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 <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, cfs_page_t *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 /*****************************************************************************
  82  *
  83  * Lov object layout operations.
  84  *
  85  */
  86
  87 static void lov_install_empty(const struct lu_env *env,
  88                               struct lov_object *lov,
  89                               union  lov_layout_state *state)
  90 {
  91         /*
  92          * File without objects.
  93          */
  94 }
  95
  96 static int lov_init_empty(const struct lu_env *env,
  97                           struct lov_device *dev, struct lov_object *lov,
  98                           const struct cl_object_conf *conf,
  99                           union  lov_layout_state *state)
 100 {
 101         return 0;
 102 }
 103
 104 static void lov_install_raid0(const struct lu_env *env,
 105                               struct lov_object *lov,
 106                               union  lov_layout_state *state)
 107 {
 108 }
 109
 110 static struct cl_object *lov_sub_find(const struct lu_env *env,
 111                                       struct cl_device *dev,
 112                                       const struct lu_fid *fid,
 113                                       const struct cl_object_conf *conf)
 114 {
 115         struct lu_object *o;
 116
 117         ENTRY;
 118         o = lu_object_find_at(env, cl2lu_dev(dev), fid, &conf->coc_lu);
 119         LASSERT(ergo(!IS_ERR(o), o->lo_dev->ld_type == &lovsub_device_type));
 120         RETURN(lu2cl(o));
 121 }
 122
 123 static int lov_init_sub(const struct lu_env *env, struct lov_object *lov,
 124                         struct cl_object *stripe,
 125                         struct lov_layout_raid0 *r0, int idx)
 126 {
 127         struct cl_object_header *hdr;
 128         struct cl_object_header *subhdr;
 129         struct cl_object_header *parent;
 130         struct lov_oinfo        *oinfo;
 131         int result;
 132
 133         if (OBD_FAIL_CHECK(OBD_FAIL_LOV_INIT)) {
 134                 /* For sanity:test_206.
 135                  * Do not leave the object in cache to avoid accessing
 136                  * freed memory. This is because osc_object is referring to
 137                  * lov_oinfo of lsm_stripe_data which will be freed due to
 138                  * this failure. */
 139                 cl_object_kill(env, stripe);
 140                 cl_object_put(env, stripe);
 141                 return -EIO;
 142         }
 143
 144         hdr    = cl_object_header(lov2cl(lov));
 145         subhdr = cl_object_header(stripe);
 146         parent = subhdr->coh_parent;
 147
 148         oinfo = lov->lo_lsm->lsm_oinfo[idx];
 149         CDEBUG(D_INODE, DFID"@%p[%d] -> "DFID"@%p: id: "LPU64" seq: "LPU64
 150                " idx: %d gen: %d\n",
 151                PFID(&subhdr->coh_lu.loh_fid), subhdr, idx,
 152                PFID(&hdr->coh_lu.loh_fid), hdr,
 153                oinfo->loi_id, oinfo->loi_seq,
 154                oinfo->loi_ost_idx, oinfo->loi_ost_gen);
 155
 156         if (parent == NULL) {
 157                 subhdr->coh_parent = hdr;
 158                 subhdr->coh_nesting = hdr->coh_nesting + 1;
 159                 lu_object_ref_add(&stripe->co_lu, "lov-parent", lov);
 160                 r0->lo_sub[idx] = cl2lovsub(stripe);
 161                 r0->lo_sub[idx]->lso_super = lov;
 162                 r0->lo_sub[idx]->lso_index = idx;
 163                 result = 0;
 164         } else {
 165                 struct lu_object  *old_obj;
 166                 struct lov_object *old_lov;
 167                 unsigned int mask = D_INODE;
 168
 169                 old_obj = lu_object_locate(&parent->coh_lu, &lov_device_type);
 170                 LASSERT(old_obj != NULL);
 171                 old_lov = cl2lov(lu2cl(old_obj));
 172                 if (old_lov->lo_layout_invalid) {
 173                         /* the object's layout has already changed but isn't
 174                          * refreshed */
 175                         lu_object_unhash(env, &stripe->co_lu);
 176                         result = -EAGAIN;
 177                 } else {
 178                         mask = D_ERROR;
 179                         result = -EIO;
 180                 }
 181
 182                 LU_OBJECT_DEBUG(mask, env, &stripe->co_lu,
 183                                 "stripe %d is already owned.\n", idx);
 184                 LU_OBJECT_DEBUG(mask, env, old_obj, "owned.\n");
 185                 LU_OBJECT_HEADER(mask, env, lov2lu(lov), "try to own.\n");
 186                 cl_object_put(env, stripe);
 187         }
 188         return result;
 189 }
 190
 191 static int lov_init_raid0(const struct lu_env *env,
 192                           struct lov_device *dev, struct lov_object *lov,
 193                           const struct cl_object_conf *conf,
 194                           union  lov_layout_state *state)
 195 {
 196         int result;
 197         int i;
 198
 199         struct cl_object        *stripe;
 200         struct lov_thread_info  *lti     = lov_env_info(env);
 201         struct cl_object_conf   *subconf = &lti->lti_stripe_conf;
 202         struct lov_stripe_md    *lsm     = conf->u.coc_md->lsm;
 203         struct lu_fid           *ofid    = &lti->lti_fid;
 204         struct lov_layout_raid0 *r0      = &state->raid0;
 205
 206         ENTRY;
 207
 208         if (lsm->lsm_magic != LOV_MAGIC_V1 && lsm->lsm_magic != LOV_MAGIC_V3) {
 209                 dump_lsm(D_ERROR, lsm);
 210                 LASSERTF(0, "magic mismatch, expected %d/%d, actual %d.\n",
 211                          LOV_MAGIC_V1, LOV_MAGIC_V3, lsm->lsm_magic);
 212         }
 213
 214         LASSERT(lov->lo_lsm == NULL);
 215         lov->lo_lsm = lsm_addref(lsm);
 216         r0->lo_nr  = lsm->lsm_stripe_count;
 217         LASSERT(r0->lo_nr <= lov_targets_nr(dev));
 218
 219         OBD_ALLOC_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 220         if (r0->lo_sub != NULL) {
 221                 result = 0;
 222                 subconf->coc_inode = conf->coc_inode;
 223                 spin_lock_init(&r0->lo_sub_lock);
 224                 /*
 225                  * Create stripe cl_objects.
 226                  */
 227                 for (i = 0; i < r0->lo_nr && result == 0; ++i) {
 228                         struct cl_device *subdev;
 229                         struct lov_oinfo *oinfo = lsm->lsm_oinfo[i];
 230                         int ost_idx = oinfo->loi_ost_idx;
 231
 232                         fid_ostid_unpack(ofid, &oinfo->loi_oi,
 233                                          oinfo->loi_ost_idx);
 234                         subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
 235                         subconf->u.coc_oinfo = oinfo;
 236                         LASSERTF(subdev != NULL, "not init ost %d\n", ost_idx);
 237                         /* In the function below, .hs_keycmp resolves to
 238                          * lu_obj_hop_keycmp() */
 239                         /* coverity[overrun-buffer-val] */
 240                         stripe = lov_sub_find(env, subdev, ofid, subconf);
 241                         if (!IS_ERR(stripe)) {
 242                                 result = lov_init_sub(env, lov, stripe, r0, i);
 243                                 if (result == -EAGAIN) { /* try again */
 244                                         --i;
 245                                         result = 0;
 246                                 }
 247                         } else {
 248                                 result = PTR_ERR(stripe);
 249                         }
 250                 }
 251         } else
 252                 result = -ENOMEM;
 253         RETURN(result);
 254 }
 255
 256 static int lov_delete_empty(const struct lu_env *env, struct lov_object *lov,
 257                             union lov_layout_state *state)
 258 {
 259         LASSERT(lov->lo_type == LLT_EMPTY);
 260         if (cfs_atomic_read(&lov->lo_active_ios) > 0)
 261                 RETURN(-EBUSY);
 262
 263         cl_object_prune(env, &lov->lo_cl);
 264         return 0;
 265 }
 266
 267 static void lov_subobject_kill(const struct lu_env *env, struct lov_object *lov,
 268                                struct lovsub_object *los, int idx)
 269 {
 270         struct cl_object        *sub;
 271         struct lov_layout_raid0 *r0;
 272         struct lu_site          *site;
 273         struct lu_site_bkt_data *bkt;
 274         cfs_waitlink_t          *waiter;
 275
 276         r0  = &lov->u.raid0;
 277         LASSERT(r0->lo_sub[idx] == los);
 278
 279         sub  = lovsub2cl(los);
 280         site = sub->co_lu.lo_dev->ld_site;
 281         bkt  = lu_site_bkt_from_fid(site, &sub->co_lu.lo_header->loh_fid);
 282
 283         cl_object_kill(env, sub);
 284         /* release a reference to the sub-object and ... */
 285         lu_object_ref_del(&sub->co_lu, "lov-parent", lov);
 286         cl_object_put(env, sub);
 287
 288         /* ... wait until it is actually destroyed---sub-object clears its
 289          * ->lo_sub[] slot in lovsub_object_fini() */
 290         if (r0->lo_sub[idx] == los) {
 291                 waiter = &lov_env_info(env)->lti_waiter;
 292                 cfs_waitlink_init(waiter);
 293                 cfs_waitq_add(&bkt->lsb_marche_funebre, waiter);
 294                 cfs_set_current_state(CFS_TASK_UNINT);
 295                 while (1) {
 296                         /* this wait-queue is signaled at the end of
 297                          * lu_object_free(). */
 298                         cfs_set_current_state(CFS_TASK_UNINT);
 299                         spin_lock(&r0->lo_sub_lock);
 300                         if (r0->lo_sub[idx] == los) {
 301                                 spin_unlock(&r0->lo_sub_lock);
 302                                 cfs_waitq_wait(waiter, CFS_TASK_UNINT);
 303                         } else {
 304                                 spin_unlock(&r0->lo_sub_lock);
 305                                 cfs_set_current_state(CFS_TASK_RUNNING);
 306                                 break;
 307                         }
 308                 }
 309                 cfs_waitq_del(&bkt->lsb_marche_funebre, waiter);
 310         }
 311         LASSERT(r0->lo_sub[idx] == NULL);
 312 }
 313
 314 static int lov_delete_raid0(const struct lu_env *env, struct lov_object *lov,
 315                             union lov_layout_state *state)
 316 {
 317         struct lov_layout_raid0 *r0 = &state->raid0;
 318         struct lov_stripe_md    *lsm = lov->lo_lsm;
 319         int i;
 320
 321         ENTRY;
 322
 323         dump_lsm(D_INODE, lsm);
 324         if (cfs_atomic_read(&lov->lo_active_ios) > 0)
 325                 RETURN(-EBUSY);
 326
 327         if (r0->lo_sub != NULL) {
 328                 for (i = 0; i < r0->lo_nr; ++i) {
 329                         struct lovsub_object *los = r0->lo_sub[i];
 330
 331                         if (los != NULL) {
 332                                 cl_locks_prune(env, &los->lso_cl, 1);
 333                                 /*
 334                                  * If top-level object is to be evicted from
 335                                  * the cache, so are its sub-objects.
 336                                  */
 337                                 lov_subobject_kill(env, lov, los, i);
 338                         }
 339                 }
 340         }
 341         cl_object_prune(env, &lov->lo_cl);
 342         RETURN(0);
 343 }
 344
 345 static void lov_fini_empty(const struct lu_env *env, struct lov_object *lov,
 346                            union lov_layout_state *state)
 347 {
 348         LASSERT(lov->lo_type == LLT_EMPTY);
 349 }
 350
 351 static void lov_fini_raid0(const struct lu_env *env, struct lov_object *lov,
 352                            union lov_layout_state *state)
 353 {
 354         struct lov_layout_raid0 *r0 = &state->raid0;
 355         ENTRY;
 356
 357         if (r0->lo_sub != NULL) {
 358                 OBD_FREE_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 359                 r0->lo_sub = NULL;
 360         }
 361
 362         dump_lsm(D_INODE, lov->lo_lsm);
 363         lov_free_memmd(&lov->lo_lsm);
 364
 365         EXIT;
 366 }
 367
 368 static int lov_print_empty(const struct lu_env *env, void *cookie,
 369                            lu_printer_t p, const struct lu_object *o)
 370 {
 371         (*p)(env, cookie, "empty %d\n", lu2lov(o)->lo_layout_invalid);
 372         return 0;
 373 }
 374
 375 static int lov_print_raid0(const struct lu_env *env, void *cookie,
 376                            lu_printer_t p, const struct lu_object *o)
 377 {
 378         struct lov_object       *lov = lu2lov(o);
 379         struct lov_layout_raid0 *r0  = lov_r0(lov);
 380         struct lov_stripe_md    *lsm = lov->lo_lsm;
 381         int i;
 382
 383         (*p)(env, cookie, "stripes: %d, %svalid, lsm{%p 0x%08X %d %u %u}: \n",
 384                 r0->lo_nr, lov->lo_layout_invalid ? "in" : "", lsm,
 385                 lsm->lsm_magic, cfs_atomic_read(&lsm->lsm_refc),
 386                 lsm->lsm_stripe_count, lsm->lsm_layout_gen);
 387         for (i = 0; i < r0->lo_nr; ++i) {
 388                 struct lu_object *sub;
 389
 390                 if (r0->lo_sub[i] != NULL) {
 391                         sub = lovsub2lu(r0->lo_sub[i]);
 392                         lu_object_print(env, cookie, p, sub);
 393                 } else
 394                         (*p)(env, cookie, "sub %d absent\n", i);
 395         }
 396         return 0;
 397 }
 398
 399 /**
 400  * Implements cl_object_operations::coo_attr_get() method for an object
 401  * without stripes (LLT_EMPTY layout type).
 402  *
 403  * The only attributes this layer is authoritative in this case is
 404  * cl_attr::cat_blocks---it's 0.
 405  */
 406 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
 407                               struct cl_attr *attr)
 408 {
 409         attr->cat_blocks = 0;
 410         return 0;
 411 }
 412
 413 static int lov_attr_get_raid0(const struct lu_env *env, struct cl_object *obj,
 414                               struct cl_attr *attr)
 415 {
 416         struct lov_object       *lov = cl2lov(obj);
 417         struct lov_layout_raid0 *r0 = lov_r0(lov);
 418         struct cl_attr          *lov_attr = &r0->lo_attr;
 419         int                      result = 0;
 420
 421         ENTRY;
 422
 423         /* this is called w/o holding type guard mutex, so it must be inside
 424          * an on going IO otherwise lsm may be replaced.
 425          * LU-2117: it turns out there exists one exception. For mmaped files,
 426          * the lock of those files may be requested in the other file's IO
 427          * context, and this function is called in ccc_lock_state(), it will
 428          * hit this assertion.
 429          * Anyway, it's still okay to call attr_get w/o type guard as layout
 430          * can't go if locks exist. */
 431         /* LASSERT(cfs_atomic_read(&lsm->lsm_refc) > 1); */
 432
 433         if (!r0->lo_attr_valid) {
 434                 struct lov_stripe_md    *lsm = lov->lo_lsm;
 435                 struct ost_lvb          *lvb = &lov_env_info(env)->lti_lvb;
 436                 __u64                    kms = 0;
 437
 438                 memset(lvb, 0, sizeof(*lvb));
 439                 /* XXX: timestamps can be negative by sanity:test_39m,
 440                  * how can it be? */
 441                 lvb->lvb_atime = LLONG_MIN;
 442                 lvb->lvb_ctime = LLONG_MIN;
 443                 lvb->lvb_mtime = LLONG_MIN;
 444
 445                 /*
 446                  * XXX that should be replaced with a loop over sub-objects,
 447                  * doing cl_object_attr_get() on them. But for now, let's
 448                  * reuse old lov code.
 449                  */
 450
 451                 /*
 452                  * XXX take lsm spin-lock to keep lov_merge_lvb_kms()
 453                  * happy. It's not needed, because new code uses
 454                  * ->coh_attr_guard spin-lock to protect consistency of
 455                  * sub-object attributes.
 456                  */
 457                 lov_stripe_lock(lsm);
 458                 result = lov_merge_lvb_kms(lsm, lvb, &kms);
 459                 lov_stripe_unlock(lsm);
 460                 if (result == 0) {
 461                         cl_lvb2attr(lov_attr, lvb);
 462                         lov_attr->cat_kms = kms;
 463                         r0->lo_attr_valid = 1;
 464                 }
 465         }
 466         if (result == 0) { /* merge results */
 467                 attr->cat_blocks = lov_attr->cat_blocks;
 468                 attr->cat_size = lov_attr->cat_size;
 469                 attr->cat_kms = lov_attr->cat_kms;
 470                 if (attr->cat_atime < lov_attr->cat_atime)
 471                         attr->cat_atime = lov_attr->cat_atime;
 472                 if (attr->cat_ctime < lov_attr->cat_ctime)
 473                         attr->cat_ctime = lov_attr->cat_ctime;
 474                 if (attr->cat_mtime < lov_attr->cat_mtime)
 475                         attr->cat_mtime = lov_attr->cat_mtime;
 476         }
 477         RETURN(result);
 478 }
 479
 480 const static struct lov_layout_operations lov_dispatch[] = {
 481         [LLT_EMPTY] = {
 482                 .llo_init      = lov_init_empty,
 483                 .llo_delete    = lov_delete_empty,
 484                 .llo_fini      = lov_fini_empty,
 485                 .llo_install   = lov_install_empty,
 486                 .llo_print     = lov_print_empty,
 487                 .llo_page_init = lov_page_init_empty,
 488                 .llo_lock_init = lov_lock_init_empty,
 489                 .llo_io_init   = lov_io_init_empty,
 490                 .llo_getattr   = lov_attr_get_empty
 491         },
 492         [LLT_RAID0] = {
 493                 .llo_init      = lov_init_raid0,
 494                 .llo_delete    = lov_delete_raid0,
 495                 .llo_fini      = lov_fini_raid0,
 496                 .llo_install   = lov_install_raid0,
 497                 .llo_print     = lov_print_raid0,
 498                 .llo_page_init = lov_page_init_raid0,
 499                 .llo_lock_init = lov_lock_init_raid0,
 500                 .llo_io_init   = lov_io_init_raid0,
 501                 .llo_getattr   = lov_attr_get_raid0
 502         }
 503 };
 504
 505
 506 /**
 507  * Performs a double-dispatch based on the layout type of an object.
 508  */
 509 #define LOV_2DISPATCH_NOLOCK(obj, op, ...)                              \
 510 ({                                                                      \
 511         struct lov_object                      *__obj = (obj);          \
 512         enum lov_layout_type                    __llt;                  \
 513                                                                         \
 514         __llt = __obj->lo_type;                                         \
 515         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 516         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 517 })
 518
 519 static inline void lov_conf_freeze(struct lov_object *lov)
 520 {
 521         if (lov->lo_owner != cfs_current())
 522                 down_read(&lov->lo_type_guard);
 523 }
 524
 525 static inline void lov_conf_thaw(struct lov_object *lov)
 526 {
 527         if (lov->lo_owner != cfs_current())
 528                 up_read(&lov->lo_type_guard);
 529 }
 530
 531 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...)                       \
 532 ({                                                                      \
 533         struct lov_object                      *__obj = (obj);          \
 534         int                                     __lock = !!(lock);      \
 535         typeof(lov_dispatch[0].op(__VA_ARGS__)) __result;               \
 536                                                                         \
 537         if (__lock)                                                     \
 538                 lov_conf_freeze(__obj);                                 \
 539         __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__);          \
 540         if (__lock)                                                     \
 541                 lov_conf_thaw(__obj);                                   \
 542         __result;                                                       \
 543 })
 544
 545 /**
 546  * Performs a locked double-dispatch based on the layout type of an object.
 547  */
 548 #define LOV_2DISPATCH(obj, op, ...)                     \
 549         LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
 550
 551 #define LOV_2DISPATCH_VOID(obj, op, ...)                                \
 552 do {                                                                    \
 553         struct lov_object                      *__obj = (obj);          \
 554         enum lov_layout_type                    __llt;                  \
 555                                                                         \
 556         lov_conf_freeze(__obj);                                         \
 557         __llt = __obj->lo_type;                                         \
 558         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 559         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 560         lov_conf_thaw(__obj);                                           \
 561 } while (0)
 562
 563 static void lov_conf_lock(struct lov_object *lov)
 564 {
 565         LASSERT(lov->lo_owner != cfs_current());
 566         down_write(&lov->lo_type_guard);
 567         LASSERT(lov->lo_owner == NULL);
 568         lov->lo_owner = cfs_current();
 569 }
 570
 571 static void lov_conf_unlock(struct lov_object *lov)
 572 {
 573         lov->lo_owner = NULL;
 574         up_write(&lov->lo_type_guard);
 575 }
 576
 577 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov)
 578 {
 579         struct l_wait_info lwi = { 0 };
 580         ENTRY;
 581
 582         if (!lov->lo_layout_invalid)
 583                 RETURN(0);
 584
 585         while (cfs_atomic_read(&lov->lo_active_ios) > 0) {
 586                 lov_conf_unlock(lov);
 587
 588                 CDEBUG(D_INODE, "file:"DFID" wait for active IO, now: %d.\n",
 589                         PFID(lu_object_fid(lov2lu(lov))),
 590                         cfs_atomic_read(&lov->lo_active_ios));
 591
 592                 l_wait_event(lov->lo_waitq,
 593                              cfs_atomic_read(&lov->lo_active_ios) == 0, &lwi);
 594                 lov_conf_lock(lov);
 595         }
 596         RETURN(0);
 597 }
 598
 599 static int lov_layout_change(const struct lu_env *unused,
 600                              struct lov_object *lov,
 601                              const struct cl_object_conf *conf)
 602 {
 603         int result;
 604         enum lov_layout_type llt = LLT_EMPTY;
 605         union lov_layout_state *state = &lov->u;
 606         const struct lov_layout_operations *old_ops;
 607         const struct lov_layout_operations *new_ops;
 608
 609         struct cl_object_header *hdr = cl_object_header(&lov->lo_cl);
 610         void *cookie;
 611         struct lu_env *env;
 612         int refcheck;
 613         ENTRY;
 614
 615         LASSERT(0 <= lov->lo_type && lov->lo_type < ARRAY_SIZE(lov_dispatch));
 616
 617         if (conf->u.coc_md != NULL && conf->u.coc_md->lsm != NULL)
 618                 llt = LLT_RAID0; /* only raid0 is supported. */
 619         LASSERT(0 <= llt && llt < ARRAY_SIZE(lov_dispatch));
 620
 621         cookie = cl_env_reenter();
 622         env = cl_env_get(&refcheck);
 623         if (IS_ERR(env)) {
 624                 cl_env_reexit(cookie);
 625                 RETURN(PTR_ERR(env));
 626         }
 627
 628         old_ops = &lov_dispatch[lov->lo_type];
 629         new_ops = &lov_dispatch[llt];
 630
 631         result = old_ops->llo_delete(env, lov, &lov->u);
 632         if (result == 0) {
 633                 old_ops->llo_fini(env, lov, &lov->u);
 634
 635                 LASSERT(cfs_atomic_read(&lov->lo_active_ios) == 0);
 636                 LASSERT(hdr->coh_tree.rnode == NULL);
 637                 LASSERT(hdr->coh_pages == 0);
 638
 639                 lov->lo_type = LLT_EMPTY;
 640                 result = new_ops->llo_init(env,
 641                                         lu2lov_dev(lov->lo_cl.co_lu.lo_dev),
 642                                         lov, conf, state);
 643                 if (result == 0) {
 644                         new_ops->llo_install(env, lov, state);
 645                         lov->lo_type = llt;
 646                 } else {
 647                         new_ops->llo_delete(env, lov, state);
 648                         new_ops->llo_fini(env, lov, state);
 649                         /* this file becomes an EMPTY file. */
 650                 }
 651         }
 652
 653         cl_env_put(env, &refcheck);
 654         cl_env_reexit(cookie);
 655         RETURN(result);
 656 }
 657
 658 /*****************************************************************************
 659  *
 660  * Lov object operations.
 661  *
 662  */
 663
 664 int lov_object_init(const struct lu_env *env, struct lu_object *obj,
 665                     const struct lu_object_conf *conf)
 666 {
 667         struct lov_device            *dev   = lu2lov_dev(obj->lo_dev);
 668         struct lov_object            *lov   = lu2lov(obj);
 669         const struct cl_object_conf  *cconf = lu2cl_conf(conf);
 670         union  lov_layout_state      *set   = &lov->u;
 671         const struct lov_layout_operations *ops;
 672         int result;
 673
 674         ENTRY;
 675         init_rwsem(&lov->lo_type_guard);
 676         cfs_atomic_set(&lov->lo_active_ios, 0);
 677         cfs_waitq_init(&lov->lo_waitq);
 678
 679         cl_object_page_init(lu2cl(obj), sizeof(struct lov_page));
 680
 681         /* no locking is necessary, as object is being created */
 682         lov->lo_type = cconf->u.coc_md->lsm != NULL ? LLT_RAID0 : LLT_EMPTY;
 683         ops = &lov_dispatch[lov->lo_type];
 684         result = ops->llo_init(env, dev, lov, cconf, set);
 685         if (result == 0)
 686                 ops->llo_install(env, lov, set);
 687         RETURN(result);
 688 }
 689
 690 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
 691                         const struct cl_object_conf *conf)
 692 {
 693         struct lov_stripe_md *lsm = NULL;
 694         struct lov_object *lov = cl2lov(obj);
 695         int result = 0;
 696         ENTRY;
 697
 698         lov_conf_lock(lov);
 699         if (conf->coc_opc == OBJECT_CONF_INVALIDATE) {
 700                 lov->lo_layout_invalid = true;
 701                 GOTO(out, result = 0);
 702         }
 703
 704         if (conf->coc_opc == OBJECT_CONF_WAIT) {
 705                 result = lov_layout_wait(env, lov);
 706                 GOTO(out, result);
 707         }
 708
 709         LASSERT(conf->coc_opc == OBJECT_CONF_SET);
 710
 711         if (conf->u.coc_md != NULL)
 712                 lsm = conf->u.coc_md->lsm;
 713         if ((lsm == NULL && lov->lo_lsm == NULL) ||
 714             (lsm != NULL && lov->lo_lsm != NULL &&
 715              lov->lo_lsm->lsm_layout_gen == lsm->lsm_layout_gen)) {
 716                 /* same version of layout */
 717                 lov->lo_layout_invalid = false;
 718                 GOTO(out, result = 0);
 719         }
 720
 721         /* will change layout - check if there still exists active IO. */
 722         if (cfs_atomic_read(&lov->lo_active_ios) > 1) {
 723                 lov->lo_layout_invalid = true;
 724                 GOTO(out, result = -EBUSY);
 725         }
 726
 727         lov->lo_layout_invalid = lov_layout_change(env, lov, conf);
 728         EXIT;
 729
 730 out:
 731         lov_conf_unlock(lov);
 732         RETURN(result);
 733 }
 734
 735 static void lov_object_delete(const struct lu_env *env, struct lu_object *obj)
 736 {
 737         struct lov_object *lov = lu2lov(obj);
 738
 739         ENTRY;
 740         LOV_2DISPATCH_VOID(lov, llo_delete, env, lov, &lov->u);
 741         EXIT;
 742 }
 743
 744 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
 745 {
 746         struct lov_object *lov = lu2lov(obj);
 747
 748         ENTRY;
 749         LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
 750         lu_object_fini(obj);
 751         OBD_SLAB_FREE_PTR(lov, lov_object_kmem);
 752         EXIT;
 753 }
 754
 755 static int lov_object_print(const struct lu_env *env, void *cookie,
 756                             lu_printer_t p, const struct lu_object *o)
 757 {
 758         return LOV_2DISPATCH_NOLOCK(lu2lov(o), llo_print, env, cookie, p, o);
 759 }
 760
 761 int lov_page_init(const struct lu_env *env, struct cl_object *obj,
 762                 struct cl_page *page, cfs_page_t *vmpage)
 763 {
 764         return LOV_2DISPATCH_NOLOCK(cl2lov(obj),
 765                                     llo_page_init, env, obj, page, vmpage);
 766 }
 767
 768 /**
 769  * Implements cl_object_operations::clo_io_init() method for lov
 770  * layer. Dispatches to the appropriate layout io initialization method.
 771  */
 772 int lov_io_init(const struct lu_env *env, struct cl_object *obj,
 773                 struct cl_io *io)
 774 {
 775         CL_IO_SLICE_CLEAN(lov_env_io(env), lis_cl);
 776         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
 777                                      !io->ci_ignore_layout, env, obj, io);
 778 }
 779
 780 /**
 781  * An implementation of cl_object_operations::clo_attr_get() method for lov
 782  * layer. For raid0 layout this collects and merges attributes of all
 783  * sub-objects.
 784  */
 785 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
 786                         struct cl_attr *attr)
 787 {
 788         /* do not take lock, as this function is called under a
 789          * spin-lock. Layout is protected from changing by ongoing IO. */
 790         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
 791 }
 792
 793 static int lov_attr_set(const struct lu_env *env, struct cl_object *obj,
 794                         const struct cl_attr *attr, unsigned valid)
 795 {
 796         /*
 797          * No dispatch is required here, as no layout implements this.
 798          */
 799         return 0;
 800 }
 801
 802 int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
 803                   struct cl_lock *lock, const struct cl_io *io)
 804 {
 805         /* No need to lock because we've taken one refcount of layout.  */
 806         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_lock_init, env, obj, lock,
 807                                     io);
 808 }
 809
 810 static const struct cl_object_operations lov_ops = {
 811         .coo_page_init = lov_page_init,
 812         .coo_lock_init = lov_lock_init,
 813         .coo_io_init   = lov_io_init,
 814         .coo_attr_get  = lov_attr_get,
 815         .coo_attr_set  = lov_attr_set,
 816         .coo_conf_set  = lov_conf_set
 817 };
 818
 819 static const struct lu_object_operations lov_lu_obj_ops = {
 820         .loo_object_init      = lov_object_init,
 821         .loo_object_delete    = lov_object_delete,
 822         .loo_object_release   = NULL,
 823         .loo_object_free      = lov_object_free,
 824         .loo_object_print     = lov_object_print,
 825         .loo_object_invariant = NULL
 826 };
 827
 828 struct lu_object *lov_object_alloc(const struct lu_env *env,
 829                                    const struct lu_object_header *unused,
 830                                    struct lu_device *dev)
 831 {
 832         struct lov_object *lov;
 833         struct lu_object  *obj;
 834
 835         ENTRY;
 836         OBD_SLAB_ALLOC_PTR_GFP(lov, lov_object_kmem, CFS_ALLOC_IO);
 837         if (lov != NULL) {
 838                 obj = lov2lu(lov);
 839                 lu_object_init(obj, NULL, dev);
 840                 lov->lo_cl.co_ops = &lov_ops;
 841                 lov->lo_type = -1; /* invalid, to catch uninitialized type */
 842                 /*
 843                  * object io operation vector (cl_object::co_iop) is installed
 844                  * later in lov_object_init(), as different vectors are used
 845                  * for object with different layouts.
 846                  */
 847                 obj->lo_ops = &lov_lu_obj_ops;
 848         } else
 849                 obj = NULL;
 850         RETURN(obj);
 851 }
 852
 853 struct lov_stripe_md *lov_lsm_addref(struct lov_object *lov)
 854 {
 855         struct lov_stripe_md *lsm = NULL;
 856
 857         lov_conf_freeze(lov);
 858         if (lov->lo_lsm != NULL) {
 859                 lsm = lsm_addref(lov->lo_lsm);
 860                 CDEBUG(D_INODE, "lsm %p addref %d/%d by %p.\n",
 861                         lsm, cfs_atomic_read(&lsm->lsm_refc),
 862                         lov->lo_layout_invalid, cfs_current());
 863         }
 864         lov_conf_thaw(lov);
 865         return lsm;
 866 }
 867
 868 void lov_lsm_decref(struct lov_object *lov, struct lov_stripe_md *lsm)
 869 {
 870         if (lsm == NULL)
 871                 return;
 872
 873         CDEBUG(D_INODE, "lsm %p decref %d by %p.\n",
 874                 lsm, cfs_atomic_read(&lsm->lsm_refc), cfs_current());
 875
 876         lov_free_memmd(&lsm);
 877 }
 878
 879 struct lov_stripe_md *lov_lsm_get(struct cl_object *clobj)
 880 {
 881         struct lu_object *luobj;
 882         struct lov_stripe_md *lsm = NULL;
 883
 884         if (clobj == NULL)
 885                 return NULL;
 886
 887         luobj = lu_object_locate(&cl_object_header(clobj)->coh_lu,
 888                                  &lov_device_type);
 889         if (luobj != NULL)
 890                 lsm = lov_lsm_addref(lu2lov(luobj));
 891         return lsm;
 892 }
 893 EXPORT_SYMBOL(lov_lsm_get);
 894
 895 void lov_lsm_put(struct cl_object *unused, struct lov_stripe_md *lsm)
 896 {
 897         if (lsm != NULL)
 898                 lov_free_memmd(&lsm);
 899 }
 900 EXPORT_SYMBOL(lov_lsm_put);
 901
 902 int lov_read_and_clear_async_rc(struct cl_object *clob)
 903 {
 904         struct lu_object *luobj;
 905         int rc = 0;
 906         ENTRY;
 907
 908         luobj = lu_object_locate(&cl_object_header(clob)->coh_lu,
 909                                  &lov_device_type);
 910         if (luobj != NULL) {
 911                 struct lov_object *lov = lu2lov(luobj);
 912
 913                 lov_conf_freeze(lov);
 914                 switch (lov->lo_type) {
 915                 case LLT_RAID0: {
 916                         struct lov_stripe_md *lsm;
 917                         int i;
 918
 919                         lsm = lov->lo_lsm;
 920                         LASSERT(lsm != NULL);
 921                         for (i = 0; i < lsm->lsm_stripe_count; i++) {
 922                                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
 923                                 if (loi->loi_ar.ar_rc && !rc)
 924                                         rc = loi->loi_ar.ar_rc;
 925                                 loi->loi_ar.ar_rc = 0;
 926                         }
 927                 }
 928                 case LLT_EMPTY:
 929                         break;
 930                 default:
 931                         LBUG();
 932                 }
 933                 lov_conf_thaw(lov);
 934         }
 935         RETURN(rc);
 936 }
 937 EXPORT_SYMBOL(lov_read_and_clear_async_rc);
 938
 939 /** @} lov */