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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2015, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * Implementation of cl_object for LOV layer.
  33  *
  34  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  35  *   Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_LOV
  39
  40 #include "lov_cl_internal.h"
  41
  42 /** \addtogroup lov
  43  *  @{
  44  */
  45
  46 /*****************************************************************************
  47  *
  48  * Layout operations.
  49  *
  50  */
  51
  52 struct lov_layout_operations {
  53         int (*llo_init)(const struct lu_env *env, struct lov_device *dev,
  54                         struct lov_object *lov,
  55                         const struct cl_object_conf *conf,
  56                         union lov_layout_state *state);
  57         int (*llo_delete)(const struct lu_env *env, struct lov_object *lov,
  58                           union lov_layout_state *state);
  59         void (*llo_fini)(const struct lu_env *env, struct lov_object *lov,
  60                          union lov_layout_state *state);
  61         void (*llo_install)(const struct lu_env *env, struct lov_object *lov,
  62                             union lov_layout_state *state);
  63         int  (*llo_print)(const struct lu_env *env, void *cookie,
  64                           lu_printer_t p, const struct lu_object *o);
  65         int  (*llo_page_init)(const struct lu_env *env, struct cl_object *obj,
  66                               struct cl_page *page, pgoff_t index);
  67         int  (*llo_lock_init)(const struct lu_env *env,
  68                               struct cl_object *obj, struct cl_lock *lock,
  69                               const struct cl_io *io);
  70         int  (*llo_io_init)(const struct lu_env *env,
  71                             struct cl_object *obj, struct cl_io *io);
  72         int  (*llo_getattr)(const struct lu_env *env, struct cl_object *obj,
  73                             struct cl_attr *attr);
  74 };
  75
  76 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov);
  77
  78 /*****************************************************************************
  79  *
  80  * Lov object layout operations.
  81  *
  82  */
  83
  84 static void lov_install_empty(const struct lu_env *env,
  85                               struct lov_object *lov,
  86                               union  lov_layout_state *state)
  87 {
  88         /*
  89          * File without objects.
  90          */
  91 }
  92
  93 static int lov_init_empty(const struct lu_env *env,
  94                           struct lov_device *dev, struct lov_object *lov,
  95                           const struct cl_object_conf *conf,
  96                           union  lov_layout_state *state)
  97 {
  98         return 0;
  99 }
 100
 101 static void lov_install_raid0(const struct lu_env *env,
 102                               struct lov_object *lov,
 103                               union  lov_layout_state *state)
 104 {
 105 }
 106
 107 static struct cl_object *lov_sub_find(const struct lu_env *env,
 108                                       struct cl_device *dev,
 109                                       const struct lu_fid *fid,
 110                                       const struct cl_object_conf *conf)
 111 {
 112         struct lu_object *o;
 113
 114         o = lu_object_find_at(env, cl2lu_dev(dev), fid, &conf->coc_lu);
 115         LASSERT(ergo(!IS_ERR(o), o->lo_dev->ld_type == &lovsub_device_type));
 116         return lu2cl(o);
 117 }
 118
 119 static int lov_init_sub(const struct lu_env *env, struct lov_object *lov,
 120                         struct cl_object *stripe, struct lov_layout_raid0 *r0,
 121                         int idx)
 122 {
 123         struct cl_object_header *hdr;
 124         struct cl_object_header *subhdr;
 125         struct cl_object_header *parent;
 126         struct lov_oinfo        *oinfo;
 127         int result;
 128
 129         if (OBD_FAIL_CHECK(OBD_FAIL_LOV_INIT)) {
 130                 /* For sanity:test_206.
 131                  * Do not leave the object in cache to avoid accessing
 132                  * freed memory. This is because osc_object is referring to
 133                  * lov_oinfo of lsm_stripe_data which will be freed due to
 134                  * this failure.
 135                  */
 136                 cl_object_kill(env, stripe);
 137                 cl_object_put(env, stripe);
 138                 return -EIO;
 139         }
 140
 141         hdr    = cl_object_header(lov2cl(lov));
 142         subhdr = cl_object_header(stripe);
 143
 144         oinfo = lov->lo_lsm->lsm_oinfo[idx];
 145         CDEBUG(D_INODE, DFID"@%p[%d] -> "DFID"@%p: ostid: "DOSTID
 146                " idx: %d gen: %d\n",
 147                PFID(&subhdr->coh_lu.loh_fid), subhdr, idx,
 148                PFID(&hdr->coh_lu.loh_fid), hdr, POSTID(&oinfo->loi_oi),
 149                oinfo->loi_ost_idx, oinfo->loi_ost_gen);
 150
 151         /* reuse ->coh_attr_guard to protect coh_parent change */
 152         spin_lock(&subhdr->coh_attr_guard);
 153         parent = subhdr->coh_parent;
 154         if (!parent) {
 155                 subhdr->coh_parent = hdr;
 156                 spin_unlock(&subhdr->coh_attr_guard);
 157                 subhdr->coh_nesting = hdr->coh_nesting + 1;
 158                 lu_object_ref_add(&stripe->co_lu, "lov-parent", lov);
 159                 r0->lo_sub[idx] = cl2lovsub(stripe);
 160                 r0->lo_sub[idx]->lso_super = lov;
 161                 r0->lo_sub[idx]->lso_index = idx;
 162                 result = 0;
 163         } else {
 164                 struct lu_object  *old_obj;
 165                 struct lov_object *old_lov;
 166                 unsigned int mask = D_INODE;
 167
 168                 spin_unlock(&subhdr->coh_attr_guard);
 169                 old_obj = lu_object_locate(&parent->coh_lu, &lov_device_type);
 170                 LASSERT(old_obj);
 171                 old_lov = cl2lov(lu2cl(old_obj));
 172                 if (old_lov->lo_layout_invalid) {
 173                         /* the object's layout has already changed but isn't
 174                          * refreshed
 175                          */
 176                         lu_object_unhash(env, &stripe->co_lu);
 177                         result = -EAGAIN;
 178                 } else {
 179                         mask = D_ERROR;
 180                         result = -EIO;
 181                 }
 182
 183                 LU_OBJECT_DEBUG(mask, env, &stripe->co_lu,
 184                                 "stripe %d is already owned.", idx);
 185                 LU_OBJECT_DEBUG(mask, env, old_obj, "owned.");
 186                 LU_OBJECT_HEADER(mask, env, lov2lu(lov), "try to own.\n");
 187                 cl_object_put(env, stripe);
 188         }
 189         return result;
 190 }
 191
 192 static int lov_page_slice_fixup(struct lov_object *lov,
 193                                 struct cl_object *stripe)
 194 {
 195         struct cl_object_header *hdr = cl_object_header(&lov->lo_cl);
 196         struct cl_object *o;
 197
 198         cl_object_for_each(o, stripe)
 199                 o->co_slice_off += hdr->coh_page_bufsize;
 200
 201         return cl_object_header(stripe)->coh_page_bufsize;
 202 }
 203
 204 static int lov_init_raid0(const struct lu_env *env,
 205                           struct lov_device *dev, struct lov_object *lov,
 206                           const struct cl_object_conf *conf,
 207                           union  lov_layout_state *state)
 208 {
 209         int result;
 210         int i;
 211
 212         struct cl_object        *stripe;
 213         struct lov_thread_info  *lti     = lov_env_info(env);
 214         struct cl_object_conf   *subconf = &lti->lti_stripe_conf;
 215         struct lov_stripe_md    *lsm     = conf->u.coc_md->lsm;
 216         struct lu_fid      *ofid    = &lti->lti_fid;
 217         struct lov_layout_raid0 *r0      = &state->raid0;
 218
 219         if (lsm->lsm_magic != LOV_MAGIC_V1 && lsm->lsm_magic != LOV_MAGIC_V3) {
 220                 dump_lsm(D_ERROR, lsm);
 221                 LASSERTF(0, "magic mismatch, expected %d/%d, actual %d.\n",
 222                          LOV_MAGIC_V1, LOV_MAGIC_V3, lsm->lsm_magic);
 223         }
 224
 225         LASSERT(!lov->lo_lsm);
 226         lov->lo_lsm = lsm_addref(lsm);
 227         r0->lo_nr  = lsm->lsm_stripe_count;
 228         LASSERT(r0->lo_nr <= lov_targets_nr(dev));
 229
 230         r0->lo_sub = libcfs_kvzalloc(r0->lo_nr * sizeof(r0->lo_sub[0]),
 231                                      GFP_NOFS);
 232         if (r0->lo_sub) {
 233                 int psz = 0;
 234
 235                 result = 0;
 236                 subconf->coc_inode = conf->coc_inode;
 237                 spin_lock_init(&r0->lo_sub_lock);
 238                 /*
 239                  * Create stripe cl_objects.
 240                  */
 241                 for (i = 0; i < r0->lo_nr && result == 0; ++i) {
 242                         struct cl_device *subdev;
 243                         struct lov_oinfo *oinfo = lsm->lsm_oinfo[i];
 244                         int ost_idx = oinfo->loi_ost_idx;
 245
 246                         if (lov_oinfo_is_dummy(oinfo))
 247                                 continue;
 248
 249                         result = ostid_to_fid(ofid, &oinfo->loi_oi,
 250                                               oinfo->loi_ost_idx);
 251                         if (result != 0)
 252                                 goto out;
 253
 254                         subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
 255                         subconf->u.coc_oinfo = oinfo;
 256                         LASSERTF(subdev, "not init ost %d\n", ost_idx);
 257                         /* In the function below, .hs_keycmp resolves to
 258                          * lu_obj_hop_keycmp()
 259                          */
 260                         /* coverity[overrun-buffer-val] */
 261                         stripe = lov_sub_find(env, subdev, ofid, subconf);
 262                         if (!IS_ERR(stripe)) {
 263                                 result = lov_init_sub(env, lov, stripe, r0, i);
 264                                 if (result == -EAGAIN) { /* try again */
 265                                         --i;
 266                                         result = 0;
 267                                         continue;
 268                                 }
 269                         } else {
 270                                 result = PTR_ERR(stripe);
 271                         }
 272
 273                         if (result == 0) {
 274                                 int sz = lov_page_slice_fixup(lov, stripe);
 275
 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         }
 285 out:
 286         return result;
 287 }
 288
 289 static int lov_init_released(const struct lu_env *env,
 290                              struct lov_device *dev, struct lov_object *lov,
 291                              const struct cl_object_conf *conf,
 292                              union  lov_layout_state *state)
 293 {
 294         struct lov_stripe_md *lsm = conf->u.coc_md->lsm;
 295
 296         LASSERT(lsm);
 297         LASSERT(lsm_is_released(lsm));
 298         LASSERT(!lov->lo_lsm);
 299
 300         lov->lo_lsm = lsm_addref(lsm);
 301         return 0;
 302 }
 303
 304 static int lov_delete_empty(const struct lu_env *env, struct lov_object *lov,
 305                             union lov_layout_state *state)
 306 {
 307         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 308
 309         lov_layout_wait(env, lov);
 310         return 0;
 311 }
 312
 313 static void lov_subobject_kill(const struct lu_env *env, struct lov_object *lov,
 314                                struct lovsub_object *los, int idx)
 315 {
 316         struct cl_object        *sub;
 317         struct lov_layout_raid0 *r0;
 318         struct lu_site    *site;
 319         struct lu_site_bkt_data *bkt;
 320         wait_queue_t      *waiter;
 321
 322         r0  = &lov->u.raid0;
 323         LASSERT(r0->lo_sub[idx] == los);
 324
 325         sub  = lovsub2cl(los);
 326         site = sub->co_lu.lo_dev->ld_site;
 327         bkt  = lu_site_bkt_from_fid(site, &sub->co_lu.lo_header->loh_fid);
 328
 329         cl_object_kill(env, sub);
 330         /* release a reference to the sub-object and ... */
 331         lu_object_ref_del(&sub->co_lu, "lov-parent", lov);
 332         cl_object_put(env, sub);
 333
 334         /* ... wait until it is actually destroyed---sub-object clears its
 335          * ->lo_sub[] slot in lovsub_object_fini()
 336          */
 337         if (r0->lo_sub[idx] == los) {
 338                 waiter = &lov_env_info(env)->lti_waiter;
 339                 init_waitqueue_entry(waiter, current);
 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                          */
 346                         set_current_state(TASK_UNINTERRUPTIBLE);
 347                         spin_lock(&r0->lo_sub_lock);
 348                         if (r0->lo_sub[idx] == los) {
 349                                 spin_unlock(&r0->lo_sub_lock);
 350                                 schedule();
 351                         } else {
 352                                 spin_unlock(&r0->lo_sub_lock);
 353                                 set_current_state(TASK_RUNNING);
 354                                 break;
 355                         }
 356                 }
 357                 remove_wait_queue(&bkt->lsb_marche_funebre, waiter);
 358         }
 359         LASSERT(!r0->lo_sub[idx]);
 360 }
 361
 362 static int lov_delete_raid0(const struct lu_env *env, struct lov_object *lov,
 363                             union lov_layout_state *state)
 364 {
 365         struct lov_layout_raid0 *r0 = &state->raid0;
 366         struct lov_stripe_md    *lsm = lov->lo_lsm;
 367         int i;
 368
 369         dump_lsm(D_INODE, lsm);
 370
 371         lov_layout_wait(env, lov);
 372         if (r0->lo_sub) {
 373                 for (i = 0; i < r0->lo_nr; ++i) {
 374                         struct lovsub_object *los = r0->lo_sub[i];
 375
 376                         if (los) {
 377                                 cl_object_prune(env, &los->lso_cl);
 378                                 /*
 379                                  * If top-level object is to be evicted from
 380                                  * the cache, so are its sub-objects.
 381                                  */
 382                                 lov_subobject_kill(env, lov, los, i);
 383                         }
 384                 }
 385         }
 386         return 0;
 387 }
 388
 389 static void lov_fini_empty(const struct lu_env *env, struct lov_object *lov,
 390                            union lov_layout_state *state)
 391 {
 392         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 393 }
 394
 395 static void lov_fini_raid0(const struct lu_env *env, struct lov_object *lov,
 396                            union lov_layout_state *state)
 397 {
 398         struct lov_layout_raid0 *r0 = &state->raid0;
 399
 400         if (r0->lo_sub) {
 401                 kvfree(r0->lo_sub);
 402                 r0->lo_sub = NULL;
 403         }
 404
 405         dump_lsm(D_INODE, lov->lo_lsm);
 406         lov_free_memmd(&lov->lo_lsm);
 407 }
 408
 409 static void lov_fini_released(const struct lu_env *env, struct lov_object *lov,
 410                               union lov_layout_state *state)
 411 {
 412         dump_lsm(D_INODE, lov->lo_lsm);
 413         lov_free_memmd(&lov->lo_lsm);
 414 }
 415
 416 static int lov_print_empty(const struct lu_env *env, void *cookie,
 417                            lu_printer_t p, const struct lu_object *o)
 418 {
 419         (*p)(env, cookie, "empty %d\n", lu2lov(o)->lo_layout_invalid);
 420         return 0;
 421 }
 422
 423 static int lov_print_raid0(const struct lu_env *env, void *cookie,
 424                            lu_printer_t p, const struct lu_object *o)
 425 {
 426         struct lov_object       *lov = lu2lov(o);
 427         struct lov_layout_raid0 *r0  = lov_r0(lov);
 428         struct lov_stripe_md    *lsm = lov->lo_lsm;
 429         int                      i;
 430
 431         (*p)(env, cookie, "stripes: %d, %s, lsm{%p 0x%08X %d %u %u}:\n",
 432              r0->lo_nr, 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         for (i = 0; i < r0->lo_nr; ++i) {
 436                 struct lu_object *sub;
 437
 438                 if (r0->lo_sub[i]) {
 439                         sub = lovsub2lu(r0->lo_sub[i]);
 440                         lu_object_print(env, cookie, p, sub);
 441                 } else {
 442                         (*p)(env, cookie, "sub %d absent\n", i);
 443                 }
 444         }
 445         return 0;
 446 }
 447
 448 static int lov_print_released(const struct lu_env *env, void *cookie,
 449                               lu_printer_t p, const struct lu_object *o)
 450 {
 451         struct lov_object       *lov = lu2lov(o);
 452         struct lov_stripe_md    *lsm = lov->lo_lsm;
 453
 454         (*p)(env, cookie,
 455              "released: %s, lsm{%p 0x%08X %d %u %u}:\n",
 456              lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 457              lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 458              lsm->lsm_stripe_count, lsm->lsm_layout_gen);
 459         return 0;
 460 }
 461
 462 /**
 463  * Implements cl_object_operations::coo_attr_get() method for an object
 464  * without stripes (LLT_EMPTY layout type).
 465  *
 466  * The only attributes this layer is authoritative in this case is
 467  * cl_attr::cat_blocks---it's 0.
 468  */
 469 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
 470                               struct cl_attr *attr)
 471 {
 472         attr->cat_blocks = 0;
 473         return 0;
 474 }
 475
 476 static int lov_attr_get_raid0(const struct lu_env *env, struct cl_object *obj,
 477                               struct cl_attr *attr)
 478 {
 479         struct lov_object       *lov = cl2lov(obj);
 480         struct lov_layout_raid0 *r0 = lov_r0(lov);
 481         struct cl_attr          *lov_attr = &r0->lo_attr;
 482         int                      result = 0;
 483
 484         /* this is called w/o holding type guard mutex, so it must be inside
 485          * an on going IO otherwise lsm may be replaced.
 486          * LU-2117: it turns out there exists one exception. For mmaped files,
 487          * the lock of those files may be requested in the other file's IO
 488          * context, and this function is called in ccc_lock_state(), it will
 489          * hit this assertion.
 490          * Anyway, it's still okay to call attr_get w/o type guard as layout
 491          * can't go if locks exist.
 492          */
 493         /* LASSERT(atomic_read(&lsm->lsm_refc) > 1); */
 494
 495         if (!r0->lo_attr_valid) {
 496                 struct lov_stripe_md    *lsm = lov->lo_lsm;
 497                 struct ost_lvb    *lvb = &lov_env_info(env)->lti_lvb;
 498                 __u64               kms = 0;
 499
 500                 memset(lvb, 0, sizeof(*lvb));
 501                 /* XXX: timestamps can be negative by sanity:test_39m,
 502                  * how can it be?
 503                  */
 504                 lvb->lvb_atime = LLONG_MIN;
 505                 lvb->lvb_ctime = LLONG_MIN;
 506                 lvb->lvb_mtime = LLONG_MIN;
 507
 508                 /*
 509                  * XXX that should be replaced with a loop over sub-objects,
 510                  * doing cl_object_attr_get() on them. But for now, let's
 511                  * reuse old lov code.
 512                  */
 513
 514                 /*
 515                  * XXX take lsm spin-lock to keep lov_merge_lvb_kms()
 516                  * happy. It's not needed, because new code uses
 517                  * ->coh_attr_guard spin-lock to protect consistency of
 518                  * sub-object attributes.
 519                  */
 520                 lov_stripe_lock(lsm);
 521                 result = lov_merge_lvb_kms(lsm, lvb, &kms);
 522                 lov_stripe_unlock(lsm);
 523                 if (result == 0) {
 524                         cl_lvb2attr(lov_attr, lvb);
 525                         lov_attr->cat_kms = kms;
 526                         r0->lo_attr_valid = 1;
 527                 }
 528         }
 529         if (result == 0) { /* merge results */
 530                 attr->cat_blocks = lov_attr->cat_blocks;
 531                 attr->cat_size = lov_attr->cat_size;
 532                 attr->cat_kms = lov_attr->cat_kms;
 533                 if (attr->cat_atime < lov_attr->cat_atime)
 534                         attr->cat_atime = lov_attr->cat_atime;
 535                 if (attr->cat_ctime < lov_attr->cat_ctime)
 536                         attr->cat_ctime = lov_attr->cat_ctime;
 537                 if (attr->cat_mtime < lov_attr->cat_mtime)
 538                         attr->cat_mtime = lov_attr->cat_mtime;
 539         }
 540         return result;
 541 }
 542
 543 static const struct lov_layout_operations lov_dispatch[] = {
 544         [LLT_EMPTY] = {
 545                 .llo_init      = lov_init_empty,
 546                 .llo_delete    = lov_delete_empty,
 547                 .llo_fini      = lov_fini_empty,
 548                 .llo_install   = lov_install_empty,
 549                 .llo_print     = lov_print_empty,
 550                 .llo_page_init = lov_page_init_empty,
 551                 .llo_lock_init = lov_lock_init_empty,
 552                 .llo_io_init   = lov_io_init_empty,
 553                 .llo_getattr   = lov_attr_get_empty
 554         },
 555         [LLT_RAID0] = {
 556                 .llo_init      = lov_init_raid0,
 557                 .llo_delete    = lov_delete_raid0,
 558                 .llo_fini      = lov_fini_raid0,
 559                 .llo_install   = lov_install_raid0,
 560                 .llo_print     = lov_print_raid0,
 561                 .llo_page_init = lov_page_init_raid0,
 562                 .llo_lock_init = lov_lock_init_raid0,
 563                 .llo_io_init   = lov_io_init_raid0,
 564                 .llo_getattr   = lov_attr_get_raid0
 565         },
 566         [LLT_RELEASED] = {
 567                 .llo_init      = lov_init_released,
 568                 .llo_delete    = lov_delete_empty,
 569                 .llo_fini      = lov_fini_released,
 570                 .llo_install   = lov_install_empty,
 571                 .llo_print     = lov_print_released,
 572                 .llo_page_init = lov_page_init_empty,
 573                 .llo_lock_init = lov_lock_init_empty,
 574                 .llo_io_init   = lov_io_init_released,
 575                 .llo_getattr   = lov_attr_get_empty
 576         }
 577 };
 578
 579 /**
 580  * Performs a double-dispatch based on the layout type of an object.
 581  */
 582 #define LOV_2DISPATCH_NOLOCK(obj, op, ...)                            \
 583 ({                                                                    \
 584         struct lov_object                     *__obj = (obj);     \
 585         enum lov_layout_type                __llt;                \
 586                                                                         \
 587         __llt = __obj->lo_type;                                  \
 588         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 589         lov_dispatch[__llt].op(__VA_ARGS__);                        \
 590 })
 591
 592 /**
 593  * Return lov_layout_type associated with a given lsm
 594  */
 595 static enum lov_layout_type lov_type(struct lov_stripe_md *lsm)
 596 {
 597         if (!lsm)
 598                 return LLT_EMPTY;
 599         if (lsm_is_released(lsm))
 600                 return LLT_RELEASED;
 601         return LLT_RAID0;
 602 }
 603
 604 static inline void lov_conf_freeze(struct lov_object *lov)
 605 {
 606         if (lov->lo_owner != current)
 607                 down_read(&lov->lo_type_guard);
 608 }
 609
 610 static inline void lov_conf_thaw(struct lov_object *lov)
 611 {
 612         if (lov->lo_owner != current)
 613                 up_read(&lov->lo_type_guard);
 614 }
 615
 616 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...)                      \
 617 ({                                                                    \
 618         struct lov_object                     *__obj = (obj);     \
 619         int                                  __lock = !!(lock);      \
 620         typeof(lov_dispatch[0].op(__VA_ARGS__)) __result;              \
 621                                                                         \
 622         if (__lock)                                                  \
 623                 lov_conf_freeze(__obj);                                 \
 624         __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__);    \
 625         if (__lock)                                                  \
 626                 lov_conf_thaw(__obj);                                   \
 627         __result;                                                      \
 628 })
 629
 630 /**
 631  * Performs a locked double-dispatch based on the layout type of an object.
 632  */
 633 #define LOV_2DISPATCH(obj, op, ...)                  \
 634         LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
 635
 636 #define LOV_2DISPATCH_VOID(obj, op, ...)                                \
 637 do {                                                                \
 638         struct lov_object                     *__obj = (obj);     \
 639         enum lov_layout_type                __llt;                \
 640                                                                         \
 641         lov_conf_freeze(__obj);                                         \
 642         __llt = __obj->lo_type;                                  \
 643         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 644         lov_dispatch[__llt].op(__VA_ARGS__);                        \
 645         lov_conf_thaw(__obj);                                           \
 646 } while (0)
 647
 648 static void lov_conf_lock(struct lov_object *lov)
 649 {
 650         LASSERT(lov->lo_owner != current);
 651         down_write(&lov->lo_type_guard);
 652         LASSERT(!lov->lo_owner);
 653         lov->lo_owner = current;
 654 }
 655
 656 static void lov_conf_unlock(struct lov_object *lov)
 657 {
 658         lov->lo_owner = NULL;
 659         up_write(&lov->lo_type_guard);
 660 }
 661
 662 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov)
 663 {
 664         struct l_wait_info lwi = { 0 };
 665
 666         while (atomic_read(&lov->lo_active_ios) > 0) {
 667                 CDEBUG(D_INODE, "file:" DFID " wait for active IO, now: %d.\n",
 668                        PFID(lu_object_fid(lov2lu(lov))),
 669                        atomic_read(&lov->lo_active_ios));
 670
 671                 l_wait_event(lov->lo_waitq,
 672                              atomic_read(&lov->lo_active_ios) == 0, &lwi);
 673         }
 674         return 0;
 675 }
 676
 677 static int lov_layout_change(const struct lu_env *unused,
 678                              struct lov_object *lov,
 679                              const struct cl_object_conf *conf)
 680 {
 681         int result;
 682         enum lov_layout_type llt = LLT_EMPTY;
 683         union lov_layout_state *state = &lov->u;
 684         const struct lov_layout_operations *old_ops;
 685         const struct lov_layout_operations *new_ops;
 686
 687         void *cookie;
 688         struct lu_env *env;
 689         int refcheck;
 690
 691         LASSERT(0 <= lov->lo_type && lov->lo_type < ARRAY_SIZE(lov_dispatch));
 692
 693         if (conf->u.coc_md)
 694                 llt = lov_type(conf->u.coc_md->lsm);
 695         LASSERT(0 <= llt && llt < ARRAY_SIZE(lov_dispatch));
 696
 697         cookie = cl_env_reenter();
 698         env = cl_env_get(&refcheck);
 699         if (IS_ERR(env)) {
 700                 cl_env_reexit(cookie);
 701                 return PTR_ERR(env);
 702         }
 703
 704         CDEBUG(D_INODE, DFID" from %s to %s\n",
 705                PFID(lu_object_fid(lov2lu(lov))),
 706                llt2str(lov->lo_type), llt2str(llt));
 707
 708         old_ops = &lov_dispatch[lov->lo_type];
 709         new_ops = &lov_dispatch[llt];
 710
 711         result = cl_object_prune(env, &lov->lo_cl);
 712         if (result != 0)
 713                 goto out;
 714
 715         result = old_ops->llo_delete(env, lov, &lov->u);
 716         if (result == 0) {
 717                 old_ops->llo_fini(env, lov, &lov->u);
 718
 719                 LASSERT(atomic_read(&lov->lo_active_ios) == 0);
 720
 721                 lov->lo_type = LLT_EMPTY;
 722                 result = new_ops->llo_init(env,
 723                                         lu2lov_dev(lov->lo_cl.co_lu.lo_dev),
 724                                         lov, conf, state);
 725                 if (result == 0) {
 726                         new_ops->llo_install(env, lov, state);
 727                         lov->lo_type = llt;
 728                 } else {
 729                         new_ops->llo_delete(env, lov, state);
 730                         new_ops->llo_fini(env, lov, state);
 731                         /* this file becomes an EMPTY file. */
 732                 }
 733         }
 734
 735 out:
 736         cl_env_put(env, &refcheck);
 737         cl_env_reexit(cookie);
 738         return result;
 739 }
 740
 741 /*****************************************************************************
 742  *
 743  * Lov object operations.
 744  *
 745  */
 746 int lov_object_init(const struct lu_env *env, struct lu_object *obj,
 747                     const struct lu_object_conf *conf)
 748 {
 749         struct lov_device           *dev   = lu2lov_dev(obj->lo_dev);
 750         struct lov_object           *lov   = lu2lov(obj);
 751         const struct cl_object_conf  *cconf = lu2cl_conf(conf);
 752         union  lov_layout_state      *set   = &lov->u;
 753         const struct lov_layout_operations *ops;
 754         int result;
 755
 756         init_rwsem(&lov->lo_type_guard);
 757         atomic_set(&lov->lo_active_ios, 0);
 758         init_waitqueue_head(&lov->lo_waitq);
 759
 760         cl_object_page_init(lu2cl(obj), sizeof(struct lov_page));
 761
 762         /* no locking is necessary, as object is being created */
 763         lov->lo_type = lov_type(cconf->u.coc_md->lsm);
 764         ops = &lov_dispatch[lov->lo_type];
 765         result = ops->llo_init(env, dev, lov, cconf, set);
 766         if (result == 0)
 767                 ops->llo_install(env, lov, set);
 768         return result;
 769 }
 770
 771 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
 772                         const struct cl_object_conf *conf)
 773 {
 774         struct lov_stripe_md    *lsm = NULL;
 775         struct lov_object       *lov = cl2lov(obj);
 776         int                      result = 0;
 777
 778         lov_conf_lock(lov);
 779         if (conf->coc_opc == OBJECT_CONF_INVALIDATE) {
 780                 lov->lo_layout_invalid = true;
 781                 result = 0;
 782                 goto out;
 783         }
 784
 785         if (conf->coc_opc == OBJECT_CONF_WAIT) {
 786                 if (lov->lo_layout_invalid &&
 787                     atomic_read(&lov->lo_active_ios) > 0) {
 788                         lov_conf_unlock(lov);
 789                         result = lov_layout_wait(env, lov);
 790                         lov_conf_lock(lov);
 791                 }
 792                 goto out;
 793         }
 794
 795         LASSERT(conf->coc_opc == OBJECT_CONF_SET);
 796
 797         if (conf->u.coc_md)
 798                 lsm = conf->u.coc_md->lsm;
 799         if ((!lsm && !lov->lo_lsm) ||
 800             ((lsm && lov->lo_lsm) &&
 801              (lov->lo_lsm->lsm_layout_gen == lsm->lsm_layout_gen) &&
 802              (lov->lo_lsm->lsm_pattern == lsm->lsm_pattern))) {
 803                 /* same version of layout */
 804                 lov->lo_layout_invalid = false;
 805                 result = 0;
 806                 goto out;
 807         }
 808
 809         /* will change layout - check if there still exists active IO. */
 810         if (atomic_read(&lov->lo_active_ios) > 0) {
 811                 lov->lo_layout_invalid = true;
 812                 result = -EBUSY;
 813                 goto out;
 814         }
 815
 816         result = lov_layout_change(env, lov, conf);
 817         lov->lo_layout_invalid = result != 0;
 818
 819 out:
 820         lov_conf_unlock(lov);
 821         CDEBUG(D_INODE, DFID" lo_layout_invalid=%d\n",
 822                PFID(lu_object_fid(lov2lu(lov))), lov->lo_layout_invalid);
 823         return result;
 824 }
 825
 826 static void lov_object_delete(const struct lu_env *env, struct lu_object *obj)
 827 {
 828         struct lov_object *lov = lu2lov(obj);
 829
 830         LOV_2DISPATCH_VOID(lov, llo_delete, env, lov, &lov->u);
 831 }
 832
 833 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
 834 {
 835         struct lov_object *lov = lu2lov(obj);
 836
 837         LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
 838         lu_object_fini(obj);
 839         kmem_cache_free(lov_object_kmem, lov);
 840 }
 841
 842 static int lov_object_print(const struct lu_env *env, void *cookie,
 843                             lu_printer_t p, const struct lu_object *o)
 844 {
 845         return LOV_2DISPATCH_NOLOCK(lu2lov(o), llo_print, env, cookie, p, o);
 846 }
 847
 848 int lov_page_init(const struct lu_env *env, struct cl_object *obj,
 849                   struct cl_page *page, pgoff_t index)
 850 {
 851         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_page_init, env, obj, page,
 852                                     index);
 853 }
 854
 855 /**
 856  * Implements cl_object_operations::clo_io_init() method for lov
 857  * layer. Dispatches to the appropriate layout io initialization method.
 858  */
 859 int lov_io_init(const struct lu_env *env, struct cl_object *obj,
 860                 struct cl_io *io)
 861 {
 862         CL_IO_SLICE_CLEAN(lov_env_io(env), lis_cl);
 863         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
 864                                      !io->ci_ignore_layout, env, obj, io);
 865 }
 866
 867 /**
 868  * An implementation of cl_object_operations::clo_attr_get() method for lov
 869  * layer. For raid0 layout this collects and merges attributes of all
 870  * sub-objects.
 871  */
 872 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
 873                         struct cl_attr *attr)
 874 {
 875         /* do not take lock, as this function is called under a
 876          * spin-lock. Layout is protected from changing by ongoing IO.
 877          */
 878         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
 879 }
 880
 881 static int lov_attr_set(const struct lu_env *env, struct cl_object *obj,
 882                         const struct cl_attr *attr, unsigned valid)
 883 {
 884         /*
 885          * No dispatch is required here, as no layout implements this.
 886          */
 887         return 0;
 888 }
 889
 890 int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
 891                   struct cl_lock *lock, const struct cl_io *io)
 892 {
 893         /* No need to lock because we've taken one refcount of layout.  */
 894         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_lock_init, env, obj, lock,
 895                                     io);
 896 }
 897
 898 static const struct cl_object_operations lov_ops = {
 899         .coo_page_init = lov_page_init,
 900         .coo_lock_init = lov_lock_init,
 901         .coo_io_init   = lov_io_init,
 902         .coo_attr_get  = lov_attr_get,
 903         .coo_attr_set  = lov_attr_set,
 904         .coo_conf_set  = lov_conf_set
 905 };
 906
 907 static const struct lu_object_operations lov_lu_obj_ops = {
 908         .loo_object_init      = lov_object_init,
 909         .loo_object_delete    = lov_object_delete,
 910         .loo_object_release   = NULL,
 911         .loo_object_free      = lov_object_free,
 912         .loo_object_print     = lov_object_print,
 913         .loo_object_invariant = NULL
 914 };
 915
 916 struct lu_object *lov_object_alloc(const struct lu_env *env,
 917                                    const struct lu_object_header *unused,
 918                                    struct lu_device *dev)
 919 {
 920         struct lov_object *lov;
 921         struct lu_object  *obj;
 922
 923         lov = kmem_cache_zalloc(lov_object_kmem, GFP_NOFS);
 924         if (lov) {
 925                 obj = lov2lu(lov);
 926                 lu_object_init(obj, NULL, dev);
 927                 lov->lo_cl.co_ops = &lov_ops;
 928                 lov->lo_type = -1; /* invalid, to catch uninitialized type */
 929                 /*
 930                  * object io operation vector (cl_object::co_iop) is installed
 931                  * later in lov_object_init(), as different vectors are used
 932                  * for object with different layouts.
 933                  */
 934                 obj->lo_ops = &lov_lu_obj_ops;
 935         } else {
 936                 obj = NULL;
 937         }
 938         return obj;
 939 }
 940
 941 static struct lov_stripe_md *lov_lsm_addref(struct lov_object *lov)
 942 {
 943         struct lov_stripe_md *lsm = NULL;
 944
 945         lov_conf_freeze(lov);
 946         if (lov->lo_lsm) {
 947                 lsm = lsm_addref(lov->lo_lsm);
 948                 CDEBUG(D_INODE, "lsm %p addref %d/%d by %p.\n",
 949                        lsm, atomic_read(&lsm->lsm_refc),
 950                        lov->lo_layout_invalid, current);
 951         }
 952         lov_conf_thaw(lov);
 953         return lsm;
 954 }
 955
 956 struct lov_stripe_md *lov_lsm_get(struct cl_object *clobj)
 957 {
 958         struct lu_object *luobj;
 959         struct lov_stripe_md *lsm = NULL;
 960
 961         if (!clobj)
 962                 return NULL;
 963
 964         luobj = lu_object_locate(&cl_object_header(clobj)->coh_lu,
 965                                  &lov_device_type);
 966         if (luobj)
 967                 lsm = lov_lsm_addref(lu2lov(luobj));
 968         return lsm;
 969 }
 970 EXPORT_SYMBOL(lov_lsm_get);
 971
 972 void lov_lsm_put(struct cl_object *unused, struct lov_stripe_md *lsm)
 973 {
 974         if (lsm)
 975                 lov_free_memmd(&lsm);
 976 }
 977 EXPORT_SYMBOL(lov_lsm_put);
 978
 979 int lov_read_and_clear_async_rc(struct cl_object *clob)
 980 {
 981         struct lu_object *luobj;
 982         int rc = 0;
 983
 984         luobj = lu_object_locate(&cl_object_header(clob)->coh_lu,
 985                                  &lov_device_type);
 986         if (luobj) {
 987                 struct lov_object *lov = lu2lov(luobj);
 988
 989                 lov_conf_freeze(lov);
 990                 switch (lov->lo_type) {
 991                 case LLT_RAID0: {
 992                         struct lov_stripe_md *lsm;
 993                         int i;
 994
 995                         lsm = lov->lo_lsm;
 996                         for (i = 0; i < lsm->lsm_stripe_count; i++) {
 997                                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
 998
 999                                 if (lov_oinfo_is_dummy(loi))
1000                                         continue;
1001
1002                                 if (loi->loi_ar.ar_rc && !rc)
1003                                         rc = loi->loi_ar.ar_rc;
1004                                 loi->loi_ar.ar_rc = 0;
1005                         }
1006                 }
1007                 case LLT_RELEASED:
1008                 case LLT_EMPTY:
1009                         break;
1010                 default:
1011                         LBUG();
1012                 }
1013                 lov_conf_thaw(lov);
1014         }
1015         return rc;
1016 }
1017 EXPORT_SYMBOL(lov_read_and_clear_async_rc);
1018
1019 /** @} lov */