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, 2016, 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 <lustre_osc.h>
  41
  42 #include "osc_internal.h"
  43
  44 static int extent_debug; /* set it to be true for more debug */
  45
  46 static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
  47 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
  48                            enum osc_extent_state state);
  49 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
  50                               struct osc_async_page *oap, int sent, int rc);
  51 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
  52                           int cmd);
  53 static int osc_refresh_count(const struct lu_env *env,
  54                              struct osc_async_page *oap, int cmd);
  55 static int osc_io_unplug_async(const struct lu_env *env,
  56                                struct client_obd *cli, struct osc_object *osc);
  57 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
  58                            unsigned int lost_grant, unsigned int dirty_grant);
  59
  60 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
  61                                   const char *func, int line);
  62 #define osc_extent_tree_dump(lvl, obj) \
  63         osc_extent_tree_dump0(lvl, obj, __func__, __LINE__)
  64
  65 static void osc_unreserve_grant(struct client_obd *cli, unsigned int reserved,
  66                                 unsigned int unused);
  67
  68 /** \addtogroup osc
  69  *  @{
  70  */
  71
  72 /* ------------------ osc extent ------------------ */
  73 static inline char *ext_flags(struct osc_extent *ext, char *flags)
  74 {
  75         char *buf = flags;
  76         *buf++ = ext->oe_rw ? 'r' : 'w';
  77         if (ext->oe_intree)
  78                 *buf++ = 'i';
  79         if (ext->oe_sync)
  80                 *buf++ = 'S';
  81         if (ext->oe_srvlock)
  82                 *buf++ = 's';
  83         if (ext->oe_hp)
  84                 *buf++ = 'h';
  85         if (ext->oe_urgent)
  86                 *buf++ = 'u';
  87         if (ext->oe_memalloc)
  88                 *buf++ = 'm';
  89         if (ext->oe_trunc_pending)
  90                 *buf++ = 't';
  91         if (ext->oe_fsync_wait)
  92                 *buf++ = 'Y';
  93         *buf = 0;
  94         return flags;
  95 }
  96
  97 static inline char list_empty_marker(struct list_head *list)
  98 {
  99         return list_empty(list) ? '-' : '+';
 100 }
 101
 102 #define EXTSTR       "[%lu -> %lu/%lu]"
 103 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
 104 static const char *oes_strings[] = {
 105         "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
 106
 107 #define OSC_EXTENT_DUMP(lvl, extent, fmt, ...) do {                           \
 108         struct osc_extent *__ext = (extent);                                  \
 109         char __buf[16];                                                       \
 110                                                                               \
 111         CDEBUG(lvl,                                                           \
 112                 "extent %p@{" EXTSTR ", "                                     \
 113                 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt,          \
 114                 /* ----- extent part 0 ----- */                               \
 115                 __ext, EXTPARA(__ext),                                        \
 116                 /* ----- part 1 ----- */                                      \
 117                 atomic_read(&__ext->oe_refc),                         \
 118                 atomic_read(&__ext->oe_users),                        \
 119                 list_empty_marker(&__ext->oe_link),                           \
 120                 oes_strings[__ext->oe_state], ext_flags(__ext, __buf),        \
 121                 __ext->oe_obj,                                                \
 122                 /* ----- part 2 ----- */                                      \
 123                 __ext->oe_grants, __ext->oe_nr_pages,                         \
 124                 list_empty_marker(&__ext->oe_pages),                          \
 125                 waitqueue_active(&__ext->oe_waitq) ? '+' : '-',               \
 126                 __ext->oe_dlmlock, __ext->oe_mppr, __ext->oe_owner,           \
 127                 /* ----- part 4 ----- */                                      \
 128                 ## __VA_ARGS__);                                              \
 129         if (lvl == D_ERROR && __ext->oe_dlmlock != NULL)                      \
 130                 LDLM_ERROR(__ext->oe_dlmlock, "extent: %p", __ext);           \
 131         else                                                                  \
 132                 LDLM_DEBUG(__ext->oe_dlmlock, "extent: %p", __ext);           \
 133 } while (0)
 134
 135 #undef EASSERTF
 136 #define EASSERTF(expr, ext, fmt, args...) do {                          \
 137         if (!(expr)) {                                                  \
 138                 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args);           \
 139                 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj);           \
 140                 LASSERT(expr);                                          \
 141         }                                                               \
 142 } while (0)
 143
 144 #undef EASSERT
 145 #define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
 146
 147 static inline struct osc_extent *rb_extent(struct rb_node *n)
 148 {
 149         if (n == NULL)
 150                 return NULL;
 151
 152         return container_of(n, struct osc_extent, oe_node);
 153 }
 154
 155 static inline struct osc_extent *next_extent(struct osc_extent *ext)
 156 {
 157         if (ext == NULL)
 158                 return NULL;
 159
 160         LASSERT(ext->oe_intree);
 161         return rb_extent(rb_next(&ext->oe_node));
 162 }
 163
 164 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
 165 {
 166         if (ext == NULL)
 167                 return NULL;
 168
 169         LASSERT(ext->oe_intree);
 170         return rb_extent(rb_prev(&ext->oe_node));
 171 }
 172
 173 static inline struct osc_extent *first_extent(struct osc_object *obj)
 174 {
 175         return rb_extent(rb_first(&obj->oo_root));
 176 }
 177
 178 /* object must be locked by caller. */
 179 static int osc_extent_sanity_check0(struct osc_extent *ext,
 180                                     const char *func, const int line)
 181 {
 182         struct osc_object *obj = ext->oe_obj;
 183         struct osc_async_page *oap;
 184         size_t page_count;
 185         int rc = 0;
 186
 187         if (!osc_object_is_locked(obj))
 188                 GOTO(out, rc = 9);
 189
 190         if (ext->oe_state >= OES_STATE_MAX)
 191                 GOTO(out, rc = 10);
 192
 193         if (atomic_read(&ext->oe_refc) <= 0)
 194                 GOTO(out, rc = 20);
 195
 196         if (atomic_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
 197                 GOTO(out, rc = 30);
 198
 199         switch (ext->oe_state) {
 200         case OES_INV:
 201                 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
 202                         GOTO(out, rc = 35);
 203                 GOTO(out, rc = 0);
 204                 break;
 205         case OES_ACTIVE:
 206                 if (atomic_read(&ext->oe_users) == 0)
 207                         GOTO(out, rc = 40);
 208                 if (ext->oe_hp)
 209                         GOTO(out, rc = 50);
 210                 if (ext->oe_fsync_wait && !ext->oe_urgent)
 211                         GOTO(out, rc = 55);
 212                 break;
 213         case OES_CACHE:
 214                 if (ext->oe_grants == 0)
 215                         GOTO(out, rc = 60);
 216                 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
 217                         GOTO(out, rc = 65);
 218         default:
 219                 if (atomic_read(&ext->oe_users) > 0)
 220                         GOTO(out, rc = 70);
 221         }
 222
 223         if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
 224                 GOTO(out, rc = 80);
 225
 226         if (ext->oe_sync && ext->oe_grants > 0)
 227                 GOTO(out, rc = 90);
 228
 229         if (ext->oe_dlmlock != NULL &&
 230             ext->oe_dlmlock->l_resource->lr_type == LDLM_EXTENT &&
 231             !ldlm_is_failed(ext->oe_dlmlock)) {
 232                 struct ldlm_extent *extent;
 233
 234                 extent = &ext->oe_dlmlock->l_policy_data.l_extent;
 235                 if (!(extent->start <= cl_offset(osc2cl(obj), ext->oe_start) &&
 236                       extent->end   >= cl_offset(osc2cl(obj), ext->oe_max_end)))
 237                         GOTO(out, rc = 100);
 238
 239                 if (!(ext->oe_dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP)))
 240                         GOTO(out, rc = 102);
 241         }
 242
 243         if (ext->oe_nr_pages > ext->oe_mppr)
 244                 GOTO(out, rc = 105);
 245
 246         /* Do not verify page list if extent is in RPC. This is because an
 247          * in-RPC extent is supposed to be exclusively accessible w/o lock. */
 248         if (ext->oe_state > OES_CACHE)
 249                 GOTO(out, rc = 0);
 250
 251         if (!extent_debug)
 252                 GOTO(out, rc = 0);
 253
 254         page_count = 0;
 255         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
 256                 pgoff_t index = osc_index(oap2osc(oap));
 257                 ++page_count;
 258                 if (index > ext->oe_end || index < ext->oe_start)
 259                         GOTO(out, rc = 110);
 260         }
 261         if (page_count != ext->oe_nr_pages)
 262                 GOTO(out, rc = 120);
 263
 264 out:
 265         if (rc != 0)
 266                 OSC_EXTENT_DUMP(D_ERROR, ext,
 267                                 "%s:%d sanity check %p failed with rc = %d\n",
 268                                 func, line, ext, rc);
 269         return rc;
 270 }
 271
 272 #define sanity_check_nolock(ext) \
 273         osc_extent_sanity_check0(ext, __func__, __LINE__)
 274
 275 #define sanity_check(ext) ({                                                   \
 276         int __res;                                                             \
 277         osc_object_lock((ext)->oe_obj);                                        \
 278         __res = sanity_check_nolock(ext);                                      \
 279         osc_object_unlock((ext)->oe_obj);                                      \
 280         __res;                                                                 \
 281 })
 282
 283
 284 /**
 285  * sanity check - to make sure there is no overlapped extent in the tree.
 286  */
 287 static int osc_extent_is_overlapped(struct osc_object *obj,
 288                                     struct osc_extent *ext)
 289 {
 290         struct osc_extent *tmp;
 291
 292         LASSERT(osc_object_is_locked(obj));
 293
 294         if (!extent_debug)
 295                 return 0;
 296
 297         for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
 298                 if (tmp == ext)
 299                         continue;
 300                 if (tmp->oe_end >= ext->oe_start &&
 301                     tmp->oe_start <= ext->oe_end)
 302                         return 1;
 303         }
 304         return 0;
 305 }
 306
 307 static void osc_extent_state_set(struct osc_extent *ext, int state)
 308 {
 309         LASSERT(osc_object_is_locked(ext->oe_obj));
 310         LASSERT(state >= OES_INV && state < OES_STATE_MAX);
 311
 312         /* Never try to sanity check a state changing extent :-) */
 313         /* LASSERT(sanity_check_nolock(ext) == 0); */
 314
 315         /* TODO: validate the state machine */
 316         ext->oe_state = state;
 317         wake_up_all(&ext->oe_waitq);
 318 }
 319
 320 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
 321 {
 322         struct osc_extent *ext;
 323
 324         OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_NOFS);
 325         if (ext == NULL)
 326                 return NULL;
 327
 328         RB_CLEAR_NODE(&ext->oe_node);
 329         ext->oe_obj = obj;
 330         cl_object_get(osc2cl(obj));
 331         atomic_set(&ext->oe_refc, 1);
 332         atomic_set(&ext->oe_users, 0);
 333         INIT_LIST_HEAD(&ext->oe_link);
 334         ext->oe_state = OES_INV;
 335         INIT_LIST_HEAD(&ext->oe_pages);
 336         init_waitqueue_head(&ext->oe_waitq);
 337         ext->oe_dlmlock = NULL;
 338
 339         return ext;
 340 }
 341
 342 static void osc_extent_free(struct osc_extent *ext)
 343 {
 344         OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
 345 }
 346
 347 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
 348 {
 349         LASSERT(atomic_read(&ext->oe_refc) >= 0);
 350         atomic_inc(&ext->oe_refc);
 351         return ext;
 352 }
 353
 354 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
 355 {
 356         LASSERT(atomic_read(&ext->oe_refc) > 0);
 357         if (atomic_dec_and_test(&ext->oe_refc)) {
 358                 LASSERT(list_empty(&ext->oe_link));
 359                 LASSERT(atomic_read(&ext->oe_users) == 0);
 360                 LASSERT(ext->oe_state == OES_INV);
 361                 LASSERT(!ext->oe_intree);
 362
 363                 if (ext->oe_dlmlock != NULL) {
 364                         lu_ref_add(&ext->oe_dlmlock->l_reference,
 365                                    "osc_extent", ext);
 366                         LDLM_LOCK_PUT(ext->oe_dlmlock);
 367                         ext->oe_dlmlock = NULL;
 368                 }
 369                 cl_object_put(env, osc2cl(ext->oe_obj));
 370                 osc_extent_free(ext);
 371         }
 372 }
 373
 374 /**
 375  * osc_extent_put_trust() is a special version of osc_extent_put() when
 376  * it's known that the caller is not the last user. This is to address the
 377  * problem of lacking of lu_env ;-).
 378  */
 379 static void osc_extent_put_trust(struct osc_extent *ext)
 380 {
 381         LASSERT(atomic_read(&ext->oe_refc) > 1);
 382         LASSERT(osc_object_is_locked(ext->oe_obj));
 383         atomic_dec(&ext->oe_refc);
 384 }
 385
 386 /**
 387  * Return the extent which includes pgoff @index, or return the greatest
 388  * previous extent in the tree.
 389  */
 390 static struct osc_extent *osc_extent_search(struct osc_object *obj,
 391                                             pgoff_t index)
 392 {
 393         struct rb_node    *n = obj->oo_root.rb_node;
 394         struct osc_extent *tmp, *p = NULL;
 395
 396         LASSERT(osc_object_is_locked(obj));
 397         while (n != NULL) {
 398                 tmp = rb_extent(n);
 399                 if (index < tmp->oe_start) {
 400                         n = n->rb_left;
 401                 } else if (index > tmp->oe_end) {
 402                         p = rb_extent(n);
 403                         n = n->rb_right;
 404                 } else {
 405                         return tmp;
 406                 }
 407         }
 408         return p;
 409 }
 410
 411 /*
 412  * Return the extent covering @index, otherwise return NULL.
 413  * caller must have held object lock.
 414  */
 415 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
 416                                             pgoff_t index)
 417 {
 418         struct osc_extent *ext;
 419
 420         ext = osc_extent_search(obj, index);
 421         if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
 422                 return osc_extent_get(ext);
 423         return NULL;
 424 }
 425
 426 /* caller must have held object lock. */
 427 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
 428 {
 429         struct rb_node   **n      = &obj->oo_root.rb_node;
 430         struct rb_node    *parent = NULL;
 431         struct osc_extent *tmp;
 432
 433         LASSERT(ext->oe_intree == 0);
 434         LASSERT(ext->oe_obj == obj);
 435         LASSERT(osc_object_is_locked(obj));
 436         while (*n != NULL) {
 437                 tmp = rb_extent(*n);
 438                 parent = *n;
 439
 440                 if (ext->oe_end < tmp->oe_start)
 441                         n = &(*n)->rb_left;
 442                 else if (ext->oe_start > tmp->oe_end)
 443                         n = &(*n)->rb_right;
 444                 else
 445                         EASSERTF(0, tmp, EXTSTR"\n", EXTPARA(ext));
 446         }
 447         rb_link_node(&ext->oe_node, parent, n);
 448         rb_insert_color(&ext->oe_node, &obj->oo_root);
 449         osc_extent_get(ext);
 450         ext->oe_intree = 1;
 451 }
 452
 453 /* caller must have held object lock. */
 454 static void osc_extent_erase(struct osc_extent *ext)
 455 {
 456         struct osc_object *obj = ext->oe_obj;
 457         LASSERT(osc_object_is_locked(obj));
 458         if (ext->oe_intree) {
 459                 rb_erase(&ext->oe_node, &obj->oo_root);
 460                 ext->oe_intree = 0;
 461                 /* rbtree held a refcount */
 462                 osc_extent_put_trust(ext);
 463         }
 464 }
 465
 466 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
 467 {
 468         struct osc_object *obj = ext->oe_obj;
 469
 470         LASSERT(osc_object_is_locked(obj));
 471         LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
 472         if (ext->oe_state == OES_CACHE) {
 473                 osc_extent_state_set(ext, OES_ACTIVE);
 474                 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
 475         }
 476         atomic_inc(&ext->oe_users);
 477         list_del_init(&ext->oe_link);
 478         return osc_extent_get(ext);
 479 }
 480
 481 static void __osc_extent_remove(struct osc_extent *ext)
 482 {
 483         LASSERT(osc_object_is_locked(ext->oe_obj));
 484         LASSERT(list_empty(&ext->oe_pages));
 485         osc_extent_erase(ext);
 486         list_del_init(&ext->oe_link);
 487         osc_extent_state_set(ext, OES_INV);
 488         OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
 489 }
 490
 491 static void osc_extent_remove(struct osc_extent *ext)
 492 {
 493         struct osc_object *obj = ext->oe_obj;
 494
 495         osc_object_lock(obj);
 496         __osc_extent_remove(ext);
 497         osc_object_unlock(obj);
 498 }
 499
 500 /**
 501  * This function is used to merge extents to get better performance. It checks
 502  * if @cur and @victim are contiguous at block level.
 503  */
 504 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
 505                             struct osc_extent *victim)
 506 {
 507         struct osc_object       *obj = cur->oe_obj;
 508         struct client_obd       *cli = osc_cli(obj);
 509         pgoff_t                  chunk_start;
 510         pgoff_t                  chunk_end;
 511         int                      ppc_bits;
 512
 513         LASSERT(cur->oe_state == OES_CACHE);
 514         LASSERT(osc_object_is_locked(obj));
 515         if (victim == NULL)
 516                 return -EINVAL;
 517
 518         if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
 519                 return -EBUSY;
 520
 521         if (cur->oe_max_end != victim->oe_max_end)
 522                 return -ERANGE;
 523
 524         LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
 525         ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_SHIFT;
 526         chunk_start = cur->oe_start >> ppc_bits;
 527         chunk_end   = cur->oe_end   >> ppc_bits;
 528         if (chunk_start   != (victim->oe_end >> ppc_bits) + 1 &&
 529             chunk_end + 1 != victim->oe_start >> ppc_bits)
 530                 return -ERANGE;
 531
 532         /* overall extent size should not exceed the max supported limit
 533          * reported by the server */
 534         if (cur->oe_end - cur->oe_start + 1 +
 535             victim->oe_end - victim->oe_start + 1 > cli->cl_max_extent_pages)
 536                 return -ERANGE;
 537
 538         OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
 539
 540         cur->oe_start     = min(cur->oe_start, victim->oe_start);
 541         cur->oe_end       = max(cur->oe_end,   victim->oe_end);
 542         /* per-extent tax should be accounted only once for the whole extent */
 543         cur->oe_grants   += victim->oe_grants - cli->cl_grant_extent_tax;
 544         cur->oe_nr_pages += victim->oe_nr_pages;
 545         /* only the following bits are needed to merge */
 546         cur->oe_urgent   |= victim->oe_urgent;
 547         cur->oe_memalloc |= victim->oe_memalloc;
 548         list_splice_init(&victim->oe_pages, &cur->oe_pages);
 549         list_del_init(&victim->oe_link);
 550         victim->oe_nr_pages = 0;
 551
 552         osc_extent_get(victim);
 553         __osc_extent_remove(victim);
 554         osc_extent_put(env, victim);
 555
 556         OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
 557         return 0;
 558 }
 559
 560 /**
 561  * Drop user count of osc_extent, and unplug IO asynchronously.
 562  */
 563 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
 564 {
 565         struct osc_object *obj = ext->oe_obj;
 566         struct client_obd *cli = osc_cli(obj);
 567         int rc = 0;
 568         ENTRY;
 569
 570         LASSERT(atomic_read(&ext->oe_users) > 0);
 571         LASSERT(sanity_check(ext) == 0);
 572         LASSERT(ext->oe_grants > 0);
 573
 574         if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
 575                 LASSERT(ext->oe_state == OES_ACTIVE);
 576                 if (ext->oe_trunc_pending) {
 577                         /* a truncate process is waiting for this extent.
 578                          * This may happen due to a race, check
 579                          * osc_cache_truncate_start(). */
 580                         osc_extent_state_set(ext, OES_TRUNC);
 581                         ext->oe_trunc_pending = 0;
 582                 } else {
 583                         int grant = 0;
 584
 585                         osc_extent_state_set(ext, OES_CACHE);
 586                         osc_update_pending(obj, OBD_BRW_WRITE,
 587                                            ext->oe_nr_pages);
 588
 589                         /* try to merge the previous and next extent. */
 590                         if (osc_extent_merge(env, ext, prev_extent(ext)) == 0)
 591                                 grant += cli->cl_grant_extent_tax;
 592                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 593                                 grant += cli->cl_grant_extent_tax;
 594                         if (grant > 0)
 595                                 osc_unreserve_grant(cli, 0, grant);
 596
 597                         if (ext->oe_urgent)
 598                                 list_move_tail(&ext->oe_link,
 599                                                &obj->oo_urgent_exts);
 600                         else if (ext->oe_nr_pages == ext->oe_mppr) {
 601                                 list_move_tail(&ext->oe_link,
 602                                                &obj->oo_full_exts);
 603                         }
 604                 }
 605                 osc_object_unlock(obj);
 606
 607                 osc_io_unplug_async(env, cli, obj);
 608         }
 609         osc_extent_put(env, ext);
 610         RETURN(rc);
 611 }
 612
 613 static inline int overlapped(struct osc_extent *ex1, struct osc_extent *ex2)
 614 {
 615         return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
 616 }
 617
 618 /**
 619  * Find or create an extent which includes @index, core function to manage
 620  * extent tree.
 621  */
 622 static struct osc_extent *osc_extent_find(const struct lu_env *env,
 623                                           struct osc_object *obj, pgoff_t index,
 624                                           unsigned int *grants)
 625 {
 626         struct client_obd *cli = osc_cli(obj);
 627         struct osc_lock   *olck;
 628         struct cl_lock_descr *descr;
 629         struct osc_extent *cur;
 630         struct osc_extent *ext;
 631         struct osc_extent *conflict = NULL;
 632         struct osc_extent *found = NULL;
 633         pgoff_t    chunk;
 634         pgoff_t    max_end;
 635         unsigned int max_pages; /* max_pages_per_rpc */
 636         unsigned int chunksize;
 637         int        ppc_bits; /* pages per chunk bits */
 638         pgoff_t    chunk_mask;
 639         int        rc;
 640         ENTRY;
 641
 642         cur = osc_extent_alloc(obj);
 643         if (cur == NULL)
 644                 RETURN(ERR_PTR(-ENOMEM));
 645
 646         olck = osc_env_io(env)->oi_write_osclock;
 647         LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
 648         LASSERT(olck->ols_state == OLS_GRANTED);
 649
 650         descr = &olck->ols_cl.cls_lock->cll_descr;
 651         LASSERT(descr->cld_mode >= CLM_WRITE);
 652
 653         LASSERTF(cli->cl_chunkbits >= PAGE_SHIFT,
 654                  "chunkbits: %u\n", cli->cl_chunkbits);
 655         ppc_bits   = cli->cl_chunkbits - PAGE_SHIFT;
 656         chunk_mask = ~((1 << ppc_bits) - 1);
 657         chunksize  = 1 << cli->cl_chunkbits;
 658         chunk      = index >> ppc_bits;
 659
 660         /* align end to RPC edge. */
 661         max_pages = cli->cl_max_pages_per_rpc;
 662         if ((max_pages & ~chunk_mask) != 0) {
 663                 CERROR("max_pages: %#x chunkbits: %u chunk_mask: %#lx\n",
 664                        max_pages, cli->cl_chunkbits, chunk_mask);
 665                 RETURN(ERR_PTR(-EINVAL));
 666         }
 667         max_end = index - (index % max_pages) + max_pages - 1;
 668         max_end = min_t(pgoff_t, max_end, descr->cld_end);
 669
 670         /* initialize new extent by parameters so far */
 671         cur->oe_max_end = max_end;
 672         cur->oe_start   = index & chunk_mask;
 673         cur->oe_end     = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
 674         if (cur->oe_start < descr->cld_start)
 675                 cur->oe_start = descr->cld_start;
 676         if (cur->oe_end > max_end)
 677                 cur->oe_end = max_end;
 678         cur->oe_grants  = 0;
 679         cur->oe_mppr    = max_pages;
 680         if (olck->ols_dlmlock != NULL) {
 681                 LASSERT(olck->ols_hold);
 682                 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
 683                 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
 684         }
 685
 686         /* grants has been allocated by caller */
 687         LASSERTF(*grants >= chunksize + cli->cl_grant_extent_tax,
 688                  "%u/%u/%u.\n", *grants, chunksize, cli->cl_grant_extent_tax);
 689         LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
 690                  EXTPARA(cur));
 691
 692 restart:
 693         osc_object_lock(obj);
 694         ext = osc_extent_search(obj, cur->oe_start);
 695         if (ext == NULL)
 696                 ext = first_extent(obj);
 697         while (ext != NULL) {
 698                 pgoff_t ext_chk_start = ext->oe_start >> ppc_bits;
 699                 pgoff_t ext_chk_end   = ext->oe_end   >> ppc_bits;
 700
 701                 LASSERT(sanity_check_nolock(ext) == 0);
 702                 if (chunk > ext_chk_end + 1)
 703                         break;
 704
 705                 /* if covering by different locks, no chance to match */
 706                 if (olck->ols_dlmlock != ext->oe_dlmlock) {
 707                         EASSERTF(!overlapped(ext, cur), ext,
 708                                  EXTSTR"\n", EXTPARA(cur));
 709
 710                         ext = next_extent(ext);
 711                         continue;
 712                 }
 713
 714                 /* discontiguous chunks? */
 715                 if (chunk + 1 < ext_chk_start) {
 716                         ext = next_extent(ext);
 717                         continue;
 718                 }
 719
 720                 /* ok, from now on, ext and cur have these attrs:
 721                  * 1. covered by the same lock
 722                  * 2. contiguous at chunk level or overlapping. */
 723
 724                 if (overlapped(ext, cur)) {
 725                         /* cur is the minimum unit, so overlapping means
 726                          * full contain. */
 727                         EASSERTF((ext->oe_start <= cur->oe_start &&
 728                                   ext->oe_end >= cur->oe_end),
 729                                  ext, EXTSTR"\n", EXTPARA(cur));
 730
 731                         if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
 732                                 /* for simplicity, we wait for this extent to
 733                                  * finish before going forward. */
 734                                 conflict = osc_extent_get(ext);
 735                                 break;
 736                         }
 737
 738                         found = osc_extent_hold(ext);
 739                         break;
 740                 }
 741
 742                 /* non-overlapped extent */
 743                 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
 744                         /* we can't do anything for a non OES_CACHE extent, or
 745                          * if there is someone waiting for this extent to be
 746                          * flushed, try next one. */
 747                         ext = next_extent(ext);
 748                         continue;
 749                 }
 750
 751                 /* check if they belong to the same rpc slot before trying to
 752                  * merge. the extents are not overlapped and contiguous at
 753                  * chunk level to get here. */
 754                 if (ext->oe_max_end != max_end) {
 755                         /* if they don't belong to the same RPC slot or
 756                          * max_pages_per_rpc has ever changed, do not merge. */
 757                         ext = next_extent(ext);
 758                         continue;
 759                 }
 760
 761                 /* check whether maximum extent size will be hit */
 762                 if ((ext_chk_end - ext_chk_start + 1 + 1) << ppc_bits >
 763                     cli->cl_max_extent_pages) {
 764                         ext = next_extent(ext);
 765                         continue;
 766                 }
 767
 768                 /* it's required that an extent must be contiguous at chunk
 769                  * level so that we know the whole extent is covered by grant
 770                  * (the pages in the extent are NOT required to be contiguous).
 771                  * Otherwise, it will be too much difficult to know which
 772                  * chunks have grants allocated. */
 773
 774                 /* try to do front merge - extend ext's start */
 775                 if (chunk + 1 == ext_chk_start) {
 776                         /* ext must be chunk size aligned */
 777                         EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
 778
 779                         /* pull ext's start back to cover cur */
 780                         ext->oe_start   = cur->oe_start;
 781                         ext->oe_grants += chunksize;
 782                         LASSERT(*grants >= chunksize);
 783                         *grants -= chunksize;
 784
 785                         found = osc_extent_hold(ext);
 786                 } else if (chunk == ext_chk_end + 1) {
 787                         /* rear merge */
 788                         ext->oe_end     = cur->oe_end;
 789                         ext->oe_grants += chunksize;
 790                         LASSERT(*grants >= chunksize);
 791                         *grants -= chunksize;
 792
 793                         /* try to merge with the next one because we just fill
 794                          * in a gap */
 795                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 796                                 /* we can save extent tax from next extent */
 797                                 *grants += cli->cl_grant_extent_tax;
 798
 799                         found = osc_extent_hold(ext);
 800                 }
 801                 if (found != NULL)
 802                         break;
 803
 804                 ext = next_extent(ext);
 805         }
 806
 807         osc_extent_tree_dump(D_CACHE, obj);
 808         if (found != NULL) {
 809                 LASSERT(conflict == NULL);
 810                 if (!IS_ERR(found)) {
 811                         LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
 812                         OSC_EXTENT_DUMP(D_CACHE, found,
 813                                         "found caching ext for %lu.\n", index);
 814                 }
 815         } else if (conflict == NULL) {
 816                 /* create a new extent */
 817                 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
 818                 cur->oe_grants = chunksize + cli->cl_grant_extent_tax;
 819                 LASSERT(*grants >= cur->oe_grants);
 820                 *grants -= cur->oe_grants;
 821
 822                 cur->oe_state = OES_CACHE;
 823                 found = osc_extent_hold(cur);
 824                 osc_extent_insert(obj, cur);
 825                 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
 826                                 index, descr->cld_end);
 827         }
 828         osc_object_unlock(obj);
 829
 830         if (conflict != NULL) {
 831                 LASSERT(found == NULL);
 832
 833                 /* waiting for IO to finish. Please notice that it's impossible
 834                  * to be an OES_TRUNC extent. */
 835                 rc = osc_extent_wait(env, conflict, OES_INV);
 836                 osc_extent_put(env, conflict);
 837                 conflict = NULL;
 838                 if (rc < 0)
 839                         GOTO(out, found = ERR_PTR(rc));
 840
 841                 goto restart;
 842         }
 843         EXIT;
 844
 845 out:
 846         osc_extent_put(env, cur);
 847         return found;
 848 }
 849
 850 /**
 851  * Called when IO is finished to an extent.
 852  */
 853 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
 854                       int sent, int rc)
 855 {
 856         struct client_obd *cli = osc_cli(ext->oe_obj);
 857         struct osc_async_page *oap;
 858         struct osc_async_page *tmp;
 859         int nr_pages = ext->oe_nr_pages;
 860         int lost_grant = 0;
 861         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
 862         loff_t last_off = 0;
 863         int last_count = -1;
 864         ENTRY;
 865
 866         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
 867
 868         ext->oe_rc = rc ?: ext->oe_nr_pages;
 869         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
 870
 871         osc_lru_add_batch(cli, &ext->oe_pages);
 872         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 873                                      oap_pending_item) {
 874                 list_del_init(&oap->oap_rpc_item);
 875                 list_del_init(&oap->oap_pending_item);
 876                 if (last_off <= oap->oap_obj_off) {
 877                         last_off = oap->oap_obj_off;
 878                         last_count = oap->oap_count;
 879                 }
 880
 881                 --ext->oe_nr_pages;
 882                 osc_ap_completion(env, cli, oap, sent, rc);
 883         }
 884         EASSERT(ext->oe_nr_pages == 0, ext);
 885
 886         if (!sent) {
 887                 lost_grant = ext->oe_grants;
 888         } else if (blocksize < PAGE_SIZE &&
 889                    last_count != PAGE_SIZE) {
 890                 /* For short writes we shouldn't count parts of pages that
 891                  * span a whole chunk on the OST side, or our accounting goes
 892                  * wrong.  Should match the code in filter_grant_check. */
 893                 int offset = last_off & ~PAGE_MASK;
 894                 int count = last_count + (offset & (blocksize - 1));
 895                 int end = (offset + last_count) & (blocksize - 1);
 896                 if (end)
 897                         count += blocksize - end;
 898
 899                 lost_grant = PAGE_SIZE - count;
 900         }
 901         if (ext->oe_grants > 0)
 902                 osc_free_grant(cli, nr_pages, lost_grant, ext->oe_grants);
 903
 904         osc_extent_remove(ext);
 905         /* put the refcount for RPC */
 906         osc_extent_put(env, ext);
 907         RETURN(0);
 908 }
 909
 910 static int extent_wait_cb(struct osc_extent *ext, enum osc_extent_state state)
 911 {
 912         int ret;
 913
 914         osc_object_lock(ext->oe_obj);
 915         ret = ext->oe_state == state;
 916         osc_object_unlock(ext->oe_obj);
 917
 918         return ret;
 919 }
 920
 921 /**
 922  * Wait for the extent's state to become @state.
 923  */
 924 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
 925                            enum osc_extent_state state)
 926 {
 927         struct osc_object *obj = ext->oe_obj;
 928         struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
 929                                                   LWI_ON_SIGNAL_NOOP, NULL);
 930         int rc = 0;
 931         ENTRY;
 932
 933         osc_object_lock(obj);
 934         LASSERT(sanity_check_nolock(ext) == 0);
 935         /* `Kick' this extent only if the caller is waiting for it to be
 936          * written out. */
 937         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
 938                 if (ext->oe_state == OES_ACTIVE) {
 939                         ext->oe_urgent = 1;
 940                 } else if (ext->oe_state == OES_CACHE) {
 941                         ext->oe_urgent = 1;
 942                         osc_extent_hold(ext);
 943                         rc = 1;
 944                 }
 945         }
 946         osc_object_unlock(obj);
 947         if (rc == 1)
 948                 osc_extent_release(env, ext);
 949
 950         /* wait for the extent until its state becomes @state */
 951         rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
 952         if (rc == -ETIMEDOUT) {
 953                 OSC_EXTENT_DUMP(D_ERROR, ext,
 954                         "%s: wait ext to %u timedout, recovery in progress?\n",
 955                         cli_name(osc_cli(obj)), state);
 956
 957                 lwi = LWI_INTR(NULL, NULL);
 958                 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state),
 959                                   &lwi);
 960         }
 961         if (rc == 0 && ext->oe_rc < 0)
 962                 rc = ext->oe_rc;
 963         RETURN(rc);
 964 }
 965
 966 /**
 967  * Discard pages with index greater than @size. If @ext is overlapped with
 968  * @size, then partial truncate happens.
 969  */
 970 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
 971                                 bool partial)
 972 {
 973         struct lu_env         *env;
 974         struct cl_io          *io;
 975         struct osc_object     *obj = ext->oe_obj;
 976         struct client_obd     *cli = osc_cli(obj);
 977         struct osc_async_page *oap;
 978         struct osc_async_page *tmp;
 979         int                    pages_in_chunk = 0;
 980         int                    ppc_bits    = cli->cl_chunkbits -
 981                                              PAGE_SHIFT;
 982         __u64                  trunc_chunk = trunc_index >> ppc_bits;
 983         int                    grants   = 0;
 984         int                    nr_pages = 0;
 985         int                    rc       = 0;
 986         __u16                  refcheck;
 987         ENTRY;
 988
 989         LASSERT(sanity_check(ext) == 0);
 990         LASSERT(ext->oe_state == OES_TRUNC);
 991         LASSERT(!ext->oe_urgent);
 992
 993         /* Request new lu_env.
 994          * We can't use that env from osc_cache_truncate_start() because
 995          * it's from lov_io_sub and not fully initialized. */
 996         env = cl_env_get(&refcheck);
 997         if (IS_ERR(env))
 998                 RETURN(PTR_ERR(env));
 999
1000         io  = osc_env_thread_io(env);
1001         io->ci_obj = cl_object_top(osc2cl(obj));
1002         io->ci_ignore_layout = 1;
1003         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
1004         if (rc < 0)
1005                 GOTO(out, rc);
1006
1007         /* discard all pages with index greater than trunc_index */
1008         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
1009                                      oap_pending_item) {
1010                 pgoff_t index = osc_index(oap2osc(oap));
1011                 struct cl_page  *page = oap2cl_page(oap);
1012
1013                 LASSERT(list_empty(&oap->oap_rpc_item));
1014
1015                 /* only discard the pages with their index greater than
1016                  * trunc_index, and ... */
1017                 if (index < trunc_index ||
1018                     (index == trunc_index && partial)) {
1019                         /* accounting how many pages remaining in the chunk
1020                          * so that we can calculate grants correctly. */
1021                         if (index >> ppc_bits == trunc_chunk)
1022                                 ++pages_in_chunk;
1023                         continue;
1024                 }
1025
1026                 list_del_init(&oap->oap_pending_item);
1027
1028                 cl_page_get(page);
1029                 lu_ref_add(&page->cp_reference, "truncate", current);
1030
1031                 if (cl_page_own(env, io, page) == 0) {
1032                         cl_page_discard(env, io, page);
1033                         cl_page_disown(env, io, page);
1034                 } else {
1035                         LASSERT(page->cp_state == CPS_FREEING);
1036                         LASSERT(0);
1037                 }
1038
1039                 lu_ref_del(&page->cp_reference, "truncate", current);
1040                 cl_page_put(env, page);
1041
1042                 --ext->oe_nr_pages;
1043                 ++nr_pages;
1044         }
1045         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1046                       ext->oe_nr_pages == 0),
1047                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1048
1049         osc_object_lock(obj);
1050         if (ext->oe_nr_pages == 0) {
1051                 LASSERT(pages_in_chunk == 0);
1052                 grants = ext->oe_grants;
1053                 ext->oe_grants = 0;
1054         } else { /* calculate how many grants we can free */
1055                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1056                 pgoff_t last_index;
1057
1058
1059                 /* if there is no pages in this chunk, we can also free grants
1060                  * for the last chunk */
1061                 if (pages_in_chunk == 0) {
1062                         /* if this is the 1st chunk and no pages in this chunk,
1063                          * ext->oe_nr_pages must be zero, so we should be in
1064                          * the other if-clause. */
1065                         LASSERT(trunc_chunk > 0);
1066                         --trunc_chunk;
1067                         ++chunks;
1068                 }
1069
1070                 /* this is what we can free from this extent */
1071                 grants          = chunks << cli->cl_chunkbits;
1072                 ext->oe_grants -= grants;
1073                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1074                 ext->oe_end     = min(last_index, ext->oe_max_end);
1075                 LASSERT(ext->oe_end >= ext->oe_start);
1076                 LASSERT(ext->oe_grants > 0);
1077         }
1078         osc_object_unlock(obj);
1079
1080         if (grants > 0 || nr_pages > 0)
1081                 osc_free_grant(cli, nr_pages, grants, grants);
1082
1083 out:
1084         cl_io_fini(env, io);
1085         cl_env_put(env, &refcheck);
1086         RETURN(rc);
1087 }
1088
1089 /**
1090  * This function is used to make the extent prepared for transfer.
1091  * A race with flusing page - ll_writepage() has to be handled cautiously.
1092  */
1093 static int osc_extent_make_ready(const struct lu_env *env,
1094                                  struct osc_extent *ext)
1095 {
1096         struct osc_async_page *oap;
1097         struct osc_async_page *last = NULL;
1098         struct osc_object *obj = ext->oe_obj;
1099         unsigned int page_count = 0;
1100         int rc;
1101         ENTRY;
1102
1103         /* we're going to grab page lock, so object lock must not be taken. */
1104         LASSERT(sanity_check(ext) == 0);
1105         /* in locking state, any process should not touch this extent. */
1106         EASSERT(ext->oe_state == OES_LOCKING, ext);
1107         EASSERT(ext->oe_owner != NULL, ext);
1108
1109         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1110
1111         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1112                 ++page_count;
1113                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1114                         last = oap;
1115
1116                 /* checking ASYNC_READY is race safe */
1117                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1118                         continue;
1119
1120                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1121                 switch (rc) {
1122                 case 0:
1123                         spin_lock(&oap->oap_lock);
1124                         oap->oap_async_flags |= ASYNC_READY;
1125                         spin_unlock(&oap->oap_lock);
1126                         break;
1127                 case -EALREADY:
1128                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1129                         break;
1130                 default:
1131                         LASSERTF(0, "unknown return code: %d\n", rc);
1132                 }
1133         }
1134
1135         LASSERT(page_count == ext->oe_nr_pages);
1136         LASSERT(last != NULL);
1137         /* the last page is the only one we need to refresh its count by
1138          * the size of file. */
1139         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1140                 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1141                 LASSERT(last_oap_count > 0);
1142                 LASSERT(last->oap_page_off + last_oap_count <= PAGE_SIZE);
1143                 last->oap_count = last_oap_count;
1144                 spin_lock(&last->oap_lock);
1145                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1146                 spin_unlock(&last->oap_lock);
1147         }
1148
1149         /* for the rest of pages, we don't need to call osf_refresh_count()
1150          * because it's known they are not the last page */
1151         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1152                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1153                         oap->oap_count = PAGE_SIZE - oap->oap_page_off;
1154                         spin_lock(&oap->oap_lock);
1155                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1156                         spin_unlock(&oap->oap_lock);
1157                 }
1158         }
1159
1160         osc_object_lock(obj);
1161         osc_extent_state_set(ext, OES_RPC);
1162         osc_object_unlock(obj);
1163         /* get a refcount for RPC. */
1164         osc_extent_get(ext);
1165
1166         RETURN(0);
1167 }
1168
1169 /**
1170  * Quick and simple version of osc_extent_find(). This function is frequently
1171  * called to expand the extent for the same IO. To expand the extent, the
1172  * page index must be in the same or next chunk of ext->oe_end.
1173  */
1174 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index,
1175                              unsigned int *grants)
1176 {
1177         struct osc_object *obj = ext->oe_obj;
1178         struct client_obd *cli = osc_cli(obj);
1179         struct osc_extent *next;
1180         int ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1181         pgoff_t chunk = index >> ppc_bits;
1182         pgoff_t end_chunk;
1183         pgoff_t end_index;
1184         unsigned int chunksize = 1 << cli->cl_chunkbits;
1185         int rc = 0;
1186         ENTRY;
1187
1188         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1189         osc_object_lock(obj);
1190         LASSERT(sanity_check_nolock(ext) == 0);
1191         end_chunk = ext->oe_end >> ppc_bits;
1192         if (chunk > end_chunk + 1)
1193                 GOTO(out, rc = -ERANGE);
1194
1195         if (end_chunk >= chunk)
1196                 GOTO(out, rc = 0);
1197
1198         LASSERT(end_chunk + 1 == chunk);
1199
1200         /* try to expand this extent to cover @index */
1201         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1202
1203         /* don't go over the maximum extent size reported by server */
1204         if (end_index - ext->oe_start + 1 > cli->cl_max_extent_pages)
1205                 GOTO(out, rc = -ERANGE);
1206
1207         next = next_extent(ext);
1208         if (next != NULL && next->oe_start <= end_index)
1209                 /* complex mode - overlapped with the next extent,
1210                  * this case will be handled by osc_extent_find() */
1211                 GOTO(out, rc = -EAGAIN);
1212
1213         ext->oe_end = end_index;
1214         ext->oe_grants += chunksize;
1215         LASSERT(*grants >= chunksize);
1216         *grants -= chunksize;
1217         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1218                  "overlapped after expanding for %lu.\n", index);
1219         EXIT;
1220
1221 out:
1222         osc_object_unlock(obj);
1223         RETURN(rc);
1224 }
1225
1226 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
1227                                   const char *func, int line)
1228 {
1229         struct osc_extent *ext;
1230         int cnt;
1231
1232         if (!cfs_cdebug_show(level, DEBUG_SUBSYSTEM))
1233                 return;
1234
1235         CDEBUG(level, "Dump object %p extents at %s:%d, mppr: %u.\n",
1236                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1237
1238         /* osc_object_lock(obj); */
1239         cnt = 1;
1240         for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1241                 OSC_EXTENT_DUMP(level, ext, "in tree %d.\n", cnt++);
1242
1243         cnt = 1;
1244         list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1245                 OSC_EXTENT_DUMP(level, ext, "hp %d.\n", cnt++);
1246
1247         cnt = 1;
1248         list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1249                 OSC_EXTENT_DUMP(level, ext, "urgent %d.\n", cnt++);
1250
1251         cnt = 1;
1252         list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1253                 OSC_EXTENT_DUMP(level, ext, "reading %d.\n", cnt++);
1254         /* osc_object_unlock(obj); */
1255 }
1256
1257 /* ------------------ osc extent end ------------------ */
1258
1259 static inline int osc_is_ready(struct osc_object *osc)
1260 {
1261         return !list_empty(&osc->oo_ready_item) ||
1262                !list_empty(&osc->oo_hp_ready_item);
1263 }
1264
1265 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1266         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1267                (OSC), osc_is_ready(OSC),                                       \
1268                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1269                list_empty_marker(&(OSC)->oo_ready_item),                       \
1270                atomic_read(&(OSC)->oo_nr_writes),                              \
1271                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1272                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1273                atomic_read(&(OSC)->oo_nr_reads),                               \
1274                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1275                ##args)
1276
1277 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1278                           int cmd)
1279 {
1280         struct osc_page *opg  = oap2osc_page(oap);
1281         struct cl_page  *page = oap2cl_page(oap);
1282         int result;
1283
1284         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1285
1286         ENTRY;
1287         result = cl_page_make_ready(env, page, CRT_WRITE);
1288         if (result == 0)
1289                 opg->ops_submit_time = ktime_get_seconds();
1290         RETURN(result);
1291 }
1292
1293 static int osc_refresh_count(const struct lu_env *env,
1294                              struct osc_async_page *oap, int cmd)
1295 {
1296         struct osc_page  *opg = oap2osc_page(oap);
1297         pgoff_t index = osc_index(oap2osc(oap));
1298         struct cl_object *obj;
1299         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
1300         int result;
1301         loff_t kms;
1302
1303         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1304         LASSERT(!(cmd & OBD_BRW_READ));
1305         LASSERT(opg != NULL);
1306         obj = opg->ops_cl.cpl_obj;
1307
1308         cl_object_attr_lock(obj);
1309         result = cl_object_attr_get(env, obj, attr);
1310         cl_object_attr_unlock(obj);
1311         if (result < 0)
1312                 return result;
1313         kms = attr->cat_kms;
1314         if (cl_offset(obj, index) >= kms)
1315                 /* catch race with truncate */
1316                 return 0;
1317         else if (cl_offset(obj, index + 1) > kms)
1318                 /* catch sub-page write at end of file */
1319                 return kms % PAGE_SIZE;
1320         else
1321                 return PAGE_SIZE;
1322 }
1323
1324 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1325                           int cmd, int rc)
1326 {
1327         struct osc_page   *opg  = oap2osc_page(oap);
1328         struct cl_page    *page = oap2cl_page(oap);
1329         enum cl_req_type   crt;
1330         int srvlock;
1331
1332         ENTRY;
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         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1342         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1343         opg->ops_transfer_pinned = 0;
1344
1345         opg->ops_submit_time = 0;
1346         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1347
1348         /* statistic */
1349         if (rc == 0 && srvlock) {
1350                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1351                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1352                 size_t bytes = oap->oap_count;
1353
1354                 if (crt == CRT_READ)
1355                         stats->os_lockless_reads += bytes;
1356                 else
1357                         stats->os_lockless_writes += bytes;
1358         }
1359
1360         /*
1361          * This has to be the last operation with the page, as locks are
1362          * released in cl_page_completion() and nothing except for the
1363          * reference counter protects page from concurrent reclaim.
1364          */
1365         lu_ref_del(&page->cp_reference, "transfer", page);
1366
1367         cl_page_completion(env, page, crt, rc);
1368         cl_page_put(env, page);
1369
1370         RETURN(0);
1371 }
1372
1373 #define OSC_DUMP_GRANT(lvl, cli, fmt, args...) do {                     \
1374         struct client_obd *__tmp = (cli);                               \
1375         CDEBUG(lvl, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu "  \
1376                "dropped: %ld avail: %ld, dirty_grant: %ld, "            \
1377                "reserved: %ld, flight: %d } lru {in list: %ld, "        \
1378                "left: %ld, waiters: %d }" fmt "\n",                     \
1379                cli_name(__tmp),                                         \
1380                __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages,        \
1381                atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1382                __tmp->cl_lost_grant, __tmp->cl_avail_grant,             \
1383                __tmp->cl_dirty_grant,                                   \
1384                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,         \
1385                atomic_long_read(&__tmp->cl_lru_in_list),                \
1386                atomic_long_read(&__tmp->cl_lru_busy),                   \
1387                atomic_read(&__tmp->cl_lru_shrinkers), ##args);          \
1388 } while (0)
1389
1390 /* caller must hold loi_list_lock */
1391 static void osc_consume_write_grant(struct client_obd *cli,
1392                                     struct brw_page *pga)
1393 {
1394         assert_spin_locked(&cli->cl_loi_list_lock);
1395         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1396         atomic_long_inc(&obd_dirty_pages);
1397         cli->cl_dirty_pages++;
1398         pga->flag |= OBD_BRW_FROM_GRANT;
1399         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1400                PAGE_SIZE, pga, pga->pg);
1401         osc_update_next_shrink(cli);
1402 }
1403
1404 /* the companion to osc_consume_write_grant, called when a brw has completed.
1405  * must be called with the loi lock held. */
1406 static void osc_release_write_grant(struct client_obd *cli,
1407                                     struct brw_page *pga)
1408 {
1409         ENTRY;
1410
1411         assert_spin_locked(&cli->cl_loi_list_lock);
1412         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1413                 EXIT;
1414                 return;
1415         }
1416
1417         pga->flag &= ~OBD_BRW_FROM_GRANT;
1418         atomic_long_dec(&obd_dirty_pages);
1419         cli->cl_dirty_pages--;
1420         if (pga->flag & OBD_BRW_NOCACHE) {
1421                 pga->flag &= ~OBD_BRW_NOCACHE;
1422                 atomic_long_dec(&obd_dirty_transit_pages);
1423                 cli->cl_dirty_transit--;
1424         }
1425         EXIT;
1426 }
1427
1428 /**
1429  * To avoid sleeping with object lock held, it's good for us allocate enough
1430  * grants before entering into critical section.
1431  *
1432  * client_obd_list_lock held by caller
1433  */
1434 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1435 {
1436         int rc = -EDQUOT;
1437
1438         if (cli->cl_avail_grant >= bytes) {
1439                 cli->cl_avail_grant    -= bytes;
1440                 cli->cl_reserved_grant += bytes;
1441                 rc = 0;
1442         }
1443         return rc;
1444 }
1445
1446 static void __osc_unreserve_grant(struct client_obd *cli,
1447                                   unsigned int reserved, unsigned int unused)
1448 {
1449         /* it's quite normal for us to get more grant than reserved.
1450          * Thinking about a case that two extents merged by adding a new
1451          * chunk, we can save one extent tax. If extent tax is greater than
1452          * one chunk, we can save more grant by adding a new chunk */
1453         cli->cl_reserved_grant -= reserved;
1454         if (unused > reserved) {
1455                 cli->cl_avail_grant += reserved;
1456                 cli->cl_lost_grant  += unused - reserved;
1457                 cli->cl_dirty_grant -= unused - reserved;
1458         } else {
1459                 cli->cl_avail_grant += unused;
1460                 cli->cl_dirty_grant += reserved - unused;
1461         }
1462 }
1463
1464 static void osc_unreserve_grant(struct client_obd *cli,
1465                                 unsigned int reserved, unsigned int unused)
1466 {
1467         spin_lock(&cli->cl_loi_list_lock);
1468         __osc_unreserve_grant(cli, reserved, unused);
1469         if (unused > 0)
1470                 osc_wake_cache_waiters(cli);
1471         spin_unlock(&cli->cl_loi_list_lock);
1472 }
1473
1474 /**
1475  * Free grant after IO is finished or canceled.
1476  *
1477  * @lost_grant is used to remember how many grants we have allocated but not
1478  * used, we should return these grants to OST. There're two cases where grants
1479  * can be lost:
1480  * 1. truncate;
1481  * 2. blocksize at OST is less than PAGE_SIZE and a partial page was
1482  *    written. In this case OST may use less chunks to serve this partial
1483  *    write. OSTs don't actually know the page size on the client side. so
1484  *    clients have to calculate lost grant by the blocksize on the OST.
1485  *    See filter_grant_check() for details.
1486  */
1487 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1488                            unsigned int lost_grant, unsigned int dirty_grant)
1489 {
1490         unsigned long grant;
1491
1492         grant = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
1493
1494         spin_lock(&cli->cl_loi_list_lock);
1495         atomic_long_sub(nr_pages, &obd_dirty_pages);
1496         cli->cl_dirty_pages -= nr_pages;
1497         cli->cl_lost_grant += lost_grant;
1498         cli->cl_dirty_grant -= dirty_grant;
1499         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1500                 /* borrow some grant from truncate to avoid the case that
1501                  * truncate uses up all avail grant */
1502                 cli->cl_lost_grant -= grant;
1503                 cli->cl_avail_grant += grant;
1504         }
1505         osc_wake_cache_waiters(cli);
1506         spin_unlock(&cli->cl_loi_list_lock);
1507         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu/%lu\n",
1508                lost_grant, cli->cl_lost_grant,
1509                cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_SHIFT,
1510                cli->cl_dirty_grant);
1511 }
1512
1513 /**
1514  * The companion to osc_enter_cache(), called when @oap is no longer part of
1515  * the dirty accounting due to error.
1516  */
1517 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1518 {
1519         spin_lock(&cli->cl_loi_list_lock);
1520         osc_release_write_grant(cli, &oap->oap_brw_page);
1521         spin_unlock(&cli->cl_loi_list_lock);
1522 }
1523
1524 /**
1525  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1526  * is available.
1527  */
1528 static int osc_enter_cache_try(struct client_obd *cli,
1529                                struct osc_async_page *oap,
1530                                int bytes, int transient)
1531 {
1532         int rc;
1533
1534         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1535
1536         rc = osc_reserve_grant(cli, bytes);
1537         if (rc < 0)
1538                 return 0;
1539
1540         if (cli->cl_dirty_pages < cli->cl_dirty_max_pages &&
1541             1 + atomic_long_read(&obd_dirty_pages) <= obd_max_dirty_pages) {
1542                 osc_consume_write_grant(cli, &oap->oap_brw_page);
1543                 if (transient) {
1544                         cli->cl_dirty_transit++;
1545                         atomic_long_inc(&obd_dirty_transit_pages);
1546                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
1547                 }
1548                 rc = 1;
1549         } else {
1550                 __osc_unreserve_grant(cli, bytes, bytes);
1551                 rc = 0;
1552         }
1553         return rc;
1554 }
1555
1556 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1557 {
1558         int rc;
1559         spin_lock(&cli->cl_loi_list_lock);
1560         rc = list_empty(&ocw->ocw_entry);
1561         spin_unlock(&cli->cl_loi_list_lock);
1562         return rc;
1563 }
1564
1565 /**
1566  * The main entry to reserve dirty page accounting. Usually the grant reserved
1567  * in this function will be freed in bulk in osc_free_grant() unless it fails
1568  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1569  *
1570  * The process will be put into sleep if it's already run out of grant.
1571  */
1572 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1573                            struct osc_async_page *oap, int bytes)
1574 {
1575         struct osc_object       *osc = oap->oap_obj;
1576         struct lov_oinfo        *loi = osc->oo_oinfo;
1577         struct osc_cache_waiter  ocw;
1578         struct l_wait_info       lwi;
1579         int                      rc = -EDQUOT;
1580         ENTRY;
1581
1582         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(AT_OFF ? obd_timeout : at_max),
1583                                NULL, LWI_ON_SIGNAL_NOOP, NULL);
1584
1585         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1586
1587         spin_lock(&cli->cl_loi_list_lock);
1588
1589         /* force the caller to try sync io.  this can jump the list
1590          * of queued writes and create a discontiguous rpc stream */
1591         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1592             cli->cl_dirty_max_pages == 0 ||
1593             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) {
1594                 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1595                 GOTO(out, rc = -EDQUOT);
1596         }
1597
1598         /* Hopefully normal case - cache space and write credits available */
1599         if (osc_enter_cache_try(cli, oap, bytes, 0)) {
1600                 OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1601                 GOTO(out, rc = 0);
1602         }
1603
1604         /* We can get here for two reasons: too many dirty pages in cache, or
1605          * run out of grants. In both cases we should write dirty pages out.
1606          * Adding a cache waiter will trigger urgent write-out no matter what
1607          * RPC size will be.
1608          * The exiting condition is no avail grants and no dirty pages caching,
1609          * that really means there is no space on the OST. */
1610         init_waitqueue_head(&ocw.ocw_waitq);
1611         ocw.ocw_oap   = oap;
1612         ocw.ocw_grant = bytes;
1613         while (cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0) {
1614                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1615                 ocw.ocw_rc = 0;
1616                 spin_unlock(&cli->cl_loi_list_lock);
1617
1618                 osc_io_unplug_async(env, cli, NULL);
1619
1620                 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1621                        cli_name(cli), &ocw, oap);
1622
1623                 rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1624
1625                 spin_lock(&cli->cl_loi_list_lock);
1626
1627                 if (rc < 0) {
1628                         /* l_wait_event is interrupted by signal or timed out */
1629                         list_del_init(&ocw.ocw_entry);
1630                         break;
1631                 }
1632                 LASSERT(list_empty(&ocw.ocw_entry));
1633                 rc = ocw.ocw_rc;
1634
1635                 if (rc != -EDQUOT)
1636                         break;
1637                 if (osc_enter_cache_try(cli, oap, bytes, 0)) {
1638                         rc = 0;
1639                         break;
1640                 }
1641         }
1642
1643         switch (rc) {
1644         case 0:
1645                 OSC_DUMP_GRANT(D_CACHE, cli, "finally got grant space\n");
1646                 break;
1647         case -ETIMEDOUT:
1648                 OSC_DUMP_GRANT(D_CACHE, cli,
1649                                "timeout, fall back to sync i/o\n");
1650                 osc_extent_tree_dump(D_CACHE, osc);
1651                 /* fall back to synchronous I/O */
1652                 rc = -EDQUOT;
1653                 break;
1654         case -EINTR:
1655                 /* Ensures restartability - LU-3581 */
1656                 OSC_DUMP_GRANT(D_CACHE, cli, "interrupted\n");
1657                 rc = -ERESTARTSYS;
1658                 break;
1659         case -EDQUOT:
1660                 OSC_DUMP_GRANT(D_CACHE, cli,
1661                                "no grant space, fall back to sync i/o\n");
1662                 break;
1663         default:
1664                 CDEBUG(D_CACHE, "%s: event for cache space @ %p never arrived "
1665                        "due to %d, fall back to sync i/o\n",
1666                        cli_name(cli), &ocw, rc);
1667                 break;
1668         }
1669         EXIT;
1670 out:
1671         spin_unlock(&cli->cl_loi_list_lock);
1672         RETURN(rc);
1673 }
1674
1675 /* caller must hold loi_list_lock */
1676 void osc_wake_cache_waiters(struct client_obd *cli)
1677 {
1678         struct list_head *l, *tmp;
1679         struct osc_cache_waiter *ocw;
1680
1681         ENTRY;
1682         list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1683                 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
1684                 list_del_init(&ocw->ocw_entry);
1685
1686                 ocw->ocw_rc = -EDQUOT;
1687                 /* we can't dirty more */
1688                 if ((cli->cl_dirty_pages  >= cli->cl_dirty_max_pages) ||
1689                     (1 + atomic_long_read(&obd_dirty_pages) >
1690                      obd_max_dirty_pages)) {
1691                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
1692                                "osc max %ld, sys max %ld\n",
1693                                cli->cl_dirty_pages, cli->cl_dirty_max_pages,
1694                                obd_max_dirty_pages);
1695                         goto wakeup;
1696                 }
1697
1698                 if (osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
1699                         ocw->ocw_rc = 0;
1700 wakeup:
1701                 CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld, %d\n",
1702                        ocw, ocw->ocw_oap, cli->cl_avail_grant, ocw->ocw_rc);
1703
1704                 wake_up(&ocw->ocw_waitq);
1705         }
1706
1707         EXIT;
1708 }
1709 EXPORT_SYMBOL(osc_wake_cache_waiters);
1710
1711 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1712 {
1713         int hprpc = !!list_empty(&osc->oo_hp_exts);
1714         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1715 }
1716
1717 /* This maintains the lists of pending pages to read/write for a given object
1718  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1719  * to quickly find objects that are ready to send an RPC. */
1720 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1721                          int cmd)
1722 {
1723         int invalid_import = 0;
1724         ENTRY;
1725
1726         /* if we have an invalid import we want to drain the queued pages
1727          * by forcing them through rpcs that immediately fail and complete
1728          * the pages.  recovery relies on this to empty the queued pages
1729          * before canceling the locks and evicting down the llite pages */
1730         if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1731                 invalid_import = 1;
1732
1733         if (cmd & OBD_BRW_WRITE) {
1734                 if (atomic_read(&osc->oo_nr_writes) == 0)
1735                         RETURN(0);
1736                 if (invalid_import) {
1737                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1738                         RETURN(1);
1739                 }
1740                 if (!list_empty(&osc->oo_hp_exts)) {
1741                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1742                         RETURN(1);
1743                 }
1744                 if (!list_empty(&osc->oo_urgent_exts)) {
1745                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1746                         RETURN(1);
1747                 }
1748                 /* trigger a write rpc stream as long as there are dirtiers
1749                  * waiting for space.  as they're waiting, they're not going to
1750                  * create more pages to coalesce with what's waiting.. */
1751                 if (!list_empty(&cli->cl_cache_waiters)) {
1752                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1753                         RETURN(1);
1754                 }
1755                 if (!list_empty(&osc->oo_full_exts)) {
1756                         CDEBUG(D_CACHE, "full extent ready, make an RPC\n");
1757                         RETURN(1);
1758                 }
1759         } else {
1760                 if (atomic_read(&osc->oo_nr_reads) == 0)
1761                         RETURN(0);
1762                 if (invalid_import) {
1763                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1764                         RETURN(1);
1765                 }
1766                 /* all read are urgent. */
1767                 if (!list_empty(&osc->oo_reading_exts))
1768                         RETURN(1);
1769         }
1770
1771         RETURN(0);
1772 }
1773
1774 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1775 {
1776         struct client_obd *cli = osc_cli(obj);
1777         if (cmd & OBD_BRW_WRITE) {
1778                 atomic_add(delta, &obj->oo_nr_writes);
1779                 atomic_add(delta, &cli->cl_pending_w_pages);
1780                 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1781         } else {
1782                 atomic_add(delta, &obj->oo_nr_reads);
1783                 atomic_add(delta, &cli->cl_pending_r_pages);
1784                 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1785         }
1786         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1787 }
1788
1789 static int osc_makes_hprpc(struct osc_object *obj)
1790 {
1791         return !list_empty(&obj->oo_hp_exts);
1792 }
1793
1794 static void on_list(struct list_head *item, struct list_head *list,
1795                     int should_be_on)
1796 {
1797         if (list_empty(item) && should_be_on)
1798                 list_add_tail(item, list);
1799         else if (!list_empty(item) && !should_be_on)
1800                 list_del_init(item);
1801 }
1802
1803 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1804  * can find pages to build into rpcs quickly */
1805 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1806 {
1807         if (osc_makes_hprpc(osc)) {
1808                 /* HP rpc */
1809                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1810                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1811         } else {
1812                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1813                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1814                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1815                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1816         }
1817
1818         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1819                 atomic_read(&osc->oo_nr_writes) > 0);
1820
1821         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1822                 atomic_read(&osc->oo_nr_reads) > 0);
1823
1824         return osc_is_ready(osc);
1825 }
1826
1827 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1828 {
1829         int is_ready;
1830
1831         spin_lock(&cli->cl_loi_list_lock);
1832         is_ready = __osc_list_maint(cli, osc);
1833         spin_unlock(&cli->cl_loi_list_lock);
1834
1835         return is_ready;
1836 }
1837
1838 /* this is trying to propogate async writeback errors back up to the
1839  * application.  As an async write fails we record the error code for later if
1840  * the app does an fsync.  As long as errors persist we force future rpcs to be
1841  * sync so that the app can get a sync error and break the cycle of queueing
1842  * pages for which writeback will fail. */
1843 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1844                            int rc)
1845 {
1846         if (rc) {
1847                 if (!ar->ar_rc)
1848                         ar->ar_rc = rc;
1849
1850                 ar->ar_force_sync = 1;
1851                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1852                 return;
1853
1854         }
1855
1856         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1857                 ar->ar_force_sync = 0;
1858 }
1859
1860 /* this must be called holding the loi list lock to give coverage to exit_cache,
1861  * async_flag maintenance, and oap_request */
1862 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1863                               struct osc_async_page *oap, int sent, int rc)
1864 {
1865         struct osc_object *osc = oap->oap_obj;
1866         struct lov_oinfo  *loi = osc->oo_oinfo;
1867         __u64 xid = 0;
1868
1869         ENTRY;
1870         if (oap->oap_request != NULL) {
1871                 xid = ptlrpc_req_xid(oap->oap_request);
1872                 ptlrpc_req_finished(oap->oap_request);
1873                 oap->oap_request = NULL;
1874         }
1875
1876         /* As the transfer for this page is being done, clear the flags */
1877         spin_lock(&oap->oap_lock);
1878         oap->oap_async_flags = 0;
1879         spin_unlock(&oap->oap_lock);
1880         oap->oap_interrupted = 0;
1881
1882         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1883                 spin_lock(&cli->cl_loi_list_lock);
1884                 osc_process_ar(&cli->cl_ar, xid, rc);
1885                 osc_process_ar(&loi->loi_ar, xid, rc);
1886                 spin_unlock(&cli->cl_loi_list_lock);
1887         }
1888
1889         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1890         if (rc)
1891                 CERROR("completion on oap %p obj %p returns %d.\n",
1892                        oap, osc, rc);
1893
1894         EXIT;
1895 }
1896
1897 struct extent_rpc_data {
1898         struct list_head        *erd_rpc_list;
1899         unsigned int            erd_page_count;
1900         unsigned int            erd_max_pages;
1901         unsigned int            erd_max_chunks;
1902         unsigned int            erd_max_extents;
1903 };
1904
1905 static inline unsigned osc_extent_chunks(const struct osc_extent *ext)
1906 {
1907         struct client_obd *cli = osc_cli(ext->oe_obj);
1908         unsigned ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1909
1910         return (ext->oe_end >> ppc_bits) - (ext->oe_start >> ppc_bits) + 1;
1911 }
1912
1913 /**
1914  * Try to add extent to one RPC. We need to think about the following things:
1915  * - # of pages must not be over max_pages_per_rpc
1916  * - extent must be compatible with previous ones
1917  */
1918 static int try_to_add_extent_for_io(struct client_obd *cli,
1919                                     struct osc_extent *ext,
1920                                     struct extent_rpc_data *data)
1921 {
1922         struct osc_extent *tmp;
1923         unsigned int chunk_count;
1924         struct osc_async_page *oap = list_first_entry(&ext->oe_pages,
1925                                                       struct osc_async_page,
1926                                                       oap_pending_item);
1927         ENTRY;
1928
1929         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1930                 ext);
1931         OSC_EXTENT_DUMP(D_CACHE, ext, "trying to add this extent\n");
1932
1933         if (data->erd_max_extents == 0)
1934                 RETURN(0);
1935
1936         chunk_count = osc_extent_chunks(ext);
1937         EASSERTF(data->erd_page_count != 0 ||
1938                  chunk_count <= data->erd_max_chunks, ext,
1939                  "The first extent to be fit in a RPC contains %u chunks, "
1940                  "which is over the limit %u.\n", chunk_count,
1941                  data->erd_max_chunks);
1942         if (chunk_count > data->erd_max_chunks)
1943                 RETURN(0);
1944
1945         data->erd_max_pages = max(ext->oe_mppr, data->erd_max_pages);
1946         EASSERTF(data->erd_page_count != 0 ||
1947                 ext->oe_nr_pages <= data->erd_max_pages, ext,
1948                 "The first extent to be fit in a RPC contains %u pages, "
1949                 "which is over the limit %u.\n", ext->oe_nr_pages,
1950                 data->erd_max_pages);
1951         if (data->erd_page_count + ext->oe_nr_pages > data->erd_max_pages)
1952                 RETURN(0);
1953
1954         list_for_each_entry(tmp, data->erd_rpc_list, oe_link) {
1955                 struct osc_async_page *oap2;
1956                 oap2 = list_first_entry(&tmp->oe_pages, struct osc_async_page,
1957                                         oap_pending_item);
1958                 EASSERT(tmp->oe_owner == current, tmp);
1959 #if 0
1960                 if (overlapped(tmp, ext)) {
1961                         OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
1962                         EASSERT(0, ext);
1963                 }
1964 #endif
1965                 if (oap2cl_page(oap)->cp_type != oap2cl_page(oap2)->cp_type) {
1966                         CDEBUG(D_CACHE, "Do not permit different types of IO "
1967                                "in one RPC\n");
1968                         RETURN(0);
1969                 }
1970
1971                 if (tmp->oe_srvlock != ext->oe_srvlock ||
1972                     !tmp->oe_grants != !ext->oe_grants ||
1973                     tmp->oe_ndelay != ext->oe_ndelay ||
1974                     tmp->oe_no_merge || ext->oe_no_merge)
1975                         RETURN(0);
1976
1977                 /* remove break for strict check */
1978                 break;
1979         }
1980
1981         data->erd_max_extents--;
1982         data->erd_max_chunks -= chunk_count;
1983         data->erd_page_count += ext->oe_nr_pages;
1984         list_move_tail(&ext->oe_link, data->erd_rpc_list);
1985         ext->oe_owner = current;
1986         RETURN(1);
1987 }
1988
1989 static inline unsigned osc_max_write_chunks(const struct client_obd *cli)
1990 {
1991         /*
1992          * LU-8135:
1993          *
1994          * The maximum size of a single transaction is about 64MB in ZFS.
1995          * #define DMU_MAX_ACCESS (64 * 1024 * 1024)
1996          *
1997          * Since ZFS is a copy-on-write file system, a single dirty page in
1998          * a chunk will result in the rewrite of the whole chunk, therefore
1999          * an RPC shouldn't be allowed to contain too many chunks otherwise
2000          * it will make transaction size much bigger than 64MB, especially
2001          * with big block size for ZFS.
2002          *
2003          * This piece of code is to make sure that OSC won't send write RPCs
2004          * with too many chunks. The maximum chunk size that an RPC can cover
2005          * is set to PTLRPC_MAX_BRW_SIZE, which is defined to 16MB. Ideally
2006          * OST should tell the client what the biggest transaction size is,
2007          * but it's good enough for now.
2008          *
2009          * This limitation doesn't apply to ldiskfs, which allows as many
2010          * chunks in one RPC as we want. However, it won't have any benefits
2011          * to have too many discontiguous pages in one RPC.
2012          *
2013          * An osc_extent won't cover over a RPC size, so the chunks in an
2014          * osc_extent won't bigger than PTLRPC_MAX_BRW_SIZE >> chunkbits.
2015          */
2016         return PTLRPC_MAX_BRW_SIZE >> cli->cl_chunkbits;
2017 }
2018
2019 /**
2020  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
2021  * get_write_extent() takes all appropriate extents in atomic.
2022  *
2023  * The following policy is used to collect extents for IO:
2024  * 1. Add as many HP extents as possible;
2025  * 2. Add the first urgent extent in urgent extent list and take it out of
2026  *    urgent list;
2027  * 3. Add subsequent extents of this urgent extent;
2028  * 4. If urgent list is not empty, goto 2;
2029  * 5. Traverse the extent tree from the 1st extent;
2030  * 6. Above steps exit if there is no space in this RPC.
2031  */
2032 static unsigned int get_write_extents(struct osc_object *obj,
2033                                       struct list_head *rpclist)
2034 {
2035         struct client_obd *cli = osc_cli(obj);
2036         struct osc_extent *ext;
2037         struct extent_rpc_data data = {
2038                 .erd_rpc_list   = rpclist,
2039                 .erd_page_count = 0,
2040                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
2041                 .erd_max_chunks = osc_max_write_chunks(cli),
2042                 .erd_max_extents = 256,
2043         };
2044
2045         LASSERT(osc_object_is_locked(obj));
2046         while (!list_empty(&obj->oo_hp_exts)) {
2047                 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
2048                                  oe_link);
2049                 LASSERT(ext->oe_state == OES_CACHE);
2050                 if (!try_to_add_extent_for_io(cli, ext, &data))
2051                         return data.erd_page_count;
2052                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2053         }
2054         if (data.erd_page_count == data.erd_max_pages)
2055                 return data.erd_page_count;
2056
2057         while (!list_empty(&obj->oo_urgent_exts)) {
2058                 ext = list_entry(obj->oo_urgent_exts.next,
2059                                  struct osc_extent, oe_link);
2060                 if (!try_to_add_extent_for_io(cli, ext, &data))
2061                         return data.erd_page_count;
2062         }
2063         if (data.erd_page_count == data.erd_max_pages)
2064                 return data.erd_page_count;
2065
2066         /* One key difference between full extents and other extents: full
2067          * extents can usually only be added if the rpclist was empty, so if we
2068          * can't add one, we continue on to trying to add normal extents.  This
2069          * is so we don't miss adding extra extents to an RPC containing high
2070          * priority or urgent extents. */
2071         while (!list_empty(&obj->oo_full_exts)) {
2072                 ext = list_entry(obj->oo_full_exts.next,
2073                                  struct osc_extent, oe_link);
2074                 if (!try_to_add_extent_for_io(cli, ext, &data))
2075                         break;
2076         }
2077         if (data.erd_page_count == data.erd_max_pages)
2078                 return data.erd_page_count;
2079
2080         ext = first_extent(obj);
2081         while (ext != NULL) {
2082                 if ((ext->oe_state != OES_CACHE) ||
2083                     /* this extent may be already in current rpclist */
2084                     (!list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
2085                         ext = next_extent(ext);
2086                         continue;
2087                 }
2088
2089                 if (!try_to_add_extent_for_io(cli, ext, &data))
2090                         return data.erd_page_count;
2091
2092                 ext = next_extent(ext);
2093         }
2094         return data.erd_page_count;
2095 }
2096
2097 static int
2098 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
2099                    struct osc_object *osc)
2100 __must_hold(osc)
2101 {
2102         struct list_head   rpclist = LIST_HEAD_INIT(rpclist);
2103         struct osc_extent *ext;
2104         struct osc_extent *tmp;
2105         struct osc_extent *first = NULL;
2106         unsigned int page_count = 0;
2107         int srvlock = 0;
2108         int rc = 0;
2109         ENTRY;
2110
2111         LASSERT(osc_object_is_locked(osc));
2112
2113         page_count = get_write_extents(osc, &rpclist);
2114         LASSERT(equi(page_count == 0, list_empty(&rpclist)));
2115
2116         if (list_empty(&rpclist))
2117                 RETURN(0);
2118
2119         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2120
2121         list_for_each_entry(ext, &rpclist, oe_link) {
2122                 LASSERT(ext->oe_state == OES_CACHE ||
2123                         ext->oe_state == OES_LOCK_DONE);
2124                 if (ext->oe_state == OES_CACHE)
2125                         osc_extent_state_set(ext, OES_LOCKING);
2126                 else
2127                         osc_extent_state_set(ext, OES_RPC);
2128         }
2129
2130         /* we're going to grab page lock, so release object lock because
2131          * lock order is page lock -> object lock. */
2132         osc_object_unlock(osc);
2133
2134         list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2135                 if (ext->oe_state == OES_LOCKING) {
2136                         rc = osc_extent_make_ready(env, ext);
2137                         if (unlikely(rc < 0)) {
2138                                 list_del_init(&ext->oe_link);
2139                                 osc_extent_finish(env, ext, 0, rc);
2140                                 continue;
2141                         }
2142                 }
2143                 if (first == NULL) {
2144                         first = ext;
2145                         srvlock = ext->oe_srvlock;
2146                 } else {
2147                         LASSERT(srvlock == ext->oe_srvlock);
2148                 }
2149         }
2150
2151         if (!list_empty(&rpclist)) {
2152                 LASSERT(page_count > 0);
2153                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2154                 LASSERT(list_empty(&rpclist));
2155         }
2156
2157         osc_object_lock(osc);
2158         RETURN(rc);
2159 }
2160
2161 /**
2162  * prepare pages for ASYNC io and put pages in send queue.
2163  *
2164  * \param cmd OBD_BRW_* macroses
2165  * \param lop pending pages
2166  *
2167  * \return zero if no page added to send queue.
2168  * \return 1 if pages successfully added to send queue.
2169  * \return negative on errors.
2170  */
2171 static int
2172 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2173                   struct osc_object *osc)
2174 __must_hold(osc)
2175 {
2176         struct osc_extent *ext;
2177         struct osc_extent *next;
2178         struct list_head rpclist = LIST_HEAD_INIT(rpclist);
2179         struct extent_rpc_data data = {
2180                 .erd_rpc_list   = &rpclist,
2181                 .erd_page_count = 0,
2182                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
2183                 .erd_max_chunks = UINT_MAX,
2184                 .erd_max_extents = UINT_MAX,
2185         };
2186         int rc = 0;
2187         ENTRY;
2188
2189         LASSERT(osc_object_is_locked(osc));
2190         list_for_each_entry_safe(ext, next, &osc->oo_reading_exts, oe_link) {
2191                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2192                 if (!try_to_add_extent_for_io(cli, ext, &data))
2193                         break;
2194                 osc_extent_state_set(ext, OES_RPC);
2195                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2196         }
2197         LASSERT(data.erd_page_count <= data.erd_max_pages);
2198
2199         osc_update_pending(osc, OBD_BRW_READ, -data.erd_page_count);
2200
2201         if (!list_empty(&rpclist)) {
2202                 osc_object_unlock(osc);
2203
2204                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2205                 LASSERT(list_empty(&rpclist));
2206
2207                 osc_object_lock(osc);
2208         }
2209         RETURN(rc);
2210 }
2211
2212 #define list_to_obj(list, item) ({                                            \
2213         struct list_head *__tmp = (list)->next;                               \
2214         list_del_init(__tmp);                                         \
2215         list_entry(__tmp, struct osc_object, oo_##item);                      \
2216 })
2217
2218 /* This is called by osc_check_rpcs() to find which objects have pages that
2219  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2220 static struct osc_object *osc_next_obj(struct client_obd *cli)
2221 {
2222         ENTRY;
2223
2224         /* First return objects that have blocked locks so that they
2225          * will be flushed quickly and other clients can get the lock,
2226          * then objects which have pages ready to be stuffed into RPCs */
2227         if (!list_empty(&cli->cl_loi_hp_ready_list))
2228                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2229         if (!list_empty(&cli->cl_loi_ready_list))
2230                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2231
2232         /* then if we have cache waiters, return all objects with queued
2233          * writes.  This is especially important when many small files
2234          * have filled up the cache and not been fired into rpcs because
2235          * they don't pass the nr_pending/object threshhold */
2236         if (!list_empty(&cli->cl_cache_waiters) &&
2237             !list_empty(&cli->cl_loi_write_list))
2238                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2239
2240         /* then return all queued objects when we have an invalid import
2241          * so that they get flushed */
2242         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2243                 if (!list_empty(&cli->cl_loi_write_list))
2244                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2245                                            write_item));
2246                 if (!list_empty(&cli->cl_loi_read_list))
2247                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2248                                            read_item));
2249         }
2250         RETURN(NULL);
2251 }
2252
2253 /* called with the loi list lock held */
2254 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2255 __must_hold(&cli->cl_loi_list_lock)
2256 {
2257         struct osc_object *osc;
2258         int rc = 0;
2259         ENTRY;
2260
2261         while ((osc = osc_next_obj(cli)) != NULL) {
2262                 struct cl_object *obj = osc2cl(osc);
2263                 struct lu_ref_link link;
2264
2265                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2266
2267                 if (osc_max_rpc_in_flight(cli, osc)) {
2268                         __osc_list_maint(cli, osc);
2269                         break;
2270                 }
2271
2272                 cl_object_get(obj);
2273                 spin_unlock(&cli->cl_loi_list_lock);
2274                 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2275
2276                 /* attempt some read/write balancing by alternating between
2277                  * reads and writes in an object.  The makes_rpc checks here
2278                  * would be redundant if we were getting read/write work items
2279                  * instead of objects.  we don't want send_oap_rpc to drain a
2280                  * partial read pending queue when we're given this object to
2281                  * do io on writes while there are cache waiters */
2282                 osc_object_lock(osc);
2283                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2284                         rc = osc_send_write_rpc(env, cli, osc);
2285                         if (rc < 0) {
2286                                 CERROR("Write request failed with %d\n", rc);
2287
2288                                 /* osc_send_write_rpc failed, mostly because of
2289                                  * memory pressure.
2290                                  *
2291                                  * It can't break here, because if:
2292                                  *  - a page was submitted by osc_io_submit, so
2293                                  *    page locked;
2294                                  *  - no request in flight
2295                                  *  - no subsequent request
2296                                  * The system will be in live-lock state,
2297                                  * because there is no chance to call
2298                                  * osc_io_unplug() and osc_check_rpcs() any
2299                                  * more. pdflush can't help in this case,
2300                                  * because it might be blocked at grabbing
2301                                  * the page lock as we mentioned.
2302                                  *
2303                                  * Anyway, continue to drain pages. */
2304                                 /* break; */
2305                         }
2306                 }
2307                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2308                         rc = osc_send_read_rpc(env, cli, osc);
2309                         if (rc < 0)
2310                                 CERROR("Read request failed with %d\n", rc);
2311                 }
2312                 osc_object_unlock(osc);
2313
2314                 osc_list_maint(cli, osc);
2315                 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2316                 cl_object_put(env, obj);
2317
2318                 spin_lock(&cli->cl_loi_list_lock);
2319         }
2320 }
2321
2322 int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2323                    struct osc_object *osc, int async)
2324 {
2325         int rc = 0;
2326
2327         if (osc != NULL && osc_list_maint(cli, osc) == 0)
2328                 return 0;
2329
2330         if (!async) {
2331                 spin_lock(&cli->cl_loi_list_lock);
2332                 osc_check_rpcs(env, cli);
2333                 spin_unlock(&cli->cl_loi_list_lock);
2334         } else {
2335                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2336                 LASSERT(cli->cl_writeback_work != NULL);
2337                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2338         }
2339         return rc;
2340 }
2341 EXPORT_SYMBOL(osc_io_unplug0);
2342
2343 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2344                         struct page *page, loff_t offset)
2345 {
2346         struct obd_export     *exp = osc_export(osc);
2347         struct osc_async_page *oap = &ops->ops_oap;
2348         ENTRY;
2349
2350         if (!page)
2351                 return cfs_size_round(sizeof(*oap));
2352
2353         oap->oap_magic = OAP_MAGIC;
2354         oap->oap_cli = &exp->exp_obd->u.cli;
2355         oap->oap_obj = osc;
2356
2357         oap->oap_page = page;
2358         oap->oap_obj_off = offset;
2359         LASSERT(!(offset & ~PAGE_MASK));
2360
2361         if (cfs_capable(CFS_CAP_SYS_RESOURCE))
2362                 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2363
2364         INIT_LIST_HEAD(&oap->oap_pending_item);
2365         INIT_LIST_HEAD(&oap->oap_rpc_item);
2366
2367         spin_lock_init(&oap->oap_lock);
2368         CDEBUG(D_INFO, "oap %p page %p obj off %llu\n",
2369                oap, page, oap->oap_obj_off);
2370         RETURN(0);
2371 }
2372 EXPORT_SYMBOL(osc_prep_async_page);
2373
2374 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2375                        struct osc_page *ops)
2376 {
2377         struct osc_io *oio = osc_env_io(env);
2378         struct osc_extent     *ext = NULL;
2379         struct osc_async_page *oap = &ops->ops_oap;
2380         struct client_obd     *cli = oap->oap_cli;
2381         struct osc_object     *osc = oap->oap_obj;
2382         pgoff_t index;
2383         unsigned int tmp;
2384         unsigned int grants = 0;
2385         u32    brw_flags = OBD_BRW_ASYNC;
2386         int    cmd = OBD_BRW_WRITE;
2387         int    need_release = 0;
2388         int    rc = 0;
2389         ENTRY;
2390
2391         if (oap->oap_magic != OAP_MAGIC)
2392                 RETURN(-EINVAL);
2393
2394         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2395                 RETURN(-EIO);
2396
2397         if (!list_empty(&oap->oap_pending_item) ||
2398             !list_empty(&oap->oap_rpc_item))
2399                 RETURN(-EBUSY);
2400
2401         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2402         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2403         if (cfs_capable(CFS_CAP_SYS_RESOURCE)) {
2404                 brw_flags |= OBD_BRW_NOQUOTA;
2405                 cmd |= OBD_BRW_NOQUOTA;
2406         }
2407
2408         /* check if the file's owner/group is over quota */
2409         if (!(cmd & OBD_BRW_NOQUOTA)) {
2410                 struct cl_object *obj;
2411                 struct cl_attr   *attr;
2412                 unsigned int qid[LL_MAXQUOTAS];
2413
2414                 obj = cl_object_top(&osc->oo_cl);
2415                 attr = &osc_env_info(env)->oti_attr;
2416
2417                 cl_object_attr_lock(obj);
2418                 rc = cl_object_attr_get(env, obj, attr);
2419                 cl_object_attr_unlock(obj);
2420
2421                 qid[USRQUOTA] = attr->cat_uid;
2422                 qid[GRPQUOTA] = attr->cat_gid;
2423                 qid[PRJQUOTA] = attr->cat_projid;
2424                 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2425                         rc = -EDQUOT;
2426                 if (rc)
2427                         RETURN(rc);
2428         }
2429
2430         oap->oap_cmd = cmd;
2431         oap->oap_page_off = ops->ops_from;
2432         oap->oap_count = ops->ops_to - ops->ops_from;
2433         /* No need to hold a lock here,
2434          * since this page is not in any list yet. */
2435         oap->oap_async_flags = 0;
2436         oap->oap_brw_flags = brw_flags;
2437
2438         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2439                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2440
2441         index = osc_index(oap2osc(oap));
2442
2443         /* Add this page into extent by the following steps:
2444          * 1. if there exists an active extent for this IO, mostly this page
2445          *    can be added to the active extent and sometimes we need to
2446          *    expand extent to accomodate this page;
2447          * 2. otherwise, a new extent will be allocated. */
2448
2449         ext = oio->oi_active;
2450         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2451                 /* one chunk plus extent overhead must be enough to write this
2452                  * page */
2453                 grants = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2454                 if (ext->oe_end >= index)
2455                         grants = 0;
2456
2457                 /* it doesn't need any grant to dirty this page */
2458                 spin_lock(&cli->cl_loi_list_lock);
2459                 rc = osc_enter_cache_try(cli, oap, grants, 0);
2460                 spin_unlock(&cli->cl_loi_list_lock);
2461                 if (rc == 0) { /* try failed */
2462                         grants = 0;
2463                         need_release = 1;
2464                 } else if (ext->oe_end < index) {
2465                         tmp = grants;
2466                         /* try to expand this extent */
2467                         rc = osc_extent_expand(ext, index, &tmp);
2468                         if (rc < 0) {
2469                                 need_release = 1;
2470                                 /* don't free reserved grant */
2471                         } else {
2472                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2473                                                 "expanded for %lu.\n", index);
2474                                 osc_unreserve_grant(cli, grants, tmp);
2475                                 grants = 0;
2476                         }
2477                 }
2478                 rc = 0;
2479         } else if (ext != NULL) {
2480                 /* index is located outside of active extent */
2481                 need_release = 1;
2482         }
2483         if (need_release) {
2484                 osc_extent_release(env, ext);
2485                 oio->oi_active = NULL;
2486                 ext = NULL;
2487         }
2488
2489         if (ext == NULL) {
2490                 tmp = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2491
2492                 /* try to find new extent to cover this page */
2493                 LASSERT(oio->oi_active == NULL);
2494                 /* we may have allocated grant for this page if we failed
2495                  * to expand the previous active extent. */
2496                 LASSERT(ergo(grants > 0, grants >= tmp));
2497
2498                 rc = 0;
2499                 if (grants == 0) {
2500                         /* we haven't allocated grant for this page. */
2501                         rc = osc_enter_cache(env, cli, oap, tmp);
2502                         if (rc == 0)
2503                                 grants = tmp;
2504                 }
2505
2506                 tmp = grants;
2507                 if (rc == 0) {
2508                         ext = osc_extent_find(env, osc, index, &tmp);
2509                         if (IS_ERR(ext)) {
2510                                 LASSERT(tmp == grants);
2511                                 osc_exit_cache(cli, oap);
2512                                 rc = PTR_ERR(ext);
2513                                 ext = NULL;
2514                         } else {
2515                                 oio->oi_active = ext;
2516                         }
2517                 }
2518                 if (grants > 0)
2519                         osc_unreserve_grant(cli, grants, tmp);
2520         }
2521
2522         LASSERT(ergo(rc == 0, ext != NULL));
2523         if (ext != NULL) {
2524                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2525                          ext, "index = %lu.\n", index);
2526                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2527
2528                 osc_object_lock(osc);
2529                 if (ext->oe_nr_pages == 0)
2530                         ext->oe_srvlock = ops->ops_srvlock;
2531                 else
2532                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2533                 ++ext->oe_nr_pages;
2534                 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2535                 osc_object_unlock(osc);
2536
2537                 if (!ext->oe_layout_version)
2538                         ext->oe_layout_version = io->ci_layout_version;
2539         }
2540
2541         RETURN(rc);
2542 }
2543
2544 int osc_teardown_async_page(const struct lu_env *env,
2545                             struct osc_object *obj, struct osc_page *ops)
2546 {
2547         struct osc_async_page *oap = &ops->ops_oap;
2548         int rc = 0;
2549         ENTRY;
2550
2551         LASSERT(oap->oap_magic == OAP_MAGIC);
2552
2553         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2554                oap, ops, osc_index(oap2osc(oap)));
2555
2556         if (!list_empty(&oap->oap_rpc_item)) {
2557                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2558                 rc = -EBUSY;
2559         } else if (!list_empty(&oap->oap_pending_item)) {
2560                 struct osc_extent *ext = NULL;
2561
2562                 osc_object_lock(obj);
2563                 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2564                 osc_object_unlock(obj);
2565                 /* only truncated pages are allowed to be taken out.
2566                  * See osc_extent_truncate() and osc_cache_truncate_start()
2567                  * for details. */
2568                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2569                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2570                                         osc_index(oap2osc(oap)));
2571                         rc = -EBUSY;
2572                 }
2573                 if (ext != NULL)
2574                         osc_extent_put(env, ext);
2575         }
2576         RETURN(rc);
2577 }
2578
2579 /**
2580  * This is called when a page is picked up by kernel to write out.
2581  *
2582  * We should find out the corresponding extent and add the whole extent
2583  * into urgent list. The extent may be being truncated or used, handle it
2584  * carefully.
2585  */
2586 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2587                          struct osc_page *ops)
2588 {
2589         struct osc_extent *ext   = NULL;
2590         struct osc_object *obj   = cl2osc(ops->ops_cl.cpl_obj);
2591         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2592         pgoff_t            index = osc_index(ops);
2593         struct osc_async_page *oap = &ops->ops_oap;
2594         bool unplug = false;
2595         int rc = 0;
2596         ENTRY;
2597
2598         osc_object_lock(obj);
2599         ext = osc_extent_lookup(obj, index);
2600         if (ext == NULL) {
2601                 osc_extent_tree_dump(D_ERROR, obj);
2602                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2603         }
2604
2605         switch (ext->oe_state) {
2606         case OES_RPC:
2607         case OES_LOCK_DONE:
2608                 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2609                 LASSERT(0);
2610                 break;
2611         case OES_LOCKING:
2612                 /* If we know this extent is being written out, we should abort
2613                  * so that the writer can make this page ready. Otherwise, there
2614                  * exists a deadlock problem because other process can wait for
2615                  * page writeback bit holding page lock; and meanwhile in
2616                  * vvp_page_make_ready(), we need to grab page lock before
2617                  * really sending the RPC. */
2618         case OES_TRUNC:
2619                 /* race with truncate, page will be redirtied */
2620         case OES_ACTIVE:
2621                 /* The extent is active so we need to abort and let the caller
2622                  * re-dirty the page. If we continued on here, and we were the
2623                  * one making the extent active, we could deadlock waiting for
2624                  * the page writeback to clear but it won't because the extent
2625                  * is active and won't be written out. */
2626                 GOTO(out, rc = -EAGAIN);
2627         default:
2628                 break;
2629         }
2630
2631         rc = cl_page_prep(env, io, cp, CRT_WRITE);
2632         if (rc)
2633                 GOTO(out, rc);
2634
2635         spin_lock(&oap->oap_lock);
2636         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2637         spin_unlock(&oap->oap_lock);
2638
2639         if (memory_pressure_get())
2640                 ext->oe_memalloc = 1;
2641
2642         ext->oe_urgent = 1;
2643         if (ext->oe_state == OES_CACHE) {
2644                 OSC_EXTENT_DUMP(D_CACHE, ext,
2645                                 "flush page %p make it urgent.\n", oap);
2646                 if (list_empty(&ext->oe_link))
2647                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2648                 unplug = true;
2649         }
2650         rc = 0;
2651         EXIT;
2652
2653 out:
2654         osc_object_unlock(obj);
2655         osc_extent_put(env, ext);
2656         if (unplug)
2657                 osc_io_unplug_async(env, osc_cli(obj), obj);
2658         return rc;
2659 }
2660
2661 /**
2662  * this is called when a sync waiter receives an interruption.  Its job is to
2663  * get the caller woken as soon as possible.  If its page hasn't been put in an
2664  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2665  * desiring interruption which will forcefully complete the rpc once the rpc
2666  * has timed out.
2667  */
2668 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2669 {
2670         struct osc_async_page *oap = &ops->ops_oap;
2671         struct osc_object     *obj = oap->oap_obj;
2672         struct client_obd     *cli = osc_cli(obj);
2673         struct osc_extent     *ext;
2674         struct osc_extent     *found = NULL;
2675         struct list_head            *plist;
2676         pgoff_t index = osc_index(ops);
2677         int     rc = -EBUSY;
2678         int     cmd;
2679         ENTRY;
2680
2681         LASSERT(!oap->oap_interrupted);
2682         oap->oap_interrupted = 1;
2683
2684         /* Find out the caching extent */
2685         osc_object_lock(obj);
2686         if (oap->oap_cmd & OBD_BRW_WRITE) {
2687                 plist = &obj->oo_urgent_exts;
2688                 cmd   = OBD_BRW_WRITE;
2689         } else {
2690                 plist = &obj->oo_reading_exts;
2691                 cmd   = OBD_BRW_READ;
2692         }
2693         list_for_each_entry(ext, plist, oe_link) {
2694                 if (ext->oe_start <= index && ext->oe_end >= index) {
2695                         LASSERT(ext->oe_state == OES_LOCK_DONE);
2696                         /* For OES_LOCK_DONE state extent, it has already held
2697                          * a refcount for RPC. */
2698                         found = osc_extent_get(ext);
2699                         break;
2700                 }
2701         }
2702         if (found != NULL) {
2703                 list_del_init(&found->oe_link);
2704                 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2705                 osc_object_unlock(obj);
2706
2707                 osc_extent_finish(env, found, 0, -EINTR);
2708                 osc_extent_put(env, found);
2709                 rc = 0;
2710         } else {
2711                 osc_object_unlock(obj);
2712                 /* ok, it's been put in an rpc. only one oap gets a request
2713                  * reference */
2714                 if (oap->oap_request != NULL) {
2715                         ptlrpc_mark_interrupted(oap->oap_request);
2716                         ptlrpcd_wake(oap->oap_request);
2717                         ptlrpc_req_finished(oap->oap_request);
2718                         oap->oap_request = NULL;
2719                 }
2720         }
2721
2722         osc_list_maint(cli, obj);
2723         RETURN(rc);
2724 }
2725
2726 int osc_queue_sync_pages(const struct lu_env *env, const struct cl_io *io,
2727                          struct osc_object *obj, struct list_head *list,
2728                          int brw_flags)
2729 {
2730         struct client_obd     *cli = osc_cli(obj);
2731         struct osc_extent     *ext;
2732         struct osc_async_page *oap;
2733         int     page_count = 0;
2734         int     mppr       = cli->cl_max_pages_per_rpc;
2735         bool    can_merge   = true;
2736         pgoff_t start      = CL_PAGE_EOF;
2737         pgoff_t end        = 0;
2738         ENTRY;
2739
2740         list_for_each_entry(oap, list, oap_pending_item) {
2741                 struct osc_page *opg = oap2osc_page(oap);
2742                 pgoff_t index = osc_index(opg);
2743
2744                 if (index > end)
2745                         end = index;
2746                 if (index < start)
2747                         start = index;
2748                 ++page_count;
2749                 mppr <<= (page_count > mppr);
2750
2751                 if (unlikely(opg->ops_from > 0 || opg->ops_to < PAGE_SIZE))
2752                         can_merge = false;
2753         }
2754
2755         ext = osc_extent_alloc(obj);
2756         if (ext == NULL) {
2757                 struct osc_async_page *tmp;
2758
2759                 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2760                         list_del_init(&oap->oap_pending_item);
2761                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2762                 }
2763                 RETURN(-ENOMEM);
2764         }
2765
2766         ext->oe_rw = !!(brw_flags & OBD_BRW_READ);
2767         ext->oe_sync = 1;
2768         ext->oe_no_merge = !can_merge;
2769         ext->oe_urgent = 1;
2770         ext->oe_start = start;
2771         ext->oe_end = ext->oe_max_end = end;
2772         ext->oe_obj = obj;
2773         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2774         ext->oe_ndelay = !!(brw_flags & OBD_BRW_NDELAY);
2775         ext->oe_nr_pages = page_count;
2776         ext->oe_mppr = mppr;
2777         list_splice_init(list, &ext->oe_pages);
2778         ext->oe_layout_version = io->ci_layout_version;
2779
2780         osc_object_lock(obj);
2781         /* Reuse the initial refcount for RPC, don't drop it */
2782         osc_extent_state_set(ext, OES_LOCK_DONE);
2783         if (!ext->oe_rw) { /* write */
2784                 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2785                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2786         } else {
2787                 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2788                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2789         }
2790         osc_object_unlock(obj);
2791
2792         osc_io_unplug_async(env, cli, obj);
2793         RETURN(0);
2794 }
2795
2796 /**
2797  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2798  */
2799 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
2800                              __u64 size, struct osc_extent **extp)
2801 {
2802         struct client_obd *cli = osc_cli(obj);
2803         struct osc_extent *ext;
2804         struct osc_extent *waiting = NULL;
2805         pgoff_t index;
2806         struct list_head list = LIST_HEAD_INIT(list);
2807         int result = 0;
2808         bool partial;
2809         ENTRY;
2810
2811         /* pages with index greater or equal to index will be truncated. */
2812         index = cl_index(osc2cl(obj), size);
2813         partial = size > cl_offset(osc2cl(obj), index);
2814
2815 again:
2816         osc_object_lock(obj);
2817         ext = osc_extent_search(obj, index);
2818         if (ext == NULL)
2819                 ext = first_extent(obj);
2820         else if (ext->oe_end < index)
2821                 ext = next_extent(ext);
2822         while (ext != NULL) {
2823                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2824
2825                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2826                         /* if ext is in urgent state, it means there must exist
2827                          * a page already having been flushed by write_page().
2828                          * We have to wait for this extent because we can't
2829                          * truncate that page. */
2830                         OSC_EXTENT_DUMP(D_CACHE, ext,
2831                                         "waiting for busy extent\n");
2832                         waiting = osc_extent_get(ext);
2833                         break;
2834                 }
2835
2836                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:%llu.\n", size);
2837
2838                 osc_extent_get(ext);
2839                 if (ext->oe_state == OES_ACTIVE) {
2840                         /* though we grab inode mutex for write path, but we
2841                          * release it before releasing extent(in osc_io_end()),
2842                          * so there is a race window that an extent is still
2843                          * in OES_ACTIVE when truncate starts. */
2844                         LASSERT(!ext->oe_trunc_pending);
2845                         ext->oe_trunc_pending = 1;
2846                 } else {
2847                         EASSERT(ext->oe_state == OES_CACHE, ext);
2848                         osc_extent_state_set(ext, OES_TRUNC);
2849                         osc_update_pending(obj, OBD_BRW_WRITE,
2850                                            -ext->oe_nr_pages);
2851                 }
2852                 /* This extent could be on the full extents list, that's OK */
2853                 EASSERT(!ext->oe_hp && !ext->oe_urgent, ext);
2854                 if (!list_empty(&ext->oe_link))
2855                         list_move_tail(&ext->oe_link, &list);
2856                 else
2857                         list_add_tail(&ext->oe_link, &list);
2858
2859                 ext = next_extent(ext);
2860         }
2861         osc_object_unlock(obj);
2862
2863         osc_list_maint(cli, obj);
2864
2865         while (!list_empty(&list)) {
2866                 int rc;
2867
2868                 ext = list_entry(list.next, struct osc_extent, oe_link);
2869                 list_del_init(&ext->oe_link);
2870
2871                 /* extent may be in OES_ACTIVE state because inode mutex
2872                  * is released before osc_io_end() in file write case */
2873                 if (ext->oe_state != OES_TRUNC)
2874                         osc_extent_wait(env, ext, OES_TRUNC);
2875
2876                 rc = osc_extent_truncate(ext, index, partial);
2877                 if (rc < 0) {
2878                         if (result == 0)
2879                                 result = rc;
2880
2881                         OSC_EXTENT_DUMP(D_ERROR, ext,
2882                                         "truncate error %d\n", rc);
2883                 } else if (ext->oe_nr_pages == 0) {
2884                         osc_extent_remove(ext);
2885                 } else {
2886                         /* this must be an overlapped extent which means only
2887                          * part of pages in this extent have been truncated.
2888                          */
2889                         EASSERTF(ext->oe_start <= index, ext,
2890                                  "trunc index = %lu/%d.\n", index, partial);
2891                         /* fix index to skip this partially truncated extent */
2892                         index = ext->oe_end + 1;
2893                         partial = false;
2894
2895                         /* we need to hold this extent in OES_TRUNC state so
2896                          * that no writeback will happen. This is to avoid
2897                          * BUG 17397.
2898                          * Only partial truncate can reach here, if @size is
2899                          * not zero, the caller should provide a valid @extp. */
2900                         LASSERT(*extp == NULL);
2901                         *extp = osc_extent_get(ext);
2902                         OSC_EXTENT_DUMP(D_CACHE, ext,
2903                                         "trunc at %llu\n", size);
2904                 }
2905                 osc_extent_put(env, ext);
2906         }
2907         if (waiting != NULL) {
2908                 int rc;
2909
2910                 /* ignore the result of osc_extent_wait the write initiator
2911                  * should take care of it. */
2912                 rc = osc_extent_wait(env, waiting, OES_INV);
2913                 if (rc < 0)
2914                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2915
2916                 osc_extent_put(env, waiting);
2917                 waiting = NULL;
2918                 goto again;
2919         }
2920         RETURN(result);
2921 }
2922 EXPORT_SYMBOL(osc_cache_truncate_start);
2923
2924 /**
2925  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2926  */
2927 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext)
2928 {
2929         if (ext != NULL) {
2930                 struct osc_object *obj = ext->oe_obj;
2931                 bool unplug = false;
2932
2933                 EASSERT(ext->oe_nr_pages > 0, ext);
2934                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2935                 EASSERT(!ext->oe_urgent, ext);
2936
2937                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2938                 osc_object_lock(obj);
2939                 osc_extent_state_set(ext, OES_CACHE);
2940                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2941                         ext->oe_urgent = 1;
2942                         list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2943                         unplug = true;
2944                 }
2945                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2946                 osc_object_unlock(obj);
2947                 osc_extent_put(env, ext);
2948
2949                 if (unplug)
2950                         osc_io_unplug_async(env, osc_cli(obj), obj);
2951         }
2952 }
2953
2954 /**
2955  * Wait for extents in a specific range to be written out.
2956  * The caller must have called osc_cache_writeback_range() to issue IO
2957  * otherwise it will take a long time for this function to finish.
2958  *
2959  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2960  * nobody else can dirty this range of file while we're waiting for
2961  * extents to be written.
2962  */
2963 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2964                          pgoff_t start, pgoff_t end)
2965 {
2966         struct osc_extent *ext;
2967         pgoff_t index = start;
2968         int     result = 0;
2969         ENTRY;
2970
2971 again:
2972         osc_object_lock(obj);
2973         ext = osc_extent_search(obj, index);
2974         if (ext == NULL)
2975                 ext = first_extent(obj);
2976         else if (ext->oe_end < index)
2977                 ext = next_extent(ext);
2978         while (ext != NULL) {
2979                 int rc;
2980
2981                 if (ext->oe_start > end)
2982                         break;
2983
2984                 if (!ext->oe_fsync_wait) {
2985                         ext = next_extent(ext);
2986                         continue;
2987                 }
2988
2989                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2990                              ext->oe_hp || ext->oe_urgent), ext);
2991                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2992                              !ext->oe_hp && ext->oe_urgent), ext);
2993
2994                 index = ext->oe_end + 1;
2995                 osc_extent_get(ext);
2996                 osc_object_unlock(obj);
2997
2998                 rc = osc_extent_wait(env, ext, OES_INV);
2999                 if (result == 0)
3000                         result = rc;
3001                 osc_extent_put(env, ext);
3002                 goto again;
3003         }
3004         osc_object_unlock(obj);
3005
3006         OSC_IO_DEBUG(obj, "sync file range.\n");
3007         RETURN(result);
3008 }
3009 EXPORT_SYMBOL(osc_cache_wait_range);
3010
3011 /**
3012  * Called to write out a range of osc object.
3013  *
3014  * @hp     : should be set this is caused by lock cancel;
3015  * @discard: is set if dirty pages should be dropped - file will be deleted or
3016  *         truncated, this implies there is no partially discarding extents.
3017  *
3018  * Return how many pages will be issued, or error code if error occurred.
3019  */
3020 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
3021                               pgoff_t start, pgoff_t end, int hp, int discard)
3022 {
3023         struct osc_extent *ext;
3024         struct list_head discard_list = LIST_HEAD_INIT(discard_list);
3025         bool unplug = false;
3026         int result = 0;
3027         ENTRY;
3028
3029         osc_object_lock(obj);
3030         ext = osc_extent_search(obj, start);
3031         if (ext == NULL)
3032                 ext = first_extent(obj);
3033         else if (ext->oe_end < start)
3034                 ext = next_extent(ext);
3035         while (ext != NULL) {
3036                 if (ext->oe_start > end)
3037                         break;
3038
3039                 ext->oe_fsync_wait = 1;
3040                 switch (ext->oe_state) {
3041                 case OES_CACHE:
3042                         result += ext->oe_nr_pages;
3043                         if (!discard) {
3044                                 struct list_head *list = NULL;
3045                                 if (hp) {
3046                                         EASSERT(!ext->oe_hp, ext);
3047                                         ext->oe_hp = 1;
3048                                         list = &obj->oo_hp_exts;
3049                                 } else if (!ext->oe_urgent) {
3050                                         ext->oe_urgent = 1;
3051                                         list = &obj->oo_urgent_exts;
3052                                 }
3053                                 if (list != NULL)
3054                                         list_move_tail(&ext->oe_link, list);
3055                                 unplug = true;
3056                         } else {
3057                                 /* the only discarder is lock cancelling, so
3058                                  * [start, end] must contain this extent */
3059                                 EASSERT(ext->oe_start >= start &&
3060                                         ext->oe_max_end <= end, ext);
3061                                 osc_extent_state_set(ext, OES_LOCKING);
3062                                 ext->oe_owner = current;
3063                                 list_move_tail(&ext->oe_link,
3064                                                    &discard_list);
3065                                 osc_update_pending(obj, OBD_BRW_WRITE,
3066                                                    -ext->oe_nr_pages);
3067                         }
3068                         break;
3069                 case OES_ACTIVE:
3070                         /* It's pretty bad to wait for ACTIVE extents, because
3071                          * we don't know how long we will wait for it to be
3072                          * flushed since it may be blocked at awaiting more
3073                          * grants. We do this for the correctness of fsync. */
3074                         LASSERT(hp == 0 && discard == 0);
3075                         ext->oe_urgent = 1;
3076                         break;
3077                 case OES_TRUNC:
3078                         /* this extent is being truncated, can't do anything
3079                          * for it now. it will be set to urgent after truncate
3080                          * is finished in osc_cache_truncate_end(). */
3081                 default:
3082                         break;
3083                 }
3084                 ext = next_extent(ext);
3085         }
3086         osc_object_unlock(obj);
3087
3088         LASSERT(ergo(!discard, list_empty(&discard_list)));
3089         if (!list_empty(&discard_list)) {
3090                 struct osc_extent *tmp;
3091                 int rc;
3092
3093                 osc_list_maint(osc_cli(obj), obj);
3094                 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
3095                         list_del_init(&ext->oe_link);
3096                         EASSERT(ext->oe_state == OES_LOCKING, ext);
3097
3098                         /* Discard caching pages. We don't actually write this
3099                          * extent out but we complete it as if we did. */
3100                         rc = osc_extent_make_ready(env, ext);
3101                         if (unlikely(rc < 0)) {
3102                                 OSC_EXTENT_DUMP(D_ERROR, ext,
3103                                                 "make_ready returned %d\n", rc);
3104                                 if (result >= 0)
3105                                         result = rc;
3106                         }
3107
3108                         /* finish the extent as if the pages were sent */
3109                         osc_extent_finish(env, ext, 0, 0);
3110                 }
3111         }
3112
3113         if (unplug)
3114                 osc_io_unplug(env, osc_cli(obj), obj);
3115
3116         if (hp || discard) {
3117                 int rc;
3118                 rc = osc_cache_wait_range(env, obj, start, end);
3119                 if (result >= 0 && rc < 0)
3120                         result = rc;
3121         }
3122
3123         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3124         RETURN(result);
3125 }
3126 EXPORT_SYMBOL(osc_cache_writeback_range);
3127
3128 /**
3129  * Returns a list of pages by a given [start, end] of \a obj.
3130  *
3131  * \param resched If not NULL, then we give up before hogging CPU for too
3132  * long and set *resched = 1, in that case caller should implement a retry
3133  * logic.
3134  *
3135  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3136  * crucial in the face of [offset, EOF] locks.
3137  *
3138  * Return at least one page in @queue unless there is no covered page.
3139  */
3140 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3141                         struct osc_object *osc, pgoff_t start, pgoff_t end,
3142                         osc_page_gang_cbt cb, void *cbdata)
3143 {
3144         struct osc_page *ops;
3145         void            **pvec;
3146         pgoff_t         idx;
3147         unsigned int    nr;
3148         unsigned int    i;
3149         unsigned int    j;
3150         int             res = CLP_GANG_OKAY;
3151         bool            tree_lock = true;
3152         ENTRY;
3153
3154         idx = start;
3155         pvec = osc_env_info(env)->oti_pvec;
3156         spin_lock(&osc->oo_tree_lock);
3157         while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3158                                             idx, OTI_PVEC_SIZE)) > 0) {
3159                 struct cl_page *page;
3160                 bool end_of_region = false;
3161
3162                 for (i = 0, j = 0; i < nr; ++i) {
3163                         ops = pvec[i];
3164                         pvec[i] = NULL;
3165
3166                         idx = osc_index(ops);
3167                         if (idx > end) {
3168                                 end_of_region = true;
3169                                 break;
3170                         }
3171
3172                         page = ops->ops_cl.cpl_page;
3173                         LASSERT(page->cp_type == CPT_CACHEABLE);
3174                         if (page->cp_state == CPS_FREEING)
3175                                 continue;
3176
3177                         cl_page_get(page);
3178                         lu_ref_add_atomic(&page->cp_reference,
3179                                           "gang_lookup", current);
3180                         pvec[j++] = ops;
3181                 }
3182                 ++idx;
3183
3184                 /*
3185                  * Here a delicate locking dance is performed. Current thread
3186                  * holds a reference to a page, but has to own it before it
3187                  * can be placed into queue. Owning implies waiting, so
3188                  * radix-tree lock is to be released. After a wait one has to
3189                  * check that pages weren't truncated (cl_page_own() returns
3190                  * error in the latter case).
3191                  */
3192                 spin_unlock(&osc->oo_tree_lock);
3193                 tree_lock = false;
3194
3195                 for (i = 0; i < j; ++i) {
3196                         ops = pvec[i];
3197                         if (res == CLP_GANG_OKAY)
3198                                 res = (*cb)(env, io, ops, cbdata);
3199
3200                         page = ops->ops_cl.cpl_page;
3201                         lu_ref_del(&page->cp_reference, "gang_lookup", current);
3202                         cl_page_put(env, page);
3203                 }
3204                 if (nr < OTI_PVEC_SIZE || end_of_region)
3205                         break;
3206
3207                 if (res == CLP_GANG_OKAY && need_resched())
3208                         res = CLP_GANG_RESCHED;
3209                 if (res != CLP_GANG_OKAY)
3210                         break;
3211
3212                 spin_lock(&osc->oo_tree_lock);
3213                 tree_lock = true;
3214         }
3215         if (tree_lock)
3216                 spin_unlock(&osc->oo_tree_lock);
3217         RETURN(res);
3218 }
3219 EXPORT_SYMBOL(osc_page_gang_lookup);
3220
3221 /**
3222  * Check if page @page is covered by an extra lock or discard it.
3223  */
3224 static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3225                                 struct osc_page *ops, void *cbdata)
3226 {
3227         struct osc_thread_info *info = osc_env_info(env);
3228         struct osc_object *osc = cbdata;
3229         pgoff_t index;
3230
3231         index = osc_index(ops);
3232         if (index >= info->oti_fn_index) {
3233                 struct ldlm_lock *tmp;
3234                 struct cl_page *page = ops->ops_cl.cpl_page;
3235
3236                 /* refresh non-overlapped index */
3237                 tmp = osc_dlmlock_at_pgoff(env, osc, index,
3238                                            OSC_DAP_FL_TEST_LOCK);
3239                 if (tmp != NULL) {
3240                         __u64 end = tmp->l_policy_data.l_extent.end;
3241                         /* Cache the first-non-overlapped index so as to skip
3242                          * all pages within [index, oti_fn_index). This is safe
3243                          * because if tmp lock is canceled, it will discard
3244                          * these pages. */
3245                         info->oti_fn_index = cl_index(osc2cl(osc), end + 1);
3246                         if (end == OBD_OBJECT_EOF)
3247                                 info->oti_fn_index = CL_PAGE_EOF;
3248                         LDLM_LOCK_PUT(tmp);
3249                 } else if (cl_page_own(env, io, page) == 0) {
3250                         /* discard the page */
3251                         cl_page_discard(env, io, page);
3252                         cl_page_disown(env, io, page);
3253                 } else {
3254                         LASSERT(page->cp_state == CPS_FREEING);
3255                 }
3256         }
3257
3258         info->oti_next_index = index + 1;
3259         return CLP_GANG_OKAY;
3260 }
3261
3262 int osc_discard_cb(const struct lu_env *env, struct cl_io *io,
3263                    struct osc_page *ops, void *cbdata)
3264 {
3265         struct osc_thread_info *info = osc_env_info(env);
3266         struct cl_page *page = ops->ops_cl.cpl_page;
3267
3268         /* page is top page. */
3269         info->oti_next_index = osc_index(ops) + 1;
3270         if (cl_page_own(env, io, page) == 0) {
3271                 if (!ergo(page->cp_type == CPT_CACHEABLE,
3272                           !PageDirty(cl_page_vmpage(page))))
3273                         CL_PAGE_DEBUG(D_ERROR, env, page,
3274                                         "discard dirty page?\n");
3275
3276                 /* discard the page */
3277                 cl_page_discard(env, io, page);
3278                 cl_page_disown(env, io, page);
3279         } else {
3280                 LASSERT(page->cp_state == CPS_FREEING);
3281         }
3282
3283         return CLP_GANG_OKAY;
3284 }
3285 EXPORT_SYMBOL(osc_discard_cb);
3286
3287 /**
3288  * Discard pages protected by the given lock. This function traverses radix
3289  * tree to find all covering pages and discard them. If a page is being covered
3290  * by other locks, it should remain in cache.
3291  *
3292  * If error happens on any step, the process continues anyway (the reasoning
3293  * behind this being that lock cancellation cannot be delayed indefinitely).
3294  */
3295 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3296                            pgoff_t start, pgoff_t end, bool discard)
3297 {
3298         struct osc_thread_info *info = osc_env_info(env);
3299         struct cl_io *io = osc_env_thread_io(env);
3300         osc_page_gang_cbt cb;
3301         int res;
3302         int result;
3303
3304         ENTRY;
3305
3306         io->ci_obj = cl_object_top(osc2cl(osc));
3307         io->ci_ignore_layout = 1;
3308         result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3309         if (result != 0)
3310                 GOTO(out, result);
3311
3312         cb = discard ? osc_discard_cb : check_and_discard_cb;
3313         info->oti_fn_index = info->oti_next_index = start;
3314         do {
3315                 res = osc_page_gang_lookup(env, io, osc,
3316                                            info->oti_next_index, end, cb, osc);
3317                 if (info->oti_next_index > end)
3318                         break;
3319
3320                 if (res == CLP_GANG_RESCHED)
3321                         cond_resched();
3322         } while (res != CLP_GANG_OKAY);
3323 out:
3324         cl_io_fini(env, io);
3325         RETURN(result);
3326 }
3327
3328
3329 /** @} osc */