drivers/staging/lustre/lustre/osc/osc_cache.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2015, Intel Corporation.
  27  *
  28  */
  29 /*
  30  * This file is part of Lustre, http://www.lustre.org/
  31  * Lustre is a trademark of Sun Microsystems, Inc.
  32  *
  33  * osc cache management.
  34  *
  35  * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_OSC
  39
  40 #include "osc_cl_internal.h"
  41 #include "osc_internal.h"
  42
  43 static int extent_debug; /* set it to be true for more debug */
  44
  45 static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
  46 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
  47                            enum osc_extent_state state);
  48 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
  49                               struct osc_async_page *oap, int sent, int rc);
  50 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
  51                           int cmd);
  52 static int osc_refresh_count(const struct lu_env *env,
  53                              struct osc_async_page *oap, int cmd);
  54 static int osc_io_unplug_async(const struct lu_env *env,
  55                                struct client_obd *cli, struct osc_object *osc);
  56 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
  57                            unsigned int lost_grant);
  58
  59 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
  60                                   const char *func, int line);
  61 #define osc_extent_tree_dump(lvl, obj) \
  62         osc_extent_tree_dump0(lvl, obj, __func__, __LINE__)
  63
  64 /** \addtogroup osc
  65  *  @{
  66  */
  67
  68 /* ------------------ osc extent ------------------ */
  69 static inline char *ext_flags(struct osc_extent *ext, char *flags)
  70 {
  71         char *buf = flags;
  72         *buf++ = ext->oe_rw ? 'r' : 'w';
  73         if (ext->oe_intree)
  74                 *buf++ = 'i';
  75         if (ext->oe_sync)
  76                 *buf++ = 'S';
  77         if (ext->oe_srvlock)
  78                 *buf++ = 's';
  79         if (ext->oe_hp)
  80                 *buf++ = 'h';
  81         if (ext->oe_urgent)
  82                 *buf++ = 'u';
  83         if (ext->oe_memalloc)
  84                 *buf++ = 'm';
  85         if (ext->oe_trunc_pending)
  86                 *buf++ = 't';
  87         if (ext->oe_fsync_wait)
  88                 *buf++ = 'Y';
  89         *buf = 0;
  90         return flags;
  91 }
  92
  93 static inline char list_empty_marker(struct list_head *list)
  94 {
  95         return list_empty(list) ? '-' : '+';
  96 }
  97
  98 #define EXTSTR       "[%lu -> %lu/%lu]"
  99 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
 100 static const char *oes_strings[] = {
 101         "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
 102
 103 #define OSC_EXTENT_DUMP(lvl, extent, fmt, ...) do {                           \
 104         struct osc_extent *__ext = (extent);                                  \
 105         char __buf[16];                                                       \
 106                                                                               \
 107         CDEBUG(lvl,                                                           \
 108                 "extent %p@{" EXTSTR ", "                                     \
 109                 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt,          \
 110                 /* ----- extent part 0 ----- */                               \
 111                 __ext, EXTPARA(__ext),                                        \
 112                 /* ----- part 1 ----- */                                      \
 113                 atomic_read(&__ext->oe_refc),                                 \
 114                 atomic_read(&__ext->oe_users),                                \
 115                 list_empty_marker(&__ext->oe_link),                           \
 116                 oes_strings[__ext->oe_state], ext_flags(__ext, __buf),        \
 117                 __ext->oe_obj,                                                \
 118                 /* ----- part 2 ----- */                                      \
 119                 __ext->oe_grants, __ext->oe_nr_pages,                         \
 120                 list_empty_marker(&__ext->oe_pages),                          \
 121                 waitqueue_active(&__ext->oe_waitq) ? '+' : '-',               \
 122                 __ext->oe_dlmlock, __ext->oe_mppr, __ext->oe_owner,           \
 123                 /* ----- part 4 ----- */                                      \
 124                 ## __VA_ARGS__);                                              \
 125         if (lvl == D_ERROR && __ext->oe_dlmlock)                              \
 126                 LDLM_ERROR(__ext->oe_dlmlock, "extent: %p", __ext);           \
 127         else                                                                  \
 128                 LDLM_DEBUG(__ext->oe_dlmlock, "extent: %p", __ext);           \
 129 } while (0)
 130
 131 #undef EASSERTF
 132 #define EASSERTF(expr, ext, fmt, args...) do {                          \
 133         if (!(expr)) {                                                  \
 134                 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args);           \
 135                 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj);           \
 136                 LASSERT(expr);                                          \
 137         }                                                               \
 138 } while (0)
 139
 140 #undef EASSERT
 141 #define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
 142
 143 static inline struct osc_extent *rb_extent(struct rb_node *n)
 144 {
 145         return rb_entry_safe(n, struct osc_extent, oe_node);
 146 }
 147
 148 static inline struct osc_extent *next_extent(struct osc_extent *ext)
 149 {
 150         if (!ext)
 151                 return NULL;
 152
 153         LASSERT(ext->oe_intree);
 154         return rb_extent(rb_next(&ext->oe_node));
 155 }
 156
 157 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
 158 {
 159         if (!ext)
 160                 return NULL;
 161
 162         LASSERT(ext->oe_intree);
 163         return rb_extent(rb_prev(&ext->oe_node));
 164 }
 165
 166 static inline struct osc_extent *first_extent(struct osc_object *obj)
 167 {
 168         return rb_extent(rb_first(&obj->oo_root));
 169 }
 170
 171 /* object must be locked by caller. */
 172 static int osc_extent_sanity_check0(struct osc_extent *ext,
 173                                     const char *func, const int line)
 174 {
 175         struct osc_object *obj = ext->oe_obj;
 176         struct osc_async_page *oap;
 177         size_t page_count;
 178         int rc = 0;
 179
 180         if (!osc_object_is_locked(obj)) {
 181                 rc = 9;
 182                 goto out;
 183         }
 184
 185         if (ext->oe_state >= OES_STATE_MAX) {
 186                 rc = 10;
 187                 goto out;
 188         }
 189
 190         if (atomic_read(&ext->oe_refc) <= 0) {
 191                 rc = 20;
 192                 goto out;
 193         }
 194
 195         if (atomic_read(&ext->oe_refc) < atomic_read(&ext->oe_users)) {
 196                 rc = 30;
 197                 goto out;
 198         }
 199
 200         switch (ext->oe_state) {
 201         case OES_INV:
 202                 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
 203                         rc = 35;
 204                 else
 205                         rc = 0;
 206                 goto out;
 207         case OES_ACTIVE:
 208                 if (atomic_read(&ext->oe_users) == 0) {
 209                         rc = 40;
 210                         goto out;
 211                 }
 212                 if (ext->oe_hp) {
 213                         rc = 50;
 214                         goto out;
 215                 }
 216                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
 217                         rc = 55;
 218                         goto out;
 219                 }
 220                 break;
 221         case OES_CACHE:
 222                 if (ext->oe_grants == 0) {
 223                         rc = 60;
 224                         goto out;
 225                 }
 226                 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp) {
 227                         rc = 65;
 228                         goto out;
 229                 }
 230         default:
 231                 if (atomic_read(&ext->oe_users) > 0) {
 232                         rc = 70;
 233                         goto out;
 234                 }
 235         }
 236
 237         if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start) {
 238                 rc = 80;
 239                 goto out;
 240         }
 241
 242         if (ext->oe_sync && ext->oe_grants > 0) {
 243                 rc = 90;
 244                 goto out;
 245         }
 246
 247         if (ext->oe_dlmlock && !ldlm_is_failed(ext->oe_dlmlock)) {
 248                 struct ldlm_extent *extent;
 249
 250                 extent = &ext->oe_dlmlock->l_policy_data.l_extent;
 251                 if (!(extent->start <= cl_offset(osc2cl(obj), ext->oe_start) &&
 252                       extent->end >= cl_offset(osc2cl(obj), ext->oe_max_end))) {
 253                         rc = 100;
 254                         goto out;
 255                 }
 256
 257                 if (!(ext->oe_dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP))) {
 258                         rc = 102;
 259                         goto out;
 260                 }
 261         }
 262
 263         if (ext->oe_nr_pages > ext->oe_mppr) {
 264                 rc = 105;
 265                 goto out;
 266         }
 267
 268         /* Do not verify page list if extent is in RPC. This is because an
 269          * in-RPC extent is supposed to be exclusively accessible w/o lock.
 270          */
 271         if (ext->oe_state > OES_CACHE) {
 272                 rc = 0;
 273                 goto out;
 274         }
 275
 276         if (!extent_debug) {
 277                 rc = 0;
 278                 goto out;
 279         }
 280
 281         page_count = 0;
 282         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
 283                 pgoff_t index = osc_index(oap2osc(oap));
 284                 ++page_count;
 285                 if (index > ext->oe_end || index < ext->oe_start) {
 286                         rc = 110;
 287                         goto out;
 288                 }
 289         }
 290         if (page_count != ext->oe_nr_pages) {
 291                 rc = 120;
 292                 goto out;
 293         }
 294
 295 out:
 296         if (rc != 0)
 297                 OSC_EXTENT_DUMP(D_ERROR, ext,
 298                                 "%s:%d sanity check %p failed with rc = %d\n",
 299                                 func, line, ext, rc);
 300         return rc;
 301 }
 302
 303 #define sanity_check_nolock(ext) \
 304         osc_extent_sanity_check0(ext, __func__, __LINE__)
 305
 306 #define sanity_check(ext) ({                                            \
 307         int __res;                                                      \
 308         osc_object_lock((ext)->oe_obj);                                 \
 309         __res = sanity_check_nolock(ext);                               \
 310         osc_object_unlock((ext)->oe_obj);                               \
 311         __res;                                                          \
 312 })
 313
 314 /**
 315  * sanity check - to make sure there is no overlapped extent in the tree.
 316  */
 317 static int osc_extent_is_overlapped(struct osc_object *obj,
 318                                     struct osc_extent *ext)
 319 {
 320         struct osc_extent *tmp;
 321
 322         LASSERT(osc_object_is_locked(obj));
 323
 324         if (!extent_debug)
 325                 return 0;
 326
 327         for (tmp = first_extent(obj); tmp; tmp = next_extent(tmp)) {
 328                 if (tmp == ext)
 329                         continue;
 330                 if (tmp->oe_end >= ext->oe_start &&
 331                     tmp->oe_start <= ext->oe_end)
 332                         return 1;
 333         }
 334         return 0;
 335 }
 336
 337 static void osc_extent_state_set(struct osc_extent *ext, int state)
 338 {
 339         LASSERT(osc_object_is_locked(ext->oe_obj));
 340         LASSERT(state >= OES_INV && state < OES_STATE_MAX);
 341
 342         /* Never try to sanity check a state changing extent :-) */
 343         /* LASSERT(sanity_check_nolock(ext) == 0); */
 344
 345         /* TODO: validate the state machine */
 346         ext->oe_state = state;
 347         wake_up_all(&ext->oe_waitq);
 348 }
 349
 350 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
 351 {
 352         struct osc_extent *ext;
 353
 354         ext = kmem_cache_zalloc(osc_extent_kmem, GFP_NOFS);
 355         if (!ext)
 356                 return NULL;
 357
 358         RB_CLEAR_NODE(&ext->oe_node);
 359         ext->oe_obj = obj;
 360         cl_object_get(osc2cl(obj));
 361         atomic_set(&ext->oe_refc, 1);
 362         atomic_set(&ext->oe_users, 0);
 363         INIT_LIST_HEAD(&ext->oe_link);
 364         ext->oe_state = OES_INV;
 365         INIT_LIST_HEAD(&ext->oe_pages);
 366         init_waitqueue_head(&ext->oe_waitq);
 367         ext->oe_dlmlock = NULL;
 368
 369         return ext;
 370 }
 371
 372 static void osc_extent_free(struct osc_extent *ext)
 373 {
 374         kmem_cache_free(osc_extent_kmem, ext);
 375 }
 376
 377 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
 378 {
 379         LASSERT(atomic_read(&ext->oe_refc) >= 0);
 380         atomic_inc(&ext->oe_refc);
 381         return ext;
 382 }
 383
 384 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
 385 {
 386         LASSERT(atomic_read(&ext->oe_refc) > 0);
 387         if (atomic_dec_and_test(&ext->oe_refc)) {
 388                 LASSERT(list_empty(&ext->oe_link));
 389                 LASSERT(atomic_read(&ext->oe_users) == 0);
 390                 LASSERT(ext->oe_state == OES_INV);
 391                 LASSERT(!ext->oe_intree);
 392
 393                 if (ext->oe_dlmlock) {
 394                         lu_ref_add(&ext->oe_dlmlock->l_reference,
 395                                    "osc_extent", ext);
 396                         LDLM_LOCK_PUT(ext->oe_dlmlock);
 397                         ext->oe_dlmlock = NULL;
 398                 }
 399                 cl_object_put(env, osc2cl(ext->oe_obj));
 400                 osc_extent_free(ext);
 401         }
 402 }
 403
 404 /**
 405  * osc_extent_put_trust() is a special version of osc_extent_put() when
 406  * it's known that the caller is not the last user. This is to address the
 407  * problem of lacking of lu_env ;-).
 408  */
 409 static void osc_extent_put_trust(struct osc_extent *ext)
 410 {
 411         LASSERT(atomic_read(&ext->oe_refc) > 1);
 412         LASSERT(osc_object_is_locked(ext->oe_obj));
 413         atomic_dec(&ext->oe_refc);
 414 }
 415
 416 /**
 417  * Return the extent which includes pgoff @index, or return the greatest
 418  * previous extent in the tree.
 419  */
 420 static struct osc_extent *osc_extent_search(struct osc_object *obj,
 421                                             pgoff_t index)
 422 {
 423         struct rb_node *n = obj->oo_root.rb_node;
 424         struct osc_extent *tmp, *p = NULL;
 425
 426         LASSERT(osc_object_is_locked(obj));
 427         while (n) {
 428                 tmp = rb_extent(n);
 429                 if (index < tmp->oe_start) {
 430                         n = n->rb_left;
 431                 } else if (index > tmp->oe_end) {
 432                         p = rb_extent(n);
 433                         n = n->rb_right;
 434                 } else {
 435                         return tmp;
 436                 }
 437         }
 438         return p;
 439 }
 440
 441 /*
 442  * Return the extent covering @index, otherwise return NULL.
 443  * caller must have held object lock.
 444  */
 445 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
 446                                             pgoff_t index)
 447 {
 448         struct osc_extent *ext;
 449
 450         ext = osc_extent_search(obj, index);
 451         if (ext && ext->oe_start <= index && index <= ext->oe_end)
 452                 return osc_extent_get(ext);
 453         return NULL;
 454 }
 455
 456 /* caller must have held object lock. */
 457 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
 458 {
 459         struct rb_node **n = &obj->oo_root.rb_node;
 460         struct rb_node *parent = NULL;
 461         struct osc_extent *tmp;
 462
 463         LASSERT(ext->oe_intree == 0);
 464         LASSERT(ext->oe_obj == obj);
 465         LASSERT(osc_object_is_locked(obj));
 466         while (*n) {
 467                 tmp = rb_extent(*n);
 468                 parent = *n;
 469
 470                 if (ext->oe_end < tmp->oe_start)
 471                         n = &(*n)->rb_left;
 472                 else if (ext->oe_start > tmp->oe_end)
 473                         n = &(*n)->rb_right;
 474                 else
 475                         EASSERTF(0, tmp, EXTSTR "\n", EXTPARA(ext));
 476         }
 477         rb_link_node(&ext->oe_node, parent, n);
 478         rb_insert_color(&ext->oe_node, &obj->oo_root);
 479         osc_extent_get(ext);
 480         ext->oe_intree = 1;
 481 }
 482
 483 /* caller must have held object lock. */
 484 static void osc_extent_erase(struct osc_extent *ext)
 485 {
 486         struct osc_object *obj = ext->oe_obj;
 487
 488         LASSERT(osc_object_is_locked(obj));
 489         if (ext->oe_intree) {
 490                 rb_erase(&ext->oe_node, &obj->oo_root);
 491                 ext->oe_intree = 0;
 492                 /* rbtree held a refcount */
 493                 osc_extent_put_trust(ext);
 494         }
 495 }
 496
 497 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
 498 {
 499         struct osc_object *obj = ext->oe_obj;
 500
 501         LASSERT(osc_object_is_locked(obj));
 502         LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
 503         if (ext->oe_state == OES_CACHE) {
 504                 osc_extent_state_set(ext, OES_ACTIVE);
 505                 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
 506         }
 507         atomic_inc(&ext->oe_users);
 508         list_del_init(&ext->oe_link);
 509         return osc_extent_get(ext);
 510 }
 511
 512 static void __osc_extent_remove(struct osc_extent *ext)
 513 {
 514         LASSERT(osc_object_is_locked(ext->oe_obj));
 515         LASSERT(list_empty(&ext->oe_pages));
 516         osc_extent_erase(ext);
 517         list_del_init(&ext->oe_link);
 518         osc_extent_state_set(ext, OES_INV);
 519         OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
 520 }
 521
 522 static void osc_extent_remove(struct osc_extent *ext)
 523 {
 524         struct osc_object *obj = ext->oe_obj;
 525
 526         osc_object_lock(obj);
 527         __osc_extent_remove(ext);
 528         osc_object_unlock(obj);
 529 }
 530
 531 /**
 532  * This function is used to merge extents to get better performance. It checks
 533  * if @cur and @victim are contiguous at chunk level.
 534  */
 535 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
 536                             struct osc_extent *victim)
 537 {
 538         struct osc_object *obj = cur->oe_obj;
 539         pgoff_t chunk_start;
 540         pgoff_t chunk_end;
 541         int ppc_bits;
 542
 543         LASSERT(cur->oe_state == OES_CACHE);
 544         LASSERT(osc_object_is_locked(obj));
 545         if (!victim)
 546                 return -EINVAL;
 547
 548         if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
 549                 return -EBUSY;
 550
 551         if (cur->oe_max_end != victim->oe_max_end)
 552                 return -ERANGE;
 553
 554         LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
 555         ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_SHIFT;
 556         chunk_start = cur->oe_start >> ppc_bits;
 557         chunk_end = cur->oe_end >> ppc_bits;
 558         if (chunk_start != (victim->oe_end >> ppc_bits) + 1 &&
 559             chunk_end + 1 != victim->oe_start >> ppc_bits)
 560                 return -ERANGE;
 561
 562         OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
 563
 564         cur->oe_start = min(cur->oe_start, victim->oe_start);
 565         cur->oe_end = max(cur->oe_end, victim->oe_end);
 566         cur->oe_grants += victim->oe_grants;
 567         cur->oe_nr_pages += victim->oe_nr_pages;
 568         /* only the following bits are needed to merge */
 569         cur->oe_urgent |= victim->oe_urgent;
 570         cur->oe_memalloc |= victim->oe_memalloc;
 571         list_splice_init(&victim->oe_pages, &cur->oe_pages);
 572         list_del_init(&victim->oe_link);
 573         victim->oe_nr_pages = 0;
 574
 575         osc_extent_get(victim);
 576         __osc_extent_remove(victim);
 577         osc_extent_put(env, victim);
 578
 579         OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
 580         return 0;
 581 }
 582
 583 /**
 584  * Drop user count of osc_extent, and unplug IO asynchronously.
 585  */
 586 void osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
 587 {
 588         struct osc_object *obj = ext->oe_obj;
 589
 590         LASSERT(atomic_read(&ext->oe_users) > 0);
 591         LASSERT(sanity_check(ext) == 0);
 592         LASSERT(ext->oe_grants > 0);
 593
 594         if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
 595                 LASSERT(ext->oe_state == OES_ACTIVE);
 596                 if (ext->oe_trunc_pending) {
 597                         /* a truncate process is waiting for this extent.
 598                          * This may happen due to a race, check
 599                          * osc_cache_truncate_start().
 600                          */
 601                         osc_extent_state_set(ext, OES_TRUNC);
 602                         ext->oe_trunc_pending = 0;
 603                 } else {
 604                         osc_extent_state_set(ext, OES_CACHE);
 605                         osc_update_pending(obj, OBD_BRW_WRITE,
 606                                            ext->oe_nr_pages);
 607
 608                         /* try to merge the previous and next extent. */
 609                         osc_extent_merge(env, ext, prev_extent(ext));
 610                         osc_extent_merge(env, ext, next_extent(ext));
 611
 612                         if (ext->oe_urgent)
 613                                 list_move_tail(&ext->oe_link,
 614                                                &obj->oo_urgent_exts);
 615                 }
 616                 osc_object_unlock(obj);
 617
 618                 osc_io_unplug_async(env, osc_cli(obj), obj);
 619         }
 620         osc_extent_put(env, ext);
 621 }
 622
 623 static inline int overlapped(struct osc_extent *ex1, struct osc_extent *ex2)
 624 {
 625         return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
 626 }
 627
 628 /**
 629  * Find or create an extent which includes @index, core function to manage
 630  * extent tree.
 631  */
 632 static struct osc_extent *osc_extent_find(const struct lu_env *env,
 633                                           struct osc_object *obj, pgoff_t index,
 634                                           unsigned int *grants)
 635 {
 636         struct client_obd *cli = osc_cli(obj);
 637         struct osc_lock   *olck;
 638         struct cl_lock_descr *descr;
 639         struct osc_extent *cur;
 640         struct osc_extent *ext;
 641         struct osc_extent *conflict = NULL;
 642         struct osc_extent *found = NULL;
 643         pgoff_t chunk;
 644         pgoff_t max_end;
 645         unsigned int max_pages; /* max_pages_per_rpc */
 646         unsigned int chunksize;
 647         int ppc_bits; /* pages per chunk bits */
 648         pgoff_t chunk_mask;
 649         int rc;
 650
 651         cur = osc_extent_alloc(obj);
 652         if (!cur)
 653                 return ERR_PTR(-ENOMEM);
 654
 655         olck = osc_env_io(env)->oi_write_osclock;
 656         LASSERTF(olck, "page %lu is not covered by lock\n", index);
 657         LASSERT(olck->ols_state == OLS_GRANTED);
 658
 659         descr = &olck->ols_cl.cls_lock->cll_descr;
 660         LASSERT(descr->cld_mode >= CLM_WRITE);
 661
 662         LASSERT(cli->cl_chunkbits >= PAGE_SHIFT);
 663         ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
 664         chunk_mask = ~((1 << ppc_bits) - 1);
 665         chunksize = 1 << cli->cl_chunkbits;
 666         chunk = index >> ppc_bits;
 667
 668         /* align end to rpc edge, rpc size may not be a power 2 integer. */
 669         max_pages = cli->cl_max_pages_per_rpc;
 670         LASSERT((max_pages & ~chunk_mask) == 0);
 671         max_end = index - (index % max_pages) + max_pages - 1;
 672         max_end = min_t(pgoff_t, max_end, descr->cld_end);
 673
 674         /* initialize new extent by parameters so far */
 675         cur->oe_max_end = max_end;
 676         cur->oe_start = index & chunk_mask;
 677         cur->oe_end = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
 678         if (cur->oe_start < descr->cld_start)
 679                 cur->oe_start = descr->cld_start;
 680         if (cur->oe_end > max_end)
 681                 cur->oe_end = max_end;
 682         cur->oe_grants = 0;
 683         cur->oe_mppr = max_pages;
 684         if (olck->ols_dlmlock) {
 685                 LASSERT(olck->ols_hold);
 686                 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
 687                 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
 688         }
 689
 690         /* grants has been allocated by caller */
 691         LASSERTF(*grants >= chunksize + cli->cl_extent_tax,
 692                  "%u/%u/%u.\n", *grants, chunksize, cli->cl_extent_tax);
 693         LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR "\n",
 694                  EXTPARA(cur));
 695
 696 restart:
 697         osc_object_lock(obj);
 698         ext = osc_extent_search(obj, cur->oe_start);
 699         if (!ext)
 700                 ext = first_extent(obj);
 701         while (ext) {
 702                 pgoff_t ext_chk_start = ext->oe_start >> ppc_bits;
 703                 pgoff_t ext_chk_end = ext->oe_end >> ppc_bits;
 704
 705                 LASSERT(sanity_check_nolock(ext) == 0);
 706                 if (chunk > ext_chk_end + 1)
 707                         break;
 708
 709                 /* if covering by different locks, no chance to match */
 710                 if (olck->ols_dlmlock != ext->oe_dlmlock) {
 711                         EASSERTF(!overlapped(ext, cur), ext,
 712                                  EXTSTR "\n", EXTPARA(cur));
 713
 714                         ext = next_extent(ext);
 715                         continue;
 716                 }
 717
 718                 /* discontiguous chunks? */
 719                 if (chunk + 1 < ext_chk_start) {
 720                         ext = next_extent(ext);
 721                         continue;
 722                 }
 723
 724                 /* ok, from now on, ext and cur have these attrs:
 725                  * 1. covered by the same lock
 726                  * 2. contiguous at chunk level or overlapping.
 727                  */
 728
 729                 if (overlapped(ext, cur)) {
 730                         /* cur is the minimum unit, so overlapping means
 731                          * full contain.
 732                          */
 733                         EASSERTF((ext->oe_start <= cur->oe_start &&
 734                                   ext->oe_end >= cur->oe_end),
 735                                  ext, EXTSTR "\n", EXTPARA(cur));
 736
 737                         if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
 738                                 /* for simplicity, we wait for this extent to
 739                                  * finish before going forward.
 740                                  */
 741                                 conflict = osc_extent_get(ext);
 742                                 break;
 743                         }
 744
 745                         found = osc_extent_hold(ext);
 746                         break;
 747                 }
 748
 749                 /* non-overlapped extent */
 750                 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
 751                         /* we can't do anything for a non OES_CACHE extent, or
 752                          * if there is someone waiting for this extent to be
 753                          * flushed, try next one.
 754                          */
 755                         ext = next_extent(ext);
 756                         continue;
 757                 }
 758
 759                 /* check if they belong to the same rpc slot before trying to
 760                  * merge. the extents are not overlapped and contiguous at
 761                  * chunk level to get here.
 762                  */
 763                 if (ext->oe_max_end != max_end) {
 764                         /* if they don't belong to the same RPC slot or
 765                          * max_pages_per_rpc has ever changed, do not merge.
 766                          */
 767                         ext = next_extent(ext);
 768                         continue;
 769                 }
 770
 771                 /* it's required that an extent must be contiguous at chunk
 772                  * level so that we know the whole extent is covered by grant
 773                  * (the pages in the extent are NOT required to be contiguous).
 774                  * Otherwise, it will be too much difficult to know which
 775                  * chunks have grants allocated.
 776                  */
 777
 778                 /* try to do front merge - extend ext's start */
 779                 if (chunk + 1 == ext_chk_start) {
 780                         /* ext must be chunk size aligned */
 781                         EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
 782
 783                         /* pull ext's start back to cover cur */
 784                         ext->oe_start = cur->oe_start;
 785                         ext->oe_grants += chunksize;
 786                         LASSERT(*grants >= chunksize);
 787                         *grants -= chunksize;
 788
 789                         found = osc_extent_hold(ext);
 790                 } else if (chunk == ext_chk_end + 1) {
 791                         /* rear merge */
 792                         ext->oe_end = cur->oe_end;
 793                         ext->oe_grants += chunksize;
 794                         LASSERT(*grants >= chunksize);
 795                         *grants -= chunksize;
 796
 797                         /* try to merge with the next one because we just fill
 798                          * in a gap
 799                          */
 800                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 801                                 /* we can save extent tax from next extent */
 802                                 *grants += cli->cl_extent_tax;
 803
 804                         found = osc_extent_hold(ext);
 805                 }
 806                 if (found)
 807                         break;
 808
 809                 ext = next_extent(ext);
 810         }
 811
 812         osc_extent_tree_dump(D_CACHE, obj);
 813         if (found) {
 814                 LASSERT(!conflict);
 815                 if (!IS_ERR(found)) {
 816                         LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
 817                         OSC_EXTENT_DUMP(D_CACHE, found,
 818                                         "found caching ext for %lu.\n", index);
 819                 }
 820         } else if (!conflict) {
 821                 /* create a new extent */
 822                 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
 823                 cur->oe_grants = chunksize + cli->cl_extent_tax;
 824                 LASSERT(*grants >= cur->oe_grants);
 825                 *grants -= cur->oe_grants;
 826
 827                 cur->oe_state = OES_CACHE;
 828                 found = osc_extent_hold(cur);
 829                 osc_extent_insert(obj, cur);
 830                 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
 831                                 index, descr->cld_end);
 832         }
 833         osc_object_unlock(obj);
 834
 835         if (conflict) {
 836                 LASSERT(!found);
 837
 838                 /* waiting for IO to finish. Please notice that it's impossible
 839                  * to be an OES_TRUNC extent.
 840                  */
 841                 rc = osc_extent_wait(env, conflict, OES_INV);
 842                 osc_extent_put(env, conflict);
 843                 conflict = NULL;
 844                 if (rc < 0) {
 845                         found = ERR_PTR(rc);
 846                         goto out;
 847                 }
 848
 849                 goto restart;
 850         }
 851
 852 out:
 853         osc_extent_put(env, cur);
 854         return found;
 855 }
 856
 857 /**
 858  * Called when IO is finished to an extent.
 859  */
 860 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
 861                       int sent, int rc)
 862 {
 863         struct client_obd *cli = osc_cli(ext->oe_obj);
 864         struct osc_async_page *oap;
 865         struct osc_async_page *tmp;
 866         int nr_pages = ext->oe_nr_pages;
 867         int lost_grant = 0;
 868         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
 869         __u64 last_off = 0;
 870         int last_count = -1;
 871
 872         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
 873
 874         ext->oe_rc = rc ?: ext->oe_nr_pages;
 875         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
 876
 877         osc_lru_add_batch(cli, &ext->oe_pages);
 878         list_for_each_entry_safe(oap, tmp, &ext->oe_pages, oap_pending_item) {
 879                 list_del_init(&oap->oap_rpc_item);
 880                 list_del_init(&oap->oap_pending_item);
 881                 if (last_off <= oap->oap_obj_off) {
 882                         last_off = oap->oap_obj_off;
 883                         last_count = oap->oap_count;
 884                 }
 885
 886                 --ext->oe_nr_pages;
 887                 osc_ap_completion(env, cli, oap, sent, rc);
 888         }
 889         EASSERT(ext->oe_nr_pages == 0, ext);
 890
 891         if (!sent) {
 892                 lost_grant = ext->oe_grants;
 893         } else if (blocksize < PAGE_SIZE &&
 894                    last_count != PAGE_SIZE) {
 895                 /* For short writes we shouldn't count parts of pages that
 896                  * span a whole chunk on the OST side, or our accounting goes
 897                  * wrong.  Should match the code in filter_grant_check.
 898                  */
 899                 int offset = last_off & ~PAGE_MASK;
 900                 int count = last_count + (offset & (blocksize - 1));
 901                 int end = (offset + last_count) & (blocksize - 1);
 902
 903                 if (end)
 904                         count += blocksize - end;
 905
 906                 lost_grant = PAGE_SIZE - count;
 907         }
 908         if (ext->oe_grants > 0)
 909                 osc_free_grant(cli, nr_pages, lost_grant);
 910
 911         osc_extent_remove(ext);
 912         /* put the refcount for RPC */
 913         osc_extent_put(env, ext);
 914         return 0;
 915 }
 916
 917 static int extent_wait_cb(struct osc_extent *ext, enum osc_extent_state state)
 918 {
 919         int ret;
 920
 921         osc_object_lock(ext->oe_obj);
 922         ret = ext->oe_state == state;
 923         osc_object_unlock(ext->oe_obj);
 924
 925         return ret;
 926 }
 927
 928 /**
 929  * Wait for the extent's state to become @state.
 930  */
 931 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
 932                            enum osc_extent_state state)
 933 {
 934         struct osc_object *obj = ext->oe_obj;
 935         struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
 936                                                   LWI_ON_SIGNAL_NOOP, NULL);
 937         int rc = 0;
 938
 939         osc_object_lock(obj);
 940         LASSERT(sanity_check_nolock(ext) == 0);
 941         /* `Kick' this extent only if the caller is waiting for it to be
 942          * written out.
 943          */
 944         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp &&
 945             !ext->oe_trunc_pending) {
 946                 if (ext->oe_state == OES_ACTIVE) {
 947                         ext->oe_urgent = 1;
 948                 } else if (ext->oe_state == OES_CACHE) {
 949                         ext->oe_urgent = 1;
 950                         osc_extent_hold(ext);
 951                         rc = 1;
 952                 }
 953         }
 954         osc_object_unlock(obj);
 955         if (rc == 1)
 956                 osc_extent_release(env, ext);
 957
 958         /* wait for the extent until its state becomes @state */
 959         rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
 960         if (rc == -ETIMEDOUT) {
 961                 OSC_EXTENT_DUMP(D_ERROR, ext,
 962                                 "%s: wait ext to %u timedout, recovery in progress?\n",
 963                                 cli_name(osc_cli(obj)), state);
 964
 965                 lwi = LWI_INTR(NULL, NULL);
 966                 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state),
 967                                   &lwi);
 968         }
 969         if (rc == 0 && ext->oe_rc < 0)
 970                 rc = ext->oe_rc;
 971         return rc;
 972 }
 973
 974 /**
 975  * Discard pages with index greater than @size. If @ext is overlapped with
 976  * @size, then partial truncate happens.
 977  */
 978 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
 979                                bool partial)
 980 {
 981         struct lu_env *env;
 982         struct cl_io *io;
 983         struct osc_object *obj = ext->oe_obj;
 984         struct client_obd *cli = osc_cli(obj);
 985         struct osc_async_page *oap;
 986         struct osc_async_page *tmp;
 987         int pages_in_chunk = 0;
 988         int ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
 989         __u64 trunc_chunk = trunc_index >> ppc_bits;
 990         int grants = 0;
 991         int nr_pages = 0;
 992         int rc = 0;
 993         u16 refcheck;
 994
 995         LASSERT(sanity_check(ext) == 0);
 996         EASSERT(ext->oe_state == OES_TRUNC, ext);
 997         EASSERT(!ext->oe_urgent, ext);
 998
 999         /* Request new lu_env.
1000          * We can't use that env from osc_cache_truncate_start() because
1001          * it's from lov_io_sub and not fully initialized.
1002          */
1003         env = cl_env_get(&refcheck);
1004         io  = &osc_env_info(env)->oti_io;
1005         io->ci_obj = cl_object_top(osc2cl(obj));
1006         io->ci_ignore_layout = 1;
1007         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
1008         if (rc < 0)
1009                 goto out;
1010
1011         /* discard all pages with index greater then trunc_index */
1012         list_for_each_entry_safe(oap, tmp, &ext->oe_pages, oap_pending_item) {
1013                 pgoff_t index = osc_index(oap2osc(oap));
1014                 struct cl_page *page = oap2cl_page(oap);
1015
1016                 LASSERT(list_empty(&oap->oap_rpc_item));
1017
1018                 /* only discard the pages with their index greater than
1019                  * trunc_index, and ...
1020                  */
1021                 if (index < trunc_index ||
1022                     (index == trunc_index && partial)) {
1023                         /* accounting how many pages remaining in the chunk
1024                          * so that we can calculate grants correctly. */
1025                         if (index >> ppc_bits == trunc_chunk)
1026                                 ++pages_in_chunk;
1027                         continue;
1028                 }
1029
1030                 list_del_init(&oap->oap_pending_item);
1031
1032                 cl_page_get(page);
1033                 lu_ref_add(&page->cp_reference, "truncate", current);
1034
1035                 if (cl_page_own(env, io, page) == 0) {
1036                         cl_page_discard(env, io, page);
1037                         cl_page_disown(env, io, page);
1038                 } else {
1039                         LASSERT(page->cp_state == CPS_FREEING);
1040                         LASSERT(0);
1041                 }
1042
1043                 lu_ref_del(&page->cp_reference, "truncate", current);
1044                 cl_page_put(env, page);
1045
1046                 --ext->oe_nr_pages;
1047                 ++nr_pages;
1048         }
1049         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1050                       ext->oe_nr_pages == 0),
1051                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1052
1053         osc_object_lock(obj);
1054         if (ext->oe_nr_pages == 0) {
1055                 LASSERT(pages_in_chunk == 0);
1056                 grants = ext->oe_grants;
1057                 ext->oe_grants = 0;
1058         } else { /* calculate how many grants we can free */
1059                 int chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1060                 pgoff_t last_index;
1061
1062                 /* if there is no pages in this chunk, we can also free grants
1063                  * for the last chunk
1064                  */
1065                 if (pages_in_chunk == 0) {
1066                         /* if this is the 1st chunk and no pages in this chunk,
1067                          * ext->oe_nr_pages must be zero, so we should be in
1068                          * the other if-clause.
1069                          */
1070                         LASSERT(trunc_chunk > 0);
1071                         --trunc_chunk;
1072                         ++chunks;
1073                 }
1074
1075                 /* this is what we can free from this extent */
1076                 grants = chunks << cli->cl_chunkbits;
1077                 ext->oe_grants -= grants;
1078                 last_index = ((trunc_chunk + 1) << ppc_bits) - 1;
1079                 ext->oe_end = min(last_index, ext->oe_max_end);
1080                 LASSERT(ext->oe_end >= ext->oe_start);
1081                 LASSERT(ext->oe_grants > 0);
1082         }
1083         osc_object_unlock(obj);
1084
1085         if (grants > 0 || nr_pages > 0)
1086                 osc_free_grant(cli, nr_pages, grants);
1087
1088 out:
1089         cl_io_fini(env, io);
1090         cl_env_put(env, &refcheck);
1091         return rc;
1092 }
1093
1094 /**
1095  * This function is used to make the extent prepared for transfer.
1096  * A race with flushing page - ll_writepage() has to be handled cautiously.
1097  */
1098 static int osc_extent_make_ready(const struct lu_env *env,
1099                                  struct osc_extent *ext)
1100 {
1101         struct osc_async_page *oap;
1102         struct osc_async_page *last = NULL;
1103         struct osc_object *obj = ext->oe_obj;
1104         unsigned int page_count = 0;
1105         int rc;
1106
1107         /* we're going to grab page lock, so object lock must not be taken. */
1108         LASSERT(sanity_check(ext) == 0);
1109         /* in locking state, any process should not touch this extent. */
1110         EASSERT(ext->oe_state == OES_LOCKING, ext);
1111         EASSERT(ext->oe_owner, ext);
1112
1113         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1114
1115         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1116                 ++page_count;
1117                 if (!last || last->oap_obj_off < oap->oap_obj_off)
1118                         last = oap;
1119
1120                 /* checking ASYNC_READY is race safe */
1121                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1122                         continue;
1123
1124                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1125                 switch (rc) {
1126                 case 0:
1127                         spin_lock(&oap->oap_lock);
1128                         oap->oap_async_flags |= ASYNC_READY;
1129                         spin_unlock(&oap->oap_lock);
1130                         break;
1131                 case -EALREADY:
1132                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1133                         break;
1134                 default:
1135                         LASSERTF(0, "unknown return code: %d\n", rc);
1136                 }
1137         }
1138
1139         LASSERT(page_count == ext->oe_nr_pages);
1140         LASSERT(last);
1141         /* the last page is the only one we need to refresh its count by
1142          * the size of file.
1143          */
1144         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1145                 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1146
1147                 LASSERT(last_oap_count > 0);
1148                 LASSERT(last->oap_page_off + last_oap_count <= PAGE_SIZE);
1149                 last->oap_count = last_oap_count;
1150                 spin_lock(&last->oap_lock);
1151                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1152                 spin_unlock(&last->oap_lock);
1153         }
1154
1155         /* for the rest of pages, we don't need to call osf_refresh_count()
1156          * because it's known they are not the last page
1157          */
1158         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1159                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1160                         oap->oap_count = PAGE_SIZE - oap->oap_page_off;
1161                         spin_lock(&last->oap_lock);
1162                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1163                         spin_unlock(&last->oap_lock);
1164                 }
1165         }
1166
1167         osc_object_lock(obj);
1168         osc_extent_state_set(ext, OES_RPC);
1169         osc_object_unlock(obj);
1170         /* get a refcount for RPC. */
1171         osc_extent_get(ext);
1172
1173         return 0;
1174 }
1175
1176 /**
1177  * Quick and simple version of osc_extent_find(). This function is frequently
1178  * called to expand the extent for the same IO. To expand the extent, the
1179  * page index must be in the same or next chunk of ext->oe_end.
1180  */
1181 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index,
1182                              unsigned int *grants)
1183 {
1184         struct osc_object *obj = ext->oe_obj;
1185         struct client_obd *cli = osc_cli(obj);
1186         struct osc_extent *next;
1187         int ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1188         pgoff_t chunk = index >> ppc_bits;
1189         pgoff_t end_chunk;
1190         pgoff_t end_index;
1191         unsigned int chunksize = 1 << cli->cl_chunkbits;
1192         int rc = 0;
1193
1194         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1195         osc_object_lock(obj);
1196         LASSERT(sanity_check_nolock(ext) == 0);
1197         end_chunk = ext->oe_end >> ppc_bits;
1198         if (chunk > end_chunk + 1) {
1199                 rc = -ERANGE;
1200                 goto out;
1201         }
1202
1203         if (end_chunk >= chunk) {
1204                 rc = 0;
1205                 goto out;
1206         }
1207
1208         LASSERT(end_chunk + 1 == chunk);
1209         /* try to expand this extent to cover @index */
1210         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1211
1212         next = next_extent(ext);
1213         if (next && next->oe_start <= end_index) {
1214                 /* complex mode - overlapped with the next extent,
1215                  * this case will be handled by osc_extent_find()
1216                  */
1217                 rc = -EAGAIN;
1218                 goto out;
1219         }
1220
1221         ext->oe_end = end_index;
1222         ext->oe_grants += chunksize;
1223         LASSERT(*grants >= chunksize);
1224         *grants -= chunksize;
1225         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1226                  "overlapped after expanding for %lu.\n", index);
1227
1228 out:
1229         osc_object_unlock(obj);
1230         return rc;
1231 }
1232
1233 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
1234                                   const char *func, int line)
1235 {
1236         struct osc_extent *ext;
1237         int cnt;
1238
1239         CDEBUG(level, "Dump object %p extents at %s:%d, mppr: %u.\n",
1240                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1241
1242         /* osc_object_lock(obj); */
1243         cnt = 1;
1244         for (ext = first_extent(obj); ext; ext = next_extent(ext))
1245                 OSC_EXTENT_DUMP(level, ext, "in tree %d.\n", cnt++);
1246
1247         cnt = 1;
1248         list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1249                 OSC_EXTENT_DUMP(level, ext, "hp %d.\n", cnt++);
1250
1251         cnt = 1;
1252         list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1253                 OSC_EXTENT_DUMP(level, ext, "urgent %d.\n", cnt++);
1254
1255         cnt = 1;
1256         list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1257                 OSC_EXTENT_DUMP(level, ext, "reading %d.\n", cnt++);
1258         /* osc_object_unlock(obj); */
1259 }
1260
1261 /* ------------------ osc extent end ------------------ */
1262
1263 static inline int osc_is_ready(struct osc_object *osc)
1264 {
1265         return !list_empty(&osc->oo_ready_item) ||
1266                !list_empty(&osc->oo_hp_ready_item);
1267 }
1268
1269 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1270         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1271                (OSC), osc_is_ready(OSC),                                       \
1272                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1273                list_empty_marker(&(OSC)->oo_ready_item),                       \
1274                atomic_read(&(OSC)->oo_nr_writes),                              \
1275                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1276                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1277                atomic_read(&(OSC)->oo_nr_reads),                               \
1278                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1279                ##args)
1280
1281 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1282                           int cmd)
1283 {
1284         struct osc_page *opg = oap2osc_page(oap);
1285         struct cl_page  *page = oap2cl_page(oap);
1286         int result;
1287
1288         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1289
1290         result = cl_page_make_ready(env, page, CRT_WRITE);
1291         if (result == 0)
1292                 opg->ops_submit_time = cfs_time_current();
1293         return result;
1294 }
1295
1296 static int osc_refresh_count(const struct lu_env *env,
1297                              struct osc_async_page *oap, int cmd)
1298 {
1299         struct osc_page *opg = oap2osc_page(oap);
1300         pgoff_t index = osc_index(oap2osc(oap));
1301         struct cl_object *obj;
1302         struct cl_attr *attr = &osc_env_info(env)->oti_attr;
1303
1304         int result;
1305         loff_t kms;
1306
1307         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1308         LASSERT(!(cmd & OBD_BRW_READ));
1309         obj = opg->ops_cl.cpl_obj;
1310
1311         cl_object_attr_lock(obj);
1312         result = cl_object_attr_get(env, obj, attr);
1313         cl_object_attr_unlock(obj);
1314         if (result < 0)
1315                 return result;
1316         kms = attr->cat_kms;
1317         if (cl_offset(obj, index) >= kms)
1318                 /* catch race with truncate */
1319                 return 0;
1320         else if (cl_offset(obj, index + 1) > kms)
1321                 /* catch sub-page write at end of file */
1322                 return kms % PAGE_SIZE;
1323         else
1324                 return PAGE_SIZE;
1325 }
1326
1327 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1328                           int cmd, int rc)
1329 {
1330         struct osc_page *opg = oap2osc_page(oap);
1331         struct cl_page    *page = oap2cl_page(oap);
1332         enum cl_req_type crt;
1333         int srvlock;
1334
1335         cmd &= ~OBD_BRW_NOQUOTA;
1336         LASSERTF(equi(page->cp_state == CPS_PAGEIN, cmd == OBD_BRW_READ),
1337                  "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1338         LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1339                  "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1340         LASSERT(opg->ops_transfer_pinned);
1341
1342         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1343         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1344         opg->ops_transfer_pinned = 0;
1345
1346         opg->ops_submit_time = 0;
1347         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1348
1349         /* statistic */
1350         if (rc == 0 && srvlock) {
1351                 struct lu_device *ld = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1352                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1353                 size_t bytes = oap->oap_count;
1354
1355                 if (crt == CRT_READ)
1356                         stats->os_lockless_reads += bytes;
1357                 else
1358                         stats->os_lockless_writes += bytes;
1359         }
1360
1361         /*
1362          * This has to be the last operation with the page, as locks are
1363          * released in cl_page_completion() and nothing except for the
1364          * reference counter protects page from concurrent reclaim.
1365          */
1366         lu_ref_del(&page->cp_reference, "transfer", page);
1367
1368         cl_page_completion(env, page, crt, rc);
1369         cl_page_put(env, page);
1370
1371         return 0;
1372 }
1373
1374 #define OSC_DUMP_GRANT(lvl, cli, fmt, args...) do {                           \
1375         struct client_obd *__tmp = (cli);                                     \
1376         CDEBUG(lvl, "%s: grant { dirty: %lu/%lu dirty_pages: %ld/%lu "        \
1377                "dropped: %ld avail: %ld, reserved: %ld, flight: %d }"         \
1378                "lru {in list: %ld, left: %ld, waiters: %d }" fmt "\n",        \
1379                cli_name(__tmp),                                               \
1380                __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages,              \
1381                atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages,       \
1382                __tmp->cl_lost_grant, __tmp->cl_avail_grant,                   \
1383                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,               \
1384                atomic_long_read(&__tmp->cl_lru_in_list),                      \
1385                atomic_long_read(&__tmp->cl_lru_busy),                         \
1386                atomic_read(&__tmp->cl_lru_shrinkers), ##args);                \
1387 } while (0)
1388
1389 /* caller must hold loi_list_lock */
1390 static void osc_consume_write_grant(struct client_obd *cli,
1391                                     struct brw_page *pga)
1392 {
1393         assert_spin_locked(&cli->cl_loi_list_lock);
1394         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1395         atomic_long_inc(&obd_dirty_pages);
1396         cli->cl_dirty_pages++;
1397         pga->flag |= OBD_BRW_FROM_GRANT;
1398         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1399                PAGE_SIZE, pga, pga->pg);
1400         osc_update_next_shrink(cli);
1401 }
1402
1403 /* the companion to osc_consume_write_grant, called when a brw has completed.
1404  * must be called with the loi lock held.
1405  */
1406 static void osc_release_write_grant(struct client_obd *cli,
1407                                     struct brw_page *pga)
1408 {
1409         assert_spin_locked(&cli->cl_loi_list_lock);
1410         if (!(pga->flag & OBD_BRW_FROM_GRANT))
1411                 return;
1412
1413         pga->flag &= ~OBD_BRW_FROM_GRANT;
1414         atomic_long_dec(&obd_dirty_pages);
1415         cli->cl_dirty_pages--;
1416         if (pga->flag & OBD_BRW_NOCACHE) {
1417                 pga->flag &= ~OBD_BRW_NOCACHE;
1418                 atomic_long_dec(&obd_dirty_transit_pages);
1419                 cli->cl_dirty_transit--;
1420         }
1421 }
1422
1423 /**
1424  * To avoid sleeping with object lock held, it's good for us allocate enough
1425  * grants before entering into critical section.
1426  *
1427  * spin_lock held by caller
1428  */
1429 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1430 {
1431         int rc = -EDQUOT;
1432
1433         if (cli->cl_avail_grant >= bytes) {
1434                 cli->cl_avail_grant -= bytes;
1435                 cli->cl_reserved_grant += bytes;
1436                 rc = 0;
1437         }
1438         return rc;
1439 }
1440
1441 static void __osc_unreserve_grant(struct client_obd *cli,
1442                                   unsigned int reserved, unsigned int unused)
1443 {
1444         /* it's quite normal for us to get more grant than reserved.
1445          * Thinking about a case that two extents merged by adding a new
1446          * chunk, we can save one extent tax. If extent tax is greater than
1447          * one chunk, we can save more grant by adding a new chunk
1448          */
1449         cli->cl_reserved_grant -= reserved;
1450         if (unused > reserved) {
1451                 cli->cl_avail_grant += reserved;
1452                 cli->cl_lost_grant  += unused - reserved;
1453         } else {
1454                 cli->cl_avail_grant += unused;
1455         }
1456 }
1457
1458 static void osc_unreserve_grant(struct client_obd *cli,
1459                                 unsigned int reserved, unsigned int unused)
1460 {
1461         spin_lock(&cli->cl_loi_list_lock);
1462         __osc_unreserve_grant(cli, reserved, unused);
1463         if (unused > 0)
1464                 osc_wake_cache_waiters(cli);
1465         spin_unlock(&cli->cl_loi_list_lock);
1466 }
1467
1468 /**
1469  * Free grant after IO is finished or canceled.
1470  *
1471  * @lost_grant is used to remember how many grants we have allocated but not
1472  * used, we should return these grants to OST. There're two cases where grants
1473  * can be lost:
1474  * 1. truncate;
1475  * 2. blocksize at OST is less than PAGE_SIZE and a partial page was
1476  *    written. In this case OST may use less chunks to serve this partial
1477  *    write. OSTs don't actually know the page size on the client side. so
1478  *    clients have to calculate lost grant by the blocksize on the OST.
1479  *    See filter_grant_check() for details.
1480  */
1481 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1482                            unsigned int lost_grant)
1483 {
1484         unsigned long grant = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
1485
1486         spin_lock(&cli->cl_loi_list_lock);
1487         atomic_long_sub(nr_pages, &obd_dirty_pages);
1488         cli->cl_dirty_pages -= nr_pages;
1489         cli->cl_lost_grant += lost_grant;
1490         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1491                 /* borrow some grant from truncate to avoid the case that
1492                  * truncate uses up all avail grant
1493                  */
1494                 cli->cl_lost_grant -= grant;
1495                 cli->cl_avail_grant += grant;
1496         }
1497         osc_wake_cache_waiters(cli);
1498         spin_unlock(&cli->cl_loi_list_lock);
1499         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu\n",
1500                lost_grant, cli->cl_lost_grant,
1501                cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_SHIFT);
1502 }
1503
1504 /**
1505  * The companion to osc_enter_cache(), called when @oap is no longer part of
1506  * the dirty accounting due to error.
1507  */
1508 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1509 {
1510         spin_lock(&cli->cl_loi_list_lock);
1511         osc_release_write_grant(cli, &oap->oap_brw_page);
1512         spin_unlock(&cli->cl_loi_list_lock);
1513 }
1514
1515 /**
1516  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1517  * is available.
1518  */
1519 static int osc_enter_cache_try(struct client_obd *cli,
1520                                struct osc_async_page *oap,
1521                                int bytes, int transient)
1522 {
1523         int rc;
1524
1525         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1526
1527         rc = osc_reserve_grant(cli, bytes);
1528         if (rc < 0)
1529                 return 0;
1530
1531         if (cli->cl_dirty_pages < cli->cl_dirty_max_pages &&
1532             atomic_long_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages) {
1533                 osc_consume_write_grant(cli, &oap->oap_brw_page);
1534                 if (transient) {
1535                         cli->cl_dirty_transit++;
1536                         atomic_long_inc(&obd_dirty_transit_pages);
1537                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
1538                 }
1539                 rc = 1;
1540         } else {
1541                 __osc_unreserve_grant(cli, bytes, bytes);
1542                 rc = 0;
1543         }
1544         return rc;
1545 }
1546
1547 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1548 {
1549         int rc;
1550
1551         spin_lock(&cli->cl_loi_list_lock);
1552         rc = list_empty(&ocw->ocw_entry);
1553         spin_unlock(&cli->cl_loi_list_lock);
1554         return rc;
1555 }
1556
1557 /**
1558  * The main entry to reserve dirty page accounting. Usually the grant reserved
1559  * in this function will be freed in bulk in osc_free_grant() unless it fails
1560  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1561  *
1562  * The process will be put into sleep if it's already run out of grant.
1563  */
1564 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1565                            struct osc_async_page *oap, int bytes)
1566 {
1567         struct osc_object *osc = oap->oap_obj;
1568         struct lov_oinfo *loi = osc->oo_oinfo;
1569         struct osc_cache_waiter ocw;
1570         struct l_wait_info lwi;
1571         int rc = -EDQUOT;
1572
1573         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(AT_OFF ? obd_timeout : at_max),
1574                                NULL, LWI_ON_SIGNAL_NOOP, NULL);
1575
1576         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1577
1578         spin_lock(&cli->cl_loi_list_lock);
1579
1580         /* force the caller to try sync io.  this can jump the list
1581          * of queued writes and create a discontiguous rpc stream
1582          */
1583         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1584             !cli->cl_dirty_max_pages || cli->cl_ar.ar_force_sync ||
1585             loi->loi_ar.ar_force_sync) {
1586                 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1587                 rc = -EDQUOT;
1588                 goto out;
1589         }
1590
1591         /* Hopefully normal case - cache space and write credits available */
1592         if (osc_enter_cache_try(cli, oap, bytes, 0)) {
1593                 OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1594                 rc = 0;
1595                 goto out;
1596         }
1597
1598         /* We can get here for two reasons: too many dirty pages in cache, or
1599          * run out of grants. In both cases we should write dirty pages out.
1600          * Adding a cache waiter will trigger urgent write-out no matter what
1601          * RPC size will be.
1602          * The exiting condition is no avail grants and no dirty pages caching,
1603          * that really means there is no space on the OST.
1604          */
1605         init_waitqueue_head(&ocw.ocw_waitq);
1606         ocw.ocw_oap   = oap;
1607         ocw.ocw_grant = bytes;
1608         while (cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0) {
1609                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1610                 ocw.ocw_rc = 0;
1611                 spin_unlock(&cli->cl_loi_list_lock);
1612
1613                 osc_io_unplug_async(env, cli, NULL);
1614
1615                 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1616                        cli_name(cli), &ocw, oap);
1617
1618                 rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1619
1620                 spin_lock(&cli->cl_loi_list_lock);
1621
1622                 if (rc < 0) {
1623                         /* l_wait_event is interrupted by signal, or timed out */
1624                         list_del_init(&ocw.ocw_entry);
1625                         break;
1626                 }
1627                 LASSERT(list_empty(&ocw.ocw_entry));
1628                 rc = ocw.ocw_rc;
1629
1630                 if (rc != -EDQUOT)
1631                         break;
1632                 if (osc_enter_cache_try(cli, oap, bytes, 0)) {
1633                         rc = 0;
1634                         break;
1635                 }
1636         }
1637
1638         switch (rc) {
1639         case 0:
1640                 OSC_DUMP_GRANT(D_CACHE, cli, "finally got grant space\n");
1641                 break;
1642         case -ETIMEDOUT:
1643                 OSC_DUMP_GRANT(D_CACHE, cli,
1644                                "timeout, fall back to sync i/o\n");
1645                 osc_extent_tree_dump(D_CACHE, osc);
1646                 /* fall back to synchronous I/O */
1647                 rc = -EDQUOT;
1648                 break;
1649         case -EINTR:
1650                 /* Ensures restartability - LU-3581 */
1651                 OSC_DUMP_GRANT(D_CACHE, cli, "interrupted\n");
1652                 rc = -ERESTARTSYS;
1653                 break;
1654         case -EDQUOT:
1655                 OSC_DUMP_GRANT(D_CACHE, cli,
1656                                "no grant space, fall back to sync i/o\n");
1657                 break;
1658         default:
1659                 CDEBUG(D_CACHE, "%s: event for cache space @ %p never arrived due to %d, fall back to sync i/o\n",
1660                        cli_name(cli), &ocw, rc);
1661                 break;
1662         }
1663 out:
1664         spin_unlock(&cli->cl_loi_list_lock);
1665         return rc;
1666 }
1667
1668 /* caller must hold loi_list_lock */
1669 void osc_wake_cache_waiters(struct client_obd *cli)
1670 {
1671         struct list_head *l, *tmp;
1672         struct osc_cache_waiter *ocw;
1673
1674         list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1675                 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
1676                 list_del_init(&ocw->ocw_entry);
1677
1678                 ocw->ocw_rc = -EDQUOT;
1679                 /* we can't dirty more */
1680                 if ((cli->cl_dirty_pages > cli->cl_dirty_max_pages) ||
1681                     (atomic_long_read(&obd_dirty_pages) + 1 >
1682                      obd_max_dirty_pages)) {
1683                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld osc max %ld, sys max %ld\n",
1684                                cli->cl_dirty_pages, cli->cl_dirty_max_pages,
1685                                obd_max_dirty_pages);
1686                         goto wakeup;
1687                 }
1688
1689                 if (osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
1690                         ocw->ocw_rc = 0;
1691 wakeup:
1692                 CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld, %d\n",
1693                        ocw, ocw->ocw_oap, cli->cl_avail_grant, ocw->ocw_rc);
1694
1695                 wake_up(&ocw->ocw_waitq);
1696         }
1697 }
1698
1699 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1700 {
1701         int hprpc = !!list_empty(&osc->oo_hp_exts);
1702
1703         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1704 }
1705
1706 /* This maintains the lists of pending pages to read/write for a given object
1707  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1708  * to quickly find objects that are ready to send an RPC.
1709  */
1710 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1711                          int cmd)
1712 {
1713         int invalid_import = 0;
1714
1715         /* if we have an invalid import we want to drain the queued pages
1716          * by forcing them through rpcs that immediately fail and complete
1717          * the pages.  recovery relies on this to empty the queued pages
1718          * before canceling the locks and evicting down the llite pages
1719          */
1720         if (!cli->cl_import || cli->cl_import->imp_invalid)
1721                 invalid_import = 1;
1722
1723         if (cmd & OBD_BRW_WRITE) {
1724                 if (atomic_read(&osc->oo_nr_writes) == 0)
1725                         return 0;
1726                 if (invalid_import) {
1727                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1728                         return 1;
1729                 }
1730                 if (!list_empty(&osc->oo_hp_exts)) {
1731                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1732                         return 1;
1733                 }
1734                 if (!list_empty(&osc->oo_urgent_exts)) {
1735                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1736                         return 1;
1737                 }
1738                 /* trigger a write rpc stream as long as there are dirtiers
1739                  * waiting for space.  as they're waiting, they're not going to
1740                  * create more pages to coalesce with what's waiting..
1741                  */
1742                 if (!list_empty(&cli->cl_cache_waiters)) {
1743                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1744                         return 1;
1745                 }
1746                 if (atomic_read(&osc->oo_nr_writes) >=
1747                     cli->cl_max_pages_per_rpc)
1748                         return 1;
1749         } else {
1750                 if (atomic_read(&osc->oo_nr_reads) == 0)
1751                         return 0;
1752                 if (invalid_import) {
1753                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1754                         return 1;
1755                 }
1756                 /* all read are urgent. */
1757                 if (!list_empty(&osc->oo_reading_exts))
1758                         return 1;
1759         }
1760
1761         return 0;
1762 }
1763
1764 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1765 {
1766         struct client_obd *cli = osc_cli(obj);
1767
1768         if (cmd & OBD_BRW_WRITE) {
1769                 atomic_add(delta, &obj->oo_nr_writes);
1770                 atomic_add(delta, &cli->cl_pending_w_pages);
1771                 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1772         } else {
1773                 atomic_add(delta, &obj->oo_nr_reads);
1774                 atomic_add(delta, &cli->cl_pending_r_pages);
1775                 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1776         }
1777         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1778 }
1779
1780 static int osc_makes_hprpc(struct osc_object *obj)
1781 {
1782         return !list_empty(&obj->oo_hp_exts);
1783 }
1784
1785 static void on_list(struct list_head *item, struct list_head *list, int should_be_on)
1786 {
1787         if (list_empty(item) && should_be_on)
1788                 list_add_tail(item, list);
1789         else if (!list_empty(item) && !should_be_on)
1790                 list_del_init(item);
1791 }
1792
1793 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1794  * can find pages to build into rpcs quickly
1795  */
1796 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1797 {
1798         if (osc_makes_hprpc(osc)) {
1799                 /* HP rpc */
1800                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1801                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1802         } else {
1803                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1804                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1805                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1806                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1807         }
1808
1809         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1810                 atomic_read(&osc->oo_nr_writes) > 0);
1811
1812         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1813                 atomic_read(&osc->oo_nr_reads) > 0);
1814
1815         return osc_is_ready(osc);
1816 }
1817
1818 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1819 {
1820         int is_ready;
1821
1822         spin_lock(&cli->cl_loi_list_lock);
1823         is_ready = __osc_list_maint(cli, osc);
1824         spin_unlock(&cli->cl_loi_list_lock);
1825
1826         return is_ready;
1827 }
1828
1829 /* this is trying to propagate async writeback errors back up to the
1830  * application.  As an async write fails we record the error code for later if
1831  * the app does an fsync.  As long as errors persist we force future rpcs to be
1832  * sync so that the app can get a sync error and break the cycle of queueing
1833  * pages for which writeback will fail.
1834  */
1835 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1836                            int rc)
1837 {
1838         if (rc) {
1839                 if (!ar->ar_rc)
1840                         ar->ar_rc = rc;
1841
1842                 ar->ar_force_sync = 1;
1843                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1844                 return;
1845         }
1846
1847         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1848                 ar->ar_force_sync = 0;
1849 }
1850
1851 /* this must be called holding the loi list lock to give coverage to exit_cache,
1852  * async_flag maintenance, and oap_request
1853  */
1854 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1855                               struct osc_async_page *oap, int sent, int rc)
1856 {
1857         struct osc_object *osc = oap->oap_obj;
1858         struct lov_oinfo *loi = osc->oo_oinfo;
1859         __u64 xid = 0;
1860
1861         if (oap->oap_request) {
1862                 xid = ptlrpc_req_xid(oap->oap_request);
1863                 ptlrpc_req_finished(oap->oap_request);
1864                 oap->oap_request = NULL;
1865         }
1866
1867         /* As the transfer for this page is being done, clear the flags */
1868         spin_lock(&oap->oap_lock);
1869         oap->oap_async_flags = 0;
1870         spin_unlock(&oap->oap_lock);
1871         oap->oap_interrupted = 0;
1872
1873         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1874                 spin_lock(&cli->cl_loi_list_lock);
1875                 osc_process_ar(&cli->cl_ar, xid, rc);
1876                 osc_process_ar(&loi->loi_ar, xid, rc);
1877                 spin_unlock(&cli->cl_loi_list_lock);
1878         }
1879
1880         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1881         if (rc)
1882                 CERROR("completion on oap %p obj %p returns %d.\n",
1883                        oap, osc, rc);
1884 }
1885
1886 struct extent_rpc_data {
1887         struct list_head       *erd_rpc_list;
1888         unsigned int            erd_page_count;
1889         unsigned int            erd_max_pages;
1890         unsigned int            erd_max_chunks;
1891         unsigned int            erd_max_extents;
1892 };
1893
1894 static inline unsigned int osc_extent_chunks(const struct osc_extent *ext)
1895 {
1896         struct client_obd *cli = osc_cli(ext->oe_obj);
1897         unsigned int ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1898
1899         return (ext->oe_end >> ppc_bits) - (ext->oe_start >> ppc_bits) + 1;
1900 }
1901
1902 /**
1903  * Try to add extent to one RPC. We need to think about the following things:
1904  * - # of pages must not be over max_pages_per_rpc
1905  * - extent must be compatible with previous ones
1906  */
1907 static int try_to_add_extent_for_io(struct client_obd *cli,
1908                                     struct osc_extent *ext,
1909                                     struct extent_rpc_data *data)
1910 {
1911         struct osc_extent *tmp;
1912         unsigned int chunk_count;
1913         struct osc_async_page *oap = list_first_entry(&ext->oe_pages,
1914                                                       struct osc_async_page,
1915                                                       oap_pending_item);
1916
1917         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1918                 ext);
1919
1920         if (!data->erd_max_extents)
1921                 return 0;
1922
1923         chunk_count = osc_extent_chunks(ext);
1924         EASSERTF(data->erd_page_count != 0 ||
1925                  chunk_count <= data->erd_max_chunks, ext,
1926                  "The first extent to be fit in a RPC contains %u chunks, which is over the limit %u.\n",
1927                  chunk_count, data->erd_max_chunks);
1928
1929         if (chunk_count > data->erd_max_chunks)
1930                 return 0;
1931
1932         data->erd_max_pages = max(ext->oe_mppr, data->erd_max_pages);
1933         EASSERTF(data->erd_page_count != 0 ||
1934                  ext->oe_nr_pages <= data->erd_max_pages, ext,
1935                  "The first extent to be fit in a RPC contains %u pages, which is over the limit %u.\n",
1936                  ext->oe_nr_pages, data->erd_max_pages);
1937         if (data->erd_page_count + ext->oe_nr_pages > data->erd_max_pages)
1938                 return 0;
1939
1940         list_for_each_entry(tmp, data->erd_rpc_list, oe_link) {
1941                 struct osc_async_page *oap2;
1942
1943                 oap2 = list_first_entry(&tmp->oe_pages, struct osc_async_page,
1944                                         oap_pending_item);
1945                 EASSERT(tmp->oe_owner == current, tmp);
1946                 if (oap2cl_page(oap)->cp_type != oap2cl_page(oap2)->cp_type) {
1947                         CDEBUG(D_CACHE, "Do not permit different type of IO in one RPC\n");
1948                         return 0;
1949                 }
1950
1951                 if (tmp->oe_srvlock != ext->oe_srvlock ||
1952                     !tmp->oe_grants != !ext->oe_grants ||
1953                     tmp->oe_no_merge || ext->oe_no_merge)
1954                         return 0;
1955
1956                 /* remove break for strict check */
1957                 break;
1958         }
1959
1960         data->erd_max_extents--;
1961         data->erd_max_chunks -= chunk_count;
1962         data->erd_page_count += ext->oe_nr_pages;
1963         list_move_tail(&ext->oe_link, data->erd_rpc_list);
1964         ext->oe_owner = current;
1965         return 1;
1966 }
1967
1968 static inline unsigned int osc_max_write_chunks(const struct client_obd *cli)
1969 {
1970         /*
1971          * LU-8135:
1972          *
1973          * The maximum size of a single transaction is about 64MB in ZFS.
1974          * #define DMU_MAX_ACCESS (64 * 1024 * 1024)
1975          *
1976          * Since ZFS is a copy-on-write file system, a single dirty page in
1977          * a chunk will result in the rewrite of the whole chunk, therefore
1978          * an RPC shouldn't be allowed to contain too many chunks otherwise
1979          * it will make transaction size much bigger than 64MB, especially
1980          * with big block size for ZFS.
1981          *
1982          * This piece of code is to make sure that OSC won't send write RPCs
1983          * with too many chunks. The maximum chunk size that an RPC can cover
1984          * is set to PTLRPC_MAX_BRW_SIZE, which is defined to 16MB. Ideally
1985          * OST should tell the client what the biggest transaction size is,
1986          * but it's good enough for now.
1987          *
1988          * This limitation doesn't apply to ldiskfs, which allows as many
1989          * chunks in one RPC as we want. However, it won't have any benefits
1990          * to have too many discontiguous pages in one RPC.
1991          *
1992          * An osc_extent won't cover over a RPC size, so the chunks in an
1993          * osc_extent won't bigger than PTLRPC_MAX_BRW_SIZE >> chunkbits.
1994          */
1995         return PTLRPC_MAX_BRW_SIZE >> cli->cl_chunkbits;
1996 }
1997
1998 /**
1999  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
2000  * get_write_extent() takes all appropriate extents in atomic.
2001  *
2002  * The following policy is used to collect extents for IO:
2003  * 1. Add as many HP extents as possible;
2004  * 2. Add the first urgent extent in urgent extent list and take it out of
2005  *    urgent list;
2006  * 3. Add subsequent extents of this urgent extent;
2007  * 4. If urgent list is not empty, goto 2;
2008  * 5. Traverse the extent tree from the 1st extent;
2009  * 6. Above steps exit if there is no space in this RPC.
2010  */
2011 static unsigned int get_write_extents(struct osc_object *obj,
2012                                       struct list_head *rpclist)
2013 {
2014         struct client_obd *cli = osc_cli(obj);
2015         struct osc_extent *ext;
2016         struct osc_extent *temp;
2017         struct extent_rpc_data data = {
2018                 .erd_rpc_list = rpclist,
2019                 .erd_page_count = 0,
2020                 .erd_max_pages = cli->cl_max_pages_per_rpc,
2021                 .erd_max_chunks = osc_max_write_chunks(cli),
2022                 .erd_max_extents = 256,
2023         };
2024
2025         LASSERT(osc_object_is_locked(obj));
2026         list_for_each_entry_safe(ext, temp, &obj->oo_hp_exts, oe_link) {
2027                 LASSERT(ext->oe_state == OES_CACHE);
2028                 if (!try_to_add_extent_for_io(cli, ext, &data))
2029                         return data.erd_page_count;
2030                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2031         }
2032         if (data.erd_page_count == data.erd_max_pages)
2033                 return data.erd_page_count;
2034
2035         while (!list_empty(&obj->oo_urgent_exts)) {
2036                 ext = list_entry(obj->oo_urgent_exts.next,
2037                                  struct osc_extent, oe_link);
2038                 if (!try_to_add_extent_for_io(cli, ext, &data))
2039                         return data.erd_page_count;
2040
2041                 if (!ext->oe_intree)
2042                         continue;
2043
2044                 while ((ext = next_extent(ext)) != NULL) {
2045                         if ((ext->oe_state != OES_CACHE) ||
2046                             (!list_empty(&ext->oe_link) &&
2047                              ext->oe_owner))
2048                                 continue;
2049
2050                         if (!try_to_add_extent_for_io(cli, ext, &data))
2051                                 return data.erd_page_count;
2052                 }
2053         }
2054         if (data.erd_page_count == data.erd_max_pages)
2055                 return data.erd_page_count;
2056
2057         ext = first_extent(obj);
2058         while (ext) {
2059                 if ((ext->oe_state != OES_CACHE) ||
2060                     /* this extent may be already in current rpclist */
2061                     (!list_empty(&ext->oe_link) && ext->oe_owner)) {
2062                         ext = next_extent(ext);
2063                         continue;
2064                 }
2065
2066                 if (!try_to_add_extent_for_io(cli, ext, &data))
2067                         return data.erd_page_count;
2068
2069                 ext = next_extent(ext);
2070         }
2071         return data.erd_page_count;
2072 }
2073
2074 static int
2075 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
2076                    struct osc_object *osc)
2077         __must_hold(osc)
2078 {
2079         LIST_HEAD(rpclist);
2080         struct osc_extent *ext;
2081         struct osc_extent *tmp;
2082         struct osc_extent *first = NULL;
2083         u32 page_count = 0;
2084         int srvlock = 0;
2085         int rc = 0;
2086
2087         LASSERT(osc_object_is_locked(osc));
2088
2089         page_count = get_write_extents(osc, &rpclist);
2090         LASSERT(equi(page_count == 0, list_empty(&rpclist)));
2091
2092         if (list_empty(&rpclist))
2093                 return 0;
2094
2095         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2096
2097         list_for_each_entry(ext, &rpclist, oe_link) {
2098                 LASSERT(ext->oe_state == OES_CACHE ||
2099                         ext->oe_state == OES_LOCK_DONE);
2100                 if (ext->oe_state == OES_CACHE)
2101                         osc_extent_state_set(ext, OES_LOCKING);
2102                 else
2103                         osc_extent_state_set(ext, OES_RPC);
2104         }
2105
2106         /* we're going to grab page lock, so release object lock because
2107          * lock order is page lock -> object lock.
2108          */
2109         osc_object_unlock(osc);
2110
2111         list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2112                 if (ext->oe_state == OES_LOCKING) {
2113                         rc = osc_extent_make_ready(env, ext);
2114                         if (unlikely(rc < 0)) {
2115                                 list_del_init(&ext->oe_link);
2116                                 osc_extent_finish(env, ext, 0, rc);
2117                                 continue;
2118                         }
2119                 }
2120                 if (!first) {
2121                         first = ext;
2122                         srvlock = ext->oe_srvlock;
2123                 } else {
2124                         LASSERT(srvlock == ext->oe_srvlock);
2125                 }
2126         }
2127
2128         if (!list_empty(&rpclist)) {
2129                 LASSERT(page_count > 0);
2130                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2131                 LASSERT(list_empty(&rpclist));
2132         }
2133
2134         osc_object_lock(osc);
2135         return rc;
2136 }
2137
2138 /**
2139  * prepare pages for ASYNC io and put pages in send queue.
2140  *
2141  * \param cmd OBD_BRW_* macroses
2142  * \param lop pending pages
2143  *
2144  * \return zero if no page added to send queue.
2145  * \return 1 if pages successfully added to send queue.
2146  * \return negative on errors.
2147  */
2148 static int
2149 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2150                   struct osc_object *osc)
2151         __must_hold(osc)
2152 {
2153         struct osc_extent *ext;
2154         struct osc_extent *next;
2155         LIST_HEAD(rpclist);
2156         struct extent_rpc_data data = {
2157                 .erd_rpc_list = &rpclist,
2158                 .erd_page_count = 0,
2159                 .erd_max_pages = cli->cl_max_pages_per_rpc,
2160                 .erd_max_chunks = UINT_MAX,
2161                 .erd_max_extents = UINT_MAX,
2162         };
2163         int rc = 0;
2164
2165         LASSERT(osc_object_is_locked(osc));
2166         list_for_each_entry_safe(ext, next, &osc->oo_reading_exts, oe_link) {
2167                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2168                 if (!try_to_add_extent_for_io(cli, ext, &data))
2169                         break;
2170                 osc_extent_state_set(ext, OES_RPC);
2171                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2172         }
2173         LASSERT(data.erd_page_count <= data.erd_max_pages);
2174
2175         osc_update_pending(osc, OBD_BRW_READ, -data.erd_page_count);
2176
2177         if (!list_empty(&rpclist)) {
2178                 osc_object_unlock(osc);
2179
2180                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2181                 LASSERT(list_empty(&rpclist));
2182
2183                 osc_object_lock(osc);
2184         }
2185         return rc;
2186 }
2187
2188 #define list_to_obj(list, item) ({                                            \
2189         struct list_head *__tmp = (list)->next;                               \
2190         list_del_init(__tmp);                                                 \
2191         list_entry(__tmp, struct osc_object, oo_##item);                      \
2192 })
2193
2194 /* This is called by osc_check_rpcs() to find which objects have pages that
2195  * we could be sending.  These lists are maintained by osc_makes_rpc().
2196  */
2197 static struct osc_object *osc_next_obj(struct client_obd *cli)
2198 {
2199         /* First return objects that have blocked locks so that they
2200          * will be flushed quickly and other clients can get the lock,
2201          * then objects which have pages ready to be stuffed into RPCs
2202          */
2203         if (!list_empty(&cli->cl_loi_hp_ready_list))
2204                 return list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item);
2205         if (!list_empty(&cli->cl_loi_ready_list))
2206                 return list_to_obj(&cli->cl_loi_ready_list, ready_item);
2207
2208         /* then if we have cache waiters, return all objects with queued
2209          * writes.  This is especially important when many small files
2210          * have filled up the cache and not been fired into rpcs because
2211          * they don't pass the nr_pending/object threshold
2212          */
2213         if (!list_empty(&cli->cl_cache_waiters) &&
2214             !list_empty(&cli->cl_loi_write_list))
2215                 return list_to_obj(&cli->cl_loi_write_list, write_item);
2216
2217         /* then return all queued objects when we have an invalid import
2218          * so that they get flushed
2219          */
2220         if (!cli->cl_import || cli->cl_import->imp_invalid) {
2221                 if (!list_empty(&cli->cl_loi_write_list))
2222                         return list_to_obj(&cli->cl_loi_write_list, write_item);
2223                 if (!list_empty(&cli->cl_loi_read_list))
2224                         return list_to_obj(&cli->cl_loi_read_list, read_item);
2225         }
2226         return NULL;
2227 }
2228
2229 /* called with the loi list lock held */
2230 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2231         __must_hold(&cli->cl_loi_list_lock)
2232 {
2233         struct osc_object *osc;
2234         int rc = 0;
2235
2236         while ((osc = osc_next_obj(cli)) != NULL) {
2237                 struct cl_object *obj = osc2cl(osc);
2238                 struct lu_ref_link link;
2239
2240                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2241
2242                 if (osc_max_rpc_in_flight(cli, osc)) {
2243                         __osc_list_maint(cli, osc);
2244                         break;
2245                 }
2246
2247                 cl_object_get(obj);
2248                 spin_unlock(&cli->cl_loi_list_lock);
2249                 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2250
2251                 /* attempt some read/write balancing by alternating between
2252                  * reads and writes in an object.  The makes_rpc checks here
2253                  * would be redundant if we were getting read/write work items
2254                  * instead of objects.  we don't want send_oap_rpc to drain a
2255                  * partial read pending queue when we're given this object to
2256                  * do io on writes while there are cache waiters
2257                  */
2258                 osc_object_lock(osc);
2259                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2260                         rc = osc_send_write_rpc(env, cli, osc);
2261                         if (rc < 0) {
2262                                 CERROR("Write request failed with %d\n", rc);
2263
2264                                 /* osc_send_write_rpc failed, mostly because of
2265                                  * memory pressure.
2266                                  *
2267                                  * It can't break here, because if:
2268                                  *  - a page was submitted by osc_io_submit, so
2269                                  *    page locked;
2270                                  *  - no request in flight
2271                                  *  - no subsequent request
2272                                  * The system will be in live-lock state,
2273                                  * because there is no chance to call
2274                                  * osc_io_unplug() and osc_check_rpcs() any
2275                                  * more. pdflush can't help in this case,
2276                                  * because it might be blocked at grabbing
2277                                  * the page lock as we mentioned.
2278                                  *
2279                                  * Anyway, continue to drain pages.
2280                                  */
2281                                 /* break; */
2282                         }
2283                 }
2284                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2285                         rc = osc_send_read_rpc(env, cli, osc);
2286                         if (rc < 0)
2287                                 CERROR("Read request failed with %d\n", rc);
2288                 }
2289                 osc_object_unlock(osc);
2290
2291                 osc_list_maint(cli, osc);
2292                 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2293                 cl_object_put(env, obj);
2294
2295                 spin_lock(&cli->cl_loi_list_lock);
2296         }
2297 }
2298
2299 static int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2300                           struct osc_object *osc, int async)
2301 {
2302         int rc = 0;
2303
2304         if (osc && osc_list_maint(cli, osc) == 0)
2305                 return 0;
2306
2307         if (!async) {
2308                 spin_lock(&cli->cl_loi_list_lock);
2309                 osc_check_rpcs(env, cli);
2310                 spin_unlock(&cli->cl_loi_list_lock);
2311         } else {
2312                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2313                 LASSERT(cli->cl_writeback_work);
2314                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2315         }
2316         return rc;
2317 }
2318
2319 static int osc_io_unplug_async(const struct lu_env *env,
2320                                struct client_obd *cli, struct osc_object *osc)
2321 {
2322         return osc_io_unplug0(env, cli, osc, 1);
2323 }
2324
2325 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
2326                    struct osc_object *osc)
2327 {
2328         (void)osc_io_unplug0(env, cli, osc, 0);
2329 }
2330
2331 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2332                         struct page *page, loff_t offset)
2333 {
2334         struct obd_export *exp = osc_export(osc);
2335         struct osc_async_page *oap = &ops->ops_oap;
2336
2337         if (!page)
2338                 return cfs_size_round(sizeof(*oap));
2339
2340         oap->oap_magic = OAP_MAGIC;
2341         oap->oap_cli = &exp->exp_obd->u.cli;
2342         oap->oap_obj = osc;
2343
2344         oap->oap_page = page;
2345         oap->oap_obj_off = offset;
2346         LASSERT(!(offset & ~PAGE_MASK));
2347
2348         if (capable(CFS_CAP_SYS_RESOURCE))
2349                 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2350
2351         INIT_LIST_HEAD(&oap->oap_pending_item);
2352         INIT_LIST_HEAD(&oap->oap_rpc_item);
2353
2354         spin_lock_init(&oap->oap_lock);
2355         CDEBUG(D_INFO, "oap %p page %p obj off %llu\n",
2356                oap, page, oap->oap_obj_off);
2357         return 0;
2358 }
2359
2360 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2361                        struct osc_page *ops)
2362 {
2363         struct osc_io *oio = osc_env_io(env);
2364         struct osc_extent *ext = NULL;
2365         struct osc_async_page *oap = &ops->ops_oap;
2366         struct client_obd *cli = oap->oap_cli;
2367         struct osc_object *osc = oap->oap_obj;
2368         pgoff_t index;
2369         unsigned int grants = 0, tmp;
2370         int brw_flags = OBD_BRW_ASYNC;
2371         int cmd = OBD_BRW_WRITE;
2372         int need_release = 0;
2373         int rc = 0;
2374
2375         if (oap->oap_magic != OAP_MAGIC)
2376                 return -EINVAL;
2377
2378         if (!cli->cl_import || cli->cl_import->imp_invalid)
2379                 return -EIO;
2380
2381         if (!list_empty(&oap->oap_pending_item) ||
2382             !list_empty(&oap->oap_rpc_item))
2383                 return -EBUSY;
2384
2385         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2386         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2387         if (capable(CFS_CAP_SYS_RESOURCE)) {
2388                 brw_flags |= OBD_BRW_NOQUOTA;
2389                 cmd |= OBD_BRW_NOQUOTA;
2390         }
2391
2392         /* check if the file's owner/group is over quota */
2393         if (!(cmd & OBD_BRW_NOQUOTA)) {
2394                 struct cl_object *obj;
2395                 struct cl_attr *attr;
2396                 unsigned int qid[MAXQUOTAS];
2397
2398                 obj = cl_object_top(&osc->oo_cl);
2399                 attr = &osc_env_info(env)->oti_attr;
2400
2401                 cl_object_attr_lock(obj);
2402                 rc = cl_object_attr_get(env, obj, attr);
2403                 cl_object_attr_unlock(obj);
2404
2405                 qid[USRQUOTA] = attr->cat_uid;
2406                 qid[GRPQUOTA] = attr->cat_gid;
2407                 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2408                         rc = -EDQUOT;
2409                 if (rc)
2410                         return rc;
2411         }
2412
2413         oap->oap_cmd = cmd;
2414         oap->oap_page_off = ops->ops_from;
2415         oap->oap_count = ops->ops_to - ops->ops_from;
2416         /*
2417          * No need to hold a lock here,
2418          * since this page is not in any list yet.
2419          */
2420         oap->oap_async_flags = 0;
2421         oap->oap_brw_flags = brw_flags;
2422
2423         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2424                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2425
2426         index = osc_index(oap2osc(oap));
2427
2428         /* Add this page into extent by the following steps:
2429          * 1. if there exists an active extent for this IO, mostly this page
2430          *    can be added to the active extent and sometimes we need to
2431          *    expand extent to accommodate this page;
2432          * 2. otherwise, a new extent will be allocated.
2433          */
2434
2435         ext = oio->oi_active;
2436         if (ext && ext->oe_start <= index && ext->oe_max_end >= index) {
2437                 /* one chunk plus extent overhead must be enough to write this
2438                  * page
2439                  */
2440                 grants = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2441                 if (ext->oe_end >= index)
2442                         grants = 0;
2443
2444                 /* it doesn't need any grant to dirty this page */
2445                 spin_lock(&cli->cl_loi_list_lock);
2446                 rc = osc_enter_cache_try(cli, oap, grants, 0);
2447                 spin_unlock(&cli->cl_loi_list_lock);
2448                 if (rc == 0) { /* try failed */
2449                         grants = 0;
2450                         need_release = 1;
2451                 } else if (ext->oe_end < index) {
2452                         tmp = grants;
2453                         /* try to expand this extent */
2454                         rc = osc_extent_expand(ext, index, &tmp);
2455                         if (rc < 0) {
2456                                 need_release = 1;
2457                                 /* don't free reserved grant */
2458                         } else {
2459                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2460                                                 "expanded for %lu.\n", index);
2461                                 osc_unreserve_grant(cli, grants, tmp);
2462                                 grants = 0;
2463                         }
2464                 }
2465                 rc = 0;
2466         } else if (ext) {
2467                 /* index is located outside of active extent */
2468                 need_release = 1;
2469         }
2470         if (need_release) {
2471                 osc_extent_release(env, ext);
2472                 oio->oi_active = NULL;
2473                 ext = NULL;
2474         }
2475
2476         if (!ext) {
2477                 tmp = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2478
2479                 /* try to find new extent to cover this page */
2480                 LASSERT(!oio->oi_active);
2481                 /* we may have allocated grant for this page if we failed
2482                  * to expand the previous active extent.
2483                  */
2484                 LASSERT(ergo(grants > 0, grants >= tmp));
2485
2486                 rc = 0;
2487                 if (grants == 0) {
2488                         /* we haven't allocated grant for this page. */
2489                         rc = osc_enter_cache(env, cli, oap, tmp);
2490                         if (rc == 0)
2491                                 grants = tmp;
2492                 }
2493
2494                 tmp = grants;
2495                 if (rc == 0) {
2496                         ext = osc_extent_find(env, osc, index, &tmp);
2497                         if (IS_ERR(ext)) {
2498                                 LASSERT(tmp == grants);
2499                                 osc_exit_cache(cli, oap);
2500                                 rc = PTR_ERR(ext);
2501                                 ext = NULL;
2502                         } else {
2503                                 oio->oi_active = ext;
2504                         }
2505                 }
2506                 if (grants > 0)
2507                         osc_unreserve_grant(cli, grants, tmp);
2508         }
2509
2510         LASSERT(ergo(rc == 0, ext));
2511         if (ext) {
2512                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2513                          ext, "index = %lu.\n", index);
2514                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2515
2516                 osc_object_lock(osc);
2517                 if (ext->oe_nr_pages == 0)
2518                         ext->oe_srvlock = ops->ops_srvlock;
2519                 else
2520                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2521                 ++ext->oe_nr_pages;
2522                 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2523                 osc_object_unlock(osc);
2524         }
2525         return rc;
2526 }
2527
2528 int osc_teardown_async_page(const struct lu_env *env,
2529                             struct osc_object *obj, struct osc_page *ops)
2530 {
2531         struct osc_async_page *oap = &ops->ops_oap;
2532         int rc = 0;
2533
2534         LASSERT(oap->oap_magic == OAP_MAGIC);
2535
2536         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2537                oap, ops, osc_index(oap2osc(oap)));
2538
2539         if (!list_empty(&oap->oap_rpc_item)) {
2540                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2541                 rc = -EBUSY;
2542         } else if (!list_empty(&oap->oap_pending_item)) {
2543                 struct osc_extent *ext = NULL;
2544
2545                 osc_object_lock(obj);
2546                 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2547                 osc_object_unlock(obj);
2548                 /* only truncated pages are allowed to be taken out.
2549                  * See osc_extent_truncate() and osc_cache_truncate_start()
2550                  * for details.
2551                  */
2552                 if (ext && ext->oe_state != OES_TRUNC) {
2553                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2554                                         osc_index(oap2osc(oap)));
2555                         rc = -EBUSY;
2556                 }
2557                 if (ext)
2558                         osc_extent_put(env, ext);
2559         }
2560         return rc;
2561 }
2562
2563 /**
2564  * This is called when a page is picked up by kernel to write out.
2565  *
2566  * We should find out the corresponding extent and add the whole extent
2567  * into urgent list. The extent may be being truncated or used, handle it
2568  * carefully.
2569  */
2570 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2571                          struct osc_page *ops)
2572 {
2573         struct osc_extent *ext = NULL;
2574         struct osc_object *obj = cl2osc(ops->ops_cl.cpl_obj);
2575         struct cl_page *cp = ops->ops_cl.cpl_page;
2576         pgoff_t            index = osc_index(ops);
2577         struct osc_async_page *oap = &ops->ops_oap;
2578         bool unplug = false;
2579         int rc = 0;
2580
2581         osc_object_lock(obj);
2582         ext = osc_extent_lookup(obj, index);
2583         if (!ext) {
2584                 osc_extent_tree_dump(D_ERROR, obj);
2585                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2586         }
2587
2588         switch (ext->oe_state) {
2589         case OES_RPC:
2590         case OES_LOCK_DONE:
2591                 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2592                 LASSERT(0);
2593                 break;
2594         case OES_LOCKING:
2595                 /* If we know this extent is being written out, we should abort
2596                  * so that the writer can make this page ready. Otherwise, there
2597                  * exists a deadlock problem because other process can wait for
2598                  * page writeback bit holding page lock; and meanwhile in
2599                  * vvp_page_make_ready(), we need to grab page lock before
2600                  * really sending the RPC.
2601                  */
2602         case OES_TRUNC:
2603                 /* race with truncate, page will be redirtied */
2604         case OES_ACTIVE:
2605                 /* The extent is active so we need to abort and let the caller
2606                  * re-dirty the page. If we continued on here, and we were the
2607                  * one making the extent active, we could deadlock waiting for
2608                  * the page writeback to clear but it won't because the extent
2609                  * is active and won't be written out.
2610                  */
2611                 rc = -EAGAIN;
2612                 goto out;
2613         default:
2614                 break;
2615         }
2616
2617         rc = cl_page_prep(env, io, cp, CRT_WRITE);
2618         if (rc)
2619                 goto out;
2620
2621         spin_lock(&oap->oap_lock);
2622         oap->oap_async_flags |= ASYNC_READY | ASYNC_URGENT;
2623         spin_unlock(&oap->oap_lock);
2624
2625         if (memory_pressure_get())
2626                 ext->oe_memalloc = 1;
2627
2628         ext->oe_urgent = 1;
2629         if (ext->oe_state == OES_CACHE) {
2630                 OSC_EXTENT_DUMP(D_CACHE, ext,
2631                                 "flush page %p make it urgent.\n", oap);
2632                 if (list_empty(&ext->oe_link))
2633                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2634                 unplug = true;
2635         }
2636         rc = 0;
2637
2638 out:
2639         osc_object_unlock(obj);
2640         osc_extent_put(env, ext);
2641         if (unplug)
2642                 osc_io_unplug_async(env, osc_cli(obj), obj);
2643         return rc;
2644 }
2645
2646 /**
2647  * this is called when a sync waiter receives an interruption.  Its job is to
2648  * get the caller woken as soon as possible.  If its page hasn't been put in an
2649  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2650  * desiring interruption which will forcefully complete the rpc once the rpc
2651  * has timed out.
2652  */
2653 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2654 {
2655         struct osc_async_page *oap = &ops->ops_oap;
2656         struct osc_object *obj = oap->oap_obj;
2657         struct client_obd *cli = osc_cli(obj);
2658         struct osc_extent *ext;
2659         struct osc_extent *found = NULL;
2660         struct list_head *plist;
2661         pgoff_t index = osc_index(ops);
2662         int rc = -EBUSY;
2663         int cmd;
2664
2665         LASSERT(!oap->oap_interrupted);
2666         oap->oap_interrupted = 1;
2667
2668         /* Find out the caching extent */
2669         osc_object_lock(obj);
2670         if (oap->oap_cmd & OBD_BRW_WRITE) {
2671                 plist = &obj->oo_urgent_exts;
2672                 cmd = OBD_BRW_WRITE;
2673         } else {
2674                 plist = &obj->oo_reading_exts;
2675                 cmd = OBD_BRW_READ;
2676         }
2677         list_for_each_entry(ext, plist, oe_link) {
2678                 if (ext->oe_start <= index && ext->oe_end >= index) {
2679                         LASSERT(ext->oe_state == OES_LOCK_DONE);
2680                         /* For OES_LOCK_DONE state extent, it has already held
2681                          * a refcount for RPC.
2682                          */
2683                         found = osc_extent_get(ext);
2684                         break;
2685                 }
2686         }
2687         if (found) {
2688                 list_del_init(&found->oe_link);
2689                 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2690                 osc_object_unlock(obj);
2691
2692                 osc_extent_finish(env, found, 0, -EINTR);
2693                 osc_extent_put(env, found);
2694                 rc = 0;
2695         } else {
2696                 osc_object_unlock(obj);
2697                 /* ok, it's been put in an rpc. only one oap gets a request
2698                  * reference
2699                  */
2700                 if (oap->oap_request) {
2701                         ptlrpc_mark_interrupted(oap->oap_request);
2702                         ptlrpcd_wake(oap->oap_request);
2703                         ptlrpc_req_finished(oap->oap_request);
2704                         oap->oap_request = NULL;
2705                 }
2706         }
2707
2708         osc_list_maint(cli, obj);
2709         return rc;
2710 }
2711
2712 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
2713                          struct list_head *list, int cmd, int brw_flags)
2714 {
2715         struct client_obd *cli = osc_cli(obj);
2716         struct osc_extent *ext;
2717         struct osc_async_page *oap, *tmp;
2718         int page_count = 0;
2719         int mppr = cli->cl_max_pages_per_rpc;
2720         bool can_merge = true;
2721         pgoff_t start = CL_PAGE_EOF;
2722         pgoff_t end = 0;
2723
2724         list_for_each_entry(oap, list, oap_pending_item) {
2725                 struct osc_page *opg = oap2osc_page(oap);
2726                 pgoff_t index = osc_index(opg);
2727
2728                 if (index > end)
2729                         end = index;
2730                 if (index < start)
2731                         start = index;
2732                 ++page_count;
2733                 mppr <<= (page_count > mppr);
2734
2735                 if (unlikely(opg->ops_from > 0 || opg->ops_to < PAGE_SIZE))
2736                         can_merge = false;
2737         }
2738
2739         ext = osc_extent_alloc(obj);
2740         if (!ext) {
2741                 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2742                         list_del_init(&oap->oap_pending_item);
2743                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2744                 }
2745                 return -ENOMEM;
2746         }
2747
2748         ext->oe_rw = !!(cmd & OBD_BRW_READ);
2749         ext->oe_sync = 1;
2750         ext->oe_no_merge = !can_merge;
2751         ext->oe_urgent = 1;
2752         ext->oe_start = start;
2753         ext->oe_end = end;
2754         ext->oe_max_end = end;
2755         ext->oe_obj = obj;
2756         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2757         ext->oe_nr_pages = page_count;
2758         ext->oe_mppr = mppr;
2759         list_splice_init(list, &ext->oe_pages);
2760
2761         osc_object_lock(obj);
2762         /* Reuse the initial refcount for RPC, don't drop it */
2763         osc_extent_state_set(ext, OES_LOCK_DONE);
2764         if (cmd & OBD_BRW_WRITE) {
2765                 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2766                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2767         } else {
2768                 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2769                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2770         }
2771         osc_object_unlock(obj);
2772
2773         osc_io_unplug_async(env, cli, obj);
2774         return 0;
2775 }
2776
2777 /**
2778  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2779  */
2780 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
2781                              u64 size, struct osc_extent **extp)
2782 {
2783         struct client_obd *cli = osc_cli(obj);
2784         struct osc_extent *ext;
2785         struct osc_extent *temp;
2786         struct osc_extent *waiting = NULL;
2787         pgoff_t index;
2788         LIST_HEAD(list);
2789         int result = 0;
2790         bool partial;
2791
2792         /* pages with index greater or equal to index will be truncated. */
2793         index = cl_index(osc2cl(obj), size);
2794         partial = size > cl_offset(osc2cl(obj), index);
2795
2796 again:
2797         osc_object_lock(obj);
2798         ext = osc_extent_search(obj, index);
2799         if (!ext)
2800                 ext = first_extent(obj);
2801         else if (ext->oe_end < index)
2802                 ext = next_extent(ext);
2803         while (ext) {
2804                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2805
2806                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2807                         /* if ext is in urgent state, it means there must exist
2808                          * a page already having been flushed by write_page().
2809                          * We have to wait for this extent because we can't
2810                          * truncate that page.
2811                          */
2812                         OSC_EXTENT_DUMP(D_CACHE, ext,
2813                                         "waiting for busy extent\n");
2814                         waiting = osc_extent_get(ext);
2815                         break;
2816                 }
2817
2818                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:%llu.\n", size);
2819
2820                 osc_extent_get(ext);
2821                 if (ext->oe_state == OES_ACTIVE) {
2822                         /* though we grab inode mutex for write path, but we
2823                          * release it before releasing extent(in osc_io_end()),
2824                          * so there is a race window that an extent is still
2825                          * in OES_ACTIVE when truncate starts.
2826                          */
2827                         LASSERT(!ext->oe_trunc_pending);
2828                         ext->oe_trunc_pending = 1;
2829                 } else {
2830                         EASSERT(ext->oe_state == OES_CACHE, ext);
2831                         osc_extent_state_set(ext, OES_TRUNC);
2832                         osc_update_pending(obj, OBD_BRW_WRITE,
2833                                            -ext->oe_nr_pages);
2834                 }
2835                 EASSERT(list_empty(&ext->oe_link), ext);
2836                 list_add_tail(&ext->oe_link, &list);
2837
2838                 ext = next_extent(ext);
2839         }
2840         osc_object_unlock(obj);
2841
2842         osc_list_maint(cli, obj);
2843
2844         list_for_each_entry_safe(ext, temp, &list, oe_link) {
2845                 int rc;
2846
2847                 list_del_init(&ext->oe_link);
2848
2849                 /* extent may be in OES_ACTIVE state because inode mutex
2850                  * is released before osc_io_end() in file write case
2851                  */
2852                 if (ext->oe_state != OES_TRUNC)
2853                         osc_extent_wait(env, ext, OES_TRUNC);
2854
2855                 rc = osc_extent_truncate(ext, index, partial);
2856                 if (rc < 0) {
2857                         if (result == 0)
2858                                 result = rc;
2859
2860                         OSC_EXTENT_DUMP(D_ERROR, ext,
2861                                         "truncate error %d\n", rc);
2862                 } else if (ext->oe_nr_pages == 0) {
2863                         osc_extent_remove(ext);
2864                 } else {
2865                         /* this must be an overlapped extent which means only
2866                          * part of pages in this extent have been truncated.
2867                          */
2868                         EASSERTF(ext->oe_start <= index, ext,
2869                                  "trunc index = %lu/%d.\n", index, partial);
2870                         /* fix index to skip this partially truncated extent */
2871                         index = ext->oe_end + 1;
2872                         partial = false;
2873
2874                         /* we need to hold this extent in OES_TRUNC state so
2875                          * that no writeback will happen. This is to avoid
2876                          * BUG 17397.
2877                          * Only partial truncate can reach here, if @size is
2878                          * not zero, the caller should provide a valid @extp.
2879                          */
2880                         LASSERT(!*extp);
2881                         *extp = osc_extent_get(ext);
2882                         OSC_EXTENT_DUMP(D_CACHE, ext,
2883                                         "trunc at %llu\n", size);
2884                 }
2885                 osc_extent_put(env, ext);
2886         }
2887         if (waiting) {
2888                 int rc;
2889
2890                 /* ignore the result of osc_extent_wait the write initiator
2891                  * should take care of it.
2892                  */
2893                 rc = osc_extent_wait(env, waiting, OES_INV);
2894                 if (rc < 0)
2895                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2896
2897                 osc_extent_put(env, waiting);
2898                 waiting = NULL;
2899                 goto again;
2900         }
2901         return result;
2902 }
2903
2904 /**
2905  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2906  */
2907 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext)
2908 {
2909         if (ext) {
2910                 struct osc_object *obj = ext->oe_obj;
2911                 bool unplug = false;
2912
2913                 EASSERT(ext->oe_nr_pages > 0, ext);
2914                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2915                 EASSERT(!ext->oe_urgent, ext);
2916
2917                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2918                 osc_object_lock(obj);
2919                 osc_extent_state_set(ext, OES_CACHE);
2920                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2921                         ext->oe_urgent = 1;
2922                         list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2923                         unplug = true;
2924                 }
2925                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2926                 osc_object_unlock(obj);
2927                 osc_extent_put(env, ext);
2928
2929                 if (unplug)
2930                         osc_io_unplug_async(env, osc_cli(obj), obj);
2931         }
2932 }
2933
2934 /**
2935  * Wait for extents in a specific range to be written out.
2936  * The caller must have called osc_cache_writeback_range() to issue IO
2937  * otherwise it will take a long time for this function to finish.
2938  *
2939  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2940  * nobody else can dirty this range of file while we're waiting for
2941  * extents to be written.
2942  */
2943 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2944                          pgoff_t start, pgoff_t end)
2945 {
2946         struct osc_extent *ext;
2947         pgoff_t index = start;
2948         int result = 0;
2949
2950 again:
2951         osc_object_lock(obj);
2952         ext = osc_extent_search(obj, index);
2953         if (!ext)
2954                 ext = first_extent(obj);
2955         else if (ext->oe_end < index)
2956                 ext = next_extent(ext);
2957         while (ext) {
2958                 int rc;
2959
2960                 if (ext->oe_start > end)
2961                         break;
2962
2963                 if (!ext->oe_fsync_wait) {
2964                         ext = next_extent(ext);
2965                         continue;
2966                 }
2967
2968                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2969                              ext->oe_hp || ext->oe_urgent), ext);
2970                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2971                              !ext->oe_hp && ext->oe_urgent), ext);
2972
2973                 index = ext->oe_end + 1;
2974                 osc_extent_get(ext);
2975                 osc_object_unlock(obj);
2976
2977                 rc = osc_extent_wait(env, ext, OES_INV);
2978                 if (result == 0)
2979                         result = rc;
2980                 osc_extent_put(env, ext);
2981                 goto again;
2982         }
2983         osc_object_unlock(obj);
2984
2985         OSC_IO_DEBUG(obj, "sync file range.\n");
2986         return result;
2987 }
2988
2989 /**
2990  * Called to write out a range of osc object.
2991  *
2992  * @hp     : should be set this is caused by lock cancel;
2993  * @discard: is set if dirty pages should be dropped - file will be deleted or
2994  *         truncated, this implies there is no partially discarding extents.
2995  *
2996  * Return how many pages will be issued, or error code if error occurred.
2997  */
2998 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2999                               pgoff_t start, pgoff_t end, int hp, int discard)
3000 {
3001         struct osc_extent *ext;
3002         LIST_HEAD(discard_list);
3003         bool unplug = false;
3004         int result = 0;
3005
3006         osc_object_lock(obj);
3007         ext = osc_extent_search(obj, start);
3008         if (!ext)
3009                 ext = first_extent(obj);
3010         else if (ext->oe_end < start)
3011                 ext = next_extent(ext);
3012         while (ext) {
3013                 if (ext->oe_start > end)
3014                         break;
3015
3016                 ext->oe_fsync_wait = 1;
3017                 switch (ext->oe_state) {
3018                 case OES_CACHE:
3019                         result += ext->oe_nr_pages;
3020                         if (!discard) {
3021                                 struct list_head *list = NULL;
3022
3023                                 if (hp) {
3024                                         EASSERT(!ext->oe_hp, ext);
3025                                         ext->oe_hp = 1;
3026                                         list = &obj->oo_hp_exts;
3027                                 } else if (!ext->oe_urgent) {
3028                                         ext->oe_urgent = 1;
3029                                         list = &obj->oo_urgent_exts;
3030                                 }
3031                                 if (list)
3032                                         list_move_tail(&ext->oe_link, list);
3033                                 unplug = true;
3034                         } else {
3035                                 /* the only discarder is lock cancelling, so
3036                                  * [start, end] must contain this extent
3037                                  */
3038                                 EASSERT(ext->oe_start >= start &&
3039                                         ext->oe_max_end <= end, ext);
3040                                 osc_extent_state_set(ext, OES_LOCKING);
3041                                 ext->oe_owner = current;
3042                                 list_move_tail(&ext->oe_link, &discard_list);
3043                                 osc_update_pending(obj, OBD_BRW_WRITE,
3044                                                    -ext->oe_nr_pages);
3045                         }
3046                         break;
3047                 case OES_ACTIVE:
3048                         /* It's pretty bad to wait for ACTIVE extents, because
3049                          * we don't know how long we will wait for it to be
3050                          * flushed since it may be blocked at awaiting more
3051                          * grants. We do this for the correctness of fsync.
3052                          */
3053                         LASSERT(hp == 0 && discard == 0);
3054                         ext->oe_urgent = 1;
3055                         break;
3056                 case OES_TRUNC:
3057                         /* this extent is being truncated, can't do anything
3058                          * for it now. it will be set to urgent after truncate
3059                          * is finished in osc_cache_truncate_end().
3060                          */
3061                 default:
3062                         break;
3063                 }
3064                 ext = next_extent(ext);
3065         }
3066         osc_object_unlock(obj);
3067
3068         LASSERT(ergo(!discard, list_empty(&discard_list)));
3069         if (!list_empty(&discard_list)) {
3070                 struct osc_extent *tmp;
3071                 int rc;
3072
3073                 osc_list_maint(osc_cli(obj), obj);
3074                 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
3075                         list_del_init(&ext->oe_link);
3076                         EASSERT(ext->oe_state == OES_LOCKING, ext);
3077
3078                         /* Discard caching pages. We don't actually write this
3079                          * extent out but we complete it as if we did.
3080                          */
3081                         rc = osc_extent_make_ready(env, ext);
3082                         if (unlikely(rc < 0)) {
3083                                 OSC_EXTENT_DUMP(D_ERROR, ext,
3084                                                 "make_ready returned %d\n", rc);
3085                                 if (result >= 0)
3086                                         result = rc;
3087                         }
3088
3089                         /* finish the extent as if the pages were sent */
3090                         osc_extent_finish(env, ext, 0, 0);
3091                 }
3092         }
3093
3094         if (unplug)
3095                 osc_io_unplug(env, osc_cli(obj), obj);
3096
3097         if (hp || discard) {
3098                 int rc;
3099
3100                 rc = osc_cache_wait_range(env, obj, start, end);
3101                 if (result >= 0 && rc < 0)
3102                         result = rc;
3103         }
3104
3105         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3106         return result;
3107 }
3108
3109 /**
3110  * Returns a list of pages by a given [start, end] of \a obj.
3111  *
3112  * \param resched If not NULL, then we give up before hogging CPU for too
3113  * long and set *resched = 1, in that case caller should implement a retry
3114  * logic.
3115  *
3116  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3117  * crucial in the face of [offset, EOF] locks.
3118  *
3119  * Return at least one page in @queue unless there is no covered page.
3120  */
3121 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3122                          struct osc_object *osc, pgoff_t start, pgoff_t end,
3123                          osc_page_gang_cbt cb, void *cbdata)
3124 {
3125         struct osc_page *ops;
3126         void            **pvec;
3127         pgoff_t         idx;
3128         unsigned int    nr;
3129         unsigned int    i;
3130         unsigned int    j;
3131         int             res = CLP_GANG_OKAY;
3132         bool            tree_lock = true;
3133
3134         idx = start;
3135         pvec = osc_env_info(env)->oti_pvec;
3136         spin_lock(&osc->oo_tree_lock);
3137         while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3138                                             idx, OTI_PVEC_SIZE)) > 0) {
3139                 struct cl_page *page;
3140                 bool end_of_region = false;
3141
3142                 for (i = 0, j = 0; i < nr; ++i) {
3143                         ops = pvec[i];
3144                         pvec[i] = NULL;
3145
3146                         idx = osc_index(ops);
3147                         if (idx > end) {
3148                                 end_of_region = true;
3149                                 break;
3150                         }
3151
3152                         page = ops->ops_cl.cpl_page;
3153                         LASSERT(page->cp_type == CPT_CACHEABLE);
3154                         if (page->cp_state == CPS_FREEING)
3155                                 continue;
3156
3157                         cl_page_get(page);
3158                         lu_ref_add_atomic(&page->cp_reference,
3159                                           "gang_lookup", current);
3160                         pvec[j++] = ops;
3161                 }
3162                 ++idx;
3163
3164                 /*
3165                  * Here a delicate locking dance is performed. Current thread
3166                  * holds a reference to a page, but has to own it before it
3167                  * can be placed into queue. Owning implies waiting, so
3168                  * radix-tree lock is to be released. After a wait one has to
3169                  * check that pages weren't truncated (cl_page_own() returns
3170                  * error in the latter case).
3171                  */
3172                 spin_unlock(&osc->oo_tree_lock);
3173                 tree_lock = false;
3174
3175                 for (i = 0; i < j; ++i) {
3176                         ops = pvec[i];
3177                         if (res == CLP_GANG_OKAY)
3178                                 res = (*cb)(env, io, ops, cbdata);
3179
3180                         page = ops->ops_cl.cpl_page;
3181                         lu_ref_del(&page->cp_reference, "gang_lookup", current);
3182                         cl_page_put(env, page);
3183                 }
3184                 if (nr < OTI_PVEC_SIZE || end_of_region)
3185                         break;
3186
3187                 if (res == CLP_GANG_OKAY && need_resched())
3188                         res = CLP_GANG_RESCHED;
3189                 if (res != CLP_GANG_OKAY)
3190                         break;
3191
3192                 spin_lock(&osc->oo_tree_lock);
3193                 tree_lock = true;
3194         }
3195         if (tree_lock)
3196                 spin_unlock(&osc->oo_tree_lock);
3197         return res;
3198 }
3199
3200 /**
3201  * Check if page @page is covered by an extra lock or discard it.
3202  */
3203 static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3204                                 struct osc_page *ops, void *cbdata)
3205 {
3206         struct osc_thread_info *info = osc_env_info(env);
3207         struct osc_object *osc = cbdata;
3208         pgoff_t index;
3209
3210         index = osc_index(ops);
3211         if (index >= info->oti_fn_index) {
3212                 struct ldlm_lock *tmp;
3213                 struct cl_page *page = ops->ops_cl.cpl_page;
3214
3215                 /* refresh non-overlapped index */
3216                 tmp = osc_dlmlock_at_pgoff(env, osc, index,
3217                                            OSC_DAP_FL_TEST_LOCK);
3218                 if (tmp) {
3219                         __u64 end = tmp->l_policy_data.l_extent.end;
3220                         /* Cache the first-non-overlapped index so as to skip
3221                          * all pages within [index, oti_fn_index). This is safe
3222                          * because if tmp lock is canceled, it will discard
3223                          * these pages.
3224                          */
3225                         info->oti_fn_index = cl_index(osc2cl(osc), end + 1);
3226                         if (end == OBD_OBJECT_EOF)
3227                                 info->oti_fn_index = CL_PAGE_EOF;
3228                         LDLM_LOCK_PUT(tmp);
3229                 } else if (cl_page_own(env, io, page) == 0) {
3230                         /* discard the page */
3231                         cl_page_discard(env, io, page);
3232                         cl_page_disown(env, io, page);
3233                 } else {
3234                         LASSERT(page->cp_state == CPS_FREEING);
3235                 }
3236         }
3237
3238         info->oti_next_index = index + 1;
3239         return CLP_GANG_OKAY;
3240 }
3241
3242 static int discard_cb(const struct lu_env *env, struct cl_io *io,
3243                       struct osc_page *ops, void *cbdata)
3244 {
3245         struct osc_thread_info *info = osc_env_info(env);
3246         struct cl_page *page = ops->ops_cl.cpl_page;
3247
3248         /* page is top page. */
3249         info->oti_next_index = osc_index(ops) + 1;
3250         if (cl_page_own(env, io, page) == 0) {
3251                 if (page->cp_type == CPT_CACHEABLE &&
3252                     PageDirty(cl_page_vmpage(page)))
3253                         CL_PAGE_DEBUG(D_ERROR, env, page,
3254                                       "discard dirty page?\n");
3255
3256                 /* discard the page */
3257                 cl_page_discard(env, io, page);
3258                 cl_page_disown(env, io, page);
3259         } else {
3260                 LASSERT(page->cp_state == CPS_FREEING);
3261         }
3262
3263         return CLP_GANG_OKAY;
3264 }
3265
3266 /**
3267  * Discard pages protected by the given lock. This function traverses radix
3268  * tree to find all covering pages and discard them. If a page is being covered
3269  * by other locks, it should remain in cache.
3270  *
3271  * If error happens on any step, the process continues anyway (the reasoning
3272  * behind this being that lock cancellation cannot be delayed indefinitely).
3273  */
3274 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3275                            pgoff_t start, pgoff_t end, enum cl_lock_mode mode)
3276 {
3277         struct osc_thread_info *info = osc_env_info(env);
3278         struct cl_io *io = &info->oti_io;
3279         osc_page_gang_cbt cb;
3280         int res;
3281         int result;
3282
3283         io->ci_obj = cl_object_top(osc2cl(osc));
3284         io->ci_ignore_layout = 1;
3285         result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3286         if (result != 0)
3287                 goto out;
3288
3289         cb = mode == CLM_READ ? check_and_discard_cb : discard_cb;
3290         info->oti_fn_index = start;
3291         info->oti_next_index = start;
3292         do {
3293                 res = osc_page_gang_lookup(env, io, osc,
3294                                            info->oti_next_index, end, cb, osc);
3295                 if (info->oti_next_index > end)
3296                         break;
3297
3298                 if (res == CLP_GANG_RESCHED)
3299                         cond_resched();
3300         } while (res != CLP_GANG_OKAY);
3301 out:
3302         cl_io_fini(env, io);
3303         return result;
3304 }
3305
3306 /** @} osc */