drivers/staging/lustre/lustre/lov/lov_object.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * Implementation of cl_object for LOV layer.
  37  *
  38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  39  *   Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_LOV
  43
  44 #include "lov_cl_internal.h"
  45 #include "../include/lclient.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, struct page *vmpage);
  72         int  (*llo_lock_init)(const struct lu_env *env,
  73                               struct cl_object *obj, struct cl_lock *lock,
  74                               const struct cl_io *io);
  75         int  (*llo_io_init)(const struct lu_env *env,
  76                             struct cl_object *obj, struct cl_io *io);
  77         int  (*llo_getattr)(const struct lu_env *env, struct cl_object *obj,
  78                             struct cl_attr *attr);
  79 };
  80
  81 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov);
  82
  83 /*****************************************************************************
  84  *
  85  * Lov object layout operations.
  86  *
  87  */
  88
  89 static void lov_install_empty(const struct lu_env *env,
  90                               struct lov_object *lov,
  91                               union  lov_layout_state *state)
  92 {
  93         /*
  94          * File without objects.
  95          */
  96 }
  97
  98 static int lov_init_empty(const struct lu_env *env,
  99                           struct lov_device *dev, struct lov_object *lov,
 100                           const struct cl_object_conf *conf,
 101                           union  lov_layout_state *state)
 102 {
 103         return 0;
 104 }
 105
 106 static void lov_install_raid0(const struct lu_env *env,
 107                               struct lov_object *lov,
 108                               union  lov_layout_state *state)
 109 {
 110 }
 111
 112 static struct cl_object *lov_sub_find(const struct lu_env *env,
 113                                       struct cl_device *dev,
 114                                       const struct lu_fid *fid,
 115                                       const struct cl_object_conf *conf)
 116 {
 117         struct lu_object *o;
 118
 119         o = lu_object_find_at(env, cl2lu_dev(dev), fid, &conf->coc_lu);
 120         LASSERT(ergo(!IS_ERR(o), o->lo_dev->ld_type == &lovsub_device_type));
 121         return lu2cl(o);
 122 }
 123
 124 static int lov_init_sub(const struct lu_env *env, struct lov_object *lov,
 125                         struct cl_object *stripe, struct lov_layout_raid0 *r0,
 126                         int idx)
 127 {
 128         struct cl_object_header *hdr;
 129         struct cl_object_header *subhdr;
 130         struct cl_object_header *parent;
 131         struct lov_oinfo        *oinfo;
 132         int result;
 133
 134         if (OBD_FAIL_CHECK(OBD_FAIL_LOV_INIT)) {
 135                 /* For sanity:test_206.
 136                  * Do not leave the object in cache to avoid accessing
 137                  * freed memory. This is because osc_object is referring to
 138                  * lov_oinfo of lsm_stripe_data which will be freed due to
 139                  * this failure. */
 140                 cl_object_kill(env, stripe);
 141                 cl_object_put(env, stripe);
 142                 return -EIO;
 143         }
 144
 145         hdr    = cl_object_header(lov2cl(lov));
 146         subhdr = cl_object_header(stripe);
 147
 148         oinfo = lov->lo_lsm->lsm_oinfo[idx];
 149         CDEBUG(D_INODE, DFID"@%p[%d] -> "DFID"@%p: ostid: "DOSTID
 150                " idx: %d gen: %d\n",
 151                PFID(&subhdr->coh_lu.loh_fid), subhdr, idx,
 152                PFID(&hdr->coh_lu.loh_fid), hdr, POSTID(&oinfo->loi_oi),
 153                oinfo->loi_ost_idx, oinfo->loi_ost_gen);
 154
 155         /* reuse ->coh_attr_guard to protect coh_parent change */
 156         spin_lock(&subhdr->coh_attr_guard);
 157         parent = subhdr->coh_parent;
 158         if (parent == NULL) {
 159                 subhdr->coh_parent = hdr;
 160                 spin_unlock(&subhdr->coh_attr_guard);
 161                 subhdr->coh_nesting = hdr->coh_nesting + 1;
 162                 lu_object_ref_add(&stripe->co_lu, "lov-parent", lov);
 163                 r0->lo_sub[idx] = cl2lovsub(stripe);
 164                 r0->lo_sub[idx]->lso_super = lov;
 165                 r0->lo_sub[idx]->lso_index = idx;
 166                 result = 0;
 167         } else {
 168                 struct lu_object  *old_obj;
 169                 struct lov_object *old_lov;
 170                 unsigned int mask = D_INODE;
 171
 172                 spin_unlock(&subhdr->coh_attr_guard);
 173                 old_obj = lu_object_locate(&parent->coh_lu, &lov_device_type);
 174                 LASSERT(old_obj != NULL);
 175                 old_lov = cl2lov(lu2cl(old_obj));
 176                 if (old_lov->lo_layout_invalid) {
 177                         /* the object's layout has already changed but isn't
 178                          * refreshed */
 179                         lu_object_unhash(env, &stripe->co_lu);
 180                         result = -EAGAIN;
 181                 } else {
 182                         mask = D_ERROR;
 183                         result = -EIO;
 184                 }
 185
 186                 LU_OBJECT_DEBUG(mask, env, &stripe->co_lu,
 187                                 "stripe %d is already owned.\n", idx);
 188                 LU_OBJECT_DEBUG(mask, env, old_obj, "owned.\n");
 189                 LU_OBJECT_HEADER(mask, env, lov2lu(lov), "try to own.\n");
 190                 cl_object_put(env, stripe);
 191         }
 192         return result;
 193 }
 194
 195 static int lov_init_raid0(const struct lu_env *env,
 196                           struct lov_device *dev, struct lov_object *lov,
 197                           const struct cl_object_conf *conf,
 198                           union  lov_layout_state *state)
 199 {
 200         int result;
 201         int i;
 202
 203         struct cl_object        *stripe;
 204         struct lov_thread_info  *lti     = lov_env_info(env);
 205         struct cl_object_conf   *subconf = &lti->lti_stripe_conf;
 206         struct lov_stripe_md    *lsm     = conf->u.coc_md->lsm;
 207         struct lu_fid      *ofid    = &lti->lti_fid;
 208         struct lov_layout_raid0 *r0      = &state->raid0;
 209
 210         if (lsm->lsm_magic != LOV_MAGIC_V1 && lsm->lsm_magic != LOV_MAGIC_V3) {
 211                 dump_lsm(D_ERROR, lsm);
 212                 LASSERTF(0, "magic mismatch, expected %d/%d, actual %d.\n",
 213                          LOV_MAGIC_V1, LOV_MAGIC_V3, lsm->lsm_magic);
 214         }
 215
 216         LASSERT(lov->lo_lsm == NULL);
 217         lov->lo_lsm = lsm_addref(lsm);
 218         r0->lo_nr  = lsm->lsm_stripe_count;
 219         LASSERT(r0->lo_nr <= lov_targets_nr(dev));
 220
 221         r0->lo_sub = libcfs_kvzalloc(r0->lo_nr * sizeof(r0->lo_sub[0]),
 222                                      GFP_NOFS);
 223         if (r0->lo_sub != NULL) {
 224                 result = 0;
 225                 subconf->coc_inode = conf->coc_inode;
 226                 spin_lock_init(&r0->lo_sub_lock);
 227                 /*
 228                  * Create stripe cl_objects.
 229                  */
 230                 for (i = 0; i < r0->lo_nr && result == 0; ++i) {
 231                         struct cl_device *subdev;
 232                         struct lov_oinfo *oinfo = lsm->lsm_oinfo[i];
 233                         int ost_idx = oinfo->loi_ost_idx;
 234
 235                         if (lov_oinfo_is_dummy(oinfo))
 236                                 continue;
 237
 238                         result = ostid_to_fid(ofid, &oinfo->loi_oi,
 239                                               oinfo->loi_ost_idx);
 240                         if (result != 0)
 241                                 goto out;
 242
 243                         subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
 244                         subconf->u.coc_oinfo = oinfo;
 245                         LASSERTF(subdev != NULL, "not init ost %d\n", ost_idx);
 246                         /* In the function below, .hs_keycmp resolves to
 247                          * lu_obj_hop_keycmp() */
 248                         /* coverity[overrun-buffer-val] */
 249                         stripe = lov_sub_find(env, subdev, ofid, subconf);
 250                         if (!IS_ERR(stripe)) {
 251                                 result = lov_init_sub(env, lov, stripe, r0, i);
 252                                 if (result == -EAGAIN) { /* try again */
 253                                         --i;
 254                                         result = 0;
 255                                 }
 256                         } else {
 257                                 result = PTR_ERR(stripe);
 258                         }
 259                 }
 260         } else
 261                 result = -ENOMEM;
 262 out:
 263         return result;
 264 }
 265
 266 static int lov_init_released(const struct lu_env *env,
 267                         struct lov_device *dev, struct lov_object *lov,
 268                         const struct cl_object_conf *conf,
 269                         union  lov_layout_state *state)
 270 {
 271         struct lov_stripe_md *lsm = conf->u.coc_md->lsm;
 272
 273         LASSERT(lsm != NULL);
 274         LASSERT(lsm_is_released(lsm));
 275         LASSERT(lov->lo_lsm == NULL);
 276
 277         lov->lo_lsm = lsm_addref(lsm);
 278         return 0;
 279 }
 280
 281 static int lov_delete_empty(const struct lu_env *env, struct lov_object *lov,
 282                             union lov_layout_state *state)
 283 {
 284         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 285
 286         lov_layout_wait(env, lov);
 287
 288         cl_object_prune(env, &lov->lo_cl);
 289         return 0;
 290 }
 291
 292 static void lov_subobject_kill(const struct lu_env *env, struct lov_object *lov,
 293                                struct lovsub_object *los, int idx)
 294 {
 295         struct cl_object        *sub;
 296         struct lov_layout_raid0 *r0;
 297         struct lu_site    *site;
 298         struct lu_site_bkt_data *bkt;
 299         wait_queue_t      *waiter;
 300
 301         r0  = &lov->u.raid0;
 302         LASSERT(r0->lo_sub[idx] == los);
 303
 304         sub  = lovsub2cl(los);
 305         site = sub->co_lu.lo_dev->ld_site;
 306         bkt  = lu_site_bkt_from_fid(site, &sub->co_lu.lo_header->loh_fid);
 307
 308         cl_object_kill(env, sub);
 309         /* release a reference to the sub-object and ... */
 310         lu_object_ref_del(&sub->co_lu, "lov-parent", lov);
 311         cl_object_put(env, sub);
 312
 313         /* ... wait until it is actually destroyed---sub-object clears its
 314          * ->lo_sub[] slot in lovsub_object_fini() */
 315         if (r0->lo_sub[idx] == los) {
 316                 waiter = &lov_env_info(env)->lti_waiter;
 317                 init_waitqueue_entry(waiter, current);
 318                 add_wait_queue(&bkt->lsb_marche_funebre, waiter);
 319                 set_current_state(TASK_UNINTERRUPTIBLE);
 320                 while (1) {
 321                         /* this wait-queue is signaled at the end of
 322                          * lu_object_free(). */
 323                         set_current_state(TASK_UNINTERRUPTIBLE);
 324                         spin_lock(&r0->lo_sub_lock);
 325                         if (r0->lo_sub[idx] == los) {
 326                                 spin_unlock(&r0->lo_sub_lock);
 327                                 schedule();
 328                         } else {
 329                                 spin_unlock(&r0->lo_sub_lock);
 330                                 set_current_state(TASK_RUNNING);
 331                                 break;
 332                         }
 333                 }
 334                 remove_wait_queue(&bkt->lsb_marche_funebre, waiter);
 335         }
 336         LASSERT(r0->lo_sub[idx] == NULL);
 337 }
 338
 339 static int lov_delete_raid0(const struct lu_env *env, struct lov_object *lov,
 340                             union lov_layout_state *state)
 341 {
 342         struct lov_layout_raid0 *r0 = &state->raid0;
 343         struct lov_stripe_md    *lsm = lov->lo_lsm;
 344         int i;
 345
 346         dump_lsm(D_INODE, lsm);
 347
 348         lov_layout_wait(env, lov);
 349         if (r0->lo_sub != NULL) {
 350                 for (i = 0; i < r0->lo_nr; ++i) {
 351                         struct lovsub_object *los = r0->lo_sub[i];
 352
 353                         if (los != NULL) {
 354                                 cl_locks_prune(env, &los->lso_cl, 1);
 355                                 /*
 356                                  * If top-level object is to be evicted from
 357                                  * the cache, so are its sub-objects.
 358                                  */
 359                                 lov_subobject_kill(env, lov, los, i);
 360                         }
 361                 }
 362         }
 363         cl_object_prune(env, &lov->lo_cl);
 364         return 0;
 365 }
 366
 367 static void lov_fini_empty(const struct lu_env *env, struct lov_object *lov,
 368                            union lov_layout_state *state)
 369 {
 370         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 371 }
 372
 373 static void lov_fini_raid0(const struct lu_env *env, struct lov_object *lov,
 374                            union lov_layout_state *state)
 375 {
 376         struct lov_layout_raid0 *r0 = &state->raid0;
 377
 378         if (r0->lo_sub != NULL) {
 379                 kvfree(r0->lo_sub);
 380                 r0->lo_sub = NULL;
 381         }
 382
 383         dump_lsm(D_INODE, lov->lo_lsm);
 384         lov_free_memmd(&lov->lo_lsm);
 385 }
 386
 387 static void lov_fini_released(const struct lu_env *env, struct lov_object *lov,
 388                                 union lov_layout_state *state)
 389 {
 390         dump_lsm(D_INODE, lov->lo_lsm);
 391         lov_free_memmd(&lov->lo_lsm);
 392 }
 393
 394 static int lov_print_empty(const struct lu_env *env, void *cookie,
 395                            lu_printer_t p, const struct lu_object *o)
 396 {
 397         (*p)(env, cookie, "empty %d\n", lu2lov(o)->lo_layout_invalid);
 398         return 0;
 399 }
 400
 401 static int lov_print_raid0(const struct lu_env *env, void *cookie,
 402                            lu_printer_t p, const struct lu_object *o)
 403 {
 404         struct lov_object       *lov = lu2lov(o);
 405         struct lov_layout_raid0 *r0  = lov_r0(lov);
 406         struct lov_stripe_md    *lsm = lov->lo_lsm;
 407         int                      i;
 408
 409         (*p)(env, cookie, "stripes: %d, %s, lsm{%p 0x%08X %d %u %u}:\n",
 410                 r0->lo_nr, lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 411                 lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 412                 lsm->lsm_stripe_count, lsm->lsm_layout_gen);
 413         for (i = 0; i < r0->lo_nr; ++i) {
 414                 struct lu_object *sub;
 415
 416                 if (r0->lo_sub[i] != NULL) {
 417                         sub = lovsub2lu(r0->lo_sub[i]);
 418                         lu_object_print(env, cookie, p, sub);
 419                 } else {
 420                         (*p)(env, cookie, "sub %d absent\n", i);
 421                 }
 422         }
 423         return 0;
 424 }
 425
 426 static int lov_print_released(const struct lu_env *env, void *cookie,
 427                                 lu_printer_t p, const struct lu_object *o)
 428 {
 429         struct lov_object       *lov = lu2lov(o);
 430         struct lov_stripe_md    *lsm = lov->lo_lsm;
 431
 432         (*p)(env, cookie,
 433                 "released: %s, lsm{%p 0x%08X %d %u %u}:\n",
 434                 lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 435                 lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 436                 lsm->lsm_stripe_count, lsm->lsm_layout_gen);
 437         return 0;
 438 }
 439
 440 /**
 441  * Implements cl_object_operations::coo_attr_get() method for an object
 442  * without stripes (LLT_EMPTY layout type).
 443  *
 444  * The only attributes this layer is authoritative in this case is
 445  * cl_attr::cat_blocks---it's 0.
 446  */
 447 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
 448                               struct cl_attr *attr)
 449 {
 450         attr->cat_blocks = 0;
 451         return 0;
 452 }
 453
 454 static int lov_attr_get_raid0(const struct lu_env *env, struct cl_object *obj,
 455                               struct cl_attr *attr)
 456 {
 457         struct lov_object       *lov = cl2lov(obj);
 458         struct lov_layout_raid0 *r0 = lov_r0(lov);
 459         struct cl_attr          *lov_attr = &r0->lo_attr;
 460         int                      result = 0;
 461
 462         /* this is called w/o holding type guard mutex, so it must be inside
 463          * an on going IO otherwise lsm may be replaced.
 464          * LU-2117: it turns out there exists one exception. For mmaped files,
 465          * the lock of those files may be requested in the other file's IO
 466          * context, and this function is called in ccc_lock_state(), it will
 467          * hit this assertion.
 468          * Anyway, it's still okay to call attr_get w/o type guard as layout
 469          * can't go if locks exist. */
 470         /* LASSERT(atomic_read(&lsm->lsm_refc) > 1); */
 471
 472         if (!r0->lo_attr_valid) {
 473                 struct lov_stripe_md    *lsm = lov->lo_lsm;
 474                 struct ost_lvb    *lvb = &lov_env_info(env)->lti_lvb;
 475                 __u64               kms = 0;
 476
 477                 memset(lvb, 0, sizeof(*lvb));
 478                 /* XXX: timestamps can be negative by sanity:test_39m,
 479                  * how can it be? */
 480                 lvb->lvb_atime = LLONG_MIN;
 481                 lvb->lvb_ctime = LLONG_MIN;
 482                 lvb->lvb_mtime = LLONG_MIN;
 483
 484                 /*
 485                  * XXX that should be replaced with a loop over sub-objects,
 486                  * doing cl_object_attr_get() on them. But for now, let's
 487                  * reuse old lov code.
 488                  */
 489
 490                 /*
 491                  * XXX take lsm spin-lock to keep lov_merge_lvb_kms()
 492                  * happy. It's not needed, because new code uses
 493                  * ->coh_attr_guard spin-lock to protect consistency of
 494                  * sub-object attributes.
 495                  */
 496                 lov_stripe_lock(lsm);
 497                 result = lov_merge_lvb_kms(lsm, lvb, &kms);
 498                 lov_stripe_unlock(lsm);
 499                 if (result == 0) {
 500                         cl_lvb2attr(lov_attr, lvb);
 501                         lov_attr->cat_kms = kms;
 502                         r0->lo_attr_valid = 1;
 503                 }
 504         }
 505         if (result == 0) { /* merge results */
 506                 attr->cat_blocks = lov_attr->cat_blocks;
 507                 attr->cat_size = lov_attr->cat_size;
 508                 attr->cat_kms = lov_attr->cat_kms;
 509                 if (attr->cat_atime < lov_attr->cat_atime)
 510                         attr->cat_atime = lov_attr->cat_atime;
 511                 if (attr->cat_ctime < lov_attr->cat_ctime)
 512                         attr->cat_ctime = lov_attr->cat_ctime;
 513                 if (attr->cat_mtime < lov_attr->cat_mtime)
 514                         attr->cat_mtime = lov_attr->cat_mtime;
 515         }
 516         return result;
 517 }
 518
 519 static const struct lov_layout_operations lov_dispatch[] = {
 520         [LLT_EMPTY] = {
 521                 .llo_init      = lov_init_empty,
 522                 .llo_delete    = lov_delete_empty,
 523                 .llo_fini      = lov_fini_empty,
 524                 .llo_install   = lov_install_empty,
 525                 .llo_print     = lov_print_empty,
 526                 .llo_page_init = lov_page_init_empty,
 527                 .llo_lock_init = lov_lock_init_empty,
 528                 .llo_io_init   = lov_io_init_empty,
 529                 .llo_getattr   = lov_attr_get_empty
 530         },
 531         [LLT_RAID0] = {
 532                 .llo_init      = lov_init_raid0,
 533                 .llo_delete    = lov_delete_raid0,
 534                 .llo_fini      = lov_fini_raid0,
 535                 .llo_install   = lov_install_raid0,
 536                 .llo_print     = lov_print_raid0,
 537                 .llo_page_init = lov_page_init_raid0,
 538                 .llo_lock_init = lov_lock_init_raid0,
 539                 .llo_io_init   = lov_io_init_raid0,
 540                 .llo_getattr   = lov_attr_get_raid0
 541         },
 542         [LLT_RELEASED] = {
 543                 .llo_init      = lov_init_released,
 544                 .llo_delete    = lov_delete_empty,
 545                 .llo_fini      = lov_fini_released,
 546                 .llo_install   = lov_install_empty,
 547                 .llo_print     = lov_print_released,
 548                 .llo_page_init = lov_page_init_empty,
 549                 .llo_lock_init = lov_lock_init_empty,
 550                 .llo_io_init   = lov_io_init_released,
 551                 .llo_getattr   = lov_attr_get_empty
 552         }
 553 };
 554
 555 /**
 556  * Performs a double-dispatch based on the layout type of an object.
 557  */
 558 #define LOV_2DISPATCH_NOLOCK(obj, op, ...)                            \
 559 ({                                                                    \
 560         struct lov_object                     *__obj = (obj);     \
 561         enum lov_layout_type                __llt;                \
 562                                                                         \
 563         __llt = __obj->lo_type;                                  \
 564         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 565         lov_dispatch[__llt].op(__VA_ARGS__);                        \
 566 })
 567
 568 /**
 569  * Return lov_layout_type associated with a given lsm
 570  */
 571 static enum lov_layout_type lov_type(struct lov_stripe_md *lsm)
 572 {
 573         if (lsm == NULL)
 574                 return LLT_EMPTY;
 575         if (lsm_is_released(lsm))
 576                 return LLT_RELEASED;
 577         return LLT_RAID0;
 578 }
 579
 580 static inline void lov_conf_freeze(struct lov_object *lov)
 581 {
 582         if (lov->lo_owner != current)
 583                 down_read(&lov->lo_type_guard);
 584 }
 585
 586 static inline void lov_conf_thaw(struct lov_object *lov)
 587 {
 588         if (lov->lo_owner != current)
 589                 up_read(&lov->lo_type_guard);
 590 }
 591
 592 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...)                      \
 593 ({                                                                    \
 594         struct lov_object                     *__obj = (obj);     \
 595         int                                  __lock = !!(lock);      \
 596         typeof(lov_dispatch[0].op(__VA_ARGS__)) __result;              \
 597                                                                         \
 598         if (__lock)                                                  \
 599                 lov_conf_freeze(__obj);                                 \
 600         __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__);    \
 601         if (__lock)                                                  \
 602                 lov_conf_thaw(__obj);                                   \
 603         __result;                                                      \
 604 })
 605
 606 /**
 607  * Performs a locked double-dispatch based on the layout type of an object.
 608  */
 609 #define LOV_2DISPATCH(obj, op, ...)                  \
 610         LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
 611
 612 #define LOV_2DISPATCH_VOID(obj, op, ...)                                \
 613 do {                                                                \
 614         struct lov_object                     *__obj = (obj);     \
 615         enum lov_layout_type                __llt;                \
 616                                                                         \
 617         lov_conf_freeze(__obj);                                         \
 618         __llt = __obj->lo_type;                                  \
 619         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 620         lov_dispatch[__llt].op(__VA_ARGS__);                        \
 621         lov_conf_thaw(__obj);                                           \
 622 } while (0)
 623
 624 static void lov_conf_lock(struct lov_object *lov)
 625 {
 626         LASSERT(lov->lo_owner != current);
 627         down_write(&lov->lo_type_guard);
 628         LASSERT(lov->lo_owner == NULL);
 629         lov->lo_owner = current;
 630 }
 631
 632 static void lov_conf_unlock(struct lov_object *lov)
 633 {
 634         lov->lo_owner = NULL;
 635         up_write(&lov->lo_type_guard);
 636 }
 637
 638 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov)
 639 {
 640         struct l_wait_info lwi = { 0 };
 641
 642         while (atomic_read(&lov->lo_active_ios) > 0) {
 643                 CDEBUG(D_INODE, "file:"DFID" wait for active IO, now: %d.\n",
 644                         PFID(lu_object_fid(lov2lu(lov))),
 645                         atomic_read(&lov->lo_active_ios));
 646
 647                 l_wait_event(lov->lo_waitq,
 648                              atomic_read(&lov->lo_active_ios) == 0, &lwi);
 649         }
 650         return 0;
 651 }
 652
 653 static int lov_layout_change(const struct lu_env *unused,
 654                              struct lov_object *lov,
 655                              const struct cl_object_conf *conf)
 656 {
 657         int result;
 658         enum lov_layout_type llt = LLT_EMPTY;
 659         union lov_layout_state *state = &lov->u;
 660         const struct lov_layout_operations *old_ops;
 661         const struct lov_layout_operations *new_ops;
 662
 663         struct cl_object_header *hdr = cl_object_header(&lov->lo_cl);
 664         void *cookie;
 665         struct lu_env *env;
 666         int refcheck;
 667
 668         LASSERT(0 <= lov->lo_type && lov->lo_type < ARRAY_SIZE(lov_dispatch));
 669
 670         if (conf->u.coc_md != NULL)
 671                 llt = lov_type(conf->u.coc_md->lsm);
 672         LASSERT(0 <= llt && llt < ARRAY_SIZE(lov_dispatch));
 673
 674         cookie = cl_env_reenter();
 675         env = cl_env_get(&refcheck);
 676         if (IS_ERR(env)) {
 677                 cl_env_reexit(cookie);
 678                 return PTR_ERR(env);
 679         }
 680
 681         CDEBUG(D_INODE, DFID" from %s to %s\n",
 682                PFID(lu_object_fid(lov2lu(lov))),
 683                llt2str(lov->lo_type), llt2str(llt));
 684
 685         old_ops = &lov_dispatch[lov->lo_type];
 686         new_ops = &lov_dispatch[llt];
 687
 688         result = old_ops->llo_delete(env, lov, &lov->u);
 689         if (result == 0) {
 690                 old_ops->llo_fini(env, lov, &lov->u);
 691
 692                 LASSERT(atomic_read(&lov->lo_active_ios) == 0);
 693                 LASSERT(hdr->coh_tree.rnode == NULL);
 694                 LASSERT(hdr->coh_pages == 0);
 695
 696                 lov->lo_type = LLT_EMPTY;
 697                 result = new_ops->llo_init(env,
 698                                         lu2lov_dev(lov->lo_cl.co_lu.lo_dev),
 699                                         lov, conf, state);
 700                 if (result == 0) {
 701                         new_ops->llo_install(env, lov, state);
 702                         lov->lo_type = llt;
 703                 } else {
 704                         new_ops->llo_delete(env, lov, state);
 705                         new_ops->llo_fini(env, lov, state);
 706                         /* this file becomes an EMPTY file. */
 707                 }
 708         }
 709
 710         cl_env_put(env, &refcheck);
 711         cl_env_reexit(cookie);
 712         return result;
 713 }
 714
 715 /*****************************************************************************
 716  *
 717  * Lov object operations.
 718  *
 719  */
 720 int lov_object_init(const struct lu_env *env, struct lu_object *obj,
 721                     const struct lu_object_conf *conf)
 722 {
 723         struct lov_device           *dev   = lu2lov_dev(obj->lo_dev);
 724         struct lov_object           *lov   = lu2lov(obj);
 725         const struct cl_object_conf  *cconf = lu2cl_conf(conf);
 726         union  lov_layout_state      *set   = &lov->u;
 727         const struct lov_layout_operations *ops;
 728         int result;
 729
 730         init_rwsem(&lov->lo_type_guard);
 731         atomic_set(&lov->lo_active_ios, 0);
 732         init_waitqueue_head(&lov->lo_waitq);
 733
 734         cl_object_page_init(lu2cl(obj), sizeof(struct lov_page));
 735
 736         /* no locking is necessary, as object is being created */
 737         lov->lo_type = lov_type(cconf->u.coc_md->lsm);
 738         ops = &lov_dispatch[lov->lo_type];
 739         result = ops->llo_init(env, dev, lov, cconf, set);
 740         if (result == 0)
 741                 ops->llo_install(env, lov, set);
 742         return result;
 743 }
 744
 745 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
 746                         const struct cl_object_conf *conf)
 747 {
 748         struct lov_stripe_md    *lsm = NULL;
 749         struct lov_object       *lov = cl2lov(obj);
 750         int                      result = 0;
 751
 752         lov_conf_lock(lov);
 753         if (conf->coc_opc == OBJECT_CONF_INVALIDATE) {
 754                 lov->lo_layout_invalid = true;
 755                 result = 0;
 756                 goto out;
 757         }
 758
 759         if (conf->coc_opc == OBJECT_CONF_WAIT) {
 760                 if (lov->lo_layout_invalid &&
 761                     atomic_read(&lov->lo_active_ios) > 0) {
 762                         lov_conf_unlock(lov);
 763                         result = lov_layout_wait(env, lov);
 764                         lov_conf_lock(lov);
 765                 }
 766                 goto out;
 767         }
 768
 769         LASSERT(conf->coc_opc == OBJECT_CONF_SET);
 770
 771         if (conf->u.coc_md != NULL)
 772                 lsm = conf->u.coc_md->lsm;
 773         if ((lsm == NULL && lov->lo_lsm == NULL) ||
 774             ((lsm != NULL && lov->lo_lsm != NULL) &&
 775              (lov->lo_lsm->lsm_layout_gen == lsm->lsm_layout_gen) &&
 776              (lov->lo_lsm->lsm_pattern == lsm->lsm_pattern))) {
 777                 /* same version of layout */
 778                 lov->lo_layout_invalid = false;
 779                 result = 0;
 780                 goto out;
 781         }
 782
 783         /* will change layout - check if there still exists active IO. */
 784         if (atomic_read(&lov->lo_active_ios) > 0) {
 785                 lov->lo_layout_invalid = true;
 786                 result = -EBUSY;
 787                 goto out;
 788         }
 789
 790         lov->lo_layout_invalid = lov_layout_change(env, lov, conf);
 791
 792 out:
 793         lov_conf_unlock(lov);
 794         CDEBUG(D_INODE, DFID" lo_layout_invalid=%d\n",
 795                PFID(lu_object_fid(lov2lu(lov))), lov->lo_layout_invalid);
 796         return result;
 797 }
 798
 799 static void lov_object_delete(const struct lu_env *env, struct lu_object *obj)
 800 {
 801         struct lov_object *lov = lu2lov(obj);
 802
 803         LOV_2DISPATCH_VOID(lov, llo_delete, env, lov, &lov->u);
 804 }
 805
 806 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
 807 {
 808         struct lov_object *lov = lu2lov(obj);
 809
 810         LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
 811         lu_object_fini(obj);
 812         OBD_SLAB_FREE_PTR(lov, lov_object_kmem);
 813 }
 814
 815 static int lov_object_print(const struct lu_env *env, void *cookie,
 816                             lu_printer_t p, const struct lu_object *o)
 817 {
 818         return LOV_2DISPATCH_NOLOCK(lu2lov(o), llo_print, env, cookie, p, o);
 819 }
 820
 821 int lov_page_init(const struct lu_env *env, struct cl_object *obj,
 822                 struct cl_page *page, struct page *vmpage)
 823 {
 824         return LOV_2DISPATCH_NOLOCK(cl2lov(obj),
 825                                     llo_page_init, env, obj, page, vmpage);
 826 }
 827
 828 /**
 829  * Implements cl_object_operations::clo_io_init() method for lov
 830  * layer. Dispatches to the appropriate layout io initialization method.
 831  */
 832 int lov_io_init(const struct lu_env *env, struct cl_object *obj,
 833                 struct cl_io *io)
 834 {
 835         CL_IO_SLICE_CLEAN(lov_env_io(env), lis_cl);
 836         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
 837                                      !io->ci_ignore_layout, env, obj, io);
 838 }
 839
 840 /**
 841  * An implementation of cl_object_operations::clo_attr_get() method for lov
 842  * layer. For raid0 layout this collects and merges attributes of all
 843  * sub-objects.
 844  */
 845 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
 846                         struct cl_attr *attr)
 847 {
 848         /* do not take lock, as this function is called under a
 849          * spin-lock. Layout is protected from changing by ongoing IO. */
 850         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
 851 }
 852
 853 static int lov_attr_set(const struct lu_env *env, struct cl_object *obj,
 854                         const struct cl_attr *attr, unsigned valid)
 855 {
 856         /*
 857          * No dispatch is required here, as no layout implements this.
 858          */
 859         return 0;
 860 }
 861
 862 int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
 863                   struct cl_lock *lock, const struct cl_io *io)
 864 {
 865         /* No need to lock because we've taken one refcount of layout.  */
 866         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_lock_init, env, obj, lock,
 867                                     io);
 868 }
 869
 870 static const struct cl_object_operations lov_ops = {
 871         .coo_page_init = lov_page_init,
 872         .coo_lock_init = lov_lock_init,
 873         .coo_io_init   = lov_io_init,
 874         .coo_attr_get  = lov_attr_get,
 875         .coo_attr_set  = lov_attr_set,
 876         .coo_conf_set  = lov_conf_set
 877 };
 878
 879 static const struct lu_object_operations lov_lu_obj_ops = {
 880         .loo_object_init      = lov_object_init,
 881         .loo_object_delete    = lov_object_delete,
 882         .loo_object_release   = NULL,
 883         .loo_object_free      = lov_object_free,
 884         .loo_object_print     = lov_object_print,
 885         .loo_object_invariant = NULL
 886 };
 887
 888 struct lu_object *lov_object_alloc(const struct lu_env *env,
 889                                    const struct lu_object_header *unused,
 890                                    struct lu_device *dev)
 891 {
 892         struct lov_object *lov;
 893         struct lu_object  *obj;
 894
 895         OBD_SLAB_ALLOC_PTR_GFP(lov, lov_object_kmem, GFP_NOFS);
 896         if (lov != NULL) {
 897                 obj = lov2lu(lov);
 898                 lu_object_init(obj, NULL, dev);
 899                 lov->lo_cl.co_ops = &lov_ops;
 900                 lov->lo_type = -1; /* invalid, to catch uninitialized type */
 901                 /*
 902                  * object io operation vector (cl_object::co_iop) is installed
 903                  * later in lov_object_init(), as different vectors are used
 904                  * for object with different layouts.
 905                  */
 906                 obj->lo_ops = &lov_lu_obj_ops;
 907         } else
 908                 obj = NULL;
 909         return obj;
 910 }
 911
 912 struct lov_stripe_md *lov_lsm_addref(struct lov_object *lov)
 913 {
 914         struct lov_stripe_md *lsm = NULL;
 915
 916         lov_conf_freeze(lov);
 917         if (lov->lo_lsm != NULL) {
 918                 lsm = lsm_addref(lov->lo_lsm);
 919                 CDEBUG(D_INODE, "lsm %p addref %d/%d by %p.\n",
 920                         lsm, atomic_read(&lsm->lsm_refc),
 921                         lov->lo_layout_invalid, current);
 922         }
 923         lov_conf_thaw(lov);
 924         return lsm;
 925 }
 926
 927 void lov_lsm_decref(struct lov_object *lov, struct lov_stripe_md *lsm)
 928 {
 929         if (lsm == NULL)
 930                 return;
 931
 932         CDEBUG(D_INODE, "lsm %p decref %d by %p.\n",
 933                 lsm, atomic_read(&lsm->lsm_refc), current);
 934
 935         lov_free_memmd(&lsm);
 936 }
 937
 938 struct lov_stripe_md *lov_lsm_get(struct cl_object *clobj)
 939 {
 940         struct lu_object *luobj;
 941         struct lov_stripe_md *lsm = NULL;
 942
 943         if (clobj == NULL)
 944                 return NULL;
 945
 946         luobj = lu_object_locate(&cl_object_header(clobj)->coh_lu,
 947                                  &lov_device_type);
 948         if (luobj != NULL)
 949                 lsm = lov_lsm_addref(lu2lov(luobj));
 950         return lsm;
 951 }
 952 EXPORT_SYMBOL(lov_lsm_get);
 953
 954 void lov_lsm_put(struct cl_object *unused, struct lov_stripe_md *lsm)
 955 {
 956         if (lsm != NULL)
 957                 lov_free_memmd(&lsm);
 958 }
 959 EXPORT_SYMBOL(lov_lsm_put);
 960
 961 int lov_read_and_clear_async_rc(struct cl_object *clob)
 962 {
 963         struct lu_object *luobj;
 964         int rc = 0;
 965
 966         luobj = lu_object_locate(&cl_object_header(clob)->coh_lu,
 967                                  &lov_device_type);
 968         if (luobj != NULL) {
 969                 struct lov_object *lov = lu2lov(luobj);
 970
 971                 lov_conf_freeze(lov);
 972                 switch (lov->lo_type) {
 973                 case LLT_RAID0: {
 974                         struct lov_stripe_md *lsm;
 975                         int i;
 976
 977                         lsm = lov->lo_lsm;
 978                         LASSERT(lsm != NULL);
 979                         for (i = 0; i < lsm->lsm_stripe_count; i++) {
 980                                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
 981
 982                                 if (lov_oinfo_is_dummy(loi))
 983                                         continue;
 984
 985                                 if (loi->loi_ar.ar_rc && !rc)
 986                                         rc = loi->loi_ar.ar_rc;
 987                                 loi->loi_ar.ar_rc = 0;
 988                         }
 989                 }
 990                 case LLT_RELEASED:
 991                 case LLT_EMPTY:
 992                         break;
 993                 default:
 994                         LBUG();
 995                 }
 996                 lov_conf_thaw(lov);
 997         }
 998         return rc;
 999 }
1000 EXPORT_SYMBOL(lov_read_and_clear_async_rc);
1001
1002 /** @} lov */