Whamcloud - gitweb
LU-2139 osc: Use SOFT_SYNC to urge server commit
[fs/lustre-release.git] / lustre / osc / osc_cache.c
index 9f12ad6..4d97b3e 100644 (file)
@@ -1,6 +1,4 @@
-/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
- * vim:expandtab:shiftwidth=8:tabstop=8:
- *
+/*
  * GPL HEADER START
  *
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
@@ -29,7 +27,7 @@
  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  *
- * Copyright (c) 2011 Whamcloud, Inc.
+ * Copyright (c) 2012, 2013, Intel Corporation.
  *
  */
 /*
  * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  */
 
-#define DEBUG_SUBSYSTEM S_OSC
+#define DEBUG_SUBSYSTEM S_OSC
+
+#include "osc_cl_internal.h"
+#include "osc_internal.h"
+
+static int extent_debug; /* set it to be true for more debug */
+
+static void osc_update_pending(struct osc_object *obj, int cmd, int delta);
+static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
+                          int state);
+static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
+                             struct osc_async_page *oap, int sent, int rc);
+static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
+                         int cmd);
+static int osc_refresh_count(const struct lu_env *env,
+                            struct osc_async_page *oap, int cmd);
+static int osc_io_unplug_async(const struct lu_env *env,
+                              struct client_obd *cli, struct osc_object *osc);
+static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
+                          unsigned int lost_grant);
+
+static void osc_extent_tree_dump0(int level, struct osc_object *obj,
+                                 const char *func, int line);
+#define osc_extent_tree_dump(lvl, obj) \
+       osc_extent_tree_dump0(lvl, obj, __func__, __LINE__)
+
+/** \addtogroup osc
+ *  @{
+ */
+
+/* ------------------ osc extent ------------------ */
+static inline char *ext_flags(struct osc_extent *ext, char *flags)
+{
+       char *buf = flags;
+       *buf++ = ext->oe_rw ? 'r' : 'w';
+       if (ext->oe_intree)
+               *buf++ = 'i';
+       if (ext->oe_srvlock)
+               *buf++ = 's';
+       if (ext->oe_hp)
+               *buf++ = 'h';
+       if (ext->oe_urgent)
+               *buf++ = 'u';
+       if (ext->oe_memalloc)
+               *buf++ = 'm';
+       if (ext->oe_trunc_pending)
+               *buf++ = 't';
+       if (ext->oe_fsync_wait)
+               *buf++ = 'Y';
+       *buf = 0;
+       return flags;
+}
+
+static inline char list_empty_marker(cfs_list_t *list)
+{
+       return cfs_list_empty(list) ? '-' : '+';
+}
+
+#define EXTSTR       "[%lu -> %lu/%lu]"
+#define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
+static const char *oes_strings[] = {
+       "inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
+
+#define OSC_EXTENT_DUMP(lvl, extent, fmt, ...) do {                          \
+       struct osc_extent *__ext = (extent);                                  \
+       char __buf[16];                                                       \
+                                                                             \
+       CDEBUG(lvl,                                                           \
+               "extent %p@{" EXTSTR ", "                                     \
+               "[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt,          \
+               /* ----- extent part 0 ----- */                               \
+               __ext, EXTPARA(__ext),                                        \
+               /* ----- part 1 ----- */                                      \
+               cfs_atomic_read(&__ext->oe_refc),                             \
+               cfs_atomic_read(&__ext->oe_users),                            \
+               list_empty_marker(&__ext->oe_link),                           \
+               oes_strings[__ext->oe_state], ext_flags(__ext, __buf),        \
+               __ext->oe_obj,                                                \
+               /* ----- part 2 ----- */                                      \
+               __ext->oe_grants, __ext->oe_nr_pages,                         \
+               list_empty_marker(&__ext->oe_pages),                          \
+               waitqueue_active(&__ext->oe_waitq) ? '+' : '-',               \
+               __ext->oe_osclock, __ext->oe_mppr, __ext->oe_owner,           \
+               /* ----- part 4 ----- */                                      \
+               ## __VA_ARGS__);                                              \
+} while (0)
+
+#undef EASSERTF
+#define EASSERTF(expr, ext, fmt, args...) do {                         \
+       if (!(expr)) {                                                  \
+               OSC_EXTENT_DUMP(D_ERROR, (ext), fmt, ##args);           \
+               osc_extent_tree_dump(D_ERROR, (ext)->oe_obj);           \
+               LASSERT(expr);                                          \
+       }                                                               \
+} while (0)
+
+#undef EASSERT
+#define EASSERT(expr, ext) EASSERTF(expr, ext, "\n")
+
+static inline struct osc_extent *rb_extent(struct rb_node *n)
+{
+       if (n == NULL)
+               return NULL;
+
+       return container_of(n, struct osc_extent, oe_node);
+}
+
+static inline struct osc_extent *next_extent(struct osc_extent *ext)
+{
+       if (ext == NULL)
+               return NULL;
+
+       LASSERT(ext->oe_intree);
+       return rb_extent(rb_next(&ext->oe_node));
+}
+
+static inline struct osc_extent *prev_extent(struct osc_extent *ext)
+{
+       if (ext == NULL)
+               return NULL;
+
+       LASSERT(ext->oe_intree);
+       return rb_extent(rb_prev(&ext->oe_node));
+}
+
+static inline struct osc_extent *first_extent(struct osc_object *obj)
+{
+       return rb_extent(rb_first(&obj->oo_root));
+}
+
+/* object must be locked by caller. */
+static int osc_extent_sanity_check0(struct osc_extent *ext,
+                                   const char *func, const int line)
+{
+       struct osc_object *obj = ext->oe_obj;
+       struct osc_async_page *oap;
+       int page_count;
+       int rc = 0;
+
+       if (!osc_object_is_locked(obj))
+               GOTO(out, rc = 9);
+
+       if (ext->oe_state >= OES_STATE_MAX)
+               GOTO(out, rc = 10);
+
+       if (cfs_atomic_read(&ext->oe_refc) <= 0)
+               GOTO(out, rc = 20);
+
+       if (cfs_atomic_read(&ext->oe_refc) < cfs_atomic_read(&ext->oe_users))
+               GOTO(out, rc = 30);
+
+       switch (ext->oe_state) {
+       case OES_INV:
+               if (ext->oe_nr_pages > 0 || !cfs_list_empty(&ext->oe_pages))
+                       GOTO(out, rc = 35);
+               GOTO(out, rc = 0);
+               break;
+       case OES_ACTIVE:
+               if (cfs_atomic_read(&ext->oe_users) == 0)
+                       GOTO(out, rc = 40);
+               if (ext->oe_hp)
+                       GOTO(out, rc = 50);
+               if (ext->oe_fsync_wait && !ext->oe_urgent)
+                       GOTO(out, rc = 55);
+               break;
+       case OES_CACHE:
+               if (ext->oe_grants == 0)
+                       GOTO(out, rc = 60);
+               if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
+                       GOTO(out, rc = 65);
+       default:
+               if (cfs_atomic_read(&ext->oe_users) > 0)
+                       GOTO(out, rc = 70);
+       }
+
+       if (ext->oe_max_end < ext->oe_end || ext->oe_end < ext->oe_start)
+               GOTO(out, rc = 80);
+
+       if (ext->oe_osclock == NULL && ext->oe_grants > 0)
+               GOTO(out, rc = 90);
+
+       if (ext->oe_osclock) {
+               struct cl_lock_descr *descr;
+               descr = &ext->oe_osclock->cll_descr;
+               if (!(descr->cld_start <= ext->oe_start &&
+                     descr->cld_end >= ext->oe_max_end))
+                       GOTO(out, rc = 100);
+       }
+
+       if (ext->oe_nr_pages > ext->oe_mppr)
+               GOTO(out, rc = 105);
+
+       /* Do not verify page list if extent is in RPC. This is because an
+        * in-RPC extent is supposed to be exclusively accessible w/o lock. */
+       if (ext->oe_state > OES_CACHE)
+               GOTO(out, rc = 0);
+
+       if (!extent_debug)
+               GOTO(out, rc = 0);
+
+       page_count = 0;
+       cfs_list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
+               pgoff_t index = oap2cl_page(oap)->cp_index;
+               ++page_count;
+               if (index > ext->oe_end || index < ext->oe_start)
+                       GOTO(out, rc = 110);
+       }
+       if (page_count != ext->oe_nr_pages)
+               GOTO(out, rc = 120);
+
+out:
+       if (rc != 0)
+               OSC_EXTENT_DUMP(D_ERROR, ext,
+                               "%s:%d sanity check %p failed with rc = %d\n",
+                               func, line, ext, rc);
+       return rc;
+}
+
+#define sanity_check_nolock(ext) \
+       osc_extent_sanity_check0(ext, __func__, __LINE__)
+
+#define sanity_check(ext) ({                                                   \
+       int __res;                                                             \
+       osc_object_lock((ext)->oe_obj);                                        \
+       __res = sanity_check_nolock(ext);                                      \
+       osc_object_unlock((ext)->oe_obj);                                      \
+       __res;                                                                 \
+})
+
+
+/**
+ * sanity check - to make sure there is no overlapped extent in the tree.
+ */
+static int osc_extent_is_overlapped(struct osc_object *obj,
+                                   struct osc_extent *ext)
+{
+       struct osc_extent *tmp;
+
+       LASSERT(osc_object_is_locked(obj));
+
+       if (!extent_debug)
+               return 0;
+
+       for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
+               if (tmp == ext)
+                       continue;
+               if (tmp->oe_end >= ext->oe_start &&
+                   tmp->oe_start <= ext->oe_end)
+                       return 1;
+       }
+       return 0;
+}
+
+static void osc_extent_state_set(struct osc_extent *ext, int state)
+{
+       LASSERT(osc_object_is_locked(ext->oe_obj));
+       LASSERT(state >= OES_INV && state < OES_STATE_MAX);
+
+       /* Never try to sanity check a state changing extent :-) */
+       /* LASSERT(sanity_check_nolock(ext) == 0); */
+
+       /* TODO: validate the state machine */
+       ext->oe_state = state;
+       wake_up_all(&ext->oe_waitq);
+}
+
+static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
+{
+       struct osc_extent *ext;
+
+       OBD_SLAB_ALLOC_PTR_GFP(ext, osc_extent_kmem, GFP_IOFS);
+       if (ext == NULL)
+               return NULL;
+
+       RB_CLEAR_NODE(&ext->oe_node);
+       ext->oe_obj = obj;
+       cfs_atomic_set(&ext->oe_refc, 1);
+       cfs_atomic_set(&ext->oe_users, 0);
+       CFS_INIT_LIST_HEAD(&ext->oe_link);
+       ext->oe_state = OES_INV;
+       CFS_INIT_LIST_HEAD(&ext->oe_pages);
+       init_waitqueue_head(&ext->oe_waitq);
+       ext->oe_osclock = NULL;
+
+       return ext;
+}
+
+static void osc_extent_free(struct osc_extent *ext)
+{
+       OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
+}
+
+static struct osc_extent *osc_extent_get(struct osc_extent *ext)
+{
+       LASSERT(cfs_atomic_read(&ext->oe_refc) >= 0);
+       cfs_atomic_inc(&ext->oe_refc);
+       return ext;
+}
+
+static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
+{
+       LASSERT(cfs_atomic_read(&ext->oe_refc) > 0);
+       if (cfs_atomic_dec_and_test(&ext->oe_refc)) {
+               LASSERT(cfs_list_empty(&ext->oe_link));
+               LASSERT(cfs_atomic_read(&ext->oe_users) == 0);
+               LASSERT(ext->oe_state == OES_INV);
+               LASSERT(!ext->oe_intree);
+
+               if (ext->oe_osclock) {
+                       cl_lock_put(env, ext->oe_osclock);
+                       ext->oe_osclock = NULL;
+               }
+               osc_extent_free(ext);
+       }
+}
+
+/**
+ * osc_extent_put_trust() is a special version of osc_extent_put() when
+ * it's known that the caller is not the last user. This is to address the
+ * problem of lacking of lu_env ;-).
+ */
+static void osc_extent_put_trust(struct osc_extent *ext)
+{
+       LASSERT(cfs_atomic_read(&ext->oe_refc) > 1);
+       LASSERT(osc_object_is_locked(ext->oe_obj));
+       cfs_atomic_dec(&ext->oe_refc);
+}
+
+/**
+ * Return the extent which includes pgoff @index, or return the greatest
+ * previous extent in the tree.
+ */
+static struct osc_extent *osc_extent_search(struct osc_object *obj,
+                                           pgoff_t index)
+{
+       struct rb_node    *n = obj->oo_root.rb_node;
+       struct osc_extent *tmp, *p = NULL;
+
+       LASSERT(osc_object_is_locked(obj));
+       while (n != NULL) {
+               tmp = rb_extent(n);
+               if (index < tmp->oe_start) {
+                       n = n->rb_left;
+               } else if (index > tmp->oe_end) {
+                       p = rb_extent(n);
+                       n = n->rb_right;
+               } else {
+                       return tmp;
+               }
+       }
+       return p;
+}
+
+/*
+ * Return the extent covering @index, otherwise return NULL.
+ * caller must have held object lock.
+ */
+static struct osc_extent *osc_extent_lookup(struct osc_object *obj,
+                                           pgoff_t index)
+{
+       struct osc_extent *ext;
+
+       ext = osc_extent_search(obj, index);
+       if (ext != NULL && ext->oe_start <= index && index <= ext->oe_end)
+               return osc_extent_get(ext);
+       return NULL;
+}
+
+/* caller must have held object lock. */
+static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
+{
+       struct rb_node   **n      = &obj->oo_root.rb_node;
+       struct rb_node    *parent = NULL;
+       struct osc_extent *tmp;
+
+       LASSERT(ext->oe_intree == 0);
+       LASSERT(ext->oe_obj == obj);
+       LASSERT(osc_object_is_locked(obj));
+       while (*n != NULL) {
+               tmp = rb_extent(*n);
+               parent = *n;
+
+               if (ext->oe_end < tmp->oe_start)
+                       n = &(*n)->rb_left;
+               else if (ext->oe_start > tmp->oe_end)
+                       n = &(*n)->rb_right;
+               else
+                       EASSERTF(0, tmp, EXTSTR, EXTPARA(ext));
+       }
+       rb_link_node(&ext->oe_node, parent, n);
+       rb_insert_color(&ext->oe_node, &obj->oo_root);
+       osc_extent_get(ext);
+       ext->oe_intree = 1;
+}
+
+/* caller must have held object lock. */
+static void osc_extent_erase(struct osc_extent *ext)
+{
+       struct osc_object *obj = ext->oe_obj;
+       LASSERT(osc_object_is_locked(obj));
+       if (ext->oe_intree) {
+               rb_erase(&ext->oe_node, &obj->oo_root);
+               ext->oe_intree = 0;
+               /* rbtree held a refcount */
+               osc_extent_put_trust(ext);
+       }
+}
+
+static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
+{
+       struct osc_object *obj = ext->oe_obj;
+
+       LASSERT(osc_object_is_locked(obj));
+       LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
+       if (ext->oe_state == OES_CACHE) {
+               osc_extent_state_set(ext, OES_ACTIVE);
+               osc_update_pending(obj, OBD_BRW_WRITE, -ext->oe_nr_pages);
+       }
+       cfs_atomic_inc(&ext->oe_users);
+       cfs_list_del_init(&ext->oe_link);
+       return osc_extent_get(ext);
+}
+
+static void __osc_extent_remove(struct osc_extent *ext)
+{
+       LASSERT(osc_object_is_locked(ext->oe_obj));
+       LASSERT(cfs_list_empty(&ext->oe_pages));
+       osc_extent_erase(ext);
+       cfs_list_del_init(&ext->oe_link);
+       osc_extent_state_set(ext, OES_INV);
+       OSC_EXTENT_DUMP(D_CACHE, ext, "destroyed.\n");
+}
+
+static void osc_extent_remove(struct osc_extent *ext)
+{
+       struct osc_object *obj = ext->oe_obj;
+
+       osc_object_lock(obj);
+       __osc_extent_remove(ext);
+       osc_object_unlock(obj);
+}
+
+/**
+ * This function is used to merge extents to get better performance. It checks
+ * if @cur and @victim are contiguous at chunk level.
+ */
+static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
+                           struct osc_extent *victim)
+{
+       struct osc_object *obj = cur->oe_obj;
+       pgoff_t chunk_start;
+       pgoff_t chunk_end;
+       int ppc_bits;
+
+       LASSERT(cur->oe_state == OES_CACHE);
+       LASSERT(osc_object_is_locked(obj));
+       if (victim == NULL)
+               return -EINVAL;
+
+       if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
+               return -EBUSY;
+
+       if (cur->oe_max_end != victim->oe_max_end)
+               return -ERANGE;
+
+       LASSERT(cur->oe_osclock == victim->oe_osclock);
+       ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_CACHE_SHIFT;
+       chunk_start = cur->oe_start >> ppc_bits;
+       chunk_end   = cur->oe_end   >> ppc_bits;
+       if (chunk_start   != (victim->oe_end >> ppc_bits) + 1 &&
+           chunk_end + 1 != victim->oe_start >> ppc_bits)
+               return -ERANGE;
+
+       OSC_EXTENT_DUMP(D_CACHE, victim, "will be merged by %p.\n", cur);
+
+       cur->oe_start     = min(cur->oe_start, victim->oe_start);
+       cur->oe_end       = max(cur->oe_end,   victim->oe_end);
+       cur->oe_grants   += victim->oe_grants;
+       cur->oe_nr_pages += victim->oe_nr_pages;
+       /* only the following bits are needed to merge */
+       cur->oe_urgent   |= victim->oe_urgent;
+       cur->oe_memalloc |= victim->oe_memalloc;
+       cfs_list_splice_init(&victim->oe_pages, &cur->oe_pages);
+       cfs_list_del_init(&victim->oe_link);
+       victim->oe_nr_pages = 0;
+
+       osc_extent_get(victim);
+       __osc_extent_remove(victim);
+       osc_extent_put(env, victim);
+
+       OSC_EXTENT_DUMP(D_CACHE, cur, "after merging %p.\n", victim);
+       return 0;
+}
+
+/**
+ * Drop user count of osc_extent, and unplug IO asynchronously.
+ */
+int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
+{
+       struct osc_object *obj = ext->oe_obj;
+       int rc = 0;
+       ENTRY;
+
+       LASSERT(cfs_atomic_read(&ext->oe_users) > 0);
+       LASSERT(sanity_check(ext) == 0);
+       LASSERT(ext->oe_grants > 0);
+
+       if (cfs_atomic_dec_and_lock(&ext->oe_users, &obj->oo_lock)) {
+               LASSERT(ext->oe_state == OES_ACTIVE);
+               if (ext->oe_trunc_pending) {
+                       /* a truncate process is waiting for this extent.
+                        * This may happen due to a race, check
+                        * osc_cache_truncate_start(). */
+                       osc_extent_state_set(ext, OES_TRUNC);
+                       ext->oe_trunc_pending = 0;
+               } else {
+                       osc_extent_state_set(ext, OES_CACHE);
+                       osc_update_pending(obj, OBD_BRW_WRITE,
+                                          ext->oe_nr_pages);
+
+                       /* try to merge the previous and next extent. */
+                       osc_extent_merge(env, ext, prev_extent(ext));
+                       osc_extent_merge(env, ext, next_extent(ext));
+
+                       if (ext->oe_urgent)
+                               cfs_list_move_tail(&ext->oe_link,
+                                                  &obj->oo_urgent_exts);
+               }
+               osc_object_unlock(obj);
+
+               osc_io_unplug_async(env, osc_cli(obj), obj);
+       }
+       osc_extent_put(env, ext);
+       RETURN(rc);
+}
+
+static inline int overlapped(struct osc_extent *ex1, struct osc_extent *ex2)
+{
+       return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
+}
+
+/**
+ * Find or create an extent which includes @index, core function to manage
+ * extent tree.
+ */
+struct osc_extent *osc_extent_find(const struct lu_env *env,
+                                  struct osc_object *obj, pgoff_t index,
+                                  int *grants)
+
+{
+       struct client_obd *cli = osc_cli(obj);
+       struct cl_lock    *lock;
+       struct osc_extent *cur;
+       struct osc_extent *ext;
+       struct osc_extent *conflict = NULL;
+       struct osc_extent *found = NULL;
+       pgoff_t    chunk;
+       pgoff_t    max_end;
+       int        max_pages; /* max_pages_per_rpc */
+       int        chunksize;
+       int        ppc_bits; /* pages per chunk bits */
+       int        chunk_mask;
+       int        rc;
+       ENTRY;
+
+       cur = osc_extent_alloc(obj);
+       if (cur == NULL)
+               RETURN(ERR_PTR(-ENOMEM));
+
+       lock = cl_lock_at_pgoff(env, osc2cl(obj), index, NULL, 1, 0);
+       LASSERT(lock != NULL);
+       LASSERT(lock->cll_descr.cld_mode >= CLM_WRITE);
+
+       LASSERT(cli->cl_chunkbits >= PAGE_CACHE_SHIFT);
+       ppc_bits   = cli->cl_chunkbits - PAGE_CACHE_SHIFT;
+       chunk_mask = ~((1 << ppc_bits) - 1);
+       chunksize  = 1 << cli->cl_chunkbits;
+       chunk      = index >> ppc_bits;
+
+       /* align end to rpc edge, rpc size may not be a power 2 integer. */
+       max_pages = cli->cl_max_pages_per_rpc;
+       LASSERT((max_pages & ~chunk_mask) == 0);
+       max_end = index - (index % max_pages) + max_pages - 1;
+       max_end = min_t(pgoff_t, max_end, lock->cll_descr.cld_end);
+
+       /* initialize new extent by parameters so far */
+       cur->oe_max_end = max_end;
+       cur->oe_start   = index & chunk_mask;
+       cur->oe_end     = ((index + ~chunk_mask + 1) & chunk_mask) - 1;
+       if (cur->oe_start < lock->cll_descr.cld_start)
+               cur->oe_start = lock->cll_descr.cld_start;
+       if (cur->oe_end > max_end)
+               cur->oe_end = max_end;
+       cur->oe_osclock = lock;
+       cur->oe_grants  = 0;
+       cur->oe_mppr    = max_pages;
+
+       /* grants has been allocated by caller */
+       LASSERTF(*grants >= chunksize + cli->cl_extent_tax,
+                "%u/%u/%u.\n", *grants, chunksize, cli->cl_extent_tax);
+       LASSERTF((max_end - cur->oe_start) < max_pages, EXTSTR, EXTPARA(cur));
+
+restart:
+       osc_object_lock(obj);
+       ext = osc_extent_search(obj, cur->oe_start);
+       if (ext == NULL)
+               ext = first_extent(obj);
+       while (ext != NULL) {
+               loff_t ext_chk_start = ext->oe_start >> ppc_bits;
+               loff_t ext_chk_end   = ext->oe_end   >> ppc_bits;
+
+               LASSERT(sanity_check_nolock(ext) == 0);
+               if (chunk > ext_chk_end + 1)
+                       break;
+
+               /* if covering by different locks, no chance to match */
+               if (lock != ext->oe_osclock) {
+                       EASSERTF(!overlapped(ext, cur), ext,
+                                EXTSTR, EXTPARA(cur));
+
+                       ext = next_extent(ext);
+                       continue;
+               }
+
+               /* discontiguous chunks? */
+               if (chunk + 1 < ext_chk_start) {
+                       ext = next_extent(ext);
+                       continue;
+               }
+
+               /* ok, from now on, ext and cur have these attrs:
+                * 1. covered by the same lock
+                * 2. contiguous at chunk level or overlapping. */
+
+               if (overlapped(ext, cur)) {
+                       /* cur is the minimum unit, so overlapping means
+                        * full contain. */
+                       EASSERTF((ext->oe_start <= cur->oe_start &&
+                                 ext->oe_end >= cur->oe_end),
+                                ext, EXTSTR, EXTPARA(cur));
+
+                       if (ext->oe_state > OES_CACHE || ext->oe_fsync_wait) {
+                               /* for simplicity, we wait for this extent to
+                                * finish before going forward. */
+                               conflict = osc_extent_get(ext);
+                               break;
+                       }
+
+                       found = osc_extent_hold(ext);
+                       break;
+               }
+
+               /* non-overlapped extent */
+               if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
+                       /* we can't do anything for a non OES_CACHE extent, or
+                        * if there is someone waiting for this extent to be
+                        * flushed, try next one. */
+                       ext = next_extent(ext);
+                       continue;
+               }
+
+               /* check if they belong to the same rpc slot before trying to
+                * merge. the extents are not overlapped and contiguous at
+                * chunk level to get here. */
+               if (ext->oe_max_end != max_end) {
+                       /* if they don't belong to the same RPC slot or
+                        * max_pages_per_rpc has ever changed, do not merge. */
+                       ext = next_extent(ext);
+                       continue;
+               }
+
+               /* it's required that an extent must be contiguous at chunk
+                * level so that we know the whole extent is covered by grant
+                * (the pages in the extent are NOT required to be contiguous).
+                * Otherwise, it will be too much difficult to know which
+                * chunks have grants allocated. */
+
+               /* try to do front merge - extend ext's start */
+               if (chunk + 1 == ext_chk_start) {
+                       /* ext must be chunk size aligned */
+                       EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
+
+                       /* pull ext's start back to cover cur */
+                       ext->oe_start   = cur->oe_start;
+                       ext->oe_grants += chunksize;
+                       *grants -= chunksize;
+
+                       found = osc_extent_hold(ext);
+               } else if (chunk == ext_chk_end + 1) {
+                       /* rear merge */
+                       ext->oe_end     = cur->oe_end;
+                       ext->oe_grants += chunksize;
+                       *grants -= chunksize;
+
+                       /* try to merge with the next one because we just fill
+                        * in a gap */
+                       if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
+                               /* we can save extent tax from next extent */
+                               *grants += cli->cl_extent_tax;
+
+                       found = osc_extent_hold(ext);
+               }
+               if (found != NULL)
+                       break;
+
+               ext = next_extent(ext);
+       }
+
+       osc_extent_tree_dump(D_CACHE, obj);
+       if (found != NULL) {
+               LASSERT(conflict == NULL);
+               if (!IS_ERR(found)) {
+                       LASSERT(found->oe_osclock == cur->oe_osclock);
+                       OSC_EXTENT_DUMP(D_CACHE, found,
+                                       "found caching ext for %lu.\n", index);
+               }
+       } else if (conflict == NULL) {
+               /* create a new extent */
+               EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
+               cur->oe_grants = chunksize + cli->cl_extent_tax;
+               *grants -= cur->oe_grants;
+               LASSERT(*grants >= 0);
+
+               cur->oe_state = OES_CACHE;
+               found = osc_extent_hold(cur);
+               osc_extent_insert(obj, cur);
+               OSC_EXTENT_DUMP(D_CACHE, cur, "add into tree %lu/%lu.\n",
+                               index, lock->cll_descr.cld_end);
+       }
+       osc_object_unlock(obj);
+
+       if (conflict != NULL) {
+               LASSERT(found == NULL);
+
+               /* waiting for IO to finish. Please notice that it's impossible
+                * to be an OES_TRUNC extent. */
+               rc = osc_extent_wait(env, conflict, OES_INV);
+               osc_extent_put(env, conflict);
+               conflict = NULL;
+               if (rc < 0)
+                       GOTO(out, found = ERR_PTR(rc));
+
+               goto restart;
+       }
+       EXIT;
+
+out:
+       osc_extent_put(env, cur);
+       LASSERT(*grants >= 0);
+       return found;
+}
+
+/**
+ * Called when IO is finished to an extent.
+ */
+int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
+                     int sent, int rc)
+{
+       struct client_obd *cli = osc_cli(ext->oe_obj);
+       struct osc_async_page *oap;
+       struct osc_async_page *tmp;
+       int nr_pages = ext->oe_nr_pages;
+       int lost_grant = 0;
+       int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
+       __u64 last_off = 0;
+       int last_count = -1;
+       ENTRY;
+
+       OSC_EXTENT_DUMP(D_CACHE, ext, "extent finished.\n");
+
+       ext->oe_rc = rc ?: ext->oe_nr_pages;
+       EASSERT(ergo(rc == 0, ext->oe_state == OES_RPC), ext);
+       cfs_list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
+                                    oap_pending_item) {
+               cfs_list_del_init(&oap->oap_rpc_item);
+               cfs_list_del_init(&oap->oap_pending_item);
+               if (last_off <= oap->oap_obj_off) {
+                       last_off = oap->oap_obj_off;
+                       last_count = oap->oap_count;
+               }
+
+               --ext->oe_nr_pages;
+               osc_ap_completion(env, cli, oap, sent, rc);
+       }
+       EASSERT(ext->oe_nr_pages == 0, ext);
+
+       if (!sent) {
+               lost_grant = ext->oe_grants;
+       } else if (blocksize < PAGE_CACHE_SIZE &&
+                  last_count != PAGE_CACHE_SIZE) {
+               /* For short writes we shouldn't count parts of pages that
+                * span a whole chunk on the OST side, or our accounting goes
+                * wrong.  Should match the code in filter_grant_check. */
+               int offset = oap->oap_page_off & ~CFS_PAGE_MASK;
+               int count = oap->oap_count + (offset & (blocksize - 1));
+               int end = (offset + oap->oap_count) & (blocksize - 1);
+               if (end)
+                       count += blocksize - end;
+
+               lost_grant = PAGE_CACHE_SIZE - count;
+       }
+       if (ext->oe_grants > 0)
+               osc_free_grant(cli, nr_pages, lost_grant);
+
+       osc_extent_remove(ext);
+       /* put the refcount for RPC */
+       osc_extent_put(env, ext);
+       RETURN(0);
+}
+
+static int extent_wait_cb(struct osc_extent *ext, int state)
+{
+       int ret;
+
+       osc_object_lock(ext->oe_obj);
+       ret = ext->oe_state == state;
+       osc_object_unlock(ext->oe_obj);
+
+       return ret;
+}
+
+/**
+ * Wait for the extent's state to become @state.
+ */
+static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
+                          int state)
+{
+       struct osc_object *obj = ext->oe_obj;
+       struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
+                                                 LWI_ON_SIGNAL_NOOP, NULL);
+       int rc = 0;
+       ENTRY;
+
+       osc_object_lock(obj);
+       LASSERT(sanity_check_nolock(ext) == 0);
+       /* `Kick' this extent only if the caller is waiting for it to be
+        * written out. */
+       if (state == OES_INV && !ext->oe_urgent && !ext->oe_hp) {
+               if (ext->oe_state == OES_ACTIVE) {
+                       ext->oe_urgent = 1;
+               } else if (ext->oe_state == OES_CACHE) {
+                       ext->oe_urgent = 1;
+                       osc_extent_hold(ext);
+                       rc = 1;
+               }
+       }
+       osc_object_unlock(obj);
+       if (rc == 1)
+               osc_extent_release(env, ext);
+
+       /* wait for the extent until its state becomes @state */
+       rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
+       if (rc == -ETIMEDOUT) {
+               OSC_EXTENT_DUMP(D_ERROR, ext,
+                       "%s: wait ext to %d timedout, recovery in progress?\n",
+                       osc_export(obj)->exp_obd->obd_name, state);
+
+               lwi = LWI_INTR(NULL, NULL);
+               rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state),
+                                 &lwi);
+       }
+       if (rc == 0 && ext->oe_rc < 0)
+               rc = ext->oe_rc;
+       RETURN(rc);
+}
+
+/**
+ * Discard pages with index greater than @size. If @ext is overlapped with
+ * @size, then partial truncate happens.
+ */
+static int osc_extent_truncate(struct osc_extent *ext, pgoff_t trunc_index,
+                               bool partial)
+{
+       struct cl_env_nest     nest;
+       struct lu_env         *env;
+       struct cl_io          *io;
+       struct osc_object     *obj = ext->oe_obj;
+       struct client_obd     *cli = osc_cli(obj);
+       struct osc_async_page *oap;
+       struct osc_async_page *tmp;
+       int                    pages_in_chunk = 0;
+       int                    ppc_bits    = cli->cl_chunkbits -
+                                            PAGE_CACHE_SHIFT;
+       __u64                  trunc_chunk = trunc_index >> ppc_bits;
+       int                    grants   = 0;
+       int                    nr_pages = 0;
+       int                    rc       = 0;
+       ENTRY;
+
+       LASSERT(sanity_check(ext) == 0);
+       LASSERT(ext->oe_state == OES_TRUNC);
+       LASSERT(!ext->oe_urgent);
+
+       /* Request new lu_env.
+        * We can't use that env from osc_cache_truncate_start() because
+        * it's from lov_io_sub and not fully initialized. */
+       env = cl_env_nested_get(&nest);
+       io  = &osc_env_info(env)->oti_io;
+       io->ci_obj = cl_object_top(osc2cl(obj));
+       rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
+       if (rc < 0)
+               GOTO(out, rc);
+
+       /* discard all pages with index greater then trunc_index */
+       cfs_list_for_each_entry_safe(oap, tmp, &ext->oe_pages,
+                                    oap_pending_item) {
+               struct cl_page  *sub  = oap2cl_page(oap);
+               struct cl_page  *page = cl_page_top(sub);
+
+               LASSERT(cfs_list_empty(&oap->oap_rpc_item));
+
+               /* only discard the pages with their index greater than
+                * trunc_index, and ... */
+               if (sub->cp_index < trunc_index ||
+                   (sub->cp_index == trunc_index && partial)) {
+                       /* accounting how many pages remaining in the chunk
+                        * so that we can calculate grants correctly. */
+                       if (sub->cp_index >> ppc_bits == trunc_chunk)
+                               ++pages_in_chunk;
+                       continue;
+               }
+
+               cfs_list_del_init(&oap->oap_pending_item);
+
+               cl_page_get(page);
+               lu_ref_add(&page->cp_reference, "truncate", current);
+
+               if (cl_page_own(env, io, page) == 0) {
+                       cl_page_unmap(env, io, page);
+                       cl_page_discard(env, io, page);
+                       cl_page_disown(env, io, page);
+               } else {
+                       LASSERT(page->cp_state == CPS_FREEING);
+                       LASSERT(0);
+               }
+
+               lu_ref_del(&page->cp_reference, "truncate", current);
+               cl_page_put(env, page);
+
+               --ext->oe_nr_pages;
+               ++nr_pages;
+       }
+       EASSERTF(ergo(ext->oe_start >= trunc_index + !!partial,
+                     ext->oe_nr_pages == 0),
+               ext, "trunc_index %lu, partial %d\n", trunc_index, partial);
+
+       osc_object_lock(obj);
+       if (ext->oe_nr_pages == 0) {
+               LASSERT(pages_in_chunk == 0);
+               grants = ext->oe_grants;
+               ext->oe_grants = 0;
+       } else { /* calculate how many grants we can free */
+               int     chunks = (ext->oe_end >> ppc_bits) - trunc_chunk;
+               pgoff_t last_index;
+
+
+               /* if there is no pages in this chunk, we can also free grants
+                * for the last chunk */
+               if (pages_in_chunk == 0) {
+                       /* if this is the 1st chunk and no pages in this chunk,
+                        * ext->oe_nr_pages must be zero, so we should be in
+                        * the other if-clause. */
+                       LASSERT(trunc_chunk > 0);
+                       --trunc_chunk;
+                       ++chunks;
+               }
+
+               /* this is what we can free from this extent */
+               grants          = chunks << cli->cl_chunkbits;
+               ext->oe_grants -= grants;
+               last_index      = ((trunc_chunk + 1) << ppc_bits) - 1;
+               ext->oe_end     = min(last_index, ext->oe_max_end);
+               LASSERT(ext->oe_end >= ext->oe_start);
+               LASSERT(ext->oe_grants > 0);
+       }
+       osc_object_unlock(obj);
+
+       if (grants > 0 || nr_pages > 0)
+               osc_free_grant(cli, nr_pages, grants);
+
+out:
+       cl_io_fini(env, io);
+       cl_env_nested_put(&nest, env);
+       RETURN(rc);
+}
+
+/**
+ * This function is used to make the extent prepared for transfer.
+ * A race with flusing page - ll_writepage() has to be handled cautiously.
+ */
+static int osc_extent_make_ready(const struct lu_env *env,
+                                struct osc_extent *ext)
+{
+       struct osc_async_page *oap;
+       struct osc_async_page *last = NULL;
+       struct osc_object *obj = ext->oe_obj;
+       int page_count = 0;
+       int rc;
+       ENTRY;
+
+       /* we're going to grab page lock, so object lock must not be taken. */
+       LASSERT(sanity_check(ext) == 0);
+       /* in locking state, any process should not touch this extent. */
+       EASSERT(ext->oe_state == OES_LOCKING, ext);
+       EASSERT(ext->oe_owner != NULL, ext);
+
+       OSC_EXTENT_DUMP(D_CACHE, ext, "make ready\n");
+
+       cfs_list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
+               ++page_count;
+               if (last == NULL || last->oap_obj_off < oap->oap_obj_off)
+                       last = oap;
+
+               /* checking ASYNC_READY is race safe */
+               if ((oap->oap_async_flags & ASYNC_READY) != 0)
+                       continue;
+
+               rc = osc_make_ready(env, oap, OBD_BRW_WRITE);
+               switch (rc) {
+               case 0:
+                       spin_lock(&oap->oap_lock);
+                       oap->oap_async_flags |= ASYNC_READY;
+                       spin_unlock(&oap->oap_lock);
+                       break;
+               case -EALREADY:
+                       LASSERT((oap->oap_async_flags & ASYNC_READY) != 0);
+                       break;
+               default:
+                       LASSERTF(0, "unknown return code: %d\n", rc);
+               }
+       }
+
+       LASSERT(page_count == ext->oe_nr_pages);
+       LASSERT(last != NULL);
+       /* the last page is the only one we need to refresh its count by
+        * the size of file. */
+       if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
+               last->oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
+               LASSERT(last->oap_count > 0);
+               LASSERT(last->oap_page_off + last->oap_count <= PAGE_CACHE_SIZE);
+               spin_lock(&last->oap_lock);
+               last->oap_async_flags |= ASYNC_COUNT_STABLE;
+               spin_unlock(&last->oap_lock);
+       }
+
+       /* for the rest of pages, we don't need to call osf_refresh_count()
+        * because it's known they are not the last page */
+       cfs_list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
+               if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
+                       oap->oap_count = PAGE_CACHE_SIZE - oap->oap_page_off;
+                       spin_lock(&oap->oap_lock);
+                       oap->oap_async_flags |= ASYNC_COUNT_STABLE;
+                       spin_unlock(&oap->oap_lock);
+               }
+       }
+
+       osc_object_lock(obj);
+       osc_extent_state_set(ext, OES_RPC);
+       osc_object_unlock(obj);
+       /* get a refcount for RPC. */
+       osc_extent_get(ext);
+
+       RETURN(0);
+}
+
+/**
+ * Quick and simple version of osc_extent_find(). This function is frequently
+ * called to expand the extent for the same IO. To expand the extent, the
+ * page index must be in the same or next chunk of ext->oe_end.
+ */
+static int osc_extent_expand(struct osc_extent *ext, pgoff_t index, int *grants)
+{
+       struct osc_object *obj = ext->oe_obj;
+       struct client_obd *cli = osc_cli(obj);
+       struct osc_extent *next;
+       int ppc_bits = cli->cl_chunkbits - PAGE_CACHE_SHIFT;
+       pgoff_t chunk = index >> ppc_bits;
+       pgoff_t end_chunk;
+       pgoff_t end_index;
+       int chunksize = 1 << cli->cl_chunkbits;
+       int rc = 0;
+       ENTRY;
+
+       LASSERT(ext->oe_max_end >= index && ext->oe_start <= index);
+       osc_object_lock(obj);
+       LASSERT(sanity_check_nolock(ext) == 0);
+       end_chunk = ext->oe_end >> ppc_bits;
+       if (chunk > end_chunk + 1)
+               GOTO(out, rc = -ERANGE);
+
+       if (end_chunk >= chunk)
+               GOTO(out, rc = 0);
+
+       LASSERT(end_chunk + 1 == chunk);
+       /* try to expand this extent to cover @index */
+       end_index = min(ext->oe_max_end, ((chunk + 1) << ppc_bits) - 1);
+
+       next = next_extent(ext);
+       if (next != NULL && next->oe_start <= end_index)
+               /* complex mode - overlapped with the next extent,
+                * this case will be handled by osc_extent_find() */
+               GOTO(out, rc = -EAGAIN);
+
+       ext->oe_end = end_index;
+       ext->oe_grants += chunksize;
+       *grants -= chunksize;
+       LASSERT(*grants >= 0);
+       EASSERTF(osc_extent_is_overlapped(obj, ext) == 0, ext,
+                "overlapped after expanding for %lu.\n", index);
+       EXIT;
 
-#include "osc_cl_internal.h"
-#include "osc_internal.h"
+out:
+       osc_object_unlock(obj);
+       RETURN(rc);
+}
 
-static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
-                          struct osc_async_page *oap);
-static int osc_enter_cache_try(const struct lu_env *env, struct client_obd *cli,
-                              struct osc_async_page *oap, int transient);
-static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
-                          int sent);
+static void osc_extent_tree_dump0(int level, struct osc_object *obj,
+                                 const char *func, int line)
+{
+       struct osc_extent *ext;
+       int cnt;
 
-/** \addtogroup osc
- *  @{
- */
+       CDEBUG(level, "Dump object %p extents at %s:%d, mppr: %u.\n",
+              obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
+
+       /* osc_object_lock(obj); */
+       cnt = 1;
+       for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
+               OSC_EXTENT_DUMP(level, ext, "in tree %d.\n", cnt++);
+
+       cnt = 1;
+       cfs_list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
+               OSC_EXTENT_DUMP(level, ext, "hp %d.\n", cnt++);
 
-#define OSC_IO_DEBUG(OSC, STR, args...)                                  \
-       CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR,     \
-              !cfs_list_empty(&(OSC)->oo_ready_item) ||                 \
-              !cfs_list_empty(&(OSC)->oo_hp_ready_item),               \
-              (OSC)->oo_write_pages.oop_num_pending,               \
-              !cfs_list_empty(&(OSC)->oo_write_pages.oop_urgent),       \
-              (OSC)->oo_read_pages.oop_num_pending,                 \
-              !cfs_list_empty(&(OSC)->oo_read_pages.oop_urgent),       \
-              args)
+       cnt = 1;
+       cfs_list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
+               OSC_EXTENT_DUMP(level, ext, "urgent %d.\n", cnt++);
 
-static inline struct osc_page *oap2osc_page(struct osc_async_page *oap)
+       cnt = 1;
+       cfs_list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
+               OSC_EXTENT_DUMP(level, ext, "reading %d.\n", cnt++);
+       /* osc_object_unlock(obj); */
+}
+
+/* ------------------ osc extent end ------------------ */
+
+static inline int osc_is_ready(struct osc_object *osc)
 {
-       return (struct osc_page *)container_of(oap, struct osc_page, ops_oap);
+       return !cfs_list_empty(&osc->oo_ready_item) ||
+              !cfs_list_empty(&osc->oo_hp_ready_item);
 }
 
+#define OSC_IO_DEBUG(OSC, STR, args...)                                               \
+       CDEBUG(D_CACHE, "obj %p ready %d|%c|%c wr %d|%c|%c rd %d|%c " STR,     \
+              (OSC), osc_is_ready(OSC),                                       \
+              list_empty_marker(&(OSC)->oo_hp_ready_item),                    \
+              list_empty_marker(&(OSC)->oo_ready_item),                       \
+              cfs_atomic_read(&(OSC)->oo_nr_writes),                          \
+              list_empty_marker(&(OSC)->oo_hp_exts),                          \
+              list_empty_marker(&(OSC)->oo_urgent_exts),                      \
+              cfs_atomic_read(&(OSC)->oo_nr_reads),                           \
+              list_empty_marker(&(OSC)->oo_reading_exts),                     \
+              ##args)
+
 static int osc_make_ready(const struct lu_env *env, struct osc_async_page *oap,
                          int cmd)
 {
        struct osc_page *opg  = oap2osc_page(oap);
-       struct cl_page  *page = cl_page_top(opg->ops_cl.cpl_page);
+       struct cl_page  *page = cl_page_top(oap2cl_page(oap));
        int result;
 
        LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
@@ -92,7 +1223,7 @@ static int osc_refresh_count(const struct lu_env *env,
                             struct osc_async_page *oap, int cmd)
 {
        struct osc_page  *opg = oap2osc_page(oap);
-       struct cl_page   *page;
+       struct cl_page   *page = oap2cl_page(oap);
        struct cl_object *obj;
        struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
 
@@ -102,7 +1233,6 @@ static int osc_refresh_count(const struct lu_env *env,
        /* readpage queues with _COUNT_STABLE, shouldn't get here. */
        LASSERT(!(cmd & OBD_BRW_READ));
        LASSERT(opg != NULL);
-       page = opg->ops_cl.cpl_page;
        obj = opg->ops_cl.cpl_obj;
 
        cl_object_attr_lock(obj);
@@ -116,16 +1246,16 @@ static int osc_refresh_count(const struct lu_env *env,
                return 0;
        else if (cl_offset(obj, page->cp_index + 1) > kms)
                /* catch sub-page write at end of file */
-               return kms % CFS_PAGE_SIZE;
+               return kms % PAGE_CACHE_SIZE;
        else
-               return CFS_PAGE_SIZE;
+               return PAGE_CACHE_SIZE;
 }
 
 static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
-                         int cmd, struct obdo *oa, int rc)
+                         int cmd, int rc)
 {
        struct osc_page   *opg  = oap2osc_page(oap);
-       struct cl_page    *page = cl_page_top(opg->ops_cl.cpl_page);
+       struct cl_page    *page = cl_page_top(oap2cl_page(oap));
        struct osc_object *obj  = cl2osc(opg->ops_cl.cpl_obj);
        enum cl_req_type   crt;
        int srvlock;
@@ -133,8 +1263,10 @@ static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
        ENTRY;
 
        cmd &= ~OBD_BRW_NOQUOTA;
-       LASSERT(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ));
-       LASSERT(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE));
+       LASSERTF(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ),
+                "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
+       LASSERTF(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE),
+               "cp_state:%u, cmd:%d\n", page->cp_state, cmd);
        LASSERT(opg->ops_transfer_pinned);
 
        /*
@@ -145,26 +1277,20 @@ static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
                cl_req_page_done(env, page);
        LASSERT(page->cp_req == NULL);
 
-       /* As the transfer for this page is being done, clear the flags */
-       cfs_spin_lock(&oap->oap_lock);
-       oap->oap_async_flags = 0;
-       cfs_spin_unlock(&oap->oap_lock);
-
        crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
        /* Clear opg->ops_transfer_pinned before VM lock is released. */
        opg->ops_transfer_pinned = 0;
 
-       cfs_spin_lock(&obj->oo_seatbelt);
+       spin_lock(&obj->oo_seatbelt);
        LASSERT(opg->ops_submitter != NULL);
        LASSERT(!cfs_list_empty(&opg->ops_inflight));
        cfs_list_del_init(&opg->ops_inflight);
-       cfs_spin_unlock(&obj->oo_seatbelt);
+       opg->ops_submitter = NULL;
+       spin_unlock(&obj->oo_seatbelt);
 
        opg->ops_submit_time = 0;
        srvlock = oap->oap_brw_flags & OBD_BRW_SRVLOCK;
 
-       cl_page_completion(env, page, crt, rc);
-
        /* statistic */
        if (rc == 0 && srvlock) {
                struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
@@ -183,40 +1309,51 @@ static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
         * reference counter protects page from concurrent reclaim.
         */
        lu_ref_del(&page->cp_reference, "transfer", page);
-       /*
-        * As page->cp_obj is pinned by a reference from page->cp_req, it is
-        * safe to call cl_page_put() without risking object destruction in a
-        * non-blocking context.
-        */
-       cl_page_put(env, page);
+
+       cl_page_completion(env, page, crt, rc);
+
        RETURN(0);
 }
 
+#define OSC_DUMP_GRANT(lvl, cli, fmt, args...) do {                          \
+       struct client_obd *__tmp = (cli);                                     \
+       CDEBUG(lvl, "%s: grant { dirty: %ld/%ld dirty_pages: %d/%d "          \
+              "unstable_pages: %d/%d dropped: %ld avail: %ld, "              \
+              "reserved: %ld, flight: %d } lru {in list: %d, "               \
+              "left: %d, waiters: %d }" fmt,                                 \
+              __tmp->cl_import->imp_obd->obd_name,                           \
+              __tmp->cl_dirty, __tmp->cl_dirty_max,                          \
+              cfs_atomic_read(&obd_dirty_pages), obd_max_dirty_pages,        \
+              cfs_atomic_read(&obd_unstable_pages), obd_max_dirty_pages,     \
+              __tmp->cl_lost_grant, __tmp->cl_avail_grant,                   \
+              __tmp->cl_reserved_grant, __tmp->cl_w_in_flight,               \
+              cfs_atomic_read(&__tmp->cl_lru_in_list),                       \
+              cfs_atomic_read(&__tmp->cl_lru_busy),                          \
+              cfs_atomic_read(&__tmp->cl_lru_shrinkers), ##args);            \
+} while (0)
+
 /* caller must hold loi_list_lock */
 static void osc_consume_write_grant(struct client_obd *cli,
                                    struct brw_page *pga)
 {
-       LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
+       LASSERT(spin_is_locked(&cli->cl_loi_list_lock.lock));
        LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
        cfs_atomic_inc(&obd_dirty_pages);
-       cli->cl_dirty += CFS_PAGE_SIZE;
-       cli->cl_avail_grant -= CFS_PAGE_SIZE;
+       cli->cl_dirty += PAGE_CACHE_SIZE;
        pga->flag |= OBD_BRW_FROM_GRANT;
        CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
-              CFS_PAGE_SIZE, pga, pga->pg);
-       LASSERT(cli->cl_avail_grant >= 0);
+              PAGE_CACHE_SIZE, pga, pga->pg);
        osc_update_next_shrink(cli);
 }
 
 /* the companion to osc_consume_write_grant, called when a brw has completed.
  * must be called with the loi lock held. */
 static void osc_release_write_grant(struct client_obd *cli,
-                                   struct brw_page *pga, int sent)
+                                   struct brw_page *pga)
 {
-       int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
        ENTRY;
 
-       LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
+       LASSERT(spin_is_locked(&cli->cl_loi_list_lock.lock));
        if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
                EXIT;
                return;
@@ -224,103 +1361,179 @@ static void osc_release_write_grant(struct client_obd *cli,
 
        pga->flag &= ~OBD_BRW_FROM_GRANT;
        cfs_atomic_dec(&obd_dirty_pages);
-       cli->cl_dirty -= CFS_PAGE_SIZE;
+       cli->cl_dirty -= PAGE_CACHE_SIZE;
        if (pga->flag & OBD_BRW_NOCACHE) {
                pga->flag &= ~OBD_BRW_NOCACHE;
                cfs_atomic_dec(&obd_dirty_transit_pages);
-               cli->cl_dirty_transit -= CFS_PAGE_SIZE;
+               cli->cl_dirty_transit -= PAGE_CACHE_SIZE;
        }
-       if (!sent) {
-               /* Reclaim grant from truncated pages. This is used to solve
-                * write-truncate and grant all gone(to lost_grant) problem.
-                * For a vfs write this problem can be easily solved by a sync
-                * write, however, this is not an option for page_mkwrite()
-                * because grant has to be allocated before a page becomes
-                * dirty. */
-               if (cli->cl_avail_grant < PTLRPC_MAX_BRW_SIZE)
-                       cli->cl_avail_grant += CFS_PAGE_SIZE;
-               else
-                       cli->cl_lost_grant += CFS_PAGE_SIZE;
-               CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
-                      cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
-       } else if (CFS_PAGE_SIZE != blocksize && pga->count != CFS_PAGE_SIZE) {
-               /* For short writes we shouldn't count parts of pages that
-                * span a whole block on the OST side, or our accounting goes
-                * wrong.  Should match the code in filter_grant_check. */
-               int offset = pga->off & ~CFS_PAGE_MASK;
-               int count = pga->count + (offset & (blocksize - 1));
-               int end = (offset + pga->count) & (blocksize - 1);
-               if (end)
-                       count += blocksize - end;
+       EXIT;
+}
+
+/**
+ * To avoid sleeping with object lock held, it's good for us allocate enough
+ * grants before entering into critical section.
+ *
+ * client_obd_list_lock held by caller
+ */
+static int osc_reserve_grant(struct client_obd *cli, unsigned int bytes)
+{
+       int rc = -EDQUOT;
+
+       if (cli->cl_avail_grant >= bytes) {
+               cli->cl_avail_grant    -= bytes;
+               cli->cl_reserved_grant += bytes;
+               rc = 0;
+       }
+       return rc;
+}
 
-               cli->cl_lost_grant += CFS_PAGE_SIZE - count;
-               CDEBUG(D_CACHE, "lost %lu grant: %lu avail: %lu dirty: %lu\n",
-                      CFS_PAGE_SIZE - count, cli->cl_lost_grant,
-                      cli->cl_avail_grant, cli->cl_dirty);
+static void __osc_unreserve_grant(struct client_obd *cli,
+                                 unsigned int reserved, unsigned int unused)
+{
+       /* it's quite normal for us to get more grant than reserved.
+        * Thinking about a case that two extents merged by adding a new
+        * chunk, we can save one extent tax. If extent tax is greater than
+        * one chunk, we can save more grant by adding a new chunk */
+       cli->cl_reserved_grant -= reserved;
+       if (unused > reserved) {
+               cli->cl_avail_grant += reserved;
+               cli->cl_lost_grant  += unused - reserved;
+       } else {
+               cli->cl_avail_grant += unused;
        }
+}
 
-       EXIT;
+void osc_unreserve_grant(struct client_obd *cli,
+                        unsigned int reserved, unsigned int unused)
+{
+       client_obd_list_lock(&cli->cl_loi_list_lock);
+       __osc_unreserve_grant(cli, reserved, unused);
+       if (unused > 0)
+               osc_wake_cache_waiters(cli);
+       client_obd_list_unlock(&cli->cl_loi_list_lock);
+}
+
+/**
+ * Free grant after IO is finished or canceled.
+ *
+ * @lost_grant is used to remember how many grants we have allocated but not
+ * used, we should return these grants to OST. There're two cases where grants
+ * can be lost:
+ * 1. truncate;
+ * 2. blocksize at OST is less than PAGE_CACHE_SIZE and a partial page was
+ *    written. In this case OST may use less chunks to serve this partial
+ *    write. OSTs don't actually know the page size on the client side. so
+ *    clients have to calculate lost grant by the blocksize on the OST.
+ *    See filter_grant_check() for details.
+ */
+static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
+                          unsigned int lost_grant)
+{
+       int grant = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
+
+       client_obd_list_lock(&cli->cl_loi_list_lock);
+       cfs_atomic_sub(nr_pages, &obd_dirty_pages);
+       cli->cl_dirty -= nr_pages << PAGE_CACHE_SHIFT;
+       cli->cl_lost_grant += lost_grant;
+       if (cli->cl_avail_grant < grant && cli->cl_lost_grant >= grant) {
+               /* borrow some grant from truncate to avoid the case that
+                * truncate uses up all avail grant */
+               cli->cl_lost_grant -= grant;
+               cli->cl_avail_grant += grant;
+       }
+       osc_wake_cache_waiters(cli);
+       client_obd_list_unlock(&cli->cl_loi_list_lock);
+       CDEBUG(D_CACHE, "lost %u grant: %lu avail: %lu dirty: %lu\n",
+              lost_grant, cli->cl_lost_grant,
+              cli->cl_avail_grant, cli->cl_dirty);
 }
 
-/* The companion to osc_enter_cache(), called when @oap is no longer part of
- * the dirty accounting.  Writeback completes or truncate happens before
- * writing starts.  Must be called with the loi lock held. */
-static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
-                          int sent)
+/**
+ * The companion to osc_enter_cache(), called when @oap is no longer part of
+ * the dirty accounting due to error.
+ */
+static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
 {
-       osc_release_write_grant(cli, &oap->oap_brw_page, sent);
+       client_obd_list_lock(&cli->cl_loi_list_lock);
+       osc_release_write_grant(cli, &oap->oap_brw_page);
+       client_obd_list_unlock(&cli->cl_loi_list_lock);
 }
 
 /**
  * Non-blocking version of osc_enter_cache() that consumes grant only when it
  * is available.
  */
-static int osc_enter_cache_try(const struct lu_env *env, struct client_obd *cli,
-                              struct osc_async_page *oap, int transient)
+static int osc_enter_cache_try(struct client_obd *cli,
+                              struct osc_async_page *oap,
+                              int bytes, int transient)
 {
-       int has_grant;
+       int rc;
+
+       OSC_DUMP_GRANT(D_CACHE, cli, "need:%d.\n", bytes);
+
+       rc = osc_reserve_grant(cli, bytes);
+       if (rc < 0)
+               return 0;
 
-       has_grant = cli->cl_avail_grant >= CFS_PAGE_SIZE;
-       if (has_grant) {
+       if (cli->cl_dirty + PAGE_CACHE_SIZE <= cli->cl_dirty_max &&
+           cfs_atomic_read(&obd_unstable_pages) + 1 +
+           cfs_atomic_read(&obd_dirty_pages) <= obd_max_dirty_pages) {
                osc_consume_write_grant(cli, &oap->oap_brw_page);
                if (transient) {
-                       cli->cl_dirty_transit += CFS_PAGE_SIZE;
+                       cli->cl_dirty_transit += PAGE_CACHE_SIZE;
                        cfs_atomic_inc(&obd_dirty_transit_pages);
                        oap->oap_brw_flags |= OBD_BRW_NOCACHE;
                }
+               rc = 1;
+       } else {
+               __osc_unreserve_grant(cli, bytes, bytes);
+               rc = 0;
        }
-       return has_grant;
+       return rc;
+}
+
+static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
+{
+       int rc;
+       client_obd_list_lock(&cli->cl_loi_list_lock);
+       rc = cfs_list_empty(&ocw->ocw_entry);
+       client_obd_list_unlock(&cli->cl_loi_list_lock);
+       return rc;
 }
 
-/* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
- * grant or cache space. */
+/**
+ * The main entry to reserve dirty page accounting. Usually the grant reserved
+ * in this function will be freed in bulk in osc_free_grant() unless it fails
+ * to add osc cache, in that case, it will be freed in osc_exit_cache().
+ *
+ * The process will be put into sleep if it's already run out of grant.
+ */
 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
-                          struct osc_async_page *oap)
+                          struct osc_async_page *oap, int bytes)
 {
        struct osc_object *osc = oap->oap_obj;
        struct lov_oinfo  *loi = osc->oo_oinfo;
        struct osc_cache_waiter ocw;
-       struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
+       struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
+                                                 LWI_ON_SIGNAL_NOOP, NULL);
        int rc = -EDQUOT;
        ENTRY;
 
-       CDEBUG(D_CACHE, "dirty: %ld/%d dirty_max: %ld/%d dropped: %lu "
-              "grant: %lu\n", cli->cl_dirty, cfs_atomic_read(&obd_dirty_pages),
-              cli->cl_dirty_max, obd_max_dirty_pages,
-              cli->cl_lost_grant, cli->cl_avail_grant);
+       OSC_DUMP_GRANT(D_CACHE, cli, "need:%d.\n", bytes);
+
+       client_obd_list_lock(&cli->cl_loi_list_lock);
 
        /* force the caller to try sync io.  this can jump the list
         * of queued writes and create a discontiguous rpc stream */
        if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
-           cli->cl_dirty_max < CFS_PAGE_SIZE     ||
+           cli->cl_dirty_max < PAGE_CACHE_SIZE     ||
            cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync)
-               RETURN(-EDQUOT);
+               GOTO(out, rc = -EDQUOT);
 
        /* Hopefully normal case - cache space and write credits available */
-       if (cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max &&
-           cfs_atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages &&
-           osc_enter_cache_try(env, cli, oap, 0))
-               RETURN(0);
+       if (osc_enter_cache_try(cli, oap, bytes, 0))
+               GOTO(out, rc = 0);
 
        /* We can get here for two reasons: too many dirty pages in cache, or
         * run out of grants. In both cases we should write dirty pages out.
@@ -328,31 +1541,59 @@ static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
         * RPC size will be.
         * The exiting condition is no avail grants and no dirty pages caching,
         * that really means there is no space on the OST. */
-       cfs_waitq_init(&ocw.ocw_waitq);
-       ocw.ocw_oap = oap;
-       while (cli->cl_dirty > 0) {
+       init_waitqueue_head(&ocw.ocw_waitq);
+       ocw.ocw_oap   = oap;
+       ocw.ocw_grant = bytes;
+       while (cli->cl_dirty > 0 || cli->cl_w_in_flight > 0) {
                cfs_list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
                ocw.ocw_rc = 0;
-
-               osc_io_unplug(env, cli, osc, PDL_POLICY_ROUND);
                client_obd_list_unlock(&cli->cl_loi_list_lock);
 
+               osc_io_unplug_async(env, cli, NULL);
+
                CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
                       cli->cl_import->imp_obd->obd_name, &ocw, oap);
 
-               rc = l_wait_event(ocw.ocw_waitq, cfs_list_empty(&ocw.ocw_entry),
-                                 &lwi);
+               rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
 
                client_obd_list_lock(&cli->cl_loi_list_lock);
-               cfs_list_del_init(&ocw.ocw_entry);
-               if (rc < 0)
-                       break;
 
+               /* l_wait_event is interrupted by signal, or timed out */
+               if (rc < 0) {
+                       switch (rc) {
+                       case -ETIMEDOUT:
+                               OSC_DUMP_GRANT(D_ERROR, cli,
+                                               "try to reserve %d.\n", bytes);
+                               osc_extent_tree_dump(D_ERROR, osc);
+                               rc = -EDQUOT;
+                               break;
+                       case -EINTR:
+                               /* Ensures restartability - LU-3581 */
+                               rc = -ERESTARTSYS;
+                               break;
+                       default:
+                               CDEBUG(D_CACHE, "%s: event for cache space @"
+                                      " %p never arrived due to %d\n",
+                                      cli->cl_import->imp_obd->obd_name,
+                                      &ocw, rc);
+                               break;
+                       }
+                       cfs_list_del_init(&ocw.ocw_entry);
+                       GOTO(out, rc);
+               }
+
+               LASSERT(cfs_list_empty(&ocw.ocw_entry));
                rc = ocw.ocw_rc;
+
                if (rc != -EDQUOT)
-                       break;
+                       GOTO(out, rc);
+               if (osc_enter_cache_try(cli, oap, bytes, 0))
+                       GOTO(out, rc = 0);
        }
-
+       EXIT;
+out:
+       client_obd_list_unlock(&cli->cl_loi_list_lock);
+       OSC_DUMP_GRANT(D_CACHE, cli, "returned %d.\n", rc);
        RETURN(rc);
 }
 
@@ -364,39 +1605,29 @@ void osc_wake_cache_waiters(struct client_obd *cli)
 
        ENTRY;
        cfs_list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
-               /* if we can't dirty more, we must wait until some is written */
-               if ((cli->cl_dirty + CFS_PAGE_SIZE > cli->cl_dirty_max) ||
-                  (cfs_atomic_read(&obd_dirty_pages) + 1 >
-                   obd_max_dirty_pages)) {
+               ocw = cfs_list_entry(l, struct osc_cache_waiter, ocw_entry);
+               cfs_list_del_init(&ocw->ocw_entry);
+
+               ocw->ocw_rc = -EDQUOT;
+               /* we can't dirty more */
+               if ((cli->cl_dirty + PAGE_CACHE_SIZE > cli->cl_dirty_max) ||
+                   (cfs_atomic_read(&obd_unstable_pages) + 1 +
+                    cfs_atomic_read(&obd_dirty_pages) > obd_max_dirty_pages)) {
                        CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
                               "osc max %ld, sys max %d\n", cli->cl_dirty,
                               cli->cl_dirty_max, obd_max_dirty_pages);
-                       return;
+                       goto wakeup;
                }
 
-               /* if still dirty cache but no grant wait for pending RPCs that
-                * may yet return us some grant before doing sync writes */
-               if (cli->cl_w_in_flight && cli->cl_avail_grant < CFS_PAGE_SIZE) {
-                       CDEBUG(D_CACHE, "%u BRW writes in flight, no grant\n",
-                              cli->cl_w_in_flight);
-                       return;
-               }
-
-               ocw = cfs_list_entry(l, struct osc_cache_waiter, ocw_entry);
-               cfs_list_del_init(&ocw->ocw_entry);
-               if (cli->cl_avail_grant < CFS_PAGE_SIZE) {
-                       /* no more RPCs in flight to return grant, do sync IO */
+               ocw->ocw_rc = 0;
+               if (!osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
                        ocw->ocw_rc = -EDQUOT;
-                       CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
-               } else {
-                       osc_consume_write_grant(cli,
-                                               &ocw->ocw_oap->oap_brw_page);
-               }
 
-               CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld\n",
-                      ocw, ocw->ocw_oap, cli->cl_avail_grant);
+wakeup:
+               CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld, %d\n",
+                      ocw, ocw->ocw_oap, cli->cl_avail_grant, ocw->ocw_rc);
 
-               cfs_waitq_signal(&ocw->ocw_waitq);
+               wake_up(&ocw->ocw_waitq);
        }
 
        EXIT;
@@ -404,21 +1635,7 @@ void osc_wake_cache_waiters(struct client_obd *cli)
 
 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
 {
-       struct osc_async_page *oap;
-       int hprpc = 0;
-
-       if (!cfs_list_empty(&osc->oo_write_pages.oop_urgent)) {
-               oap = cfs_list_entry(osc->oo_write_pages.oop_urgent.next,
-                                    struct osc_async_page, oap_urgent_item);
-               hprpc = !!(oap->oap_async_flags & ASYNC_HP);
-       }
-
-       if (!hprpc && !cfs_list_empty(&osc->oo_read_pages.oop_urgent)) {
-               oap = cfs_list_entry(osc->oo_read_pages.oop_urgent.next,
-                                    struct osc_async_page, oap_urgent_item);
-               hprpc = !!(oap->oap_async_flags & ASYNC_HP);
-       }
-
+       int hprpc = !!cfs_list_empty(&osc->oo_hp_exts);
        return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
 }
 
@@ -428,35 +1645,31 @@ static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
 static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
                         int cmd)
 {
-       struct osc_oap_pages *lop;
+       int invalid_import = 0;
        ENTRY;
 
-       if (cmd & OBD_BRW_WRITE) {
-               lop = &osc->oo_write_pages;
-       } else {
-               lop = &osc->oo_read_pages;
-       }
-
-       if (lop->oop_num_pending == 0)
-               RETURN(0);
-
        /* if we have an invalid import we want to drain the queued pages
         * by forcing them through rpcs that immediately fail and complete
         * the pages.  recovery relies on this to empty the queued pages
         * before canceling the locks and evicting down the llite pages */
-       if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
-               RETURN(1);
-
-       /* stream rpcs in queue order as long as as there is an urgent page
-        * queued.  this is our cheap solution for good batching in the case
-        * where writepage marks some random page in the middle of the file
-        * as urgent because of, say, memory pressure */
-       if (!cfs_list_empty(&lop->oop_urgent)) {
-               CDEBUG(D_CACHE, "urgent request forcing RPC\n");
-               RETURN(1);
-       }
+       if ((cli->cl_import == NULL || cli->cl_import->imp_invalid))
+               invalid_import = 1;
 
        if (cmd & OBD_BRW_WRITE) {
+               if (cfs_atomic_read(&osc->oo_nr_writes) == 0)
+                       RETURN(0);
+               if (invalid_import) {
+                       CDEBUG(D_CACHE, "invalid import forcing RPC\n");
+                       RETURN(1);
+               }
+               if (!cfs_list_empty(&osc->oo_hp_exts)) {
+                       CDEBUG(D_CACHE, "high prio request forcing RPC\n");
+                       RETURN(1);
+               }
+               if (!cfs_list_empty(&osc->oo_urgent_exts)) {
+                       CDEBUG(D_CACHE, "urgent request forcing RPC\n");
+                       RETURN(1);
+               }
                /* trigger a write rpc stream as long as there are dirtiers
                 * waiting for space.  as they're waiting, they're not going to
                 * create more pages to coalesce with what's waiting.. */
@@ -464,40 +1677,42 @@ static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
                        CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
                        RETURN(1);
                }
+               if (cfs_atomic_read(&osc->oo_nr_writes) >=
+                   cli->cl_max_pages_per_rpc)
+                       RETURN(1);
+       } else {
+               if (cfs_atomic_read(&osc->oo_nr_reads) == 0)
+                       RETURN(0);
+               if (invalid_import) {
+                       CDEBUG(D_CACHE, "invalid import forcing RPC\n");
+                       RETURN(1);
+               }
+               /* all read are urgent. */
+               if (!cfs_list_empty(&osc->oo_reading_exts))
+                       RETURN(1);
        }
-       if (lop->oop_num_pending >= cli->cl_max_pages_per_rpc)
-               RETURN(1);
 
        RETURN(0);
 }
 
-static void lop_update_pending(struct client_obd *cli,
-                              struct osc_oap_pages *lop, int cmd, int delta)
+static void osc_update_pending(struct osc_object *obj, int cmd, int delta)
 {
-       lop->oop_num_pending += delta;
-       if (cmd & OBD_BRW_WRITE)
-               cli->cl_pending_w_pages += delta;
-       else
-               cli->cl_pending_r_pages += delta;
+       struct client_obd *cli = osc_cli(obj);
+       if (cmd & OBD_BRW_WRITE) {
+               cfs_atomic_add(delta, &obj->oo_nr_writes);
+               cfs_atomic_add(delta, &cli->cl_pending_w_pages);
+               LASSERT(cfs_atomic_read(&obj->oo_nr_writes) >= 0);
+       } else {
+               cfs_atomic_add(delta, &obj->oo_nr_reads);
+               cfs_atomic_add(delta, &cli->cl_pending_r_pages);
+               LASSERT(cfs_atomic_read(&obj->oo_nr_reads) >= 0);
+       }
+       OSC_IO_DEBUG(obj, "update pending cmd %d delta %d.\n", cmd, delta);
 }
 
-static int osc_makes_hprpc(struct osc_oap_pages *lop)
+static int osc_makes_hprpc(struct osc_object *obj)
 {
-       struct osc_async_page *oap;
-       ENTRY;
-
-       if (cfs_list_empty(&lop->oop_urgent))
-               RETURN(0);
-
-       oap = cfs_list_entry(lop->oop_urgent.next,
-                        struct osc_async_page, oap_urgent_item);
-
-       if (oap->oap_async_flags & ASYNC_HP) {
-               CDEBUG(D_CACHE, "hp request forcing RPC\n");
-               RETURN(1);
-       }
-
-       RETURN(0);
+       return !cfs_list_empty(&obj->oo_hp_exts);
 }
 
 static void on_list(cfs_list_t *item, cfs_list_t *list, int should_be_on)
@@ -510,10 +1725,9 @@ static void on_list(cfs_list_t *item, cfs_list_t *list, int should_be_on)
 
 /* maintain the osc's cli list membership invariants so that osc_send_oap_rpc
  * can find pages to build into rpcs quickly */
-static void osc_list_maint(struct client_obd *cli, struct osc_object *osc)
+static int __osc_list_maint(struct client_obd *cli, struct osc_object *osc)
 {
-       if (osc_makes_hprpc(&osc->oo_write_pages) ||
-           osc_makes_hprpc(&osc->oo_read_pages)) {
+       if (osc_makes_hprpc(osc)) {
                /* HP rpc */
                on_list(&osc->oo_ready_item, &cli->cl_loi_ready_list, 0);
                on_list(&osc->oo_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
@@ -525,10 +1739,23 @@ static void osc_list_maint(struct client_obd *cli, struct osc_object *osc)
        }
 
        on_list(&osc->oo_write_item, &cli->cl_loi_write_list,
-               osc->oo_write_pages.oop_num_pending);
+               cfs_atomic_read(&osc->oo_nr_writes) > 0);
 
        on_list(&osc->oo_read_item, &cli->cl_loi_read_list,
-               osc->oo_read_pages.oop_num_pending);
+               cfs_atomic_read(&osc->oo_nr_reads) > 0);
+
+       return osc_is_ready(osc);
+}
+
+static int osc_list_maint(struct client_obd *cli, struct osc_object *osc)
+{
+       int is_ready;
+
+       client_obd_list_lock(&cli->cl_loi_list_lock);
+       is_ready = __osc_list_maint(cli, osc);
+       client_obd_list_unlock(&cli->cl_loi_list_lock);
+
+       return is_ready;
 }
 
 /* this is trying to propogate async writeback errors back up to the
@@ -553,29 +1780,95 @@ static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
                ar->ar_force_sync = 0;
 }
 
-static void osc_oap_to_pending(struct osc_async_page *oap)
+/* Performs "unstable" page accounting. This function balances the
+ * increment operations performed in osc_inc_unstable_pages. It is
+ * registered as the RPC request callback, and is executed when the
+ * bulk RPC is committed on the server. Thus at this point, the pages
+ * involved in the bulk transfer are no longer considered unstable. */
+void osc_dec_unstable_pages(struct ptlrpc_request *req)
 {
-       struct osc_object    *osc = oap->oap_obj;
-       struct osc_oap_pages *lop;
+       struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
+       struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
+       obd_count                page_count = desc->bd_iov_count;
+       int i;
 
-       if (oap->oap_cmd & OBD_BRW_WRITE)
-               lop = &osc->oo_write_pages;
-       else
-               lop = &osc->oo_read_pages;
+       /* No unstable page tracking */
+       if (cli->cl_cache == NULL)
+               return;
+
+       LASSERT(page_count >= 0);
+
+       for (i = 0; i < page_count; i++)
+               dec_zone_page_state(desc->bd_iov[i].kiov_page, NR_UNSTABLE_NFS);
+
+       cfs_atomic_sub(page_count, &cli->cl_cache->ccc_unstable_nr);
+       LASSERT(cfs_atomic_read(&cli->cl_cache->ccc_unstable_nr) >= 0);
+
+       cfs_atomic_sub(page_count, &cli->cl_unstable_count);
+       LASSERT(cfs_atomic_read(&cli->cl_unstable_count) >= 0);
+
+       cfs_atomic_sub(page_count, &obd_unstable_pages);
+       LASSERT(cfs_atomic_read(&obd_unstable_pages) >= 0);
+
+       spin_lock(&req->rq_lock);
+       req->rq_committed = 1;
+       req->rq_unstable  = 0;
+       spin_unlock(&req->rq_lock);
+
+       wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
+}
+
+/* "unstable" page accounting. See: osc_dec_unstable_pages. */
+void osc_inc_unstable_pages(struct ptlrpc_request *req)
+{
+       struct ptlrpc_bulk_desc *desc = req->rq_bulk;
+       struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
+       obd_count                page_count = desc->bd_iov_count;
+       int i;
+
+       /* No unstable page tracking */
+       if (cli->cl_cache == NULL)
+               return;
+
+       LASSERT(page_count >= 0);
+
+       for (i = 0; i < page_count; i++)
+               inc_zone_page_state(desc->bd_iov[i].kiov_page, NR_UNSTABLE_NFS);
+
+       LASSERT(cfs_atomic_read(&cli->cl_cache->ccc_unstable_nr) >= 0);
+       cfs_atomic_add(page_count, &cli->cl_cache->ccc_unstable_nr);
+
+       LASSERT(cfs_atomic_read(&cli->cl_unstable_count) >= 0);
+       cfs_atomic_add(page_count, &cli->cl_unstable_count);
+
+       LASSERT(cfs_atomic_read(&obd_unstable_pages) >= 0);
+       cfs_atomic_add(page_count, &obd_unstable_pages);
 
-       if (oap->oap_async_flags & ASYNC_HP)
-               cfs_list_add(&oap->oap_urgent_item, &lop->oop_urgent);
-       else if (oap->oap_async_flags & ASYNC_URGENT)
-               cfs_list_add_tail(&oap->oap_urgent_item, &lop->oop_urgent);
-       cfs_list_add_tail(&oap->oap_pending_item, &lop->oop_pending);
-       lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, 1);
+       spin_lock(&req->rq_lock);
+
+       /* If the request has already been committed (i.e. brw_commit
+        * called via rq_commit_cb), we need to undo the unstable page
+        * increments we just performed because rq_commit_cb wont be
+        * called again. Otherwise, just set the commit callback so the
+        * unstable page accounting is properly updated when the request
+        * is committed */
+       if (req->rq_committed) {
+               /* Drop lock before calling osc_dec_unstable_pages */
+               spin_unlock(&req->rq_lock);
+               osc_dec_unstable_pages(req);
+               spin_lock(&req->rq_lock);
+       } else {
+               req->rq_unstable  = 1;
+               req->rq_commit_cb = osc_dec_unstable_pages;
+       }
+
+       spin_unlock(&req->rq_lock);
 }
 
 /* this must be called holding the loi list lock to give coverage to exit_cache,
  * async_flag maintenance, and oap_request */
-void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
-                      struct obdo *oa, struct osc_async_page *oap,
-                      int sent, int rc)
+static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
+                             struct osc_async_page *oap, int sent, int rc)
 {
        struct osc_object *osc = oap->oap_obj;
        struct lov_oinfo  *loi = osc->oo_oinfo;
@@ -583,43 +1876,213 @@ void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
 
        ENTRY;
        if (oap->oap_request != NULL) {
+               if (rc == 0)
+                       osc_inc_unstable_pages(oap->oap_request);
+
                xid = ptlrpc_req_xid(oap->oap_request);
                ptlrpc_req_finished(oap->oap_request);
                oap->oap_request = NULL;
        }
 
-       cfs_spin_lock(&oap->oap_lock);
-       oap->oap_async_flags = 0;
-       cfs_spin_unlock(&oap->oap_lock);
-       oap->oap_interrupted = 0;
+       /* As the transfer for this page is being done, clear the flags */
+       spin_lock(&oap->oap_lock);
+       oap->oap_async_flags = 0;
+       spin_unlock(&oap->oap_lock);
+       oap->oap_interrupted = 0;
+
+       if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
+               client_obd_list_lock(&cli->cl_loi_list_lock);
+               osc_process_ar(&cli->cl_ar, xid, rc);
+               osc_process_ar(&loi->loi_ar, xid, rc);
+               client_obd_list_unlock(&cli->cl_loi_list_lock);
+       }
+
+       rc = osc_completion(env, oap, oap->oap_cmd, rc);
+       if (rc)
+               CERROR("completion on oap %p obj %p returns %d.\n",
+                      oap, osc, rc);
+
+       EXIT;
+}
+
+/**
+ * Try to add extent to one RPC. We need to think about the following things:
+ * - # of pages must not be over max_pages_per_rpc
+ * - extent must be compatible with previous ones
+ */
+static int try_to_add_extent_for_io(struct client_obd *cli,
+                                   struct osc_extent *ext, cfs_list_t *rpclist,
+                                   int *pc, unsigned int *max_pages)
+{
+       struct osc_extent *tmp;
+       ENTRY;
+
+       EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
+               ext);
+
+       *max_pages = max(ext->oe_mppr, *max_pages);
+       if (*pc + ext->oe_nr_pages > *max_pages)
+               RETURN(0);
+
+       cfs_list_for_each_entry(tmp, rpclist, oe_link) {
+               EASSERT(tmp->oe_owner == current, tmp);
+#if 0
+               if (overlapped(tmp, ext)) {
+                       OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
+                       EASSERT(0, ext);
+               }
+#endif
+
+               if (tmp->oe_srvlock != ext->oe_srvlock ||
+                   !tmp->oe_grants != !ext->oe_grants)
+                       RETURN(0);
+
+               /* remove break for strict check */
+               break;
+       }
+
+       *pc += ext->oe_nr_pages;
+       cfs_list_move_tail(&ext->oe_link, rpclist);
+       ext->oe_owner = current;
+       RETURN(1);
+}
+
+/**
+ * In order to prevent multiple ptlrpcd from breaking contiguous extents,
+ * get_write_extent() takes all appropriate extents in atomic.
+ *
+ * The following policy is used to collect extents for IO:
+ * 1. Add as many HP extents as possible;
+ * 2. Add the first urgent extent in urgent extent list and take it out of
+ *    urgent list;
+ * 3. Add subsequent extents of this urgent extent;
+ * 4. If urgent list is not empty, goto 2;
+ * 5. Traverse the extent tree from the 1st extent;
+ * 6. Above steps exit if there is no space in this RPC.
+ */
+static int get_write_extents(struct osc_object *obj, cfs_list_t *rpclist)
+{
+       struct client_obd *cli = osc_cli(obj);
+       struct osc_extent *ext;
+       int page_count = 0;
+       unsigned int max_pages = cli->cl_max_pages_per_rpc;
+
+       LASSERT(osc_object_is_locked(obj));
+       while (!cfs_list_empty(&obj->oo_hp_exts)) {
+               ext = cfs_list_entry(obj->oo_hp_exts.next, struct osc_extent,
+                                    oe_link);
+               LASSERT(ext->oe_state == OES_CACHE);
+               if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
+                                             &max_pages))
+                       return page_count;
+               EASSERT(ext->oe_nr_pages <= max_pages, ext);
+       }
+       if (page_count == max_pages)
+               return page_count;
+
+       while (!cfs_list_empty(&obj->oo_urgent_exts)) {
+               ext = cfs_list_entry(obj->oo_urgent_exts.next,
+                                    struct osc_extent, oe_link);
+               if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
+                                             &max_pages))
+                       return page_count;
+
+               if (!ext->oe_intree)
+                       continue;
+
+               while ((ext = next_extent(ext)) != NULL) {
+                       if ((ext->oe_state != OES_CACHE) ||
+                           (!cfs_list_empty(&ext->oe_link) &&
+                            ext->oe_owner != NULL))
+                               continue;
+
+                       if (!try_to_add_extent_for_io(cli, ext, rpclist,
+                                                     &page_count, &max_pages))
+                               return page_count;
+               }
+       }
+       if (page_count == max_pages)
+               return page_count;
+
+       ext = first_extent(obj);
+       while (ext != NULL) {
+               if ((ext->oe_state != OES_CACHE) ||
+                   /* this extent may be already in current rpclist */
+                   (!cfs_list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
+                       ext = next_extent(ext);
+                       continue;
+               }
+
+               if (!try_to_add_extent_for_io(cli, ext, rpclist, &page_count,
+                                             &max_pages))
+                       return page_count;
+
+               ext = next_extent(ext);
+       }
+       return page_count;
+}
+
+static int
+osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
+                  struct osc_object *osc, pdl_policy_t pol)
+{
+       CFS_LIST_HEAD(rpclist);
+       struct osc_extent *ext;
+       struct osc_extent *tmp;
+       struct osc_extent *first = NULL;
+       obd_count page_count = 0;
+       int srvlock = 0;
+       int rc = 0;
+       ENTRY;
+
+       LASSERT(osc_object_is_locked(osc));
+
+       page_count = get_write_extents(osc, &rpclist);
+       LASSERT(equi(page_count == 0, cfs_list_empty(&rpclist)));
+
+       if (cfs_list_empty(&rpclist))
+               RETURN(0);
 
-       if (oap->oap_cmd & OBD_BRW_WRITE) {
-               osc_process_ar(&cli->cl_ar, xid, rc);
-               osc_process_ar(&loi->loi_ar, xid, rc);
+       osc_update_pending(osc, OBD_BRW_WRITE, -page_count);
+
+       cfs_list_for_each_entry(ext, &rpclist, oe_link) {
+               LASSERT(ext->oe_state == OES_CACHE ||
+                       ext->oe_state == OES_LOCK_DONE);
+               if (ext->oe_state == OES_CACHE)
+                       osc_extent_state_set(ext, OES_LOCKING);
+               else
+                       osc_extent_state_set(ext, OES_RPC);
        }
 
-       if (rc == 0 && oa != NULL) {
-               if (oa->o_valid & OBD_MD_FLBLOCKS)
-                       loi->loi_lvb.lvb_blocks = oa->o_blocks;
-               if (oa->o_valid & OBD_MD_FLMTIME)
-                       loi->loi_lvb.lvb_mtime = oa->o_mtime;
-               if (oa->o_valid & OBD_MD_FLATIME)
-                       loi->loi_lvb.lvb_atime = oa->o_atime;
-               if (oa->o_valid & OBD_MD_FLCTIME)
-                       loi->loi_lvb.lvb_ctime = oa->o_ctime;
+       /* we're going to grab page lock, so release object lock because
+        * lock order is page lock -> object lock. */
+       osc_object_unlock(osc);
+
+       cfs_list_for_each_entry_safe(ext, tmp, &rpclist, oe_link) {
+               if (ext->oe_state == OES_LOCKING) {
+                       rc = osc_extent_make_ready(env, ext);
+                       if (unlikely(rc < 0)) {
+                               cfs_list_del_init(&ext->oe_link);
+                               osc_extent_finish(env, ext, 0, rc);
+                               continue;
+                       }
+               }
+               if (first == NULL) {
+                       first = ext;
+                       srvlock = ext->oe_srvlock;
+               } else {
+                       LASSERT(srvlock == ext->oe_srvlock);
+               }
        }
 
-       rc = osc_completion(env, oap, oap->oap_cmd, oa, rc);
+       if (!cfs_list_empty(&rpclist)) {
+               LASSERT(page_count > 0);
+               rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_WRITE, pol);
+               LASSERT(cfs_list_empty(&rpclist));
+       }
 
-       /* cl_page_completion() drops PG_locked. so, a new I/O on the page could
-        * start, but OSC calls it under lock and thus we can add oap back to
-        * pending safely */
-       if (rc)
-               /* upper layer wants to leave the page on pending queue */
-               osc_oap_to_pending(oap);
-       else
-               osc_exit_cache(cli, oap, sent);
-       EXIT;
+       osc_object_lock(osc);
+       RETURN(rc);
 }
 
 /**
@@ -633,193 +2096,48 @@ void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
  * \return negative on errors.
  */
 static int
-osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
-                struct osc_object *osc, int cmd,
-                struct osc_oap_pages *lop, pdl_policy_t pol)
+osc_send_read_rpc(const struct lu_env *env, struct client_obd *cli,
+                 struct osc_object *osc, pdl_policy_t pol)
 {
-       obd_count page_count = 0;
-       struct osc_async_page *oap = NULL, *tmp;
-       CFS_LIST_HEAD(rpc_list);
-       int srvlock = 0, mem_tight = 0;
-       obd_off starting_offset = OBD_OBJECT_EOF;
-       unsigned int ending_offset;
-       int starting_page_off = 0;
-       int rc;
+       struct osc_extent *ext;
+       struct osc_extent *next;
+       CFS_LIST_HEAD(rpclist);
+       int page_count = 0;
+       unsigned int max_pages = cli->cl_max_pages_per_rpc;
+       int rc = 0;
        ENTRY;
 
-       /* ASYNC_HP pages first. At present, when the lock the pages is
-        * to be canceled, the pages covered by the lock will be sent out
-        * with ASYNC_HP. We have to send out them as soon as possible. */
-       cfs_list_for_each_entry_safe(oap, tmp, &lop->oop_urgent, oap_urgent_item) {
-               if (oap->oap_async_flags & ASYNC_HP)
-                       cfs_list_move(&oap->oap_pending_item, &rpc_list);
-               else if (!(oap->oap_brw_flags & OBD_BRW_SYNC))
-                       /* only do this for writeback pages. */
-                       cfs_list_move_tail(&oap->oap_pending_item, &rpc_list);
-               if (++page_count >= cli->cl_max_pages_per_rpc)
-                       break;
-       }
-       cfs_list_splice_init(&rpc_list, &lop->oop_pending);
-       page_count = 0;
-
-       /* first we find the pages we're allowed to work with */
-       cfs_list_for_each_entry_safe(oap, tmp, &lop->oop_pending,
-                                    oap_pending_item) {
-               LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
-                        "magic 0x%x\n", oap, oap->oap_magic);
-
-               if (page_count != 0 &&
-                   srvlock != !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK)) {
-                       CDEBUG(D_PAGE, "SRVLOCK flag mismatch,"
-                              " oap %p, page %p, srvlock %u\n",
-                              oap, oap->oap_brw_page.pg, (unsigned)!srvlock);
-                       break;
-               }
-
-               /* If there is a gap at the start of this page, it can't merge
-                * with any previous page, so we'll hand the network a
-                * "fragmented" page array that it can't transfer in 1 RDMA */
-               if (oap->oap_obj_off < starting_offset) {
-                       if (starting_page_off != 0)
-                               break;
-
-                       starting_page_off = oap->oap_page_off;
-                       starting_offset = oap->oap_obj_off + starting_page_off;
-               } else if (oap->oap_page_off != 0)
-                       break;
-
-               /* in llite being 'ready' equates to the page being locked
-                * until completion unlocks it.  commit_write submits a page
-                * as not ready because its unlock will happen unconditionally
-                * as the call returns.  if we race with commit_write giving
-                * us that page we don't want to create a hole in the page
-                * stream, so we stop and leave the rpc to be fired by
-                * another dirtier or kupdated interval (the not ready page
-                * will still be on the dirty list).  we could call in
-                * at the end of ll_file_write to process the queue again. */
-               if (!(oap->oap_async_flags & ASYNC_READY)) {
-                       int rc = osc_make_ready(env, oap, cmd);
-                       if (rc < 0)
-                               CDEBUG(D_INODE, "oap %p page %p returned %d "
-                                               "instead of ready\n", oap,
-                                               oap->oap_page, rc);
-                       switch (rc) {
-                       case -EAGAIN:
-                               /* llite is telling us that the page is still
-                                * in commit_write and that we should try
-                                * and put it in an rpc again later.  we
-                                * break out of the loop so we don't create
-                                * a hole in the sequence of pages in the rpc
-                                * stream.*/
-                               oap = NULL;
-                               break;
-                       case -EINTR:
-                               /* the io isn't needed.. tell the checks
-                                * below to complete the rpc with EINTR */
-                               cfs_spin_lock(&oap->oap_lock);
-                               oap->oap_async_flags |= ASYNC_COUNT_STABLE;
-                               cfs_spin_unlock(&oap->oap_lock);
-                               oap->oap_count = -EINTR;
-                               break;
-                       case 0:
-                               cfs_spin_lock(&oap->oap_lock);
-                               oap->oap_async_flags |= ASYNC_READY;
-                               cfs_spin_unlock(&oap->oap_lock);
-                               break;
-                       default:
-                               LASSERTF(0, "oap %p page %p returned %d "
-                                           "from make_ready\n", oap,
-                                           oap->oap_page, rc);
-                               break;
-                       }
-               }
-               if (oap == NULL)
-                       break;
-
-               /* take the page out of our book-keeping */
-               cfs_list_del_init(&oap->oap_pending_item);
-               lop_update_pending(cli, lop, cmd, -1);
-               cfs_list_del_init(&oap->oap_urgent_item);
-
-               /* ask the caller for the size of the io as the rpc leaves. */
-               if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
-                       oap->oap_count = osc_refresh_count(env, oap, cmd);
-                       LASSERT(oap->oap_page_off + oap->oap_count <= CFS_PAGE_SIZE);
-               }
-               if (oap->oap_count <= 0) {
-                       CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
-                              oap->oap_count);
-                       osc_ap_completion(env, cli, NULL,
-                                         oap, 0, oap->oap_count);
-                       continue;
-               }
-
-               /* now put the page back in our accounting */
-               cfs_list_add_tail(&oap->oap_rpc_item, &rpc_list);
-               if (page_count++ == 0)
-                       srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
-
-               if (oap->oap_brw_flags & OBD_BRW_MEMALLOC)
-                       mem_tight = 1;
-
-               /* End on a PTLRPC_MAX_BRW_SIZE boundary.  We want full-sized
-                * RPCs aligned on PTLRPC_MAX_BRW_SIZE boundaries to help reads
-                * have the same alignment as the initial writes that allocated
-                * extents on the server. */
-               ending_offset = oap->oap_obj_off + oap->oap_page_off +
-                               oap->oap_count;
-               if (!(ending_offset & (PTLRPC_MAX_BRW_SIZE - 1)))
-                       break;
-
-               if (page_count >= cli->cl_max_pages_per_rpc)
-                       break;
-
-               /* If there is a gap at the end of this page, it can't merge
-                * with any subsequent pages, so we'll hand the network a
-                * "fragmented" page array that it can't transfer in 1 RDMA */
-               if (oap->oap_page_off + oap->oap_count < CFS_PAGE_SIZE)
+       LASSERT(osc_object_is_locked(osc));
+       cfs_list_for_each_entry_safe(ext, next,
+                                    &osc->oo_reading_exts, oe_link) {
+               EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
+               if (!try_to_add_extent_for_io(cli, ext, &rpclist, &page_count,
+                                             &max_pages))
                        break;
+               osc_extent_state_set(ext, OES_RPC);
+               EASSERT(ext->oe_nr_pages <= max_pages, ext);
        }
+       LASSERT(page_count <= max_pages);
 
-       osc_list_maint(cli, osc);
-
-       client_obd_list_unlock(&cli->cl_loi_list_lock);
+       osc_update_pending(osc, OBD_BRW_READ, -page_count);
 
-       if (page_count == 0) {
-               client_obd_list_lock(&cli->cl_loi_list_lock);
-               RETURN(0);
-       }
+       if (!cfs_list_empty(&rpclist)) {
+               osc_object_unlock(osc);
 
-       if (mem_tight)
-               cmd |= OBD_BRW_MEMALLOC;
-       rc = osc_build_rpc(env, cli, &rpc_list, page_count, cmd, pol);
-       if (rc != 0) {
-               LASSERT(cfs_list_empty(&rpc_list));
-               osc_list_maint(cli, osc);
-               RETURN(rc);
-       }
+               LASSERT(page_count > 0);
+               rc = osc_build_rpc(env, cli, &rpclist, OBD_BRW_READ, pol);
+               LASSERT(cfs_list_empty(&rpclist));
 
-       starting_offset &= PTLRPC_MAX_BRW_SIZE - 1;
-       if (cmd == OBD_BRW_READ) {
-               cli->cl_r_in_flight++;
-               lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
-               lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
-               lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
-                                     (starting_offset >> CFS_PAGE_SHIFT) + 1);
-       } else {
-               cli->cl_w_in_flight++;
-               lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
-               lprocfs_oh_tally(&cli->cl_write_rpc_hist,
-                                cli->cl_w_in_flight);
-               lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
-                                     (starting_offset >> CFS_PAGE_SHIFT) + 1);
+               osc_object_lock(osc);
        }
-
-       RETURN(1);
+       RETURN(rc);
 }
 
-#define list_to_obj(list, item) \
-       cfs_list_entry((list)->next, struct osc_object, oo_##item)
+#define list_to_obj(list, item) ({                                           \
+       cfs_list_t *__tmp = (list)->next;                                     \
+       cfs_list_del_init(__tmp);                                             \
+       cfs_list_entry(__tmp, struct osc_object, oo_##item);                  \
+})
 
 /* This is called by osc_check_rpcs() to find which objects have pages that
  * we could be sending.  These lists are maintained by osc_makes_rpc(). */
@@ -861,14 +2179,24 @@ static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli,
                           pdl_policy_t pol)
 {
        struct osc_object *osc;
-       int rc = 0, race_counter = 0;
+       int rc = 0;
        ENTRY;
 
        while ((osc = osc_next_obj(cli)) != NULL) {
+               struct cl_object *obj = osc2cl(osc);
+               struct lu_ref_link link;
+
                OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
 
-               if (osc_max_rpc_in_flight(cli, osc))
+               if (osc_max_rpc_in_flight(cli, osc)) {
+                       __osc_list_maint(cli, osc);
                        break;
+               }
+
+               cl_object_get(obj);
+               client_obd_list_unlock(&cli->cl_loi_list_lock);
+               lu_object_ref_add_at(&obj->co_lu, &link, "check",
+                                    current);
 
                /* attempt some read/write balancing by alternating between
                 * reads and writes in an object.  The makes_rpc checks here
@@ -876,13 +2204,13 @@ static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli,
                 * instead of objects.  we don't want send_oap_rpc to drain a
                 * partial read pending queue when we're given this object to
                 * do io on writes while there are cache waiters */
+               osc_object_lock(osc);
                if (osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
-                       rc = osc_send_oap_rpc(env, cli, osc, OBD_BRW_WRITE,
-                                             &osc->oo_write_pages, pol);
+                       rc = osc_send_write_rpc(env, cli, osc, pol);
                        if (rc < 0) {
                                CERROR("Write request failed with %d\n", rc);
 
-                               /* osc_send_oap_rpc failed, mostly because of
+                               /* osc_send_write_rpc failed, mostly because of
                                 * memory pressure.
                                 *
                                 * It can't break here, because if:
@@ -900,57 +2228,62 @@ static void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli,
                                 * Anyway, continue to drain pages. */
                                /* break; */
                        }
-
-                       if (rc > 0)
-                               race_counter = 0;
-                       else if (rc == 0)
-                               race_counter++;
                }
                if (osc_makes_rpc(cli, osc, OBD_BRW_READ)) {
-                       rc = osc_send_oap_rpc(env, cli, osc, OBD_BRW_READ,
-                                             &osc->oo_read_pages, pol);
+                       rc = osc_send_read_rpc(env, cli, osc, pol);
                        if (rc < 0)
                                CERROR("Read request failed with %d\n", rc);
-
-                       if (rc > 0)
-                               race_counter = 0;
-                       else if (rc == 0)
-                               race_counter++;
                }
-
-               /* attempt some inter-object balancing by issuing rpcs
-                * for each object in turn */
-               if (!cfs_list_empty(&osc->oo_hp_ready_item))
-                       cfs_list_del_init(&osc->oo_hp_ready_item);
-               if (!cfs_list_empty(&osc->oo_ready_item))
-                       cfs_list_del_init(&osc->oo_ready_item);
-               if (!cfs_list_empty(&osc->oo_write_item))
-                       cfs_list_del_init(&osc->oo_write_item);
-               if (!cfs_list_empty(&osc->oo_read_item))
-                       cfs_list_del_init(&osc->oo_read_item);
+               osc_object_unlock(osc);
 
                osc_list_maint(cli, osc);
+               lu_object_ref_del_at(&obj->co_lu, &link, "check",
+                                    current);
+               cl_object_put(env, obj);
 
-               /* send_oap_rpc fails with 0 when make_ready tells it to
-                * back off.  llite's make_ready does this when it tries
-                * to lock a page queued for write that is already locked.
-                * we want to try sending rpcs from many objects, but we
-                * don't want to spin failing with 0.  */
-               if (race_counter == 10)
-                       break;
+               client_obd_list_lock(&cli->cl_loi_list_lock);
+       }
+}
+
+static int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
+                         struct osc_object *osc, pdl_policy_t pol, int async)
+{
+       int rc = 0;
+
+       if (osc != NULL && osc_list_maint(cli, osc) == 0)
+               return 0;
+
+       if (!async) {
+               /* disable osc_lru_shrink() temporarily to avoid
+                * potential stack overrun problem. LU-2859 */
+               cfs_atomic_inc(&cli->cl_lru_shrinkers);
+               client_obd_list_lock(&cli->cl_loi_list_lock);
+               osc_check_rpcs(env, cli, pol);
+               client_obd_list_unlock(&cli->cl_loi_list_lock);
+               cfs_atomic_dec(&cli->cl_lru_shrinkers);
+       } else {
+               CDEBUG(D_CACHE, "Queue writeback work for client %p.\n", cli);
+               LASSERT(cli->cl_writeback_work != NULL);
+               rc = ptlrpcd_queue_work(cli->cl_writeback_work);
        }
+       return rc;
+}
+
+static int osc_io_unplug_async(const struct lu_env *env,
+                               struct client_obd *cli, struct osc_object *osc)
+{
+       /* XXX: policy is no use actually. */
+       return osc_io_unplug0(env, cli, osc, PDL_POLICY_ROUND, 1);
 }
 
 void osc_io_unplug(const struct lu_env *env, struct client_obd *cli,
                   struct osc_object *osc, pdl_policy_t pol)
 {
-       if (osc)
-               osc_list_maint(cli, osc);
-       osc_check_rpcs(env, cli, pol);
+       (void)osc_io_unplug0(env, cli, osc, pol, 0);
 }
 
 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
-                       cfs_page_t *page, loff_t offset)
+                       struct page *page, loff_t offset)
 {
        struct obd_export     *exp = osc_export(osc);
        struct osc_async_page *oap = &ops->ops_oap;
@@ -971,24 +2304,28 @@ int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
                oap->oap_brw_flags = OBD_BRW_NOQUOTA;
 
        CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
-       CFS_INIT_LIST_HEAD(&oap->oap_urgent_item);
        CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
 
-       cfs_spin_lock_init(&oap->oap_lock);
-       CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n",
+       spin_lock_init(&oap->oap_lock);
+       CDEBUG(D_INFO, "oap %p page %p obj off "LPU64"\n",
               oap, page, oap->oap_obj_off);
        RETURN(0);
 }
 
-int osc_queue_async_io(const struct lu_env *env, struct osc_page *ops)
+int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
+                      struct osc_page *ops)
 {
+       struct osc_io *oio = osc_env_io(env);
+       struct osc_extent     *ext = NULL;
        struct osc_async_page *oap = &ops->ops_oap;
        struct client_obd     *cli = oap->oap_cli;
        struct osc_object     *osc = oap->oap_obj;
-       struct obd_export     *exp = osc_export(osc);
-       int brw_flags = OBD_BRW_ASYNC;
-       int cmd = OBD_BRW_WRITE;
-       int rc = 0;
+       pgoff_t index;
+       int    grants = 0;
+       int    brw_flags = OBD_BRW_ASYNC;
+       int    cmd = OBD_BRW_WRITE;
+       int    need_release = 0;
+       int    rc = 0;
        ENTRY;
 
        if (oap->oap_magic != OAP_MAGIC)
@@ -998,13 +2335,13 @@ int osc_queue_async_io(const struct lu_env *env, struct osc_page *ops)
                RETURN(-EIO);
 
        if (!cfs_list_empty(&oap->oap_pending_item) ||
-           !cfs_list_empty(&oap->oap_urgent_item) ||
            !cfs_list_empty(&oap->oap_rpc_item))
                RETURN(-EBUSY);
 
        /* Set the OBD_BRW_SRVLOCK before the page is queued. */
        brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
-       if (!client_is_remote(exp) && cfs_capable(CFS_CAP_SYS_RESOURCE)) {
+       if (!client_is_remote(osc_export(osc)) &&
+           cfs_capable(CFS_CAP_SYS_RESOURCE)) {
                brw_flags |= OBD_BRW_NOQUOTA;
                cmd |= OBD_BRW_NOQUOTA;
        }
@@ -1024,213 +2361,683 @@ int osc_queue_async_io(const struct lu_env *env, struct osc_page *ops)
 
                qid[USRQUOTA] = attr->cat_uid;
                qid[GRPQUOTA] = attr->cat_gid;
-               if (rc == 0 &&
-                   osc_quota_chkdq(cli, qid) == NO_QUOTA)
+               if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
                        rc = -EDQUOT;
                if (rc)
                        RETURN(rc);
        }
 
-       client_obd_list_lock(&cli->cl_loi_list_lock);
+       if (osc_over_unstable_soft_limit(cli))
+               brw_flags |= OBD_BRW_SOFT_SYNC;
 
        oap->oap_cmd = cmd;
        oap->oap_page_off = ops->ops_from;
        oap->oap_count = ops->ops_to - ops->ops_from;
+       /* No need to hold a lock here,
+        * since this page is not in any list yet. */
        oap->oap_async_flags = 0;
        oap->oap_brw_flags = brw_flags;
-       /* Give a hint to OST that requests are coming from kswapd - bug19529 */
-       if (cfs_memory_pressure_get())
-               oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
 
-       rc = osc_enter_cache(env, cli, oap);
-       if (rc) {
+       OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
+                    oap, oap->oap_page, oap->oap_cmd & OBD_BRW_RWMASK);
+
+       index = oap2cl_page(oap)->cp_index;
+
+       /* Add this page into extent by the following steps:
+        * 1. if there exists an active extent for this IO, mostly this page
+        *    can be added to the active extent and sometimes we need to
+        *    expand extent to accomodate this page;
+        * 2. otherwise, a new extent will be allocated. */
+
+       ext = oio->oi_active;
+       if (ext != NULL && ext->oe_start <= index && ext->oe_max_end >= index) {
+               /* one chunk plus extent overhead must be enough to write this
+                * page */
+               grants = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
+               if (ext->oe_end >= index)
+                       grants = 0;
+
+               /* it doesn't need any grant to dirty this page */
+               client_obd_list_lock(&cli->cl_loi_list_lock);
+               rc = osc_enter_cache_try(cli, oap, grants, 0);
                client_obd_list_unlock(&cli->cl_loi_list_lock);
-               RETURN(rc);
+               if (rc == 0) { /* try failed */
+                       grants = 0;
+                       need_release = 1;
+               } else if (ext->oe_end < index) {
+                       int tmp = grants;
+                       /* try to expand this extent */
+                       rc = osc_extent_expand(ext, index, &tmp);
+                       if (rc < 0) {
+                               need_release = 1;
+                               /* don't free reserved grant */
+                       } else {
+                               OSC_EXTENT_DUMP(D_CACHE, ext,
+                                               "expanded for %lu.\n", index);
+                               osc_unreserve_grant(cli, grants, tmp);
+                               grants = 0;
+                       }
+               }
+               rc = 0;
+       } else if (ext != NULL) {
+               /* index is located outside of active extent */
+               need_release = 1;
+       }
+       if (need_release) {
+               osc_extent_release(env, ext);
+               oio->oi_active = NULL;
+               ext = NULL;
        }
 
-       OSC_IO_DEBUG(osc, "oap %p page %p added for cmd %d\n",
-                    oap, oap->oap_page, cmd);
+       if (ext == NULL) {
+               int tmp = (1 << cli->cl_chunkbits) + cli->cl_extent_tax;
+
+               /* try to find new extent to cover this page */
+               LASSERT(oio->oi_active == NULL);
+               /* we may have allocated grant for this page if we failed
+                * to expand the previous active extent. */
+               LASSERT(ergo(grants > 0, grants >= tmp));
+
+               rc = 0;
+               if (grants == 0) {
+                       /* we haven't allocated grant for this page. */
+                       rc = osc_enter_cache(env, cli, oap, tmp);
+                       if (rc == 0)
+                               grants = tmp;
+               }
 
-       osc_oap_to_pending(oap);
-       osc_list_maint(cli, osc);
-       if (!osc_max_rpc_in_flight(cli, osc) &&
-           osc_makes_rpc(cli, osc, OBD_BRW_WRITE)) {
-               LASSERT(cli->cl_writeback_work != NULL);
-               rc = ptlrpcd_queue_work(cli->cl_writeback_work);
+               tmp = grants;
+               if (rc == 0) {
+                       ext = osc_extent_find(env, osc, index, &tmp);
+                       if (IS_ERR(ext)) {
+                               LASSERT(tmp == grants);
+                               osc_exit_cache(cli, oap);
+                               rc = PTR_ERR(ext);
+                               ext = NULL;
+                       } else {
+                               oio->oi_active = ext;
+                       }
+               }
+               if (grants > 0)
+                       osc_unreserve_grant(cli, grants, tmp);
+       }
+
+       LASSERT(ergo(rc == 0, ext != NULL));
+       if (ext != NULL) {
+               EASSERTF(ext->oe_end >= index && ext->oe_start <= index,
+                        ext, "index = %lu.\n", index);
+               LASSERT((oap->oap_brw_flags & OBD_BRW_FROM_GRANT) != 0);
 
-               CDEBUG(D_CACHE, "Queued writeback work for client obd %p/%d.\n",
-                      cli, rc);
+               osc_object_lock(osc);
+               if (ext->oe_nr_pages == 0)
+                       ext->oe_srvlock = ops->ops_srvlock;
+               else
+                       LASSERT(ext->oe_srvlock == ops->ops_srvlock);
+               ++ext->oe_nr_pages;
+               cfs_list_add_tail(&oap->oap_pending_item, &ext->oe_pages);
+               osc_object_unlock(osc);
        }
-       client_obd_list_unlock(&cli->cl_loi_list_lock);
+       RETURN(rc);
+}
 
-       RETURN(0);
+int osc_teardown_async_page(const struct lu_env *env,
+                           struct osc_object *obj, struct osc_page *ops)
+{
+       struct osc_async_page *oap = &ops->ops_oap;
+       struct osc_extent     *ext = NULL;
+       int rc = 0;
+       ENTRY;
+
+       LASSERT(oap->oap_magic == OAP_MAGIC);
+
+       CDEBUG(D_INFO, "teardown oap %p page %p at index %lu.\n",
+              oap, ops, oap2cl_page(oap)->cp_index);
+
+       osc_object_lock(obj);
+       if (!cfs_list_empty(&oap->oap_rpc_item)) {
+               CDEBUG(D_CACHE, "oap %p is not in cache.\n", oap);
+               rc = -EBUSY;
+       } else if (!cfs_list_empty(&oap->oap_pending_item)) {
+               ext = osc_extent_lookup(obj, oap2cl_page(oap)->cp_index);
+               /* only truncated pages are allowed to be taken out.
+                * See osc_extent_truncate() and osc_cache_truncate_start()
+                * for details. */
+               if (ext != NULL && ext->oe_state != OES_TRUNC) {
+                       OSC_EXTENT_DUMP(D_ERROR, ext, "trunc at %lu.\n",
+                                       oap2cl_page(oap)->cp_index);
+                       rc = -EBUSY;
+               }
+       }
+       osc_object_unlock(obj);
+       if (ext != NULL)
+               osc_extent_put(env, ext);
+       RETURN(rc);
 }
 
-int osc_teardown_async_page(struct osc_object *obj, struct osc_page *ops)
+/**
+ * This is called when a page is picked up by kernel to write out.
+ *
+ * We should find out the corresponding extent and add the whole extent
+ * into urgent list. The extent may be being truncated or used, handle it
+ * carefully.
+ */
+int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
+                        struct osc_page *ops)
 {
+       struct osc_extent *ext   = NULL;
+       struct osc_object *obj   = cl2osc(ops->ops_cl.cpl_obj);
+       struct cl_page    *cp    = ops->ops_cl.cpl_page;
+       pgoff_t            index = cp->cp_index;
        struct osc_async_page *oap = &ops->ops_oap;
-       struct client_obd     *cli = oap->oap_cli;
-       struct osc_oap_pages  *lop;
+       bool unplug = false;
        int rc = 0;
        ENTRY;
 
-       if (oap->oap_magic != OAP_MAGIC)
-               RETURN(-EINVAL);
+       osc_object_lock(obj);
+       ext = osc_extent_lookup(obj, index);
+       if (ext == NULL) {
+               osc_extent_tree_dump(D_ERROR, obj);
+               LASSERTF(0, "page index %lu is NOT covered.\n", index);
+       }
 
-       if (oap->oap_cmd & OBD_BRW_WRITE) {
-               lop = &obj->oo_write_pages;
-       } else {
-               lop = &obj->oo_read_pages;
+       switch (ext->oe_state) {
+       case OES_RPC:
+       case OES_LOCK_DONE:
+               CL_PAGE_DEBUG(D_ERROR, env, cl_page_top(cp),
+                             "flush an in-rpc page?\n");
+               LASSERT(0);
+               break;
+       case OES_LOCKING:
+               /* If we know this extent is being written out, we should abort
+                * so that the writer can make this page ready. Otherwise, there
+                * exists a deadlock problem because other process can wait for
+                * page writeback bit holding page lock; and meanwhile in
+                * vvp_page_make_ready(), we need to grab page lock before
+                * really sending the RPC. */
+       case OES_TRUNC:
+               /* race with truncate, page will be redirtied */
+               GOTO(out, rc = -EAGAIN);
+       default:
+               break;
        }
 
-       client_obd_list_lock(&cli->cl_loi_list_lock);
+       rc = cl_page_prep(env, io, cl_page_top(cp), CRT_WRITE);
+       if (rc)
+               GOTO(out, rc);
+
+       spin_lock(&oap->oap_lock);
+       oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
+       spin_unlock(&oap->oap_lock);
+
+       if (memory_pressure_get())
+               ext->oe_memalloc = 1;
+
+       ext->oe_urgent = 1;
+       if (ext->oe_state == OES_CACHE) {
+               OSC_EXTENT_DUMP(D_CACHE, ext,
+                               "flush page %p make it urgent.\n", oap);
+               if (cfs_list_empty(&ext->oe_link))
+                       cfs_list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
+               unplug = true;
+       }
+       rc = 0;
+       EXIT;
+
+out:
+       osc_object_unlock(obj);
+       osc_extent_put(env, ext);
+       if (unplug)
+               osc_io_unplug_async(env, osc_cli(obj), obj);
+       return rc;
+}
 
-       if (!cfs_list_empty(&oap->oap_rpc_item))
-               GOTO(out, rc = -EBUSY);
+/**
+ * this is called when a sync waiter receives an interruption.  Its job is to
+ * get the caller woken as soon as possible.  If its page hasn't been put in an
+ * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
+ * desiring interruption which will forcefully complete the rpc once the rpc
+ * has timed out.
+ */
+int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
+{
+       struct osc_async_page *oap = &ops->ops_oap;
+       struct osc_object     *obj = oap->oap_obj;
+       struct client_obd     *cli = osc_cli(obj);
+       struct osc_extent     *ext;
+       struct osc_extent     *found = NULL;
+       cfs_list_t            *plist;
+       pgoff_t index = oap2cl_page(oap)->cp_index;
+       int     rc = -EBUSY;
+       int     cmd;
+       ENTRY;
 
-       osc_exit_cache(cli, oap, 0);
-       osc_wake_cache_waiters(cli);
+       LASSERT(!oap->oap_interrupted);
+       oap->oap_interrupted = 1;
 
-       if (!cfs_list_empty(&oap->oap_urgent_item)) {
-               cfs_list_del_init(&oap->oap_urgent_item);
-               cfs_spin_lock(&oap->oap_lock);
-               oap->oap_async_flags &= ~(ASYNC_URGENT | ASYNC_HP);
-               cfs_spin_unlock(&oap->oap_lock);
+       /* Find out the caching extent */
+       osc_object_lock(obj);
+       if (oap->oap_cmd & OBD_BRW_WRITE) {
+               plist = &obj->oo_urgent_exts;
+               cmd   = OBD_BRW_WRITE;
+       } else {
+               plist = &obj->oo_reading_exts;
+               cmd   = OBD_BRW_READ;
        }
-       if (!cfs_list_empty(&oap->oap_pending_item)) {
-               cfs_list_del_init(&oap->oap_pending_item);
-               lop_update_pending(cli, lop, oap->oap_cmd, -1);
+       cfs_list_for_each_entry(ext, plist, oe_link) {
+               if (ext->oe_start <= index && ext->oe_end >= index) {
+                       LASSERT(ext->oe_state == OES_LOCK_DONE);
+                       /* For OES_LOCK_DONE state extent, it has already held
+                        * a refcount for RPC. */
+                       found = osc_extent_get(ext);
+                       break;
+               }
        }
+       if (found != NULL) {
+               cfs_list_del_init(&found->oe_link);
+               osc_update_pending(obj, cmd, -found->oe_nr_pages);
+               osc_object_unlock(obj);
+
+               osc_extent_finish(env, found, 0, -EINTR);
+               osc_extent_put(env, found);
+               rc = 0;
+       } else {
+               osc_object_unlock(obj);
+               /* ok, it's been put in an rpc. only one oap gets a request
+                * reference */
+               if (oap->oap_request != NULL) {
+                       ptlrpc_mark_interrupted(oap->oap_request);
+                       ptlrpcd_wake(oap->oap_request);
+                       ptlrpc_req_finished(oap->oap_request);
+                       oap->oap_request = NULL;
+               }
+       }
+
        osc_list_maint(cli, obj);
-       OSC_IO_DEBUG(obj, "oap %p page %p torn down\n", oap, oap->oap_page);
-out:
-       client_obd_list_unlock(&cli->cl_loi_list_lock);
        RETURN(rc);
 }
 
-/* aka (~was & now & flag), but this is more clear :) */
-#define SETTING(was, now, flag) (!(was & flag) && (now & flag))
-
-int osc_set_async_flags(struct osc_object *obj, struct osc_page *opg,
-                       obd_flag async_flags)
+int osc_queue_sync_pages(const struct lu_env *env, struct osc_object *obj,
+                        cfs_list_t *list, int cmd, int brw_flags)
 {
-       struct osc_async_page *oap = &opg->ops_oap;
-       struct osc_oap_pages *lop;
-       int flags = 0;
+       struct client_obd     *cli = osc_cli(obj);
+       struct osc_extent     *ext;
+       struct osc_async_page *oap;
+       int     page_count = 0;
+       int     mppr       = cli->cl_max_pages_per_rpc;
+       pgoff_t start      = CL_PAGE_EOF;
+       pgoff_t end        = 0;
        ENTRY;
 
-       LASSERT(!cfs_list_empty(&oap->oap_pending_item));
+       cfs_list_for_each_entry(oap, list, oap_pending_item) {
+               struct cl_page *cp = oap2cl_page(oap);
+               if (cp->cp_index > end)
+                       end = cp->cp_index;
+               if (cp->cp_index < start)
+                       start = cp->cp_index;
+               ++page_count;
+               mppr <<= (page_count > mppr);
+       }
 
-       if (oap->oap_cmd & OBD_BRW_WRITE) {
-               lop = &obj->oo_write_pages;
+       ext = osc_extent_alloc(obj);
+       if (ext == NULL) {
+               cfs_list_for_each_entry(oap, list, oap_pending_item) {
+                       cfs_list_del_init(&oap->oap_pending_item);
+                       osc_ap_completion(env, cli, oap, 0, -ENOMEM);
+               }
+               RETURN(-ENOMEM);
+       }
+
+       ext->oe_rw = !!(cmd & OBD_BRW_READ);
+       ext->oe_urgent = 1;
+       ext->oe_start = start;
+       ext->oe_end = ext->oe_max_end = end;
+       ext->oe_obj = obj;
+       ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
+       ext->oe_nr_pages = page_count;
+       ext->oe_mppr = mppr;
+       cfs_list_splice_init(list, &ext->oe_pages);
+
+       osc_object_lock(obj);
+       /* Reuse the initial refcount for RPC, don't drop it */
+       osc_extent_state_set(ext, OES_LOCK_DONE);
+       if (cmd & OBD_BRW_WRITE) {
+               cfs_list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
+               osc_update_pending(obj, OBD_BRW_WRITE, page_count);
        } else {
-               lop = &obj->oo_read_pages;
+               cfs_list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
+               osc_update_pending(obj, OBD_BRW_READ, page_count);
        }
+       osc_object_unlock(obj);
 
-       if ((oap->oap_async_flags & async_flags) == async_flags)
-               RETURN(0);
+       osc_io_unplug(env, cli, obj, PDL_POLICY_ROUND);
+       RETURN(0);
+}
+
+/**
+ * Called by osc_io_setattr_start() to freeze and destroy covering extents.
+ */
+int osc_cache_truncate_start(const struct lu_env *env, struct osc_io *oio,
+                            struct osc_object *obj, __u64 size)
+{
+       struct client_obd *cli = osc_cli(obj);
+       struct osc_extent *ext;
+       struct osc_extent *waiting = NULL;
+       pgoff_t index;
+       CFS_LIST_HEAD(list);
+       int result = 0;
+       bool partial;
+       ENTRY;
+
+       /* pages with index greater or equal to index will be truncated. */
+       index = cl_index(osc2cl(obj), size);
+       partial = size > cl_offset(osc2cl(obj), index);
+
+again:
+       osc_object_lock(obj);
+       ext = osc_extent_search(obj, index);
+       if (ext == NULL)
+               ext = first_extent(obj);
+       else if (ext->oe_end < index)
+               ext = next_extent(ext);
+       while (ext != NULL) {
+               EASSERT(ext->oe_state != OES_TRUNC, ext);
+
+               if (ext->oe_state > OES_CACHE || ext->oe_urgent) {
+                       /* if ext is in urgent state, it means there must exist
+                        * a page already having been flushed by write_page().
+                        * We have to wait for this extent because we can't
+                        * truncate that page. */
+                       LASSERT(!ext->oe_hp);
+                       OSC_EXTENT_DUMP(D_CACHE, ext,
+                                       "waiting for busy extent\n");
+                       waiting = osc_extent_get(ext);
+                       break;
+               }
 
-       if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
-               flags |= ASYNC_READY;
+               OSC_EXTENT_DUMP(D_CACHE, ext, "try to trunc:"LPU64".\n", size);
 
-       if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT) &&
-           cfs_list_empty(&oap->oap_rpc_item)) {
-               if (oap->oap_async_flags & ASYNC_HP)
-                       cfs_list_add(&oap->oap_urgent_item, &lop->oop_urgent);
-               else
-                       cfs_list_add_tail(&oap->oap_urgent_item,
-                                         &lop->oop_urgent);
-               flags |= ASYNC_URGENT;
-               osc_list_maint(oap->oap_cli, obj);
+               osc_extent_get(ext);
+               if (ext->oe_state == OES_ACTIVE) {
+                       /* though we grab inode mutex for write path, but we
+                        * release it before releasing extent(in osc_io_end()),
+                        * so there is a race window that an extent is still
+                        * in OES_ACTIVE when truncate starts. */
+                       LASSERT(!ext->oe_trunc_pending);
+                       ext->oe_trunc_pending = 1;
+               } else {
+                       EASSERT(ext->oe_state == OES_CACHE, ext);
+                       osc_extent_state_set(ext, OES_TRUNC);
+                       osc_update_pending(obj, OBD_BRW_WRITE,
+                                          -ext->oe_nr_pages);
+               }
+               EASSERT(cfs_list_empty(&ext->oe_link), ext);
+               cfs_list_add_tail(&ext->oe_link, &list);
+
+               ext = next_extent(ext);
        }
-       cfs_spin_lock(&oap->oap_lock);
-       oap->oap_async_flags |= flags;
-       cfs_spin_unlock(&oap->oap_lock);
+       osc_object_unlock(obj);
 
-       OSC_IO_DEBUG(obj, "oap %p page %p has flags %x\n", oap,
-                    oap->oap_page, oap->oap_async_flags);
-       RETURN(0);
+       osc_list_maint(cli, obj);
+
+       while (!cfs_list_empty(&list)) {
+               int rc;
+
+               ext = cfs_list_entry(list.next, struct osc_extent, oe_link);
+               cfs_list_del_init(&ext->oe_link);
+
+               /* extent may be in OES_ACTIVE state because inode mutex
+                * is released before osc_io_end() in file write case */
+               if (ext->oe_state != OES_TRUNC)
+                       osc_extent_wait(env, ext, OES_TRUNC);
+
+               rc = osc_extent_truncate(ext, index, partial);
+               if (rc < 0) {
+                       if (result == 0)
+                               result = rc;
+
+                       OSC_EXTENT_DUMP(D_ERROR, ext,
+                                       "truncate error %d\n", rc);
+               } else if (ext->oe_nr_pages == 0) {
+                       osc_extent_remove(ext);
+               } else {
+                       /* this must be an overlapped extent which means only
+                        * part of pages in this extent have been truncated.
+                        */
+                       EASSERTF(ext->oe_start <= index, ext,
+                                "trunc index = %lu/%d.\n", index, partial);
+                       /* fix index to skip this partially truncated extent */
+                       index = ext->oe_end + 1;
+                       partial = false;
+
+                       /* we need to hold this extent in OES_TRUNC state so
+                        * that no writeback will happen. This is to avoid
+                        * BUG 17397. */
+                       LASSERT(oio->oi_trunc == NULL);
+                       oio->oi_trunc = osc_extent_get(ext);
+                       OSC_EXTENT_DUMP(D_CACHE, ext,
+                                       "trunc at "LPU64"\n", size);
+               }
+               osc_extent_put(env, ext);
+       }
+       if (waiting != NULL) {
+               int rc;
+
+               /* ignore the result of osc_extent_wait the write initiator
+                * should take care of it. */
+               rc = osc_extent_wait(env, waiting, OES_INV);
+               if (rc < 0)
+                       OSC_EXTENT_DUMP(D_CACHE, waiting, "error: %d.\n", rc);
+
+               osc_extent_put(env, waiting);
+               waiting = NULL;
+               goto again;
+       }
+       RETURN(result);
 }
 
 /**
- * this is called when a sync waiter receives an interruption.  Its job is to
- * get the caller woken as soon as possible.  If its page hasn't been put in an
- * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
- * desiring interruption which will forcefully complete the rpc once the rpc
- * has timed out.
+ * Called after osc_io_setattr_end to add oio->oi_trunc back to cache.
  */
-int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
+void osc_cache_truncate_end(const struct lu_env *env, struct osc_io *oio,
+                           struct osc_object *obj)
 {
-       struct osc_async_page *oap = &ops->ops_oap;
-       int rc = -EBUSY;
+       struct osc_extent *ext = oio->oi_trunc;
+
+       oio->oi_trunc = NULL;
+       if (ext != NULL) {
+               bool unplug = false;
+
+               EASSERT(ext->oe_nr_pages > 0, ext);
+               EASSERT(ext->oe_state == OES_TRUNC, ext);
+               EASSERT(!ext->oe_urgent, ext);
+
+               OSC_EXTENT_DUMP(D_CACHE, ext, "trunc -> cache.\n");
+               osc_object_lock(obj);
+               osc_extent_state_set(ext, OES_CACHE);
+               if (ext->oe_fsync_wait && !ext->oe_urgent) {
+                       ext->oe_urgent = 1;
+                       cfs_list_move_tail(&ext->oe_link, &obj->oo_urgent_exts);
+                       unplug = true;
+               }
+               osc_update_pending(obj, OBD_BRW_WRITE, ext->oe_nr_pages);
+               osc_object_unlock(obj);
+               osc_extent_put(env, ext);
+
+               if (unplug)
+                       osc_io_unplug_async(env, osc_cli(obj), obj);
+       }
+}
+
+/**
+ * Wait for extents in a specific range to be written out.
+ * The caller must have called osc_cache_writeback_range() to issue IO
+ * otherwise it will take a long time for this function to finish.
+ *
+ * Caller must hold inode_mutex , or cancel exclusive dlm lock so that
+ * nobody else can dirty this range of file while we're waiting for
+ * extents to be written.
+ */
+int osc_cache_wait_range(const struct lu_env *env, struct osc_object *obj,
+                        pgoff_t start, pgoff_t end)
+{
+       struct osc_extent *ext;
+       pgoff_t index = start;
+       int     result = 0;
        ENTRY;
 
-       LASSERT(!oap->oap_interrupted);
-       oap->oap_interrupted = 1;
+again:
+       osc_object_lock(obj);
+       ext = osc_extent_search(obj, index);
+       if (ext == NULL)
+               ext = first_extent(obj);
+       else if (ext->oe_end < index)
+               ext = next_extent(ext);
+       while (ext != NULL) {
+               int rc;
+
+               if (ext->oe_start > end)
+                       break;
 
-       /* ok, it's been put in an rpc. only one oap gets a request reference */
-       if (oap->oap_request != NULL) {
-               ptlrpc_mark_interrupted(oap->oap_request);
-               ptlrpcd_wake(oap->oap_request);
-               ptlrpc_req_finished(oap->oap_request);
-               oap->oap_request = NULL;
-       }
+               if (!ext->oe_fsync_wait) {
+                       ext = next_extent(ext);
+                       continue;
+               }
 
-       /*
-        * page completion may be called only if ->cpo_prep() method was
-        * executed by osc_io_submit(), that also adds page the to pending list
-        */
-       if (!cfs_list_empty(&oap->oap_pending_item)) {
-               struct osc_oap_pages *lop;
-               struct osc_object *osc = oap->oap_obj;
+               EASSERT(ergo(ext->oe_state == OES_CACHE,
+                            ext->oe_hp || ext->oe_urgent), ext);
+               EASSERT(ergo(ext->oe_state == OES_ACTIVE,
+                            !ext->oe_hp && ext->oe_urgent), ext);
 
-               cfs_list_del_init(&oap->oap_pending_item);
-               cfs_list_del_init(&oap->oap_urgent_item);
+               index = ext->oe_end + 1;
+               osc_extent_get(ext);
+               osc_object_unlock(obj);
 
-               lop = (oap->oap_cmd & OBD_BRW_WRITE) ?
-                       &osc->oo_write_pages : &osc->oo_read_pages;
-               lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
-               osc_list_maint(oap->oap_cli, osc);
-               rc = osc_completion(env, oap, oap->oap_cmd, NULL, -EINTR);
+               rc = osc_extent_wait(env, ext, OES_INV);
+               if (result == 0)
+                       result = rc;
+               osc_extent_put(env, ext);
+               goto again;
        }
+       osc_object_unlock(obj);
 
-       RETURN(rc);
+       OSC_IO_DEBUG(obj, "sync file range.\n");
+       RETURN(result);
 }
 
-int osc_queue_sync_page(const struct lu_env *env, struct osc_page *opg,
-                       int cmd, int brw_flags)
+/**
+ * Called to write out a range of osc object.
+ *
+ * @hp     : should be set this is caused by lock cancel;
+ * @discard: is set if dirty pages should be dropped - file will be deleted or
+ *        truncated, this implies there is no partially discarding extents.
+ *
+ * Return how many pages will be issued, or error code if error occurred.
+ */
+int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
+                             pgoff_t start, pgoff_t end, int hp, int discard)
 {
-       struct osc_async_page *oap = &opg->ops_oap;
-       struct client_obd     *cli = oap->oap_cli;
-       int flags = 0;
+       struct osc_extent *ext;
+       CFS_LIST_HEAD(discard_list);
+       bool unplug = false;
+       int result = 0;
        ENTRY;
 
-       oap->oap_cmd       = cmd;
-       oap->oap_page_off  = opg->ops_from;
-       oap->oap_count     = opg->ops_to - opg->ops_from;
-       oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
+       osc_object_lock(obj);
+       ext = osc_extent_search(obj, start);
+       if (ext == NULL)
+               ext = first_extent(obj);
+       else if (ext->oe_end < start)
+               ext = next_extent(ext);
+       while (ext != NULL) {
+               if (ext->oe_start > end)
+                       break;
 
-       /* Give a hint to OST that requests are coming from kswapd - bug19529 */
-       if (cfs_memory_pressure_get())
-               oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
+               ext->oe_fsync_wait = 1;
+               switch (ext->oe_state) {
+               case OES_CACHE:
+                       result += ext->oe_nr_pages;
+                       if (!discard) {
+                               cfs_list_t *list = NULL;
+                               if (hp) {
+                                       EASSERT(!ext->oe_hp, ext);
+                                       ext->oe_hp = 1;
+                                       list = &obj->oo_hp_exts;
+                               } else if (!ext->oe_urgent) {
+                                       ext->oe_urgent = 1;
+                                       list = &obj->oo_urgent_exts;
+                               }
+                               if (list != NULL)
+                                       cfs_list_move_tail(&ext->oe_link, list);
+                               unplug = true;
+                       } else {
+                               /* the only discarder is lock cancelling, so
+                                * [start, end] must contain this extent */
+                               EASSERT(ext->oe_start >= start &&
+                                       ext->oe_max_end <= end, ext);
+                               osc_extent_state_set(ext, OES_LOCKING);
+                               ext->oe_owner = current;
+                               cfs_list_move_tail(&ext->oe_link,
+                                                  &discard_list);
+                               osc_update_pending(obj, OBD_BRW_WRITE,
+                                                  -ext->oe_nr_pages);
+                       }
+                       break;
+               case OES_ACTIVE:
+                       /* It's pretty bad to wait for ACTIVE extents, because
+                        * we don't know how long we will wait for it to be
+                        * flushed since it may be blocked at awaiting more
+                        * grants. We do this for the correctness of fsync. */
+                       LASSERT(hp == 0 && discard == 0);
+                       ext->oe_urgent = 1;
+                       break;
+               case OES_TRUNC:
+                       /* this extent is being truncated, can't do anything
+                        * for it now. it will be set to urgent after truncate
+                        * is finished in osc_cache_truncate_end(). */
+               default:
+                       break;
+               }
+               ext = next_extent(ext);
+       }
+       osc_object_unlock(obj);
+
+       LASSERT(ergo(!discard, cfs_list_empty(&discard_list)));
+       if (!cfs_list_empty(&discard_list)) {
+               struct osc_extent *tmp;
+               int rc;
+
+               osc_list_maint(osc_cli(obj), obj);
+               cfs_list_for_each_entry_safe(ext, tmp, &discard_list, oe_link) {
+                       cfs_list_del_init(&ext->oe_link);
+                       EASSERT(ext->oe_state == OES_LOCKING, ext);
+
+                       /* Discard caching pages. We don't actually write this
+                        * extent out but we complete it as if we did. */
+                       rc = osc_extent_make_ready(env, ext);
+                       if (unlikely(rc < 0)) {
+                               OSC_EXTENT_DUMP(D_ERROR, ext,
+                                               "make_ready returned %d\n", rc);
+                               if (result >= 0)
+                                       result = rc;
+                       }
 
-       if (!client_is_remote(osc_export(cl2osc(opg->ops_cl.cpl_obj))) &&
-           cfs_capable(CFS_CAP_SYS_RESOURCE)) {
-               oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
-               oap->oap_cmd |= OBD_BRW_NOQUOTA;
+                       /* finish the extent as if the pages were sent */
+                       osc_extent_finish(env, ext, 0, 0);
+               }
        }
 
-       if (oap->oap_cmd & OBD_BRW_READ)
-               flags = ASYNC_COUNT_STABLE;
-       else if (!(oap->oap_brw_page.flag & OBD_BRW_FROM_GRANT))
-               osc_enter_cache_try(env, cli, oap, 1);
+       if (unplug)
+               osc_io_unplug(env, osc_cli(obj), obj, PDL_POLICY_ROUND);
 
-       cfs_spin_lock(&oap->oap_lock);
-       oap->oap_async_flags |= OSC_FLAGS | flags;
-       cfs_spin_unlock(&oap->oap_lock);
+       if (hp || discard) {
+               int rc;
+               rc = osc_cache_wait_range(env, obj, start, end);
+               if (result >= 0 && rc < 0)
+                       result = rc;
+       }
 
-       osc_oap_to_pending(oap);
-       RETURN(0);
+       OSC_IO_DEBUG(obj, "pageout [%lu, %lu], %d.\n", start, end, result);
+       RETURN(result);
 }
 
 /** @} osc */