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 static struct lov_stripe_md *lov_lsm_addref(struct lov_object *lov);
  85
  86 static void lov_lsm_put(struct lov_stripe_md *lsm)
  87 {
  88         if (lsm != NULL)
  89                 lov_free_memmd(&lsm);
  90 }
  91
  92 /*****************************************************************************
  93  *
  94  * Lov object layout operations.
  95  *
  96  */
  97
  98 static struct cl_object *lov_sub_find(const struct lu_env *env,
  99                                       struct cl_device *dev,
 100                                       const struct lu_fid *fid,
 101                                       const struct cl_object_conf *conf)
 102 {
 103         struct lu_object *o;
 104
 105         ENTRY;
 106
 107         o = lu_object_find_at(env, cl2lu_dev(dev), fid, &conf->coc_lu);
 108         LASSERT(ergo(!IS_ERR(o), o->lo_dev->ld_type == &lovsub_device_type));
 109         RETURN(lu2cl(o));
 110 }
 111
 112 static int lov_page_slice_fixup(struct lov_object *lov,
 113                                 struct cl_object *stripe)
 114 {
 115         struct cl_object_header *hdr = cl_object_header(&lov->lo_cl);
 116         struct cl_object *o;
 117
 118         if (stripe == NULL)
 119                 return hdr->coh_page_bufsize - lov->lo_cl.co_slice_off -
 120                        cfs_size_round(sizeof(struct lov_page));
 121
 122         cl_object_for_each(o, stripe)
 123                 o->co_slice_off += hdr->coh_page_bufsize;
 124
 125         return cl_object_header(stripe)->coh_page_bufsize;
 126 }
 127
 128 static int lov_init_sub(const struct lu_env *env, struct lov_object *lov,
 129                         struct cl_object *subobj, struct lov_oinfo *oinfo,
 130                         int idx)
 131 {
 132         struct cl_object_header *hdr;
 133         struct cl_object_header *subhdr;
 134         struct cl_object_header *parent;
 135         int entry = lov_comp_entry(idx);
 136         int stripe = lov_comp_stripe(idx);
 137         int result;
 138
 139         if (OBD_FAIL_CHECK(OBD_FAIL_LOV_INIT)) {
 140                 /* For sanity:test_206.
 141                  * Do not leave the object in cache to avoid accessing
 142                  * freed memory. This is because osc_object is referring to
 143                  * lov_oinfo of lsm_stripe_data which will be freed due to
 144                  * this failure. */
 145                 cl_object_kill(env, subobj);
 146                 cl_object_put(env, subobj);
 147                 return -EIO;
 148         }
 149
 150         hdr = cl_object_header(lov2cl(lov));
 151         subhdr = cl_object_header(subobj);
 152
 153         CDEBUG(D_INODE, DFID"@%p[%d:%d] -> "DFID"@%p: ostid: "DOSTID
 154                " ost idx: %d gen: %d\n",
 155                PFID(lu_object_fid(&subobj->co_lu)), subhdr, entry, stripe,
 156                PFID(lu_object_fid(lov2lu(lov))), hdr, POSTID(&oinfo->loi_oi),
 157                oinfo->loi_ost_idx, oinfo->loi_ost_gen);
 158
 159         /* reuse ->coh_attr_guard to protect coh_parent change */
 160         spin_lock(&subhdr->coh_attr_guard);
 161         parent = subhdr->coh_parent;
 162         if (parent == NULL) {
 163                 struct lovsub_object *lso = cl2lovsub(subobj);
 164
 165                 subhdr->coh_parent = hdr;
 166                 spin_unlock(&subhdr->coh_attr_guard);
 167                 subhdr->coh_nesting = hdr->coh_nesting + 1;
 168                 lu_object_ref_add(&subobj->co_lu, "lov-parent", lov);
 169                 lso->lso_super = lov;
 170                 lso->lso_index = idx;
 171                 result = 0;
 172         } else {
 173                 struct lu_object  *old_obj;
 174                 struct lov_object *old_lov;
 175                 unsigned int mask = D_INODE;
 176
 177                 spin_unlock(&subhdr->coh_attr_guard);
 178                 old_obj = lu_object_locate(&parent->coh_lu, &lov_device_type);
 179                 LASSERT(old_obj != NULL);
 180                 old_lov = cl2lov(lu2cl(old_obj));
 181                 if (old_lov->lo_layout_invalid) {
 182                         /* the object's layout has already changed but isn't
 183                          * refreshed */
 184                         lu_object_unhash(env, &subobj->co_lu);
 185                         result = -EAGAIN;
 186                 } else {
 187                         mask = D_ERROR;
 188                         result = -EIO;
 189                 }
 190
 191                 LU_OBJECT_DEBUG(mask, env, &subobj->co_lu,
 192                                 "stripe %d is already owned.", idx);
 193                 LU_OBJECT_DEBUG(mask, env, old_obj, "owned.");
 194                 LU_OBJECT_HEADER(mask, env, lov2lu(lov), "try to own.\n");
 195                 cl_object_put(env, subobj);
 196         }
 197         return result;
 198 }
 199
 200 static int lov_init_raid0(const struct lu_env *env, struct lov_device *dev,
 201                           struct lov_object *lov, unsigned int index,
 202                           const struct cl_object_conf *conf,
 203                           struct lov_layout_entry *lle)
 204 {
 205         struct lov_layout_raid0 *r0 = &lle->lle_raid0;
 206         struct lov_thread_info *lti = lov_env_info(env);
 207         struct cl_object_conf *subconf = &lti->lti_stripe_conf;
 208         struct lu_fid *ofid = &lti->lti_fid;
 209         struct cl_object *stripe;
 210         struct lov_stripe_md_entry *lse  = lov_lse(lov, index);
 211         int result;
 212         int psz, sz;
 213         int i;
 214
 215         ENTRY;
 216
 217         spin_lock_init(&r0->lo_sub_lock);
 218         r0->lo_nr = lse->lsme_stripe_count;
 219         r0->lo_trunc_stripeno = -1;
 220
 221         OBD_ALLOC_PTR_ARRAY_LARGE(r0->lo_sub, r0->lo_nr);
 222         if (r0->lo_sub == NULL)
 223                 GOTO(out, result = -ENOMEM);
 224
 225         psz = 0;
 226         result = 0;
 227         memset(subconf, 0, sizeof(*subconf));
 228
 229         /*
 230          * Create stripe cl_objects.
 231          */
 232         for (i = 0; i < r0->lo_nr; ++i) {
 233                 struct cl_device *subdev;
 234                 struct lov_oinfo *oinfo = lse->lsme_oinfo[i];
 235                 int ost_idx = oinfo->loi_ost_idx;
 236
 237                 if (lov_oinfo_is_dummy(oinfo))
 238                         continue;
 239
 240                 result = ostid_to_fid(ofid, &oinfo->loi_oi, oinfo->loi_ost_idx);
 241                 if (result != 0)
 242                         GOTO(out, result);
 243
 244                 if (dev->ld_target[ost_idx] == NULL) {
 245                         CERROR("%s: OST %04x is not initialized\n",
 246                                lov2obd(dev->ld_lov)->obd_name, ost_idx);
 247                         GOTO(out, result = -EIO);
 248                 }
 249
 250                 subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
 251                 subconf->u.coc_oinfo = oinfo;
 252                 LASSERTF(subdev != NULL, "not init ost %d\n", ost_idx);
 253                 /* In the function below, .hs_keycmp resolves to
 254                  * lu_obj_hop_keycmp() */
 255                 /* coverity[overrun-buffer-val] */
 256                 stripe = lov_sub_find(env, subdev, ofid, subconf);
 257                 if (IS_ERR(stripe))
 258                         GOTO(out, result = PTR_ERR(stripe));
 259
 260                 result = lov_init_sub(env, lov, stripe, oinfo,
 261                                       lov_comp_index(index, i));
 262                 if (result == -EAGAIN) { /* try again */
 263                         --i;
 264                         result = 0;
 265                         continue;
 266                 }
 267
 268                 if (result == 0) {
 269                         r0->lo_sub[i] = cl2lovsub(stripe);
 270
 271                         sz = lov_page_slice_fixup(lov, stripe);
 272                         LASSERT(ergo(psz > 0, psz == sz));
 273                         psz = sz;
 274                 }
 275         }
 276         if (result == 0)
 277                 result = psz;
 278 out:
 279         RETURN(result);
 280 }
 281
 282 static void lov_subobject_kill(const struct lu_env *env, struct lov_object *lov,
 283                                struct lov_layout_raid0 *r0,
 284                                struct lovsub_object *los, int idx)
 285 {
 286         struct cl_object        *sub;
 287         struct lu_site          *site;
 288         wait_queue_head_t *wq;
 289
 290         LASSERT(r0->lo_sub[idx] == los);
 291
 292         sub = lovsub2cl(los);
 293         site = sub->co_lu.lo_dev->ld_site;
 294         wq = lu_site_wq_from_fid(site, &sub->co_lu.lo_header->loh_fid);
 295
 296         cl_object_kill(env, sub);
 297         /* release a reference to the sub-object and ... */
 298         lu_object_ref_del(&sub->co_lu, "lov-parent", lov);
 299         cl_object_put(env, sub);
 300
 301         /* ... wait until it is actually destroyed---sub-object clears its
 302          * ->lo_sub[] slot in lovsub_object_free() */
 303         wait_event(*wq, r0->lo_sub[idx] != los);
 304         LASSERT(r0->lo_sub[idx] == NULL);
 305 }
 306
 307 static void lov_delete_raid0(const struct lu_env *env, struct lov_object *lov,
 308                              struct lov_layout_entry *lle)
 309 {
 310         struct lov_layout_raid0 *r0 = &lle->lle_raid0;
 311
 312         ENTRY;
 313
 314         if (r0->lo_sub != NULL) {
 315                 int i;
 316
 317                 for (i = 0; i < r0->lo_nr; ++i) {
 318                         struct lovsub_object *los = r0->lo_sub[i];
 319
 320                         if (los != NULL) {
 321                                 cl_object_prune(env, &los->lso_cl);
 322                                 /*
 323                                  * If top-level object is to be evicted from
 324                                  * the cache, so are its sub-objects.
 325                                  */
 326                                 lov_subobject_kill(env, lov, r0, los, i);
 327                         }
 328                 }
 329         }
 330
 331         EXIT;
 332 }
 333
 334 static void lov_fini_raid0(const struct lu_env *env,
 335                            struct lov_layout_entry *lle)
 336 {
 337         struct lov_layout_raid0 *r0 = &lle->lle_raid0;
 338
 339         if (r0->lo_sub != NULL) {
 340                 OBD_FREE_PTR_ARRAY_LARGE(r0->lo_sub, r0->lo_nr);
 341                 r0->lo_sub = NULL;
 342         }
 343 }
 344
 345 static int lov_print_raid0(const struct lu_env *env, void *cookie,
 346                            lu_printer_t p, const struct lov_layout_entry *lle)
 347 {
 348         const struct lov_layout_raid0 *r0 = &lle->lle_raid0;
 349         int i;
 350
 351         for (i = 0; i < r0->lo_nr; ++i) {
 352                 struct lu_object *sub;
 353
 354                 if (r0->lo_sub[i] != NULL) {
 355                         sub = lovsub2lu(r0->lo_sub[i]);
 356                         lu_object_print(env, cookie, p, sub);
 357                 } else {
 358                         (*p)(env, cookie, "sub %d absent\n", i);
 359                 }
 360         }
 361         return 0;
 362 }
 363
 364 static int lov_attr_get_raid0(const struct lu_env *env, struct lov_object *lov,
 365                               unsigned int index, struct lov_layout_entry *lle,
 366                               struct cl_attr **lov_attr)
 367 {
 368         struct lov_layout_raid0 *r0 = &lle->lle_raid0;
 369         struct lov_stripe_md *lsm = lov->lo_lsm;
 370         struct ost_lvb *lvb = &lov_env_info(env)->lti_lvb;
 371         struct cl_attr *attr = &r0->lo_attr;
 372         __u64 kms = 0;
 373         int result = 0;
 374
 375         if (r0->lo_attr_valid) {
 376                 *lov_attr = attr;
 377                 return 0;
 378         }
 379
 380         memset(lvb, 0, sizeof(*lvb));
 381
 382         /* XXX: timestamps can be negative by sanity:test_39m,
 383          * how can it be? */
 384         lvb->lvb_atime = LLONG_MIN;
 385         lvb->lvb_ctime = LLONG_MIN;
 386         lvb->lvb_mtime = LLONG_MIN;
 387
 388         /*
 389          * XXX that should be replaced with a loop over sub-objects,
 390          * doing cl_object_attr_get() on them. But for now, let's
 391          * reuse old lov code.
 392          */
 393
 394         /*
 395          * XXX take lsm spin-lock to keep lov_merge_lvb_kms()
 396          * happy. It's not needed, because new code uses
 397          * ->coh_attr_guard spin-lock to protect consistency of
 398          * sub-object attributes.
 399          */
 400         lov_stripe_lock(lsm);
 401         result = lov_merge_lvb_kms(lsm, index, lvb, &kms);
 402         lov_stripe_unlock(lsm);
 403         if (result == 0) {
 404                 cl_lvb2attr(attr, lvb);
 405                 attr->cat_kms = kms;
 406                 r0->lo_attr_valid = 1;
 407                 *lov_attr = attr;
 408         }
 409
 410         return result;
 411 }
 412
 413 static struct lov_comp_layout_entry_ops raid0_ops = {
 414         .lco_init      = lov_init_raid0,
 415         .lco_fini      = lov_fini_raid0,
 416         .lco_getattr   = lov_attr_get_raid0,
 417 };
 418
 419 static int lov_attr_get_dom(const struct lu_env *env, struct lov_object *lov,
 420                             unsigned int index, struct lov_layout_entry *lle,
 421                             struct cl_attr **lov_attr)
 422 {
 423         struct lov_layout_dom *dom = &lle->lle_dom;
 424         struct lov_oinfo *loi = dom->lo_loi;
 425         struct cl_attr *attr = &dom->lo_dom_r0.lo_attr;
 426
 427         if (dom->lo_dom_r0.lo_attr_valid) {
 428                 *lov_attr = attr;
 429                 return 0;
 430         }
 431
 432         if (OST_LVB_IS_ERR(loi->loi_lvb.lvb_blocks))
 433                 return OST_LVB_GET_ERR(loi->loi_lvb.lvb_blocks);
 434
 435         cl_lvb2attr(attr, &loi->loi_lvb);
 436
 437         /* DoM component size can be bigger than stripe size after
 438          * client's setattr RPC, so do not count anything beyond
 439          * component end. Alternatively, check that limit on server
 440          * and do not allow size overflow there. */
 441         if (attr->cat_size > lle->lle_extent->e_end)
 442                 attr->cat_size = lle->lle_extent->e_end;
 443
 444         attr->cat_kms = attr->cat_size;
 445
 446         dom->lo_dom_r0.lo_attr_valid = 1;
 447         *lov_attr = attr;
 448
 449         return 0;
 450 }
 451
 452 /**
 453  * Lookup FLD to get MDS index of the given DOM object FID.
 454  *
 455  * \param[in]  ld       LOV device
 456  * \param[in]  fid      FID to lookup
 457  * \param[out] nr       index in MDC array to return back
 458  *
 459  * \retval              0 and \a mds filled with MDS index if successful
 460  * \retval              negative value on error
 461  */
 462 static int lov_fld_lookup(struct lov_device *ld, const struct lu_fid *fid,
 463                           __u32 *nr)
 464 {
 465         __u32 mds_idx;
 466         int i, rc;
 467
 468         ENTRY;
 469
 470         rc = fld_client_lookup(&ld->ld_lmv->u.lmv.lmv_fld, fid_seq(fid),
 471                                &mds_idx, LU_SEQ_RANGE_MDT, NULL);
 472         if (rc) {
 473                 CERROR("%s: error while looking for mds number. Seq %#llx"
 474                        ", err = %d\n", lu_dev_name(cl2lu_dev(&ld->ld_cl)),
 475                        fid_seq(fid), rc);
 476                 RETURN(rc);
 477         }
 478
 479         CDEBUG(D_INODE, "FLD lookup got mds #%x for fid="DFID"\n",
 480                mds_idx, PFID(fid));
 481
 482         /* find proper MDC device in the array */
 483         for (i = 0; i < ld->ld_md_tgts_nr; i++) {
 484                 if (ld->ld_md_tgts[i].ldm_mdc != NULL &&
 485                     ld->ld_md_tgts[i].ldm_idx == mds_idx)
 486                         break;
 487         }
 488
 489         if (i == ld->ld_md_tgts_nr) {
 490                 CERROR("%s: cannot find corresponding MDC device for mds #%x "
 491                        "for fid="DFID"\n", lu_dev_name(cl2lu_dev(&ld->ld_cl)),
 492                        mds_idx, PFID(fid));
 493                 rc = -EINVAL;
 494         } else {
 495                 *nr = i;
 496         }
 497         RETURN(rc);
 498 }
 499
 500 /**
 501  * Implementation of lov_comp_layout_entry_ops::lco_init for DOM object.
 502  *
 503  * Init the DOM object for the first time. It prepares also RAID0 entry
 504  * for it to use in common methods with ordinary RAID0 layout entries.
 505  *
 506  * \param[in] env       execution environment
 507  * \param[in] dev       LOV device
 508  * \param[in] lov       LOV object
 509  * \param[in] index     Composite layout entry index in LSM
 510  * \param[in] lle       Composite LOV layout entry
 511  */
 512 static int lov_init_dom(const struct lu_env *env, struct lov_device *dev,
 513                         struct lov_object *lov, unsigned int index,
 514                         const struct cl_object_conf *conf,
 515                         struct lov_layout_entry *lle)
 516 {
 517         struct lov_thread_info *lti = lov_env_info(env);
 518         struct lov_stripe_md_entry *lsme = lov_lse(lov, index);
 519         struct cl_object *clo;
 520         struct lu_object *o = lov2lu(lov);
 521         const struct lu_fid *fid = lu_object_fid(o);
 522         struct cl_device *mdcdev;
 523         struct lov_oinfo *loi = NULL;
 524         struct cl_object_conf *sconf = &lti->lti_stripe_conf;
 525         int rc;
 526         __u32 idx = 0;
 527
 528         ENTRY;
 529
 530         /* DOM entry may be not zero index due to FLR but must start from 0 */
 531         if (unlikely(lle->lle_extent->e_start != 0)) {
 532                 CERROR("%s: DOM entry must be the first stripe in a mirror\n",
 533                        lov2obd(dev->ld_lov)->obd_name);
 534                 dump_lsm(D_ERROR, lov->lo_lsm);
 535                 RETURN(-EINVAL);
 536         }
 537
 538         /* find proper MDS device */
 539         rc = lov_fld_lookup(dev, fid, &idx);
 540         if (rc)
 541                 RETURN(rc);
 542
 543         LASSERTF(dev->ld_md_tgts[idx].ldm_mdc != NULL,
 544                  "LOV md target[%u] is NULL\n", idx);
 545
 546         /* check lsm is DOM, more checks are needed */
 547         LASSERT(lsme->lsme_stripe_count == 0);
 548
 549         /*
 550          * Create lower cl_objects.
 551          */
 552         mdcdev = dev->ld_md_tgts[idx].ldm_mdc;
 553
 554         LASSERTF(mdcdev != NULL, "non-initialized mdc subdev\n");
 555
 556         /* DoM object has no oinfo in LSM entry, create it exclusively */
 557         OBD_SLAB_ALLOC_PTR_GFP(loi, lov_oinfo_slab, GFP_NOFS);
 558         if (loi == NULL)
 559                 RETURN(-ENOMEM);
 560
 561         fid_to_ostid(lu_object_fid(lov2lu(lov)), &loi->loi_oi);
 562
 563         sconf->u.coc_oinfo = loi;
 564 again:
 565         clo = lov_sub_find(env, mdcdev, fid, sconf);
 566         if (IS_ERR(clo))
 567                 GOTO(out, rc = PTR_ERR(clo));
 568
 569         rc = lov_init_sub(env, lov, clo, loi, lov_comp_index(index, 0));
 570         if (rc == -EAGAIN) /* try again */
 571                 goto again;
 572         else if (rc != 0)
 573                 GOTO(out, rc);
 574
 575         lle->lle_dom.lo_dom = cl2lovsub(clo);
 576         spin_lock_init(&lle->lle_dom.lo_dom_r0.lo_sub_lock);
 577         lle->lle_dom.lo_dom_r0.lo_nr = 1;
 578         lle->lle_dom.lo_dom_r0.lo_sub = &lle->lle_dom.lo_dom;
 579         lle->lle_dom.lo_loi = loi;
 580
 581         rc = lov_page_slice_fixup(lov, clo);
 582         RETURN(rc);
 583
 584 out:
 585         if (loi != NULL)
 586                 OBD_SLAB_FREE_PTR(loi, lov_oinfo_slab);
 587         return rc;
 588 }
 589
 590 /**
 591  * Implementation of lov_layout_operations::llo_fini for DOM object.
 592  *
 593  * Finish the DOM object and free related memory.
 594  *
 595  * \param[in] env       execution environment
 596  * \param[in] lov       LOV object
 597  * \param[in] state     LOV layout state
 598  */
 599 static void lov_fini_dom(const struct lu_env *env,
 600                          struct lov_layout_entry *lle)
 601 {
 602         if (lle->lle_dom.lo_dom != NULL)
 603                 lle->lle_dom.lo_dom = NULL;
 604         if (lle->lle_dom.lo_loi != NULL)
 605                 OBD_SLAB_FREE_PTR(lle->lle_dom.lo_loi, lov_oinfo_slab);
 606 }
 607
 608 static struct lov_comp_layout_entry_ops dom_ops = {
 609         .lco_init = lov_init_dom,
 610         .lco_fini = lov_fini_dom,
 611         .lco_getattr = lov_attr_get_dom,
 612 };
 613
 614 static int lov_init_composite(const struct lu_env *env, struct lov_device *dev,
 615                               struct lov_object *lov, struct lov_stripe_md *lsm,
 616                               const struct cl_object_conf *conf,
 617                               union lov_layout_state *state)
 618 {
 619         struct lov_layout_composite *comp = &state->composite;
 620         struct lov_layout_entry *lle;
 621         struct lov_mirror_entry *lre;
 622         unsigned int entry_count;
 623         unsigned int psz = 0;
 624         unsigned int mirror_count;
 625         int flr_state = lsm->lsm_flags & LCM_FL_FLR_MASK;
 626         int result = 0;
 627         unsigned int seq;
 628         int i, j;
 629         bool dom_size = 0;
 630
 631         ENTRY;
 632
 633         LASSERT(lsm->lsm_entry_count > 0);
 634         LASSERT(lov->lo_lsm == NULL);
 635         lov->lo_lsm = lsm_addref(lsm);
 636         lov->lo_layout_invalid = true;
 637
 638         dump_lsm(D_INODE, lsm);
 639
 640         entry_count = lsm->lsm_entry_count;
 641
 642         spin_lock_init(&comp->lo_write_lock);
 643         comp->lo_flags = lsm->lsm_flags;
 644         comp->lo_mirror_count = lsm->lsm_mirror_count + 1;
 645         comp->lo_entry_count = lsm->lsm_entry_count;
 646         comp->lo_preferred_mirror = -1;
 647
 648         if (equi(flr_state == LCM_FL_NONE, comp->lo_mirror_count > 1))
 649                 RETURN(-EINVAL);
 650
 651         OBD_ALLOC_PTR_ARRAY(comp->lo_mirrors, comp->lo_mirror_count);
 652         if (comp->lo_mirrors == NULL)
 653                 RETURN(-ENOMEM);
 654
 655         OBD_ALLOC_PTR_ARRAY(comp->lo_entries, entry_count);
 656         if (comp->lo_entries == NULL)
 657                 RETURN(-ENOMEM);
 658
 659         /* Initiate all entry types and extents data at first */
 660         for (i = 0, j = 0, mirror_count = 1; i < entry_count; i++) {
 661                 int mirror_id = 0;
 662
 663                 lle = &comp->lo_entries[i];
 664
 665                 lle->lle_lsme = lsm->lsm_entries[i];
 666                 lle->lle_type = lov_entry_type(lle->lle_lsme);
 667                 switch (lle->lle_type) {
 668                 case LOV_PATTERN_RAID0:
 669                         lle->lle_comp_ops = &raid0_ops;
 670                         break;
 671                 case LOV_PATTERN_MDT:
 672                         /* Allowed to have several DOM stripes in different
 673                          * mirrors with the same DoM size.
 674                          */
 675                         if (!dom_size) {
 676                                 dom_size = lle->lle_lsme->lsme_extent.e_end;
 677                         } else if (dom_size !=
 678                                    lle->lle_lsme->lsme_extent.e_end) {
 679                                 CERROR("%s: DOM entries with different sizes\n",
 680                                        lov2obd(dev->ld_lov)->obd_name);
 681                                 dump_lsm(D_ERROR, lsm);
 682                                 RETURN(-EINVAL);
 683                         }
 684                         lle->lle_comp_ops = &dom_ops;
 685                         break;
 686                 case LOV_PATTERN_FOREIGN:
 687                         lle->lle_comp_ops = NULL;
 688                         break;
 689                 default:
 690                         CERROR("%s: unknown composite layout entry type %i\n",
 691                                lov2obd(dev->ld_lov)->obd_name,
 692                                lsm->lsm_entries[i]->lsme_pattern);
 693                         dump_lsm(D_ERROR, lsm);
 694                         RETURN(-EIO);
 695                 }
 696
 697                 lle->lle_extent = &lle->lle_lsme->lsme_extent;
 698                 lle->lle_valid = !(lle->lle_lsme->lsme_flags & LCME_FL_STALE);
 699
 700                 if (flr_state != LCM_FL_NONE)
 701                         mirror_id = mirror_id_of(lle->lle_lsme->lsme_id);
 702
 703                 lre = &comp->lo_mirrors[j];
 704                 if (i > 0) {
 705                         if (mirror_id == lre->lre_mirror_id) {
 706                                 lre->lre_valid |= lle->lle_valid;
 707                                 lre->lre_stale |= !lle->lle_valid;
 708                                 lre->lre_foreign |=
 709                                         lsme_is_foreign(lle->lle_lsme);
 710                                 lre->lre_end = i;
 711                                 continue;
 712                         }
 713
 714                         /* new mirror detected, assume that the mirrors
 715                          * are shorted in layout */
 716                         ++mirror_count;
 717                         ++j;
 718                         if (j >= comp->lo_mirror_count)
 719                                 break;
 720
 721                         lre = &comp->lo_mirrors[j];
 722                 }
 723
 724                 /* entries must be sorted by mirrors */
 725                 lre->lre_mirror_id = mirror_id;
 726                 lre->lre_start = lre->lre_end = i;
 727                 lre->lre_preferred = !!(lle->lle_lsme->lsme_flags &
 728                                         LCME_FL_PREF_RD);
 729                 lre->lre_valid = lle->lle_valid;
 730                 lre->lre_stale = !lle->lle_valid;
 731                 lre->lre_foreign = lsme_is_foreign(lle->lle_lsme);
 732         }
 733
 734         /* sanity check for FLR */
 735         if (mirror_count != comp->lo_mirror_count) {
 736                 CDEBUG(D_INODE, DFID
 737                        " doesn't have the # of mirrors it claims, %u/%u\n",
 738                        PFID(lu_object_fid(lov2lu(lov))), mirror_count,
 739                        comp->lo_mirror_count + 1);
 740
 741                 GOTO(out, result = -EINVAL);
 742         }
 743
 744         lov_foreach_layout_entry(lov, lle) {
 745                 int index = lov_layout_entry_index(lov, lle);
 746
 747                 /**
 748                  * If the component has not been init-ed on MDS side, for
 749                  * PFL layout, we'd know that the components beyond this one
 750                  * will be dynamically init-ed later on file write/trunc ops.
 751                  */
 752                 if (!lsme_inited(lle->lle_lsme))
 753                         continue;
 754
 755                 if (lsme_is_foreign(lle->lle_lsme))
 756                         continue;
 757
 758                 result = lle->lle_comp_ops->lco_init(env, dev, lov, index,
 759                                                      conf, lle);
 760                 if (result < 0)
 761                         break;
 762
 763                 LASSERT(ergo(psz > 0, psz == result));
 764                 psz = result;
 765         }
 766
 767         if (psz > 0)
 768                 cl_object_header(&lov->lo_cl)->coh_page_bufsize += psz;
 769
 770         /* decide the preferred mirror. It uses the hash value of lov_object
 771          * so that different clients would use different mirrors for read. */
 772         mirror_count = 0;
 773         seq = hash_long((unsigned long)lov, 8);
 774         for (i = 0; i < comp->lo_mirror_count; i++) {
 775                 unsigned int idx = (i + seq) % comp->lo_mirror_count;
 776
 777                 lre = lov_mirror_entry(lov, idx);
 778                 if (lre->lre_stale)
 779                         continue;
 780
 781                 if (lre->lre_foreign)
 782                         continue;
 783
 784                 mirror_count++; /* valid mirror */
 785
 786                 if (lre->lre_preferred || comp->lo_preferred_mirror < 0)
 787                         comp->lo_preferred_mirror = idx;
 788         }
 789         if (!mirror_count) {
 790                 CDEBUG(D_INODE, DFID
 791                        " doesn't have any valid mirrors\n",
 792                        PFID(lu_object_fid(lov2lu(lov))));
 793
 794                 comp->lo_preferred_mirror = 0;
 795         }
 796
 797         LASSERT(comp->lo_preferred_mirror >= 0);
 798
 799         EXIT;
 800 out:
 801         return result > 0 ? 0 : result;
 802 }
 803
 804 static int lov_init_empty(const struct lu_env *env, struct lov_device *dev,
 805                           struct lov_object *lov, struct lov_stripe_md *lsm,
 806                           const struct cl_object_conf *conf,
 807                           union lov_layout_state *state)
 808 {
 809         return 0;
 810 }
 811
 812 static int lov_init_released(const struct lu_env *env,
 813                              struct lov_device *dev, struct lov_object *lov,
 814                              struct lov_stripe_md *lsm,
 815                              const struct cl_object_conf *conf,
 816                              union lov_layout_state *state)
 817 {
 818         LASSERT(lsm != NULL);
 819         LASSERT(lsm->lsm_is_released);
 820         LASSERT(lov->lo_lsm == NULL);
 821
 822         lov->lo_lsm = lsm_addref(lsm);
 823         return 0;
 824 }
 825
 826 static int lov_init_foreign(const struct lu_env *env,
 827                             struct lov_device *dev, struct lov_object *lov,
 828                             struct lov_stripe_md *lsm,
 829                             const struct cl_object_conf *conf,
 830                             union lov_layout_state *state)
 831 {
 832         LASSERT(lsm != NULL);
 833         LASSERT(lov->lo_type == LLT_FOREIGN);
 834         LASSERT(lov->lo_lsm == NULL);
 835
 836         lov->lo_lsm = lsm_addref(lsm);
 837         return 0;
 838 }
 839
 840 static int lov_delete_empty(const struct lu_env *env, struct lov_object *lov,
 841                             union lov_layout_state *state)
 842 {
 843         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED ||
 844                 lov->lo_type == LLT_FOREIGN);
 845
 846         lov_layout_wait(env, lov);
 847         return 0;
 848 }
 849
 850 static int lov_delete_composite(const struct lu_env *env,
 851                                 struct lov_object *lov,
 852                                 union lov_layout_state *state)
 853 {
 854         struct lov_layout_entry *entry;
 855         struct lov_layout_composite *comp = &state->composite;
 856
 857         ENTRY;
 858
 859         dump_lsm(D_INODE, lov->lo_lsm);
 860
 861         lov_layout_wait(env, lov);
 862         if (comp->lo_entries)
 863                 lov_foreach_layout_entry(lov, entry) {
 864                         if (entry->lle_lsme && lsme_is_foreign(entry->lle_lsme))
 865                                 continue;
 866
 867                         lov_delete_raid0(env, lov, entry);
 868         }
 869
 870         RETURN(0);
 871 }
 872
 873 static void lov_fini_empty(const struct lu_env *env, struct lov_object *lov,
 874                            union lov_layout_state *state)
 875 {
 876         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 877 }
 878
 879 static void lov_fini_composite(const struct lu_env *env,
 880                                struct lov_object *lov,
 881                                union lov_layout_state *state)
 882 {
 883         struct lov_layout_composite *comp = &state->composite;
 884         ENTRY;
 885
 886         if (comp->lo_entries != NULL) {
 887                 struct lov_layout_entry *entry;
 888
 889                 lov_foreach_layout_entry(lov, entry)
 890                         if (entry->lle_comp_ops)
 891                                 entry->lle_comp_ops->lco_fini(env, entry);
 892
 893                 OBD_FREE_PTR_ARRAY(comp->lo_entries, comp->lo_entry_count);
 894                 comp->lo_entries = NULL;
 895         }
 896
 897         if (comp->lo_mirrors != NULL) {
 898                 OBD_FREE_PTR_ARRAY(comp->lo_mirrors, comp->lo_mirror_count);
 899                 comp->lo_mirrors = NULL;
 900         }
 901
 902         memset(comp, 0, sizeof(*comp));
 903
 904         dump_lsm(D_INODE, lov->lo_lsm);
 905         lov_free_memmd(&lov->lo_lsm);
 906
 907         EXIT;
 908 }
 909
 910 static void lov_fini_released(const struct lu_env *env, struct lov_object *lov,
 911                                 union lov_layout_state *state)
 912 {
 913         ENTRY;
 914         dump_lsm(D_INODE, lov->lo_lsm);
 915         lov_free_memmd(&lov->lo_lsm);
 916         EXIT;
 917 }
 918
 919 static int lov_print_empty(const struct lu_env *env, void *cookie,
 920                            lu_printer_t p, const struct lu_object *o)
 921 {
 922         (*p)(env, cookie, "empty %d\n", lu2lov(o)->lo_layout_invalid);
 923         return 0;
 924 }
 925
 926 static int lov_print_composite(const struct lu_env *env, void *cookie,
 927                                lu_printer_t p, const struct lu_object *o)
 928 {
 929         struct lov_object *lov = lu2lov(o);
 930         struct lov_stripe_md *lsm = lov->lo_lsm;
 931         int i;
 932
 933         (*p)(env, cookie, "entries: %d, %s, lsm{%p 0x%08X %d %u}:\n",
 934              lsm->lsm_entry_count,
 935              lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 936              lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 937              lsm->lsm_layout_gen);
 938
 939         for (i = 0; i < lsm->lsm_entry_count; i++) {
 940                 struct lov_stripe_md_entry *lse = lsm->lsm_entries[i];
 941                 struct lov_layout_entry *lle = lov_entry(lov, i);
 942
 943                 (*p)(env, cookie,
 944                      DEXT ": { 0x%08X, %u, %#x, %u, %#x, %u, %u }\n",
 945                      PEXT(&lse->lsme_extent), lse->lsme_magic,
 946                      lse->lsme_id, lse->lsme_pattern, lse->lsme_layout_gen,
 947                      lse->lsme_flags, lse->lsme_stripe_count,
 948                      lse->lsme_stripe_size);
 949
 950                 if (!lsme_is_foreign(lse))
 951                         lov_print_raid0(env, cookie, p, lle);
 952         }
 953
 954         return 0;
 955 }
 956
 957 static int lov_print_released(const struct lu_env *env, void *cookie,
 958                                 lu_printer_t p, const struct lu_object *o)
 959 {
 960         struct lov_object       *lov = lu2lov(o);
 961         struct lov_stripe_md    *lsm = lov->lo_lsm;
 962
 963         (*p)(env, cookie,
 964                 "released: %s, lsm{%p 0x%08X %d %u}:\n",
 965                 lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 966                 lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 967                 lsm->lsm_layout_gen);
 968         return 0;
 969 }
 970
 971 static int lov_print_foreign(const struct lu_env *env, void *cookie,
 972                                 lu_printer_t p, const struct lu_object *o)
 973 {
 974         struct lov_object       *lov = lu2lov(o);
 975         struct lov_stripe_md    *lsm = lov->lo_lsm;
 976
 977         (*p)(env, cookie,
 978                 "foreign: %s, lsm{%p 0x%08X %d %u}:\n",
 979                 lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 980                 lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 981                 lsm->lsm_layout_gen);
 982         (*p)(env, cookie,
 983                 "raw_ea_content '%.*s'\n",
 984                 (int)lsm->lsm_foreign_size, (char *)lsm_foreign(lsm));
 985         return 0;
 986 }
 987
 988 /**
 989  * Implements cl_object_operations::coo_attr_get() method for an object
 990  * without stripes (LLT_EMPTY layout type).
 991  *
 992  * The only attributes this layer is authoritative in this case is
 993  * cl_attr::cat_blocks---it's 0.
 994  */
 995 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
 996                               struct cl_attr *attr)
 997 {
 998         attr->cat_blocks = 0;
 999         return 0;
1000 }
1001
1002 static int lov_attr_get_composite(const struct lu_env *env,
1003                                   struct cl_object *obj,
1004                                   struct cl_attr *attr)
1005 {
1006         struct lov_object       *lov = cl2lov(obj);
1007         struct lov_layout_entry *entry;
1008         int                      result = 0;
1009
1010         ENTRY;
1011
1012         attr->cat_size = 0;
1013         attr->cat_blocks = 0;
1014         lov_foreach_layout_entry(lov, entry) {
1015                 struct cl_attr *lov_attr = NULL;
1016                 int index = lov_layout_entry_index(lov, entry);
1017
1018                 if (!entry->lle_valid)
1019                         continue;
1020
1021                 /* PFL: This component has not been init-ed. */
1022                 if (!lsm_entry_inited(lov->lo_lsm, index))
1023                         continue;
1024
1025                 result = entry->lle_comp_ops->lco_getattr(env, lov, index,
1026                                                           entry, &lov_attr);
1027                 if (result < 0)
1028                         RETURN(result);
1029
1030                 if (lov_attr == NULL)
1031                         continue;
1032
1033                 CDEBUG(D_INODE, "COMP ID #%i: s=%llu m=%llu a=%llu c=%llu "
1034                        "b=%llu\n", index - 1, lov_attr->cat_size,
1035                        lov_attr->cat_mtime, lov_attr->cat_atime,
1036                        lov_attr->cat_ctime, lov_attr->cat_blocks);
1037
1038                 /* merge results */
1039                 attr->cat_blocks += lov_attr->cat_blocks;
1040                 if (attr->cat_size < lov_attr->cat_size)
1041                         attr->cat_size = lov_attr->cat_size;
1042                 if (attr->cat_kms < lov_attr->cat_kms)
1043                         attr->cat_kms = lov_attr->cat_kms;
1044                 if (attr->cat_atime < lov_attr->cat_atime)
1045                         attr->cat_atime = lov_attr->cat_atime;
1046                 if (attr->cat_ctime < lov_attr->cat_ctime)
1047                         attr->cat_ctime = lov_attr->cat_ctime;
1048                 if (attr->cat_mtime < lov_attr->cat_mtime)
1049                         attr->cat_mtime = lov_attr->cat_mtime;
1050         }
1051
1052         RETURN(0);
1053 }
1054
1055 static int lov_flush_composite(const struct lu_env *env,
1056                                struct cl_object *obj,
1057                                struct ldlm_lock *lock)
1058 {
1059         struct lov_object *lov = cl2lov(obj);
1060         struct lov_layout_entry *lle;
1061         int rc = -ENODATA;
1062
1063         ENTRY;
1064
1065         lov_foreach_layout_entry(lov, lle) {
1066                 if (!lsme_is_dom(lle->lle_lsme))
1067                         continue;
1068                 rc = cl_object_flush(env, lovsub2cl(lle->lle_dom.lo_dom), lock);
1069                 break;
1070         }
1071
1072         RETURN(rc);
1073 }
1074
1075 static int lov_flush_empty(const struct lu_env *env, struct cl_object *obj,
1076                            struct ldlm_lock *lock)
1077 {
1078         return 0;
1079 }
1080
1081 const static struct lov_layout_operations lov_dispatch[] = {
1082         [LLT_EMPTY] = {
1083                 .llo_init      = lov_init_empty,
1084                 .llo_delete    = lov_delete_empty,
1085                 .llo_fini      = lov_fini_empty,
1086                 .llo_print     = lov_print_empty,
1087                 .llo_page_init = lov_page_init_empty,
1088                 .llo_lock_init = lov_lock_init_empty,
1089                 .llo_io_init   = lov_io_init_empty,
1090                 .llo_getattr   = lov_attr_get_empty,
1091                 .llo_flush     = lov_flush_empty,
1092         },
1093         [LLT_RELEASED] = {
1094                 .llo_init      = lov_init_released,
1095                 .llo_delete    = lov_delete_empty,
1096                 .llo_fini      = lov_fini_released,
1097                 .llo_print     = lov_print_released,
1098                 .llo_page_init = lov_page_init_empty,
1099                 .llo_lock_init = lov_lock_init_empty,
1100                 .llo_io_init   = lov_io_init_released,
1101                 .llo_getattr   = lov_attr_get_empty,
1102                 .llo_flush     = lov_flush_empty,
1103         },
1104         [LLT_COMP] = {
1105                 .llo_init      = lov_init_composite,
1106                 .llo_delete    = lov_delete_composite,
1107                 .llo_fini      = lov_fini_composite,
1108                 .llo_print     = lov_print_composite,
1109                 .llo_page_init = lov_page_init_composite,
1110                 .llo_lock_init = lov_lock_init_composite,
1111                 .llo_io_init   = lov_io_init_composite,
1112                 .llo_getattr   = lov_attr_get_composite,
1113                 .llo_flush     = lov_flush_composite,
1114         },
1115         [LLT_FOREIGN] = {
1116                 .llo_init      = lov_init_foreign,
1117                 .llo_delete    = lov_delete_empty,
1118                 .llo_fini      = lov_fini_released,
1119                 .llo_print     = lov_print_foreign,
1120                 .llo_page_init = lov_page_init_foreign,
1121                 .llo_lock_init = lov_lock_init_empty,
1122                 .llo_io_init   = lov_io_init_empty,
1123                 .llo_getattr   = lov_attr_get_empty,
1124                 .llo_flush     = lov_flush_empty,
1125         },
1126 };
1127
1128 /**
1129  * Performs a double-dispatch based on the layout type of an object.
1130  */
1131 #define LOV_2DISPATCH_NOLOCK(obj, op, ...)              \
1132 ({                                                      \
1133         struct lov_object *__obj = (obj);               \
1134         enum lov_layout_type __llt;                     \
1135                                                         \
1136         __llt = __obj->lo_type;                         \
1137         LASSERT(__llt < ARRAY_SIZE(lov_dispatch));      \
1138         lov_dispatch[__llt].op(__VA_ARGS__);            \
1139 })
1140
1141 /**
1142  * Return lov_layout_type associated with a given lsm
1143  */
1144 static enum lov_layout_type lov_type(struct lov_stripe_md *lsm)
1145 {
1146         if (lsm == NULL)
1147                 return LLT_EMPTY;
1148
1149         if (lsm->lsm_is_released)
1150                 return LLT_RELEASED;
1151
1152         if (lsm->lsm_magic == LOV_MAGIC_V1 ||
1153             lsm->lsm_magic == LOV_MAGIC_V3 ||
1154             lsm->lsm_magic == LOV_MAGIC_COMP_V1)
1155                 return LLT_COMP;
1156
1157         if (lsm->lsm_magic == LOV_MAGIC_FOREIGN)
1158                 return LLT_FOREIGN;
1159
1160         return LLT_EMPTY;
1161 }
1162
1163 static inline void lov_conf_freeze(struct lov_object *lov)
1164 {
1165         CDEBUG(D_INODE, "To take share lov(%p) owner %p/%p\n",
1166                 lov, lov->lo_owner, current);
1167         if (lov->lo_owner != current)
1168                 down_read(&lov->lo_type_guard);
1169 }
1170
1171 static inline void lov_conf_thaw(struct lov_object *lov)
1172 {
1173         CDEBUG(D_INODE, "To release share lov(%p) owner %p/%p\n",
1174                 lov, lov->lo_owner, current);
1175         if (lov->lo_owner != current)
1176                 up_read(&lov->lo_type_guard);
1177 }
1178
1179 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...)                       \
1180 ({                                                                      \
1181         struct lov_object                      *__obj = (obj);          \
1182         int                                     __lock = !!(lock);      \
1183         typeof(lov_dispatch[0].op(__VA_ARGS__)) __result;               \
1184                                                                         \
1185         if (__lock)                                                     \
1186                 lov_conf_freeze(__obj);                                 \
1187         __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__);          \
1188         if (__lock)                                                     \
1189                 lov_conf_thaw(__obj);                                   \
1190         __result;                                                       \
1191 })
1192
1193 /**
1194  * Performs a locked double-dispatch based on the layout type of an object.
1195  */
1196 #define LOV_2DISPATCH(obj, op, ...)                     \
1197         LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
1198
1199 #define LOV_2DISPATCH_VOID(obj, op, ...)                                \
1200 do {                                                                    \
1201         struct lov_object                      *__obj = (obj);          \
1202         enum lov_layout_type                    __llt;                  \
1203                                                                         \
1204         lov_conf_freeze(__obj);                                         \
1205         __llt = __obj->lo_type;                                         \
1206         LASSERT(__llt < ARRAY_SIZE(lov_dispatch));                      \
1207         lov_dispatch[__llt].op(__VA_ARGS__);                            \
1208         lov_conf_thaw(__obj);                                           \
1209 } while (0)
1210
1211 static void lov_conf_lock(struct lov_object *lov)
1212 {
1213         LASSERT(lov->lo_owner != current);
1214         down_write(&lov->lo_type_guard);
1215         LASSERT(lov->lo_owner == NULL);
1216         lov->lo_owner = current;
1217         CDEBUG(D_INODE, "Took exclusive lov(%p) owner %p\n",
1218                 lov, lov->lo_owner);
1219 }
1220
1221 static void lov_conf_unlock(struct lov_object *lov)
1222 {
1223         CDEBUG(D_INODE, "To release exclusive lov(%p) owner %p\n",
1224                 lov, lov->lo_owner);
1225         lov->lo_owner = NULL;
1226         up_write(&lov->lo_type_guard);
1227 }
1228
1229 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov)
1230 {
1231         ENTRY;
1232
1233         while (atomic_read(&lov->lo_active_ios) > 0) {
1234                 CDEBUG(D_INODE, "file:"DFID" wait for active IO, now: %d.\n",
1235                         PFID(lu_object_fid(lov2lu(lov))),
1236                         atomic_read(&lov->lo_active_ios));
1237
1238                 wait_event_idle(lov->lo_waitq,
1239                                 atomic_read(&lov->lo_active_ios) == 0);
1240         }
1241         RETURN(0);
1242 }
1243
1244 static int lov_layout_change(const struct lu_env *unused,
1245                              struct lov_object *lov, struct lov_stripe_md *lsm,
1246                              const struct cl_object_conf *conf)
1247 {
1248         enum lov_layout_type llt = lov_type(lsm);
1249         union lov_layout_state *state = &lov->u;
1250         const struct lov_layout_operations *old_ops;
1251         const struct lov_layout_operations *new_ops;
1252         struct lov_device *lov_dev = lov_object_dev(lov);
1253         struct lu_env *env;
1254         __u16 refcheck;
1255         int rc;
1256         ENTRY;
1257
1258         LASSERT(lov->lo_type < ARRAY_SIZE(lov_dispatch));
1259
1260         env = cl_env_get(&refcheck);
1261         if (IS_ERR(env))
1262                 RETURN(PTR_ERR(env));
1263
1264         LASSERT(llt < ARRAY_SIZE(lov_dispatch));
1265
1266         CDEBUG(D_INODE, DFID" from %s to %s\n",
1267                PFID(lu_object_fid(lov2lu(lov))),
1268                llt2str(lov->lo_type), llt2str(llt));
1269
1270         old_ops = &lov_dispatch[lov->lo_type];
1271         new_ops = &lov_dispatch[llt];
1272
1273         rc = cl_object_prune(env, &lov->lo_cl);
1274         if (rc != 0)
1275                 GOTO(out, rc);
1276
1277         rc = old_ops->llo_delete(env, lov, &lov->u);
1278         if (rc != 0)
1279                 GOTO(out, rc);
1280
1281         old_ops->llo_fini(env, lov, &lov->u);
1282
1283         LASSERT(atomic_read(&lov->lo_active_ios) == 0);
1284
1285         CDEBUG(D_INODE, DFID "Apply new layout lov %p, type %d\n",
1286                PFID(lu_object_fid(lov2lu(lov))), lov, llt);
1287
1288         /* page bufsize fixup */
1289         cl_object_header(&lov->lo_cl)->coh_page_bufsize -=
1290                 lov_page_slice_fixup(lov, NULL);
1291
1292         lov->lo_type = llt;
1293         rc = new_ops->llo_init(env, lov_dev, lov, lsm, conf, state);
1294         if (rc != 0) {
1295                 struct obd_device *obd = lov2obd(lov_dev->ld_lov);
1296
1297                 CERROR("%s: cannot apply new layout on "DFID" : rc = %d\n",
1298                        obd->obd_name, PFID(lu_object_fid(lov2lu(lov))), rc);
1299                 new_ops->llo_delete(env, lov, state);
1300                 new_ops->llo_fini(env, lov, state);
1301                 /* this file becomes an EMPTY file. */
1302                 lov->lo_type = LLT_EMPTY;
1303                 GOTO(out, rc);
1304         }
1305
1306 out:
1307         cl_env_put(env, &refcheck);
1308         RETURN(rc);
1309 }
1310
1311 /*****************************************************************************
1312  *
1313  * Lov object operations.
1314  *
1315  */
1316 static int lov_object_init(const struct lu_env *env, struct lu_object *obj,
1317                            const struct lu_object_conf *conf)
1318 {
1319         struct lov_object            *lov   = lu2lov(obj);
1320         struct lov_device            *dev   = lov_object_dev(lov);
1321         const struct cl_object_conf  *cconf = lu2cl_conf(conf);
1322         union lov_layout_state       *set   = &lov->u;
1323         const struct lov_layout_operations *ops;
1324         struct lov_stripe_md *lsm = NULL;
1325         int rc;
1326         ENTRY;
1327
1328         init_rwsem(&lov->lo_type_guard);
1329         atomic_set(&lov->lo_active_ios, 0);
1330         init_waitqueue_head(&lov->lo_waitq);
1331         cl_object_page_init(lu2cl(obj), sizeof(struct lov_page));
1332
1333         lov->lo_type = LLT_EMPTY;
1334         if (cconf->u.coc_layout.lb_buf != NULL) {
1335                 lsm = lov_unpackmd(dev->ld_lov,
1336                                    cconf->u.coc_layout.lb_buf,
1337                                    cconf->u.coc_layout.lb_len);
1338                 if (IS_ERR(lsm))
1339                         RETURN(PTR_ERR(lsm));
1340
1341                 dump_lsm(D_INODE, lsm);
1342         }
1343
1344         /* no locking is necessary, as object is being created */
1345         lov->lo_type = lov_type(lsm);
1346         ops = &lov_dispatch[lov->lo_type];
1347         rc = ops->llo_init(env, dev, lov, lsm, cconf, set);
1348         if (rc != 0)
1349                 GOTO(out_lsm, rc);
1350
1351 out_lsm:
1352         lov_lsm_put(lsm);
1353
1354         RETURN(rc);
1355 }
1356
1357 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
1358                         const struct cl_object_conf *conf)
1359 {
1360         struct lov_stripe_md    *lsm = NULL;
1361         struct lov_object       *lov = cl2lov(obj);
1362         int                      result = 0;
1363         ENTRY;
1364
1365         if (conf->coc_opc == OBJECT_CONF_SET &&
1366             conf->u.coc_layout.lb_buf != NULL) {
1367                 lsm = lov_unpackmd(lov_object_dev(lov)->ld_lov,
1368                                    conf->u.coc_layout.lb_buf,
1369                                    conf->u.coc_layout.lb_len);
1370                 if (IS_ERR(lsm))
1371                         RETURN(PTR_ERR(lsm));
1372                 dump_lsm(D_INODE, lsm);
1373         }
1374
1375         lov_conf_lock(lov);
1376         if (conf->coc_opc == OBJECT_CONF_INVALIDATE) {
1377                 lov->lo_layout_invalid = true;
1378                 GOTO(out, result = 0);
1379         }
1380
1381         if (conf->coc_opc == OBJECT_CONF_WAIT) {
1382                 if (lov->lo_layout_invalid &&
1383                     atomic_read(&lov->lo_active_ios) > 0) {
1384                         lov_conf_unlock(lov);
1385                         result = lov_layout_wait(env, lov);
1386                         lov_conf_lock(lov);
1387                 }
1388                 GOTO(out, result);
1389         }
1390
1391         LASSERT(conf->coc_opc == OBJECT_CONF_SET);
1392
1393         if ((lsm == NULL && lov->lo_lsm == NULL) ||
1394             ((lsm != NULL && lov->lo_lsm != NULL) &&
1395              (lov->lo_lsm->lsm_layout_gen == lsm->lsm_layout_gen) &&
1396              (lov->lo_lsm->lsm_flags == lsm->lsm_flags) &&
1397              (lov->lo_lsm->lsm_entries[0]->lsme_pattern ==
1398               lsm->lsm_entries[0]->lsme_pattern))) {
1399                 /* same version of layout */
1400                 lov->lo_layout_invalid = false;
1401                 GOTO(out, result = 0);
1402         }
1403
1404         /* will change layout - check if there still exists active IO. */
1405         if (atomic_read(&lov->lo_active_ios) > 0) {
1406                 lov->lo_layout_invalid = true;
1407                 GOTO(out, result = -EBUSY);
1408         }
1409
1410         result = lov_layout_change(env, lov, lsm, conf);
1411         lov->lo_layout_invalid = result != 0;
1412         EXIT;
1413
1414 out:
1415         lov_conf_unlock(lov);
1416         lov_lsm_put(lsm);
1417         CDEBUG(D_INODE, DFID" lo_layout_invalid=%d\n",
1418                PFID(lu_object_fid(lov2lu(lov))), lov->lo_layout_invalid);
1419         RETURN(result);
1420 }
1421
1422 static void lov_object_delete(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_delete, env, lov, &lov->u);
1428         EXIT;
1429 }
1430
1431 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
1432 {
1433         struct lov_object *lov = lu2lov(obj);
1434
1435         ENTRY;
1436         LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
1437         lu_object_fini(obj);
1438         OBD_SLAB_FREE_PTR(lov, lov_object_kmem);
1439         EXIT;
1440 }
1441
1442 static int lov_object_print(const struct lu_env *env, void *cookie,
1443                             lu_printer_t p, const struct lu_object *o)
1444 {
1445         return LOV_2DISPATCH_NOLOCK(lu2lov(o), llo_print, env, cookie, p, o);
1446 }
1447
1448 static int lov_page_init(const struct lu_env *env, struct cl_object *obj,
1449                          struct cl_page *page, pgoff_t index)
1450 {
1451         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_page_init, env, obj, page,
1452                                     index);
1453 }
1454
1455 /**
1456  * Implements cl_object_operations::clo_io_init() method for lov
1457  * layer. Dispatches to the appropriate layout io initialization method.
1458  */
1459 static int lov_io_init(const struct lu_env *env, struct cl_object *obj,
1460                        struct cl_io *io)
1461 {
1462         CL_IO_SLICE_CLEAN(lov_env_io(env), lis_preserved);
1463
1464         CDEBUG(D_INODE, DFID "io %p type %d ignore/verify layout %d/%d\n",
1465                PFID(lu_object_fid(&obj->co_lu)), io, io->ci_type,
1466                io->ci_ignore_layout, io->ci_verify_layout);
1467
1468         /* IO type CIT_MISC with ci_ignore_layout set are usually invoked from
1469          * the OSC layer. It shouldn't take lov layout conf lock in that case,
1470          * because as long as the OSC object exists, the layout can't be
1471          * reconfigured. */
1472         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
1473                         !(io->ci_ignore_layout && io->ci_type == CIT_MISC),
1474                         env, obj, io);
1475 }
1476
1477 /**
1478  * An implementation of cl_object_operations::clo_attr_get() method for lov
1479  * layer. For raid0 layout this collects and merges attributes of all
1480  * sub-objects.
1481  */
1482 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
1483                         struct cl_attr *attr)
1484 {
1485         /* do not take lock, as this function is called under a
1486          * spin-lock. Layout is protected from changing by ongoing IO. */
1487         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
1488 }
1489
1490 static int lov_attr_update(const struct lu_env *env, struct cl_object *obj,
1491                            const struct cl_attr *attr, unsigned valid)
1492 {
1493         /*
1494          * No dispatch is required here, as no layout implements this.
1495          */
1496         return 0;
1497 }
1498
1499 static int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
1500                   struct cl_lock *lock, const struct cl_io *io)
1501 {
1502         /* No need to lock because we've taken one refcount of layout.  */
1503         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_lock_init, env, obj, lock,
1504                                     io);
1505 }
1506
1507 /**
1508  * We calculate on which OST the mapping will end. If the length of mapping
1509  * is greater than (stripe_size * stripe_count) then the last_stripe will
1510  * will be one just before start_stripe. Else we check if the mapping
1511  * intersects each OST and find last_stripe.
1512  * This function returns the last_stripe and also sets the stripe_count
1513  * over which the mapping is spread
1514  *
1515  * \param lsm [in]              striping information for the file
1516  * \param index [in]            stripe component index
1517  * \param ext [in]              logical extent of mapping
1518  * \param start_stripe [in]     starting stripe of the mapping
1519  * \param stripe_count [out]    the number of stripes across which to map is
1520  *                              returned
1521  *
1522  * \retval last_stripe          return the last stripe of the mapping
1523  */
1524 static int fiemap_calc_last_stripe(struct lov_stripe_md *lsm, int index,
1525                                    struct lu_extent *ext,
1526                                    int start_stripe, int *stripe_count)
1527 {
1528         struct lov_stripe_md_entry *lsme = lsm->lsm_entries[index];
1529         int init_stripe;
1530         int last_stripe;
1531         int i, j;
1532
1533         init_stripe = lov_stripe_number(lsm, index, ext->e_start);
1534
1535         if (ext->e_end - ext->e_start >
1536             lsme->lsme_stripe_size * lsme->lsme_stripe_count) {
1537                 if (init_stripe == start_stripe) {
1538                         last_stripe = (start_stripe < 1) ?
1539                                 lsme->lsme_stripe_count - 1 : start_stripe - 1;
1540                         *stripe_count = lsme->lsme_stripe_count;
1541                 } else if (init_stripe < start_stripe) {
1542                         last_stripe = (init_stripe < 1) ?
1543                                 lsme->lsme_stripe_count - 1 : init_stripe - 1;
1544                         *stripe_count = lsme->lsme_stripe_count -
1545                                         (start_stripe - init_stripe);
1546                 } else {
1547                         last_stripe = init_stripe - 1;
1548                         *stripe_count = init_stripe - start_stripe;
1549                 }
1550         } else {
1551                 for (j = 0, i = start_stripe; j < lsme->lsme_stripe_count;
1552                      i = (i + 1) % lsme->lsme_stripe_count, j++) {
1553                         if (!lov_stripe_intersects(lsm, index,  i, ext, NULL,
1554                                                    NULL))
1555                                 break;
1556                         if ((start_stripe != init_stripe) && (i == init_stripe))
1557                                 break;
1558                 }
1559                 *stripe_count = j;
1560                 last_stripe = (start_stripe + j - 1) % lsme->lsme_stripe_count;
1561         }
1562
1563         return last_stripe;
1564 }
1565
1566 /**
1567  * Set fe_device and copy extents from local buffer into main return buffer.
1568  *
1569  * \param fiemap [out]          fiemap to hold all extents
1570  * \param lcl_fm_ext [in]       array of fiemap extents get from OSC layer
1571  * \param ost_index [in]        OST index to be written into the fm_device
1572  *                              field for each extent
1573  * \param ext_count [in]        number of extents to be copied
1574  * \param current_extent [in]   where to start copying in the extent array
1575  */
1576 static void fiemap_prepare_and_copy_exts(struct fiemap *fiemap,
1577                                          struct fiemap_extent *lcl_fm_ext,
1578                                          int ost_index, unsigned int ext_count,
1579                                          int current_extent, int abs_stripeno)
1580 {
1581         char            *to;
1582         unsigned int    ext;
1583
1584         for (ext = 0; ext < ext_count; ext++) {
1585                 set_fe_device_stripenr(&lcl_fm_ext[ext], ost_index,
1586                                        abs_stripeno);
1587                 lcl_fm_ext[ext].fe_flags |= FIEMAP_EXTENT_NET;
1588         }
1589
1590         /* Copy fm_extent's from fm_local to return buffer */
1591         to = (char *)fiemap + fiemap_count_to_size(current_extent);
1592         memcpy(to, lcl_fm_ext, ext_count * sizeof(struct fiemap_extent));
1593 }
1594
1595 #define FIEMAP_BUFFER_SIZE 4096
1596
1597 /**
1598  * Non-zero fe_logical indicates that this is a continuation FIEMAP
1599  * call. The local end offset and the device are sent in the first
1600  * fm_extent. This function calculates the stripe number from the index.
1601  * This function returns a stripe_no on which mapping is to be restarted.
1602  *
1603  * This function returns fm_end_offset which is the in-OST offset at which
1604  * mapping should be restarted. If fm_end_offset=0 is returned then caller
1605  * will re-calculate proper offset in next stripe.
1606  * Note that the first extent is passed to lov_get_info via the value field.
1607  *
1608  * \param fiemap [in]           fiemap request header
1609  * \param lsm [in]              striping information for the file
1610  * \param index [in]            stripe component index
1611  * \param ext [in]              logical extent of mapping
1612  * \param start_stripe [out]    starting stripe will be returned in this
1613  */
1614 static u64 fiemap_calc_fm_end_offset(struct fiemap *fiemap,
1615                                      struct lov_stripe_md *lsm,
1616                                      int index, struct lu_extent *ext,
1617                                      int *start_stripe)
1618 {
1619         struct lov_stripe_md_entry *lsme = lsm->lsm_entries[index];
1620         u64 local_end = fiemap->fm_extents[0].fe_logical;
1621         u64 lun_end;
1622         u64 fm_end_offset;
1623         int stripe_no = -1;
1624
1625         if (fiemap->fm_extent_count == 0 ||
1626             fiemap->fm_extents[0].fe_logical == 0)
1627                 return 0;
1628
1629         stripe_no = *start_stripe;
1630
1631         if (stripe_no == -1)
1632                 return -EINVAL;
1633
1634         /* If we have finished mapping on previous device, shift logical
1635          * offset to start of next device */
1636         if (lov_stripe_intersects(lsm, index, stripe_no, ext, NULL, &lun_end) &&
1637             local_end < lun_end) {
1638                 fm_end_offset = local_end;
1639         } else {
1640                 /* This is a special value to indicate that caller should
1641                  * calculate offset in next stripe. */
1642                 fm_end_offset = 0;
1643                 *start_stripe = (stripe_no + 1) % lsme->lsme_stripe_count;
1644         }
1645
1646         return fm_end_offset;
1647 }
1648
1649 struct fiemap_state {
1650         struct fiemap           *fs_fm;
1651         struct lu_extent        fs_ext;         /* current entry extent */
1652         u64                     fs_length;
1653         u64                     fs_end_offset;  /* last iteration offset */
1654         int                     fs_cur_extent;  /* collected exts so far */
1655         int                     fs_cnt_need;    /* # of extents buf can hold */
1656         int                     fs_start_stripe;
1657         int                     fs_last_stripe;
1658         bool                    fs_device_done; /* enough for this OST */
1659         bool                    fs_finish_stripe; /* reached fs_last_stripe */
1660         bool                    fs_enough;      /* enough for this call */
1661 };
1662
1663 static struct cl_object *lov_find_subobj(const struct lu_env *env,
1664                                          struct lov_object *lov,
1665                                          struct lov_stripe_md *lsm,
1666                                          int index)
1667 {
1668         struct lov_device       *dev = lu2lov_dev(lov2lu(lov)->lo_dev);
1669         struct lov_thread_info  *lti = lov_env_info(env);
1670         struct lu_fid           *ofid = &lti->lti_fid;
1671         struct lov_oinfo        *oinfo;
1672         struct cl_device        *subdev;
1673         int                     entry = lov_comp_entry(index);
1674         int                     stripe = lov_comp_stripe(index);
1675         int                     ost_idx;
1676         int                     rc;
1677         struct cl_object        *result;
1678
1679         if (lov->lo_type != LLT_COMP)
1680                 GOTO(out, result = NULL);
1681
1682         if (entry >= lsm->lsm_entry_count ||
1683             stripe >= lsm->lsm_entries[entry]->lsme_stripe_count)
1684                 GOTO(out, result = NULL);
1685
1686         oinfo = lsm->lsm_entries[entry]->lsme_oinfo[stripe];
1687         ost_idx = oinfo->loi_ost_idx;
1688         rc = ostid_to_fid(ofid, &oinfo->loi_oi, ost_idx);
1689         if (rc != 0)
1690                 GOTO(out, result = NULL);
1691
1692         subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
1693         result = lov_sub_find(env, subdev, ofid, NULL);
1694 out:
1695         if (result == NULL)
1696                 result = ERR_PTR(-EINVAL);
1697         return result;
1698 }
1699
1700 static int fiemap_for_stripe(const struct lu_env *env, struct cl_object *obj,
1701                              struct lov_stripe_md *lsm, struct fiemap *fiemap,
1702                              size_t *buflen, struct ll_fiemap_info_key *fmkey,
1703                              int index, int stripe_last, int stripeno,
1704                              struct fiemap_state *fs)
1705 {
1706         struct lov_stripe_md_entry *lsme = lsm->lsm_entries[index];
1707         struct cl_object *subobj;
1708         struct lov_obd *lov = lu2lov_dev(obj->co_lu.lo_dev)->ld_lov;
1709         struct fiemap_extent *fm_ext = &fs->fs_fm->fm_extents[0];
1710         u64 req_fm_len; /* max requested extent coverage */
1711         u64 len_mapped_single_call;
1712         u64 obd_start;
1713         u64 obd_end;
1714         unsigned int ext_count;
1715         /* EOF for object */
1716         bool ost_eof = false;
1717         /* done with required mapping for this OST? */
1718         bool ost_done = false;
1719         int ost_index;
1720         int rc = 0;
1721
1722         fs->fs_device_done = false;
1723         /* Find out range of mapping on this stripe */
1724         if ((lov_stripe_intersects(lsm, index, stripeno, &fs->fs_ext,
1725                                    &obd_start, &obd_end)) == 0)
1726                 return 0;
1727
1728         if (lov_oinfo_is_dummy(lsme->lsme_oinfo[stripeno]))
1729                 return -EIO;
1730
1731         /* If this is a continuation FIEMAP call and we are on
1732          * starting stripe then obd_start needs to be set to
1733          * end_offset */
1734         if (fs->fs_end_offset != 0 && stripeno == fs->fs_start_stripe)
1735                 obd_start = fs->fs_end_offset;
1736
1737         if (lov_size_to_stripe(lsm, index, fs->fs_ext.e_end, stripeno) ==
1738             obd_start)
1739                 return 0;
1740
1741         req_fm_len = obd_end - obd_start + 1;
1742         fs->fs_fm->fm_length = 0;
1743         len_mapped_single_call = 0;
1744
1745         /* find lobsub object */
1746         subobj = lov_find_subobj(env, cl2lov(obj), lsm,
1747                                  lov_comp_index(index, stripeno));
1748         if (IS_ERR(subobj))
1749                 return PTR_ERR(subobj);
1750         /* If the output buffer is very large and the objects have many
1751          * extents we may need to loop on a single OST repeatedly */
1752         do {
1753                 if (fiemap->fm_extent_count > 0) {
1754                         /* Don't get too many extents. */
1755                         if (fs->fs_cur_extent + fs->fs_cnt_need >
1756                             fiemap->fm_extent_count)
1757                                 fs->fs_cnt_need = fiemap->fm_extent_count -
1758                                                   fs->fs_cur_extent;
1759                 }
1760
1761                 obd_start += len_mapped_single_call;
1762                 fs->fs_fm->fm_length = req_fm_len - len_mapped_single_call;
1763                 req_fm_len = fs->fs_fm->fm_length;
1764                 /**
1765                  * If we've collected enough extent map, we'd request 1 more,
1766                  * to see whether we coincidentally finished all available
1767                  * extent map, so that FIEMAP_EXTENT_LAST would be set.
1768                  */
1769                 fs->fs_fm->fm_extent_count = fs->fs_enough ?
1770                                              1 : fs->fs_cnt_need;
1771                 fs->fs_fm->fm_mapped_extents = 0;
1772                 fs->fs_fm->fm_flags = fiemap->fm_flags;
1773
1774                 ost_index = lsme->lsme_oinfo[stripeno]->loi_ost_idx;
1775
1776                 if (ost_index < 0 || ost_index >= lov->desc.ld_tgt_count)
1777                         GOTO(obj_put, rc = -EINVAL);
1778                 /* If OST is inactive, return extent with UNKNOWN flag. */
1779                 if (!lov->lov_tgts[ost_index]->ltd_active) {
1780                         fs->fs_fm->fm_flags |= FIEMAP_EXTENT_LAST;
1781                         fs->fs_fm->fm_mapped_extents = 1;
1782
1783                         fm_ext[0].fe_logical = obd_start;
1784                         fm_ext[0].fe_length = obd_end - obd_start + 1;
1785                         fm_ext[0].fe_flags |= FIEMAP_EXTENT_UNKNOWN;
1786
1787                         goto inactive_tgt;
1788                 }
1789
1790                 fs->fs_fm->fm_start = obd_start;
1791                 fs->fs_fm->fm_flags &= ~FIEMAP_FLAG_DEVICE_ORDER;
1792                 memcpy(&fmkey->lfik_fiemap, fs->fs_fm, sizeof(*fs->fs_fm));
1793                 *buflen = fiemap_count_to_size(fs->fs_fm->fm_extent_count);
1794
1795                 rc = cl_object_fiemap(env, subobj, fmkey, fs->fs_fm, buflen);
1796                 if (rc != 0)
1797                         GOTO(obj_put, rc);
1798 inactive_tgt:
1799                 ext_count = fs->fs_fm->fm_mapped_extents;
1800                 if (ext_count == 0) {
1801                         ost_done = true;
1802                         fs->fs_device_done = true;
1803                         /* If last stripe has hold at the end,
1804                          * we need to return */
1805                         if (stripeno == fs->fs_last_stripe) {
1806                                 fiemap->fm_mapped_extents = 0;
1807                                 fs->fs_finish_stripe = true;
1808                                 GOTO(obj_put, rc);
1809                         }
1810                         break;
1811                 } else if (fs->fs_enough) {
1812                         /*
1813                          * We've collected enough extents and there are
1814                          * more extents after it.
1815                          */
1816                         GOTO(obj_put, rc);
1817                 }
1818
1819                 /* If we just need num of extents, got to next device */
1820                 if (fiemap->fm_extent_count == 0) {
1821                         fs->fs_cur_extent += ext_count;
1822                         break;
1823                 }
1824
1825                 /* prepare to copy retrived map extents */
1826                 len_mapped_single_call = fm_ext[ext_count - 1].fe_logical +
1827                                          fm_ext[ext_count - 1].fe_length -
1828                                          obd_start;
1829
1830                 /* Have we finished mapping on this device? */
1831                 if (req_fm_len <= len_mapped_single_call) {
1832                         ost_done = true;
1833                         fs->fs_device_done = true;
1834                 }
1835
1836                 /* Clear the EXTENT_LAST flag which can be present on
1837                  * the last extent */
1838                 if (fm_ext[ext_count - 1].fe_flags & FIEMAP_EXTENT_LAST)
1839                         fm_ext[ext_count - 1].fe_flags &= ~FIEMAP_EXTENT_LAST;
1840                 if (lov_stripe_size(lsm, index,
1841                                     fm_ext[ext_count - 1].fe_logical +
1842                                     fm_ext[ext_count - 1].fe_length,
1843                                     stripeno) >= fmkey->lfik_oa.o_size) {
1844                         ost_eof = true;
1845                         fs->fs_device_done = true;
1846                 }
1847
1848                 fiemap_prepare_and_copy_exts(fiemap, fm_ext, ost_index,
1849                                              ext_count, fs->fs_cur_extent,
1850                                              stripe_last + stripeno);
1851                 fs->fs_cur_extent += ext_count;
1852
1853                 /* Ran out of available extents? */
1854                 if (fs->fs_cur_extent >= fiemap->fm_extent_count)
1855                         fs->fs_enough = true;
1856         } while (!ost_done && !ost_eof);
1857
1858         if (stripeno == fs->fs_last_stripe)
1859                 fs->fs_finish_stripe = true;
1860 obj_put:
1861         cl_object_put(env, subobj);
1862
1863         return rc;
1864 }
1865
1866 /**
1867  * Break down the FIEMAP request and send appropriate calls to individual OSTs.
1868  * This also handles the restarting of FIEMAP calls in case mapping overflows
1869  * the available number of extents in single call.
1870  *
1871  * \param env [in]              lustre environment
1872  * \param obj [in]              file object
1873  * \param fmkey [in]            fiemap request header and other info
1874  * \param fiemap [out]          fiemap buffer holding retrived map extents
1875  * \param buflen [in/out]       max buffer length of @fiemap, when iterate
1876  *                              each OST, it is used to limit max map needed
1877  * \retval 0    success
1878  * \retval < 0  error
1879  */
1880 static int lov_object_fiemap(const struct lu_env *env, struct cl_object *obj,
1881                              struct ll_fiemap_info_key *fmkey,
1882                              struct fiemap *fiemap, size_t *buflen)
1883 {
1884         struct lov_stripe_md_entry *lsme;
1885         struct lov_stripe_md *lsm;
1886         struct fiemap *fm_local = NULL;
1887         loff_t whole_start;
1888         loff_t whole_end;
1889         int entry;
1890         int start_entry = -1;
1891         int end_entry;
1892         int cur_stripe = 0;
1893         int stripe_count;
1894         unsigned int buffer_size = FIEMAP_BUFFER_SIZE;
1895         int rc = 0;
1896         struct fiemap_state fs = { 0 };
1897         struct lu_extent range;
1898         int cur_ext;
1899         int stripe_last;
1900         int start_stripe = 0;
1901         bool resume = false;
1902         ENTRY;
1903
1904         lsm = lov_lsm_addref(cl2lov(obj));
1905         if (lsm == NULL) {
1906                 /* no extent: there is no object for mapping */
1907                 fiemap->fm_mapped_extents = 0;
1908                 return 0;
1909         }
1910
1911         if (!(fiemap->fm_flags & FIEMAP_FLAG_DEVICE_ORDER)) {
1912                 /**
1913                  * If the entry count > 1 or stripe_count > 1 and the
1914                  * application does not understand DEVICE_ORDER flag,
1915                  * it cannot interpret the extents correctly.
1916                  */
1917                 if (lsm->lsm_entry_count > 1 ||
1918                     (lsm->lsm_entry_count == 1 &&
1919                      lsm->lsm_entries[0]->lsme_stripe_count > 1))
1920                         GOTO(out_lsm, rc = -ENOTSUPP);
1921         }
1922
1923         /* No support for DOM layout yet. */
1924         if (lsme_is_dom(lsm->lsm_entries[0]))
1925                 GOTO(out_lsm, rc = -ENOTSUPP);
1926
1927         if (lsm->lsm_is_released) {
1928                 if (fiemap->fm_start < fmkey->lfik_oa.o_size) {
1929                         /**
1930                          * released file, return a minimal FIEMAP if
1931                          * request fits in file-size.
1932                          */
1933                         fiemap->fm_mapped_extents = 1;
1934                         fiemap->fm_extents[0].fe_logical = fiemap->fm_start;
1935                         if (fiemap->fm_start + fiemap->fm_length <
1936                             fmkey->lfik_oa.o_size)
1937                                 fiemap->fm_extents[0].fe_length =
1938                                         fiemap->fm_length;
1939                         else
1940                                 fiemap->fm_extents[0].fe_length =
1941                                         fmkey->lfik_oa.o_size -
1942                                         fiemap->fm_start;
1943                         fiemap->fm_extents[0].fe_flags |=
1944                                 FIEMAP_EXTENT_UNKNOWN | FIEMAP_EXTENT_LAST;
1945                 }
1946                 GOTO(out_lsm, rc = 0);
1947         }
1948
1949         /* buffer_size is small to hold fm_extent_count of extents. */
1950         if (fiemap_count_to_size(fiemap->fm_extent_count) < buffer_size)
1951                 buffer_size = fiemap_count_to_size(fiemap->fm_extent_count);
1952
1953         OBD_ALLOC_LARGE(fm_local, buffer_size);
1954         if (fm_local == NULL)
1955                 GOTO(out_lsm, rc = -ENOMEM);
1956
1957         /**
1958          * Requested extent count exceeds the fiemap buffer size, shrink our
1959          * ambition.
1960          */
1961         if (fiemap_count_to_size(fiemap->fm_extent_count) > *buflen)
1962                 fiemap->fm_extent_count = fiemap_size_to_count(*buflen);
1963
1964         fs.fs_enough = false;
1965         fs.fs_cur_extent = 0;
1966         fs.fs_fm = fm_local;
1967         fs.fs_cnt_need = fiemap_size_to_count(buffer_size);
1968
1969         whole_start = fiemap->fm_start;
1970         /* whole_start is beyond the end of the file */
1971         if (whole_start > fmkey->lfik_oa.o_size)
1972                 GOTO(out_fm_local, rc = -EINVAL);
1973         whole_end = (fiemap->fm_length == OBD_OBJECT_EOF) ?
1974                                         fmkey->lfik_oa.o_size + 1 :
1975                                         whole_start + fiemap->fm_length;
1976         /**
1977          * If fiemap->fm_length != OBD_OBJECT_EOF but whole_end exceeds file
1978          * size
1979          */
1980         if (whole_end > fmkey->lfik_oa.o_size + 1)
1981                 whole_end = fmkey->lfik_oa.o_size + 1;
1982
1983         /**
1984          * the high 16bits of fe_device remember which stripe the last
1985          * call has been arrived, we'd continue from there in this call.
1986          */
1987         if (fiemap->fm_extent_count && fiemap->fm_extents[0].fe_logical)
1988                 resume = true;
1989         stripe_last = get_fe_stripenr(&fiemap->fm_extents[0]);
1990         /**
1991          * stripe_last records stripe number we've been processed in the last
1992          * call
1993          */
1994         end_entry = lsm->lsm_entry_count - 1;
1995         cur_stripe = 0;
1996         for (entry = 0; entry <= end_entry; entry++) {
1997                 lsme = lsm->lsm_entries[entry];
1998                 if (cur_stripe + lsme->lsme_stripe_count >= stripe_last) {
1999                         start_entry = entry;
2000                         start_stripe = stripe_last - cur_stripe;
2001                         break;
2002                 }
2003
2004                 cur_stripe += lsme->lsme_stripe_count;
2005         }
2006         if (start_entry == -1) {
2007                 CERROR(DFID": FIEMAP does not init start entry, cur_stripe=%d, "
2008                        "stripe_last=%d\n", PFID(lu_object_fid(&obj->co_lu)),
2009                        cur_stripe, stripe_last);
2010                 GOTO(out_fm_local, rc = -EINVAL);
2011         }
2012         /**
2013          * @start_entry & @start_stripe records the position of fiemap
2014          * resumption @stripe_last keeps recording the absolution position
2015          * we'are processing. @resume indicates we'd honor @start_stripe.
2016          */
2017
2018         range.e_start = whole_start;
2019         range.e_end = whole_end;
2020
2021         for (entry = start_entry; entry <= end_entry; entry++) {
2022                 /* remeber to update stripe_last accordingly */
2023                 lsme = lsm->lsm_entries[entry];
2024
2025                 /* FLR could contain component holes between entries */
2026                 if (!lsme_inited(lsme)) {
2027                         stripe_last += lsme->lsme_stripe_count;
2028                         resume = false;
2029                         continue;
2030                 }
2031
2032                 if (!lu_extent_is_overlapped(&range, &lsme->lsme_extent)) {
2033                         stripe_last += lsme->lsme_stripe_count;
2034                         resume = false;
2035                         continue;
2036                 }
2037
2038                 /* prepare for a component entry iteration */
2039                 if (lsme->lsme_extent.e_start > whole_start)
2040                         fs.fs_ext.e_start = lsme->lsme_extent.e_start;
2041                 else
2042                         fs.fs_ext.e_start = whole_start;
2043                 if (lsme->lsme_extent.e_end > whole_end)
2044                         fs.fs_ext.e_end = whole_end;
2045                 else
2046                         fs.fs_ext.e_end = lsme->lsme_extent.e_end;
2047
2048                 /* Calculate start stripe, last stripe and length of mapping */
2049                 if (resume) {
2050                         fs.fs_start_stripe = start_stripe;
2051                         /* put stripe_last to the first stripe of the comp */
2052                         stripe_last -= start_stripe;
2053                         resume = false;
2054                 } else {
2055                         fs.fs_start_stripe = lov_stripe_number(lsm, entry,
2056                                                         fs.fs_ext.e_start);
2057                 }
2058                 fs.fs_last_stripe = fiemap_calc_last_stripe(lsm, entry,
2059                                         &fs.fs_ext, fs.fs_start_stripe,
2060                                         &stripe_count);
2061                 /**
2062                  * A new mirror component is under process, reset
2063                  * fs.fs_end_offset and then fiemap_for_stripe() starts from
2064                  * the overlapping extent, otherwise starts from
2065                  * fs.fs_end_offset.
2066                  */
2067                 if (entry > start_entry && lsme->lsme_extent.e_start == 0) {
2068                         /* new mirror */
2069                         fs.fs_end_offset = 0;
2070                 } else {
2071                         fs.fs_end_offset = fiemap_calc_fm_end_offset(fiemap,
2072                                                 lsm, entry, &fs.fs_ext,
2073                                                 &fs.fs_start_stripe);
2074                 }
2075
2076                 /* Check each stripe */
2077                 for (cur_stripe = fs.fs_start_stripe; stripe_count > 0;
2078                      --stripe_count,
2079                      cur_stripe = (cur_stripe + 1) % lsme->lsme_stripe_count) {
2080                         /* reset fs_finish_stripe */
2081                         fs.fs_finish_stripe = false;
2082                         rc = fiemap_for_stripe(env, obj, lsm, fiemap, buflen,
2083                                                fmkey, entry, stripe_last,
2084                                                cur_stripe, &fs);
2085                         if (rc < 0)
2086                                 GOTO(out_fm_local, rc);
2087                         if (fs.fs_enough) {
2088                                 stripe_last += cur_stripe;
2089                                 GOTO(finish, rc);
2090                         }
2091                         if (fs.fs_finish_stripe)
2092                                 break;
2093                 } /* for each stripe */
2094                 stripe_last += lsme->lsme_stripe_count;
2095         } /* for covering layout component entry */
2096
2097 finish:
2098         if (fs.fs_cur_extent > 0)
2099                 cur_ext = fs.fs_cur_extent - 1;
2100         else
2101                 cur_ext = 0;
2102
2103         /* done all the processing */
2104         if (entry > end_entry)
2105                 fiemap->fm_extents[cur_ext].fe_flags |= FIEMAP_EXTENT_LAST;
2106
2107         /* Indicate that we are returning device offsets unless file just has
2108          * single stripe */
2109         if (lsm->lsm_entry_count > 1 ||
2110             (lsm->lsm_entry_count == 1 &&
2111              lsm->lsm_entries[0]->lsme_stripe_count > 1))
2112                 fiemap->fm_flags |= FIEMAP_FLAG_DEVICE_ORDER;
2113
2114         if (fiemap->fm_extent_count == 0)
2115                 goto skip_last_device_calc;
2116
2117 skip_last_device_calc:
2118         fiemap->fm_mapped_extents = fs.fs_cur_extent;
2119 out_fm_local:
2120         OBD_FREE_LARGE(fm_local, buffer_size);
2121
2122 out_lsm:
2123         lov_lsm_put(lsm);
2124         return rc;
2125 }
2126
2127 static int lov_object_getstripe(const struct lu_env *env, struct cl_object *obj,
2128                                 struct lov_user_md __user *lum, size_t size)
2129 {
2130         struct lov_object       *lov = cl2lov(obj);
2131         struct lov_stripe_md    *lsm;
2132         int                     rc = 0;
2133         ENTRY;
2134
2135         lsm = lov_lsm_addref(lov);
2136         if (lsm == NULL)
2137                 RETURN(-ENODATA);
2138
2139         rc = lov_getstripe(env, cl2lov(obj), lsm, lum, size);
2140         lov_lsm_put(lsm);
2141         RETURN(rc);
2142 }
2143
2144 static int lov_object_layout_get(const struct lu_env *env,
2145                                  struct cl_object *obj,
2146                                  struct cl_layout *cl)
2147 {
2148         struct lov_object *lov = cl2lov(obj);
2149         struct lov_stripe_md *lsm = lov_lsm_addref(lov);
2150         struct lu_buf *buf = &cl->cl_buf;
2151         ssize_t rc;
2152         ENTRY;
2153
2154         if (lsm == NULL) {
2155                 cl->cl_size = 0;
2156                 cl->cl_layout_gen = CL_LAYOUT_GEN_EMPTY;
2157
2158                 RETURN(0);
2159         }
2160
2161         cl->cl_size = lov_comp_md_size(lsm);
2162         cl->cl_layout_gen = lsm->lsm_layout_gen;
2163         cl->cl_is_released = lsm->lsm_is_released;
2164         cl->cl_is_composite = lsm_is_composite(lsm->lsm_magic);
2165
2166         rc = lov_lsm_pack(lsm, buf->lb_buf, buf->lb_len);
2167         lov_lsm_put(lsm);
2168
2169         /* return error or number of bytes */
2170         RETURN(rc);
2171 }
2172
2173 static loff_t lov_object_maxbytes(struct cl_object *obj)
2174 {
2175         struct lov_object *lov = cl2lov(obj);
2176         struct lov_stripe_md *lsm = lov_lsm_addref(lov);
2177         loff_t maxbytes;
2178
2179         if (lsm == NULL)
2180                 return LLONG_MAX;
2181
2182         maxbytes = lsm->lsm_maxbytes;
2183
2184         lov_lsm_put(lsm);
2185
2186         return maxbytes;
2187 }
2188
2189 static int lov_object_flush(const struct lu_env *env, struct cl_object *obj,
2190                             struct ldlm_lock *lock)
2191 {
2192         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_flush, true, env, obj,
2193                                      lock);
2194 }
2195
2196 static const struct cl_object_operations lov_ops = {
2197         .coo_page_init    = lov_page_init,
2198         .coo_lock_init    = lov_lock_init,
2199         .coo_io_init      = lov_io_init,
2200         .coo_attr_get     = lov_attr_get,
2201         .coo_attr_update  = lov_attr_update,
2202         .coo_conf_set     = lov_conf_set,
2203         .coo_getstripe    = lov_object_getstripe,
2204         .coo_layout_get   = lov_object_layout_get,
2205         .coo_maxbytes     = lov_object_maxbytes,
2206         .coo_fiemap       = lov_object_fiemap,
2207         .coo_object_flush = lov_object_flush
2208 };
2209
2210 static const struct lu_object_operations lov_lu_obj_ops = {
2211         .loo_object_init        = lov_object_init,
2212         .loo_object_delete      = lov_object_delete,
2213         .loo_object_release     = NULL,
2214         .loo_object_free        = lov_object_free,
2215         .loo_object_print       = lov_object_print,
2216         .loo_object_invariant   = NULL,
2217 };
2218
2219 struct lu_object *lov_object_alloc(const struct lu_env *env,
2220                                    const struct lu_object_header *unused,
2221                                    struct lu_device *dev)
2222 {
2223         struct lov_object *lov;
2224         struct lu_object  *obj;
2225
2226         ENTRY;
2227         OBD_SLAB_ALLOC_PTR_GFP(lov, lov_object_kmem, GFP_NOFS);
2228         if (lov != NULL) {
2229                 obj = lov2lu(lov);
2230                 lu_object_init(obj, NULL, dev);
2231                 lov->lo_cl.co_ops = &lov_ops;
2232                 lov->lo_type = -1; /* invalid, to catch uninitialized type */
2233                 /*
2234                  * object io operation vector (cl_object::co_iop) is installed
2235                  * later in lov_object_init(), as different vectors are used
2236                  * for object with different layouts.
2237                  */
2238                 obj->lo_ops = &lov_lu_obj_ops;
2239         } else
2240                 obj = NULL;
2241         RETURN(obj);
2242 }
2243
2244 static struct lov_stripe_md *lov_lsm_addref(struct lov_object *lov)
2245 {
2246         struct lov_stripe_md *lsm = NULL;
2247
2248         lov_conf_freeze(lov);
2249         if (lov->lo_lsm != NULL) {
2250                 lsm = lsm_addref(lov->lo_lsm);
2251                 CDEBUG(D_INODE, "lsm %p addref %d/%d by %p.\n",
2252                         lsm, atomic_read(&lsm->lsm_refc),
2253                         lov->lo_layout_invalid, current);
2254         }
2255         lov_conf_thaw(lov);
2256         return lsm;
2257 }
2258
2259 int lov_read_and_clear_async_rc(struct cl_object *clob)
2260 {
2261         struct lu_object *luobj;
2262         int rc = 0;
2263         ENTRY;
2264
2265         luobj = lu_object_locate(&cl_object_header(clob)->coh_lu,
2266                                  &lov_device_type);
2267         if (luobj != NULL) {
2268                 struct lov_object *lov = lu2lov(luobj);
2269
2270                 lov_conf_freeze(lov);
2271                 switch (lov->lo_type) {
2272                 case LLT_COMP: {
2273                         struct lov_stripe_md *lsm;
2274                         int i;
2275
2276                         lsm = lov->lo_lsm;
2277                         LASSERT(lsm != NULL);
2278                         for (i = 0; i < lsm->lsm_entry_count; i++) {
2279                                 struct lov_stripe_md_entry *lse =
2280                                                 lsm->lsm_entries[i];
2281                                 int j;
2282
2283                                 if (!lsme_inited(lse))
2284                                         break;
2285
2286                                 for (j = 0; j < lse->lsme_stripe_count; j++) {
2287                                         struct lov_oinfo *loi =
2288                                                         lse->lsme_oinfo[j];
2289
2290                                         if (lov_oinfo_is_dummy(loi))
2291                                                 continue;
2292
2293                                         if (loi->loi_ar.ar_rc && !rc)
2294                                                 rc = loi->loi_ar.ar_rc;
2295                                         loi->loi_ar.ar_rc = 0;
2296                                 }
2297                         }
2298                 }
2299                 case LLT_RELEASED:
2300                 case LLT_EMPTY:
2301                         /* fall through */
2302                 case LLT_FOREIGN:
2303                         break;
2304                 default:
2305                         LBUG();
2306                 }
2307                 lov_conf_thaw(lov);
2308         }
2309         RETURN(rc);
2310 }
2311 EXPORT_SYMBOL(lov_read_and_clear_async_rc);
2312
2313 /** @} lov */