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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, 2015, Intel Corporation.
  31  *
  32  */
  33 /*
  34  * This file is part of Lustre, http://www.lustre.org/
  35  * Lustre is a trademark of Sun Microsystems, Inc.
  36  *
  37  * osc cache management.
  38  *
  39  * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_OSC
  43
  44 #include "osc_cl_internal.h"
  45 #include "osc_internal.h"
  46
  47 static int extent_debug; /* set it to be true for more debug */
  48
  49 static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
  50 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
  51                            enum osc_extent_state state);
  52 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
  53                               struct osc_async_page *oap, int sent, int rc);
  54 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
  55                           int cmd);
  56 static int osc_refresh_count(const struct lu_env *env,
  57                              struct osc_async_page *oap, int cmd);
  58 static int osc_io_unplug_async(const struct lu_env *env,
  59                                struct client_obd *cli, struct osc_object *osc);
  60 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
  61                            unsigned int lost_grant);
  62
  63 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
  64                                   const char *func, int line);
  65 #define osc_extent_tree_dump(lvl, obj) \
  66         osc_extent_tree_dump0(lvl, obj, __func__, __LINE__)
  67
  68 /** \addtogroup osc
  69  *  @{
  70  */
  71
  72 /* ------------------ osc extent ------------------ */
  73 static inline char *ext_flags(struct osc_extent *ext, char *flags)
  74 {
  75         char *buf = flags;
  76         *buf++ = ext->oe_rw ? 'r' : 'w';
  77         if (ext->oe_intree)
  78                 *buf++ = 'i';
  79         if (ext->oe_sync)
  80                 *buf++ = 'S';
  81         if (ext->oe_srvlock)
  82                 *buf++ = 's';
  83         if (ext->oe_hp)
  84                 *buf++ = 'h';
  85         if (ext->oe_urgent)
  86                 *buf++ = 'u';
  87         if (ext->oe_memalloc)
  88                 *buf++ = 'm';
  89         if (ext->oe_trunc_pending)
  90                 *buf++ = 't';
  91         if (ext->oe_fsync_wait)
  92                 *buf++ = 'Y';
  93         *buf = 0;
  94         return flags;
  95 }
  96
  97 static inline char list_empty_marker(struct list_head *list)
  98 {
  99         return list_empty(list) ? '-' : '+';
 100 }
 101
 102 #define EXTSTR       "[%lu -> %lu/%lu]"
 103 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
 104 static const char *oes_strings[] = {
 105         "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
 106
 107 #define OSC_EXTENT_DUMP(lvl, extent, fmt, ...) do {                           \
 108         struct osc_extent *__ext = (extent);                                  \
 109         char __buf[16];                                                       \
 110                                                                               \
 111         CDEBUG(lvl,                                                           \
 112                 "extent %p@{" EXTSTR ", "                                     \
 113                 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt,          \
 114                 /* ----- extent part 0 ----- */                               \
 115                 __ext, EXTPARA(__ext),                                        \
 116                 /* ----- part 1 ----- */                                      \
 117                 atomic_read(&__ext->oe_refc),                         \
 118                 atomic_read(&__ext->oe_users),                        \
 119                 list_empty_marker(&__ext->oe_link),                           \
 120                 oes_strings[__ext->oe_state], ext_flags(__ext, __buf),        \
 121                 __ext->oe_obj,                                                \
 122                 /* ----- part 2 ----- */                                      \
 123                 __ext->oe_grants, __ext->oe_nr_pages,                         \
 124                 list_empty_marker(&__ext->oe_pages),                          \
 125                 waitqueue_active(&__ext->oe_waitq) ? '+' : '-',               \
 126                 __ext->oe_dlmlock, __ext->oe_mppr, __ext->oe_owner,           \
 127                 /* ----- part 4 ----- */                                      \
 128                 ## __VA_ARGS__);                                              \
 129         if (lvl == D_ERROR && __ext->oe_dlmlock != NULL)                      \
 130                 LDLM_ERROR(__ext->oe_dlmlock, "extent: %p\n", __ext);         \
 131         else                                                                  \
 132                 LDLM_DEBUG(__ext->oe_dlmlock, "extent: %p\n", __ext);         \
 133 } while (0)
 134
 135 #undef EASSERTF
 136 #define EASSERTF(expr, ext, fmt, args...) do {                          \
 137         if (!(expr)) {                                                  \
 138                 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args);           \
 139                 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj);           \
 140                 LASSERT(expr);                                          \
 141         }                                                               \
 142 } while (0)
 143
 144 #undef EASSERT
 145 #define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
 146
 147 static inline struct osc_extent *rb_extent(struct rb_node *n)
 148 {
 149         if (n == NULL)
 150                 return NULL;
 151
 152         return container_of(n, struct osc_extent, oe_node);
 153 }
 154
 155 static inline struct osc_extent *next_extent(struct osc_extent *ext)
 156 {
 157         if (ext == NULL)
 158                 return NULL;
 159
 160         LASSERT(ext->oe_intree);
 161         return rb_extent(rb_next(&ext->oe_node));
 162 }
 163
 164 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
 165 {
 166         if (ext == NULL)
 167                 return NULL;
 168
 169         LASSERT(ext->oe_intree);
 170         return rb_extent(rb_prev(&ext->oe_node));
 171 }
 172
 173 static inline struct osc_extent *first_extent(struct osc_object *obj)
 174 {
 175         return rb_extent(rb_first(&obj->oo_root));
 176 }
 177
 178 /* object must be locked by caller. */
 179 static int osc_extent_sanity_check0(struct osc_extent *ext,
 180                                     const char *func, const int line)
 181 {
 182         struct osc_object *obj = ext->oe_obj;
 183         struct osc_async_page *oap;
 184         size_t page_count;
 185         int rc = 0;
 186
 187         if (!osc_object_is_locked(obj))
 188                 GOTO(out, rc = 9);
 189
 190         if (ext->oe_state >= OES_STATE_MAX)
 191                 GOTO(out, rc = 10);
 192
 193         if (atomic_read(&ext->oe_refc) <= 0)
 194                 GOTO(out, rc = 20);
 195
 196         if (atomic_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
 197                 GOTO(out, rc = 30);
 198
 199         switch (ext->oe_state) {
 200         case OES_INV:
 201                 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
 202                         GOTO(out, rc = 35);
 203                 GOTO(out, rc = 0);
 204                 break;
 205         case OES_ACTIVE:
 206                 if (atomic_read(&ext->oe_users) == 0)
 207                         GOTO(out, rc = 40);
 208                 if (ext->oe_hp)
 209                         GOTO(out, rc = 50);
 210                 if (ext->oe_fsync_wait && !ext->oe_urgent)
 211                         GOTO(out, rc = 55);
 212                 break;
 213         case OES_CACHE:
 214                 if (ext->oe_grants == 0)
 215                         GOTO(out, rc = 60);
 216                 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
 217                         GOTO(out, rc = 65);
 218         default:
 219                 if (atomic_read(&ext->oe_users) > 0)
 220                         GOTO(out, rc = 70);
 221         }
 222
 223         if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
 224                 GOTO(out, rc = 80);
 225
 226         if (ext->oe_sync && ext->oe_grants > 0)
 227                 GOTO(out, rc = 90);
 228
 229         if (ext->oe_dlmlock != NULL && !ldlm_is_failed(ext->oe_dlmlock)) {
 230                 struct ldlm_extent *extent;
 231
 232                 extent = &ext->oe_dlmlock->l_policy_data.l_extent;
 233                 if (!(extent->start <= cl_offset(osc2cl(obj), ext->oe_start) &&
 234                       extent->end   >= cl_offset(osc2cl(obj), ext->oe_max_end)))
 235                         GOTO(out, rc = 100);
 236
 237                 if (!(ext->oe_dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP)))
 238                         GOTO(out, rc = 102);
 239         }
 240
 241         if (ext->oe_nr_pages > ext->oe_mppr)
 242                 GOTO(out, rc = 105);
 243
 244         /* Do not verify page list if extent is in RPC. This is because an
 245          * in-RPC extent is supposed to be exclusively accessible w/o lock. */
 246         if (ext->oe_state > OES_CACHE)
 247                 GOTO(out, rc = 0);
 248
 249         if (!extent_debug)
 250                 GOTO(out, rc = 0);
 251
 252         page_count = 0;
 253         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
 254                 pgoff_t index = osc_index(oap2osc(oap));
 255                 ++page_count;
 256                 if (index > ext->oe_end || index < ext->oe_start)
 257                         GOTO(out, rc = 110);
 258         }
 259         if (page_count != ext->oe_nr_pages)
 260                 GOTO(out, rc = 120);
 261
 262 out:
 263         if (rc != 0)
 264                 OSC_EXTENT_DUMP(D_ERROR, ext,
 265                                 "%s:%d sanity check %p failed with rc = %d\n",
 266                                 func, line, ext, rc);
 267         return rc;
 268 }
 269
 270 #define sanity_check_nolock(ext) \
 271         osc_extent_sanity_check0(ext, __func__, __LINE__)
 272
 273 #define sanity_check(ext) ({                                                   \
 274         int __res;                                                             \
 275         osc_object_lock((ext)->oe_obj);                                        \
 276         __res = sanity_check_nolock(ext);                                      \
 277         osc_object_unlock((ext)->oe_obj);                                      \
 278         __res;                                                                 \
 279 })
 280
 281
 282 /**
 283  * sanity check - to make sure there is no overlapped extent in the tree.
 284  */
 285 static int osc_extent_is_overlapped(struct osc_object *obj,
 286                                     struct osc_extent *ext)
 287 {
 288         struct osc_extent *tmp;
 289
 290         LASSERT(osc_object_is_locked(obj));
 291
 292         if (!extent_debug)
 293                 return 0;
 294
 295         for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
 296                 if (tmp == ext)
 297                         continue;
 298                 if (tmp->oe_end >= ext->oe_start &&
 299                     tmp->oe_start <= ext->oe_end)
 300                         return 1;
 301         }
 302         return 0;
 303 }
 304
 305 static void osc_extent_state_set(struct osc_extent *ext, int state)
 306 {
 307         LASSERT(osc_object_is_locked(ext->oe_obj));
 308         LASSERT(state >= OES_INV && state < OES_STATE_MAX);
 309
 310         /* Never try to sanity check a state changing extent :-) */
 311         /* LASSERT(sanity_check_nolock(ext) == 0); */
 312
 313         /* TODO: validate the state machine */
 314         ext->oe_state = state;
 315         wake_up_all(&ext->oe_waitq);
 316 }
 317
 318 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
 319 {
 320         struct osc_extent *ext;
 321
 322         OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_IOFS);
 323         if (ext == NULL)
 324                 return NULL;
 325
 326         RB_CLEAR_NODE(&ext->oe_node);
 327         ext->oe_obj = obj;
 328         cl_object_get(osc2cl(obj));
 329         atomic_set(&ext->oe_refc, 1);
 330         atomic_set(&ext->oe_users, 0);
 331         INIT_LIST_HEAD(&ext->oe_link);
 332         ext->oe_state = OES_INV;
 333         INIT_LIST_HEAD(&ext->oe_pages);
 334         init_waitqueue_head(&ext->oe_waitq);
 335         ext->oe_dlmlock = NULL;
 336
 337         return ext;
 338 }
 339
 340 static void osc_extent_free(struct osc_extent *ext)
 341 {
 342         OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
 343 }
 344
 345 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
 346 {
 347         LASSERT(atomic_read(&ext->oe_refc) >= 0);
 348         atomic_inc(&ext->oe_refc);
 349         return ext;
 350 }
 351
 352 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
 353 {
 354         LASSERT(atomic_read(&ext->oe_refc) > 0);
 355         if (atomic_dec_and_test(&ext->oe_refc)) {
 356                 LASSERT(list_empty(&ext->oe_link));
 357                 LASSERT(atomic_read(&ext->oe_users) == 0);
 358                 LASSERT(ext->oe_state == OES_INV);
 359                 LASSERT(!ext->oe_intree);
 360
 361                 if (ext->oe_dlmlock != NULL) {
 362                         lu_ref_add(&ext->oe_dlmlock->l_reference,
 363                                    "osc_extent", ext);
 364                         LDLM_LOCK_PUT(ext->oe_dlmlock);
 365                         ext->oe_dlmlock = NULL;
 366                 }
 367                 cl_object_put(env, osc2cl(ext->oe_obj));
 368                 osc_extent_free(ext);
 369         }
 370 }
 371
 372 /**
 373  * osc_extent_put_trust() is a special version of osc_extent_put() when
 374  * it's known that the caller is not the last user. This is to address the
 375  * problem of lacking of lu_env ;-).
 376  */
 377 static void osc_extent_put_trust(struct osc_extent *ext)
 378 {
 379         LASSERT(atomic_read(&ext->oe_refc) > 1);
 380         LASSERT(osc_object_is_locked(ext->oe_obj));
 381         atomic_dec(&ext->oe_refc);
 382 }
 383
 384 /**
 385  * Return the extent which includes pgoff @index, or return the greatest
 386  * previous extent in the tree.
 387  */
 388 static struct osc_extent *osc_extent_search(struct osc_object *obj,
 389                                             pgoff_t index)
 390 {
 391         struct rb_node    *n = obj->oo_root.rb_node;
 392         struct osc_extent *tmp, *p = NULL;
 393
 394         LASSERT(osc_object_is_locked(obj));
 395         while (n != NULL) {
 396                 tmp = rb_extent(n);
 397                 if (index < tmp->oe_start) {
 398                         n = n->rb_left;
 399                 } else if (index > tmp->oe_end) {
 400                         p = rb_extent(n);
 401                         n = n->rb_right;
 402                 } else {
 403                         return tmp;
 404                 }
 405         }
 406         return p;
 407 }
 408
 409 /*
 410  * Return the extent covering @index, otherwise return NULL.
 411  * caller must have held object lock.
 412  */
 413 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
 414                                             pgoff_t index)
 415 {
 416         struct osc_extent *ext;
 417
 418         ext = osc_extent_search(obj, index);
 419         if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
 420                 return osc_extent_get(ext);
 421         return NULL;
 422 }
 423
 424 /* caller must have held object lock. */
 425 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
 426 {
 427         struct rb_node   **n      = &obj->oo_root.rb_node;
 428         struct rb_node    *parent = NULL;
 429         struct osc_extent *tmp;
 430
 431         LASSERT(ext->oe_intree == 0);
 432         LASSERT(ext->oe_obj == obj);
 433         LASSERT(osc_object_is_locked(obj));
 434         while (*n != NULL) {
 435                 tmp = rb_extent(*n);
 436                 parent = *n;
 437
 438                 if (ext->oe_end < tmp->oe_start)
 439                         n = &(*n)->rb_left;
 440                 else if (ext->oe_start > tmp->oe_end)
 441                         n = &(*n)->rb_right;
 442                 else
 443                         EASSERTF(0, tmp, EXTSTR"\n", EXTPARA(ext));
 444         }
 445         rb_link_node(&ext->oe_node, parent, n);
 446         rb_insert_color(&ext->oe_node, &obj->oo_root);
 447         osc_extent_get(ext);
 448         ext->oe_intree = 1;
 449 }
 450
 451 /* caller must have held object lock. */
 452 static void osc_extent_erase(struct osc_extent *ext)
 453 {
 454         struct osc_object *obj = ext->oe_obj;
 455         LASSERT(osc_object_is_locked(obj));
 456         if (ext->oe_intree) {
 457                 rb_erase(&ext->oe_node, &obj->oo_root);
 458                 ext->oe_intree = 0;
 459                 /* rbtree held a refcount */
 460                 osc_extent_put_trust(ext);
 461         }
 462 }
 463
 464 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
 465 {
 466         struct osc_object *obj = ext->oe_obj;
 467
 468         LASSERT(osc_object_is_locked(obj));
 469         LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
 470         if (ext->oe_state == OES_CACHE) {
 471                 osc_extent_state_set(ext, OES_ACTIVE);
 472                 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
 473         }
 474         atomic_inc(&ext->oe_users);
 475         list_del_init(&ext->oe_link);
 476         return osc_extent_get(ext);
 477 }
 478
 479 static void __osc_extent_remove(struct osc_extent *ext)
 480 {
 481         LASSERT(osc_object_is_locked(ext->oe_obj));
 482         LASSERT(list_empty(&ext->oe_pages));
 483         osc_extent_erase(ext);
 484         list_del_init(&ext->oe_link);
 485         osc_extent_state_set(ext, OES_INV);
 486         OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
 487 }
 488
 489 static void osc_extent_remove(struct osc_extent *ext)
 490 {
 491         struct osc_object *obj = ext->oe_obj;
 492
 493         osc_object_lock(obj);
 494         __osc_extent_remove(ext);
 495         osc_object_unlock(obj);
 496 }
 497
 498 /**
 499  * This function is used to merge extents to get better performance. It checks
 500  * if @cur and @victim are contiguous at chunk level.
 501  */
 502 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
 503                             struct osc_extent *victim)
 504 {
 505         struct osc_object *obj = cur->oe_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_CACHE_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         OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
 530
 531         cur->oe_start     = min(cur->oe_start, victim->oe_start);
 532         cur->oe_end       = max(cur->oe_end,   victim->oe_end);
 533         cur->oe_grants   += victim->oe_grants;
 534         cur->oe_nr_pages += victim->oe_nr_pages;
 535         /* only the following bits are needed to merge */
 536         cur->oe_urgent   |= victim->oe_urgent;
 537         cur->oe_memalloc |= victim->oe_memalloc;
 538         list_splice_init(&victim->oe_pages, &cur->oe_pages);
 539         list_del_init(&victim->oe_link);
 540         victim->oe_nr_pages = 0;
 541
 542         osc_extent_get(victim);
 543         __osc_extent_remove(victim);
 544         osc_extent_put(env, victim);
 545
 546         OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
 547         return 0;
 548 }
 549
 550 /**
 551  * Drop user count of osc_extent, and unplug IO asynchronously.
 552  */
 553 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
 554 {
 555         struct osc_object *obj = ext->oe_obj;
 556         int rc = 0;
 557         ENTRY;
 558
 559         LASSERT(atomic_read(&ext->oe_users) > 0);
 560         LASSERT(sanity_check(ext) == 0);
 561         LASSERT(ext->oe_grants > 0);
 562
 563         if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
 564                 LASSERT(ext->oe_state == OES_ACTIVE);
 565                 if (ext->oe_trunc_pending) {
 566                         /* a truncate process is waiting for this extent.
 567                          * This may happen due to a race, check
 568                          * osc_cache_truncate_start(). */
 569                         osc_extent_state_set(ext, OES_TRUNC);
 570                         ext->oe_trunc_pending = 0;
 571                 } else {
 572                         osc_extent_state_set(ext, OES_CACHE);
 573                         osc_update_pending(obj, OBD_BRW_WRITE,
 574                                            ext->oe_nr_pages);
 575
 576                         /* try to merge the previous and next extent. */
 577                         osc_extent_merge(env, ext, prev_extent(ext));
 578                         osc_extent_merge(env, ext, next_extent(ext));
 579
 580                         if (ext->oe_urgent)
 581                                 list_move_tail(&ext->oe_link,
 582                                                &obj->oo_urgent_exts);
 583                 }
 584                 osc_object_unlock(obj);
 585
 586                 osc_io_unplug_async(env, osc_cli(obj), obj);
 587         }
 588         osc_extent_put(env, ext);
 589         RETURN(rc);
 590 }
 591
 592 static inline int overlapped(struct osc_extent *ex1, struct osc_extent *ex2)
 593 {
 594         return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
 595 }
 596
 597 /**
 598  * Find or create an extent which includes @index, core function to manage
 599  * extent tree.
 600  */
 601 static struct osc_extent *osc_extent_find(const struct lu_env *env,
 602                                           struct osc_object *obj, pgoff_t index,
 603                                           unsigned int *grants)
 604 {
 605         struct client_obd *cli = osc_cli(obj);
 606         struct osc_lock   *olck;
 607         struct cl_lock_descr *descr;
 608         struct osc_extent *cur;
 609         struct osc_extent *ext;
 610         struct osc_extent *conflict = NULL;
 611         struct osc_extent *found = NULL;
 612         pgoff_t    chunk;
 613         pgoff_t    max_end;
 614         unsigned int max_pages; /* max_pages_per_rpc */
 615         unsigned int chunksize;
 616         int        ppc_bits; /* pages per chunk bits */
 617         pgoff_t    chunk_mask;
 618         int        rc;
 619         ENTRY;
 620
 621         cur = osc_extent_alloc(obj);
 622         if (cur == NULL)
 623                 RETURN(ERR_PTR(-ENOMEM));
 624
 625         olck = osc_env_io(env)->oi_write_osclock;
 626         LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
 627         LASSERT(olck->ols_state == OLS_GRANTED);
 628
 629         descr = &olck->ols_cl.cls_lock->cll_descr;
 630         LASSERT(descr->cld_mode >= CLM_WRITE);
 631
 632         LASSERT(cli->cl_chunkbits >= PAGE_CACHE_SHIFT);
 633         ppc_bits   = cli->cl_chunkbits - PAGE_CACHE_SHIFT;
 634         chunk_mask = ~((1 << ppc_bits) - 1);
 635         chunksize  = 1 << cli->cl_chunkbits;
 636         chunk      = index >> ppc_bits;
 637
 638         /* align end to rpc edge, rpc size may not be a power 2 integer. */
 639         max_pages = cli->cl_max_pages_per_rpc;
 640         LASSERT((max_pages & ~chunk_mask) == 0);
 641         max_end = index - (index % max_pages) + max_pages - 1;
 642         max_end = min_t(pgoff_t, max_end, descr->cld_end);
 643
 644         /* initialize new extent by parameters so far */
 645         cur->oe_max_end = max_end;
 646         cur->oe_start   = index & chunk_mask;
 647         cur->oe_end     = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
 648         if (cur->oe_start < descr->cld_start)
 649                 cur->oe_start = descr->cld_start;
 650         if (cur->oe_end > max_end)
 651                 cur->oe_end = max_end;
 652         cur->oe_grants  = 0;
 653         cur->oe_mppr    = max_pages;
 654         if (olck->ols_dlmlock != NULL) {
 655                 LASSERT(olck->ols_hold);
 656                 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
 657                 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
 658         }
 659
 660         /* grants has been allocated by caller */
 661         LASSERTF(*grants >= chunksize + cli->cl_extent_tax,
 662                  "%u/%u/%u.\n", *grants, chunksize, cli->cl_extent_tax);
 663         LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
 664                  EXTPARA(cur));
 665
 666 restart:
 667         osc_object_lock(obj);
 668         ext = osc_extent_search(obj, cur->oe_start);
 669         if (ext == NULL)
 670                 ext = first_extent(obj);
 671         while (ext != NULL) {
 672                 pgoff_t ext_chk_start = ext->oe_start >> ppc_bits;
 673                 pgoff_t ext_chk_end   = ext->oe_end   >> ppc_bits;
 674
 675                 LASSERT(sanity_check_nolock(ext) == 0);
 676                 if (chunk > ext_chk_end + 1)
 677                         break;
 678
 679                 /* if covering by different locks, no chance to match */
 680                 if (olck->ols_dlmlock != ext->oe_dlmlock) {
 681                         EASSERTF(!overlapped(ext, cur), ext,
 682                                  EXTSTR"\n", EXTPARA(cur));
 683
 684                         ext = next_extent(ext);
 685                         continue;
 686                 }
 687
 688                 /* discontiguous chunks? */
 689                 if (chunk + 1 < ext_chk_start) {
 690                         ext = next_extent(ext);
 691                         continue;
 692                 }
 693
 694                 /* ok, from now on, ext and cur have these attrs:
 695                  * 1. covered by the same lock
 696                  * 2. contiguous at chunk level or overlapping. */
 697
 698                 if (overlapped(ext, cur)) {
 699                         /* cur is the minimum unit, so overlapping means
 700                          * full contain. */
 701                         EASSERTF((ext->oe_start <= cur->oe_start &&
 702                                   ext->oe_end >= cur->oe_end),
 703                                  ext, EXTSTR"\n", EXTPARA(cur));
 704
 705                         if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
 706                                 /* for simplicity, we wait for this extent to
 707                                  * finish before going forward. */
 708                                 conflict = osc_extent_get(ext);
 709                                 break;
 710                         }
 711
 712                         found = osc_extent_hold(ext);
 713                         break;
 714                 }
 715
 716                 /* non-overlapped extent */
 717                 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
 718                         /* we can't do anything for a non OES_CACHE extent, or
 719                          * if there is someone waiting for this extent to be
 720                          * flushed, try next one. */
 721                         ext = next_extent(ext);
 722                         continue;
 723                 }
 724
 725                 /* check if they belong to the same rpc slot before trying to
 726                  * merge. the extents are not overlapped and contiguous at
 727                  * chunk level to get here. */
 728                 if (ext->oe_max_end != max_end) {
 729                         /* if they don't belong to the same RPC slot or
 730                          * max_pages_per_rpc has ever changed, do not merge. */
 731                         ext = next_extent(ext);
 732                         continue;
 733                 }
 734
 735                 /* it's required that an extent must be contiguous at chunk
 736                  * level so that we know the whole extent is covered by grant
 737                  * (the pages in the extent are NOT required to be contiguous).
 738                  * Otherwise, it will be too much difficult to know which
 739                  * chunks have grants allocated. */
 740
 741                 /* try to do front merge - extend ext's start */
 742                 if (chunk + 1 == ext_chk_start) {
 743                         /* ext must be chunk size aligned */
 744                         EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
 745
 746                         /* pull ext's start back to cover cur */
 747                         ext->oe_start   = cur->oe_start;
 748                         ext->oe_grants += chunksize;
 749                         *grants -= chunksize;
 750
 751                         found = osc_extent_hold(ext);
 752                 } else if (chunk == ext_chk_end + 1) {
 753                         /* rear merge */
 754                         ext->oe_end     = cur->oe_end;
 755                         ext->oe_grants += chunksize;
 756                         *grants -= chunksize;
 757
 758                         /* try to merge with the next one because we just fill
 759                          * in a gap */
 760                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 761                                 /* we can save extent tax from next extent */
 762                                 *grants += cli->cl_extent_tax;
 763
 764                         found = osc_extent_hold(ext);
 765                 }
 766                 if (found != NULL)
 767                         break;
 768
 769                 ext = next_extent(ext);
 770         }
 771
 772         osc_extent_tree_dump(D_CACHE, obj);
 773         if (found != NULL) {
 774                 LASSERT(conflict == NULL);
 775                 if (!IS_ERR(found)) {
 776                         LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
 777                         OSC_EXTENT_DUMP(D_CACHE, found,
 778                                         "found caching ext for %lu.\n", index);
 779                 }
 780         } else if (conflict == NULL) {
 781                 /* create a new extent */
 782                 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
 783                 cur->oe_grants = chunksize + cli->cl_extent_tax;
 784                 *grants -= cur->oe_grants;
 785                 LASSERT(*grants >= 0);
 786
 787                 cur->oe_state = OES_CACHE;
 788                 found = osc_extent_hold(cur);
 789                 osc_extent_insert(obj, cur);
 790                 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
 791                                 index, descr->cld_end);
 792         }
 793         osc_object_unlock(obj);
 794
 795         if (conflict != NULL) {
 796                 LASSERT(found == NULL);
 797
 798                 /* waiting for IO to finish. Please notice that it's impossible
 799                  * to be an OES_TRUNC extent. */
 800                 rc = osc_extent_wait(env, conflict, OES_INV);
 801                 osc_extent_put(env, conflict);
 802                 conflict = NULL;
 803                 if (rc < 0)
 804                         GOTO(out, found = ERR_PTR(rc));
 805
 806                 goto restart;
 807         }
 808         EXIT;
 809
 810 out:
 811         osc_extent_put(env, cur);
 812         LASSERT(*grants >= 0);
 813         return found;
 814 }
 815
 816 /**
 817  * Called when IO is finished to an extent.
 818  */
 819 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
 820                       int sent, int rc)
 821 {
 822         struct client_obd *cli = osc_cli(ext->oe_obj);
 823         struct osc_async_page *oap;
 824         struct osc_async_page *tmp;
 825         int nr_pages = ext->oe_nr_pages;
 826         int lost_grant = 0;
 827         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
 828         loff_t last_off = 0;
 829         int last_count = -1;
 830         ENTRY;
 831
 832         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
 833
 834         ext->oe_rc = rc ?: ext->oe_nr_pages;
 835         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
 836
 837         osc_lru_add_batch(cli, &ext->oe_pages);
 838         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 839                                      oap_pending_item) {
 840                 list_del_init(&oap->oap_rpc_item);
 841                 list_del_init(&oap->oap_pending_item);
 842                 if (last_off <= oap->oap_obj_off) {
 843                         last_off = oap->oap_obj_off;
 844                         last_count = oap->oap_count;
 845                 }
 846
 847                 --ext->oe_nr_pages;
 848                 osc_ap_completion(env, cli, oap, sent, rc);
 849         }
 850         EASSERT(ext->oe_nr_pages == 0, ext);
 851
 852         if (!sent) {
 853                 lost_grant = ext->oe_grants;
 854         } else if (blocksize < PAGE_CACHE_SIZE &&
 855                    last_count != PAGE_CACHE_SIZE) {
 856                 /* For short writes we shouldn't count parts of pages that
 857                  * span a whole chunk on the OST side, or our accounting goes
 858                  * wrong.  Should match the code in filter_grant_check. */
 859                 int offset = last_off & ~PAGE_MASK;
 860                 int count = last_count + (offset & (blocksize - 1));
 861                 int end = (offset + last_count) & (blocksize - 1);
 862                 if (end)
 863                         count += blocksize - end;
 864
 865                 lost_grant = PAGE_CACHE_SIZE - count;
 866         }
 867         if (ext->oe_grants > 0)
 868                 osc_free_grant(cli, nr_pages, lost_grant);
 869
 870         osc_extent_remove(ext);
 871         /* put the refcount for RPC */
 872         osc_extent_put(env, ext);
 873         RETURN(0);
 874 }
 875
 876 static int extent_wait_cb(struct osc_extent *ext, enum osc_extent_state state)
 877 {
 878         int ret;
 879
 880         osc_object_lock(ext->oe_obj);
 881         ret = ext->oe_state == state;
 882         osc_object_unlock(ext->oe_obj);
 883
 884         return ret;
 885 }
 886
 887 /**
 888  * Wait for the extent's state to become @state.
 889  */
 890 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
 891                            enum osc_extent_state state)
 892 {
 893         struct osc_object *obj = ext->oe_obj;
 894         struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
 895                                                   LWI_ON_SIGNAL_NOOP, NULL);
 896         int rc = 0;
 897         ENTRY;
 898
 899         osc_object_lock(obj);
 900         LASSERT(sanity_check_nolock(ext) == 0);
 901         /* `Kick' this extent only if the caller is waiting for it to be
 902          * written out. */
 903         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
 904                 if (ext->oe_state == OES_ACTIVE) {
 905                         ext->oe_urgent = 1;
 906                 } else if (ext->oe_state == OES_CACHE) {
 907                         ext->oe_urgent = 1;
 908                         osc_extent_hold(ext);
 909                         rc = 1;
 910                 }
 911         }
 912         osc_object_unlock(obj);
 913         if (rc == 1)
 914                 osc_extent_release(env, ext);
 915
 916         /* wait for the extent until its state becomes @state */
 917         rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
 918         if (rc == -ETIMEDOUT) {
 919                 OSC_EXTENT_DUMP(D_ERROR, ext,
 920                         "%s: wait ext to %u timedout, recovery in progress?\n",
 921                         cli_name(osc_cli(obj)), state);
 922
 923                 lwi = LWI_INTR(NULL, NULL);
 924                 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state),
 925                                   &lwi);
 926         }
 927         if (rc == 0 && ext->oe_rc < 0)
 928                 rc = ext->oe_rc;
 929         RETURN(rc);
 930 }
 931
 932 /**
 933  * Discard pages with index greater than @size. If @ext is overlapped with
 934  * @size, then partial truncate happens.
 935  */
 936 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
 937                                 bool partial)
 938 {
 939         struct cl_env_nest     nest;
 940         struct lu_env         *env;
 941         struct cl_io          *io;
 942         struct osc_object     *obj = ext->oe_obj;
 943         struct client_obd     *cli = osc_cli(obj);
 944         struct osc_async_page *oap;
 945         struct osc_async_page *tmp;
 946         int                    pages_in_chunk = 0;
 947         int                    ppc_bits    = cli->cl_chunkbits -
 948                                              PAGE_CACHE_SHIFT;
 949         __u64                  trunc_chunk = trunc_index >> ppc_bits;
 950         int                    grants   = 0;
 951         int                    nr_pages = 0;
 952         int                    rc       = 0;
 953         ENTRY;
 954
 955         LASSERT(sanity_check(ext) == 0);
 956         LASSERT(ext->oe_state == OES_TRUNC);
 957         LASSERT(!ext->oe_urgent);
 958
 959         /* Request new lu_env.
 960          * We can't use that env from osc_cache_truncate_start() because
 961          * it's from lov_io_sub and not fully initialized. */
 962         env = cl_env_nested_get(&nest);
 963         io  = &osc_env_info(env)->oti_io;
 964         io->ci_obj = cl_object_top(osc2cl(obj));
 965         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
 966         if (rc < 0)
 967                 GOTO(out, rc);
 968
 969         /* discard all pages with index greater than trunc_index */
 970         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 971                                      oap_pending_item) {
 972                 pgoff_t index = osc_index(oap2osc(oap));
 973                 struct cl_page  *page = oap2cl_page(oap);
 974
 975                 LASSERT(list_empty(&oap->oap_rpc_item));
 976
 977                 /* only discard the pages with their index greater than
 978                  * trunc_index, and ... */
 979                 if (index < trunc_index ||
 980                     (index == trunc_index && partial)) {
 981                         /* accounting how many pages remaining in the chunk
 982                          * so that we can calculate grants correctly. */
 983                         if (index >> ppc_bits == trunc_chunk)
 984                                 ++pages_in_chunk;
 985                         continue;
 986                 }
 987
 988                 list_del_init(&oap->oap_pending_item);
 989
 990                 cl_page_get(page);
 991                 lu_ref_add(&page->cp_reference, "truncate", current);
 992
 993                 if (cl_page_own(env, io, page) == 0) {
 994                         cl_page_discard(env, io, page);
 995                         cl_page_disown(env, io, page);
 996                 } else {
 997                         LASSERT(page->cp_state == CPS_FREEING);
 998                         LASSERT(0);
 999                 }
1000
1001                 lu_ref_del(&page->cp_reference, "truncate", current);
1002                 cl_page_put(env, page);
1003
1004                 --ext->oe_nr_pages;
1005                 ++nr_pages;
1006         }
1007         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1008                       ext->oe_nr_pages == 0),
1009                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1010
1011         osc_object_lock(obj);
1012         if (ext->oe_nr_pages == 0) {
1013                 LASSERT(pages_in_chunk == 0);
1014                 grants = ext->oe_grants;
1015                 ext->oe_grants = 0;
1016         } else { /* calculate how many grants we can free */
1017                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1018                 pgoff_t last_index;
1019
1020
1021                 /* if there is no pages in this chunk, we can also free grants
1022                  * for the last chunk */
1023                 if (pages_in_chunk == 0) {
1024                         /* if this is the 1st chunk and no pages in this chunk,
1025                          * ext->oe_nr_pages must be zero, so we should be in
1026                          * the other if-clause. */
1027                         LASSERT(trunc_chunk > 0);
1028                         --trunc_chunk;
1029                         ++chunks;
1030                 }
1031
1032                 /* this is what we can free from this extent */
1033                 grants          = chunks << cli->cl_chunkbits;
1034                 ext->oe_grants -= grants;
1035                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1036                 ext->oe_end     = min(last_index, ext->oe_max_end);
1037                 LASSERT(ext->oe_end >= ext->oe_start);
1038                 LASSERT(ext->oe_grants > 0);
1039         }
1040         osc_object_unlock(obj);
1041
1042         if (grants > 0 || nr_pages > 0)
1043                 osc_free_grant(cli, nr_pages, grants);
1044
1045 out:
1046         cl_io_fini(env, io);
1047         cl_env_nested_put(&nest, env);
1048         RETURN(rc);
1049 }
1050
1051 /**
1052  * This function is used to make the extent prepared for transfer.
1053  * A race with flusing page - ll_writepage() has to be handled cautiously.
1054  */
1055 static int osc_extent_make_ready(const struct lu_env *env,
1056                                  struct osc_extent *ext)
1057 {
1058         struct osc_async_page *oap;
1059         struct osc_async_page *last = NULL;
1060         struct osc_object *obj = ext->oe_obj;
1061         unsigned int page_count = 0;
1062         int rc;
1063         ENTRY;
1064
1065         /* we're going to grab page lock, so object lock must not be taken. */
1066         LASSERT(sanity_check(ext) == 0);
1067         /* in locking state, any process should not touch this extent. */
1068         EASSERT(ext->oe_state == OES_LOCKING, ext);
1069         EASSERT(ext->oe_owner != NULL, ext);
1070
1071         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1072
1073         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1074                 ++page_count;
1075                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1076                         last = oap;
1077
1078                 /* checking ASYNC_READY is race safe */
1079                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1080                         continue;
1081
1082                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1083                 switch (rc) {
1084                 case 0:
1085                         spin_lock(&oap->oap_lock);
1086                         oap->oap_async_flags |= ASYNC_READY;
1087                         spin_unlock(&oap->oap_lock);
1088                         break;
1089                 case -EALREADY:
1090                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1091                         break;
1092                 default:
1093                         LASSERTF(0, "unknown return code: %d\n", rc);
1094                 }
1095         }
1096
1097         LASSERT(page_count == ext->oe_nr_pages);
1098         LASSERT(last != NULL);
1099         /* the last page is the only one we need to refresh its count by
1100          * the size of file. */
1101         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1102                 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1103                 LASSERT(last_oap_count > 0);
1104                 LASSERT(last->oap_page_off + last_oap_count <= PAGE_CACHE_SIZE);
1105                 last->oap_count = last_oap_count;
1106                 spin_lock(&last->oap_lock);
1107                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1108                 spin_unlock(&last->oap_lock);
1109         }
1110
1111         /* for the rest of pages, we don't need to call osf_refresh_count()
1112          * because it's known they are not the last page */
1113         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1114                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1115                         oap->oap_count = PAGE_CACHE_SIZE - oap->oap_page_off;
1116                         spin_lock(&oap->oap_lock);
1117                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1118                         spin_unlock(&oap->oap_lock);
1119                 }
1120         }
1121
1122         osc_object_lock(obj);
1123         osc_extent_state_set(ext, OES_RPC);
1124         osc_object_unlock(obj);
1125         /* get a refcount for RPC. */
1126         osc_extent_get(ext);
1127
1128         RETURN(0);
1129 }
1130
1131 /**
1132  * Quick and simple version of osc_extent_find(). This function is frequently
1133  * called to expand the extent for the same IO. To expand the extent, the
1134  * page index must be in the same or next chunk of ext->oe_end.
1135  */
1136 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index,
1137                              unsigned int *grants)
1138 {
1139         struct osc_object *obj = ext->oe_obj;
1140         struct client_obd *cli = osc_cli(obj);
1141         struct osc_extent *next;
1142         int ppc_bits = cli->cl_chunkbits - PAGE_CACHE_SHIFT;
1143         pgoff_t chunk = index >> ppc_bits;
1144         pgoff_t end_chunk;
1145         pgoff_t end_index;
1146         unsigned int chunksize = 1 << cli->cl_chunkbits;
1147         int rc = 0;
1148         ENTRY;
1149
1150         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1151         osc_object_lock(obj);
1152         LASSERT(sanity_check_nolock(ext) == 0);
1153         end_chunk = ext->oe_end >> ppc_bits;
1154         if (chunk > end_chunk + 1)
1155                 GOTO(out, rc = -ERANGE);
1156
1157         if (end_chunk >= chunk)
1158                 GOTO(out, rc = 0);
1159
1160         LASSERT(end_chunk + 1 == chunk);
1161         /* try to expand this extent to cover @index */
1162         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1163
1164         next = next_extent(ext);
1165         if (next != NULL && next->oe_start <= end_index)
1166                 /* complex mode - overlapped with the next extent,
1167                  * this case will be handled by osc_extent_find() */
1168                 GOTO(out, rc = -EAGAIN);
1169
1170         ext->oe_end = end_index;
1171         ext->oe_grants += chunksize;
1172         *grants -= chunksize;
1173         LASSERT(*grants >= 0);
1174         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1175                  "overlapped after expanding for %lu.\n", index);
1176         EXIT;
1177
1178 out:
1179         osc_object_unlock(obj);
1180         RETURN(rc);
1181 }
1182
1183 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
1184                                   const char *func, int line)
1185 {
1186         struct osc_extent *ext;
1187         int cnt;
1188
1189         CDEBUG(level, "Dump object %p extents at %s:%d, mppr: %u.\n",
1190                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1191
1192         /* osc_object_lock(obj); */
1193         cnt = 1;
1194         for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1195                 OSC_EXTENT_DUMP(level, ext, "in tree %d.\n", cnt++);
1196
1197         cnt = 1;
1198         list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1199                 OSC_EXTENT_DUMP(level, ext, "hp %d.\n", cnt++);
1200
1201         cnt = 1;
1202         list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1203                 OSC_EXTENT_DUMP(level, ext, "urgent %d.\n", cnt++);
1204
1205         cnt = 1;
1206         list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1207                 OSC_EXTENT_DUMP(level, ext, "reading %d.\n", cnt++);
1208         /* osc_object_unlock(obj); */
1209 }
1210
1211 /* ------------------ osc extent end ------------------ */
1212
1213 static inline int osc_is_ready(struct osc_object *osc)
1214 {
1215         return !list_empty(&osc->oo_ready_item) ||
1216                !list_empty(&osc->oo_hp_ready_item);
1217 }
1218
1219 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1220         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1221                (OSC), osc_is_ready(OSC),                                       \
1222                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1223                list_empty_marker(&(OSC)->oo_ready_item),                       \
1224                atomic_read(&(OSC)->oo_nr_writes),                              \
1225                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1226                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1227                atomic_read(&(OSC)->oo_nr_reads),                               \
1228                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1229                ##args)
1230
1231 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1232                           int cmd)
1233 {
1234         struct osc_page *opg  = oap2osc_page(oap);
1235         struct cl_page  *page = oap2cl_page(oap);
1236         int result;
1237
1238         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1239
1240         ENTRY;
1241         result = cl_page_make_ready(env, page, CRT_WRITE);
1242         if (result == 0)
1243                 opg->ops_submit_time = cfs_time_current();
1244         RETURN(result);
1245 }
1246
1247 static int osc_refresh_count(const struct lu_env *env,
1248                              struct osc_async_page *oap, int cmd)
1249 {
1250         struct osc_page  *opg = oap2osc_page(oap);
1251         pgoff_t index = osc_index(oap2osc(oap));
1252         struct cl_object *obj;
1253         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
1254
1255         int result;
1256         loff_t kms;
1257
1258         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1259         LASSERT(!(cmd & OBD_BRW_READ));
1260         LASSERT(opg != NULL);
1261         obj = opg->ops_cl.cpl_obj;
1262
1263         cl_object_attr_lock(obj);
1264         result = cl_object_attr_get(env, obj, attr);
1265         cl_object_attr_unlock(obj);
1266         if (result < 0)
1267                 return result;
1268         kms = attr->cat_kms;
1269         if (cl_offset(obj, index) >= kms)
1270                 /* catch race with truncate */
1271                 return 0;
1272         else if (cl_offset(obj, index + 1) > kms)
1273                 /* catch sub-page write at end of file */
1274                 return kms % PAGE_CACHE_SIZE;
1275         else
1276                 return PAGE_CACHE_SIZE;
1277 }
1278
1279 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1280                           int cmd, int rc)
1281 {
1282         struct osc_page   *opg  = oap2osc_page(oap);
1283         struct cl_page    *page = oap2cl_page(oap);
1284         enum cl_req_type   crt;
1285         int srvlock;
1286
1287         ENTRY;
1288
1289         cmd &= ~OBD_BRW_NOQUOTA;
1290         LASSERTF(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ),
1291                  "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1292         LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1293                 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1294         LASSERT(opg->ops_transfer_pinned);
1295
1296         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1297         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1298         opg->ops_transfer_pinned = 0;
1299
1300         opg->ops_submit_time = 0;
1301         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1302
1303         /* statistic */
1304         if (rc == 0 && srvlock) {
1305                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1306                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1307                 size_t bytes = oap->oap_count;
1308
1309                 if (crt == CRT_READ)
1310                         stats->os_lockless_reads += bytes;
1311                 else
1312                         stats->os_lockless_writes += bytes;
1313         }
1314
1315         /*
1316          * This has to be the last operation with the page, as locks are
1317          * released in cl_page_completion() and nothing except for the
1318          * reference counter protects page from concurrent reclaim.
1319          */
1320         lu_ref_del(&page->cp_reference, "transfer", page);
1321
1322         cl_page_completion(env, page, crt, rc);
1323         cl_page_put(env, page);
1324
1325         RETURN(0);
1326 }
1327
1328 #define OSC_DUMP_GRANT(lvl, cli, fmt, args...) do {                     \
1329         struct client_obd *__tmp = (cli);                               \
1330         CDEBUG(lvl, "%s: grant { dirty: %lu/%lu dirty_pages: %ld/%lu "  \
1331                "dropped: %ld avail: %ld, reserved: %ld, flight: %d }"   \
1332                "lru {in list: %ld, left: %ld, waiters: %d }"fmt"\n",    \
1333                cli_name(__tmp),                                         \
1334                __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages,        \
1335                atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1336                __tmp->cl_lost_grant, __tmp->cl_avail_grant,             \
1337                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,         \
1338                atomic_long_read(&__tmp->cl_lru_in_list),                \
1339                atomic_long_read(&__tmp->cl_lru_busy),                   \
1340                atomic_read(&__tmp->cl_lru_shrinkers), ##args);          \
1341 } while (0)
1342
1343 /* caller must hold loi_list_lock */
1344 static void osc_consume_write_grant(struct client_obd *cli,
1345                                     struct brw_page *pga)
1346 {
1347         assert_spin_locked(&cli->cl_loi_list_lock);
1348         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1349         atomic_long_inc(&obd_dirty_pages);
1350         cli->cl_dirty_pages++;
1351         pga->flag |= OBD_BRW_FROM_GRANT;
1352         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1353                PAGE_CACHE_SIZE, pga, pga->pg);
1354         osc_update_next_shrink(cli);
1355 }
1356
1357 /* the companion to osc_consume_write_grant, called when a brw has completed.
1358  * must be called with the loi lock held. */
1359 static void osc_release_write_grant(struct client_obd *cli,
1360                                     struct brw_page *pga)
1361 {
1362         ENTRY;
1363
1364         assert_spin_locked(&cli->cl_loi_list_lock);
1365         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1366                 EXIT;
1367                 return;
1368         }
1369
1370         pga->flag &= ~OBD_BRW_FROM_GRANT;
1371         atomic_long_dec(&obd_dirty_pages);
1372         cli->cl_dirty_pages--;
1373         if (pga->flag & OBD_BRW_NOCACHE) {
1374                 pga->flag &= ~OBD_BRW_NOCACHE;
1375                 atomic_long_dec(&obd_dirty_transit_pages);
1376                 cli->cl_dirty_transit--;
1377         }
1378         EXIT;
1379 }
1380
1381 /**
1382  * To avoid sleeping with object lock held, it's good for us allocate enough
1383  * grants before entering into critical section.
1384  *
1385  * client_obd_list_lock held by caller
1386  */
1387 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1388 {
1389         int rc = -EDQUOT;
1390
1391         if (cli->cl_avail_grant >= bytes) {
1392                 cli->cl_avail_grant    -= bytes;
1393                 cli->cl_reserved_grant += bytes;
1394                 rc = 0;
1395         }
1396         return rc;
1397 }
1398
1399 static void __osc_unreserve_grant(struct client_obd *cli,
1400                                   unsigned int reserved, unsigned int unused)
1401 {
1402         /* it's quite normal for us to get more grant than reserved.
1403          * Thinking about a case that two extents merged by adding a new
1404          * chunk, we can save one extent tax. If extent tax is greater than
1405          * one chunk, we can save more grant by adding a new chunk */
1406         cli->cl_reserved_grant -= reserved;
1407         if (unused > reserved) {
1408                 cli->cl_avail_grant += reserved;
1409                 cli->cl_lost_grant  += unused - reserved;
1410         } else {
1411                 cli->cl_avail_grant += unused;
1412         }
1413 }
1414
1415 static void osc_unreserve_grant(struct client_obd *cli,
1416                                 unsigned int reserved, unsigned int unused)
1417 {
1418         spin_lock(&cli->cl_loi_list_lock);
1419         __osc_unreserve_grant(cli, reserved, unused);
1420         if (unused > 0)
1421                 osc_wake_cache_waiters(cli);
1422         spin_unlock(&cli->cl_loi_list_lock);
1423 }
1424
1425 /**
1426  * Free grant after IO is finished or canceled.
1427  *
1428  * @lost_grant is used to remember how many grants we have allocated but not
1429  * used, we should return these grants to OST. There're two cases where grants
1430  * can be lost:
1431  * 1. truncate;
1432  * 2. blocksize at OST is less than PAGE_CACHE_SIZE and a partial page was
1433  *    written. In this case OST may use less chunks to serve this partial
1434  *    write. OSTs don't actually know the page size on the client side. so
1435  *    clients have to calculate lost grant by the blocksize on the OST.
1436  *    See filter_grant_check() for details.
1437  */
1438 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1439                            unsigned int lost_grant)
1440 {
1441         unsigned long grant = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
1442
1443         spin_lock(&cli->cl_loi_list_lock);
1444         atomic_long_sub(nr_pages, &obd_dirty_pages);
1445         cli->cl_dirty_pages -= nr_pages;
1446         cli->cl_lost_grant += lost_grant;
1447         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1448                 /* borrow some grant from truncate to avoid the case that
1449                  * truncate uses up all avail grant */
1450                 cli->cl_lost_grant -= grant;
1451                 cli->cl_avail_grant += grant;
1452         }
1453         osc_wake_cache_waiters(cli);
1454         spin_unlock(&cli->cl_loi_list_lock);
1455         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu\n",
1456                lost_grant, cli->cl_lost_grant,
1457                cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_CACHE_SHIFT);
1458 }
1459
1460 /**
1461  * The companion to osc_enter_cache(), called when @oap is no longer part of
1462  * the dirty accounting due to error.
1463  */
1464 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1465 {
1466         spin_lock(&cli->cl_loi_list_lock);
1467         osc_release_write_grant(cli, &oap->oap_brw_page);
1468         spin_unlock(&cli->cl_loi_list_lock);
1469 }
1470
1471 /**
1472  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1473  * is available.
1474  */
1475 static int osc_enter_cache_try(struct client_obd *cli,
1476                                struct osc_async_page *oap,
1477                                int bytes, int transient)
1478 {
1479         int rc;
1480
1481         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1482
1483         rc = osc_reserve_grant(cli, bytes);
1484         if (rc < 0)
1485                 return 0;
1486
1487         if (cli->cl_dirty_pages < cli->cl_dirty_max_pages &&
1488             1 + atomic_long_read(&obd_dirty_pages) <= obd_max_dirty_pages) {
1489                 osc_consume_write_grant(cli, &oap->oap_brw_page);
1490                 if (transient) {
1491                         cli->cl_dirty_transit++;
1492                         atomic_long_inc(&obd_dirty_transit_pages);
1493                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
1494                 }
1495                 rc = 1;
1496         } else {
1497                 __osc_unreserve_grant(cli, bytes, bytes);
1498                 rc = 0;
1499         }
1500         return rc;
1501 }
1502
1503 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1504 {
1505         int rc;
1506         spin_lock(&cli->cl_loi_list_lock);
1507         rc = list_empty(&ocw->ocw_entry);
1508         spin_unlock(&cli->cl_loi_list_lock);
1509         return rc;
1510 }
1511
1512 /**
1513  * The main entry to reserve dirty page accounting. Usually the grant reserved
1514  * in this function will be freed in bulk in osc_free_grant() unless it fails
1515  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1516  *
1517  * The process will be put into sleep if it's already run out of grant.
1518  */
1519 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1520                            struct osc_async_page *oap, int bytes)
1521 {
1522         struct osc_object       *osc = oap->oap_obj;
1523         struct lov_oinfo        *loi = osc->oo_oinfo;
1524         struct osc_cache_waiter  ocw;
1525         struct l_wait_info       lwi;
1526         int                      rc = -EDQUOT;
1527         ENTRY;
1528
1529         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(AT_OFF ? obd_timeout : at_max),
1530                                NULL, LWI_ON_SIGNAL_NOOP, NULL);
1531
1532         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1533
1534         spin_lock(&cli->cl_loi_list_lock);
1535
1536         /* force the caller to try sync io.  this can jump the list
1537          * of queued writes and create a discontiguous rpc stream */
1538         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1539             cli->cl_dirty_max_pages == 0 ||
1540             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) {
1541                 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1542                 GOTO(out, rc = -EDQUOT);
1543         }
1544
1545         /* Hopefully normal case - cache space and write credits available */
1546         if (osc_enter_cache_try(cli, oap, bytes, 0)) {
1547                 OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1548                 GOTO(out, rc = 0);
1549         }
1550
1551         /* We can get here for two reasons: too many dirty pages in cache, or
1552          * run out of grants. In both cases we should write dirty pages out.
1553          * Adding a cache waiter will trigger urgent write-out no matter what
1554          * RPC size will be.
1555          * The exiting condition is no avail grants and no dirty pages caching,
1556          * that really means there is no space on the OST. */
1557         init_waitqueue_head(&ocw.ocw_waitq);
1558         ocw.ocw_oap   = oap;
1559         ocw.ocw_grant = bytes;
1560         while (cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0) {
1561                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1562                 ocw.ocw_rc = 0;
1563                 spin_unlock(&cli->cl_loi_list_lock);
1564
1565                 osc_io_unplug_async(env, cli, NULL);
1566
1567                 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1568                        cli_name(cli), &ocw, oap);
1569
1570                 rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1571
1572                 spin_lock(&cli->cl_loi_list_lock);
1573
1574                 if (rc < 0) {
1575                         /* l_wait_event is interrupted by signal or timed out */
1576                         list_del_init(&ocw.ocw_entry);
1577                         break;
1578                 }
1579                 LASSERT(list_empty(&ocw.ocw_entry));
1580                 rc = ocw.ocw_rc;
1581
1582                 if (rc != -EDQUOT)
1583                         break;
1584                 if (osc_enter_cache_try(cli, oap, bytes, 0)) {
1585                         rc = 0;
1586                         break;
1587                 }
1588         }
1589
1590         switch (rc) {
1591         case 0:
1592                 OSC_DUMP_GRANT(D_CACHE, cli, "finally got grant space\n");
1593                 break;
1594         case -ETIMEDOUT:
1595                 OSC_DUMP_GRANT(D_CACHE, cli,
1596                                "timeout, fall back to sync i/o\n");
1597                 osc_extent_tree_dump(D_CACHE, osc);
1598                 /* fall back to synchronous I/O */
1599                 rc = -EDQUOT;
1600                 break;
1601         case -EINTR:
1602                 /* Ensures restartability - LU-3581 */
1603                 OSC_DUMP_GRANT(D_CACHE, cli, "interrupted\n");
1604                 rc = -ERESTARTSYS;
1605                 break;
1606         case -EDQUOT:
1607                 OSC_DUMP_GRANT(D_CACHE, cli,
1608                                "no grant space, fall back to sync i/o\n");
1609                 break;
1610         default:
1611                 CDEBUG(D_CACHE, "%s: event for cache space @ %p never arrived "
1612                        "due to %d, fall back to sync i/o\n",
1613                        cli_name(cli), &ocw, rc);
1614                 break;
1615         }
1616         EXIT;
1617 out:
1618         spin_unlock(&cli->cl_loi_list_lock);
1619         RETURN(rc);
1620 }
1621
1622 /* caller must hold loi_list_lock */
1623 void osc_wake_cache_waiters(struct client_obd *cli)
1624 {
1625         struct list_head *l, *tmp;
1626         struct osc_cache_waiter *ocw;
1627
1628         ENTRY;
1629         list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1630                 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
1631                 list_del_init(&ocw->ocw_entry);
1632
1633                 ocw->ocw_rc = -EDQUOT;
1634                 /* we can't dirty more */
1635                 if ((cli->cl_dirty_pages  >= cli->cl_dirty_max_pages) ||
1636                     (1 + atomic_long_read(&obd_dirty_pages) >
1637                      obd_max_dirty_pages)) {
1638                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
1639                                "osc max %ld, sys max %ld\n",
1640                                cli->cl_dirty_pages, cli->cl_dirty_max_pages,
1641                                obd_max_dirty_pages);
1642                         goto wakeup;
1643                 }
1644
1645                 if (osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
1646                         ocw->ocw_rc = 0;
1647 wakeup:
1648                 CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld, %d\n",
1649                        ocw, ocw->ocw_oap, cli->cl_avail_grant, ocw->ocw_rc);
1650
1651                 wake_up(&ocw->ocw_waitq);
1652         }
1653
1654         EXIT;
1655 }
1656
1657 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1658 {
1659         int hprpc = !!list_empty(&osc->oo_hp_exts);
1660         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1661 }
1662
1663 /* This maintains the lists of pending pages to read/write for a given object
1664  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1665  * to quickly find objects that are ready to send an RPC. */
1666 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1667                          int cmd)
1668 {
1669         int invalid_import = 0;
1670         ENTRY;
1671
1672         /* if we have an invalid import we want to drain the queued pages
1673          * by forcing them through rpcs that immediately fail and complete
1674          * the pages.  recovery relies on this to empty the queued pages
1675          * before canceling the locks and evicting down the llite pages */
1676         if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1677                 invalid_import = 1;
1678
1679         if (cmd & OBD_BRW_WRITE) {
1680                 if (atomic_read(&osc->oo_nr_writes) == 0)
1681                         RETURN(0);
1682                 if (invalid_import) {
1683                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1684                         RETURN(1);
1685                 }
1686                 if (!list_empty(&osc->oo_hp_exts)) {
1687                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1688                         RETURN(1);
1689                 }
1690                 if (!list_empty(&osc->oo_urgent_exts)) {
1691                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1692                         RETURN(1);
1693                 }
1694                 /* trigger a write rpc stream as long as there are dirtiers
1695                  * waiting for space.  as they're waiting, they're not going to
1696                  * create more pages to coalesce with what's waiting.. */
1697                 if (!list_empty(&cli->cl_cache_waiters)) {
1698                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1699                         RETURN(1);
1700                 }
1701                 if (atomic_read(&osc->oo_nr_writes) >=
1702                     cli->cl_max_pages_per_rpc)
1703                         RETURN(1);
1704         } else {
1705                 if (atomic_read(&osc->oo_nr_reads) == 0)
1706                         RETURN(0);
1707                 if (invalid_import) {
1708                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1709                         RETURN(1);
1710                 }
1711                 /* all read are urgent. */
1712                 if (!list_empty(&osc->oo_reading_exts))
1713                         RETURN(1);
1714         }
1715
1716         RETURN(0);
1717 }
1718
1719 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1720 {
1721         struct client_obd *cli = osc_cli(obj);
1722         if (cmd & OBD_BRW_WRITE) {
1723                 atomic_add(delta, &obj->oo_nr_writes);
1724                 atomic_add(delta, &cli->cl_pending_w_pages);
1725                 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1726         } else {
1727                 atomic_add(delta, &obj->oo_nr_reads);
1728                 atomic_add(delta, &cli->cl_pending_r_pages);
1729                 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1730         }
1731         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1732 }
1733
1734 static int osc_makes_hprpc(struct osc_object *obj)
1735 {
1736         return !list_empty(&obj->oo_hp_exts);
1737 }
1738
1739 static void on_list(struct list_head *item, struct list_head *list,
1740                     int should_be_on)
1741 {
1742         if (list_empty(item) && should_be_on)
1743                 list_add_tail(item, list);
1744         else if (!list_empty(item) && !should_be_on)
1745                 list_del_init(item);
1746 }
1747
1748 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1749  * can find pages to build into rpcs quickly */
1750 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1751 {
1752         if (osc_makes_hprpc(osc)) {
1753                 /* HP rpc */
1754                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1755                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1756         } else {
1757                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1758                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1759                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1760                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1761         }
1762
1763         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1764                 atomic_read(&osc->oo_nr_writes) > 0);
1765
1766         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1767                 atomic_read(&osc->oo_nr_reads) > 0);
1768
1769         return osc_is_ready(osc);
1770 }
1771
1772 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1773 {
1774         int is_ready;
1775
1776         spin_lock(&cli->cl_loi_list_lock);
1777         is_ready = __osc_list_maint(cli, osc);
1778         spin_unlock(&cli->cl_loi_list_lock);
1779
1780         return is_ready;
1781 }
1782
1783 /* this is trying to propogate async writeback errors back up to the
1784  * application.  As an async write fails we record the error code for later if
1785  * the app does an fsync.  As long as errors persist we force future rpcs to be
1786  * sync so that the app can get a sync error and break the cycle of queueing
1787  * pages for which writeback will fail. */
1788 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1789                            int rc)
1790 {
1791         if (rc) {
1792                 if (!ar->ar_rc)
1793                         ar->ar_rc = rc;
1794
1795                 ar->ar_force_sync = 1;
1796                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1797                 return;
1798
1799         }
1800
1801         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1802                 ar->ar_force_sync = 0;
1803 }
1804
1805 /* this must be called holding the loi list lock to give coverage to exit_cache,
1806  * async_flag maintenance, and oap_request */
1807 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1808                               struct osc_async_page *oap, int sent, int rc)
1809 {
1810         struct osc_object *osc = oap->oap_obj;
1811         struct lov_oinfo  *loi = osc->oo_oinfo;
1812         __u64 xid = 0;
1813
1814         ENTRY;
1815         if (oap->oap_request != NULL) {
1816                 xid = ptlrpc_req_xid(oap->oap_request);
1817                 ptlrpc_req_finished(oap->oap_request);
1818                 oap->oap_request = NULL;
1819         }
1820
1821         /* As the transfer for this page is being done, clear the flags */
1822         spin_lock(&oap->oap_lock);
1823         oap->oap_async_flags = 0;
1824         spin_unlock(&oap->oap_lock);
1825         oap->oap_interrupted = 0;
1826
1827         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1828                 spin_lock(&cli->cl_loi_list_lock);
1829                 osc_process_ar(&cli->cl_ar, xid, rc);
1830                 osc_process_ar(&loi->loi_ar, xid, rc);
1831                 spin_unlock(&cli->cl_loi_list_lock);
1832         }
1833
1834         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1835         if (rc)
1836                 CERROR("completion on oap %p obj %p returns %d.\n",
1837                        oap, osc, rc);
1838
1839         EXIT;
1840 }
1841
1842 /**
1843  * Try to add extent to one RPC. We need to think about the following things:
1844  * - # of pages must not be over max_pages_per_rpc
1845  * - extent must be compatible with previous ones
1846  */
1847 static int try_to_add_extent_for_io(struct client_obd *cli,
1848                                     struct osc_extent *ext,
1849                                     struct list_head *rpclist,
1850                                     unsigned int *pc, unsigned int *max_pages)
1851 {
1852         struct osc_extent *tmp;
1853         struct osc_async_page *oap = list_first_entry(&ext->oe_pages,
1854                                                       struct osc_async_page,
1855                                                       oap_pending_item);
1856         ENTRY;
1857
1858         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1859                 ext);
1860
1861         *max_pages = max(ext->oe_mppr, *max_pages);
1862         if (*pc + ext->oe_nr_pages > *max_pages)
1863                 RETURN(0);
1864
1865         list_for_each_entry(tmp, rpclist, oe_link) {
1866                 struct osc_async_page *oap2;
1867                 oap2 = list_first_entry(&tmp->oe_pages, struct osc_async_page,
1868                                         oap_pending_item);
1869                 EASSERT(tmp->oe_owner == current, tmp);
1870 #if 0
1871                 if (overlapped(tmp, ext)) {
1872                         OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
1873                         EASSERT(0, ext);
1874                 }
1875 #endif
1876                 if (oap2cl_page(oap)->cp_type != oap2cl_page(oap2)->cp_type) {
1877                         CDEBUG(D_CACHE, "Do not permit different type of IO"
1878                                         " for a same RPC\n");
1879                         RETURN(0);
1880                 }
1881
1882                 if (tmp->oe_srvlock != ext->oe_srvlock ||
1883                     !tmp->oe_grants != !ext->oe_grants ||
1884                     tmp->oe_no_merge || ext->oe_no_merge)
1885                         RETURN(0);
1886
1887                 /* remove break for strict check */
1888                 break;
1889         }
1890
1891         *pc += ext->oe_nr_pages;
1892         list_move_tail(&ext->oe_link, rpclist);
1893         ext->oe_owner = current;
1894         RETURN(1);
1895 }
1896
1897 /**
1898  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1899  * get_write_extent() takes all appropriate extents in atomic.
1900  *
1901  * The following policy is used to collect extents for IO:
1902  * 1. Add as many HP extents as possible;
1903  * 2. Add the first urgent extent in urgent extent list and take it out of
1904  *    urgent list;
1905  * 3. Add subsequent extents of this urgent extent;
1906  * 4. If urgent list is not empty, goto 2;
1907  * 5. Traverse the extent tree from the 1st extent;
1908  * 6. Above steps exit if there is no space in this RPC.
1909  */
1910 static unsigned int get_write_extents(struct osc_object *obj,
1911                                       struct list_head *rpclist)
1912 {
1913         struct client_obd *cli = osc_cli(obj);
1914         struct osc_extent *ext;
1915         unsigned int page_count = 0;
1916         unsigned int max_pages = cli->cl_max_pages_per_rpc;
1917
1918         LASSERT(osc_object_is_locked(obj));
1919         while (!list_empty(&obj->oo_hp_exts)) {
1920                 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
1921                                  oe_link);
1922                 LASSERT(ext->oe_state == OES_CACHE);
1923                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1924                                               &max_pages))
1925                         return page_count;
1926                 EASSERT(ext->oe_nr_pages <= max_pages, ext);
1927         }
1928         if (page_count == max_pages)
1929                 return page_count;
1930
1931         while (!list_empty(&obj->oo_urgent_exts)) {
1932                 ext = list_entry(obj->oo_urgent_exts.next,
1933                                  struct osc_extent, oe_link);
1934                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1935                                               &max_pages))
1936                         return page_count;
1937
1938                 if (!ext->oe_intree)
1939                         continue;
1940
1941                 while ((ext = next_extent(ext)) != NULL) {
1942                         if ((ext->oe_state != OES_CACHE) ||
1943                             (!list_empty(&ext->oe_link) &&
1944                              ext->oe_owner != NULL))
1945                                 continue;
1946
1947                         if (!try_to_add_extent_for_io(cli, ext, rpclist,
1948                                                       &page_count, &max_pages))
1949                                 return page_count;
1950                 }
1951         }
1952         if (page_count == max_pages)
1953                 return page_count;
1954
1955         ext = first_extent(obj);
1956         while (ext != NULL) {
1957                 if ((ext->oe_state != OES_CACHE) ||
1958                     /* this extent may be already in current rpclist */
1959                     (!list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
1960                         ext = next_extent(ext);
1961                         continue;
1962                 }
1963
1964                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1965                                               &max_pages))
1966                         return page_count;
1967
1968                 ext = next_extent(ext);
1969         }
1970         return page_count;
1971 }
1972
1973 static int
1974 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
1975                    struct osc_object *osc)
1976 __must_hold(osc)
1977 {
1978         struct list_head   rpclist = LIST_HEAD_INIT(rpclist);
1979         struct osc_extent *ext;
1980         struct osc_extent *tmp;
1981         struct osc_extent *first = NULL;
1982         unsigned int page_count = 0;
1983         int srvlock = 0;
1984         int rc = 0;
1985         ENTRY;
1986
1987         LASSERT(osc_object_is_locked(osc));
1988
1989         page_count = get_write_extents(osc, &rpclist);
1990         LASSERT(equi(page_count == 0, list_empty(&rpclist)));
1991
1992         if (list_empty(&rpclist))
1993                 RETURN(0);
1994
1995         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
1996
1997         list_for_each_entry(ext, &rpclist, oe_link) {
1998                 LASSERT(ext->oe_state == OES_CACHE ||
1999                         ext->oe_state == OES_LOCK_DONE);
2000                 if (ext->oe_state == OES_CACHE)
2001                         osc_extent_state_set(ext, OES_LOCKING);
2002                 else
2003                         osc_extent_state_set(ext, OES_RPC);
2004         }
2005
2006         /* we're going to grab page lock, so release object lock because
2007          * lock order is page lock -> object lock. */
2008         osc_object_unlock(osc);
2009
2010         list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2011                 if (ext->oe_state == OES_LOCKING) {
2012                         rc = osc_extent_make_ready(env, ext);
2013                         if (unlikely(rc < 0)) {
2014                                 list_del_init(&ext->oe_link);
2015                                 osc_extent_finish(env, ext, 0, rc);
2016                                 continue;
2017                         }
2018                 }
2019                 if (first == NULL) {
2020                         first = ext;
2021                         srvlock = ext->oe_srvlock;
2022                 } else {
2023                         LASSERT(srvlock == ext->oe_srvlock);
2024                 }
2025         }
2026
2027         if (!list_empty(&rpclist)) {
2028                 LASSERT(page_count > 0);
2029                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2030                 LASSERT(list_empty(&rpclist));
2031         }
2032
2033         osc_object_lock(osc);
2034         RETURN(rc);
2035 }
2036
2037 /**
2038  * prepare pages for ASYNC io and put pages in send queue.
2039  *
2040  * \param cmd OBD_BRW_* macroses
2041  * \param lop pending pages
2042  *
2043  * \return zero if no page added to send queue.
2044  * \return 1 if pages successfully added to send queue.
2045  * \return negative on errors.
2046  */
2047 static int
2048 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2049                   struct osc_object *osc)
2050 __must_hold(osc)
2051 {
2052         struct osc_extent *ext;
2053         struct osc_extent *next;
2054         struct list_head rpclist = LIST_HEAD_INIT(rpclist);
2055         unsigned int page_count = 0;
2056         unsigned int max_pages = cli->cl_max_pages_per_rpc;
2057         int rc = 0;
2058         ENTRY;
2059
2060         LASSERT(osc_object_is_locked(osc));
2061         list_for_each_entry_safe(ext, next,
2062                                      &osc->oo_reading_exts, oe_link) {
2063                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2064                 if (!try_to_add_extent_for_io(cli, ext, &rpclist, &page_count,
2065                                               &max_pages))
2066                         break;
2067                 osc_extent_state_set(ext, OES_RPC);
2068                 EASSERT(ext->oe_nr_pages <= max_pages, ext);
2069         }
2070         LASSERT(page_count <= max_pages);
2071
2072         osc_update_pending(osc, OBD_BRW_READ, -page_count);
2073
2074         if (!list_empty(&rpclist)) {
2075                 osc_object_unlock(osc);
2076
2077                 LASSERT(page_count > 0);
2078                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2079                 LASSERT(list_empty(&rpclist));
2080
2081                 osc_object_lock(osc);
2082         }
2083         RETURN(rc);
2084 }
2085
2086 #define list_to_obj(list, item) ({                                            \
2087         struct list_head *__tmp = (list)->next;                               \
2088         list_del_init(__tmp);                                         \
2089         list_entry(__tmp, struct osc_object, oo_##item);                      \
2090 })
2091
2092 /* This is called by osc_check_rpcs() to find which objects have pages that
2093  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2094 static struct osc_object *osc_next_obj(struct client_obd *cli)
2095 {
2096         ENTRY;
2097
2098         /* First return objects that have blocked locks so that they
2099          * will be flushed quickly and other clients can get the lock,
2100          * then objects which have pages ready to be stuffed into RPCs */
2101         if (!list_empty(&cli->cl_loi_hp_ready_list))
2102                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2103         if (!list_empty(&cli->cl_loi_ready_list))
2104                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2105
2106         /* then if we have cache waiters, return all objects with queued
2107          * writes.  This is especially important when many small files
2108          * have filled up the cache and not been fired into rpcs because
2109          * they don't pass the nr_pending/object threshhold */
2110         if (!list_empty(&cli->cl_cache_waiters) &&
2111             !list_empty(&cli->cl_loi_write_list))
2112                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2113
2114         /* then return all queued objects when we have an invalid import
2115          * so that they get flushed */
2116         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2117                 if (!list_empty(&cli->cl_loi_write_list))
2118                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2119                                            write_item));
2120                 if (!list_empty(&cli->cl_loi_read_list))
2121                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2122                                            read_item));
2123         }
2124         RETURN(NULL);
2125 }
2126
2127 /* called with the loi list lock held */
2128 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2129 __must_hold(&cli->cl_loi_list_lock)
2130 {
2131         struct osc_object *osc;
2132         int rc = 0;
2133         ENTRY;
2134
2135         while ((osc = osc_next_obj(cli)) != NULL) {
2136                 struct cl_object *obj = osc2cl(osc);
2137                 struct lu_ref_link link;
2138
2139                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2140
2141                 if (osc_max_rpc_in_flight(cli, osc)) {
2142                         __osc_list_maint(cli, osc);
2143                         break;
2144                 }
2145
2146                 cl_object_get(obj);
2147                 spin_unlock(&cli->cl_loi_list_lock);
2148                 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2149
2150                 /* attempt some read/write balancing by alternating between
2151                  * reads and writes in an object.  The makes_rpc checks here
2152                  * would be redundant if we were getting read/write work items
2153                  * instead of objects.  we don't want send_oap_rpc to drain a
2154                  * partial read pending queue when we're given this object to
2155                  * do io on writes while there are cache waiters */
2156                 osc_object_lock(osc);
2157                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2158                         rc = osc_send_write_rpc(env, cli, osc);
2159                         if (rc < 0) {
2160                                 CERROR("Write request failed with %d\n", rc);
2161
2162                                 /* osc_send_write_rpc failed, mostly because of
2163                                  * memory pressure.
2164                                  *
2165                                  * It can't break here, because if:
2166                                  *  - a page was submitted by osc_io_submit, so
2167                                  *    page locked;
2168                                  *  - no request in flight
2169                                  *  - no subsequent request
2170                                  * The system will be in live-lock state,
2171                                  * because there is no chance to call
2172                                  * osc_io_unplug() and osc_check_rpcs() any
2173                                  * more. pdflush can't help in this case,
2174                                  * because it might be blocked at grabbing
2175                                  * the page lock as we mentioned.
2176                                  *
2177                                  * Anyway, continue to drain pages. */
2178                                 /* break; */
2179                         }
2180                 }
2181                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2182                         rc = osc_send_read_rpc(env, cli, osc);
2183                         if (rc < 0)
2184                                 CERROR("Read request failed with %d\n", rc);
2185                 }
2186                 osc_object_unlock(osc);
2187
2188                 osc_list_maint(cli, osc);
2189                 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2190                 cl_object_put(env, obj);
2191
2192                 spin_lock(&cli->cl_loi_list_lock);
2193         }
2194 }
2195
2196 static int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2197                           struct osc_object *osc, int async)
2198 {
2199         int rc = 0;
2200
2201         if (osc != NULL && osc_list_maint(cli, osc) == 0)
2202                 return 0;
2203
2204         if (!async) {
2205                 spin_lock(&cli->cl_loi_list_lock);
2206                 osc_check_rpcs(env, cli);
2207                 spin_unlock(&cli->cl_loi_list_lock);
2208         } else {
2209                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2210                 LASSERT(cli->cl_writeback_work != NULL);
2211                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2212         }
2213         return rc;
2214 }
2215
2216 static int osc_io_unplug_async(const struct lu_env *env,
2217                                 struct client_obd *cli, struct osc_object *osc)
2218 {
2219         return osc_io_unplug0(env, cli, osc, 1);
2220 }
2221
2222 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
2223                    struct osc_object *osc)
2224 {
2225         (void)osc_io_unplug0(env, cli, osc, 0);
2226 }
2227
2228 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2229                         struct page *page, loff_t offset)
2230 {
2231         struct obd_export     *exp = osc_export(osc);
2232         struct osc_async_page *oap = &ops->ops_oap;
2233         ENTRY;
2234
2235         if (!page)
2236                 return cfs_size_round(sizeof(*oap));
2237
2238         oap->oap_magic = OAP_MAGIC;
2239         oap->oap_cli = &exp->exp_obd->u.cli;
2240         oap->oap_obj = osc;
2241
2242         oap->oap_page = page;
2243         oap->oap_obj_off = offset;
2244         LASSERT(!(offset & ~PAGE_MASK));
2245
2246         if (!client_is_remote(exp) && cfs_capable(CFS_CAP_SYS_RESOURCE))
2247                 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2248
2249         INIT_LIST_HEAD(&oap->oap_pending_item);
2250         INIT_LIST_HEAD(&oap->oap_rpc_item);
2251
2252         spin_lock_init(&oap->oap_lock);
2253         CDEBUG(D_INFO, "oap %p page %p obj off "LPU64"\n",
2254                oap, page, oap->oap_obj_off);
2255         RETURN(0);
2256 }
2257
2258 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2259                        struct osc_page *ops)
2260 {
2261         struct osc_io *oio = osc_env_io(env);
2262         struct osc_extent     *ext = NULL;
2263         struct osc_async_page *oap = &ops->ops_oap;
2264         struct client_obd     *cli = oap->oap_cli;
2265         struct osc_object     *osc = oap->oap_obj;
2266         pgoff_t index;
2267         unsigned int tmp;
2268         unsigned int grants = 0;
2269         u32    brw_flags = OBD_BRW_ASYNC;
2270         int    cmd = OBD_BRW_WRITE;
2271         int    need_release = 0;
2272         int    rc = 0;
2273         ENTRY;
2274
2275         if (oap->oap_magic != OAP_MAGIC)
2276                 RETURN(-EINVAL);
2277
2278         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2279                 RETURN(-EIO);
2280
2281         if (!list_empty(&oap->oap_pending_item) ||
2282             !list_empty(&oap->oap_rpc_item))
2283                 RETURN(-EBUSY);
2284
2285         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2286         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2287         if (!client_is_remote(osc_export(osc)) &&
2288             cfs_capable(CFS_CAP_SYS_RESOURCE)) {
2289                 brw_flags |= OBD_BRW_NOQUOTA;
2290                 cmd |= OBD_BRW_NOQUOTA;
2291         }
2292
2293         /* check if the file's owner/group is over quota */
2294         if (!(cmd & OBD_BRW_NOQUOTA)) {
2295                 struct cl_object *obj;
2296                 struct cl_attr   *attr;
2297                 unsigned int qid[MAXQUOTAS];
2298
2299                 obj = cl_object_top(&osc->oo_cl);
2300                 attr = &osc_env_info(env)->oti_attr;
2301
2302                 cl_object_attr_lock(obj);
2303                 rc = cl_object_attr_get(env, obj, attr);
2304                 cl_object_attr_unlock(obj);
2305
2306                 qid[USRQUOTA] = attr->cat_uid;
2307                 qid[GRPQUOTA] = attr->cat_gid;
2308                 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2309                         rc = -EDQUOT;
2310                 if (rc)
2311                         RETURN(rc);
2312         }
2313
2314         oap->oap_cmd = cmd;
2315         oap->oap_page_off = ops->ops_from;
2316         oap->oap_count = ops->ops_to - ops->ops_from;
2317         /* No need to hold a lock here,
2318          * since this page is not in any list yet. */
2319         oap->oap_async_flags = 0;
2320         oap->oap_brw_flags = brw_flags;
2321
2322         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2323                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2324
2325         index = osc_index(oap2osc(oap));
2326
2327         /* Add this page into extent by the following steps:
2328          * 1. if there exists an active extent for this IO, mostly this page
2329          *    can be added to the active extent and sometimes we need to
2330          *    expand extent to accomodate this page;
2331          * 2. otherwise, a new extent will be allocated. */
2332
2333         ext = oio->oi_active;
2334         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2335                 /* one chunk plus extent overhead must be enough to write this
2336                  * page */
2337                 grants = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2338                 if (ext->oe_end >= index)
2339                         grants = 0;
2340
2341                 /* it doesn't need any grant to dirty this page */
2342                 spin_lock(&cli->cl_loi_list_lock);
2343                 rc = osc_enter_cache_try(cli, oap, grants, 0);
2344                 spin_unlock(&cli->cl_loi_list_lock);
2345                 if (rc == 0) { /* try failed */
2346                         grants = 0;
2347                         need_release = 1;
2348                 } else if (ext->oe_end < index) {
2349                         tmp = grants;
2350                         /* try to expand this extent */
2351                         rc = osc_extent_expand(ext, index, &tmp);
2352                         if (rc < 0) {
2353                                 need_release = 1;
2354                                 /* don't free reserved grant */
2355                         } else {
2356                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2357                                                 "expanded for %lu.\n", index);
2358                                 osc_unreserve_grant(cli, grants, tmp);
2359                                 grants = 0;
2360                         }
2361                 }
2362                 rc = 0;
2363         } else if (ext != NULL) {
2364                 /* index is located outside of active extent */
2365                 need_release = 1;
2366         }
2367         if (need_release) {
2368                 osc_extent_release(env, ext);
2369                 oio->oi_active = NULL;
2370                 ext = NULL;
2371         }
2372
2373         if (ext == NULL) {
2374                 tmp = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2375
2376                 /* try to find new extent to cover this page */
2377                 LASSERT(oio->oi_active == NULL);
2378                 /* we may have allocated grant for this page if we failed
2379                  * to expand the previous active extent. */
2380                 LASSERT(ergo(grants > 0, grants >= tmp));
2381
2382                 rc = 0;
2383                 if (grants == 0) {
2384                         /* we haven't allocated grant for this page. */
2385                         rc = osc_enter_cache(env, cli, oap, tmp);
2386                         if (rc == 0)
2387                                 grants = tmp;
2388                 }
2389
2390                 tmp = grants;
2391                 if (rc == 0) {
2392                         ext = osc_extent_find(env, osc, index, &tmp);
2393                         if (IS_ERR(ext)) {
2394                                 LASSERT(tmp == grants);
2395                                 osc_exit_cache(cli, oap);
2396                                 rc = PTR_ERR(ext);
2397                                 ext = NULL;
2398                         } else {
2399                                 oio->oi_active = ext;
2400                         }
2401                 }
2402                 if (grants > 0)
2403                         osc_unreserve_grant(cli, grants, tmp);
2404         }
2405
2406         LASSERT(ergo(rc == 0, ext != NULL));
2407         if (ext != NULL) {
2408                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2409                          ext, "index = %lu.\n", index);
2410                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2411
2412                 osc_object_lock(osc);
2413                 if (ext->oe_nr_pages == 0)
2414                         ext->oe_srvlock = ops->ops_srvlock;
2415                 else
2416                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2417                 ++ext->oe_nr_pages;
2418                 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2419                 osc_object_unlock(osc);
2420         }
2421         RETURN(rc);
2422 }
2423
2424 int osc_teardown_async_page(const struct lu_env *env,
2425                             struct osc_object *obj, struct osc_page *ops)
2426 {
2427         struct osc_async_page *oap = &ops->ops_oap;
2428         int rc = 0;
2429         ENTRY;
2430
2431         LASSERT(oap->oap_magic == OAP_MAGIC);
2432
2433         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2434                oap, ops, osc_index(oap2osc(oap)));
2435
2436         if (!list_empty(&oap->oap_rpc_item)) {
2437                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2438                 rc = -EBUSY;
2439         } else if (!list_empty(&oap->oap_pending_item)) {
2440                 struct osc_extent *ext = NULL;
2441
2442                 osc_object_lock(obj);
2443                 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2444                 osc_object_unlock(obj);
2445                 /* only truncated pages are allowed to be taken out.
2446                  * See osc_extent_truncate() and osc_cache_truncate_start()
2447                  * for details. */
2448                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2449                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2450                                         osc_index(oap2osc(oap)));
2451                         rc = -EBUSY;
2452                 }
2453                 if (ext != NULL)
2454                         osc_extent_put(env, ext);
2455         }
2456         RETURN(rc);
2457 }
2458
2459 /**
2460  * This is called when a page is picked up by kernel to write out.
2461  *
2462  * We should find out the corresponding extent and add the whole extent
2463  * into urgent list. The extent may be being truncated or used, handle it
2464  * carefully.
2465  */
2466 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2467                          struct osc_page *ops)
2468 {
2469         struct osc_extent *ext   = NULL;
2470         struct osc_object *obj   = cl2osc(ops->ops_cl.cpl_obj);
2471         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2472         pgoff_t            index = osc_index(ops);
2473         struct osc_async_page *oap = &ops->ops_oap;
2474         bool unplug = false;
2475         int rc = 0;
2476         ENTRY;
2477
2478         osc_object_lock(obj);
2479         ext = osc_extent_lookup(obj, index);
2480         if (ext == NULL) {
2481                 osc_extent_tree_dump(D_ERROR, obj);
2482                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2483         }
2484
2485         switch (ext->oe_state) {
2486         case OES_RPC:
2487         case OES_LOCK_DONE:
2488                 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2489                 LASSERT(0);
2490                 break;
2491         case OES_LOCKING:
2492                 /* If we know this extent is being written out, we should abort
2493                  * so that the writer can make this page ready. Otherwise, there
2494                  * exists a deadlock problem because other process can wait for
2495                  * page writeback bit holding page lock; and meanwhile in
2496                  * vvp_page_make_ready(), we need to grab page lock before
2497                  * really sending the RPC. */
2498         case OES_TRUNC:
2499                 /* race with truncate, page will be redirtied */
2500         case OES_ACTIVE:
2501                 /* The extent is active so we need to abort and let the caller
2502                  * re-dirty the page. If we continued on here, and we were the
2503                  * one making the extent active, we could deadlock waiting for
2504                  * the page writeback to clear but it won't because the extent
2505                  * is active and won't be written out. */
2506                 GOTO(out, rc = -EAGAIN);
2507         default:
2508                 break;
2509         }
2510
2511         rc = cl_page_prep(env, io, cp, CRT_WRITE);
2512         if (rc)
2513                 GOTO(out, rc);
2514
2515         spin_lock(&oap->oap_lock);
2516         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2517         spin_unlock(&oap->oap_lock);
2518
2519         if (memory_pressure_get())
2520                 ext->oe_memalloc = 1;
2521
2522         ext->oe_urgent = 1;
2523         if (ext->oe_state == OES_CACHE) {
2524                 OSC_EXTENT_DUMP(D_CACHE, ext,
2525                                 "flush page %p make it urgent.\n", oap);
2526                 if (list_empty(&ext->oe_link))
2527                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2528                 unplug = true;
2529         }
2530         rc = 0;
2531         EXIT;
2532
2533 out:
2534         osc_object_unlock(obj);
2535         osc_extent_put(env, ext);
2536         if (unplug)
2537                 osc_io_unplug_async(env, osc_cli(obj), obj);
2538         return rc;
2539 }
2540
2541 /**
2542  * this is called when a sync waiter receives an interruption.  Its job is to
2543  * get the caller woken as soon as possible.  If its page hasn't been put in an
2544  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2545  * desiring interruption which will forcefully complete the rpc once the rpc
2546  * has timed out.
2547  */
2548 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2549 {
2550         struct osc_async_page *oap = &ops->ops_oap;
2551         struct osc_object     *obj = oap->oap_obj;
2552         struct client_obd     *cli = osc_cli(obj);
2553         struct osc_extent     *ext;
2554         struct osc_extent     *found = NULL;
2555         struct list_head            *plist;
2556         pgoff_t index = osc_index(ops);
2557         int     rc = -EBUSY;
2558         int     cmd;
2559         ENTRY;
2560
2561         LASSERT(!oap->oap_interrupted);
2562         oap->oap_interrupted = 1;
2563
2564         /* Find out the caching extent */
2565         osc_object_lock(obj);
2566         if (oap->oap_cmd & OBD_BRW_WRITE) {
2567                 plist = &obj->oo_urgent_exts;
2568                 cmd   = OBD_BRW_WRITE;
2569         } else {
2570                 plist = &obj->oo_reading_exts;
2571                 cmd   = OBD_BRW_READ;
2572         }
2573         list_for_each_entry(ext, plist, oe_link) {
2574                 if (ext->oe_start <= index && ext->oe_end >= index) {
2575                         LASSERT(ext->oe_state == OES_LOCK_DONE);
2576                         /* For OES_LOCK_DONE state extent, it has already held
2577                          * a refcount for RPC. */
2578                         found = osc_extent_get(ext);
2579                         break;
2580                 }
2581         }
2582         if (found != NULL) {
2583                 list_del_init(&found->oe_link);
2584                 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2585                 osc_object_unlock(obj);
2586
2587                 osc_extent_finish(env, found, 0, -EINTR);
2588                 osc_extent_put(env, found);
2589                 rc = 0;
2590         } else {
2591                 osc_object_unlock(obj);
2592                 /* ok, it's been put in an rpc. only one oap gets a request
2593                  * reference */
2594                 if (oap->oap_request != NULL) {
2595                         ptlrpc_mark_interrupted(oap->oap_request);
2596                         ptlrpcd_wake(oap->oap_request);
2597                         ptlrpc_req_finished(oap->oap_request);
2598                         oap->oap_request = NULL;
2599                 }
2600         }
2601
2602         osc_list_maint(cli, obj);
2603         RETURN(rc);
2604 }
2605
2606 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
2607                          struct list_head *list, int cmd, int brw_flags)
2608 {
2609         struct client_obd     *cli = osc_cli(obj);
2610         struct osc_extent     *ext;
2611         struct osc_async_page *oap;
2612         int     page_count = 0;
2613         int     mppr       = cli->cl_max_pages_per_rpc;
2614         bool    can_merge   = true;
2615         pgoff_t start      = CL_PAGE_EOF;
2616         pgoff_t end        = 0;
2617         ENTRY;
2618
2619         list_for_each_entry(oap, list, oap_pending_item) {
2620                 struct osc_page *opg = oap2osc_page(oap);
2621                 pgoff_t index = osc_index(opg);
2622
2623                 if (index > end)
2624                         end = index;
2625                 if (index < start)
2626                         start = index;
2627                 ++page_count;
2628                 mppr <<= (page_count > mppr);
2629
2630                 if (unlikely(opg->ops_from > 0 || opg->ops_to < PAGE_SIZE))
2631                         can_merge = false;
2632         }
2633
2634         ext = osc_extent_alloc(obj);
2635         if (ext == NULL) {
2636                 struct osc_async_page *tmp;
2637
2638                 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2639                         list_del_init(&oap->oap_pending_item);
2640                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2641                 }
2642                 RETURN(-ENOMEM);
2643         }
2644
2645         ext->oe_rw = !!(cmd & OBD_BRW_READ);
2646         ext->oe_sync = 1;
2647         ext->oe_no_merge = !can_merge;
2648         ext->oe_urgent = 1;
2649         ext->oe_start = start;
2650         ext->oe_end = ext->oe_max_end = end;
2651         ext->oe_obj = obj;
2652         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2653         ext->oe_nr_pages = page_count;
2654         ext->oe_mppr = mppr;
2655         list_splice_init(list, &ext->oe_pages);
2656
2657         osc_object_lock(obj);
2658         /* Reuse the initial refcount for RPC, don't drop it */
2659         osc_extent_state_set(ext, OES_LOCK_DONE);
2660         if (cmd & OBD_BRW_WRITE) {
2661                 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2662                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2663         } else {
2664                 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2665                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2666         }
2667         osc_object_unlock(obj);
2668
2669         osc_io_unplug_async(env, cli, obj);
2670         RETURN(0);
2671 }
2672
2673 /**
2674  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2675  */
2676 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
2677                              __u64 size, struct osc_extent **extp)
2678 {
2679         struct client_obd *cli = osc_cli(obj);
2680         struct osc_extent *ext;
2681         struct osc_extent *waiting = NULL;
2682         pgoff_t index;
2683         struct list_head list = LIST_HEAD_INIT(list);
2684         int result = 0;
2685         bool partial;
2686         ENTRY;
2687
2688         /* pages with index greater or equal to index will be truncated. */
2689         index = cl_index(osc2cl(obj), size);
2690         partial = size > cl_offset(osc2cl(obj), index);
2691
2692 again:
2693         osc_object_lock(obj);
2694         ext = osc_extent_search(obj, index);
2695         if (ext == NULL)
2696                 ext = first_extent(obj);
2697         else if (ext->oe_end < index)
2698                 ext = next_extent(ext);
2699         while (ext != NULL) {
2700                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2701
2702                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2703                         /* if ext is in urgent state, it means there must exist
2704                          * a page already having been flushed by write_page().
2705                          * We have to wait for this extent because we can't
2706                          * truncate that page. */
2707                         OSC_EXTENT_DUMP(D_CACHE, ext,
2708                                         "waiting for busy extent\n");
2709                         waiting = osc_extent_get(ext);
2710                         break;
2711                 }
2712
2713                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:"LPU64".\n", size);
2714
2715                 osc_extent_get(ext);
2716                 if (ext->oe_state == OES_ACTIVE) {
2717                         /* though we grab inode mutex for write path, but we
2718                          * release it before releasing extent(in osc_io_end()),
2719                          * so there is a race window that an extent is still
2720                          * in OES_ACTIVE when truncate starts. */
2721                         LASSERT(!ext->oe_trunc_pending);
2722                         ext->oe_trunc_pending = 1;
2723                 } else {
2724                         EASSERT(ext->oe_state == OES_CACHE, ext);
2725                         osc_extent_state_set(ext, OES_TRUNC);
2726                         osc_update_pending(obj, OBD_BRW_WRITE,
2727                                            -ext->oe_nr_pages);
2728                 }
2729                 EASSERT(list_empty(&ext->oe_link), ext);
2730                 list_add_tail(&ext->oe_link, &list);
2731
2732                 ext = next_extent(ext);
2733         }
2734         osc_object_unlock(obj);
2735
2736         osc_list_maint(cli, obj);
2737
2738         while (!list_empty(&list)) {
2739                 int rc;
2740
2741                 ext = list_entry(list.next, struct osc_extent, oe_link);
2742                 list_del_init(&ext->oe_link);
2743
2744                 /* extent may be in OES_ACTIVE state because inode mutex
2745                  * is released before osc_io_end() in file write case */
2746                 if (ext->oe_state != OES_TRUNC)
2747                         osc_extent_wait(env, ext, OES_TRUNC);
2748
2749                 rc = osc_extent_truncate(ext, index, partial);
2750                 if (rc < 0) {
2751                         if (result == 0)
2752                                 result = rc;
2753
2754                         OSC_EXTENT_DUMP(D_ERROR, ext,
2755                                         "truncate error %d\n", rc);
2756                 } else if (ext->oe_nr_pages == 0) {
2757                         osc_extent_remove(ext);
2758                 } else {
2759                         /* this must be an overlapped extent which means only
2760                          * part of pages in this extent have been truncated.
2761                          */
2762                         EASSERTF(ext->oe_start <= index, ext,
2763                                  "trunc index = %lu/%d.\n", index, partial);
2764                         /* fix index to skip this partially truncated extent */
2765                         index = ext->oe_end + 1;
2766                         partial = false;
2767
2768                         /* we need to hold this extent in OES_TRUNC state so
2769                          * that no writeback will happen. This is to avoid
2770                          * BUG 17397.
2771                          * Only partial truncate can reach here, if @size is
2772                          * not zero, the caller should provide a valid @extp. */
2773                         LASSERT(*extp == NULL);
2774                         *extp = osc_extent_get(ext);
2775                         OSC_EXTENT_DUMP(D_CACHE, ext,
2776                                         "trunc at "LPU64"\n", size);
2777                 }
2778                 osc_extent_put(env, ext);
2779         }
2780         if (waiting != NULL) {
2781                 int rc;
2782
2783                 /* ignore the result of osc_extent_wait the write initiator
2784                  * should take care of it. */
2785                 rc = osc_extent_wait(env, waiting, OES_INV);
2786                 if (rc < 0)
2787                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2788
2789                 osc_extent_put(env, waiting);
2790                 waiting = NULL;
2791                 goto again;
2792         }
2793         RETURN(result);
2794 }
2795
2796 /**
2797  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2798  */
2799 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext)
2800 {
2801         if (ext != NULL) {
2802                 struct osc_object *obj = ext->oe_obj;
2803                 bool unplug = false;
2804
2805                 EASSERT(ext->oe_nr_pages > 0, ext);
2806                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2807                 EASSERT(!ext->oe_urgent, ext);
2808
2809                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2810                 osc_object_lock(obj);
2811                 osc_extent_state_set(ext, OES_CACHE);
2812                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2813                         ext->oe_urgent = 1;
2814                         list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2815                         unplug = true;
2816                 }
2817                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2818                 osc_object_unlock(obj);
2819                 osc_extent_put(env, ext);
2820
2821                 if (unplug)
2822                         osc_io_unplug_async(env, osc_cli(obj), obj);
2823         }
2824 }
2825
2826 /**
2827  * Wait for extents in a specific range to be written out.
2828  * The caller must have called osc_cache_writeback_range() to issue IO
2829  * otherwise it will take a long time for this function to finish.
2830  *
2831  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2832  * nobody else can dirty this range of file while we're waiting for
2833  * extents to be written.
2834  */
2835 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2836                          pgoff_t start, pgoff_t end)
2837 {
2838         struct osc_extent *ext;
2839         pgoff_t index = start;
2840         int     result = 0;
2841         ENTRY;
2842
2843 again:
2844         osc_object_lock(obj);
2845         ext = osc_extent_search(obj, index);
2846         if (ext == NULL)
2847                 ext = first_extent(obj);
2848         else if (ext->oe_end < index)
2849                 ext = next_extent(ext);
2850         while (ext != NULL) {
2851                 int rc;
2852
2853                 if (ext->oe_start > end)
2854                         break;
2855
2856                 if (!ext->oe_fsync_wait) {
2857                         ext = next_extent(ext);
2858                         continue;
2859                 }
2860
2861                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2862                              ext->oe_hp || ext->oe_urgent), ext);
2863                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2864                              !ext->oe_hp && ext->oe_urgent), ext);
2865
2866                 index = ext->oe_end + 1;
2867                 osc_extent_get(ext);
2868                 osc_object_unlock(obj);
2869
2870                 rc = osc_extent_wait(env, ext, OES_INV);
2871                 if (result == 0)
2872                         result = rc;
2873                 osc_extent_put(env, ext);
2874                 goto again;
2875         }
2876         osc_object_unlock(obj);
2877
2878         OSC_IO_DEBUG(obj, "sync file range.\n");
2879         RETURN(result);
2880 }
2881
2882 /**
2883  * Called to write out a range of osc object.
2884  *
2885  * @hp     : should be set this is caused by lock cancel;
2886  * @discard: is set if dirty pages should be dropped - file will be deleted or
2887  *         truncated, this implies there is no partially discarding extents.
2888  *
2889  * Return how many pages will be issued, or error code if error occurred.
2890  */
2891 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2892                               pgoff_t start, pgoff_t end, int hp, int discard)
2893 {
2894         struct osc_extent *ext;
2895         struct list_head discard_list = LIST_HEAD_INIT(discard_list);
2896         bool unplug = false;
2897         int result = 0;
2898         ENTRY;
2899
2900         osc_object_lock(obj);
2901         ext = osc_extent_search(obj, start);
2902         if (ext == NULL)
2903                 ext = first_extent(obj);
2904         else if (ext->oe_end < start)
2905                 ext = next_extent(ext);
2906         while (ext != NULL) {
2907                 if (ext->oe_start > end)
2908                         break;
2909
2910                 ext->oe_fsync_wait = 1;
2911                 switch (ext->oe_state) {
2912                 case OES_CACHE:
2913                         result += ext->oe_nr_pages;
2914                         if (!discard) {
2915                                 struct list_head *list = NULL;
2916                                 if (hp) {
2917                                         EASSERT(!ext->oe_hp, ext);
2918                                         ext->oe_hp = 1;
2919                                         list = &obj->oo_hp_exts;
2920                                 } else if (!ext->oe_urgent) {
2921                                         ext->oe_urgent = 1;
2922                                         list = &obj->oo_urgent_exts;
2923                                 }
2924                                 if (list != NULL)
2925                                         list_move_tail(&ext->oe_link, list);
2926                                 unplug = true;
2927                         } else {
2928                                 /* the only discarder is lock cancelling, so
2929                                  * [start, end] must contain this extent */
2930                                 EASSERT(ext->oe_start >= start &&
2931                                         ext->oe_max_end <= end, ext);
2932                                 osc_extent_state_set(ext, OES_LOCKING);
2933                                 ext->oe_owner = current;
2934                                 list_move_tail(&ext->oe_link,
2935                                                    &discard_list);
2936                                 osc_update_pending(obj, OBD_BRW_WRITE,
2937                                                    -ext->oe_nr_pages);
2938                         }
2939                         break;
2940                 case OES_ACTIVE:
2941                         /* It's pretty bad to wait for ACTIVE extents, because
2942                          * we don't know how long we will wait for it to be
2943                          * flushed since it may be blocked at awaiting more
2944                          * grants. We do this for the correctness of fsync. */
2945                         LASSERT(hp == 0 && discard == 0);
2946                         ext->oe_urgent = 1;
2947                         break;
2948                 case OES_TRUNC:
2949                         /* this extent is being truncated, can't do anything
2950                          * for it now. it will be set to urgent after truncate
2951                          * is finished in osc_cache_truncate_end(). */
2952                 default:
2953                         break;
2954                 }
2955                 ext = next_extent(ext);
2956         }
2957         osc_object_unlock(obj);
2958
2959         LASSERT(ergo(!discard, list_empty(&discard_list)));
2960         if (!list_empty(&discard_list)) {
2961                 struct osc_extent *tmp;
2962                 int rc;
2963
2964                 osc_list_maint(osc_cli(obj), obj);
2965                 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
2966                         list_del_init(&ext->oe_link);
2967                         EASSERT(ext->oe_state == OES_LOCKING, ext);
2968
2969                         /* Discard caching pages. We don't actually write this
2970                          * extent out but we complete it as if we did. */
2971                         rc = osc_extent_make_ready(env, ext);
2972                         if (unlikely(rc < 0)) {
2973                                 OSC_EXTENT_DUMP(D_ERROR, ext,
2974                                                 "make_ready returned %d\n", rc);
2975                                 if (result >= 0)
2976                                         result = rc;
2977                         }
2978
2979                         /* finish the extent as if the pages were sent */
2980                         osc_extent_finish(env, ext, 0, 0);
2981                 }
2982         }
2983
2984         if (unplug)
2985                 osc_io_unplug(env, osc_cli(obj), obj);
2986
2987         if (hp || discard) {
2988                 int rc;
2989                 rc = osc_cache_wait_range(env, obj, start, end);
2990                 if (result >= 0 && rc < 0)
2991                         result = rc;
2992         }
2993
2994         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
2995         RETURN(result);
2996 }
2997
2998 /**
2999  * Returns a list of pages by a given [start, end] of \a obj.
3000  *
3001  * \param resched If not NULL, then we give up before hogging CPU for too
3002  * long and set *resched = 1, in that case caller should implement a retry
3003  * logic.
3004  *
3005  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3006  * crucial in the face of [offset, EOF] locks.
3007  *
3008  * Return at least one page in @queue unless there is no covered page.
3009  */
3010 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3011                         struct osc_object *osc, pgoff_t start, pgoff_t end,
3012                         osc_page_gang_cbt cb, void *cbdata)
3013 {
3014         struct osc_page *ops;
3015         void            **pvec;
3016         pgoff_t         idx;
3017         unsigned int    nr;
3018         unsigned int    i;
3019         unsigned int    j;
3020         int             res = CLP_GANG_OKAY;
3021         bool            tree_lock = true;
3022         ENTRY;
3023
3024         idx = start;
3025         pvec = osc_env_info(env)->oti_pvec;
3026         spin_lock(&osc->oo_tree_lock);
3027         while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3028                                             idx, OTI_PVEC_SIZE)) > 0) {
3029                 struct cl_page *page;
3030                 bool end_of_region = false;
3031
3032                 for (i = 0, j = 0; i < nr; ++i) {
3033                         ops = pvec[i];
3034                         pvec[i] = NULL;
3035
3036                         idx = osc_index(ops);
3037                         if (idx > end) {
3038                                 end_of_region = true;
3039                                 break;
3040                         }
3041
3042                         page = ops->ops_cl.cpl_page;
3043                         LASSERT(page->cp_type == CPT_CACHEABLE);
3044                         if (page->cp_state == CPS_FREEING)
3045                                 continue;
3046
3047                         cl_page_get(page);
3048                         lu_ref_add_atomic(&page->cp_reference,
3049                                           "gang_lookup", current);
3050                         pvec[j++] = ops;
3051                 }
3052                 ++idx;
3053
3054                 /*
3055                  * Here a delicate locking dance is performed. Current thread
3056                  * holds a reference to a page, but has to own it before it
3057                  * can be placed into queue. Owning implies waiting, so
3058                  * radix-tree lock is to be released. After a wait one has to
3059                  * check that pages weren't truncated (cl_page_own() returns
3060                  * error in the latter case).
3061                  */
3062                 spin_unlock(&osc->oo_tree_lock);
3063                 tree_lock = false;
3064
3065                 for (i = 0; i < j; ++i) {
3066                         ops = pvec[i];
3067                         if (res == CLP_GANG_OKAY)
3068                                 res = (*cb)(env, io, ops, cbdata);
3069
3070                         page = ops->ops_cl.cpl_page;
3071                         lu_ref_del(&page->cp_reference, "gang_lookup", current);
3072                         cl_page_put(env, page);
3073                 }
3074                 if (nr < OTI_PVEC_SIZE || end_of_region)
3075                         break;
3076
3077                 if (res == CLP_GANG_OKAY && need_resched())
3078                         res = CLP_GANG_RESCHED;
3079                 if (res != CLP_GANG_OKAY)
3080                         break;
3081
3082                 spin_lock(&osc->oo_tree_lock);
3083                 tree_lock = true;
3084         }
3085         if (tree_lock)
3086                 spin_unlock(&osc->oo_tree_lock);
3087         RETURN(res);
3088 }
3089
3090 /**
3091  * Check if page @page is covered by an extra lock or discard it.
3092  */
3093 static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3094                                 struct osc_page *ops, void *cbdata)
3095 {
3096         struct osc_thread_info *info = osc_env_info(env);
3097         struct osc_object *osc = cbdata;
3098         pgoff_t index;
3099
3100         index = osc_index(ops);
3101         if (index >= info->oti_fn_index) {
3102                 struct ldlm_lock *tmp;
3103                 struct cl_page *page = ops->ops_cl.cpl_page;
3104
3105                 /* refresh non-overlapped index */
3106                 tmp = osc_dlmlock_at_pgoff(env, osc, index,
3107                                            OSC_DAP_FL_TEST_LOCK);
3108                 if (tmp != NULL) {
3109                         __u64 end = tmp->l_policy_data.l_extent.end;
3110                         /* Cache the first-non-overlapped index so as to skip
3111                          * all pages within [index, oti_fn_index). This is safe
3112                          * because if tmp lock is canceled, it will discard
3113                          * these pages. */
3114                         info->oti_fn_index = cl_index(osc2cl(osc), end + 1);
3115                         if (end == OBD_OBJECT_EOF)
3116                                 info->oti_fn_index = CL_PAGE_EOF;
3117                         LDLM_LOCK_PUT(tmp);
3118                 } else if (cl_page_own(env, io, page) == 0) {
3119                         /* discard the page */
3120                         cl_page_discard(env, io, page);
3121                         cl_page_disown(env, io, page);
3122                 } else {
3123                         LASSERT(page->cp_state == CPS_FREEING);
3124                 }
3125         }
3126
3127         info->oti_next_index = index + 1;
3128         return CLP_GANG_OKAY;
3129 }
3130
3131 static int discard_cb(const struct lu_env *env, struct cl_io *io,
3132                       struct osc_page *ops, void *cbdata)
3133 {
3134         struct osc_thread_info *info = osc_env_info(env);
3135         struct cl_page *page = ops->ops_cl.cpl_page;
3136
3137         /* page is top page. */
3138         info->oti_next_index = osc_index(ops) + 1;
3139         if (cl_page_own(env, io, page) == 0) {
3140                 if (!ergo(page->cp_type == CPT_CACHEABLE,
3141                           !PageDirty(cl_page_vmpage(page))))
3142                         CL_PAGE_DEBUG(D_ERROR, env, page,
3143                                         "discard dirty page?\n");
3144
3145                 /* discard the page */
3146                 cl_page_discard(env, io, page);
3147                 cl_page_disown(env, io, page);
3148         } else {
3149                 LASSERT(page->cp_state == CPS_FREEING);
3150         }
3151
3152         return CLP_GANG_OKAY;
3153 }
3154
3155 /**
3156  * Discard pages protected by the given lock. This function traverses radix
3157  * tree to find all covering pages and discard them. If a page is being covered
3158  * by other locks, it should remain in cache.
3159  *
3160  * If error happens on any step, the process continues anyway (the reasoning
3161  * behind this being that lock cancellation cannot be delayed indefinitely).
3162  */
3163 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3164                            pgoff_t start, pgoff_t end, enum cl_lock_mode mode)
3165 {
3166         struct osc_thread_info *info = osc_env_info(env);
3167         struct cl_io *io = &info->oti_io;
3168         osc_page_gang_cbt cb;
3169         int res;
3170         int result;
3171
3172         ENTRY;
3173
3174         io->ci_obj = cl_object_top(osc2cl(osc));
3175         io->ci_ignore_layout = 1;
3176         result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3177         if (result != 0)
3178                 GOTO(out, result);
3179
3180         cb = mode == CLM_READ ? check_and_discard_cb : discard_cb;
3181         info->oti_fn_index = info->oti_next_index = start;
3182         do {
3183                 res = osc_page_gang_lookup(env, io, osc,
3184                                            info->oti_next_index, end, cb, osc);
3185                 if (info->oti_next_index > end)
3186                         break;
3187
3188                 if (res == CLP_GANG_RESCHED)
3189                         cond_resched();
3190         } while (res != CLP_GANG_OKAY);
3191 out:
3192         cl_io_fini(env, io);
3193         RETURN(result);
3194 }
3195
3196
3197 /** @} osc */