lustre/osc/osc_cache.c

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