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, 2012, Intel Corporation.
  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                 /* For sanity:test_206.
 137                  * Do not leave the object in cache to avoid accessing
 138                  * freed memory. This is because osc_object is referring to
 139                  * lov_oinfo of lsm_stripe_data which will be freed due to
 140                  * this failure. */
 141                 cl_object_kill(env, stripe);
 142                 cl_object_put(env, stripe);
 143                 return -EIO;
 144         }
 145
 146         hdr    = cl_object_header(lov2cl(lov));
 147         subhdr = cl_object_header(stripe);
 148         parent = subhdr->coh_parent;
 149
 150         oinfo = lov->lo_lsm->lsm_oinfo[idx];
 151         CDEBUG(D_INODE, DFID"@%p[%d] -> "DFID"@%p: id: "LPU64" seq: "LPU64
 152                " idx: %d gen: %d\n",
 153                PFID(&subhdr->coh_lu.loh_fid), subhdr, idx,
 154                PFID(&hdr->coh_lu.loh_fid), hdr,
 155                oinfo->loi_id, oinfo->loi_seq,
 156                oinfo->loi_ost_idx, oinfo->loi_ost_gen);
 157
 158         if (parent == NULL) {
 159                 subhdr->coh_parent = hdr;
 160                 subhdr->coh_nesting = hdr->coh_nesting + 1;
 161                 lu_object_ref_add(&stripe->co_lu, "lov-parent", lov);
 162                 r0->lo_sub[idx] = cl2lovsub(stripe);
 163                 r0->lo_sub[idx]->lso_super = lov;
 164                 r0->lo_sub[idx]->lso_index = idx;
 165                 result = 0;
 166         } else {
 167                 CERROR("Stripe is already owned by other file (%d).\n", idx);
 168                 LU_OBJECT_DEBUG(D_ERROR, env, &stripe->co_lu, "\n");
 169                 LU_OBJECT_DEBUG(D_ERROR, env, lu_object_top(&parent->coh_lu),
 170                                 "old\n");
 171                 LU_OBJECT_HEADER(D_ERROR, env, lov2lu(lov), "new\n");
 172                 cl_object_put(env, stripe);
 173                 result = -EIO;
 174         }
 175         return result;
 176 }
 177
 178 static int lov_init_raid0(const struct lu_env *env,
 179                           struct lov_device *dev, struct lov_object *lov,
 180                           const struct cl_object_conf *conf,
 181                           union  lov_layout_state *state)
 182 {
 183         int result;
 184         int i;
 185
 186         struct cl_object        *stripe;
 187         struct lov_thread_info  *lti     = lov_env_info(env);
 188         struct cl_object_conf   *subconf = &lti->lti_stripe_conf;
 189         struct lov_stripe_md    *lsm     = conf->u.coc_md->lsm;
 190         struct lu_fid           *ofid    = &lti->lti_fid;
 191         struct lov_layout_raid0 *r0      = &state->raid0;
 192
 193         ENTRY;
 194
 195         if (lsm->lsm_magic != LOV_MAGIC_V1 && lsm->lsm_magic != LOV_MAGIC_V3) {
 196                 dump_lsm(D_ERROR, lsm);
 197                 LASSERTF(0, "magic mismatch, expected %d/%d, actual %d.\n",
 198                          LOV_MAGIC_V1, LOV_MAGIC_V3, lsm->lsm_magic);
 199         }
 200
 201         LASSERT(lov->lo_lsm == NULL);
 202         lov->lo_lsm = lsm_addref(lsm);
 203         r0->lo_nr  = lsm->lsm_stripe_count;
 204         LASSERT(r0->lo_nr <= lov_targets_nr(dev));
 205
 206         OBD_ALLOC_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 207         if (r0->lo_sub != NULL) {
 208                 result = 0;
 209                 subconf->coc_inode = conf->coc_inode;
 210                 spin_lock_init(&r0->lo_sub_lock);
 211                 /*
 212                  * Create stripe cl_objects.
 213                  */
 214                 for (i = 0; i < r0->lo_nr && result == 0; ++i) {
 215                         struct cl_device *subdev;
 216                         struct lov_oinfo *oinfo = lsm->lsm_oinfo[i];
 217                         int ost_idx = oinfo->loi_ost_idx;
 218
 219                         fid_ostid_unpack(ofid, &oinfo->loi_oi,
 220                                          oinfo->loi_ost_idx);
 221                         subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
 222                         subconf->u.coc_oinfo = oinfo;
 223                         LASSERTF(subdev != NULL, "not init ost %d\n", ost_idx);
 224                         /* In the function below, .hs_keycmp resolves to
 225                          * lu_obj_hop_keycmp() */
 226                         /* coverity[overrun-buffer-val] */
 227                         stripe = lov_sub_find(env, subdev, ofid, subconf);
 228                         if (!IS_ERR(stripe))
 229                                 result = lov_init_sub(env, lov, stripe, r0, i);
 230                         else
 231                                 result = PTR_ERR(stripe);
 232                 }
 233         } else
 234                 result = -ENOMEM;
 235         RETURN(result);
 236 }
 237
 238 static int lov_delete_empty(const struct lu_env *env, struct lov_object *lov,
 239                             union lov_layout_state *state)
 240 {
 241         LASSERT(lov->lo_type == LLT_EMPTY);
 242         cl_object_prune(env, &lov->lo_cl);
 243         return 0;
 244 }
 245
 246 static void lov_subobject_kill(const struct lu_env *env, struct lov_object *lov,
 247                                struct lovsub_object *los, int idx)
 248 {
 249         struct cl_object        *sub;
 250         struct lov_layout_raid0 *r0;
 251         struct lu_site          *site;
 252         struct lu_site_bkt_data *bkt;
 253         cfs_waitlink_t          *waiter;
 254
 255         r0  = &lov->u.raid0;
 256         LASSERT(r0->lo_sub[idx] == los);
 257
 258         sub  = lovsub2cl(los);
 259         site = sub->co_lu.lo_dev->ld_site;
 260         bkt  = lu_site_bkt_from_fid(site, &sub->co_lu.lo_header->loh_fid);
 261
 262         cl_object_kill(env, sub);
 263         /* release a reference to the sub-object and ... */
 264         lu_object_ref_del(&sub->co_lu, "lov-parent", lov);
 265         cl_object_put(env, sub);
 266
 267         /* ... wait until it is actually destroyed---sub-object clears its
 268          * ->lo_sub[] slot in lovsub_object_fini() */
 269         if (r0->lo_sub[idx] == los) {
 270                 waiter = &lov_env_info(env)->lti_waiter;
 271                 cfs_waitlink_init(waiter);
 272                 cfs_waitq_add(&bkt->lsb_marche_funebre, waiter);
 273                 cfs_set_current_state(CFS_TASK_UNINT);
 274                 while (1) {
 275                         /* this wait-queue is signaled at the end of
 276                          * lu_object_free(). */
 277                         cfs_set_current_state(CFS_TASK_UNINT);
 278                         spin_lock(&r0->lo_sub_lock);
 279                         if (r0->lo_sub[idx] == los) {
 280                                 spin_unlock(&r0->lo_sub_lock);
 281                                 cfs_waitq_wait(waiter, CFS_TASK_UNINT);
 282                         } else {
 283                                 spin_unlock(&r0->lo_sub_lock);
 284                                 cfs_set_current_state(CFS_TASK_RUNNING);
 285                                 break;
 286                         }
 287                 }
 288                 cfs_waitq_del(&bkt->lsb_marche_funebre, waiter);
 289         }
 290         LASSERT(r0->lo_sub[idx] == NULL);
 291 }
 292
 293 static int lov_delete_raid0(const struct lu_env *env, struct lov_object *lov,
 294                             union lov_layout_state *state)
 295 {
 296         struct lov_layout_raid0 *r0 = &state->raid0;
 297         struct lov_stripe_md    *lsm = lov->lo_lsm;
 298         int i;
 299
 300         ENTRY;
 301
 302         dump_lsm(D_INODE, lsm);
 303         if (lov->lo_lsm_invalid && cfs_atomic_read(&lsm->lsm_refc) > 1)
 304                 RETURN(-EBUSY);
 305
 306         if (r0->lo_sub != NULL) {
 307                 for (i = 0; i < r0->lo_nr; ++i) {
 308                         struct lovsub_object *los = r0->lo_sub[i];
 309
 310                         if (los != NULL) {
 311                                 cl_locks_prune(env, &los->lso_cl, 1);
 312                                 /*
 313                                  * If top-level object is to be evicted from
 314                                  * the cache, so are its sub-objects.
 315                                  */
 316                                 lov_subobject_kill(env, lov, los, i);
 317                         }
 318                 }
 319         }
 320         RETURN(0);
 321 }
 322
 323 static void lov_fini_empty(const struct lu_env *env, struct lov_object *lov,
 324                            union lov_layout_state *state)
 325 {
 326         LASSERT(lov->lo_type == LLT_EMPTY);
 327 }
 328
 329 static void lov_fini_raid0(const struct lu_env *env, struct lov_object *lov,
 330                            union lov_layout_state *state)
 331 {
 332         struct lov_layout_raid0 *r0 = &state->raid0;
 333         ENTRY;
 334
 335         if (r0->lo_sub != NULL) {
 336                 OBD_FREE_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 337                 r0->lo_sub = NULL;
 338         }
 339
 340         dump_lsm(D_INODE, lov->lo_lsm);
 341         lov_free_memmd(&lov->lo_lsm);
 342
 343         EXIT;
 344 }
 345
 346 static int lov_print_empty(const struct lu_env *env, void *cookie,
 347                            lu_printer_t p, const struct lu_object *o)
 348 {
 349         (*p)(env, cookie, "empty\n");
 350         return 0;
 351 }
 352
 353 static int lov_print_raid0(const struct lu_env *env, void *cookie,
 354                            lu_printer_t p, const struct lu_object *o)
 355 {
 356         struct lov_object       *lov = lu2lov(o);
 357         struct lov_layout_raid0 *r0  = lov_r0(lov);
 358         int i;
 359
 360         (*p)(env, cookie, "stripes: %d:\n", r0->lo_nr);
 361         for (i = 0; i < r0->lo_nr; ++i) {
 362                 struct lu_object *sub;
 363
 364                 if (r0->lo_sub[i] != NULL) {
 365                         sub = lovsub2lu(r0->lo_sub[i]);
 366                         lu_object_print(env, cookie, p, sub);
 367                 } else
 368                         (*p)(env, cookie, "sub %d absent\n", i);
 369         }
 370         return 0;
 371 }
 372
 373 /**
 374  * Implements cl_object_operations::coo_attr_get() method for an object
 375  * without stripes (LLT_EMPTY layout type).
 376  *
 377  * The only attributes this layer is authoritative in this case is
 378  * cl_attr::cat_blocks---it's 0.
 379  */
 380 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
 381                               struct cl_attr *attr)
 382 {
 383         attr->cat_blocks = 0;
 384         return 0;
 385 }
 386
 387 static int lov_attr_get_raid0(const struct lu_env *env, struct cl_object *obj,
 388                               struct cl_attr *attr)
 389 {
 390         struct lov_object       *lov = cl2lov(obj);
 391         struct lov_layout_raid0 *r0 = lov_r0(lov);
 392         struct lov_stripe_md    *lsm = lov->lo_lsm;
 393         struct ost_lvb          *lvb = &lov_env_info(env)->lti_lvb;
 394         __u64                    kms;
 395         int                      result = 0;
 396
 397         ENTRY;
 398
 399         /* this is called w/o holding type guard mutex, so it must be inside
 400          * an on going IO otherwise lsm may be replaced.
 401          * LU-2117: it turns out there exists one exception. For mmaped files,
 402          * the lock of those files may be requested in the other file's IO
 403          * context, and this function is called in ccc_lock_state(), it will
 404          * hit this assertion.
 405          * Anyway, it's still okay to call attr_get w/o type guard as layout
 406          * can't go if locks exist. */
 407         /* LASSERT(cfs_atomic_read(&lsm->lsm_refc) > 1); */
 408
 409         if (!r0->lo_attr_valid) {
 410                 /*
 411                  * Fill LVB with attributes already initialized by the upper
 412                  * layer.
 413                  */
 414                 cl_attr2lvb(lvb, attr);
 415                 kms = attr->cat_kms;
 416
 417                 /*
 418                  * XXX that should be replaced with a loop over sub-objects,
 419                  * doing cl_object_attr_get() on them. But for now, let's
 420                  * reuse old lov code.
 421                  */
 422
 423                 /*
 424                  * XXX take lsm spin-lock to keep lov_merge_lvb_kms()
 425                  * happy. It's not needed, because new code uses
 426                  * ->coh_attr_guard spin-lock to protect consistency of
 427                  * sub-object attributes.
 428                  */
 429                 lov_stripe_lock(lsm);
 430                 result = lov_merge_lvb_kms(lsm, lvb, &kms);
 431                 lov_stripe_unlock(lsm);
 432                 if (result == 0) {
 433                         cl_lvb2attr(attr, lvb);
 434                         attr->cat_kms = kms;
 435                         r0->lo_attr_valid = 1;
 436                         r0->lo_attr = *attr;
 437                 }
 438         } else
 439                 *attr = r0->lo_attr;
 440         RETURN(result);
 441 }
 442
 443 const static struct lov_layout_operations lov_dispatch[] = {
 444         [LLT_EMPTY] = {
 445                 .llo_init      = lov_init_empty,
 446                 .llo_delete    = lov_delete_empty,
 447                 .llo_fini      = lov_fini_empty,
 448                 .llo_install   = lov_install_empty,
 449                 .llo_print     = lov_print_empty,
 450                 .llo_page_init = lov_page_init_empty,
 451                 .llo_lock_init = lov_lock_init_empty,
 452                 .llo_io_init   = lov_io_init_empty,
 453                 .llo_getattr   = lov_attr_get_empty
 454         },
 455         [LLT_RAID0] = {
 456                 .llo_init      = lov_init_raid0,
 457                 .llo_delete    = lov_delete_raid0,
 458                 .llo_fini      = lov_fini_raid0,
 459                 .llo_install   = lov_install_raid0,
 460                 .llo_print     = lov_print_raid0,
 461                 .llo_page_init = lov_page_init_raid0,
 462                 .llo_lock_init = lov_lock_init_raid0,
 463                 .llo_io_init   = lov_io_init_raid0,
 464                 .llo_getattr   = lov_attr_get_raid0
 465         }
 466 };
 467
 468
 469 /**
 470  * Performs a double-dispatch based on the layout type of an object.
 471  */
 472 #define LOV_2DISPATCH_NOLOCK(obj, op, ...)                              \
 473 ({                                                                      \
 474         struct lov_object                      *__obj = (obj);          \
 475         enum lov_layout_type                    __llt;                  \
 476                                                                         \
 477         __llt = __obj->lo_type;                                         \
 478         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 479         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 480 })
 481
 482 static inline void lov_conf_freeze(struct lov_object *lov)
 483 {
 484         if (lov->lo_owner != cfs_current())
 485                 down_read(&lov->lo_type_guard);
 486 }
 487
 488 static inline void lov_conf_thaw(struct lov_object *lov)
 489 {
 490         if (lov->lo_owner != cfs_current())
 491                 up_read(&lov->lo_type_guard);
 492 }
 493
 494 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...)                       \
 495 ({                                                                      \
 496         struct lov_object                      *__obj = (obj);          \
 497         int                                     __lock = !!(lock);      \
 498         typeof(lov_dispatch[0].op(__VA_ARGS__)) __result;               \
 499                                                                         \
 500         if (__lock)                                                     \
 501                 lov_conf_freeze(__obj);                                 \
 502         __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__);          \
 503         if (__lock)                                                     \
 504                 lov_conf_thaw(__obj);                                   \
 505         __result;                                                       \
 506 })
 507
 508 /**
 509  * Performs a locked double-dispatch based on the layout type of an object.
 510  */
 511 #define LOV_2DISPATCH(obj, op, ...)                     \
 512         LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
 513
 514 #define LOV_2DISPATCH_VOID(obj, op, ...)                                \
 515 do {                                                                    \
 516         struct lov_object                      *__obj = (obj);          \
 517         enum lov_layout_type                    __llt;                  \
 518                                                                         \
 519         lov_conf_freeze(__obj);                                         \
 520         __llt = __obj->lo_type;                                         \
 521         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 522         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 523         lov_conf_thaw(__obj);                                           \
 524 } while (0)
 525
 526 static void lov_conf_lock(struct lov_object *lov)
 527 {
 528         LASSERT(lov->lo_owner != cfs_current());
 529         down_write(&lov->lo_type_guard);
 530         LASSERT(lov->lo_owner == NULL);
 531         lov->lo_owner = cfs_current();
 532 }
 533
 534 static void lov_conf_unlock(struct lov_object *lov)
 535 {
 536         lov->lo_owner = NULL;
 537         up_write(&lov->lo_type_guard);
 538 }
 539
 540 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov)
 541 {
 542         struct l_wait_info lwi = { 0 };
 543         struct lov_stripe_md *lsm = lov->lo_lsm;
 544         ENTRY;
 545
 546         if (!lov->lo_lsm_invalid || lsm == NULL)
 547                 RETURN(0);
 548
 549         LASSERT(cfs_atomic_read(&lsm->lsm_refc) > 0);
 550         while (cfs_atomic_read(&lsm->lsm_refc) > 1 && lov->lo_lsm_invalid) {
 551                 lov_conf_unlock(lov);
 552
 553                 CDEBUG(D_INODE, "file:"DFID" wait for active IO, now: %d.\n",
 554                         PFID(lu_object_fid(lov2lu(lov))),
 555                         cfs_atomic_read(&lsm->lsm_refc));
 556
 557                 l_wait_event(lov->lo_waitq,
 558                              cfs_atomic_read(&lsm->lsm_refc) == 1, &lwi);
 559                 lov_conf_lock(lov);
 560         }
 561         RETURN(0);
 562 }
 563
 564 static int lov_layout_change(const struct lu_env *unused,
 565                              struct lov_object *lov, enum lov_layout_type llt,
 566                              const struct cl_object_conf *conf)
 567 {
 568         int result;
 569         union lov_layout_state *state = &lov->u;
 570         const struct lov_layout_operations *old_ops;
 571         const struct lov_layout_operations *new_ops;
 572
 573         struct cl_object_header *hdr = cl_object_header(&lov->lo_cl);
 574         void *cookie;
 575         struct lu_env *env;
 576         int refcheck;
 577
 578         LASSERT(0 <= lov->lo_type && lov->lo_type < ARRAY_SIZE(lov_dispatch));
 579         LASSERT(0 <= llt && llt < ARRAY_SIZE(lov_dispatch));
 580         ENTRY;
 581
 582         cookie = cl_env_reenter();
 583         env = cl_env_get(&refcheck);
 584         if (IS_ERR(env)) {
 585                 cl_env_reexit(cookie);
 586                 RETURN(PTR_ERR(env));
 587         }
 588
 589         old_ops = &lov_dispatch[lov->lo_type];
 590         new_ops = &lov_dispatch[llt];
 591
 592         result = old_ops->llo_delete(env, lov, &lov->u);
 593         if (result == 0) {
 594                 old_ops->llo_fini(env, lov, &lov->u);
 595                 LASSERT(cfs_list_empty(&hdr->coh_locks));
 596                 LASSERT(hdr->coh_tree.rnode == NULL);
 597                 LASSERT(hdr->coh_pages == 0);
 598
 599                 lov->lo_type = LLT_EMPTY;
 600                 result = new_ops->llo_init(env,
 601                                         lu2lov_dev(lov->lo_cl.co_lu.lo_dev),
 602                                         lov, conf, state);
 603                 if (result == 0) {
 604                         new_ops->llo_install(env, lov, state);
 605                         lov->lo_type = llt;
 606                 } else {
 607                         new_ops->llo_delete(env, lov, state);
 608                         new_ops->llo_fini(env, lov, state);
 609                         /* this file becomes an EMPTY file. */
 610                 }
 611         }
 612
 613         cl_env_put(env, &refcheck);
 614         cl_env_reexit(cookie);
 615         RETURN(result);
 616 }
 617
 618 /*****************************************************************************
 619  *
 620  * Lov object operations.
 621  *
 622  */
 623
 624 int lov_object_init(const struct lu_env *env, struct lu_object *obj,
 625                     const struct lu_object_conf *conf)
 626 {
 627         struct lov_device            *dev   = lu2lov_dev(obj->lo_dev);
 628         struct lov_object            *lov   = lu2lov(obj);
 629         const struct cl_object_conf  *cconf = lu2cl_conf(conf);
 630         union  lov_layout_state      *set   = &lov->u;
 631         const struct lov_layout_operations *ops;
 632         int result;
 633
 634         ENTRY;
 635         init_rwsem(&lov->lo_type_guard);
 636         cfs_waitq_init(&lov->lo_waitq);
 637
 638         /* no locking is necessary, as object is being created */
 639         lov->lo_type = cconf->u.coc_md->lsm != NULL ? LLT_RAID0 : LLT_EMPTY;
 640         ops = &lov_dispatch[lov->lo_type];
 641         result = ops->llo_init(env, dev, lov, cconf, set);
 642         if (result == 0)
 643                 ops->llo_install(env, lov, set);
 644         RETURN(result);
 645 }
 646
 647 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
 648                         const struct cl_object_conf *conf)
 649 {
 650         struct lov_stripe_md *lsm = NULL;
 651         struct lov_object *lov = cl2lov(obj);
 652         int result = 0;
 653         ENTRY;
 654
 655         lov_conf_lock(lov);
 656         if (conf->coc_opc == OBJECT_CONF_INVALIDATE) {
 657                 lov->lo_lsm_invalid = 1;
 658                 GOTO(out, result = 0);
 659         }
 660
 661         if (conf->coc_opc == OBJECT_CONF_WAIT) {
 662                 result = lov_layout_wait(env, lov);
 663                 GOTO(out, result);
 664         }
 665
 666         LASSERT(conf->coc_opc == OBJECT_CONF_SET);
 667
 668         if (conf->u.coc_md != NULL)
 669                 lsm = conf->u.coc_md->lsm;
 670         if ((lsm == NULL && lov->lo_lsm == NULL) ||
 671             (lsm != NULL && lov->lo_lsm != NULL &&
 672              lov->lo_lsm->lsm_layout_gen == lsm->lsm_layout_gen)) {
 673                 /* same version of layout */
 674                 lov->lo_lsm_invalid = 0;
 675                 GOTO(out, result = 0);
 676         }
 677
 678         /* will change layout - check if there still exists active IO. */
 679         if (lov->lo_lsm != NULL &&
 680             cfs_atomic_read(&lov->lo_lsm->lsm_refc) > 1) {
 681                 lov->lo_lsm_invalid = 1;
 682                 GOTO(out, result = -EBUSY);
 683         }
 684
 685         /*
 686          * Only LLT_EMPTY <-> LLT_RAID0 transitions are supported.
 687          */
 688         switch (lov->lo_type) {
 689         case LLT_EMPTY:
 690                 if (lsm != NULL)
 691                         result = lov_layout_change(env, lov, LLT_RAID0, conf);
 692                 break;
 693         case LLT_RAID0:
 694                 if (lsm == NULL)
 695                         result = lov_layout_change(env, lov, LLT_EMPTY, conf);
 696                 else if (lov_stripe_md_cmp(lov->lo_lsm, lsm))
 697                         result = -EOPNOTSUPP;
 698                 break;
 699         default:
 700                 LBUG();
 701         }
 702         lov->lo_lsm_invalid = result != 0;
 703         EXIT;
 704
 705 out:
 706         lov_conf_unlock(lov);
 707         RETURN(result);
 708 }
 709
 710 static void lov_object_delete(const struct lu_env *env, struct lu_object *obj)
 711 {
 712         struct lov_object *lov = lu2lov(obj);
 713
 714         ENTRY;
 715         LOV_2DISPATCH_VOID(lov, llo_delete, env, lov, &lov->u);
 716         EXIT;
 717 }
 718
 719 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
 720 {
 721         struct lov_object *lov = lu2lov(obj);
 722
 723         ENTRY;
 724         LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
 725         lu_object_fini(obj);
 726         OBD_SLAB_FREE_PTR(lov, lov_object_kmem);
 727         EXIT;
 728 }
 729
 730 static int lov_object_print(const struct lu_env *env, void *cookie,
 731                             lu_printer_t p, const struct lu_object *o)
 732 {
 733         return LOV_2DISPATCH(lu2lov(o), llo_print, env, cookie, p, o);
 734 }
 735
 736 struct cl_page *lov_page_init(const struct lu_env *env, struct cl_object *obj,
 737                               struct cl_page *page, cfs_page_t *vmpage)
 738 {
 739         return LOV_2DISPATCH_NOLOCK(cl2lov(obj),
 740                                     llo_page_init, env, obj, page, vmpage);
 741 }
 742
 743 /**
 744  * Implements cl_object_operations::clo_io_init() method for lov
 745  * layer. Dispatches to the appropriate layout io initialization method.
 746  */
 747 int lov_io_init(const struct lu_env *env, struct cl_object *obj,
 748                 struct cl_io *io)
 749 {
 750         CL_IO_SLICE_CLEAN(lov_env_io(env), lis_cl);
 751         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
 752                                      !io->ci_ignore_layout, env, obj, io);
 753 }
 754
 755 /**
 756  * An implementation of cl_object_operations::clo_attr_get() method for lov
 757  * layer. For raid0 layout this collects and merges attributes of all
 758  * sub-objects.
 759  */
 760 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
 761                         struct cl_attr *attr)
 762 {
 763         /* do not take lock, as this function is called under a
 764          * spin-lock. Layout is protected from changing by ongoing IO. */
 765         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
 766 }
 767
 768 static int lov_attr_set(const struct lu_env *env, struct cl_object *obj,
 769                         const struct cl_attr *attr, unsigned valid)
 770 {
 771         /*
 772          * No dispatch is required here, as no layout implements this.
 773          */
 774         return 0;
 775 }
 776
 777 int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
 778                   struct cl_lock *lock, const struct cl_io *io)
 779 {
 780         /* No need to lock because we've taken one refcount of layout.  */
 781         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_lock_init, env, obj, lock,
 782                                     io);
 783 }
 784
 785 static const struct cl_object_operations lov_ops = {
 786         .coo_page_init = lov_page_init,
 787         .coo_lock_init = lov_lock_init,
 788         .coo_io_init   = lov_io_init,
 789         .coo_attr_get  = lov_attr_get,
 790         .coo_attr_set  = lov_attr_set,
 791         .coo_conf_set  = lov_conf_set
 792 };
 793
 794 static const struct lu_object_operations lov_lu_obj_ops = {
 795         .loo_object_init      = lov_object_init,
 796         .loo_object_delete    = lov_object_delete,
 797         .loo_object_release   = NULL,
 798         .loo_object_free      = lov_object_free,
 799         .loo_object_print     = lov_object_print,
 800         .loo_object_invariant = NULL
 801 };
 802
 803 struct lu_object *lov_object_alloc(const struct lu_env *env,
 804                                    const struct lu_object_header *unused,
 805                                    struct lu_device *dev)
 806 {
 807         struct lov_object *lov;
 808         struct lu_object  *obj;
 809
 810         ENTRY;
 811         OBD_SLAB_ALLOC_PTR_GFP(lov, lov_object_kmem, CFS_ALLOC_IO);
 812         if (lov != NULL) {
 813                 obj = lov2lu(lov);
 814                 lu_object_init(obj, NULL, dev);
 815                 lov->lo_cl.co_ops = &lov_ops;
 816                 lov->lo_type = -1; /* invalid, to catch uninitialized type */
 817                 /*
 818                  * object io operation vector (cl_object::co_iop) is installed
 819                  * later in lov_object_init(), as different vectors are used
 820                  * for object with different layouts.
 821                  */
 822                 obj->lo_ops = &lov_lu_obj_ops;
 823         } else
 824                 obj = NULL;
 825         RETURN(obj);
 826 }
 827
 828 struct lov_stripe_md *lov_lsm_addref(struct lov_object *lov)
 829 {
 830         struct lov_stripe_md *lsm = NULL;
 831
 832         lov_conf_freeze(lov);
 833         if (lov->lo_lsm != NULL) {
 834                 lsm = lsm_addref(lov->lo_lsm);
 835                 CDEBUG(D_INODE, "lsm %p addref %d/%d by %p.\n",
 836                         lsm, cfs_atomic_read(&lsm->lsm_refc),
 837                         lov->lo_lsm_invalid, cfs_current());
 838         }
 839         lov_conf_thaw(lov);
 840         return lsm;
 841 }
 842
 843 void lov_lsm_decref(struct lov_object *lov, struct lov_stripe_md *lsm)
 844 {
 845         if (lsm == NULL)
 846                 return;
 847
 848         CDEBUG(D_INODE, "lsm %p decref %d by %p.\n",
 849                 lsm, cfs_atomic_read(&lsm->lsm_refc), cfs_current());
 850
 851         if (lov_free_memmd(&lsm) <= 1 && lov->lo_lsm_invalid)
 852                 cfs_waitq_signal(&lov->lo_waitq);
 853 }
 854
 855 struct lov_stripe_md *lov_lsm_get(struct cl_object *clobj)
 856 {
 857         struct lu_object *luobj;
 858         struct lov_stripe_md *lsm = NULL;
 859
 860         if (clobj == NULL)
 861                 return NULL;
 862
 863         luobj = lu_object_locate(&cl_object_header(clobj)->coh_lu,
 864                                  &lov_device_type);
 865         if (luobj != NULL)
 866                 lsm = lov_lsm_addref(lu2lov(luobj));
 867         return lsm;
 868 }
 869 EXPORT_SYMBOL(lov_lsm_get);
 870
 871 void lov_lsm_put(struct cl_object *clobj, struct lov_stripe_md *lsm)
 872 {
 873         struct lu_object *luobj;
 874
 875         if (clobj == NULL || lsm == NULL)
 876                 return;
 877
 878         luobj = lu_object_locate(&cl_object_header(clobj)->coh_lu,
 879                                  &lov_device_type);
 880         LASSERT(luobj != NULL);
 881
 882         lov_lsm_decref(lu2lov(luobj), lsm);
 883 }
 884 EXPORT_SYMBOL(lov_lsm_put);
 885
 886 int lov_read_and_clear_async_rc(struct cl_object *clob)
 887 {
 888         struct lu_object *luobj;
 889         int rc = 0;
 890         ENTRY;
 891
 892         luobj = lu_object_locate(&cl_object_header(clob)->coh_lu,
 893                                  &lov_device_type);
 894         if (luobj != NULL) {
 895                 struct lov_object *lov = lu2lov(luobj);
 896
 897                 lov_conf_freeze(lov);
 898                 switch (lov->lo_type) {
 899                 case LLT_RAID0: {
 900                         struct lov_stripe_md *lsm;
 901                         int i;
 902
 903                         lsm = lov->lo_lsm;
 904                         LASSERT(lsm != NULL);
 905                         for (i = 0; i < lsm->lsm_stripe_count; i++) {
 906                                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
 907                                 if (loi->loi_ar.ar_rc && !rc)
 908                                         rc = loi->loi_ar.ar_rc;
 909                                 loi->loi_ar.ar_rc = 0;
 910                         }
 911                 }
 912                 case LLT_EMPTY:
 913                         break;
 914                 default:
 915                         LBUG();
 916                 }
 917                 lov_conf_thaw(lov);
 918         }
 919         RETURN(rc);
 920 }
 921 EXPORT_SYMBOL(lov_read_and_clear_async_rc);
 922
 923 /** @} lov */