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