Whamcloud - gitweb
LU-925 agl: trigger async glimpse lock when statahead
authorFan Yong <yong.fan@whamcloud.com>
Fri, 16 Dec 2011 07:34:25 +0000 (15:34 +0800)
committerOleg Drokin <green@whamcloud.com>
Wed, 4 Jan 2012 19:59:28 +0000 (14:59 -0500)
Client will send async glimpse lock RPCs to OSTs for file size
attribute before stat files to accelerate traversing large dir.

Signed-off-by: Fan Yong <yong.fan@whamcloud.com>
Change-Id: I9bf850abdd3c02c8470ddfcd91a3dc3ef7819c6d
Reviewed-on: http://review.whamcloud.com/1692
Reviewed-by: Lai Siyao <laisiyao@whamcloud.com>
Tested-by: Hudson
Tested-by: Maloo <whamcloud.maloo@gmail.com>
Reviewed-by: Oleg Drokin <green@whamcloud.com>
lustre/llite/file.c
lustre/llite/llite_internal.h
lustre/llite/llite_lib.c
lustre/llite/llite_mmap.c
lustre/llite/lproc_llite.c
lustre/llite/namei.c
lustre/llite/statahead.c

index 50763ed..42d0164 100644 (file)
@@ -1888,7 +1888,7 @@ loff_t ll_file_seek(struct file *file, loff_t offset, int origin)
         if (origin == 2) { /* SEEK_END */
                 int rc;
 
-                rc = cl_glimpse_size(inode);
+                rc = ll_glimpse_size(inode);
                 if (rc != 0)
                         RETURN(rc);
 
@@ -2351,11 +2351,11 @@ int ll_inode_revalidate_it(struct dentry *dentry, struct lookup_intent *it,
                 RETURN(0);
         }
 
-        /* cl_glimpse_size will prefer locally cached writes if they extend
+        /* ll_glimpse_size will prefer locally cached writes if they extend
          * the file */
 
         if (rc == 0)
-                rc = cl_glimpse_size(inode);
+                rc = ll_glimpse_size(inode);
 
         RETURN(rc);
 }
index 4803907..7cfba13 100644 (file)
@@ -172,6 +172,7 @@ struct ll_inode_info {
 
         /* the most recent timestamps obtained from mds */
         struct ost_lvb                  lli_lvb;
+        cfs_spinlock_t                  lli_agl_lock;
 
         /* Try to make the d::member and f::member are aligned. Before using
          * these members, make clear whether it is directory or not. */
@@ -224,6 +225,11 @@ struct ll_inode_info {
                         /* for writepage() only to communicate to fsync */
                         int                             f_async_rc;
                         int                             f_write_rc;
+
+                        cfs_rw_semaphore_t              f_glimpse_sem;
+                        cfs_time_t                      f_glimpse_time;
+                        cfs_list_t                      f_agl_list;
+                        __u64                           f_agl_index;
                 } f;
 
 #define lli_size_sem            u.f.f_size_sem
@@ -234,6 +240,10 @@ struct ll_inode_info {
 #define lli_write_sem           u.f.f_write_sem
 #define lli_async_rc            u.f.f_async_rc
 #define lli_write_rc            u.f.f_write_rc
+#define lli_glimpse_sem         u.f.f_glimpse_sem
+#define lli_glimpse_time        u.f.f_glimpse_time
+#define lli_agl_list            u.f.f_agl_list
+#define lli_agl_index           u.f.f_agl_index
 
         } u;
 
@@ -369,6 +379,7 @@ enum stats_track_type {
 #define LL_SBI_SOM_PREVIEW     0x1000 /* SOM preview mount option */
 #define LL_SBI_32BIT_API       0x2000 /* generate 32 bit inodes. */
 #define LL_SBI_64BIT_HASH      0x4000 /* support 64-bits dir hash/offset */
+#define LL_SBI_AGL_ENABLED     0x8000 /* enable agl */
 
 /* default value for ll_sb_info->contention_time */
 #define SBI_DEFAULT_CONTENTION_SECONDS     60
@@ -447,8 +458,8 @@ struct ll_sb_info {
         unsigned int              ll_offset_process_count;
         struct ll_rw_process_info ll_rw_offset_info[LL_OFFSET_HIST_MAX];
         unsigned int              ll_rw_offset_entry_count;
-        enum stats_track_type     ll_stats_track_type;
         int                       ll_stats_track_id;
+        enum stats_track_type     ll_stats_track_type;
         int                       ll_rw_stats_on;
 
         /* metadata stat-ahead */
@@ -457,6 +468,7 @@ struct ll_sb_info {
                                                   * count */
         atomic_t                  ll_sa_wrong;   /* statahead thread stopped for
                                                   * low hit ratio */
+        atomic_t                  ll_agl_total;  /* AGL thread started count */
 
         dev_t                     ll_sdev_orig; /* save s_dev before assign for
                                                  * clustred nfs */
@@ -1196,12 +1208,15 @@ struct ll_statahead_info {
         unsigned int            sai_skip_hidden;/* skipped hidden dentry count */
         unsigned int            sai_ls_all:1,   /* "ls -al", do stat-ahead for
                                                  * hidden entries */
-                                sai_in_readpage:1;/* statahead is in readdir()*/
+                                sai_in_readpage:1,/* statahead is in readdir()*/
+                                sai_agl_valid:1;/* AGL is valid for the dir */
         cfs_waitq_t             sai_waitq;      /* stat-ahead wait queue */
         struct ptlrpc_thread    sai_thread;     /* stat-ahead thread */
+        struct ptlrpc_thread    sai_agl_thread; /* AGL thread */
         cfs_list_t              sai_entries_sent;     /* entries sent out */
         cfs_list_t              sai_entries_received; /* entries returned */
         cfs_list_t              sai_entries_stated;   /* entries stated */
+        cfs_list_t              sai_entries_agl; /* AGL entries to be sent */
         cfs_list_t              sai_cache[LL_SA_CACHE_SIZE];
         cfs_spinlock_t          sai_cache_lock[LL_SA_CACHE_SIZE];
         cfs_atomic_t            sai_cache_count;      /* entry count in cache */
@@ -1211,6 +1226,18 @@ int do_statahead_enter(struct inode *dir, struct dentry **dentry,
                        int only_unplug);
 void ll_stop_statahead(struct inode *dir, void *key);
 
+static inline int ll_glimpse_size(struct inode *inode)
+{
+        struct ll_inode_info *lli = ll_i2info(inode);
+        int rc;
+
+        cfs_down_read(&lli->lli_glimpse_sem);
+        rc = cl_glimpse_size(inode);
+        lli->lli_glimpse_time = cfs_time_current();
+        cfs_up_read(&lli->lli_glimpse_sem);
+        return rc;
+}
+
 static inline void
 ll_statahead_mark(struct inode *dir, struct dentry *dentry)
 {
index 70d0fcb..41624bb 100644 (file)
@@ -142,6 +142,8 @@ static struct ll_sb_info *ll_init_sbi(void)
         sbi->ll_sa_max = LL_SA_RPC_DEF;
         cfs_atomic_set(&sbi->ll_sa_total, 0);
         cfs_atomic_set(&sbi->ll_sa_wrong, 0);
+        cfs_atomic_set(&sbi->ll_agl_total, 0);
+        sbi->ll_flags |= LL_SBI_AGL_ENABLED;
 
         RETURN(sbi);
 }
@@ -875,6 +877,7 @@ void ll_lli_init(struct ll_inode_info *lli)
         lli->lli_open_fd_write_count = 0;
         lli->lli_open_fd_exec_count = 0;
         cfs_sema_init(&lli->lli_och_sem, 1);
+        cfs_spin_lock_init(&lli->lli_agl_lock);
         lli->lli_smd = NULL;
         lli->lli_clob = NULL;
 
@@ -896,6 +899,10 @@ void ll_lli_init(struct ll_inode_info *lli)
                 cfs_sema_init(&lli->lli_write_sem, 1);
                 lli->lli_async_rc = 0;
                 lli->lli_write_rc = 0;
+                cfs_init_rwsem(&lli->lli_glimpse_sem);
+                lli->lli_glimpse_time = 0;
+                CFS_INIT_LIST_HEAD(&lli->lli_agl_list);
+                lli->lli_agl_index = 0;
         }
 }
 
@@ -1168,6 +1175,9 @@ void ll_clear_inode(struct inode *inode)
         lli->lli_inode_magic = LLI_INODE_DEAD;
 
         ll_clear_inode_capas(inode);
+        if (!S_ISDIR(inode->i_mode))
+                LASSERT(cfs_list_empty(&lli->lli_agl_list));
+
         /*
          * XXX This has to be done before lsm is freed below, because
          * cl_object still uses inode lsm.
index 821f96f..9de415d 100644 (file)
@@ -603,7 +603,7 @@ int ll_file_mmap(struct file *file, struct vm_area_struct * vma)
                 vma->vm_ops = &ll_file_vm_ops;
                 vma->vm_ops->open(vma);
                 /* update the inode's size and mtime */
-                rc = cl_glimpse_size(inode);
+                rc = ll_glimpse_size(inode);
         }
 
         RETURN(rc);
index 0845ba7..3bfe608 100644 (file)
@@ -30,6 +30,9 @@
  * Use is subject to license terms.
  */
 /*
+ * Copyright (c) 2011 Whamcloud, Inc.
+ */
+/*
  * This file is part of Lustre, http://www.lustre.org/
  * Lustre is a trademark of Sun Microsystems, Inc.
  */
@@ -562,6 +565,35 @@ static int ll_wr_statahead_max(struct file *file, const char *buffer,
         return count;
 }
 
+static int ll_rd_statahead_agl(char *page, char **start, off_t off,
+                               int count, int *eof, void *data)
+{
+        struct super_block *sb = data;
+        struct ll_sb_info *sbi = ll_s2sbi(sb);
+
+        return snprintf(page, count, "%u\n",
+                        sbi->ll_flags & LL_SBI_AGL_ENABLED ? 1 : 0);
+}
+
+static int ll_wr_statahead_agl(struct file *file, const char *buffer,
+                               unsigned long count, void *data)
+{
+        struct super_block *sb = data;
+        struct ll_sb_info *sbi = ll_s2sbi(sb);
+        int val, rc;
+
+        rc = lprocfs_write_helper(buffer, count, &val);
+        if (rc)
+                return rc;
+
+        if (val)
+                sbi->ll_flags |= LL_SBI_AGL_ENABLED;
+        else
+                sbi->ll_flags &= ~LL_SBI_AGL_ENABLED;
+
+        return count;
+}
+
 static int ll_rd_statahead_stats(char *page, char **start, off_t off,
                                  int count, int *eof, void *data)
 {
@@ -570,9 +602,11 @@ static int ll_rd_statahead_stats(char *page, char **start, off_t off,
 
         return snprintf(page, count,
                         "statahead total: %u\n"
-                        "statahead wrong: %u\n",
+                        "statahead wrong: %u\n"
+                        "agl total: %u\n",
                         atomic_read(&sbi->ll_sa_total),
-                        atomic_read(&sbi->ll_sa_wrong));
+                        atomic_read(&sbi->ll_sa_wrong),
+                        atomic_read(&sbi->ll_agl_total));
 }
 
 static int ll_rd_lazystatfs(char *page, char **start, off_t off,
@@ -630,6 +664,7 @@ static struct lprocfs_vars lprocfs_llite_obd_vars[] = {
         { "stats_track_ppid", ll_rd_track_ppid, ll_wr_track_ppid, 0 },
         { "stats_track_gid",  ll_rd_track_gid, ll_wr_track_gid, 0 },
         { "statahead_max",    ll_rd_statahead_max, ll_wr_statahead_max, 0 },
+        { "statahead_agl",    ll_rd_statahead_agl, ll_wr_statahead_agl, 0 },
         { "statahead_stats",  ll_rd_statahead_stats, 0, 0 },
         { "lazystatfs",         ll_rd_lazystatfs, ll_wr_lazystatfs, 0 },
         { 0 }
index 52945c4..363bcb4 100644 (file)
@@ -469,7 +469,7 @@ int ll_lookup_it_finish(struct ptlrpc_request *request,
                    2.4 and
                    vfs_getattr_it->ll_getattr()->ll_inode_revalidate_it() in 2.6
                    Everybody else who needs correct file size would call
-                   cl_glimpse_size or some equivalent themselves anyway.
+                   ll_glimpse_size or some equivalent themselves anyway.
                    Also see bug 7198. */
                 ll_dops_init(*de, 1, 1);
                 *de = ll_find_alias(inode, *de);
index cb1883b..537b44d 100644 (file)
@@ -140,14 +140,42 @@ ll_sa_entry_unhash(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
         cfs_spin_unlock(&sai->sai_cache_lock[i]);
 }
 
+static inline int agl_should_run(struct ll_statahead_info *sai,
+                                 struct inode *inode)
+{
+        if (inode != NULL && S_ISREG(inode->i_mode) &&
+            ll_i2info(inode)->lli_smd != NULL && sai->sai_agl_valid)
+                return 1;
+        return 0;
+}
+
+static inline struct ll_sa_entry *
+sa_first_received_entry(struct ll_statahead_info *sai)
+{
+        return cfs_list_entry(sai->sai_entries_received.next,
+                              struct ll_sa_entry, se_list);
+}
+
+static inline struct ll_inode_info *
+agl_first_entry(struct ll_statahead_info *sai)
+{
+        return cfs_list_entry(sai->sai_entries_agl.next,
+                              struct ll_inode_info, lli_agl_list);
+}
+
+static inline int sa_sent_full(struct ll_statahead_info *sai)
+{
+        return cfs_atomic_read(&sai->sai_cache_count) >= sai->sai_max;
+}
+
 static inline int sa_received_empty(struct ll_statahead_info *sai)
 {
         return cfs_list_empty(&sai->sai_entries_received);
 }
 
-static inline int sa_not_full(struct ll_statahead_info *sai)
+static inline int agl_list_empty(struct ll_statahead_info *sai)
 {
-        return (cfs_atomic_read(&sai->sai_cache_count) < sai->sai_max);
+        return cfs_list_empty(&sai->sai_entries_agl);
 }
 
 /**
@@ -417,6 +445,37 @@ ll_sa_entry_to_stated(struct ll_statahead_info *sai,
         return ret;
 }
 
+/*
+ * Insert inode into the list of sai_entries_agl.
+ */
+static void ll_agl_add(struct ll_statahead_info *sai,
+                       struct inode *inode, int index)
+{
+        struct ll_inode_info *child  = ll_i2info(inode);
+        struct ll_inode_info *parent = ll_i2info(sai->sai_inode);
+        int                   added  = 0;
+
+        cfs_spin_lock(&child->lli_agl_lock);
+        if (child->lli_agl_index == 0) {
+                child->lli_agl_index = index;
+                cfs_spin_unlock(&child->lli_agl_lock);
+
+                LASSERT(cfs_list_empty(&child->lli_agl_list));
+
+                igrab(inode);
+                cfs_spin_lock(&parent->lli_agl_lock);
+                if (agl_list_empty(sai))
+                        added = 1;
+                cfs_list_add_tail(&child->lli_agl_list, &sai->sai_entries_agl);
+                cfs_spin_unlock(&parent->lli_agl_lock);
+        } else {
+                cfs_spin_unlock(&child->lli_agl_lock);
+        }
+
+        if (added > 0)
+                cfs_waitq_signal(&sai->sai_agl_thread.t_ctl_waitq);
+}
+
 static struct ll_statahead_info *ll_sai_alloc(void)
 {
         struct ll_statahead_info *sai;
@@ -428,17 +487,24 @@ static struct ll_statahead_info *ll_sai_alloc(void)
                 RETURN(NULL);
 
         cfs_atomic_set(&sai->sai_refcount, 1);
+
         cfs_spin_lock(&sai_generation_lock);
         sai->sai_generation = ++sai_generation;
         if (unlikely(sai_generation == 0))
                 sai->sai_generation = ++sai_generation;
         cfs_spin_unlock(&sai_generation_lock);
+
         sai->sai_max = LL_SA_RPC_MIN;
+        sai->sai_index = 1;
         cfs_waitq_init(&sai->sai_waitq);
         cfs_waitq_init(&sai->sai_thread.t_ctl_waitq);
+        cfs_waitq_init(&sai->sai_agl_thread.t_ctl_waitq);
+
         CFS_INIT_LIST_HEAD(&sai->sai_entries_sent);
         CFS_INIT_LIST_HEAD(&sai->sai_entries_received);
         CFS_INIT_LIST_HEAD(&sai->sai_entries_stated);
+        CFS_INIT_LIST_HEAD(&sai->sai_entries_agl);
+
         for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
                 CFS_INIT_LIST_HEAD(&sai->sai_cache[i]);
                 cfs_spin_lock_init(&sai->sai_cache_lock[i]);
@@ -472,12 +538,13 @@ static void ll_sai_put(struct ll_statahead_info *sai)
                 }
 
                 LASSERT(lli->lli_opendir_key == NULL);
+                LASSERT(thread_is_stopped(&sai->sai_thread));
+                LASSERT(thread_is_stopped(&sai->sai_agl_thread));
+
                 lli->lli_sai = NULL;
                 lli->lli_opendir_pid = 0;
                 cfs_spin_unlock(&lli->lli_sa_lock);
 
-                LASSERT(thread_is_stopped(&sai->sai_thread));
-
                 if (sai->sai_sent > sai->sai_replied)
                         CDEBUG(D_READA,"statahead for dir "DFID" does not "
                               "finish: [sent:"LPU64"] [replied:"LPU64"]\n",
@@ -495,6 +562,7 @@ static void ll_sai_put(struct ll_statahead_info *sai)
                         do_sai_entry_fini(sai, entry);
 
                 LASSERT(cfs_atomic_read(&sai->sai_cache_count) == 0);
+                LASSERT(agl_list_empty(sai));
 
                 iput(inode);
                 OBD_FREE_PTR(sai);
@@ -503,6 +571,73 @@ static void ll_sai_put(struct ll_statahead_info *sai)
         EXIT;
 }
 
+#ifndef HAVE_AGL_SUPPORT
+# define cl_agl(inode)  do {} while (0)
+#endif
+
+/* Do NOT forget to drop inode refcount when into sai_entries_agl. */
+static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai)
+{
+        struct ll_inode_info *lli   = ll_i2info(inode);
+        __u64                 index = lli->lli_agl_index;
+        int                   rc;
+        ENTRY;
+
+        LASSERT(cfs_list_empty(&lli->lli_agl_list));
+
+        /* AGL maybe fall behind statahead with one entry */
+        if (is_omitted_entry(sai, index + 1)) {
+                lli->lli_agl_index = 0;
+                iput(inode);
+                RETURN_EXIT;
+        }
+
+        /* Someone is in glimpse (sync or async), do nothing. */
+        rc = cfs_down_write_trylock(&lli->lli_glimpse_sem);
+        if (rc == 0) {
+                lli->lli_agl_index = 0;
+                iput(inode);
+                RETURN_EXIT;
+        }
+
+        /*
+         * Someone triggered glimpse within 1 sec before.
+         * 1) The former glimpse succeeded with glimpse lock granted by OST, and
+         *    if the lock is still cached on client, AGL needs to do nothing. If
+         *    it is cancelled by other client, AGL maybe cannot obtaion new lock
+         *    for no glimpse callback triggered by AGL.
+         * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
+         *    Under such case, it is quite possible that the OST will not grant
+         *    glimpse lock for AGL also.
+         * 3) The former glimpse failed, compared with other two cases, it is
+         *    relative rare. AGL can ignore such case, and it will not muchly
+         *    affect the performance.
+         */
+        if (lli->lli_glimpse_time != 0 &&
+            cfs_time_before(cfs_time_shift(-1), lli->lli_glimpse_time)) {
+                cfs_up_write(&lli->lli_glimpse_sem);
+                lli->lli_agl_index = 0;
+                iput(inode);
+                RETURN_EXIT;
+        }
+
+        CDEBUG(D_READA, "Handling (init) async glimpse: inode = "
+               DFID", idx = "LPU64"\n", PFID(&lli->lli_fid), index);
+
+        cl_agl(inode);
+        lli->lli_agl_index = 0;
+        lli->lli_glimpse_time = cfs_time_current();
+        cfs_up_write(&lli->lli_glimpse_sem);
+
+        CDEBUG(D_READA, "Handled (init) async glimpse: inode= "
+               DFID", idx = "LPU64", rc = %d\n",
+               PFID(&lli->lli_fid), index, rc);
+
+        iput(inode);
+
+        EXIT;
+}
+
 static void do_statahead_interpret(struct ll_statahead_info *sai,
                                    struct ll_sa_entry *target)
 {
@@ -525,8 +660,7 @@ static void do_statahead_interpret(struct ll_statahead_info *sai,
                 cfs_spin_unlock(&lli->lli_sa_lock);
                 RETURN_EXIT;
         } else {
-                entry = cfs_list_entry(sai->sai_entries_received.next,
-                                       struct ll_sa_entry, se_list);
+                entry = sa_first_received_entry(sai);
         }
 
         cfs_atomic_inc(&entry->se_refcount);
@@ -580,6 +714,9 @@ static void do_statahead_interpret(struct ll_statahead_info *sai,
 
         entry->se_inode = child;
 
+        if (agl_should_run(sai, child))
+                ll_agl_add(sai, child, entry->se_index);
+
         EXIT;
 
 out:
@@ -689,11 +826,12 @@ static void sa_args_fini(struct md_enqueue_info *minfo,
  * "md_intent_getattr_async".
  */
 static int sa_args_init(struct inode *dir, struct inode *child,
-                        struct qstr *qstr, struct md_enqueue_info **pmi,
+                        struct ll_sa_entry *entry, struct md_enqueue_info **pmi,
                         struct ldlm_enqueue_info **pei,
                         struct obd_capa **pcapa)
 {
-        struct ll_inode_info     *lli = ll_i2info(dir);
+        struct qstr              *qstr = &entry->se_qstr;
+        struct ll_inode_info     *lli  = ll_i2info(dir);
         struct md_enqueue_info   *minfo;
         struct ldlm_enqueue_info *einfo;
         struct md_op_data        *op_data;
@@ -720,7 +858,7 @@ static int sa_args_init(struct inode *dir, struct inode *child,
         minfo->mi_dir = igrab(dir);
         minfo->mi_cb = ll_statahead_interpret;
         minfo->mi_generation = lli->lli_sai->sai_generation;
-        minfo->mi_cbdata = lli->lli_sai->sai_index;
+        minfo->mi_cbdata = entry->se_index;
 
         einfo->ei_type   = LDLM_IBITS;
         einfo->ei_mode   = it_to_lock_mode(&minfo->mi_it);
@@ -745,7 +883,7 @@ static int do_sa_lookup(struct inode *dir, struct ll_sa_entry *entry)
         int                       rc;
         ENTRY;
 
-        rc = sa_args_init(dir, NULL, &entry->se_qstr, &minfo, &einfo, capas);
+        rc = sa_args_init(dir, NULL, entry, &minfo, &einfo, capas);
         if (rc)
                 RETURN(rc);
 
@@ -795,7 +933,7 @@ static int do_sa_revalidate(struct inode *dir, struct ll_sa_entry *entry,
                 RETURN(1);
         }
 
-        rc = sa_args_init(dir, inode, &entry->se_qstr, &minfo, &einfo, capas);
+        rc = sa_args_init(dir, inode, entry, &minfo, &einfo, capas);
         if (rc) {
                 entry->se_inode = NULL;
                 iput(inode);
@@ -818,9 +956,9 @@ static int do_sa_revalidate(struct inode *dir, struct ll_sa_entry *entry,
 static void ll_statahead_one(struct dentry *parent, const char* entry_name,
                              int entry_name_len)
 {
-        struct inode             *dir = parent->d_inode;
-        struct ll_inode_info     *lli = ll_i2info(dir);
-        struct ll_statahead_info *sai = lli->lli_sai;
+        struct inode             *dir    = parent->d_inode;
+        struct ll_inode_info     *lli    = ll_i2info(dir);
+        struct ll_statahead_info *sai    = lli->lli_sai;
         struct dentry            *dentry = NULL;
         struct ll_sa_entry       *entry;
         int                       rc;
@@ -833,10 +971,13 @@ static void ll_statahead_one(struct dentry *parent, const char* entry_name,
                 RETURN_EXIT;
 
         dentry = d_lookup(parent, &entry->se_qstr);
-        if (!dentry)
+        if (!dentry) {
                 rc = do_sa_lookup(dir, entry);
-        else
+        } else {
                 rc = do_sa_revalidate(dir, entry, dentry);
+                if (rc == 1 && agl_should_run(sai, dentry->d_inode))
+                        ll_agl_add(sai, dentry->d_inode, entry->se_index);
+        }
 
         if (dentry != NULL)
                 dput(dentry);
@@ -857,40 +998,137 @@ static void ll_statahead_one(struct dentry *parent, const char* entry_name,
         EXIT;
 }
 
+static int ll_agl_thread(void *arg)
+{
+        struct dentry            *parent = (struct dentry *)arg;
+        struct inode             *dir    = parent->d_inode;
+        struct ll_inode_info     *plli   = ll_i2info(dir);
+        struct ll_inode_info     *clli;
+        struct ll_sb_info        *sbi    = ll_i2sbi(dir);
+        struct ll_statahead_info *sai    = ll_sai_get(plli->lli_sai);
+        struct ptlrpc_thread     *thread = &sai->sai_agl_thread;
+        struct l_wait_info        lwi    = { 0 };
+        ENTRY;
+
+        {
+                char pname[16];
+                snprintf(pname, 15, "ll_agl_%u", plli->lli_opendir_pid);
+                cfs_daemonize(pname);
+        }
+
+        CDEBUG(D_READA, "agl thread started: [pid %d] [parent %.*s]\n",
+               cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
+
+        atomic_inc(&sbi->ll_agl_total);
+        cfs_spin_lock(&plli->lli_agl_lock);
+        sai->sai_agl_valid = 1;
+        thread_set_flags(thread, SVC_RUNNING);
+        cfs_spin_unlock(&plli->lli_agl_lock);
+        cfs_waitq_signal(&thread->t_ctl_waitq);
+
+        while (1) {
+                l_wait_event(thread->t_ctl_waitq,
+                             !agl_list_empty(sai) ||
+                             !thread_is_running(thread),
+                             &lwi);
+
+                if (!thread_is_running(thread))
+                        break;
+
+                cfs_spin_lock(&plli->lli_agl_lock);
+                /* The statahead thread maybe help to process AGL entries,
+                 * so check whether list empty again. */
+                if (!agl_list_empty(sai)) {
+                        clli = agl_first_entry(sai);
+                        cfs_list_del_init(&clli->lli_agl_list);
+                        cfs_spin_unlock(&plli->lli_agl_lock);
+                        ll_agl_trigger(&clli->lli_vfs_inode, sai);
+                } else {
+                        cfs_spin_unlock(&plli->lli_agl_lock);
+                }
+        }
+
+        cfs_spin_lock(&plli->lli_agl_lock);
+        sai->sai_agl_valid = 0;
+        while (!agl_list_empty(sai)) {
+                clli = agl_first_entry(sai);
+                cfs_list_del_init(&clli->lli_agl_list);
+                cfs_spin_unlock(&plli->lli_agl_lock);
+                clli->lli_agl_index = 0;
+                iput(&clli->lli_vfs_inode);
+                cfs_spin_lock(&plli->lli_agl_lock);
+        }
+        thread_set_flags(thread, SVC_STOPPED);
+        cfs_spin_unlock(&plli->lli_agl_lock);
+        cfs_waitq_signal(&thread->t_ctl_waitq);
+        ll_sai_put(sai);
+        CDEBUG(D_READA, "agl thread stopped: [pid %d] [parent %.*s]\n",
+               cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
+        RETURN(0);
+}
+
+static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
+{
+        struct ptlrpc_thread *thread = &sai->sai_agl_thread;
+        struct l_wait_info    lwi    = { 0 };
+        int                   rc;
+        ENTRY;
+
+        CDEBUG(D_READA, "start agl thread: [pid %d] [parent %.*s]\n",
+               cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
+
+        rc = cfs_create_thread(ll_agl_thread, parent, 0);
+        if (rc < 0) {
+                CERROR("can't start ll_agl thread, rc: %d\n", rc);
+                RETURN_EXIT;
+        }
+
+        l_wait_event(thread->t_ctl_waitq,
+                     thread_is_running(thread) || thread_is_stopped(thread),
+                     &lwi);
+        EXIT;
+}
+
 static int ll_statahead_thread(void *arg)
 {
         struct dentry            *parent = (struct dentry *)arg;
-        struct inode             *dir = parent->d_inode;
-        struct ll_inode_info     *lli = ll_i2info(dir);
-        struct ll_sb_info        *sbi = ll_i2sbi(dir);
-        struct ll_statahead_info *sai = ll_sai_get(lli->lli_sai);
+        struct inode             *dir    = parent->d_inode;
+        struct ll_inode_info     *plli   = ll_i2info(dir);
+        struct ll_inode_info     *clli;
+        struct ll_sb_info        *sbi    = ll_i2sbi(dir);
+        struct ll_statahead_info *sai    = ll_sai_get(plli->lli_sai);
         struct ptlrpc_thread     *thread = &sai->sai_thread;
         struct page              *page;
-        __u64                     pos = 0;
-        int                       first = 0;
-        int                       rc = 0;
+        __u64                     pos    = 0;
+        int                       first  = 0;
+        int                       rc     = 0;
         struct ll_dir_chain       chain;
+        struct l_wait_info        lwi    = { 0 };
         ENTRY;
 
         {
                 char pname[16];
-                snprintf(pname, 15, "ll_sa_%u", lli->lli_opendir_pid);
+                snprintf(pname, 15, "ll_sa_%u", plli->lli_opendir_pid);
                 cfs_daemonize(pname);
         }
 
+        CDEBUG(D_READA, "statahead thread started: [pid %d] [parent %.*s]\n",
+               cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
+
+        if (sbi->ll_flags & LL_SBI_AGL_ENABLED)
+                ll_start_agl(parent, sai);
+
         atomic_inc(&sbi->ll_sa_total);
-        cfs_spin_lock(&lli->lli_sa_lock);
+        cfs_spin_lock(&plli->lli_sa_lock);
         thread_set_flags(thread, SVC_RUNNING);
-        cfs_spin_unlock(&lli->lli_sa_lock);
+        cfs_spin_unlock(&plli->lli_sa_lock);
         cfs_waitq_signal(&thread->t_ctl_waitq);
-        CDEBUG(D_READA, "start doing statahead for %s\n", parent->d_name.name);
 
-        lli->lli_sa_pos = 0;
+        plli->lli_sa_pos = 0;
         ll_dir_chain_init(&chain);
         page = ll_get_dir_page(NULL, dir, pos, &chain);
 
         while (1) {
-                struct l_wait_info lwi = { 0 };
                 struct lu_dirpage *dp;
                 struct lu_dirent  *ent;
 
@@ -899,7 +1137,7 @@ static int ll_statahead_thread(void *arg)
                         CDEBUG(D_READA, "error reading dir "DFID" at "LPU64
                                "/"LPU64": [rc %d] [parent %u]\n",
                                PFID(ll_inode2fid(dir)), pos, sai->sai_index,
-                               rc, lli->lli_opendir_pid);
+                               rc, plli->lli_opendir_pid);
                         GOTO(out, rc);
                 }
 
@@ -951,27 +1189,55 @@ static int ll_statahead_thread(void *arg)
                         if (unlikely(++first == 1))
                                 continue;
 
-keep_de:
+keep_it:
                         l_wait_event(thread->t_ctl_waitq,
-                                     sa_not_full(sai) ||
+                                     !sa_sent_full(sai) ||
                                      !sa_received_empty(sai) ||
+                                     !agl_list_empty(sai) ||
                                      !thread_is_running(thread),
                                      &lwi);
 
+interpret_it:
                         while (!sa_received_empty(sai))
                                 do_statahead_interpret(sai, NULL);
 
                         if (unlikely(!thread_is_running(thread))) {
                                 ll_release_page(page, 0);
-                                GOTO(out, rc);
+                                GOTO(out, rc = 0);
                         }
 
-                        if (!sa_not_full(sai))
-                                /*
-                                 * do not skip the current de.
-                                 */
-                                goto keep_de;
+                        /* If no window for metadata statahead, but there are
+                         * some AGL entries to be triggered, then try to help
+                         * to process the AGL entries. */
+                        if (sa_sent_full(sai)) {
+                                cfs_spin_lock(&plli->lli_agl_lock);
+                                while (!agl_list_empty(sai)) {
+                                        clli = agl_first_entry(sai);
+                                        cfs_list_del_init(&clli->lli_agl_list);
+                                        cfs_spin_unlock(&plli->lli_agl_lock);
+                                        ll_agl_trigger(&clli->lli_vfs_inode,
+                                                       sai);
+
+                                        if (!sa_received_empty(sai))
+                                                goto interpret_it;
+
+                                        if (unlikely(
+                                                !thread_is_running(thread))) {
+                                                ll_release_page(page, 0);
+                                                GOTO(out, rc = 0);
+                                        }
+
+                                        if (!sa_sent_full(sai))
+                                                goto do_it;
+
+                                        cfs_spin_lock(&plli->lli_agl_lock);
+                                }
+                                cfs_spin_unlock(&plli->lli_agl_lock);
+
+                                goto keep_it;
+                        }
 
+do_it:
                         ll_statahead_one(parent, name, namelen);
                 }
                 pos = le64_to_cpu(dp->ldp_hash_end);
@@ -990,11 +1256,26 @@ keep_de:
                                 while (!sa_received_empty(sai))
                                         do_statahead_interpret(sai, NULL);
 
-                                if ((sai->sai_sent == sai->sai_replied &&
-                                     sa_received_empty(sai)) ||
-                                    !thread_is_running(thread))
+                                if (unlikely(!thread_is_running(thread)))
                                         GOTO(out, rc = 0);
+
+                                if (sai->sai_sent == sai->sai_replied &&
+                                    sa_received_empty(sai))
+                                        break;
+                        }
+
+                        cfs_spin_lock(&plli->lli_agl_lock);
+                        while (!agl_list_empty(sai) &&
+                               thread_is_running(thread)) {
+                                clli = agl_first_entry(sai);
+                                cfs_list_del_init(&clli->lli_agl_list);
+                                cfs_spin_unlock(&plli->lli_agl_lock);
+                                ll_agl_trigger(&clli->lli_vfs_inode, sai);
+                                cfs_spin_lock(&plli->lli_agl_lock);
                         }
+                        cfs_spin_unlock(&plli->lli_agl_lock);
+
+                        GOTO(out, rc = 0);
                 } else if (1) {
                         /*
                          * chain is exhausted.
@@ -1002,7 +1283,7 @@ keep_de:
                          */
                         ll_release_page(page, le32_to_cpu(dp->ldp_flags) &
                                               LDF_COLLIDE);
-                        lli->lli_sa_pos = pos;
+                        plli->lli_sa_pos = pos;
                         sai->sai_in_readpage = 1;
                         page = ll_get_dir_page(NULL, dir, pos, &chain);
                         sai->sai_in_readpage = 0;
@@ -1017,26 +1298,41 @@ keep_de:
         EXIT;
 
 out:
+        if (sai->sai_agl_valid) {
+                struct ptlrpc_thread *agl_thread = &sai->sai_agl_thread;
+
+                cfs_spin_lock(&plli->lli_agl_lock);
+                thread_set_flags(agl_thread, SVC_STOPPING);
+                cfs_spin_unlock(&plli->lli_agl_lock);
+                cfs_waitq_signal(&agl_thread->t_ctl_waitq);
+
+                CDEBUG(D_READA, "stop agl thread: [pid %d]\n",
+                       cfs_curproc_pid());
+                l_wait_event(agl_thread->t_ctl_waitq,
+                             thread_is_stopped(agl_thread),
+                             &lwi);
+        }
+
         ll_dir_chain_fini(&chain);
-        cfs_spin_lock(&lli->lli_sa_lock);
+        cfs_spin_lock(&plli->lli_sa_lock);
         if (!sa_received_empty(sai)) {
                 thread_set_flags(thread, SVC_STOPPING);
-                cfs_spin_unlock(&lli->lli_sa_lock);
+                cfs_spin_unlock(&plli->lli_sa_lock);
 
                 /* To release the resources held by received entries. */
                 while (!sa_received_empty(sai))
                         do_statahead_interpret(sai, NULL);
 
-                cfs_spin_lock(&lli->lli_sa_lock);
+                cfs_spin_lock(&plli->lli_sa_lock);
         }
         thread_set_flags(thread, SVC_STOPPED);
-        cfs_spin_unlock(&lli->lli_sa_lock);
+        cfs_spin_unlock(&plli->lli_sa_lock);
         cfs_waitq_signal(&sai->sai_waitq);
         cfs_waitq_signal(&thread->t_ctl_waitq);
         ll_sai_put(sai);
         dput(parent);
-        CDEBUG(D_READA, "statahead thread stopped, pid %d\n",
-               cfs_curproc_pid());
+        CDEBUG(D_READA, "statahead thread stopped: [pid %d] [parent %.*s]\n",
+               cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
         return rc;
 }
 
@@ -1067,7 +1363,7 @@ void ll_stop_statahead(struct inode *dir, void *key)
                         cfs_spin_unlock(&lli->lli_sa_lock);
                         cfs_waitq_signal(&thread->t_ctl_waitq);
 
-                        CDEBUG(D_READA, "stopping statahead thread, pid %d\n",
+                        CDEBUG(D_READA, "stop statahead thread: [pid %d]\n",
                                cfs_curproc_pid());
                         l_wait_event(thread->t_ctl_waitq,
                                      thread_is_stopped(thread),
@@ -1105,13 +1401,13 @@ enum {
 
 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
 {
-        struct ll_inode_info *lli = ll_i2info(dir);
-        struct ll_dir_chain chain;
-        struct qstr        *target = &dentry->d_name;
-        struct page        *page;
-        __u64               pos = 0;
-        int                 dot_de;
-        int                 rc = LS_NONE_FIRST_DE;
+        struct ll_inode_info *lli    = ll_i2info(dir);
+        struct ll_dir_chain   chain;
+        struct qstr          *target = &dentry->d_name;
+        struct page          *page;
+        __u64                 pos    = 0;
+        int                   dot_de;
+        int                   rc     = LS_NONE_FIRST_DE;
         ENTRY;
 
         lli->lli_sa_pos = 0;
@@ -1225,7 +1521,7 @@ ll_sai_unplug(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 {
         struct ptlrpc_thread *thread = &sai->sai_thread;
         struct ll_sb_info    *sbi    = ll_i2sbi(sai->sai_inode);
-        int                  hit;
+        int                   hit;
         ENTRY;
 
         if (entry != NULL && entry->se_stat == SA_ENTRY_SUCC)
@@ -1416,6 +1712,9 @@ int do_statahead_enter(struct inode *dir, struct dentry **dentryp,
                 GOTO(out, rc = -EAGAIN);
         }
 
+        CDEBUG(D_READA, "start statahead thread: [pid %d] [parent %.*s]\n",
+               cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
+
         lli->lli_sai = sai;
         rc = cfs_create_thread(ll_statahead_thread, parent, 0);
         thread = &sai->sai_thread;