lustre/osc/osc_cache.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  *
  28  */
  29 /*
  30  * This file is part of Lustre, http://www.lustre.org/
  31  * Lustre is a trademark of Sun Microsystems, Inc.
  32  *
  33  * osc cache management.
  34  *
  35  * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_OSC
  39
  40 #include <lustre_osc.h>
  41
  42 #include "osc_internal.h"
  43
  44 static int extent_debug; /* set it to be true for more debug */
  45
  46 static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
  47 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
  48                            enum osc_extent_state state);
  49 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
  50                               struct osc_async_page *oap, int sent, int rc);
  51 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
  52                           int cmd);
  53 static int osc_refresh_count(const struct lu_env *env,
  54                              struct osc_async_page *oap, int cmd);
  55 static int osc_io_unplug_async(const struct lu_env *env,
  56                                struct client_obd *cli, struct osc_object *osc);
  57 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
  58                            unsigned int lost_grant, unsigned int dirty_grant);
  59
  60 static void osc_extent_tree_dump0(int mask, struct osc_object *obj,
  61                                   const char *func, int line);
  62 #define osc_extent_tree_dump(mask, obj) \
  63         osc_extent_tree_dump0(mask, obj, __func__, __LINE__)
  64
  65 static void osc_unreserve_grant(struct client_obd *cli, unsigned int reserved,
  66                                 unsigned int unused);
  67
  68 /** \addtogroup osc
  69  *  @{
  70  */
  71
  72 /* ------------------ osc extent ------------------ */
  73 static inline char *ext_flags(struct osc_extent *ext, char *flags)
  74 {
  75         char *buf = flags;
  76         *buf++ = ext->oe_rw ? 'r' : 'w';
  77         if (ext->oe_intree)
  78                 *buf++ = 'i';
  79         if (ext->oe_sync)
  80                 *buf++ = 'S';
  81         if (ext->oe_srvlock)
  82                 *buf++ = 's';
  83         if (ext->oe_hp)
  84                 *buf++ = 'h';
  85         if (ext->oe_urgent)
  86                 *buf++ = 'u';
  87         if (ext->oe_memalloc)
  88                 *buf++ = 'm';
  89         if (ext->oe_trunc_pending)
  90                 *buf++ = 't';
  91         if (ext->oe_fsync_wait)
  92                 *buf++ = 'Y';
  93         *buf = 0;
  94         return flags;
  95 }
  96
  97 static inline char list_empty_marker(struct list_head *list)
  98 {
  99         return list_empty(list) ? '-' : '+';
 100 }
 101
 102 #define EXTSTR       "[%lu -> %lu/%lu]"
 103 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
 104 static const char *oes_strings[] = {
 105         "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
 106
 107 #define OSC_EXTENT_DUMP_WITH_LOC(file, func, line, mask, extent, fmt, ...) do {\
 108         static struct cfs_debug_limit_state cdls;                             \
 109         struct osc_extent *__ext = (extent);                                  \
 110         char __buf[16];                                                       \
 111                                                                               \
 112         __CDEBUG_WITH_LOC(file, func, line, mask, &cdls,                      \
 113                 "extent %p@{" EXTSTR ", "                                     \
 114                 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt,          \
 115                 /* ----- extent part 0 ----- */                               \
 116                 __ext, EXTPARA(__ext),                                        \
 117                 /* ----- part 1 ----- */                                      \
 118                 atomic_read(&__ext->oe_refc),                                 \
 119                 atomic_read(&__ext->oe_users),                                \
 120                 list_empty_marker(&__ext->oe_link),                           \
 121                 oes_strings[__ext->oe_state], ext_flags(__ext, __buf),        \
 122                 __ext->oe_obj,                                                \
 123                 /* ----- part 2 ----- */                                      \
 124                 __ext->oe_grants, __ext->oe_nr_pages,                         \
 125                 list_empty_marker(&__ext->oe_pages),                          \
 126                 waitqueue_active(&__ext->oe_waitq) ? '+' : '-',               \
 127                 __ext->oe_dlmlock, __ext->oe_mppr, __ext->oe_owner,           \
 128                 /* ----- part 4 ----- */                                      \
 129                 ## __VA_ARGS__);                                              \
 130         if (mask == D_ERROR && __ext->oe_dlmlock != NULL)                     \
 131                 LDLM_ERROR(__ext->oe_dlmlock, "extent: %p", __ext);           \
 132         else                                                                  \
 133                 LDLM_DEBUG(__ext->oe_dlmlock, "extent: %p", __ext);           \
 134 } while (0)
 135
 136 #define OSC_EXTENT_DUMP(mask, ext, fmt, ...)                            \
 137         OSC_EXTENT_DUMP_WITH_LOC(__FILE__, __func__, __LINE__,          \
 138                                  mask, ext, fmt, ## __VA_ARGS__)
 139
 140 #undef EASSERTF
 141 #define EASSERTF(expr, ext, fmt, args...) do {                          \
 142         if (!(expr)) {                                                  \
 143                 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args);           \
 144                 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj);           \
 145                 LASSERT(expr);                                          \
 146         }                                                               \
 147 } while (0)
 148
 149 #undef EASSERT
 150 #define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
 151
 152 static inline struct osc_extent *rb_extent(struct rb_node *n)
 153 {
 154         if (n == NULL)
 155                 return NULL;
 156
 157         return container_of(n, struct osc_extent, oe_node);
 158 }
 159
 160 static inline struct osc_extent *next_extent(struct osc_extent *ext)
 161 {
 162         if (ext == NULL)
 163                 return NULL;
 164
 165         LASSERT(ext->oe_intree);
 166         return rb_extent(rb_next(&ext->oe_node));
 167 }
 168
 169 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
 170 {
 171         if (ext == NULL)
 172                 return NULL;
 173
 174         LASSERT(ext->oe_intree);
 175         return rb_extent(rb_prev(&ext->oe_node));
 176 }
 177
 178 static inline struct osc_extent *first_extent(struct osc_object *obj)
 179 {
 180         return rb_extent(rb_first(&obj->oo_root));
 181 }
 182
 183 /* object must be locked by caller. */
 184 static int osc_extent_sanity_check0(struct osc_extent *ext,
 185                                     const char *func, const int line)
 186 {
 187         struct osc_object *obj = ext->oe_obj;
 188         struct osc_async_page *oap;
 189         size_t page_count;
 190         int rc = 0;
 191
 192         if (!osc_object_is_locked(obj))
 193                 GOTO(out, rc = 9);
 194
 195         if (ext->oe_state >= OES_STATE_MAX)
 196                 GOTO(out, rc = 10);
 197
 198         if (atomic_read(&ext->oe_refc) <= 0)
 199                 GOTO(out, rc = 20);
 200
 201         if (atomic_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
 202                 GOTO(out, rc = 30);
 203
 204         switch (ext->oe_state) {
 205         case OES_INV:
 206                 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
 207                         GOTO(out, rc = 35);
 208                 GOTO(out, rc = 0);
 209                 break;
 210         case OES_ACTIVE:
 211                 if (atomic_read(&ext->oe_users) == 0)
 212                         GOTO(out, rc = 40);
 213                 if (ext->oe_hp)
 214                         GOTO(out, rc = 50);
 215                 if (ext->oe_fsync_wait && !ext->oe_urgent)
 216                         GOTO(out, rc = 55);
 217                 break;
 218         case OES_CACHE:
 219                 if (ext->oe_grants == 0)
 220                         GOTO(out, rc = 60);
 221                 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
 222                         GOTO(out, rc = 65);
 223                 /* fallthrough */
 224         default:
 225                 if (atomic_read(&ext->oe_users) > 0)
 226                         GOTO(out, rc = 70);
 227         }
 228
 229         if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
 230                 GOTO(out, rc = 80);
 231
 232         if (ext->oe_sync && ext->oe_grants > 0)
 233                 GOTO(out, rc = 90);
 234
 235         if (ext->oe_dlmlock != NULL &&
 236             ext->oe_dlmlock->l_resource->lr_type == LDLM_EXTENT &&
 237             !ldlm_is_failed(ext->oe_dlmlock)) {
 238                 struct ldlm_extent *extent;
 239
 240                 extent = &ext->oe_dlmlock->l_policy_data.l_extent;
 241                 if (!(extent->start <= cl_offset(osc2cl(obj), ext->oe_start) &&
 242                       extent->end   >= cl_offset(osc2cl(obj), ext->oe_max_end)))
 243                         GOTO(out, rc = 100);
 244
 245                 if (!(ext->oe_dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP)))
 246                         GOTO(out, rc = 102);
 247         }
 248
 249         if (ext->oe_nr_pages > ext->oe_mppr)
 250                 GOTO(out, rc = 105);
 251
 252         /* Do not verify page list if extent is in RPC. This is because an
 253          * in-RPC extent is supposed to be exclusively accessible w/o lock. */
 254         if (ext->oe_state > OES_CACHE)
 255                 GOTO(out, rc = 0);
 256
 257         if (!extent_debug)
 258                 GOTO(out, rc = 0);
 259
 260         page_count = 0;
 261         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
 262                 pgoff_t index = osc_index(oap2osc(oap));
 263                 ++page_count;
 264                 if (index > ext->oe_end || index < ext->oe_start)
 265                         GOTO(out, rc = 110);
 266         }
 267         if (page_count != ext->oe_nr_pages)
 268                 GOTO(out, rc = 120);
 269
 270 out:
 271         if (rc != 0)
 272                 OSC_EXTENT_DUMP_WITH_LOC(__FILE__, func, line, D_ERROR, ext,
 273                                          "sanity check %p failed: rc = %d\n",
 274                                          ext, rc);
 275         return rc;
 276 }
 277
 278 #define sanity_check_nolock(ext) \
 279         osc_extent_sanity_check0(ext, __func__, __LINE__)
 280
 281 #define sanity_check(ext) ({                                                   \
 282         int __res;                                                             \
 283         osc_object_lock((ext)->oe_obj);                                        \
 284         __res = sanity_check_nolock(ext);                                      \
 285         osc_object_unlock((ext)->oe_obj);                                      \
 286         __res;                                                                 \
 287 })
 288
 289
 290 /**
 291  * sanity check - to make sure there is no overlapped extent in the tree.
 292  */
 293 static int osc_extent_is_overlapped(struct osc_object *obj,
 294                                     struct osc_extent *ext)
 295 {
 296         struct osc_extent *tmp;
 297
 298         LASSERT(osc_object_is_locked(obj));
 299
 300         if (!extent_debug)
 301                 return 0;
 302
 303         for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
 304                 if (tmp == ext)
 305                         continue;
 306                 if (tmp->oe_end >= ext->oe_start &&
 307                     tmp->oe_start <= ext->oe_end)
 308                         return 1;
 309         }
 310         return 0;
 311 }
 312
 313 static void osc_extent_state_set(struct osc_extent *ext, int state)
 314 {
 315         LASSERT(osc_object_is_locked(ext->oe_obj));
 316         LASSERT(state >= OES_INV && state < OES_STATE_MAX);
 317
 318         /* Never try to sanity check a state changing extent :-) */
 319         /* LASSERT(sanity_check_nolock(ext) == 0); */
 320
 321         /* TODO: validate the state machine */
 322         ext->oe_state = state;
 323         wake_up_all(&ext->oe_waitq);
 324 }
 325
 326 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
 327 {
 328         struct osc_extent *ext;
 329
 330         OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_NOFS);
 331         if (ext == NULL)
 332                 return NULL;
 333
 334         RB_CLEAR_NODE(&ext->oe_node);
 335         ext->oe_obj = obj;
 336         cl_object_get(osc2cl(obj));
 337         atomic_set(&ext->oe_refc, 1);
 338         atomic_set(&ext->oe_users, 0);
 339         INIT_LIST_HEAD(&ext->oe_link);
 340         ext->oe_state = OES_INV;
 341         INIT_LIST_HEAD(&ext->oe_pages);
 342         init_waitqueue_head(&ext->oe_waitq);
 343         ext->oe_dlmlock = NULL;
 344
 345         return ext;
 346 }
 347
 348 static void osc_extent_free(struct osc_extent *ext)
 349 {
 350         OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
 351 }
 352
 353 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
 354 {
 355         LASSERT(atomic_read(&ext->oe_refc) >= 0);
 356         atomic_inc(&ext->oe_refc);
 357         return ext;
 358 }
 359
 360 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
 361 {
 362         LASSERT(atomic_read(&ext->oe_refc) > 0);
 363         if (atomic_dec_and_test(&ext->oe_refc)) {
 364                 LASSERT(list_empty(&ext->oe_link));
 365                 LASSERT(atomic_read(&ext->oe_users) == 0);
 366                 LASSERT(ext->oe_state == OES_INV);
 367                 LASSERT(!ext->oe_intree);
 368
 369                 if (ext->oe_dlmlock != NULL) {
 370                         lu_ref_add(&ext->oe_dlmlock->l_reference,
 371                                    "osc_extent", ext);
 372                         LDLM_LOCK_PUT(ext->oe_dlmlock);
 373                         ext->oe_dlmlock = NULL;
 374                 }
 375                 cl_object_put(env, osc2cl(ext->oe_obj));
 376                 osc_extent_free(ext);
 377         }
 378 }
 379
 380 /**
 381  * osc_extent_put_trust() is a special version of osc_extent_put() when
 382  * it's known that the caller is not the last user. This is to address the
 383  * problem of lacking of lu_env ;-).
 384  */
 385 static void osc_extent_put_trust(struct osc_extent *ext)
 386 {
 387         LASSERT(atomic_read(&ext->oe_refc) > 1);
 388         LASSERT(osc_object_is_locked(ext->oe_obj));
 389         atomic_dec(&ext->oe_refc);
 390 }
 391
 392 /**
 393  * Return the extent which includes pgoff @index, or return the greatest
 394  * previous extent in the tree.
 395  */
 396 static struct osc_extent *osc_extent_search(struct osc_object *obj,
 397                                             pgoff_t index)
 398 {
 399         struct rb_node    *n = obj->oo_root.rb_node;
 400         struct osc_extent *tmp, *p = NULL;
 401
 402         LASSERT(osc_object_is_locked(obj));
 403         while (n != NULL) {
 404                 tmp = rb_extent(n);
 405                 if (index < tmp->oe_start) {
 406                         n = n->rb_left;
 407                 } else if (index > tmp->oe_end) {
 408                         p = rb_extent(n);
 409                         n = n->rb_right;
 410                 } else {
 411                         return tmp;
 412                 }
 413         }
 414         return p;
 415 }
 416
 417 /*
 418  * Return the extent covering @index, otherwise return NULL.
 419  * caller must have held object lock.
 420  */
 421 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
 422                                             pgoff_t index)
 423 {
 424         struct osc_extent *ext;
 425
 426         ext = osc_extent_search(obj, index);
 427         if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
 428                 return osc_extent_get(ext);
 429         return NULL;
 430 }
 431
 432 /* caller must have held object lock. */
 433 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
 434 {
 435         struct rb_node   **n      = &obj->oo_root.rb_node;
 436         struct rb_node    *parent = NULL;
 437         struct osc_extent *tmp;
 438
 439         LASSERT(ext->oe_intree == 0);
 440         LASSERT(ext->oe_obj == obj);
 441         LASSERT(osc_object_is_locked(obj));
 442         while (*n != NULL) {
 443                 tmp = rb_extent(*n);
 444                 parent = *n;
 445
 446                 if (ext->oe_end < tmp->oe_start)
 447                         n = &(*n)->rb_left;
 448                 else if (ext->oe_start > tmp->oe_end)
 449                         n = &(*n)->rb_right;
 450                 else
 451                         EASSERTF(0, tmp, EXTSTR"\n", EXTPARA(ext));
 452         }
 453         rb_link_node(&ext->oe_node, parent, n);
 454         rb_insert_color(&ext->oe_node, &obj->oo_root);
 455         osc_extent_get(ext);
 456         ext->oe_intree = 1;
 457 }
 458
 459 /* caller must have held object lock. */
 460 static void osc_extent_erase(struct osc_extent *ext)
 461 {
 462         struct osc_object *obj = ext->oe_obj;
 463         LASSERT(osc_object_is_locked(obj));
 464         if (ext->oe_intree) {
 465                 rb_erase(&ext->oe_node, &obj->oo_root);
 466                 ext->oe_intree = 0;
 467                 /* rbtree held a refcount */
 468                 osc_extent_put_trust(ext);
 469         }
 470 }
 471
 472 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
 473 {
 474         struct osc_object *obj = ext->oe_obj;
 475
 476         LASSERT(osc_object_is_locked(obj));
 477         LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
 478         if (ext->oe_state == OES_CACHE) {
 479                 osc_extent_state_set(ext, OES_ACTIVE);
 480                 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
 481         }
 482         atomic_inc(&ext->oe_users);
 483         list_del_init(&ext->oe_link);
 484         return osc_extent_get(ext);
 485 }
 486
 487 static void __osc_extent_remove(struct osc_extent *ext)
 488 {
 489         LASSERT(osc_object_is_locked(ext->oe_obj));
 490         LASSERT(list_empty(&ext->oe_pages));
 491         osc_extent_erase(ext);
 492         list_del_init(&ext->oe_link);
 493         osc_extent_state_set(ext, OES_INV);
 494         OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
 495 }
 496
 497 static void osc_extent_remove(struct osc_extent *ext)
 498 {
 499         struct osc_object *obj = ext->oe_obj;
 500
 501         osc_object_lock(obj);
 502         __osc_extent_remove(ext);
 503         osc_object_unlock(obj);
 504 }
 505
 506 /**
 507  * This function is used to merge extents to get better performance. It checks
 508  * if @cur and @victim are contiguous at block level.
 509  */
 510 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
 511                             struct osc_extent *victim)
 512 {
 513         struct osc_object       *obj = cur->oe_obj;
 514         struct client_obd       *cli = osc_cli(obj);
 515         pgoff_t                  chunk_start;
 516         pgoff_t                  chunk_end;
 517         int                      ppc_bits;
 518
 519         LASSERT(cur->oe_state == OES_CACHE);
 520         LASSERT(osc_object_is_locked(obj));
 521         if (victim == NULL)
 522                 return -EINVAL;
 523
 524         if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
 525                 return -EBUSY;
 526
 527         if (cur->oe_max_end != victim->oe_max_end)
 528                 return -ERANGE;
 529
 530         LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
 531         ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_SHIFT;
 532         chunk_start = cur->oe_start >> ppc_bits;
 533         chunk_end   = cur->oe_end   >> ppc_bits;
 534         if (chunk_start   != (victim->oe_end >> ppc_bits) + 1 &&
 535             chunk_end + 1 != victim->oe_start >> ppc_bits)
 536                 return -ERANGE;
 537
 538         /* overall extent size should not exceed the max supported limit
 539          * reported by the server */
 540         if (cur->oe_end - cur->oe_start + 1 +
 541             victim->oe_end - victim->oe_start + 1 > cli->cl_max_extent_pages)
 542                 return -ERANGE;
 543
 544         OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
 545
 546         cur->oe_start     = min(cur->oe_start, victim->oe_start);
 547         cur->oe_end       = max(cur->oe_end,   victim->oe_end);
 548         /* per-extent tax should be accounted only once for the whole extent */
 549         cur->oe_grants   += victim->oe_grants - cli->cl_grant_extent_tax;
 550         cur->oe_nr_pages += victim->oe_nr_pages;
 551         /* only the following bits are needed to merge */
 552         cur->oe_urgent   |= victim->oe_urgent;
 553         cur->oe_memalloc |= victim->oe_memalloc;
 554         list_splice_init(&victim->oe_pages, &cur->oe_pages);
 555         list_del_init(&victim->oe_link);
 556         victim->oe_nr_pages = 0;
 557
 558         osc_extent_get(victim);
 559         __osc_extent_remove(victim);
 560         osc_extent_put(env, victim);
 561
 562         OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
 563         return 0;
 564 }
 565
 566 /**
 567  * Drop user count of osc_extent, and unplug IO asynchronously.
 568  */
 569 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
 570 {
 571         struct osc_object *obj = ext->oe_obj;
 572         struct client_obd *cli = osc_cli(obj);
 573         int rc = 0;
 574         ENTRY;
 575
 576         LASSERT(atomic_read(&ext->oe_users) > 0);
 577         LASSERT(sanity_check(ext) == 0);
 578         LASSERT(ext->oe_grants > 0);
 579
 580         if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
 581                 LASSERT(ext->oe_state == OES_ACTIVE);
 582                 if (ext->oe_trunc_pending) {
 583                         /* a truncate process is waiting for this extent.
 584                          * This may happen due to a race, check
 585                          * osc_cache_truncate_start(). */
 586                         osc_extent_state_set(ext, OES_TRUNC);
 587                         ext->oe_trunc_pending = 0;
 588                         osc_object_unlock(obj);
 589                 } else {
 590                         int grant = 0;
 591
 592                         osc_extent_state_set(ext, OES_CACHE);
 593                         osc_update_pending(obj, OBD_BRW_WRITE,
 594                                            ext->oe_nr_pages);
 595
 596                         /* try to merge the previous and next extent. */
 597                         if (osc_extent_merge(env, ext, prev_extent(ext)) == 0)
 598                                 grant += cli->cl_grant_extent_tax;
 599                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 600                                 grant += cli->cl_grant_extent_tax;
 601
 602                         if (ext->oe_urgent)
 603                                 list_move_tail(&ext->oe_link,
 604                                                &obj->oo_urgent_exts);
 605                         else if (ext->oe_nr_pages == ext->oe_mppr) {
 606                                 list_move_tail(&ext->oe_link,
 607                                                &obj->oo_full_exts);
 608                         }
 609                         osc_object_unlock(obj);
 610                         if (grant > 0)
 611                                 osc_unreserve_grant(cli, 0, grant);
 612                 }
 613
 614                 osc_io_unplug_async(env, cli, obj);
 615         }
 616         osc_extent_put(env, ext);
 617         RETURN(rc);
 618 }
 619
 620 static inline int overlapped(struct osc_extent *ex1, struct osc_extent *ex2)
 621 {
 622         return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
 623 }
 624
 625 /**
 626  * Find or create an extent which includes @index, core function to manage
 627  * extent tree.
 628  */
 629 static struct osc_extent *osc_extent_find(const struct lu_env *env,
 630                                           struct osc_object *obj, pgoff_t index,
 631                                           unsigned int *grants)
 632 {
 633         struct client_obd *cli = osc_cli(obj);
 634         struct osc_lock   *olck;
 635         struct cl_lock_descr *descr;
 636         struct osc_extent *cur;
 637         struct osc_extent *ext;
 638         struct osc_extent *conflict = NULL;
 639         struct osc_extent *found = NULL;
 640         pgoff_t    chunk;
 641         pgoff_t    max_end;
 642         unsigned int max_pages; /* max_pages_per_rpc */
 643         unsigned int chunksize;
 644         int        ppc_bits; /* pages per chunk bits */
 645         pgoff_t    chunk_mask;
 646         int        rc;
 647         ENTRY;
 648
 649         cur = osc_extent_alloc(obj);
 650         if (cur == NULL)
 651                 RETURN(ERR_PTR(-ENOMEM));
 652
 653         olck = osc_env_io(env)->oi_write_osclock;
 654         LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
 655         LASSERT(olck->ols_state == OLS_GRANTED);
 656
 657         descr = &olck->ols_cl.cls_lock->cll_descr;
 658         LASSERT(descr->cld_mode >= CLM_WRITE);
 659
 660         LASSERTF(cli->cl_chunkbits >= PAGE_SHIFT,
 661                  "chunkbits: %u\n", cli->cl_chunkbits);
 662         ppc_bits   = cli->cl_chunkbits - PAGE_SHIFT;
 663         chunk_mask = ~((1 << ppc_bits) - 1);
 664         chunksize  = 1 << cli->cl_chunkbits;
 665         chunk      = index >> ppc_bits;
 666
 667         /* align end to RPC edge. */
 668         max_pages = cli->cl_max_pages_per_rpc;
 669         if ((max_pages & ~chunk_mask) != 0) {
 670                 CERROR("max_pages: %#x chunkbits: %u chunk_mask: %#lx\n",
 671                        max_pages, cli->cl_chunkbits, chunk_mask);
 672                 RETURN(ERR_PTR(-EINVAL));
 673         }
 674         max_end = index - (index % max_pages) + max_pages - 1;
 675         max_end = min_t(pgoff_t, max_end, descr->cld_end);
 676
 677         /* initialize new extent by parameters so far */
 678         cur->oe_max_end = max_end;
 679         cur->oe_start   = index & chunk_mask;
 680         cur->oe_end     = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
 681         if (cur->oe_start < descr->cld_start)
 682                 cur->oe_start = descr->cld_start;
 683         if (cur->oe_end > max_end)
 684                 cur->oe_end = max_end;
 685         cur->oe_grants  = 0;
 686         cur->oe_mppr    = max_pages;
 687         if (olck->ols_dlmlock != NULL) {
 688                 LASSERT(olck->ols_hold);
 689                 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
 690                 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
 691         }
 692
 693         /* grants has been allocated by caller */
 694         LASSERTF(*grants >= chunksize + cli->cl_grant_extent_tax,
 695                  "%u/%u/%u.\n", *grants, chunksize, cli->cl_grant_extent_tax);
 696         LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
 697                  EXTPARA(cur));
 698
 699 restart:
 700         osc_object_lock(obj);
 701         ext = osc_extent_search(obj, cur->oe_start);
 702         if (ext == NULL)
 703                 ext = first_extent(obj);
 704         while (ext != NULL) {
 705                 pgoff_t ext_chk_start = ext->oe_start >> ppc_bits;
 706                 pgoff_t ext_chk_end   = ext->oe_end   >> ppc_bits;
 707
 708                 LASSERT(sanity_check_nolock(ext) == 0);
 709                 if (chunk > ext_chk_end + 1 || chunk < ext_chk_start)
 710                         break;
 711
 712                 /* if covering by different locks, no chance to match */
 713                 if (olck->ols_dlmlock != ext->oe_dlmlock) {
 714                         EASSERTF(!overlapped(ext, cur), ext,
 715                                  EXTSTR"\n", EXTPARA(cur));
 716
 717                         ext = next_extent(ext);
 718                         continue;
 719                 }
 720
 721                 /* discontiguous chunks? */
 722                 if (chunk + 1 < ext_chk_start) {
 723                         ext = next_extent(ext);
 724                         continue;
 725                 }
 726
 727                 /* ok, from now on, ext and cur have these attrs:
 728                  * 1. covered by the same lock
 729                  * 2. contiguous at chunk level or overlapping. */
 730
 731                 if (overlapped(ext, cur)) {
 732                         /* cur is the minimum unit, so overlapping means
 733                          * full contain. */
 734                         EASSERTF((ext->oe_start <= cur->oe_start &&
 735                                   ext->oe_end >= cur->oe_end),
 736                                  ext, EXTSTR"\n", EXTPARA(cur));
 737
 738                         if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
 739                                 /* for simplicity, we wait for this extent to
 740                                  * finish before going forward. */
 741                                 conflict = osc_extent_get(ext);
 742                                 break;
 743                         }
 744
 745                         found = osc_extent_hold(ext);
 746                         break;
 747                 }
 748
 749                 /* non-overlapped extent */
 750                 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
 751                         /* we can't do anything for a non OES_CACHE extent, or
 752                          * if there is someone waiting for this extent to be
 753                          * flushed, try next one. */
 754                         ext = next_extent(ext);
 755                         continue;
 756                 }
 757
 758                 /* check if they belong to the same rpc slot before trying to
 759                  * merge. the extents are not overlapped and contiguous at
 760                  * chunk level to get here. */
 761                 if (ext->oe_max_end != max_end) {
 762                         /* if they don't belong to the same RPC slot or
 763                          * max_pages_per_rpc has ever changed, do not merge. */
 764                         ext = next_extent(ext);
 765                         continue;
 766                 }
 767
 768                 /* check whether maximum extent size will be hit */
 769                 if ((ext_chk_end - ext_chk_start + 1 + 1) << ppc_bits >
 770                     cli->cl_max_extent_pages) {
 771                         ext = next_extent(ext);
 772                         continue;
 773                 }
 774
 775                 /* it's required that an extent must be contiguous at chunk
 776                  * level so that we know the whole extent is covered by grant
 777                  * (the pages in the extent are NOT required to be contiguous).
 778                  * Otherwise, it will be too much difficult to know which
 779                  * chunks have grants allocated. */
 780
 781                 /* try to do front merge - extend ext's start */
 782                 if (chunk + 1 == ext_chk_start) {
 783                         /* ext must be chunk size aligned */
 784                         EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
 785
 786                         /* pull ext's start back to cover cur */
 787                         ext->oe_start   = cur->oe_start;
 788                         ext->oe_grants += chunksize;
 789                         LASSERT(*grants >= chunksize);
 790                         *grants -= chunksize;
 791
 792                         found = osc_extent_hold(ext);
 793                 } else if (chunk == ext_chk_end + 1) {
 794                         /* rear merge */
 795                         ext->oe_end     = cur->oe_end;
 796                         ext->oe_grants += chunksize;
 797                         LASSERT(*grants >= chunksize);
 798                         *grants -= chunksize;
 799
 800                         /* try to merge with the next one because we just fill
 801                          * in a gap */
 802                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 803                                 /* we can save extent tax from next extent */
 804                                 *grants += cli->cl_grant_extent_tax;
 805
 806                         found = osc_extent_hold(ext);
 807                 }
 808                 if (found != NULL)
 809                         break;
 810
 811                 ext = next_extent(ext);
 812         }
 813
 814         osc_extent_tree_dump(D_CACHE, obj);
 815         if (found != NULL) {
 816                 LASSERT(conflict == NULL);
 817                 if (!IS_ERR(found)) {
 818                         LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
 819                         OSC_EXTENT_DUMP(D_CACHE, found,
 820                                         "found caching ext for %lu.\n", index);
 821                 }
 822         } else if (conflict == NULL) {
 823                 /* create a new extent */
 824                 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
 825                 cur->oe_grants = chunksize + cli->cl_grant_extent_tax;
 826                 LASSERT(*grants >= cur->oe_grants);
 827                 *grants -= cur->oe_grants;
 828
 829                 cur->oe_state = OES_CACHE;
 830                 found = osc_extent_hold(cur);
 831                 osc_extent_insert(obj, cur);
 832                 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
 833                                 index, descr->cld_end);
 834         }
 835         osc_object_unlock(obj);
 836
 837         if (conflict != NULL) {
 838                 LASSERT(found == NULL);
 839
 840                 /* waiting for IO to finish. Please notice that it's impossible
 841                  * to be an OES_TRUNC extent. */
 842                 rc = osc_extent_wait(env, conflict, OES_INV);
 843                 osc_extent_put(env, conflict);
 844                 conflict = NULL;
 845                 if (rc < 0)
 846                         GOTO(out, found = ERR_PTR(rc));
 847
 848                 goto restart;
 849         }
 850         EXIT;
 851
 852 out:
 853         osc_extent_put(env, cur);
 854         return found;
 855 }
 856
 857 /**
 858  * Called when IO is finished to an extent.
 859  */
 860 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
 861                       int sent, int rc)
 862 {
 863         struct client_obd *cli = osc_cli(ext->oe_obj);
 864         struct osc_async_page *oap;
 865         struct osc_async_page *tmp;
 866         int nr_pages = ext->oe_nr_pages;
 867         int lost_grant = 0;
 868         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
 869         loff_t last_off = 0;
 870         int last_count = -1;
 871         ENTRY;
 872
 873         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
 874
 875         ext->oe_rc = rc ?: ext->oe_nr_pages;
 876         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
 877
 878         osc_lru_add_batch(cli, &ext->oe_pages);
 879         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 880                                      oap_pending_item) {
 881                 list_del_init(&oap->oap_rpc_item);
 882                 list_del_init(&oap->oap_pending_item);
 883                 if (last_off <= oap->oap_obj_off) {
 884                         last_off = oap->oap_obj_off;
 885                         last_count = oap->oap_count;
 886                 }
 887
 888                 --ext->oe_nr_pages;
 889                 osc_ap_completion(env, cli, oap, sent, rc);
 890         }
 891         EASSERT(ext->oe_nr_pages == 0, ext);
 892
 893         if (!sent) {
 894                 lost_grant = ext->oe_grants;
 895         } else if (blocksize < PAGE_SIZE &&
 896                    last_count != PAGE_SIZE) {
 897                 /* For short writes we shouldn't count parts of pages that
 898                  * span a whole chunk on the OST side, or our accounting goes
 899                  * wrong.  Should match the code in filter_grant_check. */
 900                 int offset = last_off & ~PAGE_MASK;
 901                 int count = last_count + (offset & (blocksize - 1));
 902                 int end = (offset + last_count) & (blocksize - 1);
 903                 if (end)
 904                         count += blocksize - end;
 905
 906                 lost_grant = PAGE_SIZE - count;
 907         }
 908         if (ext->oe_grants > 0)
 909                 osc_free_grant(cli, nr_pages, lost_grant, ext->oe_grants);
 910
 911         osc_extent_remove(ext);
 912         /* put the refcount for RPC */
 913         osc_extent_put(env, ext);
 914         RETURN(0);
 915 }
 916
 917 static int extent_wait_cb(struct osc_extent *ext, enum osc_extent_state state)
 918 {
 919         int ret;
 920
 921         osc_object_lock(ext->oe_obj);
 922         ret = ext->oe_state == state;
 923         osc_object_unlock(ext->oe_obj);
 924
 925         return ret;
 926 }
 927
 928 /**
 929  * Wait for the extent's state to become @state.
 930  */
 931 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
 932                            enum osc_extent_state state)
 933 {
 934         struct osc_object *obj = ext->oe_obj;
 935         struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
 936                                                   LWI_ON_SIGNAL_NOOP, NULL);
 937         int rc = 0;
 938         ENTRY;
 939
 940         osc_object_lock(obj);
 941         LASSERT(sanity_check_nolock(ext) == 0);
 942         /* `Kick' this extent only if the caller is waiting for it to be
 943          * written out. */
 944         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
 945                 if (ext->oe_state == OES_ACTIVE) {
 946                         ext->oe_urgent = 1;
 947                 } else if (ext->oe_state == OES_CACHE) {
 948                         ext->oe_urgent = 1;
 949                         osc_extent_hold(ext);
 950                         rc = 1;
 951                 }
 952         }
 953         osc_object_unlock(obj);
 954         if (rc == 1)
 955                 osc_extent_release(env, ext);
 956
 957         /* wait for the extent until its state becomes @state */
 958         rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
 959         if (rc == -ETIMEDOUT) {
 960                 OSC_EXTENT_DUMP(D_ERROR, ext,
 961                         "%s: wait ext to %u timedout, recovery in progress?\n",
 962                         cli_name(osc_cli(obj)), state);
 963
 964                 wait_event_idle(ext->oe_waitq, extent_wait_cb(ext, state));
 965                 rc = 0;
 966         }
 967         if (rc == 0 && ext->oe_rc < 0)
 968                 rc = ext->oe_rc;
 969         RETURN(rc);
 970 }
 971
 972 /**
 973  * Discard pages with index greater than @size. If @ext is overlapped with
 974  * @size, then partial truncate happens.
 975  */
 976 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
 977                                 bool partial)
 978 {
 979         struct lu_env         *env;
 980         struct cl_io          *io;
 981         struct osc_object     *obj = ext->oe_obj;
 982         struct client_obd     *cli = osc_cli(obj);
 983         struct osc_async_page *oap;
 984         struct osc_async_page *tmp;
 985         struct pagevec        *pvec;
 986         int                    pages_in_chunk = 0;
 987         int                    ppc_bits    = cli->cl_chunkbits -
 988                                              PAGE_SHIFT;
 989         __u64                  trunc_chunk = trunc_index >> ppc_bits;
 990         int                    grants   = 0;
 991         int                    nr_pages = 0;
 992         int                    rc       = 0;
 993         __u16                  refcheck;
 994         ENTRY;
 995
 996         LASSERT(sanity_check(ext) == 0);
 997         LASSERT(ext->oe_state == OES_TRUNC);
 998         LASSERT(!ext->oe_urgent);
 999
1000         /* Request new lu_env.
1001          * We can't use that env from osc_cache_truncate_start() because
1002          * it's from lov_io_sub and not fully initialized. */
1003         env = cl_env_get(&refcheck);
1004         if (IS_ERR(env))
1005                 RETURN(PTR_ERR(env));
1006
1007         io  = osc_env_thread_io(env);
1008         io->ci_obj = cl_object_top(osc2cl(obj));
1009         io->ci_ignore_layout = 1;
1010         pvec = &osc_env_info(env)->oti_pagevec;
1011         ll_pagevec_init(pvec, 0);
1012         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
1013         if (rc < 0)
1014                 GOTO(out, rc);
1015
1016         /* discard all pages with index greater than trunc_index */
1017         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
1018                                      oap_pending_item) {
1019                 pgoff_t index = osc_index(oap2osc(oap));
1020                 struct cl_page  *page = oap2cl_page(oap);
1021
1022                 LASSERT(list_empty(&oap->oap_rpc_item));
1023
1024                 /* only discard the pages with their index greater than
1025                  * trunc_index, and ... */
1026                 if (index < trunc_index ||
1027                     (index == trunc_index && partial)) {
1028                         /* accounting how many pages remaining in the chunk
1029                          * so that we can calculate grants correctly. */
1030                         if (index >> ppc_bits == trunc_chunk)
1031                                 ++pages_in_chunk;
1032                         continue;
1033                 }
1034
1035                 list_del_init(&oap->oap_pending_item);
1036
1037                 cl_page_get(page);
1038                 lu_ref_add(&page->cp_reference, "truncate", current);
1039
1040                 if (cl_page_own(env, io, page) == 0) {
1041                         cl_page_discard(env, io, page);
1042                         cl_page_disown(env, io, page);
1043                 } else {
1044                         LASSERT(page->cp_state == CPS_FREEING);
1045                         LASSERT(0);
1046                 }
1047
1048                 lu_ref_del(&page->cp_reference, "truncate", current);
1049                 cl_pagevec_put(env, page, pvec);
1050
1051                 --ext->oe_nr_pages;
1052                 ++nr_pages;
1053         }
1054         pagevec_release(pvec);
1055
1056         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1057                       ext->oe_nr_pages == 0),
1058                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1059
1060         osc_object_lock(obj);
1061         if (ext->oe_nr_pages == 0) {
1062                 LASSERT(pages_in_chunk == 0);
1063                 grants = ext->oe_grants;
1064                 ext->oe_grants = 0;
1065         } else { /* calculate how many grants we can free */
1066                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1067                 pgoff_t last_index;
1068
1069
1070                 /* if there is no pages in this chunk, we can also free grants
1071                  * for the last chunk */
1072                 if (pages_in_chunk == 0) {
1073                         /* if this is the 1st chunk and no pages in this chunk,
1074                          * ext->oe_nr_pages must be zero, so we should be in
1075                          * the other if-clause. */
1076                         LASSERT(trunc_chunk > 0);
1077                         --trunc_chunk;
1078                         ++chunks;
1079                 }
1080
1081                 /* this is what we can free from this extent */
1082                 grants          = chunks << cli->cl_chunkbits;
1083                 ext->oe_grants -= grants;
1084                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1085                 ext->oe_end     = min(last_index, ext->oe_max_end);
1086                 LASSERT(ext->oe_end >= ext->oe_start);
1087                 LASSERT(ext->oe_grants > 0);
1088         }
1089         osc_object_unlock(obj);
1090
1091         if (grants > 0 || nr_pages > 0)
1092                 osc_free_grant(cli, nr_pages, grants, grants);
1093
1094 out:
1095         cl_io_fini(env, io);
1096         cl_env_put(env, &refcheck);
1097         RETURN(rc);
1098 }
1099
1100 /**
1101  * This function is used to make the extent prepared for transfer.
1102  * A race with flusing page - ll_writepage() has to be handled cautiously.
1103  */
1104 static int osc_extent_make_ready(const struct lu_env *env,
1105                                  struct osc_extent *ext)
1106 {
1107         struct osc_async_page *oap;
1108         struct osc_async_page *last = NULL;
1109         struct osc_object *obj = ext->oe_obj;
1110         unsigned int page_count = 0;
1111         int rc;
1112         ENTRY;
1113
1114         /* we're going to grab page lock, so object lock must not be taken. */
1115         LASSERT(sanity_check(ext) == 0);
1116         /* in locking state, any process should not touch this extent. */
1117         EASSERT(ext->oe_state == OES_LOCKING, ext);
1118         EASSERT(ext->oe_owner != NULL, ext);
1119
1120         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1121
1122         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1123                 ++page_count;
1124                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1125                         last = oap;
1126
1127                 /* checking ASYNC_READY is race safe */
1128                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1129                         continue;
1130
1131                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1132                 switch (rc) {
1133                 case 0:
1134                         spin_lock(&oap->oap_lock);
1135                         oap->oap_async_flags |= ASYNC_READY;
1136                         spin_unlock(&oap->oap_lock);
1137                         break;
1138                 case -EALREADY:
1139                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1140                         break;
1141                 default:
1142                         LASSERTF(0, "unknown return code: %d\n", rc);
1143                 }
1144         }
1145
1146         LASSERT(page_count == ext->oe_nr_pages);
1147         LASSERT(last != NULL);
1148         /* the last page is the only one we need to refresh its count by
1149          * the size of file. */
1150         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1151                 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1152                 LASSERT(last_oap_count > 0);
1153                 LASSERT(last->oap_page_off + last_oap_count <= PAGE_SIZE);
1154                 last->oap_count = last_oap_count;
1155                 spin_lock(&last->oap_lock);
1156                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1157                 spin_unlock(&last->oap_lock);
1158         }
1159
1160         /* for the rest of pages, we don't need to call osf_refresh_count()
1161          * because it's known they are not the last page */
1162         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1163                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1164                         oap->oap_count = PAGE_SIZE - oap->oap_page_off;
1165                         spin_lock(&oap->oap_lock);
1166                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1167                         spin_unlock(&oap->oap_lock);
1168                 }
1169         }
1170
1171         osc_object_lock(obj);
1172         osc_extent_state_set(ext, OES_RPC);
1173         osc_object_unlock(obj);
1174         /* get a refcount for RPC. */
1175         osc_extent_get(ext);
1176
1177         RETURN(0);
1178 }
1179
1180 /**
1181  * Quick and simple version of osc_extent_find(). This function is frequently
1182  * called to expand the extent for the same IO. To expand the extent, the
1183  * page index must be in the same or next chunk of ext->oe_end.
1184  */
1185 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index,
1186                              unsigned int *grants)
1187 {
1188         struct osc_object *obj = ext->oe_obj;
1189         struct client_obd *cli = osc_cli(obj);
1190         struct osc_extent *next;
1191         int ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1192         pgoff_t chunk = index >> ppc_bits;
1193         pgoff_t end_chunk;
1194         pgoff_t end_index;
1195         unsigned int chunksize = 1 << cli->cl_chunkbits;
1196         int rc = 0;
1197         ENTRY;
1198
1199         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1200         osc_object_lock(obj);
1201         LASSERT(sanity_check_nolock(ext) == 0);
1202         end_chunk = ext->oe_end >> ppc_bits;
1203         if (chunk > end_chunk + 1)
1204                 GOTO(out, rc = -ERANGE);
1205
1206         if (end_chunk >= chunk)
1207                 GOTO(out, rc = 0);
1208
1209         LASSERT(end_chunk + 1 == chunk);
1210
1211         /* try to expand this extent to cover @index */
1212         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1213
1214         /* don't go over the maximum extent size reported by server */
1215         if (end_index - ext->oe_start + 1 > cli->cl_max_extent_pages)
1216                 GOTO(out, rc = -ERANGE);
1217
1218         next = next_extent(ext);
1219         if (next != NULL && next->oe_start <= end_index)
1220                 /* complex mode - overlapped with the next extent,
1221                  * this case will be handled by osc_extent_find() */
1222                 GOTO(out, rc = -EAGAIN);
1223
1224         ext->oe_end = end_index;
1225         ext->oe_grants += chunksize;
1226         LASSERT(*grants >= chunksize);
1227         *grants -= chunksize;
1228         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1229                  "overlapped after expanding for %lu.\n", index);
1230         EXIT;
1231
1232 out:
1233         osc_object_unlock(obj);
1234         RETURN(rc);
1235 }
1236
1237 static void osc_extent_tree_dump0(int mask, struct osc_object *obj,
1238                                   const char *func, int line)
1239 {
1240         struct osc_extent *ext;
1241         int cnt;
1242
1243         if (!cfs_cdebug_show(mask, DEBUG_SUBSYSTEM))
1244                 return;
1245
1246         CDEBUG(mask, "Dump object %p extents at %s:%d, mppr: %u.\n",
1247                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1248
1249         /* osc_object_lock(obj); */
1250         cnt = 1;
1251         for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1252                 OSC_EXTENT_DUMP(mask, ext, "in tree %d.\n", cnt++);
1253
1254         cnt = 1;
1255         list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1256                 OSC_EXTENT_DUMP(mask, ext, "hp %d.\n", cnt++);
1257
1258         cnt = 1;
1259         list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1260                 OSC_EXTENT_DUMP(mask, ext, "urgent %d.\n", cnt++);
1261
1262         cnt = 1;
1263         list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1264                 OSC_EXTENT_DUMP(mask, ext, "reading %d.\n", cnt++);
1265         /* osc_object_unlock(obj); */
1266 }
1267
1268 /* ------------------ osc extent end ------------------ */
1269
1270 static inline int osc_is_ready(struct osc_object *osc)
1271 {
1272         return !list_empty(&osc->oo_ready_item) ||
1273                !list_empty(&osc->oo_hp_ready_item);
1274 }
1275
1276 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1277         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1278                (OSC), osc_is_ready(OSC),                                       \
1279                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1280                list_empty_marker(&(OSC)->oo_ready_item),                       \
1281                atomic_read(&(OSC)->oo_nr_writes),                              \
1282                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1283                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1284                atomic_read(&(OSC)->oo_nr_reads),                               \
1285                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1286                ##args)
1287
1288 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1289                           int cmd)
1290 {
1291         struct osc_page *opg  = oap2osc_page(oap);
1292         struct cl_page  *page = oap2cl_page(oap);
1293         int result;
1294
1295         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1296
1297         ENTRY;
1298         result = cl_page_make_ready(env, page, CRT_WRITE);
1299         if (result == 0)
1300                 opg->ops_submit_time = ktime_get();
1301         RETURN(result);
1302 }
1303
1304 static int osc_refresh_count(const struct lu_env *env,
1305                              struct osc_async_page *oap, int cmd)
1306 {
1307         struct osc_page  *opg = oap2osc_page(oap);
1308         pgoff_t index = osc_index(oap2osc(oap));
1309         struct cl_object *obj;
1310         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
1311         int result;
1312         loff_t kms;
1313
1314         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1315         LASSERT(!(cmd & OBD_BRW_READ));
1316         LASSERT(opg != NULL);
1317         obj = opg->ops_cl.cpl_obj;
1318
1319         cl_object_attr_lock(obj);
1320         result = cl_object_attr_get(env, obj, attr);
1321         cl_object_attr_unlock(obj);
1322         if (result < 0)
1323                 return result;
1324         kms = attr->cat_kms;
1325         if (cl_offset(obj, index) >= kms)
1326                 /* catch race with truncate */
1327                 return 0;
1328         else if (cl_offset(obj, index + 1) > kms)
1329                 /* catch sub-page write at end of file */
1330                 return kms % PAGE_SIZE;
1331         else
1332                 return PAGE_SIZE;
1333 }
1334
1335 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1336                           int cmd, int rc)
1337 {
1338         struct osc_page   *opg  = oap2osc_page(oap);
1339         struct cl_page    *page = oap2cl_page(oap);
1340         enum cl_req_type   crt;
1341         int srvlock;
1342
1343         ENTRY;
1344
1345         cmd &= ~OBD_BRW_NOQUOTA;
1346         LASSERTF(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ),
1347                  "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1348         LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1349                 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1350         LASSERT(opg->ops_transfer_pinned);
1351
1352         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1353         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1354         opg->ops_transfer_pinned = 0;
1355
1356         opg->ops_submit_time = ktime_set(0, 0);
1357         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1358
1359         /* statistic */
1360         if (rc == 0 && srvlock) {
1361                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1362                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1363                 size_t bytes = oap->oap_count;
1364
1365                 if (crt == CRT_READ)
1366                         stats->os_lockless_reads += bytes;
1367                 else
1368                         stats->os_lockless_writes += bytes;
1369         }
1370
1371         /*
1372          * This has to be the last operation with the page, as locks are
1373          * released in cl_page_completion() and nothing except for the
1374          * reference counter protects page from concurrent reclaim.
1375          */
1376         lu_ref_del(&page->cp_reference, "transfer", page);
1377
1378         cl_page_completion(env, page, crt, rc);
1379         cl_page_put(env, page);
1380
1381         RETURN(0);
1382 }
1383
1384 #define OSC_DUMP_GRANT(mask, cli, fmt, args...) do {                    \
1385         struct client_obd *__tmp = (cli);                               \
1386         CDEBUG(mask, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu " \
1387                "dropped: %ld avail: %ld, dirty_grant: %ld, "            \
1388                "reserved: %ld, flight: %d } lru {in list: %ld, "        \
1389                "left: %ld, waiters: %d }" fmt "\n",                     \
1390                cli_name(__tmp),                                         \
1391                __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages,        \
1392                atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1393                __tmp->cl_lost_grant, __tmp->cl_avail_grant,             \
1394                __tmp->cl_dirty_grant,                                   \
1395                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,         \
1396                atomic_long_read(&__tmp->cl_lru_in_list),                \
1397                atomic_long_read(&__tmp->cl_lru_busy),                   \
1398                atomic_read(&__tmp->cl_lru_shrinkers), ##args);          \
1399 } while (0)
1400
1401 /* caller must hold loi_list_lock */
1402 static void osc_consume_write_grant(struct client_obd *cli,
1403                                     struct brw_page *pga)
1404 {
1405         assert_spin_locked(&cli->cl_loi_list_lock);
1406         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1407         cli->cl_dirty_pages++;
1408         pga->flag |= OBD_BRW_FROM_GRANT;
1409         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1410                PAGE_SIZE, pga, pga->pg);
1411         osc_update_next_shrink(cli);
1412 }
1413
1414 /* the companion to osc_consume_write_grant, called when a brw has completed.
1415  * must be called with the loi lock held. */
1416 static void osc_release_write_grant(struct client_obd *cli,
1417                                     struct brw_page *pga)
1418 {
1419         ENTRY;
1420
1421         assert_spin_locked(&cli->cl_loi_list_lock);
1422         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1423                 EXIT;
1424                 return;
1425         }
1426
1427         pga->flag &= ~OBD_BRW_FROM_GRANT;
1428         atomic_long_dec(&obd_dirty_pages);
1429         cli->cl_dirty_pages--;
1430         EXIT;
1431 }
1432
1433 /**
1434  * To avoid sleeping with object lock held, it's good for us allocate enough
1435  * grants before entering into critical section.
1436  *
1437  * client_obd_list_lock held by caller
1438  */
1439 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1440 {
1441         int rc = -EDQUOT;
1442
1443         if (cli->cl_avail_grant >= bytes) {
1444                 cli->cl_avail_grant    -= bytes;
1445                 cli->cl_reserved_grant += bytes;
1446                 rc = 0;
1447         }
1448         return rc;
1449 }
1450
1451 static void __osc_unreserve_grant(struct client_obd *cli,
1452                                   unsigned int reserved, unsigned int unused)
1453 {
1454         /* it's quite normal for us to get more grant than reserved.
1455          * Thinking about a case that two extents merged by adding a new
1456          * chunk, we can save one extent tax. If extent tax is greater than
1457          * one chunk, we can save more grant by adding a new chunk */
1458         cli->cl_reserved_grant -= reserved;
1459         if (unused > reserved) {
1460                 cli->cl_avail_grant += reserved;
1461                 cli->cl_lost_grant  += unused - reserved;
1462                 cli->cl_dirty_grant -= unused - reserved;
1463         } else {
1464                 cli->cl_avail_grant += unused;
1465                 cli->cl_dirty_grant += reserved - unused;
1466         }
1467 }
1468
1469 static void osc_unreserve_grant_nolock(struct client_obd *cli,
1470                                        unsigned int reserved,
1471                                        unsigned int unused)
1472 {
1473         __osc_unreserve_grant(cli, reserved, unused);
1474         if (unused > 0)
1475                 osc_wake_cache_waiters(cli);
1476 }
1477
1478 static void osc_unreserve_grant(struct client_obd *cli,
1479                                 unsigned int reserved, unsigned int unused)
1480 {
1481         spin_lock(&cli->cl_loi_list_lock);
1482         osc_unreserve_grant_nolock(cli, reserved, unused);
1483         spin_unlock(&cli->cl_loi_list_lock);
1484 }
1485
1486 /**
1487  * Free grant after IO is finished or canceled.
1488  *
1489  * @lost_grant is used to remember how many grants we have allocated but not
1490  * used, we should return these grants to OST. There're two cases where grants
1491  * can be lost:
1492  * 1. truncate;
1493  * 2. blocksize at OST is less than PAGE_SIZE and a partial page was
1494  *    written. In this case OST may use less chunks to serve this partial
1495  *    write. OSTs don't actually know the page size on the client side. so
1496  *    clients have to calculate lost grant by the blocksize on the OST.
1497  *    See filter_grant_check() for details.
1498  */
1499 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1500                            unsigned int lost_grant, unsigned int dirty_grant)
1501 {
1502         unsigned long grant;
1503
1504         grant = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
1505
1506         spin_lock(&cli->cl_loi_list_lock);
1507         atomic_long_sub(nr_pages, &obd_dirty_pages);
1508         cli->cl_dirty_pages -= nr_pages;
1509         cli->cl_lost_grant += lost_grant;
1510         cli->cl_dirty_grant -= dirty_grant;
1511         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1512                 /* borrow some grant from truncate to avoid the case that
1513                  * truncate uses up all avail grant */
1514                 cli->cl_lost_grant -= grant;
1515                 cli->cl_avail_grant += grant;
1516         }
1517         osc_wake_cache_waiters(cli);
1518         spin_unlock(&cli->cl_loi_list_lock);
1519         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu/%lu\n",
1520                lost_grant, cli->cl_lost_grant,
1521                cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_SHIFT,
1522                cli->cl_dirty_grant);
1523 }
1524
1525 /**
1526  * The companion to osc_enter_cache(), called when @oap is no longer part of
1527  * the dirty accounting due to error.
1528  */
1529 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1530 {
1531         spin_lock(&cli->cl_loi_list_lock);
1532         osc_release_write_grant(cli, &oap->oap_brw_page);
1533         spin_unlock(&cli->cl_loi_list_lock);
1534 }
1535
1536 /**
1537  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1538  * is available.
1539  */
1540 static int osc_enter_cache_try(struct client_obd *cli,
1541                                struct osc_async_page *oap,
1542                                int bytes)
1543 {
1544         int rc;
1545
1546         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1547
1548         rc = osc_reserve_grant(cli, bytes);
1549         if (rc < 0)
1550                 return 0;
1551
1552         if (cli->cl_dirty_pages < cli->cl_dirty_max_pages) {
1553                 if (atomic_long_add_return(1, &obd_dirty_pages) <=
1554                     obd_max_dirty_pages) {
1555                         osc_consume_write_grant(cli, &oap->oap_brw_page);
1556                         rc = 1;
1557                         goto out;
1558                 } else
1559                         atomic_long_dec(&obd_dirty_pages);
1560         }
1561         __osc_unreserve_grant(cli, bytes, bytes);
1562
1563 out:
1564         return rc;
1565 }
1566
1567 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1568 {
1569         int rc;
1570         spin_lock(&cli->cl_loi_list_lock);
1571         rc = list_empty(&ocw->ocw_entry);
1572         spin_unlock(&cli->cl_loi_list_lock);
1573         return rc;
1574 }
1575
1576 /**
1577  * The main entry to reserve dirty page accounting. Usually the grant reserved
1578  * in this function will be freed in bulk in osc_free_grant() unless it fails
1579  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1580  *
1581  * The process will be put into sleep if it's already run out of grant.
1582  */
1583 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1584                            struct osc_async_page *oap, int bytes)
1585 {
1586         struct osc_object       *osc = oap->oap_obj;
1587         struct lov_oinfo        *loi = osc->oo_oinfo;
1588         struct osc_cache_waiter  ocw;
1589         struct l_wait_info       lwi;
1590         int                      rc = -EDQUOT;
1591         ENTRY;
1592
1593         lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(AT_OFF ? obd_timeout : at_max),
1594                                NULL, LWI_ON_SIGNAL_NOOP, NULL);
1595
1596         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1597
1598         spin_lock(&cli->cl_loi_list_lock);
1599
1600         /* force the caller to try sync io.  this can jump the list
1601          * of queued writes and create a discontiguous rpc stream */
1602         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1603             cli->cl_dirty_max_pages == 0 ||
1604             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) {
1605                 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1606                 GOTO(out, rc = -EDQUOT);
1607         }
1608
1609         /* Hopefully normal case - cache space and write credits available */
1610         if (list_empty(&cli->cl_cache_waiters) &&
1611             osc_enter_cache_try(cli, oap, bytes)) {
1612                 OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1613                 GOTO(out, rc = 0);
1614         }
1615
1616         /* We can get here for two reasons: too many dirty pages in cache, or
1617          * run out of grants. In both cases we should write dirty pages out.
1618          * Adding a cache waiter will trigger urgent write-out no matter what
1619          * RPC size will be.
1620          * The exiting condition is no avail grants and no dirty pages caching,
1621          * that really means there is no space on the OST. */
1622         init_waitqueue_head(&ocw.ocw_waitq);
1623         ocw.ocw_oap   = oap;
1624         ocw.ocw_grant = bytes;
1625         while (cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0) {
1626                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1627                 ocw.ocw_rc = 0;
1628                 spin_unlock(&cli->cl_loi_list_lock);
1629
1630                 osc_io_unplug_async(env, cli, NULL);
1631
1632                 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1633                        cli_name(cli), &ocw, oap);
1634
1635                 rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1636
1637                 spin_lock(&cli->cl_loi_list_lock);
1638
1639                 if (rc < 0) {
1640                         /* l_wait_event is interrupted by signal or timed out */
1641                         list_del_init(&ocw.ocw_entry);
1642                         break;
1643                 }
1644                 LASSERT(list_empty(&ocw.ocw_entry));
1645                 rc = ocw.ocw_rc;
1646
1647                 if (rc != -EDQUOT)
1648                         break;
1649                 if (osc_enter_cache_try(cli, oap, bytes)) {
1650                         rc = 0;
1651                         break;
1652                 }
1653         }
1654
1655         switch (rc) {
1656         case 0:
1657                 OSC_DUMP_GRANT(D_CACHE, cli, "finally got grant space\n");
1658                 break;
1659         case -ETIMEDOUT:
1660                 OSC_DUMP_GRANT(D_CACHE, cli,
1661                                "timeout, fall back to sync i/o\n");
1662                 osc_extent_tree_dump(D_CACHE, osc);
1663                 /* fall back to synchronous I/O */
1664                 rc = -EDQUOT;
1665                 break;
1666         case -EINTR:
1667                 /* Ensures restartability - LU-3581 */
1668                 OSC_DUMP_GRANT(D_CACHE, cli, "interrupted\n");
1669                 rc = -ERESTARTSYS;
1670                 break;
1671         case -EDQUOT:
1672                 OSC_DUMP_GRANT(D_CACHE, cli,
1673                                "no grant space, fall back to sync i/o\n");
1674                 break;
1675         default:
1676                 CDEBUG(D_CACHE, "%s: event for cache space @ %p never arrived "
1677                        "due to %d, fall back to sync i/o\n",
1678                        cli_name(cli), &ocw, rc);
1679                 break;
1680         }
1681         EXIT;
1682 out:
1683         spin_unlock(&cli->cl_loi_list_lock);
1684         RETURN(rc);
1685 }
1686
1687 /* caller must hold loi_list_lock */
1688 void osc_wake_cache_waiters(struct client_obd *cli)
1689 {
1690         struct list_head *l, *tmp;
1691         struct osc_cache_waiter *ocw;
1692
1693         ENTRY;
1694         list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1695                 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
1696
1697                 ocw->ocw_rc = -EDQUOT;
1698
1699                 if (osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant))
1700                         ocw->ocw_rc = 0;
1701
1702                 if (ocw->ocw_rc == 0 ||
1703                     !(cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0)) {
1704                         list_del_init(&ocw->ocw_entry);
1705                         CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant "
1706                                "%ld, %d\n", ocw, ocw->ocw_oap,
1707                                cli->cl_avail_grant, ocw->ocw_rc);
1708
1709                         wake_up(&ocw->ocw_waitq);
1710                 }
1711         }
1712
1713         EXIT;
1714 }
1715 EXPORT_SYMBOL(osc_wake_cache_waiters);
1716
1717 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1718 {
1719         int hprpc = !!list_empty(&osc->oo_hp_exts);
1720         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1721 }
1722
1723 /* This maintains the lists of pending pages to read/write for a given object
1724  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1725  * to quickly find objects that are ready to send an RPC. */
1726 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1727                          int cmd)
1728 {
1729         int invalid_import = 0;
1730         ENTRY;
1731
1732         /* if we have an invalid import we want to drain the queued pages
1733          * by forcing them through rpcs that immediately fail and complete
1734          * the pages.  recovery relies on this to empty the queued pages
1735          * before canceling the locks and evicting down the llite pages */
1736         if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1737                 invalid_import = 1;
1738
1739         if (cmd & OBD_BRW_WRITE) {
1740                 if (atomic_read(&osc->oo_nr_writes) == 0)
1741                         RETURN(0);
1742                 if (invalid_import) {
1743                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1744                         RETURN(1);
1745                 }
1746                 if (!list_empty(&osc->oo_hp_exts)) {
1747                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1748                         RETURN(1);
1749                 }
1750                 if (!list_empty(&osc->oo_urgent_exts)) {
1751                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1752                         RETURN(1);
1753                 }
1754                 /* trigger a write rpc stream as long as there are dirtiers
1755                  * waiting for space.  as they're waiting, they're not going to
1756                  * create more pages to coalesce with what's waiting.. */
1757                 if (!list_empty(&cli->cl_cache_waiters)) {
1758                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1759                         RETURN(1);
1760                 }
1761                 if (!list_empty(&osc->oo_full_exts)) {
1762                         CDEBUG(D_CACHE, "full extent ready, make an RPC\n");
1763                         RETURN(1);
1764                 }
1765         } else {
1766                 if (atomic_read(&osc->oo_nr_reads) == 0)
1767                         RETURN(0);
1768                 if (invalid_import) {
1769                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1770                         RETURN(1);
1771                 }
1772                 /* all read are urgent. */
1773                 if (!list_empty(&osc->oo_reading_exts))
1774                         RETURN(1);
1775         }
1776
1777         RETURN(0);
1778 }
1779
1780 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1781 {
1782         struct client_obd *cli = osc_cli(obj);
1783         if (cmd & OBD_BRW_WRITE) {
1784                 atomic_add(delta, &obj->oo_nr_writes);
1785                 atomic_add(delta, &cli->cl_pending_w_pages);
1786                 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1787         } else {
1788                 atomic_add(delta, &obj->oo_nr_reads);
1789                 atomic_add(delta, &cli->cl_pending_r_pages);
1790                 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1791         }
1792         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1793 }
1794
1795 static int osc_makes_hprpc(struct osc_object *obj)
1796 {
1797         return !list_empty(&obj->oo_hp_exts);
1798 }
1799
1800 static void on_list(struct list_head *item, struct list_head *list,
1801                     int should_be_on)
1802 {
1803         if (list_empty(item) && should_be_on)
1804                 list_add_tail(item, list);
1805         else if (!list_empty(item) && !should_be_on)
1806                 list_del_init(item);
1807 }
1808
1809 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1810  * can find pages to build into rpcs quickly */
1811 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1812 {
1813         if (osc_makes_hprpc(osc)) {
1814                 /* HP rpc */
1815                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1816                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1817         } else {
1818                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1819                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1820                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1821                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1822         }
1823
1824         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1825                 atomic_read(&osc->oo_nr_writes) > 0);
1826
1827         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1828                 atomic_read(&osc->oo_nr_reads) > 0);
1829
1830         return osc_is_ready(osc);
1831 }
1832
1833 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1834 {
1835         int is_ready;
1836
1837         spin_lock(&cli->cl_loi_list_lock);
1838         is_ready = __osc_list_maint(cli, osc);
1839         spin_unlock(&cli->cl_loi_list_lock);
1840
1841         return is_ready;
1842 }
1843
1844 /* this is trying to propogate async writeback errors back up to the
1845  * application.  As an async write fails we record the error code for later if
1846  * the app does an fsync.  As long as errors persist we force future rpcs to be
1847  * sync so that the app can get a sync error and break the cycle of queueing
1848  * pages for which writeback will fail. */
1849 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1850                            int rc)
1851 {
1852         if (rc) {
1853                 if (!ar->ar_rc)
1854                         ar->ar_rc = rc;
1855
1856                 ar->ar_force_sync = 1;
1857                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1858                 return;
1859
1860         }
1861
1862         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1863                 ar->ar_force_sync = 0;
1864 }
1865
1866 /* this must be called holding the loi list lock to give coverage to exit_cache,
1867  * async_flag maintenance, and oap_request */
1868 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1869                               struct osc_async_page *oap, int sent, int rc)
1870 {
1871         struct osc_object *osc = oap->oap_obj;
1872         struct lov_oinfo  *loi = osc->oo_oinfo;
1873         __u64 xid = 0;
1874
1875         ENTRY;
1876         if (oap->oap_request != NULL) {
1877                 xid = ptlrpc_req_xid(oap->oap_request);
1878                 ptlrpc_req_finished(oap->oap_request);
1879                 oap->oap_request = NULL;
1880         }
1881
1882         /* As the transfer for this page is being done, clear the flags */
1883         spin_lock(&oap->oap_lock);
1884         oap->oap_async_flags = 0;
1885         spin_unlock(&oap->oap_lock);
1886         oap->oap_interrupted = 0;
1887
1888         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1889                 spin_lock(&cli->cl_loi_list_lock);
1890                 osc_process_ar(&cli->cl_ar, xid, rc);
1891                 osc_process_ar(&loi->loi_ar, xid, rc);
1892                 spin_unlock(&cli->cl_loi_list_lock);
1893         }
1894
1895         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1896         if (rc)
1897                 CERROR("completion on oap %p obj %p returns %d.\n",
1898                        oap, osc, rc);
1899
1900         EXIT;
1901 }
1902
1903 struct extent_rpc_data {
1904         struct list_head        *erd_rpc_list;
1905         unsigned int            erd_page_count;
1906         unsigned int            erd_max_pages;
1907         unsigned int            erd_max_chunks;
1908         unsigned int            erd_max_extents;
1909 };
1910
1911 static inline unsigned osc_extent_chunks(const struct osc_extent *ext)
1912 {
1913         struct client_obd *cli = osc_cli(ext->oe_obj);
1914         unsigned ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1915
1916         return (ext->oe_end >> ppc_bits) - (ext->oe_start >> ppc_bits) + 1;
1917 }
1918
1919 /**
1920  * Try to add extent to one RPC. We need to think about the following things:
1921  * - # of pages must not be over max_pages_per_rpc
1922  * - extent must be compatible with previous ones
1923  */
1924 static int try_to_add_extent_for_io(struct client_obd *cli,
1925                                     struct osc_extent *ext,
1926                                     struct extent_rpc_data *data)
1927 {
1928         struct osc_extent *tmp;
1929         unsigned int chunk_count;
1930         struct osc_async_page *oap = list_first_entry(&ext->oe_pages,
1931                                                       struct osc_async_page,
1932                                                       oap_pending_item);
1933         ENTRY;
1934
1935         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1936                 ext);
1937         OSC_EXTENT_DUMP(D_CACHE, ext, "trying to add this extent\n");
1938
1939         if (data->erd_max_extents == 0)
1940                 RETURN(0);
1941
1942         chunk_count = osc_extent_chunks(ext);
1943         EASSERTF(data->erd_page_count != 0 ||
1944                  chunk_count <= data->erd_max_chunks, ext,
1945                  "The first extent to be fit in a RPC contains %u chunks, "
1946                  "which is over the limit %u.\n", chunk_count,
1947                  data->erd_max_chunks);
1948         if (chunk_count > data->erd_max_chunks)
1949                 RETURN(0);
1950
1951         data->erd_max_pages = max(ext->oe_mppr, data->erd_max_pages);
1952         EASSERTF(data->erd_page_count != 0 ||
1953                 ext->oe_nr_pages <= data->erd_max_pages, ext,
1954                 "The first extent to be fit in a RPC contains %u pages, "
1955                 "which is over the limit %u.\n", ext->oe_nr_pages,
1956                 data->erd_max_pages);
1957         if (data->erd_page_count + ext->oe_nr_pages > data->erd_max_pages)
1958                 RETURN(0);
1959
1960         list_for_each_entry(tmp, data->erd_rpc_list, oe_link) {
1961                 struct osc_async_page *oap2;
1962                 oap2 = list_first_entry(&tmp->oe_pages, struct osc_async_page,
1963                                         oap_pending_item);
1964                 EASSERT(tmp->oe_owner == current, tmp);
1965 #if 0
1966                 if (overlapped(tmp, ext)) {
1967                         OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
1968                         EASSERT(0, ext);
1969                 }
1970 #endif
1971                 if (oap2cl_page(oap)->cp_type != oap2cl_page(oap2)->cp_type) {
1972                         CDEBUG(D_CACHE, "Do not permit different types of IO "
1973                                "in one RPC\n");
1974                         RETURN(0);
1975                 }
1976
1977                 if (tmp->oe_srvlock != ext->oe_srvlock ||
1978                     !tmp->oe_grants != !ext->oe_grants ||
1979                     tmp->oe_ndelay != ext->oe_ndelay ||
1980                     tmp->oe_no_merge || ext->oe_no_merge)
1981                         RETURN(0);
1982
1983                 /* remove break for strict check */
1984                 break;
1985         }
1986
1987         data->erd_max_extents--;
1988         data->erd_max_chunks -= chunk_count;
1989         data->erd_page_count += ext->oe_nr_pages;
1990         list_move_tail(&ext->oe_link, data->erd_rpc_list);
1991         ext->oe_owner = current;
1992         RETURN(1);
1993 }
1994
1995 /**
1996  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1997  * get_write_extent() takes all appropriate extents in atomic.
1998  *
1999  * The following policy is used to collect extents for IO:
2000  * 1. Add as many HP extents as possible;
2001  * 2. Add the first urgent extent in urgent extent list and take it out of
2002  *    urgent list;
2003  * 3. Add subsequent extents of this urgent extent;
2004  * 4. If urgent list is not empty, goto 2;
2005  * 5. Traverse the extent tree from the 1st extent;
2006  * 6. Above steps exit if there is no space in this RPC.
2007  */
2008 static unsigned int get_write_extents(struct osc_object *obj,
2009                                       struct list_head *rpclist)
2010 {
2011         struct client_obd *cli = osc_cli(obj);
2012         struct osc_extent *ext;
2013         struct extent_rpc_data data = {
2014                 .erd_rpc_list   = rpclist,
2015                 .erd_page_count = 0,
2016                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
2017                 .erd_max_chunks = osc_max_write_chunks(cli),
2018                 .erd_max_extents = 256,
2019         };
2020
2021         LASSERT(osc_object_is_locked(obj));
2022         while (!list_empty(&obj->oo_hp_exts)) {
2023                 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
2024                                  oe_link);
2025                 LASSERT(ext->oe_state == OES_CACHE);
2026                 if (!try_to_add_extent_for_io(cli, ext, &data))
2027                         return data.erd_page_count;
2028                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2029         }
2030         if (data.erd_page_count == data.erd_max_pages)
2031                 return data.erd_page_count;
2032
2033         while (!list_empty(&obj->oo_urgent_exts)) {
2034                 ext = list_entry(obj->oo_urgent_exts.next,
2035                                  struct osc_extent, oe_link);
2036                 if (!try_to_add_extent_for_io(cli, ext, &data))
2037                         return data.erd_page_count;
2038         }
2039         if (data.erd_page_count == data.erd_max_pages)
2040                 return data.erd_page_count;
2041
2042         /* One key difference between full extents and other extents: full
2043          * extents can usually only be added if the rpclist was empty, so if we
2044          * can't add one, we continue on to trying to add normal extents.  This
2045          * is so we don't miss adding extra extents to an RPC containing high
2046          * priority or urgent extents. */
2047         while (!list_empty(&obj->oo_full_exts)) {
2048                 ext = list_entry(obj->oo_full_exts.next,
2049                                  struct osc_extent, oe_link);
2050                 if (!try_to_add_extent_for_io(cli, ext, &data))
2051                         break;
2052         }
2053         if (data.erd_page_count == data.erd_max_pages)
2054                 return data.erd_page_count;
2055
2056         ext = first_extent(obj);
2057         while (ext != NULL) {
2058                 if ((ext->oe_state != OES_CACHE) ||
2059                     /* this extent may be already in current rpclist */
2060                     (!list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
2061                         ext = next_extent(ext);
2062                         continue;
2063                 }
2064
2065                 if (!try_to_add_extent_for_io(cli, ext, &data))
2066                         return data.erd_page_count;
2067
2068                 ext = next_extent(ext);
2069         }
2070         return data.erd_page_count;
2071 }
2072
2073 static int
2074 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
2075                    struct osc_object *osc)
2076 __must_hold(osc)
2077 {
2078         LIST_HEAD(rpclist);
2079         struct osc_extent *ext;
2080         struct osc_extent *tmp;
2081         struct osc_extent *first = NULL;
2082         unsigned int page_count = 0;
2083         int srvlock = 0;
2084         int rc = 0;
2085         ENTRY;
2086
2087         LASSERT(osc_object_is_locked(osc));
2088
2089         page_count = get_write_extents(osc, &rpclist);
2090         LASSERT(equi(page_count == 0, list_empty(&rpclist)));
2091
2092         if (list_empty(&rpclist))
2093                 RETURN(0);
2094
2095         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2096
2097         list_for_each_entry(ext, &rpclist, oe_link) {
2098                 LASSERT(ext->oe_state == OES_CACHE ||
2099                         ext->oe_state == OES_LOCK_DONE);
2100                 if (ext->oe_state == OES_CACHE)
2101                         osc_extent_state_set(ext, OES_LOCKING);
2102                 else
2103                         osc_extent_state_set(ext, OES_RPC);
2104         }
2105
2106         /* we're going to grab page lock, so release object lock because
2107          * lock order is page lock -> object lock. */
2108         osc_object_unlock(osc);
2109
2110         list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2111                 if (ext->oe_state == OES_LOCKING) {
2112                         rc = osc_extent_make_ready(env, ext);
2113                         if (unlikely(rc < 0)) {
2114                                 list_del_init(&ext->oe_link);
2115                                 osc_extent_finish(env, ext, 0, rc);
2116                                 continue;
2117                         }
2118                 }
2119                 if (first == NULL) {
2120                         first = ext;
2121                         srvlock = ext->oe_srvlock;
2122                 } else {
2123                         LASSERT(srvlock == ext->oe_srvlock);
2124                 }
2125         }
2126
2127         if (!list_empty(&rpclist)) {
2128                 LASSERT(page_count > 0);
2129                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2130                 LASSERT(list_empty(&rpclist));
2131         }
2132
2133         osc_object_lock(osc);
2134         RETURN(rc);
2135 }
2136
2137 /**
2138  * prepare pages for ASYNC io and put pages in send queue.
2139  *
2140  * \param cmd OBD_BRW_* macroses
2141  * \param lop pending pages
2142  *
2143  * \return zero if no page added to send queue.
2144  * \return 1 if pages successfully added to send queue.
2145  * \return negative on errors.
2146  */
2147 static int
2148 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2149                   struct osc_object *osc)
2150 __must_hold(osc)
2151 {
2152         struct osc_extent *ext;
2153         struct osc_extent *next;
2154         LIST_HEAD(rpclist);
2155         struct extent_rpc_data data = {
2156                 .erd_rpc_list   = &rpclist,
2157                 .erd_page_count = 0,
2158                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
2159                 .erd_max_chunks = UINT_MAX,
2160                 .erd_max_extents = UINT_MAX,
2161         };
2162         int rc = 0;
2163         ENTRY;
2164
2165         LASSERT(osc_object_is_locked(osc));
2166         list_for_each_entry_safe(ext, next, &osc->oo_reading_exts, oe_link) {
2167                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2168                 if (!try_to_add_extent_for_io(cli, ext, &data))
2169                         break;
2170                 osc_extent_state_set(ext, OES_RPC);
2171                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2172         }
2173         LASSERT(data.erd_page_count <= data.erd_max_pages);
2174
2175         osc_update_pending(osc, OBD_BRW_READ, -data.erd_page_count);
2176
2177         if (!list_empty(&rpclist)) {
2178                 osc_object_unlock(osc);
2179
2180                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2181                 LASSERT(list_empty(&rpclist));
2182
2183                 osc_object_lock(osc);
2184         }
2185         RETURN(rc);
2186 }
2187
2188 #define list_to_obj(list, item) ({                                            \
2189         struct list_head *__tmp = (list)->next;                               \
2190         list_del_init(__tmp);                                         \
2191         list_entry(__tmp, struct osc_object, oo_##item);                      \
2192 })
2193
2194 /* This is called by osc_check_rpcs() to find which objects have pages that
2195  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2196 static struct osc_object *osc_next_obj(struct client_obd *cli)
2197 {
2198         ENTRY;
2199
2200         /* First return objects that have blocked locks so that they
2201          * will be flushed quickly and other clients can get the lock,
2202          * then objects which have pages ready to be stuffed into RPCs */
2203         if (!list_empty(&cli->cl_loi_hp_ready_list))
2204                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2205         if (!list_empty(&cli->cl_loi_ready_list))
2206                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2207
2208         /* then if we have cache waiters, return all objects with queued
2209          * writes.  This is especially important when many small files
2210          * have filled up the cache and not been fired into rpcs because
2211          * they don't pass the nr_pending/object threshhold */
2212         if (!list_empty(&cli->cl_cache_waiters) &&
2213             !list_empty(&cli->cl_loi_write_list))
2214                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2215
2216         /* then return all queued objects when we have an invalid import
2217          * so that they get flushed */
2218         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2219                 if (!list_empty(&cli->cl_loi_write_list))
2220                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2221                                            write_item));
2222                 if (!list_empty(&cli->cl_loi_read_list))
2223                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2224                                            read_item));
2225         }
2226         RETURN(NULL);
2227 }
2228
2229 /* called with the loi list lock held */
2230 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2231 __must_hold(&cli->cl_loi_list_lock)
2232 {
2233         struct osc_object *osc;
2234         int rc = 0;
2235         ENTRY;
2236
2237         while ((osc = osc_next_obj(cli)) != NULL) {
2238                 struct cl_object *obj = osc2cl(osc);
2239                 struct lu_ref_link link;
2240
2241                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2242
2243                 if (osc_max_rpc_in_flight(cli, osc)) {
2244                         __osc_list_maint(cli, osc);
2245                         break;
2246                 }
2247
2248                 cl_object_get(obj);
2249                 spin_unlock(&cli->cl_loi_list_lock);
2250                 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2251
2252                 /* attempt some read/write balancing by alternating between
2253                  * reads and writes in an object.  The makes_rpc checks here
2254                  * would be redundant if we were getting read/write work items
2255                  * instead of objects.  we don't want send_oap_rpc to drain a
2256                  * partial read pending queue when we're given this object to
2257                  * do io on writes while there are cache waiters */
2258                 osc_object_lock(osc);
2259                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2260                         rc = osc_send_write_rpc(env, cli, osc);
2261                         if (rc < 0) {
2262                                 CERROR("Write request failed with %d\n", rc);
2263
2264                                 /* osc_send_write_rpc failed, mostly because of
2265                                  * memory pressure.
2266                                  *
2267                                  * It can't break here, because if:
2268                                  *  - a page was submitted by osc_io_submit, so
2269                                  *    page locked;
2270                                  *  - no request in flight
2271                                  *  - no subsequent request
2272                                  * The system will be in live-lock state,
2273                                  * because there is no chance to call
2274                                  * osc_io_unplug() and osc_check_rpcs() any
2275                                  * more. pdflush can't help in this case,
2276                                  * because it might be blocked at grabbing
2277                                  * the page lock as we mentioned.
2278                                  *
2279                                  * Anyway, continue to drain pages. */
2280                                 /* break; */
2281                         }
2282                 }
2283                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2284                         rc = osc_send_read_rpc(env, cli, osc);
2285                         if (rc < 0)
2286                                 CERROR("Read request failed with %d\n", rc);
2287                 }
2288                 osc_object_unlock(osc);
2289
2290                 osc_list_maint(cli, osc);
2291                 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2292                 cl_object_put(env, obj);
2293
2294                 spin_lock(&cli->cl_loi_list_lock);
2295         }
2296 }
2297
2298 int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2299                    struct osc_object *osc, int async)
2300 {
2301         int rc = 0;
2302
2303         if (osc != NULL && osc_list_maint(cli, osc) == 0)
2304                 return 0;
2305
2306         if (!async) {
2307                 spin_lock(&cli->cl_loi_list_lock);
2308                 osc_check_rpcs(env, cli);
2309                 spin_unlock(&cli->cl_loi_list_lock);
2310         } else {
2311                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2312                 LASSERT(cli->cl_writeback_work != NULL);
2313                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2314         }
2315         return rc;
2316 }
2317 EXPORT_SYMBOL(osc_io_unplug0);
2318
2319 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2320                         struct page *page, loff_t offset)
2321 {
2322         struct obd_export     *exp = osc_export(osc);
2323         struct osc_async_page *oap = &ops->ops_oap;
2324         ENTRY;
2325
2326         if (!page)
2327                 return cfs_size_round(sizeof(*oap));
2328
2329         oap->oap_magic = OAP_MAGIC;
2330         oap->oap_cli = &exp->exp_obd->u.cli;
2331         oap->oap_obj = osc;
2332
2333         oap->oap_page = page;
2334         oap->oap_obj_off = offset;
2335         LASSERT(!(offset & ~PAGE_MASK));
2336
2337         INIT_LIST_HEAD(&oap->oap_pending_item);
2338         INIT_LIST_HEAD(&oap->oap_rpc_item);
2339
2340         spin_lock_init(&oap->oap_lock);
2341         CDEBUG(D_INFO, "oap %p page %p obj off %llu\n",
2342                oap, page, oap->oap_obj_off);
2343         RETURN(0);
2344 }
2345 EXPORT_SYMBOL(osc_prep_async_page);
2346
2347 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2348                        struct osc_page *ops, cl_commit_cbt cb)
2349 {
2350         struct osc_io *oio = osc_env_io(env);
2351         struct osc_extent     *ext = NULL;
2352         struct osc_async_page *oap = &ops->ops_oap;
2353         struct client_obd     *cli = oap->oap_cli;
2354         struct osc_object     *osc = oap->oap_obj;
2355         struct pagevec        *pvec = &osc_env_info(env)->oti_pagevec;
2356         pgoff_t index;
2357         unsigned int tmp;
2358         unsigned int grants = 0;
2359         u32    brw_flags = OBD_BRW_ASYNC;
2360         int    cmd = OBD_BRW_WRITE;
2361         int    need_release = 0;
2362         int    rc = 0;
2363         ENTRY;
2364
2365         if (oap->oap_magic != OAP_MAGIC)
2366                 RETURN(-EINVAL);
2367
2368         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2369                 RETURN(-EIO);
2370
2371         if (!list_empty(&oap->oap_pending_item) ||
2372             !list_empty(&oap->oap_rpc_item))
2373                 RETURN(-EBUSY);
2374
2375         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2376         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2377         if (oio->oi_cap_sys_resource) {
2378                 brw_flags |= OBD_BRW_NOQUOTA;
2379                 cmd |= OBD_BRW_NOQUOTA;
2380         }
2381
2382         /* check if the file's owner/group is over quota */
2383         if (!(cmd & OBD_BRW_NOQUOTA)) {
2384                 struct cl_object *obj;
2385                 struct cl_attr   *attr;
2386                 unsigned int qid[LL_MAXQUOTAS];
2387
2388                 obj = cl_object_top(&osc->oo_cl);
2389                 attr = &osc_env_info(env)->oti_attr;
2390
2391                 cl_object_attr_lock(obj);
2392                 rc = cl_object_attr_get(env, obj, attr);
2393                 cl_object_attr_unlock(obj);
2394
2395                 qid[USRQUOTA] = attr->cat_uid;
2396                 qid[GRPQUOTA] = attr->cat_gid;
2397                 qid[PRJQUOTA] = attr->cat_projid;
2398                 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2399                         rc = -EDQUOT;
2400                 if (rc)
2401                         RETURN(rc);
2402         }
2403
2404         oap->oap_cmd = cmd;
2405         oap->oap_page_off = ops->ops_from;
2406         oap->oap_count = ops->ops_to - ops->ops_from;
2407         /* No need to hold a lock here,
2408          * since this page is not in any list yet. */
2409         oap->oap_async_flags = 0;
2410         oap->oap_brw_flags = brw_flags;
2411
2412         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2413                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2414
2415         index = osc_index(oap2osc(oap));
2416
2417         /* Add this page into extent by the following steps:
2418          * 1. if there exists an active extent for this IO, mostly this page
2419          *    can be added to the active extent and sometimes we need to
2420          *    expand extent to accomodate this page;
2421          * 2. otherwise, a new extent will be allocated. */
2422
2423         ext = oio->oi_active;
2424         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2425                 /* one chunk plus extent overhead must be enough to write this
2426                  * page */
2427                 grants = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2428                 if (ext->oe_end >= index)
2429                         grants = 0;
2430
2431                 /* it doesn't need any grant to dirty this page */
2432                 spin_lock(&cli->cl_loi_list_lock);
2433                 rc = osc_enter_cache_try(cli, oap, grants);
2434                 if (rc == 0) { /* try failed */
2435                         grants = 0;
2436                         need_release = 1;
2437                 } else if (ext->oe_end < index) {
2438                         tmp = grants;
2439                         /* try to expand this extent */
2440                         rc = osc_extent_expand(ext, index, &tmp);
2441                         if (rc < 0) {
2442                                 need_release = 1;
2443                                 /* don't free reserved grant */
2444                         } else {
2445                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2446                                                 "expanded for %lu.\n", index);
2447                                 osc_unreserve_grant_nolock(cli, grants, tmp);
2448                                 grants = 0;
2449                         }
2450                 }
2451                 spin_unlock(&cli->cl_loi_list_lock);
2452                 rc = 0;
2453         } else if (ext != NULL) {
2454                 /* index is located outside of active extent */
2455                 need_release = 1;
2456         }
2457         if (need_release) {
2458                 osc_extent_release(env, ext);
2459                 oio->oi_active = NULL;
2460                 ext = NULL;
2461         }
2462
2463         if (ext == NULL) {
2464                 tmp = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2465
2466                 /* try to find new extent to cover this page */
2467                 LASSERT(oio->oi_active == NULL);
2468                 /* we may have allocated grant for this page if we failed
2469                  * to expand the previous active extent. */
2470                 LASSERT(ergo(grants > 0, grants >= tmp));
2471
2472                 rc = 0;
2473                 if (grants == 0) {
2474                         /* We haven't allocated grant for this page, and we
2475                          * must not hold a page lock while we do enter_cache,
2476                          * so we must mark dirty & unlock any pages in the
2477                          * write commit pagevec. */
2478                         if (pagevec_count(pvec)) {
2479                                 cb(env, io, pvec);
2480                                 pagevec_reinit(pvec);
2481                         }
2482                         rc = osc_enter_cache(env, cli, oap, tmp);
2483                         if (rc == 0)
2484                                 grants = tmp;
2485                 }
2486
2487                 tmp = grants;
2488                 if (rc == 0) {
2489                         ext = osc_extent_find(env, osc, index, &tmp);
2490                         if (IS_ERR(ext)) {
2491                                 LASSERT(tmp == grants);
2492                                 osc_exit_cache(cli, oap);
2493                                 rc = PTR_ERR(ext);
2494                                 ext = NULL;
2495                         } else {
2496                                 oio->oi_active = ext;
2497                         }
2498                 }
2499                 if (grants > 0)
2500                         osc_unreserve_grant(cli, grants, tmp);
2501         }
2502
2503         LASSERT(ergo(rc == 0, ext != NULL));
2504         if (ext != NULL) {
2505                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2506                          ext, "index = %lu.\n", index);
2507                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2508
2509                 osc_object_lock(osc);
2510                 if (ext->oe_nr_pages == 0)
2511                         ext->oe_srvlock = ops->ops_srvlock;
2512                 else
2513                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2514                 ++ext->oe_nr_pages;
2515                 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2516                 osc_object_unlock(osc);
2517
2518                 if (!ext->oe_layout_version)
2519                         ext->oe_layout_version = io->ci_layout_version;
2520         }
2521
2522         RETURN(rc);
2523 }
2524
2525 int osc_teardown_async_page(const struct lu_env *env,
2526                             struct osc_object *obj, struct osc_page *ops)
2527 {
2528         struct osc_async_page *oap = &ops->ops_oap;
2529         int rc = 0;
2530         ENTRY;
2531
2532         LASSERT(oap->oap_magic == OAP_MAGIC);
2533
2534         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2535                oap, ops, osc_index(oap2osc(oap)));
2536
2537         if (!list_empty(&oap->oap_rpc_item)) {
2538                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2539                 rc = -EBUSY;
2540         } else if (!list_empty(&oap->oap_pending_item)) {
2541                 struct osc_extent *ext = NULL;
2542
2543                 osc_object_lock(obj);
2544                 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2545                 osc_object_unlock(obj);
2546                 /* only truncated pages are allowed to be taken out.
2547                  * See osc_extent_truncate() and osc_cache_truncate_start()
2548                  * for details. */
2549                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2550                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2551                                         osc_index(oap2osc(oap)));
2552                         rc = -EBUSY;
2553                 }
2554                 if (ext != NULL)
2555                         osc_extent_put(env, ext);
2556         }
2557         RETURN(rc);
2558 }
2559
2560 /**
2561  * This is called when a page is picked up by kernel to write out.
2562  *
2563  * We should find out the corresponding extent and add the whole extent
2564  * into urgent list. The extent may be being truncated or used, handle it
2565  * carefully.
2566  */
2567 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2568                          struct osc_page *ops)
2569 {
2570         struct osc_extent *ext   = NULL;
2571         struct osc_object *obj   = cl2osc(ops->ops_cl.cpl_obj);
2572         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2573         pgoff_t            index = osc_index(ops);
2574         struct osc_async_page *oap = &ops->ops_oap;
2575         bool unplug = false;
2576         int rc = 0;
2577         ENTRY;
2578
2579         osc_object_lock(obj);
2580         ext = osc_extent_lookup(obj, index);
2581         if (ext == NULL) {
2582                 osc_extent_tree_dump(D_ERROR, obj);
2583                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2584         }
2585
2586         switch (ext->oe_state) {
2587         case OES_RPC:
2588         case OES_LOCK_DONE:
2589                 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2590                 LASSERT(0);
2591                 break;
2592         case OES_LOCKING:
2593                 /* If we know this extent is being written out, we should abort
2594                  * so that the writer can make this page ready. Otherwise, there
2595                  * exists a deadlock problem because other process can wait for
2596                  * page writeback bit holding page lock; and meanwhile in
2597                  * vvp_page_make_ready(), we need to grab page lock before
2598                  * really sending the RPC. */
2599         case OES_TRUNC:
2600                 /* race with truncate, page will be redirtied */
2601         case OES_ACTIVE:
2602                 /* The extent is active so we need to abort and let the caller
2603                  * re-dirty the page. If we continued on here, and we were the
2604                  * one making the extent active, we could deadlock waiting for
2605                  * the page writeback to clear but it won't because the extent
2606                  * is active and won't be written out. */
2607                 GOTO(out, rc = -EAGAIN);
2608         default:
2609                 break;
2610         }
2611
2612         rc = cl_page_prep(env, io, cp, CRT_WRITE);
2613         if (rc)
2614                 GOTO(out, rc);
2615
2616         spin_lock(&oap->oap_lock);
2617         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2618         spin_unlock(&oap->oap_lock);
2619
2620         if (memory_pressure_get())
2621                 ext->oe_memalloc = 1;
2622
2623         ext->oe_urgent = 1;
2624         if (ext->oe_state == OES_CACHE) {
2625                 OSC_EXTENT_DUMP(D_CACHE, ext,
2626                                 "flush page %p make it urgent.\n", oap);
2627                 if (list_empty(&ext->oe_link))
2628                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2629                 unplug = true;
2630         }
2631         rc = 0;
2632         EXIT;
2633
2634 out:
2635         osc_object_unlock(obj);
2636         osc_extent_put(env, ext);
2637         if (unplug)
2638                 osc_io_unplug_async(env, osc_cli(obj), obj);
2639         return rc;
2640 }
2641
2642 /**
2643  * this is called when a sync waiter receives an interruption.  Its job is to
2644  * get the caller woken as soon as possible.  If its page hasn't been put in an
2645  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2646  * desiring interruption which will forcefully complete the rpc once the rpc
2647  * has timed out.
2648  */
2649 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2650 {
2651         struct osc_async_page *oap = &ops->ops_oap;
2652         struct osc_object     *obj = oap->oap_obj;
2653         struct client_obd     *cli = osc_cli(obj);
2654         struct osc_extent     *ext;
2655         struct osc_extent     *found = NULL;
2656         struct list_head            *plist;
2657         pgoff_t index = osc_index(ops);
2658         int     rc = -EBUSY;
2659         int     cmd;
2660         ENTRY;
2661
2662         LASSERT(!oap->oap_interrupted);
2663         oap->oap_interrupted = 1;
2664
2665         /* Find out the caching extent */
2666         osc_object_lock(obj);
2667         if (oap->oap_cmd & OBD_BRW_WRITE) {
2668                 plist = &obj->oo_urgent_exts;
2669                 cmd   = OBD_BRW_WRITE;
2670         } else {
2671                 plist = &obj->oo_reading_exts;
2672                 cmd   = OBD_BRW_READ;
2673         }
2674         list_for_each_entry(ext, plist, oe_link) {
2675                 if (ext->oe_start <= index && ext->oe_end >= index) {
2676                         LASSERT(ext->oe_state == OES_LOCK_DONE);
2677                         /* For OES_LOCK_DONE state extent, it has already held
2678                          * a refcount for RPC. */
2679                         found = osc_extent_get(ext);
2680                         break;
2681                 }
2682         }
2683         if (found != NULL) {
2684                 list_del_init(&found->oe_link);
2685                 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2686                 osc_object_unlock(obj);
2687
2688                 osc_extent_finish(env, found, 0, -EINTR);
2689                 osc_extent_put(env, found);
2690                 rc = 0;
2691         } else {
2692                 osc_object_unlock(obj);
2693                 /* ok, it's been put in an rpc. only one oap gets a request
2694                  * reference */
2695                 if (oap->oap_request != NULL) {
2696                         ptlrpc_mark_interrupted(oap->oap_request);
2697                         ptlrpcd_wake(oap->oap_request);
2698                         ptlrpc_req_finished(oap->oap_request);
2699                         oap->oap_request = NULL;
2700                 }
2701         }
2702
2703         osc_list_maint(cli, obj);
2704         RETURN(rc);
2705 }
2706
2707 int osc_queue_sync_pages(const struct lu_env *env, const struct cl_io *io,
2708                          struct osc_object *obj, struct list_head *list,
2709                          int brw_flags)
2710 {
2711         struct client_obd     *cli = osc_cli(obj);
2712         struct osc_extent     *ext;
2713         struct osc_async_page *oap;
2714         int     page_count = 0;
2715         int     mppr       = cli->cl_max_pages_per_rpc;
2716         bool    can_merge   = true;
2717         pgoff_t start      = CL_PAGE_EOF;
2718         pgoff_t end        = 0;
2719         ENTRY;
2720
2721         list_for_each_entry(oap, list, oap_pending_item) {
2722                 struct osc_page *opg = oap2osc_page(oap);
2723                 pgoff_t index = osc_index(opg);
2724
2725                 if (index > end)
2726                         end = index;
2727                 if (index < start)
2728                         start = index;
2729                 ++page_count;
2730                 mppr <<= (page_count > mppr);
2731
2732                 if (unlikely(opg->ops_from > 0 || opg->ops_to < PAGE_SIZE))
2733                         can_merge = false;
2734         }
2735
2736         ext = osc_extent_alloc(obj);
2737         if (ext == NULL) {
2738                 struct osc_async_page *tmp;
2739
2740                 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2741                         list_del_init(&oap->oap_pending_item);
2742                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2743                 }
2744                 RETURN(-ENOMEM);
2745         }
2746
2747         ext->oe_rw = !!(brw_flags & OBD_BRW_READ);
2748         ext->oe_sync = 1;
2749         ext->oe_no_merge = !can_merge;
2750         ext->oe_urgent = 1;
2751         ext->oe_start = start;
2752         ext->oe_end = ext->oe_max_end = end;
2753         ext->oe_obj = obj;
2754         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2755         ext->oe_ndelay = !!(brw_flags & OBD_BRW_NDELAY);
2756         ext->oe_nr_pages = page_count;
2757         ext->oe_mppr = mppr;
2758         list_splice_init(list, &ext->oe_pages);
2759         ext->oe_layout_version = io->ci_layout_version;
2760
2761         osc_object_lock(obj);
2762         /* Reuse the initial refcount for RPC, don't drop it */
2763         osc_extent_state_set(ext, OES_LOCK_DONE);
2764         if (!ext->oe_rw) { /* write */
2765                 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2766                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2767         } else {
2768                 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2769                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2770         }
2771         osc_object_unlock(obj);
2772
2773         osc_io_unplug_async(env, cli, obj);
2774         RETURN(0);
2775 }
2776
2777 /**
2778  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2779  */
2780 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
2781                              __u64 size, struct osc_extent **extp)
2782 {
2783         struct client_obd *cli = osc_cli(obj);
2784         struct osc_extent *ext;
2785         struct osc_extent *waiting = NULL;
2786         pgoff_t index;
2787         LIST_HEAD(list);
2788         int result = 0;
2789         bool partial;
2790         ENTRY;
2791
2792         /* pages with index greater or equal to index will be truncated. */
2793         index = cl_index(osc2cl(obj), size);
2794         partial = size > cl_offset(osc2cl(obj), index);
2795
2796 again:
2797         osc_object_lock(obj);
2798         ext = osc_extent_search(obj, index);
2799         if (ext == NULL)
2800                 ext = first_extent(obj);
2801         else if (ext->oe_end < index)
2802                 ext = next_extent(ext);
2803         while (ext != NULL) {
2804                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2805
2806                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2807                         /* if ext is in urgent state, it means there must exist
2808                          * a page already having been flushed by write_page().
2809                          * We have to wait for this extent because we can't
2810                          * truncate that page. */
2811                         OSC_EXTENT_DUMP(D_CACHE, ext,
2812                                         "waiting for busy extent\n");
2813                         waiting = osc_extent_get(ext);
2814                         break;
2815                 }
2816
2817                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:%llu.\n", size);
2818
2819                 osc_extent_get(ext);
2820                 if (ext->oe_state == OES_ACTIVE) {
2821                         /* though we grab inode mutex for write path, but we
2822                          * release it before releasing extent(in osc_io_end()),
2823                          * so there is a race window that an extent is still
2824                          * in OES_ACTIVE when truncate starts. */
2825                         LASSERT(!ext->oe_trunc_pending);
2826                         ext->oe_trunc_pending = 1;
2827                 } else {
2828                         EASSERT(ext->oe_state == OES_CACHE, ext);
2829                         osc_extent_state_set(ext, OES_TRUNC);
2830                         osc_update_pending(obj, OBD_BRW_WRITE,
2831                                            -ext->oe_nr_pages);
2832                 }
2833                 /* This extent could be on the full extents list, that's OK */
2834                 EASSERT(!ext->oe_hp && !ext->oe_urgent, ext);
2835                 if (!list_empty(&ext->oe_link))
2836                         list_move_tail(&ext->oe_link, &list);
2837                 else
2838                         list_add_tail(&ext->oe_link, &list);
2839
2840                 ext = next_extent(ext);
2841         }
2842         osc_object_unlock(obj);
2843
2844         osc_list_maint(cli, obj);
2845
2846         while (!list_empty(&list)) {
2847                 int rc;
2848
2849                 ext = list_entry(list.next, struct osc_extent, oe_link);
2850                 list_del_init(&ext->oe_link);
2851
2852                 /* extent may be in OES_ACTIVE state because inode mutex
2853                  * is released before osc_io_end() in file write case */
2854                 if (ext->oe_state != OES_TRUNC)
2855                         osc_extent_wait(env, ext, OES_TRUNC);
2856
2857                 rc = osc_extent_truncate(ext, index, partial);
2858                 if (rc < 0) {
2859                         if (result == 0)
2860                                 result = rc;
2861
2862                         OSC_EXTENT_DUMP(D_ERROR, ext,
2863                                         "truncate error %d\n", rc);
2864                 } else if (ext->oe_nr_pages == 0) {
2865                         osc_extent_remove(ext);
2866                 } else {
2867                         /* this must be an overlapped extent which means only
2868                          * part of pages in this extent have been truncated.
2869                          */
2870                         EASSERTF(ext->oe_start <= index, ext,
2871                                  "trunc index = %lu/%d.\n", index, partial);
2872                         /* fix index to skip this partially truncated extent */
2873                         index = ext->oe_end + 1;
2874                         partial = false;
2875
2876                         /* we need to hold this extent in OES_TRUNC state so
2877                          * that no writeback will happen. This is to avoid
2878                          * BUG 17397.
2879                          * Only partial truncate can reach here, if @size is
2880                          * not zero, the caller should provide a valid @extp. */
2881                         LASSERT(*extp == NULL);
2882                         *extp = osc_extent_get(ext);
2883                         OSC_EXTENT_DUMP(D_CACHE, ext,
2884                                         "trunc at %llu\n", size);
2885                 }
2886                 osc_extent_put(env, ext);
2887         }
2888         if (waiting != NULL) {
2889                 int rc;
2890
2891                 /* ignore the result of osc_extent_wait the write initiator
2892                  * should take care of it. */
2893                 rc = osc_extent_wait(env, waiting, OES_INV);
2894                 if (rc < 0)
2895                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2896
2897                 osc_extent_put(env, waiting);
2898                 waiting = NULL;
2899                 goto again;
2900         }
2901         RETURN(result);
2902 }
2903 EXPORT_SYMBOL(osc_cache_truncate_start);
2904
2905 /**
2906  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2907  */
2908 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext)
2909 {
2910         if (ext != NULL) {
2911                 struct osc_object *obj = ext->oe_obj;
2912                 bool unplug = false;
2913
2914                 EASSERT(ext->oe_nr_pages > 0, ext);
2915                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2916                 EASSERT(!ext->oe_urgent, ext);
2917
2918                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2919                 osc_object_lock(obj);
2920                 osc_extent_state_set(ext, OES_CACHE);
2921                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2922                         ext->oe_urgent = 1;
2923                         list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2924                         unplug = true;
2925                 }
2926                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2927                 osc_object_unlock(obj);
2928                 osc_extent_put(env, ext);
2929
2930                 if (unplug)
2931                         osc_io_unplug_async(env, osc_cli(obj), obj);
2932         }
2933 }
2934
2935 /**
2936  * Wait for extents in a specific range to be written out.
2937  * The caller must have called osc_cache_writeback_range() to issue IO
2938  * otherwise it will take a long time for this function to finish.
2939  *
2940  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2941  * nobody else can dirty this range of file while we're waiting for
2942  * extents to be written.
2943  */
2944 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2945                          pgoff_t start, pgoff_t end)
2946 {
2947         struct osc_extent *ext;
2948         pgoff_t index = start;
2949         int     result = 0;
2950         ENTRY;
2951
2952 again:
2953         osc_object_lock(obj);
2954         ext = osc_extent_search(obj, index);
2955         if (ext == NULL)
2956                 ext = first_extent(obj);
2957         else if (ext->oe_end < index)
2958                 ext = next_extent(ext);
2959         while (ext != NULL) {
2960                 int rc;
2961
2962                 if (ext->oe_start > end)
2963                         break;
2964
2965                 if (!ext->oe_fsync_wait) {
2966                         ext = next_extent(ext);
2967                         continue;
2968                 }
2969
2970                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2971                              ext->oe_hp || ext->oe_urgent), ext);
2972                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2973                              !ext->oe_hp && ext->oe_urgent), ext);
2974
2975                 index = ext->oe_end + 1;
2976                 osc_extent_get(ext);
2977                 osc_object_unlock(obj);
2978
2979                 rc = osc_extent_wait(env, ext, OES_INV);
2980                 if (result == 0)
2981                         result = rc;
2982                 osc_extent_put(env, ext);
2983                 goto again;
2984         }
2985         osc_object_unlock(obj);
2986
2987         OSC_IO_DEBUG(obj, "sync file range.\n");
2988         RETURN(result);
2989 }
2990 EXPORT_SYMBOL(osc_cache_wait_range);
2991
2992 /**
2993  * Called to write out a range of osc object.
2994  *
2995  * @hp     : should be set this is caused by lock cancel;
2996  * @discard: is set if dirty pages should be dropped - file will be deleted or
2997  *         truncated, this implies there is no partially discarding extents.
2998  *
2999  * Return how many pages will be issued, or error code if error occurred.
3000  */
3001 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
3002                               pgoff_t start, pgoff_t end, int hp, int discard)
3003 {
3004         struct osc_extent *ext;
3005         LIST_HEAD(discard_list);
3006         bool unplug = false;
3007         int result = 0;
3008         ENTRY;
3009
3010         osc_object_lock(obj);
3011         ext = osc_extent_search(obj, start);
3012         if (ext == NULL)
3013                 ext = first_extent(obj);
3014         else if (ext->oe_end < start)
3015                 ext = next_extent(ext);
3016         while (ext != NULL) {
3017                 if (ext->oe_start > end)
3018                         break;
3019
3020                 ext->oe_fsync_wait = 1;
3021                 switch (ext->oe_state) {
3022                 case OES_CACHE:
3023                         result += ext->oe_nr_pages;
3024                         if (!discard) {
3025                                 struct list_head *list = NULL;
3026                                 if (hp) {
3027                                         EASSERT(!ext->oe_hp, ext);
3028                                         ext->oe_hp = 1;
3029                                         list = &obj->oo_hp_exts;
3030                                 } else if (!ext->oe_urgent) {
3031                                         ext->oe_urgent = 1;
3032                                         list = &obj->oo_urgent_exts;
3033                                 }
3034                                 if (list != NULL)
3035                                         list_move_tail(&ext->oe_link, list);
3036                                 unplug = true;
3037                         } else {
3038                                 struct client_obd *cli = osc_cli(obj);
3039                                 int pcc_bits = cli->cl_chunkbits - PAGE_SHIFT;
3040                                 pgoff_t align_by = (1 << pcc_bits);
3041                                 pgoff_t a_start = round_down(start, align_by);
3042                                 pgoff_t a_end = round_up(end, align_by);
3043
3044                                 /* overflow case */
3045                                 if (end && !a_end)
3046                                         a_end = CL_PAGE_EOF;
3047                                 /* the only discarder is lock cancelling, so
3048                                  * [start, end], aligned by chunk size, must
3049                                  * contain this extent */
3050                                 LASSERTF(ext->oe_start >= a_start &&
3051                                          ext->oe_end <= a_end,
3052                                          "ext [%lu, %lu] reg [%lu, %lu] "
3053                                          "orig [%lu %lu] align %lu bits "
3054                                          "%d\n", ext->oe_start, ext->oe_end,
3055                                          a_start, a_end, start, end,
3056                                          align_by, pcc_bits);
3057                                 osc_extent_state_set(ext, OES_LOCKING);
3058                                 ext->oe_owner = current;
3059                                 list_move_tail(&ext->oe_link,
3060                                                    &discard_list);
3061                                 osc_update_pending(obj, OBD_BRW_WRITE,
3062                                                    -ext->oe_nr_pages);
3063                         }
3064                         break;
3065                 case OES_ACTIVE:
3066                         /* It's pretty bad to wait for ACTIVE extents, because
3067                          * we don't know how long we will wait for it to be
3068                          * flushed since it may be blocked at awaiting more
3069                          * grants. We do this for the correctness of fsync. */
3070                         LASSERT(hp == 0 && discard == 0);
3071                         ext->oe_urgent = 1;
3072                         break;
3073                 case OES_TRUNC:
3074                         /* this extent is being truncated, can't do anything
3075                          * for it now. it will be set to urgent after truncate
3076                          * is finished in osc_cache_truncate_end(). */
3077                 default:
3078                         break;
3079                 }
3080                 ext = next_extent(ext);
3081         }
3082         osc_object_unlock(obj);
3083
3084         LASSERT(ergo(!discard, list_empty(&discard_list)));
3085         if (!list_empty(&discard_list)) {
3086                 struct osc_extent *tmp;
3087                 int rc;
3088
3089                 osc_list_maint(osc_cli(obj), obj);
3090                 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
3091                         list_del_init(&ext->oe_link);
3092                         EASSERT(ext->oe_state == OES_LOCKING, ext);
3093
3094                         /* Discard caching pages. We don't actually write this
3095                          * extent out but we complete it as if we did. */
3096                         rc = osc_extent_make_ready(env, ext);
3097                         if (unlikely(rc < 0)) {
3098                                 OSC_EXTENT_DUMP(D_ERROR, ext,
3099                                                 "make_ready returned %d\n", rc);
3100                                 if (result >= 0)
3101                                         result = rc;
3102                         }
3103
3104                         /* finish the extent as if the pages were sent */
3105                         osc_extent_finish(env, ext, 0, 0);
3106                 }
3107         }
3108
3109         if (unplug)
3110                 osc_io_unplug(env, osc_cli(obj), obj);
3111
3112         if (hp || discard) {
3113                 int rc;
3114                 rc = osc_cache_wait_range(env, obj, start, end);
3115                 if (result >= 0 && rc < 0)
3116                         result = rc;
3117         }
3118
3119         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3120         RETURN(result);
3121 }
3122 EXPORT_SYMBOL(osc_cache_writeback_range);
3123
3124 /**
3125  * Returns a list of pages by a given [start, end] of \a obj.
3126  *
3127  * \param resched If not NULL, then we give up before hogging CPU for too
3128  * long and set *resched = 1, in that case caller should implement a retry
3129  * logic.
3130  *
3131  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3132  * crucial in the face of [offset, EOF] locks.
3133  *
3134  * Return at least one page in @queue unless there is no covered page.
3135  */
3136 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3137                         struct osc_object *osc, pgoff_t start, pgoff_t end,
3138                         osc_page_gang_cbt cb, void *cbdata)
3139 {
3140         struct osc_page *ops;
3141         struct pagevec  *pagevec;
3142         void            **pvec;
3143         pgoff_t         idx;
3144         unsigned int    nr;
3145         unsigned int    i;
3146         unsigned int    j;
3147         int             res = CLP_GANG_OKAY;
3148         bool            tree_lock = true;
3149         ENTRY;
3150
3151         idx = start;
3152         pvec = osc_env_info(env)->oti_pvec;
3153         pagevec = &osc_env_info(env)->oti_pagevec;
3154         ll_pagevec_init(pagevec, 0);
3155         spin_lock(&osc->oo_tree_lock);
3156         while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3157                                             idx, OTI_PVEC_SIZE)) > 0) {
3158                 struct cl_page *page;
3159                 bool end_of_region = false;
3160
3161                 for (i = 0, j = 0; i < nr; ++i) {
3162                         ops = pvec[i];
3163                         pvec[i] = NULL;
3164
3165                         idx = osc_index(ops);
3166                         if (idx > end) {
3167                                 end_of_region = true;
3168                                 break;
3169                         }
3170
3171                         page = ops->ops_cl.cpl_page;
3172                         LASSERT(page->cp_type == CPT_CACHEABLE);
3173                         if (page->cp_state == CPS_FREEING)
3174                                 continue;
3175
3176                         cl_page_get(page);
3177                         lu_ref_add_atomic(&page->cp_reference,
3178                                           "gang_lookup", current);
3179                         pvec[j++] = ops;
3180                 }
3181                 ++idx;
3182
3183                 /*
3184                  * Here a delicate locking dance is performed. Current thread
3185                  * holds a reference to a page, but has to own it before it
3186                  * can be placed into queue. Owning implies waiting, so
3187                  * radix-tree lock is to be released. After a wait one has to
3188                  * check that pages weren't truncated (cl_page_own() returns
3189                  * error in the latter case).
3190                  */
3191                 spin_unlock(&osc->oo_tree_lock);
3192                 tree_lock = false;
3193
3194                 for (i = 0; i < j; ++i) {
3195                         ops = pvec[i];
3196                         if (res == CLP_GANG_OKAY)
3197                                 res = (*cb)(env, io, ops, cbdata);
3198
3199                         page = ops->ops_cl.cpl_page;
3200                         lu_ref_del(&page->cp_reference, "gang_lookup", current);
3201                         cl_pagevec_put(env, page, pagevec);
3202                 }
3203                 pagevec_release(pagevec);
3204
3205                 if (nr < OTI_PVEC_SIZE || end_of_region)
3206                         break;
3207
3208                 if (res == CLP_GANG_OKAY && need_resched())
3209                         res = CLP_GANG_RESCHED;
3210                 if (res != CLP_GANG_OKAY)
3211                         break;
3212
3213                 spin_lock(&osc->oo_tree_lock);
3214                 tree_lock = true;
3215         }
3216         if (tree_lock)
3217                 spin_unlock(&osc->oo_tree_lock);
3218         RETURN(res);
3219 }
3220 EXPORT_SYMBOL(osc_page_gang_lookup);
3221
3222 /**
3223  * Check if page @page is covered by an extra lock or discard it.
3224  */
3225 static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3226                                 struct osc_page *ops, void *cbdata)
3227 {
3228         struct osc_thread_info *info = osc_env_info(env);
3229         struct osc_object *osc = cbdata;
3230         pgoff_t index;
3231
3232         index = osc_index(ops);
3233         if (index >= info->oti_fn_index) {
3234                 struct ldlm_lock *tmp;
3235                 struct cl_page *page = ops->ops_cl.cpl_page;
3236
3237                 /* refresh non-overlapped index */
3238                 tmp = osc_dlmlock_at_pgoff(env, osc, index,
3239                                            OSC_DAP_FL_TEST_LOCK);
3240                 if (tmp != NULL) {
3241                         __u64 end = tmp->l_policy_data.l_extent.end;
3242                         /* Cache the first-non-overlapped index so as to skip
3243                          * all pages within [index, oti_fn_index). This is safe
3244                          * because if tmp lock is canceled, it will discard
3245                          * these pages. */
3246                         info->oti_fn_index = cl_index(osc2cl(osc), end + 1);
3247                         if (end == OBD_OBJECT_EOF)
3248                                 info->oti_fn_index = CL_PAGE_EOF;
3249                         LDLM_LOCK_PUT(tmp);
3250                 } else if (cl_page_own(env, io, page) == 0) {
3251                         /* discard the page */
3252                         cl_page_discard(env, io, page);
3253                         cl_page_disown(env, io, page);
3254                 } else {
3255                         LASSERT(page->cp_state == CPS_FREEING);
3256                 }
3257         }
3258
3259         info->oti_next_index = index + 1;
3260         return CLP_GANG_OKAY;
3261 }
3262
3263 int osc_discard_cb(const struct lu_env *env, struct cl_io *io,
3264                    struct osc_page *ops, void *cbdata)
3265 {
3266         struct osc_thread_info *info = osc_env_info(env);
3267         struct cl_page *page = ops->ops_cl.cpl_page;
3268
3269         /* page is top page. */
3270         info->oti_next_index = osc_index(ops) + 1;
3271         if (cl_page_own(env, io, page) == 0) {
3272                 if (!ergo(page->cp_type == CPT_CACHEABLE,
3273                           !PageDirty(cl_page_vmpage(page))))
3274                         CL_PAGE_DEBUG(D_ERROR, env, page,
3275                                         "discard dirty page?\n");
3276
3277                 /* discard the page */
3278                 cl_page_discard(env, io, page);
3279                 cl_page_disown(env, io, page);
3280         } else {
3281                 LASSERT(page->cp_state == CPS_FREEING);
3282         }
3283
3284         return CLP_GANG_OKAY;
3285 }
3286 EXPORT_SYMBOL(osc_discard_cb);
3287
3288 /**
3289  * Discard pages protected by the given lock. This function traverses radix
3290  * tree to find all covering pages and discard them. If a page is being covered
3291  * by other locks, it should remain in cache.
3292  *
3293  * If error happens on any step, the process continues anyway (the reasoning
3294  * behind this being that lock cancellation cannot be delayed indefinitely).
3295  */
3296 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3297                            pgoff_t start, pgoff_t end, bool discard)
3298 {
3299         struct osc_thread_info *info = osc_env_info(env);
3300         struct cl_io *io = osc_env_thread_io(env);
3301         osc_page_gang_cbt cb;
3302         int res;
3303         int result;
3304
3305         ENTRY;
3306
3307         io->ci_obj = cl_object_top(osc2cl(osc));
3308         io->ci_ignore_layout = 1;
3309         result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3310         if (result != 0)
3311                 GOTO(out, result);
3312
3313         cb = discard ? osc_discard_cb : check_and_discard_cb;
3314         info->oti_fn_index = info->oti_next_index = start;
3315         do {
3316                 res = osc_page_gang_lookup(env, io, osc,
3317                                            info->oti_next_index, end, cb, osc);
3318                 if (info->oti_next_index > end)
3319                         break;
3320
3321                 if (res == CLP_GANG_RESCHED)
3322                         cond_resched();
3323         } while (res != CLP_GANG_OKAY);
3324 out:
3325         cl_io_fini(env, io);
3326         RETURN(result);
3327 }
3328
3329
3330 /** @} osc */