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"\n", 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"\n",
 643                  EXTPARA(cur));
 644
 645 restart:
 646         osc_object_lock(obj);
 647         ext = osc_extent_search(obj, cur->oe_start);
 648         if (ext == NULL)
 649                 ext = first_extent(obj);
 650         while (ext != NULL) {
 651                 loff_t ext_chk_start = ext->oe_start >> ppc_bits;
 652                 loff_t ext_chk_end   = ext->oe_end   >> ppc_bits;
 653
 654                 LASSERT(sanity_check_nolock(ext) == 0);
 655                 if (chunk > ext_chk_end + 1)
 656                         break;
 657
 658                 /* if covering by different locks, no chance to match */
 659                 if (lock != ext->oe_osclock) {
 660                         EASSERTF(!overlapped(ext, cur), ext,
 661                                  EXTSTR"\n", EXTPARA(cur));
 662
 663                         ext = next_extent(ext);
 664                         continue;
 665                 }
 666
 667                 /* discontiguous chunks? */
 668                 if (chunk + 1 < ext_chk_start) {
 669                         ext = next_extent(ext);
 670                         continue;
 671                 }
 672
 673                 /* ok, from now on, ext and cur have these attrs:
 674                  * 1. covered by the same lock
 675                  * 2. contiguous at chunk level or overlapping. */
 676
 677                 if (overlapped(ext, cur)) {
 678                         /* cur is the minimum unit, so overlapping means
 679                          * full contain. */
 680                         EASSERTF((ext->oe_start <= cur->oe_start &&
 681                                   ext->oe_end >= cur->oe_end),
 682                                  ext, EXTSTR"\n", EXTPARA(cur));
 683
 684                         if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
 685                                 /* for simplicity, we wait for this extent to
 686                                  * finish before going forward. */
 687                                 conflict = osc_extent_get(ext);
 688                                 break;
 689                         }
 690
 691                         found = osc_extent_hold(ext);
 692                         break;
 693                 }
 694
 695                 /* non-overlapped extent */
 696                 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
 697                         /* we can't do anything for a non OES_CACHE extent, or
 698                          * if there is someone waiting for this extent to be
 699                          * flushed, try next one. */
 700                         ext = next_extent(ext);
 701                         continue;
 702                 }
 703
 704                 /* check if they belong to the same rpc slot before trying to
 705                  * merge. the extents are not overlapped and contiguous at
 706                  * chunk level to get here. */
 707                 if (ext->oe_max_end != max_end) {
 708                         /* if they don't belong to the same RPC slot or
 709                          * max_pages_per_rpc has ever changed, do not merge. */
 710                         ext = next_extent(ext);
 711                         continue;
 712                 }
 713
 714                 /* it's required that an extent must be contiguous at chunk
 715                  * level so that we know the whole extent is covered by grant
 716                  * (the pages in the extent are NOT required to be contiguous).
 717                  * Otherwise, it will be too much difficult to know which
 718                  * chunks have grants allocated. */
 719
 720                 /* try to do front merge - extend ext's start */
 721                 if (chunk + 1 == ext_chk_start) {
 722                         /* ext must be chunk size aligned */
 723                         EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
 724
 725                         /* pull ext's start back to cover cur */
 726                         ext->oe_start   = cur->oe_start;
 727                         ext->oe_grants += chunksize;
 728                         *grants -= chunksize;
 729
 730                         found = osc_extent_hold(ext);
 731                 } else if (chunk == ext_chk_end + 1) {
 732                         /* rear merge */
 733                         ext->oe_end     = cur->oe_end;
 734                         ext->oe_grants += chunksize;
 735                         *grants -= chunksize;
 736
 737                         /* try to merge with the next one because we just fill
 738                          * in a gap */
 739                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 740                                 /* we can save extent tax from next extent */
 741                                 *grants += cli->cl_extent_tax;
 742
 743                         found = osc_extent_hold(ext);
 744                 }
 745                 if (found != NULL)
 746                         break;
 747
 748                 ext = next_extent(ext);
 749         }
 750
 751         osc_extent_tree_dump(D_CACHE, obj);
 752         if (found != NULL) {
 753                 LASSERT(conflict == NULL);
 754                 if (!IS_ERR(found)) {
 755                         LASSERT(found->oe_osclock == cur->oe_osclock);
 756                         OSC_EXTENT_DUMP(D_CACHE, found,
 757                                         "found caching ext for %lu.\n", index);
 758                 }
 759         } else if (conflict == NULL) {
 760                 /* create a new extent */
 761                 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
 762                 cur->oe_grants = chunksize + cli->cl_extent_tax;
 763                 *grants -= cur->oe_grants;
 764                 LASSERT(*grants >= 0);
 765
 766                 cur->oe_state = OES_CACHE;
 767                 found = osc_extent_hold(cur);
 768                 osc_extent_insert(obj, cur);
 769                 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
 770                                 index, lock->cll_descr.cld_end);
 771         }
 772         osc_object_unlock(obj);
 773
 774         if (conflict != NULL) {
 775                 LASSERT(found == NULL);
 776
 777                 /* waiting for IO to finish. Please notice that it's impossible
 778                  * to be an OES_TRUNC extent. */
 779                 rc = osc_extent_wait(env, conflict, OES_INV);
 780                 osc_extent_put(env, conflict);
 781                 conflict = NULL;
 782                 if (rc < 0)
 783                         GOTO(out, found = ERR_PTR(rc));
 784
 785                 goto restart;
 786         }
 787         EXIT;
 788
 789 out:
 790         osc_extent_put(env, cur);
 791         LASSERT(*grants >= 0);
 792         return found;
 793 }
 794
 795 /**
 796  * Called when IO is finished to an extent.
 797  */
 798 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
 799                       int sent, int rc)
 800 {
 801         struct client_obd *cli = osc_cli(ext->oe_obj);
 802         struct osc_async_page *oap;
 803         struct osc_async_page *tmp;
 804         int nr_pages = ext->oe_nr_pages;
 805         int lost_grant = 0;
 806         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
 807         __u64 last_off = 0;
 808         int last_count = -1;
 809         ENTRY;
 810
 811         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
 812
 813         ext->oe_rc = rc ?: ext->oe_nr_pages;
 814         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
 815
 816         osc_lru_add_batch(cli, &ext->oe_pages);
 817         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 818                                      oap_pending_item) {
 819                 list_del_init(&oap->oap_rpc_item);
 820                 list_del_init(&oap->oap_pending_item);
 821                 if (last_off <= oap->oap_obj_off) {
 822                         last_off = oap->oap_obj_off;
 823                         last_count = oap->oap_count;
 824                 }
 825
 826                 --ext->oe_nr_pages;
 827                 osc_ap_completion(env, cli, oap, sent, rc);
 828         }
 829         EASSERT(ext->oe_nr_pages == 0, ext);
 830
 831         if (!sent) {
 832                 lost_grant = ext->oe_grants;
 833         } else if (blocksize < PAGE_CACHE_SIZE &&
 834                    last_count != PAGE_CACHE_SIZE) {
 835                 /* For short writes we shouldn't count parts of pages that
 836                  * span a whole chunk on the OST side, or our accounting goes
 837                  * wrong.  Should match the code in filter_grant_check. */
 838                 int offset = last_off & ~CFS_PAGE_MASK;
 839                 int count = last_count + (offset & (blocksize - 1));
 840                 int end = (offset + last_count) & (blocksize - 1);
 841                 if (end)
 842                         count += blocksize - end;
 843
 844                 lost_grant = PAGE_CACHE_SIZE - count;
 845         }
 846         if (ext->oe_grants > 0)
 847                 osc_free_grant(cli, nr_pages, lost_grant);
 848
 849         osc_extent_remove(ext);
 850         /* put the refcount for RPC */
 851         osc_extent_put(env, ext);
 852         RETURN(0);
 853 }
 854
 855 static int extent_wait_cb(struct osc_extent *ext, int state)
 856 {
 857         int ret;
 858
 859         osc_object_lock(ext->oe_obj);
 860         ret = ext->oe_state == state;
 861         osc_object_unlock(ext->oe_obj);
 862
 863         return ret;
 864 }
 865
 866 /**
 867  * Wait for the extent's state to become @state.
 868  */
 869 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
 870                            int state)
 871 {
 872         struct osc_object *obj = ext->oe_obj;
 873         struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
 874                                                   LWI_ON_SIGNAL_NOOP, NULL);
 875         int rc = 0;
 876         ENTRY;
 877
 878         osc_object_lock(obj);
 879         LASSERT(sanity_check_nolock(ext) == 0);
 880         /* `Kick' this extent only if the caller is waiting for it to be
 881          * written out. */
 882         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
 883                 if (ext->oe_state == OES_ACTIVE) {
 884                         ext->oe_urgent = 1;
 885                 } else if (ext->oe_state == OES_CACHE) {
 886                         ext->oe_urgent = 1;
 887                         osc_extent_hold(ext);
 888                         rc = 1;
 889                 }
 890         }
 891         osc_object_unlock(obj);
 892         if (rc == 1)
 893                 osc_extent_release(env, ext);
 894
 895         /* wait for the extent until its state becomes @state */
 896         rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
 897         if (rc == -ETIMEDOUT) {
 898                 OSC_EXTENT_DUMP(D_ERROR, ext,
 899                         "%s: wait ext to %d timedout, recovery in progress?\n",
 900                         osc_export(obj)->exp_obd->obd_name, state);
 901
 902                 lwi = LWI_INTR(NULL, NULL);
 903                 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state),
 904                                   &lwi);
 905         }
 906         if (rc == 0 && ext->oe_rc < 0)
 907                 rc = ext->oe_rc;
 908         RETURN(rc);
 909 }
 910
 911 /**
 912  * Discard pages with index greater than @size. If @ext is overlapped with
 913  * @size, then partial truncate happens.
 914  */
 915 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
 916                                 bool partial)
 917 {
 918         struct cl_env_nest     nest;
 919         struct lu_env         *env;
 920         struct cl_io          *io;
 921         struct osc_object     *obj = ext->oe_obj;
 922         struct client_obd     *cli = osc_cli(obj);
 923         struct osc_async_page *oap;
 924         struct osc_async_page *tmp;
 925         int                    pages_in_chunk = 0;
 926         int                    ppc_bits    = cli->cl_chunkbits -
 927                                              PAGE_CACHE_SHIFT;
 928         __u64                  trunc_chunk = trunc_index >> ppc_bits;
 929         int                    grants   = 0;
 930         int                    nr_pages = 0;
 931         int                    rc       = 0;
 932         ENTRY;
 933
 934         LASSERT(sanity_check(ext) == 0);
 935         LASSERT(ext->oe_state == OES_TRUNC);
 936         LASSERT(!ext->oe_urgent);
 937
 938         /* Request new lu_env.
 939          * We can't use that env from osc_cache_truncate_start() because
 940          * it's from lov_io_sub and not fully initialized. */
 941         env = cl_env_nested_get(&nest);
 942         io  = &osc_env_info(env)->oti_io;
 943         io->ci_obj = cl_object_top(osc2cl(obj));
 944         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
 945         if (rc < 0)
 946                 GOTO(out, rc);
 947
 948         /* discard all pages with index greater then trunc_index */
 949         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 950                                      oap_pending_item) {
 951                 pgoff_t index = osc_index(oap2osc(oap));
 952                 struct cl_page  *page = oap2cl_page(oap);
 953
 954                 LASSERT(list_empty(&oap->oap_rpc_item));
 955
 956                 /* only discard the pages with their index greater than
 957                  * trunc_index, and ... */
 958                 if (index < trunc_index ||
 959                     (index == trunc_index && partial)) {
 960                         /* accounting how many pages remaining in the chunk
 961                          * so that we can calculate grants correctly. */
 962                         if (index >> ppc_bits == trunc_chunk)
 963                                 ++pages_in_chunk;
 964                         continue;
 965                 }
 966
 967                 list_del_init(&oap->oap_pending_item);
 968
 969                 cl_page_get(page);
 970                 lu_ref_add(&page->cp_reference, "truncate", current);
 971
 972                 if (cl_page_own(env, io, page) == 0) {
 973                         cl_page_discard(env, io, page);
 974                         cl_page_disown(env, io, page);
 975                 } else {
 976                         LASSERT(page->cp_state == CPS_FREEING);
 977                         LASSERT(0);
 978                 }
 979
 980                 lu_ref_del(&page->cp_reference, "truncate", current);
 981                 cl_page_put(env, page);
 982
 983                 --ext->oe_nr_pages;
 984                 ++nr_pages;
 985         }
 986         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
 987                       ext->oe_nr_pages == 0),
 988                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
 989
 990         osc_object_lock(obj);
 991         if (ext->oe_nr_pages == 0) {
 992                 LASSERT(pages_in_chunk == 0);
 993                 grants = ext->oe_grants;
 994                 ext->oe_grants = 0;
 995         } else { /* calculate how many grants we can free */
 996                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
 997                 pgoff_t last_index;
 998
 999
1000                 /* if there is no pages in this chunk, we can also free grants
1001                  * for the last chunk */
1002                 if (pages_in_chunk == 0) {
1003                         /* if this is the 1st chunk and no pages in this chunk,
1004                          * ext->oe_nr_pages must be zero, so we should be in
1005                          * the other if-clause. */
1006                         LASSERT(trunc_chunk > 0);
1007                         --trunc_chunk;
1008                         ++chunks;
1009                 }
1010
1011                 /* this is what we can free from this extent */
1012                 grants          = chunks << cli->cl_chunkbits;
1013                 ext->oe_grants -= grants;
1014                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1015                 ext->oe_end     = min(last_index, ext->oe_max_end);
1016                 LASSERT(ext->oe_end >= ext->oe_start);
1017                 LASSERT(ext->oe_grants > 0);
1018         }
1019         osc_object_unlock(obj);
1020
1021         if (grants > 0 || nr_pages > 0)
1022                 osc_free_grant(cli, nr_pages, grants);
1023
1024 out:
1025         cl_io_fini(env, io);
1026         cl_env_nested_put(&nest, env);
1027         RETURN(rc);
1028 }
1029
1030 /**
1031  * This function is used to make the extent prepared for transfer.
1032  * A race with flusing page - ll_writepage() has to be handled cautiously.
1033  */
1034 static int osc_extent_make_ready(const struct lu_env *env,
1035                                  struct osc_extent *ext)
1036 {
1037         struct osc_async_page *oap;
1038         struct osc_async_page *last = NULL;
1039         struct osc_object *obj = ext->oe_obj;
1040         int page_count = 0;
1041         int rc;
1042         ENTRY;
1043
1044         /* we're going to grab page lock, so object lock must not be taken. */
1045         LASSERT(sanity_check(ext) == 0);
1046         /* in locking state, any process should not touch this extent. */
1047         EASSERT(ext->oe_state == OES_LOCKING, ext);
1048         EASSERT(ext->oe_owner != NULL, ext);
1049
1050         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1051
1052         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1053                 ++page_count;
1054                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1055                         last = oap;
1056
1057                 /* checking ASYNC_READY is race safe */
1058                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1059                         continue;
1060
1061                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1062                 switch (rc) {
1063                 case 0:
1064                         spin_lock(&oap->oap_lock);
1065                         oap->oap_async_flags |= ASYNC_READY;
1066                         spin_unlock(&oap->oap_lock);
1067                         break;
1068                 case -EALREADY:
1069                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1070                         break;
1071                 default:
1072                         LASSERTF(0, "unknown return code: %d\n", rc);
1073                 }
1074         }
1075
1076         LASSERT(page_count == ext->oe_nr_pages);
1077         LASSERT(last != NULL);
1078         /* the last page is the only one we need to refresh its count by
1079          * the size of file. */
1080         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1081                 last->oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1082                 LASSERT(last->oap_count > 0);
1083                 LASSERT(last->oap_page_off + last->oap_count <= PAGE_CACHE_SIZE);
1084                 spin_lock(&last->oap_lock);
1085                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1086                 spin_unlock(&last->oap_lock);
1087         }
1088
1089         /* for the rest of pages, we don't need to call osf_refresh_count()
1090          * because it's known they are not the last page */
1091         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1092                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1093                         oap->oap_count = PAGE_CACHE_SIZE - oap->oap_page_off;
1094                         spin_lock(&oap->oap_lock);
1095                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1096                         spin_unlock(&oap->oap_lock);
1097                 }
1098         }
1099
1100         osc_object_lock(obj);
1101         osc_extent_state_set(ext, OES_RPC);
1102         osc_object_unlock(obj);
1103         /* get a refcount for RPC. */
1104         osc_extent_get(ext);
1105
1106         RETURN(0);
1107 }
1108
1109 /**
1110  * Quick and simple version of osc_extent_find(). This function is frequently
1111  * called to expand the extent for the same IO. To expand the extent, the
1112  * page index must be in the same or next chunk of ext->oe_end.
1113  */
1114 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index, int *grants)
1115 {
1116         struct osc_object *obj = ext->oe_obj;
1117         struct client_obd *cli = osc_cli(obj);
1118         struct osc_extent *next;
1119         int ppc_bits = cli->cl_chunkbits - PAGE_CACHE_SHIFT;
1120         pgoff_t chunk = index >> ppc_bits;
1121         pgoff_t end_chunk;
1122         pgoff_t end_index;
1123         int chunksize = 1 << cli->cl_chunkbits;
1124         int rc = 0;
1125         ENTRY;
1126
1127         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1128         osc_object_lock(obj);
1129         LASSERT(sanity_check_nolock(ext) == 0);
1130         end_chunk = ext->oe_end >> ppc_bits;
1131         if (chunk > end_chunk + 1)
1132                 GOTO(out, rc = -ERANGE);
1133
1134         if (end_chunk >= chunk)
1135                 GOTO(out, rc = 0);
1136
1137         LASSERT(end_chunk + 1 == chunk);
1138         /* try to expand this extent to cover @index */
1139         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1140
1141         next = next_extent(ext);
1142         if (next != NULL && next->oe_start <= end_index)
1143                 /* complex mode - overlapped with the next extent,
1144                  * this case will be handled by osc_extent_find() */
1145                 GOTO(out, rc = -EAGAIN);
1146
1147         ext->oe_end = end_index;
1148         ext->oe_grants += chunksize;
1149         *grants -= chunksize;
1150         LASSERT(*grants >= 0);
1151         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1152                  "overlapped after expanding for %lu.\n", index);
1153         EXIT;
1154
1155 out:
1156         osc_object_unlock(obj);
1157         RETURN(rc);
1158 }
1159
1160 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
1161                                   const char *func, int line)
1162 {
1163         struct osc_extent *ext;
1164         int cnt;
1165
1166         CDEBUG(level, "Dump object %p extents at %s:%d, mppr: %u.\n",
1167                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1168
1169         /* osc_object_lock(obj); */
1170         cnt = 1;
1171         for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1172                 OSC_EXTENT_DUMP(level, ext, "in tree %d.\n", cnt++);
1173
1174         cnt = 1;
1175         list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1176                 OSC_EXTENT_DUMP(level, ext, "hp %d.\n", cnt++);
1177
1178         cnt = 1;
1179         list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1180                 OSC_EXTENT_DUMP(level, ext, "urgent %d.\n", cnt++);
1181
1182         cnt = 1;
1183         list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1184                 OSC_EXTENT_DUMP(level, ext, "reading %d.\n", cnt++);
1185         /* osc_object_unlock(obj); */
1186 }
1187
1188 /* ------------------ osc extent end ------------------ */
1189
1190 static inline int osc_is_ready(struct osc_object *osc)
1191 {
1192         return !list_empty(&osc->oo_ready_item) ||
1193                !list_empty(&osc->oo_hp_ready_item);
1194 }
1195
1196 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1197         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1198                (OSC), osc_is_ready(OSC),                                       \
1199                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1200                list_empty_marker(&(OSC)->oo_ready_item),                       \
1201                atomic_read(&(OSC)->oo_nr_writes),                              \
1202                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1203                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1204                atomic_read(&(OSC)->oo_nr_reads),                               \
1205                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1206                ##args)
1207
1208 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1209                           int cmd)
1210 {
1211         struct osc_page *opg  = oap2osc_page(oap);
1212         struct cl_page  *page = oap2cl_page(oap);
1213         int result;
1214
1215         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1216
1217         ENTRY;
1218         result = cl_page_make_ready(env, page, CRT_WRITE);
1219         if (result == 0)
1220                 opg->ops_submit_time = cfs_time_current();
1221         RETURN(result);
1222 }
1223
1224 static int osc_refresh_count(const struct lu_env *env,
1225                              struct osc_async_page *oap, int cmd)
1226 {
1227         struct osc_page  *opg = oap2osc_page(oap);
1228         pgoff_t index = osc_index(oap2osc(oap));
1229         struct cl_object *obj;
1230         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
1231
1232         int result;
1233         loff_t kms;
1234
1235         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1236         LASSERT(!(cmd & OBD_BRW_READ));
1237         LASSERT(opg != NULL);
1238         obj = opg->ops_cl.cpl_obj;
1239
1240         cl_object_attr_lock(obj);
1241         result = cl_object_attr_get(env, obj, attr);
1242         cl_object_attr_unlock(obj);
1243         if (result < 0)
1244                 return result;
1245         kms = attr->cat_kms;
1246         if (cl_offset(obj, index) >= kms)
1247                 /* catch race with truncate */
1248                 return 0;
1249         else if (cl_offset(obj, index + 1) > kms)
1250                 /* catch sub-page write at end of file */
1251                 return kms % PAGE_CACHE_SIZE;
1252         else
1253                 return PAGE_CACHE_SIZE;
1254 }
1255
1256 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1257                           int cmd, int rc)
1258 {
1259         struct osc_page   *opg  = oap2osc_page(oap);
1260         struct cl_page    *page = oap2cl_page(oap);
1261         struct osc_object *obj  = cl2osc(opg->ops_cl.cpl_obj);
1262         enum cl_req_type   crt;
1263         int srvlock;
1264
1265         ENTRY;
1266
1267         cmd &= ~OBD_BRW_NOQUOTA;
1268         LASSERTF(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ),
1269                  "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1270         LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1271                 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1272         LASSERT(opg->ops_transfer_pinned);
1273
1274         /*
1275          * page->cp_req can be NULL if io submission failed before
1276          * cl_req was allocated.
1277          */
1278         if (page->cp_req != NULL)
1279                 cl_req_page_done(env, page);
1280         LASSERT(page->cp_req == NULL);
1281
1282         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1283         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1284         opg->ops_transfer_pinned = 0;
1285
1286         spin_lock(&obj->oo_seatbelt);
1287         LASSERT(opg->ops_submitter != NULL);
1288         LASSERT(!list_empty(&opg->ops_inflight));
1289         list_del_init(&opg->ops_inflight);
1290         opg->ops_submitter = NULL;
1291         spin_unlock(&obj->oo_seatbelt);
1292
1293         opg->ops_submit_time = 0;
1294         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1295
1296         /* statistic */
1297         if (rc == 0 && srvlock) {
1298                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1299                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1300                 int bytes = oap->oap_count;
1301
1302                 if (crt == CRT_READ)
1303                         stats->os_lockless_reads += bytes;
1304                 else
1305                         stats->os_lockless_writes += bytes;
1306         }
1307
1308         /*
1309          * This has to be the last operation with the page, as locks are
1310          * released in cl_page_completion() and nothing except for the
1311          * reference counter protects page from concurrent reclaim.
1312          */
1313         lu_ref_del(&page->cp_reference, "transfer", page);
1314
1315         cl_page_completion(env, page, crt, rc);
1316
1317         RETURN(0);
1318 }
1319
1320 #define OSC_DUMP_GRANT(lvl, cli, fmt, args...) do {                     \
1321         struct client_obd *__tmp = (cli);                               \
1322         CDEBUG(lvl, "%s: grant { dirty: %ld/%ld dirty_pages: %d/%d "    \
1323                "dropped: %ld avail: %ld, reserved: %ld, flight: %d }"   \
1324                "lru {in list: %d, left: %d, waiters: %d }" fmt,         \
1325                __tmp->cl_import->imp_obd->obd_name,                     \
1326                __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages,        \
1327                atomic_read(&obd_dirty_pages), obd_max_dirty_pages,      \
1328                __tmp->cl_lost_grant, __tmp->cl_avail_grant,             \
1329                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,         \
1330                atomic_read(&__tmp->cl_lru_in_list),                     \
1331                atomic_read(&__tmp->cl_lru_busy),                        \
1332                atomic_read(&__tmp->cl_lru_shrinkers), ##args);          \
1333 } while (0)
1334
1335 /* caller must hold loi_list_lock */
1336 static void osc_consume_write_grant(struct client_obd *cli,
1337                                     struct brw_page *pga)
1338 {
1339         assert_spin_locked(&cli->cl_loi_list_lock.lock);
1340         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1341         atomic_inc(&obd_dirty_pages);
1342         cli->cl_dirty_pages++;
1343         pga->flag |= OBD_BRW_FROM_GRANT;
1344         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1345                PAGE_CACHE_SIZE, pga, pga->pg);
1346         osc_update_next_shrink(cli);
1347 }
1348
1349 /* the companion to osc_consume_write_grant, called when a brw has completed.
1350  * must be called with the loi lock held. */
1351 static void osc_release_write_grant(struct client_obd *cli,
1352                                     struct brw_page *pga)
1353 {
1354         ENTRY;
1355
1356         assert_spin_locked(&cli->cl_loi_list_lock.lock);
1357         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1358                 EXIT;
1359                 return;
1360         }
1361
1362         pga->flag &= ~OBD_BRW_FROM_GRANT;
1363         atomic_dec(&obd_dirty_pages);
1364         cli->cl_dirty_pages--;
1365         if (pga->flag & OBD_BRW_NOCACHE) {
1366                 pga->flag &= ~OBD_BRW_NOCACHE;
1367                 atomic_dec(&obd_dirty_transit_pages);
1368                 cli->cl_dirty_transit--;
1369         }
1370         EXIT;
1371 }
1372
1373 /**
1374  * To avoid sleeping with object lock held, it's good for us allocate enough
1375  * grants before entering into critical section.
1376  *
1377  * client_obd_list_lock held by caller
1378  */
1379 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1380 {
1381         int rc = -EDQUOT;
1382
1383         if (cli->cl_avail_grant >= bytes) {
1384                 cli->cl_avail_grant    -= bytes;
1385                 cli->cl_reserved_grant += bytes;
1386                 rc = 0;
1387         }
1388         return rc;
1389 }
1390
1391 static void __osc_unreserve_grant(struct client_obd *cli,
1392                                   unsigned int reserved, unsigned int unused)
1393 {
1394         /* it's quite normal for us to get more grant than reserved.
1395          * Thinking about a case that two extents merged by adding a new
1396          * chunk, we can save one extent tax. If extent tax is greater than
1397          * one chunk, we can save more grant by adding a new chunk */
1398         cli->cl_reserved_grant -= reserved;
1399         if (unused > reserved) {
1400                 cli->cl_avail_grant += reserved;
1401                 cli->cl_lost_grant  += unused - reserved;
1402         } else {
1403                 cli->cl_avail_grant += unused;
1404         }
1405 }
1406
1407 void osc_unreserve_grant(struct client_obd *cli,
1408                          unsigned int reserved, unsigned int unused)
1409 {
1410         client_obd_list_lock(&cli->cl_loi_list_lock);
1411         __osc_unreserve_grant(cli, reserved, unused);
1412         if (unused > 0)
1413                 osc_wake_cache_waiters(cli);
1414         client_obd_list_unlock(&cli->cl_loi_list_lock);
1415 }
1416
1417 /**
1418  * Free grant after IO is finished or canceled.
1419  *
1420  * @lost_grant is used to remember how many grants we have allocated but not
1421  * used, we should return these grants to OST. There're two cases where grants
1422  * can be lost:
1423  * 1. truncate;
1424  * 2. blocksize at OST is less than PAGE_CACHE_SIZE and a partial page was
1425  *    written. In this case OST may use less chunks to serve this partial
1426  *    write. OSTs don't actually know the page size on the client side. so
1427  *    clients have to calculate lost grant by the blocksize on the OST.
1428  *    See filter_grant_check() for details.
1429  */
1430 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1431                            unsigned int lost_grant)
1432 {
1433         int grant = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
1434
1435         client_obd_list_lock(&cli->cl_loi_list_lock);
1436         atomic_sub(nr_pages, &obd_dirty_pages);
1437         cli->cl_dirty_pages -= nr_pages;
1438         cli->cl_lost_grant += lost_grant;
1439         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1440                 /* borrow some grant from truncate to avoid the case that
1441                  * truncate uses up all avail grant */
1442                 cli->cl_lost_grant -= grant;
1443                 cli->cl_avail_grant += grant;
1444         }
1445         osc_wake_cache_waiters(cli);
1446         client_obd_list_unlock(&cli->cl_loi_list_lock);
1447         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu\n",
1448                lost_grant, cli->cl_lost_grant,
1449                cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_CACHE_SHIFT);
1450 }
1451
1452 /**
1453  * The companion to osc_enter_cache(), called when @oap is no longer part of
1454  * the dirty accounting due to error.
1455  */
1456 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1457 {
1458         client_obd_list_lock(&cli->cl_loi_list_lock);
1459         osc_release_write_grant(cli, &oap->oap_brw_page);
1460         client_obd_list_unlock(&cli->cl_loi_list_lock);
1461 }
1462
1463 /**
1464  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1465  * is available.
1466  */
1467 static int osc_enter_cache_try(struct client_obd *cli,
1468                                struct osc_async_page *oap,
1469                                int bytes, int transient)
1470 {
1471         int rc;
1472
1473         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d.\n", bytes);
1474
1475         rc = osc_reserve_grant(cli, bytes);
1476         if (rc < 0)
1477                 return 0;
1478
1479         if (cli->cl_dirty_pages < cli->cl_dirty_max_pages &&
1480             1 + atomic_read(&obd_dirty_pages) <= obd_max_dirty_pages) {
1481                 osc_consume_write_grant(cli, &oap->oap_brw_page);
1482                 if (transient) {
1483                         cli->cl_dirty_transit++;
1484                         atomic_inc(&obd_dirty_transit_pages);
1485                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
1486                 }
1487                 rc = 1;
1488         } else {
1489                 __osc_unreserve_grant(cli, bytes, bytes);
1490                 rc = 0;
1491         }
1492         return rc;
1493 }
1494
1495 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1496 {
1497         int rc;
1498         client_obd_list_lock(&cli->cl_loi_list_lock);
1499         rc = list_empty(&ocw->ocw_entry);
1500         client_obd_list_unlock(&cli->cl_loi_list_lock);
1501         return rc;
1502 }
1503
1504 /**
1505  * The main entry to reserve dirty page accounting. Usually the grant reserved
1506  * in this function will be freed in bulk in osc_free_grant() unless it fails
1507  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1508  *
1509  * The process will be put into sleep if it's already run out of grant.
1510  */
1511 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1512                            struct osc_async_page *oap, int bytes)
1513 {
1514         struct osc_object *osc = oap->oap_obj;
1515         struct lov_oinfo  *loi = osc->oo_oinfo;
1516         struct osc_cache_waiter ocw;
1517         struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
1518                                                   LWI_ON_SIGNAL_NOOP, NULL);
1519         int rc = -EDQUOT;
1520         ENTRY;
1521
1522         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d.\n", bytes);
1523
1524         client_obd_list_lock(&cli->cl_loi_list_lock);
1525
1526         /* force the caller to try sync io.  this can jump the list
1527          * of queued writes and create a discontiguous rpc stream */
1528         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1529             cli->cl_dirty_max_pages == 0 ||
1530             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync)
1531                 GOTO(out, rc = -EDQUOT);
1532
1533         /* Hopefully normal case - cache space and write credits available */
1534         if (osc_enter_cache_try(cli, oap, bytes, 0))
1535                 GOTO(out, rc = 0);
1536
1537         /* We can get here for two reasons: too many dirty pages in cache, or
1538          * run out of grants. In both cases we should write dirty pages out.
1539          * Adding a cache waiter will trigger urgent write-out no matter what
1540          * RPC size will be.
1541          * The exiting condition is no avail grants and no dirty pages caching,
1542          * that really means there is no space on the OST. */
1543         init_waitqueue_head(&ocw.ocw_waitq);
1544         ocw.ocw_oap   = oap;
1545         ocw.ocw_grant = bytes;
1546         while (cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0) {
1547                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1548                 ocw.ocw_rc = 0;
1549                 client_obd_list_unlock(&cli->cl_loi_list_lock);
1550
1551                 osc_io_unplug_async(env, cli, NULL);
1552
1553                 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1554                        cli->cl_import->imp_obd->obd_name, &ocw, oap);
1555
1556                 rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1557
1558                 client_obd_list_lock(&cli->cl_loi_list_lock);
1559
1560                 /* l_wait_event is interrupted by signal, or timed out */
1561                 if (rc < 0) {
1562                         switch (rc) {
1563                         case -ETIMEDOUT:
1564                                 OSC_DUMP_GRANT(D_ERROR, cli,
1565                                                 "try to reserve %d.\n", bytes);
1566                                 osc_extent_tree_dump(D_ERROR, osc);
1567                                 rc = -EDQUOT;
1568                                 break;
1569                         case -EINTR:
1570                                 /* Ensures restartability - LU-3581 */
1571                                 rc = -ERESTARTSYS;
1572                                 break;
1573                         default:
1574                                 CDEBUG(D_CACHE, "%s: event for cache space @"
1575                                        " %p never arrived due to %d\n",
1576                                        cli->cl_import->imp_obd->obd_name,
1577                                        &ocw, rc);
1578                                 break;
1579                         }
1580                         list_del_init(&ocw.ocw_entry);
1581                         GOTO(out, rc);
1582                 }
1583
1584                 LASSERT(list_empty(&ocw.ocw_entry));
1585                 rc = ocw.ocw_rc;
1586
1587                 if (rc != -EDQUOT)
1588                         GOTO(out, rc);
1589                 if (osc_enter_cache_try(cli, oap, bytes, 0))
1590                         GOTO(out, rc = 0);
1591         }
1592         EXIT;
1593 out:
1594         client_obd_list_unlock(&cli->cl_loi_list_lock);
1595         OSC_DUMP_GRANT(D_CACHE, cli, "returned %d.\n", rc);
1596         RETURN(rc);
1597 }
1598
1599 /* caller must hold loi_list_lock */
1600 void osc_wake_cache_waiters(struct client_obd *cli)
1601 {
1602         struct list_head *l, *tmp;
1603         struct osc_cache_waiter *ocw;
1604
1605         ENTRY;
1606         list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1607                 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
1608                 list_del_init(&ocw->ocw_entry);
1609
1610                 ocw->ocw_rc = -EDQUOT;
1611                 /* we can't dirty more */
1612                 if ((cli->cl_dirty_pages  >= cli->cl_dirty_max_pages) ||
1613                     (1 + atomic_read(&obd_dirty_pages) > obd_max_dirty_pages)) {
1614                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
1615                                "osc max %ld, sys max %d\n", cli->cl_dirty_pages,
1616                                cli->cl_dirty_max_pages, obd_max_dirty_pages);
1617                         goto wakeup;
1618                 }
1619
1620                 ocw->ocw_rc = 0;
1621                 if (!osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
1622                         ocw->ocw_rc = -EDQUOT;
1623
1624 wakeup:
1625                 CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld, %d\n",
1626                        ocw, ocw->ocw_oap, cli->cl_avail_grant, ocw->ocw_rc);
1627
1628                 wake_up(&ocw->ocw_waitq);
1629         }
1630
1631         EXIT;
1632 }
1633
1634 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1635 {
1636         int hprpc = !!list_empty(&osc->oo_hp_exts);
1637         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1638 }
1639
1640 /* This maintains the lists of pending pages to read/write for a given object
1641  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1642  * to quickly find objects that are ready to send an RPC. */
1643 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1644                          int cmd)
1645 {
1646         int invalid_import = 0;
1647         ENTRY;
1648
1649         /* if we have an invalid import we want to drain the queued pages
1650          * by forcing them through rpcs that immediately fail and complete
1651          * the pages.  recovery relies on this to empty the queued pages
1652          * before canceling the locks and evicting down the llite pages */
1653         if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1654                 invalid_import = 1;
1655
1656         if (cmd & OBD_BRW_WRITE) {
1657                 if (atomic_read(&osc->oo_nr_writes) == 0)
1658                         RETURN(0);
1659                 if (invalid_import) {
1660                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1661                         RETURN(1);
1662                 }
1663                 if (!list_empty(&osc->oo_hp_exts)) {
1664                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1665                         RETURN(1);
1666                 }
1667                 if (!list_empty(&osc->oo_urgent_exts)) {
1668                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1669                         RETURN(1);
1670                 }
1671                 /* trigger a write rpc stream as long as there are dirtiers
1672                  * waiting for space.  as they're waiting, they're not going to
1673                  * create more pages to coalesce with what's waiting.. */
1674                 if (!list_empty(&cli->cl_cache_waiters)) {
1675                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1676                         RETURN(1);
1677                 }
1678                 if (atomic_read(&osc->oo_nr_writes) >=
1679                     cli->cl_max_pages_per_rpc)
1680                         RETURN(1);
1681         } else {
1682                 if (atomic_read(&osc->oo_nr_reads) == 0)
1683                         RETURN(0);
1684                 if (invalid_import) {
1685                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1686                         RETURN(1);
1687                 }
1688                 /* all read are urgent. */
1689                 if (!list_empty(&osc->oo_reading_exts))
1690                         RETURN(1);
1691         }
1692
1693         RETURN(0);
1694 }
1695
1696 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1697 {
1698         struct client_obd *cli = osc_cli(obj);
1699         if (cmd & OBD_BRW_WRITE) {
1700                 atomic_add(delta, &obj->oo_nr_writes);
1701                 atomic_add(delta, &cli->cl_pending_w_pages);
1702                 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1703         } else {
1704                 atomic_add(delta, &obj->oo_nr_reads);
1705                 atomic_add(delta, &cli->cl_pending_r_pages);
1706                 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1707         }
1708         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1709 }
1710
1711 static int osc_makes_hprpc(struct osc_object *obj)
1712 {
1713         return !list_empty(&obj->oo_hp_exts);
1714 }
1715
1716 static void on_list(struct list_head *item, struct list_head *list,
1717                     int should_be_on)
1718 {
1719         if (list_empty(item) && should_be_on)
1720                 list_add_tail(item, list);
1721         else if (!list_empty(item) && !should_be_on)
1722                 list_del_init(item);
1723 }
1724
1725 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1726  * can find pages to build into rpcs quickly */
1727 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1728 {
1729         if (osc_makes_hprpc(osc)) {
1730                 /* HP rpc */
1731                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1732                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1733         } else {
1734                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1735                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1736                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1737                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1738         }
1739
1740         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1741                 atomic_read(&osc->oo_nr_writes) > 0);
1742
1743         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1744                 atomic_read(&osc->oo_nr_reads) > 0);
1745
1746         return osc_is_ready(osc);
1747 }
1748
1749 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1750 {
1751         int is_ready;
1752
1753         client_obd_list_lock(&cli->cl_loi_list_lock);
1754         is_ready = __osc_list_maint(cli, osc);
1755         client_obd_list_unlock(&cli->cl_loi_list_lock);
1756
1757         return is_ready;
1758 }
1759
1760 /* this is trying to propogate async writeback errors back up to the
1761  * application.  As an async write fails we record the error code for later if
1762  * the app does an fsync.  As long as errors persist we force future rpcs to be
1763  * sync so that the app can get a sync error and break the cycle of queueing
1764  * pages for which writeback will fail. */
1765 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1766                            int rc)
1767 {
1768         if (rc) {
1769                 if (!ar->ar_rc)
1770                         ar->ar_rc = rc;
1771
1772                 ar->ar_force_sync = 1;
1773                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1774                 return;
1775
1776         }
1777
1778         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1779                 ar->ar_force_sync = 0;
1780 }
1781
1782 /* this must be called holding the loi list lock to give coverage to exit_cache,
1783  * async_flag maintenance, and oap_request */
1784 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1785                               struct osc_async_page *oap, int sent, int rc)
1786 {
1787         struct osc_object *osc = oap->oap_obj;
1788         struct lov_oinfo  *loi = osc->oo_oinfo;
1789         __u64 xid = 0;
1790
1791         ENTRY;
1792         if (oap->oap_request != NULL) {
1793                 xid = ptlrpc_req_xid(oap->oap_request);
1794                 ptlrpc_req_finished(oap->oap_request);
1795                 oap->oap_request = NULL;
1796         }
1797
1798         /* As the transfer for this page is being done, clear the flags */
1799         spin_lock(&oap->oap_lock);
1800         oap->oap_async_flags = 0;
1801         spin_unlock(&oap->oap_lock);
1802         oap->oap_interrupted = 0;
1803
1804         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1805                 client_obd_list_lock(&cli->cl_loi_list_lock);
1806                 osc_process_ar(&cli->cl_ar, xid, rc);
1807                 osc_process_ar(&loi->loi_ar, xid, rc);
1808                 client_obd_list_unlock(&cli->cl_loi_list_lock);
1809         }
1810
1811         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1812         if (rc)
1813                 CERROR("completion on oap %p obj %p returns %d.\n",
1814                        oap, osc, rc);
1815
1816         EXIT;
1817 }
1818
1819 /**
1820  * Try to add extent to one RPC. We need to think about the following things:
1821  * - # of pages must not be over max_pages_per_rpc
1822  * - extent must be compatible with previous ones
1823  */
1824 static int try_to_add_extent_for_io(struct client_obd *cli,
1825                                     struct osc_extent *ext,
1826                                     struct list_head *rpclist,
1827                                     int *pc, unsigned int *max_pages)
1828 {
1829         struct osc_extent *tmp;
1830         struct osc_async_page *oap = list_first_entry(&ext->oe_pages,
1831                                                       struct osc_async_page,
1832                                                       oap_pending_item);
1833         ENTRY;
1834
1835         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1836                 ext);
1837
1838         *max_pages = max(ext->oe_mppr, *max_pages);
1839         if (*pc + ext->oe_nr_pages > *max_pages)
1840                 RETURN(0);
1841
1842         list_for_each_entry(tmp, rpclist, oe_link) {
1843                 struct osc_async_page *oap2;
1844                 oap2 = list_first_entry(&tmp->oe_pages, struct osc_async_page,
1845                                         oap_pending_item);
1846                 EASSERT(tmp->oe_owner == current, tmp);
1847 #if 0
1848                 if (overlapped(tmp, ext)) {
1849                         OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
1850                         EASSERT(0, ext);
1851                 }
1852 #endif
1853                 if (oap2cl_page(oap)->cp_type != oap2cl_page(oap2)->cp_type) {
1854                         CDEBUG(D_CACHE, "Do not permit different type of IO"
1855                                         " for a same RPC\n");
1856                         RETURN(0);
1857                 }
1858
1859                 if (tmp->oe_srvlock != ext->oe_srvlock ||
1860                     !tmp->oe_grants != !ext->oe_grants)
1861                         RETURN(0);
1862
1863                 /* remove break for strict check */
1864                 break;
1865         }
1866
1867         *pc += ext->oe_nr_pages;
1868         list_move_tail(&ext->oe_link, rpclist);
1869         ext->oe_owner = current;
1870         RETURN(1);
1871 }
1872
1873 /**
1874  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1875  * get_write_extent() takes all appropriate extents in atomic.
1876  *
1877  * The following policy is used to collect extents for IO:
1878  * 1. Add as many HP extents as possible;
1879  * 2. Add the first urgent extent in urgent extent list and take it out of
1880  *    urgent list;
1881  * 3. Add subsequent extents of this urgent extent;
1882  * 4. If urgent list is not empty, goto 2;
1883  * 5. Traverse the extent tree from the 1st extent;
1884  * 6. Above steps exit if there is no space in this RPC.
1885  */
1886 static int get_write_extents(struct osc_object *obj, struct list_head *rpclist)
1887 {
1888         struct client_obd *cli = osc_cli(obj);
1889         struct osc_extent *ext;
1890         int page_count = 0;
1891         unsigned int max_pages = cli->cl_max_pages_per_rpc;
1892
1893         LASSERT(osc_object_is_locked(obj));
1894         while (!list_empty(&obj->oo_hp_exts)) {
1895                 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
1896                                  oe_link);
1897                 LASSERT(ext->oe_state == OES_CACHE);
1898                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1899                                               &max_pages))
1900                         return page_count;
1901                 EASSERT(ext->oe_nr_pages <= max_pages, ext);
1902         }
1903         if (page_count == max_pages)
1904                 return page_count;
1905
1906         while (!list_empty(&obj->oo_urgent_exts)) {
1907                 ext = list_entry(obj->oo_urgent_exts.next,
1908                                  struct osc_extent, oe_link);
1909                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1910                                               &max_pages))
1911                         return page_count;
1912
1913                 if (!ext->oe_intree)
1914                         continue;
1915
1916                 while ((ext = next_extent(ext)) != NULL) {
1917                         if ((ext->oe_state != OES_CACHE) ||
1918                             (!list_empty(&ext->oe_link) &&
1919                              ext->oe_owner != NULL))
1920                                 continue;
1921
1922                         if (!try_to_add_extent_for_io(cli, ext, rpclist,
1923                                                       &page_count, &max_pages))
1924                                 return page_count;
1925                 }
1926         }
1927         if (page_count == max_pages)
1928                 return page_count;
1929
1930         ext = first_extent(obj);
1931         while (ext != NULL) {
1932                 if ((ext->oe_state != OES_CACHE) ||
1933                     /* this extent may be already in current rpclist */
1934                     (!list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
1935                         ext = next_extent(ext);
1936                         continue;
1937                 }
1938
1939                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1940                                               &max_pages))
1941                         return page_count;
1942
1943                 ext = next_extent(ext);
1944         }
1945         return page_count;
1946 }
1947
1948 static int
1949 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
1950                    struct osc_object *osc, pdl_policy_t pol)
1951 __must_hold(osc)
1952 {
1953         struct list_head   rpclist = LIST_HEAD_INIT(rpclist);
1954         struct osc_extent *ext;
1955         struct osc_extent *tmp;
1956         struct osc_extent *first = NULL;
1957         obd_count page_count = 0;
1958         int srvlock = 0;
1959         int rc = 0;
1960         ENTRY;
1961
1962         LASSERT(osc_object_is_locked(osc));
1963
1964         page_count = get_write_extents(osc, &rpclist);
1965         LASSERT(equi(page_count == 0, list_empty(&rpclist)));
1966
1967         if (list_empty(&rpclist))
1968                 RETURN(0);
1969
1970         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
1971
1972         list_for_each_entry(ext, &rpclist, oe_link) {
1973                 LASSERT(ext->oe_state == OES_CACHE ||
1974                         ext->oe_state == OES_LOCK_DONE);
1975                 if (ext->oe_state == OES_CACHE)
1976                         osc_extent_state_set(ext, OES_LOCKING);
1977                 else
1978                         osc_extent_state_set(ext, OES_RPC);
1979         }
1980
1981         /* we're going to grab page lock, so release object lock because
1982          * lock order is page lock -> object lock. */
1983         osc_object_unlock(osc);
1984
1985         list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
1986                 if (ext->oe_state == OES_LOCKING) {
1987                         rc = osc_extent_make_ready(env, ext);
1988                         if (unlikely(rc < 0)) {
1989                                 list_del_init(&ext->oe_link);
1990                                 osc_extent_finish(env, ext, 0, rc);
1991                                 continue;
1992                         }
1993                 }
1994                 if (first == NULL) {
1995                         first = ext;
1996                         srvlock = ext->oe_srvlock;
1997                 } else {
1998                         LASSERT(srvlock == ext->oe_srvlock);
1999                 }
2000         }
2001
2002         if (!list_empty(&rpclist)) {
2003                 LASSERT(page_count > 0);
2004                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE, pol);
2005                 LASSERT(list_empty(&rpclist));
2006         }
2007
2008         osc_object_lock(osc);
2009         RETURN(rc);
2010 }
2011
2012 /**
2013  * prepare pages for ASYNC io and put pages in send queue.
2014  *
2015  * \param cmd OBD_BRW_* macroses
2016  * \param lop pending pages
2017  *
2018  * \return zero if no page added to send queue.
2019  * \return 1 if pages successfully added to send queue.
2020  * \return negative on errors.
2021  */
2022 static int
2023 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2024                   struct osc_object *osc, pdl_policy_t pol)
2025 __must_hold(osc)
2026 {
2027         struct osc_extent *ext;
2028         struct osc_extent *next;
2029         struct list_head rpclist = LIST_HEAD_INIT(rpclist);
2030         int page_count = 0;
2031         unsigned int max_pages = cli->cl_max_pages_per_rpc;
2032         int rc = 0;
2033         ENTRY;
2034
2035         LASSERT(osc_object_is_locked(osc));
2036         list_for_each_entry_safe(ext, next,
2037                                      &osc->oo_reading_exts, oe_link) {
2038                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2039                 if (!try_to_add_extent_for_io(cli, ext, &rpclist, &page_count,
2040                                               &max_pages))
2041                         break;
2042                 osc_extent_state_set(ext, OES_RPC);
2043                 EASSERT(ext->oe_nr_pages <= max_pages, ext);
2044         }
2045         LASSERT(page_count <= max_pages);
2046
2047         osc_update_pending(osc, OBD_BRW_READ, -page_count);
2048
2049         if (!list_empty(&rpclist)) {
2050                 osc_object_unlock(osc);
2051
2052                 LASSERT(page_count > 0);
2053                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ, pol);
2054                 LASSERT(list_empty(&rpclist));
2055
2056                 osc_object_lock(osc);
2057         }
2058         RETURN(rc);
2059 }
2060
2061 #define list_to_obj(list, item) ({                                            \
2062         struct list_head *__tmp = (list)->next;                               \
2063         list_del_init(__tmp);                                         \
2064         list_entry(__tmp, struct osc_object, oo_##item);                      \
2065 })
2066
2067 /* This is called by osc_check_rpcs() to find which objects have pages that
2068  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2069 static struct osc_object *osc_next_obj(struct client_obd *cli)
2070 {
2071         ENTRY;
2072
2073         /* First return objects that have blocked locks so that they
2074          * will be flushed quickly and other clients can get the lock,
2075          * then objects which have pages ready to be stuffed into RPCs */
2076         if (!list_empty(&cli->cl_loi_hp_ready_list))
2077                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2078         if (!list_empty(&cli->cl_loi_ready_list))
2079                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2080
2081         /* then if we have cache waiters, return all objects with queued
2082          * writes.  This is especially important when many small files
2083          * have filled up the cache and not been fired into rpcs because
2084          * they don't pass the nr_pending/object threshhold */
2085         if (!list_empty(&cli->cl_cache_waiters) &&
2086             !list_empty(&cli->cl_loi_write_list))
2087                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2088
2089         /* then return all queued objects when we have an invalid import
2090          * so that they get flushed */
2091         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2092                 if (!list_empty(&cli->cl_loi_write_list))
2093                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2094                                            write_item));
2095                 if (!list_empty(&cli->cl_loi_read_list))
2096                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2097                                            read_item));
2098         }
2099         RETURN(NULL);
2100 }
2101
2102 /* called with the loi list lock held */
2103 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli,
2104                            pdl_policy_t pol)
2105 __must_hold(&cli->cl_loi_list_lock)
2106 {
2107         struct osc_object *osc;
2108         int rc = 0;
2109         ENTRY;
2110
2111         while ((osc = osc_next_obj(cli)) != NULL) {
2112                 struct cl_object *obj = osc2cl(osc);
2113                 struct lu_ref_link link;
2114
2115                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2116
2117                 if (osc_max_rpc_in_flight(cli, osc)) {
2118                         __osc_list_maint(cli, osc);
2119                         break;
2120                 }
2121
2122                 cl_object_get(obj);
2123                 client_obd_list_unlock(&cli->cl_loi_list_lock);
2124                 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2125
2126                 /* attempt some read/write balancing by alternating between
2127                  * reads and writes in an object.  The makes_rpc checks here
2128                  * would be redundant if we were getting read/write work items
2129                  * instead of objects.  we don't want send_oap_rpc to drain a
2130                  * partial read pending queue when we're given this object to
2131                  * do io on writes while there are cache waiters */
2132                 osc_object_lock(osc);
2133                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2134                         rc = osc_send_write_rpc(env, cli, osc, pol);
2135                         if (rc < 0) {
2136                                 CERROR("Write request failed with %d\n", rc);
2137
2138                                 /* osc_send_write_rpc failed, mostly because of
2139                                  * memory pressure.
2140                                  *
2141                                  * It can't break here, because if:
2142                                  *  - a page was submitted by osc_io_submit, so
2143                                  *    page locked;
2144                                  *  - no request in flight
2145                                  *  - no subsequent request
2146                                  * The system will be in live-lock state,
2147                                  * because there is no chance to call
2148                                  * osc_io_unplug() and osc_check_rpcs() any
2149                                  * more. pdflush can't help in this case,
2150                                  * because it might be blocked at grabbing
2151                                  * the page lock as we mentioned.
2152                                  *
2153                                  * Anyway, continue to drain pages. */
2154                                 /* break; */
2155                         }
2156                 }
2157                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2158                         rc = osc_send_read_rpc(env, cli, osc, pol);
2159                         if (rc < 0)
2160                                 CERROR("Read request failed with %d\n", rc);
2161                 }
2162                 osc_object_unlock(osc);
2163
2164                 osc_list_maint(cli, osc);
2165                 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2166                 cl_object_put(env, obj);
2167
2168                 client_obd_list_lock(&cli->cl_loi_list_lock);
2169         }
2170 }
2171
2172 static int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2173                           struct osc_object *osc, pdl_policy_t pol, int async)
2174 {
2175         int rc = 0;
2176
2177         if (osc != NULL && osc_list_maint(cli, osc) == 0)
2178                 return 0;
2179
2180         if (!async) {
2181                 /* disable osc_lru_shrink() temporarily to avoid
2182                  * potential stack overrun problem. LU-2859 */
2183                 atomic_inc(&cli->cl_lru_shrinkers);
2184                 client_obd_list_lock(&cli->cl_loi_list_lock);
2185                 osc_check_rpcs(env, cli, pol);
2186                 client_obd_list_unlock(&cli->cl_loi_list_lock);
2187                 atomic_dec(&cli->cl_lru_shrinkers);
2188         } else {
2189                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2190                 LASSERT(cli->cl_writeback_work != NULL);
2191                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2192         }
2193         return rc;
2194 }
2195
2196 static int osc_io_unplug_async(const struct lu_env *env,
2197                                 struct client_obd *cli, struct osc_object *osc)
2198 {
2199         /* XXX: policy is no use actually. */
2200         return osc_io_unplug0(env, cli, osc, PDL_POLICY_ROUND, 1);
2201 }
2202
2203 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
2204                    struct osc_object *osc, pdl_policy_t pol)
2205 {
2206         (void)osc_io_unplug0(env, cli, osc, pol, 0);
2207 }
2208
2209 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2210                         struct page *page, loff_t offset)
2211 {
2212         struct obd_export     *exp = osc_export(osc);
2213         struct osc_async_page *oap = &ops->ops_oap;
2214         ENTRY;
2215
2216         if (!page)
2217                 return cfs_size_round(sizeof(*oap));
2218
2219         oap->oap_magic = OAP_MAGIC;
2220         oap->oap_cli = &exp->exp_obd->u.cli;
2221         oap->oap_obj = osc;
2222
2223         oap->oap_page = page;
2224         oap->oap_obj_off = offset;
2225         LASSERT(!(offset & ~CFS_PAGE_MASK));
2226
2227         if (!client_is_remote(exp) && cfs_capable(CFS_CAP_SYS_RESOURCE))
2228                 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2229
2230         INIT_LIST_HEAD(&oap->oap_pending_item);
2231         INIT_LIST_HEAD(&oap->oap_rpc_item);
2232
2233         spin_lock_init(&oap->oap_lock);
2234         CDEBUG(D_INFO, "oap %p page %p obj off "LPU64"\n",
2235                oap, page, oap->oap_obj_off);
2236         RETURN(0);
2237 }
2238
2239 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2240                        struct osc_page *ops)
2241 {
2242         struct osc_io *oio = osc_env_io(env);
2243         struct osc_extent     *ext = NULL;
2244         struct osc_async_page *oap = &ops->ops_oap;
2245         struct client_obd     *cli = oap->oap_cli;
2246         struct osc_object     *osc = oap->oap_obj;
2247         pgoff_t index;
2248         int    grants = 0;
2249         int    brw_flags = OBD_BRW_ASYNC;
2250         int    cmd = OBD_BRW_WRITE;
2251         int    need_release = 0;
2252         int    rc = 0;
2253         ENTRY;
2254
2255         if (oap->oap_magic != OAP_MAGIC)
2256                 RETURN(-EINVAL);
2257
2258         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2259                 RETURN(-EIO);
2260
2261         if (!list_empty(&oap->oap_pending_item) ||
2262             !list_empty(&oap->oap_rpc_item))
2263                 RETURN(-EBUSY);
2264
2265         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2266         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2267         if (!client_is_remote(osc_export(osc)) &&
2268             cfs_capable(CFS_CAP_SYS_RESOURCE)) {
2269                 brw_flags |= OBD_BRW_NOQUOTA;
2270                 cmd |= OBD_BRW_NOQUOTA;
2271         }
2272
2273         /* check if the file's owner/group is over quota */
2274         if (!(cmd & OBD_BRW_NOQUOTA)) {
2275                 struct cl_object *obj;
2276                 struct cl_attr   *attr;
2277                 unsigned int qid[MAXQUOTAS];
2278
2279                 obj = cl_object_top(&osc->oo_cl);
2280                 attr = &osc_env_info(env)->oti_attr;
2281
2282                 cl_object_attr_lock(obj);
2283                 rc = cl_object_attr_get(env, obj, attr);
2284                 cl_object_attr_unlock(obj);
2285
2286                 qid[USRQUOTA] = attr->cat_uid;
2287                 qid[GRPQUOTA] = attr->cat_gid;
2288                 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2289                         rc = -EDQUOT;
2290                 if (rc)
2291                         RETURN(rc);
2292         }
2293
2294         oap->oap_cmd = cmd;
2295         oap->oap_page_off = ops->ops_from;
2296         oap->oap_count = ops->ops_to - ops->ops_from;
2297         /* No need to hold a lock here,
2298          * since this page is not in any list yet. */
2299         oap->oap_async_flags = 0;
2300         oap->oap_brw_flags = brw_flags;
2301
2302         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2303                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2304
2305         index = osc_index(oap2osc(oap));
2306
2307         /* Add this page into extent by the following steps:
2308          * 1. if there exists an active extent for this IO, mostly this page
2309          *    can be added to the active extent and sometimes we need to
2310          *    expand extent to accomodate this page;
2311          * 2. otherwise, a new extent will be allocated. */
2312
2313         ext = oio->oi_active;
2314         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2315                 /* one chunk plus extent overhead must be enough to write this
2316                  * page */
2317                 grants = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2318                 if (ext->oe_end >= index)
2319                         grants = 0;
2320
2321                 /* it doesn't need any grant to dirty this page */
2322                 client_obd_list_lock(&cli->cl_loi_list_lock);
2323                 rc = osc_enter_cache_try(cli, oap, grants, 0);
2324                 client_obd_list_unlock(&cli->cl_loi_list_lock);
2325                 if (rc == 0) { /* try failed */
2326                         grants = 0;
2327                         need_release = 1;
2328                 } else if (ext->oe_end < index) {
2329                         int tmp = grants;
2330                         /* try to expand this extent */
2331                         rc = osc_extent_expand(ext, index, &tmp);
2332                         if (rc < 0) {
2333                                 need_release = 1;
2334                                 /* don't free reserved grant */
2335                         } else {
2336                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2337                                                 "expanded for %lu.\n", index);
2338                                 osc_unreserve_grant(cli, grants, tmp);
2339                                 grants = 0;
2340                         }
2341                 }
2342                 rc = 0;
2343         } else if (ext != NULL) {
2344                 /* index is located outside of active extent */
2345                 need_release = 1;
2346         }
2347         if (need_release) {
2348                 osc_extent_release(env, ext);
2349                 oio->oi_active = NULL;
2350                 ext = NULL;
2351         }
2352
2353         if (ext == NULL) {
2354                 int tmp = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2355
2356                 /* try to find new extent to cover this page */
2357                 LASSERT(oio->oi_active == NULL);
2358                 /* we may have allocated grant for this page if we failed
2359                  * to expand the previous active extent. */
2360                 LASSERT(ergo(grants > 0, grants >= tmp));
2361
2362                 rc = 0;
2363                 if (grants == 0) {
2364                         /* we haven't allocated grant for this page. */
2365                         rc = osc_enter_cache(env, cli, oap, tmp);
2366                         if (rc == 0)
2367                                 grants = tmp;
2368                 }
2369
2370                 tmp = grants;
2371                 if (rc == 0) {
2372                         ext = osc_extent_find(env, osc, index, &tmp);
2373                         if (IS_ERR(ext)) {
2374                                 LASSERT(tmp == grants);
2375                                 osc_exit_cache(cli, oap);
2376                                 rc = PTR_ERR(ext);
2377                                 ext = NULL;
2378                         } else {
2379                                 oio->oi_active = ext;
2380                         }
2381                 }
2382                 if (grants > 0)
2383                         osc_unreserve_grant(cli, grants, tmp);
2384         }
2385
2386         LASSERT(ergo(rc == 0, ext != NULL));
2387         if (ext != NULL) {
2388                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2389                          ext, "index = %lu.\n", index);
2390                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2391
2392                 osc_object_lock(osc);
2393                 if (ext->oe_nr_pages == 0)
2394                         ext->oe_srvlock = ops->ops_srvlock;
2395                 else
2396                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2397                 ++ext->oe_nr_pages;
2398                 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2399                 osc_object_unlock(osc);
2400         }
2401         RETURN(rc);
2402 }
2403
2404 int osc_teardown_async_page(const struct lu_env *env,
2405                             struct osc_object *obj, struct osc_page *ops)
2406 {
2407         struct osc_async_page *oap = &ops->ops_oap;
2408         struct osc_extent     *ext = NULL;
2409         int rc = 0;
2410         ENTRY;
2411
2412         LASSERT(oap->oap_magic == OAP_MAGIC);
2413
2414         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2415                oap, ops, osc_index(oap2osc(oap)));
2416
2417         osc_object_lock(obj);
2418         if (!list_empty(&oap->oap_rpc_item)) {
2419                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2420                 rc = -EBUSY;
2421         } else if (!list_empty(&oap->oap_pending_item)) {
2422                 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2423                 /* only truncated pages are allowed to be taken out.
2424                  * See osc_extent_truncate() and osc_cache_truncate_start()
2425                  * for details. */
2426                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2427                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2428                                         osc_index(oap2osc(oap)));
2429                         rc = -EBUSY;
2430                 }
2431         }
2432         osc_object_unlock(obj);
2433         if (ext != NULL)
2434                 osc_extent_put(env, ext);
2435         RETURN(rc);
2436 }
2437
2438 /**
2439  * This is called when a page is picked up by kernel to write out.
2440  *
2441  * We should find out the corresponding extent and add the whole extent
2442  * into urgent list. The extent may be being truncated or used, handle it
2443  * carefully.
2444  */
2445 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2446                          struct osc_page *ops)
2447 {
2448         struct osc_extent *ext   = NULL;
2449         struct osc_object *obj   = cl2osc(ops->ops_cl.cpl_obj);
2450         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2451         pgoff_t            index = osc_index(ops);
2452         struct osc_async_page *oap = &ops->ops_oap;
2453         bool unplug = false;
2454         int rc = 0;
2455         ENTRY;
2456
2457         osc_object_lock(obj);
2458         ext = osc_extent_lookup(obj, index);
2459         if (ext == NULL) {
2460                 osc_extent_tree_dump(D_ERROR, obj);
2461                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2462         }
2463
2464         switch (ext->oe_state) {
2465         case OES_RPC:
2466         case OES_LOCK_DONE:
2467                 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2468                 LASSERT(0);
2469                 break;
2470         case OES_LOCKING:
2471                 /* If we know this extent is being written out, we should abort
2472                  * so that the writer can make this page ready. Otherwise, there
2473                  * exists a deadlock problem because other process can wait for
2474                  * page writeback bit holding page lock; and meanwhile in
2475                  * vvp_page_make_ready(), we need to grab page lock before
2476                  * really sending the RPC. */
2477         case OES_TRUNC:
2478                 /* race with truncate, page will be redirtied */
2479         case OES_ACTIVE:
2480                 /* The extent is active so we need to abort and let the caller
2481                  * re-dirty the page. If we continued on here, and we were the
2482                  * one making the extent active, we could deadlock waiting for
2483                  * the page writeback to clear but it won't because the extent
2484                  * is active and won't be written out. */
2485                 GOTO(out, rc = -EAGAIN);
2486         default:
2487                 break;
2488         }
2489
2490         rc = cl_page_prep(env, io, cp, CRT_WRITE);
2491         if (rc)
2492                 GOTO(out, rc);
2493
2494         spin_lock(&oap->oap_lock);
2495         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2496         spin_unlock(&oap->oap_lock);
2497
2498         if (memory_pressure_get())
2499                 ext->oe_memalloc = 1;
2500
2501         ext->oe_urgent = 1;
2502         if (ext->oe_state == OES_CACHE) {
2503                 OSC_EXTENT_DUMP(D_CACHE, ext,
2504                                 "flush page %p make it urgent.\n", oap);
2505                 if (list_empty(&ext->oe_link))
2506                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2507                 unplug = true;
2508         }
2509         rc = 0;
2510         EXIT;
2511
2512 out:
2513         osc_object_unlock(obj);
2514         osc_extent_put(env, ext);
2515         if (unplug)
2516                 osc_io_unplug_async(env, osc_cli(obj), obj);
2517         return rc;
2518 }
2519
2520 /**
2521  * this is called when a sync waiter receives an interruption.  Its job is to
2522  * get the caller woken as soon as possible.  If its page hasn't been put in an
2523  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2524  * desiring interruption which will forcefully complete the rpc once the rpc
2525  * has timed out.
2526  */
2527 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2528 {
2529         struct osc_async_page *oap = &ops->ops_oap;
2530         struct osc_object     *obj = oap->oap_obj;
2531         struct client_obd     *cli = osc_cli(obj);
2532         struct osc_extent     *ext;
2533         struct osc_extent     *found = NULL;
2534         struct list_head            *plist;
2535         pgoff_t index = osc_index(ops);
2536         int     rc = -EBUSY;
2537         int     cmd;
2538         ENTRY;
2539
2540         LASSERT(!oap->oap_interrupted);
2541         oap->oap_interrupted = 1;
2542
2543         /* Find out the caching extent */
2544         osc_object_lock(obj);
2545         if (oap->oap_cmd & OBD_BRW_WRITE) {
2546                 plist = &obj->oo_urgent_exts;
2547                 cmd   = OBD_BRW_WRITE;
2548         } else {
2549                 plist = &obj->oo_reading_exts;
2550                 cmd   = OBD_BRW_READ;
2551         }
2552         list_for_each_entry(ext, plist, oe_link) {
2553                 if (ext->oe_start <= index && ext->oe_end >= index) {
2554                         LASSERT(ext->oe_state == OES_LOCK_DONE);
2555                         /* For OES_LOCK_DONE state extent, it has already held
2556                          * a refcount for RPC. */
2557                         found = osc_extent_get(ext);
2558                         break;
2559                 }
2560         }
2561         if (found != NULL) {
2562                 list_del_init(&found->oe_link);
2563                 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2564                 osc_object_unlock(obj);
2565
2566                 osc_extent_finish(env, found, 0, -EINTR);
2567                 osc_extent_put(env, found);
2568                 rc = 0;
2569         } else {
2570                 osc_object_unlock(obj);
2571                 /* ok, it's been put in an rpc. only one oap gets a request
2572                  * reference */
2573                 if (oap->oap_request != NULL) {
2574                         ptlrpc_mark_interrupted(oap->oap_request);
2575                         ptlrpcd_wake(oap->oap_request);
2576                         ptlrpc_req_finished(oap->oap_request);
2577                         oap->oap_request = NULL;
2578                 }
2579         }
2580
2581         osc_list_maint(cli, obj);
2582         RETURN(rc);
2583 }
2584
2585 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
2586                          struct list_head *list, int cmd, int brw_flags)
2587 {
2588         struct client_obd     *cli = osc_cli(obj);
2589         struct osc_extent     *ext;
2590         struct osc_async_page *oap;
2591         int     page_count = 0;
2592         int     mppr       = cli->cl_max_pages_per_rpc;
2593         pgoff_t start      = CL_PAGE_EOF;
2594         pgoff_t end        = 0;
2595         ENTRY;
2596
2597         list_for_each_entry(oap, list, oap_pending_item) {
2598                 pgoff_t index = osc_index(oap2osc(oap));
2599                 if (index > end)
2600                         end = index;
2601                 if (index < start)
2602                         start = index;
2603                 ++page_count;
2604                 mppr <<= (page_count > mppr);
2605         }
2606
2607         ext = osc_extent_alloc(obj);
2608         if (ext == NULL) {
2609                 list_for_each_entry(oap, list, oap_pending_item) {
2610                         list_del_init(&oap->oap_pending_item);
2611                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2612                 }
2613                 RETURN(-ENOMEM);
2614         }
2615
2616         ext->oe_rw = !!(cmd & OBD_BRW_READ);
2617         ext->oe_urgent = 1;
2618         ext->oe_start = start;
2619         ext->oe_end = ext->oe_max_end = end;
2620         ext->oe_obj = obj;
2621         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2622         ext->oe_nr_pages = page_count;
2623         ext->oe_mppr = mppr;
2624         list_splice_init(list, &ext->oe_pages);
2625
2626         osc_object_lock(obj);
2627         /* Reuse the initial refcount for RPC, don't drop it */
2628         osc_extent_state_set(ext, OES_LOCK_DONE);
2629         if (cmd & OBD_BRW_WRITE) {
2630                 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2631                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2632         } else {
2633                 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2634                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2635         }
2636         osc_object_unlock(obj);
2637
2638         osc_io_unplug_async(env, cli, obj);
2639         RETURN(0);
2640 }
2641
2642 /**
2643  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2644  */
2645 int osc_cache_truncate_start(const struct lu_env *env, struct osc_io *oio,
2646                              struct osc_object *obj, __u64 size)
2647 {
2648         struct client_obd *cli = osc_cli(obj);
2649         struct osc_extent *ext;
2650         struct osc_extent *waiting = NULL;
2651         pgoff_t index;
2652         struct list_head list = LIST_HEAD_INIT(list);
2653         int result = 0;
2654         bool partial;
2655         ENTRY;
2656
2657         /* pages with index greater or equal to index will be truncated. */
2658         index = cl_index(osc2cl(obj), size);
2659         partial = size > cl_offset(osc2cl(obj), index);
2660
2661 again:
2662         osc_object_lock(obj);
2663         ext = osc_extent_search(obj, index);
2664         if (ext == NULL)
2665                 ext = first_extent(obj);
2666         else if (ext->oe_end < index)
2667                 ext = next_extent(ext);
2668         while (ext != NULL) {
2669                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2670
2671                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2672                         /* if ext is in urgent state, it means there must exist
2673                          * a page already having been flushed by write_page().
2674                          * We have to wait for this extent because we can't
2675                          * truncate that page. */
2676                         LASSERT(!ext->oe_hp);
2677                         OSC_EXTENT_DUMP(D_CACHE, ext,
2678                                         "waiting for busy extent\n");
2679                         waiting = osc_extent_get(ext);
2680                         break;
2681                 }
2682
2683                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:"LPU64".\n", size);
2684
2685                 osc_extent_get(ext);
2686                 if (ext->oe_state == OES_ACTIVE) {
2687                         /* though we grab inode mutex for write path, but we
2688                          * release it before releasing extent(in osc_io_end()),
2689                          * so there is a race window that an extent is still
2690                          * in OES_ACTIVE when truncate starts. */
2691                         LASSERT(!ext->oe_trunc_pending);
2692                         ext->oe_trunc_pending = 1;
2693                 } else {
2694                         EASSERT(ext->oe_state == OES_CACHE, ext);
2695                         osc_extent_state_set(ext, OES_TRUNC);
2696                         osc_update_pending(obj, OBD_BRW_WRITE,
2697                                            -ext->oe_nr_pages);
2698                 }
2699                 EASSERT(list_empty(&ext->oe_link), ext);
2700                 list_add_tail(&ext->oe_link, &list);
2701
2702                 ext = next_extent(ext);
2703         }
2704         osc_object_unlock(obj);
2705
2706         osc_list_maint(cli, obj);
2707
2708         while (!list_empty(&list)) {
2709                 int rc;
2710
2711                 ext = list_entry(list.next, struct osc_extent, oe_link);
2712                 list_del_init(&ext->oe_link);
2713
2714                 /* extent may be in OES_ACTIVE state because inode mutex
2715                  * is released before osc_io_end() in file write case */
2716                 if (ext->oe_state != OES_TRUNC)
2717                         osc_extent_wait(env, ext, OES_TRUNC);
2718
2719                 rc = osc_extent_truncate(ext, index, partial);
2720                 if (rc < 0) {
2721                         if (result == 0)
2722                                 result = rc;
2723
2724                         OSC_EXTENT_DUMP(D_ERROR, ext,
2725                                         "truncate error %d\n", rc);
2726                 } else if (ext->oe_nr_pages == 0) {
2727                         osc_extent_remove(ext);
2728                 } else {
2729                         /* this must be an overlapped extent which means only
2730                          * part of pages in this extent have been truncated.
2731                          */
2732                         EASSERTF(ext->oe_start <= index, ext,
2733                                  "trunc index = %lu/%d.\n", index, partial);
2734                         /* fix index to skip this partially truncated extent */
2735                         index = ext->oe_end + 1;
2736                         partial = false;
2737
2738                         /* we need to hold this extent in OES_TRUNC state so
2739                          * that no writeback will happen. This is to avoid
2740                          * BUG 17397. */
2741                         LASSERT(oio->oi_trunc == NULL);
2742                         oio->oi_trunc = osc_extent_get(ext);
2743                         OSC_EXTENT_DUMP(D_CACHE, ext,
2744                                         "trunc at "LPU64"\n", size);
2745                 }
2746                 osc_extent_put(env, ext);
2747         }
2748         if (waiting != NULL) {
2749                 int rc;
2750
2751                 /* ignore the result of osc_extent_wait the write initiator
2752                  * should take care of it. */
2753                 rc = osc_extent_wait(env, waiting, OES_INV);
2754                 if (rc < 0)
2755                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2756
2757                 osc_extent_put(env, waiting);
2758                 waiting = NULL;
2759                 goto again;
2760         }
2761         RETURN(result);
2762 }
2763
2764 /**
2765  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2766  */
2767 void osc_cache_truncate_end(const struct lu_env *env, struct osc_io *oio,
2768                             struct osc_object *obj)
2769 {
2770         struct osc_extent *ext = oio->oi_trunc;
2771
2772         oio->oi_trunc = NULL;
2773         if (ext != NULL) {
2774                 bool unplug = false;
2775
2776                 EASSERT(ext->oe_nr_pages > 0, ext);
2777                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2778                 EASSERT(!ext->oe_urgent, ext);
2779
2780                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2781                 osc_object_lock(obj);
2782                 osc_extent_state_set(ext, OES_CACHE);
2783                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2784                         ext->oe_urgent = 1;
2785                         list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2786                         unplug = true;
2787                 }
2788                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2789                 osc_object_unlock(obj);
2790                 osc_extent_put(env, ext);
2791
2792                 if (unplug)
2793                         osc_io_unplug_async(env, osc_cli(obj), obj);
2794         }
2795 }
2796
2797 /**
2798  * Wait for extents in a specific range to be written out.
2799  * The caller must have called osc_cache_writeback_range() to issue IO
2800  * otherwise it will take a long time for this function to finish.
2801  *
2802  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2803  * nobody else can dirty this range of file while we're waiting for
2804  * extents to be written.
2805  */
2806 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2807                          pgoff_t start, pgoff_t end)
2808 {
2809         struct osc_extent *ext;
2810         pgoff_t index = start;
2811         int     result = 0;
2812         ENTRY;
2813
2814 again:
2815         osc_object_lock(obj);
2816         ext = osc_extent_search(obj, index);
2817         if (ext == NULL)
2818                 ext = first_extent(obj);
2819         else if (ext->oe_end < index)
2820                 ext = next_extent(ext);
2821         while (ext != NULL) {
2822                 int rc;
2823
2824                 if (ext->oe_start > end)
2825                         break;
2826
2827                 if (!ext->oe_fsync_wait) {
2828                         ext = next_extent(ext);
2829                         continue;
2830                 }
2831
2832                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2833                              ext->oe_hp || ext->oe_urgent), ext);
2834                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2835                              !ext->oe_hp && ext->oe_urgent), ext);
2836
2837                 index = ext->oe_end + 1;
2838                 osc_extent_get(ext);
2839                 osc_object_unlock(obj);
2840
2841                 rc = osc_extent_wait(env, ext, OES_INV);
2842                 if (result == 0)
2843                         result = rc;
2844                 osc_extent_put(env, ext);
2845                 goto again;
2846         }
2847         osc_object_unlock(obj);
2848
2849         OSC_IO_DEBUG(obj, "sync file range.\n");
2850         RETURN(result);
2851 }
2852
2853 /**
2854  * Called to write out a range of osc object.
2855  *
2856  * @hp     : should be set this is caused by lock cancel;
2857  * @discard: is set if dirty pages should be dropped - file will be deleted or
2858  *         truncated, this implies there is no partially discarding extents.
2859  *
2860  * Return how many pages will be issued, or error code if error occurred.
2861  */
2862 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2863                               pgoff_t start, pgoff_t end, int hp, int discard)
2864 {
2865         struct osc_extent *ext;
2866         struct list_head discard_list = LIST_HEAD_INIT(discard_list);
2867         bool unplug = false;
2868         int result = 0;
2869         ENTRY;
2870
2871         osc_object_lock(obj);
2872         ext = osc_extent_search(obj, start);
2873         if (ext == NULL)
2874                 ext = first_extent(obj);
2875         else if (ext->oe_end < start)
2876                 ext = next_extent(ext);
2877         while (ext != NULL) {
2878                 if (ext->oe_start > end)
2879                         break;
2880
2881                 ext->oe_fsync_wait = 1;
2882                 switch (ext->oe_state) {
2883                 case OES_CACHE:
2884                         result += ext->oe_nr_pages;
2885                         if (!discard) {
2886                                 struct list_head *list = NULL;
2887                                 if (hp) {
2888                                         EASSERT(!ext->oe_hp, ext);
2889                                         ext->oe_hp = 1;
2890                                         list = &obj->oo_hp_exts;
2891                                 } else if (!ext->oe_urgent) {
2892                                         ext->oe_urgent = 1;
2893                                         list = &obj->oo_urgent_exts;
2894                                 }
2895                                 if (list != NULL)
2896                                         list_move_tail(&ext->oe_link, list);
2897                                 unplug = true;
2898                         } else {
2899                                 /* the only discarder is lock cancelling, so
2900                                  * [start, end] must contain this extent */
2901                                 EASSERT(ext->oe_start >= start &&
2902                                         ext->oe_max_end <= end, ext);
2903                                 osc_extent_state_set(ext, OES_LOCKING);
2904                                 ext->oe_owner = current;
2905                                 list_move_tail(&ext->oe_link,
2906                                                    &discard_list);
2907                                 osc_update_pending(obj, OBD_BRW_WRITE,
2908                                                    -ext->oe_nr_pages);
2909                         }
2910                         break;
2911                 case OES_ACTIVE:
2912                         /* It's pretty bad to wait for ACTIVE extents, because
2913                          * we don't know how long we will wait for it to be
2914                          * flushed since it may be blocked at awaiting more
2915                          * grants. We do this for the correctness of fsync. */
2916                         LASSERT(hp == 0 && discard == 0);
2917                         ext->oe_urgent = 1;
2918                         break;
2919                 case OES_TRUNC:
2920                         /* this extent is being truncated, can't do anything
2921                          * for it now. it will be set to urgent after truncate
2922                          * is finished in osc_cache_truncate_end(). */
2923                 default:
2924                         break;
2925                 }
2926                 ext = next_extent(ext);
2927         }
2928         osc_object_unlock(obj);
2929
2930         LASSERT(ergo(!discard, list_empty(&discard_list)));
2931         if (!list_empty(&discard_list)) {
2932                 struct osc_extent *tmp;
2933                 int rc;
2934
2935                 osc_list_maint(osc_cli(obj), obj);
2936                 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
2937                         list_del_init(&ext->oe_link);
2938                         EASSERT(ext->oe_state == OES_LOCKING, ext);
2939
2940                         /* Discard caching pages. We don't actually write this
2941                          * extent out but we complete it as if we did. */
2942                         rc = osc_extent_make_ready(env, ext);
2943                         if (unlikely(rc < 0)) {
2944                                 OSC_EXTENT_DUMP(D_ERROR, ext,
2945                                                 "make_ready returned %d\n", rc);
2946                                 if (result >= 0)
2947                                         result = rc;
2948                         }
2949
2950                         /* finish the extent as if the pages were sent */
2951                         osc_extent_finish(env, ext, 0, 0);
2952                 }
2953         }
2954
2955         if (unplug)
2956                 osc_io_unplug(env, osc_cli(obj), obj, PDL_POLICY_ROUND);
2957
2958         if (hp || discard) {
2959                 int rc;
2960                 rc = osc_cache_wait_range(env, obj, start, end);
2961                 if (result >= 0 && rc < 0)
2962                         result = rc;
2963         }
2964
2965         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
2966         RETURN(result);
2967 }
2968
2969 /**
2970  * Returns a list of pages by a given [start, end] of \a obj.
2971  *
2972  * \param resched If not NULL, then we give up before hogging CPU for too
2973  * long and set *resched = 1, in that case caller should implement a retry
2974  * logic.
2975  *
2976  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
2977  * crucial in the face of [offset, EOF] locks.
2978  *
2979  * Return at least one page in @queue unless there is no covered page.
2980  */
2981 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
2982                         struct osc_object *osc, pgoff_t start, pgoff_t end,
2983                         osc_page_gang_cbt cb, void *cbdata)
2984 {
2985         struct osc_page *ops;
2986         void            **pvec;
2987         pgoff_t         idx;
2988         unsigned int    nr;
2989         unsigned int    i;
2990         unsigned int    j;
2991         int             res = CLP_GANG_OKAY;
2992         bool            tree_lock = true;
2993         ENTRY;
2994
2995         idx = start;
2996         pvec = osc_env_info(env)->oti_pvec;
2997         spin_lock(&osc->oo_tree_lock);
2998         while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
2999                                             idx, OTI_PVEC_SIZE)) > 0) {
3000                 struct cl_page *page;
3001                 bool end_of_region = false;
3002
3003                 for (i = 0, j = 0; i < nr; ++i) {
3004                         ops = pvec[i];
3005                         pvec[i] = NULL;
3006
3007                         idx = osc_index(ops);
3008                         if (idx > end) {
3009                                 end_of_region = true;
3010                                 break;
3011                         }
3012
3013                         page = ops->ops_cl.cpl_page;
3014                         LASSERT(page->cp_type == CPT_CACHEABLE);
3015                         if (page->cp_state == CPS_FREEING)
3016                                 continue;
3017
3018                         cl_page_get(page);
3019                         lu_ref_add_atomic(&page->cp_reference,
3020                                           "gang_lookup", current);
3021                         pvec[j++] = ops;
3022                 }
3023                 ++idx;
3024
3025                 /*
3026                  * Here a delicate locking dance is performed. Current thread
3027                  * holds a reference to a page, but has to own it before it
3028                  * can be placed into queue. Owning implies waiting, so
3029                  * radix-tree lock is to be released. After a wait one has to
3030                  * check that pages weren't truncated (cl_page_own() returns
3031                  * error in the latter case).
3032                  */
3033                 spin_unlock(&osc->oo_tree_lock);
3034                 tree_lock = false;
3035
3036                 for (i = 0; i < j; ++i) {
3037                         ops = pvec[i];
3038                         if (res == CLP_GANG_OKAY)
3039                                 res = (*cb)(env, io, ops, cbdata);
3040
3041                         page = ops->ops_cl.cpl_page;
3042                         lu_ref_del(&page->cp_reference, "gang_lookup", current);
3043                         cl_page_put(env, page);
3044                 }
3045                 if (nr < OTI_PVEC_SIZE || end_of_region)
3046                         break;
3047
3048                 if (res == CLP_GANG_OKAY && need_resched())
3049                         res = CLP_GANG_RESCHED;
3050                 if (res != CLP_GANG_OKAY)
3051                         break;
3052
3053                 spin_lock(&osc->oo_tree_lock);
3054                 tree_lock = true;
3055         }
3056         if (tree_lock)
3057                 spin_unlock(&osc->oo_tree_lock);
3058         RETURN(res);
3059 }
3060
3061 /**
3062  * Check if page @page is covered by an extra lock or discard it.
3063  */
3064 static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3065                                 struct osc_page *ops, void *cbdata)
3066 {
3067         struct osc_thread_info *info = osc_env_info(env);
3068         struct cl_lock *lock = cbdata;
3069         pgoff_t index;
3070
3071         index = osc_index(ops);
3072         if (index >= info->oti_fn_index) {
3073                 struct cl_lock *tmp;
3074                 struct cl_page *page = ops->ops_cl.cpl_page;
3075
3076                 /* refresh non-overlapped index */
3077                 tmp = cl_lock_at_pgoff(env, lock->cll_descr.cld_obj, index,
3078                                        lock, 1, 0);
3079                 if (tmp != NULL) {
3080                         /* Cache the first-non-overlapped index so as to skip
3081                          * all pages within [index, oti_fn_index). This
3082                          * is safe because if tmp lock is canceled, it will
3083                          * discard these pages. */
3084                         info->oti_fn_index = tmp->cll_descr.cld_end + 1;
3085                         if (tmp->cll_descr.cld_end == CL_PAGE_EOF)
3086                                 info->oti_fn_index = CL_PAGE_EOF;
3087                         cl_lock_put(env, tmp);
3088                 } else if (cl_page_own(env, io, page) == 0) {
3089                         /* discard the page */
3090                         cl_page_discard(env, io, page);
3091                         cl_page_disown(env, io, page);
3092                 } else {
3093                         LASSERT(page->cp_state == CPS_FREEING);
3094                 }
3095         }
3096
3097         info->oti_next_index = index + 1;
3098         return CLP_GANG_OKAY;
3099 }
3100
3101 static int discard_cb(const struct lu_env *env, struct cl_io *io,
3102                       struct osc_page *ops, void *cbdata)
3103 {
3104         struct osc_thread_info *info = osc_env_info(env);
3105         struct cl_lock *lock = cbdata;
3106         struct cl_page *page = ops->ops_cl.cpl_page;
3107
3108         LASSERT(lock->cll_descr.cld_mode >= CLM_WRITE);
3109
3110         /* page is top page. */
3111         info->oti_next_index = osc_index(ops) + 1;
3112         if (cl_page_own(env, io, page) == 0) {
3113                 KLASSERT(ergo(page->cp_type == CPT_CACHEABLE,
3114                               !PageDirty(cl_page_vmpage(page))));
3115
3116                 /* discard the page */
3117                 cl_page_discard(env, io, page);
3118                 cl_page_disown(env, io, page);
3119         } else {
3120                 LASSERT(page->cp_state == CPS_FREEING);
3121         }
3122
3123         return CLP_GANG_OKAY;
3124 }
3125
3126 /**
3127  * Discard pages protected by the given lock. This function traverses radix
3128  * tree to find all covering pages and discard them. If a page is being covered
3129  * by other locks, it should remain in cache.
3130  *
3131  * If error happens on any step, the process continues anyway (the reasoning
3132  * behind this being that lock cancellation cannot be delayed indefinitely).
3133  */
3134 int osc_lock_discard_pages(const struct lu_env *env, struct osc_lock *ols)
3135 {
3136         struct osc_thread_info *info = osc_env_info(env);
3137         struct cl_io *io = &info->oti_io;
3138         struct cl_object *osc = ols->ols_cl.cls_obj;
3139         struct cl_lock *lock = ols->ols_cl.cls_lock;
3140         struct cl_lock_descr *descr = &lock->cll_descr;
3141         osc_page_gang_cbt cb;
3142         int res;
3143         int result;
3144
3145         ENTRY;
3146
3147         io->ci_obj = cl_object_top(osc);
3148         io->ci_ignore_layout = 1;
3149         result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3150         if (result != 0)
3151                 GOTO(out, result);
3152
3153         cb = descr->cld_mode == CLM_READ ? check_and_discard_cb : discard_cb;
3154         info->oti_fn_index = info->oti_next_index = descr->cld_start;
3155         do {
3156                 res = osc_page_gang_lookup(env, io, cl2osc(osc),
3157                                            info->oti_next_index, descr->cld_end,
3158                                            cb, (void *)lock);
3159                 if (info->oti_next_index > descr->cld_end)
3160                         break;
3161
3162                 if (res == CLP_GANG_RESCHED)
3163                         cond_resched();
3164         } while (res != CLP_GANG_OKAY);
3165 out:
3166         cl_io_fini(env, io);
3167         RETURN(result);
3168 }
3169
3170
3171 /** @} osc */