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 static inline char list_empty_marker(struct list_head *list)
  98 {
  99         return list_empty(list) ? '-' : '+';
 100 }
 101
 102 #define EXTSTR       "[%lu -> %lu/%lu]"
 103 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
 104 static const char *oes_strings[] = {
 105         "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
 106
 107 #define OSC_EXTENT_DUMP_WITH_LOC(file, func, line, mask, extent, fmt, ...) do {\
 108         static struct cfs_debug_limit_state cdls;                             \
 109         struct osc_extent *__ext = (extent);                                  \
 110         char __buf[16];                                                       \
 111                                                                               \
 112         __CDEBUG_WITH_LOC(file, func, line, mask, &cdls,                      \
 113                 "extent %p@{" EXTSTR ", "                                     \
 114                 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt,          \
 115                 /* ----- extent part 0 ----- */                               \
 116                 __ext, EXTPARA(__ext),                                        \
 117                 /* ----- part 1 ----- */                                      \
 118                 atomic_read(&__ext->oe_refc),                                 \
 119                 atomic_read(&__ext->oe_users),                                \
 120                 list_empty_marker(&__ext->oe_link),                           \
 121                 oes_strings[__ext->oe_state], ext_flags(__ext, __buf),        \
 122                 __ext->oe_obj,                                                \
 123                 /* ----- part 2 ----- */                                      \
 124                 __ext->oe_grants, __ext->oe_nr_pages,                         \
 125                 list_empty_marker(&__ext->oe_pages),                          \
 126                 waitqueue_active(&__ext->oe_waitq) ? '+' : '-',               \
 127                 __ext->oe_dlmlock, __ext->oe_mppr, __ext->oe_owner,           \
 128                 /* ----- part 4 ----- */                                      \
 129                 ## __VA_ARGS__);                                              \
 130         if (mask == D_ERROR && __ext->oe_dlmlock != NULL)                     \
 131                 LDLM_ERROR(__ext->oe_dlmlock, "extent: %p", __ext);           \
 132         else                                                                  \
 133                 LDLM_DEBUG(__ext->oe_dlmlock, "extent: %p", __ext);           \
 134 } while (0)
 135
 136 #define OSC_EXTENT_DUMP(mask, ext, fmt, ...)                            \
 137         OSC_EXTENT_DUMP_WITH_LOC(__FILE__, __func__, __LINE__,          \
 138                                  mask, ext, fmt, ## __VA_ARGS__)
 139
 140 #undef EASSERTF
 141 #define EASSERTF(expr, ext, fmt, args...) do {                          \
 142         if (!(expr)) {                                                  \
 143                 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args);           \
 144                 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj);           \
 145                 LASSERT(expr);                                          \
 146         }                                                               \
 147 } while (0)
 148
 149 #undef EASSERT
 150 #define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
 151
 152 static inline struct osc_extent *rb_extent(struct rb_node *n)
 153 {
 154         if (n == NULL)
 155                 return NULL;
 156
 157         return container_of(n, struct osc_extent, oe_node);
 158 }
 159
 160 static inline struct osc_extent *next_extent(struct osc_extent *ext)
 161 {
 162         if (ext == NULL)
 163                 return NULL;
 164
 165         LASSERT(ext->oe_intree);
 166         return rb_extent(rb_next(&ext->oe_node));
 167 }
 168
 169 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
 170 {
 171         if (ext == NULL)
 172                 return NULL;
 173
 174         LASSERT(ext->oe_intree);
 175         return rb_extent(rb_prev(&ext->oe_node));
 176 }
 177
 178 static inline struct osc_extent *first_extent(struct osc_object *obj)
 179 {
 180         return rb_extent(rb_first(&obj->oo_root));
 181 }
 182
 183 /* object must be locked by caller. */
 184 static int osc_extent_sanity_check0(struct osc_extent *ext,
 185                                     const char *func, const int line)
 186 {
 187         struct osc_object *obj = ext->oe_obj;
 188         struct osc_async_page *oap;
 189         size_t page_count;
 190         int rc = 0;
 191
 192         if (!osc_object_is_locked(obj))
 193                 GOTO(out, rc = 9);
 194
 195         if (ext->oe_state >= OES_STATE_MAX)
 196                 GOTO(out, rc = 10);
 197
 198         if (atomic_read(&ext->oe_refc) <= 0)
 199                 GOTO(out, rc = 20);
 200
 201         if (atomic_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
 202                 GOTO(out, rc = 30);
 203
 204         switch (ext->oe_state) {
 205         case OES_INV:
 206                 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
 207                         GOTO(out, rc = 35);
 208                 GOTO(out, rc = 0);
 209                 break;
 210         case OES_ACTIVE:
 211                 if (atomic_read(&ext->oe_users) == 0)
 212                         GOTO(out, rc = 40);
 213                 if (ext->oe_hp)
 214                         GOTO(out, rc = 50);
 215                 if (ext->oe_fsync_wait && !ext->oe_urgent)
 216                         GOTO(out, rc = 55);
 217                 break;
 218         case OES_CACHE:
 219                 if (ext->oe_grants == 0)
 220                         GOTO(out, rc = 60);
 221                 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
 222                         GOTO(out, rc = 65);
 223         default:
 224                 if (atomic_read(&ext->oe_users) > 0)
 225                         GOTO(out, rc = 70);
 226         }
 227
 228         if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
 229                 GOTO(out, rc = 80);
 230
 231         if (ext->oe_sync && ext->oe_grants > 0)
 232                 GOTO(out, rc = 90);
 233
 234         if (ext->oe_dlmlock != NULL &&
 235             ext->oe_dlmlock->l_resource->lr_type == LDLM_EXTENT &&
 236             !ldlm_is_failed(ext->oe_dlmlock)) {
 237                 struct ldlm_extent *extent;
 238
 239                 extent = &ext->oe_dlmlock->l_policy_data.l_extent;
 240                 if (!(extent->start <= cl_offset(osc2cl(obj), ext->oe_start) &&
 241                       extent->end   >= cl_offset(osc2cl(obj), ext->oe_max_end)))
 242                         GOTO(out, rc = 100);
 243
 244                 if (!(ext->oe_dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP)))
 245                         GOTO(out, rc = 102);
 246         }
 247
 248         if (ext->oe_nr_pages > ext->oe_mppr)
 249                 GOTO(out, rc = 105);
 250
 251         /* Do not verify page list if extent is in RPC. This is because an
 252          * in-RPC extent is supposed to be exclusively accessible w/o lock. */
 253         if (ext->oe_state > OES_CACHE)
 254                 GOTO(out, rc = 0);
 255
 256         if (!extent_debug)
 257                 GOTO(out, rc = 0);
 258
 259         page_count = 0;
 260         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
 261                 pgoff_t index = osc_index(oap2osc(oap));
 262                 ++page_count;
 263                 if (index > ext->oe_end || index < ext->oe_start)
 264                         GOTO(out, rc = 110);
 265         }
 266         if (page_count != ext->oe_nr_pages)
 267                 GOTO(out, rc = 120);
 268
 269 out:
 270         if (rc != 0)
 271                 OSC_EXTENT_DUMP_WITH_LOC(__FILE__, func, line, D_ERROR, ext,
 272                                          "sanity check %p failed: rc = %d\n",
 273                                          ext, rc);
 274         return rc;
 275 }
 276
 277 #define sanity_check_nolock(ext) \
 278         osc_extent_sanity_check0(ext, __func__, __LINE__)
 279
 280 #define sanity_check(ext) ({                                                   \
 281         int __res;                                                             \
 282         osc_object_lock((ext)->oe_obj);                                        \
 283         __res = sanity_check_nolock(ext);                                      \
 284         osc_object_unlock((ext)->oe_obj);                                      \
 285         __res;                                                                 \
 286 })
 287
 288
 289 /**
 290  * sanity check - to make sure there is no overlapped extent in the tree.
 291  */
 292 static int osc_extent_is_overlapped(struct osc_object *obj,
 293                                     struct osc_extent *ext)
 294 {
 295         struct osc_extent *tmp;
 296
 297         LASSERT(osc_object_is_locked(obj));
 298
 299         if (!extent_debug)
 300                 return 0;
 301
 302         for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
 303                 if (tmp == ext)
 304                         continue;
 305                 if (tmp->oe_end >= ext->oe_start &&
 306                     tmp->oe_start <= ext->oe_end)
 307                         return 1;
 308         }
 309         return 0;
 310 }
 311
 312 static void osc_extent_state_set(struct osc_extent *ext, int state)
 313 {
 314         LASSERT(osc_object_is_locked(ext->oe_obj));
 315         LASSERT(state >= OES_INV && state < OES_STATE_MAX);
 316
 317         /* Never try to sanity check a state changing extent :-) */
 318         /* LASSERT(sanity_check_nolock(ext) == 0); */
 319
 320         /* TODO: validate the state machine */
 321         ext->oe_state = state;
 322         wake_up_all(&ext->oe_waitq);
 323 }
 324
 325 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
 326 {
 327         struct osc_extent *ext;
 328
 329         OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_NOFS);
 330         if (ext == NULL)
 331                 return NULL;
 332
 333         RB_CLEAR_NODE(&ext->oe_node);
 334         ext->oe_obj = obj;
 335         cl_object_get(osc2cl(obj));
 336         atomic_set(&ext->oe_refc, 1);
 337         atomic_set(&ext->oe_users, 0);
 338         INIT_LIST_HEAD(&ext->oe_link);
 339         ext->oe_state = OES_INV;
 340         INIT_LIST_HEAD(&ext->oe_pages);
 341         init_waitqueue_head(&ext->oe_waitq);
 342         ext->oe_dlmlock = NULL;
 343
 344         return ext;
 345 }
 346
 347 static void osc_extent_free(struct osc_extent *ext)
 348 {
 349         OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
 350 }
 351
 352 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
 353 {
 354         LASSERT(atomic_read(&ext->oe_refc) >= 0);
 355         atomic_inc(&ext->oe_refc);
 356         return ext;
 357 }
 358
 359 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
 360 {
 361         LASSERT(atomic_read(&ext->oe_refc) > 0);
 362         if (atomic_dec_and_test(&ext->oe_refc)) {
 363                 LASSERT(list_empty(&ext->oe_link));
 364                 LASSERT(atomic_read(&ext->oe_users) == 0);
 365                 LASSERT(ext->oe_state == OES_INV);
 366                 LASSERT(!ext->oe_intree);
 367
 368                 if (ext->oe_dlmlock != NULL) {
 369                         lu_ref_add(&ext->oe_dlmlock->l_reference,
 370                                    "osc_extent", ext);
 371                         LDLM_LOCK_PUT(ext->oe_dlmlock);
 372                         ext->oe_dlmlock = NULL;
 373                 }
 374                 cl_object_put(env, osc2cl(ext->oe_obj));
 375                 osc_extent_free(ext);
 376         }
 377 }
 378
 379 /**
 380  * osc_extent_put_trust() is a special version of osc_extent_put() when
 381  * it's known that the caller is not the last user. This is to address the
 382  * problem of lacking of lu_env ;-).
 383  */
 384 static void osc_extent_put_trust(struct osc_extent *ext)
 385 {
 386         LASSERT(atomic_read(&ext->oe_refc) > 1);
 387         LASSERT(osc_object_is_locked(ext->oe_obj));
 388         atomic_dec(&ext->oe_refc);
 389 }
 390
 391 /**
 392  * Return the extent which includes pgoff @index, or return the greatest
 393  * previous extent in the tree.
 394  */
 395 static struct osc_extent *osc_extent_search(struct osc_object *obj,
 396                                             pgoff_t index)
 397 {
 398         struct rb_node    *n = obj->oo_root.rb_node;
 399         struct osc_extent *tmp, *p = NULL;
 400
 401         LASSERT(osc_object_is_locked(obj));
 402         while (n != NULL) {
 403                 tmp = rb_extent(n);
 404                 if (index < tmp->oe_start) {
 405                         n = n->rb_left;
 406                 } else if (index > tmp->oe_end) {
 407                         p = rb_extent(n);
 408                         n = n->rb_right;
 409                 } else {
 410                         return tmp;
 411                 }
 412         }
 413         return p;
 414 }
 415
 416 /*
 417  * Return the extent covering @index, otherwise return NULL.
 418  * caller must have held object lock.
 419  */
 420 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
 421                                             pgoff_t index)
 422 {
 423         struct osc_extent *ext;
 424
 425         ext = osc_extent_search(obj, index);
 426         if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
 427                 return osc_extent_get(ext);
 428         return NULL;
 429 }
 430
 431 /* caller must have held object lock. */
 432 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
 433 {
 434         struct rb_node   **n      = &obj->oo_root.rb_node;
 435         struct rb_node    *parent = NULL;
 436         struct osc_extent *tmp;
 437
 438         LASSERT(ext->oe_intree == 0);
 439         LASSERT(ext->oe_obj == obj);
 440         LASSERT(osc_object_is_locked(obj));
 441         while (*n != NULL) {
 442                 tmp = rb_extent(*n);
 443                 parent = *n;
 444
 445                 if (ext->oe_end < tmp->oe_start)
 446                         n = &(*n)->rb_left;
 447                 else if (ext->oe_start > tmp->oe_end)
 448                         n = &(*n)->rb_right;
 449                 else
 450                         EASSERTF(0, tmp, EXTSTR"\n", EXTPARA(ext));
 451         }
 452         rb_link_node(&ext->oe_node, parent, n);
 453         rb_insert_color(&ext->oe_node, &obj->oo_root);
 454         osc_extent_get(ext);
 455         ext->oe_intree = 1;
 456 }
 457
 458 /* caller must have held object lock. */
 459 static void osc_extent_erase(struct osc_extent *ext)
 460 {
 461         struct osc_object *obj = ext->oe_obj;
 462         LASSERT(osc_object_is_locked(obj));
 463         if (ext->oe_intree) {
 464                 rb_erase(&ext->oe_node, &obj->oo_root);
 465                 ext->oe_intree = 0;
 466                 /* rbtree held a refcount */
 467                 osc_extent_put_trust(ext);
 468         }
 469 }
 470
 471 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
 472 {
 473         struct osc_object *obj = ext->oe_obj;
 474
 475         LASSERT(osc_object_is_locked(obj));
 476         LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
 477         if (ext->oe_state == OES_CACHE) {
 478                 osc_extent_state_set(ext, OES_ACTIVE);
 479                 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
 480         }
 481         atomic_inc(&ext->oe_users);
 482         list_del_init(&ext->oe_link);
 483         return osc_extent_get(ext);
 484 }
 485
 486 static void __osc_extent_remove(struct osc_extent *ext)
 487 {
 488         LASSERT(osc_object_is_locked(ext->oe_obj));
 489         LASSERT(list_empty(&ext->oe_pages));
 490         osc_extent_erase(ext);
 491         list_del_init(&ext->oe_link);
 492         osc_extent_state_set(ext, OES_INV);
 493         OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
 494 }
 495
 496 static void osc_extent_remove(struct osc_extent *ext)
 497 {
 498         struct osc_object *obj = ext->oe_obj;
 499
 500         osc_object_lock(obj);
 501         __osc_extent_remove(ext);
 502         osc_object_unlock(obj);
 503 }
 504
 505 /**
 506  * This function is used to merge extents to get better performance. It checks
 507  * if @cur and @victim are contiguous at block level.
 508  */
 509 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
 510                             struct osc_extent *victim)
 511 {
 512         struct osc_object       *obj = cur->oe_obj;
 513         struct client_obd       *cli = osc_cli(obj);
 514         pgoff_t                  chunk_start;
 515         pgoff_t                  chunk_end;
 516         int                      ppc_bits;
 517
 518         LASSERT(cur->oe_state == OES_CACHE);
 519         LASSERT(osc_object_is_locked(obj));
 520         if (victim == NULL)
 521                 return -EINVAL;
 522
 523         if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
 524                 return -EBUSY;
 525
 526         if (cur->oe_max_end != victim->oe_max_end)
 527                 return -ERANGE;
 528
 529         LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
 530         ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_SHIFT;
 531         chunk_start = cur->oe_start >> ppc_bits;
 532         chunk_end   = cur->oe_end   >> ppc_bits;
 533         if (chunk_start   != (victim->oe_end >> ppc_bits) + 1 &&
 534             chunk_end + 1 != victim->oe_start >> ppc_bits)
 535                 return -ERANGE;
 536
 537         /* overall extent size should not exceed the max supported limit
 538          * reported by the server */
 539         if (cur->oe_end - cur->oe_start + 1 +
 540             victim->oe_end - victim->oe_start + 1 > cli->cl_max_extent_pages)
 541                 return -ERANGE;
 542
 543         OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
 544
 545         cur->oe_start     = min(cur->oe_start, victim->oe_start);
 546         cur->oe_end       = max(cur->oe_end,   victim->oe_end);
 547         /* per-extent tax should be accounted only once for the whole extent */
 548         cur->oe_grants   += victim->oe_grants - cli->cl_grant_extent_tax;
 549         cur->oe_nr_pages += victim->oe_nr_pages;
 550         /* only the following bits are needed to merge */
 551         cur->oe_urgent   |= victim->oe_urgent;
 552         cur->oe_memalloc |= victim->oe_memalloc;
 553         list_splice_init(&victim->oe_pages, &cur->oe_pages);
 554         list_del_init(&victim->oe_link);
 555         victim->oe_nr_pages = 0;
 556
 557         osc_extent_get(victim);
 558         __osc_extent_remove(victim);
 559         osc_extent_put(env, victim);
 560
 561         OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
 562         return 0;
 563 }
 564
 565 /**
 566  * Drop user count of osc_extent, and unplug IO asynchronously.
 567  */
 568 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
 569 {
 570         struct osc_object *obj = ext->oe_obj;
 571         struct client_obd *cli = osc_cli(obj);
 572         int rc = 0;
 573         ENTRY;
 574
 575         LASSERT(atomic_read(&ext->oe_users) > 0);
 576         LASSERT(sanity_check(ext) == 0);
 577         LASSERT(ext->oe_grants > 0);
 578
 579         if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
 580                 LASSERT(ext->oe_state == OES_ACTIVE);
 581                 if (ext->oe_trunc_pending) {
 582                         /* a truncate process is waiting for this extent.
 583                          * This may happen due to a race, check
 584                          * osc_cache_truncate_start(). */
 585                         osc_extent_state_set(ext, OES_TRUNC);
 586                         ext->oe_trunc_pending = 0;
 587                 } else {
 588                         int grant = 0;
 589
 590                         osc_extent_state_set(ext, OES_CACHE);
 591                         osc_update_pending(obj, OBD_BRW_WRITE,
 592                                            ext->oe_nr_pages);
 593
 594                         /* try to merge the previous and next extent. */
 595                         if (osc_extent_merge(env, ext, prev_extent(ext)) == 0)
 596                                 grant += cli->cl_grant_extent_tax;
 597                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 598                                 grant += cli->cl_grant_extent_tax;
 599                         if (grant > 0)
 600                                 osc_unreserve_grant(cli, 0, grant);
 601
 602                         if (ext->oe_urgent)
 603                                 list_move_tail(&ext->oe_link,
 604                                                &obj->oo_urgent_exts);
 605                         else if (ext->oe_nr_pages == ext->oe_mppr) {
 606                                 list_move_tail(&ext->oe_link,
 607                                                &obj->oo_full_exts);
 608                         }
 609                 }
 610                 osc_object_unlock(obj);
 611
 612                 osc_io_unplug_async(env, cli, obj);
 613         }
 614         osc_extent_put(env, ext);
 615         RETURN(rc);
 616 }
 617
 618 static inline int overlapped(struct osc_extent *ex1, struct osc_extent *ex2)
 619 {
 620         return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
 621 }
 622
 623 /**
 624  * Find or create an extent which includes @index, core function to manage
 625  * extent tree.
 626  */
 627 static struct osc_extent *osc_extent_find(const struct lu_env *env,
 628                                           struct osc_object *obj, pgoff_t index,
 629                                           unsigned int *grants)
 630 {
 631         struct client_obd *cli = osc_cli(obj);
 632         struct osc_lock   *olck;
 633         struct cl_lock_descr *descr;
 634         struct osc_extent *cur;
 635         struct osc_extent *ext;
 636         struct osc_extent *conflict = NULL;
 637         struct osc_extent *found = NULL;
 638         pgoff_t    chunk;
 639         pgoff_t    max_end;
 640         unsigned int max_pages; /* max_pages_per_rpc */
 641         unsigned int chunksize;
 642         int        ppc_bits; /* pages per chunk bits */
 643         pgoff_t    chunk_mask;
 644         int        rc;
 645         ENTRY;
 646
 647         cur = osc_extent_alloc(obj);
 648         if (cur == NULL)
 649                 RETURN(ERR_PTR(-ENOMEM));
 650
 651         olck = osc_env_io(env)->oi_write_osclock;
 652         LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
 653         LASSERT(olck->ols_state == OLS_GRANTED);
 654
 655         descr = &olck->ols_cl.cls_lock->cll_descr;
 656         LASSERT(descr->cld_mode >= CLM_WRITE);
 657
 658         LASSERTF(cli->cl_chunkbits >= PAGE_SHIFT,
 659                  "chunkbits: %u\n", cli->cl_chunkbits);
 660         ppc_bits   = cli->cl_chunkbits - PAGE_SHIFT;
 661         chunk_mask = ~((1 << ppc_bits) - 1);
 662         chunksize  = 1 << cli->cl_chunkbits;
 663         chunk      = index >> ppc_bits;
 664
 665         /* align end to RPC edge. */
 666         max_pages = cli->cl_max_pages_per_rpc;
 667         if ((max_pages & ~chunk_mask) != 0) {
 668                 CERROR("max_pages: %#x chunkbits: %u chunk_mask: %#lx\n",
 669                        max_pages, cli->cl_chunkbits, chunk_mask);
 670                 RETURN(ERR_PTR(-EINVAL));
 671         }
 672         max_end = index - (index % max_pages) + max_pages - 1;
 673         max_end = min_t(pgoff_t, max_end, descr->cld_end);
 674
 675         /* initialize new extent by parameters so far */
 676         cur->oe_max_end = max_end;
 677         cur->oe_start   = index & chunk_mask;
 678         cur->oe_end     = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
 679         if (cur->oe_start < descr->cld_start)
 680                 cur->oe_start = descr->cld_start;
 681         if (cur->oe_end > max_end)
 682                 cur->oe_end = max_end;
 683         cur->oe_grants  = 0;
 684         cur->oe_mppr    = max_pages;
 685         if (olck->ols_dlmlock != NULL) {
 686                 LASSERT(olck->ols_hold);
 687                 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
 688                 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
 689         }
 690
 691         /* grants has been allocated by caller */
 692         LASSERTF(*grants >= chunksize + cli->cl_grant_extent_tax,
 693                  "%u/%u/%u.\n", *grants, chunksize, cli->cl_grant_extent_tax);
 694         LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
 695                  EXTPARA(cur));
 696
 697 restart:
 698         osc_object_lock(obj);
 699         ext = osc_extent_search(obj, cur->oe_start);
 700         if (ext == NULL)
 701                 ext = first_extent(obj);
 702         while (ext != NULL) {
 703                 pgoff_t ext_chk_start = ext->oe_start >> ppc_bits;
 704                 pgoff_t ext_chk_end   = ext->oe_end   >> ppc_bits;
 705
 706                 LASSERT(sanity_check_nolock(ext) == 0);
 707                 if (chunk > ext_chk_end + 1 || chunk < ext_chk_start)
 708                         break;
 709
 710                 /* if covering by different locks, no chance to match */
 711                 if (olck->ols_dlmlock != ext->oe_dlmlock) {
 712                         EASSERTF(!overlapped(ext, cur), ext,
 713                                  EXTSTR"\n", EXTPARA(cur));
 714
 715                         ext = next_extent(ext);
 716                         continue;
 717                 }
 718
 719                 /* discontiguous chunks? */
 720                 if (chunk + 1 < ext_chk_start) {
 721                         ext = next_extent(ext);
 722                         continue;
 723                 }
 724
 725                 /* ok, from now on, ext and cur have these attrs:
 726                  * 1. covered by the same lock
 727                  * 2. contiguous at chunk level or overlapping. */
 728
 729                 if (overlapped(ext, cur)) {
 730                         /* cur is the minimum unit, so overlapping means
 731                          * full contain. */
 732                         EASSERTF((ext->oe_start <= cur->oe_start &&
 733                                   ext->oe_end >= cur->oe_end),
 734                                  ext, EXTSTR"\n", EXTPARA(cur));
 735
 736                         if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
 737                                 /* for simplicity, we wait for this extent to
 738                                  * finish before going forward. */
 739                                 conflict = osc_extent_get(ext);
 740                                 break;
 741                         }
 742
 743                         found = osc_extent_hold(ext);
 744                         break;
 745                 }
 746
 747                 /* non-overlapped extent */
 748                 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
 749                         /* we can't do anything for a non OES_CACHE extent, or
 750                          * if there is someone waiting for this extent to be
 751                          * flushed, try next one. */
 752                         ext = next_extent(ext);
 753                         continue;
 754                 }
 755
 756                 /* check if they belong to the same rpc slot before trying to
 757                  * merge. the extents are not overlapped and contiguous at
 758                  * chunk level to get here. */
 759                 if (ext->oe_max_end != max_end) {
 760                         /* if they don't belong to the same RPC slot or
 761                          * max_pages_per_rpc has ever changed, do not merge. */
 762                         ext = next_extent(ext);
 763                         continue;
 764                 }
 765
 766                 /* check whether maximum extent size will be hit */
 767                 if ((ext_chk_end - ext_chk_start + 1 + 1) << ppc_bits >
 768                     cli->cl_max_extent_pages) {
 769                         ext = next_extent(ext);
 770                         continue;
 771                 }
 772
 773                 /* it's required that an extent must be contiguous at chunk
 774                  * level so that we know the whole extent is covered by grant
 775                  * (the pages in the extent are NOT required to be contiguous).
 776                  * Otherwise, it will be too much difficult to know which
 777                  * chunks have grants allocated. */
 778
 779                 /* try to do front merge - extend ext's start */
 780                 if (chunk + 1 == ext_chk_start) {
 781                         /* ext must be chunk size aligned */
 782                         EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
 783
 784                         /* pull ext's start back to cover cur */
 785                         ext->oe_start   = cur->oe_start;
 786                         ext->oe_grants += chunksize;
 787                         LASSERT(*grants >= chunksize);
 788                         *grants -= chunksize;
 789
 790                         found = osc_extent_hold(ext);
 791                 } else if (chunk == ext_chk_end + 1) {
 792                         /* rear merge */
 793                         ext->oe_end     = cur->oe_end;
 794                         ext->oe_grants += chunksize;
 795                         LASSERT(*grants >= chunksize);
 796                         *grants -= chunksize;
 797
 798                         /* try to merge with the next one because we just fill
 799                          * in a gap */
 800                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 801                                 /* we can save extent tax from next extent */
 802                                 *grants += cli->cl_grant_extent_tax;
 803
 804                         found = osc_extent_hold(ext);
 805                 }
 806                 if (found != NULL)
 807                         break;
 808
 809                 ext = next_extent(ext);
 810         }
 811
 812         osc_extent_tree_dump(D_CACHE, obj);
 813         if (found != NULL) {
 814                 LASSERT(conflict == NULL);
 815                 if (!IS_ERR(found)) {
 816                         LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
 817                         OSC_EXTENT_DUMP(D_CACHE, found,
 818                                         "found caching ext for %lu.\n", index);
 819                 }
 820         } else if (conflict == NULL) {
 821                 /* create a new extent */
 822                 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
 823                 cur->oe_grants = chunksize + cli->cl_grant_extent_tax;
 824                 LASSERT(*grants >= cur->oe_grants);
 825                 *grants -= cur->oe_grants;
 826
 827                 cur->oe_state = OES_CACHE;
 828                 found = osc_extent_hold(cur);
 829                 osc_extent_insert(obj, cur);
 830                 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
 831                                 index, descr->cld_end);
 832         }
 833         osc_object_unlock(obj);
 834
 835         if (conflict != NULL) {
 836                 LASSERT(found == NULL);
 837
 838                 /* waiting for IO to finish. Please notice that it's impossible
 839                  * to be an OES_TRUNC extent. */
 840                 rc = osc_extent_wait(env, conflict, OES_INV);
 841                 osc_extent_put(env, conflict);
 842                 conflict = NULL;
 843                 if (rc < 0)
 844                         GOTO(out, found = ERR_PTR(rc));
 845
 846                 goto restart;
 847         }
 848         EXIT;
 849
 850 out:
 851         osc_extent_put(env, cur);
 852         return found;
 853 }
 854
 855 /**
 856  * Called when IO is finished to an extent.
 857  */
 858 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
 859                       int sent, int rc)
 860 {
 861         struct client_obd *cli = osc_cli(ext->oe_obj);
 862         struct osc_async_page *oap;
 863         struct osc_async_page *tmp;
 864         int nr_pages = ext->oe_nr_pages;
 865         int lost_grant = 0;
 866         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
 867         loff_t last_off = 0;
 868         int last_count = -1;
 869         ENTRY;
 870
 871         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
 872
 873         ext->oe_rc = rc ?: ext->oe_nr_pages;
 874         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
 875
 876         osc_lru_add_batch(cli, &ext->oe_pages);
 877         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 878                                      oap_pending_item) {
 879                 list_del_init(&oap->oap_rpc_item);
 880                 list_del_init(&oap->oap_pending_item);
 881                 if (last_off <= oap->oap_obj_off) {
 882                         last_off = oap->oap_obj_off;
 883                         last_count = oap->oap_count;
 884                 }
 885
 886                 --ext->oe_nr_pages;
 887                 osc_ap_completion(env, cli, oap, sent, rc);
 888         }
 889         EASSERT(ext->oe_nr_pages == 0, ext);
 890
 891         if (!sent) {
 892                 lost_grant = ext->oe_grants;
 893         } else if (blocksize < PAGE_SIZE &&
 894                    last_count != PAGE_SIZE) {
 895                 /* For short writes we shouldn't count parts of pages that
 896                  * span a whole chunk on the OST side, or our accounting goes
 897                  * wrong.  Should match the code in filter_grant_check. */
 898                 int offset = last_off & ~PAGE_MASK;
 899                 int count = last_count + (offset & (blocksize - 1));
 900                 int end = (offset + last_count) & (blocksize - 1);
 901                 if (end)
 902                         count += blocksize - end;
 903
 904                 lost_grant = PAGE_SIZE - count;
 905         }
 906         if (ext->oe_grants > 0)
 907                 osc_free_grant(cli, nr_pages, lost_grant, ext->oe_grants);
 908
 909         osc_extent_remove(ext);
 910         /* put the refcount for RPC */
 911         osc_extent_put(env, ext);
 912         RETURN(0);
 913 }
 914
 915 static int extent_wait_cb(struct osc_extent *ext, enum osc_extent_state state)
 916 {
 917         int ret;
 918
 919         osc_object_lock(ext->oe_obj);
 920         ret = ext->oe_state == state;
 921         osc_object_unlock(ext->oe_obj);
 922
 923         return ret;
 924 }
 925
 926 /**
 927  * Wait for the extent's state to become @state.
 928  */
 929 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
 930                            enum osc_extent_state state)
 931 {
 932         struct osc_object *obj = ext->oe_obj;
 933         struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
 934                                                   LWI_ON_SIGNAL_NOOP, NULL);
 935         int rc = 0;
 936         ENTRY;
 937
 938         osc_object_lock(obj);
 939         LASSERT(sanity_check_nolock(ext) == 0);
 940         /* `Kick' this extent only if the caller is waiting for it to be
 941          * written out. */
 942         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
 943                 if (ext->oe_state == OES_ACTIVE) {
 944                         ext->oe_urgent = 1;
 945                 } else if (ext->oe_state == OES_CACHE) {
 946                         ext->oe_urgent = 1;
 947                         osc_extent_hold(ext);
 948                         rc = 1;
 949                 }
 950         }
 951         osc_object_unlock(obj);
 952         if (rc == 1)
 953                 osc_extent_release(env, ext);
 954
 955         /* wait for the extent until its state becomes @state */
 956         rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
 957         if (rc == -ETIMEDOUT) {
 958                 OSC_EXTENT_DUMP(D_ERROR, ext,
 959                         "%s: wait ext to %u timedout, recovery in progress?\n",
 960                         cli_name(osc_cli(obj)), state);
 961
 962                 lwi = LWI_INTR(NULL, NULL);
 963                 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state),
 964                                   &lwi);
 965         }
 966         if (rc == 0 && ext->oe_rc < 0)
 967                 rc = ext->oe_rc;
 968         RETURN(rc);
 969 }
 970
 971 /**
 972  * Discard pages with index greater than @size. If @ext is overlapped with
 973  * @size, then partial truncate happens.
 974  */
 975 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
 976                                 bool partial)
 977 {
 978         struct lu_env         *env;
 979         struct cl_io          *io;
 980         struct osc_object     *obj = ext->oe_obj;
 981         struct client_obd     *cli = osc_cli(obj);
 982         struct osc_async_page *oap;
 983         struct osc_async_page *tmp;
 984         struct pagevec        *pvec;
 985         int                    pages_in_chunk = 0;
 986         int                    ppc_bits    = cli->cl_chunkbits -
 987                                              PAGE_SHIFT;
 988         __u64                  trunc_chunk = trunc_index >> ppc_bits;
 989         int                    grants   = 0;
 990         int                    nr_pages = 0;
 991         int                    rc       = 0;
 992         __u16                  refcheck;
 993         ENTRY;
 994
 995         LASSERT(sanity_check(ext) == 0);
 996         LASSERT(ext->oe_state == OES_TRUNC);
 997         LASSERT(!ext->oe_urgent);
 998
 999         /* Request new lu_env.
1000          * We can't use that env from osc_cache_truncate_start() because
1001          * it's from lov_io_sub and not fully initialized. */
1002         env = cl_env_get(&refcheck);
1003         if (IS_ERR(env))
1004                 RETURN(PTR_ERR(env));
1005
1006         io  = osc_env_thread_io(env);
1007         io->ci_obj = cl_object_top(osc2cl(obj));
1008         io->ci_ignore_layout = 1;
1009         pvec = &osc_env_info(env)->oti_pagevec;
1010         ll_pagevec_init(pvec, 0);
1011         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
1012         if (rc < 0)
1013                 GOTO(out, rc);
1014
1015         /* discard all pages with index greater than trunc_index */
1016         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
1017                                      oap_pending_item) {
1018                 pgoff_t index = osc_index(oap2osc(oap));
1019                 struct cl_page  *page = oap2cl_page(oap);
1020
1021                 LASSERT(list_empty(&oap->oap_rpc_item));
1022
1023                 /* only discard the pages with their index greater than
1024                  * trunc_index, and ... */
1025                 if (index < trunc_index ||
1026                     (index == trunc_index && partial)) {
1027                         /* accounting how many pages remaining in the chunk
1028                          * so that we can calculate grants correctly. */
1029                         if (index >> ppc_bits == trunc_chunk)
1030                                 ++pages_in_chunk;
1031                         continue;
1032                 }
1033
1034                 list_del_init(&oap->oap_pending_item);
1035
1036                 cl_page_get(page);
1037                 lu_ref_add(&page->cp_reference, "truncate", current);
1038
1039                 if (cl_page_own(env, io, page) == 0) {
1040                         cl_page_discard(env, io, page);
1041                         cl_page_disown(env, io, page);
1042                 } else {
1043                         LASSERT(page->cp_state == CPS_FREEING);
1044                         LASSERT(0);
1045                 }
1046
1047                 lu_ref_del(&page->cp_reference, "truncate", current);
1048                 cl_pagevec_put(env, page, pvec);
1049
1050                 --ext->oe_nr_pages;
1051                 ++nr_pages;
1052         }
1053         pagevec_release(pvec);
1054
1055         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1056                       ext->oe_nr_pages == 0),
1057                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1058
1059         osc_object_lock(obj);
1060         if (ext->oe_nr_pages == 0) {
1061                 LASSERT(pages_in_chunk == 0);
1062                 grants = ext->oe_grants;
1063                 ext->oe_grants = 0;
1064         } else { /* calculate how many grants we can free */
1065                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1066                 pgoff_t last_index;
1067
1068
1069                 /* if there is no pages in this chunk, we can also free grants
1070                  * for the last chunk */
1071                 if (pages_in_chunk == 0) {
1072                         /* if this is the 1st chunk and no pages in this chunk,
1073                          * ext->oe_nr_pages must be zero, so we should be in
1074                          * the other if-clause. */
1075                         LASSERT(trunc_chunk > 0);
1076                         --trunc_chunk;
1077                         ++chunks;
1078                 }
1079
1080                 /* this is what we can free from this extent */
1081                 grants          = chunks << cli->cl_chunkbits;
1082                 ext->oe_grants -= grants;
1083                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1084                 ext->oe_end     = min(last_index, ext->oe_max_end);
1085                 LASSERT(ext->oe_end >= ext->oe_start);
1086                 LASSERT(ext->oe_grants > 0);
1087         }
1088         osc_object_unlock(obj);
1089
1090         if (grants > 0 || nr_pages > 0)
1091                 osc_free_grant(cli, nr_pages, grants, grants);
1092
1093 out:
1094         cl_io_fini(env, io);
1095         cl_env_put(env, &refcheck);
1096         RETURN(rc);
1097 }
1098
1099 /**
1100  * This function is used to make the extent prepared for transfer.
1101  * A race with flusing page - ll_writepage() has to be handled cautiously.
1102  */
1103 static int osc_extent_make_ready(const struct lu_env *env,
1104                                  struct osc_extent *ext)
1105 {
1106         struct osc_async_page *oap;
1107         struct osc_async_page *last = NULL;
1108         struct osc_object *obj = ext->oe_obj;
1109         unsigned int page_count = 0;
1110         int rc;
1111         ENTRY;
1112
1113         /* we're going to grab page lock, so object lock must not be taken. */
1114         LASSERT(sanity_check(ext) == 0);
1115         /* in locking state, any process should not touch this extent. */
1116         EASSERT(ext->oe_state == OES_LOCKING, ext);
1117         EASSERT(ext->oe_owner != NULL, ext);
1118
1119         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1120
1121         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1122                 ++page_count;
1123                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1124                         last = oap;
1125
1126                 /* checking ASYNC_READY is race safe */
1127                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1128                         continue;
1129
1130                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1131                 switch (rc) {
1132                 case 0:
1133                         spin_lock(&oap->oap_lock);
1134                         oap->oap_async_flags |= ASYNC_READY;
1135                         spin_unlock(&oap->oap_lock);
1136                         break;
1137                 case -EALREADY:
1138                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1139                         break;
1140                 default:
1141                         LASSERTF(0, "unknown return code: %d\n", rc);
1142                 }
1143         }
1144
1145         LASSERT(page_count == ext->oe_nr_pages);
1146         LASSERT(last != NULL);
1147         /* the last page is the only one we need to refresh its count by
1148          * the size of file. */
1149         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1150                 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1151                 LASSERT(last_oap_count > 0);
1152                 LASSERT(last->oap_page_off + last_oap_count <= PAGE_SIZE);
1153                 last->oap_count = last_oap_count;
1154                 spin_lock(&last->oap_lock);
1155                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1156                 spin_unlock(&last->oap_lock);
1157         }
1158
1159         /* for the rest of pages, we don't need to call osf_refresh_count()
1160          * because it's known they are not the last page */
1161         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1162                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1163                         oap->oap_count = PAGE_SIZE - oap->oap_page_off;
1164                         spin_lock(&oap->oap_lock);
1165                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1166                         spin_unlock(&oap->oap_lock);
1167                 }
1168         }
1169
1170         osc_object_lock(obj);
1171         osc_extent_state_set(ext, OES_RPC);
1172         osc_object_unlock(obj);
1173         /* get a refcount for RPC. */
1174         osc_extent_get(ext);
1175
1176         RETURN(0);
1177 }
1178
1179 /**
1180  * Quick and simple version of osc_extent_find(). This function is frequently
1181  * called to expand the extent for the same IO. To expand the extent, the
1182  * page index must be in the same or next chunk of ext->oe_end.
1183  */
1184 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index,
1185                              unsigned int *grants)
1186 {
1187         struct osc_object *obj = ext->oe_obj;
1188         struct client_obd *cli = osc_cli(obj);
1189         struct osc_extent *next;
1190         int ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1191         pgoff_t chunk = index >> ppc_bits;
1192         pgoff_t end_chunk;
1193         pgoff_t end_index;
1194         unsigned int chunksize = 1 << cli->cl_chunkbits;
1195         int rc = 0;
1196         ENTRY;
1197
1198         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1199         osc_object_lock(obj);
1200         LASSERT(sanity_check_nolock(ext) == 0);
1201         end_chunk = ext->oe_end >> ppc_bits;
1202         if (chunk > end_chunk + 1)
1203                 GOTO(out, rc = -ERANGE);
1204
1205         if (end_chunk >= chunk)
1206                 GOTO(out, rc = 0);
1207
1208         LASSERT(end_chunk + 1 == chunk);
1209
1210         /* try to expand this extent to cover @index */
1211         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1212
1213         /* don't go over the maximum extent size reported by server */
1214         if (end_index - ext->oe_start + 1 > cli->cl_max_extent_pages)
1215                 GOTO(out, rc = -ERANGE);
1216
1217         next = next_extent(ext);
1218         if (next != NULL && next->oe_start <= end_index)
1219                 /* complex mode - overlapped with the next extent,
1220                  * this case will be handled by osc_extent_find() */
1221                 GOTO(out, rc = -EAGAIN);
1222
1223         ext->oe_end = end_index;
1224         ext->oe_grants += chunksize;
1225         LASSERT(*grants >= chunksize);
1226         *grants -= chunksize;
1227         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1228                  "overlapped after expanding for %lu.\n", index);
1229         EXIT;
1230
1231 out:
1232         osc_object_unlock(obj);
1233         RETURN(rc);
1234 }
1235
1236 static void osc_extent_tree_dump0(int mask, struct osc_object *obj,
1237                                   const char *func, int line)
1238 {
1239         struct osc_extent *ext;
1240         int cnt;
1241
1242         if (!cfs_cdebug_show(mask, DEBUG_SUBSYSTEM))
1243                 return;
1244
1245         CDEBUG(mask, "Dump object %p extents at %s:%d, mppr: %u.\n",
1246                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1247
1248         /* osc_object_lock(obj); */
1249         cnt = 1;
1250         for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1251                 OSC_EXTENT_DUMP(mask, ext, "in tree %d.\n", cnt++);
1252
1253         cnt = 1;
1254         list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1255                 OSC_EXTENT_DUMP(mask, ext, "hp %d.\n", cnt++);
1256
1257         cnt = 1;
1258         list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1259                 OSC_EXTENT_DUMP(mask, ext, "urgent %d.\n", cnt++);
1260
1261         cnt = 1;
1262         list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1263                 OSC_EXTENT_DUMP(mask, ext, "reading %d.\n", cnt++);
1264         /* osc_object_unlock(obj); */
1265 }
1266
1267 /* ------------------ osc extent end ------------------ */
1268
1269 static inline int osc_is_ready(struct osc_object *osc)
1270 {
1271         return !list_empty(&osc->oo_ready_item) ||
1272                !list_empty(&osc->oo_hp_ready_item);
1273 }
1274
1275 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1276         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1277                (OSC), osc_is_ready(OSC),                                       \
1278                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1279                list_empty_marker(&(OSC)->oo_ready_item),                       \
1280                atomic_read(&(OSC)->oo_nr_writes),                              \
1281                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1282                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1283                atomic_read(&(OSC)->oo_nr_reads),                               \
1284                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1285                ##args)
1286
1287 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1288                           int cmd)
1289 {
1290         struct osc_page *opg  = oap2osc_page(oap);
1291         struct cl_page  *page = oap2cl_page(oap);
1292         int result;
1293
1294         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1295
1296         ENTRY;
1297         result = cl_page_make_ready(env, page, CRT_WRITE);
1298         if (result == 0)
1299                 opg->ops_submit_time = ktime_get();
1300         RETURN(result);
1301 }
1302
1303 static int osc_refresh_count(const struct lu_env *env,
1304                              struct osc_async_page *oap, int cmd)
1305 {
1306         struct osc_page  *opg = oap2osc_page(oap);
1307         pgoff_t index = osc_index(oap2osc(oap));
1308         struct cl_object *obj;
1309         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
1310         int result;
1311         loff_t kms;
1312
1313         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1314         LASSERT(!(cmd & OBD_BRW_READ));
1315         LASSERT(opg != NULL);
1316         obj = opg->ops_cl.cpl_obj;
1317
1318         cl_object_attr_lock(obj);
1319         result = cl_object_attr_get(env, obj, attr);
1320         cl_object_attr_unlock(obj);
1321         if (result < 0)
1322                 return result;
1323         kms = attr->cat_kms;
1324         if (cl_offset(obj, index) >= kms)
1325                 /* catch race with truncate */
1326                 return 0;
1327         else if (cl_offset(obj, index + 1) > kms)
1328                 /* catch sub-page write at end of file */
1329                 return kms % PAGE_SIZE;
1330         else
1331                 return PAGE_SIZE;
1332 }
1333
1334 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1335                           int cmd, int rc)
1336 {
1337         struct osc_page   *opg  = oap2osc_page(oap);
1338         struct cl_page    *page = oap2cl_page(oap);
1339         enum cl_req_type   crt;
1340         int srvlock;
1341
1342         ENTRY;
1343
1344         cmd &= ~OBD_BRW_NOQUOTA;
1345         LASSERTF(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ),
1346                  "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1347         LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1348                 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1349         LASSERT(opg->ops_transfer_pinned);
1350
1351         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1352         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1353         opg->ops_transfer_pinned = 0;
1354
1355         opg->ops_submit_time = ktime_set(0, 0);
1356         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1357
1358         /* statistic */
1359         if (rc == 0 && srvlock) {
1360                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1361                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1362                 size_t bytes = oap->oap_count;
1363
1364                 if (crt == CRT_READ)
1365                         stats->os_lockless_reads += bytes;
1366                 else
1367                         stats->os_lockless_writes += bytes;
1368         }
1369
1370         /*
1371          * This has to be the last operation with the page, as locks are
1372          * released in cl_page_completion() and nothing except for the
1373          * reference counter protects page from concurrent reclaim.
1374          */
1375         lu_ref_del(&page->cp_reference, "transfer", page);
1376
1377         cl_page_completion(env, page, crt, rc);
1378         cl_page_put(env, page);
1379
1380         RETURN(0);
1381 }
1382
1383 #define OSC_DUMP_GRANT(mask, cli, fmt, args...) do {                    \
1384         struct client_obd *__tmp = (cli);                               \
1385         CDEBUG(mask, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu " \
1386                "dropped: %ld avail: %ld, dirty_grant: %ld, "            \
1387                "reserved: %ld, flight: %d } lru {in list: %ld, "        \
1388                "left: %ld, waiters: %d }" fmt "\n",                     \
1389                cli_name(__tmp),                                         \
1390                __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages,        \
1391                atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1392                __tmp->cl_lost_grant, __tmp->cl_avail_grant,             \
1393                __tmp->cl_dirty_grant,                                   \
1394                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,         \
1395                atomic_long_read(&__tmp->cl_lru_in_list),                \
1396                atomic_long_read(&__tmp->cl_lru_busy),                   \
1397                atomic_read(&__tmp->cl_lru_shrinkers), ##args);          \
1398 } while (0)
1399
1400 /* caller must hold loi_list_lock */
1401 static void osc_consume_write_grant(struct client_obd *cli,
1402                                     struct brw_page *pga)
1403 {
1404         assert_spin_locked(&cli->cl_loi_list_lock);
1405         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1406         atomic_long_inc(&obd_dirty_pages);
1407         cli->cl_dirty_pages++;
1408         pga->flag |= OBD_BRW_FROM_GRANT;
1409         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1410                PAGE_SIZE, pga, pga->pg);
1411         osc_update_next_shrink(cli);
1412 }
1413
1414 /* the companion to osc_consume_write_grant, called when a brw has completed.
1415  * must be called with the loi lock held. */
1416 static void osc_release_write_grant(struct client_obd *cli,
1417                                     struct brw_page *pga)
1418 {
1419         ENTRY;
1420
1421         assert_spin_locked(&cli->cl_loi_list_lock);
1422         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1423                 EXIT;
1424                 return;
1425         }
1426
1427         pga->flag &= ~OBD_BRW_FROM_GRANT;
1428         atomic_long_dec(&obd_dirty_pages);
1429         cli->cl_dirty_pages--;
1430         if (pga->flag & OBD_BRW_NOCACHE) {
1431                 pga->flag &= ~OBD_BRW_NOCACHE;
1432                 atomic_long_dec(&obd_dirty_transit_pages);
1433                 cli->cl_dirty_transit--;
1434         }
1435         EXIT;
1436 }
1437
1438 /**
1439  * To avoid sleeping with object lock held, it's good for us allocate enough
1440  * grants before entering into critical section.
1441  *
1442  * client_obd_list_lock held by caller
1443  */
1444 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1445 {
1446         int rc = -EDQUOT;
1447
1448         if (cli->cl_avail_grant >= bytes) {
1449                 cli->cl_avail_grant    -= bytes;
1450                 cli->cl_reserved_grant += bytes;
1451                 rc = 0;
1452         }
1453         return rc;
1454 }
1455
1456 static void __osc_unreserve_grant(struct client_obd *cli,
1457                                   unsigned int reserved, unsigned int unused)
1458 {
1459         /* it's quite normal for us to get more grant than reserved.
1460          * Thinking about a case that two extents merged by adding a new
1461          * chunk, we can save one extent tax. If extent tax is greater than
1462          * one chunk, we can save more grant by adding a new chunk */
1463         cli->cl_reserved_grant -= reserved;
1464         if (unused > reserved) {
1465                 cli->cl_avail_grant += reserved;
1466                 cli->cl_lost_grant  += unused - reserved;
1467                 cli->cl_dirty_grant -= unused - reserved;
1468         } else {
1469                 cli->cl_avail_grant += unused;
1470                 cli->cl_dirty_grant += reserved - unused;
1471         }
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(cli, reserved, unused);
1479         if (unused > 0)
1480                 osc_wake_cache_waiters(cli);
1481         spin_unlock(&cli->cl_loi_list_lock);
1482 }
1483
1484 /**
1485  * Free grant after IO is finished or canceled.
1486  *
1487  * @lost_grant is used to remember how many grants we have allocated but not
1488  * used, we should return these grants to OST. There're two cases where grants
1489  * can be lost:
1490  * 1. truncate;
1491  * 2. blocksize at OST is less than PAGE_SIZE and a partial page was
1492  *    written. In this case OST may use less chunks to serve this partial
1493  *    write. OSTs don't actually know the page size on the client side. so
1494  *    clients have to calculate lost grant by the blocksize on the OST.
1495  *    See filter_grant_check() for details.
1496  */
1497 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1498                            unsigned int lost_grant, unsigned int dirty_grant)
1499 {
1500         unsigned long grant;
1501
1502         grant = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
1503
1504         spin_lock(&cli->cl_loi_list_lock);
1505         atomic_long_sub(nr_pages, &obd_dirty_pages);
1506         cli->cl_dirty_pages -= nr_pages;
1507         cli->cl_lost_grant += lost_grant;
1508         cli->cl_dirty_grant -= dirty_grant;
1509         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1510                 /* borrow some grant from truncate to avoid the case that
1511                  * truncate uses up all avail grant */
1512                 cli->cl_lost_grant -= grant;
1513                 cli->cl_avail_grant += grant;
1514         }
1515         osc_wake_cache_waiters(cli);
1516         spin_unlock(&cli->cl_loi_list_lock);
1517         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu/%lu\n",
1518                lost_grant, cli->cl_lost_grant,
1519                cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_SHIFT,
1520                cli->cl_dirty_grant);
1521 }
1522
1523 /**
1524  * The companion to osc_enter_cache(), called when @oap is no longer part of
1525  * the dirty accounting due to error.
1526  */
1527 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1528 {
1529         spin_lock(&cli->cl_loi_list_lock);
1530         osc_release_write_grant(cli, &oap->oap_brw_page);
1531         spin_unlock(&cli->cl_loi_list_lock);
1532 }
1533
1534 /**
1535  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1536  * is available.
1537  */
1538 static int osc_enter_cache_try(struct client_obd *cli,
1539                                struct osc_async_page *oap,
1540                                int bytes, int transient)
1541 {
1542         int rc;
1543
1544         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1545
1546         rc = osc_reserve_grant(cli, bytes);
1547         if (rc < 0)
1548                 return 0;
1549
1550         if (cli->cl_dirty_pages < cli->cl_dirty_max_pages &&
1551             1 + atomic_long_read(&obd_dirty_pages) <= obd_max_dirty_pages) {
1552                 osc_consume_write_grant(cli, &oap->oap_brw_page);
1553                 if (transient) {
1554                         cli->cl_dirty_transit++;
1555                         atomic_long_inc(&obd_dirty_transit_pages);
1556                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
1557                 }
1558                 rc = 1;
1559         } else {
1560                 __osc_unreserve_grant(cli, bytes, bytes);
1561                 rc = 0;
1562         }
1563         return rc;
1564 }
1565
1566 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1567 {
1568         int rc;
1569         spin_lock(&cli->cl_loi_list_lock);
1570         rc = list_empty(&ocw->ocw_entry);
1571         spin_unlock(&cli->cl_loi_list_lock);
1572         return rc;
1573 }
1574
1575 /**
1576  * The main entry to reserve dirty page accounting. Usually the grant reserved
1577  * in this function will be freed in bulk in osc_free_grant() unless it fails
1578  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1579  *
1580  * The process will be put into sleep if it's already run out of grant.
1581  */
1582 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1583                            struct osc_async_page *oap, int bytes)
1584 {
1585         struct osc_object       *osc = oap->oap_obj;
1586         struct lov_oinfo        *loi = osc->oo_oinfo;
1587         struct osc_cache_waiter  ocw;
1588         struct l_wait_info       lwi;
1589         int                      rc = -EDQUOT;
1590         ENTRY;
1591
1592         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(AT_OFF ? obd_timeout : at_max),
1593                                NULL, LWI_ON_SIGNAL_NOOP, NULL);
1594
1595         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1596
1597         spin_lock(&cli->cl_loi_list_lock);
1598
1599         /* force the caller to try sync io.  this can jump the list
1600          * of queued writes and create a discontiguous rpc stream */
1601         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1602             cli->cl_dirty_max_pages == 0 ||
1603             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) {
1604                 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1605                 GOTO(out, rc = -EDQUOT);
1606         }
1607
1608         /* Hopefully normal case - cache space and write credits available */
1609         if (list_empty(&cli->cl_cache_waiters) &&
1610             osc_enter_cache_try(cli, oap, bytes, 0)) {
1611                 OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1612                 GOTO(out, rc = 0);
1613         }
1614
1615         /* We can get here for two reasons: too many dirty pages in cache, or
1616          * run out of grants. In both cases we should write dirty pages out.
1617          * Adding a cache waiter will trigger urgent write-out no matter what
1618          * RPC size will be.
1619          * The exiting condition is no avail grants and no dirty pages caching,
1620          * that really means there is no space on the OST. */
1621         init_waitqueue_head(&ocw.ocw_waitq);
1622         ocw.ocw_oap   = oap;
1623         ocw.ocw_grant = bytes;
1624         while (cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0) {
1625                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1626                 ocw.ocw_rc = 0;
1627                 spin_unlock(&cli->cl_loi_list_lock);
1628
1629                 osc_io_unplug_async(env, cli, NULL);
1630
1631                 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1632                        cli_name(cli), &ocw, oap);
1633
1634                 rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1635
1636                 spin_lock(&cli->cl_loi_list_lock);
1637
1638                 if (rc < 0) {
1639                         /* l_wait_event is interrupted by signal or timed out */
1640                         list_del_init(&ocw.ocw_entry);
1641                         break;
1642                 }
1643                 LASSERT(list_empty(&ocw.ocw_entry));
1644                 rc = ocw.ocw_rc;
1645
1646                 if (rc != -EDQUOT)
1647                         break;
1648                 if (osc_enter_cache_try(cli, oap, bytes, 0)) {
1649                         rc = 0;
1650                         break;
1651                 }
1652         }
1653
1654         switch (rc) {
1655         case 0:
1656                 OSC_DUMP_GRANT(D_CACHE, cli, "finally got grant space\n");
1657                 break;
1658         case -ETIMEDOUT:
1659                 OSC_DUMP_GRANT(D_CACHE, cli,
1660                                "timeout, fall back to sync i/o\n");
1661                 osc_extent_tree_dump(D_CACHE, osc);
1662                 /* fall back to synchronous I/O */
1663                 rc = -EDQUOT;
1664                 break;
1665         case -EINTR:
1666                 /* Ensures restartability - LU-3581 */
1667                 OSC_DUMP_GRANT(D_CACHE, cli, "interrupted\n");
1668                 rc = -ERESTARTSYS;
1669                 break;
1670         case -EDQUOT:
1671                 OSC_DUMP_GRANT(D_CACHE, cli,
1672                                "no grant space, fall back to sync i/o\n");
1673                 break;
1674         default:
1675                 CDEBUG(D_CACHE, "%s: event for cache space @ %p never arrived "
1676                        "due to %d, fall back to sync i/o\n",
1677                        cli_name(cli), &ocw, rc);
1678                 break;
1679         }
1680         EXIT;
1681 out:
1682         spin_unlock(&cli->cl_loi_list_lock);
1683         RETURN(rc);
1684 }
1685
1686 /* caller must hold loi_list_lock */
1687 void osc_wake_cache_waiters(struct client_obd *cli)
1688 {
1689         struct list_head *l, *tmp;
1690         struct osc_cache_waiter *ocw;
1691
1692         ENTRY;
1693         list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1694                 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
1695
1696                 ocw->ocw_rc = -EDQUOT;
1697
1698                 if (osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
1699                         ocw->ocw_rc = 0;
1700
1701                 if (ocw->ocw_rc == 0 ||
1702                     !(cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0)) {
1703                         list_del_init(&ocw->ocw_entry);
1704                         CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant "
1705                                "%ld, %d\n", ocw, ocw->ocw_oap,
1706                                cli->cl_avail_grant, ocw->ocw_rc);
1707
1708                         wake_up(&ocw->ocw_waitq);
1709                 }
1710         }
1711
1712         EXIT;
1713 }
1714 EXPORT_SYMBOL(osc_wake_cache_waiters);
1715
1716 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1717 {
1718         int hprpc = !!list_empty(&osc->oo_hp_exts);
1719         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1720 }
1721
1722 /* This maintains the lists of pending pages to read/write for a given object
1723  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1724  * to quickly find objects that are ready to send an RPC. */
1725 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1726                          int cmd)
1727 {
1728         int invalid_import = 0;
1729         ENTRY;
1730
1731         /* if we have an invalid import we want to drain the queued pages
1732          * by forcing them through rpcs that immediately fail and complete
1733          * the pages.  recovery relies on this to empty the queued pages
1734          * before canceling the locks and evicting down the llite pages */
1735         if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1736                 invalid_import = 1;
1737
1738         if (cmd & OBD_BRW_WRITE) {
1739                 if (atomic_read(&osc->oo_nr_writes) == 0)
1740                         RETURN(0);
1741                 if (invalid_import) {
1742                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1743                         RETURN(1);
1744                 }
1745                 if (!list_empty(&osc->oo_hp_exts)) {
1746                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1747                         RETURN(1);
1748                 }
1749                 if (!list_empty(&osc->oo_urgent_exts)) {
1750                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1751                         RETURN(1);
1752                 }
1753                 /* trigger a write rpc stream as long as there are dirtiers
1754                  * waiting for space.  as they're waiting, they're not going to
1755                  * create more pages to coalesce with what's waiting.. */
1756                 if (!list_empty(&cli->cl_cache_waiters)) {
1757                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1758                         RETURN(1);
1759                 }
1760                 if (!list_empty(&osc->oo_full_exts)) {
1761                         CDEBUG(D_CACHE, "full extent ready, make an RPC\n");
1762                         RETURN(1);
1763                 }
1764         } else {
1765                 if (atomic_read(&osc->oo_nr_reads) == 0)
1766                         RETURN(0);
1767                 if (invalid_import) {
1768                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1769                         RETURN(1);
1770                 }
1771                 /* all read are urgent. */
1772                 if (!list_empty(&osc->oo_reading_exts))
1773                         RETURN(1);
1774         }
1775
1776         RETURN(0);
1777 }
1778
1779 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1780 {
1781         struct client_obd *cli = osc_cli(obj);
1782         if (cmd & OBD_BRW_WRITE) {
1783                 atomic_add(delta, &obj->oo_nr_writes);
1784                 atomic_add(delta, &cli->cl_pending_w_pages);
1785                 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1786         } else {
1787                 atomic_add(delta, &obj->oo_nr_reads);
1788                 atomic_add(delta, &cli->cl_pending_r_pages);
1789                 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1790         }
1791         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1792 }
1793
1794 static int osc_makes_hprpc(struct osc_object *obj)
1795 {
1796         return !list_empty(&obj->oo_hp_exts);
1797 }
1798
1799 static void on_list(struct list_head *item, struct list_head *list,
1800                     int should_be_on)
1801 {
1802         if (list_empty(item) && should_be_on)
1803                 list_add_tail(item, list);
1804         else if (!list_empty(item) && !should_be_on)
1805                 list_del_init(item);
1806 }
1807
1808 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1809  * can find pages to build into rpcs quickly */
1810 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1811 {
1812         if (osc_makes_hprpc(osc)) {
1813                 /* HP rpc */
1814                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1815                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1816         } else {
1817                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1818                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1819                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1820                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1821         }
1822
1823         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1824                 atomic_read(&osc->oo_nr_writes) > 0);
1825
1826         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1827                 atomic_read(&osc->oo_nr_reads) > 0);
1828
1829         return osc_is_ready(osc);
1830 }
1831
1832 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1833 {
1834         int is_ready;
1835
1836         spin_lock(&cli->cl_loi_list_lock);
1837         is_ready = __osc_list_maint(cli, osc);
1838         spin_unlock(&cli->cl_loi_list_lock);
1839
1840         return is_ready;
1841 }
1842
1843 /* this is trying to propogate async writeback errors back up to the
1844  * application.  As an async write fails we record the error code for later if
1845  * the app does an fsync.  As long as errors persist we force future rpcs to be
1846  * sync so that the app can get a sync error and break the cycle of queueing
1847  * pages for which writeback will fail. */
1848 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1849                            int rc)
1850 {
1851         if (rc) {
1852                 if (!ar->ar_rc)
1853                         ar->ar_rc = rc;
1854
1855                 ar->ar_force_sync = 1;
1856                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1857                 return;
1858
1859         }
1860
1861         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1862                 ar->ar_force_sync = 0;
1863 }
1864
1865 /* this must be called holding the loi list lock to give coverage to exit_cache,
1866  * async_flag maintenance, and oap_request */
1867 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1868                               struct osc_async_page *oap, int sent, int rc)
1869 {
1870         struct osc_object *osc = oap->oap_obj;
1871         struct lov_oinfo  *loi = osc->oo_oinfo;
1872         __u64 xid = 0;
1873
1874         ENTRY;
1875         if (oap->oap_request != NULL) {
1876                 xid = ptlrpc_req_xid(oap->oap_request);
1877                 ptlrpc_req_finished(oap->oap_request);
1878                 oap->oap_request = NULL;
1879         }
1880
1881         /* As the transfer for this page is being done, clear the flags */
1882         spin_lock(&oap->oap_lock);
1883         oap->oap_async_flags = 0;
1884         spin_unlock(&oap->oap_lock);
1885         oap->oap_interrupted = 0;
1886
1887         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1888                 spin_lock(&cli->cl_loi_list_lock);
1889                 osc_process_ar(&cli->cl_ar, xid, rc);
1890                 osc_process_ar(&loi->loi_ar, xid, rc);
1891                 spin_unlock(&cli->cl_loi_list_lock);
1892         }
1893
1894         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1895         if (rc)
1896                 CERROR("completion on oap %p obj %p returns %d.\n",
1897                        oap, osc, rc);
1898
1899         EXIT;
1900 }
1901
1902 struct extent_rpc_data {
1903         struct list_head        *erd_rpc_list;
1904         unsigned int            erd_page_count;
1905         unsigned int            erd_max_pages;
1906         unsigned int            erd_max_chunks;
1907         unsigned int            erd_max_extents;
1908 };
1909
1910 static inline unsigned osc_extent_chunks(const struct osc_extent *ext)
1911 {
1912         struct client_obd *cli = osc_cli(ext->oe_obj);
1913         unsigned ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1914
1915         return (ext->oe_end >> ppc_bits) - (ext->oe_start >> ppc_bits) + 1;
1916 }
1917
1918 /**
1919  * Try to add extent to one RPC. We need to think about the following things:
1920  * - # of pages must not be over max_pages_per_rpc
1921  * - extent must be compatible with previous ones
1922  */
1923 static int try_to_add_extent_for_io(struct client_obd *cli,
1924                                     struct osc_extent *ext,
1925                                     struct extent_rpc_data *data)
1926 {
1927         struct osc_extent *tmp;
1928         unsigned int chunk_count;
1929         struct osc_async_page *oap = list_first_entry(&ext->oe_pages,
1930                                                       struct osc_async_page,
1931                                                       oap_pending_item);
1932         ENTRY;
1933
1934         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1935                 ext);
1936         OSC_EXTENT_DUMP(D_CACHE, ext, "trying to add this extent\n");
1937
1938         if (data->erd_max_extents == 0)
1939                 RETURN(0);
1940
1941         chunk_count = osc_extent_chunks(ext);
1942         EASSERTF(data->erd_page_count != 0 ||
1943                  chunk_count <= data->erd_max_chunks, ext,
1944                  "The first extent to be fit in a RPC contains %u chunks, "
1945                  "which is over the limit %u.\n", chunk_count,
1946                  data->erd_max_chunks);
1947         if (chunk_count > data->erd_max_chunks)
1948                 RETURN(0);
1949
1950         data->erd_max_pages = max(ext->oe_mppr, data->erd_max_pages);
1951         EASSERTF(data->erd_page_count != 0 ||
1952                 ext->oe_nr_pages <= data->erd_max_pages, ext,
1953                 "The first extent to be fit in a RPC contains %u pages, "
1954                 "which is over the limit %u.\n", ext->oe_nr_pages,
1955                 data->erd_max_pages);
1956         if (data->erd_page_count + ext->oe_nr_pages > data->erd_max_pages)
1957                 RETURN(0);
1958
1959         list_for_each_entry(tmp, data->erd_rpc_list, oe_link) {
1960                 struct osc_async_page *oap2;
1961                 oap2 = list_first_entry(&tmp->oe_pages, struct osc_async_page,
1962                                         oap_pending_item);
1963                 EASSERT(tmp->oe_owner == current, tmp);
1964 #if 0
1965                 if (overlapped(tmp, ext)) {
1966                         OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
1967                         EASSERT(0, ext);
1968                 }
1969 #endif
1970                 if (oap2cl_page(oap)->cp_type != oap2cl_page(oap2)->cp_type) {
1971                         CDEBUG(D_CACHE, "Do not permit different types of IO "
1972                                "in one RPC\n");
1973                         RETURN(0);
1974                 }
1975
1976                 if (tmp->oe_srvlock != ext->oe_srvlock ||
1977                     !tmp->oe_grants != !ext->oe_grants ||
1978                     tmp->oe_ndelay != ext->oe_ndelay ||
1979                     tmp->oe_no_merge || ext->oe_no_merge)
1980                         RETURN(0);
1981
1982                 /* remove break for strict check */
1983                 break;
1984         }
1985
1986         data->erd_max_extents--;
1987         data->erd_max_chunks -= chunk_count;
1988         data->erd_page_count += ext->oe_nr_pages;
1989         list_move_tail(&ext->oe_link, data->erd_rpc_list);
1990         ext->oe_owner = current;
1991         RETURN(1);
1992 }
1993
1994 /**
1995  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1996  * get_write_extent() takes all appropriate extents in atomic.
1997  *
1998  * The following policy is used to collect extents for IO:
1999  * 1. Add as many HP extents as possible;
2000  * 2. Add the first urgent extent in urgent extent list and take it out of
2001  *    urgent list;
2002  * 3. Add subsequent extents of this urgent extent;
2003  * 4. If urgent list is not empty, goto 2;
2004  * 5. Traverse the extent tree from the 1st extent;
2005  * 6. Above steps exit if there is no space in this RPC.
2006  */
2007 static unsigned int get_write_extents(struct osc_object *obj,
2008                                       struct list_head *rpclist)
2009 {
2010         struct client_obd *cli = osc_cli(obj);
2011         struct osc_extent *ext;
2012         struct extent_rpc_data data = {
2013                 .erd_rpc_list   = rpclist,
2014                 .erd_page_count = 0,
2015                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
2016                 .erd_max_chunks = osc_max_write_chunks(cli),
2017                 .erd_max_extents = 256,
2018         };
2019
2020         LASSERT(osc_object_is_locked(obj));
2021         while (!list_empty(&obj->oo_hp_exts)) {
2022                 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
2023                                  oe_link);
2024                 LASSERT(ext->oe_state == OES_CACHE);
2025                 if (!try_to_add_extent_for_io(cli, ext, &data))
2026                         return data.erd_page_count;
2027                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2028         }
2029         if (data.erd_page_count == data.erd_max_pages)
2030                 return data.erd_page_count;
2031
2032         while (!list_empty(&obj->oo_urgent_exts)) {
2033                 ext = list_entry(obj->oo_urgent_exts.next,
2034                                  struct osc_extent, oe_link);
2035                 if (!try_to_add_extent_for_io(cli, ext, &data))
2036                         return data.erd_page_count;
2037         }
2038         if (data.erd_page_count == data.erd_max_pages)
2039                 return data.erd_page_count;
2040
2041         /* One key difference between full extents and other extents: full
2042          * extents can usually only be added if the rpclist was empty, so if we
2043          * can't add one, we continue on to trying to add normal extents.  This
2044          * is so we don't miss adding extra extents to an RPC containing high
2045          * priority or urgent extents. */
2046         while (!list_empty(&obj->oo_full_exts)) {
2047                 ext = list_entry(obj->oo_full_exts.next,
2048                                  struct osc_extent, oe_link);
2049                 if (!try_to_add_extent_for_io(cli, ext, &data))
2050                         break;
2051         }
2052         if (data.erd_page_count == data.erd_max_pages)
2053                 return data.erd_page_count;
2054
2055         ext = first_extent(obj);
2056         while (ext != NULL) {
2057                 if ((ext->oe_state != OES_CACHE) ||
2058                     /* this extent may be already in current rpclist */
2059                     (!list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
2060                         ext = next_extent(ext);
2061                         continue;
2062                 }
2063
2064                 if (!try_to_add_extent_for_io(cli, ext, &data))
2065                         return data.erd_page_count;
2066
2067                 ext = next_extent(ext);
2068         }
2069         return data.erd_page_count;
2070 }
2071
2072 static int
2073 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
2074                    struct osc_object *osc)
2075 __must_hold(osc)
2076 {
2077         struct list_head   rpclist = LIST_HEAD_INIT(rpclist);
2078         struct osc_extent *ext;
2079         struct osc_extent *tmp;
2080         struct osc_extent *first = NULL;
2081         unsigned int page_count = 0;
2082         int srvlock = 0;
2083         int rc = 0;
2084         ENTRY;
2085
2086         LASSERT(osc_object_is_locked(osc));
2087
2088         page_count = get_write_extents(osc, &rpclist);
2089         LASSERT(equi(page_count == 0, list_empty(&rpclist)));
2090
2091         if (list_empty(&rpclist))
2092                 RETURN(0);
2093
2094         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2095
2096         list_for_each_entry(ext, &rpclist, oe_link) {
2097                 LASSERT(ext->oe_state == OES_CACHE ||
2098                         ext->oe_state == OES_LOCK_DONE);
2099                 if (ext->oe_state == OES_CACHE)
2100                         osc_extent_state_set(ext, OES_LOCKING);
2101                 else
2102                         osc_extent_state_set(ext, OES_RPC);
2103         }
2104
2105         /* we're going to grab page lock, so release object lock because
2106          * lock order is page lock -> object lock. */
2107         osc_object_unlock(osc);
2108
2109         list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2110                 if (ext->oe_state == OES_LOCKING) {
2111                         rc = osc_extent_make_ready(env, ext);
2112                         if (unlikely(rc < 0)) {
2113                                 list_del_init(&ext->oe_link);
2114                                 osc_extent_finish(env, ext, 0, rc);
2115                                 continue;
2116                         }
2117                 }
2118                 if (first == NULL) {
2119                         first = ext;
2120                         srvlock = ext->oe_srvlock;
2121                 } else {
2122                         LASSERT(srvlock == ext->oe_srvlock);
2123                 }
2124         }
2125
2126         if (!list_empty(&rpclist)) {
2127                 LASSERT(page_count > 0);
2128                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2129                 LASSERT(list_empty(&rpclist));
2130         }
2131
2132         osc_object_lock(osc);
2133         RETURN(rc);
2134 }
2135
2136 /**
2137  * prepare pages for ASYNC io and put pages in send queue.
2138  *
2139  * \param cmd OBD_BRW_* macroses
2140  * \param lop pending pages
2141  *
2142  * \return zero if no page added to send queue.
2143  * \return 1 if pages successfully added to send queue.
2144  * \return negative on errors.
2145  */
2146 static int
2147 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2148                   struct osc_object *osc)
2149 __must_hold(osc)
2150 {
2151         struct osc_extent *ext;
2152         struct osc_extent *next;
2153         struct list_head rpclist = LIST_HEAD_INIT(rpclist);
2154         struct extent_rpc_data data = {
2155                 .erd_rpc_list   = &rpclist,
2156                 .erd_page_count = 0,
2157                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
2158                 .erd_max_chunks = UINT_MAX,
2159                 .erd_max_extents = UINT_MAX,
2160         };
2161         int rc = 0;
2162         ENTRY;
2163
2164         LASSERT(osc_object_is_locked(osc));
2165         list_for_each_entry_safe(ext, next, &osc->oo_reading_exts, oe_link) {
2166                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2167                 if (!try_to_add_extent_for_io(cli, ext, &data))
2168                         break;
2169                 osc_extent_state_set(ext, OES_RPC);
2170                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2171         }
2172         LASSERT(data.erd_page_count <= data.erd_max_pages);
2173
2174         osc_update_pending(osc, OBD_BRW_READ, -data.erd_page_count);
2175
2176         if (!list_empty(&rpclist)) {
2177                 osc_object_unlock(osc);
2178
2179                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2180                 LASSERT(list_empty(&rpclist));
2181
2182                 osc_object_lock(osc);
2183         }
2184         RETURN(rc);
2185 }
2186
2187 #define list_to_obj(list, item) ({                                            \
2188         struct list_head *__tmp = (list)->next;                               \
2189         list_del_init(__tmp);                                         \
2190         list_entry(__tmp, struct osc_object, oo_##item);                      \
2191 })
2192
2193 /* This is called by osc_check_rpcs() to find which objects have pages that
2194  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2195 static struct osc_object *osc_next_obj(struct client_obd *cli)
2196 {
2197         ENTRY;
2198
2199         /* First return objects that have blocked locks so that they
2200          * will be flushed quickly and other clients can get the lock,
2201          * then objects which have pages ready to be stuffed into RPCs */
2202         if (!list_empty(&cli->cl_loi_hp_ready_list))
2203                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2204         if (!list_empty(&cli->cl_loi_ready_list))
2205                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2206
2207         /* then if we have cache waiters, return all objects with queued
2208          * writes.  This is especially important when many small files
2209          * have filled up the cache and not been fired into rpcs because
2210          * they don't pass the nr_pending/object threshhold */
2211         if (!list_empty(&cli->cl_cache_waiters) &&
2212             !list_empty(&cli->cl_loi_write_list))
2213                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2214
2215         /* then return all queued objects when we have an invalid import
2216          * so that they get flushed */
2217         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2218                 if (!list_empty(&cli->cl_loi_write_list))
2219                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2220                                            write_item));
2221                 if (!list_empty(&cli->cl_loi_read_list))
2222                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2223                                            read_item));
2224         }
2225         RETURN(NULL);
2226 }
2227
2228 /* called with the loi list lock held */
2229 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2230 __must_hold(&cli->cl_loi_list_lock)
2231 {
2232         struct osc_object *osc;
2233         int rc = 0;
2234         ENTRY;
2235
2236         while ((osc = osc_next_obj(cli)) != NULL) {
2237                 struct cl_object *obj = osc2cl(osc);
2238                 struct lu_ref_link link;
2239
2240                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2241
2242                 if (osc_max_rpc_in_flight(cli, osc)) {
2243                         __osc_list_maint(cli, osc);
2244                         break;
2245                 }
2246
2247                 cl_object_get(obj);
2248                 spin_unlock(&cli->cl_loi_list_lock);
2249                 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2250
2251                 /* attempt some read/write balancing by alternating between
2252                  * reads and writes in an object.  The makes_rpc checks here
2253                  * would be redundant if we were getting read/write work items
2254                  * instead of objects.  we don't want send_oap_rpc to drain a
2255                  * partial read pending queue when we're given this object to
2256                  * do io on writes while there are cache waiters */
2257                 osc_object_lock(osc);
2258                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2259                         rc = osc_send_write_rpc(env, cli, osc);
2260                         if (rc < 0) {
2261                                 CERROR("Write request failed with %d\n", rc);
2262
2263                                 /* osc_send_write_rpc failed, mostly because of
2264                                  * memory pressure.
2265                                  *
2266                                  * It can't break here, because if:
2267                                  *  - a page was submitted by osc_io_submit, so
2268                                  *    page locked;
2269                                  *  - no request in flight
2270                                  *  - no subsequent request
2271                                  * The system will be in live-lock state,
2272                                  * because there is no chance to call
2273                                  * osc_io_unplug() and osc_check_rpcs() any
2274                                  * more. pdflush can't help in this case,
2275                                  * because it might be blocked at grabbing
2276                                  * the page lock as we mentioned.
2277                                  *
2278                                  * Anyway, continue to drain pages. */
2279                                 /* break; */
2280                         }
2281                 }
2282                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2283                         rc = osc_send_read_rpc(env, cli, osc);
2284                         if (rc < 0)
2285                                 CERROR("Read request failed with %d\n", rc);
2286                 }
2287                 osc_object_unlock(osc);
2288
2289                 osc_list_maint(cli, osc);
2290                 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2291                 cl_object_put(env, obj);
2292
2293                 spin_lock(&cli->cl_loi_list_lock);
2294         }
2295 }
2296
2297 int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2298                    struct osc_object *osc, int async)
2299 {
2300         int rc = 0;
2301
2302         if (osc != NULL && osc_list_maint(cli, osc) == 0)
2303                 return 0;
2304
2305         if (!async) {
2306                 spin_lock(&cli->cl_loi_list_lock);
2307                 osc_check_rpcs(env, cli);
2308                 spin_unlock(&cli->cl_loi_list_lock);
2309         } else {
2310                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2311                 LASSERT(cli->cl_writeback_work != NULL);
2312                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2313         }
2314         return rc;
2315 }
2316 EXPORT_SYMBOL(osc_io_unplug0);
2317
2318 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2319                         struct page *page, loff_t offset)
2320 {
2321         struct obd_export     *exp = osc_export(osc);
2322         struct osc_async_page *oap = &ops->ops_oap;
2323         ENTRY;
2324
2325         if (!page)
2326                 return cfs_size_round(sizeof(*oap));
2327
2328         oap->oap_magic = OAP_MAGIC;
2329         oap->oap_cli = &exp->exp_obd->u.cli;
2330         oap->oap_obj = osc;
2331
2332         oap->oap_page = page;
2333         oap->oap_obj_off = offset;
2334         LASSERT(!(offset & ~PAGE_MASK));
2335
2336         if (cfs_capable(CFS_CAP_SYS_RESOURCE))
2337                 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2338
2339         INIT_LIST_HEAD(&oap->oap_pending_item);
2340         INIT_LIST_HEAD(&oap->oap_rpc_item);
2341
2342         spin_lock_init(&oap->oap_lock);
2343         CDEBUG(D_INFO, "oap %p page %p obj off %llu\n",
2344                oap, page, oap->oap_obj_off);
2345         RETURN(0);
2346 }
2347 EXPORT_SYMBOL(osc_prep_async_page);
2348
2349 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2350                        struct osc_page *ops)
2351 {
2352         struct osc_io *oio = osc_env_io(env);
2353         struct osc_extent     *ext = NULL;
2354         struct osc_async_page *oap = &ops->ops_oap;
2355         struct client_obd     *cli = oap->oap_cli;
2356         struct osc_object     *osc = oap->oap_obj;
2357         pgoff_t index;
2358         unsigned int tmp;
2359         unsigned int grants = 0;
2360         u32    brw_flags = OBD_BRW_ASYNC;
2361         int    cmd = OBD_BRW_WRITE;
2362         int    need_release = 0;
2363         int    rc = 0;
2364         ENTRY;
2365
2366         if (oap->oap_magic != OAP_MAGIC)
2367                 RETURN(-EINVAL);
2368
2369         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2370                 RETURN(-EIO);
2371
2372         if (!list_empty(&oap->oap_pending_item) ||
2373             !list_empty(&oap->oap_rpc_item))
2374                 RETURN(-EBUSY);
2375
2376         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2377         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2378         if (cfs_capable(CFS_CAP_SYS_RESOURCE)) {
2379                 brw_flags |= OBD_BRW_NOQUOTA;
2380                 cmd |= OBD_BRW_NOQUOTA;
2381         }
2382
2383         /* check if the file's owner/group is over quota */
2384         if (!(cmd & OBD_BRW_NOQUOTA)) {
2385                 struct cl_object *obj;
2386                 struct cl_attr   *attr;
2387                 unsigned int qid[LL_MAXQUOTAS];
2388
2389                 obj = cl_object_top(&osc->oo_cl);
2390                 attr = &osc_env_info(env)->oti_attr;
2391
2392                 cl_object_attr_lock(obj);
2393                 rc = cl_object_attr_get(env, obj, attr);
2394                 cl_object_attr_unlock(obj);
2395
2396                 qid[USRQUOTA] = attr->cat_uid;
2397                 qid[GRPQUOTA] = attr->cat_gid;
2398                 qid[PRJQUOTA] = attr->cat_projid;
2399                 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2400                         rc = -EDQUOT;
2401                 if (rc)
2402                         RETURN(rc);
2403         }
2404
2405         oap->oap_cmd = cmd;
2406         oap->oap_page_off = ops->ops_from;
2407         oap->oap_count = ops->ops_to - ops->ops_from;
2408         /* No need to hold a lock here,
2409          * since this page is not in any list yet. */
2410         oap->oap_async_flags = 0;
2411         oap->oap_brw_flags = brw_flags;
2412
2413         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2414                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2415
2416         index = osc_index(oap2osc(oap));
2417
2418         /* Add this page into extent by the following steps:
2419          * 1. if there exists an active extent for this IO, mostly this page
2420          *    can be added to the active extent and sometimes we need to
2421          *    expand extent to accomodate this page;
2422          * 2. otherwise, a new extent will be allocated. */
2423
2424         ext = oio->oi_active;
2425         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2426                 /* one chunk plus extent overhead must be enough to write this
2427                  * page */
2428                 grants = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2429                 if (ext->oe_end >= index)
2430                         grants = 0;
2431
2432                 /* it doesn't need any grant to dirty this page */
2433                 spin_lock(&cli->cl_loi_list_lock);
2434                 rc = osc_enter_cache_try(cli, oap, grants, 0);
2435                 spin_unlock(&cli->cl_loi_list_lock);
2436                 if (rc == 0) { /* try failed */
2437                         grants = 0;
2438                         need_release = 1;
2439                 } else if (ext->oe_end < index) {
2440                         tmp = grants;
2441                         /* try to expand this extent */
2442                         rc = osc_extent_expand(ext, index, &tmp);
2443                         if (rc < 0) {
2444                                 need_release = 1;
2445                                 /* don't free reserved grant */
2446                         } else {
2447                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2448                                                 "expanded for %lu.\n", index);
2449                                 osc_unreserve_grant(cli, grants, tmp);
2450                                 grants = 0;
2451                         }
2452                 }
2453                 rc = 0;
2454         } else if (ext != NULL) {
2455                 /* index is located outside of active extent */
2456                 need_release = 1;
2457         }
2458         if (need_release) {
2459                 osc_extent_release(env, ext);
2460                 oio->oi_active = NULL;
2461                 ext = NULL;
2462         }
2463
2464         if (ext == NULL) {
2465                 tmp = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2466
2467                 /* try to find new extent to cover this page */
2468                 LASSERT(oio->oi_active == NULL);
2469                 /* we may have allocated grant for this page if we failed
2470                  * to expand the previous active extent. */
2471                 LASSERT(ergo(grants > 0, grants >= tmp));
2472
2473                 rc = 0;
2474                 if (grants == 0) {
2475                         /* we haven't allocated grant for this page. */
2476                         rc = osc_enter_cache(env, cli, oap, tmp);
2477                         if (rc == 0)
2478                                 grants = tmp;
2479                 }
2480
2481                 tmp = grants;
2482                 if (rc == 0) {
2483                         ext = osc_extent_find(env, osc, index, &tmp);
2484                         if (IS_ERR(ext)) {
2485                                 LASSERT(tmp == grants);
2486                                 osc_exit_cache(cli, oap);
2487                                 rc = PTR_ERR(ext);
2488                                 ext = NULL;
2489                         } else {
2490                                 oio->oi_active = ext;
2491                         }
2492                 }
2493                 if (grants > 0)
2494                         osc_unreserve_grant(cli, grants, tmp);
2495         }
2496
2497         LASSERT(ergo(rc == 0, ext != NULL));
2498         if (ext != NULL) {
2499                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2500                          ext, "index = %lu.\n", index);
2501                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2502
2503                 osc_object_lock(osc);
2504                 if (ext->oe_nr_pages == 0)
2505                         ext->oe_srvlock = ops->ops_srvlock;
2506                 else
2507                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2508                 ++ext->oe_nr_pages;
2509                 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2510                 osc_object_unlock(osc);
2511
2512                 if (!ext->oe_layout_version)
2513                         ext->oe_layout_version = io->ci_layout_version;
2514         }
2515
2516         RETURN(rc);
2517 }
2518
2519 int osc_teardown_async_page(const struct lu_env *env,
2520                             struct osc_object *obj, struct osc_page *ops)
2521 {
2522         struct osc_async_page *oap = &ops->ops_oap;
2523         int rc = 0;
2524         ENTRY;
2525
2526         LASSERT(oap->oap_magic == OAP_MAGIC);
2527
2528         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2529                oap, ops, osc_index(oap2osc(oap)));
2530
2531         if (!list_empty(&oap->oap_rpc_item)) {
2532                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2533                 rc = -EBUSY;
2534         } else if (!list_empty(&oap->oap_pending_item)) {
2535                 struct osc_extent *ext = NULL;
2536
2537                 osc_object_lock(obj);
2538                 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2539                 osc_object_unlock(obj);
2540                 /* only truncated pages are allowed to be taken out.
2541                  * See osc_extent_truncate() and osc_cache_truncate_start()
2542                  * for details. */
2543                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2544                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2545                                         osc_index(oap2osc(oap)));
2546                         rc = -EBUSY;
2547                 }
2548                 if (ext != NULL)
2549                         osc_extent_put(env, ext);
2550         }
2551         RETURN(rc);
2552 }
2553
2554 /**
2555  * This is called when a page is picked up by kernel to write out.
2556  *
2557  * We should find out the corresponding extent and add the whole extent
2558  * into urgent list. The extent may be being truncated or used, handle it
2559  * carefully.
2560  */
2561 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2562                          struct osc_page *ops)
2563 {
2564         struct osc_extent *ext   = NULL;
2565         struct osc_object *obj   = cl2osc(ops->ops_cl.cpl_obj);
2566         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2567         pgoff_t            index = osc_index(ops);
2568         struct osc_async_page *oap = &ops->ops_oap;
2569         bool unplug = false;
2570         int rc = 0;
2571         ENTRY;
2572
2573         osc_object_lock(obj);
2574         ext = osc_extent_lookup(obj, index);
2575         if (ext == NULL) {
2576                 osc_extent_tree_dump(D_ERROR, obj);
2577                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2578         }
2579
2580         switch (ext->oe_state) {
2581         case OES_RPC:
2582         case OES_LOCK_DONE:
2583                 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2584                 LASSERT(0);
2585                 break;
2586         case OES_LOCKING:
2587                 /* If we know this extent is being written out, we should abort
2588                  * so that the writer can make this page ready. Otherwise, there
2589                  * exists a deadlock problem because other process can wait for
2590                  * page writeback bit holding page lock; and meanwhile in
2591                  * vvp_page_make_ready(), we need to grab page lock before
2592                  * really sending the RPC. */
2593         case OES_TRUNC:
2594                 /* race with truncate, page will be redirtied */
2595         case OES_ACTIVE:
2596                 /* The extent is active so we need to abort and let the caller
2597                  * re-dirty the page. If we continued on here, and we were the
2598                  * one making the extent active, we could deadlock waiting for
2599                  * the page writeback to clear but it won't because the extent
2600                  * is active and won't be written out. */
2601                 GOTO(out, rc = -EAGAIN);
2602         default:
2603                 break;
2604         }
2605
2606         rc = cl_page_prep(env, io, cp, CRT_WRITE);
2607         if (rc)
2608                 GOTO(out, rc);
2609
2610         spin_lock(&oap->oap_lock);
2611         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2612         spin_unlock(&oap->oap_lock);
2613
2614         if (memory_pressure_get())
2615                 ext->oe_memalloc = 1;
2616
2617         ext->oe_urgent = 1;
2618         if (ext->oe_state == OES_CACHE) {
2619                 OSC_EXTENT_DUMP(D_CACHE, ext,
2620                                 "flush page %p make it urgent.\n", oap);
2621                 if (list_empty(&ext->oe_link))
2622                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2623                 unplug = true;
2624         }
2625         rc = 0;
2626         EXIT;
2627
2628 out:
2629         osc_object_unlock(obj);
2630         osc_extent_put(env, ext);
2631         if (unplug)
2632                 osc_io_unplug_async(env, osc_cli(obj), obj);
2633         return rc;
2634 }
2635
2636 /**
2637  * this is called when a sync waiter receives an interruption.  Its job is to
2638  * get the caller woken as soon as possible.  If its page hasn't been put in an
2639  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2640  * desiring interruption which will forcefully complete the rpc once the rpc
2641  * has timed out.
2642  */
2643 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2644 {
2645         struct osc_async_page *oap = &ops->ops_oap;
2646         struct osc_object     *obj = oap->oap_obj;
2647         struct client_obd     *cli = osc_cli(obj);
2648         struct osc_extent     *ext;
2649         struct osc_extent     *found = NULL;
2650         struct list_head            *plist;
2651         pgoff_t index = osc_index(ops);
2652         int     rc = -EBUSY;
2653         int     cmd;
2654         ENTRY;
2655
2656         LASSERT(!oap->oap_interrupted);
2657         oap->oap_interrupted = 1;
2658
2659         /* Find out the caching extent */
2660         osc_object_lock(obj);
2661         if (oap->oap_cmd & OBD_BRW_WRITE) {
2662                 plist = &obj->oo_urgent_exts;
2663                 cmd   = OBD_BRW_WRITE;
2664         } else {
2665                 plist = &obj->oo_reading_exts;
2666                 cmd   = OBD_BRW_READ;
2667         }
2668         list_for_each_entry(ext, plist, oe_link) {
2669                 if (ext->oe_start <= index && ext->oe_end >= index) {
2670                         LASSERT(ext->oe_state == OES_LOCK_DONE);
2671                         /* For OES_LOCK_DONE state extent, it has already held
2672                          * a refcount for RPC. */
2673                         found = osc_extent_get(ext);
2674                         break;
2675                 }
2676         }
2677         if (found != NULL) {
2678                 list_del_init(&found->oe_link);
2679                 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2680                 osc_object_unlock(obj);
2681
2682                 osc_extent_finish(env, found, 0, -EINTR);
2683                 osc_extent_put(env, found);
2684                 rc = 0;
2685         } else {
2686                 osc_object_unlock(obj);
2687                 /* ok, it's been put in an rpc. only one oap gets a request
2688                  * reference */
2689                 if (oap->oap_request != NULL) {
2690                         ptlrpc_mark_interrupted(oap->oap_request);
2691                         ptlrpcd_wake(oap->oap_request);
2692                         ptlrpc_req_finished(oap->oap_request);
2693                         oap->oap_request = NULL;
2694                 }
2695         }
2696
2697         osc_list_maint(cli, obj);
2698         RETURN(rc);
2699 }
2700
2701 int osc_queue_sync_pages(const struct lu_env *env, const struct cl_io *io,
2702                          struct osc_object *obj, struct list_head *list,
2703                          int brw_flags)
2704 {
2705         struct client_obd     *cli = osc_cli(obj);
2706         struct osc_extent     *ext;
2707         struct osc_async_page *oap;
2708         int     page_count = 0;
2709         int     mppr       = cli->cl_max_pages_per_rpc;
2710         bool    can_merge   = true;
2711         pgoff_t start      = CL_PAGE_EOF;
2712         pgoff_t end        = 0;
2713         ENTRY;
2714
2715         list_for_each_entry(oap, list, oap_pending_item) {
2716                 struct osc_page *opg = oap2osc_page(oap);
2717                 pgoff_t index = osc_index(opg);
2718
2719                 if (index > end)
2720                         end = index;
2721                 if (index < start)
2722                         start = index;
2723                 ++page_count;
2724                 mppr <<= (page_count > mppr);
2725
2726                 if (unlikely(opg->ops_from > 0 || opg->ops_to < PAGE_SIZE))
2727                         can_merge = false;
2728         }
2729
2730         ext = osc_extent_alloc(obj);
2731         if (ext == NULL) {
2732                 struct osc_async_page *tmp;
2733
2734                 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2735                         list_del_init(&oap->oap_pending_item);
2736                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2737                 }
2738                 RETURN(-ENOMEM);
2739         }
2740
2741         ext->oe_rw = !!(brw_flags & OBD_BRW_READ);
2742         ext->oe_sync = 1;
2743         ext->oe_no_merge = !can_merge;
2744         ext->oe_urgent = 1;
2745         ext->oe_start = start;
2746         ext->oe_end = ext->oe_max_end = end;
2747         ext->oe_obj = obj;
2748         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2749         ext->oe_ndelay = !!(brw_flags & OBD_BRW_NDELAY);
2750         ext->oe_nr_pages = page_count;
2751         ext->oe_mppr = mppr;
2752         list_splice_init(list, &ext->oe_pages);
2753         ext->oe_layout_version = io->ci_layout_version;
2754
2755         osc_object_lock(obj);
2756         /* Reuse the initial refcount for RPC, don't drop it */
2757         osc_extent_state_set(ext, OES_LOCK_DONE);
2758         if (!ext->oe_rw) { /* write */
2759                 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2760                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2761         } else {
2762                 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2763                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2764         }
2765         osc_object_unlock(obj);
2766
2767         osc_io_unplug_async(env, cli, obj);
2768         RETURN(0);
2769 }
2770
2771 /**
2772  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2773  */
2774 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
2775                              __u64 size, struct osc_extent **extp)
2776 {
2777         struct client_obd *cli = osc_cli(obj);
2778         struct osc_extent *ext;
2779         struct osc_extent *waiting = NULL;
2780         pgoff_t index;
2781         struct list_head list = LIST_HEAD_INIT(list);
2782         int result = 0;
2783         bool partial;
2784         ENTRY;
2785
2786         /* pages with index greater or equal to index will be truncated. */
2787         index = cl_index(osc2cl(obj), size);
2788         partial = size > cl_offset(osc2cl(obj), index);
2789
2790 again:
2791         osc_object_lock(obj);
2792         ext = osc_extent_search(obj, index);
2793         if (ext == NULL)
2794                 ext = first_extent(obj);
2795         else if (ext->oe_end < index)
2796                 ext = next_extent(ext);
2797         while (ext != NULL) {
2798                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2799
2800                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2801                         /* if ext is in urgent state, it means there must exist
2802                          * a page already having been flushed by write_page().
2803                          * We have to wait for this extent because we can't
2804                          * truncate that page. */
2805                         OSC_EXTENT_DUMP(D_CACHE, ext,
2806                                         "waiting for busy extent\n");
2807                         waiting = osc_extent_get(ext);
2808                         break;
2809                 }
2810
2811                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:%llu.\n", size);
2812
2813                 osc_extent_get(ext);
2814                 if (ext->oe_state == OES_ACTIVE) {
2815                         /* though we grab inode mutex for write path, but we
2816                          * release it before releasing extent(in osc_io_end()),
2817                          * so there is a race window that an extent is still
2818                          * in OES_ACTIVE when truncate starts. */
2819                         LASSERT(!ext->oe_trunc_pending);
2820                         ext->oe_trunc_pending = 1;
2821                 } else {
2822                         EASSERT(ext->oe_state == OES_CACHE, ext);
2823                         osc_extent_state_set(ext, OES_TRUNC);
2824                         osc_update_pending(obj, OBD_BRW_WRITE,
2825                                            -ext->oe_nr_pages);
2826                 }
2827                 /* This extent could be on the full extents list, that's OK */
2828                 EASSERT(!ext->oe_hp && !ext->oe_urgent, ext);
2829                 if (!list_empty(&ext->oe_link))
2830                         list_move_tail(&ext->oe_link, &list);
2831                 else
2832                         list_add_tail(&ext->oe_link, &list);
2833
2834                 ext = next_extent(ext);
2835         }
2836         osc_object_unlock(obj);
2837
2838         osc_list_maint(cli, obj);
2839
2840         while (!list_empty(&list)) {
2841                 int rc;
2842
2843                 ext = list_entry(list.next, struct osc_extent, oe_link);
2844                 list_del_init(&ext->oe_link);
2845
2846                 /* extent may be in OES_ACTIVE state because inode mutex
2847                  * is released before osc_io_end() in file write case */
2848                 if (ext->oe_state != OES_TRUNC)
2849                         osc_extent_wait(env, ext, OES_TRUNC);
2850
2851                 rc = osc_extent_truncate(ext, index, partial);
2852                 if (rc < 0) {
2853                         if (result == 0)
2854                                 result = rc;
2855
2856                         OSC_EXTENT_DUMP(D_ERROR, ext,
2857                                         "truncate error %d\n", rc);
2858                 } else if (ext->oe_nr_pages == 0) {
2859                         osc_extent_remove(ext);
2860                 } else {
2861                         /* this must be an overlapped extent which means only
2862                          * part of pages in this extent have been truncated.
2863                          */
2864                         EASSERTF(ext->oe_start <= index, ext,
2865                                  "trunc index = %lu/%d.\n", index, partial);
2866                         /* fix index to skip this partially truncated extent */
2867                         index = ext->oe_end + 1;
2868                         partial = false;
2869
2870                         /* we need to hold this extent in OES_TRUNC state so
2871                          * that no writeback will happen. This is to avoid
2872                          * BUG 17397.
2873                          * Only partial truncate can reach here, if @size is
2874                          * not zero, the caller should provide a valid @extp. */
2875                         LASSERT(*extp == NULL);
2876                         *extp = osc_extent_get(ext);
2877                         OSC_EXTENT_DUMP(D_CACHE, ext,
2878                                         "trunc at %llu\n", size);
2879                 }
2880                 osc_extent_put(env, ext);
2881         }
2882         if (waiting != NULL) {
2883                 int rc;
2884
2885                 /* ignore the result of osc_extent_wait the write initiator
2886                  * should take care of it. */
2887                 rc = osc_extent_wait(env, waiting, OES_INV);
2888                 if (rc < 0)
2889                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2890
2891                 osc_extent_put(env, waiting);
2892                 waiting = NULL;
2893                 goto again;
2894         }
2895         RETURN(result);
2896 }
2897 EXPORT_SYMBOL(osc_cache_truncate_start);
2898
2899 /**
2900  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2901  */
2902 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext)
2903 {
2904         if (ext != NULL) {
2905                 struct osc_object *obj = ext->oe_obj;
2906                 bool unplug = false;
2907
2908                 EASSERT(ext->oe_nr_pages > 0, ext);
2909                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2910                 EASSERT(!ext->oe_urgent, ext);
2911
2912                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2913                 osc_object_lock(obj);
2914                 osc_extent_state_set(ext, OES_CACHE);
2915                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2916                         ext->oe_urgent = 1;
2917                         list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2918                         unplug = true;
2919                 }
2920                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2921                 osc_object_unlock(obj);
2922                 osc_extent_put(env, ext);
2923
2924                 if (unplug)
2925                         osc_io_unplug_async(env, osc_cli(obj), obj);
2926         }
2927 }
2928
2929 /**
2930  * Wait for extents in a specific range to be written out.
2931  * The caller must have called osc_cache_writeback_range() to issue IO
2932  * otherwise it will take a long time for this function to finish.
2933  *
2934  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2935  * nobody else can dirty this range of file while we're waiting for
2936  * extents to be written.
2937  */
2938 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2939                          pgoff_t start, pgoff_t end)
2940 {
2941         struct osc_extent *ext;
2942         pgoff_t index = start;
2943         int     result = 0;
2944         ENTRY;
2945
2946 again:
2947         osc_object_lock(obj);
2948         ext = osc_extent_search(obj, index);
2949         if (ext == NULL)
2950                 ext = first_extent(obj);
2951         else if (ext->oe_end < index)
2952                 ext = next_extent(ext);
2953         while (ext != NULL) {
2954                 int rc;
2955
2956                 if (ext->oe_start > end)
2957                         break;
2958
2959                 if (!ext->oe_fsync_wait) {
2960                         ext = next_extent(ext);
2961                         continue;
2962                 }
2963
2964                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2965                              ext->oe_hp || ext->oe_urgent), ext);
2966                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2967                              !ext->oe_hp && ext->oe_urgent), ext);
2968
2969                 index = ext->oe_end + 1;
2970                 osc_extent_get(ext);
2971                 osc_object_unlock(obj);
2972
2973                 rc = osc_extent_wait(env, ext, OES_INV);
2974                 if (result == 0)
2975                         result = rc;
2976                 osc_extent_put(env, ext);
2977                 goto again;
2978         }
2979         osc_object_unlock(obj);
2980
2981         OSC_IO_DEBUG(obj, "sync file range.\n");
2982         RETURN(result);
2983 }
2984 EXPORT_SYMBOL(osc_cache_wait_range);
2985
2986 /**
2987  * Called to write out a range of osc object.
2988  *
2989  * @hp     : should be set this is caused by lock cancel;
2990  * @discard: is set if dirty pages should be dropped - file will be deleted or
2991  *         truncated, this implies there is no partially discarding extents.
2992  *
2993  * Return how many pages will be issued, or error code if error occurred.
2994  */
2995 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2996                               pgoff_t start, pgoff_t end, int hp, int discard)
2997 {
2998         struct osc_extent *ext;
2999         struct list_head discard_list = LIST_HEAD_INIT(discard_list);
3000         bool unplug = false;
3001         int result = 0;
3002         ENTRY;
3003
3004         osc_object_lock(obj);
3005         ext = osc_extent_search(obj, start);
3006         if (ext == NULL)
3007                 ext = first_extent(obj);
3008         else if (ext->oe_end < start)
3009                 ext = next_extent(ext);
3010         while (ext != NULL) {
3011                 if (ext->oe_start > end)
3012                         break;
3013
3014                 ext->oe_fsync_wait = 1;
3015                 switch (ext->oe_state) {
3016                 case OES_CACHE:
3017                         result += ext->oe_nr_pages;
3018                         if (!discard) {
3019                                 struct list_head *list = NULL;
3020                                 if (hp) {
3021                                         EASSERT(!ext->oe_hp, ext);
3022                                         ext->oe_hp = 1;
3023                                         list = &obj->oo_hp_exts;
3024                                 } else if (!ext->oe_urgent) {
3025                                         ext->oe_urgent = 1;
3026                                         list = &obj->oo_urgent_exts;
3027                                 }
3028                                 if (list != NULL)
3029                                         list_move_tail(&ext->oe_link, list);
3030                                 unplug = true;
3031                         } else {
3032                                 /* the only discarder is lock cancelling, so
3033                                  * [start, end] must contain this extent */
3034                                 EASSERT(ext->oe_start >= start &&
3035                                         ext->oe_max_end <= end, ext);
3036                                 osc_extent_state_set(ext, OES_LOCKING);
3037                                 ext->oe_owner = current;
3038                                 list_move_tail(&ext->oe_link,
3039                                                    &discard_list);
3040                                 osc_update_pending(obj, OBD_BRW_WRITE,
3041                                                    -ext->oe_nr_pages);
3042                         }
3043                         break;
3044                 case OES_ACTIVE:
3045                         /* It's pretty bad to wait for ACTIVE extents, because
3046                          * we don't know how long we will wait for it to be
3047                          * flushed since it may be blocked at awaiting more
3048                          * grants. We do this for the correctness of fsync. */
3049                         LASSERT(hp == 0 && discard == 0);
3050                         ext->oe_urgent = 1;
3051                         break;
3052                 case OES_TRUNC:
3053                         /* this extent is being truncated, can't do anything
3054                          * for it now. it will be set to urgent after truncate
3055                          * is finished in osc_cache_truncate_end(). */
3056                 default:
3057                         break;
3058                 }
3059                 ext = next_extent(ext);
3060         }
3061         osc_object_unlock(obj);
3062
3063         LASSERT(ergo(!discard, list_empty(&discard_list)));
3064         if (!list_empty(&discard_list)) {
3065                 struct osc_extent *tmp;
3066                 int rc;
3067
3068                 osc_list_maint(osc_cli(obj), obj);
3069                 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
3070                         list_del_init(&ext->oe_link);
3071                         EASSERT(ext->oe_state == OES_LOCKING, ext);
3072
3073                         /* Discard caching pages. We don't actually write this
3074                          * extent out but we complete it as if we did. */
3075                         rc = osc_extent_make_ready(env, ext);
3076                         if (unlikely(rc < 0)) {
3077                                 OSC_EXTENT_DUMP(D_ERROR, ext,
3078                                                 "make_ready returned %d\n", rc);
3079                                 if (result >= 0)
3080                                         result = rc;
3081                         }
3082
3083                         /* finish the extent as if the pages were sent */
3084                         osc_extent_finish(env, ext, 0, 0);
3085                 }
3086         }
3087
3088         if (unplug)
3089                 osc_io_unplug(env, osc_cli(obj), obj);
3090
3091         if (hp || discard) {
3092                 int rc;
3093                 rc = osc_cache_wait_range(env, obj, start, end);
3094                 if (result >= 0 && rc < 0)
3095                         result = rc;
3096         }
3097
3098         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3099         RETURN(result);
3100 }
3101 EXPORT_SYMBOL(osc_cache_writeback_range);
3102
3103 /**
3104  * Returns a list of pages by a given [start, end] of \a obj.
3105  *
3106  * \param resched If not NULL, then we give up before hogging CPU for too
3107  * long and set *resched = 1, in that case caller should implement a retry
3108  * logic.
3109  *
3110  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3111  * crucial in the face of [offset, EOF] locks.
3112  *
3113  * Return at least one page in @queue unless there is no covered page.
3114  */
3115 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3116                         struct osc_object *osc, pgoff_t start, pgoff_t end,
3117                         osc_page_gang_cbt cb, void *cbdata)
3118 {
3119         struct osc_page *ops;
3120         struct pagevec  *pagevec;
3121         void            **pvec;
3122         pgoff_t         idx;
3123         unsigned int    nr;
3124         unsigned int    i;
3125         unsigned int    j;
3126         int             res = CLP_GANG_OKAY;
3127         bool            tree_lock = true;
3128         ENTRY;
3129
3130         idx = start;
3131         pvec = osc_env_info(env)->oti_pvec;
3132         pagevec = &osc_env_info(env)->oti_pagevec;
3133         ll_pagevec_init(pagevec, 0);
3134         spin_lock(&osc->oo_tree_lock);
3135         while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3136                                             idx, OTI_PVEC_SIZE)) > 0) {
3137                 struct cl_page *page;
3138                 bool end_of_region = false;
3139
3140                 for (i = 0, j = 0; i < nr; ++i) {
3141                         ops = pvec[i];
3142                         pvec[i] = NULL;
3143
3144                         idx = osc_index(ops);
3145                         if (idx > end) {
3146                                 end_of_region = true;
3147                                 break;
3148                         }
3149
3150                         page = ops->ops_cl.cpl_page;
3151                         LASSERT(page->cp_type == CPT_CACHEABLE);
3152                         if (page->cp_state == CPS_FREEING)
3153                                 continue;
3154
3155                         cl_page_get(page);
3156                         lu_ref_add_atomic(&page->cp_reference,
3157                                           "gang_lookup", current);
3158                         pvec[j++] = ops;
3159                 }
3160                 ++idx;
3161
3162                 /*
3163                  * Here a delicate locking dance is performed. Current thread
3164                  * holds a reference to a page, but has to own it before it
3165                  * can be placed into queue. Owning implies waiting, so
3166                  * radix-tree lock is to be released. After a wait one has to
3167                  * check that pages weren't truncated (cl_page_own() returns
3168                  * error in the latter case).
3169                  */
3170                 spin_unlock(&osc->oo_tree_lock);
3171                 tree_lock = false;
3172
3173                 for (i = 0; i < j; ++i) {
3174                         ops = pvec[i];
3175                         if (res == CLP_GANG_OKAY)
3176                                 res = (*cb)(env, io, ops, cbdata);
3177
3178                         page = ops->ops_cl.cpl_page;
3179                         lu_ref_del(&page->cp_reference, "gang_lookup", current);
3180                         cl_pagevec_put(env, page, pagevec);
3181                 }
3182                 pagevec_release(pagevec);
3183
3184                 if (nr < OTI_PVEC_SIZE || end_of_region)
3185                         break;
3186
3187                 if (res == CLP_GANG_OKAY && need_resched())
3188                         res = CLP_GANG_RESCHED;
3189                 if (res != CLP_GANG_OKAY)
3190                         break;
3191
3192                 spin_lock(&osc->oo_tree_lock);
3193                 tree_lock = true;
3194         }
3195         if (tree_lock)
3196                 spin_unlock(&osc->oo_tree_lock);
3197         RETURN(res);
3198 }
3199 EXPORT_SYMBOL(osc_page_gang_lookup);
3200
3201 /**
3202  * Check if page @page is covered by an extra lock or discard it.
3203  */
3204 static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3205                                 struct osc_page *ops, void *cbdata)
3206 {
3207         struct osc_thread_info *info = osc_env_info(env);
3208         struct osc_object *osc = cbdata;
3209         pgoff_t index;
3210
3211         index = osc_index(ops);
3212         if (index >= info->oti_fn_index) {
3213                 struct ldlm_lock *tmp;
3214                 struct cl_page *page = ops->ops_cl.cpl_page;
3215
3216                 /* refresh non-overlapped index */
3217                 tmp = osc_dlmlock_at_pgoff(env, osc, index,
3218                                            OSC_DAP_FL_TEST_LOCK);
3219                 if (tmp != NULL) {
3220                         __u64 end = tmp->l_policy_data.l_extent.end;
3221                         /* Cache the first-non-overlapped index so as to skip
3222                          * all pages within [index, oti_fn_index). This is safe
3223                          * because if tmp lock is canceled, it will discard
3224                          * these pages. */
3225                         info->oti_fn_index = cl_index(osc2cl(osc), end + 1);
3226                         if (end == OBD_OBJECT_EOF)
3227                                 info->oti_fn_index = CL_PAGE_EOF;
3228                         LDLM_LOCK_PUT(tmp);
3229                 } else if (cl_page_own(env, io, page) == 0) {
3230                         /* discard the page */
3231                         cl_page_discard(env, io, page);
3232                         cl_page_disown(env, io, page);
3233                 } else {
3234                         LASSERT(page->cp_state == CPS_FREEING);
3235                 }
3236         }
3237
3238         info->oti_next_index = index + 1;
3239         return CLP_GANG_OKAY;
3240 }
3241
3242 int osc_discard_cb(const struct lu_env *env, struct cl_io *io,
3243                    struct osc_page *ops, void *cbdata)
3244 {
3245         struct osc_thread_info *info = osc_env_info(env);
3246         struct cl_page *page = ops->ops_cl.cpl_page;
3247
3248         /* page is top page. */
3249         info->oti_next_index = osc_index(ops) + 1;
3250         if (cl_page_own(env, io, page) == 0) {
3251                 if (!ergo(page->cp_type == CPT_CACHEABLE,
3252                           !PageDirty(cl_page_vmpage(page))))
3253                         CL_PAGE_DEBUG(D_ERROR, env, page,
3254                                         "discard dirty page?\n");
3255
3256                 /* discard the page */
3257                 cl_page_discard(env, io, page);
3258                 cl_page_disown(env, io, page);
3259         } else {
3260                 LASSERT(page->cp_state == CPS_FREEING);
3261         }
3262
3263         return CLP_GANG_OKAY;
3264 }
3265 EXPORT_SYMBOL(osc_discard_cb);
3266
3267 /**
3268  * Discard pages protected by the given lock. This function traverses radix
3269  * tree to find all covering pages and discard them. If a page is being covered
3270  * by other locks, it should remain in cache.
3271  *
3272  * If error happens on any step, the process continues anyway (the reasoning
3273  * behind this being that lock cancellation cannot be delayed indefinitely).
3274  */
3275 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3276                            pgoff_t start, pgoff_t end, bool discard)
3277 {
3278         struct osc_thread_info *info = osc_env_info(env);
3279         struct cl_io *io = osc_env_thread_io(env);
3280         osc_page_gang_cbt cb;
3281         int res;
3282         int result;
3283
3284         ENTRY;
3285
3286         io->ci_obj = cl_object_top(osc2cl(osc));
3287         io->ci_ignore_layout = 1;
3288         result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3289         if (result != 0)
3290                 GOTO(out, result);
3291
3292         cb = discard ? osc_discard_cb : check_and_discard_cb;
3293         info->oti_fn_index = info->oti_next_index = start;
3294         do {
3295                 res = osc_page_gang_lookup(env, io, osc,
3296                                            info->oti_next_index, end, cb, osc);
3297                 if (info->oti_next_index > end)
3298                         break;
3299
3300                 if (res == CLP_GANG_RESCHED)
3301                         cond_resched();
3302         } while (res != CLP_GANG_OKAY);
3303 out:
3304         cl_io_fini(env, io);
3305         RETURN(result);
3306 }
3307
3308
3309 /** @} osc */