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         if (!osc_object_is_locked(obj))
 188                 GOTO(out, rc = 9);
 189
 190         if (ext->oe_state >= OES_STATE_MAX)
 191                 GOTO(out, rc = 10);
 192
 193         if (atomic_read(&ext->oe_refc) <= 0)
 194                 GOTO(out, rc = 20);
 195
 196         if (atomic_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
 197                 GOTO(out, rc = 30);
 198
 199         switch (ext->oe_state) {
 200         case OES_INV:
 201                 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
 202                         GOTO(out, rc = 35);
 203                 GOTO(out, rc = 0);
 204                 break;
 205         case OES_ACTIVE:
 206                 if (atomic_read(&ext->oe_users) == 0)
 207                         GOTO(out, rc = 40);
 208                 if (ext->oe_hp)
 209                         GOTO(out, rc = 50);
 210                 if (ext->oe_fsync_wait && !ext->oe_urgent)
 211                         GOTO(out, rc = 55);
 212                 break;
 213         case OES_CACHE:
 214                 if (ext->oe_grants == 0)
 215                         GOTO(out, rc = 60);
 216                 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
 217                         GOTO(out, rc = 65);
 218                 /* fallthrough */
 219         default:
 220                 if (atomic_read(&ext->oe_users) > 0)
 221                         GOTO(out, rc = 70);
 222         }
 223
 224         if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
 225                 GOTO(out, rc = 80);
 226
 227         if (ext->oe_sync && ext->oe_grants > 0)
 228                 GOTO(out, rc = 90);
 229
 230         if (ext->oe_dlmlock != NULL &&
 231             ext->oe_dlmlock->l_resource->lr_type == LDLM_EXTENT &&
 232             !ldlm_is_failed(ext->oe_dlmlock)) {
 233                 struct ldlm_extent *extent;
 234
 235                 extent = &ext->oe_dlmlock->l_policy_data.l_extent;
 236                 if (!(extent->start <= cl_offset(osc2cl(obj), ext->oe_start) &&
 237                       extent->end   >= cl_offset(osc2cl(obj), ext->oe_max_end)))
 238                         GOTO(out, rc = 100);
 239
 240                 if (!(ext->oe_dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP)))
 241                         GOTO(out, rc = 102);
 242         }
 243
 244         if (ext->oe_nr_pages > ext->oe_mppr)
 245                 GOTO(out, rc = 105);
 246
 247         /* Do not verify page list if extent is in RPC. This is because an
 248          * in-RPC extent is supposed to be exclusively accessible w/o lock. */
 249         if (ext->oe_state > OES_CACHE)
 250                 GOTO(out, rc = 0);
 251
 252         if (!extent_debug)
 253                 GOTO(out, rc = 0);
 254
 255         page_count = 0;
 256         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
 257                 pgoff_t index = osc_index(oap2osc(oap));
 258                 ++page_count;
 259                 if (index > ext->oe_end || index < ext->oe_start)
 260                         GOTO(out, rc = 110);
 261         }
 262         if (page_count != ext->oe_nr_pages)
 263                 GOTO(out, rc = 120);
 264
 265 out:
 266         if (rc != 0)
 267                 OSC_EXTENT_DUMP_WITH_LOC(__FILE__, func, line, D_ERROR, ext,
 268                                          "sanity check %p failed: rc = %d\n",
 269                                          ext, rc);
 270         return rc;
 271 }
 272
 273 #define sanity_check_nolock(ext) \
 274         osc_extent_sanity_check0(ext, __func__, __LINE__)
 275
 276 #define sanity_check(ext) ({                                                   \
 277         int __res;                                                             \
 278         osc_object_lock((ext)->oe_obj);                                        \
 279         __res = sanity_check_nolock(ext);                                      \
 280         osc_object_unlock((ext)->oe_obj);                                      \
 281         __res;                                                                 \
 282 })
 283
 284
 285 /**
 286  * sanity check - to make sure there is no overlapped extent in the tree.
 287  */
 288 static int osc_extent_is_overlapped(struct osc_object *obj,
 289                                     struct osc_extent *ext)
 290 {
 291         struct osc_extent *tmp;
 292
 293         LASSERT(osc_object_is_locked(obj));
 294
 295         if (!extent_debug)
 296                 return 0;
 297
 298         for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
 299                 if (tmp == ext)
 300                         continue;
 301                 if (tmp->oe_end >= ext->oe_start &&
 302                     tmp->oe_start <= ext->oe_end)
 303                         return 1;
 304         }
 305         return 0;
 306 }
 307
 308 static void osc_extent_state_set(struct osc_extent *ext, int state)
 309 {
 310         LASSERT(osc_object_is_locked(ext->oe_obj));
 311         LASSERT(state >= OES_INV && state < OES_STATE_MAX);
 312
 313         /* Never try to sanity check a state changing extent :-) */
 314         /* LASSERT(sanity_check_nolock(ext) == 0); */
 315
 316         /* TODO: validate the state machine */
 317         ext->oe_state = state;
 318         wake_up_all(&ext->oe_waitq);
 319 }
 320
 321 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
 322 {
 323         struct osc_extent *ext;
 324
 325         OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_NOFS);
 326         if (ext == NULL)
 327                 return NULL;
 328
 329         RB_CLEAR_NODE(&ext->oe_node);
 330         ext->oe_obj = obj;
 331         cl_object_get(osc2cl(obj));
 332         atomic_set(&ext->oe_refc, 1);
 333         atomic_set(&ext->oe_users, 0);
 334         INIT_LIST_HEAD(&ext->oe_link);
 335         ext->oe_state = OES_INV;
 336         INIT_LIST_HEAD(&ext->oe_pages);
 337         init_waitqueue_head(&ext->oe_waitq);
 338         ext->oe_dlmlock = NULL;
 339
 340         return ext;
 341 }
 342
 343 static void osc_extent_free(struct osc_extent *ext)
 344 {
 345         OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
 346 }
 347
 348 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
 349 {
 350         LASSERT(atomic_read(&ext->oe_refc) >= 0);
 351         atomic_inc(&ext->oe_refc);
 352         return ext;
 353 }
 354
 355 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
 356 {
 357         LASSERT(atomic_read(&ext->oe_refc) > 0);
 358         if (atomic_dec_and_test(&ext->oe_refc)) {
 359                 LASSERT(list_empty(&ext->oe_link));
 360                 LASSERT(atomic_read(&ext->oe_users) == 0);
 361                 LASSERT(ext->oe_state == OES_INV);
 362                 LASSERT(!ext->oe_intree);
 363
 364                 if (ext->oe_dlmlock != NULL) {
 365                         lu_ref_add(&ext->oe_dlmlock->l_reference,
 366                                    "osc_extent", ext);
 367                         LDLM_LOCK_PUT(ext->oe_dlmlock);
 368                         ext->oe_dlmlock = NULL;
 369                 }
 370                 cl_object_put(env, osc2cl(ext->oe_obj));
 371                 osc_extent_free(ext);
 372         }
 373 }
 374
 375 /**
 376  * osc_extent_put_trust() is a special version of osc_extent_put() when
 377  * it's known that the caller is not the last user. This is to address the
 378  * problem of lacking of lu_env ;-).
 379  */
 380 static void osc_extent_put_trust(struct osc_extent *ext)
 381 {
 382         LASSERT(atomic_read(&ext->oe_refc) > 1);
 383         LASSERT(osc_object_is_locked(ext->oe_obj));
 384         atomic_dec(&ext->oe_refc);
 385 }
 386
 387 /**
 388  * Return the extent which includes pgoff @index, or return the greatest
 389  * previous extent in the tree.
 390  */
 391 static struct osc_extent *osc_extent_search(struct osc_object *obj,
 392                                             pgoff_t index)
 393 {
 394         struct rb_node    *n = obj->oo_root.rb_node;
 395         struct osc_extent *tmp, *p = NULL;
 396
 397         LASSERT(osc_object_is_locked(obj));
 398         while (n != NULL) {
 399                 tmp = rb_extent(n);
 400                 if (index < tmp->oe_start) {
 401                         n = n->rb_left;
 402                 } else if (index > tmp->oe_end) {
 403                         p = rb_extent(n);
 404                         n = n->rb_right;
 405                 } else {
 406                         return tmp;
 407                 }
 408         }
 409         return p;
 410 }
 411
 412 /*
 413  * Return the extent covering @index, otherwise return NULL.
 414  * caller must have held object lock.
 415  */
 416 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
 417                                             pgoff_t index)
 418 {
 419         struct osc_extent *ext;
 420
 421         ext = osc_extent_search(obj, index);
 422         if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
 423                 return osc_extent_get(ext);
 424         return NULL;
 425 }
 426
 427 /* caller must have held object lock. */
 428 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
 429 {
 430         struct rb_node   **n      = &obj->oo_root.rb_node;
 431         struct rb_node    *parent = NULL;
 432         struct osc_extent *tmp;
 433
 434         LASSERT(ext->oe_intree == 0);
 435         LASSERT(ext->oe_obj == obj);
 436         LASSERT(osc_object_is_locked(obj));
 437         while (*n != NULL) {
 438                 tmp = rb_extent(*n);
 439                 parent = *n;
 440
 441                 if (ext->oe_end < tmp->oe_start)
 442                         n = &(*n)->rb_left;
 443                 else if (ext->oe_start > tmp->oe_end)
 444                         n = &(*n)->rb_right;
 445                 else
 446                         EASSERTF(0, tmp, EXTSTR"\n", EXTPARA(ext));
 447         }
 448         rb_link_node(&ext->oe_node, parent, n);
 449         rb_insert_color(&ext->oe_node, &obj->oo_root);
 450         osc_extent_get(ext);
 451         ext->oe_intree = 1;
 452 }
 453
 454 /* caller must have held object lock. */
 455 static void osc_extent_erase(struct osc_extent *ext)
 456 {
 457         struct osc_object *obj = ext->oe_obj;
 458         LASSERT(osc_object_is_locked(obj));
 459         if (ext->oe_intree) {
 460                 rb_erase(&ext->oe_node, &obj->oo_root);
 461                 ext->oe_intree = 0;
 462                 /* rbtree held a refcount */
 463                 osc_extent_put_trust(ext);
 464         }
 465 }
 466
 467 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
 468 {
 469         struct osc_object *obj = ext->oe_obj;
 470
 471         LASSERT(osc_object_is_locked(obj));
 472         LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
 473         if (ext->oe_state == OES_CACHE) {
 474                 osc_extent_state_set(ext, OES_ACTIVE);
 475                 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
 476         }
 477         atomic_inc(&ext->oe_users);
 478         list_del_init(&ext->oe_link);
 479         return osc_extent_get(ext);
 480 }
 481
 482 static void __osc_extent_remove(struct osc_extent *ext)
 483 {
 484         LASSERT(osc_object_is_locked(ext->oe_obj));
 485         LASSERT(list_empty(&ext->oe_pages));
 486         osc_extent_erase(ext);
 487         list_del_init(&ext->oe_link);
 488         osc_extent_state_set(ext, OES_INV);
 489         OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
 490 }
 491
 492 static void osc_extent_remove(struct osc_extent *ext)
 493 {
 494         struct osc_object *obj = ext->oe_obj;
 495
 496         osc_object_lock(obj);
 497         __osc_extent_remove(ext);
 498         osc_object_unlock(obj);
 499 }
 500
 501 /**
 502  * This function is used to merge extents to get better performance. It checks
 503  * if @cur and @victim are contiguous at block level.
 504  */
 505 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
 506                             struct osc_extent *victim)
 507 {
 508         struct osc_object       *obj = cur->oe_obj;
 509         struct client_obd       *cli = osc_cli(obj);
 510         pgoff_t                  chunk_start;
 511         pgoff_t                  chunk_end;
 512         int                      ppc_bits;
 513
 514         LASSERT(cur->oe_state == OES_CACHE);
 515         LASSERT(osc_object_is_locked(obj));
 516         if (victim == NULL)
 517                 return -EINVAL;
 518
 519         if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
 520                 return -EBUSY;
 521
 522         if (cur->oe_max_end != victim->oe_max_end)
 523                 return -ERANGE;
 524
 525         LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
 526         ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_SHIFT;
 527         chunk_start = cur->oe_start >> ppc_bits;
 528         chunk_end   = cur->oe_end   >> ppc_bits;
 529         if (chunk_start   != (victim->oe_end >> ppc_bits) + 1 &&
 530             chunk_end + 1 != victim->oe_start >> ppc_bits)
 531                 return -ERANGE;
 532
 533         /* overall extent size should not exceed the max supported limit
 534          * reported by the server */
 535         if (cur->oe_end - cur->oe_start + 1 +
 536             victim->oe_end - victim->oe_start + 1 > cli->cl_max_extent_pages)
 537                 return -ERANGE;
 538
 539         OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
 540
 541         cur->oe_start     = min(cur->oe_start, victim->oe_start);
 542         cur->oe_end       = max(cur->oe_end,   victim->oe_end);
 543         /* per-extent tax should be accounted only once for the whole extent */
 544         cur->oe_grants   += victim->oe_grants - cli->cl_grant_extent_tax;
 545         cur->oe_nr_pages += victim->oe_nr_pages;
 546         /* only the following bits are needed to merge */
 547         cur->oe_urgent   |= victim->oe_urgent;
 548         cur->oe_memalloc |= victim->oe_memalloc;
 549         list_splice_init(&victim->oe_pages, &cur->oe_pages);
 550         list_del_init(&victim->oe_link);
 551         victim->oe_nr_pages = 0;
 552
 553         osc_extent_get(victim);
 554         __osc_extent_remove(victim);
 555         osc_extent_put(env, victim);
 556
 557         OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
 558         return 0;
 559 }
 560
 561 /**
 562  * Drop user count of osc_extent, and unplug IO asynchronously.
 563  */
 564 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
 565 {
 566         struct osc_object *obj = ext->oe_obj;
 567         struct client_obd *cli = osc_cli(obj);
 568         int rc = 0;
 569         ENTRY;
 570
 571         LASSERT(atomic_read(&ext->oe_users) > 0);
 572         LASSERT(sanity_check(ext) == 0);
 573         LASSERT(ext->oe_grants > 0);
 574
 575         if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
 576                 LASSERT(ext->oe_state == OES_ACTIVE);
 577                 if (ext->oe_trunc_pending) {
 578                         /* a truncate process is waiting for this extent.
 579                          * This may happen due to a race, check
 580                          * osc_cache_truncate_start(). */
 581                         osc_extent_state_set(ext, OES_TRUNC);
 582                         ext->oe_trunc_pending = 0;
 583                         osc_object_unlock(obj);
 584                 } else {
 585                         int grant = 0;
 586
 587                         osc_extent_state_set(ext, OES_CACHE);
 588                         osc_update_pending(obj, OBD_BRW_WRITE,
 589                                            ext->oe_nr_pages);
 590
 591                         /* try to merge the previous and next extent. */
 592                         if (osc_extent_merge(env, ext, prev_extent(ext)) == 0)
 593                                 grant += cli->cl_grant_extent_tax;
 594                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 595                                 grant += cli->cl_grant_extent_tax;
 596
 597                         if (ext->oe_urgent)
 598                                 list_move_tail(&ext->oe_link,
 599                                                &obj->oo_urgent_exts);
 600                         else if (ext->oe_nr_pages == ext->oe_mppr) {
 601                                 list_move_tail(&ext->oe_link,
 602                                                &obj->oo_full_exts);
 603                         }
 604                         osc_object_unlock(obj);
 605                         if (grant > 0)
 606                                 osc_unreserve_grant(cli, 0, grant);
 607                 }
 608
 609                 osc_io_unplug_async(env, cli, obj);
 610         }
 611         osc_extent_put(env, ext);
 612         RETURN(rc);
 613 }
 614
 615 static inline bool
 616 overlapped(const struct osc_extent *ex1, const struct osc_extent *ex2)
 617 {
 618         return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
 619 }
 620
 621 /**
 622  * Find or create an extent which includes @index, core function to manage
 623  * extent tree.
 624  */
 625 static struct osc_extent *osc_extent_find(const struct lu_env *env,
 626                                           struct osc_object *obj, pgoff_t index,
 627                                           unsigned int *grants)
 628 {
 629         struct client_obd *cli = osc_cli(obj);
 630         struct osc_lock   *olck;
 631         struct cl_lock_descr *descr;
 632         struct osc_extent *cur;
 633         struct osc_extent *ext;
 634         struct osc_extent *conflict = NULL;
 635         struct osc_extent *found = NULL;
 636         pgoff_t    chunk;
 637         pgoff_t    max_end;
 638         unsigned int max_pages; /* max_pages_per_rpc */
 639         unsigned int chunksize;
 640         int        ppc_bits; /* pages per chunk bits */
 641         pgoff_t    chunk_mask;
 642         int        rc;
 643         ENTRY;
 644
 645         cur = osc_extent_alloc(obj);
 646         if (cur == NULL)
 647                 RETURN(ERR_PTR(-ENOMEM));
 648
 649         olck = osc_env_io(env)->oi_write_osclock;
 650         LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
 651         LASSERT(olck->ols_state == OLS_GRANTED);
 652
 653         descr = &olck->ols_cl.cls_lock->cll_descr;
 654         LASSERT(descr->cld_mode >= CLM_WRITE);
 655
 656         LASSERTF(cli->cl_chunkbits >= PAGE_SHIFT,
 657                  "chunkbits: %u\n", cli->cl_chunkbits);
 658         ppc_bits   = cli->cl_chunkbits - PAGE_SHIFT;
 659         chunk_mask = ~((1 << ppc_bits) - 1);
 660         chunksize  = 1 << cli->cl_chunkbits;
 661         chunk      = index >> ppc_bits;
 662
 663         /* align end to RPC edge. */
 664         max_pages = cli->cl_max_pages_per_rpc;
 665         if ((max_pages & ~chunk_mask) != 0) {
 666                 CERROR("max_pages: %#x chunkbits: %u chunk_mask: %#lx\n",
 667                        max_pages, cli->cl_chunkbits, chunk_mask);
 668                 RETURN(ERR_PTR(-EINVAL));
 669         }
 670         max_end = index - (index % max_pages) + max_pages - 1;
 671         max_end = min_t(pgoff_t, max_end, descr->cld_end);
 672
 673         /* initialize new extent by parameters so far */
 674         cur->oe_max_end = max_end;
 675         cur->oe_start   = index & chunk_mask;
 676         cur->oe_end     = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
 677         if (cur->oe_start < descr->cld_start)
 678                 cur->oe_start = descr->cld_start;
 679         if (cur->oe_end > max_end)
 680                 cur->oe_end = max_end;
 681         cur->oe_grants  = 0;
 682         cur->oe_mppr    = max_pages;
 683         if (olck->ols_dlmlock != NULL) {
 684                 LASSERT(olck->ols_hold);
 685                 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
 686                 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
 687         }
 688
 689         /* grants has been allocated by caller */
 690         LASSERTF(*grants >= chunksize + cli->cl_grant_extent_tax,
 691                  "%u/%u/%u.\n", *grants, chunksize, cli->cl_grant_extent_tax);
 692         LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
 693                  EXTPARA(cur));
 694
 695 restart:
 696         osc_object_lock(obj);
 697         ext = osc_extent_search(obj, cur->oe_start);
 698         if (ext == NULL)
 699                 ext = first_extent(obj);
 700         while (ext != NULL) {
 701                 pgoff_t ext_chk_start = ext->oe_start >> ppc_bits;
 702                 pgoff_t ext_chk_end   = ext->oe_end   >> ppc_bits;
 703
 704                 LASSERT(sanity_check_nolock(ext) == 0);
 705                 if (chunk > ext_chk_end + 1 || chunk < ext_chk_start)
 706                         break;
 707
 708                 /* if covering by different locks, no chance to match */
 709                 if (olck->ols_dlmlock != ext->oe_dlmlock) {
 710                         EASSERTF(!overlapped(ext, cur), ext,
 711                                  EXTSTR"\n", EXTPARA(cur));
 712
 713                         ext = next_extent(ext);
 714                         continue;
 715                 }
 716
 717                 /* discontiguous chunks? */
 718                 if (chunk + 1 < ext_chk_start) {
 719                         ext = next_extent(ext);
 720                         continue;
 721                 }
 722
 723                 /* ok, from now on, ext and cur have these attrs:
 724                  * 1. covered by the same lock
 725                  * 2. contiguous at chunk level or overlapping. */
 726
 727                 if (overlapped(ext, cur)) {
 728                         /* cur is the minimum unit, so overlapping means
 729                          * full contain. */
 730                         EASSERTF((ext->oe_start <= cur->oe_start &&
 731                                   ext->oe_end >= cur->oe_end),
 732                                  ext, EXTSTR"\n", EXTPARA(cur));
 733
 734                         if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
 735                                 /* for simplicity, we wait for this extent to
 736                                  * finish before going forward. */
 737                                 conflict = osc_extent_get(ext);
 738                                 break;
 739                         }
 740
 741                         found = osc_extent_hold(ext);
 742                         break;
 743                 }
 744
 745                 /* non-overlapped extent */
 746                 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
 747                         /* we can't do anything for a non OES_CACHE extent, or
 748                          * if there is someone waiting for this extent to be
 749                          * flushed, try next one. */
 750                         ext = next_extent(ext);
 751                         continue;
 752                 }
 753
 754                 /* check if they belong to the same rpc slot before trying to
 755                  * merge. the extents are not overlapped and contiguous at
 756                  * chunk level to get here. */
 757                 if (ext->oe_max_end != max_end) {
 758                         /* if they don't belong to the same RPC slot or
 759                          * max_pages_per_rpc has ever changed, do not merge. */
 760                         ext = next_extent(ext);
 761                         continue;
 762                 }
 763
 764                 /* check whether maximum extent size will be hit */
 765                 if ((ext_chk_end - ext_chk_start + 1 + 1) << ppc_bits >
 766                     cli->cl_max_extent_pages) {
 767                         ext = next_extent(ext);
 768                         continue;
 769                 }
 770
 771                 /* it's required that an extent must be contiguous at chunk
 772                  * level so that we know the whole extent is covered by grant
 773                  * (the pages in the extent are NOT required to be contiguous).
 774                  * Otherwise, it will be too much difficult to know which
 775                  * chunks have grants allocated. */
 776
 777                 /* try to do front merge - extend ext's start */
 778                 if (chunk + 1 == ext_chk_start) {
 779                         /* ext must be chunk size aligned */
 780                         EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
 781
 782                         /* pull ext's start back to cover cur */
 783                         ext->oe_start   = cur->oe_start;
 784                         ext->oe_grants += chunksize;
 785                         LASSERT(*grants >= chunksize);
 786                         *grants -= chunksize;
 787
 788                         found = osc_extent_hold(ext);
 789                 } else if (chunk == ext_chk_end + 1) {
 790                         /* rear merge */
 791                         ext->oe_end     = cur->oe_end;
 792                         ext->oe_grants += chunksize;
 793                         LASSERT(*grants >= chunksize);
 794                         *grants -= chunksize;
 795
 796                         /* try to merge with the next one because we just fill
 797                          * in a gap */
 798                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 799                                 /* we can save extent tax from next extent */
 800                                 *grants += cli->cl_grant_extent_tax;
 801
 802                         found = osc_extent_hold(ext);
 803                 }
 804                 if (found != NULL)
 805                         break;
 806
 807                 ext = next_extent(ext);
 808         }
 809
 810         osc_extent_tree_dump(D_CACHE, obj);
 811         if (found != NULL) {
 812                 LASSERT(conflict == NULL);
 813                 if (!IS_ERR(found)) {
 814                         LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
 815                         OSC_EXTENT_DUMP(D_CACHE, found,
 816                                         "found caching ext for %lu.\n", index);
 817                 }
 818         } else if (conflict == NULL) {
 819                 /* create a new extent */
 820                 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
 821                 cur->oe_grants = chunksize + cli->cl_grant_extent_tax;
 822                 LASSERT(*grants >= cur->oe_grants);
 823                 *grants -= cur->oe_grants;
 824
 825                 cur->oe_state = OES_CACHE;
 826                 found = osc_extent_hold(cur);
 827                 osc_extent_insert(obj, cur);
 828                 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
 829                                 index, descr->cld_end);
 830         }
 831         osc_object_unlock(obj);
 832
 833         if (conflict != NULL) {
 834                 LASSERT(found == NULL);
 835
 836                 /* waiting for IO to finish. Please notice that it's impossible
 837                  * to be an OES_TRUNC extent. */
 838                 rc = osc_extent_wait(env, conflict, OES_INV);
 839                 osc_extent_put(env, conflict);
 840                 conflict = NULL;
 841                 if (rc < 0)
 842                         GOTO(out, found = ERR_PTR(rc));
 843
 844                 goto restart;
 845         }
 846         EXIT;
 847
 848 out:
 849         osc_extent_put(env, cur);
 850         return found;
 851 }
 852
 853 /**
 854  * Called when IO is finished to an extent.
 855  */
 856 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
 857                       int sent, int rc)
 858 {
 859         struct client_obd *cli = osc_cli(ext->oe_obj);
 860         struct osc_async_page *oap;
 861         struct osc_async_page *tmp;
 862         int nr_pages = ext->oe_nr_pages;
 863         int lost_grant = 0;
 864         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
 865         loff_t last_off = 0;
 866         int last_count = -1;
 867         ENTRY;
 868
 869         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
 870
 871         ext->oe_rc = rc ?: ext->oe_nr_pages;
 872         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
 873
 874         osc_lru_add_batch(cli, &ext->oe_pages);
 875         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 876                                      oap_pending_item) {
 877                 list_del_init(&oap->oap_rpc_item);
 878                 list_del_init(&oap->oap_pending_item);
 879                 if (last_off <= oap->oap_obj_off) {
 880                         last_off = oap->oap_obj_off;
 881                         last_count = oap->oap_count;
 882                 }
 883
 884                 --ext->oe_nr_pages;
 885                 osc_ap_completion(env, cli, oap, sent, rc);
 886         }
 887         EASSERT(ext->oe_nr_pages == 0, ext);
 888
 889         if (!sent) {
 890                 lost_grant = ext->oe_grants;
 891         } else if (blocksize < PAGE_SIZE &&
 892                    last_count != PAGE_SIZE) {
 893                 /* For short writes we shouldn't count parts of pages that
 894                  * span a whole chunk on the OST side, or our accounting goes
 895                  * wrong.  Should match the code in filter_grant_check. */
 896                 int offset = last_off & ~PAGE_MASK;
 897                 int count = last_count + (offset & (blocksize - 1));
 898                 int end = (offset + last_count) & (blocksize - 1);
 899                 if (end)
 900                         count += blocksize - end;
 901
 902                 lost_grant = PAGE_SIZE - count;
 903         }
 904         if (ext->oe_grants > 0)
 905                 osc_free_grant(cli, nr_pages, lost_grant, ext->oe_grants);
 906
 907         osc_extent_remove(ext);
 908         /* put the refcount for RPC */
 909         osc_extent_put(env, ext);
 910         RETURN(0);
 911 }
 912
 913 static int extent_wait_cb(struct osc_extent *ext, enum osc_extent_state state)
 914 {
 915         int ret;
 916
 917         osc_object_lock(ext->oe_obj);
 918         ret = ext->oe_state == state;
 919         osc_object_unlock(ext->oe_obj);
 920
 921         return ret;
 922 }
 923
 924 /**
 925  * Wait for the extent's state to become @state.
 926  */
 927 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
 928                            enum osc_extent_state state)
 929 {
 930         struct osc_object *obj = ext->oe_obj;
 931         int rc = 0;
 932         ENTRY;
 933
 934         osc_object_lock(obj);
 935         LASSERT(sanity_check_nolock(ext) == 0);
 936         /* `Kick' this extent only if the caller is waiting for it to be
 937          * written out. */
 938         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
 939                 if (ext->oe_state == OES_ACTIVE) {
 940                         ext->oe_urgent = 1;
 941                 } else if (ext->oe_state == OES_CACHE) {
 942                         ext->oe_urgent = 1;
 943                         osc_extent_hold(ext);
 944                         rc = 1;
 945                 }
 946         }
 947         osc_object_unlock(obj);
 948         if (rc == 1)
 949                 osc_extent_release(env, ext);
 950
 951         /* wait for the extent until its state becomes @state */
 952         rc = wait_event_idle_timeout(ext->oe_waitq, extent_wait_cb(ext, state),
 953                                      cfs_time_seconds(600));
 954         if (rc == 0) {
 955                 OSC_EXTENT_DUMP(D_ERROR, ext,
 956                         "%s: wait ext to %u timedout, recovery in progress?\n",
 957                         cli_name(osc_cli(obj)), state);
 958
 959                 wait_event_idle(ext->oe_waitq, extent_wait_cb(ext, state));
 960         }
 961         if (ext->oe_rc < 0)
 962                 rc = ext->oe_rc;
 963         else
 964                 rc = 0;
 965         RETURN(rc);
 966 }
 967
 968 /**
 969  * Discard pages with index greater than @size. If @ext is overlapped with
 970  * @size, then partial truncate happens.
 971  */
 972 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
 973                                 bool partial)
 974 {
 975         struct lu_env         *env;
 976         struct cl_io          *io;
 977         struct osc_object     *obj = ext->oe_obj;
 978         struct client_obd     *cli = osc_cli(obj);
 979         struct osc_async_page *oap;
 980         struct osc_async_page *tmp;
 981         struct pagevec        *pvec;
 982         int                    pages_in_chunk = 0;
 983         int                    ppc_bits    = cli->cl_chunkbits -
 984                                              PAGE_SHIFT;
 985         __u64                  trunc_chunk = trunc_index >> ppc_bits;
 986         int                    grants   = 0;
 987         int                    nr_pages = 0;
 988         int                    rc       = 0;
 989         __u16                  refcheck;
 990         ENTRY;
 991
 992         LASSERT(sanity_check(ext) == 0);
 993         LASSERT(ext->oe_state == OES_TRUNC);
 994         LASSERT(!ext->oe_urgent);
 995
 996         /* Request new lu_env.
 997          * We can't use that env from osc_cache_truncate_start() because
 998          * it's from lov_io_sub and not fully initialized. */
 999         env = cl_env_get(&refcheck);
1000         if (IS_ERR(env))
1001                 RETURN(PTR_ERR(env));
1002
1003         io  = osc_env_thread_io(env);
1004         io->ci_obj = cl_object_top(osc2cl(obj));
1005         io->ci_ignore_layout = 1;
1006         pvec = &osc_env_info(env)->oti_pagevec;
1007         ll_pagevec_init(pvec, 0);
1008         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
1009         if (rc < 0)
1010                 GOTO(out, rc);
1011
1012         /* discard all pages with index greater than trunc_index */
1013         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
1014                                      oap_pending_item) {
1015                 pgoff_t index = osc_index(oap2osc(oap));
1016                 struct cl_page  *page = oap2cl_page(oap);
1017
1018                 LASSERT(list_empty(&oap->oap_rpc_item));
1019
1020                 /* only discard the pages with their index greater than
1021                  * trunc_index, and ... */
1022                 if (index < trunc_index ||
1023                     (index == trunc_index && partial)) {
1024                         /* accounting how many pages remaining in the chunk
1025                          * so that we can calculate grants correctly. */
1026                         if (index >> ppc_bits == trunc_chunk)
1027                                 ++pages_in_chunk;
1028                         continue;
1029                 }
1030
1031                 list_del_init(&oap->oap_pending_item);
1032
1033                 cl_page_get(page);
1034                 lu_ref_add(&page->cp_reference, "truncate", current);
1035
1036                 if (cl_page_own(env, io, page) == 0) {
1037                         cl_page_discard(env, io, page);
1038                         cl_page_disown(env, io, page);
1039                 } else {
1040                         LASSERT(page->cp_state == CPS_FREEING);
1041                         LASSERT(0);
1042                 }
1043
1044                 lu_ref_del(&page->cp_reference, "truncate", current);
1045                 cl_pagevec_put(env, page, pvec);
1046
1047                 --ext->oe_nr_pages;
1048                 ++nr_pages;
1049         }
1050         pagevec_release(pvec);
1051
1052         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1053                       ext->oe_nr_pages == 0),
1054                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1055
1056         osc_object_lock(obj);
1057         if (ext->oe_nr_pages == 0) {
1058                 LASSERT(pages_in_chunk == 0);
1059                 grants = ext->oe_grants;
1060                 ext->oe_grants = 0;
1061         } else { /* calculate how many grants we can free */
1062                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1063                 pgoff_t last_index;
1064
1065
1066                 /* if there is no pages in this chunk, we can also free grants
1067                  * for the last chunk */
1068                 if (pages_in_chunk == 0) {
1069                         /* if this is the 1st chunk and no pages in this chunk,
1070                          * ext->oe_nr_pages must be zero, so we should be in
1071                          * the other if-clause. */
1072                         LASSERT(trunc_chunk > 0);
1073                         --trunc_chunk;
1074                         ++chunks;
1075                 }
1076
1077                 /* this is what we can free from this extent */
1078                 grants          = chunks << cli->cl_chunkbits;
1079                 ext->oe_grants -= grants;
1080                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1081                 ext->oe_end     = min(last_index, ext->oe_max_end);
1082                 LASSERT(ext->oe_end >= ext->oe_start);
1083                 LASSERT(ext->oe_grants > 0);
1084         }
1085         osc_object_unlock(obj);
1086
1087         if (grants > 0 || nr_pages > 0)
1088                 osc_free_grant(cli, nr_pages, grants, grants);
1089
1090 out:
1091         cl_io_fini(env, io);
1092         cl_env_put(env, &refcheck);
1093         RETURN(rc);
1094 }
1095
1096 /**
1097  * This function is used to make the extent prepared for transfer.
1098  * A race with flusing page - ll_writepage() has to be handled cautiously.
1099  */
1100 static int osc_extent_make_ready(const struct lu_env *env,
1101                                  struct osc_extent *ext)
1102 {
1103         struct osc_async_page *oap;
1104         struct osc_async_page *last = NULL;
1105         struct osc_object *obj = ext->oe_obj;
1106         unsigned int page_count = 0;
1107         int rc;
1108         ENTRY;
1109
1110         /* we're going to grab page lock, so object lock must not be taken. */
1111         LASSERT(sanity_check(ext) == 0);
1112         /* in locking state, any process should not touch this extent. */
1113         EASSERT(ext->oe_state == OES_LOCKING, ext);
1114         EASSERT(ext->oe_owner != NULL, ext);
1115
1116         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1117
1118         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1119                 ++page_count;
1120                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1121                         last = oap;
1122
1123                 /* checking ASYNC_READY is race safe */
1124                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1125                         continue;
1126
1127                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1128                 switch (rc) {
1129                 case 0:
1130                         spin_lock(&oap->oap_lock);
1131                         oap->oap_async_flags |= ASYNC_READY;
1132                         spin_unlock(&oap->oap_lock);
1133                         break;
1134                 case -EALREADY:
1135                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1136                         break;
1137                 default:
1138                         LASSERTF(0, "unknown return code: %d\n", rc);
1139                 }
1140         }
1141
1142         LASSERT(page_count == ext->oe_nr_pages);
1143         LASSERT(last != NULL);
1144         /* the last page is the only one we need to refresh its count by
1145          * the size of file. */
1146         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1147                 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1148                 LASSERT(last_oap_count > 0);
1149                 LASSERT(last->oap_page_off + last_oap_count <= PAGE_SIZE);
1150                 last->oap_count = last_oap_count;
1151                 spin_lock(&last->oap_lock);
1152                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1153                 spin_unlock(&last->oap_lock);
1154         }
1155
1156         /* for the rest of pages, we don't need to call osf_refresh_count()
1157          * because it's known they are not the last page */
1158         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1159                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1160                         oap->oap_count = PAGE_SIZE - oap->oap_page_off;
1161                         spin_lock(&oap->oap_lock);
1162                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1163                         spin_unlock(&oap->oap_lock);
1164                 }
1165         }
1166
1167         osc_object_lock(obj);
1168         osc_extent_state_set(ext, OES_RPC);
1169         osc_object_unlock(obj);
1170         /* get a refcount for RPC. */
1171         osc_extent_get(ext);
1172
1173         RETURN(0);
1174 }
1175
1176 /**
1177  * Quick and simple version of osc_extent_find(). This function is frequently
1178  * called to expand the extent for the same IO. To expand the extent, the
1179  * page index must be in the same or next chunk of ext->oe_end.
1180  */
1181 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index,
1182                              unsigned int *grants)
1183 {
1184         struct osc_object *obj = ext->oe_obj;
1185         struct client_obd *cli = osc_cli(obj);
1186         struct osc_extent *next;
1187         int ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1188         pgoff_t chunk = index >> ppc_bits;
1189         pgoff_t end_chunk;
1190         pgoff_t end_index;
1191         unsigned int chunksize = 1 << cli->cl_chunkbits;
1192         int rc = 0;
1193         ENTRY;
1194
1195         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1196         osc_object_lock(obj);
1197         LASSERT(sanity_check_nolock(ext) == 0);
1198         end_chunk = ext->oe_end >> ppc_bits;
1199         if (chunk > end_chunk + 1)
1200                 GOTO(out, rc = -ERANGE);
1201
1202         if (end_chunk >= chunk)
1203                 GOTO(out, rc = 0);
1204
1205         LASSERT(end_chunk + 1 == chunk);
1206
1207         /* try to expand this extent to cover @index */
1208         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1209
1210         /* don't go over the maximum extent size reported by server */
1211         if (end_index - ext->oe_start + 1 > cli->cl_max_extent_pages)
1212                 GOTO(out, rc = -ERANGE);
1213
1214         next = next_extent(ext);
1215         if (next != NULL && next->oe_start <= end_index)
1216                 /* complex mode - overlapped with the next extent,
1217                  * this case will be handled by osc_extent_find() */
1218                 GOTO(out, rc = -EAGAIN);
1219
1220         ext->oe_end = end_index;
1221         ext->oe_grants += chunksize;
1222         LASSERT(*grants >= chunksize);
1223         *grants -= chunksize;
1224         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1225                  "overlapped after expanding for %lu.\n", index);
1226         EXIT;
1227
1228 out:
1229         osc_object_unlock(obj);
1230         RETURN(rc);
1231 }
1232
1233 static void osc_extent_tree_dump0(int mask, struct osc_object *obj,
1234                                   const char *func, int line)
1235 {
1236         struct osc_extent *ext;
1237         int cnt;
1238
1239         if (!cfs_cdebug_show(mask, DEBUG_SUBSYSTEM))
1240                 return;
1241
1242         CDEBUG(mask, "Dump object %p extents at %s:%d, mppr: %u.\n",
1243                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1244
1245         /* osc_object_lock(obj); */
1246         cnt = 1;
1247         for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1248                 OSC_EXTENT_DUMP(mask, ext, "in tree %d.\n", cnt++);
1249
1250         cnt = 1;
1251         list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1252                 OSC_EXTENT_DUMP(mask, ext, "hp %d.\n", cnt++);
1253
1254         cnt = 1;
1255         list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1256                 OSC_EXTENT_DUMP(mask, ext, "urgent %d.\n", cnt++);
1257
1258         cnt = 1;
1259         list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1260                 OSC_EXTENT_DUMP(mask, ext, "reading %d.\n", cnt++);
1261         /* osc_object_unlock(obj); */
1262 }
1263
1264 /* ------------------ osc extent end ------------------ */
1265
1266 static inline int osc_is_ready(struct osc_object *osc)
1267 {
1268         return !list_empty(&osc->oo_ready_item) ||
1269                !list_empty(&osc->oo_hp_ready_item);
1270 }
1271
1272 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1273         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1274                (OSC), osc_is_ready(OSC),                                       \
1275                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1276                list_empty_marker(&(OSC)->oo_ready_item),                       \
1277                atomic_read(&(OSC)->oo_nr_writes),                              \
1278                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1279                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1280                atomic_read(&(OSC)->oo_nr_reads),                               \
1281                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1282                ##args)
1283
1284 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1285                           int cmd)
1286 {
1287         struct osc_page *opg  = oap2osc_page(oap);
1288         struct cl_page  *page = oap2cl_page(oap);
1289         int result;
1290
1291         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1292
1293         ENTRY;
1294         result = cl_page_make_ready(env, page, CRT_WRITE);
1295         if (result == 0)
1296                 opg->ops_submit_time = ktime_get();
1297         RETURN(result);
1298 }
1299
1300 static int osc_refresh_count(const struct lu_env *env,
1301                              struct osc_async_page *oap, int cmd)
1302 {
1303         struct osc_page  *opg = oap2osc_page(oap);
1304         pgoff_t index = osc_index(oap2osc(oap));
1305         struct cl_object *obj;
1306         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
1307         int result;
1308         loff_t kms;
1309
1310         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1311         LASSERT(!(cmd & OBD_BRW_READ));
1312         LASSERT(opg != NULL);
1313         obj = opg->ops_cl.cpl_obj;
1314
1315         cl_object_attr_lock(obj);
1316         result = cl_object_attr_get(env, obj, attr);
1317         cl_object_attr_unlock(obj);
1318         if (result < 0)
1319                 return result;
1320         kms = attr->cat_kms;
1321         if (cl_offset(obj, index) >= kms)
1322                 /* catch race with truncate */
1323                 return 0;
1324         else if (cl_offset(obj, index + 1) > kms)
1325                 /* catch sub-page write at end of file */
1326                 return kms & ~PAGE_MASK;
1327         else
1328                 return PAGE_SIZE;
1329 }
1330
1331 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1332                           int cmd, int rc)
1333 {
1334         struct osc_page   *opg  = oap2osc_page(oap);
1335         struct cl_page    *page = oap2cl_page(oap);
1336         enum cl_req_type   crt;
1337         int srvlock;
1338
1339         ENTRY;
1340
1341         cmd &= ~OBD_BRW_NOQUOTA;
1342         LASSERTF(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ),
1343                  "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1344         LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1345                 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1346         LASSERT(opg->ops_transfer_pinned);
1347
1348         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1349         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1350         opg->ops_transfer_pinned = 0;
1351
1352         opg->ops_submit_time = ktime_set(0, 0);
1353         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1354
1355         /* statistic */
1356         if (rc == 0 && srvlock) {
1357                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1358                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1359                 size_t bytes = oap->oap_count;
1360
1361                 if (crt == CRT_READ)
1362                         stats->os_lockless_reads += bytes;
1363                 else
1364                         stats->os_lockless_writes += bytes;
1365         }
1366
1367         /*
1368          * This has to be the last operation with the page, as locks are
1369          * released in cl_page_completion() and nothing except for the
1370          * reference counter protects page from concurrent reclaim.
1371          */
1372         lu_ref_del(&page->cp_reference, "transfer", page);
1373
1374         cl_page_completion(env, page, crt, rc);
1375         cl_page_put(env, page);
1376
1377         RETURN(0);
1378 }
1379
1380 #define OSC_DUMP_GRANT(mask, cli, fmt, args...) do {                    \
1381         struct client_obd *__tmp = (cli);                               \
1382         CDEBUG(mask, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu " \
1383                "dropped: %ld avail: %ld, dirty_grant: %ld, "            \
1384                "reserved: %ld, flight: %d } lru {in list: %ld, "        \
1385                "left: %ld, waiters: %d }" fmt "\n",                     \
1386                cli_name(__tmp),                                         \
1387                __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages,        \
1388                atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1389                __tmp->cl_lost_grant, __tmp->cl_avail_grant,             \
1390                __tmp->cl_dirty_grant,                                   \
1391                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,         \
1392                atomic_long_read(&__tmp->cl_lru_in_list),                \
1393                atomic_long_read(&__tmp->cl_lru_busy),                   \
1394                atomic_read(&__tmp->cl_lru_shrinkers), ##args);          \
1395 } while (0)
1396
1397 /* caller must hold loi_list_lock */
1398 static void osc_consume_write_grant(struct client_obd *cli,
1399                                     struct brw_page *pga)
1400 {
1401         assert_spin_locked(&cli->cl_loi_list_lock);
1402         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1403         cli->cl_dirty_pages++;
1404         pga->flag |= OBD_BRW_FROM_GRANT;
1405         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1406                PAGE_SIZE, pga, pga->pg);
1407         osc_update_next_shrink(cli);
1408 }
1409
1410 /* the companion to osc_consume_write_grant, called when a brw has completed.
1411  * must be called with the loi lock held. */
1412 static void osc_release_write_grant(struct client_obd *cli,
1413                                     struct brw_page *pga)
1414 {
1415         ENTRY;
1416
1417         assert_spin_locked(&cli->cl_loi_list_lock);
1418         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1419                 EXIT;
1420                 return;
1421         }
1422
1423         pga->flag &= ~OBD_BRW_FROM_GRANT;
1424         atomic_long_dec(&obd_dirty_pages);
1425         cli->cl_dirty_pages--;
1426         EXIT;
1427 }
1428
1429 /**
1430  * To avoid sleeping with object lock held, it's good for us allocate enough
1431  * grants before entering into critical section.
1432  *
1433  * client_obd_list_lock held by caller
1434  */
1435 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1436 {
1437         int rc = -EDQUOT;
1438
1439         if (cli->cl_avail_grant >= bytes) {
1440                 cli->cl_avail_grant    -= bytes;
1441                 cli->cl_reserved_grant += bytes;
1442                 rc = 0;
1443         }
1444         return rc;
1445 }
1446
1447 static void __osc_unreserve_grant(struct client_obd *cli,
1448                                   unsigned int reserved, unsigned int unused)
1449 {
1450         /* it's quite normal for us to get more grant than reserved.
1451          * Thinking about a case that two extents merged by adding a new
1452          * chunk, we can save one extent tax. If extent tax is greater than
1453          * one chunk, we can save more grant by adding a new chunk */
1454         cli->cl_reserved_grant -= reserved;
1455         if (unused > reserved) {
1456                 cli->cl_avail_grant += reserved;
1457                 cli->cl_lost_grant  += unused - reserved;
1458                 cli->cl_dirty_grant -= unused - reserved;
1459         } else {
1460                 cli->cl_avail_grant += unused;
1461                 cli->cl_dirty_grant += reserved - unused;
1462         }
1463 }
1464
1465 static void osc_unreserve_grant_nolock(struct client_obd *cli,
1466                                        unsigned int reserved,
1467                                        unsigned int unused)
1468 {
1469         __osc_unreserve_grant(cli, reserved, unused);
1470         if (unused > 0)
1471                 osc_wake_cache_waiters(cli);
1472 }
1473
1474 static void osc_unreserve_grant(struct client_obd *cli,
1475                                 unsigned int reserved, unsigned int unused)
1476 {
1477         spin_lock(&cli->cl_loi_list_lock);
1478         osc_unreserve_grant_nolock(cli, reserved, unused);
1479         spin_unlock(&cli->cl_loi_list_lock);
1480 }
1481
1482 /**
1483  * Free grant after IO is finished or canceled.
1484  *
1485  * @lost_grant is used to remember how many grants we have allocated but not
1486  * used, we should return these grants to OST. There're two cases where grants
1487  * can be lost:
1488  * 1. truncate;
1489  * 2. blocksize at OST is less than PAGE_SIZE and a partial page was
1490  *    written. In this case OST may use less chunks to serve this partial
1491  *    write. OSTs don't actually know the page size on the client side. so
1492  *    clients have to calculate lost grant by the blocksize on the OST.
1493  *    See filter_grant_check() for details.
1494  */
1495 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1496                            unsigned int lost_grant, unsigned int dirty_grant)
1497 {
1498         unsigned long grant;
1499
1500         grant = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
1501
1502         spin_lock(&cli->cl_loi_list_lock);
1503         atomic_long_sub(nr_pages, &obd_dirty_pages);
1504         cli->cl_dirty_pages -= nr_pages;
1505         cli->cl_lost_grant += lost_grant;
1506         cli->cl_dirty_grant -= dirty_grant;
1507         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1508                 /* borrow some grant from truncate to avoid the case that
1509                  * truncate uses up all avail grant */
1510                 cli->cl_lost_grant -= grant;
1511                 cli->cl_avail_grant += grant;
1512         }
1513         osc_wake_cache_waiters(cli);
1514         spin_unlock(&cli->cl_loi_list_lock);
1515         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu/%lu\n",
1516                lost_grant, cli->cl_lost_grant,
1517                cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_SHIFT,
1518                cli->cl_dirty_grant);
1519 }
1520
1521 /**
1522  * The companion to osc_enter_cache(), called when @oap is no longer part of
1523  * the dirty accounting due to error.
1524  */
1525 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1526 {
1527         spin_lock(&cli->cl_loi_list_lock);
1528         osc_release_write_grant(cli, &oap->oap_brw_page);
1529         spin_unlock(&cli->cl_loi_list_lock);
1530 }
1531
1532 /**
1533  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1534  * is available.
1535  */
1536 static int osc_enter_cache_try(struct client_obd *cli,
1537                                struct osc_async_page *oap,
1538                                int bytes)
1539 {
1540         int rc;
1541
1542         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1543
1544         rc = osc_reserve_grant(cli, bytes);
1545         if (rc < 0)
1546                 return 0;
1547
1548         if (cli->cl_dirty_pages < cli->cl_dirty_max_pages) {
1549                 if (atomic_long_add_return(1, &obd_dirty_pages) <=
1550                     obd_max_dirty_pages) {
1551                         osc_consume_write_grant(cli, &oap->oap_brw_page);
1552                         rc = 1;
1553                         goto out;
1554                 } else
1555                         atomic_long_dec(&obd_dirty_pages);
1556         }
1557         __osc_unreserve_grant(cli, bytes, bytes);
1558
1559 out:
1560         return rc;
1561 }
1562
1563 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1564 {
1565         int rc;
1566         spin_lock(&cli->cl_loi_list_lock);
1567         rc = list_empty(&ocw->ocw_entry);
1568         spin_unlock(&cli->cl_loi_list_lock);
1569         return rc;
1570 }
1571
1572 /**
1573  * The main entry to reserve dirty page accounting. Usually the grant reserved
1574  * in this function will be freed in bulk in osc_free_grant() unless it fails
1575  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1576  *
1577  * The process will be put into sleep if it's already run out of grant.
1578  */
1579 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1580                            struct osc_async_page *oap, int bytes)
1581 {
1582         struct osc_object       *osc = oap->oap_obj;
1583         struct lov_oinfo        *loi = osc->oo_oinfo;
1584         struct osc_cache_waiter  ocw;
1585         int                      rc = -EDQUOT;
1586         ENTRY;
1587
1588         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1589
1590         spin_lock(&cli->cl_loi_list_lock);
1591
1592         /* force the caller to try sync io.  this can jump the list
1593          * of queued writes and create a discontiguous rpc stream */
1594         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1595             cli->cl_dirty_max_pages == 0 ||
1596             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) {
1597                 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1598                 GOTO(out, rc = -EDQUOT);
1599         }
1600
1601         /* Hopefully normal case - cache space and write credits available */
1602         if (list_empty(&cli->cl_cache_waiters) &&
1603             osc_enter_cache_try(cli, oap, bytes)) {
1604                 OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1605                 GOTO(out, rc = 0);
1606         }
1607
1608         /* We can get here for two reasons: too many dirty pages in cache, or
1609          * run out of grants. In both cases we should write dirty pages out.
1610          * Adding a cache waiter will trigger urgent write-out no matter what
1611          * RPC size will be.
1612          * The exiting condition is no avail grants and no dirty pages caching,
1613          * that really means there is no space on the OST. */
1614         init_waitqueue_head(&ocw.ocw_waitq);
1615         ocw.ocw_oap   = oap;
1616         ocw.ocw_grant = bytes;
1617         while (cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0) {
1618                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1619                 ocw.ocw_rc = 0;
1620                 spin_unlock(&cli->cl_loi_list_lock);
1621
1622                 osc_io_unplug_async(env, cli, NULL);
1623
1624                 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1625                        cli_name(cli), &ocw, oap);
1626
1627                 rc = wait_event_idle_timeout(ocw.ocw_waitq,
1628                                              ocw_granted(cli, &ocw),
1629                                              cfs_time_seconds(AT_OFF ?
1630                                                               obd_timeout :
1631                                                               at_max));
1632
1633                 spin_lock(&cli->cl_loi_list_lock);
1634
1635                 if (rc <= 0) {
1636                         /* l_wait_event is interrupted by signal or timed out */
1637                         list_del_init(&ocw.ocw_entry);
1638                         if (rc == 0)
1639                                 rc = -ETIMEDOUT;
1640                         break;
1641                 }
1642                 LASSERT(list_empty(&ocw.ocw_entry));
1643                 rc = ocw.ocw_rc;
1644
1645                 if (rc != -EDQUOT)
1646                         break;
1647                 if (osc_enter_cache_try(cli, oap, bytes)) {
1648                         rc = 0;
1649                         break;
1650                 }
1651         }
1652
1653         switch (rc) {
1654         case 0:
1655                 OSC_DUMP_GRANT(D_CACHE, cli, "finally got grant space\n");
1656                 break;
1657         case -ETIMEDOUT:
1658                 OSC_DUMP_GRANT(D_CACHE, cli,
1659                                "timeout, fall back to sync i/o\n");
1660                 osc_extent_tree_dump(D_CACHE, osc);
1661                 /* fall back to synchronous I/O */
1662                 rc = -EDQUOT;
1663                 break;
1664         case -EINTR:
1665                 /* Ensures restartability - LU-3581 */
1666                 OSC_DUMP_GRANT(D_CACHE, cli, "interrupted\n");
1667                 rc = -ERESTARTSYS;
1668                 break;
1669         case -EDQUOT:
1670                 OSC_DUMP_GRANT(D_CACHE, cli,
1671                                "no grant space, fall back to sync i/o\n");
1672                 break;
1673         default:
1674                 CDEBUG(D_CACHE, "%s: event for cache space @ %p never arrived "
1675                        "due to %d, fall back to sync i/o\n",
1676                        cli_name(cli), &ocw, rc);
1677                 break;
1678         }
1679         EXIT;
1680 out:
1681         spin_unlock(&cli->cl_loi_list_lock);
1682         RETURN(rc);
1683 }
1684
1685 /* caller must hold loi_list_lock */
1686 void osc_wake_cache_waiters(struct client_obd *cli)
1687 {
1688         struct list_head *l, *tmp;
1689         struct osc_cache_waiter *ocw;
1690
1691         ENTRY;
1692         list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1693                 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
1694
1695                 ocw->ocw_rc = -EDQUOT;
1696
1697                 if (osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant))
1698                         ocw->ocw_rc = 0;
1699
1700                 if (ocw->ocw_rc == 0 ||
1701                     !(cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0)) {
1702                         list_del_init(&ocw->ocw_entry);
1703                         CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant "
1704                                "%ld, %d\n", ocw, ocw->ocw_oap,
1705                                cli->cl_avail_grant, ocw->ocw_rc);
1706
1707                         wake_up(&ocw->ocw_waitq);
1708                 }
1709         }
1710
1711         EXIT;
1712 }
1713 EXPORT_SYMBOL(osc_wake_cache_waiters);
1714
1715 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1716 {
1717         int hprpc = !!list_empty(&osc->oo_hp_exts);
1718         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1719 }
1720
1721 /* This maintains the lists of pending pages to read/write for a given object
1722  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1723  * to quickly find objects that are ready to send an RPC. */
1724 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1725                          int cmd)
1726 {
1727         int invalid_import = 0;
1728         ENTRY;
1729
1730         /* if we have an invalid import we want to drain the queued pages
1731          * by forcing them through rpcs that immediately fail and complete
1732          * the pages.  recovery relies on this to empty the queued pages
1733          * before canceling the locks and evicting down the llite pages */
1734         if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1735                 invalid_import = 1;
1736
1737         if (cmd & OBD_BRW_WRITE) {
1738                 if (atomic_read(&osc->oo_nr_writes) == 0)
1739                         RETURN(0);
1740                 if (invalid_import) {
1741                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1742                         RETURN(1);
1743                 }
1744                 if (!list_empty(&osc->oo_hp_exts)) {
1745                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1746                         RETURN(1);
1747                 }
1748                 if (!list_empty(&osc->oo_urgent_exts)) {
1749                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1750                         RETURN(1);
1751                 }
1752                 /* trigger a write rpc stream as long as there are dirtiers
1753                  * waiting for space.  as they're waiting, they're not going to
1754                  * create more pages to coalesce with what's waiting.. */
1755                 if (!list_empty(&cli->cl_cache_waiters)) {
1756                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1757                         RETURN(1);
1758                 }
1759                 if (!list_empty(&osc->oo_full_exts)) {
1760                         CDEBUG(D_CACHE, "full extent ready, make an RPC\n");
1761                         RETURN(1);
1762                 }
1763         } else {
1764                 if (atomic_read(&osc->oo_nr_reads) == 0)
1765                         RETURN(0);
1766                 if (invalid_import) {
1767                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1768                         RETURN(1);
1769                 }
1770                 /* all read are urgent. */
1771                 if (!list_empty(&osc->oo_reading_exts))
1772                         RETURN(1);
1773         }
1774
1775         RETURN(0);
1776 }
1777
1778 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1779 {
1780         struct client_obd *cli = osc_cli(obj);
1781         if (cmd & OBD_BRW_WRITE) {
1782                 atomic_add(delta, &obj->oo_nr_writes);
1783                 atomic_add(delta, &cli->cl_pending_w_pages);
1784                 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1785         } else {
1786                 atomic_add(delta, &obj->oo_nr_reads);
1787                 atomic_add(delta, &cli->cl_pending_r_pages);
1788                 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1789         }
1790         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1791 }
1792
1793 static int osc_makes_hprpc(struct osc_object *obj)
1794 {
1795         return !list_empty(&obj->oo_hp_exts);
1796 }
1797
1798 static void on_list(struct list_head *item, struct list_head *list,
1799                     int should_be_on)
1800 {
1801         if (list_empty(item) && should_be_on)
1802                 list_add_tail(item, list);
1803         else if (!list_empty(item) && !should_be_on)
1804                 list_del_init(item);
1805 }
1806
1807 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1808  * can find pages to build into rpcs quickly */
1809 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1810 {
1811         if (osc_makes_hprpc(osc)) {
1812                 /* HP rpc */
1813                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1814                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1815         } else {
1816                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1817                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1818                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1819                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1820         }
1821
1822         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1823                 atomic_read(&osc->oo_nr_writes) > 0);
1824
1825         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1826                 atomic_read(&osc->oo_nr_reads) > 0);
1827
1828         return osc_is_ready(osc);
1829 }
1830
1831 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1832 {
1833         int is_ready;
1834
1835         spin_lock(&cli->cl_loi_list_lock);
1836         is_ready = __osc_list_maint(cli, osc);
1837         spin_unlock(&cli->cl_loi_list_lock);
1838
1839         return is_ready;
1840 }
1841
1842 /* this is trying to propogate async writeback errors back up to the
1843  * application.  As an async write fails we record the error code for later if
1844  * the app does an fsync.  As long as errors persist we force future rpcs to be
1845  * sync so that the app can get a sync error and break the cycle of queueing
1846  * pages for which writeback will fail. */
1847 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1848                            int rc)
1849 {
1850         if (rc) {
1851                 if (!ar->ar_rc)
1852                         ar->ar_rc = rc;
1853
1854                 ar->ar_force_sync = 1;
1855                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1856                 return;
1857
1858         }
1859
1860         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1861                 ar->ar_force_sync = 0;
1862 }
1863
1864 /* this must be called holding the loi list lock to give coverage to exit_cache,
1865  * async_flag maintenance, and oap_request */
1866 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1867                               struct osc_async_page *oap, int sent, int rc)
1868 {
1869         struct osc_object *osc = oap->oap_obj;
1870         struct lov_oinfo  *loi = osc->oo_oinfo;
1871         __u64 xid = 0;
1872
1873         ENTRY;
1874         if (oap->oap_request != NULL) {
1875                 xid = ptlrpc_req_xid(oap->oap_request);
1876                 ptlrpc_req_finished(oap->oap_request);
1877                 oap->oap_request = NULL;
1878         }
1879
1880         /* As the transfer for this page is being done, clear the flags */
1881         spin_lock(&oap->oap_lock);
1882         oap->oap_async_flags = 0;
1883         spin_unlock(&oap->oap_lock);
1884         oap->oap_interrupted = 0;
1885
1886         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1887                 spin_lock(&cli->cl_loi_list_lock);
1888                 osc_process_ar(&cli->cl_ar, xid, rc);
1889                 osc_process_ar(&loi->loi_ar, xid, rc);
1890                 spin_unlock(&cli->cl_loi_list_lock);
1891         }
1892
1893         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1894         if (rc)
1895                 CERROR("completion on oap %p obj %p returns %d.\n",
1896                        oap, osc, rc);
1897
1898         EXIT;
1899 }
1900
1901 struct extent_rpc_data {
1902         struct list_head        *erd_rpc_list;
1903         unsigned int            erd_page_count;
1904         unsigned int            erd_max_pages;
1905         unsigned int            erd_max_chunks;
1906         unsigned int            erd_max_extents;
1907 };
1908
1909 static inline unsigned osc_extent_chunks(const struct osc_extent *ext)
1910 {
1911         struct client_obd *cli = osc_cli(ext->oe_obj);
1912         unsigned ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1913
1914         return (ext->oe_end >> ppc_bits) - (ext->oe_start >> ppc_bits) + 1;
1915 }
1916
1917 static inline bool
1918 can_merge(const struct osc_extent *ext, const struct osc_extent *in_rpc)
1919 {
1920         if (ext->oe_no_merge || in_rpc->oe_no_merge)
1921                 return false;
1922
1923         if (ext->oe_srvlock != in_rpc->oe_srvlock)
1924                 return false;
1925
1926         if (ext->oe_ndelay != in_rpc->oe_ndelay)
1927                 return false;
1928
1929         if (!ext->oe_grants != !in_rpc->oe_grants)
1930                 return false;
1931
1932         if (ext->oe_dio != in_rpc->oe_dio)
1933                 return false;
1934
1935         /* It's possible to have overlap on DIO */
1936         if (in_rpc->oe_dio && overlapped(ext, in_rpc))
1937                 return false;
1938
1939         return true;
1940 }
1941
1942 /**
1943  * Try to add extent to one RPC. We need to think about the following things:
1944  * - # of pages must not be over max_pages_per_rpc
1945  * - extent must be compatible with previous ones
1946  */
1947 static int try_to_add_extent_for_io(struct client_obd *cli,
1948                                     struct osc_extent *ext,
1949                                     struct extent_rpc_data *data)
1950 {
1951         struct osc_extent *tmp;
1952         unsigned int chunk_count;
1953         ENTRY;
1954
1955         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1956                 ext);
1957         OSC_EXTENT_DUMP(D_CACHE, ext, "trying to add this extent\n");
1958
1959         if (data->erd_max_extents == 0)
1960                 RETURN(0);
1961
1962         chunk_count = osc_extent_chunks(ext);
1963         EASSERTF(data->erd_page_count != 0 ||
1964                  chunk_count <= data->erd_max_chunks, ext,
1965                  "The first extent to be fit in a RPC contains %u chunks, "
1966                  "which is over the limit %u.\n", chunk_count,
1967                  data->erd_max_chunks);
1968         if (chunk_count > data->erd_max_chunks)
1969                 RETURN(0);
1970
1971         data->erd_max_pages = max(ext->oe_mppr, data->erd_max_pages);
1972         EASSERTF(data->erd_page_count != 0 ||
1973                 ext->oe_nr_pages <= data->erd_max_pages, ext,
1974                 "The first extent to be fit in a RPC contains %u pages, "
1975                 "which is over the limit %u.\n", ext->oe_nr_pages,
1976                 data->erd_max_pages);
1977         if (data->erd_page_count + ext->oe_nr_pages > data->erd_max_pages)
1978                 RETURN(0);
1979
1980         list_for_each_entry(tmp, data->erd_rpc_list, oe_link) {
1981                 EASSERT(tmp->oe_owner == current, tmp);
1982
1983                 if (!can_merge(ext, tmp))
1984                         RETURN(0);
1985         }
1986
1987         data->erd_max_extents--;
1988         data->erd_max_chunks -= chunk_count;
1989         data->erd_page_count += ext->oe_nr_pages;
1990         list_move_tail(&ext->oe_link, data->erd_rpc_list);
1991         ext->oe_owner = current;
1992         RETURN(1);
1993 }
1994
1995 /**
1996  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1997  * get_write_extent() takes all appropriate extents in atomic.
1998  *
1999  * The following policy is used to collect extents for IO:
2000  * 1. Add as many HP extents as possible;
2001  * 2. Add the first urgent extent in urgent extent list and take it out of
2002  *    urgent list;
2003  * 3. Add subsequent extents of this urgent extent;
2004  * 4. If urgent list is not empty, goto 2;
2005  * 5. Traverse the extent tree from the 1st extent;
2006  * 6. Above steps exit if there is no space in this RPC.
2007  */
2008 static unsigned int get_write_extents(struct osc_object *obj,
2009                                       struct list_head *rpclist)
2010 {
2011         struct client_obd *cli = osc_cli(obj);
2012         struct osc_extent *ext;
2013         struct extent_rpc_data data = {
2014                 .erd_rpc_list   = rpclist,
2015                 .erd_page_count = 0,
2016                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
2017                 .erd_max_chunks = osc_max_write_chunks(cli),
2018                 .erd_max_extents = 256,
2019         };
2020
2021         LASSERT(osc_object_is_locked(obj));
2022         while (!list_empty(&obj->oo_hp_exts)) {
2023                 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
2024                                  oe_link);
2025                 LASSERT(ext->oe_state == OES_CACHE);
2026                 if (!try_to_add_extent_for_io(cli, ext, &data))
2027                         return data.erd_page_count;
2028                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2029         }
2030         if (data.erd_page_count == data.erd_max_pages)
2031                 return data.erd_page_count;
2032
2033         while (!list_empty(&obj->oo_urgent_exts)) {
2034                 ext = list_entry(obj->oo_urgent_exts.next,
2035                                  struct osc_extent, oe_link);
2036                 if (!try_to_add_extent_for_io(cli, ext, &data))
2037                         return data.erd_page_count;
2038         }
2039         if (data.erd_page_count == data.erd_max_pages)
2040                 return data.erd_page_count;
2041
2042         /* One key difference between full extents and other extents: full
2043          * extents can usually only be added if the rpclist was empty, so if we
2044          * can't add one, we continue on to trying to add normal extents.  This
2045          * is so we don't miss adding extra extents to an RPC containing high
2046          * priority or urgent extents. */
2047         while (!list_empty(&obj->oo_full_exts)) {
2048                 ext = list_entry(obj->oo_full_exts.next,
2049                                  struct osc_extent, oe_link);
2050                 if (!try_to_add_extent_for_io(cli, ext, &data))
2051                         break;
2052         }
2053         if (data.erd_page_count == data.erd_max_pages)
2054                 return data.erd_page_count;
2055
2056         ext = first_extent(obj);
2057         while (ext != NULL) {
2058                 if ((ext->oe_state != OES_CACHE) ||
2059                     /* this extent may be already in current rpclist */
2060                     (!list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
2061                         ext = next_extent(ext);
2062                         continue;
2063                 }
2064
2065                 if (!try_to_add_extent_for_io(cli, ext, &data))
2066                         return data.erd_page_count;
2067
2068                 ext = next_extent(ext);
2069         }
2070         return data.erd_page_count;
2071 }
2072
2073 static int
2074 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
2075                    struct osc_object *osc)
2076 __must_hold(osc)
2077 {
2078         LIST_HEAD(rpclist);
2079         struct osc_extent *ext;
2080         struct osc_extent *tmp;
2081         struct osc_extent *first = NULL;
2082         unsigned int page_count = 0;
2083         int srvlock = 0;
2084         int rc = 0;
2085         ENTRY;
2086
2087         LASSERT(osc_object_is_locked(osc));
2088
2089         page_count = get_write_extents(osc, &rpclist);
2090         LASSERT(equi(page_count == 0, list_empty(&rpclist)));
2091
2092         if (list_empty(&rpclist))
2093                 RETURN(0);
2094
2095         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2096
2097         list_for_each_entry(ext, &rpclist, oe_link) {
2098                 LASSERT(ext->oe_state == OES_CACHE ||
2099                         ext->oe_state == OES_LOCK_DONE);
2100                 if (ext->oe_state == OES_CACHE)
2101                         osc_extent_state_set(ext, OES_LOCKING);
2102                 else
2103                         osc_extent_state_set(ext, OES_RPC);
2104         }
2105
2106         /* we're going to grab page lock, so release object lock because
2107          * lock order is page lock -> object lock. */
2108         osc_object_unlock(osc);
2109
2110         list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2111                 if (ext->oe_state == OES_LOCKING) {
2112                         rc = osc_extent_make_ready(env, ext);
2113                         if (unlikely(rc < 0)) {
2114                                 list_del_init(&ext->oe_link);
2115                                 osc_extent_finish(env, ext, 0, rc);
2116                                 continue;
2117                         }
2118                 }
2119                 if (first == NULL) {
2120                         first = ext;
2121                         srvlock = ext->oe_srvlock;
2122                 } else {
2123                         LASSERT(srvlock == ext->oe_srvlock);
2124                 }
2125         }
2126
2127         if (!list_empty(&rpclist)) {
2128                 LASSERT(page_count > 0);
2129                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2130                 LASSERT(list_empty(&rpclist));
2131         }
2132
2133         osc_object_lock(osc);
2134         RETURN(rc);
2135 }
2136
2137 /**
2138  * prepare pages for ASYNC io and put pages in send queue.
2139  *
2140  * \param cmd OBD_BRW_* macroses
2141  * \param lop pending pages
2142  *
2143  * \return zero if no page added to send queue.
2144  * \return 1 if pages successfully added to send queue.
2145  * \return negative on errors.
2146  */
2147 static int
2148 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2149                   struct osc_object *osc)
2150 __must_hold(osc)
2151 {
2152         struct osc_extent *ext;
2153         struct osc_extent *next;
2154         LIST_HEAD(rpclist);
2155         struct extent_rpc_data data = {
2156                 .erd_rpc_list   = &rpclist,
2157                 .erd_page_count = 0,
2158                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
2159                 .erd_max_chunks = UINT_MAX,
2160                 .erd_max_extents = UINT_MAX,
2161         };
2162         int rc = 0;
2163         ENTRY;
2164
2165         LASSERT(osc_object_is_locked(osc));
2166         list_for_each_entry_safe(ext, next, &osc->oo_reading_exts, oe_link) {
2167                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2168                 if (!try_to_add_extent_for_io(cli, ext, &data))
2169                         break;
2170                 osc_extent_state_set(ext, OES_RPC);
2171                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2172         }
2173         LASSERT(data.erd_page_count <= data.erd_max_pages);
2174
2175         osc_update_pending(osc, OBD_BRW_READ, -data.erd_page_count);
2176
2177         if (!list_empty(&rpclist)) {
2178                 osc_object_unlock(osc);
2179
2180                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2181                 LASSERT(list_empty(&rpclist));
2182
2183                 osc_object_lock(osc);
2184         }
2185         RETURN(rc);
2186 }
2187
2188 #define list_to_obj(list, item) ({                                            \
2189         struct list_head *__tmp = (list)->next;                               \
2190         list_del_init(__tmp);                                         \
2191         list_entry(__tmp, struct osc_object, oo_##item);                      \
2192 })
2193
2194 /* This is called by osc_check_rpcs() to find which objects have pages that
2195  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2196 static struct osc_object *osc_next_obj(struct client_obd *cli)
2197 {
2198         ENTRY;
2199
2200         /* First return objects that have blocked locks so that they
2201          * will be flushed quickly and other clients can get the lock,
2202          * then objects which have pages ready to be stuffed into RPCs */
2203         if (!list_empty(&cli->cl_loi_hp_ready_list))
2204                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2205         if (!list_empty(&cli->cl_loi_ready_list))
2206                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2207
2208         /* then if we have cache waiters, return all objects with queued
2209          * writes.  This is especially important when many small files
2210          * have filled up the cache and not been fired into rpcs because
2211          * they don't pass the nr_pending/object threshhold */
2212         if (!list_empty(&cli->cl_cache_waiters) &&
2213             !list_empty(&cli->cl_loi_write_list))
2214                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2215
2216         /* then return all queued objects when we have an invalid import
2217          * so that they get flushed */
2218         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2219                 if (!list_empty(&cli->cl_loi_write_list))
2220                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2221                                            write_item));
2222                 if (!list_empty(&cli->cl_loi_read_list))
2223                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2224                                            read_item));
2225         }
2226         RETURN(NULL);
2227 }
2228
2229 /* called with the loi list lock held */
2230 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2231 __must_hold(&cli->cl_loi_list_lock)
2232 {
2233         struct osc_object *osc;
2234         int rc = 0;
2235         ENTRY;
2236
2237         while ((osc = osc_next_obj(cli)) != NULL) {
2238                 struct cl_object *obj = osc2cl(osc);
2239                 struct lu_ref_link link;
2240
2241                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2242
2243                 if (osc_max_rpc_in_flight(cli, osc)) {
2244                         __osc_list_maint(cli, osc);
2245                         break;
2246                 }
2247
2248                 cl_object_get(obj);
2249                 spin_unlock(&cli->cl_loi_list_lock);
2250                 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2251
2252                 /* attempt some read/write balancing by alternating between
2253                  * reads and writes in an object.  The makes_rpc checks here
2254                  * would be redundant if we were getting read/write work items
2255                  * instead of objects.  we don't want send_oap_rpc to drain a
2256                  * partial read pending queue when we're given this object to
2257                  * do io on writes while there are cache waiters */
2258                 osc_object_lock(osc);
2259                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2260                         rc = osc_send_write_rpc(env, cli, osc);
2261                         if (rc < 0) {
2262                                 CERROR("Write request failed with %d\n", rc);
2263
2264                                 /* osc_send_write_rpc failed, mostly because of
2265                                  * memory pressure.
2266                                  *
2267                                  * It can't break here, because if:
2268                                  *  - a page was submitted by osc_io_submit, so
2269                                  *    page locked;
2270                                  *  - no request in flight
2271                                  *  - no subsequent request
2272                                  * The system will be in live-lock state,
2273                                  * because there is no chance to call
2274                                  * osc_io_unplug() and osc_check_rpcs() any
2275                                  * more. pdflush can't help in this case,
2276                                  * because it might be blocked at grabbing
2277                                  * the page lock as we mentioned.
2278                                  *
2279                                  * Anyway, continue to drain pages. */
2280                                 /* break; */
2281                         }
2282                 }
2283                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2284                         rc = osc_send_read_rpc(env, cli, osc);
2285                         if (rc < 0)
2286                                 CERROR("Read request failed with %d\n", rc);
2287                 }
2288                 osc_object_unlock(osc);
2289
2290                 osc_list_maint(cli, osc);
2291                 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2292                 cl_object_put(env, obj);
2293
2294                 spin_lock(&cli->cl_loi_list_lock);
2295         }
2296 }
2297
2298 int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2299                    struct osc_object *osc, int async)
2300 {
2301         int rc = 0;
2302
2303         if (osc != NULL && osc_list_maint(cli, osc) == 0)
2304                 return 0;
2305
2306         if (!async) {
2307                 spin_lock(&cli->cl_loi_list_lock);
2308                 osc_check_rpcs(env, cli);
2309                 spin_unlock(&cli->cl_loi_list_lock);
2310         } else {
2311                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2312                 LASSERT(cli->cl_writeback_work != NULL);
2313                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2314         }
2315         return rc;
2316 }
2317 EXPORT_SYMBOL(osc_io_unplug0);
2318
2319 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2320                         struct page *page, loff_t offset)
2321 {
2322         struct obd_export     *exp = osc_export(osc);
2323         struct osc_async_page *oap = &ops->ops_oap;
2324         ENTRY;
2325
2326         if (!page)
2327                 return cfs_size_round(sizeof(*oap));
2328
2329         oap->oap_magic = OAP_MAGIC;
2330         oap->oap_cli = &exp->exp_obd->u.cli;
2331         oap->oap_obj = osc;
2332
2333         oap->oap_page = page;
2334         oap->oap_obj_off = offset;
2335         LASSERT(!(offset & ~PAGE_MASK));
2336
2337         INIT_LIST_HEAD(&oap->oap_pending_item);
2338         INIT_LIST_HEAD(&oap->oap_rpc_item);
2339
2340         spin_lock_init(&oap->oap_lock);
2341         CDEBUG(D_INFO, "oap %p page %p obj off %llu\n",
2342                oap, page, oap->oap_obj_off);
2343         RETURN(0);
2344 }
2345 EXPORT_SYMBOL(osc_prep_async_page);
2346
2347 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2348                        struct osc_page *ops, cl_commit_cbt cb)
2349 {
2350         struct osc_io *oio = osc_env_io(env);
2351         struct osc_extent     *ext = NULL;
2352         struct osc_async_page *oap = &ops->ops_oap;
2353         struct client_obd     *cli = oap->oap_cli;
2354         struct osc_object     *osc = oap->oap_obj;
2355         struct pagevec        *pvec = &osc_env_info(env)->oti_pagevec;
2356         pgoff_t index;
2357         unsigned int tmp;
2358         unsigned int grants = 0;
2359         u32    brw_flags = OBD_BRW_ASYNC;
2360         int    cmd = OBD_BRW_WRITE;
2361         int    need_release = 0;
2362         int    rc = 0;
2363         ENTRY;
2364
2365         if (oap->oap_magic != OAP_MAGIC)
2366                 RETURN(-EINVAL);
2367
2368         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2369                 RETURN(-EIO);
2370
2371         if (!list_empty(&oap->oap_pending_item) ||
2372             !list_empty(&oap->oap_rpc_item))
2373                 RETURN(-EBUSY);
2374
2375         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2376         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2377         if (oio->oi_cap_sys_resource || io->ci_noquota) {
2378                 brw_flags |= OBD_BRW_NOQUOTA;
2379                 cmd |= OBD_BRW_NOQUOTA;
2380         }
2381
2382         /* check if the file's owner/group is over quota */
2383         if (!(cmd & OBD_BRW_NOQUOTA)) {
2384                 struct cl_object *obj;
2385                 struct cl_attr   *attr;
2386                 unsigned int qid[LL_MAXQUOTAS];
2387
2388                 obj = cl_object_top(&osc->oo_cl);
2389                 attr = &osc_env_info(env)->oti_attr;
2390
2391                 cl_object_attr_lock(obj);
2392                 rc = cl_object_attr_get(env, obj, attr);
2393                 cl_object_attr_unlock(obj);
2394
2395                 qid[USRQUOTA] = attr->cat_uid;
2396                 qid[GRPQUOTA] = attr->cat_gid;
2397                 qid[PRJQUOTA] = attr->cat_projid;
2398                 if (rc == 0 && osc_quota_chkdq(cli, qid) == -EDQUOT)
2399                         rc = -EDQUOT;
2400                 if (rc)
2401                         RETURN(rc);
2402         }
2403
2404         oap->oap_cmd = cmd;
2405         oap->oap_page_off = ops->ops_from;
2406         oap->oap_count = ops->ops_to - ops->ops_from;
2407         /* No need to hold a lock here,
2408          * since this page is not in any list yet. */
2409         oap->oap_async_flags = 0;
2410         oap->oap_brw_flags = brw_flags;
2411
2412         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2413                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2414
2415         index = osc_index(oap2osc(oap));
2416
2417         /* Add this page into extent by the following steps:
2418          * 1. if there exists an active extent for this IO, mostly this page
2419          *    can be added to the active extent and sometimes we need to
2420          *    expand extent to accomodate this page;
2421          * 2. otherwise, a new extent will be allocated. */
2422
2423         ext = oio->oi_active;
2424         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2425                 /* one chunk plus extent overhead must be enough to write this
2426                  * page */
2427                 grants = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2428                 if (ext->oe_end >= index)
2429                         grants = 0;
2430
2431                 /* it doesn't need any grant to dirty this page */
2432                 spin_lock(&cli->cl_loi_list_lock);
2433                 rc = osc_enter_cache_try(cli, oap, grants);
2434                 if (rc == 0) { /* try failed */
2435                         grants = 0;
2436                         need_release = 1;
2437                 } else if (ext->oe_end < index) {
2438                         tmp = grants;
2439                         /* try to expand this extent */
2440                         rc = osc_extent_expand(ext, index, &tmp);
2441                         if (rc < 0) {
2442                                 need_release = 1;
2443                                 /* don't free reserved grant */
2444                         } else {
2445                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2446                                                 "expanded for %lu.\n", index);
2447                                 osc_unreserve_grant_nolock(cli, grants, tmp);
2448                                 grants = 0;
2449                         }
2450                 }
2451                 spin_unlock(&cli->cl_loi_list_lock);
2452                 rc = 0;
2453         } else if (ext != NULL) {
2454                 /* index is located outside of active extent */
2455                 need_release = 1;
2456         }
2457         if (need_release) {
2458                 osc_extent_release(env, ext);
2459                 oio->oi_active = NULL;
2460                 ext = NULL;
2461         }
2462
2463         if (ext == NULL) {
2464                 tmp = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2465
2466                 /* try to find new extent to cover this page */
2467                 LASSERT(oio->oi_active == NULL);
2468                 /* we may have allocated grant for this page if we failed
2469                  * to expand the previous active extent. */
2470                 LASSERT(ergo(grants > 0, grants >= tmp));
2471
2472                 rc = 0;
2473                 if (grants == 0) {
2474                         /* We haven't allocated grant for this page, and we
2475                          * must not hold a page lock while we do enter_cache,
2476                          * so we must mark dirty & unlock any pages in the
2477                          * write commit pagevec. */
2478                         if (pagevec_count(pvec)) {
2479                                 cb(env, io, pvec);
2480                                 pagevec_reinit(pvec);
2481                         }
2482                         rc = osc_enter_cache(env, cli, oap, tmp);
2483                         if (rc == 0)
2484                                 grants = tmp;
2485                 }
2486
2487                 tmp = grants;
2488                 if (rc == 0) {
2489                         ext = osc_extent_find(env, osc, index, &tmp);
2490                         if (IS_ERR(ext)) {
2491                                 LASSERT(tmp == grants);
2492                                 osc_exit_cache(cli, oap);
2493                                 rc = PTR_ERR(ext);
2494                                 ext = NULL;
2495                         } else {
2496                                 oio->oi_active = ext;
2497                         }
2498                 }
2499                 if (grants > 0)
2500                         osc_unreserve_grant(cli, grants, tmp);
2501         }
2502
2503         LASSERT(ergo(rc == 0, ext != NULL));
2504         if (ext != NULL) {
2505                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2506                          ext, "index = %lu.\n", index);
2507                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2508
2509                 osc_object_lock(osc);
2510                 if (ext->oe_nr_pages == 0)
2511                         ext->oe_srvlock = ops->ops_srvlock;
2512                 else
2513                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2514                 ++ext->oe_nr_pages;
2515                 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2516                 osc_object_unlock(osc);
2517
2518                 if (!ext->oe_layout_version)
2519                         ext->oe_layout_version = io->ci_layout_version;
2520         }
2521
2522         RETURN(rc);
2523 }
2524
2525 int osc_teardown_async_page(const struct lu_env *env,
2526                             struct osc_object *obj, struct osc_page *ops)
2527 {
2528         struct osc_async_page *oap = &ops->ops_oap;
2529         int rc = 0;
2530         ENTRY;
2531
2532         LASSERT(oap->oap_magic == OAP_MAGIC);
2533
2534         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2535                oap, ops, osc_index(oap2osc(oap)));
2536
2537         if (!list_empty(&oap->oap_rpc_item)) {
2538                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2539                 rc = -EBUSY;
2540         } else if (!list_empty(&oap->oap_pending_item)) {
2541                 struct osc_extent *ext = NULL;
2542
2543                 osc_object_lock(obj);
2544                 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2545                 osc_object_unlock(obj);
2546                 /* only truncated pages are allowed to be taken out.
2547                  * See osc_extent_truncate() and osc_cache_truncate_start()
2548                  * for details. */
2549                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2550                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2551                                         osc_index(oap2osc(oap)));
2552                         rc = -EBUSY;
2553                 }
2554                 if (ext != NULL)
2555                         osc_extent_put(env, ext);
2556         }
2557         RETURN(rc);
2558 }
2559
2560 /**
2561  * This is called when a page is picked up by kernel to write out.
2562  *
2563  * We should find out the corresponding extent and add the whole extent
2564  * into urgent list. The extent may be being truncated or used, handle it
2565  * carefully.
2566  */
2567 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2568                          struct osc_page *ops)
2569 {
2570         struct osc_extent *ext   = NULL;
2571         struct osc_object *obj   = cl2osc(ops->ops_cl.cpl_obj);
2572         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2573         pgoff_t            index = osc_index(ops);
2574         struct osc_async_page *oap = &ops->ops_oap;
2575         bool unplug = false;
2576         int rc = 0;
2577         ENTRY;
2578
2579         osc_object_lock(obj);
2580         ext = osc_extent_lookup(obj, index);
2581         if (ext == NULL) {
2582                 osc_extent_tree_dump(D_ERROR, obj);
2583                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2584         }
2585
2586         switch (ext->oe_state) {
2587         case OES_RPC:
2588         case OES_LOCK_DONE:
2589                 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2590                 LASSERT(0);
2591                 break;
2592         case OES_LOCKING:
2593                 /* If we know this extent is being written out, we should abort
2594                  * so that the writer can make this page ready. Otherwise, there
2595                  * exists a deadlock problem because other process can wait for
2596                  * page writeback bit holding page lock; and meanwhile in
2597                  * vvp_page_make_ready(), we need to grab page lock before
2598                  * really sending the RPC. */
2599         case OES_TRUNC:
2600                 /* race with truncate, page will be redirtied */
2601         case OES_ACTIVE:
2602                 /* The extent is active so we need to abort and let the caller
2603                  * re-dirty the page. If we continued on here, and we were the
2604                  * one making the extent active, we could deadlock waiting for
2605                  * the page writeback to clear but it won't because the extent
2606                  * is active and won't be written out. */
2607                 GOTO(out, rc = -EAGAIN);
2608         default:
2609                 break;
2610         }
2611
2612         rc = cl_page_prep(env, io, cp, CRT_WRITE);
2613         if (rc)
2614                 GOTO(out, rc);
2615
2616         spin_lock(&oap->oap_lock);
2617         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2618         spin_unlock(&oap->oap_lock);
2619
2620         if (memory_pressure_get())
2621                 ext->oe_memalloc = 1;
2622
2623         ext->oe_urgent = 1;
2624         if (ext->oe_state == OES_CACHE) {
2625                 OSC_EXTENT_DUMP(D_CACHE, ext,
2626                                 "flush page %p make it urgent.\n", oap);
2627                 if (list_empty(&ext->oe_link))
2628                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2629                 unplug = true;
2630         }
2631         rc = 0;
2632         EXIT;
2633
2634 out:
2635         osc_object_unlock(obj);
2636         osc_extent_put(env, ext);
2637         if (unplug)
2638                 osc_io_unplug_async(env, osc_cli(obj), obj);
2639         return rc;
2640 }
2641
2642 /**
2643  * this is called when a sync waiter receives an interruption.  Its job is to
2644  * get the caller woken as soon as possible.  If its page hasn't been put in an
2645  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2646  * desiring interruption which will forcefully complete the rpc once the rpc
2647  * has timed out.
2648  */
2649 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2650 {
2651         struct osc_async_page *oap = &ops->ops_oap;
2652         struct osc_object     *obj = oap->oap_obj;
2653         struct client_obd     *cli = osc_cli(obj);
2654         struct osc_extent     *ext;
2655         struct osc_extent     *found = NULL;
2656         struct list_head            *plist;
2657         pgoff_t index = osc_index(ops);
2658         int     rc = -EBUSY;
2659         int     cmd;
2660         ENTRY;
2661
2662         LASSERT(!oap->oap_interrupted);
2663         oap->oap_interrupted = 1;
2664
2665         /* Find out the caching extent */
2666         osc_object_lock(obj);
2667         if (oap->oap_cmd & OBD_BRW_WRITE) {
2668                 plist = &obj->oo_urgent_exts;
2669                 cmd   = OBD_BRW_WRITE;
2670         } else {
2671                 plist = &obj->oo_reading_exts;
2672                 cmd   = OBD_BRW_READ;
2673         }
2674         list_for_each_entry(ext, plist, oe_link) {
2675                 if (ext->oe_start <= index && ext->oe_end >= index) {
2676                         LASSERT(ext->oe_state == OES_LOCK_DONE);
2677                         /* For OES_LOCK_DONE state extent, it has already held
2678                          * a refcount for RPC. */
2679                         found = osc_extent_get(ext);
2680                         break;
2681                 }
2682         }
2683         if (found != NULL) {
2684                 list_del_init(&found->oe_link);
2685                 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2686                 osc_object_unlock(obj);
2687
2688                 osc_extent_finish(env, found, 0, -EINTR);
2689                 osc_extent_put(env, found);
2690                 rc = 0;
2691         } else {
2692                 osc_object_unlock(obj);
2693                 /* ok, it's been put in an rpc. only one oap gets a request
2694                  * reference */
2695                 if (oap->oap_request != NULL) {
2696                         ptlrpc_mark_interrupted(oap->oap_request);
2697                         ptlrpcd_wake(oap->oap_request);
2698                         ptlrpc_req_finished(oap->oap_request);
2699                         oap->oap_request = NULL;
2700                 }
2701         }
2702
2703         osc_list_maint(cli, obj);
2704         RETURN(rc);
2705 }
2706
2707 int osc_queue_sync_pages(const struct lu_env *env, const struct cl_io *io,
2708                          struct osc_object *obj, struct list_head *list,
2709                          int brw_flags)
2710 {
2711         struct client_obd     *cli = osc_cli(obj);
2712         struct osc_extent     *ext;
2713         struct osc_async_page *oap;
2714         int     page_count = 0;
2715         int     mppr       = cli->cl_max_pages_per_rpc;
2716         bool    can_merge   = true;
2717         pgoff_t start      = CL_PAGE_EOF;
2718         pgoff_t end        = 0;
2719         ENTRY;
2720
2721         list_for_each_entry(oap, list, oap_pending_item) {
2722                 struct osc_page *opg = oap2osc_page(oap);
2723                 pgoff_t index = osc_index(opg);
2724
2725                 if (index > end)
2726                         end = index;
2727                 if (index < start)
2728                         start = index;
2729                 ++page_count;
2730                 mppr <<= (page_count > mppr);
2731
2732                 if (unlikely(opg->ops_from > 0 || opg->ops_to < PAGE_SIZE))
2733                         can_merge = false;
2734         }
2735
2736         ext = osc_extent_alloc(obj);
2737         if (ext == NULL) {
2738                 struct osc_async_page *tmp;
2739
2740                 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2741                         list_del_init(&oap->oap_pending_item);
2742                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2743                 }
2744                 RETURN(-ENOMEM);
2745         }
2746
2747         ext->oe_rw = !!(brw_flags & OBD_BRW_READ);
2748         ext->oe_sync = 1;
2749         ext->oe_no_merge = !can_merge;
2750         ext->oe_urgent = 1;
2751         ext->oe_start = start;
2752         ext->oe_end = ext->oe_max_end = end;
2753         ext->oe_obj = obj;
2754         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2755         ext->oe_ndelay = !!(brw_flags & OBD_BRW_NDELAY);
2756         ext->oe_dio = !!(brw_flags & OBD_BRW_NOCACHE);
2757         ext->oe_nr_pages = page_count;
2758         ext->oe_mppr = mppr;
2759         list_splice_init(list, &ext->oe_pages);
2760         ext->oe_layout_version = io->ci_layout_version;
2761
2762         osc_object_lock(obj);
2763         /* Reuse the initial refcount for RPC, don't drop it */
2764         osc_extent_state_set(ext, OES_LOCK_DONE);
2765         if (!ext->oe_rw) { /* write */
2766                 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2767                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2768         } else {
2769                 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2770                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2771         }
2772         osc_object_unlock(obj);
2773
2774         osc_io_unplug_async(env, cli, obj);
2775         RETURN(0);
2776 }
2777
2778 /**
2779  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2780  */
2781 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
2782                              __u64 size, struct osc_extent **extp)
2783 {
2784         struct client_obd *cli = osc_cli(obj);
2785         struct osc_extent *ext;
2786         struct osc_extent *waiting = NULL;
2787         pgoff_t index;
2788         LIST_HEAD(list);
2789         int result = 0;
2790         bool partial;
2791         ENTRY;
2792
2793         /* pages with index greater or equal to index will be truncated. */
2794         index = cl_index(osc2cl(obj), size);
2795         partial = size > cl_offset(osc2cl(obj), index);
2796
2797 again:
2798         osc_object_lock(obj);
2799         ext = osc_extent_search(obj, index);
2800         if (ext == NULL)
2801                 ext = first_extent(obj);
2802         else if (ext->oe_end < index)
2803                 ext = next_extent(ext);
2804         while (ext != NULL) {
2805                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2806
2807                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2808                         /* if ext is in urgent state, it means there must exist
2809                          * a page already having been flushed by write_page().
2810                          * We have to wait for this extent because we can't
2811                          * truncate that page. */
2812                         OSC_EXTENT_DUMP(D_CACHE, ext,
2813                                         "waiting for busy extent\n");
2814                         waiting = osc_extent_get(ext);
2815                         break;
2816                 }
2817
2818                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:%llu.\n", size);
2819
2820                 osc_extent_get(ext);
2821                 if (ext->oe_state == OES_ACTIVE) {
2822                         /* though we grab inode mutex for write path, but we
2823                          * release it before releasing extent(in osc_io_end()),
2824                          * so there is a race window that an extent is still
2825                          * in OES_ACTIVE when truncate starts. */
2826                         LASSERT(!ext->oe_trunc_pending);
2827                         ext->oe_trunc_pending = 1;
2828                 } else {
2829                         EASSERT(ext->oe_state == OES_CACHE, ext);
2830                         osc_extent_state_set(ext, OES_TRUNC);
2831                         osc_update_pending(obj, OBD_BRW_WRITE,
2832                                            -ext->oe_nr_pages);
2833                 }
2834                 /* This extent could be on the full extents list, that's OK */
2835                 EASSERT(!ext->oe_hp && !ext->oe_urgent, ext);
2836                 if (!list_empty(&ext->oe_link))
2837                         list_move_tail(&ext->oe_link, &list);
2838                 else
2839                         list_add_tail(&ext->oe_link, &list);
2840
2841                 ext = next_extent(ext);
2842         }
2843         osc_object_unlock(obj);
2844
2845         osc_list_maint(cli, obj);
2846
2847         while (!list_empty(&list)) {
2848                 int rc;
2849
2850                 ext = list_entry(list.next, struct osc_extent, oe_link);
2851                 list_del_init(&ext->oe_link);
2852
2853                 /* extent may be in OES_ACTIVE state because inode mutex
2854                  * is released before osc_io_end() in file write case */
2855                 if (ext->oe_state != OES_TRUNC)
2856                         osc_extent_wait(env, ext, OES_TRUNC);
2857
2858                 rc = osc_extent_truncate(ext, index, partial);
2859                 if (rc < 0) {
2860                         if (result == 0)
2861                                 result = rc;
2862
2863                         OSC_EXTENT_DUMP(D_ERROR, ext,
2864                                         "truncate error %d\n", rc);
2865                 } else if (ext->oe_nr_pages == 0) {
2866                         osc_extent_remove(ext);
2867                 } else {
2868                         /* this must be an overlapped extent which means only
2869                          * part of pages in this extent have been truncated.
2870                          */
2871                         EASSERTF(ext->oe_start <= index, ext,
2872                                  "trunc index = %lu/%d.\n", index, partial);
2873                         /* fix index to skip this partially truncated extent */
2874                         index = ext->oe_end + 1;
2875                         partial = false;
2876
2877                         /* we need to hold this extent in OES_TRUNC state so
2878                          * that no writeback will happen. This is to avoid
2879                          * BUG 17397.
2880                          * Only partial truncate can reach here, if @size is
2881                          * not zero, the caller should provide a valid @extp. */
2882                         LASSERT(*extp == NULL);
2883                         *extp = osc_extent_get(ext);
2884                         OSC_EXTENT_DUMP(D_CACHE, ext,
2885                                         "trunc at %llu\n", size);
2886                 }
2887                 osc_extent_put(env, ext);
2888         }
2889         if (waiting != NULL) {
2890                 int rc;
2891
2892                 /* ignore the result of osc_extent_wait the write initiator
2893                  * should take care of it. */
2894                 rc = osc_extent_wait(env, waiting, OES_INV);
2895                 if (rc < 0)
2896                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2897
2898                 osc_extent_put(env, waiting);
2899                 waiting = NULL;
2900                 goto again;
2901         }
2902         RETURN(result);
2903 }
2904 EXPORT_SYMBOL(osc_cache_truncate_start);
2905
2906 /**
2907  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2908  */
2909 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext)
2910 {
2911         if (ext != NULL) {
2912                 struct osc_object *obj = ext->oe_obj;
2913                 bool unplug = false;
2914
2915                 EASSERT(ext->oe_nr_pages > 0, ext);
2916                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2917                 EASSERT(!ext->oe_urgent, ext);
2918
2919                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2920                 osc_object_lock(obj);
2921                 osc_extent_state_set(ext, OES_CACHE);
2922                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2923                         ext->oe_urgent = 1;
2924                         list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2925                         unplug = true;
2926                 }
2927                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2928                 osc_object_unlock(obj);
2929                 osc_extent_put(env, ext);
2930
2931                 if (unplug)
2932                         osc_io_unplug_async(env, osc_cli(obj), obj);
2933         }
2934 }
2935
2936 /**
2937  * Wait for extents in a specific range to be written out.
2938  * The caller must have called osc_cache_writeback_range() to issue IO
2939  * otherwise it will take a long time for this function to finish.
2940  *
2941  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2942  * nobody else can dirty this range of file while we're waiting for
2943  * extents to be written.
2944  */
2945 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2946                          pgoff_t start, pgoff_t end)
2947 {
2948         struct osc_extent *ext;
2949         pgoff_t index = start;
2950         int     result = 0;
2951         ENTRY;
2952
2953 again:
2954         osc_object_lock(obj);
2955         ext = osc_extent_search(obj, index);
2956         if (ext == NULL)
2957                 ext = first_extent(obj);
2958         else if (ext->oe_end < index)
2959                 ext = next_extent(ext);
2960         while (ext != NULL) {
2961                 int rc;
2962
2963                 if (ext->oe_start > end)
2964                         break;
2965
2966                 if (!ext->oe_fsync_wait) {
2967                         ext = next_extent(ext);
2968                         continue;
2969                 }
2970
2971                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2972                              ext->oe_hp || ext->oe_urgent), ext);
2973                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2974                              !ext->oe_hp && ext->oe_urgent), ext);
2975
2976                 index = ext->oe_end + 1;
2977                 osc_extent_get(ext);
2978                 osc_object_unlock(obj);
2979
2980                 rc = osc_extent_wait(env, ext, OES_INV);
2981                 if (result == 0)
2982                         result = rc;
2983                 osc_extent_put(env, ext);
2984                 goto again;
2985         }
2986         osc_object_unlock(obj);
2987
2988         OSC_IO_DEBUG(obj, "sync file range.\n");
2989         RETURN(result);
2990 }
2991 EXPORT_SYMBOL(osc_cache_wait_range);
2992
2993 /**
2994  * Called to write out a range of osc object.
2995  *
2996  * @hp     : should be set this is caused by lock cancel;
2997  * @discard: is set if dirty pages should be dropped - file will be deleted or
2998  *         truncated, this implies there is no partially discarding extents.
2999  *
3000  * Return how many pages will be issued, or error code if error occurred.
3001  */
3002 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
3003                               pgoff_t start, pgoff_t end, int hp, int discard)
3004 {
3005         struct osc_extent *ext;
3006         LIST_HEAD(discard_list);
3007         bool unplug = false;
3008         int result = 0;
3009         ENTRY;
3010
3011         osc_object_lock(obj);
3012         ext = osc_extent_search(obj, start);
3013         if (ext == NULL)
3014                 ext = first_extent(obj);
3015         else if (ext->oe_end < start)
3016                 ext = next_extent(ext);
3017         while (ext != NULL) {
3018                 if (ext->oe_start > end)
3019                         break;
3020
3021                 ext->oe_fsync_wait = 1;
3022                 switch (ext->oe_state) {
3023                 case OES_CACHE:
3024                         result += ext->oe_nr_pages;
3025                         if (!discard) {
3026                                 struct list_head *list = NULL;
3027                                 if (hp) {
3028                                         EASSERT(!ext->oe_hp, ext);
3029                                         ext->oe_hp = 1;
3030                                         list = &obj->oo_hp_exts;
3031                                 } else if (!ext->oe_urgent) {
3032                                         ext->oe_urgent = 1;
3033                                         list = &obj->oo_urgent_exts;
3034                                 }
3035                                 if (list != NULL)
3036                                         list_move_tail(&ext->oe_link, list);
3037                                 unplug = true;
3038                         } else {
3039                                 struct client_obd *cli = osc_cli(obj);
3040                                 int pcc_bits = cli->cl_chunkbits - PAGE_SHIFT;
3041                                 pgoff_t align_by = (1 << pcc_bits);
3042                                 pgoff_t a_start = round_down(start, align_by);
3043                                 pgoff_t a_end = round_up(end, align_by);
3044
3045                                 /* overflow case */
3046                                 if (end && !a_end)
3047                                         a_end = CL_PAGE_EOF;
3048                                 /* the only discarder is lock cancelling, so
3049                                  * [start, end], aligned by chunk size, must
3050                                  * contain this extent */
3051                                 LASSERTF(ext->oe_start >= a_start &&
3052                                          ext->oe_end <= a_end,
3053                                          "ext [%lu, %lu] reg [%lu, %lu] "
3054                                          "orig [%lu %lu] align %lu bits "
3055                                          "%d\n", ext->oe_start, ext->oe_end,
3056                                          a_start, a_end, start, end,
3057                                          align_by, pcc_bits);
3058                                 osc_extent_state_set(ext, OES_LOCKING);
3059                                 ext->oe_owner = current;
3060                                 list_move_tail(&ext->oe_link,
3061                                                    &discard_list);
3062                                 osc_update_pending(obj, OBD_BRW_WRITE,
3063                                                    -ext->oe_nr_pages);
3064                         }
3065                         break;
3066                 case OES_ACTIVE:
3067                         /* It's pretty bad to wait for ACTIVE extents, because
3068                          * we don't know how long we will wait for it to be
3069                          * flushed since it may be blocked at awaiting more
3070                          * grants. We do this for the correctness of fsync. */
3071                         LASSERT(hp == 0 && discard == 0);
3072                         ext->oe_urgent = 1;
3073                         break;
3074                 case OES_TRUNC:
3075                         /* this extent is being truncated, can't do anything
3076                          * for it now. it will be set to urgent after truncate
3077                          * is finished in osc_cache_truncate_end(). */
3078                 default:
3079                         break;
3080                 }
3081                 ext = next_extent(ext);
3082         }
3083         osc_object_unlock(obj);
3084
3085         LASSERT(ergo(!discard, list_empty(&discard_list)));
3086         if (!list_empty(&discard_list)) {
3087                 struct osc_extent *tmp;
3088                 int rc;
3089
3090                 osc_list_maint(osc_cli(obj), obj);
3091                 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
3092                         list_del_init(&ext->oe_link);
3093                         EASSERT(ext->oe_state == OES_LOCKING, ext);
3094
3095                         /* Discard caching pages. We don't actually write this
3096                          * extent out but we complete it as if we did. */
3097                         rc = osc_extent_make_ready(env, ext);
3098                         if (unlikely(rc < 0)) {
3099                                 OSC_EXTENT_DUMP(D_ERROR, ext,
3100                                                 "make_ready returned %d\n", rc);
3101                                 if (result >= 0)
3102                                         result = rc;
3103                         }
3104
3105                         /* finish the extent as if the pages were sent */
3106                         osc_extent_finish(env, ext, 0, 0);
3107                 }
3108         }
3109
3110         if (unplug)
3111                 osc_io_unplug(env, osc_cli(obj), obj);
3112
3113         if (hp || discard) {
3114                 int rc;
3115                 rc = osc_cache_wait_range(env, obj, start, end);
3116                 if (result >= 0 && rc < 0)
3117                         result = rc;
3118         }
3119
3120         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3121         RETURN(result);
3122 }
3123 EXPORT_SYMBOL(osc_cache_writeback_range);
3124
3125 /**
3126  * Returns a list of pages by a given [start, end] of \a obj.
3127  *
3128  * \param resched If not NULL, then we give up before hogging CPU for too
3129  * long and set *resched = 1, in that case caller should implement a retry
3130  * logic.
3131  *
3132  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3133  * crucial in the face of [offset, EOF] locks.
3134  *
3135  * Return at least one page in @queue unless there is no covered page.
3136  */
3137 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3138                         struct osc_object *osc, pgoff_t start, pgoff_t end,
3139                         osc_page_gang_cbt cb, void *cbdata)
3140 {
3141         struct osc_page *ops;
3142         struct pagevec  *pagevec;
3143         void            **pvec;
3144         pgoff_t         idx;
3145         unsigned int    nr;
3146         unsigned int    i;
3147         unsigned int    j;
3148         int             res = CLP_GANG_OKAY;
3149         bool            tree_lock = true;
3150         ENTRY;
3151
3152         idx = start;
3153         pvec = osc_env_info(env)->oti_pvec;
3154         pagevec = &osc_env_info(env)->oti_pagevec;
3155         ll_pagevec_init(pagevec, 0);
3156         spin_lock(&osc->oo_tree_lock);
3157         while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3158                                             idx, OTI_PVEC_SIZE)) > 0) {
3159                 struct cl_page *page;
3160                 bool end_of_region = false;
3161
3162                 for (i = 0, j = 0; i < nr; ++i) {
3163                         ops = pvec[i];
3164                         pvec[i] = NULL;
3165
3166                         idx = osc_index(ops);
3167                         if (idx > end) {
3168                                 end_of_region = true;
3169                                 break;
3170                         }
3171
3172                         page = ops->ops_cl.cpl_page;
3173                         LASSERT(page->cp_type == CPT_CACHEABLE);
3174                         if (page->cp_state == CPS_FREEING)
3175                                 continue;
3176
3177                         cl_page_get(page);
3178                         lu_ref_add_atomic(&page->cp_reference,
3179                                           "gang_lookup", current);
3180                         pvec[j++] = ops;
3181                 }
3182                 ++idx;
3183
3184                 /*
3185                  * Here a delicate locking dance is performed. Current thread
3186                  * holds a reference to a page, but has to own it before it
3187                  * can be placed into queue. Owning implies waiting, so
3188                  * radix-tree lock is to be released. After a wait one has to
3189                  * check that pages weren't truncated (cl_page_own() returns
3190                  * error in the latter case).
3191                  */
3192                 spin_unlock(&osc->oo_tree_lock);
3193                 tree_lock = false;
3194
3195                 for (i = 0; i < j; ++i) {
3196                         ops = pvec[i];
3197                         if (res == CLP_GANG_OKAY)
3198                                 res = (*cb)(env, io, ops, cbdata);
3199
3200                         page = ops->ops_cl.cpl_page;
3201                         lu_ref_del(&page->cp_reference, "gang_lookup", current);
3202                         cl_pagevec_put(env, page, pagevec);
3203                 }
3204                 pagevec_release(pagevec);
3205
3206                 if (nr < OTI_PVEC_SIZE || end_of_region)
3207                         break;
3208
3209                 if (res == CLP_GANG_OKAY && need_resched())
3210                         res = CLP_GANG_RESCHED;
3211                 if (res != CLP_GANG_OKAY)
3212                         break;
3213
3214                 spin_lock(&osc->oo_tree_lock);
3215                 tree_lock = true;
3216         }
3217         if (tree_lock)
3218                 spin_unlock(&osc->oo_tree_lock);
3219         RETURN(res);
3220 }
3221 EXPORT_SYMBOL(osc_page_gang_lookup);
3222
3223 /**
3224  * Check if page @page is covered by an extra lock or discard it.
3225  */
3226 static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3227                                 struct osc_page *ops, void *cbdata)
3228 {
3229         struct osc_thread_info *info = osc_env_info(env);
3230         struct osc_object *osc = cbdata;
3231         pgoff_t index;
3232
3233         index = osc_index(ops);
3234         if (index >= info->oti_fn_index) {
3235                 struct ldlm_lock *tmp;
3236                 struct cl_page *page = ops->ops_cl.cpl_page;
3237
3238                 /* refresh non-overlapped index */
3239                 tmp = osc_dlmlock_at_pgoff(env, osc, index,
3240                                            OSC_DAP_FL_TEST_LOCK);
3241                 if (tmp != NULL) {
3242                         __u64 end = tmp->l_policy_data.l_extent.end;
3243                         /* Cache the first-non-overlapped index so as to skip
3244                          * all pages within [index, oti_fn_index). This is safe
3245                          * because if tmp lock is canceled, it will discard
3246                          * these pages. */
3247                         info->oti_fn_index = cl_index(osc2cl(osc), end + 1);
3248                         if (end == OBD_OBJECT_EOF)
3249                                 info->oti_fn_index = CL_PAGE_EOF;
3250                         LDLM_LOCK_PUT(tmp);
3251                 } else if (cl_page_own(env, io, page) == 0) {
3252                         /* discard the page */
3253                         cl_page_discard(env, io, page);
3254                         cl_page_disown(env, io, page);
3255                 } else {
3256                         LASSERT(page->cp_state == CPS_FREEING);
3257                 }
3258         }
3259
3260         info->oti_next_index = index + 1;
3261         return CLP_GANG_OKAY;
3262 }
3263
3264 int osc_discard_cb(const struct lu_env *env, struct cl_io *io,
3265                    struct osc_page *ops, void *cbdata)
3266 {
3267         struct osc_thread_info *info = osc_env_info(env);
3268         struct cl_page *page = ops->ops_cl.cpl_page;
3269
3270         /* page is top page. */
3271         info->oti_next_index = osc_index(ops) + 1;
3272         if (cl_page_own(env, io, page) == 0) {
3273                 if (!ergo(page->cp_type == CPT_CACHEABLE,
3274                           !PageDirty(cl_page_vmpage(page))))
3275                         CL_PAGE_DEBUG(D_ERROR, env, page,
3276                                         "discard dirty page?\n");
3277
3278                 /* discard the page */
3279                 cl_page_discard(env, io, page);
3280                 cl_page_disown(env, io, page);
3281         } else {
3282                 LASSERT(page->cp_state == CPS_FREEING);
3283         }
3284
3285         return CLP_GANG_OKAY;
3286 }
3287 EXPORT_SYMBOL(osc_discard_cb);
3288
3289 /**
3290  * Discard pages protected by the given lock. This function traverses radix
3291  * tree to find all covering pages and discard them. If a page is being covered
3292  * by other locks, it should remain in cache.
3293  *
3294  * If error happens on any step, the process continues anyway (the reasoning
3295  * behind this being that lock cancellation cannot be delayed indefinitely).
3296  */
3297 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3298                            pgoff_t start, pgoff_t end, bool discard)
3299 {
3300         struct osc_thread_info *info = osc_env_info(env);
3301         struct cl_io *io = osc_env_thread_io(env);
3302         osc_page_gang_cbt cb;
3303         int res;
3304         int result;
3305
3306         ENTRY;
3307
3308         io->ci_obj = cl_object_top(osc2cl(osc));
3309         io->ci_ignore_layout = 1;
3310         result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3311         if (result != 0)
3312                 GOTO(out, result);
3313
3314         cb = discard ? osc_discard_cb : check_and_discard_cb;
3315         info->oti_fn_index = info->oti_next_index = start;
3316         do {
3317                 res = osc_page_gang_lookup(env, io, osc,
3318                                            info->oti_next_index, end, cb, osc);
3319                 if (info->oti_next_index > end)
3320                         break;
3321
3322                 if (res == CLP_GANG_RESCHED)
3323                         cond_resched();
3324         } while (res != CLP_GANG_OKAY);
3325 out:
3326         cl_io_fini(env, io);
3327         RETURN(result);
3328 }
3329
3330
3331 /** @} osc */