Whamcloud - gitweb
LU-3270 statahead: use dcache-like interface for sa entry
[fs/lustre-release.git] / lustre / llite / statahead.c
index f781f7c..45bfe84 100644 (file)
 #define SA_OMITTED_ENTRY_MAX 8ULL
 
 typedef enum {
-        /** negative values are for error cases */
-        SA_ENTRY_INIT = 0,      /** init entry */
-        SA_ENTRY_SUCC = 1,      /** stat succeed */
-        SA_ENTRY_INVA = 2,      /** invalid entry */
-        SA_ENTRY_DEST = 3,      /** entry to be destroyed */
-} se_stat_t;
-
-struct ll_sa_entry {
-       /* link into sai->sai_entries */
-       struct list_head        se_link;
-       /* link into sai->sai_entries_{received,stated} */
+       /** negative values are for error cases */
+       SA_ENTRY_INIT = 0,      /** init entry */
+       SA_ENTRY_SUCC = 1,      /** stat succeed */
+       SA_ENTRY_INVA = 2,      /** invalid entry */
+} se_state_t;
+
+/* sa_entry is not refcounted: statahead thread allocates it and do async stat,
+ * and in async stat callback ll_statahead_interpret() will add it into
+ * sai_cb_entries, later statahead thread will call sa_handle_callback() to
+ * instantiate entry and move it into sai_entries, and then only scanner process
+ * can access and free it. */
+struct sa_entry {
+       /* link into sai_cb_entries or sai_entries */
        struct list_head        se_list;
        /* link into sai hash table locally */
        struct list_head        se_hash;
-       /* entry reference count */
-       atomic_t                se_refcount;
        /* entry index in the sai */
-       __u64                   se_index;
+       __u64                   se_index;
        /* low layer ldlm lock handle */
-       __u64                   se_handle;
+       __u64                   se_handle;
        /* entry status */
-       se_stat_t               se_stat;
+       se_state_t              se_state;
        /* entry size, contains name */
-       int                     se_size;
+       int                     se_size;
        /* pointer to async getattr enqueue info */
        struct md_enqueue_info *se_minfo;
        /* pointer to the async getattr request */
        struct ptlrpc_request  *se_req;
        /* pointer to the target inode */
-       struct inode           *se_inode;
+       struct inode           *se_inode;
        /* entry name */
-       struct qstr             se_qstr;
+       struct qstr             se_qstr;
 };
 
 static unsigned int sai_generation = 0;
 static DEFINE_SPINLOCK(sai_generation_lock);
 
-static inline int ll_sa_entry_unhashed(struct ll_sa_entry *entry)
+static inline int sa_unhashed(struct sa_entry *entry)
 {
        return list_empty(&entry->se_hash);
 }
@@ -94,24 +94,24 @@ static inline int ll_sa_entry_unhashed(struct ll_sa_entry *entry)
 /*
  * The entry only can be released by the caller, it is necessary to hold lock.
  */
-static inline int ll_sa_entry_stated(struct ll_sa_entry *entry)
+static inline int sa_ready(struct sa_entry *entry)
 {
-        smp_rmb();
-        return (entry->se_stat != SA_ENTRY_INIT);
+       smp_rmb();
+       return (entry->se_state != SA_ENTRY_INIT);
 }
 
-static inline int ll_sa_entry_hash(int val)
+static inline int sa_hash(int val)
 {
-        return val & LL_SA_CACHE_MASK;
+       return val & LL_SA_CACHE_MASK;
 }
 
 /*
  * Insert entry to hash SA table.
  */
 static inline void
-ll_sa_entry_enhash(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
+sa_rehash(struct ll_statahead_info *sai, struct sa_entry *entry)
 {
-       int i = ll_sa_entry_hash(entry->se_qstr.hash);
+       int i = sa_hash(entry->se_qstr.hash);
 
        spin_lock(&sai->sai_cache_lock[i]);
        list_add_tail(&entry->se_hash, &sai->sai_cache[i]);
@@ -122,9 +122,9 @@ ll_sa_entry_enhash(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
  * Remove entry from SA table.
  */
 static inline void
-ll_sa_entry_unhash(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
+sa_unhash(struct ll_statahead_info *sai, struct sa_entry *entry)
 {
-       int i = ll_sa_entry_hash(entry->se_qstr.hash);
+       int i = sa_hash(entry->se_qstr.hash);
 
        spin_lock(&sai->sai_cache_lock[i]);
        list_del_init(&entry->se_hash);
@@ -137,18 +137,11 @@ static inline int agl_should_run(struct ll_statahead_info *sai,
        return (inode != NULL && S_ISREG(inode->i_mode) && sai->sai_agl_valid);
 }
 
-static inline struct ll_sa_entry *
-sa_first_received_entry(struct ll_statahead_info *sai)
-{
-       return 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 list_entry(sai->sai_entries_agl.next,
-                         struct ll_inode_info, lli_agl_list);
+       return list_entry(sai->sai_agls.next, struct ll_inode_info,
+                         lli_agl_list);
 }
 
 static inline int sa_sent_full(struct ll_statahead_info *sai)
@@ -156,14 +149,14 @@ static inline int sa_sent_full(struct ll_statahead_info *sai)
        return atomic_read(&sai->sai_cache_count) >= sai->sai_max;
 }
 
-static inline int sa_received_empty(struct ll_statahead_info *sai)
+static inline int sa_has_callback(struct ll_statahead_info *sai)
 {
-       return list_empty(&sai->sai_entries_received);
+       return !list_empty(&sai->sai_interim_entries);
 }
 
 static inline int agl_list_empty(struct ll_statahead_info *sai)
 {
-       return list_empty(&sai->sai_entries_agl);
+       return list_empty(&sai->sai_agls);
 }
 
 /**
@@ -188,56 +181,29 @@ static inline int is_omitted_entry(struct ll_statahead_info *sai, __u64 index)
                  sai->sai_index);
 }
 
-/*
- * Insert it into sai_entries tail when init.
- */
-static struct ll_sa_entry *
-ll_sa_entry_alloc(struct ll_statahead_info *sai, __u64 index, const char *name,
-                int len)
+/* allocate sa_entry and add it into hash to let scanner process to find it */
+static struct sa_entry *
+sa_alloc(struct ll_statahead_info *sai, __u64 index, const char *name, int len)
 {
-        struct ll_inode_info *lli;
-        struct ll_sa_entry   *entry;
-        int                   entry_size;
-        char                 *dname;
-        ENTRY;
+       struct ll_inode_info *lli;
+       struct sa_entry *entry;
+       int entry_size;
+       char *dname;
+       ENTRY;
 
-        entry_size = sizeof(struct ll_sa_entry) + (len & ~3) + 4;
-        OBD_ALLOC(entry, entry_size);
-        if (unlikely(entry == NULL))
-                RETURN(ERR_PTR(-ENOMEM));
+       entry_size = sizeof(struct sa_entry) + (len & ~3) + 4;
+       OBD_ALLOC(entry, entry_size);
+       if (unlikely(entry == NULL))
+               RETURN(ERR_PTR(-ENOMEM));
 
        CDEBUG(D_READA, "alloc sa entry %.*s(%p) index "LPU64"\n",
               len, name, entry, index);
 
-        entry->se_index = index;
+       entry->se_index = index;
 
-       /*
-        * Statahead entry reference rules:
-        *
-        * 1) When statahead entry is initialized, its reference is set as 2.
-        *    One reference is used by the directory scanner. When the scanner
-        *    searches the statahead cache for the given name, it can perform
-        *    lockless hash lookup (only the scanner can remove entry from hash
-        *    list), and once found, it needn't to call "atomic_inc()" for the
-        *    entry reference. So the performance is improved. After using the
-        *    statahead entry, the scanner will call "atomic_dec()" to drop the
-        *    reference held when initialization. If it is the last reference,
-        *    the statahead entry will be freed.
-        *
-        * 2) All other threads, including statahead thread and ptlrpcd thread,
-        *    when they process the statahead entry, the reference for target
-        *    should be held to guarantee the entry will not be released by the
-        *    directory scanner. After processing the entry, these threads will
-        *    drop the entry reference. If it is the last reference, the entry
-        *    will be freed.
-        *
-        *    The second reference when initializes the statahead entry is used
-        *    by the statahead thread, following the rule 2).
-        */
-       atomic_set(&entry->se_refcount, 2);
-       entry->se_stat = SA_ENTRY_INIT;
+       entry->se_state = SA_ENTRY_INIT;
        entry->se_size = entry_size;
-       dname = (char *)entry + sizeof(struct ll_sa_entry);
+       dname = (char *)entry + sizeof(struct sa_entry);
        memcpy(dname, name, len);
        dname[len] = 0;
        entry->se_qstr.hash = full_name_hash(name, len);
@@ -246,9 +212,8 @@ ll_sa_entry_alloc(struct ll_statahead_info *sai, __u64 index, const char *name,
 
        lli = ll_i2info(sai->sai_inode);
        spin_lock(&lli->lli_sa_lock);
-       list_add_tail(&entry->se_link, &sai->sai_entries);
        INIT_LIST_HEAD(&entry->se_list);
-       ll_sa_entry_enhash(sai, entry);
+       sa_rehash(sai, entry);
        spin_unlock(&lli->lli_sa_lock);
 
        atomic_inc(&sai->sai_cache_count);
@@ -256,152 +221,123 @@ ll_sa_entry_alloc(struct ll_statahead_info *sai, __u64 index, const char *name,
        RETURN(entry);
 }
 
-/*
- * Used by the directory scanner to search entry with name.
- *
- * Only the caller can remove the entry from hash, so it is unnecessary to hold
- * hash lock. It is caller's duty to release the init refcount on the entry, so
- * it is also unnecessary to increase refcount on the entry.
- */
-static struct ll_sa_entry *
-ll_sa_entry_get_byname(struct ll_statahead_info *sai, const struct qstr *qstr)
+/* free sa_entry which should have been unhashed and not in any list */
+static void sa_free(struct ll_statahead_info *sai, struct sa_entry *entry)
 {
-        struct ll_sa_entry *entry;
-        int i = ll_sa_entry_hash(qstr->hash);
+       CDEBUG(D_READA, "free sa entry %.*s(%p) index "LPU64"\n",
+              entry->se_qstr.len, entry->se_qstr.name, entry,
+              entry->se_index);
 
-       list_for_each_entry(entry, &sai->sai_cache[i], se_hash) {
-                if (entry->se_qstr.hash == qstr->hash &&
-                    entry->se_qstr.len == qstr->len &&
-                    memcmp(entry->se_qstr.name, qstr->name, qstr->len) == 0)
-                        return entry;
-        }
-        return NULL;
+       LASSERT(list_empty(&entry->se_list));
+       LASSERT(sa_unhashed(entry));
+
+       OBD_FREE(entry, entry->se_size);
+       atomic_dec(&sai->sai_cache_count);
 }
 
-/*
- * Used by the async getattr request callback to find entry with index.
- *
- * Inside lli_sa_lock to prevent others to change the list during the search.
- * It needs to increase entry refcount before returning to guarantee that the
- * entry cannot be freed by others.
+/* find sa_entry by name, used by directory scanner, lock is not needed because
+ * only scanner can remove the entry from hash.
  */
-static struct ll_sa_entry *
-ll_sa_entry_get_byindex(struct ll_statahead_info *sai, __u64 index)
+static struct sa_entry *
+sa_get(struct ll_statahead_info *sai, const struct qstr *qstr)
 {
-       struct ll_sa_entry *entry;
+       struct sa_entry *entry;
+       int i = sa_hash(qstr->hash);
 
-       list_for_each_entry(entry, &sai->sai_entries, se_link) {
-               if (entry->se_index == index) {
-                       LASSERT(atomic_read(&entry->se_refcount) > 0);
-                       atomic_inc(&entry->se_refcount);
+       list_for_each_entry(entry, &sai->sai_cache[i], se_hash) {
+               if (entry->se_qstr.hash == qstr->hash &&
+                   entry->se_qstr.len == qstr->len &&
+                   memcmp(entry->se_qstr.name, qstr->name, qstr->len) == 0)
                        return entry;
-                }
-               if (entry->se_index > index)
-                       break;
        }
        return NULL;
 }
 
-static void ll_sa_entry_put(struct ll_statahead_info *sai,
-                             struct ll_sa_entry *entry)
-{
-       if (atomic_dec_and_test(&entry->se_refcount)) {
-               CDEBUG(D_READA, "free sa entry %.*s(%p) index "LPU64"\n",
-                      entry->se_qstr.len, entry->se_qstr.name, entry,
-                      entry->se_index);
-
-               LASSERT(list_empty(&entry->se_link));
-               LASSERT(list_empty(&entry->se_list));
-               LASSERT(ll_sa_entry_unhashed(entry));
-
-               if (entry->se_inode)
-                       iput(entry->se_inode);
-
-               OBD_FREE(entry, entry->se_size);
-               atomic_dec(&sai->sai_cache_count);
-       }
-}
-
 static inline void
-do_sa_entry_fini(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
+sa_kill(struct ll_statahead_info *sai, struct sa_entry *entry)
 {
        struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 
-       LASSERT(!ll_sa_entry_unhashed(entry));
-       LASSERT(!list_empty(&entry->se_link));
+       LASSERT(!sa_unhashed(entry));
+       LASSERT(!list_empty(&entry->se_list));
+       LASSERT(sa_ready(entry));
 
-       ll_sa_entry_unhash(sai, entry);
+       sa_unhash(sai, entry);
 
        spin_lock(&lli->lli_sa_lock);
-       entry->se_stat = SA_ENTRY_DEST;
-       list_del_init(&entry->se_link);
-       if (likely(!list_empty(&entry->se_list)))
-               list_del_init(&entry->se_list);
+       list_del_init(&entry->se_list);
        spin_unlock(&lli->lli_sa_lock);
 
-       ll_sa_entry_put(sai, entry);
+       if (entry->se_inode != NULL)
+               iput(entry->se_inode);
+
+       sa_free(sai, entry);
 }
 
-/*
- * Delete it from sai_entries_stated list when fini.
- */
+/* called by scanner after use, sa_entry will be killed */
 static void
-ll_sa_entry_fini(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
+sa_put(struct ll_statahead_info *sai, struct sa_entry *entry)
 {
-       struct ll_sa_entry *pos, *next;
+       struct sa_entry *tmp, *next;
 
-       if (entry)
-               do_sa_entry_fini(sai, entry);
+       if (entry != NULL && entry->se_state == SA_ENTRY_SUCC) {
+               struct ll_sb_info *sbi = ll_i2sbi(sai->sai_inode);
 
-       /* drop old entry, only 'scanner' process does this, no need to lock */
-       list_for_each_entry_safe(pos, next, &sai->sai_entries, se_link) {
-               if (!is_omitted_entry(sai, pos->se_index))
+               sai->sai_hit++;
+               sai->sai_consecutive_miss = 0;
+               sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
+       } else {
+               sai->sai_miss++;
+               sai->sai_consecutive_miss++;
+       }
+
+       if (entry != NULL)
+               sa_kill(sai, entry);
+
+       /* kill old completed entries, only scanner process does this, no need
+        * to lock */
+       list_for_each_entry_safe(tmp, next, &sai->sai_entries, se_list) {
+               if (!is_omitted_entry(sai, tmp->se_index))
                        break;
-               /* keep those whose statahead RPC not finished */
-               if (pos->se_stat == SA_ENTRY_SUCC ||
-                   pos->se_stat == SA_ENTRY_INVA)
-                       do_sa_entry_fini(sai, pos);
+               sa_kill(sai, tmp);
        }
+
+       wake_up(&sai->sai_thread.t_ctl_waitq);
 }
 
-/*
- * Inside lli_sa_lock.
- */
-static void
-__sa_entry_post_stat(struct ll_statahead_info *sai, struct ll_sa_entry *entry,
-                    se_stat_t stat)
+/* update state and sort add entry to sai_entries by index, return true if
+ * scanner is waiting on this entry. */
+static bool
+__sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret)
 {
-       struct ll_sa_entry *se;
-       struct list_head *pos = &sai->sai_entries_stated;
+       struct sa_entry *se;
+       struct list_head *pos = &sai->sai_entries;
 
-       LASSERT(entry->se_stat == SA_ENTRY_INIT);
+       LASSERT(!sa_ready(entry));
+       LASSERT(list_empty(&entry->se_list));
 
-       if (!list_empty(&entry->se_list))
-               list_del_init(&entry->se_list);
+       entry->se_state = ret < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC;
 
-       list_for_each_entry_reverse(se, &sai->sai_entries_stated, se_list) {
+       list_for_each_entry_reverse(se, &sai->sai_entries, se_list) {
                if (se->se_index < entry->se_index) {
                        pos = &se->se_list;
                        break;
                }
        }
-
        list_add(&entry->se_list, pos);
-       entry->se_stat = stat;
+
+       return (entry->se_index == sai->sai_index_wait);
 }
 
-/*
- * Move entry to sai_entries_stated and sort with the index.
- * \retval 1    -- entry to be destroyed.
- * \retval 0    -- entry is inserted into stated list.
- */
+/* release resources used in async stat RPC, complete entry information and
+ * wakeup if necessary */
 static void
-sa_entry_post_stat(struct ll_statahead_info *sai, struct ll_sa_entry *entry,
-                  se_stat_t stat)
+sa_make_ready(struct ll_statahead_info *sai, struct sa_entry *entry, int ret)
 {
        struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
        struct md_enqueue_info *minfo = entry->se_minfo;
        struct ptlrpc_request *req = entry->se_req;
+       bool wakeup;
 
        /* release resources used in RPC */
        if (minfo) {
@@ -417,12 +353,15 @@ sa_entry_post_stat(struct ll_statahead_info *sai, struct ll_sa_entry *entry,
        }
 
        spin_lock(&lli->lli_sa_lock);
-       __sa_entry_post_stat(sai, entry, stat);
+       wakeup = __sa_make_ready(sai, entry, ret);
        spin_unlock(&lli->lli_sa_lock);
+
+       if (wakeup)
+               wake_up(&sai->sai_waitq);
 }
 
 /*
- * Insert inode into the list of sai_entries_agl.
+ * Insert inode into the list of sai_agls.
  */
 static void ll_agl_add(struct ll_statahead_info *sai,
                        struct inode *inode, int index)
@@ -442,7 +381,7 @@ static void ll_agl_add(struct ll_statahead_info *sai,
                spin_lock(&parent->lli_agl_lock);
                if (agl_list_empty(sai))
                        added = 1;
-               list_add_tail(&child->lli_agl_list, &sai->sai_entries_agl);
+               list_add_tail(&child->lli_agl_list, &sai->sai_agls);
                spin_unlock(&parent->lli_agl_lock);
        } else {
                spin_unlock(&child->lli_agl_lock);
@@ -476,10 +415,9 @@ static struct ll_statahead_info *ll_sai_alloc(void)
        init_waitqueue_head(&sai->sai_thread.t_ctl_waitq);
        init_waitqueue_head(&sai->sai_agl_thread.t_ctl_waitq);
 
+       INIT_LIST_HEAD(&sai->sai_interim_entries);
        INIT_LIST_HEAD(&sai->sai_entries);
-       INIT_LIST_HEAD(&sai->sai_entries_received);
-       INIT_LIST_HEAD(&sai->sai_entries_stated);
-       INIT_LIST_HEAD(&sai->sai_entries_agl);
+       INIT_LIST_HEAD(&sai->sai_agls);
 
        for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
                INIT_LIST_HEAD(&sai->sai_cache[i]);
@@ -509,7 +447,7 @@ static void ll_sai_put(struct ll_statahead_info *sai)
        struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
 
        if (atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
-               struct ll_sa_entry *entry, *next;
+               struct sa_entry *entry, *next;
                struct ll_sb_info *sbi = ll_i2sbi(sai->sai_inode);
 
                lli->lli_sai = NULL;
@@ -518,14 +456,14 @@ static void ll_sai_put(struct ll_statahead_info *sai)
                LASSERT(thread_is_stopped(&sai->sai_thread));
                LASSERT(thread_is_stopped(&sai->sai_agl_thread));
                LASSERT(sai->sai_sent == sai->sai_replied);
+               LASSERT(!sa_has_callback(sai));
 
                list_for_each_entry_safe(entry, next, &sai->sai_entries,
-                                        se_link)
-                       do_sa_entry_fini(sai, entry);
+                                        se_list)
+                       sa_kill(sai, entry);
 
                LASSERT(atomic_read(&sai->sai_cache_count) == 0);
                LASSERT(agl_list_empty(sai));
-               LASSERT(atomic_read(&sai->sai_refcount) == 0);
 
                iput(sai->sai_inode);
                OBD_FREE_PTR(sai);
@@ -533,13 +471,13 @@ static void ll_sai_put(struct ll_statahead_info *sai)
        }
 }
 
-/* Do NOT forget to drop inode refcount when into sai_entries_agl. */
+/* Do NOT forget to drop inode refcount when into sai_agls. */
 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;
+       struct ll_inode_info *lli = ll_i2info(inode);
+       __u64 index = lli->lli_agl_index;
+       int rc;
+       ENTRY;
 
        LASSERT(list_empty(&lli->lli_agl_list));
 
@@ -596,18 +534,19 @@ static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai)
         EXIT;
 }
 
-/* prepare inode for received statahead entry, and add it into agl list */
-static void sa_post_one(struct ll_statahead_info *sai,
-                       struct ll_sa_entry *entry)
+/* prepare inode for sa entry, add it into agl list, now sa_entry is ready
+ * to be used by scanner process. */
+static void sa_instantiate(struct ll_statahead_info *sai,
+                                struct sa_entry *entry)
 {
-        struct inode           *dir   = sai->sai_inode;
-        struct inode           *child;
-        struct md_enqueue_info *minfo;
-        struct lookup_intent   *it;
-        struct ptlrpc_request  *req;
-        struct mdt_body        *body;
-        int                     rc    = 0;
-        ENTRY;
+       struct inode *dir = sai->sai_inode;
+       struct inode *child;
+       struct md_enqueue_info *minfo;
+       struct lookup_intent *it;
+       struct ptlrpc_request *req;
+       struct mdt_body *body;
+       int rc = 0;
+       ENTRY;
 
         LASSERT(entry->se_handle != 0);
 
@@ -651,8 +590,9 @@ static void sa_post_one(struct ll_statahead_info *sai,
         if (rc)
                 GOTO(out, rc);
 
-       CDEBUG(D_DLMTRACE, "%s: setting l_data to inode "DFID"(%p)\n",
+       CDEBUG(D_READA, "%s: setting %.*s"DFID" l_data to inode %p\n",
               ll_get_fsname(child->i_sb, NULL, 0),
+              entry->se_qstr.len, entry->se_qstr.name,
               PFID(ll_inode2fid(child)), child);
         ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
 
@@ -664,36 +604,33 @@ static void sa_post_one(struct ll_statahead_info *sai,
         EXIT;
 
 out:
-       /* The "sa_entry_post_stat()" will drop related ldlm ibits lock
-        * reference count by calling "ll_intent_drop_lock()" in spite of the
-        * above operations failed or not. Do not worry about calling
-        * "ll_intent_drop_lock()" more than once. */
-       sa_entry_post_stat(sai, entry, rc < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
-       if (entry->se_index == sai->sai_index_wait)
-               wake_up(&sai->sai_waitq);
-       ll_sa_entry_put(sai, entry);
+       /* sa_make_ready() will drop ldlm ibits lock refcount by calling
+        * ll_intent_drop_lock() in spite of failures. Do not worry about
+        * calling ll_intent_drop_lock() more than once. */
+       sa_make_ready(sai, entry, rc);
 }
 
-static void ll_post_statahead(struct ll_statahead_info *sai)
+/* once there are async stat replies, instantiate sa_entry */
+static void sa_handle_callback(struct ll_statahead_info *sai)
 {
        struct ll_inode_info *lli;
 
        lli = ll_i2info(sai->sai_inode);
 
-       while (!sa_received_empty(sai)) {
-               struct ll_sa_entry *entry;
+       while (sa_has_callback(sai)) {
+               struct sa_entry *entry;
 
                spin_lock(&lli->lli_sa_lock);
-               if (unlikely(sa_received_empty(sai))) {
+               if (unlikely(!sa_has_callback(sai))) {
                        spin_unlock(&lli->lli_sa_lock);
                        break;
                }
-               entry = sa_first_received_entry(sai);
-               atomic_inc(&entry->se_refcount);
+               entry = list_entry(sai->sai_interim_entries.next,
+                                  struct sa_entry, se_list);
                list_del_init(&entry->se_list);
                spin_unlock(&lli->lli_sa_lock);
 
-               sa_post_one(sai, entry);
+               sa_instantiate(sai, entry);
        }
 
        spin_lock(&lli->lli_agl_lock);
@@ -711,30 +648,35 @@ static void ll_post_statahead(struct ll_statahead_info *sai)
        spin_unlock(&lli->lli_agl_lock);
 }
 
+/* callback for async stat, because this is called in ptlrpcd context, we only
+ * put sa_entry in sai_cb_entries list, and let sa_handle_callback() to really
+ * prepare inode and instantiate sa_entry later. */
 static int ll_statahead_interpret(struct ptlrpc_request *req,
                                  struct md_enqueue_info *minfo, int rc)
 {
        struct lookup_intent *it = &minfo->mi_it;
        struct inode *dir = minfo->mi_dir;
        struct ll_inode_info *lli = ll_i2info(dir);
-       struct ll_statahead_info *sai;
-       struct ll_sa_entry *entry;
-       int wakeup;
+       struct ll_statahead_info *sai = lli->lli_sai;
+       struct sa_entry *entry = (struct sa_entry *)minfo->mi_cbdata;
+       bool wakeup;
        ENTRY;
 
        if (it_disposition(it, DISP_LOOKUP_NEG))
                rc = -ENOENT;
 
-       sai = ll_sai_get(dir);
+       /* because statahead thread will wait for all inflight RPC to finish,
+        * sai should be always valid, no need to refcount */
        LASSERT(sai != NULL);
        LASSERT(!thread_is_stopped(&sai->sai_thread));
+       LASSERT(entry != NULL);
+
+       CDEBUG(D_READA, "sa_entry %.*s rc %d\n",
+              entry->se_qstr.len, entry->se_qstr.name, rc);
 
        spin_lock(&lli->lli_sa_lock);
-       entry = ll_sa_entry_get_byindex(sai, minfo->mi_cbdata);
-       LASSERT(entry != NULL);
        if (rc != 0) {
-               __sa_entry_post_stat(sai, entry, SA_ENTRY_INVA);
-               wakeup = (entry->se_index == sai->sai_index_wait);
+               wakeup = __sa_make_ready(sai, entry, rc);
        } else {
                entry->se_minfo = minfo;
                entry->se_req = ptlrpc_request_addref(req);
@@ -744,26 +686,23 @@ static int ll_statahead_interpret(struct ptlrpc_request *req,
                 * with parent's lock held, for example: unlink. */
                entry->se_handle = it->d.lustre.it_lock_handle;
                ll_intent_drop_lock(it);
-               wakeup = sa_received_empty(sai);
-               list_add_tail(&entry->se_list, &sai->sai_entries_received);
+               wakeup = !sa_has_callback(sai);
+               list_add_tail(&entry->se_list, &sai->sai_interim_entries);
        }
        sai->sai_replied++;
-       spin_unlock(&lli->lli_sa_lock);
-
-       ll_sa_entry_put(sai, entry);
        if (wakeup)
                wake_up(&sai->sai_thread.t_ctl_waitq);
+       spin_unlock(&lli->lli_sa_lock);
 
        if (rc != 0) {
                ll_intent_release(it);
                iput(dir);
                OBD_FREE_PTR(minfo);
        }
-       ll_sai_put(sai);
        RETURN(rc);
 }
 
-static void sa_args_fini(struct md_enqueue_info *minfo,
+static void sa_fini_data(struct md_enqueue_info *minfo,
                          struct ldlm_enqueue_info *einfo)
 {
         LASSERT(minfo && einfo);
@@ -783,10 +722,10 @@ static void sa_args_fini(struct md_enqueue_info *minfo,
  * "ocapa". So here reserve "op_data.op_capa[1,2]" in "pcapa" before calling
  * "md_intent_getattr_async".
  */
-static int sa_args_init(struct inode *dir, struct inode *child,
-                        struct ll_sa_entry *entry, struct md_enqueue_info **pmi,
-                        struct ldlm_enqueue_info **pei,
-                        struct obd_capa **pcapa)
+static int sa_prep_data(struct inode *dir, struct inode *child,
+                       struct sa_entry *entry, struct md_enqueue_info **pmi,
+                       struct ldlm_enqueue_info **pei,
+                       struct obd_capa **pcapa)
 {
         struct qstr              *qstr = &entry->se_qstr;
         struct md_enqueue_info   *minfo;
@@ -814,7 +753,7 @@ static int sa_args_init(struct inode *dir, struct inode *child,
        minfo->mi_it.it_op = IT_GETATTR;
        minfo->mi_dir = igrab(dir);
        minfo->mi_cb = ll_statahead_interpret;
-       minfo->mi_cbdata = entry->se_index;
+       minfo->mi_cbdata = entry;
 
         einfo->ei_type   = LDLM_IBITS;
         einfo->ei_mode   = it_to_lock_mode(&minfo->mi_it);
@@ -831,27 +770,27 @@ static int sa_args_init(struct inode *dir, struct inode *child,
         return 0;
 }
 
-static int do_sa_lookup(struct inode *dir, struct ll_sa_entry *entry)
+static int sa_lookup(struct inode *dir, struct sa_entry *entry)
 {
-        struct md_enqueue_info   *minfo;
-        struct ldlm_enqueue_info *einfo;
-        struct obd_capa          *capas[2];
-        int                       rc;
-        ENTRY;
+       struct md_enqueue_info   *minfo;
+       struct ldlm_enqueue_info *einfo;
+       struct obd_capa          *capas[2];
+       int                       rc;
+       ENTRY;
 
-        rc = sa_args_init(dir, NULL, entry, &minfo, &einfo, capas);
-        if (rc)
-                RETURN(rc);
-
-        rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
-        if (!rc) {
-                capa_put(capas[0]);
-                capa_put(capas[1]);
-        } else {
-                sa_args_fini(minfo, einfo);
-        }
+       rc = sa_prep_data(dir, NULL, entry, &minfo, &einfo, capas);
+       if (rc)
+               RETURN(rc);
 
-        RETURN(rc);
+       rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
+       if (!rc) {
+               capa_put(capas[0]);
+               capa_put(capas[1]);
+       } else {
+               sa_fini_data(minfo, einfo);
+       }
+
+       RETURN(rc);
 }
 
 /**
@@ -860,72 +799,72 @@ static int do_sa_lookup(struct inode *dir, struct ll_sa_entry *entry)
  * \retval      0 -- will send stat-ahead request
  * \retval others -- prepare stat-ahead request failed
  */
-static int do_sa_revalidate(struct inode *dir, struct ll_sa_entry *entry,
-                            struct dentry *dentry)
+static int sa_revalidate(struct inode *dir, struct sa_entry *entry,
+                        struct dentry *dentry)
 {
-        struct inode             *inode = dentry->d_inode;
-        struct lookup_intent      it = { .it_op = IT_GETATTR,
-                                         .d.lustre.it_lock_handle = 0 };
-        struct md_enqueue_info   *minfo;
-        struct ldlm_enqueue_info *einfo;
-        struct obd_capa          *capas[2];
-        int rc;
-        ENTRY;
-
-        if (unlikely(inode == NULL))
-                RETURN(1);
-
-        if (d_mountpoint(dentry))
-                RETURN(1);
-
-        entry->se_inode = igrab(inode);
-        rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),NULL);
-        if (rc == 1) {
-                entry->se_handle = it.d.lustre.it_lock_handle;
-                ll_intent_release(&it);
-                RETURN(1);
-        }
+       struct inode *inode = dentry->d_inode;
+       struct lookup_intent it = { .it_op = IT_GETATTR,
+                                   .d.lustre.it_lock_handle = 0 };
+       struct md_enqueue_info *minfo;
+       struct ldlm_enqueue_info *einfo;
+       struct obd_capa *capas[2];
+       int rc;
+       ENTRY;
 
-        rc = sa_args_init(dir, inode, entry, &minfo, &einfo, capas);
-        if (rc) {
-                entry->se_inode = NULL;
-                iput(inode);
-                RETURN(rc);
-        }
+       if (unlikely(inode == NULL))
+               RETURN(1);
 
-        rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
-        if (!rc) {
-                capa_put(capas[0]);
-                capa_put(capas[1]);
-        } else {
-                entry->se_inode = NULL;
-                iput(inode);
-                sa_args_fini(minfo, einfo);
-        }
+       if (d_mountpoint(dentry))
+               RETURN(1);
+
+       entry->se_inode = igrab(inode);
+       rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),
+                               NULL);
+       if (rc == 1) {
+               entry->se_handle = it.d.lustre.it_lock_handle;
+               ll_intent_release(&it);
+               RETURN(1);
+       }
+
+       rc = sa_prep_data(dir, inode, entry, &minfo, &einfo, capas);
+       if (rc) {
+               entry->se_inode = NULL;
+               iput(inode);
+               RETURN(rc);
+       }
 
-        RETURN(rc);
+       rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
+       if (!rc) {
+               capa_put(capas[0]);
+               capa_put(capas[1]);
+       } else {
+               entry->se_inode = NULL;
+               iput(inode);
+               sa_fini_data(minfo, einfo);
+       }
+
+       RETURN(rc);
 }
 
-static void ll_statahead_one(struct dentry *parent, const char *name,
-                            const int namelen)
+static void sa_statahead(struct dentry *parent, const char *name, int len)
 {
-       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;
+       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 sa_entry *entry;
+       int rc;
        ENTRY;
 
-       entry = ll_sa_entry_alloc(sai, sai->sai_index, name,namelen);
+       entry = sa_alloc(sai, sai->sai_index, name, len);
        if (IS_ERR(entry))
                RETURN_EXIT;
 
        dentry = d_lookup(parent, &entry->se_qstr);
        if (!dentry) {
-               rc = do_sa_lookup(dir, entry);
+               rc = sa_lookup(dir, entry);
        } else {
-               rc = do_sa_revalidate(dir, entry, dentry);
+               rc = 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);
        }
@@ -933,18 +872,12 @@ static void ll_statahead_one(struct dentry *parent, const char *name,
        if (dentry != NULL)
                dput(dentry);
 
-       if (rc) {
-               sa_entry_post_stat(sai, entry,
-                                  rc < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
-               if (entry->se_index == sai->sai_index_wait)
-                       wake_up(&sai->sai_waitq);
-       } else {
+       if (rc != 0)
+               sa_make_ready(sai, entry, rc);
+       else
                sai->sai_sent++;
-       }
 
        sai->sai_index++;
-       /* drop one refcount on entry by ll_sa_entry_alloc */
-       ll_sa_entry_put(sai, entry);
 
        EXIT;
 }
@@ -1162,16 +1095,16 @@ static int ll_statahead_thread(void *arg)
                        do {
                                l_wait_event(thread->t_ctl_waitq,
                                             !sa_sent_full(sai) ||
-                                            !sa_received_empty(sai) ||
+                                            sa_has_callback(sai) ||
                                             !agl_list_empty(sai) ||
                                             !thread_is_running(thread),
                                             &lwi);
 
-                               ll_post_statahead(sai);
+                               sa_handle_callback(sai);
                        } while (sa_sent_full(sai) &&
                                 thread_is_running(thread));
 
-                       ll_statahead_one(parent, name, namelen);
+                       sa_statahead(parent, name, namelen);
                }
 
                pos = le64_to_cpu(dp->ldp_hash_end);
@@ -1205,12 +1138,12 @@ static int ll_statahead_thread(void *arg)
         * for file release to stop me. */
        while (thread_is_running(thread)) {
                l_wait_event(thread->t_ctl_waitq,
-                            !sa_received_empty(sai) ||
+                            sa_has_callback(sai) ||
                             !agl_list_empty(sai) ||
                             !thread_is_running(thread),
                             &lwi);
 
-               ll_post_statahead(sai);
+               sa_handle_callback(sai);
        }
 
        EXIT;
@@ -1240,8 +1173,8 @@ out:
                        sai->sai_sent == sai->sai_replied, &lwi);
        }
 
-       /* release resources held by received entries. */
-       ll_post_statahead(sai);
+       /* release resources held by statahead RPCs */
+       sa_handle_callback(sai);
 
        spin_lock(&lli->lli_sa_lock);
        thread_set_flags(thread, SVC_STOPPED);
@@ -1296,7 +1229,7 @@ void ll_deauthorize_statahead(struct inode *dir, void *key)
        if (sai != NULL && thread_is_running(&sai->sai_thread)) {
                /*
                 * statahead thread may not quit yet because it needs to cache
-                * stated entries, now it's time to tell it to quit.
+                * entries, now it's time to tell it to quit.
                 */
                thread_set_flags(&sai->sai_thread, SVC_STOPPING);
                wake_up(&sai->sai_thread.t_ctl_waitq);
@@ -1439,29 +1372,12 @@ out:
         return rc;
 }
 
-static void
-ll_sai_unplug(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
-{
-       if (entry != NULL && entry->se_stat == SA_ENTRY_SUCC) {
-               struct ll_sb_info *sbi = ll_i2sbi(sai->sai_inode);
-
-               sai->sai_hit++;
-               sai->sai_consecutive_miss = 0;
-               sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
-       } else {
-               sai->sai_miss++;
-               sai->sai_consecutive_miss++;
-       }
-       ll_sa_entry_fini(sai, entry);
-       wake_up(&sai->sai_thread.t_ctl_waitq);
-}
-
 static int revalidate_statahead_dentry(struct inode *dir,
                                        struct ll_statahead_info *sai,
                                        struct dentry **dentryp,
                                        int only_unplug)
 {
-       struct ll_sa_entry *entry = NULL;
+       struct sa_entry *entry = NULL;
        struct l_wait_info lwi = { 0 };
        int rc = 0;
        ENTRY;
@@ -1493,31 +1409,31 @@ static int revalidate_statahead_dentry(struct inode *dir,
                }
        }
 
-       entry = ll_sa_entry_get_byname(sai, &(*dentryp)->d_name);
+       entry = sa_get(sai, &(*dentryp)->d_name);
        if (entry == NULL || only_unplug) {
-               ll_sai_unplug(sai, entry);
+               sa_put(sai, entry);
                RETURN(entry ? 1 : -EAGAIN);
        }
 
        /* if statahead is busy in readdir, help it do post-work */
-       if (!ll_sa_entry_stated(entry) && sai->sai_in_readpage)
-               ll_post_statahead(sai);
+       if (!sa_ready(entry) && sai->sai_in_readpage)
+               sa_handle_callback(sai);
 
-       if (!ll_sa_entry_stated(entry)) {
+       if (!sa_ready(entry)) {
                sai->sai_index_wait = entry->se_index;
                lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(30), NULL,
                                        LWI_ON_SIGNAL_NOOP, NULL);
                rc = l_wait_event(sai->sai_waitq,
-                               ll_sa_entry_stated(entry) ||
+                               sa_ready(entry) ||
                                thread_is_stopped(&sai->sai_thread),
                                &lwi);
                if (rc < 0) {
-                       ll_sai_unplug(sai, entry);
+                       sa_put(sai, entry);
                        RETURN(-EAGAIN);
                }
        }
 
-       if (entry->se_stat == SA_ENTRY_SUCC && entry->se_inode != NULL) {
+       if (entry->se_state == SA_ENTRY_SUCC && entry->se_inode != NULL) {
                struct inode *inode = entry->se_inode;
                struct lookup_intent it = { .it_op = IT_GETATTR,
                                            .d.lustre.it_lock_handle =
@@ -1532,7 +1448,7 @@ static int revalidate_statahead_dentry(struct inode *dir,
 
                                alias = ll_splice_alias(inode, *dentryp);
                                if (IS_ERR(alias)) {
-                                       ll_sai_unplug(sai, entry);
+                                       sa_put(sai, entry);
                                        RETURN(PTR_ERR(alias));
                                }
                                *dentryp = alias;
@@ -1548,7 +1464,7 @@ static int revalidate_statahead_dentry(struct inode *dir,
                                        (*dentryp)->d_name.name,
                                        PFID(ll_inode2fid((*dentryp)->d_inode)),
                                        PFID(ll_inode2fid(inode)));
-                               ll_sai_unplug(sai, entry);
+                               sa_put(sai, entry);
                                RETURN(-ESTALE);
                        } else {
                                iput(inode);
@@ -1562,7 +1478,7 @@ static int revalidate_statahead_dentry(struct inode *dir,
                }
        }
 
-       ll_sai_unplug(sai, entry);
+       sa_put(sai, entry);
        RETURN(rc);
 }