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