lustre/osc/osc_cache.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, 2013, Intel Corporation.
  31  *
  32  */
  33 /*
  34  * This file is part of Lustre, http://www.lustre.org/
  35  * Lustre is a trademark of Sun Microsystems, Inc.
  36  *
  37  * osc cache management.
  38  *
  39  * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_OSC
  43
  44 #include "osc_cl_internal.h"
  45 #include "osc_internal.h"
  46
  47 static int extent_debug; /* set it to be true for more debug */
  48
  49 static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
  50 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
  51                            int state);
  52 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
  53                               struct osc_async_page *oap, int sent, int rc);
  54 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
  55                           int cmd);
  56 static int osc_refresh_count(const struct lu_env *env,
  57                              struct osc_async_page *oap, int cmd);
  58 static int osc_io_unplug_async(const struct lu_env *env,
  59                                struct client_obd *cli, struct osc_object *osc);
  60 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
  61                            unsigned int lost_grant);
  62
  63 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
  64                                   const char *func, int line);
  65 #define osc_extent_tree_dump(lvl, obj) \
  66         osc_extent_tree_dump0(lvl, obj, __func__, __LINE__)
  67
  68 /** \addtogroup osc
  69  *  @{
  70  */
  71
  72 /* ------------------ osc extent ------------------ */
  73 static inline char *ext_flags(struct osc_extent *ext, char *flags)
  74 {
  75         char *buf = flags;
  76         *buf++ = ext->oe_rw ? 'r' : 'w';
  77         if (ext->oe_intree)
  78                 *buf++ = 'i';
  79         if (ext->oe_sync)
  80                 *buf++ = 'S';
  81         if (ext->oe_srvlock)
  82                 *buf++ = 's';
  83         if (ext->oe_hp)
  84                 *buf++ = 'h';
  85         if (ext->oe_urgent)
  86                 *buf++ = 'u';
  87         if (ext->oe_memalloc)
  88                 *buf++ = 'm';
  89         if (ext->oe_trunc_pending)
  90                 *buf++ = 't';
  91         if (ext->oe_fsync_wait)
  92                 *buf++ = 'Y';
  93         *buf = 0;
  94         return flags;
  95 }
  96
  97 static inline char list_empty_marker(struct list_head *list)
  98 {
  99         return list_empty(list) ? '-' : '+';
 100 }
 101
 102 #define EXTSTR       "[%lu -> %lu/%lu]"
 103 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
 104 static const char *oes_strings[] = {
 105         "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
 106
 107 #define OSC_EXTENT_DUMP(lvl, extent, fmt, ...) do {                           \
 108         struct osc_extent *__ext = (extent);                                  \
 109         char __buf[16];                                                       \
 110                                                                               \
 111         CDEBUG(lvl,                                                           \
 112                 "extent %p@{" EXTSTR ", "                                     \
 113                 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt,          \
 114                 /* ----- extent part 0 ----- */                               \
 115                 __ext, EXTPARA(__ext),                                        \
 116                 /* ----- part 1 ----- */                                      \
 117                 atomic_read(&__ext->oe_refc),                         \
 118                 atomic_read(&__ext->oe_users),                        \
 119                 list_empty_marker(&__ext->oe_link),                           \
 120                 oes_strings[__ext->oe_state], ext_flags(__ext, __buf),        \
 121                 __ext->oe_obj,                                                \
 122                 /* ----- part 2 ----- */                                      \
 123                 __ext->oe_grants, __ext->oe_nr_pages,                         \
 124                 list_empty_marker(&__ext->oe_pages),                          \
 125                 waitqueue_active(&__ext->oe_waitq) ? '+' : '-',               \
 126                 __ext->oe_dlmlock, __ext->oe_mppr, __ext->oe_owner,           \
 127                 /* ----- part 4 ----- */                                      \
 128                 ## __VA_ARGS__);                                              \
 129         if (lvl == D_ERROR && __ext->oe_dlmlock != NULL)                      \
 130                 LDLM_ERROR(__ext->oe_dlmlock, "extent: %p\n", __ext);         \
 131         else                                                                  \
 132                 LDLM_DEBUG(__ext->oe_dlmlock, "extent: %p\n", __ext);         \
 133 } while (0)
 134
 135 #undef EASSERTF
 136 #define EASSERTF(expr, ext, fmt, args...) do {                          \
 137         if (!(expr)) {                                                  \
 138                 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args);           \
 139                 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj);           \
 140                 LASSERT(expr);                                          \
 141         }                                                               \
 142 } while (0)
 143
 144 #undef EASSERT
 145 #define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
 146
 147 static inline struct osc_extent *rb_extent(struct rb_node *n)
 148 {
 149         if (n == NULL)
 150                 return NULL;
 151
 152         return container_of(n, struct osc_extent, oe_node);
 153 }
 154
 155 static inline struct osc_extent *next_extent(struct osc_extent *ext)
 156 {
 157         if (ext == NULL)
 158                 return NULL;
 159
 160         LASSERT(ext->oe_intree);
 161         return rb_extent(rb_next(&ext->oe_node));
 162 }
 163
 164 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
 165 {
 166         if (ext == NULL)
 167                 return NULL;
 168
 169         LASSERT(ext->oe_intree);
 170         return rb_extent(rb_prev(&ext->oe_node));
 171 }
 172
 173 static inline struct osc_extent *first_extent(struct osc_object *obj)
 174 {
 175         return rb_extent(rb_first(&obj->oo_root));
 176 }
 177
 178 /* object must be locked by caller. */
 179 static int osc_extent_sanity_check0(struct osc_extent *ext,
 180                                     const char *func, const int line)
 181 {
 182         struct osc_object *obj = ext->oe_obj;
 183         struct osc_async_page *oap;
 184         int page_count;
 185         int rc = 0;
 186
 187         if (!osc_object_is_locked(obj))
 188                 GOTO(out, rc = 9);
 189
 190         if (ext->oe_state >= OES_STATE_MAX)
 191                 GOTO(out, rc = 10);
 192
 193         if (atomic_read(&ext->oe_refc) <= 0)
 194                 GOTO(out, rc = 20);
 195
 196         if (atomic_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
 197                 GOTO(out, rc = 30);
 198
 199         switch (ext->oe_state) {
 200         case OES_INV:
 201                 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
 202                         GOTO(out, rc = 35);
 203                 GOTO(out, rc = 0);
 204                 break;
 205         case OES_ACTIVE:
 206                 if (atomic_read(&ext->oe_users) == 0)
 207                         GOTO(out, rc = 40);
 208                 if (ext->oe_hp)
 209                         GOTO(out, rc = 50);
 210                 if (ext->oe_fsync_wait && !ext->oe_urgent)
 211                         GOTO(out, rc = 55);
 212                 break;
 213         case OES_CACHE:
 214                 if (ext->oe_grants == 0)
 215                         GOTO(out, rc = 60);
 216                 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
 217                         GOTO(out, rc = 65);
 218         default:
 219                 if (atomic_read(&ext->oe_users) > 0)
 220                         GOTO(out, rc = 70);
 221         }
 222
 223         if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
 224                 GOTO(out, rc = 80);
 225
 226         if (ext->oe_sync && ext->oe_grants > 0)
 227                 GOTO(out, rc = 90);
 228
 229         if (ext->oe_dlmlock != NULL) {
 230                 struct ldlm_extent *extent;
 231
 232                 extent = &ext->oe_dlmlock->l_policy_data.l_extent;
 233                 if (!(extent->start <= cl_offset(osc2cl(obj), ext->oe_start) &&
 234                       extent->end   >= cl_offset(osc2cl(obj), ext->oe_max_end)))
 235                         GOTO(out, rc = 100);
 236
 237                 if (!(ext->oe_dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP)))
 238                         GOTO(out, rc = 102);
 239         }
 240
 241         if (ext->oe_nr_pages > ext->oe_mppr)
 242                 GOTO(out, rc = 105);
 243
 244         /* Do not verify page list if extent is in RPC. This is because an
 245          * in-RPC extent is supposed to be exclusively accessible w/o lock. */
 246         if (ext->oe_state > OES_CACHE)
 247                 GOTO(out, rc = 0);
 248
 249         if (!extent_debug)
 250                 GOTO(out, rc = 0);
 251
 252         page_count = 0;
 253         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
 254                 pgoff_t index = osc_index(oap2osc(oap));
 255                 ++page_count;
 256                 if (index > ext->oe_end || index < ext->oe_start)
 257                         GOTO(out, rc = 110);
 258         }
 259         if (page_count != ext->oe_nr_pages)
 260                 GOTO(out, rc = 120);
 261
 262 out:
 263         if (rc != 0)
 264                 OSC_EXTENT_DUMP(D_ERROR, ext,
 265                                 "%s:%d sanity check %p failed with rc = %d\n",
 266                                 func, line, ext, rc);
 267         return rc;
 268 }
 269
 270 #define sanity_check_nolock(ext) \
 271         osc_extent_sanity_check0(ext, __func__, __LINE__)
 272
 273 #define sanity_check(ext) ({                                                   \
 274         int __res;                                                             \
 275         osc_object_lock((ext)->oe_obj);                                        \
 276         __res = sanity_check_nolock(ext);                                      \
 277         osc_object_unlock((ext)->oe_obj);                                      \
 278         __res;                                                                 \
 279 })
 280
 281
 282 /**
 283  * sanity check - to make sure there is no overlapped extent in the tree.
 284  */
 285 static int osc_extent_is_overlapped(struct osc_object *obj,
 286                                     struct osc_extent *ext)
 287 {
 288         struct osc_extent *tmp;
 289
 290         LASSERT(osc_object_is_locked(obj));
 291
 292         if (!extent_debug)
 293                 return 0;
 294
 295         for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
 296                 if (tmp == ext)
 297                         continue;
 298                 if (tmp->oe_end >= ext->oe_start &&
 299                     tmp->oe_start <= ext->oe_end)
 300                         return 1;
 301         }
 302         return 0;
 303 }
 304
 305 static void osc_extent_state_set(struct osc_extent *ext, int state)
 306 {
 307         LASSERT(osc_object_is_locked(ext->oe_obj));
 308         LASSERT(state >= OES_INV && state < OES_STATE_MAX);
 309
 310         /* Never try to sanity check a state changing extent :-) */
 311         /* LASSERT(sanity_check_nolock(ext) == 0); */
 312
 313         /* TODO: validate the state machine */
 314         ext->oe_state = state;
 315         wake_up_all(&ext->oe_waitq);
 316 }
 317
 318 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
 319 {
 320         struct osc_extent *ext;
 321
 322         OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_IOFS);
 323         if (ext == NULL)
 324                 return NULL;
 325
 326         RB_CLEAR_NODE(&ext->oe_node);
 327         ext->oe_obj = obj;
 328         atomic_set(&ext->oe_refc, 1);
 329         atomic_set(&ext->oe_users, 0);
 330         INIT_LIST_HEAD(&ext->oe_link);
 331         ext->oe_state = OES_INV;
 332         INIT_LIST_HEAD(&ext->oe_pages);
 333         init_waitqueue_head(&ext->oe_waitq);
 334         ext->oe_dlmlock = NULL;
 335
 336         return ext;
 337 }
 338
 339 static void osc_extent_free(struct osc_extent *ext)
 340 {
 341         OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
 342 }
 343
 344 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
 345 {
 346         LASSERT(atomic_read(&ext->oe_refc) >= 0);
 347         atomic_inc(&ext->oe_refc);
 348         return ext;
 349 }
 350
 351 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
 352 {
 353         LASSERT(atomic_read(&ext->oe_refc) > 0);
 354         if (atomic_dec_and_test(&ext->oe_refc)) {
 355                 LASSERT(list_empty(&ext->oe_link));
 356                 LASSERT(atomic_read(&ext->oe_users) == 0);
 357                 LASSERT(ext->oe_state == OES_INV);
 358                 LASSERT(!ext->oe_intree);
 359
 360                 if (ext->oe_dlmlock != NULL) {
 361                         lu_ref_add(&ext->oe_dlmlock->l_reference,
 362                                    "osc_extent", ext);
 363                         LDLM_LOCK_PUT(ext->oe_dlmlock);
 364                         ext->oe_dlmlock = NULL;
 365                 }
 366                 osc_extent_free(ext);
 367         }
 368 }
 369
 370 /**
 371  * osc_extent_put_trust() is a special version of osc_extent_put() when
 372  * it's known that the caller is not the last user. This is to address the
 373  * problem of lacking of lu_env ;-).
 374  */
 375 static void osc_extent_put_trust(struct osc_extent *ext)
 376 {
 377         LASSERT(atomic_read(&ext->oe_refc) > 1);
 378         LASSERT(osc_object_is_locked(ext->oe_obj));
 379         atomic_dec(&ext->oe_refc);
 380 }
 381
 382 /**
 383  * Return the extent which includes pgoff @index, or return the greatest
 384  * previous extent in the tree.
 385  */
 386 static struct osc_extent *osc_extent_search(struct osc_object *obj,
 387                                             pgoff_t index)
 388 {
 389         struct rb_node    *n = obj->oo_root.rb_node;
 390         struct osc_extent *tmp, *p = NULL;
 391
 392         LASSERT(osc_object_is_locked(obj));
 393         while (n != NULL) {
 394                 tmp = rb_extent(n);
 395                 if (index < tmp->oe_start) {
 396                         n = n->rb_left;
 397                 } else if (index > tmp->oe_end) {
 398                         p = rb_extent(n);
 399                         n = n->rb_right;
 400                 } else {
 401                         return tmp;
 402                 }
 403         }
 404         return p;
 405 }
 406
 407 /*
 408  * Return the extent covering @index, otherwise return NULL.
 409  * caller must have held object lock.
 410  */
 411 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
 412                                             pgoff_t index)
 413 {
 414         struct osc_extent *ext;
 415
 416         ext = osc_extent_search(obj, index);
 417         if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
 418                 return osc_extent_get(ext);
 419         return NULL;
 420 }
 421
 422 /* caller must have held object lock. */
 423 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
 424 {
 425         struct rb_node   **n      = &obj->oo_root.rb_node;
 426         struct rb_node    *parent = NULL;
 427         struct osc_extent *tmp;
 428
 429         LASSERT(ext->oe_intree == 0);
 430         LASSERT(ext->oe_obj == obj);
 431         LASSERT(osc_object_is_locked(obj));
 432         while (*n != NULL) {
 433                 tmp = rb_extent(*n);
 434                 parent = *n;
 435
 436                 if (ext->oe_end < tmp->oe_start)
 437                         n = &(*n)->rb_left;
 438                 else if (ext->oe_start > tmp->oe_end)
 439                         n = &(*n)->rb_right;
 440                 else
 441                         EASSERTF(0, tmp, EXTSTR"\n", EXTPARA(ext));
 442         }
 443         rb_link_node(&ext->oe_node, parent, n);
 444         rb_insert_color(&ext->oe_node, &obj->oo_root);
 445         osc_extent_get(ext);
 446         ext->oe_intree = 1;
 447 }
 448
 449 /* caller must have held object lock. */
 450 static void osc_extent_erase(struct osc_extent *ext)
 451 {
 452         struct osc_object *obj = ext->oe_obj;
 453         LASSERT(osc_object_is_locked(obj));
 454         if (ext->oe_intree) {
 455                 rb_erase(&ext->oe_node, &obj->oo_root);
 456                 ext->oe_intree = 0;
 457                 /* rbtree held a refcount */
 458                 osc_extent_put_trust(ext);
 459         }
 460 }
 461
 462 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
 463 {
 464         struct osc_object *obj = ext->oe_obj;
 465
 466         LASSERT(osc_object_is_locked(obj));
 467         LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
 468         if (ext->oe_state == OES_CACHE) {
 469                 osc_extent_state_set(ext, OES_ACTIVE);
 470                 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
 471         }
 472         atomic_inc(&ext->oe_users);
 473         list_del_init(&ext->oe_link);
 474         return osc_extent_get(ext);
 475 }
 476
 477 static void __osc_extent_remove(struct osc_extent *ext)
 478 {
 479         LASSERT(osc_object_is_locked(ext->oe_obj));
 480         LASSERT(list_empty(&ext->oe_pages));
 481         osc_extent_erase(ext);
 482         list_del_init(&ext->oe_link);
 483         osc_extent_state_set(ext, OES_INV);
 484         OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
 485 }
 486
 487 static void osc_extent_remove(struct osc_extent *ext)
 488 {
 489         struct osc_object *obj = ext->oe_obj;
 490
 491         osc_object_lock(obj);
 492         __osc_extent_remove(ext);
 493         osc_object_unlock(obj);
 494 }
 495
 496 /**
 497  * This function is used to merge extents to get better performance. It checks
 498  * if @cur and @victim are contiguous at chunk level.
 499  */
 500 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
 501                             struct osc_extent *victim)
 502 {
 503         struct osc_object *obj = cur->oe_obj;
 504         pgoff_t chunk_start;
 505         pgoff_t chunk_end;
 506         int ppc_bits;
 507
 508         LASSERT(cur->oe_state == OES_CACHE);
 509         LASSERT(osc_object_is_locked(obj));
 510         if (victim == NULL)
 511                 return -EINVAL;
 512
 513         if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
 514                 return -EBUSY;
 515
 516         if (cur->oe_max_end != victim->oe_max_end)
 517                 return -ERANGE;
 518
 519         LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
 520         ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_CACHE_SHIFT;
 521         chunk_start = cur->oe_start >> ppc_bits;
 522         chunk_end   = cur->oe_end   >> ppc_bits;
 523         if (chunk_start   != (victim->oe_end >> ppc_bits) + 1 &&
 524             chunk_end + 1 != victim->oe_start >> ppc_bits)
 525                 return -ERANGE;
 526
 527         OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
 528
 529         cur->oe_start     = min(cur->oe_start, victim->oe_start);
 530         cur->oe_end       = max(cur->oe_end,   victim->oe_end);
 531         cur->oe_grants   += victim->oe_grants;
 532         cur->oe_nr_pages += victim->oe_nr_pages;
 533         /* only the following bits are needed to merge */
 534         cur->oe_urgent   |= victim->oe_urgent;
 535         cur->oe_memalloc |= victim->oe_memalloc;
 536         list_splice_init(&victim->oe_pages, &cur->oe_pages);
 537         list_del_init(&victim->oe_link);
 538         victim->oe_nr_pages = 0;
 539
 540         osc_extent_get(victim);
 541         __osc_extent_remove(victim);
 542         osc_extent_put(env, victim);
 543
 544         OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
 545         return 0;
 546 }
 547
 548 /**
 549  * Drop user count of osc_extent, and unplug IO asynchronously.
 550  */
 551 int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
 552 {
 553         struct osc_object *obj = ext->oe_obj;
 554         int rc = 0;
 555         ENTRY;
 556
 557         LASSERT(atomic_read(&ext->oe_users) > 0);
 558         LASSERT(sanity_check(ext) == 0);
 559         LASSERT(ext->oe_grants > 0);
 560
 561         if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
 562                 LASSERT(ext->oe_state == OES_ACTIVE);
 563                 if (ext->oe_trunc_pending) {
 564                         /* a truncate process is waiting for this extent.
 565                          * This may happen due to a race, check
 566                          * osc_cache_truncate_start(). */
 567                         osc_extent_state_set(ext, OES_TRUNC);
 568                         ext->oe_trunc_pending = 0;
 569                 } else {
 570                         osc_extent_state_set(ext, OES_CACHE);
 571                         osc_update_pending(obj, OBD_BRW_WRITE,
 572                                            ext->oe_nr_pages);
 573
 574                         /* try to merge the previous and next extent. */
 575                         osc_extent_merge(env, ext, prev_extent(ext));
 576                         osc_extent_merge(env, ext, next_extent(ext));
 577
 578                         if (ext->oe_urgent)
 579                                 list_move_tail(&ext->oe_link,
 580                                                &obj->oo_urgent_exts);
 581                 }
 582                 osc_object_unlock(obj);
 583
 584                 osc_io_unplug_async(env, osc_cli(obj), obj);
 585         }
 586         osc_extent_put(env, ext);
 587         RETURN(rc);
 588 }
 589
 590 static inline int overlapped(struct osc_extent *ex1, struct osc_extent *ex2)
 591 {
 592         return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
 593 }
 594
 595 /**
 596  * Find or create an extent which includes @index, core function to manage
 597  * extent tree.
 598  */
 599 static struct osc_extent *osc_extent_find(const struct lu_env *env,
 600                                           struct osc_object *obj, pgoff_t index,
 601                                           int *grants)
 602
 603 {
 604         struct client_obd *cli = osc_cli(obj);
 605         struct osc_lock   *olck;
 606         struct cl_lock_descr *descr;
 607         struct osc_extent *cur;
 608         struct osc_extent *ext;
 609         struct osc_extent *conflict = NULL;
 610         struct osc_extent *found = NULL;
 611         pgoff_t    chunk;
 612         pgoff_t    max_end;
 613         int        max_pages; /* max_pages_per_rpc */
 614         int        chunksize;
 615         int        ppc_bits; /* pages per chunk bits */
 616         int        chunk_mask;
 617         int        rc;
 618         ENTRY;
 619
 620         cur = osc_extent_alloc(obj);
 621         if (cur == NULL)
 622                 RETURN(ERR_PTR(-ENOMEM));
 623
 624         olck = osc_env_io(env)->oi_write_osclock;
 625         LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
 626         LASSERT(olck->ols_state == OLS_GRANTED);
 627
 628         descr = &olck->ols_cl.cls_lock->cll_descr;
 629         LASSERT(descr->cld_mode >= CLM_WRITE);
 630
 631         LASSERT(cli->cl_chunkbits >= PAGE_CACHE_SHIFT);
 632         ppc_bits   = cli->cl_chunkbits - PAGE_CACHE_SHIFT;
 633         chunk_mask = ~((1 << ppc_bits) - 1);
 634         chunksize  = 1 << cli->cl_chunkbits;
 635         chunk      = index >> ppc_bits;
 636
 637         /* align end to rpc edge, rpc size may not be a power 2 integer. */
 638         max_pages = cli->cl_max_pages_per_rpc;
 639         LASSERT((max_pages & ~chunk_mask) == 0);
 640         max_end = index - (index % max_pages) + max_pages - 1;
 641         max_end = min_t(pgoff_t, max_end, descr->cld_end);
 642
 643         /* initialize new extent by parameters so far */
 644         cur->oe_max_end = max_end;
 645         cur->oe_start   = index & chunk_mask;
 646         cur->oe_end     = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
 647         if (cur->oe_start < descr->cld_start)
 648                 cur->oe_start = descr->cld_start;
 649         if (cur->oe_end > max_end)
 650                 cur->oe_end = max_end;
 651         cur->oe_grants  = 0;
 652         cur->oe_mppr    = max_pages;
 653         if (olck->ols_dlmlock != NULL) {
 654                 LASSERT(olck->ols_hold);
 655                 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
 656                 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
 657         }
 658
 659         /* grants has been allocated by caller */
 660         LASSERTF(*grants >= chunksize + cli->cl_extent_tax,
 661                  "%u/%u/%u.\n", *grants, chunksize, cli->cl_extent_tax);
 662         LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
 663                  EXTPARA(cur));
 664
 665 restart:
 666         osc_object_lock(obj);
 667         ext = osc_extent_search(obj, cur->oe_start);
 668         if (ext == NULL)
 669                 ext = first_extent(obj);
 670         while (ext != NULL) {
 671                 loff_t ext_chk_start = ext->oe_start >> ppc_bits;
 672                 loff_t ext_chk_end   = ext->oe_end   >> ppc_bits;
 673
 674                 LASSERT(sanity_check_nolock(ext) == 0);
 675                 if (chunk > ext_chk_end + 1)
 676                         break;
 677
 678                 /* if covering by different locks, no chance to match */
 679                 if (olck->ols_dlmlock != ext->oe_dlmlock) {
 680                         EASSERTF(!overlapped(ext, cur), ext,
 681                                  EXTSTR"\n", EXTPARA(cur));
 682
 683                         ext = next_extent(ext);
 684                         continue;
 685                 }
 686
 687                 /* discontiguous chunks? */
 688                 if (chunk + 1 < ext_chk_start) {
 689                         ext = next_extent(ext);
 690                         continue;
 691                 }
 692
 693                 /* ok, from now on, ext and cur have these attrs:
 694                  * 1. covered by the same lock
 695                  * 2. contiguous at chunk level or overlapping. */
 696
 697                 if (overlapped(ext, cur)) {
 698                         /* cur is the minimum unit, so overlapping means
 699                          * full contain. */
 700                         EASSERTF((ext->oe_start <= cur->oe_start &&
 701                                   ext->oe_end >= cur->oe_end),
 702                                  ext, EXTSTR"\n", EXTPARA(cur));
 703
 704                         if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
 705                                 /* for simplicity, we wait for this extent to
 706                                  * finish before going forward. */
 707                                 conflict = osc_extent_get(ext);
 708                                 break;
 709                         }
 710
 711                         found = osc_extent_hold(ext);
 712                         break;
 713                 }
 714
 715                 /* non-overlapped extent */
 716                 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
 717                         /* we can't do anything for a non OES_CACHE extent, or
 718                          * if there is someone waiting for this extent to be
 719                          * flushed, try next one. */
 720                         ext = next_extent(ext);
 721                         continue;
 722                 }
 723
 724                 /* check if they belong to the same rpc slot before trying to
 725                  * merge. the extents are not overlapped and contiguous at
 726                  * chunk level to get here. */
 727                 if (ext->oe_max_end != max_end) {
 728                         /* if they don't belong to the same RPC slot or
 729                          * max_pages_per_rpc has ever changed, do not merge. */
 730                         ext = next_extent(ext);
 731                         continue;
 732                 }
 733
 734                 /* it's required that an extent must be contiguous at chunk
 735                  * level so that we know the whole extent is covered by grant
 736                  * (the pages in the extent are NOT required to be contiguous).
 737                  * Otherwise, it will be too much difficult to know which
 738                  * chunks have grants allocated. */
 739
 740                 /* try to do front merge - extend ext's start */
 741                 if (chunk + 1 == ext_chk_start) {
 742                         /* ext must be chunk size aligned */
 743                         EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
 744
 745                         /* pull ext's start back to cover cur */
 746                         ext->oe_start   = cur->oe_start;
 747                         ext->oe_grants += chunksize;
 748                         *grants -= chunksize;
 749
 750                         found = osc_extent_hold(ext);
 751                 } else if (chunk == ext_chk_end + 1) {
 752                         /* rear merge */
 753                         ext->oe_end     = cur->oe_end;
 754                         ext->oe_grants += chunksize;
 755                         *grants -= chunksize;
 756
 757                         /* try to merge with the next one because we just fill
 758                          * in a gap */
 759                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 760                                 /* we can save extent tax from next extent */
 761                                 *grants += cli->cl_extent_tax;
 762
 763                         found = osc_extent_hold(ext);
 764                 }
 765                 if (found != NULL)
 766                         break;
 767
 768                 ext = next_extent(ext);
 769         }
 770
 771         osc_extent_tree_dump(D_CACHE, obj);
 772         if (found != NULL) {
 773                 LASSERT(conflict == NULL);
 774                 if (!IS_ERR(found)) {
 775                         LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
 776                         OSC_EXTENT_DUMP(D_CACHE, found,
 777                                         "found caching ext for %lu.\n", index);
 778                 }
 779         } else if (conflict == NULL) {
 780                 /* create a new extent */
 781                 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
 782                 cur->oe_grants = chunksize + cli->cl_extent_tax;
 783                 *grants -= cur->oe_grants;
 784                 LASSERT(*grants >= 0);
 785
 786                 cur->oe_state = OES_CACHE;
 787                 found = osc_extent_hold(cur);
 788                 osc_extent_insert(obj, cur);
 789                 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
 790                                 index, descr->cld_end);
 791         }
 792         osc_object_unlock(obj);
 793
 794         if (conflict != NULL) {
 795                 LASSERT(found == NULL);
 796
 797                 /* waiting for IO to finish. Please notice that it's impossible
 798                  * to be an OES_TRUNC extent. */
 799                 rc = osc_extent_wait(env, conflict, OES_INV);
 800                 osc_extent_put(env, conflict);
 801                 conflict = NULL;
 802                 if (rc < 0)
 803                         GOTO(out, found = ERR_PTR(rc));
 804
 805                 goto restart;
 806         }
 807         EXIT;
 808
 809 out:
 810         osc_extent_put(env, cur);
 811         LASSERT(*grants >= 0);
 812         return found;
 813 }
 814
 815 /**
 816  * Called when IO is finished to an extent.
 817  */
 818 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
 819                       int sent, int rc)
 820 {
 821         struct client_obd *cli = osc_cli(ext->oe_obj);
 822         struct osc_async_page *oap;
 823         struct osc_async_page *tmp;
 824         int nr_pages = ext->oe_nr_pages;
 825         int lost_grant = 0;
 826         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
 827         __u64 last_off = 0;
 828         int last_count = -1;
 829         ENTRY;
 830
 831         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
 832
 833         ext->oe_rc = rc ?: ext->oe_nr_pages;
 834         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
 835
 836         osc_lru_add_batch(cli, &ext->oe_pages);
 837         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 838                                      oap_pending_item) {
 839                 list_del_init(&oap->oap_rpc_item);
 840                 list_del_init(&oap->oap_pending_item);
 841                 if (last_off <= oap->oap_obj_off) {
 842                         last_off = oap->oap_obj_off;
 843                         last_count = oap->oap_count;
 844                 }
 845
 846                 --ext->oe_nr_pages;
 847                 osc_ap_completion(env, cli, oap, sent, rc);
 848         }
 849         EASSERT(ext->oe_nr_pages == 0, ext);
 850
 851         if (!sent) {
 852                 lost_grant = ext->oe_grants;
 853         } else if (blocksize < PAGE_CACHE_SIZE &&
 854                    last_count != PAGE_CACHE_SIZE) {
 855                 /* For short writes we shouldn't count parts of pages that
 856                  * span a whole chunk on the OST side, or our accounting goes
 857                  * wrong.  Should match the code in filter_grant_check. */
 858                 int offset = last_off & ~CFS_PAGE_MASK;
 859                 int count = last_count + (offset & (blocksize - 1));
 860                 int end = (offset + last_count) & (blocksize - 1);
 861                 if (end)
 862                         count += blocksize - end;
 863
 864                 lost_grant = PAGE_CACHE_SIZE - count;
 865         }
 866         if (ext->oe_grants > 0)
 867                 osc_free_grant(cli, nr_pages, lost_grant);
 868
 869         osc_extent_remove(ext);
 870         /* put the refcount for RPC */
 871         osc_extent_put(env, ext);
 872         RETURN(0);
 873 }
 874
 875 static int extent_wait_cb(struct osc_extent *ext, int state)
 876 {
 877         int ret;
 878
 879         osc_object_lock(ext->oe_obj);
 880         ret = ext->oe_state == state;
 881         osc_object_unlock(ext->oe_obj);
 882
 883         return ret;
 884 }
 885
 886 /**
 887  * Wait for the extent's state to become @state.
 888  */
 889 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
 890                            int state)
 891 {
 892         struct osc_object *obj = ext->oe_obj;
 893         struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
 894                                                   LWI_ON_SIGNAL_NOOP, NULL);
 895         int rc = 0;
 896         ENTRY;
 897
 898         osc_object_lock(obj);
 899         LASSERT(sanity_check_nolock(ext) == 0);
 900         /* `Kick' this extent only if the caller is waiting for it to be
 901          * written out. */
 902         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
 903                 if (ext->oe_state == OES_ACTIVE) {
 904                         ext->oe_urgent = 1;
 905                 } else if (ext->oe_state == OES_CACHE) {
 906                         ext->oe_urgent = 1;
 907                         osc_extent_hold(ext);
 908                         rc = 1;
 909                 }
 910         }
 911         osc_object_unlock(obj);
 912         if (rc == 1)
 913                 osc_extent_release(env, ext);
 914
 915         /* wait for the extent until its state becomes @state */
 916         rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
 917         if (rc == -ETIMEDOUT) {
 918                 OSC_EXTENT_DUMP(D_ERROR, ext,
 919                         "%s: wait ext to %d timedout, recovery in progress?\n",
 920                         osc_export(obj)->exp_obd->obd_name, state);
 921
 922                 lwi = LWI_INTR(NULL, NULL);
 923                 rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state),
 924                                   &lwi);
 925         }
 926         if (rc == 0 && ext->oe_rc < 0)
 927                 rc = ext->oe_rc;
 928         RETURN(rc);
 929 }
 930
 931 /**
 932  * Discard pages with index greater than @size. If @ext is overlapped with
 933  * @size, then partial truncate happens.
 934  */
 935 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
 936                                 bool partial)
 937 {
 938         struct cl_env_nest     nest;
 939         struct lu_env         *env;
 940         struct cl_io          *io;
 941         struct osc_object     *obj = ext->oe_obj;
 942         struct client_obd     *cli = osc_cli(obj);
 943         struct osc_async_page *oap;
 944         struct osc_async_page *tmp;
 945         int                    pages_in_chunk = 0;
 946         int                    ppc_bits    = cli->cl_chunkbits -
 947                                              PAGE_CACHE_SHIFT;
 948         __u64                  trunc_chunk = trunc_index >> ppc_bits;
 949         int                    grants   = 0;
 950         int                    nr_pages = 0;
 951         int                    rc       = 0;
 952         ENTRY;
 953
 954         LASSERT(sanity_check(ext) == 0);
 955         LASSERT(ext->oe_state == OES_TRUNC);
 956         LASSERT(!ext->oe_urgent);
 957
 958         /* Request new lu_env.
 959          * We can't use that env from osc_cache_truncate_start() because
 960          * it's from lov_io_sub and not fully initialized. */
 961         env = cl_env_nested_get(&nest);
 962         io  = &osc_env_info(env)->oti_io;
 963         io->ci_obj = cl_object_top(osc2cl(obj));
 964         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
 965         if (rc < 0)
 966                 GOTO(out, rc);
 967
 968         /* discard all pages with index greater then trunc_index */
 969         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 970                                      oap_pending_item) {
 971                 pgoff_t index = osc_index(oap2osc(oap));
 972                 struct cl_page  *page = oap2cl_page(oap);
 973
 974                 LASSERT(list_empty(&oap->oap_rpc_item));
 975
 976                 /* only discard the pages with their index greater than
 977                  * trunc_index, and ... */
 978                 if (index < trunc_index ||
 979                     (index == trunc_index && partial)) {
 980                         /* accounting how many pages remaining in the chunk
 981                          * so that we can calculate grants correctly. */
 982                         if (index >> ppc_bits == trunc_chunk)
 983                                 ++pages_in_chunk;
 984                         continue;
 985                 }
 986
 987                 list_del_init(&oap->oap_pending_item);
 988
 989                 cl_page_get(page);
 990                 lu_ref_add(&page->cp_reference, "truncate", current);
 991
 992                 if (cl_page_own(env, io, page) == 0) {
 993                         cl_page_discard(env, io, page);
 994                         cl_page_disown(env, io, page);
 995                 } else {
 996                         LASSERT(page->cp_state == CPS_FREEING);
 997                         LASSERT(0);
 998                 }
 999
1000                 lu_ref_del(&page->cp_reference, "truncate", current);
1001                 cl_page_put(env, page);
1002
1003                 --ext->oe_nr_pages;
1004                 ++nr_pages;
1005         }
1006         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1007                       ext->oe_nr_pages == 0),
1008                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1009
1010         osc_object_lock(obj);
1011         if (ext->oe_nr_pages == 0) {
1012                 LASSERT(pages_in_chunk == 0);
1013                 grants = ext->oe_grants;
1014                 ext->oe_grants = 0;
1015         } else { /* calculate how many grants we can free */
1016                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1017                 pgoff_t last_index;
1018
1019
1020                 /* if there is no pages in this chunk, we can also free grants
1021                  * for the last chunk */
1022                 if (pages_in_chunk == 0) {
1023                         /* if this is the 1st chunk and no pages in this chunk,
1024                          * ext->oe_nr_pages must be zero, so we should be in
1025                          * the other if-clause. */
1026                         LASSERT(trunc_chunk > 0);
1027                         --trunc_chunk;
1028                         ++chunks;
1029                 }
1030
1031                 /* this is what we can free from this extent */
1032                 grants          = chunks << cli->cl_chunkbits;
1033                 ext->oe_grants -= grants;
1034                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1035                 ext->oe_end     = min(last_index, ext->oe_max_end);
1036                 LASSERT(ext->oe_end >= ext->oe_start);
1037                 LASSERT(ext->oe_grants > 0);
1038         }
1039         osc_object_unlock(obj);
1040
1041         if (grants > 0 || nr_pages > 0)
1042                 osc_free_grant(cli, nr_pages, grants);
1043
1044 out:
1045         cl_io_fini(env, io);
1046         cl_env_nested_put(&nest, env);
1047         RETURN(rc);
1048 }
1049
1050 /**
1051  * This function is used to make the extent prepared for transfer.
1052  * A race with flusing page - ll_writepage() has to be handled cautiously.
1053  */
1054 static int osc_extent_make_ready(const struct lu_env *env,
1055                                  struct osc_extent *ext)
1056 {
1057         struct osc_async_page *oap;
1058         struct osc_async_page *last = NULL;
1059         struct osc_object *obj = ext->oe_obj;
1060         int page_count = 0;
1061         int rc;
1062         ENTRY;
1063
1064         /* we're going to grab page lock, so object lock must not be taken. */
1065         LASSERT(sanity_check(ext) == 0);
1066         /* in locking state, any process should not touch this extent. */
1067         EASSERT(ext->oe_state == OES_LOCKING, ext);
1068         EASSERT(ext->oe_owner != NULL, ext);
1069
1070         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1071
1072         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1073                 ++page_count;
1074                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1075                         last = oap;
1076
1077                 /* checking ASYNC_READY is race safe */
1078                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1079                         continue;
1080
1081                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1082                 switch (rc) {
1083                 case 0:
1084                         spin_lock(&oap->oap_lock);
1085                         oap->oap_async_flags |= ASYNC_READY;
1086                         spin_unlock(&oap->oap_lock);
1087                         break;
1088                 case -EALREADY:
1089                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1090                         break;
1091                 default:
1092                         LASSERTF(0, "unknown return code: %d\n", rc);
1093                 }
1094         }
1095
1096         LASSERT(page_count == ext->oe_nr_pages);
1097         LASSERT(last != NULL);
1098         /* the last page is the only one we need to refresh its count by
1099          * the size of file. */
1100         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1101                 last->oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1102                 LASSERT(last->oap_count > 0);
1103                 LASSERT(last->oap_page_off + last->oap_count <= PAGE_CACHE_SIZE);
1104                 spin_lock(&last->oap_lock);
1105                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1106                 spin_unlock(&last->oap_lock);
1107         }
1108
1109         /* for the rest of pages, we don't need to call osf_refresh_count()
1110          * because it's known they are not the last page */
1111         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1112                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1113                         oap->oap_count = PAGE_CACHE_SIZE - oap->oap_page_off;
1114                         spin_lock(&oap->oap_lock);
1115                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1116                         spin_unlock(&oap->oap_lock);
1117                 }
1118         }
1119
1120         osc_object_lock(obj);
1121         osc_extent_state_set(ext, OES_RPC);
1122         osc_object_unlock(obj);
1123         /* get a refcount for RPC. */
1124         osc_extent_get(ext);
1125
1126         RETURN(0);
1127 }
1128
1129 /**
1130  * Quick and simple version of osc_extent_find(). This function is frequently
1131  * called to expand the extent for the same IO. To expand the extent, the
1132  * page index must be in the same or next chunk of ext->oe_end.
1133  */
1134 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index, int *grants)
1135 {
1136         struct osc_object *obj = ext->oe_obj;
1137         struct client_obd *cli = osc_cli(obj);
1138         struct osc_extent *next;
1139         int ppc_bits = cli->cl_chunkbits - PAGE_CACHE_SHIFT;
1140         pgoff_t chunk = index >> ppc_bits;
1141         pgoff_t end_chunk;
1142         pgoff_t end_index;
1143         int chunksize = 1 << cli->cl_chunkbits;
1144         int rc = 0;
1145         ENTRY;
1146
1147         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1148         osc_object_lock(obj);
1149         LASSERT(sanity_check_nolock(ext) == 0);
1150         end_chunk = ext->oe_end >> ppc_bits;
1151         if (chunk > end_chunk + 1)
1152                 GOTO(out, rc = -ERANGE);
1153
1154         if (end_chunk >= chunk)
1155                 GOTO(out, rc = 0);
1156
1157         LASSERT(end_chunk + 1 == chunk);
1158         /* try to expand this extent to cover @index */
1159         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1160
1161         next = next_extent(ext);
1162         if (next != NULL && next->oe_start <= end_index)
1163                 /* complex mode - overlapped with the next extent,
1164                  * this case will be handled by osc_extent_find() */
1165                 GOTO(out, rc = -EAGAIN);
1166
1167         ext->oe_end = end_index;
1168         ext->oe_grants += chunksize;
1169         *grants -= chunksize;
1170         LASSERT(*grants >= 0);
1171         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1172                  "overlapped after expanding for %lu.\n", index);
1173         EXIT;
1174
1175 out:
1176         osc_object_unlock(obj);
1177         RETURN(rc);
1178 }
1179
1180 static void osc_extent_tree_dump0(int level, struct osc_object *obj,
1181                                   const char *func, int line)
1182 {
1183         struct osc_extent *ext;
1184         int cnt;
1185
1186         CDEBUG(level, "Dump object %p extents at %s:%d, mppr: %u.\n",
1187                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1188
1189         /* osc_object_lock(obj); */
1190         cnt = 1;
1191         for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1192                 OSC_EXTENT_DUMP(level, ext, "in tree %d.\n", cnt++);
1193
1194         cnt = 1;
1195         list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1196                 OSC_EXTENT_DUMP(level, ext, "hp %d.\n", cnt++);
1197
1198         cnt = 1;
1199         list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1200                 OSC_EXTENT_DUMP(level, ext, "urgent %d.\n", cnt++);
1201
1202         cnt = 1;
1203         list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1204                 OSC_EXTENT_DUMP(level, ext, "reading %d.\n", cnt++);
1205         /* osc_object_unlock(obj); */
1206 }
1207
1208 /* ------------------ osc extent end ------------------ */
1209
1210 static inline int osc_is_ready(struct osc_object *osc)
1211 {
1212         return !list_empty(&osc->oo_ready_item) ||
1213                !list_empty(&osc->oo_hp_ready_item);
1214 }
1215
1216 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1217         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1218                (OSC), osc_is_ready(OSC),                                       \
1219                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1220                list_empty_marker(&(OSC)->oo_ready_item),                       \
1221                atomic_read(&(OSC)->oo_nr_writes),                              \
1222                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1223                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1224                atomic_read(&(OSC)->oo_nr_reads),                               \
1225                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1226                ##args)
1227
1228 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1229                           int cmd)
1230 {
1231         struct osc_page *opg  = oap2osc_page(oap);
1232         struct cl_page  *page = oap2cl_page(oap);
1233         int result;
1234
1235         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1236
1237         ENTRY;
1238         result = cl_page_make_ready(env, page, CRT_WRITE);
1239         if (result == 0)
1240                 opg->ops_submit_time = cfs_time_current();
1241         RETURN(result);
1242 }
1243
1244 static int osc_refresh_count(const struct lu_env *env,
1245                              struct osc_async_page *oap, int cmd)
1246 {
1247         struct osc_page  *opg = oap2osc_page(oap);
1248         pgoff_t index = osc_index(oap2osc(oap));
1249         struct cl_object *obj;
1250         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
1251
1252         int result;
1253         loff_t kms;
1254
1255         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1256         LASSERT(!(cmd & OBD_BRW_READ));
1257         LASSERT(opg != NULL);
1258         obj = opg->ops_cl.cpl_obj;
1259
1260         cl_object_attr_lock(obj);
1261         result = cl_object_attr_get(env, obj, attr);
1262         cl_object_attr_unlock(obj);
1263         if (result < 0)
1264                 return result;
1265         kms = attr->cat_kms;
1266         if (cl_offset(obj, index) >= kms)
1267                 /* catch race with truncate */
1268                 return 0;
1269         else if (cl_offset(obj, index + 1) > kms)
1270                 /* catch sub-page write at end of file */
1271                 return kms % PAGE_CACHE_SIZE;
1272         else
1273                 return PAGE_CACHE_SIZE;
1274 }
1275
1276 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1277                           int cmd, int rc)
1278 {
1279         struct osc_page   *opg  = oap2osc_page(oap);
1280         struct cl_page    *page = oap2cl_page(oap);
1281         struct osc_object *obj  = cl2osc(opg->ops_cl.cpl_obj);
1282         enum cl_req_type   crt;
1283         int srvlock;
1284
1285         ENTRY;
1286
1287         cmd &= ~OBD_BRW_NOQUOTA;
1288         LASSERTF(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ),
1289                  "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1290         LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1291                 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1292         LASSERT(opg->ops_transfer_pinned);
1293
1294         /*
1295          * page->cp_req can be NULL if io submission failed before
1296          * cl_req was allocated.
1297          */
1298         if (page->cp_req != NULL)
1299                 cl_req_page_done(env, page);
1300         LASSERT(page->cp_req == NULL);
1301
1302         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1303         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1304         opg->ops_transfer_pinned = 0;
1305
1306         spin_lock(&obj->oo_seatbelt);
1307         LASSERT(opg->ops_submitter != NULL);
1308         LASSERT(!list_empty(&opg->ops_inflight));
1309         list_del_init(&opg->ops_inflight);
1310         opg->ops_submitter = NULL;
1311         spin_unlock(&obj->oo_seatbelt);
1312
1313         opg->ops_submit_time = 0;
1314         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1315
1316         /* statistic */
1317         if (rc == 0 && srvlock) {
1318                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1319                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1320                 int bytes = oap->oap_count;
1321
1322                 if (crt == CRT_READ)
1323                         stats->os_lockless_reads += bytes;
1324                 else
1325                         stats->os_lockless_writes += bytes;
1326         }
1327
1328         /*
1329          * This has to be the last operation with the page, as locks are
1330          * released in cl_page_completion() and nothing except for the
1331          * reference counter protects page from concurrent reclaim.
1332          */
1333         lu_ref_del(&page->cp_reference, "transfer", page);
1334
1335         cl_page_completion(env, page, crt, rc);
1336
1337         RETURN(0);
1338 }
1339
1340 #define OSC_DUMP_GRANT(lvl, cli, fmt, args...) do {                     \
1341         struct client_obd *__tmp = (cli);                               \
1342         CDEBUG(lvl, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu "  \
1343                "dropped: %ld avail: %ld, reserved: %ld, flight: %d }"   \
1344                "lru {in list: %ld, left: %ld, waiters: %d }"fmt"\n",    \
1345                __tmp->cl_import->imp_obd->obd_name,                     \
1346                __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages,        \
1347                atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1348                __tmp->cl_lost_grant, __tmp->cl_avail_grant,             \
1349                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,         \
1350                atomic_long_read(&__tmp->cl_lru_in_list),                \
1351                atomic_long_read(&__tmp->cl_lru_busy),                   \
1352                atomic_read(&__tmp->cl_lru_shrinkers), ##args);          \
1353 } while (0)
1354
1355 /* caller must hold loi_list_lock */
1356 static void osc_consume_write_grant(struct client_obd *cli,
1357                                     struct brw_page *pga)
1358 {
1359         assert_spin_locked(&cli->cl_loi_list_lock);
1360         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1361         atomic_long_inc(&obd_dirty_pages);
1362         cli->cl_dirty_pages++;
1363         pga->flag |= OBD_BRW_FROM_GRANT;
1364         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1365                PAGE_CACHE_SIZE, pga, pga->pg);
1366         osc_update_next_shrink(cli);
1367 }
1368
1369 /* the companion to osc_consume_write_grant, called when a brw has completed.
1370  * must be called with the loi lock held. */
1371 static void osc_release_write_grant(struct client_obd *cli,
1372                                     struct brw_page *pga)
1373 {
1374         ENTRY;
1375
1376         assert_spin_locked(&cli->cl_loi_list_lock);
1377         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1378                 EXIT;
1379                 return;
1380         }
1381
1382         pga->flag &= ~OBD_BRW_FROM_GRANT;
1383         atomic_long_dec(&obd_dirty_pages);
1384         cli->cl_dirty_pages--;
1385         if (pga->flag & OBD_BRW_NOCACHE) {
1386                 pga->flag &= ~OBD_BRW_NOCACHE;
1387                 atomic_long_dec(&obd_dirty_transit_pages);
1388                 cli->cl_dirty_transit--;
1389         }
1390         EXIT;
1391 }
1392
1393 /**
1394  * To avoid sleeping with object lock held, it's good for us allocate enough
1395  * grants before entering into critical section.
1396  *
1397  * client_obd_list_lock held by caller
1398  */
1399 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1400 {
1401         int rc = -EDQUOT;
1402
1403         if (cli->cl_avail_grant >= bytes) {
1404                 cli->cl_avail_grant    -= bytes;
1405                 cli->cl_reserved_grant += bytes;
1406                 rc = 0;
1407         }
1408         return rc;
1409 }
1410
1411 static void __osc_unreserve_grant(struct client_obd *cli,
1412                                   unsigned int reserved, unsigned int unused)
1413 {
1414         /* it's quite normal for us to get more grant than reserved.
1415          * Thinking about a case that two extents merged by adding a new
1416          * chunk, we can save one extent tax. If extent tax is greater than
1417          * one chunk, we can save more grant by adding a new chunk */
1418         cli->cl_reserved_grant -= reserved;
1419         if (unused > reserved) {
1420                 cli->cl_avail_grant += reserved;
1421                 cli->cl_lost_grant  += unused - reserved;
1422         } else {
1423                 cli->cl_avail_grant += unused;
1424         }
1425 }
1426
1427 static void osc_unreserve_grant(struct client_obd *cli,
1428                                 unsigned int reserved, unsigned int unused)
1429 {
1430         spin_lock(&cli->cl_loi_list_lock);
1431         __osc_unreserve_grant(cli, reserved, unused);
1432         if (unused > 0)
1433                 osc_wake_cache_waiters(cli);
1434         spin_unlock(&cli->cl_loi_list_lock);
1435 }
1436
1437 /**
1438  * Free grant after IO is finished or canceled.
1439  *
1440  * @lost_grant is used to remember how many grants we have allocated but not
1441  * used, we should return these grants to OST. There're two cases where grants
1442  * can be lost:
1443  * 1. truncate;
1444  * 2. blocksize at OST is less than PAGE_CACHE_SIZE and a partial page was
1445  *    written. In this case OST may use less chunks to serve this partial
1446  *    write. OSTs don't actually know the page size on the client side. so
1447  *    clients have to calculate lost grant by the blocksize on the OST.
1448  *    See filter_grant_check() for details.
1449  */
1450 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1451                            unsigned int lost_grant)
1452 {
1453         int grant = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
1454
1455         spin_lock(&cli->cl_loi_list_lock);
1456         atomic_long_sub(nr_pages, &obd_dirty_pages);
1457         cli->cl_dirty_pages -= nr_pages;
1458         cli->cl_lost_grant += lost_grant;
1459         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1460                 /* borrow some grant from truncate to avoid the case that
1461                  * truncate uses up all avail grant */
1462                 cli->cl_lost_grant -= grant;
1463                 cli->cl_avail_grant += grant;
1464         }
1465         osc_wake_cache_waiters(cli);
1466         spin_unlock(&cli->cl_loi_list_lock);
1467         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu\n",
1468                lost_grant, cli->cl_lost_grant,
1469                cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_CACHE_SHIFT);
1470 }
1471
1472 /**
1473  * The companion to osc_enter_cache(), called when @oap is no longer part of
1474  * the dirty accounting due to error.
1475  */
1476 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1477 {
1478         spin_lock(&cli->cl_loi_list_lock);
1479         osc_release_write_grant(cli, &oap->oap_brw_page);
1480         spin_unlock(&cli->cl_loi_list_lock);
1481 }
1482
1483 /**
1484  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1485  * is available.
1486  */
1487 static int osc_enter_cache_try(struct client_obd *cli,
1488                                struct osc_async_page *oap,
1489                                int bytes, int transient)
1490 {
1491         int rc;
1492
1493         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d.\n", bytes);
1494
1495         rc = osc_reserve_grant(cli, bytes);
1496         if (rc < 0)
1497                 return 0;
1498
1499         if (cli->cl_dirty_pages < cli->cl_dirty_max_pages &&
1500             1 + atomic_long_read(&obd_dirty_pages) <= obd_max_dirty_pages) {
1501                 osc_consume_write_grant(cli, &oap->oap_brw_page);
1502                 if (transient) {
1503                         cli->cl_dirty_transit++;
1504                         atomic_long_inc(&obd_dirty_transit_pages);
1505                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
1506                 }
1507                 rc = 1;
1508         } else {
1509                 __osc_unreserve_grant(cli, bytes, bytes);
1510                 rc = 0;
1511         }
1512         return rc;
1513 }
1514
1515 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1516 {
1517         int rc;
1518         spin_lock(&cli->cl_loi_list_lock);
1519         rc = list_empty(&ocw->ocw_entry);
1520         spin_unlock(&cli->cl_loi_list_lock);
1521         return rc;
1522 }
1523
1524 /**
1525  * The main entry to reserve dirty page accounting. Usually the grant reserved
1526  * in this function will be freed in bulk in osc_free_grant() unless it fails
1527  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1528  *
1529  * The process will be put into sleep if it's already run out of grant.
1530  */
1531 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1532                            struct osc_async_page *oap, int bytes)
1533 {
1534         struct osc_object *osc = oap->oap_obj;
1535         struct lov_oinfo  *loi = osc->oo_oinfo;
1536         struct osc_cache_waiter ocw;
1537         struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
1538                                                   LWI_ON_SIGNAL_NOOP, NULL);
1539         int rc = -EDQUOT;
1540         ENTRY;
1541
1542         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d.\n", bytes);
1543
1544         spin_lock(&cli->cl_loi_list_lock);
1545
1546         /* force the caller to try sync io.  this can jump the list
1547          * of queued writes and create a discontiguous rpc stream */
1548         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1549             cli->cl_dirty_max_pages == 0 ||
1550             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync)
1551                 GOTO(out, rc = -EDQUOT);
1552
1553         /* Hopefully normal case - cache space and write credits available */
1554         if (osc_enter_cache_try(cli, oap, bytes, 0))
1555                 GOTO(out, rc = 0);
1556
1557         /* We can get here for two reasons: too many dirty pages in cache, or
1558          * run out of grants. In both cases we should write dirty pages out.
1559          * Adding a cache waiter will trigger urgent write-out no matter what
1560          * RPC size will be.
1561          * The exiting condition is no avail grants and no dirty pages caching,
1562          * that really means there is no space on the OST. */
1563         init_waitqueue_head(&ocw.ocw_waitq);
1564         ocw.ocw_oap   = oap;
1565         ocw.ocw_grant = bytes;
1566         while (cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0) {
1567                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1568                 ocw.ocw_rc = 0;
1569                 spin_unlock(&cli->cl_loi_list_lock);
1570
1571                 osc_io_unplug_async(env, cli, NULL);
1572
1573                 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
1574                        cli->cl_import->imp_obd->obd_name, &ocw, oap);
1575
1576                 rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1577
1578                 spin_lock(&cli->cl_loi_list_lock);
1579
1580                 /* l_wait_event is interrupted by signal, or timed out */
1581                 if (rc < 0) {
1582                         switch (rc) {
1583                         case -ETIMEDOUT:
1584                                 OSC_DUMP_GRANT(D_ERROR, cli,
1585                                                 "try to reserve %d.\n", bytes);
1586                                 osc_extent_tree_dump(D_ERROR, osc);
1587                                 rc = -EDQUOT;
1588                                 break;
1589                         case -EINTR:
1590                                 /* Ensures restartability - LU-3581 */
1591                                 rc = -ERESTARTSYS;
1592                                 break;
1593                         default:
1594                                 CDEBUG(D_CACHE, "%s: event for cache space @"
1595                                        " %p never arrived due to %d\n",
1596                                        cli->cl_import->imp_obd->obd_name,
1597                                        &ocw, rc);
1598                                 break;
1599                         }
1600                         list_del_init(&ocw.ocw_entry);
1601                         GOTO(out, rc);
1602                 }
1603
1604                 LASSERT(list_empty(&ocw.ocw_entry));
1605                 rc = ocw.ocw_rc;
1606
1607                 if (rc != -EDQUOT)
1608                         GOTO(out, rc);
1609                 if (osc_enter_cache_try(cli, oap, bytes, 0))
1610                         GOTO(out, rc = 0);
1611         }
1612         EXIT;
1613 out:
1614         spin_unlock(&cli->cl_loi_list_lock);
1615         OSC_DUMP_GRANT(D_CACHE, cli, "returned %d.\n", rc);
1616         RETURN(rc);
1617 }
1618
1619 /* caller must hold loi_list_lock */
1620 void osc_wake_cache_waiters(struct client_obd *cli)
1621 {
1622         struct list_head *l, *tmp;
1623         struct osc_cache_waiter *ocw;
1624
1625         ENTRY;
1626         list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
1627                 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
1628                 list_del_init(&ocw->ocw_entry);
1629
1630                 ocw->ocw_rc = -EDQUOT;
1631                 /* we can't dirty more */
1632                 if ((cli->cl_dirty_pages  >= cli->cl_dirty_max_pages) ||
1633                     (1 + atomic_long_read(&obd_dirty_pages) >
1634                      obd_max_dirty_pages)) {
1635                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
1636                                "osc max %ld, sys max %ld\n",
1637                                cli->cl_dirty_pages, cli->cl_dirty_max_pages,
1638                                obd_max_dirty_pages);
1639                         goto wakeup;
1640                 }
1641
1642                 ocw->ocw_rc = 0;
1643                 if (!osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
1644                         ocw->ocw_rc = -EDQUOT;
1645
1646 wakeup:
1647                 CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld, %d\n",
1648                        ocw, ocw->ocw_oap, cli->cl_avail_grant, ocw->ocw_rc);
1649
1650                 wake_up(&ocw->ocw_waitq);
1651         }
1652
1653         EXIT;
1654 }
1655
1656 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1657 {
1658         int hprpc = !!list_empty(&osc->oo_hp_exts);
1659         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1660 }
1661
1662 /* This maintains the lists of pending pages to read/write for a given object
1663  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1664  * to quickly find objects that are ready to send an RPC. */
1665 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1666                          int cmd)
1667 {
1668         int invalid_import = 0;
1669         ENTRY;
1670
1671         /* if we have an invalid import we want to drain the queued pages
1672          * by forcing them through rpcs that immediately fail and complete
1673          * the pages.  recovery relies on this to empty the queued pages
1674          * before canceling the locks and evicting down the llite pages */
1675         if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1676                 invalid_import = 1;
1677
1678         if (cmd & OBD_BRW_WRITE) {
1679                 if (atomic_read(&osc->oo_nr_writes) == 0)
1680                         RETURN(0);
1681                 if (invalid_import) {
1682                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1683                         RETURN(1);
1684                 }
1685                 if (!list_empty(&osc->oo_hp_exts)) {
1686                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1687                         RETURN(1);
1688                 }
1689                 if (!list_empty(&osc->oo_urgent_exts)) {
1690                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1691                         RETURN(1);
1692                 }
1693                 /* trigger a write rpc stream as long as there are dirtiers
1694                  * waiting for space.  as they're waiting, they're not going to
1695                  * create more pages to coalesce with what's waiting.. */
1696                 if (!list_empty(&cli->cl_cache_waiters)) {
1697                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1698                         RETURN(1);
1699                 }
1700                 if (atomic_read(&osc->oo_nr_writes) >=
1701                     cli->cl_max_pages_per_rpc)
1702                         RETURN(1);
1703         } else {
1704                 if (atomic_read(&osc->oo_nr_reads) == 0)
1705                         RETURN(0);
1706                 if (invalid_import) {
1707                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1708                         RETURN(1);
1709                 }
1710                 /* all read are urgent. */
1711                 if (!list_empty(&osc->oo_reading_exts))
1712                         RETURN(1);
1713         }
1714
1715         RETURN(0);
1716 }
1717
1718 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1719 {
1720         struct client_obd *cli = osc_cli(obj);
1721         if (cmd & OBD_BRW_WRITE) {
1722                 atomic_add(delta, &obj->oo_nr_writes);
1723                 atomic_add(delta, &cli->cl_pending_w_pages);
1724                 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1725         } else {
1726                 atomic_add(delta, &obj->oo_nr_reads);
1727                 atomic_add(delta, &cli->cl_pending_r_pages);
1728                 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1729         }
1730         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1731 }
1732
1733 static int osc_makes_hprpc(struct osc_object *obj)
1734 {
1735         return !list_empty(&obj->oo_hp_exts);
1736 }
1737
1738 static void on_list(struct list_head *item, struct list_head *list,
1739                     int should_be_on)
1740 {
1741         if (list_empty(item) && should_be_on)
1742                 list_add_tail(item, list);
1743         else if (!list_empty(item) && !should_be_on)
1744                 list_del_init(item);
1745 }
1746
1747 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1748  * can find pages to build into rpcs quickly */
1749 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1750 {
1751         if (osc_makes_hprpc(osc)) {
1752                 /* HP rpc */
1753                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1754                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1755         } else {
1756                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1757                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1758                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1759                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1760         }
1761
1762         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1763                 atomic_read(&osc->oo_nr_writes) > 0);
1764
1765         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1766                 atomic_read(&osc->oo_nr_reads) > 0);
1767
1768         return osc_is_ready(osc);
1769 }
1770
1771 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1772 {
1773         int is_ready;
1774
1775         spin_lock(&cli->cl_loi_list_lock);
1776         is_ready = __osc_list_maint(cli, osc);
1777         spin_unlock(&cli->cl_loi_list_lock);
1778
1779         return is_ready;
1780 }
1781
1782 /* this is trying to propogate async writeback errors back up to the
1783  * application.  As an async write fails we record the error code for later if
1784  * the app does an fsync.  As long as errors persist we force future rpcs to be
1785  * sync so that the app can get a sync error and break the cycle of queueing
1786  * pages for which writeback will fail. */
1787 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1788                            int rc)
1789 {
1790         if (rc) {
1791                 if (!ar->ar_rc)
1792                         ar->ar_rc = rc;
1793
1794                 ar->ar_force_sync = 1;
1795                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1796                 return;
1797
1798         }
1799
1800         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1801                 ar->ar_force_sync = 0;
1802 }
1803
1804 /* this must be called holding the loi list lock to give coverage to exit_cache,
1805  * async_flag maintenance, and oap_request */
1806 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1807                               struct osc_async_page *oap, int sent, int rc)
1808 {
1809         struct osc_object *osc = oap->oap_obj;
1810         struct lov_oinfo  *loi = osc->oo_oinfo;
1811         __u64 xid = 0;
1812
1813         ENTRY;
1814         if (oap->oap_request != NULL) {
1815                 xid = ptlrpc_req_xid(oap->oap_request);
1816                 ptlrpc_req_finished(oap->oap_request);
1817                 oap->oap_request = NULL;
1818         }
1819
1820         /* As the transfer for this page is being done, clear the flags */
1821         spin_lock(&oap->oap_lock);
1822         oap->oap_async_flags = 0;
1823         spin_unlock(&oap->oap_lock);
1824         oap->oap_interrupted = 0;
1825
1826         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1827                 spin_lock(&cli->cl_loi_list_lock);
1828                 osc_process_ar(&cli->cl_ar, xid, rc);
1829                 osc_process_ar(&loi->loi_ar, xid, rc);
1830                 spin_unlock(&cli->cl_loi_list_lock);
1831         }
1832
1833         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1834         if (rc)
1835                 CERROR("completion on oap %p obj %p returns %d.\n",
1836                        oap, osc, rc);
1837
1838         EXIT;
1839 }
1840
1841 /**
1842  * Try to add extent to one RPC. We need to think about the following things:
1843  * - # of pages must not be over max_pages_per_rpc
1844  * - extent must be compatible with previous ones
1845  */
1846 static int try_to_add_extent_for_io(struct client_obd *cli,
1847                                     struct osc_extent *ext,
1848                                     struct list_head *rpclist,
1849                                     int *pc, unsigned int *max_pages)
1850 {
1851         struct osc_extent *tmp;
1852         struct osc_async_page *oap = list_first_entry(&ext->oe_pages,
1853                                                       struct osc_async_page,
1854                                                       oap_pending_item);
1855         ENTRY;
1856
1857         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1858                 ext);
1859
1860         *max_pages = max(ext->oe_mppr, *max_pages);
1861         if (*pc + ext->oe_nr_pages > *max_pages)
1862                 RETURN(0);
1863
1864         list_for_each_entry(tmp, rpclist, oe_link) {
1865                 struct osc_async_page *oap2;
1866                 oap2 = list_first_entry(&tmp->oe_pages, struct osc_async_page,
1867                                         oap_pending_item);
1868                 EASSERT(tmp->oe_owner == current, tmp);
1869 #if 0
1870                 if (overlapped(tmp, ext)) {
1871                         OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
1872                         EASSERT(0, ext);
1873                 }
1874 #endif
1875                 if (oap2cl_page(oap)->cp_type != oap2cl_page(oap2)->cp_type) {
1876                         CDEBUG(D_CACHE, "Do not permit different type of IO"
1877                                         " for a same RPC\n");
1878                         RETURN(0);
1879                 }
1880
1881                 if (tmp->oe_srvlock != ext->oe_srvlock ||
1882                     !tmp->oe_grants != !ext->oe_grants)
1883                         RETURN(0);
1884
1885                 /* remove break for strict check */
1886                 break;
1887         }
1888
1889         *pc += ext->oe_nr_pages;
1890         list_move_tail(&ext->oe_link, rpclist);
1891         ext->oe_owner = current;
1892         RETURN(1);
1893 }
1894
1895 /**
1896  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1897  * get_write_extent() takes all appropriate extents in atomic.
1898  *
1899  * The following policy is used to collect extents for IO:
1900  * 1. Add as many HP extents as possible;
1901  * 2. Add the first urgent extent in urgent extent list and take it out of
1902  *    urgent list;
1903  * 3. Add subsequent extents of this urgent extent;
1904  * 4. If urgent list is not empty, goto 2;
1905  * 5. Traverse the extent tree from the 1st extent;
1906  * 6. Above steps exit if there is no space in this RPC.
1907  */
1908 static int get_write_extents(struct osc_object *obj, struct list_head *rpclist)
1909 {
1910         struct client_obd *cli = osc_cli(obj);
1911         struct osc_extent *ext;
1912         int page_count = 0;
1913         unsigned int max_pages = cli->cl_max_pages_per_rpc;
1914
1915         LASSERT(osc_object_is_locked(obj));
1916         while (!list_empty(&obj->oo_hp_exts)) {
1917                 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
1918                                  oe_link);
1919                 LASSERT(ext->oe_state == OES_CACHE);
1920                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1921                                               &max_pages))
1922                         return page_count;
1923                 EASSERT(ext->oe_nr_pages <= max_pages, ext);
1924         }
1925         if (page_count == max_pages)
1926                 return page_count;
1927
1928         while (!list_empty(&obj->oo_urgent_exts)) {
1929                 ext = list_entry(obj->oo_urgent_exts.next,
1930                                  struct osc_extent, oe_link);
1931                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1932                                               &max_pages))
1933                         return page_count;
1934
1935                 if (!ext->oe_intree)
1936                         continue;
1937
1938                 while ((ext = next_extent(ext)) != NULL) {
1939                         if ((ext->oe_state != OES_CACHE) ||
1940                             (!list_empty(&ext->oe_link) &&
1941                              ext->oe_owner != NULL))
1942                                 continue;
1943
1944                         if (!try_to_add_extent_for_io(cli, ext, rpclist,
1945                                                       &page_count, &max_pages))
1946                                 return page_count;
1947                 }
1948         }
1949         if (page_count == max_pages)
1950                 return page_count;
1951
1952         ext = first_extent(obj);
1953         while (ext != NULL) {
1954                 if ((ext->oe_state != OES_CACHE) ||
1955                     /* this extent may be already in current rpclist */
1956                     (!list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
1957                         ext = next_extent(ext);
1958                         continue;
1959                 }
1960
1961                 if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
1962                                               &max_pages))
1963                         return page_count;
1964
1965                 ext = next_extent(ext);
1966         }
1967         return page_count;
1968 }
1969
1970 static int
1971 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
1972                    struct osc_object *osc, pdl_policy_t pol)
1973 __must_hold(osc)
1974 {
1975         struct list_head   rpclist = LIST_HEAD_INIT(rpclist);
1976         struct osc_extent *ext;
1977         struct osc_extent *tmp;
1978         struct osc_extent *first = NULL;
1979         obd_count page_count = 0;
1980         int srvlock = 0;
1981         int rc = 0;
1982         ENTRY;
1983
1984         LASSERT(osc_object_is_locked(osc));
1985
1986         page_count = get_write_extents(osc, &rpclist);
1987         LASSERT(equi(page_count == 0, list_empty(&rpclist)));
1988
1989         if (list_empty(&rpclist))
1990                 RETURN(0);
1991
1992         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
1993
1994         list_for_each_entry(ext, &rpclist, oe_link) {
1995                 LASSERT(ext->oe_state == OES_CACHE ||
1996                         ext->oe_state == OES_LOCK_DONE);
1997                 if (ext->oe_state == OES_CACHE)
1998                         osc_extent_state_set(ext, OES_LOCKING);
1999                 else
2000                         osc_extent_state_set(ext, OES_RPC);
2001         }
2002
2003         /* we're going to grab page lock, so release object lock because
2004          * lock order is page lock -> object lock. */
2005         osc_object_unlock(osc);
2006
2007         list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2008                 if (ext->oe_state == OES_LOCKING) {
2009                         rc = osc_extent_make_ready(env, ext);
2010                         if (unlikely(rc < 0)) {
2011                                 list_del_init(&ext->oe_link);
2012                                 osc_extent_finish(env, ext, 0, rc);
2013                                 continue;
2014                         }
2015                 }
2016                 if (first == NULL) {
2017                         first = ext;
2018                         srvlock = ext->oe_srvlock;
2019                 } else {
2020                         LASSERT(srvlock == ext->oe_srvlock);
2021                 }
2022         }
2023
2024         if (!list_empty(&rpclist)) {
2025                 LASSERT(page_count > 0);
2026                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE, pol);
2027                 LASSERT(list_empty(&rpclist));
2028         }
2029
2030         osc_object_lock(osc);
2031         RETURN(rc);
2032 }
2033
2034 /**
2035  * prepare pages for ASYNC io and put pages in send queue.
2036  *
2037  * \param cmd OBD_BRW_* macroses
2038  * \param lop pending pages
2039  *
2040  * \return zero if no page added to send queue.
2041  * \return 1 if pages successfully added to send queue.
2042  * \return negative on errors.
2043  */
2044 static int
2045 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2046                   struct osc_object *osc, pdl_policy_t pol)
2047 __must_hold(osc)
2048 {
2049         struct osc_extent *ext;
2050         struct osc_extent *next;
2051         struct list_head rpclist = LIST_HEAD_INIT(rpclist);
2052         int page_count = 0;
2053         unsigned int max_pages = cli->cl_max_pages_per_rpc;
2054         int rc = 0;
2055         ENTRY;
2056
2057         LASSERT(osc_object_is_locked(osc));
2058         list_for_each_entry_safe(ext, next,
2059                                      &osc->oo_reading_exts, oe_link) {
2060                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2061                 if (!try_to_add_extent_for_io(cli, ext, &rpclist, &page_count,
2062                                               &max_pages))
2063                         break;
2064                 osc_extent_state_set(ext, OES_RPC);
2065                 EASSERT(ext->oe_nr_pages <= max_pages, ext);
2066         }
2067         LASSERT(page_count <= max_pages);
2068
2069         osc_update_pending(osc, OBD_BRW_READ, -page_count);
2070
2071         if (!list_empty(&rpclist)) {
2072                 osc_object_unlock(osc);
2073
2074                 LASSERT(page_count > 0);
2075                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ, pol);
2076                 LASSERT(list_empty(&rpclist));
2077
2078                 osc_object_lock(osc);
2079         }
2080         RETURN(rc);
2081 }
2082
2083 #define list_to_obj(list, item) ({                                            \
2084         struct list_head *__tmp = (list)->next;                               \
2085         list_del_init(__tmp);                                         \
2086         list_entry(__tmp, struct osc_object, oo_##item);                      \
2087 })
2088
2089 /* This is called by osc_check_rpcs() to find which objects have pages that
2090  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2091 static struct osc_object *osc_next_obj(struct client_obd *cli)
2092 {
2093         ENTRY;
2094
2095         /* First return objects that have blocked locks so that they
2096          * will be flushed quickly and other clients can get the lock,
2097          * then objects which have pages ready to be stuffed into RPCs */
2098         if (!list_empty(&cli->cl_loi_hp_ready_list))
2099                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2100         if (!list_empty(&cli->cl_loi_ready_list))
2101                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2102
2103         /* then if we have cache waiters, return all objects with queued
2104          * writes.  This is especially important when many small files
2105          * have filled up the cache and not been fired into rpcs because
2106          * they don't pass the nr_pending/object threshhold */
2107         if (!list_empty(&cli->cl_cache_waiters) &&
2108             !list_empty(&cli->cl_loi_write_list))
2109                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2110
2111         /* then return all queued objects when we have an invalid import
2112          * so that they get flushed */
2113         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2114                 if (!list_empty(&cli->cl_loi_write_list))
2115                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2116                                            write_item));
2117                 if (!list_empty(&cli->cl_loi_read_list))
2118                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2119                                            read_item));
2120         }
2121         RETURN(NULL);
2122 }
2123
2124 /* called with the loi list lock held */
2125 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli,
2126                            pdl_policy_t pol)
2127 __must_hold(&cli->cl_loi_list_lock)
2128 {
2129         struct osc_object *osc;
2130         int rc = 0;
2131         ENTRY;
2132
2133         while ((osc = osc_next_obj(cli)) != NULL) {
2134                 struct cl_object *obj = osc2cl(osc);
2135                 struct lu_ref_link link;
2136
2137                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2138
2139                 if (osc_max_rpc_in_flight(cli, osc)) {
2140                         __osc_list_maint(cli, osc);
2141                         break;
2142                 }
2143
2144                 cl_object_get(obj);
2145                 spin_unlock(&cli->cl_loi_list_lock);
2146                 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2147
2148                 /* attempt some read/write balancing by alternating between
2149                  * reads and writes in an object.  The makes_rpc checks here
2150                  * would be redundant if we were getting read/write work items
2151                  * instead of objects.  we don't want send_oap_rpc to drain a
2152                  * partial read pending queue when we're given this object to
2153                  * do io on writes while there are cache waiters */
2154                 osc_object_lock(osc);
2155                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2156                         rc = osc_send_write_rpc(env, cli, osc, pol);
2157                         if (rc < 0) {
2158                                 CERROR("Write request failed with %d\n", rc);
2159
2160                                 /* osc_send_write_rpc failed, mostly because of
2161                                  * memory pressure.
2162                                  *
2163                                  * It can't break here, because if:
2164                                  *  - a page was submitted by osc_io_submit, so
2165                                  *    page locked;
2166                                  *  - no request in flight
2167                                  *  - no subsequent request
2168                                  * The system will be in live-lock state,
2169                                  * because there is no chance to call
2170                                  * osc_io_unplug() and osc_check_rpcs() any
2171                                  * more. pdflush can't help in this case,
2172                                  * because it might be blocked at grabbing
2173                                  * the page lock as we mentioned.
2174                                  *
2175                                  * Anyway, continue to drain pages. */
2176                                 /* break; */
2177                         }
2178                 }
2179                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2180                         rc = osc_send_read_rpc(env, cli, osc, pol);
2181                         if (rc < 0)
2182                                 CERROR("Read request failed with %d\n", rc);
2183                 }
2184                 osc_object_unlock(osc);
2185
2186                 osc_list_maint(cli, osc);
2187                 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2188                 cl_object_put(env, obj);
2189
2190                 spin_lock(&cli->cl_loi_list_lock);
2191         }
2192 }
2193
2194 static int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2195                           struct osc_object *osc, pdl_policy_t pol, int async)
2196 {
2197         int rc = 0;
2198
2199         if (osc != NULL && osc_list_maint(cli, osc) == 0)
2200                 return 0;
2201
2202         if (!async) {
2203                 /* disable osc_lru_shrink() temporarily to avoid
2204                  * potential stack overrun problem. LU-2859 */
2205                 atomic_inc(&cli->cl_lru_shrinkers);
2206                 spin_lock(&cli->cl_loi_list_lock);
2207                 osc_check_rpcs(env, cli, pol);
2208                 spin_unlock(&cli->cl_loi_list_lock);
2209                 atomic_dec(&cli->cl_lru_shrinkers);
2210         } else {
2211                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2212                 LASSERT(cli->cl_writeback_work != NULL);
2213                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2214         }
2215         return rc;
2216 }
2217
2218 static int osc_io_unplug_async(const struct lu_env *env,
2219                                 struct client_obd *cli, struct osc_object *osc)
2220 {
2221         /* XXX: policy is no use actually. */
2222         return osc_io_unplug0(env, cli, osc, PDL_POLICY_ROUND, 1);
2223 }
2224
2225 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
2226                    struct osc_object *osc, pdl_policy_t pol)
2227 {
2228         (void)osc_io_unplug0(env, cli, osc, pol, 0);
2229 }
2230
2231 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2232                         struct page *page, loff_t offset)
2233 {
2234         struct obd_export     *exp = osc_export(osc);
2235         struct osc_async_page *oap = &ops->ops_oap;
2236         ENTRY;
2237
2238         if (!page)
2239                 return cfs_size_round(sizeof(*oap));
2240
2241         oap->oap_magic = OAP_MAGIC;
2242         oap->oap_cli = &exp->exp_obd->u.cli;
2243         oap->oap_obj = osc;
2244
2245         oap->oap_page = page;
2246         oap->oap_obj_off = offset;
2247         LASSERT(!(offset & ~CFS_PAGE_MASK));
2248
2249         if (!client_is_remote(exp) && cfs_capable(CFS_CAP_SYS_RESOURCE))
2250                 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2251
2252         INIT_LIST_HEAD(&oap->oap_pending_item);
2253         INIT_LIST_HEAD(&oap->oap_rpc_item);
2254
2255         spin_lock_init(&oap->oap_lock);
2256         CDEBUG(D_INFO, "oap %p page %p obj off "LPU64"\n",
2257                oap, page, oap->oap_obj_off);
2258         RETURN(0);
2259 }
2260
2261 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2262                        struct osc_page *ops)
2263 {
2264         struct osc_io *oio = osc_env_io(env);
2265         struct osc_extent     *ext = NULL;
2266         struct osc_async_page *oap = &ops->ops_oap;
2267         struct client_obd     *cli = oap->oap_cli;
2268         struct osc_object     *osc = oap->oap_obj;
2269         pgoff_t index;
2270         int    grants = 0;
2271         int    brw_flags = OBD_BRW_ASYNC;
2272         int    cmd = OBD_BRW_WRITE;
2273         int    need_release = 0;
2274         int    rc = 0;
2275         ENTRY;
2276
2277         if (oap->oap_magic != OAP_MAGIC)
2278                 RETURN(-EINVAL);
2279
2280         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2281                 RETURN(-EIO);
2282
2283         if (!list_empty(&oap->oap_pending_item) ||
2284             !list_empty(&oap->oap_rpc_item))
2285                 RETURN(-EBUSY);
2286
2287         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2288         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2289         if (!client_is_remote(osc_export(osc)) &&
2290             cfs_capable(CFS_CAP_SYS_RESOURCE)) {
2291                 brw_flags |= OBD_BRW_NOQUOTA;
2292                 cmd |= OBD_BRW_NOQUOTA;
2293         }
2294
2295         /* check if the file's owner/group is over quota */
2296         if (!(cmd & OBD_BRW_NOQUOTA)) {
2297                 struct cl_object *obj;
2298                 struct cl_attr   *attr;
2299                 unsigned int qid[MAXQUOTAS];
2300
2301                 obj = cl_object_top(&osc->oo_cl);
2302                 attr = &osc_env_info(env)->oti_attr;
2303
2304                 cl_object_attr_lock(obj);
2305                 rc = cl_object_attr_get(env, obj, attr);
2306                 cl_object_attr_unlock(obj);
2307
2308                 qid[USRQUOTA] = attr->cat_uid;
2309                 qid[GRPQUOTA] = attr->cat_gid;
2310                 if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
2311                         rc = -EDQUOT;
2312                 if (rc)
2313                         RETURN(rc);
2314         }
2315
2316         oap->oap_cmd = cmd;
2317         oap->oap_page_off = ops->ops_from;
2318         oap->oap_count = ops->ops_to - ops->ops_from;
2319         /* No need to hold a lock here,
2320          * since this page is not in any list yet. */
2321         oap->oap_async_flags = 0;
2322         oap->oap_brw_flags = brw_flags;
2323
2324         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2325                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2326
2327         index = osc_index(oap2osc(oap));
2328
2329         /* Add this page into extent by the following steps:
2330          * 1. if there exists an active extent for this IO, mostly this page
2331          *    can be added to the active extent and sometimes we need to
2332          *    expand extent to accomodate this page;
2333          * 2. otherwise, a new extent will be allocated. */
2334
2335         ext = oio->oi_active;
2336         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2337                 /* one chunk plus extent overhead must be enough to write this
2338                  * page */
2339                 grants = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2340                 if (ext->oe_end >= index)
2341                         grants = 0;
2342
2343                 /* it doesn't need any grant to dirty this page */
2344                 spin_lock(&cli->cl_loi_list_lock);
2345                 rc = osc_enter_cache_try(cli, oap, grants, 0);
2346                 spin_unlock(&cli->cl_loi_list_lock);
2347                 if (rc == 0) { /* try failed */
2348                         grants = 0;
2349                         need_release = 1;
2350                 } else if (ext->oe_end < index) {
2351                         int tmp = grants;
2352                         /* try to expand this extent */
2353                         rc = osc_extent_expand(ext, index, &tmp);
2354                         if (rc < 0) {
2355                                 need_release = 1;
2356                                 /* don't free reserved grant */
2357                         } else {
2358                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2359                                                 "expanded for %lu.\n", index);
2360                                 osc_unreserve_grant(cli, grants, tmp);
2361                                 grants = 0;
2362                         }
2363                 }
2364                 rc = 0;
2365         } else if (ext != NULL) {
2366                 /* index is located outside of active extent */
2367                 need_release = 1;
2368         }
2369         if (need_release) {
2370                 osc_extent_release(env, ext);
2371                 oio->oi_active = NULL;
2372                 ext = NULL;
2373         }
2374
2375         if (ext == NULL) {
2376                 int tmp = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
2377
2378                 /* try to find new extent to cover this page */
2379                 LASSERT(oio->oi_active == NULL);
2380                 /* we may have allocated grant for this page if we failed
2381                  * to expand the previous active extent. */
2382                 LASSERT(ergo(grants > 0, grants >= tmp));
2383
2384                 rc = 0;
2385                 if (grants == 0) {
2386                         /* we haven't allocated grant for this page. */
2387                         rc = osc_enter_cache(env, cli, oap, tmp);
2388                         if (rc == 0)
2389                                 grants = tmp;
2390                 }
2391
2392                 tmp = grants;
2393                 if (rc == 0) {
2394                         ext = osc_extent_find(env, osc, index, &tmp);
2395                         if (IS_ERR(ext)) {
2396                                 LASSERT(tmp == grants);
2397                                 osc_exit_cache(cli, oap);
2398                                 rc = PTR_ERR(ext);
2399                                 ext = NULL;
2400                         } else {
2401                                 oio->oi_active = ext;
2402                         }
2403                 }
2404                 if (grants > 0)
2405                         osc_unreserve_grant(cli, grants, tmp);
2406         }
2407
2408         LASSERT(ergo(rc == 0, ext != NULL));
2409         if (ext != NULL) {
2410                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2411                          ext, "index = %lu.\n", index);
2412                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2413
2414                 osc_object_lock(osc);
2415                 if (ext->oe_nr_pages == 0)
2416                         ext->oe_srvlock = ops->ops_srvlock;
2417                 else
2418                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2419                 ++ext->oe_nr_pages;
2420                 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2421                 osc_object_unlock(osc);
2422         }
2423         RETURN(rc);
2424 }
2425
2426 int osc_teardown_async_page(const struct lu_env *env,
2427                             struct osc_object *obj, struct osc_page *ops)
2428 {
2429         struct osc_async_page *oap = &ops->ops_oap;
2430         struct osc_extent     *ext = NULL;
2431         int rc = 0;
2432         ENTRY;
2433
2434         LASSERT(oap->oap_magic == OAP_MAGIC);
2435
2436         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2437                oap, ops, osc_index(oap2osc(oap)));
2438
2439         osc_object_lock(obj);
2440         if (!list_empty(&oap->oap_rpc_item)) {
2441                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2442                 rc = -EBUSY;
2443         } else if (!list_empty(&oap->oap_pending_item)) {
2444                 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2445                 /* only truncated pages are allowed to be taken out.
2446                  * See osc_extent_truncate() and osc_cache_truncate_start()
2447                  * for details. */
2448                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2449                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2450                                         osc_index(oap2osc(oap)));
2451                         rc = -EBUSY;
2452                 }
2453         }
2454         osc_object_unlock(obj);
2455         if (ext != NULL)
2456                 osc_extent_put(env, ext);
2457         RETURN(rc);
2458 }
2459
2460 /**
2461  * This is called when a page is picked up by kernel to write out.
2462  *
2463  * We should find out the corresponding extent and add the whole extent
2464  * into urgent list. The extent may be being truncated or used, handle it
2465  * carefully.
2466  */
2467 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2468                          struct osc_page *ops)
2469 {
2470         struct osc_extent *ext   = NULL;
2471         struct osc_object *obj   = cl2osc(ops->ops_cl.cpl_obj);
2472         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2473         pgoff_t            index = osc_index(ops);
2474         struct osc_async_page *oap = &ops->ops_oap;
2475         bool unplug = false;
2476         int rc = 0;
2477         ENTRY;
2478
2479         osc_object_lock(obj);
2480         ext = osc_extent_lookup(obj, index);
2481         if (ext == NULL) {
2482                 osc_extent_tree_dump(D_ERROR, obj);
2483                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2484         }
2485
2486         switch (ext->oe_state) {
2487         case OES_RPC:
2488         case OES_LOCK_DONE:
2489                 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2490                 LASSERT(0);
2491                 break;
2492         case OES_LOCKING:
2493                 /* If we know this extent is being written out, we should abort
2494                  * so that the writer can make this page ready. Otherwise, there
2495                  * exists a deadlock problem because other process can wait for
2496                  * page writeback bit holding page lock; and meanwhile in
2497                  * vvp_page_make_ready(), we need to grab page lock before
2498                  * really sending the RPC. */
2499         case OES_TRUNC:
2500                 /* race with truncate, page will be redirtied */
2501         case OES_ACTIVE:
2502                 /* The extent is active so we need to abort and let the caller
2503                  * re-dirty the page. If we continued on here, and we were the
2504                  * one making the extent active, we could deadlock waiting for
2505                  * the page writeback to clear but it won't because the extent
2506                  * is active and won't be written out. */
2507                 GOTO(out, rc = -EAGAIN);
2508         default:
2509                 break;
2510         }
2511
2512         rc = cl_page_prep(env, io, cp, CRT_WRITE);
2513         if (rc)
2514                 GOTO(out, rc);
2515
2516         spin_lock(&oap->oap_lock);
2517         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2518         spin_unlock(&oap->oap_lock);
2519
2520         if (memory_pressure_get())
2521                 ext->oe_memalloc = 1;
2522
2523         ext->oe_urgent = 1;
2524         if (ext->oe_state == OES_CACHE) {
2525                 OSC_EXTENT_DUMP(D_CACHE, ext,
2526                                 "flush page %p make it urgent.\n", oap);
2527                 if (list_empty(&ext->oe_link))
2528                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2529                 unplug = true;
2530         }
2531         rc = 0;
2532         EXIT;
2533
2534 out:
2535         osc_object_unlock(obj);
2536         osc_extent_put(env, ext);
2537         if (unplug)
2538                 osc_io_unplug_async(env, osc_cli(obj), obj);
2539         return rc;
2540 }
2541
2542 /**
2543  * this is called when a sync waiter receives an interruption.  Its job is to
2544  * get the caller woken as soon as possible.  If its page hasn't been put in an
2545  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2546  * desiring interruption which will forcefully complete the rpc once the rpc
2547  * has timed out.
2548  */
2549 int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
2550 {
2551         struct osc_async_page *oap = &ops->ops_oap;
2552         struct osc_object     *obj = oap->oap_obj;
2553         struct client_obd     *cli = osc_cli(obj);
2554         struct osc_extent     *ext;
2555         struct osc_extent     *found = NULL;
2556         struct list_head            *plist;
2557         pgoff_t index = osc_index(ops);
2558         int     rc = -EBUSY;
2559         int     cmd;
2560         ENTRY;
2561
2562         LASSERT(!oap->oap_interrupted);
2563         oap->oap_interrupted = 1;
2564
2565         /* Find out the caching extent */
2566         osc_object_lock(obj);
2567         if (oap->oap_cmd & OBD_BRW_WRITE) {
2568                 plist = &obj->oo_urgent_exts;
2569                 cmd   = OBD_BRW_WRITE;
2570         } else {
2571                 plist = &obj->oo_reading_exts;
2572                 cmd   = OBD_BRW_READ;
2573         }
2574         list_for_each_entry(ext, plist, oe_link) {
2575                 if (ext->oe_start <= index && ext->oe_end >= index) {
2576                         LASSERT(ext->oe_state == OES_LOCK_DONE);
2577                         /* For OES_LOCK_DONE state extent, it has already held
2578                          * a refcount for RPC. */
2579                         found = osc_extent_get(ext);
2580                         break;
2581                 }
2582         }
2583         if (found != NULL) {
2584                 list_del_init(&found->oe_link);
2585                 osc_update_pending(obj, cmd, -found->oe_nr_pages);
2586                 osc_object_unlock(obj);
2587
2588                 osc_extent_finish(env, found, 0, -EINTR);
2589                 osc_extent_put(env, found);
2590                 rc = 0;
2591         } else {
2592                 osc_object_unlock(obj);
2593                 /* ok, it's been put in an rpc. only one oap gets a request
2594                  * reference */
2595                 if (oap->oap_request != NULL) {
2596                         ptlrpc_mark_interrupted(oap->oap_request);
2597                         ptlrpcd_wake(oap->oap_request);
2598                         ptlrpc_req_finished(oap->oap_request);
2599                         oap->oap_request = NULL;
2600                 }
2601         }
2602
2603         osc_list_maint(cli, obj);
2604         RETURN(rc);
2605 }
2606
2607 int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
2608                          struct list_head *list, int cmd, int brw_flags)
2609 {
2610         struct client_obd     *cli = osc_cli(obj);
2611         struct osc_extent     *ext;
2612         struct osc_async_page *oap;
2613         int     page_count = 0;
2614         int     mppr       = cli->cl_max_pages_per_rpc;
2615         pgoff_t start      = CL_PAGE_EOF;
2616         pgoff_t end        = 0;
2617         ENTRY;
2618
2619         list_for_each_entry(oap, list, oap_pending_item) {
2620                 pgoff_t index = osc_index(oap2osc(oap));
2621                 if (index > end)
2622                         end = index;
2623                 if (index < start)
2624                         start = index;
2625                 ++page_count;
2626                 mppr <<= (page_count > mppr);
2627         }
2628
2629         ext = osc_extent_alloc(obj);
2630         if (ext == NULL) {
2631                 list_for_each_entry(oap, list, oap_pending_item) {
2632                         list_del_init(&oap->oap_pending_item);
2633                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2634                 }
2635                 RETURN(-ENOMEM);
2636         }
2637
2638         ext->oe_rw = !!(cmd & OBD_BRW_READ);
2639         ext->oe_sync = 1;
2640         ext->oe_urgent = 1;
2641         ext->oe_start = start;
2642         ext->oe_end = ext->oe_max_end = end;
2643         ext->oe_obj = obj;
2644         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2645         ext->oe_nr_pages = page_count;
2646         ext->oe_mppr = mppr;
2647         list_splice_init(list, &ext->oe_pages);
2648
2649         osc_object_lock(obj);
2650         /* Reuse the initial refcount for RPC, don't drop it */
2651         osc_extent_state_set(ext, OES_LOCK_DONE);
2652         if (cmd & OBD_BRW_WRITE) {
2653                 list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2654                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2655         } else {
2656                 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2657                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2658         }
2659         osc_object_unlock(obj);
2660
2661         osc_io_unplug_async(env, cli, obj);
2662         RETURN(0);
2663 }
2664
2665 /**
2666  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2667  */
2668 int osc_cache_truncate_start(const struct lu_env *env, struct osc_io *oio,
2669                              struct osc_object *obj, __u64 size)
2670 {
2671         struct client_obd *cli = osc_cli(obj);
2672         struct osc_extent *ext;
2673         struct osc_extent *waiting = NULL;
2674         pgoff_t index;
2675         struct list_head list = LIST_HEAD_INIT(list);
2676         int result = 0;
2677         bool partial;
2678         ENTRY;
2679
2680         /* pages with index greater or equal to index will be truncated. */
2681         index = cl_index(osc2cl(obj), size);
2682         partial = size > cl_offset(osc2cl(obj), index);
2683
2684 again:
2685         osc_object_lock(obj);
2686         ext = osc_extent_search(obj, index);
2687         if (ext == NULL)
2688                 ext = first_extent(obj);
2689         else if (ext->oe_end < index)
2690                 ext = next_extent(ext);
2691         while (ext != NULL) {
2692                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2693
2694                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2695                         /* if ext is in urgent state, it means there must exist
2696                          * a page already having been flushed by write_page().
2697                          * We have to wait for this extent because we can't
2698                          * truncate that page. */
2699                         LASSERT(!ext->oe_hp);
2700                         OSC_EXTENT_DUMP(D_CACHE, ext,
2701                                         "waiting for busy extent\n");
2702                         waiting = osc_extent_get(ext);
2703                         break;
2704                 }
2705
2706                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:"LPU64".\n", size);
2707
2708                 osc_extent_get(ext);
2709                 if (ext->oe_state == OES_ACTIVE) {
2710                         /* though we grab inode mutex for write path, but we
2711                          * release it before releasing extent(in osc_io_end()),
2712                          * so there is a race window that an extent is still
2713                          * in OES_ACTIVE when truncate starts. */
2714                         LASSERT(!ext->oe_trunc_pending);
2715                         ext->oe_trunc_pending = 1;
2716                 } else {
2717                         EASSERT(ext->oe_state == OES_CACHE, ext);
2718                         osc_extent_state_set(ext, OES_TRUNC);
2719                         osc_update_pending(obj, OBD_BRW_WRITE,
2720                                            -ext->oe_nr_pages);
2721                 }
2722                 EASSERT(list_empty(&ext->oe_link), ext);
2723                 list_add_tail(&ext->oe_link, &list);
2724
2725                 ext = next_extent(ext);
2726         }
2727         osc_object_unlock(obj);
2728
2729         osc_list_maint(cli, obj);
2730
2731         while (!list_empty(&list)) {
2732                 int rc;
2733
2734                 ext = list_entry(list.next, struct osc_extent, oe_link);
2735                 list_del_init(&ext->oe_link);
2736
2737                 /* extent may be in OES_ACTIVE state because inode mutex
2738                  * is released before osc_io_end() in file write case */
2739                 if (ext->oe_state != OES_TRUNC)
2740                         osc_extent_wait(env, ext, OES_TRUNC);
2741
2742                 rc = osc_extent_truncate(ext, index, partial);
2743                 if (rc < 0) {
2744                         if (result == 0)
2745                                 result = rc;
2746
2747                         OSC_EXTENT_DUMP(D_ERROR, ext,
2748                                         "truncate error %d\n", rc);
2749                 } else if (ext->oe_nr_pages == 0) {
2750                         osc_extent_remove(ext);
2751                 } else {
2752                         /* this must be an overlapped extent which means only
2753                          * part of pages in this extent have been truncated.
2754                          */
2755                         EASSERTF(ext->oe_start <= index, ext,
2756                                  "trunc index = %lu/%d.\n", index, partial);
2757                         /* fix index to skip this partially truncated extent */
2758                         index = ext->oe_end + 1;
2759                         partial = false;
2760
2761                         /* we need to hold this extent in OES_TRUNC state so
2762                          * that no writeback will happen. This is to avoid
2763                          * BUG 17397. */
2764                         LASSERT(oio->oi_trunc == NULL);
2765                         oio->oi_trunc = osc_extent_get(ext);
2766                         OSC_EXTENT_DUMP(D_CACHE, ext,
2767                                         "trunc at "LPU64"\n", size);
2768                 }
2769                 osc_extent_put(env, ext);
2770         }
2771         if (waiting != NULL) {
2772                 int rc;
2773
2774                 /* ignore the result of osc_extent_wait the write initiator
2775                  * should take care of it. */
2776                 rc = osc_extent_wait(env, waiting, OES_INV);
2777                 if (rc < 0)
2778                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2779
2780                 osc_extent_put(env, waiting);
2781                 waiting = NULL;
2782                 goto again;
2783         }
2784         RETURN(result);
2785 }
2786
2787 /**
2788  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2789  */
2790 void osc_cache_truncate_end(const struct lu_env *env, struct osc_io *oio,
2791                             struct osc_object *obj)
2792 {
2793         struct osc_extent *ext = oio->oi_trunc;
2794
2795         oio->oi_trunc = NULL;
2796         if (ext != NULL) {
2797                 bool unplug = false;
2798
2799                 EASSERT(ext->oe_nr_pages > 0, ext);
2800                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2801                 EASSERT(!ext->oe_urgent, ext);
2802
2803                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2804                 osc_object_lock(obj);
2805                 osc_extent_state_set(ext, OES_CACHE);
2806                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2807                         ext->oe_urgent = 1;
2808                         list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2809                         unplug = true;
2810                 }
2811                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2812                 osc_object_unlock(obj);
2813                 osc_extent_put(env, ext);
2814
2815                 if (unplug)
2816                         osc_io_unplug_async(env, osc_cli(obj), obj);
2817         }
2818 }
2819
2820 /**
2821  * Wait for extents in a specific range to be written out.
2822  * The caller must have called osc_cache_writeback_range() to issue IO
2823  * otherwise it will take a long time for this function to finish.
2824  *
2825  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2826  * nobody else can dirty this range of file while we're waiting for
2827  * extents to be written.
2828  */
2829 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2830                          pgoff_t start, pgoff_t end)
2831 {
2832         struct osc_extent *ext;
2833         pgoff_t index = start;
2834         int     result = 0;
2835         ENTRY;
2836
2837 again:
2838         osc_object_lock(obj);
2839         ext = osc_extent_search(obj, index);
2840         if (ext == NULL)
2841                 ext = first_extent(obj);
2842         else if (ext->oe_end < index)
2843                 ext = next_extent(ext);
2844         while (ext != NULL) {
2845                 int rc;
2846
2847                 if (ext->oe_start > end)
2848                         break;
2849
2850                 if (!ext->oe_fsync_wait) {
2851                         ext = next_extent(ext);
2852                         continue;
2853                 }
2854
2855                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2856                              ext->oe_hp || ext->oe_urgent), ext);
2857                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2858                              !ext->oe_hp && ext->oe_urgent), ext);
2859
2860                 index = ext->oe_end + 1;
2861                 osc_extent_get(ext);
2862                 osc_object_unlock(obj);
2863
2864                 rc = osc_extent_wait(env, ext, OES_INV);
2865                 if (result == 0)
2866                         result = rc;
2867                 osc_extent_put(env, ext);
2868                 goto again;
2869         }
2870         osc_object_unlock(obj);
2871
2872         OSC_IO_DEBUG(obj, "sync file range.\n");
2873         RETURN(result);
2874 }
2875
2876 /**
2877  * Called to write out a range of osc object.
2878  *
2879  * @hp     : should be set this is caused by lock cancel;
2880  * @discard: is set if dirty pages should be dropped - file will be deleted or
2881  *         truncated, this implies there is no partially discarding extents.
2882  *
2883  * Return how many pages will be issued, or error code if error occurred.
2884  */
2885 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2886                               pgoff_t start, pgoff_t end, int hp, int discard)
2887 {
2888         struct osc_extent *ext;
2889         struct list_head discard_list = LIST_HEAD_INIT(discard_list);
2890         bool unplug = false;
2891         int result = 0;
2892         ENTRY;
2893
2894         osc_object_lock(obj);
2895         ext = osc_extent_search(obj, start);
2896         if (ext == NULL)
2897                 ext = first_extent(obj);
2898         else if (ext->oe_end < start)
2899                 ext = next_extent(ext);
2900         while (ext != NULL) {
2901                 if (ext->oe_start > end)
2902                         break;
2903
2904                 ext->oe_fsync_wait = 1;
2905                 switch (ext->oe_state) {
2906                 case OES_CACHE:
2907                         result += ext->oe_nr_pages;
2908                         if (!discard) {
2909                                 struct list_head *list = NULL;
2910                                 if (hp) {
2911                                         EASSERT(!ext->oe_hp, ext);
2912                                         ext->oe_hp = 1;
2913                                         list = &obj->oo_hp_exts;
2914                                 } else if (!ext->oe_urgent) {
2915                                         ext->oe_urgent = 1;
2916                                         list = &obj->oo_urgent_exts;
2917                                 }
2918                                 if (list != NULL)
2919                                         list_move_tail(&ext->oe_link, list);
2920                                 unplug = true;
2921                         } else {
2922                                 /* the only discarder is lock cancelling, so
2923                                  * [start, end] must contain this extent */
2924                                 EASSERT(ext->oe_start >= start &&
2925                                         ext->oe_max_end <= end, ext);
2926                                 osc_extent_state_set(ext, OES_LOCKING);
2927                                 ext->oe_owner = current;
2928                                 list_move_tail(&ext->oe_link,
2929                                                    &discard_list);
2930                                 osc_update_pending(obj, OBD_BRW_WRITE,
2931                                                    -ext->oe_nr_pages);
2932                         }
2933                         break;
2934                 case OES_ACTIVE:
2935                         /* It's pretty bad to wait for ACTIVE extents, because
2936                          * we don't know how long we will wait for it to be
2937                          * flushed since it may be blocked at awaiting more
2938                          * grants. We do this for the correctness of fsync. */
2939                         LASSERT(hp == 0 && discard == 0);
2940                         ext->oe_urgent = 1;
2941                         break;
2942                 case OES_TRUNC:
2943                         /* this extent is being truncated, can't do anything
2944                          * for it now. it will be set to urgent after truncate
2945                          * is finished in osc_cache_truncate_end(). */
2946                 default:
2947                         break;
2948                 }
2949                 ext = next_extent(ext);
2950         }
2951         osc_object_unlock(obj);
2952
2953         LASSERT(ergo(!discard, list_empty(&discard_list)));
2954         if (!list_empty(&discard_list)) {
2955                 struct osc_extent *tmp;
2956                 int rc;
2957
2958                 osc_list_maint(osc_cli(obj), obj);
2959                 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
2960                         list_del_init(&ext->oe_link);
2961                         EASSERT(ext->oe_state == OES_LOCKING, ext);
2962
2963                         /* Discard caching pages. We don't actually write this
2964                          * extent out but we complete it as if we did. */
2965                         rc = osc_extent_make_ready(env, ext);
2966                         if (unlikely(rc < 0)) {
2967                                 OSC_EXTENT_DUMP(D_ERROR, ext,
2968                                                 "make_ready returned %d\n", rc);
2969                                 if (result >= 0)
2970                                         result = rc;
2971                         }
2972
2973                         /* finish the extent as if the pages were sent */
2974                         osc_extent_finish(env, ext, 0, 0);
2975                 }
2976         }
2977
2978         if (unplug)
2979                 osc_io_unplug(env, osc_cli(obj), obj, PDL_POLICY_ROUND);
2980
2981         if (hp || discard) {
2982                 int rc;
2983                 rc = osc_cache_wait_range(env, obj, start, end);
2984                 if (result >= 0 && rc < 0)
2985                         result = rc;
2986         }
2987
2988         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
2989         RETURN(result);
2990 }
2991
2992 /**
2993  * Returns a list of pages by a given [start, end] of \a obj.
2994  *
2995  * \param resched If not NULL, then we give up before hogging CPU for too
2996  * long and set *resched = 1, in that case caller should implement a retry
2997  * logic.
2998  *
2999  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3000  * crucial in the face of [offset, EOF] locks.
3001  *
3002  * Return at least one page in @queue unless there is no covered page.
3003  */
3004 int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3005                         struct osc_object *osc, pgoff_t start, pgoff_t end,
3006                         osc_page_gang_cbt cb, void *cbdata)
3007 {
3008         struct osc_page *ops;
3009         void            **pvec;
3010         pgoff_t         idx;
3011         unsigned int    nr;
3012         unsigned int    i;
3013         unsigned int    j;
3014         int             res = CLP_GANG_OKAY;
3015         bool            tree_lock = true;
3016         ENTRY;
3017
3018         idx = start;
3019         pvec = osc_env_info(env)->oti_pvec;
3020         spin_lock(&osc->oo_tree_lock);
3021         while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3022                                             idx, OTI_PVEC_SIZE)) > 0) {
3023                 struct cl_page *page;
3024                 bool end_of_region = false;
3025
3026                 for (i = 0, j = 0; i < nr; ++i) {
3027                         ops = pvec[i];
3028                         pvec[i] = NULL;
3029
3030                         idx = osc_index(ops);
3031                         if (idx > end) {
3032                                 end_of_region = true;
3033                                 break;
3034                         }
3035
3036                         page = ops->ops_cl.cpl_page;
3037                         LASSERT(page->cp_type == CPT_CACHEABLE);
3038                         if (page->cp_state == CPS_FREEING)
3039                                 continue;
3040
3041                         cl_page_get(page);
3042                         lu_ref_add_atomic(&page->cp_reference,
3043                                           "gang_lookup", current);
3044                         pvec[j++] = ops;
3045                 }
3046                 ++idx;
3047
3048                 /*
3049                  * Here a delicate locking dance is performed. Current thread
3050                  * holds a reference to a page, but has to own it before it
3051                  * can be placed into queue. Owning implies waiting, so
3052                  * radix-tree lock is to be released. After a wait one has to
3053                  * check that pages weren't truncated (cl_page_own() returns
3054                  * error in the latter case).
3055                  */
3056                 spin_unlock(&osc->oo_tree_lock);
3057                 tree_lock = false;
3058
3059                 for (i = 0; i < j; ++i) {
3060                         ops = pvec[i];
3061                         if (res == CLP_GANG_OKAY)
3062                                 res = (*cb)(env, io, ops, cbdata);
3063
3064                         page = ops->ops_cl.cpl_page;
3065                         lu_ref_del(&page->cp_reference, "gang_lookup", current);
3066                         cl_page_put(env, page);
3067                 }
3068                 if (nr < OTI_PVEC_SIZE || end_of_region)
3069                         break;
3070
3071                 if (res == CLP_GANG_OKAY && need_resched())
3072                         res = CLP_GANG_RESCHED;
3073                 if (res != CLP_GANG_OKAY)
3074                         break;
3075
3076                 spin_lock(&osc->oo_tree_lock);
3077                 tree_lock = true;
3078         }
3079         if (tree_lock)
3080                 spin_unlock(&osc->oo_tree_lock);
3081         RETURN(res);
3082 }
3083
3084 /**
3085  * Check if page @page is covered by an extra lock or discard it.
3086  */
3087 static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3088                                 struct osc_page *ops, void *cbdata)
3089 {
3090         struct osc_thread_info *info = osc_env_info(env);
3091         struct osc_object *osc = cbdata;
3092         pgoff_t index;
3093
3094         index = osc_index(ops);
3095         if (index >= info->oti_fn_index) {
3096                 struct ldlm_lock *tmp;
3097                 struct cl_page *page = ops->ops_cl.cpl_page;
3098
3099                 /* refresh non-overlapped index */
3100                 tmp = osc_dlmlock_at_pgoff(env, osc, index, 0, 0);
3101                 if (tmp != NULL) {
3102                         __u64 end = tmp->l_policy_data.l_extent.end;
3103                         /* Cache the first-non-overlapped index so as to skip
3104                          * all pages within [index, oti_fn_index). This is safe
3105                          * because if tmp lock is canceled, it will discard
3106                          * these pages. */
3107                         info->oti_fn_index = cl_index(osc2cl(osc), end + 1);
3108                         if (end == OBD_OBJECT_EOF)
3109                                 info->oti_fn_index = CL_PAGE_EOF;
3110                         LDLM_LOCK_PUT(tmp);
3111                 } else if (cl_page_own(env, io, page) == 0) {
3112                         /* discard the page */
3113                         cl_page_discard(env, io, page);
3114                         cl_page_disown(env, io, page);
3115                 } else {
3116                         LASSERT(page->cp_state == CPS_FREEING);
3117                 }
3118         }
3119
3120         info->oti_next_index = index + 1;
3121         return CLP_GANG_OKAY;
3122 }
3123
3124 static int discard_cb(const struct lu_env *env, struct cl_io *io,
3125                       struct osc_page *ops, void *cbdata)
3126 {
3127         struct osc_thread_info *info = osc_env_info(env);
3128         struct cl_page *page = ops->ops_cl.cpl_page;
3129
3130         /* page is top page. */
3131         info->oti_next_index = osc_index(ops) + 1;
3132         if (cl_page_own(env, io, page) == 0) {
3133                 KLASSERT(ergo(page->cp_type == CPT_CACHEABLE,
3134                               !PageDirty(cl_page_vmpage(page))));
3135
3136                 /* discard the page */
3137                 cl_page_discard(env, io, page);
3138                 cl_page_disown(env, io, page);
3139         } else {
3140                 LASSERT(page->cp_state == CPS_FREEING);
3141         }
3142
3143         return CLP_GANG_OKAY;
3144 }
3145
3146 /**
3147  * Discard pages protected by the given lock. This function traverses radix
3148  * tree to find all covering pages and discard them. If a page is being covered
3149  * by other locks, it should remain in cache.
3150  *
3151  * If error happens on any step, the process continues anyway (the reasoning
3152  * behind this being that lock cancellation cannot be delayed indefinitely).
3153  */
3154 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3155                            pgoff_t start, pgoff_t end, enum cl_lock_mode mode)
3156 {
3157         struct osc_thread_info *info = osc_env_info(env);
3158         struct cl_io *io = &info->oti_io;
3159         osc_page_gang_cbt cb;
3160         int res;
3161         int result;
3162
3163         ENTRY;
3164
3165         io->ci_obj = cl_object_top(osc2cl(osc));
3166         io->ci_ignore_layout = 1;
3167         result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3168         if (result != 0)
3169                 GOTO(out, result);
3170
3171         cb = mode == CLM_READ ? check_and_discard_cb : discard_cb;
3172         info->oti_fn_index = info->oti_next_index = start;
3173         do {
3174                 res = osc_page_gang_lookup(env, io, osc,
3175                                            info->oti_next_index, end, cb, osc);
3176                 if (info->oti_next_index > end)
3177                         break;
3178
3179                 if (res == CLP_GANG_RESCHED)
3180                         cond_resched();
3181         } while (res != CLP_GANG_OKAY);
3182 out:
3183         cl_io_fini(env, io);
3184         RETURN(result);
3185 }
3186
3187
3188 /** @} osc */