drivers/staging/lustre/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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2015, Intel Corporation.
  27  *
  28  */
  29 /*
  30  * This file is part of Lustre, http://www.lustre.org/
  31  * Lustre is a trademark of Sun Microsystems, Inc.
  32  *
  33  * osc cache management.
  34  *
  35  * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_OSC
  39
  40 #include "osc_cl_internal.h"
  41 #include "osc_internal.h"
  42
  43 static int extent_debug; /* set it to be true for more debug */
  44
  45 static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
  46 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
  47                            enum osc_extent_state state);
  48 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
  49                               struct osc_async_page *oap, int sent, int rc);
  50 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
  51                           int cmd);
  52 static int osc_refresh_count(const struct lu_env *env,
  53                              struct osc_async_page *oap, int cmd);
  54 static int osc_io_unplug_async(const struct lu_env *env,
  55                                struct client_obd *cli, struct osc_object *osc);
  56 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
  57                            unsigned int lost_grant);
  58
  59 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
  60                                   const char *func, int line);
  61 #define osc_extent_tree_dump(lvl, obj) \
  62         osc_extent_tree_dump0(lvl, obj, __func__, __LINE__)
  63
  64 /** \addtogroup osc
  65  *  @{
  66  */
  67
  68 /* ------------------ osc extent ------------------ */
  69 static inline char *ext_flags(struct osc_extent *ext, char *flags)
  70 {
  71         char *buf = flags;
  72         *buf++ = ext->oe_rw ? 'r' : 'w';
  73         if (ext->oe_intree)
  74                 *buf++ = 'i';
  75         if (ext->oe_sync)
  76                 *buf++ = 'S';
  77         if (ext->oe_srvlock)
  78                 *buf++ = 's';
  79         if (ext->oe_hp)
  80                 *buf++ = 'h';
  81         if (ext->oe_urgent)
  82                 *buf++ = 'u';
  83         if (ext->oe_memalloc)
  84                 *buf++ = 'm';
  85         if (ext->oe_trunc_pending)
  86                 *buf++ = 't';
  87         if (ext->oe_fsync_wait)
  88                 *buf++ = 'Y';
  89         *buf = 0;
  90         return flags;
  91 }
  92
  93 static inline char list_empty_marker(struct list_head *list)
  94 {
  95         return list_empty(list) ? '-' : '+';
  96 }
  97
  98 #define EXTSTR       "[%lu -> %lu/%lu]"
  99 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
 100 static const char *oes_strings[] = {
 101         "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
 102
 103 #define OSC_EXTENT_DUMP(lvl, extent, fmt, ...) do {                           \
 104         struct osc_extent *__ext = (extent);                                  \
 105         char __buf[16];                                                       \
 106                                                                               \
 107         CDEBUG(lvl,                                                           \
 108                 "extent %p@{" EXTSTR ", "                                     \
 109                 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt,          \
 110                 /* ----- extent part 0 ----- */                               \
 111                 __ext, EXTPARA(__ext),                                        \
 112                 /* ----- part 1 ----- */                                      \
 113                 atomic_read(&__ext->oe_refc),                                 \
 114                 atomic_read(&__ext->oe_users),                                \
 115                 list_empty_marker(&__ext->oe_link),                           \
 116                 oes_strings[__ext->oe_state], ext_flags(__ext, __buf),        \
 117                 __ext->oe_obj,                                                \
 118                 /* ----- part 2 ----- */                                      \
 119                 __ext->oe_grants, __ext->oe_nr_pages,                         \
 120                 list_empty_marker(&__ext->oe_pages),                          \
 121                 waitqueue_active(&__ext->oe_waitq) ? '+' : '-',               \
 122                 __ext->oe_dlmlock, __ext->oe_mppr, __ext->oe_owner,           \
 123                 /* ----- part 4 ----- */                                      \
 124                 ## __VA_ARGS__);                                              \
 125         if (lvl == D_ERROR && __ext->oe_dlmlock)                              \
 126                 LDLM_ERROR(__ext->oe_dlmlock, "extent: %p", __ext);           \
 127         else                                                                  \
 128                 LDLM_DEBUG(__ext->oe_dlmlock, "extent: %p", __ext);           \
 129 } while (0)
 130
 131 #undef EASSERTF
 132 #define EASSERTF(expr, ext, fmt, args...) do {                          \
 133         if (!(expr)) {                                                  \
 134                 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args);           \
 135                 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj);           \
 136                 LASSERT(expr);                                          \
 137         }                                                               \
 138 } while (0)
 139
 140 #undef EASSERT
 141 #define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
 142
 143 static inline struct osc_extent *rb_extent(struct rb_node *n)
 144 {
 145         if (!n)
 146                 return NULL;
 147
 148         return container_of(n, struct osc_extent, oe_node);
 149 }
 150
 151 static inline struct osc_extent *next_extent(struct osc_extent *ext)
 152 {
 153         if (!ext)
 154                 return NULL;
 155
 156         LASSERT(ext->oe_intree);
 157         return rb_extent(rb_next(&ext->oe_node));
 158 }
 159
 160 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
 161 {
 162         if (!ext)
 163                 return NULL;
 164
 165         LASSERT(ext->oe_intree);
 166         return rb_extent(rb_prev(&ext->oe_node));
 167 }
 168
 169 static inline struct osc_extent *first_extent(struct osc_object *obj)
 170 {
 171         return rb_extent(rb_first(&obj->oo_root));
 172 }
 173
 174 /* object must be locked by caller. */
 175 static int osc_extent_sanity_check0(struct osc_extent *ext,
 176                                     const char *func, const int line)
 177 {
 178         struct osc_object *obj = ext->oe_obj;
 179         struct osc_async_page *oap;
 180         size_t page_count;
 181         int rc = 0;
 182
 183         if (!osc_object_is_locked(obj)) {
 184                 rc = 9;
 185                 goto out;
 186         }
 187
 188         if (ext->oe_state >= OES_STATE_MAX) {
 189                 rc = 10;
 190                 goto out;
 191         }
 192
 193         if (atomic_read(&ext->oe_refc) <= 0) {
 194                 rc = 20;
 195                 goto out;
 196         }
 197
 198         if (atomic_read(&ext->oe_refc) < atomic_read(&ext->oe_users)) {
 199                 rc = 30;
 200                 goto out;
 201         }
 202
 203         switch (ext->oe_state) {
 204         case OES_INV:
 205                 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
 206                         rc = 35;
 207                 else
 208                         rc = 0;
 209                 goto out;
 210         case OES_ACTIVE:
 211                 if (atomic_read(&ext->oe_users) == 0) {
 212                         rc = 40;
 213                         goto out;
 214                 }
 215                 if (ext->oe_hp) {
 216                         rc = 50;
 217                         goto out;
 218                 }
 219                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
 220                         rc = 55;
 221                         goto out;
 222                 }
 223                 break;
 224         case OES_CACHE:
 225                 if (ext->oe_grants == 0) {
 226                         rc = 60;
 227                         goto out;
 228                 }
 229                 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp) {
 230                         rc = 65;
 231                         goto out;
 232                 }
 233         default:
 234                 if (atomic_read(&ext->oe_users) > 0) {
 235                         rc = 70;
 236                         goto out;
 237                 }
 238         }
 239
 240         if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start) {
 241                 rc = 80;
 242                 goto out;
 243         }
 244
 245         if (ext->oe_sync && ext->oe_grants > 0) {
 246                 rc = 90;
 247                 goto out;
 248         }
 249
 250         if (ext->oe_dlmlock) {
 251                 struct ldlm_extent *extent;
 252
 253                 extent = &ext->oe_dlmlock->l_policy_data.l_extent;
 254                 if (!(extent->start <= cl_offset(osc2cl(obj), ext->oe_start) &&
 255                       extent->end >= cl_offset(osc2cl(obj), ext->oe_max_end))) {
 256                         rc = 100;
 257                         goto out;
 258                 }
 259
 260                 if (!(ext->oe_dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP))) {
 261                         rc = 102;
 262                         goto out;
 263                 }
 264         }
 265
 266         if (ext->oe_nr_pages > ext->oe_mppr) {
 267                 rc = 105;
 268                 goto out;
 269         }
 270
 271         /* Do not verify page list if extent is in RPC. This is because an
 272          * in-RPC extent is supposed to be exclusively accessible w/o lock.
 273          */
 274         if (ext->oe_state > OES_CACHE) {
 275                 rc = 0;
 276                 goto out;
 277         }
 278
 279         if (!extent_debug) {
 280                 rc = 0;
 281                 goto out;
 282         }
 283
 284         page_count = 0;
 285         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
 286                 pgoff_t index = osc_index(oap2osc(oap));
 287                 ++page_count;
 288                 if (index > ext->oe_end || index < ext->oe_start) {
 289                         rc = 110;
 290                         goto out;
 291                 }
 292         }
 293         if (page_count != ext->oe_nr_pages) {
 294                 rc = 120;
 295                 goto out;
 296         }
 297
 298 out:
 299         if (rc != 0)
 300                 OSC_EXTENT_DUMP(D_ERROR, ext,
 301                                 "%s:%d sanity check %p failed with rc = %d\n",
 302                                 func, line, ext, rc);
 303         return rc;
 304 }
 305
 306 #define sanity_check_nolock(ext) \
 307         osc_extent_sanity_check0(ext, __func__, __LINE__)
 308
 309 #define sanity_check(ext) ({                                            \
 310         int __res;                                                      \
 311         osc_object_lock((ext)->oe_obj);                                 \
 312         __res = sanity_check_nolock(ext);                               \
 313         osc_object_unlock((ext)->oe_obj);                               \
 314         __res;                                                          \
 315 })
 316
 317 /**
 318  * sanity check - to make sure there is no overlapped extent in the tree.
 319  */
 320 static int osc_extent_is_overlapped(struct osc_object *obj,
 321                                     struct osc_extent *ext)
 322 {
 323         struct osc_extent *tmp;
 324
 325         LASSERT(osc_object_is_locked(obj));
 326
 327         if (!extent_debug)
 328                 return 0;
 329
 330         for (tmp = first_extent(obj); tmp; tmp = next_extent(tmp)) {
 331                 if (tmp == ext)
 332                         continue;
 333                 if (tmp->oe_end >= ext->oe_start &&
 334                     tmp->oe_start <= ext->oe_end)
 335                         return 1;
 336         }
 337         return 0;
 338 }
 339
 340 static void osc_extent_state_set(struct osc_extent *ext, int state)
 341 {
 342         LASSERT(osc_object_is_locked(ext->oe_obj));
 343         LASSERT(state >= OES_INV && state < OES_STATE_MAX);
 344
 345         /* Never try to sanity check a state changing extent :-) */
 346         /* LASSERT(sanity_check_nolock(ext) == 0); */
 347
 348         /* TODO: validate the state machine */
 349         ext->oe_state = state;
 350         wake_up_all(&ext->oe_waitq);
 351 }
 352
 353 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
 354 {
 355         struct osc_extent *ext;
 356
 357         ext = kmem_cache_zalloc(osc_extent_kmem, GFP_NOFS);
 358         if (!ext)
 359                 return NULL;
 360
 361         RB_CLEAR_NODE(&ext->oe_node);
 362         ext->oe_obj = obj;
 363         atomic_set(&ext->oe_refc, 1);
 364         atomic_set(&ext->oe_users, 0);
 365         INIT_LIST_HEAD(&ext->oe_link);
 366         ext->oe_state = OES_INV;
 367         INIT_LIST_HEAD(&ext->oe_pages);
 368         init_waitqueue_head(&ext->oe_waitq);
 369         ext->oe_dlmlock = NULL;
 370
 371         return ext;
 372 }
 373
 374 static void osc_extent_free(struct osc_extent *ext)
 375 {
 376         kmem_cache_free(osc_extent_kmem, ext);
 377 }
 378
 379 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
 380 {
 381         LASSERT(atomic_read(&ext->oe_refc) >= 0);
 382         atomic_inc(&ext->oe_refc);
 383         return ext;
 384 }
 385
 386 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
 387 {
 388         LASSERT(atomic_read(&ext->oe_refc) > 0);
 389         if (atomic_dec_and_test(&ext->oe_refc)) {
 390                 LASSERT(list_empty(&ext->oe_link));
 391                 LASSERT(atomic_read(&ext->oe_users) == 0);
 392                 LASSERT(ext->oe_state == OES_INV);
 393                 LASSERT(!ext->oe_intree);
 394
 395                 if (ext->oe_dlmlock) {
 396                         lu_ref_add(&ext->oe_dlmlock->l_reference,
 397                                    "osc_extent", ext);
 398                         LDLM_LOCK_PUT(ext->oe_dlmlock);
 399                         ext->oe_dlmlock = NULL;
 400                 }
 401                 osc_extent_free(ext);
 402         }
 403 }
 404
 405 /**
 406  * osc_extent_put_trust() is a special version of osc_extent_put() when
 407  * it's known that the caller is not the last user. This is to address the
 408  * problem of lacking of lu_env ;-).
 409  */
 410 static void osc_extent_put_trust(struct osc_extent *ext)
 411 {
 412         LASSERT(atomic_read(&ext->oe_refc) > 1);
 413         LASSERT(osc_object_is_locked(ext->oe_obj));
 414         atomic_dec(&ext->oe_refc);
 415 }
 416
 417 /**
 418  * Return the extent which includes pgoff @index, or return the greatest
 419  * previous extent in the tree.
 420  */
 421 static struct osc_extent *osc_extent_search(struct osc_object *obj,
 422                                             pgoff_t index)
 423 {
 424         struct rb_node *n = obj->oo_root.rb_node;
 425         struct osc_extent *tmp, *p = NULL;
 426
 427         LASSERT(osc_object_is_locked(obj));
 428         while (n) {
 429                 tmp = rb_extent(n);
 430                 if (index < tmp->oe_start) {
 431                         n = n->rb_left;
 432                 } else if (index > tmp->oe_end) {
 433                         p = rb_extent(n);
 434                         n = n->rb_right;
 435                 } else {
 436                         return tmp;
 437                 }
 438         }
 439         return p;
 440 }
 441
 442 /*
 443  * Return the extent covering @index, otherwise return NULL.
 444  * caller must have held object lock.
 445  */
 446 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
 447                                             pgoff_t index)
 448 {
 449         struct osc_extent *ext;
 450
 451         ext = osc_extent_search(obj, index);
 452         if (ext && ext->oe_start <= index && index <= ext->oe_end)
 453                 return osc_extent_get(ext);
 454         return NULL;
 455 }
 456
 457 /* caller must have held object lock. */
 458 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
 459 {
 460         struct rb_node **n = &obj->oo_root.rb_node;
 461         struct rb_node *parent = NULL;
 462         struct osc_extent *tmp;
 463
 464         LASSERT(ext->oe_intree == 0);
 465         LASSERT(ext->oe_obj == obj);
 466         LASSERT(osc_object_is_locked(obj));
 467         while (*n) {
 468                 tmp = rb_extent(*n);
 469                 parent = *n;
 470
 471                 if (ext->oe_end < tmp->oe_start)
 472                         n = &(*n)->rb_left;
 473                 else if (ext->oe_start > tmp->oe_end)
 474                         n = &(*n)->rb_right;
 475                 else
 476                         EASSERTF(0, tmp, EXTSTR"\n", EXTPARA(ext));
 477         }
 478         rb_link_node(&ext->oe_node, parent, n);
 479         rb_insert_color(&ext->oe_node, &obj->oo_root);
 480         osc_extent_get(ext);
 481         ext->oe_intree = 1;
 482 }
 483
 484 /* caller must have held object lock. */
 485 static void osc_extent_erase(struct osc_extent *ext)
 486 {
 487         struct osc_object *obj = ext->oe_obj;
 488
 489         LASSERT(osc_object_is_locked(obj));
 490         if (ext->oe_intree) {
 491                 rb_erase(&ext->oe_node, &obj->oo_root);
 492                 ext->oe_intree = 0;
 493                 /* rbtree held a refcount */
 494                 osc_extent_put_trust(ext);
 495         }
 496 }
 497
 498 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
 499 {
 500         struct osc_object *obj = ext->oe_obj;
 501
 502         LASSERT(osc_object_is_locked(obj));
 503         LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
 504         if (ext->oe_state == OES_CACHE) {
 505                 osc_extent_state_set(ext, OES_ACTIVE);
 506                 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
 507         }
 508         atomic_inc(&ext->oe_users);
 509         list_del_init(&ext->oe_link);
 510         return osc_extent_get(ext);
 511 }
 512
 513 static void __osc_extent_remove(struct osc_extent *ext)
 514 {
 515         LASSERT(osc_object_is_locked(ext->oe_obj));
 516         LASSERT(list_empty(&ext->oe_pages));
 517         osc_extent_erase(ext);
 518         list_del_init(&ext->oe_link);
 519         osc_extent_state_set(ext, OES_INV);
 520         OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
 521 }
 522
 523 static void osc_extent_remove(struct osc_extent *ext)
 524 {
 525         struct osc_object *obj = ext->oe_obj;
 526
 527         osc_object_lock(obj);
 528         __osc_extent_remove(ext);
 529         osc_object_unlock(obj);
 530 }
 531
 532 /**
 533  * This function is used to merge extents to get better performance. It checks
 534  * if @cur and @victim are contiguous at chunk level.
 535  */
 536 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
 537                             struct osc_extent *victim)
 538 {
 539         struct osc_object *obj = cur->oe_obj;
 540         pgoff_t chunk_start;
 541         pgoff_t chunk_end;
 542         int ppc_bits;
 543
 544         LASSERT(cur->oe_state == OES_CACHE);
 545         LASSERT(osc_object_is_locked(obj));
 546         if (!victim)
 547                 return -EINVAL;
 548
 549         if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
 550                 return -EBUSY;
 551
 552         if (cur->oe_max_end != victim->oe_max_end)
 553                 return -ERANGE;
 554
 555         LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
 556         ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_SHIFT;
 557         chunk_start = cur->oe_start >> ppc_bits;
 558         chunk_end = cur->oe_end >> ppc_bits;
 559         if (chunk_start != (victim->oe_end >> ppc_bits) + 1 &&
 560             chunk_end + 1 != victim->oe_start >> ppc_bits)
 561                 return -ERANGE;
 562
 563         OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
 564
 565         cur->oe_start = min(cur->oe_start, victim->oe_start);
 566         cur->oe_end = max(cur->oe_end, victim->oe_end);
 567         cur->oe_grants += victim->oe_grants;
 568         cur->oe_nr_pages += victim->oe_nr_pages;
 569         /* only the following bits are needed to merge */
 570         cur->oe_urgent |= victim->oe_urgent;
 571         cur->oe_memalloc |= victim->oe_memalloc;
 572         list_splice_init(&victim->oe_pages, &cur->oe_pages);
 573         list_del_init(&victim->oe_link);
 574         victim->oe_nr_pages = 0;
 575
 576         osc_extent_get(victim);
 577         __osc_extent_remove(victim);
 578         osc_extent_put(env, victim);
 579
 580         OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
 581         return 0;
 582 }
 583
 584 /**
 585  * Drop user count of osc_extent, and unplug IO asynchronously.
 586  */
 587 void osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
 588 {
 589         struct osc_object *obj = ext->oe_obj;
 590
 591         LASSERT(atomic_read(&ext->oe_users) > 0);
 592         LASSERT(sanity_check(ext) == 0);
 593         LASSERT(ext->oe_grants > 0);
 594
 595         if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
 596                 LASSERT(ext->oe_state == OES_ACTIVE);
 597                 if (ext->oe_trunc_pending) {
 598                         /* a truncate process is waiting for this extent.
 599                          * This may happen due to a race, check
 600                          * osc_cache_truncate_start().
 601                          */
 602                         osc_extent_state_set(ext, OES_TRUNC);
 603                         ext->oe_trunc_pending = 0;
 604                 } else {
 605                         osc_extent_state_set(ext, OES_CACHE);
 606                         osc_update_pending(obj, OBD_BRW_WRITE,
 607                                            ext->oe_nr_pages);
 608
 609                         /* try to merge the previous and next extent. */
 610                         osc_extent_merge(env, ext, prev_extent(ext));
 611                         osc_extent_merge(env, ext, next_extent(ext));
 612
 613                         if (ext->oe_urgent)
 614                                 list_move_tail(&ext->oe_link,
 615                                                &obj->oo_urgent_exts);
 616                 }
 617                 osc_object_unlock(obj);
 618
 619                 osc_io_unplug_async(env, osc_cli(obj), obj);
 620         }
 621         osc_extent_put(env, ext);
 622 }
 623
 624 static inline int overlapped(struct osc_extent *ex1, struct osc_extent *ex2)
 625 {
 626         return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
 627 }
 628
 629 /**
 630  * Find or create an extent which includes @index, core function to manage
 631  * extent tree.
 632  */
 633 static struct osc_extent *osc_extent_find(const struct lu_env *env,
 634                                           struct osc_object *obj, pgoff_t index,
 635                                           unsigned int *grants)
 636 {
 637         struct client_obd *cli = osc_cli(obj);
 638         struct osc_lock   *olck;
 639         struct cl_lock_descr *descr;
 640         struct osc_extent *cur;
 641         struct osc_extent *ext;
 642         struct osc_extent *conflict = NULL;
 643         struct osc_extent *found = NULL;
 644         pgoff_t chunk;
 645         pgoff_t max_end;
 646         unsigned int max_pages; /* max_pages_per_rpc */
 647         unsigned int chunksize;
 648         int ppc_bits; /* pages per chunk bits */
 649         pgoff_t chunk_mask;
 650         int rc;
 651
 652         cur = osc_extent_alloc(obj);
 653         if (!cur)
 654                 return ERR_PTR(-ENOMEM);
 655
 656         olck = osc_env_io(env)->oi_write_osclock;
 657         LASSERTF(olck, "page %lu is not covered by lock\n", index);
 658         LASSERT(olck->ols_state == OLS_GRANTED);
 659
 660         descr = &olck->ols_cl.cls_lock->cll_descr;
 661         LASSERT(descr->cld_mode >= CLM_WRITE);
 662
 663         LASSERT(cli->cl_chunkbits >= PAGE_SHIFT);
 664         ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
 665         chunk_mask = ~((1 << ppc_bits) - 1);
 666         chunksize = 1 << cli->cl_chunkbits;
 667         chunk = index >> ppc_bits;
 668
 669         /* align end to rpc edge, rpc size may not be a power 2 integer. */
 670         max_pages = cli->cl_max_pages_per_rpc;
 671         LASSERT((max_pages & ~chunk_mask) == 0);
 672         max_end = index - (index % max_pages) + max_pages - 1;
 673         max_end = min_t(pgoff_t, max_end, descr->cld_end);
 674
 675         /* initialize new extent by parameters so far */
 676         cur->oe_max_end = max_end;
 677         cur->oe_start = index & chunk_mask;
 678         cur->oe_end = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
 679         if (cur->oe_start < descr->cld_start)
 680                 cur->oe_start = descr->cld_start;
 681         if (cur->oe_end > max_end)
 682                 cur->oe_end = max_end;
 683         cur->oe_grants = 0;
 684         cur->oe_mppr = max_pages;
 685         if (olck->ols_dlmlock) {
 686                 LASSERT(olck->ols_hold);
 687                 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
 688                 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
 689         }
 690
 691         /* grants has been allocated by caller */
 692         LASSERTF(*grants >= chunksize + cli->cl_extent_tax,
 693                  "%u/%u/%u.\n", *grants, chunksize, cli->cl_extent_tax);
 694         LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
 695                  EXTPARA(cur));
 696
 697 restart:
 698         osc_object_lock(obj);
 699         ext = osc_extent_search(obj, cur->oe_start);
 700         if (!ext)
 701                 ext = first_extent(obj);
 702         while (ext) {
 703                 pgoff_t ext_chk_start = ext->oe_start >> ppc_bits;
 704                 pgoff_t ext_chk_end = ext->oe_end >> ppc_bits;
 705
 706                 LASSERT(sanity_check_nolock(ext) == 0);
 707                 if (chunk > ext_chk_end + 1)
 708                         break;
 709
 710                 /* if covering by different locks, no chance to match */
 711                 if (olck->ols_dlmlock != ext->oe_dlmlock) {
 712                         EASSERTF(!overlapped(ext, cur), ext,
 713                                  EXTSTR"\n", EXTPARA(cur));
 714
 715                         ext = next_extent(ext);
 716                         continue;
 717                 }
 718
 719                 /* discontiguous chunks? */
 720                 if (chunk + 1 < ext_chk_start) {
 721                         ext = next_extent(ext);
 722                         continue;
 723                 }
 724
 725                 /* ok, from now on, ext and cur have these attrs:
 726                  * 1. covered by the same lock
 727                  * 2. contiguous at chunk level or overlapping.
 728                  */
 729
 730                 if (overlapped(ext, cur)) {
 731                         /* cur is the minimum unit, so overlapping means
 732                          * full contain.
 733                          */
 734                         EASSERTF((ext->oe_start <= cur->oe_start &&
 735                                   ext->oe_end >= cur->oe_end),
 736                                  ext, EXTSTR"\n", EXTPARA(cur));
 737
 738                         if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
 739                                 /* for simplicity, we wait for this extent to
 740                                  * finish before going forward.
 741                                  */
 742                                 conflict = osc_extent_get(ext);
 743                                 break;
 744                         }
 745
 746                         found = osc_extent_hold(ext);
 747                         break;
 748                 }
 749
 750                 /* non-overlapped extent */
 751                 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
 752                         /* we can't do anything for a non OES_CACHE extent, or
 753                          * if there is someone waiting for this extent to be
 754                          * flushed, try next one.
 755                          */
 756                         ext = next_extent(ext);
 757                         continue;
 758                 }
 759
 760                 /* check if they belong to the same rpc slot before trying to
 761                  * merge. the extents are not overlapped and contiguous at
 762                  * chunk level to get here.
 763                  */
 764                 if (ext->oe_max_end != max_end) {
 765                         /* if they don't belong to the same RPC slot or
 766                          * max_pages_per_rpc has ever changed, do not merge.
 767                          */
 768                         ext = next_extent(ext);
 769                         continue;
 770                 }
 771
 772                 /* it's required that an extent must be contiguous at chunk
 773                  * level so that we know the whole extent is covered by grant
 774                  * (the pages in the extent are NOT required to be contiguous).
 775                  * Otherwise, it will be too much difficult to know which
 776                  * chunks have grants allocated.
 777                  */
 778
 779                 /* try to do front merge - extend ext's start */
 780                 if (chunk + 1 == ext_chk_start) {
 781                         /* ext must be chunk size aligned */
 782                         EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
 783
 784                         /* pull ext's start back to cover cur */
 785                         ext->oe_start = cur->oe_start;
 786                         ext->oe_grants += chunksize;
 787                         *grants -= chunksize;
 788
 789                         found = osc_extent_hold(ext);
 790                 } else if (chunk == ext_chk_end + 1) {
 791                         /* rear merge */
 792                         ext->oe_end = cur->oe_end;
 793                         ext->oe_grants += chunksize;
 794                         *grants -= chunksize;
 795
 796                         /* try to merge with the next one because we just fill
 797                          * in a gap
 798                          */
 799                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 800                                 /* we can save extent tax from next extent */
 801                                 *grants += cli->cl_extent_tax;
 802
 803                         found = osc_extent_hold(ext);
 804                 }
 805                 if (found)
 806                         break;
 807
 808                 ext = next_extent(ext);
 809         }
 810
 811         osc_extent_tree_dump(D_CACHE, obj);
 812         if (found) {
 813                 LASSERT(!conflict);
 814                 if (!IS_ERR(found)) {
 815                         LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
 816                         OSC_EXTENT_DUMP(D_CACHE, found,
 817                                         "found caching ext for %lu.\n", index);
 818                 }
 819         } else if (!conflict) {
 820                 /* create a new extent */
 821                 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
 822                 cur->oe_grants = chunksize + cli->cl_extent_tax;
 823                 *grants -= cur->oe_grants;
 824                 LASSERT(*grants >= 0);
 825
 826                 cur->oe_state = OES_CACHE;
 827                 found = osc_extent_hold(cur);
 828                 osc_extent_insert(obj, cur);
 829                 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
 830                                 index, descr->cld_end);
 831         }
 832         osc_object_unlock(obj);
 833
 834         if (conflict) {
 835                 LASSERT(!found);
 836
 837                 /* waiting for IO to finish. Please notice that it's impossible
 838                  * to be an OES_TRUNC extent.
 839                  */
 840                 rc = osc_extent_wait(env, conflict, OES_INV);
 841                 osc_extent_put(env, conflict);
 842                 conflict = NULL;
 843                 if (rc < 0) {
 844                         found = ERR_PTR(rc);
 845                         goto out;
 846                 }
 847
 848                 goto restart;
 849         }
 850
 851 out:
 852         osc_extent_put(env, cur);
 853         LASSERT(*grants >= 0);
 854         return found;
 855 }
 856
 857 /**
 858  * Called when IO is finished to an extent.
 859  */
 860 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
 861                       int sent, int rc)
 862 {
 863         struct client_obd *cli = osc_cli(ext->oe_obj);
 864         struct osc_async_page *oap;
 865         struct osc_async_page *tmp;
 866         int nr_pages = ext->oe_nr_pages;
 867         int lost_grant = 0;
 868         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
 869         __u64 last_off = 0;
 870         int last_count = -1;
 871
 872         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
 873
 874         ext->oe_rc = rc ?: ext->oe_nr_pages;
 875         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
 876
 877         osc_lru_add_batch(cli, &ext->oe_pages);
 878         list_for_each_entry_safe(oap, tmp, &ext->oe_pages, oap_pending_item) {
 879                 list_del_init(&oap->oap_rpc_item);
 880                 list_del_init(&oap->oap_pending_item);
 881                 if (last_off <= oap->oap_obj_off) {
 882                         last_off = oap->oap_obj_off;
 883                         last_count = oap->oap_count;
 884                 }
 885
 886                 --ext->oe_nr_pages;
 887                 osc_ap_completion(env, cli, oap, sent, rc);
 888         }
 889         EASSERT(ext->oe_nr_pages == 0, ext);
 890
 891         if (!sent) {
 892                 lost_grant = ext->oe_grants;
 893         } else if (blocksize < PAGE_SIZE &&
 894                    last_count != PAGE_SIZE) {
 895                 /* For short writes we shouldn't count parts of pages that
 896                  * span a whole chunk on the OST side, or our accounting goes
 897                  * wrong.  Should match the code in filter_grant_check.
 898                  */
 899                 int offset = last_off & ~PAGE_MASK;
 900                 int count = last_count + (offset & (blocksize - 1));
 901                 int end = (offset + last_count) & (blocksize - 1);
 902                 if (end)
 903                         count += blocksize - end;
 904
 905                 lost_grant = PAGE_SIZE - count;
 906         }
 907         if (ext->oe_grants > 0)
 908                 osc_free_grant(cli, nr_pages, lost_grant);
 909
 910         osc_extent_remove(ext);
 911         /* put the refcount for RPC */
 912         osc_extent_put(env, ext);
 913         return 0;
 914 }
 915
 916 static int extent_wait_cb(struct osc_extent *ext, enum osc_extent_state state)
 917 {
 918         int ret;
 919
 920         osc_object_lock(ext->oe_obj);
 921         ret = ext->oe_state == state;
 922         osc_object_unlock(ext->oe_obj);
 923
 924         return ret;
 925 }
 926
 927 /**
 928  * Wait for the extent's state to become @state.
 929  */
 930 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
 931                            enum osc_extent_state state)
 932 {
 933         struct osc_object *obj = ext->oe_obj;
 934         struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
 935                                                   LWI_ON_SIGNAL_NOOP, NULL);
 936         int rc = 0;
 937
 938         osc_object_lock(obj);
 939         LASSERT(sanity_check_nolock(ext) == 0);
 940         /* `Kick' this extent only if the caller is waiting for it to be
 941          * written out.
 942          */
 943         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp &&
 944             !ext->oe_trunc_pending) {
 945                 if (ext->oe_state == OES_ACTIVE) {
 946                         ext->oe_urgent = 1;
 947                 } else if (ext->oe_state == OES_CACHE) {
 948                         ext->oe_urgent = 1;
 949                         osc_extent_hold(ext);
 950                         rc = 1;
 951                 }
 952         }
 953         osc_object_unlock(obj);
 954         if (rc == 1)
 955                 osc_extent_release(env, ext);
 956
 957         /* wait for the extent until its state becomes @state */
 958         rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
 959         if (rc == -ETIMEDOUT) {
 960                 OSC_EXTENT_DUMP(D_ERROR, ext,
 961                                 "%s: wait ext to %u timedout, recovery in progress?\n",
 962                                 osc_export(obj)->exp_obd->obd_name, state);
 963
 964                 lwi = LWI_INTR(NULL, NULL);
 965                 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state),
 966                                   &lwi);
 967         }
 968         if (rc == 0 && ext->oe_rc < 0)
 969                 rc = ext->oe_rc;
 970         return rc;
 971 }
 972
 973 /**
 974  * Discard pages with index greater than @size. If @ext is overlapped with
 975  * @size, then partial truncate happens.
 976  */
 977 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
 978                                bool partial)
 979 {
 980         struct cl_env_nest nest;
 981         struct lu_env *env;
 982         struct cl_io *io;
 983         struct osc_object *obj = ext->oe_obj;
 984         struct client_obd *cli = osc_cli(obj);
 985         struct osc_async_page *oap;
 986         struct osc_async_page *tmp;
 987         int pages_in_chunk = 0;
 988         int ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
 989         __u64 trunc_chunk = trunc_index >> ppc_bits;
 990         int grants = 0;
 991         int nr_pages = 0;
 992         int rc = 0;
 993
 994         LASSERT(sanity_check(ext) == 0);
 995         EASSERT(ext->oe_state == OES_TRUNC, ext);
 996         EASSERT(!ext->oe_urgent, ext);
 997
 998         /* Request new lu_env.
 999          * We can't use that env from osc_cache_truncate_start() because
1000          * it's from lov_io_sub and not fully initialized.
1001          */
1002         env = cl_env_nested_get(&nest);
1003         io  = &osc_env_info(env)->oti_io;
1004         io->ci_obj = cl_object_top(osc2cl(obj));
1005         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
1006         if (rc < 0)
1007                 goto out;
1008
1009         /* discard all pages with index greater then trunc_index */
1010         list_for_each_entry_safe(oap, tmp, &ext->oe_pages, oap_pending_item) {
1011                 pgoff_t index = osc_index(oap2osc(oap));
1012                 struct cl_page *page = oap2cl_page(oap);
1013
1014                 LASSERT(list_empty(&oap->oap_rpc_item));
1015
1016                 /* only discard the pages with their index greater than
1017                  * trunc_index, and ...
1018                  */
1019                 if (index < trunc_index ||
1020                     (index == trunc_index && partial)) {
1021                         /* accounting how many pages remaining in the chunk
1022                          * so that we can calculate grants correctly. */
1023                         if (index >> ppc_bits == trunc_chunk)
1024                                 ++pages_in_chunk;
1025                         continue;
1026                 }
1027
1028                 list_del_init(&oap->oap_pending_item);
1029
1030                 cl_page_get(page);
1031                 lu_ref_add(&page->cp_reference, "truncate", current);
1032
1033                 if (cl_page_own(env, io, page) == 0) {
1034                         cl_page_discard(env, io, page);
1035                         cl_page_disown(env, io, page);
1036                 } else {
1037                         LASSERT(page->cp_state == CPS_FREEING);
1038                         LASSERT(0);
1039                 }
1040
1041                 lu_ref_del(&page->cp_reference, "truncate", current);
1042                 cl_page_put(env, page);
1043
1044                 --ext->oe_nr_pages;
1045                 ++nr_pages;
1046         }
1047         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1048                       ext->oe_nr_pages == 0),
1049                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1050
1051         osc_object_lock(obj);
1052         if (ext->oe_nr_pages == 0) {
1053                 LASSERT(pages_in_chunk == 0);
1054                 grants = ext->oe_grants;
1055                 ext->oe_grants = 0;
1056         } else { /* calculate how many grants we can free */
1057                 int chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1058                 pgoff_t last_index;
1059
1060                 /* if there is no pages in this chunk, we can also free grants
1061                  * for the last chunk
1062                  */
1063                 if (pages_in_chunk == 0) {
1064                         /* if this is the 1st chunk and no pages in this chunk,
1065                          * ext->oe_nr_pages must be zero, so we should be in
1066                          * the other if-clause.
1067                          */
1068                         LASSERT(trunc_chunk > 0);
1069                         --trunc_chunk;
1070                         ++chunks;
1071                 }
1072
1073                 /* this is what we can free from this extent */
1074                 grants = chunks << cli->cl_chunkbits;
1075                 ext->oe_grants -= grants;
1076                 last_index = ((trunc_chunk + 1) << ppc_bits) - 1;
1077                 ext->oe_end = min(last_index, ext->oe_max_end);
1078                 LASSERT(ext->oe_end >= ext->oe_start);
1079                 LASSERT(ext->oe_grants > 0);
1080         }
1081         osc_object_unlock(obj);
1082
1083         if (grants > 0 || nr_pages > 0)
1084                 osc_free_grant(cli, nr_pages, grants);
1085
1086 out:
1087         cl_io_fini(env, io);
1088         cl_env_nested_put(&nest, env);
1089         return rc;
1090 }
1091
1092 /**
1093  * This function is used to make the extent prepared for transfer.
1094  * A race with flushing page - ll_writepage() has to be handled cautiously.
1095  */
1096 static int osc_extent_make_ready(const struct lu_env *env,
1097                                  struct osc_extent *ext)
1098 {
1099         struct osc_async_page *oap;
1100         struct osc_async_page *last = NULL;
1101         struct osc_object *obj = ext->oe_obj;
1102         unsigned int page_count = 0;
1103         int rc;
1104
1105         /* we're going to grab page lock, so object lock must not be taken. */
1106         LASSERT(sanity_check(ext) == 0);
1107         /* in locking state, any process should not touch this extent. */
1108         EASSERT(ext->oe_state == OES_LOCKING, ext);
1109         EASSERT(ext->oe_owner, ext);
1110
1111         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1112
1113         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1114                 ++page_count;
1115                 if (!last || last->oap_obj_off < oap->oap_obj_off)
1116                         last = oap;
1117
1118                 /* checking ASYNC_READY is race safe */
1119                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1120                         continue;
1121
1122                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1123                 switch (rc) {
1124                 case 0:
1125                         spin_lock(&oap->oap_lock);
1126                         oap->oap_async_flags |= ASYNC_READY;
1127                         spin_unlock(&oap->oap_lock);
1128                         break;
1129                 case -EALREADY:
1130                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1131                         break;
1132                 default:
1133                         LASSERTF(0, "unknown return code: %d\n", rc);
1134                 }
1135         }
1136
1137         LASSERT(page_count == ext->oe_nr_pages);
1138         LASSERT(last);
1139         /* the last page is the only one we need to refresh its count by
1140          * the size of file.
1141          */
1142         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1143                 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1144
1145                 LASSERT(last_oap_count > 0);
1146                 LASSERT(last->oap_page_off + last_oap_count <= PAGE_SIZE);
1147                 last->oap_count = last_oap_count;
1148                 spin_lock(&last->oap_lock);
1149                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1150                 spin_unlock(&last->oap_lock);
1151         }
1152
1153         /* for the rest of pages, we don't need to call osf_refresh_count()
1154          * because it's known they are not the last page
1155          */
1156         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1157                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1158                         oap->oap_count = PAGE_SIZE - oap->oap_page_off;
1159                         spin_lock(&last->oap_lock);
1160                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1161                         spin_unlock(&last->oap_lock);
1162                 }
1163         }
1164
1165         osc_object_lock(obj);
1166         osc_extent_state_set(ext, OES_RPC);
1167         osc_object_unlock(obj);
1168         /* get a refcount for RPC. */
1169         osc_extent_get(ext);
1170
1171         return 0;
1172 }
1173
1174 /**
1175  * Quick and simple version of osc_extent_find(). This function is frequently
1176  * called to expand the extent for the same IO. To expand the extent, the
1177  * page index must be in the same or next chunk of ext->oe_end.
1178  */
1179 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index,
1180                              unsigned int *grants)
1181 {
1182         struct osc_object *obj = ext->oe_obj;
1183         struct client_obd *cli = osc_cli(obj);
1184         struct osc_extent *next;
1185         int ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1186         pgoff_t chunk = index >> ppc_bits;
1187         pgoff_t end_chunk;
1188         pgoff_t end_index;
1189         unsigned int chunksize = 1 << cli->cl_chunkbits;
1190         int rc = 0;
1191
1192         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1193         osc_object_lock(obj);
1194         LASSERT(sanity_check_nolock(ext) == 0);
1195         end_chunk = ext->oe_end >> ppc_bits;
1196         if (chunk > end_chunk + 1) {
1197                 rc = -ERANGE;
1198                 goto out;
1199         }
1200
1201         if (end_chunk >= chunk) {
1202                 rc = 0;
1203                 goto out;
1204         }
1205
1206         LASSERT(end_chunk + 1 == chunk);
1207         /* try to expand this extent to cover @index */
1208         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1209
1210         next = next_extent(ext);
1211         if (next && next->oe_start <= end_index) {
1212                 /* complex mode - overlapped with the next extent,
1213                  * this case will be handled by osc_extent_find()
1214                  */
1215                 rc = -EAGAIN;
1216                 goto out;
1217         }
1218
1219         ext->oe_end = end_index;
1220         ext->oe_grants += chunksize;
1221         *grants -= chunksize;
1222         LASSERT(*grants >= 0);
1223         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1224                  "overlapped after expanding for %lu.\n", index);
1225
1226 out:
1227         osc_object_unlock(obj);
1228         return rc;
1229 }
1230
1231 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
1232                                   const char *func, int line)
1233 {
1234         struct osc_extent *ext;
1235         int cnt;
1236
1237         CDEBUG(level, "Dump object %p extents at %s:%d, mppr: %u.\n",
1238                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1239
1240         /* osc_object_lock(obj); */
1241         cnt = 1;
1242         for (ext = first_extent(obj); ext; ext = next_extent(ext))
1243                 OSC_EXTENT_DUMP(level, ext, "in tree %d.\n", cnt++);
1244
1245         cnt = 1;
1246         list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1247                 OSC_EXTENT_DUMP(level, ext, "hp %d.\n", cnt++);
1248
1249         cnt = 1;
1250         list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1251                 OSC_EXTENT_DUMP(level, ext, "urgent %d.\n", cnt++);
1252
1253         cnt = 1;
1254         list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1255                 OSC_EXTENT_DUMP(level, ext, "reading %d.\n", cnt++);
1256         /* osc_object_unlock(obj); */
1257 }
1258
1259 /* ------------------ osc extent end ------------------ */
1260
1261 static inline int osc_is_ready(struct osc_object *osc)
1262 {
1263         return !list_empty(&osc->oo_ready_item) ||
1264                !list_empty(&osc->oo_hp_ready_item);
1265 }
1266
1267 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1268         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1269                (OSC), osc_is_ready(OSC),                                       \
1270                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1271                list_empty_marker(&(OSC)->oo_ready_item),                       \
1272                atomic_read(&(OSC)->oo_nr_writes),                              \
1273                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1274                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1275                atomic_read(&(OSC)->oo_nr_reads),                               \
1276                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1277                ##args)
1278
1279 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1280                           int cmd)
1281 {
1282         struct osc_page *opg = oap2osc_page(oap);
1283         struct cl_page  *page = oap2cl_page(oap);
1284         int result;
1285
1286         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1287
1288         result = cl_page_make_ready(env, page, CRT_WRITE);
1289         if (result == 0)
1290                 opg->ops_submit_time = cfs_time_current();
1291         return result;
1292 }
1293
1294 static int osc_refresh_count(const struct lu_env *env,
1295                              struct osc_async_page *oap, int cmd)
1296 {
1297         struct osc_page *opg = oap2osc_page(oap);
1298         pgoff_t index = osc_index(oap2osc(oap));
1299         struct cl_object *obj;
1300         struct cl_attr *attr = &osc_env_info(env)->oti_attr;
1301
1302         int result;
1303         loff_t kms;
1304
1305         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1306         LASSERT(!(cmd & OBD_BRW_READ));
1307         obj = opg->ops_cl.cpl_obj;
1308
1309         cl_object_attr_lock(obj);
1310         result = cl_object_attr_get(env, obj, attr);
1311         cl_object_attr_unlock(obj);
1312         if (result < 0)
1313                 return result;
1314         kms = attr->cat_kms;
1315         if (cl_offset(obj, index) >= kms)
1316                 /* catch race with truncate */
1317                 return 0;
1318         else if (cl_offset(obj, index + 1) > kms)
1319                 /* catch sub-page write at end of file */
1320                 return kms % PAGE_SIZE;
1321         else
1322                 return PAGE_SIZE;
1323 }
1324
1325 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1326                           int cmd, int rc)
1327 {
1328         struct osc_page *opg = oap2osc_page(oap);
1329         struct cl_page    *page = oap2cl_page(oap);
1330         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
1331         enum cl_req_type crt;
1332         int srvlock;
1333
1334         cmd &= ~OBD_BRW_NOQUOTA;
1335         LASSERTF(equi(page->cp_state == CPS_PAGEIN, cmd == OBD_BRW_READ),
1336                  "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1337         LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1338                  "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1339         LASSERT(opg->ops_transfer_pinned);
1340
1341         /*
1342          * page->cp_req can be NULL if io submission failed before
1343          * cl_req was allocated.
1344          */
1345         if (page->cp_req)
1346                 cl_req_page_done(env, page);
1347         LASSERT(!page->cp_req);
1348
1349         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1350         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1351         opg->ops_transfer_pinned = 0;
1352
1353         spin_lock(&obj->oo_seatbelt);
1354         LASSERT(opg->ops_submitter);
1355         LASSERT(!list_empty(&opg->ops_inflight));
1356         list_del_init(&opg->ops_inflight);
1357         opg->ops_submitter = NULL;
1358         spin_unlock(&obj->oo_seatbelt);
1359
1360         opg->ops_submit_time = 0;
1361         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1362
1363         /* statistic */
1364         if (rc == 0 && srvlock) {
1365                 struct lu_device *ld = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1366                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1367                 size_t bytes = oap->oap_count;
1368
1369                 if (crt == CRT_READ)
1370                         stats->os_lockless_reads += bytes;
1371                 else
1372                         stats->os_lockless_writes += bytes;
1373         }
1374
1375         /*
1376          * This has to be the last operation with the page, as locks are
1377          * released in cl_page_completion() and nothing except for the
1378          * reference counter protects page from concurrent reclaim.
1379          */
1380         lu_ref_del(&page->cp_reference, "transfer", page);
1381
1382         cl_page_completion(env, page, crt, rc);
1383
1384         return 0;
1385 }
1386
1387 #define OSC_DUMP_GRANT(lvl, cli, fmt, args...) do {                           \
1388         struct client_obd *__tmp = (cli);                                     \
1389         CDEBUG(lvl, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu "        \
1390                "dropped: %ld avail: %ld, reserved: %ld, flight: %d }"         \
1391                "lru {in list: %ld, left: %ld, waiters: %d }" fmt "\n",        \
1392                __tmp->cl_import->imp_obd->obd_name,                           \
1393                __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages,              \
1394                atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages,       \
1395                __tmp->cl_lost_grant, __tmp->cl_avail_grant,                   \
1396                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,               \
1397                atomic_long_read(&__tmp->cl_lru_in_list),                      \
1398                atomic_long_read(&__tmp->cl_lru_busy),                         \
1399                atomic_read(&__tmp->cl_lru_shrinkers), ##args);                \
1400 } while (0)
1401
1402 /* caller must hold loi_list_lock */
1403 static void osc_consume_write_grant(struct client_obd *cli,
1404                                     struct brw_page *pga)
1405 {
1406         assert_spin_locked(&cli->cl_loi_list_lock);
1407         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1408         atomic_long_inc(&obd_dirty_pages);
1409         cli->cl_dirty_pages++;
1410         pga->flag |= OBD_BRW_FROM_GRANT;
1411         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1412                PAGE_SIZE, pga, pga->pg);
1413         osc_update_next_shrink(cli);
1414 }
1415
1416 /* the companion to osc_consume_write_grant, called when a brw has completed.
1417  * must be called with the loi lock held.
1418  */
1419 static void osc_release_write_grant(struct client_obd *cli,
1420                                     struct brw_page *pga)
1421 {
1422         assert_spin_locked(&cli->cl_loi_list_lock);
1423         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1424                 return;
1425         }
1426
1427         pga->flag &= ~OBD_BRW_FROM_GRANT;
1428         atomic_long_dec(&obd_dirty_pages);
1429         cli->cl_dirty_pages--;
1430         if (pga->flag & OBD_BRW_NOCACHE) {
1431                 pga->flag &= ~OBD_BRW_NOCACHE;
1432                 atomic_long_dec(&obd_dirty_transit_pages);
1433                 cli->cl_dirty_transit--;
1434         }
1435 }
1436
1437 /**
1438  * To avoid sleeping with object lock held, it's good for us allocate enough
1439  * grants before entering into critical section.
1440  *
1441  * spin_lock held by caller
1442  */
1443 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1444 {
1445         int rc = -EDQUOT;
1446
1447         if (cli->cl_avail_grant >= bytes) {
1448                 cli->cl_avail_grant -= bytes;
1449                 cli->cl_reserved_grant += bytes;
1450                 rc = 0;
1451         }
1452         return rc;
1453 }
1454
1455 static void __osc_unreserve_grant(struct client_obd *cli,
1456                                   unsigned int reserved, unsigned int unused)
1457 {
1458         /* it's quite normal for us to get more grant than reserved.
1459          * Thinking about a case that two extents merged by adding a new
1460          * chunk, we can save one extent tax. If extent tax is greater than
1461          * one chunk, we can save more grant by adding a new chunk
1462          */
1463         cli->cl_reserved_grant -= reserved;
1464         if (unused > reserved) {
1465                 cli->cl_avail_grant += reserved;
1466                 cli->cl_lost_grant  += unused - reserved;
1467         } else {
1468                 cli->cl_avail_grant += unused;
1469         }
1470 }
1471
1472 static void osc_unreserve_grant(struct client_obd *cli,
1473                                 unsigned int reserved, unsigned int unused)
1474 {
1475         spin_lock(&cli->cl_loi_list_lock);
1476         __osc_unreserve_grant(cli, reserved, unused);
1477         if (unused > 0)
1478                 osc_wake_cache_waiters(cli);
1479         spin_unlock(&cli->cl_loi_list_lock);
1480 }
1481
1482 /**
1483  * Free grant after IO is finished or canceled.
1484  *
1485  * @lost_grant is used to remember how many grants we have allocated but not
1486  * used, we should return these grants to OST. There're two cases where grants
1487  * can be lost:
1488  * 1. truncate;
1489  * 2. blocksize at OST is less than PAGE_SIZE and a partial page was
1490  *    written. In this case OST may use less chunks to serve this partial
1491  *    write. OSTs don't actually know the page size on the client side. so
1492  *    clients have to calculate lost grant by the blocksize on the OST.
1493  *    See filter_grant_check() for details.
1494  */
1495 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1496                            unsigned int lost_grant)
1497 {
1498         unsigned long grant = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
1499
1500         spin_lock(&cli->cl_loi_list_lock);
1501         atomic_long_sub(nr_pages, &obd_dirty_pages);
1502         cli->cl_dirty_pages -= nr_pages;
1503         cli->cl_lost_grant += lost_grant;
1504         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1505                 /* borrow some grant from truncate to avoid the case that
1506                  * truncate uses up all avail grant
1507                  */
1508                 cli->cl_lost_grant -= grant;
1509                 cli->cl_avail_grant += grant;
1510         }
1511         osc_wake_cache_waiters(cli);
1512         spin_unlock(&cli->cl_loi_list_lock);
1513         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu\n",
1514                lost_grant, cli->cl_lost_grant,
1515                cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_SHIFT);
1516 }
1517
1518 /**
1519  * The companion to osc_enter_cache(), called when @oap is no longer part of
1520  * the dirty accounting due to error.
1521  */
1522 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1523 {
1524         spin_lock(&cli->cl_loi_list_lock);
1525         osc_release_write_grant(cli, &oap->oap_brw_page);
1526         spin_unlock(&cli->cl_loi_list_lock);
1527 }
1528
1529 /**
1530  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1531  * is available.
1532  */
1533 static int osc_enter_cache_try(struct client_obd *cli,
1534                                struct osc_async_page *oap,
1535                                int bytes, int transient)
1536 {
1537         int rc;
1538
1539         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1540
1541         rc = osc_reserve_grant(cli, bytes);
1542         if (rc < 0)
1543                 return 0;
1544
1545         if (cli->cl_dirty_pages <= cli->cl_dirty_max_pages &&
1546             atomic_long_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages) {
1547                 osc_consume_write_grant(cli, &oap->oap_brw_page);
1548                 if (transient) {
1549                         cli->cl_dirty_transit++;
1550                         atomic_long_inc(&obd_dirty_transit_pages);
1551                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
1552                 }
1553                 rc = 1;
1554         } else {
1555                 __osc_unreserve_grant(cli, bytes, bytes);
1556                 rc = 0;
1557         }
1558         return rc;
1559 }
1560
1561 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1562 {
1563         int rc;
1564
1565         spin_lock(&cli->cl_loi_list_lock);
1566         rc = list_empty(&ocw->ocw_entry);
1567         spin_unlock(&cli->cl_loi_list_lock);
1568         return rc;
1569 }
1570
1571 /**
1572  * The main entry to reserve dirty page accounting. Usually the grant reserved
1573  * in this function will be freed in bulk in osc_free_grant() unless it fails
1574  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1575  *
1576  * The process will be put into sleep if it's already run out of grant.
1577  */
1578 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1579                            struct osc_async_page *oap, int bytes)
1580 {
1581         struct osc_object *osc = oap->oap_obj;
1582         struct lov_oinfo *loi = osc->oo_oinfo;
1583         struct osc_cache_waiter ocw;
1584         struct l_wait_info lwi;
1585         int rc = -EDQUOT;
1586
1587         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(AT_OFF ? obd_timeout : at_max),
1588                                NULL, LWI_ON_SIGNAL_NOOP, NULL);
1589
1590         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1591
1592         spin_lock(&cli->cl_loi_list_lock);
1593
1594         /* force the caller to try sync io.  this can jump the list
1595          * of queued writes and create a discontiguous rpc stream
1596          */
1597         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1598             !cli->cl_dirty_max_pages || cli->cl_ar.ar_force_sync ||
1599             loi->loi_ar.ar_force_sync) {
1600                 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1601                 rc = -EDQUOT;
1602                 goto out;
1603         }
1604
1605         /* Hopefully normal case - cache space and write credits available */
1606         if (osc_enter_cache_try(cli, oap, bytes, 0)) {
1607                 OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1608                 rc = 0;
1609                 goto out;
1610         }
1611
1612         /* We can get here for two reasons: too many dirty pages in cache, or
1613          * run out of grants. In both cases we should write dirty pages out.
1614          * Adding a cache waiter will trigger urgent write-out no matter what
1615          * RPC size will be.
1616          * The exiting condition is no avail grants and no dirty pages caching,
1617          * that really means there is no space on the OST.
1618          */
1619         init_waitqueue_head(&ocw.ocw_waitq);
1620         ocw.ocw_oap   = oap;
1621         ocw.ocw_grant = bytes;
1622         while (cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0) {
1623                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1624                 ocw.ocw_rc = 0;
1625                 spin_unlock(&cli->cl_loi_list_lock);
1626
1627                 osc_io_unplug_async(env, cli, NULL);
1628
1629                 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1630                        cli->cl_import->imp_obd->obd_name, &ocw, oap);
1631
1632                 rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1633
1634                 spin_lock(&cli->cl_loi_list_lock);
1635
1636                 if (rc < 0) {
1637                         /* l_wait_event is interrupted by signal, or timed out */
1638                         list_del_init(&ocw.ocw_entry);
1639                         break;
1640                 }
1641                 LASSERT(list_empty(&ocw.ocw_entry));
1642                 rc = ocw.ocw_rc;
1643
1644                 if (rc != -EDQUOT)
1645                         break;
1646                 if (osc_enter_cache_try(cli, oap, bytes, 0)) {
1647                         rc = 0;
1648                         break;
1649                 }
1650         }
1651
1652         switch (rc) {
1653         case 0:
1654                 OSC_DUMP_GRANT(D_CACHE, cli, "finally got grant space\n");
1655                 break;
1656         case -ETIMEDOUT:
1657                 OSC_DUMP_GRANT(D_CACHE, cli,
1658                                "timeout, fall back to sync i/o\n");
1659                 osc_extent_tree_dump(D_CACHE, osc);
1660                 /* fall back to synchronous I/O */
1661                 rc = -EDQUOT;
1662                 break;
1663         case -EINTR:
1664                 /* Ensures restartability - LU-3581 */
1665                 OSC_DUMP_GRANT(D_CACHE, cli, "interrupted\n");
1666                 rc = -ERESTARTSYS;
1667                 break;
1668         case -EDQUOT:
1669                 OSC_DUMP_GRANT(D_CACHE, cli,
1670                                "no grant space, fall back to sync i/o\n");
1671                 break;
1672         default:
1673                 CDEBUG(D_CACHE, "%s: event for cache space @ %p never arrived due to %d, fall back to sync i/o\n",
1674                        cli->cl_import->imp_obd->obd_name, &ocw, rc);
1675                 break;
1676         }
1677 out:
1678         spin_unlock(&cli->cl_loi_list_lock);
1679         return rc;
1680 }
1681
1682 /* caller must hold loi_list_lock */
1683 void osc_wake_cache_waiters(struct client_obd *cli)
1684 {
1685         struct list_head *l, *tmp;
1686         struct osc_cache_waiter *ocw;
1687
1688         list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1689                 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
1690                 list_del_init(&ocw->ocw_entry);
1691
1692                 ocw->ocw_rc = -EDQUOT;
1693                 /* we can't dirty more */
1694                 if ((cli->cl_dirty_pages > cli->cl_dirty_max_pages) ||
1695                     (atomic_long_read(&obd_dirty_pages) + 1 >
1696                      obd_max_dirty_pages)) {
1697                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld osc max %ld, sys max %ld\n",
1698                                cli->cl_dirty_pages, cli->cl_dirty_max_pages,
1699                                obd_max_dirty_pages);
1700                         goto wakeup;
1701                 }
1702
1703                 if (osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
1704                         ocw->ocw_rc = 0;
1705 wakeup:
1706                 CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld, %d\n",
1707                        ocw, ocw->ocw_oap, cli->cl_avail_grant, ocw->ocw_rc);
1708
1709                 wake_up(&ocw->ocw_waitq);
1710         }
1711 }
1712
1713 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1714 {
1715         int hprpc = !!list_empty(&osc->oo_hp_exts);
1716
1717         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1718 }
1719
1720 /* This maintains the lists of pending pages to read/write for a given object
1721  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1722  * to quickly find objects that are ready to send an RPC.
1723  */
1724 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1725                          int cmd)
1726 {
1727         int invalid_import = 0;
1728
1729         /* if we have an invalid import we want to drain the queued pages
1730          * by forcing them through rpcs that immediately fail and complete
1731          * the pages.  recovery relies on this to empty the queued pages
1732          * before canceling the locks and evicting down the llite pages
1733          */
1734         if (!cli->cl_import || cli->cl_import->imp_invalid)
1735                 invalid_import = 1;
1736
1737         if (cmd & OBD_BRW_WRITE) {
1738                 if (atomic_read(&osc->oo_nr_writes) == 0)
1739                         return 0;
1740                 if (invalid_import) {
1741                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1742                         return 1;
1743                 }
1744                 if (!list_empty(&osc->oo_hp_exts)) {
1745                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1746                         return 1;
1747                 }
1748                 if (!list_empty(&osc->oo_urgent_exts)) {
1749                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1750                         return 1;
1751                 }
1752                 /* trigger a write rpc stream as long as there are dirtiers
1753                  * waiting for space.  as they're waiting, they're not going to
1754                  * create more pages to coalesce with what's waiting..
1755                  */
1756                 if (!list_empty(&cli->cl_cache_waiters)) {
1757                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1758                         return 1;
1759                 }
1760                 if (atomic_read(&osc->oo_nr_writes) >=
1761                     cli->cl_max_pages_per_rpc)
1762                         return 1;
1763         } else {
1764                 if (atomic_read(&osc->oo_nr_reads) == 0)
1765                         return 0;
1766                 if (invalid_import) {
1767                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1768                         return 1;
1769                 }
1770                 /* all read are urgent. */
1771                 if (!list_empty(&osc->oo_reading_exts))
1772                         return 1;
1773         }
1774
1775         return 0;
1776 }
1777
1778 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1779 {
1780         struct client_obd *cli = osc_cli(obj);
1781
1782         if (cmd & OBD_BRW_WRITE) {
1783                 atomic_add(delta, &obj->oo_nr_writes);
1784                 atomic_add(delta, &cli->cl_pending_w_pages);
1785                 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1786         } else {
1787                 atomic_add(delta, &obj->oo_nr_reads);
1788                 atomic_add(delta, &cli->cl_pending_r_pages);
1789                 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1790         }
1791         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1792 }
1793
1794 static int osc_makes_hprpc(struct osc_object *obj)
1795 {
1796         return !list_empty(&obj->oo_hp_exts);
1797 }
1798
1799 static void on_list(struct list_head *item, struct list_head *list, int should_be_on)
1800 {
1801         if (list_empty(item) && should_be_on)
1802                 list_add_tail(item, list);
1803         else if (!list_empty(item) && !should_be_on)
1804                 list_del_init(item);
1805 }
1806
1807 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1808  * can find pages to build into rpcs quickly
1809  */
1810 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1811 {
1812         if (osc_makes_hprpc(osc)) {
1813                 /* HP rpc */
1814                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1815                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1816         } else {
1817                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1818                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1819                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1820                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1821         }
1822
1823         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1824                 atomic_read(&osc->oo_nr_writes) > 0);
1825
1826         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1827                 atomic_read(&osc->oo_nr_reads) > 0);
1828
1829         return osc_is_ready(osc);
1830 }
1831
1832 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1833 {
1834         int is_ready;
1835
1836         spin_lock(&cli->cl_loi_list_lock);
1837         is_ready = __osc_list_maint(cli, osc);
1838         spin_unlock(&cli->cl_loi_list_lock);
1839
1840         return is_ready;
1841 }
1842
1843 /* this is trying to propagate async writeback errors back up to the
1844  * application.  As an async write fails we record the error code for later if
1845  * the app does an fsync.  As long as errors persist we force future rpcs to be
1846  * sync so that the app can get a sync error and break the cycle of queueing
1847  * pages for which writeback will fail.
1848  */
1849 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1850                            int rc)
1851 {
1852         if (rc) {
1853                 if (!ar->ar_rc)
1854                         ar->ar_rc = rc;
1855
1856                 ar->ar_force_sync = 1;
1857                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1858                 return;
1859         }
1860
1861         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1862                 ar->ar_force_sync = 0;
1863 }
1864
1865 /* this must be called holding the loi list lock to give coverage to exit_cache,
1866  * async_flag maintenance, and oap_request
1867  */
1868 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1869                               struct osc_async_page *oap, int sent, int rc)
1870 {
1871         struct osc_object *osc = oap->oap_obj;
1872         struct lov_oinfo *loi = osc->oo_oinfo;
1873         __u64 xid = 0;
1874
1875         if (oap->oap_request) {
1876                 xid = ptlrpc_req_xid(oap->oap_request);
1877                 ptlrpc_req_finished(oap->oap_request);
1878                 oap->oap_request = NULL;
1879         }
1880
1881         /* As the transfer for this page is being done, clear the flags */
1882         spin_lock(&oap->oap_lock);
1883         oap->oap_async_flags = 0;
1884         spin_unlock(&oap->oap_lock);
1885         oap->oap_interrupted = 0;
1886
1887         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1888                 spin_lock(&cli->cl_loi_list_lock);
1889                 osc_process_ar(&cli->cl_ar, xid, rc);
1890                 osc_process_ar(&loi->loi_ar, xid, rc);
1891                 spin_unlock(&cli->cl_loi_list_lock);
1892         }
1893
1894         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1895         if (rc)
1896                 CERROR("completion on oap %p obj %p returns %d.\n",
1897                        oap, osc, rc);
1898 }
1899
1900 /**
1901  * Try to add extent to one RPC. We need to think about the following things:
1902  * - # of pages must not be over max_pages_per_rpc
1903  * - extent must be compatible with previous ones
1904  */
1905 static int try_to_add_extent_for_io(struct client_obd *cli,
1906                                     struct osc_extent *ext, struct list_head *rpclist,
1907                                     unsigned int *pc, unsigned int *max_pages)
1908 {
1909         struct osc_extent *tmp;
1910         struct osc_async_page *oap = list_first_entry(&ext->oe_pages,
1911                                                       struct osc_async_page,
1912                                                       oap_pending_item);
1913
1914         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1915                 ext);
1916
1917         *max_pages = max(ext->oe_mppr, *max_pages);
1918         if (*pc + ext->oe_nr_pages > *max_pages)
1919                 return 0;
1920
1921         list_for_each_entry(tmp, rpclist, oe_link) {
1922                 struct osc_async_page *oap2;
1923
1924                 oap2 = list_first_entry(&tmp->oe_pages, struct osc_async_page,
1925                                         oap_pending_item);
1926                 EASSERT(tmp->oe_owner == current, tmp);
1927                 if (oap2cl_page(oap)->cp_type != oap2cl_page(oap2)->cp_type) {
1928                         CDEBUG(D_CACHE, "Do not permit different type of IO"
1929                                         " for a same RPC\n");
1930                         return 0;
1931                 }
1932
1933                 if (tmp->oe_srvlock != ext->oe_srvlock ||
1934                     !tmp->oe_grants != !ext->oe_grants)
1935                         return 0;
1936
1937                 /* remove break for strict check */
1938                 break;
1939         }
1940
1941         *pc += ext->oe_nr_pages;
1942         list_move_tail(&ext->oe_link, rpclist);
1943         ext->oe_owner = current;
1944         return 1;
1945 }
1946
1947 /**
1948  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1949  * get_write_extent() takes all appropriate extents in atomic.
1950  *
1951  * The following policy is used to collect extents for IO:
1952  * 1. Add as many HP extents as possible;
1953  * 2. Add the first urgent extent in urgent extent list and take it out of
1954  *    urgent list;
1955  * 3. Add subsequent extents of this urgent extent;
1956  * 4. If urgent list is not empty, goto 2;
1957  * 5. Traverse the extent tree from the 1st extent;
1958  * 6. Above steps exit if there is no space in this RPC.
1959  */
1960 static unsigned int get_write_extents(struct osc_object *obj,
1961                                       struct list_head *rpclist)
1962 {
1963         struct client_obd *cli = osc_cli(obj);
1964         struct osc_extent *ext;
1965         struct osc_extent *temp;
1966         unsigned int page_count = 0;
1967         unsigned int max_pages = cli->cl_max_pages_per_rpc;
1968
1969         LASSERT(osc_object_is_locked(obj));
1970         list_for_each_entry_safe(ext, temp, &obj->oo_hp_exts, oe_link) {
1971                 LASSERT(ext->oe_state == OES_CACHE);
1972                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1973                                               &max_pages))
1974                         return page_count;
1975                 EASSERT(ext->oe_nr_pages <= max_pages, ext);
1976         }
1977         if (page_count == max_pages)
1978                 return page_count;
1979
1980         while (!list_empty(&obj->oo_urgent_exts)) {
1981                 ext = list_entry(obj->oo_urgent_exts.next,
1982                                  struct osc_extent, oe_link);
1983                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1984                                               &max_pages))
1985                         return page_count;
1986
1987                 if (!ext->oe_intree)
1988                         continue;
1989
1990                 while ((ext = next_extent(ext)) != NULL) {
1991                         if ((ext->oe_state != OES_CACHE) ||
1992                             (!list_empty(&ext->oe_link) &&
1993                              ext->oe_owner))
1994                                 continue;
1995
1996                         if (!try_to_add_extent_for_io(cli, ext, rpclist,
1997                                                       &page_count, &max_pages))
1998                                 return page_count;
1999                 }
2000         }
2001         if (page_count == max_pages)
2002                 return page_count;
2003
2004         ext = first_extent(obj);
2005         while (ext) {
2006                 if ((ext->oe_state != OES_CACHE) ||
2007                     /* this extent may be already in current rpclist */
2008                     (!list_empty(&ext->oe_link) && ext->oe_owner)) {
2009                         ext = next_extent(ext);
2010                         continue;
2011                 }
2012
2013                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
2014                                               &max_pages))
2015                         return page_count;
2016
2017                 ext = next_extent(ext);
2018         }
2019         return page_count;
2020 }
2021
2022 static int
2023 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
2024                    struct osc_object *osc)
2025         __must_hold(osc)
2026 {
2027         LIST_HEAD(rpclist);
2028         struct osc_extent *ext;
2029         struct osc_extent *tmp;
2030         struct osc_extent *first = NULL;
2031         u32 page_count = 0;
2032         int srvlock = 0;
2033         int rc = 0;
2034
2035         LASSERT(osc_object_is_locked(osc));
2036
2037         page_count = get_write_extents(osc, &rpclist);
2038         LASSERT(equi(page_count == 0, list_empty(&rpclist)));
2039
2040         if (list_empty(&rpclist))
2041                 return 0;
2042
2043         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2044
2045         list_for_each_entry(ext, &rpclist, oe_link) {
2046                 LASSERT(ext->oe_state == OES_CACHE ||
2047                         ext->oe_state == OES_LOCK_DONE);
2048                 if (ext->oe_state == OES_CACHE)
2049                         osc_extent_state_set(ext, OES_LOCKING);
2050                 else
2051                         osc_extent_state_set(ext, OES_RPC);
2052         }
2053
2054         /* we're going to grab page lock, so release object lock because
2055          * lock order is page lock -> object lock.
2056          */
2057         osc_object_unlock(osc);
2058
2059         list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2060                 if (ext->oe_state == OES_LOCKING) {
2061                         rc = osc_extent_make_ready(env, ext);
2062                         if (unlikely(rc < 0)) {
2063                                 list_del_init(&ext->oe_link);
2064                                 osc_extent_finish(env, ext, 0, rc);
2065                                 continue;
2066                         }
2067                 }
2068                 if (!first) {
2069                         first = ext;
2070                         srvlock = ext->oe_srvlock;
2071                 } else {
2072                         LASSERT(srvlock == ext->oe_srvlock);
2073                 }
2074         }
2075
2076         if (!list_empty(&rpclist)) {
2077                 LASSERT(page_count > 0);
2078                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2079                 LASSERT(list_empty(&rpclist));
2080         }
2081
2082         osc_object_lock(osc);
2083         return rc;
2084 }
2085
2086 /**
2087  * prepare pages for ASYNC io and put pages in send queue.
2088  *
2089  * \param cmd OBD_BRW_* macroses
2090  * \param lop pending pages
2091  *
2092  * \return zero if no page added to send queue.
2093  * \return 1 if pages successfully added to send queue.
2094  * \return negative on errors.
2095  */
2096 static int
2097 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2098                   struct osc_object *osc)
2099         __must_hold(osc)
2100 {
2101         struct osc_extent *ext;
2102         struct osc_extent *next;
2103         LIST_HEAD(rpclist);
2104         unsigned int page_count = 0;
2105         unsigned int max_pages = cli->cl_max_pages_per_rpc;
2106         int rc = 0;
2107
2108         LASSERT(osc_object_is_locked(osc));
2109         list_for_each_entry_safe(ext, next, &osc->oo_reading_exts, oe_link) {
2110                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2111                 if (!try_to_add_extent_for_io(cli, ext, &rpclist, &page_count,
2112                                               &max_pages))
2113                         break;
2114                 osc_extent_state_set(ext, OES_RPC);
2115                 EASSERT(ext->oe_nr_pages <= max_pages, ext);
2116         }
2117         LASSERT(page_count <= max_pages);
2118
2119         osc_update_pending(osc, OBD_BRW_READ, -page_count);
2120
2121         if (!list_empty(&rpclist)) {
2122                 osc_object_unlock(osc);
2123
2124                 LASSERT(page_count > 0);
2125                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2126                 LASSERT(list_empty(&rpclist));
2127
2128                 osc_object_lock(osc);
2129         }
2130         return rc;
2131 }
2132
2133 #define list_to_obj(list, item) ({                                            \
2134         struct list_head *__tmp = (list)->next;                               \
2135         list_del_init(__tmp);                                                 \
2136         list_entry(__tmp, struct osc_object, oo_##item);                      \
2137 })
2138
2139 /* This is called by osc_check_rpcs() to find which objects have pages that
2140  * we could be sending.  These lists are maintained by osc_makes_rpc().
2141  */
2142 static struct osc_object *osc_next_obj(struct client_obd *cli)
2143 {
2144         /* First return objects that have blocked locks so that they
2145          * will be flushed quickly and other clients can get the lock,
2146          * then objects which have pages ready to be stuffed into RPCs
2147          */
2148         if (!list_empty(&cli->cl_loi_hp_ready_list))
2149                 return list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item);
2150         if (!list_empty(&cli->cl_loi_ready_list))
2151                 return list_to_obj(&cli->cl_loi_ready_list, ready_item);
2152
2153         /* then if we have cache waiters, return all objects with queued
2154          * writes.  This is especially important when many small files
2155          * have filled up the cache and not been fired into rpcs because
2156          * they don't pass the nr_pending/object threshold
2157          */
2158         if (!list_empty(&cli->cl_cache_waiters) &&
2159             !list_empty(&cli->cl_loi_write_list))
2160                 return list_to_obj(&cli->cl_loi_write_list, write_item);
2161
2162         /* then return all queued objects when we have an invalid import
2163          * so that they get flushed
2164          */
2165         if (!cli->cl_import || cli->cl_import->imp_invalid) {
2166                 if (!list_empty(&cli->cl_loi_write_list))
2167                         return list_to_obj(&cli->cl_loi_write_list, write_item);
2168                 if (!list_empty(&cli->cl_loi_read_list))
2169                         return list_to_obj(&cli->cl_loi_read_list, read_item);
2170         }
2171         return NULL;
2172 }
2173
2174 /* called with the loi list lock held */
2175 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2176         __must_hold(&cli->cl_loi_list_lock)
2177 {
2178         struct osc_object *osc;
2179         int rc = 0;
2180
2181         while ((osc = osc_next_obj(cli)) != NULL) {
2182                 struct cl_object *obj = osc2cl(osc);
2183                 struct lu_ref_link link;
2184
2185                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2186
2187                 if (osc_max_rpc_in_flight(cli, osc)) {
2188                         __osc_list_maint(cli, osc);
2189                         break;
2190                 }
2191
2192                 cl_object_get(obj);
2193                 spin_unlock(&cli->cl_loi_list_lock);
2194                 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2195
2196                 /* attempt some read/write balancing by alternating between
2197                  * reads and writes in an object.  The makes_rpc checks here
2198                  * would be redundant if we were getting read/write work items
2199                  * instead of objects.  we don't want send_oap_rpc to drain a
2200                  * partial read pending queue when we're given this object to
2201                  * do io on writes while there are cache waiters
2202                  */
2203                 osc_object_lock(osc);
2204                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2205                         rc = osc_send_write_rpc(env, cli, osc);
2206                         if (rc < 0) {
2207                                 CERROR("Write request failed with %d\n", rc);
2208
2209                                 /* osc_send_write_rpc failed, mostly because of
2210                                  * memory pressure.
2211                                  *
2212                                  * It can't break here, because if:
2213                                  *  - a page was submitted by osc_io_submit, so
2214                                  *    page locked;
2215                                  *  - no request in flight
2216                                  *  - no subsequent request
2217                                  * The system will be in live-lock state,
2218                                  * because there is no chance to call
2219                                  * osc_io_unplug() and osc_check_rpcs() any
2220                                  * more. pdflush can't help in this case,
2221                                  * because it might be blocked at grabbing
2222                                  * the page lock as we mentioned.
2223                                  *
2224                                  * Anyway, continue to drain pages.
2225                                  */
2226                                 /* break; */
2227                         }
2228                 }
2229                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2230                         rc = osc_send_read_rpc(env, cli, osc);
2231                         if (rc < 0)
2232                                 CERROR("Read request failed with %d\n", rc);
2233                 }
2234                 osc_object_unlock(osc);
2235
2236                 osc_list_maint(cli, osc);
2237                 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2238                 cl_object_put(env, obj);
2239
2240                 spin_lock(&cli->cl_loi_list_lock);
2241         }
2242 }
2243
2244 static int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2245                           struct osc_object *osc, int async)
2246 {
2247         int rc = 0;
2248
2249         if (osc && osc_list_maint(cli, osc) == 0)
2250                 return 0;
2251
2252         if (!async) {
2253                 /* disable osc_lru_shrink() temporarily to avoid
2254                  * potential stack overrun problem. LU-2859
2255                  */
2256                 atomic_inc(&cli->cl_lru_shrinkers);
2257                 spin_lock(&cli->cl_loi_list_lock);
2258                 osc_check_rpcs(env, cli);
2259                 spin_unlock(&cli->cl_loi_list_lock);
2260                 atomic_dec(&cli->cl_lru_shrinkers);
2261         } else {
2262                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2263                 LASSERT(cli->cl_writeback_work);
2264                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2265         }
2266         return rc;
2267 }
2268
2269 static int osc_io_unplug_async(const struct lu_env *env,
2270                                struct client_obd *cli, struct osc_object *osc)
2271 {
2272         return osc_io_unplug0(env, cli, osc, 1);
2273 }
2274
2275 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
2276                    struct osc_object *osc)
2277 {
2278         (void)osc_io_unplug0(env, cli, osc, 0);
2279 }
2280
2281 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2282                         struct page *page, loff_t offset)
2283 {
2284         struct obd_export *exp = osc_export(osc);
2285         struct osc_async_page *oap = &ops->ops_oap;
2286
2287         if (!page)
2288                 return cfs_size_round(sizeof(*oap));
2289
2290         oap->oap_magic = OAP_MAGIC;
2291         oap->oap_cli = &exp->exp_obd->u.cli;
2292         oap->oap_obj = osc;
2293
2294         oap->oap_page = page;
2295         oap->oap_obj_off = offset;
2296         LASSERT(!(offset & ~PAGE_MASK));
2297
2298         if (capable(CFS_CAP_SYS_RESOURCE))
2299                 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2300
2301         INIT_LIST_HEAD(&oap->oap_pending_item);
2302         INIT_LIST_HEAD(&oap->oap_rpc_item);
2303
2304         spin_lock_init(&oap->oap_lock);
2305         CDEBUG(D_INFO, "oap %p page %p obj off %llu\n",
2306                oap, page, oap->oap_obj_off);
2307         return 0;
2308 }
2309
2310 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2311                        struct osc_page *ops)
2312 {
2313         struct osc_io *oio = osc_env_io(env);
2314         struct osc_extent *ext = NULL;
2315         struct osc_async_page *oap = &ops->ops_oap;
2316         struct client_obd *cli = oap->oap_cli;
2317         struct osc_object *osc = oap->oap_obj;
2318         pgoff_t index;
2319         unsigned int grants = 0, tmp;
2320         int brw_flags = OBD_BRW_ASYNC;
2321         int cmd = OBD_BRW_WRITE;
2322         int need_release = 0;
2323         int rc = 0;
2324
2325         if (oap->oap_magic != OAP_MAGIC)
2326                 return -EINVAL;
2327
2328         if (!cli->cl_import || cli->cl_import->imp_invalid)
2329                 return -EIO;
2330
2331         if (!list_empty(&oap->oap_pending_item) ||
2332             !list_empty(&oap->oap_rpc_item))
2333                 return -EBUSY;
2334
2335         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2336         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2337         if (capable(CFS_CAP_SYS_RESOURCE)) {
2338                 brw_flags |= OBD_BRW_NOQUOTA;
2339                 cmd |= OBD_BRW_NOQUOTA;
2340         }
2341
2342         /* check if the file's owner/group is over quota */
2343         if (!(cmd & OBD_BRW_NOQUOTA)) {
2344                 struct cl_object *obj;
2345                 struct cl_attr *attr;
2346                 unsigned int qid[MAXQUOTAS];
2347
2348                 obj = cl_object_top(&osc->oo_cl);
2349                 attr = &osc_env_info(env)->oti_attr;
2350
2351                 cl_object_attr_lock(obj);
2352                 rc = cl_object_attr_get(env, obj, attr);
2353                 cl_object_attr_unlock(obj);
2354
2355                 qid[USRQUOTA] = attr->cat_uid;
2356                 qid[GRPQUOTA] = attr->cat_gid;
2357                 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2358                         rc = -EDQUOT;
2359                 if (rc)
2360                         return rc;
2361         }
2362
2363         oap->oap_cmd = cmd;
2364         oap->oap_page_off = ops->ops_from;
2365         oap->oap_count = ops->ops_to - ops->ops_from;
2366         /*
2367          * No need to hold a lock here,
2368          * since this page is not in any list yet.
2369          */
2370         oap->oap_async_flags = 0;
2371         oap->oap_brw_flags = brw_flags;
2372
2373         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2374                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2375
2376         index = osc_index(oap2osc(oap));
2377
2378         /* Add this page into extent by the following steps:
2379          * 1. if there exists an active extent for this IO, mostly this page
2380          *    can be added to the active extent and sometimes we need to
2381          *    expand extent to accommodate this page;
2382          * 2. otherwise, a new extent will be allocated.
2383          */
2384
2385         ext = oio->oi_active;
2386         if (ext && ext->oe_start <= index && ext->oe_max_end >= index) {
2387                 /* one chunk plus extent overhead must be enough to write this
2388                  * page
2389                  */
2390                 grants = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2391                 if (ext->oe_end >= index)
2392                         grants = 0;
2393
2394                 /* it doesn't need any grant to dirty this page */
2395                 spin_lock(&cli->cl_loi_list_lock);
2396                 rc = osc_enter_cache_try(cli, oap, grants, 0);
2397                 spin_unlock(&cli->cl_loi_list_lock);
2398                 if (rc == 0) { /* try failed */
2399                         grants = 0;
2400                         need_release = 1;
2401                 } else if (ext->oe_end < index) {
2402                         tmp = grants;
2403                         /* try to expand this extent */
2404                         rc = osc_extent_expand(ext, index, &tmp);
2405                         if (rc < 0) {
2406                                 need_release = 1;
2407                                 /* don't free reserved grant */
2408                         } else {
2409                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2410                                                 "expanded for %lu.\n", index);
2411                                 osc_unreserve_grant(cli, grants, tmp);
2412                                 grants = 0;
2413                         }
2414                 }
2415                 rc = 0;
2416         } else if (ext) {
2417                 /* index is located outside of active extent */
2418                 need_release = 1;
2419         }
2420         if (need_release) {
2421                 osc_extent_release(env, ext);
2422                 oio->oi_active = NULL;
2423                 ext = NULL;
2424         }
2425
2426         if (!ext) {
2427                 tmp = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2428
2429                 /* try to find new extent to cover this page */
2430                 LASSERT(!oio->oi_active);
2431                 /* we may have allocated grant for this page if we failed
2432                  * to expand the previous active extent.
2433                  */
2434                 LASSERT(ergo(grants > 0, grants >= tmp));
2435
2436                 rc = 0;
2437                 if (grants == 0) {
2438                         /* we haven't allocated grant for this page. */
2439                         rc = osc_enter_cache(env, cli, oap, tmp);
2440                         if (rc == 0)
2441                                 grants = tmp;
2442                 }
2443
2444                 tmp = grants;
2445                 if (rc == 0) {
2446                         ext = osc_extent_find(env, osc, index, &tmp);
2447                         if (IS_ERR(ext)) {
2448                                 LASSERT(tmp == grants);
2449                                 osc_exit_cache(cli, oap);
2450                                 rc = PTR_ERR(ext);
2451                                 ext = NULL;
2452                         } else {
2453                                 oio->oi_active = ext;
2454                         }
2455                 }
2456                 if (grants > 0)
2457                         osc_unreserve_grant(cli, grants, tmp);
2458         }
2459
2460         LASSERT(ergo(rc == 0, ext));
2461         if (ext) {
2462                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2463                          ext, "index = %lu.\n", index);
2464                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2465
2466                 osc_object_lock(osc);
2467                 if (ext->oe_nr_pages == 0)
2468                         ext->oe_srvlock = ops->ops_srvlock;
2469                 else
2470                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2471                 ++ext->oe_nr_pages;
2472                 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2473                 osc_object_unlock(osc);
2474         }
2475         return rc;
2476 }
2477
2478 int osc_teardown_async_page(const struct lu_env *env,
2479                             struct osc_object *obj, struct osc_page *ops)
2480 {
2481         struct osc_async_page *oap = &ops->ops_oap;
2482         struct osc_extent *ext = NULL;
2483         int rc = 0;
2484
2485         LASSERT(oap->oap_magic == OAP_MAGIC);
2486
2487         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2488                oap, ops, osc_index(oap2osc(oap)));
2489
2490         osc_object_lock(obj);
2491         if (!list_empty(&oap->oap_rpc_item)) {
2492                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2493                 rc = -EBUSY;
2494         } else if (!list_empty(&oap->oap_pending_item)) {
2495                 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2496                 /* only truncated pages are allowed to be taken out.
2497                  * See osc_extent_truncate() and osc_cache_truncate_start()
2498                  * for details.
2499                  */
2500                 if (ext && ext->oe_state != OES_TRUNC) {
2501                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2502                                         osc_index(oap2osc(oap)));
2503                         rc = -EBUSY;
2504                 }
2505         }
2506         osc_object_unlock(obj);
2507         if (ext)
2508                 osc_extent_put(env, ext);
2509         return rc;
2510 }
2511
2512 /**
2513  * This is called when a page is picked up by kernel to write out.
2514  *
2515  * We should find out the corresponding extent and add the whole extent
2516  * into urgent list. The extent may be being truncated or used, handle it
2517  * carefully.
2518  */
2519 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2520                          struct osc_page *ops)
2521 {
2522         struct osc_extent *ext = NULL;
2523         struct osc_object *obj = cl2osc(ops->ops_cl.cpl_obj);
2524         struct cl_page *cp = ops->ops_cl.cpl_page;
2525         pgoff_t            index = osc_index(ops);
2526         struct osc_async_page *oap = &ops->ops_oap;
2527         bool unplug = false;
2528         int rc = 0;
2529
2530         osc_object_lock(obj);
2531         ext = osc_extent_lookup(obj, index);
2532         if (!ext) {
2533                 osc_extent_tree_dump(D_ERROR, obj);
2534                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2535         }
2536
2537         switch (ext->oe_state) {
2538         case OES_RPC:
2539         case OES_LOCK_DONE:
2540                 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2541                 LASSERT(0);
2542                 break;
2543         case OES_LOCKING:
2544                 /* If we know this extent is being written out, we should abort
2545                  * so that the writer can make this page ready. Otherwise, there
2546                  * exists a deadlock problem because other process can wait for
2547                  * page writeback bit holding page lock; and meanwhile in
2548                  * vvp_page_make_ready(), we need to grab page lock before
2549                  * really sending the RPC.
2550                  */
2551         case OES_TRUNC:
2552                 /* race with truncate, page will be redirtied */
2553         case OES_ACTIVE:
2554                 /* The extent is active so we need to abort and let the caller
2555                  * re-dirty the page. If we continued on here, and we were the
2556                  * one making the extent active, we could deadlock waiting for
2557                  * the page writeback to clear but it won't because the extent
2558                  * is active and won't be written out.
2559                  */
2560                 rc = -EAGAIN;
2561                 goto out;
2562         default:
2563                 break;
2564         }
2565
2566         rc = cl_page_prep(env, io, cp, CRT_WRITE);
2567         if (rc)
2568                 goto out;
2569
2570         spin_lock(&oap->oap_lock);
2571         oap->oap_async_flags |= ASYNC_READY | ASYNC_URGENT;
2572         spin_unlock(&oap->oap_lock);
2573
2574         if (memory_pressure_get())
2575                 ext->oe_memalloc = 1;
2576
2577         ext->oe_urgent = 1;
2578         if (ext->oe_state == OES_CACHE) {
2579                 OSC_EXTENT_DUMP(D_CACHE, ext,
2580                                 "flush page %p make it urgent.\n", oap);
2581                 if (list_empty(&ext->oe_link))
2582                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2583                 unplug = true;
2584         }
2585         rc = 0;
2586
2587 out:
2588         osc_object_unlock(obj);
2589         osc_extent_put(env, ext);
2590         if (unplug)
2591                 osc_io_unplug_async(env, osc_cli(obj), obj);
2592         return rc;
2593 }
2594
2595 /**
2596  * this is called when a sync waiter receives an interruption.  Its job is to
2597  * get the caller woken as soon as possible.  If its page hasn't been put in an
2598  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2599  * desiring interruption which will forcefully complete the rpc once the rpc
2600  * has timed out.
2601  */
2602 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2603 {
2604         struct osc_async_page *oap = &ops->ops_oap;
2605         struct osc_object *obj = oap->oap_obj;
2606         struct client_obd *cli = osc_cli(obj);
2607         struct osc_extent *ext;
2608         struct osc_extent *found = NULL;
2609         struct list_head *plist;
2610         pgoff_t index = osc_index(ops);
2611         int rc = -EBUSY;
2612         int cmd;
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         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                          */
2632                         found = osc_extent_get(ext);
2633                         break;
2634                 }
2635         }
2636         if (found) {
2637                 list_del_init(&found->oe_link);
2638                 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2639                 osc_object_unlock(obj);
2640
2641                 osc_extent_finish(env, found, 0, -EINTR);
2642                 osc_extent_put(env, found);
2643                 rc = 0;
2644         } else {
2645                 osc_object_unlock(obj);
2646                 /* ok, it's been put in an rpc. only one oap gets a request
2647                  * reference
2648                  */
2649                 if (oap->oap_request) {
2650                         ptlrpc_mark_interrupted(oap->oap_request);
2651                         ptlrpcd_wake(oap->oap_request);
2652                         ptlrpc_req_finished(oap->oap_request);
2653                         oap->oap_request = NULL;
2654                 }
2655         }
2656
2657         osc_list_maint(cli, obj);
2658         return rc;
2659 }
2660
2661 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
2662                          struct list_head *list, int cmd, int brw_flags)
2663 {
2664         struct client_obd *cli = osc_cli(obj);
2665         struct osc_extent *ext;
2666         struct osc_async_page *oap, *tmp;
2667         int page_count = 0;
2668         int mppr = cli->cl_max_pages_per_rpc;
2669         pgoff_t start = CL_PAGE_EOF;
2670         pgoff_t end = 0;
2671
2672         list_for_each_entry(oap, list, oap_pending_item) {
2673                 pgoff_t index = osc_index(oap2osc(oap));
2674
2675                 if (index > end)
2676                         end = index;
2677                 if (index < start)
2678                         start = index;
2679                 ++page_count;
2680                 mppr <<= (page_count > mppr);
2681         }
2682
2683         ext = osc_extent_alloc(obj);
2684         if (!ext) {
2685                 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2686                         list_del_init(&oap->oap_pending_item);
2687                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2688                 }
2689                 return -ENOMEM;
2690         }
2691
2692         ext->oe_rw = !!(cmd & OBD_BRW_READ);
2693         ext->oe_sync = 1;
2694         ext->oe_urgent = 1;
2695         ext->oe_start = start;
2696         ext->oe_end = end;
2697         ext->oe_max_end = end;
2698         ext->oe_obj = obj;
2699         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2700         ext->oe_nr_pages = page_count;
2701         ext->oe_mppr = mppr;
2702         list_splice_init(list, &ext->oe_pages);
2703
2704         osc_object_lock(obj);
2705         /* Reuse the initial refcount for RPC, don't drop it */
2706         osc_extent_state_set(ext, OES_LOCK_DONE);
2707         if (cmd & OBD_BRW_WRITE) {
2708                 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2709                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2710         } else {
2711                 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2712                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2713         }
2714         osc_object_unlock(obj);
2715
2716         osc_io_unplug_async(env, cli, obj);
2717         return 0;
2718 }
2719
2720 /**
2721  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2722  */
2723 int osc_cache_truncate_start(const struct lu_env *env, struct osc_io *oio,
2724                              struct osc_object *obj, __u64 size)
2725 {
2726         struct client_obd *cli = osc_cli(obj);
2727         struct osc_extent *ext;
2728         struct osc_extent *temp;
2729         struct osc_extent *waiting = NULL;
2730         pgoff_t index;
2731         LIST_HEAD(list);
2732         int result = 0;
2733         bool partial;
2734
2735         /* pages with index greater or equal to index will be truncated. */
2736         index = cl_index(osc2cl(obj), size);
2737         partial = size > cl_offset(osc2cl(obj), index);
2738
2739 again:
2740         osc_object_lock(obj);
2741         ext = osc_extent_search(obj, index);
2742         if (!ext)
2743                 ext = first_extent(obj);
2744         else if (ext->oe_end < index)
2745                 ext = next_extent(ext);
2746         while (ext) {
2747                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2748
2749                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2750                         /* if ext is in urgent state, it means there must exist
2751                          * a page already having been flushed by write_page().
2752                          * We have to wait for this extent because we can't
2753                          * truncate that page.
2754                          */
2755                         LASSERT(!ext->oe_hp);
2756                         OSC_EXTENT_DUMP(D_CACHE, ext,
2757                                         "waiting for busy extent\n");
2758                         waiting = osc_extent_get(ext);
2759                         break;
2760                 }
2761
2762                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:%llu.\n", size);
2763
2764                 osc_extent_get(ext);
2765                 if (ext->oe_state == OES_ACTIVE) {
2766                         /* though we grab inode mutex for write path, but we
2767                          * release it before releasing extent(in osc_io_end()),
2768                          * so there is a race window that an extent is still
2769                          * in OES_ACTIVE when truncate starts.
2770                          */
2771                         LASSERT(!ext->oe_trunc_pending);
2772                         ext->oe_trunc_pending = 1;
2773                 } else {
2774                         EASSERT(ext->oe_state == OES_CACHE, ext);
2775                         osc_extent_state_set(ext, OES_TRUNC);
2776                         osc_update_pending(obj, OBD_BRW_WRITE,
2777                                            -ext->oe_nr_pages);
2778                 }
2779                 EASSERT(list_empty(&ext->oe_link), ext);
2780                 list_add_tail(&ext->oe_link, &list);
2781
2782                 ext = next_extent(ext);
2783         }
2784         osc_object_unlock(obj);
2785
2786         osc_list_maint(cli, obj);
2787
2788         list_for_each_entry_safe(ext, temp, &list, oe_link) {
2789                 int rc;
2790
2791                 list_del_init(&ext->oe_link);
2792
2793                 /* extent may be in OES_ACTIVE state because inode mutex
2794                  * is released before osc_io_end() in file write case
2795                  */
2796                 if (ext->oe_state != OES_TRUNC)
2797                         osc_extent_wait(env, ext, OES_TRUNC);
2798
2799                 rc = osc_extent_truncate(ext, index, partial);
2800                 if (rc < 0) {
2801                         if (result == 0)
2802                                 result = rc;
2803
2804                         OSC_EXTENT_DUMP(D_ERROR, ext,
2805                                         "truncate error %d\n", rc);
2806                 } else if (ext->oe_nr_pages == 0) {
2807                         osc_extent_remove(ext);
2808                 } else {
2809                         /* this must be an overlapped extent which means only
2810                          * part of pages in this extent have been truncated.
2811                          */
2812                         EASSERTF(ext->oe_start <= index, ext,
2813                                  "trunc index = %lu/%d.\n", index, partial);
2814                         /* fix index to skip this partially truncated extent */
2815                         index = ext->oe_end + 1;
2816                         partial = false;
2817
2818                         /* we need to hold this extent in OES_TRUNC state so
2819                          * that no writeback will happen. This is to avoid
2820                          * BUG 17397.
2821                          */
2822                         LASSERT(!oio->oi_trunc);
2823                         oio->oi_trunc = osc_extent_get(ext);
2824                         OSC_EXTENT_DUMP(D_CACHE, ext,
2825                                         "trunc at %llu\n", size);
2826                 }
2827                 osc_extent_put(env, ext);
2828         }
2829         if (waiting) {
2830                 int rc;
2831
2832                 /* ignore the result of osc_extent_wait the write initiator
2833                  * should take care of it.
2834                  */
2835                 rc = osc_extent_wait(env, waiting, OES_INV);
2836                 if (rc < 0)
2837                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2838
2839                 osc_extent_put(env, waiting);
2840                 waiting = NULL;
2841                 goto again;
2842         }
2843         return result;
2844 }
2845
2846 /**
2847  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2848  */
2849 void osc_cache_truncate_end(const struct lu_env *env, struct osc_io *oio,
2850                             struct osc_object *obj)
2851 {
2852         struct osc_extent *ext = oio->oi_trunc;
2853
2854         oio->oi_trunc = NULL;
2855         if (ext) {
2856                 bool unplug = false;
2857
2858                 EASSERT(ext->oe_nr_pages > 0, ext);
2859                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2860                 EASSERT(!ext->oe_urgent, ext);
2861
2862                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2863                 osc_object_lock(obj);
2864                 osc_extent_state_set(ext, OES_CACHE);
2865                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2866                         ext->oe_urgent = 1;
2867                         list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2868                         unplug = true;
2869                 }
2870                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2871                 osc_object_unlock(obj);
2872                 osc_extent_put(env, ext);
2873
2874                 if (unplug)
2875                         osc_io_unplug_async(env, osc_cli(obj), obj);
2876         }
2877 }
2878
2879 /**
2880  * Wait for extents in a specific range to be written out.
2881  * The caller must have called osc_cache_writeback_range() to issue IO
2882  * otherwise it will take a long time for this function to finish.
2883  *
2884  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2885  * nobody else can dirty this range of file while we're waiting for
2886  * extents to be written.
2887  */
2888 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2889                          pgoff_t start, pgoff_t end)
2890 {
2891         struct osc_extent *ext;
2892         pgoff_t index = start;
2893         int result = 0;
2894
2895 again:
2896         osc_object_lock(obj);
2897         ext = osc_extent_search(obj, index);
2898         if (!ext)
2899                 ext = first_extent(obj);
2900         else if (ext->oe_end < index)
2901                 ext = next_extent(ext);
2902         while (ext) {
2903                 int rc;
2904
2905                 if (ext->oe_start > end)
2906                         break;
2907
2908                 if (!ext->oe_fsync_wait) {
2909                         ext = next_extent(ext);
2910                         continue;
2911                 }
2912
2913                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2914                              ext->oe_hp || ext->oe_urgent), ext);
2915                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2916                              !ext->oe_hp && ext->oe_urgent), ext);
2917
2918                 index = ext->oe_end + 1;
2919                 osc_extent_get(ext);
2920                 osc_object_unlock(obj);
2921
2922                 rc = osc_extent_wait(env, ext, OES_INV);
2923                 if (result == 0)
2924                         result = rc;
2925                 osc_extent_put(env, ext);
2926                 goto again;
2927         }
2928         osc_object_unlock(obj);
2929
2930         OSC_IO_DEBUG(obj, "sync file range.\n");
2931         return result;
2932 }
2933
2934 /**
2935  * Called to write out a range of osc object.
2936  *
2937  * @hp     : should be set this is caused by lock cancel;
2938  * @discard: is set if dirty pages should be dropped - file will be deleted or
2939  *         truncated, this implies there is no partially discarding extents.
2940  *
2941  * Return how many pages will be issued, or error code if error occurred.
2942  */
2943 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2944                               pgoff_t start, pgoff_t end, int hp, int discard)
2945 {
2946         struct osc_extent *ext;
2947         LIST_HEAD(discard_list);
2948         bool unplug = false;
2949         int result = 0;
2950
2951         osc_object_lock(obj);
2952         ext = osc_extent_search(obj, start);
2953         if (!ext)
2954                 ext = first_extent(obj);
2955         else if (ext->oe_end < start)
2956                 ext = next_extent(ext);
2957         while (ext) {
2958                 if (ext->oe_start > end)
2959                         break;
2960
2961                 ext->oe_fsync_wait = 1;
2962                 switch (ext->oe_state) {
2963                 case OES_CACHE:
2964                         result += ext->oe_nr_pages;
2965                         if (!discard) {
2966                                 struct list_head *list = NULL;
2967
2968                                 if (hp) {
2969                                         EASSERT(!ext->oe_hp, ext);
2970                                         ext->oe_hp = 1;
2971                                         list = &obj->oo_hp_exts;
2972                                 } else if (!ext->oe_urgent) {
2973                                         ext->oe_urgent = 1;
2974                                         list = &obj->oo_urgent_exts;
2975                                 }
2976                                 if (list)
2977                                         list_move_tail(&ext->oe_link, list);
2978                                 unplug = true;
2979                         } else {
2980                                 /* the only discarder is lock cancelling, so
2981                                  * [start, end] must contain this extent
2982                                  */
2983                                 EASSERT(ext->oe_start >= start &&
2984                                         ext->oe_max_end <= end, ext);
2985                                 osc_extent_state_set(ext, OES_LOCKING);
2986                                 ext->oe_owner = current;
2987                                 list_move_tail(&ext->oe_link, &discard_list);
2988                                 osc_update_pending(obj, OBD_BRW_WRITE,
2989                                                    -ext->oe_nr_pages);
2990                         }
2991                         break;
2992                 case OES_ACTIVE:
2993                         /* It's pretty bad to wait for ACTIVE extents, because
2994                          * we don't know how long we will wait for it to be
2995                          * flushed since it may be blocked at awaiting more
2996                          * grants. We do this for the correctness of fsync.
2997                          */
2998                         LASSERT(hp == 0 && discard == 0);
2999                         ext->oe_urgent = 1;
3000                         break;
3001                 case OES_TRUNC:
3002                         /* this extent is being truncated, can't do anything
3003                          * for it now. it will be set to urgent after truncate
3004                          * is finished in osc_cache_truncate_end().
3005                          */
3006                 default:
3007                         break;
3008                 }
3009                 ext = next_extent(ext);
3010         }
3011         osc_object_unlock(obj);
3012
3013         LASSERT(ergo(!discard, list_empty(&discard_list)));
3014         if (!list_empty(&discard_list)) {
3015                 struct osc_extent *tmp;
3016                 int rc;
3017
3018                 osc_list_maint(osc_cli(obj), obj);
3019                 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
3020                         list_del_init(&ext->oe_link);
3021                         EASSERT(ext->oe_state == OES_LOCKING, ext);
3022
3023                         /* Discard caching pages. We don't actually write this
3024                          * extent out but we complete it as if we did.
3025                          */
3026                         rc = osc_extent_make_ready(env, ext);
3027                         if (unlikely(rc < 0)) {
3028                                 OSC_EXTENT_DUMP(D_ERROR, ext,
3029                                                 "make_ready returned %d\n", rc);
3030                                 if (result >= 0)
3031                                         result = rc;
3032                         }
3033
3034                         /* finish the extent as if the pages were sent */
3035                         osc_extent_finish(env, ext, 0, 0);
3036                 }
3037         }
3038
3039         if (unplug)
3040                 osc_io_unplug(env, osc_cli(obj), obj);
3041
3042         if (hp || discard) {
3043                 int rc;
3044
3045                 rc = osc_cache_wait_range(env, obj, start, end);
3046                 if (result >= 0 && rc < 0)
3047                         result = rc;
3048         }
3049
3050         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3051         return result;
3052 }
3053
3054 /**
3055  * Returns a list of pages by a given [start, end] of \a obj.
3056  *
3057  * \param resched If not NULL, then we give up before hogging CPU for too
3058  * long and set *resched = 1, in that case caller should implement a retry
3059  * logic.
3060  *
3061  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3062  * crucial in the face of [offset, EOF] locks.
3063  *
3064  * Return at least one page in @queue unless there is no covered page.
3065  */
3066 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3067                          struct osc_object *osc, pgoff_t start, pgoff_t end,
3068                          osc_page_gang_cbt cb, void *cbdata)
3069 {
3070         struct osc_page *ops;
3071         void            **pvec;
3072         pgoff_t         idx;
3073         unsigned int    nr;
3074         unsigned int    i;
3075         unsigned int    j;
3076         int             res = CLP_GANG_OKAY;
3077         bool            tree_lock = true;
3078
3079         idx = start;
3080         pvec = osc_env_info(env)->oti_pvec;
3081         spin_lock(&osc->oo_tree_lock);
3082         while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3083                                             idx, OTI_PVEC_SIZE)) > 0) {
3084                 struct cl_page *page;
3085                 bool end_of_region = false;
3086
3087                 for (i = 0, j = 0; i < nr; ++i) {
3088                         ops = pvec[i];
3089                         pvec[i] = NULL;
3090
3091                         idx = osc_index(ops);
3092                         if (idx > end) {
3093                                 end_of_region = true;
3094                                 break;
3095                         }
3096
3097                         page = ops->ops_cl.cpl_page;
3098                         LASSERT(page->cp_type == CPT_CACHEABLE);
3099                         if (page->cp_state == CPS_FREEING)
3100                                 continue;
3101
3102                         cl_page_get(page);
3103                         lu_ref_add_atomic(&page->cp_reference,
3104                                           "gang_lookup", current);
3105                         pvec[j++] = ops;
3106                 }
3107                 ++idx;
3108
3109                 /*
3110                  * Here a delicate locking dance is performed. Current thread
3111                  * holds a reference to a page, but has to own it before it
3112                  * can be placed into queue. Owning implies waiting, so
3113                  * radix-tree lock is to be released. After a wait one has to
3114                  * check that pages weren't truncated (cl_page_own() returns
3115                  * error in the latter case).
3116                  */
3117                 spin_unlock(&osc->oo_tree_lock);
3118                 tree_lock = false;
3119
3120                 for (i = 0; i < j; ++i) {
3121                         ops = pvec[i];
3122                         if (res == CLP_GANG_OKAY)
3123                                 res = (*cb)(env, io, ops, cbdata);
3124
3125                         page = ops->ops_cl.cpl_page;
3126                         lu_ref_del(&page->cp_reference, "gang_lookup", current);
3127                         cl_page_put(env, page);
3128                 }
3129                 if (nr < OTI_PVEC_SIZE || end_of_region)
3130                         break;
3131
3132                 if (res == CLP_GANG_OKAY && need_resched())
3133                         res = CLP_GANG_RESCHED;
3134                 if (res != CLP_GANG_OKAY)
3135                         break;
3136
3137                 spin_lock(&osc->oo_tree_lock);
3138                 tree_lock = true;
3139         }
3140         if (tree_lock)
3141                 spin_unlock(&osc->oo_tree_lock);
3142         return res;
3143 }
3144
3145 /**
3146  * Check if page @page is covered by an extra lock or discard it.
3147  */
3148 static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3149                                 struct osc_page *ops, void *cbdata)
3150 {
3151         struct osc_thread_info *info = osc_env_info(env);
3152         struct osc_object *osc = cbdata;
3153         pgoff_t index;
3154
3155         index = osc_index(ops);
3156         if (index >= info->oti_fn_index) {
3157                 struct ldlm_lock *tmp;
3158                 struct cl_page *page = ops->ops_cl.cpl_page;
3159
3160                 /* refresh non-overlapped index */
3161                 tmp = osc_dlmlock_at_pgoff(env, osc, index, 0, 0);
3162                 if (tmp) {
3163                         __u64 end = tmp->l_policy_data.l_extent.end;
3164                         /* Cache the first-non-overlapped index so as to skip
3165                          * all pages within [index, oti_fn_index). This is safe
3166                          * because if tmp lock is canceled, it will discard
3167                          * these pages.
3168                          */
3169                         info->oti_fn_index = cl_index(osc2cl(osc), end + 1);
3170                         if (end == OBD_OBJECT_EOF)
3171                                 info->oti_fn_index = CL_PAGE_EOF;
3172                         LDLM_LOCK_PUT(tmp);
3173                 } else if (cl_page_own(env, io, page) == 0) {
3174                         /* discard the page */
3175                         cl_page_discard(env, io, page);
3176                         cl_page_disown(env, io, page);
3177                 } else {
3178                         LASSERT(page->cp_state == CPS_FREEING);
3179                 }
3180         }
3181
3182         info->oti_next_index = index + 1;
3183         return CLP_GANG_OKAY;
3184 }
3185
3186 static int discard_cb(const struct lu_env *env, struct cl_io *io,
3187                       struct osc_page *ops, void *cbdata)
3188 {
3189         struct osc_thread_info *info = osc_env_info(env);
3190         struct cl_page *page = ops->ops_cl.cpl_page;
3191
3192         /* page is top page. */
3193         info->oti_next_index = osc_index(ops) + 1;
3194         if (cl_page_own(env, io, page) == 0) {
3195                 KLASSERT(ergo(page->cp_type == CPT_CACHEABLE,
3196                               !PageDirty(cl_page_vmpage(page))));
3197
3198                 /* discard the page */
3199                 cl_page_discard(env, io, page);
3200                 cl_page_disown(env, io, page);
3201         } else {
3202                 LASSERT(page->cp_state == CPS_FREEING);
3203         }
3204
3205         return CLP_GANG_OKAY;
3206 }
3207
3208 /**
3209  * Discard pages protected by the given lock. This function traverses radix
3210  * tree to find all covering pages and discard them. If a page is being covered
3211  * by other locks, it should remain in cache.
3212  *
3213  * If error happens on any step, the process continues anyway (the reasoning
3214  * behind this being that lock cancellation cannot be delayed indefinitely).
3215  */
3216 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3217                            pgoff_t start, pgoff_t end, enum cl_lock_mode mode)
3218 {
3219         struct osc_thread_info *info = osc_env_info(env);
3220         struct cl_io *io = &info->oti_io;
3221         osc_page_gang_cbt cb;
3222         int res;
3223         int result;
3224
3225         io->ci_obj = cl_object_top(osc2cl(osc));
3226         io->ci_ignore_layout = 1;
3227         result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3228         if (result != 0)
3229                 goto out;
3230
3231         cb = mode == CLM_READ ? check_and_discard_cb : discard_cb;
3232         info->oti_fn_index = start;
3233         info->oti_next_index = start;
3234         do {
3235                 res = osc_page_gang_lookup(env, io, osc,
3236                                            info->oti_next_index, end, cb, osc);
3237                 if (info->oti_next_index > end)
3238                         break;
3239
3240                 if (res == CLP_GANG_RESCHED)
3241                         cond_resched();
3242         } while (res != CLP_GANG_OKAY);
3243 out:
3244         cl_io_fini(env, io);
3245         return result;
3246 }
3247
3248 /** @} osc */