVE_SERVER_SUPPORT */
 529
 530         if (!IS_ERR_OR_NULL(fld_debugfs_dir))
 531                 ldebugfs_remove(&fld_debugfs_dir);
 532 }
 533
 534 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
 535 MODULE_DESCRIPTION("Lustre FID Location Database");
 536 MODULE_VERSION(LUSTRE_VERSION_STRING);
 537 MODULE_LICENSE("GPL");
 538
 539 module_init(fld_init);
 540 module_exit(fld_exit);