Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/penberg/slab-2.6

* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/penberg/slab-2.6: slob: fix lockup in slob_free() slub: use get_track() slub: rename calculate_min_partial() to set_min_partial() slub: add min_partial sysfs tunable slub: move min_partial to struct kmem_cache SLUB: Fix default slab order for big object sizes SLUB: Do not pass 8k objects through to the page allocator SLUB: Introduce and use SLUB_MAX_SIZE and SLUB_PAGE_SHIFT constants slob: clean up the code SLUB: Use ->objsize from struct kmem_cache_cpu in slab_free()

Linus Torvalds committed 17 years ago

Commit be0ea692 parents 4496d93 15a5b0a3 files · +96 −50Browse files

Unified · Split

3 files

include/linux/slub_def.h +17 −4
mm/slob.c +27 −16
mm/slub.c +52 −30

include/linux/slub_def.h+17−4

@@ -46,7 +46,6 @@ struct kmem_cache_cpu { · 45 unmodified lines
46	46	struct kmem_cache_node {
47	47	spinlock_t list_lock; /* Protect partial list and nr_partial */
48	48	unsigned long nr_partial;
49		− unsigned long min_partial;
50	49	struct list_head partial;
51	50	#ifdef CONFIG_SLUB_DEBUG
52	51	atomic_long_t nr_slabs;
@@ -89,6 +88,7 @@ struct kmem_cache { · 36 unmodified lines
89	88	void (ctor)(void );
90	89	int inuse; /* Offset to metadata */
91	90	int align; /* Alignment */
	91	+ unsigned long min_partial;
92	92	const char name; / Name (only for display!) */
93	93	struct list_head list; /* List of slab caches */
94	94	#ifdef CONFIG_SLUB_DEBUG
@@ -120,11 +120,24 @@ struct kmem_cache { · 25 unmodified lines
120	120
121	121	#define KMALLOC_SHIFT_LOW ilog2(KMALLOC_MIN_SIZE)
122	122
	123	+/*
	124	+ * Maximum kmalloc object size handled by SLUB. Larger object allocations
	125	+ * are passed through to the page allocator. The page allocator "fastpath"
	126	+ * is relatively slow so we need this value sufficiently high so that
	127	+ * performance critical objects are allocated through the SLUB fastpath.
	128	+ *
	129	+ * This should be dropped to PAGE_SIZE / 2 once the page allocator
	130	+ * "fastpath" becomes competitive with the slab allocator fastpaths.
	131	+ */
	132	+#define SLUB_MAX_SIZE (2 * PAGE_SIZE)
	133	+
	134	+#define SLUB_PAGE_SHIFT (PAGE_SHIFT + 2)
	135	+
123	136	/*
124	137	* We keep the general caches in an array of slab caches that are used for
125	138	* 2^x bytes of allocations.
126	139	*/
127		−extern struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1];
	140	+extern struct kmem_cache kmalloc_caches[SLUB_PAGE_SHIFT];
128	141
129	142	/*
130	143	* Sorry that the following has to be that ugly but some versions of GCC
@@ -212,7 +225,7 @@ static __always_inline void *kmalloc_large(size_t size, gfp_t flags) · 81 unmodified lines
212	225	static __always_inline void *kmalloc(size_t size, gfp_t flags)
213	226	{
214	227	if (__builtin_constant_p(size)) {
215		− if (size > PAGE_SIZE)
	228	+ if (size > SLUB_MAX_SIZE)
216	229	return kmalloc_large(size, flags);
217	230
218	231	if (!(flags & SLUB_DMA)) {
@@ -234,7 +247,7 @@ void kmem_cache_alloc_node(struct kmem_cache , gfp_t flags, int node); · 15 unmodified lines
234	247	static __always_inline void *kmalloc_node(size_t size, gfp_t flags, int node)
235	248	{
236	249	if (__builtin_constant_p(size) &&
237		− size <= PAGE_SIZE && !(flags & SLUB_DMA)) {
	250	+ size <= SLUB_MAX_SIZE && !(flags & SLUB_DMA)) {
238	251	struct kmem_cache *s = kmalloc_slab(size);
239	252
240	253	if (!s)

mm/slob.c+27−16

@@ -126,9 +126,9 @@ static LIST_HEAD(free_slob_medium); · 125 unmodified lines
126	126	static LIST_HEAD(free_slob_large);
127	127
128	128	/*
129		− * slob_page: True for all slob pages (false for bigblock pages)
	129	+ * is_slob_page: True for all slob pages (false for bigblock pages)
130	130	*/
131		−static inline int slob_page(struct slob_page *sp)
	131	+static inline int is_slob_page(struct slob_page *sp)
132	132	{
133	133	return PageSlobPage((struct page *)sp);
134	134	}
@@ -143,6 +143,11 @@ static inline void clear_slob_page(struct slob_page *sp) · 8 unmodified lines
143	143	__ClearPageSlobPage((struct page *)sp);
144	144	}
145	145
	146	+static inline struct slob_page slob_page(const void addr)
	147	+{
	148	+ return (struct slob_page *)virt_to_page(addr);
	149	+}
	150	+
146	151	/*
147	152	* slob_page_free: true for pages on free_slob_pages list.
148	153	*/
@@ -230,7 +235,7 @@ static int slob_last(slob_t *s) · 81 unmodified lines
230	235	return !((unsigned long)slob_next(s) & ~PAGE_MASK);
231	236	}
232	237
233		−static void *slob_new_page(gfp_t gfp, int order, int node)
	238	+static void *slob_new_pages(gfp_t gfp, int order, int node)
234	239	{
235	240	void *page;
236	241
@@ -247,12 +252,17 @@ static void *slob_new_page(gfp_t gfp, int order, int node) · 10 unmodified lines
247	252	return page_address(page);
248	253	}
249	254
	255	+static void slob_free_pages(void *b, int order)
	256	+{
	257	+ free_pages((unsigned long)b, order);
	258	+}
	259	+
250	260	/*
251	261	* Allocate a slob block within a given slob_page sp.
252	262	*/
253	263	static void slob_page_alloc(struct slob_page sp, size_t size, int align)
254	264	{
255		− slob_t prev, cur, *aligned = 0;
	265	+ slob_t prev, cur, *aligned = NULL;
256	266	int delta = 0, units = SLOB_UNITS(size);
257	267
258	268	for (prev = NULL, cur = sp->free; ; prev = cur, cur = slob_next(cur)) {
@@ -349,10 +359,10 @@ static void *slob_alloc(size_t size, gfp_t gfp, int align, int node) · 90 unmodified lines
349	359
350	360	/* Not enough space: must allocate a new page */
351	361	if (!b) {
352		− b = slob_new_page(gfp & ~__GFP_ZERO, 0, node);
	362	+ b = slob_new_pages(gfp & ~__GFP_ZERO, 0, node);
353	363	if (!b)
354		− return 0;
355		− sp = (struct slob_page *)virt_to_page(b);
	364	+ return NULL;
	365	+ sp = slob_page(b);
356	366	set_slob_page(sp);
357	367
358	368	spin_lock_irqsave(&slob_lock, flags);
@@ -384,7 +394,7 @@ static void slob_free(void *block, int size) · 25 unmodified lines
384	394	return;
385	395	BUG_ON(!size);
386	396
387		− sp = (struct slob_page *)virt_to_page(block);
	397	+ sp = slob_page(block);
388	398	units = SLOB_UNITS(size);
389	399
390	400	spin_lock_irqsave(&slob_lock, flags);
@@ -393,10 +403,11 @@ static void slob_free(void *block, int size) · 2 unmodified lines
393	403	/* Go directly to page allocator. Do not pass slob allocator */
394	404	if (slob_page_free(sp))
395	405	clear_slob_page_free(sp);
	406	+ spin_unlock_irqrestore(&slob_lock, flags);
396	407	clear_slob_page(sp);
397	408	free_slob_page(sp);
398	409	free_page((unsigned long)b);
399		− goto out;
	410	+ return;
400	411	}
401	412
402	413	if (!slob_page_free(sp)) {
@@ -476,7 +487,7 @@ void *__kmalloc_node(size_t size, gfp_t gfp, int node) · 73 unmodified lines
476	487	} else {
477	488	void *ret;
478	489
479		− ret = slob_new_page(gfp \| __GFP_COMP, get_order(size), node);
	490	+ ret = slob_new_pages(gfp \| __GFP_COMP, get_order(size), node);
480	491	if (ret) {
481	492	struct page *page;
482	493	page = virt_to_page(ret);
@@ -494,8 +505,8 @@ void kfree(const void *block) · 11 unmodified lines
494	505	if (unlikely(ZERO_OR_NULL_PTR(block)))
495	506	return;
496	507
497		− sp = (struct slob_page *)virt_to_page(block);
498		− if (slob_page(sp)) {
	508	+ sp = slob_page(block);
	509	+ if (is_slob_page(sp)) {
499	510	int align = max(ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
500	511	unsigned int m = (unsigned int )(block - align);
501	512	slob_free(m, *m + align);
@@ -513,8 +524,8 @@ size_t ksize(const void *block) · 11 unmodified lines
513	524	if (unlikely(block == ZERO_SIZE_PTR))
514	525	return 0;
515	526
516		− sp = (struct slob_page *)virt_to_page(block);
517		− if (slob_page(sp)) {
	527	+ sp = slob_page(block);
	528	+ if (is_slob_page(sp)) {
518	529	int align = max(ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
519	530	unsigned int m = (unsigned int )(block - align);
520	531	return SLOB_UNITS(m) SLOB_UNIT;
@@ -573,7 +584,7 @@ void kmem_cache_alloc_node(struct kmem_cache c, gfp_t flags, int node) · 52 unmodified lines
573	584	if (c->size < PAGE_SIZE)
574	585	b = slob_alloc(c->size, flags, c->align, node);
575	586	else
576		− b = slob_new_page(flags, get_order(c->size), node);
	587	+ b = slob_new_pages(flags, get_order(c->size), node);
577	588
578	589	if (c->ctor)
579	590	c->ctor(b);
@@ -587,7 +598,7 @@ static void __kmem_cache_free(void *b, int size) · 7 unmodified lines
587	598	if (size < PAGE_SIZE)
588	599	slob_free(b, size);
589	600	else
590		− free_pages((unsigned long)b, get_order(size));
	601	+ slob_free_pages(b, get_order(size));
591	602	}
592	603
593	604	static void kmem_rcu_free(struct rcu_head *head)

mm/slub.c+52−30

@@ -374,14 +374,8 @@ static struct track get_track(struct kmem_cache s, void *object, · 373 unmodified lines
374	374	static void set_track(struct kmem_cache s, void object,
375	375	enum track_item alloc, unsigned long addr)
376	376	{
377		− struct track *p;
378		−
379		− if (s->offset)
380		− p = object + s->offset + sizeof(void *);
381		− else
382		− p = object + s->inuse;
	377	+ struct track *p = get_track(s, object, alloc);
383	378
384		− p += alloc;
385	379	if (addr) {
386	380	p->addr = addr;
387	381	p->cpu = smp_processor_id();
@@ -1335,7 +1329,7 @@ static struct page get_any_partial(struct kmem_cache s, gfp_t flags) · 947 unmodified lines
1335	1329	n = get_node(s, zone_to_nid(zone));
1336	1330
1337	1331	if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1338		− n->nr_partial > n->min_partial) {
	1332	+ n->nr_partial > s->min_partial) {
1339	1333	page = get_partial_node(n);
1340	1334	if (page)
1341	1335	return page;
@@ -1387,7 +1381,7 @@ static void unfreeze_slab(struct kmem_cache s, struct page page, int tail) · 45 unmodified lines
1387	1381	slab_unlock(page);
1388	1382	} else {
1389	1383	stat(c, DEACTIVATE_EMPTY);
1390		− if (n->nr_partial < n->min_partial) {
	1384	+ if (n->nr_partial < s->min_partial) {
1391	1385	/*
1392	1386	* Adding an empty slab to the partial slabs in order
1393	1387	* to avoid page allocator overhead. This slab needs
@@ -1724,7 +1718,7 @@ static __always_inline void slab_free(struct kmem_cache *s, · 330 unmodified lines
1724	1718	c = get_cpu_slab(s, smp_processor_id());
1725	1719	debug_check_no_locks_freed(object, c->objsize);
1726	1720	if (!(s->flags & SLAB_DEBUG_OBJECTS))
1727		− debug_check_no_obj_freed(object, s->objsize);
	1721	+ debug_check_no_obj_freed(object, c->objsize);
1728	1722	if (likely(page == c->page && c->node >= 0)) {
1729	1723	object[c->offset] = c->freelist;
1730	1724	c->freelist = object;
@@ -1844,6 +1838,7 @@ static inline int calculate_order(int size) · 113 unmodified lines
1844	1838	int order;
1845	1839	int min_objects;
1846	1840	int fraction;
	1841	+ int max_objects;
1847	1842
1848	1843	/*
1849	1844	* Attempt to find best configuration for a slab. This
@@ -1856,6 +1851,9 @@ static inline int calculate_order(int size) · 6 unmodified lines
1856	1851	min_objects = slub_min_objects;
1857	1852	if (!min_objects)
1858	1853	min_objects = 4 * (fls(nr_cpu_ids) + 1);
	1854	+ max_objects = (PAGE_SIZE << slub_max_order)/size;
	1855	+ min_objects = min(min_objects, max_objects);
	1856	+
1859	1857	while (min_objects > 1) {
1860	1858	fraction = 16;
1861	1859	while (fraction >= 4) {
@@ -1865,7 +1863,7 @@ static inline int calculate_order(int size) · 3 unmodified lines
1865	1863	return order;
1866	1864	fraction /= 2;
1867	1865	}
1868		− min_objects /= 2;
	1866	+ min_objects --;
1869	1867	}
1870	1868
1871	1869	/*
@@ -1928,17 +1926,6 @@ static void · 56 unmodified lines
1928	1926	init_kmem_cache_node(struct kmem_cache_node n, struct kmem_cache s)
1929	1927	{
1930	1928	n->nr_partial = 0;
1931		−
1932		− /*
1933		− * The larger the object size is, the more pages we want on the partial
1934		− * list to avoid pounding the page allocator excessively.
1935		− */
1936		− n->min_partial = ilog2(s->size);
1937		− if (n->min_partial < MIN_PARTIAL)
1938		− n->min_partial = MIN_PARTIAL;
1939		− else if (n->min_partial > MAX_PARTIAL)
1940		− n->min_partial = MAX_PARTIAL;
1941		−
1942	1929	spin_lock_init(&n->list_lock);
1943	1930	INIT_LIST_HEAD(&n->partial);
1944	1931	#ifdef CONFIG_SLUB_DEBUG
@@ -2181,6 +2168,15 @@ static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags) · 236 unmodified lines
2181	2168	}
2182	2169	#endif
2183	2170
	2171	+static void set_min_partial(struct kmem_cache *s, unsigned long min)
	2172	+{
	2173	+ if (min < MIN_PARTIAL)
	2174	+ min = MIN_PARTIAL;
	2175	+ else if (min > MAX_PARTIAL)
	2176	+ min = MAX_PARTIAL;
	2177	+ s->min_partial = min;
	2178	+}
	2179	+
2184	2180	/*
2185	2181	* calculate_sizes() determines the order and the distribution of data within
2186	2182	* a slab object.
@@ -2319,6 +2315,11 @@ static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags, · 132 unmodified lines
2319	2315	if (!calculate_sizes(s, -1))
2320	2316	goto error;
2321	2317
	2318	+ /*
	2319	+ * The larger the object size is, the more pages we want on the partial
	2320	+ * list to avoid pounding the page allocator excessively.
	2321	+ */
	2322	+ set_min_partial(s, ilog2(s->size));
2322	2323	s->refcount = 1;
2323	2324	#ifdef CONFIG_NUMA
2324	2325	s->remote_node_defrag_ratio = 1000;
@@ -2475,7 +2476,7 @@ EXPORT_SYMBOL(kmem_cache_destroy); · 150 unmodified lines
2475	2476	* Kmalloc subsystem
2476	2477	*******************************************************************/
2477	2478
2478		−struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned;
	2479	+struct kmem_cache kmalloc_caches[SLUB_PAGE_SHIFT] __cacheline_aligned;
2479	2480	EXPORT_SYMBOL(kmalloc_caches);
2480	2481
2481	2482	static int __init setup_slub_min_order(char *str)
@@ -2537,7 +2538,7 @@ panic: · 55 unmodified lines
2537	2538	}
2538	2539
2539	2540	#ifdef CONFIG_ZONE_DMA
2540		−static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1];
	2541	+static struct kmem_cache *kmalloc_caches_dma[SLUB_PAGE_SHIFT];
2541	2542
2542	2543	static void sysfs_add_func(struct work_struct *w)
2543	2544	{
@@ -2658,7 +2659,7 @@ void *__kmalloc(size_t size, gfp_t flags) · 114 unmodified lines
2658	2659	{
2659	2660	struct kmem_cache *s;
2660	2661
2661		− if (unlikely(size > PAGE_SIZE))
	2662	+ if (unlikely(size > SLUB_MAX_SIZE))
2662	2663	return kmalloc_large(size, flags);
2663	2664
2664	2665	s = get_slab(size, flags);
@@ -2686,7 +2687,7 @@ void *__kmalloc_node(size_t size, gfp_t flags, int node) · 21 unmodified lines
2686	2687	{
2687	2688	struct kmem_cache *s;
2688	2689
2689		− if (unlikely(size > PAGE_SIZE))
	2690	+ if (unlikely(size > SLUB_MAX_SIZE))
2690	2691	return kmalloc_large_node(size, flags, node);
2691	2692
2692	2693	s = get_slab(size, flags);
@@ -2986,7 +2987,7 @@ void __init kmem_cache_init(void) · 293 unmodified lines
2986	2987	caches++;
2987	2988	}
2988	2989
2989		− for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) {
	2990	+ for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
2990	2991	create_kmalloc_cache(&kmalloc_caches[i],
2991	2992	"kmalloc", 1 << i, GFP_KERNEL);
2992	2993	caches++;
@@ -3023,7 +3024,7 @@ void __init kmem_cache_init(void) · 30 unmodified lines
3023	3024	slab_state = UP;
3024	3025
3025	3026	/* Provide the correct kmalloc names now that the caches are up */
3026		− for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++)
	3027	+ for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++)
3027	3028	kmalloc_caches[i]. name =
3028	3029	kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
3029	3030
@@ -3223,7 +3224,7 @@ void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller) · 193 unmodified lines
3223	3224	{
3224	3225	struct kmem_cache *s;
3225	3226
3226		− if (unlikely(size > PAGE_SIZE))
	3227	+ if (unlikely(size > SLUB_MAX_SIZE))
3227	3228	return kmalloc_large(size, gfpflags);
3228	3229
3229	3230	s = get_slab(size, gfpflags);
@@ -3239,7 +3240,7 @@ void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, · 9 unmodified lines
3239	3240	{
3240	3241	struct kmem_cache *s;
3241	3242
3242		− if (unlikely(size > PAGE_SIZE))
	3243	+ if (unlikely(size > SLUB_MAX_SIZE))
3243	3244	return kmalloc_large_node(size, gfpflags, node);
3244	3245
3245	3246	s = get_slab(size, gfpflags);
@@ -3836,6 +3837,26 @@ static ssize_t order_show(struct kmem_cache s, char buf) · 590 unmodified lines
3836	3837	}
3837	3838	SLAB_ATTR(order);
3838	3839
	3840	+static ssize_t min_partial_show(struct kmem_cache s, char buf)
	3841	+{
	3842	+ return sprintf(buf, "%lu\n", s->min_partial);
	3843	+}
	3844	+
	3845	+static ssize_t min_partial_store(struct kmem_cache s, const char buf,
	3846	+ size_t length)
	3847	+{
	3848	+ unsigned long min;
	3849	+ int err;
	3850	+
	3851	+ err = strict_strtoul(buf, 10, &min);
	3852	+ if (err)
	3853	+ return err;
	3854	+
	3855	+ set_min_partial(s, min);
	3856	+ return length;
	3857	+}
	3858	+SLAB_ATTR(min_partial);
	3859	+
3839	3860	static ssize_t ctor_show(struct kmem_cache s, char buf)
3840	3861	{
3841	3862	if (s->ctor) {
@@ -4151,6 +4172,7 @@ static struct attribute *slab_attrs[] = { · 309 unmodified lines
4151	4172	&object_size_attr.attr,
4152	4173	&objs_per_slab_attr.attr,
4153	4174	&order_attr.attr,
	4175	+ &min_partial_attr.attr,
4154	4176	&objects_attr.attr,
4155	4177	&objects_partial_attr.attr,
4156	4178	&total_objects_attr.attr,

Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/penberg/slab-2.6

Keyboard shortcuts

General

Issues and pull requests

Issues, pull requests and Inbox

Repository code