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