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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2016, 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  * Implementation of cl_object for LOV layer.
  33  *
  34  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  35  *   Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_LOV
  39
  40 #include "lov_cl_internal.h"
  41
  42 static inline struct lov_device *lov_object_dev(struct lov_object *obj)
  43 {
  44         return lu2lov_dev(obj->lo_cl.co_lu.lo_dev);
  45 }
  46
  47 /** \addtogroup lov
  48  *  @{
  49  */
  50
  51 /*****************************************************************************
  52  *
  53  * Layout operations.
  54  *
  55  */
  56
  57 struct lov_layout_operations {
  58         int (*llo_init)(const struct lu_env *env, struct lov_device *dev,
  59                         struct lov_object *lov, struct lov_stripe_md *lsm,
  60                         const struct cl_object_conf *conf,
  61                         union lov_layout_state *state);
  62         int (*llo_delete)(const struct lu_env *env, struct lov_object *lov,
  63                            union lov_layout_state *state);
  64         void (*llo_fini)(const struct lu_env *env, struct lov_object *lov,
  65                          union lov_layout_state *state);
  66         int  (*llo_print)(const struct lu_env *env, void *cookie,
  67                           lu_printer_t p, const struct lu_object *o);
  68         int  (*llo_page_init)(const struct lu_env *env, struct cl_object *obj,
  69                               struct cl_page *page, pgoff_t index);
  70         int  (*llo_lock_init)(const struct lu_env *env,
  71                               struct cl_object *obj, struct cl_lock *lock,
  72                               const struct cl_io *io);
  73         int  (*llo_io_init)(const struct lu_env *env,
  74                             struct cl_object *obj, struct cl_io *io);
  75         int  (*llo_getattr)(const struct lu_env *env, struct cl_object *obj,
  76                             struct cl_attr *attr);
  77 };
  78
  79 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov);
  80
  81 static void lov_lsm_put(struct lov_stripe_md *lsm)
  82 {
  83         if (lsm != NULL)
  84                 lov_free_memmd(&lsm);
  85 }
  86
  87 /*****************************************************************************
  88  *
  89  * Lov object layout operations.
  90  *
  91  */
  92
  93 static struct cl_object *lov_sub_find(const struct lu_env *env,
  94                                       struct cl_device *dev,
  95                                       const struct lu_fid *fid,
  96                                       const struct cl_object_conf *conf)
  97 {
  98         struct lu_object *o;
  99
 100         ENTRY;
 101
 102         o = lu_object_find_at(env, cl2lu_dev(dev), fid, &conf->coc_lu);
 103         LASSERT(ergo(!IS_ERR(o), o->lo_dev->ld_type == &lovsub_device_type));
 104         RETURN(lu2cl(o));
 105 }
 106
 107 static int lov_page_slice_fixup(struct lov_object *lov,
 108                                 struct cl_object *stripe)
 109 {
 110         struct cl_object_header *hdr = cl_object_header(&lov->lo_cl);
 111         struct cl_object *o;
 112
 113         if (stripe == NULL)
 114                 return hdr->coh_page_bufsize - lov->lo_cl.co_slice_off -
 115                        cfs_size_round(sizeof(struct lov_page));
 116
 117         cl_object_for_each(o, stripe)
 118                 o->co_slice_off += hdr->coh_page_bufsize;
 119
 120         return cl_object_header(stripe)->coh_page_bufsize;
 121 }
 122
 123 static int lov_init_sub(const struct lu_env *env, struct lov_object *lov,
 124                         struct cl_object *subobj, struct lov_oinfo *oinfo,
 125                         int idx)
 126 {
 127         struct cl_object_header *hdr;
 128         struct cl_object_header *subhdr;
 129         struct cl_object_header *parent;
 130         int entry = lov_comp_entry(idx);
 131         int stripe = lov_comp_stripe(idx);
 132         int result;
 133
 134         if (OBD_FAIL_CHECK(OBD_FAIL_LOV_INIT)) {
 135                 /* For sanity:test_206.
 136                  * Do not leave the object in cache to avoid accessing
 137                  * freed memory. This is because osc_object is referring to
 138                  * lov_oinfo of lsm_stripe_data which will be freed due to
 139                  * this failure. */
 140                 cl_object_kill(env, subobj);
 141                 cl_object_put(env, subobj);
 142                 return -EIO;
 143         }
 144
 145         hdr = cl_object_header(lov2cl(lov));
 146         subhdr = cl_object_header(subobj);
 147
 148         CDEBUG(D_INODE, DFID"@%p[%d:%d] -> "DFID"@%p: ostid: "DOSTID
 149                " ost idx: %d gen: %d\n",
 150                PFID(lu_object_fid(&subobj->co_lu)), subhdr, entry, stripe,
 151                PFID(lu_object_fid(lov2lu(lov))), hdr, POSTID(&oinfo->loi_oi),
 152                oinfo->loi_ost_idx, oinfo->loi_ost_gen);
 153
 154         /* reuse ->coh_attr_guard to protect coh_parent change */
 155         spin_lock(&subhdr->coh_attr_guard);
 156         parent = subhdr->coh_parent;
 157         if (parent == NULL) {
 158                 struct lovsub_object *lso = cl2lovsub(subobj);
 159
 160                 subhdr->coh_parent = hdr;
 161                 spin_unlock(&subhdr->coh_attr_guard);
 162                 subhdr->coh_nesting = hdr->coh_nesting + 1;
 163                 lu_object_ref_add(&subobj->co_lu, "lov-parent", lov);
 164                 lso->lso_super = lov;
 165                 lso->lso_index = idx;
 166                 result = 0;
 167         } else {
 168                 struct lu_object  *old_obj;
 169                 struct lov_object *old_lov;
 170                 unsigned int mask = D_INODE;
 171
 172                 spin_unlock(&subhdr->coh_attr_guard);
 173                 old_obj = lu_object_locate(&parent->coh_lu, &lov_device_type);
 174                 LASSERT(old_obj != NULL);
 175                 old_lov = cl2lov(lu2cl(old_obj));
 176                 if (old_lov->lo_layout_invalid) {
 177                         /* the object's layout has already changed but isn't
 178                          * refreshed */
 179                         lu_object_unhash(env, &subobj->co_lu);
 180                         result = -EAGAIN;
 181                 } else {
 182                         mask = D_ERROR;
 183                         result = -EIO;
 184                 }
 185
 186                 LU_OBJECT_DEBUG(mask, env, &subobj->co_lu,
 187                                 "stripe %d is already owned.", idx);
 188                 LU_OBJECT_DEBUG(mask, env, old_obj, "owned.");
 189                 LU_OBJECT_HEADER(mask, env, lov2lu(lov), "try to own.\n");
 190                 cl_object_put(env, subobj);
 191         }
 192         return result;
 193 }
 194
 195 static int lov_init_raid0(const struct lu_env *env, struct lov_device *dev,
 196                           struct lov_object *lov, unsigned int index,
 197                           const struct cl_object_conf *conf,
 198                           struct lov_layout_entry *lle)
 199 {
 200         struct lov_layout_raid0 *r0 = &lle->lle_raid0;
 201         struct lov_thread_info *lti = lov_env_info(env);
 202         struct cl_object_conf *subconf = &lti->lti_stripe_conf;
 203         struct lu_fid *ofid = &lti->lti_fid;
 204         struct cl_object *stripe;
 205         struct lov_stripe_md_entry *lse  = lov_lse(lov, index);
 206         int result;
 207         int psz, sz;
 208         int i;
 209
 210         ENTRY;
 211
 212         spin_lock_init(&r0->lo_sub_lock);
 213         r0->lo_nr = lse->lsme_stripe_count;
 214         LASSERT(r0->lo_nr <= lov_targets_nr(dev));
 215
 216         OBD_ALLOC_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 217         if (r0->lo_sub == NULL)
 218                 GOTO(out, result = -ENOMEM);
 219
 220         psz = 0;
 221         result = 0;
 222         memset(subconf, 0, sizeof(*subconf));
 223
 224         /*
 225          * Create stripe cl_objects.
 226          */
 227         for (i = 0; i < r0->lo_nr; ++i) {
 228                 struct cl_device *subdev;
 229                 struct lov_oinfo *oinfo = lse->lsme_oinfo[i];
 230                 int ost_idx = oinfo->loi_ost_idx;
 231
 232                 if (lov_oinfo_is_dummy(oinfo))
 233                         continue;
 234
 235                 result = ostid_to_fid(ofid, &oinfo->loi_oi, oinfo->loi_ost_idx);
 236                 if (result != 0)
 237                         GOTO(out, result);
 238
 239                 if (dev->ld_target[ost_idx] == NULL) {
 240                         CERROR("%s: OST %04x is not initialized\n",
 241                                lov2obd(dev->ld_lov)->obd_name, ost_idx);
 242                         GOTO(out, result = -EIO);
 243                 }
 244
 245                 subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
 246                 subconf->u.coc_oinfo = oinfo;
 247                 LASSERTF(subdev != NULL, "not init ost %d\n", ost_idx);
 248                 /* In the function below, .hs_keycmp resolves to
 249                  * lu_obj_hop_keycmp() */
 250                 /* coverity[overrun-buffer-val] */
 251                 stripe = lov_sub_find(env, subdev, ofid, subconf);
 252                 if (IS_ERR(stripe))
 253                         GOTO(out, result = PTR_ERR(stripe));
 254
 255                 result = lov_init_sub(env, lov, stripe, oinfo,
 256                                       lov_comp_index(index, i));
 257                 if (result == -EAGAIN) { /* try again */
 258                         --i;
 259                         result = 0;
 260                         continue;
 261                 }
 262
 263                 if (result == 0) {
 264                         r0->lo_sub[i] = cl2lovsub(stripe);
 265
 266                         sz = lov_page_slice_fixup(lov, stripe);
 267                         LASSERT(ergo(psz > 0, psz == sz));
 268                         psz = sz;
 269                 }
 270         }
 271         if (result == 0)
 272                 result = psz;
 273 out:
 274         RETURN(result);
 275 }
 276
 277 static void lov_subobject_kill(const struct lu_env *env, struct lov_object *lov,
 278                                struct lov_layout_raid0 *r0,
 279                                struct lovsub_object *los, int idx)
 280 {
 281         struct cl_object        *sub;
 282         struct lu_site          *site;
 283         struct lu_site_bkt_data *bkt;
 284         wait_queue_t          *waiter;
 285
 286         LASSERT(r0->lo_sub[idx] == los);
 287
 288         sub  = lovsub2cl(los);
 289         site = sub->co_lu.lo_dev->ld_site;
 290         bkt  = lu_site_bkt_from_fid(site, &sub->co_lu.lo_header->loh_fid);
 291
 292         cl_object_kill(env, sub);
 293         /* release a reference to the sub-object and ... */
 294         lu_object_ref_del(&sub->co_lu, "lov-parent", lov);
 295         cl_object_put(env, sub);
 296
 297         /* ... wait until it is actually destroyed---sub-object clears its
 298          * ->lo_sub[] slot in lovsub_object_free() */
 299         if (r0->lo_sub[idx] == los) {
 300                 waiter = &lov_env_info(env)->lti_waiter;
 301                 init_waitqueue_entry(waiter, current);
 302                 add_wait_queue(&bkt->lsb_marche_funebre, waiter);
 303                 set_current_state(TASK_UNINTERRUPTIBLE);
 304                 while (1) {
 305                         /* this wait-queue is signaled at the end of
 306                          * lu_object_free(). */
 307                         set_current_state(TASK_UNINTERRUPTIBLE);
 308                         spin_lock(&r0->lo_sub_lock);
 309                         if (r0->lo_sub[idx] == los) {
 310                                 spin_unlock(&r0->lo_sub_lock);
 311                                 schedule();
 312                         } else {
 313                                 spin_unlock(&r0->lo_sub_lock);
 314                                 set_current_state(TASK_RUNNING);
 315                                 break;
 316                         }
 317                 }
 318                 remove_wait_queue(&bkt->lsb_marche_funebre, waiter);
 319         }
 320         LASSERT(r0->lo_sub[idx] == NULL);
 321 }
 322
 323 static void lov_delete_raid0(const struct lu_env *env, struct lov_object *lov,
 324                              struct lov_layout_entry *lle)
 325 {
 326         struct lov_layout_raid0 *r0 = &lle->lle_raid0;
 327
 328         ENTRY;
 329
 330         if (r0->lo_sub != NULL) {
 331                 int i;
 332
 333                 for (i = 0; i < r0->lo_nr; ++i) {
 334                         struct lovsub_object *los = r0->lo_sub[i];
 335
 336                         if (los != NULL) {
 337                                 cl_object_prune(env, &los->lso_cl);
 338                                 /*
 339                                  * If top-level object is to be evicted from
 340                                  * the cache, so are its sub-objects.
 341                                  */
 342                                 lov_subobject_kill(env, lov, r0, los, i);
 343                         }
 344                 }
 345         }
 346
 347         EXIT;
 348 }
 349
 350 static void lov_fini_raid0(const struct lu_env *env,
 351                            struct lov_layout_entry *lle)
 352 {
 353         struct lov_layout_raid0 *r0 = &lle->lle_raid0;
 354
 355         if (r0->lo_sub != NULL) {
 356                 OBD_FREE_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 357                 r0->lo_sub = NULL;
 358         }
 359 }
 360
 361 static int lov_print_raid0(const struct lu_env *env, void *cookie,
 362                            lu_printer_t p, const struct lov_layout_entry *lle)
 363 {
 364         const struct lov_layout_raid0 *r0 = &lle->lle_raid0;
 365         int i;
 366
 367         for (i = 0; i < r0->lo_nr; ++i) {
 368                 struct lu_object *sub;
 369
 370                 if (r0->lo_sub[i] != NULL) {
 371                         sub = lovsub2lu(r0->lo_sub[i]);
 372                         lu_object_print(env, cookie, p, sub);
 373                 } else {
 374                         (*p)(env, cookie, "sub %d absent\n", i);
 375                 }
 376         }
 377         return 0;
 378 }
 379
 380 static int lov_attr_get_raid0(const struct lu_env *env, struct lov_object *lov,
 381                               unsigned int index, struct lov_layout_entry *lle,
 382                               struct cl_attr **lov_attr)
 383 {
 384         struct lov_layout_raid0 *r0 = &lle->lle_raid0;
 385         struct lov_stripe_md *lsm = lov->lo_lsm;
 386         struct ost_lvb *lvb = &lov_env_info(env)->lti_lvb;
 387         struct cl_attr *attr = &r0->lo_attr;
 388         __u64 kms = 0;
 389         int result = 0;
 390
 391         if (r0->lo_attr_valid) {
 392                 *lov_attr = attr;
 393                 return 0;
 394         }
 395
 396         memset(lvb, 0, sizeof(*lvb));
 397
 398         /* XXX: timestamps can be negative by sanity:test_39m,
 399          * how can it be? */
 400         lvb->lvb_atime = LLONG_MIN;
 401         lvb->lvb_ctime = LLONG_MIN;
 402         lvb->lvb_mtime = LLONG_MIN;
 403
 404         /*
 405          * XXX that should be replaced with a loop over sub-objects,
 406          * doing cl_object_attr_get() on them. But for now, let's
 407          * reuse old lov code.
 408          */
 409
 410         /*
 411          * XXX take lsm spin-lock to keep lov_merge_lvb_kms()
 412          * happy. It's not needed, because new code uses
 413          * ->coh_attr_guard spin-lock to protect consistency of
 414          * sub-object attributes.
 415          */
 416         lov_stripe_lock(lsm);
 417         result = lov_merge_lvb_kms(lsm, index, lvb, &kms);
 418         lov_stripe_unlock(lsm);
 419         if (result == 0) {
 420                 cl_lvb2attr(attr, lvb);
 421                 attr->cat_kms = kms;
 422                 r0->lo_attr_valid = 1;
 423                 *lov_attr = attr;
 424         }
 425
 426         return result;
 427 }
 428
 429 static struct lov_comp_layout_entry_ops raid0_ops = {
 430         .lco_init      = lov_init_raid0,
 431         .lco_fini      = lov_fini_raid0,
 432         .lco_getattr   = lov_attr_get_raid0,
 433 };
 434
 435 static int lov_attr_get_dom(const struct lu_env *env, struct lov_object *lov,
 436                             unsigned int index, struct lov_layout_entry *lle,
 437                             struct cl_attr **lov_attr)
 438 {
 439         struct lov_layout_dom *dom = &lle->lle_dom;
 440         struct lov_oinfo *loi = dom->lo_loi;
 441         struct cl_attr *attr = &dom->lo_dom_r0.lo_attr;
 442
 443         if (dom->lo_dom_r0.lo_attr_valid) {
 444                 *lov_attr = attr;
 445                 return 0;
 446         }
 447
 448         if (OST_LVB_IS_ERR(loi->loi_lvb.lvb_blocks))
 449                 return OST_LVB_GET_ERR(loi->loi_lvb.lvb_blocks);
 450
 451         cl_lvb2attr(attr, &loi->loi_lvb);
 452
 453         /* DoM component size can be bigger than stripe size after
 454          * client's setattr RPC, so do not count anything beyond
 455          * component end. Alternatively, check that limit on server
 456          * and do not allow size overflow there. */
 457         if (attr->cat_size > lle->lle_extent.e_end)
 458                 attr->cat_size = lle->lle_extent.e_end;
 459
 460         attr->cat_kms = attr->cat_size;
 461
 462         dom->lo_dom_r0.lo_attr_valid = 1;
 463         *lov_attr = attr;
 464
 465         return 0;
 466 }
 467
 468 /**
 469  * Lookup FLD to get MDS index of the given DOM object FID.
 470  *
 471  * \param[in]  ld       LOV device
 472  * \param[in]  fid      FID to lookup
 473  * \param[out] nr       index in MDC array to return back
 474  *
 475  * \retval              0 and \a mds filled with MDS index if successful
 476  * \retval              negative value on error
 477  */
 478 static int lov_fld_lookup(struct lov_device *ld, const struct lu_fid *fid,
 479                           __u32 *nr)
 480 {
 481         __u32 mds_idx;
 482         int i, rc;
 483
 484         ENTRY;
 485
 486         rc = fld_client_lookup(&ld->ld_lmv->u.lmv.lmv_fld, fid_seq(fid),
 487                                &mds_idx, LU_SEQ_RANGE_MDT, NULL);
 488         if (rc) {
 489                 CERROR("%s: error while looking for mds number. Seq %#llx"
 490                        ", err = %d\n", lu_dev_name(cl2lu_dev(&ld->ld_cl)),
 491                        fid_seq(fid), rc);
 492                 RETURN(rc);
 493         }
 494
 495         CDEBUG(D_INODE, "FLD lookup got mds #%x for fid="DFID"\n",
 496                mds_idx, PFID(fid));
 497
 498         /* find proper MDC device in the array */
 499         for (i = 0; i < ld->ld_md_tgts_nr; i++) {
 500                 if (ld->ld_md_tgts[i].ldm_mdc != NULL &&
 501                     ld->ld_md_tgts[i].ldm_idx == mds_idx)
 502                         break;
 503         }
 504
 505         if (i == ld->ld_md_tgts_nr) {
 506                 CERROR("%s: cannot find corresponding MDC device for mds #%x "
 507                        "for fid="DFID"\n", lu_dev_name(cl2lu_dev(&ld->ld_cl)),
 508                        mds_idx, PFID(fid));
 509                 rc = -EINVAL;
 510         } else {
 511                 *nr = i;
 512         }
 513         RETURN(rc);
 514 }
 515
 516 /**
 517  * Implementation of lov_comp_layout_entry_ops::lco_init for DOM object.
 518  *
 519  * Init the DOM object for the first time. It prepares also RAID0 entry
 520  * for it to use in common methods with ordinary RAID0 layout entries.
 521  *
 522  * \param[in] env       execution environment
 523  * \param[in] dev       LOV device
 524  * \param[in] lov       LOV object
 525  * \param[in] index     Composite layout entry index in LSM
 526  * \param[in] lle       Composite LOV layout entry
 527  */
 528 static int lov_init_dom(const struct lu_env *env, struct lov_device *dev,
 529                         struct lov_object *lov, unsigned int index,
 530                         const struct cl_object_conf *conf,
 531                         struct lov_layout_entry *lle)
 532 {
 533         struct lov_thread_info *lti = lov_env_info(env);
 534         struct lov_stripe_md_entry *lsme = lov_lse(lov, index);
 535         struct cl_object *clo;
 536         struct lu_object *o = lov2lu(lov);
 537         const struct lu_fid *fid = lu_object_fid(o);
 538         struct cl_device *mdcdev;
 539         struct lov_oinfo *loi = NULL;
 540         struct cl_object_conf *sconf = &lti->lti_stripe_conf;
 541
 542         int rc;
 543         __u32 idx = 0;
 544
 545         ENTRY;
 546
 547         LASSERT(index == 0);
 548
 549         /* find proper MDS device */
 550         rc = lov_fld_lookup(dev, fid, &idx);
 551         if (rc)
 552                 RETURN(rc);
 553
 554         LASSERTF(dev->ld_md_tgts[idx].ldm_mdc != NULL,
 555                  "LOV md target[%u] is NULL\n", idx);
 556
 557         /* check lsm is DOM, more checks are needed */
 558         LASSERT(lsme->lsme_stripe_count == 0);
 559
 560         /*
 561          * Create lower cl_objects.
 562          */
 563         mdcdev = dev->ld_md_tgts[idx].ldm_mdc;
 564
 565         LASSERTF(mdcdev != NULL, "non-initialized mdc subdev\n");
 566
 567         /* DoM object has no oinfo in LSM entry, create it exclusively */
 568         OBD_SLAB_ALLOC_PTR_GFP(loi, lov_oinfo_slab, GFP_NOFS);
 569         if (loi == NULL)
 570                 RETURN(-ENOMEM);
 571
 572         fid_to_ostid(lu_object_fid(lov2lu(lov)), &loi->loi_oi);
 573
 574         sconf->u.coc_oinfo = loi;
 575 again:
 576         clo = lov_sub_find(env, mdcdev, fid, sconf);
 577         if (IS_ERR(clo))
 578                 GOTO(out, rc = PTR_ERR(clo));
 579
 580         rc = lov_init_sub(env, lov, clo, loi, lov_comp_index(index, 0));
 581         if (rc == -EAGAIN) /* try again */
 582                 goto again;
 583         else if (rc != 0)
 584                 GOTO(out, rc);
 585
 586         lle->lle_dom.lo_dom = cl2lovsub(clo);
 587         spin_lock_init(&lle->lle_dom.lo_dom_r0.lo_sub_lock);
 588         lle->lle_dom.lo_dom_r0.lo_nr = 1;
 589         lle->lle_dom.lo_dom_r0.lo_sub = &lle->lle_dom.lo_dom;
 590         lle->lle_dom.lo_loi = loi;
 591
 592         rc = lov_page_slice_fixup(lov, clo);
 593         RETURN(rc);
 594
 595 out:
 596         if (loi != NULL)
 597                 OBD_SLAB_FREE_PTR(loi, lov_oinfo_slab);
 598         return rc;
 599 }
 600
 601 /**
 602  * Implementation of lov_layout_operations::llo_fini for DOM object.
 603  *
 604  * Finish the DOM object and free related memory.
 605  *
 606  * \param[in] env       execution environment
 607  * \param[in] lov       LOV object
 608  * \param[in] state     LOV layout state
 609  */
 610 static void lov_fini_dom(const struct lu_env *env,
 611                          struct lov_layout_entry *lle)
 612 {
 613         if (lle->lle_dom.lo_dom != NULL)
 614                 lle->lle_dom.lo_dom = NULL;
 615         if (lle->lle_dom.lo_loi != NULL)
 616                 OBD_SLAB_FREE_PTR(lle->lle_dom.lo_loi, lov_oinfo_slab);
 617 }
 618
 619 static struct lov_comp_layout_entry_ops dom_ops = {
 620         .lco_init = lov_init_dom,
 621         .lco_fini = lov_fini_dom,
 622         .lco_getattr = lov_attr_get_dom,
 623 };
 624
 625 static int lov_init_composite(const struct lu_env *env, struct lov_device *dev,
 626                               struct lov_object *lov, struct lov_stripe_md *lsm,
 627                               const struct cl_object_conf *conf,
 628                               union lov_layout_state *state)
 629 {
 630         struct lov_layout_composite *comp = &state->composite;
 631         struct lov_layout_entry *lle;
 632         unsigned int entry_count;
 633         unsigned int psz = 0;
 634         int result = 0;
 635         int i;
 636
 637         ENTRY;
 638
 639         LASSERT(lsm->lsm_entry_count > 0);
 640         LASSERT(lov->lo_lsm == NULL);
 641         lov->lo_lsm = lsm_addref(lsm);
 642         lov->lo_layout_invalid = true;
 643
 644         entry_count = lsm->lsm_entry_count;
 645         comp->lo_entry_count = entry_count;
 646
 647         OBD_ALLOC(comp->lo_entries, entry_count * sizeof(*comp->lo_entries));
 648         if (comp->lo_entries == NULL)
 649                 RETURN(-ENOMEM);
 650
 651         /* Initiate all entry types and extents data at first */
 652         for (i = 0; i < entry_count; i++) {
 653                 lle = &comp->lo_entries[i];
 654
 655                 lle->lle_type = lov_entry_type(lsm->lsm_entries[i]);
 656                 switch (lle->lle_type) {
 657                 case LOV_PATTERN_RAID0:
 658                         lle->lle_comp_ops = &raid0_ops;
 659                         break;
 660                 case LOV_PATTERN_MDT:
 661                         lle->lle_comp_ops = &dom_ops;
 662                         break;
 663                 default:
 664                         CERROR("%s: unknown composite layout entry type %i\n",
 665                                lov2obd(dev->ld_lov)->obd_name,
 666                                lsm->lsm_entries[i]->lsme_pattern);
 667                         dump_lsm(D_ERROR, lsm);
 668                         RETURN(-EIO);
 669                 }
 670                 lle->lle_extent = lsm->lsm_entries[i]->lsme_extent;
 671         }
 672
 673         i = 0;
 674         lov_foreach_layout_entry(lov, lle) {
 675                 /**
 676                  * If the component has not been init-ed on MDS side, for
 677                  * PFL layout, we'd know that the components beyond this one
 678                  * will be dynamically init-ed later on file write/trunc ops.
 679                  */
 680                 if (lsm_entry_inited(lsm, i)) {
 681                         result = lle->lle_comp_ops->lco_init(env, dev, lov, i,
 682                                                              conf, lle);
 683                         if (result < 0)
 684                                 break;
 685
 686                         LASSERT(ergo(psz > 0, psz == result));
 687                         psz = result;
 688                 }
 689                 i++;
 690         }
 691         if (psz > 0)
 692                 cl_object_header(&lov->lo_cl)->coh_page_bufsize += psz;
 693
 694         return result > 0 ? 0 : result;
 695 }
 696
 697 static int lov_init_empty(const struct lu_env *env, struct lov_device *dev,
 698                           struct lov_object *lov, struct lov_stripe_md *lsm,
 699                           const struct cl_object_conf *conf,
 700                           union lov_layout_state *state)
 701 {
 702         return 0;
 703 }
 704
 705 static int lov_init_released(const struct lu_env *env,
 706                              struct lov_device *dev, struct lov_object *lov,
 707                              struct lov_stripe_md *lsm,
 708                              const struct cl_object_conf *conf,
 709                              union lov_layout_state *state)
 710 {
 711         LASSERT(lsm != NULL);
 712         LASSERT(lsm->lsm_is_released);
 713         LASSERT(lov->lo_lsm == NULL);
 714
 715         lov->lo_lsm = lsm_addref(lsm);
 716         return 0;
 717 }
 718
 719 static int lov_delete_empty(const struct lu_env *env, struct lov_object *lov,
 720                             union lov_layout_state *state)
 721 {
 722         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 723
 724         lov_layout_wait(env, lov);
 725         return 0;
 726 }
 727
 728 static int lov_delete_composite(const struct lu_env *env,
 729                                 struct lov_object *lov,
 730                                 union lov_layout_state *state)
 731 {
 732         struct lov_layout_entry *entry;
 733         struct lov_layout_composite *comp = &state->composite;
 734
 735         ENTRY;
 736
 737         dump_lsm(D_INODE, lov->lo_lsm);
 738
 739         lov_layout_wait(env, lov);
 740         if (comp->lo_entries)
 741                 lov_foreach_layout_entry(lov, entry)
 742                         lov_delete_raid0(env, lov, entry);
 743
 744         RETURN(0);
 745 }
 746
 747 static void lov_fini_empty(const struct lu_env *env, struct lov_object *lov,
 748                            union lov_layout_state *state)
 749 {
 750         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 751 }
 752
 753 static void lov_fini_composite(const struct lu_env *env,
 754                                struct lov_object *lov,
 755                                union lov_layout_state *state)
 756 {
 757         struct lov_layout_composite *comp = &state->composite;
 758         ENTRY;
 759
 760         if (comp->lo_entries != NULL) {
 761                 struct lov_layout_entry *entry;
 762
 763                 lov_foreach_layout_entry(lov, entry)
 764                         entry->lle_comp_ops->lco_fini(env, entry);
 765
 766                 OBD_FREE(comp->lo_entries,
 767                          comp->lo_entry_count * sizeof(*comp->lo_entries));
 768                 comp->lo_entries = NULL;
 769         }
 770
 771         dump_lsm(D_INODE, lov->lo_lsm);
 772         lov_free_memmd(&lov->lo_lsm);
 773
 774         EXIT;
 775 }
 776
 777 static void lov_fini_released(const struct lu_env *env, struct lov_object *lov,
 778                                 union lov_layout_state *state)
 779 {
 780         ENTRY;
 781         dump_lsm(D_INODE, lov->lo_lsm);
 782         lov_free_memmd(&lov->lo_lsm);
 783         EXIT;
 784 }
 785
 786 static int lov_print_empty(const struct lu_env *env, void *cookie,
 787                            lu_printer_t p, const struct lu_object *o)
 788 {
 789         (*p)(env, cookie, "empty %d\n", lu2lov(o)->lo_layout_invalid);
 790         return 0;
 791 }
 792
 793 static int lov_print_composite(const struct lu_env *env, void *cookie,
 794                                lu_printer_t p, const struct lu_object *o)
 795 {
 796         struct lov_object *lov = lu2lov(o);
 797         struct lov_stripe_md *lsm = lov->lo_lsm;
 798         int i;
 799
 800         (*p)(env, cookie, "entries: %d, %s, lsm{%p 0x%08X %d %u}:\n",
 801              lsm->lsm_entry_count,
 802              lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 803              lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 804              lsm->lsm_layout_gen);
 805
 806         for (i = 0; i < lsm->lsm_entry_count; i++) {
 807                 struct lov_stripe_md_entry *lse = lsm->lsm_entries[i];
 808                 struct lov_layout_entry *lle = lov_entry(lov, i);
 809
 810                 (*p)(env, cookie,
 811                      DEXT ": { 0x%08X, %u, %#x, %u, %#x, %u, %u }\n",
 812                      PEXT(&lse->lsme_extent), lse->lsme_magic,
 813                      lse->lsme_id, lse->lsme_pattern, lse->lsme_layout_gen,
 814                      lse->lsme_flags, lse->lsme_stripe_count,
 815                      lse->lsme_stripe_size);
 816                 lov_print_raid0(env, cookie, p, lle);
 817         }
 818
 819         return 0;
 820 }
 821
 822 static int lov_print_released(const struct lu_env *env, void *cookie,
 823                                 lu_printer_t p, const struct lu_object *o)
 824 {
 825         struct lov_object       *lov = lu2lov(o);
 826         struct lov_stripe_md    *lsm = lov->lo_lsm;
 827
 828         (*p)(env, cookie,
 829                 "released: %s, lsm{%p 0x%08X %d %u}:\n",
 830                 lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 831                 lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 832                 lsm->lsm_layout_gen);
 833         return 0;
 834 }
 835
 836 /**
 837  * Implements cl_object_operations::coo_attr_get() method for an object
 838  * without stripes (LLT_EMPTY layout type).
 839  *
 840  * The only attributes this layer is authoritative in this case is
 841  * cl_attr::cat_blocks---it's 0.
 842  */
 843 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
 844                               struct cl_attr *attr)
 845 {
 846         attr->cat_blocks = 0;
 847         return 0;
 848 }
 849
 850 static int lov_attr_get_composite(const struct lu_env *env,
 851                                   struct cl_object *obj,
 852                                   struct cl_attr *attr)
 853 {
 854         struct lov_object       *lov = cl2lov(obj);
 855         struct lov_layout_entry *entry;
 856         int                      result = 0;
 857         int                      index = 0;
 858
 859         ENTRY;
 860
 861         attr->cat_size = 0;
 862         attr->cat_blocks = 0;
 863         lov_foreach_layout_entry(lov, entry) {
 864                 struct cl_attr *lov_attr = NULL;
 865
 866                 /* PFL: This component has not been init-ed. */
 867                 if (!lsm_entry_inited(lov->lo_lsm, index))
 868                         break;
 869
 870                 result = entry->lle_comp_ops->lco_getattr(env, lov, index,
 871                                                           entry, &lov_attr);
 872                 if (result < 0)
 873                         RETURN(result);
 874
 875                 index++;
 876
 877                 if (lov_attr == NULL)
 878                         continue;
 879
 880                 CDEBUG(D_INODE, "COMP ID #%i: s=%llu m=%llu a=%llu c=%llu "
 881                        "b=%llu\n", index - 1, lov_attr->cat_size,
 882                        lov_attr->cat_mtime, lov_attr->cat_atime,
 883                        lov_attr->cat_ctime, lov_attr->cat_blocks);
 884
 885                 /* merge results */
 886                 attr->cat_blocks += lov_attr->cat_blocks;
 887                 if (attr->cat_size < lov_attr->cat_size)
 888                         attr->cat_size = lov_attr->cat_size;
 889                 if (attr->cat_kms < lov_attr->cat_kms)
 890                         attr->cat_kms = lov_attr->cat_kms;
 891                 if (attr->cat_atime < lov_attr->cat_atime)
 892                         attr->cat_atime = lov_attr->cat_atime;
 893                 if (attr->cat_ctime < lov_attr->cat_ctime)
 894                         attr->cat_ctime = lov_attr->cat_ctime;
 895                 if (attr->cat_mtime < lov_attr->cat_mtime)
 896                         attr->cat_mtime = lov_attr->cat_mtime;
 897         }
 898         RETURN(0);
 899 }
 900
 901 const static struct lov_layout_operations lov_dispatch[] = {
 902         [LLT_EMPTY] = {
 903                 .llo_init      = lov_init_empty,
 904                 .llo_delete    = lov_delete_empty,
 905                 .llo_fini      = lov_fini_empty,
 906                 .llo_print     = lov_print_empty,
 907                 .llo_page_init = lov_page_init_empty,
 908                 .llo_lock_init = lov_lock_init_empty,
 909                 .llo_io_init   = lov_io_init_empty,
 910                 .llo_getattr   = lov_attr_get_empty,
 911         },
 912         [LLT_RELEASED] = {
 913                 .llo_init      = lov_init_released,
 914                 .llo_delete    = lov_delete_empty,
 915                 .llo_fini      = lov_fini_released,
 916                 .llo_print     = lov_print_released,
 917                 .llo_page_init = lov_page_init_empty,
 918                 .llo_lock_init = lov_lock_init_empty,
 919                 .llo_io_init   = lov_io_init_released,
 920                 .llo_getattr   = lov_attr_get_empty,
 921         },
 922         [LLT_COMP] = {
 923                 .llo_init      = lov_init_composite,
 924                 .llo_delete    = lov_delete_composite,
 925                 .llo_fini      = lov_fini_composite,
 926                 .llo_print     = lov_print_composite,
 927                 .llo_page_init = lov_page_init_composite,
 928                 .llo_lock_init = lov_lock_init_composite,
 929                 .llo_io_init   = lov_io_init_composite,
 930                 .llo_getattr   = lov_attr_get_composite,
 931         },
 932 };
 933
 934 /**
 935  * Performs a double-dispatch based on the layout type of an object.
 936  */
 937 #define LOV_2DISPATCH_NOLOCK(obj, op, ...)              \
 938 ({                                                      \
 939         struct lov_object *__obj = (obj);               \
 940         enum lov_layout_type __llt;                     \
 941                                                         \
 942         __llt = __obj->lo_type;                         \
 943         LASSERT(__llt < ARRAY_SIZE(lov_dispatch));      \
 944         lov_dispatch[__llt].op(__VA_ARGS__);            \
 945 })
 946
 947 /**
 948  * Return lov_layout_type associated with a given lsm
 949  */
 950 static enum lov_layout_type lov_type(struct lov_stripe_md *lsm)
 951 {
 952         if (lsm == NULL)
 953                 return LLT_EMPTY;
 954
 955         if (lsm->lsm_is_released)
 956                 return LLT_RELEASED;
 957
 958         if (lsm->lsm_magic == LOV_MAGIC_V1 ||
 959             lsm->lsm_magic == LOV_MAGIC_V3 ||
 960             lsm->lsm_magic == LOV_MAGIC_COMP_V1)
 961                 return LLT_COMP;
 962
 963         return LLT_EMPTY;
 964 }
 965
 966 static inline void lov_conf_freeze(struct lov_object *lov)
 967 {
 968         CDEBUG(D_INODE, "To take share lov(%p) owner %p/%p\n",
 969                 lov, lov->lo_owner, current);
 970         if (lov->lo_owner != current)
 971                 down_read(&lov->lo_type_guard);
 972 }
 973
 974 static inline void lov_conf_thaw(struct lov_object *lov)
 975 {
 976         CDEBUG(D_INODE, "To release share lov(%p) owner %p/%p\n",
 977                 lov, lov->lo_owner, current);
 978         if (lov->lo_owner != current)
 979                 up_read(&lov->lo_type_guard);
 980 }
 981
 982 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...)                       \
 983 ({                                                                      \
 984         struct lov_object                      *__obj = (obj);          \
 985         int                                     __lock = !!(lock);      \
 986         typeof(lov_dispatch[0].op(__VA_ARGS__)) __result;               \
 987                                                                         \
 988         if (__lock)                                                     \
 989                 lov_conf_freeze(__obj);                                 \
 990         __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__);          \
 991         if (__lock)                                                     \
 992                 lov_conf_thaw(__obj);                                   \
 993         __result;                                                       \
 994 })
 995
 996 /**
 997  * Performs a locked double-dispatch based on the layout type of an object.
 998  */
 999 #define LOV_2DISPATCH(obj, op, ...)                     \
1000         LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
1001
1002 #define LOV_2DISPATCH_VOID(obj, op, ...)                                \
1003 do {                                                                    \
1004         struct lov_object                      *__obj = (obj);          \
1005         enum lov_layout_type                    __llt;                  \
1006                                                                         \
1007         lov_conf_freeze(__obj);                                         \
1008         __llt = __obj->lo_type;                                         \
1009         LASSERT(__llt < ARRAY_SIZE(lov_dispatch));                      \
1010         lov_dispatch[__llt].op(__VA_ARGS__);                            \
1011         lov_conf_thaw(__obj);                                           \
1012 } while (0)
1013
1014 static void lov_conf_lock(struct lov_object *lov)
1015 {
1016         LASSERT(lov->lo_owner != current);
1017         down_write(&lov->lo_type_guard);
1018         LASSERT(lov->lo_owner == NULL);
1019         lov->lo_owner = current;
1020         CDEBUG(D_INODE, "Took exclusive lov(%p) owner %p\n",
1021                 lov, lov->lo_owner);
1022 }
1023
1024 static void lov_conf_unlock(struct lov_object *lov)
1025 {
1026         CDEBUG(D_INODE, "To release exclusive lov(%p) owner %p\n",
1027                 lov, lov->lo_owner);
1028         lov->lo_owner = NULL;
1029         up_write(&lov->lo_type_guard);
1030 }
1031
1032 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov)
1033 {
1034         struct l_wait_info lwi = { 0 };
1035         ENTRY;
1036
1037         while (atomic_read(&lov->lo_active_ios) > 0) {
1038                 CDEBUG(D_INODE, "file:"DFID" wait for active IO, now: %d.\n",
1039                         PFID(lu_object_fid(lov2lu(lov))),
1040                         atomic_read(&lov->lo_active_ios));
1041
1042                 l_wait_event(lov->lo_waitq,
1043                              atomic_read(&lov->lo_active_ios) == 0, &lwi);
1044         }
1045         RETURN(0);
1046 }
1047
1048 static int lov_layout_change(const struct lu_env *unused,
1049                              struct lov_object *lov, struct lov_stripe_md *lsm,
1050                              const struct cl_object_conf *conf)
1051 {
1052         enum lov_layout_type llt = lov_type(lsm);
1053         union lov_layout_state *state = &lov->u;
1054         const struct lov_layout_operations *old_ops;
1055         const struct lov_layout_operations *new_ops;
1056         struct lov_device *lov_dev = lov_object_dev(lov);
1057         struct lu_env *env;
1058         __u16 refcheck;
1059         int rc;
1060         ENTRY;
1061
1062         LASSERT(lov->lo_type < ARRAY_SIZE(lov_dispatch));
1063
1064         env = cl_env_get(&refcheck);
1065         if (IS_ERR(env))
1066                 RETURN(PTR_ERR(env));
1067
1068         LASSERT(llt < ARRAY_SIZE(lov_dispatch));
1069
1070         CDEBUG(D_INODE, DFID" from %s to %s\n",
1071                PFID(lu_object_fid(lov2lu(lov))),
1072                llt2str(lov->lo_type), llt2str(llt));
1073
1074         old_ops = &lov_dispatch[lov->lo_type];
1075         new_ops = &lov_dispatch[llt];
1076
1077         rc = cl_object_prune(env, &lov->lo_cl);
1078         if (rc != 0)
1079                 GOTO(out, rc);
1080
1081         rc = old_ops->llo_delete(env, lov, &lov->u);
1082         if (rc != 0)
1083                 GOTO(out, rc);
1084
1085         old_ops->llo_fini(env, lov, &lov->u);
1086
1087         LASSERT(atomic_read(&lov->lo_active_ios) == 0);
1088
1089         CDEBUG(D_INODE, DFID "Apply new layout lov %p, type %d\n",
1090                PFID(lu_object_fid(lov2lu(lov))), lov, llt);
1091
1092         lov->lo_type = LLT_EMPTY;
1093
1094         /* page bufsize fixup */
1095         cl_object_header(&lov->lo_cl)->coh_page_bufsize -=
1096                 lov_page_slice_fixup(lov, NULL);
1097
1098         rc = new_ops->llo_init(env, lov_dev, lov, lsm, conf, state);
1099         if (rc != 0) {
1100                 struct obd_device *obd = lov2obd(lov_dev->ld_lov);
1101
1102                 CERROR("%s: cannot apply new layout on "DFID" : rc = %d\n",
1103                        obd->obd_name, PFID(lu_object_fid(lov2lu(lov))), rc);
1104                 new_ops->llo_delete(env, lov, state);
1105                 new_ops->llo_fini(env, lov, state);
1106                 /* this file becomes an EMPTY file. */
1107                 GOTO(out, rc);
1108         }
1109
1110         lov->lo_type = llt;
1111
1112 out:
1113         cl_env_put(env, &refcheck);
1114         RETURN(rc);
1115 }
1116
1117 /*****************************************************************************
1118  *
1119  * Lov object operations.
1120  *
1121  */
1122 int lov_object_init(const struct lu_env *env, struct lu_object *obj,
1123                     const struct lu_object_conf *conf)
1124 {
1125         struct lov_object            *lov   = lu2lov(obj);
1126         struct lov_device            *dev   = lov_object_dev(lov);
1127         const struct cl_object_conf  *cconf = lu2cl_conf(conf);
1128         union lov_layout_state       *set   = &lov->u;
1129         const struct lov_layout_operations *ops;
1130         struct lov_stripe_md *lsm = NULL;
1131         int rc;
1132         ENTRY;
1133
1134         init_rwsem(&lov->lo_type_guard);
1135         atomic_set(&lov->lo_active_ios, 0);
1136         init_waitqueue_head(&lov->lo_waitq);
1137         cl_object_page_init(lu2cl(obj), sizeof(struct lov_page));
1138
1139         lov->lo_type = LLT_EMPTY;
1140         if (cconf->u.coc_layout.lb_buf != NULL) {
1141                 lsm = lov_unpackmd(dev->ld_lov,
1142                                    cconf->u.coc_layout.lb_buf,
1143                                    cconf->u.coc_layout.lb_len);
1144                 if (IS_ERR(lsm))
1145                         RETURN(PTR_ERR(lsm));
1146
1147                 dump_lsm(D_INODE, lsm);
1148         }
1149
1150         /* no locking is necessary, as object is being created */
1151         lov->lo_type = lov_type(lsm);
1152         ops = &lov_dispatch[lov->lo_type];
1153         rc = ops->llo_init(env, dev, lov, lsm, cconf, set);
1154         if (rc != 0)
1155                 GOTO(out_lsm, rc);
1156
1157 out_lsm:
1158         lov_lsm_put(lsm);
1159
1160         RETURN(rc);
1161 }
1162
1163 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
1164                         const struct cl_object_conf *conf)
1165 {
1166         struct lov_stripe_md    *lsm = NULL;
1167         struct lov_object       *lov = cl2lov(obj);
1168         int                      result = 0;
1169         ENTRY;
1170
1171         if (conf->coc_opc == OBJECT_CONF_SET &&
1172             conf->u.coc_layout.lb_buf != NULL) {
1173                 lsm = lov_unpackmd(lov_object_dev(lov)->ld_lov,
1174                                    conf->u.coc_layout.lb_buf,
1175                                    conf->u.coc_layout.lb_len);
1176                 if (IS_ERR(lsm))
1177                         RETURN(PTR_ERR(lsm));
1178                 dump_lsm(D_INODE, lsm);
1179         }
1180
1181         lov_conf_lock(lov);
1182         if (conf->coc_opc == OBJECT_CONF_INVALIDATE) {
1183                 lov->lo_layout_invalid = true;
1184                 GOTO(out, result = 0);
1185         }
1186
1187         if (conf->coc_opc == OBJECT_CONF_WAIT) {
1188                 if (lov->lo_layout_invalid &&
1189                     atomic_read(&lov->lo_active_ios) > 0) {
1190                         lov_conf_unlock(lov);
1191                         result = lov_layout_wait(env, lov);
1192                         lov_conf_lock(lov);
1193                 }
1194                 GOTO(out, result);
1195         }
1196
1197         LASSERT(conf->coc_opc == OBJECT_CONF_SET);
1198
1199         if ((lsm == NULL && lov->lo_lsm == NULL) ||
1200             ((lsm != NULL && lov->lo_lsm != NULL) &&
1201              (lov->lo_lsm->lsm_layout_gen == lsm->lsm_layout_gen) &&
1202              (lov->lo_lsm->lsm_entries[0]->lsme_pattern ==
1203               lsm->lsm_entries[0]->lsme_pattern))) {
1204                 /* same version of layout */
1205                 lov->lo_layout_invalid = false;
1206                 GOTO(out, result = 0);
1207         }
1208
1209         /* will change layout - check if there still exists active IO. */
1210         if (atomic_read(&lov->lo_active_ios) > 0) {
1211                 lov->lo_layout_invalid = true;
1212                 GOTO(out, result = -EBUSY);
1213         }
1214
1215         result = lov_layout_change(env, lov, lsm, conf);
1216         lov->lo_layout_invalid = result != 0;
1217         EXIT;
1218
1219 out:
1220         lov_conf_unlock(lov);
1221         lov_lsm_put(lsm);
1222         CDEBUG(D_INODE, DFID" lo_layout_invalid=%d\n",
1223                PFID(lu_object_fid(lov2lu(lov))), lov->lo_layout_invalid);
1224         RETURN(result);
1225 }
1226
1227 static void lov_object_delete(const struct lu_env *env, struct lu_object *obj)
1228 {
1229         struct lov_object *lov = lu2lov(obj);
1230
1231         ENTRY;
1232         LOV_2DISPATCH_VOID(lov, llo_delete, env, lov, &lov->u);
1233         EXIT;
1234 }
1235
1236 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
1237 {
1238         struct lov_object *lov = lu2lov(obj);
1239
1240         ENTRY;
1241         LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
1242         lu_object_fini(obj);
1243         OBD_SLAB_FREE_PTR(lov, lov_object_kmem);
1244         EXIT;
1245 }
1246
1247 static int lov_object_print(const struct lu_env *env, void *cookie,
1248                             lu_printer_t p, const struct lu_object *o)
1249 {
1250         return LOV_2DISPATCH_NOLOCK(lu2lov(o), llo_print, env, cookie, p, o);
1251 }
1252
1253 int lov_page_init(const struct lu_env *env, struct cl_object *obj,
1254                   struct cl_page *page, pgoff_t index)
1255 {
1256         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_page_init, env, obj, page,
1257                                     index);
1258 }
1259
1260 /**
1261  * Implements cl_object_operations::clo_io_init() method for lov
1262  * layer. Dispatches to the appropriate layout io initialization method.
1263  */
1264 int lov_io_init(const struct lu_env *env, struct cl_object *obj,
1265                 struct cl_io *io)
1266 {
1267         CL_IO_SLICE_CLEAN(lov_env_io(env), lis_cl);
1268
1269         CDEBUG(D_INODE, DFID "io %p type %d ignore/verify layout %d/%d\n",
1270                PFID(lu_object_fid(&obj->co_lu)), io, io->ci_type,
1271                io->ci_ignore_layout, io->ci_verify_layout);
1272
1273         /* IO type CIT_MISC with ci_ignore_layout set are usually invoked from
1274          * the OSC layer. It shouldn't take lov layout conf lock in that case,
1275          * because as long as the OSC object exists, the layout can't be
1276          * reconfigured. */
1277         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
1278                         !(io->ci_ignore_layout && io->ci_type == CIT_MISC),
1279                         env, obj, io);
1280 }
1281
1282 /**
1283  * An implementation of cl_object_operations::clo_attr_get() method for lov
1284  * layer. For raid0 layout this collects and merges attributes of all
1285  * sub-objects.
1286  */
1287 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
1288                         struct cl_attr *attr)
1289 {
1290         /* do not take lock, as this function is called under a
1291          * spin-lock. Layout is protected from changing by ongoing IO. */
1292         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
1293 }
1294
1295 static int lov_attr_update(const struct lu_env *env, struct cl_object *obj,
1296                            const struct cl_attr *attr, unsigned valid)
1297 {
1298         /*
1299          * No dispatch is required here, as no layout implements this.
1300          */
1301         return 0;
1302 }
1303
1304 int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
1305                   struct cl_lock *lock, const struct cl_io *io)
1306 {
1307         /* No need to lock because we've taken one refcount of layout.  */
1308         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_lock_init, env, obj, lock,
1309                                     io);
1310 }
1311
1312 /**
1313  * We calculate on which OST the mapping will end. If the length of mapping
1314  * is greater than (stripe_size * stripe_count) then the last_stripe will
1315  * will be one just before start_stripe. Else we check if the mapping
1316  * intersects each OST and find last_stripe.
1317  * This function returns the last_stripe and also sets the stripe_count
1318  * over which the mapping is spread
1319  *
1320  * \param lsm [in]              striping information for the file
1321  * \param index [in]            stripe component index
1322  * \param ext [in]              logical extent of mapping
1323  * \param start_stripe [in]     starting stripe of the mapping
1324  * \param stripe_count [out]    the number of stripes across which to map is
1325  *                              returned
1326  *
1327  * \retval last_stripe          return the last stripe of the mapping
1328  */
1329 static int fiemap_calc_last_stripe(struct lov_stripe_md *lsm, int index,
1330                                    struct lu_extent *ext,
1331                                    int start_stripe, int *stripe_count)
1332 {
1333         struct lov_stripe_md_entry *lsme = lsm->lsm_entries[index];
1334         int last_stripe;
1335         u64 obd_start;
1336         u64 obd_end;
1337         int i, j;
1338
1339         if (ext->e_end - ext->e_start >
1340             lsme->lsme_stripe_size * lsme->lsme_stripe_count) {
1341                 last_stripe = (start_stripe < 1 ? lsme->lsme_stripe_count - 1 :
1342                                                   start_stripe - 1);
1343                 *stripe_count = lsme->lsme_stripe_count;
1344         } else {
1345                 for (j = 0, i = start_stripe; j < lsme->lsme_stripe_count;
1346                      i = (i + 1) % lsme->lsme_stripe_count, j++) {
1347                         if ((lov_stripe_intersects(lsm, index,  i, ext,
1348                                                    &obd_start, &obd_end)) == 0)
1349                                 break;
1350                 }
1351                 *stripe_count = j;
1352                 last_stripe = (start_stripe + j - 1) % lsme->lsme_stripe_count;
1353         }
1354
1355         return last_stripe;
1356 }
1357
1358 /**
1359  * Set fe_device and copy extents from local buffer into main return buffer.
1360  *
1361  * \param fiemap [out]          fiemap to hold all extents
1362  * \param lcl_fm_ext [in]       array of fiemap extents get from OSC layer
1363  * \param ost_index [in]        OST index to be written into the fm_device
1364  *                              field for each extent
1365  * \param ext_count [in]        number of extents to be copied
1366  * \param current_extent [in]   where to start copying in the extent array
1367  */
1368 static void fiemap_prepare_and_copy_exts(struct fiemap *fiemap,
1369                                          struct fiemap_extent *lcl_fm_ext,
1370                                          int ost_index, unsigned int ext_count,
1371                                          int current_extent)
1372 {
1373         char            *to;
1374         unsigned int    ext;
1375
1376         for (ext = 0; ext < ext_count; ext++) {
1377                 lcl_fm_ext[ext].fe_device = ost_index;
1378                 lcl_fm_ext[ext].fe_flags |= FIEMAP_EXTENT_NET;
1379         }
1380
1381         /* Copy fm_extent's from fm_local to return buffer */
1382         to = (char *)fiemap + fiemap_count_to_size(current_extent);
1383         memcpy(to, lcl_fm_ext, ext_count * sizeof(struct fiemap_extent));
1384 }
1385
1386 #define FIEMAP_BUFFER_SIZE 4096
1387
1388 /**
1389  * Non-zero fe_logical indicates that this is a continuation FIEMAP
1390  * call. The local end offset and the device are sent in the first
1391  * fm_extent. This function calculates the stripe number from the index.
1392  * This function returns a stripe_no on which mapping is to be restarted.
1393  *
1394  * This function returns fm_end_offset which is the in-OST offset at which
1395  * mapping should be restarted. If fm_end_offset=0 is returned then caller
1396  * will re-calculate proper offset in next stripe.
1397  * Note that the first extent is passed to lov_get_info via the value field.
1398  *
1399  * \param fiemap [in]           fiemap request header
1400  * \param lsm [in]              striping information for the file
1401  * \param index [in]            stripe component index
1402  * \param ext [in]              logical extent of mapping
1403  * \param start_stripe [out]    starting stripe will be returned in this
1404  */
1405 static u64 fiemap_calc_fm_end_offset(struct fiemap *fiemap,
1406                                      struct lov_stripe_md *lsm,
1407                                      int index, struct lu_extent *ext,
1408                                      int *start_stripe)
1409 {
1410         struct lov_stripe_md_entry *lsme = lsm->lsm_entries[index];
1411         u64 local_end = fiemap->fm_extents[0].fe_logical;
1412         u64 lun_start;
1413         u64 lun_end;
1414         u64 fm_end_offset;
1415         int stripe_no = -1;
1416         int i;
1417
1418         if (fiemap->fm_extent_count == 0 ||
1419             fiemap->fm_extents[0].fe_logical == 0)
1420                 return 0;
1421
1422         /* Find out stripe_no from ost_index saved in the fe_device */
1423         for (i = 0; i < lsme->lsme_stripe_count; i++) {
1424                 struct lov_oinfo *oinfo = lsme->lsme_oinfo[i];
1425
1426                 if (lov_oinfo_is_dummy(oinfo))
1427                         continue;
1428
1429                 if (oinfo->loi_ost_idx == fiemap->fm_extents[0].fe_device) {
1430                         stripe_no = i;
1431                         break;
1432                 }
1433         }
1434
1435         if (stripe_no == -1)
1436                 return -EINVAL;
1437
1438         /* If we have finished mapping on previous device, shift logical
1439          * offset to start of next device */
1440         if (lov_stripe_intersects(lsm, index, stripe_no, ext,
1441                                    &lun_start, &lun_end) != 0 &&
1442             local_end < lun_end) {
1443                 fm_end_offset = local_end;
1444                 *start_stripe = stripe_no;
1445         } else {
1446                 /* This is a special value to indicate that caller should
1447                  * calculate offset in next stripe. */
1448                 fm_end_offset = 0;
1449                 *start_stripe = (stripe_no + 1) % lsme->lsme_stripe_count;
1450         }
1451
1452         return fm_end_offset;
1453 }
1454
1455 struct fiemap_state {
1456         struct fiemap           *fs_fm;
1457         struct lu_extent        fs_ext;
1458         u64                     fs_length;
1459         u64                     fs_end_offset;
1460         int                     fs_cur_extent;
1461         int                     fs_cnt_need;
1462         int                     fs_start_stripe;
1463         int                     fs_last_stripe;
1464         bool                    fs_device_done;
1465         bool                    fs_finish_stripe;
1466         bool                    fs_enough;
1467 };
1468
1469 static struct cl_object *lov_find_subobj(const struct lu_env *env,
1470                                          struct lov_object *lov,
1471                                          struct lov_stripe_md *lsm,
1472                                          int index)
1473 {
1474         struct lov_device       *dev = lu2lov_dev(lov2lu(lov)->lo_dev);
1475         struct lov_thread_info  *lti = lov_env_info(env);
1476         struct lu_fid           *ofid = &lti->lti_fid;
1477         struct lov_oinfo        *oinfo;
1478         struct cl_device        *subdev;
1479         int                     entry = lov_comp_entry(index);
1480         int                     stripe = lov_comp_stripe(index);
1481         int                     ost_idx;
1482         int                     rc;
1483         struct cl_object        *result;
1484
1485         if (lov->lo_type != LLT_COMP)
1486                 GOTO(out, result = NULL);
1487
1488         if (entry >= lsm->lsm_entry_count ||
1489             stripe >= lsm->lsm_entries[entry]->lsme_stripe_count)
1490                 GOTO(out, result = NULL);
1491
1492         oinfo = lsm->lsm_entries[entry]->lsme_oinfo[stripe];
1493         ost_idx = oinfo->loi_ost_idx;
1494         rc = ostid_to_fid(ofid, &oinfo->loi_oi, ost_idx);
1495         if (rc != 0)
1496                 GOTO(out, result = NULL);
1497
1498         subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
1499         result = lov_sub_find(env, subdev, ofid, NULL);
1500 out:
1501         if (result == NULL)
1502                 result = ERR_PTR(-EINVAL);
1503         return result;
1504 }
1505
1506 int fiemap_for_stripe(const struct lu_env *env, struct cl_object *obj,
1507                       struct lov_stripe_md *lsm, struct fiemap *fiemap,
1508                       size_t *buflen, struct ll_fiemap_info_key *fmkey,
1509                       int index, int stripeno, struct fiemap_state *fs)
1510 {
1511         struct lov_stripe_md_entry *lsme = lsm->lsm_entries[index];
1512         struct cl_object *subobj;
1513         struct lov_obd *lov = lu2lov_dev(obj->co_lu.lo_dev)->ld_lov;
1514         struct fiemap_extent *fm_ext = &fs->fs_fm->fm_extents[0];
1515         u64 req_fm_len; /* Stores length of required mapping */
1516         u64 len_mapped_single_call;
1517         u64 lun_start;
1518         u64 lun_end;
1519         u64 obd_object_end;
1520         unsigned int ext_count;
1521         /* EOF for object */
1522         bool ost_eof = false;
1523         /* done with required mapping for this OST? */
1524         bool ost_done = false;
1525         int ost_index;
1526         int rc = 0;
1527
1528         fs->fs_device_done = false;
1529         /* Find out range of mapping on this stripe */
1530         if ((lov_stripe_intersects(lsm, index, stripeno, &fs->fs_ext,
1531                                    &lun_start, &obd_object_end)) == 0)
1532                 return 0;
1533
1534         if (lov_oinfo_is_dummy(lsme->lsme_oinfo[stripeno]))
1535                 return -EIO;
1536
1537         /* If this is a continuation FIEMAP call and we are on
1538          * starting stripe then lun_start needs to be set to
1539          * end_offset */
1540         if (fs->fs_end_offset != 0 && stripeno == fs->fs_start_stripe)
1541                 lun_start = fs->fs_end_offset;
1542         lun_end = lov_size_to_stripe(lsm, index, fs->fs_ext.e_end, stripeno);
1543         if (lun_start == lun_end)
1544                 return 0;
1545
1546         req_fm_len = obd_object_end - lun_start;
1547         fs->fs_fm->fm_length = 0;
1548         len_mapped_single_call = 0;
1549
1550         /* find lobsub object */
1551         subobj = lov_find_subobj(env, cl2lov(obj), lsm,
1552                                  lov_comp_index(index, stripeno));
1553         if (IS_ERR(subobj))
1554                 return PTR_ERR(subobj);
1555         /* If the output buffer is very large and the objects have many
1556          * extents we may need to loop on a single OST repeatedly */
1557         do {
1558                 if (fiemap->fm_extent_count > 0) {
1559                         /* Don't get too many extents. */
1560                         if (fs->fs_cur_extent + fs->fs_cnt_need >
1561                             fiemap->fm_extent_count)
1562                                 fs->fs_cnt_need = fiemap->fm_extent_count -
1563                                                   fs->fs_cur_extent;
1564                 }
1565
1566                 lun_start += len_mapped_single_call;
1567                 fs->fs_fm->fm_length = req_fm_len - len_mapped_single_call;
1568                 req_fm_len = fs->fs_fm->fm_length;
1569                 /**
1570                  * If we've collected enough extent map, we'd request 1 more,
1571                  * to see whether we coincidentally finished all available
1572                  * extent map, so that FIEMAP_EXTENT_LAST would be set.
1573                  */
1574                 fs->fs_fm->fm_extent_count = fs->fs_enough ?
1575                                              1 : fs->fs_cnt_need;
1576                 fs->fs_fm->fm_mapped_extents = 0;
1577                 fs->fs_fm->fm_flags = fiemap->fm_flags;
1578
1579                 ost_index = lsme->lsme_oinfo[stripeno]->loi_ost_idx;
1580
1581                 if (ost_index < 0 || ost_index >= lov->desc.ld_tgt_count)
1582                         GOTO(obj_put, rc = -EINVAL);
1583                 /* If OST is inactive, return extent with UNKNOWN flag. */
1584                 if (!lov->lov_tgts[ost_index]->ltd_active) {
1585                         fs->fs_fm->fm_flags |= FIEMAP_EXTENT_LAST;
1586                         fs->fs_fm->fm_mapped_extents = 1;
1587
1588                         fm_ext[0].fe_logical = lun_start;
1589                         fm_ext[0].fe_length = obd_object_end - lun_start;
1590                         fm_ext[0].fe_flags |= FIEMAP_EXTENT_UNKNOWN;
1591
1592                         goto inactive_tgt;
1593                 }
1594
1595                 fs->fs_fm->fm_start = lun_start;
1596                 fs->fs_fm->fm_flags &= ~FIEMAP_FLAG_DEVICE_ORDER;
1597                 memcpy(&fmkey->lfik_fiemap, fs->fs_fm, sizeof(*fs->fs_fm));
1598                 *buflen = fiemap_count_to_size(fs->fs_fm->fm_extent_count);
1599
1600                 rc = cl_object_fiemap(env, subobj, fmkey, fs->fs_fm, buflen);
1601                 if (rc != 0)
1602                         GOTO(obj_put, rc);
1603 inactive_tgt:
1604                 ext_count = fs->fs_fm->fm_mapped_extents;
1605                 if (ext_count == 0) {
1606                         ost_done = true;
1607                         fs->fs_device_done = true;
1608                         /* If last stripe has hold at the end,
1609                          * we need to return */
1610                         if (stripeno == fs->fs_last_stripe) {
1611                                 fiemap->fm_mapped_extents = 0;
1612                                 fs->fs_finish_stripe = true;
1613                                 GOTO(obj_put, rc);
1614                         }
1615                         break;
1616                 } else if (fs->fs_enough) {
1617                         /*
1618                          * We've collected enough extents and there are
1619                          * more extents after it.
1620                          */
1621                         GOTO(obj_put, rc);
1622                 }
1623
1624                 /* If we just need num of extents, got to next device */
1625                 if (fiemap->fm_extent_count == 0) {
1626                         fs->fs_cur_extent += ext_count;
1627                         break;
1628                 }
1629
1630                 /* prepare to copy retrived map extents */
1631                 len_mapped_single_call = fm_ext[ext_count - 1].fe_logical +
1632                                          fm_ext[ext_count - 1].fe_length -
1633                                          lun_start;
1634
1635                 /* Have we finished mapping on this device? */
1636                 if (req_fm_len <= len_mapped_single_call) {
1637                         ost_done = true;
1638                         fs->fs_device_done = true;
1639                 }
1640
1641                 /* Clear the EXTENT_LAST flag which can be present on
1642                  * the last extent */
1643                 if (fm_ext[ext_count - 1].fe_flags & FIEMAP_EXTENT_LAST)
1644                         fm_ext[ext_count - 1].fe_flags &= ~FIEMAP_EXTENT_LAST;
1645                 if (lov_stripe_size(lsm, index,
1646                                     fm_ext[ext_count - 1].fe_logical +
1647                                     fm_ext[ext_count - 1].fe_length,
1648                                     stripeno) >= fmkey->lfik_oa.o_size) {
1649                         ost_eof = true;
1650                         fs->fs_device_done = true;
1651                 }
1652
1653                 fiemap_prepare_and_copy_exts(fiemap, fm_ext, ost_index,
1654                                              ext_count, fs->fs_cur_extent);
1655                 fs->fs_cur_extent += ext_count;
1656
1657                 /* Ran out of available extents? */
1658                 if (fs->fs_cur_extent >= fiemap->fm_extent_count)
1659                         fs->fs_enough = true;
1660         } while (!ost_done && !ost_eof);
1661
1662         if (stripeno == fs->fs_last_stripe)
1663                 fs->fs_finish_stripe = true;
1664 obj_put:
1665         cl_object_put(env, subobj);
1666
1667         return rc;
1668 }
1669
1670 /**
1671  * Break down the FIEMAP request and send appropriate calls to individual OSTs.
1672  * This also handles the restarting of FIEMAP calls in case mapping overflows
1673  * the available number of extents in single call.
1674  *
1675  * \param env [in]              lustre environment
1676  * \param obj [in]              file object
1677  * \param fmkey [in]            fiemap request header and other info
1678  * \param fiemap [out]          fiemap buffer holding retrived map extents
1679  * \param buflen [in/out]       max buffer length of @fiemap, when iterate
1680  *                              each OST, it is used to limit max map needed
1681  * \retval 0    success
1682  * \retval < 0  error
1683  */
1684 static int lov_object_fiemap(const struct lu_env *env, struct cl_object *obj,
1685                              struct ll_fiemap_info_key *fmkey,
1686                              struct fiemap *fiemap, size_t *buflen)
1687 {
1688         struct lov_stripe_md_entry *lsme;
1689         struct lov_stripe_md *lsm;
1690         struct fiemap *fm_local = NULL;
1691         loff_t whole_start;
1692         loff_t whole_end;
1693         int entry;
1694         int start_entry;
1695         int end_entry;
1696         int cur_stripe = 0;
1697         int stripe_count;
1698         unsigned int buffer_size = FIEMAP_BUFFER_SIZE;
1699         int rc = 0;
1700         struct fiemap_state fs = { 0 };
1701         ENTRY;
1702
1703         lsm = lov_lsm_addref(cl2lov(obj));
1704         if (lsm == NULL)
1705                 RETURN(-ENODATA);
1706
1707         if (!(fiemap->fm_flags & FIEMAP_FLAG_DEVICE_ORDER)) {
1708                 /**
1709                  * If the entry count > 1 or stripe_count > 1 and the
1710                  * application does not understand DEVICE_ORDER flag,
1711                  * it cannot interpret the extents correctly.
1712                  */
1713                 if (lsm->lsm_entry_count > 1 ||
1714                     (lsm->lsm_entry_count == 1 &&
1715                      lsm->lsm_entries[0]->lsme_stripe_count > 1))
1716                         GOTO(out_lsm, rc = -ENOTSUPP);
1717         }
1718
1719         /* No support for DOM layout yet. */
1720         if (lsme_is_dom(lsm->lsm_entries[0]))
1721                 GOTO(out_lsm, rc = -ENOTSUPP);
1722
1723         if (lsm->lsm_is_released) {
1724                 if (fiemap->fm_start < fmkey->lfik_oa.o_size) {
1725                         /**
1726                          * released file, return a minimal FIEMAP if
1727                          * request fits in file-size.
1728                          */
1729                         fiemap->fm_mapped_extents = 1;
1730                         fiemap->fm_extents[0].fe_logical = fiemap->fm_start;
1731                         if (fiemap->fm_start + fiemap->fm_length <
1732                             fmkey->lfik_oa.o_size)
1733                                 fiemap->fm_extents[0].fe_length =
1734                                         fiemap->fm_length;
1735                         else
1736                                 fiemap->fm_extents[0].fe_length =
1737                                         fmkey->lfik_oa.o_size -
1738                                         fiemap->fm_start;
1739                         fiemap->fm_extents[0].fe_flags |=
1740                                 FIEMAP_EXTENT_UNKNOWN | FIEMAP_EXTENT_LAST;
1741                 }
1742                 GOTO(out_lsm, rc = 0);
1743         }
1744
1745         /* buffer_size is small to hold fm_extent_count of extents. */
1746         if (fiemap_count_to_size(fiemap->fm_extent_count) < buffer_size)
1747                 buffer_size = fiemap_count_to_size(fiemap->fm_extent_count);
1748
1749         OBD_ALLOC_LARGE(fm_local, buffer_size);
1750         if (fm_local == NULL)
1751                 GOTO(out_lsm, rc = -ENOMEM);
1752
1753         /**
1754          * Requested extent count exceeds the fiemap buffer size, shrink our
1755          * ambition.
1756          */
1757         if (fiemap_count_to_size(fiemap->fm_extent_count) > *buflen)
1758                 fiemap->fm_extent_count = fiemap_size_to_count(*buflen);
1759         if (fiemap->fm_extent_count == 0)
1760                 fs.fs_cnt_need = 0;
1761
1762         fs.fs_enough = false;
1763         fs.fs_cur_extent = 0;
1764         fs.fs_fm = fm_local;
1765         fs.fs_cnt_need = fiemap_size_to_count(buffer_size);
1766
1767         whole_start = fiemap->fm_start;
1768         /* whole_start is beyond the end of the file */
1769         if (whole_start > fmkey->lfik_oa.o_size)
1770                 GOTO(out_fm_local, rc = -EINVAL);
1771         whole_end = (fiemap->fm_length == OBD_OBJECT_EOF) ?
1772                                         fmkey->lfik_oa.o_size :
1773                                         whole_start + fiemap->fm_length - 1;
1774         /**
1775          * If fiemap->fm_length != OBD_OBJECT_EOF but whole_end exceeds file
1776          * size
1777          */
1778         if (whole_end > fmkey->lfik_oa.o_size)
1779                 whole_end = fmkey->lfik_oa.o_size;
1780
1781         start_entry = lov_lsm_entry(lsm, whole_start);
1782         end_entry = lov_lsm_entry(lsm, whole_end);
1783         if (end_entry == -1)
1784                 end_entry = lsm->lsm_entry_count - 1;
1785
1786         if (start_entry == -1 || end_entry == -1)
1787                 GOTO(out_fm_local, rc = -EINVAL);
1788
1789         for (entry = start_entry; entry <= end_entry; entry++) {
1790                 lsme = lsm->lsm_entries[entry];
1791
1792                 if (!lsme_inited(lsme))
1793                         break;
1794
1795                 if (entry == start_entry)
1796                         fs.fs_ext.e_start = whole_start;
1797                 else
1798                         fs.fs_ext.e_start = lsme->lsme_extent.e_start;
1799                 if (entry == end_entry)
1800                         fs.fs_ext.e_end = whole_end;
1801                 else
1802                         fs.fs_ext.e_end = lsme->lsme_extent.e_end - 1;
1803                 fs.fs_length = fs.fs_ext.e_end - fs.fs_ext.e_start + 1;
1804
1805                 /* Calculate start stripe, last stripe and length of mapping */
1806                 fs.fs_start_stripe = lov_stripe_number(lsm, entry,
1807                                                        fs.fs_ext.e_start);
1808                 fs.fs_last_stripe = fiemap_calc_last_stripe(lsm, entry,
1809                                         &fs.fs_ext, fs.fs_start_stripe,
1810                                         &stripe_count);
1811                 fs.fs_end_offset = fiemap_calc_fm_end_offset(fiemap, lsm, entry,
1812                                         &fs.fs_ext, &fs.fs_start_stripe);
1813                 /* Check each stripe */
1814                 for (cur_stripe = fs.fs_start_stripe; stripe_count > 0;
1815                      --stripe_count,
1816                      cur_stripe = (cur_stripe + 1) % lsme->lsme_stripe_count) {
1817                         rc = fiemap_for_stripe(env, obj, lsm, fiemap, buflen,
1818                                                fmkey, entry, cur_stripe, &fs);
1819                         if (rc < 0)
1820                                 GOTO(out_fm_local, rc);
1821                         if (fs.fs_enough)
1822                                 GOTO(finish, rc);
1823                         if (fs.fs_finish_stripe)
1824                                 break;
1825                 } /* for each stripe */
1826         } /* for covering layout component */
1827         /*
1828          * We've traversed all components, set @entry to the last component
1829          * entry, it's for the last stripe check.
1830          */
1831         entry--;
1832 finish:
1833         /* Indicate that we are returning device offsets unless file just has
1834          * single stripe */
1835         if (lsm->lsm_entry_count > 1 ||
1836             (lsm->lsm_entry_count == 1 &&
1837              lsm->lsm_entries[0]->lsme_stripe_count > 1))
1838                 fiemap->fm_flags |= FIEMAP_FLAG_DEVICE_ORDER;
1839
1840         if (fiemap->fm_extent_count == 0)
1841                 goto skip_last_device_calc;
1842
1843         /* Check if we have reached the last stripe and whether mapping for that
1844          * stripe is done. */
1845         if ((cur_stripe == fs.fs_last_stripe) && fs.fs_device_done)
1846                 fiemap->fm_extents[fs.fs_cur_extent - 1].fe_flags |=
1847                                                              FIEMAP_EXTENT_LAST;
1848 skip_last_device_calc:
1849         fiemap->fm_mapped_extents = fs.fs_cur_extent;
1850 out_fm_local:
1851         OBD_FREE_LARGE(fm_local, buffer_size);
1852
1853 out_lsm:
1854         lov_lsm_put(lsm);
1855         return rc;
1856 }
1857
1858 static int lov_object_getstripe(const struct lu_env *env, struct cl_object *obj,
1859                                 struct lov_user_md __user *lum, size_t size)
1860 {
1861         struct lov_object       *lov = cl2lov(obj);
1862         struct lov_stripe_md    *lsm;
1863         int                     rc = 0;
1864         ENTRY;
1865
1866         lsm = lov_lsm_addref(lov);
1867         if (lsm == NULL)
1868                 RETURN(-ENODATA);
1869
1870         rc = lov_getstripe(env, cl2lov(obj), lsm, lum, size);
1871         lov_lsm_put(lsm);
1872         RETURN(rc);
1873 }
1874
1875 static int lov_object_layout_get(const struct lu_env *env,
1876                                  struct cl_object *obj,
1877                                  struct cl_layout *cl)
1878 {
1879         struct lov_object *lov = cl2lov(obj);
1880         struct lov_stripe_md *lsm = lov_lsm_addref(lov);
1881         struct lu_buf *buf = &cl->cl_buf;
1882         ssize_t rc;
1883         ENTRY;
1884
1885         if (lsm == NULL) {
1886                 cl->cl_size = 0;
1887                 cl->cl_layout_gen = CL_LAYOUT_GEN_EMPTY;
1888
1889                 RETURN(0);
1890         }
1891
1892         cl->cl_size = lov_comp_md_size(lsm);
1893         cl->cl_layout_gen = lsm->lsm_layout_gen;
1894         cl->cl_dom_comp_size = 0;
1895         if (lsm_is_composite(lsm->lsm_magic)) {
1896                 struct lov_stripe_md_entry *lsme = lsm->lsm_entries[0];
1897
1898                 cl->cl_is_composite = true;
1899
1900                 if (lsme_is_dom(lsme))
1901                         cl->cl_dom_comp_size = lsme->lsme_extent.e_end;
1902         } else {
1903                 cl->cl_is_composite = false;
1904         }
1905
1906         rc = lov_lsm_pack(lsm, buf->lb_buf, buf->lb_len);
1907         lov_lsm_put(lsm);
1908
1909         RETURN(rc < 0 ? rc : 0);
1910 }
1911
1912 static loff_t lov_object_maxbytes(struct cl_object *obj)
1913 {
1914         struct lov_object *lov = cl2lov(obj);
1915         struct lov_stripe_md *lsm = lov_lsm_addref(lov);
1916         loff_t maxbytes;
1917
1918         if (lsm == NULL)
1919                 return LLONG_MAX;
1920
1921         maxbytes = lsm->lsm_maxbytes;
1922
1923         lov_lsm_put(lsm);
1924
1925         return maxbytes;
1926 }
1927
1928 static const struct cl_object_operations lov_ops = {
1929         .coo_page_init    = lov_page_init,
1930         .coo_lock_init    = lov_lock_init,
1931         .coo_io_init      = lov_io_init,
1932         .coo_attr_get     = lov_attr_get,
1933         .coo_attr_update  = lov_attr_update,
1934         .coo_conf_set     = lov_conf_set,
1935         .coo_getstripe    = lov_object_getstripe,
1936         .coo_layout_get   = lov_object_layout_get,
1937         .coo_maxbytes     = lov_object_maxbytes,
1938         .coo_fiemap       = lov_object_fiemap,
1939 };
1940
1941 static const struct lu_object_operations lov_lu_obj_ops = {
1942         .loo_object_init      = lov_object_init,
1943         .loo_object_delete    = lov_object_delete,
1944         .loo_object_release   = NULL,
1945         .loo_object_free      = lov_object_free,
1946         .loo_object_print     = lov_object_print,
1947         .loo_object_invariant = NULL
1948 };
1949
1950 struct lu_object *lov_object_alloc(const struct lu_env *env,
1951                                    const struct lu_object_header *unused,
1952                                    struct lu_device *dev)
1953 {
1954         struct lov_object *lov;
1955         struct lu_object  *obj;
1956
1957         ENTRY;
1958         OBD_SLAB_ALLOC_PTR_GFP(lov, lov_object_kmem, GFP_NOFS);
1959         if (lov != NULL) {
1960                 obj = lov2lu(lov);
1961                 lu_object_init(obj, NULL, dev);
1962                 lov->lo_cl.co_ops = &lov_ops;
1963                 lov->lo_type = -1; /* invalid, to catch uninitialized type */
1964                 /*
1965                  * object io operation vector (cl_object::co_iop) is installed
1966                  * later in lov_object_init(), as different vectors are used
1967                  * for object with different layouts.
1968                  */
1969                 obj->lo_ops = &lov_lu_obj_ops;
1970         } else
1971                 obj = NULL;
1972         RETURN(obj);
1973 }
1974
1975 struct lov_stripe_md *lov_lsm_addref(struct lov_object *lov)
1976 {
1977         struct lov_stripe_md *lsm = NULL;
1978
1979         lov_conf_freeze(lov);
1980         if (lov->lo_lsm != NULL) {
1981                 lsm = lsm_addref(lov->lo_lsm);
1982                 CDEBUG(D_INODE, "lsm %p addref %d/%d by %p.\n",
1983                         lsm, atomic_read(&lsm->lsm_refc),
1984                         lov->lo_layout_invalid, current);
1985         }
1986         lov_conf_thaw(lov);
1987         return lsm;
1988 }
1989
1990 int lov_read_and_clear_async_rc(struct cl_object *clob)
1991 {
1992         struct lu_object *luobj;
1993         int rc = 0;
1994         ENTRY;
1995
1996         luobj = lu_object_locate(&cl_object_header(clob)->coh_lu,
1997                                  &lov_device_type);
1998         if (luobj != NULL) {
1999                 struct lov_object *lov = lu2lov(luobj);
2000
2001                 lov_conf_freeze(lov);
2002                 switch (lov->lo_type) {
2003                 case LLT_COMP: {
2004                         struct lov_stripe_md *lsm;
2005                         int i;
2006
2007                         lsm = lov->lo_lsm;
2008                         LASSERT(lsm != NULL);
2009                         for (i = 0; i < lsm->lsm_entry_count; i++) {
2010                                 struct lov_stripe_md_entry *lse =
2011                                                 lsm->lsm_entries[i];
2012                                 int j;
2013
2014                                 if (!lsme_inited(lse))
2015                                         break;
2016
2017                                 for (j = 0; j < lse->lsme_stripe_count; j++) {
2018                                         struct lov_oinfo *loi =
2019                                                         lse->lsme_oinfo[j];
2020
2021                                         if (lov_oinfo_is_dummy(loi))
2022                                                 continue;
2023
2024                                         if (loi->loi_ar.ar_rc && !rc)
2025                                                 rc = loi->loi_ar.ar_rc;
2026                                         loi->loi_ar.ar_rc = 0;
2027                                 }
2028                         }
2029                 }
2030                 case LLT_RELEASED:
2031                 case LLT_EMPTY:
2032                         break;
2033                 default:
2034                         LBUG();
2035                 }
2036                 lov_conf_thaw(lov);
2037         }
2038         RETURN(rc);
2039 }
2040 EXPORT_SYMBOL(lov_read_and_clear_async_rc);
2041
2042 /** @} lov */