lustre/osc/osc_cache.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  *
  28  */
  29 /*
  30  * This file is part of Lustre, http://www.lustre.org/
  31  * Lustre is a trademark of Sun Microsystems, Inc.
  32  *
  33  * osc cache management.
  34  *
  35  * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_OSC
  39
  40 #include <lustre_osc.h>
  41 #include <lustre_dlm.h>
  42
  43 #include "osc_internal.h"
  44
  45 static int extent_debug; /* set it to be true for more debug */
  46
  47 static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
  48 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
  49                            enum osc_extent_state state);
  50 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
  51                               struct osc_async_page *oap, int sent, int rc);
  52 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
  53                           int cmd);
  54 static int osc_refresh_count(const struct lu_env *env,
  55                              struct osc_async_page *oap, int cmd);
  56 static int osc_io_unplug_async(const struct lu_env *env,
  57                                struct client_obd *cli, struct osc_object *osc);
  58 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
  59                            unsigned int lost_grant, unsigned int dirty_grant);
  60
  61 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
  62                                   const char *func, int line);
  63 #define osc_extent_tree_dump(lvl, obj) \
  64         osc_extent_tree_dump0(lvl, obj, __func__, __LINE__)
  65
  66 static void osc_unreserve_grant(struct client_obd *cli, unsigned int reserved,
  67                                 unsigned int unused);
  68
  69 /** \addtogroup osc
  70  *  @{
  71  */
  72
  73 /* ------------------ osc extent ------------------ */
  74 static inline char *ext_flags(struct osc_extent *ext, char *flags)
  75 {
  76         char *buf = flags;
  77         *buf++ = ext->oe_rw ? 'r' : 'w';
  78         if (ext->oe_intree)
  79                 *buf++ = 'i';
  80         if (ext->oe_sync)
  81                 *buf++ = 'S';
  82         if (ext->oe_srvlock)
  83                 *buf++ = 's';
  84         if (ext->oe_hp)
  85                 *buf++ = 'h';
  86         if (ext->oe_urgent)
  87                 *buf++ = 'u';
  88         if (ext->oe_memalloc)
  89                 *buf++ = 'm';
  90         if (ext->oe_trunc_pending)
  91                 *buf++ = 't';
  92         if (ext->oe_fsync_wait)
  93                 *buf++ = 'Y';
  94         *buf = 0;
  95         return flags;
  96 }
  97
  98 static inline char list_empty_marker(struct list_head *list)
  99 {
 100         return list_empty(list) ? '-' : '+';
 101 }
 102
 103 #define EXTSTR       "[%lu -> %lu/%lu]"
 104 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
 105 static const char *oes_strings[] = {
 106         "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
 107
 108 #define OSC_EXTENT_DUMP(lvl, extent, fmt, ...) do {                           \
 109         struct osc_extent *__ext = (extent);                                  \
 110         char __buf[16];                                                       \
 111                                                                               \
 112         CDEBUG(lvl,                                                           \
 113                 "extent %p@{" EXTSTR ", "                                     \
 114                 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt,          \
 115                 /* ----- extent part 0 ----- */                               \
 116                 __ext, EXTPARA(__ext),                                        \
 117                 /* ----- part 1 ----- */                                      \
 118                 atomic_read(&__ext->oe_refc),                         \
 119                 atomic_read(&__ext->oe_users),                        \
 120                 list_empty_marker(&__ext->oe_link),                           \
 121                 oes_strings[__ext->oe_state], ext_flags(__ext, __buf),        \
 122                 __ext->oe_obj,                                                \
 123                 /* ----- part 2 ----- */                                      \
 124                 __ext->oe_grants, __ext->oe_nr_pages,                         \
 125                 list_empty_marker(&__ext->oe_pages),                          \
 126                 waitqueue_active(&__ext->oe_waitq) ? '+' : '-',               \
 127                 __ext->oe_dlmlock, __ext->oe_mppr, __ext->oe_owner,           \
 128                 /* ----- part 4 ----- */                                      \
 129                 ## __VA_ARGS__);                                              \
 130         if (lvl == D_ERROR && __ext->oe_dlmlock != NULL)                      \
 131                 LDLM_ERROR(__ext->oe_dlmlock, "extent: %p", __ext);           \
 132         else                                                                  \
 133                 LDLM_DEBUG(__ext->oe_dlmlock, "extent: %p", __ext);           \
 134 } while (0)
 135
 136 #undef EASSERTF
 137 #define EASSERTF(expr, ext, fmt, args...) do {                          \
 138         if (!(expr)) {                                                  \
 139                 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args);           \
 140                 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj);           \
 141                 LASSERT(expr);                                          \
 142         }                                                               \
 143 } while (0)
 144
 145 #undef EASSERT
 146 #define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
 147
 148 static inline struct osc_extent *rb_extent(struct rb_node *n)
 149 {
 150         if (n == NULL)
 151                 return NULL;
 152
 153         return container_of(n, struct osc_extent, oe_node);
 154 }
 155
 156 static inline struct osc_extent *next_extent(struct osc_extent *ext)
 157 {
 158         if (ext == NULL)
 159                 return NULL;
 160
 161         LASSERT(ext->oe_intree);
 162         return rb_extent(rb_next(&ext->oe_node));
 163 }
 164
 165 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
 166 {
 167         if (ext == NULL)
 168                 return NULL;
 169
 170         LASSERT(ext->oe_intree);
 171         return rb_extent(rb_prev(&ext->oe_node));
 172 }
 173
 174 static inline struct osc_extent *first_extent(struct osc_object *obj)
 175 {
 176         return rb_extent(rb_first(&obj->oo_root));
 177 }
 178
 179 /* object must be locked by caller. */
 180 static int osc_extent_sanity_check0(struct osc_extent *ext,
 181                                     const char *func, const int line)
 182 {
 183         struct osc_object *obj = ext->oe_obj;
 184         struct osc_async_page *oap;
 185         size_t page_count;
 186         int rc = 0;
 187
 188         if (!osc_object_is_locked(obj))
 189                 GOTO(out, rc = 9);
 190
 191         if (ext->oe_state >= OES_STATE_MAX)
 192                 GOTO(out, rc = 10);
 193
 194         if (atomic_read(&ext->oe_refc) <= 0)
 195                 GOTO(out, rc = 20);
 196
 197         if (atomic_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
 198                 GOTO(out, rc = 30);
 199
 200         switch (ext->oe_state) {
 201         case OES_INV:
 202                 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
 203                         GOTO(out, rc = 35);
 204                 GOTO(out, rc = 0);
 205                 break;
 206         case OES_ACTIVE:
 207                 if (atomic_read(&ext->oe_users) == 0)
 208                         GOTO(out, rc = 40);
 209                 if (ext->oe_hp)
 210                         GOTO(out, rc = 50);
 211                 if (ext->oe_fsync_wait && !ext->oe_urgent)
 212                         GOTO(out, rc = 55);
 213                 break;
 214         case OES_CACHE:
 215                 if (ext->oe_grants == 0)
 216                         GOTO(out, rc = 60);
 217                 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
 218                         GOTO(out, rc = 65);
 219         default:
 220                 if (atomic_read(&ext->oe_users) > 0)
 221                         GOTO(out, rc = 70);
 222         }
 223
 224         if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
 225                 GOTO(out, rc = 80);
 226
 227         if (ext->oe_sync && ext->oe_grants > 0)
 228                 GOTO(out, rc = 90);
 229
 230         if (ext->oe_dlmlock != NULL &&
 231             ext->oe_dlmlock->l_resource->lr_type == LDLM_EXTENT &&
 232             !ldlm_is_failed(ext->oe_dlmlock)) {
 233                 struct ldlm_extent *extent;
 234
 235                 extent = &ext->oe_dlmlock->l_policy_data.l_extent;
 236                 if (!(extent->start <= cl_offset(osc2cl(obj), ext->oe_start) &&
 237                       extent->end   >= cl_offset(osc2cl(obj), ext->oe_max_end)))
 238                         GOTO(out, rc = 100);
 239
 240                 if (!(ext->oe_dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP)))
 241                         GOTO(out, rc = 102);
 242         }
 243
 244         if (ext->oe_nr_pages > ext->oe_mppr)
 245                 GOTO(out, rc = 105);
 246
 247         /* Do not verify page list if extent is in RPC. This is because an
 248          * in-RPC extent is supposed to be exclusively accessible w/o lock. */
 249         if (ext->oe_state > OES_CACHE)
 250                 GOTO(out, rc = 0);
 251
 252         if (!extent_debug)
 253                 GOTO(out, rc = 0);
 254
 255         page_count = 0;
 256         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
 257                 pgoff_t index = osc_index(oap2osc(oap));
 258                 ++page_count;
 259                 if (index > ext->oe_end || index < ext->oe_start)
 260                         GOTO(out, rc = 110);
 261         }
 262         if (page_count != ext->oe_nr_pages)
 263                 GOTO(out, rc = 120);
 264
 265 out:
 266         if (rc != 0)
 267                 OSC_EXTENT_DUMP(D_ERROR, ext,
 268                                 "%s:%d sanity check %p failed with rc = %d\n",
 269                                 func, line, ext, rc);
 270         return rc;
 271 }
 272
 273 #define sanity_check_nolock(ext) \
 274         osc_extent_sanity_check0(ext, __func__, __LINE__)
 275
 276 #define sanity_check(ext) ({                                                   \
 277         int __res;                                                             \
 278         osc_object_lock((ext)->oe_obj);                                        \
 279         __res = sanity_check_nolock(ext);                                      \
 280         osc_object_unlock((ext)->oe_obj);                                      \
 281         __res;                                                                 \
 282 })
 283
 284
 285 /**
 286  * sanity check - to make sure there is no overlapped extent in the tree.
 287  */
 288 static int osc_extent_is_overlapped(struct osc_object *obj,
 289                                     struct osc_extent *ext)
 290 {
 291         struct osc_extent *tmp;
 292
 293         LASSERT(osc_object_is_locked(obj));
 294
 295         if (!extent_debug)
 296                 return 0;
 297
 298         for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
 299                 if (tmp == ext)
 300                         continue;
 301                 if (tmp->oe_end >= ext->oe_start &&
 302                     tmp->oe_start <= ext->oe_end)
 303                         return 1;
 304         }
 305         return 0;
 306 }
 307
 308 static void osc_extent_state_set(struct osc_extent *ext, int state)
 309 {
 310         LASSERT(osc_object_is_locked(ext->oe_obj));
 311         LASSERT(state >= OES_INV && state < OES_STATE_MAX);
 312
 313         /* Never try to sanity check a state changing extent :-) */
 314         /* LASSERT(sanity_check_nolock(ext) == 0); */
 315
 316         /* TODO: validate the state machine */
 317         ext->oe_state = state;
 318         wake_up_all(&ext->oe_waitq);
 319 }
 320
 321 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
 322 {
 323         struct osc_extent *ext;
 324
 325         OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_NOFS);
 326         if (ext == NULL)
 327                 return NULL;
 328
 329         RB_CLEAR_NODE(&ext->oe_node);
 330         ext->oe_obj = obj;
 331         cl_object_get(osc2cl(obj));
 332         atomic_set(&ext->oe_refc, 1);
 333         atomic_set(&ext->oe_users, 0);
 334         INIT_LIST_HEAD(&ext->oe_link);
 335         ext->oe_state = OES_INV;
 336         INIT_LIST_HEAD(&ext->oe_pages);
 337         init_waitqueue_head(&ext->oe_waitq);
 338         ext->oe_dlmlock = NULL;
 339
 340         return ext;
 341 }
 342
 343 static void osc_extent_free(struct osc_extent *ext)
 344 {
 345         OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
 346 }
 347
 348 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
 349 {
 350         LASSERT(atomic_read(&ext->oe_refc) >= 0);
 351         atomic_inc(&ext->oe_refc);
 352         return ext;
 353 }
 354
 355 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
 356 {
 357         LASSERT(atomic_read(&ext->oe_refc) > 0);
 358         if (atomic_dec_and_test(&ext->oe_refc)) {
 359                 LASSERT(list_empty(&ext->oe_link));
 360                 LASSERT(atomic_read(&ext->oe_users) == 0);
 361                 LASSERT(ext->oe_state == OES_INV);
 362                 LASSERT(!ext->oe_intree);
 363
 364                 if (ext->oe_dlmlock != NULL) {
 365                         lu_ref_add(&ext->oe_dlmlock->l_reference,
 366                                    "osc_extent", ext);
 367                         LDLM_LOCK_PUT(ext->oe_dlmlock);
 368                         ext->oe_dlmlock = NULL;
 369                 }
 370                 cl_object_put(env, osc2cl(ext->oe_obj));
 371                 osc_extent_free(ext);
 372         }
 373 }
 374
 375 /**
 376  * osc_extent_put_trust() is a special version of osc_extent_put() when
 377  * it's known that the caller is not the last user. This is to address the
 378  * problem of lacking of lu_env ;-).
 379  */
 380 static void osc_extent_put_trust(struct osc_extent *ext)
 381 {
 382         LASSERT(atomic_read(&ext->oe_refc) > 1);
 383         LASSERT(osc_object_is_locked(ext->oe_obj));
 384         atomic_dec(&ext->oe_refc);
 385 }
 386
 387 /**
 388  * Return the extent which includes pgoff @index, or return the greatest
 389  * previous extent in the tree.
 390  */
 391 static struct osc_extent *osc_extent_search(struct osc_object *obj,
 392                                             pgoff_t index)
 393 {
 394         struct rb_node    *n = obj->oo_root.rb_node;
 395         struct osc_extent *tmp, *p = NULL;
 396
 397         LASSERT(osc_object_is_locked(obj));
 398         while (n != NULL) {
 399                 tmp = rb_extent(n);
 400                 if (index < tmp->oe_start) {
 401                         n = n->rb_left;
 402                 } else if (index > tmp->oe_end) {
 403                         p = rb_extent(n);
 404                         n = n->rb_right;
 405                 } else {
 406                         return tmp;
 407                 }
 408         }
 409         return p;
 410 }
 411
 412 /*
 413  * Return the extent covering @index, otherwise return NULL.
 414  * caller must have held object lock.
 415  */
 416 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
 417                                             pgoff_t index)
 418 {
 419         struct osc_extent *ext;
 420
 421         ext = osc_extent_search(obj, index);
 422         if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
 423                 return osc_extent_get(ext);
 424         return NULL;
 425 }
 426
 427 /* caller must have held object lock. */
 428 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
 429 {
 430         struct rb_node   **n      = &obj->oo_root.rb_node;
 431         struct rb_node    *parent = NULL;
 432         struct osc_extent *tmp;
 433
 434         LASSERT(ext->oe_intree == 0);
 435         LASSERT(ext->oe_obj == obj);
 436         LASSERT(osc_object_is_locked(obj));
 437         while (*n != NULL) {
 438                 tmp = rb_extent(*n);
 439                 parent = *n;
 440
 441                 if (ext->oe_end < tmp->oe_start)
 442                         n = &(*n)->rb_left;
 443                 else if (ext->oe_start > tmp->oe_end)
 444                         n = &(*n)->rb_right;
 445                 else
 446                         EASSERTF(0, tmp, EXTSTR"\n", EXTPARA(ext));
 447         }
 448         rb_link_node(&ext->oe_node, parent, n);
 449         rb_insert_color(&ext->oe_node, &obj->oo_root);
 450         osc_extent_get(ext);
 451         ext->oe_intree = 1;
 452 }
 453
 454 /* caller must have held object lock. */
 455 static void osc_extent_erase(struct osc_extent *ext)
 456 {
 457         struct osc_object *obj = ext->oe_obj;
 458         LASSERT(osc_object_is_locked(obj));
 459         if (ext->oe_intree) {
 460                 rb_erase(&ext->oe_node, &obj->oo_root);
 461                 ext->oe_intree = 0;
 462                 /* rbtree held a refcount */
 463                 osc_extent_put_trust(ext);
 464         }
 465 }
 466
 467 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
 468 {
 469         struct osc_object *obj = ext->oe_obj;
 470
 471         LASSERT(osc_object_is_locked(obj));
 472         LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
 473         if (ext->oe_state == OES_CACHE) {
 474                 osc_extent_state_set(ext, OES_ACTIVE);
 475                 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
 476         }
 477         atomic_inc(&ext->oe_users);
 478         list_del_init(&ext->oe_link);
 479         return osc_extent_get(ext);
 480 }
 481
 482 static void __osc_extent_remove(struct osc_extent *ext)
 483 {
 484         LASSERT(osc_object_is_locked(ext->oe_obj));
 485         LASSERT(list_empty(&ext->oe_pages));
 486         osc_extent_erase(ext);
 487         list_del_init(&ext->oe_link);
 488         osc_extent_state_set(ext, OES_INV);
 489         OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
 490 }
 491
 492 static void osc_extent_remove(struct osc_extent *ext)
 493 {
 494         struct osc_object *obj = ext->oe_obj;
 495
 496         osc_object_lock(obj);
 497         __osc_extent_remove(ext);
 498         osc_object_unlock(obj);
 499 }
 500
 501 /**
 502  * This function is used to merge extents to get better performance. It checks
 503  * if @cur and @victim are contiguous at block level.
 504  */
 505 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
 506                             struct osc_extent *victim)
 507 {
 508         struct osc_object       *obj = cur->oe_obj;
 509         struct client_obd       *cli = osc_cli(obj);
 510         pgoff_t                  chunk_start;
 511         pgoff_t                  chunk_end;
 512         int                      ppc_bits;
 513
 514         LASSERT(cur->oe_state == OES_CACHE);
 515         LASSERT(osc_object_is_locked(obj));
 516         if (victim == NULL)
 517                 return -EINVAL;
 518
 519         if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
 520                 return -EBUSY;
 521
 522         if (cur->oe_max_end != victim->oe_max_end)
 523                 return -ERANGE;
 524
 525         LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
 526         ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_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_hp)
 599                                 list_move_tail(&ext->oe_link,
 600                                                &obj->oo_hp_exts);
 601                         else if (ext->oe_urgent)
 602                                 list_move_tail(&ext->oe_link,
 603                                                &obj->oo_urgent_exts);
 604                         else if (ext->oe_nr_pages == ext->oe_mppr) {
 605                                 list_move_tail(&ext->oe_link,
 606                                                &obj->oo_full_exts);
 607                         }
 608                 }
 609                 osc_object_unlock(obj);
 610
 611                 osc_io_unplug_async(env, cli, obj);
 612         }
 613         osc_extent_put(env, ext);
 614         RETURN(rc);
 615 }
 616
 617 static inline int overlapped(struct osc_extent *ex1, struct osc_extent *ex2)
 618 {
 619         return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
 620 }
 621
 622 /**
 623  * Find or create an extent which includes @index, core function to manage
 624  * extent tree.
 625  */
 626 static struct osc_extent *osc_extent_find(const struct lu_env *env,
 627                                           struct osc_object *obj, pgoff_t index,
 628                                           unsigned int *grants)
 629 {
 630         struct client_obd *cli = osc_cli(obj);
 631         struct osc_lock   *olck;
 632         struct cl_lock_descr *descr;
 633         struct osc_extent *cur;
 634         struct osc_extent *ext;
 635         struct osc_extent *conflict = NULL;
 636         struct osc_extent *found = NULL;
 637         pgoff_t    chunk;
 638         pgoff_t    max_end;
 639         unsigned int max_pages; /* max_pages_per_rpc */
 640         unsigned int chunksize;
 641         int        ppc_bits; /* pages per chunk bits */
 642         pgoff_t    chunk_mask;
 643         int        rc;
 644         ENTRY;
 645
 646         cur = osc_extent_alloc(obj);
 647         if (cur == NULL)
 648                 RETURN(ERR_PTR(-ENOMEM));
 649
 650         olck = osc_env_io(env)->oi_write_osclock;
 651         LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
 652         LASSERT(olck->ols_state == OLS_GRANTED);
 653
 654         descr = &olck->ols_cl.cls_lock->cll_descr;
 655         LASSERT(descr->cld_mode >= CLM_WRITE);
 656
 657         LASSERTF(cli->cl_chunkbits >= PAGE_SHIFT,
 658                  "chunkbits: %u\n", cli->cl_chunkbits);
 659         ppc_bits   = cli->cl_chunkbits - PAGE_SHIFT;
 660         chunk_mask = ~((1 << ppc_bits) - 1);
 661         chunksize  = 1 << cli->cl_chunkbits;
 662         chunk      = index >> ppc_bits;
 663
 664         /* align end to RPC edge. */
 665         max_pages = cli->cl_max_pages_per_rpc;
 666         if ((max_pages & ~chunk_mask) != 0) {
 667                 CERROR("max_pages: %#x chunkbits: %u chunk_mask: %#lx\n",
 668                        max_pages, cli->cl_chunkbits, chunk_mask);
 669                 RETURN(ERR_PTR(-EINVAL));
 670         }
 671         max_end = index - (index % max_pages) + max_pages - 1;
 672         max_end = min_t(pgoff_t, max_end, descr->cld_end);
 673
 674         /* initialize new extent by parameters so far */
 675         cur->oe_max_end = max_end;
 676         cur->oe_start   = index & chunk_mask;
 677         cur->oe_end     = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
 678         if (cur->oe_start < descr->cld_start)
 679                 cur->oe_start = descr->cld_start;
 680         if (cur->oe_end > max_end)
 681                 cur->oe_end = max_end;
 682         cur->oe_grants  = 0;
 683         cur->oe_mppr    = max_pages;
 684         if (olck->ols_dlmlock != NULL) {
 685                 LASSERT(olck->ols_hold);
 686                 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
 687                 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
 688         }
 689
 690         /* grants has been allocated by caller */
 691         LASSERTF(*grants >= chunksize + cli->cl_grant_extent_tax,
 692                  "%u/%u/%u.\n", *grants, chunksize, cli->cl_grant_extent_tax);
 693         LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
 694                  EXTPARA(cur));
 695
 696 restart:
 697         osc_object_lock(obj);
 698         ext = osc_extent_search(obj, cur->oe_start);
 699         if (ext == NULL)
 700                 ext = first_extent(obj);
 701         while (ext != NULL) {
 702                 pgoff_t ext_chk_start = ext->oe_start >> ppc_bits;
 703                 pgoff_t ext_chk_end   = ext->oe_end   >> ppc_bits;
 704
 705                 LASSERT(sanity_check_nolock(ext) == 0);
 706                 if (chunk > ext_chk_end + 1 || chunk < ext_chk_start)
 707                         break;
 708
 709                 /* if covering by different locks, no chance to match */
 710                 if (olck->ols_dlmlock != ext->oe_dlmlock) {
 711                         EASSERTF(!overlapped(ext, cur), ext,
 712                                  EXTSTR"\n", EXTPARA(cur));
 713
 714                         ext = next_extent(ext);
 715                         continue;
 716                 }
 717
 718                 /* discontiguous chunks? */
 719                 if (chunk + 1 < ext_chk_start) {
 720                         ext = next_extent(ext);
 721                         continue;
 722                 }
 723
 724                 /* ok, from now on, ext and cur have these attrs:
 725                  * 1. covered by the same lock
 726                  * 2. contiguous at chunk level or overlapping. */
 727
 728                 if (overlapped(ext, cur)) {
 729                         /* cur is the minimum unit, so overlapping means
 730                          * full contain. */
 731                         EASSERTF((ext->oe_start <= cur->oe_start &&
 732                                   ext->oe_end >= cur->oe_end),
 733                                  ext, EXTSTR"\n", EXTPARA(cur));
 734
 735                         if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
 736                                 /* for simplicity, we wait for this extent to
 737                                  * finish before going forward. */
 738                                 conflict = osc_extent_get(ext);
 739                                 break;
 740                         }
 741
 742                         found = osc_extent_hold(ext);
 743                         break;
 744                 }
 745
 746                 /* non-overlapped extent */
 747                 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
 748                         /* we can't do anything for a non OES_CACHE extent, or
 749                          * if there is someone waiting for this extent to be
 750                          * flushed, try next one. */
 751                         ext = next_extent(ext);
 752                         continue;
 753                 }
 754
 755                 /* check if they belong to the same rpc slot before trying to
 756                  * merge. the extents are not overlapped and contiguous at
 757                  * chunk level to get here. */
 758                 if (ext->oe_max_end != max_end) {
 759                         /* if they don't belong to the same RPC slot or
 760                          * max_pages_per_rpc has ever changed, do not merge. */
 761                         ext = next_extent(ext);
 762                         continue;
 763                 }
 764
 765                 /* check whether maximum extent size will be hit */
 766                 if ((ext_chk_end - ext_chk_start + 1 + 1) << ppc_bits >
 767                     cli->cl_max_extent_pages) {
 768                         ext = next_extent(ext);
 769                         continue;
 770                 }
 771
 772                 /* it's required that an extent must be contiguous at chunk
 773                  * level so that we know the whole extent is covered by grant
 774                  * (the pages in the extent are NOT required to be contiguous).
 775                  * Otherwise, it will be too much difficult to know which
 776                  * chunks have grants allocated. */
 777
 778                 /* try to do front merge - extend ext's start */
 779                 if (chunk + 1 == ext_chk_start) {
 780                         /* ext must be chunk size aligned */
 781                         EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
 782
 783                         /* pull ext's start back to cover cur */
 784                         ext->oe_start   = cur->oe_start;
 785                         ext->oe_grants += chunksize;
 786                         LASSERT(*grants >= chunksize);
 787                         *grants -= chunksize;
 788
 789                         found = osc_extent_hold(ext);
 790                 } else if (chunk == ext_chk_end + 1) {
 791                         /* rear merge */
 792                         ext->oe_end     = cur->oe_end;
 793                         ext->oe_grants += chunksize;
 794                         LASSERT(*grants >= chunksize);
 795                         *grants -= chunksize;
 796
 797                         /* try to merge with the next one because we just fill
 798                          * in a gap */
 799                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 800                                 /* we can save extent tax from next extent */
 801                                 *grants += cli->cl_grant_extent_tax;
 802
 803                         found = osc_extent_hold(ext);
 804                 }
 805                 if (found != NULL)
 806                         break;
 807
 808                 ext = next_extent(ext);
 809         }
 810
 811         osc_extent_tree_dump(D_CACHE, obj);
 812         if (found != NULL) {
 813                 LASSERT(conflict == NULL);
 814                 if (!IS_ERR(found)) {
 815                         LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
 816                         OSC_EXTENT_DUMP(D_CACHE, found,
 817                                         "found caching ext for %lu.\n", index);
 818                 }
 819         } else if (conflict == NULL) {
 820                 /* create a new extent */
 821                 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
 822                 cur->oe_grants = chunksize + cli->cl_grant_extent_tax;
 823                 LASSERT(*grants >= cur->oe_grants);
 824                 *grants -= cur->oe_grants;
 825
 826                 cur->oe_state = OES_CACHE;
 827                 found = osc_extent_hold(cur);
 828                 osc_extent_insert(obj, cur);
 829                 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
 830                                 index, descr->cld_end);
 831         }
 832         osc_object_unlock(obj);
 833
 834         if (conflict != NULL) {
 835                 LASSERT(found == NULL);
 836
 837                 /* waiting for IO to finish. Please notice that it's impossible
 838                  * to be an OES_TRUNC extent. */
 839                 rc = osc_extent_wait(env, conflict, OES_INV);
 840                 osc_extent_put(env, conflict);
 841                 conflict = NULL;
 842                 if (rc < 0)
 843                         GOTO(out, found = ERR_PTR(rc));
 844
 845                 goto restart;
 846         }
 847         EXIT;
 848
 849 out:
 850         osc_extent_put(env, cur);
 851         return found;
 852 }
 853
 854 /**
 855  * Called when IO is finished to an extent.
 856  */
 857 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
 858                       int sent, int rc)
 859 {
 860         struct client_obd *cli = osc_cli(ext->oe_obj);
 861         struct osc_async_page *oap;
 862         struct osc_async_page *tmp;
 863         int nr_pages = ext->oe_nr_pages;
 864         int lost_grant = 0;
 865         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
 866         loff_t last_off = 0;
 867         int last_count = -1;
 868         ENTRY;
 869
 870         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
 871
 872         ext->oe_rc = rc ?: ext->oe_nr_pages;
 873         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
 874
 875         osc_lru_add_batch(cli, &ext->oe_pages);
 876         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 877                                      oap_pending_item) {
 878                 list_del_init(&oap->oap_rpc_item);
 879                 list_del_init(&oap->oap_pending_item);
 880                 if (last_off <= oap->oap_obj_off) {
 881                         last_off = oap->oap_obj_off;
 882                         last_count = oap->oap_count;
 883                 }
 884
 885                 --ext->oe_nr_pages;
 886                 osc_ap_completion(env, cli, oap, sent, rc);
 887         }
 888         EASSERT(ext->oe_nr_pages == 0, ext);
 889
 890         if (!sent) {
 891                 lost_grant = ext->oe_grants;
 892         } else if (blocksize < PAGE_SIZE &&
 893                    last_count != PAGE_SIZE) {
 894                 /* For short writes we shouldn't count parts of pages that
 895                  * span a whole chunk on the OST side, or our accounting goes
 896                  * wrong.  Should match the code in filter_grant_check. */
 897                 int offset = last_off & ~PAGE_MASK;
 898                 int count = last_count + (offset & (blocksize - 1));
 899                 int end = (offset + last_count) & (blocksize - 1);
 900                 if (end)
 901                         count += blocksize - end;
 902
 903                 lost_grant = PAGE_SIZE - count;
 904         }
 905         if (ext->oe_grants > 0)
 906                 osc_free_grant(cli, nr_pages, lost_grant, ext->oe_grants);
 907
 908         osc_extent_remove(ext);
 909         /* put the refcount for RPC */
 910         osc_extent_put(env, ext);
 911         RETURN(0);
 912 }
 913
 914 static int extent_wait_cb(struct osc_extent *ext, enum osc_extent_state state)
 915 {
 916         int ret;
 917
 918         osc_object_lock(ext->oe_obj);
 919         ret = ext->oe_state == state;
 920         osc_object_unlock(ext->oe_obj);
 921
 922         return ret;
 923 }
 924
 925 /**
 926  * Wait for the extent's state to become @state.
 927  */
 928 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
 929                            enum osc_extent_state state)
 930 {
 931         struct osc_object *obj = ext->oe_obj;
 932         struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
 933                                                   LWI_ON_SIGNAL_NOOP, NULL);
 934         int rc = 0;
 935         ENTRY;
 936
 937         osc_object_lock(obj);
 938         LASSERT(sanity_check_nolock(ext) == 0);
 939         /* `Kick' this extent only if the caller is waiting for it to be
 940          * written out. */
 941         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
 942                 if (ext->oe_state == OES_ACTIVE) {
 943                         ext->oe_urgent = 1;
 944                 } else if (ext->oe_state == OES_CACHE) {
 945                         ext->oe_urgent = 1;
 946                         osc_extent_hold(ext);
 947                         rc = 1;
 948                 }
 949         }
 950         osc_object_unlock(obj);
 951         if (rc == 1)
 952                 osc_extent_release(env, ext);
 953
 954         /* wait for the extent until its state becomes @state */
 955         rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
 956         if (rc == -ETIMEDOUT) {
 957                 OSC_EXTENT_DUMP(D_ERROR, ext,
 958                         "%s: wait ext to %u timedout, recovery in progress?\n",
 959                         cli_name(osc_cli(obj)), state);
 960
 961                 lwi = LWI_INTR(NULL, NULL);
 962                 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state),
 963                                   &lwi);
 964         }
 965         if (rc == 0 && ext->oe_rc < 0)
 966                 rc = ext->oe_rc;
 967         RETURN(rc);
 968 }
 969
 970 /**
 971  * Discard pages with index greater than @size. If @ext is overlapped with
 972  * @size, then partial truncate happens.
 973  */
 974 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
 975                                 bool partial)
 976 {
 977         struct lu_env         *env;
 978         struct cl_io          *io;
 979         struct osc_object     *obj = ext->oe_obj;
 980         struct client_obd     *cli = osc_cli(obj);
 981         struct osc_async_page *oap;
 982         struct osc_async_page *tmp;
 983         struct pagevec        *pvec;
 984         int                    pages_in_chunk = 0;
 985         int                    ppc_bits    = cli->cl_chunkbits -
 986                                              PAGE_SHIFT;
 987         __u64                  trunc_chunk = trunc_index >> ppc_bits;
 988         int                    grants   = 0;
 989         int                    nr_pages = 0;
 990         int                    rc       = 0;
 991         __u16                  refcheck;
 992         ENTRY;
 993
 994         LASSERT(sanity_check(ext) == 0);
 995         LASSERT(ext->oe_state == OES_TRUNC);
 996         LASSERT(!ext->oe_urgent);
 997
 998         /* Request new lu_env.
 999          * We can't use that env from osc_cache_truncate_start() because
1000          * it's from lov_io_sub and not fully initialized. */
1001         env = cl_env_get(&refcheck);
1002         if (IS_ERR(env))
1003                 RETURN(PTR_ERR(env));
1004
1005         io  = osc_env_thread_io(env);
1006         io->ci_obj = cl_object_top(osc2cl(obj));
1007         io->ci_ignore_layout = 1;
1008         pvec = &osc_env_info(env)->oti_pagevec;
1009         ll_pagevec_init(pvec, 0);
1010         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
1011         if (rc < 0)
1012                 GOTO(out, rc);
1013
1014         /* discard all pages with index greater than trunc_index */
1015         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
1016                                      oap_pending_item) {
1017                 pgoff_t index = osc_index(oap2osc(oap));
1018                 struct cl_page  *page = oap2cl_page(oap);
1019
1020                 LASSERT(list_empty(&oap->oap_rpc_item));
1021
1022                 /* only discard the pages with their index greater than
1023                  * trunc_index, and ... */
1024                 if (index < trunc_index ||
1025                     (index == trunc_index && partial)) {
1026                         /* accounting how many pages remaining in the chunk
1027                          * so that we can calculate grants correctly. */
1028                         if (index >> ppc_bits == trunc_chunk)
1029                                 ++pages_in_chunk;
1030                         continue;
1031                 }
1032
1033                 list_del_init(&oap->oap_pending_item);
1034
1035                 cl_page_get(page);
1036                 lu_ref_add(&page->cp_reference, "truncate", current);
1037
1038                 if (cl_page_own(env, io, page) == 0) {
1039                         cl_page_discard(env, io, page);
1040                         cl_page_disown(env, io, page);
1041                 } else {
1042                         LASSERT(page->cp_state == CPS_FREEING);
1043                         LASSERT(0);
1044                 }
1045
1046                 lu_ref_del(&page->cp_reference, "truncate", current);
1047                 cl_pagevec_put(env, page, pvec);
1048
1049                 --ext->oe_nr_pages;
1050                 ++nr_pages;
1051         }
1052         pagevec_release(pvec);
1053
1054         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1055                       ext->oe_nr_pages == 0),
1056                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1057
1058         osc_object_lock(obj);
1059         if (ext->oe_nr_pages == 0) {
1060                 LASSERT(pages_in_chunk == 0);
1061                 grants = ext->oe_grants;
1062                 ext->oe_grants = 0;
1063         } else { /* calculate how many grants we can free */
1064                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1065                 pgoff_t last_index;
1066
1067
1068                 /* if there is no pages in this chunk, we can also free grants
1069                  * for the last chunk */
1070                 if (pages_in_chunk == 0) {
1071                         /* if this is the 1st chunk and no pages in this chunk,
1072                          * ext->oe_nr_pages must be zero, so we should be in
1073                          * the other if-clause. */
1074                         LASSERT(trunc_chunk > 0);
1075                         --trunc_chunk;
1076                         ++chunks;
1077                 }
1078
1079                 /* this is what we can free from this extent */
1080                 grants          = chunks << cli->cl_chunkbits;
1081                 ext->oe_grants -= grants;
1082                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1083                 ext->oe_end     = min(last_index, ext->oe_max_end);
1084                 LASSERT(ext->oe_end >= ext->oe_start);
1085                 LASSERT(ext->oe_grants > 0);
1086         }
1087         osc_object_unlock(obj);
1088
1089         if (grants > 0 || nr_pages > 0)
1090                 osc_free_grant(cli, nr_pages, grants, grants);
1091
1092 out:
1093         cl_io_fini(env, io);
1094         cl_env_put(env, &refcheck);
1095         RETURN(rc);
1096 }
1097
1098 /**
1099  * This function is used to make the extent prepared for transfer.
1100  * A race with flusing page - ll_writepage() has to be handled cautiously.
1101  */
1102 static int osc_extent_make_ready(const struct lu_env *env,
1103                                  struct osc_extent *ext)
1104 {
1105         struct osc_async_page *oap;
1106         struct osc_async_page *last = NULL;
1107         struct osc_object *obj = ext->oe_obj;
1108         unsigned int page_count = 0;
1109         int rc;
1110         ENTRY;
1111
1112         /* we're going to grab page lock, so object lock must not be taken. */
1113         LASSERT(sanity_check(ext) == 0);
1114         /* in locking state, any process should not touch this extent. */
1115         EASSERT(ext->oe_state == OES_LOCKING, ext);
1116         EASSERT(ext->oe_owner != NULL, ext);
1117
1118         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1119
1120         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1121                 ++page_count;
1122                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1123                         last = oap;
1124
1125                 /* checking ASYNC_READY is race safe */
1126                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1127                         continue;
1128
1129                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1130                 switch (rc) {
1131                 case 0:
1132                         spin_lock(&oap->oap_lock);
1133                         oap->oap_async_flags |= ASYNC_READY;
1134                         spin_unlock(&oap->oap_lock);
1135                         break;
1136                 case -EALREADY:
1137                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1138                         break;
1139                 default:
1140                         LASSERTF(0, "unknown return code: %d\n", rc);
1141                 }
1142         }
1143
1144         LASSERT(page_count == ext->oe_nr_pages);
1145         LASSERT(last != NULL);
1146         /* the last page is the only one we need to refresh its count by
1147          * the size of file. */
1148         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1149                 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1150                 LASSERT(last_oap_count > 0);
1151                 LASSERT(last->oap_page_off + last_oap_count <= PAGE_SIZE);
1152                 last->oap_count = last_oap_count;
1153                 spin_lock(&last->oap_lock);
1154                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1155                 spin_unlock(&last->oap_lock);
1156         }
1157
1158         /* for the rest of pages, we don't need to call osf_refresh_count()
1159          * because it's known they are not the last page */
1160         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1161                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1162                         oap->oap_count = PAGE_SIZE - oap->oap_page_off;
1163                         spin_lock(&oap->oap_lock);
1164                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1165                         spin_unlock(&oap->oap_lock);
1166                 }
1167         }
1168
1169         osc_object_lock(obj);
1170         osc_extent_state_set(ext, OES_RPC);
1171         osc_object_unlock(obj);
1172         /* get a refcount for RPC. */
1173         osc_extent_get(ext);
1174
1175         RETURN(0);
1176 }
1177
1178 /**
1179  * Quick and simple version of osc_extent_find(). This function is frequently
1180  * called to expand the extent for the same IO. To expand the extent, the
1181  * page index must be in the same or next chunk of ext->oe_end.
1182  */
1183 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index,
1184                              unsigned int *grants)
1185 {
1186         struct osc_object *obj = ext->oe_obj;
1187         struct client_obd *cli = osc_cli(obj);
1188         struct osc_extent *next;
1189         int ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1190         pgoff_t chunk = index >> ppc_bits;
1191         pgoff_t end_chunk;
1192         pgoff_t end_index;
1193         unsigned int chunksize = 1 << cli->cl_chunkbits;
1194         int rc = 0;
1195         ENTRY;
1196
1197         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1198         osc_object_lock(obj);
1199         LASSERT(sanity_check_nolock(ext) == 0);
1200         end_chunk = ext->oe_end >> ppc_bits;
1201         if (chunk > end_chunk + 1)
1202                 GOTO(out, rc = -ERANGE);
1203
1204         if (end_chunk >= chunk)
1205                 GOTO(out, rc = 0);
1206
1207         LASSERT(end_chunk + 1 == chunk);
1208
1209         /* try to expand this extent to cover @index */
1210         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1211
1212         /* don't go over the maximum extent size reported by server */
1213         if (end_index - ext->oe_start + 1 > cli->cl_max_extent_pages)
1214                 GOTO(out, rc = -ERANGE);
1215
1216         next = next_extent(ext);
1217         if (next != NULL && next->oe_start <= end_index)
1218                 /* complex mode - overlapped with the next extent,
1219                  * this case will be handled by osc_extent_find() */
1220                 GOTO(out, rc = -EAGAIN);
1221
1222         ext->oe_end = end_index;
1223         ext->oe_grants += chunksize;
1224         LASSERT(*grants >= chunksize);
1225         *grants -= chunksize;
1226         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1227                  "overlapped after expanding for %lu.\n", index);
1228         EXIT;
1229
1230 out:
1231         osc_object_unlock(obj);
1232         RETURN(rc);
1233 }
1234
1235 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
1236                                   const char *func, int line)
1237 {
1238         struct osc_extent *ext;
1239         int cnt;
1240
1241         if (!cfs_cdebug_show(level, DEBUG_SUBSYSTEM))
1242                 return;
1243
1244         CDEBUG(level, "Dump object %p extents at %s:%d, mppr: %u.\n",
1245                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1246
1247         /* osc_object_lock(obj); */
1248         cnt = 1;
1249         for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1250                 OSC_EXTENT_DUMP(level, ext, "in tree %d.\n", cnt++);
1251
1252         cnt = 1;
1253         list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1254                 OSC_EXTENT_DUMP(level, ext, "hp %d.\n", cnt++);
1255
1256         cnt = 1;
1257         list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1258                 OSC_EXTENT_DUMP(level, ext, "urgent %d.\n", cnt++);
1259
1260         cnt = 1;
1261         list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1262                 OSC_EXTENT_DUMP(level, ext, "reading %d.\n", cnt++);
1263         /* osc_object_unlock(obj); */
1264 }
1265
1266 /* ------------------ osc extent end ------------------ */
1267
1268 static inline int osc_is_ready(struct osc_object *osc)
1269 {
1270         return !list_empty(&osc->oo_ready_item) ||
1271                !list_empty(&osc->oo_hp_ready_item);
1272 }
1273
1274 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1275         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1276                (OSC), osc_is_ready(OSC),                                       \
1277                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1278                list_empty_marker(&(OSC)->oo_ready_item),                       \
1279                atomic_read(&(OSC)->oo_nr_writes),                              \
1280                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1281                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1282                atomic_read(&(OSC)->oo_nr_reads),                               \
1283                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1284                ##args)
1285
1286 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1287                           int cmd)
1288 {
1289         struct osc_page *opg  = oap2osc_page(oap);
1290         struct cl_page  *page = oap2cl_page(oap);
1291         int result;
1292
1293         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1294
1295         ENTRY;
1296         result = cl_page_make_ready(env, page, CRT_WRITE);
1297         if (result == 0)
1298                 opg->ops_submit_time = ktime_get();
1299         RETURN(result);
1300 }
1301
1302 static int osc_refresh_count(const struct lu_env *env,
1303                              struct osc_async_page *oap, int cmd)
1304 {
1305         struct osc_page  *opg = oap2osc_page(oap);
1306         pgoff_t index = osc_index(oap2osc(oap));
1307         struct cl_object *obj;
1308         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
1309         int result;
1310         loff_t kms;
1311
1312         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1313         LASSERT(!(cmd & OBD_BRW_READ));
1314         LASSERT(opg != NULL);
1315         obj = opg->ops_cl.cpl_obj;
1316
1317         cl_object_attr_lock(obj);
1318         result = cl_object_attr_get(env, obj, attr);
1319         cl_object_attr_unlock(obj);
1320         if (result < 0)
1321                 return result;
1322         kms = attr->cat_kms;
1323         if (cl_offset(obj, index) >= kms)
1324                 /* catch race with truncate */
1325                 return 0;
1326         else if (cl_offset(obj, index + 1) > kms)
1327                 /* catch sub-page write at end of file */
1328                 return kms % PAGE_SIZE;
1329         else
1330                 return PAGE_SIZE;
1331 }
1332
1333 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1334                           int cmd, int rc)
1335 {
1336         struct osc_page   *opg  = oap2osc_page(oap);
1337         struct cl_page    *page = oap2cl_page(oap);
1338         enum cl_req_type   crt;
1339         int srvlock;
1340
1341         ENTRY;
1342
1343         cmd &= ~OBD_BRW_NOQUOTA;
1344         LASSERTF(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ),
1345                  "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1346         LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1347                 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1348         LASSERT(opg->ops_transfer_pinned);
1349
1350         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1351         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1352         opg->ops_transfer_pinned = 0;
1353
1354         opg->ops_submit_time = ktime_set(0, 0);
1355         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1356
1357         /* statistic */
1358         if (rc == 0 && srvlock) {
1359                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1360                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1361                 size_t bytes = oap->oap_count;
1362
1363                 if (crt == CRT_READ)
1364                         stats->os_lockless_reads += bytes;
1365                 else
1366                         stats->os_lockless_writes += bytes;
1367         }
1368
1369         /*
1370          * This has to be the last operation with the page, as locks are
1371          * released in cl_page_completion() and nothing except for the
1372          * reference counter protects page from concurrent reclaim.
1373          */
1374         lu_ref_del(&page->cp_reference, "transfer", page);
1375
1376         cl_page_completion(env, page, crt, rc);
1377         cl_page_put(env, page);
1378
1379         RETURN(0);
1380 }
1381
1382 #define OSC_DUMP_GRANT(lvl, cli, fmt, args...) do {                     \
1383         struct client_obd *__tmp = (cli);                               \
1384         CDEBUG(lvl, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu "  \
1385                "dropped: %ld avail: %ld, dirty_grant: %ld, "            \
1386                "reserved: %ld, flight: %d } lru {in list: %ld, "        \
1387                "left: %ld, waiters: %d }" fmt "\n",                     \
1388                cli_name(__tmp),                                         \
1389                __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages,        \
1390                atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1391                __tmp->cl_lost_grant, __tmp->cl_avail_grant,             \
1392                __tmp->cl_dirty_grant,                                   \
1393                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,         \
1394                atomic_long_read(&__tmp->cl_lru_in_list),                \
1395                atomic_long_read(&__tmp->cl_lru_busy),                   \
1396                atomic_read(&__tmp->cl_lru_shrinkers), ##args);          \
1397 } while (0)
1398
1399 /* caller must hold loi_list_lock */
1400 static void osc_consume_write_grant(struct client_obd *cli,
1401                                     struct brw_page *pga)
1402 {
1403         assert_spin_locked(&cli->cl_loi_list_lock);
1404         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1405         atomic_long_inc(&obd_dirty_pages);
1406         cli->cl_dirty_pages++;
1407         pga->flag |= OBD_BRW_FROM_GRANT;
1408         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1409                PAGE_SIZE, pga, pga->pg);
1410         osc_update_next_shrink(cli);
1411 }
1412
1413 /* the companion to osc_consume_write_grant, called when a brw has completed.
1414  * must be called with the loi lock held. */
1415 static void osc_release_write_grant(struct client_obd *cli,
1416                                     struct brw_page *pga)
1417 {
1418         ENTRY;
1419
1420         assert_spin_locked(&cli->cl_loi_list_lock);
1421         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1422                 EXIT;
1423                 return;
1424         }
1425
1426         pga->flag &= ~OBD_BRW_FROM_GRANT;
1427         atomic_long_dec(&obd_dirty_pages);
1428         cli->cl_dirty_pages--;
1429         if (pga->flag & OBD_BRW_NOCACHE) {
1430                 pga->flag &= ~OBD_BRW_NOCACHE;
1431                 atomic_long_dec(&obd_dirty_transit_pages);
1432                 cli->cl_dirty_transit--;
1433         }
1434         EXIT;
1435 }
1436
1437 /**
1438  * To avoid sleeping with object lock held, it's good for us allocate enough
1439  * grants before entering into critical section.
1440  *
1441  * client_obd_list_lock held by caller
1442  */
1443 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1444 {
1445         int rc = -EDQUOT;
1446
1447         if (cli->cl_avail_grant >= bytes) {
1448                 cli->cl_avail_grant    -= bytes;
1449                 cli->cl_reserved_grant += bytes;
1450                 rc = 0;
1451         }
1452         return rc;
1453 }
1454
1455 static void __osc_unreserve_grant(struct client_obd *cli,
1456                                   unsigned int reserved, unsigned int unused)
1457 {
1458         /* it's quite normal for us to get more grant than reserved.
1459          * Thinking about a case that two extents merged by adding a new
1460          * chunk, we can save one extent tax. If extent tax is greater than
1461          * one chunk, we can save more grant by adding a new chunk */
1462         cli->cl_reserved_grant -= reserved;
1463         if (unused > reserved) {
1464                 cli->cl_avail_grant += reserved;
1465                 cli->cl_lost_grant  += unused - reserved;
1466                 cli->cl_dirty_grant -= unused - reserved;
1467         } else {
1468                 cli->cl_avail_grant += unused;
1469                 cli->cl_dirty_grant += reserved - unused;
1470         }
1471 }
1472
1473 static void osc_unreserve_grant(struct client_obd *cli,
1474                                 unsigned int reserved, unsigned int unused)
1475 {
1476         spin_lock(&cli->cl_loi_list_lock);
1477         __osc_unreserve_grant(cli, reserved, unused);
1478         if (unused > 0)
1479                 osc_wake_cache_waiters(cli);
1480         spin_unlock(&cli->cl_loi_list_lock);
1481 }
1482
1483 /**
1484  * Free grant after IO is finished or canceled.
1485  *
1486  * @lost_grant is used to remember how many grants we have allocated but not
1487  * used, we should return these grants to OST. There're two cases where grants
1488  * can be lost:
1489  * 1. truncate;
1490  * 2. blocksize at OST is less than PAGE_SIZE and a partial page was
1491  *    written. In this case OST may use less chunks to serve this partial
1492  *    write. OSTs don't actually know the page size on the client side. so
1493  *    clients have to calculate lost grant by the blocksize on the OST.
1494  *    See filter_grant_check() for details.
1495  */
1496 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1497                            unsigned int lost_grant, unsigned int dirty_grant)
1498 {
1499         unsigned long grant;
1500
1501         grant = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
1502
1503         spin_lock(&cli->cl_loi_list_lock);
1504         atomic_long_sub(nr_pages, &obd_dirty_pages);
1505         cli->cl_dirty_pages -= nr_pages;
1506         cli->cl_lost_grant += lost_grant;
1507         cli->cl_dirty_grant -= dirty_grant;
1508         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1509                 /* borrow some grant from truncate to avoid the case that
1510                  * truncate uses up all avail grant */
1511                 cli->cl_lost_grant -= grant;
1512                 cli->cl_avail_grant += grant;
1513         }
1514         osc_wake_cache_waiters(cli);
1515         spin_unlock(&cli->cl_loi_list_lock);
1516         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu/%lu\n",
1517                lost_grant, cli->cl_lost_grant,
1518                cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_SHIFT,
1519                cli->cl_dirty_grant);
1520 }
1521
1522 /**
1523  * The companion to osc_enter_cache(), called when @oap is no longer part of
1524  * the dirty accounting due to error.
1525  */
1526 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1527 {
1528         spin_lock(&cli->cl_loi_list_lock);
1529         osc_release_write_grant(cli, &oap->oap_brw_page);
1530         spin_unlock(&cli->cl_loi_list_lock);
1531 }
1532
1533 /**
1534  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1535  * is available.
1536  */
1537 static int osc_enter_cache_try(struct client_obd *cli,
1538                                struct osc_async_page *oap,
1539                                int bytes, int transient)
1540 {
1541         int rc;
1542
1543         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1544
1545         rc = osc_reserve_grant(cli, bytes);
1546         if (rc < 0)
1547                 return 0;
1548
1549         if (cli->cl_dirty_pages < cli->cl_dirty_max_pages &&
1550             1 + atomic_long_read(&obd_dirty_pages) <= obd_max_dirty_pages) {
1551                 osc_consume_write_grant(cli, &oap->oap_brw_page);
1552                 if (transient) {
1553                         cli->cl_dirty_transit++;
1554                         atomic_long_inc(&obd_dirty_transit_pages);
1555                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
1556                 }
1557                 rc = 1;
1558         } else {
1559                 __osc_unreserve_grant(cli, bytes, bytes);
1560                 rc = 0;
1561         }
1562         return rc;
1563 }
1564
1565 /* Following two inlines exist to pass code fragments
1566  * to wait_event_idle_exclusive_timeout_cmd().  Passing
1567  * code fragments as macro args can look confusing, so
1568  * we provide inlines to encapsulate them.
1569  */
1570 static inline void cli_unlock_and_unplug(const struct lu_env *env,
1571                                          struct client_obd *cli,
1572                                          struct osc_async_page *oap)
1573 {
1574         spin_unlock(&cli->cl_loi_list_lock);
1575         osc_io_unplug_async(env, cli, NULL);
1576         CDEBUG(D_CACHE,
1577                "%s: sleeping for cache space for %p\n",
1578                cli_name(cli), oap);
1579 }
1580
1581 static inline void cli_lock_after_unplug(struct client_obd *cli)
1582 {
1583         spin_lock(&cli->cl_loi_list_lock);
1584 }
1585 /**
1586  * The main entry to reserve dirty page accounting. Usually the grant reserved
1587  * in this function will be freed in bulk in osc_free_grant() unless it fails
1588  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1589  *
1590  * The process will be put into sleep if it's already run out of grant.
1591  */
1592 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1593                            struct osc_async_page *oap, int bytes)
1594 {
1595         struct osc_object *osc = oap->oap_obj;
1596         struct lov_oinfo *loi = osc->oo_oinfo;
1597         int rc = -EDQUOT;
1598         int remain;
1599         bool entered = false;
1600         /* We cannot wait for a long time here since we are holding ldlm lock
1601          * across the actual IO. If no requests complete fast (e.g. due to
1602          * overloaded OST that takes a long time to process everything, we'd
1603          * get evicted if we wait for a normal obd_timeout or some such.
1604          * So we try to wait half the time it would take the client to be
1605          * evicted by server which is half obd_timeout when AT is off
1606          * or at least ldlm_enqueue_min with AT on.
1607          * See LU-13131 */
1608         unsigned long timeout = cfs_time_seconds(AT_OFF ? obd_timeout / 2 :
1609                                                           ldlm_enqueue_min / 2);
1610
1611         ENTRY;
1612
1613         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1614
1615         spin_lock(&cli->cl_loi_list_lock);
1616
1617         /* force the caller to try sync io.  this can jump the list
1618          * of queued writes and create a discontiguous rpc stream */
1619         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1620             cli->cl_dirty_max_pages == 0 ||
1621             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) {
1622                 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1623                 GOTO(out, rc = -EDQUOT);
1624         }
1625
1626         /*
1627          * We can wait here for two reasons: too many dirty pages in cache, or
1628          * run out of grants. In both cases we should write dirty pages out.
1629          * Adding a cache waiter will trigger urgent write-out no matter what
1630          * RPC size will be.
1631          * The exiting condition (other than success) is no avail grants
1632          * and no dirty pages caching, that really means there is no space
1633          * on the OST.
1634          */
1635         remain = wait_event_idle_exclusive_timeout_cmd(
1636                 cli->cl_cache_waiters,
1637                 (entered = osc_enter_cache_try(cli, oap, bytes, 0)) ||
1638                 (cli->cl_dirty_pages == 0 && cli->cl_w_in_flight == 0),
1639                 timeout,
1640                 cli_unlock_and_unplug(env, cli, oap),
1641                 cli_lock_after_unplug(cli));
1642
1643         if (entered) {
1644                 if (remain == timeout)
1645                         OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1646                 else
1647                         OSC_DUMP_GRANT(D_CACHE, cli,
1648                                        "finally got grant space\n");
1649                 wake_up(&cli->cl_cache_waiters);
1650                 rc = 0;
1651         } else if (remain == 0) {
1652                 OSC_DUMP_GRANT(D_CACHE, cli,
1653                                "timeout, fall back to sync i/o\n");
1654                 osc_extent_tree_dump(D_CACHE, osc);
1655                 /* fall back to synchronous I/O */
1656         } else {
1657                 OSC_DUMP_GRANT(D_CACHE, cli,
1658                                "no grant space, fall back to sync i/o\n");
1659                 wake_up_all(&cli->cl_cache_waiters);
1660         }
1661         EXIT;
1662 out:
1663         spin_unlock(&cli->cl_loi_list_lock);
1664         RETURN(rc);
1665 }
1666
1667 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1668 {
1669         int hprpc = !!list_empty(&osc->oo_hp_exts);
1670         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1671 }
1672
1673 /* This maintains the lists of pending pages to read/write for a given object
1674  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1675  * to quickly find objects that are ready to send an RPC. */
1676 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1677                          int cmd)
1678 {
1679         int invalid_import = 0;
1680         ENTRY;
1681
1682         /* if we have an invalid import we want to drain the queued pages
1683          * by forcing them through rpcs that immediately fail and complete
1684          * the pages.  recovery relies on this to empty the queued pages
1685          * before canceling the locks and evicting down the llite pages */
1686         if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1687                 invalid_import = 1;
1688
1689         if (cmd & OBD_BRW_WRITE) {
1690                 if (atomic_read(&osc->oo_nr_writes) == 0)
1691                         RETURN(0);
1692                 if (invalid_import) {
1693                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1694                         RETURN(1);
1695                 }
1696                 if (!list_empty(&osc->oo_hp_exts)) {
1697                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1698                         RETURN(1);
1699                 }
1700                 if (!list_empty(&osc->oo_urgent_exts)) {
1701                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1702                         RETURN(1);
1703                 }
1704                 /* trigger a write rpc stream as long as there are dirtiers
1705                  * waiting for space.  as they're waiting, they're not going to
1706                  * create more pages to coalesce with what's waiting..
1707                  */
1708                 if (waitqueue_active(&cli->cl_cache_waiters)) {
1709                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1710                         RETURN(1);
1711                 }
1712                 if (!list_empty(&osc->oo_full_exts)) {
1713                         CDEBUG(D_CACHE, "full extent ready, make an RPC\n");
1714                         RETURN(1);
1715                 }
1716         } else {
1717                 if (atomic_read(&osc->oo_nr_reads) == 0)
1718                         RETURN(0);
1719                 if (invalid_import) {
1720                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1721                         RETURN(1);
1722                 }
1723                 /* all read are urgent. */
1724                 if (!list_empty(&osc->oo_reading_exts))
1725                         RETURN(1);
1726         }
1727
1728         RETURN(0);
1729 }
1730
1731 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1732 {
1733         struct client_obd *cli = osc_cli(obj);
1734         if (cmd & OBD_BRW_WRITE) {
1735                 atomic_add(delta, &obj->oo_nr_writes);
1736                 atomic_add(delta, &cli->cl_pending_w_pages);
1737                 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1738         } else {
1739                 atomic_add(delta, &obj->oo_nr_reads);
1740                 atomic_add(delta, &cli->cl_pending_r_pages);
1741                 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1742         }
1743         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1744 }
1745
1746 static int osc_makes_hprpc(struct osc_object *obj)
1747 {
1748         return !list_empty(&obj->oo_hp_exts);
1749 }
1750
1751 static void on_list(struct list_head *item, struct list_head *list,
1752                     int should_be_on)
1753 {
1754         if (list_empty(item) && should_be_on)
1755                 list_add_tail(item, list);
1756         else if (!list_empty(item) && !should_be_on)
1757                 list_del_init(item);
1758 }
1759
1760 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1761  * can find pages to build into rpcs quickly */
1762 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1763 {
1764         if (osc_makes_hprpc(osc)) {
1765                 /* HP rpc */
1766                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1767                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1768         } else {
1769                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1770                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1771                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1772                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1773         }
1774
1775         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1776                 atomic_read(&osc->oo_nr_writes) > 0);
1777
1778         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1779                 atomic_read(&osc->oo_nr_reads) > 0);
1780
1781         return osc_is_ready(osc);
1782 }
1783
1784 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1785 {
1786         int is_ready;
1787
1788         spin_lock(&cli->cl_loi_list_lock);
1789         is_ready = __osc_list_maint(cli, osc);
1790         spin_unlock(&cli->cl_loi_list_lock);
1791
1792         return is_ready;
1793 }
1794
1795 /* this is trying to propogate async writeback errors back up to the
1796  * application.  As an async write fails we record the error code for later if
1797  * the app does an fsync.  As long as errors persist we force future rpcs to be
1798  * sync so that the app can get a sync error and break the cycle of queueing
1799  * pages for which writeback will fail. */
1800 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1801                            int rc)
1802 {
1803         if (rc) {
1804                 if (!ar->ar_rc)
1805                         ar->ar_rc = rc;
1806
1807                 ar->ar_force_sync = 1;
1808                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1809                 return;
1810
1811         }
1812
1813         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1814                 ar->ar_force_sync = 0;
1815 }
1816
1817 /* this must be called holding the loi list lock to give coverage to exit_cache,
1818  * async_flag maintenance, and oap_request */
1819 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1820                               struct osc_async_page *oap, int sent, int rc)
1821 {
1822         struct osc_object *osc = oap->oap_obj;
1823         struct lov_oinfo  *loi = osc->oo_oinfo;
1824         __u64 xid = 0;
1825
1826         ENTRY;
1827         if (oap->oap_request != NULL) {
1828                 xid = ptlrpc_req_xid(oap->oap_request);
1829                 ptlrpc_req_finished(oap->oap_request);
1830                 oap->oap_request = NULL;
1831         }
1832
1833         /* As the transfer for this page is being done, clear the flags */
1834         spin_lock(&oap->oap_lock);
1835         oap->oap_async_flags = 0;
1836         spin_unlock(&oap->oap_lock);
1837         oap->oap_interrupted = 0;
1838
1839         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1840                 spin_lock(&cli->cl_loi_list_lock);
1841                 osc_process_ar(&cli->cl_ar, xid, rc);
1842                 osc_process_ar(&loi->loi_ar, xid, rc);
1843                 spin_unlock(&cli->cl_loi_list_lock);
1844         }
1845
1846         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1847         if (rc)
1848                 CERROR("completion on oap %p obj %p returns %d.\n",
1849                        oap, osc, rc);
1850
1851         EXIT;
1852 }
1853
1854 struct extent_rpc_data {
1855         struct list_head        *erd_rpc_list;
1856         unsigned int            erd_page_count;
1857         unsigned int            erd_max_pages;
1858         unsigned int            erd_max_chunks;
1859         unsigned int            erd_max_extents;
1860 };
1861
1862 static inline unsigned osc_extent_chunks(const struct osc_extent *ext)
1863 {
1864         struct client_obd *cli = osc_cli(ext->oe_obj);
1865         unsigned ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1866
1867         return (ext->oe_end >> ppc_bits) - (ext->oe_start >> ppc_bits) + 1;
1868 }
1869
1870 /**
1871  * Try to add extent to one RPC. We need to think about the following things:
1872  * - # of pages must not be over max_pages_per_rpc
1873  * - extent must be compatible with previous ones
1874  */
1875 static int try_to_add_extent_for_io(struct client_obd *cli,
1876                                     struct osc_extent *ext,
1877                                     struct extent_rpc_data *data)
1878 {
1879         struct osc_extent *tmp;
1880         unsigned int chunk_count;
1881         struct osc_async_page *oap = list_first_entry(&ext->oe_pages,
1882                                                       struct osc_async_page,
1883                                                       oap_pending_item);
1884         ENTRY;
1885
1886         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1887                 ext);
1888         OSC_EXTENT_DUMP(D_CACHE, ext, "trying to add this extent\n");
1889
1890         if (data->erd_max_extents == 0)
1891                 RETURN(0);
1892
1893         chunk_count = osc_extent_chunks(ext);
1894         EASSERTF(data->erd_page_count != 0 ||
1895                  chunk_count <= data->erd_max_chunks, ext,
1896                  "The first extent to be fit in a RPC contains %u chunks, "
1897                  "which is over the limit %u.\n", chunk_count,
1898                  data->erd_max_chunks);
1899         if (chunk_count > data->erd_max_chunks)
1900                 RETURN(0);
1901
1902         data->erd_max_pages = max(ext->oe_mppr, data->erd_max_pages);
1903         EASSERTF(data->erd_page_count != 0 ||
1904                 ext->oe_nr_pages <= data->erd_max_pages, ext,
1905                 "The first extent to be fit in a RPC contains %u pages, "
1906                 "which is over the limit %u.\n", ext->oe_nr_pages,
1907                 data->erd_max_pages);
1908         if (data->erd_page_count + ext->oe_nr_pages > data->erd_max_pages)
1909                 RETURN(0);
1910
1911         list_for_each_entry(tmp, data->erd_rpc_list, oe_link) {
1912                 struct osc_async_page *oap2;
1913                 oap2 = list_first_entry(&tmp->oe_pages, struct osc_async_page,
1914                                         oap_pending_item);
1915                 EASSERT(tmp->oe_owner == current, tmp);
1916 #if 0
1917                 if (overlapped(tmp, ext)) {
1918                         OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
1919                         EASSERT(0, ext);
1920                 }
1921 #endif
1922                 if (oap2cl_page(oap)->cp_type != oap2cl_page(oap2)->cp_type) {
1923                         CDEBUG(D_CACHE, "Do not permit different types of IO "
1924                                "in one RPC\n");
1925                         RETURN(0);
1926                 }
1927
1928                 if (tmp->oe_srvlock != ext->oe_srvlock ||
1929                     !tmp->oe_grants != !ext->oe_grants ||
1930                     tmp->oe_ndelay != ext->oe_ndelay ||
1931                     tmp->oe_no_merge || ext->oe_no_merge)
1932                         RETURN(0);
1933
1934                 /* remove break for strict check */
1935                 break;
1936         }
1937
1938         data->erd_max_extents--;
1939         data->erd_max_chunks -= chunk_count;
1940         data->erd_page_count += ext->oe_nr_pages;
1941         list_move_tail(&ext->oe_link, data->erd_rpc_list);
1942         ext->oe_owner = current;
1943         RETURN(1);
1944 }
1945
1946 static inline unsigned osc_max_write_chunks(const struct client_obd *cli)
1947 {
1948         /*
1949          * LU-8135:
1950          *
1951          * The maximum size of a single transaction is about 64MB in ZFS.
1952          * #define DMU_MAX_ACCESS (64 * 1024 * 1024)
1953          *
1954          * Since ZFS is a copy-on-write file system, a single dirty page in
1955          * a chunk will result in the rewrite of the whole chunk, therefore
1956          * an RPC shouldn't be allowed to contain too many chunks otherwise
1957          * it will make transaction size much bigger than 64MB, especially
1958          * with big block size for ZFS.
1959          *
1960          * This piece of code is to make sure that OSC won't send write RPCs
1961          * with too many chunks. The maximum chunk size that an RPC can cover
1962          * is set to PTLRPC_MAX_BRW_SIZE, which is defined to 16MB. Ideally
1963          * OST should tell the client what the biggest transaction size is,
1964          * but it's good enough for now.
1965          *
1966          * This limitation doesn't apply to ldiskfs, which allows as many
1967          * chunks in one RPC as we want. However, it won't have any benefits
1968          * to have too many discontiguous pages in one RPC.
1969          *
1970          * An osc_extent won't cover over a RPC size, so the chunks in an
1971          * osc_extent won't bigger than PTLRPC_MAX_BRW_SIZE >> chunkbits.
1972          */
1973         return PTLRPC_MAX_BRW_SIZE >> cli->cl_chunkbits;
1974 }
1975
1976 /**
1977  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1978  * get_write_extent() takes all appropriate extents in atomic.
1979  *
1980  * The following policy is used to collect extents for IO:
1981  * 1. Add as many HP extents as possible;
1982  * 2. Add the first urgent extent in urgent extent list and take it out of
1983  *    urgent list;
1984  * 3. Add subsequent extents of this urgent extent;
1985  * 4. If urgent list is not empty, goto 2;
1986  * 5. Traverse the extent tree from the 1st extent;
1987  * 6. Above steps exit if there is no space in this RPC.
1988  */
1989 static unsigned int get_write_extents(struct osc_object *obj,
1990                                       struct list_head *rpclist)
1991 {
1992         struct client_obd *cli = osc_cli(obj);
1993         struct osc_extent *ext;
1994         struct extent_rpc_data data = {
1995                 .erd_rpc_list   = rpclist,
1996                 .erd_page_count = 0,
1997                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
1998                 .erd_max_chunks = osc_max_write_chunks(cli),
1999                 .erd_max_extents = 256,
2000         };
2001
2002         LASSERT(osc_object_is_locked(obj));
2003         while (!list_empty(&obj->oo_hp_exts)) {
2004                 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
2005                                  oe_link);
2006                 LASSERT(ext->oe_state == OES_CACHE);
2007                 if (!try_to_add_extent_for_io(cli, ext, &data))
2008                         return data.erd_page_count;
2009                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2010         }
2011         if (data.erd_page_count == data.erd_max_pages)
2012                 return data.erd_page_count;
2013
2014         while (!list_empty(&obj->oo_urgent_exts)) {
2015                 ext = list_entry(obj->oo_urgent_exts.next,
2016                                  struct osc_extent, oe_link);
2017                 if (!try_to_add_extent_for_io(cli, ext, &data))
2018                         return data.erd_page_count;
2019         }
2020         if (data.erd_page_count == data.erd_max_pages)
2021                 return data.erd_page_count;
2022
2023         /* One key difference between full extents and other extents: full
2024          * extents can usually only be added if the rpclist was empty, so if we
2025          * can't add one, we continue on to trying to add normal extents.  This
2026          * is so we don't miss adding extra extents to an RPC containing high
2027          * priority or urgent extents. */
2028         while (!list_empty(&obj->oo_full_exts)) {
2029                 ext = list_entry(obj->oo_full_exts.next,
2030                                  struct osc_extent, oe_link);
2031                 if (!try_to_add_extent_for_io(cli, ext, &data))
2032                         break;
2033         }
2034         if (data.erd_page_count == data.erd_max_pages)
2035                 return data.erd_page_count;
2036
2037         ext = first_extent(obj);
2038         while (ext != NULL) {
2039                 if ((ext->oe_state != OES_CACHE) ||
2040                     /* this extent may be already in current rpclist */
2041                     (!list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
2042                         ext = next_extent(ext);
2043                         continue;
2044                 }
2045
2046                 if (!try_to_add_extent_for_io(cli, ext, &data))
2047                         return data.erd_page_count;
2048
2049                 ext = next_extent(ext);
2050         }
2051         return data.erd_page_count;
2052 }
2053
2054 static int
2055 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
2056                    struct osc_object *osc)
2057 __must_hold(osc)
2058 {
2059         struct list_head   rpclist = LIST_HEAD_INIT(rpclist);
2060         struct osc_extent *ext;
2061         struct osc_extent *tmp;
2062         struct osc_extent *first = NULL;
2063         unsigned int page_count = 0;
2064         int srvlock = 0;
2065         int rc = 0;
2066         ENTRY;
2067
2068         LASSERT(osc_object_is_locked(osc));
2069
2070         page_count = get_write_extents(osc, &rpclist);
2071         LASSERT(equi(page_count == 0, list_empty(&rpclist)));
2072
2073         if (list_empty(&rpclist))
2074                 RETURN(0);
2075
2076         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2077
2078         list_for_each_entry(ext, &rpclist, oe_link) {
2079                 LASSERT(ext->oe_state == OES_CACHE ||
2080                         ext->oe_state == OES_LOCK_DONE);
2081                 if (ext->oe_state == OES_CACHE)
2082                         osc_extent_state_set(ext, OES_LOCKING);
2083                 else
2084                         osc_extent_state_set(ext, OES_RPC);
2085         }
2086
2087         /* we're going to grab page lock, so release object lock because
2088          * lock order is page lock -> object lock. */
2089         osc_object_unlock(osc);
2090
2091         list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2092                 if (ext->oe_state == OES_LOCKING) {
2093                         rc = osc_extent_make_ready(env, ext);
2094                         if (unlikely(rc < 0)) {
2095                                 list_del_init(&ext->oe_link);
2096                                 osc_extent_finish(env, ext, 0, rc);
2097                                 continue;
2098                         }
2099                 }
2100                 if (first == NULL) {
2101                         first = ext;
2102                         srvlock = ext->oe_srvlock;
2103                 } else {
2104                         LASSERT(srvlock == ext->oe_srvlock);
2105                 }
2106         }
2107
2108         if (!list_empty(&rpclist)) {
2109                 LASSERT(page_count > 0);
2110                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2111                 LASSERT(list_empty(&rpclist));
2112         }
2113
2114         osc_object_lock(osc);
2115         RETURN(rc);
2116 }
2117
2118 /**
2119  * prepare pages for ASYNC io and put pages in send queue.
2120  *
2121  * \param cmd OBD_BRW_* macroses
2122  * \param lop pending pages
2123  *
2124  * \return zero if no page added to send queue.
2125  * \return 1 if pages successfully added to send queue.
2126  * \return negative on errors.
2127  */
2128 static int
2129 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2130                   struct osc_object *osc)
2131 __must_hold(osc)
2132 {
2133         struct osc_extent *ext;
2134         struct osc_extent *next;
2135         struct list_head rpclist = LIST_HEAD_INIT(rpclist);
2136         struct extent_rpc_data data = {
2137                 .erd_rpc_list   = &rpclist,
2138                 .erd_page_count = 0,
2139                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
2140                 .erd_max_chunks = UINT_MAX,
2141                 .erd_max_extents = UINT_MAX,
2142         };
2143         int rc = 0;
2144         ENTRY;
2145
2146         LASSERT(osc_object_is_locked(osc));
2147         list_for_each_entry_safe(ext, next, &osc->oo_reading_exts, oe_link) {
2148                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2149                 if (!try_to_add_extent_for_io(cli, ext, &data))
2150                         break;
2151                 osc_extent_state_set(ext, OES_RPC);
2152                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2153         }
2154         LASSERT(data.erd_page_count <= data.erd_max_pages);
2155
2156         osc_update_pending(osc, OBD_BRW_READ, -data.erd_page_count);
2157
2158         if (!list_empty(&rpclist)) {
2159                 osc_object_unlock(osc);
2160
2161                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2162                 LASSERT(list_empty(&rpclist));
2163
2164                 osc_object_lock(osc);
2165         }
2166         RETURN(rc);
2167 }
2168
2169 #define list_to_obj(list, item) ({                                            \
2170         struct list_head *__tmp = (list)->next;                               \
2171         list_del_init(__tmp);                                         \
2172         list_entry(__tmp, struct osc_object, oo_##item);                      \
2173 })
2174
2175 /* This is called by osc_check_rpcs() to find which objects have pages that
2176  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2177 static struct osc_object *osc_next_obj(struct client_obd *cli)
2178 {
2179         ENTRY;
2180
2181         /* First return objects that have blocked locks so that they
2182          * will be flushed quickly and other clients can get the lock,
2183          * then objects which have pages ready to be stuffed into RPCs */
2184         if (!list_empty(&cli->cl_loi_hp_ready_list))
2185                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2186         if (!list_empty(&cli->cl_loi_ready_list))
2187                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2188
2189         /* then if we have cache waiters, return all objects with queued
2190          * writes.  This is especially important when many small files
2191          * have filled up the cache and not been fired into rpcs because
2192          * they don't pass the nr_pending/object threshhold
2193          */
2194         if (waitqueue_active(&cli->cl_cache_waiters) &&
2195             !list_empty(&cli->cl_loi_write_list))
2196                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2197
2198         /* then return all queued objects when we have an invalid import
2199          * so that they get flushed */
2200         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2201                 if (!list_empty(&cli->cl_loi_write_list))
2202                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2203                                            write_item));
2204                 if (!list_empty(&cli->cl_loi_read_list))
2205                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2206                                            read_item));
2207         }
2208         RETURN(NULL);
2209 }
2210
2211 /* called with the loi list lock held */
2212 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2213 __must_hold(&cli->cl_loi_list_lock)
2214 {
2215         struct osc_object *osc;
2216         int rc = 0;
2217         ENTRY;
2218
2219         while ((osc = osc_next_obj(cli)) != NULL) {
2220                 struct cl_object *obj = osc2cl(osc);
2221                 struct lu_ref_link link;
2222
2223                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2224
2225                 /* even if we have reached our max in flight RPCs, we still
2226                  * allow all high-priority RPCs through to prevent their
2227                  * starvation and leading to server evicting us for not
2228                  * writing out pages in a timely manner LU-13131 */
2229                 if (osc_max_rpc_in_flight(cli, osc) &&
2230                     list_empty(&osc->oo_hp_exts)) {
2231                         __osc_list_maint(cli, osc);
2232                         break;
2233                 }
2234
2235                 cl_object_get(obj);
2236                 spin_unlock(&cli->cl_loi_list_lock);
2237                 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2238
2239                 /* attempt some read/write balancing by alternating between
2240                  * reads and writes in an object.  The makes_rpc checks here
2241                  * would be redundant if we were getting read/write work items
2242                  * instead of objects.  we don't want send_oap_rpc to drain a
2243                  * partial read pending queue when we're given this object to
2244                  * do io on writes while there are cache waiters */
2245                 osc_object_lock(osc);
2246                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2247                         rc = osc_send_write_rpc(env, cli, osc);
2248                         if (rc < 0) {
2249                                 CERROR("Write request failed with %d\n", rc);
2250
2251                                 /* osc_send_write_rpc failed, mostly because of
2252                                  * memory pressure.
2253                                  *
2254                                  * It can't break here, because if:
2255                                  *  - a page was submitted by osc_io_submit, so
2256                                  *    page locked;
2257                                  *  - no request in flight
2258                                  *  - no subsequent request
2259                                  * The system will be in live-lock state,
2260                                  * because there is no chance to call
2261                                  * osc_io_unplug() and osc_check_rpcs() any
2262                                  * more. pdflush can't help in this case,
2263                                  * because it might be blocked at grabbing
2264                                  * the page lock as we mentioned.
2265                                  *
2266                                  * Anyway, continue to drain pages. */
2267                                 /* break; */
2268                         }
2269                 }
2270                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2271                         rc = osc_send_read_rpc(env, cli, osc);
2272                         if (rc < 0)
2273                                 CERROR("Read request failed with %d\n", rc);
2274                 }
2275                 osc_object_unlock(osc);
2276
2277                 osc_list_maint(cli, osc);
2278                 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2279                 cl_object_put(env, obj);
2280
2281                 spin_lock(&cli->cl_loi_list_lock);
2282         }
2283 }
2284
2285 int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2286                    struct osc_object *osc, int async)
2287 {
2288         int rc = 0;
2289
2290         if (osc != NULL && osc_list_maint(cli, osc) == 0)
2291                 return 0;
2292
2293         if (!async) {
2294                 spin_lock(&cli->cl_loi_list_lock);
2295                 osc_check_rpcs(env, cli);
2296                 spin_unlock(&cli->cl_loi_list_lock);
2297         } else {
2298                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2299                 LASSERT(cli->cl_writeback_work != NULL);
2300                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2301         }
2302         return rc;
2303 }
2304 EXPORT_SYMBOL(osc_io_unplug0);
2305
2306 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2307                         struct page *page, loff_t offset)
2308 {
2309         struct obd_export     *exp = osc_export(osc);
2310         struct osc_async_page *oap = &ops->ops_oap;
2311         ENTRY;
2312
2313         if (!page)
2314                 return cfs_size_round(sizeof(*oap));
2315
2316         oap->oap_magic = OAP_MAGIC;
2317         oap->oap_cli = &exp->exp_obd->u.cli;
2318         oap->oap_obj = osc;
2319
2320         oap->oap_page = page;
2321         oap->oap_obj_off = offset;
2322         LASSERT(!(offset & ~PAGE_MASK));
2323
2324         INIT_LIST_HEAD(&oap->oap_pending_item);
2325         INIT_LIST_HEAD(&oap->oap_rpc_item);
2326
2327         spin_lock_init(&oap->oap_lock);
2328         CDEBUG(D_INFO, "oap %p page %p obj off %llu\n",
2329                oap, page, oap->oap_obj_off);
2330         RETURN(0);
2331 }
2332 EXPORT_SYMBOL(osc_prep_async_page);
2333
2334 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2335                        struct osc_page *ops)
2336 {
2337         struct osc_io *oio = osc_env_io(env);
2338         struct osc_extent     *ext = NULL;
2339         struct osc_async_page *oap = &ops->ops_oap;
2340         struct client_obd     *cli = oap->oap_cli;
2341         struct osc_object     *osc = oap->oap_obj;
2342         pgoff_t index;
2343         unsigned int tmp;
2344         unsigned int grants = 0;
2345         u32    brw_flags = OBD_BRW_ASYNC;
2346         int    cmd = OBD_BRW_WRITE;
2347         int    need_release = 0;
2348         int    rc = 0;
2349         ENTRY;
2350
2351         if (oap->oap_magic != OAP_MAGIC)
2352                 RETURN(-EINVAL);
2353
2354         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2355                 RETURN(-EIO);
2356
2357         if (!list_empty(&oap->oap_pending_item) ||
2358             !list_empty(&oap->oap_rpc_item))
2359                 RETURN(-EBUSY);
2360
2361         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2362         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2363         if (oio->oi_cap_sys_resource || io->ci_noquota) {
2364                 brw_flags |= OBD_BRW_NOQUOTA;
2365                 cmd |= OBD_BRW_NOQUOTA;
2366         }
2367
2368         /* check if the file's owner/group is over quota */
2369         if (!(cmd & OBD_BRW_NOQUOTA)) {
2370                 struct cl_object *obj;
2371                 struct cl_attr   *attr;
2372                 unsigned int qid[LL_MAXQUOTAS];
2373
2374                 obj = cl_object_top(&osc->oo_cl);
2375                 attr = &osc_env_info(env)->oti_attr;
2376
2377                 cl_object_attr_lock(obj);
2378                 rc = cl_object_attr_get(env, obj, attr);
2379                 cl_object_attr_unlock(obj);
2380
2381                 qid[USRQUOTA] = attr->cat_uid;
2382                 qid[GRPQUOTA] = attr->cat_gid;
2383                 qid[PRJQUOTA] = attr->cat_projid;
2384                 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2385                         rc = -EDQUOT;
2386                 if (rc)
2387                         RETURN(rc);
2388         }
2389
2390         oap->oap_cmd = cmd;
2391         oap->oap_page_off = ops->ops_from;
2392         oap->oap_count = ops->ops_to - ops->ops_from;
2393         /* No need to hold a lock here,
2394          * since this page is not in any list yet. */
2395         oap->oap_async_flags = 0;
2396         oap->oap_brw_flags = brw_flags;
2397
2398         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2399                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2400
2401         index = osc_index(oap2osc(oap));
2402
2403         /* Add this page into extent by the following steps:
2404          * 1. if there exists an active extent for this IO, mostly this page
2405          *    can be added to the active extent and sometimes we need to
2406          *    expand extent to accomodate this page;
2407          * 2. otherwise, a new extent will be allocated. */
2408
2409         ext = oio->oi_active;
2410         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2411                 /* one chunk plus extent overhead must be enough to write this
2412                  * page */
2413                 grants = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2414                 if (ext->oe_end >= index)
2415                         grants = 0;
2416
2417                 /* it doesn't need any grant to dirty this page */
2418                 spin_lock(&cli->cl_loi_list_lock);
2419                 rc = osc_enter_cache_try(cli, oap, grants, 0);
2420                 spin_unlock(&cli->cl_loi_list_lock);
2421                 if (rc == 0) { /* try failed */
2422                         grants = 0;
2423                         need_release = 1;
2424                 } else if (ext->oe_end < index) {
2425                         tmp = grants;
2426                         /* try to expand this extent */
2427                         rc = osc_extent_expand(ext, index, &tmp);
2428                         if (rc < 0) {
2429                                 need_release = 1;
2430                                 /* don't free reserved grant */
2431                         } else {
2432                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2433                                                 "expanded for %lu.\n", index);
2434                                 osc_unreserve_grant(cli, grants, tmp);
2435                                 grants = 0;
2436                         }
2437                 }
2438                 rc = 0;
2439         } else if (ext != NULL) {
2440                 /* index is located outside of active extent */
2441                 need_release = 1;
2442         }
2443         if (need_release) {
2444                 osc_extent_release(env, ext);
2445                 oio->oi_active = NULL;
2446                 ext = NULL;
2447         }
2448
2449         if (ext == NULL) {
2450                 tmp = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2451
2452                 /* try to find new extent to cover this page */
2453                 LASSERT(oio->oi_active == NULL);
2454                 /* we may have allocated grant for this page if we failed
2455                  * to expand the previous active extent. */
2456                 LASSERT(ergo(grants > 0, grants >= tmp));
2457
2458                 rc = 0;
2459                 if (grants == 0) {
2460                         /* we haven't allocated grant for this page. */
2461                         rc = osc_enter_cache(env, cli, oap, tmp);
2462                         if (rc == 0)
2463                                 grants = tmp;
2464                 }
2465
2466                 tmp = grants;
2467                 if (rc == 0) {
2468                         ext = osc_extent_find(env, osc, index, &tmp);
2469                         if (IS_ERR(ext)) {
2470                                 LASSERT(tmp == grants);
2471                                 osc_exit_cache(cli, oap);
2472                                 rc = PTR_ERR(ext);
2473                                 ext = NULL;
2474                         } else {
2475                                 oio->oi_active = ext;
2476                         }
2477                 }
2478                 if (grants > 0)
2479                         osc_unreserve_grant(cli, grants, tmp);
2480         }
2481
2482         LASSERT(ergo(rc == 0, ext != NULL));
2483         if (ext != NULL) {
2484                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2485                          ext, "index = %lu.\n", index);
2486                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2487
2488                 osc_object_lock(osc);
2489                 if (ext->oe_nr_pages == 0)
2490                         ext->oe_srvlock = ops->ops_srvlock;
2491                 else
2492                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2493                 ++ext->oe_nr_pages;
2494                 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2495                 osc_object_unlock(osc);
2496
2497                 if (!ext->oe_layout_version)
2498                         ext->oe_layout_version = io->ci_layout_version;
2499         }
2500
2501         RETURN(rc);
2502 }
2503
2504 int osc_teardown_async_page(const struct lu_env *env,
2505                             struct osc_object *obj, struct osc_page *ops)
2506 {
2507         struct osc_async_page *oap = &ops->ops_oap;
2508         int rc = 0;
2509         ENTRY;
2510
2511         LASSERT(oap->oap_magic == OAP_MAGIC);
2512
2513         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2514                oap, ops, osc_index(oap2osc(oap)));
2515
2516         if (!list_empty(&oap->oap_rpc_item)) {
2517                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2518                 rc = -EBUSY;
2519         } else if (!list_empty(&oap->oap_pending_item)) {
2520                 struct osc_extent *ext = NULL;
2521
2522                 osc_object_lock(obj);
2523                 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2524                 osc_object_unlock(obj);
2525                 /* only truncated pages are allowed to be taken out.
2526                  * See osc_extent_truncate() and osc_cache_truncate_start()
2527                  * for details. */
2528                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2529                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2530                                         osc_index(oap2osc(oap)));
2531                         rc = -EBUSY;
2532                 }
2533                 if (ext != NULL)
2534                         osc_extent_put(env, ext);
2535         }
2536         RETURN(rc);
2537 }
2538
2539 /**
2540  * This is called when a page is picked up by kernel to write out.
2541  *
2542  * We should find out the corresponding extent and add the whole extent
2543  * into urgent list. The extent may be being truncated or used, handle it
2544  * carefully.
2545  */
2546 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2547                          struct osc_page *ops)
2548 {
2549         struct osc_extent *ext   = NULL;
2550         struct osc_object *obj   = cl2osc(ops->ops_cl.cpl_obj);
2551         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2552         pgoff_t            index = osc_index(ops);
2553         struct osc_async_page *oap = &ops->ops_oap;
2554         bool unplug = false;
2555         int rc = 0;
2556         ENTRY;
2557
2558         osc_object_lock(obj);
2559         ext = osc_extent_lookup(obj, index);
2560         if (ext == NULL) {
2561                 osc_extent_tree_dump(D_ERROR, obj);
2562                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2563         }
2564
2565         switch (ext->oe_state) {
2566         case OES_RPC:
2567         case OES_LOCK_DONE:
2568                 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2569                 LASSERT(0);
2570                 break;
2571         case OES_LOCKING:
2572                 /* If we know this extent is being written out, we should abort
2573                  * so that the writer can make this page ready. Otherwise, there
2574                  * exists a deadlock problem because other process can wait for
2575                  * page writeback bit holding page lock; and meanwhile in
2576                  * vvp_page_make_ready(), we need to grab page lock before
2577                  * really sending the RPC. */
2578         case OES_TRUNC:
2579                 /* race with truncate, page will be redirtied */
2580         case OES_ACTIVE:
2581                 /* The extent is active so we need to abort and let the caller
2582                  * re-dirty the page. If we continued on here, and we were the
2583                  * one making the extent active, we could deadlock waiting for
2584                  * the page writeback to clear but it won't because the extent
2585                  * is active and won't be written out. */
2586                 GOTO(out, rc = -EAGAIN);
2587         default:
2588                 break;
2589         }
2590
2591         rc = cl_page_prep(env, io, cp, CRT_WRITE);
2592         if (rc)
2593                 GOTO(out, rc);
2594
2595         spin_lock(&oap->oap_lock);
2596         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2597         spin_unlock(&oap->oap_lock);
2598
2599         if (memory_pressure_get())
2600                 ext->oe_memalloc = 1;
2601
2602         ext->oe_urgent = 1;
2603         if (ext->oe_state == OES_CACHE) {
2604                 OSC_EXTENT_DUMP(D_CACHE, ext,
2605                                 "flush page %p make it urgent.\n", oap);
2606                 if (list_empty(&ext->oe_link))
2607                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2608                 unplug = true;
2609         }
2610         rc = 0;
2611         EXIT;
2612
2613 out:
2614         osc_object_unlock(obj);
2615         osc_extent_put(env, ext);
2616         if (unplug)
2617                 osc_io_unplug_async(env, osc_cli(obj), obj);
2618         return rc;
2619 }
2620
2621 /**
2622  * this is called when a sync waiter receives an interruption.  Its job is to
2623  * get the caller woken as soon as possible.  If its page hasn't been put in an
2624  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2625  * desiring interruption which will forcefully complete the rpc once the rpc
2626  * has timed out.
2627  */
2628 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2629 {
2630         struct osc_async_page *oap = &ops->ops_oap;
2631         struct osc_object     *obj = oap->oap_obj;
2632         struct client_obd     *cli = osc_cli(obj);
2633         struct osc_extent     *ext;
2634         struct osc_extent     *found = NULL;
2635         struct list_head            *plist;
2636         pgoff_t index = osc_index(ops);
2637         int     rc = -EBUSY;
2638         int     cmd;
2639         ENTRY;
2640
2641         LASSERT(!oap->oap_interrupted);
2642         oap->oap_interrupted = 1;
2643
2644         /* Find out the caching extent */
2645         osc_object_lock(obj);
2646         if (oap->oap_cmd & OBD_BRW_WRITE) {
2647                 plist = &obj->oo_urgent_exts;
2648                 cmd   = OBD_BRW_WRITE;
2649         } else {
2650                 plist = &obj->oo_reading_exts;
2651                 cmd   = OBD_BRW_READ;
2652         }
2653         list_for_each_entry(ext, plist, oe_link) {
2654                 if (ext->oe_start <= index && ext->oe_end >= index) {
2655                         LASSERT(ext->oe_state == OES_LOCK_DONE);
2656                         /* For OES_LOCK_DONE state extent, it has already held
2657                          * a refcount for RPC. */
2658                         found = osc_extent_get(ext);
2659                         break;
2660                 }
2661         }
2662         if (found != NULL) {
2663                 list_del_init(&found->oe_link);
2664                 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2665                 osc_object_unlock(obj);
2666
2667                 osc_extent_finish(env, found, 0, -EINTR);
2668                 osc_extent_put(env, found);
2669                 rc = 0;
2670         } else {
2671                 osc_object_unlock(obj);
2672                 /* ok, it's been put in an rpc. only one oap gets a request
2673                  * reference */
2674                 if (oap->oap_request != NULL) {
2675                         ptlrpc_mark_interrupted(oap->oap_request);
2676                         ptlrpcd_wake(oap->oap_request);
2677                         ptlrpc_req_finished(oap->oap_request);
2678                         oap->oap_request = NULL;
2679                 }
2680         }
2681
2682         osc_list_maint(cli, obj);
2683         RETURN(rc);
2684 }
2685
2686 int osc_queue_sync_pages(const struct lu_env *env, const struct cl_io *io,
2687                          struct osc_object *obj, struct list_head *list,
2688                          int brw_flags)
2689 {
2690         struct client_obd     *cli = osc_cli(obj);
2691         struct osc_extent     *ext;
2692         struct osc_async_page *oap;
2693         int     page_count = 0;
2694         int     mppr       = cli->cl_max_pages_per_rpc;
2695         bool    can_merge   = true;
2696         pgoff_t start      = CL_PAGE_EOF;
2697         pgoff_t end        = 0;
2698         ENTRY;
2699
2700         list_for_each_entry(oap, list, oap_pending_item) {
2701                 struct osc_page *opg = oap2osc_page(oap);
2702                 pgoff_t index = osc_index(opg);
2703
2704                 if (index > end)
2705                         end = index;
2706                 if (index < start)
2707                         start = index;
2708                 ++page_count;
2709                 mppr <<= (page_count > mppr);
2710
2711                 if (unlikely(opg->ops_from > 0 || opg->ops_to < PAGE_SIZE))
2712                         can_merge = false;
2713         }
2714
2715         ext = osc_extent_alloc(obj);
2716         if (ext == NULL) {
2717                 struct osc_async_page *tmp;
2718
2719                 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2720                         list_del_init(&oap->oap_pending_item);
2721                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2722                 }
2723                 RETURN(-ENOMEM);
2724         }
2725
2726         ext->oe_rw = !!(brw_flags & OBD_BRW_READ);
2727         ext->oe_sync = 1;
2728         ext->oe_no_merge = !can_merge;
2729         ext->oe_urgent = 1;
2730         ext->oe_start = start;
2731         ext->oe_end = ext->oe_max_end = end;
2732         ext->oe_obj = obj;
2733         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2734         ext->oe_ndelay = !!(brw_flags & OBD_BRW_NDELAY);
2735         ext->oe_nr_pages = page_count;
2736         ext->oe_mppr = mppr;
2737         list_splice_init(list, &ext->oe_pages);
2738         ext->oe_layout_version = io->ci_layout_version;
2739
2740         osc_object_lock(obj);
2741         /* Reuse the initial refcount for RPC, don't drop it */
2742         osc_extent_state_set(ext, OES_LOCK_DONE);
2743         if (!ext->oe_rw) { /* write */
2744                 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2745                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2746         } else {
2747                 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2748                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2749         }
2750         osc_object_unlock(obj);
2751
2752         osc_io_unplug_async(env, cli, obj);
2753         RETURN(0);
2754 }
2755
2756 /**
2757  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2758  */
2759 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
2760                              __u64 size, struct osc_extent **extp)
2761 {
2762         struct client_obd *cli = osc_cli(obj);
2763         struct osc_extent *ext;
2764         struct osc_extent *waiting = NULL;
2765         pgoff_t index;
2766         struct list_head list = LIST_HEAD_INIT(list);
2767         int result = 0;
2768         bool partial;
2769         ENTRY;
2770
2771         /* pages with index greater or equal to index will be truncated. */
2772         index = cl_index(osc2cl(obj), size);
2773         partial = size > cl_offset(osc2cl(obj), index);
2774
2775 again:
2776         osc_object_lock(obj);
2777         ext = osc_extent_search(obj, index);
2778         if (ext == NULL)
2779                 ext = first_extent(obj);
2780         else if (ext->oe_end < index)
2781                 ext = next_extent(ext);
2782         while (ext != NULL) {
2783                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2784
2785                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2786                         /* if ext is in urgent state, it means there must exist
2787                          * a page already having been flushed by write_page().
2788                          * We have to wait for this extent because we can't
2789                          * truncate that page. */
2790                         OSC_EXTENT_DUMP(D_CACHE, ext,
2791                                         "waiting for busy extent\n");
2792                         waiting = osc_extent_get(ext);
2793                         break;
2794                 }
2795
2796                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:%llu.\n", size);
2797
2798                 osc_extent_get(ext);
2799                 if (ext->oe_state == OES_ACTIVE) {
2800                         /* though we grab inode mutex for write path, but we
2801                          * release it before releasing extent(in osc_io_end()),
2802                          * so there is a race window that an extent is still
2803                          * in OES_ACTIVE when truncate starts. */
2804                         LASSERT(!ext->oe_trunc_pending);
2805                         ext->oe_trunc_pending = 1;
2806                 } else {
2807                         EASSERT(ext->oe_state == OES_CACHE, ext);
2808                         osc_extent_state_set(ext, OES_TRUNC);
2809                         osc_update_pending(obj, OBD_BRW_WRITE,
2810                                            -ext->oe_nr_pages);
2811                 }
2812                 /* This extent could be on the full extents list, that's OK */
2813                 EASSERT(!ext->oe_hp && !ext->oe_urgent, ext);
2814                 if (!list_empty(&ext->oe_link))
2815                         list_move_tail(&ext->oe_link, &list);
2816                 else
2817                         list_add_tail(&ext->oe_link, &list);
2818
2819                 ext = next_extent(ext);
2820         }
2821         osc_object_unlock(obj);
2822
2823         osc_list_maint(cli, obj);
2824
2825         while (!list_empty(&list)) {
2826                 int rc;
2827
2828                 ext = list_entry(list.next, struct osc_extent, oe_link);
2829                 list_del_init(&ext->oe_link);
2830
2831                 /* extent may be in OES_ACTIVE state because inode mutex
2832                  * is released before osc_io_end() in file write case */
2833                 if (ext->oe_state != OES_TRUNC)
2834                         osc_extent_wait(env, ext, OES_TRUNC);
2835
2836                 rc = osc_extent_truncate(ext, index, partial);
2837                 if (rc < 0) {
2838                         if (result == 0)
2839                                 result = rc;
2840
2841                         OSC_EXTENT_DUMP(D_ERROR, ext,
2842                                         "truncate error %d\n", rc);
2843                 } else if (ext->oe_nr_pages == 0) {
2844                         osc_extent_remove(ext);
2845                 } else {
2846                         /* this must be an overlapped extent which means only
2847                          * part of pages in this extent have been truncated.
2848                          */
2849                         EASSERTF(ext->oe_start <= index, ext,
2850                                  "trunc index = %lu/%d.\n", index, partial);
2851                         /* fix index to skip this partially truncated extent */
2852                         index = ext->oe_end + 1;
2853                         partial = false;
2854
2855                         /* we need to hold this extent in OES_TRUNC state so
2856                          * that no writeback will happen. This is to avoid
2857                          * BUG 17397.
2858                          * Only partial truncate can reach here, if @size is
2859                          * not zero, the caller should provide a valid @extp. */
2860                         LASSERT(*extp == NULL);
2861                         *extp = osc_extent_get(ext);
2862                         OSC_EXTENT_DUMP(D_CACHE, ext,
2863                                         "trunc at %llu\n", size);
2864                 }
2865                 osc_extent_put(env, ext);
2866         }
2867         if (waiting != NULL) {
2868                 int rc;
2869
2870                 /* ignore the result of osc_extent_wait the write initiator
2871                  * should take care of it. */
2872                 rc = osc_extent_wait(env, waiting, OES_INV);
2873                 if (rc < 0)
2874                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2875
2876                 osc_extent_put(env, waiting);
2877                 waiting = NULL;
2878                 goto again;
2879         }
2880         RETURN(result);
2881 }
2882 EXPORT_SYMBOL(osc_cache_truncate_start);
2883
2884 /**
2885  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2886  */
2887 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext)
2888 {
2889         if (ext != NULL) {
2890                 struct osc_object *obj = ext->oe_obj;
2891                 bool unplug = false;
2892
2893                 EASSERT(ext->oe_nr_pages > 0, ext);
2894                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2895                 EASSERT(!ext->oe_urgent, ext);
2896
2897                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2898                 osc_object_lock(obj);
2899                 osc_extent_state_set(ext, OES_CACHE);
2900                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2901                         ext->oe_urgent = 1;
2902                         list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2903                         unplug = true;
2904                 }
2905                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2906                 osc_object_unlock(obj);
2907                 osc_extent_put(env, ext);
2908
2909                 if (unplug)
2910                         osc_io_unplug_async(env, osc_cli(obj), obj);
2911         }
2912 }
2913
2914 /**
2915  * Wait for extents in a specific range to be written out.
2916  * The caller must have called osc_cache_writeback_range() to issue IO
2917  * otherwise it will take a long time for this function to finish.
2918  *
2919  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2920  * nobody else can dirty this range of file while we're waiting for
2921  * extents to be written.
2922  */
2923 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2924                          pgoff_t start, pgoff_t end)
2925 {
2926         struct osc_extent *ext;
2927         pgoff_t index = start;
2928         int     result = 0;
2929         ENTRY;
2930
2931 again:
2932         osc_object_lock(obj);
2933         ext = osc_extent_search(obj, index);
2934         if (ext == NULL)
2935                 ext = first_extent(obj);
2936         else if (ext->oe_end < index)
2937                 ext = next_extent(ext);
2938         while (ext != NULL) {
2939                 int rc;
2940
2941                 if (ext->oe_start > end)
2942                         break;
2943
2944                 if (!ext->oe_fsync_wait) {
2945                         ext = next_extent(ext);
2946                         continue;
2947                 }
2948
2949                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2950                              ext->oe_hp || ext->oe_urgent), ext);
2951                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2952                              !ext->oe_hp && ext->oe_urgent), ext);
2953
2954                 index = ext->oe_end + 1;
2955                 osc_extent_get(ext);
2956                 osc_object_unlock(obj);
2957
2958                 rc = osc_extent_wait(env, ext, OES_INV);
2959                 if (result == 0)
2960                         result = rc;
2961                 osc_extent_put(env, ext);
2962                 goto again;
2963         }
2964         osc_object_unlock(obj);
2965
2966         OSC_IO_DEBUG(obj, "sync file range.\n");
2967         RETURN(result);
2968 }
2969 EXPORT_SYMBOL(osc_cache_wait_range);
2970
2971 /**
2972  * Called to write out a range of osc object.
2973  *
2974  * @hp     : should be set this is caused by lock cancel;
2975  * @discard: is set if dirty pages should be dropped - file will be deleted or
2976  *         truncated, this implies there is no partially discarding extents.
2977  *
2978  * Return how many pages will be issued, or error code if error occurred.
2979  */
2980 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2981                               pgoff_t start, pgoff_t end, int hp, int discard)
2982 {
2983         struct osc_extent *ext;
2984         struct list_head discard_list = LIST_HEAD_INIT(discard_list);
2985         bool unplug = false;
2986         int result = 0;
2987         ENTRY;
2988
2989         osc_object_lock(obj);
2990         ext = osc_extent_search(obj, start);
2991         if (ext == NULL)
2992                 ext = first_extent(obj);
2993         else if (ext->oe_end < start)
2994                 ext = next_extent(ext);
2995         while (ext != NULL) {
2996                 if (ext->oe_start > end)
2997                         break;
2998
2999                 ext->oe_fsync_wait = 1;
3000                 switch (ext->oe_state) {
3001                 case OES_CACHE:
3002                         result += ext->oe_nr_pages;
3003                         if (!discard) {
3004                                 struct list_head *list = NULL;
3005                                 if (hp) {
3006                                         EASSERT(!ext->oe_hp, ext);
3007                                         ext->oe_hp = 1;
3008                                         list = &obj->oo_hp_exts;
3009                                 } else if (!ext->oe_urgent && !ext->oe_hp) {
3010                                         ext->oe_urgent = 1;
3011                                         list = &obj->oo_urgent_exts;
3012                                 }
3013                                 if (list != NULL)
3014                                         list_move_tail(&ext->oe_link, list);
3015                                 unplug = true;
3016                         } else {
3017                                 struct client_obd *cli = osc_cli(obj);
3018                                 int pcc_bits = cli->cl_chunkbits - PAGE_SHIFT;
3019                                 pgoff_t align_by = (1 << pcc_bits);
3020                                 pgoff_t a_start = round_down(start, align_by);
3021                                 pgoff_t a_end = round_up(end, align_by);
3022
3023                                 /* overflow case */
3024                                 if (end && !a_end)
3025                                         a_end = CL_PAGE_EOF;
3026                                 /* the only discarder is lock cancelling, so
3027                                  * [start, end], aligned by chunk size, must
3028                                  * contain this extent */
3029                                 LASSERTF(ext->oe_start >= a_start &&
3030                                          ext->oe_end <= a_end,
3031                                          "ext [%lu, %lu] reg [%lu, %lu] "
3032                                          "orig [%lu %lu] align %lu bits "
3033                                          "%d\n", ext->oe_start, ext->oe_end,
3034                                          a_start, a_end, start, end,
3035                                          align_by, pcc_bits);
3036                                 osc_extent_state_set(ext, OES_LOCKING);
3037                                 ext->oe_owner = current;
3038                                 list_move_tail(&ext->oe_link,
3039                                                    &discard_list);
3040                                 osc_update_pending(obj, OBD_BRW_WRITE,
3041                                                    -ext->oe_nr_pages);
3042                         }
3043                         break;
3044                 case OES_ACTIVE:
3045                         /* It's pretty bad to wait for ACTIVE extents, because
3046                          * we don't know how long we will wait for it to be
3047                          * flushed since it may be blocked at awaiting more
3048                          * grants. We do this for the correctness of fsync. */
3049                         LASSERT(hp == 0 && discard == 0);
3050                         ext->oe_urgent = 1;
3051                         break;
3052                 case OES_TRUNC:
3053                         /* this extent is being truncated, can't do anything
3054                          * for it now. it will be set to urgent after truncate
3055                          * is finished in osc_cache_truncate_end(). */
3056                 default:
3057                         break;
3058                 }
3059                 ext = next_extent(ext);
3060         }
3061         osc_object_unlock(obj);
3062
3063         LASSERT(ergo(!discard, list_empty(&discard_list)));
3064         if (!list_empty(&discard_list)) {
3065                 struct osc_extent *tmp;
3066                 int rc;
3067
3068                 osc_list_maint(osc_cli(obj), obj);
3069                 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
3070                         list_del_init(&ext->oe_link);
3071                         EASSERT(ext->oe_state == OES_LOCKING, ext);
3072
3073                         /* Discard caching pages. We don't actually write this
3074                          * extent out but we complete it as if we did. */
3075                         rc = osc_extent_make_ready(env, ext);
3076                         if (unlikely(rc < 0)) {
3077                                 OSC_EXTENT_DUMP(D_ERROR, ext,
3078                                                 "make_ready returned %d\n", rc);
3079                                 if (result >= 0)
3080                                         result = rc;
3081                         }
3082
3083                         /* finish the extent as if the pages were sent */
3084                         osc_extent_finish(env, ext, 0, 0);
3085                 }
3086         }
3087
3088         if (unplug)
3089                 osc_io_unplug(env, osc_cli(obj), obj);
3090
3091         if (hp || discard) {
3092                 int rc;
3093                 rc = osc_cache_wait_range(env, obj, start, end);
3094                 if (result >= 0 && rc < 0)
3095                         result = rc;
3096         }
3097
3098         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3099         RETURN(result);
3100 }
3101 EXPORT_SYMBOL(osc_cache_writeback_range);
3102
3103 /**
3104  * Returns a list of pages by a given [start, end] of \a obj.
3105  *
3106  * \param resched If not NULL, then we give up before hogging CPU for too
3107  * long and set *resched = 1, in that case caller should implement a retry
3108  * logic.
3109  *
3110  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3111  * crucial in the face of [offset, EOF] locks.
3112  *
3113  * Return at least one page in @queue unless there is no covered page.
3114  */
3115 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3116                         struct osc_object *osc, pgoff_t start, pgoff_t end,
3117                         osc_page_gang_cbt cb, void *cbdata)
3118 {
3119         struct osc_page *ops;
3120         struct pagevec  *pagevec;
3121         void            **pvec;
3122         pgoff_t         idx;
3123         unsigned int    nr;
3124         unsigned int    i;
3125         unsigned int    j;
3126         int             res = CLP_GANG_OKAY;
3127         bool            tree_lock = true;
3128         ENTRY;
3129
3130         idx = start;
3131         pvec = osc_env_info(env)->oti_pvec;
3132         pagevec = &osc_env_info(env)->oti_pagevec;
3133         ll_pagevec_init(pagevec, 0);
3134         spin_lock(&osc->oo_tree_lock);
3135         while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3136                                             idx, OTI_PVEC_SIZE)) > 0) {
3137                 struct cl_page *page;
3138                 bool end_of_region = false;
3139
3140                 for (i = 0, j = 0; i < nr; ++i) {
3141                         ops = pvec[i];
3142                         pvec[i] = NULL;
3143
3144                         idx = osc_index(ops);
3145                         if (idx > end) {
3146                                 end_of_region = true;
3147                                 break;
3148                         }
3149
3150                         page = ops->ops_cl.cpl_page;
3151                         LASSERT(page->cp_type == CPT_CACHEABLE);
3152                         if (page->cp_state == CPS_FREEING)
3153                                 continue;
3154
3155                         cl_page_get(page);
3156                         lu_ref_add_atomic(&page->cp_reference,
3157                                           "gang_lookup", current);
3158                         pvec[j++] = ops;
3159                 }
3160                 ++idx;
3161
3162                 /*
3163                  * Here a delicate locking dance is performed. Current thread
3164                  * holds a reference to a page, but has to own it before it
3165                  * can be placed into queue. Owning implies waiting, so
3166                  * radix-tree lock is to be released. After a wait one has to
3167                  * check that pages weren't truncated (cl_page_own() returns
3168                  * error in the latter case).
3169                  */
3170                 spin_unlock(&osc->oo_tree_lock);
3171                 tree_lock = false;
3172
3173                 for (i = 0; i < j; ++i) {
3174                         ops = pvec[i];
3175                         if (res == CLP_GANG_OKAY)
3176                                 res = (*cb)(env, io, ops, cbdata);
3177
3178                         page = ops->ops_cl.cpl_page;
3179                         lu_ref_del(&page->cp_reference, "gang_lookup", current);
3180                         cl_pagevec_put(env, page, pagevec);
3181                 }
3182                 pagevec_release(pagevec);
3183
3184                 if (nr < OTI_PVEC_SIZE || end_of_region)
3185                         break;
3186
3187                 if (res == CLP_GANG_OKAY && need_resched())
3188                         res = CLP_GANG_RESCHED;
3189                 if (res != CLP_GANG_OKAY)
3190                         break;
3191
3192                 spin_lock(&osc->oo_tree_lock);
3193                 tree_lock = true;
3194         }
3195         if (tree_lock)
3196                 spin_unlock(&osc->oo_tree_lock);
3197         RETURN(res);
3198 }
3199 EXPORT_SYMBOL(osc_page_gang_lookup);
3200
3201 /**
3202  * Check if page @page is covered by an extra lock or discard it.
3203  */
3204 static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3205                                 struct osc_page *ops, void *cbdata)
3206 {
3207         struct osc_thread_info *info = osc_env_info(env);
3208         struct osc_object *osc = cbdata;
3209         pgoff_t index;
3210
3211         index = osc_index(ops);
3212         if (index >= info->oti_fn_index) {
3213                 struct ldlm_lock *tmp;
3214                 struct cl_page *page = ops->ops_cl.cpl_page;
3215
3216                 /* refresh non-overlapped index */
3217                 tmp = osc_dlmlock_at_pgoff(env, osc, index,
3218                                            OSC_DAP_FL_TEST_LOCK);
3219                 if (tmp != NULL) {
3220                         __u64 end = tmp->l_policy_data.l_extent.end;
3221                         /* Cache the first-non-overlapped index so as to skip
3222                          * all pages within [index, oti_fn_index). This is safe
3223                          * because if tmp lock is canceled, it will discard
3224                          * these pages. */
3225                         info->oti_fn_index = cl_index(osc2cl(osc), end + 1);
3226                         if (end == OBD_OBJECT_EOF)
3227                                 info->oti_fn_index = CL_PAGE_EOF;
3228                         LDLM_LOCK_PUT(tmp);
3229                 } else if (cl_page_own(env, io, page) == 0) {
3230                         /* discard the page */
3231                         cl_page_discard(env, io, page);
3232                         cl_page_disown(env, io, page);
3233                 } else {
3234                         LASSERT(page->cp_state == CPS_FREEING);
3235                 }
3236         }
3237
3238         info->oti_next_index = index + 1;
3239         return CLP_GANG_OKAY;
3240 }
3241
3242 int osc_discard_cb(const struct lu_env *env, struct cl_io *io,
3243                    struct osc_page *ops, void *cbdata)
3244 {
3245         struct osc_thread_info *info = osc_env_info(env);
3246         struct cl_page *page = ops->ops_cl.cpl_page;
3247
3248         /* page is top page. */
3249         info->oti_next_index = osc_index(ops) + 1;
3250         if (cl_page_own(env, io, page) == 0) {
3251                 if (!ergo(page->cp_type == CPT_CACHEABLE,
3252                           !PageDirty(cl_page_vmpage(page))))
3253                         CL_PAGE_DEBUG(D_ERROR, env, page,
3254                                         "discard dirty page?\n");
3255
3256                 /* discard the page */
3257                 cl_page_discard(env, io, page);
3258                 cl_page_disown(env, io, page);
3259         } else {
3260                 LASSERT(page->cp_state == CPS_FREEING);
3261         }
3262
3263         return CLP_GANG_OKAY;
3264 }
3265 EXPORT_SYMBOL(osc_discard_cb);
3266
3267 /**
3268  * Discard pages protected by the given lock. This function traverses radix
3269  * tree to find all covering pages and discard them. If a page is being covered
3270  * by other locks, it should remain in cache.
3271  *
3272  * If error happens on any step, the process continues anyway (the reasoning
3273  * behind this being that lock cancellation cannot be delayed indefinitely).
3274  */
3275 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3276                            pgoff_t start, pgoff_t end, bool discard)
3277 {
3278         struct osc_thread_info *info = osc_env_info(env);
3279         struct cl_io *io = osc_env_thread_io(env);
3280         osc_page_gang_cbt cb;
3281         int res;
3282         int result;
3283
3284         ENTRY;
3285
3286         io->ci_obj = cl_object_top(osc2cl(osc));
3287         io->ci_ignore_layout = 1;
3288         result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3289         if (result != 0)
3290                 GOTO(out, result);
3291
3292         cb = discard ? osc_discard_cb : check_and_discard_cb;
3293         info->oti_fn_index = info->oti_next_index = start;
3294         do {
3295                 res = osc_page_gang_lookup(env, io, osc,
3296                                            info->oti_next_index, end, cb, osc);
3297                 if (info->oti_next_index > end)
3298                         break;
3299
3300                 if (res == CLP_GANG_RESCHED)
3301                         cond_resched();
3302         } while (res != CLP_GANG_OKAY);
3303 out:
3304         cl_io_fini(env, io);
3305         RETURN(result);
3306 }
3307
3308
3309 /** @} osc */