lustre/lov/lov_object.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) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2014, 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  * Implementation of cl_object for LOV layer.
  37  *
  38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  39  *   Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_LOV
  43
  44 #include "lov_cl_internal.h"
  45
  46 /** \addtogroup lov
  47  *  @{
  48  */
  49
  50 /*****************************************************************************
  51  *
  52  * Layout operations.
  53  *
  54  */
  55
  56 struct lov_layout_operations {
  57         int (*llo_init)(const struct lu_env *env, struct lov_device *dev,
  58                         struct lov_object *lov,
  59                         const struct cl_object_conf *conf,
  60                         union lov_layout_state *state);
  61         int (*llo_delete)(const struct lu_env *env, struct lov_object *lov,
  62                            union lov_layout_state *state);
  63         void (*llo_fini)(const struct lu_env *env, struct lov_object *lov,
  64                          union lov_layout_state *state);
  65         void (*llo_install)(const struct lu_env *env, struct lov_object *lov,
  66                             union lov_layout_state *state);
  67         int  (*llo_print)(const struct lu_env *env, void *cookie,
  68                           lu_printer_t p, const struct lu_object *o);
  69         int  (*llo_page_init)(const struct lu_env *env, struct cl_object *obj,
  70                               struct cl_page *page, pgoff_t index);
  71         int  (*llo_lock_init)(const struct lu_env *env,
  72                               struct cl_object *obj, struct cl_lock *lock,
  73                               const struct cl_io *io);
  74         int  (*llo_io_init)(const struct lu_env *env,
  75                             struct cl_object *obj, struct cl_io *io);
  76         int  (*llo_getattr)(const struct lu_env *env, struct cl_object *obj,
  77                             struct cl_attr *attr);
  78         int  (*llo_find_cbdata)(const struct lu_env *env, struct cl_object *obj,
  79                                 ldlm_iterator_t iter, void *data);
  80 };
  81
  82 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov);
  83
  84 struct lov_stripe_md *lov_lsm_get(struct cl_object *clobj)
  85 {
  86         struct lu_object *luobj;
  87         struct lov_stripe_md *lsm = NULL;
  88
  89         if (clobj == NULL)
  90                 return NULL;
  91
  92         luobj = lu_object_locate(&cl_object_header(clobj)->coh_lu,
  93                                  &lov_device_type);
  94         if (luobj != NULL)
  95                 lsm = lov_lsm_addref(lu2lov(luobj));
  96         return lsm;
  97 }
  98 EXPORT_SYMBOL(lov_lsm_get);
  99
 100 void lov_lsm_put(struct cl_object *unused, struct lov_stripe_md *lsm)
 101 {
 102         if (lsm != NULL)
 103                 lov_free_memmd(&lsm);
 104 }
 105 EXPORT_SYMBOL(lov_lsm_put);
 106
 107 /*****************************************************************************
 108  *
 109  * Lov object layout operations.
 110  *
 111  */
 112
 113 static void lov_install_empty(const struct lu_env *env,
 114                               struct lov_object *lov,
 115                               union  lov_layout_state *state)
 116 {
 117         /*
 118          * File without objects.
 119          */
 120 }
 121
 122 static int lov_init_empty(const struct lu_env *env,
 123                           struct lov_device *dev, struct lov_object *lov,
 124                           const struct cl_object_conf *conf,
 125                           union  lov_layout_state *state)
 126 {
 127         return 0;
 128 }
 129
 130 static void lov_install_raid0(const struct lu_env *env,
 131                               struct lov_object *lov,
 132                               union  lov_layout_state *state)
 133 {
 134 }
 135
 136 static struct cl_object *lov_sub_find(const struct lu_env *env,
 137                                       struct cl_device *dev,
 138                                       const struct lu_fid *fid,
 139                                       const struct cl_object_conf *conf)
 140 {
 141         struct lu_object *o;
 142
 143         ENTRY;
 144         o = lu_object_find_at(env, cl2lu_dev(dev), fid, &conf->coc_lu);
 145         LASSERT(ergo(!IS_ERR(o), o->lo_dev->ld_type == &lovsub_device_type));
 146         RETURN(lu2cl(o));
 147 }
 148
 149 static int lov_init_sub(const struct lu_env *env, struct lov_object *lov,
 150                         struct cl_object *stripe, struct lov_layout_raid0 *r0,
 151                         int idx)
 152 {
 153         struct cl_object_header *hdr;
 154         struct cl_object_header *subhdr;
 155         struct cl_object_header *parent;
 156         struct lov_oinfo        *oinfo;
 157         int result;
 158
 159         if (OBD_FAIL_CHECK(OBD_FAIL_LOV_INIT)) {
 160                 /* For sanity:test_206.
 161                  * Do not leave the object in cache to avoid accessing
 162                  * freed memory. This is because osc_object is referring to
 163                  * lov_oinfo of lsm_stripe_data which will be freed due to
 164                  * this failure. */
 165                 cl_object_kill(env, stripe);
 166                 cl_object_put(env, stripe);
 167                 return -EIO;
 168         }
 169
 170         hdr    = cl_object_header(lov2cl(lov));
 171         subhdr = cl_object_header(stripe);
 172
 173         oinfo = lov->lo_lsm->lsm_oinfo[idx];
 174         CDEBUG(D_INODE, DFID"@%p[%d] -> "DFID"@%p: ostid: "DOSTID
 175                " idx: %d gen: %d\n",
 176                PFID(&subhdr->coh_lu.loh_fid), subhdr, idx,
 177                PFID(&hdr->coh_lu.loh_fid), hdr, POSTID(&oinfo->loi_oi),
 178                oinfo->loi_ost_idx, oinfo->loi_ost_gen);
 179
 180         /* reuse ->coh_attr_guard to protect coh_parent change */
 181         spin_lock(&subhdr->coh_attr_guard);
 182         parent = subhdr->coh_parent;
 183         if (parent == NULL) {
 184                 subhdr->coh_parent = hdr;
 185                 spin_unlock(&subhdr->coh_attr_guard);
 186                 subhdr->coh_nesting = hdr->coh_nesting + 1;
 187                 lu_object_ref_add(&stripe->co_lu, "lov-parent", lov);
 188                 r0->lo_sub[idx] = cl2lovsub(stripe);
 189                 r0->lo_sub[idx]->lso_super = lov;
 190                 r0->lo_sub[idx]->lso_index = idx;
 191                 result = 0;
 192         } else {
 193                 struct lu_object  *old_obj;
 194                 struct lov_object *old_lov;
 195                 unsigned int mask = D_INODE;
 196
 197                 spin_unlock(&subhdr->coh_attr_guard);
 198                 old_obj = lu_object_locate(&parent->coh_lu, &lov_device_type);
 199                 LASSERT(old_obj != NULL);
 200                 old_lov = cl2lov(lu2cl(old_obj));
 201                 if (old_lov->lo_layout_invalid) {
 202                         /* the object's layout has already changed but isn't
 203                          * refreshed */
 204                         lu_object_unhash(env, &stripe->co_lu);
 205                         result = -EAGAIN;
 206                 } else {
 207                         mask = D_ERROR;
 208                         result = -EIO;
 209                 }
 210
 211                 LU_OBJECT_DEBUG(mask, env, &stripe->co_lu,
 212                                 "stripe %d is already owned.\n", idx);
 213                 LU_OBJECT_DEBUG(mask, env, old_obj, "owned.\n");
 214                 LU_OBJECT_HEADER(mask, env, lov2lu(lov), "try to own.\n");
 215                 cl_object_put(env, stripe);
 216         }
 217         return result;
 218 }
 219
 220 static int lov_page_slice_fixup(struct lov_object *lov,
 221                                 struct cl_object *stripe)
 222 {
 223         struct cl_object_header *hdr = cl_object_header(&lov->lo_cl);
 224         struct cl_object *o;
 225
 226         if (stripe == NULL)
 227                 return hdr->coh_page_bufsize - lov->lo_cl.co_slice_off -
 228                        cfs_size_round(sizeof(struct lov_page));
 229
 230         cl_object_for_each(o, stripe)
 231                 o->co_slice_off += hdr->coh_page_bufsize;
 232
 233         return cl_object_header(stripe)->coh_page_bufsize;
 234 }
 235
 236 static int lov_init_raid0(const struct lu_env *env,
 237                           struct lov_device *dev, struct lov_object *lov,
 238                           const struct cl_object_conf *conf,
 239                           union  lov_layout_state *state)
 240 {
 241         int result;
 242         int i;
 243
 244         struct cl_object        *stripe;
 245         struct lov_thread_info  *lti     = lov_env_info(env);
 246         struct cl_object_conf   *subconf = &lti->lti_stripe_conf;
 247         struct lov_stripe_md    *lsm     = conf->u.coc_md->lsm;
 248         struct lu_fid           *ofid    = &lti->lti_fid;
 249         struct lov_layout_raid0 *r0      = &state->raid0;
 250
 251         ENTRY;
 252
 253         if (lsm->lsm_magic != LOV_MAGIC_V1 && lsm->lsm_magic != LOV_MAGIC_V3) {
 254                 dump_lsm(D_ERROR, lsm);
 255                 LASSERTF(0, "magic mismatch, expected %d/%d, actual %d.\n",
 256                          LOV_MAGIC_V1, LOV_MAGIC_V3, lsm->lsm_magic);
 257         }
 258
 259         LASSERT(lov->lo_lsm == NULL);
 260         lov->lo_lsm = lsm_addref(lsm);
 261         r0->lo_nr  = lsm->lsm_stripe_count;
 262         LASSERT(r0->lo_nr <= lov_targets_nr(dev));
 263
 264         lov->lo_layout_invalid = true;
 265
 266         OBD_ALLOC_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 267         if (r0->lo_sub != NULL) {
 268                 int psz = 0;
 269
 270                 result = 0;
 271                 subconf->coc_inode = conf->coc_inode;
 272                 spin_lock_init(&r0->lo_sub_lock);
 273                 /*
 274                  * Create stripe cl_objects.
 275                  */
 276                 for (i = 0; i < r0->lo_nr && result == 0; ++i) {
 277                         struct cl_device *subdev;
 278                         struct lov_oinfo *oinfo = lsm->lsm_oinfo[i];
 279                         int ost_idx = oinfo->loi_ost_idx;
 280
 281                         if (lov_oinfo_is_dummy(oinfo))
 282                                 continue;
 283
 284                         result = ostid_to_fid(ofid, &oinfo->loi_oi,
 285                                               oinfo->loi_ost_idx);
 286                         if (result != 0)
 287                                 GOTO(out, result);
 288
 289                         subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
 290                         subconf->u.coc_oinfo = oinfo;
 291                         LASSERTF(subdev != NULL, "not init ost %d\n", ost_idx);
 292                         /* In the function below, .hs_keycmp resolves to
 293                          * lu_obj_hop_keycmp() */
 294                         /* coverity[overrun-buffer-val] */
 295                         stripe = lov_sub_find(env, subdev, ofid, subconf);
 296                         if (!IS_ERR(stripe)) {
 297                                 result = lov_init_sub(env, lov, stripe, r0, i);
 298                                 if (result == -EAGAIN) { /* try again */
 299                                         --i;
 300                                         result = 0;
 301                                         continue;
 302                                 }
 303                         } else {
 304                                 result = PTR_ERR(stripe);
 305                         }
 306
 307                         if (result == 0) {
 308                                 int sz = lov_page_slice_fixup(lov, stripe);
 309                                 LASSERT(ergo(psz > 0, psz == sz));
 310                                 psz = sz;
 311                         }
 312                 }
 313                 if (result == 0)
 314                         cl_object_header(&lov->lo_cl)->coh_page_bufsize += psz;
 315         } else
 316                 result = -ENOMEM;
 317 out:
 318         RETURN(result);
 319 }
 320
 321 static int lov_init_released(const struct lu_env *env,
 322                         struct lov_device *dev, struct lov_object *lov,
 323                         const struct cl_object_conf *conf,
 324                         union  lov_layout_state *state)
 325 {
 326         struct lov_stripe_md *lsm = conf->u.coc_md->lsm;
 327
 328         LASSERT(lsm != NULL);
 329         LASSERT(lsm_is_released(lsm));
 330         LASSERT(lov->lo_lsm == NULL);
 331
 332         lov->lo_lsm = lsm_addref(lsm);
 333         return 0;
 334 }
 335
 336 static struct cl_object *lov_find_subobj(const struct lu_env *env,
 337                                          struct lov_object *lov,
 338                                          struct lov_stripe_md *lsm,
 339                                          int stripe_idx)
 340 {
 341         struct lov_device       *dev = lu2lov_dev(lov2lu(lov)->lo_dev);
 342         struct lov_oinfo        *oinfo = lsm->lsm_oinfo[stripe_idx];
 343         struct lov_thread_info  *lti = lov_env_info(env);
 344         struct lu_fid           *ofid = &lti->lti_fid;
 345         struct cl_device        *subdev;
 346         int                     ost_idx;
 347         int                     rc;
 348         struct cl_object        *result;
 349
 350         if (lov->lo_type != LLT_RAID0)
 351                 GOTO(out, result = NULL);
 352
 353         ost_idx = oinfo->loi_ost_idx;
 354         rc = ostid_to_fid(ofid, &oinfo->loi_oi, ost_idx);
 355         if (rc != 0)
 356                 GOTO(out, result = NULL);
 357
 358         subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
 359         result = lov_sub_find(env, subdev, ofid, NULL);
 360 out:
 361         if (result == NULL)
 362                 result = ERR_PTR(-EINVAL);
 363         return result;
 364 }
 365
 366 static int lov_delete_empty(const struct lu_env *env, struct lov_object *lov,
 367                             union lov_layout_state *state)
 368 {
 369         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 370
 371         lov_layout_wait(env, lov);
 372         return 0;
 373 }
 374
 375 static void lov_subobject_kill(const struct lu_env *env, struct lov_object *lov,
 376                                struct lovsub_object *los, int idx)
 377 {
 378         struct cl_object        *sub;
 379         struct lov_layout_raid0 *r0;
 380         struct lu_site          *site;
 381         struct lu_site_bkt_data *bkt;
 382         wait_queue_t          *waiter;
 383
 384         r0  = &lov->u.raid0;
 385         LASSERT(r0->lo_sub[idx] == los);
 386
 387         sub  = lovsub2cl(los);
 388         site = sub->co_lu.lo_dev->ld_site;
 389         bkt  = lu_site_bkt_from_fid(site, &sub->co_lu.lo_header->loh_fid);
 390
 391         cl_object_kill(env, sub);
 392         /* release a reference to the sub-object and ... */
 393         lu_object_ref_del(&sub->co_lu, "lov-parent", lov);
 394         cl_object_put(env, sub);
 395
 396         /* ... wait until it is actually destroyed---sub-object clears its
 397          * ->lo_sub[] slot in lovsub_object_fini() */
 398         if (r0->lo_sub[idx] == los) {
 399                 waiter = &lov_env_info(env)->lti_waiter;
 400                 init_waitqueue_entry_current(waiter);
 401                 add_wait_queue(&bkt->lsb_marche_funebre, waiter);
 402                 set_current_state(TASK_UNINTERRUPTIBLE);
 403                 while (1) {
 404                         /* this wait-queue is signaled at the end of
 405                          * lu_object_free(). */
 406                         set_current_state(TASK_UNINTERRUPTIBLE);
 407                         spin_lock(&r0->lo_sub_lock);
 408                         if (r0->lo_sub[idx] == los) {
 409                                 spin_unlock(&r0->lo_sub_lock);
 410                                 waitq_wait(waiter, TASK_UNINTERRUPTIBLE);
 411                         } else {
 412                                 spin_unlock(&r0->lo_sub_lock);
 413                                 set_current_state(TASK_RUNNING);
 414                                 break;
 415                         }
 416                 }
 417                 remove_wait_queue(&bkt->lsb_marche_funebre, waiter);
 418         }
 419         LASSERT(r0->lo_sub[idx] == NULL);
 420 }
 421
 422 static int lov_delete_raid0(const struct lu_env *env, struct lov_object *lov,
 423                             union lov_layout_state *state)
 424 {
 425         struct lov_layout_raid0 *r0 = &state->raid0;
 426         struct lov_stripe_md    *lsm = lov->lo_lsm;
 427         int i;
 428
 429         ENTRY;
 430
 431         dump_lsm(D_INODE, lsm);
 432
 433         lov_layout_wait(env, lov);
 434         if (r0->lo_sub != NULL) {
 435                 for (i = 0; i < r0->lo_nr; ++i) {
 436                         struct lovsub_object *los = r0->lo_sub[i];
 437
 438                         if (los != NULL) {
 439                                 cl_object_prune(env, &los->lso_cl);
 440                                 /*
 441                                  * If top-level object is to be evicted from
 442                                  * the cache, so are its sub-objects.
 443                                  */
 444                                 lov_subobject_kill(env, lov, los, i);
 445                         }
 446                 }
 447         }
 448         RETURN(0);
 449 }
 450
 451 static void lov_fini_empty(const struct lu_env *env, struct lov_object *lov,
 452                            union lov_layout_state *state)
 453 {
 454         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 455 }
 456
 457 static void lov_fini_raid0(const struct lu_env *env, struct lov_object *lov,
 458                            union lov_layout_state *state)
 459 {
 460         struct lov_layout_raid0 *r0 = &state->raid0;
 461         ENTRY;
 462
 463         if (r0->lo_sub != NULL) {
 464                 OBD_FREE_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 465                 r0->lo_sub = NULL;
 466         }
 467
 468         dump_lsm(D_INODE, lov->lo_lsm);
 469         lov_free_memmd(&lov->lo_lsm);
 470
 471         EXIT;
 472 }
 473
 474 static void lov_fini_released(const struct lu_env *env, struct lov_object *lov,
 475                                 union lov_layout_state *state)
 476 {
 477         ENTRY;
 478         dump_lsm(D_INODE, lov->lo_lsm);
 479         lov_free_memmd(&lov->lo_lsm);
 480         EXIT;
 481 }
 482
 483 static int lov_print_empty(const struct lu_env *env, void *cookie,
 484                            lu_printer_t p, const struct lu_object *o)
 485 {
 486         (*p)(env, cookie, "empty %d\n", lu2lov(o)->lo_layout_invalid);
 487         return 0;
 488 }
 489
 490 static int lov_print_raid0(const struct lu_env *env, void *cookie,
 491                            lu_printer_t p, const struct lu_object *o)
 492 {
 493         struct lov_object       *lov = lu2lov(o);
 494         struct lov_layout_raid0 *r0  = lov_r0(lov);
 495         struct lov_stripe_md    *lsm = lov->lo_lsm;
 496         int                      i;
 497
 498         (*p)(env, cookie, "stripes: %d, %s, lsm{%p 0x%08X %d %u %u}:\n",
 499                 r0->lo_nr, lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 500                 lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 501                 lsm->lsm_stripe_count, lsm->lsm_layout_gen);
 502         for (i = 0; i < r0->lo_nr; ++i) {
 503                 struct lu_object *sub;
 504
 505                 if (r0->lo_sub[i] != NULL) {
 506                         sub = lovsub2lu(r0->lo_sub[i]);
 507                         lu_object_print(env, cookie, p, sub);
 508                 } else {
 509                         (*p)(env, cookie, "sub %d absent\n", i);
 510                 }
 511         }
 512         return 0;
 513 }
 514
 515 static int lov_print_released(const struct lu_env *env, void *cookie,
 516                                 lu_printer_t p, const struct lu_object *o)
 517 {
 518         struct lov_object       *lov = lu2lov(o);
 519         struct lov_stripe_md    *lsm = lov->lo_lsm;
 520
 521         (*p)(env, cookie,
 522                 "released: %s, lsm{%p 0x%08X %d %u %u}:\n",
 523                 lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 524                 lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 525                 lsm->lsm_stripe_count, lsm->lsm_layout_gen);
 526         return 0;
 527 }
 528
 529 /**
 530  * Implements cl_object_operations::coo_attr_get() method for an object
 531  * without stripes (LLT_EMPTY layout type).
 532  *
 533  * The only attributes this layer is authoritative in this case is
 534  * cl_attr::cat_blocks---it's 0.
 535  */
 536 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
 537                               struct cl_attr *attr)
 538 {
 539         attr->cat_blocks = 0;
 540         return 0;
 541 }
 542
 543 static int lov_attr_get_raid0(const struct lu_env *env, struct cl_object *obj,
 544                               struct cl_attr *attr)
 545 {
 546         struct lov_object       *lov = cl2lov(obj);
 547         struct lov_layout_raid0 *r0 = lov_r0(lov);
 548         struct cl_attr          *lov_attr = &r0->lo_attr;
 549         int                      result = 0;
 550
 551         ENTRY;
 552
 553         /* this is called w/o holding type guard mutex, so it must be inside
 554          * an on going IO otherwise lsm may be replaced.
 555          * LU-2117: it turns out there exists one exception. For mmaped files,
 556          * the lock of those files may be requested in the other file's IO
 557          * context, and this function is called in ccc_lock_state(), it will
 558          * hit this assertion.
 559          * Anyway, it's still okay to call attr_get w/o type guard as layout
 560          * can't go if locks exist. */
 561         /* LASSERT(atomic_read(&lsm->lsm_refc) > 1); */
 562
 563         if (!r0->lo_attr_valid) {
 564                 struct lov_stripe_md    *lsm = lov->lo_lsm;
 565                 struct ost_lvb          *lvb = &lov_env_info(env)->lti_lvb;
 566                 __u64                    kms = 0;
 567
 568                 memset(lvb, 0, sizeof(*lvb));
 569                 /* XXX: timestamps can be negative by sanity:test_39m,
 570                  * how can it be? */
 571                 lvb->lvb_atime = LLONG_MIN;
 572                 lvb->lvb_ctime = LLONG_MIN;
 573                 lvb->lvb_mtime = LLONG_MIN;
 574
 575                 /*
 576                  * XXX that should be replaced with a loop over sub-objects,
 577                  * doing cl_object_attr_get() on them. But for now, let's
 578                  * reuse old lov code.
 579                  */
 580
 581                 /*
 582                  * XXX take lsm spin-lock to keep lov_merge_lvb_kms()
 583                  * happy. It's not needed, because new code uses
 584                  * ->coh_attr_guard spin-lock to protect consistency of
 585                  * sub-object attributes.
 586                  */
 587                 lov_stripe_lock(lsm);
 588                 result = lov_merge_lvb_kms(lsm, lvb, &kms);
 589                 lov_stripe_unlock(lsm);
 590                 if (result == 0) {
 591                         cl_lvb2attr(lov_attr, lvb);
 592                         lov_attr->cat_kms = kms;
 593                         r0->lo_attr_valid = 1;
 594                 }
 595         }
 596         if (result == 0) { /* merge results */
 597                 attr->cat_blocks = lov_attr->cat_blocks;
 598                 attr->cat_size = lov_attr->cat_size;
 599                 attr->cat_kms = lov_attr->cat_kms;
 600                 if (attr->cat_atime < lov_attr->cat_atime)
 601                         attr->cat_atime = lov_attr->cat_atime;
 602                 if (attr->cat_ctime < lov_attr->cat_ctime)
 603                         attr->cat_ctime = lov_attr->cat_ctime;
 604                 if (attr->cat_mtime < lov_attr->cat_mtime)
 605                         attr->cat_mtime = lov_attr->cat_mtime;
 606         }
 607         RETURN(result);
 608 }
 609
 610 static int lov_find_cbdata_empty(const struct lu_env *env,
 611                                  struct cl_object *obj, ldlm_iterator_t iter,
 612                                  void *data)
 613 {
 614         return 0;
 615 }
 616
 617 static int lov_find_cbdata_raid0(const struct lu_env *env,
 618                                  struct cl_object *obj, ldlm_iterator_t iter,
 619                                  void *data)
 620 {
 621         struct lov_object       *lov = cl2lov(obj);
 622         struct lov_layout_raid0 *r0 = lov_r0(lov);
 623         struct cl_object        *subobj;
 624         int                     i;
 625         int                     rc = 0;
 626
 627         for (i = 0; i < r0->lo_nr; ++i) {
 628                 if (r0->lo_sub[i] == NULL)
 629                         continue;
 630
 631                 subobj = lovsub2cl(r0->lo_sub[i]);
 632
 633                 rc = cl_object_find_cbdata(env, subobj, iter, data);
 634                 if (rc != 0)
 635                         break;
 636         }
 637
 638         return rc;
 639 }
 640
 641 const static struct lov_layout_operations lov_dispatch[] = {
 642         [LLT_EMPTY] = {
 643                 .llo_init      = lov_init_empty,
 644                 .llo_delete    = lov_delete_empty,
 645                 .llo_fini      = lov_fini_empty,
 646                 .llo_install   = lov_install_empty,
 647                 .llo_print     = lov_print_empty,
 648                 .llo_page_init = lov_page_init_empty,
 649                 .llo_lock_init = lov_lock_init_empty,
 650                 .llo_io_init   = lov_io_init_empty,
 651                 .llo_getattr   = lov_attr_get_empty,
 652                 .llo_find_cbdata = lov_find_cbdata_empty
 653         },
 654         [LLT_RAID0] = {
 655                 .llo_init      = lov_init_raid0,
 656                 .llo_delete    = lov_delete_raid0,
 657                 .llo_fini      = lov_fini_raid0,
 658                 .llo_install   = lov_install_raid0,
 659                 .llo_print     = lov_print_raid0,
 660                 .llo_page_init = lov_page_init_raid0,
 661                 .llo_lock_init = lov_lock_init_raid0,
 662                 .llo_io_init   = lov_io_init_raid0,
 663                 .llo_getattr   = lov_attr_get_raid0,
 664                 .llo_find_cbdata = lov_find_cbdata_raid0
 665         },
 666         [LLT_RELEASED] = {
 667                 .llo_init      = lov_init_released,
 668                 .llo_delete    = lov_delete_empty,
 669                 .llo_fini      = lov_fini_released,
 670                 .llo_install   = lov_install_empty,
 671                 .llo_print     = lov_print_released,
 672                 .llo_page_init = lov_page_init_empty,
 673                 .llo_lock_init = lov_lock_init_empty,
 674                 .llo_io_init   = lov_io_init_released,
 675                 .llo_getattr   = lov_attr_get_empty,
 676                 .llo_find_cbdata = lov_find_cbdata_empty
 677         }
 678 };
 679
 680 /**
 681  * Performs a double-dispatch based on the layout type of an object.
 682  */
 683 #define LOV_2DISPATCH_NOLOCK(obj, op, ...)                              \
 684 ({                                                                      \
 685         struct lov_object                      *__obj = (obj);          \
 686         enum lov_layout_type                    __llt;                  \
 687                                                                         \
 688         __llt = __obj->lo_type;                                         \
 689         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 690         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 691 })
 692
 693 /**
 694  * Return lov_layout_type associated with a given lsm
 695  */
 696 static enum lov_layout_type lov_type(struct lov_stripe_md *lsm)
 697 {
 698         if (lsm == NULL)
 699                 return LLT_EMPTY;
 700         if (lsm_is_released(lsm))
 701                 return LLT_RELEASED;
 702         return LLT_RAID0;
 703 }
 704
 705 static inline void lov_conf_freeze(struct lov_object *lov)
 706 {
 707         if (lov->lo_owner != current)
 708                 down_read(&lov->lo_type_guard);
 709 }
 710
 711 static inline void lov_conf_thaw(struct lov_object *lov)
 712 {
 713         if (lov->lo_owner != current)
 714                 up_read(&lov->lo_type_guard);
 715 }
 716
 717 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...)                       \
 718 ({                                                                      \
 719         struct lov_object                      *__obj = (obj);          \
 720         int                                     __lock = !!(lock);      \
 721         typeof(lov_dispatch[0].op(__VA_ARGS__)) __result;               \
 722                                                                         \
 723         if (__lock)                                                     \
 724                 lov_conf_freeze(__obj);                                 \
 725         __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__);          \
 726         if (__lock)                                                     \
 727                 lov_conf_thaw(__obj);                                   \
 728         __result;                                                       \
 729 })
 730
 731 /**
 732  * Performs a locked double-dispatch based on the layout type of an object.
 733  */
 734 #define LOV_2DISPATCH(obj, op, ...)                     \
 735         LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
 736
 737 #define LOV_2DISPATCH_VOID(obj, op, ...)                                \
 738 do {                                                                    \
 739         struct lov_object                      *__obj = (obj);          \
 740         enum lov_layout_type                    __llt;                  \
 741                                                                         \
 742         lov_conf_freeze(__obj);                                         \
 743         __llt = __obj->lo_type;                                         \
 744         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 745         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 746         lov_conf_thaw(__obj);                                           \
 747 } while (0)
 748
 749 static void lov_conf_lock(struct lov_object *lov)
 750 {
 751         LASSERT(lov->lo_owner != current);
 752         down_write(&lov->lo_type_guard);
 753         LASSERT(lov->lo_owner == NULL);
 754         lov->lo_owner = current;
 755 }
 756
 757 static void lov_conf_unlock(struct lov_object *lov)
 758 {
 759         lov->lo_owner = NULL;
 760         up_write(&lov->lo_type_guard);
 761 }
 762
 763 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov)
 764 {
 765         struct l_wait_info lwi = { 0 };
 766         ENTRY;
 767
 768         while (atomic_read(&lov->lo_active_ios) > 0) {
 769                 CDEBUG(D_INODE, "file:"DFID" wait for active IO, now: %d.\n",
 770                         PFID(lu_object_fid(lov2lu(lov))),
 771                         atomic_read(&lov->lo_active_ios));
 772
 773                 l_wait_event(lov->lo_waitq,
 774                              atomic_read(&lov->lo_active_ios) == 0, &lwi);
 775         }
 776         RETURN(0);
 777 }
 778
 779 static int lov_layout_change(const struct lu_env *unused,
 780                              struct lov_object *lov,
 781                              const struct cl_object_conf *conf)
 782 {
 783         int result;
 784         enum lov_layout_type llt = LLT_EMPTY;
 785         union lov_layout_state *state = &lov->u;
 786         const struct lov_layout_operations *old_ops;
 787         const struct lov_layout_operations *new_ops;
 788
 789         void *cookie;
 790         struct lu_env *env;
 791         int refcheck;
 792         ENTRY;
 793
 794         LASSERT(0 <= lov->lo_type && lov->lo_type < ARRAY_SIZE(lov_dispatch));
 795
 796         if (conf->u.coc_md != NULL)
 797                 llt = lov_type(conf->u.coc_md->lsm);
 798         LASSERT(0 <= llt && llt < ARRAY_SIZE(lov_dispatch));
 799
 800         cookie = cl_env_reenter();
 801         env = cl_env_get(&refcheck);
 802         if (IS_ERR(env)) {
 803                 cl_env_reexit(cookie);
 804                 RETURN(PTR_ERR(env));
 805         }
 806
 807         CDEBUG(D_INODE, DFID" from %s to %s\n",
 808                PFID(lu_object_fid(lov2lu(lov))),
 809                llt2str(lov->lo_type), llt2str(llt));
 810
 811         old_ops = &lov_dispatch[lov->lo_type];
 812         new_ops = &lov_dispatch[llt];
 813
 814         result = cl_object_prune(env, &lov->lo_cl);
 815         if (result != 0)
 816                 GOTO(out, result);
 817
 818         result = old_ops->llo_delete(env, lov, &lov->u);
 819         if (result == 0) {
 820                 old_ops->llo_fini(env, lov, &lov->u);
 821
 822                 LASSERT(atomic_read(&lov->lo_active_ios) == 0);
 823
 824                 lov->lo_type = LLT_EMPTY;
 825                 /* page bufsize fixup */
 826                 cl_object_header(&lov->lo_cl)->coh_page_bufsize -=
 827                         lov_page_slice_fixup(lov, NULL);
 828
 829                 result = new_ops->llo_init(env,
 830                                         lu2lov_dev(lov->lo_cl.co_lu.lo_dev),
 831                                         lov, conf, state);
 832                 if (result == 0) {
 833                         new_ops->llo_install(env, lov, state);
 834                         lov->lo_type = llt;
 835                 } else {
 836                         new_ops->llo_delete(env, lov, state);
 837                         new_ops->llo_fini(env, lov, state);
 838                         /* this file becomes an EMPTY file. */
 839                 }
 840         }
 841
 842 out:
 843         cl_env_put(env, &refcheck);
 844         cl_env_reexit(cookie);
 845         RETURN(result);
 846 }
 847
 848 /*****************************************************************************
 849  *
 850  * Lov object operations.
 851  *
 852  */
 853 int lov_object_init(const struct lu_env *env, struct lu_object *obj,
 854                     const struct lu_object_conf *conf)
 855 {
 856         struct lov_device            *dev   = lu2lov_dev(obj->lo_dev);
 857         struct lov_object            *lov   = lu2lov(obj);
 858         const struct cl_object_conf  *cconf = lu2cl_conf(conf);
 859         union  lov_layout_state      *set   = &lov->u;
 860         const struct lov_layout_operations *ops;
 861         int result;
 862
 863         ENTRY;
 864         init_rwsem(&lov->lo_type_guard);
 865         atomic_set(&lov->lo_active_ios, 0);
 866         init_waitqueue_head(&lov->lo_waitq);
 867
 868         cl_object_page_init(lu2cl(obj), sizeof(struct lov_page));
 869
 870         /* no locking is necessary, as object is being created */
 871         lov->lo_type = lov_type(cconf->u.coc_md->lsm);
 872         ops = &lov_dispatch[lov->lo_type];
 873         result = ops->llo_init(env, dev, lov, cconf, set);
 874         if (result == 0)
 875                 ops->llo_install(env, lov, set);
 876         RETURN(result);
 877 }
 878
 879 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
 880                         const struct cl_object_conf *conf)
 881 {
 882         struct lov_stripe_md    *lsm = NULL;
 883         struct lov_object       *lov = cl2lov(obj);
 884         int                      result = 0;
 885         ENTRY;
 886
 887         lov_conf_lock(lov);
 888         if (conf->coc_opc == OBJECT_CONF_INVALIDATE) {
 889                 lov->lo_layout_invalid = true;
 890                 GOTO(out, result = 0);
 891         }
 892
 893         if (conf->coc_opc == OBJECT_CONF_WAIT) {
 894                 if (lov->lo_layout_invalid &&
 895                     atomic_read(&lov->lo_active_ios) > 0) {
 896                         lov_conf_unlock(lov);
 897                         result = lov_layout_wait(env, lov);
 898                         lov_conf_lock(lov);
 899                 }
 900                 GOTO(out, result);
 901         }
 902
 903         LASSERT(conf->coc_opc == OBJECT_CONF_SET);
 904
 905         if (conf->u.coc_md != NULL)
 906                 lsm = conf->u.coc_md->lsm;
 907         if ((lsm == NULL && lov->lo_lsm == NULL) ||
 908             ((lsm != NULL && lov->lo_lsm != NULL) &&
 909              (lov->lo_lsm->lsm_layout_gen == lsm->lsm_layout_gen) &&
 910              (lov->lo_lsm->lsm_pattern == lsm->lsm_pattern))) {
 911                 /* same version of layout */
 912                 lov->lo_layout_invalid = false;
 913                 GOTO(out, result = 0);
 914         }
 915
 916         /* will change layout - check if there still exists active IO. */
 917         if (atomic_read(&lov->lo_active_ios) > 0) {
 918                 lov->lo_layout_invalid = true;
 919                 GOTO(out, result = -EBUSY);
 920         }
 921
 922         result = lov_layout_change(env, lov, conf);
 923         lov->lo_layout_invalid = result != 0;
 924         EXIT;
 925
 926 out:
 927         lov_conf_unlock(lov);
 928         CDEBUG(D_INODE, DFID" lo_layout_invalid=%d\n",
 929                PFID(lu_object_fid(lov2lu(lov))), lov->lo_layout_invalid);
 930         RETURN(result);
 931 }
 932
 933 static void lov_object_delete(const struct lu_env *env, struct lu_object *obj)
 934 {
 935         struct lov_object *lov = lu2lov(obj);
 936
 937         ENTRY;
 938         LOV_2DISPATCH_VOID(lov, llo_delete, env, lov, &lov->u);
 939         EXIT;
 940 }
 941
 942 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
 943 {
 944         struct lov_object *lov = lu2lov(obj);
 945
 946         ENTRY;
 947         LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
 948         lu_object_fini(obj);
 949         OBD_SLAB_FREE_PTR(lov, lov_object_kmem);
 950         EXIT;
 951 }
 952
 953 static int lov_object_print(const struct lu_env *env, void *cookie,
 954                             lu_printer_t p, const struct lu_object *o)
 955 {
 956         return LOV_2DISPATCH_NOLOCK(lu2lov(o), llo_print, env, cookie, p, o);
 957 }
 958
 959 int lov_page_init(const struct lu_env *env, struct cl_object *obj,
 960                   struct cl_page *page, pgoff_t index)
 961 {
 962         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_page_init, env, obj, page,
 963                                     index);
 964 }
 965
 966 /**
 967  * Implements cl_object_operations::clo_io_init() method for lov
 968  * layer. Dispatches to the appropriate layout io initialization method.
 969  */
 970 int lov_io_init(const struct lu_env *env, struct cl_object *obj,
 971                 struct cl_io *io)
 972 {
 973         CL_IO_SLICE_CLEAN(lov_env_io(env), lis_cl);
 974         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
 975                                      !io->ci_ignore_layout, env, obj, io);
 976 }
 977
 978 /**
 979  * An implementation of cl_object_operations::clo_attr_get() method for lov
 980  * layer. For raid0 layout this collects and merges attributes of all
 981  * sub-objects.
 982  */
 983 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
 984                         struct cl_attr *attr)
 985 {
 986         /* do not take lock, as this function is called under a
 987          * spin-lock. Layout is protected from changing by ongoing IO. */
 988         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
 989 }
 990
 991 static int lov_attr_update(const struct lu_env *env, struct cl_object *obj,
 992                            const struct cl_attr *attr, unsigned valid)
 993 {
 994         /*
 995          * No dispatch is required here, as no layout implements this.
 996          */
 997         return 0;
 998 }
 999
1000 int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
1001                   struct cl_lock *lock, const struct cl_io *io)
1002 {
1003         /* No need to lock because we've taken one refcount of layout.  */
1004         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_lock_init, env, obj, lock,
1005                                     io);
1006 }
1007
1008 /**
1009  * We calculate on which OST the mapping will end. If the length of mapping
1010  * is greater than (stripe_size * stripe_count) then the last_stripe will
1011  * will be one just before start_stripe. Else we check if the mapping
1012  * intersects each OST and find last_stripe.
1013  * This function returns the last_stripe and also sets the stripe_count
1014  * over which the mapping is spread
1015  *
1016  * \param lsm [in]              striping information for the file
1017  * \param fm_start [in]         logical start of mapping
1018  * \param fm_end [in]           logical end of mapping
1019  * \param start_stripe [in]     starting stripe of the mapping
1020  * \param stripe_count [out]    the number of stripes across which to map is
1021  *                              returned
1022  *
1023  * \retval last_stripe          return the last stripe of the mapping
1024  */
1025 static int fiemap_calc_last_stripe(struct lov_stripe_md *lsm,
1026                                    loff_t fm_start, loff_t fm_end,
1027                                    int start_stripe, int *stripe_count)
1028 {
1029         int last_stripe;
1030         loff_t obd_start;
1031         loff_t obd_end;
1032         int i, j;
1033
1034         if (fm_end - fm_start > lsm->lsm_stripe_size * lsm->lsm_stripe_count) {
1035                 last_stripe = (start_stripe < 1 ? lsm->lsm_stripe_count - 1 :
1036                                                               start_stripe - 1);
1037                 *stripe_count = lsm->lsm_stripe_count;
1038         } else {
1039                 for (j = 0, i = start_stripe; j < lsm->lsm_stripe_count;
1040                      i = (i + 1) % lsm->lsm_stripe_count, j++) {
1041                         if ((lov_stripe_intersects(lsm, i, fm_start, fm_end,
1042                                                    &obd_start, &obd_end)) == 0)
1043                                 break;
1044                 }
1045                 *stripe_count = j;
1046                 last_stripe = (start_stripe + j - 1) % lsm->lsm_stripe_count;
1047         }
1048
1049         return last_stripe;
1050 }
1051
1052 /**
1053  * Set fe_device and copy extents from local buffer into main return buffer.
1054  *
1055  * \param fiemap [out]          fiemap to hold all extents
1056  * \param lcl_fm_ext [in]       array of fiemap extents get from OSC layer
1057  * \param ost_index [in]        OST index to be written into the fm_device
1058  *                              field for each extent
1059  * \param ext_count [in]        number of extents to be copied
1060  * \param current_extent [in]   where to start copying in the extent array
1061  */
1062 static void fiemap_prepare_and_copy_exts(struct fiemap *fiemap,
1063                                          struct ll_fiemap_extent *lcl_fm_ext,
1064                                          int ost_index, unsigned int ext_count,
1065                                          int current_extent)
1066 {
1067         char            *to;
1068         unsigned int    ext;
1069
1070         for (ext = 0; ext < ext_count; ext++) {
1071                 lcl_fm_ext[ext].fe_device = ost_index;
1072                 lcl_fm_ext[ext].fe_flags |= FIEMAP_EXTENT_NET;
1073         }
1074
1075         /* Copy fm_extent's from fm_local to return buffer */
1076         to = (char *)fiemap + fiemap_count_to_size(current_extent);
1077         memcpy(to, lcl_fm_ext, ext_count * sizeof(struct ll_fiemap_extent));
1078 }
1079
1080 #define FIEMAP_BUFFER_SIZE 4096
1081
1082 /**
1083  * Non-zero fe_logical indicates that this is a continuation FIEMAP
1084  * call. The local end offset and the device are sent in the first
1085  * fm_extent. This function calculates the stripe number from the index.
1086  * This function returns a stripe_no on which mapping is to be restarted.
1087  *
1088  * This function returns fm_end_offset which is the in-OST offset at which
1089  * mapping should be restarted. If fm_end_offset=0 is returned then caller
1090  * will re-calculate proper offset in next stripe.
1091  * Note that the first extent is passed to lov_get_info via the value field.
1092  *
1093  * \param fiemap [in]           fiemap request header
1094  * \param lsm [in]              striping information for the file
1095  * \param fm_start [in]         logical start of mapping
1096  * \param fm_end [in]           logical end of mapping
1097  * \param start_stripe [out]    starting stripe will be returned in this
1098  */
1099 static loff_t fiemap_calc_fm_end_offset(struct fiemap *fiemap,
1100                                         struct lov_stripe_md *lsm,
1101                                         loff_t fm_start, loff_t fm_end,
1102                                         int *start_stripe)
1103 {
1104         loff_t local_end = fiemap->fm_extents[0].fe_logical;
1105         loff_t lun_start;
1106         loff_t lun_end;
1107         loff_t fm_end_offset;
1108         int stripe_no = -1;
1109         int i;
1110
1111         if (fiemap->fm_extent_count == 0 ||
1112             fiemap->fm_extents[0].fe_logical == 0)
1113                 return 0;
1114
1115         /* Find out stripe_no from ost_index saved in the fe_device */
1116         for (i = 0; i < lsm->lsm_stripe_count; i++) {
1117                 struct lov_oinfo *oinfo = lsm->lsm_oinfo[i];
1118
1119                 if (lov_oinfo_is_dummy(oinfo))
1120                         continue;
1121
1122                 if (oinfo->loi_ost_idx == fiemap->fm_extents[0].fe_device) {
1123                         stripe_no = i;
1124                         break;
1125                 }
1126         }
1127
1128         if (stripe_no == -1)
1129                 return -EINVAL;
1130
1131         /* If we have finished mapping on previous device, shift logical
1132          * offset to start of next device */
1133         if (lov_stripe_intersects(lsm, stripe_no, fm_start, fm_end,
1134                                   &lun_start, &lun_end) != 0 &&
1135             local_end < lun_end) {
1136                 fm_end_offset = local_end;
1137                 *start_stripe = stripe_no;
1138         } else {
1139                 /* This is a special value to indicate that caller should
1140                  * calculate offset in next stripe. */
1141                 fm_end_offset = 0;
1142                 *start_stripe = (stripe_no + 1) % lsm->lsm_stripe_count;
1143         }
1144
1145         return fm_end_offset;
1146 }
1147
1148 /**
1149  * Break down the FIEMAP request and send appropriate calls to individual OSTs.
1150  * This also handles the restarting of FIEMAP calls in case mapping overflows
1151  * the available number of extents in single call.
1152  *
1153  * \param env [in]              lustre environment
1154  * \param obj [in]              file object
1155  * \param fmkey [in]            fiemap request header and other info
1156  * \param fiemap [out]          fiemap buffer holding retrived map extents
1157  * \param buflen [in/out]       max buffer length of @fiemap, when iterate
1158  *                              each OST, it is used to limit max map needed
1159  * \retval 0    success
1160  * \retval < 0  error
1161  */
1162 static int lov_object_fiemap(const struct lu_env *env, struct cl_object *obj,
1163                              struct ll_fiemap_info_key *fmkey,
1164                              struct fiemap *fiemap, size_t *buflen)
1165 {
1166         struct lov_stripe_md    *lsm;
1167         struct cl_object        *subobj = NULL;
1168         struct lov_obd          *lov = lu2lov_dev(obj->co_lu.lo_dev)->ld_lov;
1169         struct fiemap           *fm_local = NULL;
1170         struct ll_fiemap_extent *lcl_fm_ext;
1171         loff_t                  fm_start;
1172         loff_t                  fm_end;
1173         loff_t                  fm_length;
1174         loff_t                  fm_end_offset;
1175         int                     count_local;
1176         int                     ost_index = 0;
1177         int                     start_stripe;
1178         int                     current_extent = 0;
1179         int                     rc = 0;
1180         int                     last_stripe;
1181         int                     cur_stripe = 0;
1182         int                     cur_stripe_wrap = 0;
1183         int                     stripe_count;
1184         unsigned int            buffer_size = FIEMAP_BUFFER_SIZE;
1185         /* Whether have we collected enough extents */
1186         bool                    enough = false;
1187         /* EOF for object */
1188         bool                    ost_eof = false;
1189         /* done with required mapping for this OST? */
1190         bool                    ost_done = false;
1191         ENTRY;
1192
1193         lsm = lov_lsm_addref(cl2lov(obj));
1194         if (lsm == NULL)
1195                 RETURN(-ENODATA);
1196
1197         /**
1198          * If the stripe_count > 1 and the application does not understand
1199          * DEVICE_ORDER flag, it cannot interpret the extents correctly.
1200          */
1201         if (lsm->lsm_stripe_count > 1 && !(fiemap->fm_flags &
1202                                            FIEMAP_FLAG_DEVICE_ORDER))
1203                 GOTO(out, rc = -ENOTSUPP);
1204
1205         if (lsm_is_released(lsm)) {
1206                 if (fiemap->fm_start < fmkey->oa.o_size) {
1207                         /**
1208                          * released file, return a minimal FIEMAP if
1209                          * request fits in file-size.
1210                          */
1211                         fiemap->fm_mapped_extents = 1;
1212                         fiemap->fm_extents[0].fe_logical = fiemap->fm_start;
1213                         if (fiemap->fm_start + fiemap->fm_length <
1214                             fmkey->oa.o_size)
1215                                 fiemap->fm_extents[0].fe_length =
1216                                         fiemap->fm_length;
1217                         else
1218                                 fiemap->fm_extents[0].fe_length =
1219                                         fmkey->oa.o_size - fiemap->fm_start;
1220                         fiemap->fm_extents[0].fe_flags |=
1221                                 FIEMAP_EXTENT_UNKNOWN | FIEMAP_EXTENT_LAST;
1222                 }
1223                 GOTO(out, rc = 0);
1224         }
1225
1226         if (fiemap_count_to_size(fiemap->fm_extent_count) < buffer_size)
1227                 buffer_size = fiemap_count_to_size(fiemap->fm_extent_count);
1228
1229         OBD_ALLOC_LARGE(fm_local, buffer_size);
1230         if (fm_local == NULL)
1231                 GOTO(out, rc = -ENOMEM);
1232         lcl_fm_ext = &fm_local->fm_extents[0];
1233         count_local = fiemap_size_to_count(buffer_size);
1234
1235         fm_start = fiemap->fm_start;
1236         fm_length = fiemap->fm_length;
1237         /* Calculate start stripe, last stripe and length of mapping */
1238         start_stripe = lov_stripe_number(lsm, fm_start);
1239         fm_end = (fm_length == ~0ULL) ? fmkey->oa.o_size :
1240                                         fm_start + fm_length - 1;
1241         /* If fm_length != ~0ULL but fm_start_fm_length-1 exceeds file size */
1242         if (fm_end > fmkey->oa.o_size)
1243                 fm_end = fmkey->oa.o_size;
1244
1245         last_stripe = fiemap_calc_last_stripe(lsm, fm_start, fm_end,
1246                                               start_stripe, &stripe_count);
1247         fm_end_offset = fiemap_calc_fm_end_offset(fiemap, lsm, fm_start, fm_end,
1248                                                   &start_stripe);
1249         if (fm_end_offset == -EINVAL)
1250                 GOTO(out, rc = -EINVAL);
1251
1252         /**
1253          * Requested extent count exceeds the fiemap buffer size, shrink our
1254          * ambition.
1255          */
1256         if (fiemap_count_to_size(fiemap->fm_extent_count) > *buflen)
1257                 fiemap->fm_extent_count = fiemap_size_to_count(*buflen);
1258         if (fiemap->fm_extent_count == 0)
1259                 count_local = 0;
1260
1261         /* Check each stripe */
1262         for (cur_stripe = start_stripe; stripe_count > 0;
1263              --stripe_count,
1264              cur_stripe = (cur_stripe + 1) % lsm->lsm_stripe_count) {
1265                 loff_t req_fm_len; /* Stores length of required mapping */
1266                 loff_t len_mapped_single_call;
1267                 loff_t lun_start;
1268                 loff_t lun_end;
1269                 loff_t obd_object_end;
1270                 unsigned int ext_count;
1271
1272                 cur_stripe_wrap = cur_stripe;
1273
1274                 /* Find out range of mapping on this stripe */
1275                 if ((lov_stripe_intersects(lsm, cur_stripe, fm_start, fm_end,
1276                                            &lun_start, &obd_object_end)) == 0)
1277                         continue;
1278
1279                 if (lov_oinfo_is_dummy(lsm->lsm_oinfo[cur_stripe]))
1280                         GOTO(out, rc = -EIO);
1281
1282                 /* If this is a continuation FIEMAP call and we are on
1283                  * starting stripe then lun_start needs to be set to
1284                  * fm_end_offset */
1285                 if (fm_end_offset != 0 && cur_stripe == start_stripe)
1286                         lun_start = fm_end_offset;
1287
1288                 if (fm_length != ~0ULL) {
1289                         /* Handle fm_start + fm_length overflow */
1290                         if (fm_start + fm_length < fm_start)
1291                                 fm_length = ~0ULL - fm_start;
1292                         lun_end = lov_size_to_stripe(lsm, fm_start + fm_length,
1293                                                      cur_stripe);
1294                 } else {
1295                         lun_end = ~0ULL;
1296                 }
1297
1298                 if (lun_start == lun_end)
1299                         continue;
1300
1301                 req_fm_len = obd_object_end - lun_start;
1302                 fm_local->fm_length = 0;
1303                 len_mapped_single_call = 0;
1304
1305                 /* find lobsub object */
1306                 subobj = lov_find_subobj(env, cl2lov(obj), lsm,
1307                                              cur_stripe);
1308                 if (IS_ERR(subobj))
1309                         GOTO(out, rc = PTR_ERR(subobj));
1310                 /* If the output buffer is very large and the objects have many
1311                  * extents we may need to loop on a single OST repeatedly */
1312                 ost_eof = false;
1313                 ost_done = false;
1314                 do {
1315                         if (fiemap->fm_extent_count > 0) {
1316                                 /* Don't get too many extents. */
1317                                 if (current_extent + count_local >
1318                                     fiemap->fm_extent_count)
1319                                         count_local = fiemap->fm_extent_count -
1320                                                       current_extent;
1321                         }
1322
1323                         lun_start += len_mapped_single_call;
1324                         fm_local->fm_length = req_fm_len -
1325                                               len_mapped_single_call;
1326                         req_fm_len = fm_local->fm_length;
1327                         fm_local->fm_extent_count = enough ? 1 : count_local;
1328                         fm_local->fm_mapped_extents = 0;
1329                         fm_local->fm_flags = fiemap->fm_flags;
1330
1331                         ost_index = lsm->lsm_oinfo[cur_stripe]->loi_ost_idx;
1332
1333                         if (ost_index < 0 ||
1334                             ost_index >= lov->desc.ld_tgt_count)
1335                                 GOTO(obj_put, rc = -EINVAL);
1336                         /* If OST is inactive, return extent with UNKNOWN
1337                          * flag. */
1338                         if (!lov->lov_tgts[ost_index]->ltd_active) {
1339                                 fm_local->fm_flags |= FIEMAP_EXTENT_LAST;
1340                                 fm_local->fm_mapped_extents = 1;
1341
1342                                 lcl_fm_ext[0].fe_logical = lun_start;
1343                                 lcl_fm_ext[0].fe_length = obd_object_end -
1344                                                           lun_start;
1345                                 lcl_fm_ext[0].fe_flags |= FIEMAP_EXTENT_UNKNOWN;
1346
1347                                 goto inactive_tgt;
1348                         }
1349
1350                         fm_local->fm_start = lun_start;
1351                         fm_local->fm_flags &= ~FIEMAP_FLAG_DEVICE_ORDER;
1352                         memcpy(&fmkey->fiemap, fm_local, sizeof(*fm_local));
1353                         *buflen = fiemap_count_to_size(
1354                                                 fm_local->fm_extent_count);
1355
1356                         rc = cl_object_fiemap(env, subobj, fmkey, fm_local,
1357                                               buflen);
1358                         if (rc != 0)
1359                                 GOTO(obj_put, rc);
1360 inactive_tgt:
1361                         ext_count = fm_local->fm_mapped_extents;
1362                         if (ext_count == 0) {
1363                                 ost_done = true;
1364                                 /* If last stripe has hold at the end,
1365                                  * we need to return */
1366                                 if (cur_stripe_wrap == last_stripe) {
1367                                         fiemap->fm_mapped_extents = 0;
1368                                         goto finish;
1369                                 }
1370                                 break;
1371                         } else if (enough) {
1372                                 /*
1373                                  * We've collected enough extents and there are
1374                                  * more extents after it.
1375                                  */
1376                                 goto finish;
1377                         }
1378
1379                         /* If we just need num of extents, got to next device */
1380                         if (fiemap->fm_extent_count == 0) {
1381                                 current_extent += ext_count;
1382                                 break;
1383                         }
1384
1385                         /* prepare to copy retrived map extents */
1386                         len_mapped_single_call =
1387                                 lcl_fm_ext[ext_count - 1].fe_logical -
1388                                 lun_start + lcl_fm_ext[ext_count - 1].fe_length;
1389
1390                         /* Have we finished mapping on this device? */
1391                         if (req_fm_len <= len_mapped_single_call)
1392                                 ost_done = true;
1393
1394                         /* Clear the EXTENT_LAST flag which can be present on
1395                          * the last extent */
1396                         if (lcl_fm_ext[ext_count - 1].fe_flags &
1397                             FIEMAP_EXTENT_LAST)
1398                                 lcl_fm_ext[ext_count - 1].fe_flags &=
1399                                                         ~FIEMAP_EXTENT_LAST;
1400                         if (lov_stripe_size(lsm,
1401                                         lcl_fm_ext[ext_count - 1].fe_logical +
1402                                         lcl_fm_ext[ext_count - 1].fe_length,
1403                                         cur_stripe) >= fmkey->oa.o_size)
1404                                 ost_eof = true;
1405
1406                         fiemap_prepare_and_copy_exts(fiemap, lcl_fm_ext,
1407                                                      ost_index, ext_count,
1408                                                      current_extent);
1409                         current_extent += ext_count;
1410
1411                         /* Ran out of available extents? */
1412                         if (current_extent >= fiemap->fm_extent_count)
1413                                 enough = true;
1414                 } while (!ost_done && !ost_eof);
1415
1416                 cl_object_put(env, subobj);
1417                 subobj = NULL;
1418
1419                 if (cur_stripe_wrap == last_stripe)
1420                         goto finish;
1421         } /* for each stripe */
1422 finish:
1423         /* Indicate that we are returning device offsets unless file just has
1424          * single stripe */
1425         if (lsm->lsm_stripe_count > 1)
1426                 fiemap->fm_flags |= FIEMAP_FLAG_DEVICE_ORDER;
1427
1428         if (fiemap->fm_extent_count == 0)
1429                 goto skip_last_device_calc;
1430
1431         /* Check if we have reached the last stripe and whether mapping for that
1432          * stripe is done. */
1433         if ((cur_stripe_wrap == last_stripe) && (ost_done || ost_eof))
1434                 fiemap->fm_extents[current_extent - 1].fe_flags |=
1435                                                              FIEMAP_EXTENT_LAST;
1436 skip_last_device_calc:
1437         fiemap->fm_mapped_extents = current_extent;
1438 obj_put:
1439         if (subobj != NULL)
1440                 cl_object_put(env, subobj);
1441 out:
1442         if (fm_local != NULL)
1443                 OBD_FREE_LARGE(fm_local, buffer_size);
1444         lov_lsm_put(obj, lsm);
1445         RETURN(rc);
1446 }
1447
1448 static int lov_object_getstripe(const struct lu_env *env, struct cl_object *obj,
1449                                 struct lov_user_md __user *lum)
1450 {
1451         struct lov_object       *lov = cl2lov(obj);
1452         struct lov_stripe_md    *lsm;
1453         int                     rc = 0;
1454         ENTRY;
1455
1456         lsm = lov_lsm_addref(lov);
1457         if (lsm == NULL)
1458                 RETURN(-ENODATA);
1459
1460         rc = lov_getstripe(cl2lov(obj), lsm, lum);
1461         lov_lsm_put(obj, lsm);
1462         RETURN(rc);
1463 }
1464
1465 static int lov_object_find_cbdata(const struct lu_env *env,
1466                                   struct cl_object *obj, ldlm_iterator_t iter,
1467                                   void *data)
1468 {
1469         int rc;
1470         ENTRY;
1471
1472         /* call cl_object_find_cbdata for sub obj */
1473         rc = LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_find_cbdata,
1474                                   env, obj, iter, data);
1475         RETURN(rc);
1476 }
1477
1478 static const struct cl_object_operations lov_ops = {
1479         .coo_page_init   = lov_page_init,
1480         .coo_lock_init   = lov_lock_init,
1481         .coo_io_init     = lov_io_init,
1482         .coo_attr_get    = lov_attr_get,
1483         .coo_attr_update = lov_attr_update,
1484         .coo_conf_set    = lov_conf_set,
1485         .coo_getstripe   = lov_object_getstripe,
1486         .coo_find_cbdata = lov_object_find_cbdata,
1487         .coo_fiemap      = lov_object_fiemap,
1488 };
1489
1490 static const struct lu_object_operations lov_lu_obj_ops = {
1491         .loo_object_init      = lov_object_init,
1492         .loo_object_delete    = lov_object_delete,
1493         .loo_object_release   = NULL,
1494         .loo_object_free      = lov_object_free,
1495         .loo_object_print     = lov_object_print,
1496         .loo_object_invariant = NULL
1497 };
1498
1499 struct lu_object *lov_object_alloc(const struct lu_env *env,
1500                                    const struct lu_object_header *unused,
1501                                    struct lu_device *dev)
1502 {
1503         struct lov_object *lov;
1504         struct lu_object  *obj;
1505
1506         ENTRY;
1507         OBD_SLAB_ALLOC_PTR_GFP(lov, lov_object_kmem, GFP_NOFS);
1508         if (lov != NULL) {
1509                 obj = lov2lu(lov);
1510                 lu_object_init(obj, NULL, dev);
1511                 lov->lo_cl.co_ops = &lov_ops;
1512                 lov->lo_type = -1; /* invalid, to catch uninitialized type */
1513                 /*
1514                  * object io operation vector (cl_object::co_iop) is installed
1515                  * later in lov_object_init(), as different vectors are used
1516                  * for object with different layouts.
1517                  */
1518                 obj->lo_ops = &lov_lu_obj_ops;
1519         } else
1520                 obj = NULL;
1521         RETURN(obj);
1522 }
1523
1524 struct lov_stripe_md *lov_lsm_addref(struct lov_object *lov)
1525 {
1526         struct lov_stripe_md *lsm = NULL;
1527
1528         lov_conf_freeze(lov);
1529         if (lov->lo_lsm != NULL) {
1530                 lsm = lsm_addref(lov->lo_lsm);
1531                 CDEBUG(D_INODE, "lsm %p addref %d/%d by %p.\n",
1532                         lsm, atomic_read(&lsm->lsm_refc),
1533                         lov->lo_layout_invalid, current);
1534         }
1535         lov_conf_thaw(lov);
1536         return lsm;
1537 }
1538
1539 int lov_read_and_clear_async_rc(struct cl_object *clob)
1540 {
1541         struct lu_object *luobj;
1542         int rc = 0;
1543         ENTRY;
1544
1545         luobj = lu_object_locate(&cl_object_header(clob)->coh_lu,
1546                                  &lov_device_type);
1547         if (luobj != NULL) {
1548                 struct lov_object *lov = lu2lov(luobj);
1549
1550                 lov_conf_freeze(lov);
1551                 switch (lov->lo_type) {
1552                 case LLT_RAID0: {
1553                         struct lov_stripe_md *lsm;
1554                         int i;
1555
1556                         lsm = lov->lo_lsm;
1557                         LASSERT(lsm != NULL);
1558                         for (i = 0; i < lsm->lsm_stripe_count; i++) {
1559                                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
1560
1561                                 if (lov_oinfo_is_dummy(loi))
1562                                         continue;
1563
1564                                 if (loi->loi_ar.ar_rc && !rc)
1565                                         rc = loi->loi_ar.ar_rc;
1566                                 loi->loi_ar.ar_rc = 0;
1567                         }
1568                 }
1569                 case LLT_RELEASED:
1570                 case LLT_EMPTY:
1571                         break;
1572                 default:
1573                         LBUG();
1574                 }
1575                 lov_conf_thaw(lov);
1576         }
1577         RETURN(rc);
1578 }
1579 EXPORT_SYMBOL(lov_read_and_clear_async_rc);
1580
1581 /** @} lov */