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, 2013, 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
  46 /** \addtogroup lov
  47  *  @{
  48  */
  49
  50 /*****************************************************************************
  51  *
  52  * Layout operations.
  53  *
  54  */
  55
  56 struct lov_layout_operations {
  57         int (*llo_init)(const struct lu_env *env, struct lov_device *dev,
  58                         struct lov_object *lov,
  59                         const struct cl_object_conf *conf,
  60                         union lov_layout_state *state);
  61         int (*llo_delete)(const struct lu_env *env, struct lov_object *lov,
  62                            union lov_layout_state *state);
  63         void (*llo_fini)(const struct lu_env *env, struct lov_object *lov,
  64                          union lov_layout_state *state);
  65         void (*llo_install)(const struct lu_env *env, struct lov_object *lov,
  66                             union lov_layout_state *state);
  67         int  (*llo_print)(const struct lu_env *env, void *cookie,
  68                           lu_printer_t p, const struct lu_object *o);
  69         int  (*llo_page_init)(const struct lu_env *env, struct cl_object *obj,
  70                               struct cl_page *page, pgoff_t index);
  71         int  (*llo_lock_init)(const struct lu_env *env,
  72                               struct cl_object *obj, struct cl_lock *lock,
  73                               const struct cl_io *io);
  74         int  (*llo_io_init)(const struct lu_env *env,
  75                             struct cl_object *obj, struct cl_io *io);
  76         int  (*llo_getattr)(const struct lu_env *env, struct cl_object *obj,
  77                             struct cl_attr *attr);
  78 };
  79
  80 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov);
  81
  82 /*****************************************************************************
  83  *
  84  * Lov object layout operations.
  85  *
  86  */
  87
  88 static void lov_install_empty(const struct lu_env *env,
  89                               struct lov_object *lov,
  90                               union  lov_layout_state *state)
  91 {
  92         /*
  93          * File without objects.
  94          */
  95 }
  96
  97 static int lov_init_empty(const struct lu_env *env,
  98                           struct lov_device *dev, struct lov_object *lov,
  99                           const struct cl_object_conf *conf,
 100                           union  lov_layout_state *state)
 101 {
 102         return 0;
 103 }
 104
 105 static void lov_install_raid0(const struct lu_env *env,
 106                               struct lov_object *lov,
 107                               union  lov_layout_state *state)
 108 {
 109 }
 110
 111 static struct cl_object *lov_sub_find(const struct lu_env *env,
 112                                       struct cl_device *dev,
 113                                       const struct lu_fid *fid,
 114                                       const struct cl_object_conf *conf)
 115 {
 116         struct lu_object *o;
 117
 118         ENTRY;
 119         o = lu_object_find_at(env, cl2lu_dev(dev), fid, &conf->coc_lu);
 120         LASSERT(ergo(!IS_ERR(o), o->lo_dev->ld_type == &lovsub_device_type));
 121         RETURN(lu2cl(o));
 122 }
 123
 124 static int lov_init_sub(const struct lu_env *env, struct lov_object *lov,
 125                         struct cl_object *stripe, struct lov_layout_raid0 *r0,
 126                         int idx)
 127 {
 128         struct cl_object_header *hdr;
 129         struct cl_object_header *subhdr;
 130         struct cl_object_header *parent;
 131         struct lov_oinfo        *oinfo;
 132         int result;
 133
 134         if (OBD_FAIL_CHECK(OBD_FAIL_LOV_INIT)) {
 135                 /* For sanity:test_206.
 136                  * Do not leave the object in cache to avoid accessing
 137                  * freed memory. This is because osc_object is referring to
 138                  * lov_oinfo of lsm_stripe_data which will be freed due to
 139                  * this failure. */
 140                 cl_object_kill(env, stripe);
 141                 cl_object_put(env, stripe);
 142                 return -EIO;
 143         }
 144
 145         hdr    = cl_object_header(lov2cl(lov));
 146         subhdr = cl_object_header(stripe);
 147
 148         oinfo = lov->lo_lsm->lsm_oinfo[idx];
 149         CDEBUG(D_INODE, DFID"@%p[%d] -> "DFID"@%p: ostid: "DOSTID
 150                " idx: %d gen: %d\n",
 151                PFID(&subhdr->coh_lu.loh_fid), subhdr, idx,
 152                PFID(&hdr->coh_lu.loh_fid), hdr, POSTID(&oinfo->loi_oi),
 153                oinfo->loi_ost_idx, oinfo->loi_ost_gen);
 154
 155         /* reuse ->coh_attr_guard to protect coh_parent change */
 156         spin_lock(&subhdr->coh_attr_guard);
 157         parent = subhdr->coh_parent;
 158         if (parent == NULL) {
 159                 subhdr->coh_parent = hdr;
 160                 spin_unlock(&subhdr->coh_attr_guard);
 161                 subhdr->coh_nesting = hdr->coh_nesting + 1;
 162                 lu_object_ref_add(&stripe->co_lu, "lov-parent", lov);
 163                 r0->lo_sub[idx] = cl2lovsub(stripe);
 164                 r0->lo_sub[idx]->lso_super = lov;
 165                 r0->lo_sub[idx]->lso_index = idx;
 166                 result = 0;
 167         } else {
 168                 struct lu_object  *old_obj;
 169                 struct lov_object *old_lov;
 170                 unsigned int mask = D_INODE;
 171
 172                 spin_unlock(&subhdr->coh_attr_guard);
 173                 old_obj = lu_object_locate(&parent->coh_lu, &lov_device_type);
 174                 LASSERT(old_obj != NULL);
 175                 old_lov = cl2lov(lu2cl(old_obj));
 176                 if (old_lov->lo_layout_invalid) {
 177                         /* the object's layout has already changed but isn't
 178                          * refreshed */
 179                         lu_object_unhash(env, &stripe->co_lu);
 180                         result = -EAGAIN;
 181                 } else {
 182                         mask = D_ERROR;
 183                         result = -EIO;
 184                 }
 185
 186                 LU_OBJECT_DEBUG(mask, env, &stripe->co_lu,
 187                                 "stripe %d is already owned.\n", idx);
 188                 LU_OBJECT_DEBUG(mask, env, old_obj, "owned.\n");
 189                 LU_OBJECT_HEADER(mask, env, lov2lu(lov), "try to own.\n");
 190                 cl_object_put(env, stripe);
 191         }
 192         return result;
 193 }
 194
 195 static int lov_page_slice_fixup(struct lov_object *lov,
 196                                 struct cl_object *stripe)
 197 {
 198         struct cl_object_header *hdr = cl_object_header(&lov->lo_cl);
 199         struct cl_object *o;
 200
 201         if (stripe == NULL)
 202                 return hdr->coh_page_bufsize - lov->lo_cl.co_slice_off -
 203                        cfs_size_round(sizeof(struct lov_page));
 204
 205         cl_object_for_each(o, stripe)
 206                 o->co_slice_off += hdr->coh_page_bufsize;
 207
 208         return cl_object_header(stripe)->coh_page_bufsize;
 209 }
 210
 211 static int lov_init_raid0(const struct lu_env *env,
 212                           struct lov_device *dev, struct lov_object *lov,
 213                           const struct cl_object_conf *conf,
 214                           union  lov_layout_state *state)
 215 {
 216         int result;
 217         int i;
 218
 219         struct cl_object        *stripe;
 220         struct lov_thread_info  *lti     = lov_env_info(env);
 221         struct cl_object_conf   *subconf = &lti->lti_stripe_conf;
 222         struct lov_stripe_md    *lsm     = conf->u.coc_md->lsm;
 223         struct lu_fid           *ofid    = &lti->lti_fid;
 224         struct lov_layout_raid0 *r0      = &state->raid0;
 225
 226         ENTRY;
 227
 228         if (lsm->lsm_magic != LOV_MAGIC_V1 && lsm->lsm_magic != LOV_MAGIC_V3) {
 229                 dump_lsm(D_ERROR, lsm);
 230                 LASSERTF(0, "magic mismatch, expected %d/%d, actual %d.\n",
 231                          LOV_MAGIC_V1, LOV_MAGIC_V3, lsm->lsm_magic);
 232         }
 233
 234         LASSERT(lov->lo_lsm == NULL);
 235         lov->lo_lsm = lsm_addref(lsm);
 236         r0->lo_nr  = lsm->lsm_stripe_count;
 237         LASSERT(r0->lo_nr <= lov_targets_nr(dev));
 238
 239         lov->lo_layout_invalid = true;
 240
 241         OBD_ALLOC_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 242         if (r0->lo_sub != NULL) {
 243                 int psz = 0;
 244
 245                 result = 0;
 246                 subconf->coc_inode = conf->coc_inode;
 247                 spin_lock_init(&r0->lo_sub_lock);
 248                 /*
 249                  * Create stripe cl_objects.
 250                  */
 251                 for (i = 0; i < r0->lo_nr && result == 0; ++i) {
 252                         struct cl_device *subdev;
 253                         struct lov_oinfo *oinfo = lsm->lsm_oinfo[i];
 254                         int ost_idx = oinfo->loi_ost_idx;
 255
 256                         if (lov_oinfo_is_dummy(oinfo))
 257                                 continue;
 258
 259                         result = ostid_to_fid(ofid, &oinfo->loi_oi,
 260                                               oinfo->loi_ost_idx);
 261                         if (result != 0)
 262                                 GOTO(out, result);
 263
 264                         subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
 265                         subconf->u.coc_oinfo = oinfo;
 266                         LASSERTF(subdev != NULL, "not init ost %d\n", ost_idx);
 267                         /* In the function below, .hs_keycmp resolves to
 268                          * lu_obj_hop_keycmp() */
 269                         /* coverity[overrun-buffer-val] */
 270                         stripe = lov_sub_find(env, subdev, ofid, subconf);
 271                         if (!IS_ERR(stripe)) {
 272                                 result = lov_init_sub(env, lov, stripe, r0, i);
 273                                 if (result == -EAGAIN) { /* try again */
 274                                         --i;
 275                                         result = 0;
 276                                         continue;
 277                                 }
 278                         } else {
 279                                 result = PTR_ERR(stripe);
 280                         }
 281
 282                         if (result == 0) {
 283                                 int sz = lov_page_slice_fixup(lov, stripe);
 284                                 LASSERT(ergo(psz > 0, psz == sz));
 285                                 psz = sz;
 286                         }
 287                 }
 288                 if (result == 0)
 289                         cl_object_header(&lov->lo_cl)->coh_page_bufsize += psz;
 290         } else
 291                 result = -ENOMEM;
 292 out:
 293         RETURN(result);
 294 }
 295
 296 static int lov_init_released(const struct lu_env *env,
 297                         struct lov_device *dev, struct lov_object *lov,
 298                         const struct cl_object_conf *conf,
 299                         union  lov_layout_state *state)
 300 {
 301         struct lov_stripe_md *lsm = conf->u.coc_md->lsm;
 302
 303         LASSERT(lsm != NULL);
 304         LASSERT(lsm_is_released(lsm));
 305         LASSERT(lov->lo_lsm == NULL);
 306
 307         lov->lo_lsm = lsm_addref(lsm);
 308         return 0;
 309 }
 310
 311 static int lov_delete_empty(const struct lu_env *env, struct lov_object *lov,
 312                             union lov_layout_state *state)
 313 {
 314         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 315
 316         lov_layout_wait(env, lov);
 317         return 0;
 318 }
 319
 320 static void lov_subobject_kill(const struct lu_env *env, struct lov_object *lov,
 321                                struct lovsub_object *los, int idx)
 322 {
 323         struct cl_object        *sub;
 324         struct lov_layout_raid0 *r0;
 325         struct lu_site          *site;
 326         struct lu_site_bkt_data *bkt;
 327         wait_queue_t          *waiter;
 328
 329         r0  = &lov->u.raid0;
 330         LASSERT(r0->lo_sub[idx] == los);
 331
 332         sub  = lovsub2cl(los);
 333         site = sub->co_lu.lo_dev->ld_site;
 334         bkt  = lu_site_bkt_from_fid(site, &sub->co_lu.lo_header->loh_fid);
 335
 336         cl_object_kill(env, sub);
 337         /* release a reference to the sub-object and ... */
 338         lu_object_ref_del(&sub->co_lu, "lov-parent", lov);
 339         cl_object_put(env, sub);
 340
 341         /* ... wait until it is actually destroyed---sub-object clears its
 342          * ->lo_sub[] slot in lovsub_object_fini() */
 343         if (r0->lo_sub[idx] == los) {
 344                 waiter = &lov_env_info(env)->lti_waiter;
 345                 init_waitqueue_entry_current(waiter);
 346                 add_wait_queue(&bkt->lsb_marche_funebre, waiter);
 347                 set_current_state(TASK_UNINTERRUPTIBLE);
 348                 while (1) {
 349                         /* this wait-queue is signaled at the end of
 350                          * lu_object_free(). */
 351                         set_current_state(TASK_UNINTERRUPTIBLE);
 352                         spin_lock(&r0->lo_sub_lock);
 353                         if (r0->lo_sub[idx] == los) {
 354                                 spin_unlock(&r0->lo_sub_lock);
 355                                 waitq_wait(waiter, TASK_UNINTERRUPTIBLE);
 356                         } else {
 357                                 spin_unlock(&r0->lo_sub_lock);
 358                                 set_current_state(TASK_RUNNING);
 359                                 break;
 360                         }
 361                 }
 362                 remove_wait_queue(&bkt->lsb_marche_funebre, waiter);
 363         }
 364         LASSERT(r0->lo_sub[idx] == NULL);
 365 }
 366
 367 static int lov_delete_raid0(const struct lu_env *env, struct lov_object *lov,
 368                             union lov_layout_state *state)
 369 {
 370         struct lov_layout_raid0 *r0 = &state->raid0;
 371         struct lov_stripe_md    *lsm = lov->lo_lsm;
 372         int i;
 373
 374         ENTRY;
 375
 376         dump_lsm(D_INODE, lsm);
 377
 378         lov_layout_wait(env, lov);
 379         if (r0->lo_sub != NULL) {
 380                 for (i = 0; i < r0->lo_nr; ++i) {
 381                         struct lovsub_object *los = r0->lo_sub[i];
 382
 383                         if (los != NULL) {
 384                                 cl_object_prune(env, &los->lso_cl);
 385                                 /*
 386                                  * If top-level object is to be evicted from
 387                                  * the cache, so are its sub-objects.
 388                                  */
 389                                 lov_subobject_kill(env, lov, los, i);
 390                         }
 391                 }
 392         }
 393         RETURN(0);
 394 }
 395
 396 static void lov_fini_empty(const struct lu_env *env, struct lov_object *lov,
 397                            union lov_layout_state *state)
 398 {
 399         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 400 }
 401
 402 static void lov_fini_raid0(const struct lu_env *env, struct lov_object *lov,
 403                            union lov_layout_state *state)
 404 {
 405         struct lov_layout_raid0 *r0 = &state->raid0;
 406         ENTRY;
 407
 408         if (r0->lo_sub != NULL) {
 409                 OBD_FREE_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 410                 r0->lo_sub = NULL;
 411         }
 412
 413         dump_lsm(D_INODE, lov->lo_lsm);
 414         lov_free_memmd(&lov->lo_lsm);
 415
 416         EXIT;
 417 }
 418
 419 static void lov_fini_released(const struct lu_env *env, struct lov_object *lov,
 420                                 union lov_layout_state *state)
 421 {
 422         ENTRY;
 423         dump_lsm(D_INODE, lov->lo_lsm);
 424         lov_free_memmd(&lov->lo_lsm);
 425         EXIT;
 426 }
 427
 428 static int lov_print_empty(const struct lu_env *env, void *cookie,
 429                            lu_printer_t p, const struct lu_object *o)
 430 {
 431         (*p)(env, cookie, "empty %d\n", lu2lov(o)->lo_layout_invalid);
 432         return 0;
 433 }
 434
 435 static int lov_print_raid0(const struct lu_env *env, void *cookie,
 436                            lu_printer_t p, const struct lu_object *o)
 437 {
 438         struct lov_object       *lov = lu2lov(o);
 439         struct lov_layout_raid0 *r0  = lov_r0(lov);
 440         struct lov_stripe_md    *lsm = lov->lo_lsm;
 441         int                      i;
 442
 443         (*p)(env, cookie, "stripes: %d, %s, lsm{%p 0x%08X %d %u %u}:\n",
 444                 r0->lo_nr, lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 445                 lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 446                 lsm->lsm_stripe_count, lsm->lsm_layout_gen);
 447         for (i = 0; i < r0->lo_nr; ++i) {
 448                 struct lu_object *sub;
 449
 450                 if (r0->lo_sub[i] != NULL) {
 451                         sub = lovsub2lu(r0->lo_sub[i]);
 452                         lu_object_print(env, cookie, p, sub);
 453                 } else {
 454                         (*p)(env, cookie, "sub %d absent\n", i);
 455                 }
 456         }
 457         return 0;
 458 }
 459
 460 static int lov_print_released(const struct lu_env *env, void *cookie,
 461                                 lu_printer_t p, const struct lu_object *o)
 462 {
 463         struct lov_object       *lov = lu2lov(o);
 464         struct lov_stripe_md    *lsm = lov->lo_lsm;
 465
 466         (*p)(env, cookie,
 467                 "released: %s, lsm{%p 0x%08X %d %u %u}:\n",
 468                 lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 469                 lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 470                 lsm->lsm_stripe_count, lsm->lsm_layout_gen);
 471         return 0;
 472 }
 473
 474 /**
 475  * Implements cl_object_operations::coo_attr_get() method for an object
 476  * without stripes (LLT_EMPTY layout type).
 477  *
 478  * The only attributes this layer is authoritative in this case is
 479  * cl_attr::cat_blocks---it's 0.
 480  */
 481 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
 482                               struct cl_attr *attr)
 483 {
 484         attr->cat_blocks = 0;
 485         return 0;
 486 }
 487
 488 static int lov_attr_get_raid0(const struct lu_env *env, struct cl_object *obj,
 489                               struct cl_attr *attr)
 490 {
 491         struct lov_object       *lov = cl2lov(obj);
 492         struct lov_layout_raid0 *r0 = lov_r0(lov);
 493         struct cl_attr          *lov_attr = &r0->lo_attr;
 494         int                      result = 0;
 495
 496         ENTRY;
 497
 498         /* this is called w/o holding type guard mutex, so it must be inside
 499          * an on going IO otherwise lsm may be replaced.
 500          * LU-2117: it turns out there exists one exception. For mmaped files,
 501          * the lock of those files may be requested in the other file's IO
 502          * context, and this function is called in ccc_lock_state(), it will
 503          * hit this assertion.
 504          * Anyway, it's still okay to call attr_get w/o type guard as layout
 505          * can't go if locks exist. */
 506         /* LASSERT(atomic_read(&lsm->lsm_refc) > 1); */
 507
 508         if (!r0->lo_attr_valid) {
 509                 struct lov_stripe_md    *lsm = lov->lo_lsm;
 510                 struct ost_lvb          *lvb = &lov_env_info(env)->lti_lvb;
 511                 __u64                    kms = 0;
 512
 513                 memset(lvb, 0, sizeof(*lvb));
 514                 /* XXX: timestamps can be negative by sanity:test_39m,
 515                  * how can it be? */
 516                 lvb->lvb_atime = LLONG_MIN;
 517                 lvb->lvb_ctime = LLONG_MIN;
 518                 lvb->lvb_mtime = LLONG_MIN;
 519
 520                 /*
 521                  * XXX that should be replaced with a loop over sub-objects,
 522                  * doing cl_object_attr_get() on them. But for now, let's
 523                  * reuse old lov code.
 524                  */
 525
 526                 /*
 527                  * XXX take lsm spin-lock to keep lov_merge_lvb_kms()
 528                  * happy. It's not needed, because new code uses
 529                  * ->coh_attr_guard spin-lock to protect consistency of
 530                  * sub-object attributes.
 531                  */
 532                 lov_stripe_lock(lsm);
 533                 result = lov_merge_lvb_kms(lsm, lvb, &kms);
 534                 lov_stripe_unlock(lsm);
 535                 if (result == 0) {
 536                         cl_lvb2attr(lov_attr, lvb);
 537                         lov_attr->cat_kms = kms;
 538                         r0->lo_attr_valid = 1;
 539                 }
 540         }
 541         if (result == 0) { /* merge results */
 542                 attr->cat_blocks = lov_attr->cat_blocks;
 543                 attr->cat_size = lov_attr->cat_size;
 544                 attr->cat_kms = lov_attr->cat_kms;
 545                 if (attr->cat_atime < lov_attr->cat_atime)
 546                         attr->cat_atime = lov_attr->cat_atime;
 547                 if (attr->cat_ctime < lov_attr->cat_ctime)
 548                         attr->cat_ctime = lov_attr->cat_ctime;
 549                 if (attr->cat_mtime < lov_attr->cat_mtime)
 550                         attr->cat_mtime = lov_attr->cat_mtime;
 551         }
 552         RETURN(result);
 553 }
 554
 555 const static struct lov_layout_operations lov_dispatch[] = {
 556         [LLT_EMPTY] = {
 557                 .llo_init      = lov_init_empty,
 558                 .llo_delete    = lov_delete_empty,
 559                 .llo_fini      = lov_fini_empty,
 560                 .llo_install   = lov_install_empty,
 561                 .llo_print     = lov_print_empty,
 562                 .llo_page_init = lov_page_init_empty,
 563                 .llo_lock_init = lov_lock_init_empty,
 564                 .llo_io_init   = lov_io_init_empty,
 565                 .llo_getattr   = lov_attr_get_empty
 566         },
 567         [LLT_RAID0] = {
 568                 .llo_init      = lov_init_raid0,
 569                 .llo_delete    = lov_delete_raid0,
 570                 .llo_fini      = lov_fini_raid0,
 571                 .llo_install   = lov_install_raid0,
 572                 .llo_print     = lov_print_raid0,
 573                 .llo_page_init = lov_page_init_raid0,
 574                 .llo_lock_init = lov_lock_init_raid0,
 575                 .llo_io_init   = lov_io_init_raid0,
 576                 .llo_getattr   = lov_attr_get_raid0
 577         },
 578         [LLT_RELEASED] = {
 579                 .llo_init      = lov_init_released,
 580                 .llo_delete    = lov_delete_empty,
 581                 .llo_fini      = lov_fini_released,
 582                 .llo_install   = lov_install_empty,
 583                 .llo_print     = lov_print_released,
 584                 .llo_page_init = lov_page_init_empty,
 585                 .llo_lock_init = lov_lock_init_empty,
 586                 .llo_io_init   = lov_io_init_released,
 587                 .llo_getattr   = lov_attr_get_empty
 588         }
 589 };
 590
 591 /**
 592  * Performs a double-dispatch based on the layout type of an object.
 593  */
 594 #define LOV_2DISPATCH_NOLOCK(obj, op, ...)                              \
 595 ({                                                                      \
 596         struct lov_object                      *__obj = (obj);          \
 597         enum lov_layout_type                    __llt;                  \
 598                                                                         \
 599         __llt = __obj->lo_type;                                         \
 600         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 601         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 602 })
 603
 604 /**
 605  * Return lov_layout_type associated with a given lsm
 606  */
 607 enum lov_layout_type lov_type(struct lov_stripe_md *lsm)
 608 {
 609         if (lsm == NULL)
 610                 return LLT_EMPTY;
 611         if (lsm_is_released(lsm))
 612                 return LLT_RELEASED;
 613         return LLT_RAID0;
 614 }
 615
 616 static inline void lov_conf_freeze(struct lov_object *lov)
 617 {
 618         if (lov->lo_owner != current)
 619                 down_read(&lov->lo_type_guard);
 620 }
 621
 622 static inline void lov_conf_thaw(struct lov_object *lov)
 623 {
 624         if (lov->lo_owner != current)
 625                 up_read(&lov->lo_type_guard);
 626 }
 627
 628 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...)                       \
 629 ({                                                                      \
 630         struct lov_object                      *__obj = (obj);          \
 631         int                                     __lock = !!(lock);      \
 632         typeof(lov_dispatch[0].op(__VA_ARGS__)) __result;               \
 633                                                                         \
 634         if (__lock)                                                     \
 635                 lov_conf_freeze(__obj);                                 \
 636         __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__);          \
 637         if (__lock)                                                     \
 638                 lov_conf_thaw(__obj);                                   \
 639         __result;                                                       \
 640 })
 641
 642 /**
 643  * Performs a locked double-dispatch based on the layout type of an object.
 644  */
 645 #define LOV_2DISPATCH(obj, op, ...)                     \
 646         LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
 647
 648 #define LOV_2DISPATCH_VOID(obj, op, ...)                                \
 649 do {                                                                    \
 650         struct lov_object                      *__obj = (obj);          \
 651         enum lov_layout_type                    __llt;                  \
 652                                                                         \
 653         lov_conf_freeze(__obj);                                         \
 654         __llt = __obj->lo_type;                                         \
 655         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 656         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 657         lov_conf_thaw(__obj);                                           \
 658 } while (0)
 659
 660 static void lov_conf_lock(struct lov_object *lov)
 661 {
 662         LASSERT(lov->lo_owner != current);
 663         down_write(&lov->lo_type_guard);
 664         LASSERT(lov->lo_owner == NULL);
 665         lov->lo_owner = current;
 666 }
 667
 668 static void lov_conf_unlock(struct lov_object *lov)
 669 {
 670         lov->lo_owner = NULL;
 671         up_write(&lov->lo_type_guard);
 672 }
 673
 674 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov)
 675 {
 676         struct l_wait_info lwi = { 0 };
 677         ENTRY;
 678
 679         while (atomic_read(&lov->lo_active_ios) > 0) {
 680                 CDEBUG(D_INODE, "file:"DFID" wait for active IO, now: %d.\n",
 681                         PFID(lu_object_fid(lov2lu(lov))),
 682                         atomic_read(&lov->lo_active_ios));
 683
 684                 l_wait_event(lov->lo_waitq,
 685                              atomic_read(&lov->lo_active_ios) == 0, &lwi);
 686         }
 687         RETURN(0);
 688 }
 689
 690 static int lov_layout_change(const struct lu_env *unused,
 691                              struct lov_object *lov,
 692                              const struct cl_object_conf *conf)
 693 {
 694         int result;
 695         enum lov_layout_type llt = LLT_EMPTY;
 696         union lov_layout_state *state = &lov->u;
 697         const struct lov_layout_operations *old_ops;
 698         const struct lov_layout_operations *new_ops;
 699
 700         void *cookie;
 701         struct lu_env *env;
 702         int refcheck;
 703         ENTRY;
 704
 705         LASSERT(0 <= lov->lo_type && lov->lo_type < ARRAY_SIZE(lov_dispatch));
 706
 707         if (conf->u.coc_md != NULL)
 708                 llt = lov_type(conf->u.coc_md->lsm);
 709         LASSERT(0 <= llt && llt < ARRAY_SIZE(lov_dispatch));
 710
 711         cookie = cl_env_reenter();
 712         env = cl_env_get(&refcheck);
 713         if (IS_ERR(env)) {
 714                 cl_env_reexit(cookie);
 715                 RETURN(PTR_ERR(env));
 716         }
 717
 718         CDEBUG(D_INODE, DFID" from %s to %s\n",
 719                PFID(lu_object_fid(lov2lu(lov))),
 720                llt2str(lov->lo_type), llt2str(llt));
 721
 722         old_ops = &lov_dispatch[lov->lo_type];
 723         new_ops = &lov_dispatch[llt];
 724
 725         result = cl_object_prune(env, &lov->lo_cl);
 726         if (result != 0)
 727                 GOTO(out, result);
 728
 729         result = old_ops->llo_delete(env, lov, &lov->u);
 730         if (result == 0) {
 731                 old_ops->llo_fini(env, lov, &lov->u);
 732
 733                 LASSERT(atomic_read(&lov->lo_active_ios) == 0);
 734
 735                 lov->lo_type = LLT_EMPTY;
 736                 /* page bufsize fixup */
 737                 cl_object_header(&lov->lo_cl)->coh_page_bufsize -=
 738                         lov_page_slice_fixup(lov, NULL);
 739
 740                 result = new_ops->llo_init(env,
 741                                         lu2lov_dev(lov->lo_cl.co_lu.lo_dev),
 742                                         lov, conf, state);
 743                 if (result == 0) {
 744                         new_ops->llo_install(env, lov, state);
 745                         lov->lo_type = llt;
 746                 } else {
 747                         new_ops->llo_delete(env, lov, state);
 748                         new_ops->llo_fini(env, lov, state);
 749                         /* this file becomes an EMPTY file. */
 750                 }
 751         }
 752
 753 out:
 754         cl_env_put(env, &refcheck);
 755         cl_env_reexit(cookie);
 756         RETURN(result);
 757 }
 758
 759 /*****************************************************************************
 760  *
 761  * Lov object operations.
 762  *
 763  */
 764 int lov_object_init(const struct lu_env *env, struct lu_object *obj,
 765                     const struct lu_object_conf *conf)
 766 {
 767         struct lov_device            *dev   = lu2lov_dev(obj->lo_dev);
 768         struct lov_object            *lov   = lu2lov(obj);
 769         const struct cl_object_conf  *cconf = lu2cl_conf(conf);
 770         union  lov_layout_state      *set   = &lov->u;
 771         const struct lov_layout_operations *ops;
 772         int result;
 773
 774         ENTRY;
 775         init_rwsem(&lov->lo_type_guard);
 776         atomic_set(&lov->lo_active_ios, 0);
 777         init_waitqueue_head(&lov->lo_waitq);
 778
 779         cl_object_page_init(lu2cl(obj), sizeof(struct lov_page));
 780
 781         /* no locking is necessary, as object is being created */
 782         lov->lo_type = lov_type(cconf->u.coc_md->lsm);
 783         ops = &lov_dispatch[lov->lo_type];
 784         result = ops->llo_init(env, dev, lov, cconf, set);
 785         if (result == 0)
 786                 ops->llo_install(env, lov, set);
 787         RETURN(result);
 788 }
 789
 790 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
 791                         const struct cl_object_conf *conf)
 792 {
 793         struct lov_stripe_md    *lsm = NULL;
 794         struct lov_object       *lov = cl2lov(obj);
 795         int                      result = 0;
 796         ENTRY;
 797
 798         lov_conf_lock(lov);
 799         if (conf->coc_opc == OBJECT_CONF_INVALIDATE) {
 800                 lov->lo_layout_invalid = true;
 801                 GOTO(out, result = 0);
 802         }
 803
 804         if (conf->coc_opc == OBJECT_CONF_WAIT) {
 805                 if (lov->lo_layout_invalid &&
 806                     atomic_read(&lov->lo_active_ios) > 0) {
 807                         lov_conf_unlock(lov);
 808                         result = lov_layout_wait(env, lov);
 809                         lov_conf_lock(lov);
 810                 }
 811                 GOTO(out, result);
 812         }
 813
 814         LASSERT(conf->coc_opc == OBJECT_CONF_SET);
 815
 816         if (conf->u.coc_md != NULL)
 817                 lsm = conf->u.coc_md->lsm;
 818         if ((lsm == NULL && lov->lo_lsm == NULL) ||
 819             ((lsm != NULL && lov->lo_lsm != NULL) &&
 820              (lov->lo_lsm->lsm_layout_gen == lsm->lsm_layout_gen) &&
 821              (lov->lo_lsm->lsm_pattern == lsm->lsm_pattern))) {
 822                 /* same version of layout */
 823                 lov->lo_layout_invalid = false;
 824                 GOTO(out, result = 0);
 825         }
 826
 827         /* will change layout - check if there still exists active IO. */
 828         if (atomic_read(&lov->lo_active_ios) > 0) {
 829                 lov->lo_layout_invalid = true;
 830                 GOTO(out, result = -EBUSY);
 831         }
 832
 833         result = lov_layout_change(env, lov, conf);
 834         lov->lo_layout_invalid = result != 0;
 835         EXIT;
 836
 837 out:
 838         lov_conf_unlock(lov);
 839         CDEBUG(D_INODE, DFID" lo_layout_invalid=%d\n",
 840                PFID(lu_object_fid(lov2lu(lov))), lov->lo_layout_invalid);
 841         RETURN(result);
 842 }
 843
 844 static void lov_object_delete(const struct lu_env *env, struct lu_object *obj)
 845 {
 846         struct lov_object *lov = lu2lov(obj);
 847
 848         ENTRY;
 849         LOV_2DISPATCH_VOID(lov, llo_delete, env, lov, &lov->u);
 850         EXIT;
 851 }
 852
 853 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
 854 {
 855         struct lov_object *lov = lu2lov(obj);
 856
 857         ENTRY;
 858         LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
 859         lu_object_fini(obj);
 860         OBD_SLAB_FREE_PTR(lov, lov_object_kmem);
 861         EXIT;
 862 }
 863
 864 static int lov_object_print(const struct lu_env *env, void *cookie,
 865                             lu_printer_t p, const struct lu_object *o)
 866 {
 867         return LOV_2DISPATCH_NOLOCK(lu2lov(o), llo_print, env, cookie, p, o);
 868 }
 869
 870 int lov_page_init(const struct lu_env *env, struct cl_object *obj,
 871                   struct cl_page *page, pgoff_t index)
 872 {
 873         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_page_init, env, obj, page,
 874                                     index);
 875 }
 876
 877 /**
 878  * Implements cl_object_operations::clo_io_init() method for lov
 879  * layer. Dispatches to the appropriate layout io initialization method.
 880  */
 881 int lov_io_init(const struct lu_env *env, struct cl_object *obj,
 882                 struct cl_io *io)
 883 {
 884         CL_IO_SLICE_CLEAN(lov_env_io(env), lis_cl);
 885         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
 886                                      !io->ci_ignore_layout, env, obj, io);
 887 }
 888
 889 /**
 890  * An implementation of cl_object_operations::clo_attr_get() method for lov
 891  * layer. For raid0 layout this collects and merges attributes of all
 892  * sub-objects.
 893  */
 894 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
 895                         struct cl_attr *attr)
 896 {
 897         /* do not take lock, as this function is called under a
 898          * spin-lock. Layout is protected from changing by ongoing IO. */
 899         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
 900 }
 901
 902 static int lov_attr_set(const struct lu_env *env, struct cl_object *obj,
 903                         const struct cl_attr *attr, unsigned valid)
 904 {
 905         /*
 906          * No dispatch is required here, as no layout implements this.
 907          */
 908         return 0;
 909 }
 910
 911 int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
 912                   struct cl_lock *lock, const struct cl_io *io)
 913 {
 914         /* No need to lock because we've taken one refcount of layout.  */
 915         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_lock_init, env, obj, lock,
 916                                     io);
 917 }
 918
 919 static const struct cl_object_operations lov_ops = {
 920         .coo_page_init = lov_page_init,
 921         .coo_lock_init = lov_lock_init,
 922         .coo_io_init   = lov_io_init,
 923         .coo_attr_get  = lov_attr_get,
 924         .coo_attr_set  = lov_attr_set,
 925         .coo_conf_set  = lov_conf_set
 926 };
 927
 928 static const struct lu_object_operations lov_lu_obj_ops = {
 929         .loo_object_init      = lov_object_init,
 930         .loo_object_delete    = lov_object_delete,
 931         .loo_object_release   = NULL,
 932         .loo_object_free      = lov_object_free,
 933         .loo_object_print     = lov_object_print,
 934         .loo_object_invariant = NULL
 935 };
 936
 937 struct lu_object *lov_object_alloc(const struct lu_env *env,
 938                                    const struct lu_object_header *unused,
 939                                    struct lu_device *dev)
 940 {
 941         struct lov_object *lov;
 942         struct lu_object  *obj;
 943
 944         ENTRY;
 945         OBD_SLAB_ALLOC_PTR_GFP(lov, lov_object_kmem, GFP_NOFS);
 946         if (lov != NULL) {
 947                 obj = lov2lu(lov);
 948                 lu_object_init(obj, NULL, dev);
 949                 lov->lo_cl.co_ops = &lov_ops;
 950                 lov->lo_type = -1; /* invalid, to catch uninitialized type */
 951                 /*
 952                  * object io operation vector (cl_object::co_iop) is installed
 953                  * later in lov_object_init(), as different vectors are used
 954                  * for object with different layouts.
 955                  */
 956                 obj->lo_ops = &lov_lu_obj_ops;
 957         } else
 958                 obj = NULL;
 959         RETURN(obj);
 960 }
 961
 962 struct lov_stripe_md *lov_lsm_addref(struct lov_object *lov)
 963 {
 964         struct lov_stripe_md *lsm = NULL;
 965
 966         lov_conf_freeze(lov);
 967         if (lov->lo_lsm != NULL) {
 968                 lsm = lsm_addref(lov->lo_lsm);
 969                 CDEBUG(D_INODE, "lsm %p addref %d/%d by %p.\n",
 970                         lsm, atomic_read(&lsm->lsm_refc),
 971                         lov->lo_layout_invalid, current);
 972         }
 973         lov_conf_thaw(lov);
 974         return lsm;
 975 }
 976
 977 struct lov_stripe_md *lov_lsm_get(struct cl_object *clobj)
 978 {
 979         struct lu_object *luobj;
 980         struct lov_stripe_md *lsm = NULL;
 981
 982         if (clobj == NULL)
 983                 return NULL;
 984
 985         luobj = lu_object_locate(&cl_object_header(clobj)->coh_lu,
 986                                  &lov_device_type);
 987         if (luobj != NULL)
 988                 lsm = lov_lsm_addref(lu2lov(luobj));
 989         return lsm;
 990 }
 991 EXPORT_SYMBOL(lov_lsm_get);
 992
 993 void lov_lsm_put(struct cl_object *unused, struct lov_stripe_md *lsm)
 994 {
 995         if (lsm != NULL)
 996                 lov_free_memmd(&lsm);
 997 }
 998 EXPORT_SYMBOL(lov_lsm_put);
 999
1000 int lov_read_and_clear_async_rc(struct cl_object *clob)
1001 {
1002         struct lu_object *luobj;
1003         int rc = 0;
1004         ENTRY;
1005
1006         luobj = lu_object_locate(&cl_object_header(clob)->coh_lu,
1007                                  &lov_device_type);
1008         if (luobj != NULL) {
1009                 struct lov_object *lov = lu2lov(luobj);
1010
1011                 lov_conf_freeze(lov);
1012                 switch (lov->lo_type) {
1013                 case LLT_RAID0: {
1014                         struct lov_stripe_md *lsm;
1015                         int i;
1016
1017                         lsm = lov->lo_lsm;
1018                         LASSERT(lsm != NULL);
1019                         for (i = 0; i < lsm->lsm_stripe_count; i++) {
1020                                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
1021
1022                                 if (lov_oinfo_is_dummy(loi))
1023                                         continue;
1024
1025                                 if (loi->loi_ar.ar_rc && !rc)
1026                                         rc = loi->loi_ar.ar_rc;
1027                                 loi->loi_ar.ar_rc = 0;
1028                         }
1029                 }
1030                 case LLT_RELEASED:
1031                 case LLT_EMPTY:
1032                         break;
1033                 default:
1034                         LBUG();
1035                 }
1036                 lov_conf_thaw(lov);
1037         }
1038         RETURN(rc);
1039 }
1040 EXPORT_SYMBOL(lov_read_and_clear_async_rc);
1041
1042 /** @} lov */