[Ocfs2-devel] [PATCH 1/1] OCFS2: fasten dlm_lock_resource hash_table lookups

[Ocfs2-devel] [PATCH 1/1] OCFS2: fasten dlm_lock_resource hash_table lookups

[wengang wang]

#backporting the 3 patches at http://kernel.us.oracle.com/~smushran/srini/ to 1.2.

enlarge hash_table capacity to fasten hash_table lookups.

Signed-off-by: Wengang wang <wen.gang.wang@oracle.com>
--
diff -up ./svnocfs2-1.2/fs/ocfs2/dlm/dlmdebug.c.orig ./svnocfs2-1.2/fs/ocfs2/dlm/dlmdebug.c
--- ./svnocfs2-1.2/fs/ocfs2/dlm/dlmdebug.c.orig	2009-04-22 11:00:37.000000000 +0800
+++ ./svnocfs2-1.2/fs/ocfs2/dlm/dlmdebug.c	2009-04-22 11:08:27.000000000 +0800
@@ -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);
 	}
diff -up ./svnocfs2-1.2/fs/ocfs2/dlm/dlmrecovery.c.orig ./svnocfs2-1.2/fs/ocfs2/dlm/dlmrecovery.c
--- ./svnocfs2-1.2/fs/ocfs2/dlm/dlmrecovery.c.orig	2009-04-22 11:01:18.000000000 +0800
+++ ./svnocfs2-1.2/fs/ocfs2/dlm/dlmrecovery.c	2009-04-22 11:21:12.000000000 +0800
@@ -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 */
diff -up ./svnocfs2-1.2/fs/ocfs2/dlm/dlmdomain.c.orig ./svnocfs2-1.2/fs/ocfs2/dlm/dlmdomain.c
--- ./svnocfs2-1.2/fs/ocfs2/dlm/dlmdomain.c.orig	2009-04-22 11:00:47.000000000 +0800
+++ ./svnocfs2-1.2/fs/ocfs2/dlm/dlmdomain.c	2009-04-22 11:33:26.000000000 +0800
@@ -49,6 +49,34 @@
 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_DOMAIN)
 #include "cluster/masklog.h"
 
+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;
+	}
+	mlog(0, "Allocated DLM hash pagevec; %d pages (%lu expected), %Zd "
+	     "buckets per page\n", pages, DLM_HASH_PAGES, DLM_BUCKETS_PER_PAGE);
+	return vec;
+out_free:
+	dlm_free_pagevec(vec, i);
+	return NULL;
+}
+
 /*
  * ocfs2 node maps are array of long int, which limits to send them freely
  * across the wire due to endianness issues. To workaround this, we convert
@@ -127,7 +155,7 @@ 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 +179,7 @@ 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) {
@@ -394,7 +422,7 @@ 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);
 		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);
@@ -1366,7 +1395,7 @@ static struct dlm_ctxt *dlm_alloc_ctxt(c
 	}
 
 	for (i=0; i<DLM_HASH_BUCKETS; i++)
-		INIT_HLIST_HEAD(&dlm->lockres_hash[i]);
+		INIT_HLIST_HEAD(dlm_lockres_hash(dlm, i));
 
 	strcpy(dlm->name, domain);
 	dlm->key = key;
diff -up ./svnocfs2-1.2/fs/ocfs2/dlm/dlmcommon.h.orig ./svnocfs2-1.2/fs/ocfs2/dlm/dlmcommon.h
--- ./svnocfs2-1.2/fs/ocfs2/dlm/dlmcommon.h.orig	2009-04-22 10:59:51.000000000 +0800
+++ ./svnocfs2-1.2/fs/ocfs2/dlm/dlmcommon.h	2009-04-22 11:26:06.000000000 +0800
@@ -37,7 +37,14 @@
 #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_DEFAULT	(1 << 17)
+#if DLM_HASH_SIZE_DEFAULT < PAGE_SIZE
+# define DLM_HASH_PAGES		1
+#else
+# define DLM_HASH_PAGES		(DLM_HASH_SIZE_DEFAULT / PAGE_SIZE)
+#endif
+#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 +93,7 @@ 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 +143,13 @@ 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

[Ocfs2-devel] [PATCH 1/1] OCFS2: fasten dlm_lock_resource hash_table lookups

[Sunil Mushran]

Wengang,

When we backport, we like to add bread crumbs so that we can track back
to the original patch. This helps us in cases where we, say, forget to
include a patch. Also, we want to retain original authorship.

For example:
http://oss.oracle.com/git/?p=ocfs2-1.4.git;a=commitdiff;h=174f54db9af7c67a1afaa947aed934b9c80f588c

So don't merge the patches. Add the mainline commit at the top of the
patch. As 1.2 is svn, ensure the authors name is included in an Authored-by
tag.

Has it been tested in 1.2?

Thanks
Sunil

wengang wang wrote:
> #backporting the 3 patches at http://kernel.us.oracle.com/~smushran/srini/ to 1.2.
>
> enlarge hash_table capacity to fasten hash_table lookups.
>
> Signed-off-by: Wengang wang <wen.gang.wang@oracle.com>
> --
> diff -up ./svnocfs2-1.2/fs/ocfs2/dlm/dlmdebug.c.orig ./svnocfs2-1.2/fs/ocfs2/dlm/dlmdebug.c
> --- ./svnocfs2-1.2/fs/ocfs2/dlm/dlmdebug.c.orig	2009-04-22 11:00:37.000000000 +0800
> +++ ./svnocfs2-1.2/fs/ocfs2/dlm/dlmdebug.c	2009-04-22 11:08:27.000000000 +0800
> @@ -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);
>  	}
> diff -up ./svnocfs2-1.2/fs/ocfs2/dlm/dlmrecovery.c.orig ./svnocfs2-1.2/fs/ocfs2/dlm/dlmrecovery.c
> --- ./svnocfs2-1.2/fs/ocfs2/dlm/dlmrecovery.c.orig	2009-04-22 11:01:18.000000000 +0800
> +++ ./svnocfs2-1.2/fs/ocfs2/dlm/dlmrecovery.c	2009-04-22 11:21:12.000000000 +0800
> @@ -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 */
> diff -up ./svnocfs2-1.2/fs/ocfs2/dlm/dlmdomain.c.orig ./svnocfs2-1.2/fs/ocfs2/dlm/dlmdomain.c
> --- ./svnocfs2-1.2/fs/ocfs2/dlm/dlmdomain.c.orig	2009-04-22 11:00:47.000000000 +0800
> +++ ./svnocfs2-1.2/fs/ocfs2/dlm/dlmdomain.c	2009-04-22 11:33:26.000000000 +0800
> @@ -49,6 +49,34 @@
>  #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_DOMAIN)
>  #include "cluster/masklog.h"
>  
> +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;
> +	}
> +	mlog(0, "Allocated DLM hash pagevec; %d pages (%lu expected), %Zd "
> +	     "buckets per page\n", pages, DLM_HASH_PAGES, DLM_BUCKETS_PER_PAGE);
> +	return vec;
> +out_free:
> +	dlm_free_pagevec(vec, i);
> +	return NULL;
> +}
> +
>  /*
>   * ocfs2 node maps are array of long int, which limits to send them freely
>   * across the wire due to endianness issues. To workaround this, we convert
> @@ -127,7 +155,7 @@ 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 +179,7 @@ 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) {
> @@ -394,7 +422,7 @@ 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);
>  		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);
> @@ -1366,7 +1395,7 @@ static struct dlm_ctxt *dlm_alloc_ctxt(c
>  	}
>  
>  	for (i=0; i<DLM_HASH_BUCKETS; i++)
> -		INIT_HLIST_HEAD(&dlm->lockres_hash[i]);
> +		INIT_HLIST_HEAD(dlm_lockres_hash(dlm, i));
>  
>  	strcpy(dlm->name, domain);
>  	dlm->key = key;
> diff -up ./svnocfs2-1.2/fs/ocfs2/dlm/dlmcommon.h.orig ./svnocfs2-1.2/fs/ocfs2/dlm/dlmcommon.h
> --- ./svnocfs2-1.2/fs/ocfs2/dlm/dlmcommon.h.orig	2009-04-22 10:59:51.000000000 +0800
> +++ ./svnocfs2-1.2/fs/ocfs2/dlm/dlmcommon.h	2009-04-22 11:26:06.000000000 +0800
> @@ -37,7 +37,14 @@
>  #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_DEFAULT	(1 << 17)
> +#if DLM_HASH_SIZE_DEFAULT < PAGE_SIZE
> +# define DLM_HASH_PAGES		1
> +#else
> +# define DLM_HASH_PAGES		(DLM_HASH_SIZE_DEFAULT / PAGE_SIZE)
> +#endif
> +#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 +93,7 @@ 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 +143,13 @@ 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
>
> _______________________________________________
> Ocfs2-devel mailing list
> Ocfs2-devel at oss.oracle.com
> http://oss.oracle.com/mailman/listinfo/ocfs2-devel
>   

[Ocfs2-devel] [PATCH 1/1] OCFS2: fasten dlm_lock_resource hash_table lookups

[Wengang Wang]

Hi Sunil,

Sunil Mushran wrote:
> Wengang,
>
> When we backport, we like to add bread crumbs so that we can track back
> to the original patch. This helps us in cases where we, say, forget to
> include a patch. Also, we want to retain original authorship.
>
> For example:
> http://oss.oracle.com/git/?p=ocfs2-1.4.git;a=commitdiff;h=174f54db9af7c67a1afaa947aed934b9c80f588c 
>
>
> So don't merge the patches. Add the mainline commit at the top of the
> patch. As 1.2 is svn, ensure the authors name is included in an 
> Authored-by
> tag.
>
Ok, I will follow that for later backports.
> Has it been tested in 1.2?
>
No. Since I think this part(hash_table 
construction/insertions/lookups/destruction) is independent and no 
changes from the original patch,
I didn't test it --only made sure it pass compilation.
well,  you want to test the functionility(correctness) or the performance?

thanks,
wengang.
> Thanks
> Sunil
>
> wengang wang wrote:
>> #backporting the 3 patches at 
>> http://kernel.us.oracle.com/~smushran/srini/ to 1.2.
>>
>> enlarge hash_table capacity to fasten hash_table lookups.
>>
>