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 "lclient.h"
  46
  47 /** \addtogroup lov
  48  *  @{
  49  */
  50
  51 /*****************************************************************************
  52  *
  53  * Layout operations.
  54  *
  55  */
  56
  57 struct lov_layout_operations {
  58         int (*llo_init)(const struct lu_env *env, struct lov_device *dev,
  59                         struct lov_object *lov,
  60                         const struct cl_object_conf *conf,
  61                         union lov_layout_state *state);
  62         int (*llo_delete)(const struct lu_env *env, struct lov_object *lov,
  63                            union lov_layout_state *state);
  64         void (*llo_fini)(const struct lu_env *env, struct lov_object *lov,
  65                          union lov_layout_state *state);
  66         void (*llo_install)(const struct lu_env *env, struct lov_object *lov,
  67                             union lov_layout_state *state);
  68         int  (*llo_print)(const struct lu_env *env, void *cookie,
  69                           lu_printer_t p, const struct lu_object *o);
  70         int  (*llo_page_init)(const struct lu_env *env, struct cl_object *obj,
  71                               struct cl_page *page, pgoff_t index);
  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, struct lov_layout_raid0 *r0,
 127                         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
 149         oinfo = lov->lo_lsm->lsm_oinfo[idx];
 150         CDEBUG(D_INODE, DFID"@%p[%d] -> "DFID"@%p: ostid: "DOSTID
 151                " idx: %d gen: %d\n",
 152                PFID(&subhdr->coh_lu.loh_fid), subhdr, idx,
 153                PFID(&hdr->coh_lu.loh_fid), hdr, POSTID(&oinfo->loi_oi),
 154                oinfo->loi_ost_idx, oinfo->loi_ost_gen);
 155
 156         /* reuse ->coh_attr_guard to protect coh_parent change */
 157         spin_lock(&subhdr->coh_attr_guard);
 158         parent = subhdr->coh_parent;
 159         if (parent == NULL) {
 160                 subhdr->coh_parent = hdr;
 161                 spin_unlock(&subhdr->coh_attr_guard);
 162                 subhdr->coh_nesting = hdr->coh_nesting + 1;
 163                 lu_object_ref_add(&stripe->co_lu, "lov-parent", lov);
 164                 r0->lo_sub[idx] = cl2lovsub(stripe);
 165                 r0->lo_sub[idx]->lso_super = lov;
 166                 r0->lo_sub[idx]->lso_index = idx;
 167                 result = 0;
 168         } else {
 169                 struct lu_object  *old_obj;
 170                 struct lov_object *old_lov;
 171                 unsigned int mask = D_INODE;
 172
 173                 spin_unlock(&subhdr->coh_attr_guard);
 174                 old_obj = lu_object_locate(&parent->coh_lu, &lov_device_type);
 175                 LASSERT(old_obj != NULL);
 176                 old_lov = cl2lov(lu2cl(old_obj));
 177                 if (old_lov->lo_layout_invalid) {
 178                         /* the object's layout has already changed but isn't
 179                          * refreshed */
 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         if (stripe == NULL)
 203                 return hdr->coh_page_bufsize - lov->lo_cl.co_slice_off -
 204                        cfs_size_round(sizeof(struct lov_page));
 205
 206         cl_object_for_each(o, stripe)
 207                 o->co_slice_off += hdr->coh_page_bufsize;
 208
 209         return cl_object_header(stripe)->coh_page_bufsize;
 210 }
 211
 212 static int lov_init_raid0(const struct lu_env *env,
 213                           struct lov_device *dev, struct lov_object *lov,
 214                           const struct cl_object_conf *conf,
 215                           union  lov_layout_state *state)
 216 {
 217         int result;
 218         int i;
 219
 220         struct cl_object        *stripe;
 221         struct lov_thread_info  *lti     = lov_env_info(env);
 222         struct cl_object_conf   *subconf = &lti->lti_stripe_conf;
 223         struct lov_stripe_md    *lsm     = conf->u.coc_md->lsm;
 224         struct lu_fid           *ofid    = &lti->lti_fid;
 225         struct lov_layout_raid0 *r0      = &state->raid0;
 226
 227         ENTRY;
 228
 229         if (lsm->lsm_magic != LOV_MAGIC_V1 && lsm->lsm_magic != LOV_MAGIC_V3) {
 230                 dump_lsm(D_ERROR, lsm);
 231                 LASSERTF(0, "magic mismatch, expected %d/%d, actual %d.\n",
 232                          LOV_MAGIC_V1, LOV_MAGIC_V3, lsm->lsm_magic);
 233         }
 234
 235         LASSERT(lov->lo_lsm == NULL);
 236         lov->lo_lsm = lsm_addref(lsm);
 237         r0->lo_nr  = lsm->lsm_stripe_count;
 238         LASSERT(r0->lo_nr <= lov_targets_nr(dev));
 239
 240         lov->lo_layout_invalid = true;
 241
 242         OBD_ALLOC_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 243         if (r0->lo_sub != NULL) {
 244                 int psz = 0;
 245
 246                 result = 0;
 247                 subconf->coc_inode = conf->coc_inode;
 248                 spin_lock_init(&r0->lo_sub_lock);
 249                 /*
 250                  * Create stripe cl_objects.
 251                  */
 252                 for (i = 0; i < r0->lo_nr && result == 0; ++i) {
 253                         struct cl_device *subdev;
 254                         struct lov_oinfo *oinfo = lsm->lsm_oinfo[i];
 255                         int ost_idx = oinfo->loi_ost_idx;
 256
 257                         if (lov_oinfo_is_dummy(oinfo))
 258                                 continue;
 259
 260                         result = ostid_to_fid(ofid, &oinfo->loi_oi,
 261                                               oinfo->loi_ost_idx);
 262                         if (result != 0)
 263                                 GOTO(out, result);
 264
 265                         subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
 266                         subconf->u.coc_oinfo = oinfo;
 267                         LASSERTF(subdev != NULL, "not init ost %d\n", ost_idx);
 268                         /* In the function below, .hs_keycmp resolves to
 269                          * lu_obj_hop_keycmp() */
 270                         /* coverity[overrun-buffer-val] */
 271                         stripe = lov_sub_find(env, subdev, ofid, subconf);
 272                         if (!IS_ERR(stripe)) {
 273                                 result = lov_init_sub(env, lov, stripe, r0, i);
 274                                 if (result == -EAGAIN) { /* try again */
 275                                         --i;
 276                                         result = 0;
 277                                         continue;
 278                                 }
 279                         } else {
 280                                 result = PTR_ERR(stripe);
 281                         }
 282
 283                         if (result == 0) {
 284                                 int sz = lov_page_slice_fixup(lov, stripe);
 285                                 LASSERT(ergo(psz > 0, psz == sz));
 286                                 psz = sz;
 287                         }
 288                 }
 289                 if (result == 0)
 290                         cl_object_header(&lov->lo_cl)->coh_page_bufsize += psz;
 291         } else
 292                 result = -ENOMEM;
 293 out:
 294         RETURN(result);
 295 }
 296
 297 static int lov_init_released(const struct lu_env *env,
 298                         struct lov_device *dev, struct lov_object *lov,
 299                         const struct cl_object_conf *conf,
 300                         union  lov_layout_state *state)
 301 {
 302         struct lov_stripe_md *lsm = conf->u.coc_md->lsm;
 303
 304         LASSERT(lsm != NULL);
 305         LASSERT(lsm_is_released(lsm));
 306         LASSERT(lov->lo_lsm == NULL);
 307
 308         lov->lo_lsm = lsm_addref(lsm);
 309         return 0;
 310 }
 311
 312 static int lov_delete_empty(const struct lu_env *env, struct lov_object *lov,
 313                             union lov_layout_state *state)
 314 {
 315         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 316
 317         lov_layout_wait(env, lov);
 318         return 0;
 319 }
 320
 321 static void lov_subobject_kill(const struct lu_env *env, struct lov_object *lov,
 322                                struct lovsub_object *los, int idx)
 323 {
 324         struct cl_object        *sub;
 325         struct lov_layout_raid0 *r0;
 326         struct lu_site          *site;
 327         struct lu_site_bkt_data *bkt;
 328         wait_queue_t          *waiter;
 329
 330         r0  = &lov->u.raid0;
 331         LASSERT(r0->lo_sub[idx] == los);
 332
 333         sub  = lovsub2cl(los);
 334         site = sub->co_lu.lo_dev->ld_site;
 335         bkt  = lu_site_bkt_from_fid(site, &sub->co_lu.lo_header->loh_fid);
 336
 337         cl_object_kill(env, sub);
 338         /* release a reference to the sub-object and ... */
 339         lu_object_ref_del(&sub->co_lu, "lov-parent", lov);
 340         cl_object_put(env, sub);
 341
 342         /* ... wait until it is actually destroyed---sub-object clears its
 343          * ->lo_sub[] slot in lovsub_object_fini() */
 344         if (r0->lo_sub[idx] == los) {
 345                 waiter = &lov_env_info(env)->lti_waiter;
 346                 init_waitqueue_entry_current(waiter);
 347                 add_wait_queue(&bkt->lsb_marche_funebre, waiter);
 348                 set_current_state(TASK_UNINTERRUPTIBLE);
 349                 while (1) {
 350                         /* this wait-queue is signaled at the end of
 351                          * lu_object_free(). */
 352                         set_current_state(TASK_UNINTERRUPTIBLE);
 353                         spin_lock(&r0->lo_sub_lock);
 354                         if (r0->lo_sub[idx] == los) {
 355                                 spin_unlock(&r0->lo_sub_lock);
 356                                 waitq_wait(waiter, TASK_UNINTERRUPTIBLE);
 357                         } else {
 358                                 spin_unlock(&r0->lo_sub_lock);
 359                                 set_current_state(TASK_RUNNING);
 360                                 break;
 361                         }
 362                 }
 363                 remove_wait_queue(&bkt->lsb_marche_funebre, waiter);
 364         }
 365         LASSERT(r0->lo_sub[idx] == NULL);
 366 }
 367
 368 static int lov_delete_raid0(const struct lu_env *env, struct lov_object *lov,
 369                             union lov_layout_state *state)
 370 {
 371         struct lov_layout_raid0 *r0 = &state->raid0;
 372         struct lov_stripe_md    *lsm = lov->lo_lsm;
 373         int i;
 374
 375         ENTRY;
 376
 377         dump_lsm(D_INODE, lsm);
 378
 379         lov_layout_wait(env, lov);
 380         if (r0->lo_sub != NULL) {
 381                 for (i = 0; i < r0->lo_nr; ++i) {
 382                         struct lovsub_object *los = r0->lo_sub[i];
 383
 384                         if (los != NULL) {
 385                                 cl_object_prune(env, &los->lso_cl);
 386                                 /*
 387                                  * If top-level object is to be evicted from
 388                                  * the cache, so are its sub-objects.
 389                                  */
 390                                 lov_subobject_kill(env, lov, los, i);
 391                         }
 392                 }
 393         }
 394         RETURN(0);
 395 }
 396
 397 static void lov_fini_empty(const struct lu_env *env, struct lov_object *lov,
 398                            union lov_layout_state *state)
 399 {
 400         LASSERT(lov->lo_type == LLT_EMPTY || lov->lo_type == LLT_RELEASED);
 401 }
 402
 403 static void lov_fini_raid0(const struct lu_env *env, struct lov_object *lov,
 404                            union lov_layout_state *state)
 405 {
 406         struct lov_layout_raid0 *r0 = &state->raid0;
 407         ENTRY;
 408
 409         if (r0->lo_sub != NULL) {
 410                 OBD_FREE_LARGE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
 411                 r0->lo_sub = NULL;
 412         }
 413
 414         dump_lsm(D_INODE, lov->lo_lsm);
 415         lov_free_memmd(&lov->lo_lsm);
 416
 417         EXIT;
 418 }
 419
 420 static void lov_fini_released(const struct lu_env *env, struct lov_object *lov,
 421                                 union lov_layout_state *state)
 422 {
 423         ENTRY;
 424         dump_lsm(D_INODE, lov->lo_lsm);
 425         lov_free_memmd(&lov->lo_lsm);
 426         EXIT;
 427 }
 428
 429 static int lov_print_empty(const struct lu_env *env, void *cookie,
 430                            lu_printer_t p, const struct lu_object *o)
 431 {
 432         (*p)(env, cookie, "empty %d\n", lu2lov(o)->lo_layout_invalid);
 433         return 0;
 434 }
 435
 436 static int lov_print_raid0(const struct lu_env *env, void *cookie,
 437                            lu_printer_t p, const struct lu_object *o)
 438 {
 439         struct lov_object       *lov = lu2lov(o);
 440         struct lov_layout_raid0 *r0  = lov_r0(lov);
 441         struct lov_stripe_md    *lsm = lov->lo_lsm;
 442         int                      i;
 443
 444         (*p)(env, cookie, "stripes: %d, %s, lsm{%p 0x%08X %d %u %u}:\n",
 445                 r0->lo_nr, lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 446                 lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 447                 lsm->lsm_stripe_count, lsm->lsm_layout_gen);
 448         for (i = 0; i < r0->lo_nr; ++i) {
 449                 struct lu_object *sub;
 450
 451                 if (r0->lo_sub[i] != NULL) {
 452                         sub = lovsub2lu(r0->lo_sub[i]);
 453                         lu_object_print(env, cookie, p, sub);
 454                 } else {
 455                         (*p)(env, cookie, "sub %d absent\n", i);
 456                 }
 457         }
 458         return 0;
 459 }
 460
 461 static int lov_print_released(const struct lu_env *env, void *cookie,
 462                                 lu_printer_t p, const struct lu_object *o)
 463 {
 464         struct lov_object       *lov = lu2lov(o);
 465         struct lov_stripe_md    *lsm = lov->lo_lsm;
 466
 467         (*p)(env, cookie,
 468                 "released: %s, lsm{%p 0x%08X %d %u %u}:\n",
 469                 lov->lo_layout_invalid ? "invalid" : "valid", lsm,
 470                 lsm->lsm_magic, atomic_read(&lsm->lsm_refc),
 471                 lsm->lsm_stripe_count, lsm->lsm_layout_gen);
 472         return 0;
 473 }
 474
 475 /**
 476  * Implements cl_object_operations::coo_attr_get() method for an object
 477  * without stripes (LLT_EMPTY layout type).
 478  *
 479  * The only attributes this layer is authoritative in this case is
 480  * cl_attr::cat_blocks---it's 0.
 481  */
 482 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
 483                               struct cl_attr *attr)
 484 {
 485         attr->cat_blocks = 0;
 486         return 0;
 487 }
 488
 489 static int lov_attr_get_raid0(const struct lu_env *env, struct cl_object *obj,
 490                               struct cl_attr *attr)
 491 {
 492         struct lov_object       *lov = cl2lov(obj);
 493         struct lov_layout_raid0 *r0 = lov_r0(lov);
 494         struct cl_attr          *lov_attr = &r0->lo_attr;
 495         int                      result = 0;
 496
 497         ENTRY;
 498
 499         /* this is called w/o holding type guard mutex, so it must be inside
 500          * an on going IO otherwise lsm may be replaced.
 501          * LU-2117: it turns out there exists one exception. For mmaped files,
 502          * the lock of those files may be requested in the other file's IO
 503          * context, and this function is called in ccc_lock_state(), it will
 504          * hit this assertion.
 505          * Anyway, it's still okay to call attr_get w/o type guard as layout
 506          * can't go if locks exist. */
 507         /* LASSERT(atomic_read(&lsm->lsm_refc) > 1); */
 508
 509         if (!r0->lo_attr_valid) {
 510                 struct lov_stripe_md    *lsm = lov->lo_lsm;
 511                 struct ost_lvb          *lvb = &lov_env_info(env)->lti_lvb;
 512                 __u64                    kms = 0;
 513
 514                 memset(lvb, 0, sizeof(*lvb));
 515                 /* XXX: timestamps can be negative by sanity:test_39m,
 516                  * how can it be? */
 517                 lvb->lvb_atime = LLONG_MIN;
 518                 lvb->lvb_ctime = LLONG_MIN;
 519                 lvb->lvb_mtime = LLONG_MIN;
 520
 521                 /*
 522                  * XXX that should be replaced with a loop over sub-objects,
 523                  * doing cl_object_attr_get() on them. But for now, let's
 524                  * reuse old lov code.
 525                  */
 526
 527                 /*
 528                  * XXX take lsm spin-lock to keep lov_merge_lvb_kms()
 529                  * happy. It's not needed, because new code uses
 530                  * ->coh_attr_guard spin-lock to protect consistency of
 531                  * sub-object attributes.
 532                  */
 533                 lov_stripe_lock(lsm);
 534                 result = lov_merge_lvb_kms(lsm, lvb, &kms);
 535                 lov_stripe_unlock(lsm);
 536                 if (result == 0) {
 537                         cl_lvb2attr(lov_attr, lvb);
 538                         lov_attr->cat_kms = kms;
 539                         r0->lo_attr_valid = 1;
 540                 }
 541         }
 542         if (result == 0) { /* merge results */
 543                 attr->cat_blocks = lov_attr->cat_blocks;
 544                 attr->cat_size = lov_attr->cat_size;
 545                 attr->cat_kms = lov_attr->cat_kms;
 546                 if (attr->cat_atime < lov_attr->cat_atime)
 547                         attr->cat_atime = lov_attr->cat_atime;
 548                 if (attr->cat_ctime < lov_attr->cat_ctime)
 549                         attr->cat_ctime = lov_attr->cat_ctime;
 550                 if (attr->cat_mtime < lov_attr->cat_mtime)
 551                         attr->cat_mtime = lov_attr->cat_mtime;
 552         }
 553         RETURN(result);
 554 }
 555
 556 const static struct lov_layout_operations lov_dispatch[] = {
 557         [LLT_EMPTY] = {
 558                 .llo_init      = lov_init_empty,
 559                 .llo_delete    = lov_delete_empty,
 560                 .llo_fini      = lov_fini_empty,
 561                 .llo_install   = lov_install_empty,
 562                 .llo_print     = lov_print_empty,
 563                 .llo_page_init = lov_page_init_empty,
 564                 .llo_lock_init = lov_lock_init_empty,
 565                 .llo_io_init   = lov_io_init_empty,
 566                 .llo_getattr   = lov_attr_get_empty
 567         },
 568         [LLT_RAID0] = {
 569                 .llo_init      = lov_init_raid0,
 570                 .llo_delete    = lov_delete_raid0,
 571                 .llo_fini      = lov_fini_raid0,
 572                 .llo_install   = lov_install_raid0,
 573                 .llo_print     = lov_print_raid0,
 574                 .llo_page_init = lov_page_init_raid0,
 575                 .llo_lock_init = lov_lock_init_raid0,
 576                 .llo_io_init   = lov_io_init_raid0,
 577                 .llo_getattr   = lov_attr_get_raid0
 578         },
 579         [LLT_RELEASED] = {
 580                 .llo_init      = lov_init_released,
 581                 .llo_delete    = lov_delete_empty,
 582                 .llo_fini      = lov_fini_released,
 583                 .llo_install   = lov_install_empty,
 584                 .llo_print     = lov_print_released,
 585                 .llo_page_init = lov_page_init_empty,
 586                 .llo_lock_init = lov_lock_init_empty,
 587                 .llo_io_init   = lov_io_init_released,
 588                 .llo_getattr   = lov_attr_get_empty
 589         }
 590 };
 591
 592 /**
 593  * Performs a double-dispatch based on the layout type of an object.
 594  */
 595 #define LOV_2DISPATCH_NOLOCK(obj, op, ...)                              \
 596 ({                                                                      \
 597         struct lov_object                      *__obj = (obj);          \
 598         enum lov_layout_type                    __llt;                  \
 599                                                                         \
 600         __llt = __obj->lo_type;                                         \
 601         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 602         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 603 })
 604
 605 /**
 606  * Return lov_layout_type associated with a given lsm
 607  */
 608 static enum lov_layout_type lov_type(struct lov_stripe_md *lsm)
 609 {
 610         if (lsm == NULL)
 611                 return LLT_EMPTY;
 612         if (lsm_is_released(lsm))
 613                 return LLT_RELEASED;
 614         return LLT_RAID0;
 615 }
 616
 617 static inline void lov_conf_freeze(struct lov_object *lov)
 618 {
 619         if (lov->lo_owner != current)
 620                 down_read(&lov->lo_type_guard);
 621 }
 622
 623 static inline void lov_conf_thaw(struct lov_object *lov)
 624 {
 625         if (lov->lo_owner != current)
 626                 up_read(&lov->lo_type_guard);
 627 }
 628
 629 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...)                       \
 630 ({                                                                      \
 631         struct lov_object                      *__obj = (obj);          \
 632         int                                     __lock = !!(lock);      \
 633         typeof(lov_dispatch[0].op(__VA_ARGS__)) __result;               \
 634                                                                         \
 635         if (__lock)                                                     \
 636                 lov_conf_freeze(__obj);                                 \
 637         __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__);          \
 638         if (__lock)                                                     \
 639                 lov_conf_thaw(__obj);                                   \
 640         __result;                                                       \
 641 })
 642
 643 /**
 644  * Performs a locked double-dispatch based on the layout type of an object.
 645  */
 646 #define LOV_2DISPATCH(obj, op, ...)                     \
 647         LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
 648
 649 #define LOV_2DISPATCH_VOID(obj, op, ...)                                \
 650 do {                                                                    \
 651         struct lov_object                      *__obj = (obj);          \
 652         enum lov_layout_type                    __llt;                  \
 653                                                                         \
 654         lov_conf_freeze(__obj);                                         \
 655         __llt = __obj->lo_type;                                         \
 656         LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch));        \
 657         lov_dispatch[__llt].op(__VA_ARGS__);                            \
 658         lov_conf_thaw(__obj);                                           \
 659 } while (0)
 660
 661 static void lov_conf_lock(struct lov_object *lov)
 662 {
 663         LASSERT(lov->lo_owner != current);
 664         down_write(&lov->lo_type_guard);
 665         LASSERT(lov->lo_owner == NULL);
 666         lov->lo_owner = current;
 667 }
 668
 669 static void lov_conf_unlock(struct lov_object *lov)
 670 {
 671         lov->lo_owner = NULL;
 672         up_write(&lov->lo_type_guard);
 673 }
 674
 675 static int lov_layout_wait(const struct lu_env *env, struct lov_object *lov)
 676 {
 677         struct l_wait_info lwi = { 0 };
 678         ENTRY;
 679
 680         while (atomic_read(&lov->lo_active_ios) > 0) {
 681                 CDEBUG(D_INODE, "file:"DFID" wait for active IO, now: %d.\n",
 682                         PFID(lu_object_fid(lov2lu(lov))),
 683                         atomic_read(&lov->lo_active_ios));
 684
 685                 l_wait_event(lov->lo_waitq,
 686                              atomic_read(&lov->lo_active_ios) == 0, &lwi);
 687         }
 688         RETURN(0);
 689 }
 690
 691 static int lov_layout_change(const struct lu_env *unused,
 692                              struct lov_object *lov,
 693                              const struct cl_object_conf *conf)
 694 {
 695         int result;
 696         enum lov_layout_type llt = LLT_EMPTY;
 697         union lov_layout_state *state = &lov->u;
 698         const struct lov_layout_operations *old_ops;
 699         const struct lov_layout_operations *new_ops;
 700
 701         void *cookie;
 702         struct lu_env *env;
 703         int refcheck;
 704         ENTRY;
 705
 706         LASSERT(0 <= lov->lo_type && lov->lo_type < ARRAY_SIZE(lov_dispatch));
 707
 708         if (conf->u.coc_md != NULL)
 709                 llt = lov_type(conf->u.coc_md->lsm);
 710         LASSERT(0 <= llt && llt < ARRAY_SIZE(lov_dispatch));
 711
 712         cookie = cl_env_reenter();
 713         env = cl_env_get(&refcheck);
 714         if (IS_ERR(env)) {
 715                 cl_env_reexit(cookie);
 716                 RETURN(PTR_ERR(env));
 717         }
 718
 719         CDEBUG(D_INODE, DFID" from %s to %s\n",
 720                PFID(lu_object_fid(lov2lu(lov))),
 721                llt2str(lov->lo_type), llt2str(llt));
 722
 723         old_ops = &lov_dispatch[lov->lo_type];
 724         new_ops = &lov_dispatch[llt];
 725
 726         result = cl_object_prune(env, &lov->lo_cl);
 727         if (result != 0)
 728                 GOTO(out, result);
 729
 730         result = old_ops->llo_delete(env, lov, &lov->u);
 731         if (result == 0) {
 732                 old_ops->llo_fini(env, lov, &lov->u);
 733
 734                 LASSERT(atomic_read(&lov->lo_active_ios) == 0);
 735
 736                 lov->lo_type = LLT_EMPTY;
 737                 /* page bufsize fixup */
 738                 cl_object_header(&lov->lo_cl)->coh_page_bufsize -=
 739                         lov_page_slice_fixup(lov, NULL);
 740
 741                 result = new_ops->llo_init(env,
 742                                         lu2lov_dev(lov->lo_cl.co_lu.lo_dev),
 743                                         lov, conf, state);
 744                 if (result == 0) {
 745                         new_ops->llo_install(env, lov, state);
 746                         lov->lo_type = llt;
 747                 } else {
 748                         new_ops->llo_delete(env, lov, state);
 749                         new_ops->llo_fini(env, lov, state);
 750                         /* this file becomes an EMPTY file. */
 751                 }
 752         }
 753
 754 out:
 755         cl_env_put(env, &refcheck);
 756         cl_env_reexit(cookie);
 757         RETURN(result);
 758 }
 759
 760 /*****************************************************************************
 761  *
 762  * Lov object operations.
 763  *
 764  */
 765 int lov_object_init(const struct lu_env *env, struct lu_object *obj,
 766                     const struct lu_object_conf *conf)
 767 {
 768         struct lov_device            *dev   = lu2lov_dev(obj->lo_dev);
 769         struct lov_object            *lov   = lu2lov(obj);
 770         const struct cl_object_conf  *cconf = lu2cl_conf(conf);
 771         union  lov_layout_state      *set   = &lov->u;
 772         const struct lov_layout_operations *ops;
 773         int result;
 774
 775         ENTRY;
 776         init_rwsem(&lov->lo_type_guard);
 777         atomic_set(&lov->lo_active_ios, 0);
 778         init_waitqueue_head(&lov->lo_waitq);
 779
 780         cl_object_page_init(lu2cl(obj), sizeof(struct lov_page));
 781
 782         /* no locking is necessary, as object is being created */
 783         lov->lo_type = lov_type(cconf->u.coc_md->lsm);
 784         ops = &lov_dispatch[lov->lo_type];
 785         result = ops->llo_init(env, dev, lov, cconf, set);
 786         if (result == 0)
 787                 ops->llo_install(env, lov, set);
 788         RETURN(result);
 789 }
 790
 791 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
 792                         const struct cl_object_conf *conf)
 793 {
 794         struct lov_stripe_md    *lsm = NULL;
 795         struct lov_object       *lov = cl2lov(obj);
 796         int                      result = 0;
 797         ENTRY;
 798
 799         lov_conf_lock(lov);
 800         if (conf->coc_opc == OBJECT_CONF_INVALIDATE) {
 801                 lov->lo_layout_invalid = true;
 802                 GOTO(out, result = 0);
 803         }
 804
 805         if (conf->coc_opc == OBJECT_CONF_WAIT) {
 806                 if (lov->lo_layout_invalid &&
 807                     atomic_read(&lov->lo_active_ios) > 0) {
 808                         lov_conf_unlock(lov);
 809                         result = lov_layout_wait(env, lov);
 810                         lov_conf_lock(lov);
 811                 }
 812                 GOTO(out, result);
 813         }
 814
 815         LASSERT(conf->coc_opc == OBJECT_CONF_SET);
 816
 817         if (conf->u.coc_md != NULL)
 818                 lsm = conf->u.coc_md->lsm;
 819         if ((lsm == NULL && lov->lo_lsm == NULL) ||
 820             ((lsm != NULL && lov->lo_lsm != NULL) &&
 821              (lov->lo_lsm->lsm_layout_gen == lsm->lsm_layout_gen) &&
 822              (lov->lo_lsm->lsm_pattern == lsm->lsm_pattern))) {
 823                 /* same version of layout */
 824                 lov->lo_layout_invalid = false;
 825                 GOTO(out, result = 0);
 826         }
 827
 828         /* will change layout - check if there still exists active IO. */
 829         if (atomic_read(&lov->lo_active_ios) > 0) {
 830                 lov->lo_layout_invalid = true;
 831                 GOTO(out, result = -EBUSY);
 832         }
 833
 834         result = lov_layout_change(env, lov, conf);
 835         lov->lo_layout_invalid = result != 0;
 836         EXIT;
 837
 838 out:
 839         lov_conf_unlock(lov);
 840         CDEBUG(D_INODE, DFID" lo_layout_invalid=%d\n",
 841                PFID(lu_object_fid(lov2lu(lov))), lov->lo_layout_invalid);
 842         RETURN(result);
 843 }
 844
 845 static void lov_object_delete(const struct lu_env *env, struct lu_object *obj)
 846 {
 847         struct lov_object *lov = lu2lov(obj);
 848
 849         ENTRY;
 850         LOV_2DISPATCH_VOID(lov, llo_delete, env, lov, &lov->u);
 851         EXIT;
 852 }
 853
 854 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
 855 {
 856         struct lov_object *lov = lu2lov(obj);
 857
 858         ENTRY;
 859         LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
 860         lu_object_fini(obj);
 861         OBD_SLAB_FREE_PTR(lov, lov_object_kmem);
 862         EXIT;
 863 }
 864
 865 static int lov_object_print(const struct lu_env *env, void *cookie,
 866                             lu_printer_t p, const struct lu_object *o)
 867 {
 868         return LOV_2DISPATCH_NOLOCK(lu2lov(o), llo_print, env, cookie, p, o);
 869 }
 870
 871 int lov_page_init(const struct lu_env *env, struct cl_object *obj,
 872                   struct cl_page *page, pgoff_t index)
 873 {
 874         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_page_init, env, obj, page,
 875                                     index);
 876 }
 877
 878 /**
 879  * Implements cl_object_operations::clo_io_init() method for lov
 880  * layer. Dispatches to the appropriate layout io initialization method.
 881  */
 882 int lov_io_init(const struct lu_env *env, struct cl_object *obj,
 883                 struct cl_io *io)
 884 {
 885         CL_IO_SLICE_CLEAN(lov_env_io(env), lis_cl);
 886         return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
 887                                      !io->ci_ignore_layout, env, obj, io);
 888 }
 889
 890 /**
 891  * An implementation of cl_object_operations::clo_attr_get() method for lov
 892  * layer. For raid0 layout this collects and merges attributes of all
 893  * sub-objects.
 894  */
 895 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
 896                         struct cl_attr *attr)
 897 {
 898         /* do not take lock, as this function is called under a
 899          * spin-lock. Layout is protected from changing by ongoing IO. */
 900         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
 901 }
 902
 903 static int lov_attr_set(const struct lu_env *env, struct cl_object *obj,
 904                         const struct cl_attr *attr, unsigned valid)
 905 {
 906         /*
 907          * No dispatch is required here, as no layout implements this.
 908          */
 909         return 0;
 910 }
 911
 912 int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
 913                   struct cl_lock *lock, const struct cl_io *io)
 914 {
 915         /* No need to lock because we've taken one refcount of layout.  */
 916         return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_lock_init, env, obj, lock,
 917                                     io);
 918 }
 919
 920 static const struct cl_object_operations lov_ops = {
 921         .coo_page_init = lov_page_init,
 922         .coo_lock_init = lov_lock_init,
 923         .coo_io_init   = lov_io_init,
 924         .coo_attr_get  = lov_attr_get,
 925         .coo_attr_set  = lov_attr_set,
 926         .coo_conf_set  = lov_conf_set
 927 };
 928
 929 static const struct lu_object_operations lov_lu_obj_ops = {
 930         .loo_object_init      = lov_object_init,
 931         .loo_object_delete    = lov_object_delete,
 932         .loo_object_release   = NULL,
 933         .loo_object_free      = lov_object_free,
 934         .loo_object_print     = lov_object_print,
 935         .loo_object_invariant = NULL
 936 };
 937
 938 struct lu_object *lov_object_alloc(const struct lu_env *env,
 939                                    const struct lu_object_header *unused,
 940                                    struct lu_device *dev)
 941 {
 942         struct lov_object *lov;
 943         struct lu_object  *obj;
 944
 945         ENTRY;
 946         OBD_SLAB_ALLOC_PTR_GFP(lov, lov_object_kmem, GFP_NOFS);
 947         if (lov != NULL) {
 948                 obj = lov2lu(lov);
 949                 lu_object_init(obj, NULL, dev);
 950                 lov->lo_cl.co_ops = &lov_ops;
 951                 lov->lo_type = -1; /* invalid, to catch uninitialized type */
 952                 /*
 953                  * object io operation vector (cl_object::co_iop) is installed
 954                  * later in lov_object_init(), as different vectors are used
 955                  * for object with different layouts.
 956                  */
 957                 obj->lo_ops = &lov_lu_obj_ops;
 958         } else
 959                 obj = NULL;
 960         RETURN(obj);
 961 }
 962
 963 struct lov_stripe_md *lov_lsm_addref(struct lov_object *lov)
 964 {
 965         struct lov_stripe_md *lsm = NULL;
 966
 967         lov_conf_freeze(lov);
 968         if (lov->lo_lsm != NULL) {
 969                 lsm = lsm_addref(lov->lo_lsm);
 970                 CDEBUG(D_INODE, "lsm %p addref %d/%d by %p.\n",
 971                         lsm, atomic_read(&lsm->lsm_refc),
 972                         lov->lo_layout_invalid, current);
 973         }
 974         lov_conf_thaw(lov);
 975         return lsm;
 976 }
 977
 978 struct lov_stripe_md *lov_lsm_get(struct cl_object *clobj)
 979 {
 980         struct lu_object *luobj;
 981         struct lov_stripe_md *lsm = NULL;
 982
 983         if (clobj == NULL)
 984                 return NULL;
 985
 986         luobj = lu_object_locate(&cl_object_header(clobj)->coh_lu,
 987                                  &lov_device_type);
 988         if (luobj != NULL)
 989                 lsm = lov_lsm_addref(lu2lov(luobj));
 990         return lsm;
 991 }
 992 EXPORT_SYMBOL(lov_lsm_get);
 993
 994 void lov_lsm_put(struct cl_object *unused, struct lov_stripe_md *lsm)
 995 {
 996         if (lsm != NULL)
 997                 lov_free_memmd(&lsm);
 998 }
 999 EXPORT_SYMBOL(lov_lsm_put);
1000
1001 int lov_read_and_clear_async_rc(struct cl_object *clob)
1002 {
1003         struct lu_object *luobj;
1004         int rc = 0;
1005         ENTRY;
1006
1007         luobj = lu_object_locate(&cl_object_header(clob)->coh_lu,
1008                                  &lov_device_type);
1009         if (luobj != NULL) {
1010                 struct lov_object *lov = lu2lov(luobj);
1011
1012                 lov_conf_freeze(lov);
1013                 switch (lov->lo_type) {
1014                 case LLT_RAID0: {
1015                         struct lov_stripe_md *lsm;
1016                         int i;
1017
1018                         lsm = lov->lo_lsm;
1019                         LASSERT(lsm != NULL);
1020                         for (i = 0; i < lsm->lsm_stripe_count; i++) {
1021                                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
1022
1023                                 if (lov_oinfo_is_dummy(loi))
1024                                         continue;
1025
1026                                 if (loi->loi_ar.ar_rc && !rc)
1027                                         rc = loi->loi_ar.ar_rc;
1028                                 loi->loi_ar.ar_rc = 0;
1029                         }
1030                 }
1031                 case LLT_RELEASED:
1032                 case LLT_EMPTY:
1033                         break;
1034                 default:
1035                         LBUG();
1036                 }
1037                 lov_conf_thaw(lov);
1038         }
1039         RETURN(rc);
1040 }
1041 EXPORT_SYMBOL(lov_read_and_clear_async_rc);
1042
1043 /** @} lov */