[nf-next,v2,3/8] nf_tables: Add set type for arbitrary concatenation of ranges
diff mbox series

Message ID 5e7c454e030a8ad581a12d88881f96374e96da68.1574428269.git.sbrivio@redhat.com
State Changes Requested
Delegated to: Pablo Neira
Headers show
Series
  • nftables: Set implementation for arbitrary concatenation of ranges
Related show

Commit Message

Stefano Brivio Nov. 22, 2019, 1:40 p.m. UTC
This new set type allows for intervals in concatenated fields,
which are expressed in the usual way, that is, simple byte
concatenation with padding to 32 bits for single fields, and
given as ranges by specifying start and end elements containing,
each, the full concatenation of start and end values for the
single fields.

Ranges are expanded to composing netmasks, for each field: these
are inserted as rules in per-field lookup tables. Bits to be
classified are divided in 4-bit groups, and for each group, the
lookup table contains 4^2 buckets, representing all the possible
values of a bit group. This approach was inspired by the Grouper
algorithm:
	http://www.cse.usf.edu/~ligatti/projects/grouper/

Matching is performed by a sequence of AND operations between
bucket values, with buckets selected according to the value of
packet bits, for each group. The result of this sequence tells
us which rules matched for a given field.

In order to concatenate several ranged fields, per-field rules
are mapped using mapping arrays, one per field, that specify
which rules should be considered while matching the next field.
The mapping array for the last field contains a reference to
the element originally inserted.

The notes in nft_set_pipapo.c cover the algorithm in deeper
detail.

A pure hash-based approach is of no use here, as ranges need
to be classified. An implementation based on "proxying" the
existing red-black tree set type, creating a tree for each
field, was considered, but deemed impractical due to the fact
that elements would need to be shared between trees, at least
as long as we want to keep UAPI changes to a minimum.

A stand-alone implementation of this algorithm is available at:
	https://pipapo.lameexcu.se
together with notes about possible future optimisations
(in pipapo.c).

This algorithm was designed with data locality in mind, and can
be highly optimised for SIMD instruction sets, as the bulk of
the matching work is done with repetitive, simple bitwise
operations.

v2:
 - protect access to scratch maps in nft_pipapo_lookup() with
   local_bh_disable/enable() (Florian Westphal)
 - drop rcu_read_lock/unlock() from nft_pipapo_lookup(), it's
   already implied (Florian Westphal)
 - explain why partial allocation failures don't need handling
   in pipapo_realloc_scratch(), rename 'm' to clone and update
   related kerneldoc to make it clear we're not operating on
   the live copy (Florian Westphal)
 - add expicit check for priv->start_elem in
   nft_pipapo_insert() to avoid ending up in nft_pipapo_walk()
   with a NULL start element, and also zero it out in every
   operation that might make it invalid, so that insertion
   doesn't proceed with an invalid element (Florian Westphal)

Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
---
 include/net/netfilter/nf_tables_core.h |    1 +
 net/netfilter/Makefile                 |    3 +-
 net/netfilter/nf_tables_set_core.c     |    2 +
 net/netfilter/nft_set_pipapo.c         | 2197 ++++++++++++++++++++++++
 4 files changed, 2202 insertions(+), 1 deletion(-)
 create mode 100644 net/netfilter/nft_set_pipapo.c

Comments

Pablo Neira Ayuso Nov. 27, 2019, 9:29 a.m. UTC | #1
Hi Stefano,

Just started reading, a few initial questions.

On Fri, Nov 22, 2019 at 02:40:02PM +0100, Stefano Brivio wrote:
[...]
> diff --git a/include/net/netfilter/nf_tables_core.h b/include/net/netfilter/nf_tables_core.h
> index 7281895fa6d9..9759257ec8ec 100644
> --- a/include/net/netfilter/nf_tables_core.h
> +++ b/include/net/netfilter/nf_tables_core.h
> @@ -74,6 +74,7 @@ extern struct nft_set_type nft_set_hash_type;
>  extern struct nft_set_type nft_set_hash_fast_type;
>  extern struct nft_set_type nft_set_rbtree_type;
>  extern struct nft_set_type nft_set_bitmap_type;
> +extern struct nft_set_type nft_set_pipapo_type;
>  
>  struct nft_expr;
>  struct nft_regs;
> diff --git a/net/netfilter/Makefile b/net/netfilter/Makefile
> index 5e9b2eb24349..3f572e5a975e 100644
> --- a/net/netfilter/Makefile
> +++ b/net/netfilter/Makefile
> @@ -81,7 +81,8 @@ nf_tables-objs := nf_tables_core.o nf_tables_api.o nft_chain_filter.o \
>  		  nft_chain_route.o nf_tables_offload.o
>  
>  nf_tables_set-objs := nf_tables_set_core.o \
> -		      nft_set_hash.o nft_set_bitmap.o nft_set_rbtree.o
> +		      nft_set_hash.o nft_set_bitmap.o nft_set_rbtree.o \
> +		      nft_set_pipapo.o
>  
>  obj-$(CONFIG_NF_TABLES)		+= nf_tables.o
>  obj-$(CONFIG_NF_TABLES_SET)	+= nf_tables_set.o
> diff --git a/net/netfilter/nf_tables_set_core.c b/net/netfilter/nf_tables_set_core.c
> index a9fce8d10051..586b621007eb 100644
> --- a/net/netfilter/nf_tables_set_core.c
> +++ b/net/netfilter/nf_tables_set_core.c
> @@ -9,12 +9,14 @@ static int __init nf_tables_set_module_init(void)
>  	nft_register_set(&nft_set_rhash_type);
>  	nft_register_set(&nft_set_bitmap_type);
>  	nft_register_set(&nft_set_rbtree_type);
> +	nft_register_set(&nft_set_pipapo_type);
>  
>  	return 0;
>  }
>  
>  static void __exit nf_tables_set_module_exit(void)
>  {
> +	nft_unregister_set(&nft_set_pipapo_type);
>  	nft_unregister_set(&nft_set_rbtree_type);
>  	nft_unregister_set(&nft_set_bitmap_type);
>  	nft_unregister_set(&nft_set_rhash_type);
> diff --git a/net/netfilter/nft_set_pipapo.c b/net/netfilter/nft_set_pipapo.c
> new file mode 100644
> index 000000000000..3cad9aedc168
> --- /dev/null
> +++ b/net/netfilter/nft_set_pipapo.c
> @@ -0,0 +1,2197 @@
> +// SPDX-License-Identifier: GPL-2.0-only
> +
> +/* PIPAPO: PIle PAcket POlicies: set for arbitrary concatenations of ranges
> + *
> + * Copyright (c) 2019 Red Hat GmbH
> + *
> + * Author: Stefano Brivio <sbrivio@redhat.com>
> + */
> +
> +/**
> + * DOC: Theory of Operation
> + *
> + *
> + * Problem
> + * -------
> + *
> + * Match packet bytes against entries composed of ranged or non-ranged packet
> + * field specifiers, mapping them to arbitrary references. For example:
> + *
> + * ::
> + *
> + *               --- fields --->
> + *      |    [net],[port],[net]... => [reference]
> + *   entries [net],[port],[net]... => [reference]
> + *      |    [net],[port],[net]... => [reference]
> + *      V    ...
> + *
> + * where [net] fields can be IP ranges or netmasks, and [port] fields are port
> + * ranges. Arbitrary packet fields can be matched.
> + *
> + *
> + * Algorithm Overview
> + * ------------------
> + *
> + * This algorithm is loosely inspired by [Ligatti 2010], and fundamentally
> + * relies on the consideration that every contiguous range in a space of b bits
> + * can be converted into b * 2 netmasks, from Theorem 3 in [Rottenstreich 2010],
> + * as also illustrated in Section 9 of [Kogan 2014].
> + *
> + * Classification against a number of entries, that require matching given bits
> + * of a packet field, is performed by grouping those bits in sets of arbitrary
> + * size, and classifying packet bits one group at a time.
> + *
> + * Example:
> + *   to match the source port (16 bits) of a packet, we can divide those 16 bits
> + *   in 4 groups of 4 bits each. Given the entry:
> + *      0000 0001 0101 1001
> + *   and a packet with source port:
> + *      0000 0001 1010 1001
> + *   first and second groups match, but the third doesn't. We conclude that the
> + *   packet doesn't match the given entry.
> + *
> + * Translate the set to a sequence of lookup tables, one per field. Each table
> + * has two dimensions: bit groups to be matched for a single packet field, and
> + * all the possible values of said groups (buckets). Input entries are
> + * represented as one or more rules, depending on the number of composing
> + * netmasks for the given field specifier, and a group match is indicated as a
> + * set bit, with number corresponding to the rule index, in all the buckets
> + * whose value matches the entry for a given group.
> + *
> + * Rules are mapped between fields through an array of x, n pairs, with each
> + * item mapping a matched rule to one or more rules. The position of the pair in
> + * the array indicates the matched rule to be mapped to the next field, x
> + * indicates the first rule index in the next field, and n the amount of
> + * next-field rules the current rule maps to.
> + *
> + * The mapping array for the last field maps to the desired references.
> + *
> + * To match, we perform table lookups using the values of grouped packet bits,
> + * and use a sequence of bitwise operations to progressively evaluate rule
> + * matching.
> + *
> + * A stand-alone, reference implementation, also including notes about possible
> + * future optimisations, is available at:
> + *    https://pipapo.lameexcu.se/
> + *
> + * Insertion
> + * ---------
> + *
> + * - For each packet field:
> + *
> + *   - divide the b packet bits we want to classify into groups of size t,
> + *     obtaining ceil(b / t) groups
> + *
> + *      Example: match on destination IP address, with t = 4: 32 bits, 8 groups
> + *      of 4 bits each
> + *
> + *   - allocate a lookup table with one column ("bucket") for each possible
> + *     value of a group, and with one row for each group
> + *
> + *      Example: 8 groups, 2^4 buckets:
> + *
> + * ::
> + *
> + *                     bucket
> + *      group  0   1   2   3   4   5   6   7   8   9  10  11  12  13  14  15
> + *        0
> + *        1
> + *        2
> + *        3
> + *        4
> + *        5
> + *        6
> + *        7
> + *
> + *   - map the bits we want to classify for the current field, for a given
> + *     entry, to a single rule for non-ranged and netmask set items, and to one
> + *     or multiple rules for ranges. Ranges are expanded to composing netmasks
> + *     by pipapo_expand().
> + *
> + *      Example: 2 entries, 10.0.0.5:1024 and 192.168.1.0-192.168.2.1:2048
> + *      - rule #0: 10.0.0.5
> + *      - rule #1: 192.168.1.0/24
> + *      - rule #2: 192.168.2.0/31
> + *
> + *   - insert references to the rules in the lookup table, selecting buckets
> + *     according to bit values of a rule in the given group. This is done by
> + *     pipapo_insert().
> + *
> + *      Example: given:
> + *      - rule #0: 10.0.0.5 mapping to buckets
> + *        < 0 10  0 0   0 0  0 5 >
> + *      - rule #1: 192.168.1.0/24 mapping to buckets
> + *        < 12 0  10 8  0 1  < 0..15 > < 0..15 > >
> + *      - rule #2: 192.168.2.0/31 mapping to buckets
> + *        < 12 0  10 8  0 2  0 < 0..1 > >
> + *
> + *      these bits are set in the lookup table:
> + *
> + * ::
> + *
> + *                     bucket
> + *      group  0   1   2   3   4   5   6   7   8   9  10  11  12  13  14  15
> + *        0    0                                              1,2
> + *        1   1,2                                      0
> + *        2    0                                      1,2
> + *        3    0                              1,2
> + *        4  0,1,2
> + *        5    0   1   2
> + *        6  0,1,2 1   1   1   1   1   1   1   1   1   1   1   1   1   1   1
> + *        7   1,2 1,2  1   1   1  0,1  1   1   1   1   1   1   1   1   1   1
> + *
> + *   - if this is not the last field in the set, fill a mapping array that maps
> + *     rules from the lookup table to rules belonging to the same entry in
> + *     the next lookup table, done by pipapo_map().
> + *
> + *     Note that as rules map to contiguous ranges of rules, given how netmask
> + *     expansion and insertion is performed, &union nft_pipapo_map_bucket stores
> + *     this information as pairs of first rule index, rule count.
> + *
> + *      Example: 2 entries, 10.0.0.5:1024 and 192.168.1.0-192.168.2.1:2048,
> + *      given lookup table #0 for field 0 (see example above):
> + *
> + * ::
> + *
> + *                     bucket
> + *      group  0   1   2   3   4   5   6   7   8   9  10  11  12  13  14  15
> + *        0    0                                              1,2
> + *        1   1,2                                      0
> + *        2    0                                      1,2
> + *        3    0                              1,2
> + *        4  0,1,2
> + *        5    0   1   2
> + *        6  0,1,2 1   1   1   1   1   1   1   1   1   1   1   1   1   1   1
> + *        7   1,2 1,2  1   1   1  0,1  1   1   1   1   1   1   1   1   1   1
> + *
> + *      and lookup table #1 for field 1 with:
> + *      - rule #0: 1024 mapping to buckets
> + *        < 0  0  4  0 >
> + *      - rule #1: 2048 mapping to buckets
> + *        < 0  0  5  0 >
> + *
> + * ::
> + *
> + *                     bucket
> + *      group  0   1   2   3   4   5   6   7   8   9  10  11  12  13  14  15
> + *        0   0,1
> + *        1   0,1
> + *        2                    0   1
> + *        3   0,1
> + *
> + *      we need to map rules for 10.0.0.5 in lookup table #0 (rule #0) to 1024
> + *      in lookup table #1 (rule #0) and rules for 192.168.1.0-192.168.2.1
> + *      (rules #1, #2) to 2048 in lookup table #2 (rule #1):
> + *
> + * ::
> + *
> + *       rule indices in current field: 0    1    2
> + *       map to rules in next field:    0    1    1
> + *
> + *   - if this is the last field in the set, fill a mapping array that maps
> + *     rules from the last lookup table to element pointers, also done by
> + *     pipapo_map().
> + *
> + *     Note that, in this implementation, we have two elements (start, end) for
> + *     each entry. The pointer to the end element is stored in this array, and
> + *     the pointer to the start element is linked from it.
> + *
> + *      Example: entry 10.0.0.5:1024 has a corresponding &struct nft_pipapo_elem
> + *      pointer, 0x66, and element for 192.168.1.0-192.168.2.1:2048 is at 0x42.
> + *      From the rules of lookup table #1 as mapped above:
> + *
> + * ::
> + *
> + *       rule indices in last field:    0    1
> + *       map to elements:             0x42  0x66
> + *
> + *
> + * Matching
> + * --------
> + *
> + * We use a result bitmap, with the size of a single lookup table bucket, to
> + * represent the matching state that applies at every algorithm step. This is
> + * done by pipapo_lookup().
> + *
> + * - For each packet field:
> + *
> + *   - start with an all-ones result bitmap (res_map in pipapo_lookup())
> + *
> + *   - perform a lookup into the table corresponding to the current field,
> + *     for each group, and at every group, AND the current result bitmap with
> + *     the value from the lookup table bucket
> + *
> + * ::
> + *
> + *      Example: 192.168.1.5 < 12 0  10 8  0 1  0 5 >, with lookup table from
> + *      insertion examples.
> + *      Lookup table buckets are at least 3 bits wide, we'll assume 8 bits for
> + *      convenience in this example. Initial result bitmap is 0xff, the steps
> + *      below show the value of the result bitmap after each group is processed:
> + *
> + *                     bucket
> + *      group  0   1   2   3   4   5   6   7   8   9  10  11  12  13  14  15
> + *        0    0                                              1,2
> + *        result bitmap is now: 0xff & 0x6 [bucket 12] = 0x6
> + *
> + *        1   1,2                                      0
> + *        result bitmap is now: 0x6 & 0x6 [bucket 0] = 0x6
> + *
> + *        2    0                                      1,2
> + *        result bitmap is now: 0x6 & 0x6 [bucket 10] = 0x6
> + *
> + *        3    0                              1,2
> + *        result bitmap is now: 0x6 & 0x6 [bucket 8] = 0x6
> + *
> + *        4  0,1,2
> + *        result bitmap is now: 0x6 & 0x7 [bucket 0] = 0x6
> + *
> + *        5    0   1   2
> + *        result bitmap is now: 0x6 & 0x2 [bucket 1] = 0x2
> + *
> + *        6  0,1,2 1   1   1   1   1   1   1   1   1   1   1   1   1   1   1
> + *        result bitmap is now: 0x2 & 0x7 [bucket 0] = 0x2
> + *
> + *        7   1,2 1,2  1   1   1  0,1  1   1   1   1   1   1   1   1   1   1
> + *        final result bitmap for this field is: 0x2 & 0x3 [bucket 5] = 0x2
> + *
> + *   - at the next field, start with a new, all-zeroes result bitmap. For each
> + *     bit set in the previous result bitmap, fill the new result bitmap
> + *     (fill_map in pipapo_lookup()) with the rule indices from the
> + *     corresponding buckets of the mapping field for this field, done by
> + *     pipapo_refill()
> + *
> + *      Example: with mapping table from insertion examples, with the current
> + *      result bitmap from the previous example, 0x02:
> + *
> + * ::
> + *
> + *       rule indices in current field: 0    1    2
> + *       map to rules in next field:    0    1    1
> + *
> + *      the new result bitmap will be 0x02: rule 1 was set, and rule 1 will be
> + *      set.
> + *
> + *      We can now extend this example to cover the second iteration of the step
> + *      above (lookup and AND bitmap): assuming the port field is
> + *      2048 < 0  0  5  0 >, with starting result bitmap 0x2, and lookup table
> + *      for "port" field from pre-computation example:
> + *
> + * ::
> + *
> + *                     bucket
> + *      group  0   1   2   3   4   5   6   7   8   9  10  11  12  13  14  15
> + *        0   0,1
> + *        1   0,1
> + *        2                    0   1
> + *        3   0,1
> + *
> + *       operations are: 0x2 & 0x3 [bucket 0] & 0x3 [bucket 0] & 0x2 [bucket 5]
> + *       & 0x3 [bucket 0], resulting bitmap is 0x2.
> + *
> + *   - if this is the last field in the set, look up the value from the mapping
> + *     array corresponding to the final result bitmap
> + *
> + *      Example: 0x2 resulting bitmap from 192.168.1.5:2048, mapping array for
> + *      last field from insertion example:
> + *
> + * ::
> + *
> + *       rule indices in last field:    0    1
> + *       map to elements:             0x42  0x66
> + *
> + *      the matching element is at 0x42.
> + *
> + *
> + * References
> + * ----------
> + *
> + * [Ligatti 2010]
> + *      A Packet-classification Algorithm for Arbitrary Bitmask Rules, with
> + *      Automatic Time-space Tradeoffs
> + *      Jay Ligatti, Josh Kuhn, and Chris Gage.
> + *      Proceedings of the IEEE International Conference on Computer
> + *      Communication Networks (ICCCN), August 2010.
> + *      http://www.cse.usf.edu/~ligatti/papers/grouper-conf.pdf
> + *
> + * [Rottenstreich 2010]
> + *      Worst-Case TCAM Rule Expansion
> + *      Ori Rottenstreich and Isaac Keslassy.
> + *      2010 Proceedings IEEE INFOCOM, San Diego, CA, 2010.
> + *      http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.212.4592&rep=rep1&type=pdf
> + *
> + * [Kogan 2014]
> + *      SAX-PAC (Scalable And eXpressive PAcket Classification)
> + *      Kirill Kogan, Sergey Nikolenko, Ori Rottenstreich, William Culhane,
> + *      and Patrick Eugster.
> + *      Proceedings of the 2014 ACM conference on SIGCOMM, August 2014.
> + *      http://www.sigcomm.org/sites/default/files/ccr/papers/2014/August/2619239-2626294.pdf
> + */
> +
> +#include <linux/kernel.h>
> +#include <linux/init.h>
> +#include <linux/log2.h>
> +#include <linux/module.h>
> +#include <linux/netlink.h>
> +#include <linux/netfilter.h>
> +#include <linux/netfilter/nf_tables.h>
> +#include <net/netfilter/nf_tables_core.h>
> +#include <uapi/linux/netfilter/nf_tables.h>
> +#include <net/ipv6.h>			/* For the maximum length of a field */
> +#include <linux/bitmap.h>
> +#include <linux/bitops.h>
> +
> +/* Count of concatenated fields depends on count of 32-bit nftables registers */
> +#define NFT_PIPAPO_MAX_FIELDS		NFT_REG32_COUNT
> +
> +/* Largest supported field size */
> +#define NFT_PIPAPO_MAX_BYTES		(sizeof(struct in6_addr))
> +#define NFT_PIPAPO_MAX_BITS		(NFT_PIPAPO_MAX_BYTES * BITS_PER_BYTE)
> +
> +/* Number of bits to be grouped together in lookup table buckets, arbitrary */
> +#define NFT_PIPAPO_GROUP_BITS		4
> +#define NFT_PIPAPO_GROUPS_PER_BYTE	(BITS_PER_BYTE / NFT_PIPAPO_GROUP_BITS)
> +
> +/* Fields are padded to 32 bits in input registers */
> +#define NFT_PIPAPO_GROUPS_PADDED_SIZE(x)				\
> +	(round_up((x) / NFT_PIPAPO_GROUPS_PER_BYTE, sizeof(u32)))
> +#define NFT_PIPAPO_GROUPS_PADDING(x)					\
> +	(NFT_PIPAPO_GROUPS_PADDED_SIZE((x)) - (x) / NFT_PIPAPO_GROUPS_PER_BYTE)
> +
> +/* Number of buckets, given by 2 ^ n, with n grouped bits */
> +#define NFT_PIPAPO_BUCKETS		(1 << NFT_PIPAPO_GROUP_BITS)
> +
> +/* Each n-bit range maps to up to n * 2 rules */
> +#define NFT_PIPAPO_MAP_NBITS		(const_ilog2(NFT_PIPAPO_MAX_BITS * 2))
> +
> +/* Use the rest of mapping table buckets for rule indices, but it makes no sense
> + * to exceed 32 bits
> + */
> +#if BITS_PER_LONG == 64
> +#define NFT_PIPAPO_MAP_TOBITS		32
> +#else
> +#define NFT_PIPAPO_MAP_TOBITS		(BITS_PER_LONG - NFT_PIPAPO_MAP_NBITS)
> +#endif
> +
> +/* ...which gives us the highest allowed index for a rule */
> +#define NFT_PIPAPO_RULE0_MAX		((1UL << (NFT_PIPAPO_MAP_TOBITS - 1)) \
> +					- (1UL << NFT_PIPAPO_MAP_NBITS))
> +
> +#define nft_pipapo_for_each_field(field, index, match)		\
> +	for ((field) = (match)->f, (index) = 0;			\
> +	     (index) < (match)->field_count;			\
> +	     (index)++, (field)++)
> +
> +/**
> + * union nft_pipapo_map_bucket - Bucket of mapping table
> + * @to:		First rule number (in next field) this rule maps to
> + * @n:		Number of rules (in next field) this rule maps to
> + * @e:		If there's no next field, pointer to element this rule maps to
> + */
> +union nft_pipapo_map_bucket {
> +	struct {
> +#if BITS_PER_LONG == 64
> +		static_assert(NFT_PIPAPO_MAP_TOBITS <= 32);
> +		u32 to;
> +
> +		static_assert(NFT_PIPAPO_MAP_NBITS <= 32);
> +		u32 n;
> +#else
> +		unsigned long to:NFT_PIPAPO_MAP_TOBITS;
> +		unsigned long  n:NFT_PIPAPO_MAP_NBITS;
> +#endif
> +	};
> +	struct nft_pipapo_elem *e;
> +};
> +
> +/**
> + * struct nft_pipapo_field - Lookup, mapping tables and related data for a field
> + * @groups:	Amount of 4-bit groups
> + * @rules:	Number of inserted rules
> + * @bsize:	Size of each bucket in lookup table, in longs
> + * @lt:		Lookup table: 'groups' rows of NFT_PIPAPO_BUCKETS buckets
> + * @mt:		Mapping table: one bucket per rule
> + */
> +struct nft_pipapo_field {
> +	int groups;
> +	unsigned long rules;
> +	size_t bsize;
> +	unsigned long *lt;
> +	union nft_pipapo_map_bucket *mt;
> +};
> +
> +/**
> + * struct nft_pipapo_match - Data used for lookup and matching
> + * @field_count		Amount of fields in set
> + * @scratch:		Preallocated per-CPU maps for partial matching results
> + * @bsize_max:		Maximum lookup table bucket size of all fields, in longs
> + * @rcu			Matching data is swapped on commits
> + * @f:			Fields, with lookup and mapping tables
> + */
> +struct nft_pipapo_match {
> +	int field_count;
> +	unsigned long * __percpu *scratch;
> +	size_t bsize_max;
> +	struct rcu_head rcu;
> +	struct nft_pipapo_field f[0];
> +};
> +
> +/* Current working bitmap index, toggled between field matches */
> +static DEFINE_PER_CPU(bool, nft_pipapo_scratch_index);
> +
> +/**
> + * struct nft_pipapo - Representation of a set
> + * @match:	Currently in-use matching data
> + * @clone:	Copy where pending insertions and deletions are kept
> + * @groups:	Total amount of 4-bit groups for fields in this set
> + * @width:	Total bytes to be matched for one packet, including padding
> + * @dirty:	Working copy has pending insertions or deletions
> + * @last_gc:	Timestamp of last garbage collection run, jiffies
> + * @start_data:	Key data of start element for insertion
> + * @start_elem:	Start element for insertion
> + */
> +struct nft_pipapo {
> +	struct nft_pipapo_match __rcu *match;
> +	struct nft_pipapo_match *clone;
> +	int groups;
> +	int width;
> +	bool dirty;
> +	unsigned long last_gc;
> +	u8 start_data[NFT_DATA_VALUE_MAXLEN * sizeof(u32)];
> +	struct nft_pipapo_elem *start_elem;
> +};
> +
> +struct nft_pipapo_elem;
> +
> +/**
> + * struct nft_pipapo_elem - API-facing representation of single set element
> + * @start:	Pointer to element that represents start of interval
> + * @ext:	nftables API extensions
> + */
> +struct nft_pipapo_elem {
> +	struct nft_pipapo_elem *start;
> +	struct nft_set_ext ext;
> +};
> +
> +/**
> + * pipapo_refill() - For each set bit, set bits from selected mapping table item
> + * @map:	Bitmap to be scanned for set bits
> + * @len:	Length of bitmap in longs
> + * @rules:	Number of rules in field
> + * @dst:	Destination bitmap
> + * @mt:		Mapping table containing bit set specifiers
> + * @match_only:	Find a single bit and return, don't fill
> + *
> + * Iteration over set bits with __builtin_ctzl(): Daniel Lemire, public domain.
> + *
> + * For each bit set in map, select the bucket from mapping table with index
> + * corresponding to the position of the bit set. Use start bit and amount of
> + * bits specified in bucket to fill region in dst.
> + *
> + * Return: -1 on no match, bit position on 'match_only', 0 otherwise.
> + */
> +static int pipapo_refill(unsigned long *map, int len, int rules,
> +			 unsigned long *dst, union nft_pipapo_map_bucket *mt,
> +			 bool match_only)
> +{
> +	unsigned long bitset;
> +	int k, ret = -1;
> +
> +	for (k = 0; k < len; k++) {
> +		bitset = map[k];
> +		while (bitset) {
> +			unsigned long t = bitset & -bitset;
> +			int r = __builtin_ctzl(bitset);
> +			int i = k * BITS_PER_LONG + r;
> +
> +			if (unlikely(i >= rules)) {
> +				map[k] = 0;
> +				return -1;
> +			}
> +
> +			if (unlikely(match_only)) {
> +				bitmap_clear(map, i, 1);
> +				return i;
> +			}
> +
> +			ret = 0;
> +
> +			bitmap_set(dst, mt[i].to, mt[i].n);
> +
> +			bitset ^= t;
> +		}
> +		map[k] = 0;
> +	}
> +
> +	return ret;
> +}
> +
> +/**
> + * nft_pipapo_lookup() - Lookup function
> + * @net:	Network namespace
> + * @set:	nftables API set representation
> + * @elem:	nftables API element representation containing key data
> + * @ext:	nftables API extension pointer, filled with matching reference
> + *
> + * For more details, see DOC: Theory of Operation.
> + *
> + * Return: true on match, false otherwise.
> + */
> +static bool nft_pipapo_lookup(const struct net *net, const struct nft_set *set,
> +			      const u32 *key, const struct nft_set_ext **ext)
> +{
> +	struct nft_pipapo *priv = nft_set_priv(set);
> +	unsigned long *res_map, *fill_map;
> +	u8 genmask = nft_genmask_cur(net);
> +	const u8 *rp = (const u8 *)key;
> +	struct nft_pipapo_match *m;
> +	struct nft_pipapo_field *f;
> +	bool map_index;
> +	int i;
> +
> +	local_bh_disable();
> +
> +	map_index = raw_cpu_read(nft_pipapo_scratch_index);
> +
> +	m = rcu_dereference(priv->match);
> +
> +	if (unlikely(!m || !*raw_cpu_ptr(m->scratch)))
> +		goto out;
> +
> +	res_map  = *raw_cpu_ptr(m->scratch) + (map_index ? m->bsize_max : 0);
> +	fill_map = *raw_cpu_ptr(m->scratch) + (map_index ? 0 : m->bsize_max);
> +
> +	memset(res_map, 0xff, m->bsize_max * sizeof(*res_map));
> +
> +	nft_pipapo_for_each_field(f, i, m) {
> +		bool last = i == m->field_count - 1;
> +		unsigned long *lt = f->lt;
> +		int b, group;
> +
> +		/* For each 4-bit group: select lookup table bucket depending on
> +		 * packet bytes value, then AND bucket value
> +		 */
> +		for (group = 0; group < f->groups; group++) {
> +			u8 v;
> +
> +			if (group % 2) {
> +				v = *rp & 0x0f;
> +				rp++;
> +			} else {
> +				v = *rp >> 4;
> +			}
> +			__bitmap_and(res_map, res_map, lt + v * f->bsize,
> +				     f->bsize * BITS_PER_LONG);
> +
> +			lt += f->bsize * NFT_PIPAPO_BUCKETS;
> +		}
> +
> +		/* Now populate the bitmap for the next field, unless this is
> +		 * the last field, in which case return the matched 'ext'
> +		 * pointer if any.
> +		 *
> +		 * Now res_map contains the matching bitmap, and fill_map is the
> +		 * bitmap for the next field.
> +		 */
> +next_match:
> +		b = pipapo_refill(res_map, f->bsize, f->rules, fill_map, f->mt,
> +				  last);
> +		if (b < 0) {
> +			raw_cpu_write(nft_pipapo_scratch_index, map_index);
> +			local_bh_enable();
> +
> +			return false;
> +		}
> +
> +		if (last) {
> +			*ext = &f->mt[b].e->ext;
> +			if (unlikely(nft_set_elem_expired(*ext) ||
> +				     !nft_set_elem_active(*ext, genmask)))
> +				goto next_match;
> +
> +			/* Last field: we're just returning the key without
> +			 * filling the initial bitmap for the next field, so the
> +			 * current inactive bitmap is clean and can be reused as
> +			 * *next* bitmap (not initial) for the next packet.
> +			 */
> +			raw_cpu_write(nft_pipapo_scratch_index, map_index);
> +			local_bh_enable();
> +
> +			return true;
> +		}
> +
> +		/* Swap bitmap indices: res_map is the initial bitmap for the
> +		 * next field, and fill_map is guaranteed to be all-zeroes at
> +		 * this point.
> +		 */
> +		map_index = !map_index;
> +		swap(res_map, fill_map);
> +
> +		rp += NFT_PIPAPO_GROUPS_PADDING(f->groups);
> +	}
> +
> +out:
> +	local_bh_enable();
> +	return false;
> +}
> +
> +/**
> + * pipapo_get() - Get matching start or end element reference given key data
> + * @net:	Network namespace
> + * @set:	nftables API set representation
> + * @data:	Key data to be matched against existing elements
> + * @flags:	If NFT_SET_ELEM_INTERVAL_END is passed, return the end element
> + *
> + * This is essentially the same as the lookup function, except that it matches
> + * key data against the uncommitted copy and doesn't use preallocated maps for
> + * bitmap results.
> + *
> + * Return: pointer to &struct nft_pipapo_elem on match, error pointer otherwise.
> + */
> +static void *pipapo_get(const struct net *net, const struct nft_set *set,
> +			const u8 *data, unsigned int flags)
> +{
> +	struct nft_pipapo *priv = nft_set_priv(set);
> +	struct nft_pipapo_match *m = priv->clone;
> +	unsigned long *res_map, *fill_map = NULL;
> +	void *ret = ERR_PTR(-ENOENT);
> +	struct nft_pipapo_field *f;
> +	int i;
> +
> +	res_map = kmalloc_array(m->bsize_max, sizeof(*res_map), GFP_ATOMIC);
> +	if (!res_map) {
> +		ret = ERR_PTR(-ENOMEM);
> +		goto out;
> +	}
> +
> +	fill_map = kcalloc(m->bsize_max, sizeof(*res_map), GFP_ATOMIC);
> +	if (!fill_map) {
> +		ret = ERR_PTR(-ENOMEM);
> +		goto out;
> +	}
> +
> +	memset(res_map, 0xff, m->bsize_max * sizeof(*res_map));
> +
> +	nft_pipapo_for_each_field(f, i, m) {
> +		bool last = i == m->field_count - 1;
> +		unsigned long *lt = f->lt;
> +		int b, group;
> +
> +		/* For each 4-bit group: select lookup table bucket depending on
> +		 * packet bytes value, then AND bucket value
> +		 */
> +		for (group = 0; group < f->groups; group++) {
> +			u8 v;
> +
> +			if (group % 2) {
> +				v = *data & 0x0f;
> +				data++;
> +			} else {
> +				v = *data >> 4;
> +			}
> +			__bitmap_and(res_map, res_map, lt + v * f->bsize,
> +				     f->bsize * BITS_PER_LONG);
> +
> +			lt += f->bsize * NFT_PIPAPO_BUCKETS;
> +		}
> +
> +		/* Now populate the bitmap for the next field, unless this is
> +		 * the last field, in which case return the matched 'ext'
> +		 * pointer if any.
> +		 *
> +		 * Now res_map contains the matching bitmap, and fill_map is the
> +		 * bitmap for the next field.
> +		 */
> +next_match:
> +		b = pipapo_refill(res_map, f->bsize, f->rules, fill_map, f->mt,
> +				  last);
> +		if (b < 0)
> +			goto out;
> +
> +		if (last) {
> +			if (nft_set_elem_expired(&f->mt[b].e->ext))
> +				goto next_match;
> +
> +			if (flags & NFT_SET_ELEM_INTERVAL_END)
> +				ret = f->mt[b].e;
> +			else
> +				ret = f->mt[b].e->start;
> +			goto out;
> +		}
> +
> +		data += NFT_PIPAPO_GROUPS_PADDING(f->groups);
> +
> +		/* Swap bitmap indices: fill_map will be the initial bitmap for
> +		 * the next field (i.e. the new res_map), and res_map is
> +		 * guaranteed to be all-zeroes at this point, ready to be filled
> +		 * according to the next mapping table.
> +		 */
> +		swap(res_map, fill_map);
> +	}
> +
> +out:
> +	kfree(fill_map);
> +	kfree(res_map);
> +	return ret;
> +}
> +
> +/**
> + * nft_pipapo_get() - Get matching element reference given key data
> + * @net:	Network namespace
> + * @set:	nftables API set representation
> + * @elem:	nftables API element representation containing key data
> + * @flags:	If NFT_SET_ELEM_INTERVAL_END is passed, return the end element
> + */
> +static void *nft_pipapo_get(const struct net *net, const struct nft_set *set,
> +			    const struct nft_set_elem *elem, unsigned int flags)
> +{
> +	return pipapo_get(net, set, (const u8 *)elem->key.val.data, flags);
> +}
> +
> +/**
> + * pipapo_resize() - Resize lookup or mapping table, or both
> + * @f:		Field containing lookup and mapping tables
> + * @old_rules:	Previous amount of rules in field
> + * @rules:	New amount of rules
> + *
> + * Increase, decrease or maintain tables size depending on new amount of rules,
> + * and copy data over. In case the new size is smaller, throw away data for
> + * highest-numbered rules.
> + *
> + * Return: 0 on success, -ENOMEM on allocation failure.
> + */
> +static int pipapo_resize(struct nft_pipapo_field *f, int old_rules, int rules)
> +{
> +	long *new_lt = NULL, *new_p, *old_lt = f->lt, *old_p;
> +	union nft_pipapo_map_bucket *new_mt, *old_mt = f->mt;
> +	size_t new_bucket_size, copy;
> +	int group, bucket;
> +
> +	new_bucket_size = DIV_ROUND_UP(rules, BITS_PER_LONG);
> +
> +	if (new_bucket_size == f->bsize)
> +		goto mt;
> +
> +	if (new_bucket_size > f->bsize)
> +		copy = f->bsize;
> +	else
> +		copy = new_bucket_size;
> +
> +	new_lt = kvzalloc(f->groups * NFT_PIPAPO_BUCKETS * new_bucket_size *
> +			  sizeof(*new_lt), GFP_KERNEL);
> +	if (!new_lt)
> +		return -ENOMEM;
> +
> +	new_p = new_lt;
> +	old_p = old_lt;
> +	for (group = 0; group < f->groups; group++) {
> +		for (bucket = 0; bucket < NFT_PIPAPO_BUCKETS; bucket++) {
> +			memcpy(new_p, old_p, copy * sizeof(*new_p));
> +			new_p += copy;
> +			old_p += copy;
> +
> +			if (new_bucket_size > f->bsize)
> +				new_p += new_bucket_size - f->bsize;
> +			else
> +				old_p += f->bsize - new_bucket_size;
> +		}
> +	}
> +
> +mt:
> +	new_mt = kvmalloc(rules * sizeof(*new_mt), GFP_KERNEL);
> +	if (!new_mt) {
> +		kvfree(new_lt);
> +		return -ENOMEM;
> +	}
> +
> +	memcpy(new_mt, f->mt, min(old_rules, rules) * sizeof(*new_mt));
> +	if (rules > old_rules) {
> +		memset(new_mt + old_rules, 0,
> +		       (rules - old_rules) * sizeof(*new_mt));
> +	}
> +
> +	if (new_lt) {
> +		f->bsize = new_bucket_size;
> +		f->lt = new_lt;
> +		kvfree(old_lt);
> +	}
> +
> +	f->mt = new_mt;
> +	kvfree(old_mt);
> +
> +	return 0;
> +}
> +
> +/**
> + * pipapo_bucket_set() - Set rule bit in bucket given group and group value
> + * @f:		Field containing lookup table
> + * @rule:	Rule index
> + * @group:	Group index
> + * @v:		Value of bit group
> + */
> +static void pipapo_bucket_set(struct nft_pipapo_field *f, int rule, int group,
> +			      int v)
> +{
> +	unsigned long *pos;
> +
> +	pos = f->lt + f->bsize * NFT_PIPAPO_BUCKETS * group;
> +	pos += f->bsize * v;
> +
> +	__set_bit(rule, pos);
> +}
> +
> +/**
> + * pipapo_insert() - Insert new rule in field given input key and mask length
> + * @f:		Field containing lookup table
> + * @k:		Input key for classification, without nftables padding
> + * @mask_bits:	Length of mask; matches field length for non-ranged entry
> + *
> + * Insert a new rule reference in lookup buckets corresponding to k and
> + * mask_bits.
> + *
> + * Return: 1 on success (one rule inserted), negative error code on failure.
> + */
> +static int pipapo_insert(struct nft_pipapo_field *f, const uint8_t *k,
> +			 int mask_bits)
> +{
> +	int rule = f->rules++, group, ret;
> +
> +	ret = pipapo_resize(f, f->rules - 1, f->rules);
> +	if (ret)
> +		return ret;
> +
> +	for (group = 0; group < f->groups; group++) {
> +		int i, v;
> +		u8 mask;
> +
> +		if (group % 2)
> +			v = k[group / 2] & 0x0f;
> +		else
> +			v = k[group / 2] >> 4;
> +
> +		if (mask_bits >= (group + 1) * 4) {
> +			/* Not masked */
> +			pipapo_bucket_set(f, rule, group, v);
> +		} else if (mask_bits <= group * 4) {
> +			/* Completely masked */
> +			for (i = 0; i < NFT_PIPAPO_BUCKETS; i++)
> +				pipapo_bucket_set(f, rule, group, i);
> +		} else {
> +			/* The mask limit falls on this group */
> +			mask = 0x0f >> (mask_bits - group * 4);
> +			for (i = 0; i < NFT_PIPAPO_BUCKETS; i++) {
> +				if ((i & ~mask) == (v & ~mask))
> +					pipapo_bucket_set(f, rule, group, i);
> +			}
> +		}
> +	}
> +
> +	return 1;
> +}
> +
> +/**
> + * pipapo_step_diff() - Check if setting @step bit in netmask would change it
> + * @base:	Mask we are expanding
> + * @step:	Step bit for given expansion step
> + * @len:	Total length of mask space (set and unset bits), bytes
> + *
> + * Convenience function for mask expansion.
> + *
> + * Return: true if step bit changes mask (i.e. isn't set), false otherwise.
> + */
> +static bool pipapo_step_diff(u8 *base, int step, int len)
> +{
> +	/* Network order, byte-addressed */
> +#ifdef __BIG_ENDIAN__
> +	return !(BIT(step % BITS_PER_BYTE) & base[step / BITS_PER_BYTE]);
> +#else
> +	return !(BIT(step % BITS_PER_BYTE) &
> +		 base[len - 1 - step / BITS_PER_BYTE]);
> +#endif
> +}
> +
> +/**
> + * pipapo_step_after_end() - Check if mask exceeds range end with given step
> + * @base:	Mask we are expanding
> + * @end:	End of range
> + * @step:	Step bit for given expansion step, highest bit to be set
> + * @len:	Total length of mask space (set and unset bits), bytes
> + *
> + * Convenience function for mask expansion.
> + *
> + * Return: true if mask exceeds range setting step bits, false otherwise.
> + */
> +static bool pipapo_step_after_end(const u8 *base, const u8 *end, int step,
> +				  int len)
> +{
> +	u8 tmp[NFT_PIPAPO_MAX_BYTES];
> +	int i;
> +
> +	memcpy(tmp, base, len);
> +
> +	/* Network order, byte-addressed */
> +	for (i = 0; i <= step; i++)
> +#ifdef __BIG_ENDIAN__
> +		tmp[i / BITS_PER_BYTE] |= BIT(i % BITS_PER_BYTE);
> +#else
> +		tmp[len - 1 - i / BITS_PER_BYTE] |= BIT(i % BITS_PER_BYTE);
> +#endif
> +
> +	return memcmp(tmp, end, len) > 0;
> +}
> +
> +/**
> + * pipapo_base_sum() - Sum step bit to given len-sized netmask base with carry
> + * @base:	Netmask base
> + * @step:	Step bit to sum
> + * @len:	Netmask length, bytes
> + */
> +static void pipapo_base_sum(u8 *base, int step, int len)
> +{
> +	bool carry = false;
> +	int i;
> +
> +	/* Network order, byte-addressed */
> +#ifdef __BIG_ENDIAN__
> +	for (i = step / BITS_PER_BYTE; i < len; i++) {
> +#else
> +	for (i = len - 1 - step / BITS_PER_BYTE; i >= 0; i--) {
> +#endif
> +		if (carry)
> +			base[i]++;
> +		else
> +			base[i] += 1 << (step % BITS_PER_BYTE);
> +
> +		if (base[i])
> +			break;
> +
> +		carry = true;
> +	}
> +}
> +
> +/**
> + * expand() - Expand range to composing netmasks and insert into lookup table
> + * @f:		Field containing lookup table
> + * @start:	Start of range
> + * @end:	End of range
> + * @len:	Length of value in bits
> + *
> + * Expand range to composing netmasks and insert corresponding rule references
> + * in lookup buckets.
> + *
> + * Return: number of inserted rules on success, negative error code on failure.
> + */
> +static int pipapo_expand(struct nft_pipapo_field *f,
> +			 const u8 *start, const u8 *end, int len)
> +{
> +	int step, masks = 0, bytes = DIV_ROUND_UP(len, BITS_PER_BYTE);
> +	u8 base[NFT_PIPAPO_MAX_BYTES];
> +
> +	memcpy(base, start, bytes);
> +	while (memcmp(base, end, bytes) <= 0) {
> +		int err;
> +
> +		step = 0;
> +		while (pipapo_step_diff(base, step, bytes)) {
> +			if (pipapo_step_after_end(base, end, step, bytes))
> +				break;
> +
> +			step++;
> +			if (step >= len) {
> +				if (!masks) {
> +					pipapo_insert(f, base, 0);
> +					masks = 1;
> +				}
> +				goto out;
> +			}
> +		}
> +
> +		err = pipapo_insert(f, base, len - step);
> +
> +		if (err < 0)
> +			return err;
> +
> +		masks++;
> +		pipapo_base_sum(base, step, bytes);
> +	}
> +out:
> +	return masks;
> +}
> +
> +/**
> + * pipapo_map() - Insert rules in mapping tables, mapping them between fields
> + * @m:		Matching data, including mapping table
> + * @map:	Table of rule maps: array of first rule and amount of rules
> + *		in next field a given rule maps to, for each field
> + * @ext:	For last field, nft_set_ext pointer matching rules map to
> + */
> +static void pipapo_map(struct nft_pipapo_match *m,
> +		       union nft_pipapo_map_bucket map[NFT_PIPAPO_MAX_FIELDS],
> +		       struct nft_pipapo_elem *e)
> +{
> +	struct nft_pipapo_field *f;
> +	int i, j;
> +
> +	for (i = 0, f = m->f; i < m->field_count - 1; i++, f++) {
> +		for (j = 0; j < map[i].n; j++) {
> +			f->mt[map[i].to + j].to = map[i + 1].to;
> +			f->mt[map[i].to + j].n = map[i + 1].n;
> +		}
> +	}
> +
> +	/* Last field: map to ext instead of mapping to next field */
> +	for (j = 0; j < map[i].n; j++)
> +		f->mt[map[i].to + j].e = e;
> +}
> +
> +/**
> + * pipapo_realloc_scratch() - Reallocate scratch maps for partial match results
> + * @clone:	Copy of matching data with pending insertions and deletions
> + * @bsize_max	Maximum bucket size, scratch maps cover two buckets
> + *
> + * Return: 0 on success, -ENOMEM on failure.
> + */
> +static int pipapo_realloc_scratch(struct nft_pipapo_match *clone,
> +				  unsigned long bsize_max)
> +{
> +	int i;
> +
> +	for_each_possible_cpu(i) {
> +		unsigned long *scratch;
> +
> +		scratch = kzalloc_node(bsize_max * sizeof(*scratch) * 2,
> +				       GFP_KERNEL, cpu_to_node(i));
> +		if (!scratch) {
> +			/* On failure, there's no need to undo previous
> +			 * allocations: this means that some scratch maps have
> +			 * a bigger allocated size now (this is only called on
> +			 * insertion), but the extra space won't be used by any
> +			 * CPU as new elements are not inserted and m->bsize_max
> +			 * is not updated.
> +			 */
> +			return -ENOMEM;
> +		}
> +
> +		kfree(*per_cpu_ptr(clone->scratch, i));
> +
> +		*per_cpu_ptr(clone->scratch, i) = scratch;
> +	}
> +
> +	return 0;
> +}
> +
> +/**
> + * nft_pipapo_insert() - Validate and insert ranged elements
> + * @net:	Network namespace
> + * @set:	nftables API set representation
> + * @elem:	nftables API element representation containing key data
> + * @flags:	If NFT_SET_ELEM_INTERVAL_END is passed, this is the end element
> + * @ext2:	Filled with pointer to &struct nft_set_ext in inserted element
> + *
> + * In this set implementation, this functions needs to be called twice, with
> + * start and end element, to obtain a valid entry insertion.
> + *
> + * Calls to this function are serialised with each other, so we can store
> + * element and key data on the first call with start element, and use it on the
> + * second call once we get the end element too.
> + *
> + * However, userspace could send a single NFT_SET_ELEM_INTERVAL_END element,
> + * without a start element, so we need to check for it explicitly before
> + * inserting an entry, lest we end up in nft_pipapo_walk() with an empty start
> + * element.
> + *
> + * Also, we need to make sure that the start element hasn't been deactivated or
> + * destroyed between the two calls to this function, otherwise we might link an
> + * invalid start item to the end item triggering the insertion. Clear
> + * priv->start_elem on any operation that might render it invalid.
> + *
> + * Return: 0 on success, error pointer on failure.
> + */
> +static int nft_pipapo_insert(const struct net *net, const struct nft_set *set,
> +			     const struct nft_set_elem *elem,
> +			     struct nft_set_ext **ext2)
> +{
> +	const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
> +	const u8 *data = (const u8 *)elem->key.val.data, *start, *end;
> +	union nft_pipapo_map_bucket rulemap[NFT_PIPAPO_MAX_FIELDS];
> +	struct nft_pipapo *priv = nft_set_priv(set);
> +	struct nft_pipapo_match *m = priv->clone;
> +	struct nft_pipapo_elem *e = elem->priv;
> +	struct nft_pipapo_field *f;
> +	int i, bsize_max, err = 0;
> +	void *dup;
> +
> +	dup = nft_pipapo_get(net, set, elem, 0);
> +	if (PTR_ERR(dup) != -ENOENT) {
> +		priv->start_elem = NULL;
> +		if (IS_ERR(dup))
> +			return PTR_ERR(dup);
> +		*ext2 = dup;

dup should be of nft_set_ext type. I just had a look at
nft_pipapo_get() and I think this returns nft_pipapo_elem, which is
almost good, since it contains nft_set_ext, right?

I think you also need to check if the object is active in the next
generation via nft_genmask_next() and nft_set_elem_active(), otherwise
ignore it.

Note that the datastructure needs to temporarily deal with duplicates,
ie. one inactive object (just deleted) and one active object (just
added) for the next generation.

> +		return -EEXIST;
> +	}
> +
> +	if (!nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) ||
> +	    !(*nft_set_ext_flags(ext) & NFT_SET_ELEM_INTERVAL_END)) {
> +		priv->start_elem = e;
> +		*ext2 = &e->ext;
> +		memcpy(priv->start_data, data, priv->width);
> +		return 0;
> +	}
> +
> +	if (!priv->start_elem)
> +		return -EINVAL;

I'm working on a sketch patch to extend the front-end code to make
this easier for you, will post it asap, so you don't need this special
handling to collect both ends of the interval.

So far, just spend a bit of time on this, will get back to you with
more feedback.

Thanks for working on this!
Stefano Brivio Nov. 27, 2019, 11:02 a.m. UTC | #2
On Wed, 27 Nov 2019 10:29:45 +0100
Pablo Neira Ayuso <pablo@netfilter.org> wrote:

> Hi Stefano,
> 
> Just started reading, a few initial questions.
> 
> On Fri, Nov 22, 2019 at 02:40:02PM +0100, Stefano Brivio wrote:
> [...]
>
> > +static int nft_pipapo_insert(const struct net *net, const struct nft_set *set,
> > +			     const struct nft_set_elem *elem,
> > +			     struct nft_set_ext **ext2)
> > +{
> > +	const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
> > +	const u8 *data = (const u8 *)elem->key.val.data, *start, *end;
> > +	union nft_pipapo_map_bucket rulemap[NFT_PIPAPO_MAX_FIELDS];
> > +	struct nft_pipapo *priv = nft_set_priv(set);
> > +	struct nft_pipapo_match *m = priv->clone;
> > +	struct nft_pipapo_elem *e = elem->priv;
> > +	struct nft_pipapo_field *f;
> > +	int i, bsize_max, err = 0;
> > +	void *dup;
> > +
> > +	dup = nft_pipapo_get(net, set, elem, 0);
> > +	if (PTR_ERR(dup) != -ENOENT) {
> > +		priv->start_elem = NULL;
> > +		if (IS_ERR(dup))
> > +			return PTR_ERR(dup);
> > +		*ext2 = dup;  
> 
> dup should be of nft_set_ext type. I just had a look at
> nft_pipapo_get() and I think this returns nft_pipapo_elem, which is
> almost good, since it contains nft_set_ext, right?

Right, it returns nft_pipapo_elem which contains that.

> I think you also need to check if the object is active in the next
> generation via nft_genmask_next() and nft_set_elem_active(), otherwise
> ignore it.

I guess I should actually do this in nft_pipapo_get(), also because we
don't want to return inactive elements when userspace "gets" them.

I just noticed this is currently inconsistent with the lookup, because
nft_pipapo_lookup() correctly does:

--
next_match:
		b = pipapo_refill(res_map, f->bsize, f->rules, fill_map, f->mt,
				  last);
		if (b < 0) {
			raw_cpu_write(nft_pipapo_scratch_index, map_index);
			local_bh_enable();

			return false;
		}

		if (last) {
			*ext = &f->mt[b].e->ext;
			if (unlikely(nft_set_elem_expired(*ext) ||
				     !nft_set_elem_active(*ext, genmask)))
				goto next_match;
--

but I forgot to implement the same check in pipapo_get():

--
next_match:
		b = pipapo_refill(res_map, f->bsize, f->rules, fill_map, f->mt,
				  last);
		if (b < 0)
			goto out;

		if (last) {
			if (nft_set_elem_expired(&f->mt[b].e->ext))
				goto next_match;
--

this check should simply include || !nft_set_elem_active(...), and then
I wouldn't need any further check in nft_pipapo_init(). I'd fix this in
v3.

I'm actually not sure if I need to report these elements to
nft_pipapo_remove(). If it's needed, I would add some kind of
"get_inactive" flag to pipapo_get(), which is true on the call from
nft_pipapo_remove(), and false on other paths. If the flag is true, the
nft_set_elem_active() check is then skipped.

> Note that the datastructure needs to temporarily deal with duplicates,
> ie. one inactive object (just deleted) and one active object (just
> added) for the next generation.

Yes, this is taken care of (except for the problem described above),
specifically, there can be n inactive objects, and a single active
object that are entirely overlapping.

This makes some optimisations harder to implement, namely, step 5.2.1
from:
	https://pipapo.lameexcu.se/pipapo/tree/pipapo.c#n337

because we need to allow entirely overlapping entries and map them to
possibly distinct elements.

Now, I think this would all be easier if the API implemented
transactions and commit in a way that appears more natural to me.

When I started working on this, I initially thought activate() would be
called once per transaction, not per element, so that insert() and
remove() would add or remove elements pending for a given transaction,
and activate() would commit it. Same for flush().

At that point, we would have a copy of lookup data with pending
insertions and without pending deletions, and on transaction commit,
this copy would become active, with no inactive elements into it.
Hence, no overlapping elements in live data.

This way we could also make transactions atomic. If activate() is
called once for each element in the transaction, that can't be atomic.

I plan to work on this (if it makes sense), but it looks rather
complicated to match this with existing set implementations and
especially current UAPI, that's the main reason why I "worked around"
this aspect for the moment being. I guess that having at least one set
implementation that can play along with this model would help later.

> > +		return -EEXIST;
> > +	}
> > +
> > +	if (!nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) ||
> > +	    !(*nft_set_ext_flags(ext) & NFT_SET_ELEM_INTERVAL_END)) {
> > +		priv->start_elem = e;
> > +		*ext2 = &e->ext;
> > +		memcpy(priv->start_data, data, priv->width);
> > +		return 0;
> > +	}
> > +
> > +	if (!priv->start_elem)
> > +		return -EINVAL;  
> 
> I'm working on a sketch patch to extend the front-end code to make
> this easier for you, will post it asap, so you don't need this special
> handling to collect both ends of the interval.

Nice, thanks. Mind that I think this is actually a bit ugly but fine.
As I was mentioning to Florian, it doesn't present any particular race
with bad consequences (at least in v2).

Right now I was trying to get the NFTA_SET_DESC_CONCAT >
NFTA_LIST_ELEM > NFTA_SET_FIELD_LEN nesting implemented in libnftnl in
a somewhat acceptable way. Let me know if the front-end changes would
affect this significantly, I'll wait for your patch in that case.

> So far, just spend a bit of time on this, will get back to you with
> more feedback.
> 
> Thanks for working on this!

And thanks for reviewing it!
Pablo Neira Ayuso Nov. 27, 2019, 6:29 p.m. UTC | #3
Hi Stefano,

On Wed, Nov 27, 2019 at 12:02:49PM +0100, Stefano Brivio wrote:
> On Wed, 27 Nov 2019 10:29:45 +0100
> Pablo Neira Ayuso <pablo@netfilter.org> wrote:
[...]
> > On Fri, Nov 22, 2019 at 02:40:02PM +0100, Stefano Brivio wrote:
[...]
> > I think you also need to check if the object is active in the next
> > generation via nft_genmask_next() and nft_set_elem_active(), otherwise
> > ignore it.
> 
> I guess I should actually do this in nft_pipapo_get(), also because we
> don't want to return inactive elements when userspace "gets" them.

OK. Just a side note: nft_pipapo_get() is also used to get an
interval, that needs current generation. From the insert path, one
need to check the next generation.

[...]
> > > +		return -EEXIST;
> > > +	}
> > > +
> > > +	if (!nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) ||
> > > +	    !(*nft_set_ext_flags(ext) & NFT_SET_ELEM_INTERVAL_END)) {
> > > +		priv->start_elem = e;
> > > +		*ext2 = &e->ext;
> > > +		memcpy(priv->start_data, data, priv->width);
> > > +		return 0;
> > > +	}
> > > +
> > > +	if (!priv->start_elem)
> > > +		return -EINVAL;  
> > 
> > I'm working on a sketch patch to extend the front-end code to make
> > this easier for you, will post it asap, so you don't need this special
> > handling to collect both ends of the interval.
> 
> Nice, thanks. Mind that I think this is actually a bit ugly but fine.
> As I was mentioning to Florian, it doesn't present any particular race
> with bad consequences (at least in v2).
> 
> Right now I was trying to get the NFTA_SET_DESC_CONCAT >
> NFTA_LIST_ELEM > NFTA_SET_FIELD_LEN nesting implemented in libnftnl in
> a somewhat acceptable way. Let me know if the front-end changes would
> affect this significantly, I'll wait for your patch in that case.

I'm attaching a sketch patch, I need a bit more time to finish it. The
idea is to place the interval end in the same element, instead of two.
This should also simplify the rbtree implementation. Main issue is
that this needs a bit more work to make it backward compatible.

P.S: I'm skipping the transaction discussion in this email, will come
back to it later.

Patch
diff mbox series

diff --git a/include/net/netfilter/nf_tables_core.h b/include/net/netfilter/nf_tables_core.h
index 7281895fa6d9..9759257ec8ec 100644
--- a/include/net/netfilter/nf_tables_core.h
+++ b/include/net/netfilter/nf_tables_core.h
@@ -74,6 +74,7 @@  extern struct nft_set_type nft_set_hash_type;
 extern struct nft_set_type nft_set_hash_fast_type;
 extern struct nft_set_type nft_set_rbtree_type;
 extern struct nft_set_type nft_set_bitmap_type;
+extern struct nft_set_type nft_set_pipapo_type;
 
 struct nft_expr;
 struct nft_regs;
diff --git a/net/netfilter/Makefile b/net/netfilter/Makefile
index 5e9b2eb24349..3f572e5a975e 100644
--- a/net/netfilter/Makefile
+++ b/net/netfilter/Makefile
@@ -81,7 +81,8 @@  nf_tables-objs := nf_tables_core.o nf_tables_api.o nft_chain_filter.o \
 		  nft_chain_route.o nf_tables_offload.o
 
 nf_tables_set-objs := nf_tables_set_core.o \
-		      nft_set_hash.o nft_set_bitmap.o nft_set_rbtree.o
+		      nft_set_hash.o nft_set_bitmap.o nft_set_rbtree.o \
+		      nft_set_pipapo.o
 
 obj-$(CONFIG_NF_TABLES)		+= nf_tables.o
 obj-$(CONFIG_NF_TABLES_SET)	+= nf_tables_set.o
diff --git a/net/netfilter/nf_tables_set_core.c b/net/netfilter/nf_tables_set_core.c
index a9fce8d10051..586b621007eb 100644
--- a/net/netfilter/nf_tables_set_core.c
+++ b/net/netfilter/nf_tables_set_core.c
@@ -9,12 +9,14 @@  static int __init nf_tables_set_module_init(void)
 	nft_register_set(&nft_set_rhash_type);
 	nft_register_set(&nft_set_bitmap_type);
 	nft_register_set(&nft_set_rbtree_type);
+	nft_register_set(&nft_set_pipapo_type);
 
 	return 0;
 }
 
 static void __exit nf_tables_set_module_exit(void)
 {
+	nft_unregister_set(&nft_set_pipapo_type);
 	nft_unregister_set(&nft_set_rbtree_type);
 	nft_unregister_set(&nft_set_bitmap_type);
 	nft_unregister_set(&nft_set_rhash_type);
diff --git a/net/netfilter/nft_set_pipapo.c b/net/netfilter/nft_set_pipapo.c
new file mode 100644
index 000000000000..3cad9aedc168
--- /dev/null
+++ b/net/netfilter/nft_set_pipapo.c
@@ -0,0 +1,2197 @@ 
+// SPDX-License-Identifier: GPL-2.0-only
+
+/* PIPAPO: PIle PAcket POlicies: set for arbitrary concatenations of ranges
+ *
+ * Copyright (c) 2019 Red Hat GmbH
+ *
+ * Author: Stefano Brivio <sbrivio@redhat.com>
+ */
+
+/**
+ * DOC: Theory of Operation
+ *
+ *
+ * Problem
+ * -------
+ *
+ * Match packet bytes against entries composed of ranged or non-ranged packet
+ * field specifiers, mapping them to arbitrary references. For example:
+ *
+ * ::
+ *
+ *               --- fields --->
+ *      |    [net],[port],[net]... => [reference]
+ *   entries [net],[port],[net]... => [reference]
+ *      |    [net],[port],[net]... => [reference]
+ *      V    ...
+ *
+ * where [net] fields can be IP ranges or netmasks, and [port] fields are port
+ * ranges. Arbitrary packet fields can be matched.
+ *
+ *
+ * Algorithm Overview
+ * ------------------
+ *
+ * This algorithm is loosely inspired by [Ligatti 2010], and fundamentally
+ * relies on the consideration that every contiguous range in a space of b bits
+ * can be converted into b * 2 netmasks, from Theorem 3 in [Rottenstreich 2010],
+ * as also illustrated in Section 9 of [Kogan 2014].
+ *
+ * Classification against a number of entries, that require matching given bits
+ * of a packet field, is performed by grouping those bits in sets of arbitrary
+ * size, and classifying packet bits one group at a time.
+ *
+ * Example:
+ *   to match the source port (16 bits) of a packet, we can divide those 16 bits
+ *   in 4 groups of 4 bits each. Given the entry:
+ *      0000 0001 0101 1001
+ *   and a packet with source port:
+ *      0000 0001 1010 1001
+ *   first and second groups match, but the third doesn't. We conclude that the
+ *   packet doesn't match the given entry.
+ *
+ * Translate the set to a sequence of lookup tables, one per field. Each table
+ * has two dimensions: bit groups to be matched for a single packet field, and
+ * all the possible values of said groups (buckets). Input entries are
+ * represented as one or more rules, depending on the number of composing
+ * netmasks for the given field specifier, and a group match is indicated as a
+ * set bit, with number corresponding to the rule index, in all the buckets
+ * whose value matches the entry for a given group.
+ *
+ * Rules are mapped between fields through an array of x, n pairs, with each
+ * item mapping a matched rule to one or more rules. The position of the pair in
+ * the array indicates the matched rule to be mapped to the next field, x
+ * indicates the first rule index in the next field, and n the amount of
+ * next-field rules the current rule maps to.
+ *
+ * The mapping array for the last field maps to the desired references.
+ *
+ * To match, we perform table lookups using the values of grouped packet bits,
+ * and use a sequence of bitwise operations to progressively evaluate rule
+ * matching.
+ *
+ * A stand-alone, reference implementation, also including notes about possible
+ * future optimisations, is available at:
+ *    https://pipapo.lameexcu.se/
+ *
+ * Insertion
+ * ---------
+ *
+ * - For each packet field:
+ *
+ *   - divide the b packet bits we want to classify into groups of size t,
+ *     obtaining ceil(b / t) groups
+ *
+ *      Example: match on destination IP address, with t = 4: 32 bits, 8 groups
+ *      of 4 bits each
+ *
+ *   - allocate a lookup table with one column ("bucket") for each possible
+ *     value of a group, and with one row for each group
+ *
+ *      Example: 8 groups, 2^4 buckets:
+ *
+ * ::
+ *
+ *                     bucket
+ *      group  0   1   2   3   4   5   6   7   8   9  10  11  12  13  14  15
+ *        0
+ *        1
+ *        2
+ *        3
+ *        4
+ *        5
+ *        6
+ *        7
+ *
+ *   - map the bits we want to classify for the current field, for a given
+ *     entry, to a single rule for non-ranged and netmask set items, and to one
+ *     or multiple rules for ranges. Ranges are expanded to composing netmasks
+ *     by pipapo_expand().
+ *
+ *      Example: 2 entries, 10.0.0.5:1024 and 192.168.1.0-192.168.2.1:2048
+ *      - rule #0: 10.0.0.5
+ *      - rule #1: 192.168.1.0/24
+ *      - rule #2: 192.168.2.0/31
+ *
+ *   - insert references to the rules in the lookup table, selecting buckets
+ *     according to bit values of a rule in the given group. This is done by
+ *     pipapo_insert().
+ *
+ *      Example: given:
+ *      - rule #0: 10.0.0.5 mapping to buckets
+ *        < 0 10  0 0   0 0  0 5 >
+ *      - rule #1: 192.168.1.0/24 mapping to buckets
+ *        < 12 0  10 8  0 1  < 0..15 > < 0..15 > >
+ *      - rule #2: 192.168.2.0/31 mapping to buckets
+ *        < 12 0  10 8  0 2  0 < 0..1 > >
+ *
+ *      these bits are set in the lookup table:
+ *
+ * ::
+ *
+ *                     bucket
+ *      group  0   1   2   3   4   5   6   7   8   9  10  11  12  13  14  15
+ *        0    0                                              1,2
+ *        1   1,2                                      0
+ *        2    0                                      1,2
+ *        3    0                              1,2
+ *        4  0,1,2
+ *        5    0   1   2
+ *        6  0,1,2 1   1   1   1   1   1   1   1   1   1   1   1   1   1   1
+ *        7   1,2 1,2  1   1   1  0,1  1   1   1   1   1   1   1   1   1   1
+ *
+ *   - if this is not the last field in the set, fill a mapping array that maps
+ *     rules from the lookup table to rules belonging to the same entry in
+ *     the next lookup table, done by pipapo_map().
+ *
+ *     Note that as rules map to contiguous ranges of rules, given how netmask
+ *     expansion and insertion is performed, &union nft_pipapo_map_bucket stores
+ *     this information as pairs of first rule index, rule count.
+ *
+ *      Example: 2 entries, 10.0.0.5:1024 and 192.168.1.0-192.168.2.1:2048,
+ *      given lookup table #0 for field 0 (see example above):
+ *
+ * ::
+ *
+ *                     bucket
+ *      group  0   1   2   3   4   5   6   7   8   9  10  11  12  13  14  15
+ *        0    0                                              1,2
+ *        1   1,2                                      0
+ *        2    0                                      1,2
+ *        3    0                              1,2
+ *        4  0,1,2
+ *        5    0   1   2
+ *        6  0,1,2 1   1   1   1   1   1   1   1   1   1   1   1   1   1   1
+ *        7   1,2 1,2  1   1   1  0,1  1   1   1   1   1   1   1   1   1   1
+ *
+ *      and lookup table #1 for field 1 with:
+ *      - rule #0: 1024 mapping to buckets
+ *        < 0  0  4  0 >
+ *      - rule #1: 2048 mapping to buckets
+ *        < 0  0  5  0 >
+ *
+ * ::
+ *
+ *                     bucket
+ *      group  0   1   2   3   4   5   6   7   8   9  10  11  12  13  14  15
+ *        0   0,1
+ *        1   0,1
+ *        2                    0   1
+ *        3   0,1
+ *
+ *      we need to map rules for 10.0.0.5 in lookup table #0 (rule #0) to 1024
+ *      in lookup table #1 (rule #0) and rules for 192.168.1.0-192.168.2.1
+ *      (rules #1, #2) to 2048 in lookup table #2 (rule #1):
+ *
+ * ::
+ *
+ *       rule indices in current field: 0    1    2
+ *       map to rules in next field:    0    1    1
+ *
+ *   - if this is the last field in the set, fill a mapping array that maps
+ *     rules from the last lookup table to element pointers, also done by
+ *     pipapo_map().
+ *
+ *     Note that, in this implementation, we have two elements (start, end) for
+ *     each entry. The pointer to the end element is stored in this array, and
+ *     the pointer to the start element is linked from it.
+ *
+ *      Example: entry 10.0.0.5:1024 has a corresponding &struct nft_pipapo_elem
+ *      pointer, 0x66, and element for 192.168.1.0-192.168.2.1:2048 is at 0x42.
+ *      From the rules of lookup table #1 as mapped above:
+ *
+ * ::
+ *
+ *       rule indices in last field:    0    1
+ *       map to elements:             0x42  0x66
+ *
+ *
+ * Matching
+ * --------
+ *
+ * We use a result bitmap, with the size of a single lookup table bucket, to
+ * represent the matching state that applies at every algorithm step. This is
+ * done by pipapo_lookup().
+ *
+ * - For each packet field:
+ *
+ *   - start with an all-ones result bitmap (res_map in pipapo_lookup())
+ *
+ *   - perform a lookup into the table corresponding to the current field,
+ *     for each group, and at every group, AND the current result bitmap with
+ *     the value from the lookup table bucket
+ *
+ * ::
+ *
+ *      Example: 192.168.1.5 < 12 0  10 8  0 1  0 5 >, with lookup table from
+ *      insertion examples.
+ *      Lookup table buckets are at least 3 bits wide, we'll assume 8 bits for
+ *      convenience in this example. Initial result bitmap is 0xff, the steps
+ *      below show the value of the result bitmap after each group is processed:
+ *
+ *                     bucket
+ *      group  0   1   2   3   4   5   6   7   8   9  10  11  12  13  14  15
+ *        0    0                                              1,2
+ *        result bitmap is now: 0xff & 0x6 [bucket 12] = 0x6
+ *
+ *        1   1,2                                      0
+ *        result bitmap is now: 0x6 & 0x6 [bucket 0] = 0x6
+ *
+ *        2    0                                      1,2
+ *        result bitmap is now: 0x6 & 0x6 [bucket 10] = 0x6
+ *
+ *        3    0                              1,2
+ *        result bitmap is now: 0x6 & 0x6 [bucket 8] = 0x6
+ *
+ *        4  0,1,2
+ *        result bitmap is now: 0x6 & 0x7 [bucket 0] = 0x6
+ *
+ *        5    0   1   2
+ *        result bitmap is now: 0x6 & 0x2 [bucket 1] = 0x2
+ *
+ *        6  0,1,2 1   1   1   1   1   1   1   1   1   1   1   1   1   1   1
+ *        result bitmap is now: 0x2 & 0x7 [bucket 0] = 0x2
+ *
+ *        7   1,2 1,2  1   1   1  0,1  1   1   1   1   1   1   1   1   1   1
+ *        final result bitmap for this field is: 0x2 & 0x3 [bucket 5] = 0x2
+ *
+ *   - at the next field, start with a new, all-zeroes result bitmap. For each
+ *     bit set in the previous result bitmap, fill the new result bitmap
+ *     (fill_map in pipapo_lookup()) with the rule indices from the
+ *     corresponding buckets of the mapping field for this field, done by
+ *     pipapo_refill()
+ *
+ *      Example: with mapping table from insertion examples, with the current
+ *      result bitmap from the previous example, 0x02:
+ *
+ * ::
+ *
+ *       rule indices in current field: 0    1    2
+ *       map to rules in next field:    0    1    1
+ *
+ *      the new result bitmap will be 0x02: rule 1 was set, and rule 1 will be
+ *      set.
+ *
+ *      We can now extend this example to cover the second iteration of the step
+ *      above (lookup and AND bitmap): assuming the port field is
+ *      2048 < 0  0  5  0 >, with starting result bitmap 0x2, and lookup table
+ *      for "port" field from pre-computation example:
+ *
+ * ::
+ *
+ *                     bucket
+ *      group  0   1   2   3   4   5   6   7   8   9  10  11  12  13  14  15
+ *        0   0,1
+ *        1   0,1
+ *        2                    0   1
+ *        3   0,1
+ *
+ *       operations are: 0x2 & 0x3 [bucket 0] & 0x3 [bucket 0] & 0x2 [bucket 5]
+ *       & 0x3 [bucket 0], resulting bitmap is 0x2.
+ *
+ *   - if this is the last field in the set, look up the value from the mapping
+ *     array corresponding to the final result bitmap
+ *
+ *      Example: 0x2 resulting bitmap from 192.168.1.5:2048, mapping array for
+ *      last field from insertion example:
+ *
+ * ::
+ *
+ *       rule indices in last field:    0    1
+ *       map to elements:             0x42  0x66
+ *
+ *      the matching element is at 0x42.
+ *
+ *
+ * References
+ * ----------
+ *
+ * [Ligatti 2010]
+ *      A Packet-classification Algorithm for Arbitrary Bitmask Rules, with
+ *      Automatic Time-space Tradeoffs
+ *      Jay Ligatti, Josh Kuhn, and Chris Gage.
+ *      Proceedings of the IEEE International Conference on Computer
+ *      Communication Networks (ICCCN), August 2010.
+ *      http://www.cse.usf.edu/~ligatti/papers/grouper-conf.pdf
+ *
+ * [Rottenstreich 2010]
+ *      Worst-Case TCAM Rule Expansion
+ *      Ori Rottenstreich and Isaac Keslassy.
+ *      2010 Proceedings IEEE INFOCOM, San Diego, CA, 2010.
+ *      http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.212.4592&rep=rep1&type=pdf
+ *
+ * [Kogan 2014]
+ *      SAX-PAC (Scalable And eXpressive PAcket Classification)
+ *      Kirill Kogan, Sergey Nikolenko, Ori Rottenstreich, William Culhane,
+ *      and Patrick Eugster.
+ *      Proceedings of the 2014 ACM conference on SIGCOMM, August 2014.
+ *      http://www.sigcomm.org/sites/default/files/ccr/papers/2014/August/2619239-2626294.pdf
+ */
+
+#include <linux/kernel.h>
+#include <linux/init.h>
+#include <linux/log2.h>
+#include <linux/module.h>
+#include <linux/netlink.h>
+#include <linux/netfilter.h>
+#include <linux/netfilter/nf_tables.h>
+#include <net/netfilter/nf_tables_core.h>
+#include <uapi/linux/netfilter/nf_tables.h>
+#include <net/ipv6.h>			/* For the maximum length of a field */
+#include <linux/bitmap.h>
+#include <linux/bitops.h>
+
+/* Count of concatenated fields depends on count of 32-bit nftables registers */
+#define NFT_PIPAPO_MAX_FIELDS		NFT_REG32_COUNT
+
+/* Largest supported field size */
+#define NFT_PIPAPO_MAX_BYTES		(sizeof(struct in6_addr))
+#define NFT_PIPAPO_MAX_BITS		(NFT_PIPAPO_MAX_BYTES * BITS_PER_BYTE)
+
+/* Number of bits to be grouped together in lookup table buckets, arbitrary */
+#define NFT_PIPAPO_GROUP_BITS		4
+#define NFT_PIPAPO_GROUPS_PER_BYTE	(BITS_PER_BYTE / NFT_PIPAPO_GROUP_BITS)
+
+/* Fields are padded to 32 bits in input registers */
+#define NFT_PIPAPO_GROUPS_PADDED_SIZE(x)				\
+	(round_up((x) / NFT_PIPAPO_GROUPS_PER_BYTE, sizeof(u32)))
+#define NFT_PIPAPO_GROUPS_PADDING(x)					\
+	(NFT_PIPAPO_GROUPS_PADDED_SIZE((x)) - (x) / NFT_PIPAPO_GROUPS_PER_BYTE)
+
+/* Number of buckets, given by 2 ^ n, with n grouped bits */
+#define NFT_PIPAPO_BUCKETS		(1 << NFT_PIPAPO_GROUP_BITS)
+
+/* Each n-bit range maps to up to n * 2 rules */
+#define NFT_PIPAPO_MAP_NBITS		(const_ilog2(NFT_PIPAPO_MAX_BITS * 2))
+
+/* Use the rest of mapping table buckets for rule indices, but it makes no sense
+ * to exceed 32 bits
+ */
+#if BITS_PER_LONG == 64
+#define NFT_PIPAPO_MAP_TOBITS		32
+#else
+#define NFT_PIPAPO_MAP_TOBITS		(BITS_PER_LONG - NFT_PIPAPO_MAP_NBITS)
+#endif
+
+/* ...which gives us the highest allowed index for a rule */
+#define NFT_PIPAPO_RULE0_MAX		((1UL << (NFT_PIPAPO_MAP_TOBITS - 1)) \
+					- (1UL << NFT_PIPAPO_MAP_NBITS))
+
+#define nft_pipapo_for_each_field(field, index, match)		\
+	for ((field) = (match)->f, (index) = 0;			\
+	     (index) < (match)->field_count;			\
+	     (index)++, (field)++)
+
+/**
+ * union nft_pipapo_map_bucket - Bucket of mapping table
+ * @to:		First rule number (in next field) this rule maps to
+ * @n:		Number of rules (in next field) this rule maps to
+ * @e:		If there's no next field, pointer to element this rule maps to
+ */
+union nft_pipapo_map_bucket {
+	struct {
+#if BITS_PER_LONG == 64
+		static_assert(NFT_PIPAPO_MAP_TOBITS <= 32);
+		u32 to;
+
+		static_assert(NFT_PIPAPO_MAP_NBITS <= 32);
+		u32 n;
+#else
+		unsigned long to:NFT_PIPAPO_MAP_TOBITS;
+		unsigned long  n:NFT_PIPAPO_MAP_NBITS;
+#endif
+	};
+	struct nft_pipapo_elem *e;
+};
+
+/**
+ * struct nft_pipapo_field - Lookup, mapping tables and related data for a field
+ * @groups:	Amount of 4-bit groups
+ * @rules:	Number of inserted rules
+ * @bsize:	Size of each bucket in lookup table, in longs
+ * @lt:		Lookup table: 'groups' rows of NFT_PIPAPO_BUCKETS buckets
+ * @mt:		Mapping table: one bucket per rule
+ */
+struct nft_pipapo_field {
+	int groups;
+	unsigned long rules;
+	size_t bsize;
+	unsigned long *lt;
+	union nft_pipapo_map_bucket *mt;
+};
+
+/**
+ * struct nft_pipapo_match - Data used for lookup and matching
+ * @field_count		Amount of fields in set
+ * @scratch:		Preallocated per-CPU maps for partial matching results
+ * @bsize_max:		Maximum lookup table bucket size of all fields, in longs
+ * @rcu			Matching data is swapped on commits
+ * @f:			Fields, with lookup and mapping tables
+ */
+struct nft_pipapo_match {
+	int field_count;
+	unsigned long * __percpu *scratch;
+	size_t bsize_max;
+	struct rcu_head rcu;
+	struct nft_pipapo_field f[0];
+};
+
+/* Current working bitmap index, toggled between field matches */
+static DEFINE_PER_CPU(bool, nft_pipapo_scratch_index);
+
+/**
+ * struct nft_pipapo - Representation of a set
+ * @match:	Currently in-use matching data
+ * @clone:	Copy where pending insertions and deletions are kept
+ * @groups:	Total amount of 4-bit groups for fields in this set
+ * @width:	Total bytes to be matched for one packet, including padding
+ * @dirty:	Working copy has pending insertions or deletions
+ * @last_gc:	Timestamp of last garbage collection run, jiffies
+ * @start_data:	Key data of start element for insertion
+ * @start_elem:	Start element for insertion
+ */
+struct nft_pipapo {
+	struct nft_pipapo_match __rcu *match;
+	struct nft_pipapo_match *clone;
+	int groups;
+	int width;
+	bool dirty;
+	unsigned long last_gc;
+	u8 start_data[NFT_DATA_VALUE_MAXLEN * sizeof(u32)];
+	struct nft_pipapo_elem *start_elem;
+};
+
+struct nft_pipapo_elem;
+
+/**
+ * struct nft_pipapo_elem - API-facing representation of single set element
+ * @start:	Pointer to element that represents start of interval
+ * @ext:	nftables API extensions
+ */
+struct nft_pipapo_elem {
+	struct nft_pipapo_elem *start;
+	struct nft_set_ext ext;
+};
+
+/**
+ * pipapo_refill() - For each set bit, set bits from selected mapping table item
+ * @map:	Bitmap to be scanned for set bits
+ * @len:	Length of bitmap in longs
+ * @rules:	Number of rules in field
+ * @dst:	Destination bitmap
+ * @mt:		Mapping table containing bit set specifiers
+ * @match_only:	Find a single bit and return, don't fill
+ *
+ * Iteration over set bits with __builtin_ctzl(): Daniel Lemire, public domain.
+ *
+ * For each bit set in map, select the bucket from mapping table with index
+ * corresponding to the position of the bit set. Use start bit and amount of
+ * bits specified in bucket to fill region in dst.
+ *
+ * Return: -1 on no match, bit position on 'match_only', 0 otherwise.
+ */
+static int pipapo_refill(unsigned long *map, int len, int rules,
+			 unsigned long *dst, union nft_pipapo_map_bucket *mt,
+			 bool match_only)
+{
+	unsigned long bitset;
+	int k, ret = -1;
+
+	for (k = 0; k < len; k++) {
+		bitset = map[k];
+		while (bitset) {
+			unsigned long t = bitset & -bitset;
+			int r = __builtin_ctzl(bitset);
+			int i = k * BITS_PER_LONG + r;
+
+			if (unlikely(i >= rules)) {
+				map[k] = 0;
+				return -1;
+			}
+
+			if (unlikely(match_only)) {
+				bitmap_clear(map, i, 1);
+				return i;
+			}
+
+			ret = 0;
+
+			bitmap_set(dst, mt[i].to, mt[i].n);
+
+			bitset ^= t;
+		}
+		map[k] = 0;
+	}
+
+	return ret;
+}
+
+/**
+ * nft_pipapo_lookup() - Lookup function
+ * @net:	Network namespace
+ * @set:	nftables API set representation
+ * @elem:	nftables API element representation containing key data
+ * @ext:	nftables API extension pointer, filled with matching reference
+ *
+ * For more details, see DOC: Theory of Operation.
+ *
+ * Return: true on match, false otherwise.
+ */
+static bool nft_pipapo_lookup(const struct net *net, const struct nft_set *set,
+			      const u32 *key, const struct nft_set_ext **ext)
+{
+	struct nft_pipapo *priv = nft_set_priv(set);
+	unsigned long *res_map, *fill_map;
+	u8 genmask = nft_genmask_cur(net);
+	const u8 *rp = (const u8 *)key;
+	struct nft_pipapo_match *m;
+	struct nft_pipapo_field *f;
+	bool map_index;
+	int i;
+
+	local_bh_disable();
+
+	map_index = raw_cpu_read(nft_pipapo_scratch_index);
+
+	m = rcu_dereference(priv->match);
+
+	if (unlikely(!m || !*raw_cpu_ptr(m->scratch)))
+		goto out;
+
+	res_map  = *raw_cpu_ptr(m->scratch) + (map_index ? m->bsize_max : 0);
+	fill_map = *raw_cpu_ptr(m->scratch) + (map_index ? 0 : m->bsize_max);
+
+	memset(res_map, 0xff, m->bsize_max * sizeof(*res_map));
+
+	nft_pipapo_for_each_field(f, i, m) {
+		bool last = i == m->field_count - 1;
+		unsigned long *lt = f->lt;
+		int b, group;
+
+		/* For each 4-bit group: select lookup table bucket depending on
+		 * packet bytes value, then AND bucket value
+		 */
+		for (group = 0; group < f->groups; group++) {
+			u8 v;
+
+			if (group % 2) {
+				v = *rp & 0x0f;
+				rp++;
+			} else {
+				v = *rp >> 4;
+			}
+			__bitmap_and(res_map, res_map, lt + v * f->bsize,
+				     f->bsize * BITS_PER_LONG);
+
+			lt += f->bsize * NFT_PIPAPO_BUCKETS;
+		}
+
+		/* Now populate the bitmap for the next field, unless this is
+		 * the last field, in which case return the matched 'ext'
+		 * pointer if any.
+		 *
+		 * Now res_map contains the matching bitmap, and fill_map is the
+		 * bitmap for the next field.
+		 */
+next_match:
+		b = pipapo_refill(res_map, f->bsize, f->rules, fill_map, f->mt,
+				  last);
+		if (b < 0) {
+			raw_cpu_write(nft_pipapo_scratch_index, map_index);
+			local_bh_enable();
+
+			return false;
+		}
+
+		if (last) {
+			*ext = &f->mt[b].e->ext;
+			if (unlikely(nft_set_elem_expired(*ext) ||
+				     !nft_set_elem_active(*ext, genmask)))
+				goto next_match;
+
+			/* Last field: we're just returning the key without
+			 * filling the initial bitmap for the next field, so the
+			 * current inactive bitmap is clean and can be reused as
+			 * *next* bitmap (not initial) for the next packet.
+			 */
+			raw_cpu_write(nft_pipapo_scratch_index, map_index);
+			local_bh_enable();
+
+			return true;
+		}
+
+		/* Swap bitmap indices: res_map is the initial bitmap for the
+		 * next field, and fill_map is guaranteed to be all-zeroes at
+		 * this point.
+		 */
+		map_index = !map_index;
+		swap(res_map, fill_map);
+
+		rp += NFT_PIPAPO_GROUPS_PADDING(f->groups);
+	}
+
+out:
+	local_bh_enable();
+	return false;
+}
+
+/**
+ * pipapo_get() - Get matching start or end element reference given key data
+ * @net:	Network namespace
+ * @set:	nftables API set representation
+ * @data:	Key data to be matched against existing elements
+ * @flags:	If NFT_SET_ELEM_INTERVAL_END is passed, return the end element
+ *
+ * This is essentially the same as the lookup function, except that it matches
+ * key data against the uncommitted copy and doesn't use preallocated maps for
+ * bitmap results.
+ *
+ * Return: pointer to &struct nft_pipapo_elem on match, error pointer otherwise.
+ */
+static void *pipapo_get(const struct net *net, const struct nft_set *set,
+			const u8 *data, unsigned int flags)
+{
+	struct nft_pipapo *priv = nft_set_priv(set);
+	struct nft_pipapo_match *m = priv->clone;
+	unsigned long *res_map, *fill_map = NULL;
+	void *ret = ERR_PTR(-ENOENT);
+	struct nft_pipapo_field *f;
+	int i;
+
+	res_map = kmalloc_array(m->bsize_max, sizeof(*res_map), GFP_ATOMIC);
+	if (!res_map) {
+		ret = ERR_PTR(-ENOMEM);
+		goto out;
+	}
+
+	fill_map = kcalloc(m->bsize_max, sizeof(*res_map), GFP_ATOMIC);
+	if (!fill_map) {
+		ret = ERR_PTR(-ENOMEM);
+		goto out;
+	}
+
+	memset(res_map, 0xff, m->bsize_max * sizeof(*res_map));
+
+	nft_pipapo_for_each_field(f, i, m) {
+		bool last = i == m->field_count - 1;
+		unsigned long *lt = f->lt;
+		int b, group;
+
+		/* For each 4-bit group: select lookup table bucket depending on
+		 * packet bytes value, then AND bucket value
+		 */
+		for (group = 0; group < f->groups; group++) {
+			u8 v;
+
+			if (group % 2) {
+				v = *data & 0x0f;
+				data++;
+			} else {
+				v = *data >> 4;
+			}
+			__bitmap_and(res_map, res_map, lt + v * f->bsize,
+				     f->bsize * BITS_PER_LONG);
+
+			lt += f->bsize * NFT_PIPAPO_BUCKETS;
+		}
+
+		/* Now populate the bitmap for the next field, unless this is
+		 * the last field, in which case return the matched 'ext'
+		 * pointer if any.
+		 *
+		 * Now res_map contains the matching bitmap, and fill_map is the
+		 * bitmap for the next field.
+		 */
+next_match:
+		b = pipapo_refill(res_map, f->bsize, f->rules, fill_map, f->mt,
+				  last);
+		if (b < 0)
+			goto out;
+
+		if (last) {
+			if (nft_set_elem_expired(&f->mt[b].e->ext))
+				goto next_match;
+
+			if (flags & NFT_SET_ELEM_INTERVAL_END)
+				ret = f->mt[b].e;
+			else
+				ret = f->mt[b].e->start;
+			goto out;
+		}
+
+		data += NFT_PIPAPO_GROUPS_PADDING(f->groups);
+
+		/* Swap bitmap indices: fill_map will be the initial bitmap for
+		 * the next field (i.e. the new res_map), and res_map is
+		 * guaranteed to be all-zeroes at this point, ready to be filled
+		 * according to the next mapping table.
+		 */
+		swap(res_map, fill_map);
+	}
+
+out:
+	kfree(fill_map);
+	kfree(res_map);
+	return ret;
+}
+
+/**
+ * nft_pipapo_get() - Get matching element reference given key data
+ * @net:	Network namespace
+ * @set:	nftables API set representation
+ * @elem:	nftables API element representation containing key data
+ * @flags:	If NFT_SET_ELEM_INTERVAL_END is passed, return the end element
+ */
+static void *nft_pipapo_get(const struct net *net, const struct nft_set *set,
+			    const struct nft_set_elem *elem, unsigned int flags)
+{
+	return pipapo_get(net, set, (const u8 *)elem->key.val.data, flags);
+}
+
+/**
+ * pipapo_resize() - Resize lookup or mapping table, or both
+ * @f:		Field containing lookup and mapping tables
+ * @old_rules:	Previous amount of rules in field
+ * @rules:	New amount of rules
+ *
+ * Increase, decrease or maintain tables size depending on new amount of rules,
+ * and copy data over. In case the new size is smaller, throw away data for
+ * highest-numbered rules.
+ *
+ * Return: 0 on success, -ENOMEM on allocation failure.
+ */
+static int pipapo_resize(struct nft_pipapo_field *f, int old_rules, int rules)
+{
+	long *new_lt = NULL, *new_p, *old_lt = f->lt, *old_p;
+	union nft_pipapo_map_bucket *new_mt, *old_mt = f->mt;
+	size_t new_bucket_size, copy;
+	int group, bucket;
+
+	new_bucket_size = DIV_ROUND_UP(rules, BITS_PER_LONG);
+
+	if (new_bucket_size == f->bsize)
+		goto mt;
+
+	if (new_bucket_size > f->bsize)
+		copy = f->bsize;
+	else
+		copy = new_bucket_size;
+
+	new_lt = kvzalloc(f->groups * NFT_PIPAPO_BUCKETS * new_bucket_size *
+			  sizeof(*new_lt), GFP_KERNEL);
+	if (!new_lt)
+		return -ENOMEM;
+
+	new_p = new_lt;
+	old_p = old_lt;
+	for (group = 0; group < f->groups; group++) {
+		for (bucket = 0; bucket < NFT_PIPAPO_BUCKETS; bucket++) {
+			memcpy(new_p, old_p, copy * sizeof(*new_p));
+			new_p += copy;
+			old_p += copy;
+
+			if (new_bucket_size > f->bsize)
+				new_p += new_bucket_size - f->bsize;
+			else
+				old_p += f->bsize - new_bucket_size;
+		}
+	}
+
+mt:
+	new_mt = kvmalloc(rules * sizeof(*new_mt), GFP_KERNEL);
+	if (!new_mt) {
+		kvfree(new_lt);
+		return -ENOMEM;
+	}
+
+	memcpy(new_mt, f->mt, min(old_rules, rules) * sizeof(*new_mt));
+	if (rules > old_rules) {
+		memset(new_mt + old_rules, 0,
+		       (rules - old_rules) * sizeof(*new_mt));
+	}
+
+	if (new_lt) {
+		f->bsize = new_bucket_size;
+		f->lt = new_lt;
+		kvfree(old_lt);
+	}
+
+	f->mt = new_mt;
+	kvfree(old_mt);
+
+	return 0;
+}
+
+/**
+ * pipapo_bucket_set() - Set rule bit in bucket given group and group value
+ * @f:		Field containing lookup table
+ * @rule:	Rule index
+ * @group:	Group index
+ * @v:		Value of bit group
+ */
+static void pipapo_bucket_set(struct nft_pipapo_field *f, int rule, int group,
+			      int v)
+{
+	unsigned long *pos;
+
+	pos = f->lt + f->bsize * NFT_PIPAPO_BUCKETS * group;
+	pos += f->bsize * v;
+
+	__set_bit(rule, pos);
+}
+
+/**
+ * pipapo_insert() - Insert new rule in field given input key and mask length
+ * @f:		Field containing lookup table
+ * @k:		Input key for classification, without nftables padding
+ * @mask_bits:	Length of mask; matches field length for non-ranged entry
+ *
+ * Insert a new rule reference in lookup buckets corresponding to k and
+ * mask_bits.
+ *
+ * Return: 1 on success (one rule inserted), negative error code on failure.
+ */
+static int pipapo_insert(struct nft_pipapo_field *f, const uint8_t *k,
+			 int mask_bits)
+{
+	int rule = f->rules++, group, ret;
+
+	ret = pipapo_resize(f, f->rules - 1, f->rules);
+	if (ret)
+		return ret;
+
+	for (group = 0; group < f->groups; group++) {
+		int i, v;
+		u8 mask;
+
+		if (group % 2)
+			v = k[group / 2] & 0x0f;
+		else
+			v = k[group / 2] >> 4;
+
+		if (mask_bits >= (group + 1) * 4) {
+			/* Not masked */
+			pipapo_bucket_set(f, rule, group, v);
+		} else if (mask_bits <= group * 4) {
+			/* Completely masked */
+			for (i = 0; i < NFT_PIPAPO_BUCKETS; i++)
+				pipapo_bucket_set(f, rule, group, i);
+		} else {
+			/* The mask limit falls on this group */
+			mask = 0x0f >> (mask_bits - group * 4);
+			for (i = 0; i < NFT_PIPAPO_BUCKETS; i++) {
+				if ((i & ~mask) == (v & ~mask))
+					pipapo_bucket_set(f, rule, group, i);
+			}
+		}
+	}
+
+	return 1;
+}
+
+/**
+ * pipapo_step_diff() - Check if setting @step bit in netmask would change it
+ * @base:	Mask we are expanding
+ * @step:	Step bit for given expansion step
+ * @len:	Total length of mask space (set and unset bits), bytes
+ *
+ * Convenience function for mask expansion.
+ *
+ * Return: true if step bit changes mask (i.e. isn't set), false otherwise.
+ */
+static bool pipapo_step_diff(u8 *base, int step, int len)
+{
+	/* Network order, byte-addressed */
+#ifdef __BIG_ENDIAN__
+	return !(BIT(step % BITS_PER_BYTE) & base[step / BITS_PER_BYTE]);
+#else
+	return !(BIT(step % BITS_PER_BYTE) &
+		 base[len - 1 - step / BITS_PER_BYTE]);
+#endif
+}
+
+/**
+ * pipapo_step_after_end() - Check if mask exceeds range end with given step
+ * @base:	Mask we are expanding
+ * @end:	End of range
+ * @step:	Step bit for given expansion step, highest bit to be set
+ * @len:	Total length of mask space (set and unset bits), bytes
+ *
+ * Convenience function for mask expansion.
+ *
+ * Return: true if mask exceeds range setting step bits, false otherwise.
+ */
+static bool pipapo_step_after_end(const u8 *base, const u8 *end, int step,
+				  int len)
+{
+	u8 tmp[NFT_PIPAPO_MAX_BYTES];
+	int i;
+
+	memcpy(tmp, base, len);
+
+	/* Network order, byte-addressed */
+	for (i = 0; i <= step; i++)
+#ifdef __BIG_ENDIAN__
+		tmp[i / BITS_PER_BYTE] |= BIT(i % BITS_PER_BYTE);
+#else
+		tmp[len - 1 - i / BITS_PER_BYTE] |= BIT(i % BITS_PER_BYTE);
+#endif
+
+	return memcmp(tmp, end, len) > 0;
+}
+
+/**
+ * pipapo_base_sum() - Sum step bit to given len-sized netmask base with carry
+ * @base:	Netmask base
+ * @step:	Step bit to sum
+ * @len:	Netmask length, bytes
+ */
+static void pipapo_base_sum(u8 *base, int step, int len)
+{
+	bool carry = false;
+	int i;
+
+	/* Network order, byte-addressed */
+#ifdef __BIG_ENDIAN__
+	for (i = step / BITS_PER_BYTE; i < len; i++) {
+#else
+	for (i = len - 1 - step / BITS_PER_BYTE; i >= 0; i--) {
+#endif
+		if (carry)
+			base[i]++;
+		else
+			base[i] += 1 << (step % BITS_PER_BYTE);
+
+		if (base[i])
+			break;
+
+		carry = true;
+	}
+}
+
+/**
+ * expand() - Expand range to composing netmasks and insert into lookup table
+ * @f:		Field containing lookup table
+ * @start:	Start of range
+ * @end:	End of range
+ * @len:	Length of value in bits
+ *
+ * Expand range to composing netmasks and insert corresponding rule references
+ * in lookup buckets.
+ *
+ * Return: number of inserted rules on success, negative error code on failure.
+ */
+static int pipapo_expand(struct nft_pipapo_field *f,
+			 const u8 *start, const u8 *end, int len)
+{
+	int step, masks = 0, bytes = DIV_ROUND_UP(len, BITS_PER_BYTE);
+	u8 base[NFT_PIPAPO_MAX_BYTES];
+
+	memcpy(base, start, bytes);
+	while (memcmp(base, end, bytes) <= 0) {
+		int err;
+
+		step = 0;
+		while (pipapo_step_diff(base, step, bytes)) {
+			if (pipapo_step_after_end(base, end, step, bytes))
+				break;
+
+			step++;
+			if (step >= len) {
+				if (!masks) {
+					pipapo_insert(f, base, 0);
+					masks = 1;
+				}
+				goto out;
+			}
+		}
+
+		err = pipapo_insert(f, base, len - step);
+
+		if (err < 0)
+			return err;
+
+		masks++;
+		pipapo_base_sum(base, step, bytes);
+	}
+out:
+	return masks;
+}
+
+/**
+ * pipapo_map() - Insert rules in mapping tables, mapping them between fields
+ * @m:		Matching data, including mapping table
+ * @map:	Table of rule maps: array of first rule and amount of rules
+ *		in next field a given rule maps to, for each field
+ * @ext:	For last field, nft_set_ext pointer matching rules map to
+ */
+static void pipapo_map(struct nft_pipapo_match *m,
+		       union nft_pipapo_map_bucket map[NFT_PIPAPO_MAX_FIELDS],
+		       struct nft_pipapo_elem *e)
+{
+	struct nft_pipapo_field *f;
+	int i, j;
+
+	for (i = 0, f = m->f; i < m->field_count - 1; i++, f++) {
+		for (j = 0; j < map[i].n; j++) {
+			f->mt[map[i].to + j].to = map[i + 1].to;
+			f->mt[map[i].to + j].n = map[i + 1].n;
+		}
+	}
+
+	/* Last field: map to ext instead of mapping to next field */
+	for (j = 0; j < map[i].n; j++)
+		f->mt[map[i].to + j].e = e;
+}
+
+/**
+ * pipapo_realloc_scratch() - Reallocate scratch maps for partial match results
+ * @clone:	Copy of matching data with pending insertions and deletions
+ * @bsize_max	Maximum bucket size, scratch maps cover two buckets
+ *
+ * Return: 0 on success, -ENOMEM on failure.
+ */
+static int pipapo_realloc_scratch(struct nft_pipapo_match *clone,
+				  unsigned long bsize_max)
+{
+	int i;
+
+	for_each_possible_cpu(i) {
+		unsigned long *scratch;
+
+		scratch = kzalloc_node(bsize_max * sizeof(*scratch) * 2,
+				       GFP_KERNEL, cpu_to_node(i));
+		if (!scratch) {
+			/* On failure, there's no need to undo previous
+			 * allocations: this means that some scratch maps have
+			 * a bigger allocated size now (this is only called on
+			 * insertion), but the extra space won't be used by any
+			 * CPU as new elements are not inserted and m->bsize_max
+			 * is not updated.
+			 */
+			return -ENOMEM;
+		}
+
+		kfree(*per_cpu_ptr(clone->scratch, i));
+
+		*per_cpu_ptr(clone->scratch, i) = scratch;
+	}
+
+	return 0;
+}
+
+/**
+ * nft_pipapo_insert() - Validate and insert ranged elements
+ * @net:	Network namespace
+ * @set:	nftables API set representation
+ * @elem:	nftables API element representation containing key data
+ * @flags:	If NFT_SET_ELEM_INTERVAL_END is passed, this is the end element
+ * @ext2:	Filled with pointer to &struct nft_set_ext in inserted element
+ *
+ * In this set implementation, this functions needs to be called twice, with
+ * start and end element, to obtain a valid entry insertion.
+ *
+ * Calls to this function are serialised with each other, so we can store
+ * element and key data on the first call with start element, and use it on the
+ * second call once we get the end element too.
+ *
+ * However, userspace could send a single NFT_SET_ELEM_INTERVAL_END element,
+ * without a start element, so we need to check for it explicitly before
+ * inserting an entry, lest we end up in nft_pipapo_walk() with an empty start
+ * element.
+ *
+ * Also, we need to make sure that the start element hasn't been deactivated or
+ * destroyed between the two calls to this function, otherwise we might link an
+ * invalid start item to the end item triggering the insertion. Clear
+ * priv->start_elem on any operation that might render it invalid.
+ *
+ * Return: 0 on success, error pointer on failure.
+ */
+static int nft_pipapo_insert(const struct net *net, const struct nft_set *set,
+			     const struct nft_set_elem *elem,
+			     struct nft_set_ext **ext2)
+{
+	const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
+	const u8 *data = (const u8 *)elem->key.val.data, *start, *end;
+	union nft_pipapo_map_bucket rulemap[NFT_PIPAPO_MAX_FIELDS];
+	struct nft_pipapo *priv = nft_set_priv(set);
+	struct nft_pipapo_match *m = priv->clone;
+	struct nft_pipapo_elem *e = elem->priv;
+	struct nft_pipapo_field *f;
+	int i, bsize_max, err = 0;
+	void *dup;
+
+	dup = nft_pipapo_get(net, set, elem, 0);
+	if (PTR_ERR(dup) != -ENOENT) {
+		priv->start_elem = NULL;
+		if (IS_ERR(dup))
+			return PTR_ERR(dup);
+		*ext2 = dup;
+		return -EEXIST;
+	}
+
+	if (!nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) ||
+	    !(*nft_set_ext_flags(ext) & NFT_SET_ELEM_INTERVAL_END)) {
+		priv->start_elem = e;
+		*ext2 = &e->ext;
+		memcpy(priv->start_data, data, priv->width);
+		return 0;
+	}
+
+	if (!priv->start_elem)
+		return -EINVAL;
+
+	e->start = priv->start_elem;
+
+	/* Validate */
+	start = priv->start_data;
+	end = data;
+
+	nft_pipapo_for_each_field(f, i, m) {
+		if (f->rules >= (unsigned long)NFT_PIPAPO_RULE0_MAX)
+			return -ENOSPC;
+
+		if (memcmp(start, end,
+			   f->groups / NFT_PIPAPO_GROUPS_PER_BYTE) > 0)
+			return -EINVAL;
+
+		start += NFT_PIPAPO_GROUPS_PADDED_SIZE(f->groups);
+		end += NFT_PIPAPO_GROUPS_PADDED_SIZE(f->groups);
+	}
+
+	/* Insert */
+	priv->dirty = true;
+
+	bsize_max = m->bsize_max;
+
+	start = priv->start_data;
+	end = data;
+	nft_pipapo_for_each_field(f, i, m) {
+		int ret;
+
+		rulemap[i].to = f->rules;
+
+		ret = memcmp(start, end,
+			     f->groups / NFT_PIPAPO_GROUPS_PER_BYTE);
+		if (!ret) {
+			ret = pipapo_insert(f, start,
+					    f->groups * NFT_PIPAPO_GROUP_BITS);
+		} else {
+			ret = pipapo_expand(f, start, end,
+					    f->groups * NFT_PIPAPO_GROUP_BITS);
+		}
+
+		if (f->bsize > bsize_max)
+			bsize_max = f->bsize;
+
+		rulemap[i].n = ret;
+
+		start += NFT_PIPAPO_GROUPS_PADDED_SIZE(f->groups);
+		end += NFT_PIPAPO_GROUPS_PADDED_SIZE(f->groups);
+	}
+
+	if (!*this_cpu_ptr(m->scratch) || bsize_max > m->bsize_max) {
+		err = pipapo_realloc_scratch(m, bsize_max);
+		if (err)
+			return err;
+
+		this_cpu_write(nft_pipapo_scratch_index, false);
+
+		m->bsize_max = bsize_max;
+	}
+
+	*ext2 = &e->ext;
+
+	pipapo_map(m, rulemap, e);
+
+	return 0;
+}
+
+/**
+ * pipapo_clone() - Clone matching data to create new working copy
+ * @old:	Existing matching data
+ *
+ * Return: copy of matching data passed as 'old', error pointer on failure
+ */
+static struct nft_pipapo_match *pipapo_clone(struct nft_pipapo_match *old)
+{
+	struct nft_pipapo_field *dst, *src;
+	struct nft_pipapo_match *new;
+	int i;
+
+	new = kmalloc(sizeof(*new) + sizeof(*dst) * old->field_count,
+		      GFP_KERNEL);
+	if (!new)
+		return ERR_PTR(-ENOMEM);
+
+	new->field_count = old->field_count;
+	new->bsize_max = old->bsize_max;
+
+	new->scratch = alloc_percpu(*new->scratch);
+	if (!new->scratch)
+		goto out_scratch;
+
+	rcu_head_init(&new->rcu);
+
+	src = old->f;
+	dst = new->f;
+
+	for (i = 0; i < old->field_count; i++) {
+		memcpy(dst, src, offsetof(struct nft_pipapo_field, lt));
+
+		dst->lt = kvzalloc(src->groups * NFT_PIPAPO_BUCKETS *
+				   src->bsize * sizeof(*dst->lt),
+				   GFP_KERNEL);
+		if (!dst->lt)
+			goto out_lt;
+
+		memcpy(dst->lt, src->lt,
+		       src->bsize * sizeof(*dst->lt) *
+		       src->groups * NFT_PIPAPO_BUCKETS);
+
+		dst->mt = kvmalloc(src->rules * sizeof(*src->mt), GFP_KERNEL);
+		if (!dst->mt)
+			goto out_mt;
+
+		memcpy(dst->mt, src->mt, src->rules * sizeof(*src->mt));
+		src++;
+		dst++;
+	}
+
+	return new;
+
+out_mt:
+	kvfree(dst->lt);
+out_lt:
+	for (dst--; i > 0; i--) {
+		kvfree(dst->mt);
+		kvfree(dst->lt);
+		dst--;
+	}
+	free_percpu(new->scratch);
+out_scratch:
+	kfree(new);
+
+	return ERR_PTR(-ENOMEM);
+}
+
+/**
+ * pipapo_rules_same_key() - Get number of rules originated from the same entry
+ * @f:		Field containing mapping table
+ * @first:	Index of first rule in set of rules mapping to same entry
+ *
+ * Using the fact that all rules in a field that originated from the same entry
+ * will map to the same set of rules in the next field, or to the same element
+ * reference, return the cardinality of the set of rules that originated from
+ * the same entry as the rule with index @first, @first rule included.
+ *
+ * In pictures:
+ *				rules
+ *	field #0		0    1    2    3    4
+ *		map to:		0    1   2-4  2-4  5-9
+ *				.    .    .......   . ...
+ *				|    |    |    | \   \
+ *				|    |    |    |  \   \
+ *				|    |    |    |   \   \
+ *				'    '    '    '    '   \
+ *	in field #1		0    1    2    3    4    5 ...
+ *
+ * if this is called for rule 2 on field #0, it will return 3, as also rules 2
+ * and 3 in field 0 map to the same set of rules (2, 3, 4) in the next field.
+ *
+ * For the last field in a set, we can rely on associated entries to map to the
+ * same element references.
+ *
+ * Return: Number of rules that originated from the same entry as @first.
+ */
+static int pipapo_rules_same_key(struct nft_pipapo_field *f, int first)
+{
+	struct nft_pipapo_elem *e = NULL; /* Keep gcc happy */
+	int r;
+
+	for (r = first; r < f->rules; r++) {
+		if (r != first && e != f->mt[r].e)
+			return r - first;
+
+		e = f->mt[r].e;
+	}
+
+	if (r != first)
+		return r - first;
+
+	return 0;
+}
+
+/**
+ * pipapo_unmap() - Remove rules from mapping tables, renumber remaining ones
+ * @mt:		Mapping array
+ * @rules:	Original amount of rules in mapping table
+ * @start:	First rule index to be removed
+ * @n:		Amount of rules to be removed
+ * @to_offset:	First rule index, in next field, this group of rules maps to
+ * @is_last:	If this is the last field, delete reference from mapping array
+ *
+ * This is used to unmap rules from the mapping table for a single field,
+ * maintaining consistency and compactness for the existing ones.
+ *
+ * In pictures: let's assume that we want to delete rules 2 and 3 from the
+ * following mapping array:
+ *
+ *                 rules
+ *               0      1      2      3      4
+ *      map to:  4-10   4-10   11-15  11-15  16-18
+ *
+ * the result will be:
+ *
+ *                 rules
+ *               0      1      2
+ *      map to:  4-10   4-10   11-13
+ *
+ * for fields before the last one. In case this is the mapping table for the
+ * last field in a set, and rules map to pointers to &struct nft_pipapo_elem:
+ *
+ *                      rules
+ *                        0      1      2      3      4
+ *  element pointers:  0x42   0x42   0x33   0x33   0x44
+ *
+ * the result will be:
+ *
+ *                      rules
+ *                        0      1      2
+ *  element pointers:  0x42   0x42   0x44
+ */
+static void pipapo_unmap(union nft_pipapo_map_bucket *mt, int rules,
+			 int start, int n, int to_offset, bool is_last)
+{
+	int i;
+
+	memmove(mt + start, mt + start + n, (rules - start - n) * sizeof(*mt));
+	memset(mt + rules - n, 0, n * sizeof(*mt));
+
+	if (is_last)
+		return;
+
+	for (i = start; i < rules - n; i++)
+		mt[i].to -= to_offset;
+}
+
+/**
+ * pipapo_drop() - Delete entry from lookup and mapping tables, given rule map
+ * @m:		Matching data
+ * @rulemap	Table of rule maps, arrays of first rule and amount of rules
+ *		in next field a given entry maps to, for each field
+ *
+ * For each rule in lookup table buckets mapping to this set of rules, drop
+ * all bits set in lookup table mapping. In pictures, assuming we want to drop
+ * rules 0 and 1 from this lookup table:
+ *
+ *                     bucket
+ *      group  0   1   2   3   4   5   6   7   8   9  10  11  12  13  14  15
+ *        0    0                                              1,2
+ *        1   1,2                                      0
+ *        2    0                                      1,2
+ *        3    0                              1,2
+ *        4  0,1,2
+ *        5    0   1   2
+ *        6  0,1,2 1   1   1   1   1   1   1   1   1   1   1   1   1   1   1
+ *        7   1,2 1,2  1   1   1  0,1  1   1   1   1   1   1   1   1   1   1
+ *
+ * rule 2 becomes rule 0, and the result will be:
+ *
+ *                     bucket
+ *      group  0   1   2   3   4   5   6   7   8   9  10  11  12  13  14  15
+ *        0                                                    0
+ *        1    0
+ *        2                                            0
+ *        3                                    0
+ *        4    0
+ *        5            0
+ *        6    0
+ *        7    0   0
+ *
+ * once this is done, call unmap() to drop all the corresponding rule references
+ * from mapping tables.
+ */
+static void pipapo_drop(struct nft_pipapo_match *m,
+			union nft_pipapo_map_bucket rulemap[])
+{
+	struct nft_pipapo_field *f;
+	int i;
+
+	nft_pipapo_for_each_field(f, i, m) {
+		int g;
+
+		for (g = 0; g < f->groups; g++) {
+			unsigned long *pos;
+			int b;
+
+			pos = f->lt + g * NFT_PIPAPO_BUCKETS * f->bsize;
+
+			for (b = 0; b < NFT_PIPAPO_BUCKETS; b++) {
+				bitmap_cut(pos, pos, rulemap[i].to,
+					   rulemap[i].n,
+					   f->bsize * BITS_PER_LONG);
+
+				pos += f->bsize;
+			}
+		}
+
+		pipapo_unmap(f->mt, f->rules, rulemap[i].to, rulemap[i].n,
+			     rulemap[i + 1].n, i == m->field_count - 1);
+		if (pipapo_resize(f, f->rules, f->rules - rulemap[i].n)) {
+			/* We can ignore this, a failure to shrink tables down
+			 * doesn't make tables invalid.
+			 */
+			;
+		}
+		f->rules -= rulemap[i].n;
+	}
+}
+
+/**
+ * pipapo_gc() - Drop expired entries from set, destroy start and end elements
+ * @set:	nftables API set representation
+ * @m:		Matching data
+ */
+static void pipapo_gc(const struct nft_set *set, struct nft_pipapo_match *m)
+{
+	struct nft_pipapo *priv = nft_set_priv(set);
+	int rules_f0, first_rule = 0;
+
+	while ((rules_f0 = pipapo_rules_same_key(m->f, first_rule))) {
+		union nft_pipapo_map_bucket rulemap[NFT_PIPAPO_MAX_FIELDS];
+		struct nft_pipapo_field *f;
+		struct nft_pipapo_elem *e;
+		int i, start, rules_fx;
+
+		start = first_rule;
+		rules_fx = rules_f0;
+
+		nft_pipapo_for_each_field(f, i, m) {
+			rulemap[i].to = start;
+			rulemap[i].n = rules_fx;
+
+			if (i < m->field_count - 1) {
+				rules_fx = f->mt[start].n;
+				start = f->mt[start].to;
+			}
+		}
+
+		/* Pick the last field, and its last index */
+		f--;
+		i--;
+		e = f->mt[rulemap[i].to].e;
+		if (nft_set_elem_expired(&e->ext) &&
+		    !nft_set_elem_mark_busy(&e->ext)) {
+			priv->dirty = true;
+			pipapo_drop(m, rulemap);
+
+			rcu_barrier();
+			nft_set_elem_destroy(set, e->start, true);
+			nft_set_elem_destroy(set, e, true);
+
+			/* And check again current first rule, which is now the
+			 * first we haven't checked.
+			 */
+		} else {
+			first_rule += rules_f0;
+		}
+	}
+
+	priv->last_gc = jiffies;
+}
+
+/**
+ * pipapo_free_fields() - Free per-field tables contained in matching data
+ * @m:		Matching data
+ */
+static void pipapo_free_fields(struct nft_pipapo_match *m)
+{
+	struct nft_pipapo_field *f;
+	int i;
+
+	nft_pipapo_for_each_field(f, i, m) {
+		kvfree(f->lt);
+		kvfree(f->mt);
+	}
+}
+
+/**
+ * pipapo_reclaim_match - RCU callback to free fields from old matching data
+ * @rcu:	RCU head
+ */
+static void pipapo_reclaim_match(struct rcu_head *rcu)
+{
+	struct nft_pipapo_match *m;
+	int i;
+
+	m = container_of(rcu, struct nft_pipapo_match, rcu);
+
+	for_each_possible_cpu(i)
+		kfree(*per_cpu_ptr(m->scratch, i));
+
+	free_percpu(m->scratch);
+
+	pipapo_free_fields(m);
+
+	kfree(m);
+}
+
+/**
+ * pipapo_commit() - Replace lookup data with current working copy
+ * @set:	nftables API set representation
+ *
+ * While at it, check if we should perform garbage collection on the working
+ * copy before committing it for lookup, and don't replace the table if the
+ * working copy doesn't have pending changes.
+ *
+ * We also need to create a new working copy for subsequent insertions and
+ * deletions.
+ */
+static void pipapo_commit(const struct nft_set *set)
+{
+	struct nft_pipapo *priv = nft_set_priv(set);
+	struct nft_pipapo_match *new_clone, *old;
+
+	if (time_after_eq(jiffies, priv->last_gc + nft_set_gc_interval(set)))
+		pipapo_gc(set, priv->clone);
+
+	if (!priv->dirty)
+		return;
+
+	new_clone = pipapo_clone(priv->clone);
+	if (IS_ERR(new_clone))
+		return;
+
+	priv->dirty = false;
+
+	old = rcu_access_pointer(priv->match);
+	rcu_assign_pointer(priv->match, priv->clone);
+	if (old)
+		call_rcu(&old->rcu, pipapo_reclaim_match);
+
+	priv->clone = new_clone;
+}
+
+/**
+ * nft_pipapo_activate() - Mark element reference as active given key, commit
+ * @net:	Network namespace
+ * @set:	nftables API set representation
+ * @elem:	nftables API element representation containing key data
+ *
+ * On insertion, elements are added to a copy of the matching data currently
+ * in use for lookups, and not directly inserted into current lookup data, so
+ * we'll take care of that by calling pipapo_commit() here. This is probably as
+ * close as we can get to an actual atomic transaction: both nft_pipapo_insert()
+ * and nft_pipapo_activate() are called once for each element, hence we can't
+ * purpose either one as a real commit operation.
+ */
+static void nft_pipapo_activate(const struct net *net,
+				const struct nft_set *set,
+				const struct nft_set_elem *elem)
+{
+	const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
+	struct nft_pipapo_elem *e;
+
+	if (!nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) ||
+	    !(*nft_set_ext_flags(ext) & NFT_SET_ELEM_INTERVAL_END)) {
+		e = pipapo_get(net, set, (const u8 *)elem->key.val.data, 0);
+		if (IS_ERR(e))
+			return;
+
+		nft_set_elem_change_active(net, set, &e->ext);
+		nft_set_elem_clear_busy(&e->ext);
+
+		return;
+	}
+
+	e = pipapo_get(net, set, (const u8 *)elem->key.val.data,
+		       NFT_SET_ELEM_INTERVAL_END);
+	if (IS_ERR(e))
+		return;
+
+	nft_set_elem_change_active(net, set, &e->ext);
+	nft_set_elem_clear_busy(&e->ext);
+
+	pipapo_commit(set);
+}
+
+/**
+ * pipapo_deactivate() - Check that element is in set, mark as inactive
+ * @net:	Network namespace
+ * @set:	nftables API set representation
+ * @data:	Input key data
+ * @ext:	nftables API extension pointer, used to check for end element
+ *
+ * This is a convenience function that can be called from both
+ * nft_pipapo_deactivate() and nft_pipapo_flush(), as they are in fact the same
+ * operation.
+ *
+ * Return: deactivated element if found, NULL otherwise.
+ */
+static void *pipapo_deactivate(const struct net *net, const struct nft_set *set,
+			       const u8 *data, const struct nft_set_ext *ext)
+{
+	struct nft_pipapo *priv = nft_set_priv(set);
+	u8 genmask = nft_genmask_next(net);
+	struct nft_pipapo_elem *e;
+	unsigned int flags = 0;
+
+	/* See nft_pipapo_insert() */
+	priv->start_elem = NULL;
+
+	if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS))
+		flags = *nft_set_ext_flags(ext) & NFT_SET_ELEM_INTERVAL_END;
+
+	e = pipapo_get(net, set, data, flags);
+	if (IS_ERR(e))
+		return NULL;
+
+	if (nft_set_elem_active(&e->ext, genmask))
+		nft_set_elem_change_active(net, set, &e->ext);
+
+	return e;
+}
+
+/**
+ * nft_pipapo_deactivate() - Call pipapo_deactivate() to make element inactive
+ * @net:	Network namespace
+ * @set:	nftables API set representation
+ * @elem:	nftables API element representation containing key data
+ *
+ * Return: deactivated element if found, NULL otherwise.
+ */
+static void *nft_pipapo_deactivate(const struct net *net,
+				   const struct nft_set *set,
+				   const struct nft_set_elem *elem)
+{
+	const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
+
+	return pipapo_deactivate(net, set, (const u8 *)elem->key.val.data, ext);
+}
+
+/**
+ * nft_pipapo_flush() - Call pipapo_deactivate() to make element inactive
+ * @net:	Network namespace
+ * @set:	nftables API set representation
+ * @elem:	nftables API element representation containing key data
+ *
+ * This is functionally the same as nft_pipapo_deactivate(), with a slightly
+ * different interface, and it's also called once for each element in a set
+ * being flushed, so we can't implement an atomic flush operation, which would
+ * otherwise be as simple as allocating an empty copy of the matching data.
+ *
+ * Note that we could in theory do that, mark the set as flushed, and ignore
+ * subsequent calls, but we would leak all the elements after the first one,
+ * because they wouldn't then be freed as result of API calls.
+ *
+ * Return: true if element was found and deactivated.
+ */
+static bool nft_pipapo_flush(const struct net *net, const struct nft_set *set,
+			     void *elem)
+{
+	struct nft_pipapo_elem *e = elem;
+
+	return pipapo_deactivate(net, set, (const u8 *)nft_set_ext_key(&e->ext),
+				 &e->ext);
+}
+
+/**
+ * pipapo_get_boundaries() - Get byte interval for associated rules
+ * @f:		Field including lookup table
+ * @first_rule:	First rule (lowest index)
+ * @rule_count:	Number of associated rules
+ * @left:	Byte expression for left boundary (start of range)
+ * @right:	Byte expression for right boundary (end of range)
+ *
+ * Given the first rule and amount of rules that originated from the same entry,
+ * build the original range associated with the entry, and calculate the length
+ * of the originating netmask.
+ *
+ * In pictures:
+ *
+ *                     bucket
+ *      group  0   1   2   3   4   5   6   7   8   9  10  11  12  13  14  15
+ *        0                                                   1,2
+ *        1   1,2
+ *        2                                           1,2
+ *        3                                   1,2
+ *        4   1,2
+ *        5        1   2
+ *        6   1,2  1   1   1   1   1   1   1   1   1   1   1   1   1   1   1
+ *        7   1,2 1,2  1   1   1   1   1   1   1   1   1   1   1   1   1   1
+ *
+ * this is the lookup table corresponding to the IPv4 range
+ * 192.168.1.0-192.168.2.1, which was expanded to the two composing netmasks,
+ * rule #1: 192.168.1.0/24, and rule #2: 192.168.2.0/31.
+ *
+ * This function fills @left and @right with the byte values of the leftmost
+ * and rightmost bucket indices for the lowest and highest rule indices,
+ * respectively. If @first_rule is 1 and @rule_count is 2, we obtain, in
+ * nibbles:
+ *   left:  < 12, 0, 10, 8, 0, 1, 0, 0 >
+ *   right: < 12, 0, 10, 8, 0, 2, 2, 1 >
+ * corresponding to bytes:
+ *   left:  < 192, 168, 1, 0 >
+ *   right: < 192, 168, 2, 1 >
+ * with mask length irrelevant here, unused on return, as the range is already
+ * defined by its start and end points. The mask length is relevant for a single
+ * ranged entry instead: if @first_rule is 1 and @rule_count is 1, we ignore
+ * rule 2 above: @left becomes < 192, 168, 1, 0 >, @right becomes
+ * < 192, 168, 1, 255 >, and the mask length, calculated from the distances
+ * between leftmost and rightmost bucket indices for each group, would be 24.
+ *
+ * Return: mask length, in bits.
+ */
+static int pipapo_get_boundaries(struct nft_pipapo_field *f, int first_rule,
+				 int rule_count, u8 *left, u8 *right)
+{
+	u8 *l = left, *r = right;
+	int g, mask_len = 0;
+
+	for (g = 0; g < f->groups; g++) {
+		int b, x0, x1;
+
+		x0 = -1;
+		x1 = -1;
+		for (b = 0; b < NFT_PIPAPO_BUCKETS; b++) {
+			unsigned long *pos;
+
+			pos = f->lt + (g * NFT_PIPAPO_BUCKETS + b) * f->bsize;
+			if (test_bit(first_rule, pos) && x0 == -1)
+				x0 = b;
+			if (test_bit(first_rule + rule_count - 1, pos))
+				x1 = b;
+		}
+
+		if (g % 2) {
+			*(l++) |= x0 & 0x0f;
+			*(r++) |= x1 & 0x0f;
+		} else {
+			*l |= x0 << 4;
+			*r |= x1 << 4;
+		}
+
+		if (x1 - x0 == 0)
+			mask_len += 4;
+		else if (x1 - x0 == 1)
+			mask_len += 3;
+		else if (x1 - x0 == 3)
+			mask_len += 2;
+		else if (x1 - x0 == 7)
+			mask_len += 1;
+	}
+
+	return mask_len;
+}
+
+/**
+ * pipapo_match_field() - Match rules against byte ranges
+ * @f:		Field including the lookup table
+ * @first_rule:	First of associated rules originating from same entry
+ * @rule_count:	Amount of associated rules
+ * @start:	Start of range to be matched
+ * @end:	End of range to be matched
+ *
+ * Return: true on match, false otherwise.
+ */
+static bool pipapo_match_field(struct nft_pipapo_field *f,
+			       int first_rule, int rule_count,
+			       const u8 *start, const u8 *end)
+{
+	u8 right[NFT_PIPAPO_MAX_BYTES] = { 0 };
+	u8 left[NFT_PIPAPO_MAX_BYTES] = { 0 };
+
+	pipapo_get_boundaries(f, first_rule, rule_count, left, right);
+
+	return !memcmp(start, left, f->groups / NFT_PIPAPO_GROUPS_PER_BYTE) &&
+	       !memcmp(end, right, f->groups / NFT_PIPAPO_GROUPS_PER_BYTE);
+}
+
+/**
+ * nft_pipapo_remove() - Remove element given key, commit
+ * @net:	Network namespace
+ * @set:	nftables API set representation
+ * @elem:	nftables API element representation containing key data
+ *
+ * Similarly to nft_pipapo_activate(), this is used as commit operation by the
+ * API, but it's called once per element in the pending transaction, so we can't
+ * implement an actual, atomic commit operation. Closest we can get is to remove
+ * the matched element here, if any, and commit the updated matching data.
+ */
+static void nft_pipapo_remove(const struct net *net, const struct nft_set *set,
+			      const struct nft_set_elem *elem)
+{
+	const u8 *data = (const u8 *)elem->key.val.data;
+	struct nft_pipapo *priv = nft_set_priv(set);
+	struct nft_pipapo_match *m = priv->clone;
+	const struct nft_set_ext *ext;
+	int rules_f0, first_rule = 0;
+	struct nft_pipapo_elem *e;
+
+	/* See nft_pipapo_insert() */
+	priv->start_elem = NULL;
+
+	ext = nft_set_elem_ext(set, elem->priv);
+	if (!nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) ||
+	    !(*nft_set_ext_flags(ext) & NFT_SET_ELEM_INTERVAL_END))
+		return;
+
+	e = pipapo_get(net, set, data, NFT_SET_ELEM_INTERVAL_END);
+	if (IS_ERR(e))
+		return;
+
+	while ((rules_f0 = pipapo_rules_same_key(m->f, first_rule))) {
+		union nft_pipapo_map_bucket rulemap[NFT_PIPAPO_MAX_FIELDS];
+		const u8 *match_start, *match_end;
+		struct nft_pipapo_field *f;
+		int i, start, rules_fx;
+
+		match_start = (const u8 *)nft_set_ext_key(&e->start->ext);
+		match_end = data;
+
+		start = first_rule;
+		rules_fx = rules_f0;
+
+		nft_pipapo_for_each_field(f, i, m) {
+			if (!pipapo_match_field(f, start, rules_fx,
+						match_start, match_end))
+				break;
+
+			rulemap[i].to = start;
+			rulemap[i].n = rules_fx;
+
+			rules_fx = f->mt[start].n;
+			start = f->mt[start].to;
+
+			match_start += NFT_PIPAPO_GROUPS_PADDED_SIZE(f->groups);
+			match_end += NFT_PIPAPO_GROUPS_PADDED_SIZE(f->groups);
+		}
+
+		if (i == m->field_count) {
+			priv->dirty = true;
+			pipapo_drop(m, rulemap);
+			pipapo_commit(set);
+			return;
+		}
+
+		first_rule += rules_f0;
+	}
+}
+
+/**
+ * nft_pipapo_walk() - Walk over elements
+ * @ctx:	nftables API context
+ * @set:	nftables API set representation
+ * @iter:	Iterator
+ *
+ * As elements are referenced in the mapping array for the last field, directly
+ * scan that array: there's no need to follow rule mappings from the first
+ * field.
+ *
+ * Note that we'll return two elements for each call, as each entry is
+ * represented as start and end elements.
+ */
+static void nft_pipapo_walk(const struct nft_ctx *ctx, struct nft_set *set,
+			    struct nft_set_iter *iter)
+{
+	struct nft_pipapo *priv = nft_set_priv(set);
+	struct nft_pipapo_match *m;
+	struct nft_pipapo_field *f;
+	int i, r;
+
+	rcu_read_lock();
+	m = rcu_dereference(priv->match);
+
+	if (unlikely(!m))
+		goto out;
+
+	for (i = 0, f = m->f; i < m->field_count - 1; i++, f++)
+		;
+
+	for (r = 0; r < f->rules; r++) {
+		struct nft_set_elem elem_start, elem_end;
+		struct nft_pipapo_elem *e;
+
+		if (r < f->rules - 1 && f->mt[r + 1].e == f->mt[r].e)
+			continue;
+
+		if (iter->count < iter->skip)
+			goto cont;
+
+		e = f->mt[r].e;
+		if (nft_set_elem_expired(&e->ext))
+			goto cont;
+
+		elem_start.priv = e->start;
+
+		iter->err = iter->fn(ctx, set, iter, &elem_start);
+		if (iter->err < 0)
+			goto out;
+
+		elem_end.priv = e;
+
+		iter->err = iter->fn(ctx, set, iter, &elem_end);
+		if (iter->err < 0)
+			goto out;
+
+cont:
+		iter->count += 2;
+	}
+
+out:
+	rcu_read_unlock();
+}
+
+/**
+ * nft_pipapo_privsize() - Return the size of private data for the set
+ * @nla:	netlink attributes, ignored as size doesn't depend on them
+ * @desc:	Set description, ignored as size doesn't depend on it
+ *
+ * Return: size of private data for this set implementation, in bytes
+ */
+static u64 nft_pipapo_privsize(const struct nlattr * const nla[],
+			       const struct nft_set_desc *desc)
+{
+	return sizeof(struct nft_pipapo);
+}
+
+/**
+ * nft_pipapo_estimate() - Estimate set size, space and lookup complexity
+ * @desc:	Set description, initial element count used here
+ * @features:	Flags: NFT_SET_SUBKEY needs to be there
+ * @est:	Storage for estimation data
+ *
+ * The size for this set type can vary dramatically, as it depends on the number
+ * of rules (composing netmasks) the entries expand to. We compute the worst
+ * case here, in order to ensure that other types are used if concatenation of
+ * ranges is not needed.
+ *
+ * In general, for a non-ranged entry or a single composing netmask, we need
+ * one bit in each of the sixteen NFT_PIPAPO_BUCKETS, for each 4-bit group (that
+ * is, each input bit needs four bits of matching data), plus a bucket in the
+ * mapping table for each field.
+ *
+ * Return: true
+ */
+static bool nft_pipapo_estimate(const struct nft_set_desc *desc, u32 features,
+				struct nft_set_estimate *est)
+{
+	if (!(features & NFT_SET_SUBKEY))
+		return false;
+
+	est->size = sizeof(struct nft_pipapo) + sizeof(struct nft_pipapo_match);
+
+	/* Worst-case with current amount of 32-bit VM registers (16 of them):
+	 * - 2 IPv6 addresses	8 registers
+	 * - 2 interface names	8 registers
+	 * that is, four 128-bit fields:
+	 */
+	est->size += sizeof(struct nft_pipapo_field) * 4;
+
+	/* expanding to worst-case ranges, 128 * 2 rules each, resulting in:
+	 * - 128 4-bit groups
+	 * - each set entry taking 256 bits in each bucket
+	 */
+	est->size += desc->size * NFT_PIPAPO_MAX_BITS / NFT_PIPAPO_GROUP_BITS *
+		     NFT_PIPAPO_BUCKETS * NFT_PIPAPO_MAX_BITS * 2 /
+		     BITS_PER_BYTE;
+
+	/* and we need mapping buckets, too */
+	est->size += desc->size * NFT_PIPAPO_MAP_NBITS *
+		     sizeof(union nft_pipapo_map_bucket);
+
+	est->lookup = NFT_SET_CLASS_O_LOG_N;
+
+	est->space = NFT_SET_CLASS_O_N;
+
+	return true;
+}
+
+/**
+ * nft_pipapo_init() - Initialise data for a set instance
+ * @set:	nftables API set representation
+ * @desc:	Set description
+ * @nla:	netlink attributes
+ *
+ * Validate number and size of fields passed as NFTA_SET_SUBKEY netlink
+ * attributes, initialise internal set parameters, current instance of matching
+ * data and a copy for subsequent insertions.
+ *
+ * Return: 0 on success, negative error code on failure.
+ */
+static int nft_pipapo_init(const struct nft_set *set,
+			   const struct nft_set_desc *desc,
+			   const struct nlattr * const nla[])
+{
+	struct nft_pipapo *priv = nft_set_priv(set);
+	int rem, err = -EINVAL, field_count = 0, i;
+	struct nft_pipapo_match *m;
+	struct nft_pipapo_field *f;
+	struct nlattr *attr;
+	unsigned int klen;
+
+	if (!nla || !nla[NFTA_SET_SUBKEY])
+		return -EINVAL;
+
+	nla_for_each_nested(attr, nla[NFTA_SET_SUBKEY], rem) {
+		if (++field_count >= NFT_PIPAPO_MAX_FIELDS)
+			return -EINVAL;
+
+		if (nla_len(attr) != sizeof(klen) ||
+		    nla_type(attr) != NFTA_SET_SUBKEY_LEN)
+			return -EINVAL;
+	}
+
+	if (!field_count)
+		return -EINVAL;
+
+	m = kmalloc(sizeof(*priv->match) + sizeof(*f) * field_count,
+		    GFP_KERNEL);
+	if (!m)
+		return -ENOMEM;
+
+	m->field_count = field_count;
+	m->bsize_max = 0;
+
+	m->scratch = alloc_percpu(unsigned long *);
+	if (!m->scratch) {
+		err = -ENOMEM;
+		goto out_free;
+	}
+	for_each_possible_cpu(i)
+		*per_cpu_ptr(m->scratch, i) = NULL;
+
+	rcu_head_init(&m->rcu);
+
+	f = m->f;
+	priv->width = 0;
+	nla_for_each_nested(attr, nla[NFTA_SET_SUBKEY], rem) {
+		klen = ntohl(nla_get_be32(attr));
+		if (!klen || klen % NFT_PIPAPO_GROUP_BITS)
+			goto out_free;
+
+		if (klen > NFT_PIPAPO_MAX_BITS)
+			goto out_free;
+
+		priv->groups += f->groups = klen / NFT_PIPAPO_GROUP_BITS;
+		priv->width += round_up(klen / BITS_PER_BYTE, sizeof(u32));
+
+		f->bsize = 0;
+		f->rules = 0;
+		f->lt = NULL;
+		f->mt = NULL;
+
+		f++;
+	}
+
+	/* Create an initial clone of matching data for next insertion */
+	priv->clone = pipapo_clone(m);
+	if (IS_ERR(priv->clone)) {
+		err = PTR_ERR(priv->clone);
+		goto out_free;
+	}
+
+	priv->dirty = false;
+
+	rcu_assign_pointer(priv->match, m);
+
+	return 0;
+
+out_free:
+	free_percpu(m->scratch);
+	kfree(m);
+
+	return err;
+}
+
+/**
+ * nft_pipapo_destroy() - Free private data for set and all committed elements
+ * @set:	nftables API set representation
+ */
+static void nft_pipapo_destroy(const struct nft_set *set)
+{
+	struct nft_pipapo *priv = nft_set_priv(set);
+	struct nft_pipapo_match *m;
+	struct nft_pipapo_field *f;
+	int i, r, cpu;
+
+	/* See nft_pipapo_insert() */
+	priv->start_elem = NULL;
+
+	m = rcu_dereference_protected(priv->match, true);
+	if (m) {
+		rcu_barrier();
+
+		for (i = 0, f = m->f; i < m->field_count - 1; i++, f++)
+			;
+
+		for (r = 0; r < f->rules; r++) {
+			struct nft_pipapo_elem *e;
+
+			if (r < f->rules - 1 && f->mt[r + 1].e == f->mt[r].e)
+				continue;
+
+			e = f->mt[r].e;
+
+			nft_set_elem_destroy(set, e->start, true);
+			nft_set_elem_destroy(set, e, true);
+		}
+
+		for_each_possible_cpu(cpu)
+			kfree(*per_cpu_ptr(m->scratch, cpu));
+		free_percpu(m->scratch);
+
+		pipapo_free_fields(m);
+		kfree(m);
+		priv->match = NULL;
+	}
+
+	if (priv->clone) {
+		for_each_possible_cpu(cpu)
+			kfree(*per_cpu_ptr(priv->clone->scratch, cpu));
+		free_percpu(priv->clone->scratch);
+
+		pipapo_free_fields(priv->clone);
+		kfree(priv->clone);
+		priv->clone = NULL;
+	}
+}
+
+/**
+ * nft_pipapo_gc_init() - Initialise garbage collection
+ * @set:	nftables API set representation
+ *
+ * Instead of actually setting up a periodic work for garbage collection, as
+ * this operation requires a swap of matching data with the working copy, we'll
+ * do that opportunistically with other commit operations if the interval is
+ * elapsed, so we just need to set the current jiffies timestamp here.
+ */
+static void nft_pipapo_gc_init(const struct nft_set *set)
+{
+	struct nft_pipapo *priv = nft_set_priv(set);
+
+	priv->last_gc = jiffies;
+}
+
+struct nft_set_type nft_set_pipapo_type __read_mostly = {
+	.owner		= THIS_MODULE,
+	.features	= NFT_SET_INTERVAL | NFT_SET_MAP | NFT_SET_OBJECT |
+			  NFT_SET_TIMEOUT | NFT_SET_SUBKEY,
+	.ops		= {
+		.lookup		= nft_pipapo_lookup,
+		.insert		= nft_pipapo_insert,
+		.activate	= nft_pipapo_activate,
+		.deactivate	= nft_pipapo_deactivate,
+		.flush		= nft_pipapo_flush,
+		.remove		= nft_pipapo_remove,
+		.walk		= nft_pipapo_walk,
+		.get		= nft_pipapo_get,
+		.privsize	= nft_pipapo_privsize,
+		.estimate	= nft_pipapo_estimate,
+		.init		= nft_pipapo_init,
+		.destroy	= nft_pipapo_destroy,
+		.gc_init	= nft_pipapo_gc_init,
+		.elemsize	= offsetof(struct nft_pipapo_elem, ext),
+	},
+};