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, 2017, 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 || chunk < ext_chk_start)
 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         struct pagevec        *pvec;
 980         int                    pages_in_chunk = 0;
 981         int                    ppc_bits    = cli->cl_chunkbits -
 982                                              PAGE_SHIFT;
 983         __u64                  trunc_chunk = trunc_index >> ppc_bits;
 984         int                    grants   = 0;
 985         int                    nr_pages = 0;
 986         int                    rc       = 0;
 987         __u16                  refcheck;
 988         ENTRY;
 989
 990         LASSERT(sanity_check(ext) == 0);
 991         LASSERT(ext->oe_state == OES_TRUNC);
 992         LASSERT(!ext->oe_urgent);
 993
 994         /* Request new lu_env.
 995          * We can't use that env from osc_cache_truncate_start() because
 996          * it's from lov_io_sub and not fully initialized. */
 997         env = cl_env_get(&refcheck);
 998         if (IS_ERR(env))
 999                 RETURN(PTR_ERR(env));
1000
1001         io  = osc_env_thread_io(env);
1002         io->ci_obj = cl_object_top(osc2cl(obj));
1003         io->ci_ignore_layout = 1;
1004         pvec = &osc_env_info(env)->oti_pagevec;
1005         ll_pagevec_init(pvec, 0);
1006         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
1007         if (rc < 0)
1008                 GOTO(out, rc);
1009
1010         /* discard all pages with index greater than trunc_index */
1011         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
1012                                      oap_pending_item) {
1013                 pgoff_t index = osc_index(oap2osc(oap));
1014                 struct cl_page  *page = oap2cl_page(oap);
1015
1016                 LASSERT(list_empty(&oap->oap_rpc_item));
1017
1018                 /* only discard the pages with their index greater than
1019                  * trunc_index, and ... */
1020                 if (index < trunc_index ||
1021                     (index == trunc_index && partial)) {
1022                         /* accounting how many pages remaining in the chunk
1023                          * so that we can calculate grants correctly. */
1024                         if (index >> ppc_bits == trunc_chunk)
1025                                 ++pages_in_chunk;
1026                         continue;
1027                 }
1028
1029                 list_del_init(&oap->oap_pending_item);
1030
1031                 cl_page_get(page);
1032                 lu_ref_add(&page->cp_reference, "truncate", current);
1033
1034                 if (cl_page_own(env, io, page) == 0) {
1035                         cl_page_discard(env, io, page);
1036                         cl_page_disown(env, io, page);
1037                 } else {
1038                         LASSERT(page->cp_state == CPS_FREEING);
1039                         LASSERT(0);
1040                 }
1041
1042                 lu_ref_del(&page->cp_reference, "truncate", current);
1043                 cl_pagevec_put(env, page, pvec);
1044
1045                 --ext->oe_nr_pages;
1046                 ++nr_pages;
1047         }
1048         pagevec_release(pvec);
1049
1050         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1051                       ext->oe_nr_pages == 0),
1052                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1053
1054         osc_object_lock(obj);
1055         if (ext->oe_nr_pages == 0) {
1056                 LASSERT(pages_in_chunk == 0);
1057                 grants = ext->oe_grants;
1058                 ext->oe_grants = 0;
1059         } else { /* calculate how many grants we can free */
1060                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1061                 pgoff_t last_index;
1062
1063
1064                 /* if there is no pages in this chunk, we can also free grants
1065                  * for the last chunk */
1066                 if (pages_in_chunk == 0) {
1067                         /* if this is the 1st chunk and no pages in this chunk,
1068                          * ext->oe_nr_pages must be zero, so we should be in
1069                          * the other if-clause. */
1070                         LASSERT(trunc_chunk > 0);
1071                         --trunc_chunk;
1072                         ++chunks;
1073                 }
1074
1075                 /* this is what we can free from this extent */
1076                 grants          = chunks << cli->cl_chunkbits;
1077                 ext->oe_grants -= grants;
1078                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1079                 ext->oe_end     = min(last_index, ext->oe_max_end);
1080                 LASSERT(ext->oe_end >= ext->oe_start);
1081                 LASSERT(ext->oe_grants > 0);
1082         }
1083         osc_object_unlock(obj);
1084
1085         if (grants > 0 || nr_pages > 0)
1086                 osc_free_grant(cli, nr_pages, grants, grants);
1087
1088 out:
1089         cl_io_fini(env, io);
1090         cl_env_put(env, &refcheck);
1091         RETURN(rc);
1092 }
1093
1094 /**
1095  * This function is used to make the extent prepared for transfer.
1096  * A race with flusing page - ll_writepage() has to be handled cautiously.
1097  */
1098 static int osc_extent_make_ready(const struct lu_env *env,
1099                                  struct osc_extent *ext)
1100 {
1101         struct osc_async_page *oap;
1102         struct osc_async_page *last = NULL;
1103         struct osc_object *obj = ext->oe_obj;
1104         unsigned int page_count = 0;
1105         int rc;
1106         ENTRY;
1107
1108         /* we're going to grab page lock, so object lock must not be taken. */
1109         LASSERT(sanity_check(ext) == 0);
1110         /* in locking state, any process should not touch this extent. */
1111         EASSERT(ext->oe_state == OES_LOCKING, ext);
1112         EASSERT(ext->oe_owner != NULL, ext);
1113
1114         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1115
1116         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1117                 ++page_count;
1118                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1119                         last = oap;
1120
1121                 /* checking ASYNC_READY is race safe */
1122                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1123                         continue;
1124
1125                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1126                 switch (rc) {
1127                 case 0:
1128                         spin_lock(&oap->oap_lock);
1129                         oap->oap_async_flags |= ASYNC_READY;
1130                         spin_unlock(&oap->oap_lock);
1131                         break;
1132                 case -EALREADY:
1133                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1134                         break;
1135                 default:
1136                         LASSERTF(0, "unknown return code: %d\n", rc);
1137                 }
1138         }
1139
1140         LASSERT(page_count == ext->oe_nr_pages);
1141         LASSERT(last != NULL);
1142         /* the last page is the only one we need to refresh its count by
1143          * the size of file. */
1144         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1145                 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1146                 LASSERT(last_oap_count > 0);
1147                 LASSERT(last->oap_page_off + last_oap_count <= PAGE_SIZE);
1148                 last->oap_count = last_oap_count;
1149                 spin_lock(&last->oap_lock);
1150                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1151                 spin_unlock(&last->oap_lock);
1152         }
1153
1154         /* for the rest of pages, we don't need to call osf_refresh_count()
1155          * because it's known they are not the last page */
1156         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1157                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1158                         oap->oap_count = PAGE_SIZE - oap->oap_page_off;
1159                         spin_lock(&oap->oap_lock);
1160                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1161                         spin_unlock(&oap->oap_lock);
1162                 }
1163         }
1164
1165         osc_object_lock(obj);
1166         osc_extent_state_set(ext, OES_RPC);
1167         osc_object_unlock(obj);
1168         /* get a refcount for RPC. */
1169         osc_extent_get(ext);
1170
1171         RETURN(0);
1172 }
1173
1174 /**
1175  * Quick and simple version of osc_extent_find(). This function is frequently
1176  * called to expand the extent for the same IO. To expand the extent, the
1177  * page index must be in the same or next chunk of ext->oe_end.
1178  */
1179 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index,
1180                              unsigned int *grants)
1181 {
1182         struct osc_object *obj = ext->oe_obj;
1183         struct client_obd *cli = osc_cli(obj);
1184         struct osc_extent *next;
1185         int ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1186         pgoff_t chunk = index >> ppc_bits;
1187         pgoff_t end_chunk;
1188         pgoff_t end_index;
1189         unsigned int chunksize = 1 << cli->cl_chunkbits;
1190         int rc = 0;
1191         ENTRY;
1192
1193         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1194         osc_object_lock(obj);
1195         LASSERT(sanity_check_nolock(ext) == 0);
1196         end_chunk = ext->oe_end >> ppc_bits;
1197         if (chunk > end_chunk + 1)
1198                 GOTO(out, rc = -ERANGE);
1199
1200         if (end_chunk >= chunk)
1201                 GOTO(out, rc = 0);
1202
1203         LASSERT(end_chunk + 1 == chunk);
1204
1205         /* try to expand this extent to cover @index */
1206         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1207
1208         /* don't go over the maximum extent size reported by server */
1209         if (end_index - ext->oe_start + 1 > cli->cl_max_extent_pages)
1210                 GOTO(out, rc = -ERANGE);
1211
1212         next = next_extent(ext);
1213         if (next != NULL && next->oe_start <= end_index)
1214                 /* complex mode - overlapped with the next extent,
1215                  * this case will be handled by osc_extent_find() */
1216                 GOTO(out, rc = -EAGAIN);
1217
1218         ext->oe_end = end_index;
1219         ext->oe_grants += chunksize;
1220         LASSERT(*grants >= chunksize);
1221         *grants -= chunksize;
1222         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1223                  "overlapped after expanding for %lu.\n", index);
1224         EXIT;
1225
1226 out:
1227         osc_object_unlock(obj);
1228         RETURN(rc);
1229 }
1230
1231 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
1232                                   const char *func, int line)
1233 {
1234         struct osc_extent *ext;
1235         int cnt;
1236
1237         if (!cfs_cdebug_show(level, DEBUG_SUBSYSTEM))
1238                 return;
1239
1240         CDEBUG(level, "Dump object %p extents at %s:%d, mppr: %u.\n",
1241                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1242
1243         /* osc_object_lock(obj); */
1244         cnt = 1;
1245         for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1246                 OSC_EXTENT_DUMP(level, ext, "in tree %d.\n", cnt++);
1247
1248         cnt = 1;
1249         list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1250                 OSC_EXTENT_DUMP(level, ext, "hp %d.\n", cnt++);
1251
1252         cnt = 1;
1253         list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1254                 OSC_EXTENT_DUMP(level, ext, "urgent %d.\n", cnt++);
1255
1256         cnt = 1;
1257         list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1258                 OSC_EXTENT_DUMP(level, ext, "reading %d.\n", cnt++);
1259         /* osc_object_unlock(obj); */
1260 }
1261
1262 /* ------------------ osc extent end ------------------ */
1263
1264 static inline int osc_is_ready(struct osc_object *osc)
1265 {
1266         return !list_empty(&osc->oo_ready_item) ||
1267                !list_empty(&osc->oo_hp_ready_item);
1268 }
1269
1270 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1271         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1272                (OSC), osc_is_ready(OSC),                                       \
1273                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1274                list_empty_marker(&(OSC)->oo_ready_item),                       \
1275                atomic_read(&(OSC)->oo_nr_writes),                              \
1276                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1277                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1278                atomic_read(&(OSC)->oo_nr_reads),                               \
1279                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1280                ##args)
1281
1282 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1283                           int cmd)
1284 {
1285         struct osc_page *opg  = oap2osc_page(oap);
1286         struct cl_page  *page = oap2cl_page(oap);
1287         int result;
1288
1289         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1290
1291         ENTRY;
1292         result = cl_page_make_ready(env, page, CRT_WRITE);
1293         if (result == 0)
1294                 opg->ops_submit_time = ktime_get();
1295         RETURN(result);
1296 }
1297
1298 static int osc_refresh_count(const struct lu_env *env,
1299                              struct osc_async_page *oap, int cmd)
1300 {
1301         struct osc_page  *opg = oap2osc_page(oap);
1302         pgoff_t index = osc_index(oap2osc(oap));
1303         struct cl_object *obj;
1304         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
1305         int result;
1306         loff_t kms;
1307
1308         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1309         LASSERT(!(cmd & OBD_BRW_READ));
1310         LASSERT(opg != NULL);
1311         obj = opg->ops_cl.cpl_obj;
1312
1313         cl_object_attr_lock(obj);
1314         result = cl_object_attr_get(env, obj, attr);
1315         cl_object_attr_unlock(obj);
1316         if (result < 0)
1317                 return result;
1318         kms = attr->cat_kms;
1319         if (cl_offset(obj, index) >= kms)
1320                 /* catch race with truncate */
1321                 return 0;
1322         else if (cl_offset(obj, index + 1) > kms)
1323                 /* catch sub-page write at end of file */
1324                 return kms % PAGE_SIZE;
1325         else
1326                 return PAGE_SIZE;
1327 }
1328
1329 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1330                           int cmd, int rc)
1331 {
1332         struct osc_page   *opg  = oap2osc_page(oap);
1333         struct cl_page    *page = oap2cl_page(oap);
1334         enum cl_req_type   crt;
1335         int srvlock;
1336
1337         ENTRY;
1338
1339         cmd &= ~OBD_BRW_NOQUOTA;
1340         LASSERTF(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ),
1341                  "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1342         LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1343                 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1344         LASSERT(opg->ops_transfer_pinned);
1345
1346         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1347         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1348         opg->ops_transfer_pinned = 0;
1349
1350         opg->ops_submit_time = ktime_set(0, 0);
1351         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1352
1353         /* statistic */
1354         if (rc == 0 && srvlock) {
1355                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1356                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1357                 size_t bytes = oap->oap_count;
1358
1359                 if (crt == CRT_READ)
1360                         stats->os_lockless_reads += bytes;
1361                 else
1362                         stats->os_lockless_writes += bytes;
1363         }
1364
1365         /*
1366          * This has to be the last operation with the page, as locks are
1367          * released in cl_page_completion() and nothing except for the
1368          * reference counter protects page from concurrent reclaim.
1369          */
1370         lu_ref_del(&page->cp_reference, "transfer", page);
1371
1372         cl_page_completion(env, page, crt, rc);
1373         cl_page_put(env, page);
1374
1375         RETURN(0);
1376 }
1377
1378 #define OSC_DUMP_GRANT(lvl, cli, fmt, args...) do {                     \
1379         struct client_obd *__tmp = (cli);                               \
1380         CDEBUG(lvl, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu "  \
1381                "dropped: %ld avail: %ld, dirty_grant: %ld, "            \
1382                "reserved: %ld, flight: %d } lru {in list: %ld, "        \
1383                "left: %ld, waiters: %d }" fmt "\n",                     \
1384                cli_name(__tmp),                                         \
1385                __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages,        \
1386                atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1387                __tmp->cl_lost_grant, __tmp->cl_avail_grant,             \
1388                __tmp->cl_dirty_grant,                                   \
1389                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,         \
1390                atomic_long_read(&__tmp->cl_lru_in_list),                \
1391                atomic_long_read(&__tmp->cl_lru_busy),                   \
1392                atomic_read(&__tmp->cl_lru_shrinkers), ##args);          \
1393 } while (0)
1394
1395 /* caller must hold loi_list_lock */
1396 static void osc_consume_write_grant(struct client_obd *cli,
1397                                     struct brw_page *pga)
1398 {
1399         assert_spin_locked(&cli->cl_loi_list_lock);
1400         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1401         atomic_long_inc(&obd_dirty_pages);
1402         cli->cl_dirty_pages++;
1403         pga->flag |= OBD_BRW_FROM_GRANT;
1404         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1405                PAGE_SIZE, pga, pga->pg);
1406         osc_update_next_shrink(cli);
1407 }
1408
1409 /* the companion to osc_consume_write_grant, called when a brw has completed.
1410  * must be called with the loi lock held. */
1411 static void osc_release_write_grant(struct client_obd *cli,
1412                                     struct brw_page *pga)
1413 {
1414         ENTRY;
1415
1416         assert_spin_locked(&cli->cl_loi_list_lock);
1417         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1418                 EXIT;
1419                 return;
1420         }
1421
1422         pga->flag &= ~OBD_BRW_FROM_GRANT;
1423         atomic_long_dec(&obd_dirty_pages);
1424         cli->cl_dirty_pages--;
1425         if (pga->flag & OBD_BRW_NOCACHE) {
1426                 pga->flag &= ~OBD_BRW_NOCACHE;
1427                 atomic_long_dec(&obd_dirty_transit_pages);
1428                 cli->cl_dirty_transit--;
1429         }
1430         EXIT;
1431 }
1432
1433 /**
1434  * To avoid sleeping with object lock held, it's good for us allocate enough
1435  * grants before entering into critical section.
1436  *
1437  * client_obd_list_lock held by caller
1438  */
1439 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1440 {
1441         int rc = -EDQUOT;
1442
1443         if (cli->cl_avail_grant >= bytes) {
1444                 cli->cl_avail_grant    -= bytes;
1445                 cli->cl_reserved_grant += bytes;
1446                 rc = 0;
1447         }
1448         return rc;
1449 }
1450
1451 static void __osc_unreserve_grant(struct client_obd *cli,
1452                                   unsigned int reserved, unsigned int unused)
1453 {
1454         /* it's quite normal for us to get more grant than reserved.
1455          * Thinking about a case that two extents merged by adding a new
1456          * chunk, we can save one extent tax. If extent tax is greater than
1457          * one chunk, we can save more grant by adding a new chunk */
1458         cli->cl_reserved_grant -= reserved;
1459         if (unused > reserved) {
1460                 cli->cl_avail_grant += reserved;
1461                 cli->cl_lost_grant  += unused - reserved;
1462                 cli->cl_dirty_grant -= unused - reserved;
1463         } else {
1464                 cli->cl_avail_grant += unused;
1465                 cli->cl_dirty_grant += reserved - unused;
1466         }
1467 }
1468
1469 static void osc_unreserve_grant(struct client_obd *cli,
1470                                 unsigned int reserved, unsigned int unused)
1471 {
1472         spin_lock(&cli->cl_loi_list_lock);
1473         __osc_unreserve_grant(cli, reserved, unused);
1474         if (unused > 0)
1475                 osc_wake_cache_waiters(cli);
1476         spin_unlock(&cli->cl_loi_list_lock);
1477 }
1478
1479 /**
1480  * Free grant after IO is finished or canceled.
1481  *
1482  * @lost_grant is used to remember how many grants we have allocated but not
1483  * used, we should return these grants to OST. There're two cases where grants
1484  * can be lost:
1485  * 1. truncate;
1486  * 2. blocksize at OST is less than PAGE_SIZE and a partial page was
1487  *    written. In this case OST may use less chunks to serve this partial
1488  *    write. OSTs don't actually know the page size on the client side. so
1489  *    clients have to calculate lost grant by the blocksize on the OST.
1490  *    See filter_grant_check() for details.
1491  */
1492 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1493                            unsigned int lost_grant, unsigned int dirty_grant)
1494 {
1495         unsigned long grant;
1496
1497         grant = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
1498
1499         spin_lock(&cli->cl_loi_list_lock);
1500         atomic_long_sub(nr_pages, &obd_dirty_pages);
1501         cli->cl_dirty_pages -= nr_pages;
1502         cli->cl_lost_grant += lost_grant;
1503         cli->cl_dirty_grant -= dirty_grant;
1504         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1505                 /* borrow some grant from truncate to avoid the case that
1506                  * truncate uses up all avail grant */
1507                 cli->cl_lost_grant -= grant;
1508                 cli->cl_avail_grant += grant;
1509         }
1510         osc_wake_cache_waiters(cli);
1511         spin_unlock(&cli->cl_loi_list_lock);
1512         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu/%lu\n",
1513                lost_grant, cli->cl_lost_grant,
1514                cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_SHIFT,
1515                cli->cl_dirty_grant);
1516 }
1517
1518 /**
1519  * The companion to osc_enter_cache(), called when @oap is no longer part of
1520  * the dirty accounting due to error.
1521  */
1522 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1523 {
1524         spin_lock(&cli->cl_loi_list_lock);
1525         osc_release_write_grant(cli, &oap->oap_brw_page);
1526         spin_unlock(&cli->cl_loi_list_lock);
1527 }
1528
1529 /**
1530  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1531  * is available.
1532  */
1533 static int osc_enter_cache_try(struct client_obd *cli,
1534                                struct osc_async_page *oap,
1535                                int bytes, int transient)
1536 {
1537         int rc;
1538
1539         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1540
1541         rc = osc_reserve_grant(cli, bytes);
1542         if (rc < 0)
1543                 return 0;
1544
1545         if (cli->cl_dirty_pages < cli->cl_dirty_max_pages &&
1546             1 + atomic_long_read(&obd_dirty_pages) <= obd_max_dirty_pages) {
1547                 osc_consume_write_grant(cli, &oap->oap_brw_page);
1548                 if (transient) {
1549                         cli->cl_dirty_transit++;
1550                         atomic_long_inc(&obd_dirty_transit_pages);
1551                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
1552                 }
1553                 rc = 1;
1554         } else {
1555                 __osc_unreserve_grant(cli, bytes, bytes);
1556                 rc = 0;
1557         }
1558         return rc;
1559 }
1560
1561 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1562 {
1563         int rc;
1564         spin_lock(&cli->cl_loi_list_lock);
1565         rc = list_empty(&ocw->ocw_entry);
1566         spin_unlock(&cli->cl_loi_list_lock);
1567         return rc;
1568 }
1569
1570 /**
1571  * The main entry to reserve dirty page accounting. Usually the grant reserved
1572  * in this function will be freed in bulk in osc_free_grant() unless it fails
1573  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1574  *
1575  * The process will be put into sleep if it's already run out of grant.
1576  */
1577 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1578                            struct osc_async_page *oap, int bytes)
1579 {
1580         struct osc_object       *osc = oap->oap_obj;
1581         struct lov_oinfo        *loi = osc->oo_oinfo;
1582         struct osc_cache_waiter  ocw;
1583         struct l_wait_info       lwi;
1584         int                      rc = -EDQUOT;
1585         ENTRY;
1586
1587         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(AT_OFF ? obd_timeout : at_max),
1588                                NULL, LWI_ON_SIGNAL_NOOP, NULL);
1589
1590         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1591
1592         spin_lock(&cli->cl_loi_list_lock);
1593
1594         /* force the caller to try sync io.  this can jump the list
1595          * of queued writes and create a discontiguous rpc stream */
1596         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1597             cli->cl_dirty_max_pages == 0 ||
1598             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) {
1599                 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1600                 GOTO(out, rc = -EDQUOT);
1601         }
1602
1603         /* Hopefully normal case - cache space and write credits available */
1604         if (list_empty(&cli->cl_cache_waiters) &&
1605             osc_enter_cache_try(cli, oap, bytes, 0)) {
1606                 OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1607                 GOTO(out, rc = 0);
1608         }
1609
1610         /* We can get here for two reasons: too many dirty pages in cache, or
1611          * run out of grants. In both cases we should write dirty pages out.
1612          * Adding a cache waiter will trigger urgent write-out no matter what
1613          * RPC size will be.
1614          * The exiting condition is no avail grants and no dirty pages caching,
1615          * that really means there is no space on the OST. */
1616         init_waitqueue_head(&ocw.ocw_waitq);
1617         ocw.ocw_oap   = oap;
1618         ocw.ocw_grant = bytes;
1619         while (cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0) {
1620                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1621                 ocw.ocw_rc = 0;
1622                 spin_unlock(&cli->cl_loi_list_lock);
1623
1624                 osc_io_unplug_async(env, cli, NULL);
1625
1626                 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1627                        cli_name(cli), &ocw, oap);
1628
1629                 rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1630
1631                 spin_lock(&cli->cl_loi_list_lock);
1632
1633                 if (rc < 0) {
1634                         /* l_wait_event is interrupted by signal or timed out */
1635                         list_del_init(&ocw.ocw_entry);
1636                         break;
1637                 }
1638                 LASSERT(list_empty(&ocw.ocw_entry));
1639                 rc = ocw.ocw_rc;
1640
1641                 if (rc != -EDQUOT)
1642                         break;
1643                 if (osc_enter_cache_try(cli, oap, bytes, 0)) {
1644                         rc = 0;
1645                         break;
1646                 }
1647         }
1648
1649         switch (rc) {
1650         case 0:
1651                 OSC_DUMP_GRANT(D_CACHE, cli, "finally got grant space\n");
1652                 break;
1653         case -ETIMEDOUT:
1654                 OSC_DUMP_GRANT(D_CACHE, cli,
1655                                "timeout, fall back to sync i/o\n");
1656                 osc_extent_tree_dump(D_CACHE, osc);
1657                 /* fall back to synchronous I/O */
1658                 rc = -EDQUOT;
1659                 break;
1660         case -EINTR:
1661                 /* Ensures restartability - LU-3581 */
1662                 OSC_DUMP_GRANT(D_CACHE, cli, "interrupted\n");
1663                 rc = -ERESTARTSYS;
1664                 break;
1665         case -EDQUOT:
1666                 OSC_DUMP_GRANT(D_CACHE, cli,
1667                                "no grant space, fall back to sync i/o\n");
1668                 break;
1669         default:
1670                 CDEBUG(D_CACHE, "%s: event for cache space @ %p never arrived "
1671                        "due to %d, fall back to sync i/o\n",
1672                        cli_name(cli), &ocw, rc);
1673                 break;
1674         }
1675         EXIT;
1676 out:
1677         spin_unlock(&cli->cl_loi_list_lock);
1678         RETURN(rc);
1679 }
1680
1681 /* caller must hold loi_list_lock */
1682 void osc_wake_cache_waiters(struct client_obd *cli)
1683 {
1684         struct list_head *l, *tmp;
1685         struct osc_cache_waiter *ocw;
1686
1687         ENTRY;
1688         list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1689                 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
1690
1691                 ocw->ocw_rc = -EDQUOT;
1692
1693                 if (osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
1694                         ocw->ocw_rc = 0;
1695
1696                 if (ocw->ocw_rc == 0 ||
1697                     !(cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0)) {
1698                         list_del_init(&ocw->ocw_entry);
1699                         CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant "
1700                                "%ld, %d\n", ocw, ocw->ocw_oap,
1701                                cli->cl_avail_grant, ocw->ocw_rc);
1702
1703                         wake_up(&ocw->ocw_waitq);
1704                 }
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 /**
1990  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1991  * get_write_extent() takes all appropriate extents in atomic.
1992  *
1993  * The following policy is used to collect extents for IO:
1994  * 1. Add as many HP extents as possible;
1995  * 2. Add the first urgent extent in urgent extent list and take it out of
1996  *    urgent list;
1997  * 3. Add subsequent extents of this urgent extent;
1998  * 4. If urgent list is not empty, goto 2;
1999  * 5. Traverse the extent tree from the 1st extent;
2000  * 6. Above steps exit if there is no space in this RPC.
2001  */
2002 static unsigned int get_write_extents(struct osc_object *obj,
2003                                       struct list_head *rpclist)
2004 {
2005         struct client_obd *cli = osc_cli(obj);
2006         struct osc_extent *ext;
2007         struct extent_rpc_data data = {
2008                 .erd_rpc_list   = rpclist,
2009                 .erd_page_count = 0,
2010                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
2011                 .erd_max_chunks = osc_max_write_chunks(cli),
2012                 .erd_max_extents = 256,
2013         };
2014
2015         LASSERT(osc_object_is_locked(obj));
2016         while (!list_empty(&obj->oo_hp_exts)) {
2017                 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
2018                                  oe_link);
2019                 LASSERT(ext->oe_state == OES_CACHE);
2020                 if (!try_to_add_extent_for_io(cli, ext, &data))
2021                         return data.erd_page_count;
2022                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2023         }
2024         if (data.erd_page_count == data.erd_max_pages)
2025                 return data.erd_page_count;
2026
2027         while (!list_empty(&obj->oo_urgent_exts)) {
2028                 ext = list_entry(obj->oo_urgent_exts.next,
2029                                  struct osc_extent, oe_link);
2030                 if (!try_to_add_extent_for_io(cli, ext, &data))
2031                         return data.erd_page_count;
2032         }
2033         if (data.erd_page_count == data.erd_max_pages)
2034                 return data.erd_page_count;
2035
2036         /* One key difference between full extents and other extents: full
2037          * extents can usually only be added if the rpclist was empty, so if we
2038          * can't add one, we continue on to trying to add normal extents.  This
2039          * is so we don't miss adding extra extents to an RPC containing high
2040          * priority or urgent extents. */
2041         while (!list_empty(&obj->oo_full_exts)) {
2042                 ext = list_entry(obj->oo_full_exts.next,
2043                                  struct osc_extent, oe_link);
2044                 if (!try_to_add_extent_for_io(cli, ext, &data))
2045                         break;
2046         }
2047         if (data.erd_page_count == data.erd_max_pages)
2048                 return data.erd_page_count;
2049
2050         ext = first_extent(obj);
2051         while (ext != NULL) {
2052                 if ((ext->oe_state != OES_CACHE) ||
2053                     /* this extent may be already in current rpclist */
2054                     (!list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
2055                         ext = next_extent(ext);
2056                         continue;
2057                 }
2058
2059                 if (!try_to_add_extent_for_io(cli, ext, &data))
2060                         return data.erd_page_count;
2061
2062                 ext = next_extent(ext);
2063         }
2064         return data.erd_page_count;
2065 }
2066
2067 static int
2068 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
2069                    struct osc_object *osc)
2070 __must_hold(osc)
2071 {
2072         struct list_head   rpclist = LIST_HEAD_INIT(rpclist);
2073         struct osc_extent *ext;
2074         struct osc_extent *tmp;
2075         struct osc_extent *first = NULL;
2076         unsigned int page_count = 0;
2077         int srvlock = 0;
2078         int rc = 0;
2079         ENTRY;
2080
2081         LASSERT(osc_object_is_locked(osc));
2082
2083         page_count = get_write_extents(osc, &rpclist);
2084         LASSERT(equi(page_count == 0, list_empty(&rpclist)));
2085
2086         if (list_empty(&rpclist))
2087                 RETURN(0);
2088
2089         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2090
2091         list_for_each_entry(ext, &rpclist, oe_link) {
2092                 LASSERT(ext->oe_state == OES_CACHE ||
2093                         ext->oe_state == OES_LOCK_DONE);
2094                 if (ext->oe_state == OES_CACHE)
2095                         osc_extent_state_set(ext, OES_LOCKING);
2096                 else
2097                         osc_extent_state_set(ext, OES_RPC);
2098         }
2099
2100         /* we're going to grab page lock, so release object lock because
2101          * lock order is page lock -> object lock. */
2102         osc_object_unlock(osc);
2103
2104         list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2105                 if (ext->oe_state == OES_LOCKING) {
2106                         rc = osc_extent_make_ready(env, ext);
2107                         if (unlikely(rc < 0)) {
2108                                 list_del_init(&ext->oe_link);
2109                                 osc_extent_finish(env, ext, 0, rc);
2110                                 continue;
2111                         }
2112                 }
2113                 if (first == NULL) {
2114                         first = ext;
2115                         srvlock = ext->oe_srvlock;
2116                 } else {
2117                         LASSERT(srvlock == ext->oe_srvlock);
2118                 }
2119         }
2120
2121         if (!list_empty(&rpclist)) {
2122                 LASSERT(page_count > 0);
2123                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2124                 LASSERT(list_empty(&rpclist));
2125         }
2126
2127         osc_object_lock(osc);
2128         RETURN(rc);
2129 }
2130
2131 /**
2132  * prepare pages for ASYNC io and put pages in send queue.
2133  *
2134  * \param cmd OBD_BRW_* macroses
2135  * \param lop pending pages
2136  *
2137  * \return zero if no page added to send queue.
2138  * \return 1 if pages successfully added to send queue.
2139  * \return negative on errors.
2140  */
2141 static int
2142 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2143                   struct osc_object *osc)
2144 __must_hold(osc)
2145 {
2146         struct osc_extent *ext;
2147         struct osc_extent *next;
2148         struct list_head rpclist = LIST_HEAD_INIT(rpclist);
2149         struct extent_rpc_data data = {
2150                 .erd_rpc_list   = &rpclist,
2151                 .erd_page_count = 0,
2152                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
2153                 .erd_max_chunks = UINT_MAX,
2154                 .erd_max_extents = UINT_MAX,
2155         };
2156         int rc = 0;
2157         ENTRY;
2158
2159         LASSERT(osc_object_is_locked(osc));
2160         list_for_each_entry_safe(ext, next, &osc->oo_reading_exts, oe_link) {
2161                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2162                 if (!try_to_add_extent_for_io(cli, ext, &data))
2163                         break;
2164                 osc_extent_state_set(ext, OES_RPC);
2165                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2166         }
2167         LASSERT(data.erd_page_count <= data.erd_max_pages);
2168
2169         osc_update_pending(osc, OBD_BRW_READ, -data.erd_page_count);
2170
2171         if (!list_empty(&rpclist)) {
2172                 osc_object_unlock(osc);
2173
2174                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2175                 LASSERT(list_empty(&rpclist));
2176
2177                 osc_object_lock(osc);
2178         }
2179         RETURN(rc);
2180 }
2181
2182 #define list_to_obj(list, item) ({                                            \
2183         struct list_head *__tmp = (list)->next;                               \
2184         list_del_init(__tmp);                                         \
2185         list_entry(__tmp, struct osc_object, oo_##item);                      \
2186 })
2187
2188 /* This is called by osc_check_rpcs() to find which objects have pages that
2189  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2190 static struct osc_object *osc_next_obj(struct client_obd *cli)
2191 {
2192         ENTRY;
2193
2194         /* First return objects that have blocked locks so that they
2195          * will be flushed quickly and other clients can get the lock,
2196          * then objects which have pages ready to be stuffed into RPCs */
2197         if (!list_empty(&cli->cl_loi_hp_ready_list))
2198                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2199         if (!list_empty(&cli->cl_loi_ready_list))
2200                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2201
2202         /* then if we have cache waiters, return all objects with queued
2203          * writes.  This is especially important when many small files
2204          * have filled up the cache and not been fired into rpcs because
2205          * they don't pass the nr_pending/object threshhold */
2206         if (!list_empty(&cli->cl_cache_waiters) &&
2207             !list_empty(&cli->cl_loi_write_list))
2208                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2209
2210         /* then return all queued objects when we have an invalid import
2211          * so that they get flushed */
2212         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2213                 if (!list_empty(&cli->cl_loi_write_list))
2214                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2215                                            write_item));
2216                 if (!list_empty(&cli->cl_loi_read_list))
2217                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2218                                            read_item));
2219         }
2220         RETURN(NULL);
2221 }
2222
2223 /* called with the loi list lock held */
2224 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2225 __must_hold(&cli->cl_loi_list_lock)
2226 {
2227         struct osc_object *osc;
2228         int rc = 0;
2229         ENTRY;
2230
2231         while ((osc = osc_next_obj(cli)) != NULL) {
2232                 struct cl_object *obj = osc2cl(osc);
2233                 struct lu_ref_link link;
2234
2235                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2236
2237                 if (osc_max_rpc_in_flight(cli, osc)) {
2238                         __osc_list_maint(cli, osc);
2239                         break;
2240                 }
2241
2242                 cl_object_get(obj);
2243                 spin_unlock(&cli->cl_loi_list_lock);
2244                 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2245
2246                 /* attempt some read/write balancing by alternating between
2247                  * reads and writes in an object.  The makes_rpc checks here
2248                  * would be redundant if we were getting read/write work items
2249                  * instead of objects.  we don't want send_oap_rpc to drain a
2250                  * partial read pending queue when we're given this object to
2251                  * do io on writes while there are cache waiters */
2252                 osc_object_lock(osc);
2253                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2254                         rc = osc_send_write_rpc(env, cli, osc);
2255                         if (rc < 0) {
2256                                 CERROR("Write request failed with %d\n", rc);
2257
2258                                 /* osc_send_write_rpc failed, mostly because of
2259                                  * memory pressure.
2260                                  *
2261                                  * It can't break here, because if:
2262                                  *  - a page was submitted by osc_io_submit, so
2263                                  *    page locked;
2264                                  *  - no request in flight
2265                                  *  - no subsequent request
2266                                  * The system will be in live-lock state,
2267                                  * because there is no chance to call
2268                                  * osc_io_unplug() and osc_check_rpcs() any
2269                                  * more. pdflush can't help in this case,
2270                                  * because it might be blocked at grabbing
2271                                  * the page lock as we mentioned.
2272                                  *
2273                                  * Anyway, continue to drain pages. */
2274                                 /* break; */
2275                         }
2276                 }
2277                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2278                         rc = osc_send_read_rpc(env, cli, osc);
2279                         if (rc < 0)
2280                                 CERROR("Read request failed with %d\n", rc);
2281                 }
2282                 osc_object_unlock(osc);
2283
2284                 osc_list_maint(cli, osc);
2285                 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2286                 cl_object_put(env, obj);
2287
2288                 spin_lock(&cli->cl_loi_list_lock);
2289         }
2290 }
2291
2292 int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2293                    struct osc_object *osc, int async)
2294 {
2295         int rc = 0;
2296
2297         if (osc != NULL && osc_list_maint(cli, osc) == 0)
2298                 return 0;
2299
2300         if (!async) {
2301                 spin_lock(&cli->cl_loi_list_lock);
2302                 osc_check_rpcs(env, cli);
2303                 spin_unlock(&cli->cl_loi_list_lock);
2304         } else {
2305                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2306                 LASSERT(cli->cl_writeback_work != NULL);
2307                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2308         }
2309         return rc;
2310 }
2311 EXPORT_SYMBOL(osc_io_unplug0);
2312
2313 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2314                         struct page *page, loff_t offset)
2315 {
2316         struct obd_export     *exp = osc_export(osc);
2317         struct osc_async_page *oap = &ops->ops_oap;
2318         ENTRY;
2319
2320         if (!page)
2321                 return cfs_size_round(sizeof(*oap));
2322
2323         oap->oap_magic = OAP_MAGIC;
2324         oap->oap_cli = &exp->exp_obd->u.cli;
2325         oap->oap_obj = osc;
2326
2327         oap->oap_page = page;
2328         oap->oap_obj_off = offset;
2329         LASSERT(!(offset & ~PAGE_MASK));
2330
2331         if (cfs_capable(CFS_CAP_SYS_RESOURCE))
2332                 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2333
2334         INIT_LIST_HEAD(&oap->oap_pending_item);
2335         INIT_LIST_HEAD(&oap->oap_rpc_item);
2336
2337         spin_lock_init(&oap->oap_lock);
2338         CDEBUG(D_INFO, "oap %p page %p obj off %llu\n",
2339                oap, page, oap->oap_obj_off);
2340         RETURN(0);
2341 }
2342 EXPORT_SYMBOL(osc_prep_async_page);
2343
2344 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2345                        struct osc_page *ops)
2346 {
2347         struct osc_io *oio = osc_env_io(env);
2348         struct osc_extent     *ext = NULL;
2349         struct osc_async_page *oap = &ops->ops_oap;
2350         struct client_obd     *cli = oap->oap_cli;
2351         struct osc_object     *osc = oap->oap_obj;
2352         pgoff_t index;
2353         unsigned int tmp;
2354         unsigned int grants = 0;
2355         u32    brw_flags = OBD_BRW_ASYNC;
2356         int    cmd = OBD_BRW_WRITE;
2357         int    need_release = 0;
2358         int    rc = 0;
2359         ENTRY;
2360
2361         if (oap->oap_magic != OAP_MAGIC)
2362                 RETURN(-EINVAL);
2363
2364         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2365                 RETURN(-EIO);
2366
2367         if (!list_empty(&oap->oap_pending_item) ||
2368             !list_empty(&oap->oap_rpc_item))
2369                 RETURN(-EBUSY);
2370
2371         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2372         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2373         if (cfs_capable(CFS_CAP_SYS_RESOURCE)) {
2374                 brw_flags |= OBD_BRW_NOQUOTA;
2375                 cmd |= OBD_BRW_NOQUOTA;
2376         }
2377
2378         /* check if the file's owner/group is over quota */
2379         if (!(cmd & OBD_BRW_NOQUOTA)) {
2380                 struct cl_object *obj;
2381                 struct cl_attr   *attr;
2382                 unsigned int qid[LL_MAXQUOTAS];
2383
2384                 obj = cl_object_top(&osc->oo_cl);
2385                 attr = &osc_env_info(env)->oti_attr;
2386
2387                 cl_object_attr_lock(obj);
2388                 rc = cl_object_attr_get(env, obj, attr);
2389                 cl_object_attr_unlock(obj);
2390
2391                 qid[USRQUOTA] = attr->cat_uid;
2392                 qid[GRPQUOTA] = attr->cat_gid;
2393                 qid[PRJQUOTA] = attr->cat_projid;
2394                 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2395                         rc = -EDQUOT;
2396                 if (rc)
2397                         RETURN(rc);
2398         }
2399
2400         oap->oap_cmd = cmd;
2401         oap->oap_page_off = ops->ops_from;
2402         oap->oap_count = ops->ops_to - ops->ops_from;
2403         /* No need to hold a lock here,
2404          * since this page is not in any list yet. */
2405         oap->oap_async_flags = 0;
2406         oap->oap_brw_flags = brw_flags;
2407
2408         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2409                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2410
2411         index = osc_index(oap2osc(oap));
2412
2413         /* Add this page into extent by the following steps:
2414          * 1. if there exists an active extent for this IO, mostly this page
2415          *    can be added to the active extent and sometimes we need to
2416          *    expand extent to accomodate this page;
2417          * 2. otherwise, a new extent will be allocated. */
2418
2419         ext = oio->oi_active;
2420         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2421                 /* one chunk plus extent overhead must be enough to write this
2422                  * page */
2423                 grants = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2424                 if (ext->oe_end >= index)
2425                         grants = 0;
2426
2427                 /* it doesn't need any grant to dirty this page */
2428                 spin_lock(&cli->cl_loi_list_lock);
2429                 rc = osc_enter_cache_try(cli, oap, grants, 0);
2430                 spin_unlock(&cli->cl_loi_list_lock);
2431                 if (rc == 0) { /* try failed */
2432                         grants = 0;
2433                         need_release = 1;
2434                 } else if (ext->oe_end < index) {
2435                         tmp = grants;
2436                         /* try to expand this extent */
2437                         rc = osc_extent_expand(ext, index, &tmp);
2438                         if (rc < 0) {
2439                                 need_release = 1;
2440                                 /* don't free reserved grant */
2441                         } else {
2442                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2443                                                 "expanded for %lu.\n", index);
2444                                 osc_unreserve_grant(cli, grants, tmp);
2445                                 grants = 0;
2446                         }
2447                 }
2448                 rc = 0;
2449         } else if (ext != NULL) {
2450                 /* index is located outside of active extent */
2451                 need_release = 1;
2452         }
2453         if (need_release) {
2454                 osc_extent_release(env, ext);
2455                 oio->oi_active = NULL;
2456                 ext = NULL;
2457         }
2458
2459         if (ext == NULL) {
2460                 tmp = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2461
2462                 /* try to find new extent to cover this page */
2463                 LASSERT(oio->oi_active == NULL);
2464                 /* we may have allocated grant for this page if we failed
2465                  * to expand the previous active extent. */
2466                 LASSERT(ergo(grants > 0, grants >= tmp));
2467
2468                 rc = 0;
2469                 if (grants == 0) {
2470                         /* we haven't allocated grant for this page. */
2471                         rc = osc_enter_cache(env, cli, oap, tmp);
2472                         if (rc == 0)
2473                                 grants = tmp;
2474                 }
2475
2476                 tmp = grants;
2477                 if (rc == 0) {
2478                         ext = osc_extent_find(env, osc, index, &tmp);
2479                         if (IS_ERR(ext)) {
2480                                 LASSERT(tmp == grants);
2481                                 osc_exit_cache(cli, oap);
2482                                 rc = PTR_ERR(ext);
2483                                 ext = NULL;
2484                         } else {
2485                                 oio->oi_active = ext;
2486                         }
2487                 }
2488                 if (grants > 0)
2489                         osc_unreserve_grant(cli, grants, tmp);
2490         }
2491
2492         LASSERT(ergo(rc == 0, ext != NULL));
2493         if (ext != NULL) {
2494                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2495                          ext, "index = %lu.\n", index);
2496                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2497
2498                 osc_object_lock(osc);
2499                 if (ext->oe_nr_pages == 0)
2500                         ext->oe_srvlock = ops->ops_srvlock;
2501                 else
2502                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2503                 ++ext->oe_nr_pages;
2504                 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2505                 osc_object_unlock(osc);
2506
2507                 if (!ext->oe_layout_version)
2508                         ext->oe_layout_version = io->ci_layout_version;
2509         }
2510
2511         RETURN(rc);
2512 }
2513
2514 int osc_teardown_async_page(const struct lu_env *env,
2515                             struct osc_object *obj, struct osc_page *ops)
2516 {
2517         struct osc_async_page *oap = &ops->ops_oap;
2518         int rc = 0;
2519         ENTRY;
2520
2521         LASSERT(oap->oap_magic == OAP_MAGIC);
2522
2523         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2524                oap, ops, osc_index(oap2osc(oap)));
2525
2526         if (!list_empty(&oap->oap_rpc_item)) {
2527                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2528                 rc = -EBUSY;
2529         } else if (!list_empty(&oap->oap_pending_item)) {
2530                 struct osc_extent *ext = NULL;
2531
2532                 osc_object_lock(obj);
2533                 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2534                 osc_object_unlock(obj);
2535                 /* only truncated pages are allowed to be taken out.
2536                  * See osc_extent_truncate() and osc_cache_truncate_start()
2537                  * for details. */
2538                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2539                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2540                                         osc_index(oap2osc(oap)));
2541                         rc = -EBUSY;
2542                 }
2543                 if (ext != NULL)
2544                         osc_extent_put(env, ext);
2545         }
2546         RETURN(rc);
2547 }
2548
2549 /**
2550  * This is called when a page is picked up by kernel to write out.
2551  *
2552  * We should find out the corresponding extent and add the whole extent
2553  * into urgent list. The extent may be being truncated or used, handle it
2554  * carefully.
2555  */
2556 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2557                          struct osc_page *ops)
2558 {
2559         struct osc_extent *ext   = NULL;
2560         struct osc_object *obj   = cl2osc(ops->ops_cl.cpl_obj);
2561         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2562         pgoff_t            index = osc_index(ops);
2563         struct osc_async_page *oap = &ops->ops_oap;
2564         bool unplug = false;
2565         int rc = 0;
2566         ENTRY;
2567
2568         osc_object_lock(obj);
2569         ext = osc_extent_lookup(obj, index);
2570         if (ext == NULL) {
2571                 osc_extent_tree_dump(D_ERROR, obj);
2572                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2573         }
2574
2575         switch (ext->oe_state) {
2576         case OES_RPC:
2577         case OES_LOCK_DONE:
2578                 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2579                 LASSERT(0);
2580                 break;
2581         case OES_LOCKING:
2582                 /* If we know this extent is being written out, we should abort
2583                  * so that the writer can make this page ready. Otherwise, there
2584                  * exists a deadlock problem because other process can wait for
2585                  * page writeback bit holding page lock; and meanwhile in
2586                  * vvp_page_make_ready(), we need to grab page lock before
2587                  * really sending the RPC. */
2588         case OES_TRUNC:
2589                 /* race with truncate, page will be redirtied */
2590         case OES_ACTIVE:
2591                 /* The extent is active so we need to abort and let the caller
2592                  * re-dirty the page. If we continued on here, and we were the
2593                  * one making the extent active, we could deadlock waiting for
2594                  * the page writeback to clear but it won't because the extent
2595                  * is active and won't be written out. */
2596                 GOTO(out, rc = -EAGAIN);
2597         default:
2598                 break;
2599         }
2600
2601         rc = cl_page_prep(env, io, cp, CRT_WRITE);
2602         if (rc)
2603                 GOTO(out, rc);
2604
2605         spin_lock(&oap->oap_lock);
2606         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2607         spin_unlock(&oap->oap_lock);
2608
2609         if (memory_pressure_get())
2610                 ext->oe_memalloc = 1;
2611
2612         ext->oe_urgent = 1;
2613         if (ext->oe_state == OES_CACHE) {
2614                 OSC_EXTENT_DUMP(D_CACHE, ext,
2615                                 "flush page %p make it urgent.\n", oap);
2616                 if (list_empty(&ext->oe_link))
2617                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2618                 unplug = true;
2619         }
2620         rc = 0;
2621         EXIT;
2622
2623 out:
2624         osc_object_unlock(obj);
2625         osc_extent_put(env, ext);
2626         if (unplug)
2627                 osc_io_unplug_async(env, osc_cli(obj), obj);
2628         return rc;
2629 }
2630
2631 /**
2632  * this is called when a sync waiter receives an interruption.  Its job is to
2633  * get the caller woken as soon as possible.  If its page hasn't been put in an
2634  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2635  * desiring interruption which will forcefully complete the rpc once the rpc
2636  * has timed out.
2637  */
2638 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2639 {
2640         struct osc_async_page *oap = &ops->ops_oap;
2641         struct osc_object     *obj = oap->oap_obj;
2642         struct client_obd     *cli = osc_cli(obj);
2643         struct osc_extent     *ext;
2644         struct osc_extent     *found = NULL;
2645         struct list_head            *plist;
2646         pgoff_t index = osc_index(ops);
2647         int     rc = -EBUSY;
2648         int     cmd;
2649         ENTRY;
2650
2651         LASSERT(!oap->oap_interrupted);
2652         oap->oap_interrupted = 1;
2653
2654         /* Find out the caching extent */
2655         osc_object_lock(obj);
2656         if (oap->oap_cmd & OBD_BRW_WRITE) {
2657                 plist = &obj->oo_urgent_exts;
2658                 cmd   = OBD_BRW_WRITE;
2659         } else {
2660                 plist = &obj->oo_reading_exts;
2661                 cmd   = OBD_BRW_READ;
2662         }
2663         list_for_each_entry(ext, plist, oe_link) {
2664                 if (ext->oe_start <= index && ext->oe_end >= index) {
2665                         LASSERT(ext->oe_state == OES_LOCK_DONE);
2666                         /* For OES_LOCK_DONE state extent, it has already held
2667                          * a refcount for RPC. */
2668                         found = osc_extent_get(ext);
2669                         break;
2670                 }
2671         }
2672         if (found != NULL) {
2673                 list_del_init(&found->oe_link);
2674                 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2675                 osc_object_unlock(obj);
2676
2677                 osc_extent_finish(env, found, 0, -EINTR);
2678                 osc_extent_put(env, found);
2679                 rc = 0;
2680         } else {
2681                 osc_object_unlock(obj);
2682                 /* ok, it's been put in an rpc. only one oap gets a request
2683                  * reference */
2684                 if (oap->oap_request != NULL) {
2685                         ptlrpc_mark_interrupted(oap->oap_request);
2686                         ptlrpcd_wake(oap->oap_request);
2687                         ptlrpc_req_finished(oap->oap_request);
2688                         oap->oap_request = NULL;
2689                 }
2690         }
2691
2692         osc_list_maint(cli, obj);
2693         RETURN(rc);
2694 }
2695
2696 int osc_queue_sync_pages(const struct lu_env *env, const struct cl_io *io,
2697                          struct osc_object *obj, struct list_head *list,
2698                          int brw_flags)
2699 {
2700         struct client_obd     *cli = osc_cli(obj);
2701         struct osc_extent     *ext;
2702         struct osc_async_page *oap;
2703         int     page_count = 0;
2704         int     mppr       = cli->cl_max_pages_per_rpc;
2705         bool    can_merge   = true;
2706         pgoff_t start      = CL_PAGE_EOF;
2707         pgoff_t end        = 0;
2708         ENTRY;
2709
2710         list_for_each_entry(oap, list, oap_pending_item) {
2711                 struct osc_page *opg = oap2osc_page(oap);
2712                 pgoff_t index = osc_index(opg);
2713
2714                 if (index > end)
2715                         end = index;
2716                 if (index < start)
2717                         start = index;
2718                 ++page_count;
2719                 mppr <<= (page_count > mppr);
2720
2721                 if (unlikely(opg->ops_from > 0 || opg->ops_to < PAGE_SIZE))
2722                         can_merge = false;
2723         }
2724
2725         ext = osc_extent_alloc(obj);
2726         if (ext == NULL) {
2727                 struct osc_async_page *tmp;
2728
2729                 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2730                         list_del_init(&oap->oap_pending_item);
2731                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2732                 }
2733                 RETURN(-ENOMEM);
2734         }
2735
2736         ext->oe_rw = !!(brw_flags & OBD_BRW_READ);
2737         ext->oe_sync = 1;
2738         ext->oe_no_merge = !can_merge;
2739         ext->oe_urgent = 1;
2740         ext->oe_start = start;
2741         ext->oe_end = ext->oe_max_end = end;
2742         ext->oe_obj = obj;
2743         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2744         ext->oe_ndelay = !!(brw_flags & OBD_BRW_NDELAY);
2745         ext->oe_nr_pages = page_count;
2746         ext->oe_mppr = mppr;
2747         list_splice_init(list, &ext->oe_pages);
2748         ext->oe_layout_version = io->ci_layout_version;
2749
2750         osc_object_lock(obj);
2751         /* Reuse the initial refcount for RPC, don't drop it */
2752         osc_extent_state_set(ext, OES_LOCK_DONE);
2753         if (!ext->oe_rw) { /* write */
2754                 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2755                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2756         } else {
2757                 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2758                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2759         }
2760         osc_object_unlock(obj);
2761
2762         osc_io_unplug_async(env, cli, obj);
2763         RETURN(0);
2764 }
2765
2766 /**
2767  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2768  */
2769 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
2770                              __u64 size, struct osc_extent **extp)
2771 {
2772         struct client_obd *cli = osc_cli(obj);
2773         struct osc_extent *ext;
2774         struct osc_extent *waiting = NULL;
2775         pgoff_t index;
2776         struct list_head list = LIST_HEAD_INIT(list);
2777         int result = 0;
2778         bool partial;
2779         ENTRY;
2780
2781         /* pages with index greater or equal to index will be truncated. */
2782         index = cl_index(osc2cl(obj), size);
2783         partial = size > cl_offset(osc2cl(obj), index);
2784
2785 again:
2786         osc_object_lock(obj);
2787         ext = osc_extent_search(obj, index);
2788         if (ext == NULL)
2789                 ext = first_extent(obj);
2790         else if (ext->oe_end < index)
2791                 ext = next_extent(ext);
2792         while (ext != NULL) {
2793                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2794
2795                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2796                         /* if ext is in urgent state, it means there must exist
2797                          * a page already having been flushed by write_page().
2798                          * We have to wait for this extent because we can't
2799                          * truncate that page. */
2800                         OSC_EXTENT_DUMP(D_CACHE, ext,
2801                                         "waiting for busy extent\n");
2802                         waiting = osc_extent_get(ext);
2803                         break;
2804                 }
2805
2806                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:%llu.\n", size);
2807
2808                 osc_extent_get(ext);
2809                 if (ext->oe_state == OES_ACTIVE) {
2810                         /* though we grab inode mutex for write path, but we
2811                          * release it before releasing extent(in osc_io_end()),
2812                          * so there is a race window that an extent is still
2813                          * in OES_ACTIVE when truncate starts. */
2814                         LASSERT(!ext->oe_trunc_pending);
2815                         ext->oe_trunc_pending = 1;
2816                 } else {
2817                         EASSERT(ext->oe_state == OES_CACHE, ext);
2818                         osc_extent_state_set(ext, OES_TRUNC);
2819                         osc_update_pending(obj, OBD_BRW_WRITE,
2820                                            -ext->oe_nr_pages);
2821                 }
2822                 /* This extent could be on the full extents list, that's OK */
2823                 EASSERT(!ext->oe_hp && !ext->oe_urgent, ext);
2824                 if (!list_empty(&ext->oe_link))
2825                         list_move_tail(&ext->oe_link, &list);
2826                 else
2827                         list_add_tail(&ext->oe_link, &list);
2828
2829                 ext = next_extent(ext);
2830         }
2831         osc_object_unlock(obj);
2832
2833         osc_list_maint(cli, obj);
2834
2835         while (!list_empty(&list)) {
2836                 int rc;
2837
2838                 ext = list_entry(list.next, struct osc_extent, oe_link);
2839                 list_del_init(&ext->oe_link);
2840
2841                 /* extent may be in OES_ACTIVE state because inode mutex
2842                  * is released before osc_io_end() in file write case */
2843                 if (ext->oe_state != OES_TRUNC)
2844                         osc_extent_wait(env, ext, OES_TRUNC);
2845
2846                 rc = osc_extent_truncate(ext, index, partial);
2847                 if (rc < 0) {
2848                         if (result == 0)
2849                                 result = rc;
2850
2851                         OSC_EXTENT_DUMP(D_ERROR, ext,
2852                                         "truncate error %d\n", rc);
2853                 } else if (ext->oe_nr_pages == 0) {
2854                         osc_extent_remove(ext);
2855                 } else {
2856                         /* this must be an overlapped extent which means only
2857                          * part of pages in this extent have been truncated.
2858                          */
2859                         EASSERTF(ext->oe_start <= index, ext,
2860                                  "trunc index = %lu/%d.\n", index, partial);
2861                         /* fix index to skip this partially truncated extent */
2862                         index = ext->oe_end + 1;
2863                         partial = false;
2864
2865                         /* we need to hold this extent in OES_TRUNC state so
2866                          * that no writeback will happen. This is to avoid
2867                          * BUG 17397.
2868                          * Only partial truncate can reach here, if @size is
2869                          * not zero, the caller should provide a valid @extp. */
2870                         LASSERT(*extp == NULL);
2871                         *extp = osc_extent_get(ext);
2872                         OSC_EXTENT_DUMP(D_CACHE, ext,
2873                                         "trunc at %llu\n", size);
2874                 }
2875                 osc_extent_put(env, ext);
2876         }
2877         if (waiting != NULL) {
2878                 int rc;
2879
2880                 /* ignore the result of osc_extent_wait the write initiator
2881                  * should take care of it. */
2882                 rc = osc_extent_wait(env, waiting, OES_INV);
2883                 if (rc < 0)
2884                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2885
2886                 osc_extent_put(env, waiting);
2887                 waiting = NULL;
2888                 goto again;
2889         }
2890         RETURN(result);
2891 }
2892 EXPORT_SYMBOL(osc_cache_truncate_start);
2893
2894 /**
2895  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2896  */
2897 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext)
2898 {
2899         if (ext != NULL) {
2900                 struct osc_object *obj = ext->oe_obj;
2901                 bool unplug = false;
2902
2903                 EASSERT(ext->oe_nr_pages > 0, ext);
2904                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2905                 EASSERT(!ext->oe_urgent, ext);
2906
2907                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2908                 osc_object_lock(obj);
2909                 osc_extent_state_set(ext, OES_CACHE);
2910                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2911                         ext->oe_urgent = 1;
2912                         list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2913                         unplug = true;
2914                 }
2915                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2916                 osc_object_unlock(obj);
2917                 osc_extent_put(env, ext);
2918
2919                 if (unplug)
2920                         osc_io_unplug_async(env, osc_cli(obj), obj);
2921         }
2922 }
2923
2924 /**
2925  * Wait for extents in a specific range to be written out.
2926  * The caller must have called osc_cache_writeback_range() to issue IO
2927  * otherwise it will take a long time for this function to finish.
2928  *
2929  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2930  * nobody else can dirty this range of file while we're waiting for
2931  * extents to be written.
2932  */
2933 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2934                          pgoff_t start, pgoff_t end)
2935 {
2936         struct osc_extent *ext;
2937         pgoff_t index = start;
2938         int     result = 0;
2939         ENTRY;
2940
2941 again:
2942         osc_object_lock(obj);
2943         ext = osc_extent_search(obj, index);
2944         if (ext == NULL)
2945                 ext = first_extent(obj);
2946         else if (ext->oe_end < index)
2947                 ext = next_extent(ext);
2948         while (ext != NULL) {
2949                 int rc;
2950
2951                 if (ext->oe_start > end)
2952                         break;
2953
2954                 if (!ext->oe_fsync_wait) {
2955                         ext = next_extent(ext);
2956                         continue;
2957                 }
2958
2959                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2960                              ext->oe_hp || ext->oe_urgent), ext);
2961                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2962                              !ext->oe_hp && ext->oe_urgent), ext);
2963
2964                 index = ext->oe_end + 1;
2965                 osc_extent_get(ext);
2966                 osc_object_unlock(obj);
2967
2968                 rc = osc_extent_wait(env, ext, OES_INV);
2969                 if (result == 0)
2970                         result = rc;
2971                 osc_extent_put(env, ext);
2972                 goto again;
2973         }
2974         osc_object_unlock(obj);
2975
2976         OSC_IO_DEBUG(obj, "sync file range.\n");
2977         RETURN(result);
2978 }
2979 EXPORT_SYMBOL(osc_cache_wait_range);
2980
2981 /**
2982  * Called to write out a range of osc object.
2983  *
2984  * @hp     : should be set this is caused by lock cancel;
2985  * @discard: is set if dirty pages should be dropped - file will be deleted or
2986  *         truncated, this implies there is no partially discarding extents.
2987  *
2988  * Return how many pages will be issued, or error code if error occurred.
2989  */
2990 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2991                               pgoff_t start, pgoff_t end, int hp, int discard)
2992 {
2993         struct osc_extent *ext;
2994         struct list_head discard_list = LIST_HEAD_INIT(discard_list);
2995         bool unplug = false;
2996         int result = 0;
2997         ENTRY;
2998
2999         osc_object_lock(obj);
3000         ext = osc_extent_search(obj, start);
3001         if (ext == NULL)
3002                 ext = first_extent(obj);
3003         else if (ext->oe_end < start)
3004                 ext = next_extent(ext);
3005         while (ext != NULL) {
3006                 if (ext->oe_start > end)
3007                         break;
3008
3009                 ext->oe_fsync_wait = 1;
3010                 switch (ext->oe_state) {
3011                 case OES_CACHE:
3012                         result += ext->oe_nr_pages;
3013                         if (!discard) {
3014                                 struct list_head *list = NULL;
3015                                 if (hp) {
3016                                         EASSERT(!ext->oe_hp, ext);
3017                                         ext->oe_hp = 1;
3018                                         list = &obj->oo_hp_exts;
3019                                 } else if (!ext->oe_urgent) {
3020                                         ext->oe_urgent = 1;
3021                                         list = &obj->oo_urgent_exts;
3022                                 }
3023                                 if (list != NULL)
3024                                         list_move_tail(&ext->oe_link, list);
3025                                 unplug = true;
3026                         } else {
3027                                 /* the only discarder is lock cancelling, so
3028                                  * [start, end] must contain this extent */
3029                                 EASSERT(ext->oe_start >= start &&
3030                                         ext->oe_max_end <= end, ext);
3031                                 osc_extent_state_set(ext, OES_LOCKING);
3032                                 ext->oe_owner = current;
3033                                 list_move_tail(&ext->oe_link,
3034                                                    &discard_list);
3035                                 osc_update_pending(obj, OBD_BRW_WRITE,
3036                                                    -ext->oe_nr_pages);
3037                         }
3038                         break;
3039                 case OES_ACTIVE:
3040                         /* It's pretty bad to wait for ACTIVE extents, because
3041                          * we don't know how long we will wait for it to be
3042                          * flushed since it may be blocked at awaiting more
3043                          * grants. We do this for the correctness of fsync. */
3044                         LASSERT(hp == 0 && discard == 0);
3045                         ext->oe_urgent = 1;
3046                         break;
3047                 case OES_TRUNC:
3048                         /* this extent is being truncated, can't do anything
3049                          * for it now. it will be set to urgent after truncate
3050                          * is finished in osc_cache_truncate_end(). */
3051                 default:
3052                         break;
3053                 }
3054                 ext = next_extent(ext);
3055         }
3056         osc_object_unlock(obj);
3057
3058         LASSERT(ergo(!discard, list_empty(&discard_list)));
3059         if (!list_empty(&discard_list)) {
3060                 struct osc_extent *tmp;
3061                 int rc;
3062
3063                 osc_list_maint(osc_cli(obj), obj);
3064                 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
3065                         list_del_init(&ext->oe_link);
3066                         EASSERT(ext->oe_state == OES_LOCKING, ext);
3067
3068                         /* Discard caching pages. We don't actually write this
3069                          * extent out but we complete it as if we did. */
3070                         rc = osc_extent_make_ready(env, ext);
3071                         if (unlikely(rc < 0)) {
3072                                 OSC_EXTENT_DUMP(D_ERROR, ext,
3073                                                 "make_ready returned %d\n", rc);
3074                                 if (result >= 0)
3075                                         result = rc;
3076                         }
3077
3078                         /* finish the extent as if the pages were sent */
3079                         osc_extent_finish(env, ext, 0, 0);
3080                 }
3081         }
3082
3083         if (unplug)
3084                 osc_io_unplug(env, osc_cli(obj), obj);
3085
3086         if (hp || discard) {
3087                 int rc;
3088                 rc = osc_cache_wait_range(env, obj, start, end);
3089                 if (result >= 0 && rc < 0)
3090                         result = rc;
3091         }
3092
3093         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3094         RETURN(result);
3095 }
3096 EXPORT_SYMBOL(osc_cache_writeback_range);
3097
3098 /**
3099  * Returns a list of pages by a given [start, end] of \a obj.
3100  *
3101  * \param resched If not NULL, then we give up before hogging CPU for too
3102  * long and set *resched = 1, in that case caller should implement a retry
3103  * logic.
3104  *
3105  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3106  * crucial in the face of [offset, EOF] locks.
3107  *
3108  * Return at least one page in @queue unless there is no covered page.
3109  */
3110 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3111                         struct osc_object *osc, pgoff_t start, pgoff_t end,
3112                         osc_page_gang_cbt cb, void *cbdata)
3113 {
3114         struct osc_page *ops;
3115         struct pagevec  *pagevec;
3116         void            **pvec;
3117         pgoff_t         idx;
3118         unsigned int    nr;
3119         unsigned int    i;
3120         unsigned int    j;
3121         int             res = CLP_GANG_OKAY;
3122         bool            tree_lock = true;
3123         ENTRY;
3124
3125         idx = start;
3126         pvec = osc_env_info(env)->oti_pvec;
3127         pagevec = &osc_env_info(env)->oti_pagevec;
3128         ll_pagevec_init(pagevec, 0);
3129         spin_lock(&osc->oo_tree_lock);
3130         while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3131                                             idx, OTI_PVEC_SIZE)) > 0) {
3132                 struct cl_page *page;
3133                 bool end_of_region = false;
3134
3135                 for (i = 0, j = 0; i < nr; ++i) {
3136                         ops = pvec[i];
3137                         pvec[i] = NULL;
3138
3139                         idx = osc_index(ops);
3140                         if (idx > end) {
3141                                 end_of_region = true;
3142                                 break;
3143                         }
3144
3145                         page = ops->ops_cl.cpl_page;
3146                         LASSERT(page->cp_type == CPT_CACHEABLE);
3147                         if (page->cp_state == CPS_FREEING)
3148                                 continue;
3149
3150                         cl_page_get(page);
3151                         lu_ref_add_atomic(&page->cp_reference,
3152                                           "gang_lookup", current);
3153                         pvec[j++] = ops;
3154                 }
3155                 ++idx;
3156
3157                 /*
3158                  * Here a delicate locking dance is performed. Current thread
3159                  * holds a reference to a page, but has to own it before it
3160                  * can be placed into queue. Owning implies waiting, so
3161                  * radix-tree lock is to be released. After a wait one has to
3162                  * check that pages weren't truncated (cl_page_own() returns
3163                  * error in the latter case).
3164                  */
3165                 spin_unlock(&osc->oo_tree_lock);
3166                 tree_lock = false;
3167
3168                 for (i = 0; i < j; ++i) {
3169                         ops = pvec[i];
3170                         if (res == CLP_GANG_OKAY)
3171                                 res = (*cb)(env, io, ops, cbdata);
3172
3173                         page = ops->ops_cl.cpl_page;
3174                         lu_ref_del(&page->cp_reference, "gang_lookup", current);
3175                         cl_pagevec_put(env, page, pagevec);
3176                 }
3177                 pagevec_release(pagevec);
3178
3179                 if (nr < OTI_PVEC_SIZE || end_of_region)
3180                         break;
3181
3182                 if (res == CLP_GANG_OKAY && need_resched())
3183                         res = CLP_GANG_RESCHED;
3184                 if (res != CLP_GANG_OKAY)
3185                         break;
3186
3187                 spin_lock(&osc->oo_tree_lock);
3188                 tree_lock = true;
3189         }
3190         if (tree_lock)
3191                 spin_unlock(&osc->oo_tree_lock);
3192         RETURN(res);
3193 }
3194 EXPORT_SYMBOL(osc_page_gang_lookup);
3195
3196 /**
3197  * Check if page @page is covered by an extra lock or discard it.
3198  */
3199 static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3200                                 struct osc_page *ops, void *cbdata)
3201 {
3202         struct osc_thread_info *info = osc_env_info(env);
3203         struct osc_object *osc = cbdata;
3204         pgoff_t index;
3205
3206         index = osc_index(ops);
3207         if (index >= info->oti_fn_index) {
3208                 struct ldlm_lock *tmp;
3209                 struct cl_page *page = ops->ops_cl.cpl_page;
3210
3211                 /* refresh non-overlapped index */
3212                 tmp = osc_dlmlock_at_pgoff(env, osc, index,
3213                                            OSC_DAP_FL_TEST_LOCK);
3214                 if (tmp != NULL) {
3215                         __u64 end = tmp->l_policy_data.l_extent.end;
3216                         /* Cache the first-non-overlapped index so as to skip
3217                          * all pages within [index, oti_fn_index). This is safe
3218                          * because if tmp lock is canceled, it will discard
3219                          * these pages. */
3220                         info->oti_fn_index = cl_index(osc2cl(osc), end + 1);
3221                         if (end == OBD_OBJECT_EOF)
3222                                 info->oti_fn_index = CL_PAGE_EOF;
3223                         LDLM_LOCK_PUT(tmp);
3224                 } else if (cl_page_own(env, io, page) == 0) {
3225                         /* discard the page */
3226                         cl_page_discard(env, io, page);
3227                         cl_page_disown(env, io, page);
3228                 } else {
3229                         LASSERT(page->cp_state == CPS_FREEING);
3230                 }
3231         }
3232
3233         info->oti_next_index = index + 1;
3234         return CLP_GANG_OKAY;
3235 }
3236
3237 int osc_discard_cb(const struct lu_env *env, struct cl_io *io,
3238                    struct osc_page *ops, void *cbdata)
3239 {
3240         struct osc_thread_info *info = osc_env_info(env);
3241         struct cl_page *page = ops->ops_cl.cpl_page;
3242
3243         /* page is top page. */
3244         info->oti_next_index = osc_index(ops) + 1;
3245         if (cl_page_own(env, io, page) == 0) {
3246                 if (!ergo(page->cp_type == CPT_CACHEABLE,
3247                           !PageDirty(cl_page_vmpage(page))))
3248                         CL_PAGE_DEBUG(D_ERROR, env, page,
3249                                         "discard dirty page?\n");
3250
3251                 /* discard the page */
3252                 cl_page_discard(env, io, page);
3253                 cl_page_disown(env, io, page);
3254         } else {
3255                 LASSERT(page->cp_state == CPS_FREEING);
3256         }
3257
3258         return CLP_GANG_OKAY;
3259 }
3260 EXPORT_SYMBOL(osc_discard_cb);
3261
3262 /**
3263  * Discard pages protected by the given lock. This function traverses radix
3264  * tree to find all covering pages and discard them. If a page is being covered
3265  * by other locks, it should remain in cache.
3266  *
3267  * If error happens on any step, the process continues anyway (the reasoning
3268  * behind this being that lock cancellation cannot be delayed indefinitely).
3269  */
3270 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3271                            pgoff_t start, pgoff_t end, bool discard)
3272 {
3273         struct osc_thread_info *info = osc_env_info(env);
3274         struct cl_io *io = osc_env_thread_io(env);
3275         osc_page_gang_cbt cb;
3276         int res;
3277         int result;
3278
3279         ENTRY;
3280
3281         io->ci_obj = cl_object_top(osc2cl(osc));
3282         io->ci_ignore_layout = 1;
3283         result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3284         if (result != 0)
3285                 GOTO(out, result);
3286
3287         cb = discard ? osc_discard_cb : check_and_discard_cb;
3288         info->oti_fn_index = info->oti_next_index = start;
3289         do {
3290                 res = osc_page_gang_lookup(env, io, osc,
3291                                            info->oti_next_index, end, cb, osc);
3292                 if (info->oti_next_index > end)
3293                         break;
3294
3295                 if (res == CLP_GANG_RESCHED)
3296                         cond_resched();
3297         } while (res != CLP_GANG_OKAY);
3298 out:
3299         cl_io_fini(env, io);
3300         RETURN(result);
3301 }
3302
3303
3304 /** @} osc */