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, 2015, 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 "osc_cl_internal.h"
  41 #include "osc_internal.h"
  42
  43 static int extent_debug; /* set it to be true for more debug */
  44
  45 static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
  46 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
  47                            enum osc_extent_state state);
  48 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
  49                               struct osc_async_page *oap, int sent, int rc);
  50 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
  51                           int cmd);
  52 static int osc_refresh_count(const struct lu_env *env,
  53                              struct osc_async_page *oap, int cmd);
  54 static int osc_io_unplug_async(const struct lu_env *env,
  55                                struct client_obd *cli, struct osc_object *osc);
  56 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
  57                            unsigned int lost_grant, unsigned int dirty_grant);
  58
  59 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
  60                                   const char *func, int line);
  61 #define osc_extent_tree_dump(lvl, obj) \
  62         osc_extent_tree_dump0(lvl, obj, __func__, __LINE__)
  63
  64 static void osc_unreserve_grant(struct client_obd *cli, unsigned int reserved,
  65                                 unsigned int unused);
  66
  67 /** \addtogroup osc
  68  *  @{
  69  */
  70
  71 /* ------------------ osc extent ------------------ */
  72 static inline char *ext_flags(struct osc_extent *ext, char *flags)
  73 {
  74         char *buf = flags;
  75         *buf++ = ext->oe_rw ? 'r' : 'w';
  76         if (ext->oe_intree)
  77                 *buf++ = 'i';
  78         if (ext->oe_sync)
  79                 *buf++ = 'S';
  80         if (ext->oe_srvlock)
  81                 *buf++ = 's';
  82         if (ext->oe_hp)
  83                 *buf++ = 'h';
  84         if (ext->oe_urgent)
  85                 *buf++ = 'u';
  86         if (ext->oe_memalloc)
  87                 *buf++ = 'm';
  88         if (ext->oe_trunc_pending)
  89                 *buf++ = 't';
  90         if (ext->oe_fsync_wait)
  91                 *buf++ = 'Y';
  92         *buf = 0;
  93         return flags;
  94 }
  95
  96 static inline char list_empty_marker(struct list_head *list)
  97 {
  98         return list_empty(list) ? '-' : '+';
  99 }
 100
 101 #define EXTSTR       "[%lu -> %lu/%lu]"
 102 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
 103 static const char *oes_strings[] = {
 104         "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
 105
 106 #define OSC_EXTENT_DUMP(lvl, extent, fmt, ...) do {                           \
 107         struct osc_extent *__ext = (extent);                                  \
 108         char __buf[16];                                                       \
 109                                                                               \
 110         CDEBUG(lvl,                                                           \
 111                 "extent %p@{" EXTSTR ", "                                     \
 112                 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt,          \
 113                 /* ----- extent part 0 ----- */                               \
 114                 __ext, EXTPARA(__ext),                                        \
 115                 /* ----- part 1 ----- */                                      \
 116                 atomic_read(&__ext->oe_refc),                         \
 117                 atomic_read(&__ext->oe_users),                        \
 118                 list_empty_marker(&__ext->oe_link),                           \
 119                 oes_strings[__ext->oe_state], ext_flags(__ext, __buf),        \
 120                 __ext->oe_obj,                                                \
 121                 /* ----- part 2 ----- */                                      \
 122                 __ext->oe_grants, __ext->oe_nr_pages,                         \
 123                 list_empty_marker(&__ext->oe_pages),                          \
 124                 waitqueue_active(&__ext->oe_waitq) ? '+' : '-',               \
 125                 __ext->oe_dlmlock, __ext->oe_mppr, __ext->oe_owner,           \
 126                 /* ----- part 4 ----- */                                      \
 127                 ## __VA_ARGS__);                                              \
 128         if (lvl == D_ERROR && __ext->oe_dlmlock != NULL)                      \
 129                 LDLM_ERROR(__ext->oe_dlmlock, "extent: %p", __ext);           \
 130         else                                                                  \
 131                 LDLM_DEBUG(__ext->oe_dlmlock, "extent: %p", __ext);           \
 132 } while (0)
 133
 134 #undef EASSERTF
 135 #define EASSERTF(expr, ext, fmt, args...) do {                          \
 136         if (!(expr)) {                                                  \
 137                 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args);           \
 138                 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj);           \
 139                 LASSERT(expr);                                          \
 140         }                                                               \
 141 } while (0)
 142
 143 #undef EASSERT
 144 #define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
 145
 146 static inline struct osc_extent *rb_extent(struct rb_node *n)
 147 {
 148         if (n == NULL)
 149                 return NULL;
 150
 151         return container_of(n, struct osc_extent, oe_node);
 152 }
 153
 154 static inline struct osc_extent *next_extent(struct osc_extent *ext)
 155 {
 156         if (ext == NULL)
 157                 return NULL;
 158
 159         LASSERT(ext->oe_intree);
 160         return rb_extent(rb_next(&ext->oe_node));
 161 }
 162
 163 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
 164 {
 165         if (ext == NULL)
 166                 return NULL;
 167
 168         LASSERT(ext->oe_intree);
 169         return rb_extent(rb_prev(&ext->oe_node));
 170 }
 171
 172 static inline struct osc_extent *first_extent(struct osc_object *obj)
 173 {
 174         return rb_extent(rb_first(&obj->oo_root));
 175 }
 176
 177 /* object must be locked by caller. */
 178 static int osc_extent_sanity_check0(struct osc_extent *ext,
 179                                     const char *func, const int line)
 180 {
 181         struct osc_object *obj = ext->oe_obj;
 182         struct osc_async_page *oap;
 183         size_t page_count;
 184         int rc = 0;
 185
 186         if (!osc_object_is_locked(obj))
 187                 GOTO(out, rc = 9);
 188
 189         if (ext->oe_state >= OES_STATE_MAX)
 190                 GOTO(out, rc = 10);
 191
 192         if (atomic_read(&ext->oe_refc) <= 0)
 193                 GOTO(out, rc = 20);
 194
 195         if (atomic_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
 196                 GOTO(out, rc = 30);
 197
 198         switch (ext->oe_state) {
 199         case OES_INV:
 200                 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
 201                         GOTO(out, rc = 35);
 202                 GOTO(out, rc = 0);
 203                 break;
 204         case OES_ACTIVE:
 205                 if (atomic_read(&ext->oe_users) == 0)
 206                         GOTO(out, rc = 40);
 207                 if (ext->oe_hp)
 208                         GOTO(out, rc = 50);
 209                 if (ext->oe_fsync_wait && !ext->oe_urgent)
 210                         GOTO(out, rc = 55);
 211                 break;
 212         case OES_CACHE:
 213                 if (ext->oe_grants == 0)
 214                         GOTO(out, rc = 60);
 215                 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
 216                         GOTO(out, rc = 65);
 217         default:
 218                 if (atomic_read(&ext->oe_users) > 0)
 219                         GOTO(out, rc = 70);
 220         }
 221
 222         if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
 223                 GOTO(out, rc = 80);
 224
 225         if (ext->oe_sync && ext->oe_grants > 0)
 226                 GOTO(out, rc = 90);
 227
 228         if (ext->oe_dlmlock != NULL && !ldlm_is_failed(ext->oe_dlmlock)) {
 229                 struct ldlm_extent *extent;
 230
 231                 extent = &ext->oe_dlmlock->l_policy_data.l_extent;
 232                 if (!(extent->start <= cl_offset(osc2cl(obj), ext->oe_start) &&
 233                       extent->end   >= cl_offset(osc2cl(obj), ext->oe_max_end)))
 234                         GOTO(out, rc = 100);
 235
 236                 if (!(ext->oe_dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP)))
 237                         GOTO(out, rc = 102);
 238         }
 239
 240         if (ext->oe_nr_pages > ext->oe_mppr)
 241                 GOTO(out, rc = 105);
 242
 243         /* Do not verify page list if extent is in RPC. This is because an
 244          * in-RPC extent is supposed to be exclusively accessible w/o lock. */
 245         if (ext->oe_state > OES_CACHE)
 246                 GOTO(out, rc = 0);
 247
 248         if (!extent_debug)
 249                 GOTO(out, rc = 0);
 250
 251         page_count = 0;
 252         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
 253                 pgoff_t index = osc_index(oap2osc(oap));
 254                 ++page_count;
 255                 if (index > ext->oe_end || index < ext->oe_start)
 256                         GOTO(out, rc = 110);
 257         }
 258         if (page_count != ext->oe_nr_pages)
 259                 GOTO(out, rc = 120);
 260
 261 out:
 262         if (rc != 0)
 263                 OSC_EXTENT_DUMP(D_ERROR, ext,
 264                                 "%s:%d sanity check %p failed with rc = %d\n",
 265                                 func, line, ext, rc);
 266         return rc;
 267 }
 268
 269 #define sanity_check_nolock(ext) \
 270         osc_extent_sanity_check0(ext, __func__, __LINE__)
 271
 272 #define sanity_check(ext) ({                                                   \
 273         int __res;                                                             \
 274         osc_object_lock((ext)->oe_obj);                                        \
 275         __res = sanity_check_nolock(ext);                                      \
 276         osc_object_unlock((ext)->oe_obj);                                      \
 277         __res;                                                                 \
 278 })
 279
 280
 281 /**
 282  * sanity check - to make sure there is no overlapped extent in the tree.
 283  */
 284 static int osc_extent_is_overlapped(struct osc_object *obj,
 285                                     struct osc_extent *ext)
 286 {
 287         struct osc_extent *tmp;
 288
 289         LASSERT(osc_object_is_locked(obj));
 290
 291         if (!extent_debug)
 292                 return 0;
 293
 294         for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
 295                 if (tmp == ext)
 296                         continue;
 297                 if (tmp->oe_end >= ext->oe_start &&
 298                     tmp->oe_start <= ext->oe_end)
 299                         return 1;
 300         }
 301         return 0;
 302 }
 303
 304 static void osc_extent_state_set(struct osc_extent *ext, int state)
 305 {
 306         LASSERT(osc_object_is_locked(ext->oe_obj));
 307         LASSERT(state >= OES_INV && state < OES_STATE_MAX);
 308
 309         /* Never try to sanity check a state changing extent :-) */
 310         /* LASSERT(sanity_check_nolock(ext) == 0); */
 311
 312         /* TODO: validate the state machine */
 313         ext->oe_state = state;
 314         wake_up_all(&ext->oe_waitq);
 315 }
 316
 317 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
 318 {
 319         struct osc_extent *ext;
 320
 321         OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_NOFS);
 322         if (ext == NULL)
 323                 return NULL;
 324
 325         RB_CLEAR_NODE(&ext->oe_node);
 326         ext->oe_obj = obj;
 327         cl_object_get(osc2cl(obj));
 328         atomic_set(&ext->oe_refc, 1);
 329         atomic_set(&ext->oe_users, 0);
 330         INIT_LIST_HEAD(&ext->oe_link);
 331         ext->oe_state = OES_INV;
 332         INIT_LIST_HEAD(&ext->oe_pages);
 333         init_waitqueue_head(&ext->oe_waitq);
 334         ext->oe_dlmlock = NULL;
 335
 336         return ext;
 337 }
 338
 339 static void osc_extent_free(struct osc_extent *ext)
 340 {
 341         OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
 342 }
 343
 344 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
 345 {
 346         LASSERT(atomic_read(&ext->oe_refc) >= 0);
 347         atomic_inc(&ext->oe_refc);
 348         return ext;
 349 }
 350
 351 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
 352 {
 353         LASSERT(atomic_read(&ext->oe_refc) > 0);
 354         if (atomic_dec_and_test(&ext->oe_refc)) {
 355                 LASSERT(list_empty(&ext->oe_link));
 356                 LASSERT(atomic_read(&ext->oe_users) == 0);
 357                 LASSERT(ext->oe_state == OES_INV);
 358                 LASSERT(!ext->oe_intree);
 359
 360                 if (ext->oe_dlmlock != NULL) {
 361                         lu_ref_add(&ext->oe_dlmlock->l_reference,
 362                                    "osc_extent", ext);
 363                         LDLM_LOCK_PUT(ext->oe_dlmlock);
 364                         ext->oe_dlmlock = NULL;
 365                 }
 366                 cl_object_put(env, osc2cl(ext->oe_obj));
 367                 osc_extent_free(ext);
 368         }
 369 }
 370
 371 /**
 372  * osc_extent_put_trust() is a special version of osc_extent_put() when
 373  * it's known that the caller is not the last user. This is to address the
 374  * problem of lacking of lu_env ;-).
 375  */
 376 static void osc_extent_put_trust(struct osc_extent *ext)
 377 {
 378         LASSERT(atomic_read(&ext->oe_refc) > 1);
 379         LASSERT(osc_object_is_locked(ext->oe_obj));
 380         atomic_dec(&ext->oe_refc);
 381 }
 382
 383 /**
 384  * Return the extent which includes pgoff @index, or return the greatest
 385  * previous extent in the tree.
 386  */
 387 static struct osc_extent *osc_extent_search(struct osc_object *obj,
 388                                             pgoff_t index)
 389 {
 390         struct rb_node    *n = obj->oo_root.rb_node;
 391         struct osc_extent *tmp, *p = NULL;
 392
 393         LASSERT(osc_object_is_locked(obj));
 394         while (n != NULL) {
 395                 tmp = rb_extent(n);
 396                 if (index < tmp->oe_start) {
 397                         n = n->rb_left;
 398                 } else if (index > tmp->oe_end) {
 399                         p = rb_extent(n);
 400                         n = n->rb_right;
 401                 } else {
 402                         return tmp;
 403                 }
 404         }
 405         return p;
 406 }
 407
 408 /*
 409  * Return the extent covering @index, otherwise return NULL.
 410  * caller must have held object lock.
 411  */
 412 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
 413                                             pgoff_t index)
 414 {
 415         struct osc_extent *ext;
 416
 417         ext = osc_extent_search(obj, index);
 418         if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
 419                 return osc_extent_get(ext);
 420         return NULL;
 421 }
 422
 423 /* caller must have held object lock. */
 424 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
 425 {
 426         struct rb_node   **n      = &obj->oo_root.rb_node;
 427         struct rb_node    *parent = NULL;
 428         struct osc_extent *tmp;
 429
 430         LASSERT(ext->oe_intree == 0);
 431         LASSERT(ext->oe_obj == obj);
 432         LASSERT(osc_object_is_locked(obj));
 433         while (*n != NULL) {
 434                 tmp = rb_extent(*n);
 435                 parent = *n;
 436
 437                 if (ext->oe_end < tmp->oe_start)
 438                         n = &(*n)->rb_left;
 439                 else if (ext->oe_start > tmp->oe_end)
 440                         n = &(*n)->rb_right;
 441                 else
 442                         EASSERTF(0, tmp, EXTSTR"\n", EXTPARA(ext));
 443         }
 444         rb_link_node(&ext->oe_node, parent, n);
 445         rb_insert_color(&ext->oe_node, &obj->oo_root);
 446         osc_extent_get(ext);
 447         ext->oe_intree = 1;
 448 }
 449
 450 /* caller must have held object lock. */
 451 static void osc_extent_erase(struct osc_extent *ext)
 452 {
 453         struct osc_object *obj = ext->oe_obj;
 454         LASSERT(osc_object_is_locked(obj));
 455         if (ext->oe_intree) {
 456                 rb_erase(&ext->oe_node, &obj->oo_root);
 457                 ext->oe_intree = 0;
 458                 /* rbtree held a refcount */
 459                 osc_extent_put_trust(ext);
 460         }
 461 }
 462
 463 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
 464 {
 465         struct osc_object *obj = ext->oe_obj;
 466
 467         LASSERT(osc_object_is_locked(obj));
 468         LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
 469         if (ext->oe_state == OES_CACHE) {
 470                 osc_extent_state_set(ext, OES_ACTIVE);
 471                 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
 472         }
 473         atomic_inc(&ext->oe_users);
 474         list_del_init(&ext->oe_link);
 475         return osc_extent_get(ext);
 476 }
 477
 478 static void __osc_extent_remove(struct osc_extent *ext)
 479 {
 480         LASSERT(osc_object_is_locked(ext->oe_obj));
 481         LASSERT(list_empty(&ext->oe_pages));
 482         osc_extent_erase(ext);
 483         list_del_init(&ext->oe_link);
 484         osc_extent_state_set(ext, OES_INV);
 485         OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
 486 }
 487
 488 static void osc_extent_remove(struct osc_extent *ext)
 489 {
 490         struct osc_object *obj = ext->oe_obj;
 491
 492         osc_object_lock(obj);
 493         __osc_extent_remove(ext);
 494         osc_object_unlock(obj);
 495 }
 496
 497 /**
 498  * This function is used to merge extents to get better performance. It checks
 499  * if @cur and @victim are contiguous at block level.
 500  */
 501 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
 502                             struct osc_extent *victim)
 503 {
 504         struct osc_object       *obj = cur->oe_obj;
 505         struct client_obd       *cli = osc_cli(obj);
 506         pgoff_t                  chunk_start;
 507         pgoff_t                  chunk_end;
 508         int                      ppc_bits;
 509
 510         LASSERT(cur->oe_state == OES_CACHE);
 511         LASSERT(osc_object_is_locked(obj));
 512         if (victim == NULL)
 513                 return -EINVAL;
 514
 515         if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
 516                 return -EBUSY;
 517
 518         if (cur->oe_max_end != victim->oe_max_end)
 519                 return -ERANGE;
 520
 521         LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
 522         ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_SHIFT;
 523         chunk_start = cur->oe_start >> ppc_bits;
 524         chunk_end   = cur->oe_end   >> ppc_bits;
 525         if (chunk_start   != (victim->oe_end >> ppc_bits) + 1 &&
 526             chunk_end + 1 != victim->oe_start >> ppc_bits)
 527                 return -ERANGE;
 528
 529         /* overall extent size should not exceed the max supported limit
 530          * reported by the server */
 531         if (cur->oe_end - cur->oe_start + 1 +
 532             victim->oe_end - victim->oe_start + 1 > cli->cl_max_extent_pages)
 533                 return -ERANGE;
 534
 535         OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
 536
 537         cur->oe_start     = min(cur->oe_start, victim->oe_start);
 538         cur->oe_end       = max(cur->oe_end,   victim->oe_end);
 539         /* per-extent tax should be accounted only once for the whole extent */
 540         cur->oe_grants   += victim->oe_grants - cli->cl_grant_extent_tax;
 541         cur->oe_nr_pages += victim->oe_nr_pages;
 542         /* only the following bits are needed to merge */
 543         cur->oe_urgent   |= victim->oe_urgent;
 544         cur->oe_memalloc |= victim->oe_memalloc;
 545         list_splice_init(&victim->oe_pages, &cur->oe_pages);
 546         list_del_init(&victim->oe_link);
 547         victim->oe_nr_pages = 0;
 548
 549         osc_extent_get(victim);
 550         __osc_extent_remove(victim);
 551         osc_extent_put(env, victim);
 552
 553         OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
 554         return 0;
 555 }
 556
 557 /**
 558  * Drop user count of osc_extent, and unplug IO asynchronously.
 559  */
 560 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
 561 {
 562         struct osc_object *obj = ext->oe_obj;
 563         struct client_obd *cli = osc_cli(obj);
 564         int rc = 0;
 565         ENTRY;
 566
 567         LASSERT(atomic_read(&ext->oe_users) > 0);
 568         LASSERT(sanity_check(ext) == 0);
 569         LASSERT(ext->oe_grants > 0);
 570
 571         if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
 572                 LASSERT(ext->oe_state == OES_ACTIVE);
 573                 if (ext->oe_trunc_pending) {
 574                         /* a truncate process is waiting for this extent.
 575                          * This may happen due to a race, check
 576                          * osc_cache_truncate_start(). */
 577                         osc_extent_state_set(ext, OES_TRUNC);
 578                         ext->oe_trunc_pending = 0;
 579                 } else {
 580                         int grant = 0;
 581
 582                         osc_extent_state_set(ext, OES_CACHE);
 583                         osc_update_pending(obj, OBD_BRW_WRITE,
 584                                            ext->oe_nr_pages);
 585
 586                         /* try to merge the previous and next extent. */
 587                         if (osc_extent_merge(env, ext, prev_extent(ext)) == 0)
 588                                 grant += cli->cl_grant_extent_tax;
 589                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 590                                 grant += cli->cl_grant_extent_tax;
 591                         if (grant > 0)
 592                                 osc_unreserve_grant(cli, 0, grant);
 593
 594                         if (ext->oe_urgent)
 595                                 list_move_tail(&ext->oe_link,
 596                                                &obj->oo_urgent_exts);
 597                         else if (ext->oe_nr_pages == ext->oe_mppr) {
 598                                 list_move_tail(&ext->oe_link,
 599                                                &obj->oo_full_exts);
 600                         }
 601                 }
 602                 osc_object_unlock(obj);
 603
 604                 osc_io_unplug_async(env, cli, obj);
 605         }
 606         osc_extent_put(env, ext);
 607         RETURN(rc);
 608 }
 609
 610 static inline int overlapped(struct osc_extent *ex1, struct osc_extent *ex2)
 611 {
 612         return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
 613 }
 614
 615 /**
 616  * Find or create an extent which includes @index, core function to manage
 617  * extent tree.
 618  */
 619 static struct osc_extent *osc_extent_find(const struct lu_env *env,
 620                                           struct osc_object *obj, pgoff_t index,
 621                                           unsigned int *grants)
 622 {
 623         struct client_obd *cli = osc_cli(obj);
 624         struct osc_lock   *olck;
 625         struct cl_lock_descr *descr;
 626         struct osc_extent *cur;
 627         struct osc_extent *ext;
 628         struct osc_extent *conflict = NULL;
 629         struct osc_extent *found = NULL;
 630         pgoff_t    chunk;
 631         pgoff_t    max_end;
 632         unsigned int max_pages; /* max_pages_per_rpc */
 633         unsigned int chunksize;
 634         int        ppc_bits; /* pages per chunk bits */
 635         pgoff_t    chunk_mask;
 636         int        rc;
 637         ENTRY;
 638
 639         cur = osc_extent_alloc(obj);
 640         if (cur == NULL)
 641                 RETURN(ERR_PTR(-ENOMEM));
 642
 643         olck = osc_env_io(env)->oi_write_osclock;
 644         LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
 645         LASSERT(olck->ols_state == OLS_GRANTED);
 646
 647         descr = &olck->ols_cl.cls_lock->cll_descr;
 648         LASSERT(descr->cld_mode >= CLM_WRITE);
 649
 650         LASSERTF(cli->cl_chunkbits >= PAGE_SHIFT,
 651                  "chunkbits: %u\n", cli->cl_chunkbits);
 652         ppc_bits   = cli->cl_chunkbits - PAGE_SHIFT;
 653         chunk_mask = ~((1 << ppc_bits) - 1);
 654         chunksize  = 1 << cli->cl_chunkbits;
 655         chunk      = index >> ppc_bits;
 656
 657         /* align end to RPC edge. */
 658         max_pages = cli->cl_max_pages_per_rpc;
 659         if ((max_pages & ~chunk_mask) != 0) {
 660                 CERROR("max_pages: %#x chunkbits: %u chunk_mask: %#lx\n",
 661                        max_pages, cli->cl_chunkbits, chunk_mask);
 662                 RETURN(ERR_PTR(-EINVAL));
 663         }
 664         max_end = index - (index % max_pages) + max_pages - 1;
 665         max_end = min_t(pgoff_t, max_end, descr->cld_end);
 666
 667         /* initialize new extent by parameters so far */
 668         cur->oe_max_end = max_end;
 669         cur->oe_start   = index & chunk_mask;
 670         cur->oe_end     = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
 671         if (cur->oe_start < descr->cld_start)
 672                 cur->oe_start = descr->cld_start;
 673         if (cur->oe_end > max_end)
 674                 cur->oe_end = max_end;
 675         cur->oe_grants  = 0;
 676         cur->oe_mppr    = max_pages;
 677         if (olck->ols_dlmlock != NULL) {
 678                 LASSERT(olck->ols_hold);
 679                 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
 680                 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
 681         }
 682
 683         /* grants has been allocated by caller */
 684         LASSERTF(*grants >= chunksize + cli->cl_grant_extent_tax,
 685                  "%u/%u/%u.\n", *grants, chunksize, cli->cl_grant_extent_tax);
 686         LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
 687                  EXTPARA(cur));
 688
 689 restart:
 690         osc_object_lock(obj);
 691         ext = osc_extent_search(obj, cur->oe_start);
 692         if (ext == NULL)
 693                 ext = first_extent(obj);
 694         while (ext != NULL) {
 695                 pgoff_t ext_chk_start = ext->oe_start >> ppc_bits;
 696                 pgoff_t ext_chk_end   = ext->oe_end   >> ppc_bits;
 697
 698                 LASSERT(sanity_check_nolock(ext) == 0);
 699                 if (chunk > ext_chk_end + 1)
 700                         break;
 701
 702                 /* if covering by different locks, no chance to match */
 703                 if (olck->ols_dlmlock != ext->oe_dlmlock) {
 704                         EASSERTF(!overlapped(ext, cur), ext,
 705                                  EXTSTR"\n", EXTPARA(cur));
 706
 707                         ext = next_extent(ext);
 708                         continue;
 709                 }
 710
 711                 /* discontiguous chunks? */
 712                 if (chunk + 1 < ext_chk_start) {
 713                         ext = next_extent(ext);
 714                         continue;
 715                 }
 716
 717                 /* ok, from now on, ext and cur have these attrs:
 718                  * 1. covered by the same lock
 719                  * 2. contiguous at chunk level or overlapping. */
 720
 721                 if (overlapped(ext, cur)) {
 722                         /* cur is the minimum unit, so overlapping means
 723                          * full contain. */
 724                         EASSERTF((ext->oe_start <= cur->oe_start &&
 725                                   ext->oe_end >= cur->oe_end),
 726                                  ext, EXTSTR"\n", EXTPARA(cur));
 727
 728                         if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
 729                                 /* for simplicity, we wait for this extent to
 730                                  * finish before going forward. */
 731                                 conflict = osc_extent_get(ext);
 732                                 break;
 733                         }
 734
 735                         found = osc_extent_hold(ext);
 736                         break;
 737                 }
 738
 739                 /* non-overlapped extent */
 740                 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
 741                         /* we can't do anything for a non OES_CACHE extent, or
 742                          * if there is someone waiting for this extent to be
 743                          * flushed, try next one. */
 744                         ext = next_extent(ext);
 745                         continue;
 746                 }
 747
 748                 /* check if they belong to the same rpc slot before trying to
 749                  * merge. the extents are not overlapped and contiguous at
 750                  * chunk level to get here. */
 751                 if (ext->oe_max_end != max_end) {
 752                         /* if they don't belong to the same RPC slot or
 753                          * max_pages_per_rpc has ever changed, do not merge. */
 754                         ext = next_extent(ext);
 755                         continue;
 756                 }
 757
 758                 /* check whether maximum extent size will be hit */
 759                 if ((ext_chk_end - ext_chk_start + 1 + 1) << ppc_bits >
 760                     cli->cl_max_extent_pages) {
 761                         ext = next_extent(ext);
 762                         continue;
 763                 }
 764
 765                 /* it's required that an extent must be contiguous at chunk
 766                  * level so that we know the whole extent is covered by grant
 767                  * (the pages in the extent are NOT required to be contiguous).
 768                  * Otherwise, it will be too much difficult to know which
 769                  * chunks have grants allocated. */
 770
 771                 /* try to do front merge - extend ext's start */
 772                 if (chunk + 1 == ext_chk_start) {
 773                         /* ext must be chunk size aligned */
 774                         EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
 775
 776                         /* pull ext's start back to cover cur */
 777                         ext->oe_start   = cur->oe_start;
 778                         ext->oe_grants += chunksize;
 779                         *grants -= chunksize;
 780
 781                         found = osc_extent_hold(ext);
 782                 } else if (chunk == ext_chk_end + 1) {
 783                         /* rear merge */
 784                         ext->oe_end     = cur->oe_end;
 785                         ext->oe_grants += chunksize;
 786                         *grants -= chunksize;
 787
 788                         /* try to merge with the next one because we just fill
 789                          * in a gap */
 790                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 791                                 /* we can save extent tax from next extent */
 792                                 *grants += cli->cl_grant_extent_tax;
 793
 794                         found = osc_extent_hold(ext);
 795                 }
 796                 if (found != NULL)
 797                         break;
 798
 799                 ext = next_extent(ext);
 800         }
 801
 802         osc_extent_tree_dump(D_CACHE, obj);
 803         if (found != NULL) {
 804                 LASSERT(conflict == NULL);
 805                 if (!IS_ERR(found)) {
 806                         LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
 807                         OSC_EXTENT_DUMP(D_CACHE, found,
 808                                         "found caching ext for %lu.\n", index);
 809                 }
 810         } else if (conflict == NULL) {
 811                 /* create a new extent */
 812                 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
 813                 cur->oe_grants = chunksize + cli->cl_grant_extent_tax;
 814                 *grants -= cur->oe_grants;
 815                 LASSERT(*grants >= 0);
 816
 817                 cur->oe_state = OES_CACHE;
 818                 found = osc_extent_hold(cur);
 819                 osc_extent_insert(obj, cur);
 820                 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
 821                                 index, descr->cld_end);
 822         }
 823         osc_object_unlock(obj);
 824
 825         if (conflict != NULL) {
 826                 LASSERT(found == NULL);
 827
 828                 /* waiting for IO to finish. Please notice that it's impossible
 829                  * to be an OES_TRUNC extent. */
 830                 rc = osc_extent_wait(env, conflict, OES_INV);
 831                 osc_extent_put(env, conflict);
 832                 conflict = NULL;
 833                 if (rc < 0)
 834                         GOTO(out, found = ERR_PTR(rc));
 835
 836                 goto restart;
 837         }
 838         EXIT;
 839
 840 out:
 841         osc_extent_put(env, cur);
 842         LASSERT(*grants >= 0);
 843         return found;
 844 }
 845
 846 /**
 847  * Called when IO is finished to an extent.
 848  */
 849 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
 850                       int sent, int rc)
 851 {
 852         struct client_obd *cli = osc_cli(ext->oe_obj);
 853         struct osc_async_page *oap;
 854         struct osc_async_page *tmp;
 855         int nr_pages = ext->oe_nr_pages;
 856         int lost_grant = 0;
 857         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
 858         loff_t last_off = 0;
 859         int last_count = -1;
 860         ENTRY;
 861
 862         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
 863
 864         ext->oe_rc = rc ?: ext->oe_nr_pages;
 865         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
 866
 867         osc_lru_add_batch(cli, &ext->oe_pages);
 868         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 869                                      oap_pending_item) {
 870                 list_del_init(&oap->oap_rpc_item);
 871                 list_del_init(&oap->oap_pending_item);
 872                 if (last_off <= oap->oap_obj_off) {
 873                         last_off = oap->oap_obj_off;
 874                         last_count = oap->oap_count;
 875                 }
 876
 877                 --ext->oe_nr_pages;
 878                 osc_ap_completion(env, cli, oap, sent, rc);
 879         }
 880         EASSERT(ext->oe_nr_pages == 0, ext);
 881
 882         if (!sent) {
 883                 lost_grant = ext->oe_grants;
 884         } else if (blocksize < PAGE_SIZE &&
 885                    last_count != PAGE_SIZE) {
 886                 /* For short writes we shouldn't count parts of pages that
 887                  * span a whole chunk on the OST side, or our accounting goes
 888                  * wrong.  Should match the code in filter_grant_check. */
 889                 int offset = last_off & ~PAGE_MASK;
 890                 int count = last_count + (offset & (blocksize - 1));
 891                 int end = (offset + last_count) & (blocksize - 1);
 892                 if (end)
 893                         count += blocksize - end;
 894
 895                 lost_grant = PAGE_SIZE - count;
 896         }
 897         if (ext->oe_grants > 0)
 898                 osc_free_grant(cli, nr_pages, lost_grant, ext->oe_grants);
 899
 900         osc_extent_remove(ext);
 901         /* put the refcount for RPC */
 902         osc_extent_put(env, ext);
 903         RETURN(0);
 904 }
 905
 906 static int extent_wait_cb(struct osc_extent *ext, enum osc_extent_state state)
 907 {
 908         int ret;
 909
 910         osc_object_lock(ext->oe_obj);
 911         ret = ext->oe_state == state;
 912         osc_object_unlock(ext->oe_obj);
 913
 914         return ret;
 915 }
 916
 917 /**
 918  * Wait for the extent's state to become @state.
 919  */
 920 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
 921                            enum osc_extent_state state)
 922 {
 923         struct osc_object *obj = ext->oe_obj;
 924         struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
 925                                                   LWI_ON_SIGNAL_NOOP, NULL);
 926         int rc = 0;
 927         ENTRY;
 928
 929         osc_object_lock(obj);
 930         LASSERT(sanity_check_nolock(ext) == 0);
 931         /* `Kick' this extent only if the caller is waiting for it to be
 932          * written out. */
 933         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
 934                 if (ext->oe_state == OES_ACTIVE) {
 935                         ext->oe_urgent = 1;
 936                 } else if (ext->oe_state == OES_CACHE) {
 937                         ext->oe_urgent = 1;
 938                         osc_extent_hold(ext);
 939                         rc = 1;
 940                 }
 941         }
 942         osc_object_unlock(obj);
 943         if (rc == 1)
 944                 osc_extent_release(env, ext);
 945
 946         /* wait for the extent until its state becomes @state */
 947         rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
 948         if (rc == -ETIMEDOUT) {
 949                 OSC_EXTENT_DUMP(D_ERROR, ext,
 950                         "%s: wait ext to %u timedout, recovery in progress?\n",
 951                         cli_name(osc_cli(obj)), state);
 952
 953                 lwi = LWI_INTR(NULL, NULL);
 954                 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state),
 955                                   &lwi);
 956         }
 957         if (rc == 0 && ext->oe_rc < 0)
 958                 rc = ext->oe_rc;
 959         RETURN(rc);
 960 }
 961
 962 /**
 963  * Discard pages with index greater than @size. If @ext is overlapped with
 964  * @size, then partial truncate happens.
 965  */
 966 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
 967                                 bool partial)
 968 {
 969         struct lu_env         *env;
 970         struct cl_io          *io;
 971         struct osc_object     *obj = ext->oe_obj;
 972         struct client_obd     *cli = osc_cli(obj);
 973         struct osc_async_page *oap;
 974         struct osc_async_page *tmp;
 975         int                    pages_in_chunk = 0;
 976         int                    ppc_bits    = cli->cl_chunkbits -
 977                                              PAGE_SHIFT;
 978         __u64                  trunc_chunk = trunc_index >> ppc_bits;
 979         int                    grants   = 0;
 980         int                    nr_pages = 0;
 981         int                    rc       = 0;
 982         __u16                  refcheck;
 983         ENTRY;
 984
 985         LASSERT(sanity_check(ext) == 0);
 986         LASSERT(ext->oe_state == OES_TRUNC);
 987         LASSERT(!ext->oe_urgent);
 988
 989         /* Request new lu_env.
 990          * We can't use that env from osc_cache_truncate_start() because
 991          * it's from lov_io_sub and not fully initialized. */
 992         env = cl_env_get(&refcheck);
 993         io  = &osc_env_info(env)->oti_io;
 994         io->ci_obj = cl_object_top(osc2cl(obj));
 995         io->ci_ignore_layout = 1;
 996         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
 997         if (rc < 0)
 998                 GOTO(out, rc);
 999
1000         /* discard all pages with index greater than trunc_index */
1001         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
1002                                      oap_pending_item) {
1003                 pgoff_t index = osc_index(oap2osc(oap));
1004                 struct cl_page  *page = oap2cl_page(oap);
1005
1006                 LASSERT(list_empty(&oap->oap_rpc_item));
1007
1008                 /* only discard the pages with their index greater than
1009                  * trunc_index, and ... */
1010                 if (index < trunc_index ||
1011                     (index == trunc_index && partial)) {
1012                         /* accounting how many pages remaining in the chunk
1013                          * so that we can calculate grants correctly. */
1014                         if (index >> ppc_bits == trunc_chunk)
1015                                 ++pages_in_chunk;
1016                         continue;
1017                 }
1018
1019                 list_del_init(&oap->oap_pending_item);
1020
1021                 cl_page_get(page);
1022                 lu_ref_add(&page->cp_reference, "truncate", current);
1023
1024                 if (cl_page_own(env, io, page) == 0) {
1025                         cl_page_discard(env, io, page);
1026                         cl_page_disown(env, io, page);
1027                 } else {
1028                         LASSERT(page->cp_state == CPS_FREEING);
1029                         LASSERT(0);
1030                 }
1031
1032                 lu_ref_del(&page->cp_reference, "truncate", current);
1033                 cl_page_put(env, page);
1034
1035                 --ext->oe_nr_pages;
1036                 ++nr_pages;
1037         }
1038         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1039                       ext->oe_nr_pages == 0),
1040                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1041
1042         osc_object_lock(obj);
1043         if (ext->oe_nr_pages == 0) {
1044                 LASSERT(pages_in_chunk == 0);
1045                 grants = ext->oe_grants;
1046                 ext->oe_grants = 0;
1047         } else { /* calculate how many grants we can free */
1048                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1049                 pgoff_t last_index;
1050
1051
1052                 /* if there is no pages in this chunk, we can also free grants
1053                  * for the last chunk */
1054                 if (pages_in_chunk == 0) {
1055                         /* if this is the 1st chunk and no pages in this chunk,
1056                          * ext->oe_nr_pages must be zero, so we should be in
1057                          * the other if-clause. */
1058                         LASSERT(trunc_chunk > 0);
1059                         --trunc_chunk;
1060                         ++chunks;
1061                 }
1062
1063                 /* this is what we can free from this extent */
1064                 grants          = chunks << cli->cl_chunkbits;
1065                 ext->oe_grants -= grants;
1066                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1067                 ext->oe_end     = min(last_index, ext->oe_max_end);
1068                 LASSERT(ext->oe_end >= ext->oe_start);
1069                 LASSERT(ext->oe_grants > 0);
1070         }
1071         osc_object_unlock(obj);
1072
1073         if (grants > 0 || nr_pages > 0)
1074                 osc_free_grant(cli, nr_pages, grants, grants);
1075
1076 out:
1077         cl_io_fini(env, io);
1078         cl_env_put(env, &refcheck);
1079         RETURN(rc);
1080 }
1081
1082 /**
1083  * This function is used to make the extent prepared for transfer.
1084  * A race with flusing page - ll_writepage() has to be handled cautiously.
1085  */
1086 static int osc_extent_make_ready(const struct lu_env *env,
1087                                  struct osc_extent *ext)
1088 {
1089         struct osc_async_page *oap;
1090         struct osc_async_page *last = NULL;
1091         struct osc_object *obj = ext->oe_obj;
1092         unsigned int page_count = 0;
1093         int rc;
1094         ENTRY;
1095
1096         /* we're going to grab page lock, so object lock must not be taken. */
1097         LASSERT(sanity_check(ext) == 0);
1098         /* in locking state, any process should not touch this extent. */
1099         EASSERT(ext->oe_state == OES_LOCKING, ext);
1100         EASSERT(ext->oe_owner != NULL, ext);
1101
1102         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1103
1104         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1105                 ++page_count;
1106                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1107                         last = oap;
1108
1109                 /* checking ASYNC_READY is race safe */
1110                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1111                         continue;
1112
1113                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1114                 switch (rc) {
1115                 case 0:
1116                         spin_lock(&oap->oap_lock);
1117                         oap->oap_async_flags |= ASYNC_READY;
1118                         spin_unlock(&oap->oap_lock);
1119                         break;
1120                 case -EALREADY:
1121                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1122                         break;
1123                 default:
1124                         LASSERTF(0, "unknown return code: %d\n", rc);
1125                 }
1126         }
1127
1128         LASSERT(page_count == ext->oe_nr_pages);
1129         LASSERT(last != NULL);
1130         /* the last page is the only one we need to refresh its count by
1131          * the size of file. */
1132         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1133                 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1134                 LASSERT(last_oap_count > 0);
1135                 LASSERT(last->oap_page_off + last_oap_count <= PAGE_SIZE);
1136                 last->oap_count = last_oap_count;
1137                 spin_lock(&last->oap_lock);
1138                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1139                 spin_unlock(&last->oap_lock);
1140         }
1141
1142         /* for the rest of pages, we don't need to call osf_refresh_count()
1143          * because it's known they are not the last page */
1144         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1145                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1146                         oap->oap_count = PAGE_SIZE - oap->oap_page_off;
1147                         spin_lock(&oap->oap_lock);
1148                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1149                         spin_unlock(&oap->oap_lock);
1150                 }
1151         }
1152
1153         osc_object_lock(obj);
1154         osc_extent_state_set(ext, OES_RPC);
1155         osc_object_unlock(obj);
1156         /* get a refcount for RPC. */
1157         osc_extent_get(ext);
1158
1159         RETURN(0);
1160 }
1161
1162 /**
1163  * Quick and simple version of osc_extent_find(). This function is frequently
1164  * called to expand the extent for the same IO. To expand the extent, the
1165  * page index must be in the same or next chunk of ext->oe_end.
1166  */
1167 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index,
1168                              unsigned int *grants)
1169 {
1170         struct osc_object *obj = ext->oe_obj;
1171         struct client_obd *cli = osc_cli(obj);
1172         struct osc_extent *next;
1173         int ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1174         pgoff_t chunk = index >> ppc_bits;
1175         pgoff_t end_chunk;
1176         pgoff_t end_index;
1177         unsigned int chunksize = 1 << cli->cl_chunkbits;
1178         int rc = 0;
1179         ENTRY;
1180
1181         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1182         osc_object_lock(obj);
1183         LASSERT(sanity_check_nolock(ext) == 0);
1184         end_chunk = ext->oe_end >> ppc_bits;
1185         if (chunk > end_chunk + 1)
1186                 GOTO(out, rc = -ERANGE);
1187
1188         if (end_chunk >= chunk)
1189                 GOTO(out, rc = 0);
1190
1191         LASSERT(end_chunk + 1 == chunk);
1192
1193         /* try to expand this extent to cover @index */
1194         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1195
1196         /* don't go over the maximum extent size reported by server */
1197         if (end_index - ext->oe_start + 1 > cli->cl_max_extent_pages)
1198                 GOTO(out, rc = -ERANGE);
1199
1200         next = next_extent(ext);
1201         if (next != NULL && next->oe_start <= end_index)
1202                 /* complex mode - overlapped with the next extent,
1203                  * this case will be handled by osc_extent_find() */
1204                 GOTO(out, rc = -EAGAIN);
1205
1206         ext->oe_end = end_index;
1207         ext->oe_grants += chunksize;
1208         *grants -= chunksize;
1209         LASSERT(*grants >= 0);
1210         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1211                  "overlapped after expanding for %lu.\n", index);
1212         EXIT;
1213
1214 out:
1215         osc_object_unlock(obj);
1216         RETURN(rc);
1217 }
1218
1219 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
1220                                   const char *func, int line)
1221 {
1222         struct osc_extent *ext;
1223         int cnt;
1224
1225         CDEBUG(level, "Dump object %p extents at %s:%d, mppr: %u.\n",
1226                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1227
1228         /* osc_object_lock(obj); */
1229         cnt = 1;
1230         for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1231                 OSC_EXTENT_DUMP(level, ext, "in tree %d.\n", cnt++);
1232
1233         cnt = 1;
1234         list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1235                 OSC_EXTENT_DUMP(level, ext, "hp %d.\n", cnt++);
1236
1237         cnt = 1;
1238         list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1239                 OSC_EXTENT_DUMP(level, ext, "urgent %d.\n", cnt++);
1240
1241         cnt = 1;
1242         list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1243                 OSC_EXTENT_DUMP(level, ext, "reading %d.\n", cnt++);
1244         /* osc_object_unlock(obj); */
1245 }
1246
1247 /* ------------------ osc extent end ------------------ */
1248
1249 static inline int osc_is_ready(struct osc_object *osc)
1250 {
1251         return !list_empty(&osc->oo_ready_item) ||
1252                !list_empty(&osc->oo_hp_ready_item);
1253 }
1254
1255 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1256         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1257                (OSC), osc_is_ready(OSC),                                       \
1258                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1259                list_empty_marker(&(OSC)->oo_ready_item),                       \
1260                atomic_read(&(OSC)->oo_nr_writes),                              \
1261                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1262                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1263                atomic_read(&(OSC)->oo_nr_reads),                               \
1264                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1265                ##args)
1266
1267 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1268                           int cmd)
1269 {
1270         struct osc_page *opg  = oap2osc_page(oap);
1271         struct cl_page  *page = oap2cl_page(oap);
1272         int result;
1273
1274         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1275
1276         ENTRY;
1277         result = cl_page_make_ready(env, page, CRT_WRITE);
1278         if (result == 0)
1279                 opg->ops_submit_time = cfs_time_current();
1280         RETURN(result);
1281 }
1282
1283 static int osc_refresh_count(const struct lu_env *env,
1284                              struct osc_async_page *oap, int cmd)
1285 {
1286         struct osc_page  *opg = oap2osc_page(oap);
1287         pgoff_t index = osc_index(oap2osc(oap));
1288         struct cl_object *obj;
1289         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
1290
1291         int result;
1292         loff_t kms;
1293
1294         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1295         LASSERT(!(cmd & OBD_BRW_READ));
1296         LASSERT(opg != NULL);
1297         obj = opg->ops_cl.cpl_obj;
1298
1299         cl_object_attr_lock(obj);
1300         result = cl_object_attr_get(env, obj, attr);
1301         cl_object_attr_unlock(obj);
1302         if (result < 0)
1303                 return result;
1304         kms = attr->cat_kms;
1305         if (cl_offset(obj, index) >= kms)
1306                 /* catch race with truncate */
1307                 return 0;
1308         else if (cl_offset(obj, index + 1) > kms)
1309                 /* catch sub-page write at end of file */
1310                 return kms % PAGE_SIZE;
1311         else
1312                 return PAGE_SIZE;
1313 }
1314
1315 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1316                           int cmd, int rc)
1317 {
1318         struct osc_page   *opg  = oap2osc_page(oap);
1319         struct cl_page    *page = oap2cl_page(oap);
1320         enum cl_req_type   crt;
1321         int srvlock;
1322
1323         ENTRY;
1324
1325         cmd &= ~OBD_BRW_NOQUOTA;
1326         LASSERTF(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ),
1327                  "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1328         LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1329                 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1330         LASSERT(opg->ops_transfer_pinned);
1331
1332         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1333         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1334         opg->ops_transfer_pinned = 0;
1335
1336         opg->ops_submit_time = 0;
1337         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1338
1339         /* statistic */
1340         if (rc == 0 && srvlock) {
1341                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1342                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1343                 size_t bytes = oap->oap_count;
1344
1345                 if (crt == CRT_READ)
1346                         stats->os_lockless_reads += bytes;
1347                 else
1348                         stats->os_lockless_writes += bytes;
1349         }
1350
1351         /*
1352          * This has to be the last operation with the page, as locks are
1353          * released in cl_page_completion() and nothing except for the
1354          * reference counter protects page from concurrent reclaim.
1355          */
1356         lu_ref_del(&page->cp_reference, "transfer", page);
1357
1358         cl_page_completion(env, page, crt, rc);
1359         cl_page_put(env, page);
1360
1361         RETURN(0);
1362 }
1363
1364 #define OSC_DUMP_GRANT(lvl, cli, fmt, args...) do {                     \
1365         struct client_obd *__tmp = (cli);                               \
1366         CDEBUG(lvl, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu "  \
1367                "dropped: %ld avail: %ld, dirty_grant: %ld, "            \
1368                "reserved: %ld, flight: %d } lru {in list: %ld, "        \
1369                "left: %ld, waiters: %d }" fmt "\n",                     \
1370                cli_name(__tmp),                                         \
1371                __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages,        \
1372                atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1373                __tmp->cl_lost_grant, __tmp->cl_avail_grant,             \
1374                __tmp->cl_dirty_grant,                                   \
1375                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,         \
1376                atomic_long_read(&__tmp->cl_lru_in_list),                \
1377                atomic_long_read(&__tmp->cl_lru_busy),                   \
1378                atomic_read(&__tmp->cl_lru_shrinkers), ##args);          \
1379 } while (0)
1380
1381 /* caller must hold loi_list_lock */
1382 static void osc_consume_write_grant(struct client_obd *cli,
1383                                     struct brw_page *pga)
1384 {
1385         assert_spin_locked(&cli->cl_loi_list_lock);
1386         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1387         atomic_long_inc(&obd_dirty_pages);
1388         cli->cl_dirty_pages++;
1389         pga->flag |= OBD_BRW_FROM_GRANT;
1390         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1391                PAGE_SIZE, pga, pga->pg);
1392         osc_update_next_shrink(cli);
1393 }
1394
1395 /* the companion to osc_consume_write_grant, called when a brw has completed.
1396  * must be called with the loi lock held. */
1397 static void osc_release_write_grant(struct client_obd *cli,
1398                                     struct brw_page *pga)
1399 {
1400         ENTRY;
1401
1402         assert_spin_locked(&cli->cl_loi_list_lock);
1403         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1404                 EXIT;
1405                 return;
1406         }
1407
1408         pga->flag &= ~OBD_BRW_FROM_GRANT;
1409         atomic_long_dec(&obd_dirty_pages);
1410         cli->cl_dirty_pages--;
1411         if (pga->flag & OBD_BRW_NOCACHE) {
1412                 pga->flag &= ~OBD_BRW_NOCACHE;
1413                 atomic_long_dec(&obd_dirty_transit_pages);
1414                 cli->cl_dirty_transit--;
1415         }
1416         EXIT;
1417 }
1418
1419 /**
1420  * To avoid sleeping with object lock held, it's good for us allocate enough
1421  * grants before entering into critical section.
1422  *
1423  * client_obd_list_lock held by caller
1424  */
1425 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1426 {
1427         int rc = -EDQUOT;
1428
1429         if (cli->cl_avail_grant >= bytes) {
1430                 cli->cl_avail_grant    -= bytes;
1431                 cli->cl_reserved_grant += bytes;
1432                 rc = 0;
1433         }
1434         return rc;
1435 }
1436
1437 static void __osc_unreserve_grant(struct client_obd *cli,
1438                                   unsigned int reserved, unsigned int unused)
1439 {
1440         /* it's quite normal for us to get more grant than reserved.
1441          * Thinking about a case that two extents merged by adding a new
1442          * chunk, we can save one extent tax. If extent tax is greater than
1443          * one chunk, we can save more grant by adding a new chunk */
1444         cli->cl_reserved_grant -= reserved;
1445         if (unused > reserved) {
1446                 cli->cl_avail_grant += reserved;
1447                 cli->cl_lost_grant  += unused - reserved;
1448                 cli->cl_dirty_grant -= unused - reserved;
1449         } else {
1450                 cli->cl_avail_grant += unused;
1451                 cli->cl_dirty_grant += reserved - unused;
1452         }
1453 }
1454
1455 static void osc_unreserve_grant(struct client_obd *cli,
1456                                 unsigned int reserved, unsigned int unused)
1457 {
1458         spin_lock(&cli->cl_loi_list_lock);
1459         __osc_unreserve_grant(cli, reserved, unused);
1460         if (unused > 0)
1461                 osc_wake_cache_waiters(cli);
1462         spin_unlock(&cli->cl_loi_list_lock);
1463 }
1464
1465 /**
1466  * Free grant after IO is finished or canceled.
1467  *
1468  * @lost_grant is used to remember how many grants we have allocated but not
1469  * used, we should return these grants to OST. There're two cases where grants
1470  * can be lost:
1471  * 1. truncate;
1472  * 2. blocksize at OST is less than PAGE_SIZE and a partial page was
1473  *    written. In this case OST may use less chunks to serve this partial
1474  *    write. OSTs don't actually know the page size on the client side. so
1475  *    clients have to calculate lost grant by the blocksize on the OST.
1476  *    See filter_grant_check() for details.
1477  */
1478 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1479                            unsigned int lost_grant, unsigned int dirty_grant)
1480 {
1481         unsigned long grant;
1482
1483         grant = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
1484
1485         spin_lock(&cli->cl_loi_list_lock);
1486         atomic_long_sub(nr_pages, &obd_dirty_pages);
1487         cli->cl_dirty_pages -= nr_pages;
1488         cli->cl_lost_grant += lost_grant;
1489         cli->cl_dirty_grant -= dirty_grant;
1490         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1491                 /* borrow some grant from truncate to avoid the case that
1492                  * truncate uses up all avail grant */
1493                 cli->cl_lost_grant -= grant;
1494                 cli->cl_avail_grant += grant;
1495         }
1496         osc_wake_cache_waiters(cli);
1497         spin_unlock(&cli->cl_loi_list_lock);
1498         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu/%lu\n",
1499                lost_grant, cli->cl_lost_grant,
1500                cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_SHIFT,
1501                cli->cl_dirty_grant);
1502 }
1503
1504 /**
1505  * The companion to osc_enter_cache(), called when @oap is no longer part of
1506  * the dirty accounting due to error.
1507  */
1508 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1509 {
1510         spin_lock(&cli->cl_loi_list_lock);
1511         osc_release_write_grant(cli, &oap->oap_brw_page);
1512         spin_unlock(&cli->cl_loi_list_lock);
1513 }
1514
1515 /**
1516  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1517  * is available.
1518  */
1519 static int osc_enter_cache_try(struct client_obd *cli,
1520                                struct osc_async_page *oap,
1521                                int bytes, int transient)
1522 {
1523         int rc;
1524
1525         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1526
1527         rc = osc_reserve_grant(cli, bytes);
1528         if (rc < 0)
1529                 return 0;
1530
1531         if (cli->cl_dirty_pages < cli->cl_dirty_max_pages &&
1532             1 + atomic_long_read(&obd_dirty_pages) <= obd_max_dirty_pages) {
1533                 osc_consume_write_grant(cli, &oap->oap_brw_page);
1534                 if (transient) {
1535                         cli->cl_dirty_transit++;
1536                         atomic_long_inc(&obd_dirty_transit_pages);
1537                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
1538                 }
1539                 rc = 1;
1540         } else {
1541                 __osc_unreserve_grant(cli, bytes, bytes);
1542                 rc = 0;
1543         }
1544         return rc;
1545 }
1546
1547 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1548 {
1549         int rc;
1550         spin_lock(&cli->cl_loi_list_lock);
1551         rc = list_empty(&ocw->ocw_entry);
1552         spin_unlock(&cli->cl_loi_list_lock);
1553         return rc;
1554 }
1555
1556 /**
1557  * The main entry to reserve dirty page accounting. Usually the grant reserved
1558  * in this function will be freed in bulk in osc_free_grant() unless it fails
1559  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1560  *
1561  * The process will be put into sleep if it's already run out of grant.
1562  */
1563 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1564                            struct osc_async_page *oap, int bytes)
1565 {
1566         struct osc_object       *osc = oap->oap_obj;
1567         struct lov_oinfo        *loi = osc->oo_oinfo;
1568         struct osc_cache_waiter  ocw;
1569         struct l_wait_info       lwi;
1570         int                      rc = -EDQUOT;
1571         ENTRY;
1572
1573         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(AT_OFF ? obd_timeout : at_max),
1574                                NULL, LWI_ON_SIGNAL_NOOP, NULL);
1575
1576         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1577
1578         spin_lock(&cli->cl_loi_list_lock);
1579
1580         /* force the caller to try sync io.  this can jump the list
1581          * of queued writes and create a discontiguous rpc stream */
1582         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1583             cli->cl_dirty_max_pages == 0 ||
1584             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) {
1585                 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1586                 GOTO(out, rc = -EDQUOT);
1587         }
1588
1589         /* Hopefully normal case - cache space and write credits available */
1590         if (osc_enter_cache_try(cli, oap, bytes, 0)) {
1591                 OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1592                 GOTO(out, rc = 0);
1593         }
1594
1595         /* We can get here for two reasons: too many dirty pages in cache, or
1596          * run out of grants. In both cases we should write dirty pages out.
1597          * Adding a cache waiter will trigger urgent write-out no matter what
1598          * RPC size will be.
1599          * The exiting condition is no avail grants and no dirty pages caching,
1600          * that really means there is no space on the OST. */
1601         init_waitqueue_head(&ocw.ocw_waitq);
1602         ocw.ocw_oap   = oap;
1603         ocw.ocw_grant = bytes;
1604         while (cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0) {
1605                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1606                 ocw.ocw_rc = 0;
1607                 spin_unlock(&cli->cl_loi_list_lock);
1608
1609                 osc_io_unplug_async(env, cli, NULL);
1610
1611                 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1612                        cli_name(cli), &ocw, oap);
1613
1614                 rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1615
1616                 spin_lock(&cli->cl_loi_list_lock);
1617
1618                 if (rc < 0) {
1619                         /* l_wait_event is interrupted by signal or timed out */
1620                         list_del_init(&ocw.ocw_entry);
1621                         break;
1622                 }
1623                 LASSERT(list_empty(&ocw.ocw_entry));
1624                 rc = ocw.ocw_rc;
1625
1626                 if (rc != -EDQUOT)
1627                         break;
1628                 if (osc_enter_cache_try(cli, oap, bytes, 0)) {
1629                         rc = 0;
1630                         break;
1631                 }
1632         }
1633
1634         switch (rc) {
1635         case 0:
1636                 OSC_DUMP_GRANT(D_CACHE, cli, "finally got grant space\n");
1637                 break;
1638         case -ETIMEDOUT:
1639                 OSC_DUMP_GRANT(D_CACHE, cli,
1640                                "timeout, fall back to sync i/o\n");
1641                 osc_extent_tree_dump(D_CACHE, osc);
1642                 /* fall back to synchronous I/O */
1643                 rc = -EDQUOT;
1644                 break;
1645         case -EINTR:
1646                 /* Ensures restartability - LU-3581 */
1647                 OSC_DUMP_GRANT(D_CACHE, cli, "interrupted\n");
1648                 rc = -ERESTARTSYS;
1649                 break;
1650         case -EDQUOT:
1651                 OSC_DUMP_GRANT(D_CACHE, cli,
1652                                "no grant space, fall back to sync i/o\n");
1653                 break;
1654         default:
1655                 CDEBUG(D_CACHE, "%s: event for cache space @ %p never arrived "
1656                        "due to %d, fall back to sync i/o\n",
1657                        cli_name(cli), &ocw, rc);
1658                 break;
1659         }
1660         EXIT;
1661 out:
1662         spin_unlock(&cli->cl_loi_list_lock);
1663         RETURN(rc);
1664 }
1665
1666 /* caller must hold loi_list_lock */
1667 void osc_wake_cache_waiters(struct client_obd *cli)
1668 {
1669         struct list_head *l, *tmp;
1670         struct osc_cache_waiter *ocw;
1671
1672         ENTRY;
1673         list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1674                 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
1675                 list_del_init(&ocw->ocw_entry);
1676
1677                 ocw->ocw_rc = -EDQUOT;
1678                 /* we can't dirty more */
1679                 if ((cli->cl_dirty_pages  >= cli->cl_dirty_max_pages) ||
1680                     (1 + atomic_long_read(&obd_dirty_pages) >
1681                      obd_max_dirty_pages)) {
1682                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
1683                                "osc max %ld, sys max %ld\n",
1684                                cli->cl_dirty_pages, cli->cl_dirty_max_pages,
1685                                obd_max_dirty_pages);
1686                         goto wakeup;
1687                 }
1688
1689                 if (osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
1690                         ocw->ocw_rc = 0;
1691 wakeup:
1692                 CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld, %d\n",
1693                        ocw, ocw->ocw_oap, cli->cl_avail_grant, ocw->ocw_rc);
1694
1695                 wake_up(&ocw->ocw_waitq);
1696         }
1697
1698         EXIT;
1699 }
1700
1701 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1702 {
1703         int hprpc = !!list_empty(&osc->oo_hp_exts);
1704         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1705 }
1706
1707 /* This maintains the lists of pending pages to read/write for a given object
1708  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1709  * to quickly find objects that are ready to send an RPC. */
1710 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1711                          int cmd)
1712 {
1713         int invalid_import = 0;
1714         ENTRY;
1715
1716         /* if we have an invalid import we want to drain the queued pages
1717          * by forcing them through rpcs that immediately fail and complete
1718          * the pages.  recovery relies on this to empty the queued pages
1719          * before canceling the locks and evicting down the llite pages */
1720         if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1721                 invalid_import = 1;
1722
1723         if (cmd & OBD_BRW_WRITE) {
1724                 if (atomic_read(&osc->oo_nr_writes) == 0)
1725                         RETURN(0);
1726                 if (invalid_import) {
1727                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1728                         RETURN(1);
1729                 }
1730                 if (!list_empty(&osc->oo_hp_exts)) {
1731                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1732                         RETURN(1);
1733                 }
1734                 if (!list_empty(&osc->oo_urgent_exts)) {
1735                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1736                         RETURN(1);
1737                 }
1738                 /* trigger a write rpc stream as long as there are dirtiers
1739                  * waiting for space.  as they're waiting, they're not going to
1740                  * create more pages to coalesce with what's waiting.. */
1741                 if (!list_empty(&cli->cl_cache_waiters)) {
1742                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1743                         RETURN(1);
1744                 }
1745                 if (!list_empty(&osc->oo_full_exts)) {
1746                         CDEBUG(D_CACHE, "full extent ready, make an RPC\n");
1747                         RETURN(1);
1748                 }
1749         } else {
1750                 if (atomic_read(&osc->oo_nr_reads) == 0)
1751                         RETURN(0);
1752                 if (invalid_import) {
1753                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1754                         RETURN(1);
1755                 }
1756                 /* all read are urgent. */
1757                 if (!list_empty(&osc->oo_reading_exts))
1758                         RETURN(1);
1759         }
1760
1761         RETURN(0);
1762 }
1763
1764 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1765 {
1766         struct client_obd *cli = osc_cli(obj);
1767         if (cmd & OBD_BRW_WRITE) {
1768                 atomic_add(delta, &obj->oo_nr_writes);
1769                 atomic_add(delta, &cli->cl_pending_w_pages);
1770                 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1771         } else {
1772                 atomic_add(delta, &obj->oo_nr_reads);
1773                 atomic_add(delta, &cli->cl_pending_r_pages);
1774                 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1775         }
1776         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1777 }
1778
1779 static int osc_makes_hprpc(struct osc_object *obj)
1780 {
1781         return !list_empty(&obj->oo_hp_exts);
1782 }
1783
1784 static void on_list(struct list_head *item, struct list_head *list,
1785                     int should_be_on)
1786 {
1787         if (list_empty(item) && should_be_on)
1788                 list_add_tail(item, list);
1789         else if (!list_empty(item) && !should_be_on)
1790                 list_del_init(item);
1791 }
1792
1793 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1794  * can find pages to build into rpcs quickly */
1795 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1796 {
1797         if (osc_makes_hprpc(osc)) {
1798                 /* HP rpc */
1799                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1800                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1801         } else {
1802                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1803                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1804                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1805                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1806         }
1807
1808         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1809                 atomic_read(&osc->oo_nr_writes) > 0);
1810
1811         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1812                 atomic_read(&osc->oo_nr_reads) > 0);
1813
1814         return osc_is_ready(osc);
1815 }
1816
1817 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1818 {
1819         int is_ready;
1820
1821         spin_lock(&cli->cl_loi_list_lock);
1822         is_ready = __osc_list_maint(cli, osc);
1823         spin_unlock(&cli->cl_loi_list_lock);
1824
1825         return is_ready;
1826 }
1827
1828 /* this is trying to propogate async writeback errors back up to the
1829  * application.  As an async write fails we record the error code for later if
1830  * the app does an fsync.  As long as errors persist we force future rpcs to be
1831  * sync so that the app can get a sync error and break the cycle of queueing
1832  * pages for which writeback will fail. */
1833 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1834                            int rc)
1835 {
1836         if (rc) {
1837                 if (!ar->ar_rc)
1838                         ar->ar_rc = rc;
1839
1840                 ar->ar_force_sync = 1;
1841                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1842                 return;
1843
1844         }
1845
1846         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1847                 ar->ar_force_sync = 0;
1848 }
1849
1850 /* this must be called holding the loi list lock to give coverage to exit_cache,
1851  * async_flag maintenance, and oap_request */
1852 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1853                               struct osc_async_page *oap, int sent, int rc)
1854 {
1855         struct osc_object *osc = oap->oap_obj;
1856         struct lov_oinfo  *loi = osc->oo_oinfo;
1857         __u64 xid = 0;
1858
1859         ENTRY;
1860         if (oap->oap_request != NULL) {
1861                 xid = ptlrpc_req_xid(oap->oap_request);
1862                 ptlrpc_req_finished(oap->oap_request);
1863                 oap->oap_request = NULL;
1864         }
1865
1866         /* As the transfer for this page is being done, clear the flags */
1867         spin_lock(&oap->oap_lock);
1868         oap->oap_async_flags = 0;
1869         spin_unlock(&oap->oap_lock);
1870         oap->oap_interrupted = 0;
1871
1872         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1873                 spin_lock(&cli->cl_loi_list_lock);
1874                 osc_process_ar(&cli->cl_ar, xid, rc);
1875                 osc_process_ar(&loi->loi_ar, xid, rc);
1876                 spin_unlock(&cli->cl_loi_list_lock);
1877         }
1878
1879         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1880         if (rc)
1881                 CERROR("completion on oap %p obj %p returns %d.\n",
1882                        oap, osc, rc);
1883
1884         EXIT;
1885 }
1886
1887 struct extent_rpc_data {
1888         struct list_head        *erd_rpc_list;
1889         unsigned int            erd_page_count;
1890         unsigned int            erd_max_pages;
1891         unsigned int            erd_max_chunks;
1892         unsigned int            erd_max_extents;
1893 };
1894
1895 static inline unsigned osc_extent_chunks(const struct osc_extent *ext)
1896 {
1897         struct client_obd *cli = osc_cli(ext->oe_obj);
1898         unsigned ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1899
1900         return (ext->oe_end >> ppc_bits) - (ext->oe_start >> ppc_bits) + 1;
1901 }
1902
1903 /**
1904  * Try to add extent to one RPC. We need to think about the following things:
1905  * - # of pages must not be over max_pages_per_rpc
1906  * - extent must be compatible with previous ones
1907  */
1908 static int try_to_add_extent_for_io(struct client_obd *cli,
1909                                     struct osc_extent *ext,
1910                                     struct extent_rpc_data *data)
1911 {
1912         struct osc_extent *tmp;
1913         unsigned int chunk_count;
1914         struct osc_async_page *oap = list_first_entry(&ext->oe_pages,
1915                                                       struct osc_async_page,
1916                                                       oap_pending_item);
1917         ENTRY;
1918
1919         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1920                 ext);
1921         OSC_EXTENT_DUMP(D_CACHE, ext, "trying to add this extent\n");
1922
1923         if (data->erd_max_extents == 0)
1924                 RETURN(0);
1925
1926         chunk_count = osc_extent_chunks(ext);
1927         EASSERTF(data->erd_page_count != 0 ||
1928                  chunk_count <= data->erd_max_chunks, ext,
1929                  "The first extent to be fit in a RPC contains %u chunks, "
1930                  "which is over the limit %u.\n", chunk_count,
1931                  data->erd_max_chunks);
1932         if (chunk_count > data->erd_max_chunks)
1933                 RETURN(0);
1934
1935         data->erd_max_pages = max(ext->oe_mppr, data->erd_max_pages);
1936         EASSERTF(data->erd_page_count != 0 ||
1937                 ext->oe_nr_pages <= data->erd_max_pages, ext,
1938                 "The first extent to be fit in a RPC contains %u pages, "
1939                 "which is over the limit %u.\n", ext->oe_nr_pages,
1940                 data->erd_max_pages);
1941         if (data->erd_page_count + ext->oe_nr_pages > data->erd_max_pages)
1942                 RETURN(0);
1943
1944         list_for_each_entry(tmp, data->erd_rpc_list, oe_link) {
1945                 struct osc_async_page *oap2;
1946                 oap2 = list_first_entry(&tmp->oe_pages, struct osc_async_page,
1947                                         oap_pending_item);
1948                 EASSERT(tmp->oe_owner == current, tmp);
1949 #if 0
1950                 if (overlapped(tmp, ext)) {
1951                         OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
1952                         EASSERT(0, ext);
1953                 }
1954 #endif
1955                 if (oap2cl_page(oap)->cp_type != oap2cl_page(oap2)->cp_type) {
1956                         CDEBUG(D_CACHE, "Do not permit different types of IO "
1957                                "in one RPC\n");
1958                         RETURN(0);
1959                 }
1960
1961                 if (tmp->oe_srvlock != ext->oe_srvlock ||
1962                     !tmp->oe_grants != !ext->oe_grants ||
1963                     tmp->oe_no_merge || ext->oe_no_merge)
1964                         RETURN(0);
1965
1966                 /* remove break for strict check */
1967                 break;
1968         }
1969
1970         data->erd_max_extents--;
1971         data->erd_max_chunks -= chunk_count;
1972         data->erd_page_count += ext->oe_nr_pages;
1973         list_move_tail(&ext->oe_link, data->erd_rpc_list);
1974         ext->oe_owner = current;
1975         RETURN(1);
1976 }
1977
1978 static inline unsigned osc_max_write_chunks(const struct client_obd *cli)
1979 {
1980         /*
1981          * LU-8135:
1982          *
1983          * The maximum size of a single transaction is about 64MB in ZFS.
1984          * #define DMU_MAX_ACCESS (64 * 1024 * 1024)
1985          *
1986          * Since ZFS is a copy-on-write file system, a single dirty page in
1987          * a chunk will result in the rewrite of the whole chunk, therefore
1988          * an RPC shouldn't be allowed to contain too many chunks otherwise
1989          * it will make transaction size much bigger than 64MB, especially
1990          * with big block size for ZFS.
1991          *
1992          * This piece of code is to make sure that OSC won't send write RPCs
1993          * with too many chunks. The maximum chunk size that an RPC can cover
1994          * is set to PTLRPC_MAX_BRW_SIZE, which is defined to 16MB. Ideally
1995          * OST should tell the client what the biggest transaction size is,
1996          * but it's good enough for now.
1997          *
1998          * This limitation doesn't apply to ldiskfs, which allows as many
1999          * chunks in one RPC as we want. However, it won't have any benefits
2000          * to have too many discontiguous pages in one RPC.
2001          *
2002          * An osc_extent won't cover over a RPC size, so the chunks in an
2003          * osc_extent won't bigger than PTLRPC_MAX_BRW_SIZE >> chunkbits.
2004          */
2005         return PTLRPC_MAX_BRW_SIZE >> cli->cl_chunkbits;
2006 }
2007
2008 /**
2009  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
2010  * get_write_extent() takes all appropriate extents in atomic.
2011  *
2012  * The following policy is used to collect extents for IO:
2013  * 1. Add as many HP extents as possible;
2014  * 2. Add the first urgent extent in urgent extent list and take it out of
2015  *    urgent list;
2016  * 3. Add subsequent extents of this urgent extent;
2017  * 4. If urgent list is not empty, goto 2;
2018  * 5. Traverse the extent tree from the 1st extent;
2019  * 6. Above steps exit if there is no space in this RPC.
2020  */
2021 static unsigned int get_write_extents(struct osc_object *obj,
2022                                       struct list_head *rpclist)
2023 {
2024         struct client_obd *cli = osc_cli(obj);
2025         struct osc_extent *ext;
2026         struct extent_rpc_data data = {
2027                 .erd_rpc_list   = rpclist,
2028                 .erd_page_count = 0,
2029                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
2030                 .erd_max_chunks = osc_max_write_chunks(cli),
2031                 .erd_max_extents = 256,
2032         };
2033
2034         LASSERT(osc_object_is_locked(obj));
2035         while (!list_empty(&obj->oo_hp_exts)) {
2036                 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
2037                                  oe_link);
2038                 LASSERT(ext->oe_state == OES_CACHE);
2039                 if (!try_to_add_extent_for_io(cli, ext, &data))
2040                         return data.erd_page_count;
2041                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2042         }
2043         if (data.erd_page_count == data.erd_max_pages)
2044                 return data.erd_page_count;
2045
2046         while (!list_empty(&obj->oo_urgent_exts)) {
2047                 ext = list_entry(obj->oo_urgent_exts.next,
2048                                  struct osc_extent, oe_link);
2049                 if (!try_to_add_extent_for_io(cli, ext, &data))
2050                         return data.erd_page_count;
2051         }
2052         if (data.erd_page_count == data.erd_max_pages)
2053                 return data.erd_page_count;
2054
2055         /* One key difference between full extents and other extents: full
2056          * extents can usually only be added if the rpclist was empty, so if we
2057          * can't add one, we continue on to trying to add normal extents.  This
2058          * is so we don't miss adding extra extents to an RPC containing high
2059          * priority or urgent extents. */
2060         while (!list_empty(&obj->oo_full_exts)) {
2061                 ext = list_entry(obj->oo_full_exts.next,
2062                                  struct osc_extent, oe_link);
2063                 if (!try_to_add_extent_for_io(cli, ext, &data))
2064                         break;
2065         }
2066         if (data.erd_page_count == data.erd_max_pages)
2067                 return data.erd_page_count;
2068
2069         ext = first_extent(obj);
2070         while (ext != NULL) {
2071                 if ((ext->oe_state != OES_CACHE) ||
2072                     /* this extent may be already in current rpclist */
2073                     (!list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
2074                         ext = next_extent(ext);
2075                         continue;
2076                 }
2077
2078                 if (!try_to_add_extent_for_io(cli, ext, &data))
2079                         return data.erd_page_count;
2080
2081                 ext = next_extent(ext);
2082         }
2083         return data.erd_page_count;
2084 }
2085
2086 static int
2087 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
2088                    struct osc_object *osc)
2089 __must_hold(osc)
2090 {
2091         struct list_head   rpclist = LIST_HEAD_INIT(rpclist);
2092         struct osc_extent *ext;
2093         struct osc_extent *tmp;
2094         struct osc_extent *first = NULL;
2095         unsigned int page_count = 0;
2096         int srvlock = 0;
2097         int rc = 0;
2098         ENTRY;
2099
2100         LASSERT(osc_object_is_locked(osc));
2101
2102         page_count = get_write_extents(osc, &rpclist);
2103         LASSERT(equi(page_count == 0, list_empty(&rpclist)));
2104
2105         if (list_empty(&rpclist))
2106                 RETURN(0);
2107
2108         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2109
2110         list_for_each_entry(ext, &rpclist, oe_link) {
2111                 LASSERT(ext->oe_state == OES_CACHE ||
2112                         ext->oe_state == OES_LOCK_DONE);
2113                 if (ext->oe_state == OES_CACHE)
2114                         osc_extent_state_set(ext, OES_LOCKING);
2115                 else
2116                         osc_extent_state_set(ext, OES_RPC);
2117         }
2118
2119         /* we're going to grab page lock, so release object lock because
2120          * lock order is page lock -> object lock. */
2121         osc_object_unlock(osc);
2122
2123         list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2124                 if (ext->oe_state == OES_LOCKING) {
2125                         rc = osc_extent_make_ready(env, ext);
2126                         if (unlikely(rc < 0)) {
2127                                 list_del_init(&ext->oe_link);
2128                                 osc_extent_finish(env, ext, 0, rc);
2129                                 continue;
2130                         }
2131                 }
2132                 if (first == NULL) {
2133                         first = ext;
2134                         srvlock = ext->oe_srvlock;
2135                 } else {
2136                         LASSERT(srvlock == ext->oe_srvlock);
2137                 }
2138         }
2139
2140         if (!list_empty(&rpclist)) {
2141                 LASSERT(page_count > 0);
2142                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2143                 LASSERT(list_empty(&rpclist));
2144         }
2145
2146         osc_object_lock(osc);
2147         RETURN(rc);
2148 }
2149
2150 /**
2151  * prepare pages for ASYNC io and put pages in send queue.
2152  *
2153  * \param cmd OBD_BRW_* macroses
2154  * \param lop pending pages
2155  *
2156  * \return zero if no page added to send queue.
2157  * \return 1 if pages successfully added to send queue.
2158  * \return negative on errors.
2159  */
2160 static int
2161 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2162                   struct osc_object *osc)
2163 __must_hold(osc)
2164 {
2165         struct osc_extent *ext;
2166         struct osc_extent *next;
2167         struct list_head rpclist = LIST_HEAD_INIT(rpclist);
2168         struct extent_rpc_data data = {
2169                 .erd_rpc_list   = &rpclist,
2170                 .erd_page_count = 0,
2171                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
2172                 .erd_max_chunks = UINT_MAX,
2173                 .erd_max_extents = UINT_MAX,
2174         };
2175         int rc = 0;
2176         ENTRY;
2177
2178         LASSERT(osc_object_is_locked(osc));
2179         list_for_each_entry_safe(ext, next, &osc->oo_reading_exts, oe_link) {
2180                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2181                 if (!try_to_add_extent_for_io(cli, ext, &data))
2182                         break;
2183                 osc_extent_state_set(ext, OES_RPC);
2184                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2185         }
2186         LASSERT(data.erd_page_count <= data.erd_max_pages);
2187
2188         osc_update_pending(osc, OBD_BRW_READ, -data.erd_page_count);
2189
2190         if (!list_empty(&rpclist)) {
2191                 osc_object_unlock(osc);
2192
2193                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2194                 LASSERT(list_empty(&rpclist));
2195
2196                 osc_object_lock(osc);
2197         }
2198         RETURN(rc);
2199 }
2200
2201 #define list_to_obj(list, item) ({                                            \
2202         struct list_head *__tmp = (list)->next;                               \
2203         list_del_init(__tmp);                                         \
2204         list_entry(__tmp, struct osc_object, oo_##item);                      \
2205 })
2206
2207 /* This is called by osc_check_rpcs() to find which objects have pages that
2208  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2209 static struct osc_object *osc_next_obj(struct client_obd *cli)
2210 {
2211         ENTRY;
2212
2213         /* First return objects that have blocked locks so that they
2214          * will be flushed quickly and other clients can get the lock,
2215          * then objects which have pages ready to be stuffed into RPCs */
2216         if (!list_empty(&cli->cl_loi_hp_ready_list))
2217                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2218         if (!list_empty(&cli->cl_loi_ready_list))
2219                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2220
2221         /* then if we have cache waiters, return all objects with queued
2222          * writes.  This is especially important when many small files
2223          * have filled up the cache and not been fired into rpcs because
2224          * they don't pass the nr_pending/object threshhold */
2225         if (!list_empty(&cli->cl_cache_waiters) &&
2226             !list_empty(&cli->cl_loi_write_list))
2227                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2228
2229         /* then return all queued objects when we have an invalid import
2230          * so that they get flushed */
2231         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2232                 if (!list_empty(&cli->cl_loi_write_list))
2233                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2234                                            write_item));
2235                 if (!list_empty(&cli->cl_loi_read_list))
2236                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2237                                            read_item));
2238         }
2239         RETURN(NULL);
2240 }
2241
2242 /* called with the loi list lock held */
2243 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2244 __must_hold(&cli->cl_loi_list_lock)
2245 {
2246         struct osc_object *osc;
2247         int rc = 0;
2248         ENTRY;
2249
2250         while ((osc = osc_next_obj(cli)) != NULL) {
2251                 struct cl_object *obj = osc2cl(osc);
2252                 struct lu_ref_link link;
2253
2254                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2255
2256                 if (osc_max_rpc_in_flight(cli, osc)) {
2257                         __osc_list_maint(cli, osc);
2258                         break;
2259                 }
2260
2261                 cl_object_get(obj);
2262                 spin_unlock(&cli->cl_loi_list_lock);
2263                 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2264
2265                 /* attempt some read/write balancing by alternating between
2266                  * reads and writes in an object.  The makes_rpc checks here
2267                  * would be redundant if we were getting read/write work items
2268                  * instead of objects.  we don't want send_oap_rpc to drain a
2269                  * partial read pending queue when we're given this object to
2270                  * do io on writes while there are cache waiters */
2271                 osc_object_lock(osc);
2272                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2273                         rc = osc_send_write_rpc(env, cli, osc);
2274                         if (rc < 0) {
2275                                 CERROR("Write request failed with %d\n", rc);
2276
2277                                 /* osc_send_write_rpc failed, mostly because of
2278                                  * memory pressure.
2279                                  *
2280                                  * It can't break here, because if:
2281                                  *  - a page was submitted by osc_io_submit, so
2282                                  *    page locked;
2283                                  *  - no request in flight
2284                                  *  - no subsequent request
2285                                  * The system will be in live-lock state,
2286                                  * because there is no chance to call
2287                                  * osc_io_unplug() and osc_check_rpcs() any
2288                                  * more. pdflush can't help in this case,
2289                                  * because it might be blocked at grabbing
2290                                  * the page lock as we mentioned.
2291                                  *
2292                                  * Anyway, continue to drain pages. */
2293                                 /* break; */
2294                         }
2295                 }
2296                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2297                         rc = osc_send_read_rpc(env, cli, osc);
2298                         if (rc < 0)
2299                                 CERROR("Read request failed with %d\n", rc);
2300                 }
2301                 osc_object_unlock(osc);
2302
2303                 osc_list_maint(cli, osc);
2304                 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2305                 cl_object_put(env, obj);
2306
2307                 spin_lock(&cli->cl_loi_list_lock);
2308         }
2309 }
2310
2311 static int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2312                           struct osc_object *osc, int async)
2313 {
2314         int rc = 0;
2315
2316         if (osc != NULL && osc_list_maint(cli, osc) == 0)
2317                 return 0;
2318
2319         if (!async) {
2320                 spin_lock(&cli->cl_loi_list_lock);
2321                 osc_check_rpcs(env, cli);
2322                 spin_unlock(&cli->cl_loi_list_lock);
2323         } else {
2324                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2325                 LASSERT(cli->cl_writeback_work != NULL);
2326                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2327         }
2328         return rc;
2329 }
2330
2331 static int osc_io_unplug_async(const struct lu_env *env,
2332                                 struct client_obd *cli, struct osc_object *osc)
2333 {
2334         return osc_io_unplug0(env, cli, osc, 1);
2335 }
2336
2337 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
2338                    struct osc_object *osc)
2339 {
2340         (void)osc_io_unplug0(env, cli, osc, 0);
2341 }
2342
2343 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2344                         struct page *page, loff_t offset)
2345 {
2346         struct obd_export     *exp = osc_export(osc);
2347         struct osc_async_page *oap = &ops->ops_oap;
2348         ENTRY;
2349
2350         if (!page)
2351                 return cfs_size_round(sizeof(*oap));
2352
2353         oap->oap_magic = OAP_MAGIC;
2354         oap->oap_cli = &exp->exp_obd->u.cli;
2355         oap->oap_obj = osc;
2356
2357         oap->oap_page = page;
2358         oap->oap_obj_off = offset;
2359         LASSERT(!(offset & ~PAGE_MASK));
2360
2361         if (cfs_capable(CFS_CAP_SYS_RESOURCE))
2362                 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2363
2364         INIT_LIST_HEAD(&oap->oap_pending_item);
2365         INIT_LIST_HEAD(&oap->oap_rpc_item);
2366
2367         spin_lock_init(&oap->oap_lock);
2368         CDEBUG(D_INFO, "oap %p page %p obj off %llu\n",
2369                oap, page, oap->oap_obj_off);
2370         RETURN(0);
2371 }
2372
2373 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2374                        struct osc_page *ops)
2375 {
2376         struct osc_io *oio = osc_env_io(env);
2377         struct osc_extent     *ext = NULL;
2378         struct osc_async_page *oap = &ops->ops_oap;
2379         struct client_obd     *cli = oap->oap_cli;
2380         struct osc_object     *osc = oap->oap_obj;
2381         pgoff_t index;
2382         unsigned int tmp;
2383         unsigned int grants = 0;
2384         u32    brw_flags = OBD_BRW_ASYNC;
2385         int    cmd = OBD_BRW_WRITE;
2386         int    need_release = 0;
2387         int    rc = 0;
2388         ENTRY;
2389
2390         if (oap->oap_magic != OAP_MAGIC)
2391                 RETURN(-EINVAL);
2392
2393         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2394                 RETURN(-EIO);
2395
2396         if (!list_empty(&oap->oap_pending_item) ||
2397             !list_empty(&oap->oap_rpc_item))
2398                 RETURN(-EBUSY);
2399
2400         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2401         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2402         if (cfs_capable(CFS_CAP_SYS_RESOURCE)) {
2403                 brw_flags |= OBD_BRW_NOQUOTA;
2404                 cmd |= OBD_BRW_NOQUOTA;
2405         }
2406
2407         /* check if the file's owner/group is over quota */
2408         if (!(cmd & OBD_BRW_NOQUOTA)) {
2409                 struct cl_object *obj;
2410                 struct cl_attr   *attr;
2411                 unsigned int qid[LL_MAXQUOTAS];
2412
2413                 obj = cl_object_top(&osc->oo_cl);
2414                 attr = &osc_env_info(env)->oti_attr;
2415
2416                 cl_object_attr_lock(obj);
2417                 rc = cl_object_attr_get(env, obj, attr);
2418                 cl_object_attr_unlock(obj);
2419
2420                 qid[USRQUOTA] = attr->cat_uid;
2421                 qid[GRPQUOTA] = attr->cat_gid;
2422                 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2423                         rc = -EDQUOT;
2424                 if (rc)
2425                         RETURN(rc);
2426         }
2427
2428         oap->oap_cmd = cmd;
2429         oap->oap_page_off = ops->ops_from;
2430         oap->oap_count = ops->ops_to - ops->ops_from;
2431         /* No need to hold a lock here,
2432          * since this page is not in any list yet. */
2433         oap->oap_async_flags = 0;
2434         oap->oap_brw_flags = brw_flags;
2435
2436         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2437                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2438
2439         index = osc_index(oap2osc(oap));
2440
2441         /* Add this page into extent by the following steps:
2442          * 1. if there exists an active extent for this IO, mostly this page
2443          *    can be added to the active extent and sometimes we need to
2444          *    expand extent to accomodate this page;
2445          * 2. otherwise, a new extent will be allocated. */
2446
2447         ext = oio->oi_active;
2448         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2449                 /* one chunk plus extent overhead must be enough to write this
2450                  * page */
2451                 grants = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2452                 if (ext->oe_end >= index)
2453                         grants = 0;
2454
2455                 /* it doesn't need any grant to dirty this page */
2456                 spin_lock(&cli->cl_loi_list_lock);
2457                 rc = osc_enter_cache_try(cli, oap, grants, 0);
2458                 spin_unlock(&cli->cl_loi_list_lock);
2459                 if (rc == 0) { /* try failed */
2460                         grants = 0;
2461                         need_release = 1;
2462                 } else if (ext->oe_end < index) {
2463                         tmp = grants;
2464                         /* try to expand this extent */
2465                         rc = osc_extent_expand(ext, index, &tmp);
2466                         if (rc < 0) {
2467                                 need_release = 1;
2468                                 /* don't free reserved grant */
2469                         } else {
2470                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2471                                                 "expanded for %lu.\n", index);
2472                                 osc_unreserve_grant(cli, grants, tmp);
2473                                 grants = 0;
2474                         }
2475                 }
2476                 rc = 0;
2477         } else if (ext != NULL) {
2478                 /* index is located outside of active extent */
2479                 need_release = 1;
2480         }
2481         if (need_release) {
2482                 osc_extent_release(env, ext);
2483                 oio->oi_active = NULL;
2484                 ext = NULL;
2485         }
2486
2487         if (ext == NULL) {
2488                 tmp = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2489
2490                 /* try to find new extent to cover this page */
2491                 LASSERT(oio->oi_active == NULL);
2492                 /* we may have allocated grant for this page if we failed
2493                  * to expand the previous active extent. */
2494                 LASSERT(ergo(grants > 0, grants >= tmp));
2495
2496                 rc = 0;
2497                 if (grants == 0) {
2498                         /* we haven't allocated grant for this page. */
2499                         rc = osc_enter_cache(env, cli, oap, tmp);
2500                         if (rc == 0)
2501                                 grants = tmp;
2502                 }
2503
2504                 tmp = grants;
2505                 if (rc == 0) {
2506                         ext = osc_extent_find(env, osc, index, &tmp);
2507                         if (IS_ERR(ext)) {
2508                                 LASSERT(tmp == grants);
2509                                 osc_exit_cache(cli, oap);
2510                                 rc = PTR_ERR(ext);
2511                                 ext = NULL;
2512                         } else {
2513                                 oio->oi_active = ext;
2514                         }
2515                 }
2516                 if (grants > 0)
2517                         osc_unreserve_grant(cli, grants, tmp);
2518         }
2519
2520         LASSERT(ergo(rc == 0, ext != NULL));
2521         if (ext != NULL) {
2522                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2523                          ext, "index = %lu.\n", index);
2524                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2525
2526                 osc_object_lock(osc);
2527                 if (ext->oe_nr_pages == 0)
2528                         ext->oe_srvlock = ops->ops_srvlock;
2529                 else
2530                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2531                 ++ext->oe_nr_pages;
2532                 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2533                 osc_object_unlock(osc);
2534         }
2535         RETURN(rc);
2536 }
2537
2538 int osc_teardown_async_page(const struct lu_env *env,
2539                             struct osc_object *obj, struct osc_page *ops)
2540 {
2541         struct osc_async_page *oap = &ops->ops_oap;
2542         int rc = 0;
2543         ENTRY;
2544
2545         LASSERT(oap->oap_magic == OAP_MAGIC);
2546
2547         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2548                oap, ops, osc_index(oap2osc(oap)));
2549
2550         if (!list_empty(&oap->oap_rpc_item)) {
2551                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2552                 rc = -EBUSY;
2553         } else if (!list_empty(&oap->oap_pending_item)) {
2554                 struct osc_extent *ext = NULL;
2555
2556                 osc_object_lock(obj);
2557                 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2558                 osc_object_unlock(obj);
2559                 /* only truncated pages are allowed to be taken out.
2560                  * See osc_extent_truncate() and osc_cache_truncate_start()
2561                  * for details. */
2562                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2563                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2564                                         osc_index(oap2osc(oap)));
2565                         rc = -EBUSY;
2566                 }
2567                 if (ext != NULL)
2568                         osc_extent_put(env, ext);
2569         }
2570         RETURN(rc);
2571 }
2572
2573 /**
2574  * This is called when a page is picked up by kernel to write out.
2575  *
2576  * We should find out the corresponding extent and add the whole extent
2577  * into urgent list. The extent may be being truncated or used, handle it
2578  * carefully.
2579  */
2580 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2581                          struct osc_page *ops)
2582 {
2583         struct osc_extent *ext   = NULL;
2584         struct osc_object *obj   = cl2osc(ops->ops_cl.cpl_obj);
2585         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2586         pgoff_t            index = osc_index(ops);
2587         struct osc_async_page *oap = &ops->ops_oap;
2588         bool unplug = false;
2589         int rc = 0;
2590         ENTRY;
2591
2592         osc_object_lock(obj);
2593         ext = osc_extent_lookup(obj, index);
2594         if (ext == NULL) {
2595                 osc_extent_tree_dump(D_ERROR, obj);
2596                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2597         }
2598
2599         switch (ext->oe_state) {
2600         case OES_RPC:
2601         case OES_LOCK_DONE:
2602                 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2603                 LASSERT(0);
2604                 break;
2605         case OES_LOCKING:
2606                 /* If we know this extent is being written out, we should abort
2607                  * so that the writer can make this page ready. Otherwise, there
2608                  * exists a deadlock problem because other process can wait for
2609                  * page writeback bit holding page lock; and meanwhile in
2610                  * vvp_page_make_ready(), we need to grab page lock before
2611                  * really sending the RPC. */
2612         case OES_TRUNC:
2613                 /* race with truncate, page will be redirtied */
2614         case OES_ACTIVE:
2615                 /* The extent is active so we need to abort and let the caller
2616                  * re-dirty the page. If we continued on here, and we were the
2617                  * one making the extent active, we could deadlock waiting for
2618                  * the page writeback to clear but it won't because the extent
2619                  * is active and won't be written out. */
2620                 GOTO(out, rc = -EAGAIN);
2621         default:
2622                 break;
2623         }
2624
2625         rc = cl_page_prep(env, io, cp, CRT_WRITE);
2626         if (rc)
2627                 GOTO(out, rc);
2628
2629         spin_lock(&oap->oap_lock);
2630         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2631         spin_unlock(&oap->oap_lock);
2632
2633         if (memory_pressure_get())
2634                 ext->oe_memalloc = 1;
2635
2636         ext->oe_urgent = 1;
2637         if (ext->oe_state == OES_CACHE) {
2638                 OSC_EXTENT_DUMP(D_CACHE, ext,
2639                                 "flush page %p make it urgent.\n", oap);
2640                 if (list_empty(&ext->oe_link))
2641                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2642                 unplug = true;
2643         }
2644         rc = 0;
2645         EXIT;
2646
2647 out:
2648         osc_object_unlock(obj);
2649         osc_extent_put(env, ext);
2650         if (unplug)
2651                 osc_io_unplug_async(env, osc_cli(obj), obj);
2652         return rc;
2653 }
2654
2655 /**
2656  * this is called when a sync waiter receives an interruption.  Its job is to
2657  * get the caller woken as soon as possible.  If its page hasn't been put in an
2658  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2659  * desiring interruption which will forcefully complete the rpc once the rpc
2660  * has timed out.
2661  */
2662 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2663 {
2664         struct osc_async_page *oap = &ops->ops_oap;
2665         struct osc_object     *obj = oap->oap_obj;
2666         struct client_obd     *cli = osc_cli(obj);
2667         struct osc_extent     *ext;
2668         struct osc_extent     *found = NULL;
2669         struct list_head            *plist;
2670         pgoff_t index = osc_index(ops);
2671         int     rc = -EBUSY;
2672         int     cmd;
2673         ENTRY;
2674
2675         LASSERT(!oap->oap_interrupted);
2676         oap->oap_interrupted = 1;
2677
2678         /* Find out the caching extent */
2679         osc_object_lock(obj);
2680         if (oap->oap_cmd & OBD_BRW_WRITE) {
2681                 plist = &obj->oo_urgent_exts;
2682                 cmd   = OBD_BRW_WRITE;
2683         } else {
2684                 plist = &obj->oo_reading_exts;
2685                 cmd   = OBD_BRW_READ;
2686         }
2687         list_for_each_entry(ext, plist, oe_link) {
2688                 if (ext->oe_start <= index && ext->oe_end >= index) {
2689                         LASSERT(ext->oe_state == OES_LOCK_DONE);
2690                         /* For OES_LOCK_DONE state extent, it has already held
2691                          * a refcount for RPC. */
2692                         found = osc_extent_get(ext);
2693                         break;
2694                 }
2695         }
2696         if (found != NULL) {
2697                 list_del_init(&found->oe_link);
2698                 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2699                 osc_object_unlock(obj);
2700
2701                 osc_extent_finish(env, found, 0, -EINTR);
2702                 osc_extent_put(env, found);
2703                 rc = 0;
2704         } else {
2705                 osc_object_unlock(obj);
2706                 /* ok, it's been put in an rpc. only one oap gets a request
2707                  * reference */
2708                 if (oap->oap_request != NULL) {
2709                         ptlrpc_mark_interrupted(oap->oap_request);
2710                         ptlrpcd_wake(oap->oap_request);
2711                         ptlrpc_req_finished(oap->oap_request);
2712                         oap->oap_request = NULL;
2713                 }
2714         }
2715
2716         osc_list_maint(cli, obj);
2717         RETURN(rc);
2718 }
2719
2720 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
2721                          struct list_head *list, int cmd, int brw_flags)
2722 {
2723         struct client_obd     *cli = osc_cli(obj);
2724         struct osc_extent     *ext;
2725         struct osc_async_page *oap;
2726         int     page_count = 0;
2727         int     mppr       = cli->cl_max_pages_per_rpc;
2728         bool    can_merge   = true;
2729         pgoff_t start      = CL_PAGE_EOF;
2730         pgoff_t end        = 0;
2731         ENTRY;
2732
2733         list_for_each_entry(oap, list, oap_pending_item) {
2734                 struct osc_page *opg = oap2osc_page(oap);
2735                 pgoff_t index = osc_index(opg);
2736
2737                 if (index > end)
2738                         end = index;
2739                 if (index < start)
2740                         start = index;
2741                 ++page_count;
2742                 mppr <<= (page_count > mppr);
2743
2744                 if (unlikely(opg->ops_from > 0 || opg->ops_to < PAGE_SIZE))
2745                         can_merge = false;
2746         }
2747
2748         ext = osc_extent_alloc(obj);
2749         if (ext == NULL) {
2750                 struct osc_async_page *tmp;
2751
2752                 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2753                         list_del_init(&oap->oap_pending_item);
2754                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2755                 }
2756                 RETURN(-ENOMEM);
2757         }
2758
2759         ext->oe_rw = !!(cmd & OBD_BRW_READ);
2760         ext->oe_sync = 1;
2761         ext->oe_no_merge = !can_merge;
2762         ext->oe_urgent = 1;
2763         ext->oe_start = start;
2764         ext->oe_end = ext->oe_max_end = end;
2765         ext->oe_obj = obj;
2766         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2767         ext->oe_nr_pages = page_count;
2768         ext->oe_mppr = mppr;
2769         list_splice_init(list, &ext->oe_pages);
2770
2771         osc_object_lock(obj);
2772         /* Reuse the initial refcount for RPC, don't drop it */
2773         osc_extent_state_set(ext, OES_LOCK_DONE);
2774         if (cmd & OBD_BRW_WRITE) {
2775                 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2776                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2777         } else {
2778                 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2779                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2780         }
2781         osc_object_unlock(obj);
2782
2783         osc_io_unplug_async(env, cli, obj);
2784         RETURN(0);
2785 }
2786
2787 /**
2788  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2789  */
2790 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
2791                              __u64 size, struct osc_extent **extp)
2792 {
2793         struct client_obd *cli = osc_cli(obj);
2794         struct osc_extent *ext;
2795         struct osc_extent *waiting = NULL;
2796         pgoff_t index;
2797         struct list_head list = LIST_HEAD_INIT(list);
2798         int result = 0;
2799         bool partial;
2800         ENTRY;
2801
2802         /* pages with index greater or equal to index will be truncated. */
2803         index = cl_index(osc2cl(obj), size);
2804         partial = size > cl_offset(osc2cl(obj), index);
2805
2806 again:
2807         osc_object_lock(obj);
2808         ext = osc_extent_search(obj, index);
2809         if (ext == NULL)
2810                 ext = first_extent(obj);
2811         else if (ext->oe_end < index)
2812                 ext = next_extent(ext);
2813         while (ext != NULL) {
2814                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2815
2816                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2817                         /* if ext is in urgent state, it means there must exist
2818                          * a page already having been flushed by write_page().
2819                          * We have to wait for this extent because we can't
2820                          * truncate that page. */
2821                         OSC_EXTENT_DUMP(D_CACHE, ext,
2822                                         "waiting for busy extent\n");
2823                         waiting = osc_extent_get(ext);
2824                         break;
2825                 }
2826
2827                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:%llu.\n", size);
2828
2829                 osc_extent_get(ext);
2830                 if (ext->oe_state == OES_ACTIVE) {
2831                         /* though we grab inode mutex for write path, but we
2832                          * release it before releasing extent(in osc_io_end()),
2833                          * so there is a race window that an extent is still
2834                          * in OES_ACTIVE when truncate starts. */
2835                         LASSERT(!ext->oe_trunc_pending);
2836                         ext->oe_trunc_pending = 1;
2837                 } else {
2838                         EASSERT(ext->oe_state == OES_CACHE, ext);
2839                         osc_extent_state_set(ext, OES_TRUNC);
2840                         osc_update_pending(obj, OBD_BRW_WRITE,
2841                                            -ext->oe_nr_pages);
2842                 }
2843                 /* This extent could be on the full extents list, that's OK */
2844                 EASSERT(!ext->oe_hp && !ext->oe_urgent, ext);
2845                 if (!list_empty(&ext->oe_link))
2846                         list_move_tail(&ext->oe_link, &list);
2847                 else
2848                         list_add_tail(&ext->oe_link, &list);
2849
2850                 ext = next_extent(ext);
2851         }
2852         osc_object_unlock(obj);
2853
2854         osc_list_maint(cli, obj);
2855
2856         while (!list_empty(&list)) {
2857                 int rc;
2858
2859                 ext = list_entry(list.next, struct osc_extent, oe_link);
2860                 list_del_init(&ext->oe_link);
2861
2862                 /* extent may be in OES_ACTIVE state because inode mutex
2863                  * is released before osc_io_end() in file write case */
2864                 if (ext->oe_state != OES_TRUNC)
2865                         osc_extent_wait(env, ext, OES_TRUNC);
2866
2867                 rc = osc_extent_truncate(ext, index, partial);
2868                 if (rc < 0) {
2869                         if (result == 0)
2870                                 result = rc;
2871
2872                         OSC_EXTENT_DUMP(D_ERROR, ext,
2873                                         "truncate error %d\n", rc);
2874                 } else if (ext->oe_nr_pages == 0) {
2875                         osc_extent_remove(ext);
2876                 } else {
2877                         /* this must be an overlapped extent which means only
2878                          * part of pages in this extent have been truncated.
2879                          */
2880                         EASSERTF(ext->oe_start <= index, ext,
2881                                  "trunc index = %lu/%d.\n", index, partial);
2882                         /* fix index to skip this partially truncated extent */
2883                         index = ext->oe_end + 1;
2884                         partial = false;
2885
2886                         /* we need to hold this extent in OES_TRUNC state so
2887                          * that no writeback will happen. This is to avoid
2888                          * BUG 17397.
2889                          * Only partial truncate can reach here, if @size is
2890                          * not zero, the caller should provide a valid @extp. */
2891                         LASSERT(*extp == NULL);
2892                         *extp = osc_extent_get(ext);
2893                         OSC_EXTENT_DUMP(D_CACHE, ext,
2894                                         "trunc at %llu\n", size);
2895                 }
2896                 osc_extent_put(env, ext);
2897         }
2898         if (waiting != NULL) {
2899                 int rc;
2900
2901                 /* ignore the result of osc_extent_wait the write initiator
2902                  * should take care of it. */
2903                 rc = osc_extent_wait(env, waiting, OES_INV);
2904                 if (rc < 0)
2905                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2906
2907                 osc_extent_put(env, waiting);
2908                 waiting = NULL;
2909                 goto again;
2910         }
2911         RETURN(result);
2912 }
2913
2914 /**
2915  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2916  */
2917 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext)
2918 {
2919         if (ext != NULL) {
2920                 struct osc_object *obj = ext->oe_obj;
2921                 bool unplug = false;
2922
2923                 EASSERT(ext->oe_nr_pages > 0, ext);
2924                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2925                 EASSERT(!ext->oe_urgent, ext);
2926
2927                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2928                 osc_object_lock(obj);
2929                 osc_extent_state_set(ext, OES_CACHE);
2930                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2931                         ext->oe_urgent = 1;
2932                         list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2933                         unplug = true;
2934                 }
2935                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2936                 osc_object_unlock(obj);
2937                 osc_extent_put(env, ext);
2938
2939                 if (unplug)
2940                         osc_io_unplug_async(env, osc_cli(obj), obj);
2941         }
2942 }
2943
2944 /**
2945  * Wait for extents in a specific range to be written out.
2946  * The caller must have called osc_cache_writeback_range() to issue IO
2947  * otherwise it will take a long time for this function to finish.
2948  *
2949  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2950  * nobody else can dirty this range of file while we're waiting for
2951  * extents to be written.
2952  */
2953 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2954                          pgoff_t start, pgoff_t end)
2955 {
2956         struct osc_extent *ext;
2957         pgoff_t index = start;
2958         int     result = 0;
2959         ENTRY;
2960
2961 again:
2962         osc_object_lock(obj);
2963         ext = osc_extent_search(obj, index);
2964         if (ext == NULL)
2965                 ext = first_extent(obj);
2966         else if (ext->oe_end < index)
2967                 ext = next_extent(ext);
2968         while (ext != NULL) {
2969                 int rc;
2970
2971                 if (ext->oe_start > end)
2972                         break;
2973
2974                 if (!ext->oe_fsync_wait) {
2975                         ext = next_extent(ext);
2976                         continue;
2977                 }
2978
2979                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2980                              ext->oe_hp || ext->oe_urgent), ext);
2981                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2982                              !ext->oe_hp && ext->oe_urgent), ext);
2983
2984                 index = ext->oe_end + 1;
2985                 osc_extent_get(ext);
2986                 osc_object_unlock(obj);
2987
2988                 rc = osc_extent_wait(env, ext, OES_INV);
2989                 if (result == 0)
2990                         result = rc;
2991                 osc_extent_put(env, ext);
2992                 goto again;
2993         }
2994         osc_object_unlock(obj);
2995
2996         OSC_IO_DEBUG(obj, "sync file range.\n");
2997         RETURN(result);
2998 }
2999
3000 /**
3001  * Called to write out a range of osc object.
3002  *
3003  * @hp     : should be set this is caused by lock cancel;
3004  * @discard: is set if dirty pages should be dropped - file will be deleted or
3005  *         truncated, this implies there is no partially discarding extents.
3006  *
3007  * Return how many pages will be issued, or error code if error occurred.
3008  */
3009 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
3010                               pgoff_t start, pgoff_t end, int hp, int discard)
3011 {
3012         struct osc_extent *ext;
3013         struct list_head discard_list = LIST_HEAD_INIT(discard_list);
3014         bool unplug = false;
3015         int result = 0;
3016         ENTRY;
3017
3018         osc_object_lock(obj);
3019         ext = osc_extent_search(obj, start);
3020         if (ext == NULL)
3021                 ext = first_extent(obj);
3022         else if (ext->oe_end < start)
3023                 ext = next_extent(ext);
3024         while (ext != NULL) {
3025                 if (ext->oe_start > end)
3026                         break;
3027
3028                 ext->oe_fsync_wait = 1;
3029                 switch (ext->oe_state) {
3030                 case OES_CACHE:
3031                         result += ext->oe_nr_pages;
3032                         if (!discard) {
3033                                 struct list_head *list = NULL;
3034                                 if (hp) {
3035                                         EASSERT(!ext->oe_hp, ext);
3036                                         ext->oe_hp = 1;
3037                                         list = &obj->oo_hp_exts;
3038                                 } else if (!ext->oe_urgent) {
3039                                         ext->oe_urgent = 1;
3040                                         list = &obj->oo_urgent_exts;
3041                                 }
3042                                 if (list != NULL)
3043                                         list_move_tail(&ext->oe_link, list);
3044                                 unplug = true;
3045                         } else {
3046                                 /* the only discarder is lock cancelling, so
3047                                  * [start, end] must contain this extent */
3048                                 EASSERT(ext->oe_start >= start &&
3049                                         ext->oe_max_end <= end, ext);
3050                                 osc_extent_state_set(ext, OES_LOCKING);
3051                                 ext->oe_owner = current;
3052                                 list_move_tail(&ext->oe_link,
3053                                                    &discard_list);
3054                                 osc_update_pending(obj, OBD_BRW_WRITE,
3055                                                    -ext->oe_nr_pages);
3056                         }
3057                         break;
3058                 case OES_ACTIVE:
3059                         /* It's pretty bad to wait for ACTIVE extents, because
3060                          * we don't know how long we will wait for it to be
3061                          * flushed since it may be blocked at awaiting more
3062                          * grants. We do this for the correctness of fsync. */
3063                         LASSERT(hp == 0 && discard == 0);
3064                         ext->oe_urgent = 1;
3065                         break;
3066                 case OES_TRUNC:
3067                         /* this extent is being truncated, can't do anything
3068                          * for it now. it will be set to urgent after truncate
3069                          * is finished in osc_cache_truncate_end(). */
3070                 default:
3071                         break;
3072                 }
3073                 ext = next_extent(ext);
3074         }
3075         osc_object_unlock(obj);
3076
3077         LASSERT(ergo(!discard, list_empty(&discard_list)));
3078         if (!list_empty(&discard_list)) {
3079                 struct osc_extent *tmp;
3080                 int rc;
3081
3082                 osc_list_maint(osc_cli(obj), obj);
3083                 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
3084                         list_del_init(&ext->oe_link);
3085                         EASSERT(ext->oe_state == OES_LOCKING, ext);
3086
3087                         /* Discard caching pages. We don't actually write this
3088                          * extent out but we complete it as if we did. */
3089                         rc = osc_extent_make_ready(env, ext);
3090                         if (unlikely(rc < 0)) {
3091                                 OSC_EXTENT_DUMP(D_ERROR, ext,
3092                                                 "make_ready returned %d\n", rc);
3093                                 if (result >= 0)
3094                                         result = rc;
3095                         }
3096
3097                         /* finish the extent as if the pages were sent */
3098                         osc_extent_finish(env, ext, 0, 0);
3099                 }
3100         }
3101
3102         if (unplug)
3103                 osc_io_unplug(env, osc_cli(obj), obj);
3104
3105         if (hp || discard) {
3106                 int rc;
3107                 rc = osc_cache_wait_range(env, obj, start, end);
3108                 if (result >= 0 && rc < 0)
3109                         result = rc;
3110         }
3111
3112         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3113         RETURN(result);
3114 }
3115
3116 /**
3117  * Returns a list of pages by a given [start, end] of \a obj.
3118  *
3119  * \param resched If not NULL, then we give up before hogging CPU for too
3120  * long and set *resched = 1, in that case caller should implement a retry
3121  * logic.
3122  *
3123  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3124  * crucial in the face of [offset, EOF] locks.
3125  *
3126  * Return at least one page in @queue unless there is no covered page.
3127  */
3128 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3129                         struct osc_object *osc, pgoff_t start, pgoff_t end,
3130                         osc_page_gang_cbt cb, void *cbdata)
3131 {
3132         struct osc_page *ops;
3133         void            **pvec;
3134         pgoff_t         idx;
3135         unsigned int    nr;
3136         unsigned int    i;
3137         unsigned int    j;
3138         int             res = CLP_GANG_OKAY;
3139         bool            tree_lock = true;
3140         ENTRY;
3141
3142         idx = start;
3143         pvec = osc_env_info(env)->oti_pvec;
3144         spin_lock(&osc->oo_tree_lock);
3145         while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3146                                             idx, OTI_PVEC_SIZE)) > 0) {
3147                 struct cl_page *page;
3148                 bool end_of_region = false;
3149
3150                 for (i = 0, j = 0; i < nr; ++i) {
3151                         ops = pvec[i];
3152                         pvec[i] = NULL;
3153
3154                         idx = osc_index(ops);
3155                         if (idx > end) {
3156                                 end_of_region = true;
3157                                 break;
3158                         }
3159
3160                         page = ops->ops_cl.cpl_page;
3161                         LASSERT(page->cp_type == CPT_CACHEABLE);
3162                         if (page->cp_state == CPS_FREEING)
3163                                 continue;
3164
3165                         cl_page_get(page);
3166                         lu_ref_add_atomic(&page->cp_reference,
3167                                           "gang_lookup", current);
3168                         pvec[j++] = ops;
3169                 }
3170                 ++idx;
3171
3172                 /*
3173                  * Here a delicate locking dance is performed. Current thread
3174                  * holds a reference to a page, but has to own it before it
3175                  * can be placed into queue. Owning implies waiting, so
3176                  * radix-tree lock is to be released. After a wait one has to
3177                  * check that pages weren't truncated (cl_page_own() returns
3178                  * error in the latter case).
3179                  */
3180                 spin_unlock(&osc->oo_tree_lock);
3181                 tree_lock = false;
3182
3183                 for (i = 0; i < j; ++i) {
3184                         ops = pvec[i];
3185                         if (res == CLP_GANG_OKAY)
3186                                 res = (*cb)(env, io, ops, cbdata);
3187
3188                         page = ops->ops_cl.cpl_page;
3189                         lu_ref_del(&page->cp_reference, "gang_lookup", current);
3190                         cl_page_put(env, page);
3191                 }
3192                 if (nr < OTI_PVEC_SIZE || end_of_region)
3193                         break;
3194
3195                 if (res == CLP_GANG_OKAY && need_resched())
3196                         res = CLP_GANG_RESCHED;
3197                 if (res != CLP_GANG_OKAY)
3198                         break;
3199
3200                 spin_lock(&osc->oo_tree_lock);
3201                 tree_lock = true;
3202         }
3203         if (tree_lock)
3204                 spin_unlock(&osc->oo_tree_lock);
3205         RETURN(res);
3206 }
3207
3208 /**
3209  * Check if page @page is covered by an extra lock or discard it.
3210  */
3211 static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3212                                 struct osc_page *ops, void *cbdata)
3213 {
3214         struct osc_thread_info *info = osc_env_info(env);
3215         struct osc_object *osc = cbdata;
3216         pgoff_t index;
3217
3218         index = osc_index(ops);
3219         if (index >= info->oti_fn_index) {
3220                 struct ldlm_lock *tmp;
3221                 struct cl_page *page = ops->ops_cl.cpl_page;
3222
3223                 /* refresh non-overlapped index */
3224                 tmp = osc_dlmlock_at_pgoff(env, osc, index,
3225                                            OSC_DAP_FL_TEST_LOCK);
3226                 if (tmp != NULL) {
3227                         __u64 end = tmp->l_policy_data.l_extent.end;
3228                         /* Cache the first-non-overlapped index so as to skip
3229                          * all pages within [index, oti_fn_index). This is safe
3230                          * because if tmp lock is canceled, it will discard
3231                          * these pages. */
3232                         info->oti_fn_index = cl_index(osc2cl(osc), end + 1);
3233                         if (end == OBD_OBJECT_EOF)
3234                                 info->oti_fn_index = CL_PAGE_EOF;
3235                         LDLM_LOCK_PUT(tmp);
3236                 } else if (cl_page_own(env, io, page) == 0) {
3237                         /* discard the page */
3238                         cl_page_discard(env, io, page);
3239                         cl_page_disown(env, io, page);
3240                 } else {
3241                         LASSERT(page->cp_state == CPS_FREEING);
3242                 }
3243         }
3244
3245         info->oti_next_index = index + 1;
3246         return CLP_GANG_OKAY;
3247 }
3248
3249 static int discard_cb(const struct lu_env *env, struct cl_io *io,
3250                       struct osc_page *ops, void *cbdata)
3251 {
3252         struct osc_thread_info *info = osc_env_info(env);
3253         struct cl_page *page = ops->ops_cl.cpl_page;
3254
3255         /* page is top page. */
3256         info->oti_next_index = osc_index(ops) + 1;
3257         if (cl_page_own(env, io, page) == 0) {
3258                 if (!ergo(page->cp_type == CPT_CACHEABLE,
3259                           !PageDirty(cl_page_vmpage(page))))
3260                         CL_PAGE_DEBUG(D_ERROR, env, page,
3261                                         "discard dirty page?\n");
3262
3263                 /* discard the page */
3264                 cl_page_discard(env, io, page);
3265                 cl_page_disown(env, io, page);
3266         } else {
3267                 LASSERT(page->cp_state == CPS_FREEING);
3268         }
3269
3270         return CLP_GANG_OKAY;
3271 }
3272
3273 /**
3274  * Discard pages protected by the given lock. This function traverses radix
3275  * tree to find all covering pages and discard them. If a page is being covered
3276  * by other locks, it should remain in cache.
3277  *
3278  * If error happens on any step, the process continues anyway (the reasoning
3279  * behind this being that lock cancellation cannot be delayed indefinitely).
3280  */
3281 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3282                            pgoff_t start, pgoff_t end, enum cl_lock_mode mode)
3283 {
3284         struct osc_thread_info *info = osc_env_info(env);
3285         struct cl_io *io = &info->oti_io;
3286         osc_page_gang_cbt cb;
3287         int res;
3288         int result;
3289
3290         ENTRY;
3291
3292         io->ci_obj = cl_object_top(osc2cl(osc));
3293         io->ci_ignore_layout = 1;
3294         result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3295         if (result != 0)
3296                 GOTO(out, result);
3297
3298         cb = mode == CLM_READ ? check_and_discard_cb : discard_cb;
3299         info->oti_fn_index = info->oti_next_index = start;
3300         do {
3301                 res = osc_page_gang_lookup(env, io, osc,
3302                                            info->oti_next_index, end, cb, osc);
3303                 if (info->oti_next_index > end)
3304                         break;
3305
3306                 if (res == CLP_GANG_RESCHED)
3307                         cond_resched();
3308         } while (res != CLP_GANG_OKAY);
3309 out:
3310         cl_io_fini(env, io);
3311         RETURN(result);
3312 }
3313
3314
3315 /** @} osc */