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