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, 2013, 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, pgoff_t index);
  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, struct lov_layout_raid0 *r0,
 127                         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
 149         oinfo = lov->lo_lsm->lsm_oinfo[idx];
 150         CDEBUG(D_INODE, DFID"@%p[%d] -> "DFID"@%p: ostid: "DOSTID
 151                " idx: %d gen: %d\n",
 152                PFID(&subhdr->coh_lu.loh_fid), subhdr, idx,
 153                PFID(&hdr->coh_lu.loh_fid), hdr, POSTID(&oinfo->loi_oi),
 154                oinfo->loi_ost_idx, oinfo->loi_ost_gen);
 155
 156         /* reuse ->coh_attr_guard to protect coh_parent change */
 157         spin_lock(&subhdr->coh_attr_guard);
 158         parent = subhdr->coh_parent;
 159         if (parent == NULL) {
 160                 subhdr->coh_parent = hdr;
 161                 spin_unlock(&subhdr->coh_attr_guard);
 162                 subhdr->coh_nesting = hdr->coh_nesting + 1;
 163                 lu_object_ref_add(&stripe->co_lu, "lov-parent", lov);
 164                 r0->lo_sub[idx] = cl2lovsub(stripe);
 165                 r0->lo_sub[idx]->lso_super = lov;
 166                 r0->lo_sub[idx]->lso_index = idx;
 167                 result = 0;
 168         } else {
 169                 struct lu_object  *old_obj;
 170                 struct lov_object *old_lov;
 171                 unsigned int mask = D_INODE;
 172
 173                 spin_unlock(&subhdr->coh_attr_guard);
 174                 old_obj = lu_object_locate(&parent->coh_lu, &lov_device_type);
 175                 LASSERT(old_obj != NULL);
 176                 old_lov = cl2lov(lu2cl(old_obj));
 177                 if (old_lov->lo_layout_invalid) {
 178                         /* the object's layout has already changed but isn't
 179                          * refreshed */
 180                         lu_object_unhash(env, &stripe->co_lu);
 181                         result = -EAGAIN;
 182                 } else {
 183                         mask = D_ERROR;
 184                         result = -EIO;
 185                 }
 186
 187                 LU_OBJECT_DEBUG(mask, env, &stripe->co_lu,
 188                                 "stripe %d is already owned.\n", idx);
 189                 LU_OBJECT_DEBUG(mask, env, old_obj, "owned.\n");
 190                 LU_OBJECT_HEADER(mask, env, lov2lu(lov), "try to own.\n");
 191                 cl_object_put(env, stripe);
 192         }
 193         return result;
 194 }
 195
 196 static int lov_page_slice_fixup(struct lov_object *lov,
 197                                 struct cl_object *stripe)
 198 {
 199         struct cl_object_header *hdr = cl_object_header(&lov->lo_cl);
 200         struct cl_object *o;
 201
 202         cl_object_for_each(o, stripe)
 203                 o->co_slice_off += hdr->coh_page_bufsize;
 204
 205         return cl_object_header(stripe)->coh_page_bufsize;
 206 }
 207
 208 static int lov_init_raid0(const struct lu_env *env,
 209                           struct lov_device *dev, struct lov_object *lov,
 210                           const struct cl_object_conf *conf,
 211                           union  lov_layout_state *state)
 212 {
 213         int result;
 214         int i;
 215
 216         struct cl_object        *stripe;
 217         struct lov_thread_info  *lti     = lov_env_info(env);
 218         struct cl_object_conf   *subconf = &lti->lti_stripe_conf;
 219         struct lov_stripe_md    *lsm     = conf->u.coc_md->lsm;
 220         struct lu_fid           *ofid    = &lti->lti_fid;
 221         struct lov_layout_raid0 *r0      = &state->raid0;
 222
 223         ENTRY;
 224
 225         if (lsm->lsm_magic != LOV_MAGIC_V1 && lsm->lsm_magic != LOV_MAGIC_V3) {
 226                 dump_lsm(D_ERROR, lsm);
 227                 LASSERTF(0, "magic mismatch, expected %d/%d, actual %d.\n",
 228                          LOV_MAGIC_V1, LOV_MAGIC_V3, lsm->lsm_magic);
 229         }
 230
 231         LASSERT(lov->lo_lsm == NULL);
 232         lov->lo_lsm = lsm_addref(lsm);
 233         r0->lo_nr  = lsm->lsm_stripe_count;
 234         LASSERT(r0->lo_nr <= lov_targets_nr(dev));
 235
 236         OBD_ALLOC_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 237         if (r0->lo_sub != NULL) {
 238                 int psz = 0;
 239
 240                 result = 0;
 241                 subconf->coc_inode = conf->coc_inode;
 242                 spin_lock_init(&r0->lo_sub_lock);
 243                 /*
 244                  * Create stripe cl_objects.
 245                  */
 246                 for (i = 0; i < r0->lo_nr && result == 0; ++i) {
 247                         struct cl_device *subdev;
 248                         struct lov_oinfo *oinfo = lsm->lsm_oinfo[i];
 249                         int ost_idx = oinfo->loi_ost_idx;
 250
 251                         result = ostid_to_fid(ofid, &oinfo->loi_oi,
 252                                               oinfo->loi_ost_idx);
 253                         if (result != 0)
 254                                 GOTO(out, result);
 255
 256                         subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
 257                         subconf->u.coc_oinfo = oinfo;
 258                         LASSERTF(subdev != NULL, "not init ost %d\n", ost_idx);
 259                         /* In the function below, .hs_keycmp resolves to
 260                          * lu_obj_hop_keycmp() */
 261                         /* coverity[overrun-buffer-val] */
 262                         stripe = lov_sub_find(env, subdev, ofid, subconf);
 263                         if (!IS_ERR(stripe)) {
 264                                 result = lov_init_sub(env, lov, stripe, r0, i);
 265                                 if (result == -EAGAIN) { /* try again */
 266                                         --i;
 267                                         result = 0;
 268                                         continue;
 269                                 }
 270                         } else {
 271                                 result = PTR_ERR(stripe);
 272                         }
 273
 274                         if (result == 0) {
 275                                 int sz = lov_page_slice_fixup(lov, stripe);
 276                                 LASSERT(ergo(psz > 0, psz == sz));
 277                                 psz = sz;
 278                         }
 279                 }
 280                 if (result == 0)
 281                         cl_object_header(&lov->lo_cl)->coh_page_bufsize += psz;
 282         } else
 283                 result = -ENOMEM;
 284 out:
 285         RETURN(result);
 286 }
 287
 288 static int lov_init_released(const struct lu_env *env,
 289                         struct lov_device *dev, struct lov_object *lov,
 290                         const struct cl_object_conf *conf,
 291                         union  lov_layout_state *state)
 292 {
 293         struct lov_stripe_md *lsm = conf->u.coc_md->lsm;
 294
 295         LASSERT(lsm != NULL);
 296         LASSERT(lsm_is_released(lsm));
 297         LASSERT(lov->lo_lsm == NULL);
 298
 299         lov->lo_lsm = lsm_addref(lsm);
 300         return 0;
 301 }
 302
 303 static int lov_delete_empty(const struct lu_env *env, struct lov_object *lov,
 304                             union lov_layout_state *state)
 305 {
 306         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 307
 308         lov_layout_wait(env, lov);
 309
 310         cl_locks_prune(env, &lov->lo_cl, 0);
 311         return 0;
 312 }
 313
 314 static void lov_subobject_kill(const struct lu_env *env, struct lov_object *lov,
 315                                struct lovsub_object *los, int idx)
 316 {
 317         struct cl_object        *sub;
 318         struct lov_layout_raid0 *r0;
 319         struct lu_site          *site;
 320         struct lu_site_bkt_data *bkt;
 321         wait_queue_t          *waiter;
 322
 323         r0  = &lov->u.raid0;
 324         LASSERT(r0->lo_sub[idx] == los);
 325
 326         sub  = lovsub2cl(los);
 327         site = sub->co_lu.lo_dev->ld_site;
 328         bkt  = lu_site_bkt_from_fid(site, &sub->co_lu.lo_header->loh_fid);
 329
 330         cl_object_kill(env, sub);
 331         /* release a reference to the sub-object and ... */
 332         lu_object_ref_del(&sub->co_lu, "lov-parent", lov);
 333         cl_object_put(env, sub);
 334
 335         /* ... wait until it is actually destroyed---sub-object clears its
 336          * ->lo_sub[] slot in lovsub_object_fini() */
 337         if (r0->lo_sub[idx] == los) {
 338                 waiter = &lov_env_info(env)->lti_waiter;
 339                 init_waitqueue_entry_current(waiter);
 340                 add_wait_queue(&bkt->lsb_marche_funebre, waiter);
 341                 set_current_state(TASK_UNINTERRUPTIBLE);
 342                 while (1) {
 343                         /* this wait-queue is signaled at the end of
 344                          * lu_object_free(). */
 345                         set_current_state(TASK_UNINTERRUPTIBLE);
 346                         spin_lock(&r0->lo_sub_lock);
 347                         if (r0->lo_sub[idx] == los) {
 348                                 spin_unlock(&r0->lo_sub_lock);
 349                                 waitq_wait(waiter, TASK_UNINTERRUPTIBLE);
 350                         } else {
 351                                 spin_unlock(&r0->lo_sub_lock);
 352                                 set_current_state(TASK_RUNNING);
 353                                 break;
 354                         }
 355                 }
 356                 remove_wait_queue(&bkt->lsb_marche_funebre, waiter);
 357         }
 358         LASSERT(r0->lo_sub[idx] == NULL);
 359 }
 360
 361 static int lov_delete_raid0(const struct lu_env *env, struct lov_object *lov,
 362                             union lov_layout_state *state)
 363 {
 364         struct lov_layout_raid0 *r0 = &state->raid0;
 365         struct lov_stripe_md    *lsm = lov->lo_lsm;
 366         int i;
 367
 368         ENTRY;
 369
 370         dump_lsm(D_INODE, lsm);
 371
 372         lov_layout_wait(env, lov);
 373         if (r0->lo_sub != NULL) {
 374                 for (i = 0; i < r0->lo_nr; ++i) {
 375                         struct lovsub_object *los = r0->lo_sub[i];
 376
 377                         if (los != NULL) {
 378                                 cl_locks_prune(env, &los->lso_cl, 1);
 379                                 /*
 380                                  * If top-level object is to be evicted from
 381                                  * the cache, so are its sub-objects.
 382                                  */
 383                                 lov_subobject_kill(env, lov, los, i);
 384                         }
 385                 }
 386         }
 387         cl_locks_prune(env, &lov->lo_cl, 0);
 388         RETURN(0);
 389 }
 390
 391 static void lov_fini_empty(const struct lu_env *env, struct lov_object *lov,
 392                            union lov_layout_state *state)
 393 {
 394         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 395 }
 396
 397 static void lov_fini_raid0(const struct lu_env *env, struct lov_object *lov,
 398                            union lov_layout_state *state)
 399 {
 400         struct lov_layout_raid0 *r0 = &state->raid0;
 401         ENTRY;
 402
 403         if (r0->lo_sub != NULL) {
 404                 OBD_FREE_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 405                 r0->lo_sub = NULL;
 406         }
 407
 408         dump_lsm(D_INODE, lov->lo_lsm);
 409         lov_free_memmd(&lov->lo_lsm);
 410
 411         EXIT;
 412 }
 413
 414 static void lov_fini_released(const struct lu_env *env, struct lov_object *lov,
 415                                 union lov_layout_state *state)
 416 {
 417         ENTRY;
 418         dump_lsm(D_INODE, lov->lo_lsm);
 419         lov_free_memmd(&lov->lo_lsm);
 420         EXIT;
 421 }
 422
 423 static int lov_print_empty(const struct lu_env *env, void *cookie,
 424                            lu_printer_t p, const struct lu_object *o)
 425 {
 426         (*p)(env, cookie, "empty %d\n", lu2lov(o)->lo_layout_invalid);
 427         return 0;
 428 }
 429
 430 static int lov_print_raid0(const struct lu_env *env, void *cookie,
 431                            lu_printer_t p, const struct lu_object *o)
 432 {
 433         struct lov_object       *lov = lu2lov(o);
 434         struct lov_layout_raid0 *r0  = lov_r0(lov);
 435         struct lov_stripe_md    *lsm = lov->lo_lsm;
 436         int                      i;
 437
 438         (*p)(env, cookie, "stripes: %d, %s, lsm{%p 0x%08X %d %u %u}:\n",
 439                 r0->lo_nr, lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 440                 lsm->lsm_magic, cfs_atomic_read(&lsm->lsm_refc),
 441                 lsm->lsm_stripe_count, lsm->lsm_layout_gen);
 442         for (i = 0; i < r0->lo_nr; ++i) {
 443                 struct lu_object *sub;
 444
 445                 if (r0->lo_sub[i] != NULL) {
 446                         sub = lovsub2lu(r0->lo_sub[i]);
 447                         lu_object_print(env, cookie, p, sub);
 448                 } else {
 449                         (*p)(env, cookie, "sub %d absent\n", i);
 450                 }
 451         }
 452         return 0;
 453 }
 454
 455 static int lov_print_released(const struct lu_env *env, void *cookie,
 456                                 lu_printer_t p, const struct lu_object *o)
 457 {
 458         struct lov_object       *lov = lu2lov(o);
 459         struct lov_stripe_md    *lsm = lov->lo_lsm;
 460
 461         (*p)(env, cookie,
 462                 "released: %s, lsm{%p 0x%08X %d %u %u}:\n",
 463                 lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 464                 lsm->lsm_magic, cfs_atomic_read(&lsm->lsm_refc),
 465                 lsm->lsm_stripe_count, lsm->lsm_layout_gen);
 466         return 0;
 467 }
 468
 469 /**
 470  * Implements cl_object_operations::coo_attr_get() method for an object
 471  * without stripes (LLT_EMPTY layout type).
 472  *
 473  * The only attributes this layer is authoritative in this case is
 474  * cl_attr::cat_blocks---it's 0.
 475  */
 476 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
 477                               struct cl_attr *attr)
 478 {
 479         attr->cat_blocks = 0;
 480         return 0;
 481 }
 482
 483 static int lov_attr_get_raid0(const struct lu_env *env, struct cl_object *obj,
 484                               struct cl_attr *attr)
 485 {
 486         struct lov_object       *lov = cl2lov(obj);
 487         struct lov_layout_raid0 *r0 = lov_r0(lov);
 488         struct cl_attr          *lov_attr = &r0->lo_attr;
 489         int                      result = 0;
 490
 491         ENTRY;
 492
 493         /* this is called w/o holding type guard mutex, so it must be inside
 494          * an on going IO otherwise lsm may be replaced.
 495          * LU-2117: it turns out there exists one exception. For mmaped files,
 496          * the lock of those files may be requested in the other file's IO
 497          * context, and this function is called in ccc_lock_state(), it will
 498          * hit this assertion.
 499          * Anyway, it's still okay to call attr_get w/o type guard as layout
 500          * can't go if locks exist. */
 501         /* LASSERT(cfs_atomic_read(&lsm->lsm_refc) > 1); */
 502
 503         if (!r0->lo_attr_valid) {
 504                 struct lov_stripe_md    *lsm = lov->lo_lsm;
 505                 struct ost_lvb          *lvb = &lov_env_info(env)->lti_lvb;
 506                 __u64                    kms = 0;
 507
 508                 memset(lvb, 0, sizeof(*lvb));
 509                 /* XXX: timestamps can be negative by sanity:test_39m,
 510                  * how can it be? */
 511                 lvb->lvb_atime = LLONG_MIN;
 512                 lvb->lvb_ctime = LLONG_MIN;
 513                 lvb->lvb_mtime = LLONG_MIN;
 514
 515                 /*
 516                  * XXX that should be replaced with a loop over sub-objects,
 517                  * doing cl_object_attr_get() on them. But for now, let's
 518                  * reuse old lov code.
 519                  */
 520
 521                 /*
 522                  * XXX take lsm spin-lock to keep lov_merge_lvb_kms()
 523                  * happy. It's not needed, because new code uses
 524                  * ->coh_attr_guard spin-lock to protect consistency of
 525                  * sub-object attributes.
 526                  */
 527                 lov_stripe_lock(lsm);
 528                 result = lov_merge_lvb_kms(lsm, lvb, &kms);
 529                 lov_stripe_unlock(lsm);
 530                 if (result == 0) {
 531                         cl_lvb2attr(lov_attr, lvb);
 532                         lov_attr->cat_kms = kms;
 533                         r0->lo_attr_valid = 1;
 534                 }
 535         }
 536         if (result == 0) { /* merge results */
 537                 attr->cat_blocks = lov_attr->cat_blocks;
 538                 attr->cat_size = lov_attr->cat_size;
 539                 attr->cat_kms = lov_attr->cat_kms;
 540                 if (attr->cat_atime < lov_attr->cat_atime)
 541                         attr->cat_atime = lov_attr->cat_atime;
 542                 if (attr->cat_ctime < lov_attr->cat_ctime)
 543                         attr->cat_ctime = lov_attr->cat_ctime;
 544                 if (attr->cat_mtime < lov_attr->cat_mtime)
 545                         attr->cat_mtime = lov_attr->cat_mtime;
 546         }
 547         RETURN(result);
 548 }
 549
 550 const static struct lov_layout_operations lov_dispatch[] = {
 551         [LLT_EMPTY] = {
 552                 .llo_init      = lov_init_empty,
 553                 .llo_delete    = lov_delete_empty,
 554                 .llo_fini      = lov_fini_empty,
 555                 .llo_install   = lov_install_empty,
 556                 .llo_print     = lov_print_empty,
 557                 .llo_page_init = lov_page_init_empty,
 558                 .llo_lock_init = lov_lock_init_empty,
 559                 .llo_io_init   = lov_io_init_empty,
 560                 .llo_getattr   = lov_attr_get_empty
 561         },
 562         [LLT_RAID0] = {
 563                 .llo_init      = lov_init_raid0,
 564                 .llo_delete    = lov_delete_raid0,
 565                 .llo_fini      = lov_fini_raid0,
 566                 .llo_install   = lov_install_raid0,
 567                 .llo_print     = lov_print_raid0,
 568                 .llo_page_init = lov_page_init_raid0,
 569                 .llo_lock_init = lov_lock_init_raid0,
 570                 .llo_io_init   = lov_io_init_raid0,
 571                 .llo_getattr   = lov_attr_get_raid0
 572         },
 573         [LLT_RELEASED] = {
 574                 .llo_init      = lov_init_released,
 575                 .llo_delete    = lov_delete_empty,
 576                 .llo_fini      = lov_fini_released,
 577                 .llo_install   = lov_install_empty,
 578                 .llo_print     = lov_print_released,
 579                 .llo_page_init = lov_page_init_empty,
 580                 .llo_lock_init = lov_lock_init_empty,
 581                 .llo_io_init   = lov_io_init_released,
 582                 .llo_getattr   = lov_attr_get_empty
 583         }
 584 };
 585
 586 /**
 587  * Performs a double-dispatch based on the layout type of an object.
 588  */
 589 #define LOV_2DISPATCH_NOLOCK(obj, op, ...)                              \
 590 ({                                                                      \
 591         struct lov_object                      *__obj = (obj);          \
 592         enum lov_layout_type                    __llt;                  \
 593                                                                         \
 594         __llt = __obj->lo_type;                                         \
 595         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 596         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 597 })
 598
 599 /**
 600  * Return lov_layout_type associated with a given lsm
 601  */
 602 enum lov_layout_type lov_type(struct lov_stripe_md *lsm)
 603 {
 604         if (lsm == NULL)
 605                 return LLT_EMPTY;
 606         if (lsm_is_released(lsm))
 607                 return LLT_RELEASED;
 608         return LLT_RAID0;
 609 }
 610
 611 static inline void lov_conf_freeze(struct lov_object *lov)
 612 {
 613         if (lov->lo_owner != current)
 614                 down_read(&lov->lo_type_guard);
 615 }
 616
 617 static inline void lov_conf_thaw(struct lov_object *lov)
 618 {
 619         if (lov->lo_owner != current)
 620                 up_read(&lov->lo_type_guard);
 621 }
 622
 623 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...)                       \
 624 ({                                                                      \
 625         struct lov_object                      *__obj = (obj);          \
 626         int                                     __lock = !!(lock);      \
 627         typeof(lov_dispatch[0].op(__VA_ARGS__)) __result;               \
 628                                                                         \
 629         if (__lock)                                                     \
 630                 lov_conf_freeze(__obj);                                 \
 631         __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__);          \
 632         if (__lock)                                                     \
 633                 lov_conf_thaw(__obj);                                   \
 634         __result;                                                       \
 635 })
 636
 637 /**
 638  * Performs a locked double-dispatch based on the layout type of an object.
 639  */
 640 #define LOV_2DISPATCH(obj, op, ...)                     \
 641         LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
 642
 643 #define LOV_2DISPATCH_VOID(obj, op, ...)                                \
 644 do {                                                                    \
 645         struct lov_object                      *__obj = (obj);          \
 646         enum lov_layout_type                    __llt;                  \
 647                                                                         \
 648         lov_conf_freeze(__obj);                                         \
 649         __llt = __obj->lo_type;                                         \
 650         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 651         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 652         lov_conf_thaw(__obj);                                           \
 653 } while (0)
 654
 655 static void lov_conf_lock(struct lov_object *lov)
 656 {
 657         LASSERT(lov->lo_owner != current);
 658         down_write(&lov->lo_type_guard);
 659         LASSERT(lov->lo_owner == NULL);
 660         lov->lo_owner = current;
 661 }
 662
 663 static void lov_conf_unlock(struct lov_object *lov)
 664 {
 665         lov->lo_owner = NULL;
 666         up_write(&lov->lo_type_guard);
 667 }
 668
 669 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov)
 670 {
 671         struct l_wait_info lwi = { 0 };
 672         ENTRY;
 673
 674         while (cfs_atomic_read(&lov->lo_active_ios) > 0) {
 675                 CDEBUG(D_INODE, "file:"DFID" wait for active IO, now: %d.\n",
 676                         PFID(lu_object_fid(lov2lu(lov))),
 677                         cfs_atomic_read(&lov->lo_active_ios));
 678
 679                 l_wait_event(lov->lo_waitq,
 680                              cfs_atomic_read(&lov->lo_active_ios) == 0, &lwi);
 681         }
 682         RETURN(0);
 683 }
 684
 685 static int lov_layout_change(const struct lu_env *unused,
 686                              struct lov_object *lov,
 687                              const struct cl_object_conf *conf)
 688 {
 689         int result;
 690         enum lov_layout_type llt = LLT_EMPTY;
 691         union lov_layout_state *state = &lov->u;
 692         const struct lov_layout_operations *old_ops;
 693         const struct lov_layout_operations *new_ops;
 694
 695         void *cookie;
 696         struct lu_env *env;
 697         int refcheck;
 698         ENTRY;
 699
 700         LASSERT(0 <= lov->lo_type && lov->lo_type < ARRAY_SIZE(lov_dispatch));
 701
 702         if (conf->u.coc_md != NULL)
 703                 llt = lov_type(conf->u.coc_md->lsm);
 704         LASSERT(0 <= llt && llt < ARRAY_SIZE(lov_dispatch));
 705
 706         cookie = cl_env_reenter();
 707         env = cl_env_get(&refcheck);
 708         if (IS_ERR(env)) {
 709                 cl_env_reexit(cookie);
 710                 RETURN(PTR_ERR(env));
 711         }
 712
 713         CDEBUG(D_INODE, DFID" from %s to %s\n",
 714                PFID(lu_object_fid(lov2lu(lov))),
 715                llt2str(lov->lo_type), llt2str(llt));
 716
 717         old_ops = &lov_dispatch[lov->lo_type];
 718         new_ops = &lov_dispatch[llt];
 719
 720         cl_object_prune(env, &lov->lo_cl);
 721
 722         result = old_ops->llo_delete(env, lov, &lov->u);
 723         if (result == 0) {
 724                 old_ops->llo_fini(env, lov, &lov->u);
 725
 726                 LASSERT(cfs_atomic_read(&lov->lo_active_ios) == 0);
 727
 728                 lov->lo_type = LLT_EMPTY;
 729                 result = new_ops->llo_init(env,
 730                                         lu2lov_dev(lov->lo_cl.co_lu.lo_dev),
 731                                         lov, conf, state);
 732                 if (result == 0) {
 733                         new_ops->llo_install(env, lov, state);
 734                         lov->lo_type = llt;
 735                 } else {
 736                         new_ops->llo_delete(env, lov, state);
 737                         new_ops->llo_fini(env, lov, state);
 738                         /* this file becomes an EMPTY file. */
 739                 }
 740         }
 741
 742         cl_env_put(env, &refcheck);
 743         cl_env_reexit(cookie);
 744         RETURN(result);
 745 }
 746
 747 /*****************************************************************************
 748  *
 749  * Lov object operations.
 750  *
 751  */
 752 int lov_object_init(const struct lu_env *env, struct lu_object *obj,
 753                     const struct lu_object_conf *conf)
 754 {
 755         struct lov_device            *dev   = lu2lov_dev(obj->lo_dev);
 756         struct lov_object            *lov   = lu2lov(obj);
 757         const struct cl_object_conf  *cconf = lu2cl_conf(conf);
 758         union  lov_layout_state      *set   = &lov->u;
 759         const struct lov_layout_operations *ops;
 760         int result;
 761
 762         ENTRY;
 763         init_rwsem(&lov->lo_type_guard);
 764         cfs_atomic_set(&lov->lo_active_ios, 0);
 765         init_waitqueue_head(&lov->lo_waitq);
 766
 767         cl_object_page_init(lu2cl(obj), sizeof(struct lov_page));
 768
 769         /* no locking is necessary, as object is being created */
 770         lov->lo_type = lov_type(cconf->u.coc_md->lsm);
 771         ops = &lov_dispatch[lov->lo_type];
 772         result = ops->llo_init(env, dev, lov, cconf, set);
 773         if (result == 0)
 774                 ops->llo_install(env, lov, set);
 775         RETURN(result);
 776 }
 777
 778 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
 779                         const struct cl_object_conf *conf)
 780 {
 781         struct lov_stripe_md    *lsm = NULL;
 782         struct lov_object       *lov = cl2lov(obj);
 783         int                      result = 0;
 784         ENTRY;
 785
 786         lov_conf_lock(lov);
 787         if (conf->coc_opc == OBJECT_CONF_INVALIDATE) {
 788                 lov->lo_layout_invalid = true;
 789                 GOTO(out, result = 0);
 790         }
 791
 792         if (conf->coc_opc == OBJECT_CONF_WAIT) {
 793                 if (lov->lo_layout_invalid &&
 794                     cfs_atomic_read(&lov->lo_active_ios) > 0) {
 795                         lov_conf_unlock(lov);
 796                         result = lov_layout_wait(env, lov);
 797                         lov_conf_lock(lov);
 798                 }
 799                 GOTO(out, result);
 800         }
 801
 802         LASSERT(conf->coc_opc == OBJECT_CONF_SET);
 803
 804         if (conf->u.coc_md != NULL)
 805                 lsm = conf->u.coc_md->lsm;
 806         if ((lsm == NULL && lov->lo_lsm == NULL) ||
 807             ((lsm != NULL && lov->lo_lsm != NULL) &&
 808              (lov->lo_lsm->lsm_layout_gen == lsm->lsm_layout_gen) &&
 809              (lov->lo_lsm->lsm_pattern == lsm->lsm_pattern))) {
 810                 /* same version of layout */
 811                 lov->lo_layout_invalid = false;
 812                 GOTO(out, result = 0);
 813         }
 814
 815         /* will change layout - check if there still exists active IO. */
 816         if (cfs_atomic_read(&lov->lo_active_ios) > 0) {
 817                 lov->lo_layout_invalid = true;
 818                 GOTO(out, result = -EBUSY);
 819         }
 820
 821         lov->lo_layout_invalid = lov_layout_change(env, lov, conf);
 822         EXIT;
 823
 824 out:
 825         lov_conf_unlock(lov);
 826         CDEBUG(D_INODE, DFID" lo_layout_invalid=%d\n",
 827                PFID(lu_object_fid(lov2lu(lov))), lov->lo_layout_invalid);
 828         RETURN(result);
 829 }
 830
 831 static void lov_object_delete(const struct lu_env *env, struct lu_object *obj)
 832 {
 833         struct lov_object *lov = lu2lov(obj);
 834
 835         ENTRY;
 836         LOV_2DISPATCH_VOID(lov, llo_delete, env, lov, &lov->u);
 837         EXIT;
 838 }
 839
 840 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
 841 {
 842         struct lov_object *lov = lu2lov(obj);
 843
 844         ENTRY;
 845         LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
 846         lu_object_fini(obj);
 847         OBD_SLAB_FREE_PTR(lov, lov_object_kmem);
 848         EXIT;
 849 }
 850
 851 static int lov_object_print(const struct lu_env *env, void *cookie,
 852                             lu_printer_t p, const struct lu_object *o)
 853 {
 854         return LOV_2DISPATCH_NOLOCK(lu2lov(o), llo_print, env, cookie, p, o);
 855 }
 856
 857 int lov_page_init(const struct lu_env *env, struct cl_object *obj,
 858                   struct cl_page *page, pgoff_t index)
 859 {
 860         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_page_init, env, obj, page,
 861                                     index);
 862 }
 863
 864 /**
 865  * Implements cl_object_operations::clo_io_init() method for lov
 866  * layer. Dispatches to the appropriate layout io initialization method.
 867  */
 868 int lov_io_init(const struct lu_env *env, struct cl_object *obj,
 869                 struct cl_io *io)
 870 {
 871         CL_IO_SLICE_CLEAN(lov_env_io(env), lis_cl);
 872         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
 873                                      !io->ci_ignore_layout, env, obj, io);
 874 }
 875
 876 /**
 877  * An implementation of cl_object_operations::clo_attr_get() method for lov
 878  * layer. For raid0 layout this collects and merges attributes of all
 879  * sub-objects.
 880  */
 881 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
 882                         struct cl_attr *attr)
 883 {
 884         /* do not take lock, as this function is called under a
 885          * spin-lock. Layout is protected from changing by ongoing IO. */
 886         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
 887 }
 888
 889 static int lov_attr_set(const struct lu_env *env, struct cl_object *obj,
 890                         const struct cl_attr *attr, unsigned valid)
 891 {
 892         /*
 893          * No dispatch is required here, as no layout implements this.
 894          */
 895         return 0;
 896 }
 897
 898 int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
 899                   struct cl_lock *lock, const struct cl_io *io)
 900 {
 901         /* No need to lock because we've taken one refcount of layout.  */
 902         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_lock_init, env, obj, lock,
 903                                     io);
 904 }
 905
 906 static const struct cl_object_operations lov_ops = {
 907         .coo_page_init = lov_page_init,
 908         .coo_lock_init = lov_lock_init,
 909         .coo_io_init   = lov_io_init,
 910         .coo_attr_get  = lov_attr_get,
 911         .coo_attr_set  = lov_attr_set,
 912         .coo_conf_set  = lov_conf_set
 913 };
 914
 915 static const struct lu_object_operations lov_lu_obj_ops = {
 916         .loo_object_init      = lov_object_init,
 917         .loo_object_delete    = lov_object_delete,
 918         .loo_object_release   = NULL,
 919         .loo_object_free      = lov_object_free,
 920         .loo_object_print     = lov_object_print,
 921         .loo_object_invariant = NULL
 922 };
 923
 924 struct lu_object *lov_object_alloc(const struct lu_env *env,
 925                                    const struct lu_object_header *unused,
 926                                    struct lu_device *dev)
 927 {
 928         struct lov_object *lov;
 929         struct lu_object  *obj;
 930
 931         ENTRY;
 932         OBD_SLAB_ALLOC_PTR_GFP(lov, lov_object_kmem, GFP_NOFS);
 933         if (lov != NULL) {
 934                 obj = lov2lu(lov);
 935                 lu_object_init(obj, NULL, dev);
 936                 lov->lo_cl.co_ops = &lov_ops;
 937                 lov->lo_type = -1; /* invalid, to catch uninitialized type */
 938                 /*
 939                  * object io operation vector (cl_object::co_iop) is installed
 940                  * later in lov_object_init(), as different vectors are used
 941                  * for object with different layouts.
 942                  */
 943                 obj->lo_ops = &lov_lu_obj_ops;
 944         } else
 945                 obj = NULL;
 946         RETURN(obj);
 947 }
 948
 949 struct lov_stripe_md *lov_lsm_addref(struct lov_object *lov)
 950 {
 951         struct lov_stripe_md *lsm = NULL;
 952
 953         lov_conf_freeze(lov);
 954         if (lov->lo_lsm != NULL) {
 955                 lsm = lsm_addref(lov->lo_lsm);
 956                 CDEBUG(D_INODE, "lsm %p addref %d/%d by %p.\n",
 957                         lsm, cfs_atomic_read(&lsm->lsm_refc),
 958                         lov->lo_layout_invalid, current);
 959         }
 960         lov_conf_thaw(lov);
 961         return lsm;
 962 }
 963
 964 struct lov_stripe_md *lov_lsm_get(struct cl_object *clobj)
 965 {
 966         struct lu_object *luobj;
 967         struct lov_stripe_md *lsm = NULL;
 968
 969         if (clobj == NULL)
 970                 return NULL;
 971
 972         luobj = lu_object_locate(&cl_object_header(clobj)->coh_lu,
 973                                  &lov_device_type);
 974         if (luobj != NULL)
 975                 lsm = lov_lsm_addref(lu2lov(luobj));
 976         return lsm;
 977 }
 978 EXPORT_SYMBOL(lov_lsm_get);
 979
 980 void lov_lsm_put(struct cl_object *unused, struct lov_stripe_md *lsm)
 981 {
 982         if (lsm != NULL)
 983                 lov_free_memmd(&lsm);
 984 }
 985 EXPORT_SYMBOL(lov_lsm_put);
 986
 987 int lov_read_and_clear_async_rc(struct cl_object *clob)
 988 {
 989         struct lu_object *luobj;
 990         int rc = 0;
 991         ENTRY;
 992
 993         luobj = lu_object_locate(&cl_object_header(clob)->coh_lu,
 994                                  &lov_device_type);
 995         if (luobj != NULL) {
 996                 struct lov_object *lov = lu2lov(luobj);
 997
 998                 lov_conf_freeze(lov);
 999                 switch (lov->lo_type) {
1000                 case LLT_RAID0: {
1001                         struct lov_stripe_md *lsm;
1002                         int i;
1003
1004                         lsm = lov->lo_lsm;
1005                         LASSERT(lsm != NULL);
1006                         for (i = 0; i < lsm->lsm_stripe_count; i++) {
1007                                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
1008                                 if (loi->loi_ar.ar_rc && !rc)
1009                                         rc = loi->loi_ar.ar_rc;
1010                                 loi->loi_ar.ar_rc = 0;
1011                         }
1012                 }
1013                 case LLT_RELEASED:
1014                 case LLT_EMPTY:
1015                         break;
1016                 default:
1017                         LBUG();
1018                 }
1019                 lov_conf_thaw(lov);
1020         }
1021         RETURN(rc);
1022 }
1023 EXPORT_SYMBOL(lov_read_and_clear_async_rc);
1024
1025 /** @} lov */