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 #include "lclient.h"
  46
  47 /** \addtogroup lov
  48  *  @{
  49  */
  50
  51 /*****************************************************************************
  52  *
  53  * Layout operations.
  54  *
  55  */
  56
  57 struct lov_layout_operations {
  58         int (*llo_init)(const struct lu_env *env, struct lov_device *dev,
  59                         struct lov_object *lov,
  60                         const struct cl_object_conf *conf,
  61                         union lov_layout_state *state);
  62         int (*llo_delete)(const struct lu_env *env, struct lov_object *lov,
  63                            union lov_layout_state *state);
  64         void (*llo_fini)(const struct lu_env *env, struct lov_object *lov,
  65                          union lov_layout_state *state);
  66         void (*llo_install)(const struct lu_env *env, struct lov_object *lov,
  67                             union lov_layout_state *state);
  68         int  (*llo_print)(const struct lu_env *env, void *cookie,
  69                           lu_printer_t p, const struct lu_object *o);
  70         int  (*llo_page_init)(const struct lu_env *env, struct cl_object *obj,
  71                               struct cl_page *page, pgoff_t index);
  72         int  (*llo_lock_init)(const struct lu_env *env,
  73                               struct cl_object *obj, struct cl_lock *lock,
  74                               const struct cl_io *io);
  75         int  (*llo_io_init)(const struct lu_env *env,
  76                             struct cl_object *obj, struct cl_io *io);
  77         int  (*llo_getattr)(const struct lu_env *env, struct cl_object *obj,
  78                             struct cl_attr *attr);
  79 };
  80
  81 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov);
  82
  83 struct lov_stripe_md *lov_lsm_get(struct cl_object *clobj)
  84 {
  85         struct lu_object *luobj;
  86         struct lov_stripe_md *lsm = NULL;
  87
  88         if (clobj == NULL)
  89                 return NULL;
  90
  91         luobj = lu_object_locate(&cl_object_header(clobj)->coh_lu,
  92                                  &lov_device_type);
  93         if (luobj != NULL)
  94                 lsm = lov_lsm_addref(lu2lov(luobj));
  95         return lsm;
  96 }
  97 EXPORT_SYMBOL(lov_lsm_get);
  98
  99 void lov_lsm_put(struct cl_object *unused, struct lov_stripe_md *lsm)
 100 {
 101         if (lsm != NULL)
 102                 lov_free_memmd(&lsm);
 103 }
 104 EXPORT_SYMBOL(lov_lsm_put);
 105
 106 /*****************************************************************************
 107  *
 108  * Lov object layout operations.
 109  *
 110  */
 111
 112 static void lov_install_empty(const struct lu_env *env,
 113                               struct lov_object *lov,
 114                               union  lov_layout_state *state)
 115 {
 116         /*
 117          * File without objects.
 118          */
 119 }
 120
 121 static int lov_init_empty(const struct lu_env *env,
 122                           struct lov_device *dev, struct lov_object *lov,
 123                           const struct cl_object_conf *conf,
 124                           union  lov_layout_state *state)
 125 {
 126         return 0;
 127 }
 128
 129 static void lov_install_raid0(const struct lu_env *env,
 130                               struct lov_object *lov,
 131                               union  lov_layout_state *state)
 132 {
 133 }
 134
 135 static struct cl_object *lov_sub_find(const struct lu_env *env,
 136                                       struct cl_device *dev,
 137                                       const struct lu_fid *fid,
 138                                       const struct cl_object_conf *conf)
 139 {
 140         struct lu_object *o;
 141
 142         ENTRY;
 143         o = lu_object_find_at(env, cl2lu_dev(dev), fid, &conf->coc_lu);
 144         LASSERT(ergo(!IS_ERR(o), o->lo_dev->ld_type == &lovsub_device_type));
 145         RETURN(lu2cl(o));
 146 }
 147
 148 static int lov_init_sub(const struct lu_env *env, struct lov_object *lov,
 149                         struct cl_object *stripe, struct lov_layout_raid0 *r0,
 150                         int idx)
 151 {
 152         struct cl_object_header *hdr;
 153         struct cl_object_header *subhdr;
 154         struct cl_object_header *parent;
 155         struct lov_oinfo        *oinfo;
 156         int result;
 157
 158         if (OBD_FAIL_CHECK(OBD_FAIL_LOV_INIT)) {
 159                 /* For sanity:test_206.
 160                  * Do not leave the object in cache to avoid accessing
 161                  * freed memory. This is because osc_object is referring to
 162                  * lov_oinfo of lsm_stripe_data which will be freed due to
 163                  * this failure. */
 164                 cl_object_kill(env, stripe);
 165                 cl_object_put(env, stripe);
 166                 return -EIO;
 167         }
 168
 169         hdr    = cl_object_header(lov2cl(lov));
 170         subhdr = cl_object_header(stripe);
 171
 172         oinfo = lov->lo_lsm->lsm_oinfo[idx];
 173         CDEBUG(D_INODE, DFID"@%p[%d] -> "DFID"@%p: ostid: "DOSTID
 174                " idx: %d gen: %d\n",
 175                PFID(&subhdr->coh_lu.loh_fid), subhdr, idx,
 176                PFID(&hdr->coh_lu.loh_fid), hdr, POSTID(&oinfo->loi_oi),
 177                oinfo->loi_ost_idx, oinfo->loi_ost_gen);
 178
 179         /* reuse ->coh_attr_guard to protect coh_parent change */
 180         spin_lock(&subhdr->coh_attr_guard);
 181         parent = subhdr->coh_parent;
 182         if (parent == NULL) {
 183                 subhdr->coh_parent = hdr;
 184                 spin_unlock(&subhdr->coh_attr_guard);
 185                 subhdr->coh_nesting = hdr->coh_nesting + 1;
 186                 lu_object_ref_add(&stripe->co_lu, "lov-parent", lov);
 187                 r0->lo_sub[idx] = cl2lovsub(stripe);
 188                 r0->lo_sub[idx]->lso_super = lov;
 189                 r0->lo_sub[idx]->lso_index = idx;
 190                 result = 0;
 191         } else {
 192                 struct lu_object  *old_obj;
 193                 struct lov_object *old_lov;
 194                 unsigned int mask = D_INODE;
 195
 196                 spin_unlock(&subhdr->coh_attr_guard);
 197                 old_obj = lu_object_locate(&parent->coh_lu, &lov_device_type);
 198                 LASSERT(old_obj != NULL);
 199                 old_lov = cl2lov(lu2cl(old_obj));
 200                 if (old_lov->lo_layout_invalid) {
 201                         /* the object's layout has already changed but isn't
 202                          * refreshed */
 203                         lu_object_unhash(env, &stripe->co_lu);
 204                         result = -EAGAIN;
 205                 } else {
 206                         mask = D_ERROR;
 207                         result = -EIO;
 208                 }
 209
 210                 LU_OBJECT_DEBUG(mask, env, &stripe->co_lu,
 211                                 "stripe %d is already owned.\n", idx);
 212                 LU_OBJECT_DEBUG(mask, env, old_obj, "owned.\n");
 213                 LU_OBJECT_HEADER(mask, env, lov2lu(lov), "try to own.\n");
 214                 cl_object_put(env, stripe);
 215         }
 216         return result;
 217 }
 218
 219 static int lov_page_slice_fixup(struct lov_object *lov,
 220                                 struct cl_object *stripe)
 221 {
 222         struct cl_object_header *hdr = cl_object_header(&lov->lo_cl);
 223         struct cl_object *o;
 224
 225         if (stripe == NULL)
 226                 return hdr->coh_page_bufsize - lov->lo_cl.co_slice_off -
 227                        cfs_size_round(sizeof(struct lov_page));
 228
 229         cl_object_for_each(o, stripe)
 230                 o->co_slice_off += hdr->coh_page_bufsize;
 231
 232         return cl_object_header(stripe)->coh_page_bufsize;
 233 }
 234
 235 static int lov_init_raid0(const struct lu_env *env,
 236                           struct lov_device *dev, struct lov_object *lov,
 237                           const struct cl_object_conf *conf,
 238                           union  lov_layout_state *state)
 239 {
 240         int result;
 241         int i;
 242
 243         struct cl_object        *stripe;
 244         struct lov_thread_info  *lti     = lov_env_info(env);
 245         struct cl_object_conf   *subconf = &lti->lti_stripe_conf;
 246         struct lov_stripe_md    *lsm     = conf->u.coc_md->lsm;
 247         struct lu_fid           *ofid    = &lti->lti_fid;
 248         struct lov_layout_raid0 *r0      = &state->raid0;
 249
 250         ENTRY;
 251
 252         if (lsm->lsm_magic != LOV_MAGIC_V1 && lsm->lsm_magic != LOV_MAGIC_V3) {
 253                 dump_lsm(D_ERROR, lsm);
 254                 LASSERTF(0, "magic mismatch, expected %d/%d, actual %d.\n",
 255                          LOV_MAGIC_V1, LOV_MAGIC_V3, lsm->lsm_magic);
 256         }
 257
 258         LASSERT(lov->lo_lsm == NULL);
 259         lov->lo_lsm = lsm_addref(lsm);
 260         r0->lo_nr  = lsm->lsm_stripe_count;
 261         LASSERT(r0->lo_nr <= lov_targets_nr(dev));
 262
 263         lov->lo_layout_invalid = true;
 264
 265         OBD_ALLOC_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 266         if (r0->lo_sub != NULL) {
 267                 int psz = 0;
 268
 269                 result = 0;
 270                 subconf->coc_inode = conf->coc_inode;
 271                 spin_lock_init(&r0->lo_sub_lock);
 272                 /*
 273                  * Create stripe cl_objects.
 274                  */
 275                 for (i = 0; i < r0->lo_nr && result == 0; ++i) {
 276                         struct cl_device *subdev;
 277                         struct lov_oinfo *oinfo = lsm->lsm_oinfo[i];
 278                         int ost_idx = oinfo->loi_ost_idx;
 279
 280                         if (lov_oinfo_is_dummy(oinfo))
 281                                 continue;
 282
 283                         result = ostid_to_fid(ofid, &oinfo->loi_oi,
 284                                               oinfo->loi_ost_idx);
 285                         if (result != 0)
 286                                 GOTO(out, result);
 287
 288                         subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
 289                         subconf->u.coc_oinfo = oinfo;
 290                         LASSERTF(subdev != NULL, "not init ost %d\n", ost_idx);
 291                         /* In the function below, .hs_keycmp resolves to
 292                          * lu_obj_hop_keycmp() */
 293                         /* coverity[overrun-buffer-val] */
 294                         stripe = lov_sub_find(env, subdev, ofid, subconf);
 295                         if (!IS_ERR(stripe)) {
 296                                 result = lov_init_sub(env, lov, stripe, r0, i);
 297                                 if (result == -EAGAIN) { /* try again */
 298                                         --i;
 299                                         result = 0;
 300                                         continue;
 301                                 }
 302                         } else {
 303                                 result = PTR_ERR(stripe);
 304                         }
 305
 306                         if (result == 0) {
 307                                 int sz = lov_page_slice_fixup(lov, stripe);
 308                                 LASSERT(ergo(psz > 0, psz == sz));
 309                                 psz = sz;
 310                         }
 311                 }
 312                 if (result == 0)
 313                         cl_object_header(&lov->lo_cl)->coh_page_bufsize += psz;
 314         } else
 315                 result = -ENOMEM;
 316 out:
 317         RETURN(result);
 318 }
 319
 320 static int lov_init_released(const struct lu_env *env,
 321                         struct lov_device *dev, struct lov_object *lov,
 322                         const struct cl_object_conf *conf,
 323                         union  lov_layout_state *state)
 324 {
 325         struct lov_stripe_md *lsm = conf->u.coc_md->lsm;
 326
 327         LASSERT(lsm != NULL);
 328         LASSERT(lsm_is_released(lsm));
 329         LASSERT(lov->lo_lsm == NULL);
 330
 331         lov->lo_lsm = lsm_addref(lsm);
 332         return 0;
 333 }
 334
 335 static int lov_delete_empty(const struct lu_env *env, struct lov_object *lov,
 336                             union lov_layout_state *state)
 337 {
 338         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 339
 340         lov_layout_wait(env, lov);
 341         return 0;
 342 }
 343
 344 static void lov_subobject_kill(const struct lu_env *env, struct lov_object *lov,
 345                                struct lovsub_object *los, int idx)
 346 {
 347         struct cl_object        *sub;
 348         struct lov_layout_raid0 *r0;
 349         struct lu_site          *site;
 350         struct lu_site_bkt_data *bkt;
 351         wait_queue_t          *waiter;
 352
 353         r0  = &lov->u.raid0;
 354         LASSERT(r0->lo_sub[idx] == los);
 355
 356         sub  = lovsub2cl(los);
 357         site = sub->co_lu.lo_dev->ld_site;
 358         bkt  = lu_site_bkt_from_fid(site, &sub->co_lu.lo_header->loh_fid);
 359
 360         cl_object_kill(env, sub);
 361         /* release a reference to the sub-object and ... */
 362         lu_object_ref_del(&sub->co_lu, "lov-parent", lov);
 363         cl_object_put(env, sub);
 364
 365         /* ... wait until it is actually destroyed---sub-object clears its
 366          * ->lo_sub[] slot in lovsub_object_fini() */
 367         if (r0->lo_sub[idx] == los) {
 368                 waiter = &lov_env_info(env)->lti_waiter;
 369                 init_waitqueue_entry_current(waiter);
 370                 add_wait_queue(&bkt->lsb_marche_funebre, waiter);
 371                 set_current_state(TASK_UNINTERRUPTIBLE);
 372                 while (1) {
 373                         /* this wait-queue is signaled at the end of
 374                          * lu_object_free(). */
 375                         set_current_state(TASK_UNINTERRUPTIBLE);
 376                         spin_lock(&r0->lo_sub_lock);
 377                         if (r0->lo_sub[idx] == los) {
 378                                 spin_unlock(&r0->lo_sub_lock);
 379                                 waitq_wait(waiter, TASK_UNINTERRUPTIBLE);
 380                         } else {
 381                                 spin_unlock(&r0->lo_sub_lock);
 382                                 set_current_state(TASK_RUNNING);
 383                                 break;
 384                         }
 385                 }
 386                 remove_wait_queue(&bkt->lsb_marche_funebre, waiter);
 387         }
 388         LASSERT(r0->lo_sub[idx] == NULL);
 389 }
 390
 391 static int lov_delete_raid0(const struct lu_env *env, struct lov_object *lov,
 392                             union lov_layout_state *state)
 393 {
 394         struct lov_layout_raid0 *r0 = &state->raid0;
 395         struct lov_stripe_md    *lsm = lov->lo_lsm;
 396         int i;
 397
 398         ENTRY;
 399
 400         dump_lsm(D_INODE, lsm);
 401
 402         lov_layout_wait(env, lov);
 403         if (r0->lo_sub != NULL) {
 404                 for (i = 0; i < r0->lo_nr; ++i) {
 405                         struct lovsub_object *los = r0->lo_sub[i];
 406
 407                         if (los != NULL) {
 408                                 cl_object_prune(env, &los->lso_cl);
 409                                 /*
 410                                  * If top-level object is to be evicted from
 411                                  * the cache, so are its sub-objects.
 412                                  */
 413                                 lov_subobject_kill(env, lov, los, i);
 414                         }
 415                 }
 416         }
 417         RETURN(0);
 418 }
 419
 420 static void lov_fini_empty(const struct lu_env *env, struct lov_object *lov,
 421                            union lov_layout_state *state)
 422 {
 423         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 424 }
 425
 426 static void lov_fini_raid0(const struct lu_env *env, struct lov_object *lov,
 427                            union lov_layout_state *state)
 428 {
 429         struct lov_layout_raid0 *r0 = &state->raid0;
 430         ENTRY;
 431
 432         if (r0->lo_sub != NULL) {
 433                 OBD_FREE_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 434                 r0->lo_sub = NULL;
 435         }
 436
 437         dump_lsm(D_INODE, lov->lo_lsm);
 438         lov_free_memmd(&lov->lo_lsm);
 439
 440         EXIT;
 441 }
 442
 443 static void lov_fini_released(const struct lu_env *env, struct lov_object *lov,
 444                                 union lov_layout_state *state)
 445 {
 446         ENTRY;
 447         dump_lsm(D_INODE, lov->lo_lsm);
 448         lov_free_memmd(&lov->lo_lsm);
 449         EXIT;
 450 }
 451
 452 static int lov_print_empty(const struct lu_env *env, void *cookie,
 453                            lu_printer_t p, const struct lu_object *o)
 454 {
 455         (*p)(env, cookie, "empty %d\n", lu2lov(o)->lo_layout_invalid);
 456         return 0;
 457 }
 458
 459 static int lov_print_raid0(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_layout_raid0 *r0  = lov_r0(lov);
 464         struct lov_stripe_md    *lsm = lov->lo_lsm;
 465         int                      i;
 466
 467         (*p)(env, cookie, "stripes: %d, %s, lsm{%p 0x%08X %d %u %u}:\n",
 468                 r0->lo_nr, 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         for (i = 0; i < r0->lo_nr; ++i) {
 472                 struct lu_object *sub;
 473
 474                 if (r0->lo_sub[i] != NULL) {
 475                         sub = lovsub2lu(r0->lo_sub[i]);
 476                         lu_object_print(env, cookie, p, sub);
 477                 } else {
 478                         (*p)(env, cookie, "sub %d absent\n", i);
 479                 }
 480         }
 481         return 0;
 482 }
 483
 484 static int lov_print_released(const struct lu_env *env, void *cookie,
 485                                 lu_printer_t p, const struct lu_object *o)
 486 {
 487         struct lov_object       *lov = lu2lov(o);
 488         struct lov_stripe_md    *lsm = lov->lo_lsm;
 489
 490         (*p)(env, cookie,
 491                 "released: %s, lsm{%p 0x%08X %d %u %u}:\n",
 492                 lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 493                 lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 494                 lsm->lsm_stripe_count, lsm->lsm_layout_gen);
 495         return 0;
 496 }
 497
 498 /**
 499  * Implements cl_object_operations::coo_attr_get() method for an object
 500  * without stripes (LLT_EMPTY layout type).
 501  *
 502  * The only attributes this layer is authoritative in this case is
 503  * cl_attr::cat_blocks---it's 0.
 504  */
 505 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
 506                               struct cl_attr *attr)
 507 {
 508         attr->cat_blocks = 0;
 509         return 0;
 510 }
 511
 512 static int lov_attr_get_raid0(const struct lu_env *env, struct cl_object *obj,
 513                               struct cl_attr *attr)
 514 {
 515         struct lov_object       *lov = cl2lov(obj);
 516         struct lov_layout_raid0 *r0 = lov_r0(lov);
 517         struct cl_attr          *lov_attr = &r0->lo_attr;
 518         int                      result = 0;
 519
 520         ENTRY;
 521
 522         /* this is called w/o holding type guard mutex, so it must be inside
 523          * an on going IO otherwise lsm may be replaced.
 524          * LU-2117: it turns out there exists one exception. For mmaped files,
 525          * the lock of those files may be requested in the other file's IO
 526          * context, and this function is called in ccc_lock_state(), it will
 527          * hit this assertion.
 528          * Anyway, it's still okay to call attr_get w/o type guard as layout
 529          * can't go if locks exist. */
 530         /* LASSERT(atomic_read(&lsm->lsm_refc) > 1); */
 531
 532         if (!r0->lo_attr_valid) {
 533                 struct lov_stripe_md    *lsm = lov->lo_lsm;
 534                 struct ost_lvb          *lvb = &lov_env_info(env)->lti_lvb;
 535                 __u64                    kms = 0;
 536
 537                 memset(lvb, 0, sizeof(*lvb));
 538                 /* XXX: timestamps can be negative by sanity:test_39m,
 539                  * how can it be? */
 540                 lvb->lvb_atime = LLONG_MIN;
 541                 lvb->lvb_ctime = LLONG_MIN;
 542                 lvb->lvb_mtime = LLONG_MIN;
 543
 544                 /*
 545                  * XXX that should be replaced with a loop over sub-objects,
 546                  * doing cl_object_attr_get() on them. But for now, let's
 547                  * reuse old lov code.
 548                  */
 549
 550                 /*
 551                  * XXX take lsm spin-lock to keep lov_merge_lvb_kms()
 552                  * happy. It's not needed, because new code uses
 553                  * ->coh_attr_guard spin-lock to protect consistency of
 554                  * sub-object attributes.
 555                  */
 556                 lov_stripe_lock(lsm);
 557                 result = lov_merge_lvb_kms(lsm, lvb, &kms);
 558                 lov_stripe_unlock(lsm);
 559                 if (result == 0) {
 560                         cl_lvb2attr(lov_attr, lvb);
 561                         lov_attr->cat_kms = kms;
 562                         r0->lo_attr_valid = 1;
 563                 }
 564         }
 565         if (result == 0) { /* merge results */
 566                 attr->cat_blocks = lov_attr->cat_blocks;
 567                 attr->cat_size = lov_attr->cat_size;
 568                 attr->cat_kms = lov_attr->cat_kms;
 569                 if (attr->cat_atime < lov_attr->cat_atime)
 570                         attr->cat_atime = lov_attr->cat_atime;
 571                 if (attr->cat_ctime < lov_attr->cat_ctime)
 572                         attr->cat_ctime = lov_attr->cat_ctime;
 573                 if (attr->cat_mtime < lov_attr->cat_mtime)
 574                         attr->cat_mtime = lov_attr->cat_mtime;
 575         }
 576         RETURN(result);
 577 }
 578
 579 const static struct lov_layout_operations lov_dispatch[] = {
 580         [LLT_EMPTY] = {
 581                 .llo_init      = lov_init_empty,
 582                 .llo_delete    = lov_delete_empty,
 583                 .llo_fini      = lov_fini_empty,
 584                 .llo_install   = lov_install_empty,
 585                 .llo_print     = lov_print_empty,
 586                 .llo_page_init = lov_page_init_empty,
 587                 .llo_lock_init = lov_lock_init_empty,
 588                 .llo_io_init   = lov_io_init_empty,
 589                 .llo_getattr   = lov_attr_get_empty
 590         },
 591         [LLT_RAID0] = {
 592                 .llo_init      = lov_init_raid0,
 593                 .llo_delete    = lov_delete_raid0,
 594                 .llo_fini      = lov_fini_raid0,
 595                 .llo_install   = lov_install_raid0,
 596                 .llo_print     = lov_print_raid0,
 597                 .llo_page_init = lov_page_init_raid0,
 598                 .llo_lock_init = lov_lock_init_raid0,
 599                 .llo_io_init   = lov_io_init_raid0,
 600                 .llo_getattr   = lov_attr_get_raid0
 601         },
 602         [LLT_RELEASED] = {
 603                 .llo_init      = lov_init_released,
 604                 .llo_delete    = lov_delete_empty,
 605                 .llo_fini      = lov_fini_released,
 606                 .llo_install   = lov_install_empty,
 607                 .llo_print     = lov_print_released,
 608                 .llo_page_init = lov_page_init_empty,
 609                 .llo_lock_init = lov_lock_init_empty,
 610                 .llo_io_init   = lov_io_init_released,
 611                 .llo_getattr   = lov_attr_get_empty
 612         }
 613 };
 614
 615 /**
 616  * Performs a double-dispatch based on the layout type of an object.
 617  */
 618 #define LOV_2DISPATCH_NOLOCK(obj, op, ...)                              \
 619 ({                                                                      \
 620         struct lov_object                      *__obj = (obj);          \
 621         enum lov_layout_type                    __llt;                  \
 622                                                                         \
 623         __llt = __obj->lo_type;                                         \
 624         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 625         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 626 })
 627
 628 /**
 629  * Return lov_layout_type associated with a given lsm
 630  */
 631 static enum lov_layout_type lov_type(struct lov_stripe_md *lsm)
 632 {
 633         if (lsm == NULL)
 634                 return LLT_EMPTY;
 635         if (lsm_is_released(lsm))
 636                 return LLT_RELEASED;
 637         return LLT_RAID0;
 638 }
 639
 640 static inline void lov_conf_freeze(struct lov_object *lov)
 641 {
 642         if (lov->lo_owner != current)
 643                 down_read(&lov->lo_type_guard);
 644 }
 645
 646 static inline void lov_conf_thaw(struct lov_object *lov)
 647 {
 648         if (lov->lo_owner != current)
 649                 up_read(&lov->lo_type_guard);
 650 }
 651
 652 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...)                       \
 653 ({                                                                      \
 654         struct lov_object                      *__obj = (obj);          \
 655         int                                     __lock = !!(lock);      \
 656         typeof(lov_dispatch[0].op(__VA_ARGS__)) __result;               \
 657                                                                         \
 658         if (__lock)                                                     \
 659                 lov_conf_freeze(__obj);                                 \
 660         __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__);          \
 661         if (__lock)                                                     \
 662                 lov_conf_thaw(__obj);                                   \
 663         __result;                                                       \
 664 })
 665
 666 /**
 667  * Performs a locked double-dispatch based on the layout type of an object.
 668  */
 669 #define LOV_2DISPATCH(obj, op, ...)                     \
 670         LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
 671
 672 #define LOV_2DISPATCH_VOID(obj, op, ...)                                \
 673 do {                                                                    \
 674         struct lov_object                      *__obj = (obj);          \
 675         enum lov_layout_type                    __llt;                  \
 676                                                                         \
 677         lov_conf_freeze(__obj);                                         \
 678         __llt = __obj->lo_type;                                         \
 679         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 680         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 681         lov_conf_thaw(__obj);                                           \
 682 } while (0)
 683
 684 static void lov_conf_lock(struct lov_object *lov)
 685 {
 686         LASSERT(lov->lo_owner != current);
 687         down_write(&lov->lo_type_guard);
 688         LASSERT(lov->lo_owner == NULL);
 689         lov->lo_owner = current;
 690 }
 691
 692 static void lov_conf_unlock(struct lov_object *lov)
 693 {
 694         lov->lo_owner = NULL;
 695         up_write(&lov->lo_type_guard);
 696 }
 697
 698 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov)
 699 {
 700         struct l_wait_info lwi = { 0 };
 701         ENTRY;
 702
 703         while (atomic_read(&lov->lo_active_ios) > 0) {
 704                 CDEBUG(D_INODE, "file:"DFID" wait for active IO, now: %d.\n",
 705                         PFID(lu_object_fid(lov2lu(lov))),
 706                         atomic_read(&lov->lo_active_ios));
 707
 708                 l_wait_event(lov->lo_waitq,
 709                              atomic_read(&lov->lo_active_ios) == 0, &lwi);
 710         }
 711         RETURN(0);
 712 }
 713
 714 static int lov_layout_change(const struct lu_env *unused,
 715                              struct lov_object *lov,
 716                              const struct cl_object_conf *conf)
 717 {
 718         int result;
 719         enum lov_layout_type llt = LLT_EMPTY;
 720         union lov_layout_state *state = &lov->u;
 721         const struct lov_layout_operations *old_ops;
 722         const struct lov_layout_operations *new_ops;
 723
 724         void *cookie;
 725         struct lu_env *env;
 726         int refcheck;
 727         ENTRY;
 728
 729         LASSERT(0 <= lov->lo_type && lov->lo_type < ARRAY_SIZE(lov_dispatch));
 730
 731         if (conf->u.coc_md != NULL)
 732                 llt = lov_type(conf->u.coc_md->lsm);
 733         LASSERT(0 <= llt && llt < ARRAY_SIZE(lov_dispatch));
 734
 735         cookie = cl_env_reenter();
 736         env = cl_env_get(&refcheck);
 737         if (IS_ERR(env)) {
 738                 cl_env_reexit(cookie);
 739                 RETURN(PTR_ERR(env));
 740         }
 741
 742         CDEBUG(D_INODE, DFID" from %s to %s\n",
 743                PFID(lu_object_fid(lov2lu(lov))),
 744                llt2str(lov->lo_type), llt2str(llt));
 745
 746         old_ops = &lov_dispatch[lov->lo_type];
 747         new_ops = &lov_dispatch[llt];
 748
 749         result = cl_object_prune(env, &lov->lo_cl);
 750         if (result != 0)
 751                 GOTO(out, result);
 752
 753         result = old_ops->llo_delete(env, lov, &lov->u);
 754         if (result == 0) {
 755                 old_ops->llo_fini(env, lov, &lov->u);
 756
 757                 LASSERT(atomic_read(&lov->lo_active_ios) == 0);
 758
 759                 lov->lo_type = LLT_EMPTY;
 760                 /* page bufsize fixup */
 761                 cl_object_header(&lov->lo_cl)->coh_page_bufsize -=
 762                         lov_page_slice_fixup(lov, NULL);
 763
 764                 result = new_ops->llo_init(env,
 765                                         lu2lov_dev(lov->lo_cl.co_lu.lo_dev),
 766                                         lov, conf, state);
 767                 if (result == 0) {
 768                         new_ops->llo_install(env, lov, state);
 769                         lov->lo_type = llt;
 770                 } else {
 771                         new_ops->llo_delete(env, lov, state);
 772                         new_ops->llo_fini(env, lov, state);
 773                         /* this file becomes an EMPTY file. */
 774                 }
 775         }
 776
 777 out:
 778         cl_env_put(env, &refcheck);
 779         cl_env_reexit(cookie);
 780         RETURN(result);
 781 }
 782
 783 /*****************************************************************************
 784  *
 785  * Lov object operations.
 786  *
 787  */
 788 int lov_object_init(const struct lu_env *env, struct lu_object *obj,
 789                     const struct lu_object_conf *conf)
 790 {
 791         struct lov_device            *dev   = lu2lov_dev(obj->lo_dev);
 792         struct lov_object            *lov   = lu2lov(obj);
 793         const struct cl_object_conf  *cconf = lu2cl_conf(conf);
 794         union  lov_layout_state      *set   = &lov->u;
 795         const struct lov_layout_operations *ops;
 796         int result;
 797
 798         ENTRY;
 799         init_rwsem(&lov->lo_type_guard);
 800         atomic_set(&lov->lo_active_ios, 0);
 801         init_waitqueue_head(&lov->lo_waitq);
 802
 803         cl_object_page_init(lu2cl(obj), sizeof(struct lov_page));
 804
 805         /* no locking is necessary, as object is being created */
 806         lov->lo_type = lov_type(cconf->u.coc_md->lsm);
 807         ops = &lov_dispatch[lov->lo_type];
 808         result = ops->llo_init(env, dev, lov, cconf, set);
 809         if (result == 0)
 810                 ops->llo_install(env, lov, set);
 811         RETURN(result);
 812 }
 813
 814 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
 815                         const struct cl_object_conf *conf)
 816 {
 817         struct lov_stripe_md    *lsm = NULL;
 818         struct lov_object       *lov = cl2lov(obj);
 819         int                      result = 0;
 820         ENTRY;
 821
 822         lov_conf_lock(lov);
 823         if (conf->coc_opc == OBJECT_CONF_INVALIDATE) {
 824                 lov->lo_layout_invalid = true;
 825                 GOTO(out, result = 0);
 826         }
 827
 828         if (conf->coc_opc == OBJECT_CONF_WAIT) {
 829                 if (lov->lo_layout_invalid &&
 830                     atomic_read(&lov->lo_active_ios) > 0) {
 831                         lov_conf_unlock(lov);
 832                         result = lov_layout_wait(env, lov);
 833                         lov_conf_lock(lov);
 834                 }
 835                 GOTO(out, result);
 836         }
 837
 838         LASSERT(conf->coc_opc == OBJECT_CONF_SET);
 839
 840         if (conf->u.coc_md != NULL)
 841                 lsm = conf->u.coc_md->lsm;
 842         if ((lsm == NULL && lov->lo_lsm == NULL) ||
 843             ((lsm != NULL && lov->lo_lsm != NULL) &&
 844              (lov->lo_lsm->lsm_layout_gen == lsm->lsm_layout_gen) &&
 845              (lov->lo_lsm->lsm_pattern == lsm->lsm_pattern))) {
 846                 /* same version of layout */
 847                 lov->lo_layout_invalid = false;
 848                 GOTO(out, result = 0);
 849         }
 850
 851         /* will change layout - check if there still exists active IO. */
 852         if (atomic_read(&lov->lo_active_ios) > 0) {
 853                 lov->lo_layout_invalid = true;
 854                 GOTO(out, result = -EBUSY);
 855         }
 856
 857         result = lov_layout_change(env, lov, conf);
 858         lov->lo_layout_invalid = result != 0;
 859         EXIT;
 860
 861 out:
 862         lov_conf_unlock(lov);
 863         CDEBUG(D_INODE, DFID" lo_layout_invalid=%d\n",
 864                PFID(lu_object_fid(lov2lu(lov))), lov->lo_layout_invalid);
 865         RETURN(result);
 866 }
 867
 868 static void lov_object_delete(const struct lu_env *env, struct lu_object *obj)
 869 {
 870         struct lov_object *lov = lu2lov(obj);
 871
 872         ENTRY;
 873         LOV_2DISPATCH_VOID(lov, llo_delete, env, lov, &lov->u);
 874         EXIT;
 875 }
 876
 877 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
 878 {
 879         struct lov_object *lov = lu2lov(obj);
 880
 881         ENTRY;
 882         LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
 883         lu_object_fini(obj);
 884         OBD_SLAB_FREE_PTR(lov, lov_object_kmem);
 885         EXIT;
 886 }
 887
 888 static int lov_object_print(const struct lu_env *env, void *cookie,
 889                             lu_printer_t p, const struct lu_object *o)
 890 {
 891         return LOV_2DISPATCH_NOLOCK(lu2lov(o), llo_print, env, cookie, p, o);
 892 }
 893
 894 int lov_page_init(const struct lu_env *env, struct cl_object *obj,
 895                   struct cl_page *page, pgoff_t index)
 896 {
 897         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_page_init, env, obj, page,
 898                                     index);
 899 }
 900
 901 /**
 902  * Implements cl_object_operations::clo_io_init() method for lov
 903  * layer. Dispatches to the appropriate layout io initialization method.
 904  */
 905 int lov_io_init(const struct lu_env *env, struct cl_object *obj,
 906                 struct cl_io *io)
 907 {
 908         CL_IO_SLICE_CLEAN(lov_env_io(env), lis_cl);
 909         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
 910                                      !io->ci_ignore_layout, env, obj, io);
 911 }
 912
 913 /**
 914  * An implementation of cl_object_operations::clo_attr_get() method for lov
 915  * layer. For raid0 layout this collects and merges attributes of all
 916  * sub-objects.
 917  */
 918 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
 919                         struct cl_attr *attr)
 920 {
 921         /* do not take lock, as this function is called under a
 922          * spin-lock. Layout is protected from changing by ongoing IO. */
 923         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
 924 }
 925
 926 static int lov_attr_set(const struct lu_env *env, struct cl_object *obj,
 927                         const struct cl_attr *attr, unsigned valid)
 928 {
 929         /*
 930          * No dispatch is required here, as no layout implements this.
 931          */
 932         return 0;
 933 }
 934
 935 int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
 936                   struct cl_lock *lock, const struct cl_io *io)
 937 {
 938         /* No need to lock because we've taken one refcount of layout.  */
 939         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_lock_init, env, obj, lock,
 940                                     io);
 941 }
 942
 943 static int lov_object_getstripe(const struct lu_env *env, struct cl_object *obj,
 944                                 struct lov_user_md __user *lum)
 945 {
 946         struct lov_object       *lov = cl2lov(obj);
 947         struct lov_stripe_md    *lsm;
 948         int                     rc = 0;
 949         ENTRY;
 950
 951         lsm = lov_lsm_addref(lov);
 952         if (lsm == NULL)
 953                 RETURN(-ENODATA);
 954
 955         rc = lov_getstripe(cl2lov(obj), lsm, lum);
 956         lov_lsm_put(obj, lsm);
 957         RETURN(rc);
 958 }
 959
 960 static const struct cl_object_operations lov_ops = {
 961         .coo_page_init = lov_page_init,
 962         .coo_lock_init = lov_lock_init,
 963         .coo_io_init   = lov_io_init,
 964         .coo_attr_get  = lov_attr_get,
 965         .coo_attr_set  = lov_attr_set,
 966         .coo_conf_set  = lov_conf_set,
 967         .coo_getstripe = lov_object_getstripe
 968 };
 969
 970 static const struct lu_object_operations lov_lu_obj_ops = {
 971         .loo_object_init      = lov_object_init,
 972         .loo_object_delete    = lov_object_delete,
 973         .loo_object_release   = NULL,
 974         .loo_object_free      = lov_object_free,
 975         .loo_object_print     = lov_object_print,
 976         .loo_object_invariant = NULL
 977 };
 978
 979 struct lu_object *lov_object_alloc(const struct lu_env *env,
 980                                    const struct lu_object_header *unused,
 981                                    struct lu_device *dev)
 982 {
 983         struct lov_object *lov;
 984         struct lu_object  *obj;
 985
 986         ENTRY;
 987         OBD_SLAB_ALLOC_PTR_GFP(lov, lov_object_kmem, GFP_NOFS);
 988         if (lov != NULL) {
 989                 obj = lov2lu(lov);
 990                 lu_object_init(obj, NULL, dev);
 991                 lov->lo_cl.co_ops = &lov_ops;
 992                 lov->lo_type = -1; /* invalid, to catch uninitialized type */
 993                 /*
 994                  * object io operation vector (cl_object::co_iop) is installed
 995                  * later in lov_object_init(), as different vectors are used
 996                  * for object with different layouts.
 997                  */
 998                 obj->lo_ops = &lov_lu_obj_ops;
 999         } else
1000                 obj = NULL;
1001         RETURN(obj);
1002 }
1003
1004 struct lov_stripe_md *lov_lsm_addref(struct lov_object *lov)
1005 {
1006         struct lov_stripe_md *lsm = NULL;
1007
1008         lov_conf_freeze(lov);
1009         if (lov->lo_lsm != NULL) {
1010                 lsm = lsm_addref(lov->lo_lsm);
1011                 CDEBUG(D_INODE, "lsm %p addref %d/%d by %p.\n",
1012                         lsm, atomic_read(&lsm->lsm_refc),
1013                         lov->lo_layout_invalid, current);
1014         }
1015         lov_conf_thaw(lov);
1016         return lsm;
1017 }
1018
1019 int lov_read_and_clear_async_rc(struct cl_object *clob)
1020 {
1021         struct lu_object *luobj;
1022         int rc = 0;
1023         ENTRY;
1024
1025         luobj = lu_object_locate(&cl_object_header(clob)->coh_lu,
1026                                  &lov_device_type);
1027         if (luobj != NULL) {
1028                 struct lov_object *lov = lu2lov(luobj);
1029
1030                 lov_conf_freeze(lov);
1031                 switch (lov->lo_type) {
1032                 case LLT_RAID0: {
1033                         struct lov_stripe_md *lsm;
1034                         int i;
1035
1036                         lsm = lov->lo_lsm;
1037                         LASSERT(lsm != NULL);
1038                         for (i = 0; i < lsm->lsm_stripe_count; i++) {
1039                                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
1040
1041                                 if (lov_oinfo_is_dummy(loi))
1042                                         continue;
1043
1044                                 if (loi->loi_ar.ar_rc && !rc)
1045                                         rc = loi->loi_ar.ar_rc;
1046                                 loi->loi_ar.ar_rc = 0;
1047                         }
1048                 }
1049                 case LLT_RELEASED:
1050                 case LLT_EMPTY:
1051                         break;
1052                 default:
1053                         LBUG();
1054                 }
1055                 lov_conf_thaw(lov);
1056         }
1057         RETURN(rc);
1058 }
1059 EXPORT_SYMBOL(lov_read_and_clear_async_rc);
1060
1061 /** @} lov */