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                 default:
 687                         CERROR("%s: unknown composite layout entry type %i\n",
 688                                lov2obd(dev->ld_lov)->obd_name,
 689                                lsm->lsm_entries[i]->lsme_pattern);
 690                         dump_lsm(D_ERROR, lsm);
 691                         RETURN(-EIO);
 692                 }
 693
 694                 lle->lle_extent = &lle->lle_lsme->lsme_extent;
 695                 lle->lle_valid = !(lle->lle_lsme->lsme_flags & LCME_FL_STALE);
 696
 697                 if (flr_state != LCM_FL_NONE)
 698                         mirror_id = mirror_id_of(lle->lle_lsme->lsme_id);
 699
 700                 lre = &comp->lo_mirrors[j];
 701                 if (i > 0) {
 702                         if (mirror_id == lre->lre_mirror_id) {
 703                                 lre->lre_valid |= lle->lle_valid;
 704                                 lre->lre_stale |= !lle->lle_valid;
 705                                 lre->lre_end = i;
 706                                 continue;
 707                         }
 708
 709                         /* new mirror detected, assume that the mirrors
 710                          * are shorted in layout */
 711                         ++mirror_count;
 712                         ++j;
 713                         if (j >= comp->lo_mirror_count)
 714                                 break;
 715
 716                         lre = &comp->lo_mirrors[j];
 717                 }
 718
 719                 /* entries must be sorted by mirrors */
 720                 lre->lre_mirror_id = mirror_id;
 721                 lre->lre_start = lre->lre_end = i;
 722                 lre->lre_preferred = !!(lle->lle_lsme->lsme_flags &
 723                                         LCME_FL_PREF_RD);
 724                 lre->lre_valid = lle->lle_valid;
 725                 lre->lre_stale = !lle->lle_valid;
 726         }
 727
 728         /* sanity check for FLR */
 729         if (mirror_count != comp->lo_mirror_count) {
 730                 CDEBUG(D_INODE, DFID
 731                        " doesn't have the # of mirrors it claims, %u/%u\n",
 732                        PFID(lu_object_fid(lov2lu(lov))), mirror_count,
 733                        comp->lo_mirror_count + 1);
 734
 735                 GOTO(out, result = -EINVAL);
 736         }
 737
 738         lov_foreach_layout_entry(lov, lle) {
 739                 int index = lov_layout_entry_index(lov, lle);
 740
 741                 /**
 742                  * If the component has not been init-ed on MDS side, for
 743                  * PFL layout, we'd know that the components beyond this one
 744                  * will be dynamically init-ed later on file write/trunc ops.
 745                  */
 746                 if (!lsme_inited(lle->lle_lsme))
 747                         continue;
 748
 749                 result = lle->lle_comp_ops->lco_init(env, dev, lov, index,
 750                                                      conf, lle);
 751                 if (result < 0)
 752                         break;
 753
 754                 LASSERT(ergo(psz > 0, psz == result));
 755                 psz = result;
 756         }
 757
 758         if (psz > 0)
 759                 cl_object_header(&lov->lo_cl)->coh_page_bufsize += psz;
 760
 761         /* decide the preferred mirror. It uses the hash value of lov_object
 762          * so that different clients would use different mirrors for read. */
 763         mirror_count = 0;
 764         seq = hash_long((unsigned long)lov, 8);
 765         for (i = 0; i < comp->lo_mirror_count; i++) {
 766                 unsigned int idx = (i + seq) % comp->lo_mirror_count;
 767
 768                 lre = lov_mirror_entry(lov, idx);
 769                 if (lre->lre_stale)
 770                         continue;
 771
 772                 mirror_count++; /* valid mirror */
 773
 774                 if (lre->lre_preferred || comp->lo_preferred_mirror < 0)
 775                         comp->lo_preferred_mirror = idx;
 776         }
 777         if (!mirror_count) {
 778                 CDEBUG(D_INODE, DFID
 779                        " doesn't have any valid mirrors\n",
 780                        PFID(lu_object_fid(lov2lu(lov))));
 781
 782                 comp->lo_preferred_mirror = 0;
 783         }
 784
 785         LASSERT(comp->lo_preferred_mirror >= 0);
 786
 787         EXIT;
 788 out:
 789         return result > 0 ? 0 : result;
 790 }
 791
 792 static int lov_init_empty(const struct lu_env *env, struct lov_device *dev,
 793                           struct lov_object *lov, struct lov_stripe_md *lsm,
 794                           const struct cl_object_conf *conf,
 795                           union lov_layout_state *state)
 796 {
 797         return 0;
 798 }
 799
 800 static int lov_init_released(const struct lu_env *env,
 801                              struct lov_device *dev, struct lov_object *lov,
 802                              struct lov_stripe_md *lsm,
 803                              const struct cl_object_conf *conf,
 804                              union lov_layout_state *state)
 805 {
 806         LASSERT(lsm != NULL);
 807         LASSERT(lsm->lsm_is_released);
 808         LASSERT(lov->lo_lsm == NULL);
 809
 810         lov->lo_lsm = lsm_addref(lsm);
 811         return 0;
 812 }
 813
 814 static int lov_init_foreign(const struct lu_env *env,
 815                             struct lov_device *dev, struct lov_object *lov,
 816                             struct lov_stripe_md *lsm,
 817                             const struct cl_object_conf *conf,
 818                             union lov_layout_state *state)
 819 {
 820         LASSERT(lsm != NULL);
 821         LASSERT(lov->lo_type == LLT_FOREIGN);
 822         LASSERT(lov->lo_lsm == NULL);
 823
 824         lov->lo_lsm = lsm_addref(lsm);
 825         return 0;
 826 }
 827
 828 static int lov_delete_empty(const struct lu_env *env, struct lov_object *lov,
 829                             union lov_layout_state *state)
 830 {
 831         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED ||
 832                 lov->lo_type == LLT_FOREIGN);
 833
 834         lov_layout_wait(env, lov);
 835         return 0;
 836 }
 837
 838 static int lov_delete_composite(const struct lu_env *env,
 839                                 struct lov_object *lov,
 840                                 union lov_layout_state *state)
 841 {
 842         struct lov_layout_entry *entry;
 843         struct lov_layout_composite *comp = &state->composite;
 844
 845         ENTRY;
 846
 847         dump_lsm(D_INODE, lov->lo_lsm);
 848
 849         lov_layout_wait(env, lov);
 850         if (comp->lo_entries)
 851                 lov_foreach_layout_entry(lov, entry)
 852                         lov_delete_raid0(env, lov, entry);
 853
 854         RETURN(0);
 855 }
 856
 857 static void lov_fini_empty(const struct lu_env *env, struct lov_object *lov,
 858                            union lov_layout_state *state)
 859 {
 860         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 861 }
 862
 863 static void lov_fini_composite(const struct lu_env *env,
 864                                struct lov_object *lov,
 865                                union lov_layout_state *state)
 866 {
 867         struct lov_layout_composite *comp = &state->composite;
 868         ENTRY;
 869
 870         if (comp->lo_entries != NULL) {
 871                 struct lov_layout_entry *entry;
 872
 873                 lov_foreach_layout_entry(lov, entry)
 874                         if (entry->lle_comp_ops)
 875                                 entry->lle_comp_ops->lco_fini(env, entry);
 876
 877                 OBD_FREE_PTR_ARRAY(comp->lo_entries, comp->lo_entry_count);
 878                 comp->lo_entries = NULL;
 879         }
 880
 881         if (comp->lo_mirrors != NULL) {
 882                 OBD_FREE_PTR_ARRAY(comp->lo_mirrors, comp->lo_mirror_count);
 883                 comp->lo_mirrors = NULL;
 884         }
 885
 886         memset(comp, 0, sizeof(*comp));
 887
 888         dump_lsm(D_INODE, lov->lo_lsm);
 889         lov_free_memmd(&lov->lo_lsm);
 890
 891         EXIT;
 892 }
 893
 894 static void lov_fini_released(const struct lu_env *env, struct lov_object *lov,
 895                                 union lov_layout_state *state)
 896 {
 897         ENTRY;
 898         dump_lsm(D_INODE, lov->lo_lsm);
 899         lov_free_memmd(&lov->lo_lsm);
 900         EXIT;
 901 }
 902
 903 static int lov_print_empty(const struct lu_env *env, void *cookie,
 904                            lu_printer_t p, const struct lu_object *o)
 905 {
 906         (*p)(env, cookie, "empty %d\n", lu2lov(o)->lo_layout_invalid);
 907         return 0;
 908 }
 909
 910 static int lov_print_composite(const struct lu_env *env, void *cookie,
 911                                lu_printer_t p, const struct lu_object *o)
 912 {
 913         struct lov_object *lov = lu2lov(o);
 914         struct lov_stripe_md *lsm = lov->lo_lsm;
 915         int i;
 916
 917         (*p)(env, cookie, "entries: %d, %s, lsm{%p 0x%08X %d %u}:\n",
 918              lsm->lsm_entry_count,
 919              lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 920              lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 921              lsm->lsm_layout_gen);
 922
 923         for (i = 0; i < lsm->lsm_entry_count; i++) {
 924                 struct lov_stripe_md_entry *lse = lsm->lsm_entries[i];
 925                 struct lov_layout_entry *lle = lov_entry(lov, i);
 926
 927                 (*p)(env, cookie,
 928                      DEXT ": { 0x%08X, %u, %#x, %u, %#x, %u, %u }\n",
 929                      PEXT(&lse->lsme_extent), lse->lsme_magic,
 930                      lse->lsme_id, lse->lsme_pattern, lse->lsme_layout_gen,
 931                      lse->lsme_flags, lse->lsme_stripe_count,
 932                      lse->lsme_stripe_size);
 933                 lov_print_raid0(env, cookie, p, lle);
 934         }
 935
 936         return 0;
 937 }
 938
 939 static int lov_print_released(const struct lu_env *env, void *cookie,
 940                                 lu_printer_t p, const struct lu_object *o)
 941 {
 942         struct lov_object       *lov = lu2lov(o);
 943         struct lov_stripe_md    *lsm = lov->lo_lsm;
 944
 945         (*p)(env, cookie,
 946                 "released: %s, lsm{%p 0x%08X %d %u}:\n",
 947                 lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 948                 lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 949                 lsm->lsm_layout_gen);
 950         return 0;
 951 }
 952
 953 static int lov_print_foreign(const struct lu_env *env, void *cookie,
 954                                 lu_printer_t p, const struct lu_object *o)
 955 {
 956         struct lov_object       *lov = lu2lov(o);
 957         struct lov_stripe_md    *lsm = lov->lo_lsm;
 958
 959         (*p)(env, cookie,
 960                 "foreign: %s, lsm{%p 0x%08X %d %u}:\n",
 961                 lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 962                 lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 963                 lsm->lsm_layout_gen);
 964         (*p)(env, cookie,
 965                 "raw_ea_content '%.*s'\n",
 966                 (int)lsm->lsm_foreign_size, (char *)lsm_foreign(lsm));
 967         return 0;
 968 }
 969
 970 /**
 971  * Implements cl_object_operations::coo_attr_get() method for an object
 972  * without stripes (LLT_EMPTY layout type).
 973  *
 974  * The only attributes this layer is authoritative in this case is
 975  * cl_attr::cat_blocks---it's 0.
 976  */
 977 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
 978                               struct cl_attr *attr)
 979 {
 980         attr->cat_blocks = 0;
 981         return 0;
 982 }
 983
 984 static int lov_attr_get_composite(const struct lu_env *env,
 985                                   struct cl_object *obj,
 986                                   struct cl_attr *attr)
 987 {
 988         struct lov_object       *lov = cl2lov(obj);
 989         struct lov_layout_entry *entry;
 990         int                      result = 0;
 991
 992         ENTRY;
 993
 994         attr->cat_size = 0;
 995         attr->cat_blocks = 0;
 996         lov_foreach_layout_entry(lov, entry) {
 997                 struct cl_attr *lov_attr = NULL;
 998                 int index = lov_layout_entry_index(lov, entry);
 999
1000                 if (!entry->lle_valid)
1001                         continue;
1002
1003                 /* PFL: This component has not been init-ed. */
1004                 if (!lsm_entry_inited(lov->lo_lsm, index))
1005                         continue;
1006
1007                 result = entry->lle_comp_ops->lco_getattr(env, lov, index,
1008                                                           entry, &lov_attr);
1009                 if (result < 0)
1010                         RETURN(result);
1011
1012                 if (lov_attr == NULL)
1013                         continue;
1014
1015                 CDEBUG(D_INODE, "COMP ID #%i: s=%llu m=%llu a=%llu c=%llu "
1016                        "b=%llu\n", index - 1, lov_attr->cat_size,
1017                        lov_attr->cat_mtime, lov_attr->cat_atime,
1018                        lov_attr->cat_ctime, lov_attr->cat_blocks);
1019
1020                 /* merge results */
1021                 attr->cat_blocks += lov_attr->cat_blocks;
1022                 if (attr->cat_size < lov_attr->cat_size)
1023                         attr->cat_size = lov_attr->cat_size;
1024                 if (attr->cat_kms < lov_attr->cat_kms)
1025                         attr->cat_kms = lov_attr->cat_kms;
1026                 if (attr->cat_atime < lov_attr->cat_atime)
1027                         attr->cat_atime = lov_attr->cat_atime;
1028                 if (attr->cat_ctime < lov_attr->cat_ctime)
1029                         attr->cat_ctime = lov_attr->cat_ctime;
1030                 if (attr->cat_mtime < lov_attr->cat_mtime)
1031                         attr->cat_mtime = lov_attr->cat_mtime;
1032         }
1033
1034         RETURN(0);
1035 }
1036
1037 static int lov_flush_composite(const struct lu_env *env,
1038                                struct cl_object *obj,
1039                                struct ldlm_lock *lock)
1040 {
1041         struct lov_object *lov = cl2lov(obj);
1042         struct lov_layout_entry *lle;
1043         int rc = -ENODATA;
1044
1045         ENTRY;
1046
1047         lov_foreach_layout_entry(lov, lle) {
1048                 if (!lsme_is_dom(lle->lle_lsme))
1049                         continue;
1050                 rc = cl_object_flush(env, lovsub2cl(lle->lle_dom.lo_dom), lock);
1051                 break;
1052         }
1053
1054         RETURN(rc);
1055 }
1056
1057 static int lov_flush_empty(const struct lu_env *env, struct cl_object *obj,
1058                            struct ldlm_lock *lock)
1059 {
1060         return 0;
1061 }
1062
1063 const static struct lov_layout_operations lov_dispatch[] = {
1064         [LLT_EMPTY] = {
1065                 .llo_init      = lov_init_empty,
1066                 .llo_delete    = lov_delete_empty,
1067                 .llo_fini      = lov_fini_empty,
1068                 .llo_print     = lov_print_empty,
1069                 .llo_page_init = lov_page_init_empty,
1070                 .llo_lock_init = lov_lock_init_empty,
1071                 .llo_io_init   = lov_io_init_empty,
1072                 .llo_getattr   = lov_attr_get_empty,
1073                 .llo_flush     = lov_flush_empty,
1074         },
1075         [LLT_RELEASED] = {
1076                 .llo_init      = lov_init_released,
1077                 .llo_delete    = lov_delete_empty,
1078                 .llo_fini      = lov_fini_released,
1079                 .llo_print     = lov_print_released,
1080                 .llo_page_init = lov_page_init_empty,
1081                 .llo_lock_init = lov_lock_init_empty,
1082                 .llo_io_init   = lov_io_init_released,
1083                 .llo_getattr   = lov_attr_get_empty,
1084                 .llo_flush     = lov_flush_empty,
1085         },
1086         [LLT_COMP] = {
1087                 .llo_init      = lov_init_composite,
1088                 .llo_delete    = lov_delete_composite,
1089                 .llo_fini      = lov_fini_composite,
1090                 .llo_print     = lov_print_composite,
1091                 .llo_page_init = lov_page_init_composite,
1092                 .llo_lock_init = lov_lock_init_composite,
1093                 .llo_io_init   = lov_io_init_composite,
1094                 .llo_getattr   = lov_attr_get_composite,
1095                 .llo_flush     = lov_flush_composite,
1096         },
1097         [LLT_FOREIGN] = {
1098                 .llo_init      = lov_init_foreign,
1099                 .llo_delete    = lov_delete_empty,
1100                 .llo_fini      = lov_fini_released,
1101                 .llo_print     = lov_print_foreign,
1102                 .llo_page_init = lov_page_init_foreign,
1103                 .llo_lock_init = lov_lock_init_empty,
1104                 .llo_io_init   = lov_io_init_empty,
1105                 .llo_getattr   = lov_attr_get_empty,
1106                 .llo_flush     = lov_flush_empty,
1107         },
1108 };
1109
1110 /**
1111  * Performs a double-dispatch based on the layout type of an object.
1112  */
1113 #define LOV_2DISPATCH_NOLOCK(obj, op, ...)              \
1114 ({                                                      \
1115         struct lov_object *__obj = (obj);               \
1116         enum lov_layout_type __llt;                     \
1117                                                         \
1118         __llt = __obj->lo_type;                         \
1119         LASSERT(__llt < ARRAY_SIZE(lov_dispatch));      \
1120         lov_dispatch[__llt].op(__VA_ARGS__);            \
1121 })
1122
1123 /**
1124  * Return lov_layout_type associated with a given lsm
1125  */
1126 static enum lov_layout_type lov_type(struct lov_stripe_md *lsm)
1127 {
1128         if (lsm == NULL)
1129                 return LLT_EMPTY;
1130
1131         if (lsm->lsm_is_released)
1132                 return LLT_RELEASED;
1133
1134         if (lsm->lsm_magic == LOV_MAGIC_V1 ||
1135             lsm->lsm_magic == LOV_MAGIC_V3 ||
1136             lsm->lsm_magic == LOV_MAGIC_COMP_V1)
1137                 return LLT_COMP;
1138
1139         if (lsm->lsm_magic == LOV_MAGIC_FOREIGN)
1140                 return LLT_FOREIGN;
1141
1142         return LLT_EMPTY;
1143 }
1144
1145 static inline void lov_conf_freeze(struct lov_object *lov)
1146 {
1147         CDEBUG(D_INODE, "To take share lov(%p) owner %p/%p\n",
1148                 lov, lov->lo_owner, current);
1149         if (lov->lo_owner != current)
1150                 down_read(&lov->lo_type_guard);
1151 }
1152
1153 static inline void lov_conf_thaw(struct lov_object *lov)
1154 {
1155         CDEBUG(D_INODE, "To release share lov(%p) owner %p/%p\n",
1156                 lov, lov->lo_owner, current);
1157         if (lov->lo_owner != current)
1158                 up_read(&lov->lo_type_guard);
1159 }
1160
1161 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...)                       \
1162 ({                                                                      \
1163         struct lov_object                      *__obj = (obj);          \
1164         int                                     __lock = !!(lock);      \
1165         typeof(lov_dispatch[0].op(__VA_ARGS__)) __result;               \
1166                                                                         \
1167         if (__lock)                                                     \
1168                 lov_conf_freeze(__obj);                                 \
1169         __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__);          \
1170         if (__lock)                                                     \
1171                 lov_conf_thaw(__obj);                                   \
1172         __result;                                                       \
1173 })
1174
1175 /**
1176  * Performs a locked double-dispatch based on the layout type of an object.
1177  */
1178 #define LOV_2DISPATCH(obj, op, ...)                     \
1179         LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
1180
1181 #define LOV_2DISPATCH_VOID(obj, op, ...)                                \
1182 do {                                                                    \
1183         struct lov_object                      *__obj = (obj);          \
1184         enum lov_layout_type                    __llt;                  \
1185                                                                         \
1186         lov_conf_freeze(__obj);                                         \
1187         __llt = __obj->lo_type;                                         \
1188         LASSERT(__llt < ARRAY_SIZE(lov_dispatch));                      \
1189         lov_dispatch[__llt].op(__VA_ARGS__);                            \
1190         lov_conf_thaw(__obj);                                           \
1191 } while (0)
1192
1193 static void lov_conf_lock(struct lov_object *lov)
1194 {
1195         LASSERT(lov->lo_owner != current);
1196         down_write(&lov->lo_type_guard);
1197         LASSERT(lov->lo_owner == NULL);
1198         lov->lo_owner = current;
1199         CDEBUG(D_INODE, "Took exclusive lov(%p) owner %p\n",
1200                 lov, lov->lo_owner);
1201 }
1202
1203 static void lov_conf_unlock(struct lov_object *lov)
1204 {
1205         CDEBUG(D_INODE, "To release exclusive lov(%p) owner %p\n",
1206                 lov, lov->lo_owner);
1207         lov->lo_owner = NULL;
1208         up_write(&lov->lo_type_guard);
1209 }
1210
1211 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov)
1212 {
1213         ENTRY;
1214
1215         while (atomic_read(&lov->lo_active_ios) > 0) {
1216                 CDEBUG(D_INODE, "file:"DFID" wait for active IO, now: %d.\n",
1217                         PFID(lu_object_fid(lov2lu(lov))),
1218                         atomic_read(&lov->lo_active_ios));
1219
1220                 wait_event_idle(lov->lo_waitq,
1221                                 atomic_read(&lov->lo_active_ios) == 0);
1222         }
1223         RETURN(0);
1224 }
1225
1226 static int lov_layout_change(const struct lu_env *unused,
1227                              struct lov_object *lov, struct lov_stripe_md *lsm,
1228                              const struct cl_object_conf *conf)
1229 {
1230         enum lov_layout_type llt = lov_type(lsm);
1231         union lov_layout_state *state = &lov->u;
1232         const struct lov_layout_operations *old_ops;
1233         const struct lov_layout_operations *new_ops;
1234         struct lov_device *lov_dev = lov_object_dev(lov);
1235         struct lu_env *env;
1236         __u16 refcheck;
1237         int rc;
1238         ENTRY;
1239
1240         LASSERT(lov->lo_type < ARRAY_SIZE(lov_dispatch));
1241
1242         env = cl_env_get(&refcheck);
1243         if (IS_ERR(env))
1244                 RETURN(PTR_ERR(env));
1245
1246         LASSERT(llt < ARRAY_SIZE(lov_dispatch));
1247
1248         CDEBUG(D_INODE, DFID" from %s to %s\n",
1249                PFID(lu_object_fid(lov2lu(lov))),
1250                llt2str(lov->lo_type), llt2str(llt));
1251
1252         old_ops = &lov_dispatch[lov->lo_type];
1253         new_ops = &lov_dispatch[llt];
1254
1255         rc = cl_object_prune(env, &lov->lo_cl);
1256         if (rc != 0)
1257                 GOTO(out, rc);
1258
1259         rc = old_ops->llo_delete(env, lov, &lov->u);
1260         if (rc != 0)
1261                 GOTO(out, rc);
1262
1263         old_ops->llo_fini(env, lov, &lov->u);
1264
1265         LASSERT(atomic_read(&lov->lo_active_ios) == 0);
1266
1267         CDEBUG(D_INODE, DFID "Apply new layout lov %p, type %d\n",
1268                PFID(lu_object_fid(lov2lu(lov))), lov, llt);
1269
1270         /* page bufsize fixup */
1271         cl_object_header(&lov->lo_cl)->coh_page_bufsize -=
1272                 lov_page_slice_fixup(lov, NULL);
1273
1274         lov->lo_type = llt;
1275         rc = new_ops->llo_init(env, lov_dev, lov, lsm, conf, state);
1276         if (rc != 0) {
1277                 struct obd_device *obd = lov2obd(lov_dev->ld_lov);
1278
1279                 CERROR("%s: cannot apply new layout on "DFID" : rc = %d\n",
1280                        obd->obd_name, PFID(lu_object_fid(lov2lu(lov))), rc);
1281                 new_ops->llo_delete(env, lov, state);
1282                 new_ops->llo_fini(env, lov, state);
1283                 /* this file becomes an EMPTY file. */
1284                 lov->lo_type = LLT_EMPTY;
1285                 GOTO(out, rc);
1286         }
1287
1288 out:
1289         cl_env_put(env, &refcheck);
1290         RETURN(rc);
1291 }
1292
1293 /*****************************************************************************
1294  *
1295  * Lov object operations.
1296  *
1297  */
1298 static int lov_object_init(const struct lu_env *env, struct lu_object *obj,
1299                            const struct lu_object_conf *conf)
1300 {
1301         struct lov_object            *lov   = lu2lov(obj);
1302         struct lov_device            *dev   = lov_object_dev(lov);
1303         const struct cl_object_conf  *cconf = lu2cl_conf(conf);
1304         union lov_layout_state       *set   = &lov->u;
1305         const struct lov_layout_operations *ops;
1306         struct lov_stripe_md *lsm = NULL;
1307         int rc;
1308         ENTRY;
1309
1310         init_rwsem(&lov->lo_type_guard);
1311         atomic_set(&lov->lo_active_ios, 0);
1312         init_waitqueue_head(&lov->lo_waitq);
1313         cl_object_page_init(lu2cl(obj), sizeof(struct lov_page));
1314
1315         lov->lo_type = LLT_EMPTY;
1316         if (cconf->u.coc_layout.lb_buf != NULL) {
1317                 lsm = lov_unpackmd(dev->ld_lov,
1318                                    cconf->u.coc_layout.lb_buf,
1319                                    cconf->u.coc_layout.lb_len);
1320                 if (IS_ERR(lsm))
1321                         RETURN(PTR_ERR(lsm));
1322
1323                 dump_lsm(D_INODE, lsm);
1324         }
1325
1326         /* no locking is necessary, as object is being created */
1327         lov->lo_type = lov_type(lsm);
1328         ops = &lov_dispatch[lov->lo_type];
1329         rc = ops->llo_init(env, dev, lov, lsm, cconf, set);
1330         if (rc != 0)
1331                 GOTO(out_lsm, rc);
1332
1333 out_lsm:
1334         lov_lsm_put(lsm);
1335
1336         RETURN(rc);
1337 }
1338
1339 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
1340                         const struct cl_object_conf *conf)
1341 {
1342         struct lov_stripe_md    *lsm = NULL;
1343         struct lov_object       *lov = cl2lov(obj);
1344         int                      result = 0;
1345         ENTRY;
1346
1347         if (conf->coc_opc == OBJECT_CONF_SET &&
1348             conf->u.coc_layout.lb_buf != NULL) {
1349                 lsm = lov_unpackmd(lov_object_dev(lov)->ld_lov,
1350                                    conf->u.coc_layout.lb_buf,
1351                                    conf->u.coc_layout.lb_len);
1352                 if (IS_ERR(lsm))
1353                         RETURN(PTR_ERR(lsm));
1354                 dump_lsm(D_INODE, lsm);
1355         }
1356
1357         lov_conf_lock(lov);
1358         if (conf->coc_opc == OBJECT_CONF_INVALIDATE) {
1359                 lov->lo_layout_invalid = true;
1360                 GOTO(out, result = 0);
1361         }
1362
1363         if (conf->coc_opc == OBJECT_CONF_WAIT) {
1364                 if (lov->lo_layout_invalid &&
1365                     atomic_read(&lov->lo_active_ios) > 0) {
1366                         lov_conf_unlock(lov);
1367                         result = lov_layout_wait(env, lov);
1368                         lov_conf_lock(lov);
1369                 }
1370                 GOTO(out, result);
1371         }
1372
1373         LASSERT(conf->coc_opc == OBJECT_CONF_SET);
1374
1375         if ((lsm == NULL && lov->lo_lsm == NULL) ||
1376             ((lsm != NULL && lov->lo_lsm != NULL) &&
1377              (lov->lo_lsm->lsm_layout_gen == lsm->lsm_layout_gen) &&
1378              (lov->lo_lsm->lsm_entries[0]->lsme_pattern ==
1379               lsm->lsm_entries[0]->lsme_pattern))) {
1380                 /* same version of layout */
1381                 lov->lo_layout_invalid = false;
1382                 GOTO(out, result = 0);
1383         }
1384
1385         /* will change layout - check if there still exists active IO. */
1386         if (atomic_read(&lov->lo_active_ios) > 0) {
1387                 lov->lo_layout_invalid = true;
1388                 GOTO(out, result = -EBUSY);
1389         }
1390
1391         result = lov_layout_change(env, lov, lsm, conf);
1392         lov->lo_layout_invalid = result != 0;
1393         EXIT;
1394
1395 out:
1396         lov_conf_unlock(lov);
1397         lov_lsm_put(lsm);
1398         CDEBUG(D_INODE, DFID" lo_layout_invalid=%d\n",
1399                PFID(lu_object_fid(lov2lu(lov))), lov->lo_layout_invalid);
1400         RETURN(result);
1401 }
1402
1403 static void lov_object_delete(const struct lu_env *env, struct lu_object *obj)
1404 {
1405         struct lov_object *lov = lu2lov(obj);
1406
1407         ENTRY;
1408         LOV_2DISPATCH_VOID(lov, llo_delete, env, lov, &lov->u);
1409         EXIT;
1410 }
1411
1412 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
1413 {
1414         struct lov_object *lov = lu2lov(obj);
1415
1416         ENTRY;
1417         LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
1418         lu_object_fini(obj);
1419         OBD_SLAB_FREE_PTR(lov, lov_object_kmem);
1420         EXIT;
1421 }
1422
1423 static int lov_object_print(const struct lu_env *env, void *cookie,
1424                             lu_printer_t p, const struct lu_object *o)
1425 {
1426         return LOV_2DISPATCH_NOLOCK(lu2lov(o), llo_print, env, cookie, p, o);
1427 }
1428
1429 static int lov_page_init(const struct lu_env *env, struct cl_object *obj,
1430                          struct cl_page *page, pgoff_t index)
1431 {
1432         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_page_init, env, obj, page,
1433                                     index);
1434 }
1435
1436 /**
1437  * Implements cl_object_operations::clo_io_init() method for lov
1438  * layer. Dispatches to the appropriate layout io initialization method.
1439  */
1440 static int lov_io_init(const struct lu_env *env, struct cl_object *obj,
1441                        struct cl_io *io)
1442 {
1443         CL_IO_SLICE_CLEAN(lov_env_io(env), lis_preserved);
1444
1445         CDEBUG(D_INODE, DFID "io %p type %d ignore/verify layout %d/%d\n",
1446                PFID(lu_object_fid(&obj->co_lu)), io, io->ci_type,
1447                io->ci_ignore_layout, io->ci_verify_layout);
1448
1449         /* IO type CIT_MISC with ci_ignore_layout set are usually invoked from
1450          * the OSC layer. It shouldn't take lov layout conf lock in that case,
1451          * because as long as the OSC object exists, the layout can't be
1452          * reconfigured. */
1453         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
1454                         !(io->ci_ignore_layout && io->ci_type == CIT_MISC),
1455                         env, obj, io);
1456 }
1457
1458 /**
1459  * An implementation of cl_object_operations::clo_attr_get() method for lov
1460  * layer. For raid0 layout this collects and merges attributes of all
1461  * sub-objects.
1462  */
1463 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
1464                         struct cl_attr *attr)
1465 {
1466         /* do not take lock, as this function is called under a
1467          * spin-lock. Layout is protected from changing by ongoing IO. */
1468         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
1469 }
1470
1471 static int lov_attr_update(const struct lu_env *env, struct cl_object *obj,
1472                            const struct cl_attr *attr, unsigned valid)
1473 {
1474         /*
1475          * No dispatch is required here, as no layout implements this.
1476          */
1477         return 0;
1478 }
1479
1480 static int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
1481                   struct cl_lock *lock, const struct cl_io *io)
1482 {
1483         /* No need to lock because we've taken one refcount of layout.  */
1484         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_lock_init, env, obj, lock,
1485                                     io);
1486 }
1487
1488 /**
1489  * We calculate on which OST the mapping will end. If the length of mapping
1490  * is greater than (stripe_size * stripe_count) then the last_stripe will
1491  * will be one just before start_stripe. Else we check if the mapping
1492  * intersects each OST and find last_stripe.
1493  * This function returns the last_stripe and also sets the stripe_count
1494  * over which the mapping is spread
1495  *
1496  * \param lsm [in]              striping information for the file
1497  * \param index [in]            stripe component index
1498  * \param ext [in]              logical extent of mapping
1499  * \param start_stripe [in]     starting stripe of the mapping
1500  * \param stripe_count [out]    the number of stripes across which to map is
1501  *                              returned
1502  *
1503  * \retval last_stripe          return the last stripe of the mapping
1504  */
1505 static int fiemap_calc_last_stripe(struct lov_stripe_md *lsm, int index,
1506                                    struct lu_extent *ext,
1507                                    int start_stripe, int *stripe_count)
1508 {
1509         struct lov_stripe_md_entry *lsme = lsm->lsm_entries[index];
1510         int init_stripe;
1511         int last_stripe;
1512         int i, j;
1513
1514         init_stripe = lov_stripe_number(lsm, index, ext->e_start);
1515
1516         if (ext->e_end - ext->e_start >
1517             lsme->lsme_stripe_size * lsme->lsme_stripe_count) {
1518                 if (init_stripe == start_stripe) {
1519                         last_stripe = (start_stripe < 1) ?
1520                                 lsme->lsme_stripe_count - 1 : start_stripe - 1;
1521                         *stripe_count = lsme->lsme_stripe_count;
1522                 } else if (init_stripe < start_stripe) {
1523                         last_stripe = (init_stripe < 1) ?
1524                                 lsme->lsme_stripe_count - 1 : init_stripe - 1;
1525                         *stripe_count = lsme->lsme_stripe_count -
1526                                         (start_stripe - init_stripe);
1527                 } else {
1528                         last_stripe = init_stripe - 1;
1529                         *stripe_count = init_stripe - start_stripe;
1530                 }
1531         } else {
1532                 for (j = 0, i = start_stripe; j < lsme->lsme_stripe_count;
1533                      i = (i + 1) % lsme->lsme_stripe_count, j++) {
1534                         if (!lov_stripe_intersects(lsm, index,  i, ext, NULL,
1535                                                    NULL))
1536                                 break;
1537                         if ((start_stripe != init_stripe) && (i == init_stripe))
1538                                 break;
1539                 }
1540                 *stripe_count = j;
1541                 last_stripe = (start_stripe + j - 1) % lsme->lsme_stripe_count;
1542         }
1543
1544         return last_stripe;
1545 }
1546
1547 /**
1548  * Set fe_device and copy extents from local buffer into main return buffer.
1549  *
1550  * \param fiemap [out]          fiemap to hold all extents
1551  * \param lcl_fm_ext [in]       array of fiemap extents get from OSC layer
1552  * \param ost_index [in]        OST index to be written into the fm_device
1553  *                              field for each extent
1554  * \param ext_count [in]        number of extents to be copied
1555  * \param current_extent [in]   where to start copying in the extent array
1556  */
1557 static void fiemap_prepare_and_copy_exts(struct fiemap *fiemap,
1558                                          struct fiemap_extent *lcl_fm_ext,
1559                                          int ost_index, unsigned int ext_count,
1560                                          int current_extent, int abs_stripeno)
1561 {
1562         char            *to;
1563         unsigned int    ext;
1564
1565         for (ext = 0; ext < ext_count; ext++) {
1566                 set_fe_device_stripenr(&lcl_fm_ext[ext], ost_index,
1567                                        abs_stripeno);
1568                 lcl_fm_ext[ext].fe_flags |= FIEMAP_EXTENT_NET;
1569         }
1570
1571         /* Copy fm_extent's from fm_local to return buffer */
1572         to = (char *)fiemap + fiemap_count_to_size(current_extent);
1573         memcpy(to, lcl_fm_ext, ext_count * sizeof(struct fiemap_extent));
1574 }
1575
1576 #define FIEMAP_BUFFER_SIZE 4096
1577
1578 /**
1579  * Non-zero fe_logical indicates that this is a continuation FIEMAP
1580  * call. The local end offset and the device are sent in the first
1581  * fm_extent. This function calculates the stripe number from the index.
1582  * This function returns a stripe_no on which mapping is to be restarted.
1583  *
1584  * This function returns fm_end_offset which is the in-OST offset at which
1585  * mapping should be restarted. If fm_end_offset=0 is returned then caller
1586  * will re-calculate proper offset in next stripe.
1587  * Note that the first extent is passed to lov_get_info via the value field.
1588  *
1589  * \param fiemap [in]           fiemap request header
1590  * \param lsm [in]              striping information for the file
1591  * \param index [in]            stripe component index
1592  * \param ext [in]              logical extent of mapping
1593  * \param start_stripe [out]    starting stripe will be returned in this
1594  */
1595 static u64 fiemap_calc_fm_end_offset(struct fiemap *fiemap,
1596                                      struct lov_stripe_md *lsm,
1597                                      int index, struct lu_extent *ext,
1598                                      int *start_stripe)
1599 {
1600         struct lov_stripe_md_entry *lsme = lsm->lsm_entries[index];
1601         u64 local_end = fiemap->fm_extents[0].fe_logical;
1602         u64 lun_end;
1603         u64 fm_end_offset;
1604         int stripe_no = -1;
1605
1606         if (fiemap->fm_extent_count == 0 ||
1607             fiemap->fm_extents[0].fe_logical == 0)
1608                 return 0;
1609
1610         stripe_no = *start_stripe;
1611
1612         if (stripe_no == -1)
1613                 return -EINVAL;
1614
1615         /* If we have finished mapping on previous device, shift logical
1616          * offset to start of next device */
1617         if (lov_stripe_intersects(lsm, index, stripe_no, ext, NULL, &lun_end) &&
1618             local_end < lun_end) {
1619                 fm_end_offset = local_end;
1620         } else {
1621                 /* This is a special value to indicate that caller should
1622                  * calculate offset in next stripe. */
1623                 fm_end_offset = 0;
1624                 *start_stripe = (stripe_no + 1) % lsme->lsme_stripe_count;
1625         }
1626
1627         return fm_end_offset;
1628 }
1629
1630 struct fiemap_state {
1631         struct fiemap           *fs_fm;
1632         struct lu_extent        fs_ext;         /* current entry extent */
1633         u64                     fs_length;
1634         u64                     fs_end_offset;  /* last iteration offset */
1635         int                     fs_cur_extent;  /* collected exts so far */
1636         int                     fs_cnt_need;    /* # of extents buf can hold */
1637         int                     fs_start_stripe;
1638         int                     fs_last_stripe;
1639         bool                    fs_device_done; /* enough for this OST */
1640         bool                    fs_finish_stripe; /* reached fs_last_stripe */
1641         bool                    fs_enough;      /* enough for this call */
1642 };
1643
1644 static struct cl_object *lov_find_subobj(const struct lu_env *env,
1645                                          struct lov_object *lov,
1646                                          struct lov_stripe_md *lsm,
1647                                          int index)
1648 {
1649         struct lov_device       *dev = lu2lov_dev(lov2lu(lov)->lo_dev);
1650         struct lov_thread_info  *lti = lov_env_info(env);
1651         struct lu_fid           *ofid = &lti->lti_fid;
1652         struct lov_oinfo        *oinfo;
1653         struct cl_device        *subdev;
1654         int                     entry = lov_comp_entry(index);
1655         int                     stripe = lov_comp_stripe(index);
1656         int                     ost_idx;
1657         int                     rc;
1658         struct cl_object        *result;
1659
1660         if (lov->lo_type != LLT_COMP)
1661                 GOTO(out, result = NULL);
1662
1663         if (entry >= lsm->lsm_entry_count ||
1664             stripe >= lsm->lsm_entries[entry]->lsme_stripe_count)
1665                 GOTO(out, result = NULL);
1666
1667         oinfo = lsm->lsm_entries[entry]->lsme_oinfo[stripe];
1668         ost_idx = oinfo->loi_ost_idx;
1669         rc = ostid_to_fid(ofid, &oinfo->loi_oi, ost_idx);
1670         if (rc != 0)
1671                 GOTO(out, result = NULL);
1672
1673         subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
1674         result = lov_sub_find(env, subdev, ofid, NULL);
1675 out:
1676         if (result == NULL)
1677                 result = ERR_PTR(-EINVAL);
1678         return result;
1679 }
1680
1681 static int fiemap_for_stripe(const struct lu_env *env, struct cl_object *obj,
1682                              struct lov_stripe_md *lsm, struct fiemap *fiemap,
1683                              size_t *buflen, struct ll_fiemap_info_key *fmkey,
1684                              int index, int stripe_last, int stripeno,
1685                              struct fiemap_state *fs)
1686 {
1687         struct lov_stripe_md_entry *lsme = lsm->lsm_entries[index];
1688         struct cl_object *subobj;
1689         struct lov_obd *lov = lu2lov_dev(obj->co_lu.lo_dev)->ld_lov;
1690         struct fiemap_extent *fm_ext = &fs->fs_fm->fm_extents[0];
1691         u64 req_fm_len; /* max requested extent coverage */
1692         u64 len_mapped_single_call;
1693         u64 obd_start;
1694         u64 obd_end;
1695         unsigned int ext_count;
1696         /* EOF for object */
1697         bool ost_eof = false;
1698         /* done with required mapping for this OST? */
1699         bool ost_done = false;
1700         int ost_index;
1701         int rc = 0;
1702
1703         fs->fs_device_done = false;
1704         /* Find out range of mapping on this stripe */
1705         if ((lov_stripe_intersects(lsm, index, stripeno, &fs->fs_ext,
1706                                    &obd_start, &obd_end)) == 0)
1707                 return 0;
1708
1709         if (lov_oinfo_is_dummy(lsme->lsme_oinfo[stripeno]))
1710                 return -EIO;
1711
1712         /* If this is a continuation FIEMAP call and we are on
1713          * starting stripe then obd_start needs to be set to
1714          * end_offset */
1715         if (fs->fs_end_offset != 0 && stripeno == fs->fs_start_stripe)
1716                 obd_start = fs->fs_end_offset;
1717
1718         if (lov_size_to_stripe(lsm, index, fs->fs_ext.e_end, stripeno) ==
1719             obd_start)
1720                 return 0;
1721
1722         req_fm_len = obd_end - obd_start + 1;
1723         fs->fs_fm->fm_length = 0;
1724         len_mapped_single_call = 0;
1725
1726         /* find lobsub object */
1727         subobj = lov_find_subobj(env, cl2lov(obj), lsm,
1728                                  lov_comp_index(index, stripeno));
1729         if (IS_ERR(subobj))
1730                 return PTR_ERR(subobj);
1731         /* If the output buffer is very large and the objects have many
1732          * extents we may need to loop on a single OST repeatedly */
1733         do {
1734                 if (fiemap->fm_extent_count > 0) {
1735                         /* Don't get too many extents. */
1736                         if (fs->fs_cur_extent + fs->fs_cnt_need >
1737                             fiemap->fm_extent_count)
1738                                 fs->fs_cnt_need = fiemap->fm_extent_count -
1739                                                   fs->fs_cur_extent;
1740                 }
1741
1742                 obd_start += len_mapped_single_call;
1743                 fs->fs_fm->fm_length = req_fm_len - len_mapped_single_call;
1744                 req_fm_len = fs->fs_fm->fm_length;
1745                 /**
1746                  * If we've collected enough extent map, we'd request 1 more,
1747                  * to see whether we coincidentally finished all available
1748                  * extent map, so that FIEMAP_EXTENT_LAST would be set.
1749                  */
1750                 fs->fs_fm->fm_extent_count = fs->fs_enough ?
1751                                              1 : fs->fs_cnt_need;
1752                 fs->fs_fm->fm_mapped_extents = 0;
1753                 fs->fs_fm->fm_flags = fiemap->fm_flags;
1754
1755                 ost_index = lsme->lsme_oinfo[stripeno]->loi_ost_idx;
1756
1757                 if (ost_index < 0 || ost_index >= lov->desc.ld_tgt_count)
1758                         GOTO(obj_put, rc = -EINVAL);
1759                 /* If OST is inactive, return extent with UNKNOWN flag. */
1760                 if (!lov->lov_tgts[ost_index]->ltd_active) {
1761                         fs->fs_fm->fm_flags |= FIEMAP_EXTENT_LAST;
1762                         fs->fs_fm->fm_mapped_extents = 1;
1763
1764                         fm_ext[0].fe_logical = obd_start;
1765                         fm_ext[0].fe_length = obd_end - obd_start + 1;
1766                         fm_ext[0].fe_flags |= FIEMAP_EXTENT_UNKNOWN;
1767
1768                         goto inactive_tgt;
1769                 }
1770
1771                 fs->fs_fm->fm_start = obd_start;
1772                 fs->fs_fm->fm_flags &= ~FIEMAP_FLAG_DEVICE_ORDER;
1773                 memcpy(&fmkey->lfik_fiemap, fs->fs_fm, sizeof(*fs->fs_fm));
1774                 *buflen = fiemap_count_to_size(fs->fs_fm->fm_extent_count);
1775
1776                 rc = cl_object_fiemap(env, subobj, fmkey, fs->fs_fm, buflen);
1777                 if (rc != 0)
1778                         GOTO(obj_put, rc);
1779 inactive_tgt:
1780                 ext_count = fs->fs_fm->fm_mapped_extents;
1781                 if (ext_count == 0) {
1782                         ost_done = true;
1783                         fs->fs_device_done = true;
1784                         /* If last stripe has hold at the end,
1785                          * we need to return */
1786                         if (stripeno == fs->fs_last_stripe) {
1787                                 fiemap->fm_mapped_extents = 0;
1788                                 fs->fs_finish_stripe = true;
1789                                 GOTO(obj_put, rc);
1790                         }
1791                         break;
1792                 } else if (fs->fs_enough) {
1793                         /*
1794                          * We've collected enough extents and there are
1795                          * more extents after it.
1796                          */
1797                         GOTO(obj_put, rc);
1798                 }
1799
1800                 /* If we just need num of extents, got to next device */
1801                 if (fiemap->fm_extent_count == 0) {
1802                         fs->fs_cur_extent += ext_count;
1803                         break;
1804                 }
1805
1806                 /* prepare to copy retrived map extents */
1807                 len_mapped_single_call = fm_ext[ext_count - 1].fe_logical +
1808                                          fm_ext[ext_count - 1].fe_length -
1809                                          obd_start;
1810
1811                 /* Have we finished mapping on this device? */
1812                 if (req_fm_len <= len_mapped_single_call) {
1813                         ost_done = true;
1814                         fs->fs_device_done = true;
1815                 }
1816
1817                 /* Clear the EXTENT_LAST flag which can be present on
1818                  * the last extent */
1819                 if (fm_ext[ext_count - 1].fe_flags & FIEMAP_EXTENT_LAST)
1820                         fm_ext[ext_count - 1].fe_flags &= ~FIEMAP_EXTENT_LAST;
1821                 if (lov_stripe_size(lsm, index,
1822                                     fm_ext[ext_count - 1].fe_logical +
1823                                     fm_ext[ext_count - 1].fe_length,
1824                                     stripeno) >= fmkey->lfik_oa.o_size) {
1825                         ost_eof = true;
1826                         fs->fs_device_done = true;
1827                 }
1828
1829                 fiemap_prepare_and_copy_exts(fiemap, fm_ext, ost_index,
1830                                              ext_count, fs->fs_cur_extent,
1831                                              stripe_last + stripeno);
1832                 fs->fs_cur_extent += ext_count;
1833
1834                 /* Ran out of available extents? */
1835                 if (fs->fs_cur_extent >= fiemap->fm_extent_count)
1836                         fs->fs_enough = true;
1837         } while (!ost_done && !ost_eof);
1838
1839         if (stripeno == fs->fs_last_stripe)
1840                 fs->fs_finish_stripe = true;
1841 obj_put:
1842         cl_object_put(env, subobj);
1843
1844         return rc;
1845 }
1846
1847 /**
1848  * Break down the FIEMAP request and send appropriate calls to individual OSTs.
1849  * This also handles the restarting of FIEMAP calls in case mapping overflows
1850  * the available number of extents in single call.
1851  *
1852  * \param env [in]              lustre environment
1853  * \param obj [in]              file object
1854  * \param fmkey [in]            fiemap request header and other info
1855  * \param fiemap [out]          fiemap buffer holding retrived map extents
1856  * \param buflen [in/out]       max buffer length of @fiemap, when iterate
1857  *                              each OST, it is used to limit max map needed
1858  * \retval 0    success
1859  * \retval < 0  error
1860  */
1861 static int lov_object_fiemap(const struct lu_env *env, struct cl_object *obj,
1862                              struct ll_fiemap_info_key *fmkey,
1863                              struct fiemap *fiemap, size_t *buflen)
1864 {
1865         struct lov_stripe_md_entry *lsme;
1866         struct lov_stripe_md *lsm;
1867         struct fiemap *fm_local = NULL;
1868         loff_t whole_start;
1869         loff_t whole_end;
1870         int entry;
1871         int start_entry = -1;
1872         int end_entry;
1873         int cur_stripe = 0;
1874         int stripe_count;
1875         unsigned int buffer_size = FIEMAP_BUFFER_SIZE;
1876         int rc = 0;
1877         struct fiemap_state fs = { 0 };
1878         struct lu_extent range;
1879         int cur_ext;
1880         int stripe_last;
1881         int start_stripe = 0;
1882         bool resume = false;
1883         ENTRY;
1884
1885         lsm = lov_lsm_addref(cl2lov(obj));
1886         if (lsm == NULL) {
1887                 /* no extent: there is no object for mapping */
1888                 fiemap->fm_mapped_extents = 0;
1889                 return 0;
1890         }
1891
1892         if (!(fiemap->fm_flags & FIEMAP_FLAG_DEVICE_ORDER)) {
1893                 /**
1894                  * If the entry count > 1 or stripe_count > 1 and the
1895                  * application does not understand DEVICE_ORDER flag,
1896                  * it cannot interpret the extents correctly.
1897                  */
1898                 if (lsm->lsm_entry_count > 1 ||
1899                     (lsm->lsm_entry_count == 1 &&
1900                      lsm->lsm_entries[0]->lsme_stripe_count > 1))
1901                         GOTO(out_lsm, rc = -ENOTSUPP);
1902         }
1903
1904         /* No support for DOM layout yet. */
1905         if (lsme_is_dom(lsm->lsm_entries[0]))
1906                 GOTO(out_lsm, rc = -ENOTSUPP);
1907
1908         if (lsm->lsm_is_released) {
1909                 if (fiemap->fm_start < fmkey->lfik_oa.o_size) {
1910                         /**
1911                          * released file, return a minimal FIEMAP if
1912                          * request fits in file-size.
1913                          */
1914                         fiemap->fm_mapped_extents = 1;
1915                         fiemap->fm_extents[0].fe_logical = fiemap->fm_start;
1916                         if (fiemap->fm_start + fiemap->fm_length <
1917                             fmkey->lfik_oa.o_size)
1918                                 fiemap->fm_extents[0].fe_length =
1919                                         fiemap->fm_length;
1920                         else
1921                                 fiemap->fm_extents[0].fe_length =
1922                                         fmkey->lfik_oa.o_size -
1923                                         fiemap->fm_start;
1924                         fiemap->fm_extents[0].fe_flags |=
1925                                 FIEMAP_EXTENT_UNKNOWN | FIEMAP_EXTENT_LAST;
1926                 }
1927                 GOTO(out_lsm, rc = 0);
1928         }
1929
1930         /* buffer_size is small to hold fm_extent_count of extents. */
1931         if (fiemap_count_to_size(fiemap->fm_extent_count) < buffer_size)
1932                 buffer_size = fiemap_count_to_size(fiemap->fm_extent_count);
1933
1934         OBD_ALLOC_LARGE(fm_local, buffer_size);
1935         if (fm_local == NULL)
1936                 GOTO(out_lsm, rc = -ENOMEM);
1937
1938         /**
1939          * Requested extent count exceeds the fiemap buffer size, shrink our
1940          * ambition.
1941          */
1942         if (fiemap_count_to_size(fiemap->fm_extent_count) > *buflen)
1943                 fiemap->fm_extent_count = fiemap_size_to_count(*buflen);
1944
1945         fs.fs_enough = false;
1946         fs.fs_cur_extent = 0;
1947         fs.fs_fm = fm_local;
1948         fs.fs_cnt_need = fiemap_size_to_count(buffer_size);
1949
1950         whole_start = fiemap->fm_start;
1951         /* whole_start is beyond the end of the file */
1952         if (whole_start > fmkey->lfik_oa.o_size)
1953                 GOTO(out_fm_local, rc = -EINVAL);
1954         whole_end = (fiemap->fm_length == OBD_OBJECT_EOF) ?
1955                                         fmkey->lfik_oa.o_size + 1 :
1956                                         whole_start + fiemap->fm_length;
1957         /**
1958          * If fiemap->fm_length != OBD_OBJECT_EOF but whole_end exceeds file
1959          * size
1960          */
1961         if (whole_end > fmkey->lfik_oa.o_size + 1)
1962                 whole_end = fmkey->lfik_oa.o_size + 1;
1963
1964         /**
1965          * the high 16bits of fe_device remember which stripe the last
1966          * call has been arrived, we'd continue from there in this call.
1967          */
1968         if (fiemap->fm_extent_count && fiemap->fm_extents[0].fe_logical)
1969                 resume = true;
1970         stripe_last = get_fe_stripenr(&fiemap->fm_extents[0]);
1971         /**
1972          * stripe_last records stripe number we've been processed in the last
1973          * call
1974          */
1975         end_entry = lsm->lsm_entry_count - 1;
1976         cur_stripe = 0;
1977         for (entry = 0; entry <= end_entry; entry++) {
1978                 lsme = lsm->lsm_entries[entry];
1979                 if (cur_stripe + lsme->lsme_stripe_count >= stripe_last) {
1980                         start_entry = entry;
1981                         start_stripe = stripe_last - cur_stripe;
1982                         break;
1983                 }
1984
1985                 cur_stripe += lsme->lsme_stripe_count;
1986         }
1987         if (start_entry == -1) {
1988                 CERROR(DFID": FIEMAP does not init start entry, cur_stripe=%d, "
1989                        "stripe_last=%d\n", PFID(lu_object_fid(&obj->co_lu)),
1990                        cur_stripe, stripe_last);
1991                 GOTO(out_fm_local, rc = -EINVAL);
1992         }
1993         /**
1994          * @start_entry & @start_stripe records the position of fiemap
1995          * resumption @stripe_last keeps recording the absolution position
1996          * we'are processing. @resume indicates we'd honor @start_stripe.
1997          */
1998
1999         range.e_start = whole_start;
2000         range.e_end = whole_end;
2001
2002         for (entry = start_entry; entry <= end_entry; entry++) {
2003                 /* remeber to update stripe_last accordingly */
2004                 lsme = lsm->lsm_entries[entry];
2005
2006                 /* FLR could contain component holes between entries */
2007                 if (!lsme_inited(lsme)) {
2008                         stripe_last += lsme->lsme_stripe_count;
2009                         resume = false;
2010                         continue;
2011                 }
2012
2013                 if (!lu_extent_is_overlapped(&range, &lsme->lsme_extent)) {
2014                         stripe_last += lsme->lsme_stripe_count;
2015                         resume = false;
2016                         continue;
2017                 }
2018
2019                 /* prepare for a component entry iteration */
2020                 if (lsme->lsme_extent.e_start > whole_start)
2021                         fs.fs_ext.e_start = lsme->lsme_extent.e_start;
2022                 else
2023                         fs.fs_ext.e_start = whole_start;
2024                 if (lsme->lsme_extent.e_end > whole_end)
2025                         fs.fs_ext.e_end = whole_end;
2026                 else
2027                         fs.fs_ext.e_end = lsme->lsme_extent.e_end;
2028
2029                 /* Calculate start stripe, last stripe and length of mapping */
2030                 if (resume) {
2031                         fs.fs_start_stripe = start_stripe;
2032                         /* put stripe_last to the first stripe of the comp */
2033                         stripe_last -= start_stripe;
2034                         resume = false;
2035                 } else {
2036                         fs.fs_start_stripe = lov_stripe_number(lsm, entry,
2037                                                         fs.fs_ext.e_start);
2038                 }
2039                 fs.fs_last_stripe = fiemap_calc_last_stripe(lsm, entry,
2040                                         &fs.fs_ext, fs.fs_start_stripe,
2041                                         &stripe_count);
2042                 /**
2043                  * A new mirror component is under process, reset
2044                  * fs.fs_end_offset and then fiemap_for_stripe() starts from
2045                  * the overlapping extent, otherwise starts from
2046                  * fs.fs_end_offset.
2047                  */
2048                 if (entry > start_entry && lsme->lsme_extent.e_start == 0) {
2049                         /* new mirror */
2050                         fs.fs_end_offset = 0;
2051                 } else {
2052                         fs.fs_end_offset = fiemap_calc_fm_end_offset(fiemap,
2053                                                 lsm, entry, &fs.fs_ext,
2054                                                 &fs.fs_start_stripe);
2055                 }
2056
2057                 /* Check each stripe */
2058                 for (cur_stripe = fs.fs_start_stripe; stripe_count > 0;
2059                      --stripe_count,
2060                      cur_stripe = (cur_stripe + 1) % lsme->lsme_stripe_count) {
2061                         /* reset fs_finish_stripe */
2062                         fs.fs_finish_stripe = false;
2063                         rc = fiemap_for_stripe(env, obj, lsm, fiemap, buflen,
2064                                                fmkey, entry, stripe_last,
2065                                                cur_stripe, &fs);
2066                         if (rc < 0)
2067                                 GOTO(out_fm_local, rc);
2068                         if (fs.fs_enough) {
2069                                 stripe_last += cur_stripe;
2070                                 GOTO(finish, rc);
2071                         }
2072                         if (fs.fs_finish_stripe)
2073                                 break;
2074                 } /* for each stripe */
2075                 stripe_last += lsme->lsme_stripe_count;
2076         } /* for covering layout component entry */
2077
2078 finish:
2079         if (fs.fs_cur_extent > 0)
2080                 cur_ext = fs.fs_cur_extent - 1;
2081         else
2082                 cur_ext = 0;
2083
2084         /* done all the processing */
2085         if (entry > end_entry)
2086                 fiemap->fm_extents[cur_ext].fe_flags |= FIEMAP_EXTENT_LAST;
2087
2088         /* Indicate that we are returning device offsets unless file just has
2089          * single stripe */
2090         if (lsm->lsm_entry_count > 1 ||
2091             (lsm->lsm_entry_count == 1 &&
2092              lsm->lsm_entries[0]->lsme_stripe_count > 1))
2093                 fiemap->fm_flags |= FIEMAP_FLAG_DEVICE_ORDER;
2094
2095         if (fiemap->fm_extent_count == 0)
2096                 goto skip_last_device_calc;
2097
2098 skip_last_device_calc:
2099         fiemap->fm_mapped_extents = fs.fs_cur_extent;
2100 out_fm_local:
2101         OBD_FREE_LARGE(fm_local, buffer_size);
2102
2103 out_lsm:
2104         lov_lsm_put(lsm);
2105         return rc;
2106 }
2107
2108 static int lov_object_getstripe(const struct lu_env *env, struct cl_object *obj,
2109                                 struct lov_user_md __user *lum, size_t size)
2110 {
2111         struct lov_object       *lov = cl2lov(obj);
2112         struct lov_stripe_md    *lsm;
2113         int                     rc = 0;
2114         ENTRY;
2115
2116         lsm = lov_lsm_addref(lov);
2117         if (lsm == NULL)
2118                 RETURN(-ENODATA);
2119
2120         rc = lov_getstripe(env, cl2lov(obj), lsm, lum, size);
2121         lov_lsm_put(lsm);
2122         RETURN(rc);
2123 }
2124
2125 static int lov_object_layout_get(const struct lu_env *env,
2126                                  struct cl_object *obj,
2127                                  struct cl_layout *cl)
2128 {
2129         struct lov_object *lov = cl2lov(obj);
2130         struct lov_stripe_md *lsm = lov_lsm_addref(lov);
2131         struct lu_buf *buf = &cl->cl_buf;
2132         ssize_t rc;
2133         ENTRY;
2134
2135         if (lsm == NULL) {
2136                 cl->cl_size = 0;
2137                 cl->cl_layout_gen = CL_LAYOUT_GEN_EMPTY;
2138
2139                 RETURN(0);
2140         }
2141
2142         cl->cl_size = lov_comp_md_size(lsm);
2143         cl->cl_layout_gen = lsm->lsm_layout_gen;
2144         cl->cl_is_released = lsm->lsm_is_released;
2145         cl->cl_is_composite = lsm_is_composite(lsm->lsm_magic);
2146
2147         rc = lov_lsm_pack(lsm, buf->lb_buf, buf->lb_len);
2148         lov_lsm_put(lsm);
2149
2150         RETURN(rc < 0 ? rc : 0);
2151 }
2152
2153 static loff_t lov_object_maxbytes(struct cl_object *obj)
2154 {
2155         struct lov_object *lov = cl2lov(obj);
2156         struct lov_stripe_md *lsm = lov_lsm_addref(lov);
2157         loff_t maxbytes;
2158
2159         if (lsm == NULL)
2160                 return LLONG_MAX;
2161
2162         maxbytes = lsm->lsm_maxbytes;
2163
2164         lov_lsm_put(lsm);
2165
2166         return maxbytes;
2167 }
2168
2169 static int lov_object_flush(const struct lu_env *env, struct cl_object *obj,
2170                             struct ldlm_lock *lock)
2171 {
2172         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_flush, true, env, obj,
2173                                      lock);
2174 }
2175
2176 static const struct cl_object_operations lov_ops = {
2177         .coo_page_init    = lov_page_init,
2178         .coo_lock_init    = lov_lock_init,
2179         .coo_io_init      = lov_io_init,
2180         .coo_attr_get     = lov_attr_get,
2181         .coo_attr_update  = lov_attr_update,
2182         .coo_conf_set     = lov_conf_set,
2183         .coo_getstripe    = lov_object_getstripe,
2184         .coo_layout_get   = lov_object_layout_get,
2185         .coo_maxbytes     = lov_object_maxbytes,
2186         .coo_fiemap       = lov_object_fiemap,
2187         .coo_object_flush = lov_object_flush
2188 };
2189
2190 static const struct lu_object_operations lov_lu_obj_ops = {
2191         .loo_object_init        = lov_object_init,
2192         .loo_object_delete      = lov_object_delete,
2193         .loo_object_release     = NULL,
2194         .loo_object_free        = lov_object_free,
2195         .loo_object_print       = lov_object_print,
2196         .loo_object_invariant   = NULL,
2197 };
2198
2199 struct lu_object *lov_object_alloc(const struct lu_env *env,
2200                                    const struct lu_object_header *unused,
2201                                    struct lu_device *dev)
2202 {
2203         struct lov_object *lov;
2204         struct lu_object  *obj;
2205
2206         ENTRY;
2207         OBD_SLAB_ALLOC_PTR_GFP(lov, lov_object_kmem, GFP_NOFS);
2208         if (lov != NULL) {
2209                 obj = lov2lu(lov);
2210                 lu_object_init(obj, NULL, dev);
2211                 lov->lo_cl.co_ops = &lov_ops;
2212                 lov->lo_type = -1; /* invalid, to catch uninitialized type */
2213                 /*
2214                  * object io operation vector (cl_object::co_iop) is installed
2215                  * later in lov_object_init(), as different vectors are used
2216                  * for object with different layouts.
2217                  */
2218                 obj->lo_ops = &lov_lu_obj_ops;
2219         } else
2220                 obj = NULL;
2221         RETURN(obj);
2222 }
2223
2224 static struct lov_stripe_md *lov_lsm_addref(struct lov_object *lov)
2225 {
2226         struct lov_stripe_md *lsm = NULL;
2227
2228         lov_conf_freeze(lov);
2229         if (lov->lo_lsm != NULL) {
2230                 lsm = lsm_addref(lov->lo_lsm);
2231                 CDEBUG(D_INODE, "lsm %p addref %d/%d by %p.\n",
2232                         lsm, atomic_read(&lsm->lsm_refc),
2233                         lov->lo_layout_invalid, current);
2234         }
2235         lov_conf_thaw(lov);
2236         return lsm;
2237 }
2238
2239 int lov_read_and_clear_async_rc(struct cl_object *clob)
2240 {
2241         struct lu_object *luobj;
2242         int rc = 0;
2243         ENTRY;
2244
2245         luobj = lu_object_locate(&cl_object_header(clob)->coh_lu,
2246                                  &lov_device_type);
2247         if (luobj != NULL) {
2248                 struct lov_object *lov = lu2lov(luobj);
2249
2250                 lov_conf_freeze(lov);
2251                 switch (lov->lo_type) {
2252                 case LLT_COMP: {
2253                         struct lov_stripe_md *lsm;
2254                         int i;
2255
2256                         lsm = lov->lo_lsm;
2257                         LASSERT(lsm != NULL);
2258                         for (i = 0; i < lsm->lsm_entry_count; i++) {
2259                                 struct lov_stripe_md_entry *lse =
2260                                                 lsm->lsm_entries[i];
2261                                 int j;
2262
2263                                 if (!lsme_inited(lse))
2264                                         break;
2265
2266                                 for (j = 0; j < lse->lsme_stripe_count; j++) {
2267                                         struct lov_oinfo *loi =
2268                                                         lse->lsme_oinfo[j];
2269
2270                                         if (lov_oinfo_is_dummy(loi))
2271                                                 continue;
2272
2273                                         if (loi->loi_ar.ar_rc && !rc)
2274                                                 rc = loi->loi_ar.ar_rc;
2275                                         loi->loi_ar.ar_rc = 0;
2276                                 }
2277                         }
2278                 }
2279                 case LLT_RELEASED:
2280                 case LLT_EMPTY:
2281                         /* fall through */
2282                 case LLT_FOREIGN:
2283                         break;
2284                 default:
2285                         LBUG();
2286                 }
2287                 lov_conf_thaw(lov);
2288         }
2289         RETURN(rc);
2290 }
2291 EXPORT_SYMBOL(lov_read_and_clear_async_rc);
2292
2293 /** @} lov */