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 void 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
 238 void fld_client_debugfs_fini(struct lu_client_fld *fld)
 239 {
 240         debugfs_remove_recursive(fld->lcf_debugfs_entry);
 241 }
 242 EXPORT_SYMBOL(fld_client_debugfs_fini);
 243
 244 static inline int hash_is_sane(int hash)
 245 {
 246         return (hash >= 0 && hash < ARRAY_SIZE(fld_hash));
 247 }
 248
 249 int fld_client_init(struct lu_client_fld *fld,
 250                     const char *prefix, int hash)
 251 {
 252         int cache_size, cache_threshold;
 253         int rc = 0;
 254
 255         ENTRY;
 256         snprintf(fld->lcf_name, sizeof(fld->lcf_name),
 257                  "cli-%s", prefix);
 258
 259         if (!hash_is_sane(hash)) {
 260                 CERROR("%s: Wrong hash function %#x\n",
 261                        fld->lcf_name, hash);
 262                 RETURN(-EINVAL);
 263         }
 264
 265         fld->lcf_count = 0;
 266         spin_lock_init(&fld->lcf_lock);
 267         fld->lcf_hash = &fld_hash[hash];
 268         INIT_LIST_HEAD(&fld->lcf_targets);
 269
 270         cache_size = FLD_CLIENT_CACHE_SIZE /
 271                 sizeof(struct fld_cache_entry);
 272
 273         cache_threshold = cache_size *
 274                 FLD_CLIENT_CACHE_THRESHOLD / 100;
 275
 276         fld->lcf_cache = fld_cache_init(fld->lcf_name,
 277                                         cache_size, cache_threshold);
 278         if (IS_ERR(fld->lcf_cache)) {
 279                 rc = PTR_ERR(fld->lcf_cache);
 280                 fld->lcf_cache = NULL;
 281                 GOTO(out, rc);
 282         }
 283
 284         fld_client_debugfs_init(fld);
 285         EXIT;
 286 out:
 287         if (rc)
 288                 fld_client_fini(fld);
 289         else
 290                 CDEBUG(D_INFO, "%s: Using \"%s\" hash\n",
 291                        fld->lcf_name, fld->lcf_hash->fh_name);
 292         return rc;
 293 }
 294 EXPORT_SYMBOL(fld_client_init);
 295
 296 void fld_client_fini(struct lu_client_fld *fld)
 297 {
 298         struct lu_fld_target *target, *tmp;
 299
 300         ENTRY;
 301
 302         spin_lock(&fld->lcf_lock);
 303         list_for_each_entry_safe(target, tmp, &fld->lcf_targets, ft_chain) {
 304                 fld->lcf_count--;
 305                 list_del(&target->ft_chain);
 306                 if (target->ft_exp)
 307                         class_export_put(target->ft_exp);
 308                 OBD_FREE_PTR(target);
 309         }
 310         spin_unlock(&fld->lcf_lock);
 311
 312         if (fld->lcf_cache) {
 313                 if (!IS_ERR(fld->lcf_cache))
 314                         fld_cache_fini(fld->lcf_cache);
 315                 fld->lcf_cache = NULL;
 316         }
 317
 318         EXIT;
 319 }
 320 EXPORT_SYMBOL(fld_client_fini);
 321
 322 int fld_client_rpc(struct obd_export *exp,
 323                    struct lu_seq_range *range, u32 fld_op,
 324                    struct ptlrpc_request **reqp)
 325 {
 326         struct ptlrpc_request *req = NULL;
 327         struct lu_seq_range *prange;
 328         u32 *op;
 329         int rc = 0;
 330         struct obd_import *imp;
 331
 332         ENTRY;
 333
 334         LASSERT(exp != NULL);
 335
 336         imp = class_exp2cliimp(exp);
 337         switch (fld_op) {
 338         case FLD_QUERY:
 339                 req = ptlrpc_request_alloc_pack(imp, &RQF_FLD_QUERY,
 340                                                 LUSTRE_MDS_VERSION, FLD_QUERY);
 341                 if (!req)
 342                         RETURN(-ENOMEM);
 343
 344                 /*
 345                  * XXX: only needed when talking to old server(< 2.6), it should
 346                  * be removed when < 2.6 server is not supported
 347                  */
 348                 op = req_capsule_client_get(&req->rq_pill, &RMF_FLD_OPC);
 349                 *op = FLD_LOOKUP;
 350
 351                 /*
 352                  * For MDS_MDS seq lookup, it will always use LWP connection,
 353                  * but LWP will be evicted after restart, so cause the error.
 354                  * so we will set no_delay for seq lookup request, once the
 355                  * request fails because of the eviction. always retry here
 356                  */
 357                 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
 358                         req->rq_allow_replay = 1;
 359                         req->rq_no_delay = 1;
 360                 }
 361                 break;
 362         case FLD_READ:
 363                 req = ptlrpc_request_alloc_pack(imp, &RQF_FLD_READ,
 364                                                 LUSTRE_MDS_VERSION, FLD_READ);
 365                 if (!req)
 366                         RETURN(-ENOMEM);
 367
 368                 req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA,
 369                                      RCL_SERVER, PAGE_SIZE);
 370                 break;
 371         default:
 372                 rc = -EINVAL;
 373                 break;
 374         }
 375
 376         if (rc != 0)
 377                 RETURN(rc);
 378
 379         prange = req_capsule_client_get(&req->rq_pill, &RMF_FLD_MDFLD);
 380         *prange = *range;
 381         ptlrpc_request_set_replen(req);
 382         req->rq_request_portal = FLD_REQUEST_PORTAL;
 383         req->rq_reply_portal = MDC_REPLY_PORTAL;
 384         ptlrpc_at_set_req_timeout(req);
 385
 386         if (OBD_FAIL_CHECK(OBD_FAIL_FLD_QUERY_REQ && req->rq_no_delay)) {
 387                 /* the same error returned by ptlrpc_import_delay_req */
 388                 rc = -EAGAIN;
 389                 req->rq_status = rc;
 390         } else {
 391                 obd_get_request_slot(&exp->exp_obd->u.cli);
 392                 rc = ptlrpc_queue_wait(req);
 393                 obd_put_request_slot(&exp->exp_obd->u.cli);
 394         }
 395
 396         if (rc == -ENOENT) {
 397                 /* Don't loop forever on non-existing FID sequences. */
 398                 GOTO(out_req, rc);
 399         }
 400
 401         if (rc != 0) {
 402                 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 403                     !imp->imp_deactive &&
 404                     imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS &&
 405                     OCD_HAS_FLAG(&imp->imp_connect_data, LIGHTWEIGHT) &&
 406                     rc != -ENOTSUPP) {
 407                         /*
 408                          * Since LWP is not replayable, so notify the caller
 409                          * to retry if needed after a while.
 410                          */
 411                         rc = -EAGAIN;
 412                 }
 413                 GOTO(out_req, rc);
 414         }
 415
 416         if (fld_op == FLD_QUERY) {
 417                 prange = req_capsule_server_get(&req->rq_pill,
 418                                                 &RMF_FLD_MDFLD);
 419                 if (!prange)
 420                         GOTO(out_req, rc = -EFAULT);
 421                 *range = *prange;
 422         }
 423
 424         EXIT;
 425 out_req:
 426         if (rc != 0 || !reqp) {
 427                 ptlrpc_req_finished(req);
 428                 req = NULL;
 429         }
 430
 431         if (reqp)
 432                 *reqp = req;
 433
 434         return rc;
 435 }
 436
 437 int fld_client_lookup(struct lu_client_fld *fld, u64 seq, u32 *mds,
 438                       u32 flags, const struct lu_env *env)
 439 {
 440         struct lu_seq_range res = { 0 };
 441         struct lu_fld_target *target;
 442         struct lu_fld_target *origin;
 443         int rc;
 444
 445         ENTRY;
 446
 447         rc = fld_cache_lookup(fld->lcf_cache, seq, &res);
 448         if (rc == 0) {
 449                 *mds = res.lsr_index;
 450                 RETURN(0);
 451         }
 452
 453         /* Can not find it in the cache */
 454         target = fld_client_get_target(fld, seq);
 455         LASSERT(target != NULL);
 456         origin = target;
 457 again:
 458         CDEBUG(D_INFO, "%s: Lookup fld entry (seq: %#llx) on target %s (idx %llu)\n",
 459                fld->lcf_name, seq, fld_target_name(target), target->ft_idx);
 460
 461         res.lsr_start = seq;
 462         fld_range_set_type(&res, flags);
 463
 464 #ifdef HAVE_SERVER_SUPPORT
 465         if (target->ft_srv) {
 466                 LASSERT(env != NULL);
 467                 rc = fld_server_lookup(env, target->ft_srv, seq, &res);
 468         } else
 469 #endif /* HAVE_SERVER_SUPPORT */
 470         {
 471                 rc = fld_client_rpc(target->ft_exp, &res, FLD_QUERY, NULL);
 472         }
 473
 474         if (rc == -ESHUTDOWN) {
 475                 /*
 476                  * If fld lookup failed because the target has been shutdown,
 477                  * then try next target in the list, until trying all targets
 478                  * or fld lookup succeeds
 479                  */
 480                 spin_lock(&fld->lcf_lock);
 481                 /*
 482                  * If the next entry in the list is the head of the list,
 483                  * move to the next entry after the head and retrieve
 484                  * the target. Else retreive the next target entry.
 485                  */
 486                 if (target->ft_chain.next == &fld->lcf_targets)
 487                         target = list_entry(target->ft_chain.next->next,
 488                                             struct lu_fld_target, ft_chain);
 489                 else
 490                         target = list_entry(target->ft_chain.next,
 491                                                  struct lu_fld_target,
 492                                                  ft_chain);
 493                 spin_unlock(&fld->lcf_lock);
 494                 if (target != origin)
 495                         goto again;
 496         }
 497         if (rc == 0) {
 498                 *mds = res.lsr_index;
 499                 fld_cache_insert(fld->lcf_cache, &res);
 500         }
 501
 502         RETURN(rc);
 503 }
 504 EXPORT_SYMBOL(fld_client_lookup);
 505
 506 void fld_client_flush(struct lu_client_fld *fld)
 507 {
 508         fld_cache_flush(fld->lcf_cache);
 509 }
 510
 511 static int __init fld_init(void)
 512 {
 513 #ifdef HAVE_SERVER_SUPPORT
 514         int rc;
 515
 516         rc = fld_server_mod_init();
 517         if (rc)
 518                 return rc;
 519 #endif /* HAVE_SERVER_SUPPORT */
 520
 521         fld_debugfs_dir = debugfs_create_dir(LUSTRE_FLD_NAME,
 522                                              debugfs_lustre_root);
 523         return PTR_ERR_OR_ZERO(fld_debugfs_dir);
 524 }
 525
 526 static void __exit fld_exit(void)
 527 {
 528 #ifdef HAVE_SERVER_SUPPORT
 529         fld_server_mod_exit();
 530 #endif /* HAVE_SERVER_SUPPORT */
 531
 532         debugfs_remove_recursive(fld_debugfs_dir);
 533 }
 534
 535 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
 536 MODULE_DESCRIPTION("Lustre FID Location Database");
 537 MODULE_VERSION(LUSTRE_VERSION_STRING);
 538 MODULE_LICENSE("GPL");
 539
 540 module_init(fld_init);
 541 module_exit(fld_exit);