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, cfs_page_t *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         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,
 127                         struct lov_layout_raid0 *r0, 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         parent = subhdr->coh_parent;
 149
 150         oinfo = lov->lo_lsm->lsm_oinfo[idx];
 151         CDEBUG(D_INODE, DFID"@%p[%d] -> "DFID"@%p: ostid: "DOSTID
 152                " idx: %d gen: %d\n",
 153                PFID(&subhdr->coh_lu.loh_fid), subhdr, idx,
 154                PFID(&hdr->coh_lu.loh_fid), hdr, POSTID(&oinfo->loi_oi),
 155                oinfo->loi_ost_idx, oinfo->loi_ost_gen);
 156
 157         if (parent == NULL) {
 158                 subhdr->coh_parent = hdr;
 159                 subhdr->coh_nesting = hdr->coh_nesting + 1;
 160                 lu_object_ref_add(&stripe->co_lu, "lov-parent", lov);
 161                 r0->lo_sub[idx] = cl2lovsub(stripe);
 162                 r0->lo_sub[idx]->lso_super = lov;
 163                 r0->lo_sub[idx]->lso_index = idx;
 164                 result = 0;
 165         } else {
 166                 struct lu_object  *old_obj;
 167                 struct lov_object *old_lov;
 168                 unsigned int mask = D_INODE;
 169
 170                 old_obj = lu_object_locate(&parent->coh_lu, &lov_device_type);
 171                 LASSERT(old_obj != NULL);
 172                 old_lov = cl2lov(lu2cl(old_obj));
 173                 if (old_lov->lo_layout_invalid) {
 174                         /* the object's layout has already changed but isn't
 175                          * refreshed */
 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.\n", idx);
 185                 LU_OBJECT_DEBUG(mask, env, old_obj, "owned.\n");
 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_init_raid0(const struct lu_env *env,
 193                           struct lov_device *dev, struct lov_object *lov,
 194                           const struct cl_object_conf *conf,
 195                           union  lov_layout_state *state)
 196 {
 197         int result;
 198         int i;
 199
 200         struct cl_object        *stripe;
 201         struct lov_thread_info  *lti     = lov_env_info(env);
 202         struct cl_object_conf   *subconf = &lti->lti_stripe_conf;
 203         struct lov_stripe_md    *lsm     = conf->u.coc_md->lsm;
 204         struct lu_fid           *ofid    = &lti->lti_fid;
 205         struct lov_layout_raid0 *r0      = &state->raid0;
 206
 207         ENTRY;
 208
 209         if (lsm->lsm_magic != LOV_MAGIC_V1 && lsm->lsm_magic != LOV_MAGIC_V3) {
 210                 dump_lsm(D_ERROR, lsm);
 211                 LASSERTF(0, "magic mismatch, expected %d/%d, actual %d.\n",
 212                          LOV_MAGIC_V1, LOV_MAGIC_V3, lsm->lsm_magic);
 213         }
 214
 215         LASSERT(lov->lo_lsm == NULL);
 216         lov->lo_lsm = lsm_addref(lsm);
 217         r0->lo_nr  = lsm->lsm_stripe_count;
 218         LASSERT(r0->lo_nr <= lov_targets_nr(dev));
 219
 220         OBD_ALLOC_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 221         if (r0->lo_sub != NULL) {
 222                 result = 0;
 223                 subconf->coc_inode = conf->coc_inode;
 224                 spin_lock_init(&r0->lo_sub_lock);
 225                 /*
 226                  * Create stripe cl_objects.
 227                  */
 228                 for (i = 0; i < r0->lo_nr && result == 0; ++i) {
 229                         struct cl_device *subdev;
 230                         struct lov_oinfo *oinfo = lsm->lsm_oinfo[i];
 231                         int ost_idx = oinfo->loi_ost_idx;
 232
 233                         result = ostid_to_fid(ofid, &oinfo->loi_oi,
 234                                               oinfo->loi_ost_idx);
 235                         if (result != 0)
 236                                 GOTO(out, result);
 237
 238                         subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
 239                         subconf->u.coc_oinfo = oinfo;
 240                         LASSERTF(subdev != NULL, "not init ost %d\n", ost_idx);
 241                         /* In the function below, .hs_keycmp resolves to
 242                          * lu_obj_hop_keycmp() */
 243                         /* coverity[overrun-buffer-val] */
 244                         stripe = lov_sub_find(env, subdev, ofid, subconf);
 245                         if (!IS_ERR(stripe)) {
 246                                 result = lov_init_sub(env, lov, stripe, r0, i);
 247                                 if (result == -EAGAIN) { /* try again */
 248                                         --i;
 249                                         result = 0;
 250                                 }
 251                         } else {
 252                                 result = PTR_ERR(stripe);
 253                         }
 254                 }
 255         } else
 256                 result = -ENOMEM;
 257 out:
 258         RETURN(result);
 259 }
 260
 261 static int lov_init_released(const struct lu_env *env,
 262                         struct lov_device *dev, struct lov_object *lov,
 263                         const struct cl_object_conf *conf,
 264                         union  lov_layout_state *state)
 265 {
 266         struct lov_stripe_md *lsm = conf->u.coc_md->lsm;
 267
 268         LASSERT(lsm != NULL);
 269         LASSERT(lsm_is_released(lsm));
 270         LASSERT(lov->lo_lsm == NULL);
 271
 272         lov->lo_lsm = lsm_addref(lsm);
 273         return 0;
 274 }
 275
 276 static int lov_delete_empty(const struct lu_env *env, struct lov_object *lov,
 277                             union lov_layout_state *state)
 278 {
 279         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 280
 281         lov_layout_wait(env, lov);
 282
 283         cl_object_prune(env, &lov->lo_cl);
 284         return 0;
 285 }
 286
 287 static void lov_subobject_kill(const struct lu_env *env, struct lov_object *lov,
 288                                struct lovsub_object *los, int idx)
 289 {
 290         struct cl_object        *sub;
 291         struct lov_layout_raid0 *r0;
 292         struct lu_site          *site;
 293         struct lu_site_bkt_data *bkt;
 294         cfs_waitlink_t          *waiter;
 295
 296         r0  = &lov->u.raid0;
 297         LASSERT(r0->lo_sub[idx] == los);
 298
 299         sub  = lovsub2cl(los);
 300         site = sub->co_lu.lo_dev->ld_site;
 301         bkt  = lu_site_bkt_from_fid(site, &sub->co_lu.lo_header->loh_fid);
 302
 303         cl_object_kill(env, sub);
 304         /* release a reference to the sub-object and ... */
 305         lu_object_ref_del(&sub->co_lu, "lov-parent", lov);
 306         cl_object_put(env, sub);
 307
 308         /* ... wait until it is actually destroyed---sub-object clears its
 309          * ->lo_sub[] slot in lovsub_object_fini() */
 310         if (r0->lo_sub[idx] == los) {
 311                 waiter = &lov_env_info(env)->lti_waiter;
 312                 cfs_waitlink_init(waiter);
 313                 cfs_waitq_add(&bkt->lsb_marche_funebre, waiter);
 314                 cfs_set_current_state(CFS_TASK_UNINT);
 315                 while (1) {
 316                         /* this wait-queue is signaled at the end of
 317                          * lu_object_free(). */
 318                         cfs_set_current_state(CFS_TASK_UNINT);
 319                         spin_lock(&r0->lo_sub_lock);
 320                         if (r0->lo_sub[idx] == los) {
 321                                 spin_unlock(&r0->lo_sub_lock);
 322                                 cfs_waitq_wait(waiter, CFS_TASK_UNINT);
 323                         } else {
 324                                 spin_unlock(&r0->lo_sub_lock);
 325                                 cfs_set_current_state(CFS_TASK_RUNNING);
 326                                 break;
 327                         }
 328                 }
 329                 cfs_waitq_del(&bkt->lsb_marche_funebre, waiter);
 330         }
 331         LASSERT(r0->lo_sub[idx] == NULL);
 332 }
 333
 334 static int lov_delete_raid0(const struct lu_env *env, struct lov_object *lov,
 335                             union lov_layout_state *state)
 336 {
 337         struct lov_layout_raid0 *r0 = &state->raid0;
 338         struct lov_stripe_md    *lsm = lov->lo_lsm;
 339         int i;
 340
 341         ENTRY;
 342
 343         dump_lsm(D_INODE, lsm);
 344
 345         lov_layout_wait(env, lov);
 346         if (r0->lo_sub != NULL) {
 347                 for (i = 0; i < r0->lo_nr; ++i) {
 348                         struct lovsub_object *los = r0->lo_sub[i];
 349
 350                         if (los != NULL) {
 351                                 cl_locks_prune(env, &los->lso_cl, 1);
 352                                 /*
 353                                  * If top-level object is to be evicted from
 354                                  * the cache, so are its sub-objects.
 355                                  */
 356                                 lov_subobject_kill(env, lov, los, i);
 357                         }
 358                 }
 359         }
 360         cl_object_prune(env, &lov->lo_cl);
 361         RETURN(0);
 362 }
 363
 364 static void lov_fini_empty(const struct lu_env *env, struct lov_object *lov,
 365                            union lov_layout_state *state)
 366 {
 367         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 368 }
 369
 370 static void lov_fini_raid0(const struct lu_env *env, struct lov_object *lov,
 371                            union lov_layout_state *state)
 372 {
 373         struct lov_layout_raid0 *r0 = &state->raid0;
 374         ENTRY;
 375
 376         if (r0->lo_sub != NULL) {
 377                 OBD_FREE_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 378                 r0->lo_sub = NULL;
 379         }
 380
 381         dump_lsm(D_INODE, lov->lo_lsm);
 382         lov_free_memmd(&lov->lo_lsm);
 383
 384         EXIT;
 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         ENTRY;
 391         dump_lsm(D_INODE, lov->lo_lsm);
 392         lov_free_memmd(&lov->lo_lsm);
 393         EXIT;
 394 }
 395
 396 static int lov_print_empty(const struct lu_env *env, void *cookie,
 397                            lu_printer_t p, const struct lu_object *o)
 398 {
 399         (*p)(env, cookie, "empty %d\n", lu2lov(o)->lo_layout_invalid);
 400         return 0;
 401 }
 402
 403 static int lov_print_raid0(const struct lu_env *env, void *cookie,
 404                            lu_printer_t p, const struct lu_object *o)
 405 {
 406         struct lov_object       *lov = lu2lov(o);
 407         struct lov_layout_raid0 *r0  = lov_r0(lov);
 408         struct lov_stripe_md    *lsm = lov->lo_lsm;
 409         int i;
 410
 411         (*p)(env, cookie, "stripes: %d, %svalid, lsm{%p 0x%08X %d %u %u}: \n",
 412                 r0->lo_nr, lov->lo_layout_invalid ? "in" : "", lsm,
 413                 lsm->lsm_magic, cfs_atomic_read(&lsm->lsm_refc),
 414                 lsm->lsm_stripe_count, lsm->lsm_layout_gen);
 415         for (i = 0; i < r0->lo_nr; ++i) {
 416                 struct lu_object *sub;
 417
 418                 if (r0->lo_sub[i] != NULL) {
 419                         sub = lovsub2lu(r0->lo_sub[i]);
 420                         lu_object_print(env, cookie, p, sub);
 421                 } else
 422                         (*p)(env, cookie, "sub %d absent\n", i);
 423         }
 424         return 0;
 425 }
 426
 427 static int lov_print_released(const struct lu_env *env, void *cookie,
 428                                 lu_printer_t p, const struct lu_object *o)
 429 {
 430         (*p)(env, cookie, "released\n");
 431         return 0;
 432 }
 433
 434 /**
 435  * Implements cl_object_operations::coo_attr_get() method for an object
 436  * without stripes (LLT_EMPTY layout type).
 437  *
 438  * The only attributes this layer is authoritative in this case is
 439  * cl_attr::cat_blocks---it's 0.
 440  */
 441 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
 442                               struct cl_attr *attr)
 443 {
 444         attr->cat_blocks = 0;
 445         return 0;
 446 }
 447
 448 static int lov_attr_get_raid0(const struct lu_env *env, struct cl_object *obj,
 449                               struct cl_attr *attr)
 450 {
 451         struct lov_object       *lov = cl2lov(obj);
 452         struct lov_layout_raid0 *r0 = lov_r0(lov);
 453         struct cl_attr          *lov_attr = &r0->lo_attr;
 454         int                      result = 0;
 455
 456         ENTRY;
 457
 458         /* this is called w/o holding type guard mutex, so it must be inside
 459          * an on going IO otherwise lsm may be replaced.
 460          * LU-2117: it turns out there exists one exception. For mmaped files,
 461          * the lock of those files may be requested in the other file's IO
 462          * context, and this function is called in ccc_lock_state(), it will
 463          * hit this assertion.
 464          * Anyway, it's still okay to call attr_get w/o type guard as layout
 465          * can't go if locks exist. */
 466         /* LASSERT(cfs_atomic_read(&lsm->lsm_refc) > 1); */
 467
 468         if (!r0->lo_attr_valid) {
 469                 struct lov_stripe_md    *lsm = lov->lo_lsm;
 470                 struct ost_lvb          *lvb = &lov_env_info(env)->lti_lvb;
 471                 __u64                    kms = 0;
 472
 473                 memset(lvb, 0, sizeof(*lvb));
 474                 /* XXX: timestamps can be negative by sanity:test_39m,
 475                  * how can it be? */
 476                 lvb->lvb_atime = LLONG_MIN;
 477                 lvb->lvb_ctime = LLONG_MIN;
 478                 lvb->lvb_mtime = LLONG_MIN;
 479
 480                 /*
 481                  * XXX that should be replaced with a loop over sub-objects,
 482                  * doing cl_object_attr_get() on them. But for now, let's
 483                  * reuse old lov code.
 484                  */
 485
 486                 /*
 487                  * XXX take lsm spin-lock to keep lov_merge_lvb_kms()
 488                  * happy. It's not needed, because new code uses
 489                  * ->coh_attr_guard spin-lock to protect consistency of
 490                  * sub-object attributes.
 491                  */
 492                 lov_stripe_lock(lsm);
 493                 result = lov_merge_lvb_kms(lsm, lvb, &kms);
 494                 lov_stripe_unlock(lsm);
 495                 if (result == 0) {
 496                         cl_lvb2attr(lov_attr, lvb);
 497                         lov_attr->cat_kms = kms;
 498                         r0->lo_attr_valid = 1;
 499                 }
 500         }
 501         if (result == 0) { /* merge results */
 502                 attr->cat_blocks = lov_attr->cat_blocks;
 503                 attr->cat_size = lov_attr->cat_size;
 504                 attr->cat_kms = lov_attr->cat_kms;
 505                 if (attr->cat_atime < lov_attr->cat_atime)
 506                         attr->cat_atime = lov_attr->cat_atime;
 507                 if (attr->cat_ctime < lov_attr->cat_ctime)
 508                         attr->cat_ctime = lov_attr->cat_ctime;
 509                 if (attr->cat_mtime < lov_attr->cat_mtime)
 510                         attr->cat_mtime = lov_attr->cat_mtime;
 511         }
 512         RETURN(result);
 513 }
 514
 515 const static struct lov_layout_operations lov_dispatch[] = {
 516         [LLT_EMPTY] = {
 517                 .llo_init      = lov_init_empty,
 518                 .llo_delete    = lov_delete_empty,
 519                 .llo_fini      = lov_fini_empty,
 520                 .llo_install   = lov_install_empty,
 521                 .llo_print     = lov_print_empty,
 522                 .llo_page_init = lov_page_init_empty,
 523                 .llo_lock_init = lov_lock_init_empty,
 524                 .llo_io_init   = lov_io_init_empty,
 525                 .llo_getattr   = lov_attr_get_empty
 526         },
 527         [LLT_RAID0] = {
 528                 .llo_init      = lov_init_raid0,
 529                 .llo_delete    = lov_delete_raid0,
 530                 .llo_fini      = lov_fini_raid0,
 531                 .llo_install   = lov_install_raid0,
 532                 .llo_print     = lov_print_raid0,
 533                 .llo_page_init = lov_page_init_raid0,
 534                 .llo_lock_init = lov_lock_init_raid0,
 535                 .llo_io_init   = lov_io_init_raid0,
 536                 .llo_getattr   = lov_attr_get_raid0
 537         },
 538         [LLT_RELEASED] = {
 539                 .llo_init      = lov_init_released,
 540                 .llo_delete    = lov_delete_empty,
 541                 .llo_fini      = lov_fini_released,
 542                 .llo_install   = lov_install_empty,
 543                 .llo_print     = lov_print_released,
 544                 .llo_page_init = lov_page_init_empty,
 545                 .llo_lock_init = lov_lock_init_empty,
 546                 .llo_io_init   = lov_io_init_released,
 547                 .llo_getattr   = lov_attr_get_empty
 548         }
 549 };
 550
 551 /**
 552  * Performs a double-dispatch based on the layout type of an object.
 553  */
 554 #define LOV_2DISPATCH_NOLOCK(obj, op, ...)                              \
 555 ({                                                                      \
 556         struct lov_object                      *__obj = (obj);          \
 557         enum lov_layout_type                    __llt;                  \
 558                                                                         \
 559         __llt = __obj->lo_type;                                         \
 560         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 561         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 562 })
 563
 564 /**
 565  * Return lov_layout_type associated with a given lsm
 566  */
 567 enum lov_layout_type lov_type(struct lov_stripe_md *lsm)
 568 {
 569         if (lsm == NULL)
 570                 return LLT_EMPTY;
 571         if (lsm_is_released(lsm))
 572                 return LLT_RELEASED;
 573         return LLT_RAID0;
 574 }
 575
 576 static inline void lov_conf_freeze(struct lov_object *lov)
 577 {
 578         if (lov->lo_owner != cfs_current())
 579                 down_read(&lov->lo_type_guard);
 580 }
 581
 582 static inline void lov_conf_thaw(struct lov_object *lov)
 583 {
 584         if (lov->lo_owner != cfs_current())
 585                 up_read(&lov->lo_type_guard);
 586 }
 587
 588 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...)                       \
 589 ({                                                                      \
 590         struct lov_object                      *__obj = (obj);          \
 591         int                                     __lock = !!(lock);      \
 592         typeof(lov_dispatch[0].op(__VA_ARGS__)) __result;               \
 593                                                                         \
 594         if (__lock)                                                     \
 595                 lov_conf_freeze(__obj);                                 \
 596         __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__);          \
 597         if (__lock)                                                     \
 598                 lov_conf_thaw(__obj);                                   \
 599         __result;                                                       \
 600 })
 601
 602 /**
 603  * Performs a locked double-dispatch based on the layout type of an object.
 604  */
 605 #define LOV_2DISPATCH(obj, op, ...)                     \
 606         LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
 607
 608 #define LOV_2DISPATCH_VOID(obj, op, ...)                                \
 609 do {                                                                    \
 610         struct lov_object                      *__obj = (obj);          \
 611         enum lov_layout_type                    __llt;                  \
 612                                                                         \
 613         lov_conf_freeze(__obj);                                         \
 614         __llt = __obj->lo_type;                                         \
 615         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 616         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 617         lov_conf_thaw(__obj);                                           \
 618 } while (0)
 619
 620 static void lov_conf_lock(struct lov_object *lov)
 621 {
 622         LASSERT(lov->lo_owner != cfs_current());
 623         down_write(&lov->lo_type_guard);
 624         LASSERT(lov->lo_owner == NULL);
 625         lov->lo_owner = cfs_current();
 626 }
 627
 628 static void lov_conf_unlock(struct lov_object *lov)
 629 {
 630         lov->lo_owner = NULL;
 631         up_write(&lov->lo_type_guard);
 632 }
 633
 634 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov)
 635 {
 636         struct l_wait_info lwi = { 0 };
 637         ENTRY;
 638
 639         while (cfs_atomic_read(&lov->lo_active_ios) > 0) {
 640                 CDEBUG(D_INODE, "file:"DFID" wait for active IO, now: %d.\n",
 641                         PFID(lu_object_fid(lov2lu(lov))),
 642                         cfs_atomic_read(&lov->lo_active_ios));
 643
 644                 l_wait_event(lov->lo_waitq,
 645                              cfs_atomic_read(&lov->lo_active_ios) == 0, &lwi);
 646         }
 647         RETURN(0);
 648 }
 649
 650 static int lov_layout_change(const struct lu_env *unused,
 651                              struct lov_object *lov,
 652                              const struct cl_object_conf *conf)
 653 {
 654         int result;
 655         enum lov_layout_type llt = LLT_EMPTY;
 656         union lov_layout_state *state = &lov->u;
 657         const struct lov_layout_operations *old_ops;
 658         const struct lov_layout_operations *new_ops;
 659
 660         struct cl_object_header *hdr = cl_object_header(&lov->lo_cl);
 661         void *cookie;
 662         struct lu_env *env;
 663         int refcheck;
 664         ENTRY;
 665
 666         LASSERT(0 <= lov->lo_type && lov->lo_type < ARRAY_SIZE(lov_dispatch));
 667
 668         if (conf->u.coc_md != NULL)
 669                 llt = lov_type(conf->u.coc_md->lsm);
 670         LASSERT(0 <= llt && llt < ARRAY_SIZE(lov_dispatch));
 671
 672         cookie = cl_env_reenter();
 673         env = cl_env_get(&refcheck);
 674         if (IS_ERR(env)) {
 675                 cl_env_reexit(cookie);
 676                 RETURN(PTR_ERR(env));
 677         }
 678
 679         old_ops = &lov_dispatch[lov->lo_type];
 680         new_ops = &lov_dispatch[llt];
 681
 682         result = old_ops->llo_delete(env, lov, &lov->u);
 683         if (result == 0) {
 684                 old_ops->llo_fini(env, lov, &lov->u);
 685
 686                 LASSERT(cfs_atomic_read(&lov->lo_active_ios) == 0);
 687                 LASSERT(hdr->coh_tree.rnode == NULL);
 688                 LASSERT(hdr->coh_pages == 0);
 689
 690                 lov->lo_type = LLT_EMPTY;
 691                 result = new_ops->llo_init(env,
 692                                         lu2lov_dev(lov->lo_cl.co_lu.lo_dev),
 693                                         lov, conf, state);
 694                 if (result == 0) {
 695                         new_ops->llo_install(env, lov, state);
 696                         lov->lo_type = llt;
 697                 } else {
 698                         new_ops->llo_delete(env, lov, state);
 699                         new_ops->llo_fini(env, lov, state);
 700                         /* this file becomes an EMPTY file. */
 701                 }
 702         }
 703
 704         cl_env_put(env, &refcheck);
 705         cl_env_reexit(cookie);
 706         RETURN(result);
 707 }
 708
 709 /*****************************************************************************
 710  *
 711  * Lov object operations.
 712  *
 713  */
 714 int lov_object_init(const struct lu_env *env, struct lu_object *obj,
 715                     const struct lu_object_conf *conf)
 716 {
 717         struct lov_device            *dev   = lu2lov_dev(obj->lo_dev);
 718         struct lov_object            *lov   = lu2lov(obj);
 719         const struct cl_object_conf  *cconf = lu2cl_conf(conf);
 720         union  lov_layout_state      *set   = &lov->u;
 721         const struct lov_layout_operations *ops;
 722         int result;
 723
 724         ENTRY;
 725         init_rwsem(&lov->lo_type_guard);
 726         cfs_atomic_set(&lov->lo_active_ios, 0);
 727         cfs_waitq_init(&lov->lo_waitq);
 728
 729         cl_object_page_init(lu2cl(obj), sizeof(struct lov_page));
 730
 731         /* no locking is necessary, as object is being created */
 732         lov->lo_type = lov_type(cconf->u.coc_md->lsm);
 733         ops = &lov_dispatch[lov->lo_type];
 734         result = ops->llo_init(env, dev, lov, cconf, set);
 735         if (result == 0)
 736                 ops->llo_install(env, lov, set);
 737         RETURN(result);
 738 }
 739
 740 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
 741                         const struct cl_object_conf *conf)
 742 {
 743         struct lov_stripe_md    *lsm = NULL;
 744         struct lov_object       *lov = cl2lov(obj);
 745         int                      result = 0;
 746         ENTRY;
 747
 748         lov_conf_lock(lov);
 749         if (conf->coc_opc == OBJECT_CONF_INVALIDATE) {
 750                 lov->lo_layout_invalid = true;
 751                 GOTO(out, result = 0);
 752         }
 753
 754         if (conf->coc_opc == OBJECT_CONF_WAIT) {
 755                 if (lov->lo_layout_invalid &&
 756                     cfs_atomic_read(&lov->lo_active_ios) > 0) {
 757                         lov_conf_unlock(lov);
 758                         result = lov_layout_wait(env, lov);
 759                         lov_conf_lock(lov);
 760                 }
 761                 GOTO(out, result);
 762         }
 763
 764         LASSERT(conf->coc_opc == OBJECT_CONF_SET);
 765
 766         if (conf->u.coc_md != NULL)
 767                 lsm = conf->u.coc_md->lsm;
 768         if ((lsm == NULL && lov->lo_lsm == NULL) ||
 769             (lsm != NULL && lov->lo_lsm != NULL &&
 770              lov->lo_lsm->lsm_layout_gen == lsm->lsm_layout_gen)) {
 771                 /* same version of layout */
 772                 lov->lo_layout_invalid = false;
 773                 GOTO(out, result = 0);
 774         }
 775
 776         /* will change layout - check if there still exists active IO. */
 777         if (cfs_atomic_read(&lov->lo_active_ios) > 0) {
 778                 lov->lo_layout_invalid = true;
 779                 GOTO(out, result = -EBUSY);
 780         }
 781
 782         lov->lo_layout_invalid = lov_layout_change(env, lov, conf);
 783         EXIT;
 784
 785 out:
 786         lov_conf_unlock(lov);
 787         RETURN(result);
 788 }
 789
 790 static void lov_object_delete(const struct lu_env *env, struct lu_object *obj)
 791 {
 792         struct lov_object *lov = lu2lov(obj);
 793
 794         ENTRY;
 795         LOV_2DISPATCH_VOID(lov, llo_delete, env, lov, &lov->u);
 796         EXIT;
 797 }
 798
 799 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
 800 {
 801         struct lov_object *lov = lu2lov(obj);
 802
 803         ENTRY;
 804         LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
 805         lu_object_fini(obj);
 806         OBD_SLAB_FREE_PTR(lov, lov_object_kmem);
 807         EXIT;
 808 }
 809
 810 static int lov_object_print(const struct lu_env *env, void *cookie,
 811                             lu_printer_t p, const struct lu_object *o)
 812 {
 813         return LOV_2DISPATCH_NOLOCK(lu2lov(o), llo_print, env, cookie, p, o);
 814 }
 815
 816 int lov_page_init(const struct lu_env *env, struct cl_object *obj,
 817                 struct cl_page *page, cfs_page_t *vmpage)
 818 {
 819         return LOV_2DISPATCH_NOLOCK(cl2lov(obj),
 820                                     llo_page_init, env, obj, page, vmpage);
 821 }
 822
 823 /**
 824  * Implements cl_object_operations::clo_io_init() method for lov
 825  * layer. Dispatches to the appropriate layout io initialization method.
 826  */
 827 int lov_io_init(const struct lu_env *env, struct cl_object *obj,
 828                 struct cl_io *io)
 829 {
 830         CL_IO_SLICE_CLEAN(lov_env_io(env), lis_cl);
 831         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
 832                                      !io->ci_ignore_layout, env, obj, io);
 833 }
 834
 835 /**
 836  * An implementation of cl_object_operations::clo_attr_get() method for lov
 837  * layer. For raid0 layout this collects and merges attributes of all
 838  * sub-objects.
 839  */
 840 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
 841                         struct cl_attr *attr)
 842 {
 843         /* do not take lock, as this function is called under a
 844          * spin-lock. Layout is protected from changing by ongoing IO. */
 845         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
 846 }
 847
 848 static int lov_attr_set(const struct lu_env *env, struct cl_object *obj,
 849                         const struct cl_attr *attr, unsigned valid)
 850 {
 851         /*
 852          * No dispatch is required here, as no layout implements this.
 853          */
 854         return 0;
 855 }
 856
 857 int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
 858                   struct cl_lock *lock, const struct cl_io *io)
 859 {
 860         /* No need to lock because we've taken one refcount of layout.  */
 861         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_lock_init, env, obj, lock,
 862                                     io);
 863 }
 864
 865 static const struct cl_object_operations lov_ops = {
 866         .coo_page_init = lov_page_init,
 867         .coo_lock_init = lov_lock_init,
 868         .coo_io_init   = lov_io_init,
 869         .coo_attr_get  = lov_attr_get,
 870         .coo_attr_set  = lov_attr_set,
 871         .coo_conf_set  = lov_conf_set
 872 };
 873
 874 static const struct lu_object_operations lov_lu_obj_ops = {
 875         .loo_object_init      = lov_object_init,
 876         .loo_object_delete    = lov_object_delete,
 877         .loo_object_release   = NULL,
 878         .loo_object_free      = lov_object_free,
 879         .loo_object_print     = lov_object_print,
 880         .loo_object_invariant = NULL
 881 };
 882
 883 struct lu_object *lov_object_alloc(const struct lu_env *env,
 884                                    const struct lu_object_header *unused,
 885                                    struct lu_device *dev)
 886 {
 887         struct lov_object *lov;
 888         struct lu_object  *obj;
 889
 890         ENTRY;
 891         OBD_SLAB_ALLOC_PTR_GFP(lov, lov_object_kmem, CFS_ALLOC_IO);
 892         if (lov != NULL) {
 893                 obj = lov2lu(lov);
 894                 lu_object_init(obj, NULL, dev);
 895                 lov->lo_cl.co_ops = &lov_ops;
 896                 lov->lo_type = -1; /* invalid, to catch uninitialized type */
 897                 /*
 898                  * object io operation vector (cl_object::co_iop) is installed
 899                  * later in lov_object_init(), as different vectors are used
 900                  * for object with different layouts.
 901                  */
 902                 obj->lo_ops = &lov_lu_obj_ops;
 903         } else
 904                 obj = NULL;
 905         RETURN(obj);
 906 }
 907
 908 struct lov_stripe_md *lov_lsm_addref(struct lov_object *lov)
 909 {
 910         struct lov_stripe_md *lsm = NULL;
 911
 912         lov_conf_freeze(lov);
 913         if (lov->lo_lsm != NULL) {
 914                 lsm = lsm_addref(lov->lo_lsm);
 915                 CDEBUG(D_INODE, "lsm %p addref %d/%d by %p.\n",
 916                         lsm, cfs_atomic_read(&lsm->lsm_refc),
 917                         lov->lo_layout_invalid, cfs_current());
 918         }
 919         lov_conf_thaw(lov);
 920         return lsm;
 921 }
 922
 923 void lov_lsm_decref(struct lov_object *lov, struct lov_stripe_md *lsm)
 924 {
 925         if (lsm == NULL)
 926                 return;
 927
 928         CDEBUG(D_INODE, "lsm %p decref %d by %p.\n",
 929                 lsm, cfs_atomic_read(&lsm->lsm_refc), cfs_current());
 930
 931         lov_free_memmd(&lsm);
 932 }
 933
 934 struct lov_stripe_md *lov_lsm_get(struct cl_object *clobj)
 935 {
 936         struct lu_object *luobj;
 937         struct lov_stripe_md *lsm = NULL;
 938
 939         if (clobj == NULL)
 940                 return NULL;
 941
 942         luobj = lu_object_locate(&cl_object_header(clobj)->coh_lu,
 943                                  &lov_device_type);
 944         if (luobj != NULL)
 945                 lsm = lov_lsm_addref(lu2lov(luobj));
 946         return lsm;
 947 }
 948 EXPORT_SYMBOL(lov_lsm_get);
 949
 950 void lov_lsm_put(struct cl_object *unused, struct lov_stripe_md *lsm)
 951 {
 952         if (lsm != NULL)
 953                 lov_free_memmd(&lsm);
 954 }
 955 EXPORT_SYMBOL(lov_lsm_put);
 956
 957 int lov_read_and_clear_async_rc(struct cl_object *clob)
 958 {
 959         struct lu_object *luobj;
 960         int rc = 0;
 961         ENTRY;
 962
 963         luobj = lu_object_locate(&cl_object_header(clob)->coh_lu,
 964                                  &lov_device_type);
 965         if (luobj != NULL) {
 966                 struct lov_object *lov = lu2lov(luobj);
 967
 968                 lov_conf_freeze(lov);
 969                 switch (lov->lo_type) {
 970                 case LLT_RAID0: {
 971                         struct lov_stripe_md *lsm;
 972                         int i;
 973
 974                         lsm = lov->lo_lsm;
 975                         LASSERT(lsm != NULL);
 976                         for (i = 0; i < lsm->lsm_stripe_count; i++) {
 977                                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
 978                                 if (loi->loi_ar.ar_rc && !rc)
 979                                         rc = loi->loi_ar.ar_rc;
 980                                 loi->loi_ar.ar_rc = 0;
 981                         }
 982                 }
 983                 case LLT_RELEASED:
 984                 case LLT_EMPTY:
 985                         break;
 986                 default:
 987                         LBUG();
 988                 }
 989                 lov_conf_thaw(lov);
 990         }
 991         RETURN(rc);
 992 }
 993 EXPORT_SYMBOL(lov_read_and_clear_async_rc);
 994
 995 /** @} lov */