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
  66         ENTRY;
  67
  68         /*
  69          * Because almost all of special sequence located in MDT0,
  70          * it should go to index 0 directly, instead of calculating
  71          * hash again, and also if other MDTs is not being connected,
  72          * the fld lookup requests(for seq on MDT0) should not be
  73          * blocked because of other MDTs
  74          */
  75         if (fid_seq_is_norm(seq))
  76                 hash = fld_rrb_hash(fld, seq);
  77         else
  78                 hash = 0;
  79
  80 again:
  81         list_for_each_entry(target, &fld->lcf_targets, ft_chain) {
  82                 if (target->ft_idx == hash)
  83                         RETURN(target);
  84         }
  85
  86         if (hash != 0) {
  87                 /*
  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                  */
  92                 hash = 0;
  93                 goto again;
  94         }
  95
  96         CERROR("%s: Can't find target by hash %d (seq %#llx). Targets (%d):\n",
  97                fld->lcf_name, hash, seq, 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: %llu\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
 135         ENTRY;
 136
 137         LASSERT(fld->lcf_hash != NULL);
 138
 139         spin_lock(&fld->lcf_lock);
 140         target = fld->lcf_hash->fh_scan_func(fld, seq);
 141         spin_unlock(&fld->lcf_lock);
 142
 143         if (target) {
 144                 CDEBUG(D_INFO, "%s: Found target (idx %llu) by seq %#llx\n",
 145                        fld->lcf_name, 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
 161         ENTRY;
 162
 163         LASSERT(tar != NULL);
 164         name = fld_target_name(tar);
 165         LASSERT(name != NULL);
 166         LASSERT(tar->ft_srv != NULL || tar->ft_exp != NULL);
 167
 168         CDEBUG(D_INFO, "%s: Adding target %s (idx %llu)\n", fld->lcf_name,
 169                name, tar->ft_idx);
 170
 171         OBD_ALLOC_PTR(target);
 172         if (!target)
 173                 RETURN(-ENOMEM);
 174
 175         spin_lock(&fld->lcf_lock);
 176         list_for_each_entry(tmp, &fld->lcf_targets, ft_chain) {
 177                 if (tmp->ft_idx == tar->ft_idx) {
 178                         spin_unlock(&fld->lcf_lock);
 179                         OBD_FREE_PTR(target);
 180                         CERROR("Target %s exists in FLD and known as %s:#%llu\n",
 181                                name, fld_target_name(tmp), tmp->ft_idx);
 182                         RETURN(-EEXIST);
 183                 }
 184         }
 185
 186         target->ft_exp = tar->ft_exp;
 187         if (target->ft_exp)
 188                 class_export_get(target->ft_exp);
 189         target->ft_srv = tar->ft_srv;
 190         target->ft_idx = tar->ft_idx;
 191
 192         list_add_tail(&target->ft_chain, &fld->lcf_targets);
 193
 194         fld->lcf_count++;
 195         spin_unlock(&fld->lcf_lock);
 196
 197         RETURN(0);
 198 }
 199 EXPORT_SYMBOL(fld_client_add_target);
 200
 201 /* Remove export from FLD */
 202 int fld_client_del_target(struct lu_client_fld *fld, u64 idx)
 203 {
 204         struct lu_fld_target *target, *tmp;
 205
 206         ENTRY;
 207
 208         spin_lock(&fld->lcf_lock);
 209         list_for_each_entry_safe(target, tmp, &fld->lcf_targets, ft_chain) {
 210                 if (target->ft_idx == idx) {
 211                         fld->lcf_count--;
 212                         list_del(&target->ft_chain);
 213                         spin_unlock(&fld->lcf_lock);
 214
 215                         if (target->ft_exp)
 216                                 class_export_put(target->ft_exp);
 217
 218                         OBD_FREE_PTR(target);
 219                         RETURN(0);
 220                 }
 221         }
 222         spin_unlock(&fld->lcf_lock);
 223         RETURN(-ENOENT);
 224 }
 225
 226 struct dentry *fld_debugfs_dir;
 227
 228 static int fld_client_debugfs_init(struct lu_client_fld *fld)
 229 {
 230         ENTRY;
 231         fld->lcf_debugfs_entry = debugfs_create_dir(fld->lcf_name,
 232                                                    fld_debugfs_dir);
 233         ldebugfs_add_vars(fld->lcf_debugfs_entry,
 234                           fld_client_debugfs_list,
 235                           fld);
 236
 237         return 0;
 238 }
 239
 240 void fld_client_debugfs_fini(struct lu_client_fld *fld)
 241 {
 242         debugfs_remove_recursive(fld->lcf_debugfs_entry);
 243 }
 244 EXPORT_SYMBOL(fld_client_debugfs_fini);
 245
 246 static inline int hash_is_sane(int hash)
 247 {
 248         return (hash >= 0 && hash < ARRAY_SIZE(fld_hash));
 249 }
 250
 251 int fld_client_init(struct lu_client_fld *fld,
 252                     const char *prefix, int hash)
 253 {
 254         int cache_size, cache_threshold;
 255         int rc;
 256
 257         ENTRY;
 258         snprintf(fld->lcf_name, sizeof(fld->lcf_name),
 259                  "cli-%s", prefix);
 260
 261         if (!hash_is_sane(hash)) {
 262                 CERROR("%s: Wrong hash function %#x\n",
 263                        fld->lcf_name, hash);
 264                 RETURN(-EINVAL);
 265         }
 266
 267         fld->lcf_count = 0;
 268         spin_lock_init(&fld->lcf_lock);
 269         fld->lcf_hash = &fld_hash[hash];
 270         INIT_LIST_HEAD(&fld->lcf_targets);
 271
 272         cache_size = FLD_CLIENT_CACHE_SIZE /
 273                 sizeof(struct fld_cache_entry);
 274
 275         cache_threshold = cache_size *
 276                 FLD_CLIENT_CACHE_THRESHOLD / 100;
 277
 278         fld->lcf_cache = fld_cache_init(fld->lcf_name,
 279                                         cache_size, cache_threshold);
 280         if (IS_ERR(fld->lcf_cache)) {
 281                 rc = PTR_ERR(fld->lcf_cache);
 282                 fld->lcf_cache = NULL;
 283                 GOTO(out, rc);
 284         }
 285
 286         rc = fld_client_debugfs_init(fld);
 287         if (rc)
 288                 GOTO(out, rc);
 289         EXIT;
 290 out:
 291         if (rc)
 292                 fld_client_fini(fld);
 293         else
 294                 CDEBUG(D_INFO, "%s: Using \"%s\" hash\n",
 295                        fld->lcf_name, fld->lcf_hash->fh_name);
 296         return rc;
 297 }
 298 EXPORT_SYMBOL(fld_client_init);
 299
 300 void fld_client_fini(struct lu_client_fld *fld)
 301 {
 302         struct lu_fld_target *target, *tmp;
 303
 304         ENTRY;
 305
 306         spin_lock(&fld->lcf_lock);
 307         list_for_each_entry_safe(target, tmp, &fld->lcf_targets, ft_chain) {
 308                 fld->lcf_count--;
 309                 list_del(&target->ft_chain);
 310                 if (target->ft_exp)
 311                         class_export_put(target->ft_exp);
 312                 OBD_FREE_PTR(target);
 313         }
 314         spin_unlock(&fld->lcf_lock);
 315
 316         if (fld->lcf_cache) {
 317                 if (!IS_ERR(fld->lcf_cache))
 318                         fld_cache_fini(fld->lcf_cache);
 319                 fld->lcf_cache = NULL;
 320         }
 321
 322         EXIT;
 323 }
 324 EXPORT_SYMBOL(fld_client_fini);
 325
 326 int fld_client_rpc(struct obd_export *exp,
 327                    struct lu_seq_range *range, u32 fld_op,
 328                    struct ptlrpc_request **reqp)
 329 {
 330         struct ptlrpc_request *req = NULL;
 331         struct lu_seq_range *prange;
 332         u32 *op;
 333         int rc = 0;
 334         struct obd_import *imp;
 335
 336         ENTRY;
 337
 338         LASSERT(exp != NULL);
 339
 340         imp = class_exp2cliimp(exp);
 341         switch (fld_op) {
 342         case FLD_QUERY:
 343                 req = ptlrpc_request_alloc_pack(imp, &RQF_FLD_QUERY,
 344                                                 LUSTRE_MDS_VERSION, FLD_QUERY);
 345                 if (!req)
 346                         RETURN(-ENOMEM);
 347
 348                 /*
 349                  * XXX: only needed when talking to old server(< 2.6), it should
 350                  * be removed when < 2.6 server is not supported
 351                  */
 352                 op = req_capsule_client_get(&req->rq_pill, &RMF_FLD_OPC);
 353                 *op = FLD_LOOKUP;
 354
 355                 /*
 356                  * For MDS_MDS seq lookup, it will always use LWP connection,
 357                  * but LWP will be evicted after restart, so cause the error.
 358                  * so we will set no_delay for seq lookup request, once the
 359                  * request fails because of the eviction. always retry here
 360                  */
 361                 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
 362                         req->rq_allow_replay = 1;
 363                         req->rq_no_delay = 1;
 364                 }
 365                 break;
 366         case FLD_READ:
 367                 req = ptlrpc_request_alloc_pack(imp, &RQF_FLD_READ,
 368                                                 LUSTRE_MDS_VERSION, FLD_READ);
 369                 if (!req)
 370                         RETURN(-ENOMEM);
 371
 372                 req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA,
 373                                      RCL_SERVER, PAGE_SIZE);
 374                 break;
 375         default:
 376                 rc = -EINVAL;
 377                 break;
 378         }
 379
 380         if (rc != 0)
 381                 RETURN(rc);
 382
 383         prange = req_capsule_client_get(&req->rq_pill, &RMF_FLD_MDFLD);
 384         *prange = *range;
 385         ptlrpc_request_set_replen(req);
 386         req->rq_request_portal = FLD_REQUEST_PORTAL;
 387         req->rq_reply_portal = MDC_REPLY_PORTAL;
 388         ptlrpc_at_set_req_timeout(req);
 389
 390         if (OBD_FAIL_CHECK(OBD_FAIL_FLD_QUERY_REQ && req->rq_no_delay)) {
 391                 /* the same error returned by ptlrpc_import_delay_req */
 392                 rc = -EWOULDBLOCK;
 393                 req->rq_status = rc;
 394         } else {
 395                 obd_get_request_slot(&exp->exp_obd->u.cli);
 396                 rc = ptlrpc_queue_wait(req);
 397                 obd_put_request_slot(&exp->exp_obd->u.cli);
 398         }
 399
 400         if (rc == -ENOENT) {
 401                 /* Don't loop forever on non-existing FID sequences. */
 402                 GOTO(out_req, rc);
 403         }
 404
 405         if (rc != 0) {
 406                 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 407                     !imp->imp_deactive &&
 408                     imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS &&
 409                     OCD_HAS_FLAG(&imp->imp_connect_data, LIGHTWEIGHT) &&
 410                     rc != -ENOTSUPP) {
 411                         /*
 412                          * Since LWP is not replayable, so notify the caller
 413                          * to retry if needed after a while.
 414                          */
 415                         rc = -EAGAIN;
 416                 }
 417                 GOTO(out_req, rc);
 418         }
 419
 420         if (fld_op == FLD_QUERY) {
 421                 prange = req_capsule_server_get(&req->rq_pill,
 422                                                 &RMF_FLD_MDFLD);
 423                 if (!prange)
 424                         GOTO(out_req, rc = -EFAULT);
 425                 *range = *prange;
 426         }
 427
 428         EXIT;
 429 out_req:
 430         if (rc != 0 || !reqp) {
 431                 ptlrpc_req_finished(req);
 432                 req = NULL;
 433         }
 434
 435         if (reqp)
 436                 *reqp = req;
 437
 438         return rc;
 439 }
 440
 441 int fld_client_lookup(struct lu_client_fld *fld, u64 seq, u32 *mds,
 442                       u32 flags, const struct lu_env *env)
 443 {
 444         struct lu_seq_range res = { 0 };
 445         struct lu_fld_target *target;
 446         struct lu_fld_target *origin;
 447         int rc;
 448
 449         ENTRY;
 450
 451         rc = fld_cache_lookup(fld->lcf_cache, seq, &res);
 452         if (rc == 0) {
 453                 *mds = res.lsr_index;
 454                 RETURN(0);
 455         }
 456
 457         /* Can not find it in the cache */
 458         target = fld_client_get_target(fld, seq);
 459         LASSERT(target != NULL);
 460         origin = target;
 461 again:
 462         CDEBUG(D_INFO, "%s: Lookup fld entry (seq: %#llx) on target %s (idx %llu)\n",
 463                fld->lcf_name, seq, fld_target_name(target), target->ft_idx);
 464
 465         res.lsr_start = seq;
 466         fld_range_set_type(&res, flags);
 467
 468 #ifdef HAVE_SERVER_SUPPORT
 469         if (target->ft_srv) {
 470                 LASSERT(env != NULL);
 471                 rc = fld_server_lookup(env, target->ft_srv, seq, &res);
 472         } else
 473 #endif /* HAVE_SERVER_SUPPORT */
 474         {
 475                 rc = fld_client_rpc(target->ft_exp, &res, FLD_QUERY, NULL);
 476         }
 477
 478         if (rc == -ESHUTDOWN) {
 479                 /*
 480                  * If fld lookup failed because the target has been shutdown,
 481                  * then try next target in the list, until trying all targets
 482                  * or fld lookup succeeds
 483                  */
 484                 spin_lock(&fld->lcf_lock);
 485                 /*
 486                  * If the next entry in the list is the head of the list,
 487                  * move to the next entry after the head and retrieve
 488                  * the target. Else retreive the next target entry.
 489                  */
 490                 if (target->ft_chain.next == &fld->lcf_targets)
 491                         target = list_entry(target->ft_chain.next->next,
 492                                             struct lu_fld_target, ft_chain);
 493                 else
 494                         target = list_entry(target->ft_chain.next,
 495                                                  struct lu_fld_target,
 496                                                  ft_chain);
 497                 spin_unlock(&fld->lcf_lock);
 498                 if (target != origin)
 499                         goto again;
 500         }
 501         if (rc == 0) {
 502                 *mds = res.lsr_index;
 503                 fld_cache_insert(fld->lcf_cache, &res);
 504         }
 505
 506         RETURN(rc);
 507 }
 508 EXPORT_SYMBOL(fld_client_lookup);
 509
 510 void fld_client_flush(struct lu_client_fld *fld)
 511 {
 512         fld_cache_flush(fld->lcf_cache);
 513 }
 514
 515 static int __init fld_init(void)
 516 {
 517 #ifdef HAVE_SERVER_SUPPORT
 518         int rc;
 519
 520         rc = fld_server_mod_init();
 521         if (rc)
 522                 return rc;
 523 #endif /* HAVE_SERVER_SUPPORT */
 524
 525         fld_debugfs_dir = debugfs_create_dir(LUSTRE_FLD_NAME,
 526                                              debugfs_lustre_root);
 527         return PTR_ERR_OR_ZERO(fld_debugfs_dir);
 528 }
 529
 530 static void __exit fld_exit(void)
 531 {
 532 #ifdef HAVE_SERVER_SUPPORT
 533         fld_server_mod_exit();
 534 #endif /* HAVE_SERVER_SUPPORT */
 535
 536         debugfs_remove_recursive(fld_debugfs_dir);
 537 }
 538
 539 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
 540 MODULE_DESCRIPTION("Lustre FID Location Database");
 541 MODULE_VERSION(LUSTRE_VERSION_STRING);
 542 MODULE_LICENSE("GPL");
 543
 544 module_init(fld_init);
 545 module_exit(fld_exit);