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, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * 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 <linux/random.h>
  41
  42 #include "lov_cl_internal.h"
  43
  44 static inline struct lov_device *lov_object_dev(struct lov_object *obj)
  45 {
  46         return lu2lov_dev(obj->lo_cl.co_lu.lo_dev);
  47 }
  48
  49 /** \addtogroup lov
  50  *  @{
  51  */
  52
  53 /*****************************************************************************
  54  *
  55  * Layout operations.
  56  *
  57  */
  58
  59 struct lov_layout_operations {
  60         int (*llo_init)(const struct lu_env *env, struct lov_device *dev,
  61                         struct lov_object *lov, struct lov_stripe_md *lsm,
  62                         const struct cl_object_conf *conf,
  63                         union lov_layout_state *state);
  64         int (*llo_delete)(const struct lu_env *env, struct lov_object *lov,
  65                            union lov_layout_state *state);
  66         void (*llo_fini)(const struct lu_env *env, struct lov_object *lov,
  67                          union lov_layout_state *state);
  68         int  (*llo_print)(const struct lu_env *env, void *cookie,
  69                           lu_printer_t p, const struct lu_object *o);
  70         int  (*llo_page_init)(const struct lu_env *env, struct cl_object *obj,
  71                               struct cl_page *page, pgoff_t index);
  72         int  (*llo_lock_init)(const struct lu_env *env,
  73                               struct cl_object *obj, struct cl_lock *lock,
  74                               const struct cl_io *io);
  75         int  (*llo_io_init)(const struct lu_env *env,
  76                             struct cl_object *obj, struct cl_io *io);
  77         int  (*llo_getattr)(const struct lu_env *env, struct cl_object *obj,
  78                             struct cl_attr *attr);
  79         int  (*llo_flush)(const struct lu_env *env, struct cl_object *obj,
  80                           struct ldlm_lock *lock);
  81 };
  82
  83 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov);
  84
  85 static void lov_lsm_put(struct lov_stripe_md *lsm)
  86 {
  87         if (lsm != NULL)
  88                 lov_free_memmd(&lsm);
  89 }
  90
  91 /*****************************************************************************
  92  *
  93  * Lov object layout operations.
  94  *
  95  */
  96
  97 static struct cl_object *lov_sub_find(const struct lu_env *env,
  98                                       struct cl_device *dev,
  99                                       const struct lu_fid *fid,
 100                                       const struct cl_object_conf *conf)
 101 {
 102         struct lu_object *o;
 103
 104         ENTRY;
 105
 106         o = lu_object_find_at(env, cl2lu_dev(dev), fid, &conf->coc_lu);
 107         LASSERT(ergo(!IS_ERR(o), o->lo_dev->ld_type == &lovsub_device_type));
 108         RETURN(lu2cl(o));
 109 }
 110
 111 static int lov_page_slice_fixup(struct lov_object *lov,
 112                                 struct cl_object *stripe)
 113 {
 114         struct cl_object_header *hdr = cl_object_header(&lov->lo_cl);
 115         struct cl_object *o;
 116
 117         if (stripe == NULL)
 118                 return hdr->coh_page_bufsize - lov->lo_cl.co_slice_off -
 119                        cfs_size_round(sizeof(struct lov_page));
 120
 121         cl_object_for_each(o, stripe)
 122                 o->co_slice_off += hdr->coh_page_bufsize;
 123
 124         return cl_object_header(stripe)->coh_page_bufsize;
 125 }
 126
 127 static int lov_init_sub(const struct lu_env *env, struct lov_object *lov,
 128                         struct cl_object *subobj, struct lov_oinfo *oinfo,
 129                         int idx)
 130 {
 131         struct cl_object_header *hdr;
 132         struct cl_object_header *subhdr;
 133         struct cl_object_header *parent;
 134         int entry = lov_comp_entry(idx);
 135         int stripe = lov_comp_stripe(idx);
 136         int result;
 137
 138         if (OBD_FAIL_CHECK(OBD_FAIL_LOV_INIT)) {
 139                 /* For sanity:test_206.
 140                  * Do not leave the object in cache to avoid accessing
 141                  * freed memory. This is because osc_object is referring to
 142                  * lov_oinfo of lsm_stripe_data which will be freed due to
 143                  * this failure. */
 144                 cl_object_kill(env, subobj);
 145                 cl_object_put(env, subobj);
 146                 return -EIO;
 147         }
 148
 149         hdr = cl_object_header(lov2cl(lov));
 150         subhdr = cl_object_header(subobj);
 151
 152         CDEBUG(D_INODE, DFID"@%p[%d:%d] -> "DFID"@%p: ostid: "DOSTID
 153                " ost idx: %d gen: %d\n",
 154                PFID(lu_object_fid(&subobj->co_lu)), subhdr, entry, stripe,
 155                PFID(lu_object_fid(lov2lu(lov))), hdr, POSTID(&oinfo->loi_oi),
 156                oinfo->loi_ost_idx, oinfo->loi_ost_gen);
 157
 158         /* reuse ->coh_attr_guard to protect coh_parent change */
 159         spin_lock(&subhdr->coh_attr_guard);
 160         parent = subhdr->coh_parent;
 161         if (parent == NULL) {
 162                 struct lovsub_object *lso = cl2lovsub(subobj);
 163
 164                 subhdr->coh_parent = hdr;
 165                 spin_unlock(&subhdr->coh_attr_guard);
 166                 subhdr->coh_nesting = hdr->coh_nesting + 1;
 167                 lu_object_ref_add(&subobj->co_lu, "lov-parent", lov);
 168                 lso->lso_super = lov;
 169                 lso->lso_index = idx;
 170                 result = 0;
 171         } else {
 172                 struct lu_object  *old_obj;
 173                 struct lov_object *old_lov;
 174                 unsigned int mask = D_INODE;
 175
 176                 spin_unlock(&subhdr->coh_attr_guard);
 177                 old_obj = lu_object_locate(&parent->coh_lu, &lov_device_type);
 178                 LASSERT(old_obj != NULL);
 179                 old_lov = cl2lov(lu2cl(old_obj));
 180                 if (old_lov->lo_layout_invalid) {
 181                         /* the object's layout has already changed but isn't
 182                          * refreshed */
 183                         lu_object_unhash(env, &subobj->co_lu);
 184                         result = -EAGAIN;
 185                 } else {
 186                         mask = D_ERROR;
 187                         result = -EIO;
 188                 }
 189
 190                 LU_OBJECT_DEBUG(mask, env, &subobj->co_lu,
 191                                 "stripe %d is already owned.", idx);
 192                 LU_OBJECT_DEBUG(mask, env, old_obj, "owned.");
 193                 LU_OBJECT_HEADER(mask, env, lov2lu(lov), "try to own.\n");
 194                 cl_object_put(env, subobj);
 195         }
 196         return result;
 197 }
 198
 199 static int lov_init_raid0(const struct lu_env *env, struct lov_device *dev,
 200                           struct lov_object *lov, unsigned int index,
 201                           const struct cl_object_conf *conf,
 202                           struct lov_layout_entry *lle)
 203 {
 204         struct lov_layout_raid0 *r0 = &lle->lle_raid0;
 205         struct lov_thread_info *lti = lov_env_info(env);
 206         struct cl_object_conf *subconf = &lti->lti_stripe_conf;
 207         struct lu_fid *ofid = &lti->lti_fid;
 208         struct cl_object *stripe;
 209         struct lov_stripe_md_entry *lse  = lov_lse(lov, index);
 210         int result;
 211         int psz, sz;
 212         int i;
 213
 214         ENTRY;
 215
 216         spin_lock_init(&r0->lo_sub_lock);
 217         r0->lo_nr = lse->lsme_stripe_count;
 218
 219         OBD_ALLOC_LARGE(r0->lo_sub, r0->lo_nr * sizeof(r0->lo_sub[0]));
 220         if (r0->lo_sub == NULL)
 221                 GOTO(out, result = -ENOMEM);
 222
 223         psz = 0;
 224         result = 0;
 225         memset(subconf, 0, sizeof(*subconf));
 226
 227         /*
 228          * Create stripe cl_objects.
 229          */
 230         for (i = 0; i < r0->lo_nr; ++i) {
 231                 struct cl_device *subdev;
 232                 struct lov_oinfo *oinfo = lse->lsme_oinfo[i];
 233                 int ost_idx = oinfo->loi_ost_idx;
 234
 235                 if (lov_oinfo_is_dummy(oinfo))
 236                         continue;
 237
 238                 result = ostid_to_fid(ofid, &oinfo->loi_oi, oinfo->loi_ost_idx);
 239                 if (result != 0)
 240                         GOTO(out, result);
 241
 242                 if (dev->ld_target[ost_idx] == NULL) {
 243                         CERROR("%s: OST %04x is not initialized\n",
 244                                lov2obd(dev->ld_lov)->obd_name, ost_idx);
 245                         GOTO(out, result = -EIO);
 246                 }
 247
 248                 subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
 249                 subconf->u.coc_oinfo = oinfo;
 250                 LASSERTF(subdev != NULL, "not init ost %d\n", ost_idx);
 251                 /* In the function below, .hs_keycmp resolves to
 252                  * lu_obj_hop_keycmp() */
 253                 /* coverity[overrun-buffer-val] */
 254                 stripe = lov_sub_find(env, subdev, ofid, subconf);
 255                 if (IS_ERR(stripe))
 256                         GOTO(out, result = PTR_ERR(stripe));
 257
 258                 result = lov_init_sub(env, lov, stripe, oinfo,
 259                                       lov_comp_index(index, i));
 260                 if (result == -EAGAIN) { /* try again */
 261                         --i;
 262                         result = 0;
 263                         continue;
 264                 }
 265
 266                 if (result == 0) {
 267                         r0->lo_sub[i] = cl2lovsub(stripe);
 268
 269                         sz = lov_page_slice_fixup(lov, stripe);
 270                         LASSERT(ergo(psz > 0, psz == sz));
 271                         psz = sz;
 272                 }
 273         }
 274         if (result == 0)
 275                 result = psz;
 276 out:
 277         RETURN(result);
 278 }
 279
 280 static void lov_subobject_kill(const struct lu_env *env, struct lov_object *lov,
 281                                struct lov_layout_raid0 *r0,
 282                                struct lovsub_object *los, int idx)
 283 {
 284         struct cl_object        *sub;
 285         struct lu_site          *site;
 286         wait_queue_head_t *wq;
 287         wait_queue_entry_t *waiter;
 288
 289         LASSERT(r0->lo_sub[idx] == los);
 290
 291         sub = lovsub2cl(los);
 292         site = sub->co_lu.lo_dev->ld_site;
 293         wq = lu_site_wq_from_fid(site, &sub->co_lu.lo_header->loh_fid);
 294
 295         cl_object_kill(env, sub);
 296         /* release a reference to the sub-object and ... */
 297         lu_object_ref_del(&sub->co_lu, "lov-parent", lov);
 298         cl_object_put(env, sub);
 299
 300         /* ... wait until it is actually destroyed---sub-object clears its
 301          * ->lo_sub[] slot in lovsub_object_free() */
 302         if (r0->lo_sub[idx] == los) {
 303                 waiter = &lov_env_info(env)->lti_waiter;
 304                 init_waitqueue_entry(waiter, current);
 305                 add_wait_queue(wq, waiter);
 306                 set_current_state(TASK_UNINTERRUPTIBLE);
 307                 while (1) {
 308                         /* this wait-queue is signaled at the end of
 309                          * lu_object_free(). */
 310                         set_current_state(TASK_UNINTERRUPTIBLE);
 311                         spin_lock(&r0->lo_sub_lock);
 312                         if (r0->lo_sub[idx] == los) {
 313                                 spin_unlock(&r0->lo_sub_lock);
 314                                 schedule();
 315                         } else {
 316                                 spin_unlock(&r0->lo_sub_lock);
 317                                 set_current_state(TASK_RUNNING);
 318                                 break;
 319                         }
 320                 }
 321                 remove_wait_queue(wq, waiter);
 322         }
 323         LASSERT(r0->lo_sub[idx] == NULL);
 324 }
 325
 326 static void lov_delete_raid0(const struct lu_env *env, struct lov_object *lov,
 327                              struct lov_layout_entry *lle)
 328 {
 329         struct lov_layout_raid0 *r0 = &lle->lle_raid0;
 330
 331         ENTRY;
 332
 333         if (r0->lo_sub != NULL) {
 334                 int i;
 335
 336                 for (i = 0; i < r0->lo_nr; ++i) {
 337                         struct lovsub_object *los = r0->lo_sub[i];
 338
 339                         if (los != NULL) {
 340                                 cl_object_prune(env, &los->lso_cl);
 341                                 /*
 342                                  * If top-level object is to be evicted from
 343                                  * the cache, so are its sub-objects.
 344                                  */
 345                                 lov_subobject_kill(env, lov, r0, los, i);
 346                         }
 347                 }
 348         }
 349
 350         EXIT;
 351 }
 352
 353 static void lov_fini_raid0(const struct lu_env *env,
 354                            struct lov_layout_entry *lle)
 355 {
 356         struct lov_layout_raid0 *r0 = &lle->lle_raid0;
 357
 358         if (r0->lo_sub != NULL) {
 359                 OBD_FREE_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 360                 r0->lo_sub = NULL;
 361         }
 362 }
 363
 364 static int lov_print_raid0(const struct lu_env *env, void *cookie,
 365                            lu_printer_t p, const struct lov_layout_entry *lle)
 366 {
 367         const struct lov_layout_raid0 *r0 = &lle->lle_raid0;
 368         int i;
 369
 370         for (i = 0; i < r0->lo_nr; ++i) {
 371                 struct lu_object *sub;
 372
 373                 if (r0->lo_sub[i] != NULL) {
 374                         sub = lovsub2lu(r0->lo_sub[i]);
 375                         lu_object_print(env, cookie, p, sub);
 376                 } else {
 377                         (*p)(env, cookie, "sub %d absent\n", i);
 378                 }
 379         }
 380         return 0;
 381 }
 382
 383 static int lov_attr_get_raid0(const struct lu_env *env, struct lov_object *lov,
 384                               unsigned int index, struct lov_layout_entry *lle,
 385                               struct cl_attr **lov_attr)
 386 {
 387         struct lov_layout_raid0 *r0 = &lle->lle_raid0;
 388         struct lov_stripe_md *lsm = lov->lo_lsm;
 389         struct ost_lvb *lvb = &lov_env_info(env)->lti_lvb;
 390         struct cl_attr *attr = &r0->lo_attr;
 391         __u64 kms = 0;
 392         int result = 0;
 393
 394         if (r0->lo_attr_valid) {
 395                 *lov_attr = attr;
 396                 return 0;
 397         }
 398
 399         memset(lvb, 0, sizeof(*lvb));
 400
 401         /* XXX: timestamps can be negative by sanity:test_39m,
 402          * how can it be? */
 403         lvb->lvb_atime = LLONG_MIN;
 404         lvb->lvb_ctime = LLONG_MIN;
 405         lvb->lvb_mtime = LLONG_MIN;
 406
 407         /*
 408          * XXX that should be replaced with a loop over sub-objects,
 409          * doing cl_object_attr_get() on them. But for now, let's
 410          * reuse old lov code.
 411          */
 412
 413         /*
 414          * XXX take lsm spin-lock to keep lov_merge_lvb_kms()
 415          * happy. It's not needed, because new code uses
 416          * ->coh_attr_guard spin-lock to protect consistency of
 417          * sub-object attributes.
 418          */
 419         lov_stripe_lock(lsm);
 420         result = lov_merge_lvb_kms(lsm, index, lvb, &kms);
 421         lov_stripe_unlock(lsm);
 422         if (result == 0) {
 423                 cl_lvb2attr(attr, lvb);
 424                 attr->cat_kms = kms;
 425                 r0->lo_attr_valid = 1;
 426                 *lov_attr = attr;
 427         }
 428
 429         return result;
 430 }
 431
 432 static struct lov_comp_layout_entry_ops raid0_ops = {
 433         .lco_init      = lov_init_raid0,
 434         .lco_fini      = lov_fini_raid0,
 435         .lco_getattr   = lov_attr_get_raid0,
 436 };
 437
 438 static int lov_attr_get_dom(const struct lu_env *env, struct lov_object *lov,
 439                             unsigned int index, struct lov_layout_entry *lle,
 440                             struct cl_attr **lov_attr)
 441 {
 442         struct lov_layout_dom *dom = &lle->lle_dom;
 443         struct lov_oinfo *loi = dom->lo_loi;
 444         struct cl_attr *attr = &dom->lo_dom_r0.lo_attr;
 445
 446         if (dom->lo_dom_r0.lo_attr_valid) {
 447                 *lov_attr = attr;
 448                 return 0;
 449         }
 450
 451         if (OST_LVB_IS_ERR(loi->loi_lvb.lvb_blocks))
 452                 return OST_LVB_GET_ERR(loi->loi_lvb.lvb_blocks);
 453
 454         cl_lvb2attr(attr, &loi->loi_lvb);
 455
 456         /* DoM component size can be bigger than stripe size after
 457          * client's setattr RPC, so do not count anything beyond
 458          * component end. Alternatively, check that limit on server
 459          * and do not allow size overflow there. */
 460         if (attr->cat_size > lle->lle_extent->e_end)
 461                 attr->cat_size = lle->lle_extent->e_end;
 462
 463         attr->cat_kms = attr->cat_size;
 464
 465         dom->lo_dom_r0.lo_attr_valid = 1;
 466         *lov_attr = attr;
 467
 468         return 0;
 469 }
 470
 471 /**
 472  * Lookup FLD to get MDS index of the given DOM object FID.
 473  *
 474  * \param[in]  ld       LOV device
 475  * \param[in]  fid      FID to lookup
 476  * \param[out] nr       index in MDC array to return back
 477  *
 478  * \retval              0 and \a mds filled with MDS index if successful
 479  * \retval              negative value on error
 480  */
 481 static int lov_fld_lookup(struct lov_device *ld, const struct lu_fid *fid,
 482                           __u32 *nr)
 483 {
 484         __u32 mds_idx;
 485         int i, rc;
 486
 487         ENTRY;
 488
 489         rc = fld_client_lookup(&ld->ld_lmv->u.lmv.lmv_fld, fid_seq(fid),
 490                                &mds_idx, LU_SEQ_RANGE_MDT, NULL);
 491         if (rc) {
 492                 CERROR("%s: error while looking for mds number. Seq %#llx"
 493                        ", err = %d\n", lu_dev_name(cl2lu_dev(&ld->ld_cl)),
 494                        fid_seq(fid), rc);
 495                 RETURN(rc);
 496         }
 497
 498         CDEBUG(D_INODE, "FLD lookup got mds #%x for fid="DFID"\n",
 499                mds_idx, PFID(fid));
 500
 501         /* find proper MDC device in the array */
 502         for (i = 0; i < ld->ld_md_tgts_nr; i++) {
 503                 if (ld->ld_md_tgts[i].ldm_mdc != NULL &&
 504                     ld->ld_md_tgts[i].ldm_idx == mds_idx)
 505                         break;
 506         }
 507
 508         if (i == ld->ld_md_tgts_nr) {
 509                 CERROR("%s: cannot find corresponding MDC device for mds #%x "
 510                        "for fid="DFID"\n", lu_dev_name(cl2lu_dev(&ld->ld_cl)),
 511                        mds_idx, PFID(fid));
 512                 rc = -EINVAL;
 513         } else {
 514                 *nr = i;
 515         }
 516         RETURN(rc);
 517 }
 518
 519 /**
 520  * Implementation of lov_comp_layout_entry_ops::lco_init for DOM object.
 521  *
 522  * Init the DOM object for the first time. It prepares also RAID0 entry
 523  * for it to use in common methods with ordinary RAID0 layout entries.
 524  *
 525  * \param[in] env       execution environment
 526  * \param[in] dev       LOV device
 527  * \param[in] lov       LOV object
 528  * \param[in] index     Composite layout entry index in LSM
 529  * \param[in] lle       Composite LOV layout entry
 530  */
 531 static int lov_init_dom(const struct lu_env *env, struct lov_device *dev,
 532                         struct lov_object *lov, unsigned int index,
 533                         const struct cl_object_conf *conf,
 534                         struct lov_layout_entry *lle)
 535 {
 536         struct lov_thread_info *lti = lov_env_info(env);
 537         struct lov_stripe_md_entry *lsme = lov_lse(lov, index);
 538         struct cl_object *clo;
 539         struct lu_object *o = lov2lu(lov);
 540         const struct lu_fid *fid = lu_object_fid(o);
 541         struct cl_device *mdcdev;
 542         struct lov_oinfo *loi = NULL;
 543         struct cl_object_conf *sconf = &lti->lti_stripe_conf;
 544         int rc;
 545         __u32 idx = 0;
 546
 547         ENTRY;
 548
 549         /* DOM entry may be not zero index due to FLR but must start from 0 */
 550         if (unlikely(lle->lle_extent->e_start != 0)) {
 551                 CERROR("%s: DOM entry must be the first stripe in a mirror\n",
 552                        lov2obd(dev->ld_lov)->obd_name);
 553                 dump_lsm(D_ERROR, lov->lo_lsm);
 554                 RETURN(-EINVAL);
 555         }
 556
 557         /* find proper MDS device */
 558         rc = lov_fld_lookup(dev, fid, &idx);
 559         if (rc)
 560                 RETURN(rc);
 561
 562         LASSERTF(dev->ld_md_tgts[idx].ldm_mdc != NULL,
 563                  "LOV md target[%u] is NULL\n", idx);
 564
 565         /* check lsm is DOM, more checks are needed */
 566         LASSERT(lsme->lsme_stripe_count == 0);
 567
 568         /*
 569          * Create lower cl_objects.
 570          */
 571         mdcdev = dev->ld_md_tgts[idx].ldm_mdc;
 572
 573         LASSERTF(mdcdev != NULL, "non-initialized mdc subdev\n");
 574
 575         /* DoM object has no oinfo in LSM entry, create it exclusively */
 576         OBD_SLAB_ALLOC_PTR_GFP(loi, lov_oinfo_slab, GFP_NOFS);
 577         if (loi == NULL)
 578                 RETURN(-ENOMEM);
 579
 580         fid_to_ostid(lu_object_fid(lov2lu(lov)), &loi->loi_oi);
 581
 582         sconf->u.coc_oinfo = loi;
 583 again:
 584         clo = lov_sub_find(env, mdcdev, fid, sconf);
 585         if (IS_ERR(clo))
 586                 GOTO(out, rc = PTR_ERR(clo));
 587
 588         rc = lov_init_sub(env, lov, clo, loi, lov_comp_index(index, 0));
 589         if (rc == -EAGAIN) /* try again */
 590                 goto again;
 591         else if (rc != 0)
 592                 GOTO(out, rc);
 593
 594         lle->lle_dom.lo_dom = cl2lovsub(clo);
 595         spin_lock_init(&lle->lle_dom.lo_dom_r0.lo_sub_lock);
 596         lle->lle_dom.lo_dom_r0.lo_nr = 1;
 597         lle->lle_dom.lo_dom_r0.lo_sub = &lle->lle_dom.lo_dom;
 598         lle->lle_dom.lo_loi = loi;
 599
 600         rc = lov_page_slice_fixup(lov, clo);
 601         RETURN(rc);
 602
 603 out:
 604         if (loi != NULL)
 605                 OBD_SLAB_FREE_PTR(loi, lov_oinfo_slab);
 606         return rc;
 607 }
 608
 609 /**
 610  * Implementation of lov_layout_operations::llo_fini for DOM object.
 611  *
 612  * Finish the DOM object and free related memory.
 613  *
 614  * \param[in] env       execution environment
 615  * \param[in] lov       LOV object
 616  * \param[in] state     LOV layout state
 617  */
 618 static void lov_fini_dom(const struct lu_env *env,
 619                          struct lov_layout_entry *lle)
 620 {
 621         if (lle->lle_dom.lo_dom != NULL)
 622                 lle->lle_dom.lo_dom = NULL;
 623         if (lle->lle_dom.lo_loi != NULL)
 624                 OBD_SLAB_FREE_PTR(lle->lle_dom.lo_loi, lov_oinfo_slab);
 625 }
 626
 627 static struct lov_comp_layout_entry_ops dom_ops = {
 628         .lco_init = lov_init_dom,
 629         .lco_fini = lov_fini_dom,
 630         .lco_getattr = lov_attr_get_dom,
 631 };
 632
 633 static int lov_init_composite(const struct lu_env *env, struct lov_device *dev,
 634                               struct lov_object *lov, struct lov_stripe_md *lsm,
 635                               const struct cl_object_conf *conf,
 636                               union lov_layout_state *state)
 637 {
 638         struct lov_layout_composite *comp = &state->composite;
 639         struct lov_layout_entry *lle;
 640         struct lov_mirror_entry *lre;
 641         unsigned int entry_count;
 642         unsigned int psz = 0;
 643         unsigned int mirror_count;
 644         int flr_state = lsm->lsm_flags & LCM_FL_FLR_MASK;
 645         int result = 0;
 646         unsigned int seq;
 647         int i, j;
 648         bool dom_size = 0;
 649
 650         ENTRY;
 651
 652         LASSERT(lsm->lsm_entry_count > 0);
 653         LASSERT(lov->lo_lsm == NULL);
 654         lov->lo_lsm = lsm_addref(lsm);
 655         lov->lo_layout_invalid = true;
 656
 657         dump_lsm(D_INODE, lsm);
 658
 659         entry_count = lsm->lsm_entry_count;
 660
 661         spin_lock_init(&comp->lo_write_lock);
 662         comp->lo_flags = lsm->lsm_flags;
 663         comp->lo_mirror_count = lsm->lsm_mirror_count + 1;
 664         comp->lo_entry_count = lsm->lsm_entry_count;
 665         comp->lo_preferred_mirror = -1;
 666
 667         if (equi(flr_state == LCM_FL_NONE, comp->lo_mirror_count > 1))
 668                 RETURN(-EINVAL);
 669
 670         OBD_ALLOC(comp->lo_mirrors,
 671                   comp->lo_mirror_count * sizeof(*comp->lo_mirrors));
 672         if (comp->lo_mirrors == NULL)
 673                 RETURN(-ENOMEM);
 674
 675         OBD_ALLOC(comp->lo_entries, entry_count * sizeof(*comp->lo_entries));
 676         if (comp->lo_entries == NULL)
 677                 RETURN(-ENOMEM);
 678
 679         /* Initiate all entry types and extents data at first */
 680         for (i = 0, j = 0, mirror_count = 1; i < entry_count; i++) {
 681                 int mirror_id = 0;
 682
 683                 lle = &comp->lo_entries[i];
 684
 685                 lle->lle_lsme = lsm->lsm_entries[i];
 686                 lle->lle_type = lov_entry_type(lle->lle_lsme);
 687                 switch (lle->lle_type) {
 688                 case LOV_PATTERN_RAID0:
 689                         lle->lle_comp_ops = &raid0_ops;
 690                         break;
 691                 case LOV_PATTERN_MDT:
 692                         /* Allowed to have several DOM stripes in different
 693                          * mirrors with the same DoM size.
 694                          */
 695                         if (!dom_size) {
 696                                 dom_size = lle->lle_lsme->lsme_extent.e_end;
 697                         } else if (dom_size !=
 698                                    lle->lle_lsme->lsme_extent.e_end) {
 699                                 CERROR("%s: DOM entries with different sizes\n",
 700                                        lov2obd(dev->ld_lov)->obd_name);
 701                                 dump_lsm(D_ERROR, lsm);
 702                                 RETURN(-EINVAL);
 703                         }
 704                         lle->lle_comp_ops = &dom_ops;
 705                         break;
 706                 default:
 707                         CERROR("%s: unknown composite layout entry type %i\n",
 708                                lov2obd(dev->ld_lov)->obd_name,
 709                                lsm->lsm_entries[i]->lsme_pattern);
 710                         dump_lsm(D_ERROR, lsm);
 711                         RETURN(-EIO);
 712                 }
 713
 714                 lle->lle_extent = &lle->lle_lsme->lsme_extent;
 715                 lle->lle_valid = !(lle->lle_lsme->lsme_flags & LCME_FL_STALE);
 716
 717                 if (flr_state != LCM_FL_NONE)
 718                         mirror_id = mirror_id_of(lle->lle_lsme->lsme_id);
 719
 720                 lre = &comp->lo_mirrors[j];
 721                 if (i > 0) {
 722                         if (mirror_id == lre->lre_mirror_id) {
 723                                 lre->lre_valid |= lle->lle_valid;
 724                                 lre->lre_stale |= !lle->lle_valid;
 725                                 lre->lre_end = i;
 726                                 continue;
 727                         }
 728
 729                         /* new mirror detected, assume that the mirrors
 730                          * are shorted in layout */
 731                         ++mirror_count;
 732                         ++j;
 733                         if (j >= comp->lo_mirror_count)
 734                                 break;
 735
 736                         lre = &comp->lo_mirrors[j];
 737                 }
 738
 739                 /* entries must be sorted by mirrors */
 740                 lre->lre_mirror_id = mirror_id;
 741                 lre->lre_start = lre->lre_end = i;
 742                 lre->lre_preferred = !!(lle->lle_lsme->lsme_flags &
 743                                         LCME_FL_PREF_RD);
 744                 lre->lre_valid = lle->lle_valid;
 745                 lre->lre_stale = !lle->lle_valid;
 746         }
 747
 748         /* sanity check for FLR */
 749         if (mirror_count != comp->lo_mirror_count) {
 750                 CDEBUG(D_INODE, DFID
 751                        " doesn't have the # of mirrors it claims, %u/%u\n",
 752                        PFID(lu_object_fid(lov2lu(lov))), mirror_count,
 753                        comp->lo_mirror_count + 1);
 754
 755                 GOTO(out, result = -EINVAL);
 756         }
 757
 758         lov_foreach_layout_entry(lov, lle) {
 759                 int index = lov_layout_entry_index(lov, lle);
 760
 761                 /**
 762                  * If the component has not been init-ed on MDS side, for
 763                  * PFL layout, we'd know that the components beyond this one
 764                  * will be dynamically init-ed later on file write/trunc ops.
 765                  */
 766                 if (!lsme_inited(lle->lle_lsme))
 767                         continue;
 768
 769                 result = lle->lle_comp_ops->lco_init(env, dev, lov, index,
 770                                                      conf, lle);
 771                 if (result < 0)
 772                         break;
 773
 774                 LASSERT(ergo(psz > 0, psz == result));
 775                 psz = result;
 776         }
 777
 778         if (psz > 0)
 779                 cl_object_header(&lov->lo_cl)->coh_page_bufsize += psz;
 780
 781         /* decide the preferred mirror. It uses the hash value of lov_object
 782          * so that different clients would use different mirrors for read. */
 783         mirror_count = 0;
 784         seq = hash_long((unsigned long)lov, 8);
 785         for (i = 0; i < comp->lo_mirror_count; i++) {
 786                 unsigned int idx = (i + seq) % comp->lo_mirror_count;
 787
 788                 lre = lov_mirror_entry(lov, idx);
 789                 if (lre->lre_stale)
 790                         continue;
 791
 792                 mirror_count++; /* valid mirror */
 793
 794                 if (lre->lre_preferred || comp->lo_preferred_mirror < 0)
 795                         comp->lo_preferred_mirror = idx;
 796         }
 797         if (!mirror_count) {
 798                 CDEBUG(D_INODE, DFID
 799                        " doesn't have any valid mirrors\n",
 800                        PFID(lu_object_fid(lov2lu(lov))));
 801
 802                 comp->lo_preferred_mirror = 0;
 803         }
 804
 805         LASSERT(comp->lo_preferred_mirror >= 0);
 806
 807         EXIT;
 808 out:
 809         return result > 0 ? 0 : result;
 810 }
 811
 812 static int lov_init_empty(const struct lu_env *env, struct lov_device *dev,
 813                           struct lov_object *lov, struct lov_stripe_md *lsm,
 814                           const struct cl_object_conf *conf,
 815                           union lov_layout_state *state)
 816 {
 817         return 0;
 818 }
 819
 820 static int lov_init_released(const struct lu_env *env,
 821                              struct lov_device *dev, struct lov_object *lov,
 822                              struct lov_stripe_md *lsm,
 823                              const struct cl_object_conf *conf,
 824                              union lov_layout_state *state)
 825 {
 826         LASSERT(lsm != NULL);
 827         LASSERT(lsm->lsm_is_released);
 828         LASSERT(lov->lo_lsm == NULL);
 829
 830         lov->lo_lsm = lsm_addref(lsm);
 831         return 0;
 832 }
 833
 834 static int lov_init_foreign(const struct lu_env *env,
 835                             struct lov_device *dev, struct lov_object *lov,
 836                             struct lov_stripe_md *lsm,
 837                             const struct cl_object_conf *conf,
 838                             union lov_layout_state *state)
 839 {
 840         LASSERT(lsm != NULL);
 841         LASSERT(lov->lo_type == LLT_FOREIGN);
 842         LASSERT(lov->lo_lsm == NULL);
 843
 844         lov->lo_lsm = lsm_addref(lsm);
 845         return 0;
 846 }
 847
 848 static int lov_delete_empty(const struct lu_env *env, struct lov_object *lov,
 849                             union lov_layout_state *state)
 850 {
 851         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED ||
 852                 lov->lo_type == LLT_FOREIGN);
 853
 854         lov_layout_wait(env, lov);
 855         return 0;
 856 }
 857
 858 static int lov_delete_composite(const struct lu_env *env,
 859                                 struct lov_object *lov,
 860                                 union lov_layout_state *state)
 861 {
 862         struct lov_layout_entry *entry;
 863         struct lov_layout_composite *comp = &state->composite;
 864
 865         ENTRY;
 866
 867         dump_lsm(D_INODE, lov->lo_lsm);
 868
 869         lov_layout_wait(env, lov);
 870         if (comp->lo_entries)
 871                 lov_foreach_layout_entry(lov, entry)
 872                         lov_delete_raid0(env, lov, entry);
 873
 874         RETURN(0);
 875 }
 876
 877 static void lov_fini_empty(const struct lu_env *env, struct lov_object *lov,
 878                            union lov_layout_state *state)
 879 {
 880         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 881 }
 882
 883 static void lov_fini_composite(const struct lu_env *env,
 884                                struct lov_object *lov,
 885                                union lov_layout_state *state)
 886 {
 887         struct lov_layout_composite *comp = &state->composite;
 888         ENTRY;
 889
 890         if (comp->lo_entries != NULL) {
 891                 struct lov_layout_entry *entry;
 892
 893                 lov_foreach_layout_entry(lov, entry)
 894                         if (entry->lle_comp_ops)
 895                                 entry->lle_comp_ops->lco_fini(env, entry);
 896
 897                 OBD_FREE(comp->lo_entries,
 898                          comp->lo_entry_count * sizeof(*comp->lo_entries));
 899                 comp->lo_entries = NULL;
 900         }
 901
 902         if (comp->lo_mirrors != NULL) {
 903                 OBD_FREE(comp->lo_mirrors,
 904                          comp->lo_mirror_count * sizeof(*comp->lo_mirrors));
 905                 comp->lo_mirrors = NULL;
 906         }
 907
 908         memset(comp, 0, sizeof(*comp));
 909
 910         dump_lsm(D_INODE, lov->lo_lsm);
 911         lov_free_memmd(&lov->lo_lsm);
 912
 913         EXIT;
 914 }
 915
 916 static void lov_fini_released(const struct lu_env *env, struct lov_object *lov,
 917                                 union lov_layout_state *state)
 918 {
 919         ENTRY;
 920         dump_lsm(D_INODE, lov->lo_lsm);
 921         lov_free_memmd(&lov->lo_lsm);
 922         EXIT;
 923 }
 924
 925 static int lov_print_empty(const struct lu_env *env, void *cookie,
 926                            lu_printer_t p, const struct lu_object *o)
 927 {
 928         (*p)(env, cookie, "empty %d\n", lu2lov(o)->lo_layout_invalid);
 929         return 0;
 930 }
 931
 932 static int lov_print_composite(const struct lu_env *env, void *cookie,
 933                                lu_printer_t p, const struct lu_object *o)
 934 {
 935         struct lov_object *lov = lu2lov(o);
 936         struct lov_stripe_md *lsm = lov->lo_lsm;
 937         int i;
 938
 939         (*p)(env, cookie, "entries: %d, %s, lsm{%p 0x%08X %d %u}:\n",
 940              lsm->lsm_entry_count,
 941              lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 942              lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 943              lsm->lsm_layout_gen);
 944
 945         for (i = 0; i < lsm->lsm_entry_count; i++) {
 946                 struct lov_stripe_md_entry *lse = lsm->lsm_entries[i];
 947                 struct lov_layout_entry *lle = lov_entry(lov, i);
 948
 949                 (*p)(env, cookie,
 950                      DEXT ": { 0x%08X, %u, %#x, %u, %#x, %u, %u }\n",
 951                      PEXT(&lse->lsme_extent), lse->lsme_magic,
 952                      lse->lsme_id, lse->lsme_pattern, lse->lsme_layout_gen,
 953                      lse->lsme_flags, lse->lsme_stripe_count,
 954                      lse->lsme_stripe_size);
 955                 lov_print_raid0(env, cookie, p, lle);
 956         }
 957
 958         return 0;
 959 }
 960
 961 static int lov_print_released(const struct lu_env *env, void *cookie,
 962                                 lu_printer_t p, const struct lu_object *o)
 963 {
 964         struct lov_object       *lov = lu2lov(o);
 965         struct lov_stripe_md    *lsm = lov->lo_lsm;
 966
 967         (*p)(env, cookie,
 968                 "released: %s, lsm{%p 0x%08X %d %u}:\n",
 969                 lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 970                 lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 971                 lsm->lsm_layout_gen);
 972         return 0;
 973 }
 974
 975 static int lov_print_foreign(const struct lu_env *env, void *cookie,
 976                                 lu_printer_t p, const struct lu_object *o)
 977 {
 978         struct lov_object       *lov = lu2lov(o);
 979         struct lov_stripe_md    *lsm = lov->lo_lsm;
 980
 981         (*p)(env, cookie,
 982                 "foreign: %s, lsm{%p 0x%08X %d %u}:\n",
 983                 lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 984                 lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 985                 lsm->lsm_layout_gen);
 986         (*p)(env, cookie,
 987                 "raw_ea_content '%.*s'\n",
 988                 (int)lsm->lsm_foreign_size, (char *)lsm_foreign(lsm));
 989         return 0;
 990 }
 991
 992 /**
 993  * Implements cl_object_operations::coo_attr_get() method for an object
 994  * without stripes (LLT_EMPTY layout type).
 995  *
 996  * The only attributes this layer is authoritative in this case is
 997  * cl_attr::cat_blocks---it's 0.
 998  */
 999 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
1000                               struct cl_attr *attr)
1001 {
1002         attr->cat_blocks = 0;
1003         return 0;
1004 }
1005
1006 static int lov_attr_get_composite(const struct lu_env *env,
1007                                   struct cl_object *obj,
1008                                   struct cl_attr *attr)
1009 {
1010         struct lov_object       *lov = cl2lov(obj);
1011         struct lov_layout_entry *entry;
1012         int                      result = 0;
1013
1014         ENTRY;
1015
1016         attr->cat_size = 0;
1017         attr->cat_blocks = 0;
1018         lov_foreach_layout_entry(lov, entry) {
1019                 struct cl_attr *lov_attr = NULL;
1020                 int index = lov_layout_entry_index(lov, entry);
1021
1022                 if (!entry->lle_valid)
1023                         continue;
1024
1025                 /* PFL: This component has not been init-ed. */
1026                 if (!lsm_entry_inited(lov->lo_lsm, index))
1027                         continue;
1028
1029                 result = entry->lle_comp_ops->lco_getattr(env, lov, index,
1030                                                           entry, &lov_attr);
1031                 if (result < 0)
1032                         RETURN(result);
1033
1034                 if (lov_attr == NULL)
1035                         continue;
1036
1037                 CDEBUG(D_INODE, "COMP ID #%i: s=%llu m=%llu a=%llu c=%llu "
1038                        "b=%llu\n", index - 1, lov_attr->cat_size,
1039                        lov_attr->cat_mtime, lov_attr->cat_atime,
1040                        lov_attr->cat_ctime, lov_attr->cat_blocks);
1041
1042                 /* merge results */
1043                 attr->cat_blocks += lov_attr->cat_blocks;
1044                 if (attr->cat_size < lov_attr->cat_size)
1045                         attr->cat_size = lov_attr->cat_size;
1046                 if (attr->cat_kms < lov_attr->cat_kms)
1047                         attr->cat_kms = lov_attr->cat_kms;
1048                 if (attr->cat_atime < lov_attr->cat_atime)
1049                         attr->cat_atime = lov_attr->cat_atime;
1050                 if (attr->cat_ctime < lov_attr->cat_ctime)
1051                         attr->cat_ctime = lov_attr->cat_ctime;
1052                 if (attr->cat_mtime < lov_attr->cat_mtime)
1053                         attr->cat_mtime = lov_attr->cat_mtime;
1054         }
1055
1056         RETURN(0);
1057 }
1058
1059 static int lov_flush_composite(const struct lu_env *env,
1060                                struct cl_object *obj,
1061                                struct ldlm_lock *lock)
1062 {
1063         struct lov_object *lov = cl2lov(obj);
1064         struct lovsub_object *lovsub;
1065
1066         ENTRY;
1067
1068         if (!lsme_is_dom(lov->lo_lsm->lsm_entries[0]))
1069                 RETURN(-EINVAL);
1070
1071         lovsub = lov->u.composite.lo_entries[0].lle_dom.lo_dom;
1072         RETURN(cl_object_flush(env, lovsub2cl(lovsub), lock));
1073 }
1074
1075 const static struct lov_layout_operations lov_dispatch[] = {
1076         [LLT_EMPTY] = {
1077                 .llo_init      = lov_init_empty,
1078                 .llo_delete    = lov_delete_empty,
1079                 .llo_fini      = lov_fini_empty,
1080                 .llo_print     = lov_print_empty,
1081                 .llo_page_init = lov_page_init_empty,
1082                 .llo_lock_init = lov_lock_init_empty,
1083                 .llo_io_init   = lov_io_init_empty,
1084                 .llo_getattr   = lov_attr_get_empty,
1085         },
1086         [LLT_RELEASED] = {
1087                 .llo_init      = lov_init_released,
1088                 .llo_delete    = lov_delete_empty,
1089                 .llo_fini      = lov_fini_released,
1090                 .llo_print     = lov_print_released,
1091                 .llo_page_init = lov_page_init_empty,
1092                 .llo_lock_init = lov_lock_init_empty,
1093                 .llo_io_init   = lov_io_init_released,
1094                 .llo_getattr   = lov_attr_get_empty,
1095         },
1096         [LLT_COMP] = {
1097                 .llo_init      = lov_init_composite,
1098                 .llo_delete    = lov_delete_composite,
1099                 .llo_fini      = lov_fini_composite,
1100                 .llo_print     = lov_print_composite,
1101                 .llo_page_init = lov_page_init_composite,
1102                 .llo_lock_init = lov_lock_init_composite,
1103                 .llo_io_init   = lov_io_init_composite,
1104                 .llo_getattr   = lov_attr_get_composite,
1105                 .llo_flush     = lov_flush_composite,
1106         },
1107         [LLT_FOREIGN] = {
1108                 .llo_init      = lov_init_foreign,
1109                 .llo_delete    = lov_delete_empty,
1110                 .llo_fini      = lov_fini_released,
1111                 .llo_print     = lov_print_foreign,
1112                 .llo_page_init = lov_page_init_foreign,
1113                 .llo_lock_init = lov_lock_init_empty,
1114                 .llo_io_init   = lov_io_init_empty,
1115                 .llo_getattr   = lov_attr_get_empty,
1116         },
1117 };
1118
1119 /**
1120  * Performs a double-dispatch based on the layout type of an object.
1121  */
1122 #define LOV_2DISPATCH_NOLOCK(obj, op, ...)              \
1123 ({                                                      \
1124         struct lov_object *__obj = (obj);               \
1125         enum lov_layout_type __llt;                     \
1126                                                         \
1127         __llt = __obj->lo_type;                         \
1128         LASSERT(__llt < ARRAY_SIZE(lov_dispatch));      \
1129         lov_dispatch[__llt].op(__VA_ARGS__);            \
1130 })
1131
1132 /**
1133  * Return lov_layout_type associated with a given lsm
1134  */
1135 static enum lov_layout_type lov_type(struct lov_stripe_md *lsm)
1136 {
1137         if (lsm == NULL)
1138                 return LLT_EMPTY;
1139
1140         if (lsm->lsm_is_released)
1141                 return LLT_RELEASED;
1142
1143         if (lsm->lsm_magic == LOV_MAGIC_V1 ||
1144             lsm->lsm_magic == LOV_MAGIC_V3 ||
1145             lsm->lsm_magic == LOV_MAGIC_COMP_V1)
1146                 return LLT_COMP;
1147
1148         if (lsm->lsm_magic == LOV_MAGIC_FOREIGN)
1149                 return LLT_FOREIGN;
1150
1151         return LLT_EMPTY;
1152 }
1153
1154 static inline void lov_conf_freeze(struct lov_object *lov)
1155 {
1156         CDEBUG(D_INODE, "To take share lov(%p) owner %p/%p\n",
1157                 lov, lov->lo_owner, current);
1158         if (lov->lo_owner != current)
1159                 down_read(&lov->lo_type_guard);
1160 }
1161
1162 static inline void lov_conf_thaw(struct lov_object *lov)
1163 {
1164         CDEBUG(D_INODE, "To release share lov(%p) owner %p/%p\n",
1165                 lov, lov->lo_owner, current);
1166         if (lov->lo_owner != current)
1167                 up_read(&lov->lo_type_guard);
1168 }
1169
1170 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...)                       \
1171 ({                                                                      \
1172         struct lov_object                      *__obj = (obj);          \
1173         int                                     __lock = !!(lock);      \
1174         typeof(lov_dispatch[0].op(__VA_ARGS__)) __result;               \
1175                                                                         \
1176         if (__lock)                                                     \
1177                 lov_conf_freeze(__obj);                                 \
1178         __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__);          \
1179         if (__lock)                                                     \
1180                 lov_conf_thaw(__obj);                                   \
1181         __result;                                                       \
1182 })
1183
1184 /**
1185  * Performs a locked double-dispatch based on the layout type of an object.
1186  */
1187 #define LOV_2DISPATCH(obj, op, ...)                     \
1188         LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
1189
1190 #define LOV_2DISPATCH_VOID(obj, op, ...)                                \
1191 do {                                                                    \
1192         struct lov_object                      *__obj = (obj);          \
1193         enum lov_layout_type                    __llt;                  \
1194                                                                         \
1195         lov_conf_freeze(__obj);                                         \
1196         __llt = __obj->lo_type;                                         \
1197         LASSERT(__llt < ARRAY_SIZE(lov_dispatch));                      \
1198         lov_dispatch[__llt].op(__VA_ARGS__);                            \
1199         lov_conf_thaw(__obj);                                           \
1200 } while (0)
1201
1202 static void lov_conf_lock(struct lov_object *lov)
1203 {
1204         LASSERT(lov->lo_owner != current);
1205         down_write(&lov->lo_type_guard);
1206         LASSERT(lov->lo_owner == NULL);
1207         lov->lo_owner = current;
1208         CDEBUG(D_INODE, "Took exclusive lov(%p) owner %p\n",
1209                 lov, lov->lo_owner);
1210 }
1211
1212 static void lov_conf_unlock(struct lov_object *lov)
1213 {
1214         CDEBUG(D_INODE, "To release exclusive lov(%p) owner %p\n",
1215                 lov, lov->lo_owner);
1216         lov->lo_owner = NULL;
1217         up_write(&lov->lo_type_guard);
1218 }
1219
1220 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov)
1221 {
1222         struct l_wait_info lwi = { 0 };
1223         ENTRY;
1224
1225         while (atomic_read(&lov->lo_active_ios) > 0) {
1226                 CDEBUG(D_INODE, "file:"DFID" wait for active IO, now: %d.\n",
1227                         PFID(lu_object_fid(lov2lu(lov))),
1228                         atomic_read(&lov->lo_active_ios));
1229
1230                 l_wait_event(lov->lo_waitq,
1231                              atomic_read(&lov->lo_active_ios) == 0, &lwi);
1232         }
1233         RETURN(0);
1234 }
1235
1236 static int lov_layout_change(const struct lu_env *unused,
1237                              struct lov_object *lov, struct lov_stripe_md *lsm,
1238                              const struct cl_object_conf *conf)
1239 {
1240         enum lov_layout_type llt = lov_type(lsm);
1241         union lov_layout_state *state = &lov->u;
1242         const struct lov_layout_operations *old_ops;
1243         const struct lov_layout_operations *new_ops;
1244         struct lov_device *lov_dev = lov_object_dev(lov);
1245         struct lu_env *env;
1246         __u16 refcheck;
1247         int rc;
1248         ENTRY;
1249
1250         LASSERT(lov->lo_type < ARRAY_SIZE(lov_dispatch));
1251
1252         env = cl_env_get(&refcheck);
1253         if (IS_ERR(env))
1254                 RETURN(PTR_ERR(env));
1255
1256         LASSERT(llt < ARRAY_SIZE(lov_dispatch));
1257
1258         CDEBUG(D_INODE, DFID" from %s to %s\n",
1259                PFID(lu_object_fid(lov2lu(lov))),
1260                llt2str(lov->lo_type), llt2str(llt));
1261
1262         old_ops = &lov_dispatch[lov->lo_type];
1263         new_ops = &lov_dispatch[llt];
1264
1265         rc = cl_object_prune(env, &lov->lo_cl);
1266         if (rc != 0)
1267                 GOTO(out, rc);
1268
1269         rc = old_ops->llo_delete(env, lov, &lov->u);
1270         if (rc != 0)
1271                 GOTO(out, rc);
1272
1273         old_ops->llo_fini(env, lov, &lov->u);
1274
1275         LASSERT(atomic_read(&lov->lo_active_ios) == 0);
1276
1277         CDEBUG(D_INODE, DFID "Apply new layout lov %p, type %d\n",
1278                PFID(lu_object_fid(lov2lu(lov))), lov, llt);
1279
1280         /* page bufsize fixup */
1281         cl_object_header(&lov->lo_cl)->coh_page_bufsize -=
1282                 lov_page_slice_fixup(lov, NULL);
1283
1284         lov->lo_type = llt;
1285         rc = new_ops->llo_init(env, lov_dev, lov, lsm, conf, state);
1286         if (rc != 0) {
1287                 struct obd_device *obd = lov2obd(lov_dev->ld_lov);
1288
1289                 CERROR("%s: cannot apply new layout on "DFID" : rc = %d\n",
1290                        obd->obd_name, PFID(lu_object_fid(lov2lu(lov))), rc);
1291                 new_ops->llo_delete(env, lov, state);
1292                 new_ops->llo_fini(env, lov, state);
1293                 /* this file becomes an EMPTY file. */
1294                 lov->lo_type = LLT_EMPTY;
1295                 GOTO(out, rc);
1296         }
1297
1298 out:
1299         cl_env_put(env, &refcheck);
1300         RETURN(rc);
1301 }
1302
1303 /*****************************************************************************
1304  *
1305  * Lov object operations.
1306  *
1307  */
1308 int lov_object_init(const struct lu_env *env, struct lu_object *obj,
1309                     const struct lu_object_conf *conf)
1310 {
1311         struct lov_object            *lov   = lu2lov(obj);
1312         struct lov_device            *dev   = lov_object_dev(lov);
1313         const struct cl_object_conf  *cconf = lu2cl_conf(conf);
1314         union lov_layout_state       *set   = &lov->u;
1315         const struct lov_layout_operations *ops;
1316         struct lov_stripe_md *lsm = NULL;
1317         int rc;
1318         ENTRY;
1319
1320         init_rwsem(&lov->lo_type_guard);
1321         atomic_set(&lov->lo_active_ios, 0);
1322         init_waitqueue_head(&lov->lo_waitq);
1323         cl_object_page_init(lu2cl(obj), sizeof(struct lov_page));
1324
1325         lov->lo_type = LLT_EMPTY;
1326         if (cconf->u.coc_layout.lb_buf != NULL) {
1327                 lsm = lov_unpackmd(dev->ld_lov,
1328                                    cconf->u.coc_layout.lb_buf,
1329                                    cconf->u.coc_layout.lb_len);
1330                 if (IS_ERR(lsm))
1331                         RETURN(PTR_ERR(lsm));
1332
1333                 dump_lsm(D_INODE, lsm);
1334         }
1335
1336         /* no locking is necessary, as object is being created */
1337         lov->lo_type = lov_type(lsm);
1338         ops = &lov_dispatch[lov->lo_type];
1339         rc = ops->llo_init(env, dev, lov, lsm, cconf, set);
1340         if (rc != 0)
1341                 GOTO(out_lsm, rc);
1342
1343 out_lsm:
1344         lov_lsm_put(lsm);
1345
1346         RETURN(rc);
1347 }
1348
1349 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
1350                         const struct cl_object_conf *conf)
1351 {
1352         struct lov_stripe_md    *lsm = NULL;
1353         struct lov_object       *lov = cl2lov(obj);
1354         int                      result = 0;
1355         ENTRY;
1356
1357         if (conf->coc_opc == OBJECT_CONF_SET &&
1358             conf->u.coc_layout.lb_buf != NULL) {
1359                 lsm = lov_unpackmd(lov_object_dev(lov)->ld_lov,
1360                                    conf->u.coc_layout.lb_buf,
1361                                    conf->u.coc_layout.lb_len);
1362                 if (IS_ERR(lsm))
1363                         RETURN(PTR_ERR(lsm));
1364                 dump_lsm(D_INODE, lsm);
1365         }
1366
1367         lov_conf_lock(lov);
1368         if (conf->coc_opc == OBJECT_CONF_INVALIDATE) {
1369                 lov->lo_layout_invalid = true;
1370                 GOTO(out, result = 0);
1371         }
1372
1373         if (conf->coc_opc == OBJECT_CONF_WAIT) {
1374                 if (lov->lo_layout_invalid &&
1375                     atomic_read(&lov->lo_active_ios) > 0) {
1376                         lov_conf_unlock(lov);
1377                         result = lov_layout_wait(env, lov);
1378                         lov_conf_lock(lov);
1379                 }
1380                 GOTO(out, result);
1381         }
1382
1383         LASSERT(conf->coc_opc == OBJECT_CONF_SET);
1384
1385         if ((lsm == NULL && lov->lo_lsm == NULL) ||
1386             ((lsm != NULL && lov->lo_lsm != NULL) &&
1387              (lov->lo_lsm->lsm_layout_gen == lsm->lsm_layout_gen) &&
1388              (lov->lo_lsm->lsm_entries[0]->lsme_pattern ==
1389               lsm->lsm_entries[0]->lsme_pattern))) {
1390                 /* same version of layout */
1391                 lov->lo_layout_invalid = false;
1392                 GOTO(out, result = 0);
1393         }
1394
1395         /* will change layout - check if there still exists active IO. */
1396         if (atomic_read(&lov->lo_active_ios) > 0) {
1397                 lov->lo_layout_invalid = true;
1398                 GOTO(out, result = -EBUSY);
1399         }
1400
1401         result = lov_layout_change(env, lov, lsm, conf);
1402         lov->lo_layout_invalid = result != 0;
1403         EXIT;
1404
1405 out:
1406         lov_conf_unlock(lov);
1407         lov_lsm_put(lsm);
1408         CDEBUG(D_INODE, DFID" lo_layout_invalid=%d\n",
1409                PFID(lu_object_fid(lov2lu(lov))), lov->lo_layout_invalid);
1410         RETURN(result);
1411 }
1412
1413 static void lov_object_delete(const struct lu_env *env, struct lu_object *obj)
1414 {
1415         struct lov_object *lov = lu2lov(obj);
1416
1417         ENTRY;
1418         LOV_2DISPATCH_VOID(lov, llo_delete, env, lov, &lov->u);
1419         EXIT;
1420 }
1421
1422 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
1423 {
1424         struct lov_object *lov = lu2lov(obj);
1425
1426         ENTRY;
1427         LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
1428         lu_object_fini(obj);
1429         OBD_SLAB_FREE_PTR(lov, lov_object_kmem);
1430         EXIT;
1431 }
1432
1433 static int lov_object_print(const struct lu_env *env, void *cookie,
1434                             lu_printer_t p, const struct lu_object *o)
1435 {
1436         return LOV_2DISPATCH_NOLOCK(lu2lov(o), llo_print, env, cookie, p, o);
1437 }
1438
1439 int lov_page_init(const struct lu_env *env, struct cl_object *obj,
1440                   struct cl_page *page, pgoff_t index)
1441 {
1442         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_page_init, env, obj, page,
1443                                     index);
1444 }
1445
1446 /**
1447  * Implements cl_object_operations::clo_io_init() method for lov
1448  * layer. Dispatches to the appropriate layout io initialization method.
1449  */
1450 int lov_io_init(const struct lu_env *env, struct cl_object *obj,
1451                 struct cl_io *io)
1452 {
1453         CL_IO_SLICE_CLEAN(lov_env_io(env), lis_preserved);
1454
1455         CDEBUG(D_INODE, DFID "io %p type %d ignore/verify layout %d/%d\n",
1456                PFID(lu_object_fid(&obj->co_lu)), io, io->ci_type,
1457                io->ci_ignore_layout, io->ci_verify_layout);
1458
1459         /* IO type CIT_MISC with ci_ignore_layout set are usually invoked from
1460          * the OSC layer. It shouldn't take lov layout conf lock in that case,
1461          * because as long as the OSC object exists, the layout can't be
1462          * reconfigured. */
1463         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
1464                         !(io->ci_ignore_layout && io->ci_type == CIT_MISC),
1465                         env, obj, io);
1466 }
1467
1468 /**
1469  * An implementation of cl_object_operations::clo_attr_get() method for lov
1470  * layer. For raid0 layout this collects and merges attributes of all
1471  * sub-objects.
1472  */
1473 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
1474                         struct cl_attr *attr)
1475 {
1476         /* do not take lock, as this function is called under a
1477          * spin-lock. Layout is protected from changing by ongoing IO. */
1478         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
1479 }
1480
1481 static int lov_attr_update(const struct lu_env *env, struct cl_object *obj,
1482                            const struct cl_attr *attr, unsigned valid)
1483 {
1484         /*
1485          * No dispatch is required here, as no layout implements this.
1486          */
1487         return 0;
1488 }
1489
1490 int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
1491                   struct cl_lock *lock, const struct cl_io *io)
1492 {
1493         /* No need to lock because we've taken one refcount of layout.  */
1494         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_lock_init, env, obj, lock,
1495                                     io);
1496 }
1497
1498 /**
1499  * We calculate on which OST the mapping will end. If the length of mapping
1500  * is greater than (stripe_size * stripe_count) then the last_stripe will
1501  * will be one just before start_stripe. Else we check if the mapping
1502  * intersects each OST and find last_stripe.
1503  * This function returns the last_stripe and also sets the stripe_count
1504  * over which the mapping is spread
1505  *
1506  * \param lsm [in]              striping information for the file
1507  * \param index [in]            stripe component index
1508  * \param ext [in]              logical extent of mapping
1509  * \param start_stripe [in]     starting stripe of the mapping
1510  * \param stripe_count [out]    the number of stripes across which to map is
1511  *                              returned
1512  *
1513  * \retval last_stripe          return the last stripe of the mapping
1514  */
1515 static int fiemap_calc_last_stripe(struct lov_stripe_md *lsm, int index,
1516                                    struct lu_extent *ext,
1517                                    int start_stripe, int *stripe_count)
1518 {
1519         struct lov_stripe_md_entry *lsme = lsm->lsm_entries[index];
1520         int last_stripe;
1521         u64 obd_start;
1522         u64 obd_end;
1523         int i, j;
1524
1525         if (ext->e_end - ext->e_start >
1526             lsme->lsme_stripe_size * lsme->lsme_stripe_count) {
1527                 last_stripe = (start_stripe < 1 ? lsme->lsme_stripe_count - 1 :
1528                                                   start_stripe - 1);
1529                 *stripe_count = lsme->lsme_stripe_count;
1530         } else {
1531                 for (j = 0, i = start_stripe; j < lsme->lsme_stripe_count;
1532                      i = (i + 1) % lsme->lsme_stripe_count, j++) {
1533                         if ((lov_stripe_intersects(lsm, index,  i, ext,
1534                                                    &obd_start, &obd_end)) == 0)
1535                                 break;
1536                 }
1537                 *stripe_count = j;
1538                 last_stripe = (start_stripe + j - 1) % lsme->lsme_stripe_count;
1539         }
1540
1541         return last_stripe;
1542 }
1543
1544 /**
1545  * Set fe_device and copy extents from local buffer into main return buffer.
1546  *
1547  * \param fiemap [out]          fiemap to hold all extents
1548  * \param lcl_fm_ext [in]       array of fiemap extents get from OSC layer
1549  * \param ost_index [in]        OST index to be written into the fm_device
1550  *                              field for each extent
1551  * \param ext_count [in]        number of extents to be copied
1552  * \param current_extent [in]   where to start copying in the extent array
1553  */
1554 static void fiemap_prepare_and_copy_exts(struct fiemap *fiemap,
1555                                          struct fiemap_extent *lcl_fm_ext,
1556                                          int ost_index, unsigned int ext_count,
1557                                          int current_extent)
1558 {
1559         char            *to;
1560         unsigned int    ext;
1561
1562         for (ext = 0; ext < ext_count; ext++) {
1563                 lcl_fm_ext[ext].fe_device = ost_index;
1564                 lcl_fm_ext[ext].fe_flags |= FIEMAP_EXTENT_NET;
1565         }
1566
1567         /* Copy fm_extent's from fm_local to return buffer */
1568         to = (char *)fiemap + fiemap_count_to_size(current_extent);
1569         memcpy(to, lcl_fm_ext, ext_count * sizeof(struct fiemap_extent));
1570 }
1571
1572 #define FIEMAP_BUFFER_SIZE 4096
1573
1574 /**
1575  * Non-zero fe_logical indicates that this is a continuation FIEMAP
1576  * call. The local end offset and the device are sent in the first
1577  * fm_extent. This function calculates the stripe number from the index.
1578  * This function returns a stripe_no on which mapping is to be restarted.
1579  *
1580  * This function returns fm_end_offset which is the in-OST offset at which
1581  * mapping should be restarted. If fm_end_offset=0 is returned then caller
1582  * will re-calculate proper offset in next stripe.
1583  * Note that the first extent is passed to lov_get_info via the value field.
1584  *
1585  * \param fiemap [in]           fiemap request header
1586  * \param lsm [in]              striping information for the file
1587  * \param index [in]            stripe component index
1588  * \param ext [in]              logical extent of mapping
1589  * \param start_stripe [out]    starting stripe will be returned in this
1590  */
1591 static u64 fiemap_calc_fm_end_offset(struct fiemap *fiemap,
1592                                      struct lov_stripe_md *lsm,
1593                                      int index, struct lu_extent *ext,
1594                                      int *start_stripe)
1595 {
1596         struct lov_stripe_md_entry *lsme = lsm->lsm_entries[index];
1597         u64 local_end = fiemap->fm_extents[0].fe_logical;
1598         u64 lun_start;
1599         u64 lun_end;
1600         u64 fm_end_offset;
1601         int stripe_no = -1;
1602         int i;
1603
1604         if (fiemap->fm_extent_count == 0 ||
1605             fiemap->fm_extents[0].fe_logical == 0)
1606                 return 0;
1607
1608         /* Find out stripe_no from ost_index saved in the fe_device */
1609         for (i = 0; i < lsme->lsme_stripe_count; i++) {
1610                 struct lov_oinfo *oinfo = lsme->lsme_oinfo[i];
1611
1612                 if (lov_oinfo_is_dummy(oinfo))
1613                         continue;
1614
1615                 if (oinfo->loi_ost_idx == fiemap->fm_extents[0].fe_device) {
1616                         stripe_no = i;
1617                         break;
1618                 }
1619         }
1620
1621         if (stripe_no == -1)
1622                 return -EINVAL;
1623
1624         /* If we have finished mapping on previous device, shift logical
1625          * offset to start of next device */
1626         if (lov_stripe_intersects(lsm, index, stripe_no, ext,
1627                                    &lun_start, &lun_end) != 0 &&
1628             local_end < lun_end) {
1629                 fm_end_offset = local_end;
1630                 *start_stripe = stripe_no;
1631         } else {
1632                 /* This is a special value to indicate that caller should
1633                  * calculate offset in next stripe. */
1634                 fm_end_offset = 0;
1635                 *start_stripe = (stripe_no + 1) % lsme->lsme_stripe_count;
1636         }
1637
1638         return fm_end_offset;
1639 }
1640
1641 struct fiemap_state {
1642         struct fiemap           *fs_fm;
1643         struct lu_extent        fs_ext;
1644         u64                     fs_length;
1645         u64                     fs_end_offset;
1646         int                     fs_cur_extent;
1647         int                     fs_cnt_need;
1648         int                     fs_start_stripe;
1649         int                     fs_last_stripe;
1650         bool                    fs_device_done;
1651         bool                    fs_finish_stripe;
1652         bool                    fs_enough;
1653 };
1654
1655 static struct cl_object *lov_find_subobj(const struct lu_env *env,
1656                                          struct lov_object *lov,
1657                                          struct lov_stripe_md *lsm,
1658                                          int index)
1659 {
1660         struct lov_device       *dev = lu2lov_dev(lov2lu(lov)->lo_dev);
1661         struct lov_thread_info  *lti = lov_env_info(env);
1662         struct lu_fid           *ofid = &lti->lti_fid;
1663         struct lov_oinfo        *oinfo;
1664         struct cl_device        *subdev;
1665         int                     entry = lov_comp_entry(index);
1666         int                     stripe = lov_comp_stripe(index);
1667         int                     ost_idx;
1668         int                     rc;
1669         struct cl_object        *result;
1670
1671         if (lov->lo_type != LLT_COMP)
1672                 GOTO(out, result = NULL);
1673
1674         if (entry >= lsm->lsm_entry_count ||
1675             stripe >= lsm->lsm_entries[entry]->lsme_stripe_count)
1676                 GOTO(out, result = NULL);
1677
1678         oinfo = lsm->lsm_entries[entry]->lsme_oinfo[stripe];
1679         ost_idx = oinfo->loi_ost_idx;
1680         rc = ostid_to_fid(ofid, &oinfo->loi_oi, ost_idx);
1681         if (rc != 0)
1682                 GOTO(out, result = NULL);
1683
1684         subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
1685         result = lov_sub_find(env, subdev, ofid, NULL);
1686 out:
1687         if (result == NULL)
1688                 result = ERR_PTR(-EINVAL);
1689         return result;
1690 }
1691
1692 int fiemap_for_stripe(const struct lu_env *env, struct cl_object *obj,
1693                       struct lov_stripe_md *lsm, struct fiemap *fiemap,
1694                       size_t *buflen, struct ll_fiemap_info_key *fmkey,
1695                       int index, int stripeno, struct fiemap_state *fs)
1696 {
1697         struct lov_stripe_md_entry *lsme = lsm->lsm_entries[index];
1698         struct cl_object *subobj;
1699         struct lov_obd *lov = lu2lov_dev(obj->co_lu.lo_dev)->ld_lov;
1700         struct fiemap_extent *fm_ext = &fs->fs_fm->fm_extents[0];
1701         u64 req_fm_len; /* Stores length of required mapping */
1702         u64 len_mapped_single_call;
1703         u64 lun_start;
1704         u64 lun_end;
1705         u64 obd_object_end;
1706         unsigned int ext_count;
1707         /* EOF for object */
1708         bool ost_eof = false;
1709         /* done with required mapping for this OST? */
1710         bool ost_done = false;
1711         int ost_index;
1712         int rc = 0;
1713
1714         fs->fs_device_done = false;
1715         /* Find out range of mapping on this stripe */
1716         if ((lov_stripe_intersects(lsm, index, stripeno, &fs->fs_ext,
1717                                    &lun_start, &obd_object_end)) == 0)
1718                 return 0;
1719
1720         if (lov_oinfo_is_dummy(lsme->lsme_oinfo[stripeno]))
1721                 return -EIO;
1722
1723         /* If this is a continuation FIEMAP call and we are on
1724          * starting stripe then lun_start needs to be set to
1725          * end_offset */
1726         if (fs->fs_end_offset != 0 && stripeno == fs->fs_start_stripe)
1727                 lun_start = fs->fs_end_offset;
1728         lun_end = lov_size_to_stripe(lsm, index, fs->fs_ext.e_end, stripeno);
1729         if (lun_start == lun_end)
1730                 return 0;
1731
1732         req_fm_len = obd_object_end - lun_start + 1;
1733         fs->fs_fm->fm_length = 0;
1734         len_mapped_single_call = 0;
1735
1736         /* find lobsub object */
1737         subobj = lov_find_subobj(env, cl2lov(obj), lsm,
1738                                  lov_comp_index(index, stripeno));
1739         if (IS_ERR(subobj))
1740                 return PTR_ERR(subobj);
1741         /* If the output buffer is very large and the objects have many
1742          * extents we may need to loop on a single OST repeatedly */
1743         do {
1744                 if (fiemap->fm_extent_count > 0) {
1745                         /* Don't get too many extents. */
1746                         if (fs->fs_cur_extent + fs->fs_cnt_need >
1747                             fiemap->fm_extent_count)
1748                                 fs->fs_cnt_need = fiemap->fm_extent_count -
1749                                                   fs->fs_cur_extent;
1750                 }
1751
1752                 lun_start += len_mapped_single_call;
1753                 fs->fs_fm->fm_length = req_fm_len - len_mapped_single_call;
1754                 req_fm_len = fs->fs_fm->fm_length;
1755                 /**
1756                  * If we've collected enough extent map, we'd request 1 more,
1757                  * to see whether we coincidentally finished all available
1758                  * extent map, so that FIEMAP_EXTENT_LAST would be set.
1759                  */
1760                 fs->fs_fm->fm_extent_count = fs->fs_enough ?
1761                                              1 : fs->fs_cnt_need;
1762                 fs->fs_fm->fm_mapped_extents = 0;
1763                 fs->fs_fm->fm_flags = fiemap->fm_flags;
1764
1765                 ost_index = lsme->lsme_oinfo[stripeno]->loi_ost_idx;
1766
1767                 if (ost_index < 0 || ost_index >= lov->desc.ld_tgt_count)
1768                         GOTO(obj_put, rc = -EINVAL);
1769                 /* If OST is inactive, return extent with UNKNOWN flag. */
1770                 if (!lov->lov_tgts[ost_index]->ltd_active) {
1771                         fs->fs_fm->fm_flags |= FIEMAP_EXTENT_LAST;
1772                         fs->fs_fm->fm_mapped_extents = 1;
1773
1774                         fm_ext[0].fe_logical = lun_start;
1775                         fm_ext[0].fe_length = obd_object_end - lun_start + 1;
1776                         fm_ext[0].fe_flags |= FIEMAP_EXTENT_UNKNOWN;
1777
1778                         goto inactive_tgt;
1779                 }
1780
1781                 fs->fs_fm->fm_start = lun_start;
1782                 fs->fs_fm->fm_flags &= ~FIEMAP_FLAG_DEVICE_ORDER;
1783                 memcpy(&fmkey->lfik_fiemap, fs->fs_fm, sizeof(*fs->fs_fm));
1784                 *buflen = fiemap_count_to_size(fs->fs_fm->fm_extent_count);
1785
1786                 rc = cl_object_fiemap(env, subobj, fmkey, fs->fs_fm, buflen);
1787                 if (rc != 0)
1788                         GOTO(obj_put, rc);
1789 inactive_tgt:
1790                 ext_count = fs->fs_fm->fm_mapped_extents;
1791                 if (ext_count == 0) {
1792                         ost_done = true;
1793                         fs->fs_device_done = true;
1794                         /* If last stripe has hold at the end,
1795                          * we need to return */
1796                         if (stripeno == fs->fs_last_stripe) {
1797                                 fiemap->fm_mapped_extents = 0;
1798                                 fs->fs_finish_stripe = true;
1799                                 GOTO(obj_put, rc);
1800                         }
1801                         break;
1802                 } else if (fs->fs_enough) {
1803                         /*
1804                          * We've collected enough extents and there are
1805                          * more extents after it.
1806                          */
1807                         GOTO(obj_put, rc);
1808                 }
1809
1810                 /* If we just need num of extents, got to next device */
1811                 if (fiemap->fm_extent_count == 0) {
1812                         fs->fs_cur_extent += ext_count;
1813                         break;
1814                 }
1815
1816                 /* prepare to copy retrived map extents */
1817                 len_mapped_single_call = fm_ext[ext_count - 1].fe_logical +
1818                                          fm_ext[ext_count - 1].fe_length -
1819                                          lun_start;
1820
1821                 /* Have we finished mapping on this device? */
1822                 if (req_fm_len <= len_mapped_single_call) {
1823                         ost_done = true;
1824                         fs->fs_device_done = true;
1825                 }
1826
1827                 /* Clear the EXTENT_LAST flag which can be present on
1828                  * the last extent */
1829                 if (fm_ext[ext_count - 1].fe_flags & FIEMAP_EXTENT_LAST)
1830                         fm_ext[ext_count - 1].fe_flags &= ~FIEMAP_EXTENT_LAST;
1831                 if (lov_stripe_size(lsm, index,
1832                                     fm_ext[ext_count - 1].fe_logical +
1833                                     fm_ext[ext_count - 1].fe_length,
1834                                     stripeno) >= fmkey->lfik_oa.o_size) {
1835                         ost_eof = true;
1836                         fs->fs_device_done = true;
1837                 }
1838
1839                 fiemap_prepare_and_copy_exts(fiemap, fm_ext, ost_index,
1840                                              ext_count, fs->fs_cur_extent);
1841                 fs->fs_cur_extent += ext_count;
1842
1843                 /* Ran out of available extents? */
1844                 if (fs->fs_cur_extent >= fiemap->fm_extent_count)
1845                         fs->fs_enough = true;
1846         } while (!ost_done && !ost_eof);
1847
1848         if (stripeno == fs->fs_last_stripe)
1849                 fs->fs_finish_stripe = true;
1850 obj_put:
1851         cl_object_put(env, subobj);
1852
1853         return rc;
1854 }
1855
1856 /**
1857  * Break down the FIEMAP request and send appropriate calls to individual OSTs.
1858  * This also handles the restarting of FIEMAP calls in case mapping overflows
1859  * the available number of extents in single call.
1860  *
1861  * \param env [in]              lustre environment
1862  * \param obj [in]              file object
1863  * \param fmkey [in]            fiemap request header and other info
1864  * \param fiemap [out]          fiemap buffer holding retrived map extents
1865  * \param buflen [in/out]       max buffer length of @fiemap, when iterate
1866  *                              each OST, it is used to limit max map needed
1867  * \retval 0    success
1868  * \retval < 0  error
1869  */
1870 static int lov_object_fiemap(const struct lu_env *env, struct cl_object *obj,
1871                              struct ll_fiemap_info_key *fmkey,
1872                              struct fiemap *fiemap, size_t *buflen)
1873 {
1874         struct lov_stripe_md_entry *lsme;
1875         struct lov_stripe_md *lsm;
1876         struct fiemap *fm_local = NULL;
1877         loff_t whole_start;
1878         loff_t whole_end;
1879         int entry;
1880         int start_entry;
1881         int end_entry;
1882         int cur_stripe = 0;
1883         int stripe_count;
1884         unsigned int buffer_size = FIEMAP_BUFFER_SIZE;
1885         int rc = 0;
1886         struct fiemap_state fs = { 0 };
1887         ENTRY;
1888
1889         lsm = lov_lsm_addref(cl2lov(obj));
1890         if (lsm == NULL) {
1891                 /* no extent: there is no object for mapping */
1892                 fiemap->fm_mapped_extents = 0;
1893                 return 0;
1894         }
1895
1896         if (!(fiemap->fm_flags & FIEMAP_FLAG_DEVICE_ORDER)) {
1897                 /**
1898                  * If the entry count > 1 or stripe_count > 1 and the
1899                  * application does not understand DEVICE_ORDER flag,
1900                  * it cannot interpret the extents correctly.
1901                  */
1902                 if (lsm->lsm_entry_count > 1 ||
1903                     (lsm->lsm_entry_count == 1 &&
1904                      lsm->lsm_entries[0]->lsme_stripe_count > 1))
1905                         GOTO(out_lsm, rc = -ENOTSUPP);
1906         }
1907
1908         /* No support for DOM layout yet. */
1909         if (lsme_is_dom(lsm->lsm_entries[0]))
1910                 GOTO(out_lsm, rc = -ENOTSUPP);
1911
1912         if (lsm->lsm_is_released) {
1913                 if (fiemap->fm_start < fmkey->lfik_oa.o_size) {
1914                         /**
1915                          * released file, return a minimal FIEMAP if
1916                          * request fits in file-size.
1917                          */
1918                         fiemap->fm_mapped_extents = 1;
1919                         fiemap->fm_extents[0].fe_logical = fiemap->fm_start;
1920                         if (fiemap->fm_start + fiemap->fm_length <
1921                             fmkey->lfik_oa.o_size)
1922                                 fiemap->fm_extents[0].fe_length =
1923                                         fiemap->fm_length;
1924                         else
1925                                 fiemap->fm_extents[0].fe_length =
1926                                         fmkey->lfik_oa.o_size -
1927                                         fiemap->fm_start;
1928                         fiemap->fm_extents[0].fe_flags |=
1929                                 FIEMAP_EXTENT_UNKNOWN | FIEMAP_EXTENT_LAST;
1930                 }
1931                 GOTO(out_lsm, rc = 0);
1932         }
1933
1934         /* buffer_size is small to hold fm_extent_count of extents. */
1935         if (fiemap_count_to_size(fiemap->fm_extent_count) < buffer_size)
1936                 buffer_size = fiemap_count_to_size(fiemap->fm_extent_count);
1937
1938         OBD_ALLOC_LARGE(fm_local, buffer_size);
1939         if (fm_local == NULL)
1940                 GOTO(out_lsm, rc = -ENOMEM);
1941
1942         /**
1943          * Requested extent count exceeds the fiemap buffer size, shrink our
1944          * ambition.
1945          */
1946         if (fiemap_count_to_size(fiemap->fm_extent_count) > *buflen)
1947                 fiemap->fm_extent_count = fiemap_size_to_count(*buflen);
1948         if (fiemap->fm_extent_count == 0)
1949                 fs.fs_cnt_need = 0;
1950
1951         fs.fs_enough = false;
1952         fs.fs_cur_extent = 0;
1953         fs.fs_fm = fm_local;
1954         fs.fs_cnt_need = fiemap_size_to_count(buffer_size);
1955
1956         whole_start = fiemap->fm_start;
1957         /* whole_start is beyond the end of the file */
1958         if (whole_start > fmkey->lfik_oa.o_size)
1959                 GOTO(out_fm_local, rc = -EINVAL);
1960         whole_end = (fiemap->fm_length == OBD_OBJECT_EOF) ?
1961                                         fmkey->lfik_oa.o_size :
1962                                         whole_start + fiemap->fm_length - 1;
1963         /**
1964          * If fiemap->fm_length != OBD_OBJECT_EOF but whole_end exceeds file
1965          * size
1966          */
1967         if (whole_end > fmkey->lfik_oa.o_size)
1968                 whole_end = fmkey->lfik_oa.o_size;
1969
1970         start_entry = lov_lsm_entry(lsm, whole_start);
1971         end_entry = lov_lsm_entry(lsm, whole_end);
1972         if (end_entry == -1)
1973                 end_entry = lsm->lsm_entry_count - 1;
1974
1975         if (start_entry == -1 || end_entry == -1)
1976                 GOTO(out_fm_local, rc = -EINVAL);
1977
1978         /* TODO: rewrite it with lov_foreach_io_layout() */
1979         for (entry = start_entry; entry <= end_entry; entry++) {
1980                 lsme = lsm->lsm_entries[entry];
1981
1982                 if (!lsme_inited(lsme))
1983                         break;
1984
1985                 if (entry == start_entry)
1986                         fs.fs_ext.e_start = whole_start;
1987                 else
1988                         fs.fs_ext.e_start = lsme->lsme_extent.e_start;
1989                 if (entry == end_entry)
1990                         fs.fs_ext.e_end = whole_end;
1991                 else
1992                         fs.fs_ext.e_end = lsme->lsme_extent.e_end - 1;
1993                 fs.fs_length = fs.fs_ext.e_end - fs.fs_ext.e_start + 1;
1994
1995                 /* Calculate start stripe, last stripe and length of mapping */
1996                 fs.fs_start_stripe = lov_stripe_number(lsm, entry,
1997                                                        fs.fs_ext.e_start);
1998                 fs.fs_last_stripe = fiemap_calc_last_stripe(lsm, entry,
1999                                         &fs.fs_ext, fs.fs_start_stripe,
2000                                         &stripe_count);
2001                 fs.fs_end_offset = fiemap_calc_fm_end_offset(fiemap, lsm, entry,
2002                                         &fs.fs_ext, &fs.fs_start_stripe);
2003                 /* Check each stripe */
2004                 for (cur_stripe = fs.fs_start_stripe; stripe_count > 0;
2005                      --stripe_count,
2006                      cur_stripe = (cur_stripe + 1) % lsme->lsme_stripe_count) {
2007                         rc = fiemap_for_stripe(env, obj, lsm, fiemap, buflen,
2008                                                fmkey, entry, cur_stripe, &fs);
2009                         if (rc < 0)
2010                                 GOTO(out_fm_local, rc);
2011                         if (fs.fs_enough)
2012                                 GOTO(finish, rc);
2013                         if (fs.fs_finish_stripe)
2014                                 break;
2015                 } /* for each stripe */
2016         } /* for covering layout component */
2017         /*
2018          * We've traversed all components, set @entry to the last component
2019          * entry, it's for the last stripe check.
2020          */
2021         entry--;
2022 finish:
2023         /* Indicate that we are returning device offsets unless file just has
2024          * single stripe */
2025         if (lsm->lsm_entry_count > 1 ||
2026             (lsm->lsm_entry_count == 1 &&
2027              lsm->lsm_entries[0]->lsme_stripe_count > 1))
2028                 fiemap->fm_flags |= FIEMAP_FLAG_DEVICE_ORDER;
2029
2030         if (fiemap->fm_extent_count == 0)
2031                 goto skip_last_device_calc;
2032
2033         /* Check if we have reached the last stripe and whether mapping for that
2034          * stripe is done. */
2035         if ((cur_stripe == fs.fs_last_stripe) && fs.fs_device_done)
2036                 fiemap->fm_extents[fs.fs_cur_extent - 1].fe_flags |=
2037                                                              FIEMAP_EXTENT_LAST;
2038 skip_last_device_calc:
2039         fiemap->fm_mapped_extents = fs.fs_cur_extent;
2040 out_fm_local:
2041         OBD_FREE_LARGE(fm_local, buffer_size);
2042
2043 out_lsm:
2044         lov_lsm_put(lsm);
2045         return rc;
2046 }
2047
2048 static int lov_object_getstripe(const struct lu_env *env, struct cl_object *obj,
2049                                 struct lov_user_md __user *lum, size_t size)
2050 {
2051         struct lov_object       *lov = cl2lov(obj);
2052         struct lov_stripe_md    *lsm;
2053         int                     rc = 0;
2054         ENTRY;
2055
2056         lsm = lov_lsm_addref(lov);
2057         if (lsm == NULL)
2058                 RETURN(-ENODATA);
2059
2060         rc = lov_getstripe(env, cl2lov(obj), lsm, lum, size);
2061         lov_lsm_put(lsm);
2062         RETURN(rc);
2063 }
2064
2065 static int lov_object_layout_get(const struct lu_env *env,
2066                                  struct cl_object *obj,
2067                                  struct cl_layout *cl)
2068 {
2069         struct lov_object *lov = cl2lov(obj);
2070         struct lov_stripe_md *lsm = lov_lsm_addref(lov);
2071         struct lu_buf *buf = &cl->cl_buf;
2072         ssize_t rc;
2073         ENTRY;
2074
2075         if (lsm == NULL) {
2076                 cl->cl_size = 0;
2077                 cl->cl_layout_gen = CL_LAYOUT_GEN_EMPTY;
2078
2079                 RETURN(0);
2080         }
2081
2082         cl->cl_size = lov_comp_md_size(lsm);
2083         cl->cl_layout_gen = lsm->lsm_layout_gen;
2084         cl->cl_dom_comp_size = 0;
2085         cl->cl_is_released = lsm->lsm_is_released;
2086         if (lsm_is_composite(lsm->lsm_magic)) {
2087                 struct lov_stripe_md_entry *lsme = lsm->lsm_entries[0];
2088
2089                 cl->cl_is_composite = true;
2090
2091                 if (lsme_is_dom(lsme))
2092                         cl->cl_dom_comp_size = lsme->lsme_extent.e_end;
2093         } else {
2094                 cl->cl_is_composite = false;
2095         }
2096
2097         rc = lov_lsm_pack(lsm, buf->lb_buf, buf->lb_len);
2098         lov_lsm_put(lsm);
2099
2100         RETURN(rc < 0 ? rc : 0);
2101 }
2102
2103 static loff_t lov_object_maxbytes(struct cl_object *obj)
2104 {
2105         struct lov_object *lov = cl2lov(obj);
2106         struct lov_stripe_md *lsm = lov_lsm_addref(lov);
2107         loff_t maxbytes;
2108
2109         if (lsm == NULL)
2110                 return LLONG_MAX;
2111
2112         maxbytes = lsm->lsm_maxbytes;
2113
2114         lov_lsm_put(lsm);
2115
2116         return maxbytes;
2117 }
2118
2119 static int lov_object_flush(const struct lu_env *env, struct cl_object *obj,
2120                             struct ldlm_lock *lock)
2121 {
2122         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_flush, env, obj, lock);
2123 }
2124
2125 static const struct cl_object_operations lov_ops = {
2126         .coo_page_init    = lov_page_init,
2127         .coo_lock_init    = lov_lock_init,
2128         .coo_io_init      = lov_io_init,
2129         .coo_attr_get     = lov_attr_get,
2130         .coo_attr_update  = lov_attr_update,
2131         .coo_conf_set     = lov_conf_set,
2132         .coo_getstripe    = lov_object_getstripe,
2133         .coo_layout_get   = lov_object_layout_get,
2134         .coo_maxbytes     = lov_object_maxbytes,
2135         .coo_fiemap       = lov_object_fiemap,
2136         .coo_object_flush = lov_object_flush
2137 };
2138
2139 static const struct lu_object_operations lov_lu_obj_ops = {
2140         .loo_object_init      = lov_object_init,
2141         .loo_object_delete    = lov_object_delete,
2142         .loo_object_release   = NULL,
2143         .loo_object_free      = lov_object_free,
2144         .loo_object_print     = lov_object_print,
2145         .loo_object_invariant = NULL
2146 };
2147
2148 struct lu_object *lov_object_alloc(const struct lu_env *env,
2149                                    const struct lu_object_header *unused,
2150                                    struct lu_device *dev)
2151 {
2152         struct lov_object *lov;
2153         struct lu_object  *obj;
2154
2155         ENTRY;
2156         OBD_SLAB_ALLOC_PTR_GFP(lov, lov_object_kmem, GFP_NOFS);
2157         if (lov != NULL) {
2158                 obj = lov2lu(lov);
2159                 lu_object_init(obj, NULL, dev);
2160                 lov->lo_cl.co_ops = &lov_ops;
2161                 lov->lo_type = -1; /* invalid, to catch uninitialized type */
2162                 /*
2163                  * object io operation vector (cl_object::co_iop) is installed
2164                  * later in lov_object_init(), as different vectors are used
2165                  * for object with different layouts.
2166                  */
2167                 obj->lo_ops = &lov_lu_obj_ops;
2168         } else
2169                 obj = NULL;
2170         RETURN(obj);
2171 }
2172
2173 struct lov_stripe_md *lov_lsm_addref(struct lov_object *lov)
2174 {
2175         struct lov_stripe_md *lsm = NULL;
2176
2177         lov_conf_freeze(lov);
2178         if (lov->lo_lsm != NULL) {
2179                 lsm = lsm_addref(lov->lo_lsm);
2180                 CDEBUG(D_INODE, "lsm %p addref %d/%d by %p.\n",
2181                         lsm, atomic_read(&lsm->lsm_refc),
2182                         lov->lo_layout_invalid, current);
2183         }
2184         lov_conf_thaw(lov);
2185         return lsm;
2186 }
2187
2188 int lov_read_and_clear_async_rc(struct cl_object *clob)
2189 {
2190         struct lu_object *luobj;
2191         int rc = 0;
2192         ENTRY;
2193
2194         luobj = lu_object_locate(&cl_object_header(clob)->coh_lu,
2195                                  &lov_device_type);
2196         if (luobj != NULL) {
2197                 struct lov_object *lov = lu2lov(luobj);
2198
2199                 lov_conf_freeze(lov);
2200                 switch (lov->lo_type) {
2201                 case LLT_COMP: {
2202                         struct lov_stripe_md *lsm;
2203                         int i;
2204
2205                         lsm = lov->lo_lsm;
2206                         LASSERT(lsm != NULL);
2207                         for (i = 0; i < lsm->lsm_entry_count; i++) {
2208                                 struct lov_stripe_md_entry *lse =
2209                                                 lsm->lsm_entries[i];
2210                                 int j;
2211
2212                                 if (!lsme_inited(lse))
2213                                         break;
2214
2215                                 for (j = 0; j < lse->lsme_stripe_count; j++) {
2216                                         struct lov_oinfo *loi =
2217                                                         lse->lsme_oinfo[j];
2218
2219                                         if (lov_oinfo_is_dummy(loi))
2220                                                 continue;
2221
2222                                         if (loi->loi_ar.ar_rc && !rc)
2223                                                 rc = loi->loi_ar.ar_rc;
2224                                         loi->loi_ar.ar_rc = 0;
2225                                 }
2226                         }
2227                 }
2228                 case LLT_RELEASED:
2229                 case LLT_EMPTY:
2230                         /* fall through */
2231                 case LLT_FOREIGN:
2232                         break;
2233                 default:
2234                         LBUG();
2235                 }
2236                 lov_conf_thaw(lov);
2237         }
2238         RETURN(rc);
2239 }
2240 EXPORT_SYMBOL(lov_read_and_clear_async_rc);
2241
2242 /** @} lov */