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 *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_CACHE || victim->oe_fsync_wait)
 535                 return -EBUSY;
 536
 537         if (cur->oe_max_end != victim->oe_max_end)
 538                 return -ERANGE;
 539
 540         LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
 541         ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_SHIFT;
 542         chunk_start = cur->oe_start >> ppc_bits;
 543         chunk_end   = cur->oe_end   >> ppc_bits;
 544         if (chunk_start   != (victim->oe_end >> ppc_bits) + 1 &&
 545             chunk_end + 1 != victim->oe_start >> ppc_bits)
 546                 return -ERANGE;
 547
 548         /* overall extent size should not exceed the max supported limit
 549          * reported by the server */
 550         if (cur->oe_end - cur->oe_start + 1 +
 551             victim->oe_end - victim->oe_start + 1 > cli->cl_max_extent_pages)
 552                 return -ERANGE;
 553
 554         OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
 555
 556         cur->oe_start     = min(cur->oe_start, victim->oe_start);
 557         cur->oe_end       = max(cur->oe_end,   victim->oe_end);
 558         /* per-extent tax should be accounted only once for the whole extent */
 559         cur->oe_grants   += victim->oe_grants - cli->cl_grant_extent_tax;
 560         cur->oe_nr_pages += victim->oe_nr_pages;
 561         /* only the following bits are needed to merge */
 562         cur->oe_urgent   |= victim->oe_urgent;
 563         cur->oe_memalloc |= victim->oe_memalloc;
 564         list_splice_init(&victim->oe_pages, &cur->oe_pages);
 565         list_del_init(&victim->oe_link);
 566         victim->oe_nr_pages = 0;
 567
 568         osc_extent_get(victim);
 569         __osc_extent_remove(victim);
 570         osc_extent_put(env, victim);
 571
 572         OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
 573         return 0;
 574 }
 575
 576 /**
 577  * Drop user count of osc_extent, and unplug IO asynchronously.
 578  */
 579 void osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
 580 {
 581         struct osc_object *obj = ext->oe_obj;
 582         struct client_obd *cli = osc_cli(obj);
 583         ENTRY;
 584
 585         LASSERT(atomic_read(&ext->oe_users) > 0);
 586         LASSERT(sanity_check(ext) == 0);
 587         LASSERT(ext->oe_grants > 0);
 588
 589         if (atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
 590                 LASSERT(ext->oe_state == OES_ACTIVE);
 591                 if (ext->oe_trunc_pending) {
 592                         /* a truncate process is waiting for this extent.
 593                          * This may happen due to a race, check
 594                          * osc_cache_truncate_start(). */
 595                         osc_extent_state_set(ext, OES_TRUNC);
 596                         ext->oe_trunc_pending = 0;
 597                         osc_object_unlock(obj);
 598                 } else {
 599                         int grant = 0;
 600
 601                         osc_extent_state_set(ext, OES_CACHE);
 602                         osc_update_pending(obj, OBD_BRW_WRITE,
 603                                            ext->oe_nr_pages);
 604
 605                         /* try to merge the previous and next extent. */
 606                         if (osc_extent_merge(env, ext, prev_extent(ext)) == 0)
 607                                 grant += cli->cl_grant_extent_tax;
 608                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 609                                 grant += cli->cl_grant_extent_tax;
 610
 611                         if (ext->oe_hp)
 612                                 list_move_tail(&ext->oe_link,
 613                                                &obj->oo_hp_exts);
 614                         else if (ext->oe_urgent)
 615                                 list_move_tail(&ext->oe_link,
 616                                                &obj->oo_urgent_exts);
 617                         else if (ext->oe_nr_pages == ext->oe_mppr) {
 618                                 list_move_tail(&ext->oe_link,
 619                                                &obj->oo_full_exts);
 620                         }
 621                         osc_object_unlock(obj);
 622                         if (grant > 0)
 623                                 osc_unreserve_grant(cli, 0, grant);
 624                 }
 625
 626                 osc_io_unplug_async(env, cli, obj);
 627         }
 628         osc_extent_put(env, ext);
 629
 630         RETURN_EXIT;
 631 }
 632
 633 /**
 634  * Find or create an extent which includes @index, core function to manage
 635  * extent tree.
 636  */
 637 static struct osc_extent *osc_extent_find(const struct lu_env *env,
 638                                           struct osc_object *obj, pgoff_t index,
 639                                           unsigned int *grants)
 640 {
 641         struct client_obd *cli = osc_cli(obj);
 642         struct osc_lock   *olck;
 643         struct cl_lock_descr *descr;
 644         struct osc_extent *cur;
 645         struct osc_extent *ext;
 646         struct osc_extent *conflict = NULL;
 647         struct osc_extent *found = NULL;
 648         pgoff_t    chunk;
 649         pgoff_t    max_end;
 650         unsigned int max_pages; /* max_pages_per_rpc */
 651         unsigned int chunksize;
 652         int        ppc_bits; /* pages per chunk bits */
 653         pgoff_t    chunk_mask;
 654         int        rc;
 655         ENTRY;
 656
 657         cur = osc_extent_alloc(obj);
 658         if (cur == NULL)
 659                 RETURN(ERR_PTR(-ENOMEM));
 660
 661         olck = osc_env_io(env)->oi_write_osclock;
 662         LASSERTF(olck != NULL, "page %lu is not covered by lock\n", index);
 663         LASSERT(olck->ols_state == OLS_GRANTED);
 664
 665         descr = &olck->ols_cl.cls_lock->cll_descr;
 666         LASSERT(descr->cld_mode >= CLM_WRITE);
 667
 668         LASSERTF(cli->cl_chunkbits >= PAGE_SHIFT,
 669                  "chunkbits: %u\n", cli->cl_chunkbits);
 670         ppc_bits   = cli->cl_chunkbits - PAGE_SHIFT;
 671         chunk_mask = ~((1 << ppc_bits) - 1);
 672         chunksize  = 1 << cli->cl_chunkbits;
 673         chunk      = index >> ppc_bits;
 674
 675         /* align end to RPC edge. */
 676         max_pages = cli->cl_max_pages_per_rpc;
 677         if ((max_pages & ~chunk_mask) != 0) {
 678                 CERROR("max_pages: %#x chunkbits: %u chunk_mask: %#lx\n",
 679                        max_pages, cli->cl_chunkbits, chunk_mask);
 680                 RETURN(ERR_PTR(-EINVAL));
 681         }
 682         max_end = index - (index % max_pages) + max_pages - 1;
 683         max_end = min_t(pgoff_t, max_end, descr->cld_end);
 684
 685         /* initialize new extent by parameters so far */
 686         cur->oe_max_end = max_end;
 687         cur->oe_start   = index & chunk_mask;
 688         cur->oe_end     = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
 689         if (cur->oe_start < descr->cld_start)
 690                 cur->oe_start = descr->cld_start;
 691         if (cur->oe_end > max_end)
 692                 cur->oe_end = max_end;
 693         cur->oe_grants  = 0;
 694         cur->oe_mppr    = max_pages;
 695         if (olck->ols_dlmlock != NULL) {
 696                 LASSERT(olck->ols_hold);
 697                 cur->oe_dlmlock = LDLM_LOCK_GET(olck->ols_dlmlock);
 698                 lu_ref_add(&olck->ols_dlmlock->l_reference, "osc_extent", cur);
 699         }
 700
 701         /* grants has been allocated by caller */
 702         LASSERTF(*grants >= chunksize + cli->cl_grant_extent_tax,
 703                  "%u/%u/%u.\n", *grants, chunksize, cli->cl_grant_extent_tax);
 704         LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR"\n",
 705                  EXTPARA(cur));
 706
 707 restart:
 708         osc_object_lock(obj);
 709         ext = osc_extent_search(obj, cur->oe_start);
 710         if (!ext)
 711                 ext = first_extent(obj);
 712         for (; ext; ext = next_extent(ext)) {
 713                 pgoff_t ext_chk_start = ext->oe_start >> ppc_bits;
 714                 pgoff_t ext_chk_end = ext->oe_end >> ppc_bits;
 715
 716                 LASSERT(sanity_check_nolock(ext) == 0);
 717                 if (chunk > ext_chk_end + 1 || chunk < ext_chk_start)
 718                         break;
 719
 720                 /* if covering by different locks, no chance to match */
 721                 if (olck->ols_dlmlock != ext->oe_dlmlock) {
 722                         EASSERTF(!overlapped(ext, cur), ext,
 723                                  EXTSTR"\n", EXTPARA(cur));
 724
 725                         continue;
 726                 }
 727
 728                 /* discontiguous chunks? */
 729                 if (chunk + 1 < ext_chk_start)
 730                         continue;
 731
 732                 /* ok, from now on, ext and cur have these attrs:
 733                  * 1. covered by the same lock
 734                  * 2. contiguous at chunk level or overlapping. */
 735
 736                 if (overlapped(ext, cur)) {
 737                         /* cur is the minimum unit, so overlapping means
 738                          * full contain. */
 739                         EASSERTF((ext->oe_start <= cur->oe_start &&
 740                                   ext->oe_end >= cur->oe_end),
 741                                  ext, EXTSTR"\n", EXTPARA(cur));
 742
 743                         if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
 744                                 /* for simplicity, we wait for this extent to
 745                                  * finish before going forward. */
 746                                 conflict = osc_extent_get(ext);
 747                                 break;
 748                         }
 749
 750                         found = osc_extent_hold(ext);
 751                         break;
 752                 }
 753
 754                 /* non-overlapped extent */
 755                 if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait)
 756                         /* we can't do anything for a non OES_CACHE extent, or
 757                          * if there is someone waiting for this extent to be
 758                          * flushed, try next one. */
 759                         continue;
 760
 761                 /* check if they belong to the same rpc slot before trying to
 762                  * merge. the extents are not overlapped and contiguous at
 763                  * chunk level to get here. */
 764                 if (ext->oe_max_end != max_end)
 765                         /* if they don't belong to the same RPC slot or
 766                          * max_pages_per_rpc has ever changed, do not merge. */
 767                         continue;
 768
 769                 /* check whether maximum extent size will be hit */
 770                 if ((ext_chk_end - ext_chk_start + 1 + 1) << ppc_bits >
 771                     cli->cl_max_extent_pages)
 772                         continue;
 773
 774                 /* it's required that an extent must be contiguous at chunk
 775                  * level so that we know the whole extent is covered by grant
 776                  * (the pages in the extent are NOT required to be contiguous).
 777                  * Otherwise, it will be too much difficult to know which
 778                  * chunks have grants allocated. */
 779
 780                 /* try to do front merge - extend ext's start */
 781                 if (chunk + 1 == ext_chk_start) {
 782                         /* ext must be chunk size aligned */
 783                         EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
 784
 785                         /* pull ext's start back to cover cur */
 786                         ext->oe_start   = cur->oe_start;
 787                         ext->oe_grants += chunksize;
 788                         LASSERT(*grants >= chunksize);
 789                         *grants -= chunksize;
 790
 791                         found = osc_extent_hold(ext);
 792                 } else if (chunk == ext_chk_end + 1) {
 793                         /* rear merge */
 794                         ext->oe_end     = cur->oe_end;
 795                         ext->oe_grants += chunksize;
 796                         LASSERT(*grants >= chunksize);
 797                         *grants -= chunksize;
 798
 799                         /* try to merge with the next one because we just fill
 800                          * in a gap */
 801                         if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 802                                 /* we can save extent tax from next extent */
 803                                 *grants += cli->cl_grant_extent_tax;
 804
 805                         found = osc_extent_hold(ext);
 806                 }
 807                 if (found != NULL)
 808                         break;
 809         }
 810
 811         osc_extent_tree_dump(D_CACHE, obj);
 812         if (found != NULL) {
 813                 LASSERT(conflict == NULL);
 814                 if (!IS_ERR(found)) {
 815                         LASSERT(found->oe_dlmlock == cur->oe_dlmlock);
 816                         OSC_EXTENT_DUMP(D_CACHE, found,
 817                                         "found caching ext for %lu.\n", index);
 818                 }
 819         } else if (conflict == NULL) {
 820                 /* create a new extent */
 821                 EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
 822                 cur->oe_grants = chunksize + cli->cl_grant_extent_tax;
 823                 LASSERT(*grants >= cur->oe_grants);
 824                 *grants -= cur->oe_grants;
 825
 826                 cur->oe_state = OES_CACHE;
 827                 found = osc_extent_hold(cur);
 828                 osc_extent_insert(obj, cur);
 829                 OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
 830                                 index, descr->cld_end);
 831         }
 832         osc_object_unlock(obj);
 833
 834         if (conflict != NULL) {
 835                 LASSERT(found == NULL);
 836
 837                 /* waiting for IO to finish. Please notice that it's impossible
 838                  * to be an OES_TRUNC extent. */
 839                 rc = osc_extent_wait(env, conflict, OES_INV);
 840                 osc_extent_put(env, conflict);
 841                 conflict = NULL;
 842                 if (rc < 0)
 843                         GOTO(out, found = ERR_PTR(rc));
 844
 845                 goto restart;
 846         }
 847         EXIT;
 848
 849 out:
 850         osc_extent_put(env, cur);
 851         return found;
 852 }
 853
 854 /**
 855  * Called when IO is finished to an extent.
 856  */
 857 int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
 858                       int sent, int rc)
 859 {
 860         struct client_obd *cli = osc_cli(ext->oe_obj);
 861         struct osc_async_page *oap;
 862         struct osc_async_page *tmp;
 863         int nr_pages = ext->oe_nr_pages;
 864         int lost_grant = 0;
 865         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
 866         loff_t last_off = 0;
 867         int last_count = -1;
 868         ENTRY;
 869
 870         OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
 871
 872         ext->oe_rc = rc ?: ext->oe_nr_pages;
 873         EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
 874
 875         osc_lru_add_batch(cli, &ext->oe_pages);
 876         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
 877                                      oap_pending_item) {
 878                 list_del_init(&oap->oap_rpc_item);
 879                 list_del_init(&oap->oap_pending_item);
 880                 if (last_off <= oap->oap_obj_off) {
 881                         last_off = oap->oap_obj_off;
 882                         last_count = oap->oap_count;
 883                 }
 884
 885                 --ext->oe_nr_pages;
 886                 osc_ap_completion(env, cli, oap, sent, rc);
 887         }
 888         EASSERT(ext->oe_nr_pages == 0, ext);
 889
 890         if (!sent) {
 891                 lost_grant = ext->oe_grants;
 892         } else if (blocksize < PAGE_SIZE &&
 893                    last_count != PAGE_SIZE) {
 894                 /* For short writes we shouldn't count parts of pages that
 895                  * span a whole chunk on the OST side, or our accounting goes
 896                  * wrong.  Should match the code in filter_grant_check. */
 897                 int offset = last_off & ~PAGE_MASK;
 898                 int count = last_count + (offset & (blocksize - 1));
 899                 int end = (offset + last_count) & (blocksize - 1);
 900                 if (end)
 901                         count += blocksize - end;
 902
 903                 lost_grant = PAGE_SIZE - count;
 904         }
 905         if (ext->oe_grants > 0)
 906                 osc_free_grant(cli, nr_pages, lost_grant, ext->oe_grants);
 907
 908         osc_extent_remove(ext);
 909         /* put the refcount for RPC */
 910         osc_extent_put(env, ext);
 911         RETURN(0);
 912 }
 913
 914 /**
 915  * Wait for the extent's state to become @state.
 916  */
 917 static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
 918                            enum osc_extent_state state)
 919 {
 920         struct osc_object *obj = ext->oe_obj;
 921         int rc = 0;
 922         ENTRY;
 923
 924         osc_object_lock(obj);
 925         LASSERT(sanity_check_nolock(ext) == 0);
 926         /* `Kick' this extent only if the caller is waiting for it to be
 927          * written out. */
 928         if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
 929                 if (ext->oe_state == OES_ACTIVE) {
 930                         ext->oe_urgent = 1;
 931                 } else if (ext->oe_state == OES_CACHE) {
 932                         ext->oe_urgent = 1;
 933                         osc_extent_hold(ext);
 934                         rc = 1;
 935                 }
 936         }
 937         osc_object_unlock(obj);
 938         if (rc == 1)
 939                 osc_extent_release(env, ext);
 940
 941         /* wait for the extent until its state becomes @state */
 942         rc = wait_event_idle_timeout(ext->oe_waitq,
 943                                      smp_load_acquire(&ext->oe_state) == state,
 944                                      cfs_time_seconds(600));
 945         if (rc == 0) {
 946                 OSC_EXTENT_DUMP(D_ERROR, ext,
 947                         "%s: wait ext to %u timedout, recovery in progress?\n",
 948                         cli_name(osc_cli(obj)), state);
 949
 950                 wait_event_idle(ext->oe_waitq,
 951                                 smp_load_acquire(&ext->oe_state) == state);
 952         }
 953         if (ext->oe_rc < 0)
 954                 rc = ext->oe_rc;
 955         else
 956                 rc = 0;
 957         RETURN(rc);
 958 }
 959
 960 /**
 961  * Discard pages with index greater than @size. If @ext is overlapped with
 962  * @size, then partial truncate happens.
 963  */
 964 static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
 965                                 bool partial)
 966 {
 967         struct lu_env         *env;
 968         struct cl_io          *io;
 969         struct osc_object     *obj = ext->oe_obj;
 970         struct client_obd     *cli = osc_cli(obj);
 971         struct osc_async_page *oap;
 972         struct osc_async_page *tmp;
 973         struct pagevec        *pvec;
 974         int                    pages_in_chunk = 0;
 975         int                    ppc_bits    = cli->cl_chunkbits -
 976                                              PAGE_SHIFT;
 977         __u64                  trunc_chunk = trunc_index >> ppc_bits;
 978         int                    grants   = 0;
 979         int                    nr_pages = 0;
 980         int                    rc       = 0;
 981         __u16                  refcheck;
 982         ENTRY;
 983
 984         LASSERT(sanity_check(ext) == 0);
 985         LASSERT(ext->oe_state == OES_TRUNC);
 986         LASSERT(!ext->oe_urgent);
 987
 988         /* Request new lu_env.
 989          * We can't use that env from osc_cache_truncate_start() because
 990          * it's from lov_io_sub and not fully initialized. */
 991         env = cl_env_get(&refcheck);
 992         if (IS_ERR(env))
 993                 RETURN(PTR_ERR(env));
 994
 995         io  = osc_env_thread_io(env);
 996         io->ci_obj = cl_object_top(osc2cl(obj));
 997         io->ci_ignore_layout = 1;
 998         pvec = &osc_env_info(env)->oti_pagevec;
 999         ll_pagevec_init(pvec, 0);
1000         rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
1001         if (rc < 0)
1002                 GOTO(out, rc);
1003
1004         /* discard all pages with index greater than trunc_index */
1005         list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
1006                                      oap_pending_item) {
1007                 pgoff_t index = osc_index(oap2osc(oap));
1008                 struct cl_page  *page = oap2cl_page(oap);
1009
1010                 LASSERT(list_empty(&oap->oap_rpc_item));
1011
1012                 /* only discard the pages with their index greater than
1013                  * trunc_index, and ... */
1014                 if (index < trunc_index ||
1015                     (index == trunc_index && partial)) {
1016                         /* accounting how many pages remaining in the chunk
1017                          * so that we can calculate grants correctly. */
1018                         if (index >> ppc_bits == trunc_chunk)
1019                                 ++pages_in_chunk;
1020                         continue;
1021                 }
1022
1023                 list_del_init(&oap->oap_pending_item);
1024
1025                 cl_page_get(page);
1026                 lu_ref_add(&page->cp_reference, "truncate", current);
1027
1028                 if (cl_page_own(env, io, page) == 0) {
1029                         cl_page_discard(env, io, page);
1030                         cl_page_disown(env, io, page);
1031                 } else {
1032                         LASSERT(page->cp_state == CPS_FREEING);
1033                         LASSERT(0);
1034                 }
1035
1036                 lu_ref_del(&page->cp_reference, "truncate", current);
1037                 cl_pagevec_put(env, page, pvec);
1038
1039                 --ext->oe_nr_pages;
1040                 ++nr_pages;
1041         }
1042         pagevec_release(pvec);
1043
1044         EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
1045                       ext->oe_nr_pages == 0),
1046                 ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
1047
1048         osc_object_lock(obj);
1049         if (ext->oe_nr_pages == 0) {
1050                 LASSERT(pages_in_chunk == 0);
1051                 grants = ext->oe_grants;
1052                 ext->oe_grants = 0;
1053         } else { /* calculate how many grants we can free */
1054                 int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
1055                 pgoff_t last_index;
1056
1057
1058                 /* if there is no pages in this chunk, we can also free grants
1059                  * for the last chunk */
1060                 if (pages_in_chunk == 0) {
1061                         /* if this is the 1st chunk and no pages in this chunk,
1062                          * ext->oe_nr_pages must be zero, so we should be in
1063                          * the other if-clause. */
1064                         LASSERT(trunc_chunk > 0);
1065                         --trunc_chunk;
1066                         ++chunks;
1067                 }
1068
1069                 /* this is what we can free from this extent */
1070                 grants          = chunks << cli->cl_chunkbits;
1071                 ext->oe_grants -= grants;
1072                 last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
1073                 ext->oe_end     = min(last_index, ext->oe_max_end);
1074                 LASSERT(ext->oe_end >= ext->oe_start);
1075                 LASSERT(ext->oe_grants > 0);
1076         }
1077         osc_object_unlock(obj);
1078
1079         if (grants > 0 || nr_pages > 0)
1080                 osc_free_grant(cli, nr_pages, grants, grants);
1081
1082 out:
1083         cl_io_fini(env, io);
1084         cl_env_put(env, &refcheck);
1085         RETURN(rc);
1086 }
1087
1088 /**
1089  * This function is used to make the extent prepared for transfer.
1090  * A race with flusing page - ll_writepage() has to be handled cautiously.
1091  */
1092 static int osc_extent_make_ready(const struct lu_env *env,
1093                                  struct osc_extent *ext)
1094 {
1095         struct osc_async_page *oap;
1096         struct osc_async_page *last = NULL;
1097         struct osc_object *obj = ext->oe_obj;
1098         unsigned int page_count = 0;
1099         int rc;
1100         ENTRY;
1101
1102         /* we're going to grab page lock, so object lock must not be taken. */
1103         LASSERT(sanity_check(ext) == 0);
1104         /* in locking state, any process should not touch this extent. */
1105         EASSERT(ext->oe_state == OES_LOCKING, ext);
1106         EASSERT(ext->oe_owner != NULL, ext);
1107
1108         OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
1109
1110         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1111                 ++page_count;
1112                 if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
1113                         last = oap;
1114
1115                 /* checking ASYNC_READY is race safe */
1116                 if ((oap->oap_async_flags & ASYNC_READY) != 0)
1117                         continue;
1118
1119                 rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
1120                 switch (rc) {
1121                 case 0:
1122                         spin_lock(&oap->oap_lock);
1123                         oap->oap_async_flags |= ASYNC_READY;
1124                         spin_unlock(&oap->oap_lock);
1125                         break;
1126                 case -EALREADY:
1127                         LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
1128                         break;
1129                 default:
1130                         LASSERTF(0, "unknown return code: %d\n", rc);
1131                 }
1132         }
1133
1134         LASSERT(page_count == ext->oe_nr_pages);
1135         LASSERT(last != NULL);
1136         /* the last page is the only one we need to refresh its count by
1137          * the size of file. */
1138         if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
1139                 int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
1140                 LASSERTF(last_oap_count > 0,
1141                          "last_oap_count %d\n", last_oap_count);
1142                 LASSERT(last->oap_page_off + last_oap_count <= PAGE_SIZE);
1143                 last->oap_count = last_oap_count;
1144                 spin_lock(&last->oap_lock);
1145                 last->oap_async_flags |= ASYNC_COUNT_STABLE;
1146                 spin_unlock(&last->oap_lock);
1147         }
1148
1149         /* for the rest of pages, we don't need to call osf_refresh_count()
1150          * because it's known they are not the last page */
1151         list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1152                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
1153                         oap->oap_count = PAGE_SIZE - oap->oap_page_off;
1154                         spin_lock(&oap->oap_lock);
1155                         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1156                         spin_unlock(&oap->oap_lock);
1157                 }
1158         }
1159
1160         osc_object_lock(obj);
1161         osc_extent_state_set(ext, OES_RPC);
1162         osc_object_unlock(obj);
1163         /* get a refcount for RPC. */
1164         osc_extent_get(ext);
1165
1166         RETURN(0);
1167 }
1168
1169 /**
1170  * Quick and simple version of osc_extent_find(). This function is frequently
1171  * called to expand the extent for the same IO. To expand the extent, the
1172  * page index must be in the same or next chunk of ext->oe_end.
1173  */
1174 static int osc_extent_expand(struct osc_extent *ext, pgoff_t index,
1175                              unsigned int *grants)
1176 {
1177         struct osc_object *obj = ext->oe_obj;
1178         struct client_obd *cli = osc_cli(obj);
1179         struct osc_extent *next;
1180         int ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1181         pgoff_t chunk = index >> ppc_bits;
1182         pgoff_t end_chunk;
1183         pgoff_t end_index;
1184         unsigned int chunksize = 1 << cli->cl_chunkbits;
1185         int rc = 0;
1186         ENTRY;
1187
1188         LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
1189         osc_object_lock(obj);
1190         LASSERT(sanity_check_nolock(ext) == 0);
1191         end_chunk = ext->oe_end >> ppc_bits;
1192         if (chunk > end_chunk + 1)
1193                 GOTO(out, rc = -ERANGE);
1194
1195         if (end_chunk >= chunk)
1196                 GOTO(out, rc = 0);
1197
1198         LASSERT(end_chunk + 1 == chunk);
1199
1200         /* try to expand this extent to cover @index */
1201         end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
1202
1203         /* don't go over the maximum extent size reported by server */
1204         if (end_index - ext->oe_start + 1 > cli->cl_max_extent_pages)
1205                 GOTO(out, rc = -ERANGE);
1206
1207         next = next_extent(ext);
1208         if (next != NULL && next->oe_start <= end_index)
1209                 /* complex mode - overlapped with the next extent,
1210                  * this case will be handled by osc_extent_find() */
1211                 GOTO(out, rc = -EAGAIN);
1212
1213         ext->oe_end = end_index;
1214         ext->oe_grants += chunksize;
1215         LASSERT(*grants >= chunksize);
1216         *grants -= chunksize;
1217         EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
1218                  "overlapped after expanding for %lu.\n", index);
1219         EXIT;
1220
1221 out:
1222         osc_object_unlock(obj);
1223         RETURN(rc);
1224 }
1225
1226 static void osc_extent_tree_dump0(int mask, struct osc_object *obj,
1227                                   const char *func, int line)
1228 {
1229         struct osc_extent *ext;
1230         int cnt;
1231
1232         if (!cfs_cdebug_show(mask, DEBUG_SUBSYSTEM))
1233                 return;
1234
1235         CDEBUG(mask, "Dump object %p extents at %s:%d, mppr: %u.\n",
1236                obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
1237
1238         /* osc_object_lock(obj); */
1239         cnt = 1;
1240         for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
1241                 OSC_EXTENT_DUMP(mask, ext, "in tree %d.\n", cnt++);
1242
1243         cnt = 1;
1244         list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
1245                 OSC_EXTENT_DUMP(mask, ext, "hp %d.\n", cnt++);
1246
1247         cnt = 1;
1248         list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
1249                 OSC_EXTENT_DUMP(mask, ext, "urgent %d.\n", cnt++);
1250
1251         cnt = 1;
1252         list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
1253                 OSC_EXTENT_DUMP(mask, ext, "reading %d.\n", cnt++);
1254         /* osc_object_unlock(obj); */
1255 }
1256
1257 /* ------------------ osc extent end ------------------ */
1258
1259 static inline int osc_is_ready(struct osc_object *osc)
1260 {
1261         return !list_empty(&osc->oo_ready_item) ||
1262                !list_empty(&osc->oo_hp_ready_item);
1263 }
1264
1265 #define OSC_IO_DEBUG(OSC, STR, args...)                                        \
1266         CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
1267                (OSC), osc_is_ready(OSC),                                       \
1268                list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
1269                list_empty_marker(&(OSC)->oo_ready_item),                       \
1270                atomic_read(&(OSC)->oo_nr_writes),                              \
1271                list_empty_marker(&(OSC)->oo_hp_exts),                          \
1272                list_empty_marker(&(OSC)->oo_urgent_exts),                      \
1273                atomic_read(&(OSC)->oo_nr_reads),                               \
1274                list_empty_marker(&(OSC)->oo_reading_exts),                     \
1275                ##args)
1276
1277 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
1278                           int cmd)
1279 {
1280         struct osc_page *opg  = oap2osc_page(oap);
1281         struct cl_page  *page = oap2cl_page(oap);
1282         int result;
1283
1284         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
1285
1286         ENTRY;
1287         result = cl_page_make_ready(env, page, CRT_WRITE);
1288         if (result == 0)
1289                 opg->ops_submit_time = ktime_get();
1290         RETURN(result);
1291 }
1292
1293 static int osc_refresh_count(const struct lu_env *env,
1294                              struct osc_async_page *oap, int cmd)
1295 {
1296         struct osc_page  *opg = oap2osc_page(oap);
1297         pgoff_t index = osc_index(oap2osc(oap));
1298         struct cl_object *obj;
1299         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
1300         int result;
1301         loff_t kms;
1302
1303         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
1304         LASSERT(!(cmd & OBD_BRW_READ));
1305         LASSERT(opg != NULL);
1306         obj = opg->ops_cl.cpl_obj;
1307
1308         cl_object_attr_lock(obj);
1309         result = cl_object_attr_get(env, obj, attr);
1310         cl_object_attr_unlock(obj);
1311         if (result < 0)
1312                 return result;
1313         kms = attr->cat_kms;
1314         if (cl_offset(obj, index) >= kms)
1315                 /* catch race with truncate */
1316                 return 0;
1317         else if (cl_offset(obj, index + 1) > kms)
1318                 /* catch sub-page write at end of file */
1319                 return kms & ~PAGE_MASK;
1320         else
1321                 return PAGE_SIZE;
1322 }
1323
1324 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
1325                           int cmd, int rc)
1326 {
1327         struct osc_page   *opg  = oap2osc_page(oap);
1328         struct cl_page    *page = oap2cl_page(oap);
1329         enum cl_req_type   crt;
1330         int srvlock;
1331
1332         ENTRY;
1333
1334         cmd &= ~OBD_BRW_NOQUOTA;
1335         LASSERTF(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ),
1336                  "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1337         LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
1338                 "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
1339         LASSERT(opg->ops_transfer_pinned);
1340
1341         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
1342         /* Clear opg->ops_transfer_pinned before VM lock is released. */
1343         opg->ops_transfer_pinned = 0;
1344
1345         opg->ops_submit_time = ktime_set(0, 0);
1346         srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
1347
1348         /* statistic */
1349         if (rc == 0 && srvlock) {
1350                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
1351                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
1352                 size_t bytes = oap->oap_count;
1353
1354                 if (crt == CRT_READ)
1355                         stats->os_lockless_reads += bytes;
1356                 else
1357                         stats->os_lockless_writes += bytes;
1358         }
1359
1360         /*
1361          * This has to be the last operation with the page, as locks are
1362          * released in cl_page_completion() and nothing except for the
1363          * reference counter protects page from concurrent reclaim.
1364          */
1365         lu_ref_del(&page->cp_reference, "transfer", page);
1366
1367         cl_page_completion(env, page, crt, rc);
1368         cl_page_put(env, page);
1369
1370         RETURN(0);
1371 }
1372
1373 #define OSC_DUMP_GRANT(mask, cli, fmt, args...) do {                    \
1374         struct client_obd *__tmp = (cli);                               \
1375         CDEBUG(mask, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu " \
1376                "dropped: %ld avail: %ld, dirty_grant: %ld, "            \
1377                "reserved: %ld, flight: %d } lru {in list: %ld, "        \
1378                "left: %ld, waiters: %d }" fmt "\n",                     \
1379                cli_name(__tmp),                                         \
1380                __tmp->cl_dirty_pages, __tmp->cl_dirty_max_pages,        \
1381                atomic_long_read(&obd_dirty_pages), obd_max_dirty_pages, \
1382                __tmp->cl_lost_grant, __tmp->cl_avail_grant,             \
1383                __tmp->cl_dirty_grant,                                   \
1384                __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,         \
1385                atomic_long_read(&__tmp->cl_lru_in_list),                \
1386                atomic_long_read(&__tmp->cl_lru_busy),                   \
1387                atomic_read(&__tmp->cl_lru_shrinkers), ##args);          \
1388 } while (0)
1389
1390 /* caller must hold loi_list_lock */
1391 static void osc_consume_write_grant(struct client_obd *cli,
1392                                     struct brw_page *pga)
1393 {
1394         assert_spin_locked(&cli->cl_loi_list_lock);
1395         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
1396         cli->cl_dirty_pages++;
1397         pga->flag |= OBD_BRW_FROM_GRANT;
1398         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
1399                PAGE_SIZE, pga, pga->pg);
1400         osc_update_next_shrink(cli);
1401 }
1402
1403 /* the companion to osc_consume_write_grant, called when a brw has completed.
1404  * must be called with the loi lock held. */
1405 static void osc_release_write_grant(struct client_obd *cli,
1406                                     struct brw_page *pga)
1407 {
1408         ENTRY;
1409
1410         assert_spin_locked(&cli->cl_loi_list_lock);
1411         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
1412                 EXIT;
1413                 return;
1414         }
1415
1416         pga->flag &= ~OBD_BRW_FROM_GRANT;
1417         atomic_long_dec(&obd_dirty_pages);
1418         cli->cl_dirty_pages--;
1419         EXIT;
1420 }
1421
1422 /**
1423  * To avoid sleeping with object lock held, it's good for us allocate enough
1424  * grants before entering into critical section.
1425  *
1426  * client_obd_list_lock held by caller
1427  */
1428 static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
1429 {
1430         int rc = -EDQUOT;
1431
1432         if (cli->cl_avail_grant >= bytes) {
1433                 cli->cl_avail_grant    -= bytes;
1434                 cli->cl_reserved_grant += bytes;
1435                 rc = 0;
1436         }
1437         return rc;
1438 }
1439
1440 static void __osc_unreserve_grant(struct client_obd *cli,
1441                                   unsigned int reserved, unsigned int unused)
1442 {
1443         /* it's quite normal for us to get more grant than reserved.
1444          * Thinking about a case that two extents merged by adding a new
1445          * chunk, we can save one extent tax. If extent tax is greater than
1446          * one chunk, we can save more grant by adding a new chunk */
1447         cli->cl_reserved_grant -= reserved;
1448         if (unused > reserved) {
1449                 cli->cl_avail_grant += reserved;
1450                 cli->cl_lost_grant  += unused - reserved;
1451                 cli->cl_dirty_grant -= unused - reserved;
1452         } else {
1453                 cli->cl_avail_grant += unused;
1454                 cli->cl_dirty_grant += reserved - unused;
1455         }
1456 }
1457
1458 static void osc_unreserve_grant_nolock(struct client_obd *cli,
1459                                        unsigned int reserved,
1460                                        unsigned int unused)
1461 {
1462         __osc_unreserve_grant(cli, reserved, unused);
1463         if (unused > 0)
1464                 osc_wake_cache_waiters(cli);
1465 }
1466
1467 static void osc_unreserve_grant(struct client_obd *cli,
1468                                 unsigned int reserved, unsigned int unused)
1469 {
1470         spin_lock(&cli->cl_loi_list_lock);
1471         osc_unreserve_grant_nolock(cli, reserved, unused);
1472         spin_unlock(&cli->cl_loi_list_lock);
1473 }
1474
1475 /**
1476  * Free grant after IO is finished or canceled.
1477  *
1478  * @lost_grant is used to remember how many grants we have allocated but not
1479  * used, we should return these grants to OST. There're two cases where grants
1480  * can be lost:
1481  * 1. truncate;
1482  * 2. blocksize at OST is less than PAGE_SIZE and a partial page was
1483  *    written. In this case OST may use less chunks to serve this partial
1484  *    write. OSTs don't actually know the page size on the client side. so
1485  *    clients have to calculate lost grant by the blocksize on the OST.
1486  *    See filter_grant_check() for details.
1487  */
1488 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
1489                            unsigned int lost_grant, unsigned int dirty_grant)
1490 {
1491         unsigned long grant;
1492
1493         grant = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
1494
1495         spin_lock(&cli->cl_loi_list_lock);
1496         atomic_long_sub(nr_pages, &obd_dirty_pages);
1497         cli->cl_dirty_pages -= nr_pages;
1498         cli->cl_lost_grant += lost_grant;
1499         cli->cl_dirty_grant -= dirty_grant;
1500         if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
1501                 /* borrow some grant from truncate to avoid the case that
1502                  * truncate uses up all avail grant */
1503                 cli->cl_lost_grant -= grant;
1504                 cli->cl_avail_grant += grant;
1505         }
1506         osc_wake_cache_waiters(cli);
1507         spin_unlock(&cli->cl_loi_list_lock);
1508         CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu/%lu\n",
1509                lost_grant, cli->cl_lost_grant,
1510                cli->cl_avail_grant, cli->cl_dirty_pages << PAGE_SHIFT,
1511                cli->cl_dirty_grant);
1512 }
1513
1514 /**
1515  * The companion to osc_enter_cache(), called when @oap is no longer part of
1516  * the dirty accounting due to error.
1517  */
1518 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
1519 {
1520         spin_lock(&cli->cl_loi_list_lock);
1521         osc_release_write_grant(cli, &oap->oap_brw_page);
1522         spin_unlock(&cli->cl_loi_list_lock);
1523 }
1524
1525 /**
1526  * Non-blocking version of osc_enter_cache() that consumes grant only when it
1527  * is available.
1528  */
1529 static int osc_enter_cache_try(struct client_obd *cli,
1530                                struct osc_async_page *oap,
1531                                int bytes)
1532 {
1533         int rc;
1534
1535         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1536
1537         rc = osc_reserve_grant(cli, bytes);
1538         if (rc < 0)
1539                 return 0;
1540
1541         if (cli->cl_dirty_pages < cli->cl_dirty_max_pages) {
1542                 if (atomic_long_add_return(1, &obd_dirty_pages) <=
1543                     obd_max_dirty_pages) {
1544                         osc_consume_write_grant(cli, &oap->oap_brw_page);
1545                         rc = 1;
1546                         goto out;
1547                 } else
1548                         atomic_long_dec(&obd_dirty_pages);
1549         }
1550         __osc_unreserve_grant(cli, bytes, bytes);
1551
1552 out:
1553         return rc;
1554 }
1555
1556 /* Following two inlines exist to pass code fragments
1557  * to wait_event_idle_exclusive_timeout_cmd().  Passing
1558  * code fragments as macro args can look confusing, so
1559  * we provide inlines to encapsulate them.
1560  */
1561 static inline void cli_unlock_and_unplug(const struct lu_env *env,
1562                                          struct client_obd *cli,
1563                                          struct osc_async_page *oap)
1564 {
1565         spin_unlock(&cli->cl_loi_list_lock);
1566         osc_io_unplug_async(env, cli, NULL);
1567         CDEBUG(D_CACHE,
1568                "%s: sleeping for cache space for %p\n",
1569                cli_name(cli), oap);
1570 }
1571
1572 static inline void cli_lock_after_unplug(struct client_obd *cli)
1573 {
1574         spin_lock(&cli->cl_loi_list_lock);
1575 }
1576 /**
1577  * The main entry to reserve dirty page accounting. Usually the grant reserved
1578  * in this function will be freed in bulk in osc_free_grant() unless it fails
1579  * to add osc cache, in that case, it will be freed in osc_exit_cache().
1580  *
1581  * The process will be put into sleep if it's already run out of grant.
1582  */
1583 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
1584                            struct osc_async_page *oap, int bytes)
1585 {
1586         struct osc_object *osc = oap->oap_obj;
1587         struct lov_oinfo *loi = osc->oo_oinfo;
1588         int rc = -EDQUOT;
1589         int remain;
1590         bool entered = false;
1591         /* We cannot wait for a long time here since we are holding ldlm lock
1592          * across the actual IO. If no requests complete fast (e.g. due to
1593          * overloaded OST that takes a long time to process everything, we'd
1594          * get evicted if we wait for a normal obd_timeout or some such.
1595          * So we try to wait half the time it would take the client to be
1596          * evicted by server which is half obd_timeout when AT is off
1597          * or at least ldlm_enqueue_min with AT on.
1598          * See LU-13131 */
1599         unsigned long timeout = cfs_time_seconds(AT_OFF ? obd_timeout / 2 :
1600                                                           ldlm_enqueue_min / 2);
1601
1602         ENTRY;
1603
1604         OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
1605
1606         spin_lock(&cli->cl_loi_list_lock);
1607
1608         /* force the caller to try sync io.  this can jump the list
1609          * of queued writes and create a discontiguous rpc stream */
1610         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
1611             cli->cl_dirty_max_pages == 0 ||
1612             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync) {
1613                 OSC_DUMP_GRANT(D_CACHE, cli, "forced sync i/o\n");
1614                 GOTO(out, rc = -EDQUOT);
1615         }
1616
1617         /*
1618          * We can wait here for two reasons: too many dirty pages in cache, or
1619          * run out of grants. In both cases we should write dirty pages out.
1620          * Adding a cache waiter will trigger urgent write-out no matter what
1621          * RPC size will be.
1622          * The exiting condition (other than success) is no avail grants
1623          * and no dirty pages caching, that really means there is no space
1624          * on the OST.
1625          */
1626         remain = wait_event_idle_exclusive_timeout_cmd(
1627                 cli->cl_cache_waiters,
1628                 (entered = osc_enter_cache_try(cli, oap, bytes)) ||
1629                 (cli->cl_dirty_pages == 0 && cli->cl_w_in_flight == 0),
1630                 timeout,
1631                 cli_unlock_and_unplug(env, cli, oap),
1632                 cli_lock_after_unplug(cli));
1633
1634         if (entered) {
1635                 if (remain == timeout)
1636                         OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
1637                 else
1638                         OSC_DUMP_GRANT(D_CACHE, cli,
1639                                        "finally got grant space\n");
1640                 wake_up(&cli->cl_cache_waiters);
1641                 rc = 0;
1642         } else if (remain == 0) {
1643                 OSC_DUMP_GRANT(D_CACHE, cli,
1644                                "timeout, fall back to sync i/o\n");
1645                 osc_extent_tree_dump(D_CACHE, osc);
1646                 /* fall back to synchronous I/O */
1647         } else {
1648                 OSC_DUMP_GRANT(D_CACHE, cli,
1649                                "no grant space, fall back to sync i/o\n");
1650                 wake_up_all(&cli->cl_cache_waiters);
1651         }
1652         EXIT;
1653 out:
1654         spin_unlock(&cli->cl_loi_list_lock);
1655         RETURN(rc);
1656 }
1657
1658 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
1659 {
1660         int hprpc = !!list_empty(&osc->oo_hp_exts);
1661         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
1662 }
1663
1664 /* This maintains the lists of pending pages to read/write for a given object
1665  * (lop).  This is used by osc_check_rpcs->osc_next_obj() and osc_list_maint()
1666  * to quickly find objects that are ready to send an RPC. */
1667 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
1668                          int cmd)
1669 {
1670         int invalid_import = 0;
1671         ENTRY;
1672
1673         /* if we have an invalid import we want to drain the queued pages
1674          * by forcing them through rpcs that immediately fail and complete
1675          * the pages.  recovery relies on this to empty the queued pages
1676          * before canceling the locks and evicting down the llite pages */
1677         if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
1678                 invalid_import = 1;
1679
1680         if (cmd & OBD_BRW_WRITE) {
1681                 if (atomic_read(&osc->oo_nr_writes) == 0)
1682                         RETURN(0);
1683                 if (invalid_import) {
1684                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1685                         RETURN(1);
1686                 }
1687                 if (!list_empty(&osc->oo_hp_exts)) {
1688                         CDEBUG(D_CACHE, "high prio request forcing RPC\n");
1689                         RETURN(1);
1690                 }
1691                 if (!list_empty(&osc->oo_urgent_exts)) {
1692                         CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1693                         RETURN(1);
1694                 }
1695                 /* trigger a write rpc stream as long as there are dirtiers
1696                  * waiting for space.  as they're waiting, they're not going to
1697                  * create more pages to coalesce with what's waiting..
1698                  */
1699                 if (waitqueue_active(&cli->cl_cache_waiters)) {
1700                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1701                         RETURN(1);
1702                 }
1703                 if (!list_empty(&osc->oo_full_exts)) {
1704                         CDEBUG(D_CACHE, "full extent ready, make an RPC\n");
1705                         RETURN(1);
1706                 }
1707         } else {
1708                 if (atomic_read(&osc->oo_nr_reads) == 0)
1709                         RETURN(0);
1710                 if (invalid_import) {
1711                         CDEBUG(D_CACHE, "invalid import forcing RPC\n");
1712                         RETURN(1);
1713                 }
1714                 /* all read are urgent. */
1715                 if (!list_empty(&osc->oo_reading_exts))
1716                         RETURN(1);
1717         }
1718
1719         RETURN(0);
1720 }
1721
1722 static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
1723 {
1724         struct client_obd *cli = osc_cli(obj);
1725         if (cmd & OBD_BRW_WRITE) {
1726                 atomic_add(delta, &obj->oo_nr_writes);
1727                 atomic_add(delta, &cli->cl_pending_w_pages);
1728                 LASSERT(atomic_read(&obj->oo_nr_writes) >= 0);
1729         } else {
1730                 atomic_add(delta, &obj->oo_nr_reads);
1731                 atomic_add(delta, &cli->cl_pending_r_pages);
1732                 LASSERT(atomic_read(&obj->oo_nr_reads) >= 0);
1733         }
1734         OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
1735 }
1736
1737 static int osc_makes_hprpc(struct osc_object *obj)
1738 {
1739         return !list_empty(&obj->oo_hp_exts);
1740 }
1741
1742 static void on_list(struct list_head *item, struct list_head *list,
1743                     int should_be_on)
1744 {
1745         if (list_empty(item) && should_be_on)
1746                 list_add_tail(item, list);
1747         else if (!list_empty(item) && !should_be_on)
1748                 list_del_init(item);
1749 }
1750
1751 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
1752  * can find pages to build into rpcs quickly */
1753 static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1754 {
1755         if (osc_makes_hprpc(osc)) {
1756                 /* HP rpc */
1757                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
1758                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1759         } else {
1760                 on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1761                 on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list,
1762                         osc_makes_rpc(cli, osc, OBD_BRW_WRITE) ||
1763                         osc_makes_rpc(cli, osc, OBD_BRW_READ));
1764         }
1765
1766         on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
1767                 atomic_read(&osc->oo_nr_writes) > 0);
1768
1769         on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
1770                 atomic_read(&osc->oo_nr_reads) > 0);
1771
1772         return osc_is_ready(osc);
1773 }
1774
1775 static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
1776 {
1777         int is_ready;
1778
1779         spin_lock(&cli->cl_loi_list_lock);
1780         is_ready = __osc_list_maint(cli, osc);
1781         spin_unlock(&cli->cl_loi_list_lock);
1782
1783         return is_ready;
1784 }
1785
1786 /* this is trying to propogate async writeback errors back up to the
1787  * application.  As an async write fails we record the error code for later if
1788  * the app does an fsync.  As long as errors persist we force future rpcs to be
1789  * sync so that the app can get a sync error and break the cycle of queueing
1790  * pages for which writeback will fail. */
1791 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
1792                            int rc)
1793 {
1794         if (rc) {
1795                 if (!ar->ar_rc)
1796                         ar->ar_rc = rc;
1797
1798                 ar->ar_force_sync = 1;
1799                 ar->ar_min_xid = ptlrpc_sample_next_xid();
1800                 return;
1801
1802         }
1803
1804         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
1805                 ar->ar_force_sync = 0;
1806 }
1807
1808 /* this must be called holding the loi list lock to give coverage to exit_cache,
1809  * async_flag maintenance, and oap_request */
1810 static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
1811                               struct osc_async_page *oap, int sent, int rc)
1812 {
1813         struct osc_object *osc = oap->oap_obj;
1814         struct lov_oinfo  *loi = osc->oo_oinfo;
1815         __u64 xid = 0;
1816
1817         ENTRY;
1818         if (oap->oap_request != NULL) {
1819                 xid = ptlrpc_req_xid(oap->oap_request);
1820                 ptlrpc_req_finished(oap->oap_request);
1821                 oap->oap_request = NULL;
1822         }
1823
1824         /* As the transfer for this page is being done, clear the flags */
1825         spin_lock(&oap->oap_lock);
1826         oap->oap_async_flags = 0;
1827         spin_unlock(&oap->oap_lock);
1828
1829         if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
1830                 spin_lock(&cli->cl_loi_list_lock);
1831                 osc_process_ar(&cli->cl_ar, xid, rc);
1832                 osc_process_ar(&loi->loi_ar, xid, rc);
1833                 spin_unlock(&cli->cl_loi_list_lock);
1834         }
1835
1836         rc = osc_completion(env, oap, oap->oap_cmd, rc);
1837         if (rc)
1838                 CERROR("completion on oap %p obj %p returns %d.\n",
1839                        oap, osc, rc);
1840
1841         EXIT;
1842 }
1843
1844 struct extent_rpc_data {
1845         struct list_head        *erd_rpc_list;
1846         unsigned int            erd_page_count;
1847         unsigned int            erd_max_pages;
1848         unsigned int            erd_max_chunks;
1849         unsigned int            erd_max_extents;
1850 };
1851
1852 static inline unsigned osc_extent_chunks(const struct osc_extent *ext)
1853 {
1854         struct client_obd *cli = osc_cli(ext->oe_obj);
1855         unsigned ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
1856
1857         return (ext->oe_end >> ppc_bits) - (ext->oe_start >> ppc_bits) + 1;
1858 }
1859
1860 static inline bool
1861 can_merge(const struct osc_extent *ext, const struct osc_extent *in_rpc)
1862 {
1863         if (ext->oe_no_merge || in_rpc->oe_no_merge)
1864                 return false;
1865
1866         if (ext->oe_srvlock != in_rpc->oe_srvlock)
1867                 return false;
1868
1869         if (ext->oe_ndelay != in_rpc->oe_ndelay)
1870                 return false;
1871
1872         if (!ext->oe_grants != !in_rpc->oe_grants)
1873                 return false;
1874
1875         if (ext->oe_dio != in_rpc->oe_dio)
1876                 return false;
1877
1878         /* It's possible to have overlap on DIO */
1879         if (in_rpc->oe_dio && overlapped(ext, in_rpc))
1880                 return false;
1881
1882         if (ext->oe_is_rdma_only != in_rpc->oe_is_rdma_only)
1883                 return false;
1884
1885         return true;
1886 }
1887
1888 /**
1889  * Try to add extent to one RPC. We need to think about the following things:
1890  * - # of pages must not be over max_pages_per_rpc
1891  * - extent must be compatible with previous ones
1892  */
1893 static int try_to_add_extent_for_io(struct client_obd *cli,
1894                                     struct osc_extent *ext,
1895                                     struct extent_rpc_data *data)
1896 {
1897         struct osc_extent *tmp;
1898         unsigned int chunk_count;
1899         ENTRY;
1900
1901         EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
1902                 ext);
1903         OSC_EXTENT_DUMP(D_CACHE, ext, "trying to add this extent\n");
1904
1905         if (data->erd_max_extents == 0)
1906                 RETURN(0);
1907
1908         chunk_count = osc_extent_chunks(ext);
1909         EASSERTF(data->erd_page_count != 0 ||
1910                  chunk_count <= data->erd_max_chunks, ext,
1911                  "The first extent to be fit in a RPC contains %u chunks, "
1912                  "which is over the limit %u.\n", chunk_count,
1913                  data->erd_max_chunks);
1914         if (chunk_count > data->erd_max_chunks)
1915                 RETURN(0);
1916
1917         data->erd_max_pages = max(ext->oe_mppr, data->erd_max_pages);
1918         EASSERTF(data->erd_page_count != 0 ||
1919                 ext->oe_nr_pages <= data->erd_max_pages, ext,
1920                 "The first extent to be fit in a RPC contains %u pages, "
1921                 "which is over the limit %u.\n", ext->oe_nr_pages,
1922                 data->erd_max_pages);
1923         if (data->erd_page_count + ext->oe_nr_pages > data->erd_max_pages)
1924                 RETURN(0);
1925
1926         list_for_each_entry(tmp, data->erd_rpc_list, oe_link) {
1927                 EASSERT(tmp->oe_owner == current, tmp);
1928
1929                 if (!can_merge(ext, tmp))
1930                         RETURN(0);
1931         }
1932
1933         data->erd_max_extents--;
1934         data->erd_max_chunks -= chunk_count;
1935         data->erd_page_count += ext->oe_nr_pages;
1936         list_move_tail(&ext->oe_link, data->erd_rpc_list);
1937         ext->oe_owner = current;
1938         RETURN(1);
1939 }
1940
1941 /**
1942  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
1943  * get_write_extent() takes all appropriate extents in atomic.
1944  *
1945  * The following policy is used to collect extents for IO:
1946  * 1. Add as many HP extents as possible;
1947  * 2. Add the first urgent extent in urgent extent list and take it out of
1948  *    urgent list;
1949  * 3. Add subsequent extents of this urgent extent;
1950  * 4. If urgent list is not empty, goto 2;
1951  * 5. Traverse the extent tree from the 1st extent;
1952  * 6. Above steps exit if there is no space in this RPC.
1953  */
1954 static unsigned int get_write_extents(struct osc_object *obj,
1955                                       struct list_head *rpclist)
1956 {
1957         struct client_obd *cli = osc_cli(obj);
1958         struct osc_extent *ext;
1959         struct extent_rpc_data data = {
1960                 .erd_rpc_list   = rpclist,
1961                 .erd_page_count = 0,
1962                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
1963                 .erd_max_chunks = osc_max_write_chunks(cli),
1964                 .erd_max_extents = 256,
1965         };
1966
1967         assert_osc_object_is_locked(obj);
1968         while (!list_empty(&obj->oo_hp_exts)) {
1969                 ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
1970                                  oe_link);
1971                 if (!try_to_add_extent_for_io(cli, ext, &data))
1972                         return data.erd_page_count;
1973                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
1974         }
1975         if (data.erd_page_count == data.erd_max_pages)
1976                 return data.erd_page_count;
1977
1978         while (!list_empty(&obj->oo_urgent_exts)) {
1979                 ext = list_entry(obj->oo_urgent_exts.next,
1980                                  struct osc_extent, oe_link);
1981                 if (!try_to_add_extent_for_io(cli, ext, &data))
1982                         return data.erd_page_count;
1983         }
1984         if (data.erd_page_count == data.erd_max_pages)
1985                 return data.erd_page_count;
1986
1987         /* One key difference between full extents and other extents: full
1988          * extents can usually only be added if the rpclist was empty, so if we
1989          * can't add one, we continue on to trying to add normal extents.  This
1990          * is so we don't miss adding extra extents to an RPC containing high
1991          * priority or urgent extents. */
1992         while (!list_empty(&obj->oo_full_exts)) {
1993                 ext = list_entry(obj->oo_full_exts.next,
1994                                  struct osc_extent, oe_link);
1995                 if (!try_to_add_extent_for_io(cli, ext, &data))
1996                         break;
1997         }
1998         if (data.erd_page_count == data.erd_max_pages)
1999                 return data.erd_page_count;
2000
2001         for (ext = first_extent(obj);
2002              ext;
2003              ext = next_extent(ext)) {
2004                 if ((ext->oe_state != OES_CACHE) ||
2005                     /* this extent may be already in current rpclist */
2006                     (!list_empty(&ext->oe_link) && ext->oe_owner))
2007                         continue;
2008
2009                 if (!try_to_add_extent_for_io(cli, ext, &data))
2010                         return data.erd_page_count;
2011         }
2012         return data.erd_page_count;
2013 }
2014
2015 static int
2016 osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
2017                    struct osc_object *osc)
2018 __must_hold(osc)
2019 {
2020         LIST_HEAD(rpclist);
2021         struct osc_extent *ext;
2022         struct osc_extent *tmp;
2023         struct osc_extent *first = NULL;
2024         unsigned int page_count = 0;
2025         int srvlock = 0;
2026         int rc = 0;
2027         ENTRY;
2028
2029         assert_osc_object_is_locked(osc);
2030
2031         page_count = get_write_extents(osc, &rpclist);
2032         LASSERT(equi(page_count == 0, list_empty(&rpclist)));
2033
2034         if (list_empty(&rpclist))
2035                 RETURN(0);
2036
2037         osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
2038
2039         list_for_each_entry(ext, &rpclist, oe_link) {
2040                 LASSERT(ext->oe_state == OES_CACHE ||
2041                         ext->oe_state == OES_LOCK_DONE);
2042                 if (ext->oe_state == OES_CACHE)
2043                         osc_extent_state_set(ext, OES_LOCKING);
2044                 else
2045                         osc_extent_state_set(ext, OES_RPC);
2046         }
2047
2048         /* we're going to grab page lock, so release object lock because
2049          * lock order is page lock -> object lock. */
2050         osc_object_unlock(osc);
2051
2052         list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
2053                 if (ext->oe_state == OES_LOCKING) {
2054                         rc = osc_extent_make_ready(env, ext);
2055                         if (unlikely(rc < 0)) {
2056                                 list_del_init(&ext->oe_link);
2057                                 osc_extent_finish(env, ext, 0, rc);
2058                                 continue;
2059                         }
2060                 }
2061                 if (first == NULL) {
2062                         first = ext;
2063                         srvlock = ext->oe_srvlock;
2064                 } else {
2065                         LASSERT(srvlock == ext->oe_srvlock);
2066                 }
2067         }
2068
2069         if (!list_empty(&rpclist)) {
2070                 LASSERT(page_count > 0);
2071                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE);
2072                 LASSERT(list_empty(&rpclist));
2073         }
2074
2075         osc_object_lock(osc);
2076         RETURN(rc);
2077 }
2078
2079 /**
2080  * prepare pages for ASYNC io and put pages in send queue.
2081  *
2082  * \param cmd OBD_BRW_* macroses
2083  * \param lop pending pages
2084  *
2085  * \return zero if no page added to send queue.
2086  * \return 1 if pages successfully added to send queue.
2087  * \return negative on errors.
2088  */
2089 static int
2090 osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
2091                   struct osc_object *osc)
2092 __must_hold(osc)
2093 {
2094         struct osc_extent *ext;
2095         struct osc_extent *next;
2096         LIST_HEAD(rpclist);
2097         struct extent_rpc_data data = {
2098                 .erd_rpc_list   = &rpclist,
2099                 .erd_page_count = 0,
2100                 .erd_max_pages  = cli->cl_max_pages_per_rpc,
2101                 .erd_max_chunks = UINT_MAX,
2102                 .erd_max_extents = UINT_MAX,
2103         };
2104         int rc = 0;
2105         ENTRY;
2106
2107         assert_osc_object_is_locked(osc);
2108         list_for_each_entry_safe(ext, next, &osc->oo_reading_exts, oe_link) {
2109                 EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
2110                 if (!try_to_add_extent_for_io(cli, ext, &data))
2111                         break;
2112                 osc_extent_state_set(ext, OES_RPC);
2113                 EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
2114         }
2115         LASSERT(data.erd_page_count <= data.erd_max_pages);
2116
2117         osc_update_pending(osc, OBD_BRW_READ, -data.erd_page_count);
2118
2119         if (!list_empty(&rpclist)) {
2120                 osc_object_unlock(osc);
2121
2122                 rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ);
2123                 LASSERT(list_empty(&rpclist));
2124
2125                 osc_object_lock(osc);
2126         }
2127         RETURN(rc);
2128 }
2129
2130 #define list_to_obj(list, item) ({                                            \
2131         struct list_head *__tmp = (list)->next;                               \
2132         list_del_init(__tmp);                                         \
2133         list_entry(__tmp, struct osc_object, oo_##item);                      \
2134 })
2135
2136 /* This is called by osc_check_rpcs() to find which objects have pages that
2137  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
2138 static struct osc_object *osc_next_obj(struct client_obd *cli)
2139 {
2140         ENTRY;
2141
2142         /* First return objects that have blocked locks so that they
2143          * will be flushed quickly and other clients can get the lock,
2144          * then objects which have pages ready to be stuffed into RPCs */
2145         if (!list_empty(&cli->cl_loi_hp_ready_list))
2146                 RETURN(list_to_obj(&cli->cl_loi_hp_ready_list, hp_ready_item));
2147         if (!list_empty(&cli->cl_loi_ready_list))
2148                 RETURN(list_to_obj(&cli->cl_loi_ready_list, ready_item));
2149
2150         /* then if we have cache waiters, return all objects with queued
2151          * writes.  This is especially important when many small files
2152          * have filled up the cache and not been fired into rpcs because
2153          * they don't pass the nr_pending/object threshhold
2154          */
2155         if (waitqueue_active(&cli->cl_cache_waiters) &&
2156             !list_empty(&cli->cl_loi_write_list))
2157                 RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
2158
2159         /* then return all queued objects when we have an invalid import
2160          * so that they get flushed */
2161         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2162                 if (!list_empty(&cli->cl_loi_write_list))
2163                         RETURN(list_to_obj(&cli->cl_loi_write_list,
2164                                            write_item));
2165                 if (!list_empty(&cli->cl_loi_read_list))
2166                         RETURN(list_to_obj(&cli->cl_loi_read_list,
2167                                            read_item));
2168         }
2169         RETURN(NULL);
2170 }
2171
2172 /* called with the loi list lock held */
2173 static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2174 __must_hold(&cli->cl_loi_list_lock)
2175 {
2176         struct osc_object *osc;
2177         int rc = 0;
2178         ENTRY;
2179
2180         while ((osc = osc_next_obj(cli)) != NULL) {
2181                 struct cl_object *obj = osc2cl(osc);
2182                 struct lu_ref_link link;
2183
2184                 OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
2185
2186                 /* even if we have reached our max in flight RPCs, we still
2187                  * allow all high-priority RPCs through to prevent their
2188                  * starvation and leading to server evicting us for not
2189                  * writing out pages in a timely manner LU-13131 */
2190                 if (osc_max_rpc_in_flight(cli, osc) &&
2191                     list_empty(&osc->oo_hp_exts)) {
2192                         __osc_list_maint(cli, osc);
2193                         break;
2194                 }
2195
2196                 cl_object_get(obj);
2197                 spin_unlock(&cli->cl_loi_list_lock);
2198                 lu_object_ref_add_at(&obj->co_lu, &link, "check", current);
2199
2200                 /* attempt some read/write balancing by alternating between
2201                  * reads and writes in an object.  The makes_rpc checks here
2202                  * would be redundant if we were getting read/write work items
2203                  * instead of objects.  we don't want send_oap_rpc to drain a
2204                  * partial read pending queue when we're given this object to
2205                  * do io on writes while there are cache waiters */
2206                 osc_object_lock(osc);
2207                 if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
2208                         rc = osc_send_write_rpc(env, cli, osc);
2209                         if (rc < 0) {
2210                                 CERROR("Write request failed with %d\n", rc);
2211
2212                                 /* osc_send_write_rpc failed, mostly because of
2213                                  * memory pressure.
2214                                  *
2215                                  * It can't break here, because if:
2216                                  *  - a page was submitted by osc_io_submit, so
2217                                  *    page locked;
2218                                  *  - no request in flight
2219                                  *  - no subsequent request
2220                                  * The system will be in live-lock state,
2221                                  * because there is no chance to call
2222                                  * osc_io_unplug() and osc_check_rpcs() any
2223                                  * more. pdflush can't help in this case,
2224                                  * because it might be blocked at grabbing
2225                                  * the page lock as we mentioned.
2226                                  *
2227                                  * Anyway, continue to drain pages. */
2228                                 /* break; */
2229                         }
2230                 }
2231                 if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
2232                         rc = osc_send_read_rpc(env, cli, osc);
2233                         if (rc < 0)
2234                                 CERROR("Read request failed with %d\n", rc);
2235                 }
2236                 osc_object_unlock(osc);
2237
2238                 osc_list_maint(cli, osc);
2239                 lu_object_ref_del_at(&obj->co_lu, &link, "check", current);
2240                 cl_object_put(env, obj);
2241
2242                 spin_lock(&cli->cl_loi_list_lock);
2243         }
2244 }
2245
2246 int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
2247                    struct osc_object *osc, int async)
2248 {
2249         int rc = 0;
2250
2251         if (osc != NULL && osc_list_maint(cli, osc) == 0)
2252                 return 0;
2253
2254         if (!async) {
2255                 spin_lock(&cli->cl_loi_list_lock);
2256                 osc_check_rpcs(env, cli);
2257                 spin_unlock(&cli->cl_loi_list_lock);
2258         } else {
2259                 CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
2260                 LASSERT(cli->cl_writeback_work != NULL);
2261                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
2262         }
2263         return rc;
2264 }
2265 EXPORT_SYMBOL(osc_io_unplug0);
2266
2267 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
2268                         struct page *page, loff_t offset)
2269 {
2270         struct obd_export     *exp = osc_export(osc);
2271         struct osc_async_page *oap = &ops->ops_oap;
2272         ENTRY;
2273
2274         if (!page)
2275                 return cfs_size_round(sizeof(*oap));
2276
2277         oap->oap_magic = OAP_MAGIC;
2278         oap->oap_cli = &exp->exp_obd->u.cli;
2279         oap->oap_obj = osc;
2280
2281         oap->oap_page = page;
2282         oap->oap_obj_off = offset;
2283         LASSERT(!(offset & ~PAGE_MASK));
2284
2285         INIT_LIST_HEAD(&oap->oap_pending_item);
2286         INIT_LIST_HEAD(&oap->oap_rpc_item);
2287
2288         spin_lock_init(&oap->oap_lock);
2289         CDEBUG(D_INFO, "oap %p page %p obj off %llu\n",
2290                oap, page, oap->oap_obj_off);
2291         RETURN(0);
2292 }
2293 EXPORT_SYMBOL(osc_prep_async_page);
2294
2295 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
2296                        struct osc_page *ops, cl_commit_cbt cb)
2297 {
2298         struct osc_io *oio = osc_env_io(env);
2299         struct osc_extent     *ext = NULL;
2300         struct osc_async_page *oap = &ops->ops_oap;
2301         struct client_obd     *cli = oap->oap_cli;
2302         struct osc_object     *osc = oap->oap_obj;
2303         struct pagevec        *pvec = &osc_env_info(env)->oti_pagevec;
2304         pgoff_t index;
2305         unsigned int tmp;
2306         unsigned int grants = 0;
2307         u32    brw_flags = OBD_BRW_ASYNC;
2308         int    cmd = OBD_BRW_WRITE;
2309         int    need_release = 0;
2310         int    rc = 0;
2311         ENTRY;
2312
2313         if (oap->oap_magic != OAP_MAGIC)
2314                 RETURN(-EINVAL);
2315
2316         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2317                 RETURN(-EIO);
2318
2319         if (!list_empty(&oap->oap_pending_item) ||
2320             !list_empty(&oap->oap_rpc_item))
2321                 RETURN(-EBUSY);
2322
2323         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
2324         brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
2325         if (oio->oi_cap_sys_resource || io->ci_noquota) {
2326                 brw_flags |= OBD_BRW_NOQUOTA;
2327                 cmd |= OBD_BRW_NOQUOTA;
2328         }
2329
2330         /* check if the file's owner/group is over quota */
2331         if (!(cmd & OBD_BRW_NOQUOTA)) {
2332                 struct cl_object *obj;
2333                 struct cl_attr   *attr;
2334                 unsigned int qid[LL_MAXQUOTAS];
2335
2336                 obj = cl_object_top(&osc->oo_cl);
2337                 attr = &osc_env_info(env)->oti_attr;
2338
2339                 cl_object_attr_lock(obj);
2340                 rc = cl_object_attr_get(env, obj, attr);
2341                 cl_object_attr_unlock(obj);
2342
2343                 qid[USRQUOTA] = attr->cat_uid;
2344                 qid[GRPQUOTA] = attr->cat_gid;
2345                 qid[PRJQUOTA] = attr->cat_projid;
2346                 if (rc == 0 && osc_quota_chkdq(cli, qid) == -EDQUOT)
2347                         rc = -EDQUOT;
2348                 if (rc)
2349                         RETURN(rc);
2350         }
2351
2352         oap->oap_cmd = cmd;
2353         oap->oap_page_off = ops->ops_from;
2354         oap->oap_count = ops->ops_to - ops->ops_from + 1;
2355         /* No need to hold a lock here,
2356          * since this page is not in any list yet. */
2357         oap->oap_async_flags = 0;
2358         oap->oap_brw_flags = brw_flags;
2359
2360         OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
2361                      oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
2362
2363         index = osc_index(oap2osc(oap));
2364
2365         /* Add this page into extent by the following steps:
2366          * 1. if there exists an active extent for this IO, mostly this page
2367          *    can be added to the active extent and sometimes we need to
2368          *    expand extent to accomodate this page;
2369          * 2. otherwise, a new extent will be allocated. */
2370
2371         ext = oio->oi_active;
2372         if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
2373                 /* one chunk plus extent overhead must be enough to write this
2374                  * page */
2375                 grants = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2376                 if (ext->oe_end >= index)
2377                         grants = 0;
2378
2379                 /* it doesn't need any grant to dirty this page */
2380                 spin_lock(&cli->cl_loi_list_lock);
2381                 rc = osc_enter_cache_try(cli, oap, grants);
2382                 if (rc == 0) { /* try failed */
2383                         grants = 0;
2384                         need_release = 1;
2385                 } else if (ext->oe_end < index) {
2386                         tmp = grants;
2387                         /* try to expand this extent */
2388                         rc = osc_extent_expand(ext, index, &tmp);
2389                         if (rc < 0) {
2390                                 need_release = 1;
2391                                 /* don't free reserved grant */
2392                         } else {
2393                                 OSC_EXTENT_DUMP(D_CACHE, ext,
2394                                                 "expanded for %lu.\n", index);
2395                                 osc_unreserve_grant_nolock(cli, grants, tmp);
2396                                 grants = 0;
2397                         }
2398                 }
2399                 spin_unlock(&cli->cl_loi_list_lock);
2400                 rc = 0;
2401         } else if (ext != NULL) {
2402                 /* index is located outside of active extent */
2403                 need_release = 1;
2404         }
2405         if (need_release) {
2406                 osc_extent_release(env, ext);
2407                 oio->oi_active = NULL;
2408                 ext = NULL;
2409         }
2410
2411         if (ext == NULL) {
2412                 tmp = (1 << cli->cl_chunkbits) + cli->cl_grant_extent_tax;
2413
2414                 /* try to find new extent to cover this page */
2415                 LASSERT(oio->oi_active == NULL);
2416                 /* we may have allocated grant for this page if we failed
2417                  * to expand the previous active extent. */
2418                 LASSERT(ergo(grants > 0, grants >= tmp));
2419
2420                 rc = 0;
2421                 if (grants == 0) {
2422                         /* We haven't allocated grant for this page, and we
2423                          * must not hold a page lock while we do enter_cache,
2424                          * so we must mark dirty & unlock any pages in the
2425                          * write commit pagevec. */
2426                         if (pagevec_count(pvec)) {
2427                                 cb(env, io, pvec);
2428                                 pagevec_reinit(pvec);
2429                         }
2430                         rc = osc_enter_cache(env, cli, oap, tmp);
2431                         if (rc == 0)
2432                                 grants = tmp;
2433                 }
2434
2435                 tmp = grants;
2436                 if (rc == 0) {
2437                         ext = osc_extent_find(env, osc, index, &tmp);
2438                         if (IS_ERR(ext)) {
2439                                 LASSERT(tmp == grants);
2440                                 osc_exit_cache(cli, oap);
2441                                 rc = PTR_ERR(ext);
2442                                 ext = NULL;
2443                         } else {
2444                                 oio->oi_active = ext;
2445                         }
2446                 }
2447                 if (grants > 0)
2448                         osc_unreserve_grant(cli, grants, tmp);
2449         }
2450
2451         LASSERT(ergo(rc == 0, ext != NULL));
2452         if (ext != NULL) {
2453                 EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
2454                          ext, "index = %lu.\n", index);
2455                 LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
2456
2457                 osc_object_lock(osc);
2458                 if (ext->oe_nr_pages == 0)
2459                         ext->oe_srvlock = ops->ops_srvlock;
2460                 else
2461                         LASSERT(ext->oe_srvlock == ops->ops_srvlock);
2462                 ++ext->oe_nr_pages;
2463                 list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
2464                 osc_object_unlock(osc);
2465
2466                 if (!ext->oe_layout_version)
2467                         ext->oe_layout_version = io->ci_layout_version;
2468         }
2469
2470         RETURN(rc);
2471 }
2472
2473 int osc_teardown_async_page(const struct lu_env *env,
2474                             struct osc_object *obj, struct osc_page *ops)
2475 {
2476         struct osc_async_page *oap = &ops->ops_oap;
2477         int rc = 0;
2478         ENTRY;
2479
2480         LASSERT(oap->oap_magic == OAP_MAGIC);
2481
2482         CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
2483                oap, ops, osc_index(oap2osc(oap)));
2484
2485         if (!list_empty(&oap->oap_rpc_item)) {
2486                 CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
2487                 rc = -EBUSY;
2488         } else if (!list_empty(&oap->oap_pending_item)) {
2489                 struct osc_extent *ext = NULL;
2490
2491                 osc_object_lock(obj);
2492                 ext = osc_extent_lookup(obj, osc_index(oap2osc(oap)));
2493                 osc_object_unlock(obj);
2494                 /* only truncated pages are allowed to be taken out.
2495                  * See osc_extent_truncate() and osc_cache_truncate_start()
2496                  * for details. */
2497                 if (ext != NULL && ext->oe_state != OES_TRUNC) {
2498                         OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
2499                                         osc_index(oap2osc(oap)));
2500                         rc = -EBUSY;
2501                 }
2502                 if (ext != NULL)
2503                         osc_extent_put(env, ext);
2504         }
2505         RETURN(rc);
2506 }
2507
2508 /**
2509  * This is called when a page is picked up by kernel to write out.
2510  *
2511  * We should find out the corresponding extent and add the whole extent
2512  * into urgent list. The extent may be being truncated or used, handle it
2513  * carefully.
2514  */
2515 int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
2516                          struct osc_page *ops)
2517 {
2518         struct osc_extent *ext   = NULL;
2519         struct osc_object *obj   = cl2osc(ops->ops_cl.cpl_obj);
2520         struct cl_page    *cp    = ops->ops_cl.cpl_page;
2521         pgoff_t            index = osc_index(ops);
2522         struct osc_async_page *oap = &ops->ops_oap;
2523         bool unplug = false;
2524         int rc = 0;
2525         ENTRY;
2526
2527         osc_object_lock(obj);
2528         ext = osc_extent_lookup(obj, index);
2529         if (ext == NULL) {
2530                 osc_extent_tree_dump(D_ERROR, obj);
2531                 LASSERTF(0, "page index %lu is NOT covered.\n", index);
2532         }
2533
2534         switch (ext->oe_state) {
2535         case OES_RPC:
2536         case OES_LOCK_DONE:
2537                 CL_PAGE_DEBUG(D_ERROR, env, cp, "flush an in-rpc page?\n");
2538                 LASSERT(0);
2539                 break;
2540         case OES_LOCKING:
2541                 /* If we know this extent is being written out, we should abort
2542                  * so that the writer can make this page ready. Otherwise, there
2543                  * exists a deadlock problem because other process can wait for
2544                  * page writeback bit holding page lock; and meanwhile in
2545                  * vvp_page_make_ready(), we need to grab page lock before
2546                  * really sending the RPC. */
2547         case OES_TRUNC:
2548                 /* race with truncate, page will be redirtied */
2549         case OES_ACTIVE:
2550                 /* The extent is active so we need to abort and let the caller
2551                  * re-dirty the page. If we continued on here, and we were the
2552                  * one making the extent active, we could deadlock waiting for
2553                  * the page writeback to clear but it won't because the extent
2554                  * is active and won't be written out. */
2555                 GOTO(out, rc = -EAGAIN);
2556         default:
2557                 break;
2558         }
2559
2560         rc = cl_page_prep(env, io, cp, CRT_WRITE);
2561         if (rc)
2562                 GOTO(out, rc);
2563
2564         spin_lock(&oap->oap_lock);
2565         oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
2566         spin_unlock(&oap->oap_lock);
2567
2568         if (current->flags & PF_MEMALLOC)
2569                 ext->oe_memalloc = 1;
2570
2571         ext->oe_urgent = 1;
2572         if (ext->oe_state == OES_CACHE) {
2573                 OSC_EXTENT_DUMP(D_CACHE, ext,
2574                                 "flush page %p make it urgent.\n", oap);
2575                 if (list_empty(&ext->oe_link))
2576                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2577                 unplug = true;
2578         }
2579         rc = 0;
2580         EXIT;
2581
2582 out:
2583         osc_object_unlock(obj);
2584         osc_extent_put(env, ext);
2585         if (unplug)
2586                 osc_io_unplug_async(env, osc_cli(obj), obj);
2587         return rc;
2588 }
2589
2590 int osc_queue_sync_pages(const struct lu_env *env, const struct cl_io *io,
2591                          struct osc_object *obj, struct list_head *list,
2592                          int brw_flags)
2593 {
2594         struct client_obd     *cli = osc_cli(obj);
2595         struct osc_extent     *ext;
2596         struct osc_async_page *oap;
2597         int     page_count = 0;
2598         int     mppr       = cli->cl_max_pages_per_rpc;
2599         bool    can_merge   = true;
2600         pgoff_t start      = CL_PAGE_EOF;
2601         pgoff_t end        = 0;
2602         ENTRY;
2603
2604         list_for_each_entry(oap, list, oap_pending_item) {
2605                 struct osc_page *opg = oap2osc_page(oap);
2606                 pgoff_t index = osc_index(opg);
2607
2608                 if (index > end)
2609                         end = index;
2610                 if (index < start)
2611                         start = index;
2612                 ++page_count;
2613                 mppr <<= (page_count > mppr);
2614
2615                 if (unlikely(opg->ops_from > 0 ||
2616                              opg->ops_to < PAGE_SIZE - 1))
2617                         can_merge = false;
2618         }
2619
2620         ext = osc_extent_alloc(obj);
2621         if (ext == NULL) {
2622                 struct osc_async_page *tmp;
2623
2624                 list_for_each_entry_safe(oap, tmp, list, oap_pending_item) {
2625                         list_del_init(&oap->oap_pending_item);
2626                         osc_ap_completion(env, cli, oap, 0, -ENOMEM);
2627                 }
2628                 RETURN(-ENOMEM);
2629         }
2630
2631         ext->oe_rw = !!(brw_flags & OBD_BRW_READ);
2632         ext->oe_sync = 1;
2633         ext->oe_no_merge = !can_merge;
2634         ext->oe_urgent = 1;
2635         ext->oe_start = start;
2636         ext->oe_end = ext->oe_max_end = end;
2637         ext->oe_obj = obj;
2638         ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
2639         ext->oe_ndelay = !!(brw_flags & OBD_BRW_NDELAY);
2640         ext->oe_dio = !!(brw_flags & OBD_BRW_NOCACHE);
2641         if (ext->oe_dio && !ext->oe_rw) { /* direct io write */
2642                 int grants;
2643                 int ppc;
2644
2645                 ppc = 1 << (cli->cl_chunkbits - PAGE_SHIFT);
2646                 grants = cli->cl_grant_extent_tax;
2647                 grants += (1 << cli->cl_chunkbits) *
2648                         ((page_count + ppc - 1) / ppc);
2649
2650                 spin_lock(&cli->cl_loi_list_lock);
2651                 if (osc_reserve_grant(cli, grants) == 0) {
2652                         list_for_each_entry(oap, list, oap_pending_item) {
2653                                 osc_consume_write_grant(cli,
2654                                                         &oap->oap_brw_page);
2655                                 atomic_long_inc(&obd_dirty_pages);
2656                         }
2657                         osc_unreserve_grant_nolock(cli, grants, 0);
2658                         ext->oe_grants = grants;
2659                 }
2660                 spin_unlock(&cli->cl_loi_list_lock);
2661         }
2662
2663         ext->oe_is_rdma_only = !!(brw_flags & OBD_BRW_RDMA_ONLY);
2664         ext->oe_nr_pages = page_count;
2665         ext->oe_mppr = mppr;
2666         list_splice_init(list, &ext->oe_pages);
2667         ext->oe_layout_version = io->ci_layout_version;
2668
2669         osc_object_lock(obj);
2670         /* Reuse the initial refcount for RPC, don't drop it */
2671         osc_extent_state_set(ext, OES_LOCK_DONE);
2672         if (!ext->oe_rw) { /* write */
2673                 if (!ext->oe_srvlock && !ext->oe_dio) {
2674                         /* The most likely case here is from lack of grants
2675                          * so we are either out of quota or out of space.
2676                          * Since this means we are holding locks across
2677                          * potentially multi-striped IO, we must send out
2678                          * everything out instantly to avoid prolonged
2679                          * waits resulting in lock eviction (likely since
2680                          * the extended wait in osc_cache_enter() did not
2681                          * yield any additional grant due to a timeout.
2682                          * LU-13131 */
2683                         ext->oe_hp = 1;
2684                         list_add_tail(&ext->oe_link, &obj->oo_hp_exts);
2685                 } else {
2686                         list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
2687                 }
2688                 osc_update_pending(obj, OBD_BRW_WRITE, page_count);
2689         } else {
2690                 list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
2691                 osc_update_pending(obj, OBD_BRW_READ, page_count);
2692         }
2693         osc_object_unlock(obj);
2694
2695         osc_io_unplug_async(env, cli, obj);
2696         RETURN(0);
2697 }
2698
2699 /**
2700  * Called by osc_io_setattr_start() to freeze and destroy covering extents.
2701  */
2702 int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
2703                              __u64 size, struct osc_extent **extp)
2704 {
2705         struct client_obd *cli = osc_cli(obj);
2706         struct osc_extent *ext;
2707         struct osc_extent *waiting = NULL;
2708         pgoff_t index;
2709         LIST_HEAD(list);
2710         int result = 0;
2711         bool partial;
2712         ENTRY;
2713
2714         /* pages with index greater or equal to index will be truncated. */
2715         index = cl_index(osc2cl(obj), size);
2716         partial = size > cl_offset(osc2cl(obj), index);
2717
2718 again:
2719         osc_object_lock(obj);
2720         ext = osc_extent_search(obj, index);
2721         if (ext == NULL)
2722                 ext = first_extent(obj);
2723         else if (ext->oe_end < index)
2724                 ext = next_extent(ext);
2725         while (ext != NULL) {
2726                 EASSERT(ext->oe_state != OES_TRUNC, ext);
2727
2728                 if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
2729                         /* if ext is in urgent state, it means there must exist
2730                          * a page already having been flushed by write_page().
2731                          * We have to wait for this extent because we can't
2732                          * truncate that page. */
2733                         OSC_EXTENT_DUMP(D_CACHE, ext,
2734                                         "waiting for busy extent\n");
2735                         waiting = osc_extent_get(ext);
2736                         break;
2737                 }
2738
2739                 OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:%llu.\n", size);
2740
2741                 osc_extent_get(ext);
2742                 if (ext->oe_state == OES_ACTIVE) {
2743                         /* though we grab inode mutex for write path, but we
2744                          * release it before releasing extent(in osc_io_end()),
2745                          * so there is a race window that an extent is still
2746                          * in OES_ACTIVE when truncate starts. */
2747                         LASSERT(!ext->oe_trunc_pending);
2748                         ext->oe_trunc_pending = 1;
2749                 } else {
2750                         EASSERT(ext->oe_state == OES_CACHE, ext);
2751                         osc_extent_state_set(ext, OES_TRUNC);
2752                         osc_update_pending(obj, OBD_BRW_WRITE,
2753                                            -ext->oe_nr_pages);
2754                 }
2755                 /* This extent could be on the full extents list, that's OK */
2756                 EASSERT(!ext->oe_hp && !ext->oe_urgent, ext);
2757                 if (!list_empty(&ext->oe_link))
2758                         list_move_tail(&ext->oe_link, &list);
2759                 else
2760                         list_add_tail(&ext->oe_link, &list);
2761
2762                 ext = next_extent(ext);
2763         }
2764         osc_object_unlock(obj);
2765
2766         osc_list_maint(cli, obj);
2767
2768         while (!list_empty(&list)) {
2769                 int rc;
2770
2771                 ext = list_entry(list.next, struct osc_extent, oe_link);
2772                 list_del_init(&ext->oe_link);
2773
2774                 /* extent may be in OES_ACTIVE state because inode mutex
2775                  * is released before osc_io_end() in file write case */
2776                 if (ext->oe_state != OES_TRUNC)
2777                         osc_extent_wait(env, ext, OES_TRUNC);
2778
2779                 rc = osc_extent_truncate(ext, index, partial);
2780                 if (rc < 0) {
2781                         if (result == 0)
2782                                 result = rc;
2783
2784                         OSC_EXTENT_DUMP(D_ERROR, ext,
2785                                         "truncate error %d\n", rc);
2786                 } else if (ext->oe_nr_pages == 0) {
2787                         osc_extent_remove(ext);
2788                 } else {
2789                         /* this must be an overlapped extent which means only
2790                          * part of pages in this extent have been truncated.
2791                          */
2792                         EASSERTF(ext->oe_start <= index, ext,
2793                                  "trunc index = %lu/%d.\n", index, partial);
2794                         /* fix index to skip this partially truncated extent */
2795                         index = ext->oe_end + 1;
2796                         partial = false;
2797
2798                         /* we need to hold this extent in OES_TRUNC state so
2799                          * that no writeback will happen. This is to avoid
2800                          * BUG 17397.
2801                          * Only partial truncate can reach here, if @size is
2802                          * not zero, the caller should provide a valid @extp. */
2803                         LASSERT(*extp == NULL);
2804                         *extp = osc_extent_get(ext);
2805                         OSC_EXTENT_DUMP(D_CACHE, ext,
2806                                         "trunc at %llu\n", size);
2807                 }
2808                 osc_extent_put(env, ext);
2809         }
2810         if (waiting != NULL) {
2811                 int rc;
2812
2813                 /* ignore the result of osc_extent_wait the write initiator
2814                  * should take care of it. */
2815                 rc = osc_extent_wait(env, waiting, OES_INV);
2816                 if (rc < 0)
2817                         OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
2818
2819                 osc_extent_put(env, waiting);
2820                 waiting = NULL;
2821                 goto again;
2822         }
2823         RETURN(result);
2824 }
2825 EXPORT_SYMBOL(osc_cache_truncate_start);
2826
2827 /**
2828  * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
2829  */
2830 void osc_cache_truncate_end(const struct lu_env *env, struct osc_extent *ext)
2831 {
2832         if (ext != NULL) {
2833                 struct osc_object *obj = ext->oe_obj;
2834                 bool unplug = false;
2835
2836                 EASSERT(ext->oe_nr_pages > 0, ext);
2837                 EASSERT(ext->oe_state == OES_TRUNC, ext);
2838                 EASSERT(!ext->oe_urgent, ext);
2839
2840                 OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
2841                 osc_object_lock(obj);
2842                 osc_extent_state_set(ext, OES_CACHE);
2843                 if (ext->oe_fsync_wait && !ext->oe_urgent) {
2844                         ext->oe_urgent = 1;
2845                         list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
2846                         unplug = true;
2847                 }
2848                 osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
2849                 osc_object_unlock(obj);
2850                 osc_extent_put(env, ext);
2851
2852                 if (unplug)
2853                         osc_io_unplug_async(env, osc_cli(obj), obj);
2854         }
2855 }
2856
2857 /**
2858  * Wait for extents in a specific range to be written out.
2859  * The caller must have called osc_cache_writeback_range() to issue IO
2860  * otherwise it will take a long time for this function to finish.
2861  *
2862  * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
2863  * nobody else can dirty this range of file while we're waiting for
2864  * extents to be written.
2865  */
2866 int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
2867                          pgoff_t start, pgoff_t end)
2868 {
2869         struct osc_extent *ext;
2870         pgoff_t index = start;
2871         int     result = 0;
2872         ENTRY;
2873
2874 again:
2875         osc_object_lock(obj);
2876         ext = osc_extent_search(obj, index);
2877         if (ext == NULL)
2878                 ext = first_extent(obj);
2879         else if (ext->oe_end < index)
2880                 ext = next_extent(ext);
2881         while (ext != NULL) {
2882                 int rc;
2883
2884                 if (ext->oe_start > end)
2885                         break;
2886
2887                 if (!ext->oe_fsync_wait) {
2888                         ext = next_extent(ext);
2889                         continue;
2890                 }
2891
2892                 EASSERT(ergo(ext->oe_state == OES_CACHE,
2893                              ext->oe_hp || ext->oe_urgent), ext);
2894                 EASSERT(ergo(ext->oe_state == OES_ACTIVE,
2895                              !ext->oe_hp && ext->oe_urgent), ext);
2896
2897                 index = ext->oe_end + 1;
2898                 osc_extent_get(ext);
2899                 osc_object_unlock(obj);
2900
2901                 rc = osc_extent_wait(env, ext, OES_INV);
2902                 if (result == 0)
2903                         result = rc;
2904                 osc_extent_put(env, ext);
2905                 goto again;
2906         }
2907         osc_object_unlock(obj);
2908
2909         OSC_IO_DEBUG(obj, "sync file range.\n");
2910         RETURN(result);
2911 }
2912 EXPORT_SYMBOL(osc_cache_wait_range);
2913
2914 /**
2915  * Called to write out a range of osc object.
2916  *
2917  * @hp     : should be set this is caused by lock cancel;
2918  * @discard: is set if dirty pages should be dropped - file will be deleted or
2919  *         truncated, this implies there is no partially discarding extents.
2920  *
2921  * Return how many pages will be issued, or error code if error occurred.
2922  */
2923 int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
2924                               pgoff_t start, pgoff_t end, int hp, int discard)
2925 {
2926         struct osc_extent *ext;
2927         LIST_HEAD(discard_list);
2928         bool unplug = false;
2929         int result = 0;
2930         ENTRY;
2931
2932         osc_object_lock(obj);
2933         ext = osc_extent_search(obj, start);
2934         if (ext == NULL)
2935                 ext = first_extent(obj);
2936         else if (ext->oe_end < start)
2937                 ext = next_extent(ext);
2938         while (ext != NULL) {
2939                 if (ext->oe_start > end)
2940                         break;
2941
2942                 ext->oe_fsync_wait = 1;
2943                 switch (ext->oe_state) {
2944                 case OES_CACHE:
2945                         result += ext->oe_nr_pages;
2946                         if (!discard) {
2947                                 struct list_head *list = NULL;
2948                                 if (hp) {
2949                                         EASSERT(!ext->oe_hp, ext);
2950                                         ext->oe_hp = 1;
2951                                         list = &obj->oo_hp_exts;
2952                                 } else if (!ext->oe_urgent && !ext->oe_hp) {
2953                                         ext->oe_urgent = 1;
2954                                         list = &obj->oo_urgent_exts;
2955                                 }
2956                                 if (list != NULL)
2957                                         list_move_tail(&ext->oe_link, list);
2958                                 unplug = true;
2959                         } else {
2960                                 struct client_obd *cli = osc_cli(obj);
2961                                 int pcc_bits = cli->cl_chunkbits - PAGE_SHIFT;
2962                                 pgoff_t align_by = (1 << pcc_bits);
2963                                 pgoff_t a_start = round_down(start, align_by);
2964                                 pgoff_t a_end = round_up(end, align_by);
2965
2966                                 /* overflow case */
2967                                 if (end && !a_end)
2968                                         a_end = CL_PAGE_EOF;
2969                                 /* the only discarder is lock cancelling, so
2970                                  * [start, end], aligned by chunk size, must
2971                                  * contain this extent */
2972                                 LASSERTF(ext->oe_start >= a_start &&
2973                                          ext->oe_end <= a_end,
2974                                          "ext [%lu, %lu] reg [%lu, %lu] "
2975                                          "orig [%lu %lu] align %lu bits "
2976                                          "%d\n", ext->oe_start, ext->oe_end,
2977                                          a_start, a_end, start, end,
2978                                          align_by, pcc_bits);
2979                                 osc_extent_state_set(ext, OES_LOCKING);
2980                                 ext->oe_owner = current;
2981                                 list_move_tail(&ext->oe_link,
2982                                                    &discard_list);
2983                                 osc_update_pending(obj, OBD_BRW_WRITE,
2984                                                    -ext->oe_nr_pages);
2985                         }
2986                         break;
2987                 case OES_ACTIVE:
2988                         /* It's pretty bad to wait for ACTIVE extents, because
2989                          * we don't know how long we will wait for it to be
2990                          * flushed since it may be blocked at awaiting more
2991                          * grants. We do this for the correctness of fsync. */
2992                         LASSERT(hp == 0 && discard == 0);
2993                         ext->oe_urgent = 1;
2994                         break;
2995                 case OES_TRUNC:
2996                         /* this extent is being truncated, can't do anything
2997                          * for it now. it will be set to urgent after truncate
2998                          * is finished in osc_cache_truncate_end(). */
2999                 default:
3000                         break;
3001                 }
3002                 ext = next_extent(ext);
3003         }
3004         osc_object_unlock(obj);
3005
3006         LASSERT(ergo(!discard, list_empty(&discard_list)));
3007         if (!list_empty(&discard_list)) {
3008                 struct osc_extent *tmp;
3009                 int rc;
3010
3011                 osc_list_maint(osc_cli(obj), obj);
3012                 list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
3013                         list_del_init(&ext->oe_link);
3014                         EASSERT(ext->oe_state == OES_LOCKING, ext);
3015
3016                         /* Discard caching pages. We don't actually write this
3017                          * extent out but we complete it as if we did. */
3018                         rc = osc_extent_make_ready(env, ext);
3019                         if (unlikely(rc < 0)) {
3020                                 OSC_EXTENT_DUMP(D_ERROR, ext,
3021                                                 "make_ready returned %d\n", rc);
3022                                 if (result >= 0)
3023                                         result = rc;
3024                         }
3025
3026                         /* finish the extent as if the pages were sent */
3027                         osc_extent_finish(env, ext, 0, 0);
3028                 }
3029         }
3030
3031         if (unplug)
3032                 osc_io_unplug(env, osc_cli(obj), obj);
3033
3034         if (hp || discard) {
3035                 int rc;
3036                 rc = osc_cache_wait_range(env, obj, start, end);
3037                 if (result >= 0 && rc < 0)
3038                         result = rc;
3039         }
3040
3041         OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
3042         RETURN(result);
3043 }
3044 EXPORT_SYMBOL(osc_cache_writeback_range);
3045
3046 /**
3047  * Returns a list of pages by a given [start, end] of \a obj.
3048  *
3049  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
3050  * crucial in the face of [offset, EOF] locks.
3051  *
3052  * Return at least one page in @queue unless there is no covered page.
3053  */
3054 bool osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
3055                           struct osc_object *osc, pgoff_t start, pgoff_t end,
3056                           osc_page_gang_cbt cb, void *cbdata)
3057 {
3058         struct osc_page *ops;
3059         struct pagevec  *pagevec;
3060         void            **pvec;
3061         pgoff_t         idx;
3062         unsigned int    nr;
3063         unsigned int    i;
3064         unsigned int    j;
3065         bool            res = true;
3066         bool            tree_lock = true;
3067         ENTRY;
3068
3069         idx = start;
3070         pvec = osc_env_info(env)->oti_pvec;
3071         pagevec = &osc_env_info(env)->oti_pagevec;
3072         ll_pagevec_init(pagevec, 0);
3073         spin_lock(&osc->oo_tree_lock);
3074         while ((nr = radix_tree_gang_lookup(&osc->oo_tree, pvec,
3075                                             idx, OTI_PVEC_SIZE)) > 0) {
3076                 struct cl_page *page;
3077                 bool end_of_region = false;
3078
3079                 for (i = 0, j = 0; i < nr; ++i) {
3080                         ops = pvec[i];
3081                         pvec[i] = NULL;
3082
3083                         idx = osc_index(ops);
3084                         if (idx > end) {
3085                                 end_of_region = true;
3086                                 break;
3087                         }
3088
3089                         page = ops->ops_cl.cpl_page;
3090                         LASSERT(page->cp_type == CPT_CACHEABLE);
3091                         if (page->cp_state == CPS_FREEING)
3092                                 continue;
3093
3094                         cl_page_get(page);
3095                         lu_ref_add_atomic(&page->cp_reference,
3096                                           "gang_lookup", current);
3097                         pvec[j++] = ops;
3098                 }
3099                 ++idx;
3100
3101                 /*
3102                  * Here a delicate locking dance is performed. Current thread
3103                  * holds a reference to a page, but has to own it before it
3104                  * can be placed into queue. Owning implies waiting, so
3105                  * radix-tree lock is to be released. After a wait one has to
3106                  * check that pages weren't truncated (cl_page_own() returns
3107                  * error in the latter case).
3108                  */
3109                 spin_unlock(&osc->oo_tree_lock);
3110                 tree_lock = false;
3111
3112                 for (i = 0; i < j; ++i) {
3113                         ops = pvec[i];
3114                         if (res)
3115                                 res = (*cb)(env, io, ops, cbdata);
3116
3117                         page = ops->ops_cl.cpl_page;
3118                         lu_ref_del(&page->cp_reference, "gang_lookup", current);
3119                         cl_pagevec_put(env, page, pagevec);
3120                 }
3121                 pagevec_release(pagevec);
3122
3123                 if (nr < OTI_PVEC_SIZE || end_of_region)
3124                         break;
3125
3126                 if (!res)
3127                         break;
3128                 if (need_resched())
3129                         cond_resched();
3130
3131                 spin_lock(&osc->oo_tree_lock);
3132                 tree_lock = true;
3133         }
3134         if (tree_lock)
3135                 spin_unlock(&osc->oo_tree_lock);
3136         RETURN(res);
3137 }
3138 EXPORT_SYMBOL(osc_page_gang_lookup);
3139
3140 /**
3141  * Check if page @page is covered by an extra lock or discard it.
3142  */
3143 static bool check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
3144                                 struct osc_page *ops, void *cbdata)
3145 {
3146         struct osc_thread_info *info = osc_env_info(env);
3147         struct osc_object *osc = cbdata;
3148         struct cl_page *page = ops->ops_cl.cpl_page;
3149         pgoff_t index;
3150         bool discard = false;
3151
3152         index = osc_index(ops);
3153
3154         /* negative lock caching */
3155         if (index < info->oti_ng_index) {
3156                 discard = true;
3157         } else if (index >= info->oti_fn_index) {
3158                 struct ldlm_lock *tmp;
3159                 /* refresh non-overlapped index */
3160                 tmp = osc_dlmlock_at_pgoff(env, osc, index,
3161                                            OSC_DAP_FL_TEST_LOCK |
3162                                            OSC_DAP_FL_AST | OSC_DAP_FL_RIGHT);
3163                 if (tmp != NULL) {
3164                         __u64 end = tmp->l_policy_data.l_extent.end;
3165                         __u64 start = tmp->l_policy_data.l_extent.start;
3166
3167                         /* no lock covering this page */
3168                         if (index < cl_index(osc2cl(osc), start)) {
3169                                 /* no lock at @index, first lock at @start */
3170                                 info->oti_ng_index = cl_index(osc2cl(osc),
3171                                                      start);
3172                                 discard = true;
3173                         } else {
3174                                 /* Cache the first-non-overlapped index so as to
3175                                  * skip all pages within [index, oti_fn_index).
3176                                  * This is safe because if tmp lock is canceled,
3177                                  * it will discard these pages.
3178                                  */
3179                                 info->oti_fn_index = cl_index(osc2cl(osc),
3180                                                      end + 1);
3181                                 if (end == OBD_OBJECT_EOF)
3182                                         info->oti_fn_index = CL_PAGE_EOF;
3183                         }
3184                         LDLM_LOCK_PUT(tmp);
3185                 } else {
3186                         info->oti_ng_index = CL_PAGE_EOF;
3187                         discard = true;
3188                 }
3189         }
3190
3191         if (discard) {
3192                 if (cl_page_own(env, io, page) == 0) {
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
3200         info->oti_next_index = index + 1;
3201         return true;
3202 }
3203
3204 bool osc_discard_cb(const struct lu_env *env, struct cl_io *io,
3205                    struct osc_page *ops, void *cbdata)
3206 {
3207         struct osc_thread_info *info = osc_env_info(env);
3208         struct cl_page *page = ops->ops_cl.cpl_page;
3209
3210         /* page is top page. */
3211         info->oti_next_index = osc_index(ops) + 1;
3212         if (cl_page_own(env, io, page) == 0) {
3213                 if (!ergo(page->cp_type == CPT_CACHEABLE,
3214                           !PageDirty(cl_page_vmpage(page))))
3215                         CL_PAGE_DEBUG(D_ERROR, env, page,
3216                                         "discard dirty page?\n");
3217
3218                 /* discard the page */
3219                 cl_page_discard(env, io, page);
3220                 cl_page_disown(env, io, page);
3221         } else {
3222                 LASSERT(page->cp_state == CPS_FREEING);
3223         }
3224
3225         return true;
3226 }
3227 EXPORT_SYMBOL(osc_discard_cb);
3228
3229 /**
3230  * Discard pages protected by the given lock. This function traverses radix
3231  * tree to find all covering pages and discard them. If a page is being covered
3232  * by other locks, it should remain in cache.
3233  *
3234  * If error happens on any step, the process continues anyway (the reasoning
3235  * behind this being that lock cancellation cannot be delayed indefinitely).
3236  */
3237 int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
3238                            pgoff_t start, pgoff_t end, bool discard)
3239 {
3240         struct osc_thread_info *info = osc_env_info(env);
3241         struct cl_io *io = osc_env_thread_io(env);
3242         osc_page_gang_cbt cb;
3243         int result;
3244
3245         ENTRY;
3246
3247         io->ci_obj = cl_object_top(osc2cl(osc));
3248         io->ci_ignore_layout = 1;
3249         result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
3250         if (result != 0)
3251                 GOTO(out, result);
3252
3253         cb = discard ? osc_discard_cb : check_and_discard_cb;
3254         info->oti_fn_index = info->oti_next_index = start;
3255         info->oti_ng_index = 0;
3256
3257         osc_page_gang_lookup(env, io, osc,
3258                              info->oti_next_index, end, cb, osc);
3259 out:
3260         cl_io_fini(env, io);
3261         RETURN(result);
3262 }
3263
3264
3265 /** @} osc */