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
  46 /** \addtogroup lov
  47  *  @{
  48  */
  49
  50 /*****************************************************************************
  51  *
  52  * Layout operations.
  53  *
  54  */
  55
  56 struct lov_layout_operations {
  57         int (*llo_init)(const struct lu_env *env, struct lov_device *dev,
  58                         struct lov_object *lov,
  59                         const struct cl_object_conf *conf,
  60                         union lov_layout_state *state);
  61         int (*llo_delete)(const struct lu_env *env, struct lov_object *lov,
  62                            union lov_layout_state *state);
  63         void (*llo_fini)(const struct lu_env *env, struct lov_object *lov,
  64                          union lov_layout_state *state);
  65         void (*llo_install)(const struct lu_env *env, struct lov_object *lov,
  66                             union lov_layout_state *state);
  67         int  (*llo_print)(const struct lu_env *env, void *cookie,
  68                           lu_printer_t p, const struct lu_object *o);
  69         int  (*llo_page_init)(const struct lu_env *env, struct cl_object *obj,
  70                                 struct cl_page *page, struct page *vmpage);
  71         int  (*llo_lock_init)(const struct lu_env *env,
  72                               struct cl_object *obj, struct cl_lock *lock,
  73                               const struct cl_io *io);
  74         int  (*llo_io_init)(const struct lu_env *env,
  75                             struct cl_object *obj, struct cl_io *io);
  76         int  (*llo_getattr)(const struct lu_env *env, struct cl_object *obj,
  77                             struct cl_attr *attr);
  78 };
  79
  80 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov);
  81
  82 /*****************************************************************************
  83  *
  84  * Lov object layout operations.
  85  *
  86  */
  87
  88 static void lov_install_empty(const struct lu_env *env,
  89                               struct lov_object *lov,
  90                               union  lov_layout_state *state)
  91 {
  92         /*
  93          * File without objects.
  94          */
  95 }
  96
  97 static int lov_init_empty(const struct lu_env *env,
  98                           struct lov_device *dev, struct lov_object *lov,
  99                           const struct cl_object_conf *conf,
 100                           union  lov_layout_state *state)
 101 {
 102         return 0;
 103 }
 104
 105 static void lov_install_raid0(const struct lu_env *env,
 106                               struct lov_object *lov,
 107                               union  lov_layout_state *state)
 108 {
 109 }
 110
 111 static struct cl_object *lov_sub_find(const struct lu_env *env,
 112                                       struct cl_device *dev,
 113                                       const struct lu_fid *fid,
 114                                       const struct cl_object_conf *conf)
 115 {
 116         struct lu_object *o;
 117
 118         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, struct lov_layout_raid0 *r0,
 125                         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
 147         oinfo = lov->lo_lsm->lsm_oinfo[idx];
 148         CDEBUG(D_INODE, DFID"@%p[%d] -> "DFID"@%p: ostid: "DOSTID
 149                " idx: %d gen: %d\n",
 150                PFID(&subhdr->coh_lu.loh_fid), subhdr, idx,
 151                PFID(&hdr->coh_lu.loh_fid), hdr, POSTID(&oinfo->loi_oi),
 152                oinfo->loi_ost_idx, oinfo->loi_ost_gen);
 153
 154         /* reuse ->coh_attr_guard to protect coh_parent change */
 155         spin_lock(&subhdr->coh_attr_guard);
 156         parent = subhdr->coh_parent;
 157         if (parent == NULL) {
 158                 subhdr->coh_parent = hdr;
 159                 spin_unlock(&subhdr->coh_attr_guard);
 160                 subhdr->coh_nesting = hdr->coh_nesting + 1;
 161                 lu_object_ref_add(&stripe->co_lu, "lov-parent", lov);
 162                 r0->lo_sub[idx] = cl2lovsub(stripe);
 163                 r0->lo_sub[idx]->lso_super = lov;
 164                 r0->lo_sub[idx]->lso_index = idx;
 165                 result = 0;
 166         } else {
 167                 struct lu_object  *old_obj;
 168                 struct lov_object *old_lov;
 169                 unsigned int mask = D_INODE;
 170
 171                 spin_unlock(&subhdr->coh_attr_guard);
 172                 old_obj = lu_object_locate(&parent->coh_lu, &lov_device_type);
 173                 LASSERT(old_obj != NULL);
 174                 old_lov = cl2lov(lu2cl(old_obj));
 175                 if (old_lov->lo_layout_invalid) {
 176                         /* the object's layout has already changed but isn't
 177                          * refreshed */
 178                         lu_object_unhash(env, &stripe->co_lu);
 179                         result = -EAGAIN;
 180                 } else {
 181                         mask = D_ERROR;
 182                         result = -EIO;
 183                 }
 184
 185                 LU_OBJECT_DEBUG(mask, env, &stripe->co_lu,
 186                                 "stripe %d is already owned.\n", idx);
 187                 LU_OBJECT_DEBUG(mask, env, old_obj, "owned.\n");
 188                 LU_OBJECT_HEADER(mask, env, lov2lu(lov), "try to own.\n");
 189                 cl_object_put(env, stripe);
 190         }
 191         return result;
 192 }
 193
 194 static int lov_init_raid0(const struct lu_env *env,
 195                           struct lov_device *dev, struct lov_object *lov,
 196                           const struct cl_object_conf *conf,
 197                           union  lov_layout_state *state)
 198 {
 199         int result;
 200         int i;
 201
 202         struct cl_object        *stripe;
 203         struct lov_thread_info  *lti     = lov_env_info(env);
 204         struct cl_object_conf   *subconf = &lti->lti_stripe_conf;
 205         struct lov_stripe_md    *lsm     = conf->u.coc_md->lsm;
 206         struct lu_fid      *ofid    = &lti->lti_fid;
 207         struct lov_layout_raid0 *r0      = &state->raid0;
 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                         if (lov_oinfo_is_dummy(oinfo))
 234                                 continue;
 235
 236                         result = ostid_to_fid(ofid, &oinfo->loi_oi,
 237                                               oinfo->loi_ost_idx);
 238                         if (result != 0)
 239                                 goto out;
 240
 241                         subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
 242                         subconf->u.coc_oinfo = oinfo;
 243                         LASSERTF(subdev != NULL, "not init ost %d\n", ost_idx);
 244                         /* In the function below, .hs_keycmp resolves to
 245                          * lu_obj_hop_keycmp() */
 246                         /* coverity[overrun-buffer-val] */
 247                         stripe = lov_sub_find(env, subdev, ofid, subconf);
 248                         if (!IS_ERR(stripe)) {
 249                                 result = lov_init_sub(env, lov, stripe, r0, i);
 250                                 if (result == -EAGAIN) { /* try again */
 251                                         --i;
 252                                         result = 0;
 253                                 }
 254                         } else {
 255                                 result = PTR_ERR(stripe);
 256                         }
 257                 }
 258         } else
 259                 result = -ENOMEM;
 260 out:
 261         return result;
 262 }
 263
 264 static int lov_init_released(const struct lu_env *env,
 265                         struct lov_device *dev, struct lov_object *lov,
 266                         const struct cl_object_conf *conf,
 267                         union  lov_layout_state *state)
 268 {
 269         struct lov_stripe_md *lsm = conf->u.coc_md->lsm;
 270
 271         LASSERT(lsm != NULL);
 272         LASSERT(lsm_is_released(lsm));
 273         LASSERT(lov->lo_lsm == NULL);
 274
 275         lov->lo_lsm = lsm_addref(lsm);
 276         return 0;
 277 }
 278
 279 static int lov_delete_empty(const struct lu_env *env, struct lov_object *lov,
 280                             union lov_layout_state *state)
 281 {
 282         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 283
 284         lov_layout_wait(env, lov);
 285
 286         cl_object_prune(env, &lov->lo_cl);
 287         return 0;
 288 }
 289
 290 static void lov_subobject_kill(const struct lu_env *env, struct lov_object *lov,
 291                                struct lovsub_object *los, int idx)
 292 {
 293         struct cl_object        *sub;
 294         struct lov_layout_raid0 *r0;
 295         struct lu_site    *site;
 296         struct lu_site_bkt_data *bkt;
 297         wait_queue_t      *waiter;
 298
 299         r0  = &lov->u.raid0;
 300         LASSERT(r0->lo_sub[idx] == los);
 301
 302         sub  = lovsub2cl(los);
 303         site = sub->co_lu.lo_dev->ld_site;
 304         bkt  = lu_site_bkt_from_fid(site, &sub->co_lu.lo_header->loh_fid);
 305
 306         cl_object_kill(env, sub);
 307         /* release a reference to the sub-object and ... */
 308         lu_object_ref_del(&sub->co_lu, "lov-parent", lov);
 309         cl_object_put(env, sub);
 310
 311         /* ... wait until it is actually destroyed---sub-object clears its
 312          * ->lo_sub[] slot in lovsub_object_fini() */
 313         if (r0->lo_sub[idx] == los) {
 314                 waiter = &lov_env_info(env)->lti_waiter;
 315                 init_waitqueue_entry(waiter, current);
 316                 add_wait_queue(&bkt->lsb_marche_funebre, waiter);
 317                 set_current_state(TASK_UNINTERRUPTIBLE);
 318                 while (1) {
 319                         /* this wait-queue is signaled at the end of
 320                          * lu_object_free(). */
 321                         set_current_state(TASK_UNINTERRUPTIBLE);
 322                         spin_lock(&r0->lo_sub_lock);
 323                         if (r0->lo_sub[idx] == los) {
 324                                 spin_unlock(&r0->lo_sub_lock);
 325                                 schedule();
 326                         } else {
 327                                 spin_unlock(&r0->lo_sub_lock);
 328                                 set_current_state(TASK_RUNNING);
 329                                 break;
 330                         }
 331                 }
 332                 remove_wait_queue(&bkt->lsb_marche_funebre, waiter);
 333         }
 334         LASSERT(r0->lo_sub[idx] == NULL);
 335 }
 336
 337 static int lov_delete_raid0(const struct lu_env *env, struct lov_object *lov,
 338                             union lov_layout_state *state)
 339 {
 340         struct lov_layout_raid0 *r0 = &state->raid0;
 341         struct lov_stripe_md    *lsm = lov->lo_lsm;
 342         int i;
 343
 344         dump_lsm(D_INODE, lsm);
 345
 346         lov_layout_wait(env, lov);
 347         if (r0->lo_sub != NULL) {
 348                 for (i = 0; i < r0->lo_nr; ++i) {
 349                         struct lovsub_object *los = r0->lo_sub[i];
 350
 351                         if (los != NULL) {
 352                                 cl_locks_prune(env, &los->lso_cl, 1);
 353                                 /*
 354                                  * If top-level object is to be evicted from
 355                                  * the cache, so are its sub-objects.
 356                                  */
 357                                 lov_subobject_kill(env, lov, los, i);
 358                         }
 359                 }
 360         }
 361         cl_object_prune(env, &lov->lo_cl);
 362         return 0;
 363 }
 364
 365 static void lov_fini_empty(const struct lu_env *env, struct lov_object *lov,
 366                            union lov_layout_state *state)
 367 {
 368         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 369 }
 370
 371 static void lov_fini_raid0(const struct lu_env *env, struct lov_object *lov,
 372                            union lov_layout_state *state)
 373 {
 374         struct lov_layout_raid0 *r0 = &state->raid0;
 375
 376         if (r0->lo_sub != NULL) {
 377                 OBD_FREE_LARGE(r0->lo_sub, r0->lo_nr * sizeof(r0->lo_sub[0]));
 378                 r0->lo_sub = NULL;
 379         }
 380
 381         dump_lsm(D_INODE, lov->lo_lsm);
 382         lov_free_memmd(&lov->lo_lsm);
 383 }
 384
 385 static void lov_fini_released(const struct lu_env *env, struct lov_object *lov,
 386                                 union lov_layout_state *state)
 387 {
 388         dump_lsm(D_INODE, lov->lo_lsm);
 389         lov_free_memmd(&lov->lo_lsm);
 390 }
 391
 392 static int lov_print_empty(const struct lu_env *env, void *cookie,
 393                            lu_printer_t p, const struct lu_object *o)
 394 {
 395         (*p)(env, cookie, "empty %d\n", lu2lov(o)->lo_layout_invalid);
 396         return 0;
 397 }
 398
 399 static int lov_print_raid0(const struct lu_env *env, void *cookie,
 400                            lu_printer_t p, const struct lu_object *o)
 401 {
 402         struct lov_object       *lov = lu2lov(o);
 403         struct lov_layout_raid0 *r0  = lov_r0(lov);
 404         struct lov_stripe_md    *lsm = lov->lo_lsm;
 405         int                      i;
 406
 407         (*p)(env, cookie, "stripes: %d, %s, lsm{%p 0x%08X %d %u %u}:\n",
 408                 r0->lo_nr, lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 409                 lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 410                 lsm->lsm_stripe_count, lsm->lsm_layout_gen);
 411         for (i = 0; i < r0->lo_nr; ++i) {
 412                 struct lu_object *sub;
 413
 414                 if (r0->lo_sub[i] != NULL) {
 415                         sub = lovsub2lu(r0->lo_sub[i]);
 416                         lu_object_print(env, cookie, p, sub);
 417                 } else {
 418                         (*p)(env, cookie, "sub %d absent\n", i);
 419                 }
 420         }
 421         return 0;
 422 }
 423
 424 static int lov_print_released(const struct lu_env *env, void *cookie,
 425                                 lu_printer_t p, const struct lu_object *o)
 426 {
 427         struct lov_object       *lov = lu2lov(o);
 428         struct lov_stripe_md    *lsm = lov->lo_lsm;
 429
 430         (*p)(env, cookie,
 431                 "released: %s, lsm{%p 0x%08X %d %u %u}:\n",
 432                 lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 433                 lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 434                 lsm->lsm_stripe_count, lsm->lsm_layout_gen);
 435         return 0;
 436 }
 437
 438 /**
 439  * Implements cl_object_operations::coo_attr_get() method for an object
 440  * without stripes (LLT_EMPTY layout type).
 441  *
 442  * The only attributes this layer is authoritative in this case is
 443  * cl_attr::cat_blocks---it's 0.
 444  */
 445 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
 446                               struct cl_attr *attr)
 447 {
 448         attr->cat_blocks = 0;
 449         return 0;
 450 }
 451
 452 static int lov_attr_get_raid0(const struct lu_env *env, struct cl_object *obj,
 453                               struct cl_attr *attr)
 454 {
 455         struct lov_object       *lov = cl2lov(obj);
 456         struct lov_layout_raid0 *r0 = lov_r0(lov);
 457         struct cl_attr          *lov_attr = &r0->lo_attr;
 458         int                      result = 0;
 459
 460         /* this is called w/o holding type guard mutex, so it must be inside
 461          * an on going IO otherwise lsm may be replaced.
 462          * LU-2117: it turns out there exists one exception. For mmaped files,
 463          * the lock of those files may be requested in the other file's IO
 464          * context, and this function is called in ccc_lock_state(), it will
 465          * hit this assertion.
 466          * Anyway, it's still okay to call attr_get w/o type guard as layout
 467          * can't go if locks exist. */
 468         /* LASSERT(atomic_read(&lsm->lsm_refc) > 1); */
 469
 470         if (!r0->lo_attr_valid) {
 471                 struct lov_stripe_md    *lsm = lov->lo_lsm;
 472                 struct ost_lvb    *lvb = &lov_env_info(env)->lti_lvb;
 473                 __u64               kms = 0;
 474
 475                 memset(lvb, 0, sizeof(*lvb));
 476                 /* XXX: timestamps can be negative by sanity:test_39m,
 477                  * how can it be? */
 478                 lvb->lvb_atime = LLONG_MIN;
 479                 lvb->lvb_ctime = LLONG_MIN;
 480                 lvb->lvb_mtime = LLONG_MIN;
 481
 482                 /*
 483                  * XXX that should be replaced with a loop over sub-objects,
 484                  * doing cl_object_attr_get() on them. But for now, let's
 485                  * reuse old lov code.
 486                  */
 487
 488                 /*
 489                  * XXX take lsm spin-lock to keep lov_merge_lvb_kms()
 490                  * happy. It's not needed, because new code uses
 491                  * ->coh_attr_guard spin-lock to protect consistency of
 492                  * sub-object attributes.
 493                  */
 494                 lov_stripe_lock(lsm);
 495                 result = lov_merge_lvb_kms(lsm, lvb, &kms);
 496                 lov_stripe_unlock(lsm);
 497                 if (result == 0) {
 498                         cl_lvb2attr(lov_attr, lvb);
 499                         lov_attr->cat_kms = kms;
 500                         r0->lo_attr_valid = 1;
 501                 }
 502         }
 503         if (result == 0) { /* merge results */
 504                 attr->cat_blocks = lov_attr->cat_blocks;
 505                 attr->cat_size = lov_attr->cat_size;
 506                 attr->cat_kms = lov_attr->cat_kms;
 507                 if (attr->cat_atime < lov_attr->cat_atime)
 508                         attr->cat_atime = lov_attr->cat_atime;
 509                 if (attr->cat_ctime < lov_attr->cat_ctime)
 510                         attr->cat_ctime = lov_attr->cat_ctime;
 511                 if (attr->cat_mtime < lov_attr->cat_mtime)
 512                         attr->cat_mtime = lov_attr->cat_mtime;
 513         }
 514         return result;
 515 }
 516
 517 static const struct lov_layout_operations lov_dispatch[] = {
 518         [LLT_EMPTY] = {
 519                 .llo_init      = lov_init_empty,
 520                 .llo_delete    = lov_delete_empty,
 521                 .llo_fini      = lov_fini_empty,
 522                 .llo_install   = lov_install_empty,
 523                 .llo_print     = lov_print_empty,
 524                 .llo_page_init = lov_page_init_empty,
 525                 .llo_lock_init = lov_lock_init_empty,
 526                 .llo_io_init   = lov_io_init_empty,
 527                 .llo_getattr   = lov_attr_get_empty
 528         },
 529         [LLT_RAID0] = {
 530                 .llo_init      = lov_init_raid0,
 531                 .llo_delete    = lov_delete_raid0,
 532                 .llo_fini      = lov_fini_raid0,
 533                 .llo_install   = lov_install_raid0,
 534                 .llo_print     = lov_print_raid0,
 535                 .llo_page_init = lov_page_init_raid0,
 536                 .llo_lock_init = lov_lock_init_raid0,
 537                 .llo_io_init   = lov_io_init_raid0,
 538                 .llo_getattr   = lov_attr_get_raid0
 539         },
 540         [LLT_RELEASED] = {
 541                 .llo_init      = lov_init_released,
 542                 .llo_delete    = lov_delete_empty,
 543                 .llo_fini      = lov_fini_released,
 544                 .llo_install   = lov_install_empty,
 545                 .llo_print     = lov_print_released,
 546                 .llo_page_init = lov_page_init_empty,
 547                 .llo_lock_init = lov_lock_init_empty,
 548                 .llo_io_init   = lov_io_init_released,
 549                 .llo_getattr   = lov_attr_get_empty
 550         }
 551 };
 552
 553 /**
 554  * Performs a double-dispatch based on the layout type of an object.
 555  */
 556 #define LOV_2DISPATCH_NOLOCK(obj, op, ...)                            \
 557 ({                                                                    \
 558         struct lov_object                     *__obj = (obj);     \
 559         enum lov_layout_type                __llt;                \
 560                                                                         \
 561         __llt = __obj->lo_type;                                  \
 562         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 563         lov_dispatch[__llt].op(__VA_ARGS__);                        \
 564 })
 565
 566 /**
 567  * Return lov_layout_type associated with a given lsm
 568  */
 569 static enum lov_layout_type lov_type(struct lov_stripe_md *lsm)
 570 {
 571         if (lsm == NULL)
 572                 return LLT_EMPTY;
 573         if (lsm_is_released(lsm))
 574                 return LLT_RELEASED;
 575         return LLT_RAID0;
 576 }
 577
 578 static inline void lov_conf_freeze(struct lov_object *lov)
 579 {
 580         if (lov->lo_owner != current)
 581                 down_read(&lov->lo_type_guard);
 582 }
 583
 584 static inline void lov_conf_thaw(struct lov_object *lov)
 585 {
 586         if (lov->lo_owner != current)
 587                 up_read(&lov->lo_type_guard);
 588 }
 589
 590 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...)                      \
 591 ({                                                                    \
 592         struct lov_object                     *__obj = (obj);     \
 593         int                                  __lock = !!(lock);      \
 594         typeof(lov_dispatch[0].op(__VA_ARGS__)) __result;              \
 595                                                                         \
 596         if (__lock)                                                  \
 597                 lov_conf_freeze(__obj);                                 \
 598         __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__);    \
 599         if (__lock)                                                  \
 600                 lov_conf_thaw(__obj);                                   \
 601         __result;                                                      \
 602 })
 603
 604 /**
 605  * Performs a locked double-dispatch based on the layout type of an object.
 606  */
 607 #define LOV_2DISPATCH(obj, op, ...)                  \
 608         LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
 609
 610 #define LOV_2DISPATCH_VOID(obj, op, ...)                                \
 611 do {                                                                \
 612         struct lov_object                     *__obj = (obj);     \
 613         enum lov_layout_type                __llt;                \
 614                                                                         \
 615         lov_conf_freeze(__obj);                                         \
 616         __llt = __obj->lo_type;                                  \
 617         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 618         lov_dispatch[__llt].op(__VA_ARGS__);                        \
 619         lov_conf_thaw(__obj);                                           \
 620 } while (0)
 621
 622 static void lov_conf_lock(struct lov_object *lov)
 623 {
 624         LASSERT(lov->lo_owner != current);
 625         down_write(&lov->lo_type_guard);
 626         LASSERT(lov->lo_owner == NULL);
 627         lov->lo_owner = current;
 628 }
 629
 630 static void lov_conf_unlock(struct lov_object *lov)
 631 {
 632         lov->lo_owner = NULL;
 633         up_write(&lov->lo_type_guard);
 634 }
 635
 636 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov)
 637 {
 638         struct l_wait_info lwi = { 0 };
 639
 640         while (atomic_read(&lov->lo_active_ios) > 0) {
 641                 CDEBUG(D_INODE, "file:"DFID" wait for active IO, now: %d.\n",
 642                         PFID(lu_object_fid(lov2lu(lov))),
 643                         atomic_read(&lov->lo_active_ios));
 644
 645                 l_wait_event(lov->lo_waitq,
 646                              atomic_read(&lov->lo_active_ios) == 0, &lwi);
 647         }
 648         return 0;
 649 }
 650
 651 static int lov_layout_change(const struct lu_env *unused,
 652                              struct lov_object *lov,
 653                              const struct cl_object_conf *conf)
 654 {
 655         int result;
 656         enum lov_layout_type llt = LLT_EMPTY;
 657         union lov_layout_state *state = &lov->u;
 658         const struct lov_layout_operations *old_ops;
 659         const struct lov_layout_operations *new_ops;
 660
 661         struct cl_object_header *hdr = cl_object_header(&lov->lo_cl);
 662         void *cookie;
 663         struct lu_env *env;
 664         int refcheck;
 665
 666         LASSERT(0 <= lov->lo_type && lov->lo_type < ARRAY_SIZE(lov_dispatch));
 667
 668         if (conf->u.coc_md != NULL)
 669                 llt = lov_type(conf->u.coc_md->lsm);
 670         LASSERT(0 <= llt && llt < ARRAY_SIZE(lov_dispatch));
 671
 672         cookie = cl_env_reenter();
 673         env = cl_env_get(&refcheck);
 674         if (IS_ERR(env)) {
 675                 cl_env_reexit(cookie);
 676                 return PTR_ERR(env);
 677         }
 678
 679         CDEBUG(D_INODE, DFID" from %s to %s\n",
 680                PFID(lu_object_fid(lov2lu(lov))),
 681                llt2str(lov->lo_type), llt2str(llt));
 682
 683         old_ops = &lov_dispatch[lov->lo_type];
 684         new_ops = &lov_dispatch[llt];
 685
 686         result = old_ops->llo_delete(env, lov, &lov->u);
 687         if (result == 0) {
 688                 old_ops->llo_fini(env, lov, &lov->u);
 689
 690                 LASSERT(atomic_read(&lov->lo_active_ios) == 0);
 691                 LASSERT(hdr->coh_tree.rnode == NULL);
 692                 LASSERT(hdr->coh_pages == 0);
 693
 694                 lov->lo_type = LLT_EMPTY;
 695                 result = new_ops->llo_init(env,
 696                                         lu2lov_dev(lov->lo_cl.co_lu.lo_dev),
 697                                         lov, conf, state);
 698                 if (result == 0) {
 699                         new_ops->llo_install(env, lov, state);
 700                         lov->lo_type = llt;
 701                 } else {
 702                         new_ops->llo_delete(env, lov, state);
 703                         new_ops->llo_fini(env, lov, state);
 704                         /* this file becomes an EMPTY file. */
 705                 }
 706         }
 707
 708         cl_env_put(env, &refcheck);
 709         cl_env_reexit(cookie);
 710         return result;
 711 }
 712
 713 /*****************************************************************************
 714  *
 715  * Lov object operations.
 716  *
 717  */
 718 int lov_object_init(const struct lu_env *env, struct lu_object *obj,
 719                     const struct lu_object_conf *conf)
 720 {
 721         struct lov_device           *dev   = lu2lov_dev(obj->lo_dev);
 722         struct lov_object           *lov   = lu2lov(obj);
 723         const struct cl_object_conf  *cconf = lu2cl_conf(conf);
 724         union  lov_layout_state      *set   = &lov->u;
 725         const struct lov_layout_operations *ops;
 726         int result;
 727
 728         init_rwsem(&lov->lo_type_guard);
 729         atomic_set(&lov->lo_active_ios, 0);
 730         init_waitqueue_head(&lov->lo_waitq);
 731
 732         cl_object_page_init(lu2cl(obj), sizeof(struct lov_page));
 733
 734         /* no locking is necessary, as object is being created */
 735         lov->lo_type = lov_type(cconf->u.coc_md->lsm);
 736         ops = &lov_dispatch[lov->lo_type];
 737         result = ops->llo_init(env, dev, lov, cconf, set);
 738         if (result == 0)
 739                 ops->llo_install(env, lov, set);
 740         return result;
 741 }
 742
 743 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
 744                         const struct cl_object_conf *conf)
 745 {
 746         struct lov_stripe_md    *lsm = NULL;
 747         struct lov_object       *lov = cl2lov(obj);
 748         int                      result = 0;
 749
 750         lov_conf_lock(lov);
 751         if (conf->coc_opc == OBJECT_CONF_INVALIDATE) {
 752                 lov->lo_layout_invalid = true;
 753                 result = 0;
 754                 goto out;
 755         }
 756
 757         if (conf->coc_opc == OBJECT_CONF_WAIT) {
 758                 if (lov->lo_layout_invalid &&
 759                     atomic_read(&lov->lo_active_ios) > 0) {
 760                         lov_conf_unlock(lov);
 761                         result = lov_layout_wait(env, lov);
 762                         lov_conf_lock(lov);
 763                 }
 764                 goto out;
 765         }
 766
 767         LASSERT(conf->coc_opc == OBJECT_CONF_SET);
 768
 769         if (conf->u.coc_md != NULL)
 770                 lsm = conf->u.coc_md->lsm;
 771         if ((lsm == NULL && lov->lo_lsm == NULL) ||
 772             ((lsm != NULL && lov->lo_lsm != NULL) &&
 773              (lov->lo_lsm->lsm_layout_gen == lsm->lsm_layout_gen) &&
 774              (lov->lo_lsm->lsm_pattern == lsm->lsm_pattern))) {
 775                 /* same version of layout */
 776                 lov->lo_layout_invalid = false;
 777                 result = 0;
 778                 goto out;
 779         }
 780
 781         /* will change layout - check if there still exists active IO. */
 782         if (atomic_read(&lov->lo_active_ios) > 0) {
 783                 lov->lo_layout_invalid = true;
 784                 result = -EBUSY;
 785                 goto out;
 786         }
 787
 788         lov->lo_layout_invalid = lov_layout_change(env, lov, conf);
 789
 790 out:
 791         lov_conf_unlock(lov);
 792         CDEBUG(D_INODE, DFID" lo_layout_invalid=%d\n",
 793                PFID(lu_object_fid(lov2lu(lov))), lov->lo_layout_invalid);
 794         return result;
 795 }
 796
 797 static void lov_object_delete(const struct lu_env *env, struct lu_object *obj)
 798 {
 799         struct lov_object *lov = lu2lov(obj);
 800
 801         LOV_2DISPATCH_VOID(lov, llo_delete, env, lov, &lov->u);
 802 }
 803
 804 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
 805 {
 806         struct lov_object *lov = lu2lov(obj);
 807
 808         LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
 809         lu_object_fini(obj);
 810         OBD_SLAB_FREE_PTR(lov, lov_object_kmem);
 811 }
 812
 813 static int lov_object_print(const struct lu_env *env, void *cookie,
 814                             lu_printer_t p, const struct lu_object *o)
 815 {
 816         return LOV_2DISPATCH_NOLOCK(lu2lov(o), llo_print, env, cookie, p, o);
 817 }
 818
 819 int lov_page_init(const struct lu_env *env, struct cl_object *obj,
 820                 struct cl_page *page, struct page *vmpage)
 821 {
 822         return LOV_2DISPATCH_NOLOCK(cl2lov(obj),
 823                                     llo_page_init, env, obj, page, vmpage);
 824 }
 825
 826 /**
 827  * Implements cl_object_operations::clo_io_init() method for lov
 828  * layer. Dispatches to the appropriate layout io initialization method.
 829  */
 830 int lov_io_init(const struct lu_env *env, struct cl_object *obj,
 831                 struct cl_io *io)
 832 {
 833         CL_IO_SLICE_CLEAN(lov_env_io(env), lis_cl);
 834         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
 835                                      !io->ci_ignore_layout, env, obj, io);
 836 }
 837
 838 /**
 839  * An implementation of cl_object_operations::clo_attr_get() method for lov
 840  * layer. For raid0 layout this collects and merges attributes of all
 841  * sub-objects.
 842  */
 843 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
 844                         struct cl_attr *attr)
 845 {
 846         /* do not take lock, as this function is called under a
 847          * spin-lock. Layout is protected from changing by ongoing IO. */
 848         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
 849 }
 850
 851 static int lov_attr_set(const struct lu_env *env, struct cl_object *obj,
 852                         const struct cl_attr *attr, unsigned valid)
 853 {
 854         /*
 855          * No dispatch is required here, as no layout implements this.
 856          */
 857         return 0;
 858 }
 859
 860 int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
 861                   struct cl_lock *lock, const struct cl_io *io)
 862 {
 863         /* No need to lock because we've taken one refcount of layout.  */
 864         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_lock_init, env, obj, lock,
 865                                     io);
 866 }
 867
 868 static const struct cl_object_operations lov_ops = {
 869         .coo_page_init = lov_page_init,
 870         .coo_lock_init = lov_lock_init,
 871         .coo_io_init   = lov_io_init,
 872         .coo_attr_get  = lov_attr_get,
 873         .coo_attr_set  = lov_attr_set,
 874         .coo_conf_set  = lov_conf_set
 875 };
 876
 877 static const struct lu_object_operations lov_lu_obj_ops = {
 878         .loo_object_init      = lov_object_init,
 879         .loo_object_delete    = lov_object_delete,
 880         .loo_object_release   = NULL,
 881         .loo_object_free      = lov_object_free,
 882         .loo_object_print     = lov_object_print,
 883         .loo_object_invariant = NULL
 884 };
 885
 886 struct lu_object *lov_object_alloc(const struct lu_env *env,
 887                                    const struct lu_object_header *unused,
 888                                    struct lu_device *dev)
 889 {
 890         struct lov_object *lov;
 891         struct lu_object  *obj;
 892
 893         OBD_SLAB_ALLOC_PTR_GFP(lov, lov_object_kmem, GFP_NOFS);
 894         if (lov != NULL) {
 895                 obj = lov2lu(lov);
 896                 lu_object_init(obj, NULL, dev);
 897                 lov->lo_cl.co_ops = &lov_ops;
 898                 lov->lo_type = -1; /* invalid, to catch uninitialized type */
 899                 /*
 900                  * object io operation vector (cl_object::co_iop) is installed
 901                  * later in lov_object_init(), as different vectors are used
 902                  * for object with different layouts.
 903                  */
 904                 obj->lo_ops = &lov_lu_obj_ops;
 905         } else
 906                 obj = NULL;
 907         return obj;
 908 }
 909
 910 struct lov_stripe_md *lov_lsm_addref(struct lov_object *lov)
 911 {
 912         struct lov_stripe_md *lsm = NULL;
 913
 914         lov_conf_freeze(lov);
 915         if (lov->lo_lsm != NULL) {
 916                 lsm = lsm_addref(lov->lo_lsm);
 917                 CDEBUG(D_INODE, "lsm %p addref %d/%d by %p.\n",
 918                         lsm, atomic_read(&lsm->lsm_refc),
 919                         lov->lo_layout_invalid, current);
 920         }
 921         lov_conf_thaw(lov);
 922         return lsm;
 923 }
 924
 925 void lov_lsm_decref(struct lov_object *lov, struct lov_stripe_md *lsm)
 926 {
 927         if (lsm == NULL)
 928                 return;
 929
 930         CDEBUG(D_INODE, "lsm %p decref %d by %p.\n",
 931                 lsm, atomic_read(&lsm->lsm_refc), current);
 932
 933         lov_free_memmd(&lsm);
 934 }
 935
 936 struct lov_stripe_md *lov_lsm_get(struct cl_object *clobj)
 937 {
 938         struct lu_object *luobj;
 939         struct lov_stripe_md *lsm = NULL;
 940
 941         if (clobj == NULL)
 942                 return NULL;
 943
 944         luobj = lu_object_locate(&cl_object_header(clobj)->coh_lu,
 945                                  &lov_device_type);
 946         if (luobj != NULL)
 947                 lsm = lov_lsm_addref(lu2lov(luobj));
 948         return lsm;
 949 }
 950 EXPORT_SYMBOL(lov_lsm_get);
 951
 952 void lov_lsm_put(struct cl_object *unused, struct lov_stripe_md *lsm)
 953 {
 954         if (lsm != NULL)
 955                 lov_free_memmd(&lsm);
 956 }
 957 EXPORT_SYMBOL(lov_lsm_put);
 958
 959 int lov_read_and_clear_async_rc(struct cl_object *clob)
 960 {
 961         struct lu_object *luobj;
 962         int rc = 0;
 963
 964         luobj = lu_object_locate(&cl_object_header(clob)->coh_lu,
 965                                  &lov_device_type);
 966         if (luobj != NULL) {
 967                 struct lov_object *lov = lu2lov(luobj);
 968
 969                 lov_conf_freeze(lov);
 970                 switch (lov->lo_type) {
 971                 case LLT_RAID0: {
 972                         struct lov_stripe_md *lsm;
 973                         int i;
 974
 975                         lsm = lov->lo_lsm;
 976                         LASSERT(lsm != NULL);
 977                         for (i = 0; i < lsm->lsm_stripe_count; i++) {
 978                                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
 979
 980                                 if (lov_oinfo_is_dummy(loi))
 981                                         continue;
 982
 983                                 if (loi->loi_ar.ar_rc && !rc)
 984                                         rc = loi->loi_ar.ar_rc;
 985                                 loi->loi_ar.ar_rc = 0;
 986                         }
 987                 }
 988                 case LLT_RELEASED:
 989                 case LLT_EMPTY:
 990                         break;
 991                 default:
 992                         LBUG();
 993                 }
 994                 lov_conf_thaw(lov);
 995         }
 996         return rc;
 997 }
 998 EXPORT_SYMBOL(lov_read_and_clear_async_rc);
 999
1000 /** @} lov */