1 // SPDX-License-Identifier: GPL-2.0-only
3 /* PIPAPO: PIle PAcket POlicies: set for arbitrary concatenations of ranges
5 * Copyright (c) 2019-2020 Red Hat GmbH
7 * Author: Stefano Brivio <sbrivio@redhat.com>
11 * DOC: Theory of Operation
17 * Match packet bytes against entries composed of ranged or non-ranged packet
18 * field specifiers, mapping them to arbitrary references. For example:
23 * | [net],[port],[net]... => [reference]
24 * entries [net],[port],[net]... => [reference]
25 * | [net],[port],[net]... => [reference]
28 * where [net] fields can be IP ranges or netmasks, and [port] fields are port
29 * ranges. Arbitrary packet fields can be matched.
35 * This algorithm is loosely inspired by [Ligatti 2010], and fundamentally
36 * relies on the consideration that every contiguous range in a space of b bits
37 * can be converted into b * 2 netmasks, from Theorem 3 in [Rottenstreich 2010],
38 * as also illustrated in Section 9 of [Kogan 2014].
40 * Classification against a number of entries, that require matching given bits
41 * of a packet field, is performed by grouping those bits in sets of arbitrary
42 * size, and classifying packet bits one group at a time.
45 * to match the source port (16 bits) of a packet, we can divide those 16 bits
46 * in 4 groups of 4 bits each. Given the entry:
48 * and a packet with source port:
50 * first and second groups match, but the third doesn't. We conclude that the
51 * packet doesn't match the given entry.
53 * Translate the set to a sequence of lookup tables, one per field. Each table
54 * has two dimensions: bit groups to be matched for a single packet field, and
55 * all the possible values of said groups (buckets). Input entries are
56 * represented as one or more rules, depending on the number of composing
57 * netmasks for the given field specifier, and a group match is indicated as a
58 * set bit, with number corresponding to the rule index, in all the buckets
59 * whose value matches the entry for a given group.
61 * Rules are mapped between fields through an array of x, n pairs, with each
62 * item mapping a matched rule to one or more rules. The position of the pair in
63 * the array indicates the matched rule to be mapped to the next field, x
64 * indicates the first rule index in the next field, and n the amount of
65 * next-field rules the current rule maps to.
67 * The mapping array for the last field maps to the desired references.
69 * To match, we perform table lookups using the values of grouped packet bits,
70 * and use a sequence of bitwise operations to progressively evaluate rule
73 * A stand-alone, reference implementation, also including notes about possible
74 * future optimisations, is available at:
75 * https://pipapo.lameexcu.se/
80 * - For each packet field:
82 * - divide the b packet bits we want to classify into groups of size t,
83 * obtaining ceil(b / t) groups
85 * Example: match on destination IP address, with t = 4: 32 bits, 8 groups
88 * - allocate a lookup table with one column ("bucket") for each possible
89 * value of a group, and with one row for each group
91 * Example: 8 groups, 2^4 buckets:
96 * group 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
106 * - map the bits we want to classify for the current field, for a given
107 * entry, to a single rule for non-ranged and netmask set items, and to one
108 * or multiple rules for ranges. Ranges are expanded to composing netmasks
109 * by pipapo_expand().
111 * Example: 2 entries, 10.0.0.5:1024 and 192.168.1.0-192.168.2.1:2048
112 * - rule #0: 10.0.0.5
113 * - rule #1: 192.168.1.0/24
114 * - rule #2: 192.168.2.0/31
116 * - insert references to the rules in the lookup table, selecting buckets
117 * according to bit values of a rule in the given group. This is done by
121 * - rule #0: 10.0.0.5 mapping to buckets
122 * < 0 10 0 0 0 0 0 5 >
123 * - rule #1: 192.168.1.0/24 mapping to buckets
124 * < 12 0 10 8 0 1 < 0..15 > < 0..15 > >
125 * - rule #2: 192.168.2.0/31 mapping to buckets
126 * < 12 0 10 8 0 2 0 < 0..1 > >
128 * these bits are set in the lookup table:
133 * group 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
140 * 6 0,1,2 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
141 * 7 1,2 1,2 1 1 1 0,1 1 1 1 1 1 1 1 1 1 1
143 * - if this is not the last field in the set, fill a mapping array that maps
144 * rules from the lookup table to rules belonging to the same entry in
145 * the next lookup table, done by pipapo_map().
147 * Note that as rules map to contiguous ranges of rules, given how netmask
148 * expansion and insertion is performed, &union nft_pipapo_map_bucket stores
149 * this information as pairs of first rule index, rule count.
151 * Example: 2 entries, 10.0.0.5:1024 and 192.168.1.0-192.168.2.1:2048,
152 * given lookup table #0 for field 0 (see example above):
157 * group 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
164 * 6 0,1,2 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
165 * 7 1,2 1,2 1 1 1 0,1 1 1 1 1 1 1 1 1 1 1
167 * and lookup table #1 for field 1 with:
168 * - rule #0: 1024 mapping to buckets
170 * - rule #1: 2048 mapping to buckets
176 * group 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
182 * we need to map rules for 10.0.0.5 in lookup table #0 (rule #0) to 1024
183 * in lookup table #1 (rule #0) and rules for 192.168.1.0-192.168.2.1
184 * (rules #1, #2) to 2048 in lookup table #2 (rule #1):
188 * rule indices in current field: 0 1 2
189 * map to rules in next field: 0 1 1
191 * - if this is the last field in the set, fill a mapping array that maps
192 * rules from the last lookup table to element pointers, also done by
195 * Note that, in this implementation, we have two elements (start, end) for
196 * each entry. The pointer to the end element is stored in this array, and
197 * the pointer to the start element is linked from it.
199 * Example: entry 10.0.0.5:1024 has a corresponding &struct nft_pipapo_elem
200 * pointer, 0x66, and element for 192.168.1.0-192.168.2.1:2048 is at 0x42.
201 * From the rules of lookup table #1 as mapped above:
205 * rule indices in last field: 0 1
206 * map to elements: 0x66 0x42
212 * We use a result bitmap, with the size of a single lookup table bucket, to
213 * represent the matching state that applies at every algorithm step. This is
214 * done by pipapo_lookup().
216 * - For each packet field:
218 * - start with an all-ones result bitmap (res_map in pipapo_lookup())
220 * - perform a lookup into the table corresponding to the current field,
221 * for each group, and at every group, AND the current result bitmap with
222 * the value from the lookup table bucket
226 * Example: 192.168.1.5 < 12 0 10 8 0 1 0 5 >, with lookup table from
227 * insertion examples.
228 * Lookup table buckets are at least 3 bits wide, we'll assume 8 bits for
229 * convenience in this example. Initial result bitmap is 0xff, the steps
230 * below show the value of the result bitmap after each group is processed:
233 * group 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
235 * result bitmap is now: 0xff & 0x6 [bucket 12] = 0x6
238 * result bitmap is now: 0x6 & 0x6 [bucket 0] = 0x6
241 * result bitmap is now: 0x6 & 0x6 [bucket 10] = 0x6
244 * result bitmap is now: 0x6 & 0x6 [bucket 8] = 0x6
247 * result bitmap is now: 0x6 & 0x7 [bucket 0] = 0x6
250 * result bitmap is now: 0x6 & 0x2 [bucket 1] = 0x2
252 * 6 0,1,2 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
253 * result bitmap is now: 0x2 & 0x7 [bucket 0] = 0x2
255 * 7 1,2 1,2 1 1 1 0,1 1 1 1 1 1 1 1 1 1 1
256 * final result bitmap for this field is: 0x2 & 0x3 [bucket 5] = 0x2
258 * - at the next field, start with a new, all-zeroes result bitmap. For each
259 * bit set in the previous result bitmap, fill the new result bitmap
260 * (fill_map in pipapo_lookup()) with the rule indices from the
261 * corresponding buckets of the mapping field for this field, done by
264 * Example: with mapping table from insertion examples, with the current
265 * result bitmap from the previous example, 0x02:
269 * rule indices in current field: 0 1 2
270 * map to rules in next field: 0 1 1
272 * the new result bitmap will be 0x02: rule 1 was set, and rule 1 will be
275 * We can now extend this example to cover the second iteration of the step
276 * above (lookup and AND bitmap): assuming the port field is
277 * 2048 < 0 0 5 0 >, with starting result bitmap 0x2, and lookup table
278 * for "port" field from pre-computation example:
283 * group 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
289 * operations are: 0x2 & 0x3 [bucket 0] & 0x3 [bucket 0] & 0x2 [bucket 5]
290 * & 0x3 [bucket 0], resulting bitmap is 0x2.
292 * - if this is the last field in the set, look up the value from the mapping
293 * array corresponding to the final result bitmap
295 * Example: 0x2 resulting bitmap from 192.168.1.5:2048, mapping array for
296 * last field from insertion example:
300 * rule indices in last field: 0 1
301 * map to elements: 0x66 0x42
303 * the matching element is at 0x42.
310 * A Packet-classification Algorithm for Arbitrary Bitmask Rules, with
311 * Automatic Time-space Tradeoffs
312 * Jay Ligatti, Josh Kuhn, and Chris Gage.
313 * Proceedings of the IEEE International Conference on Computer
314 * Communication Networks (ICCCN), August 2010.
315 * https://www.cse.usf.edu/~ligatti/papers/grouper-conf.pdf
317 * [Rottenstreich 2010]
318 * Worst-Case TCAM Rule Expansion
319 * Ori Rottenstreich and Isaac Keslassy.
320 * 2010 Proceedings IEEE INFOCOM, San Diego, CA, 2010.
321 * http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.212.4592&rep=rep1&type=pdf
324 * SAX-PAC (Scalable And eXpressive PAcket Classification)
325 * Kirill Kogan, Sergey Nikolenko, Ori Rottenstreich, William Culhane,
326 * and Patrick Eugster.
327 * Proceedings of the 2014 ACM conference on SIGCOMM, August 2014.
328 * https://www.sigcomm.org/sites/default/files/ccr/papers/2014/August/2619239-2626294.pdf
331 #include <linux/kernel.h>
332 #include <linux/init.h>
333 #include <linux/module.h>
334 #include <linux/netlink.h>
335 #include <linux/netfilter.h>
336 #include <linux/netfilter/nf_tables.h>
337 #include <net/netfilter/nf_tables_core.h>
338 #include <uapi/linux/netfilter/nf_tables.h>
339 #include <linux/bitmap.h>
340 #include <linux/bitops.h>
342 #include "nft_set_pipapo_avx2.h"
343 #include "nft_set_pipapo.h"
346 * pipapo_refill() - For each set bit, set bits from selected mapping table item
347 * @map: Bitmap to be scanned for set bits
348 * @len: Length of bitmap in longs
349 * @rules: Number of rules in field
350 * @dst: Destination bitmap
351 * @mt: Mapping table containing bit set specifiers
352 * @match_only: Find a single bit and return, don't fill
354 * Iteration over set bits with __builtin_ctzl(): Daniel Lemire, public domain.
356 * For each bit set in map, select the bucket from mapping table with index
357 * corresponding to the position of the bit set. Use start bit and amount of
358 * bits specified in bucket to fill region in dst.
360 * Return: -1 on no match, bit position on 'match_only', 0 otherwise.
362 int pipapo_refill(unsigned long *map, int len, int rules, unsigned long *dst,
363 union nft_pipapo_map_bucket *mt, bool match_only)
365 unsigned long bitset;
368 for (k = 0; k < len; k++) {
371 unsigned long t = bitset & -bitset;
372 int r = __builtin_ctzl(bitset);
373 int i = k * BITS_PER_LONG + r;
375 if (unlikely(i >= rules)) {
381 bitmap_clear(map, i, 1);
387 bitmap_set(dst, mt[i].to, mt[i].n);
398 * nft_pipapo_lookup() - Lookup function
399 * @net: Network namespace
400 * @set: nftables API set representation
401 * @key: nftables API element representation containing key data
402 * @ext: nftables API extension pointer, filled with matching reference
404 * For more details, see DOC: Theory of Operation.
406 * Return: true on match, false otherwise.
408 bool nft_pipapo_lookup(const struct net *net, const struct nft_set *set,
409 const u32 *key, const struct nft_set_ext **ext)
411 struct nft_pipapo *priv = nft_set_priv(set);
412 struct nft_pipapo_scratch *scratch;
413 unsigned long *res_map, *fill_map;
414 u8 genmask = nft_genmask_cur(net);
415 const u8 *rp = (const u8 *)key;
416 struct nft_pipapo_match *m;
417 struct nft_pipapo_field *f;
423 m = rcu_dereference(priv->match);
425 if (unlikely(!m || !*raw_cpu_ptr(m->scratch)))
428 scratch = *raw_cpu_ptr(m->scratch);
430 map_index = scratch->map_index;
432 res_map = scratch->map + (map_index ? m->bsize_max : 0);
433 fill_map = scratch->map + (map_index ? 0 : m->bsize_max);
435 memset(res_map, 0xff, m->bsize_max * sizeof(*res_map));
437 nft_pipapo_for_each_field(f, i, m) {
438 bool last = i == m->field_count - 1;
441 /* For each bit group: select lookup table bucket depending on
442 * packet bytes value, then AND bucket value
444 if (likely(f->bb == 8))
445 pipapo_and_field_buckets_8bit(f, res_map, rp);
447 pipapo_and_field_buckets_4bit(f, res_map, rp);
448 NFT_PIPAPO_GROUP_BITS_ARE_8_OR_4;
450 rp += f->groups / NFT_PIPAPO_GROUPS_PER_BYTE(f);
452 /* Now populate the bitmap for the next field, unless this is
453 * the last field, in which case return the matched 'ext'
456 * Now res_map contains the matching bitmap, and fill_map is the
457 * bitmap for the next field.
460 b = pipapo_refill(res_map, f->bsize, f->rules, fill_map, f->mt,
463 scratch->map_index = map_index;
470 *ext = &f->mt[b].e->ext;
471 if (unlikely(nft_set_elem_expired(*ext) ||
472 !nft_set_elem_active(*ext, genmask)))
475 /* Last field: we're just returning the key without
476 * filling the initial bitmap for the next field, so the
477 * current inactive bitmap is clean and can be reused as
478 * *next* bitmap (not initial) for the next packet.
480 scratch->map_index = map_index;
486 /* Swap bitmap indices: res_map is the initial bitmap for the
487 * next field, and fill_map is guaranteed to be all-zeroes at
490 map_index = !map_index;
491 swap(res_map, fill_map);
493 rp += NFT_PIPAPO_GROUPS_PADDING(f);
502 * pipapo_get() - Get matching element reference given key data
503 * @net: Network namespace
504 * @set: nftables API set representation
505 * @data: Key data to be matched against existing elements
506 * @genmask: If set, check that element is active in given genmask
507 * @tstamp: timestamp to check for expired elements
509 * This is essentially the same as the lookup function, except that it matches
510 * key data against the uncommitted copy and doesn't use preallocated maps for
513 * Return: pointer to &struct nft_pipapo_elem on match, error pointer otherwise.
515 static struct nft_pipapo_elem *pipapo_get(const struct net *net,
516 const struct nft_set *set,
517 const u8 *data, u8 genmask,
520 struct nft_pipapo_elem *ret = ERR_PTR(-ENOENT);
521 struct nft_pipapo *priv = nft_set_priv(set);
522 struct nft_pipapo_match *m = priv->clone;
523 unsigned long *res_map, *fill_map = NULL;
524 struct nft_pipapo_field *f;
527 res_map = kmalloc_array(m->bsize_max, sizeof(*res_map), GFP_ATOMIC);
529 ret = ERR_PTR(-ENOMEM);
533 fill_map = kcalloc(m->bsize_max, sizeof(*res_map), GFP_ATOMIC);
535 ret = ERR_PTR(-ENOMEM);
539 memset(res_map, 0xff, m->bsize_max * sizeof(*res_map));
541 nft_pipapo_for_each_field(f, i, m) {
542 bool last = i == m->field_count - 1;
545 /* For each bit group: select lookup table bucket depending on
546 * packet bytes value, then AND bucket value
549 pipapo_and_field_buckets_8bit(f, res_map, data);
551 pipapo_and_field_buckets_4bit(f, res_map, data);
555 data += f->groups / NFT_PIPAPO_GROUPS_PER_BYTE(f);
557 /* Now populate the bitmap for the next field, unless this is
558 * the last field, in which case return the matched 'ext'
561 * Now res_map contains the matching bitmap, and fill_map is the
562 * bitmap for the next field.
565 b = pipapo_refill(res_map, f->bsize, f->rules, fill_map, f->mt,
571 if (__nft_set_elem_expired(&f->mt[b].e->ext, tstamp))
574 !nft_set_elem_active(&f->mt[b].e->ext, genmask)))
581 data += NFT_PIPAPO_GROUPS_PADDING(f);
583 /* Swap bitmap indices: fill_map will be the initial bitmap for
584 * the next field (i.e. the new res_map), and res_map is
585 * guaranteed to be all-zeroes at this point, ready to be filled
586 * according to the next mapping table.
588 swap(res_map, fill_map);
598 * nft_pipapo_get() - Get matching element reference given key data
599 * @net: Network namespace
600 * @set: nftables API set representation
601 * @elem: nftables API element representation containing key data
604 static struct nft_elem_priv *
605 nft_pipapo_get(const struct net *net, const struct nft_set *set,
606 const struct nft_set_elem *elem, unsigned int flags)
608 struct nft_pipapo_elem *e;
610 e = pipapo_get(net, set, (const u8 *)elem->key.val.data,
611 nft_genmask_cur(net), get_jiffies_64());
619 * pipapo_resize() - Resize lookup or mapping table, or both
620 * @f: Field containing lookup and mapping tables
621 * @old_rules: Previous amount of rules in field
622 * @rules: New amount of rules
624 * Increase, decrease or maintain tables size depending on new amount of rules,
625 * and copy data over. In case the new size is smaller, throw away data for
626 * highest-numbered rules.
628 * Return: 0 on success, -ENOMEM on allocation failure.
630 static int pipapo_resize(struct nft_pipapo_field *f, int old_rules, int rules)
632 long *new_lt = NULL, *new_p, *old_lt = f->lt, *old_p;
633 union nft_pipapo_map_bucket *new_mt, *old_mt = f->mt;
634 size_t new_bucket_size, copy;
637 new_bucket_size = DIV_ROUND_UP(rules, BITS_PER_LONG);
638 #ifdef NFT_PIPAPO_ALIGN
639 new_bucket_size = roundup(new_bucket_size,
640 NFT_PIPAPO_ALIGN / sizeof(*new_lt));
643 if (new_bucket_size == f->bsize)
646 if (new_bucket_size > f->bsize)
649 copy = new_bucket_size;
651 new_lt = kvzalloc(f->groups * NFT_PIPAPO_BUCKETS(f->bb) *
652 new_bucket_size * sizeof(*new_lt) +
653 NFT_PIPAPO_ALIGN_HEADROOM,
658 new_p = NFT_PIPAPO_LT_ALIGN(new_lt);
659 old_p = NFT_PIPAPO_LT_ALIGN(old_lt);
661 for (group = 0; group < f->groups; group++) {
662 for (bucket = 0; bucket < NFT_PIPAPO_BUCKETS(f->bb); bucket++) {
663 memcpy(new_p, old_p, copy * sizeof(*new_p));
667 if (new_bucket_size > f->bsize)
668 new_p += new_bucket_size - f->bsize;
670 old_p += f->bsize - new_bucket_size;
675 new_mt = kvmalloc(rules * sizeof(*new_mt), GFP_KERNEL);
681 memcpy(new_mt, f->mt, min(old_rules, rules) * sizeof(*new_mt));
682 if (rules > old_rules) {
683 memset(new_mt + old_rules, 0,
684 (rules - old_rules) * sizeof(*new_mt));
688 f->bsize = new_bucket_size;
689 NFT_PIPAPO_LT_ASSIGN(f, new_lt);
700 * pipapo_bucket_set() - Set rule bit in bucket given group and group value
701 * @f: Field containing lookup table
703 * @group: Group index
704 * @v: Value of bit group
706 static void pipapo_bucket_set(struct nft_pipapo_field *f, int rule, int group,
711 pos = NFT_PIPAPO_LT_ALIGN(f->lt);
712 pos += f->bsize * NFT_PIPAPO_BUCKETS(f->bb) * group;
715 __set_bit(rule, pos);
719 * pipapo_lt_4b_to_8b() - Switch lookup table group width from 4 bits to 8 bits
720 * @old_groups: Number of current groups
721 * @bsize: Size of one bucket, in longs
722 * @old_lt: Pointer to the current lookup table
723 * @new_lt: Pointer to the new, pre-allocated lookup table
725 * Each bucket with index b in the new lookup table, belonging to group g, is
726 * filled with the bit intersection between:
727 * - bucket with index given by the upper 4 bits of b, from group g, and
728 * - bucket with index given by the lower 4 bits of b, from group g + 1
730 * That is, given buckets from the new lookup table N(x, y) and the old lookup
731 * table O(x, y), with x bucket index, and y group index:
733 * N(b, g) := O(b / 16, g) & O(b % 16, g + 1)
735 * This ensures equivalence of the matching results on lookup. Two examples in
739 * group 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 ... 254 255
746 * group 0 / 1 2 3 \ 4 5 6 7 8 9 10 11 12 13 |14 15 |
753 static void pipapo_lt_4b_to_8b(int old_groups, int bsize,
754 unsigned long *old_lt, unsigned long *new_lt)
758 for (g = 0; g < old_groups / 2; g++) {
759 int src_g0 = g * 2, src_g1 = g * 2 + 1;
761 for (b = 0; b < NFT_PIPAPO_BUCKETS(8); b++) {
762 int src_b0 = b / NFT_PIPAPO_BUCKETS(4);
763 int src_b1 = b % NFT_PIPAPO_BUCKETS(4);
764 int src_i0 = src_g0 * NFT_PIPAPO_BUCKETS(4) + src_b0;
765 int src_i1 = src_g1 * NFT_PIPAPO_BUCKETS(4) + src_b1;
767 for (i = 0; i < bsize; i++) {
768 *new_lt = old_lt[src_i0 * bsize + i] &
769 old_lt[src_i1 * bsize + i];
777 * pipapo_lt_8b_to_4b() - Switch lookup table group width from 8 bits to 4 bits
778 * @old_groups: Number of current groups
779 * @bsize: Size of one bucket, in longs
780 * @old_lt: Pointer to the current lookup table
781 * @new_lt: Pointer to the new, pre-allocated lookup table
783 * Each bucket with index b in the new lookup table, belonging to group g, is
784 * filled with the bit union of:
785 * - all the buckets with index such that the upper four bits of the lower byte
786 * equal b, from group g, with g odd
787 * - all the buckets with index such that the lower four bits equal b, from
788 * group g, with g even
790 * That is, given buckets from the new lookup table N(x, y) and the old lookup
791 * table O(x, y), with x bucket index, and y group index:
793 * - with g odd: N(b, g) := U(O(x, g) for each x : x = (b & 0xf0) >> 4)
794 * - with g even: N(b, g) := U(O(x, g) for each x : x = b & 0x0f)
796 * where U() denotes the arbitrary union operation (binary OR of n terms). This
797 * ensures equivalence of the matching results on lookup.
799 static void pipapo_lt_8b_to_4b(int old_groups, int bsize,
800 unsigned long *old_lt, unsigned long *new_lt)
804 memset(new_lt, 0, old_groups * 2 * NFT_PIPAPO_BUCKETS(4) * bsize *
805 sizeof(unsigned long));
807 for (g = 0; g < old_groups * 2; g += 2) {
810 for (b = 0; b < NFT_PIPAPO_BUCKETS(4); b++) {
811 for (bsrc = NFT_PIPAPO_BUCKETS(8) * src_g;
812 bsrc < NFT_PIPAPO_BUCKETS(8) * (src_g + 1);
814 if (((bsrc & 0xf0) >> 4) != b)
817 for (i = 0; i < bsize; i++)
818 new_lt[i] |= old_lt[bsrc * bsize + i];
824 for (b = 0; b < NFT_PIPAPO_BUCKETS(4); b++) {
825 for (bsrc = NFT_PIPAPO_BUCKETS(8) * src_g;
826 bsrc < NFT_PIPAPO_BUCKETS(8) * (src_g + 1);
828 if ((bsrc & 0x0f) != b)
831 for (i = 0; i < bsize; i++)
832 new_lt[i] |= old_lt[bsrc * bsize + i];
841 * pipapo_lt_bits_adjust() - Adjust group size for lookup table if needed
842 * @f: Field containing lookup table
844 static void pipapo_lt_bits_adjust(struct nft_pipapo_field *f)
846 unsigned long *new_lt;
850 lt_size = f->groups * NFT_PIPAPO_BUCKETS(f->bb) * f->bsize *
853 if (f->bb == NFT_PIPAPO_GROUP_BITS_SMALL_SET &&
854 lt_size > NFT_PIPAPO_LT_SIZE_HIGH) {
855 groups = f->groups * 2;
856 bb = NFT_PIPAPO_GROUP_BITS_LARGE_SET;
858 lt_size = groups * NFT_PIPAPO_BUCKETS(bb) * f->bsize *
860 } else if (f->bb == NFT_PIPAPO_GROUP_BITS_LARGE_SET &&
861 lt_size < NFT_PIPAPO_LT_SIZE_LOW) {
862 groups = f->groups / 2;
863 bb = NFT_PIPAPO_GROUP_BITS_SMALL_SET;
865 lt_size = groups * NFT_PIPAPO_BUCKETS(bb) * f->bsize *
868 /* Don't increase group width if the resulting lookup table size
869 * would exceed the upper size threshold for a "small" set.
871 if (lt_size > NFT_PIPAPO_LT_SIZE_HIGH)
877 new_lt = kvzalloc(lt_size + NFT_PIPAPO_ALIGN_HEADROOM, GFP_KERNEL);
881 NFT_PIPAPO_GROUP_BITS_ARE_8_OR_4;
882 if (f->bb == 4 && bb == 8) {
883 pipapo_lt_4b_to_8b(f->groups, f->bsize,
884 NFT_PIPAPO_LT_ALIGN(f->lt),
885 NFT_PIPAPO_LT_ALIGN(new_lt));
886 } else if (f->bb == 8 && bb == 4) {
887 pipapo_lt_8b_to_4b(f->groups, f->bsize,
888 NFT_PIPAPO_LT_ALIGN(f->lt),
889 NFT_PIPAPO_LT_ALIGN(new_lt));
897 NFT_PIPAPO_LT_ASSIGN(f, new_lt);
901 * pipapo_insert() - Insert new rule in field given input key and mask length
902 * @f: Field containing lookup table
903 * @k: Input key for classification, without nftables padding
904 * @mask_bits: Length of mask; matches field length for non-ranged entry
906 * Insert a new rule reference in lookup buckets corresponding to k and
909 * Return: 1 on success (one rule inserted), negative error code on failure.
911 static int pipapo_insert(struct nft_pipapo_field *f, const uint8_t *k,
914 int rule = f->rules, group, ret, bit_offset = 0;
916 ret = pipapo_resize(f, f->rules, f->rules + 1);
922 for (group = 0; group < f->groups; group++) {
926 v = k[group / (BITS_PER_BYTE / f->bb)];
927 v &= GENMASK(BITS_PER_BYTE - bit_offset - 1, 0);
928 v >>= (BITS_PER_BYTE - bit_offset) - f->bb;
931 bit_offset %= BITS_PER_BYTE;
933 if (mask_bits >= (group + 1) * f->bb) {
935 pipapo_bucket_set(f, rule, group, v);
936 } else if (mask_bits <= group * f->bb) {
937 /* Completely masked */
938 for (i = 0; i < NFT_PIPAPO_BUCKETS(f->bb); i++)
939 pipapo_bucket_set(f, rule, group, i);
941 /* The mask limit falls on this group */
942 mask = GENMASK(f->bb - 1, 0);
943 mask >>= mask_bits - group * f->bb;
944 for (i = 0; i < NFT_PIPAPO_BUCKETS(f->bb); i++) {
945 if ((i & ~mask) == (v & ~mask))
946 pipapo_bucket_set(f, rule, group, i);
951 pipapo_lt_bits_adjust(f);
957 * pipapo_step_diff() - Check if setting @step bit in netmask would change it
958 * @base: Mask we are expanding
959 * @step: Step bit for given expansion step
960 * @len: Total length of mask space (set and unset bits), bytes
962 * Convenience function for mask expansion.
964 * Return: true if step bit changes mask (i.e. isn't set), false otherwise.
966 static bool pipapo_step_diff(u8 *base, int step, int len)
968 /* Network order, byte-addressed */
969 #ifdef __BIG_ENDIAN__
970 return !(BIT(step % BITS_PER_BYTE) & base[step / BITS_PER_BYTE]);
972 return !(BIT(step % BITS_PER_BYTE) &
973 base[len - 1 - step / BITS_PER_BYTE]);
978 * pipapo_step_after_end() - Check if mask exceeds range end with given step
979 * @base: Mask we are expanding
981 * @step: Step bit for given expansion step, highest bit to be set
982 * @len: Total length of mask space (set and unset bits), bytes
984 * Convenience function for mask expansion.
986 * Return: true if mask exceeds range setting step bits, false otherwise.
988 static bool pipapo_step_after_end(const u8 *base, const u8 *end, int step,
991 u8 tmp[NFT_PIPAPO_MAX_BYTES];
994 memcpy(tmp, base, len);
996 /* Network order, byte-addressed */
997 for (i = 0; i <= step; i++)
998 #ifdef __BIG_ENDIAN__
999 tmp[i / BITS_PER_BYTE] |= BIT(i % BITS_PER_BYTE);
1001 tmp[len - 1 - i / BITS_PER_BYTE] |= BIT(i % BITS_PER_BYTE);
1004 return memcmp(tmp, end, len) > 0;
1008 * pipapo_base_sum() - Sum step bit to given len-sized netmask base with carry
1009 * @base: Netmask base
1010 * @step: Step bit to sum
1011 * @len: Netmask length, bytes
1013 static void pipapo_base_sum(u8 *base, int step, int len)
1018 /* Network order, byte-addressed */
1019 #ifdef __BIG_ENDIAN__
1020 for (i = step / BITS_PER_BYTE; i < len; i++) {
1022 for (i = len - 1 - step / BITS_PER_BYTE; i >= 0; i--) {
1027 base[i] += 1 << (step % BITS_PER_BYTE);
1037 * pipapo_expand() - Expand to composing netmasks, insert into lookup table
1038 * @f: Field containing lookup table
1039 * @start: Start of range
1040 * @end: End of range
1041 * @len: Length of value in bits
1043 * Expand range to composing netmasks and insert corresponding rule references
1044 * in lookup buckets.
1046 * Return: number of inserted rules on success, negative error code on failure.
1048 static int pipapo_expand(struct nft_pipapo_field *f,
1049 const u8 *start, const u8 *end, int len)
1051 int step, masks = 0, bytes = DIV_ROUND_UP(len, BITS_PER_BYTE);
1052 u8 base[NFT_PIPAPO_MAX_BYTES];
1054 memcpy(base, start, bytes);
1055 while (memcmp(base, end, bytes) <= 0) {
1059 while (pipapo_step_diff(base, step, bytes)) {
1060 if (pipapo_step_after_end(base, end, step, bytes))
1066 err = pipapo_insert(f, base, 0);
1075 err = pipapo_insert(f, base, len - step);
1081 pipapo_base_sum(base, step, bytes);
1088 * pipapo_map() - Insert rules in mapping tables, mapping them between fields
1089 * @m: Matching data, including mapping table
1090 * @map: Table of rule maps: array of first rule and amount of rules
1091 * in next field a given rule maps to, for each field
1092 * @e: For last field, nft_set_ext pointer matching rules map to
1094 static void pipapo_map(struct nft_pipapo_match *m,
1095 union nft_pipapo_map_bucket map[NFT_PIPAPO_MAX_FIELDS],
1096 struct nft_pipapo_elem *e)
1098 struct nft_pipapo_field *f;
1101 for (i = 0, f = m->f; i < m->field_count - 1; i++, f++) {
1102 for (j = 0; j < map[i].n; j++) {
1103 f->mt[map[i].to + j].to = map[i + 1].to;
1104 f->mt[map[i].to + j].n = map[i + 1].n;
1108 /* Last field: map to ext instead of mapping to next field */
1109 for (j = 0; j < map[i].n; j++)
1110 f->mt[map[i].to + j].e = e;
1114 * pipapo_free_scratch() - Free per-CPU map at original (not aligned) address
1118 static void pipapo_free_scratch(const struct nft_pipapo_match *m, unsigned int cpu)
1120 struct nft_pipapo_scratch *s;
1123 s = *per_cpu_ptr(m->scratch, cpu);
1128 mem -= s->align_off;
1133 * pipapo_realloc_scratch() - Reallocate scratch maps for partial match results
1134 * @clone: Copy of matching data with pending insertions and deletions
1135 * @bsize_max: Maximum bucket size, scratch maps cover two buckets
1137 * Return: 0 on success, -ENOMEM on failure.
1139 static int pipapo_realloc_scratch(struct nft_pipapo_match *clone,
1140 unsigned long bsize_max)
1144 for_each_possible_cpu(i) {
1145 struct nft_pipapo_scratch *scratch;
1146 #ifdef NFT_PIPAPO_ALIGN
1147 void *scratch_aligned;
1150 scratch = kzalloc_node(struct_size(scratch, map,
1152 NFT_PIPAPO_ALIGN_HEADROOM,
1153 GFP_KERNEL, cpu_to_node(i));
1155 /* On failure, there's no need to undo previous
1156 * allocations: this means that some scratch maps have
1157 * a bigger allocated size now (this is only called on
1158 * insertion), but the extra space won't be used by any
1159 * CPU as new elements are not inserted and m->bsize_max
1165 pipapo_free_scratch(clone, i);
1167 #ifdef NFT_PIPAPO_ALIGN
1168 /* Align &scratch->map (not the struct itself): the extra
1169 * %NFT_PIPAPO_ALIGN_HEADROOM bytes passed to kzalloc_node()
1170 * above guarantee we can waste up to those bytes in order
1171 * to align the map field regardless of its offset within
1174 BUILD_BUG_ON(offsetof(struct nft_pipapo_scratch, map) > NFT_PIPAPO_ALIGN_HEADROOM);
1176 scratch_aligned = NFT_PIPAPO_LT_ALIGN(&scratch->map);
1177 scratch_aligned -= offsetof(struct nft_pipapo_scratch, map);
1178 align_off = scratch_aligned - (void *)scratch;
1180 scratch = scratch_aligned;
1181 scratch->align_off = align_off;
1183 *per_cpu_ptr(clone->scratch, i) = scratch;
1190 * nft_pipapo_insert() - Validate and insert ranged elements
1191 * @net: Network namespace
1192 * @set: nftables API set representation
1193 * @elem: nftables API element representation containing key data
1194 * @elem_priv: Filled with pointer to &struct nft_set_ext in inserted element
1196 * Return: 0 on success, error pointer on failure.
1198 static int nft_pipapo_insert(const struct net *net, const struct nft_set *set,
1199 const struct nft_set_elem *elem,
1200 struct nft_elem_priv **elem_priv)
1202 const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
1203 union nft_pipapo_map_bucket rulemap[NFT_PIPAPO_MAX_FIELDS];
1204 const u8 *start = (const u8 *)elem->key.val.data, *end;
1205 struct nft_pipapo *priv = nft_set_priv(set);
1206 struct nft_pipapo_match *m = priv->clone;
1207 u8 genmask = nft_genmask_next(net);
1208 struct nft_pipapo_elem *e, *dup;
1209 u64 tstamp = nft_net_tstamp(net);
1210 struct nft_pipapo_field *f;
1211 const u8 *start_p, *end_p;
1212 int i, bsize_max, err = 0;
1214 if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY_END))
1215 end = (const u8 *)nft_set_ext_key_end(ext)->data;
1219 dup = pipapo_get(net, set, start, genmask, tstamp);
1221 /* Check if we already have the same exact entry */
1222 const struct nft_data *dup_key, *dup_end;
1224 dup_key = nft_set_ext_key(&dup->ext);
1225 if (nft_set_ext_exists(&dup->ext, NFT_SET_EXT_KEY_END))
1226 dup_end = nft_set_ext_key_end(&dup->ext);
1230 if (!memcmp(start, dup_key->data, sizeof(*dup_key->data)) &&
1231 !memcmp(end, dup_end->data, sizeof(*dup_end->data))) {
1232 *elem_priv = &dup->priv;
1239 if (PTR_ERR(dup) == -ENOENT) {
1240 /* Look for partially overlapping entries */
1241 dup = pipapo_get(net, set, end, nft_genmask_next(net), tstamp);
1244 if (PTR_ERR(dup) != -ENOENT) {
1246 return PTR_ERR(dup);
1247 *elem_priv = &dup->priv;
1254 nft_pipapo_for_each_field(f, i, m) {
1255 if (f->rules >= (unsigned long)NFT_PIPAPO_RULE0_MAX)
1258 if (memcmp(start_p, end_p,
1259 f->groups / NFT_PIPAPO_GROUPS_PER_BYTE(f)) > 0)
1262 start_p += NFT_PIPAPO_GROUPS_PADDED_SIZE(f);
1263 end_p += NFT_PIPAPO_GROUPS_PADDED_SIZE(f);
1269 bsize_max = m->bsize_max;
1271 nft_pipapo_for_each_field(f, i, m) {
1274 rulemap[i].to = f->rules;
1276 ret = memcmp(start, end,
1277 f->groups / NFT_PIPAPO_GROUPS_PER_BYTE(f));
1279 ret = pipapo_insert(f, start, f->groups * f->bb);
1281 ret = pipapo_expand(f, start, end, f->groups * f->bb);
1286 if (f->bsize > bsize_max)
1287 bsize_max = f->bsize;
1291 start += NFT_PIPAPO_GROUPS_PADDED_SIZE(f);
1292 end += NFT_PIPAPO_GROUPS_PADDED_SIZE(f);
1295 if (!*get_cpu_ptr(m->scratch) || bsize_max > m->bsize_max) {
1296 put_cpu_ptr(m->scratch);
1298 err = pipapo_realloc_scratch(m, bsize_max);
1302 m->bsize_max = bsize_max;
1304 put_cpu_ptr(m->scratch);
1307 e = nft_elem_priv_cast(elem->priv);
1308 *elem_priv = &e->priv;
1310 pipapo_map(m, rulemap, e);
1316 * pipapo_clone() - Clone matching data to create new working copy
1317 * @old: Existing matching data
1319 * Return: copy of matching data passed as 'old', error pointer on failure
1321 static struct nft_pipapo_match *pipapo_clone(struct nft_pipapo_match *old)
1323 struct nft_pipapo_field *dst, *src;
1324 struct nft_pipapo_match *new;
1327 new = kmalloc(struct_size(new, f, old->field_count), GFP_KERNEL);
1329 return ERR_PTR(-ENOMEM);
1331 new->field_count = old->field_count;
1332 new->bsize_max = old->bsize_max;
1334 new->scratch = alloc_percpu(*new->scratch);
1338 for_each_possible_cpu(i)
1339 *per_cpu_ptr(new->scratch, i) = NULL;
1341 if (pipapo_realloc_scratch(new, old->bsize_max))
1342 goto out_scratch_realloc;
1344 rcu_head_init(&new->rcu);
1349 for (i = 0; i < old->field_count; i++) {
1350 unsigned long *new_lt;
1352 memcpy(dst, src, offsetof(struct nft_pipapo_field, lt));
1354 new_lt = kvzalloc(src->groups * NFT_PIPAPO_BUCKETS(src->bb) *
1355 src->bsize * sizeof(*dst->lt) +
1356 NFT_PIPAPO_ALIGN_HEADROOM,
1361 NFT_PIPAPO_LT_ASSIGN(dst, new_lt);
1363 memcpy(NFT_PIPAPO_LT_ALIGN(new_lt),
1364 NFT_PIPAPO_LT_ALIGN(src->lt),
1365 src->bsize * sizeof(*dst->lt) *
1366 src->groups * NFT_PIPAPO_BUCKETS(src->bb));
1368 dst->mt = kvmalloc(src->rules * sizeof(*src->mt), GFP_KERNEL);
1372 memcpy(dst->mt, src->mt, src->rules * sizeof(*src->mt));
1382 for (dst--; i > 0; i--) {
1387 out_scratch_realloc:
1388 for_each_possible_cpu(i)
1389 pipapo_free_scratch(new, i);
1391 free_percpu(new->scratch);
1394 return ERR_PTR(-ENOMEM);
1398 * pipapo_rules_same_key() - Get number of rules originated from the same entry
1399 * @f: Field containing mapping table
1400 * @first: Index of first rule in set of rules mapping to same entry
1402 * Using the fact that all rules in a field that originated from the same entry
1403 * will map to the same set of rules in the next field, or to the same element
1404 * reference, return the cardinality of the set of rules that originated from
1405 * the same entry as the rule with index @first, @first rule included.
1409 * field #0 0 1 2 3 4
1410 * map to: 0 1 2-4 2-4 5-9
1416 * in field #1 0 1 2 3 4 5 ...
1418 * if this is called for rule 2 on field #0, it will return 3, as also rules 2
1419 * and 3 in field 0 map to the same set of rules (2, 3, 4) in the next field.
1421 * For the last field in a set, we can rely on associated entries to map to the
1422 * same element references.
1424 * Return: Number of rules that originated from the same entry as @first.
1426 static int pipapo_rules_same_key(struct nft_pipapo_field *f, int first)
1428 struct nft_pipapo_elem *e = NULL; /* Keep gcc happy */
1431 for (r = first; r < f->rules; r++) {
1432 if (r != first && e != f->mt[r].e)
1445 * pipapo_unmap() - Remove rules from mapping tables, renumber remaining ones
1446 * @mt: Mapping array
1447 * @rules: Original amount of rules in mapping table
1448 * @start: First rule index to be removed
1449 * @n: Amount of rules to be removed
1450 * @to_offset: First rule index, in next field, this group of rules maps to
1451 * @is_last: If this is the last field, delete reference from mapping array
1453 * This is used to unmap rules from the mapping table for a single field,
1454 * maintaining consistency and compactness for the existing ones.
1456 * In pictures: let's assume that we want to delete rules 2 and 3 from the
1457 * following mapping array:
1461 * map to: 4-10 4-10 11-15 11-15 16-18
1463 * the result will be:
1467 * map to: 4-10 4-10 11-13
1469 * for fields before the last one. In case this is the mapping table for the
1470 * last field in a set, and rules map to pointers to &struct nft_pipapo_elem:
1474 * element pointers: 0x42 0x42 0x33 0x33 0x44
1476 * the result will be:
1480 * element pointers: 0x42 0x42 0x44
1482 static void pipapo_unmap(union nft_pipapo_map_bucket *mt, int rules,
1483 int start, int n, int to_offset, bool is_last)
1487 memmove(mt + start, mt + start + n, (rules - start - n) * sizeof(*mt));
1488 memset(mt + rules - n, 0, n * sizeof(*mt));
1493 for (i = start; i < rules - n; i++)
1494 mt[i].to -= to_offset;
1498 * pipapo_drop() - Delete entry from lookup and mapping tables, given rule map
1500 * @rulemap: Table of rule maps, arrays of first rule and amount of rules
1501 * in next field a given entry maps to, for each field
1503 * For each rule in lookup table buckets mapping to this set of rules, drop
1504 * all bits set in lookup table mapping. In pictures, assuming we want to drop
1505 * rules 0 and 1 from this lookup table:
1508 * group 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
1515 * 6 0,1,2 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1516 * 7 1,2 1,2 1 1 1 0,1 1 1 1 1 1 1 1 1 1 1
1518 * rule 2 becomes rule 0, and the result will be:
1521 * group 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
1531 * once this is done, call unmap() to drop all the corresponding rule references
1532 * from mapping tables.
1534 static void pipapo_drop(struct nft_pipapo_match *m,
1535 union nft_pipapo_map_bucket rulemap[])
1537 struct nft_pipapo_field *f;
1540 nft_pipapo_for_each_field(f, i, m) {
1543 for (g = 0; g < f->groups; g++) {
1547 pos = NFT_PIPAPO_LT_ALIGN(f->lt) + g *
1548 NFT_PIPAPO_BUCKETS(f->bb) * f->bsize;
1550 for (b = 0; b < NFT_PIPAPO_BUCKETS(f->bb); b++) {
1551 bitmap_cut(pos, pos, rulemap[i].to,
1553 f->bsize * BITS_PER_LONG);
1559 pipapo_unmap(f->mt, f->rules, rulemap[i].to, rulemap[i].n,
1560 rulemap[i + 1].n, i == m->field_count - 1);
1561 if (pipapo_resize(f, f->rules, f->rules - rulemap[i].n)) {
1562 /* We can ignore this, a failure to shrink tables down
1563 * doesn't make tables invalid.
1567 f->rules -= rulemap[i].n;
1569 pipapo_lt_bits_adjust(f);
1573 static void nft_pipapo_gc_deactivate(struct net *net, struct nft_set *set,
1574 struct nft_pipapo_elem *e)
1577 nft_setelem_data_deactivate(net, set, &e->priv);
1581 * pipapo_gc() - Drop expired entries from set, destroy start and end elements
1582 * @set: nftables API set representation
1585 static void pipapo_gc(struct nft_set *set, struct nft_pipapo_match *m)
1587 struct nft_pipapo *priv = nft_set_priv(set);
1588 struct net *net = read_pnet(&set->net);
1589 u64 tstamp = nft_net_tstamp(net);
1590 int rules_f0, first_rule = 0;
1591 struct nft_pipapo_elem *e;
1592 struct nft_trans_gc *gc;
1594 gc = nft_trans_gc_alloc(set, 0, GFP_KERNEL);
1598 while ((rules_f0 = pipapo_rules_same_key(m->f, first_rule))) {
1599 union nft_pipapo_map_bucket rulemap[NFT_PIPAPO_MAX_FIELDS];
1600 struct nft_pipapo_field *f;
1601 int i, start, rules_fx;
1604 rules_fx = rules_f0;
1606 nft_pipapo_for_each_field(f, i, m) {
1607 rulemap[i].to = start;
1608 rulemap[i].n = rules_fx;
1610 if (i < m->field_count - 1) {
1611 rules_fx = f->mt[start].n;
1612 start = f->mt[start].to;
1616 /* Pick the last field, and its last index */
1619 e = f->mt[rulemap[i].to].e;
1621 /* synchronous gc never fails, there is no need to set on
1622 * NFT_SET_ELEM_DEAD_BIT.
1624 if (__nft_set_elem_expired(&e->ext, tstamp)) {
1627 gc = nft_trans_gc_queue_sync(gc, GFP_KERNEL);
1631 nft_pipapo_gc_deactivate(net, set, e);
1632 pipapo_drop(m, rulemap);
1633 nft_trans_gc_elem_add(gc, e);
1635 /* And check again current first rule, which is now the
1636 * first we haven't checked.
1639 first_rule += rules_f0;
1643 gc = nft_trans_gc_catchall_sync(gc);
1645 nft_trans_gc_queue_sync_done(gc);
1646 priv->last_gc = jiffies;
1651 * pipapo_free_fields() - Free per-field tables contained in matching data
1654 static void pipapo_free_fields(struct nft_pipapo_match *m)
1656 struct nft_pipapo_field *f;
1659 nft_pipapo_for_each_field(f, i, m) {
1665 static void pipapo_free_match(struct nft_pipapo_match *m)
1669 for_each_possible_cpu(i)
1670 pipapo_free_scratch(m, i);
1672 free_percpu(m->scratch);
1673 pipapo_free_fields(m);
1679 * pipapo_reclaim_match - RCU callback to free fields from old matching data
1682 static void pipapo_reclaim_match(struct rcu_head *rcu)
1684 struct nft_pipapo_match *m;
1686 m = container_of(rcu, struct nft_pipapo_match, rcu);
1687 pipapo_free_match(m);
1691 * nft_pipapo_commit() - Replace lookup data with current working copy
1692 * @set: nftables API set representation
1694 * While at it, check if we should perform garbage collection on the working
1695 * copy before committing it for lookup, and don't replace the table if the
1696 * working copy doesn't have pending changes.
1698 * We also need to create a new working copy for subsequent insertions and
1701 static void nft_pipapo_commit(struct nft_set *set)
1703 struct nft_pipapo *priv = nft_set_priv(set);
1704 struct nft_pipapo_match *new_clone, *old;
1706 if (time_after_eq(jiffies, priv->last_gc + nft_set_gc_interval(set)))
1707 pipapo_gc(set, priv->clone);
1712 new_clone = pipapo_clone(priv->clone);
1713 if (IS_ERR(new_clone))
1716 priv->dirty = false;
1718 old = rcu_access_pointer(priv->match);
1719 rcu_assign_pointer(priv->match, priv->clone);
1721 call_rcu(&old->rcu, pipapo_reclaim_match);
1723 priv->clone = new_clone;
1726 static bool nft_pipapo_transaction_mutex_held(const struct nft_set *set)
1728 #ifdef CONFIG_PROVE_LOCKING
1729 const struct net *net = read_pnet(&set->net);
1731 return lockdep_is_held(&nft_pernet(net)->commit_mutex);
1737 static void nft_pipapo_abort(const struct nft_set *set)
1739 struct nft_pipapo *priv = nft_set_priv(set);
1740 struct nft_pipapo_match *new_clone, *m;
1745 m = rcu_dereference_protected(priv->match, nft_pipapo_transaction_mutex_held(set));
1747 new_clone = pipapo_clone(m);
1748 if (IS_ERR(new_clone))
1751 priv->dirty = false;
1753 pipapo_free_match(priv->clone);
1754 priv->clone = new_clone;
1758 * nft_pipapo_activate() - Mark element reference as active given key, commit
1759 * @net: Network namespace
1760 * @set: nftables API set representation
1761 * @elem_priv: nftables API element representation containing key data
1763 * On insertion, elements are added to a copy of the matching data currently
1764 * in use for lookups, and not directly inserted into current lookup data. Both
1765 * nft_pipapo_insert() and nft_pipapo_activate() are called once for each
1766 * element, hence we can't purpose either one as a real commit operation.
1768 static void nft_pipapo_activate(const struct net *net,
1769 const struct nft_set *set,
1770 struct nft_elem_priv *elem_priv)
1772 struct nft_pipapo_elem *e = nft_elem_priv_cast(elem_priv);
1774 nft_set_elem_change_active(net, set, &e->ext);
1778 * pipapo_deactivate() - Check that element is in set, mark as inactive
1779 * @net: Network namespace
1780 * @set: nftables API set representation
1781 * @data: Input key data
1782 * @ext: nftables API extension pointer, used to check for end element
1784 * This is a convenience function that can be called from both
1785 * nft_pipapo_deactivate() and nft_pipapo_flush(), as they are in fact the same
1788 * Return: deactivated element if found, NULL otherwise.
1790 static void *pipapo_deactivate(const struct net *net, const struct nft_set *set,
1791 const u8 *data, const struct nft_set_ext *ext)
1793 struct nft_pipapo_elem *e;
1795 e = pipapo_get(net, set, data, nft_genmask_next(net), nft_net_tstamp(net));
1799 nft_set_elem_change_active(net, set, &e->ext);
1805 * nft_pipapo_deactivate() - Call pipapo_deactivate() to make element inactive
1806 * @net: Network namespace
1807 * @set: nftables API set representation
1808 * @elem: nftables API element representation containing key data
1810 * Return: deactivated element if found, NULL otherwise.
1812 static struct nft_elem_priv *
1813 nft_pipapo_deactivate(const struct net *net, const struct nft_set *set,
1814 const struct nft_set_elem *elem)
1816 const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
1818 return pipapo_deactivate(net, set, (const u8 *)elem->key.val.data, ext);
1822 * nft_pipapo_flush() - Call pipapo_deactivate() to make element inactive
1823 * @net: Network namespace
1824 * @set: nftables API set representation
1825 * @elem_priv: nftables API element representation containing key data
1827 * This is functionally the same as nft_pipapo_deactivate(), with a slightly
1828 * different interface, and it's also called once for each element in a set
1829 * being flushed, so we can't implement, strictly speaking, a flush operation,
1830 * which would otherwise be as simple as allocating an empty copy of the
1833 * Note that we could in theory do that, mark the set as flushed, and ignore
1834 * subsequent calls, but we would leak all the elements after the first one,
1835 * because they wouldn't then be freed as result of API calls.
1837 * Return: true if element was found and deactivated.
1839 static void nft_pipapo_flush(const struct net *net, const struct nft_set *set,
1840 struct nft_elem_priv *elem_priv)
1842 struct nft_pipapo_elem *e = nft_elem_priv_cast(elem_priv);
1844 nft_set_elem_change_active(net, set, &e->ext);
1848 * pipapo_get_boundaries() - Get byte interval for associated rules
1849 * @f: Field including lookup table
1850 * @first_rule: First rule (lowest index)
1851 * @rule_count: Number of associated rules
1852 * @left: Byte expression for left boundary (start of range)
1853 * @right: Byte expression for right boundary (end of range)
1855 * Given the first rule and amount of rules that originated from the same entry,
1856 * build the original range associated with the entry, and calculate the length
1857 * of the originating netmask.
1862 * group 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
1869 * 6 1,2 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1870 * 7 1,2 1,2 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1872 * this is the lookup table corresponding to the IPv4 range
1873 * 192.168.1.0-192.168.2.1, which was expanded to the two composing netmasks,
1874 * rule #1: 192.168.1.0/24, and rule #2: 192.168.2.0/31.
1876 * This function fills @left and @right with the byte values of the leftmost
1877 * and rightmost bucket indices for the lowest and highest rule indices,
1878 * respectively. If @first_rule is 1 and @rule_count is 2, we obtain, in
1880 * left: < 12, 0, 10, 8, 0, 1, 0, 0 >
1881 * right: < 12, 0, 10, 8, 0, 2, 2, 1 >
1882 * corresponding to bytes:
1883 * left: < 192, 168, 1, 0 >
1884 * right: < 192, 168, 2, 1 >
1885 * with mask length irrelevant here, unused on return, as the range is already
1886 * defined by its start and end points. The mask length is relevant for a single
1887 * ranged entry instead: if @first_rule is 1 and @rule_count is 1, we ignore
1888 * rule 2 above: @left becomes < 192, 168, 1, 0 >, @right becomes
1889 * < 192, 168, 1, 255 >, and the mask length, calculated from the distances
1890 * between leftmost and rightmost bucket indices for each group, would be 24.
1892 * Return: mask length, in bits.
1894 static int pipapo_get_boundaries(struct nft_pipapo_field *f, int first_rule,
1895 int rule_count, u8 *left, u8 *right)
1897 int g, mask_len = 0, bit_offset = 0;
1898 u8 *l = left, *r = right;
1900 for (g = 0; g < f->groups; g++) {
1905 for (b = 0; b < NFT_PIPAPO_BUCKETS(f->bb); b++) {
1908 pos = NFT_PIPAPO_LT_ALIGN(f->lt) +
1909 (g * NFT_PIPAPO_BUCKETS(f->bb) + b) * f->bsize;
1910 if (test_bit(first_rule, pos) && x0 == -1)
1912 if (test_bit(first_rule + rule_count - 1, pos))
1916 *l |= x0 << (BITS_PER_BYTE - f->bb - bit_offset);
1917 *r |= x1 << (BITS_PER_BYTE - f->bb - bit_offset);
1919 bit_offset += f->bb;
1920 if (bit_offset >= BITS_PER_BYTE) {
1921 bit_offset %= BITS_PER_BYTE;
1928 else if (x1 - x0 == 1)
1930 else if (x1 - x0 == 3)
1932 else if (x1 - x0 == 7)
1940 * pipapo_match_field() - Match rules against byte ranges
1941 * @f: Field including the lookup table
1942 * @first_rule: First of associated rules originating from same entry
1943 * @rule_count: Amount of associated rules
1944 * @start: Start of range to be matched
1945 * @end: End of range to be matched
1947 * Return: true on match, false otherwise.
1949 static bool pipapo_match_field(struct nft_pipapo_field *f,
1950 int first_rule, int rule_count,
1951 const u8 *start, const u8 *end)
1953 u8 right[NFT_PIPAPO_MAX_BYTES] = { 0 };
1954 u8 left[NFT_PIPAPO_MAX_BYTES] = { 0 };
1956 pipapo_get_boundaries(f, first_rule, rule_count, left, right);
1958 return !memcmp(start, left,
1959 f->groups / NFT_PIPAPO_GROUPS_PER_BYTE(f)) &&
1960 !memcmp(end, right, f->groups / NFT_PIPAPO_GROUPS_PER_BYTE(f));
1964 * nft_pipapo_remove() - Remove element given key, commit
1965 * @net: Network namespace
1966 * @set: nftables API set representation
1967 * @elem_priv: nftables API element representation containing key data
1969 * Similarly to nft_pipapo_activate(), this is used as commit operation by the
1970 * API, but it's called once per element in the pending transaction, so we can't
1971 * implement this as a single commit operation. Closest we can get is to remove
1972 * the matched element here, if any, and commit the updated matching data.
1974 static void nft_pipapo_remove(const struct net *net, const struct nft_set *set,
1975 struct nft_elem_priv *elem_priv)
1977 struct nft_pipapo *priv = nft_set_priv(set);
1978 struct nft_pipapo_match *m = priv->clone;
1979 int rules_f0, first_rule = 0;
1980 struct nft_pipapo_elem *e;
1983 e = nft_elem_priv_cast(elem_priv);
1984 data = (const u8 *)nft_set_ext_key(&e->ext);
1986 while ((rules_f0 = pipapo_rules_same_key(m->f, first_rule))) {
1987 union nft_pipapo_map_bucket rulemap[NFT_PIPAPO_MAX_FIELDS];
1988 const u8 *match_start, *match_end;
1989 struct nft_pipapo_field *f;
1990 int i, start, rules_fx;
1994 if (nft_set_ext_exists(&e->ext, NFT_SET_EXT_KEY_END))
1995 match_end = (const u8 *)nft_set_ext_key_end(&e->ext)->data;
2000 rules_fx = rules_f0;
2002 nft_pipapo_for_each_field(f, i, m) {
2003 if (!pipapo_match_field(f, start, rules_fx,
2004 match_start, match_end))
2007 rulemap[i].to = start;
2008 rulemap[i].n = rules_fx;
2010 rules_fx = f->mt[start].n;
2011 start = f->mt[start].to;
2013 match_start += NFT_PIPAPO_GROUPS_PADDED_SIZE(f);
2014 match_end += NFT_PIPAPO_GROUPS_PADDED_SIZE(f);
2017 if (i == m->field_count) {
2019 pipapo_drop(m, rulemap);
2023 first_rule += rules_f0;
2028 * nft_pipapo_walk() - Walk over elements
2029 * @ctx: nftables API context
2030 * @set: nftables API set representation
2033 * As elements are referenced in the mapping array for the last field, directly
2034 * scan that array: there's no need to follow rule mappings from the first
2037 static void nft_pipapo_walk(const struct nft_ctx *ctx, struct nft_set *set,
2038 struct nft_set_iter *iter)
2040 struct nft_pipapo *priv = nft_set_priv(set);
2041 struct net *net = read_pnet(&set->net);
2042 struct nft_pipapo_match *m;
2043 struct nft_pipapo_field *f;
2047 if (iter->genmask == nft_genmask_cur(net))
2048 m = rcu_dereference(priv->match);
2055 for (i = 0, f = m->f; i < m->field_count - 1; i++, f++)
2058 for (r = 0; r < f->rules; r++) {
2059 struct nft_pipapo_elem *e;
2061 if (r < f->rules - 1 && f->mt[r + 1].e == f->mt[r].e)
2064 if (iter->count < iter->skip)
2069 if (!nft_set_elem_active(&e->ext, iter->genmask))
2072 iter->err = iter->fn(ctx, set, iter, &e->priv);
2085 * nft_pipapo_privsize() - Return the size of private data for the set
2086 * @nla: netlink attributes, ignored as size doesn't depend on them
2087 * @desc: Set description, ignored as size doesn't depend on it
2089 * Return: size of private data for this set implementation, in bytes
2091 static u64 nft_pipapo_privsize(const struct nlattr * const nla[],
2092 const struct nft_set_desc *desc)
2094 return sizeof(struct nft_pipapo);
2098 * nft_pipapo_estimate() - Set size, space and lookup complexity
2099 * @desc: Set description, element count and field description used
2100 * @features: Flags: NFT_SET_INTERVAL needs to be there
2101 * @est: Storage for estimation data
2103 * Return: true if set description is compatible, false otherwise
2105 static bool nft_pipapo_estimate(const struct nft_set_desc *desc, u32 features,
2106 struct nft_set_estimate *est)
2108 if (!(features & NFT_SET_INTERVAL) ||
2109 desc->field_count < NFT_PIPAPO_MIN_FIELDS)
2112 est->size = pipapo_estimate_size(desc);
2116 est->lookup = NFT_SET_CLASS_O_LOG_N;
2118 est->space = NFT_SET_CLASS_O_N;
2124 * nft_pipapo_init() - Initialise data for a set instance
2125 * @set: nftables API set representation
2126 * @desc: Set description
2127 * @nla: netlink attributes
2129 * Validate number and size of fields passed as NFTA_SET_DESC_CONCAT netlink
2130 * attributes, initialise internal set parameters, current instance of matching
2131 * data and a copy for subsequent insertions.
2133 * Return: 0 on success, negative error code on failure.
2135 static int nft_pipapo_init(const struct nft_set *set,
2136 const struct nft_set_desc *desc,
2137 const struct nlattr * const nla[])
2139 struct nft_pipapo *priv = nft_set_priv(set);
2140 struct nft_pipapo_match *m;
2141 struct nft_pipapo_field *f;
2142 int err, i, field_count;
2144 BUILD_BUG_ON(offsetof(struct nft_pipapo_elem, priv) != 0);
2146 field_count = desc->field_count ? : 1;
2148 if (field_count > NFT_PIPAPO_MAX_FIELDS)
2151 m = kmalloc(struct_size(m, f, field_count), GFP_KERNEL);
2155 m->field_count = field_count;
2158 m->scratch = alloc_percpu(struct nft_pipapo_scratch *);
2163 for_each_possible_cpu(i)
2164 *per_cpu_ptr(m->scratch, i) = NULL;
2166 rcu_head_init(&m->rcu);
2168 nft_pipapo_for_each_field(f, i, m) {
2169 int len = desc->field_len[i] ? : set->klen;
2171 f->bb = NFT_PIPAPO_GROUP_BITS_INIT;
2172 f->groups = len * NFT_PIPAPO_GROUPS_PER_BYTE(f);
2174 priv->width += round_up(len, sizeof(u32));
2178 NFT_PIPAPO_LT_ASSIGN(f, NULL);
2182 /* Create an initial clone of matching data for next insertion */
2183 priv->clone = pipapo_clone(m);
2184 if (IS_ERR(priv->clone)) {
2185 err = PTR_ERR(priv->clone);
2189 priv->dirty = false;
2191 rcu_assign_pointer(priv->match, m);
2196 free_percpu(m->scratch);
2204 * nft_set_pipapo_match_destroy() - Destroy elements from key mapping array
2206 * @set: nftables API set representation
2207 * @m: matching data pointing to key mapping array
2209 static void nft_set_pipapo_match_destroy(const struct nft_ctx *ctx,
2210 const struct nft_set *set,
2211 struct nft_pipapo_match *m)
2213 struct nft_pipapo_field *f;
2216 for (i = 0, f = m->f; i < m->field_count - 1; i++, f++)
2219 for (r = 0; r < f->rules; r++) {
2220 struct nft_pipapo_elem *e;
2222 if (r < f->rules - 1 && f->mt[r + 1].e == f->mt[r].e)
2227 nf_tables_set_elem_destroy(ctx, set, &e->priv);
2232 * nft_pipapo_destroy() - Free private data for set and all committed elements
2234 * @set: nftables API set representation
2236 static void nft_pipapo_destroy(const struct nft_ctx *ctx,
2237 const struct nft_set *set)
2239 struct nft_pipapo *priv = nft_set_priv(set);
2240 struct nft_pipapo_match *m;
2243 m = rcu_dereference_protected(priv->match, true);
2247 for_each_possible_cpu(cpu)
2248 pipapo_free_scratch(m, cpu);
2249 free_percpu(m->scratch);
2250 pipapo_free_fields(m);
2258 nft_set_pipapo_match_destroy(ctx, set, m);
2260 for_each_possible_cpu(cpu)
2261 pipapo_free_scratch(priv->clone, cpu);
2262 free_percpu(priv->clone->scratch);
2264 pipapo_free_fields(priv->clone);
2271 * nft_pipapo_gc_init() - Initialise garbage collection
2272 * @set: nftables API set representation
2274 * Instead of actually setting up a periodic work for garbage collection, as
2275 * this operation requires a swap of matching data with the working copy, we'll
2276 * do that opportunistically with other commit operations if the interval is
2277 * elapsed, so we just need to set the current jiffies timestamp here.
2279 static void nft_pipapo_gc_init(const struct nft_set *set)
2281 struct nft_pipapo *priv = nft_set_priv(set);
2283 priv->last_gc = jiffies;
2286 const struct nft_set_type nft_set_pipapo_type = {
2287 .features = NFT_SET_INTERVAL | NFT_SET_MAP | NFT_SET_OBJECT |
2290 .lookup = nft_pipapo_lookup,
2291 .insert = nft_pipapo_insert,
2292 .activate = nft_pipapo_activate,
2293 .deactivate = nft_pipapo_deactivate,
2294 .flush = nft_pipapo_flush,
2295 .remove = nft_pipapo_remove,
2296 .walk = nft_pipapo_walk,
2297 .get = nft_pipapo_get,
2298 .privsize = nft_pipapo_privsize,
2299 .estimate = nft_pipapo_estimate,
2300 .init = nft_pipapo_init,
2301 .destroy = nft_pipapo_destroy,
2302 .gc_init = nft_pipapo_gc_init,
2303 .commit = nft_pipapo_commit,
2304 .abort = nft_pipapo_abort,
2305 .elemsize = offsetof(struct nft_pipapo_elem, ext),
2309 #if defined(CONFIG_X86_64) && !defined(CONFIG_UML)
2310 const struct nft_set_type nft_set_pipapo_avx2_type = {
2311 .features = NFT_SET_INTERVAL | NFT_SET_MAP | NFT_SET_OBJECT |
2314 .lookup = nft_pipapo_avx2_lookup,
2315 .insert = nft_pipapo_insert,
2316 .activate = nft_pipapo_activate,
2317 .deactivate = nft_pipapo_deactivate,
2318 .flush = nft_pipapo_flush,
2319 .remove = nft_pipapo_remove,
2320 .walk = nft_pipapo_walk,
2321 .get = nft_pipapo_get,
2322 .privsize = nft_pipapo_privsize,
2323 .estimate = nft_pipapo_avx2_estimate,
2324 .init = nft_pipapo_init,
2325 .destroy = nft_pipapo_destroy,
2326 .gc_init = nft_pipapo_gc_init,
2327 .commit = nft_pipapo_commit,
2328 .abort = nft_pipapo_abort,
2329 .elemsize = offsetof(struct nft_pipapo_elem, ext),