lustre/osc/osc_cache.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, 2013, Intel Corporation.
  31  *
  32  */
  33 /*
  34  * This file is part of Lustre, http://www.lustre.org/
  35  * Lustre is a trademark of Sun Microsystems, Inc.
  36  *
  37  * osc cache management.
  38  *
  39  * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_OSC
  43
  44 #include "osc_cl_internal.h"
  45 #include "osc_internal.h"
  46
  47 static int extent_debug; /* set it to be true for more debug */
  48
  49 static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
  50 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
  51                            int state);
  52 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
  53                               struct osc_async_page *oap, int sent, int rc);
  54 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
  55                           int cmd);
  56 static int osc_refresh_count(const struct lu_env *env,
  57                              struct osc_async_page *oap, int cmd);
  58 static int osc_io_unplug_async(const struct lu_env *env,
  59                                struct client_obd *cli, struct osc_object *osc);
  60 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
  61                            unsigned int lost_grant);
  62
  63 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
  64                                   const char *func, int line);
  65 #define osc_extent_tree_dump(lvl, obj) \
  66         osc_extent_tree_dump0(lvl, obj, __func__, __LINE__)
  67
  68 /** \addtogroup osc
  69  *  @{
  70  */
  71
  72 /* ------------------ osc extent ------------------ */
  73 static inline char *ext_flags(struct osc_extent *ext, char *flags)
  74 {
  75         char *buf = flags;
  76         *buf++ = ext->oe_rw ? 'r' : 'w';
  77         if (ext->oe_intree)
  78                 *buf++ = 'i';
  79         if (ext->oe_srvlock)
  80                 *buf++ = 's';
  81         if (ext->oe_hp)
  82                 *buf++ = 'h';
  83         if (ext->oe_urgent)
  84                 *buf++ = 'u';
  85         if (ext->oe_memalloc)
  86                 *buf++ = 'm';
  87         if (ext->oe_trunc_pending)
  88                 *buf++ = 't';
  89         if (ext->oe_fsync_wait)
  90                 *buf++ = 'Y';
  91         *buf = 0;
  92         return flags;
  93 }
  94
  95 static inline char list_empty_marker(struct list_head *list)
  96 {
  97         return list_empty(list) ? '-' : '+';
  98 }
  99
 100 #define EXTSTR       "[%lu -> %lu/%lu]"
 101 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
 102 static const char *oes_strings[] = {
 103         "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
 104
 105 #define OSC_EXTENT_DUMP(lvl, extent, fmt, ...) do {                     \
 106         struct osc_extent *__ext = (extent);                            \
 107         char __buf[16];                                                 \
 108                                                                         \
 109         CDEBUG(lvl,                                                     \
 110                 "extent %p@{" EXTSTR ", "                               \
 111                 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt,    \
 112                 /* ----- extent part 0 ----- */                         \
 113                 __ext, EXTPARA(__ext),                                  \
 114                 /* ----- part 1 ----- */                                \
 115                 atomic_read(&__ext->oe_refc),                           \
 116                 atomic_read(&__ext->oe_users),                          \
 117                 list_empty_marker(&__ext->oe_link),                     \
 118                 oes_strings[__ext->oe_state], ext_flags(__ext, __buf),  \
 119                 __ext->oe_obj,                                          \
 120                 /* ----- part 2 ----- */                                \
 121                 __ext->oe_grants, __ext->oe_nr_pages,                   \
 122                 list_empty_marker(&__ext->oe_pages),                    \
 123                 waitqueue_active(&__ext->oe_waitq) ? '+' : '-',         \
 124                 __ext->oe_osclock, __ext->oe_mppr, __ext->oe_owner,     \
 125                 /* ----- part 4 ----- */                                \
 126                 ## __VA_ARGS__);                                        \
 127 } while (0)
 128
 129 #undef EASSERTF
 130 #define EASSERTF(expr, ext, fmt, args...) do {                          \
 131         if (!(expr)) {                                                  \
 132                 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args);           \
 133                 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj);           \
 134                 LASSERT(expr);                                          \
 135         }                                                               \
 136 } while (0)
 137
 138 #undef EASSERT
 139 #define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
 140
 141 static inline struct osc_extent *rb_extent(struct rb_node *n)
 142 {
 143         if (n == NULL)
 144                 return NULL;
 145
 146         return container_of(n, struct osc_extent, oe_node);
 147 }
 148
 149 static inline struct osc_extent *next_extent(struct osc_extent *ext)
 150 {
 151         if (ext == NULL)
 152                 return NULL;
 153
 154         LASSERT(ext->oe_intree);
 155         return rb_extent(rb_next(&ext->oe_node));
 156 }
 157
 158 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
 159 {
 160         if (ext == NULL)
 161                 return NULL;
 162
 163         LASSERT(ext->oe_intree);
 164         return rb_extent(rb_prev(&ext->oe_node));
 165 }
 166
 167 static inline struct osc_extent *first_extent(struct osc_object *obj)
 168 {
 169         return rb_extent(rb_first(&obj->oo_root));
 170 }
 171
 172 /* object must be locked by caller. */
 173 static int osc_extent_sanity_check0(struct osc_extent *ext,
 174                                     const char *func, const int line)
 175 {
 176         struct osc_object *obj = ext->oe_obj;
 177         struct osc_async_page *oap;
 178         int page_count;
 179         int rc = 0;
 180
 181         if (!osc_object_is_locked(obj))
 182                 GOTO(out, rc = 9);
 183
 184         if (ext->oe_state >= OES_STATE_MAX)
 185                 GOTO(out, rc = 10);
 186
 187         if (atomic_read(&ext->oe_refc) <= 0)
 188                 GOTO(out, rc = 20);
 189
 190         if (atomic_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
 191                 GOTO(out, rc = 30);
 192
 193         switch (ext->oe_state) {
 194         case OES_INV:
 195                 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
 196                         GOTO(out, rc = 35);
 197                 GOTO(out, rc = 0);
 198                 break;
 199         case OES_ACTIVE:
 200                 if (atomic_read(&ext->oe_users) == 0)
 201                         GOTO(out, rc = 40);
 202                 if (ext->oe_hp)
 203                         GOTO(out, rc = 50);
 204                 if (ext->oe_fsync_wait && !ext->oe_urgent)
 205                         GOTO(out, rc = 55);
 206                 break;
 207         case OES_CACHE:
 208                 if (ext->oe_grants == 0)
 209                         GOTO(out, rc = 60);
 210                 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
 211                         GOTO(out, rc = 65);
 212         default:
 213                 if (atomic_read(&ext->oe_users) > 0)
 214                         GOTO(out, rc = 70);
 215         }
 216
 217         if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
 218                 GOTO(out, rc = 80);
 219
 220         if (ext->oe_osclock == NULL && ext->oe_grants > 0)
 221                 GOTO(out, rc = 90);
 222
 223         if (ext->oe_osclock) {
 224                 struct cl_lock_descr *descr;
 225                 descr = &ext->oe_osclock->cll_descr;
 226                 if (!(descr->cld_start <= ext->oe_start &&
 227                       descr->cld_end >= ext->oe_max_end))
 228                         GOTO(out, rc = 100);
 229         }
 230
 231         if (ext->oe_nr_pages > ext->oe_mppr)
 232                 GOTO(out, rc = 105);
 233
 234         /* Do not verify page list if extent is in RPC. This is because an
 235          * in-RPC extent is supposed to be exclusively accessible w/o lock. */
 236         if (ext->oe_state > OES_CACHE)
 237                 GOTO(out, rc = 0);
 238
 239         if (!extent_debug)
 240                 GOTO(out, rc = 0);
 241
 242         page_count = 0;
 243         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
 244                 pgoff_t index = osc_index(oap2osc(oap));
 245                 ++page_count;
 246                 if (index > ext->oe_end || index < ext->oe_start)
 247                         GOTO(out, rc = 110);
 248         }
 249         if (page_count != ext->oe_nr_pages)
 250                 GOTO(out, rc = 120);
 251
 252 out:
 253         if (rc != 0)
 254                 OSC_EXTENT_DUMP(D_ERROR, ext,
 255                                 "%s:%d sanity check %p failed with rc = %d\n",
 256                                 func, line, ext, rc);
 257         return rc;
 258 }
 259
 260 #define sanity_check_nolock(ext) \
 261         osc_extent_sanity_check0(ext, __func__, __LINE__)
 262
 263 #define sanity_check(ext) ({                                                   \
 264         int __res;                                                             \
 265         osc_object_lock((ext)->oe_obj);                                        \
 266         __res = sanity_check_nolock(ext);                                      \
 267         osc_object_unlock((ext)->oe_obj);                                      \
 268         __res;                                                                 \
 269 })
 270
 271
 272 /**
 273  * sanity check - to make sure there is no overlapped extent in the tree.
 274  */
 275 static int osc_extent_is_overlapped(struct osc_object *obj,
 276                                     struct osc_extent *ext)
 277 {
 278         struct osc_extent *tmp;
 279
 280         LASSERT(osc_object_is_locked(obj));
 281
 282         if (!extent_debug)
 283                 return 0;
 284
 285         for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
 286                 if (tmp == ext)
 287                         continue;
 288                 if (tmp->oe_end >= ext->oe_start &&
 289                     tmp->oe_start <= ext->oe_end)
 290                         return 1;
 291         }
 292         return 0;
 293 }
 294
 295 static void osc_extent_state_set(struct osc_extent *ext, int state)
 296 {
 297         LASSERT(osc_object_is_locked(ext->oe_obj));
 298         LASSERT(state >= OES_INV && state < OES_STATE_MAX);
 299
 300         /* Never try to sanity check a state changing extent :-) */
 301         /* LASSERT(sanity_check_nolock(ext) == 0); */
 302
 303         /* TODO: validate the state machine */
 304         ext->oe_state = state;
 305         wake_up_all(&ext->oe_waitq);
 306 }
 307
 308 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
 309 {
 310         struct osc_extent *ext;
 311
 312         OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_IOFS);
 313         if (ext == NULL)
 314                 return NULL;
 315
 316         RB_CLEAR_NODE(&ext->oe_node);
 317         ext->oe_obj = obj;
 318         atomic_set(&ext->oe_refc, 1);
 319         atomic_set(&ext->oe_users, 0);
 320         INIT_LIST_HEAD(&ext->oe_link);
 321         ext->oe_state = OES_INV;
 322         INIT_LIST_HEAD(&ext->oe_pages);
 323         init_waitqueue_head(&ext->oe_waitq);
 324         ext->oe_osclock = NULL;
 325
 326         return ext;
 327 }
 328
 329 static void osc_extent_free(struct osc_extent *ext)
 330 {
 331         OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
 332 }
 333
 334 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
 335 {
 336         LASSERT(atomic_read(&ext->oe_refc) >= 0);
 337         atomic_inc(&ext->oe_refc);
 338         return ext;
 339 }
 340
 341 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
 342 {
 343         LASSERT(atomic_read(&ext->oe_refc) > 0);
 344         if (atomic_dec_and_test(&ext->oe_refc)) {
 345                 LASSERT(list_empty(&ext->oe_link));
 346                 LASSERT(atomic_read(&ext->oe_users) == 0);
 347                 LASSERT(ext->oe_state == OES_INV);
 348                 LASSERT(!ext->oe_intree);
 349
 350                 if (ext->oe_osclock) {
 351                         cl_lock_put(env, ext->oe_osclock);
 352                         ext->oe_osclock = NULL;
 353                 }
 354                 osc_extent_free(ext);
 355         }
 356 }
 357
 358 /**
 359  * osc_extent_put_trust() is a special version of osc_extent_put() when
 360  * it's known that the caller is not the last user. This is to address the
 361  * problem of lacking of lu_env ;-).
 362  */
 363 static void osc_extent_put_trust(struct osc_extent *ext)
 364 {
 365         LASSERT(atomic_read(&ext->oe_refc) > 1);
 366         LASSERT(osc_object_is_locked(ext->oe_obj));
 367         atomic_dec(&ext->oe_refc);
 368 }
 369
 370 /**
 371  * Return the extent which includes pgoff @index, or return the greatest
 372  * previous extent in the tree.
 373  */
 374 static struct osc_extent *osc_extent_search(struct osc_object *obj,
 375                                             pgoff_t index)
 376 {
 377         struct rb_node    *n = obj->oo_root.rb_node;
 378         struct osc_extent *tmp, *p = NULL;
 379
 380         LASSERT(osc_object_is_locked(obj));
 381         while (n != NULL) {
 382                 tmp = rb_extent(n);
 383                 if (index < tmp->oe_start) {
 384                         n = n->rb_left;
 385                 } else if (index > tmp->oe_end) {
 386                         p = rb_extent(n);
 387                         n = n->rb_right;
 388                 } else {
 389                         return tmp;
 390                 }
 391         }
 392         return p;
 393 }
 394
 395 /*
 396  * Return the extent covering @index, otherwise return NULL.
 397  * caller must have held object lock.
 398  */
 399 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
 400                                             pgoff_t index)
 401 {
 402         struct osc_extent *ext;
 403
 404         ext = osc_extent_search(obj, index);
 405         if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
 406                 return osc_extent_get(ext);
 407         return NULL;
 408 }
 409
 410 /* caller must have held object lock. */
 411 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
 412 {
 413         struct rb_node   **n      = &obj->oo_root.rb_node;
 414         struct rb_node    *parent = NULL;
 415         struct osc_extent *tmp;
 416
 417         LASSERT(ext->oe_intree == 0);
 418         LASSERT(ext->oe_obj == obj);
 419         LASSERT(osc_object_is_locked(obj));
 420         while (*n != NULL) {
 421                 tmp = rb_extent(*n);
 422                 parent = *n;
 423
 424                 if (ext->oe_end < tmp->oe_start)
 425                         n = &(*n)->rb_left;
 426                 else if (ext->oe_start > tmp->oe_end)
 427                         n = &(*n)->rb_right;
 428                 else
 429                         EASSERTF(0, tmp, EXTSTR, EXTPARA(ext));
 430         }
 431         rb_link_node(&ext->oe_node, parent, n);
 432         rb_insert_color(&ext->oe_node, &obj->oo_root);
 433         osc_extent_get(ext);
 434         ext->oe_intree = 1;
 435 }
 436
 437 /* caller must have held object lock. */
 438 static void osc_extent_erase(struct osc_extent *ext)
 439 {
 440         struct osc_object *obj = ext->oe_obj;
 441         LASSERT(osc_object_is_locked(obj));
 442         if (ext->oe_intree) {
 443                 rb_erase(&ext->oe_node, &obj->oo_root);
 444                 ext->oe_intree = 0;
 445                 /* rbtree held a refcount */
 446                 osc_extent_put_trust(ext);
 447         }
 448 }
 449
 450 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
 451 {
 452         struct osc_object *obj = ext->oe_obj;
 453
 454         LASSERT(osc_object_is_locked(obj));
 455         LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
 456         if (ext->oe_state == OES_CACHE) {
 457                 osc_extent_state_set(ext, OES_ACTIVE);
 458                 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
 459         }
 460         atomic_inc(&ext->oe_users);
 461         list_del_init(&ext->oe_link);
 462         return osc_extent_get(ext);
 463 }
 464
 465 static void __osc_extent_remove(struct osc_extent *ext)
 466 {
 467         LASSERT(osc_object_is_locked(ext->oe_obj));
 468         LASSERT(list_empty(&ext->oe_pages));
 469         osc_extent_erase(ext);
 470         list_del_init(&ext->oe_link);
 471         osc_extent_state_set(ext, OES_INV);
 472         OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
 473 }
 474
 475 static void osc_extent_remove(struct osc_extent *ext)
 476 {
 477         struct osc_object *obj = ext->oe_obj;
 478
 479         osc_object_lock(obj);
 480         __osc_extent_remove(ext);
 481         osc_object_unlock(obj);
 482 }
 483
 484 /**
 485  * This function is used to merge extents to get better performance. It checks
 486  * if @cur and @victim are contiguous at chunk level.
 487  */
 488 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
 489                             struct osc_extent *victim)
 490 {
 491         struct osc_object *obj = cur->oe_obj;
 492         pgoff_t chunk_start;
 493         pgoff_t chunk_end;
 494         int ppc_bits;
 495
 496         LASSERT(cur->oe_state == OES_CACHE);
 497         LASSERT(osc_object_is_locked(obj));
 498         if (victim == NULL)
 499                 return -EINVAL;
 500
 501         if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
 502                 return -EBUSY;
 503
 504         if (cur->oe_max_end != victim->oe_max_end)
 505                 return -ERANGE;
 506
 507         LASSERT(cur->oe_osclock == victim->oe_osclock);
 508         ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_CACHE_SHIFT;
 509         chunk_start = cur->oe_start >> ppc_bits;
 510         chunk_end   = cur->oe_end   >> ppc_bits;
 511         if (chunk_start   != (victim->oe_end >> ppc_bits) + 1 &&
 512             chunk_end + 1 != victim->oe_start >> ppc_bits)
 513                 return -ERANGE;
 514
 515         OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
 516
 517         cur->oe_start     = min(cur->oe_start, victim->oe_start);
 518         cur->oe_end       = max(cur->oe_end,   victim->oe_end);
 519         cur->oe_grants   += victim->oe_grants;
 520         cur->oe_nr_pages += victim->oe_nr_pages;
 521         /* only the following bits are needed to merge */
 522         cur->oe_urgent   |= victim->oe_urgent;
 523         cur->oe_memalloc |= victim->oe_memalloc;
 524         list_splice_init(&victim->oe_pages, &cur->oe_pages);
 525         list_del_init(&victim->oe_link);
 526         victim->oe_nr_pages = 0;
 527
 528         osc_extent_get(victim);
 529         __osc_extent_remove(victim);
 530         osc_extent_put(env, victim);
 531
 532         OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
 533         return 0;
 534 }
 535
 536 /**
 537  * Drop user count of osc_extent, and unplug IO asynchronously.
 538  */
 539 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
 540 {
 541         struct osc_object *obj = ext->oe_obj;
 542         int rc = 0;
 543         ENTRY;
 544
 545         LASSERT(atomic_read(&ext->oe_users) > 0);
 546         LASSERT(sanity_check(ext) == 0);
 547         LASSERT(ext->oe_grants > 0);
 548
 549         if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
 550                 LASSERT(ext->oe_state == OES_ACTIVE);
 551                 if (ext->oe_trunc_pending) {
 552                         /* a truncate process is waiting for this extent.
 553                          * This may happen due to a race, check
 554                          * osc_cache_truncate_start(). */
 555                         osc_extent_state_set(ext, OES_TRUNC);
 556                         ext->oe_trunc_pending = 0;
 557                 } else {
 558                         osc_extent_state_set(ext, OES_CACHE);
 559                         osc_update_pending(obj, OBD_BRW_WRITE,
 560                                            ext->oe_nr_pages);
 561
 562                         /* try to merge the previous and next extent. */
 563                         osc_extent_merge(env, ext, prev_extent(ext));
 564                         osc_extent_merge(env, ext, next_extent(ext));
 565
 566                         if (ext->oe_urgent)
 567                                 list_move_tail(&ext->oe_link,
 568                                                &obj->oo_urgent_exts);
 569                 }
 570                 osc_object_unlock(obj);
 571
 572                 osc_io_unplug_async(env, osc_cli(obj), obj);
 573         }
 574         osc_extent_put(env, ext);
 575         RETURN(rc);
 576 }
 577
 578 static inline int overlapped(struct osc_extent *ex1, struct osc_extent *ex2)
 579 {
 580         return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
 581 }
 582
 583 /**
 584  * Find or create an extent which includes @index, core function to manage
 585  * extent tree.
 586  */
 587 struct osc_extent *osc_extent_find(const struct lu_env *env,
 588                                    struct osc_object *obj, pgoff_t index,
 589                                    int *grants)
 590
 591 {
 592         struct client_obd *cli = osc_cli(obj);
 593         struct cl_lock    *lock;
 594         struct osc_extent *cur;
 595         struct osc_extent *ext;
 596         struct osc_extent *conflict = NULL;
 597         struct osc_extent *found = NULL;
 598         pgoff_t    chunk;
 599         pgoff_t    max_end;
 600         int        max_pages; /* max_pages_per_rpc */
 601         int        chunksize;
 602         int        ppc_bits; /* pages per chunk bits */
 603         int        chunk_mask;
 604         int        rc;
 605         ENTRY;
 606
 607         cur = osc_extent_alloc(obj);
 608         if (cur == NULL)
 609                 RETURN(ERR_PTR(-ENOMEM));
 610
 611         lock = cl_lock_at_pgoff(env, osc2cl(obj), index, NULL, 1, 0);
 612         LASSERT(lock != NULL);
 613         LASSERT(lock->cll_descr.cld_mode >= CLM_WRITE);
 614
 615         LASSERT(cli->cl_chunkbits >= PAGE_CACHE_SHIFT);
 616         ppc_bits   = cli->cl_chunkbits - PAGE_CACHE_SHIFT;
 617         chunk_mask = ~((1 << ppc_bits) - 1);
 618         chunksize  = 1 << cli->cl_chunkbits;
 619         chunk      = index >> ppc_bits;
 620
 621         /* align end to rpc edge, rpc size may not be a power 2 integer. */
 622         max_pages = cli->cl_max_pages_per_rpc;
 623         LASSERT((max_pages & ~chunk_mask) == 0);
 624         max_end = index - (index % max_pages) + max_pages - 1;
 625         max_end = min_t(pgoff_t, max_end, lock->cll_descr.cld_end);
 626
 627         /* initialize new extent by parameters so far */
 628         cur->oe_max_end = max_end;
 629         cur->oe_start   = index & chunk_mask;
 630         cur->oe_end     = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
 631         if (cur->oe_start < lock->cll_descr.cld_start)
 632                 cur->oe_start = lock->cll_descr.cld_start;
 633         if (cur->oe_end > max_end)
 634                 cur->oe_end = max_end;
 635         cur->oe_osclock = lock;
 636         cur->oe_grants  = 0;
 637         cur->oe_mppr    = max_pages;
 638
 639         /* grants has been allocated by caller */
 640         LASSERTF(*grants >= chunksize + cli->cl_extent_tax,
 641                  "%u/%u/%u.\n", *grants, chunksize, cli->cl_extent_tax);
 642         LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR, EXTPARA(cur));
 643
 644 restart:
 645         osc_object_lock(obj);
 646         ext = osc_extent_search(obj, cur->oe_start);
 647         if (ext == NULL)
 648                 ext = first_extent(obj);
 649         while (ext != NULL) {
 650                 loff_t ext_chk_start = ext->oe_start >> ppc_bits;
 651                 loff_t ext_chk_end   = ext->oe_end   >> ppc_bits;
 652
 653                 LASSERT(sanity_check_nolock(ext) == 0);
 654                 if (chunk > ext_chk_end + 1)
 655                         break;
 656
 657                 /* if covering by different locks, no chance to match */
 658                 if (lock != ext->oe_osclock) {
 659                         EASSERTF(!overlapped(ext, cur), ext,
 660                                  EXTSTR, EXTPARA(cur));
 661
 662                         ext = next_extent(ext);
 663                         continue;
 664                 }
 665
 666                 /* discontiguous chunks? */
 667                 if (chunk + 1 < ext_chk_start) {
 668                         ext = next_extent(ext);
 669                         continue;
 670                 }
 671
 672                 /* ok, from now on, ext and cur have these attrs:
 673                  * 1. covered by the same lock
 674                  * 2. contiguous at chunk level or overlapping. */
 675
 676                 if (overlapped(ext, cur)) {
 677                         /* cur is the minimum unit, so overlapping means
 678                          * full contain. */
 679                         EASSERTF((ext->oe_start <= cur->oe_start &&
 680                                   ext->oe_end >= cur->oe_end),
 681                                  ext, EXTSTR, EXTPARA(cur));
 682
 683                         if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
 684                                 /* for simplicity, we wait for this extent to
 685                                  * finish before going forward. */
 686                                 conflict = osc_extent_get(ext);
 687                                 break;
 688                         }
 689
 690                         found = osc_extent_hold(ext);
 691                         break;
 692                 }
 693
 694                 /* non-overlapped extent */
 695                 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
 696                         /* we can't do anything for a non OES_CACHE extent, or
 697                          * if there is someone waiting for this extent to be
 698                          * flushed, try next one. */
 699                         ext = next_extent(ext);
 700                         continue;
 701                 }
 702
 703                 /* check if they belong to the same rpc slot before trying to
 704                  * merge. the extents are not overlapped and contiguous at
 705                  * chunk level to get here. */
 706                 if (ext->oe_max_end != max_end) {
 707                         /* if they don't belong to the same RPC slot or
 708                          * max_pages_per_rpc has ever changed, do not merge. */
 709                         ext = next_extent(ext);
 710                         continue;
 711                 }
 712
 713                 /* it's required that an extent must be contiguous at chunk
 714                  * level so that we know the whole extent is covered by grant
 715                  * (the pages in the extent are NOT required to be contiguous).
 716                  * Otherwise, it will be too much difficult to know which
 717                  * chunks have grants allocated. */
 718
 719                 /* try to do front merge - extend ext's start */
 720                 if (chunk + 1 == ext_chk_start) {
 721                         /* ext must be chunk size aligned */
 722                         EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
 723
 724                         /* pull ext's start back to cover cur */
 725                         ext->oe_start   = cur->oe_start;
 726                         ext->oe_grants += chunksize;
 727                         *grants -= chunksize;
 728
 729                         found = osc_extent_hold(ext);
 730                 } else if (chunk == ext_chk_end + 1) {
 731                         /* rear merge */
 732                         ext->oe_end     = cur->oe_end;
 733                         ext->oe_grants += chunksize;
 734                         *grants -= chunksize;
 735
 736                         /* try to merge with the next one because we just fill
 737                          * in a gap */
 738                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 739                                 /* we can save extent tax from next extent */
 740                                 *grants += cli->cl_extent_tax;
 741
 742                         found = osc_extent_hold(ext);
 743                 }
 744                 if (found != NULL)
 745                         break;
 746
 747                 ext = next_extent(ext);
 748         }
 749
 750         osc_extent_tree_dump(D_CACHE, obj);
 751         if (found != NULL) {
 752                 LASSERT(conflict == NULL);
 753                 if (!IS_ERR(found)) {
 754                         LASSERT(found->oe_osclock == cur->oe_osclock);
 755                         OSC_EXTENT_DUMP(D_CACHE, found,
 756                                         "found caching ext for %lu.\n", index);
 757                 }
 758         } else if (conflict == NULL) {
 759                 /* create a new extent */
 760                 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
 761                 cur->oe_grants = chunksize + cli->cl_extent_tax;
 762                 *grants -= cur->oe_grants;
 763                 LASSERT(*grants >= 0);
 764
 765                 cur->oe_state = OES_CACHE;
 766                 found = osc_extent_hold(cur);
 767                 osc_extent_insert(obj, cur);
 768                 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
 769                                 index, lock->cll_descr.cld_end);
 770         }
 771         osc_object_unlock(obj);
 772
 773         if (conflict != NULL) {
 774                 LASSERT(found == NULL);
 775
 776                 /* waiting for IO to finish. Please notice that it's impossible
 777                  * to be an OES_TRUNC extent. */
 778                 rc = osc_extent_wait(env, conflict, OES_INV);
 779                 osc_extent_put(env, conflict);
 780                 conflict = NULL;
 781                 if (rc < 0)
 782                         GOTO(out, found = ERR_PTR(rc));
 783
 784                 goto restart;
 785         }
 786         EXIT;
 787
 788 out:
 789         osc_extent_put(env, cur);
 790         LASSERT(*grants >= 0);
 791         return found;
 792 }
 793
 794 /**
 795  * Called when IO is finished to an extent.
 796  */
 797 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
 798                       int sent, int rc)
 799 {
 800         struct client_obd *cli = osc_cli(ext->oe_obj);
 801         struct osc_async_page *oap;
 802         struct osc_async_page *tmp;
 803         int nr_pages = ext->oe_nr_pages;
 804         int lost_grant = 0;
 805         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
 806         __u64 last_off = 0;
 807         int last_count = -1;
 808         ENTRY;
 809
 810         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
 811
 812         ext->oe_rc = rc ?: ext->oe_nr_pages;
 813         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
 814
 815         osc_lru_add_batch(cli, &ext->oe_pages);
 816         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 817                                      oap_pending_item) {
 818                 list_del_init(&oap->oap_rpc_item);
 819                 list_del_init(&oap->oap_pending_item);
 820                 if (last_off <= oap->oap_obj_off) {
 821                         last_off = oap->oap_obj_off;
 822                         last_count = oap->oap_count;
 823                 }
 824
 825                 --ext->oe_nr_pages;
 826                 osc_ap_completion(env, cli, oap, sent, rc);
 827         }
 828         EASSERT(ext->oe_nr_pages == 0, ext);
 829
 830         if (!sent) {
 831                 lost_grant = ext->oe_grants;
 832         } else if (blocksize < PAGE_CACHE_SIZE &&
 833                    last_count != PAGE_CACHE_SIZE) {
 834                 /* For short writes we shouldn't count parts of pages that
 835                  * span a whole chunk on the OST side, or our accounting goes
 836                  * wrong.  Should match the code in filter_grant_check. */
 837                 int offset = last_off & ~CFS_PAGE_MASK;
 838                 int count = last_count + (offset & (blocksize - 1));
 839                 int end = (offset + last_count) & (blocksize - 1);
 840                 if (end)
 841                         count += blocksize - end;
 842
 843                 lost_grant = PAGE_CACHE_SIZE - count;
 844         }
 845         if (ext->oe_grants > 0)
 846                 osc_free_grant(cli, nr_pages, lost_grant);
 847
 848         osc_extent_remove(ext);
 849         /* put the refcount for RPC */
 850         osc_extent_put(env, ext);
 851         RETURN(0);
 852 }
 853
 854 static int extent_wait_cb(struct osc_extent *ext, int state)
 855 {
 856         int ret;
 857
 858         osc_object_lock(ext->oe_obj);
 859         ret = ext->oe_state == state;
 860         osc_object_unlock(ext->oe_obj);
 861
 862         return ret;
 863 }
 864
 865 /**
 866  * Wait for the extent's state to become @state.
 867  */
 868 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
 869                            int state)
 870 {
 871         struct osc_object *obj = ext->oe_obj;
 872         struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
 873                                                   LWI_ON_SIGNAL_NOOP, NULL);
 874         int rc = 0;
 875         ENTRY;
 876
 877         osc_object_lock(obj);
 878         LASSERT(sanity_check_nolock(ext) == 0);
 879         /* `Kick' this extent only if the caller is waiting for it to be
 880          * written out. */
 881         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
 882                 if (ext->oe_state == OES_ACTIVE) {
 883                         ext->oe_urgent = 1;
 884                 } else if (ext->oe_state == OES_CACHE) {
 885                         ext->oe_urgent = 1;
 886                         osc_extent_hold(ext);
 887                         rc = 1;
 888                 }
 889         }
 890         osc_object_unlock(obj);
 891         if (rc == 1)
 892                 osc_extent_release(env, ext);
 893
 894         /* wait for the extent until its state becomes @state */
 895         rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
 896         if (rc == -ETIMEDOUT) {
 897                 OSC_EXTENT_DUMP(D_ERROR, ext,
 898                         "%s: wait ext to %d timedout, recovery in progress?\n",
 899                         osc_export(obj)->exp_obd->obd_name, state);
 900
 901                 lwi = LWI_INTR(NULL, NULL);
 902                 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state),
 903                                   &lwi);
 904         }
 905         if (rc == 0 && ext->oe_rc < 0)
 906                 rc = ext->oe_rc;
 907         RETURN(rc);
 908 }
 909
 910 /**
 911  * Discard pages with index greater than @size. If @ext is overlapped with
 912  * @size, then partial truncate happens.
 913  */
 914 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
 915                                 bool partial)
 916 {
 917         struct cl_env_nest     nest;
 918         struct lu_env         *env;
 919         struct cl_io          *io;
 920         struct osc_object     *obj = ext->oe_obj;
 921         struct client_obd     *cli = osc_cli(obj);
 922         struct osc_async_page *oap;
 923         struct osc_async_page *tmp;
 924         int                    pages_in_chunk = 0;
 925         int                    ppc_bits    = cli->cl_chunkbits -
 926                                              PAGE_CACHE_SHIFT;
 927         __u64                  trunc_chunk = trunc_index >> ppc_bits;
 928         int                    grants   = 0;
 929         int                    nr_pages = 0;
 930         int                    rc       = 0;
 931         ENTRY;
 932
 933         LASSERT(sanity_check(ext) == 0);
 934         LASSERT(ext->oe_state == OES_TRUNC);
 935         LASSERT(!ext->oe_urgent);
 936
 937         /* Request new lu_env.
 938          * We can't use that env from osc_cache_truncate_start() because
 939          * it's from lov_io_sub and not fully initialized. */
 940         env = cl_env_nested_get(&nest);
 941         io  = &osc_env_info(env)->oti_io;
 942         io->ci_obj = cl_object_top(osc2cl(obj));
 943         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
 944         if (rc < 0)
 945                 GOTO(out, rc);
 946
 947         /* discard all pages with index greater then trunc_index */
 948         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 949                                      oap_pending_item) {
 950                 pgoff_t index = osc_index(oap2osc(oap));
 951                 struct cl_page  *page = oap2cl_page(oap);
 952
 953                 LASSERT(list_empty(&oap->oap_rpc_item));
 954
 955                 /* only discard the pages with their index greater than
 956                  * trunc_index, and ... */
 957                 if (index < trunc_index ||
 958                     (index == trunc_index && partial)) {
 959                         /* accounting how many pages remaining in the chunk
 960                          * so that we can calculate grants correctly. */
 961                         if (index >> ppc_bits == trunc_chunk)
 962                                 ++pages_in_chunk;
 963                         continue;
 964                 }
 965
 966                 list_del_init(&oap->oap_pending_item);
 967
 968                 cl_page_get(page);
 969                 lu_ref_add(&page->cp_reference, "truncate", current);
 970
 971                 if (cl_page_own(env, io, page) == 0) {
 972                         cl_page_discard(env, io, page);
 973                         cl_page_disown(env, io, page);
 974                 } else {
 975                         LASSERT(page->cp_state == CPS_FREEING);
 976                         LASSERT(0);
 977                 }
 978
 979                 lu_ref_del(&page->cp_reference, "truncate", current);
 980                 cl_page_put(env, page);
 981
 982                 --ext->oe_nr_pages;
 983                 ++nr_pages;
 984         }
 985         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
 986                       ext->oe_nr_pages == 0),
 987                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
 988
 989         osc_object_lock(obj);
 990         if (ext->oe_nr_pages == 0) {
 991                 LASSERT(pages_in_chunk == 0);
 992                 grants = ext->oe_grants;
 993                 ext->oe_grants = 0;
 994         } else { /* calculate how many grants we can free */
 995                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
 996                 pgoff_t last_index;
 997
 998
 999                 /* if there is no pages in this chunk, we can also free grants
1000                  * for the last chunk */
1001                 if (pages_in_chunk == 0) {
1002                         /* if this is the 1st chunk and no pages in this chunk,
1003                          * ext->oe_nr_pages must be zero, so we should be in
1004                          * the other if-clause. */
1005                         LASSERT(trunc_chunk > 0);
1006                         --trunc_chunk;
1007                         ++chunks;
1008                 }
1009
1010                 /* this is what we can free from this extent */
1011                 grants          = chunks << cli->cl_chunkbits;
1012                 ext->oe_grants -= grants;
1013                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1014                 ext->oe_end     = min(last_index, ext->oe_max_end);
1015                 LASSERT(ext->oe_end >= ext->oe_start);
1016                 LASSERT(ext->oe_grants > 0);
1017         }
1018         osc_object_unlock(obj);
1019
1020         if (grants > 0 || nr_pages > 0)
1021                 osc_free_grant(cli, nr_pages, grants);
1022
1023 out:
1024         cl_io_fini(env, io);
1025         cl_env_nested_put(&nest, env);
1026         RETURN(rc);
1027 }
1028
1029 /**
1030  * This function is used to make the extent prepared for transfer.
1031  * A race with flusing page - ll_writepage() has to be handled cautiously.
1032  */
1033 static int osc_extent_make_ready(const struct lu_env *env,
1034                                  struct osc_extent *ext)
1035 {
1036         struct osc_async_page *oap;
1037         struct osc_async_page *last = NULL;
1038         struct osc_object *obj = ext->oe_obj;
1039         int page_count = 0;
1040         int rc;
1041         ENTRY;
1042
1043         /* we're going to grab page lock, so object lock must not be taken. */
1044         LASSERT(sanity_check(ext) == 0);
1045         /* in locking state, any process should not touch this extent. */
1046         EASSERT(ext->oe_state == OES_LOCKING, ext);
1047         EASSERT(ext->oe_owner != NULL, ext);
1048
1049         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1050
1051         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1052                 ++page_count;
1053                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1054                         last = oap;
1055
1056                 /* checking ASYNC_READY is race safe */
1057                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1058                         continue;
1059
1060                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1061                 switch (rc) {
1062                 case 0:
1063                         spin_lock(&oap->oap_lock);
1064                         oap->oap_async_flags |= ASYNC_READY;
1065                         spin_unlock(&oap->oap_lock);
1066                         break;
1067                 case -EALREADY:
1068                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1069                         break;
1070                 default:
1071                         LASSERTF(0, "unknown return code: %d\n", rc);
1072                 }
1073         }
1074
1075         LASSERT(page_count == ext->oe_nr_pages);
1076         LASSERT(last != NULL);
1077         /* the last page is the only one we need to refresh its count by
1078          * the size of file. */
1079         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1080                 last->oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1081                 LASSERT(last->oap_count > 0);
1082                 LASSERT(last->oap_page_off + last->oap_count <= PAGE_CACHE_SIZE);
1083                 spin_lock(&last->oap_lock);
1084                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1085                 spin_unlock(&last->oap_lock);
1086         }
1087
1088         /* for the rest of pages, we don't need to call osf_refresh_count()
1089          * because it's known they are not the last page */
1090         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1091                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1092                         oap->oap_count = PAGE_CACHE_SIZE - oap->oap_page_off;
1093                         spin_lock(&oap->oap_lock);
1094                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1095                         spin_unlock(&oap->oap_lock);
1096                 }
1097         }
1098
1099         osc_object_lock(obj);
1100         osc_extent_state_set(ext, OES_RPC);
1101         osc_object_unlock(obj);
1102         /* get a refcount for RPC. */
1103         osc_extent_get(ext);
1104
1105         RETURN(0);
1106 }
1107
1108 /**
1109  * Quick and simple version of osc_extent_find(). This function is frequently
1110  * called to expand the extent for the same IO. To expand the extent, the
1111  * page index must be in the same or next chunk of ext->oe_end.
1112  */
1113 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index, int *grants)
1114 {
1115         struct osc_object *obj = ext->oe_obj;
1116         struct client_obd *cli = osc_cli(obj);
1117         struct osc_extent *next;
1118         int ppc_bits = cli->cl_chunkbits - PAGE_CACHE_SHIFT;
1119         pgoff_t chunk = index >> ppc_bits;
1120         pgoff_t end_chunk;
1121         pgoff_t end_index;
1122         int chunksize = 1 << cli->cl_chunkbits;
1123         int rc = 0;
1124         ENTRY;
1125
1126         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1127         osc_object_lock(obj);
1128         LASSERT(sanity_check_nolock(ext) == 0);
1129         end_chunk = ext->oe_end >> ppc_bits;
1130         if (chunk > end_chunk + 1)
1131                 GOTO(out, rc = -ERANGE);
1132
1133         if (end_chunk >= chunk)
1134                 GOTO(out, rc = 0);
1135
1136         LASSERT(end_chunk + 1 == chunk);
1137         /* try to expand this extent to cover @index */
1138         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1139
1140         next = next_extent(ext);
1141         if (next != NULL && next->oe_start <= end_index)
1142                 /* complex mode - overlapped with the next extent,
1143                  * this case will be handled by osc_extent_find() */
1144                 GOTO(out, rc = -EAGAIN);
1145
1146         ext->oe_end = end_index;
1147         ext->oe_grants += chunksize;
1148         *grants -= chunksize;
1149         LASSERT(*grants >= 0);
1150         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1151                  "overlapped after expanding for %lu.\n", index);
1152         EXIT;
1153
1154 out:
1155         osc_object_unlock(obj);
1156         RETURN(rc);
1157 }
1158
1159 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
1160                                   const char *func, int line)
1161 {
1162         struct osc_extent *ext;
1163         int cnt;
1164
1165         CDEBUG(level, "Dump object %p extents at %s:%d, mppr: %u.\n",
1166                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1167
1168         /* osc_object_lock(obj); */
1169         cnt = 1;
1170         for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1171                 OSC_EXTENT_DUMP(level, ext, "in tree %d.\n", cnt++);
1172
1173         cnt = 1;
1174         list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1175                 OSC_EXTENT_DUMP(level, ext, "hp %d.\n", cnt++);
1176
1177         cnt = 1;
1178         list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1179                 OSC_EXTENT_DUMP(level, ext, "urgent %d.\n", cnt++);
1180
1181         cnt = 1;
1182         list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1183                 OSC_EXTENT_DUMP(level, ext, "reading %d.\n", cnt++);
1184         /* osc_object_unlock(obj); */
1185 }
1186
1187 /* ------------------ osc extent end ------------------ */
1188
1189 static inline int osc_is_ready(struct osc_object *osc)
1190 {
1191         return !list_empty(&osc->oo_ready_item) ||
1192                !list_empty(&osc->oo_hp_ready_item);
1193 }
1194
1195 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1196         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1197                (OSC), osc_is_ready(OSC),                                       \
1198                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1199                list_empty_marker(&(OSC)->oo_ready_item),                       \
1200                atomic_read(&(OSC)->oo_nr_writes),                              \
1201                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1202                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1203                atomic_read(&(OSC)->oo_nr_reads),                               \
1204                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1205                ##args)
1206
1207 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1208                           int cmd)
1209 {
1210         struct osc_page *opg  = oap2osc_page(oap);
1211         struct cl_page  *page = oap2cl_page(oap);
1212         int result;
1213
1214         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1215
1216         ENTRY;
1217         result = cl_page_make_ready(env, page, CRT_WRITE);
1218         if (result == 0)
1219                 opg->ops_submit_time = cfs_time_current();
1220         RETURN(result);
1221 }
1222
1223 static int osc_refresh_count(const struct lu_env *env,
1224                              struct osc_async_page *oap, int cmd)
1225 {
1226         struct osc_page  *opg = oap2osc_page(oap);
1227         pgoff_t index = osc_index(oap2osc(oap));
1228         struct cl_object *obj;
1229         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
1230
1231         int result;
1232         loff_t kms;
1233
1234         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1235         LASSERT(!(cmd & OBD_BRW_READ));
1236         LASSERT(opg != NULL);
1237         obj = opg->ops_cl.cpl_obj;
1238
1239         cl_object_attr_lock(obj);
1240         result = cl_object_attr_get(env, obj, attr);
1241         cl_object_attr_unlock(obj);
1242         if (result < 0)
1243                 return result;
1244         kms = attr->cat_kms;
1245         if (cl_offset(obj, index) >= kms)
1246                 /* catch race with truncate */
1247                 return 0;
1248         else if (cl_offset(obj, index + 1) > kms)
1249                 /* catch sub-page write at end of file */
1250                 return kms % PAGE_CACHE_SIZE;
1251         else
1252                 return PAGE_CACHE_SIZE;
1253 }
1254
1255 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1256                           int cmd, int rc)
1257 {
1258         struct osc_page   *opg  = oap2osc_page(oap);
1259         struct cl_page    *page = oap2cl_page(oap);
1260         struct osc_object *obj  = cl2osc(opg->ops_cl.cpl_obj);
1261         enum cl_req_type   crt;
1262         int srvlock;
1263
1264         ENTRY;
1265
1266         cmd &= ~OBD_BRW_NOQUOTA;
1267         LASSERTF(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ),
1268                  "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1269         LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1270                 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1271         LASSERT(opg->ops_transfer_pinned);
1272
1273         /*
1274          * page->cp_req can be NULL if io submission failed before
1275          * cl_req was allocated.
1276          */
1277         if (page->cp_req != NULL)
1278                 cl_req_page_done(env, page);
1279         LASSERT(page->cp_req == NULL);
1280
1281         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1282         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1283         opg->ops_transfer_pinned = 0;
1284
1285         spin_lock(&obj->oo_seatbelt);
1286         LASSERT(opg->ops_submitter != NULL);
1287         LASSERT(!list_empty(&opg->ops_inflight));
1288         list_del_init(&opg->ops_inflight);
1289         opg->ops_submitter = NULL;
1290         spin_unlock(&obj->oo_seatbelt);
1291
1292         opg->ops_submit_time = 0;
1293         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1294
1295         /* statistic */
1296         if (rc == 0 && srvlock) {
1297                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1298                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1299                 int bytes = oap->oap_count;
1300
1301                 if (crt == CRT_READ)
1302                         stats->os_lockless_reads += bytes;
1303                 else
1304                         stats->os_lockless_writes += bytes;
1305         }
1306
1307         /*
1308          * This has to be the last operation with the page, as locks are
1309          * released in cl_page_completion() and nothing except for the
1310          * reference counter protects page from concurrent reclaim.
1311          */
1312         lu_ref_del(&page->cp_reference, "transfer", page);
1313
1314         cl_page_completion(env, page, crt, rc);
1315
1316         RETURN(0);
1317 }
1318
1319 #define OSC_DUMP_GRANT(lvl, cli, fmt, args...) do {                     \
1320         struct client_obd *__tmp = (cli);                               \
1321         CDEBUG(lvl, "%s: grant { dirty: %ld/%ld dirty_pages: %d/%d "    \
1322                "unstable_pages: %d/%d dropped: %ld avail: %ld, "        \
1323                "reserved: %ld, flight: %d } lru {in list: %d, "         \
1324                "left: %d, waiters: %d }" fmt,                           \
1325                __tmp->cl_import->imp_obd->obd_name,                     \
1326                __tmp->cl_dirty, __tmp->cl_dirty_max,                    \
1327                atomic_read(&obd_dirty_pages), obd_max_dirty_pages,      \
1328                atomic_read(&obd_unstable_pages), obd_max_dirty_pages,   \
1329                __tmp->cl_lost_grant, __tmp->cl_avail_grant,             \
1330                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,         \
1331                atomic_read(&__tmp->cl_lru_in_list),                     \
1332                atomic_read(&__tmp->cl_lru_busy),                        \
1333                atomic_read(&__tmp->cl_lru_shrinkers), ##args);          \
1334 } while (0)
1335
1336 /* caller must hold loi_list_lock */
1337 static void osc_consume_write_grant(struct client_obd *cli,
1338                                     struct brw_page *pga)
1339 {
1340         assert_spin_locked(&cli->cl_loi_list_lock.lock);
1341         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1342         atomic_inc(&obd_dirty_pages);
1343         cli->cl_dirty += PAGE_CACHE_SIZE;
1344         pga->flag |= OBD_BRW_FROM_GRANT;
1345         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1346                PAGE_CACHE_SIZE, pga, pga->pg);
1347         osc_update_next_shrink(cli);
1348 }
1349
1350 /* the companion to osc_consume_write_grant, called when a brw has completed.
1351  * must be called with the loi lock held. */
1352 static void osc_release_write_grant(struct client_obd *cli,
1353                                     struct brw_page *pga)
1354 {
1355         ENTRY;
1356
1357         assert_spin_locked(&cli->cl_loi_list_lock.lock);
1358         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1359                 EXIT;
1360                 return;
1361         }
1362
1363         pga->flag &= ~OBD_BRW_FROM_GRANT;
1364         atomic_dec(&obd_dirty_pages);
1365         cli->cl_dirty -= PAGE_CACHE_SIZE;
1366         if (pga->flag & OBD_BRW_NOCACHE) {
1367                 pga->flag &= ~OBD_BRW_NOCACHE;
1368                 atomic_dec(&obd_dirty_transit_pages);
1369                 cli->cl_dirty_transit -= PAGE_CACHE_SIZE;
1370         }
1371         EXIT;
1372 }
1373
1374 /**
1375  * To avoid sleeping with object lock held, it's good for us allocate enough
1376  * grants before entering into critical section.
1377  *
1378  * client_obd_list_lock held by caller
1379  */
1380 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1381 {
1382         int rc = -EDQUOT;
1383
1384         if (cli->cl_avail_grant >= bytes) {
1385                 cli->cl_avail_grant    -= bytes;
1386                 cli->cl_reserved_grant += bytes;
1387                 rc = 0;
1388         }
1389         return rc;
1390 }
1391
1392 static void __osc_unreserve_grant(struct client_obd *cli,
1393                                   unsigned int reserved, unsigned int unused)
1394 {
1395         /* it's quite normal for us to get more grant than reserved.
1396          * Thinking about a case that two extents merged by adding a new
1397          * chunk, we can save one extent tax. If extent tax is greater than
1398          * one chunk, we can save more grant by adding a new chunk */
1399         cli->cl_reserved_grant -= reserved;
1400         if (unused > reserved) {
1401                 cli->cl_avail_grant += reserved;
1402                 cli->cl_lost_grant  += unused - reserved;
1403         } else {
1404                 cli->cl_avail_grant += unused;
1405         }
1406 }
1407
1408 void osc_unreserve_grant(struct client_obd *cli,
1409                          unsigned int reserved, unsigned int unused)
1410 {
1411         client_obd_list_lock(&cli->cl_loi_list_lock);
1412         __osc_unreserve_grant(cli, reserved, unused);
1413         if (unused > 0)
1414                 osc_wake_cache_waiters(cli);
1415         client_obd_list_unlock(&cli->cl_loi_list_lock);
1416 }
1417
1418 /**
1419  * Free grant after IO is finished or canceled.
1420  *
1421  * @lost_grant is used to remember how many grants we have allocated but not
1422  * used, we should return these grants to OST. There're two cases where grants
1423  * can be lost:
1424  * 1. truncate;
1425  * 2. blocksize at OST is less than PAGE_CACHE_SIZE and a partial page was
1426  *    written. In this case OST may use less chunks to serve this partial
1427  *    write. OSTs don't actually know the page size on the client side. so
1428  *    clients have to calculate lost grant by the blocksize on the OST.
1429  *    See filter_grant_check() for details.
1430  */
1431 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1432                            unsigned int lost_grant)
1433 {
1434         int grant = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
1435
1436         client_obd_list_lock(&cli->cl_loi_list_lock);
1437         atomic_sub(nr_pages, &obd_dirty_pages);
1438         cli->cl_dirty -= nr_pages << PAGE_CACHE_SHIFT;
1439         cli->cl_lost_grant += lost_grant;
1440         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1441                 /* borrow some grant from truncate to avoid the case that
1442                  * truncate uses up all avail grant */
1443                 cli->cl_lost_grant -= grant;
1444                 cli->cl_avail_grant += grant;
1445         }
1446         osc_wake_cache_waiters(cli);
1447         client_obd_list_unlock(&cli->cl_loi_list_lock);
1448         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu\n",
1449                lost_grant, cli->cl_lost_grant,
1450                cli->cl_avail_grant, cli->cl_dirty);
1451 }
1452
1453 /**
1454  * The companion to osc_enter_cache(), called when @oap is no longer part of
1455  * the dirty accounting due to error.
1456  */
1457 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1458 {
1459         client_obd_list_lock(&cli->cl_loi_list_lock);
1460         osc_release_write_grant(cli, &oap->oap_brw_page);
1461         client_obd_list_unlock(&cli->cl_loi_list_lock);
1462 }
1463
1464 /**
1465  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1466  * is available.
1467  */
1468 static int osc_enter_cache_try(struct client_obd *cli,
1469                                struct osc_async_page *oap,
1470                                int bytes, int transient)
1471 {
1472         int rc;
1473
1474         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d.\n", bytes);
1475
1476         rc = osc_reserve_grant(cli, bytes);
1477         if (rc < 0)
1478                 return 0;
1479
1480         if (cli->cl_dirty + PAGE_CACHE_SIZE <= cli->cl_dirty_max &&
1481             atomic_read(&obd_unstable_pages) + 1 +
1482             atomic_read(&obd_dirty_pages) <= obd_max_dirty_pages) {
1483                 osc_consume_write_grant(cli, &oap->oap_brw_page);
1484                 if (transient) {
1485                         cli->cl_dirty_transit += PAGE_CACHE_SIZE;
1486                         atomic_inc(&obd_dirty_transit_pages);
1487                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
1488                 }
1489                 rc = 1;
1490         } else {
1491                 __osc_unreserve_grant(cli, bytes, bytes);
1492                 rc = 0;
1493         }
1494         return rc;
1495 }
1496
1497 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1498 {
1499         int rc;
1500         client_obd_list_lock(&cli->cl_loi_list_lock);
1501         rc = list_empty(&ocw->ocw_entry);
1502         client_obd_list_unlock(&cli->cl_loi_list_lock);
1503         return rc;
1504 }
1505
1506 /**
1507  * The main entry to reserve dirty page accounting. Usually the grant reserved
1508  * in this function will be freed in bulk in osc_free_grant() unless it fails
1509  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1510  *
1511  * The process will be put into sleep if it's already run out of grant.
1512  */
1513 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1514                            struct osc_async_page *oap, int bytes)
1515 {
1516         struct osc_object *osc = oap->oap_obj;
1517         struct lov_oinfo  *loi = osc->oo_oinfo;
1518         struct osc_cache_waiter ocw;
1519         struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
1520                                                   LWI_ON_SIGNAL_NOOP, NULL);
1521         int rc = -EDQUOT;
1522         ENTRY;
1523
1524         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d.\n", bytes);
1525
1526         client_obd_list_lock(&cli->cl_loi_list_lock);
1527
1528         /* force the caller to try sync io.  this can jump the list
1529          * of queued writes and create a discontiguous rpc stream */
1530         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1531             cli->cl_dirty_max < PAGE_CACHE_SIZE     ||
1532             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync)
1533                 GOTO(out, rc = -EDQUOT);
1534
1535         /* Hopefully normal case - cache space and write credits available */
1536         if (osc_enter_cache_try(cli, oap, bytes, 0))
1537                 GOTO(out, rc = 0);
1538
1539         /* We can get here for two reasons: too many dirty pages in cache, or
1540          * run out of grants. In both cases we should write dirty pages out.
1541          * Adding a cache waiter will trigger urgent write-out no matter what
1542          * RPC size will be.
1543          * The exiting condition is no avail grants and no dirty pages caching,
1544          * that really means there is no space on the OST. */
1545         init_waitqueue_head(&ocw.ocw_waitq);
1546         ocw.ocw_oap   = oap;
1547         ocw.ocw_grant = bytes;
1548         while (cli->cl_dirty > 0 || cli->cl_w_in_flight > 0) {
1549                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1550                 ocw.ocw_rc = 0;
1551                 client_obd_list_unlock(&cli->cl_loi_list_lock);
1552
1553                 osc_io_unplug_async(env, cli, NULL);
1554
1555                 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1556                        cli->cl_import->imp_obd->obd_name, &ocw, oap);
1557
1558                 rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1559
1560                 client_obd_list_lock(&cli->cl_loi_list_lock);
1561
1562                 /* l_wait_event is interrupted by signal, or timed out */
1563                 if (rc < 0) {
1564                         switch (rc) {
1565                         case -ETIMEDOUT:
1566                                 OSC_DUMP_GRANT(D_ERROR, cli,
1567                                                 "try to reserve %d.\n", bytes);
1568                                 osc_extent_tree_dump(D_ERROR, osc);
1569                                 rc = -EDQUOT;
1570                                 break;
1571                         case -EINTR:
1572                                 /* Ensures restartability - LU-3581 */
1573                                 rc = -ERESTARTSYS;
1574                                 break;
1575                         default:
1576                                 CDEBUG(D_CACHE, "%s: event for cache space @"
1577                                        " %p never arrived due to %d\n",
1578                                        cli->cl_import->imp_obd->obd_name,
1579                                        &ocw, rc);
1580                                 break;
1581                         }
1582                         list_del_init(&ocw.ocw_entry);
1583                         GOTO(out, rc);
1584                 }
1585
1586                 LASSERT(list_empty(&ocw.ocw_entry));
1587                 rc = ocw.ocw_rc;
1588
1589                 if (rc != -EDQUOT)
1590                         GOTO(out, rc);
1591                 if (osc_enter_cache_try(cli, oap, bytes, 0))
1592                         GOTO(out, rc = 0);
1593         }
1594         EXIT;
1595 out:
1596         client_obd_list_unlock(&cli->cl_loi_list_lock);
1597         OSC_DUMP_GRANT(D_CACHE, cli, "returned %d.\n", rc);
1598         RETURN(rc);
1599 }
1600
1601 /* caller must hold loi_list_lock */
1602 void osc_wake_cache_waiters(struct client_obd *cli)
1603 {
1604         struct list_head *l, *tmp;
1605         struct osc_cache_waiter *ocw;
1606
1607         ENTRY;
1608         list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1609                 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
1610                 list_del_init(&ocw->ocw_entry);
1611
1612                 ocw->ocw_rc = -EDQUOT;
1613                 /* we can't dirty more */
1614                 if ((cli->cl_dirty + PAGE_CACHE_SIZE > cli->cl_dirty_max) ||
1615                     (atomic_read(&obd_unstable_pages) + 1 +
1616                      atomic_read(&obd_dirty_pages) > obd_max_dirty_pages)) {
1617                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
1618                                "osc max %ld, sys max %d\n", cli->cl_dirty,
1619                                cli->cl_dirty_max, obd_max_dirty_pages);
1620                         goto wakeup;
1621                 }
1622
1623                 ocw->ocw_rc = 0;
1624                 if (!osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
1625                         ocw->ocw_rc = -EDQUOT;
1626
1627 wakeup:
1628                 CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld, %d\n",
1629                        ocw, ocw->ocw_oap, cli->cl_avail_grant, ocw->ocw_rc);
1630
1631                 wake_up(&ocw->ocw_waitq);
1632         }
1633
1634         EXIT;
1635 }
1636
1637 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1638 {
1639         int hprpc = !!list_empty(&osc->oo_hp_exts);
1640         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1641 }
1642
1643 /* This maintains the lists of pending pages to read/write for a given object
1644  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1645  * to quickly find objects that are ready to send an RPC. */
1646 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1647                          int cmd)
1648 {
1649         int invalid_import = 0;
1650         ENTRY;
1651
1652         /* if we have an invalid import we want to drain the queued pages
1653          * by forcing them through rpcs that immediately fail and complete
1654          * the pages.  recovery relies on this to empty the queued pages
1655          * before canceling the locks and evicting down the llite pages */
1656         if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1657                 invalid_import = 1;
1658
1659         if (cmd & OBD_BRW_WRITE) {
1660                 if (atomic_read(&osc->oo_nr_writes) == 0)
1661                         RETURN(0);
1662                 if (invalid_import) {
1663                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1664                         RETURN(1);
1665                 }
1666                 if (!list_empty(&osc->oo_hp_exts)) {
1667                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1668                         RETURN(1);
1669                 }
1670                 if (!list_empty(&osc->oo_urgent_exts)) {
1671                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1672                         RETURN(1);
1673                 }
1674                 /* trigger a write rpc stream as long as there are dirtiers
1675                  * waiting for space.  as they're waiting, they're not going to
1676                  * create more pages to coalesce with what's waiting.. */
1677                 if (!list_empty(&cli->cl_cache_waiters)) {
1678                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1679                         RETURN(1);
1680                 }
1681                 if (atomic_read(&osc->oo_nr_writes) >=
1682                     cli->cl_max_pages_per_rpc)
1683                         RETURN(1);
1684         } else {
1685                 if (atomic_read(&osc->oo_nr_reads) == 0)
1686                         RETURN(0);
1687                 if (invalid_import) {
1688                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1689                         RETURN(1);
1690                 }
1691                 /* all read are urgent. */
1692                 if (!list_empty(&osc->oo_reading_exts))
1693                         RETURN(1);
1694         }
1695
1696         RETURN(0);
1697 }
1698
1699 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1700 {
1701         struct client_obd *cli = osc_cli(obj);
1702         if (cmd & OBD_BRW_WRITE) {
1703                 atomic_add(delta, &obj->oo_nr_writes);
1704                 atomic_add(delta, &cli->cl_pending_w_pages);
1705                 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1706         } else {
1707                 atomic_add(delta, &obj->oo_nr_reads);
1708                 atomic_add(delta, &cli->cl_pending_r_pages);
1709                 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1710         }
1711         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1712 }
1713
1714 static int osc_makes_hprpc(struct osc_object *obj)
1715 {
1716         return !list_empty(&obj->oo_hp_exts);
1717 }
1718
1719 static void on_list(struct list_head *item, struct list_head *list,
1720                     int should_be_on)
1721 {
1722         if (list_empty(item) && should_be_on)
1723                 list_add_tail(item, list);
1724         else if (!list_empty(item) && !should_be_on)
1725                 list_del_init(item);
1726 }
1727
1728 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1729  * can find pages to build into rpcs quickly */
1730 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1731 {
1732         if (osc_makes_hprpc(osc)) {
1733                 /* HP rpc */
1734                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1735                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1736         } else {
1737                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1738                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1739                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1740                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1741         }
1742
1743         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1744                 atomic_read(&osc->oo_nr_writes) > 0);
1745
1746         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1747                 atomic_read(&osc->oo_nr_reads) > 0);
1748
1749         return osc_is_ready(osc);
1750 }
1751
1752 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1753 {
1754         int is_ready;
1755
1756         client_obd_list_lock(&cli->cl_loi_list_lock);
1757         is_ready = __osc_list_maint(cli, osc);
1758         client_obd_list_unlock(&cli->cl_loi_list_lock);
1759
1760         return is_ready;
1761 }
1762
1763 /* this is trying to propogate async writeback errors back up to the
1764  * application.  As an async write fails we record the error code for later if
1765  * the app does an fsync.  As long as errors persist we force future rpcs to be
1766  * sync so that the app can get a sync error and break the cycle of queueing
1767  * pages for which writeback will fail. */
1768 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1769                            int rc)
1770 {
1771         if (rc) {
1772                 if (!ar->ar_rc)
1773                         ar->ar_rc = rc;
1774
1775                 ar->ar_force_sync = 1;
1776                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1777                 return;
1778
1779         }
1780
1781         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1782                 ar->ar_force_sync = 0;
1783 }
1784
1785 /* Performs "unstable" page accounting. This function balances the
1786  * increment operations performed in osc_inc_unstable_pages. It is
1787  * registered as the RPC request callback, and is executed when the
1788  * bulk RPC is committed on the server. Thus at this point, the pages
1789  * involved in the bulk transfer are no longer considered unstable. */
1790 void osc_dec_unstable_pages(struct ptlrpc_request *req)
1791 {
1792         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
1793         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
1794         int                      page_count = desc->bd_iov_count;
1795         int i;
1796
1797         /* No unstable page tracking */
1798         if (cli->cl_cache == NULL)
1799                 return;
1800
1801         LASSERT(page_count >= 0);
1802
1803         for (i = 0; i < page_count; i++)
1804                 dec_zone_page_state(desc->bd_iov[i].kiov_page, NR_UNSTABLE_NFS);
1805
1806         atomic_sub(page_count, &cli->cl_cache->ccc_unstable_nr);
1807         LASSERT(atomic_read(&cli->cl_cache->ccc_unstable_nr) >= 0);
1808
1809         atomic_sub(page_count, &cli->cl_unstable_count);
1810         LASSERT(atomic_read(&cli->cl_unstable_count) >= 0);
1811
1812         atomic_sub(page_count, &obd_unstable_pages);
1813         LASSERT(atomic_read(&obd_unstable_pages) >= 0);
1814
1815         wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
1816 }
1817
1818 /* "unstable" page accounting. See: osc_dec_unstable_pages. */
1819 void osc_inc_unstable_pages(struct ptlrpc_request *req)
1820 {
1821         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
1822         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
1823         int                      page_count = desc->bd_iov_count;
1824         int i;
1825
1826         /* No unstable page tracking */
1827         if (cli->cl_cache == NULL)
1828                 return;
1829
1830         LASSERT(page_count >= 0);
1831
1832         for (i = 0; i < page_count; i++)
1833                 inc_zone_page_state(desc->bd_iov[i].kiov_page, NR_UNSTABLE_NFS);
1834
1835         LASSERT(atomic_read(&cli->cl_cache->ccc_unstable_nr) >= 0);
1836         atomic_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1837
1838         LASSERT(atomic_read(&cli->cl_unstable_count) >= 0);
1839         atomic_add(page_count, &cli->cl_unstable_count);
1840
1841         LASSERT(atomic_read(&obd_unstable_pages) >= 0);
1842         atomic_add(page_count, &obd_unstable_pages);
1843
1844         /* If the request has already been committed (i.e. brw_commit
1845          * called via rq_commit_cb), we need to undo the unstable page
1846          * increments we just performed because rq_commit_cb wont be
1847          * called again. */
1848         spin_lock(&req->rq_lock);
1849         if (unlikely(req->rq_committed)) {
1850                 spin_unlock(&req->rq_lock);
1851
1852                 osc_dec_unstable_pages(req);
1853         } else {
1854                 req->rq_unstable = 1;
1855                 spin_unlock(&req->rq_lock);
1856         }
1857 }
1858
1859 /* this must be called holding the loi list lock to give coverage to exit_cache,
1860  * async_flag maintenance, and oap_request */
1861 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1862                               struct osc_async_page *oap, int sent, int rc)
1863 {
1864         struct osc_object *osc = oap->oap_obj;
1865         struct lov_oinfo  *loi = osc->oo_oinfo;
1866         __u64 xid = 0;
1867
1868         ENTRY;
1869         if (oap->oap_request != NULL) {
1870                 if (rc == 0)
1871                         osc_inc_unstable_pages(oap->oap_request);
1872
1873                 xid = ptlrpc_req_xid(oap->oap_request);
1874                 ptlrpc_req_finished(oap->oap_request);
1875                 oap->oap_request = NULL;
1876         }
1877
1878         /* As the transfer for this page is being done, clear the flags */
1879         spin_lock(&oap->oap_lock);
1880         oap->oap_async_flags = 0;
1881         spin_unlock(&oap->oap_lock);
1882         oap->oap_interrupted = 0;
1883
1884         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1885                 client_obd_list_lock(&cli->cl_loi_list_lock);
1886                 osc_process_ar(&cli->cl_ar, xid, rc);
1887                 osc_process_ar(&loi->loi_ar, xid, rc);
1888                 client_obd_list_unlock(&cli->cl_loi_list_lock);
1889         }
1890
1891         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1892         if (rc)
1893                 CERROR("completion on oap %p obj %p returns %d.\n",
1894                        oap, osc, rc);
1895
1896         EXIT;
1897 }
1898
1899 /**
1900  * Try to add extent to one RPC. We need to think about the following things:
1901  * - # of pages must not be over max_pages_per_rpc
1902  * - extent must be compatible with previous ones
1903  */
1904 static int try_to_add_extent_for_io(struct client_obd *cli,
1905                                     struct osc_extent *ext,
1906                                     struct list_head *rpclist,
1907                                     int *pc, unsigned int *max_pages)
1908 {
1909         struct osc_extent *tmp;
1910         ENTRY;
1911
1912         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1913                 ext);
1914
1915         *max_pages = max(ext->oe_mppr, *max_pages);
1916         if (*pc + ext->oe_nr_pages > *max_pages)
1917                 RETURN(0);
1918
1919         list_for_each_entry(tmp, rpclist, oe_link) {
1920                 EASSERT(tmp->oe_owner == current, tmp);
1921 #if 0
1922                 if (overlapped(tmp, ext)) {
1923                         OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
1924                         EASSERT(0, ext);
1925                 }
1926 #endif
1927
1928                 if (tmp->oe_srvlock != ext->oe_srvlock ||
1929                     !tmp->oe_grants != !ext->oe_grants)
1930                         RETURN(0);
1931
1932                 /* remove break for strict check */
1933                 break;
1934         }
1935
1936         *pc += ext->oe_nr_pages;
1937         list_move_tail(&ext->oe_link, rpclist);
1938         ext->oe_owner = current;
1939         RETURN(1);
1940 }
1941
1942 /**
1943  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1944  * get_write_extent() takes all appropriate extents in atomic.
1945  *
1946  * The following policy is used to collect extents for IO:
1947  * 1. Add as many HP extents as possible;
1948  * 2. Add the first urgent extent in urgent extent list and take it out of
1949  *    urgent list;
1950  * 3. Add subsequent extents of this urgent extent;
1951  * 4. If urgent list is not empty, goto 2;
1952  * 5. Traverse the extent tree from the 1st extent;
1953  * 6. Above steps exit if there is no space in this RPC.
1954  */
1955 static int get_write_extents(struct osc_object *obj, struct list_head *rpclist)
1956 {
1957         struct client_obd *cli = osc_cli(obj);
1958         struct osc_extent *ext;
1959         int page_count = 0;
1960         unsigned int max_pages = cli->cl_max_pages_per_rpc;
1961
1962         LASSERT(osc_object_is_locked(obj));
1963         while (!list_empty(&obj->oo_hp_exts)) {
1964                 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
1965                                  oe_link);
1966                 LASSERT(ext->oe_state == OES_CACHE);
1967                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1968                                               &max_pages))
1969                         return page_count;
1970                 EASSERT(ext->oe_nr_pages <= max_pages, ext);
1971         }
1972         if (page_count == max_pages)
1973                 return page_count;
1974
1975         while (!list_empty(&obj->oo_urgent_exts)) {
1976                 ext = list_entry(obj->oo_urgent_exts.next,
1977                                  struct osc_extent, oe_link);
1978                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1979                                               &max_pages))
1980                         return page_count;
1981
1982                 if (!ext->oe_intree)
1983                         continue;
1984
1985                 while ((ext = next_extent(ext)) != NULL) {
1986                         if ((ext->oe_state != OES_CACHE) ||
1987                             (!list_empty(&ext->oe_link) &&
1988                              ext->oe_owner != NULL))
1989                                 continue;
1990
1991                         if (!try_to_add_extent_for_io(cli, ext, rpclist,
1992                                                       &page_count, &max_pages))
1993                                 return page_count;
1994                 }
1995         }
1996         if (page_count == max_pages)
1997                 return page_count;
1998
1999         ext = first_extent(obj);
2000         while (ext != NULL) {
2001                 if ((ext->oe_state != OES_CACHE) ||
2002                     /* this extent may be already in current rpclist */
2003                     (!list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
2004                         ext = next_extent(ext);
2005                         continue;
2006                 }
2007
2008                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
2009                                               &max_pages))
2010                         return page_count;
2011
2012                 ext = next_extent(ext);
2013         }
2014         return page_count;
2015 }
2016
2017 static int
2018 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
2019                    struct osc_object *osc, pdl_policy_t pol)
2020 {
2021         struct list_head   rpclist = LIST_HEAD_INIT(rpclist);
2022         struct osc_extent *ext;
2023         struct osc_extent *tmp;
2024         struct osc_extent *first = NULL;
2025         obd_count page_count = 0;
2026         int srvlock = 0;
2027         int rc = 0;
2028         ENTRY;
2029
2030         LASSERT(osc_object_is_locked(osc));
2031
2032         page_count = get_write_extents(osc, &rpclist);
2033         LASSERT(equi(page_count == 0, list_empty(&rpclist)));
2034
2035         if (list_empty(&rpclist))
2036                 RETURN(0);
2037
2038         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2039
2040         list_for_each_entry(ext, &rpclist, oe_link) {
2041                 LASSERT(ext->oe_state == OES_CACHE ||
2042                         ext->oe_state == OES_LOCK_DONE);
2043                 if (ext->oe_state == OES_CACHE)
2044                         osc_extent_state_set(ext, OES_LOCKING);
2045                 else
2046                         osc_extent_state_set(ext, OES_RPC);
2047         }
2048
2049         /* we're going to grab page lock, so release object lock because
2050          * lock order is page lock -> object lock. */
2051         osc_object_unlock(osc);
2052
2053         list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2054                 if (ext->oe_state == OES_LOCKING) {
2055                         rc = osc_extent_make_ready(env, ext);
2056                         if (unlikely(rc < 0)) {
2057                                 list_del_init(&ext->oe_link);
2058                                 osc_extent_finish(env, ext, 0, rc);
2059                                 continue;
2060                         }
2061                 }
2062                 if (first == NULL) {
2063                         first = ext;
2064                         srvlock = ext->oe_srvlock;
2065                 } else {
2066                         LASSERT(srvlock == ext->oe_srvlock);
2067                 }
2068         }
2069
2070         if (!list_empty(&rpclist)) {
2071                 LASSERT(page_count > 0);
2072                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE, pol);
2073                 LASSERT(list_empty(&rpclist));
2074         }
2075
2076         osc_object_lock(osc);
2077         RETURN(rc);
2078 }
2079
2080 /**
2081  * prepare pages for ASYNC io and put pages in send queue.
2082  *
2083  * \param cmd OBD_BRW_* macroses
2084  * \param lop pending pages
2085  *
2086  * \return zero if no page added to send queue.
2087  * \return 1 if pages successfully added to send queue.
2088  * \return negative on errors.
2089  */
2090 static int
2091 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2092                   struct osc_object *osc, pdl_policy_t pol)
2093 {
2094         struct osc_extent *ext;
2095         struct osc_extent *next;
2096         struct list_head rpclist = LIST_HEAD_INIT(rpclist);
2097         int page_count = 0;
2098         unsigned int max_pages = cli->cl_max_pages_per_rpc;
2099         int rc = 0;
2100         ENTRY;
2101
2102         LASSERT(osc_object_is_locked(osc));
2103         list_for_each_entry_safe(ext, next,
2104                                      &osc->oo_reading_exts, oe_link) {
2105                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2106                 if (!try_to_add_extent_for_io(cli, ext, &rpclist, &page_count,
2107                                               &max_pages))
2108                         break;
2109                 osc_extent_state_set(ext, OES_RPC);
2110                 EASSERT(ext->oe_nr_pages <= max_pages, ext);
2111         }
2112         LASSERT(page_count <= max_pages);
2113
2114         osc_update_pending(osc, OBD_BRW_READ, -page_count);
2115
2116         if (!list_empty(&rpclist)) {
2117                 osc_object_unlock(osc);
2118
2119                 LASSERT(page_count > 0);
2120                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ, pol);
2121                 LASSERT(list_empty(&rpclist));
2122
2123                 osc_object_lock(osc);
2124         }
2125         RETURN(rc);
2126 }
2127
2128 #define list_to_obj(list, item) ({                                            \
2129         struct list_head *__tmp = (list)->next;                               \
2130         list_del_init(__tmp);                                         \
2131         list_entry(__tmp, struct osc_object, oo_##item);                      \
2132 })
2133
2134 /* This is called by osc_check_rpcs() to find which objects have pages that
2135  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2136 static struct osc_object *osc_next_obj(struct client_obd *cli)
2137 {
2138         ENTRY;
2139
2140         /* First return objects that have blocked locks so that they
2141          * will be flushed quickly and other clients can get the lock,
2142          * then objects which have pages ready to be stuffed into RPCs */
2143         if (!list_empty(&cli->cl_loi_hp_ready_list))
2144                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2145         if (!list_empty(&cli->cl_loi_ready_list))
2146                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2147
2148         /* then if we have cache waiters, return all objects with queued
2149          * writes.  This is especially important when many small files
2150          * have filled up the cache and not been fired into rpcs because
2151          * they don't pass the nr_pending/object threshhold */
2152         if (!list_empty(&cli->cl_cache_waiters) &&
2153             !list_empty(&cli->cl_loi_write_list))
2154                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2155
2156         /* then return all queued objects when we have an invalid import
2157          * so that they get flushed */
2158         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2159                 if (!list_empty(&cli->cl_loi_write_list))
2160                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2161                                            write_item));
2162                 if (!list_empty(&cli->cl_loi_read_list))
2163                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2164                                            read_item));
2165         }
2166         RETURN(NULL);
2167 }
2168
2169 /* called with the loi list lock held */
2170 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli,
2171                            pdl_policy_t pol)
2172 {
2173         struct osc_object *osc;
2174         int rc = 0;
2175         ENTRY;
2176
2177         while ((osc = osc_next_obj(cli)) != NULL) {
2178                 struct cl_object *obj = osc2cl(osc);
2179                 struct lu_ref_link link;
2180
2181                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2182
2183                 if (osc_max_rpc_in_flight(cli, osc)) {
2184                         __osc_list_maint(cli, osc);
2185                         break;
2186                 }
2187
2188                 cl_object_get(obj);
2189                 client_obd_list_unlock(&cli->cl_loi_list_lock);
2190                 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2191
2192                 /* attempt some read/write balancing by alternating between
2193                  * reads and writes in an object.  The makes_rpc checks here
2194                  * would be redundant if we were getting read/write work items
2195                  * instead of objects.  we don't want send_oap_rpc to drain a
2196                  * partial read pending queue when we're given this object to
2197                  * do io on writes while there are cache waiters */
2198                 osc_object_lock(osc);
2199                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2200                         rc = osc_send_write_rpc(env, cli, osc, pol);
2201                         if (rc < 0) {
2202                                 CERROR("Write request failed with %d\n", rc);
2203
2204                                 /* osc_send_write_rpc failed, mostly because of
2205                                  * memory pressure.
2206                                  *
2207                                  * It can't break here, because if:
2208                                  *  - a page was submitted by osc_io_submit, so
2209                                  *    page locked;
2210                                  *  - no request in flight
2211                                  *  - no subsequent request
2212                                  * The system will be in live-lock state,
2213                                  * because there is no chance to call
2214                                  * osc_io_unplug() and osc_check_rpcs() any
2215                                  * more. pdflush can't help in this case,
2216                                  * because it might be blocked at grabbing
2217                                  * the page lock as we mentioned.
2218                                  *
2219                                  * Anyway, continue to drain pages. */
2220                                 /* break; */
2221                         }
2222                 }
2223                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2224                         rc = osc_send_read_rpc(env, cli, osc, pol);
2225                         if (rc < 0)
2226                                 CERROR("Read request failed with %d\n", rc);
2227                 }
2228                 osc_object_unlock(osc);
2229
2230                 osc_list_maint(cli, osc);
2231                 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2232                 cl_object_put(env, obj);
2233
2234                 client_obd_list_lock(&cli->cl_loi_list_lock);
2235         }
2236 }
2237
2238 static int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2239                           struct osc_object *osc, pdl_policy_t pol, int async)
2240 {
2241         int rc = 0;
2242
2243         if (osc != NULL && osc_list_maint(cli, osc) == 0)
2244                 return 0;
2245
2246         if (!async) {
2247                 /* disable osc_lru_shrink() temporarily to avoid
2248                  * potential stack overrun problem. LU-2859 */
2249                 atomic_inc(&cli->cl_lru_shrinkers);
2250                 client_obd_list_lock(&cli->cl_loi_list_lock);
2251                 osc_check_rpcs(env, cli, pol);
2252                 client_obd_list_unlock(&cli->cl_loi_list_lock);
2253                 atomic_dec(&cli->cl_lru_shrinkers);
2254         } else {
2255                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2256                 LASSERT(cli->cl_writeback_work != NULL);
2257                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2258         }
2259         return rc;
2260 }
2261
2262 static int osc_io_unplug_async(const struct lu_env *env,
2263                                 struct client_obd *cli, struct osc_object *osc)
2264 {
2265         /* XXX: policy is no use actually. */
2266         return osc_io_unplug0(env, cli, osc, PDL_POLICY_ROUND, 1);
2267 }
2268
2269 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
2270                    struct osc_object *osc, pdl_policy_t pol)
2271 {
2272         (void)osc_io_unplug0(env, cli, osc, pol, 0);
2273 }
2274
2275 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2276                         struct page *page, loff_t offset)
2277 {
2278         struct obd_export     *exp = osc_export(osc);
2279         struct osc_async_page *oap = &ops->ops_oap;
2280         ENTRY;
2281
2282         if (!page)
2283                 return cfs_size_round(sizeof(*oap));
2284
2285         oap->oap_magic = OAP_MAGIC;
2286         oap->oap_cli = &exp->exp_obd->u.cli;
2287         oap->oap_obj = osc;
2288
2289         oap->oap_page = page;
2290         oap->oap_obj_off = offset;
2291         LASSERT(!(offset & ~CFS_PAGE_MASK));
2292
2293         if (!client_is_remote(exp) && cfs_capable(CFS_CAP_SYS_RESOURCE))
2294                 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2295
2296         INIT_LIST_HEAD(&oap->oap_pending_item);
2297         INIT_LIST_HEAD(&oap->oap_rpc_item);
2298
2299         spin_lock_init(&oap->oap_lock);
2300         CDEBUG(D_INFO, "oap %p page %p obj off "LPU64"\n",
2301                oap, page, oap->oap_obj_off);
2302         RETURN(0);
2303 }
2304
2305 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2306                        struct osc_page *ops)
2307 {
2308         struct osc_io *oio = osc_env_io(env);
2309         struct osc_extent     *ext = NULL;
2310         struct osc_async_page *oap = &ops->ops_oap;
2311         struct client_obd     *cli = oap->oap_cli;
2312         struct osc_object     *osc = oap->oap_obj;
2313         pgoff_t index;
2314         int    grants = 0;
2315         int    brw_flags = OBD_BRW_ASYNC;
2316         int    cmd = OBD_BRW_WRITE;
2317         int    need_release = 0;
2318         int    rc = 0;
2319         ENTRY;
2320
2321         if (oap->oap_magic != OAP_MAGIC)
2322                 RETURN(-EINVAL);
2323
2324         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2325                 RETURN(-EIO);
2326
2327         if (!list_empty(&oap->oap_pending_item) ||
2328             !list_empty(&oap->oap_rpc_item))
2329                 RETURN(-EBUSY);
2330
2331         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2332         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2333         if (!client_is_remote(osc_export(osc)) &&
2334             cfs_capable(CFS_CAP_SYS_RESOURCE)) {
2335                 brw_flags |= OBD_BRW_NOQUOTA;
2336                 cmd |= OBD_BRW_NOQUOTA;
2337         }
2338
2339         /* check if the file's owner/group is over quota */
2340         if (!(cmd & OBD_BRW_NOQUOTA)) {
2341                 struct cl_object *obj;
2342                 struct cl_attr   *attr;
2343                 unsigned int qid[MAXQUOTAS];
2344
2345                 obj = cl_object_top(&osc->oo_cl);
2346                 attr = &osc_env_info(env)->oti_attr;
2347
2348                 cl_object_attr_lock(obj);
2349                 rc = cl_object_attr_get(env, obj, attr);
2350                 cl_object_attr_unlock(obj);
2351
2352                 qid[USRQUOTA] = attr->cat_uid;
2353                 qid[GRPQUOTA] = attr->cat_gid;
2354                 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2355                         rc = -EDQUOT;
2356                 if (rc)
2357                         RETURN(rc);
2358         }
2359
2360         if (osc_over_unstable_soft_limit(cli))
2361                 brw_flags |= OBD_BRW_SOFT_SYNC;
2362
2363         oap->oap_cmd = cmd;
2364         oap->oap_page_off = ops->ops_from;
2365         oap->oap_count = ops->ops_to - ops->ops_from;
2366         /* No need to hold a lock here,
2367          * since this page is not in any list yet. */
2368         oap->oap_async_flags = 0;
2369         oap->oap_brw_flags = brw_flags;
2370
2371         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2372                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2373
2374         index = osc_index(oap2osc(oap));
2375
2376         /* Add this page into extent by the following steps:
2377          * 1. if there exists an active extent for this IO, mostly this page
2378          *    can be added to the active extent and sometimes we need to
2379          *    expand extent to accomodate this page;
2380          * 2. otherwise, a new extent will be allocated. */
2381
2382         ext = oio->oi_active;
2383         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2384                 /* one chunk plus extent overhead must be enough to write this
2385                  * page */
2386                 grants = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2387                 if (ext->oe_end >= index)
2388                         grants = 0;
2389
2390                 /* it doesn't need any grant to dirty this page */
2391                 client_obd_list_lock(&cli->cl_loi_list_lock);
2392                 rc = osc_enter_cache_try(cli, oap, grants, 0);
2393                 client_obd_list_unlock(&cli->cl_loi_list_lock);
2394                 if (rc == 0) { /* try failed */
2395                         grants = 0;
2396                         need_release = 1;
2397                 } else if (ext->oe_end < index) {
2398                         int tmp = grants;
2399                         /* try to expand this extent */
2400                         rc = osc_extent_expand(ext, index, &tmp);
2401                         if (rc < 0) {
2402                                 need_release = 1;
2403                                 /* don't free reserved grant */
2404                         } else {
2405                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2406                                                 "expanded for %lu.\n", index);
2407                                 osc_unreserve_grant(cli, grants, tmp);
2408                                 grants = 0;
2409                         }
2410                 }
2411                 rc = 0;
2412         } else if (ext != NULL) {
2413                 /* index is located outside of active extent */
2414                 need_release = 1;
2415         }
2416         if (need_release) {
2417                 osc_extent_release(env, ext);
2418                 oio->oi_active = NULL;
2419                 ext = NULL;
2420         }
2421
2422         if (ext == NULL) {
2423                 int tmp = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2424
2425                 /* try to find new extent to cover this page */
2426                 LASSERT(oio->oi_active == NULL);
2427                 /* we may have allocated grant for this page if we failed
2428                  * to expand the previous active extent. */
2429                 LASSERT(ergo(grants > 0, grants >= tmp));
2430
2431                 rc = 0;
2432                 if (grants == 0) {
2433                         /* we haven't allocated grant for this page. */
2434                         rc = osc_enter_cache(env, cli, oap, tmp);
2435                         if (rc == 0)
2436                                 grants = tmp;
2437                 }
2438
2439                 tmp = grants;
2440                 if (rc == 0) {
2441                         ext = osc_extent_find(env, osc, index, &tmp);
2442                         if (IS_ERR(ext)) {
2443                                 LASSERT(tmp == grants);
2444                                 osc_exit_cache(cli, oap);
2445                                 rc = PTR_ERR(ext);
2446                                 ext = NULL;
2447                         } else {
2448                                 oio->oi_active = ext;
2449                         }
2450                 }
2451                 if (grants > 0)
2452                         osc_unreserve_grant(cli, grants, tmp);
2453         }
2454
2455         LASSERT(ergo(rc == 0, ext != NULL));
2456         if (ext != NULL) {
2457                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2458                          ext, "index = %lu.\n", index);
2459                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2460
2461                 osc_object_lock(osc);
2462                 if (ext->oe_nr_pages == 0)
2463                         ext->oe_srvlock = ops->ops_srvlock;
2464                 else
2465                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2466                 ++ext->oe_nr_pages;
2467                 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2468                 osc_object_unlock(osc);
2469         }
2470         RETURN(rc);
2471 }
2472
2473 int osc_teardown_async_page(const struct lu_env *env,
2474                             struct osc_object *obj, struct osc_page *ops)
2475 {
2476         struct osc_async_page *oap = &ops->ops_oap;
2477         struct osc_extent     *ext = NULL;
2478         int rc = 0;
2479         ENTRY;
2480
2481         LASSERT(oap->oap_magic == OAP_MAGIC);
2482
2483         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2484                oap, ops, osc_index(oap2osc(oap)));
2485
2486         osc_object_lock(obj);
2487         if (!list_empty(&oap->oap_rpc_item)) {
2488                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2489                 rc = -EBUSY;
2490         } else if (!list_empty(&oap->oap_pending_item)) {
2491                 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2492                 /* only truncated pages are allowed to be taken out.
2493                  * See osc_extent_truncate() and osc_cache_truncate_start()
2494                  * for details. */
2495                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2496                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2497                                         osc_index(oap2osc(oap)));
2498                         rc = -EBUSY;
2499                 }
2500         }
2501         osc_object_unlock(obj);
2502         if (ext != NULL)
2503                 osc_extent_put(env, ext);
2504         RETURN(rc);
2505 }
2506
2507 /**
2508  * This is called when a page is picked up by kernel to write out.
2509  *
2510  * We should find out the corresponding extent and add the whole extent
2511  * into urgent list. The extent may be being truncated or used, handle it
2512  * carefully.
2513  */
2514 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2515                          struct osc_page *ops)
2516 {
2517         struct osc_extent *ext   = NULL;
2518         struct osc_object *obj   = cl2osc(ops->ops_cl.cpl_obj);
2519         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2520         pgoff_t            index = osc_index(ops);
2521         struct osc_async_page *oap = &ops->ops_oap;
2522         bool unplug = false;
2523         int rc = 0;
2524         ENTRY;
2525
2526         osc_object_lock(obj);
2527         ext = osc_extent_lookup(obj, index);
2528         if (ext == NULL) {
2529                 osc_extent_tree_dump(D_ERROR, obj);
2530                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2531         }
2532
2533         switch (ext->oe_state) {
2534         case OES_RPC:
2535         case OES_LOCK_DONE:
2536                 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2537                 LASSERT(0);
2538                 break;
2539         case OES_LOCKING:
2540                 /* If we know this extent is being written out, we should abort
2541                  * so that the writer can make this page ready. Otherwise, there
2542                  * exists a deadlock problem because other process can wait for
2543                  * page writeback bit holding page lock; and meanwhile in
2544                  * vvp_page_make_ready(), we need to grab page lock before
2545                  * really sending the RPC. */
2546         case OES_TRUNC:
2547                 /* race with truncate, page will be redirtied */
2548         case OES_ACTIVE:
2549                 /* The extent is active so we need to abort and let the caller
2550                  * re-dirty the page. If we continued on here, and we were the
2551                  * one making the extent active, we could deadlock waiting for
2552                  * the page writeback to clear but it won't because the extent
2553                  * is active and won't be written out. */
2554                 GOTO(out, rc = -EAGAIN);
2555         default:
2556                 break;
2557         }
2558
2559         rc = cl_page_prep(env, io, cp, CRT_WRITE);
2560         if (rc)
2561                 GOTO(out, rc);
2562
2563         spin_lock(&oap->oap_lock);
2564         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2565         spin_unlock(&oap->oap_lock);
2566
2567         if (memory_pressure_get())
2568                 ext->oe_memalloc = 1;
2569
2570         ext->oe_urgent = 1;
2571         if (ext->oe_state == OES_CACHE) {
2572                 OSC_EXTENT_DUMP(D_CACHE, ext,
2573                                 "flush page %p make it urgent.\n", oap);
2574                 if (list_empty(&ext->oe_link))
2575                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2576                 unplug = true;
2577         }
2578         rc = 0;
2579         EXIT;
2580
2581 out:
2582         osc_object_unlock(obj);
2583         osc_extent_put(env, ext);
2584         if (unplug)
2585                 osc_io_unplug_async(env, osc_cli(obj), obj);
2586         return rc;
2587 }
2588
2589 /**
2590  * this is called when a sync waiter receives an interruption.  Its job is to
2591  * get the caller woken as soon as possible.  If its page hasn't been put in an
2592  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2593  * desiring interruption which will forcefully complete the rpc once the rpc
2594  * has timed out.
2595  */
2596 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2597 {
2598         struct osc_async_page *oap = &ops->ops_oap;
2599         struct osc_object     *obj = oap->oap_obj;
2600         struct client_obd     *cli = osc_cli(obj);
2601         struct osc_extent     *ext;
2602         struct osc_extent     *found = NULL;
2603         struct list_head            *plist;
2604         pgoff_t index = osc_index(ops);
2605         int     rc = -EBUSY;
2606         int     cmd;
2607         ENTRY;
2608
2609         LASSERT(!oap->oap_interrupted);
2610         oap->oap_interrupted = 1;
2611
2612         /* Find out the caching extent */
2613         osc_object_lock(obj);
2614         if (oap->oap_cmd & OBD_BRW_WRITE) {
2615                 plist = &obj->oo_urgent_exts;
2616                 cmd   = OBD_BRW_WRITE;
2617         } else {
2618                 plist = &obj->oo_reading_exts;
2619                 cmd   = OBD_BRW_READ;
2620         }
2621         list_for_each_entry(ext, plist, oe_link) {
2622                 if (ext->oe_start <= index && ext->oe_end >= index) {
2623                         LASSERT(ext->oe_state == OES_LOCK_DONE);
2624                         /* For OES_LOCK_DONE state extent, it has already held
2625                          * a refcount for RPC. */
2626                         found = osc_extent_get(ext);
2627                         break;
2628                 }
2629         }
2630         if (found != NULL) {
2631                 list_del_init(&found->oe_link);
2632                 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2633                 osc_object_unlock(obj);
2634
2635                 osc_extent_finish(env, found, 0, -EINTR);
2636                 osc_extent_put(env, found);
2637                 rc = 0;
2638         } else {
2639                 osc_object_unlock(obj);
2640                 /* ok, it's been put in an rpc. only one oap gets a request
2641                  * reference */
2642                 if (oap->oap_request != NULL) {
2643                         ptlrpc_mark_interrupted(oap->oap_request);
2644                         ptlrpcd_wake(oap->oap_request);
2645                         ptlrpc_req_finished(oap->oap_request);
2646                         oap->oap_request = NULL;
2647                 }
2648         }
2649
2650         osc_list_maint(cli, obj);
2651         RETURN(rc);
2652 }
2653
2654 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
2655                          struct list_head *list, int cmd, int brw_flags)
2656 {
2657         struct client_obd     *cli = osc_cli(obj);
2658         struct osc_extent     *ext;
2659         struct osc_async_page *oap;
2660         int     page_count = 0;
2661         int     mppr       = cli->cl_max_pages_per_rpc;
2662         pgoff_t start      = CL_PAGE_EOF;
2663         pgoff_t end        = 0;
2664         ENTRY;
2665
2666         list_for_each_entry(oap, list, oap_pending_item) {
2667                 pgoff_t index = osc_index(oap2osc(oap));
2668                 if (index > end)
2669                         end = index;
2670                 if (index < start)
2671                         start = index;
2672                 ++page_count;
2673                 mppr <<= (page_count > mppr);
2674         }
2675
2676         ext = osc_extent_alloc(obj);
2677         if (ext == NULL) {
2678                 list_for_each_entry(oap, list, oap_pending_item) {
2679                         list_del_init(&oap->oap_pending_item);
2680                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2681                 }
2682                 RETURN(-ENOMEM);
2683         }
2684
2685         ext->oe_rw = !!(cmd & OBD_BRW_READ);
2686         ext->oe_urgent = 1;
2687         ext->oe_start = start;
2688         ext->oe_end = ext->oe_max_end = end;
2689         ext->oe_obj = obj;
2690         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2691         ext->oe_nr_pages = page_count;
2692         ext->oe_mppr = mppr;
2693         list_splice_init(list, &ext->oe_pages);
2694
2695         osc_object_lock(obj);
2696         /* Reuse the initial refcount for RPC, don't drop it */
2697         osc_extent_state_set(ext, OES_LOCK_DONE);
2698         if (cmd & OBD_BRW_WRITE) {
2699                 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2700                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2701         } else {
2702                 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2703                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2704         }
2705         osc_object_unlock(obj);
2706
2707         osc_io_unplug(env, cli, obj, PDL_POLICY_ROUND);
2708         RETURN(0);
2709 }
2710
2711 /**
2712  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2713  */
2714 int osc_cache_truncate_start(const struct lu_env *env, struct osc_io *oio,
2715                              struct osc_object *obj, __u64 size)
2716 {
2717         struct client_obd *cli = osc_cli(obj);
2718         struct osc_extent *ext;
2719         struct osc_extent *waiting = NULL;
2720         pgoff_t index;
2721         struct list_head list = LIST_HEAD_INIT(list);
2722         int result = 0;
2723         bool partial;
2724         ENTRY;
2725
2726         /* pages with index greater or equal to index will be truncated. */
2727         index = cl_index(osc2cl(obj), size);
2728         partial = size > cl_offset(osc2cl(obj), index);
2729
2730 again:
2731         osc_object_lock(obj);
2732         ext = osc_extent_search(obj, index);
2733         if (ext == NULL)
2734                 ext = first_extent(obj);
2735         else if (ext->oe_end < index)
2736                 ext = next_extent(ext);
2737         while (ext != NULL) {
2738                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2739
2740                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2741                         /* if ext is in urgent state, it means there must exist
2742                          * a page already having been flushed by write_page().
2743                          * We have to wait for this extent because we can't
2744                          * truncate that page. */
2745                         LASSERT(!ext->oe_hp);
2746                         OSC_EXTENT_DUMP(D_CACHE, ext,
2747                                         "waiting for busy extent\n");
2748                         waiting = osc_extent_get(ext);
2749                         break;
2750                 }
2751
2752                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:"LPU64".\n", size);
2753
2754                 osc_extent_get(ext);
2755                 if (ext->oe_state == OES_ACTIVE) {
2756                         /* though we grab inode mutex for write path, but we
2757                          * release it before releasing extent(in osc_io_end()),
2758                          * so there is a race window that an extent is still
2759                          * in OES_ACTIVE when truncate starts. */
2760                         LASSERT(!ext->oe_trunc_pending);
2761                         ext->oe_trunc_pending = 1;
2762                 } else {
2763                         EASSERT(ext->oe_state == OES_CACHE, ext);
2764                         osc_extent_state_set(ext, OES_TRUNC);
2765                         osc_update_pending(obj, OBD_BRW_WRITE,
2766                                            -ext->oe_nr_pages);
2767                 }
2768                 EASSERT(list_empty(&ext->oe_link), ext);
2769                 list_add_tail(&ext->oe_link, &list);
2770
2771                 ext = next_extent(ext);
2772         }
2773         osc_object_unlock(obj);
2774
2775         osc_list_maint(cli, obj);
2776
2777         while (!list_empty(&list)) {
2778                 int rc;
2779
2780                 ext = list_entry(list.next, struct osc_extent, oe_link);
2781                 list_del_init(&ext->oe_link);
2782
2783                 /* extent may be in OES_ACTIVE state because inode mutex
2784                  * is released before osc_io_end() in file write case */
2785                 if (ext->oe_state != OES_TRUNC)
2786                         osc_extent_wait(env, ext, OES_TRUNC);
2787
2788                 rc = osc_extent_truncate(ext, index, partial);
2789                 if (rc < 0) {
2790                         if (result == 0)
2791                                 result = rc;
2792
2793                         OSC_EXTENT_DUMP(D_ERROR, ext,
2794                                         "truncate error %d\n", rc);
2795                 } else if (ext->oe_nr_pages == 0) {
2796                         osc_extent_remove(ext);
2797                 } else {
2798                         /* this must be an overlapped extent which means only
2799                          * part of pages in this extent have been truncated.
2800                          */
2801                         EASSERTF(ext->oe_start <= index, ext,
2802                                  "trunc index = %lu/%d.\n", index, partial);
2803                         /* fix index to skip this partially truncated extent */
2804                         index = ext->oe_end + 1;
2805                         partial = false;
2806
2807                         /* we need to hold this extent in OES_TRUNC state so
2808                          * that no writeback will happen. This is to avoid
2809                          * BUG 17397. */
2810                         LASSERT(oio->oi_trunc == NULL);
2811                         oio->oi_trunc = osc_extent_get(ext);
2812                         OSC_EXTENT_DUMP(D_CACHE, ext,
2813                                         "trunc at "LPU64"\n", size);
2814                 }
2815                 osc_extent_put(env, ext);
2816         }
2817         if (waiting != NULL) {
2818                 int rc;
2819
2820                 /* ignore the result of osc_extent_wait the write initiator
2821                  * should take care of it. */
2822                 rc = osc_extent_wait(env, waiting, OES_INV);
2823                 if (rc < 0)
2824                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2825
2826                 osc_extent_put(env, waiting);
2827                 waiting = NULL;
2828                 goto again;
2829         }
2830         RETURN(result);
2831 }
2832
2833 /**
2834  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2835  */
2836 void osc_cache_truncate_end(const struct lu_env *env, struct osc_io *oio,
2837                             struct osc_object *obj)
2838 {
2839         struct osc_extent *ext = oio->oi_trunc;
2840
2841         oio->oi_trunc = NULL;
2842         if (ext != NULL) {
2843                 bool unplug = false;
2844
2845                 EASSERT(ext->oe_nr_pages > 0, ext);
2846                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2847                 EASSERT(!ext->oe_urgent, ext);
2848
2849                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2850                 osc_object_lock(obj);
2851                 osc_extent_state_set(ext, OES_CACHE);
2852                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2853                         ext->oe_urgent = 1;
2854                         list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2855                         unplug = true;
2856                 }
2857                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2858                 osc_object_unlock(obj);
2859                 osc_extent_put(env, ext);
2860
2861                 if (unplug)
2862                         osc_io_unplug_async(env, osc_cli(obj), obj);
2863         }
2864 }
2865
2866 /**
2867  * Wait for extents in a specific range to be written out.
2868  * The caller must have called osc_cache_writeback_range() to issue IO
2869  * otherwise it will take a long time for this function to finish.
2870  *
2871  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2872  * nobody else can dirty this range of file while we're waiting for
2873  * extents to be written.
2874  */
2875 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2876                          pgoff_t start, pgoff_t end)
2877 {
2878         struct osc_extent *ext;
2879         pgoff_t index = start;
2880         int     result = 0;
2881         ENTRY;
2882
2883 again:
2884         osc_object_lock(obj);
2885         ext = osc_extent_search(obj, index);
2886         if (ext == NULL)
2887                 ext = first_extent(obj);
2888         else if (ext->oe_end < index)
2889                 ext = next_extent(ext);
2890         while (ext != NULL) {
2891                 int rc;
2892
2893                 if (ext->oe_start > end)
2894                         break;
2895
2896                 if (!ext->oe_fsync_wait) {
2897                         ext = next_extent(ext);
2898                         continue;
2899                 }
2900
2901                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2902                              ext->oe_hp || ext->oe_urgent), ext);
2903                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2904                              !ext->oe_hp && ext->oe_urgent), ext);
2905
2906                 index = ext->oe_end + 1;
2907                 osc_extent_get(ext);
2908                 osc_object_unlock(obj);
2909
2910                 rc = osc_extent_wait(env, ext, OES_INV);
2911                 if (result == 0)
2912                         result = rc;
2913                 osc_extent_put(env, ext);
2914                 goto again;
2915         }
2916         osc_object_unlock(obj);
2917
2918         OSC_IO_DEBUG(obj, "sync file range.\n");
2919         RETURN(result);
2920 }
2921
2922 /**
2923  * Called to write out a range of osc object.
2924  *
2925  * @hp     : should be set this is caused by lock cancel;
2926  * @discard: is set if dirty pages should be dropped - file will be deleted or
2927  *         truncated, this implies there is no partially discarding extents.
2928  *
2929  * Return how many pages will be issued, or error code if error occurred.
2930  */
2931 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2932                               pgoff_t start, pgoff_t end, int hp, int discard)
2933 {
2934         struct osc_extent *ext;
2935         struct list_head discard_list = LIST_HEAD_INIT(discard_list);
2936         bool unplug = false;
2937         int result = 0;
2938         ENTRY;
2939
2940         osc_object_lock(obj);
2941         ext = osc_extent_search(obj, start);
2942         if (ext == NULL)
2943                 ext = first_extent(obj);
2944         else if (ext->oe_end < start)
2945                 ext = next_extent(ext);
2946         while (ext != NULL) {
2947                 if (ext->oe_start > end)
2948                         break;
2949
2950                 ext->oe_fsync_wait = 1;
2951                 switch (ext->oe_state) {
2952                 case OES_CACHE:
2953                         result += ext->oe_nr_pages;
2954                         if (!discard) {
2955                                 struct list_head *list = NULL;
2956                                 if (hp) {
2957                                         EASSERT(!ext->oe_hp, ext);
2958                                         ext->oe_hp = 1;
2959                                         list = &obj->oo_hp_exts;
2960                                 } else if (!ext->oe_urgent) {
2961                                         ext->oe_urgent = 1;
2962                                         list = &obj->oo_urgent_exts;
2963                                 }
2964                                 if (list != NULL)
2965                                         list_move_tail(&ext->oe_link, list);
2966                                 unplug = true;
2967                         } else {
2968                                 /* the only discarder is lock cancelling, so
2969                                  * [start, end] must contain this extent */
2970                                 EASSERT(ext->oe_start >= start &&
2971                                         ext->oe_max_end <= end, ext);
2972                                 osc_extent_state_set(ext, OES_LOCKING);
2973                                 ext->oe_owner = current;
2974                                 list_move_tail(&ext->oe_link,
2975                                                    &discard_list);
2976                                 osc_update_pending(obj, OBD_BRW_WRITE,
2977                                                    -ext->oe_nr_pages);
2978                         }
2979                         break;
2980                 case OES_ACTIVE:
2981                         /* It's pretty bad to wait for ACTIVE extents, because
2982                          * we don't know how long we will wait for it to be
2983                          * flushed since it may be blocked at awaiting more
2984                          * grants. We do this for the correctness of fsync. */
2985                         LASSERT(hp == 0 && discard == 0);
2986                         ext->oe_urgent = 1;
2987                         break;
2988                 case OES_TRUNC:
2989                         /* this extent is being truncated, can't do anything
2990                          * for it now. it will be set to urgent after truncate
2991                          * is finished in osc_cache_truncate_end(). */
2992                 default:
2993                         break;
2994                 }
2995                 ext = next_extent(ext);
2996         }
2997         osc_object_unlock(obj);
2998
2999         LASSERT(ergo(!discard, list_empty(&discard_list)));
3000         if (!list_empty(&discard_list)) {
3001                 struct osc_extent *tmp;
3002                 int rc;
3003
3004                 osc_list_maint(osc_cli(obj), obj);
3005                 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
3006                         list_del_init(&ext->oe_link);
3007                         EASSERT(ext->oe_state == OES_LOCKING, ext);
3008
3009                         /* Discard caching pages. We don't actually write this
3010                          * extent out but we complete it as if we did. */
3011                         rc = osc_extent_make_ready(env, ext);
3012                         if (unlikely(rc < 0)) {
3013                                 OSC_EXTENT_DUMP(D_ERROR, ext,
3014                                                 "make_ready returned %d\n", rc);
3015                                 if (result >= 0)
3016                                         result = rc;
3017                         }
3018
3019                         /* finish the extent as if the pages were sent */
3020                         osc_extent_finish(env, ext, 0, 0);
3021                 }
3022         }
3023
3024         if (unplug)
3025                 osc_io_unplug(env, osc_cli(obj), obj, PDL_POLICY_ROUND);
3026
3027         if (hp || discard) {
3028                 int rc;
3029                 rc = osc_cache_wait_range(env, obj, start, end);
3030                 if (result >= 0 && rc < 0)
3031                         result = rc;
3032         }
3033
3034         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3035         RETURN(result);
3036 }
3037
3038 /**
3039  * Returns a list of pages by a given [start, end] of \a obj.
3040  *
3041  * \param resched If not NULL, then we give up before hogging CPU for too
3042  * long and set *resched = 1, in that case caller should implement a retry
3043  * logic.
3044  *
3045  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3046  * crucial in the face of [offset, EOF] locks.
3047  *
3048  * Return at least one page in @queue unless there is no covered page.
3049  */
3050 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3051                         struct osc_object *osc, pgoff_t start, pgoff_t end,
3052                         osc_page_gang_cbt cb, void *cbdata)
3053 {
3054         struct osc_page *ops;
3055         void            **pvec;
3056         pgoff_t         idx;
3057         unsigned int    nr;
3058         unsigned int    i;
3059         unsigned int    j;
3060         int             res = CLP_GANG_OKAY;
3061         bool            tree_lock = true;
3062         ENTRY;
3063
3064         idx = start;
3065         pvec = osc_env_info(env)->oti_pvec;
3066         spin_lock(&osc->oo_tree_lock);
3067         while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3068                                             idx, OTI_PVEC_SIZE)) > 0) {
3069                 struct cl_page *page;
3070                 bool end_of_region = false;
3071
3072                 for (i = 0, j = 0; i < nr; ++i) {
3073                         ops = pvec[i];
3074                         pvec[i] = NULL;
3075
3076                         idx = osc_index(ops);
3077                         if (idx > end) {
3078                                 end_of_region = true;
3079                                 break;
3080                         }
3081
3082                         page = ops->ops_cl.cpl_page;
3083                         LASSERT(page->cp_type == CPT_CACHEABLE);
3084                         if (page->cp_state == CPS_FREEING)
3085                                 continue;
3086
3087                         cl_page_get(page);
3088                         lu_ref_add_atomic(&page->cp_reference,
3089                                           "gang_lookup", current);
3090                         pvec[j++] = ops;
3091                 }
3092                 ++idx;
3093
3094                 /*
3095                  * Here a delicate locking dance is performed. Current thread
3096                  * holds a reference to a page, but has to own it before it
3097                  * can be placed into queue. Owning implies waiting, so
3098                  * radix-tree lock is to be released. After a wait one has to
3099                  * check that pages weren't truncated (cl_page_own() returns
3100                  * error in the latter case).
3101                  */
3102                 spin_unlock(&osc->oo_tree_lock);
3103                 tree_lock = false;
3104
3105                 for (i = 0; i < j; ++i) {
3106                         ops = pvec[i];
3107                         if (res == CLP_GANG_OKAY)
3108                                 res = (*cb)(env, io, ops, cbdata);
3109
3110                         page = ops->ops_cl.cpl_page;
3111                         lu_ref_del(&page->cp_reference, "gang_lookup", current);
3112                         cl_page_put(env, page);
3113                 }
3114                 if (nr < OTI_PVEC_SIZE || end_of_region)
3115                         break;
3116
3117                 if (res == CLP_GANG_OKAY && need_resched())
3118                         res = CLP_GANG_RESCHED;
3119                 if (res != CLP_GANG_OKAY)
3120                         break;
3121
3122                 spin_lock(&osc->oo_tree_lock);
3123                 tree_lock = true;
3124         }
3125         if (tree_lock)
3126                 spin_unlock(&osc->oo_tree_lock);
3127         RETURN(res);
3128 }
3129
3130 /**
3131  * Check if page @page is covered by an extra lock or discard it.
3132  */
3133 static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3134                                 struct osc_page *ops, void *cbdata)
3135 {
3136         struct osc_thread_info *info = osc_env_info(env);
3137         struct cl_lock *lock = cbdata;
3138         pgoff_t index;
3139
3140         index = osc_index(ops);
3141         if (index >= info->oti_fn_index) {
3142                 struct cl_lock *tmp;
3143                 struct cl_page *page = ops->ops_cl.cpl_page;
3144
3145                 /* refresh non-overlapped index */
3146                 tmp = cl_lock_at_pgoff(env, lock->cll_descr.cld_obj, index,
3147                                        lock, 1, 0);
3148                 if (tmp != NULL) {
3149                         /* Cache the first-non-overlapped index so as to skip
3150                          * all pages within [index, oti_fn_index). This
3151                          * is safe because if tmp lock is canceled, it will
3152                          * discard these pages. */
3153                         info->oti_fn_index = tmp->cll_descr.cld_end + 1;
3154                         if (tmp->cll_descr.cld_end == CL_PAGE_EOF)
3155                                 info->oti_fn_index = CL_PAGE_EOF;
3156                         cl_lock_put(env, tmp);
3157                 } else if (cl_page_own(env, io, page) == 0) {
3158                         /* discard the page */
3159                         cl_page_discard(env, io, page);
3160                         cl_page_disown(env, io, page);
3161                 } else {
3162                         LASSERT(page->cp_state == CPS_FREEING);
3163                 }
3164         }
3165
3166         info->oti_next_index = index + 1;
3167         return CLP_GANG_OKAY;
3168 }
3169
3170 static int discard_cb(const struct lu_env *env, struct cl_io *io,
3171                       struct osc_page *ops, void *cbdata)
3172 {
3173         struct osc_thread_info *info = osc_env_info(env);
3174         struct cl_lock *lock = cbdata;
3175         struct cl_page *page = ops->ops_cl.cpl_page;
3176
3177         LASSERT(lock->cll_descr.cld_mode >= CLM_WRITE);
3178
3179         /* page is top page. */
3180         info->oti_next_index = osc_index(ops) + 1;
3181         if (cl_page_own(env, io, page) == 0) {
3182                 KLASSERT(ergo(page->cp_type == CPT_CACHEABLE,
3183                               !PageDirty(cl_page_vmpage(page))));
3184
3185                 /* discard the page */
3186                 cl_page_discard(env, io, page);
3187                 cl_page_disown(env, io, page);
3188         } else {
3189                 LASSERT(page->cp_state == CPS_FREEING);
3190         }
3191
3192         return CLP_GANG_OKAY;
3193 }
3194
3195 /**
3196  * Discard pages protected by the given lock. This function traverses radix
3197  * tree to find all covering pages and discard them. If a page is being covered
3198  * by other locks, it should remain in cache.
3199  *
3200  * If error happens on any step, the process continues anyway (the reasoning
3201  * behind this being that lock cancellation cannot be delayed indefinitely).
3202  */
3203 int osc_lock_discard_pages(const struct lu_env *env, struct osc_lock *ols)
3204 {
3205         struct osc_thread_info *info = osc_env_info(env);
3206         struct cl_io *io = &info->oti_io;
3207         struct cl_object *osc = ols->ols_cl.cls_obj;
3208         struct cl_lock *lock = ols->ols_cl.cls_lock;
3209         struct cl_lock_descr *descr = &lock->cll_descr;
3210         osc_page_gang_cbt cb;
3211         int res;
3212         int result;
3213
3214         ENTRY;
3215
3216         io->ci_obj = cl_object_top(osc);
3217         io->ci_ignore_layout = 1;
3218         result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3219         if (result != 0)
3220                 GOTO(out, result);
3221
3222         cb = descr->cld_mode == CLM_READ ? check_and_discard_cb : discard_cb;
3223         info->oti_fn_index = info->oti_next_index = descr->cld_start;
3224         do {
3225                 res = osc_page_gang_lookup(env, io, cl2osc(osc),
3226                                            info->oti_next_index, descr->cld_end,
3227                                            cb, (void *)lock);
3228                 if (info->oti_next_index > descr->cld_end)
3229                         break;
3230
3231                 if (res == CLP_GANG_RESCHED)
3232                         cond_resched();
3233         } while (res != CLP_GANG_OKAY);
3234 out:
3235         cl_io_fini(env, io);
3236         RETURN(result);
3237 }
3238
3239
3240 /** @} osc */