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, 2015, 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
  46 /** \addtogroup lov
  47  *  @{
  48  */
  49
  50 /*****************************************************************************
  51  *
  52  * Layout operations.
  53  *
  54  */
  55
  56 struct lov_layout_operations {
  57         int (*llo_init)(const struct lu_env *env, struct lov_device *dev,
  58                         struct lov_object *lov,
  59                         const struct cl_object_conf *conf,
  60                         union lov_layout_state *state);
  61         int (*llo_delete)(const struct lu_env *env, struct lov_object *lov,
  62                           union lov_layout_state *state);
  63         void (*llo_fini)(const struct lu_env *env, struct lov_object *lov,
  64                          union lov_layout_state *state);
  65         void (*llo_install)(const struct lu_env *env, struct lov_object *lov,
  66                             union lov_layout_state *state);
  67         int  (*llo_print)(const struct lu_env *env, void *cookie,
  68                           lu_printer_t p, const struct lu_object *o);
  69         int  (*llo_page_init)(const struct lu_env *env, struct cl_object *obj,
  70                               struct cl_page *page, pgoff_t index);
  71         int  (*llo_lock_init)(const struct lu_env *env,
  72                               struct cl_object *obj, struct cl_lock *lock,
  73                               const struct cl_io *io);
  74         int  (*llo_io_init)(const struct lu_env *env,
  75                             struct cl_object *obj, struct cl_io *io);
  76         int  (*llo_getattr)(const struct lu_env *env, struct cl_object *obj,
  77                             struct cl_attr *attr);
  78 };
  79
  80 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov);
  81
  82 /*****************************************************************************
  83  *
  84  * Lov object layout operations.
  85  *
  86  */
  87
  88 static void lov_install_empty(const struct lu_env *env,
  89                               struct lov_object *lov,
  90                               union  lov_layout_state *state)
  91 {
  92         /*
  93          * File without objects.
  94          */
  95 }
  96
  97 static int lov_init_empty(const struct lu_env *env,
  98                           struct lov_device *dev, struct lov_object *lov,
  99                           const struct cl_object_conf *conf,
 100                           union  lov_layout_state *state)
 101 {
 102         return 0;
 103 }
 104
 105 static void lov_install_raid0(const struct lu_env *env,
 106                               struct lov_object *lov,
 107                               union  lov_layout_state *state)
 108 {
 109 }
 110
 111 static struct cl_object *lov_sub_find(const struct lu_env *env,
 112                                       struct cl_device *dev,
 113                                       const struct lu_fid *fid,
 114                                       const struct cl_object_conf *conf)
 115 {
 116         struct lu_object *o;
 117
 118         o = lu_object_find_at(env, cl2lu_dev(dev), fid, &conf->coc_lu);
 119         LASSERT(ergo(!IS_ERR(o), o->lo_dev->ld_type == &lovsub_device_type));
 120         return lu2cl(o);
 121 }
 122
 123 static int lov_init_sub(const struct lu_env *env, struct lov_object *lov,
 124                         struct cl_object *stripe, struct lov_layout_raid0 *r0,
 125                         int idx)
 126 {
 127         struct cl_object_header *hdr;
 128         struct cl_object_header *subhdr;
 129         struct cl_object_header *parent;
 130         struct lov_oinfo        *oinfo;
 131         int result;
 132
 133         if (OBD_FAIL_CHECK(OBD_FAIL_LOV_INIT)) {
 134                 /* For sanity:test_206.
 135                  * Do not leave the object in cache to avoid accessing
 136                  * freed memory. This is because osc_object is referring to
 137                  * lov_oinfo of lsm_stripe_data which will be freed due to
 138                  * this failure.
 139                  */
 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) {
 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);
 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                          */
 180                         lu_object_unhash(env, &stripe->co_lu);
 181                         result = -EAGAIN;
 182                 } else {
 183                         mask = D_ERROR;
 184                         result = -EIO;
 185                 }
 186
 187                 LU_OBJECT_DEBUG(mask, env, &stripe->co_lu,
 188                                 "stripe %d is already owned.\n", idx);
 189                 LU_OBJECT_DEBUG(mask, env, old_obj, "owned.\n");
 190                 LU_OBJECT_HEADER(mask, env, lov2lu(lov), "try to own.\n");
 191                 cl_object_put(env, stripe);
 192         }
 193         return result;
 194 }
 195
 196 static int lov_page_slice_fixup(struct lov_object *lov,
 197                                 struct cl_object *stripe)
 198 {
 199         struct cl_object_header *hdr = cl_object_header(&lov->lo_cl);
 200         struct cl_object *o;
 201
 202         cl_object_for_each(o, stripe)
 203                 o->co_slice_off += hdr->coh_page_bufsize;
 204
 205         return cl_object_header(stripe)->coh_page_bufsize;
 206 }
 207
 208 static int lov_init_raid0(const struct lu_env *env,
 209                           struct lov_device *dev, struct lov_object *lov,
 210                           const struct cl_object_conf *conf,
 211                           union  lov_layout_state *state)
 212 {
 213         int result;
 214         int i;
 215
 216         struct cl_object        *stripe;
 217         struct lov_thread_info  *lti     = lov_env_info(env);
 218         struct cl_object_conf   *subconf = &lti->lti_stripe_conf;
 219         struct lov_stripe_md    *lsm     = conf->u.coc_md->lsm;
 220         struct lu_fid      *ofid    = &lti->lti_fid;
 221         struct lov_layout_raid0 *r0      = &state->raid0;
 222
 223         if (lsm->lsm_magic != LOV_MAGIC_V1 && lsm->lsm_magic != LOV_MAGIC_V3) {
 224                 dump_lsm(D_ERROR, lsm);
 225                 LASSERTF(0, "magic mismatch, expected %d/%d, actual %d.\n",
 226                          LOV_MAGIC_V1, LOV_MAGIC_V3, lsm->lsm_magic);
 227         }
 228
 229         LASSERT(!lov->lo_lsm);
 230         lov->lo_lsm = lsm_addref(lsm);
 231         r0->lo_nr  = lsm->lsm_stripe_count;
 232         LASSERT(r0->lo_nr <= lov_targets_nr(dev));
 233
 234         r0->lo_sub = libcfs_kvzalloc(r0->lo_nr * sizeof(r0->lo_sub[0]),
 235                                      GFP_NOFS);
 236         if (r0->lo_sub) {
 237                 int psz = 0;
 238
 239                 result = 0;
 240                 subconf->coc_inode = conf->coc_inode;
 241                 spin_lock_init(&r0->lo_sub_lock);
 242                 /*
 243                  * Create stripe cl_objects.
 244                  */
 245                 for (i = 0; i < r0->lo_nr && result == 0; ++i) {
 246                         struct cl_device *subdev;
 247                         struct lov_oinfo *oinfo = lsm->lsm_oinfo[i];
 248                         int ost_idx = oinfo->loi_ost_idx;
 249
 250                         if (lov_oinfo_is_dummy(oinfo))
 251                                 continue;
 252
 253                         result = ostid_to_fid(ofid, &oinfo->loi_oi,
 254                                               oinfo->loi_ost_idx);
 255                         if (result != 0)
 256                                 goto out;
 257
 258                         subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
 259                         subconf->u.coc_oinfo = oinfo;
 260                         LASSERTF(subdev, "not init ost %d\n", ost_idx);
 261                         /* In the function below, .hs_keycmp resolves to
 262                          * lu_obj_hop_keycmp()
 263                          */
 264                         /* coverity[overrun-buffer-val] */
 265                         stripe = lov_sub_find(env, subdev, ofid, subconf);
 266                         if (!IS_ERR(stripe)) {
 267                                 result = lov_init_sub(env, lov, stripe, r0, i);
 268                                 if (result == -EAGAIN) { /* try again */
 269                                         --i;
 270                                         result = 0;
 271                                         continue;
 272                                 }
 273                         } else {
 274                                 result = PTR_ERR(stripe);
 275                         }
 276
 277                         if (result == 0) {
 278                                 int sz = lov_page_slice_fixup(lov, stripe);
 279
 280                                 LASSERT(ergo(psz > 0, psz == sz));
 281                                 psz = sz;
 282                         }
 283                 }
 284                 if (result == 0)
 285                         cl_object_header(&lov->lo_cl)->coh_page_bufsize += psz;
 286         } else {
 287                 result = -ENOMEM;
 288         }
 289 out:
 290         return result;
 291 }
 292
 293 static int lov_init_released(const struct lu_env *env,
 294                              struct lov_device *dev, struct lov_object *lov,
 295                              const struct cl_object_conf *conf,
 296                              union  lov_layout_state *state)
 297 {
 298         struct lov_stripe_md *lsm = conf->u.coc_md->lsm;
 299
 300         LASSERT(lsm);
 301         LASSERT(lsm_is_released(lsm));
 302         LASSERT(!lov->lo_lsm);
 303
 304         lov->lo_lsm = lsm_addref(lsm);
 305         return 0;
 306 }
 307
 308 static int lov_delete_empty(const struct lu_env *env, struct lov_object *lov,
 309                             union lov_layout_state *state)
 310 {
 311         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 312
 313         lov_layout_wait(env, lov);
 314         return 0;
 315 }
 316
 317 static void lov_subobject_kill(const struct lu_env *env, struct lov_object *lov,
 318                                struct lovsub_object *los, int idx)
 319 {
 320         struct cl_object        *sub;
 321         struct lov_layout_raid0 *r0;
 322         struct lu_site    *site;
 323         struct lu_site_bkt_data *bkt;
 324         wait_queue_t      *waiter;
 325
 326         r0  = &lov->u.raid0;
 327         LASSERT(r0->lo_sub[idx] == los);
 328
 329         sub  = lovsub2cl(los);
 330         site = sub->co_lu.lo_dev->ld_site;
 331         bkt  = lu_site_bkt_from_fid(site, &sub->co_lu.lo_header->loh_fid);
 332
 333         cl_object_kill(env, sub);
 334         /* release a reference to the sub-object and ... */
 335         lu_object_ref_del(&sub->co_lu, "lov-parent", lov);
 336         cl_object_put(env, sub);
 337
 338         /* ... wait until it is actually destroyed---sub-object clears its
 339          * ->lo_sub[] slot in lovsub_object_fini()
 340          */
 341         if (r0->lo_sub[idx] == los) {
 342                 waiter = &lov_env_info(env)->lti_waiter;
 343                 init_waitqueue_entry(waiter, current);
 344                 add_wait_queue(&bkt->lsb_marche_funebre, waiter);
 345                 set_current_state(TASK_UNINTERRUPTIBLE);
 346                 while (1) {
 347                         /* this wait-queue is signaled at the end of
 348                          * lu_object_free().
 349                          */
 350                         set_current_state(TASK_UNINTERRUPTIBLE);
 351                         spin_lock(&r0->lo_sub_lock);
 352                         if (r0->lo_sub[idx] == los) {
 353                                 spin_unlock(&r0->lo_sub_lock);
 354                                 schedule();
 355                         } else {
 356                                 spin_unlock(&r0->lo_sub_lock);
 357                                 set_current_state(TASK_RUNNING);
 358                                 break;
 359                         }
 360                 }
 361                 remove_wait_queue(&bkt->lsb_marche_funebre, waiter);
 362         }
 363         LASSERT(!r0->lo_sub[idx]);
 364 }
 365
 366 static int lov_delete_raid0(const struct lu_env *env, struct lov_object *lov,
 367                             union lov_layout_state *state)
 368 {
 369         struct lov_layout_raid0 *r0 = &state->raid0;
 370         struct lov_stripe_md    *lsm = lov->lo_lsm;
 371         int i;
 372
 373         dump_lsm(D_INODE, lsm);
 374
 375         lov_layout_wait(env, lov);
 376         if (r0->lo_sub) {
 377                 for (i = 0; i < r0->lo_nr; ++i) {
 378                         struct lovsub_object *los = r0->lo_sub[i];
 379
 380                         if (los) {
 381                                 cl_object_prune(env, &los->lso_cl);
 382                                 /*
 383                                  * If top-level object is to be evicted from
 384                                  * the cache, so are its sub-objects.
 385                                  */
 386                                 lov_subobject_kill(env, lov, los, i);
 387                         }
 388                 }
 389         }
 390         return 0;
 391 }
 392
 393 static void lov_fini_empty(const struct lu_env *env, struct lov_object *lov,
 394                            union lov_layout_state *state)
 395 {
 396         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 397 }
 398
 399 static void lov_fini_raid0(const struct lu_env *env, struct lov_object *lov,
 400                            union lov_layout_state *state)
 401 {
 402         struct lov_layout_raid0 *r0 = &state->raid0;
 403
 404         if (r0->lo_sub) {
 405                 kvfree(r0->lo_sub);
 406                 r0->lo_sub = NULL;
 407         }
 408
 409         dump_lsm(D_INODE, lov->lo_lsm);
 410         lov_free_memmd(&lov->lo_lsm);
 411 }
 412
 413 static void lov_fini_released(const struct lu_env *env, struct lov_object *lov,
 414                               union lov_layout_state *state)
 415 {
 416         dump_lsm(D_INODE, lov->lo_lsm);
 417         lov_free_memmd(&lov->lo_lsm);
 418 }
 419
 420 static int lov_print_empty(const struct lu_env *env, void *cookie,
 421                            lu_printer_t p, const struct lu_object *o)
 422 {
 423         (*p)(env, cookie, "empty %d\n", lu2lov(o)->lo_layout_invalid);
 424         return 0;
 425 }
 426
 427 static int lov_print_raid0(const struct lu_env *env, void *cookie,
 428                            lu_printer_t p, const struct lu_object *o)
 429 {
 430         struct lov_object       *lov = lu2lov(o);
 431         struct lov_layout_raid0 *r0  = lov_r0(lov);
 432         struct lov_stripe_md    *lsm = lov->lo_lsm;
 433         int                      i;
 434
 435         (*p)(env, cookie, "stripes: %d, %s, lsm{%p 0x%08X %d %u %u}:\n",
 436              r0->lo_nr, lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 437              lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 438              lsm->lsm_stripe_count, lsm->lsm_layout_gen);
 439         for (i = 0; i < r0->lo_nr; ++i) {
 440                 struct lu_object *sub;
 441
 442                 if (r0->lo_sub[i]) {
 443                         sub = lovsub2lu(r0->lo_sub[i]);
 444                         lu_object_print(env, cookie, p, sub);
 445                 } else {
 446                         (*p)(env, cookie, "sub %d absent\n", i);
 447                 }
 448         }
 449         return 0;
 450 }
 451
 452 static int lov_print_released(const struct lu_env *env, void *cookie,
 453                               lu_printer_t p, const struct lu_object *o)
 454 {
 455         struct lov_object       *lov = lu2lov(o);
 456         struct lov_stripe_md    *lsm = lov->lo_lsm;
 457
 458         (*p)(env, cookie,
 459              "released: %s, lsm{%p 0x%08X %d %u %u}:\n",
 460              lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 461              lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 462              lsm->lsm_stripe_count, lsm->lsm_layout_gen);
 463         return 0;
 464 }
 465
 466 /**
 467  * Implements cl_object_operations::coo_attr_get() method for an object
 468  * without stripes (LLT_EMPTY layout type).
 469  *
 470  * The only attributes this layer is authoritative in this case is
 471  * cl_attr::cat_blocks---it's 0.
 472  */
 473 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
 474                               struct cl_attr *attr)
 475 {
 476         attr->cat_blocks = 0;
 477         return 0;
 478 }
 479
 480 static int lov_attr_get_raid0(const struct lu_env *env, struct cl_object *obj,
 481                               struct cl_attr *attr)
 482 {
 483         struct lov_object       *lov = cl2lov(obj);
 484         struct lov_layout_raid0 *r0 = lov_r0(lov);
 485         struct cl_attr          *lov_attr = &r0->lo_attr;
 486         int                      result = 0;
 487
 488         /* this is called w/o holding type guard mutex, so it must be inside
 489          * an on going IO otherwise lsm may be replaced.
 490          * LU-2117: it turns out there exists one exception. For mmaped files,
 491          * the lock of those files may be requested in the other file's IO
 492          * context, and this function is called in ccc_lock_state(), it will
 493          * hit this assertion.
 494          * Anyway, it's still okay to call attr_get w/o type guard as layout
 495          * can't go if locks exist.
 496          */
 497         /* LASSERT(atomic_read(&lsm->lsm_refc) > 1); */
 498
 499         if (!r0->lo_attr_valid) {
 500                 struct lov_stripe_md    *lsm = lov->lo_lsm;
 501                 struct ost_lvb    *lvb = &lov_env_info(env)->lti_lvb;
 502                 __u64               kms = 0;
 503
 504                 memset(lvb, 0, sizeof(*lvb));
 505                 /* XXX: timestamps can be negative by sanity:test_39m,
 506                  * how can it be?
 507                  */
 508                 lvb->lvb_atime = LLONG_MIN;
 509                 lvb->lvb_ctime = LLONG_MIN;
 510                 lvb->lvb_mtime = LLONG_MIN;
 511
 512                 /*
 513                  * XXX that should be replaced with a loop over sub-objects,
 514                  * doing cl_object_attr_get() on them. But for now, let's
 515                  * reuse old lov code.
 516                  */
 517
 518                 /*
 519                  * XXX take lsm spin-lock to keep lov_merge_lvb_kms()
 520                  * happy. It's not needed, because new code uses
 521                  * ->coh_attr_guard spin-lock to protect consistency of
 522                  * sub-object attributes.
 523                  */
 524                 lov_stripe_lock(lsm);
 525                 result = lov_merge_lvb_kms(lsm, lvb, &kms);
 526                 lov_stripe_unlock(lsm);
 527                 if (result == 0) {
 528                         cl_lvb2attr(lov_attr, lvb);
 529                         lov_attr->cat_kms = kms;
 530                         r0->lo_attr_valid = 1;
 531                 }
 532         }
 533         if (result == 0) { /* merge results */
 534                 attr->cat_blocks = lov_attr->cat_blocks;
 535                 attr->cat_size = lov_attr->cat_size;
 536                 attr->cat_kms = lov_attr->cat_kms;
 537                 if (attr->cat_atime < lov_attr->cat_atime)
 538                         attr->cat_atime = lov_attr->cat_atime;
 539                 if (attr->cat_ctime < lov_attr->cat_ctime)
 540                         attr->cat_ctime = lov_attr->cat_ctime;
 541                 if (attr->cat_mtime < lov_attr->cat_mtime)
 542                         attr->cat_mtime = lov_attr->cat_mtime;
 543         }
 544         return result;
 545 }
 546
 547 static const struct lov_layout_operations lov_dispatch[] = {
 548         [LLT_EMPTY] = {
 549                 .llo_init      = lov_init_empty,
 550                 .llo_delete    = lov_delete_empty,
 551                 .llo_fini      = lov_fini_empty,
 552                 .llo_install   = lov_install_empty,
 553                 .llo_print     = lov_print_empty,
 554                 .llo_page_init = lov_page_init_empty,
 555                 .llo_lock_init = lov_lock_init_empty,
 556                 .llo_io_init   = lov_io_init_empty,
 557                 .llo_getattr   = lov_attr_get_empty
 558         },
 559         [LLT_RAID0] = {
 560                 .llo_init      = lov_init_raid0,
 561                 .llo_delete    = lov_delete_raid0,
 562                 .llo_fini      = lov_fini_raid0,
 563                 .llo_install   = lov_install_raid0,
 564                 .llo_print     = lov_print_raid0,
 565                 .llo_page_init = lov_page_init_raid0,
 566                 .llo_lock_init = lov_lock_init_raid0,
 567                 .llo_io_init   = lov_io_init_raid0,
 568                 .llo_getattr   = lov_attr_get_raid0
 569         },
 570         [LLT_RELEASED] = {
 571                 .llo_init      = lov_init_released,
 572                 .llo_delete    = lov_delete_empty,
 573                 .llo_fini      = lov_fini_released,
 574                 .llo_install   = lov_install_empty,
 575                 .llo_print     = lov_print_released,
 576                 .llo_page_init = lov_page_init_empty,
 577                 .llo_lock_init = lov_lock_init_empty,
 578                 .llo_io_init   = lov_io_init_released,
 579                 .llo_getattr   = lov_attr_get_empty
 580         }
 581 };
 582
 583 /**
 584  * Performs a double-dispatch based on the layout type of an object.
 585  */
 586 #define LOV_2DISPATCH_NOLOCK(obj, op, ...)                            \
 587 ({                                                                    \
 588         struct lov_object                     *__obj = (obj);     \
 589         enum lov_layout_type                __llt;                \
 590                                                                         \
 591         __llt = __obj->lo_type;                                  \
 592         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 593         lov_dispatch[__llt].op(__VA_ARGS__);                        \
 594 })
 595
 596 /**
 597  * Return lov_layout_type associated with a given lsm
 598  */
 599 static enum lov_layout_type lov_type(struct lov_stripe_md *lsm)
 600 {
 601         if (!lsm)
 602                 return LLT_EMPTY;
 603         if (lsm_is_released(lsm))
 604                 return LLT_RELEASED;
 605         return LLT_RAID0;
 606 }
 607
 608 static inline void lov_conf_freeze(struct lov_object *lov)
 609 {
 610         if (lov->lo_owner != current)
 611                 down_read(&lov->lo_type_guard);
 612 }
 613
 614 static inline void lov_conf_thaw(struct lov_object *lov)
 615 {
 616         if (lov->lo_owner != current)
 617                 up_read(&lov->lo_type_guard);
 618 }
 619
 620 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...)                      \
 621 ({                                                                    \
 622         struct lov_object                     *__obj = (obj);     \
 623         int                                  __lock = !!(lock);      \
 624         typeof(lov_dispatch[0].op(__VA_ARGS__)) __result;              \
 625                                                                         \
 626         if (__lock)                                                  \
 627                 lov_conf_freeze(__obj);                                 \
 628         __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__);    \
 629         if (__lock)                                                  \
 630                 lov_conf_thaw(__obj);                                   \
 631         __result;                                                      \
 632 })
 633
 634 /**
 635  * Performs a locked double-dispatch based on the layout type of an object.
 636  */
 637 #define LOV_2DISPATCH(obj, op, ...)                  \
 638         LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
 639
 640 #define LOV_2DISPATCH_VOID(obj, op, ...)                                \
 641 do {                                                                \
 642         struct lov_object                     *__obj = (obj);     \
 643         enum lov_layout_type                __llt;                \
 644                                                                         \
 645         lov_conf_freeze(__obj);                                         \
 646         __llt = __obj->lo_type;                                  \
 647         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 648         lov_dispatch[__llt].op(__VA_ARGS__);                        \
 649         lov_conf_thaw(__obj);                                           \
 650 } while (0)
 651
 652 static void lov_conf_lock(struct lov_object *lov)
 653 {
 654         LASSERT(lov->lo_owner != current);
 655         down_write(&lov->lo_type_guard);
 656         LASSERT(!lov->lo_owner);
 657         lov->lo_owner = current;
 658 }
 659
 660 static void lov_conf_unlock(struct lov_object *lov)
 661 {
 662         lov->lo_owner = NULL;
 663         up_write(&lov->lo_type_guard);
 664 }
 665
 666 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov)
 667 {
 668         struct l_wait_info lwi = { 0 };
 669
 670         while (atomic_read(&lov->lo_active_ios) > 0) {
 671                 CDEBUG(D_INODE, "file:" DFID " wait for active IO, now: %d.\n",
 672                        PFID(lu_object_fid(lov2lu(lov))),
 673                        atomic_read(&lov->lo_active_ios));
 674
 675                 l_wait_event(lov->lo_waitq,
 676                              atomic_read(&lov->lo_active_ios) == 0, &lwi);
 677         }
 678         return 0;
 679 }
 680
 681 static int lov_layout_change(const struct lu_env *unused,
 682                              struct lov_object *lov,
 683                              const struct cl_object_conf *conf)
 684 {
 685         int result;
 686         enum lov_layout_type llt = LLT_EMPTY;
 687         union lov_layout_state *state = &lov->u;
 688         const struct lov_layout_operations *old_ops;
 689         const struct lov_layout_operations *new_ops;
 690
 691         void *cookie;
 692         struct lu_env *env;
 693         int refcheck;
 694
 695         LASSERT(0 <= lov->lo_type && lov->lo_type < ARRAY_SIZE(lov_dispatch));
 696
 697         if (conf->u.coc_md)
 698                 llt = lov_type(conf->u.coc_md->lsm);
 699         LASSERT(0 <= llt && llt < ARRAY_SIZE(lov_dispatch));
 700
 701         cookie = cl_env_reenter();
 702         env = cl_env_get(&refcheck);
 703         if (IS_ERR(env)) {
 704                 cl_env_reexit(cookie);
 705                 return PTR_ERR(env);
 706         }
 707
 708         CDEBUG(D_INODE, DFID" from %s to %s\n",
 709                PFID(lu_object_fid(lov2lu(lov))),
 710                llt2str(lov->lo_type), llt2str(llt));
 711
 712         old_ops = &lov_dispatch[lov->lo_type];
 713         new_ops = &lov_dispatch[llt];
 714
 715         result = cl_object_prune(env, &lov->lo_cl);
 716         if (result != 0)
 717                 goto out;
 718
 719         result = old_ops->llo_delete(env, lov, &lov->u);
 720         if (result == 0) {
 721                 old_ops->llo_fini(env, lov, &lov->u);
 722
 723                 LASSERT(atomic_read(&lov->lo_active_ios) == 0);
 724
 725                 lov->lo_type = LLT_EMPTY;
 726                 result = new_ops->llo_init(env,
 727                                         lu2lov_dev(lov->lo_cl.co_lu.lo_dev),
 728                                         lov, conf, state);
 729                 if (result == 0) {
 730                         new_ops->llo_install(env, lov, state);
 731                         lov->lo_type = llt;
 732                 } else {
 733                         new_ops->llo_delete(env, lov, state);
 734                         new_ops->llo_fini(env, lov, state);
 735                         /* this file becomes an EMPTY file. */
 736                 }
 737         }
 738
 739 out:
 740         cl_env_put(env, &refcheck);
 741         cl_env_reexit(cookie);
 742         return result;
 743 }
 744
 745 /*****************************************************************************
 746  *
 747  * Lov object operations.
 748  *
 749  */
 750 int lov_object_init(const struct lu_env *env, struct lu_object *obj,
 751                     const struct lu_object_conf *conf)
 752 {
 753         struct lov_device           *dev   = lu2lov_dev(obj->lo_dev);
 754         struct lov_object           *lov   = lu2lov(obj);
 755         const struct cl_object_conf  *cconf = lu2cl_conf(conf);
 756         union  lov_layout_state      *set   = &lov->u;
 757         const struct lov_layout_operations *ops;
 758         int result;
 759
 760         init_rwsem(&lov->lo_type_guard);
 761         atomic_set(&lov->lo_active_ios, 0);
 762         init_waitqueue_head(&lov->lo_waitq);
 763
 764         cl_object_page_init(lu2cl(obj), sizeof(struct lov_page));
 765
 766         /* no locking is necessary, as object is being created */
 767         lov->lo_type = lov_type(cconf->u.coc_md->lsm);
 768         ops = &lov_dispatch[lov->lo_type];
 769         result = ops->llo_init(env, dev, lov, cconf, set);
 770         if (result == 0)
 771                 ops->llo_install(env, lov, set);
 772         return result;
 773 }
 774
 775 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
 776                         const struct cl_object_conf *conf)
 777 {
 778         struct lov_stripe_md    *lsm = NULL;
 779         struct lov_object       *lov = cl2lov(obj);
 780         int                      result = 0;
 781
 782         lov_conf_lock(lov);
 783         if (conf->coc_opc == OBJECT_CONF_INVALIDATE) {
 784                 lov->lo_layout_invalid = true;
 785                 result = 0;
 786                 goto out;
 787         }
 788
 789         if (conf->coc_opc == OBJECT_CONF_WAIT) {
 790                 if (lov->lo_layout_invalid &&
 791                     atomic_read(&lov->lo_active_ios) > 0) {
 792                         lov_conf_unlock(lov);
 793                         result = lov_layout_wait(env, lov);
 794                         lov_conf_lock(lov);
 795                 }
 796                 goto out;
 797         }
 798
 799         LASSERT(conf->coc_opc == OBJECT_CONF_SET);
 800
 801         if (conf->u.coc_md)
 802                 lsm = conf->u.coc_md->lsm;
 803         if ((!lsm && !lov->lo_lsm) ||
 804             ((lsm && lov->lo_lsm) &&
 805              (lov->lo_lsm->lsm_layout_gen == lsm->lsm_layout_gen) &&
 806              (lov->lo_lsm->lsm_pattern == lsm->lsm_pattern))) {
 807                 /* same version of layout */
 808                 lov->lo_layout_invalid = false;
 809                 result = 0;
 810                 goto out;
 811         }
 812
 813         /* will change layout - check if there still exists active IO. */
 814         if (atomic_read(&lov->lo_active_ios) > 0) {
 815                 lov->lo_layout_invalid = true;
 816                 result = -EBUSY;
 817                 goto out;
 818         }
 819
 820         result = lov_layout_change(env, lov, conf);
 821         lov->lo_layout_invalid = result != 0;
 822
 823 out:
 824         lov_conf_unlock(lov);
 825         CDEBUG(D_INODE, DFID" lo_layout_invalid=%d\n",
 826                PFID(lu_object_fid(lov2lu(lov))), lov->lo_layout_invalid);
 827         return result;
 828 }
 829
 830 static void lov_object_delete(const struct lu_env *env, struct lu_object *obj)
 831 {
 832         struct lov_object *lov = lu2lov(obj);
 833
 834         LOV_2DISPATCH_VOID(lov, llo_delete, env, lov, &lov->u);
 835 }
 836
 837 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
 838 {
 839         struct lov_object *lov = lu2lov(obj);
 840
 841         LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
 842         lu_object_fini(obj);
 843         kmem_cache_free(lov_object_kmem, lov);
 844 }
 845
 846 static int lov_object_print(const struct lu_env *env, void *cookie,
 847                             lu_printer_t p, const struct lu_object *o)
 848 {
 849         return LOV_2DISPATCH_NOLOCK(lu2lov(o), llo_print, env, cookie, p, o);
 850 }
 851
 852 int lov_page_init(const struct lu_env *env, struct cl_object *obj,
 853                   struct cl_page *page, pgoff_t index)
 854 {
 855         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_page_init, env, obj, page,
 856                                     index);
 857 }
 858
 859 /**
 860  * Implements cl_object_operations::clo_io_init() method for lov
 861  * layer. Dispatches to the appropriate layout io initialization method.
 862  */
 863 int lov_io_init(const struct lu_env *env, struct cl_object *obj,
 864                 struct cl_io *io)
 865 {
 866         CL_IO_SLICE_CLEAN(lov_env_io(env), lis_cl);
 867         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
 868                                      !io->ci_ignore_layout, env, obj, io);
 869 }
 870
 871 /**
 872  * An implementation of cl_object_operations::clo_attr_get() method for lov
 873  * layer. For raid0 layout this collects and merges attributes of all
 874  * sub-objects.
 875  */
 876 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
 877                         struct cl_attr *attr)
 878 {
 879         /* do not take lock, as this function is called under a
 880          * spin-lock. Layout is protected from changing by ongoing IO.
 881          */
 882         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
 883 }
 884
 885 static int lov_attr_set(const struct lu_env *env, struct cl_object *obj,
 886                         const struct cl_attr *attr, unsigned valid)
 887 {
 888         /*
 889          * No dispatch is required here, as no layout implements this.
 890          */
 891         return 0;
 892 }
 893
 894 int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
 895                   struct cl_lock *lock, const struct cl_io *io)
 896 {
 897         /* No need to lock because we've taken one refcount of layout.  */
 898         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_lock_init, env, obj, lock,
 899                                     io);
 900 }
 901
 902 static const struct cl_object_operations lov_ops = {
 903         .coo_page_init = lov_page_init,
 904         .coo_lock_init = lov_lock_init,
 905         .coo_io_init   = lov_io_init,
 906         .coo_attr_get  = lov_attr_get,
 907         .coo_attr_set  = lov_attr_set,
 908         .coo_conf_set  = lov_conf_set
 909 };
 910
 911 static const struct lu_object_operations lov_lu_obj_ops = {
 912         .loo_object_init      = lov_object_init,
 913         .loo_object_delete    = lov_object_delete,
 914         .loo_object_release   = NULL,
 915         .loo_object_free      = lov_object_free,
 916         .loo_object_print     = lov_object_print,
 917         .loo_object_invariant = NULL
 918 };
 919
 920 struct lu_object *lov_object_alloc(const struct lu_env *env,
 921                                    const struct lu_object_header *unused,
 922                                    struct lu_device *dev)
 923 {
 924         struct lov_object *lov;
 925         struct lu_object  *obj;
 926
 927         lov = kmem_cache_zalloc(lov_object_kmem, GFP_NOFS);
 928         if (lov) {
 929                 obj = lov2lu(lov);
 930                 lu_object_init(obj, NULL, dev);
 931                 lov->lo_cl.co_ops = &lov_ops;
 932                 lov->lo_type = -1; /* invalid, to catch uninitialized type */
 933                 /*
 934                  * object io operation vector (cl_object::co_iop) is installed
 935                  * later in lov_object_init(), as different vectors are used
 936                  * for object with different layouts.
 937                  */
 938                 obj->lo_ops = &lov_lu_obj_ops;
 939         } else {
 940                 obj = NULL;
 941         }
 942         return obj;
 943 }
 944
 945 static struct lov_stripe_md *lov_lsm_addref(struct lov_object *lov)
 946 {
 947         struct lov_stripe_md *lsm = NULL;
 948
 949         lov_conf_freeze(lov);
 950         if (lov->lo_lsm) {
 951                 lsm = lsm_addref(lov->lo_lsm);
 952                 CDEBUG(D_INODE, "lsm %p addref %d/%d by %p.\n",
 953                        lsm, atomic_read(&lsm->lsm_refc),
 954                        lov->lo_layout_invalid, current);
 955         }
 956         lov_conf_thaw(lov);
 957         return lsm;
 958 }
 959
 960 struct lov_stripe_md *lov_lsm_get(struct cl_object *clobj)
 961 {
 962         struct lu_object *luobj;
 963         struct lov_stripe_md *lsm = NULL;
 964
 965         if (!clobj)
 966                 return NULL;
 967
 968         luobj = lu_object_locate(&cl_object_header(clobj)->coh_lu,
 969                                  &lov_device_type);
 970         if (luobj)
 971                 lsm = lov_lsm_addref(lu2lov(luobj));
 972         return lsm;
 973 }
 974 EXPORT_SYMBOL(lov_lsm_get);
 975
 976 void lov_lsm_put(struct cl_object *unused, struct lov_stripe_md *lsm)
 977 {
 978         if (lsm)
 979                 lov_free_memmd(&lsm);
 980 }
 981 EXPORT_SYMBOL(lov_lsm_put);
 982
 983 int lov_read_and_clear_async_rc(struct cl_object *clob)
 984 {
 985         struct lu_object *luobj;
 986         int rc = 0;
 987
 988         luobj = lu_object_locate(&cl_object_header(clob)->coh_lu,
 989                                  &lov_device_type);
 990         if (luobj) {
 991                 struct lov_object *lov = lu2lov(luobj);
 992
 993                 lov_conf_freeze(lov);
 994                 switch (lov->lo_type) {
 995                 case LLT_RAID0: {
 996                         struct lov_stripe_md *lsm;
 997                         int i;
 998
 999                         lsm = lov->lo_lsm;
1000                         for (i = 0; i < lsm->lsm_stripe_count; i++) {
1001                                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
1002
1003                                 if (lov_oinfo_is_dummy(loi))
1004                                         continue;
1005
1006                                 if (loi->loi_ar.ar_rc && !rc)
1007                                         rc = loi->loi_ar.ar_rc;
1008                                 loi->loi_ar.ar_rc = 0;
1009                         }
1010                 }
1011                 case LLT_RELEASED:
1012                 case LLT_EMPTY:
1013                         break;
1014                 default:
1015                         LBUG();
1016                 }
1017                 lov_conf_thaw(lov);
1018         }
1019         return rc;
1020 }
1021 EXPORT_SYMBOL(lov_read_and_clear_async_rc);
1022
1023 /** @} lov */