lustre/osc/osc_cache.c

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