Slab Movable Objects (SMO)

Hi,

TL;DR - Add object migration (SMO) to the SLUB allocator and implement object migration for the XArray and the dcache.

Thanks for you patience with all the RFC's of this patch set. Here it is, ready for prime time.

Internal fragmentation can occur within pages used by the slub allocator. Under some workloads large numbers of pages can be used by partial slab pages. This under-utilisation is bad simply because it wastes memory but also because if the system is under memory pressure higher order allocations may become difficult to satisfy. If we can defrag slab caches we can alleviate these problems.

In order to be able to defrag slab chaches we need to be able to migrate objects to a new slab. Slab object migration is the core functionality added by this patch series.

Internal slab fragmentation is a long known problem. This series does not claim to completely _fix_ the issue. Instead we are adding core code to the SLUB allocator to enable users of the allocator to help mitigate internal fragmentation. Object migration is on a per cache basis, with each cache being able to take advantage of object migration to varying degrees depending on the nature of the objects stored in the cache.

Series includes test modules and test code that can be used to verify the claimed behaviour.

Patch #1 - Adds the callbacks used to enable SMO for a particular cache.

Patch #2 - Updates the slabinfo tool to show operations related to SMO.

Patch #3 - Sorts the cache list putting migratable slabs at front.

Patch #4 - Adds the SMO infrastructure. This is the core patch of the
series.

Patch #5, #6 - Further update slabinfo tool for information just added.

Patch #7 - Add a module for testing SMO.

Patch #8 - Add unit test suite in Python utilising test module from #7.

Patch #9 - Add a new slab cache for the XArray (separate from radix tree).

Patch #10 - Implement SMO for the XArray.

Patch #11 - Add module for testing XArray SMO implementation.

Patch #12 - Add a dentry constructor.

Patch #13 - Use SMO to attempt to reduce fragmentation of the dcache by
selectively freeing dentry objects.

Patch #14 - Add functionality to move slab objects to a specific NUMA node.

Patch #15 - Add functionality to balance slab objects across all NUMA nodes.

The last RFC (RFCv5 and discussion on it) included code to conditionally exclude SMO for the dcache. This has been removed. IMO it is now not needed. Al sufficiently bollock'ed me during development that I believe the dentry code is good and does not negatively effect the dcache. If someone would like to prove me wrong simply remove the call to

    kmem_cache_setup_mobility(dentry_cache, d_isolate, d_partial_shrink);

Testing:

The series has been tested to verify that objects are moved using bare metal (core i5) and also Qemu. This has not been tested on big metal or on NUMA hardware.

I have no measurements on performance gains achievable with this set, I have just verified that the migration works and does not appear to break anything.

Patch #14 and #15 depend on

CONFIG_SLBU_DEBUG_ON or boot with 'slub_debug'

Thanks for taking the time to look at this.

Tobin

Tobin C. Harding (15):
slub: Add isolate() and migrate() methods tools/vm/slabinfo: Add support for -C and -M options
slub: Sort slab cache list
slub: Slab defrag core
tools/vm/slabinfo: Add remote node defrag ratio output tools/vm/slabinfo: Add defrag_used_ratio output tools/testing/slab: Add object migration test module tools/testing/slab: Add object migration test suite
lib: Separate radix_tree_node and xa_node slab cache
xarray: Implement migration function for xa_node objects tools/testing/slab: Add XArray movable objects tests
dcache: Provide a dentry constructor
dcache: Implement partial shrink via Slab Movable Objects
slub: Enable moving objects to/from specific nodes
slub: Enable balancing slabs across nodes

 Documentation/ABI/testing/sysfs-kernel-slab |  14 +
 fs/dcache.c                                 | 105 ++-
 include/linux/slab.h                        |  71 ++
 include/linux/slub_def.h                    |  10 +
 include/linux/xarray.h                      |   3 +
 init/main.c                                 |   2 +
 lib/radix-tree.c                            |   2 +-
 lib/xarray.c                                | 109 ++-
 mm/Kconfig                                  |   7 +
 mm/slab_common.c                            |   2 +-
 mm/slub.c                                   | 827 ++++++++++++++++++--
 tools/testing/slab/Makefile                 |  10 +
 tools/testing/slab/slub_defrag.c            | 567 ++++++++++++++
 tools/testing/slab/slub_defrag.py           | 451 +++++++++++
 tools/testing/slab/slub_defrag_xarray.c     | 211 +++++
 tools/vm/slabinfo.c                         |  51 +-
 16 files changed, 2339 insertions(+), 103 deletions(-)
 create mode 100644 tools/testing/slab/Makefile
 create mode 100644 tools/testing/slab/slub_defrag.c
 create mode 100755 tools/testing/slab/slub_defrag.py
 create mode 100644 tools/testing/slab/slub_defrag_xarray.c

2.21.0

Add the two methods needed for moving objects and enable the display of the callbacks via the /sys/kernel/slab interface.

Add documentation explaining the use of these methods and the prototypes for slab.h. Add functions to setup the callbacks method for a slab cache.

Add empty functions for SLAB/SLOB. The API is generic so it could be theoretically implemented for these allocators as well.

Change sysfs 'ctor' field to be 'ops' to contain all the callback operations defined for a slab cache. Display the existing 'ctor' callback in the ops fields contents along with 'isolate' and 'migrate' callbacks.

---
 include/linux/slab.h     | 70 ++++++++++++++++++++++++++++++++++++++++
 include/linux/slub_def.h |  3 ++
 mm/slub.c                | 59 +++++++++++++++++++++++++++++----
 3 files changed, 126 insertions(+), 6 deletions(-)

index 9449b19c5f10..886fc130334d 100644
@@ -154,6 +154,76 @@ void memcg_create_kmem_cache(struct mem_cgroup *, struct kmem_cache *);
 void memcg_deactivate_kmem_caches(struct mem_cgroup *);
 void memcg_destroy_kmem_caches(struct mem_cgroup *);
 
+/*
+ * Function prototypes passed to kmem_cache_setup_mobility() to enable
+ * mobile objects and targeted reclaim in slab caches.
+ */
+
+/**
+ * typedef kmem_cache_isolate_func - Object migration callback function.
+ * @s: The cache we are working on.
+ * @ptr: Pointer to an array of pointers to the objects to isolate.
+ * @nr: Number of objects in @ptr array.
+ *
+ * The purpose of kmem_cache_isolate_func() is to pin each object so that
+ * they cannot be freed until kmem_cache_migrate_func() has processed
+ * them. This may be accomplished by increasing the refcount or setting
+ * a flag.
+ *
+ * The object pointer array passed is also passed to
+ * kmem_cache_migrate_func().  The function may remove objects from the
+ * array by setting pointers to %NULL. This is useful if we can
+ * determine that an object is being freed because
+ * kmem_cache_isolate_func() was called when the subsystem was calling
+ * kmem_cache_free().  In that case it is not necessary to increase the
+ * refcount or specially mark the object because the release of the slab
+ * lock will lead to the immediate freeing of the object.
+ *
+ * Context: Called with locks held so that the slab objects cannot be
+ *          freed.  We are in an atomic context and no slab operations
+ *          may be performed.
+ * Return: A pointer that is passed to the migrate function. If any
+ *         objects cannot be touched at this point then the pointer may
+ *         indicate a failure and then the migration function can simply
+ *         remove the references that were already obtained. The private
+ *         data could be used to track the objects that were already pinned.
+ */
+typedef void *kmem_cache_isolate_func(struct kmem_cache *s, void **ptr, int nr);
+
+/**
+ * typedef kmem_cache_migrate_func - Object migration callback function.
+ * @s: The cache we are working on.
+ * @ptr: Pointer to an array of pointers to the objects to migrate.
+ * @nr: Number of objects in @ptr array.
+ * @node: The NUMA node where the object should be allocated.
+ * @private: The pointer returned by kmem_cache_isolate_func().
+ *
+ * This function is responsible for migrating objects.  Typically, for
+ * each object in the input array you will want to allocate an new
+ * object, copy the original object, update any pointers, and free the
+ * old object.
+ *
+ * After this function returns all pointers to the old object should now
+ * point to the new object.
+ *
+ * Context: Called with no locks held and interrupts enabled.  Sleeping
+ *          is possible.  Any operation may be performed.
+ */
+typedef void kmem_cache_migrate_func(struct kmem_cache *s, void **ptr,
+				     int nr, int node, void *private);
+
+/*
+ * kmem_cache_setup_mobility() is used to setup callbacks for a slab cache.
+ */
+#ifdef CONFIG_SLUB
+void kmem_cache_setup_mobility(struct kmem_cache *, kmem_cache_isolate_func,
+			       kmem_cache_migrate_func);
+#else
+static inline void
+kmem_cache_setup_mobility(struct kmem_cache *s, kmem_cache_isolate_func isolate,
+			  kmem_cache_migrate_func migrate) {}
+#endif
+
 /*
  * Please use this macro to create slab caches. Simply specify the
  * name of the structure and maybe some flags that are listed above.

index d2153789bd9f..2879a2f5f8eb 100644
@@ -99,6 +99,9 @@ struct kmem_cache {
 	gfp_t allocflags;	/* gfp flags to use on each alloc */
 	int refcount;		/* Refcount for slab cache destroy */
 	void (*ctor)(void *);
+	kmem_cache_isolate_func *isolate;
+	kmem_cache_migrate_func *migrate;
+
 	unsigned int inuse;		/* Offset to metadata */
 	unsigned int align;		/* Alignment */
 	unsigned int red_left_pad;	/* Left redzone padding size */

2.21.0

-C lists caches that use a ctor.

-M lists caches that support object migration.

Add command line options to show caches with a constructor and caches that are movable (i.e. have migrate function).

---
 tools/vm/slabinfo.c | 40 ++++++++++++++++++++++++++++++++++++----
 1 file changed, 36 insertions(+), 4 deletions(-)

index 73818f1b2ef8..cbfc56c44c2f 100644
@@ -33,6 +33,7 @@ struct slabinfo {
 	unsigned int hwcache_align, object_size, objs_per_slab;
 	unsigned int sanity_checks, slab_size, store_user, trace;
 	int order, poison, reclaim_account, red_zone;
+	int movable, ctor;
 	unsigned long partial, objects, slabs, objects_partial, objects_total;
 	unsigned long alloc_fastpath, alloc_slowpath;
 	unsigned long free_fastpath, free_slowpath;

@@ -67,6 +68,8 @@ int show_report;
 int show_alias;
 int show_slab;
 int skip_zero = 1;
+int show_movable;
+int show_ctor;
 int show_numa;
 int show_track;
 int show_first_alias;

@@ -109,11 +112,13 @@ static void fatal(const char *x, ...)
 
 static void usage(void)
 {
-	printf("slabinfo 4/15/2011. (c) 2007 sgi/(c) 2011 Linux Foundation.\n\n"
-		"slabinfo [-aADefhilnosrStTvz1LXBU] [N=K] [-dafzput] [slab-regexp]\n"
+	printf("slabinfo 4/15/2017. (c) 2007 sgi/(c) 2011 Linux Foundation/(c) 2017 Jump Trading LLC.\n\n"
+	       "slabinfo [-aACDefhilMnosrStTvz1LXBU] [N=K] [-dafzput] [slab-regexp]\n"
+
 		"-a|--aliases           Show aliases\n"
 		"-A|--activity          Most active slabs first\n"
 		"-B|--Bytes             Show size in bytes\n"
+		"-C|--ctor              Show slabs with ctors\n"
 		"-D|--display-active    Switch line format to activity\n"
 		"-e|--empty             Show empty slabs\n"
 		"-f|--first-alias       Show first alias\n"

@@ -121,6 +126,7 @@ static void usage(void)
 		"-i|--inverted          Inverted list\n"
 		"-l|--slabs             Show slabs\n"
 		"-L|--Loss              Sort by loss\n"
+		"-M|--movable           Show caches that support movable objects\n"
 		"-n|--numa              Show NUMA information\n"
 		"-N|--lines=K           Show the first K slabs\n"
 		"-o|--ops               Show kmem_cache_ops\n"

@@ -588,6 +594,12 @@ static void slabcache(struct slabinfo *s)
 	if (show_empty && s->slabs)
 		return;
 
+	if (show_ctor && !s->ctor)
+		return;
+
+	if (show_movable && !s->movable)
+		return;
+
 	if (sort_loss == 0)
 		store_size(size_str, slab_size(s));
 	else

@@ -602,6 +614,10 @@ static void slabcache(struct slabinfo *s)
 		*p++ = '*';
 	if (s->cache_dma)
 		*p++ = 'd';
+	if (s->ctor)
+		*p++ = 'C';
+	if (s->movable)
+		*p++ = 'M';
 	if (s->hwcache_align)
 		*p++ = 'A';
 	if (s->poison)

@@ -636,7 +652,8 @@ static void slabcache(struct slabinfo *s)
 		printf("%-21s %8ld %7d %15s %14s %4d %1d %3ld %3ld %s\n",
 			s->name, s->objects, s->object_size, size_str, dist_str,
 			s->objs_per_slab, s->order,
-			s->slabs ? (s->partial * 100) / s->slabs : 100,
+			s->slabs ? (s->partial * 100) /
+					(s->slabs * s->objs_per_slab) : 100,
 			s->slabs ? (s->objects * s->object_size * 100) /
 				(s->slabs * (page_size << s->order)) : 100,
 			flags);

@@ -1256,6 +1273,13 @@ static void read_slab_dir(void)
 			slab->alloc_node_mismatch = get_obj("alloc_node_mismatch");
 			slab->deactivate_bypass = get_obj("deactivate_bypass");
 			chdir("..");
+			if (read_slab_obj(slab, "ops")) {
+				if (strstr(buffer, "ctor :"))
+					slab->ctor = 1;
+				if (strstr(buffer, "migrate :"))
+					slab->movable = 1;
+			}
+
 			if (slab->name[0] == ':')
 				alias_targets++;
 			slab++;

@@ -1332,6 +1356,8 @@ static void xtotals(void)
 }
 
 struct option opts[] = {
+	{ "ctor", no_argument, NULL, 'C' },
+	{ "movable", no_argument, NULL, 'M' },
 	{ "aliases", no_argument, NULL, 'a' },
 	{ "activity", no_argument, NULL, 'A' },
 	{ "debug", optional_argument, NULL, 'd' },

@@ -1367,7 +1393,7 @@ int main(int argc, char *argv[])
 
 	page_size = getpagesize();
 
-	while ((c = getopt_long(argc, argv, "aAd::Defhil1noprstvzTSN:LXBU",
+	while ((c = getopt_long(argc, argv, "aACd::Defhil1MnoprstvzTSN:LXBU",
 						opts, NULL)) != -1)
 		switch (c) {
 		case '1':

@@ -1376,6 +1402,9 @@ int main(int argc, char *argv[])
 		case 'a':
 			show_alias = 1;
 			break;
+		case 'C':
+			show_ctor = 1;
+			break;
 		case 'A':
 			sort_active = 1;
 			break;

@@ -1399,6 +1428,9 @@ int main(int argc, char *argv[])
 		case 'i':
 			show_inverted = 1;
 			break;
+		case 'M':
+			show_movable = 1;
+			break;
 		case 'n':
 			show_numa = 1;
 			break;

2.21.0

It is advantageous to have all defragmentable slabs together at the beginning of the list of slabs so that there is no need to scan the complete list. Put defragmentable caches first when adding a slab cache and others last.

---
 mm/slab_common.c | 2 +-
 mm/slub.c        | 6 ++++++
 2 files changed, 7 insertions(+), 1 deletion(-)

index 58251ba63e4a..db5e9a0b1535 100644
@@ -393,7 +393,7 @@ static struct kmem_cache *create_cache(const char *name,
 		goto out_free_cache;
 
 	s->refcount = 1;
-	list_add(&s->list, &slab_caches);
+	list_add_tail(&s->list, &slab_caches);
 	memcg_link_cache(s);
 out:
 	if (err)

index 1c380a2bc78a..66d474397c0f 100644
@@ -4333,6 +4333,8 @@ void kmem_cache_setup_mobility(struct kmem_cache *s,
 		return;
 	}
 
+	mutex_lock(&slab_mutex);
+
 	s->isolate = isolate;
 	s->migrate = migrate;
 

@@ -4341,6 +4343,10 @@ void kmem_cache_setup_mobility(struct kmem_cache *s,
 	 * to disable fast cmpxchg based processing.
 	 */
 	s->flags &= ~__CMPXCHG_DOUBLE;
+
+	list_move(&s->list, &slab_caches);	/* Move to top */
+
+	mutex_unlock(&slab_mutex);
 }
 EXPORT_SYMBOL(kmem_cache_setup_mobility);
 

2.21.0

Internal fragmentation can occur within pages used by the slub allocator. Under some workloads large numbers of pages can be used by partial slab pages. This under-utilisation is bad simply because it wastes memory but also because if the system is under memory pressure higher order allocations may become difficult to satisfy. If we can defrag slab caches we can alleviate these problems.

Implement Slab Movable Objects in order to defragment slab caches.

Slab defragmentation may occur:

1. Unconditionally when __kmem_cache_shrink() is called on a slab cache by the kernel calling kmem_cache_shrink().

2. Unconditionally through the use of the slabinfo command.

slabinfo <cache> -s

3. Conditionally via the use of kmem_cache_defrag()

- Use Slab Movable Objects when shrinking cache.

Currently when the kernel calls kmem_cache_shrink() we curate the partial slabs list. If object migration is not enabled for the cache we still do this, if however, SMO is enabled we attempt to move objects in partially full slabs in order to defragment the cache. Shrink attempts to move all objects in order to reduce the cache to a single partial slab for each node.

- Add conditional per node defrag via new function:

kmem_defrag_slabs(int node).

kmem_defrag_slabs() attempts to defragment all slab caches for node. Defragmentation is done conditionally dependent on MAX_PARTIAL _and_ defrag_used_ratio.

Caches are only considered for defragmentation if the number of partial slabs exceeds MAX_PARTIAL (per node).

Also, defragmentation only occurs if the usage ratio of the slab is lower than the configured percentage (sysfs field added in this patch). Fragmentation ratios are measured by calculating the percentage of objects in use compared to the total number of objects that the slab page can accommodate.

The scanning of slab caches is optimized because the defragmentable slabs come first on the list. Thus we can terminate scans on the first slab encountered that does not support defragmentation.

kmem_defrag_slabs() takes a node parameter. This can either be -1 if defragmentation should be performed on all nodes, or a node number.

Defragmentation may be disabled by setting defrag ratio to 0

echo 0 > /sys/kernel/slab/<cache>/defrag_used_ratio

- Add a defrag ratio sysfs field and set it to 30% by default. A limit of 30% specifies that more than 3 out of 10 available slots for objects need to be in use otherwise slab defragmentation will be attempted on the remaining objects.

In order for a cache to be defragmentable the cache must support object migration (SMO). Enabling SMO for a cache is done via a call to the recently added function:

	void kmem_cache_setup_mobility(struct kmem_cache *,
				       kmem_cache_isolate_func,
			               kmem_cache_migrate_func);

---
 Documentation/ABI/testing/sysfs-kernel-slab |  14 +
 include/linux/slab.h                        |   1 +
 include/linux/slub_def.h                    |   7 +
 mm/slub.c                                   | 385 ++++++++++++++++----
 4 files changed, 334 insertions(+), 73 deletions(-)

index 29601d93a1c2..8bd893968e4f 100644
@@ -180,6 +180,20 @@ Description:
 		list.  It can be written to clear the current count.
 		Available when CONFIG_SLUB_STATS is enabled.
 
+What:		/sys/kernel/slab/cache/defrag_used_ratio
+Date:		June 2019
+KernelVersion:	5.2
+Contact:	Christoph Lameter <[email protected]>
+		Pekka Enberg <[email protected]>,
+Description:
+		The defrag_used_ratio file allows the control of how aggressive
+		slab fragmentation reduction works at reclaiming objects from
+		sparsely populated slabs. This is a percentage. If a slab has
+		less than this percentage of objects allocated then reclaim will
+		attempt to reclaim objects so that the whole slab page can be
+		freed. 0% specifies no reclaim attempt (defrag disabled), 100%
+		specifies attempt to reclaim all pages.  The default is 30%.
+
 What:		/sys/kernel/slab/cache/deactivate_to_tail
 Date:		February 2008
 KernelVersion:	2.6.25

index 886fc130334d..4bf381b34829 100644
@@ -149,6 +149,7 @@ struct kmem_cache *kmem_cache_create_usercopy(const char *name,
 			void (*ctor)(void *));
 void kmem_cache_destroy(struct kmem_cache *);
 int kmem_cache_shrink(struct kmem_cache *);
+unsigned long kmem_defrag_slabs(int node);
 
 void memcg_create_kmem_cache(struct mem_cgroup *, struct kmem_cache *);
 void memcg_deactivate_kmem_caches(struct mem_cgroup *);

index 2879a2f5f8eb..34c6f1250652 100644
@@ -107,6 +107,13 @@ struct kmem_cache {
 	unsigned int red_left_pad;	/* Left redzone padding size */
 	const char *name;	/* Name (only for display!) */
 	struct list_head list;	/* List of slab caches */
+	int defrag_used_ratio;	/*
+				 * Ratio used to check against the
+				 * percentage of objects allocated in a
+				 * slab page.  If less than this ratio
+				 * is allocated then reclaim attempts
+				 * are made.
+				 */
 #ifdef CONFIG_SYSFS
 	struct kobject kobj;	/* For sysfs */
 	struct work_struct kobj_remove_work;

2.21.0

Add output line for NUMA remote node defrag ratio.

---
 tools/vm/slabinfo.c | 7 +++++++
 1 file changed, 7 insertions(+)

index cbfc56c44c2f..d2c22f9ee2d8 100644
@@ -34,6 +34,7 @@ struct slabinfo {
 	unsigned int sanity_checks, slab_size, store_user, trace;
 	int order, poison, reclaim_account, red_zone;
 	int movable, ctor;
+	int remote_node_defrag_ratio;
 	unsigned long partial, objects, slabs, objects_partial, objects_total;
 	unsigned long alloc_fastpath, alloc_slowpath;
 	unsigned long free_fastpath, free_slowpath;

@@ -377,6 +378,10 @@ static void slab_numa(struct slabinfo *s, int mode)
 	if (skip_zero && !s->slabs)
 		return;
 
+	if (mode) {
+		printf("\nNUMA remote node defrag ratio: %3d\n",
+		       s->remote_node_defrag_ratio);
+	}
 	if (!line) {
 		printf("\n%-21s:", mode ? "NUMA nodes" : "Slab");
 		for(node = 0; node <= highest_node; node++)

@@ -1272,6 +1277,8 @@ static void read_slab_dir(void)
 			slab->cpu_partial_free = get_obj("cpu_partial_free");
 			slab->alloc_node_mismatch = get_obj("alloc_node_mismatch");
 			slab->deactivate_bypass = get_obj("deactivate_bypass");
+			slab->remote_node_defrag_ratio =
+					get_obj("remote_node_defrag_ratio");
 			chdir("..");
 			if (read_slab_obj(slab, "ops")) {
 				if (strstr(buffer, "ctor :"))

2.21.0

Add output for the newly added defrag_used_ratio sysfs knob.

---
 tools/vm/slabinfo.c | 4 ++++
 1 file changed, 4 insertions(+)

index d2c22f9ee2d8..ef4ff93df4cc 100644
@@ -34,6 +34,7 @@ struct slabinfo {
 	unsigned int sanity_checks, slab_size, store_user, trace;
 	int order, poison, reclaim_account, red_zone;
 	int movable, ctor;
+	int defrag_used_ratio;
 	int remote_node_defrag_ratio;
 	unsigned long partial, objects, slabs, objects_partial, objects_total;
 	unsigned long alloc_fastpath, alloc_slowpath;

@@ -549,6 +550,8 @@ static void report(struct slabinfo *s)
 		printf("** Slabs are destroyed via RCU\n");
 	if (s->reclaim_account)
 		printf("** Reclaim accounting active\n");
+	if (s->movable)
+		printf("** Defragmentation at %d%%\n", s->defrag_used_ratio);
 
 	printf("\nSizes (bytes)     Slabs              Debug                Memory\n");
 	printf("------------------------------------------------------------------------\n");

@@ -1279,6 +1282,7 @@ static void read_slab_dir(void)
 			slab->deactivate_bypass = get_obj("deactivate_bypass");
 			slab->remote_node_defrag_ratio =
 					get_obj("remote_node_defrag_ratio");
+			slab->defrag_used_ratio = get_obj("defrag_used_ratio");
 			chdir("..");
 			if (read_slab_obj(slab, "ops")) {
 				if (strstr(buffer, "ctor :"))

2.21.0

We just implemented slab movable objects for the SLUB allocator. We should test that code. In order to do so we need to be able to do a number of things

- Create a cache
- Enable Slab Movable Objects for the cache
- Allocate objects to the cache
- Free objects from within specific slabs of the cache

We can do all this via a loadable module.

Add a module that defines functions that can be triggered from userspace via a debugfs entry. From the source:

/*
* SLUB defragmentation a.k.a. Slab Movable Objects (SMO). *
* This module is used for testing the SLUB allocator. Enables
* userspace to run kernel functions via a debugfs file.
*
* debugfs: /sys/kernel/debugfs/smo/callfn (write only)
*
* String written to callfn is parsed by the module and associated
* function is called. See fn_tab for mapping of strings to functions. */

References to allocated objects are kept by the module in a linked list so that userspace can control which object to free.

We introduce the following four functions via the function table

"enable": Enables object migration for the test cache. "alloc X": Allocates X objects
"free X [Y]": Frees X objects starting at list position Y (default Y==0) "test": Runs [stress] tests from within the module (see below).

       {"enable", smo_enable_cache_mobility},
       {"alloc", smo_alloc_objects},
       {"free", smo_free_object},
       {"test", smo_run_module_tests},

Freeing from the start of the list creates a hole in the slab being freed from (i.e. creates a partial slab). The results of running these commands can be see using slabinfo (available in tools/vm/):

make -o slabinfo tools/vm/slabinfo.c

Stress tests can be run from within the module. These tests are internal to the module because we verify that object references are still good after object migration. These are called 'stress' tests because it is intended that they create/free a lot of objects. Userspace can control the number of objects to create, default is 1000.

Example test session
--------------------

Relevant /proc/slabinfo column headers:

name <active_objs> <num_objs> <objsize> <objperslab> <pagesperslab>

  # mount -t debugfs none /sys/kernel/debug/
  $ cd path/to/linux/tools/testing/slab; make
  ...

  # insmod slub_defrag.ko
  # cat /proc/slabinfo | grep smo_test | sed 's/:.*//'
  smo_test               0      0    392   20    2

From this we can see that the module created cache 'smo_test' with 20 objects per slab and 2 pages per slab (and cache is currently empty).

We can play with the slab allocator manually:

  # insmod slub_defrag.ko
  # echo 'alloc 21' > callfn
  # cat /proc/slabinfo | grep smo_test | sed 's/:.*//'
  smo_test              21     40    392   20    2

We see here that 21 active objects have been allocated creating 2 slabs (40 total objects).

  # slabinfo smo_test --report

Slabcache: smo_test Aliases: 0 Order : 1 Objects: 21

Sizes (bytes) Slabs Debug Memory
------------------------------------------------------------------------

  Object :      56  Total  :       2   Sanity Checks : On   Total:   16384
  SlabObj:     392  Full   :       1   Redzoning     : On   Used :    1176
  SlabSiz:    8192  Partial:       1   Poisoning     : On   Loss :   15208
  Loss   :     336  CpuSlab:       0   Tracking      : On   Lalig:    7056
  Align  :       8  Objects:      20   Tracing       : Off  Lpadd:     704

Now free an object from the first slot of the first slab

  # echo 'free 1' > callfn
  # cat /proc/slabinfo | grep smo_test | sed 's/:.*//'
  smo_test              20     40    392   20    2

  # slabinfo smo_test --report

Slabcache: smo_test Aliases: 0 Order : 1 Objects: 20

Sizes (bytes) Slabs Debug Memory
------------------------------------------------------------------------

  Object :      56  Total  :       2   Sanity Checks : On   Total:   16384
  SlabObj:     392  Full   :       0   Redzoning     : On   Used :    1120
  SlabSiz:    8192  Partial:       2   Poisoning     : On   Loss :   15264
  Loss   :     336  CpuSlab:       0   Tracking      : On   Lalig:    6720
  Align  :       8  Objects:      20   Tracing       : Off  Lpadd:     704

Calling shrink now on the cache does nothing because object migration is not enabled (output omitted). If we enable object migration then shrink the cache we expect the object from the second slab to me moved to the first slot in the first slab and the second slab to be removed from the partial list.

  # echo 'enable' > callfn
  # slabinfo smo_test --shrink
  # slabinfo smo_test --report

Slabcache: smo_test Aliases: 0 Order : 1 Objects: 20
** Defragmentation at 30%

Sizes (bytes) Slabs Debug Memory
------------------------------------------------------------------------

  Object :      56  Total  :       1   Sanity Checks : On   Total:    8192
  SlabObj:     392  Full   :       1   Redzoning     : On   Used :    1120
  SlabSiz:    8192  Partial:       0   Poisoning     : On   Loss :    7072
  Loss   :     336  CpuSlab:       0   Tracking      : On   Lalig:    6720
  Align  :       8  Objects:      20   Tracing       : Off  Lpadd:     352

We can run the stress tests (with the default number of objects):

  # cd /sys/kernel/debug/smo
  # echo 'test' > callfn
  [    3.576617] smo: test using nr_objs: 1000 keep: 10
  [    3.580169] smo: Module tests completed successfully

---
 tools/testing/slab/Makefile      |  10 +
 tools/testing/slab/slub_defrag.c | 566 +++++++++++++++++++++++++++++++
 2 files changed, 576 insertions(+)
 create mode 100644 tools/testing/slab/Makefile
 create mode 100644 tools/testing/slab/slub_defrag.c

new file mode 100644
index 000000000000..440c2e3e356f
@@ -0,0 +1,10 @@
+obj-m += slub_defrag.o
+
+KTREE=../../..
+
+all:
+	make -C ${KTREE} M=$(PWD) modules
+
+clean:
+	make -C ${KTREE} M=$(PWD) clean
+

2.21.0

We just added a module that enables testing the SLUB allocators ability to defrag/shrink caches via movable objects. Tests are better when they are automated.

Add automated testing via a python script for SLUB movable objects.

Example output:

  $ cd path/to/linux/tools/testing/slab
  $ /slub_defrag.py
  Please run script as root

  $ sudo ./slub_defrag.py
  <test are quiet, no output on success>

  $ sudo ./slub_defrag.py --debug
  Loading module ...
  Slab cache smo_test created
  Objects per slab: 20
  Running sanity checks ...

Running module stress test (see dmesg for additional test output) ... Removing module slub_defrag ...
Loading module ...
Slab cache smo_test created

Running test non-movable ...
testing slab 'smo_test' prior to enabling movable objects ... verified non-movable slabs are NOT shrinkable

Running test movable ...
testing slab 'smo_test' after enabling movable objects ... verified movable slabs are shrinkable

Removing module slub_defrag ...

---
 tools/testing/slab/slub_defrag.c  |   1 +
 tools/testing/slab/slub_defrag.py | 451 ++++++++++++++++++++++++++++++
 2 files changed, 452 insertions(+)
 create mode 100755 tools/testing/slab/slub_defrag.py

index 4a5c24394b96..8332e69ee868 100644
@@ -337,6 +337,7 @@ static int smo_run_module_tests(int nr_objs, int keep)
 
 /*
  * struct functions() - Map command to a function pointer.
+ * If you update this please update the documentation in slub_defrag.py
  */
 struct functions {
 	char *fn_name;

2.21.0

Earlier, Slab Movable Objects (SMO) was implemented. The XArray is now able to take advantage of SMO in order to make xarray nodes movable (when using the SLUB allocator).

Currently the radix tree uses the same slab cache as the XArray. Only XArray nodes are movable _not_ radix tree nodes. We can give the radix tree its own slab cache to overcome this.

In preparation for implementing XArray object migration (xa_node objects) via Slab Movable Objects add a slab cache solely for XArray nodes and make the XArray use this slab cache instead of the radix_tree_node slab cache.

---
 include/linux/xarray.h |  3 +++
 init/main.c            |  2 ++
 lib/radix-tree.c       |  2 +-
 lib/xarray.c           | 48 ++++++++++++++++++++++++++++++++++--------
 4 files changed, 45 insertions(+), 10 deletions(-)

index 0e01e6129145..773f91f8e1db 100644
@@ -42,6 +42,9 @@
 
 #define BITS_PER_XA_VALUE	(BITS_PER_LONG - 1)
 
+/* Called from init/main.c */
+void xarray_slabcache_init(void);
+
 /**
  * xa_mk_value() - Create an XArray entry from an integer.
  * @v: Value to store in XArray.

index 66a196c5e4c3..8c409a5dc937 100644
@@ -107,6 +107,7 @@ static int kernel_init(void *);
 
 extern void init_IRQ(void);
 extern void radix_tree_init(void);
+extern void xarray_slabcache_init(void);
 
 /*
  * Debug helper: via this flag we know that we are in 'early bootup code'

@@ -622,6 +623,7 @@ asmlinkage __visible void __init start_kernel(void)
 		 "Interrupts were enabled *very* early, fixing it\n"))
 		local_irq_disable();
 	radix_tree_init();
+	xarray_slabcache_init();
 
 	/*
 	 * Set up housekeeping before setting up workqueues to allow the unbound

index 18c1dfbb1765..e6127c4c84b5 100644
@@ -31,7 +31,7 @@
 /*
  * Radix tree node cache.
  */
-struct kmem_cache *radix_tree_node_cachep;
+static struct kmem_cache *radix_tree_node_cachep;
 
 /*
  * The radix tree is variable-height, so an insert operation not only has

2.21.0

Recently Slab Movable Objects (SMO) was implemented for the SLUB allocator. The XArray can take advantage of this and make the xa_node slab cache objects movable.

Implement functions to migrate objects and activate SMO when we initialise the XArray slab cache.

This is based on initial code by Matthew Wilcox and was modified to work with slab object migration.

---
 lib/xarray.c | 61 ++++++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 61 insertions(+)

index 861c042daa1d..9354e0f01f26 100644
@@ -1993,12 +1993,73 @@ static void xa_node_ctor(void *arg)
 	INIT_LIST_HEAD(&node->private_list);
 }
 
+static void xa_object_migrate(struct xa_node *node, int numa_node)
+{
+	struct xarray *xa = READ_ONCE(node->array);
+	void __rcu **slot;
+	struct xa_node *new_node;
+	int i;
+
+	/* Freed or not yet in tree then skip */
+	if (!xa || xa == XA_RCU_FREE)
+		return;
+
+	new_node = kmem_cache_alloc_node(xa_node_cachep, GFP_KERNEL, numa_node);
+	if (!new_node) {
+		pr_err("%s: slab cache allocation failed\n", __func__);
+		return;
+	}
+
+	xa_lock_irq(xa);
+
+	/* Check again..... */
+	if (xa != node->array) {
+		node = new_node;
+		goto unlock;
+	}
+
+	memcpy(new_node, node, sizeof(struct xa_node));
+
+	if (list_empty(&node->private_list))
+		INIT_LIST_HEAD(&new_node->private_list);
+	else
+		list_replace(&node->private_list, &new_node->private_list);
+
+	for (i = 0; i < XA_CHUNK_SIZE; i++) {
+		void *x = xa_entry_locked(xa, new_node, i);
+
+		if (xa_is_node(x))
+			rcu_assign_pointer(xa_to_node(x)->parent, new_node);
+	}
+	if (!new_node->parent)
+		slot = &xa->xa_head;
+	else
+		slot = &xa_parent_locked(xa, new_node)->slots[new_node->offset];
+	rcu_assign_pointer(*slot, xa_mk_node(new_node));
+
+unlock:
+	xa_unlock_irq(xa);
+	xa_node_free(node);
+	rcu_barrier();
+}
+
+static void xa_migrate(struct kmem_cache *s, void **objects, int nr,
+		       int node, void *_unused)
+{
+	int i;
+
+	for (i = 0; i < nr; i++)
+		xa_object_migrate(objects[i], node);
+}
+
 void __init xarray_slabcache_init(void)
 {
 	xa_node_cachep = kmem_cache_create("xarray_node",
 					   sizeof(struct xa_node), 0,
 					   SLAB_PANIC | SLAB_RECLAIM_ACCOUNT,
 					   xa_node_ctor);
+
+	kmem_cache_setup_mobility(xa_node_cachep, NULL, xa_migrate);
 }
 
 #ifdef XA_DEBUG

2.21.0

We just implemented movable objects for the XArray. Let's test it intree.

Add test module for the XArray's movable objects implementation.

Functionality of the XArray Slab Movable Object implementation can usually be seen by simply by using slabinfo on a running machine since the radix tree is typically in use on a running machine and will have partial slabs. For repeated testing we can use the test module to run to simulate a workload on the XArray then use slabinfo to test object migration is functioning.

If testing on freshly spun up VM (low radix tree workload) it may be necessary to load/unload the module a number of times to create partial slabs.

Example test session
--------------------

Relevant /proc/slabinfo column headers:

name <active_objs> <num_objs> <objsize> <objperslab> <pagesperslab>

Prior to testing slabinfo report for radix_tree_node:

  # slabinfo radix_tree_node --report

Slabcache: radix_tree_node Aliases: 0 Order : 2 Objects: 8352 ** Reclaim accounting active
** Defragmentation at 30%

Sizes (bytes) Slabs Debug Memory
------------------------------------------------------------------------

  Object :     576  Total  :     497   Sanity Checks : On   Total: 8142848
  SlabObj:     912  Full   :     473   Redzoning     : On   Used : 4810752
  SlabSiz:   16384  Partial:      24   Poisoning     : On   Loss : 3332096
  Loss   :     336  CpuSlab:       0   Tracking      : On   Lalig: 2806272
  Align  :       8  Objects:      17   Tracing       : Off  Lpadd:  437360

Here you can see the kernel was built with Slab Movable Objects enabled for the XArray (XArray uses the radix tree below the surface).

After inserting the test module (note we have triggered allocation of a number of radix tree nodes increasing the object count but decreasing the number of partial slabs):

  # slabinfo radix_tree_node --report

Slabcache: radix_tree_node Aliases: 0 Order : 2 Objects: 8442 ** Reclaim accounting active
** Defragmentation at 30%

Sizes (bytes) Slabs Debug Memory
------------------------------------------------------------------------

  Object :     576  Total  :     499   Sanity Checks : On   Total: 8175616
  SlabObj:     912  Full   :     484   Redzoning     : On   Used : 4862592
  SlabSiz:   16384  Partial:      15   Poisoning     : On   Loss : 3313024
  Loss   :     336  CpuSlab:       0   Tracking      : On   Lalig: 2836512
  Align  :       8  Objects:      17   Tracing       : Off  Lpadd:  439120

Now we can shrink the radix_tree_node cache:

  # slabinfo radix_tree_node --shrink
  # slabinfo radix_tree_node --report

Slabcache: radix_tree_node Aliases: 0 Order : 2 Objects: 8515 ** Reclaim accounting active
** Defragmentation at 30%

Sizes (bytes) Slabs Debug Memory
------------------------------------------------------------------------

  Object :     576  Total  :     501   Sanity Checks : On   Total: 8208384
  SlabObj:     912  Full   :     500   Redzoning     : On   Used : 4904640
  SlabSiz:   16384  Partial:       1   Poisoning     : On   Loss : 3303744
  Loss   :     336  CpuSlab:       0   Tracking      : On   Lalig: 2861040
  Align  :       8  Objects:      17   Tracing       : Off  Lpadd:  440880

Note the single remaining partial slab.

---
 tools/testing/slab/Makefile             |   2 +-
 tools/testing/slab/slub_defrag_xarray.c | 211 ++++++++++++++++++++++++
 2 files changed, 212 insertions(+), 1 deletion(-)
 create mode 100644 tools/testing/slab/slub_defrag_xarray.c

index 440c2e3e356f..44c18d9a4d52 100644
@@ -1,4 +1,4 @@
-obj-m += slub_defrag.o
+obj-m += slub_defrag.o slub_defrag_xarray.o
 
 KTREE=../../..
 

2.21.0

In order to support object migration on the dentry cache we need to have a determined object state at all times. Without a constructor the object would have a random state after allocation.

Provide a dentry constructor.

---
 fs/dcache.c | 30 +++++++++++++++++++++---------
 1 file changed, 21 insertions(+), 9 deletions(-)

index c435398f2c81..867d97a86940 100644
@@ -1603,6 +1603,16 @@ void d_invalidate(struct dentry *dentry)
 }
 EXPORT_SYMBOL(d_invalidate);
 
+static void dcache_ctor(void *p)
+{
+	struct dentry *dentry = p;
+
+	/* Mimic lockref_mark_dead() */
+	dentry->d_lockref.count = -128;
+
+	spin_lock_init(&dentry->d_lock);
+}
+
 /**
  * __d_alloc	-	allocate a dcache entry
  * @sb: filesystem it will belong to

@@ -1658,7 +1668,6 @@ struct dentry *__d_alloc(struct super_block *sb, const struct qstr *name)
 
 	dentry->d_lockref.count = 1;
 	dentry->d_flags = 0;
-	spin_lock_init(&dentry->d_lock);
 	seqcount_init(&dentry->d_seq);
 	dentry->d_inode = NULL;
 	dentry->d_parent = dentry;

@@ -3096,14 +3105,17 @@ static void __init dcache_init_early(void)
 
 static void __init dcache_init(void)
 {
-	/*
-	 * A constructor could be added for stable state like the lists,
-	 * but it is probably not worth it because of the cache nature
-	 * of the dcache.
-	 */
-	dentry_cache = KMEM_CACHE_USERCOPY(dentry,
-		SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_MEM_SPREAD|SLAB_ACCOUNT,
-		d_iname);
+	slab_flags_t flags =
+		SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | SLAB_MEM_SPREAD | SLAB_ACCOUNT;
+
+	dentry_cache =
+		kmem_cache_create_usercopy("dentry",
+					   sizeof(struct dentry),
+					   __alignof__(struct dentry),
+					   flags,
+					   offsetof(struct dentry, d_iname),
+					   sizeof_field(struct dentry, d_iname),
+					   dcache_ctor);
 
 	/* Hash may have been set up in dcache_init_early */
 	if (!hashdist)

2.21.0

The dentry slab cache is susceptible to internal fragmentation. Now that we have Slab Movable Objects we can attempt to defragment the dcache. Dentry objects are inherently _not_ relocatable however under some conditions they can be free'd. This is the same as shrinking the dcache but instead of shrinking the whole cache we only attempt to free those objects that are located in partially full slab pages. There is no guarantee that this will reduce the memory usage of the system, it is a compromise between fragmented memory and total cache shrinkage with the hope that some memory pressure can be alleviated.

This is implemented using the newly added Slab Movable Objects infrastructure. The dcache 'migration' function is intentionally _not_ called 'd_migrate' because we only free, we do not migrate. Call it 'd_partial_shrink' to make explicit that no reallocation is done.

In order to enable SMO a call to kmem_cache_setup_mobility() must be made, we do this during initialization of the dcache.

Implement isolate and 'migrate' functions for the dentry slab cache. Enable SMO for the dcache during initialization.

---
 fs/dcache.c | 75 +++++++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 75 insertions(+)

index 867d97a86940..3ca721752723 100644
@@ -3072,6 +3072,79 @@ void d_tmpfile(struct dentry *dentry, struct inode *inode)
 }
 EXPORT_SYMBOL(d_tmpfile);
 
+/*
+ * d_isolate() - Dentry isolation callback function.
+ * @s: The dentry cache.
+ * @v: Vector of pointers to the objects to isolate.
+ * @nr: Number of objects in @v.
+ *
+ * The slab allocator is holding off frees. We can safely examine
+ * the object without the danger of it vanishing from under us.
+ */
+static void *d_isolate(struct kmem_cache *s, void **v, int nr)
+{
+	struct list_head *dispose;
+	struct dentry *dentry;
+	int i;
+
+	dispose = kmalloc(sizeof(*dispose), GFP_KERNEL);
+	if (!dispose)
+		return NULL;
+
+	INIT_LIST_HEAD(dispose);
+
+	for (i = 0; i < nr; i++) {
+		dentry = v[i];
+		spin_lock(&dentry->d_lock);
+
+		if (dentry->d_lockref.count > 0 ||
+		    dentry->d_flags & DCACHE_SHRINK_LIST) {
+			spin_unlock(&dentry->d_lock);
+			continue;
+		}
+
+		if (dentry->d_flags & DCACHE_LRU_LIST)
+			d_lru_del(dentry);
+
+		d_shrink_add(dentry, dispose);
+		spin_unlock(&dentry->d_lock);
+	}
+
+	return dispose;
+}
+
+/*
+ * d_partial_shrink() - Dentry migration callback function.
+ * @s: The dentry cache.
+ * @_unused: We do not access the vector.
+ * @__unused: No need for length of vector.
+ * @___unused: We do not do any allocation.
+ * @private: list_head pointer representing the shrink list.
+ *
+ * Dispose of the shrink list created during isolation function.
+ *
+ * Dentry objects can _not_ be relocated and shrinking the whole dcache
+ * can be expensive.  This is an effort to free dentry objects that are
+ * stopping slab pages from being free'd without clearing the whole dcache.
+ *
+ * This callback is called from the SLUB allocator object migration
+ * infrastructure in attempt to free up slab pages by freeing dentry
+ * objects from partially full slabs.
+ */
+static void d_partial_shrink(struct kmem_cache *s, void **_unused, int __unused,
+			     int ___unused, void *private)
+{
+	struct list_head *dispose = private;
+
+	if (!private)		/* kmalloc error during isolate. */
+		return;
+
+	if (!list_empty(dispose))
+		shrink_dentry_list(dispose);
+
+	kfree(private);
+}
+
 static __initdata unsigned long dhash_entries;
 static int __init set_dhash_entries(char *str)
 {

@@ -3117,6 +3190,8 @@ static void __init dcache_init(void)
 					   sizeof_field(struct dentry, d_iname),
 					   dcache_ctor);
 
+	kmem_cache_setup_mobility(dentry_cache, d_isolate, d_partial_shrink);
+
 	/* Hash may have been set up in dcache_init_early */
 	if (!hashdist)
 		return;

2.21.0

We have just implemented Slab Movable Objects (SMO, object migration). Currently object migration is used to defrag a cache. On NUMA systems it would be nice to be able to control the source and destination nodes when moving objects.

Add CONFIG_SLUB_SMO_NODE to guard this feature. CONFIG_SLUB_SMO_NODE depends on CONFIG_SLUB_DEBUG because we use the full list.

Implement moving all objects (including those in full slabs) to a specific node. Expose this functionality to userspace via a sysfs entry.

Add sysfs entry:

/sysfs/kernel/slab/<cache>/move

With this users get access to the following functionality:

- Move all objects to specified node.

echo "N1" > move

- Move all objects from specified node to other specified node (from N1 -> to N2):

echo "N1 N2" > move

This also enables shrinking slabs on a specific node:

echo "N1 N1" > move

---
 mm/Kconfig |   7 ++
 mm/slub.c  | 247 +++++++++++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 254 insertions(+)

index f0c76ba47695..c1438b9e578b 100644
@@ -259,6 +259,13 @@ config ARCH_ENABLE_THP_MIGRATION
 config CONTIG_ALLOC
        def_bool (MEMORY_ISOLATION && COMPACTION) || CMA
 
+config SLUB_SMO_NODE
+       bool "Enable per node control of Slab Movable Objects"
+       depends on SLUB && SYSFS
+       select SLUB_DEBUG
+       help
+         On NUMA systems enable moving objects to and from a specified node.
+
 config PHYS_ADDR_T_64BIT
 	def_bool 64BIT
 

2.21.0

We have just implemented Slab Movable Objects (SMO). On NUMA systems slabs can become unbalanced i.e. many slabs on one node while other nodes have few slabs. Using SMO we can balance the slabs across all the nodes.

The algorithm used is as follows:

1. Move all objects to node 0 (this has the effect of defragmenting the

    cache).

2. Calculate the desired number of slabs for each node (this is done using the approximation nr_slabs / nr_nodes).

3. Loop over the nodes moving the desired number of slabs from node 0 to the node.

Feature is conditionally built in with CONFIG_SMO_NODE, this is because we need the full list (we enable SLUB_DEBUG to get this). Future version may separate final list out of SLUB_DEBUG.

Expose this functionality to userspace via a sysfs entry. Add sysfs entry:

/sysfs/kernel/slab/<cache>/balance

Write of '1' to this file triggers balance, no other value accepted.

This feature relies on SMO being enable for the cache, this is done with a call to, after the isolate/migrate functions have been defined.

kmem_cache_setup_mobility(s, isolate, migrate)

---
 mm/slub.c | 130 ++++++++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 130 insertions(+)

index 23566e5a712b..70e46c4db757 100644
@@ -4458,6 +4458,119 @@ static unsigned long kmem_cache_move_to_node(struct kmem_cache *s, int node)
 
 	return left;
 }
+
+/*
+ * kmem_cache_move_slabs() - Attempt to move @num slabs to target_node,
+ * @s: The cache we are working on.
+ * @node: The node to move objects from.
+ * @target_node: The node to move objects to.
+ * @num: The number of slabs to move.
+ *
+ * Attempts to move @num slabs from @node to @target_node.  This is done
+ * by migrating objects from slabs on the full_list.
+ *
+ * Return: The number of slabs moved or error code.
+ */
+static long kmem_cache_move_slabs(struct kmem_cache *s,
+				  int node, int target_node, long num)
+{
+	struct kmem_cache_node *n = get_node(s, node);
+	LIST_HEAD(move_list);
+	struct page *page, *page2;
+	unsigned long flags;
+	void **scratch;
+	long done = 0;
+
+	if (!s->migrate) {
+		pr_warn("%s SMO not enabled, cannot move objects\n", s->name);
+		goto out;
+	}
+
+	if (node == target_node)
+		return -EINVAL;
+
+	scratch = alloc_scratch(s);
+	if (!scratch)
+		return -ENOMEM;
+
+	spin_lock_irqsave(&n->list_lock, flags);
+
+	list_for_each_entry_safe(page, page2, &n->full, lru) {
+		if (!slab_trylock(page))
+			/* Busy slab. Get out of the way */
+			continue;
+
+		list_move(&page->lru, &move_list);
+		page->frozen = 1;
+		slab_unlock(page);
+
+		if (++done >= num)
+			break;
+	}
+	spin_unlock_irqrestore(&n->list_lock, flags);
+
+	list_for_each_entry(page, &move_list, lru) {
+		if (page->inuse)
+			move_slab_page(page, scratch, target_node);
+	}
+	kfree(scratch);
+
+	/* Bail here to save taking the list_lock */
+	if (list_empty(&move_list))
+		goto out;
+
+	/* Inspect results and dispose of pages */
+	spin_lock_irqsave(&n->list_lock, flags);
+	list_for_each_entry_safe(page, page2, &move_list, lru) {
+		list_del(&page->lru);
+		slab_lock(page);
+		page->frozen = 0;
+
+		if (page->inuse) {
+			/*
+			 * This is best effort only, if slab still has
+			 * objects just put it back on the partial list.
+			 */
+			n->nr_partial++;
+			list_add_tail(&page->lru, &n->partial);
+			slab_unlock(page);
+		} else {
+			slab_unlock(page);
+			discard_slab(s, page);
+		}
+	}
+	spin_unlock_irqrestore(&n->list_lock, flags);
+out:
+	return done;
+}
+
+/*
+ * kmem_cache_balance_nodes() - Balance slabs across nodes.
+ * @s: The cache we are working on.
+ */
+static void kmem_cache_balance_nodes(struct kmem_cache *s)
+{
+	struct kmem_cache_node *n = get_node(s, 0);
+	unsigned long desired_nr_slabs_per_node;
+	unsigned long nr_slabs;
+	int nr_nodes = 0;
+	int nid;
+
+	(void)kmem_cache_move_to_node(s, 0);
+
+	for_each_node_state(nid, N_NORMAL_MEMORY)
+		nr_nodes++;
+
+	nr_slabs = atomic_long_read(&n->nr_slabs);
+	desired_nr_slabs_per_node = nr_slabs / nr_nodes;
+
+	for_each_node_state(nid, N_NORMAL_MEMORY) {
+		if (nid == 0)
+			continue;
+
+		kmem_cache_move_slabs(s, 0, nid, desired_nr_slabs_per_node);
+	}
+}
 #endif	/* CONFIG_SLUB_SMO_NODE */
 
 /*

@@ -5836,6 +5949,22 @@ static ssize_t move_store(struct kmem_cache *s, const char *buf, size_t length)
 	return length;
 }
 SLAB_ATTR(move);
+
+static ssize_t balance_show(struct kmem_cache *s, char *buf)
+{
+	return 0;
+}
+
+static ssize_t balance_store(struct kmem_cache *s,
+			     const char *buf, size_t length)
+{
+	if (buf[0] == '1')
+		kmem_cache_balance_nodes(s);
+	else
+		return -EINVAL;
+	return length;
+}
+SLAB_ATTR(balance);
 #endif	/* CONFIG_SLUB_SMO_NODE */
 
 #ifdef CONFIG_NUMA

@@ -5964,6 +6093,7 @@ static struct attribute *slab_attrs[] = {
 	&shrink_attr.attr,
 #ifdef CONFIG_SLUB_SMO_NODE
 	&move_attr.attr,
+	&balance_attr.attr,
 #endif
 	&slabs_cpu_partial_attr.attr,
 #ifdef CONFIG_SLUB_DEBUG

2.21.0