lustre/osc/osc_cache.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  *
  28  */
  29 /*
  30  * This file is part of Lustre, http://www.lustre.org/
  31  * Lustre is a trademark of Sun Microsystems, Inc.
  32  *
  33  * osc cache management.
  34  *
  35  * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_OSC
  39
  40 #include <lustre_osc.h>
  41 #include <lustre_dlm.h>
  42
  43 #include "osc_internal.h"
  44
  45 static int extent_debug; /* set it to be true for more debug */
  46
  47 static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
  48 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
  49                            enum osc_extent_state state);
  50 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
  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,
  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 <= cl_offset(osc2cl(obj), ext->oe_start) &&
 234                       extent->end   >= cl_offset(osc2cl(obj), ext->oe_max_end)))
 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_PUT(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_async_page *oap;
 836         struct osc_async_page *tmp;
 837         int nr_pages = ext->oe_nr_pages;
 838         int lost_grant = 0;
 839         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
 840         loff_t last_off = 0;
 841         int last_count = -1;
 842         ENTRY;
 843
 844         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
 845
 846         ext->oe_rc = rc ?: ext->oe_nr_pages;
 847         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
 848
 849         osc_lru_add_batch(cli, &ext->oe_pages);
 850         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 851                                      oap_pending_item) {
 852                 list_del_init(&oap->oap_rpc_item);
 853                 list_del_init(&oap->oap_pending_item);
 854                 if (last_off <= oap->oap_obj_off) {
 855                         last_off = oap->oap_obj_off;
 856                         last_count = oap->oap_count;
 857                 }
 858
 859                 --ext->oe_nr_pages;
 860                 osc_ap_completion(env, cli, oap, sent, rc);
 861         }
 862         EASSERT(ext->oe_nr_pages == 0, ext);
 863
 864         if (!sent) {
 865                 lost_grant = ext->oe_grants;
 866         } else if (blocksize < PAGE_SIZE &&
 867                    last_count != PAGE_SIZE) {
 868                 /* For short writes we shouldn't count parts of pages that
 869                  * span a whole chunk on the OST side, or our accounting goes
 870                  * wrong.  Should match the code in filter_grant_check. */
 871                 int offset = last_off & ~PAGE_MASK;
 872                 int count = last_count + (offset & (blocksize - 1));
 873                 int end = (offset + last_count) & (blocksize - 1);
 874                 if (end)
 875                         count += blocksize - end;
 876
 877                 lost_grant = PAGE_SIZE - count;
 878         }
 879         if (ext->oe_grants > 0)
 880                 osc_free_grant(cli, nr_pages, lost_grant, ext->oe_grants);
 881
 882         osc_extent_remove(ext);
 883         /* put the refcount for RPC */
 884         osc_extent_put(env, ext);
 885         RETURN(0);
 886 }
 887
 888 /**
 889  * Wait for the extent's state to become @state.
 890  */
 891 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
 892                            enum osc_extent_state state)
 893 {
 894         struct osc_object *obj = ext->oe_obj;
 895         int rc = 0;
 896         ENTRY;
 897
 898         osc_object_lock(obj);
 899         LASSERT(sanity_check_nolock(ext) == 0);
 900         /* `Kick' this extent only if the caller is waiting for it to be
 901          * written out. */
 902         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
 903                 if (ext->oe_state == OES_ACTIVE) {
 904                         ext->oe_urgent = 1;
 905                 } else if (ext->oe_state == OES_CACHE) {
 906                         ext->oe_urgent = 1;
 907                         osc_extent_hold(ext);
 908                         rc = 1;
 909                 }
 910         }
 911         osc_object_unlock(obj);
 912         if (rc == 1)
 913                 osc_extent_release(env, ext);
 914
 915         /* wait for the extent until its state becomes @state */
 916         rc = wait_event_idle_timeout(ext->oe_waitq,
 917                                      smp_load_acquire(&ext->oe_state) == state,
 918                                      cfs_time_seconds(600));
 919         if (rc == 0) {
 920                 OSC_EXTENT_DUMP(D_ERROR, ext,
 921                         "%s: wait ext to %u timedout, recovery in progress?\n",
 922                         cli_name(osc_cli(obj)), state);
 923
 924                 wait_event_idle(ext->oe_waitq,
 925                                 smp_load_acquire(&ext->oe_state) == state);
 926         }
 927         if (ext->oe_rc < 0)
 928                 rc = ext->oe_rc;
 929         else
 930                 rc = 0;
 931         RETURN(rc);
 932 }
 933
 934 /**
 935  * Discard pages with index greater than @size. If @ext is overlapped with
 936  * @size, then partial truncate happens.
 937  */
 938 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
 939                                 bool partial)
 940 {
 941         struct lu_env         *env;
 942         struct cl_io          *io;
 943         struct osc_object     *obj = ext->oe_obj;
 944         struct client_obd     *cli = osc_cli(obj);
 945         struct osc_async_page *oap;
 946         struct osc_async_page *tmp;
 947         struct pagevec        *pvec;
 948         int                    pages_in_chunk = 0;
 949         int                    ppc_bits    = cli->cl_chunkbits -
 950                                              PAGE_SHIFT;
 951         __u64                  trunc_chunk = trunc_index >> ppc_bits;
 952         int                    grants   = 0;
 953         int                    nr_pages = 0;
 954         int                    rc       = 0;
 955         __u16                  refcheck;
 956         ENTRY;
 957
 958         LASSERT(sanity_check(ext) == 0);
 959         LASSERT(ext->oe_state == OES_TRUNC);
 960         LASSERT(!ext->oe_urgent);
 961
 962         /* Request new lu_env.
 963          * We can't use that env from osc_cache_truncate_start() because
 964          * it's from lov_io_sub and not fully initialized. */
 965         env = cl_env_get(&refcheck);
 966         if (IS_ERR(env))
 967                 RETURN(PTR_ERR(env));
 968
 969         io  = osc_env_thread_io(env);
 970         io->ci_obj = cl_object_top(osc2cl(obj));
 971         io->ci_ignore_layout = 1;
 972         pvec = &osc_env_info(env)->oti_pagevec;
 973         ll_pagevec_init(pvec, 0);
 974         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
 975         if (rc < 0)
 976                 GOTO(out, rc);
 977
 978         /* discard all pages with index greater than trunc_index */
 979         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 980                                      oap_pending_item) {
 981                 pgoff_t index = osc_index(oap2osc(oap));
 982                 struct cl_page  *page = oap2cl_page(oap);
 983
 984                 LASSERT(list_empty(&oap->oap_rpc_item));
 985
 986                 /* only discard the pages with their index greater than
 987                  * trunc_index, and ... */
 988                 if (index < trunc_index ||
 989                     (index == trunc_index && partial)) {
 990                         /* accounting how many pages remaining in the chunk
 991                          * so that we can calculate grants correctly. */
 992                         if (index >> ppc_bits == trunc_chunk)
 993                                 ++pages_in_chunk;
 994                         continue;
 995                 }
 996
 997                 list_del_init(&oap->oap_pending_item);
 998
 999                 cl_page_get(page);
1000                 lu_ref_add(&page->cp_reference, "truncate", current);
1001
1002                 if (cl_page_own(env, io, page) == 0) {
1003                         cl_page_discard(env, io, page);
1004                         cl_page_disown(env, io, page);
1005                 } else {
1006                         LASSERT(page->cp_state == CPS_FREEING);
1007                         LASSERT(0);
1008                 }
1009
1010                 lu_ref_del(&page->cp_reference, "truncate", current);
1011                 cl_pagevec_put(env, page, pvec);
1012
1013                 --ext->oe_nr_pages;
1014                 ++nr_pages;
1015         }
1016         pagevec_release(pvec);
1017
1018         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1019                       ext->oe_nr_pages == 0),
1020                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1021
1022         osc_object_lock(obj);
1023         if (ext->oe_nr_pages == 0) {
1024                 LASSERT(pages_in_chunk == 0);
1025                 grants = ext->oe_grants;
1026                 ext->oe_grants = 0;
1027         } else { /* calculate how many grants we can free */
1028                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1029                 pgoff_t last_index;
1030
1031
1032                 /* if there is no pages in this chunk, we can also free grants
1033                  * for the last chunk */
1034                 if (pages_in_chunk == 0) {
1035                         /* if this is the 1st chunk and no pages in this chunk,
1036                          * ext->oe_nr_pages must be zero, so we should be in
1037                          * the other if-clause. */
1038                         LASSERT(trunc_chunk > 0);
1039                         --trunc_chunk;
1040                         ++chunks;
1041                 }
1042
1043                 /* this is what we can free from this extent */
1044                 grants          = chunks << cli->cl_chunkbits;
1045                 ext->oe_grants -= grants;
1046                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1047                 ext->oe_end     = min(last_index, ext->oe_max_end);
1048                 LASSERT(ext->oe_end >= ext->oe_start);
1049                 LASSERT(ext->oe_grants > 0);
1050         }
1051         osc_object_unlock(obj);
1052
1053         if (grants > 0 || nr_pages > 0)
1054                 osc_free_grant(cli, nr_pages, grants, grants);
1055
1056 out:
1057         cl_io_fini(env, io);
1058         cl_env_put(env, &refcheck);
1059         RETURN(rc);
1060 }
1061
1062 /**
1063  * This function is used to make the extent prepared for transfer.
1064  * A race with flusing page - ll_writepage() has to be handled cautiously.
1065  */
1066 static int osc_extent_make_ready(const struct lu_env *env,
1067                                  struct osc_extent *ext)
1068 {
1069         struct osc_async_page *oap;
1070         struct osc_async_page *last = NULL;
1071         struct osc_object *obj = ext->oe_obj;
1072         unsigned int page_count = 0;
1073         int rc;
1074         ENTRY;
1075
1076         /* we're going to grab page lock, so object lock must not be taken. */
1077         LASSERT(sanity_check(ext) == 0);
1078         /* in locking state, any process should not touch this extent. */
1079         EASSERT(ext->oe_state == OES_LOCKING, ext);
1080         EASSERT(ext->oe_owner != NULL, ext);
1081
1082         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1083
1084         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1085                 ++page_count;
1086                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1087                         last = oap;
1088
1089                 /* checking ASYNC_READY is race safe */
1090                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1091                         continue;
1092
1093                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1094                 switch (rc) {
1095                 case 0:
1096                         spin_lock(&oap->oap_lock);
1097                         oap->oap_async_flags |= ASYNC_READY;
1098                         spin_unlock(&oap->oap_lock);
1099                         break;
1100                 case -EALREADY:
1101                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1102                         break;
1103                 default:
1104                         LASSERTF(0, "unknown return code: %d\n", rc);
1105                 }
1106         }
1107
1108         LASSERT(page_count == ext->oe_nr_pages);
1109         LASSERT(last != NULL);
1110         /* the last page is the only one we need to refresh its count by
1111          * the size of file. */
1112         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1113                 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1114                 LASSERTF(last_oap_count > 0,
1115                          "last_oap_count %d\n", last_oap_count);
1116                 LASSERT(last->oap_page_off + last_oap_count <= PAGE_SIZE);
1117                 last->oap_count = last_oap_count;
1118                 spin_lock(&last->oap_lock);
1119                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1120                 spin_unlock(&last->oap_lock);
1121         }
1122
1123         /* for the rest of pages, we don't need to call osf_refresh_count()
1124          * because it's known they are not the last page */
1125         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1126                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1127                         oap->oap_count = PAGE_SIZE - oap->oap_page_off;
1128                         spin_lock(&oap->oap_lock);
1129                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1130                         spin_unlock(&oap->oap_lock);
1131                 }
1132         }
1133
1134         osc_object_lock(obj);
1135         osc_extent_state_set(ext, OES_RPC);
1136         osc_object_unlock(obj);
1137         /* get a refcount for RPC. */
1138         osc_extent_get(ext);
1139
1140         RETURN(0);
1141 }
1142
1143 /**
1144  * Quick and simple version of osc_extent_find(). This function is frequently
1145  * called to expand the extent for the same IO. To expand the extent, the
1146  * page index must be in the same or next chunk of ext->oe_end.
1147  */
1148 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index,
1149                              unsigned int *grants)
1150 {
1151         struct osc_object *obj = ext->oe_obj;
1152         struct client_obd *cli = osc_cli(obj);
1153         struct osc_extent *next;
1154         int ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1155         pgoff_t chunk = index >> ppc_bits;
1156         pgoff_t end_chunk;
1157         pgoff_t end_index;
1158         unsigned int chunksize = 1 << cli->cl_chunkbits;
1159         int rc = 0;
1160         ENTRY;
1161
1162         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1163         osc_object_lock(obj);
1164         LASSERT(sanity_check_nolock(ext) == 0);
1165         end_chunk = ext->oe_end >> ppc_bits;
1166         if (chunk > end_chunk + 1)
1167                 GOTO(out, rc = -ERANGE);
1168
1169         if (end_chunk >= chunk)
1170                 GOTO(out, rc = 0);
1171
1172         LASSERT(end_chunk + 1 == chunk);
1173
1174         /* try to expand this extent to cover @index */
1175         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1176
1177         /* don't go over the maximum extent size reported by server */
1178         if (end_index - ext->oe_start + 1 > cli->cl_max_extent_pages)
1179                 GOTO(out, rc = -ERANGE);
1180
1181         next = next_extent(ext);
1182         if (next != NULL && next->oe_start <= end_index)
1183                 /* complex mode - overlapped with the next extent,
1184                  * this case will be handled by osc_extent_find() */
1185                 GOTO(out, rc = -EAGAIN);
1186
1187         ext->oe_end = end_index;
1188         ext->oe_grants += chunksize;
1189         LASSERT(*grants >= chunksize);
1190         *grants -= chunksize;
1191         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1192                  "overlapped after expanding for %lu.\n", index);
1193         EXIT;
1194
1195 out:
1196         osc_object_unlock(obj);
1197         RETURN(rc);
1198 }
1199
1200 static void osc_extent_tree_dump0(int mask, struct osc_object *obj,
1201                                   const char *func, int line)
1202 {
1203         struct osc_extent *ext;
1204         int cnt;
1205
1206         if (!cfs_cdebug_show(mask, DEBUG_SUBSYSTEM))
1207                 return;
1208
1209         CDEBUG(mask, "Dump object %p extents at %s:%d, mppr: %u.\n",
1210                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1211
1212         /* osc_object_lock(obj); */
1213         cnt = 1;
1214         for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1215                 OSC_EXTENT_DUMP(mask, ext, "in tree %d.\n", cnt++);
1216
1217         cnt = 1;
1218         list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1219                 OSC_EXTENT_DUMP(mask, ext, "hp %d.\n", cnt++);
1220
1221         cnt = 1;
1222         list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1223                 OSC_EXTENT_DUMP(mask, ext, "urgent %d.\n", cnt++);
1224
1225         cnt = 1;
1226         list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1227                 OSC_EXTENT_DUMP(mask, ext, "reading %d.\n", cnt++);
1228         /* osc_object_unlock(obj); */
1229 }
1230
1231 /* ------------------ osc extent end ------------------ */
1232
1233 static inline int osc_is_ready(struct osc_object *osc)
1234 {
1235         return !list_empty(&osc->oo_ready_item) ||
1236                !list_empty(&osc->oo_hp_ready_item);
1237 }
1238
1239 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1240         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1241                (OSC), osc_is_ready(OSC),                                       \
1242                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1243                list_empty_marker(&(OSC)->oo_ready_item),                       \
1244                atomic_read(&(OSC)->oo_nr_writes),                              \
1245                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1246                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1247                atomic_read(&(OSC)->oo_nr_reads),                               \
1248                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1249                ##args)
1250
1251 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1252                           int cmd)
1253 {
1254         struct osc_page *opg  = oap2osc_page(oap);
1255         struct cl_page  *page = oap2cl_page(oap);
1256         int result;
1257
1258         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1259
1260         ENTRY;
1261         result = cl_page_make_ready(env, page, CRT_WRITE);
1262         if (result == 0)
1263                 opg->ops_submit_time = ktime_get();
1264         RETURN(result);
1265 }
1266
1267 static int osc_refresh_count(const struct lu_env *env,
1268                              struct osc_async_page *oap, int cmd)
1269 {
1270         struct osc_page  *opg = oap2osc_page(oap);
1271         pgoff_t index = osc_index(oap2osc(oap));
1272         struct cl_object *obj;
1273         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
1274         int result;
1275         loff_t kms;
1276
1277         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1278         LASSERT(!(cmd & OBD_BRW_READ));
1279         LASSERT(opg != NULL);
1280         obj = opg->ops_cl.cpl_obj;
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 (cl_offset(obj, index) >= kms)
1289                 /* catch race with truncate */
1290                 return 0;
1291         else if (cl_offset(obj, index + 1) > 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_async_page *oap,
1299                           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         opg->ops_submit_time = ktime_set(0, 0);
1320         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1321
1322         /* statistic */
1323         if (rc == 0 && srvlock) {
1324                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1325                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1326                 size_t bytes = oap->oap_count;
1327
1328                 if (crt == CRT_READ)
1329                         stats->os_lockless_reads += bytes;
1330                 else
1331                         stats->os_lockless_writes += bytes;
1332         }
1333
1334         /*
1335          * This has to be the last operation with the page, as locks are
1336          * released in cl_page_completion() and nothing except for the
1337          * reference counter protects page from concurrent reclaim.
1338          */
1339         lu_ref_del(&page->cp_reference, "transfer", page);
1340
1341         cl_page_completion(env, page, crt, rc);
1342         cl_page_put(env, page);
1343
1344         RETURN(0);
1345 }
1346
1347 #define OSC_DUMP_GRANT(mask, cli, fmt, args...) do {                    \
1348         struct client_obd *__tmp = (cli);                               \
1349         CDEBUG(mask, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu " \
1350                "dropped: %ld avail: %ld, dirty_grant: %ld, "            \
1351                "reserved: %ld, flight: %d } lru {in list: %ld, "        \
1352                "left: %ld, waiters: %d }" fmt "\n",                     \
1353                cli_name(__tmp),                                         \
1354                __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages,        \
1355                atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1356                __tmp->cl_lost_grant, __tmp->cl_avail_grant,             \
1357                __tmp->cl_dirty_grant,                                   \
1358                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,         \
1359                atomic_long_read(&__tmp->cl_lru_in_list),                \
1360                atomic_long_read(&__tmp->cl_lru_busy),                   \
1361                atomic_read(&__tmp->cl_lru_shrinkers), ##args);          \
1362 } while (0)
1363
1364 /* caller must hold loi_list_lock */
1365 static void osc_consume_write_grant(struct client_obd *cli,
1366                                     struct brw_page *pga)
1367 {
1368         assert_spin_locked(&cli->cl_loi_list_lock);
1369         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1370         cli->cl_dirty_pages++;
1371         pga->flag |= OBD_BRW_FROM_GRANT;
1372         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1373                PAGE_SIZE, pga, pga->pg);
1374         osc_update_next_shrink(cli);
1375 }
1376
1377 /* the companion to osc_consume_write_grant, called when a brw has completed.
1378  * must be called with the loi lock held. */
1379 static void osc_release_write_grant(struct client_obd *cli,
1380                                     struct brw_page *pga)
1381 {
1382         ENTRY;
1383
1384         assert_spin_locked(&cli->cl_loi_list_lock);
1385         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1386                 EXIT;
1387                 return;
1388         }
1389
1390         pga->flag &= ~OBD_BRW_FROM_GRANT;
1391         atomic_long_dec(&obd_dirty_pages);
1392         cli->cl_dirty_pages--;
1393         EXIT;
1394 }
1395
1396 /**
1397  * To avoid sleeping with object lock held, it's good for us allocate enough
1398  * grants before entering into critical section.
1399  *
1400  * client_obd_list_lock held by caller
1401  */
1402 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1403 {
1404         int rc = -EDQUOT;
1405
1406         if (cli->cl_avail_grant >= bytes) {
1407                 cli->cl_avail_grant    -= bytes;
1408                 cli->cl_reserved_grant += bytes;
1409                 rc = 0;
1410         }
1411         return rc;
1412 }
1413
1414 static void __osc_unreserve_grant(struct client_obd *cli,
1415                                   unsigned int reserved, unsigned int unused)
1416 {
1417         /* it's quite normal for us to get more grant than reserved.
1418          * Thinking about a case that two extents merged by adding a new
1419          * chunk, we can save one extent tax. If extent tax is greater than
1420          * one chunk, we can save more grant by adding a new chunk */
1421         cli->cl_reserved_grant -= reserved;
1422         if (unused > reserved) {
1423                 cli->cl_avail_grant += reserved;
1424                 cli->cl_lost_grant  += unused - reserved;
1425                 cli->cl_dirty_grant -= unused - reserved;
1426         } else {
1427                 cli->cl_avail_grant += unused;
1428                 cli->cl_dirty_grant += reserved - unused;
1429         }
1430 }
1431
1432 static void osc_unreserve_grant_nolock(struct client_obd *cli,
1433                                        unsigned int reserved,
1434                                        unsigned int unused)
1435 {
1436         __osc_unreserve_grant(cli, reserved, unused);
1437         if (unused > 0)
1438                 osc_wake_cache_waiters(cli);
1439 }
1440
1441 static void osc_unreserve_grant(struct client_obd *cli,
1442                                 unsigned int reserved, unsigned int unused)
1443 {
1444         spin_lock(&cli->cl_loi_list_lock);
1445         osc_unreserve_grant_nolock(cli, reserved, unused);
1446         spin_unlock(&cli->cl_loi_list_lock);
1447 }
1448
1449 /**
1450  * Free grant after IO is finished or canceled.
1451  *
1452  * @lost_grant is used to remember how many grants we have allocated but not
1453  * used, we should return these grants to OST. There're two cases where grants
1454  * can be lost:
1455  * 1. truncate;
1456  * 2. blocksize at OST is less than PAGE_SIZE and a partial page was
1457  *    written. In this case OST may use less chunks to serve this partial
1458  *    write. OSTs don't actually know the page size on the client side. so
1459  *    clients have to calculate lost grant by the blocksize on the OST.
1460  *    See filter_grant_check() for details.
1461  */
1462 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1463                            unsigned int lost_grant, unsigned int dirty_grant)
1464 {
1465         unsigned long grant;
1466
1467         grant = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
1468
1469         spin_lock(&cli->cl_loi_list_lock);
1470         atomic_long_sub(nr_pages, &obd_dirty_pages);
1471         cli->cl_dirty_pages -= nr_pages;
1472         cli->cl_lost_grant += lost_grant;
1473         cli->cl_dirty_grant -= dirty_grant;
1474         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1475                 /* borrow some grant from truncate to avoid the case that
1476                  * truncate uses up all avail grant */
1477                 cli->cl_lost_grant -= grant;
1478                 cli->cl_avail_grant += grant;
1479         }
1480         osc_wake_cache_waiters(cli);
1481         spin_unlock(&cli->cl_loi_list_lock);
1482         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu/%lu\n",
1483                lost_grant, cli->cl_lost_grant,
1484                cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_SHIFT,
1485                cli->cl_dirty_grant);
1486 }
1487
1488 /**
1489  * The companion to osc_enter_cache(), called when @oap is no longer part of
1490  * the dirty accounting due to error.
1491  */
1492 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1493 {
1494         spin_lock(&cli->cl_loi_list_lock);
1495         osc_release_write_grant(cli, &oap->oap_brw_page);
1496         spin_unlock(&cli->cl_loi_list_lock);
1497 }
1498
1499 /**
1500  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1501  * is available.
1502  */
1503 static int osc_enter_cache_try(struct client_obd *cli,
1504                                struct osc_async_page *oap,
1505                                int bytes)
1506 {
1507         int rc;
1508
1509         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1510
1511         rc = osc_reserve_grant(cli, bytes);
1512         if (rc < 0)
1513                 return 0;
1514
1515         if (cli->cl_dirty_pages < cli->cl_dirty_max_pages) {
1516                 if (atomic_long_add_return(1, &obd_dirty_pages) <=
1517                     obd_max_dirty_pages) {
1518                         osc_consume_write_grant(cli, &oap->oap_brw_page);
1519                         rc = 1;
1520                         goto out;
1521                 } else
1522                         atomic_long_dec(&obd_dirty_pages);
1523         }
1524         __osc_unreserve_grant(cli, bytes, bytes);
1525
1526 out:
1527         return rc;
1528 }
1529
1530 /* Following two inlines exist to pass code fragments
1531  * to wait_event_idle_exclusive_timeout_cmd().  Passing
1532  * code fragments as macro args can look confusing, so
1533  * we provide inlines to encapsulate them.
1534  */
1535 static inline void cli_unlock_and_unplug(const struct lu_env *env,
1536                                          struct client_obd *cli,
1537                                          struct osc_async_page *oap)
1538 {
1539         spin_unlock(&cli->cl_loi_list_lock);
1540         osc_io_unplug_async(env, cli, NULL);
1541         CDEBUG(D_CACHE,
1542                "%s: sleeping for cache space for %p\n",
1543                cli_name(cli), oap);
1544 }
1545
1546 static inline void cli_lock_after_unplug(struct client_obd *cli)
1547 {
1548         spin_lock(&cli->cl_loi_list_lock);
1549 }
1550 /**
1551  * The main entry to reserve dirty page accounting. Usually the grant reserved
1552  * in this function will be freed in bulk in osc_free_grant() unless it fails
1553  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1554  *
1555  * The process will be put into sleep if it's already run out of grant.
1556  */
1557 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1558                            struct osc_async_page *oap, int bytes)
1559 {
1560         struct osc_object *osc = oap->oap_obj;
1561         struct lov_oinfo *loi = osc->oo_oinfo;
1562         int rc = -EDQUOT;
1563         int remain;
1564         bool entered = false;
1565         /* We cannot wait for a long time here since we are holding ldlm lock
1566          * across the actual IO. If no requests complete fast (e.g. due to
1567          * overloaded OST that takes a long time to process everything, we'd
1568          * get evicted if we wait for a normal obd_timeout or some such.
1569          * So we try to wait half the time it would take the client to be
1570          * evicted by server which is half obd_timeout when AT is off
1571          * or at least ldlm_enqueue_min with AT on.
1572          * See LU-13131 */
1573         unsigned long timeout = cfs_time_seconds(AT_OFF ? obd_timeout / 2 :
1574                                                           ldlm_enqueue_min / 2);
1575
1576         ENTRY;
1577
1578         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1579
1580         spin_lock(&cli->cl_loi_list_lock);
1581
1582         /* force the caller to try sync io.  this can jump the list
1583          * of queued writes and create a discontiguous rpc stream */
1584         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1585             cli->cl_dirty_max_pages == 0 ||
1586             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) {
1587                 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1588                 GOTO(out, rc = -EDQUOT);
1589         }
1590
1591         /*
1592          * We can wait here for two reasons: too many dirty pages in cache, or
1593          * run out of grants. In both cases we should write dirty pages out.
1594          * Adding a cache waiter will trigger urgent write-out no matter what
1595          * RPC size will be.
1596          * The exiting condition (other than success) is no avail grants
1597          * and no dirty pages caching, that really means there is no space
1598          * on the OST.
1599          */
1600         remain = wait_event_idle_exclusive_timeout_cmd(
1601                 cli->cl_cache_waiters,
1602                 (entered = osc_enter_cache_try(cli, oap, bytes)) ||
1603                 (cli->cl_dirty_pages == 0 && cli->cl_w_in_flight == 0),
1604                 timeout,
1605                 cli_unlock_and_unplug(env, cli, oap),
1606                 cli_lock_after_unplug(cli));
1607
1608         if (entered) {
1609                 if (remain == timeout)
1610                         OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1611                 else
1612                         OSC_DUMP_GRANT(D_CACHE, cli,
1613                                        "finally got grant space\n");
1614                 wake_up(&cli->cl_cache_waiters);
1615                 rc = 0;
1616         } else if (remain == 0) {
1617                 OSC_DUMP_GRANT(D_CACHE, cli,
1618                                "timeout, fall back to sync i/o\n");
1619                 osc_extent_tree_dump(D_CACHE, osc);
1620                 /* fall back to synchronous I/O */
1621         } else {
1622                 OSC_DUMP_GRANT(D_CACHE, cli,
1623                                "no grant space, fall back to sync i/o\n");
1624                 wake_up_all(&cli->cl_cache_waiters);
1625         }
1626         EXIT;
1627 out:
1628         spin_unlock(&cli->cl_loi_list_lock);
1629         RETURN(rc);
1630 }
1631
1632 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1633 {
1634         int hprpc = !!list_empty(&osc->oo_hp_exts);
1635         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1636 }
1637
1638 /* This maintains the lists of pending pages to read/write for a given object
1639  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1640  * to quickly find objects that are ready to send an RPC. */
1641 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1642                          int cmd)
1643 {
1644         int invalid_import = 0;
1645         ENTRY;
1646
1647         /* if we have an invalid import we want to drain the queued pages
1648          * by forcing them through rpcs that immediately fail and complete
1649          * the pages.  recovery relies on this to empty the queued pages
1650          * before canceling the locks and evicting down the llite pages */
1651         if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1652                 invalid_import = 1;
1653
1654         if (cmd & OBD_BRW_WRITE) {
1655                 if (atomic_read(&osc->oo_nr_writes) == 0)
1656                         RETURN(0);
1657                 if (invalid_import) {
1658                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1659                         RETURN(1);
1660                 }
1661                 if (!list_empty(&osc->oo_hp_exts)) {
1662                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1663                         RETURN(1);
1664                 }
1665                 if (!list_empty(&osc->oo_urgent_exts)) {
1666                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1667                         RETURN(1);
1668                 }
1669                 /* trigger a write rpc stream as long as there are dirtiers
1670                  * waiting for space.  as they're waiting, they're not going to
1671                  * create more pages to coalesce with what's waiting..
1672                  */
1673                 if (waitqueue_active(&cli->cl_cache_waiters)) {
1674                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1675                         RETURN(1);
1676                 }
1677                 if (!list_empty(&osc->oo_full_exts)) {
1678                         CDEBUG(D_CACHE, "full extent ready, make an RPC\n");
1679                         RETURN(1);
1680                 }
1681         } else {
1682                 if (atomic_read(&osc->oo_nr_reads) == 0)
1683                         RETURN(0);
1684                 if (invalid_import) {
1685                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1686                         RETURN(1);
1687                 }
1688                 /* all read are urgent. */
1689                 if (!list_empty(&osc->oo_reading_exts))
1690                         RETURN(1);
1691         }
1692
1693         RETURN(0);
1694 }
1695
1696 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1697 {
1698         struct client_obd *cli = osc_cli(obj);
1699         if (cmd & OBD_BRW_WRITE) {
1700                 atomic_add(delta, &obj->oo_nr_writes);
1701                 atomic_add(delta, &cli->cl_pending_w_pages);
1702                 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1703         } else {
1704                 atomic_add(delta, &obj->oo_nr_reads);
1705                 atomic_add(delta, &cli->cl_pending_r_pages);
1706                 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1707         }
1708         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1709 }
1710
1711 static int osc_makes_hprpc(struct osc_object *obj)
1712 {
1713         return !list_empty(&obj->oo_hp_exts);
1714 }
1715
1716 static void on_list(struct list_head *item, struct list_head *list,
1717                     int should_be_on)
1718 {
1719         if (list_empty(item) && should_be_on)
1720                 list_add_tail(item, list);
1721         else if (!list_empty(item) && !should_be_on)
1722                 list_del_init(item);
1723 }
1724
1725 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1726  * can find pages to build into rpcs quickly */
1727 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1728 {
1729         if (osc_makes_hprpc(osc)) {
1730                 /* HP rpc */
1731                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1732                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1733         } else {
1734                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1735                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1736                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1737                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1738         }
1739
1740         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1741                 atomic_read(&osc->oo_nr_writes) > 0);
1742
1743         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1744                 atomic_read(&osc->oo_nr_reads) > 0);
1745
1746         return osc_is_ready(osc);
1747 }
1748
1749 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1750 {
1751         int is_ready;
1752
1753         spin_lock(&cli->cl_loi_list_lock);
1754         is_ready = __osc_list_maint(cli, osc);
1755         spin_unlock(&cli->cl_loi_list_lock);
1756
1757         return is_ready;
1758 }
1759
1760 /* this is trying to propogate async writeback errors back up to the
1761  * application.  As an async write fails we record the error code for later if
1762  * the app does an fsync.  As long as errors persist we force future rpcs to be
1763  * sync so that the app can get a sync error and break the cycle of queueing
1764  * pages for which writeback will fail. */
1765 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1766                            int rc)
1767 {
1768         if (rc) {
1769                 if (!ar->ar_rc)
1770                         ar->ar_rc = rc;
1771
1772                 ar->ar_force_sync = 1;
1773                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1774                 return;
1775
1776         }
1777
1778         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1779                 ar->ar_force_sync = 0;
1780 }
1781
1782 /* this must be called holding the loi list lock to give coverage to exit_cache,
1783  * async_flag maintenance, and oap_request */
1784 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1785                               struct osc_async_page *oap, int sent, int rc)
1786 {
1787         struct osc_object *osc = oap->oap_obj;
1788         struct lov_oinfo  *loi = osc->oo_oinfo;
1789         __u64 xid = 0;
1790
1791         ENTRY;
1792         if (oap->oap_request != NULL) {
1793                 xid = ptlrpc_req_xid(oap->oap_request);
1794                 ptlrpc_req_finished(oap->oap_request);
1795                 oap->oap_request = NULL;
1796         }
1797
1798         /* As the transfer for this page is being done, clear the flags */
1799         spin_lock(&oap->oap_lock);
1800         oap->oap_async_flags = 0;
1801         spin_unlock(&oap->oap_lock);
1802
1803         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1804                 spin_lock(&cli->cl_loi_list_lock);
1805                 osc_process_ar(&cli->cl_ar, xid, rc);
1806                 osc_process_ar(&loi->loi_ar, xid, rc);
1807                 spin_unlock(&cli->cl_loi_list_lock);
1808         }
1809
1810         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1811         if (rc)
1812                 CERROR("completion on oap %p obj %p returns %d.\n",
1813                        oap, osc, rc);
1814
1815         EXIT;
1816 }
1817
1818 struct extent_rpc_data {
1819         struct list_head        *erd_rpc_list;
1820         unsigned int            erd_page_count;
1821         unsigned int            erd_max_pages;
1822         unsigned int            erd_max_chunks;
1823         unsigned int            erd_max_extents;
1824 };
1825
1826 static inline unsigned osc_extent_chunks(const struct osc_extent *ext)
1827 {
1828         struct client_obd *cli = osc_cli(ext->oe_obj);
1829         unsigned ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1830
1831         return (ext->oe_end >> ppc_bits) - (ext->oe_start >> ppc_bits) + 1;
1832 }
1833
1834 static inline bool
1835 can_merge(const struct osc_extent *ext, const struct osc_extent *in_rpc)
1836 {
1837         if (ext->oe_no_merge || in_rpc->oe_no_merge)
1838                 return false;
1839
1840         if (ext->oe_srvlock != in_rpc->oe_srvlock)
1841                 return false;
1842
1843         if (ext->oe_ndelay != in_rpc->oe_ndelay)
1844                 return false;
1845
1846         if (!ext->oe_grants != !in_rpc->oe_grants)
1847                 return false;
1848
1849         if (ext->oe_dio != in_rpc->oe_dio)
1850                 return false;
1851
1852         /* It's possible to have overlap on DIO */
1853         if (in_rpc->oe_dio && overlapped(ext, in_rpc))
1854                 return false;
1855
1856         if (ext->oe_is_rdma_only != in_rpc->oe_is_rdma_only)
1857                 return false;
1858
1859         return true;
1860 }
1861
1862 /**
1863  * Try to add extent to one RPC. We need to think about the following things:
1864  * - # of pages must not be over max_pages_per_rpc
1865  * - extent must be compatible with previous ones
1866  */
1867 static int try_to_add_extent_for_io(struct client_obd *cli,
1868                                     struct osc_extent *ext,
1869                                     struct extent_rpc_data *data)
1870 {
1871         struct osc_extent *tmp;
1872         unsigned int chunk_count;
1873         ENTRY;
1874
1875         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1876                 ext);
1877         OSC_EXTENT_DUMP(D_CACHE, ext, "trying to add this extent\n");
1878
1879         if (data->erd_max_extents == 0)
1880                 RETURN(0);
1881
1882         chunk_count = osc_extent_chunks(ext);
1883         EASSERTF(data->erd_page_count != 0 ||
1884                  chunk_count <= data->erd_max_chunks, ext,
1885                  "The first extent to be fit in a RPC contains %u chunks, "
1886                  "which is over the limit %u.\n", chunk_count,
1887                  data->erd_max_chunks);
1888         if (chunk_count > data->erd_max_chunks)
1889                 RETURN(0);
1890
1891         data->erd_max_pages = max(ext->oe_mppr, data->erd_max_pages);
1892         EASSERTF(data->erd_page_count != 0 ||
1893                 ext->oe_nr_pages <= data->erd_max_pages, ext,
1894                 "The first extent to be fit in a RPC contains %u pages, "
1895                 "which is over the limit %u.\n", ext->oe_nr_pages,
1896                 data->erd_max_pages);
1897         if (data->erd_page_count + ext->oe_nr_pages > data->erd_max_pages)
1898                 RETURN(0);
1899
1900         list_for_each_entry(tmp, data->erd_rpc_list, oe_link) {
1901                 EASSERT(tmp->oe_owner == current, tmp);
1902
1903                 if (!can_merge(ext, tmp))
1904                         RETURN(0);
1905         }
1906
1907         data->erd_max_extents--;
1908         data->erd_max_chunks -= chunk_count;
1909         data->erd_page_count += ext->oe_nr_pages;
1910         list_move_tail(&ext->oe_link, data->erd_rpc_list);
1911         ext->oe_owner = current;
1912         RETURN(1);
1913 }
1914
1915 /**
1916  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1917  * get_write_extent() takes all appropriate extents in atomic.
1918  *
1919  * The following policy is used to collect extents for IO:
1920  * 1. Add as many HP extents as possible;
1921  * 2. Add the first urgent extent in urgent extent list and take it out of
1922  *    urgent list;
1923  * 3. Add subsequent extents of this urgent extent;
1924  * 4. If urgent list is not empty, goto 2;
1925  * 5. Traverse the extent tree from the 1st extent;
1926  * 6. Above steps exit if there is no space in this RPC.
1927  */
1928 static unsigned int get_write_extents(struct osc_object *obj,
1929                                       struct list_head *rpclist)
1930 {
1931         struct client_obd *cli = osc_cli(obj);
1932         struct osc_extent *ext;
1933         struct extent_rpc_data data = {
1934                 .erd_rpc_list   = rpclist,
1935                 .erd_page_count = 0,
1936                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
1937                 .erd_max_chunks = osc_max_write_chunks(cli),
1938                 .erd_max_extents = 256,
1939         };
1940
1941         assert_osc_object_is_locked(obj);
1942         while (!list_empty(&obj->oo_hp_exts)) {
1943                 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
1944                                  oe_link);
1945                 if (!try_to_add_extent_for_io(cli, ext, &data))
1946                         return data.erd_page_count;
1947                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
1948         }
1949         if (data.erd_page_count == data.erd_max_pages)
1950                 return data.erd_page_count;
1951
1952         while (!list_empty(&obj->oo_urgent_exts)) {
1953                 ext = list_entry(obj->oo_urgent_exts.next,
1954                                  struct osc_extent, oe_link);
1955                 if (!try_to_add_extent_for_io(cli, ext, &data))
1956                         return data.erd_page_count;
1957         }
1958         if (data.erd_page_count == data.erd_max_pages)
1959                 return data.erd_page_count;
1960
1961         /* One key difference between full extents and other extents: full
1962          * extents can usually only be added if the rpclist was empty, so if we
1963          * can't add one, we continue on to trying to add normal extents.  This
1964          * is so we don't miss adding extra extents to an RPC containing high
1965          * priority or urgent extents. */
1966         while (!list_empty(&obj->oo_full_exts)) {
1967                 ext = list_entry(obj->oo_full_exts.next,
1968                                  struct osc_extent, oe_link);
1969                 if (!try_to_add_extent_for_io(cli, ext, &data))
1970                         break;
1971         }
1972         if (data.erd_page_count == data.erd_max_pages)
1973                 return data.erd_page_count;
1974
1975         for (ext = first_extent(obj);
1976              ext;
1977              ext = next_extent(ext)) {
1978                 if ((ext->oe_state != OES_CACHE) ||
1979                     /* this extent may be already in current rpclist */
1980                     (!list_empty(&ext->oe_link) && ext->oe_owner))
1981                         continue;
1982
1983                 if (!try_to_add_extent_for_io(cli, ext, &data))
1984                         return data.erd_page_count;
1985         }
1986         return data.erd_page_count;
1987 }
1988
1989 static int
1990 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
1991                    struct osc_object *osc)
1992 __must_hold(osc)
1993 {
1994         LIST_HEAD(rpclist);
1995         struct osc_extent *ext;
1996         struct osc_extent *tmp;
1997         struct osc_extent *first = NULL;
1998         unsigned int page_count = 0;
1999         int srvlock = 0;
2000         int rc = 0;
2001         ENTRY;
2002
2003         assert_osc_object_is_locked(osc);
2004
2005         page_count = get_write_extents(osc, &rpclist);
2006         LASSERT(equi(page_count == 0, list_empty(&rpclist)));
2007
2008         if (list_empty(&rpclist))
2009                 RETURN(0);
2010
2011         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2012
2013         list_for_each_entry(ext, &rpclist, oe_link) {
2014                 LASSERT(ext->oe_state == OES_CACHE ||
2015                         ext->oe_state == OES_LOCK_DONE);
2016                 if (ext->oe_state == OES_CACHE)
2017                         osc_extent_state_set(ext, OES_LOCKING);
2018                 else
2019                         osc_extent_state_set(ext, OES_RPC);
2020         }
2021
2022         /* we're going to grab page lock, so release object lock because
2023          * lock order is page lock -> object lock. */
2024         osc_object_unlock(osc);
2025
2026         list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2027                 if (ext->oe_state == OES_LOCKING) {
2028                         rc = osc_extent_make_ready(env, ext);
2029                         if (unlikely(rc < 0)) {
2030                                 list_del_init(&ext->oe_link);
2031                                 osc_extent_finish(env, ext, 0, rc);
2032                                 continue;
2033                         }
2034                 }
2035                 if (first == NULL) {
2036                         first = ext;
2037                         srvlock = ext->oe_srvlock;
2038                 } else {
2039                         LASSERT(srvlock == ext->oe_srvlock);
2040                 }
2041         }
2042
2043         if (!list_empty(&rpclist)) {
2044                 LASSERT(page_count > 0);
2045                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2046                 LASSERT(list_empty(&rpclist));
2047         }
2048
2049         osc_object_lock(osc);
2050         RETURN(rc);
2051 }
2052
2053 /**
2054  * prepare pages for ASYNC io and put pages in send queue.
2055  *
2056  * \param cmd OBD_BRW_* macroses
2057  * \param lop pending pages
2058  *
2059  * \return zero if no page added to send queue.
2060  * \return 1 if pages successfully added to send queue.
2061  * \return negative on errors.
2062  */
2063 static int
2064 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2065                   struct osc_object *osc)
2066 __must_hold(osc)
2067 {
2068         struct osc_extent *ext;
2069         struct osc_extent *next;
2070         LIST_HEAD(rpclist);
2071         struct extent_rpc_data data = {
2072                 .erd_rpc_list   = &rpclist,
2073                 .erd_page_count = 0,
2074                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
2075                 .erd_max_chunks = UINT_MAX,
2076                 .erd_max_extents = UINT_MAX,
2077         };
2078         int rc = 0;
2079         ENTRY;
2080
2081         assert_osc_object_is_locked(osc);
2082         list_for_each_entry_safe(ext, next, &osc->oo_reading_exts, oe_link) {
2083                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2084                 if (!try_to_add_extent_for_io(cli, ext, &data))
2085                         break;
2086                 osc_extent_state_set(ext, OES_RPC);
2087                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2088         }
2089         LASSERT(data.erd_page_count <= data.erd_max_pages);
2090
2091         osc_update_pending(osc, OBD_BRW_READ, -data.erd_page_count);
2092
2093         if (!list_empty(&rpclist)) {
2094                 osc_object_unlock(osc);
2095
2096                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2097                 LASSERT(list_empty(&rpclist));
2098
2099                 osc_object_lock(osc);
2100         }
2101         RETURN(rc);
2102 }
2103
2104 #define list_to_obj(list, item) ({                                            \
2105         struct list_head *__tmp = (list)->next;                               \
2106         list_del_init(__tmp);                                         \
2107         list_entry(__tmp, struct osc_object, oo_##item);                      \
2108 })
2109
2110 /* This is called by osc_check_rpcs() to find which objects have pages that
2111  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2112 static struct osc_object *osc_next_obj(struct client_obd *cli)
2113 {
2114         ENTRY;
2115
2116         /* First return objects that have blocked locks so that they
2117          * will be flushed quickly and other clients can get the lock,
2118          * then objects which have pages ready to be stuffed into RPCs */
2119         if (!list_empty(&cli->cl_loi_hp_ready_list))
2120                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2121         if (!list_empty(&cli->cl_loi_ready_list))
2122                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2123
2124         /* then if we have cache waiters, return all objects with queued
2125          * writes.  This is especially important when many small files
2126          * have filled up the cache and not been fired into rpcs because
2127          * they don't pass the nr_pending/object threshhold
2128          */
2129         if (waitqueue_active(&cli->cl_cache_waiters) &&
2130             !list_empty(&cli->cl_loi_write_list))
2131                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2132
2133         /* then return all queued objects when we have an invalid import
2134          * so that they get flushed */
2135         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2136                 if (!list_empty(&cli->cl_loi_write_list))
2137                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2138                                            write_item));
2139                 if (!list_empty(&cli->cl_loi_read_list))
2140                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2141                                            read_item));
2142         }
2143         RETURN(NULL);
2144 }
2145
2146 /* called with the loi list lock held */
2147 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2148 __must_hold(&cli->cl_loi_list_lock)
2149 {
2150         struct osc_object *osc;
2151         int rc = 0;
2152         ENTRY;
2153
2154         while ((osc = osc_next_obj(cli)) != NULL) {
2155                 struct cl_object *obj = osc2cl(osc);
2156                 struct lu_ref_link link;
2157
2158                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2159
2160                 /* even if we have reached our max in flight RPCs, we still
2161                  * allow all high-priority RPCs through to prevent their
2162                  * starvation and leading to server evicting us for not
2163                  * writing out pages in a timely manner LU-13131 */
2164                 if (osc_max_rpc_in_flight(cli, osc) &&
2165                     list_empty(&osc->oo_hp_exts)) {
2166                         __osc_list_maint(cli, osc);
2167                         break;
2168                 }
2169
2170                 cl_object_get(obj);
2171                 spin_unlock(&cli->cl_loi_list_lock);
2172                 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2173
2174                 /* attempt some read/write balancing by alternating between
2175                  * reads and writes in an object.  The makes_rpc checks here
2176                  * would be redundant if we were getting read/write work items
2177                  * instead of objects.  we don't want send_oap_rpc to drain a
2178                  * partial read pending queue when we're given this object to
2179                  * do io on writes while there are cache waiters */
2180                 osc_object_lock(osc);
2181                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2182                         rc = osc_send_write_rpc(env, cli, osc);
2183                         if (rc < 0) {
2184                                 CERROR("Write request failed with %d\n", rc);
2185
2186                                 /* osc_send_write_rpc failed, mostly because of
2187                                  * memory pressure.
2188                                  *
2189                                  * It can't break here, because if:
2190                                  *  - a page was submitted by osc_io_submit, so
2191                                  *    page locked;
2192                                  *  - no request in flight
2193                                  *  - no subsequent request
2194                                  * The system will be in live-lock state,
2195                                  * because there is no chance to call
2196                                  * osc_io_unplug() and osc_check_rpcs() any
2197                                  * more. pdflush can't help in this case,
2198                                  * because it might be blocked at grabbing
2199                                  * the page lock as we mentioned.
2200                                  *
2201                                  * Anyway, continue to drain pages. */
2202                                 /* break; */
2203                         }
2204                 }
2205                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2206                         rc = osc_send_read_rpc(env, cli, osc);
2207                         if (rc < 0)
2208                                 CERROR("Read request failed with %d\n", rc);
2209                 }
2210                 osc_object_unlock(osc);
2211
2212                 osc_list_maint(cli, osc);
2213                 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2214                 cl_object_put(env, obj);
2215
2216                 spin_lock(&cli->cl_loi_list_lock);
2217         }
2218 }
2219
2220 int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2221                    struct osc_object *osc, int async)
2222 {
2223         int rc = 0;
2224
2225         if (osc != NULL && osc_list_maint(cli, osc) == 0)
2226                 return 0;
2227
2228         if (!async) {
2229                 spin_lock(&cli->cl_loi_list_lock);
2230                 osc_check_rpcs(env, cli);
2231                 spin_unlock(&cli->cl_loi_list_lock);
2232         } else {
2233                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2234                 LASSERT(cli->cl_writeback_work != NULL);
2235                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2236         }
2237         return rc;
2238 }
2239 EXPORT_SYMBOL(osc_io_unplug0);
2240
2241 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2242                         struct page *page, loff_t offset)
2243 {
2244         struct obd_export     *exp = osc_export(osc);
2245         struct osc_async_page *oap = &ops->ops_oap;
2246         ENTRY;
2247
2248         if (!page)
2249                 return cfs_size_round(sizeof(*oap));
2250
2251         oap->oap_magic = OAP_MAGIC;
2252         oap->oap_cli = &exp->exp_obd->u.cli;
2253         oap->oap_obj = osc;
2254
2255         oap->oap_page = page;
2256         oap->oap_obj_off = offset;
2257         LASSERT(!(offset & ~PAGE_MASK));
2258
2259         INIT_LIST_HEAD(&oap->oap_pending_item);
2260         INIT_LIST_HEAD(&oap->oap_rpc_item);
2261
2262         spin_lock_init(&oap->oap_lock);
2263         CDEBUG(D_INFO, "oap %p page %p obj off %llu\n",
2264                oap, page, oap->oap_obj_off);
2265         RETURN(0);
2266 }
2267 EXPORT_SYMBOL(osc_prep_async_page);
2268
2269 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2270                        struct osc_page *ops, cl_commit_cbt cb)
2271 {
2272         struct osc_io *oio = osc_env_io(env);
2273         struct osc_extent     *ext = NULL;
2274         struct osc_async_page *oap = &ops->ops_oap;
2275         struct client_obd     *cli = oap->oap_cli;
2276         struct osc_object     *osc = oap->oap_obj;
2277         struct pagevec        *pvec = &osc_env_info(env)->oti_pagevec;
2278         pgoff_t index;
2279         unsigned int tmp;
2280         unsigned int grants = 0;
2281         u32    brw_flags = OBD_BRW_ASYNC;
2282         int    cmd = OBD_BRW_WRITE;
2283         int    need_release = 0;
2284         int    rc = 0;
2285         ENTRY;
2286
2287         if (oap->oap_magic != OAP_MAGIC)
2288                 RETURN(-EINVAL);
2289
2290         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2291                 RETURN(-EIO);
2292
2293         if (!list_empty(&oap->oap_pending_item) ||
2294             !list_empty(&oap->oap_rpc_item))
2295                 RETURN(-EBUSY);
2296
2297         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2298         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2299         if (oio->oi_cap_sys_resource || io->ci_noquota) {
2300                 brw_flags |= OBD_BRW_NOQUOTA;
2301                 cmd |= OBD_BRW_NOQUOTA;
2302         }
2303
2304         /* check if the file's owner/group is over quota */
2305         if (!(cmd & OBD_BRW_NOQUOTA)) {
2306                 struct cl_object *obj;
2307                 struct cl_attr   *attr;
2308                 unsigned int qid[LL_MAXQUOTAS];
2309
2310                 obj = cl_object_top(&osc->oo_cl);
2311                 attr = &osc_env_info(env)->oti_attr;
2312
2313                 cl_object_attr_lock(obj);
2314                 rc = cl_object_attr_get(env, obj, attr);
2315                 cl_object_attr_unlock(obj);
2316
2317                 qid[USRQUOTA] = attr->cat_uid;
2318                 qid[GRPQUOTA] = attr->cat_gid;
2319                 qid[PRJQUOTA] = attr->cat_projid;
2320                 if (rc == 0 && osc_quota_chkdq(cli, qid) == -EDQUOT)
2321                         rc = -EDQUOT;
2322                 if (rc)
2323                         RETURN(rc);
2324         }
2325
2326         oap->oap_cmd = cmd;
2327         oap->oap_page_off = ops->ops_from;
2328         oap->oap_count = ops->ops_to - ops->ops_from + 1;
2329         /* No need to hold a lock here,
2330          * since this page is not in any list yet. */
2331         oap->oap_async_flags = 0;
2332         oap->oap_brw_flags = brw_flags;
2333
2334         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2335                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2336
2337         index = osc_index(oap2osc(oap));
2338
2339         /* Add this page into extent by the following steps:
2340          * 1. if there exists an active extent for this IO, mostly this page
2341          *    can be added to the active extent and sometimes we need to
2342          *    expand extent to accomodate this page;
2343          * 2. otherwise, a new extent will be allocated. */
2344
2345         ext = oio->oi_active;
2346         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2347                 /* one chunk plus extent overhead must be enough to write this
2348                  * page */
2349                 grants = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2350                 if (ext->oe_end >= index)
2351                         grants = 0;
2352
2353                 /* it doesn't need any grant to dirty this page */
2354                 spin_lock(&cli->cl_loi_list_lock);
2355                 rc = osc_enter_cache_try(cli, oap, grants);
2356                 if (rc == 0) { /* try failed */
2357                         grants = 0;
2358                         need_release = 1;
2359                 } else if (ext->oe_end < index) {
2360                         tmp = grants;
2361                         /* try to expand this extent */
2362                         rc = osc_extent_expand(ext, index, &tmp);
2363                         if (rc < 0) {
2364                                 need_release = 1;
2365                                 /* don't free reserved grant */
2366                         } else {
2367                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2368                                                 "expanded for %lu.\n", index);
2369                                 osc_unreserve_grant_nolock(cli, grants, tmp);
2370                                 grants = 0;
2371                         }
2372                 }
2373                 spin_unlock(&cli->cl_loi_list_lock);
2374                 rc = 0;
2375         } else if (ext != NULL) {
2376                 /* index is located outside of active extent */
2377                 need_release = 1;
2378         }
2379         if (need_release) {
2380                 osc_extent_release(env, ext);
2381                 oio->oi_active = NULL;
2382                 ext = NULL;
2383         }
2384
2385         if (ext == NULL) {
2386                 tmp = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2387
2388                 /* try to find new extent to cover this page */
2389                 LASSERT(oio->oi_active == NULL);
2390                 /* we may have allocated grant for this page if we failed
2391                  * to expand the previous active extent. */
2392                 LASSERT(ergo(grants > 0, grants >= tmp));
2393
2394                 rc = 0;
2395                 if (grants == 0) {
2396                         /* We haven't allocated grant for this page, and we
2397                          * must not hold a page lock while we do enter_cache,
2398                          * so we must mark dirty & unlock any pages in the
2399                          * write commit pagevec. */
2400                         if (pagevec_count(pvec)) {
2401                                 cb(env, io, pvec);
2402                                 pagevec_reinit(pvec);
2403                         }
2404                         rc = osc_enter_cache(env, cli, oap, tmp);
2405                         if (rc == 0)
2406                                 grants = tmp;
2407                 }
2408
2409                 tmp = grants;
2410                 if (rc == 0) {
2411                         ext = osc_extent_find(env, osc, index, &tmp);
2412                         if (IS_ERR(ext)) {
2413                                 LASSERT(tmp == grants);
2414                                 osc_exit_cache(cli, oap);
2415                                 rc = PTR_ERR(ext);
2416                                 ext = NULL;
2417                         } else {
2418                                 oio->oi_active = ext;
2419                         }
2420                 }
2421                 if (grants > 0)
2422                         osc_unreserve_grant(cli, grants, tmp);
2423         }
2424
2425         LASSERT(ergo(rc == 0, ext != NULL));
2426         if (ext != NULL) {
2427                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2428                          ext, "index = %lu.\n", index);
2429                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2430
2431                 osc_object_lock(osc);
2432                 if (ext->oe_nr_pages == 0)
2433                         ext->oe_srvlock = ops->ops_srvlock;
2434                 else
2435                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2436                 ++ext->oe_nr_pages;
2437                 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2438                 osc_object_unlock(osc);
2439
2440                 if (!ext->oe_layout_version)
2441                         ext->oe_layout_version = io->ci_layout_version;
2442         }
2443
2444         RETURN(rc);
2445 }
2446
2447 int osc_teardown_async_page(const struct lu_env *env,
2448                             struct osc_object *obj, struct osc_page *ops)
2449 {
2450         struct osc_async_page *oap = &ops->ops_oap;
2451         int rc = 0;
2452         ENTRY;
2453
2454         LASSERT(oap->oap_magic == OAP_MAGIC);
2455
2456         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2457                oap, ops, osc_index(oap2osc(oap)));
2458
2459         if (!list_empty(&oap->oap_rpc_item)) {
2460                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2461                 rc = -EBUSY;
2462         } else if (!list_empty(&oap->oap_pending_item)) {
2463                 struct osc_extent *ext = NULL;
2464
2465                 osc_object_lock(obj);
2466                 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2467                 osc_object_unlock(obj);
2468                 /* only truncated pages are allowed to be taken out.
2469                  * See osc_extent_truncate() and osc_cache_truncate_start()
2470                  * for details. */
2471                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2472                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2473                                         osc_index(oap2osc(oap)));
2474                         rc = -EBUSY;
2475                 }
2476                 if (ext != NULL)
2477                         osc_extent_put(env, ext);
2478         }
2479         RETURN(rc);
2480 }
2481
2482 /**
2483  * This is called when a page is picked up by kernel to write out.
2484  *
2485  * We should find out the corresponding extent and add the whole extent
2486  * into urgent list. The extent may be being truncated or used, handle it
2487  * carefully.
2488  */
2489 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2490                          struct osc_page *ops)
2491 {
2492         struct osc_extent *ext   = NULL;
2493         struct osc_object *obj   = cl2osc(ops->ops_cl.cpl_obj);
2494         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2495         pgoff_t            index = osc_index(ops);
2496         struct osc_async_page *oap = &ops->ops_oap;
2497         bool unplug = false;
2498         int rc = 0;
2499         ENTRY;
2500
2501         osc_object_lock(obj);
2502         ext = osc_extent_lookup(obj, index);
2503         if (ext == NULL) {
2504                 osc_extent_tree_dump(D_ERROR, obj);
2505                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2506         }
2507
2508         switch (ext->oe_state) {
2509         case OES_RPC:
2510         case OES_LOCK_DONE:
2511                 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2512                 LASSERT(0);
2513                 break;
2514         case OES_LOCKING:
2515                 /* If we know this extent is being written out, we should abort
2516                  * so that the writer can make this page ready. Otherwise, there
2517                  * exists a deadlock problem because other process can wait for
2518                  * page writeback bit holding page lock; and meanwhile in
2519                  * vvp_page_make_ready(), we need to grab page lock before
2520                  * really sending the RPC. */
2521         case OES_TRUNC:
2522                 /* race with truncate, page will be redirtied */
2523         case OES_ACTIVE:
2524                 /* The extent is active so we need to abort and let the caller
2525                  * re-dirty the page. If we continued on here, and we were the
2526                  * one making the extent active, we could deadlock waiting for
2527                  * the page writeback to clear but it won't because the extent
2528                  * is active and won't be written out. */
2529                 GOTO(out, rc = -EAGAIN);
2530         default:
2531                 break;
2532         }
2533
2534         rc = cl_page_prep(env, io, cp, CRT_WRITE);
2535         if (rc)
2536                 GOTO(out, rc);
2537
2538         spin_lock(&oap->oap_lock);
2539         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2540         spin_unlock(&oap->oap_lock);
2541
2542         if (current->flags & PF_MEMALLOC)
2543                 ext->oe_memalloc = 1;
2544
2545         ext->oe_urgent = 1;
2546         if (ext->oe_state == OES_CACHE) {
2547                 OSC_EXTENT_DUMP(D_CACHE, ext,
2548                                 "flush page %p make it urgent.\n", oap);
2549                 if (list_empty(&ext->oe_link))
2550                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2551                 unplug = true;
2552         }
2553         rc = 0;
2554         EXIT;
2555
2556 out:
2557         osc_object_unlock(obj);
2558         osc_extent_put(env, ext);
2559         if (unplug)
2560                 osc_io_unplug_async(env, osc_cli(obj), obj);
2561         return rc;
2562 }
2563
2564 int osc_queue_sync_pages(const struct lu_env *env, const struct cl_io *io,
2565                          struct osc_object *obj, struct list_head *list,
2566                          int brw_flags)
2567 {
2568         struct client_obd     *cli = osc_cli(obj);
2569         struct osc_extent     *ext;
2570         struct osc_async_page *oap;
2571         int     page_count = 0;
2572         int     mppr       = cli->cl_max_pages_per_rpc;
2573         bool    can_merge   = true;
2574         pgoff_t start      = CL_PAGE_EOF;
2575         pgoff_t end        = 0;
2576         ENTRY;
2577
2578         list_for_each_entry(oap, list, oap_pending_item) {
2579                 struct osc_page *opg = oap2osc_page(oap);
2580                 pgoff_t index = osc_index(opg);
2581
2582                 if (index > end)
2583                         end = index;
2584                 if (index < start)
2585                         start = index;
2586                 ++page_count;
2587                 mppr <<= (page_count > mppr);
2588
2589                 if (unlikely(opg->ops_from > 0 ||
2590                              opg->ops_to < PAGE_SIZE - 1))
2591                         can_merge = false;
2592         }
2593
2594         ext = osc_extent_alloc(obj);
2595         if (ext == NULL) {
2596                 struct osc_async_page *tmp;
2597
2598                 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2599                         list_del_init(&oap->oap_pending_item);
2600                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2601                 }
2602                 RETURN(-ENOMEM);
2603         }
2604
2605         ext->oe_rw = !!(brw_flags & OBD_BRW_READ);
2606         ext->oe_sync = 1;
2607         ext->oe_no_merge = !can_merge;
2608         ext->oe_urgent = 1;
2609         ext->oe_start = start;
2610         ext->oe_end = ext->oe_max_end = end;
2611         ext->oe_obj = obj;
2612         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2613         ext->oe_ndelay = !!(brw_flags & OBD_BRW_NDELAY);
2614         ext->oe_dio = !!(brw_flags & OBD_BRW_NOCACHE);
2615         if (ext->oe_dio && !ext->oe_rw) { /* direct io write */
2616                 int grants;
2617                 int ppc;
2618
2619                 ppc = 1 << (cli->cl_chunkbits - PAGE_SHIFT);
2620                 grants = cli->cl_grant_extent_tax;
2621                 grants += (1 << cli->cl_chunkbits) *
2622                         ((page_count + ppc - 1) / ppc);
2623
2624                 spin_lock(&cli->cl_loi_list_lock);
2625                 if (osc_reserve_grant(cli, grants) == 0) {
2626                         list_for_each_entry(oap, list, oap_pending_item) {
2627                                 osc_consume_write_grant(cli,
2628                                                         &oap->oap_brw_page);
2629                                 atomic_long_inc(&obd_dirty_pages);
2630                         }
2631                         osc_unreserve_grant_nolock(cli, grants, 0);
2632                         ext->oe_grants = grants;
2633                 }
2634                 spin_unlock(&cli->cl_loi_list_lock);
2635         }
2636
2637         ext->oe_is_rdma_only = !!(brw_flags & OBD_BRW_RDMA_ONLY);
2638         ext->oe_nr_pages = page_count;
2639         ext->oe_mppr = mppr;
2640         list_splice_init(list, &ext->oe_pages);
2641         ext->oe_layout_version = io->ci_layout_version;
2642
2643         osc_object_lock(obj);
2644         /* Reuse the initial refcount for RPC, don't drop it */
2645         osc_extent_state_set(ext, OES_LOCK_DONE);
2646         if (!ext->oe_rw) { /* write */
2647                 if (!ext->oe_srvlock && !ext->oe_dio) {
2648                         /* The most likely case here is from lack of grants
2649                          * so we are either out of quota or out of space.
2650                          * Since this means we are holding locks across
2651                          * potentially multi-striped IO, we must send out
2652                          * everything out instantly to avoid prolonged
2653                          * waits resulting in lock eviction (likely since
2654                          * the extended wait in osc_cache_enter() did not
2655                          * yield any additional grant due to a timeout.
2656                          * LU-13131 */
2657                         ext->oe_hp = 1;
2658                         list_add_tail(&ext->oe_link, &obj->oo_hp_exts);
2659                 } else {
2660                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2661                 }
2662                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2663         } else {
2664                 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2665                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2666         }
2667         osc_object_unlock(obj);
2668
2669         osc_io_unplug_async(env, cli, obj);
2670         RETURN(0);
2671 }
2672
2673 /**
2674  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2675  */
2676 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
2677                              __u64 size, struct osc_extent **extp)
2678 {
2679         struct client_obd *cli = osc_cli(obj);
2680         struct osc_extent *ext;
2681         struct osc_extent *waiting = NULL;
2682         pgoff_t index;
2683         LIST_HEAD(list);
2684         int result = 0;
2685         bool partial;
2686         ENTRY;
2687
2688         /* pages with index greater or equal to index will be truncated. */
2689         index = cl_index(osc2cl(obj), size);
2690         partial = size > cl_offset(osc2cl(obj), index);
2691
2692 again:
2693         osc_object_lock(obj);
2694         ext = osc_extent_search(obj, index);
2695         if (ext == NULL)
2696                 ext = first_extent(obj);
2697         else if (ext->oe_end < index)
2698                 ext = next_extent(ext);
2699         while (ext != NULL) {
2700                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2701
2702                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2703                         /* if ext is in urgent state, it means there must exist
2704                          * a page already having been flushed by write_page().
2705                          * We have to wait for this extent because we can't
2706                          * truncate that page. */
2707                         OSC_EXTENT_DUMP(D_CACHE, ext,
2708                                         "waiting for busy extent\n");
2709                         waiting = osc_extent_get(ext);
2710                         break;
2711                 }
2712
2713                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:%llu.\n", size);
2714
2715                 osc_extent_get(ext);
2716                 if (ext->oe_state == OES_ACTIVE) {
2717                         /* though we grab inode mutex for write path, but we
2718                          * release it before releasing extent(in osc_io_end()),
2719                          * so there is a race window that an extent is still
2720                          * in OES_ACTIVE when truncate starts. */
2721                         LASSERT(!ext->oe_trunc_pending);
2722                         ext->oe_trunc_pending = 1;
2723                 } else {
2724                         EASSERT(ext->oe_state == OES_CACHE, ext);
2725                         osc_extent_state_set(ext, OES_TRUNC);
2726                         osc_update_pending(obj, OBD_BRW_WRITE,
2727                                            -ext->oe_nr_pages);
2728                 }
2729                 /* This extent could be on the full extents list, that's OK */
2730                 EASSERT(!ext->oe_hp && !ext->oe_urgent, ext);
2731                 if (!list_empty(&ext->oe_link))
2732                         list_move_tail(&ext->oe_link, &list);
2733                 else
2734                         list_add_tail(&ext->oe_link, &list);
2735
2736                 ext = next_extent(ext);
2737         }
2738         osc_object_unlock(obj);
2739
2740         osc_list_maint(cli, obj);
2741
2742         while (!list_empty(&list)) {
2743                 int rc;
2744
2745                 ext = list_entry(list.next, struct osc_extent, oe_link);
2746                 list_del_init(&ext->oe_link);
2747
2748                 /* extent may be in OES_ACTIVE state because inode mutex
2749                  * is released before osc_io_end() in file write case */
2750                 if (ext->oe_state != OES_TRUNC)
2751                         osc_extent_wait(env, ext, OES_TRUNC);
2752
2753                 rc = osc_extent_truncate(ext, index, partial);
2754                 if (rc < 0) {
2755                         if (result == 0)
2756                                 result = rc;
2757
2758                         OSC_EXTENT_DUMP(D_ERROR, ext,
2759                                         "truncate error %d\n", rc);
2760                 } else if (ext->oe_nr_pages == 0) {
2761                         osc_extent_remove(ext);
2762                 } else {
2763                         /* this must be an overlapped extent which means only
2764                          * part of pages in this extent have been truncated.
2765                          */
2766                         EASSERTF(ext->oe_start <= index, ext,
2767                                  "trunc index = %lu/%d.\n", index, partial);
2768                         /* fix index to skip this partially truncated extent */
2769                         index = ext->oe_end + 1;
2770                         partial = false;
2771
2772                         /* we need to hold this extent in OES_TRUNC state so
2773                          * that no writeback will happen. This is to avoid
2774                          * BUG 17397.
2775                          * Only partial truncate can reach here, if @size is
2776                          * not zero, the caller should provide a valid @extp. */
2777                         LASSERT(*extp == NULL);
2778                         *extp = osc_extent_get(ext);
2779                         OSC_EXTENT_DUMP(D_CACHE, ext,
2780                                         "trunc at %llu\n", size);
2781                 }
2782                 osc_extent_put(env, ext);
2783         }
2784         if (waiting != NULL) {
2785                 int rc;
2786
2787                 /* ignore the result of osc_extent_wait the write initiator
2788                  * should take care of it. */
2789                 rc = osc_extent_wait(env, waiting, OES_INV);
2790                 if (rc < 0)
2791                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2792
2793                 osc_extent_put(env, waiting);
2794                 waiting = NULL;
2795                 goto again;
2796         }
2797         RETURN(result);
2798 }
2799 EXPORT_SYMBOL(osc_cache_truncate_start);
2800
2801 /**
2802  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2803  */
2804 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext)
2805 {
2806         if (ext != NULL) {
2807                 struct osc_object *obj = ext->oe_obj;
2808                 bool unplug = false;
2809
2810                 EASSERT(ext->oe_nr_pages > 0, ext);
2811                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2812                 EASSERT(!ext->oe_urgent, ext);
2813
2814                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2815                 osc_object_lock(obj);
2816                 osc_extent_state_set(ext, OES_CACHE);
2817                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2818                         ext->oe_urgent = 1;
2819                         list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2820                         unplug = true;
2821                 }
2822                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2823                 osc_object_unlock(obj);
2824                 osc_extent_put(env, ext);
2825
2826                 if (unplug)
2827                         osc_io_unplug_async(env, osc_cli(obj), obj);
2828         }
2829 }
2830
2831 /**
2832  * Wait for extents in a specific range to be written out.
2833  * The caller must have called osc_cache_writeback_range() to issue IO
2834  * otherwise it will take a long time for this function to finish.
2835  *
2836  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2837  * nobody else can dirty this range of file while we're waiting for
2838  * extents to be written.
2839  */
2840 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2841                          pgoff_t start, pgoff_t end)
2842 {
2843         struct osc_extent *ext;
2844         pgoff_t index = start;
2845         int     result = 0;
2846         ENTRY;
2847
2848 again:
2849         osc_object_lock(obj);
2850         ext = osc_extent_search(obj, index);
2851         if (ext == NULL)
2852                 ext = first_extent(obj);
2853         else if (ext->oe_end < index)
2854                 ext = next_extent(ext);
2855         while (ext != NULL) {
2856                 int rc;
2857
2858                 if (ext->oe_start > end)
2859                         break;
2860
2861                 if (!ext->oe_fsync_wait) {
2862                         ext = next_extent(ext);
2863                         continue;
2864                 }
2865
2866                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2867                              ext->oe_hp || ext->oe_urgent), ext);
2868                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2869                              !ext->oe_hp && ext->oe_urgent), ext);
2870
2871                 index = ext->oe_end + 1;
2872                 osc_extent_get(ext);
2873                 osc_object_unlock(obj);
2874
2875                 rc = osc_extent_wait(env, ext, OES_INV);
2876                 if (result == 0)
2877                         result = rc;
2878                 osc_extent_put(env, ext);
2879                 goto again;
2880         }
2881         osc_object_unlock(obj);
2882
2883         OSC_IO_DEBUG(obj, "sync file range.\n");
2884         RETURN(result);
2885 }
2886 EXPORT_SYMBOL(osc_cache_wait_range);
2887
2888 /**
2889  * Called to write out a range of osc object.
2890  *
2891  * @hp     : should be set this is caused by lock cancel;
2892  * @discard: is set if dirty pages should be dropped - file will be deleted or
2893  *         truncated, this implies there is no partially discarding extents.
2894  *
2895  * Return how many pages will be issued, or error code if error occurred.
2896  */
2897 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2898                               pgoff_t start, pgoff_t end, int hp, int discard)
2899 {
2900         struct osc_extent *ext;
2901         LIST_HEAD(discard_list);
2902         bool unplug = false;
2903         int result = 0;
2904         ENTRY;
2905
2906         osc_object_lock(obj);
2907         ext = osc_extent_search(obj, start);
2908         if (ext == NULL)
2909                 ext = first_extent(obj);
2910         else if (ext->oe_end < start)
2911                 ext = next_extent(ext);
2912         while (ext != NULL) {
2913                 if (ext->oe_start > end)
2914                         break;
2915
2916                 ext->oe_fsync_wait = 1;
2917                 switch (ext->oe_state) {
2918                 case OES_CACHE:
2919                         result += ext->oe_nr_pages;
2920                         if (!discard) {
2921                                 struct list_head *list = NULL;
2922                                 if (hp) {
2923                                         EASSERT(!ext->oe_hp, ext);
2924                                         ext->oe_hp = 1;
2925                                         list = &obj->oo_hp_exts;
2926                                 } else if (!ext->oe_urgent && !ext->oe_hp) {
2927                                         ext->oe_urgent = 1;
2928                                         list = &obj->oo_urgent_exts;
2929                                 }
2930                                 if (list != NULL)
2931                                         list_move_tail(&ext->oe_link, list);
2932                                 unplug = true;
2933                         } else {
2934                                 struct client_obd *cli = osc_cli(obj);
2935                                 int pcc_bits = cli->cl_chunkbits - PAGE_SHIFT;
2936                                 pgoff_t align_by = (1 << pcc_bits);
2937                                 pgoff_t a_start = round_down(start, align_by);
2938                                 pgoff_t a_end = round_up(end, align_by);
2939
2940                                 /* overflow case */
2941                                 if (end && !a_end)
2942                                         a_end = CL_PAGE_EOF;
2943                                 /* the only discarder is lock cancelling, so
2944                                  * [start, end], aligned by chunk size, must
2945                                  * contain this extent */
2946                                 LASSERTF(ext->oe_start >= a_start &&
2947                                          ext->oe_end <= a_end,
2948                                          "ext [%lu, %lu] reg [%lu, %lu] "
2949                                          "orig [%lu %lu] align %lu bits "
2950                                          "%d\n", ext->oe_start, ext->oe_end,
2951                                          a_start, a_end, start, end,
2952                                          align_by, pcc_bits);
2953                                 osc_extent_state_set(ext, OES_LOCKING);
2954                                 ext->oe_owner = current;
2955                                 list_move_tail(&ext->oe_link,
2956                                                    &discard_list);
2957                                 osc_update_pending(obj, OBD_BRW_WRITE,
2958                                                    -ext->oe_nr_pages);
2959                         }
2960                         break;
2961                 case OES_ACTIVE:
2962                         /* It's pretty bad to wait for ACTIVE extents, because
2963                          * we don't know how long we will wait for it to be
2964                          * flushed since it may be blocked at awaiting more
2965                          * grants. We do this for the correctness of fsync. */
2966                         LASSERT(hp == 0 && discard == 0);
2967                         ext->oe_urgent = 1;
2968                         break;
2969                 case OES_TRUNC:
2970                         /* this extent is being truncated, can't do anything
2971                          * for it now. it will be set to urgent after truncate
2972                          * is finished in osc_cache_truncate_end(). */
2973                 default:
2974                         break;
2975                 }
2976                 ext = next_extent(ext);
2977         }
2978         osc_object_unlock(obj);
2979
2980         LASSERT(ergo(!discard, list_empty(&discard_list)));
2981         if (!list_empty(&discard_list)) {
2982                 struct osc_extent *tmp;
2983                 int rc;
2984
2985                 osc_list_maint(osc_cli(obj), obj);
2986                 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
2987                         list_del_init(&ext->oe_link);
2988                         EASSERT(ext->oe_state == OES_LOCKING, ext);
2989
2990                         /* Discard caching pages. We don't actually write this
2991                          * extent out but we complete it as if we did. */
2992                         rc = osc_extent_make_ready(env, ext);
2993                         if (unlikely(rc < 0)) {
2994                                 OSC_EXTENT_DUMP(D_ERROR, ext,
2995                                                 "make_ready returned %d\n", rc);
2996                                 if (result >= 0)
2997                                         result = rc;
2998                         }
2999
3000                         /* finish the extent as if the pages were sent */
3001                         osc_extent_finish(env, ext, 0, 0);
3002                 }
3003         }
3004
3005         if (unplug)
3006                 osc_io_unplug(env, osc_cli(obj), obj);
3007
3008         if (hp || discard) {
3009                 int rc;
3010                 rc = osc_cache_wait_range(env, obj, start, end);
3011                 if (result >= 0 && rc < 0)
3012                         result = rc;
3013         }
3014
3015         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3016         RETURN(result);
3017 }
3018 EXPORT_SYMBOL(osc_cache_writeback_range);
3019
3020 /**
3021  * Returns a list of pages by a given [start, end] of \a obj.
3022  *
3023  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3024  * crucial in the face of [offset, EOF] locks.
3025  *
3026  * Return at least one page in @queue unless there is no covered page.
3027  */
3028 bool osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3029                           struct osc_object *osc, pgoff_t start, pgoff_t end,
3030                           osc_page_gang_cbt cb, void *cbdata)
3031 {
3032         struct osc_page *ops;
3033         struct pagevec  *pagevec;
3034         void            **pvec;
3035         pgoff_t         idx;
3036         unsigned int    nr;
3037         unsigned int    i;
3038         unsigned int    j;
3039         bool            res = true;
3040         bool            tree_lock = true;
3041         ENTRY;
3042
3043         idx = start;
3044         pvec = osc_env_info(env)->oti_pvec;
3045         pagevec = &osc_env_info(env)->oti_pagevec;
3046         ll_pagevec_init(pagevec, 0);
3047         spin_lock(&osc->oo_tree_lock);
3048         while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3049                                             idx, OTI_PVEC_SIZE)) > 0) {
3050                 struct cl_page *page;
3051                 bool end_of_region = false;
3052
3053                 for (i = 0, j = 0; i < nr; ++i) {
3054                         ops = pvec[i];
3055                         pvec[i] = NULL;
3056
3057                         idx = osc_index(ops);
3058                         if (idx > end) {
3059                                 end_of_region = true;
3060                                 break;
3061                         }
3062
3063                         page = ops->ops_cl.cpl_page;
3064                         LASSERT(page->cp_type == CPT_CACHEABLE);
3065                         if (page->cp_state == CPS_FREEING)
3066                                 continue;
3067
3068                         cl_page_get(page);
3069                         lu_ref_add_atomic(&page->cp_reference,
3070                                           "gang_lookup", current);
3071                         pvec[j++] = ops;
3072                 }
3073                 ++idx;
3074
3075                 /*
3076                  * Here a delicate locking dance is performed. Current thread
3077                  * holds a reference to a page, but has to own it before it
3078                  * can be placed into queue. Owning implies waiting, so
3079                  * radix-tree lock is to be released. After a wait one has to
3080                  * check that pages weren't truncated (cl_page_own() returns
3081                  * error in the latter case).
3082                  */
3083                 spin_unlock(&osc->oo_tree_lock);
3084                 tree_lock = false;
3085
3086                 for (i = 0; i < j; ++i) {
3087                         ops = pvec[i];
3088                         if (res)
3089                                 res = (*cb)(env, io, ops, cbdata);
3090
3091                         page = ops->ops_cl.cpl_page;
3092                         lu_ref_del(&page->cp_reference, "gang_lookup", current);
3093                         cl_pagevec_put(env, page, pagevec);
3094                 }
3095                 pagevec_release(pagevec);
3096
3097                 if (nr < OTI_PVEC_SIZE || end_of_region)
3098                         break;
3099
3100                 if (!res)
3101                         break;
3102                 if (need_resched())
3103                         cond_resched();
3104
3105                 spin_lock(&osc->oo_tree_lock);
3106                 tree_lock = true;
3107         }
3108         if (tree_lock)
3109                 spin_unlock(&osc->oo_tree_lock);
3110         RETURN(res);
3111 }
3112 EXPORT_SYMBOL(osc_page_gang_lookup);
3113
3114 /**
3115  * Check if page @page is covered by an extra lock or discard it.
3116  */
3117 static bool check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3118                                 struct osc_page *ops, void *cbdata)
3119 {
3120         struct osc_thread_info *info = osc_env_info(env);
3121         struct osc_object *osc = cbdata;
3122         struct cl_page *page = ops->ops_cl.cpl_page;
3123         pgoff_t index;
3124         bool discard = false;
3125
3126         index = osc_index(ops);
3127
3128         /* negative lock caching */
3129         if (index < info->oti_ng_index) {
3130                 discard = true;
3131         } else if (index >= info->oti_fn_index) {
3132                 struct ldlm_lock *tmp;
3133                 /* refresh non-overlapped index */
3134                 tmp = osc_dlmlock_at_pgoff(env, osc, index,
3135                                            OSC_DAP_FL_TEST_LOCK |
3136                                            OSC_DAP_FL_AST | OSC_DAP_FL_RIGHT);
3137                 if (tmp != NULL) {
3138                         __u64 end = tmp->l_policy_data.l_extent.end;
3139                         __u64 start = tmp->l_policy_data.l_extent.start;
3140
3141                         /* no lock covering this page */
3142                         if (index < cl_index(osc2cl(osc), start)) {
3143                                 /* no lock at @index, first lock at @start */
3144                                 info->oti_ng_index = cl_index(osc2cl(osc),
3145                                                      start);
3146                                 discard = true;
3147                         } else {
3148                                 /* Cache the first-non-overlapped index so as to
3149                                  * skip all pages within [index, oti_fn_index).
3150                                  * This is safe because if tmp lock is canceled,
3151                                  * it will discard these pages.
3152                                  */
3153                                 info->oti_fn_index = cl_index(osc2cl(osc),
3154                                                      end + 1);
3155                                 if (end == OBD_OBJECT_EOF)
3156                                         info->oti_fn_index = CL_PAGE_EOF;
3157                         }
3158                         LDLM_LOCK_PUT(tmp);
3159                 } else {
3160                         info->oti_ng_index = CL_PAGE_EOF;
3161                         discard = true;
3162                 }
3163         }
3164
3165         if (discard) {
3166                 if (cl_page_own(env, io, page) == 0) {
3167                         cl_page_discard(env, io, page);
3168                         cl_page_disown(env, io, page);
3169                 } else {
3170                         LASSERT(page->cp_state == CPS_FREEING);
3171                 }
3172         }
3173
3174         info->oti_next_index = index + 1;
3175         return true;
3176 }
3177
3178 bool osc_discard_cb(const struct lu_env *env, struct cl_io *io,
3179                    struct osc_page *ops, void *cbdata)
3180 {
3181         struct osc_thread_info *info = osc_env_info(env);
3182         struct cl_page *page = ops->ops_cl.cpl_page;
3183
3184         /* page is top page. */
3185         info->oti_next_index = osc_index(ops) + 1;
3186         if (cl_page_own(env, io, page) == 0) {
3187                 if (!ergo(page->cp_type == CPT_CACHEABLE,
3188                           !PageDirty(cl_page_vmpage(page))))
3189                         CL_PAGE_DEBUG(D_ERROR, env, page,
3190                                         "discard dirty page?\n");
3191
3192                 /* discard the page */
3193                 cl_page_discard(env, io, page);
3194                 cl_page_disown(env, io, page);
3195         } else {
3196                 LASSERT(page->cp_state == CPS_FREEING);
3197         }
3198
3199         return true;
3200 }
3201 EXPORT_SYMBOL(osc_discard_cb);
3202
3203 /**
3204  * Discard pages protected by the given lock. This function traverses radix
3205  * tree to find all covering pages and discard them. If a page is being covered
3206  * by other locks, it should remain in cache.
3207  *
3208  * If error happens on any step, the process continues anyway (the reasoning
3209  * behind this being that lock cancellation cannot be delayed indefinitely).
3210  */
3211 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3212                            pgoff_t start, pgoff_t end, bool discard)
3213 {
3214         struct osc_thread_info *info = osc_env_info(env);
3215         struct cl_io *io = osc_env_thread_io(env);
3216         osc_page_gang_cbt cb;
3217         int result;
3218
3219         ENTRY;
3220
3221         io->ci_obj = cl_object_top(osc2cl(osc));
3222         io->ci_ignore_layout = 1;
3223         result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3224         if (result != 0)
3225                 GOTO(out, result);
3226
3227         cb = discard ? osc_discard_cb : check_and_discard_cb;
3228         info->oti_fn_index = info->oti_next_index = start;
3229         info->oti_ng_index = 0;
3230
3231         osc_page_gang_lookup(env, io, osc,
3232                              info->oti_next_index, end, cb, osc);
3233 out:
3234         cl_io_fini(env, io);
3235         RETURN(result);
3236 }
3237
3238
3239 /** @} osc */