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 <= ext->oe_start << PAGE_SHIFT &&
 233                       extent->end >= ext->oe_max_end << PAGE_SHIFT))
 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_RELEASE(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         /* dio pages do not go in the LRU */
 849         if (!ext->oe_dio)
 850                 osc_lru_add_batch(cli, &ext->oe_pages);
 851
 852         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 853                                      oap_pending_item) {
 854                 list_del_init(&oap->oap_rpc_item);
 855                 list_del_init(&oap->oap_pending_item);
 856                 if (last_off <= oap->oap_obj_off) {
 857                         last_off = oap->oap_obj_off;
 858                         last_count = oap->oap_count;
 859                 }
 860
 861                 --ext->oe_nr_pages;
 862                 osc_ap_completion(env, cli, oap, sent, rc);
 863         }
 864         EASSERT(ext->oe_nr_pages == 0, ext);
 865
 866         if (!sent) {
 867                 lost_grant = ext->oe_grants;
 868         } else if (cli->cl_ocd_grant_param == 0 &&
 869                    blocksize < PAGE_SIZE &&
 870                    last_count != PAGE_SIZE) {
 871                 /* For short writes without OBD_CONNECT_GRANT support, we
 872                  * shouldn't count parts of pages that span a whole chunk on
 873                  * the OST side, or our accounting goes wrong. Should match
 874                  * the code in tgt_grant_check.
 875                  */
 876                 int offset = last_off & ~PAGE_MASK;
 877                 int count = last_count + (offset & (blocksize - 1));
 878                 int end = (offset + last_count) & (blocksize - 1);
 879                 if (end)
 880                         count += blocksize - end;
 881
 882                 lost_grant = PAGE_SIZE - count;
 883         }
 884         if (ext->oe_grants > 0)
 885                 osc_free_grant(cli, nr_pages, lost_grant, ext->oe_grants);
 886
 887         osc_extent_remove(ext);
 888         /* put the refcount for RPC */
 889         osc_extent_put(env, ext);
 890         RETURN(0);
 891 }
 892
 893 /**
 894  * Wait for the extent's state to become @state.
 895  */
 896 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
 897                            enum osc_extent_state state)
 898 {
 899         struct osc_object *obj = ext->oe_obj;
 900         int rc = 0;
 901         ENTRY;
 902
 903         osc_object_lock(obj);
 904         LASSERT(sanity_check_nolock(ext) == 0);
 905         /* `Kick' this extent only if the caller is waiting for it to be
 906          * written out. */
 907         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
 908                 if (ext->oe_state == OES_ACTIVE) {
 909                         ext->oe_urgent = 1;
 910                 } else if (ext->oe_state == OES_CACHE) {
 911                         ext->oe_urgent = 1;
 912                         osc_extent_hold(ext);
 913                         rc = 1;
 914                 }
 915         }
 916         osc_object_unlock(obj);
 917         if (rc == 1)
 918                 osc_extent_release(env, ext);
 919
 920         /* wait for the extent until its state becomes @state */
 921         rc = wait_event_idle_timeout(ext->oe_waitq,
 922                                      smp_load_acquire(&ext->oe_state) == state,
 923                                      cfs_time_seconds(600));
 924         if (rc == 0) {
 925                 OSC_EXTENT_DUMP(D_ERROR, ext,
 926                         "%s: wait ext to %u timedout, recovery in progress?\n",
 927                         cli_name(osc_cli(obj)), state);
 928
 929                 wait_event_idle(ext->oe_waitq,
 930                                 smp_load_acquire(&ext->oe_state) == state);
 931         }
 932         if (ext->oe_rc < 0)
 933                 rc = ext->oe_rc;
 934         else
 935                 rc = 0;
 936         RETURN(rc);
 937 }
 938
 939 /**
 940  * Discard pages with index greater than @size. If @ext is overlapped with
 941  * @size, then partial truncate happens.
 942  */
 943 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
 944                                 bool partial)
 945 {
 946         struct lu_env         *env;
 947         struct cl_io          *io;
 948         struct osc_object     *obj = ext->oe_obj;
 949         struct client_obd     *cli = osc_cli(obj);
 950         struct osc_async_page *oap;
 951         struct osc_async_page *tmp;
 952         struct pagevec        *pvec;
 953         int                    pages_in_chunk = 0;
 954         int                    ppc_bits    = cli->cl_chunkbits -
 955                                              PAGE_SHIFT;
 956         __u64                  trunc_chunk = trunc_index >> ppc_bits;
 957         int                    grants   = 0;
 958         int                    nr_pages = 0;
 959         int                    rc       = 0;
 960         __u16                  refcheck;
 961         ENTRY;
 962
 963         LASSERT(sanity_check(ext) == 0);
 964         LASSERT(ext->oe_state == OES_TRUNC);
 965         LASSERT(!ext->oe_urgent);
 966
 967         /* Request new lu_env.
 968          * We can't use that env from osc_cache_truncate_start() because
 969          * it's from lov_io_sub and not fully initialized. */
 970         env = cl_env_get(&refcheck);
 971         if (IS_ERR(env))
 972                 RETURN(PTR_ERR(env));
 973
 974         io  = osc_env_thread_io(env);
 975         io->ci_obj = cl_object_top(osc2cl(obj));
 976         io->ci_ignore_layout = 1;
 977         pvec = &osc_env_info(env)->oti_pagevec;
 978         ll_pagevec_init(pvec, 0);
 979         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
 980         if (rc < 0)
 981                 GOTO(out, rc);
 982
 983         /* discard all pages with index greater than trunc_index */
 984         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 985                                      oap_pending_item) {
 986                 pgoff_t index = osc_index(oap2osc(oap));
 987                 struct cl_page  *page = oap2cl_page(oap);
 988
 989                 LASSERT(list_empty(&oap->oap_rpc_item));
 990
 991                 /* only discard the pages with their index greater than
 992                  * trunc_index, and ... */
 993                 if (index < trunc_index ||
 994                     (index == trunc_index && partial)) {
 995                         /* accounting how many pages remaining in the chunk
 996                          * so that we can calculate grants correctly. */
 997                         if (index >> ppc_bits == trunc_chunk)
 998                                 ++pages_in_chunk;
 999                         continue;
1000                 }
1001
1002                 list_del_init(&oap->oap_pending_item);
1003
1004                 cl_page_get(page);
1005                 lu_ref_add(&page->cp_reference, "truncate", current);
1006
1007                 if (cl_page_own(env, io, page) == 0) {
1008                         cl_page_discard(env, io, page);
1009                         cl_page_disown(env, io, page);
1010                 } else {
1011                         LASSERT(page->cp_state == CPS_FREEING);
1012                         LASSERT(0);
1013                 }
1014
1015                 lu_ref_del(&page->cp_reference, "truncate", current);
1016                 cl_pagevec_put(env, page, pvec);
1017
1018                 --ext->oe_nr_pages;
1019                 ++nr_pages;
1020         }
1021         pagevec_release(pvec);
1022
1023         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1024                       ext->oe_nr_pages == 0),
1025                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1026
1027         osc_object_lock(obj);
1028         if (ext->oe_nr_pages == 0) {
1029                 LASSERT(pages_in_chunk == 0);
1030                 grants = ext->oe_grants;
1031                 ext->oe_grants = 0;
1032         } else { /* calculate how many grants we can free */
1033                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1034                 pgoff_t last_index;
1035
1036
1037                 /* if there is no pages in this chunk, we can also free grants
1038                  * for the last chunk */
1039                 if (pages_in_chunk == 0) {
1040                         /* if this is the 1st chunk and no pages in this chunk,
1041                          * ext->oe_nr_pages must be zero, so we should be in
1042                          * the other if-clause. */
1043                         LASSERT(trunc_chunk > 0);
1044                         --trunc_chunk;
1045                         ++chunks;
1046                 }
1047
1048                 /* this is what we can free from this extent */
1049                 grants          = chunks << cli->cl_chunkbits;
1050                 ext->oe_grants -= grants;
1051                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1052                 ext->oe_end     = min(last_index, ext->oe_max_end);
1053                 LASSERT(ext->oe_end >= ext->oe_start);
1054                 LASSERT(ext->oe_grants > 0);
1055         }
1056         osc_object_unlock(obj);
1057
1058         if (grants > 0 || nr_pages > 0)
1059                 osc_free_grant(cli, nr_pages, grants, grants);
1060
1061 out:
1062         cl_io_fini(env, io);
1063         cl_env_put(env, &refcheck);
1064         RETURN(rc);
1065 }
1066
1067 /**
1068  * This function is used to make the extent prepared for transfer.
1069  * A race with flusing page - ll_writepage() has to be handled cautiously.
1070  */
1071 static int osc_extent_make_ready(const struct lu_env *env,
1072                                  struct osc_extent *ext)
1073 {
1074         struct osc_async_page *oap;
1075         struct osc_async_page *last = NULL;
1076         struct osc_object *obj = ext->oe_obj;
1077         unsigned int page_count = 0;
1078         int rc;
1079         ENTRY;
1080
1081         /* we're going to grab page lock, so object lock must not be taken. */
1082         LASSERT(sanity_check(ext) == 0);
1083         /* in locking state, any process should not touch this extent. */
1084         EASSERT(ext->oe_state == OES_LOCKING, ext);
1085         EASSERT(ext->oe_owner != NULL, ext);
1086
1087         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1088
1089         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1090                 ++page_count;
1091                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1092                         last = oap;
1093
1094                 /* checking ASYNC_READY is race safe */
1095                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1096                         continue;
1097
1098                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1099                 switch (rc) {
1100                 case 0:
1101                         oap->oap_async_flags |= ASYNC_READY;
1102                         break;
1103                 case -EALREADY:
1104                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1105                         break;
1106                 default:
1107                         LASSERTF(0, "unknown return code: %d\n", rc);
1108                 }
1109         }
1110
1111         LASSERT(page_count == ext->oe_nr_pages);
1112         LASSERT(last != NULL);
1113         /* the last page is the only one we need to refresh its count by
1114          * the size of file. */
1115         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1116                 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1117                 LASSERTF(last_oap_count > 0,
1118                          "last_oap_count %d\n", last_oap_count);
1119                 LASSERT(last->oap_page_off + last_oap_count <= PAGE_SIZE);
1120                 last->oap_count = last_oap_count;
1121                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1122         }
1123
1124         /* for the rest of pages, we don't need to call osf_refresh_count()
1125          * because it's known they are not the last page */
1126         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1127                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1128                         oap->oap_count = PAGE_SIZE - oap->oap_page_off;
1129                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
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 cl_page  *page = oap2cl_page(oap);
1254         int result;
1255
1256         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1257
1258         ENTRY;
1259
1260         result = cl_page_make_ready(env, page, CRT_WRITE);
1261
1262         RETURN(result);
1263 }
1264
1265 static int osc_refresh_count(const struct lu_env *env,
1266                              struct osc_async_page *oap, int cmd)
1267 {
1268         struct osc_page  *opg = oap2osc_page(oap);
1269         pgoff_t index = osc_index(oap2osc(oap));
1270         struct cl_object *obj = osc2cl(osc_page_object(opg));
1271         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
1272         int result;
1273         loff_t kms;
1274
1275         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1276         LASSERT(!(cmd & OBD_BRW_READ));
1277         LASSERT(opg != NULL);
1278
1279         cl_object_attr_lock(obj);
1280         result = cl_object_attr_get(env, obj, attr);
1281         cl_object_attr_unlock(obj);
1282         if (result < 0)
1283                 return result;
1284         kms = attr->cat_kms;
1285         if (index << PAGE_SHIFT >= kms)
1286                 /* catch race with truncate */
1287                 return 0;
1288         else if ((index + 1) << PAGE_SHIFT > kms)
1289                 /* catch sub-page write at end of file */
1290                 return kms & ~PAGE_MASK;
1291         else
1292                 return PAGE_SIZE;
1293 }
1294
1295 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1296                           int cmd, int rc)
1297 {
1298         struct osc_page   *opg  = oap2osc_page(oap);
1299         struct cl_page    *page = oap2cl_page(oap);
1300         enum cl_req_type   crt;
1301         int srvlock;
1302
1303         ENTRY;
1304
1305         cmd &= ~OBD_BRW_NOQUOTA;
1306         LASSERTF(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ),
1307                  "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1308         LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1309                 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1310         LASSERT(opg->ops_transfer_pinned);
1311
1312         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1313         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1314         opg->ops_transfer_pinned = 0;
1315
1316         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1317
1318         /* statistic */
1319         if (rc == 0 && srvlock) {
1320                 struct lu_device *ld = osc_page_object(opg)->oo_cl.co_lu.lo_dev;
1321                 struct osc_stats *stats = &lu2osc_dev(ld)->osc_stats;
1322                 size_t bytes = oap->oap_count;
1323
1324                 if (crt == CRT_READ)
1325                         stats->os_lockless_reads += bytes;
1326                 else
1327                         stats->os_lockless_writes += bytes;
1328         }
1329
1330         /*
1331          * This has to be the last operation with the page, as locks are
1332          * released in cl_page_completion() and nothing except for the
1333          * reference counter protects page from concurrent reclaim.
1334          */
1335         lu_ref_del(&page->cp_reference, "transfer", page);
1336
1337         cl_page_completion(env, page, crt, rc);
1338         cl_page_put(env, page);
1339
1340         RETURN(0);
1341 }
1342
1343 #define OSC_DUMP_GRANT(mask, cli, fmt, args...) do {                    \
1344         struct client_obd *__tmp = (cli);                               \
1345         CDEBUG(mask, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu " \
1346                "dropped: %ld avail: %ld, dirty_grant: %ld, "            \
1347                "reserved: %ld, flight: %d } lru {in list: %ld, "        \
1348                "left: %ld, waiters: %d }" fmt "\n",                     \
1349                cli_name(__tmp),                                         \
1350                __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages,        \
1351                atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1352                __tmp->cl_lost_grant, __tmp->cl_avail_grant,             \
1353                __tmp->cl_dirty_grant,                                   \
1354                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,         \
1355                atomic_long_read(&__tmp->cl_lru_in_list),                \
1356                atomic_long_read(&__tmp->cl_lru_busy),                   \
1357                atomic_read(&__tmp->cl_lru_shrinkers), ##args);          \
1358 } while (0)
1359
1360 /* caller must hold loi_list_lock */
1361 static void osc_consume_write_grant(struct client_obd *cli,
1362                                     struct brw_page *pga)
1363 {
1364         assert_spin_locked(&cli->cl_loi_list_lock);
1365         LASSERT(!(pga->bp_flag & OBD_BRW_FROM_GRANT));
1366         cli->cl_dirty_pages++;
1367         pga->bp_flag |= OBD_BRW_FROM_GRANT;
1368         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1369                PAGE_SIZE, pga, pga->bp_page);
1370 }
1371
1372 /* the companion to osc_consume_write_grant, called when a brw has completed.
1373  * must be called with the loi lock held. */
1374 static void osc_release_write_grant(struct client_obd *cli,
1375                                     struct brw_page *pga)
1376 {
1377         ENTRY;
1378
1379         assert_spin_locked(&cli->cl_loi_list_lock);
1380         if (!(pga->bp_flag & OBD_BRW_FROM_GRANT)) {
1381                 EXIT;
1382                 return;
1383         }
1384
1385         pga->bp_flag &= ~OBD_BRW_FROM_GRANT;
1386         atomic_long_dec(&obd_dirty_pages);
1387         cli->cl_dirty_pages--;
1388         EXIT;
1389 }
1390
1391 /**
1392  * To avoid sleeping with object lock held, it's good for us allocate enough
1393  * grants before entering into critical section.
1394  *
1395  * client_obd_list_lock held by caller
1396  */
1397 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1398 {
1399         int rc = -EDQUOT;
1400
1401         if (cli->cl_avail_grant >= bytes) {
1402                 cli->cl_avail_grant    -= bytes;
1403                 cli->cl_reserved_grant += bytes;
1404                 rc = 0;
1405         }
1406         return rc;
1407 }
1408
1409 static void __osc_unreserve_grant(struct client_obd *cli,
1410                                   unsigned int reserved, unsigned int unused)
1411 {
1412         /* it's quite normal for us to get more grant than reserved.
1413          * Thinking about a case that two extents merged by adding a new
1414          * chunk, we can save one extent tax. If extent tax is greater than
1415          * one chunk, we can save more grant by adding a new chunk */
1416         cli->cl_reserved_grant -= reserved;
1417         if (unused > reserved) {
1418                 cli->cl_avail_grant += reserved;
1419                 cli->cl_lost_grant  += unused - reserved;
1420                 cli->cl_dirty_grant -= unused - reserved;
1421         } else {
1422                 cli->cl_avail_grant += unused;
1423                 cli->cl_dirty_grant += reserved - unused;
1424         }
1425 }
1426
1427 static void osc_unreserve_grant_nolock(struct client_obd *cli,
1428                                        unsigned int reserved,
1429                                        unsigned int unused)
1430 {
1431         __osc_unreserve_grant(cli, reserved, unused);
1432         if (unused > 0)
1433                 osc_wake_cache_waiters(cli);
1434 }
1435
1436 static void osc_unreserve_grant(struct client_obd *cli,
1437                                 unsigned int reserved, unsigned int unused)
1438 {
1439         spin_lock(&cli->cl_loi_list_lock);
1440         osc_unreserve_grant_nolock(cli, reserved, unused);
1441         spin_unlock(&cli->cl_loi_list_lock);
1442 }
1443
1444 /**
1445  * Free grant after IO is finished or canceled.
1446  *
1447  * @lost_grant is used to remember how many grants we have allocated but not
1448  * used, we should return these grants to OST. There're two cases where grants
1449  * can be lost:
1450  * 1. truncate;
1451  * 2. Without OBD_CONNECT_GRANT support and blocksize at OST is less than
1452  *    PAGE_SIZE and a partial page was written. In this case OST may use less
1453  *    chunks to serve this partial write. OSTs don't actually know the page
1454  *    size on the client side. so clients have to calculate lost grant by the
1455  *    blocksize on the OST. See tgt_grant_check() for details.
1456  */
1457 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1458                            unsigned int lost_grant, unsigned int dirty_grant)
1459 {
1460         unsigned long grant;
1461
1462         grant = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
1463
1464         spin_lock(&cli->cl_loi_list_lock);
1465         atomic_long_sub(nr_pages, &obd_dirty_pages);
1466         cli->cl_dirty_pages -= nr_pages;
1467         cli->cl_lost_grant += lost_grant;
1468         cli->cl_dirty_grant -= dirty_grant;
1469         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1470                 /* borrow some grant from truncate to avoid the case that
1471                  * truncate uses up all avail grant */
1472                 cli->cl_lost_grant -= grant;
1473                 cli->cl_avail_grant += grant;
1474         }
1475         osc_wake_cache_waiters(cli);
1476         spin_unlock(&cli->cl_loi_list_lock);
1477         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu/%lu\n",
1478                lost_grant, cli->cl_lost_grant,
1479                cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_SHIFT,
1480                cli->cl_dirty_grant);
1481 }
1482
1483 /**
1484  * The companion to osc_enter_cache(), called when @oap is no longer part of
1485  * the dirty accounting due to error.
1486  */
1487 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1488 {
1489         spin_lock(&cli->cl_loi_list_lock);
1490         osc_release_write_grant(cli, &oap->oap_brw_page);
1491         spin_unlock(&cli->cl_loi_list_lock);
1492 }
1493
1494 /**
1495  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1496  * is available.
1497  */
1498 static int osc_enter_cache_try(struct client_obd *cli,
1499                                struct osc_async_page *oap,
1500                                int bytes)
1501 {
1502         int rc;
1503
1504         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1505
1506         rc = osc_reserve_grant(cli, bytes);
1507         if (rc < 0)
1508                 return 0;
1509
1510         if (cli->cl_dirty_pages < cli->cl_dirty_max_pages) {
1511                 if (atomic_long_add_return(1, &obd_dirty_pages) <=
1512                     obd_max_dirty_pages) {
1513                         osc_consume_write_grant(cli, &oap->oap_brw_page);
1514                         rc = 1;
1515                         goto out;
1516                 } else
1517                         atomic_long_dec(&obd_dirty_pages);
1518         }
1519         __osc_unreserve_grant(cli, bytes, bytes);
1520
1521 out:
1522         return rc;
1523 }
1524
1525 /* Following two inlines exist to pass code fragments
1526  * to wait_event_idle_exclusive_timeout_cmd().  Passing
1527  * code fragments as macro args can look confusing, so
1528  * we provide inlines to encapsulate them.
1529  */
1530 static inline void cli_unlock_and_unplug(const struct lu_env *env,
1531                                          struct client_obd *cli,
1532                                          struct osc_async_page *oap)
1533 {
1534         spin_unlock(&cli->cl_loi_list_lock);
1535         osc_io_unplug_async(env, cli, NULL);
1536         CDEBUG(D_CACHE,
1537                "%s: sleeping for cache space for %p\n",
1538                cli_name(cli), oap);
1539 }
1540
1541 static inline void cli_lock_after_unplug(struct client_obd *cli)
1542 {
1543         spin_lock(&cli->cl_loi_list_lock);
1544 }
1545 /**
1546  * The main entry to reserve dirty page accounting. Usually the grant reserved
1547  * in this function will be freed in bulk in osc_free_grant() unless it fails
1548  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1549  *
1550  * The process will be put into sleep if it's already run out of grant.
1551  */
1552 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1553                            struct osc_async_page *oap, int bytes)
1554 {
1555         struct osc_object *osc = oap->oap_obj;
1556         struct lov_oinfo *loi = osc->oo_oinfo;
1557         int rc = -EDQUOT;
1558         int remain;
1559         bool entered = false;
1560         struct obd_device *obd = cli->cl_import->imp_obd;
1561         /* We cannot wait for a long time here since we are holding ldlm lock
1562          * across the actual IO. If no requests complete fast (e.g. due to
1563          * overloaded OST that takes a long time to process everything, we'd
1564          * get evicted if we wait for a normal obd_timeout or some such.
1565          * So we try to wait half the time it would take the client to be
1566          * evicted by server which is half obd_timeout when AT is off
1567          * or at least ldlm_enqueue_min with AT on.
1568          * See LU-13131 */
1569         unsigned long timeout =
1570                 cfs_time_seconds(obd_at_off(obd) ?
1571                                  obd_timeout / 2 :
1572                                  obd_get_ldlm_enqueue_min(obd) / 2);
1573
1574         ENTRY;
1575
1576         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1577
1578         spin_lock(&cli->cl_loi_list_lock);
1579
1580         /* force the caller to try sync io.  this can jump the list
1581          * of queued writes and create a discontiguous rpc stream */
1582         if (CFS_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1583             cli->cl_dirty_max_pages == 0 ||
1584             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) {
1585                 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1586                 GOTO(out, rc = -EDQUOT);
1587         }
1588
1589         /*
1590          * We can wait here for two reasons: too many dirty pages in cache, or
1591          * run out of grants. In both cases we should write dirty pages out.
1592          * Adding a cache waiter will trigger urgent write-out no matter what
1593          * RPC size will be.
1594          * The exiting condition (other than success) is no avail grants
1595          * and no dirty pages caching, that really means there is no space
1596          * on the OST.
1597          */
1598         remain = wait_event_idle_exclusive_timeout_cmd(
1599                 cli->cl_cache_waiters,
1600                 (entered = osc_enter_cache_try(cli, oap, bytes)) ||
1601                 (cli->cl_dirty_pages == 0 && cli->cl_w_in_flight == 0),
1602                 timeout,
1603                 cli_unlock_and_unplug(env, cli, oap),
1604                 cli_lock_after_unplug(cli));
1605
1606         if (entered) {
1607                 if (remain == timeout)
1608                         OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1609                 else
1610                         OSC_DUMP_GRANT(D_CACHE, cli,
1611                                        "finally got grant space\n");
1612                 wake_up(&cli->cl_cache_waiters);
1613                 rc = 0;
1614         } else if (remain == 0) {
1615                 OSC_DUMP_GRANT(D_CACHE, cli,
1616                                "timeout, fall back to sync i/o\n");
1617                 osc_extent_tree_dump(D_CACHE, osc);
1618                 /* fall back to synchronous I/O */
1619         } else {
1620                 OSC_DUMP_GRANT(D_CACHE, cli,
1621                                "no grant space, fall back to sync i/o\n");
1622                 wake_up_all(&cli->cl_cache_waiters);
1623         }
1624         EXIT;
1625 out:
1626         spin_unlock(&cli->cl_loi_list_lock);
1627         RETURN(rc);
1628 }
1629
1630 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1631 {
1632         int hprpc = !!list_empty(&osc->oo_hp_exts);
1633         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1634 }
1635
1636 /* This maintains the lists of pending pages to read/write for a given object
1637  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1638  * to quickly find objects that are ready to send an RPC. */
1639 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1640                          int cmd)
1641 {
1642         int invalid_import = 0;
1643         ENTRY;
1644
1645         /* if we have an invalid import we want to drain the queued pages
1646          * by forcing them through rpcs that immediately fail and complete
1647          * the pages.  recovery relies on this to empty the queued pages
1648          * before canceling the locks and evicting down the llite pages */
1649         if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1650                 invalid_import = 1;
1651
1652         if (cmd & OBD_BRW_WRITE) {
1653                 if (atomic_read(&osc->oo_nr_writes) == 0)
1654                         RETURN(0);
1655                 if (invalid_import) {
1656                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1657                         RETURN(1);
1658                 }
1659                 if (!list_empty(&osc->oo_hp_exts)) {
1660                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1661                         RETURN(1);
1662                 }
1663                 if (!list_empty(&osc->oo_urgent_exts)) {
1664                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1665                         RETURN(1);
1666                 }
1667                 /* trigger a write rpc stream as long as there are dirtiers
1668                  * waiting for space.  as they're waiting, they're not going to
1669                  * create more pages to coalesce with what's waiting..
1670                  */
1671                 if (waitqueue_active(&cli->cl_cache_waiters)) {
1672                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1673                         RETURN(1);
1674                 }
1675                 if (!list_empty(&osc->oo_full_exts)) {
1676                         CDEBUG(D_CACHE, "full extent ready, make an RPC\n");
1677                         RETURN(1);
1678                 }
1679         } else {
1680                 if (atomic_read(&osc->oo_nr_reads) == 0)
1681                         RETURN(0);
1682                 if (invalid_import) {
1683                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1684                         RETURN(1);
1685                 }
1686                 /* all read are urgent. */
1687                 if (!list_empty(&osc->oo_reading_exts))
1688                         RETURN(1);
1689         }
1690
1691         RETURN(0);
1692 }
1693
1694 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1695 {
1696         struct client_obd *cli = osc_cli(obj);
1697         if (cmd & OBD_BRW_WRITE) {
1698                 atomic_add(delta, &obj->oo_nr_writes);
1699                 atomic_add(delta, &cli->cl_pending_w_pages);
1700                 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1701         } else {
1702                 atomic_add(delta, &obj->oo_nr_reads);
1703                 atomic_add(delta, &cli->cl_pending_r_pages);
1704                 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1705         }
1706         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1707 }
1708
1709 static int osc_makes_hprpc(struct osc_object *obj)
1710 {
1711         return !list_empty(&obj->oo_hp_exts);
1712 }
1713
1714 static void on_list(struct list_head *item, struct list_head *list,
1715                     int should_be_on)
1716 {
1717         if (list_empty(item) && should_be_on)
1718                 list_add_tail(item, list);
1719         else if (!list_empty(item) && !should_be_on)
1720                 list_del_init(item);
1721 }
1722
1723 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1724  * can find pages to build into rpcs quickly */
1725 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1726 {
1727         if (osc_makes_hprpc(osc)) {
1728                 /* HP rpc */
1729                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1730                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1731         } else {
1732                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1733                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1734                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1735                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1736         }
1737
1738         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1739                 atomic_read(&osc->oo_nr_writes) > 0);
1740
1741         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1742                 atomic_read(&osc->oo_nr_reads) > 0);
1743
1744         return osc_is_ready(osc);
1745 }
1746
1747 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1748 {
1749         int is_ready;
1750
1751         spin_lock(&cli->cl_loi_list_lock);
1752         is_ready = __osc_list_maint(cli, osc);
1753         spin_unlock(&cli->cl_loi_list_lock);
1754
1755         return is_ready;
1756 }
1757
1758 /* this must be called holding the loi list lock to give coverage to exit_cache,
1759  * async_flag maintenance
1760  */
1761 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1762                               struct osc_async_page *oap, int sent, int rc)
1763 {
1764         struct osc_object *osc = oap->oap_obj;
1765
1766         ENTRY;
1767
1768         /* As the transfer for this page is being done, clear the flags */
1769         oap->oap_async_flags = 0;
1770
1771         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1772         if (rc)
1773                 CERROR("completion on oap %p obj %p returns %d.\n",
1774                        oap, osc, rc);
1775
1776         EXIT;
1777 }
1778
1779 struct extent_rpc_data {
1780         struct list_head        *erd_rpc_list;
1781         unsigned int            erd_page_count;
1782         unsigned int            erd_max_pages;
1783         unsigned int            erd_max_chunks;
1784         unsigned int            erd_max_extents;
1785 };
1786
1787 static inline unsigned osc_extent_chunks(const struct osc_extent *ext)
1788 {
1789         struct client_obd *cli = osc_cli(ext->oe_obj);
1790         unsigned ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1791
1792         return (ext->oe_end >> ppc_bits) - (ext->oe_start >> ppc_bits) + 1;
1793 }
1794
1795 static inline bool
1796 can_merge(const struct osc_extent *ext, const struct osc_extent *in_rpc)
1797 {
1798         if (ext->oe_no_merge || in_rpc->oe_no_merge)
1799                 return false;
1800
1801         if (ext->oe_srvlock != in_rpc->oe_srvlock)
1802                 return false;
1803
1804         if (ext->oe_ndelay != in_rpc->oe_ndelay)
1805                 return false;
1806
1807         if (!ext->oe_grants != !in_rpc->oe_grants)
1808                 return false;
1809
1810         if (ext->oe_dio != in_rpc->oe_dio)
1811                 return false;
1812
1813         /* It's possible to have overlap on DIO */
1814         if (in_rpc->oe_dio && overlapped(ext, in_rpc))
1815                 return false;
1816
1817         if (ext->oe_is_rdma_only != in_rpc->oe_is_rdma_only)
1818                 return false;
1819
1820         return true;
1821 }
1822
1823 /**
1824  * Try to add extent to one RPC. We need to think about the following things:
1825  * - # of pages must not be over max_pages_per_rpc
1826  * - extent must be compatible with previous ones
1827  */
1828 static int try_to_add_extent_for_io(struct client_obd *cli,
1829                                     struct osc_extent *ext,
1830                                     struct extent_rpc_data *data)
1831 {
1832         struct osc_extent *tmp;
1833         unsigned int chunk_count;
1834         ENTRY;
1835
1836         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1837                 ext);
1838         OSC_EXTENT_DUMP(D_CACHE, ext, "trying to add this extent\n");
1839
1840         if (data->erd_max_extents == 0)
1841                 RETURN(0);
1842
1843         chunk_count = osc_extent_chunks(ext);
1844         EASSERTF(data->erd_page_count != 0 ||
1845                  chunk_count <= data->erd_max_chunks, ext,
1846                  "The first extent to be fit in a RPC contains %u chunks, "
1847                  "which is over the limit %u.\n", chunk_count,
1848                  data->erd_max_chunks);
1849         if (chunk_count > data->erd_max_chunks)
1850                 RETURN(0);
1851
1852         data->erd_max_pages = max(ext->oe_mppr, data->erd_max_pages);
1853         EASSERTF(data->erd_page_count != 0 ||
1854                 ext->oe_nr_pages <= data->erd_max_pages, ext,
1855                 "The first extent to be fit in a RPC contains %u pages, "
1856                 "which is over the limit %u.\n", ext->oe_nr_pages,
1857                 data->erd_max_pages);
1858         if (data->erd_page_count + ext->oe_nr_pages > data->erd_max_pages)
1859                 RETURN(0);
1860
1861         list_for_each_entry(tmp, data->erd_rpc_list, oe_link) {
1862                 EASSERT(tmp->oe_owner == current, tmp);
1863
1864                 if (!can_merge(ext, tmp))
1865                         RETURN(0);
1866         }
1867
1868         data->erd_max_extents--;
1869         data->erd_max_chunks -= chunk_count;
1870         data->erd_page_count += ext->oe_nr_pages;
1871         list_move_tail(&ext->oe_link, data->erd_rpc_list);
1872         ext->oe_owner = current;
1873         RETURN(1);
1874 }
1875
1876 /**
1877  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1878  * get_write_extent() takes all appropriate extents in atomic.
1879  *
1880  * The following policy is used to collect extents for IO:
1881  * 1. Add as many HP extents as possible;
1882  * 2. Add the first urgent extent in urgent extent list and take it out of
1883  *    urgent list;
1884  * 3. Add subsequent extents of this urgent extent;
1885  * 4. If urgent list is not empty, goto 2;
1886  * 5. Traverse the extent tree from the 1st extent;
1887  * 6. Above steps exit if there is no space in this RPC.
1888  */
1889 static unsigned int get_write_extents(struct osc_object *obj,
1890                                       struct list_head *rpclist)
1891 {
1892         struct client_obd *cli = osc_cli(obj);
1893         struct osc_extent *ext;
1894         struct extent_rpc_data data = {
1895                 .erd_rpc_list   = rpclist,
1896                 .erd_page_count = 0,
1897                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
1898                 .erd_max_chunks = osc_max_write_chunks(cli),
1899                 .erd_max_extents = 256,
1900         };
1901
1902         assert_osc_object_is_locked(obj);
1903         while ((ext = list_first_entry_or_null(&obj->oo_hp_exts,
1904                                                struct osc_extent,
1905                                                oe_link)) != NULL) {
1906                 if (!try_to_add_extent_for_io(cli, ext, &data))
1907                         return data.erd_page_count;
1908                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
1909         }
1910         if (data.erd_page_count == data.erd_max_pages)
1911                 return data.erd_page_count;
1912
1913         while ((ext = list_first_entry_or_null(&obj->oo_urgent_exts,
1914                                                struct osc_extent,
1915                                                oe_link)) != NULL) {
1916                 if (!try_to_add_extent_for_io(cli, ext, &data))
1917                         return data.erd_page_count;
1918         }
1919         if (data.erd_page_count == data.erd_max_pages)
1920                 return data.erd_page_count;
1921
1922         /* One key difference between full extents and other extents: full
1923          * extents can usually only be added if the rpclist was empty, so if we
1924          * can't add one, we continue on to trying to add normal extents.  This
1925          * is so we don't miss adding extra extents to an RPC containing high
1926          * priority or urgent extents.
1927          */
1928         while ((ext = list_first_entry_or_null(&obj->oo_full_exts,
1929                                                struct osc_extent,
1930                                                oe_link)) != NULL) {
1931                 if (!try_to_add_extent_for_io(cli, ext, &data))
1932                         break;
1933         }
1934         if (data.erd_page_count == data.erd_max_pages)
1935                 return data.erd_page_count;
1936
1937         for (ext = first_extent(obj);
1938              ext;
1939              ext = next_extent(ext)) {
1940                 if ((ext->oe_state != OES_CACHE) ||
1941                     /* this extent may be already in current rpclist */
1942                     (!list_empty(&ext->oe_link) && ext->oe_owner))
1943                         continue;
1944
1945                 if (!try_to_add_extent_for_io(cli, ext, &data))
1946                         return data.erd_page_count;
1947         }
1948         return data.erd_page_count;
1949 }
1950
1951 static int
1952 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
1953                    struct osc_object *osc)
1954 __must_hold(osc)
1955 {
1956         LIST_HEAD(rpclist);
1957         struct osc_extent *ext;
1958         struct osc_extent *tmp;
1959         struct osc_extent *first = NULL;
1960         unsigned int page_count = 0;
1961         int srvlock = 0;
1962         int rc = 0;
1963         ENTRY;
1964
1965         assert_osc_object_is_locked(osc);
1966
1967         page_count = get_write_extents(osc, &rpclist);
1968         LASSERT(equi(page_count == 0, list_empty(&rpclist)));
1969
1970         if (list_empty(&rpclist))
1971                 RETURN(0);
1972
1973         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
1974
1975         list_for_each_entry(ext, &rpclist, oe_link) {
1976                 LASSERT(ext->oe_state == OES_CACHE ||
1977                         ext->oe_state == OES_LOCK_DONE);
1978                 if (ext->oe_state == OES_CACHE)
1979                         osc_extent_state_set(ext, OES_LOCKING);
1980                 else
1981                         osc_extent_state_set(ext, OES_RPC);
1982         }
1983
1984         /* we're going to grab page lock, so release object lock because
1985          * lock order is page lock -> object lock. */
1986         osc_object_unlock(osc);
1987
1988         list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
1989                 if (ext->oe_state == OES_LOCKING) {
1990                         rc = osc_extent_make_ready(env, ext);
1991                         if (unlikely(rc < 0)) {
1992                                 list_del_init(&ext->oe_link);
1993                                 osc_extent_finish(env, ext, 0, rc);
1994                                 continue;
1995                         }
1996                 }
1997                 if (first == NULL) {
1998                         first = ext;
1999                         srvlock = ext->oe_srvlock;
2000                 } else {
2001                         LASSERT(srvlock == ext->oe_srvlock);
2002                 }
2003         }
2004
2005         if (!list_empty(&rpclist)) {
2006                 LASSERT(page_count > 0);
2007                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2008                 LASSERT(list_empty(&rpclist));
2009         }
2010
2011         osc_object_lock(osc);
2012         RETURN(rc);
2013 }
2014
2015 /**
2016  * prepare pages for ASYNC io and put pages in send queue.
2017  *
2018  * \param cmd OBD_BRW_* macroses
2019  * \param lop pending pages
2020  *
2021  * \return zero if no page added to send queue.
2022  * \return 1 if pages successfully added to send queue.
2023  * \return negative on errors.
2024  */
2025 static int
2026 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2027                   struct osc_object *osc)
2028 __must_hold(osc)
2029 {
2030         struct osc_extent *ext;
2031         struct osc_extent *next;
2032         LIST_HEAD(rpclist);
2033         struct extent_rpc_data data = {
2034                 .erd_rpc_list   = &rpclist,
2035                 .erd_page_count = 0,
2036                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
2037                 .erd_max_chunks = UINT_MAX,
2038                 .erd_max_extents = UINT_MAX,
2039         };
2040         int rc = 0;
2041         ENTRY;
2042
2043         assert_osc_object_is_locked(osc);
2044         list_for_each_entry_safe(ext, next, &osc->oo_reading_exts, oe_link) {
2045                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2046                 if (!try_to_add_extent_for_io(cli, ext, &data))
2047                         break;
2048                 osc_extent_state_set(ext, OES_RPC);
2049                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2050         }
2051         LASSERT(data.erd_page_count <= data.erd_max_pages);
2052
2053         osc_update_pending(osc, OBD_BRW_READ, -data.erd_page_count);
2054
2055         if (!list_empty(&rpclist)) {
2056                 osc_object_unlock(osc);
2057
2058                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2059                 LASSERT(list_empty(&rpclist));
2060
2061                 osc_object_lock(osc);
2062         }
2063         RETURN(rc);
2064 }
2065
2066 #define list_to_obj(list, item) ({                                            \
2067         struct list_head *__tmp = (list)->next;                               \
2068         list_del_init(__tmp);                                         \
2069         list_entry(__tmp, struct osc_object, oo_##item);                      \
2070 })
2071
2072 /* This is called by osc_check_rpcs() to find which objects have pages that
2073  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2074 static struct osc_object *osc_next_obj(struct client_obd *cli)
2075 {
2076         ENTRY;
2077
2078         /* First return objects that have blocked locks so that they
2079          * will be flushed quickly and other clients can get the lock,
2080          * then objects which have pages ready to be stuffed into RPCs */
2081         if (!list_empty(&cli->cl_loi_hp_ready_list))
2082                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2083         if (!list_empty(&cli->cl_loi_ready_list))
2084                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2085
2086         /* then if we have cache waiters, return all objects with queued
2087          * writes.  This is especially important when many small files
2088          * have filled up the cache and not been fired into rpcs because
2089          * they don't pass the nr_pending/object threshhold
2090          */
2091         if (waitqueue_active(&cli->cl_cache_waiters) &&
2092             !list_empty(&cli->cl_loi_write_list))
2093                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2094
2095         /* then return all queued objects when we have an invalid import
2096          * so that they get flushed */
2097         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2098                 if (!list_empty(&cli->cl_loi_write_list))
2099                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2100                                            write_item));
2101                 if (!list_empty(&cli->cl_loi_read_list))
2102                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2103                                            read_item));
2104         }
2105         RETURN(NULL);
2106 }
2107
2108 /* called with the loi list lock held */
2109 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2110 __must_hold(&cli->cl_loi_list_lock)
2111 {
2112         struct osc_object *osc;
2113         int rc = 0;
2114         ENTRY;
2115
2116         while ((osc = osc_next_obj(cli)) != NULL) {
2117                 struct cl_object *obj = osc2cl(osc);
2118                 struct lu_ref_link link;
2119
2120                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2121
2122                 /* even if we have reached our max in flight RPCs, we still
2123                  * allow all high-priority RPCs through to prevent their
2124                  * starvation and leading to server evicting us for not
2125                  * writing out pages in a timely manner LU-13131 */
2126                 if (osc_max_rpc_in_flight(cli, osc) &&
2127                     list_empty(&osc->oo_hp_exts)) {
2128                         __osc_list_maint(cli, osc);
2129                         break;
2130                 }
2131
2132                 cl_object_get(obj);
2133                 spin_unlock(&cli->cl_loi_list_lock);
2134                 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2135
2136                 /* attempt some read/write balancing by alternating between
2137                  * reads and writes in an object.  The makes_rpc checks here
2138                  * would be redundant if we were getting read/write work items
2139                  * instead of objects.  we don't want send_oap_rpc to drain a
2140                  * partial read pending queue when we're given this object to
2141                  * do io on writes while there are cache waiters */
2142                 osc_object_lock(osc);
2143                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2144                         rc = osc_send_write_rpc(env, cli, osc);
2145                         if (rc < 0) {
2146                                 CERROR("Write request failed with %d\n", rc);
2147
2148                                 /* osc_send_write_rpc failed, mostly because of
2149                                  * memory pressure.
2150                                  *
2151                                  * It can't break here, because if:
2152                                  *  - a page was submitted by osc_io_submit, so
2153                                  *    page locked;
2154                                  *  - no request in flight
2155                                  *  - no subsequent request
2156                                  * The system will be in live-lock state,
2157                                  * because there is no chance to call
2158                                  * osc_io_unplug() and osc_check_rpcs() any
2159                                  * more. pdflush can't help in this case,
2160                                  * because it might be blocked at grabbing
2161                                  * the page lock as we mentioned.
2162                                  *
2163                                  * Anyway, continue to drain pages. */
2164                                 /* break; */
2165                         }
2166                 }
2167                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2168                         rc = osc_send_read_rpc(env, cli, osc);
2169                         if (rc < 0)
2170                                 CERROR("Read request failed with %d\n", rc);
2171                 }
2172                 osc_object_unlock(osc);
2173
2174                 osc_list_maint(cli, osc);
2175                 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2176                 cl_object_put(env, obj);
2177
2178                 spin_lock(&cli->cl_loi_list_lock);
2179         }
2180         EXIT;
2181 }
2182
2183 int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2184                    struct osc_object *osc, int async)
2185 {
2186         int rc = 0;
2187
2188         if (osc != NULL && osc_list_maint(cli, osc) == 0)
2189                 return 0;
2190
2191         if (!async) {
2192                 spin_lock(&cli->cl_loi_list_lock);
2193                 osc_check_rpcs(env, cli);
2194                 spin_unlock(&cli->cl_loi_list_lock);
2195         } else {
2196                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2197                 LASSERT(cli->cl_writeback_work != NULL);
2198                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2199         }
2200         return rc;
2201 }
2202 EXPORT_SYMBOL(osc_io_unplug0);
2203
2204 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2205                         struct cl_page *page, loff_t offset)
2206 {
2207         struct osc_async_page *oap = &ops->ops_oap;
2208
2209         ENTRY;
2210         if (!page)
2211                 return round_up(sizeof(*oap), 8);
2212
2213         oap->oap_obj = osc;
2214         oap->oap_page = page->cp_vmpage;
2215         oap->oap_obj_off = offset;
2216         LASSERT(!(offset & ~PAGE_MASK));
2217
2218         /* Count of transient (direct i/o) pages is always stable by the time
2219          * they're submitted.  Setting this here lets us avoid calling
2220          * cl_page_clip later to set this.
2221          */
2222         if (page->cp_type == CPT_TRANSIENT)
2223                 oap->oap_async_flags |= ASYNC_COUNT_STABLE|ASYNC_URGENT|
2224                                         ASYNC_READY;
2225
2226         INIT_LIST_HEAD(&oap->oap_pending_item);
2227         INIT_LIST_HEAD(&oap->oap_rpc_item);
2228
2229         CDEBUG(D_INFO, "oap %p vmpage %p obj off %llu\n",
2230                oap, oap->oap_page, oap->oap_obj_off);
2231         RETURN(0);
2232 }
2233 EXPORT_SYMBOL(osc_prep_async_page);
2234
2235 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2236                        struct osc_page *ops, cl_commit_cbt cb)
2237 {
2238         struct osc_io *oio = osc_env_io(env);
2239         struct osc_extent     *ext = NULL;
2240         struct osc_async_page *oap = &ops->ops_oap;
2241         struct osc_object     *osc = oap->oap_obj;
2242         struct client_obd     *cli = osc_cli(osc);
2243         struct pagevec        *pvec = &osc_env_info(env)->oti_pagevec;
2244         pgoff_t index;
2245         unsigned int tmp;
2246         unsigned int grants = 0;
2247         u32    brw_flags = OBD_BRW_ASYNC;
2248         int    cmd = OBD_BRW_WRITE;
2249         int    need_release = 0;
2250         int    rc = 0;
2251         ENTRY;
2252
2253         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2254                 RETURN(-EIO);
2255
2256         if (!list_empty(&oap->oap_pending_item) ||
2257             !list_empty(&oap->oap_rpc_item))
2258                 RETURN(-EBUSY);
2259
2260         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2261         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2262         if (io->ci_noquota) {
2263                 brw_flags |= OBD_BRW_NOQUOTA;
2264                 cmd |= OBD_BRW_NOQUOTA;
2265         }
2266
2267         if (oio->oi_cap_sys_resource) {
2268                 brw_flags |= OBD_BRW_SYS_RESOURCE;
2269                 cmd |= OBD_BRW_SYS_RESOURCE;
2270         }
2271
2272         /* check if the file's owner/group is over quota */
2273         /* do not check for root without root squash, because in this case
2274          * we should bypass quota
2275          */
2276         if ((!oio->oi_cap_sys_resource ||
2277              cli->cl_root_squash || cli->cl_root_prjquota) &&
2278             !io->ci_noquota) {
2279                 struct cl_object *obj;
2280                 struct cl_attr   *attr;
2281                 unsigned int qid[LL_MAXQUOTAS];
2282
2283                 obj = cl_object_top(&osc->oo_cl);
2284                 attr = &osc_env_info(env)->oti_attr;
2285
2286                 cl_object_attr_lock(obj);
2287                 rc = cl_object_attr_get(env, obj, attr);
2288                 cl_object_attr_unlock(obj);
2289
2290                 qid[USRQUOTA] = attr->cat_uid;
2291                 qid[GRPQUOTA] = attr->cat_gid;
2292                 qid[PRJQUOTA] = attr->cat_projid;
2293                 if (rc == 0 && osc_quota_chkdq(cli, qid) == -EDQUOT)
2294                         rc = -EDQUOT;
2295                 if (rc)
2296                         RETURN(rc);
2297         }
2298
2299         oap->oap_cmd = cmd;
2300         oap->oap_page_off = ops->ops_from;
2301         oap->oap_count = ops->ops_to - ops->ops_from + 1;
2302         /* No need to hold a lock here,
2303          * since this page is not in any list yet. */
2304         oap->oap_async_flags = 0;
2305         oap->oap_brw_flags = brw_flags;
2306
2307         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2308                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2309
2310         index = osc_index(oap2osc(oap));
2311
2312         /* Add this page into extent by the following steps:
2313          * 1. if there exists an active extent for this IO, mostly this page
2314          *    can be added to the active extent and sometimes we need to
2315          *    expand extent to accomodate this page;
2316          * 2. otherwise, a new extent will be allocated. */
2317
2318         ext = oio->oi_active;
2319         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2320                 /* one chunk plus extent overhead must be enough to write this
2321                  * page */
2322                 grants = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2323                 if (ext->oe_end >= index)
2324                         grants = 0;
2325
2326                 /* it doesn't need any grant to dirty this page */
2327                 spin_lock(&cli->cl_loi_list_lock);
2328                 rc = osc_enter_cache_try(cli, oap, grants);
2329                 if (rc == 0) { /* try failed */
2330                         grants = 0;
2331                         need_release = 1;
2332                 } else if (ext->oe_end < index) {
2333                         tmp = grants;
2334                         /* try to expand this extent */
2335                         rc = osc_extent_expand(ext, index, &tmp);
2336                         if (rc < 0) {
2337                                 need_release = 1;
2338                                 /* don't free reserved grant */
2339                         } else {
2340                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2341                                                 "expanded for %lu.\n", index);
2342                                 osc_unreserve_grant_nolock(cli, grants, tmp);
2343                                 grants = 0;
2344                         }
2345                 }
2346                 spin_unlock(&cli->cl_loi_list_lock);
2347                 rc = 0;
2348         } else if (ext != NULL) {
2349                 /* index is located outside of active extent */
2350                 need_release = 1;
2351         }
2352         if (need_release) {
2353                 osc_extent_release(env, ext);
2354                 oio->oi_active = NULL;
2355                 ext = NULL;
2356         }
2357
2358         if (ext == NULL) {
2359                 tmp = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2360
2361                 /* try to find new extent to cover this page */
2362                 LASSERT(oio->oi_active == NULL);
2363                 /* we may have allocated grant for this page if we failed
2364                  * to expand the previous active extent. */
2365                 LASSERT(ergo(grants > 0, grants >= tmp));
2366
2367                 rc = 0;
2368
2369                 /* We must not hold a page lock while we do osc_enter_cache()
2370                  * or osc_extent_find(), so we must mark dirty & unlock
2371                  * any pages in the write commit pagevec. */
2372                 if (pagevec_count(pvec)) {
2373                         cb(env, io, pvec);
2374                         pagevec_reinit(pvec);
2375                 }
2376
2377                 if (grants == 0) {
2378                         rc = osc_enter_cache(env, cli, oap, tmp);
2379                         if (rc == 0)
2380                                 grants = tmp;
2381                 }
2382
2383                 tmp = grants;
2384                 if (rc == 0) {
2385                         ext = osc_extent_find(env, osc, index, &tmp);
2386                         if (IS_ERR(ext)) {
2387                                 LASSERT(tmp == grants);
2388                                 osc_exit_cache(cli, oap);
2389                                 rc = PTR_ERR(ext);
2390                                 ext = NULL;
2391                         } else {
2392                                 oio->oi_active = ext;
2393                         }
2394                 }
2395                 if (grants > 0)
2396                         osc_unreserve_grant(cli, grants, tmp);
2397         }
2398
2399         LASSERT(ergo(rc == 0, ext != NULL));
2400         if (ext != NULL) {
2401                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2402                          ext, "index = %lu.\n", index);
2403                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2404
2405                 osc_object_lock(osc);
2406                 if (ext->oe_nr_pages == 0)
2407                         ext->oe_srvlock = ops->ops_srvlock;
2408                 else
2409                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2410                 ++ext->oe_nr_pages;
2411                 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2412                 osc_object_unlock(osc);
2413
2414                 if (!ext->oe_layout_version)
2415                         ext->oe_layout_version = io->ci_layout_version;
2416         }
2417
2418         RETURN(rc);
2419 }
2420
2421 int osc_teardown_async_page(const struct lu_env *env,
2422                             struct osc_object *obj, struct osc_page *ops)
2423 {
2424         struct osc_async_page *oap = &ops->ops_oap;
2425         int rc = 0;
2426         ENTRY;
2427
2428         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2429                oap, ops, osc_index(oap2osc(oap)));
2430
2431         if (!list_empty(&oap->oap_rpc_item)) {
2432                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2433                 rc = -EBUSY;
2434         } else if (!list_empty(&oap->oap_pending_item)) {
2435                 struct osc_extent *ext = NULL;
2436
2437                 osc_object_lock(obj);
2438                 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2439                 osc_object_unlock(obj);
2440                 /* only truncated pages are allowed to be taken out.
2441                  * See osc_extent_truncate() and osc_cache_truncate_start()
2442                  * for details. */
2443                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2444                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2445                                         osc_index(oap2osc(oap)));
2446                         rc = -EBUSY;
2447                 }
2448                 if (ext != NULL)
2449                         osc_extent_put(env, ext);
2450         }
2451         RETURN(rc);
2452 }
2453
2454 /**
2455  * This is called when a page is picked up by kernel to write out.
2456  *
2457  * We should find out the corresponding extent and add the whole extent
2458  * into urgent list. The extent may be being truncated or used, handle it
2459  * carefully.
2460  */
2461 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2462                          struct osc_page *ops)
2463 {
2464         struct osc_extent *ext   = NULL;
2465         struct osc_object *obj   = osc_page_object(ops);
2466         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2467         pgoff_t            index = osc_index(ops);
2468         struct osc_async_page *oap = &ops->ops_oap;
2469         bool unplug = false;
2470         int rc = 0;
2471         ENTRY;
2472
2473         osc_object_lock(obj);
2474         ext = osc_extent_lookup(obj, index);
2475         if (ext == NULL) {
2476                 osc_extent_tree_dump(D_ERROR, obj);
2477                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2478         }
2479
2480         switch (ext->oe_state) {
2481         case OES_RPC:
2482         case OES_LOCK_DONE:
2483                 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2484                 LASSERT(0);
2485                 break;
2486         case OES_LOCKING:
2487                 /* If we know this extent is being written out, we should abort
2488                  * so that the writer can make this page ready. Otherwise, there
2489                  * exists a deadlock problem because other process can wait for
2490                  * page writeback bit holding page lock; and meanwhile in
2491                  * vvp_page_make_ready(), we need to grab page lock before
2492                  * really sending the RPC. */
2493         case OES_TRUNC:
2494                 /* race with truncate, page will be redirtied */
2495         case OES_ACTIVE:
2496                 /* The extent is active so we need to abort and let the caller
2497                  * re-dirty the page. If we continued on here, and we were the
2498                  * one making the extent active, we could deadlock waiting for
2499                  * the page writeback to clear but it won't because the extent
2500                  * is active and won't be written out. */
2501                 GOTO(out, rc = -EAGAIN);
2502         default:
2503                 break;
2504         }
2505
2506         rc = cl_page_prep(env, io, cp, CRT_WRITE);
2507         if (rc)
2508                 GOTO(out, rc);
2509
2510         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2511
2512         if (current->flags & PF_MEMALLOC)
2513                 ext->oe_memalloc = 1;
2514
2515         ext->oe_urgent = 1;
2516         if (ext->oe_state == OES_CACHE) {
2517                 OSC_EXTENT_DUMP(D_CACHE, ext,
2518                                 "flush page %p make it urgent.\n", oap);
2519                 if (list_empty(&ext->oe_link))
2520                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2521                 unplug = true;
2522         }
2523         rc = 0;
2524         EXIT;
2525
2526 out:
2527         osc_object_unlock(obj);
2528         osc_extent_put(env, ext);
2529         if (unplug)
2530                 osc_io_unplug_async(env, osc_cli(obj), obj);
2531         return rc;
2532 }
2533
2534 int osc_queue_sync_pages(const struct lu_env *env, struct cl_io *io,
2535                          struct osc_object *obj, struct list_head *list,
2536                          int brw_flags)
2537 {
2538         struct osc_io *oio = osc_env_io(env);
2539         struct client_obd     *cli = osc_cli(obj);
2540         struct osc_extent     *ext;
2541         struct osc_async_page *oap;
2542         int     page_count = 0;
2543         int     mppr       = cli->cl_max_pages_per_rpc;
2544         bool    can_merge   = true;
2545         pgoff_t start      = CL_PAGE_EOF;
2546         pgoff_t end        = 0;
2547         struct osc_lock *oscl;
2548         ENTRY;
2549
2550         list_for_each_entry(oap, list, oap_pending_item) {
2551                 struct osc_page *opg = oap2osc_page(oap);
2552                 pgoff_t index = osc_index(opg);
2553
2554                 if (index > end)
2555                         end = index;
2556                 if (index < start)
2557                         start = index;
2558                 ++page_count;
2559                 mppr <<= (page_count > mppr);
2560
2561                 if (unlikely(opg->ops_from > 0 ||
2562                              opg->ops_to < PAGE_SIZE - 1))
2563                         can_merge = false;
2564         }
2565
2566         ext = osc_extent_alloc(obj);
2567         if (ext == NULL) {
2568                 struct osc_async_page *tmp;
2569
2570                 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2571                         list_del_init(&oap->oap_pending_item);
2572                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2573                 }
2574                 RETURN(-ENOMEM);
2575         }
2576
2577         ext->oe_rw = !!(brw_flags & OBD_BRW_READ);
2578         ext->oe_sync = 1;
2579         ext->oe_no_merge = !can_merge;
2580         ext->oe_urgent = 1;
2581         ext->oe_start = start;
2582         ext->oe_end = ext->oe_max_end = end;
2583         ext->oe_obj = obj;
2584         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2585         ext->oe_ndelay = !!(brw_flags & OBD_BRW_NDELAY);
2586         ext->oe_dio = !!(brw_flags & OBD_BRW_NOCACHE);
2587         oscl = oio->oi_write_osclock ? : oio->oi_read_osclock;
2588         if (oscl && oscl->ols_dlmlock != NULL) {
2589                 ext->oe_dlmlock = LDLM_LOCK_GET(oscl->ols_dlmlock);
2590                 lu_ref_add(&ext->oe_dlmlock->l_reference, "osc_extent", ext);
2591         }
2592         if (ext->oe_dio && !ext->oe_rw) { /* direct io write */
2593                 int grants;
2594                 int ppc;
2595
2596                 ppc = 1 << (cli->cl_chunkbits - PAGE_SHIFT);
2597                 grants = cli->cl_grant_extent_tax;
2598                 grants += (1 << cli->cl_chunkbits) *
2599                         ((page_count + ppc - 1) / ppc);
2600
2601                 CDEBUG(D_CACHE, "requesting %d bytes grant\n", grants);
2602                 spin_lock(&cli->cl_loi_list_lock);
2603                 if (osc_reserve_grant(cli, grants) == 0) {
2604                         list_for_each_entry(oap, list, oap_pending_item) {
2605                                 osc_consume_write_grant(cli,
2606                                                         &oap->oap_brw_page);
2607                         }
2608                         atomic_long_add(page_count, &obd_dirty_pages);
2609                         osc_unreserve_grant_nolock(cli, grants, 0);
2610                         ext->oe_grants = grants;
2611                 } else {
2612                         /* We cannot report ENOSPC correctly if we do parallel
2613                          * DIO (async RPC submission), so turn off parallel dio
2614                          * if there is not sufficient grant available.  This
2615                          * makes individual RPCs synchronous.
2616                          */
2617                         io->ci_parallel_dio = false;
2618                         CDEBUG(D_CACHE,
2619                         "not enough grant available, switching to sync for this i/o\n");
2620                 }
2621                 spin_unlock(&cli->cl_loi_list_lock);
2622                 osc_update_next_shrink(cli);
2623         }
2624
2625         ext->oe_is_rdma_only = !!(brw_flags & OBD_BRW_RDMA_ONLY);
2626         ext->oe_nr_pages = page_count;
2627         ext->oe_mppr = mppr;
2628         list_splice_init(list, &ext->oe_pages);
2629         ext->oe_layout_version = io->ci_layout_version;
2630
2631         osc_object_lock(obj);
2632         /* Reuse the initial refcount for RPC, don't drop it */
2633         osc_extent_state_set(ext, OES_LOCK_DONE);
2634         if (!ext->oe_rw) { /* write */
2635                 if (!ext->oe_srvlock && !ext->oe_dio) {
2636                         /* The most likely case here is from lack of grants
2637                          * so we are either out of quota or out of space.
2638                          * Since this means we are holding locks across
2639                          * potentially multi-striped IO, we must send out
2640                          * everything out instantly to avoid prolonged
2641                          * waits resulting in lock eviction (likely since
2642                          * the extended wait in osc_cache_enter() did not
2643                          * yield any additional grant due to a timeout.
2644                          * LU-13131 */
2645                         ext->oe_hp = 1;
2646                         list_add_tail(&ext->oe_link, &obj->oo_hp_exts);
2647                 } else {
2648                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2649                 }
2650                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2651         } else {
2652                 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2653                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2654         }
2655         osc_object_unlock(obj);
2656
2657         osc_io_unplug_async(env, cli, obj);
2658         RETURN(0);
2659 }
2660
2661 /**
2662  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2663  */
2664 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
2665                              __u64 size, struct osc_extent **extp)
2666 {
2667         struct client_obd *cli = osc_cli(obj);
2668         struct osc_extent *ext;
2669         struct osc_extent *waiting = NULL;
2670         pgoff_t index;
2671         LIST_HEAD(list);
2672         int result = 0;
2673         bool partial;
2674         ENTRY;
2675
2676         /* pages with index greater or equal to index will be truncated. */
2677         index = size >> PAGE_SHIFT;
2678         partial = size > (index << PAGE_SHIFT);
2679
2680 again:
2681         osc_object_lock(obj);
2682         ext = osc_extent_search(obj, index);
2683         if (ext == NULL)
2684                 ext = first_extent(obj);
2685         else if (ext->oe_end < index)
2686                 ext = next_extent(ext);
2687         while (ext != NULL) {
2688                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2689
2690                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2691                         /* if ext is in urgent state, it means there must exist
2692                          * a page already having been flushed by write_page().
2693                          * We have to wait for this extent because we can't
2694                          * truncate that page. */
2695                         OSC_EXTENT_DUMP(D_CACHE, ext,
2696                                         "waiting for busy extent\n");
2697                         waiting = osc_extent_get(ext);
2698                         break;
2699                 }
2700
2701                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:%llu.\n", size);
2702
2703                 osc_extent_get(ext);
2704                 if (ext->oe_state == OES_ACTIVE) {
2705                         /* though we grab inode mutex for write path, but we
2706                          * release it before releasing extent(in osc_io_end()),
2707                          * so there is a race window that an extent is still
2708                          * in OES_ACTIVE when truncate starts. */
2709                         LASSERT(!ext->oe_trunc_pending);
2710                         ext->oe_trunc_pending = 1;
2711                 } else {
2712                         EASSERT(ext->oe_state == OES_CACHE, ext);
2713                         osc_extent_state_set(ext, OES_TRUNC);
2714                         osc_update_pending(obj, OBD_BRW_WRITE,
2715                                            -ext->oe_nr_pages);
2716                 }
2717                 /* This extent could be on the full extents list, that's OK */
2718                 EASSERT(!ext->oe_hp && !ext->oe_urgent, ext);
2719                 if (!list_empty(&ext->oe_link))
2720                         list_move_tail(&ext->oe_link, &list);
2721                 else
2722                         list_add_tail(&ext->oe_link, &list);
2723
2724                 ext = next_extent(ext);
2725         }
2726         osc_object_unlock(obj);
2727
2728         osc_list_maint(cli, obj);
2729
2730         while ((ext = list_first_entry_or_null(&list,
2731                                                struct osc_extent,
2732                                                oe_link)) != NULL) {
2733                 int rc;
2734
2735                 list_del_init(&ext->oe_link);
2736
2737                 /* extent may be in OES_ACTIVE state because inode mutex
2738                  * is released before osc_io_end() in file write case */
2739                 if (ext->oe_state != OES_TRUNC)
2740                         osc_extent_wait(env, ext, OES_TRUNC);
2741
2742                 rc = osc_extent_truncate(ext, index, partial);
2743                 if (rc < 0) {
2744                         if (result == 0)
2745                                 result = rc;
2746
2747                         OSC_EXTENT_DUMP(D_ERROR, ext,
2748                                         "truncate error %d\n", rc);
2749                 } else if (ext->oe_nr_pages == 0) {
2750                         osc_extent_remove(ext);
2751                 } else {
2752                         /* this must be an overlapped extent which means only
2753                          * part of pages in this extent have been truncated.
2754                          */
2755                         EASSERTF(ext->oe_start <= index, ext,
2756                                  "trunc index = %lu/%d.\n", index, partial);
2757                         /* fix index to skip this partially truncated extent */
2758                         index = ext->oe_end + 1;
2759                         partial = false;
2760
2761                         /* we need to hold this extent in OES_TRUNC state so
2762                          * that no writeback will happen. This is to avoid
2763                          * BUG 17397.
2764                          * Only partial truncate can reach here, if @size is
2765                          * not zero, the caller should provide a valid @extp. */
2766                         LASSERT(*extp == NULL);
2767                         *extp = osc_extent_get(ext);
2768                         OSC_EXTENT_DUMP(D_CACHE, ext,
2769                                         "trunc at %llu\n", size);
2770                 }
2771                 osc_extent_put(env, ext);
2772         }
2773         if (waiting != NULL) {
2774                 int rc;
2775
2776                 /* ignore the result of osc_extent_wait the write initiator
2777                  * should take care of it. */
2778                 rc = osc_extent_wait(env, waiting, OES_INV);
2779                 if (rc < 0)
2780                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2781
2782                 osc_extent_put(env, waiting);
2783                 waiting = NULL;
2784                 goto again;
2785         }
2786         RETURN(result);
2787 }
2788 EXPORT_SYMBOL(osc_cache_truncate_start);
2789
2790 /**
2791  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2792  */
2793 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext)
2794 {
2795         if (ext != NULL) {
2796                 struct osc_object *obj = ext->oe_obj;
2797                 bool unplug = false;
2798
2799                 EASSERT(ext->oe_nr_pages > 0, ext);
2800                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2801                 EASSERT(!ext->oe_urgent, ext);
2802
2803                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2804                 osc_object_lock(obj);
2805                 osc_extent_state_set(ext, OES_CACHE);
2806                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2807                         ext->oe_urgent = 1;
2808                         list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2809                         unplug = true;
2810                 }
2811                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2812                 osc_object_unlock(obj);
2813                 osc_extent_put(env, ext);
2814
2815                 if (unplug)
2816                         osc_io_unplug_async(env, osc_cli(obj), obj);
2817         }
2818 }
2819
2820 /**
2821  * Wait for extents in a specific range to be written out.
2822  * The caller must have called osc_cache_writeback_range() to issue IO
2823  * otherwise it will take a long time for this function to finish.
2824  *
2825  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2826  * nobody else can dirty this range of file while we're waiting for
2827  * extents to be written.
2828  */
2829 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2830                          pgoff_t start, pgoff_t end)
2831 {
2832         struct osc_extent *ext;
2833         pgoff_t index = start;
2834         int     result = 0;
2835         ENTRY;
2836
2837 again:
2838         osc_object_lock(obj);
2839         ext = osc_extent_search(obj, index);
2840         if (ext == NULL)
2841                 ext = first_extent(obj);
2842         else if (ext->oe_end < index)
2843                 ext = next_extent(ext);
2844         while (ext != NULL) {
2845                 int rc;
2846
2847                 if (ext->oe_start > end)
2848                         break;
2849
2850                 if (!ext->oe_fsync_wait) {
2851                         ext = next_extent(ext);
2852                         continue;
2853                 }
2854
2855                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2856                              ext->oe_hp || ext->oe_urgent), ext);
2857                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2858                              !ext->oe_hp && ext->oe_urgent), ext);
2859
2860                 index = ext->oe_end + 1;
2861                 osc_extent_get(ext);
2862                 osc_object_unlock(obj);
2863
2864                 rc = osc_extent_wait(env, ext, OES_INV);
2865                 if (result == 0)
2866                         result = rc;
2867                 osc_extent_put(env, ext);
2868                 goto again;
2869         }
2870         osc_object_unlock(obj);
2871
2872         OSC_IO_DEBUG(obj, "sync file range.\n");
2873         RETURN(result);
2874 }
2875 EXPORT_SYMBOL(osc_cache_wait_range);
2876
2877 /**
2878  * Called to write out a range of osc object.
2879  *
2880  * @hp     : should be set this is caused by lock cancel;
2881  * @discard: is set if dirty pages should be dropped - file will be deleted or
2882  *         truncated, this implies there is no partially discarding extents.
2883  *
2884  * Return how many pages will be issued, or error code if error occurred.
2885  */
2886 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2887                               pgoff_t start, pgoff_t end, int hp, int discard)
2888 {
2889         struct osc_extent *ext;
2890         LIST_HEAD(discard_list);
2891         bool unplug = false;
2892         int result = 0;
2893         ENTRY;
2894
2895         osc_object_lock(obj);
2896         ext = osc_extent_search(obj, start);
2897         if (ext == NULL)
2898                 ext = first_extent(obj);
2899         else if (ext->oe_end < start)
2900                 ext = next_extent(ext);
2901         while (ext != NULL) {
2902                 if (ext->oe_start > end)
2903                         break;
2904
2905                 ext->oe_fsync_wait = 1;
2906                 switch (ext->oe_state) {
2907                 case OES_CACHE:
2908                         result += ext->oe_nr_pages;
2909                         if (!discard) {
2910                                 struct list_head *list = NULL;
2911                                 if (hp) {
2912                                         EASSERT(!ext->oe_hp, ext);
2913                                         ext->oe_hp = 1;
2914                                         list = &obj->oo_hp_exts;
2915                                 } else if (!ext->oe_urgent && !ext->oe_hp) {
2916                                         ext->oe_urgent = 1;
2917                                         list = &obj->oo_urgent_exts;
2918                                 }
2919                                 if (list != NULL)
2920                                         list_move_tail(&ext->oe_link, list);
2921                                 unplug = true;
2922                         } else {
2923                                 struct client_obd *cli = osc_cli(obj);
2924                                 int pcc_bits = cli->cl_chunkbits - PAGE_SHIFT;
2925                                 pgoff_t align_by = (1 << pcc_bits);
2926                                 pgoff_t a_start = round_down(start, align_by);
2927                                 pgoff_t a_end = round_up(end, align_by);
2928
2929                                 /* overflow case */
2930                                 if (end && !a_end)
2931                                         a_end = CL_PAGE_EOF;
2932                                 /* the only discarder is lock cancelling, so
2933                                  * [start, end], aligned by chunk size, must
2934                                  * contain this extent */
2935                                 LASSERTF(ext->oe_start >= a_start &&
2936                                          ext->oe_end <= a_end,
2937                                          "ext [%lu, %lu] reg [%lu, %lu] "
2938                                          "orig [%lu %lu] align %lu bits "
2939                                          "%d\n", ext->oe_start, ext->oe_end,
2940                                          a_start, a_end, start, end,
2941                                          align_by, pcc_bits);
2942                                 osc_extent_state_set(ext, OES_LOCKING);
2943                                 ext->oe_owner = current;
2944                                 list_move_tail(&ext->oe_link,
2945                                                    &discard_list);
2946                                 osc_update_pending(obj, OBD_BRW_WRITE,
2947                                                    -ext->oe_nr_pages);
2948                         }
2949                         break;
2950                 case OES_ACTIVE:
2951                         /* It's pretty bad to wait for ACTIVE extents, because
2952                          * we don't know how long we will wait for it to be
2953                          * flushed since it may be blocked at awaiting more
2954                          * grants. We do this for the correctness of fsync. */
2955                         LASSERT(hp == 0 && discard == 0);
2956                         ext->oe_urgent = 1;
2957                         break;
2958                 case OES_TRUNC:
2959                         /* this extent is being truncated, can't do anything
2960                          * for it now. it will be set to urgent after truncate
2961                          * is finished in osc_cache_truncate_end(). */
2962                 default:
2963                         break;
2964                 }
2965                 ext = next_extent(ext);
2966         }
2967         osc_object_unlock(obj);
2968
2969         LASSERT(ergo(!discard, list_empty(&discard_list)));
2970         if (!list_empty(&discard_list)) {
2971                 struct osc_extent *tmp;
2972                 int rc;
2973
2974                 osc_list_maint(osc_cli(obj), obj);
2975                 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
2976                         list_del_init(&ext->oe_link);
2977                         EASSERT(ext->oe_state == OES_LOCKING, ext);
2978
2979                         /* Discard caching pages. We don't actually write this
2980                          * extent out but we complete it as if we did. */
2981                         rc = osc_extent_make_ready(env, ext);
2982                         if (unlikely(rc < 0)) {
2983                                 OSC_EXTENT_DUMP(D_ERROR, ext,
2984                                                 "make_ready returned %d\n", rc);
2985                                 if (result >= 0)
2986                                         result = rc;
2987                         }
2988
2989                         /* finish the extent as if the pages were sent */
2990                         osc_extent_finish(env, ext, 0, 0);
2991                 }
2992         }
2993
2994         if (unplug)
2995                 osc_io_unplug(env, osc_cli(obj), obj);
2996
2997         if (hp || discard) {
2998                 int rc;
2999                 rc = osc_cache_wait_range(env, obj, start, end);
3000                 if (result >= 0 && rc < 0)
3001                         result = rc;
3002         }
3003
3004         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3005         RETURN(result);
3006 }
3007 EXPORT_SYMBOL(osc_cache_writeback_range);
3008
3009 /**
3010  * Returns a list of pages by a given [start, end] of \a obj.
3011  *
3012  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3013  * crucial in the face of [offset, EOF] locks.
3014  *
3015  * Return at least one page in @queue unless there is no covered page.
3016  */
3017 bool osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3018                           struct osc_object *osc, pgoff_t start, pgoff_t end,
3019                           osc_page_gang_cbt cb, void *cbdata)
3020 {
3021         struct osc_page *ops;
3022         struct pagevec  *pagevec;
3023         void            **pvec;
3024         pgoff_t         idx;
3025         unsigned int    nr;
3026         unsigned int    i;
3027         unsigned int    j;
3028         bool            res = true;
3029         bool            tree_lock = true;
3030         ENTRY;
3031
3032         idx = start;
3033         pvec = osc_env_info(env)->oti_pvec;
3034         pagevec = &osc_env_info(env)->oti_pagevec;
3035         ll_pagevec_init(pagevec, 0);
3036         spin_lock(&osc->oo_tree_lock);
3037         while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3038                                             idx, OTI_PVEC_SIZE)) > 0) {
3039                 struct cl_page *page;
3040                 bool end_of_region = false;
3041
3042                 for (i = 0, j = 0; i < nr; ++i) {
3043                         ops = pvec[i];
3044                         pvec[i] = NULL;
3045
3046                         idx = osc_index(ops);
3047                         if (idx > end) {
3048                                 end_of_region = true;
3049                                 break;
3050                         }
3051
3052                         page = ops->ops_cl.cpl_page;
3053                         LASSERT(page->cp_type == CPT_CACHEABLE);
3054                         if (page->cp_state == CPS_FREEING)
3055                                 continue;
3056
3057                         cl_page_get(page);
3058                         lu_ref_add_atomic(&page->cp_reference,
3059                                           "gang_lookup", current);
3060                         pvec[j++] = ops;
3061                 }
3062                 ++idx;
3063
3064                 /*
3065                  * Here a delicate locking dance is performed. Current thread
3066                  * holds a reference to a page, but has to own it before it
3067                  * can be placed into queue. Owning implies waiting, so
3068                  * radix-tree lock is to be released. After a wait one has to
3069                  * check that pages weren't truncated (cl_page_own() returns
3070                  * error in the latter case).
3071                  */
3072                 spin_unlock(&osc->oo_tree_lock);
3073                 tree_lock = false;
3074
3075                 res = (*cb)(env, io, pvec, j, cbdata);
3076
3077                 for (i = 0; i < j; ++i) {
3078                         ops = pvec[i];
3079                         page = ops->ops_cl.cpl_page;
3080                         lu_ref_del(&page->cp_reference, "gang_lookup", current);
3081                         cl_pagevec_put(env, page, pagevec);
3082                 }
3083                 pagevec_release(pagevec);
3084
3085                 if (nr < OTI_PVEC_SIZE || end_of_region)
3086                         break;
3087
3088                 if (!res)
3089                         break;
3090
3091                 CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_SLOW_PAGE_EVICT,
3092                                  cfs_fail_val ?: 20);
3093
3094                 if (io->ci_type == CIT_MISC &&
3095                     io->u.ci_misc.lm_next_rpc_time &&
3096                     ktime_get_seconds() > io->u.ci_misc.lm_next_rpc_time) {
3097                         osc_send_empty_rpc(osc, idx << PAGE_SHIFT);
3098                         io->u.ci_misc.lm_next_rpc_time = ktime_get_seconds() +
3099                                                          5 * obd_timeout / 16;
3100                 }
3101
3102                 if (need_resched())
3103                         cond_resched();
3104
3105                 spin_lock(&osc->oo_tree_lock);
3106                 tree_lock = true;
3107         }
3108         if (tree_lock)
3109                 spin_unlock(&osc->oo_tree_lock);
3110         RETURN(res);
3111 }
3112 EXPORT_SYMBOL(osc_page_gang_lookup);
3113
3114 /**
3115  * Check if page @page is covered by an extra lock or discard it.
3116  */
3117 static bool check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3118                                  void **pvec, int count, void *cbdata)
3119 {
3120         struct osc_thread_info *info = osc_env_info(env);
3121         struct osc_object *osc = cbdata;
3122         int i;
3123
3124         for (i = 0; i < count; i++) {
3125                 struct osc_page *ops = pvec[i];
3126                 struct cl_page *page = ops->ops_cl.cpl_page;
3127                 pgoff_t index = osc_index(ops);
3128                 bool discard = false;
3129
3130                 /* negative lock caching */
3131                 if (index < info->oti_ng_index) {
3132                         discard = true;
3133                 } else if (index >= info->oti_fn_index) {
3134                         struct ldlm_lock *tmp;
3135                         /* refresh non-overlapped index */
3136                         tmp = osc_dlmlock_at_pgoff(env, osc, index,
3137                                         OSC_DAP_FL_TEST_LOCK |
3138                                         OSC_DAP_FL_AST |
3139                                         OSC_DAP_FL_RIGHT);
3140                         if (tmp != NULL) {
3141                                 __u64 end =
3142                                         tmp->l_policy_data.l_extent.end;
3143                                 __u64 start =
3144                                         tmp->l_policy_data.l_extent.start;
3145
3146                                 /* no lock covering this page */
3147                                 if (index < start >> PAGE_SHIFT) {
3148                                         /* no lock at @index,
3149                                          * first lock at @start
3150                                          */
3151                                         info->oti_ng_index =
3152                                                 start >> PAGE_SHIFT;
3153                                         discard = true;
3154                                 } else {
3155                                         /* Cache the first-non-overlapped
3156                                          * index so as to skip all pages
3157                                          * within [index, oti_fn_index).
3158                                          * This is safe because if tmp lock
3159                                          * is canceled, it will discard these
3160                                          * pages.
3161                                          */
3162                                         info->oti_fn_index =
3163                                                 (end + 1) >> PAGE_SHIFT;
3164                                         if (end == OBD_OBJECT_EOF)
3165                                                 info->oti_fn_index =
3166                                                         CL_PAGE_EOF;
3167                                 }
3168                                 LDLM_LOCK_PUT(tmp);
3169                         } else {
3170                                 info->oti_ng_index = CL_PAGE_EOF;
3171                                 discard = true;
3172                         }
3173                 }
3174
3175                 if (discard) {
3176                         if (cl_page_own(env, io, page) == 0) {
3177                                 cl_page_discard(env, io, page);
3178                                 cl_page_disown(env, io, page);
3179                         } else {
3180                                 LASSERT(page->cp_state == CPS_FREEING);
3181                         }
3182                 }
3183
3184                 info->oti_next_index = index + 1;
3185         }
3186         return true;
3187 }
3188
3189 bool osc_discard_cb(const struct lu_env *env, struct cl_io *io,
3190                     void **pvec, int count, void *cbdata)
3191 {
3192         struct osc_thread_info *info = osc_env_info(env);
3193         int i;
3194
3195         for (i = 0; i < count; i++) {
3196                 struct osc_page *ops = pvec[i];
3197                 struct cl_page *page = ops->ops_cl.cpl_page;
3198
3199                 /* page is top page. */
3200                 info->oti_next_index = osc_index(ops) + 1;
3201                 if (cl_page_own(env, io, page) == 0) {
3202                         if (!ergo(page->cp_type == CPT_CACHEABLE,
3203                                   !PageDirty(cl_page_vmpage(page))))
3204                                 CL_PAGE_DEBUG(D_ERROR, env, page,
3205                                               "discard dirty page?\n");
3206
3207                         /* discard the page */
3208                         cl_page_discard(env, io, page);
3209                         cl_page_disown(env, io, page);
3210                 } else {
3211                         LASSERT(page->cp_state == CPS_FREEING);
3212                 }
3213         }
3214
3215         return true;
3216 }
3217 EXPORT_SYMBOL(osc_discard_cb);
3218
3219 /**
3220  * Discard pages protected by the given lock. This function traverses radix
3221  * tree to find all covering pages and discard them. If a page is being covered
3222  * by other locks, it should remain in cache.
3223  *
3224  * If error happens on any step, the process continues anyway (the reasoning
3225  * behind this being that lock cancellation cannot be delayed indefinitely).
3226  */
3227 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3228                            pgoff_t start, pgoff_t end, bool discard)
3229 {
3230         struct osc_thread_info *info = osc_env_info(env);
3231         struct cl_io *io = osc_env_thread_io(env);
3232         osc_page_gang_cbt cb;
3233         int result;
3234
3235         ENTRY;
3236
3237         io->ci_obj = cl_object_top(osc2cl(osc));
3238         io->ci_ignore_layout = 1;
3239         io->ci_invalidate_page_cache = 1;
3240         io->u.ci_misc.lm_next_rpc_time = ktime_get_seconds() +
3241                                          5 * obd_timeout / 16;
3242         result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3243         if (result != 0)
3244                 GOTO(out, result);
3245
3246         cb = discard ? osc_discard_cb : check_and_discard_cb;
3247         info->oti_fn_index = info->oti_next_index = start;
3248         info->oti_ng_index = 0;
3249
3250         osc_page_gang_lookup(env, io, osc,
3251                              info->oti_next_index, end, cb, osc);
3252 out:
3253         cl_io_fini(env, io);
3254         RETURN(result);
3255 }
3256
3257
3258 /** @} osc */