lustre/osc/osc_cache.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  *
  28  */
  29 /*
  30  * This file is part of Lustre, http://www.lustre.org/
  31  *
  32  * osc cache management.
  33  *
  34  * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_OSC
  38
  39 #include <lustre_osc.h>
  40 #include <lustre_dlm.h>
  41
  42 #include "osc_internal.h"
  43
  44 static int extent_debug; /* set it to be true for more debug */
  45
  46 static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
  47 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
  48                            enum osc_extent_state state);
  49 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
  50                               struct osc_async_page *oap, int sent, int rc);
  51 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
  52                           int cmd);
  53 static int osc_refresh_count(const struct lu_env *env,
  54                              struct osc_async_page *oap, int cmd);
  55 static int osc_io_unplug_async(const struct lu_env *env,
  56                                struct client_obd *cli, struct osc_object *osc);
  57 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
  58                            unsigned int lost_grant, unsigned int dirty_grant);
  59
  60 static void osc_extent_tree_dump0(int mask, struct osc_object *obj,
  61                                   const char *func, int line);
  62 #define osc_extent_tree_dump(mask, obj) \
  63         osc_extent_tree_dump0(mask, obj, __func__, __LINE__)
  64
  65 static void osc_unreserve_grant(struct client_obd *cli, unsigned int reserved,
  66                                 unsigned int unused);
  67
  68 /** \addtogroup osc
  69  *  @{
  70  */
  71
  72 /* ------------------ osc extent ------------------ */
  73 static inline char *ext_flags(struct osc_extent *ext, char *flags)
  74 {
  75         char *buf = flags;
  76         *buf++ = ext->oe_rw ? 'r' : 'w';
  77         if (!RB_EMPTY_NODE(&ext->oe_node))
  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 #define EXTSTR       "[%lu -> %lu/%lu]"
  98 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
  99 static const char *const oes_strings[] = {
 100         "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
 101
 102 #define OSC_EXTENT_DUMP_WITH_LOC(file, func, line, mask, extent, fmt, ...) do {\
 103         static struct cfs_debug_limit_state cdls;                             \
 104         struct osc_extent *__ext = (extent);                                  \
 105         char __buf[16];                                                       \
 106                                                                               \
 107         __CDEBUG_WITH_LOC(file, func, line, mask, &cdls,                      \
 108                 "extent %p@{" EXTSTR ", "                                     \
 109                 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt,          \
 110                 /* ----- extent part 0 ----- */                               \
 111                 __ext, EXTPARA(__ext),                                        \
 112                 /* ----- part 1 ----- */                                      \
 113                 kref_read(&__ext->oe_refc),                                   \
 114                 atomic_read(&__ext->oe_users),                                \
 115                 list_empty_marker(&__ext->oe_link),                           \
 116                 oes_strings[__ext->oe_state], ext_flags(__ext, __buf),        \
 117                 __ext->oe_obj,                                                \
 118                 /* ----- part 2 ----- */                                      \
 119                 __ext->oe_grants, __ext->oe_nr_pages,                         \
 120                 list_empty_marker(&__ext->oe_pages),                          \
 121                 waitqueue_active(&__ext->oe_waitq) ? '+' : '-',               \
 122                 __ext->oe_dlmlock, __ext->oe_mppr, __ext->oe_owner,           \
 123                 /* ----- part 4 ----- */                                      \
 124                 ## __VA_ARGS__);                                              \
 125         if (mask == D_ERROR && __ext->oe_dlmlock != NULL)                     \
 126                 LDLM_ERROR(__ext->oe_dlmlock, "extent: %p", __ext);           \
 127         else                                                                  \
 128                 LDLM_DEBUG(__ext->oe_dlmlock, "extent: %p", __ext);           \
 129 } while (0)
 130
 131 #define OSC_EXTENT_DUMP(mask, ext, fmt, ...)                            \
 132         OSC_EXTENT_DUMP_WITH_LOC(__FILE__, __func__, __LINE__,          \
 133                                  mask, ext, fmt, ## __VA_ARGS__)
 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         return rb_entry_safe(n, struct osc_extent, oe_node);
 150 }
 151
 152 static inline struct osc_extent *next_extent(struct osc_extent *ext)
 153 {
 154         if (ext == NULL)
 155                 return NULL;
 156
 157         LASSERT(!RB_EMPTY_NODE(&ext->oe_node));
 158         return rb_extent(rb_next(&ext->oe_node));
 159 }
 160
 161 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
 162 {
 163         if (ext == NULL)
 164                 return NULL;
 165
 166         LASSERT(!RB_EMPTY_NODE(&ext->oe_node));
 167         return rb_extent(rb_prev(&ext->oe_node));
 168 }
 169
 170 static inline struct osc_extent *first_extent(struct osc_object *obj)
 171 {
 172         return rb_extent(rb_first(&obj->oo_root));
 173 }
 174
 175 /* object must be locked by caller. */
 176 static int osc_extent_sanity_check0(struct osc_extent *ext,
 177                                     const char *func, const int line)
 178 {
 179         struct osc_object *obj = ext->oe_obj;
 180         struct osc_async_page *oap;
 181         size_t page_count;
 182         int rc = 0;
 183
 184         assert_osc_object_is_locked(obj);
 185
 186         if (ext->oe_state >= OES_STATE_MAX)
 187                 GOTO(out, rc = 10);
 188
 189         if (kref_read(&ext->oe_refc) <= 0)
 190                 GOTO(out, rc = 20);
 191
 192         if (kref_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
 193                 GOTO(out, rc = 30);
 194
 195         switch (ext->oe_state) {
 196         case OES_INV:
 197                 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
 198                         GOTO(out, rc = 35);
 199                 GOTO(out, rc = 0);
 200                 break;
 201         case OES_ACTIVE:
 202                 if (atomic_read(&ext->oe_users) == 0)
 203                         GOTO(out, rc = 40);
 204                 if (ext->oe_hp)
 205                         GOTO(out, rc = 50);
 206                 if (ext->oe_fsync_wait && !ext->oe_urgent)
 207                         GOTO(out, rc = 55);
 208                 break;
 209         case OES_CACHE:
 210                 if (ext->oe_grants == 0)
 211                         GOTO(out, rc = 60);
 212                 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
 213                         GOTO(out, rc = 65);
 214                 /* fallthrough */
 215         default:
 216                 if (atomic_read(&ext->oe_users) > 0)
 217                         GOTO(out, rc = 70);
 218         }
 219
 220         if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
 221                 GOTO(out, rc = 80);
 222
 223         if (ext->oe_sync && ext->oe_grants > 0)
 224                 GOTO(out, rc = 90);
 225
 226         if (ext->oe_dlmlock != NULL &&
 227             ext->oe_dlmlock->l_resource->lr_type == LDLM_EXTENT &&
 228             !ldlm_is_failed(ext->oe_dlmlock)) {
 229                 struct ldlm_extent *extent;
 230
 231                 extent = &ext->oe_dlmlock->l_policy_data.l_extent;
 232                 if (!(extent->start <= cl_offset(osc2cl(obj), ext->oe_start) &&
 233                       extent->end   >= cl_offset(osc2cl(obj), ext->oe_max_end)))
 234                         GOTO(out, rc = 100);
 235
 236                 if (!(ext->oe_dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP)))
 237                         GOTO(out, rc = 102);
 238         }
 239
 240         if (ext->oe_nr_pages > ext->oe_mppr)
 241                 GOTO(out, rc = 105);
 242
 243         /* Do not verify page list if extent is in RPC. This is because an
 244          * in-RPC extent is supposed to be exclusively accessible w/o lock. */
 245         if (ext->oe_state > OES_CACHE)
 246                 GOTO(out, rc = 0);
 247
 248         if (!extent_debug)
 249                 GOTO(out, rc = 0);
 250
 251         page_count = 0;
 252         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
 253                 pgoff_t index = osc_index(oap2osc(oap));
 254                 ++page_count;
 255                 if (index > ext->oe_end || index < ext->oe_start)
 256                         GOTO(out, rc = 110);
 257         }
 258         if (page_count != ext->oe_nr_pages)
 259                 GOTO(out, rc = 120);
 260
 261 out:
 262         if (rc != 0)
 263                 OSC_EXTENT_DUMP_WITH_LOC(__FILE__, func, line, D_ERROR, ext,
 264                                          "sanity check %p failed: rc = %d\n",
 265                                          ext, rc);
 266         return rc;
 267 }
 268
 269 #define sanity_check_nolock(ext) \
 270         osc_extent_sanity_check0(ext, __func__, __LINE__)
 271
 272 #define sanity_check(ext) ({                                                   \
 273         int __res;                                                             \
 274         osc_object_lock((ext)->oe_obj);                                        \
 275         __res = sanity_check_nolock(ext);                                      \
 276         osc_object_unlock((ext)->oe_obj);                                      \
 277         __res;                                                                 \
 278 })
 279
 280 static inline bool
 281 overlapped(const struct osc_extent *ex1, const struct osc_extent *ex2)
 282 {
 283         return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
 284 }
 285
 286 /**
 287  * sanity check - to make sure there is no overlapped extent in the tree.
 288  */
 289 static int osc_extent_is_overlapped(struct osc_object *obj,
 290                                     struct osc_extent *ext)
 291 {
 292         struct osc_extent *tmp;
 293
 294         assert_osc_object_is_locked(obj);
 295
 296         if (!extent_debug)
 297                 return 0;
 298
 299         for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
 300                 if (tmp == ext)
 301                         continue;
 302                 if (overlapped(tmp, ext))
 303                         return 1;
 304         }
 305         return 0;
 306 }
 307
 308 static void osc_extent_state_set(struct osc_extent *ext, int state)
 309 {
 310         assert_osc_object_is_locked(ext->oe_obj);
 311         LASSERT(state >= OES_INV && state < OES_STATE_MAX);
 312
 313         /* Never try to sanity check a state changing extent :-) */
 314         /* LASSERT(sanity_check_nolock(ext) == 0); */
 315
 316         /* TODO: validate the state machine */
 317         smp_store_release(&ext->oe_state, state);
 318         wake_up(&ext->oe_waitq);
 319 }
 320
 321 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
 322 {
 323         struct osc_extent *ext;
 324
 325         OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_NOFS);
 326         if (ext == NULL)
 327                 return NULL;
 328
 329         RB_CLEAR_NODE(&ext->oe_node);
 330         ext->oe_obj = obj;
 331         cl_object_get(osc2cl(obj));
 332         kref_init(&ext->oe_refc);
 333         atomic_set(&ext->oe_users, 0);
 334         INIT_LIST_HEAD(&ext->oe_link);
 335         ext->oe_state = OES_INV;
 336         INIT_LIST_HEAD(&ext->oe_pages);
 337         init_waitqueue_head(&ext->oe_waitq);
 338         ext->oe_dlmlock = NULL;
 339
 340         return ext;
 341 }
 342
 343 static void osc_extent_free(struct kref *kref)
 344 {
 345         struct osc_extent *ext = container_of(kref, struct osc_extent,
 346                                               oe_refc);
 347
 348         LASSERT(list_empty(&ext->oe_link));
 349         LASSERT(atomic_read(&ext->oe_users) == 0);
 350         LASSERT(ext->oe_state == OES_INV);
 351         LASSERT(RB_EMPTY_NODE(&ext->oe_node));
 352
 353         if (ext->oe_dlmlock) {
 354                 lu_ref_del(&ext->oe_dlmlock->l_reference,
 355                            "osc_extent", ext);
 356                 LDLM_LOCK_PUT(ext->oe_dlmlock);
 357                 ext->oe_dlmlock = NULL;
 358         }
 359 #if 0
 360         /* If/When cl_object_put drops the need for 'env',
 361          * this code can be enabled, and matching code in
 362          * osc_extent_put removed.
 363          */
 364         cl_object_put(osc2cl(ext->oe_obj));
 365
 366         OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
 367 #endif
 368 }
 369
 370 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
 371 {
 372         LASSERT(kref_read(&ext->oe_refc) >= 0);
 373         kref_get(&ext->oe_refc);
 374         return ext;
 375 }
 376
 377 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
 378 {
 379         LASSERT(kref_read(&ext->oe_refc) > 0);
 380         if (kref_put(&ext->oe_refc, osc_extent_free)) {
 381                 /* This should be in osc_extent_free(), but
 382                  * while we need to pass 'env' it cannot be.
 383                  */
 384                 cl_object_put(env, osc2cl(ext->oe_obj));
 385
 386                 OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
 387         }
 388 }
 389
 390 /**
 391  * osc_extent_put_trust() is a special version of osc_extent_put() when
 392  * it's known that the caller is not the last user. This is to address the
 393  * problem of lacking of lu_env ;-).
 394  */
 395 static void osc_extent_put_trust(struct osc_extent *ext)
 396 {
 397         LASSERT(kref_read(&ext->oe_refc) > 1);
 398         assert_osc_object_is_locked(ext->oe_obj);
 399         osc_extent_put(NULL, ext);
 400 }
 401
 402 /**
 403  * Return the extent which includes pgoff @index, or return the greatest
 404  * previous extent in the tree.
 405  */
 406 static struct osc_extent *osc_extent_search(struct osc_object *obj,
 407                                             pgoff_t index)
 408 {
 409         struct rb_node    *n = obj->oo_root.rb_node;
 410         struct osc_extent *tmp, *p = NULL;
 411
 412         assert_osc_object_is_locked(obj);
 413         while (n != NULL) {
 414                 tmp = rb_extent(n);
 415                 if (index < tmp->oe_start) {
 416                         n = n->rb_left;
 417                 } else if (index > tmp->oe_end) {
 418                         p = rb_extent(n);
 419                         n = n->rb_right;
 420                 } else {
 421                         return tmp;
 422                 }
 423         }
 424         return p;
 425 }
 426
 427 /*
 428  * Return the extent covering @index, otherwise return NULL.
 429  * caller must have held object lock.
 430  */
 431 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
 432                                             pgoff_t index)
 433 {
 434         struct osc_extent *ext;
 435
 436         ext = osc_extent_search(obj, index);
 437         if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
 438                 return osc_extent_get(ext);
 439         return NULL;
 440 }
 441
 442 /* caller must have held object lock. */
 443 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
 444 {
 445         struct rb_node   **n      = &obj->oo_root.rb_node;
 446         struct rb_node    *parent = NULL;
 447         struct osc_extent *tmp;
 448
 449         LASSERT(RB_EMPTY_NODE(&ext->oe_node));
 450         LASSERT(ext->oe_obj == obj);
 451         assert_osc_object_is_locked(obj);
 452         while (*n != NULL) {
 453                 tmp = rb_extent(*n);
 454                 parent = *n;
 455
 456                 if (ext->oe_end < tmp->oe_start)
 457                         n = &(*n)->rb_left;
 458                 else if (ext->oe_start > tmp->oe_end)
 459                         n = &(*n)->rb_right;
 460                 else
 461                         EASSERTF(0, tmp, EXTSTR"\n", EXTPARA(ext));
 462         }
 463         rb_link_node(&ext->oe_node, parent, n);
 464         rb_insert_color(&ext->oe_node, &obj->oo_root);
 465         osc_extent_get(ext);
 466 }
 467
 468 /* caller must have held object lock. */
 469 static void osc_extent_erase(struct osc_extent *ext)
 470 {
 471         struct osc_object *obj = ext->oe_obj;
 472         assert_osc_object_is_locked(obj);
 473         if (!RB_EMPTY_NODE(&ext->oe_node)) {
 474                 rb_erase(&ext->oe_node, &obj->oo_root);
 475                 RB_CLEAR_NODE(&ext->oe_node);
 476                 /* rbtree held a refcount */
 477                 osc_extent_put_trust(ext);
 478         }
 479 }
 480
 481 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
 482 {
 483         struct osc_object *obj = ext->oe_obj;
 484
 485         assert_osc_object_is_locked(obj);
 486         LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
 487         if (ext->oe_state == OES_CACHE) {
 488                 osc_extent_state_set(ext, OES_ACTIVE);
 489                 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
 490         }
 491         atomic_inc(&ext->oe_users);
 492         list_del_init(&ext->oe_link);
 493         return osc_extent_get(ext);
 494 }
 495
 496 static void __osc_extent_remove(struct osc_extent *ext)
 497 {
 498         assert_osc_object_is_locked(ext->oe_obj);
 499         LASSERT(list_empty(&ext->oe_pages));
 500         osc_extent_erase(ext);
 501         list_del_init(&ext->oe_link);
 502         osc_extent_state_set(ext, OES_INV);
 503         OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
 504 }
 505
 506 static void osc_extent_remove(struct osc_extent *ext)
 507 {
 508         struct osc_object *obj = ext->oe_obj;
 509
 510         osc_object_lock(obj);
 511         __osc_extent_remove(ext);
 512         osc_object_unlock(obj);
 513 }
 514
 515 /**
 516  * This function is used to merge extents to get better performance. It checks
 517  * if @cur and @victim are contiguous at block level.
 518  */
 519 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
 520                             struct osc_extent *victim)
 521 {
 522         struct osc_object       *obj = cur->oe_obj;
 523         struct client_obd       *cli = osc_cli(obj);
 524         pgoff_t                  chunk_start;
 525         pgoff_t                  chunk_end;
 526         int                      ppc_bits;
 527
 528         LASSERT(cur->oe_state == OES_CACHE);
 529         assert_osc_object_is_locked(obj);
 530         if (victim == NULL)
 531                 return -EINVAL;
 532
 533         if (victim->oe_state != OES_INV &&
 534             (victim->oe_state != OES_CACHE || victim->oe_fsync_wait))
 535                 return -EBUSY;
 536
 537         if (cur->oe_max_end != victim->oe_max_end)
 538                 return -ERANGE;
 539
 540         /*
 541          * In the rare case max_pages_per_rpc (mppr) is changed, don't
 542          * merge extents until after old ones have been sent, or the
 543          * "extents are aligned to RPCs" checks are unhappy.
 544          */
 545         if (cur->oe_mppr != victim->oe_mppr)
 546                 return -ERANGE;
 547
 548         LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
 549         ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_SHIFT;
 550         chunk_start = cur->oe_start >> ppc_bits;
 551         chunk_end   = cur->oe_end   >> ppc_bits;
 552         if (chunk_start   != (victim->oe_end >> ppc_bits) + 1 &&
 553             chunk_end + 1 != victim->oe_start >> ppc_bits)
 554                 return -ERANGE;
 555
 556         /* overall extent size should not exceed the max supported limit
 557          * reported by the server */
 558         if (cur->oe_end - cur->oe_start + 1 +
 559             victim->oe_end - victim->oe_start + 1 > cli->cl_max_extent_pages)
 560                 return -ERANGE;
 561
 562         OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
 563
 564         cur->oe_start     = min(cur->oe_start, victim->oe_start);
 565         cur->oe_end       = max(cur->oe_end,   victim->oe_end);
 566         /* per-extent tax should be accounted only once for the whole extent */
 567         cur->oe_grants   += victim->oe_grants - cli->cl_grant_extent_tax;
 568         cur->oe_nr_pages += victim->oe_nr_pages;
 569         /* only the following bits are needed to merge */
 570         cur->oe_urgent   |= victim->oe_urgent;
 571         cur->oe_memalloc |= victim->oe_memalloc;
 572         list_splice_init(&victim->oe_pages, &cur->oe_pages);
 573         victim->oe_nr_pages = 0;
 574
 575         osc_extent_get(victim);
 576         __osc_extent_remove(victim);
 577         osc_extent_put(env, victim);
 578
 579         OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
 580         return 0;
 581 }
 582
 583 /**
 584  * Drop user count of osc_extent, and unplug IO asynchronously.
 585  */
 586 void osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
 587 {
 588         struct osc_object *obj = ext->oe_obj;
 589         struct client_obd *cli = osc_cli(obj);
 590         ENTRY;
 591
 592         LASSERT(atomic_read(&ext->oe_users) > 0);
 593         LASSERT(sanity_check(ext) == 0);
 594         LASSERT(ext->oe_grants > 0);
 595
 596         if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
 597                 LASSERT(ext->oe_state == OES_ACTIVE);
 598                 if (ext->oe_trunc_pending) {
 599                         /* a truncate process is waiting for this extent.
 600                          * This may happen due to a race, check
 601                          * osc_cache_truncate_start(). */
 602                         osc_extent_state_set(ext, OES_TRUNC);
 603                         ext->oe_trunc_pending = 0;
 604                         osc_object_unlock(obj);
 605                 } else {
 606                         int grant = 0;
 607
 608                         osc_extent_state_set(ext, OES_CACHE);
 609                         osc_update_pending(obj, OBD_BRW_WRITE,
 610                                            ext->oe_nr_pages);
 611
 612                         /* try to merge the previous and next extent. */
 613                         if (osc_extent_merge(env, ext, prev_extent(ext)) == 0)
 614                                 grant += cli->cl_grant_extent_tax;
 615                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 616                                 grant += cli->cl_grant_extent_tax;
 617
 618                         if (ext->oe_hp)
 619                                 list_move_tail(&ext->oe_link,
 620                                                &obj->oo_hp_exts);
 621                         else if (ext->oe_urgent)
 622                                 list_move_tail(&ext->oe_link,
 623                                                &obj->oo_urgent_exts);
 624                         else if (ext->oe_nr_pages == ext->oe_mppr) {
 625                                 list_move_tail(&ext->oe_link,
 626                                                &obj->oo_full_exts);
 627                         }
 628                         osc_object_unlock(obj);
 629                         if (grant > 0)
 630                                 osc_unreserve_grant(cli, 0, grant);
 631                 }
 632
 633                 osc_io_unplug_async(env, cli, obj);
 634         }
 635         osc_extent_put(env, ext);
 636
 637         RETURN_EXIT;
 638 }
 639
 640 /**
 641  * Find or create an extent which includes @index, core function to manage
 642  * extent tree.
 643  */
 644 static struct osc_extent *osc_extent_find(const struct lu_env *env,
 645                                           struct osc_object *obj, pgoff_t index,
 646                                           unsigned int *grants)
 647 {
 648         struct client_obd *cli = osc_cli(obj);
 649         struct osc_lock   *olck;
 650         struct cl_lock_descr *descr;
 651         struct osc_extent *cur;
 652         struct osc_extent *ext;
 653         struct osc_extent *conflict = NULL;
 654         struct osc_extent *found = NULL;
 655         pgoff_t    chunk;
 656         pgoff_t    max_end;
 657         unsigned int max_pages; /* max_pages_per_rpc */
 658         unsigned int chunksize;
 659         int        ppc_bits; /* pages per chunk bits */
 660         pgoff_t    chunk_mask;
 661         int        rc;
 662         ENTRY;
 663
 664         cur = osc_extent_alloc(obj);
 665         if (cur == NULL)
 666                 RETURN(ERR_PTR(-ENOMEM));
 667
 668         olck = osc_env_io(env)->oi_write_osclock;
 669         LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
 670         LASSERT(olck->ols_state == OLS_GRANTED);
 671
 672         descr = &olck->ols_cl.cls_lock->cll_descr;
 673         LASSERT(descr->cld_mode >= CLM_WRITE);
 674
 675         LASSERTF(cli->cl_chunkbits >= PAGE_SHIFT,
 676                  "chunkbits: %u\n", cli->cl_chunkbits);
 677         ppc_bits   = cli->cl_chunkbits - PAGE_SHIFT;
 678         chunk_mask = ~((1 << ppc_bits) - 1);
 679         chunksize  = 1 << cli->cl_chunkbits;
 680         chunk      = index >> ppc_bits;
 681
 682         /* align end to RPC edge. */
 683         max_pages = cli->cl_max_pages_per_rpc;
 684         if ((max_pages & ~chunk_mask) != 0) {
 685                 CERROR("max_pages: %#x chunkbits: %u chunk_mask: %#lx\n",
 686                        max_pages, cli->cl_chunkbits, chunk_mask);
 687                 RETURN(ERR_PTR(-EINVAL));
 688         }
 689         max_end = index - (index % max_pages) + max_pages - 1;
 690         max_end = min_t(pgoff_t, max_end, descr->cld_end);
 691
 692         /* initialize new extent by parameters so far */
 693         cur->oe_max_end = max_end;
 694         cur->oe_start   = index & chunk_mask;
 695         cur->oe_end     = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
 696         if (cur->oe_start < descr->cld_start)
 697                 cur->oe_start = descr->cld_start;
 698         if (cur->oe_end > max_end)
 699                 cur->oe_end = max_end;
 700         cur->oe_grants  = chunksize + cli->cl_grant_extent_tax;
 701         cur->oe_mppr    = max_pages;
 702         if (olck->ols_dlmlock != NULL) {
 703                 LASSERT(olck->ols_hold);
 704                 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
 705                 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
 706         }
 707
 708         /* grants has been allocated by caller */
 709         LASSERTF(*grants >= chunksize + cli->cl_grant_extent_tax,
 710                  "%u/%u/%u.\n", *grants, chunksize, cli->cl_grant_extent_tax);
 711         LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
 712                  EXTPARA(cur));
 713
 714 restart:
 715         osc_object_lock(obj);
 716         ext = osc_extent_search(obj, cur->oe_start);
 717         if (!ext)
 718                 ext = first_extent(obj);
 719         for (; ext; ext = next_extent(ext)) {
 720                 pgoff_t ext_chk_start = ext->oe_start >> ppc_bits;
 721                 pgoff_t ext_chk_end = ext->oe_end >> ppc_bits;
 722
 723                 LASSERT(sanity_check_nolock(ext) == 0);
 724                 if (chunk > ext_chk_end + 1 || chunk < ext_chk_start)
 725                         break;
 726
 727                 /* if covering by different locks, no chance to match */
 728                 if (olck->ols_dlmlock != ext->oe_dlmlock) {
 729                         EASSERTF(!overlapped(ext, cur), ext,
 730                                  EXTSTR"\n", EXTPARA(cur));
 731
 732                         continue;
 733                 }
 734
 735                 /* discontiguous chunks? */
 736                 if (chunk + 1 < ext_chk_start)
 737                         continue;
 738
 739                 /* ok, from now on, ext and cur have these attrs:
 740                  * 1. covered by the same lock
 741                  * 2. contiguous at chunk level or overlapping. */
 742
 743                 if (overlapped(ext, cur)) {
 744                         /* cur is the minimum unit, so overlapping means
 745                          * full contain. */
 746                         EASSERTF((ext->oe_start <= cur->oe_start &&
 747                                   ext->oe_end >= cur->oe_end),
 748                                  ext, EXTSTR"\n", EXTPARA(cur));
 749
 750                         if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
 751                                 /* for simplicity, we wait for this extent to
 752                                  * finish before going forward. */
 753                                 conflict = osc_extent_get(ext);
 754                                 break;
 755                         }
 756
 757                         found = osc_extent_hold(ext);
 758                         break;
 759                 }
 760
 761                 /* non-overlapped extent */
 762                 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait)
 763                         /* we can't do anything for a non OES_CACHE extent, or
 764                          * if there is someone waiting for this extent to be
 765                          * flushed, try next one. */
 766                         continue;
 767
 768                 if (osc_extent_merge(env, ext, cur) == 0) {
 769                         LASSERT(*grants >= chunksize);
 770                         *grants -= chunksize;
 771
 772                         /*
 773                          * Try to merge with the next one too because we
 774                          * might have just filled in a gap.
 775                          */
 776                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 777                                 /* we can save extent tax from next extent */
 778                                 *grants += cli->cl_grant_extent_tax;
 779
 780                         found = osc_extent_hold(ext);
 781                         break;
 782                 }
 783         }
 784
 785         osc_extent_tree_dump(D_CACHE, obj);
 786         if (found != NULL) {
 787                 LASSERT(conflict == NULL);
 788                 if (!IS_ERR(found)) {
 789                         LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
 790                         OSC_EXTENT_DUMP(D_CACHE, found,
 791                                         "found caching ext for %lu.\n", index);
 792                 }
 793         } else if (conflict == NULL) {
 794                 /* create a new extent */
 795                 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
 796                 LASSERT(*grants >= cur->oe_grants);
 797                 *grants -= cur->oe_grants;
 798
 799                 cur->oe_state = OES_CACHE;
 800                 found = osc_extent_hold(cur);
 801                 osc_extent_insert(obj, cur);
 802                 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
 803                                 index, descr->cld_end);
 804         }
 805         osc_object_unlock(obj);
 806
 807         if (conflict != NULL) {
 808                 LASSERT(found == NULL);
 809
 810                 /* waiting for IO to finish. Please notice that it's impossible
 811                  * to be an OES_TRUNC extent. */
 812                 rc = osc_extent_wait(env, conflict, OES_INV);
 813                 osc_extent_put(env, conflict);
 814                 conflict = NULL;
 815                 if (rc < 0)
 816                         GOTO(out, found = ERR_PTR(rc));
 817
 818                 goto restart;
 819         }
 820         EXIT;
 821
 822 out:
 823         osc_extent_put(env, cur);
 824         return found;
 825 }
 826
 827 /**
 828  * Called when IO is finished to an extent.
 829  */
 830 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
 831                       int sent, int rc)
 832 {
 833         struct client_obd *cli = osc_cli(ext->oe_obj);
 834         struct osc_async_page *oap;
 835         struct osc_async_page *tmp;
 836         int nr_pages = ext->oe_nr_pages;
 837         int lost_grant = 0;
 838         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
 839         loff_t last_off = 0;
 840         int last_count = -1;
 841         ENTRY;
 842
 843         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
 844
 845         ext->oe_rc = rc ?: ext->oe_nr_pages;
 846         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
 847
 848         osc_lru_add_batch(cli, &ext->oe_pages);
 849         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 850                                      oap_pending_item) {
 851                 list_del_init(&oap->oap_rpc_item);
 852                 list_del_init(&oap->oap_pending_item);
 853                 if (last_off <= oap->oap_obj_off) {
 854                         last_off = oap->oap_obj_off;
 855                         last_count = oap->oap_count;
 856                 }
 857
 858                 --ext->oe_nr_pages;
 859                 osc_ap_completion(env, cli, oap, sent, rc);
 860         }
 861         EASSERT(ext->oe_nr_pages == 0, ext);
 862
 863         if (!sent) {
 864                 lost_grant = ext->oe_grants;
 865         } else if (blocksize < PAGE_SIZE &&
 866                    last_count != PAGE_SIZE) {
 867                 /* For short writes we shouldn't count parts of pages that
 868                  * span a whole chunk on the OST side, or our accounting goes
 869                  * wrong.  Should match the code in filter_grant_check. */
 870                 int offset = last_off & ~PAGE_MASK;
 871                 int count = last_count + (offset & (blocksize - 1));
 872                 int end = (offset + last_count) & (blocksize - 1);
 873                 if (end)
 874                         count += blocksize - end;
 875
 876                 lost_grant = PAGE_SIZE - count;
 877         }
 878         if (ext->oe_grants > 0)
 879                 osc_free_grant(cli, nr_pages, lost_grant, ext->oe_grants);
 880
 881         osc_extent_remove(ext);
 882         /* put the refcount for RPC */
 883         osc_extent_put(env, ext);
 884         RETURN(0);
 885 }
 886
 887 /**
 888  * Wait for the extent's state to become @state.
 889  */
 890 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
 891                            enum osc_extent_state state)
 892 {
 893         struct osc_object *obj = ext->oe_obj;
 894         int rc = 0;
 895         ENTRY;
 896
 897         osc_object_lock(obj);
 898         LASSERT(sanity_check_nolock(ext) == 0);
 899         /* `Kick' this extent only if the caller is waiting for it to be
 900          * written out. */
 901         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
 902                 if (ext->oe_state == OES_ACTIVE) {
 903                         ext->oe_urgent = 1;
 904                 } else if (ext->oe_state == OES_CACHE) {
 905                         ext->oe_urgent = 1;
 906                         osc_extent_hold(ext);
 907                         rc = 1;
 908                 }
 909         }
 910         osc_object_unlock(obj);
 911         if (rc == 1)
 912                 osc_extent_release(env, ext);
 913
 914         /* wait for the extent until its state becomes @state */
 915         rc = wait_event_idle_timeout(ext->oe_waitq,
 916                                      smp_load_acquire(&ext->oe_state) == state,
 917                                      cfs_time_seconds(600));
 918         if (rc == 0) {
 919                 OSC_EXTENT_DUMP(D_ERROR, ext,
 920                         "%s: wait ext to %u timedout, recovery in progress?\n",
 921                         cli_name(osc_cli(obj)), state);
 922
 923                 wait_event_idle(ext->oe_waitq,
 924                                 smp_load_acquire(&ext->oe_state) == state);
 925         }
 926         if (ext->oe_rc < 0)
 927                 rc = ext->oe_rc;
 928         else
 929                 rc = 0;
 930         RETURN(rc);
 931 }
 932
 933 /**
 934  * Discard pages with index greater than @size. If @ext is overlapped with
 935  * @size, then partial truncate happens.
 936  */
 937 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
 938                                 bool partial)
 939 {
 940         struct lu_env         *env;
 941         struct cl_io          *io;
 942         struct osc_object     *obj = ext->oe_obj;
 943         struct client_obd     *cli = osc_cli(obj);
 944         struct osc_async_page *oap;
 945         struct osc_async_page *tmp;
 946         struct pagevec        *pvec;
 947         int                    pages_in_chunk = 0;
 948         int                    ppc_bits    = cli->cl_chunkbits -
 949                                              PAGE_SHIFT;
 950         __u64                  trunc_chunk = trunc_index >> ppc_bits;
 951         int                    grants   = 0;
 952         int                    nr_pages = 0;
 953         int                    rc       = 0;
 954         __u16                  refcheck;
 955         ENTRY;
 956
 957         LASSERT(sanity_check(ext) == 0);
 958         LASSERT(ext->oe_state == OES_TRUNC);
 959         LASSERT(!ext->oe_urgent);
 960
 961         /* Request new lu_env.
 962          * We can't use that env from osc_cache_truncate_start() because
 963          * it's from lov_io_sub and not fully initialized. */
 964         env = cl_env_get(&refcheck);
 965         if (IS_ERR(env))
 966                 RETURN(PTR_ERR(env));
 967
 968         io  = osc_env_thread_io(env);
 969         io->ci_obj = cl_object_top(osc2cl(obj));
 970         io->ci_ignore_layout = 1;
 971         pvec = &osc_env_info(env)->oti_pagevec;
 972         ll_pagevec_init(pvec, 0);
 973         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
 974         if (rc < 0)
 975                 GOTO(out, rc);
 976
 977         /* discard all pages with index greater than trunc_index */
 978         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 979                                      oap_pending_item) {
 980                 pgoff_t index = osc_index(oap2osc(oap));
 981                 struct cl_page  *page = oap2cl_page(oap);
 982
 983                 LASSERT(list_empty(&oap->oap_rpc_item));
 984
 985                 /* only discard the pages with their index greater than
 986                  * trunc_index, and ... */
 987                 if (index < trunc_index ||
 988                     (index == trunc_index && partial)) {
 989                         /* accounting how many pages remaining in the chunk
 990                          * so that we can calculate grants correctly. */
 991                         if (index >> ppc_bits == trunc_chunk)
 992                                 ++pages_in_chunk;
 993                         continue;
 994                 }
 995
 996                 list_del_init(&oap->oap_pending_item);
 997
 998                 cl_page_get(page);
 999                 lu_ref_add(&page->cp_reference, "truncate", current);
1000
1001                 if (cl_page_own(env, io, page) == 0) {
1002                         cl_page_discard(env, io, page);
1003                         cl_page_disown(env, io, page);
1004                 } else {
1005                         LASSERT(page->cp_state == CPS_FREEING);
1006                         LASSERT(0);
1007                 }
1008
1009                 lu_ref_del(&page->cp_reference, "truncate", current);
1010                 cl_pagevec_put(env, page, pvec);
1011
1012                 --ext->oe_nr_pages;
1013                 ++nr_pages;
1014         }
1015         pagevec_release(pvec);
1016
1017         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1018                       ext->oe_nr_pages == 0),
1019                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1020
1021         osc_object_lock(obj);
1022         if (ext->oe_nr_pages == 0) {
1023                 LASSERT(pages_in_chunk == 0);
1024                 grants = ext->oe_grants;
1025                 ext->oe_grants = 0;
1026         } else { /* calculate how many grants we can free */
1027                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1028                 pgoff_t last_index;
1029
1030
1031                 /* if there is no pages in this chunk, we can also free grants
1032                  * for the last chunk */
1033                 if (pages_in_chunk == 0) {
1034                         /* if this is the 1st chunk and no pages in this chunk,
1035                          * ext->oe_nr_pages must be zero, so we should be in
1036                          * the other if-clause. */
1037                         LASSERT(trunc_chunk > 0);
1038                         --trunc_chunk;
1039                         ++chunks;
1040                 }
1041
1042                 /* this is what we can free from this extent */
1043                 grants          = chunks << cli->cl_chunkbits;
1044                 ext->oe_grants -= grants;
1045                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1046                 ext->oe_end     = min(last_index, ext->oe_max_end);
1047                 LASSERT(ext->oe_end >= ext->oe_start);
1048                 LASSERT(ext->oe_grants > 0);
1049         }
1050         osc_object_unlock(obj);
1051
1052         if (grants > 0 || nr_pages > 0)
1053                 osc_free_grant(cli, nr_pages, grants, grants);
1054
1055 out:
1056         cl_io_fini(env, io);
1057         cl_env_put(env, &refcheck);
1058         RETURN(rc);
1059 }
1060
1061 /**
1062  * This function is used to make the extent prepared for transfer.
1063  * A race with flusing page - ll_writepage() has to be handled cautiously.
1064  */
1065 static int osc_extent_make_ready(const struct lu_env *env,
1066                                  struct osc_extent *ext)
1067 {
1068         struct osc_async_page *oap;
1069         struct osc_async_page *last = NULL;
1070         struct osc_object *obj = ext->oe_obj;
1071         unsigned int page_count = 0;
1072         int rc;
1073         ENTRY;
1074
1075         /* we're going to grab page lock, so object lock must not be taken. */
1076         LASSERT(sanity_check(ext) == 0);
1077         /* in locking state, any process should not touch this extent. */
1078         EASSERT(ext->oe_state == OES_LOCKING, ext);
1079         EASSERT(ext->oe_owner != NULL, ext);
1080
1081         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1082
1083         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1084                 ++page_count;
1085                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1086                         last = oap;
1087
1088                 /* checking ASYNC_READY is race safe */
1089                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1090                         continue;
1091
1092                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1093                 switch (rc) {
1094                 case 0:
1095                         spin_lock(&oap->oap_lock);
1096                         oap->oap_async_flags |= ASYNC_READY;
1097                         spin_unlock(&oap->oap_lock);
1098                         break;
1099                 case -EALREADY:
1100                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1101                         break;
1102                 default:
1103                         LASSERTF(0, "unknown return code: %d\n", rc);
1104                 }
1105         }
1106
1107         LASSERT(page_count == ext->oe_nr_pages);
1108         LASSERT(last != NULL);
1109         /* the last page is the only one we need to refresh its count by
1110          * the size of file. */
1111         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1112                 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1113                 LASSERTF(last_oap_count > 0,
1114                          "last_oap_count %d\n", last_oap_count);
1115                 LASSERT(last->oap_page_off + last_oap_count <= PAGE_SIZE);
1116                 last->oap_count = last_oap_count;
1117                 spin_lock(&last->oap_lock);
1118                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1119                 spin_unlock(&last->oap_lock);
1120         }
1121
1122         /* for the rest of pages, we don't need to call osf_refresh_count()
1123          * because it's known they are not the last page */
1124         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1125                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1126                         oap->oap_count = PAGE_SIZE - oap->oap_page_off;
1127                         spin_lock(&oap->oap_lock);
1128                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1129                         spin_unlock(&oap->oap_lock);
1130                 }
1131         }
1132
1133         osc_object_lock(obj);
1134         osc_extent_state_set(ext, OES_RPC);
1135         osc_object_unlock(obj);
1136         /* get a refcount for RPC. */
1137         osc_extent_get(ext);
1138
1139         RETURN(0);
1140 }
1141
1142 /**
1143  * Quick and simple version of osc_extent_find(). This function is frequently
1144  * called to expand the extent for the same IO. To expand the extent, the
1145  * page index must be in the same or next chunk of ext->oe_end.
1146  */
1147 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index,
1148                              unsigned int *grants)
1149 {
1150         struct osc_object *obj = ext->oe_obj;
1151         struct client_obd *cli = osc_cli(obj);
1152         struct osc_extent *next;
1153         int ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1154         pgoff_t chunk = index >> ppc_bits;
1155         pgoff_t end_chunk;
1156         pgoff_t end_index;
1157         unsigned int chunksize = 1 << cli->cl_chunkbits;
1158         int rc = 0;
1159         ENTRY;
1160
1161         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1162         osc_object_lock(obj);
1163         LASSERT(sanity_check_nolock(ext) == 0);
1164         end_chunk = ext->oe_end >> ppc_bits;
1165         if (chunk > end_chunk + 1)
1166                 GOTO(out, rc = -ERANGE);
1167
1168         if (end_chunk >= chunk)
1169                 GOTO(out, rc = 0);
1170
1171         LASSERT(end_chunk + 1 == chunk);
1172
1173         /* try to expand this extent to cover @index */
1174         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1175
1176         /* don't go over the maximum extent size reported by server */
1177         if (end_index - ext->oe_start + 1 > cli->cl_max_extent_pages)
1178                 GOTO(out, rc = -ERANGE);
1179
1180         next = next_extent(ext);
1181         if (next != NULL && next->oe_start <= end_index)
1182                 /* complex mode - overlapped with the next extent,
1183                  * this case will be handled by osc_extent_find() */
1184                 GOTO(out, rc = -EAGAIN);
1185
1186         ext->oe_end = end_index;
1187         ext->oe_grants += chunksize;
1188         LASSERT(*grants >= chunksize);
1189         *grants -= chunksize;
1190         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1191                  "overlapped after expanding for %lu.\n", index);
1192         EXIT;
1193
1194 out:
1195         osc_object_unlock(obj);
1196         RETURN(rc);
1197 }
1198
1199 static void osc_extent_tree_dump0(int mask, struct osc_object *obj,
1200                                   const char *func, int line)
1201 {
1202         struct osc_extent *ext;
1203         int cnt;
1204
1205         if (!cfs_cdebug_show(mask, DEBUG_SUBSYSTEM))
1206                 return;
1207
1208         CDEBUG(mask, "Dump object %p extents at %s:%d, mppr: %u.\n",
1209                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1210
1211         /* osc_object_lock(obj); */
1212         cnt = 1;
1213         for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1214                 OSC_EXTENT_DUMP(mask, ext, "in tree %d.\n", cnt++);
1215
1216         cnt = 1;
1217         list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1218                 OSC_EXTENT_DUMP(mask, ext, "hp %d.\n", cnt++);
1219
1220         cnt = 1;
1221         list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1222                 OSC_EXTENT_DUMP(mask, ext, "urgent %d.\n", cnt++);
1223
1224         cnt = 1;
1225         list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1226                 OSC_EXTENT_DUMP(mask, ext, "reading %d.\n", cnt++);
1227         /* osc_object_unlock(obj); */
1228 }
1229
1230 /* ------------------ osc extent end ------------------ */
1231
1232 static inline int osc_is_ready(struct osc_object *osc)
1233 {
1234         return !list_empty(&osc->oo_ready_item) ||
1235                !list_empty(&osc->oo_hp_ready_item);
1236 }
1237
1238 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1239         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1240                (OSC), osc_is_ready(OSC),                                       \
1241                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1242                list_empty_marker(&(OSC)->oo_ready_item),                       \
1243                atomic_read(&(OSC)->oo_nr_writes),                              \
1244                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1245                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1246                atomic_read(&(OSC)->oo_nr_reads),                               \
1247                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1248                ##args)
1249
1250 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1251                           int cmd)
1252 {
1253         struct osc_page *opg  = oap2osc_page(oap);
1254         struct cl_page  *page = oap2cl_page(oap);
1255         int result;
1256
1257         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1258
1259         ENTRY;
1260         result = cl_page_make_ready(env, page, CRT_WRITE);
1261         if (result == 0)
1262                 opg->ops_submit_time = ktime_get();
1263         RETURN(result);
1264 }
1265
1266 static int osc_refresh_count(const struct lu_env *env,
1267                              struct osc_async_page *oap, int cmd)
1268 {
1269         struct osc_page  *opg = oap2osc_page(oap);
1270         pgoff_t index = osc_index(oap2osc(oap));
1271         struct cl_object *obj;
1272         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
1273         int result;
1274         loff_t kms;
1275
1276         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1277         LASSERT(!(cmd & OBD_BRW_READ));
1278         LASSERT(opg != NULL);
1279         obj = opg->ops_cl.cpl_obj;
1280
1281         cl_object_attr_lock(obj);
1282         result = cl_object_attr_get(env, obj, attr);
1283         cl_object_attr_unlock(obj);
1284         if (result < 0)
1285                 return result;
1286         kms = attr->cat_kms;
1287         if (cl_offset(obj, index) >= kms)
1288                 /* catch race with truncate */
1289                 return 0;
1290         else if (cl_offset(obj, index + 1) > kms)
1291                 /* catch sub-page write at end of file */
1292                 return kms & ~PAGE_MASK;
1293         else
1294                 return PAGE_SIZE;
1295 }
1296
1297 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1298                           int cmd, int rc)
1299 {
1300         struct osc_page   *opg  = oap2osc_page(oap);
1301         struct cl_page    *page = oap2cl_page(oap);
1302         enum cl_req_type   crt;
1303         int srvlock;
1304
1305         ENTRY;
1306
1307         cmd &= ~OBD_BRW_NOQUOTA;
1308         LASSERTF(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ),
1309                  "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1310         LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1311                 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1312         LASSERT(opg->ops_transfer_pinned);
1313
1314         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1315         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1316         opg->ops_transfer_pinned = 0;
1317
1318         opg->ops_submit_time = ktime_set(0, 0);
1319         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1320
1321         /* statistic */
1322         if (rc == 0 && srvlock) {
1323                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1324                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1325                 size_t bytes = oap->oap_count;
1326
1327                 if (crt == CRT_READ)
1328                         stats->os_lockless_reads += bytes;
1329                 else
1330                         stats->os_lockless_writes += bytes;
1331         }
1332
1333         /*
1334          * This has to be the last operation with the page, as locks are
1335          * released in cl_page_completion() and nothing except for the
1336          * reference counter protects page from concurrent reclaim.
1337          */
1338         lu_ref_del(&page->cp_reference, "transfer", page);
1339
1340         cl_page_completion(env, page, crt, rc);
1341         cl_page_put(env, page);
1342
1343         RETURN(0);
1344 }
1345
1346 #define OSC_DUMP_GRANT(mask, cli, fmt, args...) do {                    \
1347         struct client_obd *__tmp = (cli);                               \
1348         CDEBUG(mask, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu " \
1349                "dropped: %ld avail: %ld, dirty_grant: %ld, "            \
1350                "reserved: %ld, flight: %d } lru {in list: %ld, "        \
1351                "left: %ld, waiters: %d }" fmt "\n",                     \
1352                cli_name(__tmp),                                         \
1353                __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages,        \
1354                atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1355                __tmp->cl_lost_grant, __tmp->cl_avail_grant,             \
1356                __tmp->cl_dirty_grant,                                   \
1357                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,         \
1358                atomic_long_read(&__tmp->cl_lru_in_list),                \
1359                atomic_long_read(&__tmp->cl_lru_busy),                   \
1360                atomic_read(&__tmp->cl_lru_shrinkers), ##args);          \
1361 } while (0)
1362
1363 /* caller must hold loi_list_lock */
1364 static void osc_consume_write_grant(struct client_obd *cli,
1365                                     struct brw_page *pga)
1366 {
1367         assert_spin_locked(&cli->cl_loi_list_lock);
1368         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1369         cli->cl_dirty_pages++;
1370         pga->flag |= OBD_BRW_FROM_GRANT;
1371         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1372                PAGE_SIZE, pga, pga->pg);
1373         osc_update_next_shrink(cli);
1374 }
1375
1376 /* the companion to osc_consume_write_grant, called when a brw has completed.
1377  * must be called with the loi lock held. */
1378 static void osc_release_write_grant(struct client_obd *cli,
1379                                     struct brw_page *pga)
1380 {
1381         ENTRY;
1382
1383         assert_spin_locked(&cli->cl_loi_list_lock);
1384         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1385                 EXIT;
1386                 return;
1387         }
1388
1389         pga->flag &= ~OBD_BRW_FROM_GRANT;
1390         atomic_long_dec(&obd_dirty_pages);
1391         cli->cl_dirty_pages--;
1392         EXIT;
1393 }
1394
1395 /**
1396  * To avoid sleeping with object lock held, it's good for us allocate enough
1397  * grants before entering into critical section.
1398  *
1399  * client_obd_list_lock held by caller
1400  */
1401 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1402 {
1403         int rc = -EDQUOT;
1404
1405         if (cli->cl_avail_grant >= bytes) {
1406                 cli->cl_avail_grant    -= bytes;
1407                 cli->cl_reserved_grant += bytes;
1408                 rc = 0;
1409         }
1410         return rc;
1411 }
1412
1413 static void __osc_unreserve_grant(struct client_obd *cli,
1414                                   unsigned int reserved, unsigned int unused)
1415 {
1416         /* it's quite normal for us to get more grant than reserved.
1417          * Thinking about a case that two extents merged by adding a new
1418          * chunk, we can save one extent tax. If extent tax is greater than
1419          * one chunk, we can save more grant by adding a new chunk */
1420         cli->cl_reserved_grant -= reserved;
1421         if (unused > reserved) {
1422                 cli->cl_avail_grant += reserved;
1423                 cli->cl_lost_grant  += unused - reserved;
1424                 cli->cl_dirty_grant -= unused - reserved;
1425         } else {
1426                 cli->cl_avail_grant += unused;
1427                 cli->cl_dirty_grant += reserved - unused;
1428         }
1429 }
1430
1431 static void osc_unreserve_grant_nolock(struct client_obd *cli,
1432                                        unsigned int reserved,
1433                                        unsigned int unused)
1434 {
1435         __osc_unreserve_grant(cli, reserved, unused);
1436         if (unused > 0)
1437                 osc_wake_cache_waiters(cli);
1438 }
1439
1440 static void osc_unreserve_grant(struct client_obd *cli,
1441                                 unsigned int reserved, unsigned int unused)
1442 {
1443         spin_lock(&cli->cl_loi_list_lock);
1444         osc_unreserve_grant_nolock(cli, reserved, unused);
1445         spin_unlock(&cli->cl_loi_list_lock);
1446 }
1447
1448 /**
1449  * Free grant after IO is finished or canceled.
1450  *
1451  * @lost_grant is used to remember how many grants we have allocated but not
1452  * used, we should return these grants to OST. There're two cases where grants
1453  * can be lost:
1454  * 1. truncate;
1455  * 2. blocksize at OST is less than PAGE_SIZE and a partial page was
1456  *    written. In this case OST may use less chunks to serve this partial
1457  *    write. OSTs don't actually know the page size on the client side. so
1458  *    clients have to calculate lost grant by the blocksize on the OST.
1459  *    See filter_grant_check() for details.
1460  */
1461 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1462                            unsigned int lost_grant, unsigned int dirty_grant)
1463 {
1464         unsigned long grant;
1465
1466         grant = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
1467
1468         spin_lock(&cli->cl_loi_list_lock);
1469         atomic_long_sub(nr_pages, &obd_dirty_pages);
1470         cli->cl_dirty_pages -= nr_pages;
1471         cli->cl_lost_grant += lost_grant;
1472         cli->cl_dirty_grant -= dirty_grant;
1473         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1474                 /* borrow some grant from truncate to avoid the case that
1475                  * truncate uses up all avail grant */
1476                 cli->cl_lost_grant -= grant;
1477                 cli->cl_avail_grant += grant;
1478         }
1479         osc_wake_cache_waiters(cli);
1480         spin_unlock(&cli->cl_loi_list_lock);
1481         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu/%lu\n",
1482                lost_grant, cli->cl_lost_grant,
1483                cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_SHIFT,
1484                cli->cl_dirty_grant);
1485 }
1486
1487 /**
1488  * The companion to osc_enter_cache(), called when @oap is no longer part of
1489  * the dirty accounting due to error.
1490  */
1491 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1492 {
1493         spin_lock(&cli->cl_loi_list_lock);
1494         osc_release_write_grant(cli, &oap->oap_brw_page);
1495         spin_unlock(&cli->cl_loi_list_lock);
1496 }
1497
1498 /**
1499  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1500  * is available.
1501  */
1502 static int osc_enter_cache_try(struct client_obd *cli,
1503                                struct osc_async_page *oap,
1504                                int bytes)
1505 {
1506         int rc;
1507
1508         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1509
1510         rc = osc_reserve_grant(cli, bytes);
1511         if (rc < 0)
1512                 return 0;
1513
1514         if (cli->cl_dirty_pages < cli->cl_dirty_max_pages) {
1515                 if (atomic_long_add_return(1, &obd_dirty_pages) <=
1516                     obd_max_dirty_pages) {
1517                         osc_consume_write_grant(cli, &oap->oap_brw_page);
1518                         rc = 1;
1519                         goto out;
1520                 } else
1521                         atomic_long_dec(&obd_dirty_pages);
1522         }
1523         __osc_unreserve_grant(cli, bytes, bytes);
1524
1525 out:
1526         return rc;
1527 }
1528
1529 /* Following two inlines exist to pass code fragments
1530  * to wait_event_idle_exclusive_timeout_cmd().  Passing
1531  * code fragments as macro args can look confusing, so
1532  * we provide inlines to encapsulate them.
1533  */
1534 static inline void cli_unlock_and_unplug(const struct lu_env *env,
1535                                          struct client_obd *cli,
1536                                          struct osc_async_page *oap)
1537 {
1538         spin_unlock(&cli->cl_loi_list_lock);
1539         osc_io_unplug_async(env, cli, NULL);
1540         CDEBUG(D_CACHE,
1541                "%s: sleeping for cache space for %p\n",
1542                cli_name(cli), oap);
1543 }
1544
1545 static inline void cli_lock_after_unplug(struct client_obd *cli)
1546 {
1547         spin_lock(&cli->cl_loi_list_lock);
1548 }
1549 /**
1550  * The main entry to reserve dirty page accounting. Usually the grant reserved
1551  * in this function will be freed in bulk in osc_free_grant() unless it fails
1552  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1553  *
1554  * The process will be put into sleep if it's already run out of grant.
1555  */
1556 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1557                            struct osc_async_page *oap, int bytes)
1558 {
1559         struct osc_object *osc = oap->oap_obj;
1560         struct lov_oinfo *loi = osc->oo_oinfo;
1561         int rc = -EDQUOT;
1562         int remain;
1563         bool entered = false;
1564         /* We cannot wait for a long time here since we are holding ldlm lock
1565          * across the actual IO. If no requests complete fast (e.g. due to
1566          * overloaded OST that takes a long time to process everything, we'd
1567          * get evicted if we wait for a normal obd_timeout or some such.
1568          * So we try to wait half the time it would take the client to be
1569          * evicted by server which is half obd_timeout when AT is off
1570          * or at least ldlm_enqueue_min with AT on.
1571          * See LU-13131 */
1572         unsigned long timeout = cfs_time_seconds(AT_OFF ? obd_timeout / 2 :
1573                                                           ldlm_enqueue_min / 2);
1574
1575         ENTRY;
1576
1577         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1578
1579         spin_lock(&cli->cl_loi_list_lock);
1580
1581         /* force the caller to try sync io.  this can jump the list
1582          * of queued writes and create a discontiguous rpc stream */
1583         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1584             cli->cl_dirty_max_pages == 0 ||
1585             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) {
1586                 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1587                 GOTO(out, rc = -EDQUOT);
1588         }
1589
1590         /*
1591          * We can wait here for two reasons: too many dirty pages in cache, or
1592          * run out of grants. In both cases we should write dirty pages out.
1593          * Adding a cache waiter will trigger urgent write-out no matter what
1594          * RPC size will be.
1595          * The exiting condition (other than success) is no avail grants
1596          * and no dirty pages caching, that really means there is no space
1597          * on the OST.
1598          */
1599         remain = wait_event_idle_exclusive_timeout_cmd(
1600                 cli->cl_cache_waiters,
1601                 (entered = osc_enter_cache_try(cli, oap, bytes)) ||
1602                 (cli->cl_dirty_pages == 0 && cli->cl_w_in_flight == 0),
1603                 timeout,
1604                 cli_unlock_and_unplug(env, cli, oap),
1605                 cli_lock_after_unplug(cli));
1606
1607         if (entered) {
1608                 if (remain == timeout)
1609                         OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1610                 else
1611                         OSC_DUMP_GRANT(D_CACHE, cli,
1612                                        "finally got grant space\n");
1613                 wake_up(&cli->cl_cache_waiters);
1614                 rc = 0;
1615         } else if (remain == 0) {
1616                 OSC_DUMP_GRANT(D_CACHE, cli,
1617                                "timeout, fall back to sync i/o\n");
1618                 osc_extent_tree_dump(D_CACHE, osc);
1619                 /* fall back to synchronous I/O */
1620         } else {
1621                 OSC_DUMP_GRANT(D_CACHE, cli,
1622                                "no grant space, fall back to sync i/o\n");
1623                 wake_up_all(&cli->cl_cache_waiters);
1624         }
1625         EXIT;
1626 out:
1627         spin_unlock(&cli->cl_loi_list_lock);
1628         RETURN(rc);
1629 }
1630
1631 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1632 {
1633         int hprpc = !!list_empty(&osc->oo_hp_exts);
1634         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1635 }
1636
1637 /* This maintains the lists of pending pages to read/write for a given object
1638  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1639  * to quickly find objects that are ready to send an RPC. */
1640 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1641                          int cmd)
1642 {
1643         int invalid_import = 0;
1644         ENTRY;
1645
1646         /* if we have an invalid import we want to drain the queued pages
1647          * by forcing them through rpcs that immediately fail and complete
1648          * the pages.  recovery relies on this to empty the queued pages
1649          * before canceling the locks and evicting down the llite pages */
1650         if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1651                 invalid_import = 1;
1652
1653         if (cmd & OBD_BRW_WRITE) {
1654                 if (atomic_read(&osc->oo_nr_writes) == 0)
1655                         RETURN(0);
1656                 if (invalid_import) {
1657                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1658                         RETURN(1);
1659                 }
1660                 if (!list_empty(&osc->oo_hp_exts)) {
1661                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1662                         RETURN(1);
1663                 }
1664                 if (!list_empty(&osc->oo_urgent_exts)) {
1665                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1666                         RETURN(1);
1667                 }
1668                 /* trigger a write rpc stream as long as there are dirtiers
1669                  * waiting for space.  as they're waiting, they're not going to
1670                  * create more pages to coalesce with what's waiting..
1671                  */
1672                 if (waitqueue_active(&cli->cl_cache_waiters)) {
1673                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1674                         RETURN(1);
1675                 }
1676                 if (!list_empty(&osc->oo_full_exts)) {
1677                         CDEBUG(D_CACHE, "full extent ready, make an RPC\n");
1678                         RETURN(1);
1679                 }
1680         } else {
1681                 if (atomic_read(&osc->oo_nr_reads) == 0)
1682                         RETURN(0);
1683                 if (invalid_import) {
1684                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1685                         RETURN(1);
1686                 }
1687                 /* all read are urgent. */
1688                 if (!list_empty(&osc->oo_reading_exts))
1689                         RETURN(1);
1690         }
1691
1692         RETURN(0);
1693 }
1694
1695 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1696 {
1697         struct client_obd *cli = osc_cli(obj);
1698         if (cmd & OBD_BRW_WRITE) {
1699                 atomic_add(delta, &obj->oo_nr_writes);
1700                 atomic_add(delta, &cli->cl_pending_w_pages);
1701                 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1702         } else {
1703                 atomic_add(delta, &obj->oo_nr_reads);
1704                 atomic_add(delta, &cli->cl_pending_r_pages);
1705                 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1706         }
1707         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1708 }
1709
1710 static int osc_makes_hprpc(struct osc_object *obj)
1711 {
1712         return !list_empty(&obj->oo_hp_exts);
1713 }
1714
1715 static void on_list(struct list_head *item, struct list_head *list,
1716                     int should_be_on)
1717 {
1718         if (list_empty(item) && should_be_on)
1719                 list_add_tail(item, list);
1720         else if (!list_empty(item) && !should_be_on)
1721                 list_del_init(item);
1722 }
1723
1724 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1725  * can find pages to build into rpcs quickly */
1726 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1727 {
1728         if (osc_makes_hprpc(osc)) {
1729                 /* HP rpc */
1730                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1731                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1732         } else {
1733                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1734                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1735                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1736                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1737         }
1738
1739         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1740                 atomic_read(&osc->oo_nr_writes) > 0);
1741
1742         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1743                 atomic_read(&osc->oo_nr_reads) > 0);
1744
1745         return osc_is_ready(osc);
1746 }
1747
1748 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1749 {
1750         int is_ready;
1751
1752         spin_lock(&cli->cl_loi_list_lock);
1753         is_ready = __osc_list_maint(cli, osc);
1754         spin_unlock(&cli->cl_loi_list_lock);
1755
1756         return is_ready;
1757 }
1758
1759 /* this is trying to propogate async writeback errors back up to the
1760  * application.  As an async write fails we record the error code for later if
1761  * the app does an fsync.  As long as errors persist we force future rpcs to be
1762  * sync so that the app can get a sync error and break the cycle of queueing
1763  * pages for which writeback will fail. */
1764 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1765                            int rc)
1766 {
1767         if (rc) {
1768                 if (!ar->ar_rc)
1769                         ar->ar_rc = rc;
1770
1771                 ar->ar_force_sync = 1;
1772                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1773                 return;
1774
1775         }
1776
1777         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1778                 ar->ar_force_sync = 0;
1779 }
1780
1781 /* this must be called holding the loi list lock to give coverage to exit_cache,
1782  * async_flag maintenance, and oap_request */
1783 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1784                               struct osc_async_page *oap, int sent, int rc)
1785 {
1786         struct osc_object *osc = oap->oap_obj;
1787         struct lov_oinfo  *loi = osc->oo_oinfo;
1788         __u64 xid = 0;
1789
1790         ENTRY;
1791         if (oap->oap_request != NULL) {
1792                 xid = ptlrpc_req_xid(oap->oap_request);
1793                 ptlrpc_req_finished(oap->oap_request);
1794                 oap->oap_request = NULL;
1795         }
1796
1797         /* As the transfer for this page is being done, clear the flags */
1798         spin_lock(&oap->oap_lock);
1799         oap->oap_async_flags = 0;
1800         spin_unlock(&oap->oap_lock);
1801
1802         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1803                 spin_lock(&cli->cl_loi_list_lock);
1804                 osc_process_ar(&cli->cl_ar, xid, rc);
1805                 osc_process_ar(&loi->loi_ar, xid, rc);
1806                 spin_unlock(&cli->cl_loi_list_lock);
1807         }
1808
1809         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1810         if (rc)
1811                 CERROR("completion on oap %p obj %p returns %d.\n",
1812                        oap, osc, rc);
1813
1814         EXIT;
1815 }
1816
1817 struct extent_rpc_data {
1818         struct list_head        *erd_rpc_list;
1819         unsigned int            erd_page_count;
1820         unsigned int            erd_max_pages;
1821         unsigned int            erd_max_chunks;
1822         unsigned int            erd_max_extents;
1823 };
1824
1825 static inline unsigned osc_extent_chunks(const struct osc_extent *ext)
1826 {
1827         struct client_obd *cli = osc_cli(ext->oe_obj);
1828         unsigned ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1829
1830         return (ext->oe_end >> ppc_bits) - (ext->oe_start >> ppc_bits) + 1;
1831 }
1832
1833 static inline bool
1834 can_merge(const struct osc_extent *ext, const struct osc_extent *in_rpc)
1835 {
1836         if (ext->oe_no_merge || in_rpc->oe_no_merge)
1837                 return false;
1838
1839         if (ext->oe_srvlock != in_rpc->oe_srvlock)
1840                 return false;
1841
1842         if (ext->oe_ndelay != in_rpc->oe_ndelay)
1843                 return false;
1844
1845         if (!ext->oe_grants != !in_rpc->oe_grants)
1846                 return false;
1847
1848         if (ext->oe_dio != in_rpc->oe_dio)
1849                 return false;
1850
1851         /* It's possible to have overlap on DIO */
1852         if (in_rpc->oe_dio && overlapped(ext, in_rpc))
1853                 return false;
1854
1855         if (ext->oe_is_rdma_only != in_rpc->oe_is_rdma_only)
1856                 return false;
1857
1858         return true;
1859 }
1860
1861 /**
1862  * Try to add extent to one RPC. We need to think about the following things:
1863  * - # of pages must not be over max_pages_per_rpc
1864  * - extent must be compatible with previous ones
1865  */
1866 static int try_to_add_extent_for_io(struct client_obd *cli,
1867                                     struct osc_extent *ext,
1868                                     struct extent_rpc_data *data)
1869 {
1870         struct osc_extent *tmp;
1871         unsigned int chunk_count;
1872         ENTRY;
1873
1874         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1875                 ext);
1876         OSC_EXTENT_DUMP(D_CACHE, ext, "trying to add this extent\n");
1877
1878         if (data->erd_max_extents == 0)
1879                 RETURN(0);
1880
1881         chunk_count = osc_extent_chunks(ext);
1882         EASSERTF(data->erd_page_count != 0 ||
1883                  chunk_count <= data->erd_max_chunks, ext,
1884                  "The first extent to be fit in a RPC contains %u chunks, "
1885                  "which is over the limit %u.\n", chunk_count,
1886                  data->erd_max_chunks);
1887         if (chunk_count > data->erd_max_chunks)
1888                 RETURN(0);
1889
1890         data->erd_max_pages = max(ext->oe_mppr, data->erd_max_pages);
1891         EASSERTF(data->erd_page_count != 0 ||
1892                 ext->oe_nr_pages <= data->erd_max_pages, ext,
1893                 "The first extent to be fit in a RPC contains %u pages, "
1894                 "which is over the limit %u.\n", ext->oe_nr_pages,
1895                 data->erd_max_pages);
1896         if (data->erd_page_count + ext->oe_nr_pages > data->erd_max_pages)
1897                 RETURN(0);
1898
1899         list_for_each_entry(tmp, data->erd_rpc_list, oe_link) {
1900                 EASSERT(tmp->oe_owner == current, tmp);
1901
1902                 if (!can_merge(ext, tmp))
1903                         RETURN(0);
1904         }
1905
1906         data->erd_max_extents--;
1907         data->erd_max_chunks -= chunk_count;
1908         data->erd_page_count += ext->oe_nr_pages;
1909         list_move_tail(&ext->oe_link, data->erd_rpc_list);
1910         ext->oe_owner = current;
1911         RETURN(1);
1912 }
1913
1914 /**
1915  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1916  * get_write_extent() takes all appropriate extents in atomic.
1917  *
1918  * The following policy is used to collect extents for IO:
1919  * 1. Add as many HP extents as possible;
1920  * 2. Add the first urgent extent in urgent extent list and take it out of
1921  *    urgent list;
1922  * 3. Add subsequent extents of this urgent extent;
1923  * 4. If urgent list is not empty, goto 2;
1924  * 5. Traverse the extent tree from the 1st extent;
1925  * 6. Above steps exit if there is no space in this RPC.
1926  */
1927 static unsigned int get_write_extents(struct osc_object *obj,
1928                                       struct list_head *rpclist)
1929 {
1930         struct client_obd *cli = osc_cli(obj);
1931         struct osc_extent *ext;
1932         struct extent_rpc_data data = {
1933                 .erd_rpc_list   = rpclist,
1934                 .erd_page_count = 0,
1935                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
1936                 .erd_max_chunks = osc_max_write_chunks(cli),
1937                 .erd_max_extents = 256,
1938         };
1939
1940         assert_osc_object_is_locked(obj);
1941         while (!list_empty(&obj->oo_hp_exts)) {
1942                 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
1943                                  oe_link);
1944                 if (!try_to_add_extent_for_io(cli, ext, &data))
1945                         return data.erd_page_count;
1946                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
1947         }
1948         if (data.erd_page_count == data.erd_max_pages)
1949                 return data.erd_page_count;
1950
1951         while (!list_empty(&obj->oo_urgent_exts)) {
1952                 ext = list_entry(obj->oo_urgent_exts.next,
1953                                  struct osc_extent, oe_link);
1954                 if (!try_to_add_extent_for_io(cli, ext, &data))
1955                         return data.erd_page_count;
1956         }
1957         if (data.erd_page_count == data.erd_max_pages)
1958                 return data.erd_page_count;
1959
1960         /* One key difference between full extents and other extents: full
1961          * extents can usually only be added if the rpclist was empty, so if we
1962          * can't add one, we continue on to trying to add normal extents.  This
1963          * is so we don't miss adding extra extents to an RPC containing high
1964          * priority or urgent extents. */
1965         while (!list_empty(&obj->oo_full_exts)) {
1966                 ext = list_entry(obj->oo_full_exts.next,
1967                                  struct osc_extent, oe_link);
1968                 if (!try_to_add_extent_for_io(cli, ext, &data))
1969                         break;
1970         }
1971         if (data.erd_page_count == data.erd_max_pages)
1972                 return data.erd_page_count;
1973
1974         for (ext = first_extent(obj);
1975              ext;
1976              ext = next_extent(ext)) {
1977                 if ((ext->oe_state != OES_CACHE) ||
1978                     /* this extent may be already in current rpclist */
1979                     (!list_empty(&ext->oe_link) && ext->oe_owner))
1980                         continue;
1981
1982                 if (!try_to_add_extent_for_io(cli, ext, &data))
1983                         return data.erd_page_count;
1984         }
1985         return data.erd_page_count;
1986 }
1987
1988 static int
1989 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
1990                    struct osc_object *osc)
1991 __must_hold(osc)
1992 {
1993         LIST_HEAD(rpclist);
1994         struct osc_extent *ext;
1995         struct osc_extent *tmp;
1996         struct osc_extent *first = NULL;
1997         unsigned int page_count = 0;
1998         int srvlock = 0;
1999         int rc = 0;
2000         ENTRY;
2001
2002         assert_osc_object_is_locked(osc);
2003
2004         page_count = get_write_extents(osc, &rpclist);
2005         LASSERT(equi(page_count == 0, list_empty(&rpclist)));
2006
2007         if (list_empty(&rpclist))
2008                 RETURN(0);
2009
2010         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2011
2012         list_for_each_entry(ext, &rpclist, oe_link) {
2013                 LASSERT(ext->oe_state == OES_CACHE ||
2014                         ext->oe_state == OES_LOCK_DONE);
2015                 if (ext->oe_state == OES_CACHE)
2016                         osc_extent_state_set(ext, OES_LOCKING);
2017                 else
2018                         osc_extent_state_set(ext, OES_RPC);
2019         }
2020
2021         /* we're going to grab page lock, so release object lock because
2022          * lock order is page lock -> object lock. */
2023         osc_object_unlock(osc);
2024
2025         list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2026                 if (ext->oe_state == OES_LOCKING) {
2027                         rc = osc_extent_make_ready(env, ext);
2028                         if (unlikely(rc < 0)) {
2029                                 list_del_init(&ext->oe_link);
2030                                 osc_extent_finish(env, ext, 0, rc);
2031                                 continue;
2032                         }
2033                 }
2034                 if (first == NULL) {
2035                         first = ext;
2036                         srvlock = ext->oe_srvlock;
2037                 } else {
2038                         LASSERT(srvlock == ext->oe_srvlock);
2039                 }
2040         }
2041
2042         if (!list_empty(&rpclist)) {
2043                 LASSERT(page_count > 0);
2044                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2045                 LASSERT(list_empty(&rpclist));
2046         }
2047
2048         osc_object_lock(osc);
2049         RETURN(rc);
2050 }
2051
2052 /**
2053  * prepare pages for ASYNC io and put pages in send queue.
2054  *
2055  * \param cmd OBD_BRW_* macroses
2056  * \param lop pending pages
2057  *
2058  * \return zero if no page added to send queue.
2059  * \return 1 if pages successfully added to send queue.
2060  * \return negative on errors.
2061  */
2062 static int
2063 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2064                   struct osc_object *osc)
2065 __must_hold(osc)
2066 {
2067         struct osc_extent *ext;
2068         struct osc_extent *next;
2069         LIST_HEAD(rpclist);
2070         struct extent_rpc_data data = {
2071                 .erd_rpc_list   = &rpclist,
2072                 .erd_page_count = 0,
2073                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
2074                 .erd_max_chunks = UINT_MAX,
2075                 .erd_max_extents = UINT_MAX,
2076         };
2077         int rc = 0;
2078         ENTRY;
2079
2080         assert_osc_object_is_locked(osc);
2081         list_for_each_entry_safe(ext, next, &osc->oo_reading_exts, oe_link) {
2082                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2083                 if (!try_to_add_extent_for_io(cli, ext, &data))
2084                         break;
2085                 osc_extent_state_set(ext, OES_RPC);
2086                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2087         }
2088         LASSERT(data.erd_page_count <= data.erd_max_pages);
2089
2090         osc_update_pending(osc, OBD_BRW_READ, -data.erd_page_count);
2091
2092         if (!list_empty(&rpclist)) {
2093                 osc_object_unlock(osc);
2094
2095                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2096                 LASSERT(list_empty(&rpclist));
2097
2098                 osc_object_lock(osc);
2099         }
2100         RETURN(rc);
2101 }
2102
2103 #define list_to_obj(list, item) ({                                            \
2104         struct list_head *__tmp = (list)->next;                               \
2105         list_del_init(__tmp);                                         \
2106         list_entry(__tmp, struct osc_object, oo_##item);                      \
2107 })
2108
2109 /* This is called by osc_check_rpcs() to find which objects have pages that
2110  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2111 static struct osc_object *osc_next_obj(struct client_obd *cli)
2112 {
2113         ENTRY;
2114
2115         /* First return objects that have blocked locks so that they
2116          * will be flushed quickly and other clients can get the lock,
2117          * then objects which have pages ready to be stuffed into RPCs */
2118         if (!list_empty(&cli->cl_loi_hp_ready_list))
2119                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2120         if (!list_empty(&cli->cl_loi_ready_list))
2121                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2122
2123         /* then if we have cache waiters, return all objects with queued
2124          * writes.  This is especially important when many small files
2125          * have filled up the cache and not been fired into rpcs because
2126          * they don't pass the nr_pending/object threshhold
2127          */
2128         if (waitqueue_active(&cli->cl_cache_waiters) &&
2129             !list_empty(&cli->cl_loi_write_list))
2130                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2131
2132         /* then return all queued objects when we have an invalid import
2133          * so that they get flushed */
2134         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2135                 if (!list_empty(&cli->cl_loi_write_list))
2136                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2137                                            write_item));
2138                 if (!list_empty(&cli->cl_loi_read_list))
2139                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2140                                            read_item));
2141         }
2142         RETURN(NULL);
2143 }
2144
2145 /* called with the loi list lock held */
2146 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2147 __must_hold(&cli->cl_loi_list_lock)
2148 {
2149         struct osc_object *osc;
2150         int rc = 0;
2151         ENTRY;
2152
2153         while ((osc = osc_next_obj(cli)) != NULL) {
2154                 struct cl_object *obj = osc2cl(osc);
2155                 struct lu_ref_link link;
2156
2157                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2158
2159                 /* even if we have reached our max in flight RPCs, we still
2160                  * allow all high-priority RPCs through to prevent their
2161                  * starvation and leading to server evicting us for not
2162                  * writing out pages in a timely manner LU-13131 */
2163                 if (osc_max_rpc_in_flight(cli, osc) &&
2164                     list_empty(&osc->oo_hp_exts)) {
2165                         __osc_list_maint(cli, osc);
2166                         break;
2167                 }
2168
2169                 cl_object_get(obj);
2170                 spin_unlock(&cli->cl_loi_list_lock);
2171                 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2172
2173                 /* attempt some read/write balancing by alternating between
2174                  * reads and writes in an object.  The makes_rpc checks here
2175                  * would be redundant if we were getting read/write work items
2176                  * instead of objects.  we don't want send_oap_rpc to drain a
2177                  * partial read pending queue when we're given this object to
2178                  * do io on writes while there are cache waiters */
2179                 osc_object_lock(osc);
2180                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2181                         rc = osc_send_write_rpc(env, cli, osc);
2182                         if (rc < 0) {
2183                                 CERROR("Write request failed with %d\n", rc);
2184
2185                                 /* osc_send_write_rpc failed, mostly because of
2186                                  * memory pressure.
2187                                  *
2188                                  * It can't break here, because if:
2189                                  *  - a page was submitted by osc_io_submit, so
2190                                  *    page locked;
2191                                  *  - no request in flight
2192                                  *  - no subsequent request
2193                                  * The system will be in live-lock state,
2194                                  * because there is no chance to call
2195                                  * osc_io_unplug() and osc_check_rpcs() any
2196                                  * more. pdflush can't help in this case,
2197                                  * because it might be blocked at grabbing
2198                                  * the page lock as we mentioned.
2199                                  *
2200                                  * Anyway, continue to drain pages. */
2201                                 /* break; */
2202                         }
2203                 }
2204                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2205                         rc = osc_send_read_rpc(env, cli, osc);
2206                         if (rc < 0)
2207                                 CERROR("Read request failed with %d\n", rc);
2208                 }
2209                 osc_object_unlock(osc);
2210
2211                 osc_list_maint(cli, osc);
2212                 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2213                 cl_object_put(env, obj);
2214
2215                 spin_lock(&cli->cl_loi_list_lock);
2216         }
2217 }
2218
2219 int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2220                    struct osc_object *osc, int async)
2221 {
2222         int rc = 0;
2223
2224         if (osc != NULL && osc_list_maint(cli, osc) == 0)
2225                 return 0;
2226
2227         if (!async) {
2228                 spin_lock(&cli->cl_loi_list_lock);
2229                 osc_check_rpcs(env, cli);
2230                 spin_unlock(&cli->cl_loi_list_lock);
2231         } else {
2232                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2233                 LASSERT(cli->cl_writeback_work != NULL);
2234                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2235         }
2236         return rc;
2237 }
2238 EXPORT_SYMBOL(osc_io_unplug0);
2239
2240 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2241                         struct page *page, loff_t offset)
2242 {
2243         struct obd_export     *exp = osc_export(osc);
2244         struct osc_async_page *oap = &ops->ops_oap;
2245         ENTRY;
2246
2247         if (!page)
2248                 return cfs_size_round(sizeof(*oap));
2249
2250         oap->oap_magic = OAP_MAGIC;
2251         oap->oap_cli = &exp->exp_obd->u.cli;
2252         oap->oap_obj = osc;
2253
2254         oap->oap_page = page;
2255         oap->oap_obj_off = offset;
2256         LASSERT(!(offset & ~PAGE_MASK));
2257
2258         INIT_LIST_HEAD(&oap->oap_pending_item);
2259         INIT_LIST_HEAD(&oap->oap_rpc_item);
2260
2261         spin_lock_init(&oap->oap_lock);
2262         CDEBUG(D_INFO, "oap %p page %p obj off %llu\n",
2263                oap, page, oap->oap_obj_off);
2264         RETURN(0);
2265 }
2266 EXPORT_SYMBOL(osc_prep_async_page);
2267
2268 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2269                        struct osc_page *ops, cl_commit_cbt cb)
2270 {
2271         struct osc_io *oio = osc_env_io(env);
2272         struct osc_extent     *ext = NULL;
2273         struct osc_async_page *oap = &ops->ops_oap;
2274         struct client_obd     *cli = oap->oap_cli;
2275         struct osc_object     *osc = oap->oap_obj;
2276         struct pagevec        *pvec = &osc_env_info(env)->oti_pagevec;
2277         pgoff_t index;
2278         unsigned int tmp;
2279         unsigned int grants = 0;
2280         u32    brw_flags = OBD_BRW_ASYNC;
2281         int    cmd = OBD_BRW_WRITE;
2282         int    need_release = 0;
2283         int    rc = 0;
2284         ENTRY;
2285
2286         if (oap->oap_magic != OAP_MAGIC)
2287                 RETURN(-EINVAL);
2288
2289         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2290                 RETURN(-EIO);
2291
2292         if (!list_empty(&oap->oap_pending_item) ||
2293             !list_empty(&oap->oap_rpc_item))
2294                 RETURN(-EBUSY);
2295
2296         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2297         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2298         if (oio->oi_cap_sys_resource || io->ci_noquota) {
2299                 brw_flags |= OBD_BRW_NOQUOTA;
2300                 cmd |= OBD_BRW_NOQUOTA;
2301         }
2302
2303         /* check if the file's owner/group is over quota */
2304         if (!(cmd & OBD_BRW_NOQUOTA)) {
2305                 struct cl_object *obj;
2306                 struct cl_attr   *attr;
2307                 unsigned int qid[LL_MAXQUOTAS];
2308
2309                 obj = cl_object_top(&osc->oo_cl);
2310                 attr = &osc_env_info(env)->oti_attr;
2311
2312                 cl_object_attr_lock(obj);
2313                 rc = cl_object_attr_get(env, obj, attr);
2314                 cl_object_attr_unlock(obj);
2315
2316                 qid[USRQUOTA] = attr->cat_uid;
2317                 qid[GRPQUOTA] = attr->cat_gid;
2318                 qid[PRJQUOTA] = attr->cat_projid;
2319                 if (rc == 0 && osc_quota_chkdq(cli, qid) == -EDQUOT)
2320                         rc = -EDQUOT;
2321                 if (rc)
2322                         RETURN(rc);
2323         }
2324
2325         oap->oap_cmd = cmd;
2326         oap->oap_page_off = ops->ops_from;
2327         oap->oap_count = ops->ops_to - ops->ops_from + 1;
2328         /* No need to hold a lock here,
2329          * since this page is not in any list yet. */
2330         oap->oap_async_flags = 0;
2331         oap->oap_brw_flags = brw_flags;
2332
2333         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2334                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2335
2336         index = osc_index(oap2osc(oap));
2337
2338         /* Add this page into extent by the following steps:
2339          * 1. if there exists an active extent for this IO, mostly this page
2340          *    can be added to the active extent and sometimes we need to
2341          *    expand extent to accomodate this page;
2342          * 2. otherwise, a new extent will be allocated. */
2343
2344         ext = oio->oi_active;
2345         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2346                 /* one chunk plus extent overhead must be enough to write this
2347                  * page */
2348                 grants = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2349                 if (ext->oe_end >= index)
2350                         grants = 0;
2351
2352                 /* it doesn't need any grant to dirty this page */
2353                 spin_lock(&cli->cl_loi_list_lock);
2354                 rc = osc_enter_cache_try(cli, oap, grants);
2355                 if (rc == 0) { /* try failed */
2356                         grants = 0;
2357                         need_release = 1;
2358                 } else if (ext->oe_end < index) {
2359                         tmp = grants;
2360                         /* try to expand this extent */
2361                         rc = osc_extent_expand(ext, index, &tmp);
2362                         if (rc < 0) {
2363                                 need_release = 1;
2364                                 /* don't free reserved grant */
2365                         } else {
2366                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2367                                                 "expanded for %lu.\n", index);
2368                                 osc_unreserve_grant_nolock(cli, grants, tmp);
2369                                 grants = 0;
2370                         }
2371                 }
2372                 spin_unlock(&cli->cl_loi_list_lock);
2373                 rc = 0;
2374         } else if (ext != NULL) {
2375                 /* index is located outside of active extent */
2376                 need_release = 1;
2377         }
2378         if (need_release) {
2379                 osc_extent_release(env, ext);
2380                 oio->oi_active = NULL;
2381                 ext = NULL;
2382         }
2383
2384         if (ext == NULL) {
2385                 tmp = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2386
2387                 /* try to find new extent to cover this page */
2388                 LASSERT(oio->oi_active == NULL);
2389                 /* we may have allocated grant for this page if we failed
2390                  * to expand the previous active extent. */
2391                 LASSERT(ergo(grants > 0, grants >= tmp));
2392
2393                 rc = 0;
2394                 if (grants == 0) {
2395                         /* We haven't allocated grant for this page, and we
2396                          * must not hold a page lock while we do enter_cache,
2397                          * so we must mark dirty & unlock any pages in the
2398                          * write commit pagevec. */
2399                         if (pagevec_count(pvec)) {
2400                                 cb(env, io, pvec);
2401                                 pagevec_reinit(pvec);
2402                         }
2403                         rc = osc_enter_cache(env, cli, oap, tmp);
2404                         if (rc == 0)
2405                                 grants = tmp;
2406                 }
2407
2408                 tmp = grants;
2409                 if (rc == 0) {
2410                         ext = osc_extent_find(env, osc, index, &tmp);
2411                         if (IS_ERR(ext)) {
2412                                 LASSERT(tmp == grants);
2413                                 osc_exit_cache(cli, oap);
2414                                 rc = PTR_ERR(ext);
2415                                 ext = NULL;
2416                         } else {
2417                                 oio->oi_active = ext;
2418                         }
2419                 }
2420                 if (grants > 0)
2421                         osc_unreserve_grant(cli, grants, tmp);
2422         }
2423
2424         LASSERT(ergo(rc == 0, ext != NULL));
2425         if (ext != NULL) {
2426                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2427                          ext, "index = %lu.\n", index);
2428                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2429
2430                 osc_object_lock(osc);
2431                 if (ext->oe_nr_pages == 0)
2432                         ext->oe_srvlock = ops->ops_srvlock;
2433                 else
2434                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2435                 ++ext->oe_nr_pages;
2436                 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2437                 osc_object_unlock(osc);
2438
2439                 if (!ext->oe_layout_version)
2440                         ext->oe_layout_version = io->ci_layout_version;
2441         }
2442
2443         RETURN(rc);
2444 }
2445
2446 int osc_teardown_async_page(const struct lu_env *env,
2447                             struct osc_object *obj, struct osc_page *ops)
2448 {
2449         struct osc_async_page *oap = &ops->ops_oap;
2450         int rc = 0;
2451         ENTRY;
2452
2453         LASSERT(oap->oap_magic == OAP_MAGIC);
2454
2455         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2456                oap, ops, osc_index(oap2osc(oap)));
2457
2458         if (!list_empty(&oap->oap_rpc_item)) {
2459                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2460                 rc = -EBUSY;
2461         } else if (!list_empty(&oap->oap_pending_item)) {
2462                 struct osc_extent *ext = NULL;
2463
2464                 osc_object_lock(obj);
2465                 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2466                 osc_object_unlock(obj);
2467                 /* only truncated pages are allowed to be taken out.
2468                  * See osc_extent_truncate() and osc_cache_truncate_start()
2469                  * for details. */
2470                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2471                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2472                                         osc_index(oap2osc(oap)));
2473                         rc = -EBUSY;
2474                 }
2475                 if (ext != NULL)
2476                         osc_extent_put(env, ext);
2477         }
2478         RETURN(rc);
2479 }
2480
2481 /**
2482  * This is called when a page is picked up by kernel to write out.
2483  *
2484  * We should find out the corresponding extent and add the whole extent
2485  * into urgent list. The extent may be being truncated or used, handle it
2486  * carefully.
2487  */
2488 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2489                          struct osc_page *ops)
2490 {
2491         struct osc_extent *ext   = NULL;
2492         struct osc_object *obj   = cl2osc(ops->ops_cl.cpl_obj);
2493         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2494         pgoff_t            index = osc_index(ops);
2495         struct osc_async_page *oap = &ops->ops_oap;
2496         bool unplug = false;
2497         int rc = 0;
2498         ENTRY;
2499
2500         osc_object_lock(obj);
2501         ext = osc_extent_lookup(obj, index);
2502         if (ext == NULL) {
2503                 osc_extent_tree_dump(D_ERROR, obj);
2504                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2505         }
2506
2507         switch (ext->oe_state) {
2508         case OES_RPC:
2509         case OES_LOCK_DONE:
2510                 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2511                 LASSERT(0);
2512                 break;
2513         case OES_LOCKING:
2514                 /* If we know this extent is being written out, we should abort
2515                  * so that the writer can make this page ready. Otherwise, there
2516                  * exists a deadlock problem because other process can wait for
2517                  * page writeback bit holding page lock; and meanwhile in
2518                  * vvp_page_make_ready(), we need to grab page lock before
2519                  * really sending the RPC. */
2520         case OES_TRUNC:
2521                 /* race with truncate, page will be redirtied */
2522         case OES_ACTIVE:
2523                 /* The extent is active so we need to abort and let the caller
2524                  * re-dirty the page. If we continued on here, and we were the
2525                  * one making the extent active, we could deadlock waiting for
2526                  * the page writeback to clear but it won't because the extent
2527                  * is active and won't be written out. */
2528                 GOTO(out, rc = -EAGAIN);
2529         default:
2530                 break;
2531         }
2532
2533         rc = cl_page_prep(env, io, cp, CRT_WRITE);
2534         if (rc)
2535                 GOTO(out, rc);
2536
2537         spin_lock(&oap->oap_lock);
2538         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2539         spin_unlock(&oap->oap_lock);
2540
2541         if (current->flags & PF_MEMALLOC)
2542                 ext->oe_memalloc = 1;
2543
2544         ext->oe_urgent = 1;
2545         if (ext->oe_state == OES_CACHE) {
2546                 OSC_EXTENT_DUMP(D_CACHE, ext,
2547                                 "flush page %p make it urgent.\n", oap);
2548                 if (list_empty(&ext->oe_link))
2549                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2550                 unplug = true;
2551         }
2552         rc = 0;
2553         EXIT;
2554
2555 out:
2556         osc_object_unlock(obj);
2557         osc_extent_put(env, ext);
2558         if (unplug)
2559                 osc_io_unplug_async(env, osc_cli(obj), obj);
2560         return rc;
2561 }
2562
2563 int osc_queue_sync_pages(const struct lu_env *env, const struct cl_io *io,
2564                          struct osc_object *obj, struct list_head *list,
2565                          int brw_flags)
2566 {
2567         struct client_obd     *cli = osc_cli(obj);
2568         struct osc_extent     *ext;
2569         struct osc_async_page *oap;
2570         int     page_count = 0;
2571         int     mppr       = cli->cl_max_pages_per_rpc;
2572         bool    can_merge   = true;
2573         pgoff_t start      = CL_PAGE_EOF;
2574         pgoff_t end        = 0;
2575         ENTRY;
2576
2577         list_for_each_entry(oap, list, oap_pending_item) {
2578                 struct osc_page *opg = oap2osc_page(oap);
2579                 pgoff_t index = osc_index(opg);
2580
2581                 if (index > end)
2582                         end = index;
2583                 if (index < start)
2584                         start = index;
2585                 ++page_count;
2586                 mppr <<= (page_count > mppr);
2587
2588                 if (unlikely(opg->ops_from > 0 ||
2589                              opg->ops_to < PAGE_SIZE - 1))
2590                         can_merge = false;
2591         }
2592
2593         ext = osc_extent_alloc(obj);
2594         if (ext == NULL) {
2595                 struct osc_async_page *tmp;
2596
2597                 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2598                         list_del_init(&oap->oap_pending_item);
2599                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2600                 }
2601                 RETURN(-ENOMEM);
2602         }
2603
2604         ext->oe_rw = !!(brw_flags & OBD_BRW_READ);
2605         ext->oe_sync = 1;
2606         ext->oe_no_merge = !can_merge;
2607         ext->oe_urgent = 1;
2608         ext->oe_start = start;
2609         ext->oe_end = ext->oe_max_end = end;
2610         ext->oe_obj = obj;
2611         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2612         ext->oe_ndelay = !!(brw_flags & OBD_BRW_NDELAY);
2613         ext->oe_dio = !!(brw_flags & OBD_BRW_NOCACHE);
2614         if (ext->oe_dio && !ext->oe_rw) { /* direct io write */
2615                 int grants;
2616                 int ppc;
2617
2618                 ppc = 1 << (cli->cl_chunkbits - PAGE_SHIFT);
2619                 grants = cli->cl_grant_extent_tax;
2620                 grants += (1 << cli->cl_chunkbits) *
2621                         ((page_count + ppc - 1) / ppc);
2622
2623                 spin_lock(&cli->cl_loi_list_lock);
2624                 if (osc_reserve_grant(cli, grants) == 0) {
2625                         list_for_each_entry(oap, list, oap_pending_item) {
2626                                 osc_consume_write_grant(cli,
2627                                                         &oap->oap_brw_page);
2628                                 atomic_long_inc(&obd_dirty_pages);
2629                         }
2630                         osc_unreserve_grant_nolock(cli, grants, 0);
2631                         ext->oe_grants = grants;
2632                 }
2633                 spin_unlock(&cli->cl_loi_list_lock);
2634         }
2635
2636         ext->oe_is_rdma_only = !!(brw_flags & OBD_BRW_RDMA_ONLY);
2637         ext->oe_nr_pages = page_count;
2638         ext->oe_mppr = mppr;
2639         list_splice_init(list, &ext->oe_pages);
2640         ext->oe_layout_version = io->ci_layout_version;
2641
2642         osc_object_lock(obj);
2643         /* Reuse the initial refcount for RPC, don't drop it */
2644         osc_extent_state_set(ext, OES_LOCK_DONE);
2645         if (!ext->oe_rw) { /* write */
2646                 if (!ext->oe_srvlock && !ext->oe_dio) {
2647                         /* The most likely case here is from lack of grants
2648                          * so we are either out of quota or out of space.
2649                          * Since this means we are holding locks across
2650                          * potentially multi-striped IO, we must send out
2651                          * everything out instantly to avoid prolonged
2652                          * waits resulting in lock eviction (likely since
2653                          * the extended wait in osc_cache_enter() did not
2654                          * yield any additional grant due to a timeout.
2655                          * LU-13131 */
2656                         ext->oe_hp = 1;
2657                         list_add_tail(&ext->oe_link, &obj->oo_hp_exts);
2658                 } else {
2659                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2660                 }
2661                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2662         } else {
2663                 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2664                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2665         }
2666         osc_object_unlock(obj);
2667
2668         osc_io_unplug_async(env, cli, obj);
2669         RETURN(0);
2670 }
2671
2672 /**
2673  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2674  */
2675 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
2676                              __u64 size, struct osc_extent **extp)
2677 {
2678         struct client_obd *cli = osc_cli(obj);
2679         struct osc_extent *ext;
2680         struct osc_extent *waiting = NULL;
2681         pgoff_t index;
2682         LIST_HEAD(list);
2683         int result = 0;
2684         bool partial;
2685         ENTRY;
2686
2687         /* pages with index greater or equal to index will be truncated. */
2688         index = cl_index(osc2cl(obj), size);
2689         partial = size > cl_offset(osc2cl(obj), index);
2690
2691 again:
2692         osc_object_lock(obj);
2693         ext = osc_extent_search(obj, index);
2694         if (ext == NULL)
2695                 ext = first_extent(obj);
2696         else if (ext->oe_end < index)
2697                 ext = next_extent(ext);
2698         while (ext != NULL) {
2699                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2700
2701                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2702                         /* if ext is in urgent state, it means there must exist
2703                          * a page already having been flushed by write_page().
2704                          * We have to wait for this extent because we can't
2705                          * truncate that page. */
2706                         OSC_EXTENT_DUMP(D_CACHE, ext,
2707                                         "waiting for busy extent\n");
2708                         waiting = osc_extent_get(ext);
2709                         break;
2710                 }
2711
2712                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:%llu.\n", size);
2713
2714                 osc_extent_get(ext);
2715                 if (ext->oe_state == OES_ACTIVE) {
2716                         /* though we grab inode mutex for write path, but we
2717                          * release it before releasing extent(in osc_io_end()),
2718                          * so there is a race window that an extent is still
2719                          * in OES_ACTIVE when truncate starts. */
2720                         LASSERT(!ext->oe_trunc_pending);
2721                         ext->oe_trunc_pending = 1;
2722                 } else {
2723                         EASSERT(ext->oe_state == OES_CACHE, ext);
2724                         osc_extent_state_set(ext, OES_TRUNC);
2725                         osc_update_pending(obj, OBD_BRW_WRITE,
2726                                            -ext->oe_nr_pages);
2727                 }
2728                 /* This extent could be on the full extents list, that's OK */
2729                 EASSERT(!ext->oe_hp && !ext->oe_urgent, ext);
2730                 if (!list_empty(&ext->oe_link))
2731                         list_move_tail(&ext->oe_link, &list);
2732                 else
2733                         list_add_tail(&ext->oe_link, &list);
2734
2735                 ext = next_extent(ext);
2736         }
2737         osc_object_unlock(obj);
2738
2739         osc_list_maint(cli, obj);
2740
2741         while (!list_empty(&list)) {
2742                 int rc;
2743
2744                 ext = list_entry(list.next, struct osc_extent, oe_link);
2745                 list_del_init(&ext->oe_link);
2746
2747                 /* extent may be in OES_ACTIVE state because inode mutex
2748                  * is released before osc_io_end() in file write case */
2749                 if (ext->oe_state != OES_TRUNC)
2750                         osc_extent_wait(env, ext, OES_TRUNC);
2751
2752                 rc = osc_extent_truncate(ext, index, partial);
2753                 if (rc < 0) {
2754                         if (result == 0)
2755                                 result = rc;
2756
2757                         OSC_EXTENT_DUMP(D_ERROR, ext,
2758                                         "truncate error %d\n", rc);
2759                 } else if (ext->oe_nr_pages == 0) {
2760                         osc_extent_remove(ext);
2761                 } else {
2762                         /* this must be an overlapped extent which means only
2763                          * part of pages in this extent have been truncated.
2764                          */
2765                         EASSERTF(ext->oe_start <= index, ext,
2766                                  "trunc index = %lu/%d.\n", index, partial);
2767                         /* fix index to skip this partially truncated extent */
2768                         index = ext->oe_end + 1;
2769                         partial = false;
2770
2771                         /* we need to hold this extent in OES_TRUNC state so
2772                          * that no writeback will happen. This is to avoid
2773                          * BUG 17397.
2774                          * Only partial truncate can reach here, if @size is
2775                          * not zero, the caller should provide a valid @extp. */
2776                         LASSERT(*extp == NULL);
2777                         *extp = osc_extent_get(ext);
2778                         OSC_EXTENT_DUMP(D_CACHE, ext,
2779                                         "trunc at %llu\n", size);
2780                 }
2781                 osc_extent_put(env, ext);
2782         }
2783         if (waiting != NULL) {
2784                 int rc;
2785
2786                 /* ignore the result of osc_extent_wait the write initiator
2787                  * should take care of it. */
2788                 rc = osc_extent_wait(env, waiting, OES_INV);
2789                 if (rc < 0)
2790                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2791
2792                 osc_extent_put(env, waiting);
2793                 waiting = NULL;
2794                 goto again;
2795         }
2796         RETURN(result);
2797 }
2798 EXPORT_SYMBOL(osc_cache_truncate_start);
2799
2800 /**
2801  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2802  */
2803 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext)
2804 {
2805         if (ext != NULL) {
2806                 struct osc_object *obj = ext->oe_obj;
2807                 bool unplug = false;
2808
2809                 EASSERT(ext->oe_nr_pages > 0, ext);
2810                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2811                 EASSERT(!ext->oe_urgent, ext);
2812
2813                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2814                 osc_object_lock(obj);
2815                 osc_extent_state_set(ext, OES_CACHE);
2816                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2817                         ext->oe_urgent = 1;
2818                         list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2819                         unplug = true;
2820                 }
2821                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2822                 osc_object_unlock(obj);
2823                 osc_extent_put(env, ext);
2824
2825                 if (unplug)
2826                         osc_io_unplug_async(env, osc_cli(obj), obj);
2827         }
2828 }
2829
2830 /**
2831  * Wait for extents in a specific range to be written out.
2832  * The caller must have called osc_cache_writeback_range() to issue IO
2833  * otherwise it will take a long time for this function to finish.
2834  *
2835  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2836  * nobody else can dirty this range of file while we're waiting for
2837  * extents to be written.
2838  */
2839 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2840                          pgoff_t start, pgoff_t end)
2841 {
2842         struct osc_extent *ext;
2843         pgoff_t index = start;
2844         int     result = 0;
2845         ENTRY;
2846
2847 again:
2848         osc_object_lock(obj);
2849         ext = osc_extent_search(obj, index);
2850         if (ext == NULL)
2851                 ext = first_extent(obj);
2852         else if (ext->oe_end < index)
2853                 ext = next_extent(ext);
2854         while (ext != NULL) {
2855                 int rc;
2856
2857                 if (ext->oe_start > end)
2858                         break;
2859
2860                 if (!ext->oe_fsync_wait) {
2861                         ext = next_extent(ext);
2862                         continue;
2863                 }
2864
2865                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2866                              ext->oe_hp || ext->oe_urgent), ext);
2867                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2868                              !ext->oe_hp && ext->oe_urgent), ext);
2869
2870                 index = ext->oe_end + 1;
2871                 osc_extent_get(ext);
2872                 osc_object_unlock(obj);
2873
2874                 rc = osc_extent_wait(env, ext, OES_INV);
2875                 if (result == 0)
2876                         result = rc;
2877                 osc_extent_put(env, ext);
2878                 goto again;
2879         }
2880         osc_object_unlock(obj);
2881
2882         OSC_IO_DEBUG(obj, "sync file range.\n");
2883         RETURN(result);
2884 }
2885 EXPORT_SYMBOL(osc_cache_wait_range);
2886
2887 /**
2888  * Called to write out a range of osc object.
2889  *
2890  * @hp     : should be set this is caused by lock cancel;
2891  * @discard: is set if dirty pages should be dropped - file will be deleted or
2892  *         truncated, this implies there is no partially discarding extents.
2893  *
2894  * Return how many pages will be issued, or error code if error occurred.
2895  */
2896 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2897                               pgoff_t start, pgoff_t end, int hp, int discard)
2898 {
2899         struct osc_extent *ext;
2900         LIST_HEAD(discard_list);
2901         bool unplug = false;
2902         int result = 0;
2903         ENTRY;
2904
2905         osc_object_lock(obj);
2906         ext = osc_extent_search(obj, start);
2907         if (ext == NULL)
2908                 ext = first_extent(obj);
2909         else if (ext->oe_end < start)
2910                 ext = next_extent(ext);
2911         while (ext != NULL) {
2912                 if (ext->oe_start > end)
2913                         break;
2914
2915                 ext->oe_fsync_wait = 1;
2916                 switch (ext->oe_state) {
2917                 case OES_CACHE:
2918                         result += ext->oe_nr_pages;
2919                         if (!discard) {
2920                                 struct list_head *list = NULL;
2921                                 if (hp) {
2922                                         EASSERT(!ext->oe_hp, ext);
2923                                         ext->oe_hp = 1;
2924                                         list = &obj->oo_hp_exts;
2925                                 } else if (!ext->oe_urgent && !ext->oe_hp) {
2926                                         ext->oe_urgent = 1;
2927                                         list = &obj->oo_urgent_exts;
2928                                 }
2929                                 if (list != NULL)
2930                                         list_move_tail(&ext->oe_link, list);
2931                                 unplug = true;
2932                         } else {
2933                                 struct client_obd *cli = osc_cli(obj);
2934                                 int pcc_bits = cli->cl_chunkbits - PAGE_SHIFT;
2935                                 pgoff_t align_by = (1 << pcc_bits);
2936                                 pgoff_t a_start = round_down(start, align_by);
2937                                 pgoff_t a_end = round_up(end, align_by);
2938
2939                                 /* overflow case */
2940                                 if (end && !a_end)
2941                                         a_end = CL_PAGE_EOF;
2942                                 /* the only discarder is lock cancelling, so
2943                                  * [start, end], aligned by chunk size, must
2944                                  * contain this extent */
2945                                 LASSERTF(ext->oe_start >= a_start &&
2946                                          ext->oe_end <= a_end,
2947                                          "ext [%lu, %lu] reg [%lu, %lu] "
2948                                          "orig [%lu %lu] align %lu bits "
2949                                          "%d\n", ext->oe_start, ext->oe_end,
2950                                          a_start, a_end, start, end,
2951                                          align_by, pcc_bits);
2952                                 osc_extent_state_set(ext, OES_LOCKING);
2953                                 ext->oe_owner = current;
2954                                 list_move_tail(&ext->oe_link,
2955                                                    &discard_list);
2956                                 osc_update_pending(obj, OBD_BRW_WRITE,
2957                                                    -ext->oe_nr_pages);
2958                         }
2959                         break;
2960                 case OES_ACTIVE:
2961                         /* It's pretty bad to wait for ACTIVE extents, because
2962                          * we don't know how long we will wait for it to be
2963                          * flushed since it may be blocked at awaiting more
2964                          * grants. We do this for the correctness of fsync. */
2965                         LASSERT(hp == 0 && discard == 0);
2966                         ext->oe_urgent = 1;
2967                         break;
2968                 case OES_TRUNC:
2969                         /* this extent is being truncated, can't do anything
2970                          * for it now. it will be set to urgent after truncate
2971                          * is finished in osc_cache_truncate_end(). */
2972                 default:
2973                         break;
2974                 }
2975                 ext = next_extent(ext);
2976         }
2977         osc_object_unlock(obj);
2978
2979         LASSERT(ergo(!discard, list_empty(&discard_list)));
2980         if (!list_empty(&discard_list)) {
2981                 struct osc_extent *tmp;
2982                 int rc;
2983
2984                 osc_list_maint(osc_cli(obj), obj);
2985                 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
2986                         list_del_init(&ext->oe_link);
2987                         EASSERT(ext->oe_state == OES_LOCKING, ext);
2988
2989                         /* Discard caching pages. We don't actually write this
2990                          * extent out but we complete it as if we did. */
2991                         rc = osc_extent_make_ready(env, ext);
2992                         if (unlikely(rc < 0)) {
2993                                 OSC_EXTENT_DUMP(D_ERROR, ext,
2994                                                 "make_ready returned %d\n", rc);
2995                                 if (result >= 0)
2996                                         result = rc;
2997                         }
2998
2999                         /* finish the extent as if the pages were sent */
3000                         osc_extent_finish(env, ext, 0, 0);
3001                 }
3002         }
3003
3004         if (unplug)
3005                 osc_io_unplug(env, osc_cli(obj), obj);
3006
3007         if (hp || discard) {
3008                 int rc;
3009                 rc = osc_cache_wait_range(env, obj, start, end);
3010                 if (result >= 0 && rc < 0)
3011                         result = rc;
3012         }
3013
3014         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3015         RETURN(result);
3016 }
3017 EXPORT_SYMBOL(osc_cache_writeback_range);
3018
3019 /**
3020  * Returns a list of pages by a given [start, end] of \a obj.
3021  *
3022  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3023  * crucial in the face of [offset, EOF] locks.
3024  *
3025  * Return at least one page in @queue unless there is no covered page.
3026  */
3027 bool osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3028                           struct osc_object *osc, pgoff_t start, pgoff_t end,
3029                           osc_page_gang_cbt cb, void *cbdata)
3030 {
3031         struct osc_page *ops;
3032         struct pagevec  *pagevec;
3033         void            **pvec;
3034         pgoff_t         idx;
3035         unsigned int    nr;
3036         unsigned int    i;
3037         unsigned int    j;
3038         bool            res = true;
3039         bool            tree_lock = true;
3040         ENTRY;
3041
3042         idx = start;
3043         pvec = osc_env_info(env)->oti_pvec;
3044         pagevec = &osc_env_info(env)->oti_pagevec;
3045         ll_pagevec_init(pagevec, 0);
3046         spin_lock(&osc->oo_tree_lock);
3047         while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3048                                             idx, OTI_PVEC_SIZE)) > 0) {
3049                 struct cl_page *page;
3050                 bool end_of_region = false;
3051
3052                 for (i = 0, j = 0; i < nr; ++i) {
3053                         ops = pvec[i];
3054                         pvec[i] = NULL;
3055
3056                         idx = osc_index(ops);
3057                         if (idx > end) {
3058                                 end_of_region = true;
3059                                 break;
3060                         }
3061
3062                         page = ops->ops_cl.cpl_page;
3063                         LASSERT(page->cp_type == CPT_CACHEABLE);
3064                         if (page->cp_state == CPS_FREEING)
3065                                 continue;
3066
3067                         cl_page_get(page);
3068                         lu_ref_add_atomic(&page->cp_reference,
3069                                           "gang_lookup", current);
3070                         pvec[j++] = ops;
3071                 }
3072                 ++idx;
3073
3074                 /*
3075                  * Here a delicate locking dance is performed. Current thread
3076                  * holds a reference to a page, but has to own it before it
3077                  * can be placed into queue. Owning implies waiting, so
3078                  * radix-tree lock is to be released. After a wait one has to
3079                  * check that pages weren't truncated (cl_page_own() returns
3080                  * error in the latter case).
3081                  */
3082                 spin_unlock(&osc->oo_tree_lock);
3083                 tree_lock = false;
3084
3085                 for (i = 0; i < j; ++i) {
3086                         ops = pvec[i];
3087                         if (res)
3088                                 res = (*cb)(env, io, ops, cbdata);
3089
3090                         page = ops->ops_cl.cpl_page;
3091                         lu_ref_del(&page->cp_reference, "gang_lookup", current);
3092                         cl_pagevec_put(env, page, pagevec);
3093                 }
3094                 pagevec_release(pagevec);
3095
3096                 if (nr < OTI_PVEC_SIZE || end_of_region)
3097                         break;
3098
3099                 if (!res)
3100                         break;
3101                 if (need_resched())
3102                         cond_resched();
3103
3104                 spin_lock(&osc->oo_tree_lock);
3105                 tree_lock = true;
3106         }
3107         if (tree_lock)
3108                 spin_unlock(&osc->oo_tree_lock);
3109         RETURN(res);
3110 }
3111 EXPORT_SYMBOL(osc_page_gang_lookup);
3112
3113 /**
3114  * Check if page @page is covered by an extra lock or discard it.
3115  */
3116 static bool check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3117                                 struct osc_page *ops, void *cbdata)
3118 {
3119         struct osc_thread_info *info = osc_env_info(env);
3120         struct osc_object *osc = cbdata;
3121         struct cl_page *page = ops->ops_cl.cpl_page;
3122         pgoff_t index;
3123         bool discard = false;
3124
3125         index = osc_index(ops);
3126
3127         /* negative lock caching */
3128         if (index < info->oti_ng_index) {
3129                 discard = true;
3130         } else if (index >= info->oti_fn_index) {
3131                 struct ldlm_lock *tmp;
3132                 /* refresh non-overlapped index */
3133                 tmp = osc_dlmlock_at_pgoff(env, osc, index,
3134                                            OSC_DAP_FL_TEST_LOCK |
3135                                            OSC_DAP_FL_AST | OSC_DAP_FL_RIGHT);
3136                 if (tmp != NULL) {
3137                         __u64 end = tmp->l_policy_data.l_extent.end;
3138                         __u64 start = tmp->l_policy_data.l_extent.start;
3139
3140                         /* no lock covering this page */
3141                         if (index < cl_index(osc2cl(osc), start)) {
3142                                 /* no lock at @index, first lock at @start */
3143                                 info->oti_ng_index = cl_index(osc2cl(osc),
3144                                                      start);
3145                                 discard = true;
3146                         } else {
3147                                 /* Cache the first-non-overlapped index so as to
3148                                  * skip all pages within [index, oti_fn_index).
3149                                  * This is safe because if tmp lock is canceled,
3150                                  * it will discard these pages.
3151                                  */
3152                                 info->oti_fn_index = cl_index(osc2cl(osc),
3153                                                      end + 1);
3154                                 if (end == OBD_OBJECT_EOF)
3155                                         info->oti_fn_index = CL_PAGE_EOF;
3156                         }
3157                         LDLM_LOCK_PUT(tmp);
3158                 } else {
3159                         info->oti_ng_index = CL_PAGE_EOF;
3160                         discard = true;
3161                 }
3162         }
3163
3164         if (discard) {
3165                 if (cl_page_own(env, io, page) == 0) {
3166                         cl_page_discard(env, io, page);
3167                         cl_page_disown(env, io, page);
3168                 } else {
3169                         LASSERT(page->cp_state == CPS_FREEING);
3170                 }
3171         }
3172
3173         info->oti_next_index = index + 1;
3174         return true;
3175 }
3176
3177 bool osc_discard_cb(const struct lu_env *env, struct cl_io *io,
3178                    struct osc_page *ops, void *cbdata)
3179 {
3180         struct osc_thread_info *info = osc_env_info(env);
3181         struct cl_page *page = ops->ops_cl.cpl_page;
3182
3183         /* page is top page. */
3184         info->oti_next_index = osc_index(ops) + 1;
3185         if (cl_page_own(env, io, page) == 0) {
3186                 if (!ergo(page->cp_type == CPT_CACHEABLE,
3187                           !PageDirty(cl_page_vmpage(page))))
3188                         CL_PAGE_DEBUG(D_ERROR, env, page,
3189                                         "discard dirty page?\n");
3190
3191                 /* discard the page */
3192                 cl_page_discard(env, io, page);
3193                 cl_page_disown(env, io, page);
3194         } else {
3195                 LASSERT(page->cp_state == CPS_FREEING);
3196         }
3197
3198         return true;
3199 }
3200 EXPORT_SYMBOL(osc_discard_cb);
3201
3202 /**
3203  * Discard pages protected by the given lock. This function traverses radix
3204  * tree to find all covering pages and discard them. If a page is being covered
3205  * by other locks, it should remain in cache.
3206  *
3207  * If error happens on any step, the process continues anyway (the reasoning
3208  * behind this being that lock cancellation cannot be delayed indefinitely).
3209  */
3210 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3211                            pgoff_t start, pgoff_t end, bool discard)
3212 {
3213         struct osc_thread_info *info = osc_env_info(env);
3214         struct cl_io *io = osc_env_thread_io(env);
3215         osc_page_gang_cbt cb;
3216         int result;
3217
3218         ENTRY;
3219
3220         io->ci_obj = cl_object_top(osc2cl(osc));
3221         io->ci_ignore_layout = 1;
3222         result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3223         if (result != 0)
3224                 GOTO(out, result);
3225
3226         cb = discard ? osc_discard_cb : check_and_discard_cb;
3227         info->oti_fn_index = info->oti_next_index = start;
3228         info->oti_ng_index = 0;
3229
3230         osc_page_gang_lookup(env, io, osc,
3231                              info->oti_next_index, end, cb, osc);
3232 out:
3233         cl_io_fini(env, io);
3234         RETURN(result);
3235 }
3236
3237
3238 /** @} osc */