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         attr->cat_kms = attr->cat_size > loi->loi_kms ? attr->cat_size :
 453                                                         loi->loi_kms;
 454         dom->lo_dom_r0.lo_attr_valid = 1;
 455         *lov_attr = attr;
 456
 457         return 0;
 458 }
 459
 460 /**
 461  * Lookup FLD to get MDS index of the given DOM object FID.
 462  *
 463  * \param[in]  ld       LOV device
 464  * \param[in]  fid      FID to lookup
 465  * \param[out] nr       index in MDC array to return back
 466  *
 467  * \retval              0 and \a mds filled with MDS index if successful
 468  * \retval              negative value on error
 469  */
 470 static int lov_fld_lookup(struct lov_device *ld, const struct lu_fid *fid,
 471                           __u32 *nr)
 472 {
 473         __u32 mds_idx;
 474         int i, rc;
 475
 476         ENTRY;
 477
 478         rc = fld_client_lookup(&ld->ld_lmv->u.lmv.lmv_fld, fid_seq(fid),
 479                                &mds_idx, LU_SEQ_RANGE_MDT, NULL);
 480         if (rc) {
 481                 CERROR("%s: error while looking for mds number. Seq %#llx"
 482                        ", err = %d\n", lu_dev_name(cl2lu_dev(&ld->ld_cl)),
 483                        fid_seq(fid), rc);
 484                 RETURN(rc);
 485         }
 486
 487         CDEBUG(D_INODE, "FLD lookup got mds #%x for fid="DFID"\n",
 488                mds_idx, PFID(fid));
 489
 490         /* find proper MDC device in the array */
 491         for (i = 0; i < ld->ld_md_tgts_nr; i++) {
 492                 if (ld->ld_md_tgts[i].ldm_mdc != NULL &&
 493                     ld->ld_md_tgts[i].ldm_idx == mds_idx)
 494                         break;
 495         }
 496
 497         if (i == ld->ld_md_tgts_nr) {
 498                 CERROR("%s: cannot find corresponding MDC device for mds #%x "
 499                        "for fid="DFID"\n", lu_dev_name(cl2lu_dev(&ld->ld_cl)),
 500                        mds_idx, PFID(fid));
 501                 rc = -EINVAL;
 502         } else {
 503                 *nr = i;
 504         }
 505         RETURN(rc);
 506 }
 507
 508 /**
 509  * Implementation of lov_comp_layout_entry_ops::lco_init for DOM object.
 510  *
 511  * Init the DOM object for the first time. It prepares also RAID0 entry
 512  * for it to use in common methods with ordinary RAID0 layout entries.
 513  *
 514  * \param[in] env       execution environment
 515  * \param[in] dev       LOV device
 516  * \param[in] lov       LOV object
 517  * \param[in] index     Composite layout entry index in LSM
 518  * \param[in] lle       Composite LOV layout entry
 519  */
 520 static int lov_init_dom(const struct lu_env *env, struct lov_device *dev,
 521                         struct lov_object *lov, unsigned int index,
 522                         const struct cl_object_conf *conf,
 523                         struct lov_layout_entry *lle)
 524 {
 525         struct lov_thread_info *lti = lov_env_info(env);
 526         struct lov_stripe_md_entry *lsme = lov_lse(lov, index);
 527         struct cl_object *clo;
 528         struct lu_object *o = lov2lu(lov);
 529         const struct lu_fid *fid = lu_object_fid(o);
 530         struct cl_device *mdcdev;
 531         struct lov_oinfo *loi = NULL;
 532         struct cl_object_conf *sconf = &lti->lti_stripe_conf;
 533         struct inode *inode = conf->coc_inode;
 534
 535         int rc;
 536         __u32 idx = 0;
 537
 538         ENTRY;
 539
 540         LASSERT(index == 0);
 541
 542         /* find proper MDS device */
 543         rc = lov_fld_lookup(dev, fid, &idx);
 544         if (rc)
 545                 RETURN(rc);
 546
 547         LASSERTF(dev->ld_md_tgts[idx].ldm_mdc != NULL,
 548                  "LOV md target[%u] is NULL\n", idx);
 549
 550         /* check lsm is DOM, more checks are needed */
 551         LASSERT(lsme->lsme_stripe_count == 0);
 552
 553         /*
 554          * Create lower cl_objects.
 555          */
 556         mdcdev = dev->ld_md_tgts[idx].ldm_mdc;
 557
 558         LASSERTF(mdcdev != NULL, "non-initialized mdc subdev\n");
 559
 560         /* DoM object has no oinfo in LSM entry, create it exclusively */
 561         OBD_SLAB_ALLOC_PTR_GFP(loi, lov_oinfo_slab, GFP_NOFS);
 562         if (loi == NULL)
 563                 RETURN(-ENOMEM);
 564
 565         fid_to_ostid(lu_object_fid(lov2lu(lov)), &loi->loi_oi);
 566         /* Initialize lvb structure */
 567         loi->loi_lvb.lvb_mtime = LTIME_S(inode->i_mtime);
 568         loi->loi_lvb.lvb_atime = LTIME_S(inode->i_atime);
 569         loi->loi_lvb.lvb_ctime = LTIME_S(inode->i_ctime);
 570         loi->loi_lvb.lvb_blocks = inode->i_blocks;
 571         loi->loi_lvb.lvb_size = i_size_read(inode);
 572         if (loi->loi_lvb.lvb_size > lsme->lsme_stripe_size)
 573                 loi->loi_lvb.lvb_size = lsme->lsme_stripe_size;
 574         loi_kms_set(loi, loi->loi_lvb.lvb_size);
 575
 576         sconf->u.coc_oinfo = loi;
 577 again:
 578         clo = lov_sub_find(env, mdcdev, fid, sconf);
 579         if (IS_ERR(clo))
 580                 GOTO(out, rc = PTR_ERR(clo));
 581
 582         rc = lov_init_sub(env, lov, clo, loi, lov_comp_index(index, 0));
 583         if (rc == -EAGAIN) /* try again */
 584                 goto again;
 585         else if (rc != 0)
 586                 GOTO(out, rc);
 587
 588         lle->lle_dom.lo_dom = cl2lovsub(clo);
 589         spin_lock_init(&lle->lle_dom.lo_dom_r0.lo_sub_lock);
 590         lle->lle_dom.lo_dom_r0.lo_nr = 1;
 591         lle->lle_dom.lo_dom_r0.lo_sub = &lle->lle_dom.lo_dom;
 592         lle->lle_dom.lo_loi = loi;
 593
 594         rc = lov_page_slice_fixup(lov, clo);
 595         RETURN(rc);
 596
 597 out:
 598         if (loi != NULL)
 599                 OBD_SLAB_FREE_PTR(loi, lov_oinfo_slab);
 600         return rc;
 601 }
 602
 603 /**
 604  * Implementation of lov_layout_operations::llo_fini for DOM object.
 605  *
 606  * Finish the DOM object and free related memory.
 607  *
 608  * \param[in] env       execution environment
 609  * \param[in] lov       LOV object
 610  * \param[in] state     LOV layout state
 611  */
 612 static void lov_fini_dom(const struct lu_env *env,
 613                          struct lov_layout_entry *lle)
 614 {
 615         if (lle->lle_dom.lo_dom != NULL)
 616                 lle->lle_dom.lo_dom = NULL;
 617         if (lle->lle_dom.lo_loi != NULL)
 618                 OBD_SLAB_FREE_PTR(lle->lle_dom.lo_loi, lov_oinfo_slab);
 619 }
 620
 621 static struct lov_comp_layout_entry_ops dom_ops = {
 622         .lco_init = lov_init_dom,
 623         .lco_fini = lov_fini_dom,
 624         .lco_getattr = lov_attr_get_dom,
 625 };
 626
 627 static int lov_init_composite(const struct lu_env *env, struct lov_device *dev,
 628                               struct lov_object *lov, struct lov_stripe_md *lsm,
 629                               const struct cl_object_conf *conf,
 630                               union lov_layout_state *state)
 631 {
 632         struct lov_layout_composite *comp = &state->composite;
 633         struct lov_layout_entry *lle;
 634         unsigned int entry_count;
 635         unsigned int psz = 0;
 636         int result = 0;
 637         int i;
 638
 639         ENTRY;
 640
 641         LASSERT(lsm->lsm_entry_count > 0);
 642         LASSERT(lov->lo_lsm == NULL);
 643         lov->lo_lsm = lsm_addref(lsm);
 644         lov->lo_layout_invalid = true;
 645
 646         entry_count = lsm->lsm_entry_count;
 647         comp->lo_entry_count = entry_count;
 648
 649         OBD_ALLOC(comp->lo_entries, entry_count * sizeof(*comp->lo_entries));
 650         if (comp->lo_entries == NULL)
 651                 RETURN(-ENOMEM);
 652
 653         /* Initiate all entry types and extents data at first */
 654         for (i = 0; i < entry_count; i++) {
 655                 lle = &comp->lo_entries[i];
 656
 657                 lle->lle_type = lov_entry_type(lsm->lsm_entries[i]);
 658                 switch (lle->lle_type) {
 659                 case LOV_PATTERN_RAID0:
 660                         lle->lle_comp_ops = &raid0_ops;
 661                         break;
 662                 case LOV_PATTERN_MDT:
 663                         lle->lle_comp_ops = &dom_ops;
 664                         break;
 665                 default:
 666                         CERROR("%s: unknown composite layout entry type %i\n",
 667                                lov2obd(dev->ld_lov)->obd_name,
 668                                lsm->lsm_entries[i]->lsme_pattern);
 669                         dump_lsm(D_ERROR, lsm);
 670                         RETURN(-EIO);
 671                 }
 672                 lle->lle_extent = lsm->lsm_entries[i]->lsme_extent;
 673         }
 674
 675         i = 0;
 676         lov_foreach_layout_entry(lov, lle) {
 677                 /**
 678                  * If the component has not been init-ed on MDS side, for
 679                  * PFL layout, we'd know that the components beyond this one
 680                  * will be dynamically init-ed later on file write/trunc ops.
 681                  */
 682                 if (lsm_entry_inited(lsm, i)) {
 683                         result = lle->lle_comp_ops->lco_init(env, dev, lov, i,
 684                                                              conf, lle);
 685                         if (result < 0)
 686                                 break;
 687
 688                         LASSERT(ergo(psz > 0, psz == result));
 689                         psz = result;
 690                 }
 691                 i++;
 692         }
 693         if (psz > 0)
 694                 cl_object_header(&lov->lo_cl)->coh_page_bufsize += psz;
 695
 696         return result > 0 ? 0 : result;
 697 }
 698
 699 static int lov_init_empty(const struct lu_env *env, struct lov_device *dev,
 700                           struct lov_object *lov, struct lov_stripe_md *lsm,
 701                           const struct cl_object_conf *conf,
 702                           union lov_layout_state *state)
 703 {
 704         return 0;
 705 }
 706
 707 static int lov_init_released(const struct lu_env *env,
 708                              struct lov_device *dev, struct lov_object *lov,
 709                              struct lov_stripe_md *lsm,
 710                              const struct cl_object_conf *conf,
 711                              union lov_layout_state *state)
 712 {
 713         LASSERT(lsm != NULL);
 714         LASSERT(lsm->lsm_is_released);
 715         LASSERT(lov->lo_lsm == NULL);
 716
 717         lov->lo_lsm = lsm_addref(lsm);
 718         return 0;
 719 }
 720
 721 static int lov_delete_empty(const struct lu_env *env, struct lov_object *lov,
 722                             union lov_layout_state *state)
 723 {
 724         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 725
 726         lov_layout_wait(env, lov);
 727         return 0;
 728 }
 729
 730 static int lov_delete_composite(const struct lu_env *env,
 731                                 struct lov_object *lov,
 732                                 union lov_layout_state *state)
 733 {
 734         struct lov_layout_entry *entry;
 735         struct lov_layout_composite *comp = &state->composite;
 736
 737         ENTRY;
 738
 739         dump_lsm(D_INODE, lov->lo_lsm);
 740
 741         lov_layout_wait(env, lov);
 742         if (comp->lo_entries)
 743                 lov_foreach_layout_entry(lov, entry)
 744                         lov_delete_raid0(env, lov, entry);
 745
 746         RETURN(0);
 747 }
 748
 749 static void lov_fini_empty(const struct lu_env *env, struct lov_object *lov,
 750                            union lov_layout_state *state)
 751 {
 752         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 753 }
 754
 755 static void lov_fini_composite(const struct lu_env *env,
 756                                struct lov_object *lov,
 757                                union lov_layout_state *state)
 758 {
 759         struct lov_layout_composite *comp = &state->composite;
 760         ENTRY;
 761
 762         if (comp->lo_entries != NULL) {
 763                 struct lov_layout_entry *entry;
 764
 765                 lov_foreach_layout_entry(lov, entry)
 766                         entry->lle_comp_ops->lco_fini(env, entry);
 767
 768                 OBD_FREE(comp->lo_entries,
 769                          comp->lo_entry_count * sizeof(*comp->lo_entries));
 770                 comp->lo_entries = NULL;
 771         }
 772
 773         dump_lsm(D_INODE, lov->lo_lsm);
 774         lov_free_memmd(&lov->lo_lsm);
 775
 776         EXIT;
 777 }
 778
 779 static void lov_fini_released(const struct lu_env *env, struct lov_object *lov,
 780                                 union lov_layout_state *state)
 781 {
 782         ENTRY;
 783         dump_lsm(D_INODE, lov->lo_lsm);
 784         lov_free_memmd(&lov->lo_lsm);
 785         EXIT;
 786 }
 787
 788 static int lov_print_empty(const struct lu_env *env, void *cookie,
 789                            lu_printer_t p, const struct lu_object *o)
 790 {
 791         (*p)(env, cookie, "empty %d\n", lu2lov(o)->lo_layout_invalid);
 792         return 0;
 793 }
 794
 795 static int lov_print_composite(const struct lu_env *env, void *cookie,
 796                                lu_printer_t p, const struct lu_object *o)
 797 {
 798         struct lov_object *lov = lu2lov(o);
 799         struct lov_stripe_md *lsm = lov->lo_lsm;
 800         int i;
 801
 802         (*p)(env, cookie, "entries: %d, %s, lsm{%p 0x%08X %d %u}:\n",
 803              lsm->lsm_entry_count,
 804              lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 805              lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 806              lsm->lsm_layout_gen);
 807
 808         for (i = 0; i < lsm->lsm_entry_count; i++) {
 809                 struct lov_stripe_md_entry *lse = lsm->lsm_entries[i];
 810                 struct lov_layout_entry *lle = lov_entry(lov, i);
 811
 812                 (*p)(env, cookie,
 813                      DEXT ": { 0x%08X, %u, %#x, %u, %#x, %u, %u }\n",
 814                      PEXT(&lse->lsme_extent), lse->lsme_magic,
 815                      lse->lsme_id, lse->lsme_pattern, lse->lsme_layout_gen,
 816                      lse->lsme_flags, lse->lsme_stripe_count,
 817                      lse->lsme_stripe_size);
 818                 lov_print_raid0(env, cookie, p, lle);
 819         }
 820
 821         return 0;
 822 }
 823
 824 static int lov_print_released(const struct lu_env *env, void *cookie,
 825                                 lu_printer_t p, const struct lu_object *o)
 826 {
 827         struct lov_object       *lov = lu2lov(o);
 828         struct lov_stripe_md    *lsm = lov->lo_lsm;
 829
 830         (*p)(env, cookie,
 831                 "released: %s, lsm{%p 0x%08X %d %u}:\n",
 832                 lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 833                 lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 834                 lsm->lsm_layout_gen);
 835         return 0;
 836 }
 837
 838 /**
 839  * Implements cl_object_operations::coo_attr_get() method for an object
 840  * without stripes (LLT_EMPTY layout type).
 841  *
 842  * The only attributes this layer is authoritative in this case is
 843  * cl_attr::cat_blocks---it's 0.
 844  */
 845 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
 846                               struct cl_attr *attr)
 847 {
 848         attr->cat_blocks = 0;
 849         return 0;
 850 }
 851
 852 static int lov_attr_get_composite(const struct lu_env *env,
 853                                   struct cl_object *obj,
 854                                   struct cl_attr *attr)
 855 {
 856         struct lov_object       *lov = cl2lov(obj);
 857         struct lov_layout_entry *entry;
 858         int                      result = 0;
 859         int                      index = 0;
 860
 861         ENTRY;
 862
 863         attr->cat_size = 0;
 864         attr->cat_blocks = 0;
 865         lov_foreach_layout_entry(lov, entry) {
 866                 struct cl_attr *lov_attr = NULL;
 867
 868                 /* PFL: This component has not been init-ed. */
 869                 if (!lsm_entry_inited(lov->lo_lsm, index))
 870                         break;
 871
 872                 result = entry->lle_comp_ops->lco_getattr(env, lov, index,
 873                                                           entry, &lov_attr);
 874                 if (result < 0)
 875                         RETURN(result);
 876
 877                 index++;
 878
 879                 if (lov_attr == NULL)
 880                         continue;
 881
 882                 /* merge results */
 883                 attr->cat_blocks += lov_attr->cat_blocks;
 884                 if (attr->cat_size < lov_attr->cat_size)
 885                         attr->cat_size = lov_attr->cat_size;
 886                 if (attr->cat_kms < lov_attr->cat_kms)
 887                         attr->cat_kms = lov_attr->cat_kms;
 888                 if (attr->cat_atime < lov_attr->cat_atime)
 889                         attr->cat_atime = lov_attr->cat_atime;
 890                 if (attr->cat_ctime < lov_attr->cat_ctime)
 891                         attr->cat_ctime = lov_attr->cat_ctime;
 892                 if (attr->cat_mtime < lov_attr->cat_mtime)
 893                         attr->cat_mtime = lov_attr->cat_mtime;
 894         }
 895         RETURN(0);
 896 }
 897
 898 const static struct lov_layout_operations lov_dispatch[] = {
 899         [LLT_EMPTY] = {
 900                 .llo_init      = lov_init_empty,
 901                 .llo_delete    = lov_delete_empty,
 902                 .llo_fini      = lov_fini_empty,
 903                 .llo_print     = lov_print_empty,
 904                 .llo_page_init = lov_page_init_empty,
 905                 .llo_lock_init = lov_lock_init_empty,
 906                 .llo_io_init   = lov_io_init_empty,
 907                 .llo_getattr   = lov_attr_get_empty,
 908         },
 909         [LLT_RELEASED] = {
 910                 .llo_init      = lov_init_released,
 911                 .llo_delete    = lov_delete_empty,
 912                 .llo_fini      = lov_fini_released,
 913                 .llo_print     = lov_print_released,
 914                 .llo_page_init = lov_page_init_empty,
 915                 .llo_lock_init = lov_lock_init_empty,
 916                 .llo_io_init   = lov_io_init_released,
 917                 .llo_getattr   = lov_attr_get_empty,
 918         },
 919         [LLT_COMP] = {
 920                 .llo_init      = lov_init_composite,
 921                 .llo_delete    = lov_delete_composite,
 922                 .llo_fini      = lov_fini_composite,
 923                 .llo_print     = lov_print_composite,
 924                 .llo_page_init = lov_page_init_composite,
 925                 .llo_lock_init = lov_lock_init_composite,
 926                 .llo_io_init   = lov_io_init_composite,
 927                 .llo_getattr   = lov_attr_get_composite,
 928         },
 929 };
 930
 931 /**
 932  * Performs a double-dispatch based on the layout type of an object.
 933  */
 934 #define LOV_2DISPATCH_NOLOCK(obj, op, ...)              \
 935 ({                                                      \
 936         struct lov_object *__obj = (obj);               \
 937         enum lov_layout_type __llt;                     \
 938                                                         \
 939         __llt = __obj->lo_type;                         \
 940         LASSERT(__llt < ARRAY_SIZE(lov_dispatch));      \
 941         lov_dispatch[__llt].op(__VA_ARGS__);            \
 942 })
 943
 944 /**
 945  * Return lov_layout_type associated with a given lsm
 946  */
 947 static enum lov_layout_type lov_type(struct lov_stripe_md *lsm)
 948 {
 949         if (lsm == NULL)
 950                 return LLT_EMPTY;
 951
 952         if (lsm->lsm_is_released)
 953                 return LLT_RELEASED;
 954
 955         if (lsm->lsm_magic == LOV_MAGIC_V1 ||
 956             lsm->lsm_magic == LOV_MAGIC_V3 ||
 957             lsm->lsm_magic == LOV_MAGIC_COMP_V1)
 958                 return LLT_COMP;
 959
 960         return LLT_EMPTY;
 961 }
 962
 963 static inline void lov_conf_freeze(struct lov_object *lov)
 964 {
 965         CDEBUG(D_INODE, "To take share lov(%p) owner %p/%p\n",
 966                 lov, lov->lo_owner, current);
 967         if (lov->lo_owner != current)
 968                 down_read(&lov->lo_type_guard);
 969 }
 970
 971 static inline void lov_conf_thaw(struct lov_object *lov)
 972 {
 973         CDEBUG(D_INODE, "To release share lov(%p) owner %p/%p\n",
 974                 lov, lov->lo_owner, current);
 975         if (lov->lo_owner != current)
 976                 up_read(&lov->lo_type_guard);
 977 }
 978
 979 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...)                       \
 980 ({                                                                      \
 981         struct lov_object                      *__obj = (obj);          \
 982         int                                     __lock = !!(lock);      \
 983         typeof(lov_dispatch[0].op(__VA_ARGS__)) __result;               \
 984                                                                         \
 985         if (__lock)                                                     \
 986                 lov_conf_freeze(__obj);                                 \
 987         __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__);          \
 988         if (__lock)                                                     \
 989                 lov_conf_thaw(__obj);                                   \
 990         __result;                                                       \
 991 })
 992
 993 /**
 994  * Performs a locked double-dispatch based on the layout type of an object.
 995  */
 996 #define LOV_2DISPATCH(obj, op, ...)                     \
 997         LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
 998
 999 #define LOV_2DISPATCH_VOID(obj, op, ...)                                \
1000 do {                                                                    \
1001         struct lov_object                      *__obj = (obj);          \
1002         enum lov_layout_type                    __llt;                  \
1003                                                                         \
1004         lov_conf_freeze(__obj);                                         \
1005         __llt = __obj->lo_type;                                         \
1006         LASSERT(__llt < ARRAY_SIZE(lov_dispatch));                      \
1007         lov_dispatch[__llt].op(__VA_ARGS__);                            \
1008         lov_conf_thaw(__obj);                                           \
1009 } while (0)
1010
1011 static void lov_conf_lock(struct lov_object *lov)
1012 {
1013         LASSERT(lov->lo_owner != current);
1014         down_write(&lov->lo_type_guard);
1015         LASSERT(lov->lo_owner == NULL);
1016         lov->lo_owner = current;
1017         CDEBUG(D_INODE, "Took exclusive lov(%p) owner %p\n",
1018                 lov, lov->lo_owner);
1019 }
1020
1021 static void lov_conf_unlock(struct lov_object *lov)
1022 {
1023         CDEBUG(D_INODE, "To release exclusive lov(%p) owner %p\n",
1024                 lov, lov->lo_owner);
1025         lov->lo_owner = NULL;
1026         up_write(&lov->lo_type_guard);
1027 }
1028
1029 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov)
1030 {
1031         struct l_wait_info lwi = { 0 };
1032         ENTRY;
1033
1034         while (atomic_read(&lov->lo_active_ios) > 0) {
1035                 CDEBUG(D_INODE, "file:"DFID" wait for active IO, now: %d.\n",
1036                         PFID(lu_object_fid(lov2lu(lov))),
1037                         atomic_read(&lov->lo_active_ios));
1038
1039                 l_wait_event(lov->lo_waitq,
1040                              atomic_read(&lov->lo_active_ios) == 0, &lwi);
1041         }
1042         RETURN(0);
1043 }
1044
1045 static int lov_layout_change(const struct lu_env *unused,
1046                              struct lov_object *lov, struct lov_stripe_md *lsm,
1047                              const struct cl_object_conf *conf)
1048 {
1049         enum lov_layout_type llt = lov_type(lsm);
1050         union lov_layout_state *state = &lov->u;
1051         const struct lov_layout_operations *old_ops;
1052         const struct lov_layout_operations *new_ops;
1053         struct lov_device *lov_dev = lov_object_dev(lov);
1054         struct lu_env *env;
1055         __u16 refcheck;
1056         int rc;
1057         ENTRY;
1058
1059         LASSERT(lov->lo_type < ARRAY_SIZE(lov_dispatch));
1060
1061         env = cl_env_get(&refcheck);
1062         if (IS_ERR(env))
1063                 RETURN(PTR_ERR(env));
1064
1065         LASSERT(llt < ARRAY_SIZE(lov_dispatch));
1066
1067         CDEBUG(D_INODE, DFID" from %s to %s\n",
1068                PFID(lu_object_fid(lov2lu(lov))),
1069                llt2str(lov->lo_type), llt2str(llt));
1070
1071         old_ops = &lov_dispatch[lov->lo_type];
1072         new_ops = &lov_dispatch[llt];
1073
1074         rc = cl_object_prune(env, &lov->lo_cl);
1075         if (rc != 0)
1076                 GOTO(out, rc);
1077
1078         rc = old_ops->llo_delete(env, lov, &lov->u);
1079         if (rc != 0)
1080                 GOTO(out, rc);
1081
1082         old_ops->llo_fini(env, lov, &lov->u);
1083
1084         LASSERT(atomic_read(&lov->lo_active_ios) == 0);
1085
1086         CDEBUG(D_INODE, DFID "Apply new layout lov %p, type %d\n",
1087                PFID(lu_object_fid(lov2lu(lov))), lov, llt);
1088
1089         lov->lo_type = LLT_EMPTY;
1090
1091         /* page bufsize fixup */
1092         cl_object_header(&lov->lo_cl)->coh_page_bufsize -=
1093                 lov_page_slice_fixup(lov, NULL);
1094
1095         rc = new_ops->llo_init(env, lov_dev, lov, lsm, conf, state);
1096         if (rc != 0) {
1097                 struct obd_device *obd = lov2obd(lov_dev->ld_lov);
1098
1099                 CERROR("%s: cannot apply new layout on "DFID" : rc = %d\n",
1100                        obd->obd_name, PFID(lu_object_fid(lov2lu(lov))), rc);
1101                 new_ops->llo_delete(env, lov, state);
1102                 new_ops->llo_fini(env, lov, state);
1103                 /* this file becomes an EMPTY file. */
1104                 GOTO(out, rc);
1105         }
1106
1107         lov->lo_type = llt;
1108
1109 out:
1110         cl_env_put(env, &refcheck);
1111         RETURN(rc);
1112 }
1113
1114 /*****************************************************************************
1115  *
1116  * Lov object operations.
1117  *
1118  */
1119 int lov_object_init(const struct lu_env *env, struct lu_object *obj,
1120                     const struct lu_object_conf *conf)
1121 {
1122         struct lov_object            *lov   = lu2lov(obj);
1123         struct lov_device            *dev   = lov_object_dev(lov);
1124         const struct cl_object_conf  *cconf = lu2cl_conf(conf);
1125         union lov_layout_state       *set   = &lov->u;
1126         const struct lov_layout_operations *ops;
1127         struct lov_stripe_md *lsm = NULL;
1128         int rc;
1129         ENTRY;
1130
1131         init_rwsem(&lov->lo_type_guard);
1132         atomic_set(&lov->lo_active_ios, 0);
1133         init_waitqueue_head(&lov->lo_waitq);
1134         cl_object_page_init(lu2cl(obj), sizeof(struct lov_page));
1135
1136         lov->lo_type = LLT_EMPTY;
1137         if (cconf->u.coc_layout.lb_buf != NULL) {
1138                 lsm = lov_unpackmd(dev->ld_lov,
1139                                    cconf->u.coc_layout.lb_buf,
1140                                    cconf->u.coc_layout.lb_len);
1141                 if (IS_ERR(lsm))
1142                         RETURN(PTR_ERR(lsm));
1143
1144                 dump_lsm(D_INODE, lsm);
1145         }
1146
1147         /* no locking is necessary, as object is being created */
1148         lov->lo_type = lov_type(lsm);
1149         ops = &lov_dispatch[lov->lo_type];
1150         rc = ops->llo_init(env, dev, lov, lsm, cconf, set);
1151         if (rc != 0)
1152                 GOTO(out_lsm, rc);
1153
1154 out_lsm:
1155         lov_lsm_put(lsm);
1156
1157         RETURN(rc);
1158 }
1159
1160 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
1161                         const struct cl_object_conf *conf)
1162 {
1163         struct lov_stripe_md    *lsm = NULL;
1164         struct lov_object       *lov = cl2lov(obj);
1165         int                      result = 0;
1166         ENTRY;
1167
1168         if (conf->coc_opc == OBJECT_CONF_SET &&
1169             conf->u.coc_layout.lb_buf != NULL) {
1170                 lsm = lov_unpackmd(lov_object_dev(lov)->ld_lov,
1171                                    conf->u.coc_layout.lb_buf,
1172                                    conf->u.coc_layout.lb_len);
1173                 if (IS_ERR(lsm))
1174                         RETURN(PTR_ERR(lsm));
1175                 dump_lsm(D_INODE, lsm);
1176         }
1177
1178         lov_conf_lock(lov);
1179         if (conf->coc_opc == OBJECT_CONF_INVALIDATE) {
1180                 lov->lo_layout_invalid = true;
1181                 GOTO(out, result = 0);
1182         }
1183
1184         if (conf->coc_opc == OBJECT_CONF_WAIT) {
1185                 if (lov->lo_layout_invalid &&
1186                     atomic_read(&lov->lo_active_ios) > 0) {
1187                         lov_conf_unlock(lov);
1188                         result = lov_layout_wait(env, lov);
1189                         lov_conf_lock(lov);
1190                 }
1191                 GOTO(out, result);
1192         }
1193
1194         LASSERT(conf->coc_opc == OBJECT_CONF_SET);
1195
1196         if ((lsm == NULL && lov->lo_lsm == NULL) ||
1197             ((lsm != NULL && lov->lo_lsm != NULL) &&
1198              (lov->lo_lsm->lsm_layout_gen == lsm->lsm_layout_gen) &&
1199              (lov->lo_lsm->lsm_entries[0]->lsme_pattern ==
1200               lsm->lsm_entries[0]->lsme_pattern))) {
1201                 /* same version of layout */
1202                 lov->lo_layout_invalid = false;
1203                 GOTO(out, result = 0);
1204         }
1205
1206         /* will change layout - check if there still exists active IO. */
1207         if (atomic_read(&lov->lo_active_ios) > 0) {
1208                 lov->lo_layout_invalid = true;
1209                 GOTO(out, result = -EBUSY);
1210         }
1211
1212         result = lov_layout_change(env, lov, lsm, conf);
1213         lov->lo_layout_invalid = result != 0;
1214         EXIT;
1215
1216 out:
1217         lov_conf_unlock(lov);
1218         lov_lsm_put(lsm);
1219         CDEBUG(D_INODE, DFID" lo_layout_invalid=%d\n",
1220                PFID(lu_object_fid(lov2lu(lov))), lov->lo_layout_invalid);
1221         RETURN(result);
1222 }
1223
1224 static void lov_object_delete(const struct lu_env *env, struct lu_object *obj)
1225 {
1226         struct lov_object *lov = lu2lov(obj);
1227
1228         ENTRY;
1229         LOV_2DISPATCH_VOID(lov, llo_delete, env, lov, &lov->u);
1230         EXIT;
1231 }
1232
1233 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
1234 {
1235         struct lov_object *lov = lu2lov(obj);
1236
1237         ENTRY;
1238         LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
1239         lu_object_fini(obj);
1240         OBD_SLAB_FREE_PTR(lov, lov_object_kmem);
1241         EXIT;
1242 }
1243
1244 static int lov_object_print(const struct lu_env *env, void *cookie,
1245                             lu_printer_t p, const struct lu_object *o)
1246 {
1247         return LOV_2DISPATCH_NOLOCK(lu2lov(o), llo_print, env, cookie, p, o);
1248 }
1249
1250 int lov_page_init(const struct lu_env *env, struct cl_object *obj,
1251                   struct cl_page *page, pgoff_t index)
1252 {
1253         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_page_init, env, obj, page,
1254                                     index);
1255 }
1256
1257 /**
1258  * Implements cl_object_operations::clo_io_init() method for lov
1259  * layer. Dispatches to the appropriate layout io initialization method.
1260  */
1261 int lov_io_init(const struct lu_env *env, struct cl_object *obj,
1262                 struct cl_io *io)
1263 {
1264         CL_IO_SLICE_CLEAN(lov_env_io(env), lis_cl);
1265
1266         CDEBUG(D_INODE, DFID "io %p type %d ignore/verify layout %d/%d\n",
1267                PFID(lu_object_fid(&obj->co_lu)), io, io->ci_type,
1268                io->ci_ignore_layout, io->ci_verify_layout);
1269
1270         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
1271                                      !io->ci_ignore_layout, env, obj, io);
1272 }
1273
1274 /**
1275  * An implementation of cl_object_operations::clo_attr_get() method for lov
1276  * layer. For raid0 layout this collects and merges attributes of all
1277  * sub-objects.
1278  */
1279 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
1280                         struct cl_attr *attr)
1281 {
1282         /* do not take lock, as this function is called under a
1283          * spin-lock. Layout is protected from changing by ongoing IO. */
1284         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
1285 }
1286
1287 static int lov_attr_update(const struct lu_env *env, struct cl_object *obj,
1288                            const struct cl_attr *attr, unsigned valid)
1289 {
1290         /*
1291          * No dispatch is required here, as no layout implements this.
1292          */
1293         return 0;
1294 }
1295
1296 int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
1297                   struct cl_lock *lock, const struct cl_io *io)
1298 {
1299         /* No need to lock because we've taken one refcount of layout.  */
1300         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_lock_init, env, obj, lock,
1301                                     io);
1302 }
1303
1304 /**
1305  * We calculate on which OST the mapping will end. If the length of mapping
1306  * is greater than (stripe_size * stripe_count) then the last_stripe will
1307  * will be one just before start_stripe. Else we check if the mapping
1308  * intersects each OST and find last_stripe.
1309  * This function returns the last_stripe and also sets the stripe_count
1310  * over which the mapping is spread
1311  *
1312  * \param lsm [in]              striping information for the file
1313  * \param index [in]            stripe component index
1314  * \param ext [in]              logical extent of mapping
1315  * \param start_stripe [in]     starting stripe of the mapping
1316  * \param stripe_count [out]    the number of stripes across which to map is
1317  *                              returned
1318  *
1319  * \retval last_stripe          return the last stripe of the mapping
1320  */
1321 static int fiemap_calc_last_stripe(struct lov_stripe_md *lsm, int index,
1322                                    struct lu_extent *ext,
1323                                    int start_stripe, int *stripe_count)
1324 {
1325         struct lov_stripe_md_entry *lsme = lsm->lsm_entries[index];
1326         int last_stripe;
1327         u64 obd_start;
1328         u64 obd_end;
1329         int i, j;
1330
1331         if (ext->e_end - ext->e_start >
1332             lsme->lsme_stripe_size * lsme->lsme_stripe_count) {
1333                 last_stripe = (start_stripe < 1 ? lsme->lsme_stripe_count - 1 :
1334                                                   start_stripe - 1);
1335                 *stripe_count = lsme->lsme_stripe_count;
1336         } else {
1337                 for (j = 0, i = start_stripe; j < lsme->lsme_stripe_count;
1338                      i = (i + 1) % lsme->lsme_stripe_count, j++) {
1339                         if ((lov_stripe_intersects(lsm, index,  i, ext,
1340                                                    &obd_start, &obd_end)) == 0)
1341                                 break;
1342                 }
1343                 *stripe_count = j;
1344                 last_stripe = (start_stripe + j - 1) % lsme->lsme_stripe_count;
1345         }
1346
1347         return last_stripe;
1348 }
1349
1350 /**
1351  * Set fe_device and copy extents from local buffer into main return buffer.
1352  *
1353  * \param fiemap [out]          fiemap to hold all extents
1354  * \param lcl_fm_ext [in]       array of fiemap extents get from OSC layer
1355  * \param ost_index [in]        OST index to be written into the fm_device
1356  *                              field for each extent
1357  * \param ext_count [in]        number of extents to be copied
1358  * \param current_extent [in]   where to start copying in the extent array
1359  */
1360 static void fiemap_prepare_and_copy_exts(struct fiemap *fiemap,
1361                                          struct fiemap_extent *lcl_fm_ext,
1362                                          int ost_index, unsigned int ext_count,
1363                                          int current_extent)
1364 {
1365         char            *to;
1366         unsigned int    ext;
1367
1368         for (ext = 0; ext < ext_count; ext++) {
1369                 lcl_fm_ext[ext].fe_device = ost_index;
1370                 lcl_fm_ext[ext].fe_flags |= FIEMAP_EXTENT_NET;
1371         }
1372
1373         /* Copy fm_extent's from fm_local to return buffer */
1374         to = (char *)fiemap + fiemap_count_to_size(current_extent);
1375         memcpy(to, lcl_fm_ext, ext_count * sizeof(struct fiemap_extent));
1376 }
1377
1378 #define FIEMAP_BUFFER_SIZE 4096
1379
1380 /**
1381  * Non-zero fe_logical indicates that this is a continuation FIEMAP
1382  * call. The local end offset and the device are sent in the first
1383  * fm_extent. This function calculates the stripe number from the index.
1384  * This function returns a stripe_no on which mapping is to be restarted.
1385  *
1386  * This function returns fm_end_offset which is the in-OST offset at which
1387  * mapping should be restarted. If fm_end_offset=0 is returned then caller
1388  * will re-calculate proper offset in next stripe.
1389  * Note that the first extent is passed to lov_get_info via the value field.
1390  *
1391  * \param fiemap [in]           fiemap request header
1392  * \param lsm [in]              striping information for the file
1393  * \param index [in]            stripe component index
1394  * \param ext [in]              logical extent of mapping
1395  * \param start_stripe [out]    starting stripe will be returned in this
1396  */
1397 static u64 fiemap_calc_fm_end_offset(struct fiemap *fiemap,
1398                                      struct lov_stripe_md *lsm,
1399                                      int index, struct lu_extent *ext,
1400                                      int *start_stripe)
1401 {
1402         struct lov_stripe_md_entry *lsme = lsm->lsm_entries[index];
1403         u64 local_end = fiemap->fm_extents[0].fe_logical;
1404         u64 lun_start;
1405         u64 lun_end;
1406         u64 fm_end_offset;
1407         int stripe_no = -1;
1408         int i;
1409
1410         if (fiemap->fm_extent_count == 0 ||
1411             fiemap->fm_extents[0].fe_logical == 0)
1412                 return 0;
1413
1414         /* Find out stripe_no from ost_index saved in the fe_device */
1415         for (i = 0; i < lsme->lsme_stripe_count; i++) {
1416                 struct lov_oinfo *oinfo = lsme->lsme_oinfo[i];
1417
1418                 if (lov_oinfo_is_dummy(oinfo))
1419                         continue;
1420
1421                 if (oinfo->loi_ost_idx == fiemap->fm_extents[0].fe_device) {
1422                         stripe_no = i;
1423                         break;
1424                 }
1425         }
1426
1427         if (stripe_no == -1)
1428                 return -EINVAL;
1429
1430         /* If we have finished mapping on previous device, shift logical
1431          * offset to start of next device */
1432         if (lov_stripe_intersects(lsm, index, stripe_no, ext,
1433                                    &lun_start, &lun_end) != 0 &&
1434             local_end < lun_end) {
1435                 fm_end_offset = local_end;
1436                 *start_stripe = stripe_no;
1437         } else {
1438                 /* This is a special value to indicate that caller should
1439                  * calculate offset in next stripe. */
1440                 fm_end_offset = 0;
1441                 *start_stripe = (stripe_no + 1) % lsme->lsme_stripe_count;
1442         }
1443
1444         return fm_end_offset;
1445 }
1446
1447 struct fiemap_state {
1448         struct fiemap           *fs_fm;
1449         struct lu_extent        fs_ext;
1450         u64                     fs_length;
1451         u64                     fs_end_offset;
1452         int                     fs_cur_extent;
1453         int                     fs_cnt_need;
1454         int                     fs_start_stripe;
1455         int                     fs_last_stripe;
1456         bool                    fs_device_done;
1457         bool                    fs_finish_stripe;
1458         bool                    fs_enough;
1459 };
1460
1461 static struct cl_object *lov_find_subobj(const struct lu_env *env,
1462                                          struct lov_object *lov,
1463                                          struct lov_stripe_md *lsm,
1464                                          int index)
1465 {
1466         struct lov_device       *dev = lu2lov_dev(lov2lu(lov)->lo_dev);
1467         struct lov_thread_info  *lti = lov_env_info(env);
1468         struct lu_fid           *ofid = &lti->lti_fid;
1469         struct lov_oinfo        *oinfo;
1470         struct cl_device        *subdev;
1471         int                     entry = lov_comp_entry(index);
1472         int                     stripe = lov_comp_stripe(index);
1473         int                     ost_idx;
1474         int                     rc;
1475         struct cl_object        *result;
1476
1477         if (lov->lo_type != LLT_COMP)
1478                 GOTO(out, result = NULL);
1479
1480         if (entry >= lsm->lsm_entry_count ||
1481             stripe >= lsm->lsm_entries[entry]->lsme_stripe_count)
1482                 GOTO(out, result = NULL);
1483
1484         oinfo = lsm->lsm_entries[entry]->lsme_oinfo[stripe];
1485         ost_idx = oinfo->loi_ost_idx;
1486         rc = ostid_to_fid(ofid, &oinfo->loi_oi, ost_idx);
1487         if (rc != 0)
1488                 GOTO(out, result = NULL);
1489
1490         subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
1491         result = lov_sub_find(env, subdev, ofid, NULL);
1492 out:
1493         if (result == NULL)
1494                 result = ERR_PTR(-EINVAL);
1495         return result;
1496 }
1497
1498 int fiemap_for_stripe(const struct lu_env *env, struct cl_object *obj,
1499                       struct lov_stripe_md *lsm, struct fiemap *fiemap,
1500                       size_t *buflen, struct ll_fiemap_info_key *fmkey,
1501                       int index, int stripeno, struct fiemap_state *fs)
1502 {
1503         struct lov_stripe_md_entry *lsme = lsm->lsm_entries[index];
1504         struct cl_object *subobj;
1505         struct lov_obd *lov = lu2lov_dev(obj->co_lu.lo_dev)->ld_lov;
1506         struct fiemap_extent *fm_ext = &fs->fs_fm->fm_extents[0];
1507         u64 req_fm_len; /* Stores length of required mapping */
1508         u64 len_mapped_single_call;
1509         u64 lun_start;
1510         u64 lun_end;
1511         u64 obd_object_end;
1512         unsigned int ext_count;
1513         /* EOF for object */
1514         bool ost_eof = false;
1515         /* done with required mapping for this OST? */
1516         bool ost_done = false;
1517         int ost_index;
1518         int rc = 0;
1519
1520         fs->fs_device_done = false;
1521         /* Find out range of mapping on this stripe */
1522         if ((lov_stripe_intersects(lsm, index, stripeno, &fs->fs_ext,
1523                                    &lun_start, &obd_object_end)) == 0)
1524                 return 0;
1525
1526         if (lov_oinfo_is_dummy(lsme->lsme_oinfo[stripeno]))
1527                 return -EIO;
1528
1529         /* If this is a continuation FIEMAP call and we are on
1530          * starting stripe then lun_start needs to be set to
1531          * end_offset */
1532         if (fs->fs_end_offset != 0 && stripeno == fs->fs_start_stripe)
1533                 lun_start = fs->fs_end_offset;
1534         lun_end = lov_size_to_stripe(lsm, index, fs->fs_ext.e_end, stripeno);
1535         if (lun_start == lun_end)
1536                 return 0;
1537
1538         req_fm_len = obd_object_end - lun_start;
1539         fs->fs_fm->fm_length = 0;
1540         len_mapped_single_call = 0;
1541
1542         /* find lobsub object */
1543         subobj = lov_find_subobj(env, cl2lov(obj), lsm,
1544                                  lov_comp_index(index, stripeno));
1545         if (IS_ERR(subobj))
1546                 return PTR_ERR(subobj);
1547         /* If the output buffer is very large and the objects have many
1548          * extents we may need to loop on a single OST repeatedly */
1549         do {
1550                 if (fiemap->fm_extent_count > 0) {
1551                         /* Don't get too many extents. */
1552                         if (fs->fs_cur_extent + fs->fs_cnt_need >
1553                             fiemap->fm_extent_count)
1554                                 fs->fs_cnt_need = fiemap->fm_extent_count -
1555                                                   fs->fs_cur_extent;
1556                 }
1557
1558                 lun_start += len_mapped_single_call;
1559                 fs->fs_fm->fm_length = req_fm_len - len_mapped_single_call;
1560                 req_fm_len = fs->fs_fm->fm_length;
1561                 /**
1562                  * If we've collected enough extent map, we'd request 1 more,
1563                  * to see whether we coincidentally finished all available
1564                  * extent map, so that FIEMAP_EXTENT_LAST would be set.
1565                  */
1566                 fs->fs_fm->fm_extent_count = fs->fs_enough ?
1567                                              1 : fs->fs_cnt_need;
1568                 fs->fs_fm->fm_mapped_extents = 0;
1569                 fs->fs_fm->fm_flags = fiemap->fm_flags;
1570
1571                 ost_index = lsme->lsme_oinfo[stripeno]->loi_ost_idx;
1572
1573                 if (ost_index < 0 || ost_index >= lov->desc.ld_tgt_count)
1574                         GOTO(obj_put, rc = -EINVAL);
1575                 /* If OST is inactive, return extent with UNKNOWN flag. */
1576                 if (!lov->lov_tgts[ost_index]->ltd_active) {
1577                         fs->fs_fm->fm_flags |= FIEMAP_EXTENT_LAST;
1578                         fs->fs_fm->fm_mapped_extents = 1;
1579
1580                         fm_ext[0].fe_logical = lun_start;
1581                         fm_ext[0].fe_length = obd_object_end - lun_start;
1582                         fm_ext[0].fe_flags |= FIEMAP_EXTENT_UNKNOWN;
1583
1584                         goto inactive_tgt;
1585                 }
1586
1587                 fs->fs_fm->fm_start = lun_start;
1588                 fs->fs_fm->fm_flags &= ~FIEMAP_FLAG_DEVICE_ORDER;
1589                 memcpy(&fmkey->lfik_fiemap, fs->fs_fm, sizeof(*fs->fs_fm));
1590                 *buflen = fiemap_count_to_size(fs->fs_fm->fm_extent_count);
1591
1592                 rc = cl_object_fiemap(env, subobj, fmkey, fs->fs_fm, buflen);
1593                 if (rc != 0)
1594                         GOTO(obj_put, rc);
1595 inactive_tgt:
1596                 ext_count = fs->fs_fm->fm_mapped_extents;
1597                 if (ext_count == 0) {
1598                         ost_done = true;
1599                         fs->fs_device_done = true;
1600                         /* If last stripe has hold at the end,
1601                          * we need to return */
1602                         if (stripeno == fs->fs_last_stripe) {
1603                                 fiemap->fm_mapped_extents = 0;
1604                                 fs->fs_finish_stripe = true;
1605                                 GOTO(obj_put, rc);
1606                         }
1607                         break;
1608                 } else if (fs->fs_enough) {
1609                         /*
1610                          * We've collected enough extents and there are
1611                          * more extents after it.
1612                          */
1613                         GOTO(obj_put, rc);
1614                 }
1615
1616                 /* If we just need num of extents, got to next device */
1617                 if (fiemap->fm_extent_count == 0) {
1618                         fs->fs_cur_extent += ext_count;
1619                         break;
1620                 }
1621
1622                 /* prepare to copy retrived map extents */
1623                 len_mapped_single_call = fm_ext[ext_count - 1].fe_logical +
1624                                          fm_ext[ext_count - 1].fe_length -
1625                                          lun_start;
1626
1627                 /* Have we finished mapping on this device? */
1628                 if (req_fm_len <= len_mapped_single_call) {
1629                         ost_done = true;
1630                         fs->fs_device_done = true;
1631                 }
1632
1633                 /* Clear the EXTENT_LAST flag which can be present on
1634                  * the last extent */
1635                 if (fm_ext[ext_count - 1].fe_flags & FIEMAP_EXTENT_LAST)
1636                         fm_ext[ext_count - 1].fe_flags &= ~FIEMAP_EXTENT_LAST;
1637                 if (lov_stripe_size(lsm, index,
1638                                     fm_ext[ext_count - 1].fe_logical +
1639                                     fm_ext[ext_count - 1].fe_length,
1640                                     stripeno) >= fmkey->lfik_oa.o_size) {
1641                         ost_eof = true;
1642                         fs->fs_device_done = true;
1643                 }
1644
1645                 fiemap_prepare_and_copy_exts(fiemap, fm_ext, ost_index,
1646                                              ext_count, fs->fs_cur_extent);
1647                 fs->fs_cur_extent += ext_count;
1648
1649                 /* Ran out of available extents? */
1650                 if (fs->fs_cur_extent >= fiemap->fm_extent_count)
1651                         fs->fs_enough = true;
1652         } while (!ost_done && !ost_eof);
1653
1654         if (stripeno == fs->fs_last_stripe)
1655                 fs->fs_finish_stripe = true;
1656 obj_put:
1657         cl_object_put(env, subobj);
1658
1659         return rc;
1660 }
1661
1662 /**
1663  * Break down the FIEMAP request and send appropriate calls to individual OSTs.
1664  * This also handles the restarting of FIEMAP calls in case mapping overflows
1665  * the available number of extents in single call.
1666  *
1667  * \param env [in]              lustre environment
1668  * \param obj [in]              file object
1669  * \param fmkey [in]            fiemap request header and other info
1670  * \param fiemap [out]          fiemap buffer holding retrived map extents
1671  * \param buflen [in/out]       max buffer length of @fiemap, when iterate
1672  *                              each OST, it is used to limit max map needed
1673  * \retval 0    success
1674  * \retval < 0  error
1675  */
1676 static int lov_object_fiemap(const struct lu_env *env, struct cl_object *obj,
1677                              struct ll_fiemap_info_key *fmkey,
1678                              struct fiemap *fiemap, size_t *buflen)
1679 {
1680         struct lov_stripe_md_entry *lsme;
1681         struct lov_stripe_md *lsm;
1682         struct fiemap *fm_local = NULL;
1683         loff_t whole_start;
1684         loff_t whole_end;
1685         int entry;
1686         int start_entry;
1687         int end_entry;
1688         int cur_stripe = 0;
1689         int stripe_count;
1690         unsigned int buffer_size = FIEMAP_BUFFER_SIZE;
1691         int rc = 0;
1692         struct fiemap_state fs = { 0 };
1693         ENTRY;
1694
1695         lsm = lov_lsm_addref(cl2lov(obj));
1696         if (lsm == NULL)
1697                 RETURN(-ENODATA);
1698
1699         if (!(fiemap->fm_flags & FIEMAP_FLAG_DEVICE_ORDER)) {
1700                 /**
1701                  * If the entry count > 1 or stripe_count > 1 and the
1702                  * application does not understand DEVICE_ORDER flag,
1703                  * it cannot interpret the extents correctly.
1704                  */
1705                 if (lsm->lsm_entry_count > 1 ||
1706                     (lsm->lsm_entry_count == 1 &&
1707                      lsm->lsm_entries[0]->lsme_stripe_count > 1))
1708                         GOTO(out_lsm, rc = -ENOTSUPP);
1709         }
1710
1711         /* No support for DOM layout yet. */
1712         if (lsme_is_dom(lsm->lsm_entries[0]))
1713                 GOTO(out_lsm, rc = -ENOTSUPP);
1714
1715         if (lsm->lsm_is_released) {
1716                 if (fiemap->fm_start < fmkey->lfik_oa.o_size) {
1717                         /**
1718                          * released file, return a minimal FIEMAP if
1719                          * request fits in file-size.
1720                          */
1721                         fiemap->fm_mapped_extents = 1;
1722                         fiemap->fm_extents[0].fe_logical = fiemap->fm_start;
1723                         if (fiemap->fm_start + fiemap->fm_length <
1724                             fmkey->lfik_oa.o_size)
1725                                 fiemap->fm_extents[0].fe_length =
1726                                         fiemap->fm_length;
1727                         else
1728                                 fiemap->fm_extents[0].fe_length =
1729                                         fmkey->lfik_oa.o_size -
1730                                         fiemap->fm_start;
1731                         fiemap->fm_extents[0].fe_flags |=
1732                                 FIEMAP_EXTENT_UNKNOWN | FIEMAP_EXTENT_LAST;
1733                 }
1734                 GOTO(out_lsm, rc = 0);
1735         }
1736
1737         /* buffer_size is small to hold fm_extent_count of extents. */
1738         if (fiemap_count_to_size(fiemap->fm_extent_count) < buffer_size)
1739                 buffer_size = fiemap_count_to_size(fiemap->fm_extent_count);
1740
1741         OBD_ALLOC_LARGE(fm_local, buffer_size);
1742         if (fm_local == NULL)
1743                 GOTO(out_lsm, rc = -ENOMEM);
1744
1745         /**
1746          * Requested extent count exceeds the fiemap buffer size, shrink our
1747          * ambition.
1748          */
1749         if (fiemap_count_to_size(fiemap->fm_extent_count) > *buflen)
1750                 fiemap->fm_extent_count = fiemap_size_to_count(*buflen);
1751         if (fiemap->fm_extent_count == 0)
1752                 fs.fs_cnt_need = 0;
1753
1754         fs.fs_enough = false;
1755         fs.fs_cur_extent = 0;
1756         fs.fs_fm = fm_local;
1757         fs.fs_cnt_need = fiemap_size_to_count(buffer_size);
1758
1759         whole_start = fiemap->fm_start;
1760         /* whole_start is beyond the end of the file */
1761         if (whole_start > fmkey->lfik_oa.o_size)
1762                 GOTO(out_fm_local, rc = -EINVAL);
1763         whole_end = (fiemap->fm_length == OBD_OBJECT_EOF) ?
1764                                         fmkey->lfik_oa.o_size :
1765                                         whole_start + fiemap->fm_length - 1;
1766         /**
1767          * If fiemap->fm_length != OBD_OBJECT_EOF but whole_end exceeds file
1768          * size
1769          */
1770         if (whole_end > fmkey->lfik_oa.o_size)
1771                 whole_end = fmkey->lfik_oa.o_size;
1772
1773         start_entry = lov_lsm_entry(lsm, whole_start);
1774         end_entry = lov_lsm_entry(lsm, whole_end);
1775         if (end_entry == -1)
1776                 end_entry = lsm->lsm_entry_count - 1;
1777
1778         if (start_entry == -1 || end_entry == -1)
1779                 GOTO(out_fm_local, rc = -EINVAL);
1780
1781         for (entry = start_entry; entry <= end_entry; entry++) {
1782                 lsme = lsm->lsm_entries[entry];
1783
1784                 if (!lsme_inited(lsme))
1785                         break;
1786
1787                 if (entry == start_entry)
1788                         fs.fs_ext.e_start = whole_start;
1789                 else
1790                         fs.fs_ext.e_start = lsme->lsme_extent.e_start;
1791                 if (entry == end_entry)
1792                         fs.fs_ext.e_end = whole_end;
1793                 else
1794                         fs.fs_ext.e_end = lsme->lsme_extent.e_end - 1;
1795                 fs.fs_length = fs.fs_ext.e_end - fs.fs_ext.e_start + 1;
1796
1797                 /* Calculate start stripe, last stripe and length of mapping */
1798                 fs.fs_start_stripe = lov_stripe_number(lsm, entry,
1799                                                        fs.fs_ext.e_start);
1800                 fs.fs_last_stripe = fiemap_calc_last_stripe(lsm, entry,
1801                                         &fs.fs_ext, fs.fs_start_stripe,
1802                                         &stripe_count);
1803                 fs.fs_end_offset = fiemap_calc_fm_end_offset(fiemap, lsm, entry,
1804                                         &fs.fs_ext, &fs.fs_start_stripe);
1805                 /* Check each stripe */
1806                 for (cur_stripe = fs.fs_start_stripe; stripe_count > 0;
1807                      --stripe_count,
1808                      cur_stripe = (cur_stripe + 1) % lsme->lsme_stripe_count) {
1809                         rc = fiemap_for_stripe(env, obj, lsm, fiemap, buflen,
1810                                                fmkey, entry, cur_stripe, &fs);
1811                         if (rc < 0)
1812                                 GOTO(out_fm_local, rc);
1813                         if (fs.fs_enough)
1814                                 GOTO(finish, rc);
1815                         if (fs.fs_finish_stripe)
1816                                 break;
1817                 } /* for each stripe */
1818         } /* for covering layout component */
1819         /*
1820          * We've traversed all components, set @entry to the last component
1821          * entry, it's for the last stripe check.
1822          */
1823         entry--;
1824 finish:
1825         /* Indicate that we are returning device offsets unless file just has
1826          * single stripe */
1827         if (lsm->lsm_entry_count > 1 ||
1828             (lsm->lsm_entry_count == 1 &&
1829              lsm->lsm_entries[0]->lsme_stripe_count > 1))
1830                 fiemap->fm_flags |= FIEMAP_FLAG_DEVICE_ORDER;
1831
1832         if (fiemap->fm_extent_count == 0)
1833                 goto skip_last_device_calc;
1834
1835         /* Check if we have reached the last stripe and whether mapping for that
1836          * stripe is done. */
1837         if ((cur_stripe == fs.fs_last_stripe) && fs.fs_device_done)
1838                 fiemap->fm_extents[fs.fs_cur_extent - 1].fe_flags |=
1839                                                              FIEMAP_EXTENT_LAST;
1840 skip_last_device_calc:
1841         fiemap->fm_mapped_extents = fs.fs_cur_extent;
1842 out_fm_local:
1843         OBD_FREE_LARGE(fm_local, buffer_size);
1844
1845 out_lsm:
1846         lov_lsm_put(lsm);
1847         return rc;
1848 }
1849
1850 static int lov_object_getstripe(const struct lu_env *env, struct cl_object *obj,
1851                                 struct lov_user_md __user *lum, size_t size)
1852 {
1853         struct lov_object       *lov = cl2lov(obj);
1854         struct lov_stripe_md    *lsm;
1855         int                     rc = 0;
1856         ENTRY;
1857
1858         lsm = lov_lsm_addref(lov);
1859         if (lsm == NULL)
1860                 RETURN(-ENODATA);
1861
1862         rc = lov_getstripe(env, cl2lov(obj), lsm, lum, size);
1863         lov_lsm_put(lsm);
1864         RETURN(rc);
1865 }
1866
1867 static int lov_object_layout_get(const struct lu_env *env,
1868                                  struct cl_object *obj,
1869                                  struct cl_layout *cl)
1870 {
1871         struct lov_object *lov = cl2lov(obj);
1872         struct lov_stripe_md *lsm = lov_lsm_addref(lov);
1873         struct lu_buf *buf = &cl->cl_buf;
1874         ssize_t rc;
1875         ENTRY;
1876
1877         if (lsm == NULL) {
1878                 cl->cl_size = 0;
1879                 cl->cl_layout_gen = CL_LAYOUT_GEN_EMPTY;
1880
1881                 RETURN(0);
1882         }
1883
1884         cl->cl_size = lov_comp_md_size(lsm);
1885         cl->cl_layout_gen = lsm->lsm_layout_gen;
1886         cl->cl_is_composite = lsm_is_composite(lsm->lsm_magic);
1887
1888         rc = lov_lsm_pack(lsm, buf->lb_buf, buf->lb_len);
1889         lov_lsm_put(lsm);
1890
1891         RETURN(rc < 0 ? rc : 0);
1892 }
1893
1894 static loff_t lov_object_maxbytes(struct cl_object *obj)
1895 {
1896         struct lov_object *lov = cl2lov(obj);
1897         struct lov_stripe_md *lsm = lov_lsm_addref(lov);
1898         loff_t maxbytes;
1899
1900         if (lsm == NULL)
1901                 return LLONG_MAX;
1902
1903         maxbytes = lsm->lsm_maxbytes;
1904
1905         lov_lsm_put(lsm);
1906
1907         return maxbytes;
1908 }
1909
1910 static const struct cl_object_operations lov_ops = {
1911         .coo_page_init    = lov_page_init,
1912         .coo_lock_init    = lov_lock_init,
1913         .coo_io_init      = lov_io_init,
1914         .coo_attr_get     = lov_attr_get,
1915         .coo_attr_update  = lov_attr_update,
1916         .coo_conf_set     = lov_conf_set,
1917         .coo_getstripe    = lov_object_getstripe,
1918         .coo_layout_get   = lov_object_layout_get,
1919         .coo_maxbytes     = lov_object_maxbytes,
1920         .coo_fiemap       = lov_object_fiemap,
1921 };
1922
1923 static const struct lu_object_operations lov_lu_obj_ops = {
1924         .loo_object_init      = lov_object_init,
1925         .loo_object_delete    = lov_object_delete,
1926         .loo_object_release   = NULL,
1927         .loo_object_free      = lov_object_free,
1928         .loo_object_print     = lov_object_print,
1929         .loo_object_invariant = NULL
1930 };
1931
1932 struct lu_object *lov_object_alloc(const struct lu_env *env,
1933                                    const struct lu_object_header *unused,
1934                                    struct lu_device *dev)
1935 {
1936         struct lov_object *lov;
1937         struct lu_object  *obj;
1938
1939         ENTRY;
1940         OBD_SLAB_ALLOC_PTR_GFP(lov, lov_object_kmem, GFP_NOFS);
1941         if (lov != NULL) {
1942                 obj = lov2lu(lov);
1943                 lu_object_init(obj, NULL, dev);
1944                 lov->lo_cl.co_ops = &lov_ops;
1945                 lov->lo_type = -1; /* invalid, to catch uninitialized type */
1946                 /*
1947                  * object io operation vector (cl_object::co_iop) is installed
1948                  * later in lov_object_init(), as different vectors are used
1949                  * for object with different layouts.
1950                  */
1951                 obj->lo_ops = &lov_lu_obj_ops;
1952         } else
1953                 obj = NULL;
1954         RETURN(obj);
1955 }
1956
1957 struct lov_stripe_md *lov_lsm_addref(struct lov_object *lov)
1958 {
1959         struct lov_stripe_md *lsm = NULL;
1960
1961         lov_conf_freeze(lov);
1962         if (lov->lo_lsm != NULL) {
1963                 lsm = lsm_addref(lov->lo_lsm);
1964                 CDEBUG(D_INODE, "lsm %p addref %d/%d by %p.\n",
1965                         lsm, atomic_read(&lsm->lsm_refc),
1966                         lov->lo_layout_invalid, current);
1967         }
1968         lov_conf_thaw(lov);
1969         return lsm;
1970 }
1971
1972 int lov_read_and_clear_async_rc(struct cl_object *clob)
1973 {
1974         struct lu_object *luobj;
1975         int rc = 0;
1976         ENTRY;
1977
1978         luobj = lu_object_locate(&cl_object_header(clob)->coh_lu,
1979                                  &lov_device_type);
1980         if (luobj != NULL) {
1981                 struct lov_object *lov = lu2lov(luobj);
1982
1983                 lov_conf_freeze(lov);
1984                 switch (lov->lo_type) {
1985                 case LLT_COMP: {
1986                         struct lov_stripe_md *lsm;
1987                         int i;
1988
1989                         lsm = lov->lo_lsm;
1990                         LASSERT(lsm != NULL);
1991                         for (i = 0; i < lsm->lsm_entry_count; i++) {
1992                                 struct lov_stripe_md_entry *lse =
1993                                                 lsm->lsm_entries[i];
1994                                 int j;
1995
1996                                 if (!lsme_inited(lse))
1997                                         break;
1998
1999                                 for (j = 0; j < lse->lsme_stripe_count; j++) {
2000                                         struct lov_oinfo *loi =
2001                                                         lse->lsme_oinfo[j];
2002
2003                                         if (lov_oinfo_is_dummy(loi))
2004                                                 continue;
2005
2006                                         if (loi->loi_ar.ar_rc && !rc)
2007                                                 rc = loi->loi_ar.ar_rc;
2008                                         loi->loi_ar.ar_rc = 0;
2009                                 }
2010                         }
2011                 }
2012                 case LLT_RELEASED:
2013                 case LLT_EMPTY:
2014                         break;
2015                 default:
2016                         LBUG();
2017                 }
2018                 lov_conf_thaw(lov);
2019         }
2020         RETURN(rc);
2021 }
2022 EXPORT_SYMBOL(lov_read_and_clear_async_rc);
2023
2024 /** @} lov */