lustre/lov/lov_object.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2016, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * Implementation of cl_object for LOV layer.
  33  *
  34  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  35  *   Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_LOV
  39
  40 #include "lov_cl_internal.h"
  41
  42 static inline struct lov_device *lov_object_dev(struct lov_object *obj)
  43 {
  44         return lu2lov_dev(obj->lo_cl.co_lu.lo_dev);
  45 }
  46
  47 /** \addtogroup lov
  48  *  @{
  49  */
  50
  51 /*****************************************************************************
  52  *
  53  * Layout operations.
  54  *
  55  */
  56
  57 struct lov_layout_operations {
  58         int (*llo_init)(const struct lu_env *env, struct lov_device *dev,
  59                         struct lov_object *lov, struct lov_stripe_md *lsm,
  60                         const struct cl_object_conf *conf,
  61                         union lov_layout_state *state);
  62         int (*llo_delete)(const struct lu_env *env, struct lov_object *lov,
  63                            union lov_layout_state *state);
  64         void (*llo_fini)(const struct lu_env *env, struct lov_object *lov,
  65                          union lov_layout_state *state);
  66         void (*llo_install)(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 };
  80
  81 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov);
  82
  83 static void lov_lsm_put(struct lov_stripe_md *lsm)
  84 {
  85         if (lsm != NULL)
  86                 lov_free_memmd(&lsm);
  87 }
  88
  89 /*****************************************************************************
  90  *
  91  * Lov object layout operations.
  92  *
  93  */
  94
  95 static void lov_install_empty(const struct lu_env *env,
  96                               struct lov_object *lov,
  97                               union  lov_layout_state *state)
  98 {
  99         /*
 100          * File without objects.
 101          */
 102 }
 103
 104 static int lov_init_empty(const struct lu_env *env, struct lov_device *dev,
 105                           struct lov_object *lov, struct lov_stripe_md *lsm,
 106                           const struct cl_object_conf *conf,
 107                           union lov_layout_state *state)
 108 {
 109         return 0;
 110 }
 111
 112 static void lov_install_raid0(const struct lu_env *env,
 113                               struct lov_object *lov,
 114                               union  lov_layout_state *state)
 115 {
 116 }
 117
 118 static struct cl_object *lov_sub_find(const struct lu_env *env,
 119                                       struct cl_device *dev,
 120                                       const struct lu_fid *fid,
 121                                       const struct cl_object_conf *conf)
 122 {
 123         struct lu_object *o;
 124
 125         ENTRY;
 126         o = lu_object_find_at(env, cl2lu_dev(dev), fid, &conf->coc_lu);
 127         LASSERT(ergo(!IS_ERR(o), o->lo_dev->ld_type == &lovsub_device_type));
 128         RETURN(lu2cl(o));
 129 }
 130
 131 static int lov_init_sub(const struct lu_env *env, struct lov_object *lov,
 132                         struct cl_object *stripe, struct lov_layout_raid0 *r0,
 133                         int idx)
 134 {
 135         struct cl_object_header *hdr;
 136         struct cl_object_header *subhdr;
 137         struct cl_object_header *parent;
 138         struct lov_oinfo        *oinfo;
 139         int result;
 140
 141         if (OBD_FAIL_CHECK(OBD_FAIL_LOV_INIT)) {
 142                 /* For sanity:test_206.
 143                  * Do not leave the object in cache to avoid accessing
 144                  * freed memory. This is because osc_object is referring to
 145                  * lov_oinfo of lsm_stripe_data which will be freed due to
 146                  * this failure. */
 147                 cl_object_kill(env, stripe);
 148                 cl_object_put(env, stripe);
 149                 return -EIO;
 150         }
 151
 152         hdr    = cl_object_header(lov2cl(lov));
 153         subhdr = cl_object_header(stripe);
 154
 155         oinfo = lov->lo_lsm->lsm_oinfo[idx];
 156         CDEBUG(D_INODE, DFID"@%p[%d] -> "DFID"@%p: ostid: "DOSTID
 157                " idx: %d gen: %d\n",
 158                PFID(&subhdr->coh_lu.loh_fid), subhdr, idx,
 159                PFID(&hdr->coh_lu.loh_fid), hdr, POSTID(&oinfo->loi_oi),
 160                oinfo->loi_ost_idx, oinfo->loi_ost_gen);
 161
 162         /* reuse ->coh_attr_guard to protect coh_parent change */
 163         spin_lock(&subhdr->coh_attr_guard);
 164         parent = subhdr->coh_parent;
 165         if (parent == NULL) {
 166                 subhdr->coh_parent = hdr;
 167                 spin_unlock(&subhdr->coh_attr_guard);
 168                 subhdr->coh_nesting = hdr->coh_nesting + 1;
 169                 lu_object_ref_add(&stripe->co_lu, "lov-parent", lov);
 170                 r0->lo_sub[idx] = cl2lovsub(stripe);
 171                 r0->lo_sub[idx]->lso_super = lov;
 172                 r0->lo_sub[idx]->lso_index = idx;
 173                 result = 0;
 174         } else {
 175                 struct lu_object  *old_obj;
 176                 struct lov_object *old_lov;
 177                 unsigned int mask = D_INODE;
 178
 179                 spin_unlock(&subhdr->coh_attr_guard);
 180                 old_obj = lu_object_locate(&parent->coh_lu, &lov_device_type);
 181                 LASSERT(old_obj != NULL);
 182                 old_lov = cl2lov(lu2cl(old_obj));
 183                 if (old_lov->lo_layout_invalid) {
 184                         /* the object's layout has already changed but isn't
 185                          * refreshed */
 186                         lu_object_unhash(env, &stripe->co_lu);
 187                         result = -EAGAIN;
 188                 } else {
 189                         mask = D_ERROR;
 190                         result = -EIO;
 191                 }
 192
 193                 LU_OBJECT_DEBUG(mask, env, &stripe->co_lu,
 194                                 "stripe %d is already owned.", idx);
 195                 LU_OBJECT_DEBUG(mask, env, old_obj, "owned.");
 196                 LU_OBJECT_HEADER(mask, env, lov2lu(lov), "try to own.\n");
 197                 cl_object_put(env, stripe);
 198         }
 199         return result;
 200 }
 201
 202 static int lov_page_slice_fixup(struct lov_object *lov,
 203                                 struct cl_object *stripe)
 204 {
 205         struct cl_object_header *hdr = cl_object_header(&lov->lo_cl);
 206         struct cl_object *o;
 207
 208         if (stripe == NULL)
 209                 return hdr->coh_page_bufsize - lov->lo_cl.co_slice_off -
 210                        cfs_size_round(sizeof(struct lov_page));
 211
 212         cl_object_for_each(o, stripe)
 213                 o->co_slice_off += hdr->coh_page_bufsize;
 214
 215         return cl_object_header(stripe)->coh_page_bufsize;
 216 }
 217
 218 static int lov_init_raid0(const struct lu_env *env, struct lov_device *dev,
 219                           struct lov_object *lov, struct lov_stripe_md *lsm,
 220                           const struct cl_object_conf *conf,
 221                           union lov_layout_state *state)
 222 {
 223         int result;
 224         int i;
 225
 226         struct cl_object        *stripe;
 227         struct lov_thread_info  *lti     = lov_env_info(env);
 228         struct cl_object_conf   *subconf = &lti->lti_stripe_conf;
 229         struct lu_fid           *ofid    = &lti->lti_fid;
 230         struct lov_layout_raid0 *r0      = &state->raid0;
 231
 232         ENTRY;
 233
 234         if (lsm->lsm_magic != LOV_MAGIC_V1 && lsm->lsm_magic != LOV_MAGIC_V3) {
 235                 dump_lsm(D_ERROR, lsm);
 236                 LASSERTF(0, "magic mismatch, expected %d/%d, actual %d.\n",
 237                          LOV_MAGIC_V1, LOV_MAGIC_V3, lsm->lsm_magic);
 238         }
 239
 240         LASSERT(lov->lo_lsm == NULL);
 241         lov->lo_lsm = lsm_addref(lsm);
 242         r0->lo_nr = lsm->lsm_stripe_count;
 243         LASSERT(r0->lo_nr <= lov_targets_nr(dev));
 244
 245         lov->lo_layout_invalid = true;
 246
 247         OBD_ALLOC_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 248         if (r0->lo_sub != NULL) {
 249                 int psz = 0;
 250
 251                 result = 0;
 252                 subconf->coc_inode = conf->coc_inode;
 253                 spin_lock_init(&r0->lo_sub_lock);
 254                 /*
 255                  * Create stripe cl_objects.
 256                  */
 257                 for (i = 0; i < r0->lo_nr && result == 0; ++i) {
 258                         struct cl_device *subdev;
 259                         struct lov_oinfo *oinfo = lsm->lsm_oinfo[i];
 260                         int ost_idx = oinfo->loi_ost_idx;
 261
 262                         if (lov_oinfo_is_dummy(oinfo))
 263                                 continue;
 264
 265                         result = ostid_to_fid(ofid, &oinfo->loi_oi,
 266                                               oinfo->loi_ost_idx);
 267                         if (result != 0)
 268                                 GOTO(out, result);
 269
 270                         if (dev->ld_target[ost_idx] == NULL) {
 271                                 CERROR("%s: OST %04x is not initialized\n",
 272                                        lov2obd(dev->ld_lov)->obd_name, ost_idx);
 273                                 GOTO(out, result = -EIO);
 274                         }
 275
 276                         subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
 277                         subconf->u.coc_oinfo = oinfo;
 278                         LASSERTF(subdev != NULL, "not init ost %d\n", ost_idx);
 279                         /* In the function below, .hs_keycmp resolves to
 280                          * lu_obj_hop_keycmp() */
 281                         /* coverity[overrun-buffer-val] */
 282                         stripe = lov_sub_find(env, subdev, ofid, subconf);
 283                         if (!IS_ERR(stripe)) {
 284                                 result = lov_init_sub(env, lov, stripe, r0, i);
 285                                 if (result == -EAGAIN) { /* try again */
 286                                         --i;
 287                                         result = 0;
 288                                         continue;
 289                                 }
 290                         } else {
 291                                 result = PTR_ERR(stripe);
 292                         }
 293
 294                         if (result == 0) {
 295                                 int sz = lov_page_slice_fixup(lov, stripe);
 296                                 LASSERT(ergo(psz > 0, psz == sz));
 297                                 psz = sz;
 298                         }
 299                 }
 300                 if (result == 0)
 301                         cl_object_header(&lov->lo_cl)->coh_page_bufsize += psz;
 302         } else
 303                 result = -ENOMEM;
 304 out:
 305         RETURN(result);
 306 }
 307
 308 static int lov_init_released(const struct lu_env *env,
 309                              struct lov_device *dev, struct lov_object *lov,
 310                              struct lov_stripe_md *lsm,
 311                              const struct cl_object_conf *conf,
 312                              union lov_layout_state *state)
 313 {
 314         LASSERT(lsm != NULL);
 315         LASSERT(lsm_is_released(lsm));
 316         LASSERT(lov->lo_lsm == NULL);
 317
 318         lov->lo_lsm = lsm_addref(lsm);
 319         return 0;
 320 }
 321
 322 static struct cl_object *lov_find_subobj(const struct lu_env *env,
 323                                          struct lov_object *lov,
 324                                          struct lov_stripe_md *lsm,
 325                                          int stripe_idx)
 326 {
 327         struct lov_device       *dev = lu2lov_dev(lov2lu(lov)->lo_dev);
 328         struct lov_oinfo        *oinfo = lsm->lsm_oinfo[stripe_idx];
 329         struct lov_thread_info  *lti = lov_env_info(env);
 330         struct lu_fid           *ofid = &lti->lti_fid;
 331         struct cl_device        *subdev;
 332         int                     ost_idx;
 333         int                     rc;
 334         struct cl_object        *result;
 335
 336         if (lov->lo_type != LLT_RAID0)
 337                 GOTO(out, result = NULL);
 338
 339         ost_idx = oinfo->loi_ost_idx;
 340         rc = ostid_to_fid(ofid, &oinfo->loi_oi, ost_idx);
 341         if (rc != 0)
 342                 GOTO(out, result = NULL);
 343
 344         subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
 345         result = lov_sub_find(env, subdev, ofid, NULL);
 346 out:
 347         if (result == NULL)
 348                 result = ERR_PTR(-EINVAL);
 349         return result;
 350 }
 351
 352 static int lov_delete_empty(const struct lu_env *env, struct lov_object *lov,
 353                             union lov_layout_state *state)
 354 {
 355         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 356
 357         lov_layout_wait(env, lov);
 358         return 0;
 359 }
 360
 361 static void lov_subobject_kill(const struct lu_env *env, struct lov_object *lov,
 362                                struct lovsub_object *los, int idx)
 363 {
 364         struct cl_object        *sub;
 365         struct lov_layout_raid0 *r0;
 366         struct lu_site          *site;
 367         struct lu_site_bkt_data *bkt;
 368         wait_queue_t          *waiter;
 369
 370         r0  = &lov->u.raid0;
 371         LASSERT(r0->lo_sub[idx] == los);
 372
 373         sub  = lovsub2cl(los);
 374         site = sub->co_lu.lo_dev->ld_site;
 375         bkt  = lu_site_bkt_from_fid(site, &sub->co_lu.lo_header->loh_fid);
 376
 377         cl_object_kill(env, sub);
 378         /* release a reference to the sub-object and ... */
 379         lu_object_ref_del(&sub->co_lu, "lov-parent", lov);
 380         cl_object_put(env, sub);
 381
 382         /* ... wait until it is actually destroyed---sub-object clears its
 383          * ->lo_sub[] slot in lovsub_object_fini() */
 384         if (r0->lo_sub[idx] == los) {
 385                 waiter = &lov_env_info(env)->lti_waiter;
 386                 init_waitqueue_entry(waiter, current);
 387                 add_wait_queue(&bkt->lsb_marche_funebre, waiter);
 388                 set_current_state(TASK_UNINTERRUPTIBLE);
 389                 while (1) {
 390                         /* this wait-queue is signaled at the end of
 391                          * lu_object_free(). */
 392                         set_current_state(TASK_UNINTERRUPTIBLE);
 393                         spin_lock(&r0->lo_sub_lock);
 394                         if (r0->lo_sub[idx] == los) {
 395                                 spin_unlock(&r0->lo_sub_lock);
 396                                 schedule();
 397                         } else {
 398                                 spin_unlock(&r0->lo_sub_lock);
 399                                 set_current_state(TASK_RUNNING);
 400                                 break;
 401                         }
 402                 }
 403                 remove_wait_queue(&bkt->lsb_marche_funebre, waiter);
 404         }
 405         LASSERT(r0->lo_sub[idx] == NULL);
 406 }
 407
 408 static int lov_delete_raid0(const struct lu_env *env, struct lov_object *lov,
 409                             union lov_layout_state *state)
 410 {
 411         struct lov_layout_raid0 *r0 = &state->raid0;
 412         struct lov_stripe_md    *lsm = lov->lo_lsm;
 413         int i;
 414
 415         ENTRY;
 416
 417         dump_lsm(D_INODE, lsm);
 418
 419         lov_layout_wait(env, lov);
 420         if (r0->lo_sub != NULL) {
 421                 for (i = 0; i < r0->lo_nr; ++i) {
 422                         struct lovsub_object *los = r0->lo_sub[i];
 423
 424                         if (los != NULL) {
 425                                 cl_object_prune(env, &los->lso_cl);
 426                                 /*
 427                                  * If top-level object is to be evicted from
 428                                  * the cache, so are its sub-objects.
 429                                  */
 430                                 lov_subobject_kill(env, lov, los, i);
 431                         }
 432                 }
 433         }
 434         RETURN(0);
 435 }
 436
 437 static void lov_fini_empty(const struct lu_env *env, struct lov_object *lov,
 438                            union lov_layout_state *state)
 439 {
 440         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 441 }
 442
 443 static void lov_fini_raid0(const struct lu_env *env, struct lov_object *lov,
 444                            union lov_layout_state *state)
 445 {
 446         struct lov_layout_raid0 *r0 = &state->raid0;
 447         ENTRY;
 448
 449         if (r0->lo_sub != NULL) {
 450                 OBD_FREE_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 451                 r0->lo_sub = NULL;
 452         }
 453
 454         dump_lsm(D_INODE, lov->lo_lsm);
 455         lov_free_memmd(&lov->lo_lsm);
 456
 457         EXIT;
 458 }
 459
 460 static void lov_fini_released(const struct lu_env *env, struct lov_object *lov,
 461                                 union lov_layout_state *state)
 462 {
 463         ENTRY;
 464         dump_lsm(D_INODE, lov->lo_lsm);
 465         lov_free_memmd(&lov->lo_lsm);
 466         EXIT;
 467 }
 468
 469 static int lov_print_empty(const struct lu_env *env, void *cookie,
 470                            lu_printer_t p, const struct lu_object *o)
 471 {
 472         (*p)(env, cookie, "empty %d\n", lu2lov(o)->lo_layout_invalid);
 473         return 0;
 474 }
 475
 476 static int lov_print_raid0(const struct lu_env *env, void *cookie,
 477                            lu_printer_t p, const struct lu_object *o)
 478 {
 479         struct lov_object       *lov = lu2lov(o);
 480         struct lov_layout_raid0 *r0  = lov_r0(lov);
 481         struct lov_stripe_md    *lsm = lov->lo_lsm;
 482         int                      i;
 483
 484         (*p)(env, cookie, "stripes: %d, %s, lsm{%p 0x%08X %d %u %u}:\n",
 485                 r0->lo_nr, lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 486                 lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 487                 lsm->lsm_stripe_count, lsm->lsm_layout_gen);
 488         for (i = 0; i < r0->lo_nr; ++i) {
 489                 struct lu_object *sub;
 490
 491                 if (r0->lo_sub[i] != NULL) {
 492                         sub = lovsub2lu(r0->lo_sub[i]);
 493                         lu_object_print(env, cookie, p, sub);
 494                 } else {
 495                         (*p)(env, cookie, "sub %d absent\n", i);
 496                 }
 497         }
 498         return 0;
 499 }
 500
 501 static int lov_print_released(const struct lu_env *env, void *cookie,
 502                                 lu_printer_t p, const struct lu_object *o)
 503 {
 504         struct lov_object       *lov = lu2lov(o);
 505         struct lov_stripe_md    *lsm = lov->lo_lsm;
 506
 507         (*p)(env, cookie,
 508                 "released: %s, lsm{%p 0x%08X %d %u %u}:\n",
 509                 lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 510                 lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 511                 lsm->lsm_stripe_count, lsm->lsm_layout_gen);
 512         return 0;
 513 }
 514
 515 /**
 516  * Implements cl_object_operations::coo_attr_get() method for an object
 517  * without stripes (LLT_EMPTY layout type).
 518  *
 519  * The only attributes this layer is authoritative in this case is
 520  * cl_attr::cat_blocks---it's 0.
 521  */
 522 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
 523                               struct cl_attr *attr)
 524 {
 525         attr->cat_blocks = 0;
 526         return 0;
 527 }
 528
 529 static int lov_attr_get_raid0(const struct lu_env *env, struct cl_object *obj,
 530                               struct cl_attr *attr)
 531 {
 532         struct lov_object       *lov = cl2lov(obj);
 533         struct lov_layout_raid0 *r0 = lov_r0(lov);
 534         struct cl_attr          *lov_attr = &r0->lo_attr;
 535         int                      result = 0;
 536
 537         ENTRY;
 538
 539         /* this is called w/o holding type guard mutex, so it must be inside
 540          * an on going IO otherwise lsm may be replaced.
 541          * LU-2117: it turns out there exists one exception. For mmaped files,
 542          * the lock of those files may be requested in the other file's IO
 543          * context, and this function is called in ccc_lock_state(), it will
 544          * hit this assertion.
 545          * Anyway, it's still okay to call attr_get w/o type guard as layout
 546          * can't go if locks exist. */
 547         /* LASSERT(atomic_read(&lsm->lsm_refc) > 1); */
 548
 549         if (!r0->lo_attr_valid) {
 550                 struct lov_stripe_md    *lsm = lov->lo_lsm;
 551                 struct ost_lvb          *lvb = &lov_env_info(env)->lti_lvb;
 552                 __u64                    kms = 0;
 553
 554                 memset(lvb, 0, sizeof(*lvb));
 555                 /* XXX: timestamps can be negative by sanity:test_39m,
 556                  * how can it be? */
 557                 lvb->lvb_atime = LLONG_MIN;
 558                 lvb->lvb_ctime = LLONG_MIN;
 559                 lvb->lvb_mtime = LLONG_MIN;
 560
 561                 /*
 562                  * XXX that should be replaced with a loop over sub-objects,
 563                  * doing cl_object_attr_get() on them. But for now, let's
 564                  * reuse old lov code.
 565                  */
 566
 567                 /*
 568                  * XXX take lsm spin-lock to keep lov_merge_lvb_kms()
 569                  * happy. It's not needed, because new code uses
 570                  * ->coh_attr_guard spin-lock to protect consistency of
 571                  * sub-object attributes.
 572                  */
 573                 lov_stripe_lock(lsm);
 574                 result = lov_merge_lvb_kms(lsm, lvb, &kms);
 575                 lov_stripe_unlock(lsm);
 576                 if (result == 0) {
 577                         cl_lvb2attr(lov_attr, lvb);
 578                         lov_attr->cat_kms = kms;
 579                         r0->lo_attr_valid = 1;
 580                 }
 581         }
 582         if (result == 0) { /* merge results */
 583                 attr->cat_blocks = lov_attr->cat_blocks;
 584                 attr->cat_size = lov_attr->cat_size;
 585                 attr->cat_kms = lov_attr->cat_kms;
 586                 if (attr->cat_atime < lov_attr->cat_atime)
 587                         attr->cat_atime = lov_attr->cat_atime;
 588                 if (attr->cat_ctime < lov_attr->cat_ctime)
 589                         attr->cat_ctime = lov_attr->cat_ctime;
 590                 if (attr->cat_mtime < lov_attr->cat_mtime)
 591                         attr->cat_mtime = lov_attr->cat_mtime;
 592         }
 593         RETURN(result);
 594 }
 595
 596 const static struct lov_layout_operations lov_dispatch[] = {
 597         [LLT_EMPTY] = {
 598                 .llo_init      = lov_init_empty,
 599                 .llo_delete    = lov_delete_empty,
 600                 .llo_fini      = lov_fini_empty,
 601                 .llo_install   = lov_install_empty,
 602                 .llo_print     = lov_print_empty,
 603                 .llo_page_init = lov_page_init_empty,
 604                 .llo_lock_init = lov_lock_init_empty,
 605                 .llo_io_init   = lov_io_init_empty,
 606                 .llo_getattr   = lov_attr_get_empty,
 607         },
 608         [LLT_RAID0] = {
 609                 .llo_init      = lov_init_raid0,
 610                 .llo_delete    = lov_delete_raid0,
 611                 .llo_fini      = lov_fini_raid0,
 612                 .llo_install   = lov_install_raid0,
 613                 .llo_print     = lov_print_raid0,
 614                 .llo_page_init = lov_page_init_raid0,
 615                 .llo_lock_init = lov_lock_init_raid0,
 616                 .llo_io_init   = lov_io_init_raid0,
 617                 .llo_getattr   = lov_attr_get_raid0,
 618         },
 619         [LLT_RELEASED] = {
 620                 .llo_init      = lov_init_released,
 621                 .llo_delete    = lov_delete_empty,
 622                 .llo_fini      = lov_fini_released,
 623                 .llo_install   = lov_install_empty,
 624                 .llo_print     = lov_print_released,
 625                 .llo_page_init = lov_page_init_empty,
 626                 .llo_lock_init = lov_lock_init_empty,
 627                 .llo_io_init   = lov_io_init_released,
 628                 .llo_getattr   = lov_attr_get_empty,
 629         }
 630 };
 631
 632 /**
 633  * Performs a double-dispatch based on the layout type of an object.
 634  */
 635 #define LOV_2DISPATCH_NOLOCK(obj, op, ...)              \
 636 ({                                                      \
 637         struct lov_object *__obj = (obj);               \
 638         enum lov_layout_type __llt;                     \
 639                                                         \
 640         __llt = __obj->lo_type;                         \
 641         LASSERT(__llt < ARRAY_SIZE(lov_dispatch));      \
 642         lov_dispatch[__llt].op(__VA_ARGS__);            \
 643 })
 644
 645 /**
 646  * Return lov_layout_type associated with a given lsm
 647  */
 648 static enum lov_layout_type lov_type(struct lov_stripe_md *lsm)
 649 {
 650         if (lsm == NULL)
 651                 return LLT_EMPTY;
 652         if (lsm_is_released(lsm))
 653                 return LLT_RELEASED;
 654         return LLT_RAID0;
 655 }
 656
 657 static inline void lov_conf_freeze(struct lov_object *lov)
 658 {
 659         CDEBUG(D_INODE, "To take share lov(%p) owner %p/%p\n",
 660                 lov, lov->lo_owner, current);
 661         if (lov->lo_owner != current)
 662                 down_read(&lov->lo_type_guard);
 663 }
 664
 665 static inline void lov_conf_thaw(struct lov_object *lov)
 666 {
 667         CDEBUG(D_INODE, "To release share lov(%p) owner %p/%p\n",
 668                 lov, lov->lo_owner, current);
 669         if (lov->lo_owner != current)
 670                 up_read(&lov->lo_type_guard);
 671 }
 672
 673 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...)                       \
 674 ({                                                                      \
 675         struct lov_object                      *__obj = (obj);          \
 676         int                                     __lock = !!(lock);      \
 677         typeof(lov_dispatch[0].op(__VA_ARGS__)) __result;               \
 678                                                                         \
 679         if (__lock)                                                     \
 680                 lov_conf_freeze(__obj);                                 \
 681         __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__);          \
 682         if (__lock)                                                     \
 683                 lov_conf_thaw(__obj);                                   \
 684         __result;                                                       \
 685 })
 686
 687 /**
 688  * Performs a locked double-dispatch based on the layout type of an object.
 689  */
 690 #define LOV_2DISPATCH(obj, op, ...)                     \
 691         LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
 692
 693 #define LOV_2DISPATCH_VOID(obj, op, ...)                                \
 694 do {                                                                    \
 695         struct lov_object                      *__obj = (obj);          \
 696         enum lov_layout_type                    __llt;                  \
 697                                                                         \
 698         lov_conf_freeze(__obj);                                         \
 699         __llt = __obj->lo_type;                                         \
 700         LASSERT(__llt < ARRAY_SIZE(lov_dispatch));                      \
 701         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 702         lov_conf_thaw(__obj);                                           \
 703 } while (0)
 704
 705 static void lov_conf_lock(struct lov_object *lov)
 706 {
 707         LASSERT(lov->lo_owner != current);
 708         down_write(&lov->lo_type_guard);
 709         LASSERT(lov->lo_owner == NULL);
 710         lov->lo_owner = current;
 711         CDEBUG(D_INODE, "Took exclusive lov(%p) owner %p\n",
 712                 lov, lov->lo_owner);
 713 }
 714
 715 static void lov_conf_unlock(struct lov_object *lov)
 716 {
 717         CDEBUG(D_INODE, "To release exclusive lov(%p) owner %p\n",
 718                 lov, lov->lo_owner);
 719         lov->lo_owner = NULL;
 720         up_write(&lov->lo_type_guard);
 721 }
 722
 723 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov)
 724 {
 725         struct l_wait_info lwi = { 0 };
 726         ENTRY;
 727
 728         while (atomic_read(&lov->lo_active_ios) > 0) {
 729                 CDEBUG(D_INODE, "file:"DFID" wait for active IO, now: %d.\n",
 730                         PFID(lu_object_fid(lov2lu(lov))),
 731                         atomic_read(&lov->lo_active_ios));
 732
 733                 l_wait_event(lov->lo_waitq,
 734                              atomic_read(&lov->lo_active_ios) == 0, &lwi);
 735         }
 736         RETURN(0);
 737 }
 738
 739 static int lov_layout_change(const struct lu_env *unused,
 740                              struct lov_object *lov, struct lov_stripe_md *lsm,
 741                              const struct cl_object_conf *conf)
 742 {
 743         enum lov_layout_type llt = lov_type(lsm);
 744         union lov_layout_state *state = &lov->u;
 745         const struct lov_layout_operations *old_ops;
 746         const struct lov_layout_operations *new_ops;
 747         struct lov_device *lov_dev = lov_object_dev(lov);
 748         struct lu_env *env;
 749         __u16 refcheck;
 750         int rc;
 751         ENTRY;
 752
 753         LASSERT(lov->lo_type < ARRAY_SIZE(lov_dispatch));
 754
 755         env = cl_env_get(&refcheck);
 756         if (IS_ERR(env))
 757                 RETURN(PTR_ERR(env));
 758
 759         LASSERT(llt < ARRAY_SIZE(lov_dispatch));
 760
 761         CDEBUG(D_INODE, DFID" from %s to %s\n",
 762                PFID(lu_object_fid(lov2lu(lov))),
 763                llt2str(lov->lo_type), llt2str(llt));
 764
 765         old_ops = &lov_dispatch[lov->lo_type];
 766         new_ops = &lov_dispatch[llt];
 767
 768         rc = cl_object_prune(env, &lov->lo_cl);
 769         if (rc != 0)
 770                 GOTO(out, rc);
 771
 772         rc = old_ops->llo_delete(env, lov, &lov->u);
 773         if (rc != 0)
 774                 GOTO(out, rc);
 775
 776         old_ops->llo_fini(env, lov, &lov->u);
 777
 778         LASSERT(atomic_read(&lov->lo_active_ios) == 0);
 779
 780         CDEBUG(D_INODE, DFID "Apply new layout lov %p, type %d\n",
 781                PFID(lu_object_fid(lov2lu(lov))), lov, llt);
 782
 783         lov->lo_type = LLT_EMPTY;
 784
 785         /* page bufsize fixup */
 786         cl_object_header(&lov->lo_cl)->coh_page_bufsize -=
 787                 lov_page_slice_fixup(lov, NULL);
 788
 789         rc = new_ops->llo_init(env, lov_dev, lov, lsm, conf, state);
 790         if (rc != 0) {
 791                 struct obd_device *obd = lov2obd(lov_dev->ld_lov);
 792
 793                 CERROR("%s: cannot apply new layout on "DFID" : rc = %d\n",
 794                        obd->obd_name, PFID(lu_object_fid(lov2lu(lov))), rc);
 795                 new_ops->llo_delete(env, lov, state);
 796                 new_ops->llo_fini(env, lov, state);
 797                 /* this file becomes an EMPTY file. */
 798                 GOTO(out, rc);
 799         }
 800
 801         new_ops->llo_install(env, lov, state);
 802         lov->lo_type = llt;
 803
 804 out:
 805         cl_env_put(env, &refcheck);
 806         RETURN(rc);
 807 }
 808
 809 /*****************************************************************************
 810  *
 811  * Lov object operations.
 812  *
 813  */
 814 int lov_object_init(const struct lu_env *env, struct lu_object *obj,
 815                     const struct lu_object_conf *conf)
 816 {
 817         struct lov_object            *lov   = lu2lov(obj);
 818         struct lov_device            *dev   = lov_object_dev(lov);
 819         const struct cl_object_conf  *cconf = lu2cl_conf(conf);
 820         union lov_layout_state       *set   = &lov->u;
 821         const struct lov_layout_operations *ops;
 822         struct lov_stripe_md *lsm = NULL;
 823         int rc;
 824         ENTRY;
 825
 826         init_rwsem(&lov->lo_type_guard);
 827         atomic_set(&lov->lo_active_ios, 0);
 828         init_waitqueue_head(&lov->lo_waitq);
 829         cl_object_page_init(lu2cl(obj), sizeof(struct lov_page));
 830
 831         lov->lo_type = LLT_EMPTY;
 832         if (cconf->u.coc_layout.lb_buf != NULL) {
 833                 lsm = lov_unpackmd(dev->ld_lov,
 834                                    cconf->u.coc_layout.lb_buf,
 835                                    cconf->u.coc_layout.lb_len);
 836                 if (IS_ERR(lsm))
 837                         RETURN(PTR_ERR(lsm));
 838         }
 839
 840         /* no locking is necessary, as object is being created */
 841         lov->lo_type = lov_type(lsm);
 842         ops = &lov_dispatch[lov->lo_type];
 843         rc = ops->llo_init(env, dev, lov, lsm, cconf, set);
 844         if (rc != 0)
 845                 GOTO(out_lsm, rc);
 846
 847         ops->llo_install(env, lov, set);
 848
 849 out_lsm:
 850         lov_lsm_put(lsm);
 851
 852         RETURN(rc);
 853 }
 854
 855 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
 856                         const struct cl_object_conf *conf)
 857 {
 858         struct lov_stripe_md    *lsm = NULL;
 859         struct lov_object       *lov = cl2lov(obj);
 860         int                      result = 0;
 861         ENTRY;
 862
 863         if (conf->coc_opc == OBJECT_CONF_SET &&
 864             conf->u.coc_layout.lb_buf != NULL) {
 865                 lsm = lov_unpackmd(lov_object_dev(lov)->ld_lov,
 866                                    conf->u.coc_layout.lb_buf,
 867                                    conf->u.coc_layout.lb_len);
 868                 if (IS_ERR(lsm))
 869                         RETURN(PTR_ERR(lsm));
 870         }
 871
 872         lov_conf_lock(lov);
 873         if (conf->coc_opc == OBJECT_CONF_INVALIDATE) {
 874                 lov->lo_layout_invalid = true;
 875                 GOTO(out, result = 0);
 876         }
 877
 878         if (conf->coc_opc == OBJECT_CONF_WAIT) {
 879                 if (lov->lo_layout_invalid &&
 880                     atomic_read(&lov->lo_active_ios) > 0) {
 881                         lov_conf_unlock(lov);
 882                         result = lov_layout_wait(env, lov);
 883                         lov_conf_lock(lov);
 884                 }
 885                 GOTO(out, result);
 886         }
 887
 888         LASSERT(conf->coc_opc == OBJECT_CONF_SET);
 889
 890         if ((lsm == NULL && lov->lo_lsm == NULL) ||
 891             ((lsm != NULL && lov->lo_lsm != NULL) &&
 892              (lov->lo_lsm->lsm_layout_gen == lsm->lsm_layout_gen) &&
 893              (lov->lo_lsm->lsm_pattern == lsm->lsm_pattern))) {
 894                 /* same version of layout */
 895                 lov->lo_layout_invalid = false;
 896                 GOTO(out, result = 0);
 897         }
 898
 899         /* will change layout - check if there still exists active IO. */
 900         if (atomic_read(&lov->lo_active_ios) > 0) {
 901                 lov->lo_layout_invalid = true;
 902                 GOTO(out, result = -EBUSY);
 903         }
 904
 905         result = lov_layout_change(env, lov, lsm, conf);
 906         lov->lo_layout_invalid = result != 0;
 907         EXIT;
 908
 909 out:
 910         lov_conf_unlock(lov);
 911         lov_lsm_put(lsm);
 912         CDEBUG(D_INODE, DFID" lo_layout_invalid=%d\n",
 913                PFID(lu_object_fid(lov2lu(lov))), lov->lo_layout_invalid);
 914         RETURN(result);
 915 }
 916
 917 static void lov_object_delete(const struct lu_env *env, struct lu_object *obj)
 918 {
 919         struct lov_object *lov = lu2lov(obj);
 920
 921         ENTRY;
 922         LOV_2DISPATCH_VOID(lov, llo_delete, env, lov, &lov->u);
 923         EXIT;
 924 }
 925
 926 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
 927 {
 928         struct lov_object *lov = lu2lov(obj);
 929
 930         ENTRY;
 931         LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
 932         lu_object_fini(obj);
 933         OBD_SLAB_FREE_PTR(lov, lov_object_kmem);
 934         EXIT;
 935 }
 936
 937 static int lov_object_print(const struct lu_env *env, void *cookie,
 938                             lu_printer_t p, const struct lu_object *o)
 939 {
 940         return LOV_2DISPATCH_NOLOCK(lu2lov(o), llo_print, env, cookie, p, o);
 941 }
 942
 943 int lov_page_init(const struct lu_env *env, struct cl_object *obj,
 944                   struct cl_page *page, pgoff_t index)
 945 {
 946         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_page_init, env, obj, page,
 947                                     index);
 948 }
 949
 950 /**
 951  * Implements cl_object_operations::clo_io_init() method for lov
 952  * layer. Dispatches to the appropriate layout io initialization method.
 953  */
 954 int lov_io_init(const struct lu_env *env, struct cl_object *obj,
 955                 struct cl_io *io)
 956 {
 957         CL_IO_SLICE_CLEAN(lov_env_io(env), lis_cl);
 958
 959         CDEBUG(D_INODE, DFID "io %p type %d ignore/verify layout %d/%d\n",
 960                PFID(lu_object_fid(&obj->co_lu)), io, io->ci_type,
 961                io->ci_ignore_layout, io->ci_verify_layout);
 962
 963         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
 964                                      !io->ci_ignore_layout, env, obj, io);
 965 }
 966
 967 /**
 968  * An implementation of cl_object_operations::clo_attr_get() method for lov
 969  * layer. For raid0 layout this collects and merges attributes of all
 970  * sub-objects.
 971  */
 972 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
 973                         struct cl_attr *attr)
 974 {
 975         /* do not take lock, as this function is called under a
 976          * spin-lock. Layout is protected from changing by ongoing IO. */
 977         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
 978 }
 979
 980 static int lov_attr_update(const struct lu_env *env, struct cl_object *obj,
 981                            const struct cl_attr *attr, unsigned valid)
 982 {
 983         /*
 984          * No dispatch is required here, as no layout implements this.
 985          */
 986         return 0;
 987 }
 988
 989 int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
 990                   struct cl_lock *lock, const struct cl_io *io)
 991 {
 992         /* No need to lock because we've taken one refcount of layout.  */
 993         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_lock_init, env, obj, lock,
 994                                     io);
 995 }
 996
 997 /**
 998  * We calculate on which OST the mapping will end. If the length of mapping
 999  * is greater than (stripe_size * stripe_count) then the last_stripe will
1000  * will be one just before start_stripe. Else we check if the mapping
1001  * intersects each OST and find last_stripe.
1002  * This function returns the last_stripe and also sets the stripe_count
1003  * over which the mapping is spread
1004  *
1005  * \param lsm [in]              striping information for the file
1006  * \param fm_start [in]         logical start of mapping
1007  * \param fm_end [in]           logical end of mapping
1008  * \param start_stripe [in]     starting stripe of the mapping
1009  * \param stripe_count [out]    the number of stripes across which to map is
1010  *                              returned
1011  *
1012  * \retval last_stripe          return the last stripe of the mapping
1013  */
1014 static int fiemap_calc_last_stripe(struct lov_stripe_md *lsm,
1015                                    u64 fm_start, u64 fm_end,
1016                                    int start_stripe, int *stripe_count)
1017 {
1018         int last_stripe;
1019         u64 obd_start;
1020         u64 obd_end;
1021         int i, j;
1022
1023         if (fm_end - fm_start > lsm->lsm_stripe_size * lsm->lsm_stripe_count) {
1024                 last_stripe = (start_stripe < 1 ? lsm->lsm_stripe_count - 1 :
1025                                                               start_stripe - 1);
1026                 *stripe_count = lsm->lsm_stripe_count;
1027         } else {
1028                 for (j = 0, i = start_stripe; j < lsm->lsm_stripe_count;
1029                      i = (i + 1) % lsm->lsm_stripe_count, j++) {
1030                         if ((lov_stripe_intersects(lsm, i, fm_start, fm_end,
1031                                                    &obd_start, &obd_end)) == 0)
1032                                 break;
1033                 }
1034                 *stripe_count = j;
1035                 last_stripe = (start_stripe + j - 1) % lsm->lsm_stripe_count;
1036         }
1037
1038         return last_stripe;
1039 }
1040
1041 /**
1042  * Set fe_device and copy extents from local buffer into main return buffer.
1043  *
1044  * \param fiemap [out]          fiemap to hold all extents
1045  * \param lcl_fm_ext [in]       array of fiemap extents get from OSC layer
1046  * \param ost_index [in]        OST index to be written into the fm_device
1047  *                              field for each extent
1048  * \param ext_count [in]        number of extents to be copied
1049  * \param current_extent [in]   where to start copying in the extent array
1050  */
1051 static void fiemap_prepare_and_copy_exts(struct fiemap *fiemap,
1052                                          struct fiemap_extent *lcl_fm_ext,
1053                                          int ost_index, unsigned int ext_count,
1054                                          int current_extent)
1055 {
1056         char            *to;
1057         unsigned int    ext;
1058
1059         for (ext = 0; ext < ext_count; ext++) {
1060                 lcl_fm_ext[ext].fe_device = ost_index;
1061                 lcl_fm_ext[ext].fe_flags |= FIEMAP_EXTENT_NET;
1062         }
1063
1064         /* Copy fm_extent's from fm_local to return buffer */
1065         to = (char *)fiemap + fiemap_count_to_size(current_extent);
1066         memcpy(to, lcl_fm_ext, ext_count * sizeof(struct fiemap_extent));
1067 }
1068
1069 #define FIEMAP_BUFFER_SIZE 4096
1070
1071 /**
1072  * Non-zero fe_logical indicates that this is a continuation FIEMAP
1073  * call. The local end offset and the device are sent in the first
1074  * fm_extent. This function calculates the stripe number from the index.
1075  * This function returns a stripe_no on which mapping is to be restarted.
1076  *
1077  * This function returns fm_end_offset which is the in-OST offset at which
1078  * mapping should be restarted. If fm_end_offset=0 is returned then caller
1079  * will re-calculate proper offset in next stripe.
1080  * Note that the first extent is passed to lov_get_info via the value field.
1081  *
1082  * \param fiemap [in]           fiemap request header
1083  * \param lsm [in]              striping information for the file
1084  * \param fm_start [in]         logical start of mapping
1085  * \param fm_end [in]           logical end of mapping
1086  * \param start_stripe [out]    starting stripe will be returned in this
1087  */
1088 static u64 fiemap_calc_fm_end_offset(struct fiemap *fiemap,
1089                                      struct lov_stripe_md *lsm,
1090                                      u64 fm_start, u64 fm_end,
1091                                      int *start_stripe)
1092 {
1093         u64 local_end = fiemap->fm_extents[0].fe_logical;
1094         u64 lun_start;
1095         u64 lun_end;
1096         u64 fm_end_offset;
1097         int stripe_no = -1;
1098         int i;
1099
1100         if (fiemap->fm_extent_count == 0 ||
1101             fiemap->fm_extents[0].fe_logical == 0)
1102                 return 0;
1103
1104         /* Find out stripe_no from ost_index saved in the fe_device */
1105         for (i = 0; i < lsm->lsm_stripe_count; i++) {
1106                 struct lov_oinfo *oinfo = lsm->lsm_oinfo[i];
1107
1108                 if (lov_oinfo_is_dummy(oinfo))
1109                         continue;
1110
1111                 if (oinfo->loi_ost_idx == fiemap->fm_extents[0].fe_device) {
1112                         stripe_no = i;
1113                         break;
1114                 }
1115         }
1116
1117         if (stripe_no == -1)
1118                 return -EINVAL;
1119
1120         /* If we have finished mapping on previous device, shift logical
1121          * offset to start of next device */
1122         if (lov_stripe_intersects(lsm, stripe_no, fm_start, fm_end,
1123                                   &lun_start, &lun_end) != 0 &&
1124             local_end < lun_end) {
1125                 fm_end_offset = local_end;
1126                 *start_stripe = stripe_no;
1127         } else {
1128                 /* This is a special value to indicate that caller should
1129                  * calculate offset in next stripe. */
1130                 fm_end_offset = 0;
1131                 *start_stripe = (stripe_no + 1) % lsm->lsm_stripe_count;
1132         }
1133
1134         return fm_end_offset;
1135 }
1136
1137 struct fiemap_state {
1138         struct fiemap   *fs_fm;
1139         u64             fs_start;
1140         u64             fs_length;
1141         u64             fs_end;
1142         u64             fs_end_offset;
1143         int             fs_cur_extent;
1144         int             fs_cnt_need;
1145         int             fs_start_stripe;
1146         int             fs_last_stripe;
1147         bool            fs_device_done;
1148         bool            fs_finish;
1149         bool            fs_enough;
1150 };
1151
1152 int fiemap_for_stripe(const struct lu_env *env, struct cl_object *obj,
1153                       struct lov_stripe_md *lsm,
1154                       struct fiemap *fiemap, size_t *buflen,
1155                       struct ll_fiemap_info_key *fmkey, int stripeno,
1156                       struct fiemap_state *fs)
1157 {
1158         struct cl_object *subobj;
1159         struct lov_obd *lov = lu2lov_dev(obj->co_lu.lo_dev)->ld_lov;
1160         struct fiemap_extent *fm_ext = &fs->fs_fm->fm_extents[0];
1161         u64 req_fm_len; /* Stores length of required mapping */
1162         u64 len_mapped_single_call;
1163         u64 lun_start;
1164         u64 lun_end;
1165         u64 obd_object_end;
1166         unsigned int ext_count;
1167         /* EOF for object */
1168         bool ost_eof = false;
1169         /* done with required mapping for this OST? */
1170         bool ost_done = false;
1171         int ost_index;
1172         int rc = 0;
1173
1174         fs->fs_device_done = false;
1175         /* Find out range of mapping on this stripe */
1176         if ((lov_stripe_intersects(lsm, stripeno, fs->fs_start, fs->fs_end,
1177                                    &lun_start, &obd_object_end)) == 0)
1178                 return 0;
1179
1180         if (lov_oinfo_is_dummy(lsm->lsm_oinfo[stripeno]))
1181                 return -EIO;
1182
1183         /* If this is a continuation FIEMAP call and we are on
1184          * starting stripe then lun_start needs to be set to
1185          * end_offset */
1186         if (fs->fs_end_offset != 0 && stripeno == fs->fs_start_stripe)
1187                 lun_start = fs->fs_end_offset;
1188
1189         lun_end = fs->fs_length;
1190         if (lun_end != ~0ULL) {
1191                 /* Handle fs->fs_start + fs->fs_length overflow */
1192                 if (fs->fs_start + fs->fs_length < fs->fs_start)
1193                         fs->fs_length = ~0ULL - fs->fs_start;
1194                 lun_end = lov_size_to_stripe(lsm, fs->fs_start + fs->fs_length,
1195                                              stripeno);
1196         }
1197
1198         if (lun_start == lun_end)
1199                 return 0;
1200
1201         req_fm_len = obd_object_end - lun_start;
1202         fs->fs_fm->fm_length = 0;
1203         len_mapped_single_call = 0;
1204
1205         /* find lobsub object */
1206         subobj = lov_find_subobj(env, cl2lov(obj), lsm, stripeno);
1207         if (IS_ERR(subobj))
1208                 return PTR_ERR(subobj);
1209         /* If the output buffer is very large and the objects have many
1210          * extents we may need to loop on a single OST repeatedly */
1211         do {
1212                 if (fiemap->fm_extent_count > 0) {
1213                         /* Don't get too many extents. */
1214                         if (fs->fs_cur_extent + fs->fs_cnt_need >
1215                             fiemap->fm_extent_count)
1216                                 fs->fs_cnt_need = fiemap->fm_extent_count -
1217                                                   fs->fs_cur_extent;
1218                 }
1219
1220                 lun_start += len_mapped_single_call;
1221                 fs->fs_fm->fm_length = req_fm_len - len_mapped_single_call;
1222                 req_fm_len = fs->fs_fm->fm_length;
1223                 fs->fs_fm->fm_extent_count = fs->fs_enough ?
1224                                              1 : fs->fs_cnt_need;
1225                 fs->fs_fm->fm_mapped_extents = 0;
1226                 fs->fs_fm->fm_flags = fiemap->fm_flags;
1227
1228                 ost_index = lsm->lsm_oinfo[stripeno]->loi_ost_idx;
1229
1230                 if (ost_index < 0 || ost_index >= lov->desc.ld_tgt_count)
1231                         GOTO(obj_put, rc = -EINVAL);
1232                 /* If OST is inactive, return extent with UNKNOWN flag. */
1233                 if (!lov->lov_tgts[ost_index]->ltd_active) {
1234                         fs->fs_fm->fm_flags |= FIEMAP_EXTENT_LAST;
1235                         fs->fs_fm->fm_mapped_extents = 1;
1236
1237                         fm_ext[0].fe_logical = lun_start;
1238                         fm_ext[0].fe_length = obd_object_end - lun_start;
1239                         fm_ext[0].fe_flags |= FIEMAP_EXTENT_UNKNOWN;
1240
1241                         goto inactive_tgt;
1242                 }
1243
1244                 fs->fs_fm->fm_start = lun_start;
1245                 fs->fs_fm->fm_flags &= ~FIEMAP_FLAG_DEVICE_ORDER;
1246                 memcpy(&fmkey->lfik_fiemap, fs->fs_fm, sizeof(*fs->fs_fm));
1247                 *buflen = fiemap_count_to_size(fs->fs_fm->fm_extent_count);
1248
1249                 rc = cl_object_fiemap(env, subobj, fmkey, fs->fs_fm, buflen);
1250                 if (rc != 0)
1251                         GOTO(obj_put, rc);
1252 inactive_tgt:
1253                 ext_count = fs->fs_fm->fm_mapped_extents;
1254                 if (ext_count == 0) {
1255                         ost_done = true;
1256                         fs->fs_device_done = true;
1257                         /* If last stripe has hold at the end,
1258                          * we need to return */
1259                         if (stripeno == fs->fs_last_stripe) {
1260                                 fiemap->fm_mapped_extents = 0;
1261                                 fs->fs_finish = true;
1262                                 GOTO(obj_put, rc);
1263                         }
1264                         break;
1265                 } else if (fs->fs_enough) {
1266                         /*
1267                          * We've collected enough extents and there are
1268                          * more extents after it.
1269                          */
1270                         fs->fs_finish = true;
1271                         GOTO(obj_put, rc);
1272                 }
1273
1274                 /* If we just need num of extents, got to next device */
1275                 if (fiemap->fm_extent_count == 0) {
1276                         fs->fs_cur_extent += ext_count;
1277                         break;
1278                 }
1279
1280                 /* prepare to copy retrived map extents */
1281                 len_mapped_single_call = fm_ext[ext_count - 1].fe_logical +
1282                                          fm_ext[ext_count - 1].fe_length -
1283                                          lun_start;
1284
1285                 /* Have we finished mapping on this device? */
1286                 if (req_fm_len <= len_mapped_single_call) {
1287                         ost_done = true;
1288                         fs->fs_device_done = true;
1289                 }
1290
1291                 /* Clear the EXTENT_LAST flag which can be present on
1292                  * the last extent */
1293                 if (fm_ext[ext_count - 1].fe_flags & FIEMAP_EXTENT_LAST)
1294                         fm_ext[ext_count - 1].fe_flags &= ~FIEMAP_EXTENT_LAST;
1295                 if (lov_stripe_size(lsm, fm_ext[ext_count - 1].fe_logical +
1296                                          fm_ext[ext_count - 1].fe_length,
1297                                     stripeno) >= fmkey->lfik_oa.o_size) {
1298                         ost_eof = true;
1299                         fs->fs_device_done = true;
1300                 }
1301
1302                 fiemap_prepare_and_copy_exts(fiemap, fm_ext, ost_index,
1303                                              ext_count, fs->fs_cur_extent);
1304                 fs->fs_cur_extent += ext_count;
1305
1306                 /* Ran out of available extents? */
1307                 if (fs->fs_cur_extent >= fiemap->fm_extent_count)
1308                         fs->fs_enough = true;
1309         } while (!ost_done && !ost_eof);
1310
1311         if (stripeno == fs->fs_last_stripe)
1312                 fs->fs_finish = true;
1313 obj_put:
1314         cl_object_put(env, subobj);
1315
1316         return rc;
1317 }
1318
1319 /**
1320  * Break down the FIEMAP request and send appropriate calls to individual OSTs.
1321  * This also handles the restarting of FIEMAP calls in case mapping overflows
1322  * the available number of extents in single call.
1323  *
1324  * \param env [in]              lustre environment
1325  * \param obj [in]              file object
1326  * \param fmkey [in]            fiemap request header and other info
1327  * \param fiemap [out]          fiemap buffer holding retrived map extents
1328  * \param buflen [in/out]       max buffer length of @fiemap, when iterate
1329  *                              each OST, it is used to limit max map needed
1330  * \retval 0    success
1331  * \retval < 0  error
1332  */
1333 static int lov_object_fiemap(const struct lu_env *env, struct cl_object *obj,
1334                              struct ll_fiemap_info_key *fmkey,
1335                              struct fiemap *fiemap, size_t *buflen)
1336 {
1337         struct lov_stripe_md    *lsm;
1338         struct fiemap           *fm_local = NULL;
1339         int                     cur_stripe;
1340         int                     stripe_count;
1341         unsigned int            buffer_size = FIEMAP_BUFFER_SIZE;
1342         int                     rc = 0;
1343         struct fiemap_state fs = { 0 };
1344         ENTRY;
1345
1346         lsm = lov_lsm_addref(cl2lov(obj));
1347         if (lsm == NULL)
1348                 RETURN(-ENODATA);
1349
1350         /**
1351          * If the stripe_count > 1 and the application does not understand
1352          * DEVICE_ORDER flag, it cannot interpret the extents correctly.
1353          */
1354         if (lsm->lsm_stripe_count > 1 && !(fiemap->fm_flags &
1355                                            FIEMAP_FLAG_DEVICE_ORDER))
1356                 GOTO(out_lsm, rc = -ENOTSUPP);
1357
1358         if (lsm_is_released(lsm)) {
1359                 if (fiemap->fm_start < fmkey->lfik_oa.o_size) {
1360                         /**
1361                          * released file, return a minimal FIEMAP if
1362                          * request fits in file-size.
1363                          */
1364                         fiemap->fm_mapped_extents = 1;
1365                         fiemap->fm_extents[0].fe_logical = fiemap->fm_start;
1366                         if (fiemap->fm_start + fiemap->fm_length <
1367                             fmkey->lfik_oa.o_size)
1368                                 fiemap->fm_extents[0].fe_length =
1369                                         fiemap->fm_length;
1370                         else
1371                                 fiemap->fm_extents[0].fe_length =
1372                                         fmkey->lfik_oa.o_size -
1373                                         fiemap->fm_start;
1374                         fiemap->fm_extents[0].fe_flags |=
1375                                 FIEMAP_EXTENT_UNKNOWN | FIEMAP_EXTENT_LAST;
1376                 }
1377                 GOTO(out_lsm, rc = 0);
1378         }
1379
1380         if (fiemap_count_to_size(fiemap->fm_extent_count) < buffer_size)
1381                 buffer_size = fiemap_count_to_size(fiemap->fm_extent_count);
1382
1383         OBD_ALLOC_LARGE(fm_local, buffer_size);
1384         if (fm_local == NULL)
1385                 GOTO(out_lsm, rc = -ENOMEM);
1386
1387         fs.fs_fm = fm_local;
1388         fs.fs_cnt_need = fiemap_size_to_count(buffer_size);
1389
1390         fs.fs_start = fiemap->fm_start;
1391         /* fs.fs_start is beyond the end of the file */
1392         if (fs.fs_start > fmkey->lfik_oa.o_size)
1393                 GOTO(out_fm_local, rc = -EINVAL);
1394
1395         fs.fs_length = fiemap->fm_length;
1396         /* Calculate start stripe, last stripe and length of mapping */
1397         fs.fs_start_stripe = lov_stripe_number(lsm, fs.fs_start);
1398         fs.fs_end = (fs.fs_length == ~0ULL) ? fmkey->lfik_oa.o_size :
1399                                               fs.fs_start + fs.fs_length - 1;
1400         /* If fs_length != ~0ULL but fs_start+fs_length-1 exceeds file size */
1401         if (fs.fs_end > fmkey->lfik_oa.o_size) {
1402                 fs.fs_end = fmkey->lfik_oa.o_size;
1403                 fs.fs_length = fs.fs_end - fs.fs_start;
1404         }
1405
1406         fs.fs_last_stripe = fiemap_calc_last_stripe(lsm, fs.fs_start, fs.fs_end,
1407                                                     fs.fs_start_stripe,
1408                                                     &stripe_count);
1409         fs.fs_end_offset = fiemap_calc_fm_end_offset(fiemap, lsm, fs.fs_start,
1410                                                      fs.fs_end,
1411                                                      &fs.fs_start_stripe);
1412         if (fs.fs_end_offset == -EINVAL)
1413                 GOTO(out_fm_local, rc = -EINVAL);
1414
1415         /**
1416          * Requested extent count exceeds the fiemap buffer size, shrink our
1417          * ambition.
1418          */
1419         if (fiemap_count_to_size(fiemap->fm_extent_count) > *buflen)
1420                 fiemap->fm_extent_count = fiemap_size_to_count(*buflen);
1421         if (fiemap->fm_extent_count == 0)
1422                 fs.fs_cnt_need = 0;
1423
1424         fs.fs_finish = false;
1425         fs.fs_enough = false;
1426         fs.fs_cur_extent = 0;
1427
1428         /* Check each stripe */
1429         for (cur_stripe = fs.fs_start_stripe; stripe_count > 0;
1430              --stripe_count,
1431              cur_stripe = (cur_stripe + 1) % lsm->lsm_stripe_count) {
1432                 rc = fiemap_for_stripe(env, obj, lsm, fiemap, buflen, fmkey,
1433                                        cur_stripe, &fs);
1434                 if (rc < 0)
1435                         GOTO(out_fm_local, rc);
1436                 if (fs.fs_finish)
1437                         break;
1438         } /* for each stripe */
1439
1440         /* Indicate that we are returning device offsets unless file just has
1441          * single stripe */
1442         if (lsm->lsm_stripe_count > 1)
1443                 fiemap->fm_flags |= FIEMAP_FLAG_DEVICE_ORDER;
1444
1445         if (fiemap->fm_extent_count == 0)
1446                 goto skip_last_device_calc;
1447
1448         /* Check if we have reached the last stripe and whether mapping for that
1449          * stripe is done. */
1450         if ((cur_stripe == fs.fs_last_stripe) && fs.fs_device_done)
1451                 fiemap->fm_extents[fs.fs_cur_extent - 1].fe_flags |=
1452                                                              FIEMAP_EXTENT_LAST;
1453 skip_last_device_calc:
1454         fiemap->fm_mapped_extents = fs.fs_cur_extent;
1455 out_fm_local:
1456         OBD_FREE_LARGE(fm_local, buffer_size);
1457
1458 out_lsm:
1459         lov_lsm_put(lsm);
1460
1461         return rc;
1462 }
1463
1464 static int lov_object_getstripe(const struct lu_env *env, struct cl_object *obj,
1465                                 struct lov_user_md __user *lum)
1466 {
1467         struct lov_object       *lov = cl2lov(obj);
1468         struct lov_stripe_md    *lsm;
1469         int                     rc = 0;
1470         ENTRY;
1471
1472         lsm = lov_lsm_addref(lov);
1473         if (lsm == NULL)
1474                 RETURN(-ENODATA);
1475
1476         rc = lov_getstripe(cl2lov(obj), lsm, lum);
1477         lov_lsm_put(lsm);
1478         RETURN(rc);
1479 }
1480
1481 static int lov_object_layout_get(const struct lu_env *env,
1482                                  struct cl_object *obj,
1483                                  struct cl_layout *cl)
1484 {
1485         struct lov_object *lov = cl2lov(obj);
1486         struct lov_stripe_md *lsm = lov_lsm_addref(lov);
1487         struct lu_buf *buf = &cl->cl_buf;
1488         ssize_t rc;
1489         ENTRY;
1490
1491         if (lsm == NULL) {
1492                 cl->cl_size = 0;
1493                 cl->cl_layout_gen = CL_LAYOUT_GEN_EMPTY;
1494
1495                 RETURN(0);
1496         }
1497
1498         cl->cl_size = lov_mds_md_size(lsm->lsm_stripe_count, lsm->lsm_magic);
1499         cl->cl_layout_gen = lsm->lsm_layout_gen;
1500
1501         rc = lov_lsm_pack(lsm, buf->lb_buf, buf->lb_len);
1502         lov_lsm_put(lsm);
1503
1504         RETURN(rc < 0 ? rc : 0);
1505 }
1506
1507 static loff_t lov_object_maxbytes(struct cl_object *obj)
1508 {
1509         struct lov_object *lov = cl2lov(obj);
1510         struct lov_stripe_md *lsm = lov_lsm_addref(lov);
1511         loff_t maxbytes;
1512
1513         if (lsm == NULL)
1514                 return LLONG_MAX;
1515
1516         maxbytes = lsm->lsm_maxbytes;
1517
1518         lov_lsm_put(lsm);
1519
1520         return maxbytes;
1521 }
1522
1523 static const struct cl_object_operations lov_ops = {
1524         .coo_page_init    = lov_page_init,
1525         .coo_lock_init    = lov_lock_init,
1526         .coo_io_init      = lov_io_init,
1527         .coo_attr_get     = lov_attr_get,
1528         .coo_attr_update  = lov_attr_update,
1529         .coo_conf_set     = lov_conf_set,
1530         .coo_getstripe    = lov_object_getstripe,
1531         .coo_layout_get   = lov_object_layout_get,
1532         .coo_maxbytes     = lov_object_maxbytes,
1533         .coo_fiemap       = lov_object_fiemap,
1534 };
1535
1536 static const struct lu_object_operations lov_lu_obj_ops = {
1537         .loo_object_init      = lov_object_init,
1538         .loo_object_delete    = lov_object_delete,
1539         .loo_object_release   = NULL,
1540         .loo_object_free      = lov_object_free,
1541         .loo_object_print     = lov_object_print,
1542         .loo_object_invariant = NULL
1543 };
1544
1545 struct lu_object *lov_object_alloc(const struct lu_env *env,
1546                                    const struct lu_object_header *unused,
1547                                    struct lu_device *dev)
1548 {
1549         struct lov_object *lov;
1550         struct lu_object  *obj;
1551
1552         ENTRY;
1553         OBD_SLAB_ALLOC_PTR_GFP(lov, lov_object_kmem, GFP_NOFS);
1554         if (lov != NULL) {
1555                 obj = lov2lu(lov);
1556                 lu_object_init(obj, NULL, dev);
1557                 lov->lo_cl.co_ops = &lov_ops;
1558                 lov->lo_type = -1; /* invalid, to catch uninitialized type */
1559                 /*
1560                  * object io operation vector (cl_object::co_iop) is installed
1561                  * later in lov_object_init(), as different vectors are used
1562                  * for object with different layouts.
1563                  */
1564                 obj->lo_ops = &lov_lu_obj_ops;
1565         } else
1566                 obj = NULL;
1567         RETURN(obj);
1568 }
1569
1570 struct lov_stripe_md *lov_lsm_addref(struct lov_object *lov)
1571 {
1572         struct lov_stripe_md *lsm = NULL;
1573
1574         lov_conf_freeze(lov);
1575         if (lov->lo_lsm != NULL) {
1576                 lsm = lsm_addref(lov->lo_lsm);
1577                 CDEBUG(D_INODE, "lsm %p addref %d/%d by %p.\n",
1578                         lsm, atomic_read(&lsm->lsm_refc),
1579                         lov->lo_layout_invalid, current);
1580         }
1581         lov_conf_thaw(lov);
1582         return lsm;
1583 }
1584
1585 int lov_read_and_clear_async_rc(struct cl_object *clob)
1586 {
1587         struct lu_object *luobj;
1588         int rc = 0;
1589         ENTRY;
1590
1591         luobj = lu_object_locate(&cl_object_header(clob)->coh_lu,
1592                                  &lov_device_type);
1593         if (luobj != NULL) {
1594                 struct lov_object *lov = lu2lov(luobj);
1595
1596                 lov_conf_freeze(lov);
1597                 switch (lov->lo_type) {
1598                 case LLT_RAID0: {
1599                         struct lov_stripe_md *lsm;
1600                         int i;
1601
1602                         lsm = lov->lo_lsm;
1603                         LASSERT(lsm != NULL);
1604                         for (i = 0; i < lsm->lsm_stripe_count; i++) {
1605                                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
1606
1607                                 if (lov_oinfo_is_dummy(loi))
1608                                         continue;
1609
1610                                 if (loi->loi_ar.ar_rc && !rc)
1611                                         rc = loi->loi_ar.ar_rc;
1612                                 loi->loi_ar.ar_rc = 0;
1613                         }
1614                 }
1615                 case LLT_RELEASED:
1616                 case LLT_EMPTY:
1617                         break;
1618                 default:
1619                         LBUG();
1620                 }
1621                 lov_conf_thaw(lov);
1622         }
1623         RETURN(rc);
1624 }
1625 EXPORT_SYMBOL(lov_read_and_clear_async_rc);
1626
1627 /** @} lov */