lustre/osc/osc_cache.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, 2015, Intel Corporation.
  31  *
  32  */
  33 /*
  34  * This file is part of Lustre, http://www.lustre.org/
  35  * Lustre is a trademark of Sun Microsystems, Inc.
  36  *
  37  * osc cache management.
  38  *
  39  * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_OSC
  43
  44 #include "osc_cl_internal.h"
  45 #include "osc_internal.h"
  46
  47 static int extent_debug; /* set it to be true for more debug */
  48
  49 static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
  50 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
  51                            enum osc_extent_state state);
  52 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
  53                               struct osc_async_page *oap, int sent, int rc);
  54 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
  55                           int cmd);
  56 static int osc_refresh_count(const struct lu_env *env,
  57                              struct osc_async_page *oap, int cmd);
  58 static int osc_io_unplug_async(const struct lu_env *env,
  59                                struct client_obd *cli, struct osc_object *osc);
  60 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
  61                            unsigned int lost_grant, unsigned int dirty_grant);
  62
  63 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
  64                                   const char *func, int line);
  65 #define osc_extent_tree_dump(lvl, obj) \
  66         osc_extent_tree_dump0(lvl, obj, __func__, __LINE__)
  67
  68 static void osc_unreserve_grant(struct client_obd *cli, unsigned int reserved,
  69                                 unsigned int unused);
  70
  71 /** \addtogroup osc
  72  *  @{
  73  */
  74
  75 /* ------------------ osc extent ------------------ */
  76 static inline char *ext_flags(struct osc_extent *ext, char *flags)
  77 {
  78         char *buf = flags;
  79         *buf++ = ext->oe_rw ? 'r' : 'w';
  80         if (ext->oe_intree)
  81                 *buf++ = 'i';
  82         if (ext->oe_sync)
  83                 *buf++ = 'S';
  84         if (ext->oe_srvlock)
  85                 *buf++ = 's';
  86         if (ext->oe_hp)
  87                 *buf++ = 'h';
  88         if (ext->oe_urgent)
  89                 *buf++ = 'u';
  90         if (ext->oe_memalloc)
  91                 *buf++ = 'm';
  92         if (ext->oe_trunc_pending)
  93                 *buf++ = 't';
  94         if (ext->oe_fsync_wait)
  95                 *buf++ = 'Y';
  96         *buf = 0;
  97         return flags;
  98 }
  99
 100 static inline char list_empty_marker(struct list_head *list)
 101 {
 102         return list_empty(list) ? '-' : '+';
 103 }
 104
 105 #define EXTSTR       "[%lu -> %lu/%lu]"
 106 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
 107 static const char *oes_strings[] = {
 108         "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
 109
 110 #define OSC_EXTENT_DUMP(lvl, extent, fmt, ...) do {                           \
 111         struct osc_extent *__ext = (extent);                                  \
 112         char __buf[16];                                                       \
 113                                                                               \
 114         CDEBUG(lvl,                                                           \
 115                 "extent %p@{" EXTSTR ", "                                     \
 116                 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt,          \
 117                 /* ----- extent part 0 ----- */                               \
 118                 __ext, EXTPARA(__ext),                                        \
 119                 /* ----- part 1 ----- */                                      \
 120                 atomic_read(&__ext->oe_refc),                         \
 121                 atomic_read(&__ext->oe_users),                        \
 122                 list_empty_marker(&__ext->oe_link),                           \
 123                 oes_strings[__ext->oe_state], ext_flags(__ext, __buf),        \
 124                 __ext->oe_obj,                                                \
 125                 /* ----- part 2 ----- */                                      \
 126                 __ext->oe_grants, __ext->oe_nr_pages,                         \
 127                 list_empty_marker(&__ext->oe_pages),                          \
 128                 waitqueue_active(&__ext->oe_waitq) ? '+' : '-',               \
 129                 __ext->oe_dlmlock, __ext->oe_mppr, __ext->oe_owner,           \
 130                 /* ----- part 4 ----- */                                      \
 131                 ## __VA_ARGS__);                                              \
 132         if (lvl == D_ERROR && __ext->oe_dlmlock != NULL)                      \
 133                 LDLM_ERROR(__ext->oe_dlmlock, "extent: %p", __ext);           \
 134         else                                                                  \
 135                 LDLM_DEBUG(__ext->oe_dlmlock, "extent: %p", __ext);           \
 136 } while (0)
 137
 138 #undef EASSERTF
 139 #define EASSERTF(expr, ext, fmt, args...) do {                          \
 140         if (!(expr)) {                                                  \
 141                 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args);           \
 142                 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj);           \
 143                 LASSERT(expr);                                          \
 144         }                                                               \
 145 } while (0)
 146
 147 #undef EASSERT
 148 #define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
 149
 150 static inline struct osc_extent *rb_extent(struct rb_node *n)
 151 {
 152         if (n == NULL)
 153                 return NULL;
 154
 155         return container_of(n, struct osc_extent, oe_node);
 156 }
 157
 158 static inline struct osc_extent *next_extent(struct osc_extent *ext)
 159 {
 160         if (ext == NULL)
 161                 return NULL;
 162
 163         LASSERT(ext->oe_intree);
 164         return rb_extent(rb_next(&ext->oe_node));
 165 }
 166
 167 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
 168 {
 169         if (ext == NULL)
 170                 return NULL;
 171
 172         LASSERT(ext->oe_intree);
 173         return rb_extent(rb_prev(&ext->oe_node));
 174 }
 175
 176 static inline struct osc_extent *first_extent(struct osc_object *obj)
 177 {
 178         return rb_extent(rb_first(&obj->oo_root));
 179 }
 180
 181 /* object must be locked by caller. */
 182 static int osc_extent_sanity_check0(struct osc_extent *ext,
 183                                     const char *func, const int line)
 184 {
 185         struct osc_object *obj = ext->oe_obj;
 186         struct osc_async_page *oap;
 187         size_t page_count;
 188         int rc = 0;
 189
 190         if (!osc_object_is_locked(obj))
 191                 GOTO(out, rc = 9);
 192
 193         if (ext->oe_state >= OES_STATE_MAX)
 194                 GOTO(out, rc = 10);
 195
 196         if (atomic_read(&ext->oe_refc) <= 0)
 197                 GOTO(out, rc = 20);
 198
 199         if (atomic_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
 200                 GOTO(out, rc = 30);
 201
 202         switch (ext->oe_state) {
 203         case OES_INV:
 204                 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
 205                         GOTO(out, rc = 35);
 206                 GOTO(out, rc = 0);
 207                 break;
 208         case OES_ACTIVE:
 209                 if (atomic_read(&ext->oe_users) == 0)
 210                         GOTO(out, rc = 40);
 211                 if (ext->oe_hp)
 212                         GOTO(out, rc = 50);
 213                 if (ext->oe_fsync_wait && !ext->oe_urgent)
 214                         GOTO(out, rc = 55);
 215                 break;
 216         case OES_CACHE:
 217                 if (ext->oe_grants == 0)
 218                         GOTO(out, rc = 60);
 219                 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
 220                         GOTO(out, rc = 65);
 221         default:
 222                 if (atomic_read(&ext->oe_users) > 0)
 223                         GOTO(out, rc = 70);
 224         }
 225
 226         if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
 227                 GOTO(out, rc = 80);
 228
 229         if (ext->oe_sync && ext->oe_grants > 0)
 230                 GOTO(out, rc = 90);
 231
 232         if (ext->oe_dlmlock != NULL && !ldlm_is_failed(ext->oe_dlmlock)) {
 233                 struct ldlm_extent *extent;
 234
 235                 extent = &ext->oe_dlmlock->l_policy_data.l_extent;
 236                 if (!(extent->start <= cl_offset(osc2cl(obj), ext->oe_start) &&
 237                       extent->end   >= cl_offset(osc2cl(obj), ext->oe_max_end)))
 238                         GOTO(out, rc = 100);
 239
 240                 if (!(ext->oe_dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP)))
 241                         GOTO(out, rc = 102);
 242         }
 243
 244         if (ext->oe_nr_pages > ext->oe_mppr)
 245                 GOTO(out, rc = 105);
 246
 247         /* Do not verify page list if extent is in RPC. This is because an
 248          * in-RPC extent is supposed to be exclusively accessible w/o lock. */
 249         if (ext->oe_state > OES_CACHE)
 250                 GOTO(out, rc = 0);
 251
 252         if (!extent_debug)
 253                 GOTO(out, rc = 0);
 254
 255         page_count = 0;
 256         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
 257                 pgoff_t index = osc_index(oap2osc(oap));
 258                 ++page_count;
 259                 if (index > ext->oe_end || index < ext->oe_start)
 260                         GOTO(out, rc = 110);
 261         }
 262         if (page_count != ext->oe_nr_pages)
 263                 GOTO(out, rc = 120);
 264
 265 out:
 266         if (rc != 0)
 267                 OSC_EXTENT_DUMP(D_ERROR, ext,
 268                                 "%s:%d sanity check %p failed with rc = %d\n",
 269                                 func, line, ext, rc);
 270         return rc;
 271 }
 272
 273 #define sanity_check_nolock(ext) \
 274         osc_extent_sanity_check0(ext, __func__, __LINE__)
 275
 276 #define sanity_check(ext) ({                                                   \
 277         int __res;                                                             \
 278         osc_object_lock((ext)->oe_obj);                                        \
 279         __res = sanity_check_nolock(ext);                                      \
 280         osc_object_unlock((ext)->oe_obj);                                      \
 281         __res;                                                                 \
 282 })
 283
 284
 285 /**
 286  * sanity check - to make sure there is no overlapped extent in the tree.
 287  */
 288 static int osc_extent_is_overlapped(struct osc_object *obj,
 289                                     struct osc_extent *ext)
 290 {
 291         struct osc_extent *tmp;
 292
 293         LASSERT(osc_object_is_locked(obj));
 294
 295         if (!extent_debug)
 296                 return 0;
 297
 298         for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
 299                 if (tmp == ext)
 300                         continue;
 301                 if (tmp->oe_end >= ext->oe_start &&
 302                     tmp->oe_start <= ext->oe_end)
 303                         return 1;
 304         }
 305         return 0;
 306 }
 307
 308 static void osc_extent_state_set(struct osc_extent *ext, int state)
 309 {
 310         LASSERT(osc_object_is_locked(ext->oe_obj));
 311         LASSERT(state >= OES_INV && state < OES_STATE_MAX);
 312
 313         /* Never try to sanity check a state changing extent :-) */
 314         /* LASSERT(sanity_check_nolock(ext) == 0); */
 315
 316         /* TODO: validate the state machine */
 317         ext->oe_state = state;
 318         wake_up_all(&ext->oe_waitq);
 319 }
 320
 321 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
 322 {
 323         struct osc_extent *ext;
 324
 325         OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_NOFS);
 326         if (ext == NULL)
 327                 return NULL;
 328
 329         RB_CLEAR_NODE(&ext->oe_node);
 330         ext->oe_obj = obj;
 331         cl_object_get(osc2cl(obj));
 332         atomic_set(&ext->oe_refc, 1);
 333         atomic_set(&ext->oe_users, 0);
 334         INIT_LIST_HEAD(&ext->oe_link);
 335         ext->oe_state = OES_INV;
 336         INIT_LIST_HEAD(&ext->oe_pages);
 337         init_waitqueue_head(&ext->oe_waitq);
 338         ext->oe_dlmlock = NULL;
 339
 340         return ext;
 341 }
 342
 343 static void osc_extent_free(struct osc_extent *ext)
 344 {
 345         OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
 346 }
 347
 348 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
 349 {
 350         LASSERT(atomic_read(&ext->oe_refc) >= 0);
 351         atomic_inc(&ext->oe_refc);
 352         return ext;
 353 }
 354
 355 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
 356 {
 357         LASSERT(atomic_read(&ext->oe_refc) > 0);
 358         if (atomic_dec_and_test(&ext->oe_refc)) {
 359                 LASSERT(list_empty(&ext->oe_link));
 360                 LASSERT(atomic_read(&ext->oe_users) == 0);
 361                 LASSERT(ext->oe_state == OES_INV);
 362                 LASSERT(!ext->oe_intree);
 363
 364                 if (ext->oe_dlmlock != NULL) {
 365                         lu_ref_add(&ext->oe_dlmlock->l_reference,
 366                                    "osc_extent", ext);
 367                         LDLM_LOCK_PUT(ext->oe_dlmlock);
 368                         ext->oe_dlmlock = NULL;
 369                 }
 370                 cl_object_put(env, osc2cl(ext->oe_obj));
 371                 osc_extent_free(ext);
 372         }
 373 }
 374
 375 /**
 376  * osc_extent_put_trust() is a special version of osc_extent_put() when
 377  * it's known that the caller is not the last user. This is to address the
 378  * problem of lacking of lu_env ;-).
 379  */
 380 static void osc_extent_put_trust(struct osc_extent *ext)
 381 {
 382         LASSERT(atomic_read(&ext->oe_refc) > 1);
 383         LASSERT(osc_object_is_locked(ext->oe_obj));
 384         atomic_dec(&ext->oe_refc);
 385 }
 386
 387 /**
 388  * Return the extent which includes pgoff @index, or return the greatest
 389  * previous extent in the tree.
 390  */
 391 static struct osc_extent *osc_extent_search(struct osc_object *obj,
 392                                             pgoff_t index)
 393 {
 394         struct rb_node    *n = obj->oo_root.rb_node;
 395         struct osc_extent *tmp, *p = NULL;
 396
 397         LASSERT(osc_object_is_locked(obj));
 398         while (n != NULL) {
 399                 tmp = rb_extent(n);
 400                 if (index < tmp->oe_start) {
 401                         n = n->rb_left;
 402                 } else if (index > tmp->oe_end) {
 403                         p = rb_extent(n);
 404                         n = n->rb_right;
 405                 } else {
 406                         return tmp;
 407                 }
 408         }
 409         return p;
 410 }
 411
 412 /*
 413  * Return the extent covering @index, otherwise return NULL.
 414  * caller must have held object lock.
 415  */
 416 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
 417                                             pgoff_t index)
 418 {
 419         struct osc_extent *ext;
 420
 421         ext = osc_extent_search(obj, index);
 422         if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
 423                 return osc_extent_get(ext);
 424         return NULL;
 425 }
 426
 427 /* caller must have held object lock. */
 428 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
 429 {
 430         struct rb_node   **n      = &obj->oo_root.rb_node;
 431         struct rb_node    *parent = NULL;
 432         struct osc_extent *tmp;
 433
 434         LASSERT(ext->oe_intree == 0);
 435         LASSERT(ext->oe_obj == obj);
 436         LASSERT(osc_object_is_locked(obj));
 437         while (*n != NULL) {
 438                 tmp = rb_extent(*n);
 439                 parent = *n;
 440
 441                 if (ext->oe_end < tmp->oe_start)
 442                         n = &(*n)->rb_left;
 443                 else if (ext->oe_start > tmp->oe_end)
 444                         n = &(*n)->rb_right;
 445                 else
 446                         EASSERTF(0, tmp, EXTSTR"\n", EXTPARA(ext));
 447         }
 448         rb_link_node(&ext->oe_node, parent, n);
 449         rb_insert_color(&ext->oe_node, &obj->oo_root);
 450         osc_extent_get(ext);
 451         ext->oe_intree = 1;
 452 }
 453
 454 /* caller must have held object lock. */
 455 static void osc_extent_erase(struct osc_extent *ext)
 456 {
 457         struct osc_object *obj = ext->oe_obj;
 458         LASSERT(osc_object_is_locked(obj));
 459         if (ext->oe_intree) {
 460                 rb_erase(&ext->oe_node, &obj->oo_root);
 461                 ext->oe_intree = 0;
 462                 /* rbtree held a refcount */
 463                 osc_extent_put_trust(ext);
 464         }
 465 }
 466
 467 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
 468 {
 469         struct osc_object *obj = ext->oe_obj;
 470
 471         LASSERT(osc_object_is_locked(obj));
 472         LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
 473         if (ext->oe_state == OES_CACHE) {
 474                 osc_extent_state_set(ext, OES_ACTIVE);
 475                 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
 476         }
 477         atomic_inc(&ext->oe_users);
 478         list_del_init(&ext->oe_link);
 479         return osc_extent_get(ext);
 480 }
 481
 482 static void __osc_extent_remove(struct osc_extent *ext)
 483 {
 484         LASSERT(osc_object_is_locked(ext->oe_obj));
 485         LASSERT(list_empty(&ext->oe_pages));
 486         osc_extent_erase(ext);
 487         list_del_init(&ext->oe_link);
 488         osc_extent_state_set(ext, OES_INV);
 489         OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
 490 }
 491
 492 static void osc_extent_remove(struct osc_extent *ext)
 493 {
 494         struct osc_object *obj = ext->oe_obj;
 495
 496         osc_object_lock(obj);
 497         __osc_extent_remove(ext);
 498         osc_object_unlock(obj);
 499 }
 500
 501 /**
 502  * This function is used to merge extents to get better performance. It checks
 503  * if @cur and @victim are contiguous at block level.
 504  */
 505 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
 506                             struct osc_extent *victim)
 507 {
 508         struct osc_object       *obj = cur->oe_obj;
 509         struct client_obd       *cli = osc_cli(obj);
 510         pgoff_t                  chunk_start;
 511         pgoff_t                  chunk_end;
 512         int                      ppc_bits;
 513
 514         LASSERT(cur->oe_state == OES_CACHE);
 515         LASSERT(osc_object_is_locked(obj));
 516         if (victim == NULL)
 517                 return -EINVAL;
 518
 519         if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
 520                 return -EBUSY;
 521
 522         if (cur->oe_max_end != victim->oe_max_end)
 523                 return -ERANGE;
 524
 525         LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
 526         ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_CACHE_SHIFT;
 527         chunk_start = cur->oe_start >> ppc_bits;
 528         chunk_end   = cur->oe_end   >> ppc_bits;
 529         if (chunk_start   != (victim->oe_end >> ppc_bits) + 1 &&
 530             chunk_end + 1 != victim->oe_start >> ppc_bits)
 531                 return -ERANGE;
 532
 533         /* overall extent size should not exceed the max supported limit
 534          * reported by the server */
 535         if (cur->oe_end - cur->oe_start + 1 +
 536             victim->oe_end - victim->oe_start + 1 > cli->cl_max_extent_pages)
 537                 return -ERANGE;
 538
 539         OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
 540
 541         cur->oe_start     = min(cur->oe_start, victim->oe_start);
 542         cur->oe_end       = max(cur->oe_end,   victim->oe_end);
 543         /* per-extent tax should be accounted only once for the whole extent */
 544         cur->oe_grants   += victim->oe_grants - cli->cl_grant_extent_tax;
 545         cur->oe_nr_pages += victim->oe_nr_pages;
 546         /* only the following bits are needed to merge */
 547         cur->oe_urgent   |= victim->oe_urgent;
 548         cur->oe_memalloc |= victim->oe_memalloc;
 549         list_splice_init(&victim->oe_pages, &cur->oe_pages);
 550         list_del_init(&victim->oe_link);
 551         victim->oe_nr_pages = 0;
 552
 553         osc_extent_get(victim);
 554         __osc_extent_remove(victim);
 555         osc_extent_put(env, victim);
 556
 557         OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
 558         return 0;
 559 }
 560
 561 /**
 562  * Drop user count of osc_extent, and unplug IO asynchronously.
 563  */
 564 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
 565 {
 566         struct osc_object *obj = ext->oe_obj;
 567         struct client_obd *cli = osc_cli(obj);
 568         int rc = 0;
 569         ENTRY;
 570
 571         LASSERT(atomic_read(&ext->oe_users) > 0);
 572         LASSERT(sanity_check(ext) == 0);
 573         LASSERT(ext->oe_grants > 0);
 574
 575         if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
 576                 LASSERT(ext->oe_state == OES_ACTIVE);
 577                 if (ext->oe_trunc_pending) {
 578                         /* a truncate process is waiting for this extent.
 579                          * This may happen due to a race, check
 580                          * osc_cache_truncate_start(). */
 581                         osc_extent_state_set(ext, OES_TRUNC);
 582                         ext->oe_trunc_pending = 0;
 583                 } else {
 584                         int grant = 0;
 585
 586                         osc_extent_state_set(ext, OES_CACHE);
 587                         osc_update_pending(obj, OBD_BRW_WRITE,
 588                                            ext->oe_nr_pages);
 589
 590                         /* try to merge the previous and next extent. */
 591                         if (osc_extent_merge(env, ext, prev_extent(ext)) == 0)
 592                                 grant += cli->cl_grant_extent_tax;
 593                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 594                                 grant += cli->cl_grant_extent_tax;
 595                         if (grant > 0)
 596                                 osc_unreserve_grant(cli, 0, grant);
 597
 598                         if (ext->oe_urgent)
 599                                 list_move_tail(&ext->oe_link,
 600                                                &obj->oo_urgent_exts);
 601                 }
 602                 osc_object_unlock(obj);
 603
 604                 osc_io_unplug_async(env, cli, obj);
 605         }
 606         osc_extent_put(env, ext);
 607         RETURN(rc);
 608 }
 609
 610 static inline int overlapped(struct osc_extent *ex1, struct osc_extent *ex2)
 611 {
 612         return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
 613 }
 614
 615 /**
 616  * Find or create an extent which includes @index, core function to manage
 617  * extent tree.
 618  */
 619 static struct osc_extent *osc_extent_find(const struct lu_env *env,
 620                                           struct osc_object *obj, pgoff_t index,
 621                                           unsigned int *grants)
 622 {
 623         struct client_obd *cli = osc_cli(obj);
 624         struct osc_lock   *olck;
 625         struct cl_lock_descr *descr;
 626         struct osc_extent *cur;
 627         struct osc_extent *ext;
 628         struct osc_extent *conflict = NULL;
 629         struct osc_extent *found = NULL;
 630         pgoff_t    chunk;
 631         pgoff_t    max_end;
 632         unsigned int max_pages; /* max_pages_per_rpc */
 633         unsigned int chunksize;
 634         int        ppc_bits; /* pages per chunk bits */
 635         pgoff_t    chunk_mask;
 636         int        rc;
 637         ENTRY;
 638
 639         cur = osc_extent_alloc(obj);
 640         if (cur == NULL)
 641                 RETURN(ERR_PTR(-ENOMEM));
 642
 643         olck = osc_env_io(env)->oi_write_osclock;
 644         LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
 645         LASSERT(olck->ols_state == OLS_GRANTED);
 646
 647         descr = &olck->ols_cl.cls_lock->cll_descr;
 648         LASSERT(descr->cld_mode >= CLM_WRITE);
 649
 650         LASSERTF(cli->cl_chunkbits >= PAGE_SHIFT,
 651                  "chunkbits: %u\n", cli->cl_chunkbits);
 652         ppc_bits   = cli->cl_chunkbits - PAGE_SHIFT;
 653         chunk_mask = ~((1 << ppc_bits) - 1);
 654         chunksize  = 1 << cli->cl_chunkbits;
 655         chunk      = index >> ppc_bits;
 656
 657         /* align end to RPC edge. */
 658         max_pages = cli->cl_max_pages_per_rpc;
 659         if ((max_pages & ~chunk_mask) != 0) {
 660                 CERROR("max_pages: %#x chunkbits: %u chunk_mask: %#lx\n",
 661                        max_pages, cli->cl_chunkbits, chunk_mask);
 662                 RETURN(ERR_PTR(-EINVAL));
 663         }
 664         max_end = index - (index % max_pages) + max_pages - 1;
 665         max_end = min_t(pgoff_t, max_end, descr->cld_end);
 666
 667         /* initialize new extent by parameters so far */
 668         cur->oe_max_end = max_end;
 669         cur->oe_start   = index & chunk_mask;
 670         cur->oe_end     = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
 671         if (cur->oe_start < descr->cld_start)
 672                 cur->oe_start = descr->cld_start;
 673         if (cur->oe_end > max_end)
 674                 cur->oe_end = max_end;
 675         cur->oe_grants  = 0;
 676         cur->oe_mppr    = max_pages;
 677         if (olck->ols_dlmlock != NULL) {
 678                 LASSERT(olck->ols_hold);
 679                 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
 680                 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
 681         }
 682
 683         /* grants has been allocated by caller */
 684         LASSERTF(*grants >= chunksize + cli->cl_grant_extent_tax,
 685                  "%u/%u/%u.\n", *grants, chunksize, cli->cl_grant_extent_tax);
 686         LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
 687                  EXTPARA(cur));
 688
 689 restart:
 690         osc_object_lock(obj);
 691         ext = osc_extent_search(obj, cur->oe_start);
 692         if (ext == NULL)
 693                 ext = first_extent(obj);
 694         while (ext != NULL) {
 695                 pgoff_t ext_chk_start = ext->oe_start >> ppc_bits;
 696                 pgoff_t ext_chk_end   = ext->oe_end   >> ppc_bits;
 697
 698                 LASSERT(sanity_check_nolock(ext) == 0);
 699                 if (chunk > ext_chk_end + 1)
 700                         break;
 701
 702                 /* if covering by different locks, no chance to match */
 703                 if (olck->ols_dlmlock != ext->oe_dlmlock) {
 704                         EASSERTF(!overlapped(ext, cur), ext,
 705                                  EXTSTR"\n", EXTPARA(cur));
 706
 707                         ext = next_extent(ext);
 708                         continue;
 709                 }
 710
 711                 /* discontiguous chunks? */
 712                 if (chunk + 1 < ext_chk_start) {
 713                         ext = next_extent(ext);
 714                         continue;
 715                 }
 716
 717                 /* ok, from now on, ext and cur have these attrs:
 718                  * 1. covered by the same lock
 719                  * 2. contiguous at chunk level or overlapping. */
 720
 721                 if (overlapped(ext, cur)) {
 722                         /* cur is the minimum unit, so overlapping means
 723                          * full contain. */
 724                         EASSERTF((ext->oe_start <= cur->oe_start &&
 725                                   ext->oe_end >= cur->oe_end),
 726                                  ext, EXTSTR"\n", EXTPARA(cur));
 727
 728                         if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
 729                                 /* for simplicity, we wait for this extent to
 730                                  * finish before going forward. */
 731                                 conflict = osc_extent_get(ext);
 732                                 break;
 733                         }
 734
 735                         found = osc_extent_hold(ext);
 736                         break;
 737                 }
 738
 739                 /* non-overlapped extent */
 740                 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
 741                         /* we can't do anything for a non OES_CACHE extent, or
 742                          * if there is someone waiting for this extent to be
 743                          * flushed, try next one. */
 744                         ext = next_extent(ext);
 745                         continue;
 746                 }
 747
 748                 /* check if they belong to the same rpc slot before trying to
 749                  * merge. the extents are not overlapped and contiguous at
 750                  * chunk level to get here. */
 751                 if (ext->oe_max_end != max_end) {
 752                         /* if they don't belong to the same RPC slot or
 753                          * max_pages_per_rpc has ever changed, do not merge. */
 754                         ext = next_extent(ext);
 755                         continue;
 756                 }
 757
 758                 /* check whether maximum extent size will be hit */
 759                 if ((ext_chk_end - ext_chk_start + 1 + 1) << ppc_bits >
 760                     cli->cl_max_extent_pages) {
 761                         ext = next_extent(ext);
 762                         continue;
 763                 }
 764
 765                 /* it's required that an extent must be contiguous at chunk
 766                  * level so that we know the whole extent is covered by grant
 767                  * (the pages in the extent are NOT required to be contiguous).
 768                  * Otherwise, it will be too much difficult to know which
 769                  * chunks have grants allocated. */
 770
 771                 /* try to do front merge - extend ext's start */
 772                 if (chunk + 1 == ext_chk_start) {
 773                         /* ext must be chunk size aligned */
 774                         EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
 775
 776                         /* pull ext's start back to cover cur */
 777                         ext->oe_start   = cur->oe_start;
 778                         ext->oe_grants += chunksize;
 779                         *grants -= chunksize;
 780
 781                         found = osc_extent_hold(ext);
 782                 } else if (chunk == ext_chk_end + 1) {
 783                         /* rear merge */
 784                         ext->oe_end     = cur->oe_end;
 785                         ext->oe_grants += chunksize;
 786                         *grants -= chunksize;
 787
 788                         /* try to merge with the next one because we just fill
 789                          * in a gap */
 790                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 791                                 /* we can save extent tax from next extent */
 792                                 *grants += cli->cl_grant_extent_tax;
 793
 794                         found = osc_extent_hold(ext);
 795                 }
 796                 if (found != NULL)
 797                         break;
 798
 799                 ext = next_extent(ext);
 800         }
 801
 802         osc_extent_tree_dump(D_CACHE, obj);
 803         if (found != NULL) {
 804                 LASSERT(conflict == NULL);
 805                 if (!IS_ERR(found)) {
 806                         LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
 807                         OSC_EXTENT_DUMP(D_CACHE, found,
 808                                         "found caching ext for %lu.\n", index);
 809                 }
 810         } else if (conflict == NULL) {
 811                 /* create a new extent */
 812                 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
 813                 cur->oe_grants = chunksize + cli->cl_grant_extent_tax;
 814                 *grants -= cur->oe_grants;
 815                 LASSERT(*grants >= 0);
 816
 817                 cur->oe_state = OES_CACHE;
 818                 found = osc_extent_hold(cur);
 819                 osc_extent_insert(obj, cur);
 820                 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
 821                                 index, descr->cld_end);
 822         }
 823         osc_object_unlock(obj);
 824
 825         if (conflict != NULL) {
 826                 LASSERT(found == NULL);
 827
 828                 /* waiting for IO to finish. Please notice that it's impossible
 829                  * to be an OES_TRUNC extent. */
 830                 rc = osc_extent_wait(env, conflict, OES_INV);
 831                 osc_extent_put(env, conflict);
 832                 conflict = NULL;
 833                 if (rc < 0)
 834                         GOTO(out, found = ERR_PTR(rc));
 835
 836                 goto restart;
 837         }
 838         EXIT;
 839
 840 out:
 841         osc_extent_put(env, cur);
 842         LASSERT(*grants >= 0);
 843         return found;
 844 }
 845
 846 /**
 847  * Called when IO is finished to an extent.
 848  */
 849 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
 850                       int sent, int rc)
 851 {
 852         struct client_obd *cli = osc_cli(ext->oe_obj);
 853         struct osc_async_page *oap;
 854         struct osc_async_page *tmp;
 855         int nr_pages = ext->oe_nr_pages;
 856         int lost_grant = 0;
 857         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
 858         loff_t last_off = 0;
 859         int last_count = -1;
 860         ENTRY;
 861
 862         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
 863
 864         ext->oe_rc = rc ?: ext->oe_nr_pages;
 865         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
 866
 867         osc_lru_add_batch(cli, &ext->oe_pages);
 868         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 869                                      oap_pending_item) {
 870                 list_del_init(&oap->oap_rpc_item);
 871                 list_del_init(&oap->oap_pending_item);
 872                 if (last_off <= oap->oap_obj_off) {
 873                         last_off = oap->oap_obj_off;
 874                         last_count = oap->oap_count;
 875                 }
 876
 877                 --ext->oe_nr_pages;
 878                 osc_ap_completion(env, cli, oap, sent, rc);
 879         }
 880         EASSERT(ext->oe_nr_pages == 0, ext);
 881
 882         if (!sent) {
 883                 lost_grant = ext->oe_grants;
 884         } else if (blocksize < PAGE_CACHE_SIZE &&
 885                    last_count != PAGE_CACHE_SIZE) {
 886                 /* For short writes we shouldn't count parts of pages that
 887                  * span a whole chunk on the OST side, or our accounting goes
 888                  * wrong.  Should match the code in filter_grant_check. */
 889                 int offset = last_off & ~PAGE_MASK;
 890                 int count = last_count + (offset & (blocksize - 1));
 891                 int end = (offset + last_count) & (blocksize - 1);
 892                 if (end)
 893                         count += blocksize - end;
 894
 895                 lost_grant = PAGE_CACHE_SIZE - count;
 896         }
 897         if (ext->oe_grants > 0)
 898                 osc_free_grant(cli, nr_pages, lost_grant, ext->oe_grants);
 899
 900         osc_extent_remove(ext);
 901         /* put the refcount for RPC */
 902         osc_extent_put(env, ext);
 903         RETURN(0);
 904 }
 905
 906 static int extent_wait_cb(struct osc_extent *ext, enum osc_extent_state state)
 907 {
 908         int ret;
 909
 910         osc_object_lock(ext->oe_obj);
 911         ret = ext->oe_state == state;
 912         osc_object_unlock(ext->oe_obj);
 913
 914         return ret;
 915 }
 916
 917 /**
 918  * Wait for the extent's state to become @state.
 919  */
 920 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
 921                            enum osc_extent_state state)
 922 {
 923         struct osc_object *obj = ext->oe_obj;
 924         struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
 925                                                   LWI_ON_SIGNAL_NOOP, NULL);
 926         int rc = 0;
 927         ENTRY;
 928
 929         osc_object_lock(obj);
 930         LASSERT(sanity_check_nolock(ext) == 0);
 931         /* `Kick' this extent only if the caller is waiting for it to be
 932          * written out. */
 933         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
 934                 if (ext->oe_state == OES_ACTIVE) {
 935                         ext->oe_urgent = 1;
 936                 } else if (ext->oe_state == OES_CACHE) {
 937                         ext->oe_urgent = 1;
 938                         osc_extent_hold(ext);
 939                         rc = 1;
 940                 }
 941         }
 942         osc_object_unlock(obj);
 943         if (rc == 1)
 944                 osc_extent_release(env, ext);
 945
 946         /* wait for the extent until its state becomes @state */
 947         rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
 948         if (rc == -ETIMEDOUT) {
 949                 OSC_EXTENT_DUMP(D_ERROR, ext,
 950                         "%s: wait ext to %u timedout, recovery in progress?\n",
 951                         cli_name(osc_cli(obj)), state);
 952
 953                 lwi = LWI_INTR(NULL, NULL);
 954                 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state),
 955                                   &lwi);
 956         }
 957         if (rc == 0 && ext->oe_rc < 0)
 958                 rc = ext->oe_rc;
 959         RETURN(rc);
 960 }
 961
 962 /**
 963  * Discard pages with index greater than @size. If @ext is overlapped with
 964  * @size, then partial truncate happens.
 965  */
 966 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
 967                                 bool partial)
 968 {
 969         struct lu_env         *env;
 970         struct cl_io          *io;
 971         struct osc_object     *obj = ext->oe_obj;
 972         struct client_obd     *cli = osc_cli(obj);
 973         struct osc_async_page *oap;
 974         struct osc_async_page *tmp;
 975         int                    pages_in_chunk = 0;
 976         int                    ppc_bits    = cli->cl_chunkbits -
 977                                              PAGE_CACHE_SHIFT;
 978         __u64                  trunc_chunk = trunc_index >> ppc_bits;
 979         int                    grants   = 0;
 980         int                    nr_pages = 0;
 981         int                    rc       = 0;
 982         __u16                  refcheck;
 983         ENTRY;
 984
 985         LASSERT(sanity_check(ext) == 0);
 986         LASSERT(ext->oe_state == OES_TRUNC);
 987         LASSERT(!ext->oe_urgent);
 988
 989         /* Request new lu_env.
 990          * We can't use that env from osc_cache_truncate_start() because
 991          * it's from lov_io_sub and not fully initialized. */
 992         env = cl_env_get(&refcheck);
 993         io  = &osc_env_info(env)->oti_io;
 994         io->ci_obj = cl_object_top(osc2cl(obj));
 995         io->ci_ignore_layout = 1;
 996         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
 997         if (rc < 0)
 998                 GOTO(out, rc);
 999
1000         /* discard all pages with index greater than trunc_index */
1001         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
1002                                      oap_pending_item) {
1003                 pgoff_t index = osc_index(oap2osc(oap));
1004                 struct cl_page  *page = oap2cl_page(oap);
1005
1006                 LASSERT(list_empty(&oap->oap_rpc_item));
1007
1008                 /* only discard the pages with their index greater than
1009                  * trunc_index, and ... */
1010                 if (index < trunc_index ||
1011                     (index == trunc_index && partial)) {
1012                         /* accounting how many pages remaining in the chunk
1013                          * so that we can calculate grants correctly. */
1014                         if (index >> ppc_bits == trunc_chunk)
1015                                 ++pages_in_chunk;
1016                         continue;
1017                 }
1018
1019                 list_del_init(&oap->oap_pending_item);
1020
1021                 cl_page_get(page);
1022                 lu_ref_add(&page->cp_reference, "truncate", current);
1023
1024                 if (cl_page_own(env, io, page) == 0) {
1025                         cl_page_discard(env, io, page);
1026                         cl_page_disown(env, io, page);
1027                 } else {
1028                         LASSERT(page->cp_state == CPS_FREEING);
1029                         LASSERT(0);
1030                 }
1031
1032                 lu_ref_del(&page->cp_reference, "truncate", current);
1033                 cl_page_put(env, page);
1034
1035                 --ext->oe_nr_pages;
1036                 ++nr_pages;
1037         }
1038         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1039                       ext->oe_nr_pages == 0),
1040                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1041
1042         osc_object_lock(obj);
1043         if (ext->oe_nr_pages == 0) {
1044                 LASSERT(pages_in_chunk == 0);
1045                 grants = ext->oe_grants;
1046                 ext->oe_grants = 0;
1047         } else { /* calculate how many grants we can free */
1048                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1049                 pgoff_t last_index;
1050
1051
1052                 /* if there is no pages in this chunk, we can also free grants
1053                  * for the last chunk */
1054                 if (pages_in_chunk == 0) {
1055                         /* if this is the 1st chunk and no pages in this chunk,
1056                          * ext->oe_nr_pages must be zero, so we should be in
1057                          * the other if-clause. */
1058                         LASSERT(trunc_chunk > 0);
1059                         --trunc_chunk;
1060                         ++chunks;
1061                 }
1062
1063                 /* this is what we can free from this extent */
1064                 grants          = chunks << cli->cl_chunkbits;
1065                 ext->oe_grants -= grants;
1066                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1067                 ext->oe_end     = min(last_index, ext->oe_max_end);
1068                 LASSERT(ext->oe_end >= ext->oe_start);
1069                 LASSERT(ext->oe_grants > 0);
1070         }
1071         osc_object_unlock(obj);
1072
1073         if (grants > 0 || nr_pages > 0)
1074                 osc_free_grant(cli, nr_pages, grants, grants);
1075
1076 out:
1077         cl_io_fini(env, io);
1078         cl_env_put(env, &refcheck);
1079         RETURN(rc);
1080 }
1081
1082 /**
1083  * This function is used to make the extent prepared for transfer.
1084  * A race with flusing page - ll_writepage() has to be handled cautiously.
1085  */
1086 static int osc_extent_make_ready(const struct lu_env *env,
1087                                  struct osc_extent *ext)
1088 {
1089         struct osc_async_page *oap;
1090         struct osc_async_page *last = NULL;
1091         struct osc_object *obj = ext->oe_obj;
1092         unsigned int page_count = 0;
1093         int rc;
1094         ENTRY;
1095
1096         /* we're going to grab page lock, so object lock must not be taken. */
1097         LASSERT(sanity_check(ext) == 0);
1098         /* in locking state, any process should not touch this extent. */
1099         EASSERT(ext->oe_state == OES_LOCKING, ext);
1100         EASSERT(ext->oe_owner != NULL, ext);
1101
1102         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1103
1104         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1105                 ++page_count;
1106                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1107                         last = oap;
1108
1109                 /* checking ASYNC_READY is race safe */
1110                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1111                         continue;
1112
1113                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1114                 switch (rc) {
1115                 case 0:
1116                         spin_lock(&oap->oap_lock);
1117                         oap->oap_async_flags |= ASYNC_READY;
1118                         spin_unlock(&oap->oap_lock);
1119                         break;
1120                 case -EALREADY:
1121                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1122                         break;
1123                 default:
1124                         LASSERTF(0, "unknown return code: %d\n", rc);
1125                 }
1126         }
1127
1128         LASSERT(page_count == ext->oe_nr_pages);
1129         LASSERT(last != NULL);
1130         /* the last page is the only one we need to refresh its count by
1131          * the size of file. */
1132         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1133                 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1134                 LASSERT(last_oap_count > 0);
1135                 LASSERT(last->oap_page_off + last_oap_count <= PAGE_CACHE_SIZE);
1136                 last->oap_count = last_oap_count;
1137                 spin_lock(&last->oap_lock);
1138                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1139                 spin_unlock(&last->oap_lock);
1140         }
1141
1142         /* for the rest of pages, we don't need to call osf_refresh_count()
1143          * because it's known they are not the last page */
1144         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1145                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1146                         oap->oap_count = PAGE_CACHE_SIZE - oap->oap_page_off;
1147                         spin_lock(&oap->oap_lock);
1148                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1149                         spin_unlock(&oap->oap_lock);
1150                 }
1151         }
1152
1153         osc_object_lock(obj);
1154         osc_extent_state_set(ext, OES_RPC);
1155         osc_object_unlock(obj);
1156         /* get a refcount for RPC. */
1157         osc_extent_get(ext);
1158
1159         RETURN(0);
1160 }
1161
1162 /**
1163  * Quick and simple version of osc_extent_find(). This function is frequently
1164  * called to expand the extent for the same IO. To expand the extent, the
1165  * page index must be in the same or next chunk of ext->oe_end.
1166  */
1167 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index,
1168                              unsigned int *grants)
1169 {
1170         struct osc_object *obj = ext->oe_obj;
1171         struct client_obd *cli = osc_cli(obj);
1172         struct osc_extent *next;
1173         int ppc_bits = cli->cl_chunkbits - PAGE_CACHE_SHIFT;
1174         pgoff_t chunk = index >> ppc_bits;
1175         pgoff_t end_chunk;
1176         pgoff_t end_index;
1177         unsigned int chunksize = 1 << cli->cl_chunkbits;
1178         int rc = 0;
1179         ENTRY;
1180
1181         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1182         osc_object_lock(obj);
1183         LASSERT(sanity_check_nolock(ext) == 0);
1184         end_chunk = ext->oe_end >> ppc_bits;
1185         if (chunk > end_chunk + 1)
1186                 GOTO(out, rc = -ERANGE);
1187
1188         if (end_chunk >= chunk)
1189                 GOTO(out, rc = 0);
1190
1191         LASSERT(end_chunk + 1 == chunk);
1192
1193         /* try to expand this extent to cover @index */
1194         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1195
1196         /* don't go over the maximum extent size reported by server */
1197         if (end_index - ext->oe_start + 1 > cli->cl_max_extent_pages)
1198                 GOTO(out, rc = -ERANGE);
1199
1200         next = next_extent(ext);
1201         if (next != NULL && next->oe_start <= end_index)
1202                 /* complex mode - overlapped with the next extent,
1203                  * this case will be handled by osc_extent_find() */
1204                 GOTO(out, rc = -EAGAIN);
1205
1206         ext->oe_end = end_index;
1207         ext->oe_grants += chunksize;
1208         *grants -= chunksize;
1209         LASSERT(*grants >= 0);
1210         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1211                  "overlapped after expanding for %lu.\n", index);
1212         EXIT;
1213
1214 out:
1215         osc_object_unlock(obj);
1216         RETURN(rc);
1217 }
1218
1219 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
1220                                   const char *func, int line)
1221 {
1222         struct osc_extent *ext;
1223         int cnt;
1224
1225         CDEBUG(level, "Dump object %p extents at %s:%d, mppr: %u.\n",
1226                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1227
1228         /* osc_object_lock(obj); */
1229         cnt = 1;
1230         for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1231                 OSC_EXTENT_DUMP(level, ext, "in tree %d.\n", cnt++);
1232
1233         cnt = 1;
1234         list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1235                 OSC_EXTENT_DUMP(level, ext, "hp %d.\n", cnt++);
1236
1237         cnt = 1;
1238         list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1239                 OSC_EXTENT_DUMP(level, ext, "urgent %d.\n", cnt++);
1240
1241         cnt = 1;
1242         list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1243                 OSC_EXTENT_DUMP(level, ext, "reading %d.\n", cnt++);
1244         /* osc_object_unlock(obj); */
1245 }
1246
1247 /* ------------------ osc extent end ------------------ */
1248
1249 static inline int osc_is_ready(struct osc_object *osc)
1250 {
1251         return !list_empty(&osc->oo_ready_item) ||
1252                !list_empty(&osc->oo_hp_ready_item);
1253 }
1254
1255 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1256         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1257                (OSC), osc_is_ready(OSC),                                       \
1258                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1259                list_empty_marker(&(OSC)->oo_ready_item),                       \
1260                atomic_read(&(OSC)->oo_nr_writes),                              \
1261                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1262                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1263                atomic_read(&(OSC)->oo_nr_reads),                               \
1264                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1265                ##args)
1266
1267 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1268                           int cmd)
1269 {
1270         struct osc_page *opg  = oap2osc_page(oap);
1271         struct cl_page  *page = oap2cl_page(oap);
1272         int result;
1273
1274         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1275
1276         ENTRY;
1277         result = cl_page_make_ready(env, page, CRT_WRITE);
1278         if (result == 0)
1279                 opg->ops_submit_time = cfs_time_current();
1280         RETURN(result);
1281 }
1282
1283 static int osc_refresh_count(const struct lu_env *env,
1284                              struct osc_async_page *oap, int cmd)
1285 {
1286         struct osc_page  *opg = oap2osc_page(oap);
1287         pgoff_t index = osc_index(oap2osc(oap));
1288         struct cl_object *obj;
1289         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
1290
1291         int result;
1292         loff_t kms;
1293
1294         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1295         LASSERT(!(cmd & OBD_BRW_READ));
1296         LASSERT(opg != NULL);
1297         obj = opg->ops_cl.cpl_obj;
1298
1299         cl_object_attr_lock(obj);
1300         result = cl_object_attr_get(env, obj, attr);
1301         cl_object_attr_unlock(obj);
1302         if (result < 0)
1303                 return result;
1304         kms = attr->cat_kms;
1305         if (cl_offset(obj, index) >= kms)
1306                 /* catch race with truncate */
1307                 return 0;
1308         else if (cl_offset(obj, index + 1) > kms)
1309                 /* catch sub-page write at end of file */
1310                 return kms % PAGE_CACHE_SIZE;
1311         else
1312                 return PAGE_CACHE_SIZE;
1313 }
1314
1315 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1316                           int cmd, int rc)
1317 {
1318         struct osc_page   *opg  = oap2osc_page(oap);
1319         struct cl_page    *page = oap2cl_page(oap);
1320         enum cl_req_type   crt;
1321         int srvlock;
1322
1323         ENTRY;
1324
1325         cmd &= ~OBD_BRW_NOQUOTA;
1326         LASSERTF(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ),
1327                  "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1328         LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1329                 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1330         LASSERT(opg->ops_transfer_pinned);
1331
1332         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1333         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1334         opg->ops_transfer_pinned = 0;
1335
1336         opg->ops_submit_time = 0;
1337         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1338
1339         /* statistic */
1340         if (rc == 0 && srvlock) {
1341                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1342                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1343                 size_t bytes = oap->oap_count;
1344
1345                 if (crt == CRT_READ)
1346                         stats->os_lockless_reads += bytes;
1347                 else
1348                         stats->os_lockless_writes += bytes;
1349         }
1350
1351         /*
1352          * This has to be the last operation with the page, as locks are
1353          * released in cl_page_completion() and nothing except for the
1354          * reference counter protects page from concurrent reclaim.
1355          */
1356         lu_ref_del(&page->cp_reference, "transfer", page);
1357
1358         cl_page_completion(env, page, crt, rc);
1359         cl_page_put(env, page);
1360
1361         RETURN(0);
1362 }
1363
1364 #define OSC_DUMP_GRANT(lvl, cli, fmt, args...) do {                     \
1365         struct client_obd *__tmp = (cli);                               \
1366         CDEBUG(lvl, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu "  \
1367                "dropped: %ld avail: %ld, dirty_grant: %ld, "            \
1368                "reserved: %ld, flight: %d } lru {in list: %ld, "        \
1369                "left: %ld, waiters: %d }" fmt "\n",                     \
1370                cli_name(__tmp),                                         \
1371                __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages,        \
1372                atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1373                __tmp->cl_lost_grant, __tmp->cl_avail_grant,             \
1374                __tmp->cl_dirty_grant,                                   \
1375                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,         \
1376                atomic_long_read(&__tmp->cl_lru_in_list),                \
1377                atomic_long_read(&__tmp->cl_lru_busy),                   \
1378                atomic_read(&__tmp->cl_lru_shrinkers), ##args);          \
1379 } while (0)
1380
1381 /* caller must hold loi_list_lock */
1382 static void osc_consume_write_grant(struct client_obd *cli,
1383                                     struct brw_page *pga)
1384 {
1385         assert_spin_locked(&cli->cl_loi_list_lock);
1386         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1387         atomic_long_inc(&obd_dirty_pages);
1388         cli->cl_dirty_pages++;
1389         pga->flag |= OBD_BRW_FROM_GRANT;
1390         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1391                PAGE_CACHE_SIZE, pga, pga->pg);
1392         osc_update_next_shrink(cli);
1393 }
1394
1395 /* the companion to osc_consume_write_grant, called when a brw has completed.
1396  * must be called with the loi lock held. */
1397 static void osc_release_write_grant(struct client_obd *cli,
1398                                     struct brw_page *pga)
1399 {
1400         ENTRY;
1401
1402         assert_spin_locked(&cli->cl_loi_list_lock);
1403         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1404                 EXIT;
1405                 return;
1406         }
1407
1408         pga->flag &= ~OBD_BRW_FROM_GRANT;
1409         atomic_long_dec(&obd_dirty_pages);
1410         cli->cl_dirty_pages--;
1411         if (pga->flag & OBD_BRW_NOCACHE) {
1412                 pga->flag &= ~OBD_BRW_NOCACHE;
1413                 atomic_long_dec(&obd_dirty_transit_pages);
1414                 cli->cl_dirty_transit--;
1415         }
1416         EXIT;
1417 }
1418
1419 /**
1420  * To avoid sleeping with object lock held, it's good for us allocate enough
1421  * grants before entering into critical section.
1422  *
1423  * client_obd_list_lock held by caller
1424  */
1425 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1426 {
1427         int rc = -EDQUOT;
1428
1429         if (cli->cl_avail_grant >= bytes) {
1430                 cli->cl_avail_grant    -= bytes;
1431                 cli->cl_reserved_grant += bytes;
1432                 rc = 0;
1433         }
1434         return rc;
1435 }
1436
1437 static void __osc_unreserve_grant(struct client_obd *cli,
1438                                   unsigned int reserved, unsigned int unused)
1439 {
1440         /* it's quite normal for us to get more grant than reserved.
1441          * Thinking about a case that two extents merged by adding a new
1442          * chunk, we can save one extent tax. If extent tax is greater than
1443          * one chunk, we can save more grant by adding a new chunk */
1444         cli->cl_reserved_grant -= reserved;
1445         if (unused > reserved) {
1446                 cli->cl_avail_grant += reserved;
1447                 cli->cl_lost_grant  += unused - reserved;
1448                 cli->cl_dirty_grant -= unused - reserved;
1449         } else {
1450                 cli->cl_avail_grant += unused;
1451                 cli->cl_dirty_grant += reserved - unused;
1452         }
1453 }
1454
1455 static void osc_unreserve_grant(struct client_obd *cli,
1456                                 unsigned int reserved, unsigned int unused)
1457 {
1458         spin_lock(&cli->cl_loi_list_lock);
1459         __osc_unreserve_grant(cli, reserved, unused);
1460         if (unused > 0)
1461                 osc_wake_cache_waiters(cli);
1462         spin_unlock(&cli->cl_loi_list_lock);
1463 }
1464
1465 /**
1466  * Free grant after IO is finished or canceled.
1467  *
1468  * @lost_grant is used to remember how many grants we have allocated but not
1469  * used, we should return these grants to OST. There're two cases where grants
1470  * can be lost:
1471  * 1. truncate;
1472  * 2. blocksize at OST is less than PAGE_CACHE_SIZE and a partial page was
1473  *    written. In this case OST may use less chunks to serve this partial
1474  *    write. OSTs don't actually know the page size on the client side. so
1475  *    clients have to calculate lost grant by the blocksize on the OST.
1476  *    See filter_grant_check() for details.
1477  */
1478 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1479                            unsigned int lost_grant, unsigned int dirty_grant)
1480 {
1481         unsigned long grant;
1482
1483         grant = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
1484
1485         spin_lock(&cli->cl_loi_list_lock);
1486         atomic_long_sub(nr_pages, &obd_dirty_pages);
1487         cli->cl_dirty_pages -= nr_pages;
1488         cli->cl_lost_grant += lost_grant;
1489         cli->cl_dirty_grant -= dirty_grant;
1490         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1491                 /* borrow some grant from truncate to avoid the case that
1492                  * truncate uses up all avail grant */
1493                 cli->cl_lost_grant -= grant;
1494                 cli->cl_avail_grant += grant;
1495         }
1496         osc_wake_cache_waiters(cli);
1497         spin_unlock(&cli->cl_loi_list_lock);
1498         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu/%lu\n",
1499                lost_grant, cli->cl_lost_grant,
1500                cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_CACHE_SHIFT,
1501                cli->cl_dirty_grant);
1502 }
1503
1504 /**
1505  * The companion to osc_enter_cache(), called when @oap is no longer part of
1506  * the dirty accounting due to error.
1507  */
1508 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1509 {
1510         spin_lock(&cli->cl_loi_list_lock);
1511         osc_release_write_grant(cli, &oap->oap_brw_page);
1512         spin_unlock(&cli->cl_loi_list_lock);
1513 }
1514
1515 /**
1516  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1517  * is available.
1518  */
1519 static int osc_enter_cache_try(struct client_obd *cli,
1520                                struct osc_async_page *oap,
1521                                int bytes, int transient)
1522 {
1523         int rc;
1524
1525         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1526
1527         rc = osc_reserve_grant(cli, bytes);
1528         if (rc < 0)
1529                 return 0;
1530
1531         if (cli->cl_dirty_pages < cli->cl_dirty_max_pages &&
1532             1 + atomic_long_read(&obd_dirty_pages) <= obd_max_dirty_pages) {
1533                 osc_consume_write_grant(cli, &oap->oap_brw_page);
1534                 if (transient) {
1535                         cli->cl_dirty_transit++;
1536                         atomic_long_inc(&obd_dirty_transit_pages);
1537                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
1538                 }
1539                 rc = 1;
1540         } else {
1541                 __osc_unreserve_grant(cli, bytes, bytes);
1542                 rc = 0;
1543         }
1544         return rc;
1545 }
1546
1547 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1548 {
1549         int rc;
1550         spin_lock(&cli->cl_loi_list_lock);
1551         rc = list_empty(&ocw->ocw_entry);
1552         spin_unlock(&cli->cl_loi_list_lock);
1553         return rc;
1554 }
1555
1556 /**
1557  * The main entry to reserve dirty page accounting. Usually the grant reserved
1558  * in this function will be freed in bulk in osc_free_grant() unless it fails
1559  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1560  *
1561  * The process will be put into sleep if it's already run out of grant.
1562  */
1563 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1564                            struct osc_async_page *oap, int bytes)
1565 {
1566         struct osc_object       *osc = oap->oap_obj;
1567         struct lov_oinfo        *loi = osc->oo_oinfo;
1568         struct osc_cache_waiter  ocw;
1569         struct l_wait_info       lwi;
1570         int                      rc = -EDQUOT;
1571         ENTRY;
1572
1573         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(AT_OFF ? obd_timeout : at_max),
1574                                NULL, LWI_ON_SIGNAL_NOOP, NULL);
1575
1576         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1577
1578         spin_lock(&cli->cl_loi_list_lock);
1579
1580         /* force the caller to try sync io.  this can jump the list
1581          * of queued writes and create a discontiguous rpc stream */
1582         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1583             cli->cl_dirty_max_pages == 0 ||
1584             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) {
1585                 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1586                 GOTO(out, rc = -EDQUOT);
1587         }
1588
1589         /* Hopefully normal case - cache space and write credits available */
1590         if (osc_enter_cache_try(cli, oap, bytes, 0)) {
1591                 OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1592                 GOTO(out, rc = 0);
1593         }
1594
1595         /* We can get here for two reasons: too many dirty pages in cache, or
1596          * run out of grants. In both cases we should write dirty pages out.
1597          * Adding a cache waiter will trigger urgent write-out no matter what
1598          * RPC size will be.
1599          * The exiting condition is no avail grants and no dirty pages caching,
1600          * that really means there is no space on the OST. */
1601         init_waitqueue_head(&ocw.ocw_waitq);
1602         ocw.ocw_oap   = oap;
1603         ocw.ocw_grant = bytes;
1604         while (cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0) {
1605                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1606                 ocw.ocw_rc = 0;
1607                 spin_unlock(&cli->cl_loi_list_lock);
1608
1609                 osc_io_unplug_async(env, cli, NULL);
1610
1611                 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1612                        cli_name(cli), &ocw, oap);
1613
1614                 rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1615
1616                 spin_lock(&cli->cl_loi_list_lock);
1617
1618                 if (rc < 0) {
1619                         /* l_wait_event is interrupted by signal or timed out */
1620                         list_del_init(&ocw.ocw_entry);
1621                         break;
1622                 }
1623                 LASSERT(list_empty(&ocw.ocw_entry));
1624                 rc = ocw.ocw_rc;
1625
1626                 if (rc != -EDQUOT)
1627                         break;
1628                 if (osc_enter_cache_try(cli, oap, bytes, 0)) {
1629                         rc = 0;
1630                         break;
1631                 }
1632         }
1633
1634         switch (rc) {
1635         case 0:
1636                 OSC_DUMP_GRANT(D_CACHE, cli, "finally got grant space\n");
1637                 break;
1638         case -ETIMEDOUT:
1639                 OSC_DUMP_GRANT(D_CACHE, cli,
1640                                "timeout, fall back to sync i/o\n");
1641                 osc_extent_tree_dump(D_CACHE, osc);
1642                 /* fall back to synchronous I/O */
1643                 rc = -EDQUOT;
1644                 break;
1645         case -EINTR:
1646                 /* Ensures restartability - LU-3581 */
1647                 OSC_DUMP_GRANT(D_CACHE, cli, "interrupted\n");
1648                 rc = -ERESTARTSYS;
1649                 break;
1650         case -EDQUOT:
1651                 OSC_DUMP_GRANT(D_CACHE, cli,
1652                                "no grant space, fall back to sync i/o\n");
1653                 break;
1654         default:
1655                 CDEBUG(D_CACHE, "%s: event for cache space @ %p never arrived "
1656                        "due to %d, fall back to sync i/o\n",
1657                        cli_name(cli), &ocw, rc);
1658                 break;
1659         }
1660         EXIT;
1661 out:
1662         spin_unlock(&cli->cl_loi_list_lock);
1663         RETURN(rc);
1664 }
1665
1666 /* caller must hold loi_list_lock */
1667 void osc_wake_cache_waiters(struct client_obd *cli)
1668 {
1669         struct list_head *l, *tmp;
1670         struct osc_cache_waiter *ocw;
1671
1672         ENTRY;
1673         list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1674                 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
1675                 list_del_init(&ocw->ocw_entry);
1676
1677                 ocw->ocw_rc = -EDQUOT;
1678                 /* we can't dirty more */
1679                 if ((cli->cl_dirty_pages  >= cli->cl_dirty_max_pages) ||
1680                     (1 + atomic_long_read(&obd_dirty_pages) >
1681                      obd_max_dirty_pages)) {
1682                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
1683                                "osc max %ld, sys max %ld\n",
1684                                cli->cl_dirty_pages, cli->cl_dirty_max_pages,
1685                                obd_max_dirty_pages);
1686                         goto wakeup;
1687                 }
1688
1689                 if (osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
1690                         ocw->ocw_rc = 0;
1691 wakeup:
1692                 CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld, %d\n",
1693                        ocw, ocw->ocw_oap, cli->cl_avail_grant, ocw->ocw_rc);
1694
1695                 wake_up(&ocw->ocw_waitq);
1696         }
1697
1698         EXIT;
1699 }
1700
1701 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1702 {
1703         int hprpc = !!list_empty(&osc->oo_hp_exts);
1704         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1705 }
1706
1707 /* This maintains the lists of pending pages to read/write for a given object
1708  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1709  * to quickly find objects that are ready to send an RPC. */
1710 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1711                          int cmd)
1712 {
1713         int invalid_import = 0;
1714         ENTRY;
1715
1716         /* if we have an invalid import we want to drain the queued pages
1717          * by forcing them through rpcs that immediately fail and complete
1718          * the pages.  recovery relies on this to empty the queued pages
1719          * before canceling the locks and evicting down the llite pages */
1720         if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1721                 invalid_import = 1;
1722
1723         if (cmd & OBD_BRW_WRITE) {
1724                 if (atomic_read(&osc->oo_nr_writes) == 0)
1725                         RETURN(0);
1726                 if (invalid_import) {
1727                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1728                         RETURN(1);
1729                 }
1730                 if (!list_empty(&osc->oo_hp_exts)) {
1731                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1732                         RETURN(1);
1733                 }
1734                 if (!list_empty(&osc->oo_urgent_exts)) {
1735                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1736                         RETURN(1);
1737                 }
1738                 /* trigger a write rpc stream as long as there are dirtiers
1739                  * waiting for space.  as they're waiting, they're not going to
1740                  * create more pages to coalesce with what's waiting.. */
1741                 if (!list_empty(&cli->cl_cache_waiters)) {
1742                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1743                         RETURN(1);
1744                 }
1745                 if (atomic_read(&osc->oo_nr_writes) >=
1746                     cli->cl_max_pages_per_rpc)
1747                         RETURN(1);
1748         } else {
1749                 if (atomic_read(&osc->oo_nr_reads) == 0)
1750                         RETURN(0);
1751                 if (invalid_import) {
1752                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1753                         RETURN(1);
1754                 }
1755                 /* all read are urgent. */
1756                 if (!list_empty(&osc->oo_reading_exts))
1757                         RETURN(1);
1758         }
1759
1760         RETURN(0);
1761 }
1762
1763 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1764 {
1765         struct client_obd *cli = osc_cli(obj);
1766         if (cmd & OBD_BRW_WRITE) {
1767                 atomic_add(delta, &obj->oo_nr_writes);
1768                 atomic_add(delta, &cli->cl_pending_w_pages);
1769                 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1770         } else {
1771                 atomic_add(delta, &obj->oo_nr_reads);
1772                 atomic_add(delta, &cli->cl_pending_r_pages);
1773                 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1774         }
1775         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1776 }
1777
1778 static int osc_makes_hprpc(struct osc_object *obj)
1779 {
1780         return !list_empty(&obj->oo_hp_exts);
1781 }
1782
1783 static void on_list(struct list_head *item, struct list_head *list,
1784                     int should_be_on)
1785 {
1786         if (list_empty(item) && should_be_on)
1787                 list_add_tail(item, list);
1788         else if (!list_empty(item) && !should_be_on)
1789                 list_del_init(item);
1790 }
1791
1792 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1793  * can find pages to build into rpcs quickly */
1794 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1795 {
1796         if (osc_makes_hprpc(osc)) {
1797                 /* HP rpc */
1798                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1799                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1800         } else {
1801                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1802                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1803                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1804                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1805         }
1806
1807         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1808                 atomic_read(&osc->oo_nr_writes) > 0);
1809
1810         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1811                 atomic_read(&osc->oo_nr_reads) > 0);
1812
1813         return osc_is_ready(osc);
1814 }
1815
1816 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1817 {
1818         int is_ready;
1819
1820         spin_lock(&cli->cl_loi_list_lock);
1821         is_ready = __osc_list_maint(cli, osc);
1822         spin_unlock(&cli->cl_loi_list_lock);
1823
1824         return is_ready;
1825 }
1826
1827 /* this is trying to propogate async writeback errors back up to the
1828  * application.  As an async write fails we record the error code for later if
1829  * the app does an fsync.  As long as errors persist we force future rpcs to be
1830  * sync so that the app can get a sync error and break the cycle of queueing
1831  * pages for which writeback will fail. */
1832 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1833                            int rc)
1834 {
1835         if (rc) {
1836                 if (!ar->ar_rc)
1837                         ar->ar_rc = rc;
1838
1839                 ar->ar_force_sync = 1;
1840                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1841                 return;
1842
1843         }
1844
1845         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1846                 ar->ar_force_sync = 0;
1847 }
1848
1849 /* this must be called holding the loi list lock to give coverage to exit_cache,
1850  * async_flag maintenance, and oap_request */
1851 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1852                               struct osc_async_page *oap, int sent, int rc)
1853 {
1854         struct osc_object *osc = oap->oap_obj;
1855         struct lov_oinfo  *loi = osc->oo_oinfo;
1856         __u64 xid = 0;
1857
1858         ENTRY;
1859         if (oap->oap_request != NULL) {
1860                 xid = ptlrpc_req_xid(oap->oap_request);
1861                 ptlrpc_req_finished(oap->oap_request);
1862                 oap->oap_request = NULL;
1863         }
1864
1865         /* As the transfer for this page is being done, clear the flags */
1866         spin_lock(&oap->oap_lock);
1867         oap->oap_async_flags = 0;
1868         spin_unlock(&oap->oap_lock);
1869         oap->oap_interrupted = 0;
1870
1871         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1872                 spin_lock(&cli->cl_loi_list_lock);
1873                 osc_process_ar(&cli->cl_ar, xid, rc);
1874                 osc_process_ar(&loi->loi_ar, xid, rc);
1875                 spin_unlock(&cli->cl_loi_list_lock);
1876         }
1877
1878         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1879         if (rc)
1880                 CERROR("completion on oap %p obj %p returns %d.\n",
1881                        oap, osc, rc);
1882
1883         EXIT;
1884 }
1885
1886 /**
1887  * Try to add extent to one RPC. We need to think about the following things:
1888  * - # of pages must not be over max_pages_per_rpc
1889  * - extent must be compatible with previous ones
1890  */
1891 static int try_to_add_extent_for_io(struct client_obd *cli,
1892                                     struct osc_extent *ext,
1893                                     struct list_head *rpclist,
1894                                     unsigned int *pc, unsigned int *max_pages)
1895 {
1896         struct osc_extent *tmp;
1897         struct osc_async_page *oap = list_first_entry(&ext->oe_pages,
1898                                                       struct osc_async_page,
1899                                                       oap_pending_item);
1900         ENTRY;
1901
1902         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1903                 ext);
1904
1905         *max_pages = max(ext->oe_mppr, *max_pages);
1906         if (*pc + ext->oe_nr_pages > *max_pages)
1907                 RETURN(0);
1908
1909         list_for_each_entry(tmp, rpclist, oe_link) {
1910                 struct osc_async_page *oap2;
1911                 oap2 = list_first_entry(&tmp->oe_pages, struct osc_async_page,
1912                                         oap_pending_item);
1913                 EASSERT(tmp->oe_owner == current, tmp);
1914 #if 0
1915                 if (overlapped(tmp, ext)) {
1916                         OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
1917                         EASSERT(0, ext);
1918                 }
1919 #endif
1920                 if (oap2cl_page(oap)->cp_type != oap2cl_page(oap2)->cp_type) {
1921                         CDEBUG(D_CACHE, "Do not permit different type of IO"
1922                                         " for a same RPC\n");
1923                         RETURN(0);
1924                 }
1925
1926                 if (tmp->oe_srvlock != ext->oe_srvlock ||
1927                     !tmp->oe_grants != !ext->oe_grants ||
1928                     tmp->oe_no_merge || ext->oe_no_merge)
1929                         RETURN(0);
1930
1931                 /* remove break for strict check */
1932                 break;
1933         }
1934
1935         *pc += ext->oe_nr_pages;
1936         list_move_tail(&ext->oe_link, rpclist);
1937         ext->oe_owner = current;
1938         RETURN(1);
1939 }
1940
1941 /**
1942  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1943  * get_write_extent() takes all appropriate extents in atomic.
1944  *
1945  * The following policy is used to collect extents for IO:
1946  * 1. Add as many HP extents as possible;
1947  * 2. Add the first urgent extent in urgent extent list and take it out of
1948  *    urgent list;
1949  * 3. Add subsequent extents of this urgent extent;
1950  * 4. If urgent list is not empty, goto 2;
1951  * 5. Traverse the extent tree from the 1st extent;
1952  * 6. Above steps exit if there is no space in this RPC.
1953  */
1954 static unsigned int get_write_extents(struct osc_object *obj,
1955                                       struct list_head *rpclist)
1956 {
1957         struct client_obd *cli = osc_cli(obj);
1958         struct osc_extent *ext;
1959         unsigned int page_count = 0;
1960         unsigned int max_pages = cli->cl_max_pages_per_rpc;
1961
1962         LASSERT(osc_object_is_locked(obj));
1963         while (!list_empty(&obj->oo_hp_exts)) {
1964                 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
1965                                  oe_link);
1966                 LASSERT(ext->oe_state == OES_CACHE);
1967                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1968                                               &max_pages))
1969                         return page_count;
1970                 EASSERT(ext->oe_nr_pages <= max_pages, ext);
1971         }
1972         if (page_count == max_pages)
1973                 return page_count;
1974
1975         while (!list_empty(&obj->oo_urgent_exts)) {
1976                 ext = list_entry(obj->oo_urgent_exts.next,
1977                                  struct osc_extent, oe_link);
1978                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1979                                               &max_pages))
1980                         return page_count;
1981
1982                 if (!ext->oe_intree)
1983                         continue;
1984
1985                 while ((ext = next_extent(ext)) != NULL) {
1986                         if ((ext->oe_state != OES_CACHE) ||
1987                             (!list_empty(&ext->oe_link) &&
1988                              ext->oe_owner != NULL))
1989                                 continue;
1990
1991                         if (!try_to_add_extent_for_io(cli, ext, rpclist,
1992                                                       &page_count, &max_pages))
1993                                 return page_count;
1994                 }
1995         }
1996         if (page_count == max_pages)
1997                 return page_count;
1998
1999         ext = first_extent(obj);
2000         while (ext != NULL) {
2001                 if ((ext->oe_state != OES_CACHE) ||
2002                     /* this extent may be already in current rpclist */
2003                     (!list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
2004                         ext = next_extent(ext);
2005                         continue;
2006                 }
2007
2008                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
2009                                               &max_pages))
2010                         return page_count;
2011
2012                 ext = next_extent(ext);
2013         }
2014         return page_count;
2015 }
2016
2017 static int
2018 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
2019                    struct osc_object *osc)
2020 __must_hold(osc)
2021 {
2022         struct list_head   rpclist = LIST_HEAD_INIT(rpclist);
2023         struct osc_extent *ext;
2024         struct osc_extent *tmp;
2025         struct osc_extent *first = NULL;
2026         unsigned int page_count = 0;
2027         int srvlock = 0;
2028         int rc = 0;
2029         ENTRY;
2030
2031         LASSERT(osc_object_is_locked(osc));
2032
2033         page_count = get_write_extents(osc, &rpclist);
2034         LASSERT(equi(page_count == 0, list_empty(&rpclist)));
2035
2036         if (list_empty(&rpclist))
2037                 RETURN(0);
2038
2039         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2040
2041         list_for_each_entry(ext, &rpclist, oe_link) {
2042                 LASSERT(ext->oe_state == OES_CACHE ||
2043                         ext->oe_state == OES_LOCK_DONE);
2044                 if (ext->oe_state == OES_CACHE)
2045                         osc_extent_state_set(ext, OES_LOCKING);
2046                 else
2047                         osc_extent_state_set(ext, OES_RPC);
2048         }
2049
2050         /* we're going to grab page lock, so release object lock because
2051          * lock order is page lock -> object lock. */
2052         osc_object_unlock(osc);
2053
2054         list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2055                 if (ext->oe_state == OES_LOCKING) {
2056                         rc = osc_extent_make_ready(env, ext);
2057                         if (unlikely(rc < 0)) {
2058                                 list_del_init(&ext->oe_link);
2059                                 osc_extent_finish(env, ext, 0, rc);
2060                                 continue;
2061                         }
2062                 }
2063                 if (first == NULL) {
2064                         first = ext;
2065                         srvlock = ext->oe_srvlock;
2066                 } else {
2067                         LASSERT(srvlock == ext->oe_srvlock);
2068                 }
2069         }
2070
2071         if (!list_empty(&rpclist)) {
2072                 LASSERT(page_count > 0);
2073                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2074                 LASSERT(list_empty(&rpclist));
2075         }
2076
2077         osc_object_lock(osc);
2078         RETURN(rc);
2079 }
2080
2081 /**
2082  * prepare pages for ASYNC io and put pages in send queue.
2083  *
2084  * \param cmd OBD_BRW_* macroses
2085  * \param lop pending pages
2086  *
2087  * \return zero if no page added to send queue.
2088  * \return 1 if pages successfully added to send queue.
2089  * \return negative on errors.
2090  */
2091 static int
2092 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2093                   struct osc_object *osc)
2094 __must_hold(osc)
2095 {
2096         struct osc_extent *ext;
2097         struct osc_extent *next;
2098         struct list_head rpclist = LIST_HEAD_INIT(rpclist);
2099         unsigned int page_count = 0;
2100         unsigned int max_pages = cli->cl_max_pages_per_rpc;
2101         int rc = 0;
2102         ENTRY;
2103
2104         LASSERT(osc_object_is_locked(osc));
2105         list_for_each_entry_safe(ext, next,
2106                                      &osc->oo_reading_exts, oe_link) {
2107                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2108                 if (!try_to_add_extent_for_io(cli, ext, &rpclist, &page_count,
2109                                               &max_pages))
2110                         break;
2111                 osc_extent_state_set(ext, OES_RPC);
2112                 EASSERT(ext->oe_nr_pages <= max_pages, ext);
2113         }
2114         LASSERT(page_count <= max_pages);
2115
2116         osc_update_pending(osc, OBD_BRW_READ, -page_count);
2117
2118         if (!list_empty(&rpclist)) {
2119                 osc_object_unlock(osc);
2120
2121                 LASSERT(page_count > 0);
2122                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2123                 LASSERT(list_empty(&rpclist));
2124
2125                 osc_object_lock(osc);
2126         }
2127         RETURN(rc);
2128 }
2129
2130 #define list_to_obj(list, item) ({                                            \
2131         struct list_head *__tmp = (list)->next;                               \
2132         list_del_init(__tmp);                                         \
2133         list_entry(__tmp, struct osc_object, oo_##item);                      \
2134 })
2135
2136 /* This is called by osc_check_rpcs() to find which objects have pages that
2137  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2138 static struct osc_object *osc_next_obj(struct client_obd *cli)
2139 {
2140         ENTRY;
2141
2142         /* First return objects that have blocked locks so that they
2143          * will be flushed quickly and other clients can get the lock,
2144          * then objects which have pages ready to be stuffed into RPCs */
2145         if (!list_empty(&cli->cl_loi_hp_ready_list))
2146                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2147         if (!list_empty(&cli->cl_loi_ready_list))
2148                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2149
2150         /* then if we have cache waiters, return all objects with queued
2151          * writes.  This is especially important when many small files
2152          * have filled up the cache and not been fired into rpcs because
2153          * they don't pass the nr_pending/object threshhold */
2154         if (!list_empty(&cli->cl_cache_waiters) &&
2155             !list_empty(&cli->cl_loi_write_list))
2156                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2157
2158         /* then return all queued objects when we have an invalid import
2159          * so that they get flushed */
2160         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2161                 if (!list_empty(&cli->cl_loi_write_list))
2162                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2163                                            write_item));
2164                 if (!list_empty(&cli->cl_loi_read_list))
2165                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2166                                            read_item));
2167         }
2168         RETURN(NULL);
2169 }
2170
2171 /* called with the loi list lock held */
2172 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2173 __must_hold(&cli->cl_loi_list_lock)
2174 {
2175         struct osc_object *osc;
2176         int rc = 0;
2177         ENTRY;
2178
2179         while ((osc = osc_next_obj(cli)) != NULL) {
2180                 struct cl_object *obj = osc2cl(osc);
2181                 struct lu_ref_link link;
2182
2183                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2184
2185                 if (osc_max_rpc_in_flight(cli, osc)) {
2186                         __osc_list_maint(cli, osc);
2187                         break;
2188                 }
2189
2190                 cl_object_get(obj);
2191                 spin_unlock(&cli->cl_loi_list_lock);
2192                 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2193
2194                 /* attempt some read/write balancing by alternating between
2195                  * reads and writes in an object.  The makes_rpc checks here
2196                  * would be redundant if we were getting read/write work items
2197                  * instead of objects.  we don't want send_oap_rpc to drain a
2198                  * partial read pending queue when we're given this object to
2199                  * do io on writes while there are cache waiters */
2200                 osc_object_lock(osc);
2201                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2202                         rc = osc_send_write_rpc(env, cli, osc);
2203                         if (rc < 0) {
2204                                 CERROR("Write request failed with %d\n", rc);
2205
2206                                 /* osc_send_write_rpc failed, mostly because of
2207                                  * memory pressure.
2208                                  *
2209                                  * It can't break here, because if:
2210                                  *  - a page was submitted by osc_io_submit, so
2211                                  *    page locked;
2212                                  *  - no request in flight
2213                                  *  - no subsequent request
2214                                  * The system will be in live-lock state,
2215                                  * because there is no chance to call
2216                                  * osc_io_unplug() and osc_check_rpcs() any
2217                                  * more. pdflush can't help in this case,
2218                                  * because it might be blocked at grabbing
2219                                  * the page lock as we mentioned.
2220                                  *
2221                                  * Anyway, continue to drain pages. */
2222                                 /* break; */
2223                         }
2224                 }
2225                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2226                         rc = osc_send_read_rpc(env, cli, osc);
2227                         if (rc < 0)
2228                                 CERROR("Read request failed with %d\n", rc);
2229                 }
2230                 osc_object_unlock(osc);
2231
2232                 osc_list_maint(cli, osc);
2233                 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2234                 cl_object_put(env, obj);
2235
2236                 spin_lock(&cli->cl_loi_list_lock);
2237         }
2238 }
2239
2240 static int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2241                           struct osc_object *osc, int async)
2242 {
2243         int rc = 0;
2244
2245         if (osc != NULL && osc_list_maint(cli, osc) == 0)
2246                 return 0;
2247
2248         if (!async) {
2249                 spin_lock(&cli->cl_loi_list_lock);
2250                 osc_check_rpcs(env, cli);
2251                 spin_unlock(&cli->cl_loi_list_lock);
2252         } else {
2253                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2254                 LASSERT(cli->cl_writeback_work != NULL);
2255                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2256         }
2257         return rc;
2258 }
2259
2260 static int osc_io_unplug_async(const struct lu_env *env,
2261                                 struct client_obd *cli, struct osc_object *osc)
2262 {
2263         return osc_io_unplug0(env, cli, osc, 1);
2264 }
2265
2266 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
2267                    struct osc_object *osc)
2268 {
2269         (void)osc_io_unplug0(env, cli, osc, 0);
2270 }
2271
2272 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2273                         struct page *page, loff_t offset)
2274 {
2275         struct obd_export     *exp = osc_export(osc);
2276         struct osc_async_page *oap = &ops->ops_oap;
2277         ENTRY;
2278
2279         if (!page)
2280                 return cfs_size_round(sizeof(*oap));
2281
2282         oap->oap_magic = OAP_MAGIC;
2283         oap->oap_cli = &exp->exp_obd->u.cli;
2284         oap->oap_obj = osc;
2285
2286         oap->oap_page = page;
2287         oap->oap_obj_off = offset;
2288         LASSERT(!(offset & ~PAGE_MASK));
2289
2290         if (cfs_capable(CFS_CAP_SYS_RESOURCE))
2291                 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2292
2293         INIT_LIST_HEAD(&oap->oap_pending_item);
2294         INIT_LIST_HEAD(&oap->oap_rpc_item);
2295
2296         spin_lock_init(&oap->oap_lock);
2297         CDEBUG(D_INFO, "oap %p page %p obj off "LPU64"\n",
2298                oap, page, oap->oap_obj_off);
2299         RETURN(0);
2300 }
2301
2302 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2303                        struct osc_page *ops)
2304 {
2305         struct osc_io *oio = osc_env_io(env);
2306         struct osc_extent     *ext = NULL;
2307         struct osc_async_page *oap = &ops->ops_oap;
2308         struct client_obd     *cli = oap->oap_cli;
2309         struct osc_object     *osc = oap->oap_obj;
2310         pgoff_t index;
2311         unsigned int tmp;
2312         unsigned int grants = 0;
2313         u32    brw_flags = OBD_BRW_ASYNC;
2314         int    cmd = OBD_BRW_WRITE;
2315         int    need_release = 0;
2316         int    rc = 0;
2317         ENTRY;
2318
2319         if (oap->oap_magic != OAP_MAGIC)
2320                 RETURN(-EINVAL);
2321
2322         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2323                 RETURN(-EIO);
2324
2325         if (!list_empty(&oap->oap_pending_item) ||
2326             !list_empty(&oap->oap_rpc_item))
2327                 RETURN(-EBUSY);
2328
2329         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2330         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2331         if (cfs_capable(CFS_CAP_SYS_RESOURCE)) {
2332                 brw_flags |= OBD_BRW_NOQUOTA;
2333                 cmd |= OBD_BRW_NOQUOTA;
2334         }
2335
2336         /* check if the file's owner/group is over quota */
2337         if (!(cmd & OBD_BRW_NOQUOTA)) {
2338                 struct cl_object *obj;
2339                 struct cl_attr   *attr;
2340                 unsigned int qid[LL_MAXQUOTAS];
2341
2342                 obj = cl_object_top(&osc->oo_cl);
2343                 attr = &osc_env_info(env)->oti_attr;
2344
2345                 cl_object_attr_lock(obj);
2346                 rc = cl_object_attr_get(env, obj, attr);
2347                 cl_object_attr_unlock(obj);
2348
2349                 qid[USRQUOTA] = attr->cat_uid;
2350                 qid[GRPQUOTA] = attr->cat_gid;
2351                 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2352                         rc = -EDQUOT;
2353                 if (rc)
2354                         RETURN(rc);
2355         }
2356
2357         oap->oap_cmd = cmd;
2358         oap->oap_page_off = ops->ops_from;
2359         oap->oap_count = ops->ops_to - ops->ops_from;
2360         /* No need to hold a lock here,
2361          * since this page is not in any list yet. */
2362         oap->oap_async_flags = 0;
2363         oap->oap_brw_flags = brw_flags;
2364
2365         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2366                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2367
2368         index = osc_index(oap2osc(oap));
2369
2370         /* Add this page into extent by the following steps:
2371          * 1. if there exists an active extent for this IO, mostly this page
2372          *    can be added to the active extent and sometimes we need to
2373          *    expand extent to accomodate this page;
2374          * 2. otherwise, a new extent will be allocated. */
2375
2376         ext = oio->oi_active;
2377         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2378                 /* one chunk plus extent overhead must be enough to write this
2379                  * page */
2380                 grants = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2381                 if (ext->oe_end >= index)
2382                         grants = 0;
2383
2384                 /* it doesn't need any grant to dirty this page */
2385                 spin_lock(&cli->cl_loi_list_lock);
2386                 rc = osc_enter_cache_try(cli, oap, grants, 0);
2387                 spin_unlock(&cli->cl_loi_list_lock);
2388                 if (rc == 0) { /* try failed */
2389                         grants = 0;
2390                         need_release = 1;
2391                 } else if (ext->oe_end < index) {
2392                         tmp = grants;
2393                         /* try to expand this extent */
2394                         rc = osc_extent_expand(ext, index, &tmp);
2395                         if (rc < 0) {
2396                                 need_release = 1;
2397                                 /* don't free reserved grant */
2398                         } else {
2399                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2400                                                 "expanded for %lu.\n", index);
2401                                 osc_unreserve_grant(cli, grants, tmp);
2402                                 grants = 0;
2403                         }
2404                 }
2405                 rc = 0;
2406         } else if (ext != NULL) {
2407                 /* index is located outside of active extent */
2408                 need_release = 1;
2409         }
2410         if (need_release) {
2411                 osc_extent_release(env, ext);
2412                 oio->oi_active = NULL;
2413                 ext = NULL;
2414         }
2415
2416         if (ext == NULL) {
2417                 tmp = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2418
2419                 /* try to find new extent to cover this page */
2420                 LASSERT(oio->oi_active == NULL);
2421                 /* we may have allocated grant for this page if we failed
2422                  * to expand the previous active extent. */
2423                 LASSERT(ergo(grants > 0, grants >= tmp));
2424
2425                 rc = 0;
2426                 if (grants == 0) {
2427                         /* we haven't allocated grant for this page. */
2428                         rc = osc_enter_cache(env, cli, oap, tmp);
2429                         if (rc == 0)
2430                                 grants = tmp;
2431                 }
2432
2433                 tmp = grants;
2434                 if (rc == 0) {
2435                         ext = osc_extent_find(env, osc, index, &tmp);
2436                         if (IS_ERR(ext)) {
2437                                 LASSERT(tmp == grants);
2438                                 osc_exit_cache(cli, oap);
2439                                 rc = PTR_ERR(ext);
2440                                 ext = NULL;
2441                         } else {
2442                                 oio->oi_active = ext;
2443                         }
2444                 }
2445                 if (grants > 0)
2446                         osc_unreserve_grant(cli, grants, tmp);
2447         }
2448
2449         LASSERT(ergo(rc == 0, ext != NULL));
2450         if (ext != NULL) {
2451                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2452                          ext, "index = %lu.\n", index);
2453                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2454
2455                 osc_object_lock(osc);
2456                 if (ext->oe_nr_pages == 0)
2457                         ext->oe_srvlock = ops->ops_srvlock;
2458                 else
2459                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2460                 ++ext->oe_nr_pages;
2461                 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2462                 osc_object_unlock(osc);
2463         }
2464         RETURN(rc);
2465 }
2466
2467 int osc_teardown_async_page(const struct lu_env *env,
2468                             struct osc_object *obj, struct osc_page *ops)
2469 {
2470         struct osc_async_page *oap = &ops->ops_oap;
2471         int rc = 0;
2472         ENTRY;
2473
2474         LASSERT(oap->oap_magic == OAP_MAGIC);
2475
2476         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2477                oap, ops, osc_index(oap2osc(oap)));
2478
2479         if (!list_empty(&oap->oap_rpc_item)) {
2480                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2481                 rc = -EBUSY;
2482         } else if (!list_empty(&oap->oap_pending_item)) {
2483                 struct osc_extent *ext = NULL;
2484
2485                 osc_object_lock(obj);
2486                 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2487                 osc_object_unlock(obj);
2488                 /* only truncated pages are allowed to be taken out.
2489                  * See osc_extent_truncate() and osc_cache_truncate_start()
2490                  * for details. */
2491                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2492                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2493                                         osc_index(oap2osc(oap)));
2494                         rc = -EBUSY;
2495                 }
2496                 if (ext != NULL)
2497                         osc_extent_put(env, ext);
2498         }
2499         RETURN(rc);
2500 }
2501
2502 /**
2503  * This is called when a page is picked up by kernel to write out.
2504  *
2505  * We should find out the corresponding extent and add the whole extent
2506  * into urgent list. The extent may be being truncated or used, handle it
2507  * carefully.
2508  */
2509 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2510                          struct osc_page *ops)
2511 {
2512         struct osc_extent *ext   = NULL;
2513         struct osc_object *obj   = cl2osc(ops->ops_cl.cpl_obj);
2514         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2515         pgoff_t            index = osc_index(ops);
2516         struct osc_async_page *oap = &ops->ops_oap;
2517         bool unplug = false;
2518         int rc = 0;
2519         ENTRY;
2520
2521         osc_object_lock(obj);
2522         ext = osc_extent_lookup(obj, index);
2523         if (ext == NULL) {
2524                 osc_extent_tree_dump(D_ERROR, obj);
2525                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2526         }
2527
2528         switch (ext->oe_state) {
2529         case OES_RPC:
2530         case OES_LOCK_DONE:
2531                 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2532                 LASSERT(0);
2533                 break;
2534         case OES_LOCKING:
2535                 /* If we know this extent is being written out, we should abort
2536                  * so that the writer can make this page ready. Otherwise, there
2537                  * exists a deadlock problem because other process can wait for
2538                  * page writeback bit holding page lock; and meanwhile in
2539                  * vvp_page_make_ready(), we need to grab page lock before
2540                  * really sending the RPC. */
2541         case OES_TRUNC:
2542                 /* race with truncate, page will be redirtied */
2543         case OES_ACTIVE:
2544                 /* The extent is active so we need to abort and let the caller
2545                  * re-dirty the page. If we continued on here, and we were the
2546                  * one making the extent active, we could deadlock waiting for
2547                  * the page writeback to clear but it won't because the extent
2548                  * is active and won't be written out. */
2549                 GOTO(out, rc = -EAGAIN);
2550         default:
2551                 break;
2552         }
2553
2554         rc = cl_page_prep(env, io, cp, CRT_WRITE);
2555         if (rc)
2556                 GOTO(out, rc);
2557
2558         spin_lock(&oap->oap_lock);
2559         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2560         spin_unlock(&oap->oap_lock);
2561
2562         if (memory_pressure_get())
2563                 ext->oe_memalloc = 1;
2564
2565         ext->oe_urgent = 1;
2566         if (ext->oe_state == OES_CACHE) {
2567                 OSC_EXTENT_DUMP(D_CACHE, ext,
2568                                 "flush page %p make it urgent.\n", oap);
2569                 if (list_empty(&ext->oe_link))
2570                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2571                 unplug = true;
2572         }
2573         rc = 0;
2574         EXIT;
2575
2576 out:
2577         osc_object_unlock(obj);
2578         osc_extent_put(env, ext);
2579         if (unplug)
2580                 osc_io_unplug_async(env, osc_cli(obj), obj);
2581         return rc;
2582 }
2583
2584 /**
2585  * this is called when a sync waiter receives an interruption.  Its job is to
2586  * get the caller woken as soon as possible.  If its page hasn't been put in an
2587  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2588  * desiring interruption which will forcefully complete the rpc once the rpc
2589  * has timed out.
2590  */
2591 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2592 {
2593         struct osc_async_page *oap = &ops->ops_oap;
2594         struct osc_object     *obj = oap->oap_obj;
2595         struct client_obd     *cli = osc_cli(obj);
2596         struct osc_extent     *ext;
2597         struct osc_extent     *found = NULL;
2598         struct list_head            *plist;
2599         pgoff_t index = osc_index(ops);
2600         int     rc = -EBUSY;
2601         int     cmd;
2602         ENTRY;
2603
2604         LASSERT(!oap->oap_interrupted);
2605         oap->oap_interrupted = 1;
2606
2607         /* Find out the caching extent */
2608         osc_object_lock(obj);
2609         if (oap->oap_cmd & OBD_BRW_WRITE) {
2610                 plist = &obj->oo_urgent_exts;
2611                 cmd   = OBD_BRW_WRITE;
2612         } else {
2613                 plist = &obj->oo_reading_exts;
2614                 cmd   = OBD_BRW_READ;
2615         }
2616         list_for_each_entry(ext, plist, oe_link) {
2617                 if (ext->oe_start <= index && ext->oe_end >= index) {
2618                         LASSERT(ext->oe_state == OES_LOCK_DONE);
2619                         /* For OES_LOCK_DONE state extent, it has already held
2620                          * a refcount for RPC. */
2621                         found = osc_extent_get(ext);
2622                         break;
2623                 }
2624         }
2625         if (found != NULL) {
2626                 list_del_init(&found->oe_link);
2627                 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2628                 osc_object_unlock(obj);
2629
2630                 osc_extent_finish(env, found, 0, -EINTR);
2631                 osc_extent_put(env, found);
2632                 rc = 0;
2633         } else {
2634                 osc_object_unlock(obj);
2635                 /* ok, it's been put in an rpc. only one oap gets a request
2636                  * reference */
2637                 if (oap->oap_request != NULL) {
2638                         ptlrpc_mark_interrupted(oap->oap_request);
2639                         ptlrpcd_wake(oap->oap_request);
2640                         ptlrpc_req_finished(oap->oap_request);
2641                         oap->oap_request = NULL;
2642                 }
2643         }
2644
2645         osc_list_maint(cli, obj);
2646         RETURN(rc);
2647 }
2648
2649 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
2650                          struct list_head *list, int cmd, int brw_flags)
2651 {
2652         struct client_obd     *cli = osc_cli(obj);
2653         struct osc_extent     *ext;
2654         struct osc_async_page *oap;
2655         int     page_count = 0;
2656         int     mppr       = cli->cl_max_pages_per_rpc;
2657         bool    can_merge   = true;
2658         pgoff_t start      = CL_PAGE_EOF;
2659         pgoff_t end        = 0;
2660         ENTRY;
2661
2662         list_for_each_entry(oap, list, oap_pending_item) {
2663                 struct osc_page *opg = oap2osc_page(oap);
2664                 pgoff_t index = osc_index(opg);
2665
2666                 if (index > end)
2667                         end = index;
2668                 if (index < start)
2669                         start = index;
2670                 ++page_count;
2671                 mppr <<= (page_count > mppr);
2672
2673                 if (unlikely(opg->ops_from > 0 || opg->ops_to < PAGE_SIZE))
2674                         can_merge = false;
2675         }
2676
2677         ext = osc_extent_alloc(obj);
2678         if (ext == NULL) {
2679                 struct osc_async_page *tmp;
2680
2681                 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2682                         list_del_init(&oap->oap_pending_item);
2683                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2684                 }
2685                 RETURN(-ENOMEM);
2686         }
2687
2688         ext->oe_rw = !!(cmd & OBD_BRW_READ);
2689         ext->oe_sync = 1;
2690         ext->oe_no_merge = !can_merge;
2691         ext->oe_urgent = 1;
2692         ext->oe_start = start;
2693         ext->oe_end = ext->oe_max_end = end;
2694         ext->oe_obj = obj;
2695         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2696         ext->oe_nr_pages = page_count;
2697         ext->oe_mppr = mppr;
2698         list_splice_init(list, &ext->oe_pages);
2699
2700         osc_object_lock(obj);
2701         /* Reuse the initial refcount for RPC, don't drop it */
2702         osc_extent_state_set(ext, OES_LOCK_DONE);
2703         if (cmd & OBD_BRW_WRITE) {
2704                 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2705                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2706         } else {
2707                 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2708                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2709         }
2710         osc_object_unlock(obj);
2711
2712         osc_io_unplug_async(env, cli, obj);
2713         RETURN(0);
2714 }
2715
2716 /**
2717  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2718  */
2719 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
2720                              __u64 size, struct osc_extent **extp)
2721 {
2722         struct client_obd *cli = osc_cli(obj);
2723         struct osc_extent *ext;
2724         struct osc_extent *waiting = NULL;
2725         pgoff_t index;
2726         struct list_head list = LIST_HEAD_INIT(list);
2727         int result = 0;
2728         bool partial;
2729         ENTRY;
2730
2731         /* pages with index greater or equal to index will be truncated. */
2732         index = cl_index(osc2cl(obj), size);
2733         partial = size > cl_offset(osc2cl(obj), index);
2734
2735 again:
2736         osc_object_lock(obj);
2737         ext = osc_extent_search(obj, index);
2738         if (ext == NULL)
2739                 ext = first_extent(obj);
2740         else if (ext->oe_end < index)
2741                 ext = next_extent(ext);
2742         while (ext != NULL) {
2743                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2744
2745                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2746                         /* if ext is in urgent state, it means there must exist
2747                          * a page already having been flushed by write_page().
2748                          * We have to wait for this extent because we can't
2749                          * truncate that page. */
2750                         OSC_EXTENT_DUMP(D_CACHE, ext,
2751                                         "waiting for busy extent\n");
2752                         waiting = osc_extent_get(ext);
2753                         break;
2754                 }
2755
2756                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:"LPU64".\n", size);
2757
2758                 osc_extent_get(ext);
2759                 if (ext->oe_state == OES_ACTIVE) {
2760                         /* though we grab inode mutex for write path, but we
2761                          * release it before releasing extent(in osc_io_end()),
2762                          * so there is a race window that an extent is still
2763                          * in OES_ACTIVE when truncate starts. */
2764                         LASSERT(!ext->oe_trunc_pending);
2765                         ext->oe_trunc_pending = 1;
2766                 } else {
2767                         EASSERT(ext->oe_state == OES_CACHE, ext);
2768                         osc_extent_state_set(ext, OES_TRUNC);
2769                         osc_update_pending(obj, OBD_BRW_WRITE,
2770                                            -ext->oe_nr_pages);
2771                 }
2772                 EASSERT(list_empty(&ext->oe_link), ext);
2773                 list_add_tail(&ext->oe_link, &list);
2774
2775                 ext = next_extent(ext);
2776         }
2777         osc_object_unlock(obj);
2778
2779         osc_list_maint(cli, obj);
2780
2781         while (!list_empty(&list)) {
2782                 int rc;
2783
2784                 ext = list_entry(list.next, struct osc_extent, oe_link);
2785                 list_del_init(&ext->oe_link);
2786
2787                 /* extent may be in OES_ACTIVE state because inode mutex
2788                  * is released before osc_io_end() in file write case */
2789                 if (ext->oe_state != OES_TRUNC)
2790                         osc_extent_wait(env, ext, OES_TRUNC);
2791
2792                 rc = osc_extent_truncate(ext, index, partial);
2793                 if (rc < 0) {
2794                         if (result == 0)
2795                                 result = rc;
2796
2797                         OSC_EXTENT_DUMP(D_ERROR, ext,
2798                                         "truncate error %d\n", rc);
2799                 } else if (ext->oe_nr_pages == 0) {
2800                         osc_extent_remove(ext);
2801                 } else {
2802                         /* this must be an overlapped extent which means only
2803                          * part of pages in this extent have been truncated.
2804                          */
2805                         EASSERTF(ext->oe_start <= index, ext,
2806                                  "trunc index = %lu/%d.\n", index, partial);
2807                         /* fix index to skip this partially truncated extent */
2808                         index = ext->oe_end + 1;
2809                         partial = false;
2810
2811                         /* we need to hold this extent in OES_TRUNC state so
2812                          * that no writeback will happen. This is to avoid
2813                          * BUG 17397.
2814                          * Only partial truncate can reach here, if @size is
2815                          * not zero, the caller should provide a valid @extp. */
2816                         LASSERT(*extp == NULL);
2817                         *extp = osc_extent_get(ext);
2818                         OSC_EXTENT_DUMP(D_CACHE, ext,
2819                                         "trunc at "LPU64"\n", size);
2820                 }
2821                 osc_extent_put(env, ext);
2822         }
2823         if (waiting != NULL) {
2824                 int rc;
2825
2826                 /* ignore the result of osc_extent_wait the write initiator
2827                  * should take care of it. */
2828                 rc = osc_extent_wait(env, waiting, OES_INV);
2829                 if (rc < 0)
2830                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2831
2832                 osc_extent_put(env, waiting);
2833                 waiting = NULL;
2834                 goto again;
2835         }
2836         RETURN(result);
2837 }
2838
2839 /**
2840  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2841  */
2842 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext)
2843 {
2844         if (ext != NULL) {
2845                 struct osc_object *obj = ext->oe_obj;
2846                 bool unplug = false;
2847
2848                 EASSERT(ext->oe_nr_pages > 0, ext);
2849                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2850                 EASSERT(!ext->oe_urgent, ext);
2851
2852                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2853                 osc_object_lock(obj);
2854                 osc_extent_state_set(ext, OES_CACHE);
2855                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2856                         ext->oe_urgent = 1;
2857                         list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2858                         unplug = true;
2859                 }
2860                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2861                 osc_object_unlock(obj);
2862                 osc_extent_put(env, ext);
2863
2864                 if (unplug)
2865                         osc_io_unplug_async(env, osc_cli(obj), obj);
2866         }
2867 }
2868
2869 /**
2870  * Wait for extents in a specific range to be written out.
2871  * The caller must have called osc_cache_writeback_range() to issue IO
2872  * otherwise it will take a long time for this function to finish.
2873  *
2874  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2875  * nobody else can dirty this range of file while we're waiting for
2876  * extents to be written.
2877  */
2878 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2879                          pgoff_t start, pgoff_t end)
2880 {
2881         struct osc_extent *ext;
2882         pgoff_t index = start;
2883         int     result = 0;
2884         ENTRY;
2885
2886 again:
2887         osc_object_lock(obj);
2888         ext = osc_extent_search(obj, index);
2889         if (ext == NULL)
2890                 ext = first_extent(obj);
2891         else if (ext->oe_end < index)
2892                 ext = next_extent(ext);
2893         while (ext != NULL) {
2894                 int rc;
2895
2896                 if (ext->oe_start > end)
2897                         break;
2898
2899                 if (!ext->oe_fsync_wait) {
2900                         ext = next_extent(ext);
2901                         continue;
2902                 }
2903
2904                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2905                              ext->oe_hp || ext->oe_urgent), ext);
2906                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2907                              !ext->oe_hp && ext->oe_urgent), ext);
2908
2909                 index = ext->oe_end + 1;
2910                 osc_extent_get(ext);
2911                 osc_object_unlock(obj);
2912
2913                 rc = osc_extent_wait(env, ext, OES_INV);
2914                 if (result == 0)
2915                         result = rc;
2916                 osc_extent_put(env, ext);
2917                 goto again;
2918         }
2919         osc_object_unlock(obj);
2920
2921         OSC_IO_DEBUG(obj, "sync file range.\n");
2922         RETURN(result);
2923 }
2924
2925 /**
2926  * Called to write out a range of osc object.
2927  *
2928  * @hp     : should be set this is caused by lock cancel;
2929  * @discard: is set if dirty pages should be dropped - file will be deleted or
2930  *         truncated, this implies there is no partially discarding extents.
2931  *
2932  * Return how many pages will be issued, or error code if error occurred.
2933  */
2934 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2935                               pgoff_t start, pgoff_t end, int hp, int discard)
2936 {
2937         struct osc_extent *ext;
2938         struct list_head discard_list = LIST_HEAD_INIT(discard_list);
2939         bool unplug = false;
2940         int result = 0;
2941         ENTRY;
2942
2943         osc_object_lock(obj);
2944         ext = osc_extent_search(obj, start);
2945         if (ext == NULL)
2946                 ext = first_extent(obj);
2947         else if (ext->oe_end < start)
2948                 ext = next_extent(ext);
2949         while (ext != NULL) {
2950                 if (ext->oe_start > end)
2951                         break;
2952
2953                 ext->oe_fsync_wait = 1;
2954                 switch (ext->oe_state) {
2955                 case OES_CACHE:
2956                         result += ext->oe_nr_pages;
2957                         if (!discard) {
2958                                 struct list_head *list = NULL;
2959                                 if (hp) {
2960                                         EASSERT(!ext->oe_hp, ext);
2961                                         ext->oe_hp = 1;
2962                                         list = &obj->oo_hp_exts;
2963                                 } else if (!ext->oe_urgent) {
2964                                         ext->oe_urgent = 1;
2965                                         list = &obj->oo_urgent_exts;
2966                                 }
2967                                 if (list != NULL)
2968                                         list_move_tail(&ext->oe_link, list);
2969                                 unplug = true;
2970                         } else {
2971                                 /* the only discarder is lock cancelling, so
2972                                  * [start, end] must contain this extent */
2973                                 EASSERT(ext->oe_start >= start &&
2974                                         ext->oe_max_end <= end, ext);
2975                                 osc_extent_state_set(ext, OES_LOCKING);
2976                                 ext->oe_owner = current;
2977                                 list_move_tail(&ext->oe_link,
2978                                                    &discard_list);
2979                                 osc_update_pending(obj, OBD_BRW_WRITE,
2980                                                    -ext->oe_nr_pages);
2981                         }
2982                         break;
2983                 case OES_ACTIVE:
2984                         /* It's pretty bad to wait for ACTIVE extents, because
2985                          * we don't know how long we will wait for it to be
2986                          * flushed since it may be blocked at awaiting more
2987                          * grants. We do this for the correctness of fsync. */
2988                         LASSERT(hp == 0 && discard == 0);
2989                         ext->oe_urgent = 1;
2990                         break;
2991                 case OES_TRUNC:
2992                         /* this extent is being truncated, can't do anything
2993                          * for it now. it will be set to urgent after truncate
2994                          * is finished in osc_cache_truncate_end(). */
2995                 default:
2996                         break;
2997                 }
2998                 ext = next_extent(ext);
2999         }
3000         osc_object_unlock(obj);
3001
3002         LASSERT(ergo(!discard, list_empty(&discard_list)));
3003         if (!list_empty(&discard_list)) {
3004                 struct osc_extent *tmp;
3005                 int rc;
3006
3007                 osc_list_maint(osc_cli(obj), obj);
3008                 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
3009                         list_del_init(&ext->oe_link);
3010                         EASSERT(ext->oe_state == OES_LOCKING, ext);
3011
3012                         /* Discard caching pages. We don't actually write this
3013                          * extent out but we complete it as if we did. */
3014                         rc = osc_extent_make_ready(env, ext);
3015                         if (unlikely(rc < 0)) {
3016                                 OSC_EXTENT_DUMP(D_ERROR, ext,
3017                                                 "make_ready returned %d\n", rc);
3018                                 if (result >= 0)
3019                                         result = rc;
3020                         }
3021
3022                         /* finish the extent as if the pages were sent */
3023                         osc_extent_finish(env, ext, 0, 0);
3024                 }
3025         }
3026
3027         if (unplug)
3028                 osc_io_unplug(env, osc_cli(obj), obj);
3029
3030         if (hp || discard) {
3031                 int rc;
3032                 rc = osc_cache_wait_range(env, obj, start, end);
3033                 if (result >= 0 && rc < 0)
3034                         result = rc;
3035         }
3036
3037         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3038         RETURN(result);
3039 }
3040
3041 /**
3042  * Returns a list of pages by a given [start, end] of \a obj.
3043  *
3044  * \param resched If not NULL, then we give up before hogging CPU for too
3045  * long and set *resched = 1, in that case caller should implement a retry
3046  * logic.
3047  *
3048  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3049  * crucial in the face of [offset, EOF] locks.
3050  *
3051  * Return at least one page in @queue unless there is no covered page.
3052  */
3053 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3054                         struct osc_object *osc, pgoff_t start, pgoff_t end,
3055                         osc_page_gang_cbt cb, void *cbdata)
3056 {
3057         struct osc_page *ops;
3058         void            **pvec;
3059         pgoff_t         idx;
3060         unsigned int    nr;
3061         unsigned int    i;
3062         unsigned int    j;
3063         int             res = CLP_GANG_OKAY;
3064         bool            tree_lock = true;
3065         ENTRY;
3066
3067         idx = start;
3068         pvec = osc_env_info(env)->oti_pvec;
3069         spin_lock(&osc->oo_tree_lock);
3070         while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3071                                             idx, OTI_PVEC_SIZE)) > 0) {
3072                 struct cl_page *page;
3073                 bool end_of_region = false;
3074
3075                 for (i = 0, j = 0; i < nr; ++i) {
3076                         ops = pvec[i];
3077                         pvec[i] = NULL;
3078
3079                         idx = osc_index(ops);
3080                         if (idx > end) {
3081                                 end_of_region = true;
3082                                 break;
3083                         }
3084
3085                         page = ops->ops_cl.cpl_page;
3086                         LASSERT(page->cp_type == CPT_CACHEABLE);
3087                         if (page->cp_state == CPS_FREEING)
3088                                 continue;
3089
3090                         cl_page_get(page);
3091                         lu_ref_add_atomic(&page->cp_reference,
3092                                           "gang_lookup", current);
3093                         pvec[j++] = ops;
3094                 }
3095                 ++idx;
3096
3097                 /*
3098                  * Here a delicate locking dance is performed. Current thread
3099                  * holds a reference to a page, but has to own it before it
3100                  * can be placed into queue. Owning implies waiting, so
3101                  * radix-tree lock is to be released. After a wait one has to
3102                  * check that pages weren't truncated (cl_page_own() returns
3103                  * error in the latter case).
3104                  */
3105                 spin_unlock(&osc->oo_tree_lock);
3106                 tree_lock = false;
3107
3108                 for (i = 0; i < j; ++i) {
3109                         ops = pvec[i];
3110                         if (res == CLP_GANG_OKAY)
3111                                 res = (*cb)(env, io, ops, cbdata);
3112
3113                         page = ops->ops_cl.cpl_page;
3114                         lu_ref_del(&page->cp_reference, "gang_lookup", current);
3115                         cl_page_put(env, page);
3116                 }
3117                 if (nr < OTI_PVEC_SIZE || end_of_region)
3118                         break;
3119
3120                 if (res == CLP_GANG_OKAY && need_resched())
3121                         res = CLP_GANG_RESCHED;
3122                 if (res != CLP_GANG_OKAY)
3123                         break;
3124
3125                 spin_lock(&osc->oo_tree_lock);
3126                 tree_lock = true;
3127         }
3128         if (tree_lock)
3129                 spin_unlock(&osc->oo_tree_lock);
3130         RETURN(res);
3131 }
3132
3133 /**
3134  * Check if page @page is covered by an extra lock or discard it.
3135  */
3136 static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3137                                 struct osc_page *ops, void *cbdata)
3138 {
3139         struct osc_thread_info *info = osc_env_info(env);
3140         struct osc_object *osc = cbdata;
3141         pgoff_t index;
3142
3143         index = osc_index(ops);
3144         if (index >= info->oti_fn_index) {
3145                 struct ldlm_lock *tmp;
3146                 struct cl_page *page = ops->ops_cl.cpl_page;
3147
3148                 /* refresh non-overlapped index */
3149                 tmp = osc_dlmlock_at_pgoff(env, osc, index,
3150                                            OSC_DAP_FL_TEST_LOCK);
3151                 if (tmp != NULL) {
3152                         __u64 end = tmp->l_policy_data.l_extent.end;
3153                         /* Cache the first-non-overlapped index so as to skip
3154                          * all pages within [index, oti_fn_index). This is safe
3155                          * because if tmp lock is canceled, it will discard
3156                          * these pages. */
3157                         info->oti_fn_index = cl_index(osc2cl(osc), end + 1);
3158                         if (end == OBD_OBJECT_EOF)
3159                                 info->oti_fn_index = CL_PAGE_EOF;
3160                         LDLM_LOCK_PUT(tmp);
3161                 } else if (cl_page_own(env, io, page) == 0) {
3162                         /* discard the page */
3163                         cl_page_discard(env, io, page);
3164                         cl_page_disown(env, io, page);
3165                 } else {
3166                         LASSERT(page->cp_state == CPS_FREEING);
3167                 }
3168         }
3169
3170         info->oti_next_index = index + 1;
3171         return CLP_GANG_OKAY;
3172 }
3173
3174 static int discard_cb(const struct lu_env *env, struct cl_io *io,
3175                       struct osc_page *ops, void *cbdata)
3176 {
3177         struct osc_thread_info *info = osc_env_info(env);
3178         struct cl_page *page = ops->ops_cl.cpl_page;
3179
3180         /* page is top page. */
3181         info->oti_next_index = osc_index(ops) + 1;
3182         if (cl_page_own(env, io, page) == 0) {
3183                 if (!ergo(page->cp_type == CPT_CACHEABLE,
3184                           !PageDirty(cl_page_vmpage(page))))
3185                         CL_PAGE_DEBUG(D_ERROR, env, page,
3186                                         "discard dirty page?\n");
3187
3188                 /* discard the page */
3189                 cl_page_discard(env, io, page);
3190                 cl_page_disown(env, io, page);
3191         } else {
3192                 LASSERT(page->cp_state == CPS_FREEING);
3193         }
3194
3195         return CLP_GANG_OKAY;
3196 }
3197
3198 /**
3199  * Discard pages protected by the given lock. This function traverses radix
3200  * tree to find all covering pages and discard them. If a page is being covered
3201  * by other locks, it should remain in cache.
3202  *
3203  * If error happens on any step, the process continues anyway (the reasoning
3204  * behind this being that lock cancellation cannot be delayed indefinitely).
3205  */
3206 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3207                            pgoff_t start, pgoff_t end, enum cl_lock_mode mode)
3208 {
3209         struct osc_thread_info *info = osc_env_info(env);
3210         struct cl_io *io = &info->oti_io;
3211         osc_page_gang_cbt cb;
3212         int res;
3213         int result;
3214
3215         ENTRY;
3216
3217         io->ci_obj = cl_object_top(osc2cl(osc));
3218         io->ci_ignore_layout = 1;
3219         result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3220         if (result != 0)
3221                 GOTO(out, result);
3222
3223         cb = mode == CLM_READ ? check_and_discard_cb : discard_cb;
3224         info->oti_fn_index = info->oti_next_index = start;
3225         do {
3226                 res = osc_page_gang_lookup(env, io, osc,
3227                                            info->oti_next_index, end, cb, osc);
3228                 if (info->oti_next_index > end)
3229                         break;
3230
3231                 if (res == CLP_GANG_RESCHED)
3232                         cond_resched();
3233         } while (res != CLP_GANG_OKAY);
3234 out:
3235         cl_io_fini(env, io);
3236         RETURN(result);
3237 }
3238
3239
3240 /** @} osc */