lustre/osc/osc_cache.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  *
  28  */
  29 /*
  30  * This file is part of Lustre, http://www.lustre.org/
  31  *
  32  * osc cache management.
  33  *
  34  * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_OSC
  38
  39 #include <lustre_osc.h>
  40 #include <lustre_dlm.h>
  41
  42 #include "osc_internal.h"
  43
  44 static int extent_debug; /* set it to be true for more debug */
  45
  46 static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
  47 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
  48                            enum osc_extent_state state);
  49 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
  50                               struct osc_async_page *oap, int sent, int rc);
  51 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
  52                           int cmd);
  53 static int osc_refresh_count(const struct lu_env *env,
  54                              struct osc_async_page *oap, int cmd);
  55 static int osc_io_unplug_async(const struct lu_env *env,
  56                                struct client_obd *cli, struct osc_object *osc);
  57 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
  58                            unsigned int lost_grant, unsigned int dirty_grant);
  59
  60 static void osc_extent_tree_dump0(int mask, struct osc_object *obj,
  61                                   const char *func, int line);
  62 #define osc_extent_tree_dump(mask, obj) \
  63         osc_extent_tree_dump0(mask, obj, __func__, __LINE__)
  64
  65 static void osc_unreserve_grant(struct client_obd *cli, unsigned int reserved,
  66                                 unsigned int unused);
  67
  68 /** \addtogroup osc
  69  *  @{
  70  */
  71
  72 /* ------------------ osc extent ------------------ */
  73 static inline char *ext_flags(struct osc_extent *ext, char *flags)
  74 {
  75         char *buf = flags;
  76         *buf++ = ext->oe_rw ? 'r' : 'w';
  77         if (!RB_EMPTY_NODE(&ext->oe_node))
  78                 *buf++ = 'i';
  79         if (ext->oe_sync)
  80                 *buf++ = 'S';
  81         if (ext->oe_srvlock)
  82                 *buf++ = 's';
  83         if (ext->oe_hp)
  84                 *buf++ = 'h';
  85         if (ext->oe_urgent)
  86                 *buf++ = 'u';
  87         if (ext->oe_memalloc)
  88                 *buf++ = 'm';
  89         if (ext->oe_trunc_pending)
  90                 *buf++ = 't';
  91         if (ext->oe_fsync_wait)
  92                 *buf++ = 'Y';
  93         *buf = 0;
  94         return flags;
  95 }
  96
  97 #define EXTSTR       "[%lu -> %lu/%lu]"
  98 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
  99 static const char *const oes_strings[] = {
 100         "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
 101
 102 #define OSC_EXTENT_DUMP_WITH_LOC(file, func, line, mask, extent, fmt, ...) do {\
 103         static struct cfs_debug_limit_state cdls;                             \
 104         struct osc_extent *__ext = (extent);                                  \
 105         char __buf[16];                                                       \
 106                                                                               \
 107         __CDEBUG_WITH_LOC(file, func, line, mask, &cdls,                      \
 108                 "extent %p@{" EXTSTR ", "                                     \
 109                 "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt,          \
 110                 /* ----- extent part 0 ----- */                               \
 111                 __ext, EXTPARA(__ext),                                        \
 112                 /* ----- part 1 ----- */                                      \
 113                 kref_read(&__ext->oe_refc),                                   \
 114                 atomic_read(&__ext->oe_users),                                \
 115                 list_empty_marker(&__ext->oe_link),                           \
 116                 oes_strings[__ext->oe_state], ext_flags(__ext, __buf),        \
 117                 __ext->oe_obj,                                                \
 118                 /* ----- part 2 ----- */                                      \
 119                 __ext->oe_grants, __ext->oe_nr_pages,                         \
 120                 list_empty_marker(&__ext->oe_pages),                          \
 121                 waitqueue_active(&__ext->oe_waitq) ? '+' : '-',               \
 122                 __ext->oe_dlmlock, __ext->oe_mppr, __ext->oe_owner,           \
 123                 /* ----- part 4 ----- */                                      \
 124                 ## __VA_ARGS__);                                              \
 125         if (mask == D_ERROR && __ext->oe_dlmlock != NULL)                     \
 126                 LDLM_ERROR(__ext->oe_dlmlock, "extent: %p", __ext);           \
 127         else                                                                  \
 128                 LDLM_DEBUG(__ext->oe_dlmlock, "extent: %p", __ext);           \
 129 } while (0)
 130
 131 #define OSC_EXTENT_DUMP(mask, ext, fmt, ...)                            \
 132         OSC_EXTENT_DUMP_WITH_LOC(__FILE__, __func__, __LINE__,          \
 133                                  mask, ext, fmt, ## __VA_ARGS__)
 134
 135 #undef EASSERTF
 136 #define EASSERTF(expr, ext, fmt, args...) do {                          \
 137         if (!(expr)) {                                                  \
 138                 OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args);           \
 139                 osc_extent_tree_dump(D_ERROR, (ext)->oe_obj);           \
 140                 LASSERT(expr);                                          \
 141         }                                                               \
 142 } while (0)
 143
 144 #undef EASSERT
 145 #define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
 146
 147 static inline struct osc_extent *rb_extent(struct rb_node *n)
 148 {
 149         return rb_entry_safe(n, struct osc_extent, oe_node);
 150 }
 151
 152 static inline struct osc_extent *next_extent(struct osc_extent *ext)
 153 {
 154         if (ext == NULL)
 155                 return NULL;
 156
 157         LASSERT(!RB_EMPTY_NODE(&ext->oe_node));
 158         return rb_extent(rb_next(&ext->oe_node));
 159 }
 160
 161 static inline struct osc_extent *prev_extent(struct osc_extent *ext)
 162 {
 163         if (ext == NULL)
 164                 return NULL;
 165
 166         LASSERT(!RB_EMPTY_NODE(&ext->oe_node));
 167         return rb_extent(rb_prev(&ext->oe_node));
 168 }
 169
 170 static inline struct osc_extent *first_extent(struct osc_object *obj)
 171 {
 172         return rb_extent(rb_first(&obj->oo_root));
 173 }
 174
 175 /* object must be locked by caller. */
 176 static int osc_extent_sanity_check0(struct osc_extent *ext,
 177                                     const char *func, const int line)
 178 {
 179         struct osc_object *obj = ext->oe_obj;
 180         struct osc_async_page *oap;
 181         size_t page_count;
 182         int rc = 0;
 183
 184         assert_osc_object_is_locked(obj);
 185
 186         if (ext->oe_state >= OES_STATE_MAX)
 187                 GOTO(out, rc = 10);
 188
 189         if (kref_read(&ext->oe_refc) <= 0)
 190                 GOTO(out, rc = 20);
 191
 192         if (kref_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
 193                 GOTO(out, rc = 30);
 194
 195         switch (ext->oe_state) {
 196         case OES_INV:
 197                 if (ext->oe_nr_pages > 0 || !list_empty(&ext->oe_pages))
 198                         GOTO(out, rc = 35);
 199                 GOTO(out, rc = 0);
 200                 break;
 201         case OES_ACTIVE:
 202                 if (atomic_read(&ext->oe_users) == 0)
 203                         GOTO(out, rc = 40);
 204                 if (ext->oe_hp)
 205                         GOTO(out, rc = 50);
 206                 if (ext->oe_fsync_wait && !ext->oe_urgent)
 207                         GOTO(out, rc = 55);
 208                 break;
 209         case OES_CACHE:
 210                 if (ext->oe_grants == 0)
 211                         GOTO(out, rc = 60);
 212                 if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
 213                         GOTO(out, rc = 65);
 214                 fallthrough;
 215         default:
 216                 if (atomic_read(&ext->oe_users) > 0)
 217                         GOTO(out, rc = 70);
 218         }
 219
 220         if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
 221                 GOTO(out, rc = 80);
 222
 223         if (ext->oe_sync && ext->oe_grants > 0)
 224                 GOTO(out, rc = 90);
 225
 226         if (ext->oe_dlmlock != NULL &&
 227             ext->oe_dlmlock->l_resource->lr_type == LDLM_EXTENT &&
 228             !ldlm_is_failed(ext->oe_dlmlock)) {
 229                 struct ldlm_extent *extent;
 230
 231                 extent = &ext->oe_dlmlock->l_policy_data.l_extent;
 232                 if (!(extent->start <= cl_offset(osc2cl(obj), ext->oe_start) &&
 233                       extent->end   >= cl_offset(osc2cl(obj), ext->oe_max_end)))
 234                         GOTO(out, rc = 100);
 235
 236                 if (!(ext->oe_dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP)))
 237                         GOTO(out, rc = 102);
 238         }
 239
 240         if (ext->oe_nr_pages > ext->oe_mppr)
 241                 GOTO(out, rc = 105);
 242
 243         /* Do not verify page list if extent is in RPC. This is because an
 244          * in-RPC extent is supposed to be exclusively accessible w/o lock. */
 245         if (ext->oe_state > OES_CACHE)
 246                 GOTO(out, rc = 0);
 247
 248         if (!extent_debug)
 249                 GOTO(out, rc = 0);
 250
 251         page_count = 0;
 252         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
 253                 pgoff_t index = osc_index(oap2osc(oap));
 254                 ++page_count;
 255                 if (index > ext->oe_end || index < ext->oe_start)
 256                         GOTO(out, rc = 110);
 257         }
 258         if (page_count != ext->oe_nr_pages)
 259                 GOTO(out, rc = 120);
 260
 261 out:
 262         if (rc != 0)
 263                 OSC_EXTENT_DUMP_WITH_LOC(__FILE__, func, line, D_ERROR, ext,
 264                                          "sanity check %p failed: rc = %d\n",
 265                                          ext, rc);
 266         return rc;
 267 }
 268
 269 #define sanity_check_nolock(ext) \
 270         osc_extent_sanity_check0(ext, __func__, __LINE__)
 271
 272 #define sanity_check(ext) ({                                                   \
 273         int __res;                                                             \
 274         osc_object_lock((ext)->oe_obj);                                        \
 275         __res = sanity_check_nolock(ext);                                      \
 276         osc_object_unlock((ext)->oe_obj);                                      \
 277         __res;                                                                 \
 278 })
 279
 280 static inline bool
 281 overlapped(const struct osc_extent *ex1, const struct osc_extent *ex2)
 282 {
 283         return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
 284 }
 285
 286 /**
 287  * sanity check - to make sure there is no overlapped extent in the tree.
 288  */
 289 static int osc_extent_is_overlapped(struct osc_object *obj,
 290                                     struct osc_extent *ext)
 291 {
 292         struct osc_extent *tmp;
 293
 294         assert_osc_object_is_locked(obj);
 295
 296         if (!extent_debug)
 297                 return 0;
 298
 299         for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
 300                 if (tmp == ext)
 301                         continue;
 302                 if (overlapped(tmp, ext))
 303                         return 1;
 304         }
 305         return 0;
 306 }
 307
 308 static void osc_extent_state_set(struct osc_extent *ext, int state)
 309 {
 310         assert_osc_object_is_locked(ext->oe_obj);
 311         LASSERT(state >= OES_INV && state < OES_STATE_MAX);
 312
 313         /* Never try to sanity check a state changing extent :-) */
 314         /* LASSERT(sanity_check_nolock(ext) == 0); */
 315
 316         /* TODO: validate the state machine */
 317         smp_store_release(&ext->oe_state, state);
 318         wake_up(&ext->oe_waitq);
 319 }
 320
 321 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
 322 {
 323         struct osc_extent *ext;
 324
 325         OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_NOFS);
 326         if (ext == NULL)
 327                 return NULL;
 328
 329         RB_CLEAR_NODE(&ext->oe_node);
 330         ext->oe_obj = obj;
 331         cl_object_get(osc2cl(obj));
 332         kref_init(&ext->oe_refc);
 333         atomic_set(&ext->oe_users, 0);
 334         INIT_LIST_HEAD(&ext->oe_link);
 335         ext->oe_state = OES_INV;
 336         INIT_LIST_HEAD(&ext->oe_pages);
 337         init_waitqueue_head(&ext->oe_waitq);
 338         ext->oe_dlmlock = NULL;
 339
 340         return ext;
 341 }
 342
 343 static void osc_extent_free(struct kref *kref)
 344 {
 345         struct osc_extent *ext = container_of(kref, struct osc_extent,
 346                                               oe_refc);
 347
 348         LASSERT(list_empty(&ext->oe_link));
 349         LASSERT(atomic_read(&ext->oe_users) == 0);
 350         LASSERT(ext->oe_state == OES_INV);
 351         LASSERT(RB_EMPTY_NODE(&ext->oe_node));
 352
 353         if (ext->oe_dlmlock) {
 354                 lu_ref_del(&ext->oe_dlmlock->l_reference,
 355                            "osc_extent", ext);
 356                 LDLM_LOCK_PUT(ext->oe_dlmlock);
 357                 ext->oe_dlmlock = NULL;
 358         }
 359 #if 0
 360         /* If/When cl_object_put drops the need for 'env',
 361          * this code can be enabled, and matching code in
 362          * osc_extent_put removed.
 363          */
 364         cl_object_put(osc2cl(ext->oe_obj));
 365
 366         OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
 367 #endif
 368 }
 369
 370 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
 371 {
 372         LASSERT(kref_read(&ext->oe_refc) >= 0);
 373         kref_get(&ext->oe_refc);
 374         return ext;
 375 }
 376
 377 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
 378 {
 379         LASSERT(kref_read(&ext->oe_refc) > 0);
 380         if (kref_put(&ext->oe_refc, osc_extent_free)) {
 381                 /* This should be in osc_extent_free(), but
 382                  * while we need to pass 'env' it cannot be.
 383                  */
 384                 cl_object_put(env, osc2cl(ext->oe_obj));
 385
 386                 OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
 387         }
 388 }
 389
 390 /**
 391  * osc_extent_put_trust() is a special version of osc_extent_put() when
 392  * it's known that the caller is not the last user. This is to address the
 393  * problem of lacking of lu_env ;-).
 394  */
 395 static void osc_extent_put_trust(struct osc_extent *ext)
 396 {
 397         LASSERT(kref_read(&ext->oe_refc) > 1);
 398         assert_osc_object_is_locked(ext->oe_obj);
 399         osc_extent_put(NULL, ext);
 400 }
 401
 402 /**
 403  * Return the extent which includes pgoff @index, or return the greatest
 404  * previous extent in the tree.
 405  */
 406 static struct osc_extent *osc_extent_search(struct osc_object *obj,
 407                                             pgoff_t index)
 408 {
 409         struct rb_node    *n = obj->oo_root.rb_node;
 410         struct osc_extent *tmp, *p = NULL;
 411
 412         assert_osc_object_is_locked(obj);
 413         while (n != NULL) {
 414                 tmp = rb_extent(n);
 415                 if (index < tmp->oe_start) {
 416                         n = n->rb_left;
 417                 } else if (index > tmp->oe_end) {
 418                         p = rb_extent(n);
 419                         n = n->rb_right;
 420                 } else {
 421                         return tmp;
 422                 }
 423         }
 424         return p;
 425 }
 426
 427 /*
 428  * Return the extent covering @index, otherwise return NULL.
 429  * caller must have held object lock.
 430  */
 431 static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
 432                                             pgoff_t index)
 433 {
 434         struct osc_extent *ext;
 435
 436         ext = osc_extent_search(obj, index);
 437         if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
 438                 return osc_extent_get(ext);
 439         return NULL;
 440 }
 441
 442 /* caller must have held object lock. */
 443 static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
 444 {
 445         struct rb_node   **n      = &obj->oo_root.rb_node;
 446         struct rb_node    *parent = NULL;
 447         struct osc_extent *tmp;
 448
 449         LASSERT(RB_EMPTY_NODE(&ext->oe_node));
 450         LASSERT(ext->oe_obj == obj);
 451         assert_osc_object_is_locked(obj);
 452         while (*n != NULL) {
 453                 tmp = rb_extent(*n);
 454                 parent = *n;
 455
 456                 if (ext->oe_end < tmp->oe_start)
 457                         n = &(*n)->rb_left;
 458                 else if (ext->oe_start > tmp->oe_end)
 459                         n = &(*n)->rb_right;
 460                 else
 461                         EASSERTF(0, tmp, EXTSTR"\n", EXTPARA(ext));
 462         }
 463         rb_link_node(&ext->oe_node, parent, n);
 464         rb_insert_color(&ext->oe_node, &obj->oo_root);
 465         osc_extent_get(ext);
 466 }
 467
 468 /* caller must have held object lock. */
 469 static void osc_extent_erase(struct osc_extent *ext)
 470 {
 471         struct osc_object *obj = ext->oe_obj;
 472         assert_osc_object_is_locked(obj);
 473         if (!RB_EMPTY_NODE(&ext->oe_node)) {
 474                 rb_erase(&ext->oe_node, &obj->oo_root);
 475                 RB_CLEAR_NODE(&ext->oe_node);
 476                 /* rbtree held a refcount */
 477                 osc_extent_put_trust(ext);
 478         }
 479 }
 480
 481 static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
 482 {
 483         struct osc_object *obj = ext->oe_obj;
 484
 485         assert_osc_object_is_locked(obj);
 486         LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
 487         if (ext->oe_state == OES_CACHE) {
 488                 osc_extent_state_set(ext, OES_ACTIVE);
 489                 osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
 490         }
 491         atomic_inc(&ext->oe_users);
 492         list_del_init(&ext->oe_link);
 493         return osc_extent_get(ext);
 494 }
 495
 496 static void __osc_extent_remove(struct osc_extent *ext)
 497 {
 498         assert_osc_object_is_locked(ext->oe_obj);
 499         LASSERT(list_empty(&ext->oe_pages));
 500         osc_extent_erase(ext);
 501         list_del_init(&ext->oe_link);
 502         osc_extent_state_set(ext, OES_INV);
 503         OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
 504 }
 505
 506 static void osc_extent_remove(struct osc_extent *ext)
 507 {
 508         struct osc_object *obj = ext->oe_obj;
 509
 510         osc_object_lock(obj);
 511         __osc_extent_remove(ext);
 512         osc_object_unlock(obj);
 513 }
 514
 515 /**
 516  * This function is used to merge extents to get better performance. It checks
 517  * if @cur and @victim are contiguous at block level.
 518  */
 519 static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
 520                             struct osc_extent *victim)
 521 {
 522         struct osc_object       *obj = cur->oe_obj;
 523         struct client_obd       *cli = osc_cli(obj);
 524         pgoff_t                  chunk_start;
 525         pgoff_t                  chunk_end;
 526         int                      ppc_bits;
 527
 528         LASSERT(cur->oe_state == OES_CACHE);
 529         assert_osc_object_is_locked(obj);
 530         if (victim == NULL)
 531                 return -EINVAL;
 532
 533         if (victim->oe_state != OES_INV &&
 534             (victim->oe_state != OES_CACHE || victim->oe_fsync_wait))
 535                 return -EBUSY;
 536
 537         if (cur->oe_max_end != victim->oe_max_end)
 538                 return -ERANGE;
 539
 540         /*
 541          * In the rare case max_pages_per_rpc (mppr) is changed, don't
 542          * merge extents until after old ones have been sent, or the
 543          * "extents are aligned to RPCs" checks are unhappy.
 544          */
 545         if (cur->oe_mppr != victim->oe_mppr)
 546                 return -ERANGE;
 547
 548         LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
 549         ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_SHIFT;
 550         chunk_start = cur->oe_start >> ppc_bits;
 551         chunk_end   = cur->oe_end   >> ppc_bits;
 552         if (chunk_start   != (victim->oe_end >> ppc_bits) + 1 &&
 553             chunk_end + 1 != victim->oe_start >> ppc_bits)
 554                 return -ERANGE;
 555
 556         /* overall extent size should not exceed the max supported limit
 557          * reported by the server */
 558         if (cur->oe_end - cur->oe_start + 1 +
 559             victim->oe_end - victim->oe_start + 1 > cli->cl_max_extent_pages)
 560                 return -ERANGE;
 561
 562         OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
 563
 564         cur->oe_start     = min(cur->oe_start, victim->oe_start);
 565         cur->oe_end       = max(cur->oe_end,   victim->oe_end);
 566         /* per-extent tax should be accounted only once for the whole extent */
 567         cur->oe_grants   += victim->oe_grants - cli->cl_grant_extent_tax;
 568         cur->oe_nr_pages += victim->oe_nr_pages;
 569         /* only the following bits are needed to merge */
 570         cur->oe_urgent   |= victim->oe_urgent;
 571         cur->oe_memalloc |= victim->oe_memalloc;
 572         list_splice_init(&victim->oe_pages, &cur->oe_pages);
 573         victim->oe_nr_pages = 0;
 574
 575         osc_extent_get(victim);
 576         __osc_extent_remove(victim);
 577         osc_extent_put(env, victim);
 578
 579         OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
 580         return 0;
 581 }
 582
 583 /**
 584  * Drop user count of osc_extent, and unplug IO asynchronously.
 585  */
 586 void osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
 587 {
 588         struct osc_object *obj = ext->oe_obj;
 589         struct client_obd *cli = osc_cli(obj);
 590         ENTRY;
 591
 592         LASSERT(atomic_read(&ext->oe_users) > 0);
 593         LASSERT(sanity_check(ext) == 0);
 594         LASSERT(ext->oe_grants > 0);
 595
 596         if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
 597                 LASSERT(ext->oe_state == OES_ACTIVE);
 598                 if (ext->oe_trunc_pending) {
 599                         /* a truncate process is waiting for this extent.
 600                          * This may happen due to a race, check
 601                          * osc_cache_truncate_start(). */
 602                         osc_extent_state_set(ext, OES_TRUNC);
 603                         ext->oe_trunc_pending = 0;
 604                         osc_object_unlock(obj);
 605                 } else {
 606                         int grant = 0;
 607
 608                         osc_extent_state_set(ext, OES_CACHE);
 609                         osc_update_pending(obj, OBD_BRW_WRITE,
 610                                            ext->oe_nr_pages);
 611
 612                         /* try to merge the previous and next extent. */
 613                         if (osc_extent_merge(env, ext, prev_extent(ext)) == 0)
 614                                 grant += cli->cl_grant_extent_tax;
 615                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 616                                 grant += cli->cl_grant_extent_tax;
 617
 618                         if (ext->oe_hp)
 619                                 list_move_tail(&ext->oe_link,
 620                                                &obj->oo_hp_exts);
 621                         else if (ext->oe_urgent)
 622                                 list_move_tail(&ext->oe_link,
 623                                                &obj->oo_urgent_exts);
 624                         else if (ext->oe_nr_pages == ext->oe_mppr) {
 625                                 list_move_tail(&ext->oe_link,
 626                                                &obj->oo_full_exts);
 627                         }
 628                         osc_object_unlock(obj);
 629                         if (grant > 0)
 630                                 osc_unreserve_grant(cli, 0, grant);
 631                 }
 632
 633                 osc_io_unplug_async(env, cli, obj);
 634         }
 635         osc_extent_put(env, ext);
 636
 637         RETURN_EXIT;
 638 }
 639
 640 /**
 641  * Find or create an extent which includes @index, core function to manage
 642  * extent tree.
 643  */
 644 static struct osc_extent *osc_extent_find(const struct lu_env *env,
 645                                           struct osc_object *obj, pgoff_t index,
 646                                           unsigned int *grants)
 647 {
 648         struct client_obd *cli = osc_cli(obj);
 649         struct osc_lock   *olck;
 650         struct cl_lock_descr *descr;
 651         struct osc_extent *cur;
 652         struct osc_extent *ext;
 653         struct osc_extent *conflict = NULL;
 654         struct osc_extent *found = NULL;
 655         pgoff_t    chunk;
 656         pgoff_t    max_end;
 657         unsigned int max_pages; /* max_pages_per_rpc */
 658         unsigned int chunksize;
 659         int        ppc_bits; /* pages per chunk bits */
 660         pgoff_t    chunk_mask;
 661         int        rc;
 662         ENTRY;
 663
 664         cur = osc_extent_alloc(obj);
 665         if (cur == NULL)
 666                 RETURN(ERR_PTR(-ENOMEM));
 667
 668         olck = osc_env_io(env)->oi_write_osclock;
 669         LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
 670         LASSERT(olck->ols_state == OLS_GRANTED);
 671
 672         descr = &olck->ols_cl.cls_lock->cll_descr;
 673         LASSERT(descr->cld_mode >= CLM_WRITE);
 674
 675         LASSERTF(cli->cl_chunkbits >= PAGE_SHIFT,
 676                  "chunkbits: %u\n", cli->cl_chunkbits);
 677         ppc_bits   = cli->cl_chunkbits - PAGE_SHIFT;
 678         chunk_mask = ~((1 << ppc_bits) - 1);
 679         chunksize  = 1 << cli->cl_chunkbits;
 680         chunk      = index >> ppc_bits;
 681
 682         /* align end to RPC edge. */
 683         max_pages = cli->cl_max_pages_per_rpc;
 684         if ((max_pages & ~chunk_mask) != 0) {
 685                 CERROR("max_pages: %#x chunkbits: %u chunk_mask: %#lx\n",
 686                        max_pages, cli->cl_chunkbits, chunk_mask);
 687                 RETURN(ERR_PTR(-EINVAL));
 688         }
 689         max_end = index - (index % max_pages) + max_pages - 1;
 690         max_end = min_t(pgoff_t, max_end, descr->cld_end);
 691
 692         /* initialize new extent by parameters so far */
 693         cur->oe_max_end = max_end;
 694         cur->oe_start   = index & chunk_mask;
 695         cur->oe_end     = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
 696         if (cur->oe_start < descr->cld_start)
 697                 cur->oe_start = descr->cld_start;
 698         if (cur->oe_end > max_end)
 699                 cur->oe_end = max_end;
 700         cur->oe_grants  = chunksize + cli->cl_grant_extent_tax;
 701         cur->oe_mppr    = max_pages;
 702         if (olck->ols_dlmlock != NULL) {
 703                 LASSERT(olck->ols_hold);
 704                 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
 705                 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
 706         }
 707
 708         /* grants has been allocated by caller */
 709         LASSERTF(*grants >= chunksize + cli->cl_grant_extent_tax,
 710                  "%u/%u/%u.\n", *grants, chunksize, cli->cl_grant_extent_tax);
 711         LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
 712                  EXTPARA(cur));
 713
 714 restart:
 715         osc_object_lock(obj);
 716         ext = osc_extent_search(obj, cur->oe_start);
 717         if (!ext)
 718                 ext = first_extent(obj);
 719         for (; ext; ext = next_extent(ext)) {
 720                 pgoff_t ext_chk_start = ext->oe_start >> ppc_bits;
 721                 pgoff_t ext_chk_end = ext->oe_end >> ppc_bits;
 722
 723                 LASSERT(sanity_check_nolock(ext) == 0);
 724                 if (chunk > ext_chk_end + 1 || chunk < ext_chk_start)
 725                         break;
 726
 727                 /* if covering by different locks, no chance to match */
 728                 if (olck->ols_dlmlock != ext->oe_dlmlock) {
 729                         EASSERTF(!overlapped(ext, cur), ext,
 730                                  EXTSTR"\n", EXTPARA(cur));
 731
 732                         continue;
 733                 }
 734
 735                 /* discontiguous chunks? */
 736                 if (chunk + 1 < ext_chk_start)
 737                         continue;
 738
 739                 /* ok, from now on, ext and cur have these attrs:
 740                  * 1. covered by the same lock
 741                  * 2. contiguous at chunk level or overlapping. */
 742
 743                 if (overlapped(ext, cur)) {
 744                         /* cur is the minimum unit, so overlapping means
 745                          * full contain. */
 746                         EASSERTF((ext->oe_start <= cur->oe_start &&
 747                                   ext->oe_end >= cur->oe_end),
 748                                  ext, EXTSTR"\n", EXTPARA(cur));
 749
 750                         if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
 751                                 /* for simplicity, we wait for this extent to
 752                                  * finish before going forward. */
 753                                 conflict = osc_extent_get(ext);
 754                                 break;
 755                         }
 756
 757                         found = osc_extent_hold(ext);
 758                         break;
 759                 }
 760
 761                 /* non-overlapped extent */
 762                 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait)
 763                         /* we can't do anything for a non OES_CACHE extent, or
 764                          * if there is someone waiting for this extent to be
 765                          * flushed, try next one. */
 766                         continue;
 767
 768                 if (osc_extent_merge(env, ext, cur) == 0) {
 769                         LASSERT(*grants >= chunksize);
 770                         *grants -= chunksize;
 771
 772                         /*
 773                          * Try to merge with the next one too because we
 774                          * might have just filled in a gap.
 775                          */
 776                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 777                                 /* we can save extent tax from next extent */
 778                                 *grants += cli->cl_grant_extent_tax;
 779
 780                         found = osc_extent_hold(ext);
 781                         break;
 782                 }
 783         }
 784
 785         osc_extent_tree_dump(D_CACHE, obj);
 786         if (found != NULL) {
 787                 LASSERT(conflict == NULL);
 788                 if (!IS_ERR(found)) {
 789                         LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
 790                         OSC_EXTENT_DUMP(D_CACHE, found,
 791                                         "found caching ext for %lu.\n", index);
 792                 }
 793         } else if (conflict == NULL) {
 794                 /* create a new extent */
 795                 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
 796                 LASSERT(*grants >= cur->oe_grants);
 797                 *grants -= cur->oe_grants;
 798
 799                 cur->oe_state = OES_CACHE;
 800                 found = osc_extent_hold(cur);
 801                 osc_extent_insert(obj, cur);
 802                 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
 803                                 index, descr->cld_end);
 804         }
 805         osc_object_unlock(obj);
 806
 807         if (conflict != NULL) {
 808                 LASSERT(found == NULL);
 809
 810                 /* waiting for IO to finish. Please notice that it's impossible
 811                  * to be an OES_TRUNC extent. */
 812                 rc = osc_extent_wait(env, conflict, OES_INV);
 813                 osc_extent_put(env, conflict);
 814                 conflict = NULL;
 815                 if (rc < 0)
 816                         GOTO(out, found = ERR_PTR(rc));
 817
 818                 goto restart;
 819         }
 820         EXIT;
 821
 822 out:
 823         osc_extent_put(env, cur);
 824         return found;
 825 }
 826
 827 /**
 828  * Called when IO is finished to an extent.
 829  */
 830 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
 831                       int sent, int rc)
 832 {
 833         struct client_obd *cli = osc_cli(ext->oe_obj);
 834         struct osc_async_page *oap;
 835         struct osc_async_page *tmp;
 836         int nr_pages = ext->oe_nr_pages;
 837         int lost_grant = 0;
 838         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
 839         loff_t last_off = 0;
 840         int last_count = -1;
 841         ENTRY;
 842
 843         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
 844
 845         ext->oe_rc = rc ?: ext->oe_nr_pages;
 846         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
 847
 848         osc_lru_add_batch(cli, &ext->oe_pages);
 849         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 850                                      oap_pending_item) {
 851                 list_del_init(&oap->oap_rpc_item);
 852                 list_del_init(&oap->oap_pending_item);
 853                 if (last_off <= oap->oap_obj_off) {
 854                         last_off = oap->oap_obj_off;
 855                         last_count = oap->oap_count;
 856                 }
 857
 858                 --ext->oe_nr_pages;
 859                 osc_ap_completion(env, cli, oap, sent, rc);
 860         }
 861         EASSERT(ext->oe_nr_pages == 0, ext);
 862
 863         if (!sent) {
 864                 lost_grant = ext->oe_grants;
 865         } else if (cli->cl_ocd_grant_param == 0 &&
 866                    blocksize < PAGE_SIZE &&
 867                    last_count != PAGE_SIZE) {
 868                 /* For short writes without OBD_CONNECT_GRANT support, we
 869                  * shouldn't count parts of pages that span a whole chunk on
 870                  * the OST side, or our accounting goes wrong. Should match
 871                  * the code in tgt_grant_check.
 872                  */
 873                 int offset = last_off & ~PAGE_MASK;
 874                 int count = last_count + (offset & (blocksize - 1));
 875                 int end = (offset + last_count) & (blocksize - 1);
 876                 if (end)
 877                         count += blocksize - end;
 878
 879                 lost_grant = PAGE_SIZE - count;
 880         }
 881         if (ext->oe_grants > 0)
 882                 osc_free_grant(cli, nr_pages, lost_grant, ext->oe_grants);
 883
 884         osc_extent_remove(ext);
 885         /* put the refcount for RPC */
 886         osc_extent_put(env, ext);
 887         RETURN(0);
 888 }
 889
 890 /**
 891  * Wait for the extent's state to become @state.
 892  */
 893 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
 894                            enum osc_extent_state state)
 895 {
 896         struct osc_object *obj = ext->oe_obj;
 897         int rc = 0;
 898         ENTRY;
 899
 900         osc_object_lock(obj);
 901         LASSERT(sanity_check_nolock(ext) == 0);
 902         /* `Kick' this extent only if the caller is waiting for it to be
 903          * written out. */
 904         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
 905                 if (ext->oe_state == OES_ACTIVE) {
 906                         ext->oe_urgent = 1;
 907                 } else if (ext->oe_state == OES_CACHE) {
 908                         ext->oe_urgent = 1;
 909                         osc_extent_hold(ext);
 910                         rc = 1;
 911                 }
 912         }
 913         osc_object_unlock(obj);
 914         if (rc == 1)
 915                 osc_extent_release(env, ext);
 916
 917         /* wait for the extent until its state becomes @state */
 918         rc = wait_event_idle_timeout(ext->oe_waitq,
 919                                      smp_load_acquire(&ext->oe_state) == state,
 920                                      cfs_time_seconds(600));
 921         if (rc == 0) {
 922                 OSC_EXTENT_DUMP(D_ERROR, ext,
 923                         "%s: wait ext to %u timedout, recovery in progress?\n",
 924                         cli_name(osc_cli(obj)), state);
 925
 926                 wait_event_idle(ext->oe_waitq,
 927                                 smp_load_acquire(&ext->oe_state) == state);
 928         }
 929         if (ext->oe_rc < 0)
 930                 rc = ext->oe_rc;
 931         else
 932                 rc = 0;
 933         RETURN(rc);
 934 }
 935
 936 /**
 937  * Discard pages with index greater than @size. If @ext is overlapped with
 938  * @size, then partial truncate happens.
 939  */
 940 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
 941                                 bool partial)
 942 {
 943         struct lu_env         *env;
 944         struct cl_io          *io;
 945         struct osc_object     *obj = ext->oe_obj;
 946         struct client_obd     *cli = osc_cli(obj);
 947         struct osc_async_page *oap;
 948         struct osc_async_page *tmp;
 949         struct pagevec        *pvec;
 950         int                    pages_in_chunk = 0;
 951         int                    ppc_bits    = cli->cl_chunkbits -
 952                                              PAGE_SHIFT;
 953         __u64                  trunc_chunk = trunc_index >> ppc_bits;
 954         int                    grants   = 0;
 955         int                    nr_pages = 0;
 956         int                    rc       = 0;
 957         __u16                  refcheck;
 958         ENTRY;
 959
 960         LASSERT(sanity_check(ext) == 0);
 961         LASSERT(ext->oe_state == OES_TRUNC);
 962         LASSERT(!ext->oe_urgent);
 963
 964         /* Request new lu_env.
 965          * We can't use that env from osc_cache_truncate_start() because
 966          * it's from lov_io_sub and not fully initialized. */
 967         env = cl_env_get(&refcheck);
 968         if (IS_ERR(env))
 969                 RETURN(PTR_ERR(env));
 970
 971         io  = osc_env_thread_io(env);
 972         io->ci_obj = cl_object_top(osc2cl(obj));
 973         io->ci_ignore_layout = 1;
 974         pvec = &osc_env_info(env)->oti_pagevec;
 975         ll_pagevec_init(pvec, 0);
 976         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
 977         if (rc < 0)
 978                 GOTO(out, rc);
 979
 980         /* discard all pages with index greater than trunc_index */
 981         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 982                                      oap_pending_item) {
 983                 pgoff_t index = osc_index(oap2osc(oap));
 984                 struct cl_page  *page = oap2cl_page(oap);
 985
 986                 LASSERT(list_empty(&oap->oap_rpc_item));
 987
 988                 /* only discard the pages with their index greater than
 989                  * trunc_index, and ... */
 990                 if (index < trunc_index ||
 991                     (index == trunc_index && partial)) {
 992                         /* accounting how many pages remaining in the chunk
 993                          * so that we can calculate grants correctly. */
 994                         if (index >> ppc_bits == trunc_chunk)
 995                                 ++pages_in_chunk;
 996                         continue;
 997                 }
 998
 999                 list_del_init(&oap->oap_pending_item);
1000
1001                 cl_page_get(page);
1002                 lu_ref_add(&page->cp_reference, "truncate", current);
1003
1004                 if (cl_page_own(env, io, page) == 0) {
1005                         cl_page_discard(env, io, page);
1006                         cl_page_disown(env, io, page);
1007                 } else {
1008                         LASSERT(page->cp_state == CPS_FREEING);
1009                         LASSERT(0);
1010                 }
1011
1012                 lu_ref_del(&page->cp_reference, "truncate", current);
1013                 cl_pagevec_put(env, page, pvec);
1014
1015                 --ext->oe_nr_pages;
1016                 ++nr_pages;
1017         }
1018         pagevec_release(pvec);
1019
1020         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1021                       ext->oe_nr_pages == 0),
1022                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1023
1024         osc_object_lock(obj);
1025         if (ext->oe_nr_pages == 0) {
1026                 LASSERT(pages_in_chunk == 0);
1027                 grants = ext->oe_grants;
1028                 ext->oe_grants = 0;
1029         } else { /* calculate how many grants we can free */
1030                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1031                 pgoff_t last_index;
1032
1033
1034                 /* if there is no pages in this chunk, we can also free grants
1035                  * for the last chunk */
1036                 if (pages_in_chunk == 0) {
1037                         /* if this is the 1st chunk and no pages in this chunk,
1038                          * ext->oe_nr_pages must be zero, so we should be in
1039                          * the other if-clause. */
1040                         LASSERT(trunc_chunk > 0);
1041                         --trunc_chunk;
1042                         ++chunks;
1043                 }
1044
1045                 /* this is what we can free from this extent */
1046                 grants          = chunks << cli->cl_chunkbits;
1047                 ext->oe_grants -= grants;
1048                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1049                 ext->oe_end     = min(last_index, ext->oe_max_end);
1050                 LASSERT(ext->oe_end >= ext->oe_start);
1051                 LASSERT(ext->oe_grants > 0);
1052         }
1053         osc_object_unlock(obj);
1054
1055         if (grants > 0 || nr_pages > 0)
1056                 osc_free_grant(cli, nr_pages, grants, grants);
1057
1058 out:
1059         cl_io_fini(env, io);
1060         cl_env_put(env, &refcheck);
1061         RETURN(rc);
1062 }
1063
1064 /**
1065  * This function is used to make the extent prepared for transfer.
1066  * A race with flusing page - ll_writepage() has to be handled cautiously.
1067  */
1068 static int osc_extent_make_ready(const struct lu_env *env,
1069                                  struct osc_extent *ext)
1070 {
1071         struct osc_async_page *oap;
1072         struct osc_async_page *last = NULL;
1073         struct osc_object *obj = ext->oe_obj;
1074         unsigned int page_count = 0;
1075         int rc;
1076         ENTRY;
1077
1078         /* we're going to grab page lock, so object lock must not be taken. */
1079         LASSERT(sanity_check(ext) == 0);
1080         /* in locking state, any process should not touch this extent. */
1081         EASSERT(ext->oe_state == OES_LOCKING, ext);
1082         EASSERT(ext->oe_owner != NULL, ext);
1083
1084         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1085
1086         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1087                 ++page_count;
1088                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1089                         last = oap;
1090
1091                 /* checking ASYNC_READY is race safe */
1092                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1093                         continue;
1094
1095                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1096                 switch (rc) {
1097                 case 0:
1098                         spin_lock(&oap->oap_lock);
1099                         oap->oap_async_flags |= ASYNC_READY;
1100                         spin_unlock(&oap->oap_lock);
1101                         break;
1102                 case -EALREADY:
1103                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1104                         break;
1105                 default:
1106                         LASSERTF(0, "unknown return code: %d\n", rc);
1107                 }
1108         }
1109
1110         LASSERT(page_count == ext->oe_nr_pages);
1111         LASSERT(last != NULL);
1112         /* the last page is the only one we need to refresh its count by
1113          * the size of file. */
1114         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1115                 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1116                 LASSERTF(last_oap_count > 0,
1117                          "last_oap_count %d\n", last_oap_count);
1118                 LASSERT(last->oap_page_off + last_oap_count <= PAGE_SIZE);
1119                 last->oap_count = last_oap_count;
1120                 spin_lock(&last->oap_lock);
1121                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1122                 spin_unlock(&last->oap_lock);
1123         }
1124
1125         /* for the rest of pages, we don't need to call osf_refresh_count()
1126          * because it's known they are not the last page */
1127         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1128                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1129                         oap->oap_count = PAGE_SIZE - oap->oap_page_off;
1130                         spin_lock(&oap->oap_lock);
1131                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1132                         spin_unlock(&oap->oap_lock);
1133                 }
1134         }
1135
1136         osc_object_lock(obj);
1137         osc_extent_state_set(ext, OES_RPC);
1138         osc_object_unlock(obj);
1139         /* get a refcount for RPC. */
1140         osc_extent_get(ext);
1141
1142         RETURN(0);
1143 }
1144
1145 /**
1146  * Quick and simple version of osc_extent_find(). This function is frequently
1147  * called to expand the extent for the same IO. To expand the extent, the
1148  * page index must be in the same or next chunk of ext->oe_end.
1149  */
1150 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index,
1151                              unsigned int *grants)
1152 {
1153         struct osc_object *obj = ext->oe_obj;
1154         struct client_obd *cli = osc_cli(obj);
1155         struct osc_extent *next;
1156         int ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1157         pgoff_t chunk = index >> ppc_bits;
1158         pgoff_t end_chunk;
1159         pgoff_t end_index;
1160         unsigned int chunksize = 1 << cli->cl_chunkbits;
1161         int rc = 0;
1162         ENTRY;
1163
1164         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1165         osc_object_lock(obj);
1166         LASSERT(sanity_check_nolock(ext) == 0);
1167         end_chunk = ext->oe_end >> ppc_bits;
1168         if (chunk > end_chunk + 1)
1169                 GOTO(out, rc = -ERANGE);
1170
1171         if (end_chunk >= chunk)
1172                 GOTO(out, rc = 0);
1173
1174         LASSERT(end_chunk + 1 == chunk);
1175
1176         /* try to expand this extent to cover @index */
1177         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1178
1179         /* don't go over the maximum extent size reported by server */
1180         if (end_index - ext->oe_start + 1 > cli->cl_max_extent_pages)
1181                 GOTO(out, rc = -ERANGE);
1182
1183         next = next_extent(ext);
1184         if (next != NULL && next->oe_start <= end_index)
1185                 /* complex mode - overlapped with the next extent,
1186                  * this case will be handled by osc_extent_find() */
1187                 GOTO(out, rc = -EAGAIN);
1188
1189         ext->oe_end = end_index;
1190         ext->oe_grants += chunksize;
1191         LASSERT(*grants >= chunksize);
1192         *grants -= chunksize;
1193         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1194                  "overlapped after expanding for %lu.\n", index);
1195         EXIT;
1196
1197 out:
1198         osc_object_unlock(obj);
1199         RETURN(rc);
1200 }
1201
1202 static void osc_extent_tree_dump0(int mask, struct osc_object *obj,
1203                                   const char *func, int line)
1204 {
1205         struct osc_extent *ext;
1206         int cnt;
1207
1208         if (!cfs_cdebug_show(mask, DEBUG_SUBSYSTEM))
1209                 return;
1210
1211         CDEBUG(mask, "Dump object %p extents at %s:%d, mppr: %u.\n",
1212                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1213
1214         /* osc_object_lock(obj); */
1215         cnt = 1;
1216         for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1217                 OSC_EXTENT_DUMP(mask, ext, "in tree %d.\n", cnt++);
1218
1219         cnt = 1;
1220         list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1221                 OSC_EXTENT_DUMP(mask, ext, "hp %d.\n", cnt++);
1222
1223         cnt = 1;
1224         list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1225                 OSC_EXTENT_DUMP(mask, ext, "urgent %d.\n", cnt++);
1226
1227         cnt = 1;
1228         list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1229                 OSC_EXTENT_DUMP(mask, ext, "reading %d.\n", cnt++);
1230         /* osc_object_unlock(obj); */
1231 }
1232
1233 /* ------------------ osc extent end ------------------ */
1234
1235 static inline int osc_is_ready(struct osc_object *osc)
1236 {
1237         return !list_empty(&osc->oo_ready_item) ||
1238                !list_empty(&osc->oo_hp_ready_item);
1239 }
1240
1241 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1242         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1243                (OSC), osc_is_ready(OSC),                                       \
1244                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1245                list_empty_marker(&(OSC)->oo_ready_item),                       \
1246                atomic_read(&(OSC)->oo_nr_writes),                              \
1247                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1248                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1249                atomic_read(&(OSC)->oo_nr_reads),                               \
1250                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1251                ##args)
1252
1253 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1254                           int cmd)
1255 {
1256         struct osc_page *opg  = oap2osc_page(oap);
1257         struct cl_page  *page = oap2cl_page(oap);
1258         int result;
1259
1260         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1261
1262         ENTRY;
1263         result = cl_page_make_ready(env, page, CRT_WRITE);
1264         if (result == 0)
1265                 opg->ops_submit_time = ktime_get();
1266         RETURN(result);
1267 }
1268
1269 static int osc_refresh_count(const struct lu_env *env,
1270                              struct osc_async_page *oap, int cmd)
1271 {
1272         struct osc_page  *opg = oap2osc_page(oap);
1273         pgoff_t index = osc_index(oap2osc(oap));
1274         struct cl_object *obj = osc2cl(osc_page_object(opg));
1275         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
1276         int result;
1277         loff_t kms;
1278
1279         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1280         LASSERT(!(cmd & OBD_BRW_READ));
1281         LASSERT(opg != NULL);
1282
1283         cl_object_attr_lock(obj);
1284         result = cl_object_attr_get(env, obj, attr);
1285         cl_object_attr_unlock(obj);
1286         if (result < 0)
1287                 return result;
1288         kms = attr->cat_kms;
1289         if (cl_offset(obj, index) >= kms)
1290                 /* catch race with truncate */
1291                 return 0;
1292         else if (cl_offset(obj, index + 1) > kms)
1293                 /* catch sub-page write at end of file */
1294                 return kms & ~PAGE_MASK;
1295         else
1296                 return PAGE_SIZE;
1297 }
1298
1299 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1300                           int cmd, int rc)
1301 {
1302         struct osc_page   *opg  = oap2osc_page(oap);
1303         struct cl_page    *page = oap2cl_page(oap);
1304         enum cl_req_type   crt;
1305         int srvlock;
1306
1307         ENTRY;
1308
1309         cmd &= ~OBD_BRW_NOQUOTA;
1310         LASSERTF(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ),
1311                  "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1312         LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1313                 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1314         LASSERT(opg->ops_transfer_pinned);
1315
1316         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1317         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1318         opg->ops_transfer_pinned = 0;
1319
1320         opg->ops_submit_time = ktime_set(0, 0);
1321         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1322
1323         /* statistic */
1324         if (rc == 0 && srvlock) {
1325                 struct lu_device *ld = osc_page_object(opg)->oo_cl.co_lu.lo_dev;
1326                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1327                 size_t bytes = oap->oap_count;
1328
1329                 if (crt == CRT_READ)
1330                         stats->os_lockless_reads += bytes;
1331                 else
1332                         stats->os_lockless_writes += bytes;
1333         }
1334
1335         /*
1336          * This has to be the last operation with the page, as locks are
1337          * released in cl_page_completion() and nothing except for the
1338          * reference counter protects page from concurrent reclaim.
1339          */
1340         lu_ref_del(&page->cp_reference, "transfer", page);
1341
1342         cl_page_completion(env, page, crt, rc);
1343         cl_page_put(env, page);
1344
1345         RETURN(0);
1346 }
1347
1348 #define OSC_DUMP_GRANT(mask, cli, fmt, args...) do {                    \
1349         struct client_obd *__tmp = (cli);                               \
1350         CDEBUG(mask, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu " \
1351                "dropped: %ld avail: %ld, dirty_grant: %ld, "            \
1352                "reserved: %ld, flight: %d } lru {in list: %ld, "        \
1353                "left: %ld, waiters: %d }" fmt "\n",                     \
1354                cli_name(__tmp),                                         \
1355                __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages,        \
1356                atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1357                __tmp->cl_lost_grant, __tmp->cl_avail_grant,             \
1358                __tmp->cl_dirty_grant,                                   \
1359                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,         \
1360                atomic_long_read(&__tmp->cl_lru_in_list),                \
1361                atomic_long_read(&__tmp->cl_lru_busy),                   \
1362                atomic_read(&__tmp->cl_lru_shrinkers), ##args);          \
1363 } while (0)
1364
1365 /* caller must hold loi_list_lock */
1366 static void osc_consume_write_grant(struct client_obd *cli,
1367                                     struct brw_page *pga)
1368 {
1369         assert_spin_locked(&cli->cl_loi_list_lock);
1370         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1371         cli->cl_dirty_pages++;
1372         pga->flag |= OBD_BRW_FROM_GRANT;
1373         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1374                PAGE_SIZE, pga, pga->pg);
1375 }
1376
1377 /* the companion to osc_consume_write_grant, called when a brw has completed.
1378  * must be called with the loi lock held. */
1379 static void osc_release_write_grant(struct client_obd *cli,
1380                                     struct brw_page *pga)
1381 {
1382         ENTRY;
1383
1384         assert_spin_locked(&cli->cl_loi_list_lock);
1385         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1386                 EXIT;
1387                 return;
1388         }
1389
1390         pga->flag &= ~OBD_BRW_FROM_GRANT;
1391         atomic_long_dec(&obd_dirty_pages);
1392         cli->cl_dirty_pages--;
1393         EXIT;
1394 }
1395
1396 /**
1397  * To avoid sleeping with object lock held, it's good for us allocate enough
1398  * grants before entering into critical section.
1399  *
1400  * client_obd_list_lock held by caller
1401  */
1402 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1403 {
1404         int rc = -EDQUOT;
1405
1406         if (cli->cl_avail_grant >= bytes) {
1407                 cli->cl_avail_grant    -= bytes;
1408                 cli->cl_reserved_grant += bytes;
1409                 rc = 0;
1410         }
1411         return rc;
1412 }
1413
1414 static void __osc_unreserve_grant(struct client_obd *cli,
1415                                   unsigned int reserved, unsigned int unused)
1416 {
1417         /* it's quite normal for us to get more grant than reserved.
1418          * Thinking about a case that two extents merged by adding a new
1419          * chunk, we can save one extent tax. If extent tax is greater than
1420          * one chunk, we can save more grant by adding a new chunk */
1421         cli->cl_reserved_grant -= reserved;
1422         if (unused > reserved) {
1423                 cli->cl_avail_grant += reserved;
1424                 cli->cl_lost_grant  += unused - reserved;
1425                 cli->cl_dirty_grant -= unused - reserved;
1426         } else {
1427                 cli->cl_avail_grant += unused;
1428                 cli->cl_dirty_grant += reserved - unused;
1429         }
1430 }
1431
1432 static void osc_unreserve_grant_nolock(struct client_obd *cli,
1433                                        unsigned int reserved,
1434                                        unsigned int unused)
1435 {
1436         __osc_unreserve_grant(cli, reserved, unused);
1437         if (unused > 0)
1438                 osc_wake_cache_waiters(cli);
1439 }
1440
1441 static void osc_unreserve_grant(struct client_obd *cli,
1442                                 unsigned int reserved, unsigned int unused)
1443 {
1444         spin_lock(&cli->cl_loi_list_lock);
1445         osc_unreserve_grant_nolock(cli, reserved, unused);
1446         spin_unlock(&cli->cl_loi_list_lock);
1447 }
1448
1449 /**
1450  * Free grant after IO is finished or canceled.
1451  *
1452  * @lost_grant is used to remember how many grants we have allocated but not
1453  * used, we should return these grants to OST. There're two cases where grants
1454  * can be lost:
1455  * 1. truncate;
1456  * 2. Without OBD_CONNECT_GRANT support and blocksize at OST is less than
1457  *    PAGE_SIZE and a partial page was written. In this case OST may use less
1458  *    chunks to serve this partial write. OSTs don't actually know the page
1459  *    size on the client side. so clients have to calculate lost grant by the
1460  *    blocksize on the OST. See tgt_grant_check() for details.
1461  */
1462 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1463                            unsigned int lost_grant, unsigned int dirty_grant)
1464 {
1465         unsigned long grant;
1466
1467         grant = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
1468
1469         spin_lock(&cli->cl_loi_list_lock);
1470         atomic_long_sub(nr_pages, &obd_dirty_pages);
1471         cli->cl_dirty_pages -= nr_pages;
1472         cli->cl_lost_grant += lost_grant;
1473         cli->cl_dirty_grant -= dirty_grant;
1474         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1475                 /* borrow some grant from truncate to avoid the case that
1476                  * truncate uses up all avail grant */
1477                 cli->cl_lost_grant -= grant;
1478                 cli->cl_avail_grant += grant;
1479         }
1480         osc_wake_cache_waiters(cli);
1481         spin_unlock(&cli->cl_loi_list_lock);
1482         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu/%lu\n",
1483                lost_grant, cli->cl_lost_grant,
1484                cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_SHIFT,
1485                cli->cl_dirty_grant);
1486 }
1487
1488 /**
1489  * The companion to osc_enter_cache(), called when @oap is no longer part of
1490  * the dirty accounting due to error.
1491  */
1492 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1493 {
1494         spin_lock(&cli->cl_loi_list_lock);
1495         osc_release_write_grant(cli, &oap->oap_brw_page);
1496         spin_unlock(&cli->cl_loi_list_lock);
1497 }
1498
1499 /**
1500  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1501  * is available.
1502  */
1503 static int osc_enter_cache_try(struct client_obd *cli,
1504                                struct osc_async_page *oap,
1505                                int bytes)
1506 {
1507         int rc;
1508
1509         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1510
1511         rc = osc_reserve_grant(cli, bytes);
1512         if (rc < 0)
1513                 return 0;
1514
1515         if (cli->cl_dirty_pages < cli->cl_dirty_max_pages) {
1516                 if (atomic_long_add_return(1, &obd_dirty_pages) <=
1517                     obd_max_dirty_pages) {
1518                         osc_consume_write_grant(cli, &oap->oap_brw_page);
1519                         rc = 1;
1520                         goto out;
1521                 } else
1522                         atomic_long_dec(&obd_dirty_pages);
1523         }
1524         __osc_unreserve_grant(cli, bytes, bytes);
1525
1526 out:
1527         return rc;
1528 }
1529
1530 /* Following two inlines exist to pass code fragments
1531  * to wait_event_idle_exclusive_timeout_cmd().  Passing
1532  * code fragments as macro args can look confusing, so
1533  * we provide inlines to encapsulate them.
1534  */
1535 static inline void cli_unlock_and_unplug(const struct lu_env *env,
1536                                          struct client_obd *cli,
1537                                          struct osc_async_page *oap)
1538 {
1539         spin_unlock(&cli->cl_loi_list_lock);
1540         osc_io_unplug_async(env, cli, NULL);
1541         CDEBUG(D_CACHE,
1542                "%s: sleeping for cache space for %p\n",
1543                cli_name(cli), oap);
1544 }
1545
1546 static inline void cli_lock_after_unplug(struct client_obd *cli)
1547 {
1548         spin_lock(&cli->cl_loi_list_lock);
1549 }
1550 /**
1551  * The main entry to reserve dirty page accounting. Usually the grant reserved
1552  * in this function will be freed in bulk in osc_free_grant() unless it fails
1553  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1554  *
1555  * The process will be put into sleep if it's already run out of grant.
1556  */
1557 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1558                            struct osc_async_page *oap, int bytes)
1559 {
1560         struct osc_object *osc = oap->oap_obj;
1561         struct lov_oinfo *loi = osc->oo_oinfo;
1562         int rc = -EDQUOT;
1563         int remain;
1564         bool entered = false;
1565         /* We cannot wait for a long time here since we are holding ldlm lock
1566          * across the actual IO. If no requests complete fast (e.g. due to
1567          * overloaded OST that takes a long time to process everything, we'd
1568          * get evicted if we wait for a normal obd_timeout or some such.
1569          * So we try to wait half the time it would take the client to be
1570          * evicted by server which is half obd_timeout when AT is off
1571          * or at least ldlm_enqueue_min with AT on.
1572          * See LU-13131 */
1573         unsigned long timeout = cfs_time_seconds(AT_OFF ? obd_timeout / 2 :
1574                                                           ldlm_enqueue_min / 2);
1575
1576         ENTRY;
1577
1578         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1579
1580         spin_lock(&cli->cl_loi_list_lock);
1581
1582         /* force the caller to try sync io.  this can jump the list
1583          * of queued writes and create a discontiguous rpc stream */
1584         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1585             cli->cl_dirty_max_pages == 0 ||
1586             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) {
1587                 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1588                 GOTO(out, rc = -EDQUOT);
1589         }
1590
1591         /*
1592          * We can wait here for two reasons: too many dirty pages in cache, or
1593          * run out of grants. In both cases we should write dirty pages out.
1594          * Adding a cache waiter will trigger urgent write-out no matter what
1595          * RPC size will be.
1596          * The exiting condition (other than success) is no avail grants
1597          * and no dirty pages caching, that really means there is no space
1598          * on the OST.
1599          */
1600         remain = wait_event_idle_exclusive_timeout_cmd(
1601                 cli->cl_cache_waiters,
1602                 (entered = osc_enter_cache_try(cli, oap, bytes)) ||
1603                 (cli->cl_dirty_pages == 0 && cli->cl_w_in_flight == 0),
1604                 timeout,
1605                 cli_unlock_and_unplug(env, cli, oap),
1606                 cli_lock_after_unplug(cli));
1607
1608         if (entered) {
1609                 if (remain == timeout)
1610                         OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1611                 else
1612                         OSC_DUMP_GRANT(D_CACHE, cli,
1613                                        "finally got grant space\n");
1614                 wake_up(&cli->cl_cache_waiters);
1615                 rc = 0;
1616         } else if (remain == 0) {
1617                 OSC_DUMP_GRANT(D_CACHE, cli,
1618                                "timeout, fall back to sync i/o\n");
1619                 osc_extent_tree_dump(D_CACHE, osc);
1620                 /* fall back to synchronous I/O */
1621         } else {
1622                 OSC_DUMP_GRANT(D_CACHE, cli,
1623                                "no grant space, fall back to sync i/o\n");
1624                 wake_up_all(&cli->cl_cache_waiters);
1625         }
1626         EXIT;
1627 out:
1628         spin_unlock(&cli->cl_loi_list_lock);
1629         RETURN(rc);
1630 }
1631
1632 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1633 {
1634         int hprpc = !!list_empty(&osc->oo_hp_exts);
1635         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1636 }
1637
1638 /* This maintains the lists of pending pages to read/write for a given object
1639  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1640  * to quickly find objects that are ready to send an RPC. */
1641 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1642                          int cmd)
1643 {
1644         int invalid_import = 0;
1645         ENTRY;
1646
1647         /* if we have an invalid import we want to drain the queued pages
1648          * by forcing them through rpcs that immediately fail and complete
1649          * the pages.  recovery relies on this to empty the queued pages
1650          * before canceling the locks and evicting down the llite pages */
1651         if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1652                 invalid_import = 1;
1653
1654         if (cmd & OBD_BRW_WRITE) {
1655                 if (atomic_read(&osc->oo_nr_writes) == 0)
1656                         RETURN(0);
1657                 if (invalid_import) {
1658                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1659                         RETURN(1);
1660                 }
1661                 if (!list_empty(&osc->oo_hp_exts)) {
1662                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1663                         RETURN(1);
1664                 }
1665                 if (!list_empty(&osc->oo_urgent_exts)) {
1666                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1667                         RETURN(1);
1668                 }
1669                 /* trigger a write rpc stream as long as there are dirtiers
1670                  * waiting for space.  as they're waiting, they're not going to
1671                  * create more pages to coalesce with what's waiting..
1672                  */
1673                 if (waitqueue_active(&cli->cl_cache_waiters)) {
1674                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1675                         RETURN(1);
1676                 }
1677                 if (!list_empty(&osc->oo_full_exts)) {
1678                         CDEBUG(D_CACHE, "full extent ready, make an RPC\n");
1679                         RETURN(1);
1680                 }
1681         } else {
1682                 if (atomic_read(&osc->oo_nr_reads) == 0)
1683                         RETURN(0);
1684                 if (invalid_import) {
1685                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1686                         RETURN(1);
1687                 }
1688                 /* all read are urgent. */
1689                 if (!list_empty(&osc->oo_reading_exts))
1690                         RETURN(1);
1691         }
1692
1693         RETURN(0);
1694 }
1695
1696 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1697 {
1698         struct client_obd *cli = osc_cli(obj);
1699         if (cmd & OBD_BRW_WRITE) {
1700                 atomic_add(delta, &obj->oo_nr_writes);
1701                 atomic_add(delta, &cli->cl_pending_w_pages);
1702                 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1703         } else {
1704                 atomic_add(delta, &obj->oo_nr_reads);
1705                 atomic_add(delta, &cli->cl_pending_r_pages);
1706                 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1707         }
1708         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1709 }
1710
1711 static int osc_makes_hprpc(struct osc_object *obj)
1712 {
1713         return !list_empty(&obj->oo_hp_exts);
1714 }
1715
1716 static void on_list(struct list_head *item, struct list_head *list,
1717                     int should_be_on)
1718 {
1719         if (list_empty(item) && should_be_on)
1720                 list_add_tail(item, list);
1721         else if (!list_empty(item) && !should_be_on)
1722                 list_del_init(item);
1723 }
1724
1725 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1726  * can find pages to build into rpcs quickly */
1727 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1728 {
1729         if (osc_makes_hprpc(osc)) {
1730                 /* HP rpc */
1731                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1732                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1733         } else {
1734                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1735                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1736                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1737                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1738         }
1739
1740         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1741                 atomic_read(&osc->oo_nr_writes) > 0);
1742
1743         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1744                 atomic_read(&osc->oo_nr_reads) > 0);
1745
1746         return osc_is_ready(osc);
1747 }
1748
1749 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1750 {
1751         int is_ready;
1752
1753         spin_lock(&cli->cl_loi_list_lock);
1754         is_ready = __osc_list_maint(cli, osc);
1755         spin_unlock(&cli->cl_loi_list_lock);
1756
1757         return is_ready;
1758 }
1759
1760 /* this is trying to propogate async writeback errors back up to the
1761  * application.  As an async write fails we record the error code for later if
1762  * the app does an fsync.  As long as errors persist we force future rpcs to be
1763  * sync so that the app can get a sync error and break the cycle of queueing
1764  * pages for which writeback will fail. */
1765 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1766                            int rc)
1767 {
1768         if (rc) {
1769                 if (!ar->ar_rc)
1770                         ar->ar_rc = rc;
1771
1772                 ar->ar_force_sync = 1;
1773                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1774                 return;
1775
1776         }
1777
1778         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1779                 ar->ar_force_sync = 0;
1780 }
1781
1782 /* this must be called holding the loi list lock to give coverage to exit_cache,
1783  * async_flag maintenance, and oap_request */
1784 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1785                               struct osc_async_page *oap, int sent, int rc)
1786 {
1787         struct osc_object *osc = oap->oap_obj;
1788         struct lov_oinfo  *loi = osc->oo_oinfo;
1789         __u64 xid = 0;
1790
1791         ENTRY;
1792         if (oap->oap_request != NULL) {
1793                 xid = ptlrpc_req_xid(oap->oap_request);
1794                 ptlrpc_req_finished(oap->oap_request);
1795                 oap->oap_request = NULL;
1796         }
1797
1798         /* As the transfer for this page is being done, clear the flags */
1799         spin_lock(&oap->oap_lock);
1800         oap->oap_async_flags = 0;
1801         spin_unlock(&oap->oap_lock);
1802
1803         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1804                 spin_lock(&cli->cl_loi_list_lock);
1805                 osc_process_ar(&cli->cl_ar, xid, rc);
1806                 osc_process_ar(&loi->loi_ar, xid, rc);
1807                 spin_unlock(&cli->cl_loi_list_lock);
1808         }
1809
1810         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1811         if (rc)
1812                 CERROR("completion on oap %p obj %p returns %d.\n",
1813                        oap, osc, rc);
1814
1815         EXIT;
1816 }
1817
1818 struct extent_rpc_data {
1819         struct list_head        *erd_rpc_list;
1820         unsigned int            erd_page_count;
1821         unsigned int            erd_max_pages;
1822         unsigned int            erd_max_chunks;
1823         unsigned int            erd_max_extents;
1824 };
1825
1826 static inline unsigned osc_extent_chunks(const struct osc_extent *ext)
1827 {
1828         struct client_obd *cli = osc_cli(ext->oe_obj);
1829         unsigned ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1830
1831         return (ext->oe_end >> ppc_bits) - (ext->oe_start >> ppc_bits) + 1;
1832 }
1833
1834 static inline bool
1835 can_merge(const struct osc_extent *ext, const struct osc_extent *in_rpc)
1836 {
1837         if (ext->oe_no_merge || in_rpc->oe_no_merge)
1838                 return false;
1839
1840         if (ext->oe_srvlock != in_rpc->oe_srvlock)
1841                 return false;
1842
1843         if (ext->oe_ndelay != in_rpc->oe_ndelay)
1844                 return false;
1845
1846         if (!ext->oe_grants != !in_rpc->oe_grants)
1847                 return false;
1848
1849         if (ext->oe_dio != in_rpc->oe_dio)
1850                 return false;
1851
1852         /* It's possible to have overlap on DIO */
1853         if (in_rpc->oe_dio && overlapped(ext, in_rpc))
1854                 return false;
1855
1856         if (ext->oe_is_rdma_only != in_rpc->oe_is_rdma_only)
1857                 return false;
1858
1859         return true;
1860 }
1861
1862 /**
1863  * Try to add extent to one RPC. We need to think about the following things:
1864  * - # of pages must not be over max_pages_per_rpc
1865  * - extent must be compatible with previous ones
1866  */
1867 static int try_to_add_extent_for_io(struct client_obd *cli,
1868                                     struct osc_extent *ext,
1869                                     struct extent_rpc_data *data)
1870 {
1871         struct osc_extent *tmp;
1872         unsigned int chunk_count;
1873         ENTRY;
1874
1875         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1876                 ext);
1877         OSC_EXTENT_DUMP(D_CACHE, ext, "trying to add this extent\n");
1878
1879         if (data->erd_max_extents == 0)
1880                 RETURN(0);
1881
1882         chunk_count = osc_extent_chunks(ext);
1883         EASSERTF(data->erd_page_count != 0 ||
1884                  chunk_count <= data->erd_max_chunks, ext,
1885                  "The first extent to be fit in a RPC contains %u chunks, "
1886                  "which is over the limit %u.\n", chunk_count,
1887                  data->erd_max_chunks);
1888         if (chunk_count > data->erd_max_chunks)
1889                 RETURN(0);
1890
1891         data->erd_max_pages = max(ext->oe_mppr, data->erd_max_pages);
1892         EASSERTF(data->erd_page_count != 0 ||
1893                 ext->oe_nr_pages <= data->erd_max_pages, ext,
1894                 "The first extent to be fit in a RPC contains %u pages, "
1895                 "which is over the limit %u.\n", ext->oe_nr_pages,
1896                 data->erd_max_pages);
1897         if (data->erd_page_count + ext->oe_nr_pages > data->erd_max_pages)
1898                 RETURN(0);
1899
1900         list_for_each_entry(tmp, data->erd_rpc_list, oe_link) {
1901                 EASSERT(tmp->oe_owner == current, tmp);
1902
1903                 if (!can_merge(ext, tmp))
1904                         RETURN(0);
1905         }
1906
1907         data->erd_max_extents--;
1908         data->erd_max_chunks -= chunk_count;
1909         data->erd_page_count += ext->oe_nr_pages;
1910         list_move_tail(&ext->oe_link, data->erd_rpc_list);
1911         ext->oe_owner = current;
1912         RETURN(1);
1913 }
1914
1915 /**
1916  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1917  * get_write_extent() takes all appropriate extents in atomic.
1918  *
1919  * The following policy is used to collect extents for IO:
1920  * 1. Add as many HP extents as possible;
1921  * 2. Add the first urgent extent in urgent extent list and take it out of
1922  *    urgent list;
1923  * 3. Add subsequent extents of this urgent extent;
1924  * 4. If urgent list is not empty, goto 2;
1925  * 5. Traverse the extent tree from the 1st extent;
1926  * 6. Above steps exit if there is no space in this RPC.
1927  */
1928 static unsigned int get_write_extents(struct osc_object *obj,
1929                                       struct list_head *rpclist)
1930 {
1931         struct client_obd *cli = osc_cli(obj);
1932         struct osc_extent *ext;
1933         struct extent_rpc_data data = {
1934                 .erd_rpc_list   = rpclist,
1935                 .erd_page_count = 0,
1936                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
1937                 .erd_max_chunks = osc_max_write_chunks(cli),
1938                 .erd_max_extents = 256,
1939         };
1940
1941         assert_osc_object_is_locked(obj);
1942         while ((ext = list_first_entry_or_null(&obj->oo_hp_exts,
1943                                                struct osc_extent,
1944                                                oe_link)) != NULL) {
1945                 if (!try_to_add_extent_for_io(cli, ext, &data))
1946                         return data.erd_page_count;
1947                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
1948         }
1949         if (data.erd_page_count == data.erd_max_pages)
1950                 return data.erd_page_count;
1951
1952         while ((ext = list_first_entry_or_null(&obj->oo_urgent_exts,
1953                                                struct osc_extent,
1954                                                oe_link)) != NULL) {
1955                 if (!try_to_add_extent_for_io(cli, ext, &data))
1956                         return data.erd_page_count;
1957         }
1958         if (data.erd_page_count == data.erd_max_pages)
1959                 return data.erd_page_count;
1960
1961         /* One key difference between full extents and other extents: full
1962          * extents can usually only be added if the rpclist was empty, so if we
1963          * can't add one, we continue on to trying to add normal extents.  This
1964          * is so we don't miss adding extra extents to an RPC containing high
1965          * priority or urgent extents.
1966          */
1967         while ((ext = list_first_entry_or_null(&obj->oo_full_exts,
1968                                                struct osc_extent,
1969                                                oe_link)) != NULL) {
1970                 if (!try_to_add_extent_for_io(cli, ext, &data))
1971                         break;
1972         }
1973         if (data.erd_page_count == data.erd_max_pages)
1974                 return data.erd_page_count;
1975
1976         for (ext = first_extent(obj);
1977              ext;
1978              ext = next_extent(ext)) {
1979                 if ((ext->oe_state != OES_CACHE) ||
1980                     /* this extent may be already in current rpclist */
1981                     (!list_empty(&ext->oe_link) && ext->oe_owner))
1982                         continue;
1983
1984                 if (!try_to_add_extent_for_io(cli, ext, &data))
1985                         return data.erd_page_count;
1986         }
1987         return data.erd_page_count;
1988 }
1989
1990 static int
1991 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
1992                    struct osc_object *osc)
1993 __must_hold(osc)
1994 {
1995         LIST_HEAD(rpclist);
1996         struct osc_extent *ext;
1997         struct osc_extent *tmp;
1998         struct osc_extent *first = NULL;
1999         unsigned int page_count = 0;
2000         int srvlock = 0;
2001         int rc = 0;
2002         ENTRY;
2003
2004         assert_osc_object_is_locked(osc);
2005
2006         page_count = get_write_extents(osc, &rpclist);
2007         LASSERT(equi(page_count == 0, list_empty(&rpclist)));
2008
2009         if (list_empty(&rpclist))
2010                 RETURN(0);
2011
2012         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2013
2014         list_for_each_entry(ext, &rpclist, oe_link) {
2015                 LASSERT(ext->oe_state == OES_CACHE ||
2016                         ext->oe_state == OES_LOCK_DONE);
2017                 if (ext->oe_state == OES_CACHE)
2018                         osc_extent_state_set(ext, OES_LOCKING);
2019                 else
2020                         osc_extent_state_set(ext, OES_RPC);
2021         }
2022
2023         /* we're going to grab page lock, so release object lock because
2024          * lock order is page lock -> object lock. */
2025         osc_object_unlock(osc);
2026
2027         list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2028                 if (ext->oe_state == OES_LOCKING) {
2029                         rc = osc_extent_make_ready(env, ext);
2030                         if (unlikely(rc < 0)) {
2031                                 list_del_init(&ext->oe_link);
2032                                 osc_extent_finish(env, ext, 0, rc);
2033                                 continue;
2034                         }
2035                 }
2036                 if (first == NULL) {
2037                         first = ext;
2038                         srvlock = ext->oe_srvlock;
2039                 } else {
2040                         LASSERT(srvlock == ext->oe_srvlock);
2041                 }
2042         }
2043
2044         if (!list_empty(&rpclist)) {
2045                 LASSERT(page_count > 0);
2046                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2047                 LASSERT(list_empty(&rpclist));
2048         }
2049
2050         osc_object_lock(osc);
2051         RETURN(rc);
2052 }
2053
2054 /**
2055  * prepare pages for ASYNC io and put pages in send queue.
2056  *
2057  * \param cmd OBD_BRW_* macroses
2058  * \param lop pending pages
2059  *
2060  * \return zero if no page added to send queue.
2061  * \return 1 if pages successfully added to send queue.
2062  * \return negative on errors.
2063  */
2064 static int
2065 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2066                   struct osc_object *osc)
2067 __must_hold(osc)
2068 {
2069         struct osc_extent *ext;
2070         struct osc_extent *next;
2071         LIST_HEAD(rpclist);
2072         struct extent_rpc_data data = {
2073                 .erd_rpc_list   = &rpclist,
2074                 .erd_page_count = 0,
2075                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
2076                 .erd_max_chunks = UINT_MAX,
2077                 .erd_max_extents = UINT_MAX,
2078         };
2079         int rc = 0;
2080         ENTRY;
2081
2082         assert_osc_object_is_locked(osc);
2083         list_for_each_entry_safe(ext, next, &osc->oo_reading_exts, oe_link) {
2084                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2085                 if (!try_to_add_extent_for_io(cli, ext, &data))
2086                         break;
2087                 osc_extent_state_set(ext, OES_RPC);
2088                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2089         }
2090         LASSERT(data.erd_page_count <= data.erd_max_pages);
2091
2092         osc_update_pending(osc, OBD_BRW_READ, -data.erd_page_count);
2093
2094         if (!list_empty(&rpclist)) {
2095                 osc_object_unlock(osc);
2096
2097                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2098                 LASSERT(list_empty(&rpclist));
2099
2100                 osc_object_lock(osc);
2101         }
2102         RETURN(rc);
2103 }
2104
2105 #define list_to_obj(list, item) ({                                            \
2106         struct list_head *__tmp = (list)->next;                               \
2107         list_del_init(__tmp);                                         \
2108         list_entry(__tmp, struct osc_object, oo_##item);                      \
2109 })
2110
2111 /* This is called by osc_check_rpcs() to find which objects have pages that
2112  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2113 static struct osc_object *osc_next_obj(struct client_obd *cli)
2114 {
2115         ENTRY;
2116
2117         /* First return objects that have blocked locks so that they
2118          * will be flushed quickly and other clients can get the lock,
2119          * then objects which have pages ready to be stuffed into RPCs */
2120         if (!list_empty(&cli->cl_loi_hp_ready_list))
2121                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2122         if (!list_empty(&cli->cl_loi_ready_list))
2123                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2124
2125         /* then if we have cache waiters, return all objects with queued
2126          * writes.  This is especially important when many small files
2127          * have filled up the cache and not been fired into rpcs because
2128          * they don't pass the nr_pending/object threshhold
2129          */
2130         if (waitqueue_active(&cli->cl_cache_waiters) &&
2131             !list_empty(&cli->cl_loi_write_list))
2132                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2133
2134         /* then return all queued objects when we have an invalid import
2135          * so that they get flushed */
2136         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2137                 if (!list_empty(&cli->cl_loi_write_list))
2138                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2139                                            write_item));
2140                 if (!list_empty(&cli->cl_loi_read_list))
2141                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2142                                            read_item));
2143         }
2144         RETURN(NULL);
2145 }
2146
2147 /* called with the loi list lock held */
2148 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2149 __must_hold(&cli->cl_loi_list_lock)
2150 {
2151         struct osc_object *osc;
2152         int rc = 0;
2153         ENTRY;
2154
2155         while ((osc = osc_next_obj(cli)) != NULL) {
2156                 struct cl_object *obj = osc2cl(osc);
2157                 struct lu_ref_link link;
2158
2159                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2160
2161                 /* even if we have reached our max in flight RPCs, we still
2162                  * allow all high-priority RPCs through to prevent their
2163                  * starvation and leading to server evicting us for not
2164                  * writing out pages in a timely manner LU-13131 */
2165                 if (osc_max_rpc_in_flight(cli, osc) &&
2166                     list_empty(&osc->oo_hp_exts)) {
2167                         __osc_list_maint(cli, osc);
2168                         break;
2169                 }
2170
2171                 cl_object_get(obj);
2172                 spin_unlock(&cli->cl_loi_list_lock);
2173                 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2174
2175                 /* attempt some read/write balancing by alternating between
2176                  * reads and writes in an object.  The makes_rpc checks here
2177                  * would be redundant if we were getting read/write work items
2178                  * instead of objects.  we don't want send_oap_rpc to drain a
2179                  * partial read pending queue when we're given this object to
2180                  * do io on writes while there are cache waiters */
2181                 osc_object_lock(osc);
2182                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2183                         rc = osc_send_write_rpc(env, cli, osc);
2184                         if (rc < 0) {
2185                                 CERROR("Write request failed with %d\n", rc);
2186
2187                                 /* osc_send_write_rpc failed, mostly because of
2188                                  * memory pressure.
2189                                  *
2190                                  * It can't break here, because if:
2191                                  *  - a page was submitted by osc_io_submit, so
2192                                  *    page locked;
2193                                  *  - no request in flight
2194                                  *  - no subsequent request
2195                                  * The system will be in live-lock state,
2196                                  * because there is no chance to call
2197                                  * osc_io_unplug() and osc_check_rpcs() any
2198                                  * more. pdflush can't help in this case,
2199                                  * because it might be blocked at grabbing
2200                                  * the page lock as we mentioned.
2201                                  *
2202                                  * Anyway, continue to drain pages. */
2203                                 /* break; */
2204                         }
2205                 }
2206                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2207                         rc = osc_send_read_rpc(env, cli, osc);
2208                         if (rc < 0)
2209                                 CERROR("Read request failed with %d\n", rc);
2210                 }
2211                 osc_object_unlock(osc);
2212
2213                 osc_list_maint(cli, osc);
2214                 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2215                 cl_object_put(env, obj);
2216
2217                 spin_lock(&cli->cl_loi_list_lock);
2218         }
2219 }
2220
2221 int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2222                    struct osc_object *osc, int async)
2223 {
2224         int rc = 0;
2225
2226         if (osc != NULL && osc_list_maint(cli, osc) == 0)
2227                 return 0;
2228
2229         if (!async) {
2230                 spin_lock(&cli->cl_loi_list_lock);
2231                 osc_check_rpcs(env, cli);
2232                 spin_unlock(&cli->cl_loi_list_lock);
2233         } else {
2234                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2235                 LASSERT(cli->cl_writeback_work != NULL);
2236                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2237         }
2238         return rc;
2239 }
2240 EXPORT_SYMBOL(osc_io_unplug0);
2241
2242 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2243                         struct cl_page *page, loff_t offset)
2244 {
2245         struct osc_async_page *oap = &ops->ops_oap;
2246         struct page           *vmpage = page->cp_vmpage;
2247         ENTRY;
2248
2249         if (!page)
2250                 return cfs_size_round(sizeof(*oap));
2251
2252         oap->oap_magic = OAP_MAGIC;
2253         oap->oap_obj = osc;
2254
2255         oap->oap_page = vmpage;
2256         oap->oap_obj_off = offset;
2257         LASSERT(!(offset & ~PAGE_MASK));
2258
2259         /* Count of transient (direct i/o) pages is always stable by the time
2260          * they're submitted.  Setting this here lets us avoid calling
2261          * cl_page_clip later to set this.
2262          */
2263         if (page->cp_type == CPT_TRANSIENT)
2264                 oap->oap_async_flags |= ASYNC_COUNT_STABLE|ASYNC_URGENT|
2265                                         ASYNC_READY;
2266
2267         INIT_LIST_HEAD(&oap->oap_pending_item);
2268         INIT_LIST_HEAD(&oap->oap_rpc_item);
2269
2270         spin_lock_init(&oap->oap_lock);
2271         CDEBUG(D_INFO, "oap %p vmpage %p obj off %llu\n",
2272                oap, vmpage, oap->oap_obj_off);
2273         RETURN(0);
2274 }
2275 EXPORT_SYMBOL(osc_prep_async_page);
2276
2277 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2278                        struct osc_page *ops, cl_commit_cbt cb)
2279 {
2280         struct osc_io *oio = osc_env_io(env);
2281         struct osc_extent     *ext = NULL;
2282         struct osc_async_page *oap = &ops->ops_oap;
2283         struct osc_object     *osc = oap->oap_obj;
2284         struct client_obd     *cli = osc_cli(osc);
2285         struct pagevec        *pvec = &osc_env_info(env)->oti_pagevec;
2286         pgoff_t index;
2287         unsigned int tmp;
2288         unsigned int grants = 0;
2289         u32    brw_flags = OBD_BRW_ASYNC;
2290         int    cmd = OBD_BRW_WRITE;
2291         int    need_release = 0;
2292         int    rc = 0;
2293         ENTRY;
2294
2295         if (oap->oap_magic != OAP_MAGIC)
2296                 RETURN(-EINVAL);
2297
2298         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2299                 RETURN(-EIO);
2300
2301         if (!list_empty(&oap->oap_pending_item) ||
2302             !list_empty(&oap->oap_rpc_item))
2303                 RETURN(-EBUSY);
2304
2305         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2306         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2307         if (io->ci_noquota) {
2308                 brw_flags |= OBD_BRW_NOQUOTA;
2309                 cmd |= OBD_BRW_NOQUOTA;
2310         }
2311
2312         if (oio->oi_cap_sys_resource) {
2313                 brw_flags |= OBD_BRW_SYS_RESOURCE;
2314                 cmd |= OBD_BRW_SYS_RESOURCE;
2315         }
2316
2317         /* check if the file's owner/group is over quota */
2318         /* do not check for root without root squash, because in this case
2319          * we should bypass quota
2320          */
2321         if ((!oio->oi_cap_sys_resource ||
2322              cli->cl_root_squash) &&
2323             !io->ci_noquota) {
2324                 struct cl_object *obj;
2325                 struct cl_attr   *attr;
2326                 unsigned int qid[LL_MAXQUOTAS];
2327
2328                 obj = cl_object_top(&osc->oo_cl);
2329                 attr = &osc_env_info(env)->oti_attr;
2330
2331                 cl_object_attr_lock(obj);
2332                 rc = cl_object_attr_get(env, obj, attr);
2333                 cl_object_attr_unlock(obj);
2334
2335                 qid[USRQUOTA] = attr->cat_uid;
2336                 qid[GRPQUOTA] = attr->cat_gid;
2337                 qid[PRJQUOTA] = attr->cat_projid;
2338                 if (rc == 0 && osc_quota_chkdq(cli, qid) == -EDQUOT)
2339                         rc = -EDQUOT;
2340                 if (rc)
2341                         RETURN(rc);
2342         }
2343
2344         oap->oap_cmd = cmd;
2345         oap->oap_page_off = ops->ops_from;
2346         oap->oap_count = ops->ops_to - ops->ops_from + 1;
2347         /* No need to hold a lock here,
2348          * since this page is not in any list yet. */
2349         oap->oap_async_flags = 0;
2350         oap->oap_brw_flags = brw_flags;
2351
2352         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2353                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2354
2355         index = osc_index(oap2osc(oap));
2356
2357         /* Add this page into extent by the following steps:
2358          * 1. if there exists an active extent for this IO, mostly this page
2359          *    can be added to the active extent and sometimes we need to
2360          *    expand extent to accomodate this page;
2361          * 2. otherwise, a new extent will be allocated. */
2362
2363         ext = oio->oi_active;
2364         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2365                 /* one chunk plus extent overhead must be enough to write this
2366                  * page */
2367                 grants = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2368                 if (ext->oe_end >= index)
2369                         grants = 0;
2370
2371                 /* it doesn't need any grant to dirty this page */
2372                 spin_lock(&cli->cl_loi_list_lock);
2373                 rc = osc_enter_cache_try(cli, oap, grants);
2374                 if (rc == 0) { /* try failed */
2375                         grants = 0;
2376                         need_release = 1;
2377                 } else if (ext->oe_end < index) {
2378                         tmp = grants;
2379                         /* try to expand this extent */
2380                         rc = osc_extent_expand(ext, index, &tmp);
2381                         if (rc < 0) {
2382                                 need_release = 1;
2383                                 /* don't free reserved grant */
2384                         } else {
2385                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2386                                                 "expanded for %lu.\n", index);
2387                                 osc_unreserve_grant_nolock(cli, grants, tmp);
2388                                 grants = 0;
2389                         }
2390                 }
2391                 spin_unlock(&cli->cl_loi_list_lock);
2392                 rc = 0;
2393         } else if (ext != NULL) {
2394                 /* index is located outside of active extent */
2395                 need_release = 1;
2396         }
2397         if (need_release) {
2398                 osc_extent_release(env, ext);
2399                 oio->oi_active = NULL;
2400                 ext = NULL;
2401         }
2402
2403         if (ext == NULL) {
2404                 tmp = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2405
2406                 /* try to find new extent to cover this page */
2407                 LASSERT(oio->oi_active == NULL);
2408                 /* we may have allocated grant for this page if we failed
2409                  * to expand the previous active extent. */
2410                 LASSERT(ergo(grants > 0, grants >= tmp));
2411
2412                 rc = 0;
2413
2414                 /* We must not hold a page lock while we do osc_enter_cache()
2415                  * or osc_extent_find(), so we must mark dirty & unlock
2416                  * any pages in the write commit pagevec. */
2417                 if (pagevec_count(pvec)) {
2418                         cb(env, io, pvec);
2419                         pagevec_reinit(pvec);
2420                 }
2421
2422                 if (grants == 0) {
2423                         rc = osc_enter_cache(env, cli, oap, tmp);
2424                         if (rc == 0)
2425                                 grants = tmp;
2426                 }
2427
2428                 tmp = grants;
2429                 if (rc == 0) {
2430                         ext = osc_extent_find(env, osc, index, &tmp);
2431                         if (IS_ERR(ext)) {
2432                                 LASSERT(tmp == grants);
2433                                 osc_exit_cache(cli, oap);
2434                                 rc = PTR_ERR(ext);
2435                                 ext = NULL;
2436                         } else {
2437                                 oio->oi_active = ext;
2438                         }
2439                 }
2440                 if (grants > 0)
2441                         osc_unreserve_grant(cli, grants, tmp);
2442         }
2443
2444         LASSERT(ergo(rc == 0, ext != NULL));
2445         if (ext != NULL) {
2446                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2447                          ext, "index = %lu.\n", index);
2448                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2449
2450                 osc_object_lock(osc);
2451                 if (ext->oe_nr_pages == 0)
2452                         ext->oe_srvlock = ops->ops_srvlock;
2453                 else
2454                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2455                 ++ext->oe_nr_pages;
2456                 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2457                 osc_object_unlock(osc);
2458
2459                 if (!ext->oe_layout_version)
2460                         ext->oe_layout_version = io->ci_layout_version;
2461         }
2462
2463         RETURN(rc);
2464 }
2465
2466 int osc_teardown_async_page(const struct lu_env *env,
2467                             struct osc_object *obj, struct osc_page *ops)
2468 {
2469         struct osc_async_page *oap = &ops->ops_oap;
2470         int rc = 0;
2471         ENTRY;
2472
2473         LASSERT(oap->oap_magic == OAP_MAGIC);
2474
2475         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2476                oap, ops, osc_index(oap2osc(oap)));
2477
2478         if (!list_empty(&oap->oap_rpc_item)) {
2479                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2480                 rc = -EBUSY;
2481         } else if (!list_empty(&oap->oap_pending_item)) {
2482                 struct osc_extent *ext = NULL;
2483
2484                 osc_object_lock(obj);
2485                 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2486                 osc_object_unlock(obj);
2487                 /* only truncated pages are allowed to be taken out.
2488                  * See osc_extent_truncate() and osc_cache_truncate_start()
2489                  * for details. */
2490                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2491                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2492                                         osc_index(oap2osc(oap)));
2493                         rc = -EBUSY;
2494                 }
2495                 if (ext != NULL)
2496                         osc_extent_put(env, ext);
2497         }
2498         RETURN(rc);
2499 }
2500
2501 /**
2502  * This is called when a page is picked up by kernel to write out.
2503  *
2504  * We should find out the corresponding extent and add the whole extent
2505  * into urgent list. The extent may be being truncated or used, handle it
2506  * carefully.
2507  */
2508 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2509                          struct osc_page *ops)
2510 {
2511         struct osc_extent *ext   = NULL;
2512         struct osc_object *obj   = osc_page_object(ops);
2513         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2514         pgoff_t            index = osc_index(ops);
2515         struct osc_async_page *oap = &ops->ops_oap;
2516         bool unplug = false;
2517         int rc = 0;
2518         ENTRY;
2519
2520         osc_object_lock(obj);
2521         ext = osc_extent_lookup(obj, index);
2522         if (ext == NULL) {
2523                 osc_extent_tree_dump(D_ERROR, obj);
2524                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2525         }
2526
2527         switch (ext->oe_state) {
2528         case OES_RPC:
2529         case OES_LOCK_DONE:
2530                 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2531                 LASSERT(0);
2532                 break;
2533         case OES_LOCKING:
2534                 /* If we know this extent is being written out, we should abort
2535                  * so that the writer can make this page ready. Otherwise, there
2536                  * exists a deadlock problem because other process can wait for
2537                  * page writeback bit holding page lock; and meanwhile in
2538                  * vvp_page_make_ready(), we need to grab page lock before
2539                  * really sending the RPC. */
2540         case OES_TRUNC:
2541                 /* race with truncate, page will be redirtied */
2542         case OES_ACTIVE:
2543                 /* The extent is active so we need to abort and let the caller
2544                  * re-dirty the page. If we continued on here, and we were the
2545                  * one making the extent active, we could deadlock waiting for
2546                  * the page writeback to clear but it won't because the extent
2547                  * is active and won't be written out. */
2548                 GOTO(out, rc = -EAGAIN);
2549         default:
2550                 break;
2551         }
2552
2553         rc = cl_page_prep(env, io, cp, CRT_WRITE);
2554         if (rc)
2555                 GOTO(out, rc);
2556
2557         spin_lock(&oap->oap_lock);
2558         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2559         spin_unlock(&oap->oap_lock);
2560
2561         if (current->flags & PF_MEMALLOC)
2562                 ext->oe_memalloc = 1;
2563
2564         ext->oe_urgent = 1;
2565         if (ext->oe_state == OES_CACHE) {
2566                 OSC_EXTENT_DUMP(D_CACHE, ext,
2567                                 "flush page %p make it urgent.\n", oap);
2568                 if (list_empty(&ext->oe_link))
2569                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2570                 unplug = true;
2571         }
2572         rc = 0;
2573         EXIT;
2574
2575 out:
2576         osc_object_unlock(obj);
2577         osc_extent_put(env, ext);
2578         if (unplug)
2579                 osc_io_unplug_async(env, osc_cli(obj), obj);
2580         return rc;
2581 }
2582
2583 int osc_queue_sync_pages(const struct lu_env *env, struct cl_io *io,
2584                          struct osc_object *obj, struct list_head *list,
2585                          int brw_flags)
2586 {
2587         struct client_obd     *cli = osc_cli(obj);
2588         struct osc_extent     *ext;
2589         struct osc_async_page *oap;
2590         int     page_count = 0;
2591         int     mppr       = cli->cl_max_pages_per_rpc;
2592         bool    can_merge   = true;
2593         pgoff_t start      = CL_PAGE_EOF;
2594         pgoff_t end        = 0;
2595         ENTRY;
2596
2597         list_for_each_entry(oap, list, oap_pending_item) {
2598                 struct osc_page *opg = oap2osc_page(oap);
2599                 pgoff_t index = osc_index(opg);
2600
2601                 if (index > end)
2602                         end = index;
2603                 if (index < start)
2604                         start = index;
2605                 ++page_count;
2606                 mppr <<= (page_count > mppr);
2607
2608                 if (unlikely(opg->ops_from > 0 ||
2609                              opg->ops_to < PAGE_SIZE - 1))
2610                         can_merge = false;
2611         }
2612
2613         ext = osc_extent_alloc(obj);
2614         if (ext == NULL) {
2615                 struct osc_async_page *tmp;
2616
2617                 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2618                         list_del_init(&oap->oap_pending_item);
2619                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2620                 }
2621                 RETURN(-ENOMEM);
2622         }
2623
2624         ext->oe_rw = !!(brw_flags & OBD_BRW_READ);
2625         ext->oe_sync = 1;
2626         ext->oe_no_merge = !can_merge;
2627         ext->oe_urgent = 1;
2628         ext->oe_start = start;
2629         ext->oe_end = ext->oe_max_end = end;
2630         ext->oe_obj = obj;
2631         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2632         ext->oe_ndelay = !!(brw_flags & OBD_BRW_NDELAY);
2633         ext->oe_dio = !!(brw_flags & OBD_BRW_NOCACHE);
2634         if (ext->oe_dio && !ext->oe_rw) { /* direct io write */
2635                 int grants;
2636                 int ppc;
2637
2638                 ppc = 1 << (cli->cl_chunkbits - PAGE_SHIFT);
2639                 grants = cli->cl_grant_extent_tax;
2640                 grants += (1 << cli->cl_chunkbits) *
2641                         ((page_count + ppc - 1) / ppc);
2642
2643                 CDEBUG(D_CACHE, "requesting %d bytes grant\n", grants);
2644                 spin_lock(&cli->cl_loi_list_lock);
2645                 if (osc_reserve_grant(cli, grants) == 0) {
2646                         list_for_each_entry(oap, list, oap_pending_item) {
2647                                 osc_consume_write_grant(cli,
2648                                                         &oap->oap_brw_page);
2649                         }
2650                         atomic_long_add(page_count, &obd_dirty_pages);
2651                         osc_unreserve_grant_nolock(cli, grants, 0);
2652                         ext->oe_grants = grants;
2653                 } else {
2654                         /* We cannot report ENOSPC correctly if we do parallel
2655                          * DIO (async RPC submission), so turn off parallel dio
2656                          * if there is not sufficient grant available.  This
2657                          * makes individual RPCs synchronous.
2658                          */
2659                         io->ci_parallel_dio = false;
2660                         CDEBUG(D_CACHE,
2661                         "not enough grant available, switching to sync for this i/o\n");
2662                 }
2663                 spin_unlock(&cli->cl_loi_list_lock);
2664                 osc_update_next_shrink(cli);
2665         }
2666
2667         ext->oe_is_rdma_only = !!(brw_flags & OBD_BRW_RDMA_ONLY);
2668         ext->oe_nr_pages = page_count;
2669         ext->oe_mppr = mppr;
2670         list_splice_init(list, &ext->oe_pages);
2671         ext->oe_layout_version = io->ci_layout_version;
2672
2673         osc_object_lock(obj);
2674         /* Reuse the initial refcount for RPC, don't drop it */
2675         osc_extent_state_set(ext, OES_LOCK_DONE);
2676         if (!ext->oe_rw) { /* write */
2677                 if (!ext->oe_srvlock && !ext->oe_dio) {
2678                         /* The most likely case here is from lack of grants
2679                          * so we are either out of quota or out of space.
2680                          * Since this means we are holding locks across
2681                          * potentially multi-striped IO, we must send out
2682                          * everything out instantly to avoid prolonged
2683                          * waits resulting in lock eviction (likely since
2684                          * the extended wait in osc_cache_enter() did not
2685                          * yield any additional grant due to a timeout.
2686                          * LU-13131 */
2687                         ext->oe_hp = 1;
2688                         list_add_tail(&ext->oe_link, &obj->oo_hp_exts);
2689                 } else {
2690                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2691                 }
2692                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2693         } else {
2694                 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2695                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2696         }
2697         osc_object_unlock(obj);
2698
2699         osc_io_unplug_async(env, cli, obj);
2700         RETURN(0);
2701 }
2702
2703 /**
2704  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2705  */
2706 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
2707                              __u64 size, struct osc_extent **extp)
2708 {
2709         struct client_obd *cli = osc_cli(obj);
2710         struct osc_extent *ext;
2711         struct osc_extent *waiting = NULL;
2712         pgoff_t index;
2713         LIST_HEAD(list);
2714         int result = 0;
2715         bool partial;
2716         ENTRY;
2717
2718         /* pages with index greater or equal to index will be truncated. */
2719         index = cl_index(osc2cl(obj), size);
2720         partial = size > cl_offset(osc2cl(obj), index);
2721
2722 again:
2723         osc_object_lock(obj);
2724         ext = osc_extent_search(obj, index);
2725         if (ext == NULL)
2726                 ext = first_extent(obj);
2727         else if (ext->oe_end < index)
2728                 ext = next_extent(ext);
2729         while (ext != NULL) {
2730                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2731
2732                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2733                         /* if ext is in urgent state, it means there must exist
2734                          * a page already having been flushed by write_page().
2735                          * We have to wait for this extent because we can't
2736                          * truncate that page. */
2737                         OSC_EXTENT_DUMP(D_CACHE, ext,
2738                                         "waiting for busy extent\n");
2739                         waiting = osc_extent_get(ext);
2740                         break;
2741                 }
2742
2743                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:%llu.\n", size);
2744
2745                 osc_extent_get(ext);
2746                 if (ext->oe_state == OES_ACTIVE) {
2747                         /* though we grab inode mutex for write path, but we
2748                          * release it before releasing extent(in osc_io_end()),
2749                          * so there is a race window that an extent is still
2750                          * in OES_ACTIVE when truncate starts. */
2751                         LASSERT(!ext->oe_trunc_pending);
2752                         ext->oe_trunc_pending = 1;
2753                 } else {
2754                         EASSERT(ext->oe_state == OES_CACHE, ext);
2755                         osc_extent_state_set(ext, OES_TRUNC);
2756                         osc_update_pending(obj, OBD_BRW_WRITE,
2757                                            -ext->oe_nr_pages);
2758                 }
2759                 /* This extent could be on the full extents list, that's OK */
2760                 EASSERT(!ext->oe_hp && !ext->oe_urgent, ext);
2761                 if (!list_empty(&ext->oe_link))
2762                         list_move_tail(&ext->oe_link, &list);
2763                 else
2764                         list_add_tail(&ext->oe_link, &list);
2765
2766                 ext = next_extent(ext);
2767         }
2768         osc_object_unlock(obj);
2769
2770         osc_list_maint(cli, obj);
2771
2772         while ((ext = list_first_entry_or_null(&list,
2773                                                struct osc_extent,
2774                                                oe_link)) != NULL) {
2775                 int rc;
2776
2777                 list_del_init(&ext->oe_link);
2778
2779                 /* extent may be in OES_ACTIVE state because inode mutex
2780                  * is released before osc_io_end() in file write case */
2781                 if (ext->oe_state != OES_TRUNC)
2782                         osc_extent_wait(env, ext, OES_TRUNC);
2783
2784                 rc = osc_extent_truncate(ext, index, partial);
2785                 if (rc < 0) {
2786                         if (result == 0)
2787                                 result = rc;
2788
2789                         OSC_EXTENT_DUMP(D_ERROR, ext,
2790                                         "truncate error %d\n", rc);
2791                 } else if (ext->oe_nr_pages == 0) {
2792                         osc_extent_remove(ext);
2793                 } else {
2794                         /* this must be an overlapped extent which means only
2795                          * part of pages in this extent have been truncated.
2796                          */
2797                         EASSERTF(ext->oe_start <= index, ext,
2798                                  "trunc index = %lu/%d.\n", index, partial);
2799                         /* fix index to skip this partially truncated extent */
2800                         index = ext->oe_end + 1;
2801                         partial = false;
2802
2803                         /* we need to hold this extent in OES_TRUNC state so
2804                          * that no writeback will happen. This is to avoid
2805                          * BUG 17397.
2806                          * Only partial truncate can reach here, if @size is
2807                          * not zero, the caller should provide a valid @extp. */
2808                         LASSERT(*extp == NULL);
2809                         *extp = osc_extent_get(ext);
2810                         OSC_EXTENT_DUMP(D_CACHE, ext,
2811                                         "trunc at %llu\n", size);
2812                 }
2813                 osc_extent_put(env, ext);
2814         }
2815         if (waiting != NULL) {
2816                 int rc;
2817
2818                 /* ignore the result of osc_extent_wait the write initiator
2819                  * should take care of it. */
2820                 rc = osc_extent_wait(env, waiting, OES_INV);
2821                 if (rc < 0)
2822                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2823
2824                 osc_extent_put(env, waiting);
2825                 waiting = NULL;
2826                 goto again;
2827         }
2828         RETURN(result);
2829 }
2830 EXPORT_SYMBOL(osc_cache_truncate_start);
2831
2832 /**
2833  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2834  */
2835 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext)
2836 {
2837         if (ext != NULL) {
2838                 struct osc_object *obj = ext->oe_obj;
2839                 bool unplug = false;
2840
2841                 EASSERT(ext->oe_nr_pages > 0, ext);
2842                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2843                 EASSERT(!ext->oe_urgent, ext);
2844
2845                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2846                 osc_object_lock(obj);
2847                 osc_extent_state_set(ext, OES_CACHE);
2848                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2849                         ext->oe_urgent = 1;
2850                         list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2851                         unplug = true;
2852                 }
2853                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2854                 osc_object_unlock(obj);
2855                 osc_extent_put(env, ext);
2856
2857                 if (unplug)
2858                         osc_io_unplug_async(env, osc_cli(obj), obj);
2859         }
2860 }
2861
2862 /**
2863  * Wait for extents in a specific range to be written out.
2864  * The caller must have called osc_cache_writeback_range() to issue IO
2865  * otherwise it will take a long time for this function to finish.
2866  *
2867  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2868  * nobody else can dirty this range of file while we're waiting for
2869  * extents to be written.
2870  */
2871 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2872                          pgoff_t start, pgoff_t end)
2873 {
2874         struct osc_extent *ext;
2875         pgoff_t index = start;
2876         int     result = 0;
2877         ENTRY;
2878
2879 again:
2880         osc_object_lock(obj);
2881         ext = osc_extent_search(obj, index);
2882         if (ext == NULL)
2883                 ext = first_extent(obj);
2884         else if (ext->oe_end < index)
2885                 ext = next_extent(ext);
2886         while (ext != NULL) {
2887                 int rc;
2888
2889                 if (ext->oe_start > end)
2890                         break;
2891
2892                 if (!ext->oe_fsync_wait) {
2893                         ext = next_extent(ext);
2894                         continue;
2895                 }
2896
2897                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2898                              ext->oe_hp || ext->oe_urgent), ext);
2899                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2900                              !ext->oe_hp && ext->oe_urgent), ext);
2901
2902                 index = ext->oe_end + 1;
2903                 osc_extent_get(ext);
2904                 osc_object_unlock(obj);
2905
2906                 rc = osc_extent_wait(env, ext, OES_INV);
2907                 if (result == 0)
2908                         result = rc;
2909                 osc_extent_put(env, ext);
2910                 goto again;
2911         }
2912         osc_object_unlock(obj);
2913
2914         OSC_IO_DEBUG(obj, "sync file range.\n");
2915         RETURN(result);
2916 }
2917 EXPORT_SYMBOL(osc_cache_wait_range);
2918
2919 /**
2920  * Called to write out a range of osc object.
2921  *
2922  * @hp     : should be set this is caused by lock cancel;
2923  * @discard: is set if dirty pages should be dropped - file will be deleted or
2924  *         truncated, this implies there is no partially discarding extents.
2925  *
2926  * Return how many pages will be issued, or error code if error occurred.
2927  */
2928 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2929                               pgoff_t start, pgoff_t end, int hp, int discard)
2930 {
2931         struct osc_extent *ext;
2932         LIST_HEAD(discard_list);
2933         bool unplug = false;
2934         int result = 0;
2935         ENTRY;
2936
2937         osc_object_lock(obj);
2938         ext = osc_extent_search(obj, start);
2939         if (ext == NULL)
2940                 ext = first_extent(obj);
2941         else if (ext->oe_end < start)
2942                 ext = next_extent(ext);
2943         while (ext != NULL) {
2944                 if (ext->oe_start > end)
2945                         break;
2946
2947                 ext->oe_fsync_wait = 1;
2948                 switch (ext->oe_state) {
2949                 case OES_CACHE:
2950                         result += ext->oe_nr_pages;
2951                         if (!discard) {
2952                                 struct list_head *list = NULL;
2953                                 if (hp) {
2954                                         EASSERT(!ext->oe_hp, ext);
2955                                         ext->oe_hp = 1;
2956                                         list = &obj->oo_hp_exts;
2957                                 } else if (!ext->oe_urgent && !ext->oe_hp) {
2958                                         ext->oe_urgent = 1;
2959                                         list = &obj->oo_urgent_exts;
2960                                 }
2961                                 if (list != NULL)
2962                                         list_move_tail(&ext->oe_link, list);
2963                                 unplug = true;
2964                         } else {
2965                                 struct client_obd *cli = osc_cli(obj);
2966                                 int pcc_bits = cli->cl_chunkbits - PAGE_SHIFT;
2967                                 pgoff_t align_by = (1 << pcc_bits);
2968                                 pgoff_t a_start = round_down(start, align_by);
2969                                 pgoff_t a_end = round_up(end, align_by);
2970
2971                                 /* overflow case */
2972                                 if (end && !a_end)
2973                                         a_end = CL_PAGE_EOF;
2974                                 /* the only discarder is lock cancelling, so
2975                                  * [start, end], aligned by chunk size, must
2976                                  * contain this extent */
2977                                 LASSERTF(ext->oe_start >= a_start &&
2978                                          ext->oe_end <= a_end,
2979                                          "ext [%lu, %lu] reg [%lu, %lu] "
2980                                          "orig [%lu %lu] align %lu bits "
2981                                          "%d\n", ext->oe_start, ext->oe_end,
2982                                          a_start, a_end, start, end,
2983                                          align_by, pcc_bits);
2984                                 osc_extent_state_set(ext, OES_LOCKING);
2985                                 ext->oe_owner = current;
2986                                 list_move_tail(&ext->oe_link,
2987                                                    &discard_list);
2988                                 osc_update_pending(obj, OBD_BRW_WRITE,
2989                                                    -ext->oe_nr_pages);
2990                         }
2991                         break;
2992                 case OES_ACTIVE:
2993                         /* It's pretty bad to wait for ACTIVE extents, because
2994                          * we don't know how long we will wait for it to be
2995                          * flushed since it may be blocked at awaiting more
2996                          * grants. We do this for the correctness of fsync. */
2997                         LASSERT(hp == 0 && discard == 0);
2998                         ext->oe_urgent = 1;
2999                         break;
3000                 case OES_TRUNC:
3001                         /* this extent is being truncated, can't do anything
3002                          * for it now. it will be set to urgent after truncate
3003                          * is finished in osc_cache_truncate_end(). */
3004                 default:
3005                         break;
3006                 }
3007                 ext = next_extent(ext);
3008         }
3009         osc_object_unlock(obj);
3010
3011         LASSERT(ergo(!discard, list_empty(&discard_list)));
3012         if (!list_empty(&discard_list)) {
3013                 struct osc_extent *tmp;
3014                 int rc;
3015
3016                 osc_list_maint(osc_cli(obj), obj);
3017                 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
3018                         list_del_init(&ext->oe_link);
3019                         EASSERT(ext->oe_state == OES_LOCKING, ext);
3020
3021                         /* Discard caching pages. We don't actually write this
3022                          * extent out but we complete it as if we did. */
3023                         rc = osc_extent_make_ready(env, ext);
3024                         if (unlikely(rc < 0)) {
3025                                 OSC_EXTENT_DUMP(D_ERROR, ext,
3026                                                 "make_ready returned %d\n", rc);
3027                                 if (result >= 0)
3028                                         result = rc;
3029                         }
3030
3031                         /* finish the extent as if the pages were sent */
3032                         osc_extent_finish(env, ext, 0, 0);
3033                 }
3034         }
3035
3036         if (unplug)
3037                 osc_io_unplug(env, osc_cli(obj), obj);
3038
3039         if (hp || discard) {
3040                 int rc;
3041                 rc = osc_cache_wait_range(env, obj, start, end);
3042                 if (result >= 0 && rc < 0)
3043                         result = rc;
3044         }
3045
3046         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3047         RETURN(result);
3048 }
3049 EXPORT_SYMBOL(osc_cache_writeback_range);
3050
3051 /**
3052  * Returns a list of pages by a given [start, end] of \a obj.
3053  *
3054  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3055  * crucial in the face of [offset, EOF] locks.
3056  *
3057  * Return at least one page in @queue unless there is no covered page.
3058  */
3059 bool osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3060                           struct osc_object *osc, pgoff_t start, pgoff_t end,
3061                           osc_page_gang_cbt cb, void *cbdata)
3062 {
3063         struct osc_page *ops;
3064         struct pagevec  *pagevec;
3065         void            **pvec;
3066         pgoff_t         idx;
3067         unsigned int    nr;
3068         unsigned int    i;
3069         unsigned int    j;
3070         bool            res = true;
3071         bool            tree_lock = true;
3072         ENTRY;
3073
3074         idx = start;
3075         pvec = osc_env_info(env)->oti_pvec;
3076         pagevec = &osc_env_info(env)->oti_pagevec;
3077         ll_pagevec_init(pagevec, 0);
3078         spin_lock(&osc->oo_tree_lock);
3079         while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3080                                             idx, OTI_PVEC_SIZE)) > 0) {
3081                 struct cl_page *page;
3082                 bool end_of_region = false;
3083
3084                 for (i = 0, j = 0; i < nr; ++i) {
3085                         ops = pvec[i];
3086                         pvec[i] = NULL;
3087
3088                         idx = osc_index(ops);
3089                         if (idx > end) {
3090                                 end_of_region = true;
3091                                 break;
3092                         }
3093
3094                         page = ops->ops_cl.cpl_page;
3095                         LASSERT(page->cp_type == CPT_CACHEABLE);
3096                         if (page->cp_state == CPS_FREEING)
3097                                 continue;
3098
3099                         cl_page_get(page);
3100                         lu_ref_add_atomic(&page->cp_reference,
3101                                           "gang_lookup", current);
3102                         pvec[j++] = ops;
3103                 }
3104                 ++idx;
3105
3106                 /*
3107                  * Here a delicate locking dance is performed. Current thread
3108                  * holds a reference to a page, but has to own it before it
3109                  * can be placed into queue. Owning implies waiting, so
3110                  * radix-tree lock is to be released. After a wait one has to
3111                  * check that pages weren't truncated (cl_page_own() returns
3112                  * error in the latter case).
3113                  */
3114                 spin_unlock(&osc->oo_tree_lock);
3115                 tree_lock = false;
3116
3117                 res = (*cb)(env, io, pvec, j, cbdata);
3118
3119                 for (i = 0; i < j; ++i) {
3120                         ops = pvec[i];
3121                         page = ops->ops_cl.cpl_page;
3122                         lu_ref_del(&page->cp_reference, "gang_lookup", current);
3123                         cl_pagevec_put(env, page, pagevec);
3124                 }
3125                 pagevec_release(pagevec);
3126
3127                 if (nr < OTI_PVEC_SIZE || end_of_region)
3128                         break;
3129
3130                 if (!res)
3131                         break;
3132
3133                 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SLOW_PAGE_EVICT,
3134                                  cfs_fail_val ?: 20);
3135
3136                 if (io->ci_type == CIT_MISC &&
3137                     io->u.ci_misc.lm_next_rpc_time &&
3138                     ktime_get_seconds() > io->u.ci_misc.lm_next_rpc_time) {
3139                         osc_send_empty_rpc(osc, idx << PAGE_SHIFT);
3140                         io->u.ci_misc.lm_next_rpc_time = ktime_get_seconds() +
3141                                                          5 * obd_timeout / 16;
3142                 }
3143
3144                 if (need_resched())
3145                         cond_resched();
3146
3147                 spin_lock(&osc->oo_tree_lock);
3148                 tree_lock = true;
3149         }
3150         if (tree_lock)
3151                 spin_unlock(&osc->oo_tree_lock);
3152         RETURN(res);
3153 }
3154 EXPORT_SYMBOL(osc_page_gang_lookup);
3155
3156 /**
3157  * Check if page @page is covered by an extra lock or discard it.
3158  */
3159 static bool check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3160                                  void **pvec, int count, void *cbdata)
3161 {
3162         struct osc_thread_info *info = osc_env_info(env);
3163         struct osc_object *osc = cbdata;
3164         int i;
3165
3166         for (i = 0; i < count; i++) {
3167                 struct osc_page *ops = pvec[i];
3168                 struct cl_page *page = ops->ops_cl.cpl_page;
3169                 pgoff_t index = osc_index(ops);
3170                 bool discard = false;
3171
3172                 /* negative lock caching */
3173                 if (index < info->oti_ng_index) {
3174                         discard = true;
3175                 } else if (index >= info->oti_fn_index) {
3176                         struct ldlm_lock *tmp;
3177                         /* refresh non-overlapped index */
3178                         tmp = osc_dlmlock_at_pgoff(env, osc, index,
3179                                         OSC_DAP_FL_TEST_LOCK |
3180                                         OSC_DAP_FL_AST |
3181                                         OSC_DAP_FL_RIGHT);
3182                         if (tmp != NULL) {
3183                                 __u64 end =
3184                                         tmp->l_policy_data.l_extent.end;
3185                                 __u64 start =
3186                                         tmp->l_policy_data.l_extent.start;
3187
3188                                 /* no lock covering this page */
3189                                 if (index < cl_index(osc2cl(osc), start)) {
3190                                         /* no lock at @index,
3191                                          * first lock at @start
3192                                          */
3193                                         info->oti_ng_index =
3194                                                 cl_index(osc2cl(osc), start);
3195                                         discard = true;
3196                                 } else {
3197                                         /* Cache the first-non-overlapped
3198                                          * index so as to skip all pages
3199                                          * within [index, oti_fn_index).
3200                                          * This is safe because if tmp lock
3201                                          * is canceled, it will discard these
3202                                          * pages.
3203                                          */
3204                                         info->oti_fn_index =
3205                                                 cl_index(osc2cl(osc), end + 1);
3206                                         if (end == OBD_OBJECT_EOF)
3207                                                 info->oti_fn_index =
3208                                                         CL_PAGE_EOF;
3209                                 }
3210                                 LDLM_LOCK_PUT(tmp);
3211                         } else {
3212                                 info->oti_ng_index = CL_PAGE_EOF;
3213                                 discard = true;
3214                         }
3215                 }
3216
3217                 if (discard) {
3218                         if (cl_page_own(env, io, page) == 0) {
3219                                 cl_page_discard(env, io, page);
3220                                 cl_page_disown(env, io, page);
3221                         } else {
3222                                 LASSERT(page->cp_state == CPS_FREEING);
3223                         }
3224                 }
3225
3226                 info->oti_next_index = index + 1;
3227         }
3228         return true;
3229 }
3230
3231 bool osc_discard_cb(const struct lu_env *env, struct cl_io *io,
3232                     void **pvec, int count, void *cbdata)
3233 {
3234         struct osc_thread_info *info = osc_env_info(env);
3235         int i;
3236
3237         for (i = 0; i < count; i++) {
3238                 struct osc_page *ops = pvec[i];
3239                 struct cl_page *page = ops->ops_cl.cpl_page;
3240
3241                 /* page is top page. */
3242                 info->oti_next_index = osc_index(ops) + 1;
3243                 if (cl_page_own(env, io, page) == 0) {
3244                         if (!ergo(page->cp_type == CPT_CACHEABLE,
3245                                   !PageDirty(cl_page_vmpage(page))))
3246                                 CL_PAGE_DEBUG(D_ERROR, env, page,
3247                                               "discard dirty page?\n");
3248
3249                         /* discard the page */
3250                         cl_page_discard(env, io, page);
3251                         cl_page_disown(env, io, page);
3252                 } else {
3253                         LASSERT(page->cp_state == CPS_FREEING);
3254                 }
3255         }
3256
3257         return true;
3258 }
3259 EXPORT_SYMBOL(osc_discard_cb);
3260
3261 /**
3262  * Discard pages protected by the given lock. This function traverses radix
3263  * tree to find all covering pages and discard them. If a page is being covered
3264  * by other locks, it should remain in cache.
3265  *
3266  * If error happens on any step, the process continues anyway (the reasoning
3267  * behind this being that lock cancellation cannot be delayed indefinitely).
3268  */
3269 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3270                            pgoff_t start, pgoff_t end, bool discard)
3271 {
3272         struct osc_thread_info *info = osc_env_info(env);
3273         struct cl_io *io = osc_env_thread_io(env);
3274         osc_page_gang_cbt cb;
3275         int result;
3276
3277         ENTRY;
3278
3279         io->ci_obj = cl_object_top(osc2cl(osc));
3280         io->ci_ignore_layout = 1;
3281         io->u.ci_misc.lm_next_rpc_time = ktime_get_seconds() +
3282                                          5 * obd_timeout / 16;
3283         result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3284         if (result != 0)
3285                 GOTO(out, result);
3286
3287         cb = discard ? osc_discard_cb : check_and_discard_cb;
3288         info->oti_fn_index = info->oti_next_index = start;
3289         info->oti_ng_index = 0;
3290
3291         osc_page_gang_lookup(env, io, osc,
3292                              info->oti_next_index, end, cb, osc);
3293 out:
3294         cl_io_fini(env, io);
3295         RETURN(result);
3296 }
3297
3298
3299 /** @} osc */