Whamcloud - gitweb
LU-3531 llite: move dir cache to MDC layer
[fs/lustre-release.git] / lustre / llite / statahead.c
index a2e48f4..01521ef 100644 (file)
@@ -458,20 +458,20 @@ static void ll_agl_add(struct ll_statahead_info *sai,
        }
 
        if (added > 0)
-               cfs_waitq_signal(&sai->sai_agl_thread.t_ctl_waitq);
+               wake_up(&sai->sai_agl_thread.t_ctl_waitq);
 }
 
 static struct ll_statahead_info *ll_sai_alloc(void)
 {
-        struct ll_statahead_info *sai;
-        int                       i;
-        ENTRY;
+       struct ll_statahead_info *sai;
+       int                       i;
+       ENTRY;
 
-        OBD_ALLOC_PTR(sai);
-        if (!sai)
-                RETURN(NULL);
+       OBD_ALLOC_PTR(sai);
+       if (!sai)
+               RETURN(NULL);
 
-        cfs_atomic_set(&sai->sai_refcount, 1);
+       cfs_atomic_set(&sai->sai_refcount, 1);
 
        spin_lock(&sai_generation_lock);
        sai->sai_generation = ++sai_generation;
@@ -479,24 +479,24 @@ static struct ll_statahead_info *ll_sai_alloc(void)
                sai->sai_generation = ++sai_generation;
        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);
+       sai->sai_max = LL_SA_RPC_MIN;
+       sai->sai_index = 1;
+       init_waitqueue_head(&sai->sai_waitq);
+       init_waitqueue_head(&sai->sai_thread.t_ctl_waitq);
+       init_waitqueue_head(&sai->sai_agl_thread.t_ctl_waitq);
 
        CFS_INIT_LIST_HEAD(&sai->sai_entries);
-        CFS_INIT_LIST_HEAD(&sai->sai_entries_received);
-        CFS_INIT_LIST_HEAD(&sai->sai_entries_stated);
-        CFS_INIT_LIST_HEAD(&sai->sai_entries_agl);
+       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]);
+       for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
+               CFS_INIT_LIST_HEAD(&sai->sai_cache[i]);
                spin_lock_init(&sai->sai_cache_lock[i]);
-        }
-        cfs_atomic_set(&sai->sai_cache_count, 0);
+       }
+       cfs_atomic_set(&sai->sai_cache_count, 0);
 
-        RETURN(sai);
+       RETURN(sai);
 }
 
 static inline struct ll_statahead_info *
@@ -681,8 +681,9 @@ static void ll_post_statahead(struct ll_statahead_info *sai)
         if (rc)
                 GOTO(out, rc);
 
-        CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
-               child, child->i_ino, child->i_generation);
+       CDEBUG(D_DLMTRACE, "%s: setting l_data to inode "DFID"(%p)\n",
+              ll_get_fsname(child->i_sb, NULL, 0),
+              PFID(ll_inode2fid(child)), child);
         ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
 
         entry->se_inode = child;
@@ -693,15 +694,15 @@ static void ll_post_statahead(struct ll_statahead_info *sai)
         EXIT;
 
 out:
-        /* The "ll_sa_entry_to_stated()" 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. */
+       /* The "ll_sa_entry_to_stated()" 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. */
        rc = ll_sa_entry_to_stated(sai, entry,
                                   rc < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
        if (rc == 0 && entry->se_index == sai->sai_index_wait)
-                cfs_waitq_signal(&sai->sai_waitq);
-        ll_sa_entry_put(sai, entry);
+               wake_up(&sai->sai_waitq);
+       ll_sa_entry_put(sai, entry);
 }
 
 static int ll_statahead_interpret(struct ptlrpc_request *req,
@@ -760,7 +761,7 @@ static int ll_statahead_interpret(struct ptlrpc_request *req,
 
                ll_sa_entry_put(sai, entry);
                if (wakeup)
-                       cfs_waitq_signal(&sai->sai_thread.t_ctl_waitq);
+                       wake_up(&sai->sai_thread.t_ctl_waitq);
         }
 
         EXIT;
@@ -893,9 +894,6 @@ static int do_sa_revalidate(struct inode *dir, struct ll_sa_entry *entry,
         if (d_mountpoint(dentry))
                 RETURN(1);
 
-        if (unlikely(dentry == dentry->d_sb->s_root))
-                RETURN(1);
-
         entry->se_inode = igrab(inode);
         rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),NULL);
         if (rc == 1) {
@@ -953,14 +951,14 @@ static void ll_statahead_one(struct dentry *parent, const char* entry_name,
         if (dentry != NULL)
                 dput(dentry);
 
-        if (rc) {
-                rc1 = ll_sa_entry_to_stated(sai, entry,
-                                        rc < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
-                if (rc1 == 0 && entry->se_index == sai->sai_index_wait)
-                        cfs_waitq_signal(&sai->sai_waitq);
-        } else {
-                sai->sai_sent++;
-        }
+       if (rc) {
+               rc1 = ll_sa_entry_to_stated(sai, entry,
+                                       rc < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
+               if (rc1 == 0 && entry->se_index == sai->sai_index_wait)
+                       wake_up(&sai->sai_waitq);
+       } else {
+               sai->sai_sent++;
+       }
 
         sai->sai_index++;
         /* drop one refcount on entry by ll_sa_entry_alloc */
@@ -981,21 +979,15 @@ static int ll_agl_thread(void *arg)
         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);
+       CDEBUG(D_READA, "agl thread started: [pid %d] [parent %.*s]\n",
+              current_pid(), parent->d_name.len, parent->d_name.name);
 
-        atomic_inc(&sbi->ll_agl_total);
+       atomic_inc(&sbi->ll_agl_total);
        spin_lock(&plli->lli_agl_lock);
        sai->sai_agl_valid = 1;
        thread_set_flags(thread, SVC_RUNNING);
        spin_unlock(&plli->lli_agl_lock);
-        cfs_waitq_signal(&thread->t_ctl_waitq);
+       wake_up(&thread->t_ctl_waitq);
 
         while (1) {
                 l_wait_event(thread->t_ctl_waitq,
@@ -1031,255 +1023,223 @@ static int ll_agl_thread(void *arg)
        }
        thread_set_flags(thread, SVC_STOPPED);
        spin_unlock(&plli->lli_agl_lock);
-       cfs_waitq_signal(&thread->t_ctl_waitq);
+       wake_up(&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);
+              current_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);
-                thread_set_flags(thread, SVC_STOPPED);
-                RETURN_EXIT;
-        }
+       struct ptlrpc_thread *thread = &sai->sai_agl_thread;
+       struct l_wait_info    lwi    = { 0 };
+       struct ll_inode_info  *plli;
+       struct task_struct            *task;
+       ENTRY;
+
+       CDEBUG(D_READA, "start agl thread: [pid %d] [parent %.*s]\n",
+              current_pid(), parent->d_name.len, parent->d_name.name);
+
+       plli = ll_i2info(parent->d_inode);
+       task = kthread_run(ll_agl_thread, parent,
+                              "ll_agl_%u", plli->lli_opendir_pid);
+       if (IS_ERR(task)) {
+               CERROR("can't start ll_agl thread, rc: %ld\n", PTR_ERR(task));
+               thread_set_flags(thread, SVC_STOPPED);
+               RETURN_EXIT;
+       }
 
-        l_wait_event(thread->t_ctl_waitq,
-                     thread_is_running(thread) || thread_is_stopped(thread),
-                     &lwi);
-        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     *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 ptlrpc_thread *agl_thread = &sai->sai_agl_thread;
-        struct page              *page;
-        __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", 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);
+       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_thread;
+       struct ptlrpc_thread *agl_thread = &sai->sai_agl_thread;
+       int                       first  = 0;
+       int                       rc     = 0;
+       struct md_op_data        *op_data;
+       struct ll_dir_chain       chain;
+       struct l_wait_info        lwi    = { 0 };
+       struct lu_dirent         *ent;
+       ENTRY;
+
+       CDEBUG(D_READA, "statahead thread started: [pid %d] [parent %.*s]\n",
+              current_pid(), parent->d_name.len, parent->d_name.name);
+
+       op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
+                                    LUSTRE_OPC_ANY, dir);
+       if (IS_ERR(op_data))
+               RETURN(PTR_ERR(op_data));
+
+       op_data->op_hash_offset = 0;
+       op_data->op_max_pages =
+               ll_i2sbi(dir)->ll_md_brw_size >> PAGE_CACHE_SHIFT;
+
+       if (sbi->ll_flags & LL_SBI_AGL_ENABLED)
+               ll_start_agl(parent, sai);
+
+       atomic_inc(&sbi->ll_sa_total);
        spin_lock(&plli->lli_sa_lock);
        thread_set_flags(thread, SVC_RUNNING);
        spin_unlock(&plli->lli_sa_lock);
-       cfs_waitq_signal(&thread->t_ctl_waitq);
+       wake_up(&thread->t_ctl_waitq);
 
        ll_dir_chain_init(&chain);
-       page = ll_get_dir_page(dir, pos, &chain);
-
-        while (1) {
-                struct lu_dirpage *dp;
-                struct lu_dirent  *ent;
-
-                if (IS_ERR(page)) {
-                        rc = PTR_ERR(page);
-                        CDEBUG(D_READA, "error reading dir "DFID" at "LPU64
-                               "/"LPU64": [rc %d] [parent %u]\n",
-                               PFID(ll_inode2fid(dir)), pos, sai->sai_index,
-                               rc, plli->lli_opendir_pid);
-                        GOTO(out, rc);
-                }
-
-                dp = page_address(page);
-                for (ent = lu_dirent_start(dp); ent != NULL;
-                     ent = lu_dirent_next(ent)) {
-                        __u64 hash;
-                        int namelen;
-                        char *name;
-
-                        hash = le64_to_cpu(ent->lde_hash);
-                        if (unlikely(hash < pos))
-                                /*
-                                 * Skip until we find target hash value.
-                                 */
-                                continue;
-
-                        namelen = le16_to_cpu(ent->lde_namelen);
-                        if (unlikely(namelen == 0))
-                                /*
-                                 * Skip dummy record.
-                                 */
-                                continue;
-
-                        name = ent->lde_name;
-                        if (name[0] == '.') {
-                                if (namelen == 1) {
-                                        /*
-                                         * skip "."
-                                         */
-                                        continue;
-                                } else if (name[1] == '.' && namelen == 2) {
-                                        /*
-                                         * skip ".."
-                                         */
-                                        continue;
-                                } else if (!sai->sai_ls_all) {
-                                        /*
-                                         * skip hidden files.
-                                         */
-                                        sai->sai_skip_hidden++;
-                                        continue;
-                                }
-                        }
+       for (ent = ll_dir_entry_start(dir, op_data);
+            ent != NULL && !IS_ERR(ent);
+            ent = ll_dir_entry_next(dir, op_data, ent)) {
+               __u64 hash;
+               int namelen;
+               char *name;
+
+               hash = le64_to_cpu(ent->lde_hash);
+               if (unlikely(hash < op_data->op_hash_offset))
+                       /*
+                        * Skip until we find target hash value.
+                        */
+                       continue;
+
+               namelen = le16_to_cpu(ent->lde_namelen);
+               if (unlikely(namelen == 0))
+                       /*
+                        * Skip dummy record.
+                        */
+                       continue;
+
+               name = ent->lde_name;
+               if (name[0] == '.') {
+                       if (namelen == 1) {
+                               /*
+                                * skip "."
+                                */
+                               continue;
+                       } else if (name[1] == '.' && namelen == 2) {
+                               /*
+                                * skip ".."
+                                */
+                               continue;
+                       } else if (!sai->sai_ls_all) {
+                               /*
+                                * skip hidden files.
+                                */
+                               sai->sai_skip_hidden++;
+                               continue;
+                       }
+               }
 
-                        /*
-                         * don't stat-ahead first entry.
-                         */
-                        if (unlikely(++first == 1))
-                                continue;
+               /*
+                * don't stat-ahead first entry.
+                */
+               if (unlikely(++first == 1))
+                       continue;
 
 keep_it:
-                        l_wait_event(thread->t_ctl_waitq,
-                                     !sa_sent_full(sai) ||
-                                     !sa_received_empty(sai) ||
-                                     !agl_list_empty(sai) ||
-                                     !thread_is_running(thread),
-                                     &lwi);
+               l_wait_event(thread->t_ctl_waitq,
+                            !sa_sent_full(sai) ||
+                            !sa_received_empty(sai) ||
+                            !agl_list_empty(sai) ||
+                            !thread_is_running(thread),
+                            &lwi);
 
 interpret_it:
-                        while (!sa_received_empty(sai))
-                               ll_post_statahead(sai);
-
-                        if (unlikely(!thread_is_running(thread))) {
-                                ll_release_page(page, 0);
-                                GOTO(out, rc = 0);
-                        }
-
-                        /* 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)) {
-                               spin_lock(&plli->lli_agl_lock);
-                               while (!agl_list_empty(sai)) {
-                                       clli = agl_first_entry(sai);
-                                       cfs_list_del_init(&clli->lli_agl_list);
-                                       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;
-
-                                       spin_lock(&plli->lli_agl_lock);
-                               }
-                               spin_unlock(&plli->lli_agl_lock);
-
-                                goto keep_it;
-                        }
+               while (!sa_received_empty(sai))
+                       ll_post_statahead(sai);
 
-do_it:
-                        ll_statahead_one(parent, name, namelen);
-                }
-                pos = le64_to_cpu(dp->ldp_hash_end);
-                if (pos == MDS_DIR_END_OFF) {
-                        /*
-                         * End of directory reached.
-                         */
-                        ll_release_page(page, 0);
-                        while (1) {
-                                l_wait_event(thread->t_ctl_waitq,
-                                             !sa_received_empty(sai) ||
-                                             sai->sai_sent == sai->sai_replied||
-                                             !thread_is_running(thread),
-                                             &lwi);
-
-                                while (!sa_received_empty(sai))
-                                       ll_post_statahead(sai);
-
-                                if (unlikely(!thread_is_running(thread)))
-                                        GOTO(out, rc = 0);
-
-                                if (sai->sai_sent == sai->sai_replied &&
-                                    sa_received_empty(sai))
-                                        break;
-                        }
+               if (unlikely(!thread_is_running(thread)))
+                       GOTO(out, rc = 0);
 
+               /* 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)) {
                        spin_lock(&plli->lli_agl_lock);
-                       while (!agl_list_empty(sai) &&
-                              thread_is_running(thread)) {
+                       while (!agl_list_empty(sai)) {
                                clli = agl_first_entry(sai);
                                cfs_list_del_init(&clli->lli_agl_list);
                                spin_unlock(&plli->lli_agl_lock);
-                               ll_agl_trigger(&clli->lli_vfs_inode, sai);
+                               ll_agl_trigger(&clli->lli_vfs_inode,
+                                              sai);
+
+                               if (!sa_received_empty(sai))
+                                       goto interpret_it;
+
+                               if (unlikely(
+                                       !thread_is_running(thread)))
+                                       GOTO(out, rc = 0);
+
+                               if (!sa_sent_full(sai))
+                                       goto do_it;
+
                                spin_lock(&plli->lli_agl_lock);
                        }
                        spin_unlock(&plli->lli_agl_lock);
 
-                        GOTO(out, rc = 0);
-                } else if (1) {
-                        /*
-                         * chain is exhausted.
-                         * Normal case: continue to the next page.
-                         */
-                        ll_release_page(page, le32_to_cpu(dp->ldp_flags) &
-                                              LDF_COLLIDE);
-                        sai->sai_in_readpage = 1;
-                       page = ll_get_dir_page(dir, pos, &chain);
-                        sai->sai_in_readpage = 0;
-                } else {
-                        LASSERT(le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
-                        ll_release_page(page, 1);
-                        /*
-                         * go into overflow page.
-                         */
-                }
-        }
-        EXIT;
+                       goto keep_it;
+               }
+
+do_it:
+               ll_statahead_one(parent, name, namelen);
+       }
+
+       if (ent != NULL && !IS_ERR(ent))
+               ll_dir_entry_end(dir, op_data, ent);
 
+        /*
+        * End of directory reached.
+        */
+       while (1) {
+               l_wait_event(thread->t_ctl_waitq,
+                            !sa_received_empty(sai) ||
+                            sai->sai_sent == sai->sai_replied ||
+                            !thread_is_running(thread),
+                            &lwi);
+
+               while (!sa_received_empty(sai))
+                       ll_post_statahead(sai);
+
+               if (unlikely(!thread_is_running(thread)))
+                       GOTO(out, rc = 0);
+
+               if (sai->sai_sent == sai->sai_replied &&
+                   sa_received_empty(sai))
+                       break;
+       }
+
+       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);
+               spin_unlock(&plli->lli_agl_lock);
+               ll_agl_trigger(&clli->lli_vfs_inode, sai);
+               spin_lock(&plli->lli_agl_lock);
+       }
+       spin_unlock(&plli->lli_agl_lock);
 out:
+       EXIT;
+       ll_finish_md_op_data(op_data);
         if (sai->sai_agl_valid) {
                spin_lock(&plli->lli_agl_lock);
                thread_set_flags(agl_thread, SVC_STOPPING);
                spin_unlock(&plli->lli_agl_lock);
-                cfs_waitq_signal(&agl_thread->t_ctl_waitq);
+               wake_up(&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);
+               CDEBUG(D_READA, "stop agl thread: [pid %d]\n",
+                      current_pid());
+               l_wait_event(agl_thread->t_ctl_waitq,
+                            thread_is_stopped(agl_thread),
+                            &lwi);
         } else {
                 /* Set agl_thread flags anyway. */
                 thread_set_flags(&sai->sai_agl_thread, SVC_STOPPED);
@@ -1298,13 +1258,13 @@ out:
        }
        thread_set_flags(thread, SVC_STOPPED);
        spin_unlock(&plli->lli_sa_lock);
-        cfs_waitq_signal(&sai->sai_waitq);
-        cfs_waitq_signal(&thread->t_ctl_waitq);
+       wake_up(&sai->sai_waitq);
+       wake_up(&thread->t_ctl_waitq);
         ll_sai_put(sai);
         dput(parent);
-        CDEBUG(D_READA, "statahead thread stopped: [pid %d] [parent %.*s]\n",
-               cfs_curproc_pid(), parent->d_name.len, parent->d_name.name);
-        return rc;
+       CDEBUG(D_READA, "statahead thread stopped: [pid %d] [parent %.*s]\n",
+              current_pid(), parent->d_name.len, parent->d_name.name);
+       return rc;
 }
 
 /**
@@ -1332,10 +1292,10 @@ void ll_stop_statahead(struct inode *dir, void *key)
                 if (!thread_is_stopped(thread)) {
                         thread_set_flags(thread, SVC_STOPPING);
                        spin_unlock(&lli->lli_sa_lock);
-                       cfs_waitq_signal(&thread->t_ctl_waitq);
+                       wake_up(&thread->t_ctl_waitq);
 
                        CDEBUG(D_READA, "stop statahead thread: [pid %d]\n",
-                              cfs_curproc_pid());
+                              current_pid());
                        l_wait_event(thread->t_ctl_waitq,
                                     thread_is_stopped(thread),
                                     &lwi);
@@ -1372,115 +1332,89 @@ enum {
 
 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
 {
-        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;
-
-        ll_dir_chain_init(&chain);
-       page = ll_get_dir_page(dir, pos, &chain);
-
-        while (1) {
-                struct lu_dirpage *dp;
-                struct lu_dirent  *ent;
-
-                if (IS_ERR(page)) {
-                        struct ll_inode_info *lli = ll_i2info(dir);
-
-                        rc = PTR_ERR(page);
-                        CERROR("error reading dir "DFID" at "LPU64": "
-                               "[rc %d] [parent %u]\n",
-                               PFID(ll_inode2fid(dir)), pos,
-                               rc, lli->lli_opendir_pid);
-                        break;
-                }
+       struct ll_dir_chain   chain;
+       struct qstr          *target = &dentry->d_name;
+       struct md_op_data    *op_data;
+       int                   dot_de;
+       struct lu_dirent     *ent;
+       int                   rc     = LS_NONE_FIRST_DE;
+       ENTRY;
 
-                dp = page_address(page);
-                for (ent = lu_dirent_start(dp); ent != NULL;
-                     ent = lu_dirent_next(ent)) {
-                        __u64 hash;
-                        int namelen;
-                        char *name;
-
-                        hash = le64_to_cpu(ent->lde_hash);
-                        /* The ll_get_dir_page() can return any page containing
-                         * the given hash which may be not the start hash. */
-                        if (unlikely(hash < pos))
-                                continue;
-
-                        namelen = le16_to_cpu(ent->lde_namelen);
-                        if (unlikely(namelen == 0))
-                                /*
-                                 * skip dummy record.
-                                 */
-                                continue;
+       ll_dir_chain_init(&chain);
 
-                        name = ent->lde_name;
-                        if (name[0] == '.') {
-                                if (namelen == 1)
-                                        /*
-                                         * skip "."
-                                         */
-                                        continue;
-                                else if (name[1] == '.' && namelen == 2)
-                                        /*
-                                         * skip ".."
-                                         */
-                                        continue;
-                                else
-                                        dot_de = 1;
-                        } else {
-                                dot_de = 0;
-                        }
+       op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
+                                    LUSTRE_OPC_ANY, dir);
+       if (IS_ERR(op_data))
+               GOTO(out, rc = PTR_ERR(op_data));
+       /**
+        *FIXME choose the start offset of the readdir
+        */
+       op_data->op_stripe_offset = 0;
+       op_data->op_hash_offset = 0;
+       op_data->op_max_pages =
+               ll_i2sbi(dir)->ll_md_brw_size >> PAGE_CACHE_SHIFT;
+
+       for (ent = ll_dir_entry_start(dir, op_data);
+            ent != NULL && !IS_ERR(ent);
+            ent = ll_dir_entry_next(dir, op_data, ent)) {
+               __u64 hash;
+               int namelen;
+               char *name;
+
+               hash = le64_to_cpu(ent->lde_hash);
+               /* The ll_get_dir_page() can return any page containing
+                * the given hash which may be not the start hash. */
+               if (unlikely(hash < op_data->op_hash_offset))
+                       continue;
+
+               namelen = le16_to_cpu(ent->lde_namelen);
+               if (unlikely(namelen == 0))
+                       /*
+                        * skip dummy record.
+                        */
+                       continue;
+
+               name = ent->lde_name;
+               if (name[0] == '.') {
+                       if (namelen == 1)
+                               /*
+                                * skip "."
+                                */
+                               continue;
+                       else if (name[1] == '.' && namelen == 2)
+                               /*
+                                * skip ".."
+                                */
+                               continue;
+                       else
+                               dot_de = 1;
+               } else {
+                       dot_de = 0;
+               }
 
-                        if (dot_de && target->name[0] != '.') {
-                                CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
-                                       target->len, target->name,
-                                       namelen, name);
-                                continue;
-                        }
+               if (dot_de && target->name[0] != '.') {
+                       CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
+                              target->len, target->name,
+                              namelen, name);
+                       continue;
+               }
 
-                        if (target->len != namelen ||
-                            memcmp(target->name, name, namelen) != 0)
-                                rc = LS_NONE_FIRST_DE;
-                        else if (!dot_de)
-                                rc = LS_FIRST_DE;
-                        else
-                                rc = LS_FIRST_DOT_DE;
+               if (target->len != namelen ||
+                   memcmp(target->name, name, namelen) != 0)
+                       rc = LS_NONE_FIRST_DE;
+               else if (!dot_de)
+                       rc = LS_FIRST_DE;
+               else
+                       rc = LS_FIRST_DOT_DE;
 
-                        ll_release_page(page, 0);
-                        GOTO(out, rc);
-                }
-                pos = le64_to_cpu(dp->ldp_hash_end);
-                if (pos == MDS_DIR_END_OFF) {
-                        /*
-                         * End of directory reached.
-                         */
-                        ll_release_page(page, 0);
-                        break;
-                } else if (1) {
-                        /*
-                         * chain is exhausted
-                         * Normal case: continue to the next page.
-                         */
-                        ll_release_page(page, le32_to_cpu(dp->ldp_flags) &
-                                              LDF_COLLIDE);
-                       page = ll_get_dir_page(dir, pos, &chain);
-                } else {
-                        /*
-                         * go into overflow page.
-                         */
-                        LASSERT(le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
-                        ll_release_page(page, 1);
-                }
-        }
+               break;
+       }
         EXIT;
-
+       if (ent != NULL && !IS_ERR(ent))
+               ll_dir_entry_end(dir, op_data, ent);
+       ll_finish_md_op_data(op_data);
 out:
-        ll_dir_chain_fini(&chain);
+       ll_dir_chain_fini(&chain);
         return rc;
 }
 
@@ -1509,13 +1443,13 @@ ll_sai_unplug(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
                 sai->sai_consecutive_miss++;
                 if (sa_low_hit(sai) && thread_is_running(thread)) {
                         atomic_inc(&sbi->ll_sa_wrong);
-                        CDEBUG(D_READA, "Statahead for dir "DFID" hit "
-                               "ratio too low: hit/miss "LPU64"/"LPU64
-                               ", sent/replied "LPU64"/"LPU64", stopping "
-                               "statahead thread: pid %d\n",
-                               PFID(&lli->lli_fid), sai->sai_hit,
-                               sai->sai_miss, sai->sai_sent,
-                               sai->sai_replied, cfs_curproc_pid());
+                       CDEBUG(D_READA, "Statahead for dir "DFID" hit "
+                              "ratio too low: hit/miss "LPU64"/"LPU64
+                              ", sent/replied "LPU64"/"LPU64", stopping "
+                              "statahead thread: pid %d\n",
+                              PFID(&lli->lli_fid), sai->sai_hit,
+                              sai->sai_miss, sai->sai_sent,
+                              sai->sai_replied, current_pid());
                        spin_lock(&lli->lli_sa_lock);
                        if (!thread_is_stopped(thread))
                                thread_set_flags(thread, SVC_STOPPING);
@@ -1524,7 +1458,7 @@ ll_sai_unplug(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
        }
 
        if (!thread_is_stopped(thread))
-               cfs_waitq_signal(&thread->t_ctl_waitq);
+               wake_up(&thread->t_ctl_waitq);
 
        EXIT;
 }
@@ -1541,16 +1475,17 @@ ll_sai_unplug(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
 int do_statahead_enter(struct inode *dir, struct dentry **dentryp,
                        int only_unplug)
 {
-        struct ll_inode_info     *lli   = ll_i2info(dir);
-        struct ll_statahead_info *sai   = lli->lli_sai;
-        struct dentry            *parent;
-        struct ll_sa_entry       *entry;
-        struct ptlrpc_thread     *thread;
-        struct l_wait_info        lwi   = { 0 };
-        int                       rc    = 0;
-        ENTRY;
-
-        LASSERT(lli->lli_opendir_pid == cfs_curproc_pid());
+       struct ll_inode_info            *lli = ll_i2info(dir);
+       struct ll_statahead_info        *sai = lli->lli_sai;
+       struct dentry                   *parent;
+       struct ll_sa_entry              *entry;
+       struct ptlrpc_thread            *thread;
+       struct l_wait_info               lwi = { 0 };
+       int                              rc = 0;
+       struct ll_inode_info            *plli;
+       ENTRY;
+
+       LASSERT(lli->lli_opendir_pid == current_pid());
 
         if (sai) {
                 thread = &sai->sai_thread;
@@ -1626,19 +1561,26 @@ int do_statahead_enter(struct inode *dir, struct dentry **dentryp,
                                                ll_inode2fid(inode), &bits);
                        if (rc == 1) {
                                if ((*dentryp)->d_inode == NULL) {
-                                       *dentryp = ll_splice_alias(inode,
-                                                                  *dentryp);
+                                       struct dentry *alias;
+
+                                       alias = ll_splice_alias(inode,
+                                                               *dentryp);
+                                       if (IS_ERR(alias)) {
+                                               ll_sai_unplug(sai, entry);
+                                               RETURN(PTR_ERR(alias));
+                                       }
+                                       *dentryp = alias;
                                 } else if ((*dentryp)->d_inode != inode) {
                                         /* revalidate, but inode is recreated */
                                         CDEBUG(D_READA,
-                                              "stale dentry %.*s inode %lu/%u, "
-                                              "statahead inode %lu/%u\n",
-                                              (*dentryp)->d_name.len,
-                                              (*dentryp)->d_name.name,
-                                              (*dentryp)->d_inode->i_ino,
-                                              (*dentryp)->d_inode->i_generation,
-                                              inode->i_ino,
-                                              inode->i_generation);
+                                              "%s: stale dentry %.*s inode "
+                                              DFID", statahead inode "DFID
+                                              "\n",
+                                              ll_get_fsname((*dentryp)->d_inode->i_sb, NULL, 0),
+                                              (*dentryp)->d_name.len,
+                                              (*dentryp)->d_name.name,
+                                              PFID(ll_inode2fid((*dentryp)->d_inode)),
+                                              PFID(ll_inode2fid(inode)));
                                         ll_sai_unplug(sai, entry);
                                         RETURN(-ESTALE);
                                 } else {
@@ -1689,15 +1631,18 @@ 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);
+       CDEBUG(D_READA, "start statahead thread: [pid %d] [parent %.*s]\n",
+              current_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;
-        if (rc < 0) {
-                CERROR("can't start ll_sa thread, rc: %d\n", rc);
-                dput(parent);
+
+       plli = ll_i2info(parent->d_inode);
+       rc = PTR_ERR(kthread_run(ll_statahead_thread, parent,
+                                "ll_sa_%u", plli->lli_opendir_pid));
+       thread = &sai->sai_thread;
+       if (IS_ERR_VALUE(rc)) {
+               CERROR("can't start ll_sa thread, rc: %d\n", rc);
+               dput(parent);
                 lli->lli_opendir_key = NULL;
                 thread_set_flags(thread, SVC_STOPPED);
                 thread_set_flags(&sai->sai_agl_thread, SVC_STOPPED);