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