This commit was manufactured by cvs2svn to create branch 'vserver'.
[linux-2.6.git] / fs / dlm / recover.c
diff --git a/fs/dlm/recover.c b/fs/dlm/recover.c
new file mode 100644 (file)
index 0000000..c2cc769
--- /dev/null
@@ -0,0 +1,768 @@
+/******************************************************************************
+*******************************************************************************
+**
+**  Copyright (C) Sistina Software, Inc.  1997-2003  All rights reserved.
+**  Copyright (C) 2004-2005 Red Hat, Inc.  All rights reserved.
+**
+**  This copyrighted material is made available to anyone wishing to use,
+**  modify, copy, or redistribute it subject to the terms and conditions
+**  of the GNU General Public License v.2.
+**
+*******************************************************************************
+******************************************************************************/
+
+#include "dlm_internal.h"
+#include "lockspace.h"
+#include "dir.h"
+#include "config.h"
+#include "ast.h"
+#include "memory.h"
+#include "rcom.h"
+#include "lock.h"
+#include "lowcomms.h"
+#include "member.h"
+#include "recover.h"
+
+
+/*
+ * Recovery waiting routines: these functions wait for a particular reply from
+ * a remote node, or for the remote node to report a certain status.  They need
+ * to abort if the lockspace is stopped indicating a node has failed (perhaps
+ * the one being waited for).
+ */
+
+/*
+ * Wait until given function returns non-zero or lockspace is stopped
+ * (LS_RECOVERY_STOP set due to failure of a node in ls_nodes).  When another
+ * function thinks it could have completed the waited-on task, they should wake
+ * up ls_wait_general to get an immediate response rather than waiting for the
+ * timer to detect the result.  A timer wakes us up periodically while waiting
+ * to see if we should abort due to a node failure.  This should only be called
+ * by the dlm_recoverd thread.
+ */
+
+static void dlm_wait_timer_fn(unsigned long data)
+{
+       struct dlm_ls *ls = (struct dlm_ls *) data;
+       mod_timer(&ls->ls_timer, jiffies + (dlm_config.ci_recover_timer * HZ));
+       wake_up(&ls->ls_wait_general);
+}
+
+int dlm_wait_function(struct dlm_ls *ls, int (*testfn) (struct dlm_ls *ls))
+{
+       int error = 0;
+
+       init_timer(&ls->ls_timer);
+       ls->ls_timer.function = dlm_wait_timer_fn;
+       ls->ls_timer.data = (long) ls;
+       ls->ls_timer.expires = jiffies + (dlm_config.ci_recover_timer * HZ);
+       add_timer(&ls->ls_timer);
+
+       wait_event(ls->ls_wait_general, testfn(ls) || dlm_recovery_stopped(ls));
+       del_timer_sync(&ls->ls_timer);
+
+       if (dlm_recovery_stopped(ls)) {
+               log_debug(ls, "dlm_wait_function aborted");
+               error = -EINTR;
+       }
+       return error;
+}
+
+/*
+ * An efficient way for all nodes to wait for all others to have a certain
+ * status.  The node with the lowest nodeid polls all the others for their
+ * status (wait_status_all) and all the others poll the node with the low id
+ * for its accumulated result (wait_status_low).  When all nodes have set
+ * status flag X, then status flag X_ALL will be set on the low nodeid.
+ */
+
+uint32_t dlm_recover_status(struct dlm_ls *ls)
+{
+       uint32_t status;
+       spin_lock(&ls->ls_recover_lock);
+       status = ls->ls_recover_status;
+       spin_unlock(&ls->ls_recover_lock);
+       return status;
+}
+
+void dlm_set_recover_status(struct dlm_ls *ls, uint32_t status)
+{
+       spin_lock(&ls->ls_recover_lock);
+       ls->ls_recover_status |= status;
+       spin_unlock(&ls->ls_recover_lock);
+}
+
+static int wait_status_all(struct dlm_ls *ls, uint32_t wait_status)
+{
+       struct dlm_rcom *rc = (struct dlm_rcom *) ls->ls_recover_buf;
+       struct dlm_member *memb;
+       int error = 0, delay;
+
+       list_for_each_entry(memb, &ls->ls_nodes, list) {
+               delay = 0;
+               for (;;) {
+                       if (dlm_recovery_stopped(ls)) {
+                               error = -EINTR;
+                               goto out;
+                       }
+
+                       error = dlm_rcom_status(ls, memb->nodeid);
+                       if (error)
+                               goto out;
+
+                       if (rc->rc_result & wait_status)
+                               break;
+                       if (delay < 1000)
+                               delay += 20;
+                       msleep(delay);
+               }
+       }
+ out:
+       return error;
+}
+
+static int wait_status_low(struct dlm_ls *ls, uint32_t wait_status)
+{
+       struct dlm_rcom *rc = (struct dlm_rcom *) ls->ls_recover_buf;
+       int error = 0, delay = 0, nodeid = ls->ls_low_nodeid;
+
+       for (;;) {
+               if (dlm_recovery_stopped(ls)) {
+                       error = -EINTR;
+                       goto out;
+               }
+
+               error = dlm_rcom_status(ls, nodeid);
+               if (error)
+                       break;
+
+               if (rc->rc_result & wait_status)
+                       break;
+               if (delay < 1000)
+                       delay += 20;
+               msleep(delay);
+       }
+ out:
+       return error;
+}
+
+static int wait_status(struct dlm_ls *ls, uint32_t status)
+{
+       uint32_t status_all = status << 1;
+       int error;
+
+       if (ls->ls_low_nodeid == dlm_our_nodeid()) {
+               error = wait_status_all(ls, status);
+               if (!error)
+                       dlm_set_recover_status(ls, status_all);
+       } else
+               error = wait_status_low(ls, status_all);
+
+       return error;
+}
+
+int dlm_recover_members_wait(struct dlm_ls *ls)
+{
+       return wait_status(ls, DLM_RS_NODES);
+}
+
+int dlm_recover_directory_wait(struct dlm_ls *ls)
+{
+       return wait_status(ls, DLM_RS_DIR);
+}
+
+int dlm_recover_locks_wait(struct dlm_ls *ls)
+{
+       return wait_status(ls, DLM_RS_LOCKS);
+}
+
+int dlm_recover_done_wait(struct dlm_ls *ls)
+{
+       return wait_status(ls, DLM_RS_DONE);
+}
+
+/*
+ * The recover_list contains all the rsb's for which we've requested the new
+ * master nodeid.  As replies are returned from the resource directories the
+ * rsb's are removed from the list.  When the list is empty we're done.
+ *
+ * The recover_list is later similarly used for all rsb's for which we've sent
+ * new lkb's and need to receive new corresponding lkid's.
+ *
+ * We use the address of the rsb struct as a simple local identifier for the
+ * rsb so we can match an rcom reply with the rsb it was sent for.
+ */
+
+static int recover_list_empty(struct dlm_ls *ls)
+{
+       int empty;
+
+       spin_lock(&ls->ls_recover_list_lock);
+       empty = list_empty(&ls->ls_recover_list);
+       spin_unlock(&ls->ls_recover_list_lock);
+
+       return empty;
+}
+
+static void recover_list_add(struct dlm_rsb *r)
+{
+       struct dlm_ls *ls = r->res_ls;
+
+       spin_lock(&ls->ls_recover_list_lock);
+       if (list_empty(&r->res_recover_list)) {
+               list_add_tail(&r->res_recover_list, &ls->ls_recover_list);
+               ls->ls_recover_list_count++;
+               dlm_hold_rsb(r);
+       }
+       spin_unlock(&ls->ls_recover_list_lock);
+}
+
+static void recover_list_del(struct dlm_rsb *r)
+{
+       struct dlm_ls *ls = r->res_ls;
+
+       spin_lock(&ls->ls_recover_list_lock);
+       list_del_init(&r->res_recover_list);
+       ls->ls_recover_list_count--;
+       spin_unlock(&ls->ls_recover_list_lock);
+
+       dlm_put_rsb(r);
+}
+
+static struct dlm_rsb *recover_list_find(struct dlm_ls *ls, uint64_t id)
+{
+       struct dlm_rsb *r = NULL;
+
+       spin_lock(&ls->ls_recover_list_lock);
+
+       list_for_each_entry(r, &ls->ls_recover_list, res_recover_list) {
+               if (id == (unsigned long) r)
+                       goto out;
+       }
+       r = NULL;
+ out:
+       spin_unlock(&ls->ls_recover_list_lock);
+       return r;
+}
+
+static void recover_list_clear(struct dlm_ls *ls)
+{
+       struct dlm_rsb *r, *s;
+
+       spin_lock(&ls->ls_recover_list_lock);
+       list_for_each_entry_safe(r, s, &ls->ls_recover_list, res_recover_list) {
+               list_del_init(&r->res_recover_list);
+               r->res_recover_locks_count = 0;
+               dlm_put_rsb(r);
+               ls->ls_recover_list_count--;
+       }
+
+       if (ls->ls_recover_list_count != 0) {
+               log_error(ls, "warning: recover_list_count %d",
+                         ls->ls_recover_list_count);
+               ls->ls_recover_list_count = 0;
+       }
+       spin_unlock(&ls->ls_recover_list_lock);
+}
+
+
+/* Master recovery: find new master node for rsb's that were
+   mastered on nodes that have been removed.
+
+   dlm_recover_masters
+   recover_master
+   dlm_send_rcom_lookup            ->  receive_rcom_lookup
+                                       dlm_dir_lookup
+   receive_rcom_lookup_reply       <-
+   dlm_recover_master_reply
+   set_new_master
+   set_master_lkbs
+   set_lock_master
+*/
+
+/*
+ * Set the lock master for all LKBs in a lock queue
+ * If we are the new master of the rsb, we may have received new
+ * MSTCPY locks from other nodes already which we need to ignore
+ * when setting the new nodeid.
+ */
+
+static void set_lock_master(struct list_head *queue, int nodeid)
+{
+       struct dlm_lkb *lkb;
+
+       list_for_each_entry(lkb, queue, lkb_statequeue)
+               if (!(lkb->lkb_flags & DLM_IFL_MSTCPY))
+                       lkb->lkb_nodeid = nodeid;
+}
+
+static void set_master_lkbs(struct dlm_rsb *r)
+{
+       set_lock_master(&r->res_grantqueue, r->res_nodeid);
+       set_lock_master(&r->res_convertqueue, r->res_nodeid);
+       set_lock_master(&r->res_waitqueue, r->res_nodeid);
+}
+
+/*
+ * Propogate the new master nodeid to locks
+ * The NEW_MASTER flag tells dlm_recover_locks() which rsb's to consider.
+ * The NEW_MASTER2 flag tells recover_lvb() and set_locks_purged() which
+ * rsb's to consider.
+ */
+
+static void set_new_master(struct dlm_rsb *r, int nodeid)
+{
+       lock_rsb(r);
+       r->res_nodeid = nodeid;
+       set_master_lkbs(r);
+       rsb_set_flag(r, RSB_NEW_MASTER);
+       rsb_set_flag(r, RSB_NEW_MASTER2);
+       unlock_rsb(r);
+}
+
+/*
+ * We do async lookups on rsb's that need new masters.  The rsb's
+ * waiting for a lookup reply are kept on the recover_list.
+ */
+
+static int recover_master(struct dlm_rsb *r)
+{
+       struct dlm_ls *ls = r->res_ls;
+       int error, dir_nodeid, ret_nodeid, our_nodeid = dlm_our_nodeid();
+
+       dir_nodeid = dlm_dir_nodeid(r);
+
+       if (dir_nodeid == our_nodeid) {
+               error = dlm_dir_lookup(ls, our_nodeid, r->res_name,
+                                      r->res_length, &ret_nodeid);
+               if (error)
+                       log_error(ls, "recover dir lookup error %d", error);
+
+               if (ret_nodeid == our_nodeid)
+                       ret_nodeid = 0;
+               set_new_master(r, ret_nodeid);
+       } else {
+               recover_list_add(r);
+               error = dlm_send_rcom_lookup(r, dir_nodeid);
+       }
+
+       return error;
+}
+
+/*
+ * When not using a directory, most resource names will hash to a new static
+ * master nodeid and the resource will need to be remastered.
+ */
+
+static int recover_master_static(struct dlm_rsb *r)
+{
+       int master = dlm_dir_nodeid(r);
+
+       if (master == dlm_our_nodeid())
+               master = 0;
+
+       if (r->res_nodeid != master) {
+               if (is_master(r))
+                       dlm_purge_mstcpy_locks(r);
+               set_new_master(r, master);
+               return 1;
+       }
+       return 0;
+}
+
+/*
+ * Go through local root resources and for each rsb which has a master which
+ * has departed, get the new master nodeid from the directory.  The dir will
+ * assign mastery to the first node to look up the new master.  That means
+ * we'll discover in this lookup if we're the new master of any rsb's.
+ *
+ * We fire off all the dir lookup requests individually and asynchronously to
+ * the correct dir node.
+ */
+
+int dlm_recover_masters(struct dlm_ls *ls)
+{
+       struct dlm_rsb *r;
+       int error = 0, count = 0;
+
+       log_debug(ls, "dlm_recover_masters");
+
+       down_read(&ls->ls_root_sem);
+       list_for_each_entry(r, &ls->ls_root_list, res_root_list) {
+               if (dlm_recovery_stopped(ls)) {
+                       up_read(&ls->ls_root_sem);
+                       error = -EINTR;
+                       goto out;
+               }
+
+               if (dlm_no_directory(ls))
+                       count += recover_master_static(r);
+               else if (!is_master(r) &&
+                        (dlm_is_removed(ls, r->res_nodeid) ||
+                         rsb_flag(r, RSB_NEW_MASTER))) {
+                       recover_master(r);
+                       count++;
+               }
+
+               schedule();
+       }
+       up_read(&ls->ls_root_sem);
+
+       log_debug(ls, "dlm_recover_masters %d resources", count);
+
+       error = dlm_wait_function(ls, &recover_list_empty);
+ out:
+       if (error)
+               recover_list_clear(ls);
+       return error;
+}
+
+int dlm_recover_master_reply(struct dlm_ls *ls, struct dlm_rcom *rc)
+{
+       struct dlm_rsb *r;
+       int nodeid;
+
+       r = recover_list_find(ls, rc->rc_id);
+       if (!r) {
+               log_error(ls, "dlm_recover_master_reply no id %llx",
+                         (unsigned long long)rc->rc_id);
+               goto out;
+       }
+
+       nodeid = rc->rc_result;
+       if (nodeid == dlm_our_nodeid())
+               nodeid = 0;
+
+       set_new_master(r, nodeid);
+       recover_list_del(r);
+
+       if (recover_list_empty(ls))
+               wake_up(&ls->ls_wait_general);
+ out:
+       return 0;
+}
+
+
+/* Lock recovery: rebuild the process-copy locks we hold on a
+   remastered rsb on the new rsb master.
+
+   dlm_recover_locks
+   recover_locks
+   recover_locks_queue
+   dlm_send_rcom_lock              ->  receive_rcom_lock
+                                       dlm_recover_master_copy
+   receive_rcom_lock_reply         <-
+   dlm_recover_process_copy
+*/
+
+
+/*
+ * keep a count of the number of lkb's we send to the new master; when we get
+ * an equal number of replies then recovery for the rsb is done
+ */
+
+static int recover_locks_queue(struct dlm_rsb *r, struct list_head *head)
+{
+       struct dlm_lkb *lkb;
+       int error = 0;
+
+       list_for_each_entry(lkb, head, lkb_statequeue) {
+               error = dlm_send_rcom_lock(r, lkb);
+               if (error)
+                       break;
+               r->res_recover_locks_count++;
+       }
+
+       return error;
+}
+
+static int recover_locks(struct dlm_rsb *r)
+{
+       int error = 0;
+
+       lock_rsb(r);
+
+       DLM_ASSERT(!r->res_recover_locks_count, dlm_dump_rsb(r););
+
+       error = recover_locks_queue(r, &r->res_grantqueue);
+       if (error)
+               goto out;
+       error = recover_locks_queue(r, &r->res_convertqueue);
+       if (error)
+               goto out;
+       error = recover_locks_queue(r, &r->res_waitqueue);
+       if (error)
+               goto out;
+
+       if (r->res_recover_locks_count)
+               recover_list_add(r);
+       else
+               rsb_clear_flag(r, RSB_NEW_MASTER);
+ out:
+       unlock_rsb(r);
+       return error;
+}
+
+int dlm_recover_locks(struct dlm_ls *ls)
+{
+       struct dlm_rsb *r;
+       int error, count = 0;
+
+       log_debug(ls, "dlm_recover_locks");
+
+       down_read(&ls->ls_root_sem);
+       list_for_each_entry(r, &ls->ls_root_list, res_root_list) {
+               if (is_master(r)) {
+                       rsb_clear_flag(r, RSB_NEW_MASTER);
+                       continue;
+               }
+
+               if (!rsb_flag(r, RSB_NEW_MASTER))
+                       continue;
+
+               if (dlm_recovery_stopped(ls)) {
+                       error = -EINTR;
+                       up_read(&ls->ls_root_sem);
+                       goto out;
+               }
+
+               error = recover_locks(r);
+               if (error) {
+                       up_read(&ls->ls_root_sem);
+                       goto out;
+               }
+
+               count += r->res_recover_locks_count;
+       }
+       up_read(&ls->ls_root_sem);
+
+       log_debug(ls, "dlm_recover_locks %d locks", count);
+
+       error = dlm_wait_function(ls, &recover_list_empty);
+ out:
+       if (error)
+               recover_list_clear(ls);
+       else
+               dlm_set_recover_status(ls, DLM_RS_LOCKS);
+       return error;
+}
+
+void dlm_recovered_lock(struct dlm_rsb *r)
+{
+       DLM_ASSERT(rsb_flag(r, RSB_NEW_MASTER), dlm_dump_rsb(r););
+
+       r->res_recover_locks_count--;
+       if (!r->res_recover_locks_count) {
+               rsb_clear_flag(r, RSB_NEW_MASTER);
+               recover_list_del(r);
+       }
+
+       if (recover_list_empty(r->res_ls))
+               wake_up(&r->res_ls->ls_wait_general);
+}
+
+/*
+ * The lvb needs to be recovered on all master rsb's.  This includes setting
+ * the VALNOTVALID flag if necessary, and determining the correct lvb contents
+ * based on the lvb's of the locks held on the rsb.
+ *
+ * RSB_VALNOTVALID is set if there are only NL/CR locks on the rsb.  If it
+ * was already set prior to recovery, it's not cleared, regardless of locks.
+ *
+ * The LVB contents are only considered for changing when this is a new master
+ * of the rsb (NEW_MASTER2).  Then, the rsb's lvb is taken from any lkb with
+ * mode > CR.  If no lkb's exist with mode above CR, the lvb contents are taken
+ * from the lkb with the largest lvb sequence number.
+ */
+
+static void recover_lvb(struct dlm_rsb *r)
+{
+       struct dlm_lkb *lkb, *high_lkb = NULL;
+       uint32_t high_seq = 0;
+       int lock_lvb_exists = 0;
+       int big_lock_exists = 0;
+       int lvblen = r->res_ls->ls_lvblen;
+
+       list_for_each_entry(lkb, &r->res_grantqueue, lkb_statequeue) {
+               if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
+                       continue;
+
+               lock_lvb_exists = 1;
+
+               if (lkb->lkb_grmode > DLM_LOCK_CR) {
+                       big_lock_exists = 1;
+                       goto setflag;
+               }
+
+               if (((int)lkb->lkb_lvbseq - (int)high_seq) >= 0) {
+                       high_lkb = lkb;
+                       high_seq = lkb->lkb_lvbseq;
+               }
+       }
+
+       list_for_each_entry(lkb, &r->res_convertqueue, lkb_statequeue) {
+               if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
+                       continue;
+
+               lock_lvb_exists = 1;
+
+               if (lkb->lkb_grmode > DLM_LOCK_CR) {
+                       big_lock_exists = 1;
+                       goto setflag;
+               }
+
+               if (((int)lkb->lkb_lvbseq - (int)high_seq) >= 0) {
+                       high_lkb = lkb;
+                       high_seq = lkb->lkb_lvbseq;
+               }
+       }
+
+ setflag:
+       if (!lock_lvb_exists)
+               goto out;
+
+       if (!big_lock_exists)
+               rsb_set_flag(r, RSB_VALNOTVALID);
+
+       /* don't mess with the lvb unless we're the new master */
+       if (!rsb_flag(r, RSB_NEW_MASTER2))
+               goto out;
+
+       if (!r->res_lvbptr) {
+               r->res_lvbptr = allocate_lvb(r->res_ls);
+               if (!r->res_lvbptr)
+                       goto out;
+       }
+
+       if (big_lock_exists) {
+               r->res_lvbseq = lkb->lkb_lvbseq;
+               memcpy(r->res_lvbptr, lkb->lkb_lvbptr, lvblen);
+       } else if (high_lkb) {
+               r->res_lvbseq = high_lkb->lkb_lvbseq;
+               memcpy(r->res_lvbptr, high_lkb->lkb_lvbptr, lvblen);
+       } else {
+               r->res_lvbseq = 0;
+               memset(r->res_lvbptr, 0, lvblen);
+       }
+ out:
+       return;
+}
+
+/* All master rsb's flagged RECOVER_CONVERT need to be looked at.  The locks
+   converting PR->CW or CW->PR need to have their lkb_grmode set. */
+
+static void recover_conversion(struct dlm_rsb *r)
+{
+       struct dlm_lkb *lkb;
+       int grmode = -1;
+
+       list_for_each_entry(lkb, &r->res_grantqueue, lkb_statequeue) {
+               if (lkb->lkb_grmode == DLM_LOCK_PR ||
+                   lkb->lkb_grmode == DLM_LOCK_CW) {
+                       grmode = lkb->lkb_grmode;
+                       break;
+               }
+       }
+
+       list_for_each_entry(lkb, &r->res_convertqueue, lkb_statequeue) {
+               if (lkb->lkb_grmode != DLM_LOCK_IV)
+                       continue;
+               if (grmode == -1)
+                       lkb->lkb_grmode = lkb->lkb_rqmode;
+               else
+                       lkb->lkb_grmode = grmode;
+       }
+}
+
+/* We've become the new master for this rsb and waiting/converting locks may
+   need to be granted in dlm_grant_after_purge() due to locks that may have
+   existed from a removed node. */
+
+static void set_locks_purged(struct dlm_rsb *r)
+{
+       if (!list_empty(&r->res_waitqueue) || !list_empty(&r->res_convertqueue))
+               rsb_set_flag(r, RSB_LOCKS_PURGED);
+}
+
+void dlm_recover_rsbs(struct dlm_ls *ls)
+{
+       struct dlm_rsb *r;
+       int count = 0;
+
+       log_debug(ls, "dlm_recover_rsbs");
+
+       down_read(&ls->ls_root_sem);
+       list_for_each_entry(r, &ls->ls_root_list, res_root_list) {
+               lock_rsb(r);
+               if (is_master(r)) {
+                       if (rsb_flag(r, RSB_RECOVER_CONVERT))
+                               recover_conversion(r);
+                       if (rsb_flag(r, RSB_NEW_MASTER2))
+                               set_locks_purged(r);
+                       recover_lvb(r);
+                       count++;
+               }
+               rsb_clear_flag(r, RSB_RECOVER_CONVERT);
+               rsb_clear_flag(r, RSB_NEW_MASTER2);
+               unlock_rsb(r);
+       }
+       up_read(&ls->ls_root_sem);
+
+       log_debug(ls, "dlm_recover_rsbs %d rsbs", count);
+}
+
+/* Create a single list of all root rsb's to be used during recovery */
+
+int dlm_create_root_list(struct dlm_ls *ls)
+{
+       struct dlm_rsb *r;
+       int i, error = 0;
+
+       down_write(&ls->ls_root_sem);
+       if (!list_empty(&ls->ls_root_list)) {
+               log_error(ls, "root list not empty");
+               error = -EINVAL;
+               goto out;
+       }
+
+       for (i = 0; i < ls->ls_rsbtbl_size; i++) {
+               read_lock(&ls->ls_rsbtbl[i].lock);
+               list_for_each_entry(r, &ls->ls_rsbtbl[i].list, res_hashchain) {
+                       list_add(&r->res_root_list, &ls->ls_root_list);
+                       dlm_hold_rsb(r);
+               }
+               read_unlock(&ls->ls_rsbtbl[i].lock);
+       }
+ out:
+       up_write(&ls->ls_root_sem);
+       return error;
+}
+
+void dlm_release_root_list(struct dlm_ls *ls)
+{
+       struct dlm_rsb *r, *safe;
+
+       down_write(&ls->ls_root_sem);
+       list_for_each_entry_safe(r, safe, &ls->ls_root_list, res_root_list) {
+               list_del_init(&r->res_root_list);
+               dlm_put_rsb(r);
+       }
+       up_write(&ls->ls_root_sem);
+}
+
+void dlm_clear_toss_list(struct dlm_ls *ls)
+{
+       struct dlm_rsb *r, *safe;
+       int i;
+
+       for (i = 0; i < ls->ls_rsbtbl_size; i++) {
+               write_lock(&ls->ls_rsbtbl[i].lock);
+               list_for_each_entry_safe(r, safe, &ls->ls_rsbtbl[i].toss,
+                                        res_hashchain) {
+                       list_del(&r->res_hashchain);
+                       free_rsb(r);
+               }
+               write_unlock(&ls->ls_rsbtbl[i].lock);
+       }
+}
+