[PATCH v3 0/2] bpf: add longest prefix match map

[PATCH v3 0/2] bpf: add longest prefix match map

[Daniel Mack]

This patch set adds a longest prefix match algorithm that can be used
to match IP addresses to a stored set of ranges. It is exposed as a
bpf map type.
   
Internally, data is stored in an unbalanced tree of nodes that has a
maximum height of n, where n is the prefixlen the trie was created
with.
 
Note that this has nothing to do with fib or fib6 and is in no way meant
to replace or share code with it. It's rather a much simpler
implementation that is specifically written with bpf maps in mind.
 
Patch 1/2 adds the implementation, and 2/2 an extensive test suite.

Feedback is much appreciated.
 
 
Thanks,
Daniel

Changelog:

v2 -> v3:
	* Store both the key match data and the caller provided
	  value in the same byte array attached to a node. This
	  avoids double allocations
	* Bring back node->flags to distinguish between 'real'
	  and intermediate nodes
	* Fix comment style and some typos

v1 -> v2:
	* Turn spin lock into raw spinlock
	* Lock with irqsave options during trie_update_elem()
	* Return -ENOMEM properly from trie_alloc()
	* Force attr->flags == BPF_F_NO_PREALLOC during creation
	* Set trie->map.pages after creation to account for map memory
	* Allow arbitrary value sizes
	* Removed node->flags and denode intermediate nodes through
	  node->value == NULL instead

rfc -> v1:
	* Add __rcu pointer annotations to make sparse happy
	* Fold _lpm_trie_find_target_node() into its only caller
	* Fix some minor documentation issues


Daniel Mack (1):
  bpf: add a longest prefix match trie map implementation

David Herrmann (1):
  bpf: Add tests for the lpm trie map

 include/uapi/linux/bpf.h                   |   7 +
 kernel/bpf/Makefile                        |   2 +-
 kernel/bpf/lpm_trie.c                      | 493 +++++++++++++++++++++++++++++
 tools/testing/selftests/bpf/.gitignore     |   1 +
 tools/testing/selftests/bpf/Makefile       |   4 +-
 tools/testing/selftests/bpf/test_lpm_map.c | 358 +++++++++++++++++++++
 6 files changed, 862 insertions(+), 3 deletions(-)
 create mode 100644 kernel/bpf/lpm_trie.c
 create mode 100644 tools/testing/selftests/bpf/test_lpm_map.c

-- 
2.9.3

[PATCH v3 1/2] bpf: add a longest prefix match trie map implementation

[Daniel Mack]

This trie implements a longest prefix match algorithm that can be used
to match IP addresses to a stored set of ranges.

Internally, data is stored in an unbalanced trie of nodes that has a
maximum height of n, where n is the prefixlen the trie was created
with.

Tries may be created with prefix lengths that are multiples of 8, in
the range from 8 to 2048. The key used for lookup and update operations
is a struct bpf_lpm_trie_key, and the value is a uint64_t.

The code carries more information about the internal implementation.

Signed-off-by: Daniel Mack <daniel@zonque.org>
Reviewed-by: David Herrmann <dh.herrmann@gmail.com>
---
 include/uapi/linux/bpf.h |   7 +
 kernel/bpf/Makefile      |   2 +-
 kernel/bpf/lpm_trie.c    | 493 +++++++++++++++++++++++++++++++++++++++++++++++
 3 files changed, 501 insertions(+), 1 deletion(-)
 create mode 100644 kernel/bpf/lpm_trie.c

diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
index 0eb0e87..d564277 100644
--- a/include/uapi/linux/bpf.h
+++ b/include/uapi/linux/bpf.h
@@ -63,6 +63,12 @@ struct bpf_insn {
 	__s32	imm;		/* signed immediate constant */
 };
 
+/* Key of an a BPF_MAP_TYPE_LPM_TRIE entry */
+struct bpf_lpm_trie_key {
+	__u32	prefixlen;	/* up to 32 for AF_INET, 128 for AF_INET6 */
+	__u8	data[0];	/* Arbitrary size */
+};
+
 /* BPF syscall commands, see bpf(2) man-page for details. */
 enum bpf_cmd {
 	BPF_MAP_CREATE,
@@ -89,6 +95,7 @@ enum bpf_map_type {
 	BPF_MAP_TYPE_CGROUP_ARRAY,
 	BPF_MAP_TYPE_LRU_HASH,
 	BPF_MAP_TYPE_LRU_PERCPU_HASH,
+	BPF_MAP_TYPE_LPM_TRIE,
 };
 
 enum bpf_prog_type {
diff --git a/kernel/bpf/Makefile b/kernel/bpf/Makefile
index 1276474..e1ce4f4 100644
--- a/kernel/bpf/Makefile
+++ b/kernel/bpf/Makefile
@@ -1,7 +1,7 @@
 obj-y := core.o
 
 obj-$(CONFIG_BPF_SYSCALL) += syscall.o verifier.o inode.o helpers.o
-obj-$(CONFIG_BPF_SYSCALL) += hashtab.o arraymap.o percpu_freelist.o bpf_lru_list.o
+obj-$(CONFIG_BPF_SYSCALL) += hashtab.o arraymap.o percpu_freelist.o bpf_lru_list.o lpm_trie.o
 ifeq ($(CONFIG_PERF_EVENTS),y)
 obj-$(CONFIG_BPF_SYSCALL) += stackmap.o
 endif
diff --git a/kernel/bpf/lpm_trie.c b/kernel/bpf/lpm_trie.c
new file mode 100644
index 0000000..1c1ad27
--- /dev/null
+++ b/kernel/bpf/lpm_trie.c
@@ -0,0 +1,493 @@
+/*
+ * Longest prefix match list implementation
+ *
+ * Copyright (c) 2016,2017 Daniel Mack
+ * Copyright (c) 2016 David Herrmann
+ *
+ * This file is subject to the terms and conditions of version 2 of the GNU
+ * General Public License.  See the file COPYING in the main directory of the
+ * Linux distribution for more details.
+ */
+
+#include <linux/bpf.h>
+#include <linux/err.h>
+#include <linux/slab.h>
+#include <linux/spinlock.h>
+#include <linux/vmalloc.h>
+#include <net/ipv6.h>
+
+/* Intermediate node */
+#define LPM_TREE_NODE_FLAG_IM BIT(0)
+
+struct lpm_trie_node;
+
+struct lpm_trie_node {
+	struct rcu_head rcu;
+	struct lpm_trie_node __rcu	*child[2];
+	u32				prefixlen;
+	u32				flags;
+	u8				data[0];
+};
+
+struct lpm_trie {
+	struct bpf_map			map;
+	struct lpm_trie_node __rcu	*root;
+	size_t				n_entries;
+	size_t				max_prefixlen;
+	size_t				data_size;
+	raw_spinlock_t			lock;
+};
+
+/* This trie implements a longest prefix match algorithm that can be used to
+ * match IP addresses to a stored set of ranges.
+ *
+ * Data stored in @data of struct bpf_lpm_key and struct lpm_trie_node is
+ * interpreted as big endian, so data[0] stores the most significant byte.
+ *
+ * Match ranges are internally stored in instances of struct lpm_trie_node
+ * which each contain their prefix length as well as two pointers that may
+ * lead to more nodes containing more specific matches. Each node also stores
+ * a value that is defined by and returned to userspace via the update_elem
+ * and lookup functions.
+ *
+ * For instance, let's start with a trie that was created with a prefix length
+ * of 32, so it can be used for IPv4 addresses, and one single element that
+ * matches 192.168.0.0/16. The data array would hence contain
+ * [0xc0, 0xa8, 0x00, 0x00] in big-endian notation. This documentation will
+ * stick to IP-address notation for readability though.
+ *
+ * As the trie is empty initially, the new node (1) will be places as root
+ * node, denoted as (R) in the example below. As there are no other node, both
+ * child pointers are %NULL.
+ *
+ *              +----------------+
+ *              |       (1)  (R) |
+ *              | 192.168.0.0/16 |
+ *              |    value: 1    |
+ *              |   [0]    [1]   |
+ *              +----------------+
+ *
+ * Next, let's add a new node (2) matching 192.168.0.0/24. As there is already
+ * a node with the same data and a smaller prefix (ie, a less specific one),
+ * node (2) will become a child of (1). In child index depends on the next bit
+ * that is outside of what (1) matches, and that bit is 0, so (2) will be
+ * child[0] of (1):
+ *
+ *              +----------------+
+ *              |       (1)  (R) |
+ *              | 192.168.0.0/16 |
+ *              |    value: 1    |
+ *              |   [0]    [1]   |
+ *              +----------------+
+ *                   |
+ *    +----------------+
+ *    |       (2)      |
+ *    | 192.168.0.0/24 |
+ *    |    value: 2    |
+ *    |   [0]    [1]   |
+ *    +----------------+
+ *
+ * The child[1] slot of (1) could be filled with another node which has bit #17
+ * (the next bit after the ones that (1) matches on) set to 1. For instance,
+ * 192.168.128.0/24:
+ *
+ *              +----------------+
+ *              |       (1)  (R) |
+ *              | 192.168.0.0/16 |
+ *              |    value: 1    |
+ *              |   [0]    [1]   |
+ *              +----------------+
+ *                   |      |
+ *    +----------------+  +------------------+
+ *    |       (2)      |  |        (3)       |
+ *    | 192.168.0.0/24 |  | 192.168.128.0/24 |
+ *    |    value: 2    |  |     value: 3     |
+ *    |   [0]    [1]   |  |    [0]    [1]    |
+ *    +----------------+  +------------------+
+ *
+ * Let's add another node (4) to the game for 192.168.1.0/24. In order to place
+ * it, node (1) is looked at first, and because (4) of the semantics laid out
+ * above (bit #17 is 0), it would normally be attached to (1) as child[0].
+ * However, that slot is already allocated, so a new node is needed in between.
+ * That node does not have a value attached to it and it will never be
+ * returned to users as result of a lookup. It is only there to differentiate
+ * the traversal further. It will get a prefix as wide as necessary to
+ * distinguish its two children:
+ *
+ *                      +----------------+
+ *                      |       (1)  (R) |
+ *                      | 192.168.0.0/16 |
+ *                      |    value: 1    |
+ *                      |   [0]    [1]   |
+ *                      +----------------+
+ *                           |      |
+ *            +----------------+  +------------------+
+ *            |       (4)  (I) |  |        (3)       |
+ *            | 192.168.0.0/23 |  | 192.168.128.0/24 |
+ *            |    value: ---  |  |     value: 3     |
+ *            |   [0]    [1]   |  |    [0]    [1]    |
+ *            +----------------+  +------------------+
+ *                 |      |
+ *  +----------------+  +----------------+
+ *  |       (2)      |  |       (5)      |
+ *  | 192.168.0.0/24 |  | 192.168.1.0/24 |
+ *  |    value: 2    |  |     value: 5   |
+ *  |   [0]    [1]   |  |   [0]    [1]   |
+ *  +----------------+  +----------------+
+ *
+ * 192.168.1.1/32 would be a child of (5) etc.
+ *
+ * An intermediate node will be turned into a 'real' node on demand. In the
+ * example above, (4) would be re-used if 192.168.0.0/23 is added to the trie.
+ *
+ * A fully populated trie would have a height of 32 nodes, as the trie was
+ * created with a prefix length of 32.
+ *
+ * The lookup starts at the root node. If the current node matches and if there
+ * is a child that can be used to become more specific, the trie is traversed
+ * downwards. The last node in the traversal that is a non-intermediate one is
+ * returned.
+ */
+
+static inline int extract_bit(const u8 *data, size_t index)
+{
+	return !!(data[index / 8] & (1 << (7 - (index % 8))));
+}
+
+/**
+ * longest_prefix_match() - determine the longest prefix
+ * @trie:	The trie to get internal sizes from
+ * @node:	The node to operate on
+ * @key:	The key to compare to @node
+ *
+ * Determine the longest prefix of @node that matches the bits in @key.
+ */
+static size_t longest_prefix_match(const struct lpm_trie *trie,
+				   const struct lpm_trie_node *node,
+				   const struct bpf_lpm_trie_key *key)
+{
+	size_t prefixlen = 0;
+	size_t i;
+
+	for (i = 0; i < trie->data_size; i++) {
+		size_t b;
+
+		b = 8 - fls(node->data[i] ^ key->data[i]);
+		prefixlen += b;
+
+		if (prefixlen >= node->prefixlen || prefixlen >= key->prefixlen)
+			return min(node->prefixlen, key->prefixlen);
+
+		if (b < 8)
+			break;
+	}
+
+	return prefixlen;
+}
+
+/* Called from syscall or from eBPF program */
+static void *trie_lookup_elem(struct bpf_map *map, void *_key)
+{
+	struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
+	struct lpm_trie_node *node, *found = NULL;
+	struct bpf_lpm_trie_key *key = _key;
+
+	/* Start walking the trie from the root node ... */
+
+	for (node = rcu_dereference(trie->root); node;) {
+		unsigned int next_bit;
+		size_t matchlen;
+
+		/* Determine the longest prefix of @node that matches @key.
+		 * If it's the maximum possible prefix for this trie, we have
+		 * an exact match and can return it directly.
+		 */
+		matchlen = longest_prefix_match(trie, node, key);
+		if (matchlen == trie->max_prefixlen) {
+			found = node;
+			break;
+		}
+
+		/* If the number of bits that match is smaller than the prefix
+		 * length of @node, bail out and return the node we have seen
+		 * last in the traversal (ie, the parent).
+		 */
+		if (matchlen < node->prefixlen)
+			break;
+
+		/* Consider this node as return candidate unless it is an
+		 * artificially added intermediate one.
+		 */
+		if (!(node->flags & LPM_TREE_NODE_FLAG_IM))
+			found = node;
+
+		/* If the node match is fully satisfied, let's see if we can
+		 * become more specific. Determine the next bit in the key and
+		 * traverse down.
+		 */
+		next_bit = extract_bit(key->data, node->prefixlen);
+		node = rcu_dereference(node->child[next_bit]);
+	}
+
+	if (!found)
+		return NULL;
+
+	return found->data + trie->data_size;
+}
+
+static struct lpm_trie_node *lpm_trie_node_alloc(const struct lpm_trie *trie,
+						 const void *value)
+{
+	struct lpm_trie_node *node;
+	size_t size = sizeof(struct lpm_trie_node) + trie->data_size;
+
+	if (value)
+		size += trie->map.value_size;
+
+	node = kmalloc(size, GFP_ATOMIC | __GFP_NOWARN);
+	if (!node)
+		return NULL;
+
+	node->flags = 0;
+
+	if (value)
+		memcpy(node->data + trie->data_size, value,
+		       trie->map.value_size);
+
+	return node;
+}
+
+/* Called from syscall or from eBPF program */
+static int trie_update_elem(struct bpf_map *map,
+			    void *_key, void *value, u64 flags)
+{
+	struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
+	struct lpm_trie_node *node, *im_node, *new_node = NULL;
+	struct lpm_trie_node __rcu **slot;
+	struct bpf_lpm_trie_key *key = _key;
+	unsigned long irq_flags;
+	unsigned int next_bit;
+	size_t matchlen = 0;
+	int ret = 0;
+
+	if (unlikely(flags > BPF_EXIST))
+		return -EINVAL;
+
+	if (key->prefixlen > trie->max_prefixlen)
+		return -EINVAL;
+
+	raw_spin_lock_irqsave(&trie->lock, irq_flags);
+
+	/* Allocate and fill a new node */
+
+	if (trie->n_entries == trie->map.max_entries) {
+		ret = -ENOSPC;
+		goto out;
+	}
+
+	new_node = lpm_trie_node_alloc(trie, value);
+	if (!new_node) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	trie->n_entries++;
+
+	new_node->prefixlen = key->prefixlen;
+	RCU_INIT_POINTER(new_node->child[0], NULL);
+	RCU_INIT_POINTER(new_node->child[1], NULL);
+	memcpy(new_node->data, key->data, trie->data_size);
+
+	/* Now find a slot to attach the new node. To do that, walk the tree
+	 * from the root and match as many bits as possible for each node until
+	 * we either find an empty slot or a slot that needs to be replaced by
+	 * an intermediate node.
+	 */
+	slot = &trie->root;
+
+	while ((node = rcu_dereference_protected(*slot,
+					lockdep_is_held(&trie->lock)))) {
+		matchlen = longest_prefix_match(trie, node, key);
+
+		if (node->prefixlen != matchlen ||
+		    node->prefixlen == key->prefixlen ||
+		    node->prefixlen == trie->max_prefixlen)
+			break;
+
+		next_bit = extract_bit(key->data, node->prefixlen);
+		slot = &node->child[next_bit];
+	}
+
+	/* If the slot is empty (a free child pointer or an empty root),
+	 * simply assign the @new_node to that slot and be done.
+	 */
+	if (!node) {
+		rcu_assign_pointer(*slot, new_node);
+		goto out;
+	}
+
+	/* If the slot we picked already exists, replace it with @new_node
+	 * which already has the correct data array set.
+	 */
+	if (node->prefixlen == matchlen) {
+		new_node->child[0] = node->child[0];
+		new_node->child[1] = node->child[1];
+
+		if (!(node->flags & LPM_TREE_NODE_FLAG_IM))
+			trie->n_entries--;
+
+		rcu_assign_pointer(*slot, new_node);
+		kfree_rcu(node, rcu);
+
+		goto out;
+	}
+
+	/* If the new node matches the prefix completely, it must be inserted
+	 * as an ancestor. Simply insert it between @node and *@slot.
+	 */
+	if (matchlen == key->prefixlen) {
+		next_bit = extract_bit(node->data, matchlen);
+		rcu_assign_pointer(new_node->child[next_bit], node);
+		rcu_assign_pointer(*slot, new_node);
+		goto out;
+	}
+
+	im_node = lpm_trie_node_alloc(trie, NULL);
+	if (!im_node) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	im_node->prefixlen = matchlen;
+	im_node->flags |= LPM_TREE_NODE_FLAG_IM;
+	memcpy(im_node->data, node->data, trie->data_size);
+
+	/* Now determine which child to install in which slot */
+	if (extract_bit(key->data, matchlen)) {
+		rcu_assign_pointer(im_node->child[0], node);
+		rcu_assign_pointer(im_node->child[1], new_node);
+	} else {
+		rcu_assign_pointer(im_node->child[0], new_node);
+		rcu_assign_pointer(im_node->child[1], node);
+	}
+
+	/* Finally, assign the intermediate node to the determined spot */
+	rcu_assign_pointer(*slot, im_node);
+
+out:
+	if (ret) {
+		if (new_node)
+			trie->n_entries--;
+
+		kfree(new_node);
+		kfree(im_node);
+	}
+
+	raw_spin_unlock_irqrestore(&trie->lock, irq_flags);
+
+	return ret;
+}
+
+static struct bpf_map *trie_alloc(union bpf_attr *attr)
+{
+	size_t cost, cost_per_node;
+	struct lpm_trie *trie;
+	int ret;
+
+	/* check sanity of attributes */
+	if (attr->max_entries == 0 ||
+	    attr->map_flags != BPF_F_NO_PREALLOC ||
+	    attr->key_size < sizeof(struct bpf_lpm_trie_key) + 1   ||
+	    attr->key_size > sizeof(struct bpf_lpm_trie_key) + 256 ||
+	    attr->value_size == 0)
+		return ERR_PTR(-EINVAL);
+
+	trie = kzalloc(sizeof(*trie), GFP_USER | __GFP_NOWARN);
+	if (!trie)
+		return ERR_PTR(-ENOMEM);
+
+	/* copy mandatory map attributes */
+	trie->map.map_type = attr->map_type;
+	trie->map.key_size = attr->key_size;
+	trie->map.value_size = attr->value_size;
+	trie->map.max_entries = attr->max_entries;
+	trie->data_size = attr->key_size -
+			  offsetof(struct bpf_lpm_trie_key, data);
+	trie->max_prefixlen = trie->data_size * 8;
+
+	cost_per_node = sizeof(struct lpm_trie_node) +
+			attr->value_size + trie->data_size;
+	cost = sizeof(*trie) + attr->max_entries * cost_per_node;
+	trie->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
+
+	ret = bpf_map_precharge_memlock(trie->map.pages);
+	if (ret) {
+		kfree(trie);
+		return ERR_PTR(ret);
+	}
+
+	raw_spin_lock_init(&trie->lock);
+
+	return &trie->map;
+}
+
+static void trie_free(struct bpf_map *map)
+{
+	struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
+	struct lpm_trie_node __rcu **slot;
+	struct lpm_trie_node *node;
+
+	raw_spin_lock(&trie->lock);
+
+	/* Always start at the root and walk down to a node that has no
+	 * children. Then free that node, nullify its reference in the parent
+	 * and start over.
+	 */
+
+	for (;;) {
+		slot = &trie->root;
+
+		for (;;) {
+			node = rcu_dereference_protected(*slot,
+					lockdep_is_held(&trie->lock));
+			if (!node)
+				goto unlock;
+
+			if (rcu_access_pointer(node->child[0])) {
+				slot = &node->child[0];
+				continue;
+			}
+
+			if (rcu_access_pointer(node->child[1])) {
+				slot = &node->child[1];
+				continue;
+			}
+
+			kfree(node);
+			RCU_INIT_POINTER(*slot, NULL);
+			break;
+		}
+	}
+
+unlock:
+	raw_spin_unlock(&trie->lock);
+}
+
+static const struct bpf_map_ops trie_ops = {
+	.map_alloc = trie_alloc,
+	.map_free = trie_free,
+	.map_lookup_elem = trie_lookup_elem,
+	.map_update_elem = trie_update_elem,
+};
+
+static struct bpf_map_type_list trie_type __read_mostly = {
+	.ops = &trie_ops,
+	.type = BPF_MAP_TYPE_LPM_TRIE,
+};
+
+static int __init register_trie_map(void)
+{
+	bpf_register_map_type(&trie_type);
+	return 0;
+}
+late_initcall(register_trie_map);
-- 
2.9.3

[PATCH v3 2/2] bpf: Add tests for the lpm trie map

[Daniel Mack]

From: David Herrmann <dh.herrmann@gmail.com>

The first part of this program runs randomized tests against the
lpm-bpf-map. It implements a "Trivial Longest Prefix Match" (tlpm)
based on simple, linear, single linked lists. The implementation
should be pretty straightforward.

Based on tlpm, this inserts randomized data into bpf-lpm-maps and
verifies the trie-based bpf-map implementation behaves the same way
as tlpm.

The second part uses 'real world' IPv4 and IPv6 addresses and tests
the trie with those.

Signed-off-by: David Herrmann <dh.herrmann@gmail.com>
Signed-off-by: Daniel Mack <daniel@zonque.org>
---
 tools/testing/selftests/bpf/.gitignore     |   1 +
 tools/testing/selftests/bpf/Makefile       |   4 +-
 tools/testing/selftests/bpf/test_lpm_map.c | 358 +++++++++++++++++++++++++++++
 3 files changed, 361 insertions(+), 2 deletions(-)
 create mode 100644 tools/testing/selftests/bpf/test_lpm_map.c

diff --git a/tools/testing/selftests/bpf/.gitignore b/tools/testing/selftests/bpf/.gitignore
index 071431b..d3b1c9b 100644
--- a/tools/testing/selftests/bpf/.gitignore
+++ b/tools/testing/selftests/bpf/.gitignore
@@ -1,3 +1,4 @@
 test_verifier
 test_maps
 test_lru_map
+test_lpm_map
diff --git a/tools/testing/selftests/bpf/Makefile b/tools/testing/selftests/bpf/Makefile
index 7a5f245..064a3e5 100644
--- a/tools/testing/selftests/bpf/Makefile
+++ b/tools/testing/selftests/bpf/Makefile
@@ -1,8 +1,8 @@
 CFLAGS += -Wall -O2 -I../../../../usr/include
 
-test_objs = test_verifier test_maps test_lru_map
+test_objs = test_verifier test_maps test_lru_map test_lpm_map
 
-TEST_PROGS := test_verifier test_maps test_lru_map test_kmod.sh
+TEST_PROGS := test_verifier test_maps test_lru_map test_lpm_map test_kmod.sh
 TEST_FILES := $(test_objs)
 
 all: $(test_objs)
diff --git a/tools/testing/selftests/bpf/test_lpm_map.c b/tools/testing/selftests/bpf/test_lpm_map.c
new file mode 100644
index 0000000..dd83f0b
--- /dev/null
+++ b/tools/testing/selftests/bpf/test_lpm_map.c
@@ -0,0 +1,358 @@
+/*
+ * Randomized tests for eBPF longest-prefix-match maps
+ *
+ * This program runs randomized tests against the lpm-bpf-map. It implements a
+ * "Trivial Longest Prefix Match" (tlpm) based on simple, linear, singly linked
+ * lists. The implementation should be pretty straightforward.
+ *
+ * Based on tlpm, this inserts randomized data into bpf-lpm-maps and verifies
+ * the trie-based bpf-map implementation behaves the same way as tlpm.
+ */
+
+#include <assert.h>
+#include <errno.h>
+#include <inttypes.h>
+#include <linux/bpf.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <time.h>
+#include <unistd.h>
+#include <arpa/inet.h>
+#include <sys/time.h>
+#include <sys/resource.h>
+
+#include "bpf_sys.h"
+#include "bpf_util.h"
+
+struct tlpm_node {
+	struct tlpm_node *next;
+	size_t n_bits;
+	uint8_t key[];
+};
+
+static struct tlpm_node *tlpm_add(struct tlpm_node *list,
+				  const uint8_t *key,
+				  size_t n_bits)
+{
+	struct tlpm_node *node;
+	size_t n;
+
+	/* add new entry with @key/@n_bits to @list and return new head */
+
+	n = (n_bits + 7) / 8;
+	node = malloc(sizeof(*node) + n);
+	assert(node);
+
+	node->next = list;
+	node->n_bits = n_bits;
+	memcpy(node->key, key, n);
+
+	return node;
+}
+
+static void tlpm_clear(struct tlpm_node *list)
+{
+	struct tlpm_node *node;
+
+	/* free all entries in @list */
+
+	while ((node = list)) {
+		list = list->next;
+		free(node);
+	}
+}
+
+static struct tlpm_node *tlpm_match(struct tlpm_node *list,
+				    const uint8_t *key,
+				    size_t n_bits)
+{
+	struct tlpm_node *best = NULL;
+	size_t i;
+
+	/*
+	 * Perform longest prefix-match on @key/@n_bits. That is, iterate all
+	 * entries and match each prefix against @key. Remember the "best"
+	 * entry we find (i.e., the longest prefix that matches) and return it
+	 * to the caller when done.
+	 */
+
+	for ( ; list; list = list->next) {
+		for (i = 0; i < n_bits && i < list->n_bits; ++i) {
+			if ((key[i / 8] & (1 << (7 - i % 8))) !=
+			    (list->key[i / 8] & (1 << (7 - i % 8))))
+				break;
+		}
+
+		if (i >= list->n_bits) {
+			if (!best || i > best->n_bits)
+				best = list;
+		}
+	}
+
+	return best;
+}
+
+static void test_lpm_basic(void)
+{
+	struct tlpm_node *list = NULL, *t1, *t2;
+
+	/* very basic, static tests to verify tlpm works as expected */
+
+	assert(!tlpm_match(list, (uint8_t[]){ 0xff }, 8));
+
+	t1 = list = tlpm_add(list, (uint8_t[]){ 0xff }, 8);
+	assert(t1 == tlpm_match(list, (uint8_t[]){ 0xff }, 8));
+	assert(t1 == tlpm_match(list, (uint8_t[]){ 0xff, 0xff }, 16));
+	assert(t1 == tlpm_match(list, (uint8_t[]){ 0xff, 0x00 }, 16));
+	assert(!tlpm_match(list, (uint8_t[]){ 0x7f }, 8));
+	assert(!tlpm_match(list, (uint8_t[]){ 0xfe }, 8));
+	assert(!tlpm_match(list, (uint8_t[]){ 0xff }, 7));
+
+	t2 = list = tlpm_add(list, (uint8_t[]){ 0xff, 0xff }, 16);
+	assert(t1 == tlpm_match(list, (uint8_t[]){ 0xff }, 8));
+	assert(t2 == tlpm_match(list, (uint8_t[]){ 0xff, 0xff }, 16));
+	assert(t1 == tlpm_match(list, (uint8_t[]){ 0xff, 0xff }, 15));
+	assert(!tlpm_match(list, (uint8_t[]){ 0x7f, 0xff }, 16));
+
+	tlpm_clear(list);
+}
+
+static void test_lpm_order(void)
+{
+	struct tlpm_node *t1, *t2, *l1 = NULL, *l2 = NULL;
+	size_t i, j;
+
+	/*
+	 * Verify the tlpm implementation works correctly regardless of the
+	 * order of entries. Insert a random set of entries into @l1, and copy
+	 * the same data in reverse order into @l2. Then verify a lookup of
+	 * random keys will yield the same result in both sets.
+	 */
+
+	for (i = 0; i < (1 << 12); ++i)
+		l1 = tlpm_add(l1, (uint8_t[]){
+					rand() % 0xff,
+					rand() % 0xff,
+				}, rand() % 16 + 1);
+
+	for (t1 = l1; t1; t1 = t1->next)
+		l2 = tlpm_add(l2, t1->key, t1->n_bits);
+
+	for (i = 0; i < (1 << 8); ++i) {
+		uint8_t key[] = { rand() % 0xff, rand() % 0xff };
+
+		t1 = tlpm_match(l1, key, 16);
+		t2 = tlpm_match(l2, key, 16);
+
+		assert(!t1 == !t2);
+		if (t1) {
+			assert(t1->n_bits == t2->n_bits);
+			for (j = 0; j < t1->n_bits; ++j)
+				assert((t1->key[j / 8] & (1 << (7 - j % 8))) ==
+				       (t2->key[j / 8] & (1 << (7 - j % 8))));
+		}
+	}
+
+	tlpm_clear(l1);
+	tlpm_clear(l2);
+}
+
+static void test_lpm_map(void)
+{
+	size_t i, j, n_matches, n_nodes, n_lookups;
+	struct tlpm_node *t, *list = NULL;
+	struct bpf_lpm_trie_key *key;
+	uint8_t value[8] = {};
+	int r, map;
+
+	/*
+	 * Compare behavior of tlpm vs. bpf-lpm. Create a randomized set of
+	 * prefixes and insert it into both tlpm and bpf-lpm. Then run some
+	 * randomized lookups and verify both maps return the same result.
+	 */
+
+	n_matches = 0;
+	n_nodes = 1 << 8;
+	n_lookups = 1 << 16;
+
+	key = alloca(sizeof(*key) + 4);
+	memset(key, 0, sizeof(*key) + 4);
+
+	map = bpf_map_create(BPF_MAP_TYPE_LPM_TRIE,
+			     sizeof(*key) + 4,
+			     sizeof(value),
+			     4096,
+			     BPF_F_NO_PREALLOC);
+	assert(map >= 0);
+
+	for (i = 0; i < n_nodes; ++i) {
+		value[0] = rand() & 0xff;
+		value[1] = rand() & 0xff;
+		value[2] = rand() & 0xff;
+		value[3] = rand() & 0xff;
+		value[4] = rand() % 33;
+
+		list = tlpm_add(list, value, value[4]);
+
+		key->prefixlen = value[4];
+		memcpy(key->data, value, 4);
+		r = bpf_map_update(map, key, value, 0);
+		assert(!r);
+	}
+
+	for (i = 0; i < n_lookups; ++i) {
+		uint8_t data[] = {
+			rand() % 0xff,
+			rand() % 0xff,
+			rand() % 0xff,
+			rand() % 0xff
+		};
+
+		t = tlpm_match(list, data, 32);
+
+		key->prefixlen = 32;
+		memcpy(key->data, data, 4);
+		r = bpf_map_lookup(map, key, value);
+		assert(!r || errno == ENOENT);
+		assert(!t == !!r);
+
+		if (t) {
+			++n_matches;
+			assert(t->n_bits == value[4]);
+			for (j = 0; j < t->n_bits; ++j)
+				assert((t->key[j / 8] & (1 << (7 - j % 8))) ==
+				       (value[j / 8] & (1 << (7 - j % 8))));
+		}
+	}
+
+	close(map);
+	tlpm_clear(list);
+
+	/*
+	 * With 255 random nodes in the map, we are pretty likely to match
+	 * something on every lookup. For statistics, use this:
+	 *
+	 *     printf("  nodes: %zu\n"
+	 *            "lookups: %zu\n"
+	 *            "matches: %zu\n", n_nodes, n_lookups, n_matches);
+	 */
+}
+
+/* Test the implementation with some 'real world' examples */
+
+static void test_lpm_ipaddr(void)
+{
+	struct bpf_lpm_trie_key *key_ipv4;
+	struct bpf_lpm_trie_key *key_ipv6;
+	size_t key_size_ipv4;
+	size_t key_size_ipv6;
+	int map_fd_ipv4;
+	int map_fd_ipv6;
+	__u64 value;
+
+	key_size_ipv4 = sizeof(*key_ipv4) + sizeof(__u32);
+	key_size_ipv6 = sizeof(*key_ipv6) + sizeof(__u32) * 4;
+	key_ipv4 = alloca(key_size_ipv4);
+	key_ipv6 = alloca(key_size_ipv6);
+
+	map_fd_ipv4 = bpf_map_create(BPF_MAP_TYPE_LPM_TRIE,
+				     key_size_ipv4, sizeof(value),
+				     100, BPF_F_NO_PREALLOC);
+	assert(map_fd_ipv4 >= 0);
+
+	map_fd_ipv6 = bpf_map_create(BPF_MAP_TYPE_LPM_TRIE,
+				     key_size_ipv6, sizeof(value),
+				     100, BPF_F_NO_PREALLOC);
+	assert(map_fd_ipv6 >= 0);
+
+	/* Fill data some IPv4 and IPv6 address ranges */
+	value = 1;
+	key_ipv4->prefixlen = 16;
+	inet_pton(AF_INET, "192.168.0.0", key_ipv4->data);
+	assert(bpf_map_update(map_fd_ipv4, key_ipv4, &value, 0) == 0);
+
+	value = 2;
+	key_ipv4->prefixlen = 24;
+	inet_pton(AF_INET, "192.168.0.0", key_ipv4->data);
+	assert(bpf_map_update(map_fd_ipv4, key_ipv4, &value, 0) == 0);
+
+	value = 3;
+	key_ipv4->prefixlen = 24;
+	inet_pton(AF_INET, "192.168.128.0", key_ipv4->data);
+	assert(bpf_map_update(map_fd_ipv4, key_ipv4, &value, 0) == 0);
+
+	value = 5;
+	key_ipv4->prefixlen = 24;
+	inet_pton(AF_INET, "192.168.1.0", key_ipv4->data);
+	assert(bpf_map_update(map_fd_ipv4, key_ipv4, &value, 0) == 0);
+
+	value = 4;
+	key_ipv4->prefixlen = 23;
+	inet_pton(AF_INET, "192.168.0.0", key_ipv4->data);
+	assert(bpf_map_update(map_fd_ipv4, key_ipv4, &value, 0) == 0);
+
+	value = 0xdeadbeef;
+	key_ipv6->prefixlen = 64;
+	inet_pton(AF_INET6, "2a00:1450:4001:814::200e", key_ipv6->data);
+	assert(bpf_map_update(map_fd_ipv6, key_ipv6, &value, 0) == 0);
+
+	/* Set tprefixlen to maximum for lookups */
+	key_ipv4->prefixlen = 32;
+	key_ipv6->prefixlen = 128;
+
+	/* Test some lookups that should come back with a value */
+	inet_pton(AF_INET, "192.168.128.23", key_ipv4->data);
+	assert(bpf_map_lookup(map_fd_ipv4, key_ipv4, &value) == 0);
+	assert(value == 3);
+
+	inet_pton(AF_INET, "192.168.0.1", key_ipv4->data);
+	assert(bpf_map_lookup(map_fd_ipv4, key_ipv4, &value) == 0);
+	assert(value == 2);
+
+	inet_pton(AF_INET6, "2a00:1450:4001:814::", key_ipv6->data);
+	assert(bpf_map_lookup(map_fd_ipv6, key_ipv6, &value) == 0);
+	assert(value == 0xdeadbeef);
+
+	inet_pton(AF_INET6, "2a00:1450:4001:814::1", key_ipv6->data);
+	assert(bpf_map_lookup(map_fd_ipv6, key_ipv6, &value) == 0);
+	assert(value == 0xdeadbeef);
+
+	/* Test some lookups that should not match any entry */
+	inet_pton(AF_INET, "10.0.0.1", key_ipv4->data);
+	assert(bpf_map_lookup(map_fd_ipv4, key_ipv4, &value) == -1 &&
+	       errno == ENOENT);
+
+	inet_pton(AF_INET, "11.11.11.11", key_ipv4->data);
+	assert(bpf_map_lookup(map_fd_ipv4, key_ipv4, &value) == -1 &&
+	       errno == ENOENT);
+
+	inet_pton(AF_INET6, "2a00:ffff::", key_ipv6->data);
+	assert(bpf_map_lookup(map_fd_ipv6, key_ipv6, &value) == -1 &&
+	       errno == ENOENT);
+
+	close(map_fd_ipv4);
+	close(map_fd_ipv6);
+}
+
+int main(void)
+{
+	struct rlimit limit  = { RLIM_INFINITY, RLIM_INFINITY };
+	int ret;
+
+	/* we want predictable, pseudo random tests */
+	srand(0xf00ba1);
+
+	/* allow unlimited locked memory */
+	ret = setrlimit(RLIMIT_MEMLOCK, &limit);
+	if (ret < 0)
+		perror("Unable to lift memlock rlimit");
+
+	test_lpm_basic();
+	test_lpm_order();
+	test_lpm_map();
+	test_lpm_ipaddr();
+
+	printf("test_lpm: OK\n");
+	return 0;
+}
-- 
2.9.3

Re: [PATCH v3 1/2] bpf: add a longest prefix match trie map implementation

[Alexei Starovoitov]

On 1/14/17 4:17 AM, Daniel Mack wrote:
> +static struct bpf_map *trie_alloc(union bpf_attr *attr)
> +{
> +	size_t cost, cost_per_node;
> +	struct lpm_trie *trie;
> +	int ret;
> +
> +	/* check sanity of attributes */
> +	if (attr->max_entries == 0 ||
> +	    attr->map_flags != BPF_F_NO_PREALLOC ||
> +	    attr->key_size < sizeof(struct bpf_lpm_trie_key) + 1   ||
> +	    attr->key_size > sizeof(struct bpf_lpm_trie_key) + 256 ||
> +	    attr->value_size == 0)
> +		return ERR_PTR(-EINVAL);

could you also make it root only for now ?
Like we did for lru map:
         if (lru && !capable(CAP_SYS_ADMIN))
                 /* LRU implementation is much complicated than other
                  * maps.  Hence, limit to CAP_SYS_ADMIN for now.
                  */
                 return ERR_PTR(-EPERM);

trie is not that complicated, but I think socket_filters
(the only unpriv prog type today) can live a release or
two without ability to use it.

In patch 2/2 there are some comments not in networking style:
+	/*
+	 * Perform longest prefix-match on @key/@n_bits. That is,

As far as performance it has to be measured from datapath.
Like create tc+cls_bpf prog that does a dozen of trie lookups,
attach it to veth or tap and use
samples/pktgen/pktgen_bench_xmit_mode_netif_receive.sh
to send traffic into it.
Like: samples/bpf/test_cls_bpf.sh does.

Another alternative is to extend samples/bpf/map_perf_test
It has perf tests for most map types today (including lru)
and trie would be natural addition there.
I would prefer this latter option.

Thanks!

Re: [PATCH v3 1/2] bpf: add a longest prefix match trie map implementation

[David Herrmann]

Hi

On Sat, Jan 14, 2017 at 5:55 PM, Alexei Starovoitov <ast@fb.com> wrote:
> Another alternative is to extend samples/bpf/map_perf_test
> It has perf tests for most map types today (including lru)
> and trie would be natural addition there.
> I would prefer this latter option.

I hooked into gettid() and installed a simple kprobe bpf program that
searches for an entry in an lpm trie. The bpf program does either 0,
1, 8, or 32 lookups in a row (always the same element). The trie has
size 0, 1, or 8192. The data is below. The results vary by roughly 5%
on every run.

A single gettid() syscall with an empty bpf program takes roughly
6.5us on my system. Lookups in empty tries take ~1.8us on first try,
~0.9us on retries. Lookups in tries with 8192 entries take ~7.1us (on
the first _and_ any subsequent try).

https://gist.github.com/dvdhrm/4c90e61a1c39746d5c55ab9e0e29315e

Thanks
David


Trie-size: 0
#Lookups: 0
    0:lpm_perf kmalloc 9,230,321 events per sec
    -> 6.5us / syscall

Trie-size: 1
#Lookups: 1
    0:lpm_perf kmalloc 7,224,508 events per sec
    -> 8.3us / syscall

Trie-size: 1
#Lookups: 8
    0:lpm_perf kmalloc 4,152,740 events per sec
    -> 14.4us / syscall

Trie-size: 1
#Lookups: 32
    0:lpm_perf kmalloc 1,713,415 events per sec
    -> 35.0us / syscall

Trie-size: 8192
#Lookups: 1
    0:lpm_perf kmalloc 4,369,138 events per sec
    -> 13.7us / syscall

Trie-size: 8192
#Lookups: 8
    0:lpm_perf kmalloc 943,849 events per sec
    -> 63.6us / syscall

Trie-size: 8192
#Lookups: 32
    0:lpm_perf kmalloc 271,737 events per sec
    -> 220.8us / syscall

Re: [PATCH v3 1/2] bpf: add a longest prefix match trie map implementation

[Alexei Starovoitov]

On Wed, Jan 18, 2017 at 03:30:14PM +0100, David Herrmann wrote:
> Hi
> 
> On Sat, Jan 14, 2017 at 5:55 PM, Alexei Starovoitov <ast@fb.com> wrote:
> > Another alternative is to extend samples/bpf/map_perf_test
> > It has perf tests for most map types today (including lru)
> > and trie would be natural addition there.
> > I would prefer this latter option.
> 
> I hooked into gettid() and installed a simple kprobe bpf program that
> searches for an entry in an lpm trie. The bpf program does either 0,
> 1, 8, or 32 lookups in a row (always the same element). The trie has
> size 0, 1, or 8192. The data is below. The results vary by roughly 5%
> on every run.
> 
> A single gettid() syscall with an empty bpf program takes roughly
> 6.5us on my system. Lookups in empty tries take ~1.8us on first try,
> ~0.9us on retries. Lookups in tries with 8192 entries take ~7.1us (on
> the first _and_ any subsequent try).

thanks. please add them to commit log.

> https://gist.github.com/dvdhrm/4c90e61a1c39746d5c55ab9e0e29315e

looks good. please add it as an extra patch to this set.
and add
#pragma clang loop unroll(full)
before
+       for (i = 0; i < 8; ++i)
+               bpf_map_lookup_elem(&lpm_trie_map_alloc, &key);
to make sure it's unrolled regardless of the version of llvm.

> Trie-size: 0
> #Lookups: 0
>     0:lpm_perf kmalloc 9,230,321 events per sec
>     -> 6.5us / syscall
> 
> Trie-size: 1
> #Lookups: 1
>     0:lpm_perf kmalloc 7,224,508 events per sec
>     -> 8.3us / syscall
> 
> Trie-size: 1
> #Lookups: 8
>     0:lpm_perf kmalloc 4,152,740 events per sec
>     -> 14.4us / syscall
> 
> Trie-size: 1
> #Lookups: 32
>     0:lpm_perf kmalloc 1,713,415 events per sec
>     -> 35.0us / syscall
> 
> Trie-size: 8192
> #Lookups: 1
>     0:lpm_perf kmalloc 4,369,138 events per sec
>     -> 13.7us / syscall
> 
> Trie-size: 8192
> #Lookups: 8
>     0:lpm_perf kmalloc 943,849 events per sec
>     -> 63.6us / syscall
> 
> Trie-size: 8192
> #Lookups: 32
>     0:lpm_perf kmalloc 271,737 events per sec
>     -> 220.8us / syscall

so it's the same lookup done 32-times per each syscall.
Since the latency of single lookup in 8 vs 32 experiments
are pretty close, it means that the 32 lookup numbers
amortize the cost of syscall pretty well and no need to go higher.
So I would use 32 as a loop count in stress_lpm_trie_map_alloc()

Thanks!