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 static const char *oes_strings[] = {
 103         "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
 104
 105 #define OSC_EXTENT_DUMP(lvl, extent, fmt, ...) do {                           \
 106         struct osc_extent *__ext = (extent);                                  \
 107         char __buf[16];                                                       \
 108                                                                               \
 109         CDEBUG(lvl,                                                           \
 110                 "extent %p@{" EXTSTR ", "                                     \
 111                 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt,          \
 112                 /* ----- extent part 0 ----- */                               \
 113                 __ext, EXTPARA(__ext),                                        \
 114                 /* ----- part 1 ----- */                                      \
 115                 cfs_atomic_read(&__ext->oe_refc),                             \
 116                 cfs_atomic_read(&__ext->oe_users),                            \
 117                 list_empty_marker(&__ext->oe_link),                           \
 118                 oes_strings[__ext->oe_state], ext_flags(__ext, __buf),        \
 119                 __ext->oe_obj,                                                \
 120                 /* ----- part 2 ----- */                                      \
 121                 __ext->oe_grants, __ext->oe_nr_pages,                         \
 122                 list_empty_marker(&__ext->oe_pages),                          \
 123                 waitqueue_active(&__ext->oe_waitq) ? '+' : '-',               \
 124                 __ext->oe_osclock, __ext->oe_mppr, __ext->oe_owner,           \
 125                 /* ----- part 4 ----- */                                      \
 126                 ## __VA_ARGS__);                                              \
 127 } while (0)
 128
 129 #undef EASSERTF
 130 #define EASSERTF(expr, ext, fmt, args...) do {                          \
 131         if (!(expr)) {                                                  \
 132                 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args);           \
 133                 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj);           \
 134                 LASSERT(expr);                                          \
 135         }                                                               \
 136 } while (0)
 137
 138 #undef EASSERT
 139 #define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
 140
 141 static inline struct osc_extent *rb_extent(struct rb_node *n)
 142 {
 143         if (n == NULL)
 144                 return NULL;
 145
 146         return container_of(n, struct osc_extent, oe_node);
 147 }
 148
 149 static inline struct osc_extent *next_extent(struct osc_extent *ext)
 150 {
 151         if (ext == NULL)
 152                 return NULL;
 153
 154         LASSERT(ext->oe_intree);
 155         return rb_extent(rb_next(&ext->oe_node));
 156 }
 157
 158 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
 159 {
 160         if (ext == NULL)
 161                 return NULL;
 162
 163         LASSERT(ext->oe_intree);
 164         return rb_extent(rb_prev(&ext->oe_node));
 165 }
 166
 167 static inline struct osc_extent *first_extent(struct osc_object *obj)
 168 {
 169         return rb_extent(rb_first(&obj->oo_root));
 170 }
 171
 172 /* object must be locked by caller. */
 173 static int osc_extent_sanity_check0(struct osc_extent *ext,
 174                                     const char *func, const int line)
 175 {
 176         struct osc_object *obj = ext->oe_obj;
 177         struct osc_async_page *oap;
 178         int page_count;
 179         int rc = 0;
 180
 181         if (!osc_object_is_locked(obj))
 182                 GOTO(out, rc = 9);
 183
 184         if (ext->oe_state >= OES_STATE_MAX)
 185                 GOTO(out, rc = 10);
 186
 187         if (cfs_atomic_read(&ext->oe_refc) <= 0)
 188                 GOTO(out, rc = 20);
 189
 190         if (cfs_atomic_read(&ext->oe_refc) < cfs_atomic_read(&ext->oe_users))
 191                 GOTO(out, rc = 30);
 192
 193         switch (ext->oe_state) {
 194         case OES_INV:
 195                 if (ext->oe_nr_pages > 0 || !cfs_list_empty(&ext->oe_pages))
 196                         GOTO(out, rc = 35);
 197                 GOTO(out, rc = 0);
 198                 break;
 199         case OES_ACTIVE:
 200                 if (cfs_atomic_read(&ext->oe_users) == 0)
 201                         GOTO(out, rc = 40);
 202                 if (ext->oe_hp)
 203                         GOTO(out, rc = 50);
 204                 if (ext->oe_fsync_wait && !ext->oe_urgent)
 205                         GOTO(out, rc = 55);
 206                 break;
 207         case OES_CACHE:
 208                 if (ext->oe_grants == 0)
 209                         GOTO(out, rc = 60);
 210                 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
 211                         GOTO(out, rc = 65);
 212         default:
 213                 if (cfs_atomic_read(&ext->oe_users) > 0)
 214                         GOTO(out, rc = 70);
 215         }
 216
 217         if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
 218                 GOTO(out, rc = 80);
 219
 220         if (ext->oe_osclock == NULL && ext->oe_grants > 0)
 221                 GOTO(out, rc = 90);
 222
 223         if (ext->oe_osclock) {
 224                 struct cl_lock_descr *descr;
 225                 descr = &ext->oe_osclock->cll_descr;
 226                 if (!(descr->cld_start <= ext->oe_start &&
 227                       descr->cld_end >= ext->oe_max_end))
 228                         GOTO(out, rc = 100);
 229         }
 230
 231         if (ext->oe_nr_pages > ext->oe_mppr)
 232                 GOTO(out, rc = 105);
 233
 234         /* Do not verify page list if extent is in RPC. This is because an
 235          * in-RPC extent is supposed to be exclusively accessible w/o lock. */
 236         if (ext->oe_state > OES_CACHE)
 237                 GOTO(out, rc = 0);
 238
 239         if (!extent_debug)
 240                 GOTO(out, rc = 0);
 241
 242         page_count = 0;
 243         cfs_list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
 244                 pgoff_t index = oap2cl_page(oap)->cp_index;
 245                 ++page_count;
 246                 if (index > ext->oe_end || index < ext->oe_start)
 247                         GOTO(out, rc = 110);
 248         }
 249         if (page_count != ext->oe_nr_pages)
 250                 GOTO(out, rc = 120);
 251
 252 out:
 253         if (rc != 0)
 254                 OSC_EXTENT_DUMP(D_ERROR, ext,
 255                                 "%s:%d sanity check %p failed with rc = %d\n",
 256                                 func, line, ext, rc);
 257         return rc;
 258 }
 259
 260 #define sanity_check_nolock(ext) \
 261         osc_extent_sanity_check0(ext, __func__, __LINE__)
 262
 263 #define sanity_check(ext) ({                                                   \
 264         int __res;                                                             \
 265         osc_object_lock((ext)->oe_obj);                                        \
 266         __res = sanity_check_nolock(ext);                                      \
 267         osc_object_unlock((ext)->oe_obj);                                      \
 268         __res;                                                                 \
 269 })
 270
 271
 272 /**
 273  * sanity check - to make sure there is no overlapped extent in the tree.
 274  */
 275 static int osc_extent_is_overlapped(struct osc_object *obj,
 276                                     struct osc_extent *ext)
 277 {
 278         struct osc_extent *tmp;
 279
 280         LASSERT(osc_object_is_locked(obj));
 281
 282         if (!extent_debug)
 283                 return 0;
 284
 285         for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
 286                 if (tmp == ext)
 287                         continue;
 288                 if (tmp->oe_end >= ext->oe_start &&
 289                     tmp->oe_start <= ext->oe_end)
 290                         return 1;
 291         }
 292         return 0;
 293 }
 294
 295 static void osc_extent_state_set(struct osc_extent *ext, int state)
 296 {
 297         LASSERT(osc_object_is_locked(ext->oe_obj));
 298         LASSERT(state >= OES_INV && state < OES_STATE_MAX);
 299
 300         /* Never try to sanity check a state changing extent :-) */
 301         /* LASSERT(sanity_check_nolock(ext) == 0); */
 302
 303         /* TODO: validate the state machine */
 304         ext->oe_state = state;
 305         wake_up_all(&ext->oe_waitq);
 306 }
 307
 308 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
 309 {
 310         struct osc_extent *ext;
 311
 312         OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_IOFS);
 313         if (ext == NULL)
 314                 return NULL;
 315
 316         RB_CLEAR_NODE(&ext->oe_node);
 317         ext->oe_obj = obj;
 318         cfs_atomic_set(&ext->oe_refc, 1);
 319         cfs_atomic_set(&ext->oe_users, 0);
 320         CFS_INIT_LIST_HEAD(&ext->oe_link);
 321         ext->oe_state = OES_INV;
 322         CFS_INIT_LIST_HEAD(&ext->oe_pages);
 323         init_waitqueue_head(&ext->oe_waitq);
 324         ext->oe_osclock = NULL;
 325
 326         return ext;
 327 }
 328
 329 static void osc_extent_free(struct osc_extent *ext)
 330 {
 331         OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
 332 }
 333
 334 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
 335 {
 336         LASSERT(cfs_atomic_read(&ext->oe_refc) >= 0);
 337         cfs_atomic_inc(&ext->oe_refc);
 338         return ext;
 339 }
 340
 341 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
 342 {
 343         LASSERT(cfs_atomic_read(&ext->oe_refc) > 0);
 344         if (cfs_atomic_dec_and_test(&ext->oe_refc)) {
 345                 LASSERT(cfs_list_empty(&ext->oe_link));
 346                 LASSERT(cfs_atomic_read(&ext->oe_users) == 0);
 347                 LASSERT(ext->oe_state == OES_INV);
 348                 LASSERT(!ext->oe_intree);
 349
 350                 if (ext->oe_osclock) {
 351                         cl_lock_put(env, ext->oe_osclock);
 352                         ext->oe_osclock = NULL;
 353                 }
 354                 osc_extent_free(ext);
 355         }
 356 }
 357
 358 /**
 359  * osc_extent_put_trust() is a special version of osc_extent_put() when
 360  * it's known that the caller is not the last user. This is to address the
 361  * problem of lacking of lu_env ;-).
 362  */
 363 static void osc_extent_put_trust(struct osc_extent *ext)
 364 {
 365         LASSERT(cfs_atomic_read(&ext->oe_refc) > 1);
 366         LASSERT(osc_object_is_locked(ext->oe_obj));
 367         cfs_atomic_dec(&ext->oe_refc);
 368 }
 369
 370 /**
 371  * Return the extent which includes pgoff @index, or return the greatest
 372  * previous extent in the tree.
 373  */
 374 static struct osc_extent *osc_extent_search(struct osc_object *obj,
 375                                             pgoff_t index)
 376 {
 377         struct rb_node    *n = obj->oo_root.rb_node;
 378         struct osc_extent *tmp, *p = NULL;
 379
 380         LASSERT(osc_object_is_locked(obj));
 381         while (n != NULL) {
 382                 tmp = rb_extent(n);
 383                 if (index < tmp->oe_start) {
 384                         n = n->rb_left;
 385                 } else if (index > tmp->oe_end) {
 386                         p = rb_extent(n);
 387                         n = n->rb_right;
 388                 } else {
 389                         return tmp;
 390                 }
 391         }
 392         return p;
 393 }
 394
 395 /*
 396  * Return the extent covering @index, otherwise return NULL.
 397  * caller must have held object lock.
 398  */
 399 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
 400                                             pgoff_t index)
 401 {
 402         struct osc_extent *ext;
 403
 404         ext = osc_extent_search(obj, index);
 405         if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
 406                 return osc_extent_get(ext);
 407         return NULL;
 408 }
 409
 410 /* caller must have held object lock. */
 411 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
 412 {
 413         struct rb_node   **n      = &obj->oo_root.rb_node;
 414         struct rb_node    *parent = NULL;
 415         struct osc_extent *tmp;
 416
 417         LASSERT(ext->oe_intree == 0);
 418         LASSERT(ext->oe_obj == obj);
 419         LASSERT(osc_object_is_locked(obj));
 420         while (*n != NULL) {
 421                 tmp = rb_extent(*n);
 422                 parent = *n;
 423
 424                 if (ext->oe_end < tmp->oe_start)
 425                         n = &(*n)->rb_left;
 426                 else if (ext->oe_start > tmp->oe_end)
 427                         n = &(*n)->rb_right;
 428                 else
 429                         EASSERTF(0, tmp, EXTSTR, EXTPARA(ext));
 430         }
 431         rb_link_node(&ext->oe_node, parent, n);
 432         rb_insert_color(&ext->oe_node, &obj->oo_root);
 433         osc_extent_get(ext);
 434         ext->oe_intree = 1;
 435 }
 436
 437 /* caller must have held object lock. */
 438 static void osc_extent_erase(struct osc_extent *ext)
 439 {
 440         struct osc_object *obj = ext->oe_obj;
 441         LASSERT(osc_object_is_locked(obj));
 442         if (ext->oe_intree) {
 443                 rb_erase(&ext->oe_node, &obj->oo_root);
 444                 ext->oe_intree = 0;
 445                 /* rbtree held a refcount */
 446                 osc_extent_put_trust(ext);
 447         }
 448 }
 449
 450 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
 451 {
 452         struct osc_object *obj = ext->oe_obj;
 453
 454         LASSERT(osc_object_is_locked(obj));
 455         LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
 456         if (ext->oe_state == OES_CACHE) {
 457                 osc_extent_state_set(ext, OES_ACTIVE);
 458                 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
 459         }
 460         cfs_atomic_inc(&ext->oe_users);
 461         cfs_list_del_init(&ext->oe_link);
 462         return osc_extent_get(ext);
 463 }
 464
 465 static void __osc_extent_remove(struct osc_extent *ext)
 466 {
 467         LASSERT(osc_object_is_locked(ext->oe_obj));
 468         LASSERT(cfs_list_empty(&ext->oe_pages));
 469         osc_extent_erase(ext);
 470         cfs_list_del_init(&ext->oe_link);
 471         osc_extent_state_set(ext, OES_INV);
 472         OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
 473 }
 474
 475 static void osc_extent_remove(struct osc_extent *ext)
 476 {
 477         struct osc_object *obj = ext->oe_obj;
 478
 479         osc_object_lock(obj);
 480         __osc_extent_remove(ext);
 481         osc_object_unlock(obj);
 482 }
 483
 484 /**
 485  * This function is used to merge extents to get better performance. It checks
 486  * if @cur and @victim are contiguous at chunk level.
 487  */
 488 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
 489                             struct osc_extent *victim)
 490 {
 491         struct osc_object *obj = cur->oe_obj;
 492         pgoff_t chunk_start;
 493         pgoff_t chunk_end;
 494         int ppc_bits;
 495
 496         LASSERT(cur->oe_state == OES_CACHE);
 497         LASSERT(osc_object_is_locked(obj));
 498         if (victim == NULL)
 499                 return -EINVAL;
 500
 501         if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
 502                 return -EBUSY;
 503
 504         if (cur->oe_max_end != victim->oe_max_end)
 505                 return -ERANGE;
 506
 507         LASSERT(cur->oe_osclock == victim->oe_osclock);
 508         ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_CACHE_SHIFT;
 509         chunk_start = cur->oe_start >> ppc_bits;
 510         chunk_end   = cur->oe_end   >> ppc_bits;
 511         if (chunk_start   != (victim->oe_end >> ppc_bits) + 1 &&
 512             chunk_end + 1 != victim->oe_start >> ppc_bits)
 513                 return -ERANGE;
 514
 515         OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
 516
 517         cur->oe_start     = min(cur->oe_start, victim->oe_start);
 518         cur->oe_end       = max(cur->oe_end,   victim->oe_end);
 519         cur->oe_grants   += victim->oe_grants;
 520         cur->oe_nr_pages += victim->oe_nr_pages;
 521         /* only the following bits are needed to merge */
 522         cur->oe_urgent   |= victim->oe_urgent;
 523         cur->oe_memalloc |= victim->oe_memalloc;
 524         cfs_list_splice_init(&victim->oe_pages, &cur->oe_pages);
 525         cfs_list_del_init(&victim->oe_link);
 526         victim->oe_nr_pages = 0;
 527
 528         osc_extent_get(victim);
 529         __osc_extent_remove(victim);
 530         osc_extent_put(env, victim);
 531
 532         OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
 533         return 0;
 534 }
 535
 536 /**
 537  * Drop user count of osc_extent, and unplug IO asynchronously.
 538  */
 539 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
 540 {
 541         struct osc_object *obj = ext->oe_obj;
 542         int rc = 0;
 543         ENTRY;
 544
 545         LASSERT(cfs_atomic_read(&ext->oe_users) > 0);
 546         LASSERT(sanity_check(ext) == 0);
 547         LASSERT(ext->oe_grants > 0);
 548
 549         if (cfs_atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
 550                 LASSERT(ext->oe_state == OES_ACTIVE);
 551                 if (ext->oe_trunc_pending) {
 552                         /* a truncate process is waiting for this extent.
 553                          * This may happen due to a race, check
 554                          * osc_cache_truncate_start(). */
 555                         osc_extent_state_set(ext, OES_TRUNC);
 556                         ext->oe_trunc_pending = 0;
 557                 } else {
 558                         osc_extent_state_set(ext, OES_CACHE);
 559                         osc_update_pending(obj, OBD_BRW_WRITE,
 560                                            ext->oe_nr_pages);
 561
 562                         /* try to merge the previous and next extent. */
 563                         osc_extent_merge(env, ext, prev_extent(ext));
 564                         osc_extent_merge(env, ext, next_extent(ext));
 565
 566                         if (ext->oe_urgent)
 567                                 cfs_list_move_tail(&ext->oe_link,
 568                                                    &obj->oo_urgent_exts);
 569                 }
 570                 osc_object_unlock(obj);
 571
 572                 osc_io_unplug_async(env, osc_cli(obj), obj);
 573         }
 574         osc_extent_put(env, ext);
 575         RETURN(rc);
 576 }
 577
 578 static inline int overlapped(struct osc_extent *ex1, struct osc_extent *ex2)
 579 {
 580         return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
 581 }
 582
 583 /**
 584  * Find or create an extent which includes @index, core function to manage
 585  * extent tree.
 586  */
 587 struct osc_extent *osc_extent_find(const struct lu_env *env,
 588                                    struct osc_object *obj, pgoff_t index,
 589                                    int *grants)
 590
 591 {
 592         struct client_obd *cli = osc_cli(obj);
 593         struct cl_lock    *lock;
 594         struct osc_extent *cur;
 595         struct osc_extent *ext;
 596         struct osc_extent *conflict = NULL;
 597         struct osc_extent *found = NULL;
 598         pgoff_t    chunk;
 599         pgoff_t    max_end;
 600         int        max_pages; /* max_pages_per_rpc */
 601         int        chunksize;
 602         int        ppc_bits; /* pages per chunk bits */
 603         int        chunk_mask;
 604         int        rc;
 605         ENTRY;
 606
 607         cur = osc_extent_alloc(obj);
 608         if (cur == NULL)
 609                 RETURN(ERR_PTR(-ENOMEM));
 610
 611         lock = cl_lock_at_pgoff(env, osc2cl(obj), index, NULL, 1, 0);
 612         LASSERT(lock != NULL);
 613         LASSERT(lock->cll_descr.cld_mode >= CLM_WRITE);
 614
 615         LASSERT(cli->cl_chunkbits >= PAGE_CACHE_SHIFT);
 616         ppc_bits   = cli->cl_chunkbits - PAGE_CACHE_SHIFT;
 617         chunk_mask = ~((1 << ppc_bits) - 1);
 618         chunksize  = 1 << cli->cl_chunkbits;
 619         chunk      = index >> ppc_bits;
 620
 621         /* align end to rpc edge, rpc size may not be a power 2 integer. */
 622         max_pages = cli->cl_max_pages_per_rpc;
 623         LASSERT((max_pages & ~chunk_mask) == 0);
 624         max_end = index - (index % max_pages) + max_pages - 1;
 625         max_end = min_t(pgoff_t, max_end, lock->cll_descr.cld_end);
 626
 627         /* initialize new extent by parameters so far */
 628         cur->oe_max_end = max_end;
 629         cur->oe_start   = index & chunk_mask;
 630         cur->oe_end     = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
 631         if (cur->oe_start < lock->cll_descr.cld_start)
 632                 cur->oe_start = lock->cll_descr.cld_start;
 633         if (cur->oe_end > max_end)
 634                 cur->oe_end = max_end;
 635         cur->oe_osclock = lock;
 636         cur->oe_grants  = 0;
 637         cur->oe_mppr    = max_pages;
 638
 639         /* grants has been allocated by caller */
 640         LASSERTF(*grants >= chunksize + cli->cl_extent_tax,
 641                  "%u/%u/%u.\n", *grants, chunksize, cli->cl_extent_tax);
 642         LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR, EXTPARA(cur));
 643
 644 restart:
 645         osc_object_lock(obj);
 646         ext = osc_extent_search(obj, cur->oe_start);
 647         if (ext == NULL)
 648                 ext = first_extent(obj);
 649         while (ext != NULL) {
 650                 loff_t ext_chk_start = ext->oe_start >> ppc_bits;
 651                 loff_t ext_chk_end   = ext->oe_end   >> ppc_bits;
 652
 653                 LASSERT(sanity_check_nolock(ext) == 0);
 654                 if (chunk > ext_chk_end + 1)
 655                         break;
 656
 657                 /* if covering by different locks, no chance to match */
 658                 if (lock != ext->oe_osclock) {
 659                         EASSERTF(!overlapped(ext, cur), ext,
 660                                  EXTSTR, EXTPARA(cur));
 661
 662                         ext = next_extent(ext);
 663                         continue;
 664                 }
 665
 666                 /* discontiguous chunks? */
 667                 if (chunk + 1 < ext_chk_start) {
 668                         ext = next_extent(ext);
 669                         continue;
 670                 }
 671
 672                 /* ok, from now on, ext and cur have these attrs:
 673                  * 1. covered by the same lock
 674                  * 2. contiguous at chunk level or overlapping. */
 675
 676                 if (overlapped(ext, cur)) {
 677                         /* cur is the minimum unit, so overlapping means
 678                          * full contain. */
 679                         EASSERTF((ext->oe_start <= cur->oe_start &&
 680                                   ext->oe_end >= cur->oe_end),
 681                                  ext, EXTSTR, EXTPARA(cur));
 682
 683                         if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
 684                                 /* for simplicity, we wait for this extent to
 685                                  * finish before going forward. */
 686                                 conflict = osc_extent_get(ext);
 687                                 break;
 688                         }
 689
 690                         found = osc_extent_hold(ext);
 691                         break;
 692                 }
 693
 694                 /* non-overlapped extent */
 695                 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
 696                         /* we can't do anything for a non OES_CACHE extent, or
 697                          * if there is someone waiting for this extent to be
 698                          * flushed, try next one. */
 699                         ext = next_extent(ext);
 700                         continue;
 701                 }
 702
 703                 /* check if they belong to the same rpc slot before trying to
 704                  * merge. the extents are not overlapped and contiguous at
 705                  * chunk level to get here. */
 706                 if (ext->oe_max_end != max_end) {
 707                         /* if they don't belong to the same RPC slot or
 708                          * max_pages_per_rpc has ever changed, do not merge. */
 709                         ext = next_extent(ext);
 710                         continue;
 711                 }
 712
 713                 /* it's required that an extent must be contiguous at chunk
 714                  * level so that we know the whole extent is covered by grant
 715                  * (the pages in the extent are NOT required to be contiguous).
 716                  * Otherwise, it will be too much difficult to know which
 717                  * chunks have grants allocated. */
 718
 719                 /* try to do front merge - extend ext's start */
 720                 if (chunk + 1 == ext_chk_start) {
 721                         /* ext must be chunk size aligned */
 722                         EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
 723
 724                         /* pull ext's start back to cover cur */
 725                         ext->oe_start   = cur->oe_start;
 726                         ext->oe_grants += chunksize;
 727                         *grants -= chunksize;
 728
 729                         found = osc_extent_hold(ext);
 730                 } else if (chunk == ext_chk_end + 1) {
 731                         /* rear merge */
 732                         ext->oe_end     = cur->oe_end;
 733                         ext->oe_grants += chunksize;
 734                         *grants -= chunksize;
 735
 736                         /* try to merge with the next one because we just fill
 737                          * in a gap */
 738                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 739                                 /* we can save extent tax from next extent */
 740                                 *grants += cli->cl_extent_tax;
 741
 742                         found = osc_extent_hold(ext);
 743                 }
 744                 if (found != NULL)
 745                         break;
 746
 747                 ext = next_extent(ext);
 748         }
 749
 750         osc_extent_tree_dump(D_CACHE, obj);
 751         if (found != NULL) {
 752                 LASSERT(conflict == NULL);
 753                 if (!IS_ERR(found)) {
 754                         LASSERT(found->oe_osclock == cur->oe_osclock);
 755                         OSC_EXTENT_DUMP(D_CACHE, found,
 756                                         "found caching ext for %lu.\n", index);
 757                 }
 758         } else if (conflict == NULL) {
 759                 /* create a new extent */
 760                 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
 761                 cur->oe_grants = chunksize + cli->cl_extent_tax;
 762                 *grants -= cur->oe_grants;
 763                 LASSERT(*grants >= 0);
 764
 765                 cur->oe_state = OES_CACHE;
 766                 found = osc_extent_hold(cur);
 767                 osc_extent_insert(obj, cur);
 768                 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
 769                                 index, lock->cll_descr.cld_end);
 770         }
 771         osc_object_unlock(obj);
 772
 773         if (conflict != NULL) {
 774                 LASSERT(found == NULL);
 775
 776                 /* waiting for IO to finish. Please notice that it's impossible
 777                  * to be an OES_TRUNC extent. */
 778                 rc = osc_extent_wait(env, conflict, OES_INV);
 779                 osc_extent_put(env, conflict);
 780                 conflict = NULL;
 781                 if (rc < 0)
 782                         GOTO(out, found = ERR_PTR(rc));
 783
 784                 goto restart;
 785         }
 786         EXIT;
 787
 788 out:
 789         osc_extent_put(env, cur);
 790         LASSERT(*grants >= 0);
 791         return found;
 792 }
 793
 794 /**
 795  * Called when IO is finished to an extent.
 796  */
 797 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
 798                       int sent, int rc)
 799 {
 800         struct client_obd *cli = osc_cli(ext->oe_obj);
 801         struct osc_async_page *oap;
 802         struct osc_async_page *tmp;
 803         int nr_pages = ext->oe_nr_pages;
 804         int lost_grant = 0;
 805         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
 806         __u64 last_off = 0;
 807         int last_count = -1;
 808         ENTRY;
 809
 810         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
 811
 812         ext->oe_rc = rc ?: ext->oe_nr_pages;
 813         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
 814         cfs_list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 815                                      oap_pending_item) {
 816                 cfs_list_del_init(&oap->oap_rpc_item);
 817                 cfs_list_del_init(&oap->oap_pending_item);
 818                 if (last_off <= oap->oap_obj_off) {
 819                         last_off = oap->oap_obj_off;
 820                         last_count = oap->oap_count;
 821                 }
 822
 823                 --ext->oe_nr_pages;
 824                 osc_ap_completion(env, cli, oap, sent, rc);
 825         }
 826         EASSERT(ext->oe_nr_pages == 0, ext);
 827
 828         if (!sent) {
 829                 lost_grant = ext->oe_grants;
 830         } else if (blocksize < PAGE_CACHE_SIZE &&
 831                    last_count != PAGE_CACHE_SIZE) {
 832                 /* For short writes we shouldn't count parts of pages that
 833                  * span a whole chunk on the OST side, or our accounting goes
 834                  * wrong.  Should match the code in filter_grant_check. */
 835                 int offset = oap->oap_page_off & ~CFS_PAGE_MASK;
 836                 int count = oap->oap_count + (offset & (blocksize - 1));
 837                 int end = (offset + oap->oap_count) & (blocksize - 1);
 838                 if (end)
 839                         count += blocksize - end;
 840
 841                 lost_grant = PAGE_CACHE_SIZE - count;
 842         }
 843         if (ext->oe_grants > 0)
 844                 osc_free_grant(cli, nr_pages, lost_grant);
 845
 846         osc_extent_remove(ext);
 847         /* put the refcount for RPC */
 848         osc_extent_put(env, ext);
 849         RETURN(0);
 850 }
 851
 852 static int extent_wait_cb(struct osc_extent *ext, int state)
 853 {
 854         int ret;
 855
 856         osc_object_lock(ext->oe_obj);
 857         ret = ext->oe_state == state;
 858         osc_object_unlock(ext->oe_obj);
 859
 860         return ret;
 861 }
 862
 863 /**
 864  * Wait for the extent's state to become @state.
 865  */
 866 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
 867                            int state)
 868 {
 869         struct osc_object *obj = ext->oe_obj;
 870         struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
 871                                                   LWI_ON_SIGNAL_NOOP, NULL);
 872         int rc = 0;
 873         ENTRY;
 874
 875         osc_object_lock(obj);
 876         LASSERT(sanity_check_nolock(ext) == 0);
 877         /* `Kick' this extent only if the caller is waiting for it to be
 878          * written out. */
 879         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
 880                 if (ext->oe_state == OES_ACTIVE) {
 881                         ext->oe_urgent = 1;
 882                 } else if (ext->oe_state == OES_CACHE) {
 883                         ext->oe_urgent = 1;
 884                         osc_extent_hold(ext);
 885                         rc = 1;
 886                 }
 887         }
 888         osc_object_unlock(obj);
 889         if (rc == 1)
 890                 osc_extent_release(env, ext);
 891
 892         /* wait for the extent until its state becomes @state */
 893         rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
 894         if (rc == -ETIMEDOUT) {
 895                 OSC_EXTENT_DUMP(D_ERROR, ext,
 896                         "%s: wait ext to %d timedout, recovery in progress?\n",
 897                         osc_export(obj)->exp_obd->obd_name, state);
 898
 899                 lwi = LWI_INTR(NULL, NULL);
 900                 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state),
 901                                   &lwi);
 902         }
 903         if (rc == 0 && ext->oe_rc < 0)
 904                 rc = ext->oe_rc;
 905         RETURN(rc);
 906 }
 907
 908 /**
 909  * Discard pages with index greater than @size. If @ext is overlapped with
 910  * @size, then partial truncate happens.
 911  */
 912 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
 913                                 bool partial)
 914 {
 915         struct cl_env_nest     nest;
 916         struct lu_env         *env;
 917         struct cl_io          *io;
 918         struct osc_object     *obj = ext->oe_obj;
 919         struct client_obd     *cli = osc_cli(obj);
 920         struct osc_async_page *oap;
 921         struct osc_async_page *tmp;
 922         int                    pages_in_chunk = 0;
 923         int                    ppc_bits    = cli->cl_chunkbits -
 924                                              PAGE_CACHE_SHIFT;
 925         __u64                  trunc_chunk = trunc_index >> ppc_bits;
 926         int                    grants   = 0;
 927         int                    nr_pages = 0;
 928         int                    rc       = 0;
 929         ENTRY;
 930
 931         LASSERT(sanity_check(ext) == 0);
 932         LASSERT(ext->oe_state == OES_TRUNC);
 933         LASSERT(!ext->oe_urgent);
 934
 935         /* Request new lu_env.
 936          * We can't use that env from osc_cache_truncate_start() because
 937          * it's from lov_io_sub and not fully initialized. */
 938         env = cl_env_nested_get(&nest);
 939         io  = &osc_env_info(env)->oti_io;
 940         io->ci_obj = cl_object_top(osc2cl(obj));
 941         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
 942         if (rc < 0)
 943                 GOTO(out, rc);
 944
 945         /* discard all pages with index greater then trunc_index */
 946         cfs_list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 947                                      oap_pending_item) {
 948                 struct cl_page  *sub  = oap2cl_page(oap);
 949                 struct cl_page  *page = cl_page_top(sub);
 950
 951                 LASSERT(cfs_list_empty(&oap->oap_rpc_item));
 952
 953                 /* only discard the pages with their index greater than
 954                  * trunc_index, and ... */
 955                 if (sub->cp_index < trunc_index ||
 956                     (sub->cp_index == trunc_index && partial)) {
 957                         /* accounting how many pages remaining in the chunk
 958                          * so that we can calculate grants correctly. */
 959                         if (sub->cp_index >> ppc_bits == trunc_chunk)
 960                                 ++pages_in_chunk;
 961                         continue;
 962                 }
 963
 964                 cfs_list_del_init(&oap->oap_pending_item);
 965
 966                 cl_page_get(page);
 967                 lu_ref_add(&page->cp_reference, "truncate", current);
 968
 969                 if (cl_page_own(env, io, page) == 0) {
 970                         cl_page_unmap(env, io, page);
 971                         cl_page_discard(env, io, page);
 972                         cl_page_disown(env, io, page);
 973                 } else {
 974                         LASSERT(page->cp_state == CPS_FREEING);
 975                         LASSERT(0);
 976                 }
 977
 978                 lu_ref_del(&page->cp_reference, "truncate", current);
 979                 cl_page_put(env, page);
 980
 981                 --ext->oe_nr_pages;
 982                 ++nr_pages;
 983         }
 984         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
 985                       ext->oe_nr_pages == 0),
 986                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
 987
 988         osc_object_lock(obj);
 989         if (ext->oe_nr_pages == 0) {
 990                 LASSERT(pages_in_chunk == 0);
 991                 grants = ext->oe_grants;
 992                 ext->oe_grants = 0;
 993         } else { /* calculate how many grants we can free */
 994                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
 995                 pgoff_t last_index;
 996
 997
 998                 /* if there is no pages in this chunk, we can also free grants
 999                  * for the last chunk */
1000                 if (pages_in_chunk == 0) {
1001                         /* if this is the 1st chunk and no pages in this chunk,
1002                          * ext->oe_nr_pages must be zero, so we should be in
1003                          * the other if-clause. */
1004                         LASSERT(trunc_chunk > 0);
1005                         --trunc_chunk;
1006                         ++chunks;
1007                 }
1008
1009                 /* this is what we can free from this extent */
1010                 grants          = chunks << cli->cl_chunkbits;
1011                 ext->oe_grants -= grants;
1012                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1013                 ext->oe_end     = min(last_index, ext->oe_max_end);
1014                 LASSERT(ext->oe_end >= ext->oe_start);
1015                 LASSERT(ext->oe_grants > 0);
1016         }
1017         osc_object_unlock(obj);
1018
1019         if (grants > 0 || nr_pages > 0)
1020                 osc_free_grant(cli, nr_pages, grants);
1021
1022 out:
1023         cl_io_fini(env, io);
1024         cl_env_nested_put(&nest, env);
1025         RETURN(rc);
1026 }
1027
1028 /**
1029  * This function is used to make the extent prepared for transfer.
1030  * A race with flusing page - ll_writepage() has to be handled cautiously.
1031  */
1032 static int osc_extent_make_ready(const struct lu_env *env,
1033                                  struct osc_extent *ext)
1034 {
1035         struct osc_async_page *oap;
1036         struct osc_async_page *last = NULL;
1037         struct osc_object *obj = ext->oe_obj;
1038         int page_count = 0;
1039         int rc;
1040         ENTRY;
1041
1042         /* we're going to grab page lock, so object lock must not be taken. */
1043         LASSERT(sanity_check(ext) == 0);
1044         /* in locking state, any process should not touch this extent. */
1045         EASSERT(ext->oe_state == OES_LOCKING, ext);
1046         EASSERT(ext->oe_owner != NULL, ext);
1047
1048         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1049
1050         cfs_list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1051                 ++page_count;
1052                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1053                         last = oap;
1054
1055                 /* checking ASYNC_READY is race safe */
1056                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1057                         continue;
1058
1059                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1060                 switch (rc) {
1061                 case 0:
1062                         spin_lock(&oap->oap_lock);
1063                         oap->oap_async_flags |= ASYNC_READY;
1064                         spin_unlock(&oap->oap_lock);
1065                         break;
1066                 case -EALREADY:
1067                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1068                         break;
1069                 default:
1070                         LASSERTF(0, "unknown return code: %d\n", rc);
1071                 }
1072         }
1073
1074         LASSERT(page_count == ext->oe_nr_pages);
1075         LASSERT(last != NULL);
1076         /* the last page is the only one we need to refresh its count by
1077          * the size of file. */
1078         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1079                 last->oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1080                 LASSERT(last->oap_count > 0);
1081                 LASSERT(last->oap_page_off + last->oap_count <= PAGE_CACHE_SIZE);
1082                 spin_lock(&last->oap_lock);
1083                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1084                 spin_unlock(&last->oap_lock);
1085         }
1086
1087         /* for the rest of pages, we don't need to call osf_refresh_count()
1088          * because it's known they are not the last page */
1089         cfs_list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1090                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1091                         oap->oap_count = PAGE_CACHE_SIZE - oap->oap_page_off;
1092                         spin_lock(&oap->oap_lock);
1093                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1094                         spin_unlock(&oap->oap_lock);
1095                 }
1096         }
1097
1098         osc_object_lock(obj);
1099         osc_extent_state_set(ext, OES_RPC);
1100         osc_object_unlock(obj);
1101         /* get a refcount for RPC. */
1102         osc_extent_get(ext);
1103
1104         RETURN(0);
1105 }
1106
1107 /**
1108  * Quick and simple version of osc_extent_find(). This function is frequently
1109  * called to expand the extent for the same IO. To expand the extent, the
1110  * page index must be in the same or next chunk of ext->oe_end.
1111  */
1112 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index, int *grants)
1113 {
1114         struct osc_object *obj = ext->oe_obj;
1115         struct client_obd *cli = osc_cli(obj);
1116         struct osc_extent *next;
1117         int ppc_bits = cli->cl_chunkbits - PAGE_CACHE_SHIFT;
1118         pgoff_t chunk = index >> ppc_bits;
1119         pgoff_t end_chunk;
1120         pgoff_t end_index;
1121         int chunksize = 1 << cli->cl_chunkbits;
1122         int rc = 0;
1123         ENTRY;
1124
1125         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1126         osc_object_lock(obj);
1127         LASSERT(sanity_check_nolock(ext) == 0);
1128         end_chunk = ext->oe_end >> ppc_bits;
1129         if (chunk > end_chunk + 1)
1130                 GOTO(out, rc = -ERANGE);
1131
1132         if (end_chunk >= chunk)
1133                 GOTO(out, rc = 0);
1134
1135         LASSERT(end_chunk + 1 == chunk);
1136         /* try to expand this extent to cover @index */
1137         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1138
1139         next = next_extent(ext);
1140         if (next != NULL && next->oe_start <= end_index)
1141                 /* complex mode - overlapped with the next extent,
1142                  * this case will be handled by osc_extent_find() */
1143                 GOTO(out, rc = -EAGAIN);
1144
1145         ext->oe_end = end_index;
1146         ext->oe_grants += chunksize;
1147         *grants -= chunksize;
1148         LASSERT(*grants >= 0);
1149         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1150                  "overlapped after expanding for %lu.\n", index);
1151         EXIT;
1152
1153 out:
1154         osc_object_unlock(obj);
1155         RETURN(rc);
1156 }
1157
1158 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
1159                                   const char *func, int line)
1160 {
1161         struct osc_extent *ext;
1162         int cnt;
1163
1164         CDEBUG(level, "Dump object %p extents at %s:%d, mppr: %u.\n",
1165                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1166
1167         /* osc_object_lock(obj); */
1168         cnt = 1;
1169         for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1170                 OSC_EXTENT_DUMP(level, ext, "in tree %d.\n", cnt++);
1171
1172         cnt = 1;
1173         cfs_list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1174                 OSC_EXTENT_DUMP(level, ext, "hp %d.\n", cnt++);
1175
1176         cnt = 1;
1177         cfs_list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1178                 OSC_EXTENT_DUMP(level, ext, "urgent %d.\n", cnt++);
1179
1180         cnt = 1;
1181         cfs_list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1182                 OSC_EXTENT_DUMP(level, ext, "reading %d.\n", cnt++);
1183         /* osc_object_unlock(obj); */
1184 }
1185
1186 /* ------------------ osc extent end ------------------ */
1187
1188 static inline int osc_is_ready(struct osc_object *osc)
1189 {
1190         return !cfs_list_empty(&osc->oo_ready_item) ||
1191                !cfs_list_empty(&osc->oo_hp_ready_item);
1192 }
1193
1194 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1195         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1196                (OSC), osc_is_ready(OSC),                                       \
1197                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1198                list_empty_marker(&(OSC)->oo_ready_item),                       \
1199                cfs_atomic_read(&(OSC)->oo_nr_writes),                          \
1200                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1201                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1202                cfs_atomic_read(&(OSC)->oo_nr_reads),                           \
1203                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1204                ##args)
1205
1206 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1207                           int cmd)
1208 {
1209         struct osc_page *opg  = oap2osc_page(oap);
1210         struct cl_page  *page = cl_page_top(oap2cl_page(oap));
1211         int result;
1212
1213         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1214
1215         ENTRY;
1216         result = cl_page_make_ready(env, page, CRT_WRITE);
1217         if (result == 0)
1218                 opg->ops_submit_time = cfs_time_current();
1219         RETURN(result);
1220 }
1221
1222 static int osc_refresh_count(const struct lu_env *env,
1223                              struct osc_async_page *oap, int cmd)
1224 {
1225         struct osc_page  *opg = oap2osc_page(oap);
1226         struct cl_page   *page = oap2cl_page(oap);
1227         struct cl_object *obj;
1228         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
1229
1230         int result;
1231         loff_t kms;
1232
1233         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1234         LASSERT(!(cmd & OBD_BRW_READ));
1235         LASSERT(opg != NULL);
1236         obj = opg->ops_cl.cpl_obj;
1237
1238         cl_object_attr_lock(obj);
1239         result = cl_object_attr_get(env, obj, attr);
1240         cl_object_attr_unlock(obj);
1241         if (result < 0)
1242                 return result;
1243         kms = attr->cat_kms;
1244         if (cl_offset(obj, page->cp_index) >= kms)
1245                 /* catch race with truncate */
1246                 return 0;
1247         else if (cl_offset(obj, page->cp_index + 1) > kms)
1248                 /* catch sub-page write at end of file */
1249                 return kms % PAGE_CACHE_SIZE;
1250         else
1251                 return PAGE_CACHE_SIZE;
1252 }
1253
1254 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1255                           int cmd, int rc)
1256 {
1257         struct osc_page   *opg  = oap2osc_page(oap);
1258         struct cl_page    *page = cl_page_top(oap2cl_page(oap));
1259         struct osc_object *obj  = cl2osc(opg->ops_cl.cpl_obj);
1260         enum cl_req_type   crt;
1261         int srvlock;
1262
1263         ENTRY;
1264
1265         cmd &= ~OBD_BRW_NOQUOTA;
1266         LASSERTF(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ),
1267                  "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1268         LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1269                 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1270         LASSERT(opg->ops_transfer_pinned);
1271
1272         /*
1273          * page->cp_req can be NULL if io submission failed before
1274          * cl_req was allocated.
1275          */
1276         if (page->cp_req != NULL)
1277                 cl_req_page_done(env, page);
1278         LASSERT(page->cp_req == NULL);
1279
1280         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1281         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1282         opg->ops_transfer_pinned = 0;
1283
1284         spin_lock(&obj->oo_seatbelt);
1285         LASSERT(opg->ops_submitter != NULL);
1286         LASSERT(!cfs_list_empty(&opg->ops_inflight));
1287         cfs_list_del_init(&opg->ops_inflight);
1288         opg->ops_submitter = NULL;
1289         spin_unlock(&obj->oo_seatbelt);
1290
1291         opg->ops_submit_time = 0;
1292         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1293
1294         /* statistic */
1295         if (rc == 0 && srvlock) {
1296                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1297                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1298                 int bytes = oap->oap_count;
1299
1300                 if (crt == CRT_READ)
1301                         stats->os_lockless_reads += bytes;
1302                 else
1303                         stats->os_lockless_writes += bytes;
1304         }
1305
1306         /*
1307          * This has to be the last operation with the page, as locks are
1308          * released in cl_page_completion() and nothing except for the
1309          * reference counter protects page from concurrent reclaim.
1310          */
1311         lu_ref_del(&page->cp_reference, "transfer", page);
1312
1313         cl_page_completion(env, page, crt, rc);
1314
1315         RETURN(0);
1316 }
1317
1318 #define OSC_DUMP_GRANT(lvl, cli, fmt, args...) do {                           \
1319         struct client_obd *__tmp = (cli);                                     \
1320         CDEBUG(lvl, "%s: grant { dirty: %ld/%ld dirty_pages: %d/%d "          \
1321                "unstable_pages: %d/%d dropped: %ld avail: %ld, "              \
1322                "reserved: %ld, flight: %d } lru {in list: %d, "               \
1323                "left: %d, waiters: %d }" fmt,                                 \
1324                __tmp->cl_import->imp_obd->obd_name,                           \
1325                __tmp->cl_dirty, __tmp->cl_dirty_max,                          \
1326                cfs_atomic_read(&obd_dirty_pages), obd_max_dirty_pages,        \
1327                cfs_atomic_read(&obd_unstable_pages), obd_max_dirty_pages,     \
1328                __tmp->cl_lost_grant, __tmp->cl_avail_grant,                   \
1329                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,               \
1330                cfs_atomic_read(&__tmp->cl_lru_in_list),                       \
1331                cfs_atomic_read(&__tmp->cl_lru_busy),                          \
1332                cfs_atomic_read(&__tmp->cl_lru_shrinkers), ##args);            \
1333 } while (0)
1334
1335 /* caller must hold loi_list_lock */
1336 static void osc_consume_write_grant(struct client_obd *cli,
1337                                     struct brw_page *pga)
1338 {
1339         LASSERT(spin_is_locked(&cli->cl_loi_list_lock.lock));
1340         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1341         cfs_atomic_inc(&obd_dirty_pages);
1342         cli->cl_dirty += PAGE_CACHE_SIZE;
1343         pga->flag |= OBD_BRW_FROM_GRANT;
1344         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1345                PAGE_CACHE_SIZE, pga, pga->pg);
1346         osc_update_next_shrink(cli);
1347 }
1348
1349 /* the companion to osc_consume_write_grant, called when a brw has completed.
1350  * must be called with the loi lock held. */
1351 static void osc_release_write_grant(struct client_obd *cli,
1352                                     struct brw_page *pga)
1353 {
1354         ENTRY;
1355
1356         LASSERT(spin_is_locked(&cli->cl_loi_list_lock.lock));
1357         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1358                 EXIT;
1359                 return;
1360         }
1361
1362         pga->flag &= ~OBD_BRW_FROM_GRANT;
1363         cfs_atomic_dec(&obd_dirty_pages);
1364         cli->cl_dirty -= PAGE_CACHE_SIZE;
1365         if (pga->flag & OBD_BRW_NOCACHE) {
1366                 pga->flag &= ~OBD_BRW_NOCACHE;
1367                 cfs_atomic_dec(&obd_dirty_transit_pages);
1368                 cli->cl_dirty_transit -= PAGE_CACHE_SIZE;
1369         }
1370         EXIT;
1371 }
1372
1373 /**
1374  * To avoid sleeping with object lock held, it's good for us allocate enough
1375  * grants before entering into critical section.
1376  *
1377  * client_obd_list_lock held by caller
1378  */
1379 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1380 {
1381         int rc = -EDQUOT;
1382
1383         if (cli->cl_avail_grant >= bytes) {
1384                 cli->cl_avail_grant    -= bytes;
1385                 cli->cl_reserved_grant += bytes;
1386                 rc = 0;
1387         }
1388         return rc;
1389 }
1390
1391 static void __osc_unreserve_grant(struct client_obd *cli,
1392                                   unsigned int reserved, unsigned int unused)
1393 {
1394         /* it's quite normal for us to get more grant than reserved.
1395          * Thinking about a case that two extents merged by adding a new
1396          * chunk, we can save one extent tax. If extent tax is greater than
1397          * one chunk, we can save more grant by adding a new chunk */
1398         cli->cl_reserved_grant -= reserved;
1399         if (unused > reserved) {
1400                 cli->cl_avail_grant += reserved;
1401                 cli->cl_lost_grant  += unused - reserved;
1402         } else {
1403                 cli->cl_avail_grant += unused;
1404         }
1405 }
1406
1407 void osc_unreserve_grant(struct client_obd *cli,
1408                          unsigned int reserved, unsigned int unused)
1409 {
1410         client_obd_list_lock(&cli->cl_loi_list_lock);
1411         __osc_unreserve_grant(cli, reserved, unused);
1412         if (unused > 0)
1413                 osc_wake_cache_waiters(cli);
1414         client_obd_list_unlock(&cli->cl_loi_list_lock);
1415 }
1416
1417 /**
1418  * Free grant after IO is finished or canceled.
1419  *
1420  * @lost_grant is used to remember how many grants we have allocated but not
1421  * used, we should return these grants to OST. There're two cases where grants
1422  * can be lost:
1423  * 1. truncate;
1424  * 2. blocksize at OST is less than PAGE_CACHE_SIZE and a partial page was
1425  *    written. In this case OST may use less chunks to serve this partial
1426  *    write. OSTs don't actually know the page size on the client side. so
1427  *    clients have to calculate lost grant by the blocksize on the OST.
1428  *    See filter_grant_check() for details.
1429  */
1430 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1431                            unsigned int lost_grant)
1432 {
1433         int grant = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
1434
1435         client_obd_list_lock(&cli->cl_loi_list_lock);
1436         cfs_atomic_sub(nr_pages, &obd_dirty_pages);
1437         cli->cl_dirty -= nr_pages << PAGE_CACHE_SHIFT;
1438         cli->cl_lost_grant += lost_grant;
1439         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1440                 /* borrow some grant from truncate to avoid the case that
1441                  * truncate uses up all avail grant */
1442                 cli->cl_lost_grant -= grant;
1443                 cli->cl_avail_grant += grant;
1444         }
1445         osc_wake_cache_waiters(cli);
1446         client_obd_list_unlock(&cli->cl_loi_list_lock);
1447         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu\n",
1448                lost_grant, cli->cl_lost_grant,
1449                cli->cl_avail_grant, cli->cl_dirty);
1450 }
1451
1452 /**
1453  * The companion to osc_enter_cache(), called when @oap is no longer part of
1454  * the dirty accounting due to error.
1455  */
1456 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1457 {
1458         client_obd_list_lock(&cli->cl_loi_list_lock);
1459         osc_release_write_grant(cli, &oap->oap_brw_page);
1460         client_obd_list_unlock(&cli->cl_loi_list_lock);
1461 }
1462
1463 /**
1464  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1465  * is available.
1466  */
1467 static int osc_enter_cache_try(struct client_obd *cli,
1468                                struct osc_async_page *oap,
1469                                int bytes, int transient)
1470 {
1471         int rc;
1472
1473         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d.\n", bytes);
1474
1475         rc = osc_reserve_grant(cli, bytes);
1476         if (rc < 0)
1477                 return 0;
1478
1479         if (cli->cl_dirty + PAGE_CACHE_SIZE <= cli->cl_dirty_max &&
1480             cfs_atomic_read(&obd_unstable_pages) + 1 +
1481             cfs_atomic_read(&obd_dirty_pages) <= obd_max_dirty_pages) {
1482                 osc_consume_write_grant(cli, &oap->oap_brw_page);
1483                 if (transient) {
1484                         cli->cl_dirty_transit += PAGE_CACHE_SIZE;
1485                         cfs_atomic_inc(&obd_dirty_transit_pages);
1486                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
1487                 }
1488                 rc = 1;
1489         } else {
1490                 __osc_unreserve_grant(cli, bytes, bytes);
1491                 rc = 0;
1492         }
1493         return rc;
1494 }
1495
1496 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1497 {
1498         int rc;
1499         client_obd_list_lock(&cli->cl_loi_list_lock);
1500         rc = cfs_list_empty(&ocw->ocw_entry);
1501         client_obd_list_unlock(&cli->cl_loi_list_lock);
1502         return rc;
1503 }
1504
1505 /**
1506  * The main entry to reserve dirty page accounting. Usually the grant reserved
1507  * in this function will be freed in bulk in osc_free_grant() unless it fails
1508  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1509  *
1510  * The process will be put into sleep if it's already run out of grant.
1511  */
1512 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1513                            struct osc_async_page *oap, int bytes)
1514 {
1515         struct osc_object *osc = oap->oap_obj;
1516         struct lov_oinfo  *loi = osc->oo_oinfo;
1517         struct osc_cache_waiter ocw;
1518         struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
1519                                                   LWI_ON_SIGNAL_NOOP, NULL);
1520         int rc = -EDQUOT;
1521         ENTRY;
1522
1523         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d.\n", bytes);
1524
1525         client_obd_list_lock(&cli->cl_loi_list_lock);
1526
1527         /* force the caller to try sync io.  this can jump the list
1528          * of queued writes and create a discontiguous rpc stream */
1529         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1530             cli->cl_dirty_max < PAGE_CACHE_SIZE     ||
1531             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync)
1532                 GOTO(out, rc = -EDQUOT);
1533
1534         /* Hopefully normal case - cache space and write credits available */
1535         if (osc_enter_cache_try(cli, oap, bytes, 0))
1536                 GOTO(out, rc = 0);
1537
1538         /* We can get here for two reasons: too many dirty pages in cache, or
1539          * run out of grants. In both cases we should write dirty pages out.
1540          * Adding a cache waiter will trigger urgent write-out no matter what
1541          * RPC size will be.
1542          * The exiting condition is no avail grants and no dirty pages caching,
1543          * that really means there is no space on the OST. */
1544         init_waitqueue_head(&ocw.ocw_waitq);
1545         ocw.ocw_oap   = oap;
1546         ocw.ocw_grant = bytes;
1547         while (cli->cl_dirty > 0 || cli->cl_w_in_flight > 0) {
1548                 cfs_list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1549                 ocw.ocw_rc = 0;
1550                 client_obd_list_unlock(&cli->cl_loi_list_lock);
1551
1552                 osc_io_unplug_async(env, cli, NULL);
1553
1554                 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1555                        cli->cl_import->imp_obd->obd_name, &ocw, oap);
1556
1557                 rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1558
1559                 client_obd_list_lock(&cli->cl_loi_list_lock);
1560
1561                 /* l_wait_event is interrupted by signal, or timed out */
1562                 if (rc < 0) {
1563                         switch (rc) {
1564                         case -ETIMEDOUT:
1565                                 OSC_DUMP_GRANT(D_ERROR, cli,
1566                                                 "try to reserve %d.\n", bytes);
1567                                 osc_extent_tree_dump(D_ERROR, osc);
1568                                 rc = -EDQUOT;
1569                                 break;
1570                         case -EINTR:
1571                                 /* Ensures restartability - LU-3581 */
1572                                 rc = -ERESTARTSYS;
1573                                 break;
1574                         default:
1575                                 CDEBUG(D_CACHE, "%s: event for cache space @"
1576                                        " %p never arrived due to %d\n",
1577                                        cli->cl_import->imp_obd->obd_name,
1578                                        &ocw, rc);
1579                                 break;
1580                         }
1581                         cfs_list_del_init(&ocw.ocw_entry);
1582                         GOTO(out, rc);
1583                 }
1584
1585                 LASSERT(cfs_list_empty(&ocw.ocw_entry));
1586                 rc = ocw.ocw_rc;
1587
1588                 if (rc != -EDQUOT)
1589                         GOTO(out, rc);
1590                 if (osc_enter_cache_try(cli, oap, bytes, 0))
1591                         GOTO(out, rc = 0);
1592         }
1593         EXIT;
1594 out:
1595         client_obd_list_unlock(&cli->cl_loi_list_lock);
1596         OSC_DUMP_GRANT(D_CACHE, cli, "returned %d.\n", rc);
1597         RETURN(rc);
1598 }
1599
1600 /* caller must hold loi_list_lock */
1601 void osc_wake_cache_waiters(struct client_obd *cli)
1602 {
1603         cfs_list_t *l, *tmp;
1604         struct osc_cache_waiter *ocw;
1605
1606         ENTRY;
1607         cfs_list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1608                 ocw = cfs_list_entry(l, struct osc_cache_waiter, ocw_entry);
1609                 cfs_list_del_init(&ocw->ocw_entry);
1610
1611                 ocw->ocw_rc = -EDQUOT;
1612                 /* we can't dirty more */
1613                 if ((cli->cl_dirty + PAGE_CACHE_SIZE > cli->cl_dirty_max) ||
1614                     (cfs_atomic_read(&obd_unstable_pages) + 1 +
1615                      cfs_atomic_read(&obd_dirty_pages) > obd_max_dirty_pages)) {
1616                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
1617                                "osc max %ld, sys max %d\n", cli->cl_dirty,
1618                                cli->cl_dirty_max, obd_max_dirty_pages);
1619                         goto wakeup;
1620                 }
1621
1622                 ocw->ocw_rc = 0;
1623                 if (!osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
1624                         ocw->ocw_rc = -EDQUOT;
1625
1626 wakeup:
1627                 CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld, %d\n",
1628                        ocw, ocw->ocw_oap, cli->cl_avail_grant, ocw->ocw_rc);
1629
1630                 wake_up(&ocw->ocw_waitq);
1631         }
1632
1633         EXIT;
1634 }
1635
1636 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1637 {
1638         int hprpc = !!cfs_list_empty(&osc->oo_hp_exts);
1639         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1640 }
1641
1642 /* This maintains the lists of pending pages to read/write for a given object
1643  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1644  * to quickly find objects that are ready to send an RPC. */
1645 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1646                          int cmd)
1647 {
1648         int invalid_import = 0;
1649         ENTRY;
1650
1651         /* if we have an invalid import we want to drain the queued pages
1652          * by forcing them through rpcs that immediately fail and complete
1653          * the pages.  recovery relies on this to empty the queued pages
1654          * before canceling the locks and evicting down the llite pages */
1655         if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1656                 invalid_import = 1;
1657
1658         if (cmd & OBD_BRW_WRITE) {
1659                 if (cfs_atomic_read(&osc->oo_nr_writes) == 0)
1660                         RETURN(0);
1661                 if (invalid_import) {
1662                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1663                         RETURN(1);
1664                 }
1665                 if (!cfs_list_empty(&osc->oo_hp_exts)) {
1666                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1667                         RETURN(1);
1668                 }
1669                 if (!cfs_list_empty(&osc->oo_urgent_exts)) {
1670                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1671                         RETURN(1);
1672                 }
1673                 /* trigger a write rpc stream as long as there are dirtiers
1674                  * waiting for space.  as they're waiting, they're not going to
1675                  * create more pages to coalesce with what's waiting.. */
1676                 if (!cfs_list_empty(&cli->cl_cache_waiters)) {
1677                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1678                         RETURN(1);
1679                 }
1680                 if (cfs_atomic_read(&osc->oo_nr_writes) >=
1681                     cli->cl_max_pages_per_rpc)
1682                         RETURN(1);
1683         } else {
1684                 if (cfs_atomic_read(&osc->oo_nr_reads) == 0)
1685                         RETURN(0);
1686                 if (invalid_import) {
1687                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1688                         RETURN(1);
1689                 }
1690                 /* all read are urgent. */
1691                 if (!cfs_list_empty(&osc->oo_reading_exts))
1692                         RETURN(1);
1693         }
1694
1695         RETURN(0);
1696 }
1697
1698 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1699 {
1700         struct client_obd *cli = osc_cli(obj);
1701         if (cmd & OBD_BRW_WRITE) {
1702                 cfs_atomic_add(delta, &obj->oo_nr_writes);
1703                 cfs_atomic_add(delta, &cli->cl_pending_w_pages);
1704                 LASSERT(cfs_atomic_read(&obj->oo_nr_writes) >= 0);
1705         } else {
1706                 cfs_atomic_add(delta, &obj->oo_nr_reads);
1707                 cfs_atomic_add(delta, &cli->cl_pending_r_pages);
1708                 LASSERT(cfs_atomic_read(&obj->oo_nr_reads) >= 0);
1709         }
1710         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1711 }
1712
1713 static int osc_makes_hprpc(struct osc_object *obj)
1714 {
1715         return !cfs_list_empty(&obj->oo_hp_exts);
1716 }
1717
1718 static void on_list(cfs_list_t *item, cfs_list_t *list, int should_be_on)
1719 {
1720         if (cfs_list_empty(item) && should_be_on)
1721                 cfs_list_add_tail(item, list);
1722         else if (!cfs_list_empty(item) && !should_be_on)
1723                 cfs_list_del_init(item);
1724 }
1725
1726 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1727  * can find pages to build into rpcs quickly */
1728 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1729 {
1730         if (osc_makes_hprpc(osc)) {
1731                 /* HP rpc */
1732                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1733                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1734         } else {
1735                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1736                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1737                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1738                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1739         }
1740
1741         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1742                 cfs_atomic_read(&osc->oo_nr_writes) > 0);
1743
1744         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1745                 cfs_atomic_read(&osc->oo_nr_reads) > 0);
1746
1747         return osc_is_ready(osc);
1748 }
1749
1750 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1751 {
1752         int is_ready;
1753
1754         client_obd_list_lock(&cli->cl_loi_list_lock);
1755         is_ready = __osc_list_maint(cli, osc);
1756         client_obd_list_unlock(&cli->cl_loi_list_lock);
1757
1758         return is_ready;
1759 }
1760
1761 /* this is trying to propogate async writeback errors back up to the
1762  * application.  As an async write fails we record the error code for later if
1763  * the app does an fsync.  As long as errors persist we force future rpcs to be
1764  * sync so that the app can get a sync error and break the cycle of queueing
1765  * pages for which writeback will fail. */
1766 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1767                            int rc)
1768 {
1769         if (rc) {
1770                 if (!ar->ar_rc)
1771                         ar->ar_rc = rc;
1772
1773                 ar->ar_force_sync = 1;
1774                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1775                 return;
1776
1777         }
1778
1779         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1780                 ar->ar_force_sync = 0;
1781 }
1782
1783 /* Performs "unstable" page accounting. This function balances the
1784  * increment operations performed in osc_inc_unstable_pages. It is
1785  * registered as the RPC request callback, and is executed when the
1786  * bulk RPC is committed on the server. Thus at this point, the pages
1787  * involved in the bulk transfer are no longer considered unstable. */
1788 void osc_dec_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                 dec_zone_page_state(desc->bd_iov[i].kiov_page, NR_UNSTABLE_NFS);
1803
1804         cfs_atomic_sub(page_count, &cli->cl_cache->ccc_unstable_nr);
1805         LASSERT(cfs_atomic_read(&cli->cl_cache->ccc_unstable_nr) >= 0);
1806
1807         cfs_atomic_sub(page_count, &cli->cl_unstable_count);
1808         LASSERT(cfs_atomic_read(&cli->cl_unstable_count) >= 0);
1809
1810         cfs_atomic_sub(page_count, &obd_unstable_pages);
1811         LASSERT(cfs_atomic_read(&obd_unstable_pages) >= 0);
1812
1813         spin_lock(&req->rq_lock);
1814         req->rq_committed = 1;
1815         req->rq_unstable  = 0;
1816         spin_unlock(&req->rq_lock);
1817
1818         wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
1819 }
1820
1821 /* "unstable" page accounting. See: osc_dec_unstable_pages. */
1822 void osc_inc_unstable_pages(struct ptlrpc_request *req)
1823 {
1824         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
1825         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
1826         obd_count                page_count = desc->bd_iov_count;
1827         int i;
1828
1829         /* No unstable page tracking */
1830         if (cli->cl_cache == NULL)
1831                 return;
1832
1833         LASSERT(page_count >= 0);
1834
1835         for (i = 0; i < page_count; i++)
1836                 inc_zone_page_state(desc->bd_iov[i].kiov_page, NR_UNSTABLE_NFS);
1837
1838         LASSERT(cfs_atomic_read(&cli->cl_cache->ccc_unstable_nr) >= 0);
1839         cfs_atomic_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1840
1841         LASSERT(cfs_atomic_read(&cli->cl_unstable_count) >= 0);
1842         cfs_atomic_add(page_count, &cli->cl_unstable_count);
1843
1844         LASSERT(cfs_atomic_read(&obd_unstable_pages) >= 0);
1845         cfs_atomic_add(page_count, &obd_unstable_pages);
1846
1847         spin_lock(&req->rq_lock);
1848
1849         /* If the request has already been committed (i.e. brw_commit
1850          * called via rq_commit_cb), we need to undo the unstable page
1851          * increments we just performed because rq_commit_cb wont be
1852          * called again. Otherwise, just set the commit callback so the
1853          * unstable page accounting is properly updated when the request
1854          * is committed */
1855         if (req->rq_committed) {
1856                 /* Drop lock before calling osc_dec_unstable_pages */
1857                 spin_unlock(&req->rq_lock);
1858                 osc_dec_unstable_pages(req);
1859                 spin_lock(&req->rq_lock);
1860         } else {
1861                 req->rq_unstable  = 1;
1862                 req->rq_commit_cb = osc_dec_unstable_pages;
1863         }
1864
1865         spin_unlock(&req->rq_lock);
1866 }
1867
1868 /* this must be called holding the loi list lock to give coverage to exit_cache,
1869  * async_flag maintenance, and oap_request */
1870 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1871                               struct osc_async_page *oap, int sent, int rc)
1872 {
1873         struct osc_object *osc = oap->oap_obj;
1874         struct lov_oinfo  *loi = osc->oo_oinfo;
1875         __u64 xid = 0;
1876
1877         ENTRY;
1878         if (oap->oap_request != NULL) {
1879                 if (rc == 0)
1880                         osc_inc_unstable_pages(oap->oap_request);
1881
1882                 xid = ptlrpc_req_xid(oap->oap_request);
1883                 ptlrpc_req_finished(oap->oap_request);
1884                 oap->oap_request = NULL;
1885         }
1886
1887         /* As the transfer for this page is being done, clear the flags */
1888         spin_lock(&oap->oap_lock);
1889         oap->oap_async_flags = 0;
1890         spin_unlock(&oap->oap_lock);
1891         oap->oap_interrupted = 0;
1892
1893         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1894                 client_obd_list_lock(&cli->cl_loi_list_lock);
1895                 osc_process_ar(&cli->cl_ar, xid, rc);
1896                 osc_process_ar(&loi->loi_ar, xid, rc);
1897                 client_obd_list_unlock(&cli->cl_loi_list_lock);
1898         }
1899
1900         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1901         if (rc)
1902                 CERROR("completion on oap %p obj %p returns %d.\n",
1903                        oap, osc, rc);
1904
1905         EXIT;
1906 }
1907
1908 /**
1909  * Try to add extent to one RPC. We need to think about the following things:
1910  * - # of pages must not be over max_pages_per_rpc
1911  * - extent must be compatible with previous ones
1912  */
1913 static int try_to_add_extent_for_io(struct client_obd *cli,
1914                                     struct osc_extent *ext, cfs_list_t *rpclist,
1915                                     int *pc, unsigned int *max_pages)
1916 {
1917         struct osc_extent *tmp;
1918         ENTRY;
1919
1920         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1921                 ext);
1922
1923         *max_pages = max(ext->oe_mppr, *max_pages);
1924         if (*pc + ext->oe_nr_pages > *max_pages)
1925                 RETURN(0);
1926
1927         cfs_list_for_each_entry(tmp, rpclist, oe_link) {
1928                 EASSERT(tmp->oe_owner == current, tmp);
1929 #if 0
1930                 if (overlapped(tmp, ext)) {
1931                         OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
1932                         EASSERT(0, ext);
1933                 }
1934 #endif
1935
1936                 if (tmp->oe_srvlock != ext->oe_srvlock ||
1937                     !tmp->oe_grants != !ext->oe_grants)
1938                         RETURN(0);
1939
1940                 /* remove break for strict check */
1941                 break;
1942         }
1943
1944         *pc += ext->oe_nr_pages;
1945         cfs_list_move_tail(&ext->oe_link, rpclist);
1946         ext->oe_owner = current;
1947         RETURN(1);
1948 }
1949
1950 /**
1951  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1952  * get_write_extent() takes all appropriate extents in atomic.
1953  *
1954  * The following policy is used to collect extents for IO:
1955  * 1. Add as many HP extents as possible;
1956  * 2. Add the first urgent extent in urgent extent list and take it out of
1957  *    urgent list;
1958  * 3. Add subsequent extents of this urgent extent;
1959  * 4. If urgent list is not empty, goto 2;
1960  * 5. Traverse the extent tree from the 1st extent;
1961  * 6. Above steps exit if there is no space in this RPC.
1962  */
1963 static int get_write_extents(struct osc_object *obj, cfs_list_t *rpclist)
1964 {
1965         struct client_obd *cli = osc_cli(obj);
1966         struct osc_extent *ext;
1967         int page_count = 0;
1968         unsigned int max_pages = cli->cl_max_pages_per_rpc;
1969
1970         LASSERT(osc_object_is_locked(obj));
1971         while (!cfs_list_empty(&obj->oo_hp_exts)) {
1972                 ext = cfs_list_entry(obj->oo_hp_exts.next, struct osc_extent,
1973                                      oe_link);
1974                 LASSERT(ext->oe_state == OES_CACHE);
1975                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1976                                               &max_pages))
1977                         return page_count;
1978                 EASSERT(ext->oe_nr_pages <= max_pages, ext);
1979         }
1980         if (page_count == max_pages)
1981                 return page_count;
1982
1983         while (!cfs_list_empty(&obj->oo_urgent_exts)) {
1984                 ext = cfs_list_entry(obj->oo_urgent_exts.next,
1985                                      struct osc_extent, oe_link);
1986                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1987                                               &max_pages))
1988                         return page_count;
1989
1990                 if (!ext->oe_intree)
1991                         continue;
1992
1993                 while ((ext = next_extent(ext)) != NULL) {
1994                         if ((ext->oe_state != OES_CACHE) ||
1995                             (!cfs_list_empty(&ext->oe_link) &&
1996                              ext->oe_owner != NULL))
1997                                 continue;
1998
1999                         if (!try_to_add_extent_for_io(cli, ext, rpclist,
2000                                                       &page_count, &max_pages))
2001                                 return page_count;
2002                 }
2003         }
2004         if (page_count == max_pages)
2005                 return page_count;
2006
2007         ext = first_extent(obj);
2008         while (ext != NULL) {
2009                 if ((ext->oe_state != OES_CACHE) ||
2010                     /* this extent may be already in current rpclist */
2011                     (!cfs_list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
2012                         ext = next_extent(ext);
2013                         continue;
2014                 }
2015
2016                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
2017                                               &max_pages))
2018                         return page_count;
2019
2020                 ext = next_extent(ext);
2021         }
2022         return page_count;
2023 }
2024
2025 static int
2026 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
2027                    struct osc_object *osc, pdl_policy_t pol)
2028 {
2029         CFS_LIST_HEAD(rpclist);
2030         struct osc_extent *ext;
2031         struct osc_extent *tmp;
2032         struct osc_extent *first = NULL;
2033         obd_count page_count = 0;
2034         int srvlock = 0;
2035         int rc = 0;
2036         ENTRY;
2037
2038         LASSERT(osc_object_is_locked(osc));
2039
2040         page_count = get_write_extents(osc, &rpclist);
2041         LASSERT(equi(page_count == 0, cfs_list_empty(&rpclist)));
2042
2043         if (cfs_list_empty(&rpclist))
2044                 RETURN(0);
2045
2046         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2047
2048         cfs_list_for_each_entry(ext, &rpclist, oe_link) {
2049                 LASSERT(ext->oe_state == OES_CACHE ||
2050                         ext->oe_state == OES_LOCK_DONE);
2051                 if (ext->oe_state == OES_CACHE)
2052                         osc_extent_state_set(ext, OES_LOCKING);
2053                 else
2054                         osc_extent_state_set(ext, OES_RPC);
2055         }
2056
2057         /* we're going to grab page lock, so release object lock because
2058          * lock order is page lock -> object lock. */
2059         osc_object_unlock(osc);
2060
2061         cfs_list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2062                 if (ext->oe_state == OES_LOCKING) {
2063                         rc = osc_extent_make_ready(env, ext);
2064                         if (unlikely(rc < 0)) {
2065                                 cfs_list_del_init(&ext->oe_link);
2066                                 osc_extent_finish(env, ext, 0, rc);
2067                                 continue;
2068                         }
2069                 }
2070                 if (first == NULL) {
2071                         first = ext;
2072                         srvlock = ext->oe_srvlock;
2073                 } else {
2074                         LASSERT(srvlock == ext->oe_srvlock);
2075                 }
2076         }
2077
2078         if (!cfs_list_empty(&rpclist)) {
2079                 LASSERT(page_count > 0);
2080                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE, pol);
2081                 LASSERT(cfs_list_empty(&rpclist));
2082         }
2083
2084         osc_object_lock(osc);
2085         RETURN(rc);
2086 }
2087
2088 /**
2089  * prepare pages for ASYNC io and put pages in send queue.
2090  *
2091  * \param cmd OBD_BRW_* macroses
2092  * \param lop pending pages
2093  *
2094  * \return zero if no page added to send queue.
2095  * \return 1 if pages successfully added to send queue.
2096  * \return negative on errors.
2097  */
2098 static int
2099 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2100                   struct osc_object *osc, pdl_policy_t pol)
2101 {
2102         struct osc_extent *ext;
2103         struct osc_extent *next;
2104         CFS_LIST_HEAD(rpclist);
2105         int page_count = 0;
2106         unsigned int max_pages = cli->cl_max_pages_per_rpc;
2107         int rc = 0;
2108         ENTRY;
2109
2110         LASSERT(osc_object_is_locked(osc));
2111         cfs_list_for_each_entry_safe(ext, next,
2112                                      &osc->oo_reading_exts, oe_link) {
2113                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2114                 if (!try_to_add_extent_for_io(cli, ext, &rpclist, &page_count,
2115                                               &max_pages))
2116                         break;
2117                 osc_extent_state_set(ext, OES_RPC);
2118                 EASSERT(ext->oe_nr_pages <= max_pages, ext);
2119         }
2120         LASSERT(page_count <= max_pages);
2121
2122         osc_update_pending(osc, OBD_BRW_READ, -page_count);
2123
2124         if (!cfs_list_empty(&rpclist)) {
2125                 osc_object_unlock(osc);
2126
2127                 LASSERT(page_count > 0);
2128                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ, pol);
2129                 LASSERT(cfs_list_empty(&rpclist));
2130
2131                 osc_object_lock(osc);
2132         }
2133         RETURN(rc);
2134 }
2135
2136 #define list_to_obj(list, item) ({                                            \
2137         cfs_list_t *__tmp = (list)->next;                                     \
2138         cfs_list_del_init(__tmp);                                             \
2139         cfs_list_entry(__tmp, struct osc_object, oo_##item);                  \
2140 })
2141
2142 /* This is called by osc_check_rpcs() to find which objects have pages that
2143  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2144 static struct osc_object *osc_next_obj(struct client_obd *cli)
2145 {
2146         ENTRY;
2147
2148         /* First return objects that have blocked locks so that they
2149          * will be flushed quickly and other clients can get the lock,
2150          * then objects which have pages ready to be stuffed into RPCs */
2151         if (!cfs_list_empty(&cli->cl_loi_hp_ready_list))
2152                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2153         if (!cfs_list_empty(&cli->cl_loi_ready_list))
2154                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2155
2156         /* then if we have cache waiters, return all objects with queued
2157          * writes.  This is especially important when many small files
2158          * have filled up the cache and not been fired into rpcs because
2159          * they don't pass the nr_pending/object threshhold */
2160         if (!cfs_list_empty(&cli->cl_cache_waiters) &&
2161             !cfs_list_empty(&cli->cl_loi_write_list))
2162                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2163
2164         /* then return all queued objects when we have an invalid import
2165          * so that they get flushed */
2166         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2167                 if (!cfs_list_empty(&cli->cl_loi_write_list))
2168                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2169                                            write_item));
2170                 if (!cfs_list_empty(&cli->cl_loi_read_list))
2171                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2172                                            read_item));
2173         }
2174         RETURN(NULL);
2175 }
2176
2177 /* called with the loi list lock held */
2178 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli,
2179                            pdl_policy_t pol)
2180 {
2181         struct osc_object *osc;
2182         int rc = 0;
2183         ENTRY;
2184
2185         while ((osc = osc_next_obj(cli)) != NULL) {
2186                 struct cl_object *obj = osc2cl(osc);
2187                 struct lu_ref_link link;
2188
2189                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2190
2191                 if (osc_max_rpc_in_flight(cli, osc)) {
2192                         __osc_list_maint(cli, osc);
2193                         break;
2194                 }
2195
2196                 cl_object_get(obj);
2197                 client_obd_list_unlock(&cli->cl_loi_list_lock);
2198                 lu_object_ref_add_at(&obj->co_lu, &link, "check",
2199                                      current);
2200
2201                 /* attempt some read/write balancing by alternating between
2202                  * reads and writes in an object.  The makes_rpc checks here
2203                  * would be redundant if we were getting read/write work items
2204                  * instead of objects.  we don't want send_oap_rpc to drain a
2205                  * partial read pending queue when we're given this object to
2206                  * do io on writes while there are cache waiters */
2207                 osc_object_lock(osc);
2208                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2209                         rc = osc_send_write_rpc(env, cli, osc, pol);
2210                         if (rc < 0) {
2211                                 CERROR("Write request failed with %d\n", rc);
2212
2213                                 /* osc_send_write_rpc failed, mostly because of
2214                                  * memory pressure.
2215                                  *
2216                                  * It can't break here, because if:
2217                                  *  - a page was submitted by osc_io_submit, so
2218                                  *    page locked;
2219                                  *  - no request in flight
2220                                  *  - no subsequent request
2221                                  * The system will be in live-lock state,
2222                                  * because there is no chance to call
2223                                  * osc_io_unplug() and osc_check_rpcs() any
2224                                  * more. pdflush can't help in this case,
2225                                  * because it might be blocked at grabbing
2226                                  * the page lock as we mentioned.
2227                                  *
2228                                  * Anyway, continue to drain pages. */
2229                                 /* break; */
2230                         }
2231                 }
2232                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2233                         rc = osc_send_read_rpc(env, cli, osc, pol);
2234                         if (rc < 0)
2235                                 CERROR("Read request failed with %d\n", rc);
2236                 }
2237                 osc_object_unlock(osc);
2238
2239                 osc_list_maint(cli, osc);
2240                 lu_object_ref_del_at(&obj->co_lu, &link, "check",
2241                                      current);
2242                 cl_object_put(env, obj);
2243
2244                 client_obd_list_lock(&cli->cl_loi_list_lock);
2245         }
2246 }
2247
2248 static int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2249                           struct osc_object *osc, pdl_policy_t pol, int async)
2250 {
2251         int rc = 0;
2252
2253         if (osc != NULL && osc_list_maint(cli, osc) == 0)
2254                 return 0;
2255
2256         if (!async) {
2257                 /* disable osc_lru_shrink() temporarily to avoid
2258                  * potential stack overrun problem. LU-2859 */
2259                 cfs_atomic_inc(&cli->cl_lru_shrinkers);
2260                 client_obd_list_lock(&cli->cl_loi_list_lock);
2261                 osc_check_rpcs(env, cli, pol);
2262                 client_obd_list_unlock(&cli->cl_loi_list_lock);
2263                 cfs_atomic_dec(&cli->cl_lru_shrinkers);
2264         } else {
2265                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2266                 LASSERT(cli->cl_writeback_work != NULL);
2267                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2268         }
2269         return rc;
2270 }
2271
2272 static int osc_io_unplug_async(const struct lu_env *env,
2273                                 struct client_obd *cli, struct osc_object *osc)
2274 {
2275         /* XXX: policy is no use actually. */
2276         return osc_io_unplug0(env, cli, osc, PDL_POLICY_ROUND, 1);
2277 }
2278
2279 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
2280                    struct osc_object *osc, pdl_policy_t pol)
2281 {
2282         (void)osc_io_unplug0(env, cli, osc, pol, 0);
2283 }
2284
2285 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2286                         struct page *page, loff_t offset)
2287 {
2288         struct obd_export     *exp = osc_export(osc);
2289         struct osc_async_page *oap = &ops->ops_oap;
2290         ENTRY;
2291
2292         if (!page)
2293                 return cfs_size_round(sizeof(*oap));
2294
2295         oap->oap_magic = OAP_MAGIC;
2296         oap->oap_cli = &exp->exp_obd->u.cli;
2297         oap->oap_obj = osc;
2298
2299         oap->oap_page = page;
2300         oap->oap_obj_off = offset;
2301         LASSERT(!(offset & ~CFS_PAGE_MASK));
2302
2303         if (!client_is_remote(exp) && cfs_capable(CFS_CAP_SYS_RESOURCE))
2304                 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2305
2306         CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2307         CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2308
2309         spin_lock_init(&oap->oap_lock);
2310         CDEBUG(D_INFO, "oap %p page %p obj off "LPU64"\n",
2311                oap, page, oap->oap_obj_off);
2312         RETURN(0);
2313 }
2314
2315 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2316                        struct osc_page *ops)
2317 {
2318         struct osc_io *oio = osc_env_io(env);
2319         struct osc_extent     *ext = NULL;
2320         struct osc_async_page *oap = &ops->ops_oap;
2321         struct client_obd     *cli = oap->oap_cli;
2322         struct osc_object     *osc = oap->oap_obj;
2323         pgoff_t index;
2324         int    grants = 0;
2325         int    brw_flags = OBD_BRW_ASYNC;
2326         int    cmd = OBD_BRW_WRITE;
2327         int    need_release = 0;
2328         int    rc = 0;
2329         ENTRY;
2330
2331         if (oap->oap_magic != OAP_MAGIC)
2332                 RETURN(-EINVAL);
2333
2334         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2335                 RETURN(-EIO);
2336
2337         if (!cfs_list_empty(&oap->oap_pending_item) ||
2338             !cfs_list_empty(&oap->oap_rpc_item))
2339                 RETURN(-EBUSY);
2340
2341         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2342         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2343         if (!client_is_remote(osc_export(osc)) &&
2344             cfs_capable(CFS_CAP_SYS_RESOURCE)) {
2345                 brw_flags |= OBD_BRW_NOQUOTA;
2346                 cmd |= OBD_BRW_NOQUOTA;
2347         }
2348
2349         /* check if the file's owner/group is over quota */
2350         if (!(cmd & OBD_BRW_NOQUOTA)) {
2351                 struct cl_object *obj;
2352                 struct cl_attr   *attr;
2353                 unsigned int qid[MAXQUOTAS];
2354
2355                 obj = cl_object_top(&osc->oo_cl);
2356                 attr = &osc_env_info(env)->oti_attr;
2357
2358                 cl_object_attr_lock(obj);
2359                 rc = cl_object_attr_get(env, obj, attr);
2360                 cl_object_attr_unlock(obj);
2361
2362                 qid[USRQUOTA] = attr->cat_uid;
2363                 qid[GRPQUOTA] = attr->cat_gid;
2364                 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2365                         rc = -EDQUOT;
2366                 if (rc)
2367                         RETURN(rc);
2368         }
2369
2370         if (osc_over_unstable_soft_limit(cli))
2371                 brw_flags |= OBD_BRW_SOFT_SYNC;
2372
2373         oap->oap_cmd = cmd;
2374         oap->oap_page_off = ops->ops_from;
2375         oap->oap_count = ops->ops_to - ops->ops_from;
2376         /* No need to hold a lock here,
2377          * since this page is not in any list yet. */
2378         oap->oap_async_flags = 0;
2379         oap->oap_brw_flags = brw_flags;
2380
2381         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2382                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2383
2384         index = oap2cl_page(oap)->cp_index;
2385
2386         /* Add this page into extent by the following steps:
2387          * 1. if there exists an active extent for this IO, mostly this page
2388          *    can be added to the active extent and sometimes we need to
2389          *    expand extent to accomodate this page;
2390          * 2. otherwise, a new extent will be allocated. */
2391
2392         ext = oio->oi_active;
2393         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2394                 /* one chunk plus extent overhead must be enough to write this
2395                  * page */
2396                 grants = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2397                 if (ext->oe_end >= index)
2398                         grants = 0;
2399
2400                 /* it doesn't need any grant to dirty this page */
2401                 client_obd_list_lock(&cli->cl_loi_list_lock);
2402                 rc = osc_enter_cache_try(cli, oap, grants, 0);
2403                 client_obd_list_unlock(&cli->cl_loi_list_lock);
2404                 if (rc == 0) { /* try failed */
2405                         grants = 0;
2406                         need_release = 1;
2407                 } else if (ext->oe_end < index) {
2408                         int tmp = grants;
2409                         /* try to expand this extent */
2410                         rc = osc_extent_expand(ext, index, &tmp);
2411                         if (rc < 0) {
2412                                 need_release = 1;
2413                                 /* don't free reserved grant */
2414                         } else {
2415                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2416                                                 "expanded for %lu.\n", index);
2417                                 osc_unreserve_grant(cli, grants, tmp);
2418                                 grants = 0;
2419                         }
2420                 }
2421                 rc = 0;
2422         } else if (ext != NULL) {
2423                 /* index is located outside of active extent */
2424                 need_release = 1;
2425         }
2426         if (need_release) {
2427                 osc_extent_release(env, ext);
2428                 oio->oi_active = NULL;
2429                 ext = NULL;
2430         }
2431
2432         if (ext == NULL) {
2433                 int tmp = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2434
2435                 /* try to find new extent to cover this page */
2436                 LASSERT(oio->oi_active == NULL);
2437                 /* we may have allocated grant for this page if we failed
2438                  * to expand the previous active extent. */
2439                 LASSERT(ergo(grants > 0, grants >= tmp));
2440
2441                 rc = 0;
2442                 if (grants == 0) {
2443                         /* we haven't allocated grant for this page. */
2444                         rc = osc_enter_cache(env, cli, oap, tmp);
2445                         if (rc == 0)
2446                                 grants = tmp;
2447                 }
2448
2449                 tmp = grants;
2450                 if (rc == 0) {
2451                         ext = osc_extent_find(env, osc, index, &tmp);
2452                         if (IS_ERR(ext)) {
2453                                 LASSERT(tmp == grants);
2454                                 osc_exit_cache(cli, oap);
2455                                 rc = PTR_ERR(ext);
2456                                 ext = NULL;
2457                         } else {
2458                                 oio->oi_active = ext;
2459                         }
2460                 }
2461                 if (grants > 0)
2462                         osc_unreserve_grant(cli, grants, tmp);
2463         }
2464
2465         LASSERT(ergo(rc == 0, ext != NULL));
2466         if (ext != NULL) {
2467                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2468                          ext, "index = %lu.\n", index);
2469                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2470
2471                 osc_object_lock(osc);
2472                 if (ext->oe_nr_pages == 0)
2473                         ext->oe_srvlock = ops->ops_srvlock;
2474                 else
2475                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2476                 ++ext->oe_nr_pages;
2477                 cfs_list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2478                 osc_object_unlock(osc);
2479         }
2480         RETURN(rc);
2481 }
2482
2483 int osc_teardown_async_page(const struct lu_env *env,
2484                             struct osc_object *obj, struct osc_page *ops)
2485 {
2486         struct osc_async_page *oap = &ops->ops_oap;
2487         struct osc_extent     *ext = NULL;
2488         int rc = 0;
2489         ENTRY;
2490
2491         LASSERT(oap->oap_magic == OAP_MAGIC);
2492
2493         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2494                oap, ops, oap2cl_page(oap)->cp_index);
2495
2496         osc_object_lock(obj);
2497         if (!cfs_list_empty(&oap->oap_rpc_item)) {
2498                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2499                 rc = -EBUSY;
2500         } else if (!cfs_list_empty(&oap->oap_pending_item)) {
2501                 ext = osc_extent_lookup(obj, oap2cl_page(oap)->cp_index);
2502                 /* only truncated pages are allowed to be taken out.
2503                  * See osc_extent_truncate() and osc_cache_truncate_start()
2504                  * for details. */
2505                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2506                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2507                                         oap2cl_page(oap)->cp_index);
2508                         rc = -EBUSY;
2509                 }
2510         }
2511         osc_object_unlock(obj);
2512         if (ext != NULL)
2513                 osc_extent_put(env, ext);
2514         RETURN(rc);
2515 }
2516
2517 /**
2518  * This is called when a page is picked up by kernel to write out.
2519  *
2520  * We should find out the corresponding extent and add the whole extent
2521  * into urgent list. The extent may be being truncated or used, handle it
2522  * carefully.
2523  */
2524 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2525                          struct osc_page *ops)
2526 {
2527         struct osc_extent *ext   = NULL;
2528         struct osc_object *obj   = cl2osc(ops->ops_cl.cpl_obj);
2529         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2530         pgoff_t            index = cp->cp_index;
2531         struct osc_async_page *oap = &ops->ops_oap;
2532         bool unplug = false;
2533         int rc = 0;
2534         ENTRY;
2535
2536         osc_object_lock(obj);
2537         ext = osc_extent_lookup(obj, index);
2538         if (ext == NULL) {
2539                 osc_extent_tree_dump(D_ERROR, obj);
2540                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2541         }
2542
2543         switch (ext->oe_state) {
2544         case OES_RPC:
2545         case OES_LOCK_DONE:
2546                 CL_PAGE_DEBUG(D_ERROR, env, cl_page_top(cp),
2547                               "flush an in-rpc page?\n");
2548                 LASSERT(0);
2549                 break;
2550         case OES_LOCKING:
2551                 /* If we know this extent is being written out, we should abort
2552                  * so that the writer can make this page ready. Otherwise, there
2553                  * exists a deadlock problem because other process can wait for
2554                  * page writeback bit holding page lock; and meanwhile in
2555                  * vvp_page_make_ready(), we need to grab page lock before
2556                  * really sending the RPC. */
2557         case OES_TRUNC:
2558                 /* race with truncate, page will be redirtied */
2559                 GOTO(out, rc = -EAGAIN);
2560         default:
2561                 break;
2562         }
2563
2564         rc = cl_page_prep(env, io, cl_page_top(cp), CRT_WRITE);
2565         if (rc)
2566                 GOTO(out, rc);
2567
2568         spin_lock(&oap->oap_lock);
2569         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2570         spin_unlock(&oap->oap_lock);
2571
2572         if (memory_pressure_get())
2573                 ext->oe_memalloc = 1;
2574
2575         ext->oe_urgent = 1;
2576         if (ext->oe_state == OES_CACHE) {
2577                 OSC_EXTENT_DUMP(D_CACHE, ext,
2578                                 "flush page %p make it urgent.\n", oap);
2579                 if (cfs_list_empty(&ext->oe_link))
2580                         cfs_list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2581                 unplug = true;
2582         }
2583         rc = 0;
2584         EXIT;
2585
2586 out:
2587         osc_object_unlock(obj);
2588         osc_extent_put(env, ext);
2589         if (unplug)
2590                 osc_io_unplug_async(env, osc_cli(obj), obj);
2591         return rc;
2592 }
2593
2594 /**
2595  * this is called when a sync waiter receives an interruption.  Its job is to
2596  * get the caller woken as soon as possible.  If its page hasn't been put in an
2597  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2598  * desiring interruption which will forcefully complete the rpc once the rpc
2599  * has timed out.
2600  */
2601 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2602 {
2603         struct osc_async_page *oap = &ops->ops_oap;
2604         struct osc_object     *obj = oap->oap_obj;
2605         struct client_obd     *cli = osc_cli(obj);
2606         struct osc_extent     *ext;
2607         struct osc_extent     *found = NULL;
2608         cfs_list_t            *plist;
2609         pgoff_t index = oap2cl_page(oap)->cp_index;
2610         int     rc = -EBUSY;
2611         int     cmd;
2612         ENTRY;
2613
2614         LASSERT(!oap->oap_interrupted);
2615         oap->oap_interrupted = 1;
2616
2617         /* Find out the caching extent */
2618         osc_object_lock(obj);
2619         if (oap->oap_cmd & OBD_BRW_WRITE) {
2620                 plist = &obj->oo_urgent_exts;
2621                 cmd   = OBD_BRW_WRITE;
2622         } else {
2623                 plist = &obj->oo_reading_exts;
2624                 cmd   = OBD_BRW_READ;
2625         }
2626         cfs_list_for_each_entry(ext, plist, oe_link) {
2627                 if (ext->oe_start <= index && ext->oe_end >= index) {
2628                         LASSERT(ext->oe_state == OES_LOCK_DONE);
2629                         /* For OES_LOCK_DONE state extent, it has already held
2630                          * a refcount for RPC. */
2631                         found = osc_extent_get(ext);
2632                         break;
2633                 }
2634         }
2635         if (found != NULL) {
2636                 cfs_list_del_init(&found->oe_link);
2637                 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2638                 osc_object_unlock(obj);
2639
2640                 osc_extent_finish(env, found, 0, -EINTR);
2641                 osc_extent_put(env, found);
2642                 rc = 0;
2643         } else {
2644                 osc_object_unlock(obj);
2645                 /* ok, it's been put in an rpc. only one oap gets a request
2646                  * reference */
2647                 if (oap->oap_request != NULL) {
2648                         ptlrpc_mark_interrupted(oap->oap_request);
2649                         ptlrpcd_wake(oap->oap_request);
2650                         ptlrpc_req_finished(oap->oap_request);
2651                         oap->oap_request = NULL;
2652                 }
2653         }
2654
2655         osc_list_maint(cli, obj);
2656         RETURN(rc);
2657 }
2658
2659 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
2660                          cfs_list_t *list, int cmd, int brw_flags)
2661 {
2662         struct client_obd     *cli = osc_cli(obj);
2663         struct osc_extent     *ext;
2664         struct osc_async_page *oap;
2665         int     page_count = 0;
2666         int     mppr       = cli->cl_max_pages_per_rpc;
2667         pgoff_t start      = CL_PAGE_EOF;
2668         pgoff_t end        = 0;
2669         ENTRY;
2670
2671         cfs_list_for_each_entry(oap, list, oap_pending_item) {
2672                 struct cl_page *cp = oap2cl_page(oap);
2673                 if (cp->cp_index > end)
2674                         end = cp->cp_index;
2675                 if (cp->cp_index < start)
2676                         start = cp->cp_index;
2677                 ++page_count;
2678                 mppr <<= (page_count > mppr);
2679         }
2680
2681         ext = osc_extent_alloc(obj);
2682         if (ext == NULL) {
2683                 cfs_list_for_each_entry(oap, list, oap_pending_item) {
2684                         cfs_list_del_init(&oap->oap_pending_item);
2685                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2686                 }
2687                 RETURN(-ENOMEM);
2688         }
2689
2690         ext->oe_rw = !!(cmd & OBD_BRW_READ);
2691         ext->oe_urgent = 1;
2692         ext->oe_start = start;
2693         ext->oe_end = ext->oe_max_end = end;
2694         ext->oe_obj = obj;
2695         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2696         ext->oe_nr_pages = page_count;
2697         ext->oe_mppr = mppr;
2698         cfs_list_splice_init(list, &ext->oe_pages);
2699
2700         osc_object_lock(obj);
2701         /* Reuse the initial refcount for RPC, don't drop it */
2702         osc_extent_state_set(ext, OES_LOCK_DONE);
2703         if (cmd & OBD_BRW_WRITE) {
2704                 cfs_list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2705                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2706         } else {
2707                 cfs_list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2708                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2709         }
2710         osc_object_unlock(obj);
2711
2712         osc_io_unplug(env, cli, obj, PDL_POLICY_ROUND);
2713         RETURN(0);
2714 }
2715
2716 /**
2717  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2718  */
2719 int osc_cache_truncate_start(const struct lu_env *env, struct osc_io *oio,
2720                              struct osc_object *obj, __u64 size)
2721 {
2722         struct client_obd *cli = osc_cli(obj);
2723         struct osc_extent *ext;
2724         struct osc_extent *waiting = NULL;
2725         pgoff_t index;
2726         CFS_LIST_HEAD(list);
2727         int result = 0;
2728         bool partial;
2729         ENTRY;
2730
2731         /* pages with index greater or equal to index will be truncated. */
2732         index = cl_index(osc2cl(obj), size);
2733         partial = size > cl_offset(osc2cl(obj), index);
2734
2735 again:
2736         osc_object_lock(obj);
2737         ext = osc_extent_search(obj, index);
2738         if (ext == NULL)
2739                 ext = first_extent(obj);
2740         else if (ext->oe_end < index)
2741                 ext = next_extent(ext);
2742         while (ext != NULL) {
2743                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2744
2745                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2746                         /* if ext is in urgent state, it means there must exist
2747                          * a page already having been flushed by write_page().
2748                          * We have to wait for this extent because we can't
2749                          * truncate that page. */
2750                         LASSERT(!ext->oe_hp);
2751                         OSC_EXTENT_DUMP(D_CACHE, ext,
2752                                         "waiting for busy extent\n");
2753                         waiting = osc_extent_get(ext);
2754                         break;
2755                 }
2756
2757                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:"LPU64".\n", size);
2758
2759                 osc_extent_get(ext);
2760                 if (ext->oe_state == OES_ACTIVE) {
2761                         /* though we grab inode mutex for write path, but we
2762                          * release it before releasing extent(in osc_io_end()),
2763                          * so there is a race window that an extent is still
2764                          * in OES_ACTIVE when truncate starts. */
2765                         LASSERT(!ext->oe_trunc_pending);
2766                         ext->oe_trunc_pending = 1;
2767                 } else {
2768                         EASSERT(ext->oe_state == OES_CACHE, ext);
2769                         osc_extent_state_set(ext, OES_TRUNC);
2770                         osc_update_pending(obj, OBD_BRW_WRITE,
2771                                            -ext->oe_nr_pages);
2772                 }
2773                 EASSERT(cfs_list_empty(&ext->oe_link), ext);
2774                 cfs_list_add_tail(&ext->oe_link, &list);
2775
2776                 ext = next_extent(ext);
2777         }
2778         osc_object_unlock(obj);
2779
2780         osc_list_maint(cli, obj);
2781
2782         while (!cfs_list_empty(&list)) {
2783                 int rc;
2784
2785                 ext = cfs_list_entry(list.next, struct osc_extent, oe_link);
2786                 cfs_list_del_init(&ext->oe_link);
2787
2788                 /* extent may be in OES_ACTIVE state because inode mutex
2789                  * is released before osc_io_end() in file write case */
2790                 if (ext->oe_state != OES_TRUNC)
2791                         osc_extent_wait(env, ext, OES_TRUNC);
2792
2793                 rc = osc_extent_truncate(ext, index, partial);
2794                 if (rc < 0) {
2795                         if (result == 0)
2796                                 result = rc;
2797
2798                         OSC_EXTENT_DUMP(D_ERROR, ext,
2799                                         "truncate error %d\n", rc);
2800                 } else if (ext->oe_nr_pages == 0) {
2801                         osc_extent_remove(ext);
2802                 } else {
2803                         /* this must be an overlapped extent which means only
2804                          * part of pages in this extent have been truncated.
2805                          */
2806                         EASSERTF(ext->oe_start <= index, ext,
2807                                  "trunc index = %lu/%d.\n", index, partial);
2808                         /* fix index to skip this partially truncated extent */
2809                         index = ext->oe_end + 1;
2810                         partial = false;
2811
2812                         /* we need to hold this extent in OES_TRUNC state so
2813                          * that no writeback will happen. This is to avoid
2814                          * BUG 17397. */
2815                         LASSERT(oio->oi_trunc == NULL);
2816                         oio->oi_trunc = osc_extent_get(ext);
2817                         OSC_EXTENT_DUMP(D_CACHE, ext,
2818                                         "trunc at "LPU64"\n", size);
2819                 }
2820                 osc_extent_put(env, ext);
2821         }
2822         if (waiting != NULL) {
2823                 int rc;
2824
2825                 /* ignore the result of osc_extent_wait the write initiator
2826                  * should take care of it. */
2827                 rc = osc_extent_wait(env, waiting, OES_INV);
2828                 if (rc < 0)
2829                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2830
2831                 osc_extent_put(env, waiting);
2832                 waiting = NULL;
2833                 goto again;
2834         }
2835         RETURN(result);
2836 }
2837
2838 /**
2839  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2840  */
2841 void osc_cache_truncate_end(const struct lu_env *env, struct osc_io *oio,
2842                             struct osc_object *obj)
2843 {
2844         struct osc_extent *ext = oio->oi_trunc;
2845
2846         oio->oi_trunc = NULL;
2847         if (ext != NULL) {
2848                 bool unplug = false;
2849
2850                 EASSERT(ext->oe_nr_pages > 0, ext);
2851                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2852                 EASSERT(!ext->oe_urgent, ext);
2853
2854                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2855                 osc_object_lock(obj);
2856                 osc_extent_state_set(ext, OES_CACHE);
2857                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2858                         ext->oe_urgent = 1;
2859                         cfs_list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2860                         unplug = true;
2861                 }
2862                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2863                 osc_object_unlock(obj);
2864                 osc_extent_put(env, ext);
2865
2866                 if (unplug)
2867                         osc_io_unplug_async(env, osc_cli(obj), obj);
2868         }
2869 }
2870
2871 /**
2872  * Wait for extents in a specific range to be written out.
2873  * The caller must have called osc_cache_writeback_range() to issue IO
2874  * otherwise it will take a long time for this function to finish.
2875  *
2876  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2877  * nobody else can dirty this range of file while we're waiting for
2878  * extents to be written.
2879  */
2880 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2881                          pgoff_t start, pgoff_t end)
2882 {
2883         struct osc_extent *ext;
2884         pgoff_t index = start;
2885         int     result = 0;
2886         ENTRY;
2887
2888 again:
2889         osc_object_lock(obj);
2890         ext = osc_extent_search(obj, index);
2891         if (ext == NULL)
2892                 ext = first_extent(obj);
2893         else if (ext->oe_end < index)
2894                 ext = next_extent(ext);
2895         while (ext != NULL) {
2896                 int rc;
2897
2898                 if (ext->oe_start > end)
2899                         break;
2900
2901                 if (!ext->oe_fsync_wait) {
2902                         ext = next_extent(ext);
2903                         continue;
2904                 }
2905
2906                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2907                              ext->oe_hp || ext->oe_urgent), ext);
2908                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2909                              !ext->oe_hp && ext->oe_urgent), ext);
2910
2911                 index = ext->oe_end + 1;
2912                 osc_extent_get(ext);
2913                 osc_object_unlock(obj);
2914
2915                 rc = osc_extent_wait(env, ext, OES_INV);
2916                 if (result == 0)
2917                         result = rc;
2918                 osc_extent_put(env, ext);
2919                 goto again;
2920         }
2921         osc_object_unlock(obj);
2922
2923         OSC_IO_DEBUG(obj, "sync file range.\n");
2924         RETURN(result);
2925 }
2926
2927 /**
2928  * Called to write out a range of osc object.
2929  *
2930  * @hp     : should be set this is caused by lock cancel;
2931  * @discard: is set if dirty pages should be dropped - file will be deleted or
2932  *         truncated, this implies there is no partially discarding extents.
2933  *
2934  * Return how many pages will be issued, or error code if error occurred.
2935  */
2936 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2937                               pgoff_t start, pgoff_t end, int hp, int discard)
2938 {
2939         struct osc_extent *ext;
2940         CFS_LIST_HEAD(discard_list);
2941         bool unplug = false;
2942         int result = 0;
2943         ENTRY;
2944
2945         osc_object_lock(obj);
2946         ext = osc_extent_search(obj, start);
2947         if (ext == NULL)
2948                 ext = first_extent(obj);
2949         else if (ext->oe_end < start)
2950                 ext = next_extent(ext);
2951         while (ext != NULL) {
2952                 if (ext->oe_start > end)
2953                         break;
2954
2955                 ext->oe_fsync_wait = 1;
2956                 switch (ext->oe_state) {
2957                 case OES_CACHE:
2958                         result += ext->oe_nr_pages;
2959                         if (!discard) {
2960                                 cfs_list_t *list = NULL;
2961                                 if (hp) {
2962                                         EASSERT(!ext->oe_hp, ext);
2963                                         ext->oe_hp = 1;
2964                                         list = &obj->oo_hp_exts;
2965                                 } else if (!ext->oe_urgent) {
2966                                         ext->oe_urgent = 1;
2967                                         list = &obj->oo_urgent_exts;
2968                                 }
2969                                 if (list != NULL)
2970                                         cfs_list_move_tail(&ext->oe_link, list);
2971                                 unplug = true;
2972                         } else {
2973                                 /* the only discarder is lock cancelling, so
2974                                  * [start, end] must contain this extent */
2975                                 EASSERT(ext->oe_start >= start &&
2976                                         ext->oe_max_end <= end, ext);
2977                                 osc_extent_state_set(ext, OES_LOCKING);
2978                                 ext->oe_owner = current;
2979                                 cfs_list_move_tail(&ext->oe_link,
2980                                                    &discard_list);
2981                                 osc_update_pending(obj, OBD_BRW_WRITE,
2982                                                    -ext->oe_nr_pages);
2983                         }
2984                         break;
2985                 case OES_ACTIVE:
2986                         /* It's pretty bad to wait for ACTIVE extents, because
2987                          * we don't know how long we will wait for it to be
2988                          * flushed since it may be blocked at awaiting more
2989                          * grants. We do this for the correctness of fsync. */
2990                         LASSERT(hp == 0 && discard == 0);
2991                         ext->oe_urgent = 1;
2992                         break;
2993                 case OES_TRUNC:
2994                         /* this extent is being truncated, can't do anything
2995                          * for it now. it will be set to urgent after truncate
2996                          * is finished in osc_cache_truncate_end(). */
2997                 default:
2998                         break;
2999                 }
3000                 ext = next_extent(ext);
3001         }
3002         osc_object_unlock(obj);
3003
3004         LASSERT(ergo(!discard, cfs_list_empty(&discard_list)));
3005         if (!cfs_list_empty(&discard_list)) {
3006                 struct osc_extent *tmp;
3007                 int rc;
3008
3009                 osc_list_maint(osc_cli(obj), obj);
3010                 cfs_list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
3011                         cfs_list_del_init(&ext->oe_link);
3012                         EASSERT(ext->oe_state == OES_LOCKING, ext);
3013
3014                         /* Discard caching pages. We don't actually write this
3015                          * extent out but we complete it as if we did. */
3016                         rc = osc_extent_make_ready(env, ext);
3017                         if (unlikely(rc < 0)) {
3018                                 OSC_EXTENT_DUMP(D_ERROR, ext,
3019                                                 "make_ready returned %d\n", rc);
3020                                 if (result >= 0)
3021                                         result = rc;
3022                         }
3023
3024                         /* finish the extent as if the pages were sent */
3025                         osc_extent_finish(env, ext, 0, 0);
3026                 }
3027         }
3028
3029         if (unplug)
3030                 osc_io_unplug(env, osc_cli(obj), obj, PDL_POLICY_ROUND);
3031
3032         if (hp || discard) {
3033                 int rc;
3034                 rc = osc_cache_wait_range(env, obj, start, end);
3035                 if (result >= 0 && rc < 0)
3036                         result = rc;
3037         }
3038
3039         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3040         RETURN(result);
3041 }
3042
3043 /** @} osc */