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         lov->u = *state;
 111 }
 112
 113 static struct cl_object *lov_sub_find(const struct lu_env *env,
 114                                       struct cl_device *dev,
 115                                       const struct lu_fid *fid,
 116                                       const struct cl_object_conf *conf)
 117 {
 118         struct lu_object *o;
 119
 120         ENTRY;
 121         o = lu_object_find_at(env, cl2lu_dev(dev), fid, &conf->coc_lu);
 122         LASSERT(ergo(!IS_ERR(o), o->lo_dev->ld_type == &lovsub_device_type));
 123         RETURN(lu2cl(o));
 124 }
 125
 126 static int lov_init_sub(const struct lu_env *env, struct lov_object *lov,
 127                         struct cl_object *stripe,
 128                         struct lov_layout_raid0 *r0, int idx)
 129 {
 130         struct cl_object_header *hdr;
 131         struct cl_object_header *subhdr;
 132         struct cl_object_header *parent;
 133         struct lov_oinfo        *oinfo;
 134         int result;
 135
 136         hdr    = cl_object_header(lov2cl(lov));
 137         subhdr = cl_object_header(stripe);
 138         parent = subhdr->coh_parent;
 139
 140         oinfo = lov->lo_lsm->lsm_oinfo[idx];
 141         CDEBUG(D_INODE, DFID"@%p[%d] -> "DFID"@%p: id: "LPU64" seq: "LPU64
 142                " idx: %d gen: %d\n",
 143                PFID(&subhdr->coh_lu.loh_fid), subhdr, idx,
 144                PFID(&hdr->coh_lu.loh_fid), hdr,
 145                oinfo->loi_id, oinfo->loi_seq,
 146                oinfo->loi_ost_idx, oinfo->loi_ost_gen);
 147
 148         if (parent == NULL) {
 149                 subhdr->coh_parent = hdr;
 150                 subhdr->coh_nesting = hdr->coh_nesting + 1;
 151                 lu_object_ref_add(&stripe->co_lu, "lov-parent", lov);
 152                 r0->lo_sub[idx] = cl2lovsub(stripe);
 153                 r0->lo_sub[idx]->lso_super = lov;
 154                 r0->lo_sub[idx]->lso_index = idx;
 155                 result = 0;
 156         } else {
 157                 CERROR("Stripe is already owned by other file (%d).\n", idx);
 158                 LU_OBJECT_DEBUG(D_ERROR, env, &stripe->co_lu, "\n");
 159                 LU_OBJECT_DEBUG(D_ERROR, env, lu_object_top(&parent->coh_lu),
 160                                 "old\n");
 161                 LU_OBJECT_HEADER(D_ERROR, env, lov2lu(lov), "new\n");
 162                 cl_object_put(env, stripe);
 163                 result = -EIO;
 164         }
 165         return result;
 166 }
 167
 168 static int lov_init_raid0(const struct lu_env *env,
 169                           struct lov_device *dev, struct lov_object *lov,
 170                           const struct cl_object_conf *conf,
 171                           union  lov_layout_state *state)
 172 {
 173         int result;
 174         int i;
 175
 176         struct cl_object        *stripe;
 177         struct lov_thread_info  *lti     = lov_env_info(env);
 178         struct cl_object_conf   *subconf = &lti->lti_stripe_conf;
 179         struct lov_stripe_md    *lsm     = conf->u.coc_md->lsm;
 180         struct lu_fid           *ofid    = &lti->lti_fid;
 181         struct lov_layout_raid0 *r0      = &state->raid0;
 182
 183         ENTRY;
 184
 185         if (lsm->lsm_magic != LOV_MAGIC_V1 && lsm->lsm_magic != LOV_MAGIC_V3) {
 186                 dump_lsm(D_ERROR, lsm);
 187                 LASSERTF(0, "magic mismatch, expected %d/%d, actual %d.\n",
 188                          LOV_MAGIC_V1, LOV_MAGIC_V3, lsm->lsm_magic);
 189         }
 190
 191         LASSERT(lov->lo_lsm == NULL);
 192         lov->lo_lsm = lsm_addref(lsm);
 193         r0->lo_nr  = lsm->lsm_stripe_count;
 194         LASSERT(r0->lo_nr <= lov_targets_nr(dev));
 195
 196         OBD_ALLOC_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 197         if (r0->lo_sub != NULL) {
 198                 result = 0;
 199                 subconf->coc_inode = conf->coc_inode;
 200                 cfs_spin_lock_init(&r0->lo_sub_lock);
 201                 /*
 202                  * Create stripe cl_objects.
 203                  */
 204                 for (i = 0; i < r0->lo_nr && result == 0; ++i) {
 205                         struct cl_device *subdev;
 206                         struct lov_oinfo *oinfo = lsm->lsm_oinfo[i];
 207                         int ost_idx = oinfo->loi_ost_idx;
 208
 209                         fid_ostid_unpack(ofid, &oinfo->loi_oi,
 210                                          oinfo->loi_ost_idx);
 211                         subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
 212                         subconf->u.coc_oinfo = oinfo;
 213                         LASSERTF(subdev != NULL, "not init ost %d\n", ost_idx);
 214                         stripe = lov_sub_find(env, subdev, ofid, subconf);
 215                         if (!IS_ERR(stripe))
 216                                 result = lov_init_sub(env, lov, stripe, r0, i);
 217                         else
 218                                 result = PTR_ERR(stripe);
 219                 }
 220         } else
 221                 result = -ENOMEM;
 222         RETURN(result);
 223 }
 224
 225 static int lov_delete_empty(const struct lu_env *env, struct lov_object *lov,
 226                             union lov_layout_state *state)
 227 {
 228         LASSERT(lov->lo_type == LLT_EMPTY);
 229         return 0;
 230 }
 231
 232 static void lov_subobject_kill(const struct lu_env *env, struct lov_object *lov,
 233                                struct lovsub_object *los, int idx)
 234 {
 235         struct cl_object        *sub;
 236         struct lov_layout_raid0 *r0;
 237         struct lu_site          *site;
 238         struct lu_site_bkt_data *bkt;
 239         cfs_waitlink_t          *waiter;
 240
 241         r0  = &lov->u.raid0;
 242         LASSERT(r0->lo_sub[idx] == los);
 243
 244         sub  = lovsub2cl(los);
 245         site = sub->co_lu.lo_dev->ld_site;
 246         bkt  = lu_site_bkt_from_fid(site, &sub->co_lu.lo_header->loh_fid);
 247
 248         cl_object_kill(env, sub);
 249         /* release a reference to the sub-object and ... */
 250         lu_object_ref_del(&sub->co_lu, "lov-parent", lov);
 251         cl_object_put(env, sub);
 252
 253         /* ... wait until it is actually destroyed---sub-object clears its
 254          * ->lo_sub[] slot in lovsub_object_fini() */
 255         if (r0->lo_sub[idx] == los) {
 256                 waiter = &lov_env_info(env)->lti_waiter;
 257                 cfs_waitlink_init(waiter);
 258                 cfs_waitq_add(&bkt->lsb_marche_funebre, waiter);
 259                 cfs_set_current_state(CFS_TASK_UNINT);
 260                 while (1) {
 261                         /* this wait-queue is signaled at the end of
 262                          * lu_object_free(). */
 263                         cfs_set_current_state(CFS_TASK_UNINT);
 264                         cfs_spin_lock(&r0->lo_sub_lock);
 265                         if (r0->lo_sub[idx] == los) {
 266                                 cfs_spin_unlock(&r0->lo_sub_lock);
 267                                 cfs_waitq_wait(waiter, CFS_TASK_UNINT);
 268                         } else {
 269                                 cfs_spin_unlock(&r0->lo_sub_lock);
 270                                 cfs_set_current_state(CFS_TASK_RUNNING);
 271                                 break;
 272                         }
 273                 }
 274                 cfs_waitq_del(&bkt->lsb_marche_funebre, waiter);
 275         }
 276         LASSERT(r0->lo_sub[idx] == NULL);
 277 }
 278
 279 static int lov_delete_raid0(const struct lu_env *env, struct lov_object *lov,
 280                             union lov_layout_state *state)
 281 {
 282         struct lov_layout_raid0 *r0 = &state->raid0;
 283         struct lov_stripe_md    *lsm = lov->lo_lsm;
 284         int i;
 285
 286         ENTRY;
 287
 288         dump_lsm(D_INODE, lsm);
 289         if (cfs_atomic_read(&lsm->lsm_refc) > 1)
 290                 RETURN(-EBUSY);
 291
 292         if (r0->lo_sub != NULL) {
 293                 for (i = 0; i < r0->lo_nr; ++i) {
 294                         struct lovsub_object *los = r0->lo_sub[i];
 295
 296                         if (los != NULL)
 297                                 /*
 298                                  * If top-level object is to be evicted from
 299                                  * the cache, so are its sub-objects.
 300                                  */
 301                                 lov_subobject_kill(env, lov, los, i);
 302                 }
 303         }
 304         RETURN(0);
 305 }
 306
 307 static void lov_fini_empty(const struct lu_env *env, struct lov_object *lov,
 308                            union lov_layout_state *state)
 309 {
 310         LASSERT(lov->lo_type == LLT_EMPTY);
 311 }
 312
 313 static void lov_fini_raid0(const struct lu_env *env, struct lov_object *lov,
 314                            union lov_layout_state *state)
 315 {
 316         struct lov_layout_raid0 *r0 = &state->raid0;
 317         ENTRY;
 318
 319         if (r0->lo_sub != NULL) {
 320                 OBD_FREE_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 321                 r0->lo_sub = NULL;
 322         }
 323
 324         dump_lsm(D_INODE, lov->lo_lsm);
 325         lov_free_memmd(&lov->lo_lsm);
 326
 327         EXIT;
 328 }
 329
 330 static int lov_print_empty(const struct lu_env *env, void *cookie,
 331                            lu_printer_t p, const struct lu_object *o)
 332 {
 333         (*p)(env, cookie, "empty\n");
 334         return 0;
 335 }
 336
 337 static int lov_print_raid0(const struct lu_env *env, void *cookie,
 338                            lu_printer_t p, const struct lu_object *o)
 339 {
 340         struct lov_object       *lov = lu2lov(o);
 341         struct lov_layout_raid0 *r0  = lov_r0(lov);
 342         int i;
 343
 344         (*p)(env, cookie, "stripes: %d:\n", r0->lo_nr);
 345         for (i = 0; i < r0->lo_nr; ++i) {
 346                 struct lu_object *sub;
 347
 348                 if (r0->lo_sub[i] != NULL) {
 349                         sub = lovsub2lu(r0->lo_sub[i]);
 350                         lu_object_print(env, cookie, p, sub);
 351                 } else
 352                         (*p)(env, cookie, "sub %d absent\n", i);
 353         }
 354         return 0;
 355 }
 356
 357 /**
 358  * Implements cl_object_operations::coo_attr_get() method for an object
 359  * without stripes (LLT_EMPTY layout type).
 360  *
 361  * The only attributes this layer is authoritative in this case is
 362  * cl_attr::cat_blocks---it's 0.
 363  */
 364 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
 365                               struct cl_attr *attr)
 366 {
 367         attr->cat_blocks = 0;
 368         return 0;
 369 }
 370
 371 static int lov_attr_get_raid0(const struct lu_env *env, struct cl_object *obj,
 372                               struct cl_attr *attr)
 373 {
 374         struct lov_object       *lov = cl2lov(obj);
 375         struct lov_layout_raid0 *r0 = lov_r0(lov);
 376         struct lov_stripe_md    *lsm = lov->lo_lsm;
 377         struct ost_lvb          *lvb = &lov_env_info(env)->lti_lvb;
 378         __u64                    kms;
 379         int                      result = 0;
 380
 381         ENTRY;
 382
 383         /* this is called w/o holding type guard mutex, so it must be inside
 384          * an on going IO otherwise lsm may be replaced. */
 385         LASSERT(cfs_atomic_read(&lsm->lsm_refc) > 1);
 386
 387         if (!r0->lo_attr_valid) {
 388                 /*
 389                  * Fill LVB with attributes already initialized by the upper
 390                  * layer.
 391                  */
 392                 cl_attr2lvb(lvb, attr);
 393                 kms = attr->cat_kms;
 394
 395                 /*
 396                  * XXX that should be replaced with a loop over sub-objects,
 397                  * doing cl_object_attr_get() on them. But for now, let's
 398                  * reuse old lov code.
 399                  */
 400
 401                 /*
 402                  * XXX take lsm spin-lock to keep lov_merge_lvb_kms()
 403                  * happy. It's not needed, because new code uses
 404                  * ->coh_attr_guard spin-lock to protect consistency of
 405                  * sub-object attributes.
 406                  */
 407                 lov_stripe_lock(lsm);
 408                 result = lov_merge_lvb_kms(lsm, lvb, &kms);
 409                 lov_stripe_unlock(lsm);
 410                 if (result == 0) {
 411                         cl_lvb2attr(attr, lvb);
 412                         attr->cat_kms = kms;
 413                         r0->lo_attr_valid = 1;
 414                         r0->lo_attr = *attr;
 415                 }
 416         } else
 417                 *attr = r0->lo_attr;
 418         RETURN(result);
 419 }
 420
 421 const static struct lov_layout_operations lov_dispatch[] = {
 422         [LLT_EMPTY] = {
 423                 .llo_init      = lov_init_empty,
 424                 .llo_delete    = lov_delete_empty,
 425                 .llo_fini      = lov_fini_empty,
 426                 .llo_install   = lov_install_empty,
 427                 .llo_print     = lov_print_empty,
 428                 .llo_page_init = lov_page_init_empty,
 429                 .llo_lock_init = NULL,
 430                 .llo_io_init   = lov_io_init_empty,
 431                 .llo_getattr   = lov_attr_get_empty
 432         },
 433         [LLT_RAID0] = {
 434                 .llo_init      = lov_init_raid0,
 435                 .llo_delete    = lov_delete_raid0,
 436                 .llo_fini      = lov_fini_raid0,
 437                 .llo_install   = lov_install_raid0,
 438                 .llo_print     = lov_print_raid0,
 439                 .llo_page_init = lov_page_init_raid0,
 440                 .llo_lock_init = lov_lock_init_raid0,
 441                 .llo_io_init   = lov_io_init_raid0,
 442                 .llo_getattr   = lov_attr_get_raid0
 443         }
 444 };
 445
 446
 447 /**
 448  * Performs a double-dispatch based on the layout type of an object.
 449  */
 450 #define LOV_2DISPATCH_NOLOCK(obj, op, ...)                              \
 451 ({                                                                      \
 452         struct lov_object                      *__obj = (obj);          \
 453         enum lov_layout_type                    __llt;                  \
 454                                                                         \
 455         __llt = __obj->lo_type;                                         \
 456         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 457         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 458 })
 459
 460 static inline void lov_conf_freeze(struct lov_object *lov)
 461 {
 462         if (lov->lo_owner != cfs_current())
 463                 cfs_down_read(&lov->lo_type_guard);
 464 }
 465
 466 static inline void lov_conf_thaw(struct lov_object *lov)
 467 {
 468         if (lov->lo_owner != cfs_current())
 469                 cfs_up_read(&lov->lo_type_guard);
 470 }
 471
 472 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...)                       \
 473 ({                                                                      \
 474         struct lov_object                      *__obj = (obj);          \
 475         int                                     __lock = !!(lock);      \
 476         typeof(lov_dispatch[0].op(__VA_ARGS__)) __result;               \
 477                                                                         \
 478         if (__lock)                                                     \
 479                 lov_conf_freeze(__obj);                                 \
 480         __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__);          \
 481         if (__lock)                                                     \
 482                 lov_conf_thaw(__obj);                                   \
 483         __result;                                                       \
 484 })
 485
 486 /**
 487  * Performs a locked double-dispatch based on the layout type of an object.
 488  */
 489 #define LOV_2DISPATCH(obj, op, ...)                     \
 490         LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
 491
 492 #define LOV_2DISPATCH_VOID(obj, op, ...)                                \
 493 do {                                                                    \
 494         struct lov_object                      *__obj = (obj);          \
 495         enum lov_layout_type                    __llt;                  \
 496                                                                         \
 497         lov_conf_freeze(__obj);                                         \
 498         __llt = __obj->lo_type;                                         \
 499         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 500         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 501         lov_conf_thaw(__obj);                                           \
 502 } while (0)
 503
 504 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov)
 505 {
 506         struct l_wait_info lwi = { 0 };
 507         struct lov_stripe_md *lsm = lov->lo_lsm;
 508         ENTRY;
 509
 510         if (!lov->lo_lsm_invalid || lsm == NULL)
 511                 RETURN(0);
 512
 513         l_wait_event(lov->lo_waitq, cfs_atomic_read(&lsm->lsm_refc) == 1, &lwi);
 514         RETURN(0);
 515 }
 516
 517 static int lov_layout_change(const struct lu_env *env,
 518                              struct lov_object *lov, enum lov_layout_type llt,
 519                              const struct cl_object_conf *conf)
 520 {
 521         int result;
 522         union lov_layout_state *state = &lov_env_info(env)->lti_state;
 523         const struct lov_layout_operations *old_ops;
 524         const struct lov_layout_operations *new_ops;
 525
 526         struct cl_object_header *hdr = cl_object_header(&lov->lo_cl);
 527         void *cookie;
 528         struct lu_env *nested;
 529         int refcheck;
 530
 531         LASSERT(0 <= lov->lo_type && lov->lo_type < ARRAY_SIZE(lov_dispatch));
 532         LASSERT(0 <= llt && llt < ARRAY_SIZE(lov_dispatch));
 533         ENTRY;
 534
 535         cookie = cl_env_reenter();
 536         nested = cl_env_get(&refcheck);
 537         if (!IS_ERR(nested))
 538                 cl_object_prune(nested, &lov->lo_cl);
 539         else
 540                 result = PTR_ERR(nested);
 541         cl_env_put(nested, &refcheck);
 542         cl_env_reexit(cookie);
 543
 544         old_ops = &lov_dispatch[lov->lo_type];
 545         new_ops = &lov_dispatch[llt];
 546
 547         result = old_ops->llo_delete(env, lov, &lov->u);
 548         if (result == 0) {
 549                 old_ops->llo_fini(env, lov, &lov->u);
 550                 LASSERT(cfs_list_empty(&hdr->coh_locks));
 551                 LASSERT(hdr->coh_tree.rnode == NULL);
 552                 LASSERT(hdr->coh_pages == 0);
 553
 554                 result = new_ops->llo_init(env,
 555                                         lu2lov_dev(lov->lo_cl.co_lu.lo_dev),
 556                                         lov, conf, state);
 557                 if (result == 0) {
 558                         new_ops->llo_install(env, lov, state);
 559                         lov->lo_type = llt;
 560                 }
 561         }
 562         RETURN(result);
 563 }
 564
 565 /*****************************************************************************
 566  *
 567  * Lov object operations.
 568  *
 569  */
 570
 571 int lov_object_init(const struct lu_env *env, struct lu_object *obj,
 572                     const struct lu_object_conf *conf)
 573 {
 574         struct lov_device            *dev   = lu2lov_dev(obj->lo_dev);
 575         struct lov_object            *lov   = lu2lov(obj);
 576         const struct cl_object_conf  *cconf = lu2cl_conf(conf);
 577         union  lov_layout_state      *set   = &lov_env_info(env)->lti_state;
 578         const struct lov_layout_operations *ops;
 579         int result;
 580
 581         ENTRY;
 582         cfs_init_rwsem(&lov->lo_type_guard);
 583         cfs_waitq_init(&lov->lo_waitq);
 584
 585         /* no locking is necessary, as object is being created */
 586         lov->lo_type = cconf->u.coc_md->lsm != NULL ? LLT_RAID0 : LLT_EMPTY;
 587         ops = &lov_dispatch[lov->lo_type];
 588         result = ops->llo_init(env, dev, lov, cconf, set);
 589         if (result == 0)
 590                 ops->llo_install(env, lov, set);
 591         else
 592                 ops->llo_fini(env, lov, set);
 593         RETURN(result);
 594 }
 595
 596 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
 597                         const struct cl_object_conf *conf)
 598 {
 599         struct lov_stripe_md *lsm = conf->u.coc_md->lsm;
 600         struct lov_object *lov = cl2lov(obj);
 601         int result = 0;
 602         ENTRY;
 603
 604         /*
 605          * Only LLT_EMPTY <-> LLT_RAID0 transitions are supported.
 606          */
 607         LASSERT(lov->lo_owner != cfs_current());
 608         cfs_down_write(&lov->lo_type_guard);
 609         LASSERT(lov->lo_owner == NULL);
 610         lov->lo_owner = cfs_current();
 611
 612         if (conf->coc_invalidate) {
 613                 lov->lo_lsm_invalid = 1;
 614                 GOTO(out, result = 0);
 615         }
 616
 617         if (conf->coc_validate_only) {
 618                 if (!lov->lo_lsm_invalid)
 619                         GOTO(out, result = 0);
 620
 621                 lov_layout_wait(env, lov);
 622                 /* fall through to set up new layout */
 623         }
 624
 625         switch (lov->lo_type) {
 626         case LLT_EMPTY:
 627                 if (lsm != NULL)
 628                         result = lov_layout_change(env, lov, LLT_RAID0, conf);
 629                 break;
 630         case LLT_RAID0:
 631                 if (lsm == NULL)
 632                         result = lov_layout_change(env, lov, LLT_EMPTY, conf);
 633                 else if (lov_stripe_md_cmp(lov->lo_lsm, lsm))
 634                         result = -EOPNOTSUPP;
 635                 break;
 636         default:
 637                 LBUG();
 638         }
 639         lov->lo_lsm_invalid = result != 0;
 640         EXIT;
 641
 642 out:
 643         lov->lo_owner = NULL;
 644         cfs_up_write(&lov->lo_type_guard);
 645         RETURN(result);
 646 }
 647
 648 static void lov_object_delete(const struct lu_env *env, struct lu_object *obj)
 649 {
 650         struct lov_object *lov = lu2lov(obj);
 651
 652         ENTRY;
 653         LOV_2DISPATCH_VOID(lov, llo_delete, env, lov, &lov->u);
 654         EXIT;
 655 }
 656
 657 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
 658 {
 659         struct lov_object *lov = lu2lov(obj);
 660
 661         ENTRY;
 662         LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
 663         lu_object_fini(obj);
 664         OBD_SLAB_FREE_PTR(lov, lov_object_kmem);
 665         EXIT;
 666 }
 667
 668 static int lov_object_print(const struct lu_env *env, void *cookie,
 669                             lu_printer_t p, const struct lu_object *o)
 670 {
 671         return LOV_2DISPATCH(lu2lov(o), llo_print, env, cookie, p, o);
 672 }
 673
 674 struct cl_page *lov_page_init(const struct lu_env *env, struct cl_object *obj,
 675                               struct cl_page *page, cfs_page_t *vmpage)
 676 {
 677         return LOV_2DISPATCH(cl2lov(obj),
 678                              llo_page_init, env, obj, page, vmpage);
 679 }
 680
 681 /**
 682  * Implements cl_object_operations::clo_io_init() method for lov
 683  * layer. Dispatches to the appropriate layout io initialization method.
 684  */
 685 int lov_io_init(const struct lu_env *env, struct cl_object *obj,
 686                 struct cl_io *io)
 687 {
 688         struct lov_io *lio = lov_env_io(env);
 689
 690         CL_IO_SLICE_CLEAN(lov_env_io(env), lis_cl);
 691
 692         /* hold lsm before initializing because io relies on it */
 693         lio->lis_lsm = lov_lsm_addref(cl2lov(obj));
 694
 695         /*
 696          * Do not take lock in case of CIT_MISC io, because
 697          *
 698          *     - if this is an io for a glimpse, then we don't care;
 699          *
 700          *     - if this not a glimpse (writepage or lock cancellation), then
 701          *       layout change cannot happen because a page or a lock
 702          *       already exist; and
 703          *
 704          *     - lock ordering (lock mutex nests within layout rw-semaphore)
 705          *       is obeyed in case of lock cancellation.
 706          */
 707         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
 708                                      io->ci_type != CIT_MISC, env, obj, io);
 709 }
 710
 711 /**
 712  * An implementation of cl_object_operations::clo_attr_get() method for lov
 713  * layer. For raid0 layout this collects and merges attributes of all
 714  * sub-objects.
 715  */
 716 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
 717                         struct cl_attr *attr)
 718 {
 719         /* do not take lock, as this function is called under a
 720          * spin-lock. Layout is protected from changing by ongoing IO. */
 721         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
 722 }
 723
 724 static int lov_attr_set(const struct lu_env *env, struct cl_object *obj,
 725                         const struct cl_attr *attr, unsigned valid)
 726 {
 727         /*
 728          * No dispatch is required here, as no layout implements this.
 729          */
 730         return 0;
 731 }
 732
 733 int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
 734                   struct cl_lock *lock, const struct cl_io *io)
 735 {
 736         return LOV_2DISPATCH(cl2lov(obj), llo_lock_init, env, obj, lock, 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 */