Donate to e Foundation | Murena handsets with /e/OS | Own a part of Murena! Learn more

Commit d37cacc5 authored by Matthew Wilcox's avatar Matthew Wilcox
Browse files

ida: Use exceptional entries for small IDAs



We can use the root entry as a bitmap and save allocating a 128 byte
bitmap for an IDA that contains only a few entries (30 on a 32-bit
machine, 62 on a 64-bit machine).  This costs about 300 bytes of kernel
text on x86-64, so as long as 3 IDAs fall into this category, this
is a net win for memory consumption.

Thanks to Rasmus Villemoes for his work documenting the problem and
collecting statistics on IDAs.

Signed-off-by: default avatarMatthew Wilcox <mawilcox@microsoft.com>
parent 7ad3d4d8
Loading
Loading
Loading
Loading
+81 −6
Original line number Original line Diff line number Diff line
@@ -194,6 +194,39 @@ EXPORT_SYMBOL(idr_replace);
 * limitation, it should be quite straightforward to raise the maximum.
 * limitation, it should be quite straightforward to raise the maximum.
 */
 */


/*
 * Developer's notes:
 *
 * The IDA uses the functionality provided by the IDR & radix tree to store
 * bitmaps in each entry.  The IDR_FREE tag means there is at least one bit
 * free, unlike the IDR where it means at least one entry is free.
 *
 * I considered telling the radix tree that each slot is an order-10 node
 * and storing the bit numbers in the radix tree, but the radix tree can't
 * allow a single multiorder entry at index 0, which would significantly
 * increase memory consumption for the IDA.  So instead we divide the index
 * by the number of bits in the leaf bitmap before doing a radix tree lookup.
 *
 * As an optimisation, if there are only a few low bits set in any given
 * leaf, instead of allocating a 128-byte bitmap, we use the 'exceptional
 * entry' functionality of the radix tree to store BITS_PER_LONG - 2 bits
 * directly in the entry.  By being really tricksy, we could store
 * BITS_PER_LONG - 1 bits, but there're diminishing returns after optimising
 * for 0-3 allocated IDs.
 *
 * We allow the radix tree 'exceptional' count to get out of date.  Nothing
 * in the IDA nor the radix tree code checks it.  If it becomes important
 * to maintain an accurate exceptional count, switch the rcu_assign_pointer()
 * calls to radix_tree_iter_replace() which will correct the exceptional
 * count.
 *
 * The IDA always requires a lock to alloc/free.  If we add a 'test_bit'
 * equivalent, it will still need locking.  Going to RCU lookup would require
 * using RCU to free bitmaps, and that's not trivial without embedding an
 * RCU head in the bitmap, which adds a 2-pointer overhead to each 128-byte
 * bitmap, which is excessive.
 */

#define IDA_MAX (0x80000000U / IDA_BITMAP_BITS)
#define IDA_MAX (0x80000000U / IDA_BITMAP_BITS)


/**
/**
@@ -221,11 +254,12 @@ int ida_get_new_above(struct ida *ida, int start, int *id)
	struct radix_tree_iter iter;
	struct radix_tree_iter iter;
	struct ida_bitmap *bitmap;
	struct ida_bitmap *bitmap;
	unsigned long index;
	unsigned long index;
	unsigned bit;
	unsigned bit, ebit;
	int new;
	int new;


	index = start / IDA_BITMAP_BITS;
	index = start / IDA_BITMAP_BITS;
	bit = start % IDA_BITMAP_BITS;
	bit = start % IDA_BITMAP_BITS;
	ebit = bit + RADIX_TREE_EXCEPTIONAL_SHIFT;


	slot = radix_tree_iter_init(&iter, index);
	slot = radix_tree_iter_init(&iter, index);
	for (;;) {
	for (;;) {
@@ -240,10 +274,29 @@ int ida_get_new_above(struct ida *ida, int start, int *id)
				return PTR_ERR(slot);
				return PTR_ERR(slot);
			}
			}
		}
		}
		if (iter.index > index)
		if (iter.index > index) {
			bit = 0;
			bit = 0;
			ebit = RADIX_TREE_EXCEPTIONAL_SHIFT;
		}
		new = iter.index * IDA_BITMAP_BITS;
		new = iter.index * IDA_BITMAP_BITS;
		bitmap = rcu_dereference_raw(*slot);
		bitmap = rcu_dereference_raw(*slot);
		if (radix_tree_exception(bitmap)) {
			unsigned long tmp = (unsigned long)bitmap;
			ebit = find_next_zero_bit(&tmp, BITS_PER_LONG, ebit);
			if (ebit < BITS_PER_LONG) {
				tmp |= 1UL << ebit;
				rcu_assign_pointer(*slot, (void *)tmp);
				*id = new + ebit - RADIX_TREE_EXCEPTIONAL_SHIFT;
				return 0;
			}
			bitmap = this_cpu_xchg(ida_bitmap, NULL);
			if (!bitmap)
				return -EAGAIN;
			memset(bitmap, 0, sizeof(*bitmap));
			bitmap->bitmap[0] = tmp >> RADIX_TREE_EXCEPTIONAL_SHIFT;
			rcu_assign_pointer(*slot, bitmap);
		}

		if (bitmap) {
		if (bitmap) {
			bit = find_next_zero_bit(bitmap->bitmap,
			bit = find_next_zero_bit(bitmap->bitmap,
							IDA_BITMAP_BITS, bit);
							IDA_BITMAP_BITS, bit);
@@ -261,6 +314,14 @@ int ida_get_new_above(struct ida *ida, int start, int *id)
			new += bit;
			new += bit;
			if (new < 0)
			if (new < 0)
				return -ENOSPC;
				return -ENOSPC;
			if (ebit < BITS_PER_LONG) {
				bitmap = (void *)((1UL << ebit) |
						RADIX_TREE_EXCEPTIONAL_ENTRY);
				radix_tree_iter_replace(root, &iter, slot,
						bitmap);
				*id = new;
				return 0;
			}
			bitmap = this_cpu_xchg(ida_bitmap, NULL);
			bitmap = this_cpu_xchg(ida_bitmap, NULL);
			if (!bitmap)
			if (!bitmap)
				return -EAGAIN;
				return -EAGAIN;
@@ -287,6 +348,7 @@ void ida_remove(struct ida *ida, int id)
	unsigned long index = id / IDA_BITMAP_BITS;
	unsigned long index = id / IDA_BITMAP_BITS;
	unsigned offset = id % IDA_BITMAP_BITS;
	unsigned offset = id % IDA_BITMAP_BITS;
	struct ida_bitmap *bitmap;
	struct ida_bitmap *bitmap;
	unsigned long *btmp;
	struct radix_tree_iter iter;
	struct radix_tree_iter iter;
	void **slot;
	void **slot;


@@ -295,12 +357,24 @@ void ida_remove(struct ida *ida, int id)
		goto err;
		goto err;


	bitmap = rcu_dereference_raw(*slot);
	bitmap = rcu_dereference_raw(*slot);
	if (!test_bit(offset, bitmap->bitmap))
	if (radix_tree_exception(bitmap)) {
		btmp = (unsigned long *)slot;
		offset += RADIX_TREE_EXCEPTIONAL_SHIFT;
		if (offset >= BITS_PER_LONG)
			goto err;
	} else {
		btmp = bitmap->bitmap;
	}
	if (!test_bit(offset, btmp))
		goto err;
		goto err;


	__clear_bit(offset, bitmap->bitmap);
	__clear_bit(offset, btmp);
	radix_tree_iter_tag_set(&ida->ida_rt, &iter, IDR_FREE);
	radix_tree_iter_tag_set(&ida->ida_rt, &iter, IDR_FREE);
	if (bitmap_empty(bitmap->bitmap, IDA_BITMAP_BITS)) {
	if (radix_tree_exception(bitmap)) {
		if (rcu_dereference_raw(*slot) ==
					(void *)RADIX_TREE_EXCEPTIONAL_ENTRY)
			radix_tree_iter_delete(&ida->ida_rt, &iter, slot);
	} else if (bitmap_empty(btmp, IDA_BITMAP_BITS)) {
		kfree(bitmap);
		kfree(bitmap);
		radix_tree_iter_delete(&ida->ida_rt, &iter, slot);
		radix_tree_iter_delete(&ida->ida_rt, &iter, slot);
	}
	}
@@ -326,6 +400,7 @@ void ida_destroy(struct ida *ida)


	radix_tree_for_each_slot(slot, &ida->ida_rt, &iter, 0) {
	radix_tree_for_each_slot(slot, &ida->ida_rt, &iter, 0) {
		struct ida_bitmap *bitmap = rcu_dereference_raw(*slot);
		struct ida_bitmap *bitmap = rcu_dereference_raw(*slot);
		if (!radix_tree_exception(bitmap))
			kfree(bitmap);
			kfree(bitmap);
		radix_tree_iter_delete(&ida->ida_rt, &iter, slot);
		radix_tree_iter_delete(&ida->ida_rt, &iter, slot);
	}
	}
+8 −0
Original line number Original line Diff line number Diff line
@@ -338,6 +338,14 @@ static void dump_ida_node(void *entry, unsigned long index)
		for (i = 0; i < RADIX_TREE_MAP_SIZE; i++)
		for (i = 0; i < RADIX_TREE_MAP_SIZE; i++)
			dump_ida_node(node->slots[i],
			dump_ida_node(node->slots[i],
					index | (i << node->shift));
					index | (i << node->shift));
	} else if (radix_tree_exceptional_entry(entry)) {
		pr_debug("ida excp: %p offset %d indices %lu-%lu data %lx\n",
				entry, (int)(index & RADIX_TREE_MAP_MASK),
				index * IDA_BITMAP_BITS,
				index * IDA_BITMAP_BITS + BITS_PER_LONG -
					RADIX_TREE_EXCEPTIONAL_SHIFT,
				(unsigned long)entry >>
					RADIX_TREE_EXCEPTIONAL_SHIFT);
	} else {
	} else {
		struct ida_bitmap *bitmap = entry;
		struct ida_bitmap *bitmap = entry;


+92 −1
Original line number Original line Diff line number Diff line
@@ -11,6 +11,7 @@
 * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
 * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
 * more details.
 * more details.
 */
 */
#include <linux/bitmap.h>
#include <linux/idr.h>
#include <linux/idr.h>
#include <linux/slab.h>
#include <linux/slab.h>
#include <linux/kernel.h>
#include <linux/kernel.h>
@@ -214,7 +215,7 @@ void ida_check_nomem(void)
	DEFINE_IDA(ida);
	DEFINE_IDA(ida);
	int id, err;
	int id, err;


	err = ida_get_new(&ida, &id);
	err = ida_get_new_above(&ida, 256, &id);
	assert(err == -EAGAIN);
	assert(err == -EAGAIN);
	err = ida_get_new_above(&ida, 1UL << 30, &id);
	err = ida_get_new_above(&ida, 1UL << 30, &id);
	assert(err == -EAGAIN);
	assert(err == -EAGAIN);
@@ -246,6 +247,66 @@ void ida_check_leaf(void)
	assert(ida_is_empty(&ida));
	assert(ida_is_empty(&ida));
}
}


/*
 * Check handling of conversions between exceptional entries and full bitmaps.
 */
void ida_check_conv(void)
{
	DEFINE_IDA(ida);
	int id;
	unsigned long i;

	for (i = 0; i < IDA_BITMAP_BITS * 2; i += IDA_BITMAP_BITS) {
		assert(ida_pre_get(&ida, GFP_KERNEL));
		assert(!ida_get_new_above(&ida, i + 1, &id));
		assert(id == i + 1);
		assert(!ida_get_new_above(&ida, i + BITS_PER_LONG, &id));
		assert(id == i + BITS_PER_LONG);
		ida_remove(&ida, i + 1);
		ida_remove(&ida, i + BITS_PER_LONG);
		assert(ida_is_empty(&ida));
	}

	assert(ida_pre_get(&ida, GFP_KERNEL));

	for (i = 0; i < IDA_BITMAP_BITS * 2; i++) {
		assert(ida_pre_get(&ida, GFP_KERNEL));
		assert(!ida_get_new(&ida, &id));
		assert(id == i);
	}

	for (i = IDA_BITMAP_BITS * 2; i > 0; i--) {
		ida_remove(&ida, i - 1);
	}
	assert(ida_is_empty(&ida));

	for (i = 0; i < IDA_BITMAP_BITS + BITS_PER_LONG - 4; i++) {
		assert(ida_pre_get(&ida, GFP_KERNEL));
		assert(!ida_get_new(&ida, &id));
		assert(id == i);
	}

	for (i = IDA_BITMAP_BITS + BITS_PER_LONG - 4; i > 0; i--) {
		ida_remove(&ida, i - 1);
	}
	assert(ida_is_empty(&ida));

	radix_tree_cpu_dead(1);
	for (i = 0; i < 1000000; i++) {
		int err = ida_get_new(&ida, &id);
		if (err == -EAGAIN) {
			assert((i % IDA_BITMAP_BITS) == (BITS_PER_LONG - 2));
			assert(ida_pre_get(&ida, GFP_KERNEL));
			err = ida_get_new(&ida, &id);
		} else {
			assert((i % IDA_BITMAP_BITS) != (BITS_PER_LONG - 2));
		}
		assert(!err);
		assert(id == i);
	}
	ida_destroy(&ida);
}

/*
/*
 * Check allocations up to and slightly above the maximum allowed (2^31-1) ID.
 * Check allocations up to and slightly above the maximum allowed (2^31-1) ID.
 * Allocating up to 2^31-1 should succeed, and then allocating the next one
 * Allocating up to 2^31-1 should succeed, and then allocating the next one
@@ -273,6 +334,34 @@ void ida_check_max(void)
	}
	}
}
}


void ida_check_random(void)
{
	DEFINE_IDA(ida);
	DECLARE_BITMAP(bitmap, 2048);
	int id;
	unsigned int i;
	time_t s = time(NULL);

 repeat:
	memset(bitmap, 0, sizeof(bitmap));
	for (i = 0; i < 100000; i++) {
		int i = rand();
		int bit = i & 2047;
		if (test_bit(bit, bitmap)) {
			__clear_bit(bit, bitmap);
			ida_remove(&ida, bit);
		} else {
			__set_bit(bit, bitmap);
			ida_pre_get(&ida, GFP_KERNEL);
			assert(!ida_get_new_above(&ida, bit, &id));
			assert(id == bit);
		}
	}
	ida_destroy(&ida);
	if (time(NULL) < s + 10)
		goto repeat;
}

void ida_checks(void)
void ida_checks(void)
{
{
	DEFINE_IDA(ida);
	DEFINE_IDA(ida);
@@ -337,6 +426,8 @@ void ida_checks(void)


	ida_check_leaf();
	ida_check_leaf();
	ida_check_max();
	ida_check_max();
	ida_check_conv();
	ida_check_random();


	radix_tree_cpu_dead(1);
	radix_tree_cpu_dead(1);
}
}