nta.3

实现linux平台下零拷贝技术的软件包。

+	for (i=0; i<AVL_NODE_NUM; ++i) {
+		struct avl_node *node;
+
+		ptr = __get_free_pages(gfp_mask | __GFP_ZERO, order);
+		if (!ptr)
+			break;
+
+		node = &entry->avl_node_array[idx][off];
+
+		if (++off >= AVL_NODES_ON_PAGE) {
+			idx++;
+			off = 0;
+		}
+
+		for (j=0; j<(1<<order); ++j)
+			get_page(virt_to_page(ptr + (j<<PAGE_SHIFT)));
+
+		node->value = ptr;
+		memset(node->mask, 0, sizeof(node->mask));
+		avl_fill_bits(node->mask, ARRAY_SIZE(node->mask), 0, ((1<<order)<<PAGE_SHIFT)/AVL_MIN_SIZE, 1);
+	}
+
+	ulog("%s: entry: %p, node: %u, node_pages: %lu, node_num: %lu, order: %d, allocated: %d, container: %u, max_size: %u, min_size: %u, bits: %u.\n", 
+		__func__, entry, sizeof(struct avl_node), AVL_NODE_PAGES, AVL_NODE_NUM, order, 
+		i, AVL_CONTAINER_ARRAY_SIZE, AVL_MAX_SIZE, AVL_MIN_SIZE, ((1<<order)<<PAGE_SHIFT)/AVL_MIN_SIZE);
+
+	if (i == 0)
+		goto err_out_free;
+
+	entry->avl_node_num = i;
+	entry->avl_node_order = order;
+
+	return entry;
+
+err_out_free:
+	for (i=0; i<AVL_NODE_PAGES; ++i)
+		free_page((unsigned long)entry->avl_node_array[i]);
+err_out_free_entry:
+	kfree(entry);
+	return NULL;
+}
+
+/*
+ * Allocate memory region with given size and mode.
+ * If allocation fails due to unsupported order, otherwise
+ * allocate new node entry with given mode and try to allocate again
+ * Cache growing happens only with 0-order allocations.
+ */
+void *avl_alloc(unsigned int size, gfp_t gfp_mask)
+{
+	unsigned int i, try = 0, osize = size;
+	void *ptr = NULL;
+	unsigned long flags;
+
+	size = AVL_ALIGN(size + sizeof(struct avl_chunk));
+
+	if (size > AVL_MAX_SIZE || size < AVL_MIN_SIZE) {
+		/*
+		 * Print info about unsupported order so user could send a "bug report"
+		 * or increase initial allocation order.
+		 */
+		if (get_order(size) > AVL_ORDER && net_ratelimit()) {
+			printk(KERN_INFO "%s: Failed to allocate %u bytes with %02x mode, order %u, max order %u.\n", 
+					__func__, size, gfp_mask, get_order(size), AVL_ORDER);
+			WARN_ON(1);
+		}
+
+		return NULL;
+	}
+
+	local_irq_save(flags);
+repeat:
+	for (i=size/AVL_MIN_SIZE-1; i<AVL_CONTAINER_ARRAY_SIZE; ++i) {
+		struct list_head *head = &avl_allocator[smp_processor_id()].avl_container_array[i];
+		struct avl_container *c;
+
+		if (!list_empty(head)) {
+			struct avl_chunk *ch;
+
+			c = avl_dequeue(head);
+			ptr = c->ptr;
+
+			ch = avl_ptr_to_chunk(ptr, osize);
+			atomic_set(&ch->refcnt, 1);
+			ch->canary = AVL_CANARY;
+
+			avl_update_node(c, i, osize);
+			break;
+		}
+	}
+	local_irq_restore(flags);
+#if 1
+	if (!ptr && !try) {
+		struct avl_node_entry *entry;
+		
+		try = 1;
+
+		entry = avl_node_entry_alloc(gfp_mask, get_order(size));
+		if (entry) {
+			local_irq_save(flags);
+			avl_node_entry_commit(entry, smp_processor_id());
+			goto repeat;
+		}
+			
+	}
+#endif
+	if (unlikely(!ptr && try))
+		if (net_ratelimit())
+			printk("%s: Failed to allocate %u bytes.\n", __func__, size);
+
+
+
+	return ptr;
+}
+
+/*
+ * Remove free chunk from the list.
+ */
+static inline struct avl_container *avl_search_container(void *ptr, unsigned int idx, int cpu)
+{
+	struct avl_container *c = ptr;
+	
+	list_del(&c->centry);
+	c->ptr = ptr;
+
+	return c;
+}
+
+/*
+ * Combine neighbour free chunks into the one with bigger size
+ * and put new chunk into list of free chunks with appropriate size.
+ */
+static void avl_combine(struct avl_node *node, void *lp, unsigned int lbits, void *rp, unsigned int rbits, 
+		void *cur_ptr, unsigned int cur_bits, int cpu)
+{
+	struct avl_container *lc, *rc, *c;
+	unsigned int idx;
+	void *ptr;
+
+	lc = rc = c = NULL;
+	idx = cur_bits - 1;
+	ptr = cur_ptr;
+
+	c = (struct avl_container *)cur_ptr;
+	c->ptr = cur_ptr;
+	
+	if (rp) {
+		rc = avl_search_container(rp, rbits-1, cpu);
+		if (!rc) {
+			printk(KERN_ERR "%p.%p: Failed to find a container for right pointer %p, rbits: %u.\n", 
+					node, cur_ptr, rp, rbits);
+			BUG();
+		}
+
+		c = rc;
+		idx += rbits;
+		ptr = c->ptr;
+	}
+
+	if (lp) {
+		lc = avl_search_container(lp, lbits-1, cpu);
+		if (!lc) {
+			printk(KERN_ERR "%p.%p: Failed to find a container for left pointer %p, lbits: %u.\n", 
+					node, cur_ptr, lp, lbits);
+			BUG();
+		}
+
+		idx += lbits;
+		ptr = c->ptr;
+	}
+	avl_container_insert(c, idx, cpu);
+}
+
+/*
+ * Free memory region of given size.
+ * Must be called on the same CPU where allocation happend
+ * with disabled interrupts.
+ */
+static void __avl_free_local(void *ptr, unsigned int size)
+{
+	unsigned long val = avl_ptr_to_value(ptr);
+	unsigned int pos, idx, sbits = AVL_ALIGN(size)/AVL_MIN_SIZE;
+	unsigned int rbits, lbits, cpu = avl_get_cpu_ptr(val);
+	struct avl_node *node;
+	unsigned long p;
+	void *lp, *rp;
+
+	node = avl_get_node_ptr((unsigned long)ptr);
+
+	pos = avl_ptr_to_offset(ptr);
+	idx = pos/BITS_PER_LONG;
+
+	p = node->mask[idx] >> (pos%BITS_PER_LONG);
+	
+	if ((p & 1)) {
+		if (net_ratelimit())
+			printk(KERN_ERR "%p.%p: Broken pointer: value: %lx, pos: %u, idx: %u, mask: %lx, p: %lx.\n", 
+				node, ptr, val, pos, idx, node->mask[idx], p);
+		return;
+	}
+
+	avl_fill_bits(node->mask, ARRAY_SIZE(node->mask), pos, sbits, 1);
+
+	lp = rp = NULL;
+	rbits = lbits = 0;
+
+	idx = (pos+sbits)/BITS_PER_LONG;
+	p = (pos+sbits)%BITS_PER_LONG;
+
+	if ((node->mask[idx] >> p) & 1) {
+		lbits = avl_count_set_up(node->mask, ARRAY_SIZE(node->mask), pos+sbits);
+		if (lbits) {
+			lp = (void *)(val + (pos + sbits)*AVL_MIN_SIZE);
+		}
+	}
+
+	if (pos) {
+		idx = (pos-1)/BITS_PER_LONG;
+		p = (pos-1)%BITS_PER_LONG;
+		if ((node->mask[idx] >> p) & 1) {
+			rbits = avl_count_set_down(node->mask, pos-1);
+			if (rbits) {
+				rp = (void *)(val + (pos-rbits)*AVL_MIN_SIZE);
+			}
+		}
+	}
+
+	avl_combine(node, lp, lbits, rp, rbits, ptr, sbits, cpu);
+}
+
+/*
+ * Free memory region of given size.
+ * If freeing CPU is not the same as allocation one, chunk will 
+ * be placed into list of to-be-freed objects on allocation CPU,
+ * otherwise chunk will be freed and combined with neighbours.
+ * Must be called with disabled interrupts.
+ */
+static void __avl_free(void *ptr, unsigned int size)
+{
+	int cpu = avl_get_cpu_ptr((unsigned long)ptr);
+
+	if (cpu != smp_processor_id()) {
+		struct avl_free_list *l, *this = ptr;
+		struct avl_allocator_data *alloc = &avl_allocator[cpu];
+
+		this->cpu = smp_processor_id();
+		this->size = size;
+
+		spin_lock(&alloc->avl_free_lock);
+		l = alloc->avl_free_list_head;
+		alloc->avl_free_list_head = this;
+		this->next = l;
+		spin_unlock(&alloc->avl_free_lock);
+		return;
+	}
+
+	__avl_free_local(ptr, size);
+}
+
+/*
+ * Free memory region of given size.
+ */
+void avl_free(void *ptr, unsigned int size)
+{
+	unsigned long flags;
+	struct avl_free_list *l;
+	struct avl_allocator_data *alloc;
+	struct avl_chunk *ch = avl_ptr_to_chunk(ptr, size);
+
+	if (ch->canary != AVL_CANARY) {
+		printk("Freeing destroyed object: ptr: %p, size: %u, canary: %x, must be %x, refcnt: %d.\n",
+				ptr, size, ch->canary, AVL_CANARY, atomic_read(&ch->refcnt));
+		return;
+	}
+
+	if (atomic_dec_and_test(&ch->refcnt)) {
+		local_irq_save(flags);
+		__avl_free(ptr, size);
+		
+		alloc = &avl_allocator[smp_processor_id()];
+
+		while (alloc->avl_free_list_head) {
+			spin_lock(&alloc->avl_free_lock);
+			l = alloc->avl_free_list_head;
+			alloc->avl_free_list_head = l->next;
+			spin_unlock(&alloc->avl_free_lock);
+			__avl_free_local(l, l->size);
+		};
+		local_irq_restore(flags);
+	}
+}
+
+/*
+ * Initialize per-cpu allocator data.
+ */
+static int avl_init_cpu(int cpu)
+{
+	unsigned int i;
+	struct avl_allocator_data *alloc = &avl_allocator[cpu];
+	struct avl_node_entry *entry;
+
+	spin_lock_init(&alloc->avl_free_lock);
+	spin_lock_init(&alloc->avl_node_lock);
+	INIT_LIST_HEAD(&alloc->avl_node_list);
+
+	alloc->avl_container_array = kzalloc(sizeof(struct list_head) * AVL_CONTAINER_ARRAY_SIZE, GFP_KERNEL);
+	if (!alloc->avl_container_array)
+		goto err_out_exit;
+
+	for (i=0; i<AVL_CONTAINER_ARRAY_SIZE; ++i)
+		INIT_LIST_HEAD(&alloc->avl_container_array[i]);
+
+	entry = avl_node_entry_alloc(GFP_KERNEL, AVL_ORDER);
+	if (!entry)
+		goto err_out_free_container;
+
+	avl_node_entry_commit(entry, cpu);
+
+	return 0;
+
+err_out_free_container:
+	kfree(alloc->avl_container_array);
+err_out_exit:
+	return -ENOMEM;
+}
+
+/*
+ * Initialize network allocator.
+ */
+int avl_init(void)
+{
+	int err, cpu;
+
+	for_each_possible_cpu(cpu) {
+		err = avl_init_cpu(cpu);
+		if (err)
+			goto err_out;
+	}
+
+	err = avl_init_zc();
+
+	printk(KERN_INFO "Network tree allocator has been initialized.\n");
+	return 0;
+
+err_out:
+	panic("Failed to initialize network allocator.\n");
+
+	return -ENOMEM;
+}
diff --git a/net/core/alloc/avl.h b/net/core/alloc/avl.h
new file mode 100644
index 0000000..61982b5
--- /dev/null
+++ b/net/core/alloc/avl.h
@@ -0,0 +1,178 @@
+/*
+ * 	avl.h
+ * 
+ * 2006 Copyright (c) Evgeniy Polyakov <[e-mail]>
+ * All rights reserved.
+ * 
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHAAVLBILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+ */
+
+#ifndef __AVL_H
+#define __AVL_H
+
+struct zc_data
+{
+	union {
+		__u32		data[2];
+		void		*ptr;
+	} data;
+	__u32			off;
+	__u32			size;
+	__u32			entry;
+	__u16			number, order;
+};
+
+#ifdef __KERNEL__
+#include <linux/kernel.h>
+#include <linux/types.h>
+#include <linux/wait.h>
+#include <linux/spinlock.h>
+#include <asm/page.h>
+
+//#define AVL_DEBUG
+
+#ifdef AVL_DEBUG
+#define ulog(f, a...) printk(f, ##a)
+#else
+#define ulog(f, a...)
+#endif
+
+/*
+ * Network tree allocator variables.
+ */
+
+#define AVL_CANARY		0xc0d0e0f0
+
+#define AVL_ALIGN_SIZE		L1_CACHE_BYTES
+#define AVL_ALIGN(x) 		ALIGN(x, AVL_ALIGN_SIZE)
+
+#define AVL_ORDER		3	/* Maximum allocation order */
+#define AVL_BITS		3	/* Must cover maximum number of pages used for allocation pools */
+
+#define AVL_NODES_ON_PAGE	(PAGE_SIZE/sizeof(struct avl_node))
+#define AVL_NODE_NUM		(1UL<<AVL_BITS)
+#define AVL_NODE_PAGES		((AVL_NODE_NUM+AVL_NODES_ON_PAGE-1)/AVL_NODES_ON_PAGE)
+
+#define AVL_MIN_SIZE		AVL_ALIGN_SIZE
+#define AVL_MAX_SIZE		((1<<AVL_ORDER) << PAGE_SHIFT)