Pileus Git - ~andy/linux/blob - drivers/staging/lustre/lustre/lov/lov_object.c

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