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