#include "dlmapi.h"
#include "dlmcommon.h"
-#include "dlmdebug.h"
#include "dlmdomain.h"
#define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_MASTER)
wait_queue_head_t wq;
atomic_t woken;
struct kref mle_refs;
+ int inuse;
unsigned long maybe_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
unsigned long vote_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
unsigned long response_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
return 1;
}
-#if 0
-/* Code here is included but defined out as it aids debugging */
+#define dlm_print_nodemap(m) _dlm_print_nodemap(m,#m)
+static void _dlm_print_nodemap(unsigned long *map, const char *mapname)
+{
+ int i;
+ printk("%s=[ ", mapname);
+ for (i=0; i<O2NM_MAX_NODES; i++)
+ if (test_bit(i, map))
+ printk("%d ", i);
+ printk("]");
+}
-void dlm_print_one_mle(struct dlm_master_list_entry *mle)
+static void dlm_print_one_mle(struct dlm_master_list_entry *mle)
{
- int i = 0, refs;
+ int refs;
char *type;
char attached;
u8 master;
unsigned int namelen;
const char *name;
struct kref *k;
+ unsigned long *maybe = mle->maybe_map,
+ *vote = mle->vote_map,
+ *resp = mle->response_map,
+ *node = mle->node_map;
k = &mle->mle_refs;
if (mle->type == DLM_MLE_BLOCK)
name = mle->u.res->lockname.name;
}
- mlog(ML_NOTICE, " #%3d: %3s %3d %3u %3u %c (%d)%.*s\n",
- i, type, refs, master, mle->new_master, attached,
- namelen, namelen, name);
+ mlog(ML_NOTICE, "%.*s: %3s refs=%3d mas=%3u new=%3u evt=%c inuse=%d ",
+ namelen, name, type, refs, master, mle->new_master, attached,
+ mle->inuse);
+ dlm_print_nodemap(maybe);
+ printk(", ");
+ dlm_print_nodemap(vote);
+ printk(", ");
+ dlm_print_nodemap(resp);
+ printk(", ");
+ dlm_print_nodemap(node);
+ printk(", ");
+ printk("\n");
}
+#if 0
+/* Code here is included but defined out as it aids debugging */
+
static void dlm_dump_mles(struct dlm_ctxt *dlm)
{
struct dlm_master_list_entry *mle;
struct list_head *iter;
mlog(ML_NOTICE, "dumping all mles for domain %s:\n", dlm->name);
- mlog(ML_NOTICE, " ####: type refs owner new events? lockname nodemap votemap respmap maybemap\n");
spin_lock(&dlm->master_lock);
list_for_each(iter, &dlm->master_list) {
mle = list_entry(iter, struct dlm_master_list_entry, list);
#endif /* 0 */
-static kmem_cache_t *dlm_mle_cache = NULL;
+static struct kmem_cache *dlm_mle_cache = NULL;
static void dlm_mle_release(struct kref *kref);
static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
struct dlm_lock_resource *res,
u8 target);
+static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
+ struct dlm_lock_resource *res);
int dlm_is_host_down(int errno)
spin_unlock(&dlm->spinlock);
}
+static void dlm_get_mle_inuse(struct dlm_master_list_entry *mle)
+{
+ struct dlm_ctxt *dlm;
+ dlm = mle->dlm;
+
+ assert_spin_locked(&dlm->spinlock);
+ assert_spin_locked(&dlm->master_lock);
+ mle->inuse++;
+ kref_get(&mle->mle_refs);
+}
+
+static void dlm_put_mle_inuse(struct dlm_master_list_entry *mle)
+{
+ struct dlm_ctxt *dlm;
+ dlm = mle->dlm;
+
+ spin_lock(&dlm->spinlock);
+ spin_lock(&dlm->master_lock);
+ mle->inuse--;
+ __dlm_put_mle(mle);
+ spin_unlock(&dlm->master_lock);
+ spin_unlock(&dlm->spinlock);
+
+}
+
/* remove from list and free */
static void __dlm_put_mle(struct dlm_master_list_entry *mle)
{
assert_spin_locked(&dlm->spinlock);
assert_spin_locked(&dlm->master_lock);
- BUG_ON(!atomic_read(&mle->mle_refs.refcount));
-
- kref_put(&mle->mle_refs, dlm_mle_release);
+ if (!atomic_read(&mle->mle_refs.refcount)) {
+ /* this may or may not crash, but who cares.
+ * it's a BUG. */
+ mlog(ML_ERROR, "bad mle: %p\n", mle);
+ dlm_print_one_mle(mle);
+ BUG();
+ } else
+ kref_put(&mle->mle_refs, dlm_mle_release);
}
memset(mle->response_map, 0, sizeof(mle->response_map));
mle->master = O2NM_MAX_NODES;
mle->new_master = O2NM_MAX_NODES;
+ mle->inuse = 0;
if (mle->type == DLM_MLE_MASTER) {
BUG_ON(!res);
mlog(0, "destroying lockres %.*s\n", res->lockname.len,
res->lockname.name);
+ if (!hlist_unhashed(&res->hash_node) ||
+ !list_empty(&res->granted) ||
+ !list_empty(&res->converting) ||
+ !list_empty(&res->blocked) ||
+ !list_empty(&res->dirty) ||
+ !list_empty(&res->recovering) ||
+ !list_empty(&res->purge)) {
+ mlog(ML_ERROR,
+ "Going to BUG for resource %.*s."
+ " We're on a list! [%c%c%c%c%c%c%c]\n",
+ res->lockname.len, res->lockname.name,
+ !hlist_unhashed(&res->hash_node) ? 'H' : ' ',
+ !list_empty(&res->granted) ? 'G' : ' ',
+ !list_empty(&res->converting) ? 'C' : ' ',
+ !list_empty(&res->blocked) ? 'B' : ' ',
+ !list_empty(&res->dirty) ? 'D' : ' ',
+ !list_empty(&res->recovering) ? 'R' : ' ',
+ !list_empty(&res->purge) ? 'P' : ' ');
+
+ dlm_print_one_lock_resource(res);
+ }
+
/* By the time we're ready to blow this guy away, we shouldn't
* be on any lists. */
BUG_ON(!hlist_unhashed(&res->hash_node));
kfree(res);
}
-void dlm_lockres_get(struct dlm_lock_resource *res)
-{
- kref_get(&res->refs);
-}
-
void dlm_lockres_put(struct dlm_lock_resource *res)
{
kref_put(&res->refs, dlm_lockres_release);
memcpy(qname, name, namelen);
res->lockname.len = namelen;
- res->lockname.hash = full_name_hash(name, namelen);
+ res->lockname.hash = dlm_lockid_hash(name, namelen);
init_waitqueue_head(&res->wq);
spin_lock_init(&res->spinlock);
{
struct dlm_lock_resource *res;
- res = kmalloc(sizeof(struct dlm_lock_resource), GFP_KERNEL);
+ res = kmalloc(sizeof(struct dlm_lock_resource), GFP_NOFS);
if (!res)
return NULL;
- res->lockname.name = kmalloc(namelen, GFP_KERNEL);
+ res->lockname.name = kmalloc(namelen, GFP_NOFS);
if (!res->lockname.name) {
kfree(res);
return NULL;
*/
struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
const char *lockid,
+ int namelen,
int flags)
{
struct dlm_lock_resource *tmpres=NULL, *res=NULL;
int blocked = 0;
int ret, nodenum;
struct dlm_node_iter iter;
- unsigned int namelen;
+ unsigned int hash;
int tries = 0;
+ int bit, wait_on_recovery = 0;
BUG_ON(!lockid);
- namelen = strlen(lockid);
+ hash = dlm_lockid_hash(lockid, namelen);
mlog(0, "get lockres %s (len %d)\n", lockid, namelen);
lookup:
spin_lock(&dlm->spinlock);
- tmpres = __dlm_lookup_lockres(dlm, lockid, namelen);
+ tmpres = __dlm_lookup_lockres(dlm, lockid, namelen, hash);
if (tmpres) {
spin_unlock(&dlm->spinlock);
mlog(0, "found in hash!\n");
mlog(0, "allocating a new resource\n");
/* nothing found and we need to allocate one. */
alloc_mle = (struct dlm_master_list_entry *)
- kmem_cache_alloc(dlm_mle_cache, GFP_KERNEL);
+ kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
if (!alloc_mle)
goto leave;
res = dlm_new_lockres(dlm, lockid, namelen);
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);
+
+ /* still holding the dlm spinlock, check the recovery map
+ * to see if there are any nodes that still need to be
+ * considered. these will not appear in the mle nodemap
+ * but they might own this lockres. wait on them. */
+ bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
+ if (bit < O2NM_MAX_NODES) {
+ mlog(ML_NOTICE, "%s:%.*s: at least one node (%d) to"
+ "recover before lock mastery can begin\n",
+ dlm->name, namelen, (char *)lockid, bit);
+ wait_on_recovery = 1;
+ }
}
/* at this point there is either a DLM_MLE_BLOCK or a
* if so, the creator of the BLOCK may try to put the last
* ref at this time in the assert master handler, so we
* need an extra one to keep from a bad ptr deref. */
- dlm_get_mle(mle);
+ dlm_get_mle_inuse(mle);
spin_unlock(&dlm->master_lock);
spin_unlock(&dlm->spinlock);
+redo_request:
+ while (wait_on_recovery) {
+ /* any cluster changes that occurred after dropping the
+ * dlm spinlock would be detectable be a change on the mle,
+ * so we only need to clear out the recovery map once. */
+ if (dlm_is_recovery_lock(lockid, namelen)) {
+ mlog(ML_NOTICE, "%s: recovery map is not empty, but "
+ "must master $RECOVERY lock now\n", dlm->name);
+ if (!dlm_pre_master_reco_lockres(dlm, res))
+ wait_on_recovery = 0;
+ else {
+ mlog(0, "%s: waiting 500ms for heartbeat state "
+ "change\n", dlm->name);
+ msleep(500);
+ }
+ continue;
+ }
+
+ dlm_kick_recovery_thread(dlm);
+ msleep(1000);
+ dlm_wait_for_recovery(dlm);
+
+ spin_lock(&dlm->spinlock);
+ bit = find_next_bit(dlm->recovery_map, O2NM_MAX_NODES, 0);
+ if (bit < O2NM_MAX_NODES) {
+ mlog(ML_NOTICE, "%s:%.*s: at least one node (%d) to"
+ "recover before lock mastery can begin\n",
+ dlm->name, namelen, (char *)lockid, bit);
+ wait_on_recovery = 1;
+ } else
+ wait_on_recovery = 0;
+ spin_unlock(&dlm->spinlock);
+
+ if (wait_on_recovery)
+ dlm_wait_for_node_recovery(dlm, bit, 10000);
+ }
+
/* must wait for lock to be mastered elsewhere */
if (blocked)
goto wait;
-redo_request:
ret = -EINVAL;
dlm_node_iter_init(mle->vote_map, &iter);
while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
mlog_errno(ret);
if (mle->master != O2NM_MAX_NODES) {
/* found a master ! */
- break;
+ if (mle->master <= nodenum)
+ break;
+ /* if our master request has not reached the master
+ * yet, keep going until it does. this is how the
+ * master will know that asserts are needed back to
+ * the lower nodes. */
+ mlog(0, "%s:%.*s: requests only up to %u but master "
+ "is %u, keep going\n", dlm->name, namelen,
+ lockid, nodenum, mle->master);
}
}
/* keep going until the response map includes all nodes */
ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked);
if (ret < 0) {
+ wait_on_recovery = 1;
mlog(0, "%s:%.*s: node map changed, redo the "
"master request now, blocked=%d\n",
dlm->name, res->lockname.len,
dlm->name, res->lockname.len,
res->lockname.name, blocked);
dlm_print_one_lock_resource(res);
- /* dlm_print_one_mle(mle); */
+ dlm_print_one_mle(mle);
tries = 0;
}
goto redo_request;
dlm_mle_detach_hb_events(dlm, mle);
dlm_put_mle(mle);
/* put the extra ref */
- dlm_put_mle(mle);
+ dlm_put_mle_inuse(mle);
wake_waiters:
spin_lock(&res->spinlock);
/* check if another node has already become the owner */
spin_lock(&res->spinlock);
if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
+ mlog(0, "%s:%.*s: owner is suddenly %u\n", dlm->name,
+ res->lockname.len, res->lockname.name, res->owner);
spin_unlock(&res->spinlock);
+ /* this will cause the master to re-assert across
+ * the whole cluster, freeing up mles */
+ if (res->owner != dlm->node_num) {
+ ret = dlm_do_master_request(mle, res->owner);
+ if (ret < 0) {
+ /* give recovery a chance to run */
+ mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret);
+ msleep(500);
+ goto recheck;
+ }
+ }
+ ret = 0;
goto leave;
}
spin_unlock(&res->spinlock);
"rechecking now\n", dlm->name, res->lockname.len,
res->lockname.name);
goto recheck;
+ } else {
+ if (!voting_done) {
+ mlog(0, "map not changed and voting not done "
+ "for %s:%.*s\n", dlm->name, res->lockname.len,
+ res->lockname.name);
+ }
}
if (m != O2NM_MAX_NODES) {
set_bit(node, mle->vote_map);
} else {
mlog(ML_ERROR, "node down! %d\n", node);
-
- /* if the node wasn't involved in mastery skip it,
- * but clear it out from the maps so that it will
- * not affect mastery of this lockres */
- clear_bit(node, mle->response_map);
- clear_bit(node, mle->vote_map);
- if (!test_bit(node, mle->maybe_map))
- goto next;
-
- /* if we're already blocked on lock mastery, and the
- * dead node wasn't the expected master, or there is
- * another node in the maybe_map, keep waiting */
if (blocked) {
int lowest = find_next_bit(mle->maybe_map,
O2NM_MAX_NODES, 0);
/* act like it was never there */
clear_bit(node, mle->maybe_map);
- if (node != lowest)
- goto next;
-
- mlog(ML_ERROR, "expected master %u died while "
- "this node was blocked waiting on it!\n",
- node);
- lowest = find_next_bit(mle->maybe_map,
- O2NM_MAX_NODES,
- lowest+1);
- if (lowest < O2NM_MAX_NODES) {
- mlog(0, "still blocked. waiting "
- "on %u now\n", lowest);
- goto next;
+ if (node == lowest) {
+ mlog(0, "expected master %u died"
+ " while this node was blocked "
+ "waiting on it!\n", node);
+ lowest = find_next_bit(mle->maybe_map,
+ O2NM_MAX_NODES,
+ lowest+1);
+ if (lowest < O2NM_MAX_NODES) {
+ mlog(0, "%s:%.*s:still "
+ "blocked. waiting on %u "
+ "now\n", dlm->name,
+ res->lockname.len,
+ res->lockname.name,
+ lowest);
+ } else {
+ /* mle is an MLE_BLOCK, but
+ * there is now nothing left to
+ * block on. we need to return
+ * all the way back out and try
+ * again with an MLE_MASTER.
+ * dlm_do_local_recovery_cleanup
+ * has already run, so the mle
+ * refcount is ok */
+ mlog(0, "%s:%.*s: no "
+ "longer blocking. try to "
+ "master this here\n",
+ dlm->name,
+ res->lockname.len,
+ res->lockname.name);
+ mle->type = DLM_MLE_MASTER;
+ mle->u.res = res;
+ }
}
-
- /* mle is an MLE_BLOCK, but there is now
- * nothing left to block on. we need to return
- * all the way back out and try again with
- * an MLE_MASTER. dlm_do_local_recovery_cleanup
- * has already run, so the mle refcount is ok */
- mlog(0, "no longer blocking. we can "
- "try to master this here\n");
- mle->type = DLM_MLE_MASTER;
- memset(mle->maybe_map, 0,
- sizeof(mle->maybe_map));
- memset(mle->response_map, 0,
- sizeof(mle->maybe_map));
- memcpy(mle->vote_map, mle->node_map,
- sizeof(mle->node_map));
- mle->u.res = res;
- set_bit(dlm->node_num, mle->maybe_map);
-
- ret = -EAGAIN;
- goto next;
}
- clear_bit(node, mle->maybe_map);
- if (node > dlm->node_num)
- goto next;
-
- mlog(0, "dead node in map!\n");
- /* yuck. go back and re-contact all nodes
- * in the vote_map, removing this node. */
- memset(mle->response_map, 0,
- sizeof(mle->response_map));
+ /* now blank out everything, as if we had never
+ * contacted anyone */
+ memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
+ memset(mle->response_map, 0, sizeof(mle->response_map));
+ /* reset the vote_map to the current node_map */
+ memcpy(mle->vote_map, mle->node_map,
+ sizeof(mle->node_map));
+ /* put myself into the maybe map */
+ if (mle->type != DLM_MLE_BLOCK)
+ set_bit(dlm->node_num, mle->maybe_map);
}
ret = -EAGAIN;
-next:
node = dlm_bitmap_diff_iter_next(&bdi, &sc);
}
return ret;
{
u8 response = DLM_MASTER_RESP_MAYBE;
struct dlm_ctxt *dlm = data;
- struct dlm_lock_resource *res;
+ struct dlm_lock_resource *res = NULL;
struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
char *name;
- unsigned int namelen;
+ unsigned int namelen, hash;
int found, ret;
int set_maybe;
+ int dispatch_assert = 0;
if (!dlm_grab(dlm))
return DLM_MASTER_RESP_NO;
name = request->name;
namelen = request->namelen;
+ hash = dlm_lockid_hash(name, namelen);
if (namelen > DLM_LOCKID_NAME_MAX) {
response = DLM_IVBUFLEN;
way_up_top:
spin_lock(&dlm->spinlock);
- res = __dlm_lookup_lockres(dlm, name, namelen);
+ res = __dlm_lookup_lockres(dlm, name, namelen, hash);
if (res) {
spin_unlock(&dlm->spinlock);
}
if (res->owner == dlm->node_num) {
- u32 flags = DLM_ASSERT_MASTER_MLE_CLEANUP;
spin_unlock(&res->spinlock);
// mlog(0, "this node is the master\n");
response = DLM_MASTER_RESP_YES;
* caused all nodes up to this one to
* create mles. this node now needs to
* go back and clean those up. */
- mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
- dlm->node_num, res->lockname.len, res->lockname.name);
- ret = dlm_dispatch_assert_master(dlm, res, 1,
- request->node_idx,
- flags);
- if (ret < 0) {
- mlog(ML_ERROR, "failed to dispatch assert "
- "master work\n");
- response = DLM_MASTER_RESP_ERROR;
- }
+ dispatch_assert = 1;
goto send_response;
} else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
spin_unlock(&res->spinlock);
}
} else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
set_maybe = 0;
- if (tmpmle->master == dlm->node_num)
+ if (tmpmle->master == dlm->node_num) {
response = DLM_MASTER_RESP_YES;
- else
+ /* this node will be the owner.
+ * go back and clean the mles on any
+ * other nodes */
+ dispatch_assert = 1;
+ } else
response = DLM_MASTER_RESP_NO;
} else {
// mlog(0, "this node is attempting to "
spin_unlock(&dlm->spinlock);
mle = (struct dlm_master_list_entry *)
- kmem_cache_alloc(dlm_mle_cache, GFP_KERNEL);
+ kmem_cache_alloc(dlm_mle_cache, GFP_NOFS);
if (!mle) {
- // bad bad bad... this sucks.
response = DLM_MASTER_RESP_ERROR;
+ mlog_errno(-ENOMEM);
goto send_response;
}
- spin_lock(&dlm->spinlock);
- dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL,
- name, namelen);
- spin_unlock(&dlm->spinlock);
goto way_up_top;
}
// mlog(0, "this is second time thru, already allocated, "
// "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);
response = DLM_MASTER_RESP_NO;
// mlog(0, "mle was found\n");
set_maybe = 1;
spin_lock(&tmpmle->spinlock);
+ if (tmpmle->master == dlm->node_num) {
+ mlog(ML_ERROR, "no lockres, but an mle with this node as master!\n");
+ BUG();
+ }
if (tmpmle->type == DLM_MLE_BLOCK)
response = DLM_MASTER_RESP_NO;
else if (tmpmle->type == DLM_MLE_MIGRATION) {
mlog(0, "migration mle was found (%u->%u)\n",
tmpmle->master, tmpmle->new_master);
- if (tmpmle->master == dlm->node_num) {
- mlog(ML_ERROR, "no lockres, but migration mle "
- "says that this node is master!\n");
- BUG();
- }
/* real master can respond on its own */
response = DLM_MASTER_RESP_NO;
- } else {
- if (tmpmle->master == dlm->node_num) {
- response = DLM_MASTER_RESP_YES;
- set_maybe = 0;
- } else
- response = DLM_MASTER_RESP_MAYBE;
- }
+ } else
+ response = DLM_MASTER_RESP_MAYBE;
if (set_maybe)
set_bit(request->node_idx, tmpmle->maybe_map);
spin_unlock(&tmpmle->spinlock);
dlm_put_mle(tmpmle);
}
send_response:
+
+ if (dispatch_assert) {
+ if (response != DLM_MASTER_RESP_YES)
+ mlog(ML_ERROR, "invalid response %d\n", response);
+ if (!res) {
+ mlog(ML_ERROR, "bad lockres while trying to assert!\n");
+ BUG();
+ }
+ mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
+ dlm->node_num, res->lockname.len, res->lockname.name);
+ ret = dlm_dispatch_assert_master(dlm, res, 0, request->node_idx,
+ DLM_ASSERT_MASTER_MLE_CLEANUP);
+ if (ret < 0) {
+ mlog(ML_ERROR, "failed to dispatch assert master work\n");
+ response = DLM_MASTER_RESP_ERROR;
+ }
+ }
+
dlm_put(dlm);
return response;
}
int to, tmpret;
struct dlm_node_iter iter;
int ret = 0;
+ int reassert;
BUG_ON(namelen > O2NM_MAX_NAME_LEN);
+again:
+ reassert = 0;
/* note that if this nodemap is empty, it returns 0 */
dlm_node_iter_init(nodemap, &iter);
while ((to = dlm_node_iter_next(&iter)) >= 0) {
int r = 0;
+ struct dlm_master_list_entry *mle = NULL;
+
mlog(0, "sending assert master to %d (%.*s)\n", to,
namelen, lockname);
memset(&assert, 0, sizeof(assert));
tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
&assert, sizeof(assert), to, &r);
if (tmpret < 0) {
- mlog(ML_ERROR, "assert_master returned %d!\n", tmpret);
+ mlog(0, "assert_master returned %d!\n", tmpret);
if (!dlm_is_host_down(tmpret)) {
- mlog(ML_ERROR, "unhandled error!\n");
+ mlog(ML_ERROR, "unhandled error=%d!\n", tmpret);
BUG();
}
/* a node died. finish out the rest of the nodes. */
- mlog(ML_ERROR, "link to %d went down!\n", to);
+ mlog(0, "link to %d went down!\n", to);
/* any nonzero status return will do */
ret = tmpret;
} else if (r < 0) {
/* ok, something horribly messed. kill thyself. */
mlog(ML_ERROR,"during assert master of %.*s to %u, "
"got %d.\n", namelen, lockname, to, r);
- dlm_dump_lock_resources(dlm);
+ spin_lock(&dlm->spinlock);
+ spin_lock(&dlm->master_lock);
+ if (dlm_find_mle(dlm, &mle, (char *)lockname,
+ namelen)) {
+ dlm_print_one_mle(mle);
+ __dlm_put_mle(mle);
+ }
+ spin_unlock(&dlm->master_lock);
+ spin_unlock(&dlm->spinlock);
BUG();
+ } else if (r == EAGAIN) {
+ mlog(0, "%.*s: node %u create mles on other "
+ "nodes and requests a re-assert\n",
+ namelen, lockname, to);
+ reassert = 1;
}
}
+ if (reassert)
+ goto again;
+
return ret;
}
struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
struct dlm_lock_resource *res = NULL;
char *name;
- unsigned int namelen;
+ unsigned int namelen, hash;
u32 flags;
+ int master_request = 0;
+ int ret = 0;
if (!dlm_grab(dlm))
return 0;
name = assert->name;
namelen = assert->namelen;
+ hash = dlm_lockid_hash(name, namelen);
flags = be32_to_cpu(assert->flags);
if (namelen > DLM_LOCKID_NAME_MAX) {
if (bit >= O2NM_MAX_NODES) {
/* not necessarily an error, though less likely.
* could be master just re-asserting. */
- mlog(ML_ERROR, "no bits set in the maybe_map, but %u "
+ mlog(0, "no bits set in the maybe_map, but %u "
"is asserting! (%.*s)\n", assert->node_idx,
namelen, name);
} else if (bit != assert->node_idx) {
* number winning the mastery will respond
* YES to mastery requests, but this node
* had no way of knowing. let it pass. */
- mlog(ML_ERROR, "%u is the lowest node, "
+ mlog(0, "%u is the lowest node, "
"%u is asserting. (%.*s) %u must "
"have begun after %u won.\n", bit,
assert->node_idx, namelen, name, bit,
assert->node_idx);
}
}
+ if (mle->type == DLM_MLE_MIGRATION) {
+ if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
+ mlog(0, "%s:%.*s: got cleanup assert"
+ " from %u for migration\n",
+ dlm->name, namelen, name,
+ assert->node_idx);
+ } else if (!(flags & DLM_ASSERT_MASTER_FINISH_MIGRATION)) {
+ mlog(0, "%s:%.*s: got unrelated assert"
+ " from %u for migration, ignoring\n",
+ dlm->name, namelen, name,
+ assert->node_idx);
+ __dlm_put_mle(mle);
+ spin_unlock(&dlm->master_lock);
+ spin_unlock(&dlm->spinlock);
+ goto done;
+ }
+ }
}
spin_unlock(&dlm->master_lock);
/* ok everything checks out with the MLE
* now check to see if there is a lockres */
- res = __dlm_lookup_lockres(dlm, name, namelen);
+ res = __dlm_lookup_lockres(dlm, name, namelen, hash);
if (res) {
spin_lock(&res->spinlock);
if (res->state & DLM_LOCK_RES_RECOVERING) {
goto kill;
}
if (!mle) {
- if (res->owner != assert->node_idx) {
+ 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",
// mlog(0, "woo! got an assert_master from node %u!\n",
// assert->node_idx);
if (mle) {
- int extra_ref;
+ int extra_ref = 0;
+ int nn = -1;
+ int rr, err = 0;
spin_lock(&mle->spinlock);
- extra_ref = !!(mle->type == DLM_MLE_BLOCK
- || mle->type == DLM_MLE_MIGRATION);
+ if (mle->type == DLM_MLE_BLOCK || mle->type == DLM_MLE_MIGRATION)
+ extra_ref = 1;
+ else {
+ /* MASTER mle: if any bits set in the response map
+ * then the calling node needs to re-assert to clear
+ * up nodes that this node contacted */
+ while ((nn = find_next_bit (mle->response_map, O2NM_MAX_NODES,
+ nn+1)) < O2NM_MAX_NODES) {
+ if (nn != dlm->node_num && nn != assert->node_idx)
+ master_request = 1;
+ }
+ }
mle->master = assert->node_idx;
atomic_set(&mle->woken, 1);
wake_up(&mle->wq);
spin_unlock(&mle->spinlock);
- if (mle->type == DLM_MLE_MIGRATION && res) {
- mlog(0, "finishing off migration of lockres %.*s, "
- "from %u to %u\n",
- res->lockname.len, res->lockname.name,
- dlm->node_num, mle->new_master);
+ if (res) {
spin_lock(&res->spinlock);
- res->state &= ~DLM_LOCK_RES_MIGRATING;
- dlm_change_lockres_owner(dlm, res, mle->new_master);
- BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
+ if (mle->type == DLM_MLE_MIGRATION) {
+ mlog(0, "finishing off migration of lockres %.*s, "
+ "from %u to %u\n",
+ res->lockname.len, res->lockname.name,
+ dlm->node_num, mle->new_master);
+ res->state &= ~DLM_LOCK_RES_MIGRATING;
+ dlm_change_lockres_owner(dlm, res, mle->new_master);
+ BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
+ } else {
+ dlm_change_lockres_owner(dlm, res, mle->master);
+ }
spin_unlock(&res->spinlock);
}
- /* master is known, detach if not already detached */
- dlm_mle_detach_hb_events(dlm, mle);
- dlm_put_mle(mle);
-
+
+ /* master is known, detach if not already detached.
+ * ensures that only one assert_master call will happen
+ * on this mle. */
+ spin_lock(&dlm->spinlock);
+ spin_lock(&dlm->master_lock);
+
+ rr = atomic_read(&mle->mle_refs.refcount);
+ if (mle->inuse > 0) {
+ if (extra_ref && rr < 3)
+ err = 1;
+ else if (!extra_ref && rr < 2)
+ err = 1;
+ } else {
+ if (extra_ref && rr < 2)
+ err = 1;
+ else if (!extra_ref && rr < 1)
+ err = 1;
+ }
+ if (err) {
+ mlog(ML_ERROR, "%s:%.*s: got assert master from %u "
+ "that will mess up this node, refs=%d, extra=%d, "
+ "inuse=%d\n", dlm->name, namelen, name,
+ assert->node_idx, rr, extra_ref, mle->inuse);
+ dlm_print_one_mle(mle);
+ }
+ list_del_init(&mle->list);
+ __dlm_mle_detach_hb_events(dlm, mle);
+ __dlm_put_mle(mle);
if (extra_ref) {
/* the assert master message now balances the extra
* ref given by the master / migration request message.
* if this is the last put, it will be removed
* from the list. */
- dlm_put_mle(mle);
+ __dlm_put_mle(mle);
+ }
+ spin_unlock(&dlm->master_lock);
+ spin_unlock(&dlm->spinlock);
+ } else if (res) {
+ if (res->owner != assert->node_idx) {
+ mlog(0, "assert_master from %u, but current "
+ "owner is %u (%.*s), no mle\n", assert->node_idx,
+ res->owner, namelen, name);
}
}
done:
+ ret = 0;
if (res)
dlm_lockres_put(res);
dlm_put(dlm);
- return 0;
+ if (master_request) {
+ mlog(0, "need to tell master to reassert\n");
+ ret = EAGAIN; // positive. negative would shoot down the node.
+ }
+ return ret;
kill:
/* kill the caller! */
+ mlog(ML_ERROR, "Bad message received from another node. Dumping state "
+ "and killing the other node now! This node is OK and can continue.\n");
+ __dlm_print_one_lock_resource(res);
spin_unlock(&res->spinlock);
spin_unlock(&dlm->spinlock);
dlm_lockres_put(res);
- mlog(ML_ERROR, "Bad message received from another node. Dumping state "
- "and killing the other node now! This node is OK and can continue.\n");
- dlm_dump_lock_resources(dlm);
dlm_put(dlm);
return -EINVAL;
}
int ignore_higher, u8 request_from, u32 flags)
{
struct dlm_work_item *item;
- item = kcalloc(1, sizeof(*item), GFP_KERNEL);
+ item = kzalloc(sizeof(*item), GFP_NOFS);
if (!item)
return -ENOMEM;
item->u.am.request_from = request_from;
item->u.am.flags = flags;
+ if (ignore_higher)
+ mlog(0, "IGNORE HIGHER: %.*s\n", res->lockname.len,
+ res->lockname.name);
+
spin_lock(&dlm->work_lock);
list_add_tail(&item->list, &dlm->work_list);
spin_unlock(&dlm->work_lock);
- schedule_work(&dlm->dispatched_work);
+ queue_work(dlm->dlm_worker, &dlm->dispatched_work);
return 0;
}
}
}
+ /*
+ * If we're migrating this lock to someone else, we are no
+ * longer allowed to assert out own mastery. OTOH, we need to
+ * prevent migration from starting while we're still asserting
+ * our dominance. The reserved ast delays migration.
+ */
+ spin_lock(&res->spinlock);
+ if (res->state & DLM_LOCK_RES_MIGRATING) {
+ mlog(0, "Someone asked us to assert mastery, but we're "
+ "in the middle of migration. Skipping assert, "
+ "the new master will handle that.\n");
+ spin_unlock(&res->spinlock);
+ goto put;
+ } else
+ __dlm_lockres_reserve_ast(res);
+ spin_unlock(&res->spinlock);
+
/* this call now finishes out the nodemap
* even if one or more nodes die */
mlog(0, "worker about to master %.*s here, this=%u\n",
nodemap, flags);
if (ret < 0) {
/* no need to restart, we are done */
- mlog_errno(ret);
+ if (!dlm_is_host_down(ret))
+ mlog_errno(ret);
}
+ /* Ok, we've asserted ourselves. Let's let migration start. */
+ dlm_lockres_release_ast(dlm, res);
+
+put:
dlm_lockres_put(res);
mlog(0, "finished with dlm_assert_master_worker\n");
}
+/* SPECIAL CASE for the $RECOVERY lock used by the recovery thread.
+ * We cannot wait for node recovery to complete to begin mastering this
+ * lockres because this lockres is used to kick off recovery! ;-)
+ * So, do a pre-check on all living nodes to see if any of those nodes
+ * think that $RECOVERY is currently mastered by a dead node. If so,
+ * we wait a short time to allow that node to get notified by its own
+ * heartbeat stack, then check again. All $RECOVERY lock resources
+ * mastered by dead nodes are purged when the hearbeat callback is
+ * fired, so we can know for sure that it is safe to continue once
+ * the node returns a live node or no node. */
+static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
+ struct dlm_lock_resource *res)
+{
+ struct dlm_node_iter iter;
+ int nodenum;
+ int ret = 0;
+ u8 master = DLM_LOCK_RES_OWNER_UNKNOWN;
+
+ spin_lock(&dlm->spinlock);
+ dlm_node_iter_init(dlm->domain_map, &iter);
+ spin_unlock(&dlm->spinlock);
+
+ while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
+ /* do not send to self */
+ if (nodenum == dlm->node_num)
+ continue;
+ ret = dlm_do_master_requery(dlm, res, nodenum, &master);
+ if (ret < 0) {
+ mlog_errno(ret);
+ if (!dlm_is_host_down(ret))
+ BUG();
+ /* host is down, so answer for that node would be
+ * DLM_LOCK_RES_OWNER_UNKNOWN. continue. */
+ ret = 0;
+ }
+
+ if (master != DLM_LOCK_RES_OWNER_UNKNOWN) {
+ /* check to see if this master is in the recovery map */
+ spin_lock(&dlm->spinlock);
+ if (test_bit(master, dlm->recovery_map)) {
+ mlog(ML_NOTICE, "%s: node %u has not seen "
+ "node %u go down yet, and thinks the "
+ "dead node is mastering the recovery "
+ "lock. must wait.\n", dlm->name,
+ nodenum, master);
+ ret = -EAGAIN;
+ }
+ spin_unlock(&dlm->spinlock);
+ mlog(0, "%s: reco lock master is %u\n", dlm->name,
+ master);
+ break;
+ }
+ }
+ return ret;
+}
+
/*
* DLM_MIGRATE_LOCKRES
*/
ret = -ENOMEM;
- mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_KERNEL);
+ mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_NOFS);
if (!mres) {
mlog_errno(ret);
goto leave;
}
mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache,
- GFP_KERNEL);
+ GFP_NOFS);
if (!mle) {
mlog_errno(ret);
goto leave;
* take both dlm->spinlock and dlm->master_lock */
spin_lock(&dlm->spinlock);
spin_lock(&dlm->master_lock);
- dlm_get_mle(mle);
+ dlm_get_mle_inuse(mle);
spin_unlock(&dlm->master_lock);
spin_unlock(&dlm->spinlock);
/* migration failed, detach and clean up mle */
dlm_mle_detach_hb_events(dlm, mle);
dlm_put_mle(mle);
- dlm_put_mle(mle);
+ dlm_put_mle_inuse(mle);
+ spin_lock(&res->spinlock);
+ res->state &= ~DLM_LOCK_RES_MIGRATING;
+ spin_unlock(&res->spinlock);
goto leave;
}
/* avoid hang during shutdown when migrating lockres
* to a node which also goes down */
if (dlm_is_node_dead(dlm, target)) {
- mlog(0, "%s:%.*s: expected migration target %u "
- "is no longer up. restarting.\n",
+ mlog(0, "%s:%.*s: expected migration "
+ "target %u is no longer up, restarting\n",
dlm->name, res->lockname.len,
res->lockname.name, target);
ret = -ERESTARTSYS;
/* migration failed, detach and clean up mle */
dlm_mle_detach_hb_events(dlm, mle);
dlm_put_mle(mle);
- dlm_put_mle(mle);
+ dlm_put_mle_inuse(mle);
+ spin_lock(&res->spinlock);
+ res->state &= ~DLM_LOCK_RES_MIGRATING;
+ spin_unlock(&res->spinlock);
goto leave;
}
/* TODO: if node died: stop, clean up, return error */
/* master is known, detach if not already detached */
dlm_mle_detach_hb_events(dlm, mle);
- dlm_put_mle(mle);
+ dlm_put_mle_inuse(mle);
ret = 0;
dlm_lockres_calc_usage(dlm, res);
mlog(0, "returning %d\n", ret);
return ret;
}
-EXPORT_SYMBOL_GPL(dlm_migrate_lockres);
int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
{
struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf;
struct dlm_master_list_entry *mle = NULL, *oldmle = NULL;
const char *name;
- unsigned int namelen;
+ unsigned int namelen, hash;
int ret = 0;
if (!dlm_grab(dlm))
name = migrate->name;
namelen = migrate->namelen;
+ hash = dlm_lockid_hash(name, namelen);
/* preallocate.. if this fails, abort */
mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache,
- GFP_KERNEL);
+ GFP_NOFS);
if (!mle) {
ret = -ENOMEM;
/* check for pre-existing lock */
spin_lock(&dlm->spinlock);
- res = __dlm_lookup_lockres(dlm, name, namelen);
+ res = __dlm_lookup_lockres(dlm, name, namelen, hash);
spin_lock(&dlm->master_lock);
if (res) {
/* remove it from the list so that only one
* mle will be found */
list_del_init(&tmp->list);
+ __dlm_mle_detach_hb_events(dlm, mle);
}
spin_unlock(&tmp->spinlock);
}
struct list_head *iter, *iter2;
struct dlm_master_list_entry *mle;
struct dlm_lock_resource *res;
+ unsigned int hash;
mlog_entry("dlm=%s, dead node=%u\n", dlm->name, dead_node);
top:
* 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(ML_ERROR, "node %u was expected master\n",
+ mlog(0, "node %u was expected master\n",
dead_node);
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);
atomic_set(&mle->woken, 1);
spin_unlock(&mle->spinlock);
wake_up(&mle->wq);
- mlog(0, "node %u died during migration from "
- "%u to %u!\n", dead_node,
+ 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);
+ mle->u.name.len, hash);
if (res) {
/* unfortunately if we hit this rare case, our
* lock ordering is messed. we need to drop