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