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         struct lov_mirror_entry *lre;
 633         unsigned int entry_count;
 634         unsigned int psz = 0;
 635         unsigned int mirror_count;
 636         int flr_state = lsm->lsm_flags & LCM_FL_FLR_MASK;
 637         int result = 0;
 638         int i, j;
 639
 640         ENTRY;
 641
 642         LASSERT(lsm->lsm_entry_count > 0);
 643         LASSERT(lov->lo_lsm == NULL);
 644         lov->lo_lsm = lsm_addref(lsm);
 645         lov->lo_layout_invalid = true;
 646
 647         dump_lsm(D_INODE, lsm);
 648
 649         entry_count = lsm->lsm_entry_count;
 650
 651         spin_lock_init(&comp->lo_write_lock);
 652         comp->lo_flags = lsm->lsm_flags;
 653         comp->lo_mirror_count = lsm->lsm_mirror_count + 1;
 654         comp->lo_entry_count = lsm->lsm_entry_count;
 655         comp->lo_preferred_mirror = -1;
 656
 657         if (equi(flr_state == LCM_FL_NOT_FLR, comp->lo_mirror_count > 1))
 658                 RETURN(-EINVAL);
 659
 660         OBD_ALLOC(comp->lo_mirrors,
 661                   comp->lo_mirror_count * sizeof(*comp->lo_mirrors));
 662         if (comp->lo_mirrors == NULL)
 663                 RETURN(-ENOMEM);
 664
 665         OBD_ALLOC(comp->lo_entries, entry_count * sizeof(*comp->lo_entries));
 666         if (comp->lo_entries == NULL)
 667                 RETURN(-ENOMEM);
 668
 669         /* Initiate all entry types and extents data at first */
 670         for (i = 0, j = 0, mirror_count = 1; i < entry_count; i++) {
 671                 int mirror_id = 0;
 672
 673                 lle = &comp->lo_entries[i];
 674
 675                 lle->lle_lsme = lsm->lsm_entries[i];
 676                 lle->lle_type = lov_entry_type(lle->lle_lsme);
 677                 switch (lle->lle_type) {
 678                 case LOV_PATTERN_RAID0:
 679                         lle->lle_comp_ops = &raid0_ops;
 680                         break;
 681                 case LOV_PATTERN_MDT:
 682                         lle->lle_comp_ops = &dom_ops;
 683                         break;
 684                 default:
 685                         CERROR("%s: unknown composite layout entry type %i\n",
 686                                lov2obd(dev->ld_lov)->obd_name,
 687                                lsm->lsm_entries[i]->lsme_pattern);
 688                         dump_lsm(D_ERROR, lsm);
 689                         RETURN(-EIO);
 690                 }
 691
 692                 lle->lle_extent = &lle->lle_lsme->lsme_extent;
 693                 lle->lle_valid = !(lle->lle_lsme->lsme_flags & LCME_FL_STALE);
 694
 695                 if (flr_state != LCM_FL_NOT_FLR)
 696                         mirror_id = mirror_id_of(lle->lle_lsme->lsme_id);
 697
 698                 lre = &comp->lo_mirrors[j];
 699                 if (i > 0) {
 700                         if (mirror_id == lre->lre_mirror_id) {
 701                                 lre->lre_valid |= lle->lle_valid;
 702                                 lre->lre_end = i;
 703                                 continue;
 704                         }
 705
 706                         /* new mirror detected, assume that the mirrors
 707                          * are shorted in layout */
 708                         ++mirror_count;
 709                         ++j;
 710                         if (j >= comp->lo_mirror_count)
 711                                 break;
 712
 713                         lre = &comp->lo_mirrors[j];
 714                 }
 715
 716                 /* entries must be sorted by mirrors */
 717                 lre->lre_mirror_id = mirror_id;
 718                 lre->lre_start = lre->lre_end = i;
 719                 lre->lre_preferred = (lle->lle_lsme->lsme_flags &
 720                                         LCME_FL_PREFERRED);
 721                 lre->lre_valid = lle->lle_valid;
 722         }
 723
 724         /* sanity check for FLR */
 725         if (mirror_count != comp->lo_mirror_count) {
 726                 CDEBUG(D_INODE, DFID
 727                        " doesn't have the # of mirrors it claims, %u/%u\n",
 728                        PFID(lu_object_fid(lov2lu(lov))), mirror_count,
 729                        comp->lo_mirror_count + 1);
 730
 731                 GOTO(out, result = -EINVAL);
 732         }
 733
 734         lov_foreach_layout_entry(lov, lle) {
 735                 int index = lov_layout_entry_index(lov, lle);
 736
 737                 /**
 738                  * If the component has not been init-ed on MDS side, for
 739                  * PFL layout, we'd know that the components beyond this one
 740                  * will be dynamically init-ed later on file write/trunc ops.
 741                  */
 742                 if (!lsme_inited(lle->lle_lsme))
 743                         continue;
 744
 745                 result = lle->lle_comp_ops->lco_init(env, dev, lov, index,
 746                                                      conf, lle);
 747                 if (result < 0)
 748                         break;
 749
 750                 LASSERT(ergo(psz > 0, psz == result));
 751                 psz = result;
 752         }
 753
 754         if (psz > 0)
 755                 cl_object_header(&lov->lo_cl)->coh_page_bufsize += psz;
 756
 757         /* decide the preferred mirror */
 758         mirror_count = 0, i = 0;
 759         lov_foreach_mirror_entry(lov, lre) {
 760                 i++;
 761                 if (!lre->lre_valid)
 762                         continue;
 763
 764                 mirror_count++; /* valid mirror */
 765
 766                 if (lre->lre_preferred || comp->lo_preferred_mirror < 0)
 767                         comp->lo_preferred_mirror = i - 1;
 768         }
 769         if (mirror_count == 0) {
 770                 CDEBUG(D_INODE, DFID
 771                        " doesn't have any valid mirrors\n",
 772                        PFID(lu_object_fid(lov2lu(lov))));
 773
 774                 GOTO(out, result = -EINVAL);
 775         }
 776
 777         LASSERT(comp->lo_preferred_mirror >= 0);
 778
 779         EXIT;
 780 out:
 781         return result > 0 ? 0 : result;
 782 }
 783
 784 static int lov_init_empty(const struct lu_env *env, struct lov_device *dev,
 785                           struct lov_object *lov, struct lov_stripe_md *lsm,
 786                           const struct cl_object_conf *conf,
 787                           union lov_layout_state *state)
 788 {
 789         return 0;
 790 }
 791
 792 static int lov_init_released(const struct lu_env *env,
 793                              struct lov_device *dev, struct lov_object *lov,
 794                              struct lov_stripe_md *lsm,
 795                              const struct cl_object_conf *conf,
 796                              union lov_layout_state *state)
 797 {
 798         LASSERT(lsm != NULL);
 799         LASSERT(lsm->lsm_is_released);
 800         LASSERT(lov->lo_lsm == NULL);
 801
 802         lov->lo_lsm = lsm_addref(lsm);
 803         return 0;
 804 }
 805
 806 static int lov_delete_empty(const struct lu_env *env, struct lov_object *lov,
 807                             union lov_layout_state *state)
 808 {
 809         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 810
 811         lov_layout_wait(env, lov);
 812         return 0;
 813 }
 814
 815 static int lov_delete_composite(const struct lu_env *env,
 816                                 struct lov_object *lov,
 817                                 union lov_layout_state *state)
 818 {
 819         struct lov_layout_entry *entry;
 820         struct lov_layout_composite *comp = &state->composite;
 821
 822         ENTRY;
 823
 824         dump_lsm(D_INODE, lov->lo_lsm);
 825
 826         lov_layout_wait(env, lov);
 827         if (comp->lo_entries)
 828                 lov_foreach_layout_entry(lov, entry)
 829                         lov_delete_raid0(env, lov, entry);
 830
 831         RETURN(0);
 832 }
 833
 834 static void lov_fini_empty(const struct lu_env *env, struct lov_object *lov,
 835                            union lov_layout_state *state)
 836 {
 837         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 838 }
 839
 840 static void lov_fini_composite(const struct lu_env *env,
 841                                struct lov_object *lov,
 842                                union lov_layout_state *state)
 843 {
 844         struct lov_layout_composite *comp = &state->composite;
 845         ENTRY;
 846
 847         if (comp->lo_entries != NULL) {
 848                 struct lov_layout_entry *entry;
 849
 850                 lov_foreach_layout_entry(lov, entry)
 851                         entry->lle_comp_ops->lco_fini(env, entry);
 852
 853                 OBD_FREE(comp->lo_entries,
 854                          comp->lo_entry_count * sizeof(*comp->lo_entries));
 855                 comp->lo_entries = NULL;
 856         }
 857
 858         if (comp->lo_mirrors != NULL) {
 859                 OBD_FREE(comp->lo_mirrors,
 860                          comp->lo_mirror_count * sizeof(*comp->lo_mirrors));
 861                 comp->lo_mirrors = NULL;
 862         }
 863
 864         memset(comp, 0, sizeof(*comp));
 865
 866         dump_lsm(D_INODE, lov->lo_lsm);
 867         lov_free_memmd(&lov->lo_lsm);
 868
 869         EXIT;
 870 }
 871
 872 static void lov_fini_released(const struct lu_env *env, struct lov_object *lov,
 873                                 union lov_layout_state *state)
 874 {
 875         ENTRY;
 876         dump_lsm(D_INODE, lov->lo_lsm);
 877         lov_free_memmd(&lov->lo_lsm);
 878         EXIT;
 879 }
 880
 881 static int lov_print_empty(const struct lu_env *env, void *cookie,
 882                            lu_printer_t p, const struct lu_object *o)
 883 {
 884         (*p)(env, cookie, "empty %d\n", lu2lov(o)->lo_layout_invalid);
 885         return 0;
 886 }
 887
 888 static int lov_print_composite(const struct lu_env *env, void *cookie,
 889                                lu_printer_t p, const struct lu_object *o)
 890 {
 891         struct lov_object *lov = lu2lov(o);
 892         struct lov_stripe_md *lsm = lov->lo_lsm;
 893         int i;
 894
 895         (*p)(env, cookie, "entries: %d, %s, lsm{%p 0x%08X %d %u}:\n",
 896              lsm->lsm_entry_count,
 897              lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 898              lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 899              lsm->lsm_layout_gen);
 900
 901         for (i = 0; i < lsm->lsm_entry_count; i++) {
 902                 struct lov_stripe_md_entry *lse = lsm->lsm_entries[i];
 903                 struct lov_layout_entry *lle = lov_entry(lov, i);
 904
 905                 (*p)(env, cookie,
 906                      DEXT ": { 0x%08X, %u, %#x, %u, %#x, %u, %u }\n",
 907                      PEXT(&lse->lsme_extent), lse->lsme_magic,
 908                      lse->lsme_id, lse->lsme_pattern, lse->lsme_layout_gen,
 909                      lse->lsme_flags, lse->lsme_stripe_count,
 910                      lse->lsme_stripe_size);
 911                 lov_print_raid0(env, cookie, p, lle);
 912         }
 913
 914         return 0;
 915 }
 916
 917 static int lov_print_released(const struct lu_env *env, void *cookie,
 918                                 lu_printer_t p, const struct lu_object *o)
 919 {
 920         struct lov_object       *lov = lu2lov(o);
 921         struct lov_stripe_md    *lsm = lov->lo_lsm;
 922
 923         (*p)(env, cookie,
 924                 "released: %s, lsm{%p 0x%08X %d %u}:\n",
 925                 lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 926                 lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 927                 lsm->lsm_layout_gen);
 928         return 0;
 929 }
 930
 931 /**
 932  * Implements cl_object_operations::coo_attr_get() method for an object
 933  * without stripes (LLT_EMPTY layout type).
 934  *
 935  * The only attributes this layer is authoritative in this case is
 936  * cl_attr::cat_blocks---it's 0.
 937  */
 938 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
 939                               struct cl_attr *attr)
 940 {
 941         attr->cat_blocks = 0;
 942         return 0;
 943 }
 944
 945 static int lov_attr_get_composite(const struct lu_env *env,
 946                                   struct cl_object *obj,
 947                                   struct cl_attr *attr)
 948 {
 949         struct lov_object       *lov = cl2lov(obj);
 950         struct lov_layout_entry *entry;
 951         int                      result = 0;
 952
 953         ENTRY;
 954
 955         attr->cat_size = 0;
 956         attr->cat_blocks = 0;
 957         lov_foreach_layout_entry(lov, entry) {
 958                 struct cl_attr *lov_attr = NULL;
 959                 int index = lov_layout_entry_index(lov, entry);
 960
 961                 if (!entry->lle_valid)
 962                         continue;
 963
 964                 /* PFL: This component has not been init-ed. */
 965                 if (!lsm_entry_inited(lov->lo_lsm, index))
 966                         continue;
 967
 968                 result = entry->lle_comp_ops->lco_getattr(env, lov, index,
 969                                                           entry, &lov_attr);
 970                 if (result < 0)
 971                         RETURN(result);
 972
 973                 if (lov_attr == NULL)
 974                         continue;
 975
 976                 CDEBUG(D_INODE, "COMP ID #%i: s=%llu m=%llu a=%llu c=%llu "
 977                        "b=%llu\n", index - 1, lov_attr->cat_size,
 978                        lov_attr->cat_mtime, lov_attr->cat_atime,
 979                        lov_attr->cat_ctime, lov_attr->cat_blocks);
 980
 981                 /* merge results */
 982                 attr->cat_blocks += lov_attr->cat_blocks;
 983                 if (attr->cat_size < lov_attr->cat_size)
 984                         attr->cat_size = lov_attr->cat_size;
 985                 if (attr->cat_kms < lov_attr->cat_kms)
 986                         attr->cat_kms = lov_attr->cat_kms;
 987                 if (attr->cat_atime < lov_attr->cat_atime)
 988                         attr->cat_atime = lov_attr->cat_atime;
 989                 if (attr->cat_ctime < lov_attr->cat_ctime)
 990                         attr->cat_ctime = lov_attr->cat_ctime;
 991                 if (attr->cat_mtime < lov_attr->cat_mtime)
 992                         attr->cat_mtime = lov_attr->cat_mtime;
 993         }
 994
 995         RETURN(0);
 996 }
 997
 998 const static struct lov_layout_operations lov_dispatch[] = {
 999         [LLT_EMPTY] = {
1000                 .llo_init      = lov_init_empty,
1001                 .llo_delete    = lov_delete_empty,
1002                 .llo_fini      = lov_fini_empty,
1003                 .llo_print     = lov_print_empty,
1004                 .llo_page_init = lov_page_init_empty,
1005                 .llo_lock_init = lov_lock_init_empty,
1006                 .llo_io_init   = lov_io_init_empty,
1007                 .llo_getattr   = lov_attr_get_empty,
1008         },
1009         [LLT_RELEASED] = {
1010                 .llo_init      = lov_init_released,
1011                 .llo_delete    = lov_delete_empty,
1012                 .llo_fini      = lov_fini_released,
1013                 .llo_print     = lov_print_released,
1014                 .llo_page_init = lov_page_init_empty,
1015                 .llo_lock_init = lov_lock_init_empty,
1016                 .llo_io_init   = lov_io_init_released,
1017                 .llo_getattr   = lov_attr_get_empty,
1018         },
1019         [LLT_COMP] = {
1020                 .llo_init      = lov_init_composite,
1021                 .llo_delete    = lov_delete_composite,
1022                 .llo_fini      = lov_fini_composite,
1023                 .llo_print     = lov_print_composite,
1024                 .llo_page_init = lov_page_init_composite,
1025                 .llo_lock_init = lov_lock_init_composite,
1026                 .llo_io_init   = lov_io_init_composite,
1027                 .llo_getattr   = lov_attr_get_composite,
1028         },
1029 };
1030
1031 /**
1032  * Performs a double-dispatch based on the layout type of an object.
1033  */
1034 #define LOV_2DISPATCH_NOLOCK(obj, op, ...)              \
1035 ({                                                      \
1036         struct lov_object *__obj = (obj);               \
1037         enum lov_layout_type __llt;                     \
1038                                                         \
1039         __llt = __obj->lo_type;                         \
1040         LASSERT(__llt < ARRAY_SIZE(lov_dispatch));      \
1041         lov_dispatch[__llt].op(__VA_ARGS__);            \
1042 })
1043
1044 /**
1045  * Return lov_layout_type associated with a given lsm
1046  */
1047 static enum lov_layout_type lov_type(struct lov_stripe_md *lsm)
1048 {
1049         if (lsm == NULL)
1050                 return LLT_EMPTY;
1051
1052         if (lsm->lsm_is_released)
1053                 return LLT_RELEASED;
1054
1055         if (lsm->lsm_magic == LOV_MAGIC_V1 ||
1056             lsm->lsm_magic == LOV_MAGIC_V3 ||
1057             lsm->lsm_magic == LOV_MAGIC_COMP_V1)
1058                 return LLT_COMP;
1059
1060         return LLT_EMPTY;
1061 }
1062
1063 static inline void lov_conf_freeze(struct lov_object *lov)
1064 {
1065         CDEBUG(D_INODE, "To take share lov(%p) owner %p/%p\n",
1066                 lov, lov->lo_owner, current);
1067         if (lov->lo_owner != current)
1068                 down_read(&lov->lo_type_guard);
1069 }
1070
1071 static inline void lov_conf_thaw(struct lov_object *lov)
1072 {
1073         CDEBUG(D_INODE, "To release share lov(%p) owner %p/%p\n",
1074                 lov, lov->lo_owner, current);
1075         if (lov->lo_owner != current)
1076                 up_read(&lov->lo_type_guard);
1077 }
1078
1079 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...)                       \
1080 ({                                                                      \
1081         struct lov_object                      *__obj = (obj);          \
1082         int                                     __lock = !!(lock);      \
1083         typeof(lov_dispatch[0].op(__VA_ARGS__)) __result;               \
1084                                                                         \
1085         if (__lock)                                                     \
1086                 lov_conf_freeze(__obj);                                 \
1087         __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__);          \
1088         if (__lock)                                                     \
1089                 lov_conf_thaw(__obj);                                   \
1090         __result;                                                       \
1091 })
1092
1093 /**
1094  * Performs a locked double-dispatch based on the layout type of an object.
1095  */
1096 #define LOV_2DISPATCH(obj, op, ...)                     \
1097         LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
1098
1099 #define LOV_2DISPATCH_VOID(obj, op, ...)                                \
1100 do {                                                                    \
1101         struct lov_object                      *__obj = (obj);          \
1102         enum lov_layout_type                    __llt;                  \
1103                                                                         \
1104         lov_conf_freeze(__obj);                                         \
1105         __llt = __obj->lo_type;                                         \
1106         LASSERT(__llt < ARRAY_SIZE(lov_dispatch));                      \
1107         lov_dispatch[__llt].op(__VA_ARGS__);                            \
1108         lov_conf_thaw(__obj);                                           \
1109 } while (0)
1110
1111 static void lov_conf_lock(struct lov_object *lov)
1112 {
1113         LASSERT(lov->lo_owner != current);
1114         down_write(&lov->lo_type_guard);
1115         LASSERT(lov->lo_owner == NULL);
1116         lov->lo_owner = current;
1117         CDEBUG(D_INODE, "Took exclusive lov(%p) owner %p\n",
1118                 lov, lov->lo_owner);
1119 }
1120
1121 static void lov_conf_unlock(struct lov_object *lov)
1122 {
1123         CDEBUG(D_INODE, "To release exclusive lov(%p) owner %p\n",
1124                 lov, lov->lo_owner);
1125         lov->lo_owner = NULL;
1126         up_write(&lov->lo_type_guard);
1127 }
1128
1129 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov)
1130 {
1131         struct l_wait_info lwi = { 0 };
1132         ENTRY;
1133
1134         while (atomic_read(&lov->lo_active_ios) > 0) {
1135                 CDEBUG(D_INODE, "file:"DFID" wait for active IO, now: %d.\n",
1136                         PFID(lu_object_fid(lov2lu(lov))),
1137                         atomic_read(&lov->lo_active_ios));
1138
1139                 l_wait_event(lov->lo_waitq,
1140                              atomic_read(&lov->lo_active_ios) == 0, &lwi);
1141         }
1142         RETURN(0);
1143 }
1144
1145 static int lov_layout_change(const struct lu_env *unused,
1146                              struct lov_object *lov, struct lov_stripe_md *lsm,
1147                              const struct cl_object_conf *conf)
1148 {
1149         enum lov_layout_type llt = lov_type(lsm);
1150         union lov_layout_state *state = &lov->u;
1151         const struct lov_layout_operations *old_ops;
1152         const struct lov_layout_operations *new_ops;
1153         struct lov_device *lov_dev = lov_object_dev(lov);
1154         struct lu_env *env;
1155         __u16 refcheck;
1156         int rc;
1157         ENTRY;
1158
1159         LASSERT(lov->lo_type < ARRAY_SIZE(lov_dispatch));
1160
1161         env = cl_env_get(&refcheck);
1162         if (IS_ERR(env))
1163                 RETURN(PTR_ERR(env));
1164
1165         LASSERT(llt < ARRAY_SIZE(lov_dispatch));
1166
1167         CDEBUG(D_INODE, DFID" from %s to %s\n",
1168                PFID(lu_object_fid(lov2lu(lov))),
1169                llt2str(lov->lo_type), llt2str(llt));
1170
1171         old_ops = &lov_dispatch[lov->lo_type];
1172         new_ops = &lov_dispatch[llt];
1173
1174         rc = cl_object_prune(env, &lov->lo_cl);
1175         if (rc != 0)
1176                 GOTO(out, rc);
1177
1178         rc = old_ops->llo_delete(env, lov, &lov->u);
1179         if (rc != 0)
1180                 GOTO(out, rc);
1181
1182         old_ops->llo_fini(env, lov, &lov->u);
1183
1184         LASSERT(atomic_read(&lov->lo_active_ios) == 0);
1185
1186         CDEBUG(D_INODE, DFID "Apply new layout lov %p, type %d\n",
1187                PFID(lu_object_fid(lov2lu(lov))), lov, llt);
1188
1189         /* page bufsize fixup */
1190         cl_object_header(&lov->lo_cl)->coh_page_bufsize -=
1191                 lov_page_slice_fixup(lov, NULL);
1192
1193         lov->lo_type = llt;
1194         rc = new_ops->llo_init(env, lov_dev, lov, lsm, conf, state);
1195         if (rc != 0) {
1196                 struct obd_device *obd = lov2obd(lov_dev->ld_lov);
1197
1198                 CERROR("%s: cannot apply new layout on "DFID" : rc = %d\n",
1199                        obd->obd_name, PFID(lu_object_fid(lov2lu(lov))), rc);
1200                 new_ops->llo_delete(env, lov, state);
1201                 new_ops->llo_fini(env, lov, state);
1202                 /* this file becomes an EMPTY file. */
1203                 lov->lo_type = LLT_EMPTY;
1204                 GOTO(out, rc);
1205         }
1206
1207 out:
1208         cl_env_put(env, &refcheck);
1209         RETURN(rc);
1210 }
1211
1212 /*****************************************************************************
1213  *
1214  * Lov object operations.
1215  *
1216  */
1217 int lov_object_init(const struct lu_env *env, struct lu_object *obj,
1218                     const struct lu_object_conf *conf)
1219 {
1220         struct lov_object            *lov   = lu2lov(obj);
1221         struct lov_device            *dev   = lov_object_dev(lov);
1222         const struct cl_object_conf  *cconf = lu2cl_conf(conf);
1223         union lov_layout_state       *set   = &lov->u;
1224         const struct lov_layout_operations *ops;
1225         struct lov_stripe_md *lsm = NULL;
1226         int rc;
1227         ENTRY;
1228
1229         init_rwsem(&lov->lo_type_guard);
1230         atomic_set(&lov->lo_active_ios, 0);
1231         init_waitqueue_head(&lov->lo_waitq);
1232         cl_object_page_init(lu2cl(obj), sizeof(struct lov_page));
1233
1234         lov->lo_type = LLT_EMPTY;
1235         if (cconf->u.coc_layout.lb_buf != NULL) {
1236                 lsm = lov_unpackmd(dev->ld_lov,
1237                                    cconf->u.coc_layout.lb_buf,
1238                                    cconf->u.coc_layout.lb_len);
1239                 if (IS_ERR(lsm))
1240                         RETURN(PTR_ERR(lsm));
1241
1242                 dump_lsm(D_INODE, lsm);
1243         }
1244
1245         /* no locking is necessary, as object is being created */
1246         lov->lo_type = lov_type(lsm);
1247         ops = &lov_dispatch[lov->lo_type];
1248         rc = ops->llo_init(env, dev, lov, lsm, cconf, set);
1249         if (rc != 0)
1250                 GOTO(out_lsm, rc);
1251
1252 out_lsm:
1253         lov_lsm_put(lsm);
1254
1255         RETURN(rc);
1256 }
1257
1258 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
1259                         const struct cl_object_conf *conf)
1260 {
1261         struct lov_stripe_md    *lsm = NULL;
1262         struct lov_object       *lov = cl2lov(obj);
1263         int                      result = 0;
1264         ENTRY;
1265
1266         if (conf->coc_opc == OBJECT_CONF_SET &&
1267             conf->u.coc_layout.lb_buf != NULL) {
1268                 lsm = lov_unpackmd(lov_object_dev(lov)->ld_lov,
1269                                    conf->u.coc_layout.lb_buf,
1270                                    conf->u.coc_layout.lb_len);
1271                 if (IS_ERR(lsm))
1272                         RETURN(PTR_ERR(lsm));
1273                 dump_lsm(D_INODE, lsm);
1274         }
1275
1276         lov_conf_lock(lov);
1277         if (conf->coc_opc == OBJECT_CONF_INVALIDATE) {
1278                 lov->lo_layout_invalid = true;
1279                 GOTO(out, result = 0);
1280         }
1281
1282         if (conf->coc_opc == OBJECT_CONF_WAIT) {
1283                 if (lov->lo_layout_invalid &&
1284                     atomic_read(&lov->lo_active_ios) > 0) {
1285                         lov_conf_unlock(lov);
1286                         result = lov_layout_wait(env, lov);
1287                         lov_conf_lock(lov);
1288                 }
1289                 GOTO(out, result);
1290         }
1291
1292         LASSERT(conf->coc_opc == OBJECT_CONF_SET);
1293
1294         if ((lsm == NULL && lov->lo_lsm == NULL) ||
1295             ((lsm != NULL && lov->lo_lsm != NULL) &&
1296              (lov->lo_lsm->lsm_layout_gen == lsm->lsm_layout_gen) &&
1297              (lov->lo_lsm->lsm_entries[0]->lsme_pattern ==
1298               lsm->lsm_entries[0]->lsme_pattern))) {
1299                 /* same version of layout */
1300                 lov->lo_layout_invalid = false;
1301                 GOTO(out, result = 0);
1302         }
1303
1304         /* will change layout - check if there still exists active IO. */
1305         if (atomic_read(&lov->lo_active_ios) > 0) {
1306                 lov->lo_layout_invalid = true;
1307                 GOTO(out, result = -EBUSY);
1308         }
1309
1310         result = lov_layout_change(env, lov, lsm, conf);
1311         lov->lo_layout_invalid = result != 0;
1312         EXIT;
1313
1314 out:
1315         lov_conf_unlock(lov);
1316         lov_lsm_put(lsm);
1317         CDEBUG(D_INODE, DFID" lo_layout_invalid=%d\n",
1318                PFID(lu_object_fid(lov2lu(lov))), lov->lo_layout_invalid);
1319         RETURN(result);
1320 }
1321
1322 static void lov_object_delete(const struct lu_env *env, struct lu_object *obj)
1323 {
1324         struct lov_object *lov = lu2lov(obj);
1325
1326         ENTRY;
1327         LOV_2DISPATCH_VOID(lov, llo_delete, env, lov, &lov->u);
1328         EXIT;
1329 }
1330
1331 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
1332 {
1333         struct lov_object *lov = lu2lov(obj);
1334
1335         ENTRY;
1336         LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
1337         lu_object_fini(obj);
1338         OBD_SLAB_FREE_PTR(lov, lov_object_kmem);
1339         EXIT;
1340 }
1341
1342 static int lov_object_print(const struct lu_env *env, void *cookie,
1343                             lu_printer_t p, const struct lu_object *o)
1344 {
1345         return LOV_2DISPATCH_NOLOCK(lu2lov(o), llo_print, env, cookie, p, o);
1346 }
1347
1348 int lov_page_init(const struct lu_env *env, struct cl_object *obj,
1349                   struct cl_page *page, pgoff_t index)
1350 {
1351         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_page_init, env, obj, page,
1352                                     index);
1353 }
1354
1355 /**
1356  * Implements cl_object_operations::clo_io_init() method for lov
1357  * layer. Dispatches to the appropriate layout io initialization method.
1358  */
1359 int lov_io_init(const struct lu_env *env, struct cl_object *obj,
1360                 struct cl_io *io)
1361 {
1362         CL_IO_SLICE_CLEAN(lov_env_io(env), lis_preserved);
1363
1364         CDEBUG(D_INODE, DFID "io %p type %d ignore/verify layout %d/%d\n",
1365                PFID(lu_object_fid(&obj->co_lu)), io, io->ci_type,
1366                io->ci_ignore_layout, io->ci_verify_layout);
1367
1368         /* IO type CIT_MISC with ci_ignore_layout set are usually invoked from
1369          * the OSC layer. It shouldn't take lov layout conf lock in that case,
1370          * because as long as the OSC object exists, the layout can't be
1371          * reconfigured. */
1372         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
1373                         !(io->ci_ignore_layout && io->ci_type == CIT_MISC),
1374                         env, obj, io);
1375 }
1376
1377 /**
1378  * An implementation of cl_object_operations::clo_attr_get() method for lov
1379  * layer. For raid0 layout this collects and merges attributes of all
1380  * sub-objects.
1381  */
1382 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
1383                         struct cl_attr *attr)
1384 {
1385         /* do not take lock, as this function is called under a
1386          * spin-lock. Layout is protected from changing by ongoing IO. */
1387         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
1388 }
1389
1390 static int lov_attr_update(const struct lu_env *env, struct cl_object *obj,
1391                            const struct cl_attr *attr, unsigned valid)
1392 {
1393         /*
1394          * No dispatch is required here, as no layout implements this.
1395          */
1396         return 0;
1397 }
1398
1399 int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
1400                   struct cl_lock *lock, const struct cl_io *io)
1401 {
1402         /* No need to lock because we've taken one refcount of layout.  */
1403         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_lock_init, env, obj, lock,
1404                                     io);
1405 }
1406
1407 /**
1408  * We calculate on which OST the mapping will end. If the length of mapping
1409  * is greater than (stripe_size * stripe_count) then the last_stripe will
1410  * will be one just before start_stripe. Else we check if the mapping
1411  * intersects each OST and find last_stripe.
1412  * This function returns the last_stripe and also sets the stripe_count
1413  * over which the mapping is spread
1414  *
1415  * \param lsm [in]              striping information for the file
1416  * \param index [in]            stripe component index
1417  * \param ext [in]              logical extent of mapping
1418  * \param start_stripe [in]     starting stripe of the mapping
1419  * \param stripe_count [out]    the number of stripes across which to map is
1420  *                              returned
1421  *
1422  * \retval last_stripe          return the last stripe of the mapping
1423  */
1424 static int fiemap_calc_last_stripe(struct lov_stripe_md *lsm, int index,
1425                                    struct lu_extent *ext,
1426                                    int start_stripe, int *stripe_count)
1427 {
1428         struct lov_stripe_md_entry *lsme = lsm->lsm_entries[index];
1429         int last_stripe;
1430         u64 obd_start;
1431         u64 obd_end;
1432         int i, j;
1433
1434         if (ext->e_end - ext->e_start >
1435             lsme->lsme_stripe_size * lsme->lsme_stripe_count) {
1436                 last_stripe = (start_stripe < 1 ? lsme->lsme_stripe_count - 1 :
1437                                                   start_stripe - 1);
1438                 *stripe_count = lsme->lsme_stripe_count;
1439         } else {
1440                 for (j = 0, i = start_stripe; j < lsme->lsme_stripe_count;
1441                      i = (i + 1) % lsme->lsme_stripe_count, j++) {
1442                         if ((lov_stripe_intersects(lsm, index,  i, ext,
1443                                                    &obd_start, &obd_end)) == 0)
1444                                 break;
1445                 }
1446                 *stripe_count = j;
1447                 last_stripe = (start_stripe + j - 1) % lsme->lsme_stripe_count;
1448         }
1449
1450         return last_stripe;
1451 }
1452
1453 /**
1454  * Set fe_device and copy extents from local buffer into main return buffer.
1455  *
1456  * \param fiemap [out]          fiemap to hold all extents
1457  * \param lcl_fm_ext [in]       array of fiemap extents get from OSC layer
1458  * \param ost_index [in]        OST index to be written into the fm_device
1459  *                              field for each extent
1460  * \param ext_count [in]        number of extents to be copied
1461  * \param current_extent [in]   where to start copying in the extent array
1462  */
1463 static void fiemap_prepare_and_copy_exts(struct fiemap *fiemap,
1464                                          struct fiemap_extent *lcl_fm_ext,
1465                                          int ost_index, unsigned int ext_count,
1466                                          int current_extent)
1467 {
1468         char            *to;
1469         unsigned int    ext;
1470
1471         for (ext = 0; ext < ext_count; ext++) {
1472                 lcl_fm_ext[ext].fe_device = ost_index;
1473                 lcl_fm_ext[ext].fe_flags |= FIEMAP_EXTENT_NET;
1474         }
1475
1476         /* Copy fm_extent's from fm_local to return buffer */
1477         to = (char *)fiemap + fiemap_count_to_size(current_extent);
1478         memcpy(to, lcl_fm_ext, ext_count * sizeof(struct fiemap_extent));
1479 }
1480
1481 #define FIEMAP_BUFFER_SIZE 4096
1482
1483 /**
1484  * Non-zero fe_logical indicates that this is a continuation FIEMAP
1485  * call. The local end offset and the device are sent in the first
1486  * fm_extent. This function calculates the stripe number from the index.
1487  * This function returns a stripe_no on which mapping is to be restarted.
1488  *
1489  * This function returns fm_end_offset which is the in-OST offset at which
1490  * mapping should be restarted. If fm_end_offset=0 is returned then caller
1491  * will re-calculate proper offset in next stripe.
1492  * Note that the first extent is passed to lov_get_info via the value field.
1493  *
1494  * \param fiemap [in]           fiemap request header
1495  * \param lsm [in]              striping information for the file
1496  * \param index [in]            stripe component index
1497  * \param ext [in]              logical extent of mapping
1498  * \param start_stripe [out]    starting stripe will be returned in this
1499  */
1500 static u64 fiemap_calc_fm_end_offset(struct fiemap *fiemap,
1501                                      struct lov_stripe_md *lsm,
1502                                      int index, struct lu_extent *ext,
1503                                      int *start_stripe)
1504 {
1505         struct lov_stripe_md_entry *lsme = lsm->lsm_entries[index];
1506         u64 local_end = fiemap->fm_extents[0].fe_logical;
1507         u64 lun_start;
1508         u64 lun_end;
1509         u64 fm_end_offset;
1510         int stripe_no = -1;
1511         int i;
1512
1513         if (fiemap->fm_extent_count == 0 ||
1514             fiemap->fm_extents[0].fe_logical == 0)
1515                 return 0;
1516
1517         /* Find out stripe_no from ost_index saved in the fe_device */
1518         for (i = 0; i < lsme->lsme_stripe_count; i++) {
1519                 struct lov_oinfo *oinfo = lsme->lsme_oinfo[i];
1520
1521                 if (lov_oinfo_is_dummy(oinfo))
1522                         continue;
1523
1524                 if (oinfo->loi_ost_idx == fiemap->fm_extents[0].fe_device) {
1525                         stripe_no = i;
1526                         break;
1527                 }
1528         }
1529
1530         if (stripe_no == -1)
1531                 return -EINVAL;
1532
1533         /* If we have finished mapping on previous device, shift logical
1534          * offset to start of next device */
1535         if (lov_stripe_intersects(lsm, index, stripe_no, ext,
1536                                    &lun_start, &lun_end) != 0 &&
1537             local_end < lun_end) {
1538                 fm_end_offset = local_end;
1539                 *start_stripe = stripe_no;
1540         } else {
1541                 /* This is a special value to indicate that caller should
1542                  * calculate offset in next stripe. */
1543                 fm_end_offset = 0;
1544                 *start_stripe = (stripe_no + 1) % lsme->lsme_stripe_count;
1545         }
1546
1547         return fm_end_offset;
1548 }
1549
1550 struct fiemap_state {
1551         struct fiemap           *fs_fm;
1552         struct lu_extent        fs_ext;
1553         u64                     fs_length;
1554         u64                     fs_end_offset;
1555         int                     fs_cur_extent;
1556         int                     fs_cnt_need;
1557         int                     fs_start_stripe;
1558         int                     fs_last_stripe;
1559         bool                    fs_device_done;
1560         bool                    fs_finish_stripe;
1561         bool                    fs_enough;
1562 };
1563
1564 static struct cl_object *lov_find_subobj(const struct lu_env *env,
1565                                          struct lov_object *lov,
1566                                          struct lov_stripe_md *lsm,
1567                                          int index)
1568 {
1569         struct lov_device       *dev = lu2lov_dev(lov2lu(lov)->lo_dev);
1570         struct lov_thread_info  *lti = lov_env_info(env);
1571         struct lu_fid           *ofid = &lti->lti_fid;
1572         struct lov_oinfo        *oinfo;
1573         struct cl_device        *subdev;
1574         int                     entry = lov_comp_entry(index);
1575         int                     stripe = lov_comp_stripe(index);
1576         int                     ost_idx;
1577         int                     rc;
1578         struct cl_object        *result;
1579
1580         if (lov->lo_type != LLT_COMP)
1581                 GOTO(out, result = NULL);
1582
1583         if (entry >= lsm->lsm_entry_count ||
1584             stripe >= lsm->lsm_entries[entry]->lsme_stripe_count)
1585                 GOTO(out, result = NULL);
1586
1587         oinfo = lsm->lsm_entries[entry]->lsme_oinfo[stripe];
1588         ost_idx = oinfo->loi_ost_idx;
1589         rc = ostid_to_fid(ofid, &oinfo->loi_oi, ost_idx);
1590         if (rc != 0)
1591                 GOTO(out, result = NULL);
1592
1593         subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
1594         result = lov_sub_find(env, subdev, ofid, NULL);
1595 out:
1596         if (result == NULL)
1597                 result = ERR_PTR(-EINVAL);
1598         return result;
1599 }
1600
1601 int fiemap_for_stripe(const struct lu_env *env, struct cl_object *obj,
1602                       struct lov_stripe_md *lsm, struct fiemap *fiemap,
1603                       size_t *buflen, struct ll_fiemap_info_key *fmkey,
1604                       int index, int stripeno, struct fiemap_state *fs)
1605 {
1606         struct lov_stripe_md_entry *lsme = lsm->lsm_entries[index];
1607         struct cl_object *subobj;
1608         struct lov_obd *lov = lu2lov_dev(obj->co_lu.lo_dev)->ld_lov;
1609         struct fiemap_extent *fm_ext = &fs->fs_fm->fm_extents[0];
1610         u64 req_fm_len; /* Stores length of required mapping */
1611         u64 len_mapped_single_call;
1612         u64 lun_start;
1613         u64 lun_end;
1614         u64 obd_object_end;
1615         unsigned int ext_count;
1616         /* EOF for object */
1617         bool ost_eof = false;
1618         /* done with required mapping for this OST? */
1619         bool ost_done = false;
1620         int ost_index;
1621         int rc = 0;
1622
1623         fs->fs_device_done = false;
1624         /* Find out range of mapping on this stripe */
1625         if ((lov_stripe_intersects(lsm, index, stripeno, &fs->fs_ext,
1626                                    &lun_start, &obd_object_end)) == 0)
1627                 return 0;
1628
1629         if (lov_oinfo_is_dummy(lsme->lsme_oinfo[stripeno]))
1630                 return -EIO;
1631
1632         /* If this is a continuation FIEMAP call and we are on
1633          * starting stripe then lun_start needs to be set to
1634          * end_offset */
1635         if (fs->fs_end_offset != 0 && stripeno == fs->fs_start_stripe)
1636                 lun_start = fs->fs_end_offset;
1637         lun_end = lov_size_to_stripe(lsm, index, fs->fs_ext.e_end, stripeno);
1638         if (lun_start == lun_end)
1639                 return 0;
1640
1641         req_fm_len = obd_object_end - lun_start;
1642         fs->fs_fm->fm_length = 0;
1643         len_mapped_single_call = 0;
1644
1645         /* find lobsub object */
1646         subobj = lov_find_subobj(env, cl2lov(obj), lsm,
1647                                  lov_comp_index(index, stripeno));
1648         if (IS_ERR(subobj))
1649                 return PTR_ERR(subobj);
1650         /* If the output buffer is very large and the objects have many
1651          * extents we may need to loop on a single OST repeatedly */
1652         do {
1653                 if (fiemap->fm_extent_count > 0) {
1654                         /* Don't get too many extents. */
1655                         if (fs->fs_cur_extent + fs->fs_cnt_need >
1656                             fiemap->fm_extent_count)
1657                                 fs->fs_cnt_need = fiemap->fm_extent_count -
1658                                                   fs->fs_cur_extent;
1659                 }
1660
1661                 lun_start += len_mapped_single_call;
1662                 fs->fs_fm->fm_length = req_fm_len - len_mapped_single_call;
1663                 req_fm_len = fs->fs_fm->fm_length;
1664                 /**
1665                  * If we've collected enough extent map, we'd request 1 more,
1666                  * to see whether we coincidentally finished all available
1667                  * extent map, so that FIEMAP_EXTENT_LAST would be set.
1668                  */
1669                 fs->fs_fm->fm_extent_count = fs->fs_enough ?
1670                                              1 : fs->fs_cnt_need;
1671                 fs->fs_fm->fm_mapped_extents = 0;
1672                 fs->fs_fm->fm_flags = fiemap->fm_flags;
1673
1674                 ost_index = lsme->lsme_oinfo[stripeno]->loi_ost_idx;
1675
1676                 if (ost_index < 0 || ost_index >= lov->desc.ld_tgt_count)
1677                         GOTO(obj_put, rc = -EINVAL);
1678                 /* If OST is inactive, return extent with UNKNOWN flag. */
1679                 if (!lov->lov_tgts[ost_index]->ltd_active) {
1680                         fs->fs_fm->fm_flags |= FIEMAP_EXTENT_LAST;
1681                         fs->fs_fm->fm_mapped_extents = 1;
1682
1683                         fm_ext[0].fe_logical = lun_start;
1684                         fm_ext[0].fe_length = obd_object_end - lun_start;
1685                         fm_ext[0].fe_flags |= FIEMAP_EXTENT_UNKNOWN;
1686
1687                         goto inactive_tgt;
1688                 }
1689
1690                 fs->fs_fm->fm_start = lun_start;
1691                 fs->fs_fm->fm_flags &= ~FIEMAP_FLAG_DEVICE_ORDER;
1692                 memcpy(&fmkey->lfik_fiemap, fs->fs_fm, sizeof(*fs->fs_fm));
1693                 *buflen = fiemap_count_to_size(fs->fs_fm->fm_extent_count);
1694
1695                 rc = cl_object_fiemap(env, subobj, fmkey, fs->fs_fm, buflen);
1696                 if (rc != 0)
1697                         GOTO(obj_put, rc);
1698 inactive_tgt:
1699                 ext_count = fs->fs_fm->fm_mapped_extents;
1700                 if (ext_count == 0) {
1701                         ost_done = true;
1702                         fs->fs_device_done = true;
1703                         /* If last stripe has hold at the end,
1704                          * we need to return */
1705                         if (stripeno == fs->fs_last_stripe) {
1706                                 fiemap->fm_mapped_extents = 0;
1707                                 fs->fs_finish_stripe = true;
1708                                 GOTO(obj_put, rc);
1709                         }
1710                         break;
1711                 } else if (fs->fs_enough) {
1712                         /*
1713                          * We've collected enough extents and there are
1714                          * more extents after it.
1715                          */
1716                         GOTO(obj_put, rc);
1717                 }
1718
1719                 /* If we just need num of extents, got to next device */
1720                 if (fiemap->fm_extent_count == 0) {
1721                         fs->fs_cur_extent += ext_count;
1722                         break;
1723                 }
1724
1725                 /* prepare to copy retrived map extents */
1726                 len_mapped_single_call = fm_ext[ext_count - 1].fe_logical +
1727                                          fm_ext[ext_count - 1].fe_length -
1728                                          lun_start;
1729
1730                 /* Have we finished mapping on this device? */
1731                 if (req_fm_len <= len_mapped_single_call) {
1732                         ost_done = true;
1733                         fs->fs_device_done = true;
1734                 }
1735
1736                 /* Clear the EXTENT_LAST flag which can be present on
1737                  * the last extent */
1738                 if (fm_ext[ext_count - 1].fe_flags & FIEMAP_EXTENT_LAST)
1739                         fm_ext[ext_count - 1].fe_flags &= ~FIEMAP_EXTENT_LAST;
1740                 if (lov_stripe_size(lsm, index,
1741                                     fm_ext[ext_count - 1].fe_logical +
1742                                     fm_ext[ext_count - 1].fe_length,
1743                                     stripeno) >= fmkey->lfik_oa.o_size) {
1744                         ost_eof = true;
1745                         fs->fs_device_done = true;
1746                 }
1747
1748                 fiemap_prepare_and_copy_exts(fiemap, fm_ext, ost_index,
1749                                              ext_count, fs->fs_cur_extent);
1750                 fs->fs_cur_extent += ext_count;
1751
1752                 /* Ran out of available extents? */
1753                 if (fs->fs_cur_extent >= fiemap->fm_extent_count)
1754                         fs->fs_enough = true;
1755         } while (!ost_done && !ost_eof);
1756
1757         if (stripeno == fs->fs_last_stripe)
1758                 fs->fs_finish_stripe = true;
1759 obj_put:
1760         cl_object_put(env, subobj);
1761
1762         return rc;
1763 }
1764
1765 /**
1766  * Break down the FIEMAP request and send appropriate calls to individual OSTs.
1767  * This also handles the restarting of FIEMAP calls in case mapping overflows
1768  * the available number of extents in single call.
1769  *
1770  * \param env [in]              lustre environment
1771  * \param obj [in]              file object
1772  * \param fmkey [in]            fiemap request header and other info
1773  * \param fiemap [out]          fiemap buffer holding retrived map extents
1774  * \param buflen [in/out]       max buffer length of @fiemap, when iterate
1775  *                              each OST, it is used to limit max map needed
1776  * \retval 0    success
1777  * \retval < 0  error
1778  */
1779 static int lov_object_fiemap(const struct lu_env *env, struct cl_object *obj,
1780                              struct ll_fiemap_info_key *fmkey,
1781                              struct fiemap *fiemap, size_t *buflen)
1782 {
1783         struct lov_stripe_md_entry *lsme;
1784         struct lov_stripe_md *lsm;
1785         struct fiemap *fm_local = NULL;
1786         loff_t whole_start;
1787         loff_t whole_end;
1788         int entry;
1789         int start_entry;
1790         int end_entry;
1791         int cur_stripe = 0;
1792         int stripe_count;
1793         unsigned int buffer_size = FIEMAP_BUFFER_SIZE;
1794         int rc = 0;
1795         struct fiemap_state fs = { 0 };
1796         ENTRY;
1797
1798         lsm = lov_lsm_addref(cl2lov(obj));
1799         if (lsm == NULL)
1800                 RETURN(-ENODATA);
1801
1802         if (!(fiemap->fm_flags & FIEMAP_FLAG_DEVICE_ORDER)) {
1803                 /**
1804                  * If the entry count > 1 or stripe_count > 1 and the
1805                  * application does not understand DEVICE_ORDER flag,
1806                  * it cannot interpret the extents correctly.
1807                  */
1808                 if (lsm->lsm_entry_count > 1 ||
1809                     (lsm->lsm_entry_count == 1 &&
1810                      lsm->lsm_entries[0]->lsme_stripe_count > 1))
1811                         GOTO(out_lsm, rc = -ENOTSUPP);
1812         }
1813
1814         /* No support for DOM layout yet. */
1815         if (lsme_is_dom(lsm->lsm_entries[0]))
1816                 GOTO(out_lsm, rc = -ENOTSUPP);
1817
1818         if (lsm->lsm_is_released) {
1819                 if (fiemap->fm_start < fmkey->lfik_oa.o_size) {
1820                         /**
1821                          * released file, return a minimal FIEMAP if
1822                          * request fits in file-size.
1823                          */
1824                         fiemap->fm_mapped_extents = 1;
1825                         fiemap->fm_extents[0].fe_logical = fiemap->fm_start;
1826                         if (fiemap->fm_start + fiemap->fm_length <
1827                             fmkey->lfik_oa.o_size)
1828                                 fiemap->fm_extents[0].fe_length =
1829                                         fiemap->fm_length;
1830                         else
1831                                 fiemap->fm_extents[0].fe_length =
1832                                         fmkey->lfik_oa.o_size -
1833                                         fiemap->fm_start;
1834                         fiemap->fm_extents[0].fe_flags |=
1835                                 FIEMAP_EXTENT_UNKNOWN | FIEMAP_EXTENT_LAST;
1836                 }
1837                 GOTO(out_lsm, rc = 0);
1838         }
1839
1840         /* buffer_size is small to hold fm_extent_count of extents. */
1841         if (fiemap_count_to_size(fiemap->fm_extent_count) < buffer_size)
1842                 buffer_size = fiemap_count_to_size(fiemap->fm_extent_count);
1843
1844         OBD_ALLOC_LARGE(fm_local, buffer_size);
1845         if (fm_local == NULL)
1846                 GOTO(out_lsm, rc = -ENOMEM);
1847
1848         /**
1849          * Requested extent count exceeds the fiemap buffer size, shrink our
1850          * ambition.
1851          */
1852         if (fiemap_count_to_size(fiemap->fm_extent_count) > *buflen)
1853                 fiemap->fm_extent_count = fiemap_size_to_count(*buflen);
1854         if (fiemap->fm_extent_count == 0)
1855                 fs.fs_cnt_need = 0;
1856
1857         fs.fs_enough = false;
1858         fs.fs_cur_extent = 0;
1859         fs.fs_fm = fm_local;
1860         fs.fs_cnt_need = fiemap_size_to_count(buffer_size);
1861
1862         whole_start = fiemap->fm_start;
1863         /* whole_start is beyond the end of the file */
1864         if (whole_start > fmkey->lfik_oa.o_size)
1865                 GOTO(out_fm_local, rc = -EINVAL);
1866         whole_end = (fiemap->fm_length == OBD_OBJECT_EOF) ?
1867                                         fmkey->lfik_oa.o_size :
1868                                         whole_start + fiemap->fm_length - 1;
1869         /**
1870          * If fiemap->fm_length != OBD_OBJECT_EOF but whole_end exceeds file
1871          * size
1872          */
1873         if (whole_end > fmkey->lfik_oa.o_size)
1874                 whole_end = fmkey->lfik_oa.o_size;
1875
1876         start_entry = lov_lsm_entry(lsm, whole_start);
1877         end_entry = lov_lsm_entry(lsm, whole_end);
1878         if (end_entry == -1)
1879                 end_entry = lsm->lsm_entry_count - 1;
1880
1881         if (start_entry == -1 || end_entry == -1)
1882                 GOTO(out_fm_local, rc = -EINVAL);
1883
1884         /* TODO: rewrite it with lov_foreach_io_layout() */
1885         for (entry = start_entry; entry <= end_entry; entry++) {
1886                 lsme = lsm->lsm_entries[entry];
1887
1888                 if (!lsme_inited(lsme))
1889                         break;
1890
1891                 if (entry == start_entry)
1892                         fs.fs_ext.e_start = whole_start;
1893                 else
1894                         fs.fs_ext.e_start = lsme->lsme_extent.e_start;
1895                 if (entry == end_entry)
1896                         fs.fs_ext.e_end = whole_end;
1897                 else
1898                         fs.fs_ext.e_end = lsme->lsme_extent.e_end - 1;
1899                 fs.fs_length = fs.fs_ext.e_end - fs.fs_ext.e_start + 1;
1900
1901                 /* Calculate start stripe, last stripe and length of mapping */
1902                 fs.fs_start_stripe = lov_stripe_number(lsm, entry,
1903                                                        fs.fs_ext.e_start);
1904                 fs.fs_last_stripe = fiemap_calc_last_stripe(lsm, entry,
1905                                         &fs.fs_ext, fs.fs_start_stripe,
1906                                         &stripe_count);
1907                 fs.fs_end_offset = fiemap_calc_fm_end_offset(fiemap, lsm, entry,
1908                                         &fs.fs_ext, &fs.fs_start_stripe);
1909                 /* Check each stripe */
1910                 for (cur_stripe = fs.fs_start_stripe; stripe_count > 0;
1911                      --stripe_count,
1912                      cur_stripe = (cur_stripe + 1) % lsme->lsme_stripe_count) {
1913                         rc = fiemap_for_stripe(env, obj, lsm, fiemap, buflen,
1914                                                fmkey, entry, cur_stripe, &fs);
1915                         if (rc < 0)
1916                                 GOTO(out_fm_local, rc);
1917                         if (fs.fs_enough)
1918                                 GOTO(finish, rc);
1919                         if (fs.fs_finish_stripe)
1920                                 break;
1921                 } /* for each stripe */
1922         } /* for covering layout component */
1923         /*
1924          * We've traversed all components, set @entry to the last component
1925          * entry, it's for the last stripe check.
1926          */
1927         entry--;
1928 finish:
1929         /* Indicate that we are returning device offsets unless file just has
1930          * single stripe */
1931         if (lsm->lsm_entry_count > 1 ||
1932             (lsm->lsm_entry_count == 1 &&
1933              lsm->lsm_entries[0]->lsme_stripe_count > 1))
1934                 fiemap->fm_flags |= FIEMAP_FLAG_DEVICE_ORDER;
1935
1936         if (fiemap->fm_extent_count == 0)
1937                 goto skip_last_device_calc;
1938
1939         /* Check if we have reached the last stripe and whether mapping for that
1940          * stripe is done. */
1941         if ((cur_stripe == fs.fs_last_stripe) && fs.fs_device_done)
1942                 fiemap->fm_extents[fs.fs_cur_extent - 1].fe_flags |=
1943                                                              FIEMAP_EXTENT_LAST;
1944 skip_last_device_calc:
1945         fiemap->fm_mapped_extents = fs.fs_cur_extent;
1946 out_fm_local:
1947         OBD_FREE_LARGE(fm_local, buffer_size);
1948
1949 out_lsm:
1950         lov_lsm_put(lsm);
1951         return rc;
1952 }
1953
1954 static int lov_object_getstripe(const struct lu_env *env, struct cl_object *obj,
1955                                 struct lov_user_md __user *lum, size_t size)
1956 {
1957         struct lov_object       *lov = cl2lov(obj);
1958         struct lov_stripe_md    *lsm;
1959         int                     rc = 0;
1960         ENTRY;
1961
1962         lsm = lov_lsm_addref(lov);
1963         if (lsm == NULL)
1964                 RETURN(-ENODATA);
1965
1966         rc = lov_getstripe(env, cl2lov(obj), lsm, lum, size);
1967         lov_lsm_put(lsm);
1968         RETURN(rc);
1969 }
1970
1971 static int lov_object_layout_get(const struct lu_env *env,
1972                                  struct cl_object *obj,
1973                                  struct cl_layout *cl)
1974 {
1975         struct lov_object *lov = cl2lov(obj);
1976         struct lov_stripe_md *lsm = lov_lsm_addref(lov);
1977         struct lu_buf *buf = &cl->cl_buf;
1978         ssize_t rc;
1979         ENTRY;
1980
1981         if (lsm == NULL) {
1982                 cl->cl_size = 0;
1983                 cl->cl_layout_gen = CL_LAYOUT_GEN_EMPTY;
1984
1985                 RETURN(0);
1986         }
1987
1988         cl->cl_size = lov_comp_md_size(lsm);
1989         cl->cl_layout_gen = lsm->lsm_layout_gen;
1990         cl->cl_dom_comp_size = 0;
1991         if (lsm_is_composite(lsm->lsm_magic)) {
1992                 struct lov_stripe_md_entry *lsme = lsm->lsm_entries[0];
1993
1994                 cl->cl_is_composite = true;
1995
1996                 if (lsme_is_dom(lsme))
1997                         cl->cl_dom_comp_size = lsme->lsme_extent.e_end;
1998         } else {
1999                 cl->cl_is_composite = false;
2000         }
2001
2002         rc = lov_lsm_pack(lsm, buf->lb_buf, buf->lb_len);
2003         lov_lsm_put(lsm);
2004
2005         RETURN(rc < 0 ? rc : 0);
2006 }
2007
2008 static loff_t lov_object_maxbytes(struct cl_object *obj)
2009 {
2010         struct lov_object *lov = cl2lov(obj);
2011         struct lov_stripe_md *lsm = lov_lsm_addref(lov);
2012         loff_t maxbytes;
2013
2014         if (lsm == NULL)
2015                 return LLONG_MAX;
2016
2017         maxbytes = lsm->lsm_maxbytes;
2018
2019         lov_lsm_put(lsm);
2020
2021         return maxbytes;
2022 }
2023
2024 static const struct cl_object_operations lov_ops = {
2025         .coo_page_init    = lov_page_init,
2026         .coo_lock_init    = lov_lock_init,
2027         .coo_io_init      = lov_io_init,
2028         .coo_attr_get     = lov_attr_get,
2029         .coo_attr_update  = lov_attr_update,
2030         .coo_conf_set     = lov_conf_set,
2031         .coo_getstripe    = lov_object_getstripe,
2032         .coo_layout_get   = lov_object_layout_get,
2033         .coo_maxbytes     = lov_object_maxbytes,
2034         .coo_fiemap       = lov_object_fiemap,
2035 };
2036
2037 static const struct lu_object_operations lov_lu_obj_ops = {
2038         .loo_object_init      = lov_object_init,
2039         .loo_object_delete    = lov_object_delete,
2040         .loo_object_release   = NULL,
2041         .loo_object_free      = lov_object_free,
2042         .loo_object_print     = lov_object_print,
2043         .loo_object_invariant = NULL
2044 };
2045
2046 struct lu_object *lov_object_alloc(const struct lu_env *env,
2047                                    const struct lu_object_header *unused,
2048                                    struct lu_device *dev)
2049 {
2050         struct lov_object *lov;
2051         struct lu_object  *obj;
2052
2053         ENTRY;
2054         OBD_SLAB_ALLOC_PTR_GFP(lov, lov_object_kmem, GFP_NOFS);
2055         if (lov != NULL) {
2056                 obj = lov2lu(lov);
2057                 lu_object_init(obj, NULL, dev);
2058                 lov->lo_cl.co_ops = &lov_ops;
2059                 lov->lo_type = -1; /* invalid, to catch uninitialized type */
2060                 /*
2061                  * object io operation vector (cl_object::co_iop) is installed
2062                  * later in lov_object_init(), as different vectors are used
2063                  * for object with different layouts.
2064                  */
2065                 obj->lo_ops = &lov_lu_obj_ops;
2066         } else
2067                 obj = NULL;
2068         RETURN(obj);
2069 }
2070
2071 struct lov_stripe_md *lov_lsm_addref(struct lov_object *lov)
2072 {
2073         struct lov_stripe_md *lsm = NULL;
2074
2075         lov_conf_freeze(lov);
2076         if (lov->lo_lsm != NULL) {
2077                 lsm = lsm_addref(lov->lo_lsm);
2078                 CDEBUG(D_INODE, "lsm %p addref %d/%d by %p.\n",
2079                         lsm, atomic_read(&lsm->lsm_refc),
2080                         lov->lo_layout_invalid, current);
2081         }
2082         lov_conf_thaw(lov);
2083         return lsm;
2084 }
2085
2086 int lov_read_and_clear_async_rc(struct cl_object *clob)
2087 {
2088         struct lu_object *luobj;
2089         int rc = 0;
2090         ENTRY;
2091
2092         luobj = lu_object_locate(&cl_object_header(clob)->coh_lu,
2093                                  &lov_device_type);
2094         if (luobj != NULL) {
2095                 struct lov_object *lov = lu2lov(luobj);
2096
2097                 lov_conf_freeze(lov);
2098                 switch (lov->lo_type) {
2099                 case LLT_COMP: {
2100                         struct lov_stripe_md *lsm;
2101                         int i;
2102
2103                         lsm = lov->lo_lsm;
2104                         LASSERT(lsm != NULL);
2105                         for (i = 0; i < lsm->lsm_entry_count; i++) {
2106                                 struct lov_stripe_md_entry *lse =
2107                                                 lsm->lsm_entries[i];
2108                                 int j;
2109
2110                                 if (!lsme_inited(lse))
2111                                         break;
2112
2113                                 for (j = 0; j < lse->lsme_stripe_count; j++) {
2114                                         struct lov_oinfo *loi =
2115                                                         lse->lsme_oinfo[j];
2116
2117                                         if (lov_oinfo_is_dummy(loi))
2118                                                 continue;
2119
2120                                         if (loi->loi_ar.ar_rc && !rc)
2121                                                 rc = loi->loi_ar.ar_rc;
2122                                         loi->loi_ar.ar_rc = 0;
2123                                 }
2124                         }
2125                 }
2126                 case LLT_RELEASED:
2127                 case LLT_EMPTY:
2128                         break;
2129                 default:
2130                         LBUG();
2131                 }
2132                 lov_conf_thaw(lov);
2133         }
2134         RETURN(rc);
2135 }
2136 EXPORT_SYMBOL(lov_read_and_clear_async_rc);
2137
2138 /** @} lov */