lustre/osc/osc_cache.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) 2012, 2013, Intel Corporation.
  31  *
  32  */
  33 /*
  34  * This file is part of Lustre, http://www.lustre.org/
  35  * Lustre is a trademark of Sun Microsystems, Inc.
  36  *
  37  * osc cache management.
  38  *
  39  * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_OSC
  43
  44 #include "osc_cl_internal.h"
  45 #include "osc_internal.h"
  46
  47 static int extent_debug; /* set it to be true for more debug */
  48
  49 static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
  50 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
  51                            int state);
  52 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
  53                               struct osc_async_page *oap, int sent, int rc);
  54 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
  55                           int cmd);
  56 static int osc_refresh_count(const struct lu_env *env,
  57                              struct osc_async_page *oap, int cmd);
  58 static int osc_io_unplug_async(const struct lu_env *env,
  59                                struct client_obd *cli, struct osc_object *osc);
  60 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
  61                            unsigned int lost_grant);
  62
  63 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
  64                                   const char *func, int line);
  65 #define osc_extent_tree_dump(lvl, obj) \
  66         osc_extent_tree_dump0(lvl, obj, __func__, __LINE__)
  67
  68 /** \addtogroup osc
  69  *  @{
  70  */
  71
  72 /* ------------------ osc extent ------------------ */
  73 static inline char *ext_flags(struct osc_extent *ext, char *flags)
  74 {
  75         char *buf = flags;
  76         *buf++ = ext->oe_rw ? 'r' : 'w';
  77         if (ext->oe_intree)
  78                 *buf++ = 'i';
  79         if (ext->oe_srvlock)
  80                 *buf++ = 's';
  81         if (ext->oe_hp)
  82                 *buf++ = 'h';
  83         if (ext->oe_urgent)
  84                 *buf++ = 'u';
  85         if (ext->oe_memalloc)
  86                 *buf++ = 'm';
  87         if (ext->oe_trunc_pending)
  88                 *buf++ = 't';
  89         if (ext->oe_fsync_wait)
  90                 *buf++ = 'Y';
  91         *buf = 0;
  92         return flags;
  93 }
  94
  95 static inline char list_empty_marker(cfs_list_t *list)
  96 {
  97         return cfs_list_empty(list) ? '-' : '+';
  98 }
  99
 100 #define EXTSTR       "[%lu -> %lu/%lu]"
 101 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
 102
 103 #define OSC_EXTENT_DUMP(lvl, extent, fmt, ...) do {                           \
 104         struct osc_extent *__ext = (extent);                                  \
 105         const char *__str[] = OES_STRINGS;                                    \
 106         char __buf[16];                                                       \
 107                                                                               \
 108         CDEBUG(lvl,                                                           \
 109                 "extent %p@{" EXTSTR ", "                                     \
 110                 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt,          \
 111                 /* ----- extent part 0 ----- */                               \
 112                 __ext, EXTPARA(__ext),                                        \
 113                 /* ----- part 1 ----- */                                      \
 114                 cfs_atomic_read(&__ext->oe_refc),                             \
 115                 cfs_atomic_read(&__ext->oe_users),                            \
 116                 list_empty_marker(&__ext->oe_link),                           \
 117                 __str[__ext->oe_state], ext_flags(__ext, __buf),              \
 118                 __ext->oe_obj,                                                \
 119                 /* ----- part 2 ----- */                                      \
 120                 __ext->oe_grants, __ext->oe_nr_pages,                         \
 121                 list_empty_marker(&__ext->oe_pages),                          \
 122                 cfs_waitq_active(&__ext->oe_waitq) ? '+' : '-',               \
 123                 __ext->oe_osclock, __ext->oe_mppr, __ext->oe_owner,           \
 124                 /* ----- part 4 ----- */                                      \
 125                 ## __VA_ARGS__);                                              \
 126 } while (0)
 127
 128 #undef EASSERTF
 129 #define EASSERTF(expr, ext, fmt, args...) do {                          \
 130         if (!(expr)) {                                                  \
 131                 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args);            \
 132                 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj);            \
 133                 LASSERT(expr);                                          \
 134         }                                                                    \
 135 } while (0)
 136
 137 #undef EASSERT
 138 #define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
 139
 140 static inline struct osc_extent *rb_extent(struct rb_node *n)
 141 {
 142         if (n == NULL)
 143                 return NULL;
 144
 145         return container_of(n, struct osc_extent, oe_node);
 146 }
 147
 148 static inline struct osc_extent *next_extent(struct osc_extent *ext)
 149 {
 150         if (ext == NULL)
 151                 return NULL;
 152
 153         LASSERT(ext->oe_intree);
 154         return rb_extent(rb_next(&ext->oe_node));
 155 }
 156
 157 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
 158 {
 159         if (ext == NULL)
 160                 return NULL;
 161
 162         LASSERT(ext->oe_intree);
 163         return rb_extent(rb_prev(&ext->oe_node));
 164 }
 165
 166 static inline struct osc_extent *first_extent(struct osc_object *obj)
 167 {
 168         return rb_extent(rb_first(&obj->oo_root));
 169 }
 170
 171 /* object must be locked by caller. */
 172 static int osc_extent_sanity_check0(struct osc_extent *ext,
 173                                     const char *func, const int line)
 174 {
 175         struct osc_object *obj = ext->oe_obj;
 176         struct osc_async_page *oap;
 177         int page_count;
 178         int rc = 0;
 179
 180         if (!osc_object_is_locked(obj))
 181                 GOTO(out, rc = 9);
 182
 183         if (ext->oe_state >= OES_STATE_MAX)
 184                 GOTO(out, rc = 10);
 185
 186         if (cfs_atomic_read(&ext->oe_refc) <= 0)
 187                 GOTO(out, rc = 20);
 188
 189         if (cfs_atomic_read(&ext->oe_refc) < cfs_atomic_read(&ext->oe_users))
 190                 GOTO(out, rc = 30);
 191
 192         switch (ext->oe_state) {
 193         case OES_INV:
 194                 if (ext->oe_nr_pages > 0 || !cfs_list_empty(&ext->oe_pages))
 195                         GOTO(out, rc = 35);
 196                 GOTO(out, rc = 0);
 197                 break;
 198         case OES_ACTIVE:
 199                 if (cfs_atomic_read(&ext->oe_users) == 0)
 200                         GOTO(out, rc = 40);
 201                 if (ext->oe_hp)
 202                         GOTO(out, rc = 50);
 203                 if (ext->oe_fsync_wait && !ext->oe_urgent)
 204                         GOTO(out, rc = 55);
 205                 break;
 206         case OES_CACHE:
 207                 if (ext->oe_grants == 0)
 208                         GOTO(out, rc = 60);
 209                 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
 210                         GOTO(out, rc = 65);
 211         default:
 212                 if (cfs_atomic_read(&ext->oe_users) > 0)
 213                         GOTO(out, rc = 70);
 214         }
 215
 216         if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
 217                 GOTO(out, rc = 80);
 218
 219         if (ext->oe_osclock == NULL && ext->oe_grants > 0)
 220                 GOTO(out, rc = 90);
 221
 222         if (ext->oe_osclock) {
 223                 struct cl_lock_descr *descr;
 224                 descr = &ext->oe_osclock->cll_descr;
 225                 if (!(descr->cld_start <= ext->oe_start &&
 226                       descr->cld_end >= ext->oe_max_end))
 227                         GOTO(out, rc = 100);
 228         }
 229
 230         if (ext->oe_nr_pages > ext->oe_mppr)
 231                 GOTO(out, rc = 105);
 232
 233         /* Do not verify page list if extent is in RPC. This is because an
 234          * in-RPC extent is supposed to be exclusively accessible w/o lock. */
 235         if (ext->oe_state > OES_CACHE)
 236                 GOTO(out, rc = 0);
 237
 238         if (!extent_debug)
 239                 GOTO(out, rc = 0);
 240
 241         page_count = 0;
 242         cfs_list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
 243                 pgoff_t index = oap2cl_page(oap)->cp_index;
 244                 ++page_count;
 245                 if (index > ext->oe_end || index < ext->oe_start)
 246                         GOTO(out, rc = 110);
 247         }
 248         if (page_count != ext->oe_nr_pages)
 249                 GOTO(out, rc = 120);
 250
 251 out:
 252         if (rc != 0)
 253                 OSC_EXTENT_DUMP(D_ERROR, ext,
 254                                 "%s:%d sanity check %p failed with rc = %d\n",
 255                                 func, line, ext, rc);
 256         return rc;
 257 }
 258
 259 #define sanity_check_nolock(ext) \
 260         osc_extent_sanity_check0(ext, __func__, __LINE__)
 261
 262 #define sanity_check(ext) ({                                                   \
 263         int __res;                                                             \
 264         osc_object_lock((ext)->oe_obj);                                        \
 265         __res = sanity_check_nolock(ext);                                      \
 266         osc_object_unlock((ext)->oe_obj);                                      \
 267         __res;                                                                 \
 268 })
 269
 270
 271 /**
 272  * sanity check - to make sure there is no overlapped extent in the tree.
 273  */
 274 static int osc_extent_is_overlapped(struct osc_object *obj,
 275                                     struct osc_extent *ext)
 276 {
 277         struct osc_extent *tmp;
 278
 279         LASSERT(osc_object_is_locked(obj));
 280
 281         if (!extent_debug)
 282                 return 0;
 283
 284         for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
 285                 if (tmp == ext)
 286                         continue;
 287                 if (tmp->oe_end >= ext->oe_start &&
 288                     tmp->oe_start <= ext->oe_end)
 289                         return 1;
 290         }
 291         return 0;
 292 }
 293
 294 static void osc_extent_state_set(struct osc_extent *ext, int state)
 295 {
 296         LASSERT(osc_object_is_locked(ext->oe_obj));
 297         LASSERT(state >= OES_INV && state < OES_STATE_MAX);
 298
 299         /* Never try to sanity check a state changing extent :-) */
 300         /* LASSERT(sanity_check_nolock(ext) == 0); */
 301
 302         /* TODO: validate the state machine */
 303         ext->oe_state = state;
 304         cfs_waitq_broadcast(&ext->oe_waitq);
 305 }
 306
 307 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
 308 {
 309         struct osc_extent *ext;
 310
 311         OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, CFS_ALLOC_STD);
 312         if (ext == NULL)
 313                 return NULL;
 314
 315         RB_CLEAR_NODE(&ext->oe_node);
 316         ext->oe_obj = obj;
 317         cfs_atomic_set(&ext->oe_refc, 1);
 318         cfs_atomic_set(&ext->oe_users, 0);
 319         CFS_INIT_LIST_HEAD(&ext->oe_link);
 320         ext->oe_state = OES_INV;
 321         CFS_INIT_LIST_HEAD(&ext->oe_pages);
 322         cfs_waitq_init(&ext->oe_waitq);
 323         ext->oe_osclock = NULL;
 324
 325         return ext;
 326 }
 327
 328 static void osc_extent_free(struct osc_extent *ext)
 329 {
 330         OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
 331 }
 332
 333 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
 334 {
 335         LASSERT(cfs_atomic_read(&ext->oe_refc) >= 0);
 336         cfs_atomic_inc(&ext->oe_refc);
 337         return ext;
 338 }
 339
 340 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
 341 {
 342         LASSERT(cfs_atomic_read(&ext->oe_refc) > 0);
 343         if (cfs_atomic_dec_and_test(&ext->oe_refc)) {
 344                 LASSERT(cfs_list_empty(&ext->oe_link));
 345                 LASSERT(cfs_atomic_read(&ext->oe_users) == 0);
 346                 LASSERT(ext->oe_state == OES_INV);
 347                 LASSERT(!ext->oe_intree);
 348
 349                 if (ext->oe_osclock) {
 350                         cl_lock_put(env, ext->oe_osclock);
 351                         ext->oe_osclock = NULL;
 352                 }
 353                 osc_extent_free(ext);
 354         }
 355 }
 356
 357 /**
 358  * osc_extent_put_trust() is a special version of osc_extent_put() when
 359  * it's known that the caller is not the last user. This is to address the
 360  * problem of lacking of lu_env ;-).
 361  */
 362 static void osc_extent_put_trust(struct osc_extent *ext)
 363 {
 364         LASSERT(cfs_atomic_read(&ext->oe_refc) > 1);
 365         LASSERT(osc_object_is_locked(ext->oe_obj));
 366         cfs_atomic_dec(&ext->oe_refc);
 367 }
 368
 369 /**
 370  * Return the extent which includes pgoff @index, or return the greatest
 371  * previous extent in the tree.
 372  */
 373 static struct osc_extent *osc_extent_search(struct osc_object *obj,
 374                                             pgoff_t index)
 375 {
 376         struct rb_node    *n = obj->oo_root.rb_node;
 377         struct osc_extent *tmp, *p = NULL;
 378
 379         LASSERT(osc_object_is_locked(obj));
 380         while (n != NULL) {
 381                 tmp = rb_extent(n);
 382                 if (index < tmp->oe_start) {
 383                         n = n->rb_left;
 384                 } else if (index > tmp->oe_end) {
 385                         p = rb_extent(n);
 386                         n = n->rb_right;
 387                 } else {
 388                         return tmp;
 389                 }
 390         }
 391         return p;
 392 }
 393
 394 /*
 395  * Return the extent covering @index, otherwise return NULL.
 396  * caller must have held object lock.
 397  */
 398 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
 399                                             pgoff_t index)
 400 {
 401         struct osc_extent *ext;
 402
 403         ext = osc_extent_search(obj, index);
 404         if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
 405                 return osc_extent_get(ext);
 406         return NULL;
 407 }
 408
 409 /* caller must have held object lock. */
 410 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
 411 {
 412         struct rb_node   **n      = &obj->oo_root.rb_node;
 413         struct rb_node    *parent = NULL;
 414         struct osc_extent *tmp;
 415
 416         LASSERT(ext->oe_intree == 0);
 417         LASSERT(ext->oe_obj == obj);
 418         LASSERT(osc_object_is_locked(obj));
 419         while (*n != NULL) {
 420                 tmp = rb_extent(*n);
 421                 parent = *n;
 422
 423                 if (ext->oe_end < tmp->oe_start)
 424                         n = &(*n)->rb_left;
 425                 else if (ext->oe_start > tmp->oe_end)
 426                         n = &(*n)->rb_right;
 427                 else
 428                         EASSERTF(0, tmp, EXTSTR, EXTPARA(ext));
 429         }
 430         rb_link_node(&ext->oe_node, parent, n);
 431         rb_insert_color(&ext->oe_node, &obj->oo_root);
 432         osc_extent_get(ext);
 433         ext->oe_intree = 1;
 434 }
 435
 436 /* caller must have held object lock. */
 437 static void osc_extent_erase(struct osc_extent *ext)
 438 {
 439         struct osc_object *obj = ext->oe_obj;
 440         LASSERT(osc_object_is_locked(obj));
 441         if (ext->oe_intree) {
 442                 rb_erase(&ext->oe_node, &obj->oo_root);
 443                 ext->oe_intree = 0;
 444                 /* rbtree held a refcount */
 445                 osc_extent_put_trust(ext);
 446         }
 447 }
 448
 449 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
 450 {
 451         struct osc_object *obj = ext->oe_obj;
 452
 453         LASSERT(osc_object_is_locked(obj));
 454         LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
 455         if (ext->oe_state == OES_CACHE) {
 456                 osc_extent_state_set(ext, OES_ACTIVE);
 457                 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
 458         }
 459         cfs_atomic_inc(&ext->oe_users);
 460         cfs_list_del_init(&ext->oe_link);
 461         return osc_extent_get(ext);
 462 }
 463
 464 static void __osc_extent_remove(struct osc_extent *ext)
 465 {
 466         LASSERT(osc_object_is_locked(ext->oe_obj));
 467         LASSERT(cfs_list_empty(&ext->oe_pages));
 468         osc_extent_erase(ext);
 469         cfs_list_del_init(&ext->oe_link);
 470         osc_extent_state_set(ext, OES_INV);
 471         OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
 472 }
 473
 474 static void osc_extent_remove(struct osc_extent *ext)
 475 {
 476         struct osc_object *obj = ext->oe_obj;
 477
 478         osc_object_lock(obj);
 479         __osc_extent_remove(ext);
 480         osc_object_unlock(obj);
 481 }
 482
 483 /**
 484  * This function is used to merge extents to get better performance. It checks
 485  * if @cur and @victim are contiguous at chunk level.
 486  */
 487 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
 488                             struct osc_extent *victim)
 489 {
 490         struct osc_object *obj = cur->oe_obj;
 491         pgoff_t chunk_start;
 492         pgoff_t chunk_end;
 493         int ppc_bits;
 494
 495         LASSERT(cur->oe_state == OES_CACHE);
 496         LASSERT(osc_object_is_locked(obj));
 497         if (victim == NULL)
 498                 return -EINVAL;
 499
 500         if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
 501                 return -EBUSY;
 502
 503         if (cur->oe_max_end != victim->oe_max_end)
 504                 return -ERANGE;
 505
 506         LASSERT(cur->oe_osclock == victim->oe_osclock);
 507         ppc_bits = osc_cli(obj)->cl_chunkbits - CFS_PAGE_SHIFT;
 508         chunk_start = cur->oe_start >> ppc_bits;
 509         chunk_end   = cur->oe_end   >> ppc_bits;
 510         if (chunk_start   != (victim->oe_end >> ppc_bits) + 1 &&
 511             chunk_end + 1 != victim->oe_start >> ppc_bits)
 512                 return -ERANGE;
 513
 514         OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
 515
 516         cur->oe_start     = min(cur->oe_start, victim->oe_start);
 517         cur->oe_end       = max(cur->oe_end,   victim->oe_end);
 518         cur->oe_grants   += victim->oe_grants;
 519         cur->oe_nr_pages += victim->oe_nr_pages;
 520         /* only the following bits are needed to merge */
 521         cur->oe_urgent   |= victim->oe_urgent;
 522         cur->oe_memalloc |= victim->oe_memalloc;
 523         cfs_list_splice_init(&victim->oe_pages, &cur->oe_pages);
 524         cfs_list_del_init(&victim->oe_link);
 525         victim->oe_nr_pages = 0;
 526
 527         osc_extent_get(victim);
 528         __osc_extent_remove(victim);
 529         osc_extent_put(env, victim);
 530
 531         OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
 532         return 0;
 533 }
 534
 535 /**
 536  * Drop user count of osc_extent, and unplug IO asynchronously.
 537  */
 538 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
 539 {
 540         struct osc_object *obj = ext->oe_obj;
 541         int rc = 0;
 542         ENTRY;
 543
 544         LASSERT(cfs_atomic_read(&ext->oe_users) > 0);
 545         LASSERT(sanity_check(ext) == 0);
 546         LASSERT(ext->oe_grants > 0);
 547
 548         if (cfs_atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
 549                 LASSERT(ext->oe_state == OES_ACTIVE);
 550                 if (ext->oe_trunc_pending) {
 551                         /* a truncate process is waiting for this extent.
 552                          * This may happen due to a race, check
 553                          * osc_cache_truncate_start(). */
 554                         osc_extent_state_set(ext, OES_TRUNC);
 555                         ext->oe_trunc_pending = 0;
 556                 } else {
 557                         osc_extent_state_set(ext, OES_CACHE);
 558                         osc_update_pending(obj, OBD_BRW_WRITE,
 559                                            ext->oe_nr_pages);
 560
 561                         /* try to merge the previous and next extent. */
 562                         osc_extent_merge(env, ext, prev_extent(ext));
 563                         osc_extent_merge(env, ext, next_extent(ext));
 564
 565                         if (ext->oe_urgent)
 566                                 cfs_list_move_tail(&ext->oe_link,
 567                                                    &obj->oo_urgent_exts);
 568                 }
 569                 osc_object_unlock(obj);
 570
 571                 osc_io_unplug_async(env, osc_cli(obj), obj);
 572         }
 573         osc_extent_put(env, ext);
 574         RETURN(rc);
 575 }
 576
 577 static inline int overlapped(struct osc_extent *ex1, struct osc_extent *ex2)
 578 {
 579         return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
 580 }
 581
 582 /**
 583  * Find or create an extent which includes @index, core function to manage
 584  * extent tree.
 585  */
 586 struct osc_extent *osc_extent_find(const struct lu_env *env,
 587                                    struct osc_object *obj, pgoff_t index,
 588                                    int *grants)
 589
 590 {
 591         struct client_obd *cli = osc_cli(obj);
 592         struct cl_lock    *lock;
 593         struct osc_extent *cur;
 594         struct osc_extent *ext;
 595         struct osc_extent *conflict = NULL;
 596         struct osc_extent *found = NULL;
 597         pgoff_t    chunk;
 598         pgoff_t    max_end;
 599         int        max_pages; /* max_pages_per_rpc */
 600         int        chunksize;
 601         int        ppc_bits; /* pages per chunk bits */
 602         int        chunk_mask;
 603         int        rc;
 604         ENTRY;
 605
 606         cur = osc_extent_alloc(obj);
 607         if (cur == NULL)
 608                 RETURN(ERR_PTR(-ENOMEM));
 609
 610         lock = cl_lock_at_pgoff(env, osc2cl(obj), index, NULL, 1, 0);
 611         LASSERT(lock != NULL);
 612         LASSERT(lock->cll_descr.cld_mode >= CLM_WRITE);
 613
 614         LASSERT(cli->cl_chunkbits >= CFS_PAGE_SHIFT);
 615         ppc_bits   = cli->cl_chunkbits - CFS_PAGE_SHIFT;
 616         chunk_mask = ~((1 << ppc_bits) - 1);
 617         chunksize  = 1 << cli->cl_chunkbits;
 618         chunk      = index >> ppc_bits;
 619
 620         /* align end to rpc edge, rpc size may not be a power 2 integer. */
 621         max_pages = cli->cl_max_pages_per_rpc;
 622         LASSERT((max_pages & ~chunk_mask) == 0);
 623         max_end = index - (index % max_pages) + max_pages - 1;
 624         max_end = min_t(pgoff_t, max_end, lock->cll_descr.cld_end);
 625
 626         /* initialize new extent by parameters so far */
 627         cur->oe_max_end = max_end;
 628         cur->oe_start   = index & chunk_mask;
 629         cur->oe_end     = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
 630         if (cur->oe_start < lock->cll_descr.cld_start)
 631                 cur->oe_start = lock->cll_descr.cld_start;
 632         if (cur->oe_end > max_end)
 633                 cur->oe_end = max_end;
 634         cur->oe_osclock = lock;
 635         cur->oe_grants  = 0;
 636         cur->oe_mppr    = max_pages;
 637
 638         /* grants has been allocated by caller */
 639         LASSERTF(*grants >= chunksize + cli->cl_extent_tax,
 640                  "%u/%u/%u.\n", *grants, chunksize, cli->cl_extent_tax);
 641         LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR, EXTPARA(cur));
 642
 643 restart:
 644         osc_object_lock(obj);
 645         ext = osc_extent_search(obj, cur->oe_start);
 646         if (ext == NULL)
 647                 ext = first_extent(obj);
 648         while (ext != NULL) {
 649                 loff_t ext_chk_start = ext->oe_start >> ppc_bits;
 650                 loff_t ext_chk_end   = ext->oe_end   >> ppc_bits;
 651
 652                 LASSERT(sanity_check_nolock(ext) == 0);
 653                 if (chunk > ext_chk_end + 1)
 654                         break;
 655
 656                 /* if covering by different locks, no chance to match */
 657                 if (lock != ext->oe_osclock) {
 658                         EASSERTF(!overlapped(ext, cur), ext,
 659                                  EXTSTR, EXTPARA(cur));
 660
 661                         ext = next_extent(ext);
 662                         continue;
 663                 }
 664
 665                 /* discontiguous chunks? */
 666                 if (chunk + 1 < ext_chk_start) {
 667                         ext = next_extent(ext);
 668                         continue;
 669                 }
 670
 671                 /* ok, from now on, ext and cur have these attrs:
 672                  * 1. covered by the same lock
 673                  * 2. contiguous at chunk level or overlapping. */
 674
 675                 if (overlapped(ext, cur)) {
 676                         /* cur is the minimum unit, so overlapping means
 677                          * full contain. */
 678                         EASSERTF((ext->oe_start <= cur->oe_start &&
 679                                   ext->oe_end >= cur->oe_end),
 680                                  ext, EXTSTR, EXTPARA(cur));
 681
 682                         if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
 683                                 /* for simplicity, we wait for this extent to
 684                                  * finish before going forward. */
 685                                 conflict = osc_extent_get(ext);
 686                                 break;
 687                         }
 688
 689                         found = osc_extent_hold(ext);
 690                         break;
 691                 }
 692
 693                 /* non-overlapped extent */
 694                 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
 695                         /* we can't do anything for a non OES_CACHE extent, or
 696                          * if there is someone waiting for this extent to be
 697                          * flushed, try next one. */
 698                         ext = next_extent(ext);
 699                         continue;
 700                 }
 701
 702                 /* check if they belong to the same rpc slot before trying to
 703                  * merge. the extents are not overlapped and contiguous at
 704                  * chunk level to get here. */
 705                 if (ext->oe_max_end != max_end) {
 706                         /* if they don't belong to the same RPC slot or
 707                          * max_pages_per_rpc has ever changed, do not merge. */
 708                         ext = next_extent(ext);
 709                         continue;
 710                 }
 711
 712                 /* it's required that an extent must be contiguous at chunk
 713                  * level so that we know the whole extent is covered by grant
 714                  * (the pages in the extent are NOT required to be contiguous).
 715                  * Otherwise, it will be too much difficult to know which
 716                  * chunks have grants allocated. */
 717
 718                 /* try to do front merge - extend ext's start */
 719                 if (chunk + 1 == ext_chk_start) {
 720                         /* ext must be chunk size aligned */
 721                         EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
 722
 723                         /* pull ext's start back to cover cur */
 724                         ext->oe_start   = cur->oe_start;
 725                         ext->oe_grants += chunksize;
 726                         *grants -= chunksize;
 727
 728                         found = osc_extent_hold(ext);
 729                 } else if (chunk == ext_chk_end + 1) {
 730                         /* rear merge */
 731                         ext->oe_end     = cur->oe_end;
 732                         ext->oe_grants += chunksize;
 733                         *grants -= chunksize;
 734
 735                         /* try to merge with the next one because we just fill
 736                          * in a gap */
 737                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 738                                 /* we can save extent tax from next extent */
 739                                 *grants += cli->cl_extent_tax;
 740
 741                         found = osc_extent_hold(ext);
 742                 }
 743                 if (found != NULL)
 744                         break;
 745
 746                 ext = next_extent(ext);
 747         }
 748
 749         osc_extent_tree_dump(D_CACHE, obj);
 750         if (found != NULL) {
 751                 LASSERT(conflict == NULL);
 752                 if (!IS_ERR(found)) {
 753                         LASSERT(found->oe_osclock == cur->oe_osclock);
 754                         OSC_EXTENT_DUMP(D_CACHE, found,
 755                                         "found caching ext for %lu.\n", index);
 756                 }
 757         } else if (conflict == NULL) {
 758                 /* create a new extent */
 759                 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
 760                 cur->oe_grants = chunksize + cli->cl_extent_tax;
 761                 *grants -= cur->oe_grants;
 762                 LASSERT(*grants >= 0);
 763
 764                 cur->oe_state = OES_CACHE;
 765                 found = osc_extent_hold(cur);
 766                 osc_extent_insert(obj, cur);
 767                 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
 768                                 index, lock->cll_descr.cld_end);
 769         }
 770         osc_object_unlock(obj);
 771
 772         if (conflict != NULL) {
 773                 LASSERT(found == NULL);
 774
 775                 /* waiting for IO to finish. Please notice that it's impossible
 776                  * to be an OES_TRUNC extent. */
 777                 rc = osc_extent_wait(env, conflict, OES_INV);
 778                 osc_extent_put(env, conflict);
 779                 conflict = NULL;
 780                 if (rc < 0)
 781                         GOTO(out, found = ERR_PTR(rc));
 782
 783                 goto restart;
 784         }
 785         EXIT;
 786
 787 out:
 788         osc_extent_put(env, cur);
 789         LASSERT(*grants >= 0);
 790         return found;
 791 }
 792
 793 /**
 794  * Called when IO is finished to an extent.
 795  */
 796 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
 797                       int sent, int rc)
 798 {
 799         struct client_obd *cli = osc_cli(ext->oe_obj);
 800         struct osc_async_page *oap;
 801         struct osc_async_page *tmp;
 802         int nr_pages = ext->oe_nr_pages;
 803         int lost_grant = 0;
 804         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
 805         __u64 last_off = 0;
 806         int last_count = -1;
 807         ENTRY;
 808
 809         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
 810
 811         ext->oe_rc = rc ?: ext->oe_nr_pages;
 812         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
 813         cfs_list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 814                                      oap_pending_item) {
 815                 cfs_list_del_init(&oap->oap_rpc_item);
 816                 cfs_list_del_init(&oap->oap_pending_item);
 817                 if (last_off <= oap->oap_obj_off) {
 818                         last_off = oap->oap_obj_off;
 819                         last_count = oap->oap_count;
 820                 }
 821
 822                 --ext->oe_nr_pages;
 823                 osc_ap_completion(env, cli, oap, sent, rc);
 824         }
 825         EASSERT(ext->oe_nr_pages == 0, ext);
 826
 827         if (!sent) {
 828                 lost_grant = ext->oe_grants;
 829         } else if (blocksize < CFS_PAGE_SIZE &&
 830                    last_count != CFS_PAGE_SIZE) {
 831                 /* For short writes we shouldn't count parts of pages that
 832                  * span a whole chunk on the OST side, or our accounting goes
 833                  * wrong.  Should match the code in filter_grant_check. */
 834                 int offset = oap->oap_page_off & ~CFS_PAGE_MASK;
 835                 int count = oap->oap_count + (offset & (blocksize - 1));
 836                 int end = (offset + oap->oap_count) & (blocksize - 1);
 837                 if (end)
 838                         count += blocksize - end;
 839
 840                 lost_grant = CFS_PAGE_SIZE - count;
 841         }
 842         if (ext->oe_grants > 0)
 843                 osc_free_grant(cli, nr_pages, lost_grant);
 844
 845         osc_extent_remove(ext);
 846         /* put the refcount for RPC */
 847         osc_extent_put(env, ext);
 848         RETURN(0);
 849 }
 850
 851 static int extent_wait_cb(struct osc_extent *ext, int state)
 852 {
 853         int ret;
 854
 855         osc_object_lock(ext->oe_obj);
 856         ret = ext->oe_state == state;
 857         osc_object_unlock(ext->oe_obj);
 858
 859         return ret;
 860 }
 861
 862 /**
 863  * Wait for the extent's state to become @state.
 864  */
 865 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
 866                            int state)
 867 {
 868         struct osc_object *obj = ext->oe_obj;
 869         struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
 870                                                   LWI_ON_SIGNAL_NOOP, NULL);
 871         int rc = 0;
 872         ENTRY;
 873
 874         osc_object_lock(obj);
 875         LASSERT(sanity_check_nolock(ext) == 0);
 876         /* `Kick' this extent only if the caller is waiting for it to be
 877          * written out. */
 878         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
 879                 if (ext->oe_state == OES_ACTIVE) {
 880                         ext->oe_urgent = 1;
 881                 } else if (ext->oe_state == OES_CACHE) {
 882                         ext->oe_urgent = 1;
 883                         osc_extent_hold(ext);
 884                         rc = 1;
 885                 }
 886         }
 887         osc_object_unlock(obj);
 888         if (rc == 1)
 889                 osc_extent_release(env, ext);
 890
 891         /* wait for the extent until its state becomes @state */
 892         rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
 893         if (rc == -ETIMEDOUT) {
 894                 OSC_EXTENT_DUMP(D_ERROR, ext,
 895                         "%s: wait ext to %d timedout, recovery in progress?\n",
 896                         osc_export(obj)->exp_obd->obd_name, state);
 897
 898                 lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 899                 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state),
 900                                   &lwi);
 901         }
 902         if (rc == 0 && ext->oe_rc < 0)
 903                 rc = ext->oe_rc;
 904         RETURN(rc);
 905 }
 906
 907 /**
 908  * Discard pages with index greater than @size. If @ext is overlapped with
 909  * @size, then partial truncate happens.
 910  */
 911 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
 912                                 bool partial)
 913 {
 914         struct cl_env_nest     nest;
 915         struct lu_env         *env;
 916         struct cl_io          *io;
 917         struct osc_object     *obj = ext->oe_obj;
 918         struct client_obd     *cli = osc_cli(obj);
 919         struct osc_async_page *oap;
 920         struct osc_async_page *tmp;
 921         int                    pages_in_chunk = 0;
 922         int                    ppc_bits    = cli->cl_chunkbits - CFS_PAGE_SHIFT;
 923         __u64                  trunc_chunk = trunc_index >> ppc_bits;
 924         int                    grants   = 0;
 925         int                    nr_pages = 0;
 926         int                    rc       = 0;
 927         ENTRY;
 928
 929         LASSERT(sanity_check(ext) == 0);
 930         LASSERT(ext->oe_state == OES_TRUNC);
 931         LASSERT(!ext->oe_urgent);
 932
 933         /* Request new lu_env.
 934          * We can't use that env from osc_cache_truncate_start() because
 935          * it's from lov_io_sub and not fully initialized. */
 936         env = cl_env_nested_get(&nest);
 937         io  = &osc_env_info(env)->oti_io;
 938         io->ci_obj = cl_object_top(osc2cl(obj));
 939         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
 940         if (rc < 0)
 941                 GOTO(out, rc);
 942
 943         /* discard all pages with index greater then trunc_index */
 944         cfs_list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 945                                      oap_pending_item) {
 946                 struct cl_page  *sub  = oap2cl_page(oap);
 947                 struct cl_page  *page = cl_page_top(sub);
 948
 949                 LASSERT(cfs_list_empty(&oap->oap_rpc_item));
 950
 951                 /* only discard the pages with their index greater than
 952                  * trunc_index, and ... */
 953                 if (sub->cp_index < trunc_index ||
 954                     (sub->cp_index == trunc_index && partial)) {
 955                         /* accounting how many pages remaining in the chunk
 956                          * so that we can calculate grants correctly. */
 957                         if (sub->cp_index >> ppc_bits == trunc_chunk)
 958                                 ++pages_in_chunk;
 959                         continue;
 960                 }
 961
 962                 cfs_list_del_init(&oap->oap_pending_item);
 963
 964                 cl_page_get(page);
 965                 lu_ref_add(&page->cp_reference, "truncate", cfs_current());
 966
 967                 if (cl_page_own(env, io, page) == 0) {
 968                         cl_page_unmap(env, io, page);
 969                         cl_page_discard(env, io, page);
 970                         cl_page_disown(env, io, page);
 971                 } else {
 972                         LASSERT(page->cp_state == CPS_FREEING);
 973                         LASSERT(0);
 974                 }
 975
 976                 lu_ref_del(&page->cp_reference, "truncate", cfs_current());
 977                 cl_page_put(env, page);
 978
 979                 --ext->oe_nr_pages;
 980                 ++nr_pages;
 981         }
 982         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
 983                       ext->oe_nr_pages == 0),
 984                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
 985
 986         osc_object_lock(obj);
 987         if (ext->oe_nr_pages == 0) {
 988                 LASSERT(pages_in_chunk == 0);
 989                 grants = ext->oe_grants;
 990                 ext->oe_grants = 0;
 991         } else { /* calculate how many grants we can free */
 992                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
 993                 pgoff_t last_index;
 994
 995
 996                 /* if there is no pages in this chunk, we can also free grants
 997                  * for the last chunk */
 998                 if (pages_in_chunk == 0) {
 999                         /* if this is the 1st chunk and no pages in this chunk,
1000                          * ext->oe_nr_pages must be zero, so we should be in
1001                          * the other if-clause. */
1002                         LASSERT(trunc_chunk > 0);
1003                         --trunc_chunk;
1004                         ++chunks;
1005                 }
1006
1007                 /* this is what we can free from this extent */
1008                 grants          = chunks << cli->cl_chunkbits;
1009                 ext->oe_grants -= grants;
1010                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1011                 ext->oe_end     = min(last_index, ext->oe_max_end);
1012                 LASSERT(ext->oe_end >= ext->oe_start);
1013                 LASSERT(ext->oe_grants > 0);
1014         }
1015         osc_object_unlock(obj);
1016
1017         if (grants > 0 || nr_pages > 0)
1018                 osc_free_grant(cli, nr_pages, grants);
1019
1020 out:
1021         cl_io_fini(env, io);
1022         cl_env_nested_put(&nest, env);
1023         RETURN(rc);
1024 }
1025
1026 /**
1027  * This function is used to make the extent prepared for transfer.
1028  * A race with flusing page - ll_writepage() has to be handled cautiously.
1029  */
1030 static int osc_extent_make_ready(const struct lu_env *env,
1031                                  struct osc_extent *ext)
1032 {
1033         struct osc_async_page *oap;
1034         struct osc_async_page *last = NULL;
1035         struct osc_object *obj = ext->oe_obj;
1036         int page_count = 0;
1037         int rc;
1038         ENTRY;
1039
1040         /* we're going to grab page lock, so object lock must not be taken. */
1041         LASSERT(sanity_check(ext) == 0);
1042         /* in locking state, any process should not touch this extent. */
1043         EASSERT(ext->oe_state == OES_LOCKING, ext);
1044         EASSERT(ext->oe_owner != NULL, ext);
1045
1046         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1047
1048         cfs_list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1049                 ++page_count;
1050                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1051                         last = oap;
1052
1053                 /* checking ASYNC_READY is race safe */
1054                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1055                         continue;
1056
1057                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1058                 switch (rc) {
1059                 case 0:
1060                         spin_lock(&oap->oap_lock);
1061                         oap->oap_async_flags |= ASYNC_READY;
1062                         spin_unlock(&oap->oap_lock);
1063                         break;
1064                 case -EALREADY:
1065                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1066                         break;
1067                 default:
1068                         LASSERTF(0, "unknown return code: %d\n", rc);
1069                 }
1070         }
1071
1072         LASSERT(page_count == ext->oe_nr_pages);
1073         LASSERT(last != NULL);
1074         /* the last page is the only one we need to refresh its count by
1075          * the size of file. */
1076         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1077                 last->oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1078                 LASSERT(last->oap_count > 0);
1079                 LASSERT(last->oap_page_off + last->oap_count <= CFS_PAGE_SIZE);
1080                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1081         }
1082
1083         /* for the rest of pages, we don't need to call osf_refresh_count()
1084          * because it's known they are not the last page */
1085         cfs_list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1086                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1087                         oap->oap_count = CFS_PAGE_SIZE - oap->oap_page_off;
1088                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1089                 }
1090         }
1091
1092         osc_object_lock(obj);
1093         osc_extent_state_set(ext, OES_RPC);
1094         osc_object_unlock(obj);
1095         /* get a refcount for RPC. */
1096         osc_extent_get(ext);
1097
1098         RETURN(0);
1099 }
1100
1101 /**
1102  * Quick and simple version of osc_extent_find(). This function is frequently
1103  * called to expand the extent for the same IO. To expand the extent, the
1104  * page index must be in the same or next chunk of ext->oe_end.
1105  */
1106 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index, int *grants)
1107 {
1108         struct osc_object *obj = ext->oe_obj;
1109         struct client_obd *cli = osc_cli(obj);
1110         struct osc_extent *next;
1111         int ppc_bits = cli->cl_chunkbits - CFS_PAGE_SHIFT;
1112         pgoff_t chunk = index >> ppc_bits;
1113         pgoff_t end_chunk;
1114         pgoff_t end_index;
1115         int chunksize = 1 << cli->cl_chunkbits;
1116         int rc = 0;
1117         ENTRY;
1118
1119         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1120         osc_object_lock(obj);
1121         LASSERT(sanity_check_nolock(ext) == 0);
1122         end_chunk = ext->oe_end >> ppc_bits;
1123         if (chunk > end_chunk + 1)
1124                 GOTO(out, rc = -ERANGE);
1125
1126         if (end_chunk >= chunk)
1127                 GOTO(out, rc = 0);
1128
1129         LASSERT(end_chunk + 1 == chunk);
1130         /* try to expand this extent to cover @index */
1131         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1132
1133         next = next_extent(ext);
1134         if (next != NULL && next->oe_start <= end_index)
1135                 /* complex mode - overlapped with the next extent,
1136                  * this case will be handled by osc_extent_find() */
1137                 GOTO(out, rc = -EAGAIN);
1138
1139         ext->oe_end = end_index;
1140         ext->oe_grants += chunksize;
1141         *grants -= chunksize;
1142         LASSERT(*grants >= 0);
1143         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1144                  "overlapped after expanding for %lu.\n", index);
1145         EXIT;
1146
1147 out:
1148         osc_object_unlock(obj);
1149         RETURN(rc);
1150 }
1151
1152 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
1153                                   const char *func, int line)
1154 {
1155         struct osc_extent *ext;
1156         int cnt;
1157
1158         CDEBUG(level, "Dump object %p extents at %s:%d, mppr: %u.\n",
1159                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1160
1161         /* osc_object_lock(obj); */
1162         cnt = 1;
1163         for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1164                 OSC_EXTENT_DUMP(level, ext, "in tree %d.\n", cnt++);
1165
1166         cnt = 1;
1167         cfs_list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1168                 OSC_EXTENT_DUMP(level, ext, "hp %d.\n", cnt++);
1169
1170         cnt = 1;
1171         cfs_list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1172                 OSC_EXTENT_DUMP(level, ext, "urgent %d.\n", cnt++);
1173
1174         cnt = 1;
1175         cfs_list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1176                 OSC_EXTENT_DUMP(level, ext, "reading %d.\n", cnt++);
1177         /* osc_object_unlock(obj); */
1178 }
1179
1180 /* ------------------ osc extent end ------------------ */
1181
1182 static inline int osc_is_ready(struct osc_object *osc)
1183 {
1184         return !cfs_list_empty(&osc->oo_ready_item) ||
1185                !cfs_list_empty(&osc->oo_hp_ready_item);
1186 }
1187
1188 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1189         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1190                (OSC), osc_is_ready(OSC),                                       \
1191                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1192                list_empty_marker(&(OSC)->oo_ready_item),                       \
1193                cfs_atomic_read(&(OSC)->oo_nr_writes),                          \
1194                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1195                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1196                cfs_atomic_read(&(OSC)->oo_nr_reads),                           \
1197                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1198                ##args)
1199
1200 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1201                           int cmd)
1202 {
1203         struct osc_page *opg  = oap2osc_page(oap);
1204         struct cl_page  *page = cl_page_top(oap2cl_page(oap));
1205         int result;
1206
1207         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1208
1209         ENTRY;
1210         result = cl_page_make_ready(env, page, CRT_WRITE);
1211         if (result == 0)
1212                 opg->ops_submit_time = cfs_time_current();
1213         RETURN(result);
1214 }
1215
1216 static int osc_refresh_count(const struct lu_env *env,
1217                              struct osc_async_page *oap, int cmd)
1218 {
1219         struct osc_page  *opg = oap2osc_page(oap);
1220         struct cl_page   *page = oap2cl_page(oap);
1221         struct cl_object *obj;
1222         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
1223
1224         int result;
1225         loff_t kms;
1226
1227         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1228         LASSERT(!(cmd & OBD_BRW_READ));
1229         LASSERT(opg != NULL);
1230         obj = opg->ops_cl.cpl_obj;
1231
1232         cl_object_attr_lock(obj);
1233         result = cl_object_attr_get(env, obj, attr);
1234         cl_object_attr_unlock(obj);
1235         if (result < 0)
1236                 return result;
1237         kms = attr->cat_kms;
1238         if (cl_offset(obj, page->cp_index) >= kms)
1239                 /* catch race with truncate */
1240                 return 0;
1241         else if (cl_offset(obj, page->cp_index + 1) > kms)
1242                 /* catch sub-page write at end of file */
1243                 return kms % CFS_PAGE_SIZE;
1244         else
1245                 return CFS_PAGE_SIZE;
1246 }
1247
1248 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1249                           int cmd, int rc)
1250 {
1251         struct osc_page   *opg  = oap2osc_page(oap);
1252         struct cl_page    *page = cl_page_top(oap2cl_page(oap));
1253         struct osc_object *obj  = cl2osc(opg->ops_cl.cpl_obj);
1254         enum cl_req_type   crt;
1255         int srvlock;
1256
1257         ENTRY;
1258
1259         cmd &= ~OBD_BRW_NOQUOTA;
1260         LASSERT(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ));
1261         LASSERT(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE));
1262         LASSERT(opg->ops_transfer_pinned);
1263
1264         /*
1265          * page->cp_req can be NULL if io submission failed before
1266          * cl_req was allocated.
1267          */
1268         if (page->cp_req != NULL)
1269                 cl_req_page_done(env, page);
1270         LASSERT(page->cp_req == NULL);
1271
1272         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1273         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1274         opg->ops_transfer_pinned = 0;
1275
1276         spin_lock(&obj->oo_seatbelt);
1277         LASSERT(opg->ops_submitter != NULL);
1278         LASSERT(!cfs_list_empty(&opg->ops_inflight));
1279         cfs_list_del_init(&opg->ops_inflight);
1280         opg->ops_submitter = NULL;
1281         spin_unlock(&obj->oo_seatbelt);
1282
1283         opg->ops_submit_time = 0;
1284         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1285
1286         /* statistic */
1287         if (rc == 0 && srvlock) {
1288                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1289                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1290                 int bytes = oap->oap_count;
1291
1292                 if (crt == CRT_READ)
1293                         stats->os_lockless_reads += bytes;
1294                 else
1295                         stats->os_lockless_writes += bytes;
1296         }
1297
1298         /*
1299          * This has to be the last operation with the page, as locks are
1300          * released in cl_page_completion() and nothing except for the
1301          * reference counter protects page from concurrent reclaim.
1302          */
1303         lu_ref_del(&page->cp_reference, "transfer", page);
1304
1305         cl_page_completion(env, page, crt, rc);
1306
1307         RETURN(0);
1308 }
1309
1310 #define OSC_DUMP_GRANT(cli, fmt, args...) do {                                \
1311         struct client_obd *__tmp = (cli);                                     \
1312         CDEBUG(D_CACHE, "%s: { dirty: %ld/%ld dirty_pages: %d/%d "            \
1313                "unstable_pages: %d/%d dropped: %ld avail: %ld, "              \
1314                "reserved: %ld, flight: %d } " fmt,                            \
1315                __tmp->cl_import->imp_obd->obd_name,                           \
1316                __tmp->cl_dirty, __tmp->cl_dirty_max,                          \
1317                cfs_atomic_read(&obd_dirty_pages), obd_max_dirty_pages,        \
1318                cfs_atomic_read(&obd_unstable_pages), obd_max_dirty_pages,     \
1319                __tmp->cl_lost_grant, __tmp->cl_avail_grant,                   \
1320                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight, ##args);      \
1321 } while (0)
1322
1323 /* caller must hold loi_list_lock */
1324 static void osc_consume_write_grant(struct client_obd *cli,
1325                                     struct brw_page *pga)
1326 {
1327         LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
1328         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1329         cfs_atomic_inc(&obd_dirty_pages);
1330         cli->cl_dirty += CFS_PAGE_SIZE;
1331         pga->flag |= OBD_BRW_FROM_GRANT;
1332         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1333                CFS_PAGE_SIZE, pga, pga->pg);
1334         osc_update_next_shrink(cli);
1335 }
1336
1337 /* the companion to osc_consume_write_grant, called when a brw has completed.
1338  * must be called with the loi lock held. */
1339 static void osc_release_write_grant(struct client_obd *cli,
1340                                     struct brw_page *pga)
1341 {
1342         ENTRY;
1343
1344         LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
1345         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1346                 EXIT;
1347                 return;
1348         }
1349
1350         pga->flag &= ~OBD_BRW_FROM_GRANT;
1351         cfs_atomic_dec(&obd_dirty_pages);
1352         cli->cl_dirty -= CFS_PAGE_SIZE;
1353         if (pga->flag & OBD_BRW_NOCACHE) {
1354                 pga->flag &= ~OBD_BRW_NOCACHE;
1355                 cfs_atomic_dec(&obd_dirty_transit_pages);
1356                 cli->cl_dirty_transit -= CFS_PAGE_SIZE;
1357         }
1358         EXIT;
1359 }
1360
1361 /**
1362  * To avoid sleeping with object lock held, it's good for us allocate enough
1363  * grants before entering into critical section.
1364  *
1365  * client_obd_list_lock held by caller
1366  */
1367 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1368 {
1369         int rc = -EDQUOT;
1370
1371         if (cli->cl_avail_grant >= bytes) {
1372                 cli->cl_avail_grant    -= bytes;
1373                 cli->cl_reserved_grant += bytes;
1374                 rc = 0;
1375         }
1376         return rc;
1377 }
1378
1379 static void __osc_unreserve_grant(struct client_obd *cli,
1380                                   unsigned int reserved, unsigned int unused)
1381 {
1382         /* it's quite normal for us to get more grant than reserved.
1383          * Thinking about a case that two extents merged by adding a new
1384          * chunk, we can save one extent tax. If extent tax is greater than
1385          * one chunk, we can save more grant by adding a new chunk */
1386         cli->cl_reserved_grant -= reserved;
1387         if (unused > reserved) {
1388                 cli->cl_avail_grant += reserved;
1389                 cli->cl_lost_grant  += unused - reserved;
1390         } else {
1391                 cli->cl_avail_grant += unused;
1392         }
1393 }
1394
1395 void osc_unreserve_grant(struct client_obd *cli,
1396                          unsigned int reserved, unsigned int unused)
1397 {
1398         client_obd_list_lock(&cli->cl_loi_list_lock);
1399         __osc_unreserve_grant(cli, reserved, unused);
1400         if (unused > 0)
1401                 osc_wake_cache_waiters(cli);
1402         client_obd_list_unlock(&cli->cl_loi_list_lock);
1403 }
1404
1405 /**
1406  * Free grant after IO is finished or canceled.
1407  *
1408  * @lost_grant is used to remember how many grants we have allocated but not
1409  * used, we should return these grants to OST. There're two cases where grants
1410  * can be lost:
1411  * 1. truncate;
1412  * 2. blocksize at OST is less than CFS_PAGE_SIZE and a partial page was
1413  *    written. In this case OST may use less chunks to serve this partial
1414  *    write. OSTs don't actually know the page size on the client side. so
1415  *    clients have to calculate lost grant by the blocksize on the OST.
1416  *    See filter_grant_check() for details.
1417  */
1418 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1419                            unsigned int lost_grant)
1420 {
1421         int grant = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
1422
1423         client_obd_list_lock(&cli->cl_loi_list_lock);
1424         cfs_atomic_sub(nr_pages, &obd_dirty_pages);
1425         cli->cl_dirty -= nr_pages << CFS_PAGE_SHIFT;
1426         cli->cl_lost_grant += lost_grant;
1427         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1428                 /* borrow some grant from truncate to avoid the case that
1429                  * truncate uses up all avail grant */
1430                 cli->cl_lost_grant -= grant;
1431                 cli->cl_avail_grant += grant;
1432         }
1433         osc_wake_cache_waiters(cli);
1434         client_obd_list_unlock(&cli->cl_loi_list_lock);
1435         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu\n",
1436                lost_grant, cli->cl_lost_grant,
1437                cli->cl_avail_grant, cli->cl_dirty);
1438 }
1439
1440 /**
1441  * The companion to osc_enter_cache(), called when @oap is no longer part of
1442  * the dirty accounting due to error.
1443  */
1444 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1445 {
1446         client_obd_list_lock(&cli->cl_loi_list_lock);
1447         osc_release_write_grant(cli, &oap->oap_brw_page);
1448         client_obd_list_unlock(&cli->cl_loi_list_lock);
1449 }
1450
1451 /**
1452  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1453  * is available.
1454  */
1455 static int osc_enter_cache_try(struct client_obd *cli,
1456                                struct osc_async_page *oap,
1457                                int bytes, int transient)
1458 {
1459         int rc;
1460
1461         OSC_DUMP_GRANT(cli, "need:%d.\n", bytes);
1462
1463         rc = osc_reserve_grant(cli, bytes);
1464         if (rc < 0)
1465                 return 0;
1466
1467         if (cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max &&
1468             cfs_atomic_read(&obd_unstable_pages) + 1 +
1469             cfs_atomic_read(&obd_dirty_pages) <= obd_max_dirty_pages) {
1470                 osc_consume_write_grant(cli, &oap->oap_brw_page);
1471                 if (transient) {
1472                         cli->cl_dirty_transit += CFS_PAGE_SIZE;
1473                         cfs_atomic_inc(&obd_dirty_transit_pages);
1474                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
1475                 }
1476                 rc = 1;
1477         } else {
1478                 __osc_unreserve_grant(cli, bytes, bytes);
1479                 rc = 0;
1480         }
1481         return rc;
1482 }
1483
1484 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1485 {
1486         int rc;
1487         client_obd_list_lock(&cli->cl_loi_list_lock);
1488         rc = cfs_list_empty(&ocw->ocw_entry);
1489         client_obd_list_unlock(&cli->cl_loi_list_lock);
1490         return rc;
1491 }
1492
1493 /**
1494  * The main entry to reserve dirty page accounting. Usually the grant reserved
1495  * in this function will be freed in bulk in osc_free_grant() unless it fails
1496  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1497  *
1498  * The process will be put into sleep if it's already run out of grant.
1499  */
1500 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1501                            struct osc_async_page *oap, int bytes)
1502 {
1503         struct osc_object *osc = oap->oap_obj;
1504         struct lov_oinfo  *loi = osc->oo_oinfo;
1505         struct osc_cache_waiter ocw;
1506         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
1507         int rc = -EDQUOT;
1508         ENTRY;
1509
1510         OSC_DUMP_GRANT(cli, "need:%d.\n", bytes);
1511
1512         client_obd_list_lock(&cli->cl_loi_list_lock);
1513
1514         /* force the caller to try sync io.  this can jump the list
1515          * of queued writes and create a discontiguous rpc stream */
1516         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1517             cli->cl_dirty_max < CFS_PAGE_SIZE     ||
1518             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync)
1519                 GOTO(out, rc = -EDQUOT);
1520
1521         /* Hopefully normal case - cache space and write credits available */
1522         if (osc_enter_cache_try(cli, oap, bytes, 0))
1523                 GOTO(out, rc = 0);
1524
1525         /* We can get here for two reasons: too many dirty pages in cache, or
1526          * run out of grants. In both cases we should write dirty pages out.
1527          * Adding a cache waiter will trigger urgent write-out no matter what
1528          * RPC size will be.
1529          * The exiting condition is no avail grants and no dirty pages caching,
1530          * that really means there is no space on the OST. */
1531         cfs_waitq_init(&ocw.ocw_waitq);
1532         ocw.ocw_oap   = oap;
1533         ocw.ocw_grant = bytes;
1534         while (cli->cl_dirty > 0 || cli->cl_w_in_flight > 0) {
1535                 cfs_list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1536                 ocw.ocw_rc = 0;
1537                 client_obd_list_unlock(&cli->cl_loi_list_lock);
1538
1539                 osc_io_unplug_async(env, cli, NULL);
1540
1541                 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1542                        cli->cl_import->imp_obd->obd_name, &ocw, oap);
1543
1544                 rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1545
1546                 client_obd_list_lock(&cli->cl_loi_list_lock);
1547
1548                 /* l_wait_event is interrupted by signal */
1549                 if (rc < 0) {
1550                         cfs_list_del_init(&ocw.ocw_entry);
1551                         GOTO(out, rc);
1552                 }
1553
1554                 LASSERT(cfs_list_empty(&ocw.ocw_entry));
1555                 rc = ocw.ocw_rc;
1556
1557                 if (rc != -EDQUOT)
1558                         GOTO(out, rc);
1559                 if (osc_enter_cache_try(cli, oap, bytes, 0))
1560                         GOTO(out, rc = 0);
1561         }
1562         EXIT;
1563 out:
1564         client_obd_list_unlock(&cli->cl_loi_list_lock);
1565         OSC_DUMP_GRANT(cli, "returned %d.\n", rc);
1566         RETURN(rc);
1567 }
1568
1569 /* caller must hold loi_list_lock */
1570 void osc_wake_cache_waiters(struct client_obd *cli)
1571 {
1572         cfs_list_t *l, *tmp;
1573         struct osc_cache_waiter *ocw;
1574
1575         ENTRY;
1576         cfs_list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1577                 ocw = cfs_list_entry(l, struct osc_cache_waiter, ocw_entry);
1578                 cfs_list_del_init(&ocw->ocw_entry);
1579
1580                 ocw->ocw_rc = -EDQUOT;
1581                 /* we can't dirty more */
1582                 if (cli->cl_dirty + CFS_PAGE_SIZE > cli->cl_dirty_max ||
1583                     cfs_atomic_read(&obd_unstable_pages) + 1 +
1584                     cfs_atomic_read(&obd_dirty_pages) > obd_max_dirty_pages) {
1585                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
1586                                "osc max %ld, sys max %d\n", cli->cl_dirty,
1587                                cli->cl_dirty_max, obd_max_dirty_pages);
1588                         goto wakeup;
1589                 }
1590
1591                 ocw->ocw_rc = 0;
1592                 if (!osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
1593                         ocw->ocw_rc = -EDQUOT;
1594
1595 wakeup:
1596                 CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld, %d\n",
1597                        ocw, ocw->ocw_oap, cli->cl_avail_grant, ocw->ocw_rc);
1598
1599                 cfs_waitq_signal(&ocw->ocw_waitq);
1600         }
1601
1602         EXIT;
1603 }
1604
1605 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1606 {
1607         int hprpc = !!cfs_list_empty(&osc->oo_hp_exts);
1608         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1609 }
1610
1611 /* This maintains the lists of pending pages to read/write for a given object
1612  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1613  * to quickly find objects that are ready to send an RPC. */
1614 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1615                          int cmd)
1616 {
1617         int invalid_import = 0;
1618         ENTRY;
1619
1620         /* if we have an invalid import we want to drain the queued pages
1621          * by forcing them through rpcs that immediately fail and complete
1622          * the pages.  recovery relies on this to empty the queued pages
1623          * before canceling the locks and evicting down the llite pages */
1624         if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1625                 invalid_import = 1;
1626
1627         if (cmd & OBD_BRW_WRITE) {
1628                 if (cfs_atomic_read(&osc->oo_nr_writes) == 0)
1629                         RETURN(0);
1630                 if (invalid_import) {
1631                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1632                         RETURN(1);
1633                 }
1634                 if (!cfs_list_empty(&osc->oo_hp_exts)) {
1635                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1636                         RETURN(1);
1637                 }
1638                 if (!cfs_list_empty(&osc->oo_urgent_exts)) {
1639                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1640                         RETURN(1);
1641                 }
1642                 /* trigger a write rpc stream as long as there are dirtiers
1643                  * waiting for space.  as they're waiting, they're not going to
1644                  * create more pages to coalesce with what's waiting.. */
1645                 if (!cfs_list_empty(&cli->cl_cache_waiters)) {
1646                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1647                         RETURN(1);
1648                 }
1649                 if (cfs_atomic_read(&osc->oo_nr_writes) >=
1650                     cli->cl_max_pages_per_rpc)
1651                         RETURN(1);
1652         } else {
1653                 if (cfs_atomic_read(&osc->oo_nr_reads) == 0)
1654                         RETURN(0);
1655                 if (invalid_import) {
1656                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1657                         RETURN(1);
1658                 }
1659                 /* all read are urgent. */
1660                 if (!cfs_list_empty(&osc->oo_reading_exts))
1661                         RETURN(1);
1662         }
1663
1664         RETURN(0);
1665 }
1666
1667 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1668 {
1669         struct client_obd *cli = osc_cli(obj);
1670         if (cmd & OBD_BRW_WRITE) {
1671                 cfs_atomic_add(delta, &obj->oo_nr_writes);
1672                 cfs_atomic_add(delta, &cli->cl_pending_w_pages);
1673                 LASSERT(cfs_atomic_read(&obj->oo_nr_writes) >= 0);
1674         } else {
1675                 cfs_atomic_add(delta, &obj->oo_nr_reads);
1676                 cfs_atomic_add(delta, &cli->cl_pending_r_pages);
1677                 LASSERT(cfs_atomic_read(&obj->oo_nr_reads) >= 0);
1678         }
1679         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1680 }
1681
1682 static int osc_makes_hprpc(struct osc_object *obj)
1683 {
1684         return !cfs_list_empty(&obj->oo_hp_exts);
1685 }
1686
1687 static void on_list(cfs_list_t *item, cfs_list_t *list, int should_be_on)
1688 {
1689         if (cfs_list_empty(item) && should_be_on)
1690                 cfs_list_add_tail(item, list);
1691         else if (!cfs_list_empty(item) && !should_be_on)
1692                 cfs_list_del_init(item);
1693 }
1694
1695 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1696  * can find pages to build into rpcs quickly */
1697 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1698 {
1699         if (osc_makes_hprpc(osc)) {
1700                 /* HP rpc */
1701                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1702                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1703         } else {
1704                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1705                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1706                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1707                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1708         }
1709
1710         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1711                 cfs_atomic_read(&osc->oo_nr_writes) > 0);
1712
1713         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1714                 cfs_atomic_read(&osc->oo_nr_reads) > 0);
1715
1716         return osc_is_ready(osc);
1717 }
1718
1719 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1720 {
1721         int is_ready;
1722
1723         client_obd_list_lock(&cli->cl_loi_list_lock);
1724         is_ready = __osc_list_maint(cli, osc);
1725         client_obd_list_unlock(&cli->cl_loi_list_lock);
1726
1727         return is_ready;
1728 }
1729
1730 /* this is trying to propogate async writeback errors back up to the
1731  * application.  As an async write fails we record the error code for later if
1732  * the app does an fsync.  As long as errors persist we force future rpcs to be
1733  * sync so that the app can get a sync error and break the cycle of queueing
1734  * pages for which writeback will fail. */
1735 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1736                            int rc)
1737 {
1738         if (rc) {
1739                 if (!ar->ar_rc)
1740                         ar->ar_rc = rc;
1741
1742                 ar->ar_force_sync = 1;
1743                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1744                 return;
1745
1746         }
1747
1748         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1749                 ar->ar_force_sync = 0;
1750 }
1751
1752 /* Performs "unstable" page accounting. This function balances the
1753  * increment operations performed in osc_inc_unstable_pages. It is
1754  * registered as the RPC request callback, and is executed when the
1755  * bulk RPC is committed on the server. Thus at this point, the pages
1756  * involved in the bulk transfer are no longer considered unstable. */
1757 void osc_dec_unstable_pages(struct ptlrpc_request *req)
1758 {
1759         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
1760         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
1761         obd_count                page_count = desc->bd_iov_count;
1762         int i;
1763
1764         /* No unstable page tracking */
1765         if (cli->cl_cache == NULL)
1766                 return;
1767
1768         LASSERT(page_count >= 0);
1769
1770         for (i = 0; i < page_count; i++)
1771                 dec_zone_page_state(desc->bd_iov[i].kiov_page, NR_UNSTABLE_NFS);
1772
1773         cfs_atomic_sub(page_count, &cli->cl_cache->ccc_unstable_nr);
1774         LASSERT(cfs_atomic_read(&cli->cl_cache->ccc_unstable_nr) >= 0);
1775
1776         cfs_atomic_sub(page_count, &obd_unstable_pages);
1777         LASSERT(cfs_atomic_read(&obd_unstable_pages) >= 0);
1778
1779         spin_lock(&req->rq_lock);
1780         req->rq_committed = 1;
1781         req->rq_unstable  = 0;
1782         spin_unlock(&req->rq_lock);
1783
1784         cfs_waitq_broadcast(&cli->cl_cache->ccc_unstable_waitq);
1785 }
1786
1787 /* "unstable" page accounting. See: osc_dec_unstable_pages. */
1788 void osc_inc_unstable_pages(struct ptlrpc_request *req)
1789 {
1790         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
1791         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
1792         obd_count                page_count = desc->bd_iov_count;
1793         int i;
1794
1795         /* No unstable page tracking */
1796         if (cli->cl_cache == NULL)
1797                 return;
1798
1799         LASSERT(page_count >= 0);
1800
1801         for (i = 0; i < page_count; i++)
1802                 inc_zone_page_state(desc->bd_iov[i].kiov_page, NR_UNSTABLE_NFS);
1803
1804         LASSERT(cfs_atomic_read(&cli->cl_cache->ccc_unstable_nr) >= 0);
1805         cfs_atomic_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1806
1807         LASSERT(cfs_atomic_read(&obd_unstable_pages) >= 0);
1808         cfs_atomic_add(page_count, &obd_unstable_pages);
1809
1810         spin_lock(&req->rq_lock);
1811
1812         /* If the request has already been committed (i.e. brw_commit
1813          * called via rq_commit_cb), we need to undo the unstable page
1814          * increments we just performed because rq_commit_cb wont be
1815          * called again. Otherwise, just set the commit callback so the
1816          * unstable page accounting is properly updated when the request
1817          * is committed */
1818         if (req->rq_committed) {
1819                 /* Drop lock before calling osc_dec_unstable_pages */
1820                 spin_unlock(&req->rq_lock);
1821                 osc_dec_unstable_pages(req);
1822                 spin_lock(&req->rq_lock);
1823         } else {
1824                 req->rq_unstable  = 1;
1825                 req->rq_commit_cb = osc_dec_unstable_pages;
1826         }
1827
1828         spin_unlock(&req->rq_lock);
1829 }
1830
1831 /* this must be called holding the loi list lock to give coverage to exit_cache,
1832  * async_flag maintenance, and oap_request */
1833 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1834                               struct osc_async_page *oap, int sent, int rc)
1835 {
1836         struct osc_object *osc = oap->oap_obj;
1837         struct lov_oinfo  *loi = osc->oo_oinfo;
1838         __u64 xid = 0;
1839
1840         ENTRY;
1841         if (oap->oap_request != NULL) {
1842                 if (rc == 0)
1843                         osc_inc_unstable_pages(oap->oap_request);
1844
1845                 xid = ptlrpc_req_xid(oap->oap_request);
1846                 ptlrpc_req_finished(oap->oap_request);
1847                 oap->oap_request = NULL;
1848         }
1849
1850         /* As the transfer for this page is being done, clear the flags */
1851         spin_lock(&oap->oap_lock);
1852         oap->oap_async_flags = 0;
1853         spin_unlock(&oap->oap_lock);
1854         oap->oap_interrupted = 0;
1855
1856         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1857                 client_obd_list_lock(&cli->cl_loi_list_lock);
1858                 osc_process_ar(&cli->cl_ar, xid, rc);
1859                 osc_process_ar(&loi->loi_ar, xid, rc);
1860                 client_obd_list_unlock(&cli->cl_loi_list_lock);
1861         }
1862
1863         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1864         if (rc)
1865                 CERROR("completion on oap %p obj %p returns %d.\n",
1866                        oap, osc, rc);
1867
1868         EXIT;
1869 }
1870
1871 /**
1872  * Try to add extent to one RPC. We need to think about the following things:
1873  * - # of pages must not be over max_pages_per_rpc
1874  * - extent must be compatible with previous ones
1875  */
1876 static int try_to_add_extent_for_io(struct client_obd *cli,
1877                                     struct osc_extent *ext, cfs_list_t *rpclist,
1878                                     int *pc, unsigned int *max_pages)
1879 {
1880         struct osc_extent *tmp;
1881         ENTRY;
1882
1883         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1884                 ext);
1885
1886         *max_pages = max(ext->oe_mppr, *max_pages);
1887         if (*pc + ext->oe_nr_pages > *max_pages)
1888                 RETURN(0);
1889
1890         cfs_list_for_each_entry(tmp, rpclist, oe_link) {
1891                 EASSERT(tmp->oe_owner == cfs_current(), tmp);
1892 #if 0
1893                 if (overlapped(tmp, ext)) {
1894                         OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
1895                         EASSERT(0, ext);
1896                 }
1897 #endif
1898
1899                 if (tmp->oe_srvlock != ext->oe_srvlock ||
1900                     !tmp->oe_grants != !ext->oe_grants)
1901                         RETURN(0);
1902
1903                 /* remove break for strict check */
1904                 break;
1905         }
1906
1907         *pc += ext->oe_nr_pages;
1908         cfs_list_move_tail(&ext->oe_link, rpclist);
1909         ext->oe_owner = cfs_current();
1910         RETURN(1);
1911 }
1912
1913 /**
1914  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1915  * get_write_extent() takes all appropriate extents in atomic.
1916  *
1917  * The following policy is used to collect extents for IO:
1918  * 1. Add as many HP extents as possible;
1919  * 2. Add the first urgent extent in urgent extent list and take it out of
1920  *    urgent list;
1921  * 3. Add subsequent extents of this urgent extent;
1922  * 4. If urgent list is not empty, goto 2;
1923  * 5. Traverse the extent tree from the 1st extent;
1924  * 6. Above steps exit if there is no space in this RPC.
1925  */
1926 static int get_write_extents(struct osc_object *obj, cfs_list_t *rpclist)
1927 {
1928         struct client_obd *cli = osc_cli(obj);
1929         struct osc_extent *ext;
1930         int page_count = 0;
1931         unsigned int max_pages = cli->cl_max_pages_per_rpc;
1932
1933         LASSERT(osc_object_is_locked(obj));
1934         while (!cfs_list_empty(&obj->oo_hp_exts)) {
1935                 ext = cfs_list_entry(obj->oo_hp_exts.next, struct osc_extent,
1936                                      oe_link);
1937                 LASSERT(ext->oe_state == OES_CACHE);
1938                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1939                                               &max_pages))
1940                         return page_count;
1941                 EASSERT(ext->oe_nr_pages <= max_pages, ext);
1942         }
1943         if (page_count == max_pages)
1944                 return page_count;
1945
1946         while (!cfs_list_empty(&obj->oo_urgent_exts)) {
1947                 ext = cfs_list_entry(obj->oo_urgent_exts.next,
1948                                      struct osc_extent, oe_link);
1949                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1950                                               &max_pages))
1951                         return page_count;
1952
1953                 if (!ext->oe_intree)
1954                         continue;
1955
1956                 while ((ext = next_extent(ext)) != NULL) {
1957                         if ((ext->oe_state != OES_CACHE) ||
1958                             (!cfs_list_empty(&ext->oe_link) &&
1959                              ext->oe_owner != NULL))
1960                                 continue;
1961
1962                         if (!try_to_add_extent_for_io(cli, ext, rpclist,
1963                                                       &page_count, &max_pages))
1964                                 return page_count;
1965                 }
1966         }
1967         if (page_count == max_pages)
1968                 return page_count;
1969
1970         ext = first_extent(obj);
1971         while (ext != NULL) {
1972                 if ((ext->oe_state != OES_CACHE) ||
1973                     /* this extent may be already in current rpclist */
1974                     (!cfs_list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
1975                         ext = next_extent(ext);
1976                         continue;
1977                 }
1978
1979                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1980                                               &max_pages))
1981                         return page_count;
1982
1983                 ext = next_extent(ext);
1984         }
1985         return page_count;
1986 }
1987
1988 static int
1989 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
1990                    struct osc_object *osc, pdl_policy_t pol)
1991 {
1992         CFS_LIST_HEAD(rpclist);
1993         struct osc_extent *ext;
1994         struct osc_extent *tmp;
1995         struct osc_extent *first = NULL;
1996         obd_count page_count = 0;
1997         int srvlock = 0;
1998         int rc = 0;
1999         ENTRY;
2000
2001         LASSERT(osc_object_is_locked(osc));
2002
2003         page_count = get_write_extents(osc, &rpclist);
2004         LASSERT(equi(page_count == 0, cfs_list_empty(&rpclist)));
2005
2006         if (cfs_list_empty(&rpclist))
2007                 RETURN(0);
2008
2009         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2010
2011         cfs_list_for_each_entry(ext, &rpclist, oe_link) {
2012                 LASSERT(ext->oe_state == OES_CACHE ||
2013                         ext->oe_state == OES_LOCK_DONE);
2014                 if (ext->oe_state == OES_CACHE)
2015                         osc_extent_state_set(ext, OES_LOCKING);
2016                 else
2017                         osc_extent_state_set(ext, OES_RPC);
2018         }
2019
2020         /* we're going to grab page lock, so release object lock because
2021          * lock order is page lock -> object lock. */
2022         osc_object_unlock(osc);
2023
2024         cfs_list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2025                 if (ext->oe_state == OES_LOCKING) {
2026                         rc = osc_extent_make_ready(env, ext);
2027                         if (unlikely(rc < 0)) {
2028                                 cfs_list_del_init(&ext->oe_link);
2029                                 osc_extent_finish(env, ext, 0, rc);
2030                                 continue;
2031                         }
2032                 }
2033                 if (first == NULL) {
2034                         first = ext;
2035                         srvlock = ext->oe_srvlock;
2036                 } else {
2037                         LASSERT(srvlock == ext->oe_srvlock);
2038                 }
2039         }
2040
2041         if (!cfs_list_empty(&rpclist)) {
2042                 LASSERT(page_count > 0);
2043                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE, pol);
2044                 LASSERT(cfs_list_empty(&rpclist));
2045         }
2046
2047         osc_object_lock(osc);
2048         RETURN(rc);
2049 }
2050
2051 /**
2052  * prepare pages for ASYNC io and put pages in send queue.
2053  *
2054  * \param cmd OBD_BRW_* macroses
2055  * \param lop pending pages
2056  *
2057  * \return zero if no page added to send queue.
2058  * \return 1 if pages successfully added to send queue.
2059  * \return negative on errors.
2060  */
2061 static int
2062 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2063                   struct osc_object *osc, pdl_policy_t pol)
2064 {
2065         struct osc_extent *ext;
2066         struct osc_extent *next;
2067         CFS_LIST_HEAD(rpclist);
2068         int page_count = 0;
2069         unsigned int max_pages = cli->cl_max_pages_per_rpc;
2070         int rc = 0;
2071         ENTRY;
2072
2073         LASSERT(osc_object_is_locked(osc));
2074         cfs_list_for_each_entry_safe(ext, next,
2075                                      &osc->oo_reading_exts, oe_link) {
2076                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2077                 if (!try_to_add_extent_for_io(cli, ext, &rpclist, &page_count,
2078                                               &max_pages))
2079                         break;
2080                 osc_extent_state_set(ext, OES_RPC);
2081                 EASSERT(ext->oe_nr_pages <= max_pages, ext);
2082         }
2083         LASSERT(page_count <= max_pages);
2084
2085         osc_update_pending(osc, OBD_BRW_READ, -page_count);
2086
2087         if (!cfs_list_empty(&rpclist)) {
2088                 osc_object_unlock(osc);
2089
2090                 LASSERT(page_count > 0);
2091                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ, pol);
2092                 LASSERT(cfs_list_empty(&rpclist));
2093
2094                 osc_object_lock(osc);
2095         }
2096         RETURN(rc);
2097 }
2098
2099 #define list_to_obj(list, item) ({                                            \
2100         cfs_list_t *__tmp = (list)->next;                                     \
2101         cfs_list_del_init(__tmp);                                             \
2102         cfs_list_entry(__tmp, struct osc_object, oo_##item);                  \
2103 })
2104
2105 /* This is called by osc_check_rpcs() to find which objects have pages that
2106  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2107 static struct osc_object *osc_next_obj(struct client_obd *cli)
2108 {
2109         ENTRY;
2110
2111         /* First return objects that have blocked locks so that they
2112          * will be flushed quickly and other clients can get the lock,
2113          * then objects which have pages ready to be stuffed into RPCs */
2114         if (!cfs_list_empty(&cli->cl_loi_hp_ready_list))
2115                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2116         if (!cfs_list_empty(&cli->cl_loi_ready_list))
2117                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2118
2119         /* then if we have cache waiters, return all objects with queued
2120          * writes.  This is especially important when many small files
2121          * have filled up the cache and not been fired into rpcs because
2122          * they don't pass the nr_pending/object threshhold */
2123         if (!cfs_list_empty(&cli->cl_cache_waiters) &&
2124             !cfs_list_empty(&cli->cl_loi_write_list))
2125                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2126
2127         /* then return all queued objects when we have an invalid import
2128          * so that they get flushed */
2129         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2130                 if (!cfs_list_empty(&cli->cl_loi_write_list))
2131                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2132                                            write_item));
2133                 if (!cfs_list_empty(&cli->cl_loi_read_list))
2134                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2135                                            read_item));
2136         }
2137         RETURN(NULL);
2138 }
2139
2140 /* called with the loi list lock held */
2141 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli,
2142                            pdl_policy_t pol)
2143 {
2144         struct osc_object *osc;
2145         int rc = 0;
2146         ENTRY;
2147
2148         while ((osc = osc_next_obj(cli)) != NULL) {
2149                 struct cl_object *obj = osc2cl(osc);
2150                 struct lu_ref_link *link;
2151
2152                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2153
2154                 if (osc_max_rpc_in_flight(cli, osc)) {
2155                         __osc_list_maint(cli, osc);
2156                         break;
2157                 }
2158
2159                 cl_object_get(obj);
2160                 client_obd_list_unlock(&cli->cl_loi_list_lock);
2161                 link = lu_object_ref_add(&obj->co_lu, "check", cfs_current());
2162
2163                 /* attempt some read/write balancing by alternating between
2164                  * reads and writes in an object.  The makes_rpc checks here
2165                  * would be redundant if we were getting read/write work items
2166                  * instead of objects.  we don't want send_oap_rpc to drain a
2167                  * partial read pending queue when we're given this object to
2168                  * do io on writes while there are cache waiters */
2169                 osc_object_lock(osc);
2170                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2171                         rc = osc_send_write_rpc(env, cli, osc, pol);
2172                         if (rc < 0) {
2173                                 CERROR("Write request failed with %d\n", rc);
2174
2175                                 /* osc_send_write_rpc failed, mostly because of
2176                                  * memory pressure.
2177                                  *
2178                                  * It can't break here, because if:
2179                                  *  - a page was submitted by osc_io_submit, so
2180                                  *    page locked;
2181                                  *  - no request in flight
2182                                  *  - no subsequent request
2183                                  * The system will be in live-lock state,
2184                                  * because there is no chance to call
2185                                  * osc_io_unplug() and osc_check_rpcs() any
2186                                  * more. pdflush can't help in this case,
2187                                  * because it might be blocked at grabbing
2188                                  * the page lock as we mentioned.
2189                                  *
2190                                  * Anyway, continue to drain pages. */
2191                                 /* break; */
2192                         }
2193                 }
2194                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2195                         rc = osc_send_read_rpc(env, cli, osc, pol);
2196                         if (rc < 0)
2197                                 CERROR("Read request failed with %d\n", rc);
2198                 }
2199                 osc_object_unlock(osc);
2200
2201                 osc_list_maint(cli, osc);
2202                 lu_object_ref_del_at(&obj->co_lu, link, "check", cfs_current());
2203                 cl_object_put(env, obj);
2204
2205                 client_obd_list_lock(&cli->cl_loi_list_lock);
2206         }
2207 }
2208
2209 static int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2210                           struct osc_object *osc, pdl_policy_t pol, int async)
2211 {
2212         int has_rpcs = 1;
2213         int rc = 0;
2214
2215         client_obd_list_lock(&cli->cl_loi_list_lock);
2216         if (osc != NULL)
2217                 has_rpcs = __osc_list_maint(cli, osc);
2218         if (has_rpcs) {
2219                 if (!async) {
2220                         /* disable osc_lru_shrink() temporarily to avoid
2221                          * potential stack overrun problem. LU-2859 */
2222                         cfs_atomic_inc(&cli->cl_lru_shrinkers);
2223                         osc_check_rpcs(env, cli, pol);
2224                         cfs_atomic_dec(&cli->cl_lru_shrinkers);
2225                 } else {
2226                         CDEBUG(D_CACHE, "Queue writeback work for client %p.\n",
2227                                cli);
2228                         LASSERT(cli->cl_writeback_work != NULL);
2229                         rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2230                 }
2231         }
2232         client_obd_list_unlock(&cli->cl_loi_list_lock);
2233         return rc;
2234 }
2235
2236 static int osc_io_unplug_async(const struct lu_env *env,
2237                                 struct client_obd *cli, struct osc_object *osc)
2238 {
2239         /* XXX: policy is no use actually. */
2240         return osc_io_unplug0(env, cli, osc, PDL_POLICY_ROUND, 1);
2241 }
2242
2243 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
2244                    struct osc_object *osc, pdl_policy_t pol)
2245 {
2246         (void)osc_io_unplug0(env, cli, osc, pol, 0);
2247 }
2248
2249 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2250                         cfs_page_t *page, loff_t offset)
2251 {
2252         struct obd_export     *exp = osc_export(osc);
2253         struct osc_async_page *oap = &ops->ops_oap;
2254         ENTRY;
2255
2256         if (!page)
2257                 return cfs_size_round(sizeof(*oap));
2258
2259         oap->oap_magic = OAP_MAGIC;
2260         oap->oap_cli = &exp->exp_obd->u.cli;
2261         oap->oap_obj = osc;
2262
2263         oap->oap_page = page;
2264         oap->oap_obj_off = offset;
2265         LASSERT(!(offset & ~CFS_PAGE_MASK));
2266
2267         if (!client_is_remote(exp) && cfs_capable(CFS_CAP_SYS_RESOURCE))
2268                 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2269
2270         CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2271         CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2272
2273         spin_lock_init(&oap->oap_lock);
2274         CDEBUG(D_INFO, "oap %p page %p obj off "LPU64"\n",
2275                oap, page, oap->oap_obj_off);
2276         RETURN(0);
2277 }
2278
2279 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2280                        struct osc_page *ops)
2281 {
2282         struct osc_io *oio = osc_env_io(env);
2283         struct osc_extent     *ext = NULL;
2284         struct osc_async_page *oap = &ops->ops_oap;
2285         struct client_obd     *cli = oap->oap_cli;
2286         struct osc_object     *osc = oap->oap_obj;
2287         pgoff_t index;
2288         int    grants = 0;
2289         int    brw_flags = OBD_BRW_ASYNC;
2290         int    cmd = OBD_BRW_WRITE;
2291         int    need_release = 0;
2292         int    rc = 0;
2293         ENTRY;
2294
2295         if (oap->oap_magic != OAP_MAGIC)
2296                 RETURN(-EINVAL);
2297
2298         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2299                 RETURN(-EIO);
2300
2301         if (!cfs_list_empty(&oap->oap_pending_item) ||
2302             !cfs_list_empty(&oap->oap_rpc_item))
2303                 RETURN(-EBUSY);
2304
2305         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2306         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2307         if (!client_is_remote(osc_export(osc)) &&
2308             cfs_capable(CFS_CAP_SYS_RESOURCE)) {
2309                 brw_flags |= OBD_BRW_NOQUOTA;
2310                 cmd |= OBD_BRW_NOQUOTA;
2311         }
2312
2313         /* check if the file's owner/group is over quota */
2314         if (!(cmd & OBD_BRW_NOQUOTA)) {
2315                 struct cl_object *obj;
2316                 struct cl_attr   *attr;
2317                 unsigned int qid[MAXQUOTAS];
2318
2319                 obj = cl_object_top(&osc->oo_cl);
2320                 attr = &osc_env_info(env)->oti_attr;
2321
2322                 cl_object_attr_lock(obj);
2323                 rc = cl_object_attr_get(env, obj, attr);
2324                 cl_object_attr_unlock(obj);
2325
2326                 qid[USRQUOTA] = attr->cat_uid;
2327                 qid[GRPQUOTA] = attr->cat_gid;
2328                 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2329                         rc = -EDQUOT;
2330                 if (rc)
2331                         RETURN(rc);
2332         }
2333
2334         oap->oap_cmd = cmd;
2335         oap->oap_page_off = ops->ops_from;
2336         oap->oap_count = ops->ops_to - ops->ops_from;
2337         oap->oap_async_flags = 0;
2338         oap->oap_brw_flags = brw_flags;
2339
2340         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2341                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2342
2343         index = oap2cl_page(oap)->cp_index;
2344
2345         /* Add this page into extent by the following steps:
2346          * 1. if there exists an active extent for this IO, mostly this page
2347          *    can be added to the active extent and sometimes we need to
2348          *    expand extent to accomodate this page;
2349          * 2. otherwise, a new extent will be allocated. */
2350
2351         ext = oio->oi_active;
2352         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2353                 /* one chunk plus extent overhead must be enough to write this
2354                  * page */
2355                 grants = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2356                 if (ext->oe_end >= index)
2357                         grants = 0;
2358
2359                 /* it doesn't need any grant to dirty this page */
2360                 client_obd_list_lock(&cli->cl_loi_list_lock);
2361                 rc = osc_enter_cache_try(cli, oap, grants, 0);
2362                 client_obd_list_unlock(&cli->cl_loi_list_lock);
2363                 if (rc == 0) { /* try failed */
2364                         grants = 0;
2365                         need_release = 1;
2366                 } else if (ext->oe_end < index) {
2367                         int tmp = grants;
2368                         /* try to expand this extent */
2369                         rc = osc_extent_expand(ext, index, &tmp);
2370                         if (rc < 0) {
2371                                 need_release = 1;
2372                                 /* don't free reserved grant */
2373                         } else {
2374                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2375                                                 "expanded for %lu.\n", index);
2376                                 osc_unreserve_grant(cli, grants, tmp);
2377                                 grants = 0;
2378                         }
2379                 }
2380                 rc = 0;
2381         } else if (ext != NULL) {
2382                 /* index is located outside of active extent */
2383                 need_release = 1;
2384         }
2385         if (need_release) {
2386                 osc_extent_release(env, ext);
2387                 oio->oi_active = NULL;
2388                 ext = NULL;
2389         }
2390
2391         if (ext == NULL) {
2392                 int tmp = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2393
2394                 /* try to find new extent to cover this page */
2395                 LASSERT(oio->oi_active == NULL);
2396                 /* we may have allocated grant for this page if we failed
2397                  * to expand the previous active extent. */
2398                 LASSERT(ergo(grants > 0, grants >= tmp));
2399
2400                 rc = 0;
2401                 if (grants == 0) {
2402                         /* we haven't allocated grant for this page. */
2403                         rc = osc_enter_cache(env, cli, oap, tmp);
2404                         if (rc == 0)
2405                                 grants = tmp;
2406                 }
2407
2408                 tmp = grants;
2409                 if (rc == 0) {
2410                         ext = osc_extent_find(env, osc, index, &tmp);
2411                         if (IS_ERR(ext)) {
2412                                 LASSERT(tmp == grants);
2413                                 osc_exit_cache(cli, oap);
2414                                 rc = PTR_ERR(ext);
2415                                 ext = NULL;
2416                         } else {
2417                                 oio->oi_active = ext;
2418                         }
2419                 }
2420                 if (grants > 0)
2421                         osc_unreserve_grant(cli, grants, tmp);
2422         }
2423
2424         LASSERT(ergo(rc == 0, ext != NULL));
2425         if (ext != NULL) {
2426                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2427                          ext, "index = %lu.\n", index);
2428                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2429
2430                 osc_object_lock(osc);
2431                 if (ext->oe_nr_pages == 0)
2432                         ext->oe_srvlock = ops->ops_srvlock;
2433                 else
2434                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2435                 ++ext->oe_nr_pages;
2436                 cfs_list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2437                 osc_object_unlock(osc);
2438         }
2439         RETURN(rc);
2440 }
2441
2442 int osc_teardown_async_page(const struct lu_env *env,
2443                             struct osc_object *obj, struct osc_page *ops)
2444 {
2445         struct osc_async_page *oap = &ops->ops_oap;
2446         struct osc_extent     *ext = NULL;
2447         int rc = 0;
2448         ENTRY;
2449
2450         LASSERT(oap->oap_magic == OAP_MAGIC);
2451
2452         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2453                oap, ops, oap2cl_page(oap)->cp_index);
2454
2455         osc_object_lock(obj);
2456         if (!cfs_list_empty(&oap->oap_rpc_item)) {
2457                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2458                 rc = -EBUSY;
2459         } else if (!cfs_list_empty(&oap->oap_pending_item)) {
2460                 ext = osc_extent_lookup(obj, oap2cl_page(oap)->cp_index);
2461                 /* only truncated pages are allowed to be taken out.
2462                  * See osc_extent_truncate() and osc_cache_truncate_start()
2463                  * for details. */
2464                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2465                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2466                                         oap2cl_page(oap)->cp_index);
2467                         rc = -EBUSY;
2468                 }
2469         }
2470         osc_object_unlock(obj);
2471         if (ext != NULL)
2472                 osc_extent_put(env, ext);
2473         RETURN(rc);
2474 }
2475
2476 /**
2477  * This is called when a page is picked up by kernel to write out.
2478  *
2479  * We should find out the corresponding extent and add the whole extent
2480  * into urgent list. The extent may be being truncated or used, handle it
2481  * carefully.
2482  */
2483 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2484                          struct osc_page *ops)
2485 {
2486         struct osc_extent *ext   = NULL;
2487         struct osc_object *obj   = cl2osc(ops->ops_cl.cpl_obj);
2488         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2489         pgoff_t            index = cp->cp_index;
2490         struct osc_async_page *oap = &ops->ops_oap;
2491         bool unplug = false;
2492         int rc = 0;
2493         ENTRY;
2494
2495         osc_object_lock(obj);
2496         ext = osc_extent_lookup(obj, index);
2497         if (ext == NULL) {
2498                 osc_extent_tree_dump(D_ERROR, obj);
2499                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2500         }
2501
2502         switch (ext->oe_state) {
2503         case OES_RPC:
2504         case OES_LOCK_DONE:
2505                 CL_PAGE_DEBUG(D_ERROR, env, cl_page_top(cp),
2506                               "flush an in-rpc page?\n");
2507                 LASSERT(0);
2508                 break;
2509         case OES_LOCKING:
2510                 /* If we know this extent is being written out, we should abort
2511                  * so that the writer can make this page ready. Otherwise, there
2512                  * exists a deadlock problem because other process can wait for
2513                  * page writeback bit holding page lock; and meanwhile in
2514                  * vvp_page_make_ready(), we need to grab page lock before
2515                  * really sending the RPC. */
2516         case OES_TRUNC:
2517                 /* race with truncate, page will be redirtied */
2518                 GOTO(out, rc = -EAGAIN);
2519         default:
2520                 break;
2521         }
2522
2523         rc = cl_page_prep(env, io, cl_page_top(cp), CRT_WRITE);
2524         if (rc)
2525                 GOTO(out, rc);
2526
2527         spin_lock(&oap->oap_lock);
2528         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2529         spin_unlock(&oap->oap_lock);
2530
2531         if (cfs_memory_pressure_get())
2532                 ext->oe_memalloc = 1;
2533
2534         ext->oe_urgent = 1;
2535         if (ext->oe_state == OES_CACHE) {
2536                 OSC_EXTENT_DUMP(D_CACHE, ext,
2537                                 "flush page %p make it urgent.\n", oap);
2538                 if (cfs_list_empty(&ext->oe_link))
2539                         cfs_list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2540                 unplug = true;
2541         }
2542         rc = 0;
2543         EXIT;
2544
2545 out:
2546         osc_object_unlock(obj);
2547         osc_extent_put(env, ext);
2548         if (unplug)
2549                 osc_io_unplug_async(env, osc_cli(obj), obj);
2550         return rc;
2551 }
2552
2553 /**
2554  * this is called when a sync waiter receives an interruption.  Its job is to
2555  * get the caller woken as soon as possible.  If its page hasn't been put in an
2556  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2557  * desiring interruption which will forcefully complete the rpc once the rpc
2558  * has timed out.
2559  */
2560 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2561 {
2562         struct osc_async_page *oap = &ops->ops_oap;
2563         struct osc_object     *obj = oap->oap_obj;
2564         struct client_obd     *cli = osc_cli(obj);
2565         struct osc_extent     *ext;
2566         struct osc_extent     *found = NULL;
2567         cfs_list_t            *plist;
2568         pgoff_t index = oap2cl_page(oap)->cp_index;
2569         int     rc = -EBUSY;
2570         int     cmd;
2571         ENTRY;
2572
2573         LASSERT(!oap->oap_interrupted);
2574         oap->oap_interrupted = 1;
2575
2576         /* Find out the caching extent */
2577         osc_object_lock(obj);
2578         if (oap->oap_cmd & OBD_BRW_WRITE) {
2579                 plist = &obj->oo_urgent_exts;
2580                 cmd   = OBD_BRW_WRITE;
2581         } else {
2582                 plist = &obj->oo_reading_exts;
2583                 cmd   = OBD_BRW_READ;
2584         }
2585         cfs_list_for_each_entry(ext, plist, oe_link) {
2586                 if (ext->oe_start <= index && ext->oe_end >= index) {
2587                         LASSERT(ext->oe_state == OES_LOCK_DONE);
2588                         /* For OES_LOCK_DONE state extent, it has already held
2589                          * a refcount for RPC. */
2590                         found = osc_extent_get(ext);
2591                         break;
2592                 }
2593         }
2594         if (found != NULL) {
2595                 cfs_list_del_init(&found->oe_link);
2596                 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2597                 osc_object_unlock(obj);
2598
2599                 osc_extent_finish(env, found, 0, -EINTR);
2600                 osc_extent_put(env, found);
2601                 rc = 0;
2602         } else {
2603                 osc_object_unlock(obj);
2604                 /* ok, it's been put in an rpc. only one oap gets a request
2605                  * reference */
2606                 if (oap->oap_request != NULL) {
2607                         ptlrpc_mark_interrupted(oap->oap_request);
2608                         ptlrpcd_wake(oap->oap_request);
2609                         ptlrpc_req_finished(oap->oap_request);
2610                         oap->oap_request = NULL;
2611                 }
2612         }
2613
2614         osc_list_maint(cli, obj);
2615         RETURN(rc);
2616 }
2617
2618 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
2619                          cfs_list_t *list, int cmd, int brw_flags)
2620 {
2621         struct client_obd     *cli = osc_cli(obj);
2622         struct osc_extent     *ext;
2623         struct osc_async_page *oap;
2624         int     page_count = 0;
2625         int     mppr       = cli->cl_max_pages_per_rpc;
2626         pgoff_t start      = CL_PAGE_EOF;
2627         pgoff_t end        = 0;
2628         ENTRY;
2629
2630         cfs_list_for_each_entry(oap, list, oap_pending_item) {
2631                 struct cl_page *cp = oap2cl_page(oap);
2632                 if (cp->cp_index > end)
2633                         end = cp->cp_index;
2634                 if (cp->cp_index < start)
2635                         start = cp->cp_index;
2636                 ++page_count;
2637                 mppr <<= (page_count > mppr);
2638         }
2639
2640         ext = osc_extent_alloc(obj);
2641         if (ext == NULL) {
2642                 cfs_list_for_each_entry(oap, list, oap_pending_item) {
2643                         cfs_list_del_init(&oap->oap_pending_item);
2644                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2645                 }
2646                 RETURN(-ENOMEM);
2647         }
2648
2649         ext->oe_rw = !!(cmd & OBD_BRW_READ);
2650         ext->oe_urgent = 1;
2651         ext->oe_start = start;
2652         ext->oe_end = ext->oe_max_end = end;
2653         ext->oe_obj = obj;
2654         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2655         ext->oe_nr_pages = page_count;
2656         ext->oe_mppr = mppr;
2657         cfs_list_splice_init(list, &ext->oe_pages);
2658
2659         osc_object_lock(obj);
2660         /* Reuse the initial refcount for RPC, don't drop it */
2661         osc_extent_state_set(ext, OES_LOCK_DONE);
2662         if (cmd & OBD_BRW_WRITE) {
2663                 cfs_list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2664                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2665         } else {
2666                 cfs_list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2667                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2668         }
2669         osc_object_unlock(obj);
2670
2671         osc_io_unplug_async(env, cli, obj);
2672         RETURN(0);
2673 }
2674
2675 /**
2676  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2677  */
2678 int osc_cache_truncate_start(const struct lu_env *env, struct osc_io *oio,
2679                              struct osc_object *obj, __u64 size)
2680 {
2681         struct client_obd *cli = osc_cli(obj);
2682         struct osc_extent *ext;
2683         struct osc_extent *waiting = NULL;
2684         pgoff_t index;
2685         CFS_LIST_HEAD(list);
2686         int result = 0;
2687         bool partial;
2688         ENTRY;
2689
2690         /* pages with index greater or equal to index will be truncated. */
2691         index = cl_index(osc2cl(obj), size);
2692         partial = size > cl_offset(osc2cl(obj), index);
2693
2694 again:
2695         osc_object_lock(obj);
2696         ext = osc_extent_search(obj, index);
2697         if (ext == NULL)
2698                 ext = first_extent(obj);
2699         else if (ext->oe_end < index)
2700                 ext = next_extent(ext);
2701         while (ext != NULL) {
2702                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2703
2704                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2705                         /* if ext is in urgent state, it means there must exist
2706                          * a page already having been flushed by write_page().
2707                          * We have to wait for this extent because we can't
2708                          * truncate that page. */
2709                         LASSERT(!ext->oe_hp);
2710                         OSC_EXTENT_DUMP(D_CACHE, ext,
2711                                         "waiting for busy extent\n");
2712                         waiting = osc_extent_get(ext);
2713                         break;
2714                 }
2715
2716                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:"LPU64".\n", size);
2717
2718                 osc_extent_get(ext);
2719                 if (ext->oe_state == OES_ACTIVE) {
2720                         /* though we grab inode mutex for write path, but we
2721                          * release it before releasing extent(in osc_io_end()),
2722                          * so there is a race window that an extent is still
2723                          * in OES_ACTIVE when truncate starts. */
2724                         LASSERT(!ext->oe_trunc_pending);
2725                         ext->oe_trunc_pending = 1;
2726                 } else {
2727                         EASSERT(ext->oe_state == OES_CACHE, ext);
2728                         osc_extent_state_set(ext, OES_TRUNC);
2729                         osc_update_pending(obj, OBD_BRW_WRITE,
2730                                            -ext->oe_nr_pages);
2731                 }
2732                 EASSERT(cfs_list_empty(&ext->oe_link), ext);
2733                 cfs_list_add_tail(&ext->oe_link, &list);
2734
2735                 ext = next_extent(ext);
2736         }
2737         osc_object_unlock(obj);
2738
2739         osc_list_maint(cli, obj);
2740
2741         while (!cfs_list_empty(&list)) {
2742                 int rc;
2743
2744                 ext = cfs_list_entry(list.next, struct osc_extent, oe_link);
2745                 cfs_list_del_init(&ext->oe_link);
2746
2747                 /* extent may be in OES_ACTIVE state because inode mutex
2748                  * is released before osc_io_end() in file write case */
2749                 if (ext->oe_state != OES_TRUNC)
2750                         osc_extent_wait(env, ext, OES_TRUNC);
2751
2752                 rc = osc_extent_truncate(ext, index, partial);
2753                 if (rc < 0) {
2754                         if (result == 0)
2755                                 result = rc;
2756
2757                         OSC_EXTENT_DUMP(D_ERROR, ext,
2758                                         "truncate error %d\n", rc);
2759                 } else if (ext->oe_nr_pages == 0) {
2760                         osc_extent_remove(ext);
2761                 } else {
2762                         /* this must be an overlapped extent which means only
2763                          * part of pages in this extent have been truncated.
2764                          */
2765                         EASSERTF(ext->oe_start <= index, ext,
2766                                  "trunc index = %lu/%d.\n", index, partial);
2767                         /* fix index to skip this partially truncated extent */
2768                         index = ext->oe_end + 1;
2769                         partial = false;
2770
2771                         /* we need to hold this extent in OES_TRUNC state so
2772                          * that no writeback will happen. This is to avoid
2773                          * BUG 17397. */
2774                         LASSERT(oio->oi_trunc == NULL);
2775                         oio->oi_trunc = osc_extent_get(ext);
2776                         OSC_EXTENT_DUMP(D_CACHE, ext,
2777                                         "trunc at "LPU64"\n", size);
2778                 }
2779                 osc_extent_put(env, ext);
2780         }
2781         if (waiting != NULL) {
2782                 int rc;
2783
2784                 /* ignore the result of osc_extent_wait the write initiator
2785                  * should take care of it. */
2786                 rc = osc_extent_wait(env, waiting, OES_INV);
2787                 if (rc < 0)
2788                         OSC_EXTENT_DUMP(D_CACHE, ext, "wait error: %d.\n", rc);
2789
2790                 osc_extent_put(env, waiting);
2791                 waiting = NULL;
2792                 goto again;
2793         }
2794         RETURN(result);
2795 }
2796
2797 /**
2798  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2799  */
2800 void osc_cache_truncate_end(const struct lu_env *env, struct osc_io *oio,
2801                             struct osc_object *obj)
2802 {
2803         struct osc_extent *ext = oio->oi_trunc;
2804
2805         oio->oi_trunc = NULL;
2806         if (ext != NULL) {
2807                 bool unplug = false;
2808
2809                 EASSERT(ext->oe_nr_pages > 0, ext);
2810                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2811                 EASSERT(!ext->oe_urgent, ext);
2812
2813                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2814                 osc_object_lock(obj);
2815                 osc_extent_state_set(ext, OES_CACHE);
2816                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2817                         ext->oe_urgent = 1;
2818                         cfs_list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2819                         unplug = true;
2820                 }
2821                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2822                 osc_object_unlock(obj);
2823                 osc_extent_put(env, ext);
2824
2825                 if (unplug)
2826                         osc_io_unplug_async(env, osc_cli(obj), obj);
2827         }
2828 }
2829
2830 /**
2831  * Wait for extents in a specific range to be written out.
2832  * The caller must have called osc_cache_writeback_range() to issue IO
2833  * otherwise it will take a long time for this function to finish.
2834  *
2835  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2836  * nobody else can dirty this range of file while we're waiting for
2837  * extents to be written.
2838  */
2839 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2840                          pgoff_t start, pgoff_t end)
2841 {
2842         struct osc_extent *ext;
2843         pgoff_t index = start;
2844         int     result = 0;
2845         ENTRY;
2846
2847 again:
2848         osc_object_lock(obj);
2849         ext = osc_extent_search(obj, index);
2850         if (ext == NULL)
2851                 ext = first_extent(obj);
2852         else if (ext->oe_end < index)
2853                 ext = next_extent(ext);
2854         while (ext != NULL) {
2855                 int rc;
2856
2857                 if (ext->oe_start > end)
2858                         break;
2859
2860                 if (!ext->oe_fsync_wait) {
2861                         ext = next_extent(ext);
2862                         continue;
2863                 }
2864
2865                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2866                              ext->oe_hp || ext->oe_urgent), ext);
2867                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2868                              !ext->oe_hp && ext->oe_urgent), ext);
2869
2870                 index = ext->oe_end + 1;
2871                 osc_extent_get(ext);
2872                 osc_object_unlock(obj);
2873
2874                 rc = osc_extent_wait(env, ext, OES_INV);
2875                 if (result == 0)
2876                         result = rc;
2877                 osc_extent_put(env, ext);
2878                 goto again;
2879         }
2880         osc_object_unlock(obj);
2881
2882         OSC_IO_DEBUG(obj, "sync file range.\n");
2883         RETURN(result);
2884 }
2885
2886 /**
2887  * Called to write out a range of osc object.
2888  *
2889  * @hp     : should be set this is caused by lock cancel;
2890  * @discard: is set if dirty pages should be dropped - file will be deleted or
2891  *         truncated, this implies there is no partially discarding extents.
2892  *
2893  * Return how many pages will be issued, or error code if error occurred.
2894  */
2895 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2896                               pgoff_t start, pgoff_t end, int hp, int discard)
2897 {
2898         struct osc_extent *ext;
2899         CFS_LIST_HEAD(discard_list);
2900         bool unplug = false;
2901         int result = 0;
2902         ENTRY;
2903
2904         osc_object_lock(obj);
2905         ext = osc_extent_search(obj, start);
2906         if (ext == NULL)
2907                 ext = first_extent(obj);
2908         else if (ext->oe_end < start)
2909                 ext = next_extent(ext);
2910         while (ext != NULL) {
2911                 if (ext->oe_start > end)
2912                         break;
2913
2914                 ext->oe_fsync_wait = 1;
2915                 switch (ext->oe_state) {
2916                 case OES_CACHE:
2917                         result += ext->oe_nr_pages;
2918                         if (!discard) {
2919                                 cfs_list_t *list = NULL;
2920                                 if (hp) {
2921                                         EASSERT(!ext->oe_hp, ext);
2922                                         ext->oe_hp = 1;
2923                                         list = &obj->oo_hp_exts;
2924                                 } else if (!ext->oe_urgent) {
2925                                         ext->oe_urgent = 1;
2926                                         list = &obj->oo_urgent_exts;
2927                                 }
2928                                 if (list != NULL)
2929                                         cfs_list_move_tail(&ext->oe_link, list);
2930                                 unplug = true;
2931                         } else {
2932                                 /* the only discarder is lock cancelling, so
2933                                  * [start, end] must contain this extent */
2934                                 EASSERT(ext->oe_start >= start &&
2935                                         ext->oe_max_end <= end, ext);
2936                                 osc_extent_state_set(ext, OES_LOCKING);
2937                                 ext->oe_owner = cfs_current();
2938                                 cfs_list_move_tail(&ext->oe_link,
2939                                                    &discard_list);
2940                                 osc_update_pending(obj, OBD_BRW_WRITE,
2941                                                    -ext->oe_nr_pages);
2942                         }
2943                         break;
2944                 case OES_ACTIVE:
2945                         /* It's pretty bad to wait for ACTIVE extents, because
2946                          * we don't know how long we will wait for it to be
2947                          * flushed since it may be blocked at awaiting more
2948                          * grants. We do this for the correctness of fsync. */
2949                         LASSERT(hp == 0 && discard == 0);
2950                         ext->oe_urgent = 1;
2951                         break;
2952                 case OES_TRUNC:
2953                         /* this extent is being truncated, can't do anything
2954                          * for it now. it will be set to urgent after truncate
2955                          * is finished in osc_cache_truncate_end(). */
2956                 default:
2957                         break;
2958                 }
2959                 ext = next_extent(ext);
2960         }
2961         osc_object_unlock(obj);
2962
2963         LASSERT(ergo(!discard, cfs_list_empty(&discard_list)));
2964         if (!cfs_list_empty(&discard_list)) {
2965                 struct osc_extent *tmp;
2966                 int rc;
2967
2968                 osc_list_maint(osc_cli(obj), obj);
2969                 cfs_list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
2970                         cfs_list_del_init(&ext->oe_link);
2971                         EASSERT(ext->oe_state == OES_LOCKING, ext);
2972
2973                         /* Discard caching pages. We don't actually write this
2974                          * extent out but we complete it as if we did. */
2975                         rc = osc_extent_make_ready(env, ext);
2976                         if (unlikely(rc < 0)) {
2977                                 OSC_EXTENT_DUMP(D_ERROR, ext,
2978                                                 "make_ready returned %d\n", rc);
2979                                 if (result >= 0)
2980                                         result = rc;
2981                         }
2982
2983                         /* finish the extent as if the pages were sent */
2984                         osc_extent_finish(env, ext, 0, 0);
2985                 }
2986         }
2987
2988         if (unplug)
2989                 osc_io_unplug(env, osc_cli(obj), obj, PDL_POLICY_ROUND);
2990
2991         if (hp || discard) {
2992                 int rc;
2993                 rc = osc_cache_wait_range(env, obj, start, end);
2994                 if (result >= 0 && rc < 0)
2995                         result = rc;
2996         }
2997
2998         OSC_IO_DEBUG(obj, "cache page out.\n");
2999         RETURN(result);
3000 }
3001
3002 /** @} osc */