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