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, 2016, Intel Corporation.
  27  *
  28  */
  29 /*
  30  * This file is part of Lustre, http://www.lustre.org/
  31  * Lustre is a trademark of Sun Microsystems, Inc.
  32  *
  33  * osc cache management.
  34  *
  35  * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_OSC
  39
  40 #include <lustre_osc.h>
  41
  42 #include "osc_internal.h"
  43
  44 static int extent_debug; /* set it to be true for more debug */
  45
  46 static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
  47 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
  48                            enum osc_extent_state state);
  49 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
  50                               struct osc_async_page *oap, int sent, int rc);
  51 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
  52                           int cmd);
  53 static int osc_refresh_count(const struct lu_env *env,
  54                              struct osc_async_page *oap, int cmd);
  55 static int osc_io_unplug_async(const struct lu_env *env,
  56                                struct client_obd *cli, struct osc_object *osc);
  57 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
  58                            unsigned int lost_grant, unsigned int dirty_grant);
  59
  60 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
  61                                   const char *func, int line);
  62 #define osc_extent_tree_dump(lvl, obj) \
  63         osc_extent_tree_dump0(lvl, obj, __func__, __LINE__)
  64
  65 static void osc_unreserve_grant(struct client_obd *cli, unsigned int reserved,
  66                                 unsigned int unused);
  67
  68 /** \addtogroup osc
  69  *  @{
  70  */
  71
  72 /* ------------------ osc extent ------------------ */
  73 static inline char *ext_flags(struct osc_extent *ext, char *flags)
  74 {
  75         char *buf = flags;
  76         *buf++ = ext->oe_rw ? 'r' : 'w';
  77         if (ext->oe_intree)
  78                 *buf++ = 'i';
  79         if (ext->oe_sync)
  80                 *buf++ = 'S';
  81         if (ext->oe_srvlock)
  82                 *buf++ = 's';
  83         if (ext->oe_hp)
  84                 *buf++ = 'h';
  85         if (ext->oe_urgent)
  86                 *buf++ = 'u';
  87         if (ext->oe_memalloc)
  88                 *buf++ = 'm';
  89         if (ext->oe_trunc_pending)
  90                 *buf++ = 't';
  91         if (ext->oe_fsync_wait)
  92                 *buf++ = 'Y';
  93         *buf = 0;
  94         return flags;
  95 }
  96
  97 static inline char list_empty_marker(struct list_head *list)
  98 {
  99         return list_empty(list) ? '-' : '+';
 100 }
 101
 102 #define EXTSTR       "[%lu -> %lu/%lu]"
 103 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
 104 static const char *oes_strings[] = {
 105         "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
 106
 107 #define OSC_EXTENT_DUMP(lvl, extent, fmt, ...) do {                           \
 108         struct osc_extent *__ext = (extent);                                  \
 109         char __buf[16];                                                       \
 110                                                                               \
 111         CDEBUG(lvl,                                                           \
 112                 "extent %p@{" EXTSTR ", "                                     \
 113                 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt,          \
 114                 /* ----- extent part 0 ----- */                               \
 115                 __ext, EXTPARA(__ext),                                        \
 116                 /* ----- part 1 ----- */                                      \
 117                 atomic_read(&__ext->oe_refc),                         \
 118                 atomic_read(&__ext->oe_users),                        \
 119                 list_empty_marker(&__ext->oe_link),                           \
 120                 oes_strings[__ext->oe_state], ext_flags(__ext, __buf),        \
 121                 __ext->oe_obj,                                                \
 122                 /* ----- part 2 ----- */                                      \
 123                 __ext->oe_grants, __ext->oe_nr_pages,                         \
 124                 list_empty_marker(&__ext->oe_pages),                          \
 125                 waitqueue_active(&__ext->oe_waitq) ? '+' : '-',               \
 126                 __ext->oe_dlmlock, __ext->oe_mppr, __ext->oe_owner,           \
 127                 /* ----- part 4 ----- */                                      \
 128                 ## __VA_ARGS__);                                              \
 129         if (lvl == D_ERROR && __ext->oe_dlmlock != NULL)                      \
 130                 LDLM_ERROR(__ext->oe_dlmlock, "extent: %p", __ext);           \
 131         else                                                                  \
 132                 LDLM_DEBUG(__ext->oe_dlmlock, "extent: %p", __ext);           \
 133 } while (0)
 134
 135 #undef EASSERTF
 136 #define EASSERTF(expr, ext, fmt, args...) do {                          \
 137         if (!(expr)) {                                                  \
 138                 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args);           \
 139                 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj);           \
 140                 LASSERT(expr);                                          \
 141         }                                                               \
 142 } while (0)
 143
 144 #undef EASSERT
 145 #define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
 146
 147 static inline struct osc_extent *rb_extent(struct rb_node *n)
 148 {
 149         if (n == NULL)
 150                 return NULL;
 151
 152         return container_of(n, struct osc_extent, oe_node);
 153 }
 154
 155 static inline struct osc_extent *next_extent(struct osc_extent *ext)
 156 {
 157         if (ext == NULL)
 158                 return NULL;
 159
 160         LASSERT(ext->oe_intree);
 161         return rb_extent(rb_next(&ext->oe_node));
 162 }
 163
 164 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
 165 {
 166         if (ext == NULL)
 167                 return NULL;
 168
 169         LASSERT(ext->oe_intree);
 170         return rb_extent(rb_prev(&ext->oe_node));
 171 }
 172
 173 static inline struct osc_extent *first_extent(struct osc_object *obj)
 174 {
 175         return rb_extent(rb_first(&obj->oo_root));
 176 }
 177
 178 /* object must be locked by caller. */
 179 static int osc_extent_sanity_check0(struct osc_extent *ext,
 180                                     const char *func, const int line)
 181 {
 182         struct osc_object *obj = ext->oe_obj;
 183         struct osc_async_page *oap;
 184         size_t page_count;
 185         int rc = 0;
 186
 187         if (!osc_object_is_locked(obj))
 188                 GOTO(out, rc = 9);
 189
 190         if (ext->oe_state >= OES_STATE_MAX)
 191                 GOTO(out, rc = 10);
 192
 193         if (atomic_read(&ext->oe_refc) <= 0)
 194                 GOTO(out, rc = 20);
 195
 196         if (atomic_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
 197                 GOTO(out, rc = 30);
 198
 199         switch (ext->oe_state) {
 200         case OES_INV:
 201                 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
 202                         GOTO(out, rc = 35);
 203                 GOTO(out, rc = 0);
 204                 break;
 205         case OES_ACTIVE:
 206                 if (atomic_read(&ext->oe_users) == 0)
 207                         GOTO(out, rc = 40);
 208                 if (ext->oe_hp)
 209                         GOTO(out, rc = 50);
 210                 if (ext->oe_fsync_wait && !ext->oe_urgent)
 211                         GOTO(out, rc = 55);
 212                 break;
 213         case OES_CACHE:
 214                 if (ext->oe_grants == 0)
 215                         GOTO(out, rc = 60);
 216                 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
 217                         GOTO(out, rc = 65);
 218         default:
 219                 if (atomic_read(&ext->oe_users) > 0)
 220                         GOTO(out, rc = 70);
 221         }
 222
 223         if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
 224                 GOTO(out, rc = 80);
 225
 226         if (ext->oe_sync && ext->oe_grants > 0)
 227                 GOTO(out, rc = 90);
 228
 229         if (ext->oe_dlmlock != NULL && !ldlm_is_failed(ext->oe_dlmlock)) {
 230                 struct ldlm_extent *extent;
 231
 232                 extent = &ext->oe_dlmlock->l_policy_data.l_extent;
 233                 if (!(extent->start <= cl_offset(osc2cl(obj), ext->oe_start) &&
 234                       extent->end   >= cl_offset(osc2cl(obj), ext->oe_max_end)))
 235                         GOTO(out, rc = 100);
 236
 237                 if (!(ext->oe_dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP)))
 238                         GOTO(out, rc = 102);
 239         }
 240
 241         if (ext->oe_nr_pages > ext->oe_mppr)
 242                 GOTO(out, rc = 105);
 243
 244         /* Do not verify page list if extent is in RPC. This is because an
 245          * in-RPC extent is supposed to be exclusively accessible w/o lock. */
 246         if (ext->oe_state > OES_CACHE)
 247                 GOTO(out, rc = 0);
 248
 249         if (!extent_debug)
 250                 GOTO(out, rc = 0);
 251
 252         page_count = 0;
 253         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
 254                 pgoff_t index = osc_index(oap2osc(oap));
 255                 ++page_count;
 256                 if (index > ext->oe_end || index < ext->oe_start)
 257                         GOTO(out, rc = 110);
 258         }
 259         if (page_count != ext->oe_nr_pages)
 260                 GOTO(out, rc = 120);
 261
 262 out:
 263         if (rc != 0)
 264                 OSC_EXTENT_DUMP(D_ERROR, ext,
 265                                 "%s:%d sanity check %p failed with rc = %d\n",
 266                                 func, line, ext, rc);
 267         return rc;
 268 }
 269
 270 #define sanity_check_nolock(ext) \
 271         osc_extent_sanity_check0(ext, __func__, __LINE__)
 272
 273 #define sanity_check(ext) ({                                                   \
 274         int __res;                                                             \
 275         osc_object_lock((ext)->oe_obj);                                        \
 276         __res = sanity_check_nolock(ext);                                      \
 277         osc_object_unlock((ext)->oe_obj);                                      \
 278         __res;                                                                 \
 279 })
 280
 281
 282 /**
 283  * sanity check - to make sure there is no overlapped extent in the tree.
 284  */
 285 static int osc_extent_is_overlapped(struct osc_object *obj,
 286                                     struct osc_extent *ext)
 287 {
 288         struct osc_extent *tmp;
 289
 290         LASSERT(osc_object_is_locked(obj));
 291
 292         if (!extent_debug)
 293                 return 0;
 294
 295         for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
 296                 if (tmp == ext)
 297                         continue;
 298                 if (tmp->oe_end >= ext->oe_start &&
 299                     tmp->oe_start <= ext->oe_end)
 300                         return 1;
 301         }
 302         return 0;
 303 }
 304
 305 static void osc_extent_state_set(struct osc_extent *ext, int state)
 306 {
 307         LASSERT(osc_object_is_locked(ext->oe_obj));
 308         LASSERT(state >= OES_INV && state < OES_STATE_MAX);
 309
 310         /* Never try to sanity check a state changing extent :-) */
 311         /* LASSERT(sanity_check_nolock(ext) == 0); */
 312
 313         /* TODO: validate the state machine */
 314         ext->oe_state = state;
 315         wake_up_all(&ext->oe_waitq);
 316 }
 317
 318 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
 319 {
 320         struct osc_extent *ext;
 321
 322         OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_NOFS);
 323         if (ext == NULL)
 324                 return NULL;
 325
 326         RB_CLEAR_NODE(&ext->oe_node);
 327         ext->oe_obj = obj;
 328         cl_object_get(osc2cl(obj));
 329         atomic_set(&ext->oe_refc, 1);
 330         atomic_set(&ext->oe_users, 0);
 331         INIT_LIST_HEAD(&ext->oe_link);
 332         ext->oe_state = OES_INV;
 333         INIT_LIST_HEAD(&ext->oe_pages);
 334         init_waitqueue_head(&ext->oe_waitq);
 335         ext->oe_dlmlock = NULL;
 336
 337         return ext;
 338 }
 339
 340 static void osc_extent_free(struct osc_extent *ext)
 341 {
 342         OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
 343 }
 344
 345 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
 346 {
 347         LASSERT(atomic_read(&ext->oe_refc) >= 0);
 348         atomic_inc(&ext->oe_refc);
 349         return ext;
 350 }
 351
 352 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
 353 {
 354         LASSERT(atomic_read(&ext->oe_refc) > 0);
 355         if (atomic_dec_and_test(&ext->oe_refc)) {
 356                 LASSERT(list_empty(&ext->oe_link));
 357                 LASSERT(atomic_read(&ext->oe_users) == 0);
 358                 LASSERT(ext->oe_state == OES_INV);
 359                 LASSERT(!ext->oe_intree);
 360
 361                 if (ext->oe_dlmlock != NULL) {
 362                         lu_ref_add(&ext->oe_dlmlock->l_reference,
 363                                    "osc_extent", ext);
 364                         LDLM_LOCK_PUT(ext->oe_dlmlock);
 365                         ext->oe_dlmlock = NULL;
 366                 }
 367                 cl_object_put(env, osc2cl(ext->oe_obj));
 368                 osc_extent_free(ext);
 369         }
 370 }
 371
 372 /**
 373  * osc_extent_put_trust() is a special version of osc_extent_put() when
 374  * it's known that the caller is not the last user. This is to address the
 375  * problem of lacking of lu_env ;-).
 376  */
 377 static void osc_extent_put_trust(struct osc_extent *ext)
 378 {
 379         LASSERT(atomic_read(&ext->oe_refc) > 1);
 380         LASSERT(osc_object_is_locked(ext->oe_obj));
 381         atomic_dec(&ext->oe_refc);
 382 }
 383
 384 /**
 385  * Return the extent which includes pgoff @index, or return the greatest
 386  * previous extent in the tree.
 387  */
 388 static struct osc_extent *osc_extent_search(struct osc_object *obj,
 389                                             pgoff_t index)
 390 {
 391         struct rb_node    *n = obj->oo_root.rb_node;
 392         struct osc_extent *tmp, *p = NULL;
 393
 394         LASSERT(osc_object_is_locked(obj));
 395         while (n != NULL) {
 396                 tmp = rb_extent(n);
 397                 if (index < tmp->oe_start) {
 398                         n = n->rb_left;
 399                 } else if (index > tmp->oe_end) {
 400                         p = rb_extent(n);
 401                         n = n->rb_right;
 402                 } else {
 403                         return tmp;
 404                 }
 405         }
 406         return p;
 407 }
 408
 409 /*
 410  * Return the extent covering @index, otherwise return NULL.
 411  * caller must have held object lock.
 412  */
 413 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
 414                                             pgoff_t index)
 415 {
 416         struct osc_extent *ext;
 417
 418         ext = osc_extent_search(obj, index);
 419         if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
 420                 return osc_extent_get(ext);
 421         return NULL;
 422 }
 423
 424 /* caller must have held object lock. */
 425 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
 426 {
 427         struct rb_node   **n      = &obj->oo_root.rb_node;
 428         struct rb_node    *parent = NULL;
 429         struct osc_extent *tmp;
 430
 431         LASSERT(ext->oe_intree == 0);
 432         LASSERT(ext->oe_obj == obj);
 433         LASSERT(osc_object_is_locked(obj));
 434         while (*n != NULL) {
 435                 tmp = rb_extent(*n);
 436                 parent = *n;
 437
 438                 if (ext->oe_end < tmp->oe_start)
 439                         n = &(*n)->rb_left;
 440                 else if (ext->oe_start > tmp->oe_end)
 441                         n = &(*n)->rb_right;
 442                 else
 443                         EASSERTF(0, tmp, EXTSTR"\n", EXTPARA(ext));
 444         }
 445         rb_link_node(&ext->oe_node, parent, n);
 446         rb_insert_color(&ext->oe_node, &obj->oo_root);
 447         osc_extent_get(ext);
 448         ext->oe_intree = 1;
 449 }
 450
 451 /* caller must have held object lock. */
 452 static void osc_extent_erase(struct osc_extent *ext)
 453 {
 454         struct osc_object *obj = ext->oe_obj;
 455         LASSERT(osc_object_is_locked(obj));
 456         if (ext->oe_intree) {
 457                 rb_erase(&ext->oe_node, &obj->oo_root);
 458                 ext->oe_intree = 0;
 459                 /* rbtree held a refcount */
 460                 osc_extent_put_trust(ext);
 461         }
 462 }
 463
 464 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
 465 {
 466         struct osc_object *obj = ext->oe_obj;
 467
 468         LASSERT(osc_object_is_locked(obj));
 469         LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
 470         if (ext->oe_state == OES_CACHE) {
 471                 osc_extent_state_set(ext, OES_ACTIVE);
 472                 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
 473         }
 474         atomic_inc(&ext->oe_users);
 475         list_del_init(&ext->oe_link);
 476         return osc_extent_get(ext);
 477 }
 478
 479 static void __osc_extent_remove(struct osc_extent *ext)
 480 {
 481         LASSERT(osc_object_is_locked(ext->oe_obj));
 482         LASSERT(list_empty(&ext->oe_pages));
 483         osc_extent_erase(ext);
 484         list_del_init(&ext->oe_link);
 485         osc_extent_state_set(ext, OES_INV);
 486         OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
 487 }
 488
 489 static void osc_extent_remove(struct osc_extent *ext)
 490 {
 491         struct osc_object *obj = ext->oe_obj;
 492
 493         osc_object_lock(obj);
 494         __osc_extent_remove(ext);
 495         osc_object_unlock(obj);
 496 }
 497
 498 /**
 499  * This function is used to merge extents to get better performance. It checks
 500  * if @cur and @victim are contiguous at block level.
 501  */
 502 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
 503                             struct osc_extent *victim)
 504 {
 505         struct osc_object       *obj = cur->oe_obj;
 506         struct client_obd       *cli = osc_cli(obj);
 507         pgoff_t                  chunk_start;
 508         pgoff_t                  chunk_end;
 509         int                      ppc_bits;
 510
 511         LASSERT(cur->oe_state == OES_CACHE);
 512         LASSERT(osc_object_is_locked(obj));
 513         if (victim == NULL)
 514                 return -EINVAL;
 515
 516         if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
 517                 return -EBUSY;
 518
 519         if (cur->oe_max_end != victim->oe_max_end)
 520                 return -ERANGE;
 521
 522         LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
 523         ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_SHIFT;
 524         chunk_start = cur->oe_start >> ppc_bits;
 525         chunk_end   = cur->oe_end   >> ppc_bits;
 526         if (chunk_start   != (victim->oe_end >> ppc_bits) + 1 &&
 527             chunk_end + 1 != victim->oe_start >> ppc_bits)
 528                 return -ERANGE;
 529
 530         /* overall extent size should not exceed the max supported limit
 531          * reported by the server */
 532         if (cur->oe_end - cur->oe_start + 1 +
 533             victim->oe_end - victim->oe_start + 1 > cli->cl_max_extent_pages)
 534                 return -ERANGE;
 535
 536         OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
 537
 538         cur->oe_start     = min(cur->oe_start, victim->oe_start);
 539         cur->oe_end       = max(cur->oe_end,   victim->oe_end);
 540         /* per-extent tax should be accounted only once for the whole extent */
 541         cur->oe_grants   += victim->oe_grants - cli->cl_grant_extent_tax;
 542         cur->oe_nr_pages += victim->oe_nr_pages;
 543         /* only the following bits are needed to merge */
 544         cur->oe_urgent   |= victim->oe_urgent;
 545         cur->oe_memalloc |= victim->oe_memalloc;
 546         list_splice_init(&victim->oe_pages, &cur->oe_pages);
 547         list_del_init(&victim->oe_link);
 548         victim->oe_nr_pages = 0;
 549
 550         osc_extent_get(victim);
 551         __osc_extent_remove(victim);
 552         osc_extent_put(env, victim);
 553
 554         OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
 555         return 0;
 556 }
 557
 558 /**
 559  * Drop user count of osc_extent, and unplug IO asynchronously.
 560  */
 561 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
 562 {
 563         struct osc_object *obj = ext->oe_obj;
 564         struct client_obd *cli = osc_cli(obj);
 565         int rc = 0;
 566         ENTRY;
 567
 568         LASSERT(atomic_read(&ext->oe_users) > 0);
 569         LASSERT(sanity_check(ext) == 0);
 570         LASSERT(ext->oe_grants > 0);
 571
 572         if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
 573                 LASSERT(ext->oe_state == OES_ACTIVE);
 574                 if (ext->oe_trunc_pending) {
 575                         /* a truncate process is waiting for this extent.
 576                          * This may happen due to a race, check
 577                          * osc_cache_truncate_start(). */
 578                         osc_extent_state_set(ext, OES_TRUNC);
 579                         ext->oe_trunc_pending = 0;
 580                 } else {
 581                         int grant = 0;
 582
 583                         osc_extent_state_set(ext, OES_CACHE);
 584                         osc_update_pending(obj, OBD_BRW_WRITE,
 585                                            ext->oe_nr_pages);
 586
 587                         /* try to merge the previous and next extent. */
 588                         if (osc_extent_merge(env, ext, prev_extent(ext)) == 0)
 589                                 grant += cli->cl_grant_extent_tax;
 590                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 591                                 grant += cli->cl_grant_extent_tax;
 592                         if (grant > 0)
 593                                 osc_unreserve_grant(cli, 0, grant);
 594
 595                         if (ext->oe_urgent)
 596                                 list_move_tail(&ext->oe_link,
 597                                                &obj->oo_urgent_exts);
 598                         else if (ext->oe_nr_pages == ext->oe_mppr) {
 599                                 list_move_tail(&ext->oe_link,
 600                                                &obj->oo_full_exts);
 601                         }
 602                 }
 603                 osc_object_unlock(obj);
 604
 605                 osc_io_unplug_async(env, cli, obj);
 606         }
 607         osc_extent_put(env, ext);
 608         RETURN(rc);
 609 }
 610
 611 static inline int overlapped(struct osc_extent *ex1, struct osc_extent *ex2)
 612 {
 613         return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
 614 }
 615
 616 /**
 617  * Find or create an extent which includes @index, core function to manage
 618  * extent tree.
 619  */
 620 static struct osc_extent *osc_extent_find(const struct lu_env *env,
 621                                           struct osc_object *obj, pgoff_t index,
 622                                           unsigned int *grants)
 623 {
 624         struct client_obd *cli = osc_cli(obj);
 625         struct osc_lock   *olck;
 626         struct cl_lock_descr *descr;
 627         struct osc_extent *cur;
 628         struct osc_extent *ext;
 629         struct osc_extent *conflict = NULL;
 630         struct osc_extent *found = NULL;
 631         pgoff_t    chunk;
 632         pgoff_t    max_end;
 633         unsigned int max_pages; /* max_pages_per_rpc */
 634         unsigned int chunksize;
 635         int        ppc_bits; /* pages per chunk bits */
 636         pgoff_t    chunk_mask;
 637         int        rc;
 638         ENTRY;
 639
 640         cur = osc_extent_alloc(obj);
 641         if (cur == NULL)
 642                 RETURN(ERR_PTR(-ENOMEM));
 643
 644         olck = osc_env_io(env)->oi_write_osclock;
 645         LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
 646         LASSERT(olck->ols_state == OLS_GRANTED);
 647
 648         descr = &olck->ols_cl.cls_lock->cll_descr;
 649         LASSERT(descr->cld_mode >= CLM_WRITE);
 650
 651         LASSERTF(cli->cl_chunkbits >= PAGE_SHIFT,
 652                  "chunkbits: %u\n", cli->cl_chunkbits);
 653         ppc_bits   = cli->cl_chunkbits - PAGE_SHIFT;
 654         chunk_mask = ~((1 << ppc_bits) - 1);
 655         chunksize  = 1 << cli->cl_chunkbits;
 656         chunk      = index >> ppc_bits;
 657
 658         /* align end to RPC edge. */
 659         max_pages = cli->cl_max_pages_per_rpc;
 660         if ((max_pages & ~chunk_mask) != 0) {
 661                 CERROR("max_pages: %#x chunkbits: %u chunk_mask: %#lx\n",
 662                        max_pages, cli->cl_chunkbits, chunk_mask);
 663                 RETURN(ERR_PTR(-EINVAL));
 664         }
 665         max_end = index - (index % max_pages) + max_pages - 1;
 666         max_end = min_t(pgoff_t, max_end, descr->cld_end);
 667
 668         /* initialize new extent by parameters so far */
 669         cur->oe_max_end = max_end;
 670         cur->oe_start   = index & chunk_mask;
 671         cur->oe_end     = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
 672         if (cur->oe_start < descr->cld_start)
 673                 cur->oe_start = descr->cld_start;
 674         if (cur->oe_end > max_end)
 675                 cur->oe_end = max_end;
 676         cur->oe_grants  = 0;
 677         cur->oe_mppr    = max_pages;
 678         if (olck->ols_dlmlock != NULL) {
 679                 LASSERT(olck->ols_hold);
 680                 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
 681                 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
 682         }
 683
 684         /* grants has been allocated by caller */
 685         LASSERTF(*grants >= chunksize + cli->cl_grant_extent_tax,
 686                  "%u/%u/%u.\n", *grants, chunksize, cli->cl_grant_extent_tax);
 687         LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
 688                  EXTPARA(cur));
 689
 690 restart:
 691         osc_object_lock(obj);
 692         ext = osc_extent_search(obj, cur->oe_start);
 693         if (ext == NULL)
 694                 ext = first_extent(obj);
 695         while (ext != NULL) {
 696                 pgoff_t ext_chk_start = ext->oe_start >> ppc_bits;
 697                 pgoff_t ext_chk_end   = ext->oe_end   >> ppc_bits;
 698
 699                 LASSERT(sanity_check_nolock(ext) == 0);
 700                 if (chunk > ext_chk_end + 1)
 701                         break;
 702
 703                 /* if covering by different locks, no chance to match */
 704                 if (olck->ols_dlmlock != ext->oe_dlmlock) {
 705                         EASSERTF(!overlapped(ext, cur), ext,
 706                                  EXTSTR"\n", EXTPARA(cur));
 707
 708                         ext = next_extent(ext);
 709                         continue;
 710                 }
 711
 712                 /* discontiguous chunks? */
 713                 if (chunk + 1 < ext_chk_start) {
 714                         ext = next_extent(ext);
 715                         continue;
 716                 }
 717
 718                 /* ok, from now on, ext and cur have these attrs:
 719                  * 1. covered by the same lock
 720                  * 2. contiguous at chunk level or overlapping. */
 721
 722                 if (overlapped(ext, cur)) {
 723                         /* cur is the minimum unit, so overlapping means
 724                          * full contain. */
 725                         EASSERTF((ext->oe_start <= cur->oe_start &&
 726                                   ext->oe_end >= cur->oe_end),
 727                                  ext, EXTSTR"\n", EXTPARA(cur));
 728
 729                         if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
 730                                 /* for simplicity, we wait for this extent to
 731                                  * finish before going forward. */
 732                                 conflict = osc_extent_get(ext);
 733                                 break;
 734                         }
 735
 736                         found = osc_extent_hold(ext);
 737                         break;
 738                 }
 739
 740                 /* non-overlapped extent */
 741                 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
 742                         /* we can't do anything for a non OES_CACHE extent, or
 743                          * if there is someone waiting for this extent to be
 744                          * flushed, try next one. */
 745                         ext = next_extent(ext);
 746                         continue;
 747                 }
 748
 749                 /* check if they belong to the same rpc slot before trying to
 750                  * merge. the extents are not overlapped and contiguous at
 751                  * chunk level to get here. */
 752                 if (ext->oe_max_end != max_end) {
 753                         /* if they don't belong to the same RPC slot or
 754                          * max_pages_per_rpc has ever changed, do not merge. */
 755                         ext = next_extent(ext);
 756                         continue;
 757                 }
 758
 759                 /* check whether maximum extent size will be hit */
 760                 if ((ext_chk_end - ext_chk_start + 1 + 1) << ppc_bits >
 761                     cli->cl_max_extent_pages) {
 762                         ext = next_extent(ext);
 763                         continue;
 764                 }
 765
 766                 /* it's required that an extent must be contiguous at chunk
 767                  * level so that we know the whole extent is covered by grant
 768                  * (the pages in the extent are NOT required to be contiguous).
 769                  * Otherwise, it will be too much difficult to know which
 770                  * chunks have grants allocated. */
 771
 772                 /* try to do front merge - extend ext's start */
 773                 if (chunk + 1 == ext_chk_start) {
 774                         /* ext must be chunk size aligned */
 775                         EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
 776
 777                         /* pull ext's start back to cover cur */
 778                         ext->oe_start   = cur->oe_start;
 779                         ext->oe_grants += chunksize;
 780                         LASSERT(*grants >= chunksize);
 781                         *grants -= chunksize;
 782
 783                         found = osc_extent_hold(ext);
 784                 } else if (chunk == ext_chk_end + 1) {
 785                         /* rear merge */
 786                         ext->oe_end     = cur->oe_end;
 787                         ext->oe_grants += chunksize;
 788                         LASSERT(*grants >= chunksize);
 789                         *grants -= chunksize;
 790
 791                         /* try to merge with the next one because we just fill
 792                          * in a gap */
 793                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 794                                 /* we can save extent tax from next extent */
 795                                 *grants += cli->cl_grant_extent_tax;
 796
 797                         found = osc_extent_hold(ext);
 798                 }
 799                 if (found != NULL)
 800                         break;
 801
 802                 ext = next_extent(ext);
 803         }
 804
 805         osc_extent_tree_dump(D_CACHE, obj);
 806         if (found != NULL) {
 807                 LASSERT(conflict == NULL);
 808                 if (!IS_ERR(found)) {
 809                         LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
 810                         OSC_EXTENT_DUMP(D_CACHE, found,
 811                                         "found caching ext for %lu.\n", index);
 812                 }
 813         } else if (conflict == NULL) {
 814                 /* create a new extent */
 815                 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
 816                 cur->oe_grants = chunksize + cli->cl_grant_extent_tax;
 817                 LASSERT(*grants >= cur->oe_grants);
 818                 *grants -= cur->oe_grants;
 819
 820                 cur->oe_state = OES_CACHE;
 821                 found = osc_extent_hold(cur);
 822                 osc_extent_insert(obj, cur);
 823                 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
 824                                 index, descr->cld_end);
 825         }
 826         osc_object_unlock(obj);
 827
 828         if (conflict != NULL) {
 829                 LASSERT(found == NULL);
 830
 831                 /* waiting for IO to finish. Please notice that it's impossible
 832                  * to be an OES_TRUNC extent. */
 833                 rc = osc_extent_wait(env, conflict, OES_INV);
 834                 osc_extent_put(env, conflict);
 835                 conflict = NULL;
 836                 if (rc < 0)
 837                         GOTO(out, found = ERR_PTR(rc));
 838
 839                 goto restart;
 840         }
 841         EXIT;
 842
 843 out:
 844         osc_extent_put(env, cur);
 845         return found;
 846 }
 847
 848 /**
 849  * Called when IO is finished to an extent.
 850  */
 851 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
 852                       int sent, int rc)
 853 {
 854         struct client_obd *cli = osc_cli(ext->oe_obj);
 855         struct osc_async_page *oap;
 856         struct osc_async_page *tmp;
 857         int nr_pages = ext->oe_nr_pages;
 858         int lost_grant = 0;
 859         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
 860         loff_t last_off = 0;
 861         int last_count = -1;
 862         ENTRY;
 863
 864         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
 865
 866         ext->oe_rc = rc ?: ext->oe_nr_pages;
 867         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
 868
 869         osc_lru_add_batch(cli, &ext->oe_pages);
 870         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 871                                      oap_pending_item) {
 872                 list_del_init(&oap->oap_rpc_item);
 873                 list_del_init(&oap->oap_pending_item);
 874                 if (last_off <= oap->oap_obj_off) {
 875                         last_off = oap->oap_obj_off;
 876                         last_count = oap->oap_count;
 877                 }
 878
 879                 --ext->oe_nr_pages;
 880                 osc_ap_completion(env, cli, oap, sent, rc);
 881         }
 882         EASSERT(ext->oe_nr_pages == 0, ext);
 883
 884         if (!sent) {
 885                 lost_grant = ext->oe_grants;
 886         } else if (blocksize < PAGE_SIZE &&
 887                    last_count != PAGE_SIZE) {
 888                 /* For short writes we shouldn't count parts of pages that
 889                  * span a whole chunk on the OST side, or our accounting goes
 890                  * wrong.  Should match the code in filter_grant_check. */
 891                 int offset = last_off & ~PAGE_MASK;
 892                 int count = last_count + (offset & (blocksize - 1));
 893                 int end = (offset + last_count) & (blocksize - 1);
 894                 if (end)
 895                         count += blocksize - end;
 896
 897                 lost_grant = PAGE_SIZE - count;
 898         }
 899         if (ext->oe_grants > 0)
 900                 osc_free_grant(cli, nr_pages, lost_grant, ext->oe_grants);
 901
 902         osc_extent_remove(ext);
 903         /* put the refcount for RPC */
 904         osc_extent_put(env, ext);
 905         RETURN(0);
 906 }
 907
 908 static int extent_wait_cb(struct osc_extent *ext, enum osc_extent_state state)
 909 {
 910         int ret;
 911
 912         osc_object_lock(ext->oe_obj);
 913         ret = ext->oe_state == state;
 914         osc_object_unlock(ext->oe_obj);
 915
 916         return ret;
 917 }
 918
 919 /**
 920  * Wait for the extent's state to become @state.
 921  */
 922 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
 923                            enum osc_extent_state state)
 924 {
 925         struct osc_object *obj = ext->oe_obj;
 926         struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
 927                                                   LWI_ON_SIGNAL_NOOP, NULL);
 928         int rc = 0;
 929         ENTRY;
 930
 931         osc_object_lock(obj);
 932         LASSERT(sanity_check_nolock(ext) == 0);
 933         /* `Kick' this extent only if the caller is waiting for it to be
 934          * written out. */
 935         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
 936                 if (ext->oe_state == OES_ACTIVE) {
 937                         ext->oe_urgent = 1;
 938                 } else if (ext->oe_state == OES_CACHE) {
 939                         ext->oe_urgent = 1;
 940                         osc_extent_hold(ext);
 941                         rc = 1;
 942                 }
 943         }
 944         osc_object_unlock(obj);
 945         if (rc == 1)
 946                 osc_extent_release(env, ext);
 947
 948         /* wait for the extent until its state becomes @state */
 949         rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
 950         if (rc == -ETIMEDOUT) {
 951                 OSC_EXTENT_DUMP(D_ERROR, ext,
 952                         "%s: wait ext to %u timedout, recovery in progress?\n",
 953                         cli_name(osc_cli(obj)), state);
 954
 955                 lwi = LWI_INTR(NULL, NULL);
 956                 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state),
 957                                   &lwi);
 958         }
 959         if (rc == 0 && ext->oe_rc < 0)
 960                 rc = ext->oe_rc;
 961         RETURN(rc);
 962 }
 963
 964 /**
 965  * Discard pages with index greater than @size. If @ext is overlapped with
 966  * @size, then partial truncate happens.
 967  */
 968 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
 969                                 bool partial)
 970 {
 971         struct lu_env         *env;
 972         struct cl_io          *io;
 973         struct osc_object     *obj = ext->oe_obj;
 974         struct client_obd     *cli = osc_cli(obj);
 975         struct osc_async_page *oap;
 976         struct osc_async_page *tmp;
 977         int                    pages_in_chunk = 0;
 978         int                    ppc_bits    = cli->cl_chunkbits -
 979                                              PAGE_SHIFT;
 980         __u64                  trunc_chunk = trunc_index >> ppc_bits;
 981         int                    grants   = 0;
 982         int                    nr_pages = 0;
 983         int                    rc       = 0;
 984         __u16                  refcheck;
 985         ENTRY;
 986
 987         LASSERT(sanity_check(ext) == 0);
 988         LASSERT(ext->oe_state == OES_TRUNC);
 989         LASSERT(!ext->oe_urgent);
 990
 991         /* Request new lu_env.
 992          * We can't use that env from osc_cache_truncate_start() because
 993          * it's from lov_io_sub and not fully initialized. */
 994         env = cl_env_get(&refcheck);
 995         if (IS_ERR(env))
 996                 RETURN(PTR_ERR(env));
 997
 998         io  = &osc_env_info(env)->oti_io;
 999         io->ci_obj = cl_object_top(osc2cl(obj));
1000         io->ci_ignore_layout = 1;
1001         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
1002         if (rc < 0)
1003                 GOTO(out, rc);
1004
1005         /* discard all pages with index greater than trunc_index */
1006         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
1007                                      oap_pending_item) {
1008                 pgoff_t index = osc_index(oap2osc(oap));
1009                 struct cl_page  *page = oap2cl_page(oap);
1010
1011                 LASSERT(list_empty(&oap->oap_rpc_item));
1012
1013                 /* only discard the pages with their index greater than
1014                  * trunc_index, and ... */
1015                 if (index < trunc_index ||
1016                     (index == trunc_index && partial)) {
1017                         /* accounting how many pages remaining in the chunk
1018                          * so that we can calculate grants correctly. */
1019                         if (index >> ppc_bits == trunc_chunk)
1020                                 ++pages_in_chunk;
1021                         continue;
1022                 }
1023
1024                 list_del_init(&oap->oap_pending_item);
1025
1026                 cl_page_get(page);
1027                 lu_ref_add(&page->cp_reference, "truncate", current);
1028
1029                 if (cl_page_own(env, io, page) == 0) {
1030                         cl_page_discard(env, io, page);
1031                         cl_page_disown(env, io, page);
1032                 } else {
1033                         LASSERT(page->cp_state == CPS_FREEING);
1034                         LASSERT(0);
1035                 }
1036
1037                 lu_ref_del(&page->cp_reference, "truncate", current);
1038                 cl_page_put(env, page);
1039
1040                 --ext->oe_nr_pages;
1041                 ++nr_pages;
1042         }
1043         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1044                       ext->oe_nr_pages == 0),
1045                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1046
1047         osc_object_lock(obj);
1048         if (ext->oe_nr_pages == 0) {
1049                 LASSERT(pages_in_chunk == 0);
1050                 grants = ext->oe_grants;
1051                 ext->oe_grants = 0;
1052         } else { /* calculate how many grants we can free */
1053                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1054                 pgoff_t last_index;
1055
1056
1057                 /* if there is no pages in this chunk, we can also free grants
1058                  * for the last chunk */
1059                 if (pages_in_chunk == 0) {
1060                         /* if this is the 1st chunk and no pages in this chunk,
1061                          * ext->oe_nr_pages must be zero, so we should be in
1062                          * the other if-clause. */
1063                         LASSERT(trunc_chunk > 0);
1064                         --trunc_chunk;
1065                         ++chunks;
1066                 }
1067
1068                 /* this is what we can free from this extent */
1069                 grants          = chunks << cli->cl_chunkbits;
1070                 ext->oe_grants -= grants;
1071                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1072                 ext->oe_end     = min(last_index, ext->oe_max_end);
1073                 LASSERT(ext->oe_end >= ext->oe_start);
1074                 LASSERT(ext->oe_grants > 0);
1075         }
1076         osc_object_unlock(obj);
1077
1078         if (grants > 0 || nr_pages > 0)
1079                 osc_free_grant(cli, nr_pages, grants, grants);
1080
1081 out:
1082         cl_io_fini(env, io);
1083         cl_env_put(env, &refcheck);
1084         RETURN(rc);
1085 }
1086
1087 /**
1088  * This function is used to make the extent prepared for transfer.
1089  * A race with flusing page - ll_writepage() has to be handled cautiously.
1090  */
1091 static int osc_extent_make_ready(const struct lu_env *env,
1092                                  struct osc_extent *ext)
1093 {
1094         struct osc_async_page *oap;
1095         struct osc_async_page *last = NULL;
1096         struct osc_object *obj = ext->oe_obj;
1097         unsigned int page_count = 0;
1098         int rc;
1099         ENTRY;
1100
1101         /* we're going to grab page lock, so object lock must not be taken. */
1102         LASSERT(sanity_check(ext) == 0);
1103         /* in locking state, any process should not touch this extent. */
1104         EASSERT(ext->oe_state == OES_LOCKING, ext);
1105         EASSERT(ext->oe_owner != NULL, ext);
1106
1107         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1108
1109         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1110                 ++page_count;
1111                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1112                         last = oap;
1113
1114                 /* checking ASYNC_READY is race safe */
1115                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1116                         continue;
1117
1118                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1119                 switch (rc) {
1120                 case 0:
1121                         spin_lock(&oap->oap_lock);
1122                         oap->oap_async_flags |= ASYNC_READY;
1123                         spin_unlock(&oap->oap_lock);
1124                         break;
1125                 case -EALREADY:
1126                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1127                         break;
1128                 default:
1129                         LASSERTF(0, "unknown return code: %d\n", rc);
1130                 }
1131         }
1132
1133         LASSERT(page_count == ext->oe_nr_pages);
1134         LASSERT(last != NULL);
1135         /* the last page is the only one we need to refresh its count by
1136          * the size of file. */
1137         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1138                 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1139                 LASSERT(last_oap_count > 0);
1140                 LASSERT(last->oap_page_off + last_oap_count <= PAGE_SIZE);
1141                 last->oap_count = last_oap_count;
1142                 spin_lock(&last->oap_lock);
1143                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1144                 spin_unlock(&last->oap_lock);
1145         }
1146
1147         /* for the rest of pages, we don't need to call osf_refresh_count()
1148          * because it's known they are not the last page */
1149         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1150                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1151                         oap->oap_count = PAGE_SIZE - oap->oap_page_off;
1152                         spin_lock(&oap->oap_lock);
1153                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1154                         spin_unlock(&oap->oap_lock);
1155                 }
1156         }
1157
1158         osc_object_lock(obj);
1159         osc_extent_state_set(ext, OES_RPC);
1160         osc_object_unlock(obj);
1161         /* get a refcount for RPC. */
1162         osc_extent_get(ext);
1163
1164         RETURN(0);
1165 }
1166
1167 /**
1168  * Quick and simple version of osc_extent_find(). This function is frequently
1169  * called to expand the extent for the same IO. To expand the extent, the
1170  * page index must be in the same or next chunk of ext->oe_end.
1171  */
1172 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index,
1173                              unsigned int *grants)
1174 {
1175         struct osc_object *obj = ext->oe_obj;
1176         struct client_obd *cli = osc_cli(obj);
1177         struct osc_extent *next;
1178         int ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1179         pgoff_t chunk = index >> ppc_bits;
1180         pgoff_t end_chunk;
1181         pgoff_t end_index;
1182         unsigned int chunksize = 1 << cli->cl_chunkbits;
1183         int rc = 0;
1184         ENTRY;
1185
1186         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1187         osc_object_lock(obj);
1188         LASSERT(sanity_check_nolock(ext) == 0);
1189         end_chunk = ext->oe_end >> ppc_bits;
1190         if (chunk > end_chunk + 1)
1191                 GOTO(out, rc = -ERANGE);
1192
1193         if (end_chunk >= chunk)
1194                 GOTO(out, rc = 0);
1195
1196         LASSERT(end_chunk + 1 == chunk);
1197
1198         /* try to expand this extent to cover @index */
1199         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1200
1201         /* don't go over the maximum extent size reported by server */
1202         if (end_index - ext->oe_start + 1 > cli->cl_max_extent_pages)
1203                 GOTO(out, rc = -ERANGE);
1204
1205         next = next_extent(ext);
1206         if (next != NULL && next->oe_start <= end_index)
1207                 /* complex mode - overlapped with the next extent,
1208                  * this case will be handled by osc_extent_find() */
1209                 GOTO(out, rc = -EAGAIN);
1210
1211         ext->oe_end = end_index;
1212         ext->oe_grants += chunksize;
1213         LASSERT(*grants >= chunksize);
1214         *grants -= chunksize;
1215         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1216                  "overlapped after expanding for %lu.\n", index);
1217         EXIT;
1218
1219 out:
1220         osc_object_unlock(obj);
1221         RETURN(rc);
1222 }
1223
1224 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
1225                                   const char *func, int line)
1226 {
1227         struct osc_extent *ext;
1228         int cnt;
1229
1230         if (!cfs_cdebug_show(level, DEBUG_SUBSYSTEM))
1231                 return;
1232
1233         CDEBUG(level, "Dump object %p extents at %s:%d, mppr: %u.\n",
1234                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1235
1236         /* osc_object_lock(obj); */
1237         cnt = 1;
1238         for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1239                 OSC_EXTENT_DUMP(level, ext, "in tree %d.\n", cnt++);
1240
1241         cnt = 1;
1242         list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1243                 OSC_EXTENT_DUMP(level, ext, "hp %d.\n", cnt++);
1244
1245         cnt = 1;
1246         list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1247                 OSC_EXTENT_DUMP(level, ext, "urgent %d.\n", cnt++);
1248
1249         cnt = 1;
1250         list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1251                 OSC_EXTENT_DUMP(level, ext, "reading %d.\n", cnt++);
1252         /* osc_object_unlock(obj); */
1253 }
1254
1255 /* ------------------ osc extent end ------------------ */
1256
1257 static inline int osc_is_ready(struct osc_object *osc)
1258 {
1259         return !list_empty(&osc->oo_ready_item) ||
1260                !list_empty(&osc->oo_hp_ready_item);
1261 }
1262
1263 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1264         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1265                (OSC), osc_is_ready(OSC),                                       \
1266                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1267                list_empty_marker(&(OSC)->oo_ready_item),                       \
1268                atomic_read(&(OSC)->oo_nr_writes),                              \
1269                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1270                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1271                atomic_read(&(OSC)->oo_nr_reads),                               \
1272                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1273                ##args)
1274
1275 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1276                           int cmd)
1277 {
1278         struct osc_page *opg  = oap2osc_page(oap);
1279         struct cl_page  *page = oap2cl_page(oap);
1280         int result;
1281
1282         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1283
1284         ENTRY;
1285         result = cl_page_make_ready(env, page, CRT_WRITE);
1286         if (result == 0)
1287                 opg->ops_submit_time = cfs_time_current();
1288         RETURN(result);
1289 }
1290
1291 static int osc_refresh_count(const struct lu_env *env,
1292                              struct osc_async_page *oap, int cmd)
1293 {
1294         struct osc_page  *opg = oap2osc_page(oap);
1295         pgoff_t index = osc_index(oap2osc(oap));
1296         struct cl_object *obj;
1297         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
1298
1299         int result;
1300         loff_t kms;
1301
1302         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1303         LASSERT(!(cmd & OBD_BRW_READ));
1304         LASSERT(opg != NULL);
1305         obj = opg->ops_cl.cpl_obj;
1306
1307         cl_object_attr_lock(obj);
1308         result = cl_object_attr_get(env, obj, attr);
1309         cl_object_attr_unlock(obj);
1310         if (result < 0)
1311                 return result;
1312         kms = attr->cat_kms;
1313         if (cl_offset(obj, index) >= kms)
1314                 /* catch race with truncate */
1315                 return 0;
1316         else if (cl_offset(obj, index + 1) > kms)
1317                 /* catch sub-page write at end of file */
1318                 return kms % PAGE_SIZE;
1319         else
1320                 return PAGE_SIZE;
1321 }
1322
1323 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1324                           int cmd, int rc)
1325 {
1326         struct osc_page   *opg  = oap2osc_page(oap);
1327         struct cl_page    *page = oap2cl_page(oap);
1328         enum cl_req_type   crt;
1329         int srvlock;
1330
1331         ENTRY;
1332
1333         cmd &= ~OBD_BRW_NOQUOTA;
1334         LASSERTF(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ),
1335                  "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1336         LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1337                 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1338         LASSERT(opg->ops_transfer_pinned);
1339
1340         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1341         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1342         opg->ops_transfer_pinned = 0;
1343
1344         opg->ops_submit_time = 0;
1345         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1346
1347         /* statistic */
1348         if (rc == 0 && srvlock) {
1349                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1350                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1351                 size_t bytes = oap->oap_count;
1352
1353                 if (crt == CRT_READ)
1354                         stats->os_lockless_reads += bytes;
1355                 else
1356                         stats->os_lockless_writes += bytes;
1357         }
1358
1359         /*
1360          * This has to be the last operation with the page, as locks are
1361          * released in cl_page_completion() and nothing except for the
1362          * reference counter protects page from concurrent reclaim.
1363          */
1364         lu_ref_del(&page->cp_reference, "transfer", page);
1365
1366         cl_page_completion(env, page, crt, rc);
1367         cl_page_put(env, page);
1368
1369         RETURN(0);
1370 }
1371
1372 #define OSC_DUMP_GRANT(lvl, cli, fmt, args...) do {                     \
1373         struct client_obd *__tmp = (cli);                               \
1374         CDEBUG(lvl, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu "  \
1375                "dropped: %ld avail: %ld, dirty_grant: %ld, "            \
1376                "reserved: %ld, flight: %d } lru {in list: %ld, "        \
1377                "left: %ld, waiters: %d }" fmt "\n",                     \
1378                cli_name(__tmp),                                         \
1379                __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages,        \
1380                atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1381                __tmp->cl_lost_grant, __tmp->cl_avail_grant,             \
1382                __tmp->cl_dirty_grant,                                   \
1383                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,         \
1384                atomic_long_read(&__tmp->cl_lru_in_list),                \
1385                atomic_long_read(&__tmp->cl_lru_busy),                   \
1386                atomic_read(&__tmp->cl_lru_shrinkers), ##args);          \
1387 } while (0)
1388
1389 /* caller must hold loi_list_lock */
1390 static void osc_consume_write_grant(struct client_obd *cli,
1391                                     struct brw_page *pga)
1392 {
1393         assert_spin_locked(&cli->cl_loi_list_lock);
1394         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1395         atomic_long_inc(&obd_dirty_pages);
1396         cli->cl_dirty_pages++;
1397         pga->flag |= OBD_BRW_FROM_GRANT;
1398         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1399                PAGE_SIZE, pga, pga->pg);
1400         osc_update_next_shrink(cli);
1401 }
1402
1403 /* the companion to osc_consume_write_grant, called when a brw has completed.
1404  * must be called with the loi lock held. */
1405 static void osc_release_write_grant(struct client_obd *cli,
1406                                     struct brw_page *pga)
1407 {
1408         ENTRY;
1409
1410         assert_spin_locked(&cli->cl_loi_list_lock);
1411         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1412                 EXIT;
1413                 return;
1414         }
1415
1416         pga->flag &= ~OBD_BRW_FROM_GRANT;
1417         atomic_long_dec(&obd_dirty_pages);
1418         cli->cl_dirty_pages--;
1419         if (pga->flag & OBD_BRW_NOCACHE) {
1420                 pga->flag &= ~OBD_BRW_NOCACHE;
1421                 atomic_long_dec(&obd_dirty_transit_pages);
1422                 cli->cl_dirty_transit--;
1423         }
1424         EXIT;
1425 }
1426
1427 /**
1428  * To avoid sleeping with object lock held, it's good for us allocate enough
1429  * grants before entering into critical section.
1430  *
1431  * client_obd_list_lock held by caller
1432  */
1433 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1434 {
1435         int rc = -EDQUOT;
1436
1437         if (cli->cl_avail_grant >= bytes) {
1438                 cli->cl_avail_grant    -= bytes;
1439                 cli->cl_reserved_grant += bytes;
1440                 rc = 0;
1441         }
1442         return rc;
1443 }
1444
1445 static void __osc_unreserve_grant(struct client_obd *cli,
1446                                   unsigned int reserved, unsigned int unused)
1447 {
1448         /* it's quite normal for us to get more grant than reserved.
1449          * Thinking about a case that two extents merged by adding a new
1450          * chunk, we can save one extent tax. If extent tax is greater than
1451          * one chunk, we can save more grant by adding a new chunk */
1452         cli->cl_reserved_grant -= reserved;
1453         if (unused > reserved) {
1454                 cli->cl_avail_grant += reserved;
1455                 cli->cl_lost_grant  += unused - reserved;
1456                 cli->cl_dirty_grant -= unused - reserved;
1457         } else {
1458                 cli->cl_avail_grant += unused;
1459                 cli->cl_dirty_grant += reserved - unused;
1460         }
1461 }
1462
1463 static void osc_unreserve_grant(struct client_obd *cli,
1464                                 unsigned int reserved, unsigned int unused)
1465 {
1466         spin_lock(&cli->cl_loi_list_lock);
1467         __osc_unreserve_grant(cli, reserved, unused);
1468         if (unused > 0)
1469                 osc_wake_cache_waiters(cli);
1470         spin_unlock(&cli->cl_loi_list_lock);
1471 }
1472
1473 /**
1474  * Free grant after IO is finished or canceled.
1475  *
1476  * @lost_grant is used to remember how many grants we have allocated but not
1477  * used, we should return these grants to OST. There're two cases where grants
1478  * can be lost:
1479  * 1. truncate;
1480  * 2. blocksize at OST is less than PAGE_SIZE and a partial page was
1481  *    written. In this case OST may use less chunks to serve this partial
1482  *    write. OSTs don't actually know the page size on the client side. so
1483  *    clients have to calculate lost grant by the blocksize on the OST.
1484  *    See filter_grant_check() for details.
1485  */
1486 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1487                            unsigned int lost_grant, unsigned int dirty_grant)
1488 {
1489         unsigned long grant;
1490
1491         grant = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
1492
1493         spin_lock(&cli->cl_loi_list_lock);
1494         atomic_long_sub(nr_pages, &obd_dirty_pages);
1495         cli->cl_dirty_pages -= nr_pages;
1496         cli->cl_lost_grant += lost_grant;
1497         cli->cl_dirty_grant -= dirty_grant;
1498         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1499                 /* borrow some grant from truncate to avoid the case that
1500                  * truncate uses up all avail grant */
1501                 cli->cl_lost_grant -= grant;
1502                 cli->cl_avail_grant += grant;
1503         }
1504         osc_wake_cache_waiters(cli);
1505         spin_unlock(&cli->cl_loi_list_lock);
1506         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu/%lu\n",
1507                lost_grant, cli->cl_lost_grant,
1508                cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_SHIFT,
1509                cli->cl_dirty_grant);
1510 }
1511
1512 /**
1513  * The companion to osc_enter_cache(), called when @oap is no longer part of
1514  * the dirty accounting due to error.
1515  */
1516 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1517 {
1518         spin_lock(&cli->cl_loi_list_lock);
1519         osc_release_write_grant(cli, &oap->oap_brw_page);
1520         spin_unlock(&cli->cl_loi_list_lock);
1521 }
1522
1523 /**
1524  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1525  * is available.
1526  */
1527 static int osc_enter_cache_try(struct client_obd *cli,
1528                                struct osc_async_page *oap,
1529                                int bytes, int transient)
1530 {
1531         int rc;
1532
1533         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1534
1535         rc = osc_reserve_grant(cli, bytes);
1536         if (rc < 0)
1537                 return 0;
1538
1539         if (cli->cl_dirty_pages < cli->cl_dirty_max_pages &&
1540             1 + atomic_long_read(&obd_dirty_pages) <= obd_max_dirty_pages) {
1541                 osc_consume_write_grant(cli, &oap->oap_brw_page);
1542                 if (transient) {
1543                         cli->cl_dirty_transit++;
1544                         atomic_long_inc(&obd_dirty_transit_pages);
1545                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
1546                 }
1547                 rc = 1;
1548         } else {
1549                 __osc_unreserve_grant(cli, bytes, bytes);
1550                 rc = 0;
1551         }
1552         return rc;
1553 }
1554
1555 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1556 {
1557         int rc;
1558         spin_lock(&cli->cl_loi_list_lock);
1559         rc = list_empty(&ocw->ocw_entry);
1560         spin_unlock(&cli->cl_loi_list_lock);
1561         return rc;
1562 }
1563
1564 /**
1565  * The main entry to reserve dirty page accounting. Usually the grant reserved
1566  * in this function will be freed in bulk in osc_free_grant() unless it fails
1567  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1568  *
1569  * The process will be put into sleep if it's already run out of grant.
1570  */
1571 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1572                            struct osc_async_page *oap, int bytes)
1573 {
1574         struct osc_object       *osc = oap->oap_obj;
1575         struct lov_oinfo        *loi = osc->oo_oinfo;
1576         struct osc_cache_waiter  ocw;
1577         struct l_wait_info       lwi;
1578         int                      rc = -EDQUOT;
1579         ENTRY;
1580
1581         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(AT_OFF ? obd_timeout : at_max),
1582                                NULL, LWI_ON_SIGNAL_NOOP, NULL);
1583
1584         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1585
1586         spin_lock(&cli->cl_loi_list_lock);
1587
1588         /* force the caller to try sync io.  this can jump the list
1589          * of queued writes and create a discontiguous rpc stream */
1590         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1591             cli->cl_dirty_max_pages == 0 ||
1592             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) {
1593                 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1594                 GOTO(out, rc = -EDQUOT);
1595         }
1596
1597         /* Hopefully normal case - cache space and write credits available */
1598         if (osc_enter_cache_try(cli, oap, bytes, 0)) {
1599                 OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1600                 GOTO(out, rc = 0);
1601         }
1602
1603         /* We can get here for two reasons: too many dirty pages in cache, or
1604          * run out of grants. In both cases we should write dirty pages out.
1605          * Adding a cache waiter will trigger urgent write-out no matter what
1606          * RPC size will be.
1607          * The exiting condition is no avail grants and no dirty pages caching,
1608          * that really means there is no space on the OST. */
1609         init_waitqueue_head(&ocw.ocw_waitq);
1610         ocw.ocw_oap   = oap;
1611         ocw.ocw_grant = bytes;
1612         while (cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0) {
1613                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1614                 ocw.ocw_rc = 0;
1615                 spin_unlock(&cli->cl_loi_list_lock);
1616
1617                 osc_io_unplug_async(env, cli, NULL);
1618
1619                 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1620                        cli_name(cli), &ocw, oap);
1621
1622                 rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1623
1624                 spin_lock(&cli->cl_loi_list_lock);
1625
1626                 if (rc < 0) {
1627                         /* l_wait_event is interrupted by signal or timed out */
1628                         list_del_init(&ocw.ocw_entry);
1629                         break;
1630                 }
1631                 LASSERT(list_empty(&ocw.ocw_entry));
1632                 rc = ocw.ocw_rc;
1633
1634                 if (rc != -EDQUOT)
1635                         break;
1636                 if (osc_enter_cache_try(cli, oap, bytes, 0)) {
1637                         rc = 0;
1638                         break;
1639                 }
1640         }
1641
1642         switch (rc) {
1643         case 0:
1644                 OSC_DUMP_GRANT(D_CACHE, cli, "finally got grant space\n");
1645                 break;
1646         case -ETIMEDOUT:
1647                 OSC_DUMP_GRANT(D_CACHE, cli,
1648                                "timeout, fall back to sync i/o\n");
1649                 osc_extent_tree_dump(D_CACHE, osc);
1650                 /* fall back to synchronous I/O */
1651                 rc = -EDQUOT;
1652                 break;
1653         case -EINTR:
1654                 /* Ensures restartability - LU-3581 */
1655                 OSC_DUMP_GRANT(D_CACHE, cli, "interrupted\n");
1656                 rc = -ERESTARTSYS;
1657                 break;
1658         case -EDQUOT:
1659                 OSC_DUMP_GRANT(D_CACHE, cli,
1660                                "no grant space, fall back to sync i/o\n");
1661                 break;
1662         default:
1663                 CDEBUG(D_CACHE, "%s: event for cache space @ %p never arrived "
1664                        "due to %d, fall back to sync i/o\n",
1665                        cli_name(cli), &ocw, rc);
1666                 break;
1667         }
1668         EXIT;
1669 out:
1670         spin_unlock(&cli->cl_loi_list_lock);
1671         RETURN(rc);
1672 }
1673
1674 /* caller must hold loi_list_lock */
1675 void osc_wake_cache_waiters(struct client_obd *cli)
1676 {
1677         struct list_head *l, *tmp;
1678         struct osc_cache_waiter *ocw;
1679
1680         ENTRY;
1681         list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1682                 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
1683                 list_del_init(&ocw->ocw_entry);
1684
1685                 ocw->ocw_rc = -EDQUOT;
1686                 /* we can't dirty more */
1687                 if ((cli->cl_dirty_pages  >= cli->cl_dirty_max_pages) ||
1688                     (1 + atomic_long_read(&obd_dirty_pages) >
1689                      obd_max_dirty_pages)) {
1690                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
1691                                "osc max %ld, sys max %ld\n",
1692                                cli->cl_dirty_pages, cli->cl_dirty_max_pages,
1693                                obd_max_dirty_pages);
1694                         goto wakeup;
1695                 }
1696
1697                 if (osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
1698                         ocw->ocw_rc = 0;
1699 wakeup:
1700                 CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld, %d\n",
1701                        ocw, ocw->ocw_oap, cli->cl_avail_grant, ocw->ocw_rc);
1702
1703                 wake_up(&ocw->ocw_waitq);
1704         }
1705
1706         EXIT;
1707 }
1708
1709 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1710 {
1711         int hprpc = !!list_empty(&osc->oo_hp_exts);
1712         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1713 }
1714
1715 /* This maintains the lists of pending pages to read/write for a given object
1716  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1717  * to quickly find objects that are ready to send an RPC. */
1718 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1719                          int cmd)
1720 {
1721         int invalid_import = 0;
1722         ENTRY;
1723
1724         /* if we have an invalid import we want to drain the queued pages
1725          * by forcing them through rpcs that immediately fail and complete
1726          * the pages.  recovery relies on this to empty the queued pages
1727          * before canceling the locks and evicting down the llite pages */
1728         if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1729                 invalid_import = 1;
1730
1731         if (cmd & OBD_BRW_WRITE) {
1732                 if (atomic_read(&osc->oo_nr_writes) == 0)
1733                         RETURN(0);
1734                 if (invalid_import) {
1735                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1736                         RETURN(1);
1737                 }
1738                 if (!list_empty(&osc->oo_hp_exts)) {
1739                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1740                         RETURN(1);
1741                 }
1742                 if (!list_empty(&osc->oo_urgent_exts)) {
1743                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1744                         RETURN(1);
1745                 }
1746                 /* trigger a write rpc stream as long as there are dirtiers
1747                  * waiting for space.  as they're waiting, they're not going to
1748                  * create more pages to coalesce with what's waiting.. */
1749                 if (!list_empty(&cli->cl_cache_waiters)) {
1750                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1751                         RETURN(1);
1752                 }
1753                 if (!list_empty(&osc->oo_full_exts)) {
1754                         CDEBUG(D_CACHE, "full extent ready, make an RPC\n");
1755                         RETURN(1);
1756                 }
1757         } else {
1758                 if (atomic_read(&osc->oo_nr_reads) == 0)
1759                         RETURN(0);
1760                 if (invalid_import) {
1761                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1762                         RETURN(1);
1763                 }
1764                 /* all read are urgent. */
1765                 if (!list_empty(&osc->oo_reading_exts))
1766                         RETURN(1);
1767         }
1768
1769         RETURN(0);
1770 }
1771
1772 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1773 {
1774         struct client_obd *cli = osc_cli(obj);
1775         if (cmd & OBD_BRW_WRITE) {
1776                 atomic_add(delta, &obj->oo_nr_writes);
1777                 atomic_add(delta, &cli->cl_pending_w_pages);
1778                 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1779         } else {
1780                 atomic_add(delta, &obj->oo_nr_reads);
1781                 atomic_add(delta, &cli->cl_pending_r_pages);
1782                 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1783         }
1784         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1785 }
1786
1787 static int osc_makes_hprpc(struct osc_object *obj)
1788 {
1789         return !list_empty(&obj->oo_hp_exts);
1790 }
1791
1792 static void on_list(struct list_head *item, struct list_head *list,
1793                     int should_be_on)
1794 {
1795         if (list_empty(item) && should_be_on)
1796                 list_add_tail(item, list);
1797         else if (!list_empty(item) && !should_be_on)
1798                 list_del_init(item);
1799 }
1800
1801 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1802  * can find pages to build into rpcs quickly */
1803 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1804 {
1805         if (osc_makes_hprpc(osc)) {
1806                 /* HP rpc */
1807                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1808                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1809         } else {
1810                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1811                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1812                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1813                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1814         }
1815
1816         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1817                 atomic_read(&osc->oo_nr_writes) > 0);
1818
1819         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1820                 atomic_read(&osc->oo_nr_reads) > 0);
1821
1822         return osc_is_ready(osc);
1823 }
1824
1825 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1826 {
1827         int is_ready;
1828
1829         spin_lock(&cli->cl_loi_list_lock);
1830         is_ready = __osc_list_maint(cli, osc);
1831         spin_unlock(&cli->cl_loi_list_lock);
1832
1833         return is_ready;
1834 }
1835
1836 /* this is trying to propogate async writeback errors back up to the
1837  * application.  As an async write fails we record the error code for later if
1838  * the app does an fsync.  As long as errors persist we force future rpcs to be
1839  * sync so that the app can get a sync error and break the cycle of queueing
1840  * pages for which writeback will fail. */
1841 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1842                            int rc)
1843 {
1844         if (rc) {
1845                 if (!ar->ar_rc)
1846                         ar->ar_rc = rc;
1847
1848                 ar->ar_force_sync = 1;
1849                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1850                 return;
1851
1852         }
1853
1854         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1855                 ar->ar_force_sync = 0;
1856 }
1857
1858 /* this must be called holding the loi list lock to give coverage to exit_cache,
1859  * async_flag maintenance, and oap_request */
1860 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1861                               struct osc_async_page *oap, int sent, int rc)
1862 {
1863         struct osc_object *osc = oap->oap_obj;
1864         struct lov_oinfo  *loi = osc->oo_oinfo;
1865         __u64 xid = 0;
1866
1867         ENTRY;
1868         if (oap->oap_request != NULL) {
1869                 xid = ptlrpc_req_xid(oap->oap_request);
1870                 ptlrpc_req_finished(oap->oap_request);
1871                 oap->oap_request = NULL;
1872         }
1873
1874         /* As the transfer for this page is being done, clear the flags */
1875         spin_lock(&oap->oap_lock);
1876         oap->oap_async_flags = 0;
1877         spin_unlock(&oap->oap_lock);
1878         oap->oap_interrupted = 0;
1879
1880         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1881                 spin_lock(&cli->cl_loi_list_lock);
1882                 osc_process_ar(&cli->cl_ar, xid, rc);
1883                 osc_process_ar(&loi->loi_ar, xid, rc);
1884                 spin_unlock(&cli->cl_loi_list_lock);
1885         }
1886
1887         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1888         if (rc)
1889                 CERROR("completion on oap %p obj %p returns %d.\n",
1890                        oap, osc, rc);
1891
1892         EXIT;
1893 }
1894
1895 struct extent_rpc_data {
1896         struct list_head        *erd_rpc_list;
1897         unsigned int            erd_page_count;
1898         unsigned int            erd_max_pages;
1899         unsigned int            erd_max_chunks;
1900         unsigned int            erd_max_extents;
1901 };
1902
1903 static inline unsigned osc_extent_chunks(const struct osc_extent *ext)
1904 {
1905         struct client_obd *cli = osc_cli(ext->oe_obj);
1906         unsigned ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1907
1908         return (ext->oe_end >> ppc_bits) - (ext->oe_start >> ppc_bits) + 1;
1909 }
1910
1911 /**
1912  * Try to add extent to one RPC. We need to think about the following things:
1913  * - # of pages must not be over max_pages_per_rpc
1914  * - extent must be compatible with previous ones
1915  */
1916 static int try_to_add_extent_for_io(struct client_obd *cli,
1917                                     struct osc_extent *ext,
1918                                     struct extent_rpc_data *data)
1919 {
1920         struct osc_extent *tmp;
1921         unsigned int chunk_count;
1922         struct osc_async_page *oap = list_first_entry(&ext->oe_pages,
1923                                                       struct osc_async_page,
1924                                                       oap_pending_item);
1925         ENTRY;
1926
1927         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1928                 ext);
1929         OSC_EXTENT_DUMP(D_CACHE, ext, "trying to add this extent\n");
1930
1931         if (data->erd_max_extents == 0)
1932                 RETURN(0);
1933
1934         chunk_count = osc_extent_chunks(ext);
1935         EASSERTF(data->erd_page_count != 0 ||
1936                  chunk_count <= data->erd_max_chunks, ext,
1937                  "The first extent to be fit in a RPC contains %u chunks, "
1938                  "which is over the limit %u.\n", chunk_count,
1939                  data->erd_max_chunks);
1940         if (chunk_count > data->erd_max_chunks)
1941                 RETURN(0);
1942
1943         data->erd_max_pages = max(ext->oe_mppr, data->erd_max_pages);
1944         EASSERTF(data->erd_page_count != 0 ||
1945                 ext->oe_nr_pages <= data->erd_max_pages, ext,
1946                 "The first extent to be fit in a RPC contains %u pages, "
1947                 "which is over the limit %u.\n", ext->oe_nr_pages,
1948                 data->erd_max_pages);
1949         if (data->erd_page_count + ext->oe_nr_pages > data->erd_max_pages)
1950                 RETURN(0);
1951
1952         list_for_each_entry(tmp, data->erd_rpc_list, oe_link) {
1953                 struct osc_async_page *oap2;
1954                 oap2 = list_first_entry(&tmp->oe_pages, struct osc_async_page,
1955                                         oap_pending_item);
1956                 EASSERT(tmp->oe_owner == current, tmp);
1957 #if 0
1958                 if (overlapped(tmp, ext)) {
1959                         OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
1960                         EASSERT(0, ext);
1961                 }
1962 #endif
1963                 if (oap2cl_page(oap)->cp_type != oap2cl_page(oap2)->cp_type) {
1964                         CDEBUG(D_CACHE, "Do not permit different types of IO "
1965                                "in one RPC\n");
1966                         RETURN(0);
1967                 }
1968
1969                 if (tmp->oe_srvlock != ext->oe_srvlock ||
1970                     !tmp->oe_grants != !ext->oe_grants ||
1971                     tmp->oe_no_merge || ext->oe_no_merge)
1972                         RETURN(0);
1973
1974                 /* remove break for strict check */
1975                 break;
1976         }
1977
1978         data->erd_max_extents--;
1979         data->erd_max_chunks -= chunk_count;
1980         data->erd_page_count += ext->oe_nr_pages;
1981         list_move_tail(&ext->oe_link, data->erd_rpc_list);
1982         ext->oe_owner = current;
1983         RETURN(1);
1984 }
1985
1986 static inline unsigned osc_max_write_chunks(const struct client_obd *cli)
1987 {
1988         /*
1989          * LU-8135:
1990          *
1991          * The maximum size of a single transaction is about 64MB in ZFS.
1992          * #define DMU_MAX_ACCESS (64 * 1024 * 1024)
1993          *
1994          * Since ZFS is a copy-on-write file system, a single dirty page in
1995          * a chunk will result in the rewrite of the whole chunk, therefore
1996          * an RPC shouldn't be allowed to contain too many chunks otherwise
1997          * it will make transaction size much bigger than 64MB, especially
1998          * with big block size for ZFS.
1999          *
2000          * This piece of code is to make sure that OSC won't send write RPCs
2001          * with too many chunks. The maximum chunk size that an RPC can cover
2002          * is set to PTLRPC_MAX_BRW_SIZE, which is defined to 16MB. Ideally
2003          * OST should tell the client what the biggest transaction size is,
2004          * but it's good enough for now.
2005          *
2006          * This limitation doesn't apply to ldiskfs, which allows as many
2007          * chunks in one RPC as we want. However, it won't have any benefits
2008          * to have too many discontiguous pages in one RPC.
2009          *
2010          * An osc_extent won't cover over a RPC size, so the chunks in an
2011          * osc_extent won't bigger than PTLRPC_MAX_BRW_SIZE >> chunkbits.
2012          */
2013         return PTLRPC_MAX_BRW_SIZE >> cli->cl_chunkbits;
2014 }
2015
2016 /**
2017  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
2018  * get_write_extent() takes all appropriate extents in atomic.
2019  *
2020  * The following policy is used to collect extents for IO:
2021  * 1. Add as many HP extents as possible;
2022  * 2. Add the first urgent extent in urgent extent list and take it out of
2023  *    urgent list;
2024  * 3. Add subsequent extents of this urgent extent;
2025  * 4. If urgent list is not empty, goto 2;
2026  * 5. Traverse the extent tree from the 1st extent;
2027  * 6. Above steps exit if there is no space in this RPC.
2028  */
2029 static unsigned int get_write_extents(struct osc_object *obj,
2030                                       struct list_head *rpclist)
2031 {
2032         struct client_obd *cli = osc_cli(obj);
2033         struct osc_extent *ext;
2034         struct extent_rpc_data data = {
2035                 .erd_rpc_list   = rpclist,
2036                 .erd_page_count = 0,
2037                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
2038                 .erd_max_chunks = osc_max_write_chunks(cli),
2039                 .erd_max_extents = 256,
2040         };
2041
2042         LASSERT(osc_object_is_locked(obj));
2043         while (!list_empty(&obj->oo_hp_exts)) {
2044                 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
2045                                  oe_link);
2046                 LASSERT(ext->oe_state == OES_CACHE);
2047                 if (!try_to_add_extent_for_io(cli, ext, &data))
2048                         return data.erd_page_count;
2049                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2050         }
2051         if (data.erd_page_count == data.erd_max_pages)
2052                 return data.erd_page_count;
2053
2054         while (!list_empty(&obj->oo_urgent_exts)) {
2055                 ext = list_entry(obj->oo_urgent_exts.next,
2056                                  struct osc_extent, oe_link);
2057                 if (!try_to_add_extent_for_io(cli, ext, &data))
2058                         return data.erd_page_count;
2059         }
2060         if (data.erd_page_count == data.erd_max_pages)
2061                 return data.erd_page_count;
2062
2063         /* One key difference between full extents and other extents: full
2064          * extents can usually only be added if the rpclist was empty, so if we
2065          * can't add one, we continue on to trying to add normal extents.  This
2066          * is so we don't miss adding extra extents to an RPC containing high
2067          * priority or urgent extents. */
2068         while (!list_empty(&obj->oo_full_exts)) {
2069                 ext = list_entry(obj->oo_full_exts.next,
2070                                  struct osc_extent, oe_link);
2071                 if (!try_to_add_extent_for_io(cli, ext, &data))
2072                         break;
2073         }
2074         if (data.erd_page_count == data.erd_max_pages)
2075                 return data.erd_page_count;
2076
2077         ext = first_extent(obj);
2078         while (ext != NULL) {
2079                 if ((ext->oe_state != OES_CACHE) ||
2080                     /* this extent may be already in current rpclist */
2081                     (!list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
2082                         ext = next_extent(ext);
2083                         continue;
2084                 }
2085
2086                 if (!try_to_add_extent_for_io(cli, ext, &data))
2087                         return data.erd_page_count;
2088
2089                 ext = next_extent(ext);
2090         }
2091         return data.erd_page_count;
2092 }
2093
2094 static int
2095 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
2096                    struct osc_object *osc)
2097 __must_hold(osc)
2098 {
2099         struct list_head   rpclist = LIST_HEAD_INIT(rpclist);
2100         struct osc_extent *ext;
2101         struct osc_extent *tmp;
2102         struct osc_extent *first = NULL;
2103         unsigned int page_count = 0;
2104         int srvlock = 0;
2105         int rc = 0;
2106         ENTRY;
2107
2108         LASSERT(osc_object_is_locked(osc));
2109
2110         page_count = get_write_extents(osc, &rpclist);
2111         LASSERT(equi(page_count == 0, list_empty(&rpclist)));
2112
2113         if (list_empty(&rpclist))
2114                 RETURN(0);
2115
2116         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2117
2118         list_for_each_entry(ext, &rpclist, oe_link) {
2119                 LASSERT(ext->oe_state == OES_CACHE ||
2120                         ext->oe_state == OES_LOCK_DONE);
2121                 if (ext->oe_state == OES_CACHE)
2122                         osc_extent_state_set(ext, OES_LOCKING);
2123                 else
2124                         osc_extent_state_set(ext, OES_RPC);
2125         }
2126
2127         /* we're going to grab page lock, so release object lock because
2128          * lock order is page lock -> object lock. */
2129         osc_object_unlock(osc);
2130
2131         list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2132                 if (ext->oe_state == OES_LOCKING) {
2133                         rc = osc_extent_make_ready(env, ext);
2134                         if (unlikely(rc < 0)) {
2135                                 list_del_init(&ext->oe_link);
2136                                 osc_extent_finish(env, ext, 0, rc);
2137                                 continue;
2138                         }
2139                 }
2140                 if (first == NULL) {
2141                         first = ext;
2142                         srvlock = ext->oe_srvlock;
2143                 } else {
2144                         LASSERT(srvlock == ext->oe_srvlock);
2145                 }
2146         }
2147
2148         if (!list_empty(&rpclist)) {
2149                 LASSERT(page_count > 0);
2150                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2151                 LASSERT(list_empty(&rpclist));
2152         }
2153
2154         osc_object_lock(osc);
2155         RETURN(rc);
2156 }
2157
2158 /**
2159  * prepare pages for ASYNC io and put pages in send queue.
2160  *
2161  * \param cmd OBD_BRW_* macroses
2162  * \param lop pending pages
2163  *
2164  * \return zero if no page added to send queue.
2165  * \return 1 if pages successfully added to send queue.
2166  * \return negative on errors.
2167  */
2168 static int
2169 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2170                   struct osc_object *osc)
2171 __must_hold(osc)
2172 {
2173         struct osc_extent *ext;
2174         struct osc_extent *next;
2175         struct list_head rpclist = LIST_HEAD_INIT(rpclist);
2176         struct extent_rpc_data data = {
2177                 .erd_rpc_list   = &rpclist,
2178                 .erd_page_count = 0,
2179                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
2180                 .erd_max_chunks = UINT_MAX,
2181                 .erd_max_extents = UINT_MAX,
2182         };
2183         int rc = 0;
2184         ENTRY;
2185
2186         LASSERT(osc_object_is_locked(osc));
2187         list_for_each_entry_safe(ext, next, &osc->oo_reading_exts, oe_link) {
2188                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2189                 if (!try_to_add_extent_for_io(cli, ext, &data))
2190                         break;
2191                 osc_extent_state_set(ext, OES_RPC);
2192                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2193         }
2194         LASSERT(data.erd_page_count <= data.erd_max_pages);
2195
2196         osc_update_pending(osc, OBD_BRW_READ, -data.erd_page_count);
2197
2198         if (!list_empty(&rpclist)) {
2199                 osc_object_unlock(osc);
2200
2201                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2202                 LASSERT(list_empty(&rpclist));
2203
2204                 osc_object_lock(osc);
2205         }
2206         RETURN(rc);
2207 }
2208
2209 #define list_to_obj(list, item) ({                                            \
2210         struct list_head *__tmp = (list)->next;                               \
2211         list_del_init(__tmp);                                         \
2212         list_entry(__tmp, struct osc_object, oo_##item);                      \
2213 })
2214
2215 /* This is called by osc_check_rpcs() to find which objects have pages that
2216  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2217 static struct osc_object *osc_next_obj(struct client_obd *cli)
2218 {
2219         ENTRY;
2220
2221         /* First return objects that have blocked locks so that they
2222          * will be flushed quickly and other clients can get the lock,
2223          * then objects which have pages ready to be stuffed into RPCs */
2224         if (!list_empty(&cli->cl_loi_hp_ready_list))
2225                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2226         if (!list_empty(&cli->cl_loi_ready_list))
2227                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2228
2229         /* then if we have cache waiters, return all objects with queued
2230          * writes.  This is especially important when many small files
2231          * have filled up the cache and not been fired into rpcs because
2232          * they don't pass the nr_pending/object threshhold */
2233         if (!list_empty(&cli->cl_cache_waiters) &&
2234             !list_empty(&cli->cl_loi_write_list))
2235                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2236
2237         /* then return all queued objects when we have an invalid import
2238          * so that they get flushed */
2239         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2240                 if (!list_empty(&cli->cl_loi_write_list))
2241                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2242                                            write_item));
2243                 if (!list_empty(&cli->cl_loi_read_list))
2244                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2245                                            read_item));
2246         }
2247         RETURN(NULL);
2248 }
2249
2250 /* called with the loi list lock held */
2251 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2252 __must_hold(&cli->cl_loi_list_lock)
2253 {
2254         struct osc_object *osc;
2255         int rc = 0;
2256         ENTRY;
2257
2258         while ((osc = osc_next_obj(cli)) != NULL) {
2259                 struct cl_object *obj = osc2cl(osc);
2260                 struct lu_ref_link link;
2261
2262                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2263
2264                 if (osc_max_rpc_in_flight(cli, osc)) {
2265                         __osc_list_maint(cli, osc);
2266                         break;
2267                 }
2268
2269                 cl_object_get(obj);
2270                 spin_unlock(&cli->cl_loi_list_lock);
2271                 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2272
2273                 /* attempt some read/write balancing by alternating between
2274                  * reads and writes in an object.  The makes_rpc checks here
2275                  * would be redundant if we were getting read/write work items
2276                  * instead of objects.  we don't want send_oap_rpc to drain a
2277                  * partial read pending queue when we're given this object to
2278                  * do io on writes while there are cache waiters */
2279                 osc_object_lock(osc);
2280                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2281                         rc = osc_send_write_rpc(env, cli, osc);
2282                         if (rc < 0) {
2283                                 CERROR("Write request failed with %d\n", rc);
2284
2285                                 /* osc_send_write_rpc failed, mostly because of
2286                                  * memory pressure.
2287                                  *
2288                                  * It can't break here, because if:
2289                                  *  - a page was submitted by osc_io_submit, so
2290                                  *    page locked;
2291                                  *  - no request in flight
2292                                  *  - no subsequent request
2293                                  * The system will be in live-lock state,
2294                                  * because there is no chance to call
2295                                  * osc_io_unplug() and osc_check_rpcs() any
2296                                  * more. pdflush can't help in this case,
2297                                  * because it might be blocked at grabbing
2298                                  * the page lock as we mentioned.
2299                                  *
2300                                  * Anyway, continue to drain pages. */
2301                                 /* break; */
2302                         }
2303                 }
2304                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2305                         rc = osc_send_read_rpc(env, cli, osc);
2306                         if (rc < 0)
2307                                 CERROR("Read request failed with %d\n", rc);
2308                 }
2309                 osc_object_unlock(osc);
2310
2311                 osc_list_maint(cli, osc);
2312                 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2313                 cl_object_put(env, obj);
2314
2315                 spin_lock(&cli->cl_loi_list_lock);
2316         }
2317 }
2318
2319 static int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2320                           struct osc_object *osc, int async)
2321 {
2322         int rc = 0;
2323
2324         if (osc != NULL && osc_list_maint(cli, osc) == 0)
2325                 return 0;
2326
2327         if (!async) {
2328                 spin_lock(&cli->cl_loi_list_lock);
2329                 osc_check_rpcs(env, cli);
2330                 spin_unlock(&cli->cl_loi_list_lock);
2331         } else {
2332                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2333                 LASSERT(cli->cl_writeback_work != NULL);
2334                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2335         }
2336         return rc;
2337 }
2338
2339 static int osc_io_unplug_async(const struct lu_env *env,
2340                                 struct client_obd *cli, struct osc_object *osc)
2341 {
2342         return osc_io_unplug0(env, cli, osc, 1);
2343 }
2344
2345 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
2346                    struct osc_object *osc)
2347 {
2348         (void)osc_io_unplug0(env, cli, osc, 0);
2349 }
2350
2351 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2352                         struct page *page, loff_t offset)
2353 {
2354         struct obd_export     *exp = osc_export(osc);
2355         struct osc_async_page *oap = &ops->ops_oap;
2356         ENTRY;
2357
2358         if (!page)
2359                 return cfs_size_round(sizeof(*oap));
2360
2361         oap->oap_magic = OAP_MAGIC;
2362         oap->oap_cli = &exp->exp_obd->u.cli;
2363         oap->oap_obj = osc;
2364
2365         oap->oap_page = page;
2366         oap->oap_obj_off = offset;
2367         LASSERT(!(offset & ~PAGE_MASK));
2368
2369         if (cfs_capable(CFS_CAP_SYS_RESOURCE))
2370                 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2371
2372         INIT_LIST_HEAD(&oap->oap_pending_item);
2373         INIT_LIST_HEAD(&oap->oap_rpc_item);
2374
2375         spin_lock_init(&oap->oap_lock);
2376         CDEBUG(D_INFO, "oap %p page %p obj off %llu\n",
2377                oap, page, oap->oap_obj_off);
2378         RETURN(0);
2379 }
2380
2381 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2382                        struct osc_page *ops)
2383 {
2384         struct osc_io *oio = osc_env_io(env);
2385         struct osc_extent     *ext = NULL;
2386         struct osc_async_page *oap = &ops->ops_oap;
2387         struct client_obd     *cli = oap->oap_cli;
2388         struct osc_object     *osc = oap->oap_obj;
2389         pgoff_t index;
2390         unsigned int tmp;
2391         unsigned int grants = 0;
2392         u32    brw_flags = OBD_BRW_ASYNC;
2393         int    cmd = OBD_BRW_WRITE;
2394         int    need_release = 0;
2395         int    rc = 0;
2396         ENTRY;
2397
2398         if (oap->oap_magic != OAP_MAGIC)
2399                 RETURN(-EINVAL);
2400
2401         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2402                 RETURN(-EIO);
2403
2404         if (!list_empty(&oap->oap_pending_item) ||
2405             !list_empty(&oap->oap_rpc_item))
2406                 RETURN(-EBUSY);
2407
2408         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2409         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2410         if (cfs_capable(CFS_CAP_SYS_RESOURCE)) {
2411                 brw_flags |= OBD_BRW_NOQUOTA;
2412                 cmd |= OBD_BRW_NOQUOTA;
2413         }
2414
2415         /* check if the file's owner/group is over quota */
2416         if (!(cmd & OBD_BRW_NOQUOTA)) {
2417                 struct cl_object *obj;
2418                 struct cl_attr   *attr;
2419                 unsigned int qid[LL_MAXQUOTAS];
2420
2421                 obj = cl_object_top(&osc->oo_cl);
2422                 attr = &osc_env_info(env)->oti_attr;
2423
2424                 cl_object_attr_lock(obj);
2425                 rc = cl_object_attr_get(env, obj, attr);
2426                 cl_object_attr_unlock(obj);
2427
2428                 qid[USRQUOTA] = attr->cat_uid;
2429                 qid[GRPQUOTA] = attr->cat_gid;
2430                 qid[PRJQUOTA] = attr->cat_projid;
2431                 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2432                         rc = -EDQUOT;
2433                 if (rc)
2434                         RETURN(rc);
2435         }
2436
2437         oap->oap_cmd = cmd;
2438         oap->oap_page_off = ops->ops_from;
2439         oap->oap_count = ops->ops_to - ops->ops_from;
2440         /* No need to hold a lock here,
2441          * since this page is not in any list yet. */
2442         oap->oap_async_flags = 0;
2443         oap->oap_brw_flags = brw_flags;
2444
2445         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2446                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2447
2448         index = osc_index(oap2osc(oap));
2449
2450         /* Add this page into extent by the following steps:
2451          * 1. if there exists an active extent for this IO, mostly this page
2452          *    can be added to the active extent and sometimes we need to
2453          *    expand extent to accomodate this page;
2454          * 2. otherwise, a new extent will be allocated. */
2455
2456         ext = oio->oi_active;
2457         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2458                 /* one chunk plus extent overhead must be enough to write this
2459                  * page */
2460                 grants = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2461                 if (ext->oe_end >= index)
2462                         grants = 0;
2463
2464                 /* it doesn't need any grant to dirty this page */
2465                 spin_lock(&cli->cl_loi_list_lock);
2466                 rc = osc_enter_cache_try(cli, oap, grants, 0);
2467                 spin_unlock(&cli->cl_loi_list_lock);
2468                 if (rc == 0) { /* try failed */
2469                         grants = 0;
2470                         need_release = 1;
2471                 } else if (ext->oe_end < index) {
2472                         tmp = grants;
2473                         /* try to expand this extent */
2474                         rc = osc_extent_expand(ext, index, &tmp);
2475                         if (rc < 0) {
2476                                 need_release = 1;
2477                                 /* don't free reserved grant */
2478                         } else {
2479                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2480                                                 "expanded for %lu.\n", index);
2481                                 osc_unreserve_grant(cli, grants, tmp);
2482                                 grants = 0;
2483                         }
2484                 }
2485                 rc = 0;
2486         } else if (ext != NULL) {
2487                 /* index is located outside of active extent */
2488                 need_release = 1;
2489         }
2490         if (need_release) {
2491                 osc_extent_release(env, ext);
2492                 oio->oi_active = NULL;
2493                 ext = NULL;
2494         }
2495
2496         if (ext == NULL) {
2497                 tmp = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2498
2499                 /* try to find new extent to cover this page */
2500                 LASSERT(oio->oi_active == NULL);
2501                 /* we may have allocated grant for this page if we failed
2502                  * to expand the previous active extent. */
2503                 LASSERT(ergo(grants > 0, grants >= tmp));
2504
2505                 rc = 0;
2506                 if (grants == 0) {
2507                         /* we haven't allocated grant for this page. */
2508                         rc = osc_enter_cache(env, cli, oap, tmp);
2509                         if (rc == 0)
2510                                 grants = tmp;
2511                 }
2512
2513                 tmp = grants;
2514                 if (rc == 0) {
2515                         ext = osc_extent_find(env, osc, index, &tmp);
2516                         if (IS_ERR(ext)) {
2517                                 LASSERT(tmp == grants);
2518                                 osc_exit_cache(cli, oap);
2519                                 rc = PTR_ERR(ext);
2520                                 ext = NULL;
2521                         } else {
2522                                 oio->oi_active = ext;
2523                         }
2524                 }
2525                 if (grants > 0)
2526                         osc_unreserve_grant(cli, grants, tmp);
2527         }
2528
2529         LASSERT(ergo(rc == 0, ext != NULL));
2530         if (ext != NULL) {
2531                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2532                          ext, "index = %lu.\n", index);
2533                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2534
2535                 osc_object_lock(osc);
2536                 if (ext->oe_nr_pages == 0)
2537                         ext->oe_srvlock = ops->ops_srvlock;
2538                 else
2539                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2540                 ++ext->oe_nr_pages;
2541                 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2542                 osc_object_unlock(osc);
2543         }
2544         RETURN(rc);
2545 }
2546
2547 int osc_teardown_async_page(const struct lu_env *env,
2548                             struct osc_object *obj, struct osc_page *ops)
2549 {
2550         struct osc_async_page *oap = &ops->ops_oap;
2551         int rc = 0;
2552         ENTRY;
2553
2554         LASSERT(oap->oap_magic == OAP_MAGIC);
2555
2556         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2557                oap, ops, osc_index(oap2osc(oap)));
2558
2559         if (!list_empty(&oap->oap_rpc_item)) {
2560                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2561                 rc = -EBUSY;
2562         } else if (!list_empty(&oap->oap_pending_item)) {
2563                 struct osc_extent *ext = NULL;
2564
2565                 osc_object_lock(obj);
2566                 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2567                 osc_object_unlock(obj);
2568                 /* only truncated pages are allowed to be taken out.
2569                  * See osc_extent_truncate() and osc_cache_truncate_start()
2570                  * for details. */
2571                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2572                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2573                                         osc_index(oap2osc(oap)));
2574                         rc = -EBUSY;
2575                 }
2576                 if (ext != NULL)
2577                         osc_extent_put(env, ext);
2578         }
2579         RETURN(rc);
2580 }
2581
2582 /**
2583  * This is called when a page is picked up by kernel to write out.
2584  *
2585  * We should find out the corresponding extent and add the whole extent
2586  * into urgent list. The extent may be being truncated or used, handle it
2587  * carefully.
2588  */
2589 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2590                          struct osc_page *ops)
2591 {
2592         struct osc_extent *ext   = NULL;
2593         struct osc_object *obj   = cl2osc(ops->ops_cl.cpl_obj);
2594         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2595         pgoff_t            index = osc_index(ops);
2596         struct osc_async_page *oap = &ops->ops_oap;
2597         bool unplug = false;
2598         int rc = 0;
2599         ENTRY;
2600
2601         osc_object_lock(obj);
2602         ext = osc_extent_lookup(obj, index);
2603         if (ext == NULL) {
2604                 osc_extent_tree_dump(D_ERROR, obj);
2605                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2606         }
2607
2608         switch (ext->oe_state) {
2609         case OES_RPC:
2610         case OES_LOCK_DONE:
2611                 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2612                 LASSERT(0);
2613                 break;
2614         case OES_LOCKING:
2615                 /* If we know this extent is being written out, we should abort
2616                  * so that the writer can make this page ready. Otherwise, there
2617                  * exists a deadlock problem because other process can wait for
2618                  * page writeback bit holding page lock; and meanwhile in
2619                  * vvp_page_make_ready(), we need to grab page lock before
2620                  * really sending the RPC. */
2621         case OES_TRUNC:
2622                 /* race with truncate, page will be redirtied */
2623         case OES_ACTIVE:
2624                 /* The extent is active so we need to abort and let the caller
2625                  * re-dirty the page. If we continued on here, and we were the
2626                  * one making the extent active, we could deadlock waiting for
2627                  * the page writeback to clear but it won't because the extent
2628                  * is active and won't be written out. */
2629                 GOTO(out, rc = -EAGAIN);
2630         default:
2631                 break;
2632         }
2633
2634         rc = cl_page_prep(env, io, cp, CRT_WRITE);
2635         if (rc)
2636                 GOTO(out, rc);
2637
2638         spin_lock(&oap->oap_lock);
2639         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2640         spin_unlock(&oap->oap_lock);
2641
2642         if (memory_pressure_get())
2643                 ext->oe_memalloc = 1;
2644
2645         ext->oe_urgent = 1;
2646         if (ext->oe_state == OES_CACHE) {
2647                 OSC_EXTENT_DUMP(D_CACHE, ext,
2648                                 "flush page %p make it urgent.\n", oap);
2649                 if (list_empty(&ext->oe_link))
2650                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2651                 unplug = true;
2652         }
2653         rc = 0;
2654         EXIT;
2655
2656 out:
2657         osc_object_unlock(obj);
2658         osc_extent_put(env, ext);
2659         if (unplug)
2660                 osc_io_unplug_async(env, osc_cli(obj), obj);
2661         return rc;
2662 }
2663
2664 /**
2665  * this is called when a sync waiter receives an interruption.  Its job is to
2666  * get the caller woken as soon as possible.  If its page hasn't been put in an
2667  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2668  * desiring interruption which will forcefully complete the rpc once the rpc
2669  * has timed out.
2670  */
2671 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2672 {
2673         struct osc_async_page *oap = &ops->ops_oap;
2674         struct osc_object     *obj = oap->oap_obj;
2675         struct client_obd     *cli = osc_cli(obj);
2676         struct osc_extent     *ext;
2677         struct osc_extent     *found = NULL;
2678         struct list_head            *plist;
2679         pgoff_t index = osc_index(ops);
2680         int     rc = -EBUSY;
2681         int     cmd;
2682         ENTRY;
2683
2684         LASSERT(!oap->oap_interrupted);
2685         oap->oap_interrupted = 1;
2686
2687         /* Find out the caching extent */
2688         osc_object_lock(obj);
2689         if (oap->oap_cmd & OBD_BRW_WRITE) {
2690                 plist = &obj->oo_urgent_exts;
2691                 cmd   = OBD_BRW_WRITE;
2692         } else {
2693                 plist = &obj->oo_reading_exts;
2694                 cmd   = OBD_BRW_READ;
2695         }
2696         list_for_each_entry(ext, plist, oe_link) {
2697                 if (ext->oe_start <= index && ext->oe_end >= index) {
2698                         LASSERT(ext->oe_state == OES_LOCK_DONE);
2699                         /* For OES_LOCK_DONE state extent, it has already held
2700                          * a refcount for RPC. */
2701                         found = osc_extent_get(ext);
2702                         break;
2703                 }
2704         }
2705         if (found != NULL) {
2706                 list_del_init(&found->oe_link);
2707                 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2708                 osc_object_unlock(obj);
2709
2710                 osc_extent_finish(env, found, 0, -EINTR);
2711                 osc_extent_put(env, found);
2712                 rc = 0;
2713         } else {
2714                 osc_object_unlock(obj);
2715                 /* ok, it's been put in an rpc. only one oap gets a request
2716                  * reference */
2717                 if (oap->oap_request != NULL) {
2718                         ptlrpc_mark_interrupted(oap->oap_request);
2719                         ptlrpcd_wake(oap->oap_request);
2720                         ptlrpc_req_finished(oap->oap_request);
2721                         oap->oap_request = NULL;
2722                 }
2723         }
2724
2725         osc_list_maint(cli, obj);
2726         RETURN(rc);
2727 }
2728
2729 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
2730                          struct list_head *list, int cmd, int brw_flags)
2731 {
2732         struct client_obd     *cli = osc_cli(obj);
2733         struct osc_extent     *ext;
2734         struct osc_async_page *oap;
2735         int     page_count = 0;
2736         int     mppr       = cli->cl_max_pages_per_rpc;
2737         bool    can_merge   = true;
2738         pgoff_t start      = CL_PAGE_EOF;
2739         pgoff_t end        = 0;
2740         ENTRY;
2741
2742         list_for_each_entry(oap, list, oap_pending_item) {
2743                 struct osc_page *opg = oap2osc_page(oap);
2744                 pgoff_t index = osc_index(opg);
2745
2746                 if (index > end)
2747                         end = index;
2748                 if (index < start)
2749                         start = index;
2750                 ++page_count;
2751                 mppr <<= (page_count > mppr);
2752
2753                 if (unlikely(opg->ops_from > 0 || opg->ops_to < PAGE_SIZE))
2754                         can_merge = false;
2755         }
2756
2757         ext = osc_extent_alloc(obj);
2758         if (ext == NULL) {
2759                 struct osc_async_page *tmp;
2760
2761                 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2762                         list_del_init(&oap->oap_pending_item);
2763                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2764                 }
2765                 RETURN(-ENOMEM);
2766         }
2767
2768         ext->oe_rw = !!(cmd & OBD_BRW_READ);
2769         ext->oe_sync = 1;
2770         ext->oe_no_merge = !can_merge;
2771         ext->oe_urgent = 1;
2772         ext->oe_start = start;
2773         ext->oe_end = ext->oe_max_end = end;
2774         ext->oe_obj = obj;
2775         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2776         ext->oe_nr_pages = page_count;
2777         ext->oe_mppr = mppr;
2778         list_splice_init(list, &ext->oe_pages);
2779
2780         osc_object_lock(obj);
2781         /* Reuse the initial refcount for RPC, don't drop it */
2782         osc_extent_state_set(ext, OES_LOCK_DONE);
2783         if (cmd & OBD_BRW_WRITE) {
2784                 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2785                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2786         } else {
2787                 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2788                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2789         }
2790         osc_object_unlock(obj);
2791
2792         osc_io_unplug_async(env, cli, obj);
2793         RETURN(0);
2794 }
2795
2796 /**
2797  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2798  */
2799 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
2800                              __u64 size, struct osc_extent **extp)
2801 {
2802         struct client_obd *cli = osc_cli(obj);
2803         struct osc_extent *ext;
2804         struct osc_extent *waiting = NULL;
2805         pgoff_t index;
2806         struct list_head list = LIST_HEAD_INIT(list);
2807         int result = 0;
2808         bool partial;
2809         ENTRY;
2810
2811         /* pages with index greater or equal to index will be truncated. */
2812         index = cl_index(osc2cl(obj), size);
2813         partial = size > cl_offset(osc2cl(obj), index);
2814
2815 again:
2816         osc_object_lock(obj);
2817         ext = osc_extent_search(obj, index);
2818         if (ext == NULL)
2819                 ext = first_extent(obj);
2820         else if (ext->oe_end < index)
2821                 ext = next_extent(ext);
2822         while (ext != NULL) {
2823                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2824
2825                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2826                         /* if ext is in urgent state, it means there must exist
2827                          * a page already having been flushed by write_page().
2828                          * We have to wait for this extent because we can't
2829                          * truncate that page. */
2830                         OSC_EXTENT_DUMP(D_CACHE, ext,
2831                                         "waiting for busy extent\n");
2832                         waiting = osc_extent_get(ext);
2833                         break;
2834                 }
2835
2836                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:%llu.\n", size);
2837
2838                 osc_extent_get(ext);
2839                 if (ext->oe_state == OES_ACTIVE) {
2840                         /* though we grab inode mutex for write path, but we
2841                          * release it before releasing extent(in osc_io_end()),
2842                          * so there is a race window that an extent is still
2843                          * in OES_ACTIVE when truncate starts. */
2844                         LASSERT(!ext->oe_trunc_pending);
2845                         ext->oe_trunc_pending = 1;
2846                 } else {
2847                         EASSERT(ext->oe_state == OES_CACHE, ext);
2848                         osc_extent_state_set(ext, OES_TRUNC);
2849                         osc_update_pending(obj, OBD_BRW_WRITE,
2850                                            -ext->oe_nr_pages);
2851                 }
2852                 /* This extent could be on the full extents list, that's OK */
2853                 EASSERT(!ext->oe_hp && !ext->oe_urgent, ext);
2854                 if (!list_empty(&ext->oe_link))
2855                         list_move_tail(&ext->oe_link, &list);
2856                 else
2857                         list_add_tail(&ext->oe_link, &list);
2858
2859                 ext = next_extent(ext);
2860         }
2861         osc_object_unlock(obj);
2862
2863         osc_list_maint(cli, obj);
2864
2865         while (!list_empty(&list)) {
2866                 int rc;
2867
2868                 ext = list_entry(list.next, struct osc_extent, oe_link);
2869                 list_del_init(&ext->oe_link);
2870
2871                 /* extent may be in OES_ACTIVE state because inode mutex
2872                  * is released before osc_io_end() in file write case */
2873                 if (ext->oe_state != OES_TRUNC)
2874                         osc_extent_wait(env, ext, OES_TRUNC);
2875
2876                 rc = osc_extent_truncate(ext, index, partial);
2877                 if (rc < 0) {
2878                         if (result == 0)
2879                                 result = rc;
2880
2881                         OSC_EXTENT_DUMP(D_ERROR, ext,
2882                                         "truncate error %d\n", rc);
2883                 } else if (ext->oe_nr_pages == 0) {
2884                         osc_extent_remove(ext);
2885                 } else {
2886                         /* this must be an overlapped extent which means only
2887                          * part of pages in this extent have been truncated.
2888                          */
2889                         EASSERTF(ext->oe_start <= index, ext,
2890                                  "trunc index = %lu/%d.\n", index, partial);
2891                         /* fix index to skip this partially truncated extent */
2892                         index = ext->oe_end + 1;
2893                         partial = false;
2894
2895                         /* we need to hold this extent in OES_TRUNC state so
2896                          * that no writeback will happen. This is to avoid
2897                          * BUG 17397.
2898                          * Only partial truncate can reach here, if @size is
2899                          * not zero, the caller should provide a valid @extp. */
2900                         LASSERT(*extp == NULL);
2901                         *extp = osc_extent_get(ext);
2902                         OSC_EXTENT_DUMP(D_CACHE, ext,
2903                                         "trunc at %llu\n", size);
2904                 }
2905                 osc_extent_put(env, ext);
2906         }
2907         if (waiting != NULL) {
2908                 int rc;
2909
2910                 /* ignore the result of osc_extent_wait the write initiator
2911                  * should take care of it. */
2912                 rc = osc_extent_wait(env, waiting, OES_INV);
2913                 if (rc < 0)
2914                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2915
2916                 osc_extent_put(env, waiting);
2917                 waiting = NULL;
2918                 goto again;
2919         }
2920         RETURN(result);
2921 }
2922
2923 /**
2924  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2925  */
2926 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext)
2927 {
2928         if (ext != NULL) {
2929                 struct osc_object *obj = ext->oe_obj;
2930                 bool unplug = false;
2931
2932                 EASSERT(ext->oe_nr_pages > 0, ext);
2933                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2934                 EASSERT(!ext->oe_urgent, ext);
2935
2936                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2937                 osc_object_lock(obj);
2938                 osc_extent_state_set(ext, OES_CACHE);
2939                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2940                         ext->oe_urgent = 1;
2941                         list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2942                         unplug = true;
2943                 }
2944                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2945                 osc_object_unlock(obj);
2946                 osc_extent_put(env, ext);
2947
2948                 if (unplug)
2949                         osc_io_unplug_async(env, osc_cli(obj), obj);
2950         }
2951 }
2952
2953 /**
2954  * Wait for extents in a specific range to be written out.
2955  * The caller must have called osc_cache_writeback_range() to issue IO
2956  * otherwise it will take a long time for this function to finish.
2957  *
2958  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2959  * nobody else can dirty this range of file while we're waiting for
2960  * extents to be written.
2961  */
2962 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2963                          pgoff_t start, pgoff_t end)
2964 {
2965         struct osc_extent *ext;
2966         pgoff_t index = start;
2967         int     result = 0;
2968         ENTRY;
2969
2970 again:
2971         osc_object_lock(obj);
2972         ext = osc_extent_search(obj, index);
2973         if (ext == NULL)
2974                 ext = first_extent(obj);
2975         else if (ext->oe_end < index)
2976                 ext = next_extent(ext);
2977         while (ext != NULL) {
2978                 int rc;
2979
2980                 if (ext->oe_start > end)
2981                         break;
2982
2983                 if (!ext->oe_fsync_wait) {
2984                         ext = next_extent(ext);
2985                         continue;
2986                 }
2987
2988                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2989                              ext->oe_hp || ext->oe_urgent), ext);
2990                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2991                              !ext->oe_hp && ext->oe_urgent), ext);
2992
2993                 index = ext->oe_end + 1;
2994                 osc_extent_get(ext);
2995                 osc_object_unlock(obj);
2996
2997                 rc = osc_extent_wait(env, ext, OES_INV);
2998                 if (result == 0)
2999                         result = rc;
3000                 osc_extent_put(env, ext);
3001                 goto again;
3002         }
3003         osc_object_unlock(obj);
3004
3005         OSC_IO_DEBUG(obj, "sync file range.\n");
3006         RETURN(result);
3007 }
3008
3009 /**
3010  * Called to write out a range of osc object.
3011  *
3012  * @hp     : should be set this is caused by lock cancel;
3013  * @discard: is set if dirty pages should be dropped - file will be deleted or
3014  *         truncated, this implies there is no partially discarding extents.
3015  *
3016  * Return how many pages will be issued, or error code if error occurred.
3017  */
3018 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
3019                               pgoff_t start, pgoff_t end, int hp, int discard)
3020 {
3021         struct osc_extent *ext;
3022         struct list_head discard_list = LIST_HEAD_INIT(discard_list);
3023         bool unplug = false;
3024         int result = 0;
3025         ENTRY;
3026
3027         osc_object_lock(obj);
3028         ext = osc_extent_search(obj, start);
3029         if (ext == NULL)
3030                 ext = first_extent(obj);
3031         else if (ext->oe_end < start)
3032                 ext = next_extent(ext);
3033         while (ext != NULL) {
3034                 if (ext->oe_start > end)
3035                         break;
3036
3037                 ext->oe_fsync_wait = 1;
3038                 switch (ext->oe_state) {
3039                 case OES_CACHE:
3040                         result += ext->oe_nr_pages;
3041                         if (!discard) {
3042                                 struct list_head *list = NULL;
3043                                 if (hp) {
3044                                         EASSERT(!ext->oe_hp, ext);
3045                                         ext->oe_hp = 1;
3046                                         list = &obj->oo_hp_exts;
3047                                 } else if (!ext->oe_urgent) {
3048                                         ext->oe_urgent = 1;
3049                                         list = &obj->oo_urgent_exts;
3050                                 }
3051                                 if (list != NULL)
3052                                         list_move_tail(&ext->oe_link, list);
3053                                 unplug = true;
3054                         } else {
3055                                 /* the only discarder is lock cancelling, so
3056                                  * [start, end] must contain this extent */
3057                                 EASSERT(ext->oe_start >= start &&
3058                                         ext->oe_max_end <= end, ext);
3059                                 osc_extent_state_set(ext, OES_LOCKING);
3060                                 ext->oe_owner = current;
3061                                 list_move_tail(&ext->oe_link,
3062                                                    &discard_list);
3063                                 osc_update_pending(obj, OBD_BRW_WRITE,
3064                                                    -ext->oe_nr_pages);
3065                         }
3066                         break;
3067                 case OES_ACTIVE:
3068                         /* It's pretty bad to wait for ACTIVE extents, because
3069                          * we don't know how long we will wait for it to be
3070                          * flushed since it may be blocked at awaiting more
3071                          * grants. We do this for the correctness of fsync. */
3072                         LASSERT(hp == 0 && discard == 0);
3073                         ext->oe_urgent = 1;
3074                         break;
3075                 case OES_TRUNC:
3076                         /* this extent is being truncated, can't do anything
3077                          * for it now. it will be set to urgent after truncate
3078                          * is finished in osc_cache_truncate_end(). */
3079                 default:
3080                         break;
3081                 }
3082                 ext = next_extent(ext);
3083         }
3084         osc_object_unlock(obj);
3085
3086         LASSERT(ergo(!discard, list_empty(&discard_list)));
3087         if (!list_empty(&discard_list)) {
3088                 struct osc_extent *tmp;
3089                 int rc;
3090
3091                 osc_list_maint(osc_cli(obj), obj);
3092                 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
3093                         list_del_init(&ext->oe_link);
3094                         EASSERT(ext->oe_state == OES_LOCKING, ext);
3095
3096                         /* Discard caching pages. We don't actually write this
3097                          * extent out but we complete it as if we did. */
3098                         rc = osc_extent_make_ready(env, ext);
3099                         if (unlikely(rc < 0)) {
3100                                 OSC_EXTENT_DUMP(D_ERROR, ext,
3101                                                 "make_ready returned %d\n", rc);
3102                                 if (result >= 0)
3103                                         result = rc;
3104                         }
3105
3106                         /* finish the extent as if the pages were sent */
3107                         osc_extent_finish(env, ext, 0, 0);
3108                 }
3109         }
3110
3111         if (unplug)
3112                 osc_io_unplug(env, osc_cli(obj), obj);
3113
3114         if (hp || discard) {
3115                 int rc;
3116                 rc = osc_cache_wait_range(env, obj, start, end);
3117                 if (result >= 0 && rc < 0)
3118                         result = rc;
3119         }
3120
3121         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3122         RETURN(result);
3123 }
3124
3125 /**
3126  * Returns a list of pages by a given [start, end] of \a obj.
3127  *
3128  * \param resched If not NULL, then we give up before hogging CPU for too
3129  * long and set *resched = 1, in that case caller should implement a retry
3130  * logic.
3131  *
3132  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3133  * crucial in the face of [offset, EOF] locks.
3134  *
3135  * Return at least one page in @queue unless there is no covered page.
3136  */
3137 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3138                         struct osc_object *osc, pgoff_t start, pgoff_t end,
3139                         osc_page_gang_cbt cb, void *cbdata)
3140 {
3141         struct osc_page *ops;
3142         void            **pvec;
3143         pgoff_t         idx;
3144         unsigned int    nr;
3145         unsigned int    i;
3146         unsigned int    j;
3147         int             res = CLP_GANG_OKAY;
3148         bool            tree_lock = true;
3149         ENTRY;
3150
3151         idx = start;
3152         pvec = osc_env_info(env)->oti_pvec;
3153         spin_lock(&osc->oo_tree_lock);
3154         while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3155                                             idx, OTI_PVEC_SIZE)) > 0) {
3156                 struct cl_page *page;
3157                 bool end_of_region = false;
3158
3159                 for (i = 0, j = 0; i < nr; ++i) {
3160                         ops = pvec[i];
3161                         pvec[i] = NULL;
3162
3163                         idx = osc_index(ops);
3164                         if (idx > end) {
3165                                 end_of_region = true;
3166                                 break;
3167                         }
3168
3169                         page = ops->ops_cl.cpl_page;
3170                         LASSERT(page->cp_type == CPT_CACHEABLE);
3171                         if (page->cp_state == CPS_FREEING)
3172                                 continue;
3173
3174                         cl_page_get(page);
3175                         lu_ref_add_atomic(&page->cp_reference,
3176                                           "gang_lookup", current);
3177                         pvec[j++] = ops;
3178                 }
3179                 ++idx;
3180
3181                 /*
3182                  * Here a delicate locking dance is performed. Current thread
3183                  * holds a reference to a page, but has to own it before it
3184                  * can be placed into queue. Owning implies waiting, so
3185                  * radix-tree lock is to be released. After a wait one has to
3186                  * check that pages weren't truncated (cl_page_own() returns
3187                  * error in the latter case).
3188                  */
3189                 spin_unlock(&osc->oo_tree_lock);
3190                 tree_lock = false;
3191
3192                 for (i = 0; i < j; ++i) {
3193                         ops = pvec[i];
3194                         if (res == CLP_GANG_OKAY)
3195                                 res = (*cb)(env, io, ops, cbdata);
3196
3197                         page = ops->ops_cl.cpl_page;
3198                         lu_ref_del(&page->cp_reference, "gang_lookup", current);
3199                         cl_page_put(env, page);
3200                 }
3201                 if (nr < OTI_PVEC_SIZE || end_of_region)
3202                         break;
3203
3204                 if (res == CLP_GANG_OKAY && need_resched())
3205                         res = CLP_GANG_RESCHED;
3206                 if (res != CLP_GANG_OKAY)
3207                         break;
3208
3209                 spin_lock(&osc->oo_tree_lock);
3210                 tree_lock = true;
3211         }
3212         if (tree_lock)
3213                 spin_unlock(&osc->oo_tree_lock);
3214         RETURN(res);
3215 }
3216
3217 /**
3218  * Check if page @page is covered by an extra lock or discard it.
3219  */
3220 static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3221                                 struct osc_page *ops, void *cbdata)
3222 {
3223         struct osc_thread_info *info = osc_env_info(env);
3224         struct osc_object *osc = cbdata;
3225         pgoff_t index;
3226
3227         index = osc_index(ops);
3228         if (index >= info->oti_fn_index) {
3229                 struct ldlm_lock *tmp;
3230                 struct cl_page *page = ops->ops_cl.cpl_page;
3231
3232                 /* refresh non-overlapped index */
3233                 tmp = osc_dlmlock_at_pgoff(env, osc, index,
3234                                            OSC_DAP_FL_TEST_LOCK);
3235                 if (tmp != NULL) {
3236                         __u64 end = tmp->l_policy_data.l_extent.end;
3237                         /* Cache the first-non-overlapped index so as to skip
3238                          * all pages within [index, oti_fn_index). This is safe
3239                          * because if tmp lock is canceled, it will discard
3240                          * these pages. */
3241                         info->oti_fn_index = cl_index(osc2cl(osc), end + 1);
3242                         if (end == OBD_OBJECT_EOF)
3243                                 info->oti_fn_index = CL_PAGE_EOF;
3244                         LDLM_LOCK_PUT(tmp);
3245                 } else if (cl_page_own(env, io, page) == 0) {
3246                         /* discard the page */
3247                         cl_page_discard(env, io, page);
3248                         cl_page_disown(env, io, page);
3249                 } else {
3250                         LASSERT(page->cp_state == CPS_FREEING);
3251                 }
3252         }
3253
3254         info->oti_next_index = index + 1;
3255         return CLP_GANG_OKAY;
3256 }
3257
3258 static int discard_cb(const struct lu_env *env, struct cl_io *io,
3259                       struct osc_page *ops, void *cbdata)
3260 {
3261         struct osc_thread_info *info = osc_env_info(env);
3262         struct cl_page *page = ops->ops_cl.cpl_page;
3263
3264         /* page is top page. */
3265         info->oti_next_index = osc_index(ops) + 1;
3266         if (cl_page_own(env, io, page) == 0) {
3267                 if (!ergo(page->cp_type == CPT_CACHEABLE,
3268                           !PageDirty(cl_page_vmpage(page))))
3269                         CL_PAGE_DEBUG(D_ERROR, env, page,
3270                                         "discard dirty page?\n");
3271
3272                 /* discard the page */
3273                 cl_page_discard(env, io, page);
3274                 cl_page_disown(env, io, page);
3275         } else {
3276                 LASSERT(page->cp_state == CPS_FREEING);
3277         }
3278
3279         return CLP_GANG_OKAY;
3280 }
3281
3282 /**
3283  * Discard pages protected by the given lock. This function traverses radix
3284  * tree to find all covering pages and discard them. If a page is being covered
3285  * by other locks, it should remain in cache.
3286  *
3287  * If error happens on any step, the process continues anyway (the reasoning
3288  * behind this being that lock cancellation cannot be delayed indefinitely).
3289  */
3290 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3291                            pgoff_t start, pgoff_t end, bool discard)
3292 {
3293         struct osc_thread_info *info = osc_env_info(env);
3294         struct cl_io *io = &info->oti_io;
3295         osc_page_gang_cbt cb;
3296         int res;
3297         int result;
3298
3299         ENTRY;
3300
3301         io->ci_obj = cl_object_top(osc2cl(osc));
3302         io->ci_ignore_layout = 1;
3303         result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3304         if (result != 0)
3305                 GOTO(out, result);
3306
3307         cb = discard ? discard_cb : check_and_discard_cb;
3308         info->oti_fn_index = info->oti_next_index = start;
3309         do {
3310                 res = osc_page_gang_lookup(env, io, osc,
3311                                            info->oti_next_index, end, cb, osc);
3312                 if (info->oti_next_index > end)
3313                         break;
3314
3315                 if (res == CLP_GANG_RESCHED)
3316                         cond_resched();
3317         } while (res != CLP_GANG_OKAY);
3318 out:
3319         cl_io_fini(env, io);
3320         RETURN(result);
3321 }
3322
3323
3324 /** @} osc */