Whamcloud - gitweb
LU-243 async lov_sync() operation
[fs/lustre-release.git] / lustre / lov / lov_obd.c
index 739fee3..61fdb7b 100644 (file)
@@ -26,7 +26,7 @@
  * GPL HEADER END
  */
 /*
- * Copyright  2008 Sun Microsystems, Inc. All rights reserved
+ * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  */
 /*
 #include <obd_ost.h>
 #include <lprocfs_status.h>
 #include <lustre_param.h>
-#include <lustre_cache.h>
+#include <cl_object.h>
 #include <lustre/ll_fiemap.h>
 
 #include "lov_internal.h"
 
-
 /* Keep a refcount of lov->tgt usage to prevent racing with addition/deletion.
    Any function that expects lov_tgts to remain stationary must take a ref. */
-void lov_getref(struct obd_device *obd)
+static void lov_getref(struct obd_device *obd)
 {
         struct lov_obd *lov = &obd->u.lov;
 
         /* nobody gets through here until lov_putref is done */
-        mutex_down(&lov->lov_lock);
-        atomic_inc(&lov->lov_refcount);
-        mutex_up(&lov->lov_lock);
+        cfs_mutex_down(&lov->lov_lock);
+        cfs_atomic_inc(&lov->lov_refcount);
+        cfs_mutex_up(&lov->lov_lock);
         return;
 }
 
-static void __lov_del_obd(struct obd_device *obd, __u32 index);
+static void __lov_del_obd(struct obd_device *obd, struct lov_tgt_desc *tgt);
 
-void lov_putref(struct obd_device *obd)
+static void lov_putref(struct obd_device *obd)
 {
         struct lov_obd *lov = &obd->u.lov;
-        mutex_down(&lov->lov_lock);
+
+        cfs_mutex_down(&lov->lov_lock);
         /* ok to dec to 0 more than once -- ltd_exp's will be null */
-        if (atomic_dec_and_test(&lov->lov_refcount) && lov->lov_death_row) {
+        if (cfs_atomic_dec_and_test(&lov->lov_refcount) && lov->lov_death_row) {
+                CFS_LIST_HEAD(kill);
                 int i;
+                struct lov_tgt_desc *tgt, *n;
                 CDEBUG(D_CONFIG, "destroying %d lov targets\n",
                        lov->lov_death_row);
                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
-                        if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_reap)
+                        tgt = lov->lov_tgts[i];
+
+                        if (!tgt || !tgt->ltd_reap)
                                 continue;
-                        /* Disconnect and delete from list */
-                        __lov_del_obd(obd, i);
+                        cfs_list_add(&tgt->ltd_kill, &kill);
+                        /* XXX - right now there is a dependency on ld_tgt_count
+                         * being the maximum tgt index for computing the
+                         * mds_max_easize. So we can't shrink it. */
+                        lov_ost_pool_remove(&lov->lov_packed, i);
+                        lov->lov_tgts[i] = NULL;
                         lov->lov_death_row--;
                 }
-        }
-        mutex_up(&lov->lov_lock);
-}
-
-static int lov_register_page_removal_cb(struct obd_export *exp,
-                                        obd_page_removal_cb_t func,
-                                        obd_pin_extent_cb pin_cb)
-{
-        struct lov_obd *lov = &exp->exp_obd->u.lov;
-        int i, rc = 0;
-
-        if (lov->lov_page_removal_cb && lov->lov_page_removal_cb != func)
-                return -EBUSY;
-
-        if (lov->lov_page_pin_cb && lov->lov_page_pin_cb != pin_cb)
-                return -EBUSY;
-
-        for (i = 0; i < lov->desc.ld_tgt_count; i++) {
-                if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_exp)
-                        continue;
-                rc |= obd_register_page_removal_cb(lov->lov_tgts[i]->ltd_exp,
-                                                   func, pin_cb);
-        }
-
-        lov->lov_page_removal_cb = func;
-        lov->lov_page_pin_cb = pin_cb;
-
-        return rc;
-}
-
-static int lov_unregister_page_removal_cb(struct obd_export *exp,
-                                        obd_page_removal_cb_t func)
-{
-        struct lov_obd *lov = &exp->exp_obd->u.lov;
-        int i, rc = 0;
-
-        if (lov->lov_page_removal_cb && lov->lov_page_removal_cb != func)
-                return -EINVAL;
-
-        lov->lov_page_removal_cb = NULL;
-        lov->lov_page_pin_cb = NULL;
-
-        for (i = 0; i < lov->desc.ld_tgt_count; i++) {
-                if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_exp)
-                        continue;
-                rc |= obd_unregister_page_removal_cb(lov->lov_tgts[i]->ltd_exp,
-                                                     func);
-        }
-
-        return rc;
-}
-
-static int lov_register_lock_cancel_cb(struct obd_export *exp,
-                                         obd_lock_cancel_cb func)
-{
-        struct lov_obd *lov = &exp->exp_obd->u.lov;
-        int i, rc = 0;
-
-        if (lov->lov_lock_cancel_cb && lov->lov_lock_cancel_cb != func)
-                return -EBUSY;
+                cfs_mutex_up(&lov->lov_lock);
 
-        for (i = 0; i < lov->desc.ld_tgt_count; i++) {
-                if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_exp)
-                        continue;
-                rc |= obd_register_lock_cancel_cb(lov->lov_tgts[i]->ltd_exp,
-                                                  func);
+                cfs_list_for_each_entry_safe(tgt, n, &kill, ltd_kill) {
+                        cfs_list_del(&tgt->ltd_kill);
+                        /* Disconnect */
+                        __lov_del_obd(obd, tgt);
+                }
+        } else {
+                cfs_mutex_up(&lov->lov_lock);
         }
-
-        lov->lov_lock_cancel_cb = func;
-
-        return rc;
 }
 
-static int lov_unregister_lock_cancel_cb(struct obd_export *exp,
-                                         obd_lock_cancel_cb func)
-{
-        struct lov_obd *lov = &exp->exp_obd->u.lov;
-        int i, rc = 0;
-
-        if (lov->lov_lock_cancel_cb && lov->lov_lock_cancel_cb != func)
-                return -EINVAL;
+static int lov_set_osc_active(struct obd_device *obd, struct obd_uuid *uuid,
+                              enum obd_notify_event ev);
+static int lov_notify(struct obd_device *obd, struct obd_device *watched,
+                      enum obd_notify_event ev, void *data);
 
-        for (i = 0; i < lov->desc.ld_tgt_count; i++) {
-                if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_exp)
-                        continue;
-                rc |= obd_unregister_lock_cancel_cb(lov->lov_tgts[i]->ltd_exp,
-                                                    func);
-        }
-        lov->lov_lock_cancel_cb = NULL;
-        return rc;
-}
 
 #define MAX_STRING_SIZE 128
-static int lov_connect_obd(struct obd_device *obd, __u32 index, int activate,
-                           struct obd_connect_data *data)
+int lov_connect_obd(struct obd_device *obd, __u32 index, int activate,
+                    struct obd_connect_data *data)
 {
         struct lov_obd *lov = &obd->u.lov;
-        struct obd_uuid tgt_uuid;
+        struct obd_uuid *tgt_uuid;
         struct obd_device *tgt_obd;
-        struct obd_uuid lov_osc_uuid = { "LOV_OSC_UUID" };
-        struct lustre_handle conn = {0, };
+        static struct obd_uuid lov_osc_uuid = { "LOV_OSC_UUID" };
         struct obd_import *imp;
-
 #ifdef __KERNEL__
         cfs_proc_dir_entry_t *lov_proc_dir;
 #endif
@@ -212,20 +144,17 @@ static int lov_connect_obd(struct obd_device *obd, __u32 index, int activate,
         if (!lov->lov_tgts[index])
                 RETURN(-EINVAL);
 
-        tgt_uuid = lov->lov_tgts[index]->ltd_uuid;
-
-        tgt_obd = class_find_client_obd(&tgt_uuid, LUSTRE_OSC_NAME,
-                                        &obd->obd_uuid);
+        tgt_uuid = &lov->lov_tgts[index]->ltd_uuid;
+        tgt_obd = lov->lov_tgts[index]->ltd_obd;
 
-        if (!tgt_obd) {
-                CERROR("Target %s not attached\n", obd_uuid2str(&tgt_uuid));
-                RETURN(-EINVAL);
-        }
         if (!tgt_obd->obd_set_up) {
-                CERROR("Target %s not set up\n", obd_uuid2str(&tgt_uuid));
+                CERROR("Target %s not set up\n", obd_uuid2str(tgt_uuid));
                 RETURN(-EINVAL);
         }
 
+        /* override the sp_me from lov */
+        tgt_obd->u.cli.cl_sp_me = lov->lov_sp_me;
+
         if (data && (data->ocd_connect_flags & OBD_CONNECT_INDEX))
                 data->ocd_index = index;
 
@@ -236,96 +165,58 @@ static int lov_connect_obd(struct obd_device *obd, __u32 index, int activate,
 
         if (activate) {
                 tgt_obd->obd_no_recov = 0;
-                /* FIXME this is probably supposed to be 
+                /* FIXME this is probably supposed to be
                    ptlrpc_set_import_active.  Horrible naming. */
                 ptlrpc_activate_import(imp);
         }
 
-        if (imp->imp_invalid) {
-                CERROR("not connecting OSC %s; administratively "
-                       "disabled\n", obd_uuid2str(&tgt_uuid));
-                rc = obd_register_observer(tgt_obd, obd);
-                if (rc) {
-                        CERROR("Target %s register_observer error %d; "
-                               "will not be able to reactivate\n",
-                               obd_uuid2str(&tgt_uuid), rc);
-                }
-                RETURN(0);
-        }
-
-        rc = obd_connect(NULL, &conn, tgt_obd, &lov_osc_uuid, data, NULL);
+        rc = obd_register_observer(tgt_obd, obd);
         if (rc) {
-                CERROR("Target %s connect error %d\n",
-                       obd_uuid2str(&tgt_uuid), rc);
+                CERROR("Target %s register_observer error %d\n",
+                       obd_uuid2str(tgt_uuid), rc);
                 RETURN(rc);
         }
-        lov->lov_tgts[index]->ltd_exp = class_conn2export(&conn);
-        if (!lov->lov_tgts[index]->ltd_exp) {
-                CERROR("Target %s: null export!\n", obd_uuid2str(&tgt_uuid));
-                RETURN(-ENODEV);
-        }
 
-        rc = obd_register_page_removal_cb(lov->lov_tgts[index]->ltd_exp,
-                                          lov->lov_page_removal_cb,
-                                          lov->lov_page_pin_cb);
-        if (rc) {
-                obd_disconnect(lov->lov_tgts[index]->ltd_exp);
-                lov->lov_tgts[index]->ltd_exp = NULL;
-                RETURN(rc);
-        }
 
-        rc = obd_register_lock_cancel_cb(lov->lov_tgts[index]->ltd_exp,
-                                         lov->lov_lock_cancel_cb);
-        if (rc) {
-                obd_unregister_page_removal_cb(lov->lov_tgts[index]->ltd_exp,
-                                               lov->lov_page_removal_cb);
-                obd_disconnect(lov->lov_tgts[index]->ltd_exp);
-                lov->lov_tgts[index]->ltd_exp = NULL;
-                RETURN(rc);
+        if (imp->imp_invalid) {
+                CERROR("not connecting OSC %s; administratively "
+                       "disabled\n", obd_uuid2str(tgt_uuid));
+                RETURN(0);
         }
 
-        rc = obd_register_observer(tgt_obd, obd);
-        if (rc) {
-                CERROR("Target %s register_observer error %d\n",
-                       obd_uuid2str(&tgt_uuid), rc);
-                obd_unregister_lock_cancel_cb(lov->lov_tgts[index]->ltd_exp,
-                                              lov->lov_lock_cancel_cb);
-                obd_unregister_page_removal_cb(lov->lov_tgts[index]->ltd_exp,
-                                               lov->lov_page_removal_cb);
-                obd_disconnect(lov->lov_tgts[index]->ltd_exp);
-                lov->lov_tgts[index]->ltd_exp = NULL;
-                RETURN(rc);
+        rc = obd_connect(NULL, &lov->lov_tgts[index]->ltd_exp, tgt_obd,
+                         &lov_osc_uuid, data, NULL);
+        if (rc || !lov->lov_tgts[index]->ltd_exp) {
+                CERROR("Target %s connect error %d\n",
+                       obd_uuid2str(tgt_uuid), rc);
+                RETURN(-ENODEV);
         }
 
         lov->lov_tgts[index]->ltd_reap = 0;
-        if (activate) {
-                lov->lov_tgts[index]->ltd_active = 1;
-                lov->desc.ld_active_tgt_count++;
-                lov->lov_tgts[index]->ltd_exp->exp_obd->obd_inactive = 0;
-        }
+
         CDEBUG(D_CONFIG, "Connected tgt idx %d %s (%s) %sactive\n", index,
-               obd_uuid2str(&tgt_uuid), tgt_obd->obd_name, activate ? "":"in");
+               obd_uuid2str(tgt_uuid), tgt_obd->obd_name, activate ? "":"in");
 
 #ifdef __KERNEL__
         lov_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
         if (lov_proc_dir) {
-                struct obd_device *osc_obd = class_conn2obd(&conn);
+                struct obd_device *osc_obd = lov->lov_tgts[index]->ltd_exp->exp_obd;
                 cfs_proc_dir_entry_t *osc_symlink;
-                char name[MAX_STRING_SIZE];
 
                 LASSERT(osc_obd != NULL);
                 LASSERT(osc_obd->obd_magic == OBD_DEVICE_MAGIC);
                 LASSERT(osc_obd->obd_type->typ_name != NULL);
-                snprintf(name, MAX_STRING_SIZE, "../../../%s/%s",
-                         osc_obd->obd_type->typ_name,
-                         osc_obd->obd_name);
-                osc_symlink = lprocfs_add_symlink(osc_obd->obd_name, lov_proc_dir,
-                                                  name);
+
+                osc_symlink = lprocfs_add_symlink(osc_obd->obd_name,
+                                                  lov_proc_dir,
+                                                  "../../../%s/%s",
+                                                  osc_obd->obd_type->typ_name,
+                                                  osc_obd->obd_name);
                 if (osc_symlink == NULL) {
                         CERROR("could not register LOV target "
-                               "/proc/fs/lustre/%s/%s/target_obds/%s.",
-                               obd->obd_type->typ_name, obd->obd_name,
-                               osc_obd->obd_name);
+                                "/proc/fs/lustre/%s/%s/target_obds/%s.",
+                                obd->obd_type->typ_name, obd->obd_name,
+                                osc_obd->obd_name);
                         lprocfs_remove(&lov_proc_dir);
                 }
         }
@@ -339,21 +230,24 @@ static int lov_connect_obd(struct obd_device *obd, __u32 index, int activate,
 }
 
 static int lov_connect(const struct lu_env *env,
-                       struct lustre_handle *conn, struct obd_device *obd,
+                       struct obd_export **exp, struct obd_device *obd,
                        struct obd_uuid *cluuid, struct obd_connect_data *data,
                        void *localdata)
 {
         struct lov_obd *lov = &obd->u.lov;
         struct lov_tgt_desc *tgt;
+        struct lustre_handle conn;
         int i, rc;
         ENTRY;
 
         CDEBUG(D_CONFIG, "connect #%d\n", lov->lov_connects);
 
-        rc = class_connect(conn, obd, cluuid);
+        rc = class_connect(&conn, obd, cluuid);
         if (rc)
                 RETURN(rc);
 
+        *exp = class_conn2export(&conn);
+
         /* Why should there ever be more than 1 connect? */
         lov->lov_connects++;
         LASSERT(lov->lov_connects == 1);
@@ -362,7 +256,7 @@ static int lov_connect(const struct lu_env *env,
         if (data)
                 lov->lov_ocd = *data;
 
-        lov_getref(obd);
+        obd_getref(obd);
         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
                 tgt = lov->lov_tgts[i];
                 if (!tgt || obd_uuid_empty(&tgt->ltd_uuid))
@@ -374,37 +268,38 @@ static int lov_connect(const struct lu_env *env,
                                obd->obd_name, i, rc);
                         continue;
                 }
+                /* connect to administrative disabled ost */
+                if (!lov->lov_tgts[i]->ltd_exp)
+                        continue;
+
+                rc = lov_notify(obd, lov->lov_tgts[i]->ltd_exp->exp_obd,
+                                OBD_NOTIFY_CONNECT, (void *)&i);
+                if (rc) {
+                        CERROR("%s error sending notify %d\n",
+                               obd->obd_name, rc);
+                }
         }
-        lov_putref(obd);
+        obd_putref(obd);
 
         RETURN(0);
 }
 
-static int lov_disconnect_obd(struct obd_device *obd, __u32 index)
+static int lov_disconnect_obd(struct obd_device *obd, struct lov_tgt_desc *tgt)
 {
         cfs_proc_dir_entry_t *lov_proc_dir;
         struct lov_obd *lov = &obd->u.lov;
         struct obd_device *osc_obd;
         int rc;
-
         ENTRY;
 
-        if (lov->lov_tgts[index] == NULL)
-                RETURN(-EINVAL);
-
-        osc_obd = class_exp2obd(lov->lov_tgts[index]->ltd_exp);
+        osc_obd = class_exp2obd(tgt->ltd_exp);
         CDEBUG(D_CONFIG, "%s: disconnecting target %s\n",
                obd->obd_name, osc_obd->obd_name);
 
-        obd_unregister_lock_cancel_cb(lov->lov_tgts[index]->ltd_exp,
-                                      lov->lov_lock_cancel_cb);
-        obd_unregister_page_removal_cb(lov->lov_tgts[index]->ltd_exp,
-                                       lov->lov_page_removal_cb);
-
-        if (lov->lov_tgts[index]->ltd_active) {
-                lov->lov_tgts[index]->ltd_active = 0;
+        if (tgt->ltd_active) {
+                tgt->ltd_active = 0;
                 lov->desc.ld_active_tgt_count--;
-                lov->lov_tgts[index]->ltd_exp->exp_obd->obd_inactive = 1;
+                tgt->ltd_exp->exp_obd->obd_inactive = 1;
         }
 
         lov_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
@@ -433,22 +328,19 @@ static int lov_disconnect_obd(struct obd_device *obd, __u32 index)
 
         obd_register_observer(osc_obd, NULL);
 
-        rc = obd_disconnect(lov->lov_tgts[index]->ltd_exp);
+        rc = obd_disconnect(tgt->ltd_exp);
         if (rc) {
                 CERROR("Target %s disconnect error %d\n",
-                       lov_uuid2str(lov, index), rc);
+                       tgt->ltd_uuid.uuid, rc);
                 rc = 0;
         }
 
-        qos_del_tgt(obd, index);
+        qos_del_tgt(obd, tgt);
 
-        lov->lov_tgts[index]->ltd_exp = NULL;
+        tgt->ltd_exp = NULL;
         RETURN(0);
 }
 
-static int lov_del_target(struct obd_device *obd, __u32 index,
-                          struct obd_uuid *uuidp, int gen);
-
 static int lov_disconnect(struct obd_export *exp)
 {
         struct obd_device *obd = class_exp2obd(exp);
@@ -469,14 +361,15 @@ static int lov_disconnect(struct obd_export *exp)
 
         /* Let's hold another reference so lov_del_obd doesn't spin through
            putref every time */
-        lov_getref(obd);
+        obd_getref(obd);
+
         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
                 if (lov->lov_tgts[i] && lov->lov_tgts[i]->ltd_exp) {
                         /* Disconnection is the last we know about an obd */
                         lov_del_target(obd, i, 0, lov->lov_tgts[i]->ltd_gen);
                 }
         }
-        lov_putref(obd);
+        obd_putref(obd);
 
 out:
         rc = class_disconnect(exp); /* bz 9811 */
@@ -486,20 +379,22 @@ out:
 /* Error codes:
  *
  *  -EINVAL  : UUID can't be found in the LOV's target list
- * - any other is lov index
+ *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
+ *  -EBADF   : The UUID is found, but the OBD is the wrong type (!)
+ *  any >= 0 : is log target index
  */
 static int lov_set_osc_active(struct obd_device *obd, struct obd_uuid *uuid,
-                              int activate)
+                              enum obd_notify_event ev)
 {
         struct lov_obd *lov = &obd->u.lov;
         struct lov_tgt_desc *tgt;
-        int index;
+        int index, activate, active;
         ENTRY;
 
-        CDEBUG(D_INFO, "Searching in lov %p for uuid %s (activate=%d)\n",
-               lov, uuid->uuid, activate);
+        CDEBUG(D_INFO, "Searching in lov %p for uuid %s event(%d)\n",
+               lov, uuid->uuid, ev);
 
-        lov_getref(obd);
+        obd_getref(obd);
         for (index = 0; index < lov->desc.ld_tgt_count; index++) {
                 tgt = lov->lov_tgts[index];
                 if (!tgt || !tgt->ltd_exp)
@@ -515,29 +410,46 @@ static int lov_set_osc_active(struct obd_device *obd, struct obd_uuid *uuid,
         if (index == lov->desc.ld_tgt_count)
                 GOTO(out, index = -EINVAL);
 
-        if (lov->lov_tgts[index]->ltd_active == activate) {
-                CDEBUG(D_INFO, "OSC %s already %sactive!\n", uuid->uuid,
-                       activate ? "" : "in");
-                GOTO(out, index);
-        }
+        if (ev == OBD_NOTIFY_DEACTIVATE || ev == OBD_NOTIFY_ACTIVATE) {
+                activate = (ev == OBD_NOTIFY_ACTIVATE) ? 1 : 0;
 
-        CDEBUG(D_CONFIG, "Marking OSC %s %sactive\n", obd_uuid2str(uuid),
-               activate ? "" : "in");
+                if (lov->lov_tgts[index]->ltd_activate == activate) {
+                        CDEBUG(D_INFO, "OSC %s already %sactivate!\n",
+                               uuid->uuid, activate ? "" : "de");
+                } else {
+                        lov->lov_tgts[index]->ltd_activate = activate;
+                        CDEBUG(D_CONFIG, "%sactivate OSC %s\n",
+                               activate ? "" : "de", obd_uuid2str(uuid));
+                }
 
-        lov->lov_tgts[index]->ltd_active = activate;
+        } else if (ev == OBD_NOTIFY_INACTIVE || ev == OBD_NOTIFY_ACTIVE) {
+                active = (ev == OBD_NOTIFY_ACTIVE) ? 1 : 0;
 
-        if (activate) {
-                lov->desc.ld_active_tgt_count++;
-                lov->lov_tgts[index]->ltd_exp->exp_obd->obd_inactive = 0;
+                if (lov->lov_tgts[index]->ltd_active == active) {
+                        CDEBUG(D_INFO, "OSC %s already %sactive!\n",
+                               uuid->uuid, active ? "" : "in");
+                        GOTO(out, index);
+                } else {
+                        CDEBUG(D_CONFIG, "Marking OSC %s %sactive\n",
+                               obd_uuid2str(uuid), active ? "" : "in");
+                }
+
+                lov->lov_tgts[index]->ltd_active = active;
+                if (active) {
+                        lov->desc.ld_active_tgt_count++;
+                        lov->lov_tgts[index]->ltd_exp->exp_obd->obd_inactive = 0;
+                } else {
+                        lov->desc.ld_active_tgt_count--;
+                        lov->lov_tgts[index]->ltd_exp->exp_obd->obd_inactive = 1;
+                }
+                /* remove any old qos penalty */
+                lov->lov_tgts[index]->ltd_qos.ltq_penalty = 0;
         } else {
-                lov->desc.ld_active_tgt_count--;
-                lov->lov_tgts[index]->ltd_exp->exp_obd->obd_inactive = 1;
+                CERROR("Unknown event(%d) for uuid %s", ev, uuid->uuid);
         }
-        /* remove any old qos penalty */
-        lov->lov_tgts[index]->ltd_qos.ltq_penalty = 0;
 
  out:
-        lov_putref(obd);
+        obd_putref(obd);
         RETURN(index);
 }
 
@@ -547,7 +459,8 @@ static int lov_notify(struct obd_device *obd, struct obd_device *watched,
         int rc = 0;
         ENTRY;
 
-        if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
+        if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE ||
+            ev == OBD_NOTIFY_ACTIVATE || ev == OBD_NOTIFY_DEACTIVATE) {
                 struct obd_uuid *uuid;
 
                 LASSERT(watched);
@@ -563,13 +476,13 @@ static int lov_notify(struct obd_device *obd, struct obd_device *watched,
                 /* Set OSC as active before notifying the observer, so the
                  * observer can use the OSC normally.
                  */
-                rc = lov_set_osc_active(obd, uuid, ev == OBD_NOTIFY_ACTIVE);
+                rc = lov_set_osc_active(obd, uuid, ev);
                 if (rc < 0) {
-                        CERROR("%sactivation of %s failed: %d\n",
-                               (ev == OBD_NOTIFY_ACTIVE) ? "" : "de",
+                        CERROR("event(%d) of %s failed: %d\n", ev,
                                obd_uuid2str(uuid), rc);
                         RETURN(rc);
                 }
+                /* active event should be pass lov target index as data */
                 data = &rc;
         }
 
@@ -580,28 +493,33 @@ static int lov_notify(struct obd_device *obd, struct obd_device *watched,
                 /* NULL watched means all osc's in the lov (only for syncs) */
                 /* sync event should be send lov idx as data */
                 struct lov_obd *lov = &obd->u.lov;
-                struct obd_device *tgt_obd;
-                int i;
-                lov_getref(obd);
+                int i, is_sync;
+
+                data = &i;
+                is_sync = (ev == OBD_NOTIFY_SYNC) ||
+                          (ev == OBD_NOTIFY_SYNC_NONBLOCK);
+
+                obd_getref(obd);
                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
                         if (!lov->lov_tgts[i])
                                 continue;
 
-                        if ((ev == OBD_NOTIFY_SYNC) ||
-                            (ev == OBD_NOTIFY_SYNC_NONBLOCK))
-                                data = &i;
+                        /* don't send sync event if target not
+                         * connected/activated */
+                        if (is_sync &&  !lov->lov_tgts[i]->ltd_active)
+                                continue;
 
-                        tgt_obd = class_exp2obd(lov->lov_tgts[i]->ltd_exp);
-                        rc = obd_notify_observer(obd, tgt_obd, ev, data);
+                        rc = obd_notify_observer(obd, lov->lov_tgts[i]->ltd_obd,
+                                                 ev, data);
                         if (rc) {
                                 CERROR("%s: notify %s of %s failed %d\n",
                                        obd->obd_name,
                                        obd->obd_observer->obd_name,
-                                       tgt_obd->obd_name, rc);
-                                break;
+                                       lov->lov_tgts[i]->ltd_obd->obd_name,
+                                       rc);
                         }
                 }
-                lov_putref(obd);
+                obd_putref(obd);
         }
 
         RETURN(rc);
@@ -612,6 +530,7 @@ static int lov_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
 {
         struct lov_obd *lov = &obd->u.lov;
         struct lov_tgt_desc *tgt;
+        struct obd_device *tgt_obd;
         int rc;
         ENTRY;
 
@@ -624,13 +543,18 @@ static int lov_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
                 RETURN(-EINVAL);
         }
 
-        mutex_down(&lov->lov_lock);
+        tgt_obd = class_find_client_obd(uuidp, LUSTRE_OSC_NAME,
+                                        &obd->obd_uuid);
+        if (tgt_obd == NULL)
+                RETURN(-EINVAL);
+
+        cfs_mutex_down(&lov->lov_lock);
 
         if ((index < lov->lov_tgt_size) && (lov->lov_tgts[index] != NULL)) {
                 tgt = lov->lov_tgts[index];
                 CERROR("UUID %s already assigned at LOV target index %d\n",
                        obd_uuid2str(&tgt->ltd_uuid), index);
-                mutex_up(&lov->lov_lock);
+                cfs_mutex_up(&lov->lov_lock);
                 RETURN(-EEXIST);
         }
 
@@ -644,7 +568,7 @@ static int lov_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
                         newsize = newsize << 1;
                 OBD_ALLOC(newtgts, sizeof(*newtgts) * newsize);
                 if (newtgts == NULL) {
-                        mutex_up(&lov->lov_lock);
+                        cfs_mutex_up(&lov->lov_lock);
                         RETURN(-ENOMEM);
                 }
 
@@ -667,15 +591,21 @@ static int lov_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
                        lov->lov_tgts, lov->lov_tgt_size);
         }
 
-
         OBD_ALLOC_PTR(tgt);
         if (!tgt) {
-                mutex_up(&lov->lov_lock);
+                cfs_mutex_up(&lov->lov_lock);
                 RETURN(-ENOMEM);
         }
 
-        memset(tgt, 0, sizeof(*tgt));
+        rc = lov_ost_pool_add(&lov->lov_packed, index, lov->lov_tgt_size);
+        if (rc) {
+                cfs_mutex_up(&lov->lov_lock);
+                OBD_FREE_PTR(tgt);
+                RETURN(rc);
+        }
+
         tgt->ltd_uuid = *uuidp;
+        tgt->ltd_obd = tgt_obd;
         /* XXX - add a sanity check on the generation number. */
         tgt->ltd_gen = gen;
         tgt->ltd_index = index;
@@ -683,11 +613,14 @@ static int lov_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
         lov->lov_tgts[index] = tgt;
         if (index >= lov->desc.ld_tgt_count)
                 lov->desc.ld_tgt_count = index + 1;
-        mutex_up(&lov->lov_lock);
+
+        cfs_mutex_up(&lov->lov_lock);
 
         CDEBUG(D_CONFIG, "idx=%d ltd_gen=%d ld_tgt_count=%d\n",
                 index, tgt->ltd_gen, lov->desc.ld_tgt_count);
 
+        rc = obd_notify(obd, tgt_obd, OBD_NOTIFY_CREATE, &index);
+
         if (lov->lov_connects == 0) {
                 /* lov_connect hasn't been called yet. We'll do the
                    lov_connect_obd on this target when that fn first runs,
@@ -695,14 +628,18 @@ static int lov_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
                 RETURN(0);
         }
 
-        lov_getref(obd);
+        obd_getref(obd);
 
         rc = lov_connect_obd(obd, index, active, &lov->lov_ocd);
         if (rc)
                 GOTO(out, rc);
 
+        /* connect to administrative disabled ost */
+        if (!tgt->ltd_exp)
+                GOTO(out, rc = 0);
+
         rc = lov_notify(obd, tgt->ltd_exp->exp_obd,
-                        active ? OBD_NOTIFY_ACTIVE : OBD_NOTIFY_INACTIVE,
+                        active ? OBD_NOTIFY_CONNECT : OBD_NOTIFY_INACTIVE,
                         (void *)&index);
 
 out:
@@ -711,13 +648,13 @@ out:
                        obd_uuid2str(&tgt->ltd_uuid));
                 lov_del_target(obd, index, 0, 0);
         }
-        lov_putref(obd);
+        obd_putref(obd);
         RETURN(rc);
 }
 
 /* Schedule a target for deletion */
-static int lov_del_target(struct obd_device *obd, __u32 index,
-                          struct obd_uuid *uuidp, int gen)
+int lov_del_target(struct obd_device *obd, __u32 index,
+                   struct obd_uuid *uuidp, int gen)
 {
         struct lov_obd *lov = &obd->u.lov;
         int count = lov->desc.ld_tgt_count;
@@ -730,7 +667,7 @@ static int lov_del_target(struct obd_device *obd, __u32 index,
                 RETURN(-EINVAL);
         }
 
-        lov_getref(obd);
+        obd_getref(obd);
 
         if (!lov->lov_tgts[index]) {
                 CERROR("LOV target at index %d is not setup.\n", index);
@@ -751,19 +688,16 @@ static int lov_del_target(struct obd_device *obd, __u32 index,
 
         lov->lov_tgts[index]->ltd_reap = 1;
         lov->lov_death_row++;
-        /* we really delete it from lov_putref */
+        /* we really delete it from obd_putref */
 out:
-        lov_putref(obd);
+        obd_putref(obd);
 
         RETURN(rc);
 }
 
-/* We are holding lov_lock */
-static void __lov_del_obd(struct obd_device *obd, __u32 index)
+static void __lov_del_obd(struct obd_device *obd, struct lov_tgt_desc *tgt)
 {
-        struct lov_obd *lov = &obd->u.lov;
         struct obd_device *osc_obd;
-        struct lov_tgt_desc *tgt = lov->lov_tgts[index];
 
         LASSERT(tgt);
         LASSERT(tgt->ltd_reap);
@@ -771,17 +705,12 @@ static void __lov_del_obd(struct obd_device *obd, __u32 index)
         osc_obd = class_exp2obd(tgt->ltd_exp);
 
         CDEBUG(D_CONFIG, "Removing tgt %s : %s\n",
-               lov_uuid2str(lov, index),
+               tgt->ltd_uuid.uuid,
                osc_obd ? osc_obd->obd_name : "<no obd>");
 
         if (tgt->ltd_exp)
-                lov_disconnect_obd(obd, index);
-
-        /* XXX - right now there is a dependency on ld_tgt_count being the
-         * maximum tgt index for computing the mds_max_easize. So we can't
-         * shrink it. */
+                lov_disconnect_obd(obd, tgt);
 
-        lov->lov_tgts[index] = NULL;
         OBD_FREE_PTR(tgt);
 
         /* Manual cleanup - no cleanup logs to clean up the osc's.  We must
@@ -835,12 +764,12 @@ void lov_fix_desc(struct lov_desc *desc)
         lov_fix_desc_qos_maxage(&desc->ld_qos_maxage);
 }
 
-static int lov_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
+int lov_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 {
         struct lprocfs_static_vars lvars = { 0 };
         struct lov_desc *desc;
         struct lov_obd *lov = &obd->u.lov;
-        int count;
+        int rc;
         ENTRY;
 
         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
@@ -870,29 +799,43 @@ static int lov_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 
         lov_fix_desc(desc);
 
-        /* Because of 64-bit divide/mod operations only work with a 32-bit
-         * divisor in a 32-bit kernel, we cannot support a stripe width
-         * of 4GB or larger on 32-bit CPUs. */
-        count = desc->ld_default_stripe_count;
-        if ((count > 0 ? count : desc->ld_tgt_count) *
-            desc->ld_default_stripe_size > 0xffffffff) {
-                CERROR("LOV: stripe width "LPU64"x%u > 4294967295 bytes\n",
-                       desc->ld_default_stripe_size, count);
-                RETURN(-EINVAL);
-        }
-
         desc->ld_active_tgt_count = 0;
         lov->desc = *desc;
         lov->lov_tgt_size = 0;
-        sema_init(&lov->lov_lock, 1);
-        atomic_set(&lov->lov_refcount, 0);
+
+        cfs_sema_init(&lov->lov_lock, 1);
+        cfs_atomic_set(&lov->lov_refcount, 0);
         CFS_INIT_LIST_HEAD(&lov->lov_qos.lq_oss_list);
-        init_rwsem(&lov->lov_qos.lq_rw_sem);
+        cfs_init_rwsem(&lov->lov_qos.lq_rw_sem);
+        lov->lov_sp_me = LUSTRE_SP_CLI;
         lov->lov_qos.lq_dirty = 1;
-        lov->lov_qos.lq_dirty_rr = 1;
+        lov->lov_qos.lq_rr.lqr_dirty = 1;
         lov->lov_qos.lq_reset = 1;
         /* Default priority is toward free space balance */
         lov->lov_qos.lq_prio_free = 232;
+        /* Default threshold for rr (roughly 17%) */
+        lov->lov_qos.lq_threshold_rr = 43;
+        /* Init statfs fields */
+        OBD_ALLOC_PTR(lov->lov_qos.lq_statfs_data);
+        if (NULL == lov->lov_qos.lq_statfs_data)
+                RETURN(-ENOMEM);
+        cfs_waitq_init(&lov->lov_qos.lq_statfs_waitq);
+
+        lov->lov_pools_hash_body = cfs_hash_create("POOLS", HASH_POOLS_CUR_BITS,
+                                                   HASH_POOLS_MAX_BITS,
+                                                   HASH_POOLS_BKT_BITS, 0,
+                                                   CFS_HASH_MIN_THETA,
+                                                   CFS_HASH_MAX_THETA,
+                                                   &pool_hash_operations,
+                                                   CFS_HASH_DEFAULT);
+        CFS_INIT_LIST_HEAD(&lov->lov_pool_list);
+        lov->lov_pool_count = 0;
+        rc = lov_ost_pool_init(&lov->lov_packed, 0);
+        if (rc)
+                GOTO(out_free_statfs, rc);
+        rc = lov_ost_pool_init(&lov->lov_qos.lq_rr.lqr_pool, 0);
+        if (rc)
+                GOTO(out_free_lov_packed, rc);
 
         lprocfs_lov_init_vars(&lvars);
         lprocfs_obd_setup(obd, lvars.obd_vars);
@@ -906,18 +849,28 @@ static int lov_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
                         CWARN("Error adding the target_obd file\n");
         }
 #endif
+        lov->lov_pool_proc_entry = lprocfs_register("pools",
+                                                    obd->obd_proc_entry,
+                                                    NULL, NULL);
 
         RETURN(0);
+
+out_free_lov_packed:
+        lov_ost_pool_free(&lov->lov_packed);
+out_free_statfs:
+        OBD_FREE_PTR(lov->lov_qos.lq_statfs_data);
+        return rc;
 }
 
 static int lov_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
 {
         int rc = 0;
+        struct lov_obd *lov = &obd->u.lov;
+
         ENTRY;
 
         switch (stage) {
         case OBD_CLEANUP_EARLY: {
-                struct lov_obd *lov = &obd->u.lov;
                 int i;
                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
                         if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
@@ -939,40 +892,54 @@ static int lov_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
 static int lov_cleanup(struct obd_device *obd)
 {
         struct lov_obd *lov = &obd->u.lov;
+        cfs_list_t *pos, *tmp;
+        struct pool_desc *pool;
+
+        cfs_list_for_each_safe(pos, tmp, &lov->lov_pool_list) {
+                pool = cfs_list_entry(pos, struct pool_desc, pool_list);
+                /* free pool structs */
+                CDEBUG(D_INFO, "delete pool %p\n", pool);
+                lov_pool_del(obd, pool->pool_name);
+        }
+        cfs_hash_putref(lov->lov_pools_hash_body);
+        lov_ost_pool_free(&(lov->lov_qos.lq_rr.lqr_pool));
+        lov_ost_pool_free(&lov->lov_packed);
 
-        lprocfs_obd_cleanup(obd);
         if (lov->lov_tgts) {
                 int i;
+                obd_getref(obd);
                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
                         if (!lov->lov_tgts[i])
                                 continue;
 
                         /* Inactive targets may never have connected */
                         if (lov->lov_tgts[i]->ltd_active ||
-                            atomic_read(&lov->lov_refcount))
+                            cfs_atomic_read(&lov->lov_refcount))
                             /* We should never get here - these
                                should have been removed in the
                              disconnect. */
                                 CERROR("lov tgt %d not cleaned!"
                                        " deathrow=%d, lovrc=%d\n",
                                        i, lov->lov_death_row,
-                                       atomic_read(&lov->lov_refcount));
+                                       cfs_atomic_read(&lov->lov_refcount));
                         lov_del_target(obd, i, 0, 0);
                 }
+                obd_putref(obd);
                 OBD_FREE(lov->lov_tgts, sizeof(*lov->lov_tgts) *
                          lov->lov_tgt_size);
                 lov->lov_tgt_size = 0;
         }
 
-        if (lov->lov_qos.lq_rr_size)
-                OBD_FREE(lov->lov_qos.lq_rr_array, lov->lov_qos.lq_rr_size);
+        /* clear pools parent proc entry only after all pools is killed */
+        lprocfs_obd_cleanup(obd);
 
+        OBD_FREE_PTR(lov->lov_qos.lq_statfs_data);
         RETURN(0);
 }
 
-static int lov_process_config(struct obd_device *obd, obd_count len, void *buf)
+int lov_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg,
+                            __u32 *indexp, int *genp)
 {
-        struct lustre_cfg *lcfg = buf;
         struct obd_uuid obd_uuid;
         int cmd;
         int rc = 0;
@@ -990,10 +957,12 @@ static int lov_process_config(struct obd_device *obd, obd_count len, void *buf)
 
                 obd_str2uuid(&obd_uuid,  lustre_cfg_buf(lcfg, 1));
 
-                if (sscanf(lustre_cfg_buf(lcfg, 2), "%d", &index) != 1)
+                if (sscanf(lustre_cfg_buf(lcfg, 2), "%d", indexp) != 1)
                         GOTO(out, rc = -EINVAL);
-                if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
+                if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", genp) != 1)
                         GOTO(out, rc = -EINVAL);
+                index = *indexp;
+                gen = *genp;
                 if (cmd == LCFG_LOV_ADD_OBD)
                         rc = lov_add_target(obd, &obd_uuid, index, gen, 1);
                 else if (cmd == LCFG_LOV_ADD_INA)
@@ -1013,8 +982,16 @@ static int lov_process_config(struct obd_device *obd, obd_count len, void *buf)
 
                 rc = class_process_proc_param(PARAM_LOV, lvars.obd_vars,
                                               lcfg, obd);
+                if (rc > 0)
+                        rc = 0;
                 GOTO(out, rc);
         }
+        case LCFG_POOL_NEW:
+        case LCFG_POOL_ADD:
+        case LCFG_POOL_DEL:
+        case LCFG_POOL_REM:
+                GOTO(out, rc);
+
         default: {
                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
                 GOTO(out, rc = -EINVAL);
@@ -1026,7 +1003,7 @@ out:
 }
 
 #ifndef log2
-#define log2(n) ffz(~(n))
+#define log2(n) cfs_ffz(~(n))
 #endif
 
 static int lov_clear_orphans(struct obd_export *export, struct obdo *src_oa,
@@ -1054,7 +1031,7 @@ static int lov_clear_orphans(struct obd_export *export, struct obdo *src_oa,
                        ost_uuid->uuid);
         }
 
-        lov_getref(export->exp_obd);
+        obd_getref(export->exp_obd);
         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
                 struct lov_stripe_md obj_md;
                 struct lov_stripe_md *obj_mdp = &obj_md;
@@ -1084,16 +1061,18 @@ static int lov_clear_orphans(struct obd_export *export, struct obdo *src_oa,
                 /* XXX: LOV STACKING: use real "obj_mdp" sub-data */
                 err = obd_create(lov->lov_tgts[i]->ltd_exp,
                                  tmp_oa, &obj_mdp, oti);
-                if (err)
+                if (err) {
                         /* This export will be disabled until it is recovered,
                            and then orphan recovery will be completed. */
                         CERROR("error in orphan recovery on OST idx %d/%d: "
                                "rc = %d\n", i, lov->desc.ld_tgt_count, err);
+                        rc = err;
+                }
 
                 if (ost_uuid)
                         break;
         }
-        lov_putref(export->exp_obd);
+        obd_putref(export->exp_obd);
 
         OBDO_FREE(tmp_oa);
         RETURN(rc);
@@ -1147,8 +1126,7 @@ static int lov_create(struct obd_export *exp, struct obdo *src_oa,
         struct obd_info oinfo;
         struct lov_request_set *set = NULL;
         struct lov_request *req;
-        struct obd_statfs osfs;
-        __u64 maxage;
+        struct l_wait_info  lwi = { 0 };
         int rc = 0;
         ENTRY;
 
@@ -1166,48 +1144,64 @@ static int lov_create(struct obd_export *exp, struct obdo *src_oa,
         if (!lov->desc.ld_active_tgt_count)
                 RETURN(-EIO);
 
+        obd_getref(exp->exp_obd);
         /* Recreate a specific object id at the given OST index */
         if ((src_oa->o_valid & OBD_MD_FLFLAGS) &&
             (src_oa->o_flags & OBD_FL_RECREATE_OBJS)) {
                  rc = lov_recreate(exp, src_oa, ea, oti);
-                 RETURN(rc);
+                 GOTO(out, rc);
         }
 
-        maxage = cfs_time_shift_64(-lov->desc.ld_qos_maxage);
-        obd_statfs_rqset(exp->exp_obd, &osfs, maxage, OBD_STATFS_NODELAY);
+        /* issue statfs rpcs if the osfs data is older than qos_maxage - 1s,
+         * later in alloc_qos(), we will wait for those rpcs to complete if
+         * the osfs age is older than 2 * qos_maxage */
+        qos_statfs_update(exp->exp_obd,
+                          cfs_time_shift_64(-lov->desc.ld_qos_maxage +
+                                            OBD_STATFS_CACHE_SECONDS),
+                          0);
 
         rc = lov_prep_create_set(exp, &oinfo, ea, src_oa, oti, &set);
         if (rc)
-                RETURN(rc);
+                GOTO(out, rc);
 
-        list_for_each_entry(req, &set->set_list, rq_link) {
+        cfs_list_for_each_entry(req, &set->set_list, rq_link) {
                 /* XXX: LOV STACKING: use real "obj_mdp" sub-data */
-                rc = obd_create(lov->lov_tgts[req->rq_idx]->ltd_exp,
-                                req->rq_oi.oi_oa, &req->rq_oi.oi_md, oti);
-                lov_update_create_set(set, req, rc);
+                rc = obd_create_async(lov->lov_tgts[req->rq_idx]->ltd_exp,
+                                      &req->rq_oi, &req->rq_oi.oi_md, oti);
         }
+
+        /* osc_create have timeout equ obd_timeout/2 so waiting don't be
+         * longer then this */
+        l_wait_event(set->set_waitq, lov_finished_set(set), &lwi);
+
+        /* we not have ptlrpc set for assign set->interpret and should
+         * be call interpret function himself. calling from cb_create_update
+         * not permited because lov_fini_create_set can sleep for long time,
+         * but we must avoid sleeping in ptlrpcd interpret function. */
         rc = lov_fini_create_set(set, ea);
+out:
+        obd_putref(exp->exp_obd);
         RETURN(rc);
 }
 
 #define ASSERT_LSM_MAGIC(lsmp)                                                  \
 do {                                                                            \
         LASSERT((lsmp) != NULL);                                                \
-        LASSERTF(((lsmp)->lsm_magic == LOV_MAGIC ||                             \
-                 (lsmp)->lsm_magic == LOV_MAGIC_JOIN), "%p->lsm_magic=%x\n",    \
-                 (lsmp), (lsmp)->lsm_magic);                                    \
+        LASSERTF(((lsmp)->lsm_magic == LOV_MAGIC_V1 ||                          \
+                 (lsmp)->lsm_magic == LOV_MAGIC_V3),                            \
+                 "%p->lsm_magic=%x\n", (lsmp), (lsmp)->lsm_magic);              \
 } while (0)
 
 static int lov_destroy(struct obd_export *exp, struct obdo *oa,
                        struct lov_stripe_md *lsm, struct obd_trans_info *oti,
-                       struct obd_export *md_exp)
+                       struct obd_export *md_exp, void *capa)
 {
         struct lov_request_set *set;
         struct obd_info oinfo;
         struct lov_request *req;
-        struct list_head *pos;
+        cfs_list_t *pos;
         struct lov_obd *lov;
-        int rc = 0, err;
+        int rc = 0, err = 0;
         ENTRY;
 
         ASSERT_LSM_MAGIC(lsm);
@@ -1221,19 +1215,19 @@ static int lov_destroy(struct obd_export *exp, struct obdo *oa,
         }
 
         lov = &exp->exp_obd->u.lov;
+        obd_getref(exp->exp_obd);
         rc = lov_prep_destroy_set(exp, &oinfo, oa, lsm, oti, &set);
         if (rc)
-                RETURN(rc);
+                GOTO(out, rc);
 
-        list_for_each (pos, &set->set_list) {
-                int err;
-                req = list_entry(pos, struct lov_request, rq_link);
+        cfs_list_for_each (pos, &set->set_list) {
+                req = cfs_list_entry(pos, struct lov_request, rq_link);
 
                 if (oa->o_valid & OBD_MD_FLCOOKIE)
                         oti->oti_logcookies = set->set_cookies + req->rq_stripe;
 
                 err = obd_destroy(lov->lov_tgts[req->rq_idx]->ltd_exp,
-                                  req->rq_oi.oi_oa, NULL, oti, NULL);
+                                  req->rq_oi.oi_oa, NULL, oti, NULL, capa);
                 err = lov_update_common_set(set, req, err);
                 if (err) {
                         CERROR("error: destroying objid "LPX64" subobj "
@@ -1250,6 +1244,8 @@ static int lov_destroy(struct obd_export *exp, struct obdo *oa,
                 rc = lsm_op_find(lsm->lsm_magic)->lsm_destroy(lsm, oa, md_exp);
         }
         err = lov_fini_destroy_set(set);
+out:
+        obd_putref(exp->exp_obd);
         RETURN(rc ? rc : err);
 }
 
@@ -1257,7 +1253,7 @@ static int lov_getattr(struct obd_export *exp, struct obd_info *oinfo)
 {
         struct lov_request_set *set;
         struct lov_request *req;
-        struct list_head *pos;
+        cfs_list_t *pos;
         struct lov_obd *lov;
         int err = 0, rc = 0;
         ENTRY;
@@ -1274,8 +1270,8 @@ static int lov_getattr(struct obd_export *exp, struct obd_info *oinfo)
         if (rc)
                 RETURN(rc);
 
-        list_for_each (pos, &set->set_list) {
-                req = list_entry(pos, struct lov_request, rq_link);
+        cfs_list_for_each (pos, &set->set_list) {
+                req = cfs_list_entry(pos, struct lov_request, rq_link);
 
                 CDEBUG(D_INFO, "objid "LPX64"[%d] has subobj "LPX64" at idx "
                        "%u\n", oinfo->oi_oa->o_id, req->rq_stripe,
@@ -1318,7 +1314,7 @@ static int lov_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
 {
         struct lov_request_set *lovset;
         struct lov_obd *lov;
-        struct list_head *pos;
+        cfs_list_t *pos;
         struct lov_request *req;
         int rc = 0, err;
         ENTRY;
@@ -1339,8 +1335,8 @@ static int lov_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
                oinfo->oi_md->lsm_object_id, oinfo->oi_md->lsm_stripe_count,
                oinfo->oi_md->lsm_stripe_size);
 
-        list_for_each (pos, &lovset->set_list) {
-                req = list_entry(pos, struct lov_request, rq_link);
+        cfs_list_for_each (pos, &lovset->set_list) {
+                req = cfs_list_entry(pos, struct lov_request, rq_link);
 
                 CDEBUG(D_INFO, "objid "LPX64"[%d] has subobj "LPX64" at idx "
                        "%u\n", oinfo->oi_oa->o_id, req->rq_stripe,
@@ -1356,7 +1352,7 @@ static int lov_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
                 }
         }
 
-        if (!list_empty(&rqset->set_requests)) {
+        if (!cfs_list_empty(&rqset->set_requests)) {
                 LASSERT(rc == 0);
                 LASSERT (rqset->set_interpret == NULL);
                 rqset->set_interpret = lov_getattr_interpret;
@@ -1375,7 +1371,7 @@ static int lov_setattr(struct obd_export *exp, struct obd_info *oinfo,
 {
         struct lov_request_set *set;
         struct lov_obd *lov;
-        struct list_head *pos;
+        cfs_list_t *pos;
         struct lov_request *req;
         int err = 0, rc = 0;
         ENTRY;
@@ -1399,8 +1395,8 @@ static int lov_setattr(struct obd_export *exp, struct obd_info *oinfo,
         if (rc)
                 RETURN(rc);
 
-        list_for_each (pos, &set->set_list) {
-                req = list_entry(pos, struct lov_request, rq_link);
+        cfs_list_for_each (pos, &set->set_list) {
+                req = cfs_list_entry(pos, struct lov_request, rq_link);
 
                 rc = obd_setattr(lov->lov_tgts[req->rq_idx]->ltd_exp,
                                  &req->rq_oi, NULL);
@@ -1441,7 +1437,7 @@ static int lov_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
 {
         struct lov_request_set *set;
         struct lov_request *req;
-        struct list_head *pos;
+        cfs_list_t *pos;
         struct lov_obd *lov;
         int rc = 0;
         ENTRY;
@@ -1465,8 +1461,8 @@ static int lov_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
                oinfo->oi_md->lsm_object_id, oinfo->oi_md->lsm_stripe_count,
                oinfo->oi_md->lsm_stripe_size);
 
-        list_for_each (pos, &set->set_list) {
-                req = list_entry(pos, struct lov_request, rq_link);
+        cfs_list_for_each (pos, &set->set_list) {
+                req = cfs_list_entry(pos, struct lov_request, rq_link);
 
                 if (oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
                         oti->oti_logcookies = set->set_cookies + req->rq_stripe;
@@ -1488,7 +1484,7 @@ static int lov_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
         }
 
         /* If we are not waiting for responses on async requests, return. */
-        if (rc || !rqset || list_empty(&rqset->set_requests)) {
+        if (rc || !rqset || cfs_list_empty(&rqset->set_requests)) {
                 int err;
                 if (rc)
                         set->set_completes = 0;
@@ -1525,7 +1521,7 @@ static int lov_punch(struct obd_export *exp, struct obd_info *oinfo,
 {
         struct lov_request_set *set;
         struct lov_obd *lov;
-        struct list_head *pos;
+        cfs_list_t *pos;
         struct lov_request *req;
         int rc = 0;
         ENTRY;
@@ -1541,8 +1537,8 @@ static int lov_punch(struct obd_export *exp, struct obd_info *oinfo,
         if (rc)
                 RETURN(rc);
 
-        list_for_each (pos, &set->set_list) {
-                req = list_entry(pos, struct lov_request, rq_link);
+        cfs_list_for_each (pos, &set->set_list) {
+                req = cfs_list_entry(pos, struct lov_request, rq_link);
 
                 rc = obd_punch(lov->lov_tgts[req->rq_idx]->ltd_exp,
                                &req->rq_oi, NULL, rqset);
@@ -1555,7 +1551,7 @@ static int lov_punch(struct obd_export *exp, struct obd_info *oinfo,
                 }
         }
 
-        if (rc || list_empty(&rqset->set_requests)) {
+        if (rc || cfs_list_empty(&rqset->set_requests)) {
                 int err;
                 err = lov_fini_punch_set(set);
                 RETURN(rc ? rc : err);
@@ -1568,49 +1564,71 @@ static int lov_punch(struct obd_export *exp, struct obd_info *oinfo,
         RETURN(0);
 }
 
-static int lov_sync(struct obd_export *exp, struct obdo *oa,
-                    struct lov_stripe_md *lsm, obd_off start, obd_off end,
-                    void *capa)
+static int lov_sync_interpret(struct ptlrpc_request_set *rqset,
+                              void *data, int rc)
 {
-        struct lov_request_set *set;
-        struct obd_info oinfo;
+        struct lov_request_set *lovset = data;
+        int err;
+        ENTRY;
+
+        if (rc)
+                lovset->set_completes = 0;
+        err = lov_fini_sync_set(lovset);
+        RETURN(rc ?: err);
+}
+
+static int lov_sync(struct obd_export *exp, struct obd_info *oinfo,
+                    obd_off start, obd_off end,
+                    struct ptlrpc_request_set *rqset)
+{
+        struct lov_request_set *set = NULL;
         struct lov_obd *lov;
-        struct list_head *pos;
+        cfs_list_t *pos;
         struct lov_request *req;
-        int err = 0, rc = 0;
+        int rc = 0;
         ENTRY;
 
-        ASSERT_LSM_MAGIC(lsm);
+        ASSERT_LSM_MAGIC(oinfo->oi_md);
+        LASSERT(rqset != NULL);
 
         if (!exp->exp_obd)
                 RETURN(-ENODEV);
 
         lov = &exp->exp_obd->u.lov;
-        rc = lov_prep_sync_set(exp, &oinfo, oa, lsm, start, end, &set);
+        rc = lov_prep_sync_set(exp, oinfo, start, end, &set);
         if (rc)
                 RETURN(rc);
 
-        list_for_each (pos, &set->set_list) {
-                req = list_entry(pos, struct lov_request, rq_link);
+        CDEBUG(D_INFO, "fsync objid "LPX64" ["LPX64", "LPX64"]\n",
+               set->set_oi->oi_oa->o_id, start, end);
+
+        cfs_list_for_each (pos, &set->set_list) {
+                req = cfs_list_entry(pos, struct lov_request, rq_link);
 
-                rc = obd_sync(lov->lov_tgts[req->rq_idx]->ltd_exp,
-                              req->rq_oi.oi_oa, NULL,
+                rc = obd_sync(lov->lov_tgts[req->rq_idx]->ltd_exp, &req->rq_oi,
                               req->rq_oi.oi_policy.l_extent.start,
-                              req->rq_oi.oi_policy.l_extent.end, capa);
-                err = lov_update_common_set(set, req, rc);
-                if (err) {
+                              req->rq_oi.oi_policy.l_extent.end, rqset);
+                if (rc) {
                         CERROR("error: fsync objid "LPX64" subobj "LPX64
                                " on OST idx %d: rc = %d\n",
                                set->set_oi->oi_oa->o_id,
                                req->rq_oi.oi_oa->o_id, req->rq_idx, rc);
-                        if (!rc)
-                                rc = err;
+                        break;
                 }
         }
-        err = lov_fini_sync_set(set);
-        if (!rc)
-                rc = err;
-        RETURN(rc);
+
+        /* If we are not waiting for responses on async requests, return. */
+        if (rc || cfs_list_empty(&rqset->set_requests)) {
+                int err = lov_fini_sync_set(set);
+
+                RETURN(rc ?: err);
+        }
+
+        LASSERT(rqset->set_interpret == NULL);
+        rqset->set_interpret = lov_sync_interpret;
+        rqset->set_arg = (void *)set;
+
+        RETURN(0);
 }
 
 static int lov_brw_check(struct lov_obd *lov, struct obd_info *lov_oinfo,
@@ -1629,7 +1647,7 @@ static int lov_brw_check(struct lov_obd *lov, struct obd_info *lov_oinfo,
                 obd_off start, end;
 
                 if (!lov_stripe_intersects(lov_oinfo->oi_md, i, pga[i].off,
-                                           pga[i].off + pga[i].count,
+                                           pga[i].off + pga[i].count - 1,
                                            &start, &end))
                         continue;
 
@@ -1652,7 +1670,7 @@ static int lov_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
 {
         struct lov_request_set *set;
         struct lov_request *req;
-        struct list_head *pos;
+        cfs_list_t *pos;
         struct lov_obd *lov = &exp->exp_obd->u.lov;
         int err, rc = 0;
         ENTRY;
@@ -1668,10 +1686,10 @@ static int lov_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
         if (rc)
                 RETURN(rc);
 
-        list_for_each (pos, &set->set_list) {
+        cfs_list_for_each (pos, &set->set_list) {
                 struct obd_export *sub_exp;
                 struct brw_page *sub_pga;
-                req = list_entry(pos, struct lov_request, rq_link);
+                req = cfs_list_entry(pos, struct lov_request, rq_link);
 
                 sub_exp = lov->lov_tgts[req->rq_idx]->ltd_exp;
                 sub_pga = set->set_pga + req->rq_pgaidx;
@@ -1688,434 +1706,103 @@ static int lov_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
         RETURN(rc);
 }
 
-static int lov_brw_interpret(struct ptlrpc_request_set *reqset, void *data,
-                             int rc)
+static int lov_enqueue_interpret(struct ptlrpc_request_set *rqset,
+                                 void *data, int rc)
 {
         struct lov_request_set *lovset = (struct lov_request_set *)data;
         ENTRY;
-
-        if (rc) {
-                lovset->set_completes = 0;
-                lov_fini_brw_set(lovset);
-        } else {
-                rc = lov_fini_brw_set(lovset);
-        }
-
+        rc = lov_fini_enqueue_set(lovset, lovset->set_ei->ei_mode, rc, rqset);
         RETURN(rc);
 }
 
-static int lov_brw_async(int cmd, struct obd_export *exp,
-                         struct obd_info *oinfo, obd_count oa_bufs,
-                         struct brw_page *pga, struct obd_trans_info *oti,
-                         struct ptlrpc_request_set *set)
+static int lov_enqueue(struct obd_export *exp, struct obd_info *oinfo,
+                       struct ldlm_enqueue_info *einfo,
+                       struct ptlrpc_request_set *rqset)
 {
-        struct lov_request_set *lovset;
+        ldlm_mode_t mode = einfo->ei_mode;
+        struct lov_request_set *set;
         struct lov_request *req;
-        struct list_head *pos;
-        struct lov_obd *lov = &exp->exp_obd->u.lov;
-        int rc = 0;
+        cfs_list_t *pos;
+        struct lov_obd *lov;
+        ldlm_error_t rc;
         ENTRY;
 
         LASSERT(oinfo);
         ASSERT_LSM_MAGIC(oinfo->oi_md);
+        LASSERT(mode == (mode & -mode));
 
-        if (cmd == OBD_BRW_CHECK) {
-                rc = lov_brw_check(lov, oinfo, oa_bufs, pga);
-                RETURN(rc);
-        }
+        /* we should never be asked to replay a lock this way. */
+        LASSERT((oinfo->oi_flags & LDLM_FL_REPLAY) == 0);
 
-        rc = lov_prep_brw_set(exp, oinfo, oa_bufs, pga, oti, &lovset);
+        if (!exp || !exp->exp_obd)
+                RETURN(-ENODEV);
+
+        lov = &exp->exp_obd->u.lov;
+        rc = lov_prep_enqueue_set(exp, oinfo, einfo, &set);
         if (rc)
                 RETURN(rc);
 
-        list_for_each (pos, &lovset->set_list) {
-                struct obd_export *sub_exp;
-                struct brw_page *sub_pga;
-                req = list_entry(pos, struct lov_request, rq_link);
+        cfs_list_for_each (pos, &set->set_list) {
+                req = cfs_list_entry(pos, struct lov_request, rq_link);
 
-                sub_exp = lov->lov_tgts[req->rq_idx]->ltd_exp;
-                sub_pga = lovset->set_pga + req->rq_pgaidx;
-                rc = obd_brw_async(cmd, sub_exp, &req->rq_oi, req->rq_oabufs,
-                                   sub_pga, oti, set);
-                if (rc)
+                rc = obd_enqueue(lov->lov_tgts[req->rq_idx]->ltd_exp,
+                                 &req->rq_oi, einfo, rqset);
+                if (rc != ELDLM_OK)
                         GOTO(out, rc);
-                lov_update_common_set(lovset, req, rc);
         }
-        LASSERT(rc == 0);
-        LASSERT(set->set_interpret == NULL);
-        LASSERT(set->set_arg == NULL);
-        rc = ptlrpc_set_add_cb(set, lov_brw_interpret, lovset);
-        if (rc)
-                GOTO(out, rc);
 
-        RETURN(rc);
+        if (rqset && !cfs_list_empty(&rqset->set_requests)) {
+                LASSERT(rc == 0);
+                LASSERT(rqset->set_interpret == NULL);
+                rqset->set_interpret = lov_enqueue_interpret;
+                rqset->set_arg = (void *)set;
+                RETURN(rc);
+        }
 out:
-        lov_fini_brw_set(lovset);
+        rc = lov_fini_enqueue_set(set, mode, rc, rqset);
         RETURN(rc);
 }
 
-static int lov_ap_make_ready(void *data, int cmd)
+static int lov_change_cbdata(struct obd_export *exp,
+                             struct lov_stripe_md *lsm, ldlm_iterator_t it,
+                             void *data)
 {
-        struct lov_async_page *lap = LAP_FROM_COOKIE(data);
-
-        return lap->lap_caller_ops->ap_make_ready(lap->lap_caller_data, cmd);
-}
-
-static int lov_ap_refresh_count(void *data, int cmd)
-{
-        struct lov_async_page *lap = LAP_FROM_COOKIE(data);
-
-        return lap->lap_caller_ops->ap_refresh_count(lap->lap_caller_data,
-                                                     cmd);
-}
-
-static void lov_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
-{
-        struct lov_async_page *lap = LAP_FROM_COOKIE(data);
-
-        lap->lap_caller_ops->ap_fill_obdo(lap->lap_caller_data, cmd, oa);
-        /* XXX woah, shouldn't we be altering more here?  size? */
-        oa->o_id = lap->lap_loi_id;
-        oa->o_gr = lap->lap_loi_gr;
-        oa->o_valid |= OBD_MD_FLGROUP;
-        oa->o_stripe_idx = lap->lap_stripe;
-}
-
-static void lov_ap_update_obdo(void *data, int cmd, struct obdo *oa,
-                               obd_valid valid)
-{
-        struct lov_async_page *lap = LAP_FROM_COOKIE(data);
-
-        lap->lap_caller_ops->ap_update_obdo(lap->lap_caller_data, cmd,oa,valid);
-}
-
-static int lov_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
-{
-        struct lov_async_page *lap = LAP_FROM_COOKIE(data);
-
-        /* in a raid1 regime this would down a count of many ios
-         * in flight, onl calling the caller_ops completion when all
-         * the raid1 ios are complete */
-        rc = lap->lap_caller_ops->ap_completion(lap->lap_caller_data,cmd,oa,rc);
-        return rc;
-}
-
-static struct obd_capa *lov_ap_lookup_capa(void *data, int cmd)
-{
-        struct lov_async_page *lap = LAP_FROM_COOKIE(data);
-        return lap->lap_caller_ops->ap_lookup_capa(lap->lap_caller_data, cmd);
-}
-
-static struct obd_async_page_ops lov_async_page_ops = {
-        .ap_make_ready =        lov_ap_make_ready,
-        .ap_refresh_count =     lov_ap_refresh_count,
-        .ap_fill_obdo =         lov_ap_fill_obdo,
-        .ap_update_obdo =       lov_ap_update_obdo,
-        .ap_completion =        lov_ap_completion,
-        .ap_lookup_capa =       lov_ap_lookup_capa,
-};
-
-int lov_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
-                           struct lov_oinfo *loi, cfs_page_t *page,
-                           obd_off offset, struct obd_async_page_ops *ops,
-                           void *data, void **res, int nocache,
-                           struct lustre_handle *lockh)
-{
-        struct lov_obd *lov = &exp->exp_obd->u.lov;
-        struct lov_async_page *lap;
-        struct lov_lock_handles *lov_lockh = NULL;
-        int rc = 0;
-        ENTRY;
-
-        if (!page) {
-                int i = 0;
-                /* Find an existing osc so we can get it's stupid sizeof(*oap).
-                   Only because of this layering limitation will a client
-                   mount with no osts fail */
-                while (!lov->lov_tgts || !lov->lov_tgts[i] ||
-                       !lov->lov_tgts[i]->ltd_exp) {
-                        i++;
-                        if (i >= lov->desc.ld_tgt_count)
-                                RETURN(-ENOMEDIUM);
-                }
-                rc = size_round(sizeof(*lap)) +
-                        obd_prep_async_page(lov->lov_tgts[i]->ltd_exp, NULL,
-                                            NULL, NULL, 0, NULL, NULL, NULL, 0,
-                                            NULL);
-                RETURN(rc);
-        }
-        ASSERT_LSM_MAGIC(lsm);
-        LASSERT(loi == NULL);
-
-        lap = *res;
-        lap->lap_magic = LOV_AP_MAGIC;
-        lap->lap_caller_ops = ops;
-        lap->lap_caller_data = data;
-
-        /* for now only raid 0 which passes through */
-        lap->lap_stripe = lov_stripe_number(lsm, offset);
-        lov_stripe_offset(lsm, offset, lap->lap_stripe, &lap->lap_sub_offset);
-        loi = lsm->lsm_oinfo[lap->lap_stripe];
-
-        /* so the callback doesn't need the lsm */
-        lap->lap_loi_id = loi->loi_id;
-        lap->lap_loi_gr = lsm->lsm_object_gr;
-        LASSERT(lsm->lsm_object_gr > 0);
-        
-        lap->lap_sub_cookie = (void *)lap + size_round(sizeof(*lap));
-
-        if (lockh) {
-                lov_lockh = lov_handle2llh(lockh);
-                if (lov_lockh) {
-                        lockh = lov_lockh->llh_handles + lap->lap_stripe;
-                }
-        }
-
-        rc = obd_prep_async_page(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp,
-                                 lsm, loi, page, lap->lap_sub_offset,
-                                 &lov_async_page_ops, lap,
-                                 &lap->lap_sub_cookie, nocache, lockh);
-        if (lov_lockh)
-                lov_llh_put(lov_lockh);
-        if (rc)
-                RETURN(rc);
-        CDEBUG(D_CACHE, "lap %p page %p cookie %p off "LPU64"\n", lap, page,
-               lap->lap_sub_cookie, offset);
-        RETURN(0);
-}
-
-static int lov_queue_async_io(struct obd_export *exp,
-                              struct lov_stripe_md *lsm,
-                              struct lov_oinfo *loi, void *cookie,
-                              int cmd, obd_off off, int count,
-                              obd_flag brw_flags, obd_flag async_flags)
-{
-        struct lov_obd *lov = &exp->exp_obd->u.lov;
-        struct lov_async_page *lap;
-        int rc;
-
-        LASSERT(loi == NULL);
-
-        ASSERT_LSM_MAGIC(lsm);
-
-        lap = LAP_FROM_COOKIE(cookie);
-
-        loi = lsm->lsm_oinfo[lap->lap_stripe];
-
-        rc = obd_queue_async_io(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp, lsm,
-                                loi, lap->lap_sub_cookie, cmd, off, count,
-                                brw_flags, async_flags);
-        RETURN(rc);
-}
-
-static int lov_set_async_flags(struct obd_export *exp,
-                               struct lov_stripe_md *lsm,
-                               struct lov_oinfo *loi, void *cookie,
-                               obd_flag async_flags)
-{
-        struct lov_obd *lov = &exp->exp_obd->u.lov;
-        struct lov_async_page *lap;
-        int rc;
-
-        LASSERT(loi == NULL);
-
-        ASSERT_LSM_MAGIC(lsm);
-
-        lap = LAP_FROM_COOKIE(cookie);
-
-        loi = lsm->lsm_oinfo[lap->lap_stripe];
-
-        rc = obd_set_async_flags(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp,
-                                 lsm, loi, lap->lap_sub_cookie, async_flags);
-        RETURN(rc);
-}
-
-static int lov_queue_group_io(struct obd_export *exp,
-                              struct lov_stripe_md *lsm,
-                              struct lov_oinfo *loi,
-                              struct obd_io_group *oig, void *cookie,
-                              int cmd, obd_off off, int count,
-                              obd_flag brw_flags, obd_flag async_flags)
-{
-        struct lov_obd *lov = &exp->exp_obd->u.lov;
-        struct lov_async_page *lap;
-        int rc;
-
-        LASSERT(loi == NULL);
-
-        ASSERT_LSM_MAGIC(lsm);
-
-        lap = LAP_FROM_COOKIE(cookie);
-
-        loi = lsm->lsm_oinfo[lap->lap_stripe];
-
-        rc = obd_queue_group_io(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp, lsm,
-                                loi, oig, lap->lap_sub_cookie, cmd, off, count,
-                                brw_flags, async_flags);
-        RETURN(rc);
-}
-
-/* this isn't exactly optimal.  we may have queued sync io in oscs on
- * all stripes, but we don't record that fact at queue time.  so we
- * trigger sync io on all stripes. */
-static int lov_trigger_group_io(struct obd_export *exp,
-                                struct lov_stripe_md *lsm,
-                                struct lov_oinfo *loi,
-                                struct obd_io_group *oig)
-{
-        struct lov_obd *lov = &exp->exp_obd->u.lov;
-        int rc = 0, i, err;
-
-        LASSERT(loi == NULL);
-
-        ASSERT_LSM_MAGIC(lsm);
-
-        for (i = 0; i < lsm->lsm_stripe_count; i++) {
-                loi = lsm->lsm_oinfo[i];
-                if (!lov->lov_tgts[loi->loi_ost_idx] ||
-                    !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
-                        CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
-                        continue;
-                }
-
-                err = obd_trigger_group_io(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp,
-                                           lsm, loi, oig);
-                if (rc == 0 && err != 0)
-                        rc = err;
-        };
-        RETURN(rc);
-}
-
-static int lov_teardown_async_page(struct obd_export *exp,
-                                   struct lov_stripe_md *lsm,
-                                   struct lov_oinfo *loi, void *cookie)
-{
-        struct lov_obd *lov = &exp->exp_obd->u.lov;
-        struct lov_async_page *lap;
-        int rc;
-
-        LASSERT(loi == NULL);
-
-        ASSERT_LSM_MAGIC(lsm);
-
-        lap = LAP_FROM_COOKIE(cookie);
-
-        loi = lsm->lsm_oinfo[lap->lap_stripe];
-
-        rc = obd_teardown_async_page(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp,
-                                     lsm, loi, lap->lap_sub_cookie);
-        if (rc) {
-                CERROR("unable to teardown sub cookie %p: %d\n",
-                       lap->lap_sub_cookie, rc);
-                RETURN(rc);
-        }
-        RETURN(rc);
-}
-
-static int lov_enqueue_interpret(struct ptlrpc_request_set *rqset,
-                                 void *data, int rc)
-{
-        struct lov_request_set *lovset = (struct lov_request_set *)data;
-        ENTRY;
-        rc = lov_fini_enqueue_set(lovset, lovset->set_ei->ei_mode, rc, rqset);
-        RETURN(rc);
-}
-
-static int lov_enqueue(struct obd_export *exp, struct obd_info *oinfo,
-                       struct ldlm_enqueue_info *einfo,
-                       struct ptlrpc_request_set *rqset)
-{
-        ldlm_mode_t mode = einfo->ei_mode;
-        struct lov_request_set *set;
-        struct lov_request *req;
-        struct list_head *pos;
         struct lov_obd *lov;
-        ldlm_error_t rc;
-        ENTRY;
-
-        LASSERT(oinfo);
-        ASSERT_LSM_MAGIC(oinfo->oi_md);
-        LASSERT(mode == (mode & -mode));
-
-        /* we should never be asked to replay a lock this way. */
-        LASSERT((oinfo->oi_flags & LDLM_FL_REPLAY) == 0);
-
-        if (!exp || !exp->exp_obd)
-                RETURN(-ENODEV);
-
-        lov = &exp->exp_obd->u.lov;
-        rc = lov_prep_enqueue_set(exp, oinfo, einfo, &set);
-        if (rc)
-                RETURN(rc);
-
-        list_for_each (pos, &set->set_list) {
-                req = list_entry(pos, struct lov_request, rq_link);
-
-                rc = obd_enqueue(lov->lov_tgts[req->rq_idx]->ltd_exp,
-                                 &req->rq_oi, einfo, rqset);
-                if (rc != ELDLM_OK)
-                        GOTO(out, rc);
-        }
-
-        if (rqset && !list_empty(&rqset->set_requests)) {
-                LASSERT(rc == 0);
-                LASSERT(rqset->set_interpret == NULL);
-                rqset->set_interpret = lov_enqueue_interpret;
-                rqset->set_arg = (void *)set;
-                RETURN(rc);
-        }
-out:
-        rc = lov_fini_enqueue_set(set, mode, rc, rqset);
-        RETURN(rc);
-}
-
-static int lov_match(struct obd_export *exp, struct lov_stripe_md *lsm,
-                     __u32 type, ldlm_policy_data_t *policy, __u32 mode,
-                     int *flags, void *data, struct lustre_handle *lockh)
-{
-        struct lov_request_set *set;
-        struct obd_info oinfo;
-        struct lov_request *req;
-        struct list_head *pos;
-        struct lov_obd *lov = &exp->exp_obd->u.lov;
-        struct lustre_handle *lov_lockhp;
-        int lov_flags, rc = 0;
+        int rc = 0, i;
         ENTRY;
 
         ASSERT_LSM_MAGIC(lsm);
-        LASSERT((*flags & LDLM_FL_TEST_LOCK) || mode == (mode & -mode));
 
         if (!exp || !exp->exp_obd)
                 RETURN(-ENODEV);
 
         lov = &exp->exp_obd->u.lov;
-        rc = lov_prep_match_set(exp, &oinfo, lsm, policy, mode, lockh, &set);
-        if (rc)
-                RETURN(rc);
-
-        list_for_each (pos, &set->set_list) {
-                ldlm_policy_data_t sub_policy;
-                req = list_entry(pos, struct lov_request, rq_link);
-                lov_lockhp = set->set_lockh->llh_handles + req->rq_stripe;
-                LASSERT(lov_lockhp);
+        for (i = 0; i < lsm->lsm_stripe_count; i++) {
+                struct lov_stripe_md submd;
+                struct lov_oinfo *loi = lsm->lsm_oinfo[i];
 
-                lov_flags = *flags;
-                sub_policy.l_extent = req->rq_oi.oi_policy.l_extent;
+                if (!lov->lov_tgts[loi->loi_ost_idx]) {
+                        CDEBUG(D_HA, "lov idx %d NULL \n", loi->loi_ost_idx);
+                        continue;
+                }
 
-                rc = obd_match(lov->lov_tgts[req->rq_idx]->ltd_exp,
-                               req->rq_oi.oi_md, type, &sub_policy,
-                               mode, &lov_flags, data, lov_lockhp);
-                rc = lov_update_match_set(set, req, rc);
-                if (rc <= 0)
-                        break;
+                LASSERT_SEQ_IS_MDT(loi->loi_seq);
+                submd.lsm_object_id = loi->loi_id;
+                submd.lsm_object_seq = loi->loi_seq;
+                submd.lsm_stripe_count = 0;
+                rc = obd_change_cbdata(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp,
+                                       &submd, it, data);
         }
-        lov_fini_match_set(set, mode, *flags);
         RETURN(rc);
 }
 
-static int lov_change_cbdata(struct obd_export *exp,
-                             struct lov_stripe_md *lsm, ldlm_iterator_t it,
-                             void *data)
+/* find any ldlm lock of the inode in lov
+ * return 0    not find
+ *        1    find one
+ *      < 0    error */
+static int lov_find_cbdata(struct obd_export *exp,
+                           struct lov_stripe_md *lsm, ldlm_iterator_t it,
+                           void *data)
 {
         struct lov_obd *lov;
         int rc = 0, i;
@@ -2126,8 +1813,6 @@ static int lov_change_cbdata(struct obd_export *exp,
         if (!exp || !exp->exp_obd)
                 RETURN(-ENODEV);
 
-        LASSERT(lsm->lsm_object_gr > 0);
-
         lov = &exp->exp_obd->u.lov;
         for (i = 0; i < lsm->lsm_stripe_count; i++) {
                 struct lov_stripe_md submd;
@@ -2137,12 +1822,15 @@ static int lov_change_cbdata(struct obd_export *exp,
                         CDEBUG(D_HA, "lov idx %d NULL \n", loi->loi_ost_idx);
                         continue;
                 }
-                
+
+                LASSERT_SEQ_IS_MDT(loi->loi_seq);
                 submd.lsm_object_id = loi->loi_id;
-                submd.lsm_object_gr = lsm->lsm_object_gr;
+                submd.lsm_object_seq = loi->loi_seq;
                 submd.lsm_stripe_count = 0;
-                rc = obd_change_cbdata(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp,
-                                       &submd, it, data);
+                rc = obd_find_cbdata(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp,
+                                     &submd, it, data);
+                if (rc != 0)
+                        RETURN(rc);
         }
         RETURN(rc);
 }
@@ -2153,8 +1841,8 @@ static int lov_cancel(struct obd_export *exp, struct lov_stripe_md *lsm,
         struct lov_request_set *set;
         struct obd_info oinfo;
         struct lov_request *req;
-        struct list_head *pos;
-        struct lov_obd *lov = &exp->exp_obd->u.lov;
+        cfs_list_t *pos;
+        struct lov_obd *lov;
         struct lustre_handle *lov_lockhp;
         int err = 0, rc = 0;
         ENTRY;
@@ -2164,15 +1852,15 @@ static int lov_cancel(struct obd_export *exp, struct lov_stripe_md *lsm,
         if (!exp || !exp->exp_obd)
                 RETURN(-ENODEV);
 
-        LASSERT(lsm->lsm_object_gr > 0);
+        LASSERT_SEQ_IS_MDT(lsm->lsm_object_seq);
         LASSERT(lockh);
         lov = &exp->exp_obd->u.lov;
         rc = lov_prep_cancel_set(exp, &oinfo, lsm, mode, lockh, &set);
         if (rc)
                 RETURN(rc);
 
-        list_for_each (pos, &set->set_list) {
-                req = list_entry(pos, struct lov_request, rq_link);
+        cfs_list_for_each (pos, &set->set_list) {
+                req = cfs_list_entry(pos, struct lov_request, rq_link);
                 lov_lockhp = set->set_lockh->llh_handles + req->rq_stripe;
 
                 rc = obd_cancel(lov->lov_tgts[req->rq_idx]->ltd_exp,
@@ -2194,7 +1882,7 @@ static int lov_cancel(struct obd_export *exp, struct lov_stripe_md *lsm,
 
 static int lov_cancel_unused(struct obd_export *exp,
                              struct lov_stripe_md *lsm,
-                             int flags, void *opaque)
+                             ldlm_cancel_flags_t flags, void *opaque)
 {
         struct lov_obd *lov;
         int rc = 0, i;
@@ -2220,7 +1908,7 @@ static int lov_cancel_unused(struct obd_export *exp,
 
         ASSERT_LSM_MAGIC(lsm);
 
-        LASSERT(lsm->lsm_object_gr > 0);
+        LASSERT_SEQ_IS_MDT(lsm->lsm_object_seq);
         for (i = 0; i < lsm->lsm_stripe_count; i++) {
                 struct lov_stripe_md submd;
                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
@@ -2235,7 +1923,7 @@ static int lov_cancel_unused(struct obd_export *exp,
                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
 
                 submd.lsm_object_id = loi->loi_id;
-                submd.lsm_object_gr = lsm->lsm_object_gr;
+                submd.lsm_object_seq = loi->loi_seq;
                 submd.lsm_stripe_count = 0;
                 err = obd_cancel_unused(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp,
                                         &submd, flags, opaque);
@@ -2250,50 +1938,7 @@ static int lov_cancel_unused(struct obd_export *exp,
         RETURN(rc);
 }
 
-static int lov_join_lru(struct obd_export *exp,
-                        struct lov_stripe_md *lsm, int join)
-{
-        struct lov_obd *lov;
-        int i, count = 0;
-        ENTRY;
-
-        ASSERT_LSM_MAGIC(lsm);
-        if (!exp || !exp->exp_obd)
-                RETURN(-ENODEV);
-
-        lov = &exp->exp_obd->u.lov;
-        for (i = 0; i < lsm->lsm_stripe_count; i++) {
-                struct lov_stripe_md submd;
-                struct lov_oinfo *loi = lsm->lsm_oinfo[i];
-                int rc = 0;
-
-                if (!lov->lov_tgts[loi->loi_ost_idx]) {
-                        CDEBUG(D_HA, "lov idx %d NULL\n", loi->loi_ost_idx);
-                        continue;
-                }
-
-                if (!lov->lov_tgts[loi->loi_ost_idx]->ltd_active)
-                        CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
-
-                submd.lsm_object_id = loi->loi_id;
-                submd.lsm_object_gr = lsm->lsm_object_gr;
-                submd.lsm_stripe_count = 0;
-                rc = obd_join_lru(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp,
-                                  &submd, join);
-                if (rc < 0) {
-                        CERROR("join lru failed. objid: "LPX64" subobj: "LPX64
-                               " ostidx: %d rc: %d\n", lsm->lsm_object_id,
-                               loi->loi_id, loi->loi_ost_idx, rc);
-                        return rc;
-                } else {
-                        count += rc;
-                }
-        }
-        RETURN(count);
-}
-
-static int lov_statfs_interpret(struct ptlrpc_request_set *rqset,
-                                void *data, int rc)
+int lov_statfs_interpret(struct ptlrpc_request_set *rqset, void *data, int rc)
 {
         struct lov_request_set *lovset = (struct lov_request_set *)data;
         int err;
@@ -2311,7 +1956,7 @@ static int lov_statfs_async(struct obd_device *obd, struct obd_info *oinfo,
 {
         struct lov_request_set *set;
         struct lov_request *req;
-        struct list_head *pos;
+        cfs_list_t *pos;
         struct lov_obd *lov;
         int rc = 0;
         ENTRY;
@@ -2324,10 +1969,10 @@ static int lov_statfs_async(struct obd_device *obd, struct obd_info *oinfo,
         if (rc)
                 RETURN(rc);
 
-        list_for_each (pos, &set->set_list) {
+        cfs_list_for_each (pos, &set->set_list) {
                 struct obd_device *osc_obd;
 
-                req = list_entry(pos, struct lov_request, rq_link);
+                req = cfs_list_entry(pos, struct lov_request, rq_link);
 
                 osc_obd = class_exp2obd(lov->lov_tgts[req->rq_idx]->ltd_exp);
                 rc = obd_statfs_async(osc_obd, &req->rq_oi, max_age, rqset);
@@ -2335,7 +1980,7 @@ static int lov_statfs_async(struct obd_device *obd, struct obd_info *oinfo,
                         break;
         }
 
-        if (rc || list_empty(&rqset->set_requests)) {
+        if (rc || cfs_list_empty(&rqset->set_requests)) {
                 int err;
                 if (rc)
                         set->set_completes = 0;
@@ -2379,7 +2024,7 @@ static int lov_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
 {
         struct obd_device *obddev = class_exp2obd(exp);
         struct lov_obd *lov = &obddev->u.lov;
-        int i, rc, count = lov->desc.ld_tgt_count;
+        int i = 0, rc = 0, count = lov->desc.ld_tgt_count;
         struct obd_uuid *uuidp;
         ENTRY;
 
@@ -2391,8 +2036,6 @@ static int lov_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
                 __u32 index;
 
                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
-                LASSERT(data->ioc_plen1 == sizeof(struct obd_statfs));
-
                 if ((index >= count))
                         RETURN(-ENODEV);
 
@@ -2406,16 +2049,22 @@ static int lov_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
                 if (!osc_obd)
                         RETURN(-EINVAL);
 
+                /* copy UUID */
+                if (cfs_copy_to_user(data->ioc_pbuf2, obd2cli_tgt(osc_obd),
+                                     min((int) data->ioc_plen2,
+                                         (int) sizeof(struct obd_uuid))))
+                        RETURN(-EFAULT);
+
                 /* got statfs data */
                 rc = obd_statfs(osc_obd, &stat_buf,
-                                cfs_time_current_64() - HZ, 0);
+                                cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
+                                0);
                 if (rc)
                         RETURN(rc);
-                if (copy_to_user(data->ioc_pbuf1, &stat_buf, data->ioc_plen1))
-                        RETURN(rc);
-                /* copy UUID */
-                rc = copy_to_user(data->ioc_pbuf2, obd2cli_tgt(osc_obd),
-                                  data->ioc_plen2);
+                if (cfs_copy_to_user(data->ioc_pbuf1, &stat_buf,
+                                     min((int) data->ioc_plen1,
+                                         (int) sizeof(stat_buf))))
+                        RETURN(-EFAULT);
                 break;
         }
         case OBD_IOC_LOV_GET_CONFIG: {
@@ -2458,14 +2107,13 @@ static int lov_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
                         *genp = lov->lov_tgts[i]->ltd_gen;
                 }
 
-                rc = copy_to_user((void *)uarg, buf, len);
-                if (rc)
+                if (cfs_copy_to_user((void *)uarg, buf, len))
                         rc = -EFAULT;
                 obd_ioctl_freedata(buf, len);
                 break;
         }
         case LL_IOC_LOV_SETSTRIPE:
-                rc = lov_setstripe(exp, karg, uarg);
+                rc = lov_setstripe(exp, len, karg, uarg);
                 break;
         case LL_IOC_LOV_GETSTRIPE:
                 rc = lov_getstripe(exp, karg, uarg);
@@ -2473,20 +2121,71 @@ static int lov_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
         case LL_IOC_LOV_SETEA:
                 rc = lov_setea(exp, karg, uarg);
                 break;
+        case OBD_IOC_QUOTACTL: {
+                struct if_quotactl *qctl = karg;
+                struct lov_tgt_desc *tgt = NULL;
+                struct obd_quotactl *oqctl;
+
+                if (qctl->qc_valid == QC_OSTIDX) {
+                        if (qctl->qc_idx < 0 || count <= qctl->qc_idx)
+                                RETURN(-EINVAL);
+
+                        tgt = lov->lov_tgts[qctl->qc_idx];
+                        if (!tgt || !tgt->ltd_exp)
+                                RETURN(-EINVAL);
+                } else if (qctl->qc_valid == QC_UUID) {
+                        for (i = 0; i < count; i++) {
+                                tgt = lov->lov_tgts[i];
+                                if (!tgt ||
+                                    !obd_uuid_equals(&tgt->ltd_uuid,
+                                                     &qctl->obd_uuid))
+                                        continue;
+
+                                if (tgt->ltd_exp == NULL)
+                                        RETURN(-EINVAL);
+
+                                break;
+                        }
+                } else {
+                        RETURN(-EINVAL);
+                }
+
+                if (i >= count)
+                        RETURN(-EAGAIN);
+
+                LASSERT(tgt && tgt->ltd_exp);
+                OBD_ALLOC_PTR(oqctl);
+                if (!oqctl)
+                        RETURN(-ENOMEM);
+
+                QCTL_COPY(oqctl, qctl);
+                rc = obd_quotactl(tgt->ltd_exp, oqctl);
+                if (rc == 0) {
+                        QCTL_COPY(qctl, oqctl);
+                        qctl->qc_valid = QC_OSTIDX;
+                        qctl->obd_uuid = tgt->ltd_uuid;
+                }
+                OBD_FREE_PTR(oqctl);
+                break;
+        }
         default: {
                 int set = 0;
 
                 if (count == 0)
                         RETURN(-ENOTTY);
 
-                rc = 0;
                 for (i = 0; i < count; i++) {
                         int err;
+                        struct obd_device *osc_obd;
 
                         /* OST was disconnected */
                         if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_exp)
                                 continue;
 
+                        /* ll_umount_begin() sets force flag but for lov, not
+                         * osc. Let's pass it through */
+                        osc_obd = class_exp2obd(lov->lov_tgts[i]->ltd_exp);
+                        osc_obd->obd_force = obddev->obd_force;
                         err = obd_iocontrol(cmd, lov->lov_tgts[i]->ltd_exp,
                                             len, karg, uarg);
                         if (err == -ENODATA && cmd == OBD_IOC_POLL_QUOTACHECK) {
@@ -2672,7 +2371,7 @@ static int lov_fiemap(struct lov_obd *lov, __u32 keylen, void *key,
         if (fiemap_count_to_size(fm_key->fiemap.fm_extent_count) < buffer_size)
                 buffer_size = fiemap_count_to_size(fm_key->fiemap.fm_extent_count);
 
-        OBD_ALLOC(fm_local, buffer_size);
+        OBD_ALLOC_LARGE(fm_local, buffer_size);
         if (fm_local == NULL)
                 GOTO(out, rc = -ENOMEM);
         lcl_fm_ext = &fm_local->fm_extents[0];
@@ -2760,13 +2459,14 @@ static int lov_fiemap(struct lov_obd *lov, __u32 keylen, void *key,
                         fm_local->fm_flags = fiemap->fm_flags;
 
                         fm_key->oa.o_id = lsm->lsm_oinfo[cur_stripe]->loi_id;
+                        fm_key->oa.o_seq = lsm->lsm_oinfo[cur_stripe]->loi_seq;
                         ost_index = lsm->lsm_oinfo[cur_stripe]->loi_ost_idx;
 
                         if (ost_index < 0 || ost_index >=lov->desc.ld_tgt_count)
                                 GOTO(out, rc = -EINVAL);
 
                         /* If OST is inactive, return extent with UNKNOWN flag */
-                        if (lov && !lov->lov_tgts[ost_index]->ltd_active) {
+                        if (!lov->lov_tgts[ost_index]->ltd_active) {
                                 fm_local->fm_flags |= FIEMAP_EXTENT_LAST;
                                 fm_local->fm_mapped_extents = 1;
 
@@ -2862,7 +2562,7 @@ skip_last_device_calc:
         fiemap->fm_mapped_extents = current_extent;
 
 out:
-        OBD_FREE(fm_local, buffer_size);
+        OBD_FREE_LARGE(fm_local, buffer_size);
         return rc;
 }
 
@@ -2878,7 +2578,7 @@ static int lov_get_info(struct obd_export *exp, __u32 keylen,
         if (!vallen || !val)
                 RETURN(-EFAULT);
 
-        lov_getref(obddev);
+        obd_getref(obddev);
 
         if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
                 struct {
@@ -2905,7 +2605,7 @@ static int lov_get_info(struct obd_export *exp, __u32 keylen,
                                 continue;
                         if (lov->lov_tgts[loi->loi_ost_idx]->ltd_exp ==
                             data->lock->l_conn_export &&
-                            osc_res_name_eq(loi->loi_id, loi->loi_gr, res_id)) {
+                            osc_res_name_eq(loi->loi_id, loi->loi_seq, res_id)) {
                                 *stripe = i;
                                 GOTO(out, rc = 0);
                         }
@@ -2934,12 +2634,25 @@ static int lov_get_info(struct obd_export *exp, __u32 keylen,
         } else if (KEY_IS(KEY_FIEMAP)) {
                 rc = lov_fiemap(lov, keylen, key, vallen, val, lsm);
                 GOTO(out, rc);
+        } else if (KEY_IS(KEY_CONNECT_FLAG)) {
+                struct lov_tgt_desc *tgt;
+                __u64 ost_idx = *((__u64*)val);
+
+                LASSERT(*vallen == sizeof(__u64));
+                LASSERT(ost_idx < lov->desc.ld_tgt_count);
+                tgt = lov->lov_tgts[ost_idx];
+
+                if (!tgt || !tgt->ltd_exp)
+                        GOTO(out, rc = -ESRCH);
+
+                *((__u64*)val) = tgt->ltd_exp->exp_connect_flags;
+                GOTO(out, rc = 0);
         }
 
         rc = -EINVAL;
 
 out:
-        lov_putref(obddev);
+        obd_putref(obddev);
         RETURN(rc);
 }
 
@@ -2954,7 +2667,7 @@ static int lov_set_info_async(struct obd_export *exp, obd_count keylen,
         struct lov_tgt_desc *tgt;
         unsigned incr, check_uuid,
                  do_inactive, no_set;
-        unsigned next_id = 0,  mds_con = 0;
+        unsigned next_id = 0,  mds_con = 0, capa = 0;
         ENTRY;
 
         incr = check_uuid = do_inactive = no_set = 0;
@@ -2965,7 +2678,7 @@ static int lov_set_info_async(struct obd_export *exp, obd_count keylen,
                         RETURN(-ENOMEM);
         }
 
-        lov_getref(obddev);
+        obd_getref(obddev);
         count = lov->desc.ld_tgt_count;
 
         if (KEY_IS(KEY_NEXT_ID)) {
@@ -2976,12 +2689,12 @@ static int lov_set_info_async(struct obd_export *exp, obd_count keylen,
                 next_id = 1;
         } else if (KEY_IS(KEY_CHECKSUM)) {
                 do_inactive = 1;
-        } else if (KEY_IS(KEY_UNLINKED)) {
-                check_uuid = val ? 1 : 0;
         } else if (KEY_IS(KEY_EVICT_BY_NID)) {
                 /* use defaults:  do_inactive = incr = 0; */
         } else if (KEY_IS(KEY_MDS_CONN)) {
                 mds_con = 1;
+        } else if (KEY_IS(KEY_CAPA_KEY)) {
+                capa = 1;
         }
 
         for (i = 0; i < count; i++, val = (char *)val + incr) {
@@ -3016,7 +2729,21 @@ static int lov_set_info_async(struct obd_export *exp, obd_count keylen,
                         err = obd_set_info_async(tgt->ltd_exp,
                                          keylen, key, vallen,
                                          ((struct obd_id_info*)val)->data, set);
-                } else  {
+                } else if (capa) {
+                        struct mds_capa_info *info = (struct mds_capa_info*)val;
+
+                        LASSERT(vallen == sizeof(*info));
+
+                         /* Only want a specific OSC */
+                        if (info->uuid &&
+                            !obd_uuid_equals(info->uuid, &tgt->ltd_uuid))
+                                continue;
+
+                        err = obd_set_info_async(tgt->ltd_exp, keylen, key,
+                                                 sizeof(*info->capa),
+                                                 info->capa, set);
+                       
+                } else {
                         /* Only want a specific OSC */
                         if (check_uuid &&
                             !obd_uuid_equals(val, &tgt->ltd_uuid))
@@ -3030,7 +2757,7 @@ static int lov_set_info_async(struct obd_export *exp, obd_count keylen,
                         rc = err;
         }
 
-        lov_putref(obddev);
+        obd_putref(obddev);
         if (no_set) {
                 err = ptlrpc_set_wait(set);
                 if (!rc)
@@ -3040,21 +2767,6 @@ static int lov_set_info_async(struct obd_export *exp, obd_count keylen,
         RETURN(rc);
 }
 
-static int lov_checkmd(struct obd_export *exp, struct obd_export *md_exp,
-                       struct lov_stripe_md *lsm)
-{
-        int rc;
-        ENTRY;
-
-        if (!lsm)
-                RETURN(0);
-        LASSERT(md_exp);
-        LASSERT(lsm_op_find(lsm->lsm_magic) != NULL);
-        rc = lsm_op_find(lsm->lsm_magic)->lsm_revalidate(lsm, md_exp->exp_obd);
-
-        RETURN(rc);
-}
-
 int lov_test_and_clear_async_rc(struct lov_stripe_md *lsm)
 {
         int i, rc = 0;
@@ -3095,99 +2807,10 @@ static int lov_extent_calc(struct obd_export *exp, struct lov_stripe_md *lsm,
         RETURN(0);
 }
 
-
-#if 0
-struct lov_multi_wait {
-        struct ldlm_lock *lock;
-        wait_queue_t      wait;
-        int               completed;
-        int               generation;
-};
-
-int lov_complete_many(struct obd_export *exp, struct lov_stripe_md *lsm,
-                      struct lustre_handle *lockh)
-{
-        struct lov_lock_handles *lov_lockh = NULL;
-        struct lustre_handle *lov_lockhp;
-        struct lov_obd *lov;
-        struct lov_oinfo *loi;
-        struct lov_multi_wait *queues;
-        int rc = 0, i;
-        ENTRY;
-
-        ASSERT_LSM_MAGIC(lsm);
-
-        if (!exp || !exp->exp_obd)
-                RETURN(-ENODEV);
-
-        LASSERT(lockh != NULL);
-        if (lsm->lsm_stripe_count > 1) {
-                lov_lockh = lov_handle2llh(lockh);
-                if (lov_lockh == NULL) {
-                        CERROR("LOV: invalid lov lock handle %p\n", lockh);
-                        RETURN(-EINVAL);
-                }
-
-                lov_lockhp = lov_lockh->llh_handles;
-        } else {
-                lov_lockhp = lockh;
-        }
-
-        OBD_ALLOC(queues, lsm->lsm_stripe_count * sizeof(*queues));
-        if (queues == NULL)
-                GOTO(out, rc = -ENOMEM);
-
-        lov = &exp->exp_obd->u.lov;
-        for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
-             i++, loi++, lov_lockhp++) {
-                struct ldlm_lock *lock;
-                struct obd_device *obd;
-
-                lock = ldlm_handle2lock(lov_lockhp);
-                if (lock == NULL) {
-                        CDEBUG(D_HA, "lov idx %d subobj "LPX64" no lock?\n",
-                               loi->loi_ost_idx, loi->loi_id);
-                        queues[i].completed = 1;
-                        continue;
-                }
-
-                queues[i].lock = lock;
-                init_waitqueue_entry(&(queues[i].wait), current);
-                add_wait_queue(lock->l_waitq, &(queues[i].wait));
-
-                obd = class_exp2obd(lock->l_conn_export);
-                if (obd != NULL)
-                        imp = obd->u.cli.cl_import;
-                if (imp != NULL) {
-                        spin_lock(&imp->imp_lock);
-                        queues[i].generation = imp->imp_generation;
-                        spin_unlock(&imp->imp_lock);
-                }
-        }
-
-        lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ, ldlm_expired_completion_wait,
-                               interrupted_completion_wait, &lwd);
-        rc = l_wait_event_added(check_multi_complete(queues, lsm), &lwi);
-
-        for (i = 0; i < lsm->lsm_stripe_count; i++)
-                remove_wait_queue(lock->l_waitq, &(queues[i].wait));
-
-        if (rc == -EINTR || rc == -ETIMEDOUT) {
-
-
-        }
-
- out:
-        if (lov_lockh != NULL)
-                lov_llh_put(lov_lockh);
-        RETURN(rc);
-}
-#endif
-
 void lov_stripe_lock(struct lov_stripe_md *md)
 {
         LASSERT(md->lsm_lock_owner != cfs_curproc_pid());
-        spin_lock(&md->lsm_lock);
+        cfs_spin_lock(&md->lsm_lock);
         LASSERT(md->lsm_lock_owner == 0);
         md->lsm_lock_owner = cfs_curproc_pid();
 }
@@ -3197,104 +2820,23 @@ void lov_stripe_unlock(struct lov_stripe_md *md)
 {
         LASSERT(md->lsm_lock_owner == cfs_curproc_pid());
         md->lsm_lock_owner = 0;
-        spin_unlock(&md->lsm_lock);
+        cfs_spin_unlock(&md->lsm_lock);
 }
 EXPORT_SYMBOL(lov_stripe_unlock);
 
-/**
- * Checks if requested extent lock is compatible with a lock under the page.
- *
- * Checks if the lock under \a page is compatible with a read or write lock
- * (specified by \a rw) for an extent [\a start , \a end].
- *
- * \param exp lov export
- * \param lsm striping information for the file
- * \param res lov_async_page placeholder
- * \param rw OBD_BRW_READ if requested for reading,
- *           OBD_BRW_WRITE if requested for writing
- * \param start start of the requested extent
- * \param end end of the requested extent
- * \param cookie transparent parameter for passing locking context
- *
- * \post result == 1, *cookie == context, appropriate lock is referenced or
- * \post result == 0
- *
- * \retval 1 owned lock is reused for the request
- * \retval 0 no lock reused for the request
- *
- * \see lov_release_short_lock
- */
-static int lov_reget_short_lock(struct obd_export *exp,
-                                struct lov_stripe_md *lsm,
-                                void **res, int rw,
-                                obd_off start, obd_off end,
-                                void **cookie)
-{
-        struct lov_async_page *l = *res;
-        obd_off stripe_start, stripe_end = start;
-
-        ENTRY;
-
-        /* ensure we don't cross stripe boundaries */
-        lov_extent_calc(exp, lsm, OBD_CALC_STRIPE_END, &stripe_end);
-        if (stripe_end <= end)
-                RETURN(0);
-
-        /* map the region limits to the object limits */
-        lov_stripe_offset(lsm, start, l->lap_stripe, &stripe_start);
-        lov_stripe_offset(lsm, end, l->lap_stripe, &stripe_end);
-
-        RETURN(obd_reget_short_lock(exp->exp_obd->u.lov.lov_tgts[lsm->
-                                    lsm_oinfo[l->lap_stripe]->loi_ost_idx]->
-                                    ltd_exp, NULL, &l->lap_sub_cookie,
-                                    rw, stripe_start, stripe_end, cookie));
-}
-
-/**
- * Releases a reference to a lock taken in a "fast" way.
- *
- * Releases a read or a write (specified by \a rw) lock
- * referenced by \a cookie.
- *
- * \param exp lov export
- * \param lsm striping information for the file
- * \param end end of the locked extent
- * \param rw OBD_BRW_READ if requested for reading,
- *           OBD_BRW_WRITE if requested for writing
- * \param cookie transparent parameter for passing locking context
- *
- * \post appropriate lock is dereferenced
- *
- * \see lov_reget_short_lock
- */
-static int lov_release_short_lock(struct obd_export *exp,
-                                  struct lov_stripe_md *lsm, obd_off end,
-                                  void *cookie, int rw)
-{
-        int stripe;
-
-        ENTRY;
-
-        stripe = lov_stripe_number(lsm, end);
-
-        RETURN(obd_release_short_lock(exp->exp_obd->u.lov.lov_tgts[lsm->
-                                      lsm_oinfo[stripe]->loi_ost_idx]->
-                                      ltd_exp, NULL, end, cookie, rw));
-}
 
 struct obd_ops lov_obd_ops = {
         .o_owner               = THIS_MODULE,
         .o_setup               = lov_setup,
         .o_precleanup          = lov_precleanup,
         .o_cleanup             = lov_cleanup,
-        .o_process_config      = lov_process_config,
+        //.o_process_config      = lov_process_config,
         .o_connect             = lov_connect,
         .o_disconnect          = lov_disconnect,
         .o_statfs              = lov_statfs,
         .o_statfs_async        = lov_statfs_async,
         .o_packmd              = lov_packmd,
         .o_unpackmd            = lov_unpackmd,
-        .o_checkmd             = lov_checkmd,
         .o_create              = lov_create,
         .o_destroy             = lov_destroy,
         .o_getattr             = lov_getattr,
@@ -3302,25 +2844,15 @@ struct obd_ops lov_obd_ops = {
         .o_setattr             = lov_setattr,
         .o_setattr_async       = lov_setattr_async,
         .o_brw                 = lov_brw,
-        .o_brw_async           = lov_brw_async,
-        .o_prep_async_page     = lov_prep_async_page,
-        .o_reget_short_lock    = lov_reget_short_lock,
-        .o_release_short_lock  = lov_release_short_lock,
-        .o_queue_async_io      = lov_queue_async_io,
-        .o_set_async_flags     = lov_set_async_flags,
-        .o_queue_group_io      = lov_queue_group_io,
-        .o_trigger_group_io    = lov_trigger_group_io,
-        .o_teardown_async_page = lov_teardown_async_page,
         .o_merge_lvb           = lov_merge_lvb,
         .o_adjust_kms          = lov_adjust_kms,
         .o_punch               = lov_punch,
         .o_sync                = lov_sync,
         .o_enqueue             = lov_enqueue,
-        .o_match               = lov_match,
         .o_change_cbdata       = lov_change_cbdata,
+        .o_find_cbdata         = lov_find_cbdata,
         .o_cancel              = lov_cancel,
         .o_cancel_unused       = lov_cancel_unused,
-        .o_join_lru            = lov_join_lru,
         .o_iocontrol           = lov_iocontrol,
         .o_get_info            = lov_get_info,
         .o_set_info_async      = lov_set_info_async,
@@ -3328,10 +2860,12 @@ struct obd_ops lov_obd_ops = {
         .o_llog_init           = lov_llog_init,
         .o_llog_finish         = lov_llog_finish,
         .o_notify              = lov_notify,
-        .o_register_page_removal_cb = lov_register_page_removal_cb,
-        .o_unregister_page_removal_cb = lov_unregister_page_removal_cb,
-        .o_register_lock_cancel_cb = lov_register_lock_cancel_cb,
-        .o_unregister_lock_cancel_cb = lov_unregister_lock_cancel_cb,
+        .o_pool_new            = lov_pool_new,
+        .o_pool_rem            = lov_pool_remove,
+        .o_pool_add            = lov_pool_add,
+        .o_pool_del            = lov_pool_del,
+        .o_getref              = lov_getref,
+        .o_putref              = lov_putref,
 };
 
 static quota_interface_t *quota_interface;
@@ -3339,30 +2873,45 @@ extern quota_interface_t lov_quota_interface;
 
 cfs_mem_cache_t *lov_oinfo_slab;
 
+extern struct lu_kmem_descr lov_caches[];
+
 int __init lov_init(void)
 {
         struct lprocfs_static_vars lvars = { 0 };
         int rc, rc2;
         ENTRY;
 
+        /* print an address of _any_ initialized kernel symbol from this
+         * module, to allow debugging with gdb that doesn't support data
+         * symbols from modules.*/
+        CDEBUG(D_CONSOLE, "Lustre LOV module (%p).\n", &lov_caches);
+
+        rc = lu_kmem_init(lov_caches);
+        if (rc)
+                return rc;
+
         lov_oinfo_slab = cfs_mem_cache_create("lov_oinfo",
-                                              sizeof(struct lov_oinfo), 
-                                              0, SLAB_HWCACHE_ALIGN);
-        if (lov_oinfo_slab == NULL)
+                                              sizeof(struct lov_oinfo),
+                                              0, CFS_SLAB_HWCACHE_ALIGN);
+        if (lov_oinfo_slab == NULL) {
+                lu_kmem_fini(lov_caches);
                 return -ENOMEM;
+        }
         lprocfs_lov_init_vars(&lvars);
 
-        request_module("lquota");
+        cfs_request_module("lquota");
         quota_interface = PORTAL_SYMBOL_GET(lov_quota_interface);
         init_obd_quota_ops(quota_interface, &lov_obd_ops);
 
         rc = class_register_type(&lov_obd_ops, NULL, lvars.module_vars,
-                                 LUSTRE_LOV_NAME, NULL);
+                                 LUSTRE_LOV_NAME, &lov_device_type);
+
         if (rc) {
                 if (quota_interface)
                         PORTAL_SYMBOL_PUT(lov_quota_interface);
                 rc2 = cfs_mem_cache_destroy(lov_oinfo_slab);
                 LASSERT(rc2 == 0);
+                lu_kmem_fini(lov_caches);
         }
 
         RETURN(rc);
@@ -3372,7 +2921,10 @@ int __init lov_init(void)
 static void /*__exit*/ lov_exit(void)
 {
         int rc;
-        
+
+        lu_device_type_fini(&lov_device_type);
+        lu_kmem_fini(lov_caches);
+
         if (quota_interface)
                 PORTAL_SYMBOL_PUT(lov_quota_interface);