this_cpu_cmpxchg_double support

The following patch will make the fastpaths lockless and will no longer require interrupts to be disabled. Calling the free hook with irq disabled will no longer be possible.

Move the slab_free_hook_irq() logic into slab_free_hook. Only disable interrupts if the features are selected that require callbacks with interrupts off and reenable after calls have been made.

Signed-off-by Christoph Lameter <[email protected]>

---
 mm/slub.c |   29 +++++++++++++++++------------
 1 file changed, 17 insertions(+), 12 deletions(-)

Index: linux-2.6/mm/slub.c
===================================================================

mm/slub.c +0 -0

@@ -807,14 +807,24 @@ static inline void slab_post_alloc_hook(
 static inline void slab_free_hook(struct kmem_cache *s, void *x)
 {
 	kmemleak_free_recursive(x, s->flags);

mm/slub.c +15 -5

+	 * So in order to make the debug calls that expect irqs to be
+	 * disabled we need to disable interrupts temporarily.
+	 */
+#if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
+	{
+		unsigned long flags;
+
+		local_irq_save(flags);
+		kmemcheck_slab_free(s, x, s->objsize);
+		debug_check_no_locks_freed(x, s->objsize);
+		if (!(s->flags & SLAB_DEBUG_OBJECTS))
+			debug_check_no_obj_freed(x, s->objsize);
+		local_irq_restore(flags);
+	}
+#endif
 }
 
 /*

@@ -1101,9 +1111,6 @@ static inline void slab_post_alloc_hook(
 
 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
 
-static inline void slab_free_hook_irq(struct kmem_cache *s,
-		void *object) {}
-
 #endif /* CONFIG_SLUB_DEBUG */
 
 /*

@@ -1909,8 +1916,6 @@ static __always_inline void slab_free(st
 	local_irq_save(flags);
 	c = __this_cpu_ptr(s->cpu_slab);
 
-	slab_free_hook_irq(s, x);
-
 	if (likely(page == c->page && c->node != NUMA_NO_NODE)) {
 		set_freepointer(s, object, c->freelist);
 		c->freelist = object;

Introduce this_cpu_cmpxchg_double. this_cpu_cmpxchg_double() allows the comparision between two consecutive words and replaces them if there is

a match.

	bool this_cpu_cmpxchg_double(pcp1, pcp2,
		old_word1, old_word2, new_word1, new_word2)

this_cpu_cmpxchg_double does not return the old value (difficult since there are two words) but a boolean indicating if the operation was successful.

The first percpu variable must be double word aligned!

Signed-off-by Christoph Lameter <[email protected]>

---
 include/linux/percpu.h |  130 +++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 130 insertions(+)

Index: linux-2.6/include/linux/percpu.h
===================================================================

include/linux/percpu.h +130 -0

@@ -255,6 +255,29 @@ extern void __bad_size_call_parameter(vo
 	pscr2_ret__;							\
 })
 
+/*
+ * Special handling for cmpxchg_double. cmpxchg_double is passed two
+ * percpu variables. The first has to be aligned to a double word
+ * boundary and the second has to follow directly thereafter.
+ */
+#define __pcpu_double_call_return_int(stem, pcp1, pcp2, ...)		\
+({									\
+	int ret__;							\
+	__verify_pcpu_ptr(&pcp1);					\
+	VM_BUG_ON((unsigned long)(&pcp1) % (2 * sizeof(pcp1)));		\
+	VM_BUG_ON((unsigned long)(&pcp2) != (unsigned long)(&pcp1) + sizeof(pcp1));\
+	VM_BUG_ON(sizeof(pcp1) != sizeof(pcp2));			\
+	switch(sizeof(pcp1)) {						\
+	case 1: ret__ = stem##1(pcp1, pcp2, __VA_ARGS__);break;		\
+	case 2: ret__ = stem##2(pcp1, pcp2, __VA_ARGS__);break;		\
+	case 4: ret__ = stem##4(pcp1, pcp2, __VA_ARGS__);break;		\
+	case 8: ret__ = stem##8(pcp1, pcp2, __VA_ARGS__);break;		\
+	default:							\
+		__bad_size_call_parameter();break;			\
+	}								\
+	ret__;								\
+})
+
 #define __pcpu_size_call(stem, variable, ...)				\
 do {									\
 	__verify_pcpu_ptr(&(variable));					\

@@ -318,6 +341,80 @@ do {									\
 # define this_cpu_read(pcp)	__pcpu_size_call_return(this_cpu_read_, (pcp))
 #endif
 
+/*
+ * cmpxchg_double replaces two adjacent scalars at once. The first two
+ * parameters are per cpu variables which have to be of the same size.
+ * A truth value is returned to indicate success or
+ * failure (since a double register result is difficult to handle).
+ * There is very limited hardware support for these operations. So only certain
+ * sizes may work.
+ */
+#define __this_cpu_generic_cmpxchg_double(pcp1, pcp2, oval1, oval2, nval1, nval2)	\
+({									\
+	int __ret = 0;							\
+	if (__this_cpu_read(pcp1) == (oval1) &&				\
+			 __this_cpu_read(pcp2)  == (oval2)) {		\
+		__this_cpu_write(pcp1, (nval1));			\
+		__this_cpu_write(pcp2, (nval2));			\
+		__ret = 1;						\
+	}								\
+	(__ret);							\
+})
+
+#ifndef __this_cpu_cmpxchg_double
+# ifndef __this_cpu_cmpxchg_double_1
+#  define __this_cpu_cmpxchg_double_1(pcp1, pcp2, oval1, oval2, nval1, nval2)	\
+	__this_cpu_generic_cmpxchg_double(pcp1, pcp2, oval1, oval2, nval1, nval2)
+# endif
+# ifndef __this_cpu_cmpxchg_double_2
+#  define __this_cpu_cmpxchg_double_2(pcp1, pcp2, oval1, oval2, nval1, nval2)	\
+	__this_cpu_generic_cmpxchg_double(pcp1, pcp2, oval1, oval2, nval1, nval2)
+# endif
+# ifndef __this_cpu_cmpxchg_double_4
+#  define __this_cpu_cmpxchg_double_4(pcp1, pcp2, oval1, oval2, nval1, nval2)	\
+	__this_cpu_generic_cmpxchg_double(pcp1, pcp2, oval1, oval2, nval1, nval2)
+# endif
+# ifndef __this_cpu_cmpxchg_double_8
+#  define __this_cpu_cmpxchg_double_8(pcp1, pcp2, oval1, oval2, nval1, nval2)	\
+	__this_cpu_generic_cmpxchg_double(pcp1, pcp2, oval1, oval2, nval1, nval2)
+# endif
+# define __this_cpu_cmpxchg_double(pcp1, pcp2, oval1, oval2, nval1, nval2)	\
+	__pcpu_double_call_return_int(__this_cpu_cmpxchg_double_, (pcp1), (pcp2)	\
+					 oval1, oval2, nval1, nval2)
+#endif
+
+#define _this_cpu_generic_cmpxchg_double(pcp1, pcp2, oval1, oval2, nval1, nval2)	\
+({									\
+	int ret__;							\
+	preempt_disable();						\
+	ret__ = __this_cpu_generic_cmpxchg_double(pcp1, pcp2,		\
+			oval1, oval2, nval1, nval2);			\
+	preempt_enable();						\
+	ret__;								\
+})
+
+#ifndef this_cpu_cmpxchg_double
+# ifndef this_cpu_cmpxchg_double_1
+#  define this_cpu_cmpxchg_double_1(pcp1, pcp2, oval1, oval2, nval1, nval2)	\
+	_this_cpu_generic_cmpxchg_double(pcp1, pcp2, oval1, oval2, nval1, nval2)
+# endif
+# ifndef this_cpu_cmpxchg_double_2
+#  define this_cpu_cmpxchg_double_2(pcp1, pcp2, oval1, oval2, nval1, nval2)	\
+	_this_cpu_generic_cmpxchg_double(pcp1, pcp2, oval1, oval2, nval1, nval2)
+# endif
+# ifndef this_cpu_cmpxchg_double_4
+#  define this_cpu_cmpxchg_double_4(pcp1, pcp2, oval1, oval2, nval1, nval2)	\
+	_this_cpu_generic_cmpxchg_double(pcp1, pcp2, oval1, oval2, nval1, nval2)
+# endif
+# ifndef this_cpu_cmpxchg_double_8
+#  define this_cpu_cmpxchg_double_8(pcp1, pcp2, oval1, oval2, nval1, nval2)	\
+	_this_cpu_generic_cmpxchg_double(pcp1, pcp2, oval1, oval2, nval1, nval2)
+# endif
+# define this_cpu_cmpxchg_double(pcp1, pcp2, oval1, oval2, nval1, nval2)	\
+	__pcpu_double_call_return_int(this_cpu_cmpxchg_double_, (pcp1), (pcp2),	\
+		oval1, oval2, nval1, nval2)
+#endif
+
 #define _this_cpu_generic_to_op(pcp, val, op)				\
 do {									\
 	preempt_disable();						\

@@ -823,4 +920,37 @@ do {									\
 	__pcpu_size_call_return2(irqsafe_cpu_cmpxchg_, (pcp), oval, nval)
 #endif
 
+#define irqsafe_generic_cpu_cmpxchg_double(pcp1, pcp2, oval1, oval2, nval1, nval2)	\
+({									\
+	int ret__;							\
+	unsigned long flags;						\
+	local_irq_save(flags);						\
+	ret__ = __this_cpu_generic_cmpxchg_double(pcp1, pcp2,		\
+			oval1, oval2, nval1, nval2);			\
+	local_irq_restore(flags);					\
+	ret__;								\
+})
+
+#ifndef irqsafe_cpu_cmpxchg_double
+# ifndef irqsafe_cpu_cmpxchg_double_1
+#  define irqsafe_cpu_cmpxchg_double_1(pcp1, pcp2, oval1, oval2, nval1, nval2)	\
+	irqsafe_generic_cpu_cmpxchg_double(pcp1, pcp2, oval1, oval2, nval1, nval2)
+# endif
+# ifndef irqsafe_cpu_cmpxchg_double_2
+#  define irqsafe_cpu_cmpxchg_double_2(pcp1, pcp2, oval1, oval2, nval1, nval2)	\
+	irqsafe_generic_cpu_cmpxchg_double(pcp1, pcp2, oval1, oval2, nval1, nval2)
+# endif
+# ifndef irqsafe_cpu_cmpxchg_double_4
+#  define irqsafe_cpu_cmpxchg_double_4(pcp1, pcp2, oval1, oval2, nval1, nval2)	\
+	irqsafe_generic_cpu_cmpxchg_double(pcp1, pcp2, oval1, oval2, nval1, nval2)
+# endif
+# ifndef irqsafe_cpu_cmpxchg_double_8
+#  define irqsafe_cpu_cmpxchg_double_8(pcp1, pcp2, oval1, oval2, nval1, nval2)	\
+	irqsafe_generic_cpu_cmpxchg_double(pcp1, pcp2, oval1, oval2, nval1, nval2)
+# endif
+# define irqsafe_cpu_cmpxchg_double(pcp1, pcp2, oval1, oval2, nval1, nval2)	\
+	__pcpu_double_call_return_int(irqsafe_cpu_cmpxchg_double_, (pcp1), (pcp2),	\
+		oval1, oval2, nval1, nval2)
+#endif
+
 #endif /* __LINUX_PERCPU_H */

include/linux/percpu.h +11 -0

+	switch(sizeof(pcp1)) {						\
+	case 1: ret__ = stem##1(pcp1, pcp2, __VA_ARGS__);break;		\
+	case 2: ret__ = stem##2(pcp1, pcp2, __VA_ARGS__);break;		\
+	case 4: ret__ = stem##4(pcp1, pcp2, __VA_ARGS__);break;		\
+	case 8: ret__ = stem##8(pcp1, pcp2, __VA_ARGS__);break;		\
+	default:							\
+		__bad_size_call_parameter();break;			\
+	}								\
+	ret__;								\
+})
+

include/linux/percpu.h extern void __bad_size_call_parameter(vo · line 269

Quoted 3 lines

include/linux/percpu.h +3 -0

+	VM_BUG_ON((unsigned long)(&pcp1) % (2 * sizeof(pcp1)));		\
+	VM_BUG_ON((unsigned long)(&pcp2) != (unsigned long)(&pcp1) + sizeof(pcp1));\
+	VM_BUG_ON(sizeof(pcp1) != sizeof(pcp2));			\

Since this is a macro, and it looks like all these are constants (sizeof and addresses), couldn't you just do a BUILD_BUG_ON() instead?

Quoted 6 lines · Steven Rostedt

include/linux/percpu.h extern void __bad_size_call_parameter(vo · line 269

Quoted 3 lines

include/linux/percpu.h +3 -0

+	VM_BUG_ON((unsigned long)(&pcp1) % (2 * sizeof(pcp1)));		\
+	VM_BUG_ON((unsigned long)(&pcp2) != (unsigned long)(&pcp1) + sizeof(pcp1));\
+	VM_BUG_ON(sizeof(pcp1) != sizeof(pcp2));			\

Since this is a macro, and it looks like all these are constants (sizeof and addresses), couldn't you just do a BUILD_BUG_ON() instead?

The addresses are not constant.

I was thinking that if these are per_cpu then they would be global and thus constant. But those are done at link time, which is too late.

Quoted 2 lines · Steven Rostedt

I was thinking that if these are per_cpu then they would be global and thus constant. But those are done at link time, which is too late.

Well per_cpu data can also be dynamically allocated via alloc_percpu. Which is the use case here.

From 7c3343392172ba98d9d90a83edcc4c2e80897009 Mon Sep 17 00:00:00 2001

Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops.

Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu.

A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information.

This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end.

So there is no need even to disable preemption.

Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles.

Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double)

Kmalloc: alloc/free test

Before:

10000 times kmalloc(8)/kfree -> 134 cycles
10000 times kmalloc(16)/kfree -> 152 cycles
10000 times kmalloc(32)/kfree -> 144 cycles
10000 times kmalloc(64)/kfree -> 142 cycles
10000 times kmalloc(128)/kfree -> 142 cycles
10000 times kmalloc(256)/kfree -> 132 cycles
10000 times kmalloc(512)/kfree -> 132 cycles
10000 times kmalloc(1024)/kfree -> 135 cycles
10000 times kmalloc(2048)/kfree -> 135 cycles
10000 times kmalloc(4096)/kfree -> 135 cycles
10000 times kmalloc(8192)/kfree -> 144 cycles
10000 times kmalloc(16384)/kfree -> 754 cycles

After:

10000 times kmalloc(8)/kfree -> 78 cycles
10000 times kmalloc(16)/kfree -> 78 cycles
10000 times kmalloc(32)/kfree -> 82 cycles
10000 times kmalloc(64)/kfree -> 88 cycles
10000 times kmalloc(128)/kfree -> 79 cycles
10000 times kmalloc(256)/kfree -> 79 cycles
10000 times kmalloc(512)/kfree -> 85 cycles
10000 times kmalloc(1024)/kfree -> 82 cycles
10000 times kmalloc(2048)/kfree -> 82 cycles
10000 times kmalloc(4096)/kfree -> 85 cycles
10000 times kmalloc(8192)/kfree -> 82 cycles
10000 times kmalloc(16384)/kfree -> 706 cycles

Kmalloc: Repeatedly allocate then free test

Before:

10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles

After:

10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles

Signed-off-by Christoph Lameter <[email protected]>

---
 include/linux/slub_def.h |    5 -
 mm/slub.c                |  205 ++++++++++++++++++++++++++++++++++++++++++++++-
 2 files changed, 207 insertions(+), 3 deletions(-)

Index: linux-2.6/include/linux/slub_def.h
===================================================================

include/linux/slub_def.h +0 -0

 	struct page *page;	/* The slab from which we are allocating */
 	int node;		/* The node of the page (or -1 for debug) */
 #ifdef CONFIG_SLUB_STATS

Index: linux-2.6/mm/slub.c
===================================================================

mm/slub.c +203 -2

@@ -1494,6 +1494,77 @@ static void unfreeze_slab(struct kmem_ca
 	}
 }
 
+#ifdef CONFIG_CMPXCHG_LOCAL
+#ifdef CONFIG_PREEMPT
+/*
+ * Calculate the next globally unique transaction for disambiguiation
+ * during cmpxchg. The transactions start with the cpu number and are then
+ * incremented by CONFIG_NR_CPUS.
+ */
+#define TID_STEP  roundup_pow_of_two(CONFIG_NR_CPUS)
+#else
+/*
+ * No preemption supported therefore also no need to check for
+ * different cpus.
+ */
+#define TID_STEP 1
+#endif
+
+static inline unsigned long next_tid(unsigned long tid)
+{
+	return tid + TID_STEP;
+}
+
+static inline unsigned int tid_to_cpu(unsigned long tid)
+{
+	return tid % TID_STEP;
+}
+
+static inline unsigned long tid_to_event(unsigned long tid)
+{
+	return tid / TID_STEP;
+}
+
+static inline unsigned int init_tid(int cpu)
+{
+	return cpu;
+}
+
+static inline void note_cmpxchg_failure(const char *n,
+		const struct kmem_cache *s, unsigned long tid)
+{
+#ifdef SLUB_DEBUG_CMPXCHG
+	unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
+
+	printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
+
+#ifdef CONFIG_PREEMPT
+	if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
+		printk("due to cpu change %d -> %d\n",
+			tid_to_cpu(tid), tid_to_cpu(actual_tid));
+	else
+#endif
+	if (tid_to_event(tid) != tid_to_event(actual_tid))
+		printk("due to cpu running other code. Event %ld->%ld\n",
+			tid_to_event(tid), tid_to_event(actual_tid));
+	else
+		printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
+			actual_tid, tid, next_tid(tid));
+#endif
+}
+
+#endif
+
+void init_kmem_cache_cpus(struct kmem_cache *s)
+{
+#if defined(CONFIG_CMPXCHG_LOCAL) && defined(CONFIG_PREEMPT)
+	int cpu;
+
+	for_each_possible_cpu(cpu)
+		per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
+#endif
+
+}
 /*
  * Remove the cpu slab
  */

@@ -1525,6 +1596,9 @@ static void deactivate_slab(struct kmem_
 		page->inuse--;
 	}
 	c->page = NULL;
+#ifdef CONFIG_CMPXCHG_LOCAL
+	c->tid = next_tid(c->tid);
+#endif
 	unfreeze_slab(s, page, tail);
 }
 

@@ -1659,6 +1733,19 @@ static void *__slab_alloc(struct kmem_ca
 {
 	void **object;
 	struct page *new;
+#ifdef CONFIG_CMPXCHG_LOCAL
+	unsigned long flags;
+
+	local_irq_save(flags);
+#ifdef CONFIG_PREEMPT
+	/*
+	 * We may have been preempted and rescheduled on a different
+	 * cpu before disabling interrupts. Need to reload cpu area
+	 * pointer.
+	 */
+	c = this_cpu_ptr(s->cpu_slab);
+#endif
+#endif
 
 	/* We handle __GFP_ZERO in the caller */
 	gfpflags &= ~__GFP_ZERO;

@@ -1685,6 +1772,10 @@ load_freelist:
 	c->node = page_to_nid(c->page);
 unlock_out:
 	slab_unlock(c->page);
+#ifdef CONFIG_CMPXCHG_LOCAL
+	c->tid = next_tid(c->tid);
+	local_irq_restore(flags);
+#endif
 	stat(s, ALLOC_SLOWPATH);
 	return object;
 

@@ -1746,23 +1837,76 @@ static __always_inline void *slab_alloc(
 {
 	void **object;
 	struct kmem_cache_cpu *c;
+#ifdef CONFIG_CMPXCHG_LOCAL
+	unsigned long tid;
+#else
 	unsigned long flags;
+#endif
 
 	if (slab_pre_alloc_hook(s, gfpflags))
 		return NULL;
 
+#ifndef CONFIG_CMPXCHG_LOCAL
 	local_irq_save(flags);
+#else
+redo:
+#endif
+
+	/*
+	 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
+	 * enabled. We may switch back and forth between cpus while
+	 * reading from one cpu area. That does not matter as long
+	 * as we end up on the original cpu again when doing the cmpxchg.
+	 */
 	c = __this_cpu_ptr(s->cpu_slab);
+
+#ifdef CONFIG_CMPXCHG_LOCAL
+	/*
+	 * The transaction ids are globally unique per cpu and per operation on
+	 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
+	 * occurs on the right processor and that there was no operation on the
+	 * linked list in between.
+	 */
+	tid = c->tid;
+	barrier();
+#endif
+
 	object = c->freelist;
 	if (unlikely(!object || !node_match(c, node)))
 
 		object = __slab_alloc(s, gfpflags, node, addr, c);
 
 	else {
+#ifdef CONFIG_CMPXCHG_LOCAL
+		/*
+		 * The cmpxchg will only match if there was no additonal
+		 * operation and if we are on the right processor.
+		 *
+		 * The cmpxchg does the following atomically (without lock semantics!)
+		 * 1. Relocate first pointer to the current per cpu area.
+		 * 2. Verify that tid and freelist have not been changed
+		 * 3. If they were not changed replace tid and freelist
+		 *
+		 * Since this is without lock semantics the protection is only against
+		 * code executing on this cpu *not* from access by other cpus.
+		 */
+		if (unlikely(!this_cpu_cmpxchg_double(
+				s->cpu_slab->freelist, s->cpu_slab->tid,
+				object, tid,
+				get_freepointer(s, object), next_tid(tid)))) {
+
+			note_cmpxchg_failure("slab_alloc", s, tid);
+			goto redo;
+		}
+#else
 		c->freelist = get_freepointer(s, object);
+#endif
 		stat(s, ALLOC_FASTPATH);
 	}
+
+#ifndef CONFIG_CMPXCHG_LOCAL
 	local_irq_restore(flags);
+#endif
 
 	if (unlikely(gfpflags & __GFP_ZERO) && object)
 		memset(object, 0, s->objsize);

@@ -1840,9 +1984,13 @@ static void __slab_free(struct kmem_cach
 {
 	void *prior;
 	void **object = (void *)x;
+#ifdef CONFIG_CMPXCHG_LOCAL
+	unsigned long flags;
 
-	stat(s, FREE_SLOWPATH);
+	local_irq_save(flags);
+#endif
 	slab_lock(page);
+	stat(s, FREE_SLOWPATH);
 
 	if (kmem_cache_debug(s))
 		goto debug;

@@ -1872,6 +2020,9 @@ checks_ok:
 
 out_unlock:
 	slab_unlock(page);
+#ifdef CONFIG_CMPXCHG_LOCAL
+	local_irq_restore(flags);
+#endif
 	return;
 
 slab_empty:

@@ -1883,6 +2034,9 @@ slab_empty:
 		stat(s, FREE_REMOVE_PARTIAL);
 	}
 	slab_unlock(page);
+#ifdef CONFIG_CMPXCHG_LOCAL
+	local_irq_restore(flags);
+#endif
 	stat(s, FREE_SLAB);
 	discard_slab(s, page);
 	return;

@@ -1909,21 +2063,54 @@ static __always_inline void slab_free(st
 {
 	void **object = (void *)x;
 	struct kmem_cache_cpu *c;
+#ifdef CONFIG_CMPXCHG_LOCAL
+	unsigned long tid;
+#else
 	unsigned long flags;
+#endif
 
 	slab_free_hook(s, x);
 
+#ifndef CONFIG_CMPXCHG_LOCAL
 	local_irq_save(flags);
+#endif
+
+redo:
+	/*
+	 * Determine the currently cpus per cpu slab.
+	 * The cpu may change afterward. However that does not matter since
+	 * data is retrieved via this pointer. If we are on the same cpu
+	 * during the cmpxchg then the free will succedd.
+	 */
 	c = __this_cpu_ptr(s->cpu_slab);
 
+#ifdef CONFIG_CMPXCHG_LOCAL
+	tid = c->tid;
+	barrier();
+#endif
+
 	if (likely(page == c->page && c->node != NUMA_NO_NODE)) {
 		set_freepointer(s, object, c->freelist);
+
+#ifdef CONFIG_CMPXCHG_LOCAL
+		if (unlikely(!this_cpu_cmpxchg_double(
+				s->cpu_slab->freelist, s->cpu_slab->tid,
+				c->freelist, tid,
+				object, next_tid(tid)))) {
+
+			note_cmpxchg_failure("slab_free", s, tid);
+			goto redo;
+		}
+#else
 		c->freelist = object;
+#endif
 		stat(s, FREE_FASTPATH);
 	} else
 		__slab_free(s, page, x, addr);
 
+#ifndef CONFIG_CMPXCHG_LOCAL
 	local_irq_restore(flags);
+#endif
 }
 
 void kmem_cache_free(struct kmem_cache *s, void *x)

@@ -2115,9 +2302,23 @@ static inline int alloc_kmem_cache_cpus(
 	BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
 			SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
 
+#ifdef CONFIG_CMPXCHG_LOCAL
+	/*
+	 * Must align to double word boundary for the double cmpxchg instructions
+	 * to work.
+	 */
+	s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu), 2 * sizeof(void *));
+#else
+	/* Regular alignment is sufficient */
 	s->cpu_slab = alloc_percpu(struct kmem_cache_cpu);
+#endif
+
+	if (!s->cpu_slab)
+		return 0;
 
-	return s->cpu_slab != NULL;
+	init_kmem_cache_cpus(s);
+
+	return 1;
 }
 
 static struct kmem_cache *kmem_cache_node;

Quoted 7 lines

Index: linux-2.6/include/linux/slub_def.h
===================================================================

include/linux/slub_def.h +0 -0

@@ -35,7 +35,10 @@ enum stat_item {
 	NR_SLUB_STAT_ITEMS };
 

include/linux/slub_def.h enum stat_item { · line 41

Quoted 6 lines

include/linux/slub_def.h +4 -1

 struct kmem_cache_cpu {
-	void **freelist;	/* Pointer to first free per cpu object */
+	void **freelist;	/* Pointer to next available object */
+#ifdef CONFIG_CMPXCHG_LOCAL
+	unsigned long tid;	/* Globally unique transaction id */
+#endif

There seem to be no strong guarantee that freelist is double-word aligned here.

Quoted 19 lines

include/linux/slub_def.h +15 -1

+#ifdef CONFIG_CMPXCHG_LOCAL
+	/*
+	 * Must align to double word boundary for the double cmpxchg instructions
+	 * to work.
+	 */
+	s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu), 2 * sizeof(void *));
+#else
+	/* Regular alignment is sufficient */
 	s->cpu_slab = alloc_percpu(struct kmem_cache_cpu);
+#endif
+
+	if (!s->cpu_slab)
+		return 0;
 
-	return s->cpu_slab != NULL;
+	init_kmem_cache_cpus(s);
+
+	return 1;
 }

Quoted 7 lines · Mathieu Desnoyers

Quoted 7 lines

Index: linux-2.6/include/linux/slub_def.h
===================================================================

include/linux/slub_def.h +0 -0

@@ -35,7 +35,10 @@ enum stat_item {
 	NR_SLUB_STAT_ITEMS };
 

Quoted 8 lines

include/linux/slub_def.h enum stat_item { · line 41

Quoted 6 lines

include/linux/slub_def.h +4 -1

 struct kmem_cache_cpu {
-	void **freelist;	/* Pointer to first free per cpu object */
+	void **freelist;	/* Pointer to next available object */
+#ifdef CONFIG_CMPXCHG_LOCAL
+	unsigned long tid;	/* Globally unique transaction id */
+#endif

There seem to be no strong guarantee that freelist is double-word aligned here.

The struct kmem_cache_cpu allocation via alloc_percpu() specifies double word alignment. See the remainder of the code quoted by you:

Quoted 20 lines

Quoted 20 lines

include/linux/slub_def.h +15 -1

+#ifdef CONFIG_CMPXCHG_LOCAL
+	/*
+	 * Must align to double word boundary for the double cmpxchg instructions
+	 * to work.
+	 */
+	s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu), 2 * sizeof(void *));
+#else
+	/* Regular alignment is sufficient */
 	s->cpu_slab = alloc_percpu(struct kmem_cache_cpu);
+#endif
+
+	if (!s->cpu_slab)
+		return 0;
 
-	return s->cpu_slab != NULL;
+	init_kmem_cache_cpus(s);
+
+	return 1;
 }

From b9ec40af0e18fb7d02106be148036c2ea490fdf9 Mon Sep 17 00:00:00 2001

This patch series introduces this_cpu_cmpxchg_double().

x86 cpus support cmpxchg16b and cmpxchg8b instuction which are capable of switching two words instead of one during a cmpxchg.
Two words allow to swap more state in an atomic instruction.

this_cpu_cmpxchg_double() is used in the slub allocator to avoid interrupt disable/enable in both alloc and free fastpath. Using the new operation significantly speeds up the fastpaths.