kernel/drivers/staging/lustre/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, 2012, 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 "../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, 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         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_init_raid0(const struct lu_env *env,
 196                           struct lov_device *dev, struct lov_object *lov,
 197                           const struct cl_object_conf *conf,
 198                           union  lov_layout_state *state)
 199 {
 200         int result;
 201         int i;
 202
 203         struct cl_object        *stripe;
 204         struct lov_thread_info  *lti     = lov_env_info(env);
 205         struct cl_object_conf   *subconf = &lti->lti_stripe_conf;
 206         struct lov_stripe_md    *lsm     = conf->u.coc_md->lsm;
 207         struct lu_fid      *ofid    = &lti->lti_fid;
 208         struct lov_layout_raid0 *r0      = &state->raid0;
 209
 210         if (lsm->lsm_magic != LOV_MAGIC_V1 && lsm->lsm_magic != LOV_MAGIC_V3) {
 211                 dump_lsm(D_ERROR, lsm);
 212                 LASSERTF(0, "magic mismatch, expected %d/%d, actual %d.\n",
 213                          LOV_MAGIC_V1, LOV_MAGIC_V3, lsm->lsm_magic);
 214         }
 215
 216         LASSERT(lov->lo_lsm == NULL);
 217         lov->lo_lsm = lsm_addref(lsm);
 218         r0->lo_nr  = lsm->lsm_stripe_count;
 219         LASSERT(r0->lo_nr <= lov_targets_nr(dev));
 220
 221         OBD_ALLOC_LARGE(r0->lo_sub, r0->lo_nr * sizeof(r0->lo_sub[0]));
 222         if (r0->lo_sub != NULL) {
 223                 result = 0;
 224                 subconf->coc_inode = conf->coc_inode;
 225                 spin_lock_init(&r0->lo_sub_lock);
 226                 /*
 227                  * Create stripe cl_objects.
 228                  */
 229                 for (i = 0; i < r0->lo_nr && result == 0; ++i) {
 230                         struct cl_device *subdev;
 231                         struct lov_oinfo *oinfo = lsm->lsm_oinfo[i];
 232                         int ost_idx = oinfo->loi_ost_idx;
 233
 234                         if (lov_oinfo_is_dummy(oinfo))
 235                                 continue;
 236
 237                         result = ostid_to_fid(ofid, &oinfo->loi_oi,
 238                                               oinfo->loi_ost_idx);
 239                         if (result != 0)
 240                                 goto out;
 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(waiter, current);
 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                                 schedule();
 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         dump_lsm(D_INODE, lsm);
 346
 347         lov_layout_wait(env, lov);
 348         if (r0->lo_sub != NULL) {
 349                 for (i = 0; i < r0->lo_nr; ++i) {
 350                         struct lovsub_object *los = r0->lo_sub[i];
 351
 352                         if (los != NULL) {
 353                                 cl_locks_prune(env, &los->lso_cl, 1);
 354                                 /*
 355                                  * If top-level object is to be evicted from
 356                                  * the cache, so are its sub-objects.
 357                                  */
 358                                 lov_subobject_kill(env, lov, los, i);
 359                         }
 360                 }
 361         }
 362         cl_object_prune(env, &lov->lo_cl);
 363         return 0;
 364 }
 365
 366 static void lov_fini_empty(const struct lu_env *env, struct lov_object *lov,
 367                            union lov_layout_state *state)
 368 {
 369         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 370 }
 371
 372 static void lov_fini_raid0(const struct lu_env *env, struct lov_object *lov,
 373                            union lov_layout_state *state)
 374 {
 375         struct lov_layout_raid0 *r0 = &state->raid0;
 376
 377         if (r0->lo_sub != NULL) {
 378                 OBD_FREE_LARGE(r0->lo_sub, r0->lo_nr * sizeof(r0->lo_sub[0]));
 379                 r0->lo_sub = NULL;
 380         }
 381
 382         dump_lsm(D_INODE, lov->lo_lsm);
 383         lov_free_memmd(&lov->lo_lsm);
 384 }
 385
 386 static void lov_fini_released(const struct lu_env *env, struct lov_object *lov,
 387                                 union lov_layout_state *state)
 388 {
 389         dump_lsm(D_INODE, lov->lo_lsm);
 390         lov_free_memmd(&lov->lo_lsm);
 391 }
 392
 393 static int lov_print_empty(const struct lu_env *env, void *cookie,
 394                            lu_printer_t p, const struct lu_object *o)
 395 {
 396         (*p)(env, cookie, "empty %d\n", lu2lov(o)->lo_layout_invalid);
 397         return 0;
 398 }
 399
 400 static int lov_print_raid0(const struct lu_env *env, void *cookie,
 401                            lu_printer_t p, const struct lu_object *o)
 402 {
 403         struct lov_object       *lov = lu2lov(o);
 404         struct lov_layout_raid0 *r0  = lov_r0(lov);
 405         struct lov_stripe_md    *lsm = lov->lo_lsm;
 406         int                      i;
 407
 408         (*p)(env, cookie, "stripes: %d, %s, lsm{%p 0x%08X %d %u %u}:\n",
 409                 r0->lo_nr, lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 410                 lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 411                 lsm->lsm_stripe_count, lsm->lsm_layout_gen);
 412         for (i = 0; i < r0->lo_nr; ++i) {
 413                 struct lu_object *sub;
 414
 415                 if (r0->lo_sub[i] != NULL) {
 416                         sub = lovsub2lu(r0->lo_sub[i]);
 417                         lu_object_print(env, cookie, p, sub);
 418                 } else {
 419                         (*p)(env, cookie, "sub %d absent\n", i);
 420                 }
 421         }
 422         return 0;
 423 }
 424
 425 static int lov_print_released(const struct lu_env *env, void *cookie,
 426                                 lu_printer_t p, const struct lu_object *o)
 427 {
 428         struct lov_object       *lov = lu2lov(o);
 429         struct lov_stripe_md    *lsm = lov->lo_lsm;
 430
 431         (*p)(env, cookie,
 432                 "released: %s, lsm{%p 0x%08X %d %u %u}:\n",
 433                 lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 434                 lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 435                 lsm->lsm_stripe_count, lsm->lsm_layout_gen);
 436         return 0;
 437 }
 438
 439 /**
 440  * Implements cl_object_operations::coo_attr_get() method for an object
 441  * without stripes (LLT_EMPTY layout type).
 442  *
 443  * The only attributes this layer is authoritative in this case is
 444  * cl_attr::cat_blocks---it's 0.
 445  */
 446 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
 447                               struct cl_attr *attr)
 448 {
 449         attr->cat_blocks = 0;
 450         return 0;
 451 }
 452
 453 static int lov_attr_get_raid0(const struct lu_env *env, struct cl_object *obj,
 454                               struct cl_attr *attr)
 455 {
 456         struct lov_object       *lov = cl2lov(obj);
 457         struct lov_layout_raid0 *r0 = lov_r0(lov);
 458         struct cl_attr          *lov_attr = &r0->lo_attr;
 459         int                      result = 0;
 460
 461         /* this is called w/o holding type guard mutex, so it must be inside
 462          * an on going IO otherwise lsm may be replaced.
 463          * LU-2117: it turns out there exists one exception. For mmaped files,
 464          * the lock of those files may be requested in the other file's IO
 465          * context, and this function is called in ccc_lock_state(), it will
 466          * hit this assertion.
 467          * Anyway, it's still okay to call attr_get w/o type guard as layout
 468          * can't go if locks exist. */
 469         /* LASSERT(atomic_read(&lsm->lsm_refc) > 1); */
 470
 471         if (!r0->lo_attr_valid) {
 472                 struct lov_stripe_md    *lsm = lov->lo_lsm;
 473                 struct ost_lvb    *lvb = &lov_env_info(env)->lti_lvb;
 474                 __u64               kms = 0;
 475
 476                 memset(lvb, 0, sizeof(*lvb));
 477                 /* XXX: timestamps can be negative by sanity:test_39m,
 478                  * how can it be? */
 479                 lvb->lvb_atime = LLONG_MIN;
 480                 lvb->lvb_ctime = LLONG_MIN;
 481                 lvb->lvb_mtime = LLONG_MIN;
 482
 483                 /*
 484                  * XXX that should be replaced with a loop over sub-objects,
 485                  * doing cl_object_attr_get() on them. But for now, let's
 486                  * reuse old lov code.
 487                  */
 488
 489                 /*
 490                  * XXX take lsm spin-lock to keep lov_merge_lvb_kms()
 491                  * happy. It's not needed, because new code uses
 492                  * ->coh_attr_guard spin-lock to protect consistency of
 493                  * sub-object attributes.
 494                  */
 495                 lov_stripe_lock(lsm);
 496                 result = lov_merge_lvb_kms(lsm, lvb, &kms);
 497                 lov_stripe_unlock(lsm);
 498                 if (result == 0) {
 499                         cl_lvb2attr(lov_attr, lvb);
 500                         lov_attr->cat_kms = kms;
 501                         r0->lo_attr_valid = 1;
 502                 }
 503         }
 504         if (result == 0) { /* merge results */
 505                 attr->cat_blocks = lov_attr->cat_blocks;
 506                 attr->cat_size = lov_attr->cat_size;
 507                 attr->cat_kms = lov_attr->cat_kms;
 508                 if (attr->cat_atime < lov_attr->cat_atime)
 509                         attr->cat_atime = lov_attr->cat_atime;
 510                 if (attr->cat_ctime < lov_attr->cat_ctime)
 511                         attr->cat_ctime = lov_attr->cat_ctime;
 512                 if (attr->cat_mtime < lov_attr->cat_mtime)
 513                         attr->cat_mtime = lov_attr->cat_mtime;
 514         }
 515         return result;
 516 }
 517
 518 static const struct lov_layout_operations lov_dispatch[] = {
 519         [LLT_EMPTY] = {
 520                 .llo_init      = lov_init_empty,
 521                 .llo_delete    = lov_delete_empty,
 522                 .llo_fini      = lov_fini_empty,
 523                 .llo_install   = lov_install_empty,
 524                 .llo_print     = lov_print_empty,
 525                 .llo_page_init = lov_page_init_empty,
 526                 .llo_lock_init = lov_lock_init_empty,
 527                 .llo_io_init   = lov_io_init_empty,
 528                 .llo_getattr   = lov_attr_get_empty
 529         },
 530         [LLT_RAID0] = {
 531                 .llo_init      = lov_init_raid0,
 532                 .llo_delete    = lov_delete_raid0,
 533                 .llo_fini      = lov_fini_raid0,
 534                 .llo_install   = lov_install_raid0,
 535                 .llo_print     = lov_print_raid0,
 536                 .llo_page_init = lov_page_init_raid0,
 537                 .llo_lock_init = lov_lock_init_raid0,
 538                 .llo_io_init   = lov_io_init_raid0,
 539                 .llo_getattr   = lov_attr_get_raid0
 540         },
 541         [LLT_RELEASED] = {
 542                 .llo_init      = lov_init_released,
 543                 .llo_delete    = lov_delete_empty,
 544                 .llo_fini      = lov_fini_released,
 545                 .llo_install   = lov_install_empty,
 546                 .llo_print     = lov_print_released,
 547                 .llo_page_init = lov_page_init_empty,
 548                 .llo_lock_init = lov_lock_init_empty,
 549                 .llo_io_init   = lov_io_init_released,
 550                 .llo_getattr   = lov_attr_get_empty
 551         }
 552 };
 553
 554 /**
 555  * Performs a double-dispatch based on the layout type of an object.
 556  */
 557 #define LOV_2DISPATCH_NOLOCK(obj, op, ...)                            \
 558 ({                                                                    \
 559         struct lov_object                     *__obj = (obj);     \
 560         enum lov_layout_type                __llt;                \
 561                                                                         \
 562         __llt = __obj->lo_type;                                  \
 563         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 564         lov_dispatch[__llt].op(__VA_ARGS__);                        \
 565 })
 566
 567 /**
 568  * Return lov_layout_type associated with a given lsm
 569  */
 570 static enum lov_layout_type lov_type(struct lov_stripe_md *lsm)
 571 {
 572         if (lsm == NULL)
 573                 return LLT_EMPTY;
 574         if (lsm_is_released(lsm))
 575                 return LLT_RELEASED;
 576         return LLT_RAID0;
 577 }
 578
 579 static inline void lov_conf_freeze(struct lov_object *lov)
 580 {
 581         if (lov->lo_owner != current)
 582                 down_read(&lov->lo_type_guard);
 583 }
 584
 585 static inline void lov_conf_thaw(struct lov_object *lov)
 586 {
 587         if (lov->lo_owner != current)
 588                 up_read(&lov->lo_type_guard);
 589 }
 590
 591 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...)                      \
 592 ({                                                                    \
 593         struct lov_object                     *__obj = (obj);     \
 594         int                                  __lock = !!(lock);      \
 595         typeof(lov_dispatch[0].op(__VA_ARGS__)) __result;              \
 596                                                                         \
 597         if (__lock)                                                  \
 598                 lov_conf_freeze(__obj);                                 \
 599         __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__);    \
 600         if (__lock)                                                  \
 601                 lov_conf_thaw(__obj);                                   \
 602         __result;                                                      \
 603 })
 604
 605 /**
 606  * Performs a locked double-dispatch based on the layout type of an object.
 607  */
 608 #define LOV_2DISPATCH(obj, op, ...)                  \
 609         LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
 610
 611 #define LOV_2DISPATCH_VOID(obj, op, ...)                                \
 612 do {                                                                \
 613         struct lov_object                     *__obj = (obj);     \
 614         enum lov_layout_type                __llt;                \
 615                                                                         \
 616         lov_conf_freeze(__obj);                                         \
 617         __llt = __obj->lo_type;                                  \
 618         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 619         lov_dispatch[__llt].op(__VA_ARGS__);                        \
 620         lov_conf_thaw(__obj);                                           \
 621 } while (0)
 622
 623 static void lov_conf_lock(struct lov_object *lov)
 624 {
 625         LASSERT(lov->lo_owner != current);
 626         down_write(&lov->lo_type_guard);
 627         LASSERT(lov->lo_owner == NULL);
 628         lov->lo_owner = current;
 629 }
 630
 631 static void lov_conf_unlock(struct lov_object *lov)
 632 {
 633         lov->lo_owner = NULL;
 634         up_write(&lov->lo_type_guard);
 635 }
 636
 637 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov)
 638 {
 639         struct l_wait_info lwi = { 0 };
 640
 641         while (atomic_read(&lov->lo_active_ios) > 0) {
 642                 CDEBUG(D_INODE, "file:"DFID" wait for active IO, now: %d.\n",
 643                         PFID(lu_object_fid(lov2lu(lov))),
 644                         atomic_read(&lov->lo_active_ios));
 645
 646                 l_wait_event(lov->lo_waitq,
 647                              atomic_read(&lov->lo_active_ios) == 0, &lwi);
 648         }
 649         return 0;
 650 }
 651
 652 static int lov_layout_change(const struct lu_env *unused,
 653                              struct lov_object *lov,
 654                              const struct cl_object_conf *conf)
 655 {
 656         int result;
 657         enum lov_layout_type llt = LLT_EMPTY;
 658         union lov_layout_state *state = &lov->u;
 659         const struct lov_layout_operations *old_ops;
 660         const struct lov_layout_operations *new_ops;
 661
 662         struct cl_object_header *hdr = cl_object_header(&lov->lo_cl);
 663         void *cookie;
 664         struct lu_env *env;
 665         int refcheck;
 666
 667         LASSERT(0 <= lov->lo_type && lov->lo_type < ARRAY_SIZE(lov_dispatch));
 668
 669         if (conf->u.coc_md != NULL)
 670                 llt = lov_type(conf->u.coc_md->lsm);
 671         LASSERT(0 <= llt && llt < ARRAY_SIZE(lov_dispatch));
 672
 673         cookie = cl_env_reenter();
 674         env = cl_env_get(&refcheck);
 675         if (IS_ERR(env)) {
 676                 cl_env_reexit(cookie);
 677                 return PTR_ERR(env);
 678         }
 679
 680         CDEBUG(D_INODE, DFID" from %s to %s\n",
 681                PFID(lu_object_fid(lov2lu(lov))),
 682                llt2str(lov->lo_type), llt2str(llt));
 683
 684         old_ops = &lov_dispatch[lov->lo_type];
 685         new_ops = &lov_dispatch[llt];
 686
 687         result = old_ops->llo_delete(env, lov, &lov->u);
 688         if (result == 0) {
 689                 old_ops->llo_fini(env, lov, &lov->u);
 690
 691                 LASSERT(atomic_read(&lov->lo_active_ios) == 0);
 692                 LASSERT(hdr->coh_tree.rnode == NULL);
 693                 LASSERT(hdr->coh_pages == 0);
 694
 695                 lov->lo_type = LLT_EMPTY;
 696                 result = new_ops->llo_init(env,
 697                                         lu2lov_dev(lov->lo_cl.co_lu.lo_dev),
 698                                         lov, conf, state);
 699                 if (result == 0) {
 700                         new_ops->llo_install(env, lov, state);
 701                         lov->lo_type = llt;
 702                 } else {
 703                         new_ops->llo_delete(env, lov, state);
 704                         new_ops->llo_fini(env, lov, state);
 705                         /* this file becomes an EMPTY file. */
 706                 }
 707         }
 708
 709         cl_env_put(env, &refcheck);
 710         cl_env_reexit(cookie);
 711         return result;
 712 }
 713
 714 /*****************************************************************************
 715  *
 716  * Lov object operations.
 717  *
 718  */
 719 int lov_object_init(const struct lu_env *env, struct lu_object *obj,
 720                     const struct lu_object_conf *conf)
 721 {
 722         struct lov_device           *dev   = lu2lov_dev(obj->lo_dev);
 723         struct lov_object           *lov   = lu2lov(obj);
 724         const struct cl_object_conf  *cconf = lu2cl_conf(conf);
 725         union  lov_layout_state      *set   = &lov->u;
 726         const struct lov_layout_operations *ops;
 727         int result;
 728
 729         init_rwsem(&lov->lo_type_guard);
 730         atomic_set(&lov->lo_active_ios, 0);
 731         init_waitqueue_head(&lov->lo_waitq);
 732
 733         cl_object_page_init(lu2cl(obj), sizeof(struct lov_page));
 734
 735         /* no locking is necessary, as object is being created */
 736         lov->lo_type = lov_type(cconf->u.coc_md->lsm);
 737         ops = &lov_dispatch[lov->lo_type];
 738         result = ops->llo_init(env, dev, lov, cconf, set);
 739         if (result == 0)
 740                 ops->llo_install(env, lov, set);
 741         return result;
 742 }
 743
 744 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
 745                         const struct cl_object_conf *conf)
 746 {
 747         struct lov_stripe_md    *lsm = NULL;
 748         struct lov_object       *lov = cl2lov(obj);
 749         int                      result = 0;
 750
 751         lov_conf_lock(lov);
 752         if (conf->coc_opc == OBJECT_CONF_INVALIDATE) {
 753                 lov->lo_layout_invalid = true;
 754                 result = 0;
 755                 goto out;
 756         }
 757
 758         if (conf->coc_opc == OBJECT_CONF_WAIT) {
 759                 if (lov->lo_layout_invalid &&
 760                     atomic_read(&lov->lo_active_ios) > 0) {
 761                         lov_conf_unlock(lov);
 762                         result = lov_layout_wait(env, lov);
 763                         lov_conf_lock(lov);
 764                 }
 765                 goto out;
 766         }
 767
 768         LASSERT(conf->coc_opc == OBJECT_CONF_SET);
 769
 770         if (conf->u.coc_md != NULL)
 771                 lsm = conf->u.coc_md->lsm;
 772         if ((lsm == NULL && lov->lo_lsm == NULL) ||
 773             ((lsm != NULL && lov->lo_lsm != NULL) &&
 774              (lov->lo_lsm->lsm_layout_gen == lsm->lsm_layout_gen) &&
 775              (lov->lo_lsm->lsm_pattern == lsm->lsm_pattern))) {
 776                 /* same version of layout */
 777                 lov->lo_layout_invalid = false;
 778                 result = 0;
 779                 goto out;
 780         }
 781
 782         /* will change layout - check if there still exists active IO. */
 783         if (atomic_read(&lov->lo_active_ios) > 0) {
 784                 lov->lo_layout_invalid = true;
 785                 result = -EBUSY;
 786                 goto out;
 787         }
 788
 789         lov->lo_layout_invalid = lov_layout_change(env, lov, conf);
 790
 791 out:
 792         lov_conf_unlock(lov);
 793         CDEBUG(D_INODE, DFID" lo_layout_invalid=%d\n",
 794                PFID(lu_object_fid(lov2lu(lov))), lov->lo_layout_invalid);
 795         return result;
 796 }
 797
 798 static void lov_object_delete(const struct lu_env *env, struct lu_object *obj)
 799 {
 800         struct lov_object *lov = lu2lov(obj);
 801
 802         LOV_2DISPATCH_VOID(lov, llo_delete, env, lov, &lov->u);
 803 }
 804
 805 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
 806 {
 807         struct lov_object *lov = lu2lov(obj);
 808
 809         LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
 810         lu_object_fini(obj);
 811         OBD_SLAB_FREE_PTR(lov, lov_object_kmem);
 812 }
 813
 814 static int lov_object_print(const struct lu_env *env, void *cookie,
 815                             lu_printer_t p, const struct lu_object *o)
 816 {
 817         return LOV_2DISPATCH_NOLOCK(lu2lov(o), llo_print, env, cookie, p, o);
 818 }
 819
 820 int lov_page_init(const struct lu_env *env, struct cl_object *obj,
 821                 struct cl_page *page, struct page *vmpage)
 822 {
 823         return LOV_2DISPATCH_NOLOCK(cl2lov(obj),
 824                                     llo_page_init, env, obj, page, vmpage);
 825 }
 826
 827 /**
 828  * Implements cl_object_operations::clo_io_init() method for lov
 829  * layer. Dispatches to the appropriate layout io initialization method.
 830  */
 831 int lov_io_init(const struct lu_env *env, struct cl_object *obj,
 832                 struct cl_io *io)
 833 {
 834         CL_IO_SLICE_CLEAN(lov_env_io(env), lis_cl);
 835         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
 836                                      !io->ci_ignore_layout, env, obj, io);
 837 }
 838
 839 /**
 840  * An implementation of cl_object_operations::clo_attr_get() method for lov
 841  * layer. For raid0 layout this collects and merges attributes of all
 842  * sub-objects.
 843  */
 844 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
 845                         struct cl_attr *attr)
 846 {
 847         /* do not take lock, as this function is called under a
 848          * spin-lock. Layout is protected from changing by ongoing IO. */
 849         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
 850 }
 851
 852 static int lov_attr_set(const struct lu_env *env, struct cl_object *obj,
 853                         const struct cl_attr *attr, unsigned valid)
 854 {
 855         /*
 856          * No dispatch is required here, as no layout implements this.
 857          */
 858         return 0;
 859 }
 860
 861 int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
 862                   struct cl_lock *lock, const struct cl_io *io)
 863 {
 864         /* No need to lock because we've taken one refcount of layout.  */
 865         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_lock_init, env, obj, lock,
 866                                     io);
 867 }
 868
 869 static const struct cl_object_operations lov_ops = {
 870         .coo_page_init = lov_page_init,
 871         .coo_lock_init = lov_lock_init,
 872         .coo_io_init   = lov_io_init,
 873         .coo_attr_get  = lov_attr_get,
 874         .coo_attr_set  = lov_attr_set,
 875         .coo_conf_set  = lov_conf_set
 876 };
 877
 878 static const struct lu_object_operations lov_lu_obj_ops = {
 879         .loo_object_init      = lov_object_init,
 880         .loo_object_delete    = lov_object_delete,
 881         .loo_object_release   = NULL,
 882         .loo_object_free      = lov_object_free,
 883         .loo_object_print     = lov_object_print,
 884         .loo_object_invariant = NULL
 885 };
 886
 887 struct lu_object *lov_object_alloc(const struct lu_env *env,
 888                                    const struct lu_object_header *unused,
 889                                    struct lu_device *dev)
 890 {
 891         struct lov_object *lov;
 892         struct lu_object  *obj;
 893
 894         OBD_SLAB_ALLOC_PTR_GFP(lov, lov_object_kmem, GFP_NOFS);
 895         if (lov != NULL) {
 896                 obj = lov2lu(lov);
 897                 lu_object_init(obj, NULL, dev);
 898                 lov->lo_cl.co_ops = &lov_ops;
 899                 lov->lo_type = -1; /* invalid, to catch uninitialized type */
 900                 /*
 901                  * object io operation vector (cl_object::co_iop) is installed
 902                  * later in lov_object_init(), as different vectors are used
 903                  * for object with different layouts.
 904                  */
 905                 obj->lo_ops = &lov_lu_obj_ops;
 906         } else
 907                 obj = NULL;
 908         return obj;
 909 }
 910
 911 struct lov_stripe_md *lov_lsm_addref(struct lov_object *lov)
 912 {
 913         struct lov_stripe_md *lsm = NULL;
 914
 915         lov_conf_freeze(lov);
 916         if (lov->lo_lsm != NULL) {
 917                 lsm = lsm_addref(lov->lo_lsm);
 918                 CDEBUG(D_INODE, "lsm %p addref %d/%d by %p.\n",
 919                         lsm, atomic_read(&lsm->lsm_refc),
 920                         lov->lo_layout_invalid, current);
 921         }
 922         lov_conf_thaw(lov);
 923         return lsm;
 924 }
 925
 926 void lov_lsm_decref(struct lov_object *lov, struct lov_stripe_md *lsm)
 927 {
 928         if (lsm == NULL)
 929                 return;
 930
 931         CDEBUG(D_INODE, "lsm %p decref %d by %p.\n",
 932                 lsm, atomic_read(&lsm->lsm_refc), current);
 933
 934         lov_free_memmd(&lsm);
 935 }
 936
 937 struct lov_stripe_md *lov_lsm_get(struct cl_object *clobj)
 938 {
 939         struct lu_object *luobj;
 940         struct lov_stripe_md *lsm = NULL;
 941
 942         if (clobj == NULL)
 943                 return NULL;
 944
 945         luobj = lu_object_locate(&cl_object_header(clobj)->coh_lu,
 946                                  &lov_device_type);
 947         if (luobj != NULL)
 948                 lsm = lov_lsm_addref(lu2lov(luobj));
 949         return lsm;
 950 }
 951 EXPORT_SYMBOL(lov_lsm_get);
 952
 953 void lov_lsm_put(struct cl_object *unused, struct lov_stripe_md *lsm)
 954 {
 955         if (lsm != NULL)
 956                 lov_free_memmd(&lsm);
 957 }
 958 EXPORT_SYMBOL(lov_lsm_put);
 959
 960 int lov_read_and_clear_async_rc(struct cl_object *clob)
 961 {
 962         struct lu_object *luobj;
 963         int rc = 0;
 964
 965         luobj = lu_object_locate(&cl_object_header(clob)->coh_lu,
 966                                  &lov_device_type);
 967         if (luobj != NULL) {
 968                 struct lov_object *lov = lu2lov(luobj);
 969
 970                 lov_conf_freeze(lov);
 971                 switch (lov->lo_type) {
 972                 case LLT_RAID0: {
 973                         struct lov_stripe_md *lsm;
 974                         int i;
 975
 976                         lsm = lov->lo_lsm;
 977                         LASSERT(lsm != NULL);
 978                         for (i = 0; i < lsm->lsm_stripe_count; i++) {
 979                                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
 980
 981                                 if (lov_oinfo_is_dummy(loi))
 982                                         continue;
 983
 984                                 if (loi->loi_ar.ar_rc && !rc)
 985                                         rc = loi->loi_ar.ar_rc;
 986                                 loi->loi_ar.ar_rc = 0;
 987                         }
 988                 }
 989                 case LLT_RELEASED:
 990                 case LLT_EMPTY:
 991                         break;
 992                 default:
 993                         LBUG();
 994                 }
 995                 lov_conf_thaw(lov);
 996         }
 997         return rc;
 998 }
 999 EXPORT_SYMBOL(lov_read_and_clear_async_rc);
1000
1001 /** @} lov */