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, 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         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, %s, lsm{%p 0x%08X %d %u %u}:\n",
 412                 r0->lo_nr, lov->lo_layout_invalid ? "invalid" : "valid", 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         }
 425         return 0;
 426 }
 427
 428 static int lov_print_released(const struct lu_env *env, void *cookie,
 429                                 lu_printer_t p, const struct lu_object *o)
 430 {
 431         struct lov_object       *lov = lu2lov(o);
 432         struct lov_stripe_md    *lsm = lov->lo_lsm;
 433
 434         (*p)(env, cookie,
 435                 "released: %s, lsm{%p 0x%08X %d %u %u}:\n",
 436                 lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 437                 lsm->lsm_magic, cfs_atomic_read(&lsm->lsm_refc),
 438                 lsm->lsm_stripe_count, lsm->lsm_layout_gen);
 439         return 0;
 440 }
 441
 442 /**
 443  * Implements cl_object_operations::coo_attr_get() method for an object
 444  * without stripes (LLT_EMPTY layout type).
 445  *
 446  * The only attributes this layer is authoritative in this case is
 447  * cl_attr::cat_blocks---it's 0.
 448  */
 449 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
 450                               struct cl_attr *attr)
 451 {
 452         attr->cat_blocks = 0;
 453         return 0;
 454 }
 455
 456 static int lov_attr_get_raid0(const struct lu_env *env, struct cl_object *obj,
 457                               struct cl_attr *attr)
 458 {
 459         struct lov_object       *lov = cl2lov(obj);
 460         struct lov_layout_raid0 *r0 = lov_r0(lov);
 461         struct cl_attr          *lov_attr = &r0->lo_attr;
 462         int                      result = 0;
 463
 464         ENTRY;
 465
 466         /* this is called w/o holding type guard mutex, so it must be inside
 467          * an on going IO otherwise lsm may be replaced.
 468          * LU-2117: it turns out there exists one exception. For mmaped files,
 469          * the lock of those files may be requested in the other file's IO
 470          * context, and this function is called in ccc_lock_state(), it will
 471          * hit this assertion.
 472          * Anyway, it's still okay to call attr_get w/o type guard as layout
 473          * can't go if locks exist. */
 474         /* LASSERT(cfs_atomic_read(&lsm->lsm_refc) > 1); */
 475
 476         if (!r0->lo_attr_valid) {
 477                 struct lov_stripe_md    *lsm = lov->lo_lsm;
 478                 struct ost_lvb          *lvb = &lov_env_info(env)->lti_lvb;
 479                 __u64                    kms = 0;
 480
 481                 memset(lvb, 0, sizeof(*lvb));
 482                 /* XXX: timestamps can be negative by sanity:test_39m,
 483                  * how can it be? */
 484                 lvb->lvb_atime = LLONG_MIN;
 485                 lvb->lvb_ctime = LLONG_MIN;
 486                 lvb->lvb_mtime = LLONG_MIN;
 487
 488                 /*
 489                  * XXX that should be replaced with a loop over sub-objects,
 490                  * doing cl_object_attr_get() on them. But for now, let's
 491                  * reuse old lov code.
 492                  */
 493
 494                 /*
 495                  * XXX take lsm spin-lock to keep lov_merge_lvb_kms()
 496                  * happy. It's not needed, because new code uses
 497                  * ->coh_attr_guard spin-lock to protect consistency of
 498                  * sub-object attributes.
 499                  */
 500                 lov_stripe_lock(lsm);
 501                 result = lov_merge_lvb_kms(lsm, lvb, &kms);
 502                 lov_stripe_unlock(lsm);
 503                 if (result == 0) {
 504                         cl_lvb2attr(lov_attr, lvb);
 505                         lov_attr->cat_kms = kms;
 506                         r0->lo_attr_valid = 1;
 507                 }
 508         }
 509         if (result == 0) { /* merge results */
 510                 attr->cat_blocks = lov_attr->cat_blocks;
 511                 attr->cat_size = lov_attr->cat_size;
 512                 attr->cat_kms = lov_attr->cat_kms;
 513                 if (attr->cat_atime < lov_attr->cat_atime)
 514                         attr->cat_atime = lov_attr->cat_atime;
 515                 if (attr->cat_ctime < lov_attr->cat_ctime)
 516                         attr->cat_ctime = lov_attr->cat_ctime;
 517                 if (attr->cat_mtime < lov_attr->cat_mtime)
 518                         attr->cat_mtime = lov_attr->cat_mtime;
 519         }
 520         RETURN(result);
 521 }
 522
 523 const static struct lov_layout_operations lov_dispatch[] = {
 524         [LLT_EMPTY] = {
 525                 .llo_init      = lov_init_empty,
 526                 .llo_delete    = lov_delete_empty,
 527                 .llo_fini      = lov_fini_empty,
 528                 .llo_install   = lov_install_empty,
 529                 .llo_print     = lov_print_empty,
 530                 .llo_page_init = lov_page_init_empty,
 531                 .llo_lock_init = lov_lock_init_empty,
 532                 .llo_io_init   = lov_io_init_empty,
 533                 .llo_getattr   = lov_attr_get_empty
 534         },
 535         [LLT_RAID0] = {
 536                 .llo_init      = lov_init_raid0,
 537                 .llo_delete    = lov_delete_raid0,
 538                 .llo_fini      = lov_fini_raid0,
 539                 .llo_install   = lov_install_raid0,
 540                 .llo_print     = lov_print_raid0,
 541                 .llo_page_init = lov_page_init_raid0,
 542                 .llo_lock_init = lov_lock_init_raid0,
 543                 .llo_io_init   = lov_io_init_raid0,
 544                 .llo_getattr   = lov_attr_get_raid0
 545         },
 546         [LLT_RELEASED] = {
 547                 .llo_init      = lov_init_released,
 548                 .llo_delete    = lov_delete_empty,
 549                 .llo_fini      = lov_fini_released,
 550                 .llo_install   = lov_install_empty,
 551                 .llo_print     = lov_print_released,
 552                 .llo_page_init = lov_page_init_empty,
 553                 .llo_lock_init = lov_lock_init_empty,
 554                 .llo_io_init   = lov_io_init_released,
 555                 .llo_getattr   = lov_attr_get_empty
 556         }
 557 };
 558
 559 /**
 560  * Performs a double-dispatch based on the layout type of an object.
 561  */
 562 #define LOV_2DISPATCH_NOLOCK(obj, op, ...)                              \
 563 ({                                                                      \
 564         struct lov_object                      *__obj = (obj);          \
 565         enum lov_layout_type                    __llt;                  \
 566                                                                         \
 567         __llt = __obj->lo_type;                                         \
 568         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 569         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 570 })
 571
 572 /**
 573  * Return lov_layout_type associated with a given lsm
 574  */
 575 enum lov_layout_type lov_type(struct lov_stripe_md *lsm)
 576 {
 577         if (lsm == NULL)
 578                 return LLT_EMPTY;
 579         if (lsm_is_released(lsm))
 580                 return LLT_RELEASED;
 581         return LLT_RAID0;
 582 }
 583
 584 static inline void lov_conf_freeze(struct lov_object *lov)
 585 {
 586         if (lov->lo_owner != cfs_current())
 587                 down_read(&lov->lo_type_guard);
 588 }
 589
 590 static inline void lov_conf_thaw(struct lov_object *lov)
 591 {
 592         if (lov->lo_owner != cfs_current())
 593                 up_read(&lov->lo_type_guard);
 594 }
 595
 596 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...)                       \
 597 ({                                                                      \
 598         struct lov_object                      *__obj = (obj);          \
 599         int                                     __lock = !!(lock);      \
 600         typeof(lov_dispatch[0].op(__VA_ARGS__)) __result;               \
 601                                                                         \
 602         if (__lock)                                                     \
 603                 lov_conf_freeze(__obj);                                 \
 604         __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__);          \
 605         if (__lock)                                                     \
 606                 lov_conf_thaw(__obj);                                   \
 607         __result;                                                       \
 608 })
 609
 610 /**
 611  * Performs a locked double-dispatch based on the layout type of an object.
 612  */
 613 #define LOV_2DISPATCH(obj, op, ...)                     \
 614         LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
 615
 616 #define LOV_2DISPATCH_VOID(obj, op, ...)                                \
 617 do {                                                                    \
 618         struct lov_object                      *__obj = (obj);          \
 619         enum lov_layout_type                    __llt;                  \
 620                                                                         \
 621         lov_conf_freeze(__obj);                                         \
 622         __llt = __obj->lo_type;                                         \
 623         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 624         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 625         lov_conf_thaw(__obj);                                           \
 626 } while (0)
 627
 628 static void lov_conf_lock(struct lov_object *lov)
 629 {
 630         LASSERT(lov->lo_owner != cfs_current());
 631         down_write(&lov->lo_type_guard);
 632         LASSERT(lov->lo_owner == NULL);
 633         lov->lo_owner = cfs_current();
 634 }
 635
 636 static void lov_conf_unlock(struct lov_object *lov)
 637 {
 638         lov->lo_owner = NULL;
 639         up_write(&lov->lo_type_guard);
 640 }
 641
 642 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov)
 643 {
 644         struct l_wait_info lwi = { 0 };
 645         ENTRY;
 646
 647         while (cfs_atomic_read(&lov->lo_active_ios) > 0) {
 648                 CDEBUG(D_INODE, "file:"DFID" wait for active IO, now: %d.\n",
 649                         PFID(lu_object_fid(lov2lu(lov))),
 650                         cfs_atomic_read(&lov->lo_active_ios));
 651
 652                 l_wait_event(lov->lo_waitq,
 653                              cfs_atomic_read(&lov->lo_active_ios) == 0, &lwi);
 654         }
 655         RETURN(0);
 656 }
 657
 658 static int lov_layout_change(const struct lu_env *unused,
 659                              struct lov_object *lov,
 660                              const struct cl_object_conf *conf)
 661 {
 662         int result;
 663         enum lov_layout_type llt = LLT_EMPTY;
 664         union lov_layout_state *state = &lov->u;
 665         const struct lov_layout_operations *old_ops;
 666         const struct lov_layout_operations *new_ops;
 667
 668         struct cl_object_header *hdr = cl_object_header(&lov->lo_cl);
 669         void *cookie;
 670         struct lu_env *env;
 671         int refcheck;
 672         ENTRY;
 673
 674         LASSERT(0 <= lov->lo_type && lov->lo_type < ARRAY_SIZE(lov_dispatch));
 675
 676         if (conf->u.coc_md != NULL)
 677                 llt = lov_type(conf->u.coc_md->lsm);
 678         LASSERT(0 <= llt && llt < ARRAY_SIZE(lov_dispatch));
 679
 680         cookie = cl_env_reenter();
 681         env = cl_env_get(&refcheck);
 682         if (IS_ERR(env)) {
 683                 cl_env_reexit(cookie);
 684                 RETURN(PTR_ERR(env));
 685         }
 686
 687         CDEBUG(D_INODE, DFID" from %s to %s\n",
 688                PFID(lu_object_fid(lov2lu(lov))),
 689                llt2str(lov->lo_type), llt2str(llt));
 690
 691         old_ops = &lov_dispatch[lov->lo_type];
 692         new_ops = &lov_dispatch[llt];
 693
 694         result = old_ops->llo_delete(env, lov, &lov->u);
 695         if (result == 0) {
 696                 old_ops->llo_fini(env, lov, &lov->u);
 697
 698                 LASSERT(cfs_atomic_read(&lov->lo_active_ios) == 0);
 699                 LASSERT(hdr->coh_tree.rnode == NULL);
 700                 LASSERT(hdr->coh_pages == 0);
 701
 702                 lov->lo_type = LLT_EMPTY;
 703                 result = new_ops->llo_init(env,
 704                                         lu2lov_dev(lov->lo_cl.co_lu.lo_dev),
 705                                         lov, conf, state);
 706                 if (result == 0) {
 707                         new_ops->llo_install(env, lov, state);
 708                         lov->lo_type = llt;
 709                 } else {
 710                         new_ops->llo_delete(env, lov, state);
 711                         new_ops->llo_fini(env, lov, state);
 712                         /* this file becomes an EMPTY file. */
 713                 }
 714         }
 715
 716         cl_env_put(env, &refcheck);
 717         cl_env_reexit(cookie);
 718         RETURN(result);
 719 }
 720
 721 /*****************************************************************************
 722  *
 723  * Lov object operations.
 724  *
 725  */
 726 int lov_object_init(const struct lu_env *env, struct lu_object *obj,
 727                     const struct lu_object_conf *conf)
 728 {
 729         struct lov_device            *dev   = lu2lov_dev(obj->lo_dev);
 730         struct lov_object            *lov   = lu2lov(obj);
 731         const struct cl_object_conf  *cconf = lu2cl_conf(conf);
 732         union  lov_layout_state      *set   = &lov->u;
 733         const struct lov_layout_operations *ops;
 734         int result;
 735
 736         ENTRY;
 737         init_rwsem(&lov->lo_type_guard);
 738         cfs_atomic_set(&lov->lo_active_ios, 0);
 739         cfs_waitq_init(&lov->lo_waitq);
 740
 741         cl_object_page_init(lu2cl(obj), sizeof(struct lov_page));
 742
 743         /* no locking is necessary, as object is being created */
 744         lov->lo_type = lov_type(cconf->u.coc_md->lsm);
 745         ops = &lov_dispatch[lov->lo_type];
 746         result = ops->llo_init(env, dev, lov, cconf, set);
 747         if (result == 0)
 748                 ops->llo_install(env, lov, set);
 749         RETURN(result);
 750 }
 751
 752 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
 753                         const struct cl_object_conf *conf)
 754 {
 755         struct lov_stripe_md    *lsm = NULL;
 756         struct lov_object       *lov = cl2lov(obj);
 757         int                      result = 0;
 758         ENTRY;
 759
 760         lov_conf_lock(lov);
 761         if (conf->coc_opc == OBJECT_CONF_INVALIDATE) {
 762                 lov->lo_layout_invalid = true;
 763                 GOTO(out, result = 0);
 764         }
 765
 766         if (conf->coc_opc == OBJECT_CONF_WAIT) {
 767                 if (lov->lo_layout_invalid &&
 768                     cfs_atomic_read(&lov->lo_active_ios) > 0) {
 769                         lov_conf_unlock(lov);
 770                         result = lov_layout_wait(env, lov);
 771                         lov_conf_lock(lov);
 772                 }
 773                 GOTO(out, result);
 774         }
 775
 776         LASSERT(conf->coc_opc == OBJECT_CONF_SET);
 777
 778         if (conf->u.coc_md != NULL)
 779                 lsm = conf->u.coc_md->lsm;
 780         if ((lsm == NULL && lov->lo_lsm == NULL) ||
 781             ((lsm != NULL && lov->lo_lsm != NULL) &&
 782              (lov->lo_lsm->lsm_layout_gen == lsm->lsm_layout_gen) &&
 783              (lov->lo_lsm->lsm_pattern == lsm->lsm_pattern))) {
 784                 /* same version of layout */
 785                 lov->lo_layout_invalid = false;
 786                 GOTO(out, result = 0);
 787         }
 788
 789         /* will change layout - check if there still exists active IO. */
 790         if (cfs_atomic_read(&lov->lo_active_ios) > 0) {
 791                 lov->lo_layout_invalid = true;
 792                 GOTO(out, result = -EBUSY);
 793         }
 794
 795         lov->lo_layout_invalid = lov_layout_change(env, lov, conf);
 796         EXIT;
 797
 798 out:
 799         lov_conf_unlock(lov);
 800         CDEBUG(D_INODE, DFID" lo_layout_invalid=%d\n",
 801                PFID(lu_object_fid(lov2lu(lov))), lov->lo_layout_invalid);
 802         RETURN(result);
 803 }
 804
 805 static void lov_object_delete(const struct lu_env *env, struct lu_object *obj)
 806 {
 807         struct lov_object *lov = lu2lov(obj);
 808
 809         ENTRY;
 810         LOV_2DISPATCH_VOID(lov, llo_delete, env, lov, &lov->u);
 811         EXIT;
 812 }
 813
 814 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
 815 {
 816         struct lov_object *lov = lu2lov(obj);
 817
 818         ENTRY;
 819         LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
 820         lu_object_fini(obj);
 821         OBD_SLAB_FREE_PTR(lov, lov_object_kmem);
 822         EXIT;
 823 }
 824
 825 static int lov_object_print(const struct lu_env *env, void *cookie,
 826                             lu_printer_t p, const struct lu_object *o)
 827 {
 828         return LOV_2DISPATCH_NOLOCK(lu2lov(o), llo_print, env, cookie, p, o);
 829 }
 830
 831 int lov_page_init(const struct lu_env *env, struct cl_object *obj,
 832                   struct cl_page *page, struct page *vmpage)
 833 {
 834         return LOV_2DISPATCH_NOLOCK(cl2lov(obj),
 835                                     llo_page_init, env, obj, page, vmpage);
 836 }
 837
 838 /**
 839  * Implements cl_object_operations::clo_io_init() method for lov
 840  * layer. Dispatches to the appropriate layout io initialization method.
 841  */
 842 int lov_io_init(const struct lu_env *env, struct cl_object *obj,
 843                 struct cl_io *io)
 844 {
 845         CL_IO_SLICE_CLEAN(lov_env_io(env), lis_cl);
 846         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
 847                                      !io->ci_ignore_layout, env, obj, io);
 848 }
 849
 850 /**
 851  * An implementation of cl_object_operations::clo_attr_get() method for lov
 852  * layer. For raid0 layout this collects and merges attributes of all
 853  * sub-objects.
 854  */
 855 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
 856                         struct cl_attr *attr)
 857 {
 858         /* do not take lock, as this function is called under a
 859          * spin-lock. Layout is protected from changing by ongoing IO. */
 860         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
 861 }
 862
 863 static int lov_attr_set(const struct lu_env *env, struct cl_object *obj,
 864                         const struct cl_attr *attr, unsigned valid)
 865 {
 866         /*
 867          * No dispatch is required here, as no layout implements this.
 868          */
 869         return 0;
 870 }
 871
 872 int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
 873                   struct cl_lock *lock, const struct cl_io *io)
 874 {
 875         /* No need to lock because we've taken one refcount of layout.  */
 876         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_lock_init, env, obj, lock,
 877                                     io);
 878 }
 879
 880 static const struct cl_object_operations lov_ops = {
 881         .coo_page_init = lov_page_init,
 882         .coo_lock_init = lov_lock_init,
 883         .coo_io_init   = lov_io_init,
 884         .coo_attr_get  = lov_attr_get,
 885         .coo_attr_set  = lov_attr_set,
 886         .coo_conf_set  = lov_conf_set
 887 };
 888
 889 static const struct lu_object_operations lov_lu_obj_ops = {
 890         .loo_object_init      = lov_object_init,
 891         .loo_object_delete    = lov_object_delete,
 892         .loo_object_release   = NULL,
 893         .loo_object_free      = lov_object_free,
 894         .loo_object_print     = lov_object_print,
 895         .loo_object_invariant = NULL
 896 };
 897
 898 struct lu_object *lov_object_alloc(const struct lu_env *env,
 899                                    const struct lu_object_header *unused,
 900                                    struct lu_device *dev)
 901 {
 902         struct lov_object *lov;
 903         struct lu_object  *obj;
 904
 905         ENTRY;
 906         OBD_SLAB_ALLOC_PTR_GFP(lov, lov_object_kmem, __GFP_IO);
 907         if (lov != NULL) {
 908                 obj = lov2lu(lov);
 909                 lu_object_init(obj, NULL, dev);
 910                 lov->lo_cl.co_ops = &lov_ops;
 911                 lov->lo_type = -1; /* invalid, to catch uninitialized type */
 912                 /*
 913                  * object io operation vector (cl_object::co_iop) is installed
 914                  * later in lov_object_init(), as different vectors are used
 915                  * for object with different layouts.
 916                  */
 917                 obj->lo_ops = &lov_lu_obj_ops;
 918         } else
 919                 obj = NULL;
 920         RETURN(obj);
 921 }
 922
 923 struct lov_stripe_md *lov_lsm_addref(struct lov_object *lov)
 924 {
 925         struct lov_stripe_md *lsm = NULL;
 926
 927         lov_conf_freeze(lov);
 928         if (lov->lo_lsm != NULL) {
 929                 lsm = lsm_addref(lov->lo_lsm);
 930                 CDEBUG(D_INODE, "lsm %p addref %d/%d by %p.\n",
 931                         lsm, cfs_atomic_read(&lsm->lsm_refc),
 932                         lov->lo_layout_invalid, cfs_current());
 933         }
 934         lov_conf_thaw(lov);
 935         return lsm;
 936 }
 937
 938 void lov_lsm_decref(struct lov_object *lov, struct lov_stripe_md *lsm)
 939 {
 940         if (lsm == NULL)
 941                 return;
 942
 943         CDEBUG(D_INODE, "lsm %p decref %d by %p.\n",
 944                 lsm, cfs_atomic_read(&lsm->lsm_refc), cfs_current());
 945
 946         lov_free_memmd(&lsm);
 947 }
 948
 949 struct lov_stripe_md *lov_lsm_get(struct cl_object *clobj)
 950 {
 951         struct lu_object *luobj;
 952         struct lov_stripe_md *lsm = NULL;
 953
 954         if (clobj == NULL)
 955                 return NULL;
 956
 957         luobj = lu_object_locate(&cl_object_header(clobj)->coh_lu,
 958                                  &lov_device_type);
 959         if (luobj != NULL)
 960                 lsm = lov_lsm_addref(lu2lov(luobj));
 961         return lsm;
 962 }
 963 EXPORT_SYMBOL(lov_lsm_get);
 964
 965 void lov_lsm_put(struct cl_object *unused, struct lov_stripe_md *lsm)
 966 {
 967         if (lsm != NULL)
 968                 lov_free_memmd(&lsm);
 969 }
 970 EXPORT_SYMBOL(lov_lsm_put);
 971
 972 int lov_read_and_clear_async_rc(struct cl_object *clob)
 973 {
 974         struct lu_object *luobj;
 975         int rc = 0;
 976         ENTRY;
 977
 978         luobj = lu_object_locate(&cl_object_header(clob)->coh_lu,
 979                                  &lov_device_type);
 980         if (luobj != NULL) {
 981                 struct lov_object *lov = lu2lov(luobj);
 982
 983                 lov_conf_freeze(lov);
 984                 switch (lov->lo_type) {
 985                 case LLT_RAID0: {
 986                         struct lov_stripe_md *lsm;
 987                         int i;
 988
 989                         lsm = lov->lo_lsm;
 990                         LASSERT(lsm != NULL);
 991                         for (i = 0; i < lsm->lsm_stripe_count; i++) {
 992                                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
 993                                 if (loi->loi_ar.ar_rc && !rc)
 994                                         rc = loi->loi_ar.ar_rc;
 995                                 loi->loi_ar.ar_rc = 0;
 996                         }
 997                 }
 998                 case LLT_RELEASED:
 999                 case LLT_EMPTY:
1000                         break;
1001                 default:
1002                         LBUG();
1003                 }
1004                 lov_conf_thaw(lov);
1005         }
1006         RETURN(rc);
1007 }
1008 EXPORT_SYMBOL(lov_read_and_clear_async_rc);
1009
1010 /** @} lov */