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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, Whamcloud, Inc.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * Implementation of cl_object for LOV layer.
  37  *
  38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  39  *   Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_LOV
  43
  44 #include "lov_cl_internal.h"
  45 #include <lustre_debug.h>
  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,
  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         struct cl_page *(*llo_page_init)(const struct lu_env *env,
  71                                          struct cl_object *obj,
  72                                          struct cl_page *page,
  73                                          cfs_page_t *vmpage);
  74         int  (*llo_lock_init)(const struct lu_env *env,
  75                               struct cl_object *obj, struct cl_lock *lock,
  76                               const struct cl_io *io);
  77         int  (*llo_io_init)(const struct lu_env *env,
  78                             struct cl_object *obj, struct cl_io *io);
  79         int  (*llo_getattr)(const struct lu_env *env, struct cl_object *obj,
  80                             struct cl_attr *attr);
  81 };
  82
  83 /*****************************************************************************
  84  *
  85  * Lov object layout operations.
  86  *
  87  */
  88
  89 static void lov_install_empty(const struct lu_env *env,
  90                               struct lov_object *lov,
  91                               union  lov_layout_state *state)
  92 {
  93         /*
  94          * File without objects.
  95          */
  96 }
  97
  98 static int lov_init_empty(const struct lu_env *env,
  99                           struct lov_device *dev, struct lov_object *lov,
 100                           const struct cl_object_conf *conf,
 101                           union  lov_layout_state *state)
 102 {
 103         return 0;
 104 }
 105
 106 static void lov_install_raid0(const struct lu_env *env,
 107                               struct lov_object *lov,
 108                               union  lov_layout_state *state)
 109 {
 110 }
 111
 112 static struct cl_object *lov_sub_find(const struct lu_env *env,
 113                                       struct cl_device *dev,
 114                                       const struct lu_fid *fid,
 115                                       const struct cl_object_conf *conf)
 116 {
 117         struct lu_object *o;
 118
 119         ENTRY;
 120         o = lu_object_find_at(env, cl2lu_dev(dev), fid, &conf->coc_lu);
 121         LASSERT(ergo(!IS_ERR(o), o->lo_dev->ld_type == &lovsub_device_type));
 122         RETURN(lu2cl(o));
 123 }
 124
 125 static int lov_init_sub(const struct lu_env *env, struct lov_object *lov,
 126                         struct cl_object *stripe,
 127                         struct lov_layout_raid0 *r0, int idx)
 128 {
 129         struct cl_object_header *hdr;
 130         struct cl_object_header *subhdr;
 131         struct cl_object_header *parent;
 132         struct lov_oinfo        *oinfo;
 133         int result;
 134
 135         if (OBD_FAIL_CHECK(OBD_FAIL_LOV_INIT)) {
 136                 cl_object_put(env, stripe);
 137                 return -EIO;
 138         }
 139
 140         hdr    = cl_object_header(lov2cl(lov));
 141         subhdr = cl_object_header(stripe);
 142         parent = subhdr->coh_parent;
 143
 144         oinfo = lov->lo_lsm->lsm_oinfo[idx];
 145         CDEBUG(D_INODE, DFID"@%p[%d] -> "DFID"@%p: id: "LPU64" seq: "LPU64
 146                " idx: %d gen: %d\n",
 147                PFID(&subhdr->coh_lu.loh_fid), subhdr, idx,
 148                PFID(&hdr->coh_lu.loh_fid), hdr,
 149                oinfo->loi_id, oinfo->loi_seq,
 150                oinfo->loi_ost_idx, oinfo->loi_ost_gen);
 151
 152         if (parent == NULL) {
 153                 subhdr->coh_parent = hdr;
 154                 subhdr->coh_nesting = hdr->coh_nesting + 1;
 155                 lu_object_ref_add(&stripe->co_lu, "lov-parent", lov);
 156                 r0->lo_sub[idx] = cl2lovsub(stripe);
 157                 r0->lo_sub[idx]->lso_super = lov;
 158                 r0->lo_sub[idx]->lso_index = idx;
 159                 result = 0;
 160         } else {
 161                 CERROR("Stripe is already owned by other file (%d).\n", idx);
 162                 LU_OBJECT_DEBUG(D_ERROR, env, &stripe->co_lu, "\n");
 163                 LU_OBJECT_DEBUG(D_ERROR, env, lu_object_top(&parent->coh_lu),
 164                                 "old\n");
 165                 LU_OBJECT_HEADER(D_ERROR, env, lov2lu(lov), "new\n");
 166                 cl_object_put(env, stripe);
 167                 result = -EIO;
 168         }
 169         return result;
 170 }
 171
 172 static int lov_init_raid0(const struct lu_env *env,
 173                           struct lov_device *dev, struct lov_object *lov,
 174                           const struct cl_object_conf *conf,
 175                           union  lov_layout_state *state)
 176 {
 177         int result;
 178         int i;
 179
 180         struct cl_object        *stripe;
 181         struct lov_thread_info  *lti     = lov_env_info(env);
 182         struct cl_object_conf   *subconf = &lti->lti_stripe_conf;
 183         struct lov_stripe_md    *lsm     = conf->u.coc_md->lsm;
 184         struct lu_fid           *ofid    = &lti->lti_fid;
 185         struct lov_layout_raid0 *r0      = &state->raid0;
 186
 187         ENTRY;
 188
 189         if (lsm->lsm_magic != LOV_MAGIC_V1 && lsm->lsm_magic != LOV_MAGIC_V3) {
 190                 dump_lsm(D_ERROR, lsm);
 191                 LASSERTF(0, "magic mismatch, expected %d/%d, actual %d.\n",
 192                          LOV_MAGIC_V1, LOV_MAGIC_V3, lsm->lsm_magic);
 193         }
 194
 195         LASSERT(lov->lo_lsm == NULL);
 196         lov->lo_lsm = lsm_addref(lsm);
 197         r0->lo_nr  = lsm->lsm_stripe_count;
 198         LASSERT(r0->lo_nr <= lov_targets_nr(dev));
 199
 200         OBD_ALLOC_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 201         if (r0->lo_sub != NULL) {
 202                 result = 0;
 203                 subconf->coc_inode = conf->coc_inode;
 204                 spin_lock_init(&r0->lo_sub_lock);
 205                 /*
 206                  * Create stripe cl_objects.
 207                  */
 208                 for (i = 0; i < r0->lo_nr && result == 0; ++i) {
 209                         struct cl_device *subdev;
 210                         struct lov_oinfo *oinfo = lsm->lsm_oinfo[i];
 211                         int ost_idx = oinfo->loi_ost_idx;
 212
 213                         fid_ostid_unpack(ofid, &oinfo->loi_oi,
 214                                          oinfo->loi_ost_idx);
 215                         subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
 216                         subconf->u.coc_oinfo = oinfo;
 217                         LASSERTF(subdev != NULL, "not init ost %d\n", ost_idx);
 218                         /* In the function below, .hs_keycmp resolves to
 219                          * lu_obj_hop_keycmp() */
 220                         /* coverity[overrun-buffer-val] */
 221                         stripe = lov_sub_find(env, subdev, ofid, subconf);
 222                         if (!IS_ERR(stripe))
 223                                 result = lov_init_sub(env, lov, stripe, r0, i);
 224                         else
 225                                 result = PTR_ERR(stripe);
 226                 }
 227         } else
 228                 result = -ENOMEM;
 229         RETURN(result);
 230 }
 231
 232 static int lov_delete_empty(const struct lu_env *env, struct lov_object *lov,
 233                             union lov_layout_state *state)
 234 {
 235         LASSERT(lov->lo_type == LLT_EMPTY);
 236         cl_object_prune(env, &lov->lo_cl);
 237         return 0;
 238 }
 239
 240 static void lov_subobject_kill(const struct lu_env *env, struct lov_object *lov,
 241                                struct lovsub_object *los, int idx)
 242 {
 243         struct cl_object        *sub;
 244         struct lov_layout_raid0 *r0;
 245         struct lu_site          *site;
 246         struct lu_site_bkt_data *bkt;
 247         cfs_waitlink_t          *waiter;
 248
 249         r0  = &lov->u.raid0;
 250         LASSERT(r0->lo_sub[idx] == los);
 251
 252         sub  = lovsub2cl(los);
 253         site = sub->co_lu.lo_dev->ld_site;
 254         bkt  = lu_site_bkt_from_fid(site, &sub->co_lu.lo_header->loh_fid);
 255
 256         cl_object_kill(env, sub);
 257         /* release a reference to the sub-object and ... */
 258         lu_object_ref_del(&sub->co_lu, "lov-parent", lov);
 259         cl_object_put(env, sub);
 260
 261         /* ... wait until it is actually destroyed---sub-object clears its
 262          * ->lo_sub[] slot in lovsub_object_fini() */
 263         if (r0->lo_sub[idx] == los) {
 264                 waiter = &lov_env_info(env)->lti_waiter;
 265                 cfs_waitlink_init(waiter);
 266                 cfs_waitq_add(&bkt->lsb_marche_funebre, waiter);
 267                 cfs_set_current_state(CFS_TASK_UNINT);
 268                 while (1) {
 269                         /* this wait-queue is signaled at the end of
 270                          * lu_object_free(). */
 271                         cfs_set_current_state(CFS_TASK_UNINT);
 272                         spin_lock(&r0->lo_sub_lock);
 273                         if (r0->lo_sub[idx] == los) {
 274                                 spin_unlock(&r0->lo_sub_lock);
 275                                 cfs_waitq_wait(waiter, CFS_TASK_UNINT);
 276                         } else {
 277                                 spin_unlock(&r0->lo_sub_lock);
 278                                 cfs_set_current_state(CFS_TASK_RUNNING);
 279                                 break;
 280                         }
 281                 }
 282                 cfs_waitq_del(&bkt->lsb_marche_funebre, waiter);
 283         }
 284         LASSERT(r0->lo_sub[idx] == NULL);
 285 }
 286
 287 static int lov_delete_raid0(const struct lu_env *env, struct lov_object *lov,
 288                             union lov_layout_state *state)
 289 {
 290         struct lov_layout_raid0 *r0 = &state->raid0;
 291         struct lov_stripe_md    *lsm = lov->lo_lsm;
 292         int i;
 293
 294         ENTRY;
 295
 296         dump_lsm(D_INODE, lsm);
 297         if (lov->lo_lsm_invalid && cfs_atomic_read(&lsm->lsm_refc) > 1)
 298                 RETURN(-EBUSY);
 299
 300         if (r0->lo_sub != NULL) {
 301                 for (i = 0; i < r0->lo_nr; ++i) {
 302                         struct lovsub_object *los = r0->lo_sub[i];
 303
 304                         if (los != NULL) {
 305                                 cl_locks_prune(env, &los->lso_cl, 1);
 306                                 /*
 307                                  * If top-level object is to be evicted from
 308                                  * the cache, so are its sub-objects.
 309                                  */
 310                                 lov_subobject_kill(env, lov, los, i);
 311                         }
 312                 }
 313         }
 314         RETURN(0);
 315 }
 316
 317 static void lov_fini_empty(const struct lu_env *env, struct lov_object *lov,
 318                            union lov_layout_state *state)
 319 {
 320         LASSERT(lov->lo_type == LLT_EMPTY);
 321 }
 322
 323 static void lov_fini_raid0(const struct lu_env *env, struct lov_object *lov,
 324                            union lov_layout_state *state)
 325 {
 326         struct lov_layout_raid0 *r0 = &state->raid0;
 327         ENTRY;
 328
 329         if (r0->lo_sub != NULL) {
 330                 OBD_FREE_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 331                 r0->lo_sub = NULL;
 332         }
 333
 334         dump_lsm(D_INODE, lov->lo_lsm);
 335         lov_free_memmd(&lov->lo_lsm);
 336
 337         EXIT;
 338 }
 339
 340 static int lov_print_empty(const struct lu_env *env, void *cookie,
 341                            lu_printer_t p, const struct lu_object *o)
 342 {
 343         (*p)(env, cookie, "empty\n");
 344         return 0;
 345 }
 346
 347 static int lov_print_raid0(const struct lu_env *env, void *cookie,
 348                            lu_printer_t p, const struct lu_object *o)
 349 {
 350         struct lov_object       *lov = lu2lov(o);
 351         struct lov_layout_raid0 *r0  = lov_r0(lov);
 352         int i;
 353
 354         (*p)(env, cookie, "stripes: %d:\n", r0->lo_nr);
 355         for (i = 0; i < r0->lo_nr; ++i) {
 356                 struct lu_object *sub;
 357
 358                 if (r0->lo_sub[i] != NULL) {
 359                         sub = lovsub2lu(r0->lo_sub[i]);
 360                         lu_object_print(env, cookie, p, sub);
 361                 } else
 362                         (*p)(env, cookie, "sub %d absent\n", i);
 363         }
 364         return 0;
 365 }
 366
 367 /**
 368  * Implements cl_object_operations::coo_attr_get() method for an object
 369  * without stripes (LLT_EMPTY layout type).
 370  *
 371  * The only attributes this layer is authoritative in this case is
 372  * cl_attr::cat_blocks---it's 0.
 373  */
 374 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
 375                               struct cl_attr *attr)
 376 {
 377         attr->cat_blocks = 0;
 378         return 0;
 379 }
 380
 381 static int lov_attr_get_raid0(const struct lu_env *env, struct cl_object *obj,
 382                               struct cl_attr *attr)
 383 {
 384         struct lov_object       *lov = cl2lov(obj);
 385         struct lov_layout_raid0 *r0 = lov_r0(lov);
 386         struct lov_stripe_md    *lsm = lov->lo_lsm;
 387         struct ost_lvb          *lvb = &lov_env_info(env)->lti_lvb;
 388         __u64                    kms;
 389         int                      result = 0;
 390
 391         ENTRY;
 392
 393         /* this is called w/o holding type guard mutex, so it must be inside
 394          * an on going IO otherwise lsm may be replaced.
 395          * LU-2117: it turns out there exists one exception. For mmaped files,
 396          * the lock of those files may be requested in the other file's IO
 397          * context, and this function is called in ccc_lock_state(), it will
 398          * hit this assertion.
 399          * Anyway, it's still okay to call attr_get w/o type guard as layout
 400          * can't go if locks exist. */
 401         /* LASSERT(cfs_atomic_read(&lsm->lsm_refc) > 1); */
 402
 403         if (!r0->lo_attr_valid) {
 404                 /*
 405                  * Fill LVB with attributes already initialized by the upper
 406                  * layer.
 407                  */
 408                 cl_attr2lvb(lvb, attr);
 409                 kms = attr->cat_kms;
 410
 411                 /*
 412                  * XXX that should be replaced with a loop over sub-objects,
 413                  * doing cl_object_attr_get() on them. But for now, let's
 414                  * reuse old lov code.
 415                  */
 416
 417                 /*
 418                  * XXX take lsm spin-lock to keep lov_merge_lvb_kms()
 419                  * happy. It's not needed, because new code uses
 420                  * ->coh_attr_guard spin-lock to protect consistency of
 421                  * sub-object attributes.
 422                  */
 423                 lov_stripe_lock(lsm);
 424                 result = lov_merge_lvb_kms(lsm, lvb, &kms);
 425                 lov_stripe_unlock(lsm);
 426                 if (result == 0) {
 427                         cl_lvb2attr(attr, lvb);
 428                         attr->cat_kms = kms;
 429                         r0->lo_attr_valid = 1;
 430                         r0->lo_attr = *attr;
 431                 }
 432         } else
 433                 *attr = r0->lo_attr;
 434         RETURN(result);
 435 }
 436
 437 const static struct lov_layout_operations lov_dispatch[] = {
 438         [LLT_EMPTY] = {
 439                 .llo_init      = lov_init_empty,
 440                 .llo_delete    = lov_delete_empty,
 441                 .llo_fini      = lov_fini_empty,
 442                 .llo_install   = lov_install_empty,
 443                 .llo_print     = lov_print_empty,
 444                 .llo_page_init = lov_page_init_empty,
 445                 .llo_lock_init = lov_lock_init_empty,
 446                 .llo_io_init   = lov_io_init_empty,
 447                 .llo_getattr   = lov_attr_get_empty
 448         },
 449         [LLT_RAID0] = {
 450                 .llo_init      = lov_init_raid0,
 451                 .llo_delete    = lov_delete_raid0,
 452                 .llo_fini      = lov_fini_raid0,
 453                 .llo_install   = lov_install_raid0,
 454                 .llo_print     = lov_print_raid0,
 455                 .llo_page_init = lov_page_init_raid0,
 456                 .llo_lock_init = lov_lock_init_raid0,
 457                 .llo_io_init   = lov_io_init_raid0,
 458                 .llo_getattr   = lov_attr_get_raid0
 459         }
 460 };
 461
 462
 463 /**
 464  * Performs a double-dispatch based on the layout type of an object.
 465  */
 466 #define LOV_2DISPATCH_NOLOCK(obj, op, ...)                              \
 467 ({                                                                      \
 468         struct lov_object                      *__obj = (obj);          \
 469         enum lov_layout_type                    __llt;                  \
 470                                                                         \
 471         __llt = __obj->lo_type;                                         \
 472         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 473         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 474 })
 475
 476 static inline void lov_conf_freeze(struct lov_object *lov)
 477 {
 478         if (lov->lo_owner != cfs_current())
 479                 down_read(&lov->lo_type_guard);
 480 }
 481
 482 static inline void lov_conf_thaw(struct lov_object *lov)
 483 {
 484         if (lov->lo_owner != cfs_current())
 485                 up_read(&lov->lo_type_guard);
 486 }
 487
 488 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...)                       \
 489 ({                                                                      \
 490         struct lov_object                      *__obj = (obj);          \
 491         int                                     __lock = !!(lock);      \
 492         typeof(lov_dispatch[0].op(__VA_ARGS__)) __result;               \
 493                                                                         \
 494         if (__lock)                                                     \
 495                 lov_conf_freeze(__obj);                                 \
 496         __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__);          \
 497         if (__lock)                                                     \
 498                 lov_conf_thaw(__obj);                                   \
 499         __result;                                                       \
 500 })
 501
 502 /**
 503  * Performs a locked double-dispatch based on the layout type of an object.
 504  */
 505 #define LOV_2DISPATCH(obj, op, ...)                     \
 506         LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
 507
 508 #define LOV_2DISPATCH_VOID(obj, op, ...)                                \
 509 do {                                                                    \
 510         struct lov_object                      *__obj = (obj);          \
 511         enum lov_layout_type                    __llt;                  \
 512                                                                         \
 513         lov_conf_freeze(__obj);                                         \
 514         __llt = __obj->lo_type;                                         \
 515         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 516         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 517         lov_conf_thaw(__obj);                                           \
 518 } while (0)
 519
 520 static void lov_conf_lock(struct lov_object *lov)
 521 {
 522         LASSERT(lov->lo_owner != cfs_current());
 523         down_write(&lov->lo_type_guard);
 524         LASSERT(lov->lo_owner == NULL);
 525         lov->lo_owner = cfs_current();
 526 }
 527
 528 static void lov_conf_unlock(struct lov_object *lov)
 529 {
 530         lov->lo_owner = NULL;
 531         up_write(&lov->lo_type_guard);
 532 }
 533
 534 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov)
 535 {
 536         struct l_wait_info lwi = { 0 };
 537         struct lov_stripe_md *lsm = lov->lo_lsm;
 538         ENTRY;
 539
 540         if (!lov->lo_lsm_invalid || lsm == NULL)
 541                 RETURN(0);
 542
 543         LASSERT(cfs_atomic_read(&lsm->lsm_refc) > 0);
 544         while (cfs_atomic_read(&lsm->lsm_refc) > 1) {
 545                 lov_conf_unlock(lov);
 546                 l_wait_event(lov->lo_waitq,
 547                              cfs_atomic_read(&lsm->lsm_refc) == 1, &lwi);
 548                 lov_conf_lock(lov);
 549         }
 550         RETURN(0);
 551 }
 552
 553 static int lov_layout_change(const struct lu_env *unused,
 554                              struct lov_object *lov, enum lov_layout_type llt,
 555                              const struct cl_object_conf *conf)
 556 {
 557         int result;
 558         union lov_layout_state *state = &lov->u;
 559         const struct lov_layout_operations *old_ops;
 560         const struct lov_layout_operations *new_ops;
 561
 562         struct cl_object_header *hdr = cl_object_header(&lov->lo_cl);
 563         void *cookie;
 564         struct lu_env *env;
 565         int refcheck;
 566
 567         LASSERT(0 <= lov->lo_type && lov->lo_type < ARRAY_SIZE(lov_dispatch));
 568         LASSERT(0 <= llt && llt < ARRAY_SIZE(lov_dispatch));
 569         ENTRY;
 570
 571         cookie = cl_env_reenter();
 572         env = cl_env_get(&refcheck);
 573         if (IS_ERR(env)) {
 574                 cl_env_reexit(cookie);
 575                 RETURN(PTR_ERR(env));
 576         }
 577
 578         old_ops = &lov_dispatch[lov->lo_type];
 579         new_ops = &lov_dispatch[llt];
 580
 581         result = old_ops->llo_delete(env, lov, &lov->u);
 582         if (result == 0) {
 583                 old_ops->llo_fini(env, lov, &lov->u);
 584                 LASSERT(cfs_list_empty(&hdr->coh_locks));
 585                 LASSERT(hdr->coh_tree.rnode == NULL);
 586                 LASSERT(hdr->coh_pages == 0);
 587
 588                 lov->lo_type = LLT_EMPTY;
 589                 result = new_ops->llo_init(env,
 590                                         lu2lov_dev(lov->lo_cl.co_lu.lo_dev),
 591                                         lov, conf, state);
 592                 if (result == 0) {
 593                         new_ops->llo_install(env, lov, state);
 594                         lov->lo_type = llt;
 595                 } else {
 596                         new_ops->llo_delete(env, lov, state);
 597                         new_ops->llo_fini(env, lov, state);
 598                         /* this file becomes an EMPTY file. */
 599                 }
 600         }
 601
 602         cl_env_put(env, &refcheck);
 603         cl_env_reexit(cookie);
 604         RETURN(result);
 605 }
 606
 607 /*****************************************************************************
 608  *
 609  * Lov object operations.
 610  *
 611  */
 612
 613 int lov_object_init(const struct lu_env *env, struct lu_object *obj,
 614                     const struct lu_object_conf *conf)
 615 {
 616         struct lov_device            *dev   = lu2lov_dev(obj->lo_dev);
 617         struct lov_object            *lov   = lu2lov(obj);
 618         const struct cl_object_conf  *cconf = lu2cl_conf(conf);
 619         union  lov_layout_state      *set   = &lov->u;
 620         const struct lov_layout_operations *ops;
 621         int result;
 622
 623         ENTRY;
 624         init_rwsem(&lov->lo_type_guard);
 625         cfs_waitq_init(&lov->lo_waitq);
 626
 627         /* no locking is necessary, as object is being created */
 628         lov->lo_type = cconf->u.coc_md->lsm != NULL ? LLT_RAID0 : LLT_EMPTY;
 629         ops = &lov_dispatch[lov->lo_type];
 630         result = ops->llo_init(env, dev, lov, cconf, set);
 631         if (result == 0)
 632                 ops->llo_install(env, lov, set);
 633         RETURN(result);
 634 }
 635
 636 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
 637                         const struct cl_object_conf *conf)
 638 {
 639         struct lov_stripe_md *lsm = NULL;
 640         struct lov_object *lov = cl2lov(obj);
 641         int result = 0;
 642         ENTRY;
 643
 644         lov_conf_lock(lov);
 645         if (conf->coc_invalidate) {
 646                 lov->lo_lsm_invalid = 1;
 647                 GOTO(out, result = 0);
 648         }
 649
 650         if (conf->u.coc_md != NULL)
 651                 lsm = conf->u.coc_md->lsm;
 652
 653         if ((lsm == NULL && lov->lo_lsm == NULL) ||
 654             (lsm != NULL && lov->lo_lsm != NULL &&
 655              lov->lo_lsm->lsm_layout_gen == lsm->lsm_layout_gen)) {
 656                 lov->lo_lsm_invalid = 0;
 657                 GOTO(out, result = 0);
 658         }
 659
 660         /* will change layout */
 661         lov_layout_wait(env, lov);
 662
 663         /*
 664          * Only LLT_EMPTY <-> LLT_RAID0 transitions are supported.
 665          */
 666         switch (lov->lo_type) {
 667         case LLT_EMPTY:
 668                 if (lsm != NULL)
 669                         result = lov_layout_change(env, lov, LLT_RAID0, conf);
 670                 break;
 671         case LLT_RAID0:
 672                 if (lsm == NULL)
 673                         result = lov_layout_change(env, lov, LLT_EMPTY, conf);
 674                 else if (lov_stripe_md_cmp(lov->lo_lsm, lsm))
 675                         result = -EOPNOTSUPP;
 676                 break;
 677         default:
 678                 LBUG();
 679         }
 680         lov->lo_lsm_invalid = result != 0;
 681         EXIT;
 682
 683 out:
 684         lov_conf_unlock(lov);
 685         RETURN(result);
 686 }
 687
 688 static void lov_object_delete(const struct lu_env *env, struct lu_object *obj)
 689 {
 690         struct lov_object *lov = lu2lov(obj);
 691
 692         ENTRY;
 693         LOV_2DISPATCH_VOID(lov, llo_delete, env, lov, &lov->u);
 694         EXIT;
 695 }
 696
 697 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
 698 {
 699         struct lov_object *lov = lu2lov(obj);
 700
 701         ENTRY;
 702         LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
 703         lu_object_fini(obj);
 704         OBD_SLAB_FREE_PTR(lov, lov_object_kmem);
 705         EXIT;
 706 }
 707
 708 static int lov_object_print(const struct lu_env *env, void *cookie,
 709                             lu_printer_t p, const struct lu_object *o)
 710 {
 711         return LOV_2DISPATCH(lu2lov(o), llo_print, env, cookie, p, o);
 712 }
 713
 714 struct cl_page *lov_page_init(const struct lu_env *env, struct cl_object *obj,
 715                               struct cl_page *page, cfs_page_t *vmpage)
 716 {
 717         return LOV_2DISPATCH_NOLOCK(cl2lov(obj),
 718                                     llo_page_init, env, obj, page, vmpage);
 719 }
 720
 721 /**
 722  * Implements cl_object_operations::clo_io_init() method for lov
 723  * layer. Dispatches to the appropriate layout io initialization method.
 724  */
 725 int lov_io_init(const struct lu_env *env, struct cl_object *obj,
 726                 struct cl_io *io)
 727 {
 728         CL_IO_SLICE_CLEAN(lov_env_io(env), lis_cl);
 729         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
 730                                      !io->ci_ignore_layout, env, obj, io);
 731 }
 732
 733 /**
 734  * An implementation of cl_object_operations::clo_attr_get() method for lov
 735  * layer. For raid0 layout this collects and merges attributes of all
 736  * sub-objects.
 737  */
 738 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
 739                         struct cl_attr *attr)
 740 {
 741         /* do not take lock, as this function is called under a
 742          * spin-lock. Layout is protected from changing by ongoing IO. */
 743         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
 744 }
 745
 746 static int lov_attr_set(const struct lu_env *env, struct cl_object *obj,
 747                         const struct cl_attr *attr, unsigned valid)
 748 {
 749         /*
 750          * No dispatch is required here, as no layout implements this.
 751          */
 752         return 0;
 753 }
 754
 755 int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
 756                   struct cl_lock *lock, const struct cl_io *io)
 757 {
 758         /* No need to lock because we've taken one refcount of layout.  */
 759         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_lock_init, env, obj, lock,
 760                                     io);
 761 }
 762
 763 static const struct cl_object_operations lov_ops = {
 764         .coo_page_init = lov_page_init,
 765         .coo_lock_init = lov_lock_init,
 766         .coo_io_init   = lov_io_init,
 767         .coo_attr_get  = lov_attr_get,
 768         .coo_attr_set  = lov_attr_set,
 769         .coo_conf_set  = lov_conf_set
 770 };
 771
 772 static const struct lu_object_operations lov_lu_obj_ops = {
 773         .loo_object_init      = lov_object_init,
 774         .loo_object_delete    = lov_object_delete,
 775         .loo_object_release   = NULL,
 776         .loo_object_free      = lov_object_free,
 777         .loo_object_print     = lov_object_print,
 778         .loo_object_invariant = NULL
 779 };
 780
 781 struct lu_object *lov_object_alloc(const struct lu_env *env,
 782                                    const struct lu_object_header *unused,
 783                                    struct lu_device *dev)
 784 {
 785         struct lov_object *lov;
 786         struct lu_object  *obj;
 787
 788         ENTRY;
 789         OBD_SLAB_ALLOC_PTR_GFP(lov, lov_object_kmem, CFS_ALLOC_IO);
 790         if (lov != NULL) {
 791                 obj = lov2lu(lov);
 792                 lu_object_init(obj, NULL, dev);
 793                 lov->lo_cl.co_ops = &lov_ops;
 794                 lov->lo_type = -1; /* invalid, to catch uninitialized type */
 795                 /*
 796                  * object io operation vector (cl_object::co_iop) is installed
 797                  * later in lov_object_init(), as different vectors are used
 798                  * for object with different layouts.
 799                  */
 800                 obj->lo_ops = &lov_lu_obj_ops;
 801         } else
 802                 obj = NULL;
 803         RETURN(obj);
 804 }
 805
 806 struct lov_stripe_md *lov_lsm_addref(struct lov_object *lov)
 807 {
 808         struct lov_stripe_md *lsm = NULL;
 809
 810         lov_conf_freeze(lov);
 811         if (lov->lo_lsm != NULL) {
 812                 lsm = lsm_addref(lov->lo_lsm);
 813                 CDEBUG(D_INODE, "lsm %p addref %d/%d by %p.\n",
 814                         lsm, cfs_atomic_read(&lsm->lsm_refc),
 815                         lov->lo_lsm_invalid, cfs_current());
 816         }
 817         lov_conf_thaw(lov);
 818         return lsm;
 819 }
 820
 821 void lov_lsm_decref(struct lov_object *lov, struct lov_stripe_md *lsm)
 822 {
 823         if (lsm == NULL)
 824                 return;
 825
 826         CDEBUG(D_INODE, "lsm %p decref %d by %p.\n",
 827                 lsm, cfs_atomic_read(&lsm->lsm_refc), cfs_current());
 828
 829         if (lov_free_memmd(&lsm) <= 1 && lov->lo_lsm_invalid)
 830                 cfs_waitq_signal(&lov->lo_waitq);
 831 }
 832
 833 struct lov_stripe_md *lov_lsm_get(struct cl_object *clobj)
 834 {
 835         struct lu_object *luobj;
 836         struct lov_stripe_md *lsm = NULL;
 837
 838         if (clobj == NULL)
 839                 return NULL;
 840
 841         luobj = lu_object_locate(&cl_object_header(clobj)->coh_lu,
 842                                  &lov_device_type);
 843         if (luobj != NULL)
 844                 lsm = lov_lsm_addref(lu2lov(luobj));
 845         return lsm;
 846 }
 847 EXPORT_SYMBOL(lov_lsm_get);
 848
 849 void lov_lsm_put(struct cl_object *clobj, struct lov_stripe_md *lsm)
 850 {
 851         struct lu_object *luobj;
 852
 853         if (clobj == NULL || lsm == NULL)
 854                 return;
 855
 856         luobj = lu_object_locate(&cl_object_header(clobj)->coh_lu,
 857                                  &lov_device_type);
 858         LASSERT(luobj != NULL);
 859
 860         lov_lsm_decref(lu2lov(luobj), lsm);
 861 }
 862 EXPORT_SYMBOL(lov_lsm_put);
 863
 864 int lov_read_and_clear_async_rc(struct cl_object *clob)
 865 {
 866         struct lu_object *luobj;
 867         int rc = 0;
 868         ENTRY;
 869
 870         luobj = lu_object_locate(&cl_object_header(clob)->coh_lu,
 871                                  &lov_device_type);
 872         if (luobj != NULL) {
 873                 struct lov_object *lov = lu2lov(luobj);
 874
 875                 lov_conf_freeze(lov);
 876                 switch (lov->lo_type) {
 877                 case LLT_RAID0: {
 878                         struct lov_stripe_md *lsm;
 879                         int i;
 880
 881                         lsm = lov->lo_lsm;
 882                         LASSERT(lsm != NULL);
 883                         for (i = 0; i < lsm->lsm_stripe_count; i++) {
 884                                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
 885                                 if (loi->loi_ar.ar_rc && !rc)
 886                                         rc = loi->loi_ar.ar_rc;
 887                                 loi->loi_ar.ar_rc = 0;
 888                         }
 889                 }
 890                 case LLT_EMPTY:
 891                         break;
 892                 default:
 893                         LBUG();
 894                 }
 895                 lov_conf_thaw(lov);
 896         }
 897         RETURN(rc);
 898 }
 899 EXPORT_SYMBOL(lov_read_and_clear_async_rc);
 900
 901 /** @} lov */