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                 cfs_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                         stripe = lov_sub_find(env, subdev, ofid, subconf);
 219                         if (!IS_ERR(stripe))
 220                                 result = lov_init_sub(env, lov, stripe, r0, i);
 221                         else
 222                                 result = PTR_ERR(stripe);
 223                 }
 224         } else
 225                 result = -ENOMEM;
 226         RETURN(result);
 227 }
 228
 229 static int lov_delete_empty(const struct lu_env *env, struct lov_object *lov,
 230                             union lov_layout_state *state)
 231 {
 232         LASSERT(lov->lo_type == LLT_EMPTY);
 233         return 0;
 234 }
 235
 236 static void lov_subobject_kill(const struct lu_env *env, struct lov_object *lov,
 237                                struct lovsub_object *los, int idx)
 238 {
 239         struct cl_object        *sub;
 240         struct lov_layout_raid0 *r0;
 241         struct lu_site          *site;
 242         struct lu_site_bkt_data *bkt;
 243         cfs_waitlink_t          *waiter;
 244
 245         r0  = &lov->u.raid0;
 246         LASSERT(r0->lo_sub[idx] == los);
 247
 248         sub  = lovsub2cl(los);
 249         site = sub->co_lu.lo_dev->ld_site;
 250         bkt  = lu_site_bkt_from_fid(site, &sub->co_lu.lo_header->loh_fid);
 251
 252         cl_object_kill(env, sub);
 253         /* release a reference to the sub-object and ... */
 254         lu_object_ref_del(&sub->co_lu, "lov-parent", lov);
 255         cl_object_put(env, sub);
 256
 257         /* ... wait until it is actually destroyed---sub-object clears its
 258          * ->lo_sub[] slot in lovsub_object_fini() */
 259         if (r0->lo_sub[idx] == los) {
 260                 waiter = &lov_env_info(env)->lti_waiter;
 261                 cfs_waitlink_init(waiter);
 262                 cfs_waitq_add(&bkt->lsb_marche_funebre, waiter);
 263                 cfs_set_current_state(CFS_TASK_UNINT);
 264                 while (1) {
 265                         /* this wait-queue is signaled at the end of
 266                          * lu_object_free(). */
 267                         cfs_set_current_state(CFS_TASK_UNINT);
 268                         cfs_spin_lock(&r0->lo_sub_lock);
 269                         if (r0->lo_sub[idx] == los) {
 270                                 cfs_spin_unlock(&r0->lo_sub_lock);
 271                                 cfs_waitq_wait(waiter, CFS_TASK_UNINT);
 272                         } else {
 273                                 cfs_spin_unlock(&r0->lo_sub_lock);
 274                                 cfs_set_current_state(CFS_TASK_RUNNING);
 275                                 break;
 276                         }
 277                 }
 278                 cfs_waitq_del(&bkt->lsb_marche_funebre, waiter);
 279         }
 280         LASSERT(r0->lo_sub[idx] == NULL);
 281 }
 282
 283 static int lov_delete_raid0(const struct lu_env *env, struct lov_object *lov,
 284                             union lov_layout_state *state)
 285 {
 286         struct lov_layout_raid0 *r0 = &state->raid0;
 287         struct lov_stripe_md    *lsm = lov->lo_lsm;
 288         int i;
 289
 290         ENTRY;
 291
 292         dump_lsm(D_INODE, lsm);
 293         if (lov->lo_lsm_invalid && cfs_atomic_read(&lsm->lsm_refc) > 1)
 294                 RETURN(-EBUSY);
 295
 296         if (r0->lo_sub != NULL) {
 297                 for (i = 0; i < r0->lo_nr; ++i) {
 298                         struct lovsub_object *los = r0->lo_sub[i];
 299
 300                         if (los != NULL)
 301                                 /*
 302                                  * If top-level object is to be evicted from
 303                                  * the cache, so are its sub-objects.
 304                                  */
 305                                 lov_subobject_kill(env, lov, los, i);
 306                 }
 307         }
 308         RETURN(0);
 309 }
 310
 311 static void lov_fini_empty(const struct lu_env *env, struct lov_object *lov,
 312                            union lov_layout_state *state)
 313 {
 314         LASSERT(lov->lo_type == LLT_EMPTY);
 315 }
 316
 317 static void lov_fini_raid0(const struct lu_env *env, struct lov_object *lov,
 318                            union lov_layout_state *state)
 319 {
 320         struct lov_layout_raid0 *r0 = &state->raid0;
 321         ENTRY;
 322
 323         if (r0->lo_sub != NULL) {
 324                 OBD_FREE_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 325                 r0->lo_sub = NULL;
 326         }
 327
 328         dump_lsm(D_INODE, lov->lo_lsm);
 329         lov_free_memmd(&lov->lo_lsm);
 330
 331         EXIT;
 332 }
 333
 334 static int lov_print_empty(const struct lu_env *env, void *cookie,
 335                            lu_printer_t p, const struct lu_object *o)
 336 {
 337         (*p)(env, cookie, "empty\n");
 338         return 0;
 339 }
 340
 341 static int lov_print_raid0(const struct lu_env *env, void *cookie,
 342                            lu_printer_t p, const struct lu_object *o)
 343 {
 344         struct lov_object       *lov = lu2lov(o);
 345         struct lov_layout_raid0 *r0  = lov_r0(lov);
 346         int i;
 347
 348         (*p)(env, cookie, "stripes: %d:\n", r0->lo_nr);
 349         for (i = 0; i < r0->lo_nr; ++i) {
 350                 struct lu_object *sub;
 351
 352                 if (r0->lo_sub[i] != NULL) {
 353                         sub = lovsub2lu(r0->lo_sub[i]);
 354                         lu_object_print(env, cookie, p, sub);
 355                 } else
 356                         (*p)(env, cookie, "sub %d absent\n", i);
 357         }
 358         return 0;
 359 }
 360
 361 /**
 362  * Implements cl_object_operations::coo_attr_get() method for an object
 363  * without stripes (LLT_EMPTY layout type).
 364  *
 365  * The only attributes this layer is authoritative in this case is
 366  * cl_attr::cat_blocks---it's 0.
 367  */
 368 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
 369                               struct cl_attr *attr)
 370 {
 371         attr->cat_blocks = 0;
 372         return 0;
 373 }
 374
 375 static int lov_attr_get_raid0(const struct lu_env *env, struct cl_object *obj,
 376                               struct cl_attr *attr)
 377 {
 378         struct lov_object       *lov = cl2lov(obj);
 379         struct lov_layout_raid0 *r0 = lov_r0(lov);
 380         struct lov_stripe_md    *lsm = lov->lo_lsm;
 381         struct ost_lvb          *lvb = &lov_env_info(env)->lti_lvb;
 382         __u64                    kms;
 383         int                      result = 0;
 384
 385         ENTRY;
 386
 387         /* this is called w/o holding type guard mutex, so it must be inside
 388          * an on going IO otherwise lsm may be replaced. */
 389         LASSERT(cfs_atomic_read(&lsm->lsm_refc) > 1);
 390
 391         if (!r0->lo_attr_valid) {
 392                 /*
 393                  * Fill LVB with attributes already initialized by the upper
 394                  * layer.
 395                  */
 396                 cl_attr2lvb(lvb, attr);
 397                 kms = attr->cat_kms;
 398
 399                 /*
 400                  * XXX that should be replaced with a loop over sub-objects,
 401                  * doing cl_object_attr_get() on them. But for now, let's
 402                  * reuse old lov code.
 403                  */
 404
 405                 /*
 406                  * XXX take lsm spin-lock to keep lov_merge_lvb_kms()
 407                  * happy. It's not needed, because new code uses
 408                  * ->coh_attr_guard spin-lock to protect consistency of
 409                  * sub-object attributes.
 410                  */
 411                 lov_stripe_lock(lsm);
 412                 result = lov_merge_lvb_kms(lsm, lvb, &kms);
 413                 lov_stripe_unlock(lsm);
 414                 if (result == 0) {
 415                         cl_lvb2attr(attr, lvb);
 416                         attr->cat_kms = kms;
 417                         r0->lo_attr_valid = 1;
 418                         r0->lo_attr = *attr;
 419                 }
 420         } else
 421                 *attr = r0->lo_attr;
 422         RETURN(result);
 423 }
 424
 425 const static struct lov_layout_operations lov_dispatch[] = {
 426         [LLT_EMPTY] = {
 427                 .llo_init      = lov_init_empty,
 428                 .llo_delete    = lov_delete_empty,
 429                 .llo_fini      = lov_fini_empty,
 430                 .llo_install   = lov_install_empty,
 431                 .llo_print     = lov_print_empty,
 432                 .llo_page_init = lov_page_init_empty,
 433                 .llo_lock_init = NULL,
 434                 .llo_io_init   = lov_io_init_empty,
 435                 .llo_getattr   = lov_attr_get_empty
 436         },
 437         [LLT_RAID0] = {
 438                 .llo_init      = lov_init_raid0,
 439                 .llo_delete    = lov_delete_raid0,
 440                 .llo_fini      = lov_fini_raid0,
 441                 .llo_install   = lov_install_raid0,
 442                 .llo_print     = lov_print_raid0,
 443                 .llo_page_init = lov_page_init_raid0,
 444                 .llo_lock_init = lov_lock_init_raid0,
 445                 .llo_io_init   = lov_io_init_raid0,
 446                 .llo_getattr   = lov_attr_get_raid0
 447         }
 448 };
 449
 450
 451 /**
 452  * Performs a double-dispatch based on the layout type of an object.
 453  */
 454 #define LOV_2DISPATCH_NOLOCK(obj, op, ...)                              \
 455 ({                                                                      \
 456         struct lov_object                      *__obj = (obj);          \
 457         enum lov_layout_type                    __llt;                  \
 458                                                                         \
 459         __llt = __obj->lo_type;                                         \
 460         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 461         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 462 })
 463
 464 static inline void lov_conf_freeze(struct lov_object *lov)
 465 {
 466         if (lov->lo_owner != cfs_current())
 467                 cfs_down_read(&lov->lo_type_guard);
 468 }
 469
 470 static inline void lov_conf_thaw(struct lov_object *lov)
 471 {
 472         if (lov->lo_owner != cfs_current())
 473                 cfs_up_read(&lov->lo_type_guard);
 474 }
 475
 476 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...)                       \
 477 ({                                                                      \
 478         struct lov_object                      *__obj = (obj);          \
 479         int                                     __lock = !!(lock);      \
 480         typeof(lov_dispatch[0].op(__VA_ARGS__)) __result;               \
 481                                                                         \
 482         if (__lock)                                                     \
 483                 lov_conf_freeze(__obj);                                 \
 484         __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__);          \
 485         if (__lock)                                                     \
 486                 lov_conf_thaw(__obj);                                   \
 487         __result;                                                       \
 488 })
 489
 490 /**
 491  * Performs a locked double-dispatch based on the layout type of an object.
 492  */
 493 #define LOV_2DISPATCH(obj, op, ...)                     \
 494         LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
 495
 496 #define LOV_2DISPATCH_VOID(obj, op, ...)                                \
 497 do {                                                                    \
 498         struct lov_object                      *__obj = (obj);          \
 499         enum lov_layout_type                    __llt;                  \
 500                                                                         \
 501         lov_conf_freeze(__obj);                                         \
 502         __llt = __obj->lo_type;                                         \
 503         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 504         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 505         lov_conf_thaw(__obj);                                           \
 506 } while (0)
 507
 508 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov)
 509 {
 510         struct l_wait_info lwi = { 0 };
 511         struct lov_stripe_md *lsm = lov->lo_lsm;
 512         ENTRY;
 513
 514         if (!lov->lo_lsm_invalid || lsm == NULL)
 515                 RETURN(0);
 516
 517         l_wait_event(lov->lo_waitq, cfs_atomic_read(&lsm->lsm_refc) == 1, &lwi);
 518         RETURN(0);
 519 }
 520
 521 static int lov_layout_change(const struct lu_env *env,
 522                              struct lov_object *lov, enum lov_layout_type llt,
 523                              const struct cl_object_conf *conf)
 524 {
 525         int result;
 526         union lov_layout_state *state = &lov->u;
 527         const struct lov_layout_operations *old_ops;
 528         const struct lov_layout_operations *new_ops;
 529
 530         struct cl_object_header *hdr = cl_object_header(&lov->lo_cl);
 531         void *cookie;
 532         struct lu_env *nested;
 533         int refcheck;
 534
 535         LASSERT(0 <= lov->lo_type && lov->lo_type < ARRAY_SIZE(lov_dispatch));
 536         LASSERT(0 <= llt && llt < ARRAY_SIZE(lov_dispatch));
 537         ENTRY;
 538
 539         cookie = cl_env_reenter();
 540         nested = cl_env_get(&refcheck);
 541         if (!IS_ERR(nested))
 542                 cl_object_prune(nested, &lov->lo_cl);
 543         else
 544                 result = PTR_ERR(nested);
 545         cl_env_put(nested, &refcheck);
 546         cl_env_reexit(cookie);
 547
 548         old_ops = &lov_dispatch[lov->lo_type];
 549         new_ops = &lov_dispatch[llt];
 550
 551         result = old_ops->llo_delete(env, lov, &lov->u);
 552         if (result == 0) {
 553                 old_ops->llo_fini(env, lov, &lov->u);
 554                 LASSERT(cfs_list_empty(&hdr->coh_locks));
 555                 LASSERT(hdr->coh_tree.rnode == NULL);
 556                 LASSERT(hdr->coh_pages == 0);
 557
 558                 lov->lo_type = LLT_EMPTY;
 559                 result = new_ops->llo_init(env,
 560                                         lu2lov_dev(lov->lo_cl.co_lu.lo_dev),
 561                                         lov, conf, state);
 562                 if (result == 0) {
 563                         new_ops->llo_install(env, lov, state);
 564                         lov->lo_type = llt;
 565                 } else {
 566                         new_ops->llo_delete(env, lov, state);
 567                         new_ops->llo_fini(env, lov, state);
 568                         /* this file becomes an EMPTY file. */
 569                 }
 570         }
 571         RETURN(result);
 572 }
 573
 574 /*****************************************************************************
 575  *
 576  * Lov object operations.
 577  *
 578  */
 579
 580 int lov_object_init(const struct lu_env *env, struct lu_object *obj,
 581                     const struct lu_object_conf *conf)
 582 {
 583         struct lov_device            *dev   = lu2lov_dev(obj->lo_dev);
 584         struct lov_object            *lov   = lu2lov(obj);
 585         const struct cl_object_conf  *cconf = lu2cl_conf(conf);
 586         union  lov_layout_state      *set   = &lov->u;
 587         const struct lov_layout_operations *ops;
 588         int result;
 589
 590         ENTRY;
 591         cfs_init_rwsem(&lov->lo_type_guard);
 592         cfs_waitq_init(&lov->lo_waitq);
 593
 594         /* no locking is necessary, as object is being created */
 595         lov->lo_type = cconf->u.coc_md->lsm != NULL ? LLT_RAID0 : LLT_EMPTY;
 596         ops = &lov_dispatch[lov->lo_type];
 597         result = ops->llo_init(env, dev, lov, cconf, set);
 598         if (result == 0)
 599                 ops->llo_install(env, lov, set);
 600         RETURN(result);
 601 }
 602
 603 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
 604                         const struct cl_object_conf *conf)
 605 {
 606         struct lov_stripe_md *lsm = conf->u.coc_md->lsm;
 607         struct lov_object *lov = cl2lov(obj);
 608         int result = 0;
 609         ENTRY;
 610
 611         /*
 612          * Only LLT_EMPTY <-> LLT_RAID0 transitions are supported.
 613          */
 614         LASSERT(lov->lo_owner != cfs_current());
 615         cfs_down_write(&lov->lo_type_guard);
 616         LASSERT(lov->lo_owner == NULL);
 617         lov->lo_owner = cfs_current();
 618
 619         if (conf->coc_invalidate) {
 620                 lov->lo_lsm_invalid = 1;
 621                 GOTO(out, result = 0);
 622         }
 623
 624         if (conf->coc_validate_only) {
 625                 if (!lov->lo_lsm_invalid)
 626                         GOTO(out, result = 0);
 627
 628                 lov_layout_wait(env, lov);
 629                 /* fall through to set up new layout */
 630         }
 631
 632         switch (lov->lo_type) {
 633         case LLT_EMPTY:
 634                 if (lsm != NULL)
 635                         result = lov_layout_change(env, lov, LLT_RAID0, conf);
 636                 break;
 637         case LLT_RAID0:
 638                 if (lsm == NULL)
 639                         result = lov_layout_change(env, lov, LLT_EMPTY, conf);
 640                 else if (lov_stripe_md_cmp(lov->lo_lsm, lsm))
 641                         result = -EOPNOTSUPP;
 642                 break;
 643         default:
 644                 LBUG();
 645         }
 646         lov->lo_lsm_invalid = result != 0;
 647         EXIT;
 648
 649 out:
 650         lov->lo_owner = NULL;
 651         cfs_up_write(&lov->lo_type_guard);
 652         RETURN(result);
 653 }
 654
 655 static void lov_object_delete(const struct lu_env *env, struct lu_object *obj)
 656 {
 657         struct lov_object *lov = lu2lov(obj);
 658
 659         ENTRY;
 660         LOV_2DISPATCH_VOID(lov, llo_delete, env, lov, &lov->u);
 661         EXIT;
 662 }
 663
 664 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
 665 {
 666         struct lov_object *lov = lu2lov(obj);
 667
 668         ENTRY;
 669         LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
 670         lu_object_fini(obj);
 671         OBD_SLAB_FREE_PTR(lov, lov_object_kmem);
 672         EXIT;
 673 }
 674
 675 static int lov_object_print(const struct lu_env *env, void *cookie,
 676                             lu_printer_t p, const struct lu_object *o)
 677 {
 678         return LOV_2DISPATCH(lu2lov(o), llo_print, env, cookie, p, o);
 679 }
 680
 681 struct cl_page *lov_page_init(const struct lu_env *env, struct cl_object *obj,
 682                               struct cl_page *page, cfs_page_t *vmpage)
 683 {
 684         return LOV_2DISPATCH(cl2lov(obj),
 685                              llo_page_init, env, obj, page, vmpage);
 686 }
 687
 688 /**
 689  * Implements cl_object_operations::clo_io_init() method for lov
 690  * layer. Dispatches to the appropriate layout io initialization method.
 691  */
 692 int lov_io_init(const struct lu_env *env, struct cl_object *obj,
 693                 struct cl_io *io)
 694 {
 695         struct lov_io *lio = lov_env_io(env);
 696
 697         CL_IO_SLICE_CLEAN(lov_env_io(env), lis_cl);
 698
 699         /* hold lsm before initializing because io relies on it */
 700         lio->lis_lsm = lov_lsm_addref(cl2lov(obj));
 701
 702         /*
 703          * Do not take lock in case of CIT_MISC io, because
 704          *
 705          *     - if this is an io for a glimpse, then we don't care;
 706          *
 707          *     - if this not a glimpse (writepage or lock cancellation), then
 708          *       layout change cannot happen because a page or a lock
 709          *       already exist; and
 710          *
 711          *     - lock ordering (lock mutex nests within layout rw-semaphore)
 712          *       is obeyed in case of lock cancellation.
 713          */
 714         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
 715                                      io->ci_type != CIT_MISC, env, obj, io);
 716 }
 717
 718 /**
 719  * An implementation of cl_object_operations::clo_attr_get() method for lov
 720  * layer. For raid0 layout this collects and merges attributes of all
 721  * sub-objects.
 722  */
 723 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
 724                         struct cl_attr *attr)
 725 {
 726         /* do not take lock, as this function is called under a
 727          * spin-lock. Layout is protected from changing by ongoing IO. */
 728         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
 729 }
 730
 731 static int lov_attr_set(const struct lu_env *env, struct cl_object *obj,
 732                         const struct cl_attr *attr, unsigned valid)
 733 {
 734         /*
 735          * No dispatch is required here, as no layout implements this.
 736          */
 737         return 0;
 738 }
 739
 740 int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
 741                   struct cl_lock *lock, const struct cl_io *io)
 742 {
 743         return LOV_2DISPATCH(cl2lov(obj), llo_lock_init, env, obj, lock, io);
 744 }
 745
 746 static const struct cl_object_operations lov_ops = {
 747         .coo_page_init = lov_page_init,
 748         .coo_lock_init = lov_lock_init,
 749         .coo_io_init   = lov_io_init,
 750         .coo_attr_get  = lov_attr_get,
 751         .coo_attr_set  = lov_attr_set,
 752         .coo_conf_set  = lov_conf_set
 753 };
 754
 755 static const struct lu_object_operations lov_lu_obj_ops = {
 756         .loo_object_init      = lov_object_init,
 757         .loo_object_delete    = lov_object_delete,
 758         .loo_object_release   = NULL,
 759         .loo_object_free      = lov_object_free,
 760         .loo_object_print     = lov_object_print,
 761         .loo_object_invariant = NULL
 762 };
 763
 764 struct lu_object *lov_object_alloc(const struct lu_env *env,
 765                                    const struct lu_object_header *unused,
 766                                    struct lu_device *dev)
 767 {
 768         struct lov_object *lov;
 769         struct lu_object  *obj;
 770
 771         ENTRY;
 772         OBD_SLAB_ALLOC_PTR_GFP(lov, lov_object_kmem, CFS_ALLOC_IO);
 773         if (lov != NULL) {
 774                 obj = lov2lu(lov);
 775                 lu_object_init(obj, NULL, dev);
 776                 lov->lo_cl.co_ops = &lov_ops;
 777                 lov->lo_type = -1; /* invalid, to catch uninitialized type */
 778                 /*
 779                  * object io operation vector (cl_object::co_iop) is installed
 780                  * later in lov_object_init(), as different vectors are used
 781                  * for object with different layouts.
 782                  */
 783                 obj->lo_ops = &lov_lu_obj_ops;
 784         } else
 785                 obj = NULL;
 786         RETURN(obj);
 787 }
 788
 789 struct lov_stripe_md *lov_lsm_addref(struct lov_object *lov)
 790 {
 791         struct lov_stripe_md *lsm = NULL;
 792
 793         lov_conf_freeze(lov);
 794         if (!lov->lo_lsm_invalid && lov->lo_lsm != NULL) {
 795                 lsm = lsm_addref(lov->lo_lsm);
 796                 CDEBUG(D_INODE, "lsm %p addref %d by %p.\n",
 797                         lsm, cfs_atomic_read(&lsm->lsm_refc), cfs_current());
 798         }
 799         lov_conf_thaw(lov);
 800         return lsm;
 801 }
 802
 803 void lov_lsm_decref(struct lov_object *lov, struct lov_stripe_md *lsm)
 804 {
 805         if (lsm == NULL)
 806                 return;
 807
 808         CDEBUG(D_INODE, "lsm %p decref %d by %p.\n",
 809                 lsm, cfs_atomic_read(&lsm->lsm_refc), cfs_current());
 810
 811         if (lov_free_memmd(&lsm) <= 1 && lov->lo_lsm_invalid)
 812                 cfs_waitq_signal(&lov->lo_waitq);
 813 }
 814
 815 struct lov_stripe_md *lov_lsm_get(struct cl_object *clobj)
 816 {
 817         struct lu_object *luobj;
 818         struct lov_stripe_md *lsm = NULL;
 819
 820         if (clobj == NULL)
 821                 return NULL;
 822
 823         luobj = lu_object_locate(&cl_object_header(clobj)->coh_lu,
 824                                  &lov_device_type);
 825         if (luobj != NULL)
 826                 lsm = lov_lsm_addref(lu2lov(luobj));
 827         return lsm;
 828 }
 829 EXPORT_SYMBOL(lov_lsm_get);
 830
 831 void lov_lsm_put(struct cl_object *clobj, struct lov_stripe_md *lsm)
 832 {
 833         struct lu_object *luobj;
 834
 835         if (clobj == NULL || lsm == NULL)
 836                 return;
 837
 838         luobj = lu_object_locate(&cl_object_header(clobj)->coh_lu,
 839                                  &lov_device_type);
 840         LASSERT(luobj != NULL);
 841
 842         lov_lsm_decref(lu2lov(luobj), lsm);
 843 }
 844 EXPORT_SYMBOL(lov_lsm_put);
 845
 846 int lov_read_and_clear_async_rc(struct cl_object *clob)
 847 {
 848         struct lu_object *luobj;
 849         int rc = 0;
 850         ENTRY;
 851
 852         luobj = lu_object_locate(&cl_object_header(clob)->coh_lu,
 853                                  &lov_device_type);
 854         if (luobj != NULL) {
 855                 struct lov_object *lov = lu2lov(luobj);
 856
 857                 lov_conf_freeze(lov);
 858                 switch (lov->lo_type) {
 859                 case LLT_RAID0: {
 860                         struct lov_stripe_md *lsm;
 861                         int i;
 862
 863                         lsm = lov->lo_lsm;
 864                         LASSERT(lsm != NULL);
 865                         for (i = 0; i < lsm->lsm_stripe_count; i++) {
 866                                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
 867                                 if (loi->loi_ar.ar_rc && !rc)
 868                                         rc = loi->loi_ar.ar_rc;
 869                                 loi->loi_ar.ar_rc = 0;
 870                         }
 871                 }
 872                 case LLT_EMPTY:
 873                         break;
 874                 default:
 875                         LBUG();
 876                 }
 877                 lov_conf_thaw(lov);
 878         }
 879         RETURN(rc);
 880 }
 881 EXPORT_SYMBOL(lov_read_and_clear_async_rc);
 882
 883 /** @} lov */