#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>
const char *name,
unsigned int namelen)
{
- struct dlm_lock_resource *res;
-
if (dlm != mle->dlm)
return 0;
- if (mle->type == DLM_MLE_BLOCK ||
- mle->type == DLM_MLE_MIGRATION) {
- if (namelen != mle->u.name.len ||
- memcmp(name, mle->u.name.name, namelen)!=0)
- return 0;
- } else {
- res = mle->u.res;
- if (namelen != res->lockname.len ||
- memcmp(res->lockname.name, name, namelen) != 0)
- return 0;
- }
+ if (namelen != mle->mnamelen ||
+ memcmp(name, mle->mname, namelen) != 0)
+ return 0;
+
return 1;
}
mle->dlm = dlm;
mle->type = type;
- INIT_LIST_HEAD(&mle->list);
+ INIT_HLIST_NODE(&mle->master_hash_node);
INIT_LIST_HEAD(&mle->hb_events);
memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
spin_lock_init(&mle->spinlock);
mle->new_master = O2NM_MAX_NODES;
mle->inuse = 0;
+ BUG_ON(mle->type != DLM_MLE_BLOCK &&
+ mle->type != DLM_MLE_MASTER &&
+ mle->type != DLM_MLE_MIGRATION);
+
if (mle->type == DLM_MLE_MASTER) {
BUG_ON(!res);
- mle->u.res = res;
- } else if (mle->type == DLM_MLE_BLOCK) {
- BUG_ON(!name);
- memcpy(mle->u.name.name, name, namelen);
- mle->u.name.len = namelen;
- } else /* DLM_MLE_MIGRATION */ {
+ mle->mleres = res;
+ memcpy(mle->mname, res->lockname.name, res->lockname.len);
+ mle->mnamelen = res->lockname.len;
+ mle->mnamehash = res->lockname.hash;
+ } else {
BUG_ON(!name);
- memcpy(mle->u.name.name, name, namelen);
- mle->u.name.len = namelen;
+ mle->mleres = NULL;
+ memcpy(mle->mname, name, namelen);
+ mle->mnamelen = namelen;
+ mle->mnamehash = dlm_lockid_hash(name, namelen);
}
+ atomic_inc(&dlm->mle_tot_count[mle->type]);
+ atomic_inc(&dlm->mle_cur_count[mle->type]);
+
/* copy off the node_map and register hb callbacks on our copy */
memcpy(mle->node_map, dlm->domain_map, sizeof(mle->node_map));
memcpy(mle->vote_map, dlm->domain_map, sizeof(mle->vote_map));
__dlm_mle_attach_hb_events(dlm, mle);
}
+void __dlm_unlink_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
+{
+ assert_spin_locked(&dlm->spinlock);
+ assert_spin_locked(&dlm->master_lock);
+
+ if (!hlist_unhashed(&mle->master_hash_node))
+ hlist_del_init(&mle->master_hash_node);
+}
+
+void __dlm_insert_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle)
+{
+ struct hlist_head *bucket;
+
+ assert_spin_locked(&dlm->master_lock);
+
+ bucket = dlm_master_hash(dlm, mle->mnamehash);
+ hlist_add_head(&mle->master_hash_node, bucket);
+}
/* returns 1 if found, 0 if not */
static int dlm_find_mle(struct dlm_ctxt *dlm,
char *name, unsigned int namelen)
{
struct dlm_master_list_entry *tmpmle;
+ struct hlist_head *bucket;
+ struct hlist_node *list;
+ unsigned int hash;
assert_spin_locked(&dlm->master_lock);
- list_for_each_entry(tmpmle, &dlm->master_list, list) {
+ hash = dlm_lockid_hash(name, namelen);
+ bucket = dlm_master_hash(dlm, hash);
+ hlist_for_each(list, bucket) {
+ tmpmle = hlist_entry(list, struct dlm_master_list_entry,
+ master_hash_node);
if (!dlm_mle_equal(dlm, tmpmle, name, namelen))
continue;
dlm_get_mle(tmpmle);
mle = container_of(kref, struct dlm_master_list_entry, mle_refs);
dlm = mle->dlm;
- if (mle->type != DLM_MLE_MASTER) {
- mlog(0, "calling mle_release for %.*s, type %d\n",
- mle->u.name.len, mle->u.name.name, mle->type);
- } else {
- mlog(0, "calling mle_release for %.*s, type %d\n",
- mle->u.res->lockname.len,
- mle->u.res->lockname.name, mle->type);
- }
assert_spin_locked(&dlm->spinlock);
assert_spin_locked(&dlm->master_lock);
+ mlog(0, "Releasing mle for %.*s, type %d\n", mle->mnamelen, mle->mname,
+ mle->type);
+
/* remove from list if not already */
- if (!list_empty(&mle->list))
- list_del_init(&mle->list);
+ __dlm_unlink_mle(dlm, mle);
/* detach the mle from the domain node up/down events */
__dlm_mle_detach_hb_events(dlm, mle);
+ atomic_dec(&dlm->mle_cur_count[mle->type]);
+
/* NOTE: kfree under spinlock here.
* if this is bad, we can move this to a freelist. */
kmem_cache_free(dlm_mle_cache, mle);
kmem_cache_destroy(dlm_lockres_cache);
}
-static void dlm_set_lockres_owner(struct dlm_ctxt *dlm,
- struct dlm_lock_resource *res,
- u8 owner)
-{
- assert_spin_locked(&res->spinlock);
-
- mlog_entry("%.*s, %u\n", res->lockname.len, res->lockname.name, owner);
-
- if (owner == dlm->node_num)
- atomic_inc(&dlm->local_resources);
- else if (owner == DLM_LOCK_RES_OWNER_UNKNOWN)
- atomic_inc(&dlm->unknown_resources);
- else
- atomic_inc(&dlm->remote_resources);
-
- res->owner = owner;
-}
-
-void dlm_change_lockres_owner(struct dlm_ctxt *dlm,
- struct dlm_lock_resource *res, u8 owner)
-{
- assert_spin_locked(&res->spinlock);
-
- if (owner == res->owner)
- return;
-
- if (res->owner == dlm->node_num)
- atomic_dec(&dlm->local_resources);
- else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN)
- atomic_dec(&dlm->unknown_resources);
- else
- atomic_dec(&dlm->remote_resources);
-
- dlm_set_lockres_owner(dlm, res, owner);
-}
-
-
static void dlm_lockres_release(struct kref *kref)
{
struct dlm_lock_resource *res;
}
spin_unlock(&dlm->track_lock);
+ atomic_dec(&dlm->res_cur_count);
+
dlm_put(dlm);
if (!hlist_unhashed(&res->hash_node) ||
kref_init(&res->refs);
+ atomic_inc(&dlm->res_tot_count);
+ atomic_inc(&dlm->res_cur_count);
+
/* just for consistency */
spin_lock(&res->spinlock);
dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
alloc_mle = NULL;
dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0);
set_bit(dlm->node_num, mle->maybe_map);
- list_add(&mle->list, &dlm->master_list);
+ __dlm_insert_mle(dlm, mle);
/* still holding the dlm spinlock, check the recovery map
* to see if there are any nodes that still need to be
res->lockname.len,
res->lockname.name);
mle->type = DLM_MLE_MASTER;
- mle->u.res = res;
+ mle->mleres = res;
}
}
}
BUG_ON(mle->type == DLM_MLE_MIGRATION);
- if (mle->type != DLM_MLE_MASTER) {
- request.namelen = mle->u.name.len;
- memcpy(request.name, mle->u.name.name, request.namelen);
- } else {
- request.namelen = mle->u.res->lockname.len;
- memcpy(request.name, mle->u.res->lockname.name,
- request.namelen);
- }
+ request.namelen = (u8)mle->mnamelen;
+ memcpy(request.name, mle->mname, request.namelen);
again:
ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request,
// "add the block.\n");
dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL, name, namelen);
set_bit(request->node_idx, mle->maybe_map);
- list_add(&mle->list, &dlm->master_list);
+ __dlm_insert_mle(dlm, mle);
response = DLM_MASTER_RESP_NO;
} else {
// mlog(0, "mle was found\n");
if (!mle) {
if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN &&
res->owner != assert->node_idx) {
- mlog(ML_ERROR, "assert_master from "
- "%u, but current owner is "
- "%u! (%.*s)\n",
- assert->node_idx, res->owner,
- namelen, name);
- goto kill;
+ mlog(ML_ERROR, "DIE! Mastery assert from %u, "
+ "but current owner is %u! (%.*s)\n",
+ assert->node_idx, res->owner, namelen,
+ name);
+ __dlm_print_one_lock_resource(res);
+ BUG();
}
} else if (mle->type != DLM_MLE_MIGRATION) {
if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
assert->node_idx, rr, extra_ref, mle->inuse);
dlm_print_one_mle(mle);
}
- list_del_init(&mle->list);
+ __dlm_unlink_mle(dlm, mle);
__dlm_mle_detach_hb_events(dlm, mle);
__dlm_put_mle(mle);
if (extra_ref) {
* is complete everywhere. if the target dies while this is
* going on, some nodes could potentially see the target as the
* master, so it is important that my recovery finds the migration
- * mle and sets the master to UNKNONWN. */
+ * mle and sets the master to UNKNOWN. */
/* wait for new node to assert master */
tmp->master = master;
atomic_set(&tmp->woken, 1);
wake_up(&tmp->wq);
- /* remove it from the list so that only one
- * mle will be found */
- list_del_init(&tmp->list);
- /* this was obviously WRONG. mle is uninited here. should be tmp. */
+ /* remove it so that only one mle will be found */
+ __dlm_unlink_mle(dlm, tmp);
__dlm_mle_detach_hb_events(dlm, tmp);
ret = DLM_MIGRATE_RESPONSE_MASTERY_REF;
mlog(0, "%s:%.*s: master=%u, newmaster=%u, "
mle->master = master;
/* do this for consistency with other mle types */
set_bit(new_master, mle->maybe_map);
- list_add(&mle->list, &dlm->master_list);
+ __dlm_insert_mle(dlm, mle);
return ret;
}
-
-void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
+/*
+ * Sets the owner of the lockres, associated to the mle, to UNKNOWN
+ */
+static struct dlm_lock_resource *dlm_reset_mleres_owner(struct dlm_ctxt *dlm,
+ struct dlm_master_list_entry *mle)
{
- struct dlm_master_list_entry *mle, *next;
struct dlm_lock_resource *res;
- unsigned int hash;
- mlog_entry("dlm=%s, dead node=%u\n", dlm->name, dead_node);
-top:
- assert_spin_locked(&dlm->spinlock);
+ /* Find the lockres associated to the mle and set its owner to UNK */
+ res = __dlm_lookup_lockres(dlm, mle->mname, mle->mnamelen,
+ mle->mnamehash);
+ if (res) {
+ spin_unlock(&dlm->master_lock);
- /* clean the master list */
- spin_lock(&dlm->master_lock);
- list_for_each_entry_safe(mle, next, &dlm->master_list, list) {
- BUG_ON(mle->type != DLM_MLE_BLOCK &&
- mle->type != DLM_MLE_MASTER &&
- mle->type != DLM_MLE_MIGRATION);
-
- /* MASTER mles are initiated locally. the waiting
- * process will notice the node map change
- * shortly. let that happen as normal. */
- if (mle->type == DLM_MLE_MASTER)
- continue;
+ /* move lockres onto recovery list */
+ spin_lock(&res->spinlock);
+ dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
+ dlm_move_lockres_to_recovery_list(dlm, res);
+ spin_unlock(&res->spinlock);
+ dlm_lockres_put(res);
+ /* about to get rid of mle, detach from heartbeat */
+ __dlm_mle_detach_hb_events(dlm, mle);
- /* BLOCK mles are initiated by other nodes.
- * need to clean up if the dead node would have
- * been the master. */
- if (mle->type == DLM_MLE_BLOCK) {
- int bit;
+ /* dump the mle */
+ spin_lock(&dlm->master_lock);
+ __dlm_put_mle(mle);
+ spin_unlock(&dlm->master_lock);
+ }
- spin_lock(&mle->spinlock);
- bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
- if (bit != dead_node) {
- mlog(0, "mle found, but dead node %u would "
- "not have been master\n", dead_node);
- spin_unlock(&mle->spinlock);
- } else {
- /* must drop the refcount by one since the
- * assert_master will never arrive. this
- * may result in the mle being unlinked and
- * freed, but there may still be a process
- * waiting in the dlmlock path which is fine. */
- mlog(0, "node %u was expected master\n",
- dead_node);
- atomic_set(&mle->woken, 1);
- spin_unlock(&mle->spinlock);
- wake_up(&mle->wq);
- /* do not need events any longer, so detach
- * from heartbeat */
- __dlm_mle_detach_hb_events(dlm, mle);
- __dlm_put_mle(mle);
- }
- continue;
- }
+ return res;
+}
- /* everything else is a MIGRATION mle */
-
- /* the rule for MIGRATION mles is that the master
- * becomes UNKNOWN if *either* the original or
- * the new master dies. all UNKNOWN lockreses
- * are sent to whichever node becomes the recovery
- * master. the new master is responsible for
- * determining if there is still a master for
- * this lockres, or if he needs to take over
- * mastery. either way, this node should expect
- * another message to resolve this. */
- if (mle->master != dead_node &&
- mle->new_master != dead_node)
- continue;
+static void dlm_clean_migration_mle(struct dlm_ctxt *dlm,
+ struct dlm_master_list_entry *mle)
+{
+ __dlm_mle_detach_hb_events(dlm, mle);
- /* if we have reached this point, this mle needs to
- * be removed from the list and freed. */
+ spin_lock(&mle->spinlock);
+ __dlm_unlink_mle(dlm, mle);
+ atomic_set(&mle->woken, 1);
+ spin_unlock(&mle->spinlock);
- /* remove from the list early. NOTE: unlinking
- * list_head while in list_for_each_safe */
- __dlm_mle_detach_hb_events(dlm, mle);
- spin_lock(&mle->spinlock);
- list_del_init(&mle->list);
+ wake_up(&mle->wq);
+}
+
+static void dlm_clean_block_mle(struct dlm_ctxt *dlm,
+ struct dlm_master_list_entry *mle, u8 dead_node)
+{
+ int bit;
+
+ BUG_ON(mle->type != DLM_MLE_BLOCK);
+
+ spin_lock(&mle->spinlock);
+ bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
+ if (bit != dead_node) {
+ mlog(0, "mle found, but dead node %u would not have been "
+ "master\n", dead_node);
+ spin_unlock(&mle->spinlock);
+ } else {
+ /* Must drop the refcount by one since the assert_master will
+ * never arrive. This may result in the mle being unlinked and
+ * freed, but there may still be a process waiting in the
+ * dlmlock path which is fine. */
+ mlog(0, "node %u was expected master\n", dead_node);
atomic_set(&mle->woken, 1);
spin_unlock(&mle->spinlock);
wake_up(&mle->wq);
- mlog(0, "%s: node %u died during migration from "
- "%u to %u!\n", dlm->name, dead_node,
- mle->master, mle->new_master);
- /* if there is a lockres associated with this
- * mle, find it and set its owner to UNKNOWN */
- hash = dlm_lockid_hash(mle->u.name.name, mle->u.name.len);
- res = __dlm_lookup_lockres(dlm, mle->u.name.name,
- mle->u.name.len, hash);
- if (res) {
- /* unfortunately if we hit this rare case, our
- * lock ordering is messed. we need to drop
- * the master lock so that we can take the
- * lockres lock, meaning that we will have to
- * restart from the head of list. */
- spin_unlock(&dlm->master_lock);
+ /* Do not need events any longer, so detach from heartbeat */
+ __dlm_mle_detach_hb_events(dlm, mle);
+ __dlm_put_mle(mle);
+ }
+}
- /* move lockres onto recovery list */
- spin_lock(&res->spinlock);
- dlm_set_lockres_owner(dlm, res,
- DLM_LOCK_RES_OWNER_UNKNOWN);
- dlm_move_lockres_to_recovery_list(dlm, res);
- spin_unlock(&res->spinlock);
- dlm_lockres_put(res);
+void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
+{
+ struct dlm_master_list_entry *mle;
+ struct dlm_lock_resource *res;
+ struct hlist_head *bucket;
+ struct hlist_node *list;
+ unsigned int i;
- /* about to get rid of mle, detach from heartbeat */
- __dlm_mle_detach_hb_events(dlm, mle);
+ mlog_entry("dlm=%s, dead node=%u\n", dlm->name, dead_node);
+top:
+ assert_spin_locked(&dlm->spinlock);
- /* dump the mle */
- spin_lock(&dlm->master_lock);
- __dlm_put_mle(mle);
- spin_unlock(&dlm->master_lock);
+ /* clean the master list */
+ spin_lock(&dlm->master_lock);
+ for (i = 0; i < DLM_HASH_BUCKETS; i++) {
+ bucket = dlm_master_hash(dlm, i);
+ hlist_for_each(list, bucket) {
+ mle = hlist_entry(list, struct dlm_master_list_entry,
+ master_hash_node);
+
+ BUG_ON(mle->type != DLM_MLE_BLOCK &&
+ mle->type != DLM_MLE_MASTER &&
+ mle->type != DLM_MLE_MIGRATION);
+
+ /* MASTER mles are initiated locally. The waiting
+ * process will notice the node map change shortly.
+ * Let that happen as normal. */
+ if (mle->type == DLM_MLE_MASTER)
+ continue;
+
+ /* BLOCK mles are initiated by other nodes. Need to
+ * clean up if the dead node would have been the
+ * master. */
+ if (mle->type == DLM_MLE_BLOCK) {
+ dlm_clean_block_mle(dlm, mle, dead_node);
+ continue;
+ }
- /* restart */
- goto top;
- }
+ /* Everything else is a MIGRATION mle */
+
+ /* The rule for MIGRATION mles is that the master
+ * becomes UNKNOWN if *either* the original or the new
+ * master dies. All UNKNOWN lockres' are sent to
+ * whichever node becomes the recovery master. The new
+ * master is responsible for determining if there is
+ * still a master for this lockres, or if he needs to
+ * take over mastery. Either way, this node should
+ * expect another message to resolve this. */
+
+ if (mle->master != dead_node &&
+ mle->new_master != dead_node)
+ continue;
+
+ /* If we have reached this point, this mle needs to be
+ * removed from the list and freed. */
+ dlm_clean_migration_mle(dlm, mle);
+
+ mlog(0, "%s: node %u died during migration from "
+ "%u to %u!\n", dlm->name, dead_node, mle->master,
+ mle->new_master);
+
+ /* If we find a lockres associated with the mle, we've
+ * hit this rare case that messes up our lock ordering.
+ * If so, we need to drop the master lock so that we can
+ * take the lockres lock, meaning that we will have to
+ * restart from the head of list. */
+ res = dlm_reset_mleres_owner(dlm, mle);
+ if (res)
+ /* restart */
+ goto top;
- /* this may be the last reference */
- __dlm_put_mle(mle);
+ /* This may be the last reference */
+ __dlm_put_mle(mle);
+ }
}
spin_unlock(&dlm->master_lock);
}
-
int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
u8 old_master)
{
git://ftp.safe.ca / safe / jmp / linux-2.6 / blobdiff