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