lustre/osc/osc_cache.c

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