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, 2016, Intel Corporation.
  27  *
  28  */
  29 /*
  30  * This file is part of Lustre, http://www.lustre.org/
  31  * Lustre is a trademark of Sun Microsystems, Inc.
  32  *
  33  * osc cache management.
  34  *
  35  * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_OSC
  39
  40 #include "osc_cl_internal.h"
  41 #include "osc_internal.h"
  42
  43 static int extent_debug; /* set it to be true for more debug */
  44
  45 static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
  46 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
  47                            enum osc_extent_state state);
  48 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
  49                               struct osc_async_page *oap, int sent, int rc);
  50 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
  51                           int cmd);
  52 static int osc_refresh_count(const struct lu_env *env,
  53                              struct osc_async_page *oap, int cmd);
  54 static int osc_io_unplug_async(const struct lu_env *env,
  55                                struct client_obd *cli, struct osc_object *osc);
  56 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
  57                            unsigned int lost_grant, unsigned int dirty_grant);
  58
  59 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
  60                                   const char *func, int line);
  61 #define osc_extent_tree_dump(lvl, obj) \
  62         osc_extent_tree_dump0(lvl, obj, __func__, __LINE__)
  63
  64 static void osc_unreserve_grant(struct client_obd *cli, unsigned int reserved,
  65                                 unsigned int unused);
  66
  67 /** \addtogroup osc
  68  *  @{
  69  */
  70
  71 /* ------------------ osc extent ------------------ */
  72 static inline char *ext_flags(struct osc_extent *ext, char *flags)
  73 {
  74         char *buf = flags;
  75         *buf++ = ext->oe_rw ? 'r' : 'w';
  76         if (ext->oe_intree)
  77                 *buf++ = 'i';
  78         if (ext->oe_sync)
  79                 *buf++ = 'S';
  80         if (ext->oe_srvlock)
  81                 *buf++ = 's';
  82         if (ext->oe_hp)
  83                 *buf++ = 'h';
  84         if (ext->oe_urgent)
  85                 *buf++ = 'u';
  86         if (ext->oe_memalloc)
  87                 *buf++ = 'm';
  88         if (ext->oe_trunc_pending)
  89                 *buf++ = 't';
  90         if (ext->oe_fsync_wait)
  91                 *buf++ = 'Y';
  92         *buf = 0;
  93         return flags;
  94 }
  95
  96 static inline char list_empty_marker(struct list_head *list)
  97 {
  98         return list_empty(list) ? '-' : '+';
  99 }
 100
 101 #define EXTSTR       "[%lu -> %lu/%lu]"
 102 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
 103 static const char *oes_strings[] = {
 104         "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
 105
 106 #define OSC_EXTENT_DUMP(lvl, extent, fmt, ...) do {                           \
 107         struct osc_extent *__ext = (extent);                                  \
 108         char __buf[16];                                                       \
 109                                                                               \
 110         CDEBUG(lvl,                                                           \
 111                 "extent %p@{" EXTSTR ", "                                     \
 112                 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt,          \
 113                 /* ----- extent part 0 ----- */                               \
 114                 __ext, EXTPARA(__ext),                                        \
 115                 /* ----- part 1 ----- */                                      \
 116                 atomic_read(&__ext->oe_refc),                         \
 117                 atomic_read(&__ext->oe_users),                        \
 118                 list_empty_marker(&__ext->oe_link),                           \
 119                 oes_strings[__ext->oe_state], ext_flags(__ext, __buf),        \
 120                 __ext->oe_obj,                                                \
 121                 /* ----- part 2 ----- */                                      \
 122                 __ext->oe_grants, __ext->oe_nr_pages,                         \
 123                 list_empty_marker(&__ext->oe_pages),                          \
 124                 waitqueue_active(&__ext->oe_waitq) ? '+' : '-',               \
 125                 __ext->oe_dlmlock, __ext->oe_mppr, __ext->oe_owner,           \
 126                 /* ----- part 4 ----- */                                      \
 127                 ## __VA_ARGS__);                                              \
 128         if (lvl == D_ERROR && __ext->oe_dlmlock != NULL)                      \
 129                 LDLM_ERROR(__ext->oe_dlmlock, "extent: %p", __ext);           \
 130         else                                                                  \
 131                 LDLM_DEBUG(__ext->oe_dlmlock, "extent: %p", __ext);           \
 132 } while (0)
 133
 134 #undef EASSERTF
 135 #define EASSERTF(expr, ext, fmt, args...) do {                          \
 136         if (!(expr)) {                                                  \
 137                 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args);           \
 138                 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj);           \
 139                 LASSERT(expr);                                          \
 140         }                                                               \
 141 } while (0)
 142
 143 #undef EASSERT
 144 #define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
 145
 146 static inline struct osc_extent *rb_extent(struct rb_node *n)
 147 {
 148         if (n == NULL)
 149                 return NULL;
 150
 151         return container_of(n, struct osc_extent, oe_node);
 152 }
 153
 154 static inline struct osc_extent *next_extent(struct osc_extent *ext)
 155 {
 156         if (ext == NULL)
 157                 return NULL;
 158
 159         LASSERT(ext->oe_intree);
 160         return rb_extent(rb_next(&ext->oe_node));
 161 }
 162
 163 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
 164 {
 165         if (ext == NULL)
 166                 return NULL;
 167
 168         LASSERT(ext->oe_intree);
 169         return rb_extent(rb_prev(&ext->oe_node));
 170 }
 171
 172 static inline struct osc_extent *first_extent(struct osc_object *obj)
 173 {
 174         return rb_extent(rb_first(&obj->oo_root));
 175 }
 176
 177 /* object must be locked by caller. */
 178 static int osc_extent_sanity_check0(struct osc_extent *ext,
 179                                     const char *func, const int line)
 180 {
 181         struct osc_object *obj = ext->oe_obj;
 182         struct osc_async_page *oap;
 183         size_t page_count;
 184         int rc = 0;
 185
 186         if (!osc_object_is_locked(obj))
 187                 GOTO(out, rc = 9);
 188
 189         if (ext->oe_state >= OES_STATE_MAX)
 190                 GOTO(out, rc = 10);
 191
 192         if (atomic_read(&ext->oe_refc) <= 0)
 193                 GOTO(out, rc = 20);
 194
 195         if (atomic_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
 196                 GOTO(out, rc = 30);
 197
 198         switch (ext->oe_state) {
 199         case OES_INV:
 200                 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
 201                         GOTO(out, rc = 35);
 202                 GOTO(out, rc = 0);
 203                 break;
 204         case OES_ACTIVE:
 205                 if (atomic_read(&ext->oe_users) == 0)
 206                         GOTO(out, rc = 40);
 207                 if (ext->oe_hp)
 208                         GOTO(out, rc = 50);
 209                 if (ext->oe_fsync_wait && !ext->oe_urgent)
 210                         GOTO(out, rc = 55);
 211                 break;
 212         case OES_CACHE:
 213                 if (ext->oe_grants == 0)
 214                         GOTO(out, rc = 60);
 215                 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
 216                         GOTO(out, rc = 65);
 217         default:
 218                 if (atomic_read(&ext->oe_users) > 0)
 219                         GOTO(out, rc = 70);
 220         }
 221
 222         if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
 223                 GOTO(out, rc = 80);
 224
 225         if (ext->oe_sync && ext->oe_grants > 0)
 226                 GOTO(out, rc = 90);
 227
 228         if (ext->oe_dlmlock != NULL && !ldlm_is_failed(ext->oe_dlmlock)) {
 229                 struct ldlm_extent *extent;
 230
 231                 extent = &ext->oe_dlmlock->l_policy_data.l_extent;
 232                 if (!(extent->start <= cl_offset(osc2cl(obj), ext->oe_start) &&
 233                       extent->end   >= cl_offset(osc2cl(obj), ext->oe_max_end)))
 234                         GOTO(out, rc = 100);
 235
 236                 if (!(ext->oe_dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP)))
 237                         GOTO(out, rc = 102);
 238         }
 239
 240         if (ext->oe_nr_pages > ext->oe_mppr)
 241                 GOTO(out, rc = 105);
 242
 243         /* Do not verify page list if extent is in RPC. This is because an
 244          * in-RPC extent is supposed to be exclusively accessible w/o lock. */
 245         if (ext->oe_state > OES_CACHE)
 246                 GOTO(out, rc = 0);
 247
 248         if (!extent_debug)
 249                 GOTO(out, rc = 0);
 250
 251         page_count = 0;
 252         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
 253                 pgoff_t index = osc_index(oap2osc(oap));
 254                 ++page_count;
 255                 if (index > ext->oe_end || index < ext->oe_start)
 256                         GOTO(out, rc = 110);
 257         }
 258         if (page_count != ext->oe_nr_pages)
 259                 GOTO(out, rc = 120);
 260
 261 out:
 262         if (rc != 0)
 263                 OSC_EXTENT_DUMP(D_ERROR, ext,
 264                                 "%s:%d sanity check %p failed with rc = %d\n",
 265                                 func, line, ext, rc);
 266         return rc;
 267 }
 268
 269 #define sanity_check_nolock(ext) \
 270         osc_extent_sanity_check0(ext, __func__, __LINE__)
 271
 272 #define sanity_check(ext) ({                                                   \
 273         int __res;                                                             \
 274         osc_object_lock((ext)->oe_obj);                                        \
 275         __res = sanity_check_nolock(ext);                                      \
 276         osc_object_unlock((ext)->oe_obj);                                      \
 277         __res;                                                                 \
 278 })
 279
 280
 281 /**
 282  * sanity check - to make sure there is no overlapped extent in the tree.
 283  */
 284 static int osc_extent_is_overlapped(struct osc_object *obj,
 285                                     struct osc_extent *ext)
 286 {
 287         struct osc_extent *tmp;
 288
 289         LASSERT(osc_object_is_locked(obj));
 290
 291         if (!extent_debug)
 292                 return 0;
 293
 294         for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
 295                 if (tmp == ext)
 296                         continue;
 297                 if (tmp->oe_end >= ext->oe_start &&
 298                     tmp->oe_start <= ext->oe_end)
 299                         return 1;
 300         }
 301         return 0;
 302 }
 303
 304 static void osc_extent_state_set(struct osc_extent *ext, int state)
 305 {
 306         LASSERT(osc_object_is_locked(ext->oe_obj));
 307         LASSERT(state >= OES_INV && state < OES_STATE_MAX);
 308
 309         /* Never try to sanity check a state changing extent :-) */
 310         /* LASSERT(sanity_check_nolock(ext) == 0); */
 311
 312         /* TODO: validate the state machine */
 313         ext->oe_state = state;
 314         wake_up_all(&ext->oe_waitq);
 315 }
 316
 317 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
 318 {
 319         struct osc_extent *ext;
 320
 321         OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_NOFS);
 322         if (ext == NULL)
 323                 return NULL;
 324
 325         RB_CLEAR_NODE(&ext->oe_node);
 326         ext->oe_obj = obj;
 327         cl_object_get(osc2cl(obj));
 328         atomic_set(&ext->oe_refc, 1);
 329         atomic_set(&ext->oe_users, 0);
 330         INIT_LIST_HEAD(&ext->oe_link);
 331         ext->oe_state = OES_INV;
 332         INIT_LIST_HEAD(&ext->oe_pages);
 333         init_waitqueue_head(&ext->oe_waitq);
 334         ext->oe_dlmlock = NULL;
 335
 336         return ext;
 337 }
 338
 339 static void osc_extent_free(struct osc_extent *ext)
 340 {
 341         OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
 342 }
 343
 344 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
 345 {
 346         LASSERT(atomic_read(&ext->oe_refc) >= 0);
 347         atomic_inc(&ext->oe_refc);
 348         return ext;
 349 }
 350
 351 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
 352 {
 353         LASSERT(atomic_read(&ext->oe_refc) > 0);
 354         if (atomic_dec_and_test(&ext->oe_refc)) {
 355                 LASSERT(list_empty(&ext->oe_link));
 356                 LASSERT(atomic_read(&ext->oe_users) == 0);
 357                 LASSERT(ext->oe_state == OES_INV);
 358                 LASSERT(!ext->oe_intree);
 359
 360                 if (ext->oe_dlmlock != NULL) {
 361                         lu_ref_add(&ext->oe_dlmlock->l_reference,
 362                                    "osc_extent", ext);
 363                         LDLM_LOCK_PUT(ext->oe_dlmlock);
 364                         ext->oe_dlmlock = NULL;
 365                 }
 366                 cl_object_put(env, osc2cl(ext->oe_obj));
 367                 osc_extent_free(ext);
 368         }
 369 }
 370
 371 /**
 372  * osc_extent_put_trust() is a special version of osc_extent_put() when
 373  * it's known that the caller is not the last user. This is to address the
 374  * problem of lacking of lu_env ;-).
 375  */
 376 static void osc_extent_put_trust(struct osc_extent *ext)
 377 {
 378         LASSERT(atomic_read(&ext->oe_refc) > 1);
 379         LASSERT(osc_object_is_locked(ext->oe_obj));
 380         atomic_dec(&ext->oe_refc);
 381 }
 382
 383 /**
 384  * Return the extent which includes pgoff @index, or return the greatest
 385  * previous extent in the tree.
 386  */
 387 static struct osc_extent *osc_extent_search(struct osc_object *obj,
 388                                             pgoff_t index)
 389 {
 390         struct rb_node    *n = obj->oo_root.rb_node;
 391         struct osc_extent *tmp, *p = NULL;
 392
 393         LASSERT(osc_object_is_locked(obj));
 394         while (n != NULL) {
 395                 tmp = rb_extent(n);
 396                 if (index < tmp->oe_start) {
 397                         n = n->rb_left;
 398                 } else if (index > tmp->oe_end) {
 399                         p = rb_extent(n);
 400                         n = n->rb_right;
 401                 } else {
 402                         return tmp;
 403                 }
 404         }
 405         return p;
 406 }
 407
 408 /*
 409  * Return the extent covering @index, otherwise return NULL.
 410  * caller must have held object lock.
 411  */
 412 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
 413                                             pgoff_t index)
 414 {
 415         struct osc_extent *ext;
 416
 417         ext = osc_extent_search(obj, index);
 418         if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
 419                 return osc_extent_get(ext);
 420         return NULL;
 421 }
 422
 423 /* caller must have held object lock. */
 424 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
 425 {
 426         struct rb_node   **n      = &obj->oo_root.rb_node;
 427         struct rb_node    *parent = NULL;
 428         struct osc_extent *tmp;
 429
 430         LASSERT(ext->oe_intree == 0);
 431         LASSERT(ext->oe_obj == obj);
 432         LASSERT(osc_object_is_locked(obj));
 433         while (*n != NULL) {
 434                 tmp = rb_extent(*n);
 435                 parent = *n;
 436
 437                 if (ext->oe_end < tmp->oe_start)
 438                         n = &(*n)->rb_left;
 439                 else if (ext->oe_start > tmp->oe_end)
 440                         n = &(*n)->rb_right;
 441                 else
 442                         EASSERTF(0, tmp, EXTSTR"\n", EXTPARA(ext));
 443         }
 444         rb_link_node(&ext->oe_node, parent, n);
 445         rb_insert_color(&ext->oe_node, &obj->oo_root);
 446         osc_extent_get(ext);
 447         ext->oe_intree = 1;
 448 }
 449
 450 /* caller must have held object lock. */
 451 static void osc_extent_erase(struct osc_extent *ext)
 452 {
 453         struct osc_object *obj = ext->oe_obj;
 454         LASSERT(osc_object_is_locked(obj));
 455         if (ext->oe_intree) {
 456                 rb_erase(&ext->oe_node, &obj->oo_root);
 457                 ext->oe_intree = 0;
 458                 /* rbtree held a refcount */
 459                 osc_extent_put_trust(ext);
 460         }
 461 }
 462
 463 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
 464 {
 465         struct osc_object *obj = ext->oe_obj;
 466
 467         LASSERT(osc_object_is_locked(obj));
 468         LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
 469         if (ext->oe_state == OES_CACHE) {
 470                 osc_extent_state_set(ext, OES_ACTIVE);
 471                 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
 472         }
 473         atomic_inc(&ext->oe_users);
 474         list_del_init(&ext->oe_link);
 475         return osc_extent_get(ext);
 476 }
 477
 478 static void __osc_extent_remove(struct osc_extent *ext)
 479 {
 480         LASSERT(osc_object_is_locked(ext->oe_obj));
 481         LASSERT(list_empty(&ext->oe_pages));
 482         osc_extent_erase(ext);
 483         list_del_init(&ext->oe_link);
 484         osc_extent_state_set(ext, OES_INV);
 485         OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
 486 }
 487
 488 static void osc_extent_remove(struct osc_extent *ext)
 489 {
 490         struct osc_object *obj = ext->oe_obj;
 491
 492         osc_object_lock(obj);
 493         __osc_extent_remove(ext);
 494         osc_object_unlock(obj);
 495 }
 496
 497 /**
 498  * This function is used to merge extents to get better performance. It checks
 499  * if @cur and @victim are contiguous at block level.
 500  */
 501 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
 502                             struct osc_extent *victim)
 503 {
 504         struct osc_object       *obj = cur->oe_obj;
 505         struct client_obd       *cli = osc_cli(obj);
 506         pgoff_t                  chunk_start;
 507         pgoff_t                  chunk_end;
 508         int                      ppc_bits;
 509
 510         LASSERT(cur->oe_state == OES_CACHE);
 511         LASSERT(osc_object_is_locked(obj));
 512         if (victim == NULL)
 513                 return -EINVAL;
 514
 515         if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
 516                 return -EBUSY;
 517
 518         if (cur->oe_max_end != victim->oe_max_end)
 519                 return -ERANGE;
 520
 521         LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
 522         ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_SHIFT;
 523         chunk_start = cur->oe_start >> ppc_bits;
 524         chunk_end   = cur->oe_end   >> ppc_bits;
 525         if (chunk_start   != (victim->oe_end >> ppc_bits) + 1 &&
 526             chunk_end + 1 != victim->oe_start >> ppc_bits)
 527                 return -ERANGE;
 528
 529         /* overall extent size should not exceed the max supported limit
 530          * reported by the server */
 531         if (cur->oe_end - cur->oe_start + 1 +
 532             victim->oe_end - victim->oe_start + 1 > cli->cl_max_extent_pages)
 533                 return -ERANGE;
 534
 535         OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
 536
 537         cur->oe_start     = min(cur->oe_start, victim->oe_start);
 538         cur->oe_end       = max(cur->oe_end,   victim->oe_end);
 539         /* per-extent tax should be accounted only once for the whole extent */
 540         cur->oe_grants   += victim->oe_grants - cli->cl_grant_extent_tax;
 541         cur->oe_nr_pages += victim->oe_nr_pages;
 542         /* only the following bits are needed to merge */
 543         cur->oe_urgent   |= victim->oe_urgent;
 544         cur->oe_memalloc |= victim->oe_memalloc;
 545         list_splice_init(&victim->oe_pages, &cur->oe_pages);
 546         list_del_init(&victim->oe_link);
 547         victim->oe_nr_pages = 0;
 548
 549         osc_extent_get(victim);
 550         __osc_extent_remove(victim);
 551         osc_extent_put(env, victim);
 552
 553         OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
 554         return 0;
 555 }
 556
 557 /**
 558  * Drop user count of osc_extent, and unplug IO asynchronously.
 559  */
 560 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
 561 {
 562         struct osc_object *obj = ext->oe_obj;
 563         struct client_obd *cli = osc_cli(obj);
 564         int rc = 0;
 565         ENTRY;
 566
 567         LASSERT(atomic_read(&ext->oe_users) > 0);
 568         LASSERT(sanity_check(ext) == 0);
 569         LASSERT(ext->oe_grants > 0);
 570
 571         if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
 572                 LASSERT(ext->oe_state == OES_ACTIVE);
 573                 if (ext->oe_trunc_pending) {
 574                         /* a truncate process is waiting for this extent.
 575                          * This may happen due to a race, check
 576                          * osc_cache_truncate_start(). */
 577                         osc_extent_state_set(ext, OES_TRUNC);
 578                         ext->oe_trunc_pending = 0;
 579                 } else {
 580                         int grant = 0;
 581
 582                         osc_extent_state_set(ext, OES_CACHE);
 583                         osc_update_pending(obj, OBD_BRW_WRITE,
 584                                            ext->oe_nr_pages);
 585
 586                         /* try to merge the previous and next extent. */
 587                         if (osc_extent_merge(env, ext, prev_extent(ext)) == 0)
 588                                 grant += cli->cl_grant_extent_tax;
 589                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 590                                 grant += cli->cl_grant_extent_tax;
 591                         if (grant > 0)
 592                                 osc_unreserve_grant(cli, 0, grant);
 593
 594                         if (ext->oe_urgent)
 595                                 list_move_tail(&ext->oe_link,
 596                                                &obj->oo_urgent_exts);
 597                         else if (ext->oe_nr_pages == ext->oe_mppr) {
 598                                 list_move_tail(&ext->oe_link,
 599                                                &obj->oo_full_exts);
 600                         }
 601                 }
 602                 osc_object_unlock(obj);
 603
 604                 osc_io_unplug_async(env, cli, obj);
 605         }
 606         osc_extent_put(env, ext);
 607         RETURN(rc);
 608 }
 609
 610 static inline int overlapped(struct osc_extent *ex1, struct osc_extent *ex2)
 611 {
 612         return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
 613 }
 614
 615 /**
 616  * Find or create an extent which includes @index, core function to manage
 617  * extent tree.
 618  */
 619 static struct osc_extent *osc_extent_find(const struct lu_env *env,
 620                                           struct osc_object *obj, pgoff_t index,
 621                                           unsigned int *grants)
 622 {
 623         struct client_obd *cli = osc_cli(obj);
 624         struct osc_lock   *olck;
 625         struct cl_lock_descr *descr;
 626         struct osc_extent *cur;
 627         struct osc_extent *ext;
 628         struct osc_extent *conflict = NULL;
 629         struct osc_extent *found = NULL;
 630         pgoff_t    chunk;
 631         pgoff_t    max_end;
 632         unsigned int max_pages; /* max_pages_per_rpc */
 633         unsigned int chunksize;
 634         int        ppc_bits; /* pages per chunk bits */
 635         pgoff_t    chunk_mask;
 636         int        rc;
 637         ENTRY;
 638
 639         cur = osc_extent_alloc(obj);
 640         if (cur == NULL)
 641                 RETURN(ERR_PTR(-ENOMEM));
 642
 643         olck = osc_env_io(env)->oi_write_osclock;
 644         LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
 645         LASSERT(olck->ols_state == OLS_GRANTED);
 646
 647         descr = &olck->ols_cl.cls_lock->cll_descr;
 648         LASSERT(descr->cld_mode >= CLM_WRITE);
 649
 650         LASSERTF(cli->cl_chunkbits >= PAGE_SHIFT,
 651                  "chunkbits: %u\n", cli->cl_chunkbits);
 652         ppc_bits   = cli->cl_chunkbits - PAGE_SHIFT;
 653         chunk_mask = ~((1 << ppc_bits) - 1);
 654         chunksize  = 1 << cli->cl_chunkbits;
 655         chunk      = index >> ppc_bits;
 656
 657         /* align end to RPC edge. */
 658         max_pages = cli->cl_max_pages_per_rpc;
 659         if ((max_pages & ~chunk_mask) != 0) {
 660                 CERROR("max_pages: %#x chunkbits: %u chunk_mask: %#lx\n",
 661                        max_pages, cli->cl_chunkbits, chunk_mask);
 662                 RETURN(ERR_PTR(-EINVAL));
 663         }
 664         max_end = index - (index % max_pages) + max_pages - 1;
 665         max_end = min_t(pgoff_t, max_end, descr->cld_end);
 666
 667         /* initialize new extent by parameters so far */
 668         cur->oe_max_end = max_end;
 669         cur->oe_start   = index & chunk_mask;
 670         cur->oe_end     = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
 671         if (cur->oe_start < descr->cld_start)
 672                 cur->oe_start = descr->cld_start;
 673         if (cur->oe_end > max_end)
 674                 cur->oe_end = max_end;
 675         cur->oe_grants  = 0;
 676         cur->oe_mppr    = max_pages;
 677         if (olck->ols_dlmlock != NULL) {
 678                 LASSERT(olck->ols_hold);
 679                 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
 680                 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
 681         }
 682
 683         /* grants has been allocated by caller */
 684         LASSERTF(*grants >= chunksize + cli->cl_grant_extent_tax,
 685                  "%u/%u/%u.\n", *grants, chunksize, cli->cl_grant_extent_tax);
 686         LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
 687                  EXTPARA(cur));
 688
 689 restart:
 690         osc_object_lock(obj);
 691         ext = osc_extent_search(obj, cur->oe_start);
 692         if (ext == NULL)
 693                 ext = first_extent(obj);
 694         while (ext != NULL) {
 695                 pgoff_t ext_chk_start = ext->oe_start >> ppc_bits;
 696                 pgoff_t ext_chk_end   = ext->oe_end   >> ppc_bits;
 697
 698                 LASSERT(sanity_check_nolock(ext) == 0);
 699                 if (chunk > ext_chk_end + 1)
 700                         break;
 701
 702                 /* if covering by different locks, no chance to match */
 703                 if (olck->ols_dlmlock != ext->oe_dlmlock) {
 704                         EASSERTF(!overlapped(ext, cur), ext,
 705                                  EXTSTR"\n", EXTPARA(cur));
 706
 707                         ext = next_extent(ext);
 708                         continue;
 709                 }
 710
 711                 /* discontiguous chunks? */
 712                 if (chunk + 1 < ext_chk_start) {
 713                         ext = next_extent(ext);
 714                         continue;
 715                 }
 716
 717                 /* ok, from now on, ext and cur have these attrs:
 718                  * 1. covered by the same lock
 719                  * 2. contiguous at chunk level or overlapping. */
 720
 721                 if (overlapped(ext, cur)) {
 722                         /* cur is the minimum unit, so overlapping means
 723                          * full contain. */
 724                         EASSERTF((ext->oe_start <= cur->oe_start &&
 725                                   ext->oe_end >= cur->oe_end),
 726                                  ext, EXTSTR"\n", EXTPARA(cur));
 727
 728                         if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
 729                                 /* for simplicity, we wait for this extent to
 730                                  * finish before going forward. */
 731                                 conflict = osc_extent_get(ext);
 732                                 break;
 733                         }
 734
 735                         found = osc_extent_hold(ext);
 736                         break;
 737                 }
 738
 739                 /* non-overlapped extent */
 740                 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
 741                         /* we can't do anything for a non OES_CACHE extent, or
 742                          * if there is someone waiting for this extent to be
 743                          * flushed, try next one. */
 744                         ext = next_extent(ext);
 745                         continue;
 746                 }
 747
 748                 /* check if they belong to the same rpc slot before trying to
 749                  * merge. the extents are not overlapped and contiguous at
 750                  * chunk level to get here. */
 751                 if (ext->oe_max_end != max_end) {
 752                         /* if they don't belong to the same RPC slot or
 753                          * max_pages_per_rpc has ever changed, do not merge. */
 754                         ext = next_extent(ext);
 755                         continue;
 756                 }
 757
 758                 /* check whether maximum extent size will be hit */
 759                 if ((ext_chk_end - ext_chk_start + 1 + 1) << ppc_bits >
 760                     cli->cl_max_extent_pages) {
 761                         ext = next_extent(ext);
 762                         continue;
 763                 }
 764
 765                 /* it's required that an extent must be contiguous at chunk
 766                  * level so that we know the whole extent is covered by grant
 767                  * (the pages in the extent are NOT required to be contiguous).
 768                  * Otherwise, it will be too much difficult to know which
 769                  * chunks have grants allocated. */
 770
 771                 /* try to do front merge - extend ext's start */
 772                 if (chunk + 1 == ext_chk_start) {
 773                         /* ext must be chunk size aligned */
 774                         EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
 775
 776                         /* pull ext's start back to cover cur */
 777                         ext->oe_start   = cur->oe_start;
 778                         ext->oe_grants += chunksize;
 779                         LASSERT(*grants >= chunksize);
 780                         *grants -= chunksize;
 781
 782                         found = osc_extent_hold(ext);
 783                 } else if (chunk == ext_chk_end + 1) {
 784                         /* rear merge */
 785                         ext->oe_end     = cur->oe_end;
 786                         ext->oe_grants += chunksize;
 787                         LASSERT(*grants >= chunksize);
 788                         *grants -= chunksize;
 789
 790                         /* try to merge with the next one because we just fill
 791                          * in a gap */
 792                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 793                                 /* we can save extent tax from next extent */
 794                                 *grants += cli->cl_grant_extent_tax;
 795
 796                         found = osc_extent_hold(ext);
 797                 }
 798                 if (found != NULL)
 799                         break;
 800
 801                 ext = next_extent(ext);
 802         }
 803
 804         osc_extent_tree_dump(D_CACHE, obj);
 805         if (found != NULL) {
 806                 LASSERT(conflict == NULL);
 807                 if (!IS_ERR(found)) {
 808                         LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
 809                         OSC_EXTENT_DUMP(D_CACHE, found,
 810                                         "found caching ext for %lu.\n", index);
 811                 }
 812         } else if (conflict == NULL) {
 813                 /* create a new extent */
 814                 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
 815                 cur->oe_grants = chunksize + cli->cl_grant_extent_tax;
 816                 LASSERT(*grants >= cur->oe_grants);
 817                 *grants -= cur->oe_grants;
 818
 819                 cur->oe_state = OES_CACHE;
 820                 found = osc_extent_hold(cur);
 821                 osc_extent_insert(obj, cur);
 822                 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
 823                                 index, descr->cld_end);
 824         }
 825         osc_object_unlock(obj);
 826
 827         if (conflict != NULL) {
 828                 LASSERT(found == NULL);
 829
 830                 /* waiting for IO to finish. Please notice that it's impossible
 831                  * to be an OES_TRUNC extent. */
 832                 rc = osc_extent_wait(env, conflict, OES_INV);
 833                 osc_extent_put(env, conflict);
 834                 conflict = NULL;
 835                 if (rc < 0)
 836                         GOTO(out, found = ERR_PTR(rc));
 837
 838                 goto restart;
 839         }
 840         EXIT;
 841
 842 out:
 843         osc_extent_put(env, cur);
 844         return found;
 845 }
 846
 847 /**
 848  * Called when IO is finished to an extent.
 849  */
 850 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
 851                       int sent, int rc)
 852 {
 853         struct client_obd *cli = osc_cli(ext->oe_obj);
 854         struct osc_async_page *oap;
 855         struct osc_async_page *tmp;
 856         int nr_pages = ext->oe_nr_pages;
 857         int lost_grant = 0;
 858         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
 859         loff_t last_off = 0;
 860         int last_count = -1;
 861         ENTRY;
 862
 863         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
 864
 865         ext->oe_rc = rc ?: ext->oe_nr_pages;
 866         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
 867
 868         osc_lru_add_batch(cli, &ext->oe_pages);
 869         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 870                                      oap_pending_item) {
 871                 list_del_init(&oap->oap_rpc_item);
 872                 list_del_init(&oap->oap_pending_item);
 873                 if (last_off <= oap->oap_obj_off) {
 874                         last_off = oap->oap_obj_off;
 875                         last_count = oap->oap_count;
 876                 }
 877
 878                 --ext->oe_nr_pages;
 879                 osc_ap_completion(env, cli, oap, sent, rc);
 880         }
 881         EASSERT(ext->oe_nr_pages == 0, ext);
 882
 883         if (!sent) {
 884                 lost_grant = ext->oe_grants;
 885         } else if (blocksize < PAGE_SIZE &&
 886                    last_count != PAGE_SIZE) {
 887                 /* For short writes we shouldn't count parts of pages that
 888                  * span a whole chunk on the OST side, or our accounting goes
 889                  * wrong.  Should match the code in filter_grant_check. */
 890                 int offset = last_off & ~PAGE_MASK;
 891                 int count = last_count + (offset & (blocksize - 1));
 892                 int end = (offset + last_count) & (blocksize - 1);
 893                 if (end)
 894                         count += blocksize - end;
 895
 896                 lost_grant = PAGE_SIZE - count;
 897         }
 898         if (ext->oe_grants > 0)
 899                 osc_free_grant(cli, nr_pages, lost_grant, ext->oe_grants);
 900
 901         osc_extent_remove(ext);
 902         /* put the refcount for RPC */
 903         osc_extent_put(env, ext);
 904         RETURN(0);
 905 }
 906
 907 static int extent_wait_cb(struct osc_extent *ext, enum osc_extent_state state)
 908 {
 909         int ret;
 910
 911         osc_object_lock(ext->oe_obj);
 912         ret = ext->oe_state == state;
 913         osc_object_unlock(ext->oe_obj);
 914
 915         return ret;
 916 }
 917
 918 /**
 919  * Wait for the extent's state to become @state.
 920  */
 921 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
 922                            enum osc_extent_state state)
 923 {
 924         struct osc_object *obj = ext->oe_obj;
 925         struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
 926                                                   LWI_ON_SIGNAL_NOOP, NULL);
 927         int rc = 0;
 928         ENTRY;
 929
 930         osc_object_lock(obj);
 931         LASSERT(sanity_check_nolock(ext) == 0);
 932         /* `Kick' this extent only if the caller is waiting for it to be
 933          * written out. */
 934         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
 935                 if (ext->oe_state == OES_ACTIVE) {
 936                         ext->oe_urgent = 1;
 937                 } else if (ext->oe_state == OES_CACHE) {
 938                         ext->oe_urgent = 1;
 939                         osc_extent_hold(ext);
 940                         rc = 1;
 941                 }
 942         }
 943         osc_object_unlock(obj);
 944         if (rc == 1)
 945                 osc_extent_release(env, ext);
 946
 947         /* wait for the extent until its state becomes @state */
 948         rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
 949         if (rc == -ETIMEDOUT) {
 950                 OSC_EXTENT_DUMP(D_ERROR, ext,
 951                         "%s: wait ext to %u timedout, recovery in progress?\n",
 952                         cli_name(osc_cli(obj)), state);
 953
 954                 lwi = LWI_INTR(NULL, NULL);
 955                 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state),
 956                                   &lwi);
 957         }
 958         if (rc == 0 && ext->oe_rc < 0)
 959                 rc = ext->oe_rc;
 960         RETURN(rc);
 961 }
 962
 963 /**
 964  * Discard pages with index greater than @size. If @ext is overlapped with
 965  * @size, then partial truncate happens.
 966  */
 967 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
 968                                 bool partial)
 969 {
 970         struct lu_env         *env;
 971         struct cl_io          *io;
 972         struct osc_object     *obj = ext->oe_obj;
 973         struct client_obd     *cli = osc_cli(obj);
 974         struct osc_async_page *oap;
 975         struct osc_async_page *tmp;
 976         int                    pages_in_chunk = 0;
 977         int                    ppc_bits    = cli->cl_chunkbits -
 978                                              PAGE_SHIFT;
 979         __u64                  trunc_chunk = trunc_index >> ppc_bits;
 980         int                    grants   = 0;
 981         int                    nr_pages = 0;
 982         int                    rc       = 0;
 983         __u16                  refcheck;
 984         ENTRY;
 985
 986         LASSERT(sanity_check(ext) == 0);
 987         LASSERT(ext->oe_state == OES_TRUNC);
 988         LASSERT(!ext->oe_urgent);
 989
 990         /* Request new lu_env.
 991          * We can't use that env from osc_cache_truncate_start() because
 992          * it's from lov_io_sub and not fully initialized. */
 993         env = cl_env_get(&refcheck);
 994         if (IS_ERR(env))
 995                 RETURN(PTR_ERR(env));
 996
 997         io  = &osc_env_info(env)->oti_io;
 998         io->ci_obj = cl_object_top(osc2cl(obj));
 999         io->ci_ignore_layout = 1;
1000         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
1001         if (rc < 0)
1002                 GOTO(out, rc);
1003
1004         /* discard all pages with index greater than trunc_index */
1005         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
1006                                      oap_pending_item) {
1007                 pgoff_t index = osc_index(oap2osc(oap));
1008                 struct cl_page  *page = oap2cl_page(oap);
1009
1010                 LASSERT(list_empty(&oap->oap_rpc_item));
1011
1012                 /* only discard the pages with their index greater than
1013                  * trunc_index, and ... */
1014                 if (index < trunc_index ||
1015                     (index == trunc_index && partial)) {
1016                         /* accounting how many pages remaining in the chunk
1017                          * so that we can calculate grants correctly. */
1018                         if (index >> ppc_bits == trunc_chunk)
1019                                 ++pages_in_chunk;
1020                         continue;
1021                 }
1022
1023                 list_del_init(&oap->oap_pending_item);
1024
1025                 cl_page_get(page);
1026                 lu_ref_add(&page->cp_reference, "truncate", current);
1027
1028                 if (cl_page_own(env, io, page) == 0) {
1029                         cl_page_discard(env, io, page);
1030                         cl_page_disown(env, io, page);
1031                 } else {
1032                         LASSERT(page->cp_state == CPS_FREEING);
1033                         LASSERT(0);
1034                 }
1035
1036                 lu_ref_del(&page->cp_reference, "truncate", current);
1037                 cl_page_put(env, page);
1038
1039                 --ext->oe_nr_pages;
1040                 ++nr_pages;
1041         }
1042         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1043                       ext->oe_nr_pages == 0),
1044                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1045
1046         osc_object_lock(obj);
1047         if (ext->oe_nr_pages == 0) {
1048                 LASSERT(pages_in_chunk == 0);
1049                 grants = ext->oe_grants;
1050                 ext->oe_grants = 0;
1051         } else { /* calculate how many grants we can free */
1052                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1053                 pgoff_t last_index;
1054
1055
1056                 /* if there is no pages in this chunk, we can also free grants
1057                  * for the last chunk */
1058                 if (pages_in_chunk == 0) {
1059                         /* if this is the 1st chunk and no pages in this chunk,
1060                          * ext->oe_nr_pages must be zero, so we should be in
1061                          * the other if-clause. */
1062                         LASSERT(trunc_chunk > 0);
1063                         --trunc_chunk;
1064                         ++chunks;
1065                 }
1066
1067                 /* this is what we can free from this extent */
1068                 grants          = chunks << cli->cl_chunkbits;
1069                 ext->oe_grants -= grants;
1070                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1071                 ext->oe_end     = min(last_index, ext->oe_max_end);
1072                 LASSERT(ext->oe_end >= ext->oe_start);
1073                 LASSERT(ext->oe_grants > 0);
1074         }
1075         osc_object_unlock(obj);
1076
1077         if (grants > 0 || nr_pages > 0)
1078                 osc_free_grant(cli, nr_pages, grants, grants);
1079
1080 out:
1081         cl_io_fini(env, io);
1082         cl_env_put(env, &refcheck);
1083         RETURN(rc);
1084 }
1085
1086 /**
1087  * This function is used to make the extent prepared for transfer.
1088  * A race with flusing page - ll_writepage() has to be handled cautiously.
1089  */
1090 static int osc_extent_make_ready(const struct lu_env *env,
1091                                  struct osc_extent *ext)
1092 {
1093         struct osc_async_page *oap;
1094         struct osc_async_page *last = NULL;
1095         struct osc_object *obj = ext->oe_obj;
1096         unsigned int page_count = 0;
1097         int rc;
1098         ENTRY;
1099
1100         /* we're going to grab page lock, so object lock must not be taken. */
1101         LASSERT(sanity_check(ext) == 0);
1102         /* in locking state, any process should not touch this extent. */
1103         EASSERT(ext->oe_state == OES_LOCKING, ext);
1104         EASSERT(ext->oe_owner != NULL, ext);
1105
1106         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1107
1108         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1109                 ++page_count;
1110                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1111                         last = oap;
1112
1113                 /* checking ASYNC_READY is race safe */
1114                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1115                         continue;
1116
1117                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1118                 switch (rc) {
1119                 case 0:
1120                         spin_lock(&oap->oap_lock);
1121                         oap->oap_async_flags |= ASYNC_READY;
1122                         spin_unlock(&oap->oap_lock);
1123                         break;
1124                 case -EALREADY:
1125                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1126                         break;
1127                 default:
1128                         LASSERTF(0, "unknown return code: %d\n", rc);
1129                 }
1130         }
1131
1132         LASSERT(page_count == ext->oe_nr_pages);
1133         LASSERT(last != NULL);
1134         /* the last page is the only one we need to refresh its count by
1135          * the size of file. */
1136         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1137                 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1138                 LASSERT(last_oap_count > 0);
1139                 LASSERT(last->oap_page_off + last_oap_count <= PAGE_SIZE);
1140                 last->oap_count = last_oap_count;
1141                 spin_lock(&last->oap_lock);
1142                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1143                 spin_unlock(&last->oap_lock);
1144         }
1145
1146         /* for the rest of pages, we don't need to call osf_refresh_count()
1147          * because it's known they are not the last page */
1148         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1149                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1150                         oap->oap_count = PAGE_SIZE - oap->oap_page_off;
1151                         spin_lock(&oap->oap_lock);
1152                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1153                         spin_unlock(&oap->oap_lock);
1154                 }
1155         }
1156
1157         osc_object_lock(obj);
1158         osc_extent_state_set(ext, OES_RPC);
1159         osc_object_unlock(obj);
1160         /* get a refcount for RPC. */
1161         osc_extent_get(ext);
1162
1163         RETURN(0);
1164 }
1165
1166 /**
1167  * Quick and simple version of osc_extent_find(). This function is frequently
1168  * called to expand the extent for the same IO. To expand the extent, the
1169  * page index must be in the same or next chunk of ext->oe_end.
1170  */
1171 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index,
1172                              unsigned int *grants)
1173 {
1174         struct osc_object *obj = ext->oe_obj;
1175         struct client_obd *cli = osc_cli(obj);
1176         struct osc_extent *next;
1177         int ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1178         pgoff_t chunk = index >> ppc_bits;
1179         pgoff_t end_chunk;
1180         pgoff_t end_index;
1181         unsigned int chunksize = 1 << cli->cl_chunkbits;
1182         int rc = 0;
1183         ENTRY;
1184
1185         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1186         osc_object_lock(obj);
1187         LASSERT(sanity_check_nolock(ext) == 0);
1188         end_chunk = ext->oe_end >> ppc_bits;
1189         if (chunk > end_chunk + 1)
1190                 GOTO(out, rc = -ERANGE);
1191
1192         if (end_chunk >= chunk)
1193                 GOTO(out, rc = 0);
1194
1195         LASSERT(end_chunk + 1 == chunk);
1196
1197         /* try to expand this extent to cover @index */
1198         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1199
1200         /* don't go over the maximum extent size reported by server */
1201         if (end_index - ext->oe_start + 1 > cli->cl_max_extent_pages)
1202                 GOTO(out, rc = -ERANGE);
1203
1204         next = next_extent(ext);
1205         if (next != NULL && next->oe_start <= end_index)
1206                 /* complex mode - overlapped with the next extent,
1207                  * this case will be handled by osc_extent_find() */
1208                 GOTO(out, rc = -EAGAIN);
1209
1210         ext->oe_end = end_index;
1211         ext->oe_grants += chunksize;
1212         LASSERT(*grants >= chunksize);
1213         *grants -= chunksize;
1214         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1215                  "overlapped after expanding for %lu.\n", index);
1216         EXIT;
1217
1218 out:
1219         osc_object_unlock(obj);
1220         RETURN(rc);
1221 }
1222
1223 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
1224                                   const char *func, int line)
1225 {
1226         struct osc_extent *ext;
1227         int cnt;
1228
1229         CDEBUG(level, "Dump object %p extents at %s:%d, mppr: %u.\n",
1230                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1231
1232         /* osc_object_lock(obj); */
1233         cnt = 1;
1234         for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1235                 OSC_EXTENT_DUMP(level, ext, "in tree %d.\n", cnt++);
1236
1237         cnt = 1;
1238         list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1239                 OSC_EXTENT_DUMP(level, ext, "hp %d.\n", cnt++);
1240
1241         cnt = 1;
1242         list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1243                 OSC_EXTENT_DUMP(level, ext, "urgent %d.\n", cnt++);
1244
1245         cnt = 1;
1246         list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1247                 OSC_EXTENT_DUMP(level, ext, "reading %d.\n", cnt++);
1248         /* osc_object_unlock(obj); */
1249 }
1250
1251 /* ------------------ osc extent end ------------------ */
1252
1253 static inline int osc_is_ready(struct osc_object *osc)
1254 {
1255         return !list_empty(&osc->oo_ready_item) ||
1256                !list_empty(&osc->oo_hp_ready_item);
1257 }
1258
1259 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1260         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1261                (OSC), osc_is_ready(OSC),                                       \
1262                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1263                list_empty_marker(&(OSC)->oo_ready_item),                       \
1264                atomic_read(&(OSC)->oo_nr_writes),                              \
1265                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1266                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1267                atomic_read(&(OSC)->oo_nr_reads),                               \
1268                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1269                ##args)
1270
1271 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1272                           int cmd)
1273 {
1274         struct osc_page *opg  = oap2osc_page(oap);
1275         struct cl_page  *page = oap2cl_page(oap);
1276         int result;
1277
1278         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1279
1280         ENTRY;
1281         result = cl_page_make_ready(env, page, CRT_WRITE);
1282         if (result == 0)
1283                 opg->ops_submit_time = cfs_time_current();
1284         RETURN(result);
1285 }
1286
1287 static int osc_refresh_count(const struct lu_env *env,
1288                              struct osc_async_page *oap, int cmd)
1289 {
1290         struct osc_page  *opg = oap2osc_page(oap);
1291         pgoff_t index = osc_index(oap2osc(oap));
1292         struct cl_object *obj;
1293         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
1294
1295         int result;
1296         loff_t kms;
1297
1298         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1299         LASSERT(!(cmd & OBD_BRW_READ));
1300         LASSERT(opg != NULL);
1301         obj = opg->ops_cl.cpl_obj;
1302
1303         cl_object_attr_lock(obj);
1304         result = cl_object_attr_get(env, obj, attr);
1305         cl_object_attr_unlock(obj);
1306         if (result < 0)
1307                 return result;
1308         kms = attr->cat_kms;
1309         if (cl_offset(obj, index) >= kms)
1310                 /* catch race with truncate */
1311                 return 0;
1312         else if (cl_offset(obj, index + 1) > kms)
1313                 /* catch sub-page write at end of file */
1314                 return kms % PAGE_SIZE;
1315         else
1316                 return PAGE_SIZE;
1317 }
1318
1319 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1320                           int cmd, int rc)
1321 {
1322         struct osc_page   *opg  = oap2osc_page(oap);
1323         struct cl_page    *page = oap2cl_page(oap);
1324         enum cl_req_type   crt;
1325         int srvlock;
1326
1327         ENTRY;
1328
1329         cmd &= ~OBD_BRW_NOQUOTA;
1330         LASSERTF(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ),
1331                  "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1332         LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1333                 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1334         LASSERT(opg->ops_transfer_pinned);
1335
1336         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1337         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1338         opg->ops_transfer_pinned = 0;
1339
1340         opg->ops_submit_time = 0;
1341         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1342
1343         /* statistic */
1344         if (rc == 0 && srvlock) {
1345                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1346                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1347                 size_t bytes = oap->oap_count;
1348
1349                 if (crt == CRT_READ)
1350                         stats->os_lockless_reads += bytes;
1351                 else
1352                         stats->os_lockless_writes += bytes;
1353         }
1354
1355         /*
1356          * This has to be the last operation with the page, as locks are
1357          * released in cl_page_completion() and nothing except for the
1358          * reference counter protects page from concurrent reclaim.
1359          */
1360         lu_ref_del(&page->cp_reference, "transfer", page);
1361
1362         cl_page_completion(env, page, crt, rc);
1363         cl_page_put(env, page);
1364
1365         RETURN(0);
1366 }
1367
1368 #define OSC_DUMP_GRANT(lvl, cli, fmt, args...) do {                     \
1369         struct client_obd *__tmp = (cli);                               \
1370         CDEBUG(lvl, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu "  \
1371                "dropped: %ld avail: %ld, dirty_grant: %ld, "            \
1372                "reserved: %ld, flight: %d } lru {in list: %ld, "        \
1373                "left: %ld, waiters: %d }" fmt "\n",                     \
1374                cli_name(__tmp),                                         \
1375                __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages,        \
1376                atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1377                __tmp->cl_lost_grant, __tmp->cl_avail_grant,             \
1378                __tmp->cl_dirty_grant,                                   \
1379                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,         \
1380                atomic_long_read(&__tmp->cl_lru_in_list),                \
1381                atomic_long_read(&__tmp->cl_lru_busy),                   \
1382                atomic_read(&__tmp->cl_lru_shrinkers), ##args);          \
1383 } while (0)
1384
1385 /* caller must hold loi_list_lock */
1386 static void osc_consume_write_grant(struct client_obd *cli,
1387                                     struct brw_page *pga)
1388 {
1389         assert_spin_locked(&cli->cl_loi_list_lock);
1390         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1391         atomic_long_inc(&obd_dirty_pages);
1392         cli->cl_dirty_pages++;
1393         pga->flag |= OBD_BRW_FROM_GRANT;
1394         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1395                PAGE_SIZE, pga, pga->pg);
1396         osc_update_next_shrink(cli);
1397 }
1398
1399 /* the companion to osc_consume_write_grant, called when a brw has completed.
1400  * must be called with the loi lock held. */
1401 static void osc_release_write_grant(struct client_obd *cli,
1402                                     struct brw_page *pga)
1403 {
1404         ENTRY;
1405
1406         assert_spin_locked(&cli->cl_loi_list_lock);
1407         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1408                 EXIT;
1409                 return;
1410         }
1411
1412         pga->flag &= ~OBD_BRW_FROM_GRANT;
1413         atomic_long_dec(&obd_dirty_pages);
1414         cli->cl_dirty_pages--;
1415         if (pga->flag & OBD_BRW_NOCACHE) {
1416                 pga->flag &= ~OBD_BRW_NOCACHE;
1417                 atomic_long_dec(&obd_dirty_transit_pages);
1418                 cli->cl_dirty_transit--;
1419         }
1420         EXIT;
1421 }
1422
1423 /**
1424  * To avoid sleeping with object lock held, it's good for us allocate enough
1425  * grants before entering into critical section.
1426  *
1427  * client_obd_list_lock held by caller
1428  */
1429 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1430 {
1431         int rc = -EDQUOT;
1432
1433         if (cli->cl_avail_grant >= bytes) {
1434                 cli->cl_avail_grant    -= bytes;
1435                 cli->cl_reserved_grant += bytes;
1436                 rc = 0;
1437         }
1438         return rc;
1439 }
1440
1441 static void __osc_unreserve_grant(struct client_obd *cli,
1442                                   unsigned int reserved, unsigned int unused)
1443 {
1444         /* it's quite normal for us to get more grant than reserved.
1445          * Thinking about a case that two extents merged by adding a new
1446          * chunk, we can save one extent tax. If extent tax is greater than
1447          * one chunk, we can save more grant by adding a new chunk */
1448         cli->cl_reserved_grant -= reserved;
1449         if (unused > reserved) {
1450                 cli->cl_avail_grant += reserved;
1451                 cli->cl_lost_grant  += unused - reserved;
1452                 cli->cl_dirty_grant -= unused - reserved;
1453         } else {
1454                 cli->cl_avail_grant += unused;
1455                 cli->cl_dirty_grant += reserved - unused;
1456         }
1457 }
1458
1459 static void osc_unreserve_grant(struct client_obd *cli,
1460                                 unsigned int reserved, unsigned int unused)
1461 {
1462         spin_lock(&cli->cl_loi_list_lock);
1463         __osc_unreserve_grant(cli, reserved, unused);
1464         if (unused > 0)
1465                 osc_wake_cache_waiters(cli);
1466         spin_unlock(&cli->cl_loi_list_lock);
1467 }
1468
1469 /**
1470  * Free grant after IO is finished or canceled.
1471  *
1472  * @lost_grant is used to remember how many grants we have allocated but not
1473  * used, we should return these grants to OST. There're two cases where grants
1474  * can be lost:
1475  * 1. truncate;
1476  * 2. blocksize at OST is less than PAGE_SIZE and a partial page was
1477  *    written. In this case OST may use less chunks to serve this partial
1478  *    write. OSTs don't actually know the page size on the client side. so
1479  *    clients have to calculate lost grant by the blocksize on the OST.
1480  *    See filter_grant_check() for details.
1481  */
1482 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1483                            unsigned int lost_grant, unsigned int dirty_grant)
1484 {
1485         unsigned long grant;
1486
1487         grant = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
1488
1489         spin_lock(&cli->cl_loi_list_lock);
1490         atomic_long_sub(nr_pages, &obd_dirty_pages);
1491         cli->cl_dirty_pages -= nr_pages;
1492         cli->cl_lost_grant += lost_grant;
1493         cli->cl_dirty_grant -= dirty_grant;
1494         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1495                 /* borrow some grant from truncate to avoid the case that
1496                  * truncate uses up all avail grant */
1497                 cli->cl_lost_grant -= grant;
1498                 cli->cl_avail_grant += grant;
1499         }
1500         osc_wake_cache_waiters(cli);
1501         spin_unlock(&cli->cl_loi_list_lock);
1502         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu/%lu\n",
1503                lost_grant, cli->cl_lost_grant,
1504                cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_SHIFT,
1505                cli->cl_dirty_grant);
1506 }
1507
1508 /**
1509  * The companion to osc_enter_cache(), called when @oap is no longer part of
1510  * the dirty accounting due to error.
1511  */
1512 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1513 {
1514         spin_lock(&cli->cl_loi_list_lock);
1515         osc_release_write_grant(cli, &oap->oap_brw_page);
1516         spin_unlock(&cli->cl_loi_list_lock);
1517 }
1518
1519 /**
1520  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1521  * is available.
1522  */
1523 static int osc_enter_cache_try(struct client_obd *cli,
1524                                struct osc_async_page *oap,
1525                                int bytes, int transient)
1526 {
1527         int rc;
1528
1529         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1530
1531         rc = osc_reserve_grant(cli, bytes);
1532         if (rc < 0)
1533                 return 0;
1534
1535         if (cli->cl_dirty_pages < cli->cl_dirty_max_pages &&
1536             1 + atomic_long_read(&obd_dirty_pages) <= obd_max_dirty_pages) {
1537                 osc_consume_write_grant(cli, &oap->oap_brw_page);
1538                 if (transient) {
1539                         cli->cl_dirty_transit++;
1540                         atomic_long_inc(&obd_dirty_transit_pages);
1541                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
1542                 }
1543                 rc = 1;
1544         } else {
1545                 __osc_unreserve_grant(cli, bytes, bytes);
1546                 rc = 0;
1547         }
1548         return rc;
1549 }
1550
1551 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1552 {
1553         int rc;
1554         spin_lock(&cli->cl_loi_list_lock);
1555         rc = list_empty(&ocw->ocw_entry);
1556         spin_unlock(&cli->cl_loi_list_lock);
1557         return rc;
1558 }
1559
1560 /**
1561  * The main entry to reserve dirty page accounting. Usually the grant reserved
1562  * in this function will be freed in bulk in osc_free_grant() unless it fails
1563  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1564  *
1565  * The process will be put into sleep if it's already run out of grant.
1566  */
1567 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1568                            struct osc_async_page *oap, int bytes)
1569 {
1570         struct osc_object       *osc = oap->oap_obj;
1571         struct lov_oinfo        *loi = osc->oo_oinfo;
1572         struct osc_cache_waiter  ocw;
1573         struct l_wait_info       lwi;
1574         int                      rc = -EDQUOT;
1575         ENTRY;
1576
1577         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(AT_OFF ? obd_timeout : at_max),
1578                                NULL, LWI_ON_SIGNAL_NOOP, NULL);
1579
1580         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1581
1582         spin_lock(&cli->cl_loi_list_lock);
1583
1584         /* force the caller to try sync io.  this can jump the list
1585          * of queued writes and create a discontiguous rpc stream */
1586         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1587             cli->cl_dirty_max_pages == 0 ||
1588             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) {
1589                 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1590                 GOTO(out, rc = -EDQUOT);
1591         }
1592
1593         /* Hopefully normal case - cache space and write credits available */
1594         if (osc_enter_cache_try(cli, oap, bytes, 0)) {
1595                 OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1596                 GOTO(out, rc = 0);
1597         }
1598
1599         /* We can get here for two reasons: too many dirty pages in cache, or
1600          * run out of grants. In both cases we should write dirty pages out.
1601          * Adding a cache waiter will trigger urgent write-out no matter what
1602          * RPC size will be.
1603          * The exiting condition is no avail grants and no dirty pages caching,
1604          * that really means there is no space on the OST. */
1605         init_waitqueue_head(&ocw.ocw_waitq);
1606         ocw.ocw_oap   = oap;
1607         ocw.ocw_grant = bytes;
1608         while (cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0) {
1609                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1610                 ocw.ocw_rc = 0;
1611                 spin_unlock(&cli->cl_loi_list_lock);
1612
1613                 osc_io_unplug_async(env, cli, NULL);
1614
1615                 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1616                        cli_name(cli), &ocw, oap);
1617
1618                 rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1619
1620                 spin_lock(&cli->cl_loi_list_lock);
1621
1622                 if (rc < 0) {
1623                         /* l_wait_event is interrupted by signal or timed out */
1624                         list_del_init(&ocw.ocw_entry);
1625                         break;
1626                 }
1627                 LASSERT(list_empty(&ocw.ocw_entry));
1628                 rc = ocw.ocw_rc;
1629
1630                 if (rc != -EDQUOT)
1631                         break;
1632                 if (osc_enter_cache_try(cli, oap, bytes, 0)) {
1633                         rc = 0;
1634                         break;
1635                 }
1636         }
1637
1638         switch (rc) {
1639         case 0:
1640                 OSC_DUMP_GRANT(D_CACHE, cli, "finally got grant space\n");
1641                 break;
1642         case -ETIMEDOUT:
1643                 OSC_DUMP_GRANT(D_CACHE, cli,
1644                                "timeout, fall back to sync i/o\n");
1645                 osc_extent_tree_dump(D_CACHE, osc);
1646                 /* fall back to synchronous I/O */
1647                 rc = -EDQUOT;
1648                 break;
1649         case -EINTR:
1650                 /* Ensures restartability - LU-3581 */
1651                 OSC_DUMP_GRANT(D_CACHE, cli, "interrupted\n");
1652                 rc = -ERESTARTSYS;
1653                 break;
1654         case -EDQUOT:
1655                 OSC_DUMP_GRANT(D_CACHE, cli,
1656                                "no grant space, fall back to sync i/o\n");
1657                 break;
1658         default:
1659                 CDEBUG(D_CACHE, "%s: event for cache space @ %p never arrived "
1660                        "due to %d, fall back to sync i/o\n",
1661                        cli_name(cli), &ocw, rc);
1662                 break;
1663         }
1664         EXIT;
1665 out:
1666         spin_unlock(&cli->cl_loi_list_lock);
1667         RETURN(rc);
1668 }
1669
1670 /* caller must hold loi_list_lock */
1671 void osc_wake_cache_waiters(struct client_obd *cli)
1672 {
1673         struct list_head *l, *tmp;
1674         struct osc_cache_waiter *ocw;
1675
1676         ENTRY;
1677         list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1678                 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
1679                 list_del_init(&ocw->ocw_entry);
1680
1681                 ocw->ocw_rc = -EDQUOT;
1682                 /* we can't dirty more */
1683                 if ((cli->cl_dirty_pages  >= cli->cl_dirty_max_pages) ||
1684                     (1 + atomic_long_read(&obd_dirty_pages) >
1685                      obd_max_dirty_pages)) {
1686                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
1687                                "osc max %ld, sys max %ld\n",
1688                                cli->cl_dirty_pages, cli->cl_dirty_max_pages,
1689                                obd_max_dirty_pages);
1690                         goto wakeup;
1691                 }
1692
1693                 if (osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
1694                         ocw->ocw_rc = 0;
1695 wakeup:
1696                 CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld, %d\n",
1697                        ocw, ocw->ocw_oap, cli->cl_avail_grant, ocw->ocw_rc);
1698
1699                 wake_up(&ocw->ocw_waitq);
1700         }
1701
1702         EXIT;
1703 }
1704
1705 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1706 {
1707         int hprpc = !!list_empty(&osc->oo_hp_exts);
1708         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1709 }
1710
1711 /* This maintains the lists of pending pages to read/write for a given object
1712  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1713  * to quickly find objects that are ready to send an RPC. */
1714 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1715                          int cmd)
1716 {
1717         int invalid_import = 0;
1718         ENTRY;
1719
1720         /* if we have an invalid import we want to drain the queued pages
1721          * by forcing them through rpcs that immediately fail and complete
1722          * the pages.  recovery relies on this to empty the queued pages
1723          * before canceling the locks and evicting down the llite pages */
1724         if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1725                 invalid_import = 1;
1726
1727         if (cmd & OBD_BRW_WRITE) {
1728                 if (atomic_read(&osc->oo_nr_writes) == 0)
1729                         RETURN(0);
1730                 if (invalid_import) {
1731                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1732                         RETURN(1);
1733                 }
1734                 if (!list_empty(&osc->oo_hp_exts)) {
1735                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1736                         RETURN(1);
1737                 }
1738                 if (!list_empty(&osc->oo_urgent_exts)) {
1739                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1740                         RETURN(1);
1741                 }
1742                 /* trigger a write rpc stream as long as there are dirtiers
1743                  * waiting for space.  as they're waiting, they're not going to
1744                  * create more pages to coalesce with what's waiting.. */
1745                 if (!list_empty(&cli->cl_cache_waiters)) {
1746                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1747                         RETURN(1);
1748                 }
1749                 if (!list_empty(&osc->oo_full_exts)) {
1750                         CDEBUG(D_CACHE, "full extent ready, make an RPC\n");
1751                         RETURN(1);
1752                 }
1753         } else {
1754                 if (atomic_read(&osc->oo_nr_reads) == 0)
1755                         RETURN(0);
1756                 if (invalid_import) {
1757                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1758                         RETURN(1);
1759                 }
1760                 /* all read are urgent. */
1761                 if (!list_empty(&osc->oo_reading_exts))
1762                         RETURN(1);
1763         }
1764
1765         RETURN(0);
1766 }
1767
1768 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1769 {
1770         struct client_obd *cli = osc_cli(obj);
1771         if (cmd & OBD_BRW_WRITE) {
1772                 atomic_add(delta, &obj->oo_nr_writes);
1773                 atomic_add(delta, &cli->cl_pending_w_pages);
1774                 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1775         } else {
1776                 atomic_add(delta, &obj->oo_nr_reads);
1777                 atomic_add(delta, &cli->cl_pending_r_pages);
1778                 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1779         }
1780         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1781 }
1782
1783 static int osc_makes_hprpc(struct osc_object *obj)
1784 {
1785         return !list_empty(&obj->oo_hp_exts);
1786 }
1787
1788 static void on_list(struct list_head *item, struct list_head *list,
1789                     int should_be_on)
1790 {
1791         if (list_empty(item) && should_be_on)
1792                 list_add_tail(item, list);
1793         else if (!list_empty(item) && !should_be_on)
1794                 list_del_init(item);
1795 }
1796
1797 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1798  * can find pages to build into rpcs quickly */
1799 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1800 {
1801         if (osc_makes_hprpc(osc)) {
1802                 /* HP rpc */
1803                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1804                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1805         } else {
1806                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1807                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1808                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1809                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1810         }
1811
1812         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1813                 atomic_read(&osc->oo_nr_writes) > 0);
1814
1815         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1816                 atomic_read(&osc->oo_nr_reads) > 0);
1817
1818         return osc_is_ready(osc);
1819 }
1820
1821 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1822 {
1823         int is_ready;
1824
1825         spin_lock(&cli->cl_loi_list_lock);
1826         is_ready = __osc_list_maint(cli, osc);
1827         spin_unlock(&cli->cl_loi_list_lock);
1828
1829         return is_ready;
1830 }
1831
1832 /* this is trying to propogate async writeback errors back up to the
1833  * application.  As an async write fails we record the error code for later if
1834  * the app does an fsync.  As long as errors persist we force future rpcs to be
1835  * sync so that the app can get a sync error and break the cycle of queueing
1836  * pages for which writeback will fail. */
1837 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1838                            int rc)
1839 {
1840         if (rc) {
1841                 if (!ar->ar_rc)
1842                         ar->ar_rc = rc;
1843
1844                 ar->ar_force_sync = 1;
1845                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1846                 return;
1847
1848         }
1849
1850         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1851                 ar->ar_force_sync = 0;
1852 }
1853
1854 /* this must be called holding the loi list lock to give coverage to exit_cache,
1855  * async_flag maintenance, and oap_request */
1856 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1857                               struct osc_async_page *oap, int sent, int rc)
1858 {
1859         struct osc_object *osc = oap->oap_obj;
1860         struct lov_oinfo  *loi = osc->oo_oinfo;
1861         __u64 xid = 0;
1862
1863         ENTRY;
1864         if (oap->oap_request != NULL) {
1865                 xid = ptlrpc_req_xid(oap->oap_request);
1866                 ptlrpc_req_finished(oap->oap_request);
1867                 oap->oap_request = NULL;
1868         }
1869
1870         /* As the transfer for this page is being done, clear the flags */
1871         spin_lock(&oap->oap_lock);
1872         oap->oap_async_flags = 0;
1873         spin_unlock(&oap->oap_lock);
1874         oap->oap_interrupted = 0;
1875
1876         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1877                 spin_lock(&cli->cl_loi_list_lock);
1878                 osc_process_ar(&cli->cl_ar, xid, rc);
1879                 osc_process_ar(&loi->loi_ar, xid, rc);
1880                 spin_unlock(&cli->cl_loi_list_lock);
1881         }
1882
1883         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1884         if (rc)
1885                 CERROR("completion on oap %p obj %p returns %d.\n",
1886                        oap, osc, rc);
1887
1888         EXIT;
1889 }
1890
1891 struct extent_rpc_data {
1892         struct list_head        *erd_rpc_list;
1893         unsigned int            erd_page_count;
1894         unsigned int            erd_max_pages;
1895         unsigned int            erd_max_chunks;
1896         unsigned int            erd_max_extents;
1897 };
1898
1899 static inline unsigned osc_extent_chunks(const struct osc_extent *ext)
1900 {
1901         struct client_obd *cli = osc_cli(ext->oe_obj);
1902         unsigned ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1903
1904         return (ext->oe_end >> ppc_bits) - (ext->oe_start >> ppc_bits) + 1;
1905 }
1906
1907 /**
1908  * Try to add extent to one RPC. We need to think about the following things:
1909  * - # of pages must not be over max_pages_per_rpc
1910  * - extent must be compatible with previous ones
1911  */
1912 static int try_to_add_extent_for_io(struct client_obd *cli,
1913                                     struct osc_extent *ext,
1914                                     struct extent_rpc_data *data)
1915 {
1916         struct osc_extent *tmp;
1917         unsigned int chunk_count;
1918         struct osc_async_page *oap = list_first_entry(&ext->oe_pages,
1919                                                       struct osc_async_page,
1920                                                       oap_pending_item);
1921         ENTRY;
1922
1923         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1924                 ext);
1925         OSC_EXTENT_DUMP(D_CACHE, ext, "trying to add this extent\n");
1926
1927         if (data->erd_max_extents == 0)
1928                 RETURN(0);
1929
1930         chunk_count = osc_extent_chunks(ext);
1931         EASSERTF(data->erd_page_count != 0 ||
1932                  chunk_count <= data->erd_max_chunks, ext,
1933                  "The first extent to be fit in a RPC contains %u chunks, "
1934                  "which is over the limit %u.\n", chunk_count,
1935                  data->erd_max_chunks);
1936         if (chunk_count > data->erd_max_chunks)
1937                 RETURN(0);
1938
1939         data->erd_max_pages = max(ext->oe_mppr, data->erd_max_pages);
1940         EASSERTF(data->erd_page_count != 0 ||
1941                 ext->oe_nr_pages <= data->erd_max_pages, ext,
1942                 "The first extent to be fit in a RPC contains %u pages, "
1943                 "which is over the limit %u.\n", ext->oe_nr_pages,
1944                 data->erd_max_pages);
1945         if (data->erd_page_count + ext->oe_nr_pages > data->erd_max_pages)
1946                 RETURN(0);
1947
1948         list_for_each_entry(tmp, data->erd_rpc_list, oe_link) {
1949                 struct osc_async_page *oap2;
1950                 oap2 = list_first_entry(&tmp->oe_pages, struct osc_async_page,
1951                                         oap_pending_item);
1952                 EASSERT(tmp->oe_owner == current, tmp);
1953 #if 0
1954                 if (overlapped(tmp, ext)) {
1955                         OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
1956                         EASSERT(0, ext);
1957                 }
1958 #endif
1959                 if (oap2cl_page(oap)->cp_type != oap2cl_page(oap2)->cp_type) {
1960                         CDEBUG(D_CACHE, "Do not permit different types of IO "
1961                                "in one RPC\n");
1962                         RETURN(0);
1963                 }
1964
1965                 if (tmp->oe_srvlock != ext->oe_srvlock ||
1966                     !tmp->oe_grants != !ext->oe_grants ||
1967                     tmp->oe_no_merge || ext->oe_no_merge)
1968                         RETURN(0);
1969
1970                 /* remove break for strict check */
1971                 break;
1972         }
1973
1974         data->erd_max_extents--;
1975         data->erd_max_chunks -= chunk_count;
1976         data->erd_page_count += ext->oe_nr_pages;
1977         list_move_tail(&ext->oe_link, data->erd_rpc_list);
1978         ext->oe_owner = current;
1979         RETURN(1);
1980 }
1981
1982 static inline unsigned osc_max_write_chunks(const struct client_obd *cli)
1983 {
1984         /*
1985          * LU-8135:
1986          *
1987          * The maximum size of a single transaction is about 64MB in ZFS.
1988          * #define DMU_MAX_ACCESS (64 * 1024 * 1024)
1989          *
1990          * Since ZFS is a copy-on-write file system, a single dirty page in
1991          * a chunk will result in the rewrite of the whole chunk, therefore
1992          * an RPC shouldn't be allowed to contain too many chunks otherwise
1993          * it will make transaction size much bigger than 64MB, especially
1994          * with big block size for ZFS.
1995          *
1996          * This piece of code is to make sure that OSC won't send write RPCs
1997          * with too many chunks. The maximum chunk size that an RPC can cover
1998          * is set to PTLRPC_MAX_BRW_SIZE, which is defined to 16MB. Ideally
1999          * OST should tell the client what the biggest transaction size is,
2000          * but it's good enough for now.
2001          *
2002          * This limitation doesn't apply to ldiskfs, which allows as many
2003          * chunks in one RPC as we want. However, it won't have any benefits
2004          * to have too many discontiguous pages in one RPC.
2005          *
2006          * An osc_extent won't cover over a RPC size, so the chunks in an
2007          * osc_extent won't bigger than PTLRPC_MAX_BRW_SIZE >> chunkbits.
2008          */
2009         return PTLRPC_MAX_BRW_SIZE >> cli->cl_chunkbits;
2010 }
2011
2012 /**
2013  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
2014  * get_write_extent() takes all appropriate extents in atomic.
2015  *
2016  * The following policy is used to collect extents for IO:
2017  * 1. Add as many HP extents as possible;
2018  * 2. Add the first urgent extent in urgent extent list and take it out of
2019  *    urgent list;
2020  * 3. Add subsequent extents of this urgent extent;
2021  * 4. If urgent list is not empty, goto 2;
2022  * 5. Traverse the extent tree from the 1st extent;
2023  * 6. Above steps exit if there is no space in this RPC.
2024  */
2025 static unsigned int get_write_extents(struct osc_object *obj,
2026                                       struct list_head *rpclist)
2027 {
2028         struct client_obd *cli = osc_cli(obj);
2029         struct osc_extent *ext;
2030         struct extent_rpc_data data = {
2031                 .erd_rpc_list   = rpclist,
2032                 .erd_page_count = 0,
2033                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
2034                 .erd_max_chunks = osc_max_write_chunks(cli),
2035                 .erd_max_extents = 256,
2036         };
2037
2038         LASSERT(osc_object_is_locked(obj));
2039         while (!list_empty(&obj->oo_hp_exts)) {
2040                 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
2041                                  oe_link);
2042                 LASSERT(ext->oe_state == OES_CACHE);
2043                 if (!try_to_add_extent_for_io(cli, ext, &data))
2044                         return data.erd_page_count;
2045                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2046         }
2047         if (data.erd_page_count == data.erd_max_pages)
2048                 return data.erd_page_count;
2049
2050         while (!list_empty(&obj->oo_urgent_exts)) {
2051                 ext = list_entry(obj->oo_urgent_exts.next,
2052                                  struct osc_extent, oe_link);
2053                 if (!try_to_add_extent_for_io(cli, ext, &data))
2054                         return data.erd_page_count;
2055         }
2056         if (data.erd_page_count == data.erd_max_pages)
2057                 return data.erd_page_count;
2058
2059         /* One key difference between full extents and other extents: full
2060          * extents can usually only be added if the rpclist was empty, so if we
2061          * can't add one, we continue on to trying to add normal extents.  This
2062          * is so we don't miss adding extra extents to an RPC containing high
2063          * priority or urgent extents. */
2064         while (!list_empty(&obj->oo_full_exts)) {
2065                 ext = list_entry(obj->oo_full_exts.next,
2066                                  struct osc_extent, oe_link);
2067                 if (!try_to_add_extent_for_io(cli, ext, &data))
2068                         break;
2069         }
2070         if (data.erd_page_count == data.erd_max_pages)
2071                 return data.erd_page_count;
2072
2073         ext = first_extent(obj);
2074         while (ext != NULL) {
2075                 if ((ext->oe_state != OES_CACHE) ||
2076                     /* this extent may be already in current rpclist */
2077                     (!list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
2078                         ext = next_extent(ext);
2079                         continue;
2080                 }
2081
2082                 if (!try_to_add_extent_for_io(cli, ext, &data))
2083                         return data.erd_page_count;
2084
2085                 ext = next_extent(ext);
2086         }
2087         return data.erd_page_count;
2088 }
2089
2090 static int
2091 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
2092                    struct osc_object *osc)
2093 __must_hold(osc)
2094 {
2095         struct list_head   rpclist = LIST_HEAD_INIT(rpclist);
2096         struct osc_extent *ext;
2097         struct osc_extent *tmp;
2098         struct osc_extent *first = NULL;
2099         unsigned int page_count = 0;
2100         int srvlock = 0;
2101         int rc = 0;
2102         ENTRY;
2103
2104         LASSERT(osc_object_is_locked(osc));
2105
2106         page_count = get_write_extents(osc, &rpclist);
2107         LASSERT(equi(page_count == 0, list_empty(&rpclist)));
2108
2109         if (list_empty(&rpclist))
2110                 RETURN(0);
2111
2112         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2113
2114         list_for_each_entry(ext, &rpclist, oe_link) {
2115                 LASSERT(ext->oe_state == OES_CACHE ||
2116                         ext->oe_state == OES_LOCK_DONE);
2117                 if (ext->oe_state == OES_CACHE)
2118                         osc_extent_state_set(ext, OES_LOCKING);
2119                 else
2120                         osc_extent_state_set(ext, OES_RPC);
2121         }
2122
2123         /* we're going to grab page lock, so release object lock because
2124          * lock order is page lock -> object lock. */
2125         osc_object_unlock(osc);
2126
2127         list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2128                 if (ext->oe_state == OES_LOCKING) {
2129                         rc = osc_extent_make_ready(env, ext);
2130                         if (unlikely(rc < 0)) {
2131                                 list_del_init(&ext->oe_link);
2132                                 osc_extent_finish(env, ext, 0, rc);
2133                                 continue;
2134                         }
2135                 }
2136                 if (first == NULL) {
2137                         first = ext;
2138                         srvlock = ext->oe_srvlock;
2139                 } else {
2140                         LASSERT(srvlock == ext->oe_srvlock);
2141                 }
2142         }
2143
2144         if (!list_empty(&rpclist)) {
2145                 LASSERT(page_count > 0);
2146                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2147                 LASSERT(list_empty(&rpclist));
2148         }
2149
2150         osc_object_lock(osc);
2151         RETURN(rc);
2152 }
2153
2154 /**
2155  * prepare pages for ASYNC io and put pages in send queue.
2156  *
2157  * \param cmd OBD_BRW_* macroses
2158  * \param lop pending pages
2159  *
2160  * \return zero if no page added to send queue.
2161  * \return 1 if pages successfully added to send queue.
2162  * \return negative on errors.
2163  */
2164 static int
2165 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2166                   struct osc_object *osc)
2167 __must_hold(osc)
2168 {
2169         struct osc_extent *ext;
2170         struct osc_extent *next;
2171         struct list_head rpclist = LIST_HEAD_INIT(rpclist);
2172         struct extent_rpc_data data = {
2173                 .erd_rpc_list   = &rpclist,
2174                 .erd_page_count = 0,
2175                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
2176                 .erd_max_chunks = UINT_MAX,
2177                 .erd_max_extents = UINT_MAX,
2178         };
2179         int rc = 0;
2180         ENTRY;
2181
2182         LASSERT(osc_object_is_locked(osc));
2183         list_for_each_entry_safe(ext, next, &osc->oo_reading_exts, oe_link) {
2184                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2185                 if (!try_to_add_extent_for_io(cli, ext, &data))
2186                         break;
2187                 osc_extent_state_set(ext, OES_RPC);
2188                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2189         }
2190         LASSERT(data.erd_page_count <= data.erd_max_pages);
2191
2192         osc_update_pending(osc, OBD_BRW_READ, -data.erd_page_count);
2193
2194         if (!list_empty(&rpclist)) {
2195                 osc_object_unlock(osc);
2196
2197                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2198                 LASSERT(list_empty(&rpclist));
2199
2200                 osc_object_lock(osc);
2201         }
2202         RETURN(rc);
2203 }
2204
2205 #define list_to_obj(list, item) ({                                            \
2206         struct list_head *__tmp = (list)->next;                               \
2207         list_del_init(__tmp);                                         \
2208         list_entry(__tmp, struct osc_object, oo_##item);                      \
2209 })
2210
2211 /* This is called by osc_check_rpcs() to find which objects have pages that
2212  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2213 static struct osc_object *osc_next_obj(struct client_obd *cli)
2214 {
2215         ENTRY;
2216
2217         /* First return objects that have blocked locks so that they
2218          * will be flushed quickly and other clients can get the lock,
2219          * then objects which have pages ready to be stuffed into RPCs */
2220         if (!list_empty(&cli->cl_loi_hp_ready_list))
2221                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2222         if (!list_empty(&cli->cl_loi_ready_list))
2223                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2224
2225         /* then if we have cache waiters, return all objects with queued
2226          * writes.  This is especially important when many small files
2227          * have filled up the cache and not been fired into rpcs because
2228          * they don't pass the nr_pending/object threshhold */
2229         if (!list_empty(&cli->cl_cache_waiters) &&
2230             !list_empty(&cli->cl_loi_write_list))
2231                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2232
2233         /* then return all queued objects when we have an invalid import
2234          * so that they get flushed */
2235         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2236                 if (!list_empty(&cli->cl_loi_write_list))
2237                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2238                                            write_item));
2239                 if (!list_empty(&cli->cl_loi_read_list))
2240                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2241                                            read_item));
2242         }
2243         RETURN(NULL);
2244 }
2245
2246 /* called with the loi list lock held */
2247 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2248 __must_hold(&cli->cl_loi_list_lock)
2249 {
2250         struct osc_object *osc;
2251         int rc = 0;
2252         ENTRY;
2253
2254         while ((osc = osc_next_obj(cli)) != NULL) {
2255                 struct cl_object *obj = osc2cl(osc);
2256                 struct lu_ref_link link;
2257
2258                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2259
2260                 if (osc_max_rpc_in_flight(cli, osc)) {
2261                         __osc_list_maint(cli, osc);
2262                         break;
2263                 }
2264
2265                 cl_object_get(obj);
2266                 spin_unlock(&cli->cl_loi_list_lock);
2267                 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2268
2269                 /* attempt some read/write balancing by alternating between
2270                  * reads and writes in an object.  The makes_rpc checks here
2271                  * would be redundant if we were getting read/write work items
2272                  * instead of objects.  we don't want send_oap_rpc to drain a
2273                  * partial read pending queue when we're given this object to
2274                  * do io on writes while there are cache waiters */
2275                 osc_object_lock(osc);
2276                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2277                         rc = osc_send_write_rpc(env, cli, osc);
2278                         if (rc < 0) {
2279                                 CERROR("Write request failed with %d\n", rc);
2280
2281                                 /* osc_send_write_rpc failed, mostly because of
2282                                  * memory pressure.
2283                                  *
2284                                  * It can't break here, because if:
2285                                  *  - a page was submitted by osc_io_submit, so
2286                                  *    page locked;
2287                                  *  - no request in flight
2288                                  *  - no subsequent request
2289                                  * The system will be in live-lock state,
2290                                  * because there is no chance to call
2291                                  * osc_io_unplug() and osc_check_rpcs() any
2292                                  * more. pdflush can't help in this case,
2293                                  * because it might be blocked at grabbing
2294                                  * the page lock as we mentioned.
2295                                  *
2296                                  * Anyway, continue to drain pages. */
2297                                 /* break; */
2298                         }
2299                 }
2300                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2301                         rc = osc_send_read_rpc(env, cli, osc);
2302                         if (rc < 0)
2303                                 CERROR("Read request failed with %d\n", rc);
2304                 }
2305                 osc_object_unlock(osc);
2306
2307                 osc_list_maint(cli, osc);
2308                 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2309                 cl_object_put(env, obj);
2310
2311                 spin_lock(&cli->cl_loi_list_lock);
2312         }
2313 }
2314
2315 static int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2316                           struct osc_object *osc, int async)
2317 {
2318         int rc = 0;
2319
2320         if (osc != NULL && osc_list_maint(cli, osc) == 0)
2321                 return 0;
2322
2323         if (!async) {
2324                 spin_lock(&cli->cl_loi_list_lock);
2325                 osc_check_rpcs(env, cli);
2326                 spin_unlock(&cli->cl_loi_list_lock);
2327         } else {
2328                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2329                 LASSERT(cli->cl_writeback_work != NULL);
2330                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2331         }
2332         return rc;
2333 }
2334
2335 static int osc_io_unplug_async(const struct lu_env *env,
2336                                 struct client_obd *cli, struct osc_object *osc)
2337 {
2338         return osc_io_unplug0(env, cli, osc, 1);
2339 }
2340
2341 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
2342                    struct osc_object *osc)
2343 {
2344         (void)osc_io_unplug0(env, cli, osc, 0);
2345 }
2346
2347 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2348                         struct page *page, loff_t offset)
2349 {
2350         struct obd_export     *exp = osc_export(osc);
2351         struct osc_async_page *oap = &ops->ops_oap;
2352         ENTRY;
2353
2354         if (!page)
2355                 return cfs_size_round(sizeof(*oap));
2356
2357         oap->oap_magic = OAP_MAGIC;
2358         oap->oap_cli = &exp->exp_obd->u.cli;
2359         oap->oap_obj = osc;
2360
2361         oap->oap_page = page;
2362         oap->oap_obj_off = offset;
2363         LASSERT(!(offset & ~PAGE_MASK));
2364
2365         if (cfs_capable(CFS_CAP_SYS_RESOURCE))
2366                 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2367
2368         INIT_LIST_HEAD(&oap->oap_pending_item);
2369         INIT_LIST_HEAD(&oap->oap_rpc_item);
2370
2371         spin_lock_init(&oap->oap_lock);
2372         CDEBUG(D_INFO, "oap %p page %p obj off %llu\n",
2373                oap, page, oap->oap_obj_off);
2374         RETURN(0);
2375 }
2376
2377 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2378                        struct osc_page *ops)
2379 {
2380         struct osc_io *oio = osc_env_io(env);
2381         struct osc_extent     *ext = NULL;
2382         struct osc_async_page *oap = &ops->ops_oap;
2383         struct client_obd     *cli = oap->oap_cli;
2384         struct osc_object     *osc = oap->oap_obj;
2385         pgoff_t index;
2386         unsigned int tmp;
2387         unsigned int grants = 0;
2388         u32    brw_flags = OBD_BRW_ASYNC;
2389         int    cmd = OBD_BRW_WRITE;
2390         int    need_release = 0;
2391         int    rc = 0;
2392         ENTRY;
2393
2394         if (oap->oap_magic != OAP_MAGIC)
2395                 RETURN(-EINVAL);
2396
2397         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2398                 RETURN(-EIO);
2399
2400         if (!list_empty(&oap->oap_pending_item) ||
2401             !list_empty(&oap->oap_rpc_item))
2402                 RETURN(-EBUSY);
2403
2404         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2405         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2406         if (cfs_capable(CFS_CAP_SYS_RESOURCE)) {
2407                 brw_flags |= OBD_BRW_NOQUOTA;
2408                 cmd |= OBD_BRW_NOQUOTA;
2409         }
2410
2411         /* check if the file's owner/group is over quota */
2412         if (!(cmd & OBD_BRW_NOQUOTA)) {
2413                 struct cl_object *obj;
2414                 struct cl_attr   *attr;
2415                 unsigned int qid[LL_MAXQUOTAS];
2416
2417                 obj = cl_object_top(&osc->oo_cl);
2418                 attr = &osc_env_info(env)->oti_attr;
2419
2420                 cl_object_attr_lock(obj);
2421                 rc = cl_object_attr_get(env, obj, attr);
2422                 cl_object_attr_unlock(obj);
2423
2424                 qid[USRQUOTA] = attr->cat_uid;
2425                 qid[GRPQUOTA] = attr->cat_gid;
2426                 qid[PRJQUOTA] = attr->cat_projid;
2427                 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2428                         rc = -EDQUOT;
2429                 if (rc)
2430                         RETURN(rc);
2431         }
2432
2433         oap->oap_cmd = cmd;
2434         oap->oap_page_off = ops->ops_from;
2435         oap->oap_count = ops->ops_to - ops->ops_from;
2436         /* No need to hold a lock here,
2437          * since this page is not in any list yet. */
2438         oap->oap_async_flags = 0;
2439         oap->oap_brw_flags = brw_flags;
2440
2441         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2442                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2443
2444         index = osc_index(oap2osc(oap));
2445
2446         /* Add this page into extent by the following steps:
2447          * 1. if there exists an active extent for this IO, mostly this page
2448          *    can be added to the active extent and sometimes we need to
2449          *    expand extent to accomodate this page;
2450          * 2. otherwise, a new extent will be allocated. */
2451
2452         ext = oio->oi_active;
2453         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2454                 /* one chunk plus extent overhead must be enough to write this
2455                  * page */
2456                 grants = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2457                 if (ext->oe_end >= index)
2458                         grants = 0;
2459
2460                 /* it doesn't need any grant to dirty this page */
2461                 spin_lock(&cli->cl_loi_list_lock);
2462                 rc = osc_enter_cache_try(cli, oap, grants, 0);
2463                 spin_unlock(&cli->cl_loi_list_lock);
2464                 if (rc == 0) { /* try failed */
2465                         grants = 0;
2466                         need_release = 1;
2467                 } else if (ext->oe_end < index) {
2468                         tmp = grants;
2469                         /* try to expand this extent */
2470                         rc = osc_extent_expand(ext, index, &tmp);
2471                         if (rc < 0) {
2472                                 need_release = 1;
2473                                 /* don't free reserved grant */
2474                         } else {
2475                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2476                                                 "expanded for %lu.\n", index);
2477                                 osc_unreserve_grant(cli, grants, tmp);
2478                                 grants = 0;
2479                         }
2480                 }
2481                 rc = 0;
2482         } else if (ext != NULL) {
2483                 /* index is located outside of active extent */
2484                 need_release = 1;
2485         }
2486         if (need_release) {
2487                 osc_extent_release(env, ext);
2488                 oio->oi_active = NULL;
2489                 ext = NULL;
2490         }
2491
2492         if (ext == NULL) {
2493                 tmp = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2494
2495                 /* try to find new extent to cover this page */
2496                 LASSERT(oio->oi_active == NULL);
2497                 /* we may have allocated grant for this page if we failed
2498                  * to expand the previous active extent. */
2499                 LASSERT(ergo(grants > 0, grants >= tmp));
2500
2501                 rc = 0;
2502                 if (grants == 0) {
2503                         /* we haven't allocated grant for this page. */
2504                         rc = osc_enter_cache(env, cli, oap, tmp);
2505                         if (rc == 0)
2506                                 grants = tmp;
2507                 }
2508
2509                 tmp = grants;
2510                 if (rc == 0) {
2511                         ext = osc_extent_find(env, osc, index, &tmp);
2512                         if (IS_ERR(ext)) {
2513                                 LASSERT(tmp == grants);
2514                                 osc_exit_cache(cli, oap);
2515                                 rc = PTR_ERR(ext);
2516                                 ext = NULL;
2517                         } else {
2518                                 oio->oi_active = ext;
2519                         }
2520                 }
2521                 if (grants > 0)
2522                         osc_unreserve_grant(cli, grants, tmp);
2523         }
2524
2525         LASSERT(ergo(rc == 0, ext != NULL));
2526         if (ext != NULL) {
2527                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2528                          ext, "index = %lu.\n", index);
2529                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2530
2531                 osc_object_lock(osc);
2532                 if (ext->oe_nr_pages == 0)
2533                         ext->oe_srvlock = ops->ops_srvlock;
2534                 else
2535                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2536                 ++ext->oe_nr_pages;
2537                 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2538                 osc_object_unlock(osc);
2539         }
2540         RETURN(rc);
2541 }
2542
2543 int osc_teardown_async_page(const struct lu_env *env,
2544                             struct osc_object *obj, struct osc_page *ops)
2545 {
2546         struct osc_async_page *oap = &ops->ops_oap;
2547         int rc = 0;
2548         ENTRY;
2549
2550         LASSERT(oap->oap_magic == OAP_MAGIC);
2551
2552         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2553                oap, ops, osc_index(oap2osc(oap)));
2554
2555         if (!list_empty(&oap->oap_rpc_item)) {
2556                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2557                 rc = -EBUSY;
2558         } else if (!list_empty(&oap->oap_pending_item)) {
2559                 struct osc_extent *ext = NULL;
2560
2561                 osc_object_lock(obj);
2562                 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2563                 osc_object_unlock(obj);
2564                 /* only truncated pages are allowed to be taken out.
2565                  * See osc_extent_truncate() and osc_cache_truncate_start()
2566                  * for details. */
2567                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2568                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2569                                         osc_index(oap2osc(oap)));
2570                         rc = -EBUSY;
2571                 }
2572                 if (ext != NULL)
2573                         osc_extent_put(env, ext);
2574         }
2575         RETURN(rc);
2576 }
2577
2578 /**
2579  * This is called when a page is picked up by kernel to write out.
2580  *
2581  * We should find out the corresponding extent and add the whole extent
2582  * into urgent list. The extent may be being truncated or used, handle it
2583  * carefully.
2584  */
2585 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2586                          struct osc_page *ops)
2587 {
2588         struct osc_extent *ext   = NULL;
2589         struct osc_object *obj   = cl2osc(ops->ops_cl.cpl_obj);
2590         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2591         pgoff_t            index = osc_index(ops);
2592         struct osc_async_page *oap = &ops->ops_oap;
2593         bool unplug = false;
2594         int rc = 0;
2595         ENTRY;
2596
2597         osc_object_lock(obj);
2598         ext = osc_extent_lookup(obj, index);
2599         if (ext == NULL) {
2600                 osc_extent_tree_dump(D_ERROR, obj);
2601                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2602         }
2603
2604         switch (ext->oe_state) {
2605         case OES_RPC:
2606         case OES_LOCK_DONE:
2607                 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2608                 LASSERT(0);
2609                 break;
2610         case OES_LOCKING:
2611                 /* If we know this extent is being written out, we should abort
2612                  * so that the writer can make this page ready. Otherwise, there
2613                  * exists a deadlock problem because other process can wait for
2614                  * page writeback bit holding page lock; and meanwhile in
2615                  * vvp_page_make_ready(), we need to grab page lock before
2616                  * really sending the RPC. */
2617         case OES_TRUNC:
2618                 /* race with truncate, page will be redirtied */
2619         case OES_ACTIVE:
2620                 /* The extent is active so we need to abort and let the caller
2621                  * re-dirty the page. If we continued on here, and we were the
2622                  * one making the extent active, we could deadlock waiting for
2623                  * the page writeback to clear but it won't because the extent
2624                  * is active and won't be written out. */
2625                 GOTO(out, rc = -EAGAIN);
2626         default:
2627                 break;
2628         }
2629
2630         rc = cl_page_prep(env, io, cp, CRT_WRITE);
2631         if (rc)
2632                 GOTO(out, rc);
2633
2634         spin_lock(&oap->oap_lock);
2635         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2636         spin_unlock(&oap->oap_lock);
2637
2638         if (memory_pressure_get())
2639                 ext->oe_memalloc = 1;
2640
2641         ext->oe_urgent = 1;
2642         if (ext->oe_state == OES_CACHE) {
2643                 OSC_EXTENT_DUMP(D_CACHE, ext,
2644                                 "flush page %p make it urgent.\n", oap);
2645                 if (list_empty(&ext->oe_link))
2646                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2647                 unplug = true;
2648         }
2649         rc = 0;
2650         EXIT;
2651
2652 out:
2653         osc_object_unlock(obj);
2654         osc_extent_put(env, ext);
2655         if (unplug)
2656                 osc_io_unplug_async(env, osc_cli(obj), obj);
2657         return rc;
2658 }
2659
2660 /**
2661  * this is called when a sync waiter receives an interruption.  Its job is to
2662  * get the caller woken as soon as possible.  If its page hasn't been put in an
2663  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2664  * desiring interruption which will forcefully complete the rpc once the rpc
2665  * has timed out.
2666  */
2667 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2668 {
2669         struct osc_async_page *oap = &ops->ops_oap;
2670         struct osc_object     *obj = oap->oap_obj;
2671         struct client_obd     *cli = osc_cli(obj);
2672         struct osc_extent     *ext;
2673         struct osc_extent     *found = NULL;
2674         struct list_head            *plist;
2675         pgoff_t index = osc_index(ops);
2676         int     rc = -EBUSY;
2677         int     cmd;
2678         ENTRY;
2679
2680         LASSERT(!oap->oap_interrupted);
2681         oap->oap_interrupted = 1;
2682
2683         /* Find out the caching extent */
2684         osc_object_lock(obj);
2685         if (oap->oap_cmd & OBD_BRW_WRITE) {
2686                 plist = &obj->oo_urgent_exts;
2687                 cmd   = OBD_BRW_WRITE;
2688         } else {
2689                 plist = &obj->oo_reading_exts;
2690                 cmd   = OBD_BRW_READ;
2691         }
2692         list_for_each_entry(ext, plist, oe_link) {
2693                 if (ext->oe_start <= index && ext->oe_end >= index) {
2694                         LASSERT(ext->oe_state == OES_LOCK_DONE);
2695                         /* For OES_LOCK_DONE state extent, it has already held
2696                          * a refcount for RPC. */
2697                         found = osc_extent_get(ext);
2698                         break;
2699                 }
2700         }
2701         if (found != NULL) {
2702                 list_del_init(&found->oe_link);
2703                 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2704                 osc_object_unlock(obj);
2705
2706                 osc_extent_finish(env, found, 0, -EINTR);
2707                 osc_extent_put(env, found);
2708                 rc = 0;
2709         } else {
2710                 osc_object_unlock(obj);
2711                 /* ok, it's been put in an rpc. only one oap gets a request
2712                  * reference */
2713                 if (oap->oap_request != NULL) {
2714                         ptlrpc_mark_interrupted(oap->oap_request);
2715                         ptlrpcd_wake(oap->oap_request);
2716                         ptlrpc_req_finished(oap->oap_request);
2717                         oap->oap_request = NULL;
2718                 }
2719         }
2720
2721         osc_list_maint(cli, obj);
2722         RETURN(rc);
2723 }
2724
2725 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
2726                          struct list_head *list, int cmd, int brw_flags)
2727 {
2728         struct client_obd     *cli = osc_cli(obj);
2729         struct osc_extent     *ext;
2730         struct osc_async_page *oap;
2731         int     page_count = 0;
2732         int     mppr       = cli->cl_max_pages_per_rpc;
2733         bool    can_merge   = true;
2734         pgoff_t start      = CL_PAGE_EOF;
2735         pgoff_t end        = 0;
2736         ENTRY;
2737
2738         list_for_each_entry(oap, list, oap_pending_item) {
2739                 struct osc_page *opg = oap2osc_page(oap);
2740                 pgoff_t index = osc_index(opg);
2741
2742                 if (index > end)
2743                         end = index;
2744                 if (index < start)
2745                         start = index;
2746                 ++page_count;
2747                 mppr <<= (page_count > mppr);
2748
2749                 if (unlikely(opg->ops_from > 0 || opg->ops_to < PAGE_SIZE))
2750                         can_merge = false;
2751         }
2752
2753         ext = osc_extent_alloc(obj);
2754         if (ext == NULL) {
2755                 struct osc_async_page *tmp;
2756
2757                 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2758                         list_del_init(&oap->oap_pending_item);
2759                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2760                 }
2761                 RETURN(-ENOMEM);
2762         }
2763
2764         ext->oe_rw = !!(cmd & OBD_BRW_READ);
2765         ext->oe_sync = 1;
2766         ext->oe_no_merge = !can_merge;
2767         ext->oe_urgent = 1;
2768         ext->oe_start = start;
2769         ext->oe_end = ext->oe_max_end = end;
2770         ext->oe_obj = obj;
2771         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2772         ext->oe_nr_pages = page_count;
2773         ext->oe_mppr = mppr;
2774         list_splice_init(list, &ext->oe_pages);
2775
2776         osc_object_lock(obj);
2777         /* Reuse the initial refcount for RPC, don't drop it */
2778         osc_extent_state_set(ext, OES_LOCK_DONE);
2779         if (cmd & OBD_BRW_WRITE) {
2780                 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2781                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2782         } else {
2783                 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2784                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2785         }
2786         osc_object_unlock(obj);
2787
2788         osc_io_unplug_async(env, cli, obj);
2789         RETURN(0);
2790 }
2791
2792 /**
2793  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2794  */
2795 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
2796                              __u64 size, struct osc_extent **extp)
2797 {
2798         struct client_obd *cli = osc_cli(obj);
2799         struct osc_extent *ext;
2800         struct osc_extent *waiting = NULL;
2801         pgoff_t index;
2802         struct list_head list = LIST_HEAD_INIT(list);
2803         int result = 0;
2804         bool partial;
2805         ENTRY;
2806
2807         /* pages with index greater or equal to index will be truncated. */
2808         index = cl_index(osc2cl(obj), size);
2809         partial = size > cl_offset(osc2cl(obj), index);
2810
2811 again:
2812         osc_object_lock(obj);
2813         ext = osc_extent_search(obj, index);
2814         if (ext == NULL)
2815                 ext = first_extent(obj);
2816         else if (ext->oe_end < index)
2817                 ext = next_extent(ext);
2818         while (ext != NULL) {
2819                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2820
2821                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2822                         /* if ext is in urgent state, it means there must exist
2823                          * a page already having been flushed by write_page().
2824                          * We have to wait for this extent because we can't
2825                          * truncate that page. */
2826                         OSC_EXTENT_DUMP(D_CACHE, ext,
2827                                         "waiting for busy extent\n");
2828                         waiting = osc_extent_get(ext);
2829                         break;
2830                 }
2831
2832                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:%llu.\n", size);
2833
2834                 osc_extent_get(ext);
2835                 if (ext->oe_state == OES_ACTIVE) {
2836                         /* though we grab inode mutex for write path, but we
2837                          * release it before releasing extent(in osc_io_end()),
2838                          * so there is a race window that an extent is still
2839                          * in OES_ACTIVE when truncate starts. */
2840                         LASSERT(!ext->oe_trunc_pending);
2841                         ext->oe_trunc_pending = 1;
2842                 } else {
2843                         EASSERT(ext->oe_state == OES_CACHE, ext);
2844                         osc_extent_state_set(ext, OES_TRUNC);
2845                         osc_update_pending(obj, OBD_BRW_WRITE,
2846                                            -ext->oe_nr_pages);
2847                 }
2848                 /* This extent could be on the full extents list, that's OK */
2849                 EASSERT(!ext->oe_hp && !ext->oe_urgent, ext);
2850                 if (!list_empty(&ext->oe_link))
2851                         list_move_tail(&ext->oe_link, &list);
2852                 else
2853                         list_add_tail(&ext->oe_link, &list);
2854
2855                 ext = next_extent(ext);
2856         }
2857         osc_object_unlock(obj);
2858
2859         osc_list_maint(cli, obj);
2860
2861         while (!list_empty(&list)) {
2862                 int rc;
2863
2864                 ext = list_entry(list.next, struct osc_extent, oe_link);
2865                 list_del_init(&ext->oe_link);
2866
2867                 /* extent may be in OES_ACTIVE state because inode mutex
2868                  * is released before osc_io_end() in file write case */
2869                 if (ext->oe_state != OES_TRUNC)
2870                         osc_extent_wait(env, ext, OES_TRUNC);
2871
2872                 rc = osc_extent_truncate(ext, index, partial);
2873                 if (rc < 0) {
2874                         if (result == 0)
2875                                 result = rc;
2876
2877                         OSC_EXTENT_DUMP(D_ERROR, ext,
2878                                         "truncate error %d\n", rc);
2879                 } else if (ext->oe_nr_pages == 0) {
2880                         osc_extent_remove(ext);
2881                 } else {
2882                         /* this must be an overlapped extent which means only
2883                          * part of pages in this extent have been truncated.
2884                          */
2885                         EASSERTF(ext->oe_start <= index, ext,
2886                                  "trunc index = %lu/%d.\n", index, partial);
2887                         /* fix index to skip this partially truncated extent */
2888                         index = ext->oe_end + 1;
2889                         partial = false;
2890
2891                         /* we need to hold this extent in OES_TRUNC state so
2892                          * that no writeback will happen. This is to avoid
2893                          * BUG 17397.
2894                          * Only partial truncate can reach here, if @size is
2895                          * not zero, the caller should provide a valid @extp. */
2896                         LASSERT(*extp == NULL);
2897                         *extp = osc_extent_get(ext);
2898                         OSC_EXTENT_DUMP(D_CACHE, ext,
2899                                         "trunc at %llu\n", size);
2900                 }
2901                 osc_extent_put(env, ext);
2902         }
2903         if (waiting != NULL) {
2904                 int rc;
2905
2906                 /* ignore the result of osc_extent_wait the write initiator
2907                  * should take care of it. */
2908                 rc = osc_extent_wait(env, waiting, OES_INV);
2909                 if (rc < 0)
2910                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2911
2912                 osc_extent_put(env, waiting);
2913                 waiting = NULL;
2914                 goto again;
2915         }
2916         RETURN(result);
2917 }
2918
2919 /**
2920  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2921  */
2922 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext)
2923 {
2924         if (ext != NULL) {
2925                 struct osc_object *obj = ext->oe_obj;
2926                 bool unplug = false;
2927
2928                 EASSERT(ext->oe_nr_pages > 0, ext);
2929                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2930                 EASSERT(!ext->oe_urgent, ext);
2931
2932                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2933                 osc_object_lock(obj);
2934                 osc_extent_state_set(ext, OES_CACHE);
2935                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2936                         ext->oe_urgent = 1;
2937                         list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2938                         unplug = true;
2939                 }
2940                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2941                 osc_object_unlock(obj);
2942                 osc_extent_put(env, ext);
2943
2944                 if (unplug)
2945                         osc_io_unplug_async(env, osc_cli(obj), obj);
2946         }
2947 }
2948
2949 /**
2950  * Wait for extents in a specific range to be written out.
2951  * The caller must have called osc_cache_writeback_range() to issue IO
2952  * otherwise it will take a long time for this function to finish.
2953  *
2954  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2955  * nobody else can dirty this range of file while we're waiting for
2956  * extents to be written.
2957  */
2958 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2959                          pgoff_t start, pgoff_t end)
2960 {
2961         struct osc_extent *ext;
2962         pgoff_t index = start;
2963         int     result = 0;
2964         ENTRY;
2965
2966 again:
2967         osc_object_lock(obj);
2968         ext = osc_extent_search(obj, index);
2969         if (ext == NULL)
2970                 ext = first_extent(obj);
2971         else if (ext->oe_end < index)
2972                 ext = next_extent(ext);
2973         while (ext != NULL) {
2974                 int rc;
2975
2976                 if (ext->oe_start > end)
2977                         break;
2978
2979                 if (!ext->oe_fsync_wait) {
2980                         ext = next_extent(ext);
2981                         continue;
2982                 }
2983
2984                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2985                              ext->oe_hp || ext->oe_urgent), ext);
2986                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2987                              !ext->oe_hp && ext->oe_urgent), ext);
2988
2989                 index = ext->oe_end + 1;
2990                 osc_extent_get(ext);
2991                 osc_object_unlock(obj);
2992
2993                 rc = osc_extent_wait(env, ext, OES_INV);
2994                 if (result == 0)
2995                         result = rc;
2996                 osc_extent_put(env, ext);
2997                 goto again;
2998         }
2999         osc_object_unlock(obj);
3000
3001         OSC_IO_DEBUG(obj, "sync file range.\n");
3002         RETURN(result);
3003 }
3004
3005 /**
3006  * Called to write out a range of osc object.
3007  *
3008  * @hp     : should be set this is caused by lock cancel;
3009  * @discard: is set if dirty pages should be dropped - file will be deleted or
3010  *         truncated, this implies there is no partially discarding extents.
3011  *
3012  * Return how many pages will be issued, or error code if error occurred.
3013  */
3014 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
3015                               pgoff_t start, pgoff_t end, int hp, int discard)
3016 {
3017         struct osc_extent *ext;
3018         struct list_head discard_list = LIST_HEAD_INIT(discard_list);
3019         bool unplug = false;
3020         int result = 0;
3021         ENTRY;
3022
3023         osc_object_lock(obj);
3024         ext = osc_extent_search(obj, start);
3025         if (ext == NULL)
3026                 ext = first_extent(obj);
3027         else if (ext->oe_end < start)
3028                 ext = next_extent(ext);
3029         while (ext != NULL) {
3030                 if (ext->oe_start > end)
3031                         break;
3032
3033                 ext->oe_fsync_wait = 1;
3034                 switch (ext->oe_state) {
3035                 case OES_CACHE:
3036                         result += ext->oe_nr_pages;
3037                         if (!discard) {
3038                                 struct list_head *list = NULL;
3039                                 if (hp) {
3040                                         EASSERT(!ext->oe_hp, ext);
3041                                         ext->oe_hp = 1;
3042                                         list = &obj->oo_hp_exts;
3043                                 } else if (!ext->oe_urgent) {
3044                                         ext->oe_urgent = 1;
3045                                         list = &obj->oo_urgent_exts;
3046                                 }
3047                                 if (list != NULL)
3048                                         list_move_tail(&ext->oe_link, list);
3049                                 unplug = true;
3050                         } else {
3051                                 /* the only discarder is lock cancelling, so
3052                                  * [start, end] must contain this extent */
3053                                 EASSERT(ext->oe_start >= start &&
3054                                         ext->oe_max_end <= end, ext);
3055                                 osc_extent_state_set(ext, OES_LOCKING);
3056                                 ext->oe_owner = current;
3057                                 list_move_tail(&ext->oe_link,
3058                                                    &discard_list);
3059                                 osc_update_pending(obj, OBD_BRW_WRITE,
3060                                                    -ext->oe_nr_pages);
3061                         }
3062                         break;
3063                 case OES_ACTIVE:
3064                         /* It's pretty bad to wait for ACTIVE extents, because
3065                          * we don't know how long we will wait for it to be
3066                          * flushed since it may be blocked at awaiting more
3067                          * grants. We do this for the correctness of fsync. */
3068                         LASSERT(hp == 0 && discard == 0);
3069                         ext->oe_urgent = 1;
3070                         break;
3071                 case OES_TRUNC:
3072                         /* this extent is being truncated, can't do anything
3073                          * for it now. it will be set to urgent after truncate
3074                          * is finished in osc_cache_truncate_end(). */
3075                 default:
3076                         break;
3077                 }
3078                 ext = next_extent(ext);
3079         }
3080         osc_object_unlock(obj);
3081
3082         LASSERT(ergo(!discard, list_empty(&discard_list)));
3083         if (!list_empty(&discard_list)) {
3084                 struct osc_extent *tmp;
3085                 int rc;
3086
3087                 osc_list_maint(osc_cli(obj), obj);
3088                 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
3089                         list_del_init(&ext->oe_link);
3090                         EASSERT(ext->oe_state == OES_LOCKING, ext);
3091
3092                         /* Discard caching pages. We don't actually write this
3093                          * extent out but we complete it as if we did. */
3094                         rc = osc_extent_make_ready(env, ext);
3095                         if (unlikely(rc < 0)) {
3096                                 OSC_EXTENT_DUMP(D_ERROR, ext,
3097                                                 "make_ready returned %d\n", rc);
3098                                 if (result >= 0)
3099                                         result = rc;
3100                         }
3101
3102                         /* finish the extent as if the pages were sent */
3103                         osc_extent_finish(env, ext, 0, 0);
3104                 }
3105         }
3106
3107         if (unplug)
3108                 osc_io_unplug(env, osc_cli(obj), obj);
3109
3110         if (hp || discard) {
3111                 int rc;
3112                 rc = osc_cache_wait_range(env, obj, start, end);
3113                 if (result >= 0 && rc < 0)
3114                         result = rc;
3115         }
3116
3117         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3118         RETURN(result);
3119 }
3120
3121 /**
3122  * Returns a list of pages by a given [start, end] of \a obj.
3123  *
3124  * \param resched If not NULL, then we give up before hogging CPU for too
3125  * long and set *resched = 1, in that case caller should implement a retry
3126  * logic.
3127  *
3128  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3129  * crucial in the face of [offset, EOF] locks.
3130  *
3131  * Return at least one page in @queue unless there is no covered page.
3132  */
3133 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3134                         struct osc_object *osc, pgoff_t start, pgoff_t end,
3135                         osc_page_gang_cbt cb, void *cbdata)
3136 {
3137         struct osc_page *ops;
3138         void            **pvec;
3139         pgoff_t         idx;
3140         unsigned int    nr;
3141         unsigned int    i;
3142         unsigned int    j;
3143         int             res = CLP_GANG_OKAY;
3144         bool            tree_lock = true;
3145         ENTRY;
3146
3147         idx = start;
3148         pvec = osc_env_info(env)->oti_pvec;
3149         spin_lock(&osc->oo_tree_lock);
3150         while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3151                                             idx, OTI_PVEC_SIZE)) > 0) {
3152                 struct cl_page *page;
3153                 bool end_of_region = false;
3154
3155                 for (i = 0, j = 0; i < nr; ++i) {
3156                         ops = pvec[i];
3157                         pvec[i] = NULL;
3158
3159                         idx = osc_index(ops);
3160                         if (idx > end) {
3161                                 end_of_region = true;
3162                                 break;
3163                         }
3164
3165                         page = ops->ops_cl.cpl_page;
3166                         LASSERT(page->cp_type == CPT_CACHEABLE);
3167                         if (page->cp_state == CPS_FREEING)
3168                                 continue;
3169
3170                         cl_page_get(page);
3171                         lu_ref_add_atomic(&page->cp_reference,
3172                                           "gang_lookup", current);
3173                         pvec[j++] = ops;
3174                 }
3175                 ++idx;
3176
3177                 /*
3178                  * Here a delicate locking dance is performed. Current thread
3179                  * holds a reference to a page, but has to own it before it
3180                  * can be placed into queue. Owning implies waiting, so
3181                  * radix-tree lock is to be released. After a wait one has to
3182                  * check that pages weren't truncated (cl_page_own() returns
3183                  * error in the latter case).
3184                  */
3185                 spin_unlock(&osc->oo_tree_lock);
3186                 tree_lock = false;
3187
3188                 for (i = 0; i < j; ++i) {
3189                         ops = pvec[i];
3190                         if (res == CLP_GANG_OKAY)
3191                                 res = (*cb)(env, io, ops, cbdata);
3192
3193                         page = ops->ops_cl.cpl_page;
3194                         lu_ref_del(&page->cp_reference, "gang_lookup", current);
3195                         cl_page_put(env, page);
3196                 }
3197                 if (nr < OTI_PVEC_SIZE || end_of_region)
3198                         break;
3199
3200                 if (res == CLP_GANG_OKAY && need_resched())
3201                         res = CLP_GANG_RESCHED;
3202                 if (res != CLP_GANG_OKAY)
3203                         break;
3204
3205                 spin_lock(&osc->oo_tree_lock);
3206                 tree_lock = true;
3207         }
3208         if (tree_lock)
3209                 spin_unlock(&osc->oo_tree_lock);
3210         RETURN(res);
3211 }
3212
3213 /**
3214  * Check if page @page is covered by an extra lock or discard it.
3215  */
3216 static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3217                                 struct osc_page *ops, void *cbdata)
3218 {
3219         struct osc_thread_info *info = osc_env_info(env);
3220         struct osc_object *osc = cbdata;
3221         pgoff_t index;
3222
3223         index = osc_index(ops);
3224         if (index >= info->oti_fn_index) {
3225                 struct ldlm_lock *tmp;
3226                 struct cl_page *page = ops->ops_cl.cpl_page;
3227
3228                 /* refresh non-overlapped index */
3229                 tmp = osc_dlmlock_at_pgoff(env, osc, index,
3230                                            OSC_DAP_FL_TEST_LOCK);
3231                 if (tmp != NULL) {
3232                         __u64 end = tmp->l_policy_data.l_extent.end;
3233                         /* Cache the first-non-overlapped index so as to skip
3234                          * all pages within [index, oti_fn_index). This is safe
3235                          * because if tmp lock is canceled, it will discard
3236                          * these pages. */
3237                         info->oti_fn_index = cl_index(osc2cl(osc), end + 1);
3238                         if (end == OBD_OBJECT_EOF)
3239                                 info->oti_fn_index = CL_PAGE_EOF;
3240                         LDLM_LOCK_PUT(tmp);
3241                 } else if (cl_page_own(env, io, page) == 0) {
3242                         /* discard the page */
3243                         cl_page_discard(env, io, page);
3244                         cl_page_disown(env, io, page);
3245                 } else {
3246                         LASSERT(page->cp_state == CPS_FREEING);
3247                 }
3248         }
3249
3250         info->oti_next_index = index + 1;
3251         return CLP_GANG_OKAY;
3252 }
3253
3254 static int discard_cb(const struct lu_env *env, struct cl_io *io,
3255                       struct osc_page *ops, void *cbdata)
3256 {
3257         struct osc_thread_info *info = osc_env_info(env);
3258         struct cl_page *page = ops->ops_cl.cpl_page;
3259
3260         /* page is top page. */
3261         info->oti_next_index = osc_index(ops) + 1;
3262         if (cl_page_own(env, io, page) == 0) {
3263                 if (!ergo(page->cp_type == CPT_CACHEABLE,
3264                           !PageDirty(cl_page_vmpage(page))))
3265                         CL_PAGE_DEBUG(D_ERROR, env, page,
3266                                         "discard dirty page?\n");
3267
3268                 /* discard the page */
3269                 cl_page_discard(env, io, page);
3270                 cl_page_disown(env, io, page);
3271         } else {
3272                 LASSERT(page->cp_state == CPS_FREEING);
3273         }
3274
3275         return CLP_GANG_OKAY;
3276 }
3277
3278 /**
3279  * Discard pages protected by the given lock. This function traverses radix
3280  * tree to find all covering pages and discard them. If a page is being covered
3281  * by other locks, it should remain in cache.
3282  *
3283  * If error happens on any step, the process continues anyway (the reasoning
3284  * behind this being that lock cancellation cannot be delayed indefinitely).
3285  */
3286 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3287                            pgoff_t start, pgoff_t end, enum cl_lock_mode mode)
3288 {
3289         struct osc_thread_info *info = osc_env_info(env);
3290         struct cl_io *io = &info->oti_io;
3291         osc_page_gang_cbt cb;
3292         int res;
3293         int result;
3294
3295         ENTRY;
3296
3297         io->ci_obj = cl_object_top(osc2cl(osc));
3298         io->ci_ignore_layout = 1;
3299         result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3300         if (result != 0)
3301                 GOTO(out, result);
3302
3303         cb = mode == CLM_READ ? check_and_discard_cb : discard_cb;
3304         info->oti_fn_index = info->oti_next_index = start;
3305         do {
3306                 res = osc_page_gang_lookup(env, io, osc,
3307                                            info->oti_next_index, end, cb, osc);
3308                 if (info->oti_next_index > end)
3309                         break;
3310
3311                 if (res == CLP_GANG_RESCHED)
3312                         cond_resched();
3313         } while (res != CLP_GANG_OKAY);
3314 out:
3315         cl_io_fini(env, io);
3316         RETURN(result);
3317 }
3318
3319
3320 /** @} osc */