wengang wang
2009-May-01 04:55 UTC
[Ocfs2-devel] [PATCH 1/3] OCFS2: speed up dlm_lockr_resouce hash_table lookups
use multiple pages for the hash table.
mainline git commit: 03d864c02c3ea803b1718940ac6953a257182d7a
Authored-by: Daniel Phillips <phillips at google.com>
Signed-off-by: Wengang Wang <wen.gang.wang at oracle.com>
--
Index: ocfs2-1.2/fs/ocfs2/dlm/dlmdomain.c
==================================================================---
ocfs2-1.2/fs/ocfs2/dlm/dlmdomain.c (revision 1)
+++ ocfs2-1.2/fs/ocfs2/dlm/dlmdomain.c (revision 2)
@@ -79,6 +79,33 @@ static inline void byte_copymap(u8 dmap[
byte_set_bit(nn, dmap);
}
+static void dlm_free_pagevec(void **vec, int pages)
+{
+ while (pages--)
+ free_page((unsigned long)vec[pages]);
+ kfree(vec);
+}
+
+static void **dlm_alloc_pagevec(int pages)
+{
+ void **vec = kmalloc(pages * sizeof(void *), GFP_KERNEL);
+ int i;
+
+ if (!vec)
+ return NULL;
+
+ for (i = 0; i < pages; i++) {
+ vec[i] = (void *)__get_free_page(GFP_KERNEL);
+ if (!vec[i])
+ goto out_free;
+ }
+ return vec;
+out_free:
+ dlm_free_pagevec(vec, i);
+ return NULL;
+
+}
+
/*
*
* spinlock lock ordering: if multiple locks are needed, obey this ordering:
@@ -127,7 +154,7 @@ void __dlm_insert_lockres(struct dlm_ctx
q = &res->lockname;
q->hash = full_name_hash(q->name, q->len);
- bucket = &(dlm->lockres_hash[q->hash % DLM_HASH_BUCKETS]);
+ bucket = dlm_lockres_hash(dlm, q->hash);
/* get a reference for our hashtable */
dlm_lockres_get(res);
@@ -151,7 +178,7 @@ struct dlm_lock_resource * __dlm_lookup_
hash = full_name_hash(name, len);
- bucket = &(dlm->lockres_hash[hash % DLM_HASH_BUCKETS]);
+ bucket = dlm_lockres_hash(dlm, hash);
/* check for pre-existing lock */
hlist_for_each(iter, bucket) {
@@ -262,7 +289,7 @@ static void dlm_free_ctxt_mem(struct dlm
dlm_proc_del_domain(dlm);
if (dlm->lockres_hash)
- free_page((unsigned long) dlm->lockres_hash);
+ dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES);
if (dlm->name)
kfree(dlm->name);
@@ -394,7 +421,8 @@ static int dlm_migrate_all_locks(struct
for (i = 0; i < DLM_HASH_BUCKETS; i++) {
redo_bucket:
n = 0;
- bucket = &dlm->lockres_hash[i];
+ bucket = dlm_lockres_hash(dlm, i);
+ dlm_lockres_hash(dlm, i);
iter = bucket->first;
while (iter) {
n++;
@@ -1356,7 +1384,8 @@ static struct dlm_ctxt *dlm_alloc_ctxt(c
goto leave;
}
- dlm->lockres_hash = (struct hlist_head *) __get_free_page(GFP_KERNEL);
+ dlm->lockres_hash = (struct hlist_head **)
+ dlm_alloc_pagevec(DLM_HASH_PAGES);
if (!dlm->lockres_hash) {
mlog_errno(-ENOMEM);
kfree(dlm->name);
@@ -1365,8 +1394,8 @@ static struct dlm_ctxt *dlm_alloc_ctxt(c
goto leave;
}
- for (i=0; i<DLM_HASH_BUCKETS; i++)
- INIT_HLIST_HEAD(&dlm->lockres_hash[i]);
+ for (i = 0; i < DLM_HASH_BUCKETS; i++)
+ INIT_HLIST_HEAD(dlm_lockres_hash(dlm, i));
strcpy(dlm->name, domain);
dlm->key = key;
Index: ocfs2-1.2/fs/ocfs2/dlm/dlmdebug.c
==================================================================---
ocfs2-1.2/fs/ocfs2/dlm/dlmdebug.c (revision 1)
+++ ocfs2-1.2/fs/ocfs2/dlm/dlmdebug.c (revision 2)
@@ -547,7 +547,7 @@ void dlm_dump_lock_resources(struct dlm_
spin_lock(&dlm->spinlock);
for (i=0; i<DLM_HASH_BUCKETS; i++) {
- bucket = &(dlm->lockres_hash[i]);
+ bucket = dlm_lockres_hash(dlm, i);
hlist_for_each_entry(res, iter, bucket, hash_node)
dlm_print_one_lock_resource(res);
}
Index: ocfs2-1.2/fs/ocfs2/dlm/dlmcommon.h
==================================================================---
ocfs2-1.2/fs/ocfs2/dlm/dlmcommon.h (revision 1)
+++ ocfs2-1.2/fs/ocfs2/dlm/dlmcommon.h (revision 2)
@@ -37,7 +37,10 @@
#define DLM_THREAD_SHUFFLE_INTERVAL 5 // flush everything every 5 passes
#define DLM_THREAD_MS 200 // flush at least every 200 ms
-#define DLM_HASH_BUCKETS (PAGE_SIZE / sizeof(struct hlist_head))
+#define DLM_HASH_SIZE (1 << 14)
+#define DLM_HASH_PAGES (DLM_HASH_SIZE / PAGE_SIZE)
+#define DLM_BUCKETS_PER_PAGE (PAGE_SIZE / sizeof(struct hlist_head))
+#define DLM_HASH_BUCKETS (DLM_HASH_PAGES * DLM_BUCKETS_PER_PAGE)
enum dlm_ast_type {
DLM_AST = 0,
@@ -86,7 +89,7 @@ enum dlm_ctxt_state {
struct dlm_ctxt
{
struct list_head list;
- struct hlist_head *lockres_hash;
+ struct hlist_head **lockres_hash;
struct list_head dirty_list;
struct list_head purge_list;
struct list_head pending_asts;
@@ -136,6 +139,13 @@ struct dlm_ctxt
struct list_head dlm_eviction_callbacks;
};
+static inline
+struct hlist_head *dlm_lockres_hash(struct dlm_ctxt *dlm, unsigned i)
+{
+ return dlm->lockres_hash[(i / DLM_BUCKETS_PER_PAGE) % DLM_HASH_PAGES]
+ + (i % DLM_BUCKETS_PER_PAGE);
+}
+
/* these keventd work queue items are for less-frequently
* called functions that cannot be directly called from the
* net message handlers for some reason, usually because
Index: ocfs2-1.2/fs/ocfs2/dlm/dlmrecovery.c
==================================================================---
ocfs2-1.2/fs/ocfs2/dlm/dlmrecovery.c (revision 1)
+++ ocfs2-1.2/fs/ocfs2/dlm/dlmrecovery.c (revision 2)
@@ -2064,7 +2064,7 @@ static void dlm_finish_local_lockres_rec
* the RECOVERING state and set the owner
* if necessary */
for (i = 0; i < DLM_HASH_BUCKETS; i++) {
- bucket = &(dlm->lockres_hash[i]);
+ bucket = dlm_lockres_hash(dlm, i);
hlist_for_each_entry(res, hash_iter, bucket, hash_node) {
if (res->state & DLM_LOCK_RES_RECOVERING) {
if (res->owner == dead_node) {
@@ -2259,7 +2259,7 @@ static void dlm_do_local_recovery_cleanu
* need to be fired as a result.
*/
for (i = 0; i < DLM_HASH_BUCKETS; i++) {
- bucket = &(dlm->lockres_hash[i]);
+ bucket = dlm_lockres_hash(dlm, i);
hlist_for_each_entry(res, iter, bucket, hash_node) {
/* always prune any $RECOVERY entries for dead nodes,
* otherwise hangs can occur during later recovery */
Apparently Analagous Threads
Wisdom of the Ancients
