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         int  (*llo_page_init)(const struct lu_env *env, struct cl_object *obj,
  71                                 struct cl_page *page, cfs_page_t *vmpage);
  72         int  (*llo_lock_init)(const struct lu_env *env,
  73                               struct cl_object *obj, struct cl_lock *lock,
  74                               const struct cl_io *io);
  75         int  (*llo_io_init)(const struct lu_env *env,
  76                             struct cl_object *obj, struct cl_io *io);
  77         int  (*llo_getattr)(const struct lu_env *env, struct cl_object *obj,
  78                             struct cl_attr *attr);
  79 };
  80
  81 /*****************************************************************************
  82  *
  83  * Lov object layout operations.
  84  *
  85  */
  86
  87 static void lov_install_empty(const struct lu_env *env,
  88                               struct lov_object *lov,
  89                               union  lov_layout_state *state)
  90 {
  91         /*
  92          * File without objects.
  93          */
  94 }
  95
  96 static int lov_init_empty(const struct lu_env *env,
  97                           struct lov_device *dev, struct lov_object *lov,
  98                           const struct cl_object_conf *conf,
  99                           union  lov_layout_state *state)
 100 {
 101         return 0;
 102 }
 103
 104 static void lov_install_raid0(const struct lu_env *env,
 105                               struct lov_object *lov,
 106                               union  lov_layout_state *state)
 107 {
 108 }
 109
 110 static struct cl_object *lov_sub_find(const struct lu_env *env,
 111                                       struct cl_device *dev,
 112                                       const struct lu_fid *fid,
 113                                       const struct cl_object_conf *conf)
 114 {
 115         struct lu_object *o;
 116
 117         ENTRY;
 118         o = lu_object_find_at(env, cl2lu_dev(dev), fid, &conf->coc_lu);
 119         LASSERT(ergo(!IS_ERR(o), o->lo_dev->ld_type == &lovsub_device_type));
 120         RETURN(lu2cl(o));
 121 }
 122
 123 static int lov_init_sub(const struct lu_env *env, struct lov_object *lov,
 124                         struct cl_object *stripe,
 125                         struct lov_layout_raid0 *r0, int idx)
 126 {
 127         struct cl_object_header *hdr;
 128         struct cl_object_header *subhdr;
 129         struct cl_object_header *parent;
 130         struct lov_oinfo        *oinfo;
 131         int result;
 132
 133         if (OBD_FAIL_CHECK(OBD_FAIL_LOV_INIT)) {
 134                 /* For sanity:test_206.
 135                  * Do not leave the object in cache to avoid accessing
 136                  * freed memory. This is because osc_object is referring to
 137                  * lov_oinfo of lsm_stripe_data which will be freed due to
 138                  * this failure. */
 139                 cl_object_kill(env, stripe);
 140                 cl_object_put(env, stripe);
 141                 return -EIO;
 142         }
 143
 144         hdr    = cl_object_header(lov2cl(lov));
 145         subhdr = cl_object_header(stripe);
 146         parent = subhdr->coh_parent;
 147
 148         oinfo = lov->lo_lsm->lsm_oinfo[idx];
 149         CDEBUG(D_INODE, DFID"@%p[%d] -> "DFID"@%p: id: "LPU64" seq: "LPU64
 150                " idx: %d gen: %d\n",
 151                PFID(&subhdr->coh_lu.loh_fid), subhdr, idx,
 152                PFID(&hdr->coh_lu.loh_fid), hdr,
 153                oinfo->loi_id, oinfo->loi_seq,
 154                oinfo->loi_ost_idx, oinfo->loi_ost_gen);
 155
 156         if (parent == NULL) {
 157                 subhdr->coh_parent = hdr;
 158                 subhdr->coh_nesting = hdr->coh_nesting + 1;
 159                 lu_object_ref_add(&stripe->co_lu, "lov-parent", lov);
 160                 r0->lo_sub[idx] = cl2lovsub(stripe);
 161                 r0->lo_sub[idx]->lso_super = lov;
 162                 r0->lo_sub[idx]->lso_index = idx;
 163                 result = 0;
 164         } else {
 165                 CERROR("Stripe is already owned by other file (%d).\n", idx);
 166                 LU_OBJECT_DEBUG(D_ERROR, env, &stripe->co_lu, "\n");
 167                 LU_OBJECT_DEBUG(D_ERROR, env, lu_object_top(&parent->coh_lu),
 168                                 "old\n");
 169                 LU_OBJECT_HEADER(D_ERROR, env, lov2lu(lov), "new\n");
 170                 cl_object_put(env, stripe);
 171                 result = -EIO;
 172         }
 173         return result;
 174 }
 175
 176 static int lov_init_raid0(const struct lu_env *env,
 177                           struct lov_device *dev, struct lov_object *lov,
 178                           const struct cl_object_conf *conf,
 179                           union  lov_layout_state *state)
 180 {
 181         int result;
 182         int i;
 183
 184         struct cl_object        *stripe;
 185         struct lov_thread_info  *lti     = lov_env_info(env);
 186         struct cl_object_conf   *subconf = &lti->lti_stripe_conf;
 187         struct lov_stripe_md    *lsm     = conf->u.coc_md->lsm;
 188         struct lu_fid           *ofid    = &lti->lti_fid;
 189         struct lov_layout_raid0 *r0      = &state->raid0;
 190
 191         ENTRY;
 192
 193         if (lsm->lsm_magic != LOV_MAGIC_V1 && lsm->lsm_magic != LOV_MAGIC_V3) {
 194                 dump_lsm(D_ERROR, lsm);
 195                 LASSERTF(0, "magic mismatch, expected %d/%d, actual %d.\n",
 196                          LOV_MAGIC_V1, LOV_MAGIC_V3, lsm->lsm_magic);
 197         }
 198
 199         LASSERT(lov->lo_lsm == NULL);
 200         lov->lo_lsm = lsm_addref(lsm);
 201         r0->lo_nr  = lsm->lsm_stripe_count;
 202         LASSERT(r0->lo_nr <= lov_targets_nr(dev));
 203
 204         OBD_ALLOC_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 205         if (r0->lo_sub != NULL) {
 206                 result = 0;
 207                 subconf->coc_inode = conf->coc_inode;
 208                 spin_lock_init(&r0->lo_sub_lock);
 209                 /*
 210                  * Create stripe cl_objects.
 211                  */
 212                 for (i = 0; i < r0->lo_nr && result == 0; ++i) {
 213                         struct cl_device *subdev;
 214                         struct lov_oinfo *oinfo = lsm->lsm_oinfo[i];
 215                         int ost_idx = oinfo->loi_ost_idx;
 216
 217                         fid_ostid_unpack(ofid, &oinfo->loi_oi,
 218                                          oinfo->loi_ost_idx);
 219                         subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
 220                         subconf->u.coc_oinfo = oinfo;
 221                         LASSERTF(subdev != NULL, "not init ost %d\n", ost_idx);
 222                         /* In the function below, .hs_keycmp resolves to
 223                          * lu_obj_hop_keycmp() */
 224                         /* coverity[overrun-buffer-val] */
 225                         stripe = lov_sub_find(env, subdev, ofid, subconf);
 226                         if (!IS_ERR(stripe))
 227                                 result = lov_init_sub(env, lov, stripe, r0, i);
 228                         else
 229                                 result = PTR_ERR(stripe);
 230                 }
 231         } else
 232                 result = -ENOMEM;
 233         RETURN(result);
 234 }
 235
 236 static int lov_delete_empty(const struct lu_env *env, struct lov_object *lov,
 237                             union lov_layout_state *state)
 238 {
 239         LASSERT(lov->lo_type == LLT_EMPTY);
 240         if (cfs_atomic_read(&lov->lo_active_ios) > 0)
 241                 RETURN(-EBUSY);
 242
 243         cl_object_prune(env, &lov->lo_cl);
 244         return 0;
 245 }
 246
 247 static void lov_subobject_kill(const struct lu_env *env, struct lov_object *lov,
 248                                struct lovsub_object *los, int idx)
 249 {
 250         struct cl_object        *sub;
 251         struct lov_layout_raid0 *r0;
 252         struct lu_site          *site;
 253         struct lu_site_bkt_data *bkt;
 254         cfs_waitlink_t          *waiter;
 255
 256         r0  = &lov->u.raid0;
 257         LASSERT(r0->lo_sub[idx] == los);
 258
 259         sub  = lovsub2cl(los);
 260         site = sub->co_lu.lo_dev->ld_site;
 261         bkt  = lu_site_bkt_from_fid(site, &sub->co_lu.lo_header->loh_fid);
 262
 263         cl_object_kill(env, sub);
 264         /* release a reference to the sub-object and ... */
 265         lu_object_ref_del(&sub->co_lu, "lov-parent", lov);
 266         cl_object_put(env, sub);
 267
 268         /* ... wait until it is actually destroyed---sub-object clears its
 269          * ->lo_sub[] slot in lovsub_object_fini() */
 270         if (r0->lo_sub[idx] == los) {
 271                 waiter = &lov_env_info(env)->lti_waiter;
 272                 cfs_waitlink_init(waiter);
 273                 cfs_waitq_add(&bkt->lsb_marche_funebre, waiter);
 274                 cfs_set_current_state(CFS_TASK_UNINT);
 275                 while (1) {
 276                         /* this wait-queue is signaled at the end of
 277                          * lu_object_free(). */
 278                         cfs_set_current_state(CFS_TASK_UNINT);
 279                         spin_lock(&r0->lo_sub_lock);
 280                         if (r0->lo_sub[idx] == los) {
 281                                 spin_unlock(&r0->lo_sub_lock);
 282                                 cfs_waitq_wait(waiter, CFS_TASK_UNINT);
 283                         } else {
 284                                 spin_unlock(&r0->lo_sub_lock);
 285                                 cfs_set_current_state(CFS_TASK_RUNNING);
 286                                 break;
 287                         }
 288                 }
 289                 cfs_waitq_del(&bkt->lsb_marche_funebre, waiter);
 290         }
 291         LASSERT(r0->lo_sub[idx] == NULL);
 292 }
 293
 294 static int lov_delete_raid0(const struct lu_env *env, struct lov_object *lov,
 295                             union lov_layout_state *state)
 296 {
 297         struct lov_layout_raid0 *r0 = &state->raid0;
 298         struct lov_stripe_md    *lsm = lov->lo_lsm;
 299         int i;
 300
 301         ENTRY;
 302
 303         dump_lsm(D_INODE, lsm);
 304         if (cfs_atomic_read(&lov->lo_active_ios) > 0)
 305                 RETURN(-EBUSY);
 306
 307         if (r0->lo_sub != NULL) {
 308                 for (i = 0; i < r0->lo_nr; ++i) {
 309                         struct lovsub_object *los = r0->lo_sub[i];
 310
 311                         if (los != NULL) {
 312                                 cl_locks_prune(env, &los->lso_cl, 1);
 313                                 /*
 314                                  * If top-level object is to be evicted from
 315                                  * the cache, so are its sub-objects.
 316                                  */
 317                                 lov_subobject_kill(env, lov, los, i);
 318                         }
 319                 }
 320         }
 321         cl_object_prune(env, &lov->lo_cl);
 322         RETURN(0);
 323 }
 324
 325 static void lov_fini_empty(const struct lu_env *env, struct lov_object *lov,
 326                            union lov_layout_state *state)
 327 {
 328         LASSERT(lov->lo_type == LLT_EMPTY);
 329 }
 330
 331 static void lov_fini_raid0(const struct lu_env *env, struct lov_object *lov,
 332                            union lov_layout_state *state)
 333 {
 334         struct lov_layout_raid0 *r0 = &state->raid0;
 335         ENTRY;
 336
 337         if (r0->lo_sub != NULL) {
 338                 OBD_FREE_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 339                 r0->lo_sub = NULL;
 340         }
 341
 342         dump_lsm(D_INODE, lov->lo_lsm);
 343         lov_free_memmd(&lov->lo_lsm);
 344
 345         EXIT;
 346 }
 347
 348 static int lov_print_empty(const struct lu_env *env, void *cookie,
 349                            lu_printer_t p, const struct lu_object *o)
 350 {
 351         (*p)(env, cookie, "empty\n");
 352         return 0;
 353 }
 354
 355 static int lov_print_raid0(const struct lu_env *env, void *cookie,
 356                            lu_printer_t p, const struct lu_object *o)
 357 {
 358         struct lov_object       *lov = lu2lov(o);
 359         struct lov_layout_raid0 *r0  = lov_r0(lov);
 360         int i;
 361
 362         (*p)(env, cookie, "stripes: %d:\n", r0->lo_nr);
 363         for (i = 0; i < r0->lo_nr; ++i) {
 364                 struct lu_object *sub;
 365
 366                 if (r0->lo_sub[i] != NULL) {
 367                         sub = lovsub2lu(r0->lo_sub[i]);
 368                         lu_object_print(env, cookie, p, sub);
 369                 } else
 370                         (*p)(env, cookie, "sub %d absent\n", i);
 371         }
 372         return 0;
 373 }
 374
 375 /**
 376  * Implements cl_object_operations::coo_attr_get() method for an object
 377  * without stripes (LLT_EMPTY layout type).
 378  *
 379  * The only attributes this layer is authoritative in this case is
 380  * cl_attr::cat_blocks---it's 0.
 381  */
 382 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
 383                               struct cl_attr *attr)
 384 {
 385         attr->cat_blocks = 0;
 386         return 0;
 387 }
 388
 389 static int lov_attr_get_raid0(const struct lu_env *env, struct cl_object *obj,
 390                               struct cl_attr *attr)
 391 {
 392         struct lov_object       *lov = cl2lov(obj);
 393         struct lov_layout_raid0 *r0 = lov_r0(lov);
 394         struct cl_attr          *lov_attr = &r0->lo_attr;
 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                 struct lov_stripe_md    *lsm = lov->lo_lsm;
 411                 struct ost_lvb          *lvb = &lov_env_info(env)->lti_lvb;
 412                 __u64                    kms = 0;
 413
 414                 memset(lvb, 0, sizeof(*lvb));
 415                 /* XXX: timestamps can be negative by sanity:test_39m,
 416                  * how can it be? */
 417                 lvb->lvb_atime = LLONG_MIN;
 418                 lvb->lvb_ctime = LLONG_MIN;
 419                 lvb->lvb_mtime = LLONG_MIN;
 420
 421                 /*
 422                  * XXX that should be replaced with a loop over sub-objects,
 423                  * doing cl_object_attr_get() on them. But for now, let's
 424                  * reuse old lov code.
 425                  */
 426
 427                 /*
 428                  * XXX take lsm spin-lock to keep lov_merge_lvb_kms()
 429                  * happy. It's not needed, because new code uses
 430                  * ->coh_attr_guard spin-lock to protect consistency of
 431                  * sub-object attributes.
 432                  */
 433                 lov_stripe_lock(lsm);
 434                 result = lov_merge_lvb_kms(lsm, lvb, &kms);
 435                 lov_stripe_unlock(lsm);
 436                 if (result == 0) {
 437                         cl_lvb2attr(lov_attr, lvb);
 438                         lov_attr->cat_kms = kms;
 439                         r0->lo_attr_valid = 1;
 440                 }
 441         }
 442         if (result == 0) { /* merge results */
 443                 attr->cat_blocks = lov_attr->cat_blocks;
 444                 attr->cat_size = lov_attr->cat_size;
 445                 attr->cat_kms = lov_attr->cat_kms;
 446                 if (attr->cat_atime < lov_attr->cat_atime)
 447                         attr->cat_atime = lov_attr->cat_atime;
 448                 if (attr->cat_ctime < lov_attr->cat_ctime)
 449                         attr->cat_ctime = lov_attr->cat_ctime;
 450                 if (attr->cat_mtime < lov_attr->cat_mtime)
 451                         attr->cat_mtime = lov_attr->cat_mtime;
 452         }
 453         RETURN(result);
 454 }
 455
 456 const static struct lov_layout_operations lov_dispatch[] = {
 457         [LLT_EMPTY] = {
 458                 .llo_init      = lov_init_empty,
 459                 .llo_delete    = lov_delete_empty,
 460                 .llo_fini      = lov_fini_empty,
 461                 .llo_install   = lov_install_empty,
 462                 .llo_print     = lov_print_empty,
 463                 .llo_page_init = lov_page_init_empty,
 464                 .llo_lock_init = lov_lock_init_empty,
 465                 .llo_io_init   = lov_io_init_empty,
 466                 .llo_getattr   = lov_attr_get_empty
 467         },
 468         [LLT_RAID0] = {
 469                 .llo_init      = lov_init_raid0,
 470                 .llo_delete    = lov_delete_raid0,
 471                 .llo_fini      = lov_fini_raid0,
 472                 .llo_install   = lov_install_raid0,
 473                 .llo_print     = lov_print_raid0,
 474                 .llo_page_init = lov_page_init_raid0,
 475                 .llo_lock_init = lov_lock_init_raid0,
 476                 .llo_io_init   = lov_io_init_raid0,
 477                 .llo_getattr   = lov_attr_get_raid0
 478         }
 479 };
 480
 481
 482 /**
 483  * Performs a double-dispatch based on the layout type of an object.
 484  */
 485 #define LOV_2DISPATCH_NOLOCK(obj, op, ...)                              \
 486 ({                                                                      \
 487         struct lov_object                      *__obj = (obj);          \
 488         enum lov_layout_type                    __llt;                  \
 489                                                                         \
 490         __llt = __obj->lo_type;                                         \
 491         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 492         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 493 })
 494
 495 static inline void lov_conf_freeze(struct lov_object *lov)
 496 {
 497         if (lov->lo_owner != cfs_current())
 498                 down_read(&lov->lo_type_guard);
 499 }
 500
 501 static inline void lov_conf_thaw(struct lov_object *lov)
 502 {
 503         if (lov->lo_owner != cfs_current())
 504                 up_read(&lov->lo_type_guard);
 505 }
 506
 507 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...)                       \
 508 ({                                                                      \
 509         struct lov_object                      *__obj = (obj);          \
 510         int                                     __lock = !!(lock);      \
 511         typeof(lov_dispatch[0].op(__VA_ARGS__)) __result;               \
 512                                                                         \
 513         if (__lock)                                                     \
 514                 lov_conf_freeze(__obj);                                 \
 515         __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__);          \
 516         if (__lock)                                                     \
 517                 lov_conf_thaw(__obj);                                   \
 518         __result;                                                       \
 519 })
 520
 521 /**
 522  * Performs a locked double-dispatch based on the layout type of an object.
 523  */
 524 #define LOV_2DISPATCH(obj, op, ...)                     \
 525         LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
 526
 527 #define LOV_2DISPATCH_VOID(obj, op, ...)                                \
 528 do {                                                                    \
 529         struct lov_object                      *__obj = (obj);          \
 530         enum lov_layout_type                    __llt;                  \
 531                                                                         \
 532         lov_conf_freeze(__obj);                                         \
 533         __llt = __obj->lo_type;                                         \
 534         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 535         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 536         lov_conf_thaw(__obj);                                           \
 537 } while (0)
 538
 539 static void lov_conf_lock(struct lov_object *lov)
 540 {
 541         LASSERT(lov->lo_owner != cfs_current());
 542         down_write(&lov->lo_type_guard);
 543         LASSERT(lov->lo_owner == NULL);
 544         lov->lo_owner = cfs_current();
 545 }
 546
 547 static void lov_conf_unlock(struct lov_object *lov)
 548 {
 549         lov->lo_owner = NULL;
 550         up_write(&lov->lo_type_guard);
 551 }
 552
 553 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov)
 554 {
 555         struct l_wait_info lwi = { 0 };
 556         ENTRY;
 557
 558         if (!lov->lo_layout_invalid)
 559                 RETURN(0);
 560
 561         while (cfs_atomic_read(&lov->lo_active_ios) > 0) {
 562                 lov_conf_unlock(lov);
 563
 564                 CDEBUG(D_INODE, "file:"DFID" wait for active IO, now: %d.\n",
 565                         PFID(lu_object_fid(lov2lu(lov))),
 566                         cfs_atomic_read(&lov->lo_active_ios));
 567
 568                 l_wait_event(lov->lo_waitq,
 569                              cfs_atomic_read(&lov->lo_active_ios) == 0, &lwi);
 570                 lov_conf_lock(lov);
 571         }
 572         RETURN(0);
 573 }
 574
 575 static int lov_layout_change(const struct lu_env *unused,
 576                              struct lov_object *lov, enum lov_layout_type llt,
 577                              const struct cl_object_conf *conf)
 578 {
 579         int result;
 580         union lov_layout_state *state = &lov->u;
 581         const struct lov_layout_operations *old_ops;
 582         const struct lov_layout_operations *new_ops;
 583
 584         struct cl_object_header *hdr = cl_object_header(&lov->lo_cl);
 585         void *cookie;
 586         struct lu_env *env;
 587         int refcheck;
 588
 589         LASSERT(0 <= lov->lo_type && lov->lo_type < ARRAY_SIZE(lov_dispatch));
 590         LASSERT(0 <= llt && llt < ARRAY_SIZE(lov_dispatch));
 591         ENTRY;
 592
 593         cookie = cl_env_reenter();
 594         env = cl_env_get(&refcheck);
 595         if (IS_ERR(env)) {
 596                 cl_env_reexit(cookie);
 597                 RETURN(PTR_ERR(env));
 598         }
 599
 600         old_ops = &lov_dispatch[lov->lo_type];
 601         new_ops = &lov_dispatch[llt];
 602
 603         result = old_ops->llo_delete(env, lov, &lov->u);
 604         if (result == 0) {
 605                 old_ops->llo_fini(env, lov, &lov->u);
 606
 607                 LASSERT(cfs_atomic_read(&lov->lo_active_ios) == 0);
 608                 LASSERT(cfs_list_empty(&hdr->coh_locks));
 609                 LASSERT(hdr->coh_tree.rnode == NULL);
 610                 LASSERT(hdr->coh_pages == 0);
 611
 612                 lov->lo_type = LLT_EMPTY;
 613                 result = new_ops->llo_init(env,
 614                                         lu2lov_dev(lov->lo_cl.co_lu.lo_dev),
 615                                         lov, conf, state);
 616                 if (result == 0) {
 617                         new_ops->llo_install(env, lov, state);
 618                         lov->lo_type = llt;
 619                 } else {
 620                         new_ops->llo_delete(env, lov, state);
 621                         new_ops->llo_fini(env, lov, state);
 622                         /* this file becomes an EMPTY file. */
 623                 }
 624         }
 625
 626         cl_env_put(env, &refcheck);
 627         cl_env_reexit(cookie);
 628         RETURN(result);
 629 }
 630
 631 /*****************************************************************************
 632  *
 633  * Lov object operations.
 634  *
 635  */
 636
 637 int lov_object_init(const struct lu_env *env, struct lu_object *obj,
 638                     const struct lu_object_conf *conf)
 639 {
 640         struct lov_device            *dev   = lu2lov_dev(obj->lo_dev);
 641         struct lov_object            *lov   = lu2lov(obj);
 642         const struct cl_object_conf  *cconf = lu2cl_conf(conf);
 643         union  lov_layout_state      *set   = &lov->u;
 644         const struct lov_layout_operations *ops;
 645         int result;
 646
 647         ENTRY;
 648         init_rwsem(&lov->lo_type_guard);
 649         cfs_atomic_set(&lov->lo_active_ios, 0);
 650         cfs_waitq_init(&lov->lo_waitq);
 651
 652         cl_object_page_init(lu2cl(obj), sizeof(struct lov_page));
 653
 654         /* no locking is necessary, as object is being created */
 655         lov->lo_type = cconf->u.coc_md->lsm != NULL ? LLT_RAID0 : LLT_EMPTY;
 656         ops = &lov_dispatch[lov->lo_type];
 657         result = ops->llo_init(env, dev, lov, cconf, set);
 658         if (result == 0)
 659                 ops->llo_install(env, lov, set);
 660         RETURN(result);
 661 }
 662
 663 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
 664                         const struct cl_object_conf *conf)
 665 {
 666         struct lov_stripe_md *lsm = NULL;
 667         struct lov_object *lov = cl2lov(obj);
 668         int result = 0;
 669         ENTRY;
 670
 671         lov_conf_lock(lov);
 672         if (conf->coc_opc == OBJECT_CONF_INVALIDATE) {
 673                 lov->lo_layout_invalid = true;
 674                 GOTO(out, result = 0);
 675         }
 676
 677         if (conf->coc_opc == OBJECT_CONF_WAIT) {
 678                 result = lov_layout_wait(env, lov);
 679                 GOTO(out, result);
 680         }
 681
 682         LASSERT(conf->coc_opc == OBJECT_CONF_SET);
 683
 684         if (conf->u.coc_md != NULL)
 685                 lsm = conf->u.coc_md->lsm;
 686         if ((lsm == NULL && lov->lo_lsm == NULL) ||
 687             (lsm != NULL && lov->lo_lsm != NULL &&
 688              lov->lo_lsm->lsm_layout_gen == lsm->lsm_layout_gen)) {
 689                 /* same version of layout */
 690                 lov->lo_layout_invalid = false;
 691                 GOTO(out, result = 0);
 692         }
 693
 694         /* will change layout - check if there still exists active IO. */
 695         if (cfs_atomic_read(&lov->lo_active_ios) > 1) {
 696                 lov->lo_layout_invalid = true;
 697                 GOTO(out, result = -EBUSY);
 698         }
 699
 700         /*
 701          * Only LLT_EMPTY <-> LLT_RAID0 transitions are supported.
 702          */
 703         switch (lov->lo_type) {
 704         case LLT_EMPTY:
 705                 if (lsm != NULL)
 706                         result = lov_layout_change(env, lov, LLT_RAID0, conf);
 707                 break;
 708         case LLT_RAID0:
 709                 if (lsm == NULL)
 710                         result = lov_layout_change(env, lov, LLT_EMPTY, conf);
 711                 else if (lov_stripe_md_cmp(lov->lo_lsm, lsm))
 712                         result = -EOPNOTSUPP;
 713                 break;
 714         default:
 715                 LBUG();
 716         }
 717         lov->lo_layout_invalid = result != 0;
 718         EXIT;
 719
 720 out:
 721         lov_conf_unlock(lov);
 722         RETURN(result);
 723 }
 724
 725 static void lov_object_delete(const struct lu_env *env, struct lu_object *obj)
 726 {
 727         struct lov_object *lov = lu2lov(obj);
 728
 729         ENTRY;
 730         LOV_2DISPATCH_VOID(lov, llo_delete, env, lov, &lov->u);
 731         EXIT;
 732 }
 733
 734 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
 735 {
 736         struct lov_object *lov = lu2lov(obj);
 737
 738         ENTRY;
 739         LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
 740         lu_object_fini(obj);
 741         OBD_SLAB_FREE_PTR(lov, lov_object_kmem);
 742         EXIT;
 743 }
 744
 745 static int lov_object_print(const struct lu_env *env, void *cookie,
 746                             lu_printer_t p, const struct lu_object *o)
 747 {
 748         return LOV_2DISPATCH(lu2lov(o), llo_print, env, cookie, p, o);
 749 }
 750
 751 int lov_page_init(const struct lu_env *env, struct cl_object *obj,
 752                 struct cl_page *page, cfs_page_t *vmpage)
 753 {
 754         return LOV_2DISPATCH_NOLOCK(cl2lov(obj),
 755                                     llo_page_init, env, obj, page, vmpage);
 756 }
 757
 758 /**
 759  * Implements cl_object_operations::clo_io_init() method for lov
 760  * layer. Dispatches to the appropriate layout io initialization method.
 761  */
 762 int lov_io_init(const struct lu_env *env, struct cl_object *obj,
 763                 struct cl_io *io)
 764 {
 765         CL_IO_SLICE_CLEAN(lov_env_io(env), lis_cl);
 766         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
 767                                      !io->ci_ignore_layout, env, obj, io);
 768 }
 769
 770 /**
 771  * An implementation of cl_object_operations::clo_attr_get() method for lov
 772  * layer. For raid0 layout this collects and merges attributes of all
 773  * sub-objects.
 774  */
 775 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
 776                         struct cl_attr *attr)
 777 {
 778         /* do not take lock, as this function is called under a
 779          * spin-lock. Layout is protected from changing by ongoing IO. */
 780         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
 781 }
 782
 783 static int lov_attr_set(const struct lu_env *env, struct cl_object *obj,
 784                         const struct cl_attr *attr, unsigned valid)
 785 {
 786         /*
 787          * No dispatch is required here, as no layout implements this.
 788          */
 789         return 0;
 790 }
 791
 792 int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
 793                   struct cl_lock *lock, const struct cl_io *io)
 794 {
 795         /* No need to lock because we've taken one refcount of layout.  */
 796         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_lock_init, env, obj, lock,
 797                                     io);
 798 }
 799
 800 static const struct cl_object_operations lov_ops = {
 801         .coo_page_init = lov_page_init,
 802         .coo_lock_init = lov_lock_init,
 803         .coo_io_init   = lov_io_init,
 804         .coo_attr_get  = lov_attr_get,
 805         .coo_attr_set  = lov_attr_set,
 806         .coo_conf_set  = lov_conf_set
 807 };
 808
 809 static const struct lu_object_operations lov_lu_obj_ops = {
 810         .loo_object_init      = lov_object_init,
 811         .loo_object_delete    = lov_object_delete,
 812         .loo_object_release   = NULL,
 813         .loo_object_free      = lov_object_free,
 814         .loo_object_print     = lov_object_print,
 815         .loo_object_invariant = NULL
 816 };
 817
 818 struct lu_object *lov_object_alloc(const struct lu_env *env,
 819                                    const struct lu_object_header *unused,
 820                                    struct lu_device *dev)
 821 {
 822         struct lov_object *lov;
 823         struct lu_object  *obj;
 824
 825         ENTRY;
 826         OBD_SLAB_ALLOC_PTR_GFP(lov, lov_object_kmem, CFS_ALLOC_IO);
 827         if (lov != NULL) {
 828                 obj = lov2lu(lov);
 829                 lu_object_init(obj, NULL, dev);
 830                 lov->lo_cl.co_ops = &lov_ops;
 831                 lov->lo_type = -1; /* invalid, to catch uninitialized type */
 832                 /*
 833                  * object io operation vector (cl_object::co_iop) is installed
 834                  * later in lov_object_init(), as different vectors are used
 835                  * for object with different layouts.
 836                  */
 837                 obj->lo_ops = &lov_lu_obj_ops;
 838         } else
 839                 obj = NULL;
 840         RETURN(obj);
 841 }
 842
 843 struct lov_stripe_md *lov_lsm_addref(struct lov_object *lov)
 844 {
 845         struct lov_stripe_md *lsm = NULL;
 846
 847         lov_conf_freeze(lov);
 848         if (lov->lo_lsm != NULL) {
 849                 lsm = lsm_addref(lov->lo_lsm);
 850                 CDEBUG(D_INODE, "lsm %p addref %d/%d by %p.\n",
 851                         lsm, cfs_atomic_read(&lsm->lsm_refc),
 852                         lov->lo_layout_invalid, cfs_current());
 853         }
 854         lov_conf_thaw(lov);
 855         return lsm;
 856 }
 857
 858 void lov_lsm_decref(struct lov_object *lov, struct lov_stripe_md *lsm)
 859 {
 860         if (lsm == NULL)
 861                 return;
 862
 863         CDEBUG(D_INODE, "lsm %p decref %d by %p.\n",
 864                 lsm, cfs_atomic_read(&lsm->lsm_refc), cfs_current());
 865
 866         lov_free_memmd(&lsm);
 867 }
 868
 869 struct lov_stripe_md *lov_lsm_get(struct cl_object *clobj)
 870 {
 871         struct lu_object *luobj;
 872         struct lov_stripe_md *lsm = NULL;
 873
 874         if (clobj == NULL)
 875                 return NULL;
 876
 877         luobj = lu_object_locate(&cl_object_header(clobj)->coh_lu,
 878                                  &lov_device_type);
 879         if (luobj != NULL)
 880                 lsm = lov_lsm_addref(lu2lov(luobj));
 881         return lsm;
 882 }
 883 EXPORT_SYMBOL(lov_lsm_get);
 884
 885 void lov_lsm_put(struct cl_object *unused, struct lov_stripe_md *lsm)
 886 {
 887         if (lsm != NULL)
 888                 lov_free_memmd(&lsm);
 889 }
 890 EXPORT_SYMBOL(lov_lsm_put);
 891
 892 int lov_read_and_clear_async_rc(struct cl_object *clob)
 893 {
 894         struct lu_object *luobj;
 895         int rc = 0;
 896         ENTRY;
 897
 898         luobj = lu_object_locate(&cl_object_header(clob)->coh_lu,
 899                                  &lov_device_type);
 900         if (luobj != NULL) {
 901                 struct lov_object *lov = lu2lov(luobj);
 902
 903                 lov_conf_freeze(lov);
 904                 switch (lov->lo_type) {
 905                 case LLT_RAID0: {
 906                         struct lov_stripe_md *lsm;
 907                         int i;
 908
 909                         lsm = lov->lo_lsm;
 910                         LASSERT(lsm != NULL);
 911                         for (i = 0; i < lsm->lsm_stripe_count; i++) {
 912                                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
 913                                 if (loi->loi_ar.ar_rc && !rc)
 914                                         rc = loi->loi_ar.ar_rc;
 915                                 loi->loi_ar.ar_rc = 0;
 916                         }
 917                 }
 918                 case LLT_EMPTY:
 919                         break;
 920                 default:
 921                         LBUG();
 922                 }
 923                 lov_conf_thaw(lov);
 924         }
 925         RETURN(rc);
 926 }
 927 EXPORT_SYMBOL(lov_read_and_clear_async_rc);
 928
 929 /** @} lov */