lustre/osc/osc_cache.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  *
  28  */
  29 /*
  30  * This file is part of Lustre, http://www.lustre.org/
  31  * Lustre is a trademark of Sun Microsystems, Inc.
  32  *
  33  * osc cache management.
  34  *
  35  * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_OSC
  39
  40 #include <lustre_osc.h>
  41
  42 #include "osc_internal.h"
  43
  44 static int extent_debug; /* set it to be true for more debug */
  45
  46 static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
  47 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
  48                            enum osc_extent_state state);
  49 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
  50                               struct osc_async_page *oap, int sent, int rc);
  51 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
  52                           int cmd);
  53 static int osc_refresh_count(const struct lu_env *env,
  54                              struct osc_async_page *oap, int cmd);
  55 static int osc_io_unplug_async(const struct lu_env *env,
  56                                struct client_obd *cli, struct osc_object *osc);
  57 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
  58                            unsigned int lost_grant, unsigned int dirty_grant);
  59
  60 static void osc_extent_tree_dump0(int mask, struct osc_object *obj,
  61                                   const char *func, int line);
  62 #define osc_extent_tree_dump(mask, obj) \
  63         osc_extent_tree_dump0(mask, obj, __func__, __LINE__)
  64
  65 static void osc_unreserve_grant(struct client_obd *cli, unsigned int reserved,
  66                                 unsigned int unused);
  67
  68 /** \addtogroup osc
  69  *  @{
  70  */
  71
  72 /* ------------------ osc extent ------------------ */
  73 static inline char *ext_flags(struct osc_extent *ext, char *flags)
  74 {
  75         char *buf = flags;
  76         *buf++ = ext->oe_rw ? 'r' : 'w';
  77         if (ext->oe_intree)
  78                 *buf++ = 'i';
  79         if (ext->oe_sync)
  80                 *buf++ = 'S';
  81         if (ext->oe_srvlock)
  82                 *buf++ = 's';
  83         if (ext->oe_hp)
  84                 *buf++ = 'h';
  85         if (ext->oe_urgent)
  86                 *buf++ = 'u';
  87         if (ext->oe_memalloc)
  88                 *buf++ = 'm';
  89         if (ext->oe_trunc_pending)
  90                 *buf++ = 't';
  91         if (ext->oe_fsync_wait)
  92                 *buf++ = 'Y';
  93         *buf = 0;
  94         return flags;
  95 }
  96
  97 #define EXTSTR       "[%lu -> %lu/%lu]"
  98 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
  99 static const char *oes_strings[] = {
 100         "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
 101
 102 #define OSC_EXTENT_DUMP_WITH_LOC(file, func, line, mask, extent, fmt, ...) do {\
 103         static struct cfs_debug_limit_state cdls;                             \
 104         struct osc_extent *__ext = (extent);                                  \
 105         char __buf[16];                                                       \
 106                                                                               \
 107         __CDEBUG_WITH_LOC(file, func, line, mask, &cdls,                      \
 108                 "extent %p@{" EXTSTR ", "                                     \
 109                 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt,          \
 110                 /* ----- extent part 0 ----- */                               \
 111                 __ext, EXTPARA(__ext),                                        \
 112                 /* ----- part 1 ----- */                                      \
 113                 atomic_read(&__ext->oe_refc),                                 \
 114                 atomic_read(&__ext->oe_users),                                \
 115                 list_empty_marker(&__ext->oe_link),                           \
 116                 oes_strings[__ext->oe_state], ext_flags(__ext, __buf),        \
 117                 __ext->oe_obj,                                                \
 118                 /* ----- part 2 ----- */                                      \
 119                 __ext->oe_grants, __ext->oe_nr_pages,                         \
 120                 list_empty_marker(&__ext->oe_pages),                          \
 121                 waitqueue_active(&__ext->oe_waitq) ? '+' : '-',               \
 122                 __ext->oe_dlmlock, __ext->oe_mppr, __ext->oe_owner,           \
 123                 /* ----- part 4 ----- */                                      \
 124                 ## __VA_ARGS__);                                              \
 125         if (mask == D_ERROR && __ext->oe_dlmlock != NULL)                     \
 126                 LDLM_ERROR(__ext->oe_dlmlock, "extent: %p", __ext);           \
 127         else                                                                  \
 128                 LDLM_DEBUG(__ext->oe_dlmlock, "extent: %p", __ext);           \
 129 } while (0)
 130
 131 #define OSC_EXTENT_DUMP(mask, ext, fmt, ...)                            \
 132         OSC_EXTENT_DUMP_WITH_LOC(__FILE__, __func__, __LINE__,          \
 133                                  mask, ext, fmt, ## __VA_ARGS__)
 134
 135 #undef EASSERTF
 136 #define EASSERTF(expr, ext, fmt, args...) do {                          \
 137         if (!(expr)) {                                                  \
 138                 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args);           \
 139                 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj);           \
 140                 LASSERT(expr);                                          \
 141         }                                                               \
 142 } while (0)
 143
 144 #undef EASSERT
 145 #define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
 146
 147 static inline struct osc_extent *rb_extent(struct rb_node *n)
 148 {
 149         if (n == NULL)
 150                 return NULL;
 151
 152         return container_of(n, struct osc_extent, oe_node);
 153 }
 154
 155 static inline struct osc_extent *next_extent(struct osc_extent *ext)
 156 {
 157         if (ext == NULL)
 158                 return NULL;
 159
 160         LASSERT(ext->oe_intree);
 161         return rb_extent(rb_next(&ext->oe_node));
 162 }
 163
 164 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
 165 {
 166         if (ext == NULL)
 167                 return NULL;
 168
 169         LASSERT(ext->oe_intree);
 170         return rb_extent(rb_prev(&ext->oe_node));
 171 }
 172
 173 static inline struct osc_extent *first_extent(struct osc_object *obj)
 174 {
 175         return rb_extent(rb_first(&obj->oo_root));
 176 }
 177
 178 /* object must be locked by caller. */
 179 static int osc_extent_sanity_check0(struct osc_extent *ext,
 180                                     const char *func, const int line)
 181 {
 182         struct osc_object *obj = ext->oe_obj;
 183         struct osc_async_page *oap;
 184         size_t page_count;
 185         int rc = 0;
 186
 187         assert_osc_object_is_locked(obj);
 188
 189         if (ext->oe_state >= OES_STATE_MAX)
 190                 GOTO(out, rc = 10);
 191
 192         if (atomic_read(&ext->oe_refc) <= 0)
 193                 GOTO(out, rc = 20);
 194
 195         if (atomic_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
 196                 GOTO(out, rc = 30);
 197
 198         switch (ext->oe_state) {
 199         case OES_INV:
 200                 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
 201                         GOTO(out, rc = 35);
 202                 GOTO(out, rc = 0);
 203                 break;
 204         case OES_ACTIVE:
 205                 if (atomic_read(&ext->oe_users) == 0)
 206                         GOTO(out, rc = 40);
 207                 if (ext->oe_hp)
 208                         GOTO(out, rc = 50);
 209                 if (ext->oe_fsync_wait && !ext->oe_urgent)
 210                         GOTO(out, rc = 55);
 211                 break;
 212         case OES_CACHE:
 213                 if (ext->oe_grants == 0)
 214                         GOTO(out, rc = 60);
 215                 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
 216                         GOTO(out, rc = 65);
 217                 /* fallthrough */
 218         default:
 219                 if (atomic_read(&ext->oe_users) > 0)
 220                         GOTO(out, rc = 70);
 221         }
 222
 223         if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
 224                 GOTO(out, rc = 80);
 225
 226         if (ext->oe_sync && ext->oe_grants > 0)
 227                 GOTO(out, rc = 90);
 228
 229         if (ext->oe_dlmlock != NULL &&
 230             ext->oe_dlmlock->l_resource->lr_type == LDLM_EXTENT &&
 231             !ldlm_is_failed(ext->oe_dlmlock)) {
 232                 struct ldlm_extent *extent;
 233
 234                 extent = &ext->oe_dlmlock->l_policy_data.l_extent;
 235                 if (!(extent->start <= cl_offset(osc2cl(obj), ext->oe_start) &&
 236                       extent->end   >= cl_offset(osc2cl(obj), ext->oe_max_end)))
 237                         GOTO(out, rc = 100);
 238
 239                 if (!(ext->oe_dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP)))
 240                         GOTO(out, rc = 102);
 241         }
 242
 243         if (ext->oe_nr_pages > ext->oe_mppr)
 244                 GOTO(out, rc = 105);
 245
 246         /* Do not verify page list if extent is in RPC. This is because an
 247          * in-RPC extent is supposed to be exclusively accessible w/o lock. */
 248         if (ext->oe_state > OES_CACHE)
 249                 GOTO(out, rc = 0);
 250
 251         if (!extent_debug)
 252                 GOTO(out, rc = 0);
 253
 254         page_count = 0;
 255         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
 256                 pgoff_t index = osc_index(oap2osc(oap));
 257                 ++page_count;
 258                 if (index > ext->oe_end || index < ext->oe_start)
 259                         GOTO(out, rc = 110);
 260         }
 261         if (page_count != ext->oe_nr_pages)
 262                 GOTO(out, rc = 120);
 263
 264 out:
 265         if (rc != 0)
 266                 OSC_EXTENT_DUMP_WITH_LOC(__FILE__, func, line, D_ERROR, ext,
 267                                          "sanity check %p failed: rc = %d\n",
 268                                          ext, rc);
 269         return rc;
 270 }
 271
 272 #define sanity_check_nolock(ext) \
 273         osc_extent_sanity_check0(ext, __func__, __LINE__)
 274
 275 #define sanity_check(ext) ({                                                   \
 276         int __res;                                                             \
 277         osc_object_lock((ext)->oe_obj);                                        \
 278         __res = sanity_check_nolock(ext);                                      \
 279         osc_object_unlock((ext)->oe_obj);                                      \
 280         __res;                                                                 \
 281 })
 282
 283
 284 /**
 285  * sanity check - to make sure there is no overlapped extent in the tree.
 286  */
 287 static int osc_extent_is_overlapped(struct osc_object *obj,
 288                                     struct osc_extent *ext)
 289 {
 290         struct osc_extent *tmp;
 291
 292         assert_osc_object_is_locked(obj);
 293
 294         if (!extent_debug)
 295                 return 0;
 296
 297         for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
 298                 if (tmp == ext)
 299                         continue;
 300                 if (tmp->oe_end >= ext->oe_start &&
 301                     tmp->oe_start <= ext->oe_end)
 302                         return 1;
 303         }
 304         return 0;
 305 }
 306
 307 static void osc_extent_state_set(struct osc_extent *ext, int state)
 308 {
 309         assert_osc_object_is_locked(ext->oe_obj);
 310         LASSERT(state >= OES_INV && state < OES_STATE_MAX);
 311
 312         /* Never try to sanity check a state changing extent :-) */
 313         /* LASSERT(sanity_check_nolock(ext) == 0); */
 314
 315         /* TODO: validate the state machine */
 316         ext->oe_state = state;
 317         wake_up_all(&ext->oe_waitq);
 318 }
 319
 320 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
 321 {
 322         struct osc_extent *ext;
 323
 324         OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_NOFS);
 325         if (ext == NULL)
 326                 return NULL;
 327
 328         RB_CLEAR_NODE(&ext->oe_node);
 329         ext->oe_obj = obj;
 330         cl_object_get(osc2cl(obj));
 331         atomic_set(&ext->oe_refc, 1);
 332         atomic_set(&ext->oe_users, 0);
 333         INIT_LIST_HEAD(&ext->oe_link);
 334         ext->oe_state = OES_INV;
 335         INIT_LIST_HEAD(&ext->oe_pages);
 336         init_waitqueue_head(&ext->oe_waitq);
 337         ext->oe_dlmlock = NULL;
 338
 339         return ext;
 340 }
 341
 342 static void osc_extent_free(struct osc_extent *ext)
 343 {
 344         OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
 345 }
 346
 347 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
 348 {
 349         LASSERT(atomic_read(&ext->oe_refc) >= 0);
 350         atomic_inc(&ext->oe_refc);
 351         return ext;
 352 }
 353
 354 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
 355 {
 356         LASSERT(atomic_read(&ext->oe_refc) > 0);
 357         if (atomic_dec_and_test(&ext->oe_refc)) {
 358                 LASSERT(list_empty(&ext->oe_link));
 359                 LASSERT(atomic_read(&ext->oe_users) == 0);
 360                 LASSERT(ext->oe_state == OES_INV);
 361                 LASSERT(!ext->oe_intree);
 362
 363                 if (ext->oe_dlmlock != NULL) {
 364                         lu_ref_add(&ext->oe_dlmlock->l_reference,
 365                                    "osc_extent", ext);
 366                         LDLM_LOCK_PUT(ext->oe_dlmlock);
 367                         ext->oe_dlmlock = NULL;
 368                 }
 369                 cl_object_put(env, osc2cl(ext->oe_obj));
 370                 osc_extent_free(ext);
 371         }
 372 }
 373
 374 /**
 375  * osc_extent_put_trust() is a special version of osc_extent_put() when
 376  * it's known that the caller is not the last user. This is to address the
 377  * problem of lacking of lu_env ;-).
 378  */
 379 static void osc_extent_put_trust(struct osc_extent *ext)
 380 {
 381         LASSERT(atomic_read(&ext->oe_refc) > 1);
 382         assert_osc_object_is_locked(ext->oe_obj);
 383         atomic_dec(&ext->oe_refc);
 384 }
 385
 386 /**
 387  * Return the extent which includes pgoff @index, or return the greatest
 388  * previous extent in the tree.
 389  */
 390 static struct osc_extent *osc_extent_search(struct osc_object *obj,
 391                                             pgoff_t index)
 392 {
 393         struct rb_node    *n = obj->oo_root.rb_node;
 394         struct osc_extent *tmp, *p = NULL;
 395
 396         assert_osc_object_is_locked(obj);
 397         while (n != NULL) {
 398                 tmp = rb_extent(n);
 399                 if (index < tmp->oe_start) {
 400                         n = n->rb_left;
 401                 } else if (index > tmp->oe_end) {
 402                         p = rb_extent(n);
 403                         n = n->rb_right;
 404                 } else {
 405                         return tmp;
 406                 }
 407         }
 408         return p;
 409 }
 410
 411 /*
 412  * Return the extent covering @index, otherwise return NULL.
 413  * caller must have held object lock.
 414  */
 415 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
 416                                             pgoff_t index)
 417 {
 418         struct osc_extent *ext;
 419
 420         ext = osc_extent_search(obj, index);
 421         if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
 422                 return osc_extent_get(ext);
 423         return NULL;
 424 }
 425
 426 /* caller must have held object lock. */
 427 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
 428 {
 429         struct rb_node   **n      = &obj->oo_root.rb_node;
 430         struct rb_node    *parent = NULL;
 431         struct osc_extent *tmp;
 432
 433         LASSERT(ext->oe_intree == 0);
 434         LASSERT(ext->oe_obj == obj);
 435         assert_osc_object_is_locked(obj);
 436         while (*n != NULL) {
 437                 tmp = rb_extent(*n);
 438                 parent = *n;
 439
 440                 if (ext->oe_end < tmp->oe_start)
 441                         n = &(*n)->rb_left;
 442                 else if (ext->oe_start > tmp->oe_end)
 443                         n = &(*n)->rb_right;
 444                 else
 445                         EASSERTF(0, tmp, EXTSTR"\n", EXTPARA(ext));
 446         }
 447         rb_link_node(&ext->oe_node, parent, n);
 448         rb_insert_color(&ext->oe_node, &obj->oo_root);
 449         osc_extent_get(ext);
 450         ext->oe_intree = 1;
 451 }
 452
 453 /* caller must have held object lock. */
 454 static void osc_extent_erase(struct osc_extent *ext)
 455 {
 456         struct osc_object *obj = ext->oe_obj;
 457         assert_osc_object_is_locked(obj);
 458         if (ext->oe_intree) {
 459                 rb_erase(&ext->oe_node, &obj->oo_root);
 460                 ext->oe_intree = 0;
 461                 /* rbtree held a refcount */
 462                 osc_extent_put_trust(ext);
 463         }
 464 }
 465
 466 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
 467 {
 468         struct osc_object *obj = ext->oe_obj;
 469
 470         assert_osc_object_is_locked(obj);
 471         LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
 472         if (ext->oe_state == OES_CACHE) {
 473                 osc_extent_state_set(ext, OES_ACTIVE);
 474                 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
 475         }
 476         atomic_inc(&ext->oe_users);
 477         list_del_init(&ext->oe_link);
 478         return osc_extent_get(ext);
 479 }
 480
 481 static void __osc_extent_remove(struct osc_extent *ext)
 482 {
 483         assert_osc_object_is_locked(ext->oe_obj);
 484         LASSERT(list_empty(&ext->oe_pages));
 485         osc_extent_erase(ext);
 486         list_del_init(&ext->oe_link);
 487         osc_extent_state_set(ext, OES_INV);
 488         OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
 489 }
 490
 491 static void osc_extent_remove(struct osc_extent *ext)
 492 {
 493         struct osc_object *obj = ext->oe_obj;
 494
 495         osc_object_lock(obj);
 496         __osc_extent_remove(ext);
 497         osc_object_unlock(obj);
 498 }
 499
 500 /**
 501  * This function is used to merge extents to get better performance. It checks
 502  * if @cur and @victim are contiguous at block level.
 503  */
 504 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
 505                             struct osc_extent *victim)
 506 {
 507         struct osc_object       *obj = cur->oe_obj;
 508         struct client_obd       *cli = osc_cli(obj);
 509         pgoff_t                  chunk_start;
 510         pgoff_t                  chunk_end;
 511         int                      ppc_bits;
 512
 513         LASSERT(cur->oe_state == OES_CACHE);
 514         assert_osc_object_is_locked(obj);
 515         if (victim == NULL)
 516                 return -EINVAL;
 517
 518         if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
 519                 return -EBUSY;
 520
 521         if (cur->oe_max_end != victim->oe_max_end)
 522                 return -ERANGE;
 523
 524         LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
 525         ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_SHIFT;
 526         chunk_start = cur->oe_start >> ppc_bits;
 527         chunk_end   = cur->oe_end   >> ppc_bits;
 528         if (chunk_start   != (victim->oe_end >> ppc_bits) + 1 &&
 529             chunk_end + 1 != victim->oe_start >> ppc_bits)
 530                 return -ERANGE;
 531
 532         /* overall extent size should not exceed the max supported limit
 533          * reported by the server */
 534         if (cur->oe_end - cur->oe_start + 1 +
 535             victim->oe_end - victim->oe_start + 1 > cli->cl_max_extent_pages)
 536                 return -ERANGE;
 537
 538         OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
 539
 540         cur->oe_start     = min(cur->oe_start, victim->oe_start);
 541         cur->oe_end       = max(cur->oe_end,   victim->oe_end);
 542         /* per-extent tax should be accounted only once for the whole extent */
 543         cur->oe_grants   += victim->oe_grants - cli->cl_grant_extent_tax;
 544         cur->oe_nr_pages += victim->oe_nr_pages;
 545         /* only the following bits are needed to merge */
 546         cur->oe_urgent   |= victim->oe_urgent;
 547         cur->oe_memalloc |= victim->oe_memalloc;
 548         list_splice_init(&victim->oe_pages, &cur->oe_pages);
 549         list_del_init(&victim->oe_link);
 550         victim->oe_nr_pages = 0;
 551
 552         osc_extent_get(victim);
 553         __osc_extent_remove(victim);
 554         osc_extent_put(env, victim);
 555
 556         OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
 557         return 0;
 558 }
 559
 560 /**
 561  * Drop user count of osc_extent, and unplug IO asynchronously.
 562  */
 563 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
 564 {
 565         struct osc_object *obj = ext->oe_obj;
 566         struct client_obd *cli = osc_cli(obj);
 567         int rc = 0;
 568         ENTRY;
 569
 570         LASSERT(atomic_read(&ext->oe_users) > 0);
 571         LASSERT(sanity_check(ext) == 0);
 572         LASSERT(ext->oe_grants > 0);
 573
 574         if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
 575                 LASSERT(ext->oe_state == OES_ACTIVE);
 576                 if (ext->oe_trunc_pending) {
 577                         /* a truncate process is waiting for this extent.
 578                          * This may happen due to a race, check
 579                          * osc_cache_truncate_start(). */
 580                         osc_extent_state_set(ext, OES_TRUNC);
 581                         ext->oe_trunc_pending = 0;
 582                         osc_object_unlock(obj);
 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
 596                         if (ext->oe_urgent)
 597                                 list_move_tail(&ext->oe_link,
 598                                                &obj->oo_urgent_exts);
 599                         else if (ext->oe_nr_pages == ext->oe_mppr) {
 600                                 list_move_tail(&ext->oe_link,
 601                                                &obj->oo_full_exts);
 602                         }
 603                         osc_object_unlock(obj);
 604                         if (grant > 0)
 605                                 osc_unreserve_grant(cli, 0, grant);
 606                 }
 607
 608                 osc_io_unplug_async(env, cli, obj);
 609         }
 610         osc_extent_put(env, ext);
 611         RETURN(rc);
 612 }
 613
 614 static inline bool
 615 overlapped(const struct osc_extent *ex1, const struct osc_extent *ex2)
 616 {
 617         return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
 618 }
 619
 620 /**
 621  * Find or create an extent which includes @index, core function to manage
 622  * extent tree.
 623  */
 624 static struct osc_extent *osc_extent_find(const struct lu_env *env,
 625                                           struct osc_object *obj, pgoff_t index,
 626                                           unsigned int *grants)
 627 {
 628         struct client_obd *cli = osc_cli(obj);
 629         struct osc_lock   *olck;
 630         struct cl_lock_descr *descr;
 631         struct osc_extent *cur;
 632         struct osc_extent *ext;
 633         struct osc_extent *conflict = NULL;
 634         struct osc_extent *found = NULL;
 635         pgoff_t    chunk;
 636         pgoff_t    max_end;
 637         unsigned int max_pages; /* max_pages_per_rpc */
 638         unsigned int chunksize;
 639         int        ppc_bits; /* pages per chunk bits */
 640         pgoff_t    chunk_mask;
 641         int        rc;
 642         ENTRY;
 643
 644         cur = osc_extent_alloc(obj);
 645         if (cur == NULL)
 646                 RETURN(ERR_PTR(-ENOMEM));
 647
 648         olck = osc_env_io(env)->oi_write_osclock;
 649         LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
 650         LASSERT(olck->ols_state == OLS_GRANTED);
 651
 652         descr = &olck->ols_cl.cls_lock->cll_descr;
 653         LASSERT(descr->cld_mode >= CLM_WRITE);
 654
 655         LASSERTF(cli->cl_chunkbits >= PAGE_SHIFT,
 656                  "chunkbits: %u\n", cli->cl_chunkbits);
 657         ppc_bits   = cli->cl_chunkbits - PAGE_SHIFT;
 658         chunk_mask = ~((1 << ppc_bits) - 1);
 659         chunksize  = 1 << cli->cl_chunkbits;
 660         chunk      = index >> ppc_bits;
 661
 662         /* align end to RPC edge. */
 663         max_pages = cli->cl_max_pages_per_rpc;
 664         if ((max_pages & ~chunk_mask) != 0) {
 665                 CERROR("max_pages: %#x chunkbits: %u chunk_mask: %#lx\n",
 666                        max_pages, cli->cl_chunkbits, chunk_mask);
 667                 RETURN(ERR_PTR(-EINVAL));
 668         }
 669         max_end = index - (index % max_pages) + max_pages - 1;
 670         max_end = min_t(pgoff_t, max_end, descr->cld_end);
 671
 672         /* initialize new extent by parameters so far */
 673         cur->oe_max_end = max_end;
 674         cur->oe_start   = index & chunk_mask;
 675         cur->oe_end     = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
 676         if (cur->oe_start < descr->cld_start)
 677                 cur->oe_start = descr->cld_start;
 678         if (cur->oe_end > max_end)
 679                 cur->oe_end = max_end;
 680         cur->oe_grants  = 0;
 681         cur->oe_mppr    = max_pages;
 682         if (olck->ols_dlmlock != NULL) {
 683                 LASSERT(olck->ols_hold);
 684                 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
 685                 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
 686         }
 687
 688         /* grants has been allocated by caller */
 689         LASSERTF(*grants >= chunksize + cli->cl_grant_extent_tax,
 690                  "%u/%u/%u.\n", *grants, chunksize, cli->cl_grant_extent_tax);
 691         LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
 692                  EXTPARA(cur));
 693
 694 restart:
 695         osc_object_lock(obj);
 696         ext = osc_extent_search(obj, cur->oe_start);
 697         if (ext == NULL)
 698                 ext = first_extent(obj);
 699         while (ext != NULL) {
 700                 pgoff_t ext_chk_start = ext->oe_start >> ppc_bits;
 701                 pgoff_t ext_chk_end   = ext->oe_end   >> ppc_bits;
 702
 703                 LASSERT(sanity_check_nolock(ext) == 0);
 704                 if (chunk > ext_chk_end + 1 || chunk < ext_chk_start)
 705                         break;
 706
 707                 /* if covering by different locks, no chance to match */
 708                 if (olck->ols_dlmlock != ext->oe_dlmlock) {
 709                         EASSERTF(!overlapped(ext, cur), ext,
 710                                  EXTSTR"\n", EXTPARA(cur));
 711
 712                         ext = next_extent(ext);
 713                         continue;
 714                 }
 715
 716                 /* discontiguous chunks? */
 717                 if (chunk + 1 < ext_chk_start) {
 718                         ext = next_extent(ext);
 719                         continue;
 720                 }
 721
 722                 /* ok, from now on, ext and cur have these attrs:
 723                  * 1. covered by the same lock
 724                  * 2. contiguous at chunk level or overlapping. */
 725
 726                 if (overlapped(ext, cur)) {
 727                         /* cur is the minimum unit, so overlapping means
 728                          * full contain. */
 729                         EASSERTF((ext->oe_start <= cur->oe_start &&
 730                                   ext->oe_end >= cur->oe_end),
 731                                  ext, EXTSTR"\n", EXTPARA(cur));
 732
 733                         if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
 734                                 /* for simplicity, we wait for this extent to
 735                                  * finish before going forward. */
 736                                 conflict = osc_extent_get(ext);
 737                                 break;
 738                         }
 739
 740                         found = osc_extent_hold(ext);
 741                         break;
 742                 }
 743
 744                 /* non-overlapped extent */
 745                 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
 746                         /* we can't do anything for a non OES_CACHE extent, or
 747                          * if there is someone waiting for this extent to be
 748                          * flushed, try next one. */
 749                         ext = next_extent(ext);
 750                         continue;
 751                 }
 752
 753                 /* check if they belong to the same rpc slot before trying to
 754                  * merge. the extents are not overlapped and contiguous at
 755                  * chunk level to get here. */
 756                 if (ext->oe_max_end != max_end) {
 757                         /* if they don't belong to the same RPC slot or
 758                          * max_pages_per_rpc has ever changed, do not merge. */
 759                         ext = next_extent(ext);
 760                         continue;
 761                 }
 762
 763                 /* check whether maximum extent size will be hit */
 764                 if ((ext_chk_end - ext_chk_start + 1 + 1) << ppc_bits >
 765                     cli->cl_max_extent_pages) {
 766                         ext = next_extent(ext);
 767                         continue;
 768                 }
 769
 770                 /* it's required that an extent must be contiguous at chunk
 771                  * level so that we know the whole extent is covered by grant
 772                  * (the pages in the extent are NOT required to be contiguous).
 773                  * Otherwise, it will be too much difficult to know which
 774                  * chunks have grants allocated. */
 775
 776                 /* try to do front merge - extend ext's start */
 777                 if (chunk + 1 == ext_chk_start) {
 778                         /* ext must be chunk size aligned */
 779                         EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
 780
 781                         /* pull ext's start back to cover cur */
 782                         ext->oe_start   = cur->oe_start;
 783                         ext->oe_grants += chunksize;
 784                         LASSERT(*grants >= chunksize);
 785                         *grants -= chunksize;
 786
 787                         found = osc_extent_hold(ext);
 788                 } else if (chunk == ext_chk_end + 1) {
 789                         /* rear merge */
 790                         ext->oe_end     = cur->oe_end;
 791                         ext->oe_grants += chunksize;
 792                         LASSERT(*grants >= chunksize);
 793                         *grants -= chunksize;
 794
 795                         /* try to merge with the next one because we just fill
 796                          * in a gap */
 797                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 798                                 /* we can save extent tax from next extent */
 799                                 *grants += cli->cl_grant_extent_tax;
 800
 801                         found = osc_extent_hold(ext);
 802                 }
 803                 if (found != NULL)
 804                         break;
 805
 806                 ext = next_extent(ext);
 807         }
 808
 809         osc_extent_tree_dump(D_CACHE, obj);
 810         if (found != NULL) {
 811                 LASSERT(conflict == NULL);
 812                 if (!IS_ERR(found)) {
 813                         LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
 814                         OSC_EXTENT_DUMP(D_CACHE, found,
 815                                         "found caching ext for %lu.\n", index);
 816                 }
 817         } else if (conflict == NULL) {
 818                 /* create a new extent */
 819                 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
 820                 cur->oe_grants = chunksize + cli->cl_grant_extent_tax;
 821                 LASSERT(*grants >= cur->oe_grants);
 822                 *grants -= cur->oe_grants;
 823
 824                 cur->oe_state = OES_CACHE;
 825                 found = osc_extent_hold(cur);
 826                 osc_extent_insert(obj, cur);
 827                 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
 828                                 index, descr->cld_end);
 829         }
 830         osc_object_unlock(obj);
 831
 832         if (conflict != NULL) {
 833                 LASSERT(found == NULL);
 834
 835                 /* waiting for IO to finish. Please notice that it's impossible
 836                  * to be an OES_TRUNC extent. */
 837                 rc = osc_extent_wait(env, conflict, OES_INV);
 838                 osc_extent_put(env, conflict);
 839                 conflict = NULL;
 840                 if (rc < 0)
 841                         GOTO(out, found = ERR_PTR(rc));
 842
 843                 goto restart;
 844         }
 845         EXIT;
 846
 847 out:
 848         osc_extent_put(env, cur);
 849         return found;
 850 }
 851
 852 /**
 853  * Called when IO is finished to an extent.
 854  */
 855 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
 856                       int sent, int rc)
 857 {
 858         struct client_obd *cli = osc_cli(ext->oe_obj);
 859         struct osc_async_page *oap;
 860         struct osc_async_page *tmp;
 861         int nr_pages = ext->oe_nr_pages;
 862         int lost_grant = 0;
 863         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
 864         loff_t last_off = 0;
 865         int last_count = -1;
 866         ENTRY;
 867
 868         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
 869
 870         ext->oe_rc = rc ?: ext->oe_nr_pages;
 871         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
 872
 873         osc_lru_add_batch(cli, &ext->oe_pages);
 874         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 875                                      oap_pending_item) {
 876                 list_del_init(&oap->oap_rpc_item);
 877                 list_del_init(&oap->oap_pending_item);
 878                 if (last_off <= oap->oap_obj_off) {
 879                         last_off = oap->oap_obj_off;
 880                         last_count = oap->oap_count;
 881                 }
 882
 883                 --ext->oe_nr_pages;
 884                 osc_ap_completion(env, cli, oap, sent, rc);
 885         }
 886         EASSERT(ext->oe_nr_pages == 0, ext);
 887
 888         if (!sent) {
 889                 lost_grant = ext->oe_grants;
 890         } else if (blocksize < PAGE_SIZE &&
 891                    last_count != PAGE_SIZE) {
 892                 /* For short writes we shouldn't count parts of pages that
 893                  * span a whole chunk on the OST side, or our accounting goes
 894                  * wrong.  Should match the code in filter_grant_check. */
 895                 int offset = last_off & ~PAGE_MASK;
 896                 int count = last_count + (offset & (blocksize - 1));
 897                 int end = (offset + last_count) & (blocksize - 1);
 898                 if (end)
 899                         count += blocksize - end;
 900
 901                 lost_grant = PAGE_SIZE - count;
 902         }
 903         if (ext->oe_grants > 0)
 904                 osc_free_grant(cli, nr_pages, lost_grant, ext->oe_grants);
 905
 906         osc_extent_remove(ext);
 907         /* put the refcount for RPC */
 908         osc_extent_put(env, ext);
 909         RETURN(0);
 910 }
 911
 912 static int extent_wait_cb(struct osc_extent *ext, enum osc_extent_state state)
 913 {
 914         int ret;
 915
 916         osc_object_lock(ext->oe_obj);
 917         ret = ext->oe_state == state;
 918         osc_object_unlock(ext->oe_obj);
 919
 920         return ret;
 921 }
 922
 923 /**
 924  * Wait for the extent's state to become @state.
 925  */
 926 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
 927                            enum osc_extent_state state)
 928 {
 929         struct osc_object *obj = ext->oe_obj;
 930         int rc = 0;
 931         ENTRY;
 932
 933         osc_object_lock(obj);
 934         LASSERT(sanity_check_nolock(ext) == 0);
 935         /* `Kick' this extent only if the caller is waiting for it to be
 936          * written out. */
 937         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
 938                 if (ext->oe_state == OES_ACTIVE) {
 939                         ext->oe_urgent = 1;
 940                 } else if (ext->oe_state == OES_CACHE) {
 941                         ext->oe_urgent = 1;
 942                         osc_extent_hold(ext);
 943                         rc = 1;
 944                 }
 945         }
 946         osc_object_unlock(obj);
 947         if (rc == 1)
 948                 osc_extent_release(env, ext);
 949
 950         /* wait for the extent until its state becomes @state */
 951         rc = wait_event_idle_timeout(ext->oe_waitq, extent_wait_cb(ext, state),
 952                                      cfs_time_seconds(600));
 953         if (rc == 0) {
 954                 OSC_EXTENT_DUMP(D_ERROR, ext,
 955                         "%s: wait ext to %u timedout, recovery in progress?\n",
 956                         cli_name(osc_cli(obj)), state);
 957
 958                 wait_event_idle(ext->oe_waitq, extent_wait_cb(ext, state));
 959         }
 960         if (ext->oe_rc < 0)
 961                 rc = ext->oe_rc;
 962         else
 963                 rc = 0;
 964         RETURN(rc);
 965 }
 966
 967 /**
 968  * Discard pages with index greater than @size. If @ext is overlapped with
 969  * @size, then partial truncate happens.
 970  */
 971 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
 972                                 bool partial)
 973 {
 974         struct lu_env         *env;
 975         struct cl_io          *io;
 976         struct osc_object     *obj = ext->oe_obj;
 977         struct client_obd     *cli = osc_cli(obj);
 978         struct osc_async_page *oap;
 979         struct osc_async_page *tmp;
 980         struct pagevec        *pvec;
 981         int                    pages_in_chunk = 0;
 982         int                    ppc_bits    = cli->cl_chunkbits -
 983                                              PAGE_SHIFT;
 984         __u64                  trunc_chunk = trunc_index >> ppc_bits;
 985         int                    grants   = 0;
 986         int                    nr_pages = 0;
 987         int                    rc       = 0;
 988         __u16                  refcheck;
 989         ENTRY;
 990
 991         LASSERT(sanity_check(ext) == 0);
 992         LASSERT(ext->oe_state == OES_TRUNC);
 993         LASSERT(!ext->oe_urgent);
 994
 995         /* Request new lu_env.
 996          * We can't use that env from osc_cache_truncate_start() because
 997          * it's from lov_io_sub and not fully initialized. */
 998         env = cl_env_get(&refcheck);
 999         if (IS_ERR(env))
1000                 RETURN(PTR_ERR(env));
1001
1002         io  = osc_env_thread_io(env);
1003         io->ci_obj = cl_object_top(osc2cl(obj));
1004         io->ci_ignore_layout = 1;
1005         pvec = &osc_env_info(env)->oti_pagevec;
1006         ll_pagevec_init(pvec, 0);
1007         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
1008         if (rc < 0)
1009                 GOTO(out, rc);
1010
1011         /* discard all pages with index greater than trunc_index */
1012         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
1013                                      oap_pending_item) {
1014                 pgoff_t index = osc_index(oap2osc(oap));
1015                 struct cl_page  *page = oap2cl_page(oap);
1016
1017                 LASSERT(list_empty(&oap->oap_rpc_item));
1018
1019                 /* only discard the pages with their index greater than
1020                  * trunc_index, and ... */
1021                 if (index < trunc_index ||
1022                     (index == trunc_index && partial)) {
1023                         /* accounting how many pages remaining in the chunk
1024                          * so that we can calculate grants correctly. */
1025                         if (index >> ppc_bits == trunc_chunk)
1026                                 ++pages_in_chunk;
1027                         continue;
1028                 }
1029
1030                 list_del_init(&oap->oap_pending_item);
1031
1032                 cl_page_get(page);
1033                 lu_ref_add(&page->cp_reference, "truncate", current);
1034
1035                 if (cl_page_own(env, io, page) == 0) {
1036                         cl_page_discard(env, io, page);
1037                         cl_page_disown(env, io, page);
1038                 } else {
1039                         LASSERT(page->cp_state == CPS_FREEING);
1040                         LASSERT(0);
1041                 }
1042
1043                 lu_ref_del(&page->cp_reference, "truncate", current);
1044                 cl_pagevec_put(env, page, pvec);
1045
1046                 --ext->oe_nr_pages;
1047                 ++nr_pages;
1048         }
1049         pagevec_release(pvec);
1050
1051         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1052                       ext->oe_nr_pages == 0),
1053                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1054
1055         osc_object_lock(obj);
1056         if (ext->oe_nr_pages == 0) {
1057                 LASSERT(pages_in_chunk == 0);
1058                 grants = ext->oe_grants;
1059                 ext->oe_grants = 0;
1060         } else { /* calculate how many grants we can free */
1061                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1062                 pgoff_t last_index;
1063
1064
1065                 /* if there is no pages in this chunk, we can also free grants
1066                  * for the last chunk */
1067                 if (pages_in_chunk == 0) {
1068                         /* if this is the 1st chunk and no pages in this chunk,
1069                          * ext->oe_nr_pages must be zero, so we should be in
1070                          * the other if-clause. */
1071                         LASSERT(trunc_chunk > 0);
1072                         --trunc_chunk;
1073                         ++chunks;
1074                 }
1075
1076                 /* this is what we can free from this extent */
1077                 grants          = chunks << cli->cl_chunkbits;
1078                 ext->oe_grants -= grants;
1079                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1080                 ext->oe_end     = min(last_index, ext->oe_max_end);
1081                 LASSERT(ext->oe_end >= ext->oe_start);
1082                 LASSERT(ext->oe_grants > 0);
1083         }
1084         osc_object_unlock(obj);
1085
1086         if (grants > 0 || nr_pages > 0)
1087                 osc_free_grant(cli, nr_pages, grants, grants);
1088
1089 out:
1090         cl_io_fini(env, io);
1091         cl_env_put(env, &refcheck);
1092         RETURN(rc);
1093 }
1094
1095 /**
1096  * This function is used to make the extent prepared for transfer.
1097  * A race with flusing page - ll_writepage() has to be handled cautiously.
1098  */
1099 static int osc_extent_make_ready(const struct lu_env *env,
1100                                  struct osc_extent *ext)
1101 {
1102         struct osc_async_page *oap;
1103         struct osc_async_page *last = NULL;
1104         struct osc_object *obj = ext->oe_obj;
1105         unsigned int page_count = 0;
1106         int rc;
1107         ENTRY;
1108
1109         /* we're going to grab page lock, so object lock must not be taken. */
1110         LASSERT(sanity_check(ext) == 0);
1111         /* in locking state, any process should not touch this extent. */
1112         EASSERT(ext->oe_state == OES_LOCKING, ext);
1113         EASSERT(ext->oe_owner != NULL, ext);
1114
1115         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1116
1117         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1118                 ++page_count;
1119                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1120                         last = oap;
1121
1122                 /* checking ASYNC_READY is race safe */
1123                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1124                         continue;
1125
1126                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1127                 switch (rc) {
1128                 case 0:
1129                         spin_lock(&oap->oap_lock);
1130                         oap->oap_async_flags |= ASYNC_READY;
1131                         spin_unlock(&oap->oap_lock);
1132                         break;
1133                 case -EALREADY:
1134                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1135                         break;
1136                 default:
1137                         LASSERTF(0, "unknown return code: %d\n", rc);
1138                 }
1139         }
1140
1141         LASSERT(page_count == ext->oe_nr_pages);
1142         LASSERT(last != NULL);
1143         /* the last page is the only one we need to refresh its count by
1144          * the size of file. */
1145         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1146                 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1147                 LASSERT(last_oap_count > 0);
1148                 LASSERT(last->oap_page_off + last_oap_count <= PAGE_SIZE);
1149                 last->oap_count = last_oap_count;
1150                 spin_lock(&last->oap_lock);
1151                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1152                 spin_unlock(&last->oap_lock);
1153         }
1154
1155         /* for the rest of pages, we don't need to call osf_refresh_count()
1156          * because it's known they are not the last page */
1157         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1158                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1159                         oap->oap_count = PAGE_SIZE - oap->oap_page_off;
1160                         spin_lock(&oap->oap_lock);
1161                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1162                         spin_unlock(&oap->oap_lock);
1163                 }
1164         }
1165
1166         osc_object_lock(obj);
1167         osc_extent_state_set(ext, OES_RPC);
1168         osc_object_unlock(obj);
1169         /* get a refcount for RPC. */
1170         osc_extent_get(ext);
1171
1172         RETURN(0);
1173 }
1174
1175 /**
1176  * Quick and simple version of osc_extent_find(). This function is frequently
1177  * called to expand the extent for the same IO. To expand the extent, the
1178  * page index must be in the same or next chunk of ext->oe_end.
1179  */
1180 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index,
1181                              unsigned int *grants)
1182 {
1183         struct osc_object *obj = ext->oe_obj;
1184         struct client_obd *cli = osc_cli(obj);
1185         struct osc_extent *next;
1186         int ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1187         pgoff_t chunk = index >> ppc_bits;
1188         pgoff_t end_chunk;
1189         pgoff_t end_index;
1190         unsigned int chunksize = 1 << cli->cl_chunkbits;
1191         int rc = 0;
1192         ENTRY;
1193
1194         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1195         osc_object_lock(obj);
1196         LASSERT(sanity_check_nolock(ext) == 0);
1197         end_chunk = ext->oe_end >> ppc_bits;
1198         if (chunk > end_chunk + 1)
1199                 GOTO(out, rc = -ERANGE);
1200
1201         if (end_chunk >= chunk)
1202                 GOTO(out, rc = 0);
1203
1204         LASSERT(end_chunk + 1 == chunk);
1205
1206         /* try to expand this extent to cover @index */
1207         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1208
1209         /* don't go over the maximum extent size reported by server */
1210         if (end_index - ext->oe_start + 1 > cli->cl_max_extent_pages)
1211                 GOTO(out, rc = -ERANGE);
1212
1213         next = next_extent(ext);
1214         if (next != NULL && next->oe_start <= end_index)
1215                 /* complex mode - overlapped with the next extent,
1216                  * this case will be handled by osc_extent_find() */
1217                 GOTO(out, rc = -EAGAIN);
1218
1219         ext->oe_end = end_index;
1220         ext->oe_grants += chunksize;
1221         LASSERT(*grants >= chunksize);
1222         *grants -= chunksize;
1223         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1224                  "overlapped after expanding for %lu.\n", index);
1225         EXIT;
1226
1227 out:
1228         osc_object_unlock(obj);
1229         RETURN(rc);
1230 }
1231
1232 static void osc_extent_tree_dump0(int mask, struct osc_object *obj,
1233                                   const char *func, int line)
1234 {
1235         struct osc_extent *ext;
1236         int cnt;
1237
1238         if (!cfs_cdebug_show(mask, DEBUG_SUBSYSTEM))
1239                 return;
1240
1241         CDEBUG(mask, "Dump object %p extents at %s:%d, mppr: %u.\n",
1242                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1243
1244         /* osc_object_lock(obj); */
1245         cnt = 1;
1246         for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1247                 OSC_EXTENT_DUMP(mask, ext, "in tree %d.\n", cnt++);
1248
1249         cnt = 1;
1250         list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1251                 OSC_EXTENT_DUMP(mask, ext, "hp %d.\n", cnt++);
1252
1253         cnt = 1;
1254         list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1255                 OSC_EXTENT_DUMP(mask, ext, "urgent %d.\n", cnt++);
1256
1257         cnt = 1;
1258         list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1259                 OSC_EXTENT_DUMP(mask, ext, "reading %d.\n", cnt++);
1260         /* osc_object_unlock(obj); */
1261 }
1262
1263 /* ------------------ osc extent end ------------------ */
1264
1265 static inline int osc_is_ready(struct osc_object *osc)
1266 {
1267         return !list_empty(&osc->oo_ready_item) ||
1268                !list_empty(&osc->oo_hp_ready_item);
1269 }
1270
1271 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1272         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1273                (OSC), osc_is_ready(OSC),                                       \
1274                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1275                list_empty_marker(&(OSC)->oo_ready_item),                       \
1276                atomic_read(&(OSC)->oo_nr_writes),                              \
1277                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1278                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1279                atomic_read(&(OSC)->oo_nr_reads),                               \
1280                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1281                ##args)
1282
1283 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1284                           int cmd)
1285 {
1286         struct osc_page *opg  = oap2osc_page(oap);
1287         struct cl_page  *page = oap2cl_page(oap);
1288         int result;
1289
1290         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1291
1292         ENTRY;
1293         result = cl_page_make_ready(env, page, CRT_WRITE);
1294         if (result == 0)
1295                 opg->ops_submit_time = ktime_get();
1296         RETURN(result);
1297 }
1298
1299 static int osc_refresh_count(const struct lu_env *env,
1300                              struct osc_async_page *oap, int cmd)
1301 {
1302         struct osc_page  *opg = oap2osc_page(oap);
1303         pgoff_t index = osc_index(oap2osc(oap));
1304         struct cl_object *obj;
1305         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
1306         int result;
1307         loff_t kms;
1308
1309         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1310         LASSERT(!(cmd & OBD_BRW_READ));
1311         LASSERT(opg != NULL);
1312         obj = opg->ops_cl.cpl_obj;
1313
1314         cl_object_attr_lock(obj);
1315         result = cl_object_attr_get(env, obj, attr);
1316         cl_object_attr_unlock(obj);
1317         if (result < 0)
1318                 return result;
1319         kms = attr->cat_kms;
1320         if (cl_offset(obj, index) >= kms)
1321                 /* catch race with truncate */
1322                 return 0;
1323         else if (cl_offset(obj, index + 1) > kms)
1324                 /* catch sub-page write at end of file */
1325                 return kms & ~PAGE_MASK;
1326         else
1327                 return PAGE_SIZE;
1328 }
1329
1330 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1331                           int cmd, int rc)
1332 {
1333         struct osc_page   *opg  = oap2osc_page(oap);
1334         struct cl_page    *page = oap2cl_page(oap);
1335         enum cl_req_type   crt;
1336         int srvlock;
1337
1338         ENTRY;
1339
1340         cmd &= ~OBD_BRW_NOQUOTA;
1341         LASSERTF(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ),
1342                  "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1343         LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1344                 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1345         LASSERT(opg->ops_transfer_pinned);
1346
1347         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1348         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1349         opg->ops_transfer_pinned = 0;
1350
1351         opg->ops_submit_time = ktime_set(0, 0);
1352         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1353
1354         /* statistic */
1355         if (rc == 0 && srvlock) {
1356                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1357                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1358                 size_t bytes = oap->oap_count;
1359
1360                 if (crt == CRT_READ)
1361                         stats->os_lockless_reads += bytes;
1362                 else
1363                         stats->os_lockless_writes += bytes;
1364         }
1365
1366         /*
1367          * This has to be the last operation with the page, as locks are
1368          * released in cl_page_completion() and nothing except for the
1369          * reference counter protects page from concurrent reclaim.
1370          */
1371         lu_ref_del(&page->cp_reference, "transfer", page);
1372
1373         cl_page_completion(env, page, crt, rc);
1374         cl_page_put(env, page);
1375
1376         RETURN(0);
1377 }
1378
1379 #define OSC_DUMP_GRANT(mask, cli, fmt, args...) do {                    \
1380         struct client_obd *__tmp = (cli);                               \
1381         CDEBUG(mask, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu " \
1382                "dropped: %ld avail: %ld, dirty_grant: %ld, "            \
1383                "reserved: %ld, flight: %d } lru {in list: %ld, "        \
1384                "left: %ld, waiters: %d }" fmt "\n",                     \
1385                cli_name(__tmp),                                         \
1386                __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages,        \
1387                atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1388                __tmp->cl_lost_grant, __tmp->cl_avail_grant,             \
1389                __tmp->cl_dirty_grant,                                   \
1390                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,         \
1391                atomic_long_read(&__tmp->cl_lru_in_list),                \
1392                atomic_long_read(&__tmp->cl_lru_busy),                   \
1393                atomic_read(&__tmp->cl_lru_shrinkers), ##args);          \
1394 } while (0)
1395
1396 /* caller must hold loi_list_lock */
1397 static void osc_consume_write_grant(struct client_obd *cli,
1398                                     struct brw_page *pga)
1399 {
1400         assert_spin_locked(&cli->cl_loi_list_lock);
1401         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1402         cli->cl_dirty_pages++;
1403         pga->flag |= OBD_BRW_FROM_GRANT;
1404         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1405                PAGE_SIZE, pga, pga->pg);
1406         osc_update_next_shrink(cli);
1407 }
1408
1409 /* the companion to osc_consume_write_grant, called when a brw has completed.
1410  * must be called with the loi lock held. */
1411 static void osc_release_write_grant(struct client_obd *cli,
1412                                     struct brw_page *pga)
1413 {
1414         ENTRY;
1415
1416         assert_spin_locked(&cli->cl_loi_list_lock);
1417         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1418                 EXIT;
1419                 return;
1420         }
1421
1422         pga->flag &= ~OBD_BRW_FROM_GRANT;
1423         atomic_long_dec(&obd_dirty_pages);
1424         cli->cl_dirty_pages--;
1425         EXIT;
1426 }
1427
1428 /**
1429  * To avoid sleeping with object lock held, it's good for us allocate enough
1430  * grants before entering into critical section.
1431  *
1432  * client_obd_list_lock held by caller
1433  */
1434 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1435 {
1436         int rc = -EDQUOT;
1437
1438         if (cli->cl_avail_grant >= bytes) {
1439                 cli->cl_avail_grant    -= bytes;
1440                 cli->cl_reserved_grant += bytes;
1441                 rc = 0;
1442         }
1443         return rc;
1444 }
1445
1446 static void __osc_unreserve_grant(struct client_obd *cli,
1447                                   unsigned int reserved, unsigned int unused)
1448 {
1449         /* it's quite normal for us to get more grant than reserved.
1450          * Thinking about a case that two extents merged by adding a new
1451          * chunk, we can save one extent tax. If extent tax is greater than
1452          * one chunk, we can save more grant by adding a new chunk */
1453         cli->cl_reserved_grant -= reserved;
1454         if (unused > reserved) {
1455                 cli->cl_avail_grant += reserved;
1456                 cli->cl_lost_grant  += unused - reserved;
1457                 cli->cl_dirty_grant -= unused - reserved;
1458         } else {
1459                 cli->cl_avail_grant += unused;
1460                 cli->cl_dirty_grant += reserved - unused;
1461         }
1462 }
1463
1464 static void osc_unreserve_grant_nolock(struct client_obd *cli,
1465                                        unsigned int reserved,
1466                                        unsigned int unused)
1467 {
1468         __osc_unreserve_grant(cli, reserved, unused);
1469         if (unused > 0)
1470                 osc_wake_cache_waiters(cli);
1471 }
1472
1473 static void osc_unreserve_grant(struct client_obd *cli,
1474                                 unsigned int reserved, unsigned int unused)
1475 {
1476         spin_lock(&cli->cl_loi_list_lock);
1477         osc_unreserve_grant_nolock(cli, reserved, unused);
1478         spin_unlock(&cli->cl_loi_list_lock);
1479 }
1480
1481 /**
1482  * Free grant after IO is finished or canceled.
1483  *
1484  * @lost_grant is used to remember how many grants we have allocated but not
1485  * used, we should return these grants to OST. There're two cases where grants
1486  * can be lost:
1487  * 1. truncate;
1488  * 2. blocksize at OST is less than PAGE_SIZE and a partial page was
1489  *    written. In this case OST may use less chunks to serve this partial
1490  *    write. OSTs don't actually know the page size on the client side. so
1491  *    clients have to calculate lost grant by the blocksize on the OST.
1492  *    See filter_grant_check() for details.
1493  */
1494 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1495                            unsigned int lost_grant, unsigned int dirty_grant)
1496 {
1497         unsigned long grant;
1498
1499         grant = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
1500
1501         spin_lock(&cli->cl_loi_list_lock);
1502         atomic_long_sub(nr_pages, &obd_dirty_pages);
1503         cli->cl_dirty_pages -= nr_pages;
1504         cli->cl_lost_grant += lost_grant;
1505         cli->cl_dirty_grant -= dirty_grant;
1506         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1507                 /* borrow some grant from truncate to avoid the case that
1508                  * truncate uses up all avail grant */
1509                 cli->cl_lost_grant -= grant;
1510                 cli->cl_avail_grant += grant;
1511         }
1512         osc_wake_cache_waiters(cli);
1513         spin_unlock(&cli->cl_loi_list_lock);
1514         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu/%lu\n",
1515                lost_grant, cli->cl_lost_grant,
1516                cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_SHIFT,
1517                cli->cl_dirty_grant);
1518 }
1519
1520 /**
1521  * The companion to osc_enter_cache(), called when @oap is no longer part of
1522  * the dirty accounting due to error.
1523  */
1524 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1525 {
1526         spin_lock(&cli->cl_loi_list_lock);
1527         osc_release_write_grant(cli, &oap->oap_brw_page);
1528         spin_unlock(&cli->cl_loi_list_lock);
1529 }
1530
1531 /**
1532  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1533  * is available.
1534  */
1535 static int osc_enter_cache_try(struct client_obd *cli,
1536                                struct osc_async_page *oap,
1537                                int bytes)
1538 {
1539         int rc;
1540
1541         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1542
1543         rc = osc_reserve_grant(cli, bytes);
1544         if (rc < 0)
1545                 return 0;
1546
1547         if (cli->cl_dirty_pages < cli->cl_dirty_max_pages) {
1548                 if (atomic_long_add_return(1, &obd_dirty_pages) <=
1549                     obd_max_dirty_pages) {
1550                         osc_consume_write_grant(cli, &oap->oap_brw_page);
1551                         rc = 1;
1552                         goto out;
1553                 } else
1554                         atomic_long_dec(&obd_dirty_pages);
1555         }
1556         __osc_unreserve_grant(cli, bytes, bytes);
1557
1558 out:
1559         return rc;
1560 }
1561
1562 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1563 {
1564         int rc;
1565         spin_lock(&cli->cl_loi_list_lock);
1566         rc = list_empty(&ocw->ocw_entry);
1567         spin_unlock(&cli->cl_loi_list_lock);
1568         return rc;
1569 }
1570
1571 /**
1572  * The main entry to reserve dirty page accounting. Usually the grant reserved
1573  * in this function will be freed in bulk in osc_free_grant() unless it fails
1574  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1575  *
1576  * The process will be put into sleep if it's already run out of grant.
1577  */
1578 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1579                            struct osc_async_page *oap, int bytes)
1580 {
1581         struct osc_object       *osc = oap->oap_obj;
1582         struct lov_oinfo        *loi = osc->oo_oinfo;
1583         struct osc_cache_waiter  ocw;
1584         int                      rc = -EDQUOT;
1585         ENTRY;
1586
1587         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1588
1589         spin_lock(&cli->cl_loi_list_lock);
1590
1591         /* force the caller to try sync io.  this can jump the list
1592          * of queued writes and create a discontiguous rpc stream */
1593         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1594             cli->cl_dirty_max_pages == 0 ||
1595             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) {
1596                 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1597                 GOTO(out, rc = -EDQUOT);
1598         }
1599
1600         /* Hopefully normal case - cache space and write credits available */
1601         if (list_empty(&cli->cl_cache_waiters) &&
1602             osc_enter_cache_try(cli, oap, bytes)) {
1603                 OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1604                 GOTO(out, rc = 0);
1605         }
1606
1607         /* We can get here for two reasons: too many dirty pages in cache, or
1608          * run out of grants. In both cases we should write dirty pages out.
1609          * Adding a cache waiter will trigger urgent write-out no matter what
1610          * RPC size will be.
1611          * The exiting condition is no avail grants and no dirty pages caching,
1612          * that really means there is no space on the OST. */
1613         init_waitqueue_head(&ocw.ocw_waitq);
1614         ocw.ocw_oap   = oap;
1615         ocw.ocw_grant = bytes;
1616         while (cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0) {
1617                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1618                 ocw.ocw_rc = 0;
1619                 spin_unlock(&cli->cl_loi_list_lock);
1620
1621                 osc_io_unplug_async(env, cli, NULL);
1622
1623                 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1624                        cli_name(cli), &ocw, oap);
1625
1626                 rc = wait_event_idle_timeout(ocw.ocw_waitq,
1627                                              ocw_granted(cli, &ocw),
1628                                              cfs_time_seconds(AT_OFF ?
1629                                                               obd_timeout :
1630                                                               at_max));
1631
1632                 spin_lock(&cli->cl_loi_list_lock);
1633
1634                 if (rc <= 0) {
1635                         /* l_wait_event is interrupted by signal or timed out */
1636                         list_del_init(&ocw.ocw_entry);
1637                         if (rc == 0)
1638                                 rc = -ETIMEDOUT;
1639                         break;
1640                 }
1641                 LASSERT(list_empty(&ocw.ocw_entry));
1642                 rc = ocw.ocw_rc;
1643
1644                 if (rc != -EDQUOT)
1645                         break;
1646                 if (osc_enter_cache_try(cli, oap, bytes)) {
1647                         rc = 0;
1648                         break;
1649                 }
1650         }
1651
1652         switch (rc) {
1653         case 0:
1654                 OSC_DUMP_GRANT(D_CACHE, cli, "finally got grant space\n");
1655                 break;
1656         case -ETIMEDOUT:
1657                 OSC_DUMP_GRANT(D_CACHE, cli,
1658                                "timeout, fall back to sync i/o\n");
1659                 osc_extent_tree_dump(D_CACHE, osc);
1660                 /* fall back to synchronous I/O */
1661                 rc = -EDQUOT;
1662                 break;
1663         case -EINTR:
1664                 /* Ensures restartability - LU-3581 */
1665                 OSC_DUMP_GRANT(D_CACHE, cli, "interrupted\n");
1666                 rc = -ERESTARTSYS;
1667                 break;
1668         case -EDQUOT:
1669                 OSC_DUMP_GRANT(D_CACHE, cli,
1670                                "no grant space, fall back to sync i/o\n");
1671                 break;
1672         default:
1673                 CDEBUG(D_CACHE, "%s: event for cache space @ %p never arrived "
1674                        "due to %d, fall back to sync i/o\n",
1675                        cli_name(cli), &ocw, rc);
1676                 break;
1677         }
1678         EXIT;
1679 out:
1680         spin_unlock(&cli->cl_loi_list_lock);
1681         RETURN(rc);
1682 }
1683
1684 /* caller must hold loi_list_lock */
1685 void osc_wake_cache_waiters(struct client_obd *cli)
1686 {
1687         struct list_head *l, *tmp;
1688         struct osc_cache_waiter *ocw;
1689
1690         ENTRY;
1691         list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1692                 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
1693
1694                 ocw->ocw_rc = -EDQUOT;
1695
1696                 if (osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant))
1697                         ocw->ocw_rc = 0;
1698
1699                 if (ocw->ocw_rc == 0 ||
1700                     !(cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0)) {
1701                         list_del_init(&ocw->ocw_entry);
1702                         CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant "
1703                                "%ld, %d\n", ocw, ocw->ocw_oap,
1704                                cli->cl_avail_grant, ocw->ocw_rc);
1705
1706                         wake_up(&ocw->ocw_waitq);
1707                 }
1708         }
1709
1710         EXIT;
1711 }
1712 EXPORT_SYMBOL(osc_wake_cache_waiters);
1713
1714 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1715 {
1716         int hprpc = !!list_empty(&osc->oo_hp_exts);
1717         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1718 }
1719
1720 /* This maintains the lists of pending pages to read/write for a given object
1721  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1722  * to quickly find objects that are ready to send an RPC. */
1723 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1724                          int cmd)
1725 {
1726         int invalid_import = 0;
1727         ENTRY;
1728
1729         /* if we have an invalid import we want to drain the queued pages
1730          * by forcing them through rpcs that immediately fail and complete
1731          * the pages.  recovery relies on this to empty the queued pages
1732          * before canceling the locks and evicting down the llite pages */
1733         if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1734                 invalid_import = 1;
1735
1736         if (cmd & OBD_BRW_WRITE) {
1737                 if (atomic_read(&osc->oo_nr_writes) == 0)
1738                         RETURN(0);
1739                 if (invalid_import) {
1740                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1741                         RETURN(1);
1742                 }
1743                 if (!list_empty(&osc->oo_hp_exts)) {
1744                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1745                         RETURN(1);
1746                 }
1747                 if (!list_empty(&osc->oo_urgent_exts)) {
1748                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1749                         RETURN(1);
1750                 }
1751                 /* trigger a write rpc stream as long as there are dirtiers
1752                  * waiting for space.  as they're waiting, they're not going to
1753                  * create more pages to coalesce with what's waiting.. */
1754                 if (!list_empty(&cli->cl_cache_waiters)) {
1755                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1756                         RETURN(1);
1757                 }
1758                 if (!list_empty(&osc->oo_full_exts)) {
1759                         CDEBUG(D_CACHE, "full extent ready, make an RPC\n");
1760                         RETURN(1);
1761                 }
1762         } else {
1763                 if (atomic_read(&osc->oo_nr_reads) == 0)
1764                         RETURN(0);
1765                 if (invalid_import) {
1766                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1767                         RETURN(1);
1768                 }
1769                 /* all read are urgent. */
1770                 if (!list_empty(&osc->oo_reading_exts))
1771                         RETURN(1);
1772         }
1773
1774         RETURN(0);
1775 }
1776
1777 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1778 {
1779         struct client_obd *cli = osc_cli(obj);
1780         if (cmd & OBD_BRW_WRITE) {
1781                 atomic_add(delta, &obj->oo_nr_writes);
1782                 atomic_add(delta, &cli->cl_pending_w_pages);
1783                 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1784         } else {
1785                 atomic_add(delta, &obj->oo_nr_reads);
1786                 atomic_add(delta, &cli->cl_pending_r_pages);
1787                 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1788         }
1789         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1790 }
1791
1792 static int osc_makes_hprpc(struct osc_object *obj)
1793 {
1794         return !list_empty(&obj->oo_hp_exts);
1795 }
1796
1797 static void on_list(struct list_head *item, struct list_head *list,
1798                     int should_be_on)
1799 {
1800         if (list_empty(item) && should_be_on)
1801                 list_add_tail(item, list);
1802         else if (!list_empty(item) && !should_be_on)
1803                 list_del_init(item);
1804 }
1805
1806 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1807  * can find pages to build into rpcs quickly */
1808 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1809 {
1810         if (osc_makes_hprpc(osc)) {
1811                 /* HP rpc */
1812                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1813                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1814         } else {
1815                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1816                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1817                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1818                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1819         }
1820
1821         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1822                 atomic_read(&osc->oo_nr_writes) > 0);
1823
1824         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1825                 atomic_read(&osc->oo_nr_reads) > 0);
1826
1827         return osc_is_ready(osc);
1828 }
1829
1830 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1831 {
1832         int is_ready;
1833
1834         spin_lock(&cli->cl_loi_list_lock);
1835         is_ready = __osc_list_maint(cli, osc);
1836         spin_unlock(&cli->cl_loi_list_lock);
1837
1838         return is_ready;
1839 }
1840
1841 /* this is trying to propogate async writeback errors back up to the
1842  * application.  As an async write fails we record the error code for later if
1843  * the app does an fsync.  As long as errors persist we force future rpcs to be
1844  * sync so that the app can get a sync error and break the cycle of queueing
1845  * pages for which writeback will fail. */
1846 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1847                            int rc)
1848 {
1849         if (rc) {
1850                 if (!ar->ar_rc)
1851                         ar->ar_rc = rc;
1852
1853                 ar->ar_force_sync = 1;
1854                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1855                 return;
1856
1857         }
1858
1859         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1860                 ar->ar_force_sync = 0;
1861 }
1862
1863 /* this must be called holding the loi list lock to give coverage to exit_cache,
1864  * async_flag maintenance, and oap_request */
1865 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1866                               struct osc_async_page *oap, int sent, int rc)
1867 {
1868         struct osc_object *osc = oap->oap_obj;
1869         struct lov_oinfo  *loi = osc->oo_oinfo;
1870         __u64 xid = 0;
1871
1872         ENTRY;
1873         if (oap->oap_request != NULL) {
1874                 xid = ptlrpc_req_xid(oap->oap_request);
1875                 ptlrpc_req_finished(oap->oap_request);
1876                 oap->oap_request = NULL;
1877         }
1878
1879         /* As the transfer for this page is being done, clear the flags */
1880         spin_lock(&oap->oap_lock);
1881         oap->oap_async_flags = 0;
1882         spin_unlock(&oap->oap_lock);
1883
1884         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1885                 spin_lock(&cli->cl_loi_list_lock);
1886                 osc_process_ar(&cli->cl_ar, xid, rc);
1887                 osc_process_ar(&loi->loi_ar, xid, rc);
1888                 spin_unlock(&cli->cl_loi_list_lock);
1889         }
1890
1891         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1892         if (rc)
1893                 CERROR("completion on oap %p obj %p returns %d.\n",
1894                        oap, osc, rc);
1895
1896         EXIT;
1897 }
1898
1899 struct extent_rpc_data {
1900         struct list_head        *erd_rpc_list;
1901         unsigned int            erd_page_count;
1902         unsigned int            erd_max_pages;
1903         unsigned int            erd_max_chunks;
1904         unsigned int            erd_max_extents;
1905 };
1906
1907 static inline unsigned osc_extent_chunks(const struct osc_extent *ext)
1908 {
1909         struct client_obd *cli = osc_cli(ext->oe_obj);
1910         unsigned ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1911
1912         return (ext->oe_end >> ppc_bits) - (ext->oe_start >> ppc_bits) + 1;
1913 }
1914
1915 static inline bool
1916 can_merge(const struct osc_extent *ext, const struct osc_extent *in_rpc)
1917 {
1918         if (ext->oe_no_merge || in_rpc->oe_no_merge)
1919                 return false;
1920
1921         if (ext->oe_srvlock != in_rpc->oe_srvlock)
1922                 return false;
1923
1924         if (ext->oe_ndelay != in_rpc->oe_ndelay)
1925                 return false;
1926
1927         if (!ext->oe_grants != !in_rpc->oe_grants)
1928                 return false;
1929
1930         if (ext->oe_dio != in_rpc->oe_dio)
1931                 return false;
1932
1933         /* It's possible to have overlap on DIO */
1934         if (in_rpc->oe_dio && overlapped(ext, in_rpc))
1935                 return false;
1936
1937         return true;
1938 }
1939
1940 /**
1941  * Try to add extent to one RPC. We need to think about the following things:
1942  * - # of pages must not be over max_pages_per_rpc
1943  * - extent must be compatible with previous ones
1944  */
1945 static int try_to_add_extent_for_io(struct client_obd *cli,
1946                                     struct osc_extent *ext,
1947                                     struct extent_rpc_data *data)
1948 {
1949         struct osc_extent *tmp;
1950         unsigned int chunk_count;
1951         ENTRY;
1952
1953         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1954                 ext);
1955         OSC_EXTENT_DUMP(D_CACHE, ext, "trying to add this extent\n");
1956
1957         if (data->erd_max_extents == 0)
1958                 RETURN(0);
1959
1960         chunk_count = osc_extent_chunks(ext);
1961         EASSERTF(data->erd_page_count != 0 ||
1962                  chunk_count <= data->erd_max_chunks, ext,
1963                  "The first extent to be fit in a RPC contains %u chunks, "
1964                  "which is over the limit %u.\n", chunk_count,
1965                  data->erd_max_chunks);
1966         if (chunk_count > data->erd_max_chunks)
1967                 RETURN(0);
1968
1969         data->erd_max_pages = max(ext->oe_mppr, data->erd_max_pages);
1970         EASSERTF(data->erd_page_count != 0 ||
1971                 ext->oe_nr_pages <= data->erd_max_pages, ext,
1972                 "The first extent to be fit in a RPC contains %u pages, "
1973                 "which is over the limit %u.\n", ext->oe_nr_pages,
1974                 data->erd_max_pages);
1975         if (data->erd_page_count + ext->oe_nr_pages > data->erd_max_pages)
1976                 RETURN(0);
1977
1978         list_for_each_entry(tmp, data->erd_rpc_list, oe_link) {
1979                 EASSERT(tmp->oe_owner == current, tmp);
1980
1981                 if (!can_merge(ext, tmp))
1982                         RETURN(0);
1983         }
1984
1985         data->erd_max_extents--;
1986         data->erd_max_chunks -= chunk_count;
1987         data->erd_page_count += ext->oe_nr_pages;
1988         list_move_tail(&ext->oe_link, data->erd_rpc_list);
1989         ext->oe_owner = current;
1990         RETURN(1);
1991 }
1992
1993 /**
1994  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1995  * get_write_extent() takes all appropriate extents in atomic.
1996  *
1997  * The following policy is used to collect extents for IO:
1998  * 1. Add as many HP extents as possible;
1999  * 2. Add the first urgent extent in urgent extent list and take it out of
2000  *    urgent list;
2001  * 3. Add subsequent extents of this urgent extent;
2002  * 4. If urgent list is not empty, goto 2;
2003  * 5. Traverse the extent tree from the 1st extent;
2004  * 6. Above steps exit if there is no space in this RPC.
2005  */
2006 static unsigned int get_write_extents(struct osc_object *obj,
2007                                       struct list_head *rpclist)
2008 {
2009         struct client_obd *cli = osc_cli(obj);
2010         struct osc_extent *ext;
2011         struct extent_rpc_data data = {
2012                 .erd_rpc_list   = rpclist,
2013                 .erd_page_count = 0,
2014                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
2015                 .erd_max_chunks = osc_max_write_chunks(cli),
2016                 .erd_max_extents = 256,
2017         };
2018
2019         assert_osc_object_is_locked(obj);
2020         while (!list_empty(&obj->oo_hp_exts)) {
2021                 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
2022                                  oe_link);
2023                 LASSERT(ext->oe_state == OES_CACHE);
2024                 if (!try_to_add_extent_for_io(cli, ext, &data))
2025                         return data.erd_page_count;
2026                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2027         }
2028         if (data.erd_page_count == data.erd_max_pages)
2029                 return data.erd_page_count;
2030
2031         while (!list_empty(&obj->oo_urgent_exts)) {
2032                 ext = list_entry(obj->oo_urgent_exts.next,
2033                                  struct osc_extent, oe_link);
2034                 if (!try_to_add_extent_for_io(cli, ext, &data))
2035                         return data.erd_page_count;
2036         }
2037         if (data.erd_page_count == data.erd_max_pages)
2038                 return data.erd_page_count;
2039
2040         /* One key difference between full extents and other extents: full
2041          * extents can usually only be added if the rpclist was empty, so if we
2042          * can't add one, we continue on to trying to add normal extents.  This
2043          * is so we don't miss adding extra extents to an RPC containing high
2044          * priority or urgent extents. */
2045         while (!list_empty(&obj->oo_full_exts)) {
2046                 ext = list_entry(obj->oo_full_exts.next,
2047                                  struct osc_extent, oe_link);
2048                 if (!try_to_add_extent_for_io(cli, ext, &data))
2049                         break;
2050         }
2051         if (data.erd_page_count == data.erd_max_pages)
2052                 return data.erd_page_count;
2053
2054         ext = first_extent(obj);
2055         while (ext != NULL) {
2056                 if ((ext->oe_state != OES_CACHE) ||
2057                     /* this extent may be already in current rpclist */
2058                     (!list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
2059                         ext = next_extent(ext);
2060                         continue;
2061                 }
2062
2063                 if (!try_to_add_extent_for_io(cli, ext, &data))
2064                         return data.erd_page_count;
2065
2066                 ext = next_extent(ext);
2067         }
2068         return data.erd_page_count;
2069 }
2070
2071 static int
2072 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
2073                    struct osc_object *osc)
2074 __must_hold(osc)
2075 {
2076         LIST_HEAD(rpclist);
2077         struct osc_extent *ext;
2078         struct osc_extent *tmp;
2079         struct osc_extent *first = NULL;
2080         unsigned int page_count = 0;
2081         int srvlock = 0;
2082         int rc = 0;
2083         ENTRY;
2084
2085         assert_osc_object_is_locked(osc);
2086
2087         page_count = get_write_extents(osc, &rpclist);
2088         LASSERT(equi(page_count == 0, list_empty(&rpclist)));
2089
2090         if (list_empty(&rpclist))
2091                 RETURN(0);
2092
2093         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2094
2095         list_for_each_entry(ext, &rpclist, oe_link) {
2096                 LASSERT(ext->oe_state == OES_CACHE ||
2097                         ext->oe_state == OES_LOCK_DONE);
2098                 if (ext->oe_state == OES_CACHE)
2099                         osc_extent_state_set(ext, OES_LOCKING);
2100                 else
2101                         osc_extent_state_set(ext, OES_RPC);
2102         }
2103
2104         /* we're going to grab page lock, so release object lock because
2105          * lock order is page lock -> object lock. */
2106         osc_object_unlock(osc);
2107
2108         list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2109                 if (ext->oe_state == OES_LOCKING) {
2110                         rc = osc_extent_make_ready(env, ext);
2111                         if (unlikely(rc < 0)) {
2112                                 list_del_init(&ext->oe_link);
2113                                 osc_extent_finish(env, ext, 0, rc);
2114                                 continue;
2115                         }
2116                 }
2117                 if (first == NULL) {
2118                         first = ext;
2119                         srvlock = ext->oe_srvlock;
2120                 } else {
2121                         LASSERT(srvlock == ext->oe_srvlock);
2122                 }
2123         }
2124
2125         if (!list_empty(&rpclist)) {
2126                 LASSERT(page_count > 0);
2127                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2128                 LASSERT(list_empty(&rpclist));
2129         }
2130
2131         osc_object_lock(osc);
2132         RETURN(rc);
2133 }
2134
2135 /**
2136  * prepare pages for ASYNC io and put pages in send queue.
2137  *
2138  * \param cmd OBD_BRW_* macroses
2139  * \param lop pending pages
2140  *
2141  * \return zero if no page added to send queue.
2142  * \return 1 if pages successfully added to send queue.
2143  * \return negative on errors.
2144  */
2145 static int
2146 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2147                   struct osc_object *osc)
2148 __must_hold(osc)
2149 {
2150         struct osc_extent *ext;
2151         struct osc_extent *next;
2152         LIST_HEAD(rpclist);
2153         struct extent_rpc_data data = {
2154                 .erd_rpc_list   = &rpclist,
2155                 .erd_page_count = 0,
2156                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
2157                 .erd_max_chunks = UINT_MAX,
2158                 .erd_max_extents = UINT_MAX,
2159         };
2160         int rc = 0;
2161         ENTRY;
2162
2163         assert_osc_object_is_locked(osc);
2164         list_for_each_entry_safe(ext, next, &osc->oo_reading_exts, oe_link) {
2165                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2166                 if (!try_to_add_extent_for_io(cli, ext, &data))
2167                         break;
2168                 osc_extent_state_set(ext, OES_RPC);
2169                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2170         }
2171         LASSERT(data.erd_page_count <= data.erd_max_pages);
2172
2173         osc_update_pending(osc, OBD_BRW_READ, -data.erd_page_count);
2174
2175         if (!list_empty(&rpclist)) {
2176                 osc_object_unlock(osc);
2177
2178                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2179                 LASSERT(list_empty(&rpclist));
2180
2181                 osc_object_lock(osc);
2182         }
2183         RETURN(rc);
2184 }
2185
2186 #define list_to_obj(list, item) ({                                            \
2187         struct list_head *__tmp = (list)->next;                               \
2188         list_del_init(__tmp);                                         \
2189         list_entry(__tmp, struct osc_object, oo_##item);                      \
2190 })
2191
2192 /* This is called by osc_check_rpcs() to find which objects have pages that
2193  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2194 static struct osc_object *osc_next_obj(struct client_obd *cli)
2195 {
2196         ENTRY;
2197
2198         /* First return objects that have blocked locks so that they
2199          * will be flushed quickly and other clients can get the lock,
2200          * then objects which have pages ready to be stuffed into RPCs */
2201         if (!list_empty(&cli->cl_loi_hp_ready_list))
2202                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2203         if (!list_empty(&cli->cl_loi_ready_list))
2204                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2205
2206         /* then if we have cache waiters, return all objects with queued
2207          * writes.  This is especially important when many small files
2208          * have filled up the cache and not been fired into rpcs because
2209          * they don't pass the nr_pending/object threshhold */
2210         if (!list_empty(&cli->cl_cache_waiters) &&
2211             !list_empty(&cli->cl_loi_write_list))
2212                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2213
2214         /* then return all queued objects when we have an invalid import
2215          * so that they get flushed */
2216         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2217                 if (!list_empty(&cli->cl_loi_write_list))
2218                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2219                                            write_item));
2220                 if (!list_empty(&cli->cl_loi_read_list))
2221                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2222                                            read_item));
2223         }
2224         RETURN(NULL);
2225 }
2226
2227 /* called with the loi list lock held */
2228 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2229 __must_hold(&cli->cl_loi_list_lock)
2230 {
2231         struct osc_object *osc;
2232         int rc = 0;
2233         ENTRY;
2234
2235         while ((osc = osc_next_obj(cli)) != NULL) {
2236                 struct cl_object *obj = osc2cl(osc);
2237                 struct lu_ref_link link;
2238
2239                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2240
2241                 if (osc_max_rpc_in_flight(cli, osc)) {
2242                         __osc_list_maint(cli, osc);
2243                         break;
2244                 }
2245
2246                 cl_object_get(obj);
2247                 spin_unlock(&cli->cl_loi_list_lock);
2248                 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2249
2250                 /* attempt some read/write balancing by alternating between
2251                  * reads and writes in an object.  The makes_rpc checks here
2252                  * would be redundant if we were getting read/write work items
2253                  * instead of objects.  we don't want send_oap_rpc to drain a
2254                  * partial read pending queue when we're given this object to
2255                  * do io on writes while there are cache waiters */
2256                 osc_object_lock(osc);
2257                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2258                         rc = osc_send_write_rpc(env, cli, osc);
2259                         if (rc < 0) {
2260                                 CERROR("Write request failed with %d\n", rc);
2261
2262                                 /* osc_send_write_rpc failed, mostly because of
2263                                  * memory pressure.
2264                                  *
2265                                  * It can't break here, because if:
2266                                  *  - a page was submitted by osc_io_submit, so
2267                                  *    page locked;
2268                                  *  - no request in flight
2269                                  *  - no subsequent request
2270                                  * The system will be in live-lock state,
2271                                  * because there is no chance to call
2272                                  * osc_io_unplug() and osc_check_rpcs() any
2273                                  * more. pdflush can't help in this case,
2274                                  * because it might be blocked at grabbing
2275                                  * the page lock as we mentioned.
2276                                  *
2277                                  * Anyway, continue to drain pages. */
2278                                 /* break; */
2279                         }
2280                 }
2281                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2282                         rc = osc_send_read_rpc(env, cli, osc);
2283                         if (rc < 0)
2284                                 CERROR("Read request failed with %d\n", rc);
2285                 }
2286                 osc_object_unlock(osc);
2287
2288                 osc_list_maint(cli, osc);
2289                 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2290                 cl_object_put(env, obj);
2291
2292                 spin_lock(&cli->cl_loi_list_lock);
2293         }
2294 }
2295
2296 int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2297                    struct osc_object *osc, int async)
2298 {
2299         int rc = 0;
2300
2301         if (osc != NULL && osc_list_maint(cli, osc) == 0)
2302                 return 0;
2303
2304         if (!async) {
2305                 spin_lock(&cli->cl_loi_list_lock);
2306                 osc_check_rpcs(env, cli);
2307                 spin_unlock(&cli->cl_loi_list_lock);
2308         } else {
2309                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2310                 LASSERT(cli->cl_writeback_work != NULL);
2311                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2312         }
2313         return rc;
2314 }
2315 EXPORT_SYMBOL(osc_io_unplug0);
2316
2317 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2318                         struct page *page, loff_t offset)
2319 {
2320         struct obd_export     *exp = osc_export(osc);
2321         struct osc_async_page *oap = &ops->ops_oap;
2322         ENTRY;
2323
2324         if (!page)
2325                 return cfs_size_round(sizeof(*oap));
2326
2327         oap->oap_magic = OAP_MAGIC;
2328         oap->oap_cli = &exp->exp_obd->u.cli;
2329         oap->oap_obj = osc;
2330
2331         oap->oap_page = page;
2332         oap->oap_obj_off = offset;
2333         LASSERT(!(offset & ~PAGE_MASK));
2334
2335         INIT_LIST_HEAD(&oap->oap_pending_item);
2336         INIT_LIST_HEAD(&oap->oap_rpc_item);
2337
2338         spin_lock_init(&oap->oap_lock);
2339         CDEBUG(D_INFO, "oap %p page %p obj off %llu\n",
2340                oap, page, oap->oap_obj_off);
2341         RETURN(0);
2342 }
2343 EXPORT_SYMBOL(osc_prep_async_page);
2344
2345 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2346                        struct osc_page *ops, cl_commit_cbt cb)
2347 {
2348         struct osc_io *oio = osc_env_io(env);
2349         struct osc_extent     *ext = NULL;
2350         struct osc_async_page *oap = &ops->ops_oap;
2351         struct client_obd     *cli = oap->oap_cli;
2352         struct osc_object     *osc = oap->oap_obj;
2353         struct pagevec        *pvec = &osc_env_info(env)->oti_pagevec;
2354         pgoff_t index;
2355         unsigned int tmp;
2356         unsigned int grants = 0;
2357         u32    brw_flags = OBD_BRW_ASYNC;
2358         int    cmd = OBD_BRW_WRITE;
2359         int    need_release = 0;
2360         int    rc = 0;
2361         ENTRY;
2362
2363         if (oap->oap_magic != OAP_MAGIC)
2364                 RETURN(-EINVAL);
2365
2366         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2367                 RETURN(-EIO);
2368
2369         if (!list_empty(&oap->oap_pending_item) ||
2370             !list_empty(&oap->oap_rpc_item))
2371                 RETURN(-EBUSY);
2372
2373         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2374         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2375         if (oio->oi_cap_sys_resource || io->ci_noquota) {
2376                 brw_flags |= OBD_BRW_NOQUOTA;
2377                 cmd |= OBD_BRW_NOQUOTA;
2378         }
2379
2380         /* check if the file's owner/group is over quota */
2381         if (!(cmd & OBD_BRW_NOQUOTA)) {
2382                 struct cl_object *obj;
2383                 struct cl_attr   *attr;
2384                 unsigned int qid[LL_MAXQUOTAS];
2385
2386                 obj = cl_object_top(&osc->oo_cl);
2387                 attr = &osc_env_info(env)->oti_attr;
2388
2389                 cl_object_attr_lock(obj);
2390                 rc = cl_object_attr_get(env, obj, attr);
2391                 cl_object_attr_unlock(obj);
2392
2393                 qid[USRQUOTA] = attr->cat_uid;
2394                 qid[GRPQUOTA] = attr->cat_gid;
2395                 qid[PRJQUOTA] = attr->cat_projid;
2396                 if (rc == 0 && osc_quota_chkdq(cli, qid) == -EDQUOT)
2397                         rc = -EDQUOT;
2398                 if (rc)
2399                         RETURN(rc);
2400         }
2401
2402         oap->oap_cmd = cmd;
2403         oap->oap_page_off = ops->ops_from;
2404         oap->oap_count = ops->ops_to - ops->ops_from;
2405         /* No need to hold a lock here,
2406          * since this page is not in any list yet. */
2407         oap->oap_async_flags = 0;
2408         oap->oap_brw_flags = brw_flags;
2409
2410         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2411                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2412
2413         index = osc_index(oap2osc(oap));
2414
2415         /* Add this page into extent by the following steps:
2416          * 1. if there exists an active extent for this IO, mostly this page
2417          *    can be added to the active extent and sometimes we need to
2418          *    expand extent to accomodate this page;
2419          * 2. otherwise, a new extent will be allocated. */
2420
2421         ext = oio->oi_active;
2422         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2423                 /* one chunk plus extent overhead must be enough to write this
2424                  * page */
2425                 grants = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2426                 if (ext->oe_end >= index)
2427                         grants = 0;
2428
2429                 /* it doesn't need any grant to dirty this page */
2430                 spin_lock(&cli->cl_loi_list_lock);
2431                 rc = osc_enter_cache_try(cli, oap, grants);
2432                 if (rc == 0) { /* try failed */
2433                         grants = 0;
2434                         need_release = 1;
2435                 } else if (ext->oe_end < index) {
2436                         tmp = grants;
2437                         /* try to expand this extent */
2438                         rc = osc_extent_expand(ext, index, &tmp);
2439                         if (rc < 0) {
2440                                 need_release = 1;
2441                                 /* don't free reserved grant */
2442                         } else {
2443                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2444                                                 "expanded for %lu.\n", index);
2445                                 osc_unreserve_grant_nolock(cli, grants, tmp);
2446                                 grants = 0;
2447                         }
2448                 }
2449                 spin_unlock(&cli->cl_loi_list_lock);
2450                 rc = 0;
2451         } else if (ext != NULL) {
2452                 /* index is located outside of active extent */
2453                 need_release = 1;
2454         }
2455         if (need_release) {
2456                 osc_extent_release(env, ext);
2457                 oio->oi_active = NULL;
2458                 ext = NULL;
2459         }
2460
2461         if (ext == NULL) {
2462                 tmp = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2463
2464                 /* try to find new extent to cover this page */
2465                 LASSERT(oio->oi_active == NULL);
2466                 /* we may have allocated grant for this page if we failed
2467                  * to expand the previous active extent. */
2468                 LASSERT(ergo(grants > 0, grants >= tmp));
2469
2470                 rc = 0;
2471                 if (grants == 0) {
2472                         /* We haven't allocated grant for this page, and we
2473                          * must not hold a page lock while we do enter_cache,
2474                          * so we must mark dirty & unlock any pages in the
2475                          * write commit pagevec. */
2476                         if (pagevec_count(pvec)) {
2477                                 cb(env, io, pvec);
2478                                 pagevec_reinit(pvec);
2479                         }
2480                         rc = osc_enter_cache(env, cli, oap, tmp);
2481                         if (rc == 0)
2482                                 grants = tmp;
2483                 }
2484
2485                 tmp = grants;
2486                 if (rc == 0) {
2487                         ext = osc_extent_find(env, osc, index, &tmp);
2488                         if (IS_ERR(ext)) {
2489                                 LASSERT(tmp == grants);
2490                                 osc_exit_cache(cli, oap);
2491                                 rc = PTR_ERR(ext);
2492                                 ext = NULL;
2493                         } else {
2494                                 oio->oi_active = ext;
2495                         }
2496                 }
2497                 if (grants > 0)
2498                         osc_unreserve_grant(cli, grants, tmp);
2499         }
2500
2501         LASSERT(ergo(rc == 0, ext != NULL));
2502         if (ext != NULL) {
2503                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2504                          ext, "index = %lu.\n", index);
2505                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2506
2507                 osc_object_lock(osc);
2508                 if (ext->oe_nr_pages == 0)
2509                         ext->oe_srvlock = ops->ops_srvlock;
2510                 else
2511                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2512                 ++ext->oe_nr_pages;
2513                 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2514                 osc_object_unlock(osc);
2515
2516                 if (!ext->oe_layout_version)
2517                         ext->oe_layout_version = io->ci_layout_version;
2518         }
2519
2520         RETURN(rc);
2521 }
2522
2523 int osc_teardown_async_page(const struct lu_env *env,
2524                             struct osc_object *obj, struct osc_page *ops)
2525 {
2526         struct osc_async_page *oap = &ops->ops_oap;
2527         int rc = 0;
2528         ENTRY;
2529
2530         LASSERT(oap->oap_magic == OAP_MAGIC);
2531
2532         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2533                oap, ops, osc_index(oap2osc(oap)));
2534
2535         if (!list_empty(&oap->oap_rpc_item)) {
2536                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2537                 rc = -EBUSY;
2538         } else if (!list_empty(&oap->oap_pending_item)) {
2539                 struct osc_extent *ext = NULL;
2540
2541                 osc_object_lock(obj);
2542                 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2543                 osc_object_unlock(obj);
2544                 /* only truncated pages are allowed to be taken out.
2545                  * See osc_extent_truncate() and osc_cache_truncate_start()
2546                  * for details. */
2547                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2548                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2549                                         osc_index(oap2osc(oap)));
2550                         rc = -EBUSY;
2551                 }
2552                 if (ext != NULL)
2553                         osc_extent_put(env, ext);
2554         }
2555         RETURN(rc);
2556 }
2557
2558 /**
2559  * This is called when a page is picked up by kernel to write out.
2560  *
2561  * We should find out the corresponding extent and add the whole extent
2562  * into urgent list. The extent may be being truncated or used, handle it
2563  * carefully.
2564  */
2565 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2566                          struct osc_page *ops)
2567 {
2568         struct osc_extent *ext   = NULL;
2569         struct osc_object *obj   = cl2osc(ops->ops_cl.cpl_obj);
2570         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2571         pgoff_t            index = osc_index(ops);
2572         struct osc_async_page *oap = &ops->ops_oap;
2573         bool unplug = false;
2574         int rc = 0;
2575         ENTRY;
2576
2577         osc_object_lock(obj);
2578         ext = osc_extent_lookup(obj, index);
2579         if (ext == NULL) {
2580                 osc_extent_tree_dump(D_ERROR, obj);
2581                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2582         }
2583
2584         switch (ext->oe_state) {
2585         case OES_RPC:
2586         case OES_LOCK_DONE:
2587                 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2588                 LASSERT(0);
2589                 break;
2590         case OES_LOCKING:
2591                 /* If we know this extent is being written out, we should abort
2592                  * so that the writer can make this page ready. Otherwise, there
2593                  * exists a deadlock problem because other process can wait for
2594                  * page writeback bit holding page lock; and meanwhile in
2595                  * vvp_page_make_ready(), we need to grab page lock before
2596                  * really sending the RPC. */
2597         case OES_TRUNC:
2598                 /* race with truncate, page will be redirtied */
2599         case OES_ACTIVE:
2600                 /* The extent is active so we need to abort and let the caller
2601                  * re-dirty the page. If we continued on here, and we were the
2602                  * one making the extent active, we could deadlock waiting for
2603                  * the page writeback to clear but it won't because the extent
2604                  * is active and won't be written out. */
2605                 GOTO(out, rc = -EAGAIN);
2606         default:
2607                 break;
2608         }
2609
2610         rc = cl_page_prep(env, io, cp, CRT_WRITE);
2611         if (rc)
2612                 GOTO(out, rc);
2613
2614         spin_lock(&oap->oap_lock);
2615         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2616         spin_unlock(&oap->oap_lock);
2617
2618         if (memory_pressure_get())
2619                 ext->oe_memalloc = 1;
2620
2621         ext->oe_urgent = 1;
2622         if (ext->oe_state == OES_CACHE) {
2623                 OSC_EXTENT_DUMP(D_CACHE, ext,
2624                                 "flush page %p make it urgent.\n", oap);
2625                 if (list_empty(&ext->oe_link))
2626                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2627                 unplug = true;
2628         }
2629         rc = 0;
2630         EXIT;
2631
2632 out:
2633         osc_object_unlock(obj);
2634         osc_extent_put(env, ext);
2635         if (unplug)
2636                 osc_io_unplug_async(env, osc_cli(obj), obj);
2637         return rc;
2638 }
2639
2640 int osc_queue_sync_pages(const struct lu_env *env, const struct cl_io *io,
2641                          struct osc_object *obj, struct list_head *list,
2642                          int brw_flags)
2643 {
2644         struct client_obd     *cli = osc_cli(obj);
2645         struct osc_extent     *ext;
2646         struct osc_async_page *oap;
2647         int     page_count = 0;
2648         int     mppr       = cli->cl_max_pages_per_rpc;
2649         bool    can_merge   = true;
2650         pgoff_t start      = CL_PAGE_EOF;
2651         pgoff_t end        = 0;
2652         ENTRY;
2653
2654         list_for_each_entry(oap, list, oap_pending_item) {
2655                 struct osc_page *opg = oap2osc_page(oap);
2656                 pgoff_t index = osc_index(opg);
2657
2658                 if (index > end)
2659                         end = index;
2660                 if (index < start)
2661                         start = index;
2662                 ++page_count;
2663                 mppr <<= (page_count > mppr);
2664
2665                 if (unlikely(opg->ops_from > 0 || opg->ops_to < PAGE_SIZE))
2666                         can_merge = false;
2667         }
2668
2669         ext = osc_extent_alloc(obj);
2670         if (ext == NULL) {
2671                 struct osc_async_page *tmp;
2672
2673                 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2674                         list_del_init(&oap->oap_pending_item);
2675                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2676                 }
2677                 RETURN(-ENOMEM);
2678         }
2679
2680         ext->oe_rw = !!(brw_flags & OBD_BRW_READ);
2681         ext->oe_sync = 1;
2682         ext->oe_no_merge = !can_merge;
2683         ext->oe_urgent = 1;
2684         ext->oe_start = start;
2685         ext->oe_end = ext->oe_max_end = end;
2686         ext->oe_obj = obj;
2687         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2688         ext->oe_ndelay = !!(brw_flags & OBD_BRW_NDELAY);
2689         ext->oe_dio = !!(brw_flags & OBD_BRW_NOCACHE);
2690         ext->oe_nr_pages = page_count;
2691         ext->oe_mppr = mppr;
2692         list_splice_init(list, &ext->oe_pages);
2693         ext->oe_layout_version = io->ci_layout_version;
2694
2695         osc_object_lock(obj);
2696         /* Reuse the initial refcount for RPC, don't drop it */
2697         osc_extent_state_set(ext, OES_LOCK_DONE);
2698         if (!ext->oe_rw) { /* write */
2699                 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2700                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2701         } else {
2702                 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2703                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2704         }
2705         osc_object_unlock(obj);
2706
2707         osc_io_unplug_async(env, cli, obj);
2708         RETURN(0);
2709 }
2710
2711 /**
2712  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2713  */
2714 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
2715                              __u64 size, struct osc_extent **extp)
2716 {
2717         struct client_obd *cli = osc_cli(obj);
2718         struct osc_extent *ext;
2719         struct osc_extent *waiting = NULL;
2720         pgoff_t index;
2721         LIST_HEAD(list);
2722         int result = 0;
2723         bool partial;
2724         ENTRY;
2725
2726         /* pages with index greater or equal to index will be truncated. */
2727         index = cl_index(osc2cl(obj), size);
2728         partial = size > cl_offset(osc2cl(obj), index);
2729
2730 again:
2731         osc_object_lock(obj);
2732         ext = osc_extent_search(obj, index);
2733         if (ext == NULL)
2734                 ext = first_extent(obj);
2735         else if (ext->oe_end < index)
2736                 ext = next_extent(ext);
2737         while (ext != NULL) {
2738                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2739
2740                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2741                         /* if ext is in urgent state, it means there must exist
2742                          * a page already having been flushed by write_page().
2743                          * We have to wait for this extent because we can't
2744                          * truncate that page. */
2745                         OSC_EXTENT_DUMP(D_CACHE, ext,
2746                                         "waiting for busy extent\n");
2747                         waiting = osc_extent_get(ext);
2748                         break;
2749                 }
2750
2751                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:%llu.\n", size);
2752
2753                 osc_extent_get(ext);
2754                 if (ext->oe_state == OES_ACTIVE) {
2755                         /* though we grab inode mutex for write path, but we
2756                          * release it before releasing extent(in osc_io_end()),
2757                          * so there is a race window that an extent is still
2758                          * in OES_ACTIVE when truncate starts. */
2759                         LASSERT(!ext->oe_trunc_pending);
2760                         ext->oe_trunc_pending = 1;
2761                 } else {
2762                         EASSERT(ext->oe_state == OES_CACHE, ext);
2763                         osc_extent_state_set(ext, OES_TRUNC);
2764                         osc_update_pending(obj, OBD_BRW_WRITE,
2765                                            -ext->oe_nr_pages);
2766                 }
2767                 /* This extent could be on the full extents list, that's OK */
2768                 EASSERT(!ext->oe_hp && !ext->oe_urgent, ext);
2769                 if (!list_empty(&ext->oe_link))
2770                         list_move_tail(&ext->oe_link, &list);
2771                 else
2772                         list_add_tail(&ext->oe_link, &list);
2773
2774                 ext = next_extent(ext);
2775         }
2776         osc_object_unlock(obj);
2777
2778         osc_list_maint(cli, obj);
2779
2780         while (!list_empty(&list)) {
2781                 int rc;
2782
2783                 ext = list_entry(list.next, struct osc_extent, oe_link);
2784                 list_del_init(&ext->oe_link);
2785
2786                 /* extent may be in OES_ACTIVE state because inode mutex
2787                  * is released before osc_io_end() in file write case */
2788                 if (ext->oe_state != OES_TRUNC)
2789                         osc_extent_wait(env, ext, OES_TRUNC);
2790
2791                 rc = osc_extent_truncate(ext, index, partial);
2792                 if (rc < 0) {
2793                         if (result == 0)
2794                                 result = rc;
2795
2796                         OSC_EXTENT_DUMP(D_ERROR, ext,
2797                                         "truncate error %d\n", rc);
2798                 } else if (ext->oe_nr_pages == 0) {
2799                         osc_extent_remove(ext);
2800                 } else {
2801                         /* this must be an overlapped extent which means only
2802                          * part of pages in this extent have been truncated.
2803                          */
2804                         EASSERTF(ext->oe_start <= index, ext,
2805                                  "trunc index = %lu/%d.\n", index, partial);
2806                         /* fix index to skip this partially truncated extent */
2807                         index = ext->oe_end + 1;
2808                         partial = false;
2809
2810                         /* we need to hold this extent in OES_TRUNC state so
2811                          * that no writeback will happen. This is to avoid
2812                          * BUG 17397.
2813                          * Only partial truncate can reach here, if @size is
2814                          * not zero, the caller should provide a valid @extp. */
2815                         LASSERT(*extp == NULL);
2816                         *extp = osc_extent_get(ext);
2817                         OSC_EXTENT_DUMP(D_CACHE, ext,
2818                                         "trunc at %llu\n", size);
2819                 }
2820                 osc_extent_put(env, ext);
2821         }
2822         if (waiting != NULL) {
2823                 int rc;
2824
2825                 /* ignore the result of osc_extent_wait the write initiator
2826                  * should take care of it. */
2827                 rc = osc_extent_wait(env, waiting, OES_INV);
2828                 if (rc < 0)
2829                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2830
2831                 osc_extent_put(env, waiting);
2832                 waiting = NULL;
2833                 goto again;
2834         }
2835         RETURN(result);
2836 }
2837 EXPORT_SYMBOL(osc_cache_truncate_start);
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 EXPORT_SYMBOL(osc_cache_wait_range);
2925
2926 /**
2927  * Called to write out a range of osc object.
2928  *
2929  * @hp     : should be set this is caused by lock cancel;
2930  * @discard: is set if dirty pages should be dropped - file will be deleted or
2931  *         truncated, this implies there is no partially discarding extents.
2932  *
2933  * Return how many pages will be issued, or error code if error occurred.
2934  */
2935 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2936                               pgoff_t start, pgoff_t end, int hp, int discard)
2937 {
2938         struct osc_extent *ext;
2939         LIST_HEAD(discard_list);
2940         bool unplug = false;
2941         int result = 0;
2942         ENTRY;
2943
2944         osc_object_lock(obj);
2945         ext = osc_extent_search(obj, start);
2946         if (ext == NULL)
2947                 ext = first_extent(obj);
2948         else if (ext->oe_end < start)
2949                 ext = next_extent(ext);
2950         while (ext != NULL) {
2951                 if (ext->oe_start > end)
2952                         break;
2953
2954                 ext->oe_fsync_wait = 1;
2955                 switch (ext->oe_state) {
2956                 case OES_CACHE:
2957                         result += ext->oe_nr_pages;
2958                         if (!discard) {
2959                                 struct list_head *list = NULL;
2960                                 if (hp) {
2961                                         EASSERT(!ext->oe_hp, ext);
2962                                         ext->oe_hp = 1;
2963                                         list = &obj->oo_hp_exts;
2964                                 } else if (!ext->oe_urgent) {
2965                                         ext->oe_urgent = 1;
2966                                         list = &obj->oo_urgent_exts;
2967                                 }
2968                                 if (list != NULL)
2969                                         list_move_tail(&ext->oe_link, list);
2970                                 unplug = true;
2971                         } else {
2972                                 struct client_obd *cli = osc_cli(obj);
2973                                 int pcc_bits = cli->cl_chunkbits - PAGE_SHIFT;
2974                                 pgoff_t align_by = (1 << pcc_bits);
2975                                 pgoff_t a_start = round_down(start, align_by);
2976                                 pgoff_t a_end = round_up(end, align_by);
2977
2978                                 /* overflow case */
2979                                 if (end && !a_end)
2980                                         a_end = CL_PAGE_EOF;
2981                                 /* the only discarder is lock cancelling, so
2982                                  * [start, end], aligned by chunk size, must
2983                                  * contain this extent */
2984                                 LASSERTF(ext->oe_start >= a_start &&
2985                                          ext->oe_end <= a_end,
2986                                          "ext [%lu, %lu] reg [%lu, %lu] "
2987                                          "orig [%lu %lu] align %lu bits "
2988                                          "%d\n", ext->oe_start, ext->oe_end,
2989                                          a_start, a_end, start, end,
2990                                          align_by, pcc_bits);
2991                                 osc_extent_state_set(ext, OES_LOCKING);
2992                                 ext->oe_owner = current;
2993                                 list_move_tail(&ext->oe_link,
2994                                                    &discard_list);
2995                                 osc_update_pending(obj, OBD_BRW_WRITE,
2996                                                    -ext->oe_nr_pages);
2997                         }
2998                         break;
2999                 case OES_ACTIVE:
3000                         /* It's pretty bad to wait for ACTIVE extents, because
3001                          * we don't know how long we will wait for it to be
3002                          * flushed since it may be blocked at awaiting more
3003                          * grants. We do this for the correctness of fsync. */
3004                         LASSERT(hp == 0 && discard == 0);
3005                         ext->oe_urgent = 1;
3006                         break;
3007                 case OES_TRUNC:
3008                         /* this extent is being truncated, can't do anything
3009                          * for it now. it will be set to urgent after truncate
3010                          * is finished in osc_cache_truncate_end(). */
3011                 default:
3012                         break;
3013                 }
3014                 ext = next_extent(ext);
3015         }
3016         osc_object_unlock(obj);
3017
3018         LASSERT(ergo(!discard, list_empty(&discard_list)));
3019         if (!list_empty(&discard_list)) {
3020                 struct osc_extent *tmp;
3021                 int rc;
3022
3023                 osc_list_maint(osc_cli(obj), obj);
3024                 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
3025                         list_del_init(&ext->oe_link);
3026                         EASSERT(ext->oe_state == OES_LOCKING, ext);
3027
3028                         /* Discard caching pages. We don't actually write this
3029                          * extent out but we complete it as if we did. */
3030                         rc = osc_extent_make_ready(env, ext);
3031                         if (unlikely(rc < 0)) {
3032                                 OSC_EXTENT_DUMP(D_ERROR, ext,
3033                                                 "make_ready returned %d\n", rc);
3034                                 if (result >= 0)
3035                                         result = rc;
3036                         }
3037
3038                         /* finish the extent as if the pages were sent */
3039                         osc_extent_finish(env, ext, 0, 0);
3040                 }
3041         }
3042
3043         if (unplug)
3044                 osc_io_unplug(env, osc_cli(obj), obj);
3045
3046         if (hp || discard) {
3047                 int rc;
3048                 rc = osc_cache_wait_range(env, obj, start, end);
3049                 if (result >= 0 && rc < 0)
3050                         result = rc;
3051         }
3052
3053         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3054         RETURN(result);
3055 }
3056 EXPORT_SYMBOL(osc_cache_writeback_range);
3057
3058 /**
3059  * Returns a list of pages by a given [start, end] of \a obj.
3060  *
3061  * \param resched If not NULL, then we give up before hogging CPU for too
3062  * long and set *resched = 1, in that case caller should implement a retry
3063  * logic.
3064  *
3065  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3066  * crucial in the face of [offset, EOF] locks.
3067  *
3068  * Return at least one page in @queue unless there is no covered page.
3069  */
3070 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3071                         struct osc_object *osc, pgoff_t start, pgoff_t end,
3072                         osc_page_gang_cbt cb, void *cbdata)
3073 {
3074         struct osc_page *ops;
3075         struct pagevec  *pagevec;
3076         void            **pvec;
3077         pgoff_t         idx;
3078         unsigned int    nr;
3079         unsigned int    i;
3080         unsigned int    j;
3081         int             res = CLP_GANG_OKAY;
3082         bool            tree_lock = true;
3083         ENTRY;
3084
3085         idx = start;
3086         pvec = osc_env_info(env)->oti_pvec;
3087         pagevec = &osc_env_info(env)->oti_pagevec;
3088         ll_pagevec_init(pagevec, 0);
3089         spin_lock(&osc->oo_tree_lock);
3090         while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3091                                             idx, OTI_PVEC_SIZE)) > 0) {
3092                 struct cl_page *page;
3093                 bool end_of_region = false;
3094
3095                 for (i = 0, j = 0; i < nr; ++i) {
3096                         ops = pvec[i];
3097                         pvec[i] = NULL;
3098
3099                         idx = osc_index(ops);
3100                         if (idx > end) {
3101                                 end_of_region = true;
3102                                 break;
3103                         }
3104
3105                         page = ops->ops_cl.cpl_page;
3106                         LASSERT(page->cp_type == CPT_CACHEABLE);
3107                         if (page->cp_state == CPS_FREEING)
3108                                 continue;
3109
3110                         cl_page_get(page);
3111                         lu_ref_add_atomic(&page->cp_reference,
3112                                           "gang_lookup", current);
3113                         pvec[j++] = ops;
3114                 }
3115                 ++idx;
3116
3117                 /*
3118                  * Here a delicate locking dance is performed. Current thread
3119                  * holds a reference to a page, but has to own it before it
3120                  * can be placed into queue. Owning implies waiting, so
3121                  * radix-tree lock is to be released. After a wait one has to
3122                  * check that pages weren't truncated (cl_page_own() returns
3123                  * error in the latter case).
3124                  */
3125                 spin_unlock(&osc->oo_tree_lock);
3126                 tree_lock = false;
3127
3128                 for (i = 0; i < j; ++i) {
3129                         ops = pvec[i];
3130                         if (res == CLP_GANG_OKAY)
3131                                 res = (*cb)(env, io, ops, cbdata);
3132
3133                         page = ops->ops_cl.cpl_page;
3134                         lu_ref_del(&page->cp_reference, "gang_lookup", current);
3135                         cl_pagevec_put(env, page, pagevec);
3136                 }
3137                 pagevec_release(pagevec);
3138
3139                 if (nr < OTI_PVEC_SIZE || end_of_region)
3140                         break;
3141
3142                 if (res == CLP_GANG_OKAY && need_resched())
3143                         res = CLP_GANG_RESCHED;
3144                 if (res != CLP_GANG_OKAY)
3145                         break;
3146
3147                 spin_lock(&osc->oo_tree_lock);
3148                 tree_lock = true;
3149         }
3150         if (tree_lock)
3151                 spin_unlock(&osc->oo_tree_lock);
3152         RETURN(res);
3153 }
3154 EXPORT_SYMBOL(osc_page_gang_lookup);
3155
3156 /**
3157  * Check if page @page is covered by an extra lock or discard it.
3158  */
3159 static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3160                                 struct osc_page *ops, void *cbdata)
3161 {
3162         struct osc_thread_info *info = osc_env_info(env);
3163         struct osc_object *osc = cbdata;
3164         pgoff_t index;
3165
3166         index = osc_index(ops);
3167         if (index >= info->oti_fn_index) {
3168                 struct ldlm_lock *tmp;
3169                 struct cl_page *page = ops->ops_cl.cpl_page;
3170
3171                 /* refresh non-overlapped index */
3172                 tmp = osc_dlmlock_at_pgoff(env, osc, index,
3173                                            OSC_DAP_FL_TEST_LOCK);
3174                 if (tmp != NULL) {
3175                         __u64 end = tmp->l_policy_data.l_extent.end;
3176                         /* Cache the first-non-overlapped index so as to skip
3177                          * all pages within [index, oti_fn_index). This is safe
3178                          * because if tmp lock is canceled, it will discard
3179                          * these pages. */
3180                         info->oti_fn_index = cl_index(osc2cl(osc), end + 1);
3181                         if (end == OBD_OBJECT_EOF)
3182                                 info->oti_fn_index = CL_PAGE_EOF;
3183                         LDLM_LOCK_PUT(tmp);
3184                 } else if (cl_page_own(env, io, page) == 0) {
3185                         /* discard the page */
3186                         cl_page_discard(env, io, page);
3187                         cl_page_disown(env, io, page);
3188                 } else {
3189                         LASSERT(page->cp_state == CPS_FREEING);
3190                 }
3191         }
3192
3193         info->oti_next_index = index + 1;
3194         return CLP_GANG_OKAY;
3195 }
3196
3197 int osc_discard_cb(const struct lu_env *env, struct cl_io *io,
3198                    struct osc_page *ops, void *cbdata)
3199 {
3200         struct osc_thread_info *info = osc_env_info(env);
3201         struct cl_page *page = ops->ops_cl.cpl_page;
3202
3203         /* page is top page. */
3204         info->oti_next_index = osc_index(ops) + 1;
3205         if (cl_page_own(env, io, page) == 0) {
3206                 if (!ergo(page->cp_type == CPT_CACHEABLE,
3207                           !PageDirty(cl_page_vmpage(page))))
3208                         CL_PAGE_DEBUG(D_ERROR, env, page,
3209                                         "discard dirty page?\n");
3210
3211                 /* discard the page */
3212                 cl_page_discard(env, io, page);
3213                 cl_page_disown(env, io, page);
3214         } else {
3215                 LASSERT(page->cp_state == CPS_FREEING);
3216         }
3217
3218         return CLP_GANG_OKAY;
3219 }
3220 EXPORT_SYMBOL(osc_discard_cb);
3221
3222 /**
3223  * Discard pages protected by the given lock. This function traverses radix
3224  * tree to find all covering pages and discard them. If a page is being covered
3225  * by other locks, it should remain in cache.
3226  *
3227  * If error happens on any step, the process continues anyway (the reasoning
3228  * behind this being that lock cancellation cannot be delayed indefinitely).
3229  */
3230 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3231                            pgoff_t start, pgoff_t end, bool discard)
3232 {
3233         struct osc_thread_info *info = osc_env_info(env);
3234         struct cl_io *io = osc_env_thread_io(env);
3235         osc_page_gang_cbt cb;
3236         int res;
3237         int result;
3238
3239         ENTRY;
3240
3241         io->ci_obj = cl_object_top(osc2cl(osc));
3242         io->ci_ignore_layout = 1;
3243         result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3244         if (result != 0)
3245                 GOTO(out, result);
3246
3247         cb = discard ? osc_discard_cb : check_and_discard_cb;
3248         info->oti_fn_index = info->oti_next_index = start;
3249         do {
3250                 res = osc_page_gang_lookup(env, io, osc,
3251                                            info->oti_next_index, end, cb, osc);
3252                 if (info->oti_next_index > end)
3253                         break;
3254
3255                 if (res == CLP_GANG_RESCHED)
3256                         cond_resched();
3257         } while (res != CLP_GANG_OKAY);
3258 out:
3259         cl_io_fini(env, io);
3260         RETURN(result);
3261 }
3262
3263
3264 /** @} osc */