lustre/osc/osc_cache.c

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