aboutsummaryrefslogtreecommitdiff
path: root/fs/ocfs2/dlm/dlmrecovery.c
diff options
context:
space:
mode:
authorKurt Hackel <kurt.hackel@oracle.com>2005-12-15 14:31:23 -0800
committerJoel Becker <joel.becker@oracle.com>2006-01-03 11:45:47 -0800
commit6714d8e86bf443f6f7af50f9d432025649f091f5 (patch)
tree2c484bd1894a90cad7020869c7054f192d3bf34d /fs/ocfs2/dlm/dlmrecovery.c
parent98211489d4147e41b11703e4245846d60b3acce4 (diff)
[PATCH] OCFS2: The Second Oracle Cluster Filesystem
A distributed lock manager built with the cluster file system use case in mind. The OCFS2 dlm exposes a VMS style API, though things have been simplified internally. The only lock levels implemented currently are NLMODE, PRMODE and EXMODE. Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com> Signed-off-by: Kurt Hackel <kurt.hackel@oracle.com>
Diffstat (limited to 'fs/ocfs2/dlm/dlmrecovery.c')
-rw-r--r--fs/ocfs2/dlm/dlmrecovery.c2132
1 files changed, 2132 insertions, 0 deletions
diff --git a/fs/ocfs2/dlm/dlmrecovery.c b/fs/ocfs2/dlm/dlmrecovery.c
new file mode 100644
index 00000000000..0c8eb1093f0
--- /dev/null
+++ b/fs/ocfs2/dlm/dlmrecovery.c
@@ -0,0 +1,2132 @@
+/* -*- mode: c; c-basic-offset: 8; -*-
+ * vim: noexpandtab sw=8 ts=8 sts=0:
+ *
+ * dlmrecovery.c
+ *
+ * recovery stuff
+ *
+ * Copyright (C) 2004 Oracle. All rights reserved.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public
+ * License as published by the Free Software Foundation; either
+ * version 2 of the License, or (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public
+ * License along with this program; if not, write to the
+ * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+ * Boston, MA 021110-1307, USA.
+ *
+ */
+
+
+#include <linux/module.h>
+#include <linux/fs.h>
+#include <linux/types.h>
+#include <linux/slab.h>
+#include <linux/highmem.h>
+#include <linux/utsname.h>
+#include <linux/init.h>
+#include <linux/sysctl.h>
+#include <linux/random.h>
+#include <linux/blkdev.h>
+#include <linux/socket.h>
+#include <linux/inet.h>
+#include <linux/timer.h>
+#include <linux/kthread.h>
+
+
+#include "cluster/heartbeat.h"
+#include "cluster/nodemanager.h"
+#include "cluster/tcp.h"
+
+#include "dlmapi.h"
+#include "dlmcommon.h"
+#include "dlmdomain.h"
+
+#define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_RECOVERY)
+#include "cluster/masklog.h"
+
+static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node);
+
+static int dlm_recovery_thread(void *data);
+void dlm_complete_recovery_thread(struct dlm_ctxt *dlm);
+int dlm_launch_recovery_thread(struct dlm_ctxt *dlm);
+static void dlm_kick_recovery_thread(struct dlm_ctxt *dlm);
+static int dlm_do_recovery(struct dlm_ctxt *dlm);
+
+static int dlm_pick_recovery_master(struct dlm_ctxt *dlm);
+static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node);
+static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node);
+static int dlm_request_all_locks(struct dlm_ctxt *dlm,
+ u8 request_from, u8 dead_node);
+static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node);
+
+static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res);
+static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
+ const char *lockname, int namelen,
+ int total_locks, u64 cookie,
+ u8 flags, u8 master);
+static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
+ struct dlm_migratable_lockres *mres,
+ u8 send_to,
+ struct dlm_lock_resource *res,
+ int total_locks);
+static int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
+ struct dlm_lock_resource *res,
+ u8 *real_master);
+static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
+ struct dlm_lock_resource *res,
+ struct dlm_migratable_lockres *mres);
+static int dlm_do_master_requery(struct dlm_ctxt *dlm,
+ struct dlm_lock_resource *res,
+ u8 nodenum, u8 *real_master);
+static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm);
+static int dlm_send_all_done_msg(struct dlm_ctxt *dlm,
+ u8 dead_node, u8 send_to);
+static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node);
+static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
+ struct list_head *list, u8 dead_node);
+static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
+ u8 dead_node, u8 new_master);
+static void dlm_reco_ast(void *astdata);
+static void dlm_reco_bast(void *astdata, int blocked_type);
+static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st);
+static void dlm_request_all_locks_worker(struct dlm_work_item *item,
+ void *data);
+static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data);
+
+static u64 dlm_get_next_mig_cookie(void);
+
+static spinlock_t dlm_reco_state_lock = SPIN_LOCK_UNLOCKED;
+static spinlock_t dlm_mig_cookie_lock = SPIN_LOCK_UNLOCKED;
+static u64 dlm_mig_cookie = 1;
+
+static u64 dlm_get_next_mig_cookie(void)
+{
+ u64 c;
+ spin_lock(&dlm_mig_cookie_lock);
+ c = dlm_mig_cookie;
+ if (dlm_mig_cookie == (~0ULL))
+ dlm_mig_cookie = 1;
+ else
+ dlm_mig_cookie++;
+ spin_unlock(&dlm_mig_cookie_lock);
+ return c;
+}
+
+static inline void dlm_reset_recovery(struct dlm_ctxt *dlm)
+{
+ spin_lock(&dlm->spinlock);
+ clear_bit(dlm->reco.dead_node, dlm->recovery_map);
+ dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
+ dlm->reco.new_master = O2NM_INVALID_NODE_NUM;
+ spin_unlock(&dlm->spinlock);
+}
+
+/* Worker function used during recovery. */
+void dlm_dispatch_work(void *data)
+{
+ struct dlm_ctxt *dlm = (struct dlm_ctxt *)data;
+ LIST_HEAD(tmp_list);
+ struct list_head *iter, *iter2;
+ struct dlm_work_item *item;
+ dlm_workfunc_t *workfunc;
+
+ spin_lock(&dlm->work_lock);
+ list_splice_init(&dlm->work_list, &tmp_list);
+ spin_unlock(&dlm->work_lock);
+
+ list_for_each_safe(iter, iter2, &tmp_list) {
+ item = list_entry(iter, struct dlm_work_item, list);
+ workfunc = item->func;
+ list_del_init(&item->list);
+
+ /* already have ref on dlm to avoid having
+ * it disappear. just double-check. */
+ BUG_ON(item->dlm != dlm);
+
+ /* this is allowed to sleep and
+ * call network stuff */
+ workfunc(item, item->data);
+
+ dlm_put(dlm);
+ kfree(item);
+ }
+}
+
+/*
+ * RECOVERY THREAD
+ */
+
+static void dlm_kick_recovery_thread(struct dlm_ctxt *dlm)
+{
+ /* wake the recovery thread
+ * this will wake the reco thread in one of three places
+ * 1) sleeping with no recovery happening
+ * 2) sleeping with recovery mastered elsewhere
+ * 3) recovery mastered here, waiting on reco data */
+
+ wake_up(&dlm->dlm_reco_thread_wq);
+}
+
+/* Launch the recovery thread */
+int dlm_launch_recovery_thread(struct dlm_ctxt *dlm)
+{
+ mlog(0, "starting dlm recovery thread...\n");
+
+ dlm->dlm_reco_thread_task = kthread_run(dlm_recovery_thread, dlm,
+ "dlm_reco_thread");
+ if (IS_ERR(dlm->dlm_reco_thread_task)) {
+ mlog_errno(PTR_ERR(dlm->dlm_reco_thread_task));
+ dlm->dlm_reco_thread_task = NULL;
+ return -EINVAL;
+ }
+
+ return 0;
+}
+
+void dlm_complete_recovery_thread(struct dlm_ctxt *dlm)
+{
+ if (dlm->dlm_reco_thread_task) {
+ mlog(0, "waiting for dlm recovery thread to exit\n");
+ kthread_stop(dlm->dlm_reco_thread_task);
+ dlm->dlm_reco_thread_task = NULL;
+ }
+}
+
+
+
+/*
+ * this is lame, but here's how recovery works...
+ * 1) all recovery threads cluster wide will work on recovering
+ * ONE node at a time
+ * 2) negotiate who will take over all the locks for the dead node.
+ * thats right... ALL the locks.
+ * 3) once a new master is chosen, everyone scans all locks
+ * and moves aside those mastered by the dead guy
+ * 4) each of these locks should be locked until recovery is done
+ * 5) the new master collects up all of secondary lock queue info
+ * one lock at a time, forcing each node to communicate back
+ * before continuing
+ * 6) each secondary lock queue responds with the full known lock info
+ * 7) once the new master has run all its locks, it sends a ALLDONE!
+ * message to everyone
+ * 8) upon receiving this message, the secondary queue node unlocks
+ * and responds to the ALLDONE
+ * 9) once the new master gets responses from everyone, he unlocks
+ * everything and recovery for this dead node is done
+ *10) go back to 2) while there are still dead nodes
+ *
+ */
+
+
+#define DLM_RECO_THREAD_TIMEOUT_MS (5 * 1000)
+
+static int dlm_recovery_thread(void *data)
+{
+ int status;
+ struct dlm_ctxt *dlm = data;
+ unsigned long timeout = msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS);
+
+ mlog(0, "dlm thread running for %s...\n", dlm->name);
+
+ while (!kthread_should_stop()) {
+ if (dlm_joined(dlm)) {
+ status = dlm_do_recovery(dlm);
+ if (status == -EAGAIN) {
+ /* do not sleep, recheck immediately. */
+ continue;
+ }
+ if (status < 0)
+ mlog_errno(status);
+ }
+
+ wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
+ kthread_should_stop(),
+ timeout);
+ }
+
+ mlog(0, "quitting DLM recovery thread\n");
+ return 0;
+}
+
+/* callers of the top-level api calls (dlmlock/dlmunlock) should
+ * block on the dlm->reco.event when recovery is in progress.
+ * the dlm recovery thread will set this state when it begins
+ * recovering a dead node (as the new master or not) and clear
+ * the state and wake as soon as all affected lock resources have
+ * been marked with the RECOVERY flag */
+static int dlm_in_recovery(struct dlm_ctxt *dlm)
+{
+ int in_recovery;
+ spin_lock(&dlm->spinlock);
+ in_recovery = !!(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
+ spin_unlock(&dlm->spinlock);
+ return in_recovery;
+}
+
+
+void dlm_wait_for_recovery(struct dlm_ctxt *dlm)
+{
+ wait_event(dlm->reco.event, !dlm_in_recovery(dlm));
+}
+
+static void dlm_begin_recovery(struct dlm_ctxt *dlm)
+{
+ spin_lock(&dlm->spinlock);
+ BUG_ON(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
+ dlm->reco.state |= DLM_RECO_STATE_ACTIVE;
+ spin_unlock(&dlm->spinlock);
+}
+
+static void dlm_end_recovery(struct dlm_ctxt *dlm)
+{
+ spin_lock(&dlm->spinlock);
+ BUG_ON(!(dlm->reco.state & DLM_RECO_STATE_ACTIVE));
+ dlm->reco.state &= ~DLM_RECO_STATE_ACTIVE;
+ spin_unlock(&dlm->spinlock);
+ wake_up(&dlm->reco.event);
+}
+
+static int dlm_do_recovery(struct dlm_ctxt *dlm)
+{
+ int status = 0;
+
+ spin_lock(&dlm->spinlock);
+
+ /* check to see if the new master has died */
+ if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM &&
+ test_bit(dlm->reco.new_master, dlm->recovery_map)) {
+ mlog(0, "new master %u died while recovering %u!\n",
+ dlm->reco.new_master, dlm->reco.dead_node);
+ /* unset the new_master, leave dead_node */
+ dlm->reco.new_master = O2NM_INVALID_NODE_NUM;
+ }
+
+ /* select a target to recover */
+ if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
+ int bit;
+
+ bit = find_next_bit (dlm->recovery_map, O2NM_MAX_NODES+1, 0);
+ if (bit >= O2NM_MAX_NODES || bit < 0)
+ dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
+ else
+ dlm->reco.dead_node = bit;
+ } else if (!test_bit(dlm->reco.dead_node, dlm->recovery_map)) {
+ /* BUG? */
+ mlog(ML_ERROR, "dead_node %u no longer in recovery map!\n",
+ dlm->reco.dead_node);
+ dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
+ }
+
+ if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
+ // mlog(0, "nothing to recover! sleeping now!\n");
+ spin_unlock(&dlm->spinlock);
+ /* return to main thread loop and sleep. */
+ return 0;
+ }
+ mlog(0, "recovery thread found node %u in the recovery map!\n",
+ dlm->reco.dead_node);
+ spin_unlock(&dlm->spinlock);
+
+ /* take write barrier */
+ /* (stops the list reshuffling thread, proxy ast handling) */
+ dlm_begin_recovery(dlm);
+
+ if (dlm->reco.new_master == dlm->node_num)
+ goto master_here;
+
+ if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
+ /* choose a new master */
+ if (!dlm_pick_recovery_master(dlm)) {
+ /* already notified everyone. go. */
+ dlm->reco.new_master = dlm->node_num;
+ goto master_here;
+ }
+ mlog(0, "another node will master this recovery session.\n");
+ }
+ mlog(0, "dlm=%s, new_master=%u, this node=%u, dead_node=%u\n",
+ dlm->name, dlm->reco.new_master,
+ dlm->node_num, dlm->reco.dead_node);
+
+ /* it is safe to start everything back up here
+ * because all of the dead node's lock resources
+ * have been marked as in-recovery */
+ dlm_end_recovery(dlm);
+
+ /* sleep out in main dlm_recovery_thread loop. */
+ return 0;
+
+master_here:
+ mlog(0, "mastering recovery of %s:%u here(this=%u)!\n",
+ dlm->name, dlm->reco.dead_node, dlm->node_num);
+
+ status = dlm_remaster_locks(dlm, dlm->reco.dead_node);
+ if (status < 0) {
+ mlog(ML_ERROR, "error %d remastering locks for node %u, "
+ "retrying.\n", status, dlm->reco.dead_node);
+ } else {
+ /* success! see if any other nodes need recovery */
+ dlm_reset_recovery(dlm);
+ }
+ dlm_end_recovery(dlm);
+
+ /* continue and look for another dead node */
+ return -EAGAIN;
+}
+
+static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node)
+{
+ int status = 0;
+ struct dlm_reco_node_data *ndata;
+ struct list_head *iter;
+ int all_nodes_done;
+ int destroy = 0;
+ int pass = 0;
+
+ status = dlm_init_recovery_area(dlm, dead_node);
+ if (status < 0)
+ goto leave;
+
+ /* safe to access the node data list without a lock, since this
+ * process is the only one to change the list */
+ list_for_each(iter, &dlm->reco.node_data) {
+ ndata = list_entry (iter, struct dlm_reco_node_data, list);
+ BUG_ON(ndata->state != DLM_RECO_NODE_DATA_INIT);
+ ndata->state = DLM_RECO_NODE_DATA_REQUESTING;
+
+ mlog(0, "requesting lock info from node %u\n",
+ ndata->node_num);
+
+ if (ndata->node_num == dlm->node_num) {
+ ndata->state = DLM_RECO_NODE_DATA_DONE;
+ continue;
+ }
+
+ status = dlm_request_all_locks(dlm, ndata->node_num, dead_node);
+ if (status < 0) {
+ mlog_errno(status);
+ if (dlm_is_host_down(status))
+ ndata->state = DLM_RECO_NODE_DATA_DEAD;
+ else {
+ destroy = 1;
+ goto leave;
+ }
+ }
+
+ switch (ndata->state) {
+ case DLM_RECO_NODE_DATA_INIT:
+ case DLM_RECO_NODE_DATA_FINALIZE_SENT:
+ case DLM_RECO_NODE_DATA_REQUESTED:
+ BUG();
+ break;
+ case DLM_RECO_NODE_DATA_DEAD:
+ mlog(0, "node %u died after requesting "
+ "recovery info for node %u\n",
+ ndata->node_num, dead_node);
+ // start all over
+ destroy = 1;
+ status = -EAGAIN;
+ goto leave;
+ case DLM_RECO_NODE_DATA_REQUESTING:
+ ndata->state = DLM_RECO_NODE_DATA_REQUESTED;
+ mlog(0, "now receiving recovery data from "
+ "node %u for dead node %u\n",
+ ndata->node_num, dead_node);
+ break;
+ case DLM_RECO_NODE_DATA_RECEIVING:
+ mlog(0, "already receiving recovery data from "
+ "node %u for dead node %u\n",
+ ndata->node_num, dead_node);
+ break;
+ case DLM_RECO_NODE_DATA_DONE:
+ mlog(0, "already DONE receiving recovery data "
+ "from node %u for dead node %u\n",
+ ndata->node_num, dead_node);
+ break;
+ }
+ }
+
+ mlog(0, "done requesting all lock info\n");
+
+ /* nodes should be sending reco data now
+ * just need to wait */
+
+ while (1) {
+ /* check all the nodes now to see if we are
+ * done, or if anyone died */
+ all_nodes_done = 1;
+ spin_lock(&dlm_reco_state_lock);
+ list_for_each(iter, &dlm->reco.node_data) {
+ ndata = list_entry (iter, struct dlm_reco_node_data, list);
+
+ mlog(0, "checking recovery state of node %u\n",
+ ndata->node_num);
+ switch (ndata->state) {
+ case DLM_RECO_NODE_DATA_INIT:
+ case DLM_RECO_NODE_DATA_REQUESTING:
+ mlog(ML_ERROR, "bad ndata state for "
+ "node %u: state=%d\n",
+ ndata->node_num, ndata->state);
+ BUG();
+ break;
+ case DLM_RECO_NODE_DATA_DEAD:
+ mlog(0, "node %u died after "
+ "requesting recovery info for "
+ "node %u\n", ndata->node_num,
+ dead_node);
+ spin_unlock(&dlm_reco_state_lock);
+ // start all over
+ destroy = 1;
+ status = -EAGAIN;
+ goto leave;
+ case DLM_RECO_NODE_DATA_RECEIVING:
+ case DLM_RECO_NODE_DATA_REQUESTED:
+ all_nodes_done = 0;
+ break;
+ case DLM_RECO_NODE_DATA_DONE:
+ break;
+ case DLM_RECO_NODE_DATA_FINALIZE_SENT:
+ break;
+ }
+ }
+ spin_unlock(&dlm_reco_state_lock);
+
+ mlog(0, "pass #%d, all_nodes_done?: %s\n", ++pass,
+ all_nodes_done?"yes":"no");
+ if (all_nodes_done) {
+ int ret;
+
+ /* all nodes are now in DLM_RECO_NODE_DATA_DONE state
+ * just send a finalize message to everyone and
+ * clean up */
+ mlog(0, "all nodes are done! send finalize\n");
+ ret = dlm_send_finalize_reco_message(dlm);
+ if (ret < 0)
+ mlog_errno(ret);
+
+ spin_lock(&dlm->spinlock);
+ dlm_finish_local_lockres_recovery(dlm, dead_node,
+ dlm->node_num);
+ spin_unlock(&dlm->spinlock);
+ mlog(0, "should be done with recovery!\n");
+
+ mlog(0, "finishing recovery of %s at %lu, "
+ "dead=%u, this=%u, new=%u\n", dlm->name,
+ jiffies, dlm->reco.dead_node,
+ dlm->node_num, dlm->reco.new_master);
+ destroy = 1;
+ status = ret;
+ /* rescan everything marked dirty along the way */
+ dlm_kick_thread(dlm, NULL);
+ break;
+ }
+ /* wait to be signalled, with periodic timeout
+ * to check for node death */
+ wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
+ kthread_should_stop(),
+ msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS));
+
+ }
+
+leave:
+ if (destroy)
+ dlm_destroy_recovery_area(dlm, dead_node);
+
+ mlog_exit(status);
+ return status;
+}
+
+static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
+{
+ int num=0;
+ struct dlm_reco_node_data *ndata;
+
+ spin_lock(&dlm->spinlock);
+ memcpy(dlm->reco.node_map, dlm->domain_map, sizeof(dlm->domain_map));
+ /* nodes can only be removed (by dying) after dropping
+ * this lock, and death will be trapped later, so this should do */
+ spin_unlock(&dlm->spinlock);
+
+ while (1) {
+ num = find_next_bit (dlm->reco.node_map, O2NM_MAX_NODES, num);
+ if (num >= O2NM_MAX_NODES) {
+ break;
+ }
+ BUG_ON(num == dead_node);
+
+ ndata = kcalloc(1, sizeof(*ndata), GFP_KERNEL);
+ if (!ndata) {
+ dlm_destroy_recovery_area(dlm, dead_node);
+ return -ENOMEM;
+ }
+ ndata->node_num = num;
+ ndata->state = DLM_RECO_NODE_DATA_INIT;
+ spin_lock(&dlm_reco_state_lock);
+ list_add_tail(&ndata->list, &dlm->reco.node_data);
+ spin_unlock(&dlm_reco_state_lock);
+ num++;
+ }
+
+ return 0;
+}
+
+static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
+{
+ struct list_head *iter, *iter2;
+ struct dlm_reco_node_data *ndata;
+ LIST_HEAD(tmplist);
+
+ spin_lock(&dlm_reco_state_lock);
+ list_splice_init(&dlm->reco.node_data, &tmplist);
+ spin_unlock(&dlm_reco_state_lock);
+
+ list_for_each_safe(iter, iter2, &tmplist) {
+ ndata = list_entry (iter, struct dlm_reco_node_data, list);
+ list_del_init(&ndata->list);
+ kfree(ndata);
+ }
+}
+
+static int dlm_request_all_locks(struct dlm_ctxt *dlm, u8 request_from,
+ u8 dead_node)
+{
+ struct dlm_lock_request lr;
+ enum dlm_status ret;
+
+ mlog(0, "\n");
+
+
+ mlog(0, "dlm_request_all_locks: dead node is %u, sending request "
+ "to %u\n", dead_node, request_from);
+
+ memset(&lr, 0, sizeof(lr));
+ lr.node_idx = dlm->node_num;
+ lr.dead_node = dead_node;
+
+ // send message
+ ret = DLM_NOLOCKMGR;
+ ret = o2net_send_message(DLM_LOCK_REQUEST_MSG, dlm->key,
+ &lr, sizeof(lr), request_from, NULL);
+
+ /* negative status is handled by caller */
+ if (ret < 0)
+ mlog_errno(ret);
+
+ // return from here, then
+ // sleep until all received or error
+ return ret;
+
+}
+
+int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data)
+{
+ struct dlm_ctxt *dlm = data;
+ struct dlm_lock_request *lr = (struct dlm_lock_request *)msg->buf;
+ char *buf = NULL;
+ struct dlm_work_item *item = NULL;
+
+ if (!dlm_grab(dlm))
+ return -EINVAL;
+
+ BUG_ON(lr->dead_node != dlm->reco.dead_node);
+
+ item = kcalloc(1, sizeof(*item), GFP_KERNEL);
+ if (!item) {
+ dlm_put(dlm);
+ return -ENOMEM;
+ }
+
+ /* this will get freed by dlm_request_all_locks_worker */
+ buf = (char *) __get_free_page(GFP_KERNEL);
+ if (!buf) {
+ kfree(item);
+ dlm_put(dlm);
+ return -ENOMEM;
+ }
+
+ /* queue up work for dlm_request_all_locks_worker */
+ dlm_grab(dlm); /* get an extra ref for the work item */
+ dlm_init_work_item(dlm, item, dlm_request_all_locks_worker, buf);
+ item->u.ral.reco_master = lr->node_idx;
+ item->u.ral.dead_node = lr->dead_node;
+ spin_lock(&dlm->work_lock);
+ list_add_tail(&item->list, &dlm->work_list);
+ spin_unlock(&dlm->work_lock);
+ schedule_work(&dlm->dispatched_work);
+
+ dlm_put(dlm);
+ return 0;
+}
+
+static void dlm_request_all_locks_worker(struct dlm_work_item *item, void *data)
+{
+ struct dlm_migratable_lockres *mres;
+ struct dlm_lock_resource *res;
+ struct dlm_ctxt *dlm;
+ LIST_HEAD(resources);
+ struct list_head *iter;
+ int ret;
+ u8 dead_node, reco_master;
+
+ dlm = item->dlm;
+ dead_node = item->u.ral.dead_node;
+ reco_master = item->u.ral.reco_master;
+ BUG_ON(dead_node != dlm->reco.dead_node);
+ BUG_ON(reco_master != dlm->reco.new_master);
+
+ mres = (struct dlm_migratable_lockres *)data;
+
+ /* lock resources should have already been moved to the
+ * dlm->reco.resources list. now move items from that list
+ * to a temp list if the dead owner matches. note that the
+ * whole cluster recovers only one node at a time, so we
+ * can safely move UNKNOWN lock resources for each recovery
+ * session. */
+ dlm_move_reco_locks_to_list(dlm, &resources, dead_node);
+
+ /* now we can begin blasting lockreses without the dlm lock */
+ list_for_each(iter, &resources) {
+ res = list_entry (iter, struct dlm_lock_resource, recovering);
+ ret = dlm_send_one_lockres(dlm, res, mres, reco_master,
+ DLM_MRES_RECOVERY);
+ if (ret < 0)
+ mlog_errno(ret);
+ }
+
+ /* move the resources back to the list */
+ spin_lock(&dlm->spinlock);
+ list_splice_init(&resources, &dlm->reco.resources);
+ spin_unlock(&dlm->spinlock);
+
+ ret = dlm_send_all_done_msg(dlm, dead_node, reco_master);
+ if (ret < 0)
+ mlog_errno(ret);
+
+ free_page((unsigned long)data);
+}
+
+
+static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, u8 dead_node, u8 send_to)
+{
+ int ret, tmpret;
+ struct dlm_reco_data_done done_msg;
+
+ memset(&done_msg, 0, sizeof(done_msg));
+ done_msg.node_idx = dlm->node_num;
+ done_msg.dead_node = dead_node;
+ mlog(0, "sending DATA DONE message to %u, "
+ "my node=%u, dead node=%u\n", send_to, done_msg.node_idx,
+ done_msg.dead_node);
+
+ ret = o2net_send_message(DLM_RECO_DATA_DONE_MSG, dlm->key, &done_msg,
+ sizeof(done_msg), send_to, &tmpret);
+ /* negative status is ignored by the caller */
+ if (ret >= 0)
+ ret = tmpret;
+ return ret;
+}
+
+
+int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data)
+{
+ struct dlm_ctxt *dlm = data;
+ struct dlm_reco_data_done *done = (struct dlm_reco_data_done *)msg->buf;
+ struct list_head *iter;
+ struct dlm_reco_node_data *ndata = NULL;
+ int ret = -EINVAL;
+
+ if (!dlm_grab(dlm))
+ return -EINVAL;
+
+ mlog(0, "got DATA DONE: dead_node=%u, reco.dead_node=%u, "
+ "node_idx=%u, this node=%u\n", done->dead_node,
+ dlm->reco.dead_node, done->node_idx, dlm->node_num);
+ BUG_ON(done->dead_node != dlm->reco.dead_node);
+
+ spin_lock(&dlm_reco_state_lock);
+ list_for_each(iter, &dlm->reco.node_data) {
+ ndata = list_entry (iter, struct dlm_reco_node_data, list);
+ if (ndata->node_num != done->node_idx)
+ continue;
+
+ switch (ndata->state) {
+ case DLM_RECO_NODE_DATA_INIT:
+ case DLM_RECO_NODE_DATA_DEAD:
+ case DLM_RECO_NODE_DATA_DONE:
+ case DLM_RECO_NODE_DATA_FINALIZE_SENT:
+ mlog(ML_ERROR, "bad ndata state for node %u:"
+ " state=%d\n", ndata->node_num,
+ ndata->state);
+ BUG();
+ break;
+ case DLM_RECO_NODE_DATA_RECEIVING:
+ case DLM_RECO_NODE_DATA_REQUESTED:
+ case DLM_RECO_NODE_DATA_REQUESTING:
+ mlog(0, "node %u is DONE sending "
+ "recovery data!\n",
+ ndata->node_num);
+
+ ndata->state = DLM_RECO_NODE_DATA_DONE;
+ ret = 0;
+ break;
+ }
+ }
+ spin_unlock(&dlm_reco_state_lock);
+
+ /* wake the recovery thread, some node is done */
+ if (!ret)
+ dlm_kick_recovery_thread(dlm);
+
+ if (ret < 0)
+ mlog(ML_ERROR, "failed to find recovery node data for node "
+ "%u\n", done->node_idx);
+ dlm_put(dlm);
+
+ mlog(0, "leaving reco data done handler, ret=%d\n", ret);
+ return ret;
+}
+
+static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
+ struct list_head *list,
+ u8 dead_node)
+{
+ struct dlm_lock_resource *res;
+ struct list_head *iter, *iter2;
+
+ spin_lock(&dlm->spinlock);
+ list_for_each_safe(iter, iter2, &dlm->reco.resources) {
+ res = list_entry (iter, struct dlm_lock_resource, recovering);
+ if (dlm_is_recovery_lock(res->lockname.name,
+ res->lockname.len))
+ continue;
+ if (res->owner == dead_node) {
+ mlog(0, "found lockres owned by dead node while "
+ "doing recovery for node %u. sending it.\n",
+ dead_node);
+ list_del_init(&res->recovering);
+ list_add_tail(&res->recovering, list);
+ } else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
+ mlog(0, "found UNKNOWN owner while doing recovery "
+ "for node %u. sending it.\n", dead_node);
+ list_del_init(&res->recovering);
+ list_add_tail(&res->recovering, list);
+ }
+ }
+ spin_unlock(&dlm->spinlock);
+}
+
+static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res)
+{
+ int total_locks = 0;
+ struct list_head *iter, *queue = &res->granted;
+ int i;
+
+ for (i=0; i<3; i++) {
+ list_for_each(iter, queue)
+ total_locks++;
+ queue++;
+ }
+ return total_locks;
+}
+
+
+static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
+ struct dlm_migratable_lockres *mres,
+ u8 send_to,
+ struct dlm_lock_resource *res,
+ int total_locks)
+{
+ u64 mig_cookie = be64_to_cpu(mres->mig_cookie);
+ int mres_total_locks = be32_to_cpu(mres->total_locks);
+ int sz, ret = 0, status = 0;
+ u8 orig_flags = mres->flags,
+ orig_master = mres->master;
+
+ BUG_ON(mres->num_locks > DLM_MAX_MIGRATABLE_LOCKS);
+ if (!mres->num_locks)
+ return 0;
+
+ sz = sizeof(struct dlm_migratable_lockres) +
+ (mres->num_locks * sizeof(struct dlm_migratable_lock));
+
+ /* add an all-done flag if we reached the last lock */
+ orig_flags = mres->flags;
+ BUG_ON(total_locks > mres_total_locks);
+ if (total_locks == mres_total_locks)
+ mres->flags |= DLM_MRES_ALL_DONE;
+
+ /* send it */
+ ret = o2net_send_message(DLM_MIG_LOCKRES_MSG, dlm->key, mres,
+ sz, send_to, &status);
+ if (ret < 0) {
+ /* XXX: negative status is not handled.
+ * this will end up killing this node. */
+ mlog_errno(ret);
+ } else {
+ /* might get an -ENOMEM back here */
+ ret = status;
+ if (ret < 0) {
+ mlog_errno(ret);
+
+ if (ret == -EFAULT) {
+ mlog(ML_ERROR, "node %u told me to kill "
+ "myself!\n", send_to);
+ BUG();
+ }
+ }
+ }
+
+ /* zero and reinit the message buffer */
+ dlm_init_migratable_lockres(mres, res->lockname.name,
+ res->lockname.len, mres_total_locks,
+ mig_cookie, orig_flags, orig_master);
+ return ret;
+}
+
+static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
+ const char *lockname, int namelen,
+ int total_locks, u64 cookie,
+ u8 flags, u8 master)
+{
+ /* mres here is one full page */
+ memset(mres, 0, PAGE_SIZE);
+ mres->lockname_len = namelen;
+ memcpy(mres->lockname, lockname, namelen);
+ mres->num_locks = 0;
+ mres->total_locks = cpu_to_be32(total_locks);
+ mres->mig_cookie = cpu_to_be64(cookie);
+ mres->flags = flags;
+ mres->master = master;
+}
+
+
+/* returns 1 if this lock fills the network structure,
+ * 0 otherwise */
+static int dlm_add_lock_to_array(struct dlm_lock *lock,
+ struct dlm_migratable_lockres *mres, int queue)
+{
+ struct dlm_migratable_lock *ml;
+ int lock_num = mres->num_locks;
+
+ ml = &(mres->ml[lock_num]);
+ ml->cookie = lock->ml.cookie;
+ ml->type = lock->ml.type;
+ ml->convert_type = lock->ml.convert_type;
+ ml->highest_blocked = lock->ml.highest_blocked;
+ ml->list = queue;
+ if (lock->lksb) {
+ ml->flags = lock->lksb->flags;
+ /* send our current lvb */
+ if (ml->type == LKM_EXMODE ||
+ ml->type == LKM_PRMODE) {
+ /* if it is already set, this had better be a PR
+ * and it has to match */
+ if (mres->lvb[0] && (ml->type == LKM_EXMODE ||
+ memcmp(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN))) {
+ mlog(ML_ERROR, "mismatched lvbs!\n");
+ __dlm_print_one_lock_resource(lock->lockres);
+ BUG();
+ }
+ memcpy(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN);
+ }
+ }
+ ml->node = lock->ml.node;
+ mres->num_locks++;
+ /* we reached the max, send this network message */
+ if (mres->num_locks == DLM_MAX_MIGRATABLE_LOCKS)
+ return 1;
+ return 0;
+}
+
+
+int dlm_send_one_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
+ struct dlm_migratable_lockres *mres,
+ u8 send_to, u8 flags)
+{
+ struct list_head *queue, *iter;
+ int total_locks, i;
+ u64 mig_cookie = 0;
+ struct dlm_lock *lock;
+ int ret = 0;
+
+ BUG_ON(!(flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
+
+ mlog(0, "sending to %u\n", send_to);
+
+ total_locks = dlm_num_locks_in_lockres(res);
+ if (total_locks > DLM_MAX_MIGRATABLE_LOCKS) {
+ /* rare, but possible */
+ mlog(0, "argh. lockres has %d locks. this will "
+ "require more than one network packet to "
+ "migrate\n", total_locks);
+ mig_cookie = dlm_get_next_mig_cookie();
+ }
+
+ dlm_init_migratable_lockres(mres, res->lockname.name,
+ res->lockname.len, total_locks,
+ mig_cookie, flags, res->owner);
+
+ total_locks = 0;
+ for (i=DLM_GRANTED_LIST; i<=DLM_BLOCKED_LIST; i++) {
+ queue = dlm_list_idx_to_ptr(res, i);
+ list_for_each(iter, queue) {
+ lock = list_entry (iter, struct dlm_lock, list);
+
+ /* add another lock. */
+ total_locks++;
+ if (!dlm_add_lock_to_array(lock, mres, i))
+ continue;
+
+ /* this filled the lock message,
+ * we must send it immediately. */
+ ret = dlm_send_mig_lockres_msg(dlm, mres, send_to,
+ res, total_locks);
+ if (ret < 0) {
+ // TODO
+ mlog(ML_ERROR, "dlm_send_mig_lockres_msg "
+ "returned %d, TODO\n", ret);
+ BUG();
+ }
+ }
+ }
+ /* flush any remaining locks */
+ ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, res, total_locks);
+ if (ret < 0) {
+ // TODO
+ mlog(ML_ERROR, "dlm_send_mig_lockres_msg returned %d, "
+ "TODO\n", ret);
+ BUG();
+ }
+ return ret;
+}
+
+
+
+/*
+ * this message will contain no more than one page worth of
+ * recovery data, and it will work on only one lockres.
+ * there may be many locks in this page, and we may need to wait
+ * for additional packets to complete all the locks (rare, but
+ * possible).
+ */
+/*
+ * NOTE: the allocation error cases here are scary
+ * we really cannot afford to fail an alloc in recovery
+ * do we spin? returning an error only delays the problem really
+ */
+
+int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data)
+{
+ struct dlm_ctxt *dlm = data;
+ struct dlm_migratable_lockres *mres =
+ (struct dlm_migratable_lockres *)msg->buf;
+ int ret = 0;
+ u8 real_master;
+ char *buf = NULL;
+ struct dlm_work_item *item = NULL;
+ struct dlm_lock_resource *res = NULL;
+
+ if (!dlm_grab(dlm))
+ return -EINVAL;
+
+ BUG_ON(!(mres->flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
+
+ real_master = mres->master;
+ if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
+ /* cannot migrate a lockres with no master */
+ BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
+ }
+
+ mlog(0, "%s message received from node %u\n",
+ (mres->flags & DLM_MRES_RECOVERY) ?
+ "recovery" : "migration", mres->master);
+ if (mres->flags & DLM_MRES_ALL_DONE)
+ mlog(0, "all done flag. all lockres data received!\n");
+
+ ret = -ENOMEM;
+ buf = kmalloc(be16_to_cpu(msg->data_len), GFP_KERNEL);
+ item = kcalloc(1, sizeof(*item), GFP_KERNEL);
+ if (!buf || !item)
+ goto leave;
+
+ /* lookup the lock to see if we have a secondary queue for this
+ * already... just add the locks in and this will have its owner
+ * and RECOVERY flag changed when it completes. */
+ res = dlm_lookup_lockres(dlm, mres->lockname, mres->lockname_len);
+ if (res) {
+ /* this will get a ref on res */
+ /* mark it as recovering/migrating and hash it */
+ spin_lock(&res->spinlock);
+ if (mres->flags & DLM_MRES_RECOVERY) {
+ res->state |= DLM_LOCK_RES_RECOVERING;
+ } else {
+ if (res->state & DLM_LOCK_RES_MIGRATING) {
+ /* this is at least the second
+ * lockres message */
+ mlog(0, "lock %.*s is already migrating\n",
+ mres->lockname_len,
+ mres->lockname);
+ } else if (res->state & DLM_LOCK_RES_RECOVERING) {
+ /* caller should BUG */
+ mlog(ML_ERROR, "node is attempting to migrate "
+ "lock %.*s, but marked as recovering!\n",
+ mres->lockname_len, mres->lockname);
+ ret = -EFAULT;
+ spin_unlock(&res->spinlock);
+ goto leave;
+ }
+ res->state |= DLM_LOCK_RES_MIGRATING;
+ }
+ spin_unlock(&res->spinlock);
+ } else {
+ /* need to allocate, just like if it was
+ * mastered here normally */
+ res = dlm_new_lockres(dlm, mres->lockname, mres->lockname_len);
+ if (!res)
+ goto leave;
+
+ /* to match the ref that we would have gotten if
+ * dlm_lookup_lockres had succeeded */
+ dlm_lock