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 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov);
  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: ostid: "DOSTID
 152                " idx: %d gen: %d\n",
 153                PFID(&subhdr->coh_lu.loh_fid), subhdr, idx,
 154                PFID(&hdr->coh_lu.loh_fid), hdr, POSTID(&oinfo->loi_oi),
 155                oinfo->loi_ost_idx, oinfo->loi_ost_gen);
 156
 157         if (parent == NULL) {
 158                 subhdr->coh_parent = hdr;
 159                 subhdr->coh_nesting = hdr->coh_nesting + 1;
 160                 lu_object_ref_add(&stripe->co_lu, "lov-parent", lov);
 161                 r0->lo_sub[idx] = cl2lovsub(stripe);
 162                 r0->lo_sub[idx]->lso_super = lov;
 163                 r0->lo_sub[idx]->lso_index = idx;
 164                 result = 0;
 165         } else {
 166                 struct lu_object  *old_obj;
 167                 struct lov_object *old_lov;
 168                 unsigned int mask = D_INODE;
 169
 170                 old_obj = lu_object_locate(&parent->coh_lu, &lov_device_type);
 171                 LASSERT(old_obj != NULL);
 172                 old_lov = cl2lov(lu2cl(old_obj));
 173                 if (old_lov->lo_layout_invalid) {
 174                         /* the object's layout has already changed but isn't
 175                          * refreshed */
 176                         lu_object_unhash(env, &stripe->co_lu);
 177                         result = -EAGAIN;
 178                 } else {
 179                         mask = D_ERROR;
 180                         result = -EIO;
 181                 }
 182
 183                 LU_OBJECT_DEBUG(mask, env, &stripe->co_lu,
 184                                 "stripe %d is already owned.\n", idx);
 185                 LU_OBJECT_DEBUG(mask, env, old_obj, "owned.\n");
 186                 LU_OBJECT_HEADER(mask, env, lov2lu(lov), "try to own.\n");
 187                 cl_object_put(env, stripe);
 188         }
 189         return result;
 190 }
 191
 192 static int lov_init_raid0(const struct lu_env *env,
 193                           struct lov_device *dev, struct lov_object *lov,
 194                           const struct cl_object_conf *conf,
 195                           union  lov_layout_state *state)
 196 {
 197         int result;
 198         int i;
 199
 200         struct cl_object        *stripe;
 201         struct lov_thread_info  *lti     = lov_env_info(env);
 202         struct cl_object_conf   *subconf = &lti->lti_stripe_conf;
 203         struct lov_stripe_md    *lsm     = conf->u.coc_md->lsm;
 204         struct lu_fid           *ofid    = &lti->lti_fid;
 205         struct lov_layout_raid0 *r0      = &state->raid0;
 206
 207         ENTRY;
 208
 209         if (lsm->lsm_magic != LOV_MAGIC_V1 && lsm->lsm_magic != LOV_MAGIC_V3) {
 210                 dump_lsm(D_ERROR, lsm);
 211                 LASSERTF(0, "magic mismatch, expected %d/%d, actual %d.\n",
 212                          LOV_MAGIC_V1, LOV_MAGIC_V3, lsm->lsm_magic);
 213         }
 214
 215         LASSERT(lov->lo_lsm == NULL);
 216         lov->lo_lsm = lsm_addref(lsm);
 217         r0->lo_nr  = lsm->lsm_stripe_count;
 218         LASSERT(r0->lo_nr <= lov_targets_nr(dev));
 219
 220         OBD_ALLOC_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 221         if (r0->lo_sub != NULL) {
 222                 result = 0;
 223                 subconf->coc_inode = conf->coc_inode;
 224                 spin_lock_init(&r0->lo_sub_lock);
 225                 /*
 226                  * Create stripe cl_objects.
 227                  */
 228                 for (i = 0; i < r0->lo_nr && result == 0; ++i) {
 229                         struct cl_device *subdev;
 230                         struct lov_oinfo *oinfo = lsm->lsm_oinfo[i];
 231                         int ost_idx = oinfo->loi_ost_idx;
 232
 233                         result = ostid_to_fid(ofid, &oinfo->loi_oi,
 234                                               oinfo->loi_ost_idx);
 235                         if (result != 0)
 236                                 GOTO(out, result);
 237
 238                         subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
 239                         subconf->u.coc_oinfo = oinfo;
 240                         LASSERTF(subdev != NULL, "not init ost %d\n", ost_idx);
 241                         /* In the function below, .hs_keycmp resolves to
 242                          * lu_obj_hop_keycmp() */
 243                         /* coverity[overrun-buffer-val] */
 244                         stripe = lov_sub_find(env, subdev, ofid, subconf);
 245                         if (!IS_ERR(stripe)) {
 246                                 result = lov_init_sub(env, lov, stripe, r0, i);
 247                                 if (result == -EAGAIN) { /* try again */
 248                                         --i;
 249                                         result = 0;
 250                                 }
 251                         } else {
 252                                 result = PTR_ERR(stripe);
 253                         }
 254                 }
 255         } else
 256                 result = -ENOMEM;
 257 out:
 258         RETURN(result);
 259 }
 260
 261 static int lov_delete_empty(const struct lu_env *env, struct lov_object *lov,
 262                             union lov_layout_state *state)
 263 {
 264         LASSERT(lov->lo_type == LLT_EMPTY);
 265
 266         lov_layout_wait(env, lov);
 267
 268         cl_object_prune(env, &lov->lo_cl);
 269         return 0;
 270 }
 271
 272 static void lov_subobject_kill(const struct lu_env *env, struct lov_object *lov,
 273                                struct lovsub_object *los, int idx)
 274 {
 275         struct cl_object        *sub;
 276         struct lov_layout_raid0 *r0;
 277         struct lu_site          *site;
 278         struct lu_site_bkt_data *bkt;
 279         cfs_waitlink_t          *waiter;
 280
 281         r0  = &lov->u.raid0;
 282         LASSERT(r0->lo_sub[idx] == los);
 283
 284         sub  = lovsub2cl(los);
 285         site = sub->co_lu.lo_dev->ld_site;
 286         bkt  = lu_site_bkt_from_fid(site, &sub->co_lu.lo_header->loh_fid);
 287
 288         cl_object_kill(env, sub);
 289         /* release a reference to the sub-object and ... */
 290         lu_object_ref_del(&sub->co_lu, "lov-parent", lov);
 291         cl_object_put(env, sub);
 292
 293         /* ... wait until it is actually destroyed---sub-object clears its
 294          * ->lo_sub[] slot in lovsub_object_fini() */
 295         if (r0->lo_sub[idx] == los) {
 296                 waiter = &lov_env_info(env)->lti_waiter;
 297                 cfs_waitlink_init(waiter);
 298                 cfs_waitq_add(&bkt->lsb_marche_funebre, waiter);
 299                 cfs_set_current_state(CFS_TASK_UNINT);
 300                 while (1) {
 301                         /* this wait-queue is signaled at the end of
 302                          * lu_object_free(). */
 303                         cfs_set_current_state(CFS_TASK_UNINT);
 304                         spin_lock(&r0->lo_sub_lock);
 305                         if (r0->lo_sub[idx] == los) {
 306                                 spin_unlock(&r0->lo_sub_lock);
 307                                 cfs_waitq_wait(waiter, CFS_TASK_UNINT);
 308                         } else {
 309                                 spin_unlock(&r0->lo_sub_lock);
 310                                 cfs_set_current_state(CFS_TASK_RUNNING);
 311                                 break;
 312                         }
 313                 }
 314                 cfs_waitq_del(&bkt->lsb_marche_funebre, waiter);
 315         }
 316         LASSERT(r0->lo_sub[idx] == NULL);
 317 }
 318
 319 static int lov_delete_raid0(const struct lu_env *env, struct lov_object *lov,
 320                             union lov_layout_state *state)
 321 {
 322         struct lov_layout_raid0 *r0 = &state->raid0;
 323         struct lov_stripe_md    *lsm = lov->lo_lsm;
 324         int i;
 325
 326         ENTRY;
 327
 328         dump_lsm(D_INODE, lsm);
 329
 330         lov_layout_wait(env, lov);
 331         if (r0->lo_sub != NULL) {
 332                 for (i = 0; i < r0->lo_nr; ++i) {
 333                         struct lovsub_object *los = r0->lo_sub[i];
 334
 335                         if (los != NULL) {
 336                                 cl_locks_prune(env, &los->lso_cl, 1);
 337                                 /*
 338                                  * If top-level object is to be evicted from
 339                                  * the cache, so are its sub-objects.
 340                                  */
 341                                 lov_subobject_kill(env, lov, los, i);
 342                         }
 343                 }
 344         }
 345         cl_object_prune(env, &lov->lo_cl);
 346         RETURN(0);
 347 }
 348
 349 static void lov_fini_empty(const struct lu_env *env, struct lov_object *lov,
 350                            union lov_layout_state *state)
 351 {
 352         LASSERT(lov->lo_type == LLT_EMPTY);
 353 }
 354
 355 static void lov_fini_raid0(const struct lu_env *env, struct lov_object *lov,
 356                            union lov_layout_state *state)
 357 {
 358         struct lov_layout_raid0 *r0 = &state->raid0;
 359         ENTRY;
 360
 361         if (r0->lo_sub != NULL) {
 362                 OBD_FREE_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 363                 r0->lo_sub = NULL;
 364         }
 365
 366         dump_lsm(D_INODE, lov->lo_lsm);
 367         lov_free_memmd(&lov->lo_lsm);
 368
 369         EXIT;
 370 }
 371
 372 static int lov_print_empty(const struct lu_env *env, void *cookie,
 373                            lu_printer_t p, const struct lu_object *o)
 374 {
 375         (*p)(env, cookie, "empty %d\n", lu2lov(o)->lo_layout_invalid);
 376         return 0;
 377 }
 378
 379 static int lov_print_raid0(const struct lu_env *env, void *cookie,
 380                            lu_printer_t p, const struct lu_object *o)
 381 {
 382         struct lov_object       *lov = lu2lov(o);
 383         struct lov_layout_raid0 *r0  = lov_r0(lov);
 384         struct lov_stripe_md    *lsm = lov->lo_lsm;
 385         int i;
 386
 387         (*p)(env, cookie, "stripes: %d, %svalid, lsm{%p 0x%08X %d %u %u}: \n",
 388                 r0->lo_nr, lov->lo_layout_invalid ? "in" : "", lsm,
 389                 lsm->lsm_magic, cfs_atomic_read(&lsm->lsm_refc),
 390                 lsm->lsm_stripe_count, lsm->lsm_layout_gen);
 391         for (i = 0; i < r0->lo_nr; ++i) {
 392                 struct lu_object *sub;
 393
 394                 if (r0->lo_sub[i] != NULL) {
 395                         sub = lovsub2lu(r0->lo_sub[i]);
 396                         lu_object_print(env, cookie, p, sub);
 397                 } else
 398                         (*p)(env, cookie, "sub %d absent\n", i);
 399         }
 400         return 0;
 401 }
 402
 403 /**
 404  * Implements cl_object_operations::coo_attr_get() method for an object
 405  * without stripes (LLT_EMPTY layout type).
 406  *
 407  * The only attributes this layer is authoritative in this case is
 408  * cl_attr::cat_blocks---it's 0.
 409  */
 410 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
 411                               struct cl_attr *attr)
 412 {
 413         attr->cat_blocks = 0;
 414         return 0;
 415 }
 416
 417 static int lov_attr_get_raid0(const struct lu_env *env, struct cl_object *obj,
 418                               struct cl_attr *attr)
 419 {
 420         struct lov_object       *lov = cl2lov(obj);
 421         struct lov_layout_raid0 *r0 = lov_r0(lov);
 422         struct cl_attr          *lov_attr = &r0->lo_attr;
 423         int                      result = 0;
 424
 425         ENTRY;
 426
 427         /* this is called w/o holding type guard mutex, so it must be inside
 428          * an on going IO otherwise lsm may be replaced.
 429          * LU-2117: it turns out there exists one exception. For mmaped files,
 430          * the lock of those files may be requested in the other file's IO
 431          * context, and this function is called in ccc_lock_state(), it will
 432          * hit this assertion.
 433          * Anyway, it's still okay to call attr_get w/o type guard as layout
 434          * can't go if locks exist. */
 435         /* LASSERT(cfs_atomic_read(&lsm->lsm_refc) > 1); */
 436
 437         if (!r0->lo_attr_valid) {
 438                 struct lov_stripe_md    *lsm = lov->lo_lsm;
 439                 struct ost_lvb          *lvb = &lov_env_info(env)->lti_lvb;
 440                 __u64                    kms = 0;
 441
 442                 memset(lvb, 0, sizeof(*lvb));
 443                 /* XXX: timestamps can be negative by sanity:test_39m,
 444                  * how can it be? */
 445                 lvb->lvb_atime = LLONG_MIN;
 446                 lvb->lvb_ctime = LLONG_MIN;
 447                 lvb->lvb_mtime = LLONG_MIN;
 448
 449                 /*
 450                  * XXX that should be replaced with a loop over sub-objects,
 451                  * doing cl_object_attr_get() on them. But for now, let's
 452                  * reuse old lov code.
 453                  */
 454
 455                 /*
 456                  * XXX take lsm spin-lock to keep lov_merge_lvb_kms()
 457                  * happy. It's not needed, because new code uses
 458                  * ->coh_attr_guard spin-lock to protect consistency of
 459                  * sub-object attributes.
 460                  */
 461                 lov_stripe_lock(lsm);
 462                 result = lov_merge_lvb_kms(lsm, lvb, &kms);
 463                 lov_stripe_unlock(lsm);
 464                 if (result == 0) {
 465                         cl_lvb2attr(lov_attr, lvb);
 466                         lov_attr->cat_kms = kms;
 467                         r0->lo_attr_valid = 1;
 468                 }
 469         }
 470         if (result == 0) { /* merge results */
 471                 attr->cat_blocks = lov_attr->cat_blocks;
 472                 attr->cat_size = lov_attr->cat_size;
 473                 attr->cat_kms = lov_attr->cat_kms;
 474                 if (attr->cat_atime < lov_attr->cat_atime)
 475                         attr->cat_atime = lov_attr->cat_atime;
 476                 if (attr->cat_ctime < lov_attr->cat_ctime)
 477                         attr->cat_ctime = lov_attr->cat_ctime;
 478                 if (attr->cat_mtime < lov_attr->cat_mtime)
 479                         attr->cat_mtime = lov_attr->cat_mtime;
 480         }
 481         RETURN(result);
 482 }
 483
 484 const static struct lov_layout_operations lov_dispatch[] = {
 485         [LLT_EMPTY] = {
 486                 .llo_init      = lov_init_empty,
 487                 .llo_delete    = lov_delete_empty,
 488                 .llo_fini      = lov_fini_empty,
 489                 .llo_install   = lov_install_empty,
 490                 .llo_print     = lov_print_empty,
 491                 .llo_page_init = lov_page_init_empty,
 492                 .llo_lock_init = lov_lock_init_empty,
 493                 .llo_io_init   = lov_io_init_empty,
 494                 .llo_getattr   = lov_attr_get_empty
 495         },
 496         [LLT_RAID0] = {
 497                 .llo_init      = lov_init_raid0,
 498                 .llo_delete    = lov_delete_raid0,
 499                 .llo_fini      = lov_fini_raid0,
 500                 .llo_install   = lov_install_raid0,
 501                 .llo_print     = lov_print_raid0,
 502                 .llo_page_init = lov_page_init_raid0,
 503                 .llo_lock_init = lov_lock_init_raid0,
 504                 .llo_io_init   = lov_io_init_raid0,
 505                 .llo_getattr   = lov_attr_get_raid0
 506         }
 507 };
 508
 509
 510 /**
 511  * Performs a double-dispatch based on the layout type of an object.
 512  */
 513 #define LOV_2DISPATCH_NOLOCK(obj, op, ...)                              \
 514 ({                                                                      \
 515         struct lov_object                      *__obj = (obj);          \
 516         enum lov_layout_type                    __llt;                  \
 517                                                                         \
 518         __llt = __obj->lo_type;                                         \
 519         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 520         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 521 })
 522
 523 static inline void lov_conf_freeze(struct lov_object *lov)
 524 {
 525         if (lov->lo_owner != cfs_current())
 526                 down_read(&lov->lo_type_guard);
 527 }
 528
 529 static inline void lov_conf_thaw(struct lov_object *lov)
 530 {
 531         if (lov->lo_owner != cfs_current())
 532                 up_read(&lov->lo_type_guard);
 533 }
 534
 535 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...)                       \
 536 ({                                                                      \
 537         struct lov_object                      *__obj = (obj);          \
 538         int                                     __lock = !!(lock);      \
 539         typeof(lov_dispatch[0].op(__VA_ARGS__)) __result;               \
 540                                                                         \
 541         if (__lock)                                                     \
 542                 lov_conf_freeze(__obj);                                 \
 543         __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__);          \
 544         if (__lock)                                                     \
 545                 lov_conf_thaw(__obj);                                   \
 546         __result;                                                       \
 547 })
 548
 549 /**
 550  * Performs a locked double-dispatch based on the layout type of an object.
 551  */
 552 #define LOV_2DISPATCH(obj, op, ...)                     \
 553         LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
 554
 555 #define LOV_2DISPATCH_VOID(obj, op, ...)                                \
 556 do {                                                                    \
 557         struct lov_object                      *__obj = (obj);          \
 558         enum lov_layout_type                    __llt;                  \
 559                                                                         \
 560         lov_conf_freeze(__obj);                                         \
 561         __llt = __obj->lo_type;                                         \
 562         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 563         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 564         lov_conf_thaw(__obj);                                           \
 565 } while (0)
 566
 567 static void lov_conf_lock(struct lov_object *lov)
 568 {
 569         LASSERT(lov->lo_owner != cfs_current());
 570         down_write(&lov->lo_type_guard);
 571         LASSERT(lov->lo_owner == NULL);
 572         lov->lo_owner = cfs_current();
 573 }
 574
 575 static void lov_conf_unlock(struct lov_object *lov)
 576 {
 577         lov->lo_owner = NULL;
 578         up_write(&lov->lo_type_guard);
 579 }
 580
 581 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov)
 582 {
 583         struct l_wait_info lwi = { 0 };
 584         ENTRY;
 585
 586         while (cfs_atomic_read(&lov->lo_active_ios) > 0) {
 587                 CDEBUG(D_INODE, "file:"DFID" wait for active IO, now: %d.\n",
 588                         PFID(lu_object_fid(lov2lu(lov))),
 589                         cfs_atomic_read(&lov->lo_active_ios));
 590
 591                 l_wait_event(lov->lo_waitq,
 592                              cfs_atomic_read(&lov->lo_active_ios) == 0, &lwi);
 593         }
 594         RETURN(0);
 595 }
 596
 597 static int lov_layout_change(const struct lu_env *unused,
 598                              struct lov_object *lov,
 599                              const struct cl_object_conf *conf)
 600 {
 601         int result;
 602         enum lov_layout_type llt = LLT_EMPTY;
 603         union lov_layout_state *state = &lov->u;
 604         const struct lov_layout_operations *old_ops;
 605         const struct lov_layout_operations *new_ops;
 606
 607         struct cl_object_header *hdr = cl_object_header(&lov->lo_cl);
 608         void *cookie;
 609         struct lu_env *env;
 610         int refcheck;
 611         ENTRY;
 612
 613         LASSERT(0 <= lov->lo_type && lov->lo_type < ARRAY_SIZE(lov_dispatch));
 614
 615         if (conf->u.coc_md != NULL && conf->u.coc_md->lsm != NULL)
 616                 llt = LLT_RAID0; /* only raid0 is supported. */
 617         LASSERT(0 <= llt && llt < ARRAY_SIZE(lov_dispatch));
 618
 619         cookie = cl_env_reenter();
 620         env = cl_env_get(&refcheck);
 621         if (IS_ERR(env)) {
 622                 cl_env_reexit(cookie);
 623                 RETURN(PTR_ERR(env));
 624         }
 625
 626         old_ops = &lov_dispatch[lov->lo_type];
 627         new_ops = &lov_dispatch[llt];
 628
 629         result = old_ops->llo_delete(env, lov, &lov->u);
 630         if (result == 0) {
 631                 old_ops->llo_fini(env, lov, &lov->u);
 632
 633                 LASSERT(cfs_atomic_read(&lov->lo_active_ios) == 0);
 634                 LASSERT(hdr->coh_tree.rnode == NULL);
 635                 LASSERT(hdr->coh_pages == 0);
 636
 637                 lov->lo_type = LLT_EMPTY;
 638                 result = new_ops->llo_init(env,
 639                                         lu2lov_dev(lov->lo_cl.co_lu.lo_dev),
 640                                         lov, conf, state);
 641                 if (result == 0) {
 642                         new_ops->llo_install(env, lov, state);
 643                         lov->lo_type = llt;
 644                 } else {
 645                         new_ops->llo_delete(env, lov, state);
 646                         new_ops->llo_fini(env, lov, state);
 647                         /* this file becomes an EMPTY file. */
 648                 }
 649         }
 650
 651         cl_env_put(env, &refcheck);
 652         cl_env_reexit(cookie);
 653         RETURN(result);
 654 }
 655
 656 /*****************************************************************************
 657  *
 658  * Lov object operations.
 659  *
 660  */
 661
 662 int lov_object_init(const struct lu_env *env, struct lu_object *obj,
 663                     const struct lu_object_conf *conf)
 664 {
 665         struct lov_device            *dev   = lu2lov_dev(obj->lo_dev);
 666         struct lov_object            *lov   = lu2lov(obj);
 667         const struct cl_object_conf  *cconf = lu2cl_conf(conf);
 668         union  lov_layout_state      *set   = &lov->u;
 669         const struct lov_layout_operations *ops;
 670         int result;
 671
 672         ENTRY;
 673         init_rwsem(&lov->lo_type_guard);
 674         cfs_atomic_set(&lov->lo_active_ios, 0);
 675         cfs_waitq_init(&lov->lo_waitq);
 676
 677         cl_object_page_init(lu2cl(obj), sizeof(struct lov_page));
 678
 679         /* no locking is necessary, as object is being created */
 680         lov->lo_type = cconf->u.coc_md->lsm != NULL ? LLT_RAID0 : LLT_EMPTY;
 681         ops = &lov_dispatch[lov->lo_type];
 682         result = ops->llo_init(env, dev, lov, cconf, set);
 683         if (result == 0)
 684                 ops->llo_install(env, lov, set);
 685         RETURN(result);
 686 }
 687
 688 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
 689                         const struct cl_object_conf *conf)
 690 {
 691         struct lov_stripe_md *lsm = NULL;
 692         struct lov_object *lov = cl2lov(obj);
 693         int result = 0;
 694         ENTRY;
 695
 696         lov_conf_lock(lov);
 697         if (conf->coc_opc == OBJECT_CONF_INVALIDATE) {
 698                 lov->lo_layout_invalid = true;
 699                 GOTO(out, result = 0);
 700         }
 701
 702         if (conf->coc_opc == OBJECT_CONF_WAIT) {
 703                 if (lov->lo_layout_invalid &&
 704                     cfs_atomic_read(&lov->lo_active_ios) > 0) {
 705                         lov_conf_unlock(lov);
 706                         result = lov_layout_wait(env, lov);
 707                         lov_conf_lock(lov);
 708                 }
 709                 GOTO(out, result);
 710         }
 711
 712         LASSERT(conf->coc_opc == OBJECT_CONF_SET);
 713
 714         if (conf->u.coc_md != NULL)
 715                 lsm = conf->u.coc_md->lsm;
 716         if ((lsm == NULL && lov->lo_lsm == NULL) ||
 717             (lsm != NULL && lov->lo_lsm != NULL &&
 718              lov->lo_lsm->lsm_layout_gen == lsm->lsm_layout_gen)) {
 719                 /* same version of layout */
 720                 lov->lo_layout_invalid = false;
 721                 GOTO(out, result = 0);
 722         }
 723
 724         /* will change layout - check if there still exists active IO. */
 725         if (cfs_atomic_read(&lov->lo_active_ios) > 0) {
 726                 lov->lo_layout_invalid = true;
 727                 GOTO(out, result = -EBUSY);
 728         }
 729
 730         lov->lo_layout_invalid = lov_layout_change(env, lov, conf);
 731         EXIT;
 732
 733 out:
 734         lov_conf_unlock(lov);
 735         RETURN(result);
 736 }
 737
 738 static void lov_object_delete(const struct lu_env *env, struct lu_object *obj)
 739 {
 740         struct lov_object *lov = lu2lov(obj);
 741
 742         ENTRY;
 743         LOV_2DISPATCH_VOID(lov, llo_delete, env, lov, &lov->u);
 744         EXIT;
 745 }
 746
 747 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
 748 {
 749         struct lov_object *lov = lu2lov(obj);
 750
 751         ENTRY;
 752         LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
 753         lu_object_fini(obj);
 754         OBD_SLAB_FREE_PTR(lov, lov_object_kmem);
 755         EXIT;
 756 }
 757
 758 static int lov_object_print(const struct lu_env *env, void *cookie,
 759                             lu_printer_t p, const struct lu_object *o)
 760 {
 761         return LOV_2DISPATCH_NOLOCK(lu2lov(o), llo_print, env, cookie, p, o);
 762 }
 763
 764 int lov_page_init(const struct lu_env *env, struct cl_object *obj,
 765                 struct cl_page *page, cfs_page_t *vmpage)
 766 {
 767         return LOV_2DISPATCH_NOLOCK(cl2lov(obj),
 768                                     llo_page_init, env, obj, page, vmpage);
 769 }
 770
 771 /**
 772  * Implements cl_object_operations::clo_io_init() method for lov
 773  * layer. Dispatches to the appropriate layout io initialization method.
 774  */
 775 int lov_io_init(const struct lu_env *env, struct cl_object *obj,
 776                 struct cl_io *io)
 777 {
 778         CL_IO_SLICE_CLEAN(lov_env_io(env), lis_cl);
 779         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
 780                                      !io->ci_ignore_layout, env, obj, io);
 781 }
 782
 783 /**
 784  * An implementation of cl_object_operations::clo_attr_get() method for lov
 785  * layer. For raid0 layout this collects and merges attributes of all
 786  * sub-objects.
 787  */
 788 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
 789                         struct cl_attr *attr)
 790 {
 791         /* do not take lock, as this function is called under a
 792          * spin-lock. Layout is protected from changing by ongoing IO. */
 793         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
 794 }
 795
 796 static int lov_attr_set(const struct lu_env *env, struct cl_object *obj,
 797                         const struct cl_attr *attr, unsigned valid)
 798 {
 799         /*
 800          * No dispatch is required here, as no layout implements this.
 801          */
 802         return 0;
 803 }
 804
 805 int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
 806                   struct cl_lock *lock, const struct cl_io *io)
 807 {
 808         /* No need to lock because we've taken one refcount of layout.  */
 809         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_lock_init, env, obj, lock,
 810                                     io);
 811 }
 812
 813 static const struct cl_object_operations lov_ops = {
 814         .coo_page_init = lov_page_init,
 815         .coo_lock_init = lov_lock_init,
 816         .coo_io_init   = lov_io_init,
 817         .coo_attr_get  = lov_attr_get,
 818         .coo_attr_set  = lov_attr_set,
 819         .coo_conf_set  = lov_conf_set
 820 };
 821
 822 static const struct lu_object_operations lov_lu_obj_ops = {
 823         .loo_object_init      = lov_object_init,
 824         .loo_object_delete    = lov_object_delete,
 825         .loo_object_release   = NULL,
 826         .loo_object_free      = lov_object_free,
 827         .loo_object_print     = lov_object_print,
 828         .loo_object_invariant = NULL
 829 };
 830
 831 struct lu_object *lov_object_alloc(const struct lu_env *env,
 832                                    const struct lu_object_header *unused,
 833                                    struct lu_device *dev)
 834 {
 835         struct lov_object *lov;
 836         struct lu_object  *obj;
 837
 838         ENTRY;
 839         OBD_SLAB_ALLOC_PTR_GFP(lov, lov_object_kmem, CFS_ALLOC_IO);
 840         if (lov != NULL) {
 841                 obj = lov2lu(lov);
 842                 lu_object_init(obj, NULL, dev);
 843                 lov->lo_cl.co_ops = &lov_ops;
 844                 lov->lo_type = -1; /* invalid, to catch uninitialized type */
 845                 /*
 846                  * object io operation vector (cl_object::co_iop) is installed
 847                  * later in lov_object_init(), as different vectors are used
 848                  * for object with different layouts.
 849                  */
 850                 obj->lo_ops = &lov_lu_obj_ops;
 851         } else
 852                 obj = NULL;
 853         RETURN(obj);
 854 }
 855
 856 struct lov_stripe_md *lov_lsm_addref(struct lov_object *lov)
 857 {
 858         struct lov_stripe_md *lsm = NULL;
 859
 860         lov_conf_freeze(lov);
 861         if (lov->lo_lsm != NULL) {
 862                 lsm = lsm_addref(lov->lo_lsm);
 863                 CDEBUG(D_INODE, "lsm %p addref %d/%d by %p.\n",
 864                         lsm, cfs_atomic_read(&lsm->lsm_refc),
 865                         lov->lo_layout_invalid, cfs_current());
 866         }
 867         lov_conf_thaw(lov);
 868         return lsm;
 869 }
 870
 871 void lov_lsm_decref(struct lov_object *lov, struct lov_stripe_md *lsm)
 872 {
 873         if (lsm == NULL)
 874                 return;
 875
 876         CDEBUG(D_INODE, "lsm %p decref %d by %p.\n",
 877                 lsm, cfs_atomic_read(&lsm->lsm_refc), cfs_current());
 878
 879         lov_free_memmd(&lsm);
 880 }
 881
 882 struct lov_stripe_md *lov_lsm_get(struct cl_object *clobj)
 883 {
 884         struct lu_object *luobj;
 885         struct lov_stripe_md *lsm = NULL;
 886
 887         if (clobj == NULL)
 888                 return NULL;
 889
 890         luobj = lu_object_locate(&cl_object_header(clobj)->coh_lu,
 891                                  &lov_device_type);
 892         if (luobj != NULL)
 893                 lsm = lov_lsm_addref(lu2lov(luobj));
 894         return lsm;
 895 }
 896 EXPORT_SYMBOL(lov_lsm_get);
 897
 898 void lov_lsm_put(struct cl_object *unused, struct lov_stripe_md *lsm)
 899 {
 900         if (lsm != NULL)
 901                 lov_free_memmd(&lsm);
 902 }
 903 EXPORT_SYMBOL(lov_lsm_put);
 904
 905 int lov_read_and_clear_async_rc(struct cl_object *clob)
 906 {
 907         struct lu_object *luobj;
 908         int rc = 0;
 909         ENTRY;
 910
 911         luobj = lu_object_locate(&cl_object_header(clob)->coh_lu,
 912                                  &lov_device_type);
 913         if (luobj != NULL) {
 914                 struct lov_object *lov = lu2lov(luobj);
 915
 916                 lov_conf_freeze(lov);
 917                 switch (lov->lo_type) {
 918                 case LLT_RAID0: {
 919                         struct lov_stripe_md *lsm;
 920                         int i;
 921
 922                         lsm = lov->lo_lsm;
 923                         LASSERT(lsm != NULL);
 924                         for (i = 0; i < lsm->lsm_stripe_count; i++) {
 925                                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
 926                                 if (loi->loi_ar.ar_rc && !rc)
 927                                         rc = loi->loi_ar.ar_rc;
 928                                 loi->loi_ar.ar_rc = 0;
 929                         }
 930                 }
 931                 case LLT_EMPTY:
 932                         break;
 933                 default:
 934                         LBUG();
 935                 }
 936                 lov_conf_thaw(lov);
 937         }
 938         RETURN(rc);
 939 }
 940 EXPORT_SYMBOL(lov_read_and_clear_async_rc);
 941
 942 /** @} lov */