mm/zsmalloc: disclose statistics to debugfs

As we now talk more and more about the fragmentation of zsmalloc. But we still need to manually add some debug code to see the fragmentation. So, I think we may add the statistics of memory fragmention in zsmalloc and disclose them to debugfs. Then we can easily get and analysis them when adding or developing new feature for zsmalloc.

Below entries will be created when a zsmalloc pool is created:
/sys/kernel/debug/zsmalloc/pool-n/obj_allocated /sys/kernel/debug/zsmalloc/pool-n/obj_used

Then the status of objects usage will be:
objects_usage = obj_used / obj_allocated

Also we can collect other information and add corresponding entries in debugfs when needed.

Signed-off-by Ganesh Mahendran <[email protected]>

---
 mm/zsmalloc.c |  108 ++++++++++++++++++++++++++++++++++++++++++++++++++++++---
 1 file changed, 104 insertions(+), 4 deletions(-)

mm/zsmalloc.c +104 -4

index 4d0a063..f682ef9 100644
@@ -168,6 +168,8 @@ enum fullness_group {
 	ZS_FULL
 };
 
+static int zs_pool_num;
+
 /*
  * number of size_classes
  */

@@ -216,11 +218,19 @@ struct link_free {
 	void *next;
 };
 
+struct zs_stats {
+	atomic_long_t pages_allocated;
+	u64 obj_allocated;
+	u64 obj_used;
+};
+
 struct zs_pool {
 	struct size_class **size_class;
 
 	gfp_t flags;	/* allocation flags used when growing pool */
-	atomic_long_t pages_allocated;
+
+	struct zs_stats stats;
+	struct dentry *debugfs_dentry;
 };
 
 /*

@@ -925,12 +935,84 @@ static void init_zs_size_classes(void)
 	zs_size_classes = nr;
 }
 
+
+#ifdef CONFIG_DEBUG_FS
+#include <linux/debugfs.h>
+
+static struct dentry *zs_debugfs_root;
+
+static int __init zs_debugfs_init(void)
+{
+	if (!debugfs_initialized())
+		return -ENODEV;
+
+	zs_debugfs_root = debugfs_create_dir("zsmalloc", NULL);
+	if (!zs_debugfs_root)
+		return -ENOMEM;
+
+	return 0;
+}
+
+static void __exit zs_debugfs_exit(void)
+{
+	debugfs_remove_recursive(zs_debugfs_root);
+}
+
+static int zs_pool_debugfs_create(struct zs_pool *pool, int index)
+{
+	char name[10];
+	int ret = 0;
+
+	if (!zs_debugfs_root) {
+		ret = -ENODEV;
+		goto out;
+	}
+
+	snprintf(name, sizeof(name), "pool-%d", index);
+	pool->debugfs_dentry = debugfs_create_dir(name, zs_debugfs_root);
+	if (!pool->debugfs_dentry) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	debugfs_create_u64("obj_allocated", S_IRUGO, pool->debugfs_dentry,
+			&pool->stats.obj_allocated);
+	debugfs_create_u64("obj_used", S_IRUGO, pool->debugfs_dentry,
+			&pool->stats.obj_used);
+
+out:
+	return ret;
+}
+
+static void zs_pool_debugfs_destroy(struct zs_pool *pool)
+{
+	debugfs_remove_recursive(pool->debugfs_dentry);
+}
+
+#else
+static int __init zs_debugfs_init(void)
+{
+	return 0;
+}
+
+static void __exit zs_debugfs_exit(void) { }
+
+static int zs_pool_debugfs_create(struct zs_pool *pool, int index)
+{
+	return 0;
+}
+
+static void zs_pool_debugfs_destroy(struct zs_pool *pool) {}
+#endif
+
 static void __exit zs_exit(void)
 {
 #ifdef CONFIG_ZPOOL
 	zpool_unregister_driver(&zs_zpool_driver);
 #endif
 	zs_unregister_cpu_notifier();
+
+	zs_debugfs_exit();
 }
 
 static int __init zs_init(void)

@@ -947,6 +1029,10 @@ static int __init zs_init(void)
 #ifdef CONFIG_ZPOOL
 	zpool_register_driver(&zs_zpool_driver);
 #endif
+
+	if (zs_debugfs_init())
+		pr_warn("debugfs initialization failed\n");
+
 	return 0;
 }
 

@@ -1039,6 +1125,11 @@ struct zs_pool *zs_create_pool(gfp_t flags)
 
 	pool->flags = flags;
 
+	zs_pool_num++;
+
+	if (zs_pool_debugfs_create(pool, zs_pool_num))
+		pr_warn("zs pool debugfs initialization failed\n");
+
 	return pool;
 
 err:

@@ -1071,6 +1162,9 @@ void zs_destroy_pool(struct zs_pool *pool)
 	}
 
 	kfree(pool->size_class);
+	zs_pool_debugfs_destroy(pool);
+	zs_pool_num--;
+
 	kfree(pool);
 }
 EXPORT_SYMBOL_GPL(zs_destroy_pool);

@@ -1110,7 +1204,9 @@ unsigned long zs_malloc(struct zs_pool *pool, size_t size)
 
 		set_zspage_mapping(first_page, class->index, ZS_EMPTY);
 		atomic_long_add(class->pages_per_zspage,
-					&pool->pages_allocated);
+					&pool->stats.pages_allocated);
+		pool->stats.obj_allocated += get_maxobj_per_zspage(class->size,
+				class->pages_per_zspage);
 		spin_lock(&class->lock);
 	}
 

@@ -1125,6 +1221,7 @@ unsigned long zs_malloc(struct zs_pool *pool, size_t size)
 	kunmap_atomic(vaddr);
 
 	first_page->inuse++;
+	pool->stats.obj_used++;
 	/* Now move the zspage to another fullness group, if required */
 	fix_fullness_group(pool, first_page);
 	spin_unlock(&class->lock);

@@ -1164,12 +1261,15 @@ void zs_free(struct zs_pool *pool, unsigned long obj)
 	first_page->freelist = (void *)obj;
 
 	first_page->inuse--;
+	pool->stats.obj_used--;
 	fullness = fix_fullness_group(pool, first_page);
 	spin_unlock(&class->lock);
 
 	if (fullness == ZS_EMPTY) {
 		atomic_long_sub(class->pages_per_zspage,
-				&pool->pages_allocated);
+				&pool->stats.pages_allocated);
+		pool->stats.obj_allocated -= get_maxobj_per_zspage(class->size,
+				class->pages_per_zspage);
 		free_zspage(first_page);
 	}
 }

@@ -1267,7 +1367,7 @@ EXPORT_SYMBOL_GPL(zs_unmap_object);
 
 unsigned long zs_get_total_pages(struct zs_pool *pool)
 {
-	return atomic_long_read(&pool->pages_allocated);
+	return atomic_long_read(&pool->stats.pages_allocated);
 }
 EXPORT_SYMBOL_GPL(zs_get_total_pages);
 

Signature · 7 lines

1.7.9.5

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Hello Ganesh,

Quoted 10 lines · Ganesh Mahendran

As we now talk more and more about the fragmentation of zsmalloc. But we still need to manually add some debug code to see the fragmentation. So, I think we may add the statistics of memory fragmention in zsmalloc and disclose them to debugfs. Then we can easily get and analysis them when adding or developing new feature for zsmalloc.

Below entries will be created when a zsmalloc pool is created:
/sys/kernel/debug/zsmalloc/pool-n/obj_allocated /sys/kernel/debug/zsmalloc/pool-n/obj_used

Quoted 3 lines

Then the status of objects usage will be:
objects_usage = obj_used / obj_allocated

I didn't look at the code in detail but It would be handy for developer but not sure we should deliver it to admin so need configurable?

How about making it per-sizeclass information, not per-pool?

So we can rely on the class->lock for the locking rule.

Quoted 208 lines

Also we can collect other information and add corresponding entries in debugfs when needed.

Signed-off-by Ganesh Mahendran <[email protected]>

---
 mm/zsmalloc.c |  108 ++++++++++++++++++++++++++++++++++++++++++++++++++++++---
 1 file changed, 104 insertions(+), 4 deletions(-)

mm/zsmalloc.c +104 -5

index 4d0a063..f682ef9 100644
@@ -168,6 +168,8 @@ enum fullness_group {
      ZS_FULL
 };
 
+static int zs_pool_num;
+
 /*
  * number of size_classes
  */

@@ -216,11 +218,19 @@ struct link_free {
      void *next;
 };
 
+struct zs_stats {
+     atomic_long_t pages_allocated;
+     u64 obj_allocated;
+     u64 obj_used;
+};
+
 struct zs_pool {
      struct size_class **size_class;
 
      gfp_t flags;    /* allocation flags used when growing pool */
-     atomic_long_t pages_allocated;
+
+     struct zs_stats stats;
+     struct dentry *debugfs_dentry;
 };
 
 /*

@@ -925,12 +935,84 @@ static void init_zs_size_classes(void)
      zs_size_classes = nr;
 }
 
+
+#ifdef CONFIG_DEBUG_FS
+#include <linux/debugfs.h>
+
+static struct dentry *zs_debugfs_root;
+
+static int __init zs_debugfs_init(void)
+{
+     if (!debugfs_initialized())
+             return -ENODEV;
+
+     zs_debugfs_root = debugfs_create_dir("zsmalloc", NULL);
+     if (!zs_debugfs_root)
+             return -ENOMEM;
+
+     return 0;
+}
+
+static void __exit zs_debugfs_exit(void)
+{
+     debugfs_remove_recursive(zs_debugfs_root);
+}
+
+static int zs_pool_debugfs_create(struct zs_pool *pool, int index)
+{
+     char name[10];
+     int ret = 0;
+
+     if (!zs_debugfs_root) {
+             ret = -ENODEV;
+             goto out;
+     }
+
+     snprintf(name, sizeof(name), "pool-%d", index);
+     pool->debugfs_dentry = debugfs_create_dir(name, zs_debugfs_root);
+     if (!pool->debugfs_dentry) {
+             ret = -ENOMEM;
+             goto out;
+     }
+
+     debugfs_create_u64("obj_allocated", S_IRUGO, pool->debugfs_dentry,
+                     &pool->stats.obj_allocated);
+     debugfs_create_u64("obj_used", S_IRUGO, pool->debugfs_dentry,
+                     &pool->stats.obj_used);
+
+out:
+     return ret;
+}
+
+static void zs_pool_debugfs_destroy(struct zs_pool *pool)
+{
+     debugfs_remove_recursive(pool->debugfs_dentry);
+}
+
+#else
+static int __init zs_debugfs_init(void)
+{
+     return 0;
+}
+
+static void __exit zs_debugfs_exit(void) { }
+
+static int zs_pool_debugfs_create(struct zs_pool *pool, int index)
+{
+     return 0;
+}
+
+static void zs_pool_debugfs_destroy(struct zs_pool *pool) {}
+#endif
+
 static void __exit zs_exit(void)
 {
 #ifdef CONFIG_ZPOOL
      zpool_unregister_driver(&zs_zpool_driver);
 #endif
      zs_unregister_cpu_notifier();
+
+     zs_debugfs_exit();
 }
 
 static int __init zs_init(void)

@@ -947,6 +1029,10 @@ static int __init zs_init(void)
 #ifdef CONFIG_ZPOOL
      zpool_register_driver(&zs_zpool_driver);
 #endif
+
+     if (zs_debugfs_init())
+             pr_warn("debugfs initialization failed\n");
+
      return 0;
 }
 

@@ -1039,6 +1125,11 @@ struct zs_pool *zs_create_pool(gfp_t flags)
 
      pool->flags = flags;
 
+     zs_pool_num++;
+
+     if (zs_pool_debugfs_create(pool, zs_pool_num))
+             pr_warn("zs pool debugfs initialization failed\n");
+
      return pool;
 
 err:

@@ -1071,6 +1162,9 @@ void zs_destroy_pool(struct zs_pool *pool)
      }
 
      kfree(pool->size_class);
+     zs_pool_debugfs_destroy(pool);
+     zs_pool_num--;
+
      kfree(pool);
 }
 EXPORT_SYMBOL_GPL(zs_destroy_pool);

@@ -1110,7 +1204,9 @@ unsigned long zs_malloc(struct zs_pool *pool, size_t size)
 
              set_zspage_mapping(first_page, class->index, ZS_EMPTY);
              atomic_long_add(class->pages_per_zspage,
-                                     &pool->pages_allocated);
+                                     &pool->stats.pages_allocated);
+             pool->stats.obj_allocated += get_maxobj_per_zspage(class->size,
+                             class->pages_per_zspage);
              spin_lock(&class->lock);
      }
 

@@ -1125,6 +1221,7 @@ unsigned long zs_malloc(struct zs_pool *pool, size_t size)
      kunmap_atomic(vaddr);
 
      first_page->inuse++;
+     pool->stats.obj_used++;
      /* Now move the zspage to another fullness group, if required */
      fix_fullness_group(pool, first_page);
      spin_unlock(&class->lock);

@@ -1164,12 +1261,15 @@ void zs_free(struct zs_pool *pool, unsigned long obj)
      first_page->freelist = (void *)obj;
 
      first_page->inuse--;
+     pool->stats.obj_used--;
      fullness = fix_fullness_group(pool, first_page);
      spin_unlock(&class->lock);
 
      if (fullness == ZS_EMPTY) {
              atomic_long_sub(class->pages_per_zspage,
-                             &pool->pages_allocated);
+                             &pool->stats.pages_allocated);
+             pool->stats.obj_allocated -= get_maxobj_per_zspage(class->size,
+                             class->pages_per_zspage);
              free_zspage(first_page);
      }
 }

@@ -1267,7 +1367,7 @@ EXPORT_SYMBOL_GPL(zs_unmap_object);
 
 unsigned long zs_get_total_pages(struct zs_pool *pool)
 {
-     return atomic_long_read(&pool->pages_allocated);
+     return atomic_long_read(&pool->stats.pages_allocated);
 }
 EXPORT_SYMBOL_GPL(zs_get_total_pages);
 
--

1.7.9.5

Quoted 5 lines

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

--
Kind regards,
Minchan Kim

Hello Minchan

2014-12-12 7:40 GMT+08:00 Minchan Kim <[email protected]>:

Quoted 17 lines

Hello Ganesh,

Quoted 13 lines · Ganesh Mahendran

As we now talk more and more about the fragmentation of zsmalloc. But we still need to manually add some debug code to see the fragmentation. So, I think we may add the statistics of memory fragmention in zsmalloc and disclose them to debugfs. Then we can easily get and analysis them when adding or developing new feature for zsmalloc.

Below entries will be created when a zsmalloc pool is created:
/sys/kernel/debug/zsmalloc/pool-n/obj_allocated /sys/kernel/debug/zsmalloc/pool-n/obj_used

Then the status of objects usage will be:
objects_usage = obj_used / obj_allocated

Quoted 2 lines

I didn't look at the code in detail but It would be handy for developer but not sure we should deliver it to admin so need configurable?

What kind of configuration do you want?
I think it is reasonable to expose such information to admin like */sys/kernel/debug/usb/device*

Or maybe we can enclose these code by DEBUG macro which will be defined when CONFIG_ZSMALLOC_DEBUG is selected.

Quoted 3 lines

How about making it per-sizeclass information, not per-pool?

Yes, you are right. Per sizeclass information will be better for developers than per pool.

Is it acceptable to show 256 lines like:
#cat /sys/kernel/debug/zsmalloc/pool-1/obj_in_classes
class obj_allocated obj_used
1 ...
2 ...
....
....
255

Anyway for developers, these information is more usefull.

Quoted 26 lines

Hello Minchan

2014-12-12 7:40 GMT+08:00 Minchan Kim <[email protected]>:

Quoted 13 lines

Hello Ganesh,

Quoted 10 lines · Ganesh Mahendran

As we now talk more and more about the fragmentation of zsmalloc. But we still need to manually add some debug code to see the fragmentation. So, I think we may add the statistics of memory fragmention in zsmalloc and disclose them to debugfs. Then we can easily get and analysis them when adding or developing new feature for zsmalloc.

Below entries will be created when a zsmalloc pool is created:
/sys/kernel/debug/zsmalloc/pool-n/obj_allocated /sys/kernel/debug/zsmalloc/pool-n/obj_used

Quoted 6 lines

Quoted 3 lines

Then the status of objects usage will be:
objects_usage = obj_used / obj_allocated

I didn't look at the code in detail but It would be handy for developer but not sure we should deliver it to admin so need configurable?

What kind of configuration do you want?
I think it is reasonable to expose such information to admin like */sys/kernel/debug/usb/device*

Quoted 2 lines

Or maybe we can enclose these code by DEBUG macro which will be defined when CONFIG_ZSMALLOC_DEBUG is selected.

Hmm, I'd like to separte DEBUG and STAT because we can add some sanity checking(ex, poisoning for invalid overwriting or handle<->obj mapping verification) with DEBUG while we could count obj stat with STAT.

So, now it seems you want CONFIG_ZSMALLOC_STAT?

Quoted 16 lines

Quoted 4 lines

How about making it per-sizeclass information, not per-pool?

Yes, you are right. Per sizeclass information will be better for developers than per pool.

Is it acceptable to show 256 lines like:
#cat /sys/kernel/debug/zsmalloc/pool-1/obj_in_classes
class obj_allocated obj_used
1 ...
2 ...
....
....
255

Quoted 1 line

Anyway for developers, these information is more usefull.

It would be better to show the number of pages so we can know how many of fragment space in last subpage of zspage is wasted. But I don't want to keep pages_used in memory but you could calcurate it dynamically with obj_allocated when user access debugfs.

#cat /sys/kernel/debug/zsmalloc/pool-1/obj_in_classes
class-size obj_allocated obj_used pages_used
32
48
.
.
.

Thanks!

--
Kind regards,
Minchan Kim