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, 2014, 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, pgoff_t index);
  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         int  (*llo_find_cbdata)(const struct lu_env *env, struct cl_object *obj,
  79                                 ldlm_iterator_t iter, void *data);
  80 };
  81
  82 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov);
  83
  84 struct lov_stripe_md *lov_lsm_get(struct cl_object *clobj)
  85 {
  86         struct lu_object *luobj;
  87         struct lov_stripe_md *lsm = NULL;
  88
  89         if (clobj == NULL)
  90                 return NULL;
  91
  92         luobj = lu_object_locate(&cl_object_header(clobj)->coh_lu,
  93                                  &lov_device_type);
  94         if (luobj != NULL)
  95                 lsm = lov_lsm_addref(lu2lov(luobj));
  96         return lsm;
  97 }
  98 EXPORT_SYMBOL(lov_lsm_get);
  99
 100 void lov_lsm_put(struct cl_object *unused, struct lov_stripe_md *lsm)
 101 {
 102         if (lsm != NULL)
 103                 lov_free_memmd(&lsm);
 104 }
 105 EXPORT_SYMBOL(lov_lsm_put);
 106
 107 /*****************************************************************************
 108  *
 109  * Lov object layout operations.
 110  *
 111  */
 112
 113 static void lov_install_empty(const struct lu_env *env,
 114                               struct lov_object *lov,
 115                               union  lov_layout_state *state)
 116 {
 117         /*
 118          * File without objects.
 119          */
 120 }
 121
 122 static int lov_init_empty(const struct lu_env *env,
 123                           struct lov_device *dev, struct lov_object *lov,
 124                           const struct cl_object_conf *conf,
 125                           union  lov_layout_state *state)
 126 {
 127         return 0;
 128 }
 129
 130 static void lov_install_raid0(const struct lu_env *env,
 131                               struct lov_object *lov,
 132                               union  lov_layout_state *state)
 133 {
 134 }
 135
 136 static struct cl_object *lov_sub_find(const struct lu_env *env,
 137                                       struct cl_device *dev,
 138                                       const struct lu_fid *fid,
 139                                       const struct cl_object_conf *conf)
 140 {
 141         struct lu_object *o;
 142
 143         ENTRY;
 144         o = lu_object_find_at(env, cl2lu_dev(dev), fid, &conf->coc_lu);
 145         LASSERT(ergo(!IS_ERR(o), o->lo_dev->ld_type == &lovsub_device_type));
 146         RETURN(lu2cl(o));
 147 }
 148
 149 static int lov_init_sub(const struct lu_env *env, struct lov_object *lov,
 150                         struct cl_object *stripe, struct lov_layout_raid0 *r0,
 151                         int idx)
 152 {
 153         struct cl_object_header *hdr;
 154         struct cl_object_header *subhdr;
 155         struct cl_object_header *parent;
 156         struct lov_oinfo        *oinfo;
 157         int result;
 158
 159         if (OBD_FAIL_CHECK(OBD_FAIL_LOV_INIT)) {
 160                 /* For sanity:test_206.
 161                  * Do not leave the object in cache to avoid accessing
 162                  * freed memory. This is because osc_object is referring to
 163                  * lov_oinfo of lsm_stripe_data which will be freed due to
 164                  * this failure. */
 165                 cl_object_kill(env, stripe);
 166                 cl_object_put(env, stripe);
 167                 return -EIO;
 168         }
 169
 170         hdr    = cl_object_header(lov2cl(lov));
 171         subhdr = cl_object_header(stripe);
 172
 173         oinfo = lov->lo_lsm->lsm_oinfo[idx];
 174         CDEBUG(D_INODE, DFID"@%p[%d] -> "DFID"@%p: ostid: "DOSTID
 175                " idx: %d gen: %d\n",
 176                PFID(&subhdr->coh_lu.loh_fid), subhdr, idx,
 177                PFID(&hdr->coh_lu.loh_fid), hdr, POSTID(&oinfo->loi_oi),
 178                oinfo->loi_ost_idx, oinfo->loi_ost_gen);
 179
 180         /* reuse ->coh_attr_guard to protect coh_parent change */
 181         spin_lock(&subhdr->coh_attr_guard);
 182         parent = subhdr->coh_parent;
 183         if (parent == NULL) {
 184                 subhdr->coh_parent = hdr;
 185                 spin_unlock(&subhdr->coh_attr_guard);
 186                 subhdr->coh_nesting = hdr->coh_nesting + 1;
 187                 lu_object_ref_add(&stripe->co_lu, "lov-parent", lov);
 188                 r0->lo_sub[idx] = cl2lovsub(stripe);
 189                 r0->lo_sub[idx]->lso_super = lov;
 190                 r0->lo_sub[idx]->lso_index = idx;
 191                 result = 0;
 192         } else {
 193                 struct lu_object  *old_obj;
 194                 struct lov_object *old_lov;
 195                 unsigned int mask = D_INODE;
 196
 197                 spin_unlock(&subhdr->coh_attr_guard);
 198                 old_obj = lu_object_locate(&parent->coh_lu, &lov_device_type);
 199                 LASSERT(old_obj != NULL);
 200                 old_lov = cl2lov(lu2cl(old_obj));
 201                 if (old_lov->lo_layout_invalid) {
 202                         /* the object's layout has already changed but isn't
 203                          * refreshed */
 204                         lu_object_unhash(env, &stripe->co_lu);
 205                         result = -EAGAIN;
 206                 } else {
 207                         mask = D_ERROR;
 208                         result = -EIO;
 209                 }
 210
 211                 LU_OBJECT_DEBUG(mask, env, &stripe->co_lu,
 212                                 "stripe %d is already owned.\n", idx);
 213                 LU_OBJECT_DEBUG(mask, env, old_obj, "owned.\n");
 214                 LU_OBJECT_HEADER(mask, env, lov2lu(lov), "try to own.\n");
 215                 cl_object_put(env, stripe);
 216         }
 217         return result;
 218 }
 219
 220 static int lov_page_slice_fixup(struct lov_object *lov,
 221                                 struct cl_object *stripe)
 222 {
 223         struct cl_object_header *hdr = cl_object_header(&lov->lo_cl);
 224         struct cl_object *o;
 225
 226         if (stripe == NULL)
 227                 return hdr->coh_page_bufsize - lov->lo_cl.co_slice_off -
 228                        cfs_size_round(sizeof(struct lov_page));
 229
 230         cl_object_for_each(o, stripe)
 231                 o->co_slice_off += hdr->coh_page_bufsize;
 232
 233         return cl_object_header(stripe)->coh_page_bufsize;
 234 }
 235
 236 static int lov_init_raid0(const struct lu_env *env,
 237                           struct lov_device *dev, struct lov_object *lov,
 238                           const struct cl_object_conf *conf,
 239                           union  lov_layout_state *state)
 240 {
 241         int result;
 242         int i;
 243
 244         struct cl_object        *stripe;
 245         struct lov_thread_info  *lti     = lov_env_info(env);
 246         struct cl_object_conf   *subconf = &lti->lti_stripe_conf;
 247         struct lov_stripe_md    *lsm     = conf->u.coc_md->lsm;
 248         struct lu_fid           *ofid    = &lti->lti_fid;
 249         struct lov_layout_raid0 *r0      = &state->raid0;
 250
 251         ENTRY;
 252
 253         if (lsm->lsm_magic != LOV_MAGIC_V1 && lsm->lsm_magic != LOV_MAGIC_V3) {
 254                 dump_lsm(D_ERROR, lsm);
 255                 LASSERTF(0, "magic mismatch, expected %d/%d, actual %d.\n",
 256                          LOV_MAGIC_V1, LOV_MAGIC_V3, lsm->lsm_magic);
 257         }
 258
 259         LASSERT(lov->lo_lsm == NULL);
 260         lov->lo_lsm = lsm_addref(lsm);
 261         r0->lo_nr  = lsm->lsm_stripe_count;
 262         LASSERT(r0->lo_nr <= lov_targets_nr(dev));
 263
 264         lov->lo_layout_invalid = true;
 265
 266         OBD_ALLOC_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 267         if (r0->lo_sub != NULL) {
 268                 int psz = 0;
 269
 270                 result = 0;
 271                 subconf->coc_inode = conf->coc_inode;
 272                 spin_lock_init(&r0->lo_sub_lock);
 273                 /*
 274                  * Create stripe cl_objects.
 275                  */
 276                 for (i = 0; i < r0->lo_nr && result == 0; ++i) {
 277                         struct cl_device *subdev;
 278                         struct lov_oinfo *oinfo = lsm->lsm_oinfo[i];
 279                         int ost_idx = oinfo->loi_ost_idx;
 280
 281                         if (lov_oinfo_is_dummy(oinfo))
 282                                 continue;
 283
 284                         result = ostid_to_fid(ofid, &oinfo->loi_oi,
 285                                               oinfo->loi_ost_idx);
 286                         if (result != 0)
 287                                 GOTO(out, result);
 288
 289                         subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
 290                         subconf->u.coc_oinfo = oinfo;
 291                         LASSERTF(subdev != NULL, "not init ost %d\n", ost_idx);
 292                         /* In the function below, .hs_keycmp resolves to
 293                          * lu_obj_hop_keycmp() */
 294                         /* coverity[overrun-buffer-val] */
 295                         stripe = lov_sub_find(env, subdev, ofid, subconf);
 296                         if (!IS_ERR(stripe)) {
 297                                 result = lov_init_sub(env, lov, stripe, r0, i);
 298                                 if (result == -EAGAIN) { /* try again */
 299                                         --i;
 300                                         result = 0;
 301                                         continue;
 302                                 }
 303                         } else {
 304                                 result = PTR_ERR(stripe);
 305                         }
 306
 307                         if (result == 0) {
 308                                 int sz = lov_page_slice_fixup(lov, stripe);
 309                                 LASSERT(ergo(psz > 0, psz == sz));
 310                                 psz = sz;
 311                         }
 312                 }
 313                 if (result == 0)
 314                         cl_object_header(&lov->lo_cl)->coh_page_bufsize += psz;
 315         } else
 316                 result = -ENOMEM;
 317 out:
 318         RETURN(result);
 319 }
 320
 321 static int lov_init_released(const struct lu_env *env,
 322                         struct lov_device *dev, struct lov_object *lov,
 323                         const struct cl_object_conf *conf,
 324                         union  lov_layout_state *state)
 325 {
 326         struct lov_stripe_md *lsm = conf->u.coc_md->lsm;
 327
 328         LASSERT(lsm != NULL);
 329         LASSERT(lsm_is_released(lsm));
 330         LASSERT(lov->lo_lsm == NULL);
 331
 332         lov->lo_lsm = lsm_addref(lsm);
 333         return 0;
 334 }
 335
 336 static int lov_delete_empty(const struct lu_env *env, struct lov_object *lov,
 337                             union lov_layout_state *state)
 338 {
 339         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 340
 341         lov_layout_wait(env, lov);
 342         return 0;
 343 }
 344
 345 static void lov_subobject_kill(const struct lu_env *env, struct lov_object *lov,
 346                                struct lovsub_object *los, int idx)
 347 {
 348         struct cl_object        *sub;
 349         struct lov_layout_raid0 *r0;
 350         struct lu_site          *site;
 351         struct lu_site_bkt_data *bkt;
 352         wait_queue_t          *waiter;
 353
 354         r0  = &lov->u.raid0;
 355         LASSERT(r0->lo_sub[idx] == los);
 356
 357         sub  = lovsub2cl(los);
 358         site = sub->co_lu.lo_dev->ld_site;
 359         bkt  = lu_site_bkt_from_fid(site, &sub->co_lu.lo_header->loh_fid);
 360
 361         cl_object_kill(env, sub);
 362         /* release a reference to the sub-object and ... */
 363         lu_object_ref_del(&sub->co_lu, "lov-parent", lov);
 364         cl_object_put(env, sub);
 365
 366         /* ... wait until it is actually destroyed---sub-object clears its
 367          * ->lo_sub[] slot in lovsub_object_fini() */
 368         if (r0->lo_sub[idx] == los) {
 369                 waiter = &lov_env_info(env)->lti_waiter;
 370                 init_waitqueue_entry_current(waiter);
 371                 add_wait_queue(&bkt->lsb_marche_funebre, waiter);
 372                 set_current_state(TASK_UNINTERRUPTIBLE);
 373                 while (1) {
 374                         /* this wait-queue is signaled at the end of
 375                          * lu_object_free(). */
 376                         set_current_state(TASK_UNINTERRUPTIBLE);
 377                         spin_lock(&r0->lo_sub_lock);
 378                         if (r0->lo_sub[idx] == los) {
 379                                 spin_unlock(&r0->lo_sub_lock);
 380                                 waitq_wait(waiter, TASK_UNINTERRUPTIBLE);
 381                         } else {
 382                                 spin_unlock(&r0->lo_sub_lock);
 383                                 set_current_state(TASK_RUNNING);
 384                                 break;
 385                         }
 386                 }
 387                 remove_wait_queue(&bkt->lsb_marche_funebre, waiter);
 388         }
 389         LASSERT(r0->lo_sub[idx] == NULL);
 390 }
 391
 392 static int lov_delete_raid0(const struct lu_env *env, struct lov_object *lov,
 393                             union lov_layout_state *state)
 394 {
 395         struct lov_layout_raid0 *r0 = &state->raid0;
 396         struct lov_stripe_md    *lsm = lov->lo_lsm;
 397         int i;
 398
 399         ENTRY;
 400
 401         dump_lsm(D_INODE, lsm);
 402
 403         lov_layout_wait(env, lov);
 404         if (r0->lo_sub != NULL) {
 405                 for (i = 0; i < r0->lo_nr; ++i) {
 406                         struct lovsub_object *los = r0->lo_sub[i];
 407
 408                         if (los != NULL) {
 409                                 cl_object_prune(env, &los->lso_cl);
 410                                 /*
 411                                  * If top-level object is to be evicted from
 412                                  * the cache, so are its sub-objects.
 413                                  */
 414                                 lov_subobject_kill(env, lov, los, i);
 415                         }
 416                 }
 417         }
 418         RETURN(0);
 419 }
 420
 421 static void lov_fini_empty(const struct lu_env *env, struct lov_object *lov,
 422                            union lov_layout_state *state)
 423 {
 424         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 425 }
 426
 427 static void lov_fini_raid0(const struct lu_env *env, struct lov_object *lov,
 428                            union lov_layout_state *state)
 429 {
 430         struct lov_layout_raid0 *r0 = &state->raid0;
 431         ENTRY;
 432
 433         if (r0->lo_sub != NULL) {
 434                 OBD_FREE_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 435                 r0->lo_sub = NULL;
 436         }
 437
 438         dump_lsm(D_INODE, lov->lo_lsm);
 439         lov_free_memmd(&lov->lo_lsm);
 440
 441         EXIT;
 442 }
 443
 444 static void lov_fini_released(const struct lu_env *env, struct lov_object *lov,
 445                                 union lov_layout_state *state)
 446 {
 447         ENTRY;
 448         dump_lsm(D_INODE, lov->lo_lsm);
 449         lov_free_memmd(&lov->lo_lsm);
 450         EXIT;
 451 }
 452
 453 static int lov_print_empty(const struct lu_env *env, void *cookie,
 454                            lu_printer_t p, const struct lu_object *o)
 455 {
 456         (*p)(env, cookie, "empty %d\n", lu2lov(o)->lo_layout_invalid);
 457         return 0;
 458 }
 459
 460 static int lov_print_raid0(const struct lu_env *env, void *cookie,
 461                            lu_printer_t p, const struct lu_object *o)
 462 {
 463         struct lov_object       *lov = lu2lov(o);
 464         struct lov_layout_raid0 *r0  = lov_r0(lov);
 465         struct lov_stripe_md    *lsm = lov->lo_lsm;
 466         int                      i;
 467
 468         (*p)(env, cookie, "stripes: %d, %s, lsm{%p 0x%08X %d %u %u}:\n",
 469                 r0->lo_nr, lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 470                 lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 471                 lsm->lsm_stripe_count, lsm->lsm_layout_gen);
 472         for (i = 0; i < r0->lo_nr; ++i) {
 473                 struct lu_object *sub;
 474
 475                 if (r0->lo_sub[i] != NULL) {
 476                         sub = lovsub2lu(r0->lo_sub[i]);
 477                         lu_object_print(env, cookie, p, sub);
 478                 } else {
 479                         (*p)(env, cookie, "sub %d absent\n", i);
 480                 }
 481         }
 482         return 0;
 483 }
 484
 485 static int lov_print_released(const struct lu_env *env, void *cookie,
 486                                 lu_printer_t p, const struct lu_object *o)
 487 {
 488         struct lov_object       *lov = lu2lov(o);
 489         struct lov_stripe_md    *lsm = lov->lo_lsm;
 490
 491         (*p)(env, cookie,
 492                 "released: %s, lsm{%p 0x%08X %d %u %u}:\n",
 493                 lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 494                 lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 495                 lsm->lsm_stripe_count, lsm->lsm_layout_gen);
 496         return 0;
 497 }
 498
 499 /**
 500  * Implements cl_object_operations::coo_attr_get() method for an object
 501  * without stripes (LLT_EMPTY layout type).
 502  *
 503  * The only attributes this layer is authoritative in this case is
 504  * cl_attr::cat_blocks---it's 0.
 505  */
 506 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
 507                               struct cl_attr *attr)
 508 {
 509         attr->cat_blocks = 0;
 510         return 0;
 511 }
 512
 513 static int lov_attr_get_raid0(const struct lu_env *env, struct cl_object *obj,
 514                               struct cl_attr *attr)
 515 {
 516         struct lov_object       *lov = cl2lov(obj);
 517         struct lov_layout_raid0 *r0 = lov_r0(lov);
 518         struct cl_attr          *lov_attr = &r0->lo_attr;
 519         int                      result = 0;
 520
 521         ENTRY;
 522
 523         /* this is called w/o holding type guard mutex, so it must be inside
 524          * an on going IO otherwise lsm may be replaced.
 525          * LU-2117: it turns out there exists one exception. For mmaped files,
 526          * the lock of those files may be requested in the other file's IO
 527          * context, and this function is called in ccc_lock_state(), it will
 528          * hit this assertion.
 529          * Anyway, it's still okay to call attr_get w/o type guard as layout
 530          * can't go if locks exist. */
 531         /* LASSERT(atomic_read(&lsm->lsm_refc) > 1); */
 532
 533         if (!r0->lo_attr_valid) {
 534                 struct lov_stripe_md    *lsm = lov->lo_lsm;
 535                 struct ost_lvb          *lvb = &lov_env_info(env)->lti_lvb;
 536                 __u64                    kms = 0;
 537
 538                 memset(lvb, 0, sizeof(*lvb));
 539                 /* XXX: timestamps can be negative by sanity:test_39m,
 540                  * how can it be? */
 541                 lvb->lvb_atime = LLONG_MIN;
 542                 lvb->lvb_ctime = LLONG_MIN;
 543                 lvb->lvb_mtime = LLONG_MIN;
 544
 545                 /*
 546                  * XXX that should be replaced with a loop over sub-objects,
 547                  * doing cl_object_attr_get() on them. But for now, let's
 548                  * reuse old lov code.
 549                  */
 550
 551                 /*
 552                  * XXX take lsm spin-lock to keep lov_merge_lvb_kms()
 553                  * happy. It's not needed, because new code uses
 554                  * ->coh_attr_guard spin-lock to protect consistency of
 555                  * sub-object attributes.
 556                  */
 557                 lov_stripe_lock(lsm);
 558                 result = lov_merge_lvb_kms(lsm, lvb, &kms);
 559                 lov_stripe_unlock(lsm);
 560                 if (result == 0) {
 561                         cl_lvb2attr(lov_attr, lvb);
 562                         lov_attr->cat_kms = kms;
 563                         r0->lo_attr_valid = 1;
 564                 }
 565         }
 566         if (result == 0) { /* merge results */
 567                 attr->cat_blocks = lov_attr->cat_blocks;
 568                 attr->cat_size = lov_attr->cat_size;
 569                 attr->cat_kms = lov_attr->cat_kms;
 570                 if (attr->cat_atime < lov_attr->cat_atime)
 571                         attr->cat_atime = lov_attr->cat_atime;
 572                 if (attr->cat_ctime < lov_attr->cat_ctime)
 573                         attr->cat_ctime = lov_attr->cat_ctime;
 574                 if (attr->cat_mtime < lov_attr->cat_mtime)
 575                         attr->cat_mtime = lov_attr->cat_mtime;
 576         }
 577         RETURN(result);
 578 }
 579
 580 static int lov_find_cbdata_empty(const struct lu_env *env,
 581                                  struct cl_object *obj, ldlm_iterator_t iter,
 582                                  void *data)
 583 {
 584         return 0;
 585 }
 586
 587 static int lov_find_cbdata_raid0(const struct lu_env *env,
 588                                  struct cl_object *obj, ldlm_iterator_t iter,
 589                                  void *data)
 590 {
 591         struct lov_object       *lov = cl2lov(obj);
 592         struct lov_layout_raid0 *r0 = lov_r0(lov);
 593         struct cl_object        *subobj;
 594         int                     i;
 595         int                     rc = 0;
 596
 597         for (i = 0; i < r0->lo_nr; ++i) {
 598                 if (r0->lo_sub[i] == NULL)
 599                         continue;
 600
 601                 subobj = lovsub2cl(r0->lo_sub[i]);
 602
 603                 rc = cl_object_find_cbdata(env, subobj, iter, data);
 604                 if (rc != 0)
 605                         break;
 606         }
 607
 608         return rc;
 609 }
 610
 611 const static struct lov_layout_operations lov_dispatch[] = {
 612         [LLT_EMPTY] = {
 613                 .llo_init      = lov_init_empty,
 614                 .llo_delete    = lov_delete_empty,
 615                 .llo_fini      = lov_fini_empty,
 616                 .llo_install   = lov_install_empty,
 617                 .llo_print     = lov_print_empty,
 618                 .llo_page_init = lov_page_init_empty,
 619                 .llo_lock_init = lov_lock_init_empty,
 620                 .llo_io_init   = lov_io_init_empty,
 621                 .llo_getattr   = lov_attr_get_empty,
 622                 .llo_find_cbdata = lov_find_cbdata_empty
 623         },
 624         [LLT_RAID0] = {
 625                 .llo_init      = lov_init_raid0,
 626                 .llo_delete    = lov_delete_raid0,
 627                 .llo_fini      = lov_fini_raid0,
 628                 .llo_install   = lov_install_raid0,
 629                 .llo_print     = lov_print_raid0,
 630                 .llo_page_init = lov_page_init_raid0,
 631                 .llo_lock_init = lov_lock_init_raid0,
 632                 .llo_io_init   = lov_io_init_raid0,
 633                 .llo_getattr   = lov_attr_get_raid0,
 634                 .llo_find_cbdata = lov_find_cbdata_raid0
 635         },
 636         [LLT_RELEASED] = {
 637                 .llo_init      = lov_init_released,
 638                 .llo_delete    = lov_delete_empty,
 639                 .llo_fini      = lov_fini_released,
 640                 .llo_install   = lov_install_empty,
 641                 .llo_print     = lov_print_released,
 642                 .llo_page_init = lov_page_init_empty,
 643                 .llo_lock_init = lov_lock_init_empty,
 644                 .llo_io_init   = lov_io_init_released,
 645                 .llo_getattr   = lov_attr_get_empty,
 646                 .llo_find_cbdata = lov_find_cbdata_empty
 647         }
 648 };
 649
 650 /**
 651  * Performs a double-dispatch based on the layout type of an object.
 652  */
 653 #define LOV_2DISPATCH_NOLOCK(obj, op, ...)                              \
 654 ({                                                                      \
 655         struct lov_object                      *__obj = (obj);          \
 656         enum lov_layout_type                    __llt;                  \
 657                                                                         \
 658         __llt = __obj->lo_type;                                         \
 659         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 660         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 661 })
 662
 663 /**
 664  * Return lov_layout_type associated with a given lsm
 665  */
 666 static enum lov_layout_type lov_type(struct lov_stripe_md *lsm)
 667 {
 668         if (lsm == NULL)
 669                 return LLT_EMPTY;
 670         if (lsm_is_released(lsm))
 671                 return LLT_RELEASED;
 672         return LLT_RAID0;
 673 }
 674
 675 static inline void lov_conf_freeze(struct lov_object *lov)
 676 {
 677         if (lov->lo_owner != current)
 678                 down_read(&lov->lo_type_guard);
 679 }
 680
 681 static inline void lov_conf_thaw(struct lov_object *lov)
 682 {
 683         if (lov->lo_owner != current)
 684                 up_read(&lov->lo_type_guard);
 685 }
 686
 687 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...)                       \
 688 ({                                                                      \
 689         struct lov_object                      *__obj = (obj);          \
 690         int                                     __lock = !!(lock);      \
 691         typeof(lov_dispatch[0].op(__VA_ARGS__)) __result;               \
 692                                                                         \
 693         if (__lock)                                                     \
 694                 lov_conf_freeze(__obj);                                 \
 695         __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__);          \
 696         if (__lock)                                                     \
 697                 lov_conf_thaw(__obj);                                   \
 698         __result;                                                       \
 699 })
 700
 701 /**
 702  * Performs a locked double-dispatch based on the layout type of an object.
 703  */
 704 #define LOV_2DISPATCH(obj, op, ...)                     \
 705         LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
 706
 707 #define LOV_2DISPATCH_VOID(obj, op, ...)                                \
 708 do {                                                                    \
 709         struct lov_object                      *__obj = (obj);          \
 710         enum lov_layout_type                    __llt;                  \
 711                                                                         \
 712         lov_conf_freeze(__obj);                                         \
 713         __llt = __obj->lo_type;                                         \
 714         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 715         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 716         lov_conf_thaw(__obj);                                           \
 717 } while (0)
 718
 719 static void lov_conf_lock(struct lov_object *lov)
 720 {
 721         LASSERT(lov->lo_owner != current);
 722         down_write(&lov->lo_type_guard);
 723         LASSERT(lov->lo_owner == NULL);
 724         lov->lo_owner = current;
 725 }
 726
 727 static void lov_conf_unlock(struct lov_object *lov)
 728 {
 729         lov->lo_owner = NULL;
 730         up_write(&lov->lo_type_guard);
 731 }
 732
 733 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov)
 734 {
 735         struct l_wait_info lwi = { 0 };
 736         ENTRY;
 737
 738         while (atomic_read(&lov->lo_active_ios) > 0) {
 739                 CDEBUG(D_INODE, "file:"DFID" wait for active IO, now: %d.\n",
 740                         PFID(lu_object_fid(lov2lu(lov))),
 741                         atomic_read(&lov->lo_active_ios));
 742
 743                 l_wait_event(lov->lo_waitq,
 744                              atomic_read(&lov->lo_active_ios) == 0, &lwi);
 745         }
 746         RETURN(0);
 747 }
 748
 749 static int lov_layout_change(const struct lu_env *unused,
 750                              struct lov_object *lov,
 751                              const struct cl_object_conf *conf)
 752 {
 753         int result;
 754         enum lov_layout_type llt = LLT_EMPTY;
 755         union lov_layout_state *state = &lov->u;
 756         const struct lov_layout_operations *old_ops;
 757         const struct lov_layout_operations *new_ops;
 758
 759         void *cookie;
 760         struct lu_env *env;
 761         int refcheck;
 762         ENTRY;
 763
 764         LASSERT(0 <= lov->lo_type && lov->lo_type < ARRAY_SIZE(lov_dispatch));
 765
 766         if (conf->u.coc_md != NULL)
 767                 llt = lov_type(conf->u.coc_md->lsm);
 768         LASSERT(0 <= llt && llt < ARRAY_SIZE(lov_dispatch));
 769
 770         cookie = cl_env_reenter();
 771         env = cl_env_get(&refcheck);
 772         if (IS_ERR(env)) {
 773                 cl_env_reexit(cookie);
 774                 RETURN(PTR_ERR(env));
 775         }
 776
 777         CDEBUG(D_INODE, DFID" from %s to %s\n",
 778                PFID(lu_object_fid(lov2lu(lov))),
 779                llt2str(lov->lo_type), llt2str(llt));
 780
 781         old_ops = &lov_dispatch[lov->lo_type];
 782         new_ops = &lov_dispatch[llt];
 783
 784         result = cl_object_prune(env, &lov->lo_cl);
 785         if (result != 0)
 786                 GOTO(out, result);
 787
 788         result = old_ops->llo_delete(env, lov, &lov->u);
 789         if (result == 0) {
 790                 old_ops->llo_fini(env, lov, &lov->u);
 791
 792                 LASSERT(atomic_read(&lov->lo_active_ios) == 0);
 793
 794                 lov->lo_type = LLT_EMPTY;
 795                 /* page bufsize fixup */
 796                 cl_object_header(&lov->lo_cl)->coh_page_bufsize -=
 797                         lov_page_slice_fixup(lov, NULL);
 798
 799                 result = new_ops->llo_init(env,
 800                                         lu2lov_dev(lov->lo_cl.co_lu.lo_dev),
 801                                         lov, conf, state);
 802                 if (result == 0) {
 803                         new_ops->llo_install(env, lov, state);
 804                         lov->lo_type = llt;
 805                 } else {
 806                         new_ops->llo_delete(env, lov, state);
 807                         new_ops->llo_fini(env, lov, state);
 808                         /* this file becomes an EMPTY file. */
 809                 }
 810         }
 811
 812 out:
 813         cl_env_put(env, &refcheck);
 814         cl_env_reexit(cookie);
 815         RETURN(result);
 816 }
 817
 818 /*****************************************************************************
 819  *
 820  * Lov object operations.
 821  *
 822  */
 823 int lov_object_init(const struct lu_env *env, struct lu_object *obj,
 824                     const struct lu_object_conf *conf)
 825 {
 826         struct lov_device            *dev   = lu2lov_dev(obj->lo_dev);
 827         struct lov_object            *lov   = lu2lov(obj);
 828         const struct cl_object_conf  *cconf = lu2cl_conf(conf);
 829         union  lov_layout_state      *set   = &lov->u;
 830         const struct lov_layout_operations *ops;
 831         int result;
 832
 833         ENTRY;
 834         init_rwsem(&lov->lo_type_guard);
 835         atomic_set(&lov->lo_active_ios, 0);
 836         init_waitqueue_head(&lov->lo_waitq);
 837
 838         cl_object_page_init(lu2cl(obj), sizeof(struct lov_page));
 839
 840         /* no locking is necessary, as object is being created */
 841         lov->lo_type = lov_type(cconf->u.coc_md->lsm);
 842         ops = &lov_dispatch[lov->lo_type];
 843         result = ops->llo_init(env, dev, lov, cconf, set);
 844         if (result == 0)
 845                 ops->llo_install(env, lov, set);
 846         RETURN(result);
 847 }
 848
 849 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
 850                         const struct cl_object_conf *conf)
 851 {
 852         struct lov_stripe_md    *lsm = NULL;
 853         struct lov_object       *lov = cl2lov(obj);
 854         int                      result = 0;
 855         ENTRY;
 856
 857         lov_conf_lock(lov);
 858         if (conf->coc_opc == OBJECT_CONF_INVALIDATE) {
 859                 lov->lo_layout_invalid = true;
 860                 GOTO(out, result = 0);
 861         }
 862
 863         if (conf->coc_opc == OBJECT_CONF_WAIT) {
 864                 if (lov->lo_layout_invalid &&
 865                     atomic_read(&lov->lo_active_ios) > 0) {
 866                         lov_conf_unlock(lov);
 867                         result = lov_layout_wait(env, lov);
 868                         lov_conf_lock(lov);
 869                 }
 870                 GOTO(out, result);
 871         }
 872
 873         LASSERT(conf->coc_opc == OBJECT_CONF_SET);
 874
 875         if (conf->u.coc_md != NULL)
 876                 lsm = conf->u.coc_md->lsm;
 877         if ((lsm == NULL && lov->lo_lsm == NULL) ||
 878             ((lsm != NULL && lov->lo_lsm != NULL) &&
 879              (lov->lo_lsm->lsm_layout_gen == lsm->lsm_layout_gen) &&
 880              (lov->lo_lsm->lsm_pattern == lsm->lsm_pattern))) {
 881                 /* same version of layout */
 882                 lov->lo_layout_invalid = false;
 883                 GOTO(out, result = 0);
 884         }
 885
 886         /* will change layout - check if there still exists active IO. */
 887         if (atomic_read(&lov->lo_active_ios) > 0) {
 888                 lov->lo_layout_invalid = true;
 889                 GOTO(out, result = -EBUSY);
 890         }
 891
 892         result = lov_layout_change(env, lov, conf);
 893         lov->lo_layout_invalid = result != 0;
 894         EXIT;
 895
 896 out:
 897         lov_conf_unlock(lov);
 898         CDEBUG(D_INODE, DFID" lo_layout_invalid=%d\n",
 899                PFID(lu_object_fid(lov2lu(lov))), lov->lo_layout_invalid);
 900         RETURN(result);
 901 }
 902
 903 static void lov_object_delete(const struct lu_env *env, struct lu_object *obj)
 904 {
 905         struct lov_object *lov = lu2lov(obj);
 906
 907         ENTRY;
 908         LOV_2DISPATCH_VOID(lov, llo_delete, env, lov, &lov->u);
 909         EXIT;
 910 }
 911
 912 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
 913 {
 914         struct lov_object *lov = lu2lov(obj);
 915
 916         ENTRY;
 917         LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
 918         lu_object_fini(obj);
 919         OBD_SLAB_FREE_PTR(lov, lov_object_kmem);
 920         EXIT;
 921 }
 922
 923 static int lov_object_print(const struct lu_env *env, void *cookie,
 924                             lu_printer_t p, const struct lu_object *o)
 925 {
 926         return LOV_2DISPATCH_NOLOCK(lu2lov(o), llo_print, env, cookie, p, o);
 927 }
 928
 929 int lov_page_init(const struct lu_env *env, struct cl_object *obj,
 930                   struct cl_page *page, pgoff_t index)
 931 {
 932         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_page_init, env, obj, page,
 933                                     index);
 934 }
 935
 936 /**
 937  * Implements cl_object_operations::clo_io_init() method for lov
 938  * layer. Dispatches to the appropriate layout io initialization method.
 939  */
 940 int lov_io_init(const struct lu_env *env, struct cl_object *obj,
 941                 struct cl_io *io)
 942 {
 943         CL_IO_SLICE_CLEAN(lov_env_io(env), lis_cl);
 944         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
 945                                      !io->ci_ignore_layout, env, obj, io);
 946 }
 947
 948 /**
 949  * An implementation of cl_object_operations::clo_attr_get() method for lov
 950  * layer. For raid0 layout this collects and merges attributes of all
 951  * sub-objects.
 952  */
 953 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
 954                         struct cl_attr *attr)
 955 {
 956         /* do not take lock, as this function is called under a
 957          * spin-lock. Layout is protected from changing by ongoing IO. */
 958         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
 959 }
 960
 961 static int lov_attr_update(const struct lu_env *env, struct cl_object *obj,
 962                            const struct cl_attr *attr, unsigned valid)
 963 {
 964         /*
 965          * No dispatch is required here, as no layout implements this.
 966          */
 967         return 0;
 968 }
 969
 970 int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
 971                   struct cl_lock *lock, const struct cl_io *io)
 972 {
 973         /* No need to lock because we've taken one refcount of layout.  */
 974         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_lock_init, env, obj, lock,
 975                                     io);
 976 }
 977
 978 static int lov_object_getstripe(const struct lu_env *env, struct cl_object *obj,
 979                                 struct lov_user_md __user *lum)
 980 {
 981         struct lov_object       *lov = cl2lov(obj);
 982         struct lov_stripe_md    *lsm;
 983         int                     rc = 0;
 984         ENTRY;
 985
 986         lsm = lov_lsm_addref(lov);
 987         if (lsm == NULL)
 988                 RETURN(-ENODATA);
 989
 990         rc = lov_getstripe(cl2lov(obj), lsm, lum);
 991         lov_lsm_put(obj, lsm);
 992         RETURN(rc);
 993 }
 994
 995 static int lov_object_find_cbdata(const struct lu_env *env,
 996                                   struct cl_object *obj, ldlm_iterator_t iter,
 997                                   void *data)
 998 {
 999         int rc;
1000         ENTRY;
1001
1002         /* call cl_object_find_cbdata for sub obj */
1003         rc = LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_find_cbdata,
1004                                   env, obj, iter, data);
1005         RETURN(rc);
1006 }
1007
1008 static const struct cl_object_operations lov_ops = {
1009         .coo_page_init    = lov_page_init,
1010         .coo_lock_init    = lov_lock_init,
1011         .coo_io_init      = lov_io_init,
1012         .coo_attr_get     = lov_attr_get,
1013         .coo_attr_update  = lov_attr_update,
1014         .coo_conf_set     = lov_conf_set,
1015         .coo_getstripe    = lov_object_getstripe,
1016         .coo_find_cbdata  = lov_object_find_cbdata
1017 };
1018
1019 static const struct lu_object_operations lov_lu_obj_ops = {
1020         .loo_object_init      = lov_object_init,
1021         .loo_object_delete    = lov_object_delete,
1022         .loo_object_release   = NULL,
1023         .loo_object_free      = lov_object_free,
1024         .loo_object_print     = lov_object_print,
1025         .loo_object_invariant = NULL
1026 };
1027
1028 struct lu_object *lov_object_alloc(const struct lu_env *env,
1029                                    const struct lu_object_header *unused,
1030                                    struct lu_device *dev)
1031 {
1032         struct lov_object *lov;
1033         struct lu_object  *obj;
1034
1035         ENTRY;
1036         OBD_SLAB_ALLOC_PTR_GFP(lov, lov_object_kmem, GFP_NOFS);
1037         if (lov != NULL) {
1038                 obj = lov2lu(lov);
1039                 lu_object_init(obj, NULL, dev);
1040                 lov->lo_cl.co_ops = &lov_ops;
1041                 lov->lo_type = -1; /* invalid, to catch uninitialized type */
1042                 /*
1043                  * object io operation vector (cl_object::co_iop) is installed
1044                  * later in lov_object_init(), as different vectors are used
1045                  * for object with different layouts.
1046                  */
1047                 obj->lo_ops = &lov_lu_obj_ops;
1048         } else
1049                 obj = NULL;
1050         RETURN(obj);
1051 }
1052
1053 struct lov_stripe_md *lov_lsm_addref(struct lov_object *lov)
1054 {
1055         struct lov_stripe_md *lsm = NULL;
1056
1057         lov_conf_freeze(lov);
1058         if (lov->lo_lsm != NULL) {
1059                 lsm = lsm_addref(lov->lo_lsm);
1060                 CDEBUG(D_INODE, "lsm %p addref %d/%d by %p.\n",
1061                         lsm, atomic_read(&lsm->lsm_refc),
1062                         lov->lo_layout_invalid, current);
1063         }
1064         lov_conf_thaw(lov);
1065         return lsm;
1066 }
1067
1068 int lov_read_and_clear_async_rc(struct cl_object *clob)
1069 {
1070         struct lu_object *luobj;
1071         int rc = 0;
1072         ENTRY;
1073
1074         luobj = lu_object_locate(&cl_object_header(clob)->coh_lu,
1075                                  &lov_device_type);
1076         if (luobj != NULL) {
1077                 struct lov_object *lov = lu2lov(luobj);
1078
1079                 lov_conf_freeze(lov);
1080                 switch (lov->lo_type) {
1081                 case LLT_RAID0: {
1082                         struct lov_stripe_md *lsm;
1083                         int i;
1084
1085                         lsm = lov->lo_lsm;
1086                         LASSERT(lsm != NULL);
1087                         for (i = 0; i < lsm->lsm_stripe_count; i++) {
1088                                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
1089
1090                                 if (lov_oinfo_is_dummy(loi))
1091                                         continue;
1092
1093                                 if (loi->loi_ar.ar_rc && !rc)
1094                                         rc = loi->loi_ar.ar_rc;
1095                                 loi->loi_ar.ar_rc = 0;
1096                         }
1097                 }
1098                 case LLT_RELEASED:
1099                 case LLT_EMPTY:
1100                         break;
1101                 default:
1102                         LBUG();
1103                 }
1104                 lov_conf_thaw(lov);
1105         }
1106         RETURN(rc);
1107 }
1108 EXPORT_SYMBOL(lov_read_and_clear_async_rc);
1109
1110 /** @} lov */