1 // SPDX-License-Identifier: GPL-2.0-only
2
3 /* PIPAPO: PIle PAcket POlicies: set for arbitrary concatenations of ranges
4 *
5 * Copyright (c) 2019-2020 Red Hat GmbH
6 *
7 * Author: Stefano Brivio <sbrivio@redhat.com>
8 */
9
10 /**
11 * DOC: Theory of Operation
12 *
13 *
14 * Problem
15 * -------
16 *
17 * Match packet bytes against entries composed of ranged or non-ranged packet
18 * field specifiers, mapping them to arbitrary references. For example:
19 *
20 * ::
21 *
22 * --- fields --->
23 * | [net],[port],[net]... => [reference]
24 * entries [net],[port],[net]... => [reference]
25 * | [net],[port],[net]... => [reference]
26 * V ...
27 *
28 * where [net] fields can be IP ranges or netmasks, and [port] fields are port
29 * ranges. Arbitrary packet fields can be matched.
30 *
31 *
32 * Algorithm Overview
33 * ------------------
34 *
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].
39 *
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.
43 *
44 * Example:
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:
47 * 0000 0001 0101 1001
48 * and a packet with source port:
49 * 0000 0001 1010 1001
50 * first and second groups match, but the third doesn't. We conclude that the
51 * packet doesn't match the given entry.
52 *
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.
60 *
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.
66 *
67 * The mapping array for the last field maps to the desired references.
68 *
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
71 * matching.
72 *
73 * A stand-alone, reference implementation, also including notes about possible
74 * future optimisations, is available at:
75 * https://pipapo.lameexcu.se/
76 *
77 * Insertion
78 * ---------
79 *
80 * - For each packet field:
81 *
82 * - divide the b packet bits we want to classify into groups of size t,
83 * obtaining ceil(b / t) groups
84 *
85 * Example: match on destination IP address, with t = 4: 32 bits, 8 groups
86 * of 4 bits each
87 *
88 * - allocate a lookup table with one column ("bucket") for each possible
89 * value of a group, and with one row for each group
90 *
91 * Example: 8 groups, 2^4 buckets:
92 *
93 * ::
94 *
95 * bucket
96 * group 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
97 * 0
98 * 1
99 * 2
100 * 3
101 * 4
102 * 5
103 * 6
104 * 7
105 *
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().
110 *
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
115 *
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
118 * pipapo_insert().
119 *
120 * Example: given:
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 > >
127 *
128 * these bits are set in the lookup table:
129 *
130 * ::
131 *
132 * bucket
133 * group 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
134 * 0 0 1,2
135 * 1 1,2 0
136 * 2 0 1,2
137 * 3 0 1,2
138 * 4 0,1,2
139 * 5 0 1 2
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
142 *
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().
146 *
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.
150 *
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):
153 *
154 * ::
155 *
156 * bucket
157 * group 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
158 * 0 0 1,2
159 * 1 1,2 0
160 * 2 0 1,2
161 * 3 0 1,2
162 * 4 0,1,2
163 * 5 0 1 2
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
166 *
167 * and lookup table #1 for field 1 with:
168 * - rule #0: 1024 mapping to buckets
169 * < 0 0 4 0 >
170 * - rule #1: 2048 mapping to buckets
171 * < 0 0 5 0 >
172 *
173 * ::
174 *
175 * bucket
176 * group 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
177 * 0 0,1
178 * 1 0,1
179 * 2 0 1
180 * 3 0,1
181 *
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):
185 *
186 * ::
187 *
188 * rule indices in current field: 0 1 2
189 * map to rules in next field: 0 1 1
190 *
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
193 * pipapo_map().
194 *
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.
198 *
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:
202 *
203 * ::
204 *
205 * rule indices in last field: 0 1
206 * map to elements: 0x66 0x42
207 *
208 *
209 * Matching
210 * --------
211 *
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().
215 *
216 * - For each packet field:
217 *
218 * - start with an all-ones result bitmap (res_map in pipapo_lookup())
219 *
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
223 *
224 * ::
225 *
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:
231 *
232 * bucket
233 * group 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
234 * 0 0 1,2
235 * result bitmap is now: 0xff & 0x6 [bucket 12] = 0x6
236 *
237 * 1 1,2 0
238 * result bitmap is now: 0x6 & 0x6 [bucket 0] = 0x6
239 *
240 * 2 0 1,2
241 * result bitmap is now: 0x6 & 0x6 [bucket 10] = 0x6
242 *
243 * 3 0 1,2
244 * result bitmap is now: 0x6 & 0x6 [bucket 8] = 0x6
245 *
246 * 4 0,1,2
247 * result bitmap is now: 0x6 & 0x7 [bucket 0] = 0x6
248 *
249 * 5 0 1 2
250 * result bitmap is now: 0x6 & 0x2 [bucket 1] = 0x2
251 *
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
254 *
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
257 *
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
262 * pipapo_refill()
263 *
264 * Example: with mapping table from insertion examples, with the current
265 * result bitmap from the previous example, 0x02:
266 *
267 * ::
268 *
269 * rule indices in current field: 0 1 2
270 * map to rules in next field: 0 1 1
271 *
272 * the new result bitmap will be 0x02: rule 1 was set, and rule 1 will be
273 * set.
274 *
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:
279 *
280 * ::
281 *
282 * bucket
283 * group 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
284 * 0 0,1
285 * 1 0,1
286 * 2 0 1
287 * 3 0,1
288 *
289 * operations are: 0x2 & 0x3 [bucket 0] & 0x3 [bucket 0] & 0x2 [bucket 5]
290 * & 0x3 [bucket 0], resulting bitmap is 0x2.
291 *
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
294 *
295 * Example: 0x2 resulting bitmap from 192.168.1.5:2048, mapping array for
296 * last field from insertion example:
297 *
298 * ::
299 *
300 * rule indices in last field: 0 1
301 * map to elements: 0x66 0x42
302 *
303 * the matching element is at 0x42.
304 *
305 *
306 * References
307 * ----------
308 *
309 * [Ligatti 2010]
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
316 *
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
322 *
323 * [Kogan 2014]
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
329 */
330
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>
341
342 #include "nft_set_pipapo_avx2.h"
343 #include "nft_set_pipapo.h"
344
345 /**
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
353 *
354 * Iteration over set bits with __builtin_ctzl(): Daniel Lemire, public domain.
355 *
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.
359 *
360 * Return: -1 on no match, bit position on 'match_only', 0 otherwise.
361 */
pipapo_refill(unsigned long * map,unsigned int len,unsigned int rules,unsigned long * dst,const union nft_pipapo_map_bucket * mt,bool match_only)362 int pipapo_refill(unsigned long *map, unsigned int len, unsigned int rules,
363 unsigned long *dst,
364 const union nft_pipapo_map_bucket *mt, bool match_only)
365 {
366 unsigned long bitset;
367 unsigned int k;
368 int ret = -1;
369
370 for (k = 0; k < len; k++) {
371 bitset = map[k];
372 while (bitset) {
373 unsigned long t = bitset & -bitset;
374 int r = __builtin_ctzl(bitset);
375 int i = k * BITS_PER_LONG + r;
376
377 if (unlikely(i >= rules)) {
378 map[k] = 0;
379 return -1;
380 }
381
382 if (match_only) {
383 bitmap_clear(map, i, 1);
384 return i;
385 }
386
387 ret = 0;
388
389 bitmap_set(dst, mt[i].to, mt[i].n);
390
391 bitset ^= t;
392 }
393 map[k] = 0;
394 }
395
396 return ret;
397 }
398
399 /**
400 * nft_pipapo_lookup() - Lookup function
401 * @net: Network namespace
402 * @set: nftables API set representation
403 * @key: nftables API element representation containing key data
404 * @ext: nftables API extension pointer, filled with matching reference
405 *
406 * For more details, see DOC: Theory of Operation.
407 *
408 * Return: true on match, false otherwise.
409 */
nft_pipapo_lookup(const struct net * net,const struct nft_set * set,const u32 * key,const struct nft_set_ext ** ext)410 bool nft_pipapo_lookup(const struct net *net, const struct nft_set *set,
411 const u32 *key, const struct nft_set_ext **ext)
412 {
413 struct nft_pipapo *priv = nft_set_priv(set);
414 struct nft_pipapo_scratch *scratch;
415 unsigned long *res_map, *fill_map;
416 u8 genmask = nft_genmask_cur(net);
417 const struct nft_pipapo_match *m;
418 const struct nft_pipapo_field *f;
419 const u8 *rp = (const u8 *)key;
420 bool map_index;
421 int i;
422
423 local_bh_disable();
424
425 m = rcu_dereference(priv->match);
426
427 if (unlikely(!m || !*raw_cpu_ptr(m->scratch)))
428 goto out;
429
430 scratch = *raw_cpu_ptr(m->scratch);
431
432 map_index = scratch->map_index;
433
434 res_map = scratch->map + (map_index ? m->bsize_max : 0);
435 fill_map = scratch->map + (map_index ? 0 : m->bsize_max);
436
437 pipapo_resmap_init(m, res_map);
438
439 nft_pipapo_for_each_field(f, i, m) {
440 bool last = i == m->field_count - 1;
441 int b;
442
443 /* For each bit group: select lookup table bucket depending on
444 * packet bytes value, then AND bucket value
445 */
446 if (likely(f->bb == 8))
447 pipapo_and_field_buckets_8bit(f, res_map, rp);
448 else
449 pipapo_and_field_buckets_4bit(f, res_map, rp);
450 NFT_PIPAPO_GROUP_BITS_ARE_8_OR_4;
451
452 rp += f->groups / NFT_PIPAPO_GROUPS_PER_BYTE(f);
453
454 /* Now populate the bitmap for the next field, unless this is
455 * the last field, in which case return the matched 'ext'
456 * pointer if any.
457 *
458 * Now res_map contains the matching bitmap, and fill_map is the
459 * bitmap for the next field.
460 */
461 next_match:
462 b = pipapo_refill(res_map, f->bsize, f->rules, fill_map, f->mt,
463 last);
464 if (b < 0) {
465 scratch->map_index = map_index;
466 local_bh_enable();
467
468 return false;
469 }
470
471 if (last) {
472 *ext = &f->mt[b].e->ext;
473 if (unlikely(nft_set_elem_expired(*ext) ||
474 !nft_set_elem_active(*ext, genmask)))
475 goto next_match;
476
477 /* Last field: we're just returning the key without
478 * filling the initial bitmap for the next field, so the
479 * current inactive bitmap is clean and can be reused as
480 * *next* bitmap (not initial) for the next packet.
481 */
482 scratch->map_index = map_index;
483 local_bh_enable();
484
485 return true;
486 }
487
488 /* Swap bitmap indices: res_map is the initial bitmap for the
489 * next field, and fill_map is guaranteed to be all-zeroes at
490 * this point.
491 */
492 map_index = !map_index;
493 swap(res_map, fill_map);
494
495 rp += NFT_PIPAPO_GROUPS_PADDING(f);
496 }
497
498 out:
499 local_bh_enable();
500 return false;
501 }
502
503 /**
504 * pipapo_get() - Get matching element reference given key data
505 * @net: Network namespace
506 * @set: nftables API set representation
507 * @m: storage containing active/existing elements
508 * @data: Key data to be matched against existing elements
509 * @genmask: If set, check that element is active in given genmask
510 * @tstamp: timestamp to check for expired elements
511 * @gfp: the type of memory to allocate (see kmalloc).
512 *
513 * This is essentially the same as the lookup function, except that it matches
514 * key data against the uncommitted copy and doesn't use preallocated maps for
515 * bitmap results.
516 *
517 * Return: pointer to &struct nft_pipapo_elem on match, error pointer otherwise.
518 */
pipapo_get(const struct net * net,const struct nft_set * set,const struct nft_pipapo_match * m,const u8 * data,u8 genmask,u64 tstamp,gfp_t gfp)519 static struct nft_pipapo_elem *pipapo_get(const struct net *net,
520 const struct nft_set *set,
521 const struct nft_pipapo_match *m,
522 const u8 *data, u8 genmask,
523 u64 tstamp, gfp_t gfp)
524 {
525 struct nft_pipapo_elem *ret = ERR_PTR(-ENOENT);
526 unsigned long *res_map, *fill_map = NULL;
527 const struct nft_pipapo_field *f;
528 int i;
529
530 if (m->bsize_max == 0)
531 return ret;
532
533 res_map = kmalloc_array(m->bsize_max, sizeof(*res_map), gfp);
534 if (!res_map) {
535 ret = ERR_PTR(-ENOMEM);
536 goto out;
537 }
538
539 fill_map = kcalloc(m->bsize_max, sizeof(*res_map), gfp);
540 if (!fill_map) {
541 ret = ERR_PTR(-ENOMEM);
542 goto out;
543 }
544
545 pipapo_resmap_init(m, res_map);
546
547 nft_pipapo_for_each_field(f, i, m) {
548 bool last = i == m->field_count - 1;
549 int b;
550
551 /* For each bit group: select lookup table bucket depending on
552 * packet bytes value, then AND bucket value
553 */
554 if (f->bb == 8)
555 pipapo_and_field_buckets_8bit(f, res_map, data);
556 else if (f->bb == 4)
557 pipapo_and_field_buckets_4bit(f, res_map, data);
558 else
559 BUG();
560
561 data += f->groups / NFT_PIPAPO_GROUPS_PER_BYTE(f);
562
563 /* Now populate the bitmap for the next field, unless this is
564 * the last field, in which case return the matched 'ext'
565 * pointer if any.
566 *
567 * Now res_map contains the matching bitmap, and fill_map is the
568 * bitmap for the next field.
569 */
570 next_match:
571 b = pipapo_refill(res_map, f->bsize, f->rules, fill_map, f->mt,
572 last);
573 if (b < 0)
574 goto out;
575
576 if (last) {
577 if (__nft_set_elem_expired(&f->mt[b].e->ext, tstamp))
578 goto next_match;
579 if ((genmask &&
580 !nft_set_elem_active(&f->mt[b].e->ext, genmask)))
581 goto next_match;
582
583 ret = f->mt[b].e;
584 goto out;
585 }
586
587 data += NFT_PIPAPO_GROUPS_PADDING(f);
588
589 /* Swap bitmap indices: fill_map will be the initial bitmap for
590 * the next field (i.e. the new res_map), and res_map is
591 * guaranteed to be all-zeroes at this point, ready to be filled
592 * according to the next mapping table.
593 */
594 swap(res_map, fill_map);
595 }
596
597 out:
598 kfree(fill_map);
599 kfree(res_map);
600 return ret;
601 }
602
603 /**
604 * nft_pipapo_get() - Get matching element reference given key data
605 * @net: Network namespace
606 * @set: nftables API set representation
607 * @elem: nftables API element representation containing key data
608 * @flags: Unused
609 */
610 static struct nft_elem_priv *
nft_pipapo_get(const struct net * net,const struct nft_set * set,const struct nft_set_elem * elem,unsigned int flags)611 nft_pipapo_get(const struct net *net, const struct nft_set *set,
612 const struct nft_set_elem *elem, unsigned int flags)
613 {
614 struct nft_pipapo *priv = nft_set_priv(set);
615 struct nft_pipapo_match *m = rcu_dereference(priv->match);
616 struct nft_pipapo_elem *e;
617
618 e = pipapo_get(net, set, m, (const u8 *)elem->key.val.data,
619 nft_genmask_cur(net), get_jiffies_64(),
620 GFP_ATOMIC);
621 if (IS_ERR(e))
622 return ERR_CAST(e);
623
624 return &e->priv;
625 }
626
627 /**
628 * pipapo_realloc_mt() - Reallocate mapping table if needed upon resize
629 * @f: Field containing mapping table
630 * @old_rules: Amount of existing mapped rules
631 * @rules: Amount of new rules to map
632 *
633 * Return: 0 on success, negative error code on failure.
634 */
pipapo_realloc_mt(struct nft_pipapo_field * f,unsigned int old_rules,unsigned int rules)635 static int pipapo_realloc_mt(struct nft_pipapo_field *f,
636 unsigned int old_rules, unsigned int rules)
637 {
638 union nft_pipapo_map_bucket *new_mt = NULL, *old_mt = f->mt;
639 const unsigned int extra = PAGE_SIZE / sizeof(*new_mt);
640 unsigned int rules_alloc = rules;
641
642 might_sleep();
643
644 if (unlikely(rules == 0))
645 goto out_free;
646
647 /* growing and enough space left, no action needed */
648 if (rules > old_rules && f->rules_alloc > rules)
649 return 0;
650
651 /* downsize and extra slack has not grown too large */
652 if (rules < old_rules) {
653 unsigned int remove = f->rules_alloc - rules;
654
655 if (remove < (2u * extra))
656 return 0;
657 }
658
659 /* If set needs more than one page of memory for rules then
660 * allocate another extra page to avoid frequent reallocation.
661 */
662 if (rules > extra &&
663 check_add_overflow(rules, extra, &rules_alloc))
664 return -EOVERFLOW;
665
666 new_mt = kvmalloc_array(rules_alloc, sizeof(*new_mt), GFP_KERNEL_ACCOUNT);
667 if (!new_mt)
668 return -ENOMEM;
669
670 if (old_mt)
671 memcpy(new_mt, old_mt, min(old_rules, rules) * sizeof(*new_mt));
672
673 if (rules > old_rules) {
674 memset(new_mt + old_rules, 0,
675 (rules - old_rules) * sizeof(*new_mt));
676 }
677 out_free:
678 f->rules_alloc = rules_alloc;
679 f->mt = new_mt;
680
681 kvfree(old_mt);
682
683 return 0;
684 }
685
686 /**
687 * pipapo_resize() - Resize lookup or mapping table, or both
688 * @f: Field containing lookup and mapping tables
689 * @old_rules: Previous amount of rules in field
690 * @rules: New amount of rules
691 *
692 * Increase, decrease or maintain tables size depending on new amount of rules,
693 * and copy data over. In case the new size is smaller, throw away data for
694 * highest-numbered rules.
695 *
696 * Return: 0 on success, -ENOMEM on allocation failure.
697 */
pipapo_resize(struct nft_pipapo_field * f,unsigned int old_rules,unsigned int rules)698 static int pipapo_resize(struct nft_pipapo_field *f,
699 unsigned int old_rules, unsigned int rules)
700 {
701 long *new_lt = NULL, *new_p, *old_lt = f->lt, *old_p;
702 unsigned int new_bucket_size, copy;
703 int group, bucket, err;
704
705 if (rules >= NFT_PIPAPO_RULE0_MAX)
706 return -ENOSPC;
707
708 new_bucket_size = DIV_ROUND_UP(rules, BITS_PER_LONG);
709 #ifdef NFT_PIPAPO_ALIGN
710 new_bucket_size = roundup(new_bucket_size,
711 NFT_PIPAPO_ALIGN / sizeof(*new_lt));
712 #endif
713
714 if (new_bucket_size == f->bsize)
715 goto mt;
716
717 if (new_bucket_size > f->bsize)
718 copy = f->bsize;
719 else
720 copy = new_bucket_size;
721
722 new_lt = kvzalloc(f->groups * NFT_PIPAPO_BUCKETS(f->bb) *
723 new_bucket_size * sizeof(*new_lt) +
724 NFT_PIPAPO_ALIGN_HEADROOM,
725 GFP_KERNEL);
726 if (!new_lt)
727 return -ENOMEM;
728
729 new_p = NFT_PIPAPO_LT_ALIGN(new_lt);
730 old_p = NFT_PIPAPO_LT_ALIGN(old_lt);
731
732 for (group = 0; group < f->groups; group++) {
733 for (bucket = 0; bucket < NFT_PIPAPO_BUCKETS(f->bb); bucket++) {
734 memcpy(new_p, old_p, copy * sizeof(*new_p));
735 new_p += copy;
736 old_p += copy;
737
738 if (new_bucket_size > f->bsize)
739 new_p += new_bucket_size - f->bsize;
740 else
741 old_p += f->bsize - new_bucket_size;
742 }
743 }
744
745 mt:
746 err = pipapo_realloc_mt(f, old_rules, rules);
747 if (err) {
748 kvfree(new_lt);
749 return err;
750 }
751
752 if (new_lt) {
753 f->bsize = new_bucket_size;
754 f->lt = new_lt;
755 kvfree(old_lt);
756 }
757
758 return 0;
759 }
760
761 /**
762 * pipapo_bucket_set() - Set rule bit in bucket given group and group value
763 * @f: Field containing lookup table
764 * @rule: Rule index
765 * @group: Group index
766 * @v: Value of bit group
767 */
pipapo_bucket_set(struct nft_pipapo_field * f,int rule,int group,int v)768 static void pipapo_bucket_set(struct nft_pipapo_field *f, int rule, int group,
769 int v)
770 {
771 unsigned long *pos;
772
773 pos = NFT_PIPAPO_LT_ALIGN(f->lt);
774 pos += f->bsize * NFT_PIPAPO_BUCKETS(f->bb) * group;
775 pos += f->bsize * v;
776
777 __set_bit(rule, pos);
778 }
779
780 /**
781 * pipapo_lt_4b_to_8b() - Switch lookup table group width from 4 bits to 8 bits
782 * @old_groups: Number of current groups
783 * @bsize: Size of one bucket, in longs
784 * @old_lt: Pointer to the current lookup table
785 * @new_lt: Pointer to the new, pre-allocated lookup table
786 *
787 * Each bucket with index b in the new lookup table, belonging to group g, is
788 * filled with the bit intersection between:
789 * - bucket with index given by the upper 4 bits of b, from group g, and
790 * - bucket with index given by the lower 4 bits of b, from group g + 1
791 *
792 * That is, given buckets from the new lookup table N(x, y) and the old lookup
793 * table O(x, y), with x bucket index, and y group index:
794 *
795 * N(b, g) := O(b / 16, g) & O(b % 16, g + 1)
796 *
797 * This ensures equivalence of the matching results on lookup. Two examples in
798 * pictures:
799 *
800 * bucket
801 * group 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 ... 254 255
802 * 0 ^
803 * 1 | ^
804 * ... ( & ) |
805 * / \ |
806 * / \ .-( & )-.
807 * / bucket \ | |
808 * group 0 / 1 2 3 \ 4 5 6 7 8 9 10 11 12 13 |14 15 |
809 * 0 / \ | |
810 * 1 \ | |
811 * 2 | --'
812 * 3 '-
813 * ...
814 */
pipapo_lt_4b_to_8b(int old_groups,int bsize,unsigned long * old_lt,unsigned long * new_lt)815 static void pipapo_lt_4b_to_8b(int old_groups, int bsize,
816 unsigned long *old_lt, unsigned long *new_lt)
817 {
818 int g, b, i;
819
820 for (g = 0; g < old_groups / 2; g++) {
821 int src_g0 = g * 2, src_g1 = g * 2 + 1;
822
823 for (b = 0; b < NFT_PIPAPO_BUCKETS(8); b++) {
824 int src_b0 = b / NFT_PIPAPO_BUCKETS(4);
825 int src_b1 = b % NFT_PIPAPO_BUCKETS(4);
826 int src_i0 = src_g0 * NFT_PIPAPO_BUCKETS(4) + src_b0;
827 int src_i1 = src_g1 * NFT_PIPAPO_BUCKETS(4) + src_b1;
828
829 for (i = 0; i < bsize; i++) {
830 *new_lt = old_lt[src_i0 * bsize + i] &
831 old_lt[src_i1 * bsize + i];
832 new_lt++;
833 }
834 }
835 }
836 }
837
838 /**
839 * pipapo_lt_8b_to_4b() - Switch lookup table group width from 8 bits to 4 bits
840 * @old_groups: Number of current groups
841 * @bsize: Size of one bucket, in longs
842 * @old_lt: Pointer to the current lookup table
843 * @new_lt: Pointer to the new, pre-allocated lookup table
844 *
845 * Each bucket with index b in the new lookup table, belonging to group g, is
846 * filled with the bit union of:
847 * - all the buckets with index such that the upper four bits of the lower byte
848 * equal b, from group g, with g odd
849 * - all the buckets with index such that the lower four bits equal b, from
850 * group g, with g even
851 *
852 * That is, given buckets from the new lookup table N(x, y) and the old lookup
853 * table O(x, y), with x bucket index, and y group index:
854 *
855 * - with g odd: N(b, g) := U(O(x, g) for each x : x = (b & 0xf0) >> 4)
856 * - with g even: N(b, g) := U(O(x, g) for each x : x = b & 0x0f)
857 *
858 * where U() denotes the arbitrary union operation (binary OR of n terms). This
859 * ensures equivalence of the matching results on lookup.
860 */
pipapo_lt_8b_to_4b(int old_groups,int bsize,unsigned long * old_lt,unsigned long * new_lt)861 static void pipapo_lt_8b_to_4b(int old_groups, int bsize,
862 unsigned long *old_lt, unsigned long *new_lt)
863 {
864 int g, b, bsrc, i;
865
866 memset(new_lt, 0, old_groups * 2 * NFT_PIPAPO_BUCKETS(4) * bsize *
867 sizeof(unsigned long));
868
869 for (g = 0; g < old_groups * 2; g += 2) {
870 int src_g = g / 2;
871
872 for (b = 0; b < NFT_PIPAPO_BUCKETS(4); b++) {
873 for (bsrc = NFT_PIPAPO_BUCKETS(8) * src_g;
874 bsrc < NFT_PIPAPO_BUCKETS(8) * (src_g + 1);
875 bsrc++) {
876 if (((bsrc & 0xf0) >> 4) != b)
877 continue;
878
879 for (i = 0; i < bsize; i++)
880 new_lt[i] |= old_lt[bsrc * bsize + i];
881 }
882
883 new_lt += bsize;
884 }
885
886 for (b = 0; b < NFT_PIPAPO_BUCKETS(4); b++) {
887 for (bsrc = NFT_PIPAPO_BUCKETS(8) * src_g;
888 bsrc < NFT_PIPAPO_BUCKETS(8) * (src_g + 1);
889 bsrc++) {
890 if ((bsrc & 0x0f) != b)
891 continue;
892
893 for (i = 0; i < bsize; i++)
894 new_lt[i] |= old_lt[bsrc * bsize + i];
895 }
896
897 new_lt += bsize;
898 }
899 }
900 }
901
902 /**
903 * pipapo_lt_bits_adjust() - Adjust group size for lookup table if needed
904 * @f: Field containing lookup table
905 */
pipapo_lt_bits_adjust(struct nft_pipapo_field * f)906 static void pipapo_lt_bits_adjust(struct nft_pipapo_field *f)
907 {
908 unsigned int groups, bb;
909 unsigned long *new_lt;
910 size_t lt_size;
911
912 lt_size = f->groups * NFT_PIPAPO_BUCKETS(f->bb) * f->bsize *
913 sizeof(*f->lt);
914
915 if (f->bb == NFT_PIPAPO_GROUP_BITS_SMALL_SET &&
916 lt_size > NFT_PIPAPO_LT_SIZE_HIGH) {
917 groups = f->groups * 2;
918 bb = NFT_PIPAPO_GROUP_BITS_LARGE_SET;
919
920 lt_size = groups * NFT_PIPAPO_BUCKETS(bb) * f->bsize *
921 sizeof(*f->lt);
922 } else if (f->bb == NFT_PIPAPO_GROUP_BITS_LARGE_SET &&
923 lt_size < NFT_PIPAPO_LT_SIZE_LOW) {
924 groups = f->groups / 2;
925 bb = NFT_PIPAPO_GROUP_BITS_SMALL_SET;
926
927 lt_size = groups * NFT_PIPAPO_BUCKETS(bb) * f->bsize *
928 sizeof(*f->lt);
929
930 /* Don't increase group width if the resulting lookup table size
931 * would exceed the upper size threshold for a "small" set.
932 */
933 if (lt_size > NFT_PIPAPO_LT_SIZE_HIGH)
934 return;
935 } else {
936 return;
937 }
938
939 new_lt = kvzalloc(lt_size + NFT_PIPAPO_ALIGN_HEADROOM, GFP_KERNEL_ACCOUNT);
940 if (!new_lt)
941 return;
942
943 NFT_PIPAPO_GROUP_BITS_ARE_8_OR_4;
944 if (f->bb == 4 && bb == 8) {
945 pipapo_lt_4b_to_8b(f->groups, f->bsize,
946 NFT_PIPAPO_LT_ALIGN(f->lt),
947 NFT_PIPAPO_LT_ALIGN(new_lt));
948 } else if (f->bb == 8 && bb == 4) {
949 pipapo_lt_8b_to_4b(f->groups, f->bsize,
950 NFT_PIPAPO_LT_ALIGN(f->lt),
951 NFT_PIPAPO_LT_ALIGN(new_lt));
952 } else {
953 BUG();
954 }
955
956 f->groups = groups;
957 f->bb = bb;
958 kvfree(f->lt);
959 f->lt = new_lt;
960 }
961
962 /**
963 * pipapo_insert() - Insert new rule in field given input key and mask length
964 * @f: Field containing lookup table
965 * @k: Input key for classification, without nftables padding
966 * @mask_bits: Length of mask; matches field length for non-ranged entry
967 *
968 * Insert a new rule reference in lookup buckets corresponding to k and
969 * mask_bits.
970 *
971 * Return: 1 on success (one rule inserted), negative error code on failure.
972 */
pipapo_insert(struct nft_pipapo_field * f,const uint8_t * k,int mask_bits)973 static int pipapo_insert(struct nft_pipapo_field *f, const uint8_t *k,
974 int mask_bits)
975 {
976 unsigned int rule = f->rules, group, ret, bit_offset = 0;
977
978 ret = pipapo_resize(f, f->rules, f->rules + 1);
979 if (ret)
980 return ret;
981
982 f->rules++;
983
984 for (group = 0; group < f->groups; group++) {
985 int i, v;
986 u8 mask;
987
988 v = k[group / (BITS_PER_BYTE / f->bb)];
989 v &= GENMASK(BITS_PER_BYTE - bit_offset - 1, 0);
990 v >>= (BITS_PER_BYTE - bit_offset) - f->bb;
991
992 bit_offset += f->bb;
993 bit_offset %= BITS_PER_BYTE;
994
995 if (mask_bits >= (group + 1) * f->bb) {
996 /* Not masked */
997 pipapo_bucket_set(f, rule, group, v);
998 } else if (mask_bits <= group * f->bb) {
999 /* Completely masked */
1000 for (i = 0; i < NFT_PIPAPO_BUCKETS(f->bb); i++)
1001 pipapo_bucket_set(f, rule, group, i);
1002 } else {
1003 /* The mask limit falls on this group */
1004 mask = GENMASK(f->bb - 1, 0);
1005 mask >>= mask_bits - group * f->bb;
1006 for (i = 0; i < NFT_PIPAPO_BUCKETS(f->bb); i++) {
1007 if ((i & ~mask) == (v & ~mask))
1008 pipapo_bucket_set(f, rule, group, i);
1009 }
1010 }
1011 }
1012
1013 pipapo_lt_bits_adjust(f);
1014
1015 return 1;
1016 }
1017
1018 /**
1019 * pipapo_step_diff() - Check if setting @step bit in netmask would change it
1020 * @base: Mask we are expanding
1021 * @step: Step bit for given expansion step
1022 * @len: Total length of mask space (set and unset bits), bytes
1023 *
1024 * Convenience function for mask expansion.
1025 *
1026 * Return: true if step bit changes mask (i.e. isn't set), false otherwise.
1027 */
pipapo_step_diff(u8 * base,int step,int len)1028 static bool pipapo_step_diff(u8 *base, int step, int len)
1029 {
1030 /* Network order, byte-addressed */
1031 #ifdef __BIG_ENDIAN__
1032 return !(BIT(step % BITS_PER_BYTE) & base[step / BITS_PER_BYTE]);
1033 #else
1034 return !(BIT(step % BITS_PER_BYTE) &
1035 base[len - 1 - step / BITS_PER_BYTE]);
1036 #endif
1037 }
1038
1039 /**
1040 * pipapo_step_after_end() - Check if mask exceeds range end with given step
1041 * @base: Mask we are expanding
1042 * @end: End of range
1043 * @step: Step bit for given expansion step, highest bit to be set
1044 * @len: Total length of mask space (set and unset bits), bytes
1045 *
1046 * Convenience function for mask expansion.
1047 *
1048 * Return: true if mask exceeds range setting step bits, false otherwise.
1049 */
pipapo_step_after_end(const u8 * base,const u8 * end,int step,int len)1050 static bool pipapo_step_after_end(const u8 *base, const u8 *end, int step,
1051 int len)
1052 {
1053 u8 tmp[NFT_PIPAPO_MAX_BYTES];
1054 int i;
1055
1056 memcpy(tmp, base, len);
1057
1058 /* Network order, byte-addressed */
1059 for (i = 0; i <= step; i++)
1060 #ifdef __BIG_ENDIAN__
1061 tmp[i / BITS_PER_BYTE] |= BIT(i % BITS_PER_BYTE);
1062 #else
1063 tmp[len - 1 - i / BITS_PER_BYTE] |= BIT(i % BITS_PER_BYTE);
1064 #endif
1065
1066 return memcmp(tmp, end, len) > 0;
1067 }
1068
1069 /**
1070 * pipapo_base_sum() - Sum step bit to given len-sized netmask base with carry
1071 * @base: Netmask base
1072 * @step: Step bit to sum
1073 * @len: Netmask length, bytes
1074 */
pipapo_base_sum(u8 * base,int step,int len)1075 static void pipapo_base_sum(u8 *base, int step, int len)
1076 {
1077 bool carry = false;
1078 int i;
1079
1080 /* Network order, byte-addressed */
1081 #ifdef __BIG_ENDIAN__
1082 for (i = step / BITS_PER_BYTE; i < len; i++) {
1083 #else
1084 for (i = len - 1 - step / BITS_PER_BYTE; i >= 0; i--) {
1085 #endif
1086 if (carry)
1087 base[i]++;
1088 else
1089 base[i] += 1 << (step % BITS_PER_BYTE);
1090
1091 if (base[i])
1092 break;
1093
1094 carry = true;
1095 }
1096 }
1097
1098 /**
1099 * pipapo_expand() - Expand to composing netmasks, insert into lookup table
1100 * @f: Field containing lookup table
1101 * @start: Start of range
1102 * @end: End of range
1103 * @len: Length of value in bits
1104 *
1105 * Expand range to composing netmasks and insert corresponding rule references
1106 * in lookup buckets.
1107 *
1108 * Return: number of inserted rules on success, negative error code on failure.
1109 */
1110 static int pipapo_expand(struct nft_pipapo_field *f,
1111 const u8 *start, const u8 *end, int len)
1112 {
1113 int step, masks = 0, bytes = DIV_ROUND_UP(len, BITS_PER_BYTE);
1114 u8 base[NFT_PIPAPO_MAX_BYTES];
1115
1116 memcpy(base, start, bytes);
1117 while (memcmp(base, end, bytes) <= 0) {
1118 int err;
1119
1120 step = 0;
1121 while (pipapo_step_diff(base, step, bytes)) {
1122 if (pipapo_step_after_end(base, end, step, bytes))
1123 break;
1124
1125 step++;
1126 if (step >= len) {
1127 if (!masks) {
1128 err = pipapo_insert(f, base, 0);
1129 if (err < 0)
1130 return err;
1131 masks = 1;
1132 }
1133 goto out;
1134 }
1135 }
1136
1137 err = pipapo_insert(f, base, len - step);
1138
1139 if (err < 0)
1140 return err;
1141
1142 masks++;
1143 pipapo_base_sum(base, step, bytes);
1144 }
1145 out:
1146 return masks;
1147 }
1148
1149 /**
1150 * pipapo_map() - Insert rules in mapping tables, mapping them between fields
1151 * @m: Matching data, including mapping table
1152 * @map: Table of rule maps: array of first rule and amount of rules
1153 * in next field a given rule maps to, for each field
1154 * @e: For last field, nft_set_ext pointer matching rules map to
1155 */
1156 static void pipapo_map(struct nft_pipapo_match *m,
1157 union nft_pipapo_map_bucket map[NFT_PIPAPO_MAX_FIELDS],
1158 struct nft_pipapo_elem *e)
1159 {
1160 struct nft_pipapo_field *f;
1161 int i, j;
1162
1163 for (i = 0, f = m->f; i < m->field_count - 1; i++, f++) {
1164 for (j = 0; j < map[i].n; j++) {
1165 f->mt[map[i].to + j].to = map[i + 1].to;
1166 f->mt[map[i].to + j].n = map[i + 1].n;
1167 }
1168 }
1169
1170 /* Last field: map to ext instead of mapping to next field */
1171 for (j = 0; j < map[i].n; j++)
1172 f->mt[map[i].to + j].e = e;
1173 }
1174
1175 /**
1176 * pipapo_free_scratch() - Free per-CPU map at original (not aligned) address
1177 * @m: Matching data
1178 * @cpu: CPU number
1179 */
1180 static void pipapo_free_scratch(const struct nft_pipapo_match *m, unsigned int cpu)
1181 {
1182 struct nft_pipapo_scratch *s;
1183 void *mem;
1184
1185 s = *per_cpu_ptr(m->scratch, cpu);
1186 if (!s)
1187 return;
1188
1189 mem = s;
1190 mem -= s->align_off;
1191 kfree(mem);
1192 }
1193
1194 /**
1195 * pipapo_realloc_scratch() - Reallocate scratch maps for partial match results
1196 * @clone: Copy of matching data with pending insertions and deletions
1197 * @bsize_max: Maximum bucket size, scratch maps cover two buckets
1198 *
1199 * Return: 0 on success, -ENOMEM on failure.
1200 */
1201 static int pipapo_realloc_scratch(struct nft_pipapo_match *clone,
1202 unsigned long bsize_max)
1203 {
1204 int i;
1205
1206 for_each_possible_cpu(i) {
1207 struct nft_pipapo_scratch *scratch;
1208 #ifdef NFT_PIPAPO_ALIGN
1209 void *scratch_aligned;
1210 u32 align_off;
1211 #endif
1212 scratch = kzalloc_node(struct_size(scratch, map,
1213 bsize_max * 2) +
1214 NFT_PIPAPO_ALIGN_HEADROOM,
1215 GFP_KERNEL_ACCOUNT, cpu_to_node(i));
1216 if (!scratch) {
1217 /* On failure, there's no need to undo previous
1218 * allocations: this means that some scratch maps have
1219 * a bigger allocated size now (this is only called on
1220 * insertion), but the extra space won't be used by any
1221 * CPU as new elements are not inserted and m->bsize_max
1222 * is not updated.
1223 */
1224 return -ENOMEM;
1225 }
1226
1227 pipapo_free_scratch(clone, i);
1228
1229 #ifdef NFT_PIPAPO_ALIGN
1230 /* Align &scratch->map (not the struct itself): the extra
1231 * %NFT_PIPAPO_ALIGN_HEADROOM bytes passed to kzalloc_node()
1232 * above guarantee we can waste up to those bytes in order
1233 * to align the map field regardless of its offset within
1234 * the struct.
1235 */
1236 BUILD_BUG_ON(offsetof(struct nft_pipapo_scratch, map) > NFT_PIPAPO_ALIGN_HEADROOM);
1237
1238 scratch_aligned = NFT_PIPAPO_LT_ALIGN(&scratch->map);
1239 scratch_aligned -= offsetof(struct nft_pipapo_scratch, map);
1240 align_off = scratch_aligned - (void *)scratch;
1241
1242 scratch = scratch_aligned;
1243 scratch->align_off = align_off;
1244 #endif
1245 *per_cpu_ptr(clone->scratch, i) = scratch;
1246 }
1247
1248 return 0;
1249 }
1250
1251 static bool nft_pipapo_transaction_mutex_held(const struct nft_set *set)
1252 {
1253 #ifdef CONFIG_PROVE_LOCKING
1254 const struct net *net = read_pnet(&set->net);
1255
1256 return lockdep_is_held(&nft_pernet(net)->commit_mutex);
1257 #else
1258 return true;
1259 #endif
1260 }
1261
1262 static struct nft_pipapo_match *pipapo_clone(struct nft_pipapo_match *old);
1263
1264 /**
1265 * pipapo_maybe_clone() - Build clone for pending data changes, if not existing
1266 * @set: nftables API set representation
1267 *
1268 * Return: newly created or existing clone, if any. NULL on allocation failure
1269 */
1270 static struct nft_pipapo_match *pipapo_maybe_clone(const struct nft_set *set)
1271 {
1272 struct nft_pipapo *priv = nft_set_priv(set);
1273 struct nft_pipapo_match *m;
1274
1275 if (priv->clone)
1276 return priv->clone;
1277
1278 m = rcu_dereference_protected(priv->match,
1279 nft_pipapo_transaction_mutex_held(set));
1280 priv->clone = pipapo_clone(m);
1281
1282 return priv->clone;
1283 }
1284
1285 /**
1286 * nft_pipapo_insert() - Validate and insert ranged elements
1287 * @net: Network namespace
1288 * @set: nftables API set representation
1289 * @elem: nftables API element representation containing key data
1290 * @elem_priv: Filled with pointer to &struct nft_set_ext in inserted element
1291 *
1292 * Return: 0 on success, error pointer on failure.
1293 */
1294 static int nft_pipapo_insert(const struct net *net, const struct nft_set *set,
1295 const struct nft_set_elem *elem,
1296 struct nft_elem_priv **elem_priv)
1297 {
1298 const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
1299 union nft_pipapo_map_bucket rulemap[NFT_PIPAPO_MAX_FIELDS];
1300 const u8 *start = (const u8 *)elem->key.val.data, *end;
1301 struct nft_pipapo_match *m = pipapo_maybe_clone(set);
1302 u8 genmask = nft_genmask_next(net);
1303 struct nft_pipapo_elem *e, *dup;
1304 u64 tstamp = nft_net_tstamp(net);
1305 struct nft_pipapo_field *f;
1306 const u8 *start_p, *end_p;
1307 int i, bsize_max, err = 0;
1308
1309 if (!m)
1310 return -ENOMEM;
1311
1312 if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY_END))
1313 end = (const u8 *)nft_set_ext_key_end(ext)->data;
1314 else
1315 end = start;
1316
1317 dup = pipapo_get(net, set, m, start, genmask, tstamp, GFP_KERNEL);
1318 if (!IS_ERR(dup)) {
1319 /* Check if we already have the same exact entry */
1320 const struct nft_data *dup_key, *dup_end;
1321
1322 dup_key = nft_set_ext_key(&dup->ext);
1323 if (nft_set_ext_exists(&dup->ext, NFT_SET_EXT_KEY_END))
1324 dup_end = nft_set_ext_key_end(&dup->ext);
1325 else
1326 dup_end = dup_key;
1327
1328 if (!memcmp(start, dup_key->data, sizeof(*dup_key->data)) &&
1329 !memcmp(end, dup_end->data, sizeof(*dup_end->data))) {
1330 *elem_priv = &dup->priv;
1331 return -EEXIST;
1332 }
1333
1334 return -ENOTEMPTY;
1335 }
1336
1337 if (PTR_ERR(dup) == -ENOENT) {
1338 /* Look for partially overlapping entries */
1339 dup = pipapo_get(net, set, m, end, nft_genmask_next(net), tstamp,
1340 GFP_KERNEL);
1341 }
1342
1343 if (PTR_ERR(dup) != -ENOENT) {
1344 if (IS_ERR(dup))
1345 return PTR_ERR(dup);
1346 *elem_priv = &dup->priv;
1347 return -ENOTEMPTY;
1348 }
1349
1350 /* Validate */
1351 start_p = start;
1352 end_p = end;
1353
1354 /* some helpers return -1, or 0 >= for valid rule pos,
1355 * so we cannot support more than INT_MAX rules at this time.
1356 */
1357 BUILD_BUG_ON(NFT_PIPAPO_RULE0_MAX > INT_MAX);
1358
1359 nft_pipapo_for_each_field(f, i, m) {
1360 if (f->rules >= NFT_PIPAPO_RULE0_MAX)
1361 return -ENOSPC;
1362
1363 if (memcmp(start_p, end_p,
1364 f->groups / NFT_PIPAPO_GROUPS_PER_BYTE(f)) > 0)
1365 return -EINVAL;
1366
1367 start_p += NFT_PIPAPO_GROUPS_PADDED_SIZE(f);
1368 end_p += NFT_PIPAPO_GROUPS_PADDED_SIZE(f);
1369 }
1370
1371 /* Insert */
1372 bsize_max = m->bsize_max;
1373
1374 nft_pipapo_for_each_field(f, i, m) {
1375 int ret;
1376
1377 rulemap[i].to = f->rules;
1378
1379 ret = memcmp(start, end,
1380 f->groups / NFT_PIPAPO_GROUPS_PER_BYTE(f));
1381 if (!ret)
1382 ret = pipapo_insert(f, start, f->groups * f->bb);
1383 else
1384 ret = pipapo_expand(f, start, end, f->groups * f->bb);
1385
1386 if (ret < 0)
1387 return ret;
1388
1389 if (f->bsize > bsize_max)
1390 bsize_max = f->bsize;
1391
1392 rulemap[i].n = ret;
1393
1394 start += NFT_PIPAPO_GROUPS_PADDED_SIZE(f);
1395 end += NFT_PIPAPO_GROUPS_PADDED_SIZE(f);
1396 }
1397
1398 if (!*get_cpu_ptr(m->scratch) || bsize_max > m->bsize_max) {
1399 put_cpu_ptr(m->scratch);
1400
1401 err = pipapo_realloc_scratch(m, bsize_max);
1402 if (err)
1403 return err;
1404
1405 m->bsize_max = bsize_max;
1406 } else {
1407 put_cpu_ptr(m->scratch);
1408 }
1409
1410 e = nft_elem_priv_cast(elem->priv);
1411 *elem_priv = &e->priv;
1412
1413 pipapo_map(m, rulemap, e);
1414
1415 return 0;
1416 }
1417
1418 /**
1419 * pipapo_clone() - Clone matching data to create new working copy
1420 * @old: Existing matching data
1421 *
1422 * Return: copy of matching data passed as 'old' or NULL.
1423 */
1424 static struct nft_pipapo_match *pipapo_clone(struct nft_pipapo_match *old)
1425 {
1426 struct nft_pipapo_field *dst, *src;
1427 struct nft_pipapo_match *new;
1428 int i;
1429
1430 new = kmalloc(struct_size(new, f, old->field_count), GFP_KERNEL_ACCOUNT);
1431 if (!new)
1432 return NULL;
1433
1434 new->field_count = old->field_count;
1435 new->bsize_max = old->bsize_max;
1436
1437 new->scratch = alloc_percpu(*new->scratch);
1438 if (!new->scratch)
1439 goto out_scratch;
1440
1441 for_each_possible_cpu(i)
1442 *per_cpu_ptr(new->scratch, i) = NULL;
1443
1444 if (pipapo_realloc_scratch(new, old->bsize_max))
1445 goto out_scratch_realloc;
1446
1447 rcu_head_init(&new->rcu);
1448
1449 src = old->f;
1450 dst = new->f;
1451
1452 for (i = 0; i < old->field_count; i++) {
1453 unsigned long *new_lt;
1454
1455 memcpy(dst, src, offsetof(struct nft_pipapo_field, lt));
1456
1457 new_lt = kvzalloc(src->groups * NFT_PIPAPO_BUCKETS(src->bb) *
1458 src->bsize * sizeof(*dst->lt) +
1459 NFT_PIPAPO_ALIGN_HEADROOM,
1460 GFP_KERNEL_ACCOUNT);
1461 if (!new_lt)
1462 goto out_lt;
1463
1464 dst->lt = new_lt;
1465
1466 memcpy(NFT_PIPAPO_LT_ALIGN(new_lt),
1467 NFT_PIPAPO_LT_ALIGN(src->lt),
1468 src->bsize * sizeof(*dst->lt) *
1469 src->groups * NFT_PIPAPO_BUCKETS(src->bb));
1470
1471 if (src->rules > 0) {
1472 dst->mt = kvmalloc_array(src->rules_alloc,
1473 sizeof(*src->mt),
1474 GFP_KERNEL_ACCOUNT);
1475 if (!dst->mt)
1476 goto out_mt;
1477
1478 memcpy(dst->mt, src->mt, src->rules * sizeof(*src->mt));
1479 } else {
1480 dst->mt = NULL;
1481 dst->rules_alloc = 0;
1482 }
1483
1484 src++;
1485 dst++;
1486 }
1487
1488 return new;
1489
1490 out_mt:
1491 kvfree(dst->lt);
1492 out_lt:
1493 for (dst--; i > 0; i--) {
1494 kvfree(dst->mt);
1495 kvfree(dst->lt);
1496 dst--;
1497 }
1498 out_scratch_realloc:
1499 for_each_possible_cpu(i)
1500 pipapo_free_scratch(new, i);
1501 out_scratch:
1502 free_percpu(new->scratch);
1503 kfree(new);
1504
1505 return NULL;
1506 }
1507
1508 /**
1509 * pipapo_rules_same_key() - Get number of rules originated from the same entry
1510 * @f: Field containing mapping table
1511 * @first: Index of first rule in set of rules mapping to same entry
1512 *
1513 * Using the fact that all rules in a field that originated from the same entry
1514 * will map to the same set of rules in the next field, or to the same element
1515 * reference, return the cardinality of the set of rules that originated from
1516 * the same entry as the rule with index @first, @first rule included.
1517 *
1518 * In pictures:
1519 * rules
1520 * field #0 0 1 2 3 4
1521 * map to: 0 1 2-4 2-4 5-9
1522 * . . ....... . ...
1523 * | | | | \ \
1524 * | | | | \ \
1525 * | | | | \ \
1526 * ' ' ' ' ' \
1527 * in field #1 0 1 2 3 4 5 ...
1528 *
1529 * if this is called for rule 2 on field #0, it will return 3, as also rules 2
1530 * and 3 in field 0 map to the same set of rules (2, 3, 4) in the next field.
1531 *
1532 * For the last field in a set, we can rely on associated entries to map to the
1533 * same element references.
1534 *
1535 * Return: Number of rules that originated from the same entry as @first.
1536 */
1537 static unsigned int pipapo_rules_same_key(struct nft_pipapo_field *f, unsigned int first)
1538 {
1539 struct nft_pipapo_elem *e = NULL; /* Keep gcc happy */
1540 unsigned int r;
1541
1542 for (r = first; r < f->rules; r++) {
1543 if (r != first && e != f->mt[r].e)
1544 return r - first;
1545
1546 e = f->mt[r].e;
1547 }
1548
1549 if (r != first)
1550 return r - first;
1551
1552 return 0;
1553 }
1554
1555 /**
1556 * pipapo_unmap() - Remove rules from mapping tables, renumber remaining ones
1557 * @mt: Mapping array
1558 * @rules: Original amount of rules in mapping table
1559 * @start: First rule index to be removed
1560 * @n: Amount of rules to be removed
1561 * @to_offset: First rule index, in next field, this group of rules maps to
1562 * @is_last: If this is the last field, delete reference from mapping array
1563 *
1564 * This is used to unmap rules from the mapping table for a single field,
1565 * maintaining consistency and compactness for the existing ones.
1566 *
1567 * In pictures: let's assume that we want to delete rules 2 and 3 from the
1568 * following mapping array:
1569 *
1570 * rules
1571 * 0 1 2 3 4
1572 * map to: 4-10 4-10 11-15 11-15 16-18
1573 *
1574 * the result will be:
1575 *
1576 * rules
1577 * 0 1 2
1578 * map to: 4-10 4-10 11-13
1579 *
1580 * for fields before the last one. In case this is the mapping table for the
1581 * last field in a set, and rules map to pointers to &struct nft_pipapo_elem:
1582 *
1583 * rules
1584 * 0 1 2 3 4
1585 * element pointers: 0x42 0x42 0x33 0x33 0x44
1586 *
1587 * the result will be:
1588 *
1589 * rules
1590 * 0 1 2
1591 * element pointers: 0x42 0x42 0x44
1592 */
1593 static void pipapo_unmap(union nft_pipapo_map_bucket *mt, unsigned int rules,
1594 unsigned int start, unsigned int n,
1595 unsigned int to_offset, bool is_last)
1596 {
1597 int i;
1598
1599 memmove(mt + start, mt + start + n, (rules - start - n) * sizeof(*mt));
1600 memset(mt + rules - n, 0, n * sizeof(*mt));
1601
1602 if (is_last)
1603 return;
1604
1605 for (i = start; i < rules - n; i++)
1606 mt[i].to -= to_offset;
1607 }
1608
1609 /**
1610 * pipapo_drop() - Delete entry from lookup and mapping tables, given rule map
1611 * @m: Matching data
1612 * @rulemap: Table of rule maps, arrays of first rule and amount of rules
1613 * in next field a given entry maps to, for each field
1614 *
1615 * For each rule in lookup table buckets mapping to this set of rules, drop
1616 * all bits set in lookup table mapping. In pictures, assuming we want to drop
1617 * rules 0 and 1 from this lookup table:
1618 *
1619 * bucket
1620 * group 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
1621 * 0 0 1,2
1622 * 1 1,2 0
1623 * 2 0 1,2
1624 * 3 0 1,2
1625 * 4 0,1,2
1626 * 5 0 1 2
1627 * 6 0,1,2 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1628 * 7 1,2 1,2 1 1 1 0,1 1 1 1 1 1 1 1 1 1 1
1629 *
1630 * rule 2 becomes rule 0, and the result will be:
1631 *
1632 * bucket
1633 * group 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
1634 * 0 0
1635 * 1 0
1636 * 2 0
1637 * 3 0
1638 * 4 0
1639 * 5 0
1640 * 6 0
1641 * 7 0 0
1642 *
1643 * once this is done, call unmap() to drop all the corresponding rule references
1644 * from mapping tables.
1645 */
1646 static void pipapo_drop(struct nft_pipapo_match *m,
1647 union nft_pipapo_map_bucket rulemap[])
1648 {
1649 struct nft_pipapo_field *f;
1650 int i;
1651
1652 nft_pipapo_for_each_field(f, i, m) {
1653 int g;
1654
1655 for (g = 0; g < f->groups; g++) {
1656 unsigned long *pos;
1657 int b;
1658
1659 pos = NFT_PIPAPO_LT_ALIGN(f->lt) + g *
1660 NFT_PIPAPO_BUCKETS(f->bb) * f->bsize;
1661
1662 for (b = 0; b < NFT_PIPAPO_BUCKETS(f->bb); b++) {
1663 bitmap_cut(pos, pos, rulemap[i].to,
1664 rulemap[i].n,
1665 f->bsize * BITS_PER_LONG);
1666
1667 pos += f->bsize;
1668 }
1669 }
1670
1671 pipapo_unmap(f->mt, f->rules, rulemap[i].to, rulemap[i].n,
1672 rulemap[i + 1].n, i == m->field_count - 1);
1673 if (pipapo_resize(f, f->rules, f->rules - rulemap[i].n)) {
1674 /* We can ignore this, a failure to shrink tables down
1675 * doesn't make tables invalid.
1676 */
1677 ;
1678 }
1679 f->rules -= rulemap[i].n;
1680
1681 pipapo_lt_bits_adjust(f);
1682 }
1683 }
1684
1685 static void nft_pipapo_gc_deactivate(struct net *net, struct nft_set *set,
1686 struct nft_pipapo_elem *e)
1687
1688 {
1689 nft_setelem_data_deactivate(net, set, &e->priv);
1690 }
1691
1692 /**
1693 * pipapo_gc() - Drop expired entries from set, destroy start and end elements
1694 * @set: nftables API set representation
1695 * @m: Matching data
1696 */
1697 static void pipapo_gc(struct nft_set *set, struct nft_pipapo_match *m)
1698 {
1699 struct nft_pipapo *priv = nft_set_priv(set);
1700 struct net *net = read_pnet(&set->net);
1701 unsigned int rules_f0, first_rule = 0;
1702 u64 tstamp = nft_net_tstamp(net);
1703 struct nft_pipapo_elem *e;
1704 struct nft_trans_gc *gc;
1705
1706 gc = nft_trans_gc_alloc(set, 0, GFP_KERNEL);
1707 if (!gc)
1708 return;
1709
1710 while ((rules_f0 = pipapo_rules_same_key(m->f, first_rule))) {
1711 union nft_pipapo_map_bucket rulemap[NFT_PIPAPO_MAX_FIELDS];
1712 const struct nft_pipapo_field *f;
1713 unsigned int i, start, rules_fx;
1714
1715 start = first_rule;
1716 rules_fx = rules_f0;
1717
1718 nft_pipapo_for_each_field(f, i, m) {
1719 rulemap[i].to = start;
1720 rulemap[i].n = rules_fx;
1721
1722 if (i < m->field_count - 1) {
1723 rules_fx = f->mt[start].n;
1724 start = f->mt[start].to;
1725 }
1726 }
1727
1728 /* Pick the last field, and its last index */
1729 f--;
1730 i--;
1731 e = f->mt[rulemap[i].to].e;
1732
1733 /* synchronous gc never fails, there is no need to set on
1734 * NFT_SET_ELEM_DEAD_BIT.
1735 */
1736 if (__nft_set_elem_expired(&e->ext, tstamp)) {
1737 gc = nft_trans_gc_queue_sync(gc, GFP_KERNEL);
1738 if (!gc)
1739 return;
1740
1741 nft_pipapo_gc_deactivate(net, set, e);
1742 pipapo_drop(m, rulemap);
1743 nft_trans_gc_elem_add(gc, e);
1744
1745 /* And check again current first rule, which is now the
1746 * first we haven't checked.
1747 */
1748 } else {
1749 first_rule += rules_f0;
1750 }
1751 }
1752
1753 gc = nft_trans_gc_catchall_sync(gc);
1754 if (gc) {
1755 nft_trans_gc_queue_sync_done(gc);
1756 priv->last_gc = jiffies;
1757 }
1758 }
1759
1760 /**
1761 * pipapo_free_fields() - Free per-field tables contained in matching data
1762 * @m: Matching data
1763 */
1764 static void pipapo_free_fields(struct nft_pipapo_match *m)
1765 {
1766 struct nft_pipapo_field *f;
1767 int i;
1768
1769 nft_pipapo_for_each_field(f, i, m) {
1770 kvfree(f->lt);
1771 kvfree(f->mt);
1772 }
1773 }
1774
1775 static void pipapo_free_match(struct nft_pipapo_match *m)
1776 {
1777 int i;
1778
1779 for_each_possible_cpu(i)
1780 pipapo_free_scratch(m, i);
1781
1782 free_percpu(m->scratch);
1783 pipapo_free_fields(m);
1784
1785 kfree(m);
1786 }
1787
1788 /**
1789 * pipapo_reclaim_match - RCU callback to free fields from old matching data
1790 * @rcu: RCU head
1791 */
1792 static void pipapo_reclaim_match(struct rcu_head *rcu)
1793 {
1794 struct nft_pipapo_match *m;
1795
1796 m = container_of(rcu, struct nft_pipapo_match, rcu);
1797 pipapo_free_match(m);
1798 }
1799
1800 /**
1801 * nft_pipapo_commit() - Replace lookup data with current working copy
1802 * @set: nftables API set representation
1803 *
1804 * While at it, check if we should perform garbage collection on the working
1805 * copy before committing it for lookup, and don't replace the table if the
1806 * working copy doesn't have pending changes.
1807 *
1808 * We also need to create a new working copy for subsequent insertions and
1809 * deletions.
1810 */
1811 static void nft_pipapo_commit(struct nft_set *set)
1812 {
1813 struct nft_pipapo *priv = nft_set_priv(set);
1814 struct nft_pipapo_match *old;
1815
1816 if (!priv->clone)
1817 return;
1818
1819 if (time_after_eq(jiffies, priv->last_gc + nft_set_gc_interval(set)))
1820 pipapo_gc(set, priv->clone);
1821
1822 old = rcu_replace_pointer(priv->match, priv->clone,
1823 nft_pipapo_transaction_mutex_held(set));
1824 priv->clone = NULL;
1825
1826 if (old)
1827 call_rcu(&old->rcu, pipapo_reclaim_match);
1828 }
1829
1830 static void nft_pipapo_abort(const struct nft_set *set)
1831 {
1832 struct nft_pipapo *priv = nft_set_priv(set);
1833
1834 if (!priv->clone)
1835 return;
1836 pipapo_free_match(priv->clone);
1837 priv->clone = NULL;
1838 }
1839
1840 /**
1841 * nft_pipapo_activate() - Mark element reference as active given key, commit
1842 * @net: Network namespace
1843 * @set: nftables API set representation
1844 * @elem_priv: nftables API element representation containing key data
1845 *
1846 * On insertion, elements are added to a copy of the matching data currently
1847 * in use for lookups, and not directly inserted into current lookup data. Both
1848 * nft_pipapo_insert() and nft_pipapo_activate() are called once for each
1849 * element, hence we can't purpose either one as a real commit operation.
1850 */
1851 static void nft_pipapo_activate(const struct net *net,
1852 const struct nft_set *set,
1853 struct nft_elem_priv *elem_priv)
1854 {
1855 struct nft_pipapo_elem *e = nft_elem_priv_cast(elem_priv);
1856
1857 nft_clear(net, &e->ext);
1858 }
1859
1860 /**
1861 * nft_pipapo_deactivate() - Search for element and make it inactive
1862 * @net: Network namespace
1863 * @set: nftables API set representation
1864 * @elem: nftables API element representation containing key data
1865 *
1866 * Return: deactivated element if found, NULL otherwise.
1867 */
1868 static struct nft_elem_priv *
1869 nft_pipapo_deactivate(const struct net *net, const struct nft_set *set,
1870 const struct nft_set_elem *elem)
1871 {
1872 struct nft_pipapo_match *m = pipapo_maybe_clone(set);
1873 struct nft_pipapo_elem *e;
1874
1875 /* removal must occur on priv->clone, if we are low on memory
1876 * we have no choice and must fail the removal request.
1877 */
1878 if (!m)
1879 return NULL;
1880
1881 e = pipapo_get(net, set, m, (const u8 *)elem->key.val.data,
1882 nft_genmask_next(net), nft_net_tstamp(net), GFP_KERNEL);
1883 if (IS_ERR(e))
1884 return NULL;
1885
1886 nft_set_elem_change_active(net, set, &e->ext);
1887
1888 return &e->priv;
1889 }
1890
1891 /**
1892 * nft_pipapo_flush() - make element inactive
1893 * @net: Network namespace
1894 * @set: nftables API set representation
1895 * @elem_priv: nftables API element representation containing key data
1896 *
1897 * This is functionally the same as nft_pipapo_deactivate(), with a slightly
1898 * different interface, and it's also called once for each element in a set
1899 * being flushed, so we can't implement, strictly speaking, a flush operation,
1900 * which would otherwise be as simple as allocating an empty copy of the
1901 * matching data.
1902 *
1903 * Note that we could in theory do that, mark the set as flushed, and ignore
1904 * subsequent calls, but we would leak all the elements after the first one,
1905 * because they wouldn't then be freed as result of API calls.
1906 *
1907 * Return: true if element was found and deactivated.
1908 */
1909 static void nft_pipapo_flush(const struct net *net, const struct nft_set *set,
1910 struct nft_elem_priv *elem_priv)
1911 {
1912 struct nft_pipapo_elem *e = nft_elem_priv_cast(elem_priv);
1913
1914 nft_set_elem_change_active(net, set, &e->ext);
1915 }
1916
1917 /**
1918 * pipapo_get_boundaries() - Get byte interval for associated rules
1919 * @f: Field including lookup table
1920 * @first_rule: First rule (lowest index)
1921 * @rule_count: Number of associated rules
1922 * @left: Byte expression for left boundary (start of range)
1923 * @right: Byte expression for right boundary (end of range)
1924 *
1925 * Given the first rule and amount of rules that originated from the same entry,
1926 * build the original range associated with the entry, and calculate the length
1927 * of the originating netmask.
1928 *
1929 * In pictures:
1930 *
1931 * bucket
1932 * group 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
1933 * 0 1,2
1934 * 1 1,2
1935 * 2 1,2
1936 * 3 1,2
1937 * 4 1,2
1938 * 5 1 2
1939 * 6 1,2 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1940 * 7 1,2 1,2 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1941 *
1942 * this is the lookup table corresponding to the IPv4 range
1943 * 192.168.1.0-192.168.2.1, which was expanded to the two composing netmasks,
1944 * rule #1: 192.168.1.0/24, and rule #2: 192.168.2.0/31.
1945 *
1946 * This function fills @left and @right with the byte values of the leftmost
1947 * and rightmost bucket indices for the lowest and highest rule indices,
1948 * respectively. If @first_rule is 1 and @rule_count is 2, we obtain, in
1949 * nibbles:
1950 * left: < 12, 0, 10, 8, 0, 1, 0, 0 >
1951 * right: < 12, 0, 10, 8, 0, 2, 2, 1 >
1952 * corresponding to bytes:
1953 * left: < 192, 168, 1, 0 >
1954 * right: < 192, 168, 2, 1 >
1955 * with mask length irrelevant here, unused on return, as the range is already
1956 * defined by its start and end points. The mask length is relevant for a single
1957 * ranged entry instead: if @first_rule is 1 and @rule_count is 1, we ignore
1958 * rule 2 above: @left becomes < 192, 168, 1, 0 >, @right becomes
1959 * < 192, 168, 1, 255 >, and the mask length, calculated from the distances
1960 * between leftmost and rightmost bucket indices for each group, would be 24.
1961 *
1962 * Return: mask length, in bits.
1963 */
1964 static int pipapo_get_boundaries(struct nft_pipapo_field *f, int first_rule,
1965 int rule_count, u8 *left, u8 *right)
1966 {
1967 int g, mask_len = 0, bit_offset = 0;
1968 u8 *l = left, *r = right;
1969
1970 for (g = 0; g < f->groups; g++) {
1971 int b, x0, x1;
1972
1973 x0 = -1;
1974 x1 = -1;
1975 for (b = 0; b < NFT_PIPAPO_BUCKETS(f->bb); b++) {
1976 unsigned long *pos;
1977
1978 pos = NFT_PIPAPO_LT_ALIGN(f->lt) +
1979 (g * NFT_PIPAPO_BUCKETS(f->bb) + b) * f->bsize;
1980 if (test_bit(first_rule, pos) && x0 == -1)
1981 x0 = b;
1982 if (test_bit(first_rule + rule_count - 1, pos))
1983 x1 = b;
1984 }
1985
1986 *l |= x0 << (BITS_PER_BYTE - f->bb - bit_offset);
1987 *r |= x1 << (BITS_PER_BYTE - f->bb - bit_offset);
1988
1989 bit_offset += f->bb;
1990 if (bit_offset >= BITS_PER_BYTE) {
1991 bit_offset %= BITS_PER_BYTE;
1992 l++;
1993 r++;
1994 }
1995
1996 if (x1 - x0 == 0)
1997 mask_len += 4;
1998 else if (x1 - x0 == 1)
1999 mask_len += 3;
2000 else if (x1 - x0 == 3)
2001 mask_len += 2;
2002 else if (x1 - x0 == 7)
2003 mask_len += 1;
2004 }
2005
2006 return mask_len;
2007 }
2008
2009 /**
2010 * pipapo_match_field() - Match rules against byte ranges
2011 * @f: Field including the lookup table
2012 * @first_rule: First of associated rules originating from same entry
2013 * @rule_count: Amount of associated rules
2014 * @start: Start of range to be matched
2015 * @end: End of range to be matched
2016 *
2017 * Return: true on match, false otherwise.
2018 */
2019 static bool pipapo_match_field(struct nft_pipapo_field *f,
2020 int first_rule, int rule_count,
2021 const u8 *start, const u8 *end)
2022 {
2023 u8 right[NFT_PIPAPO_MAX_BYTES] = { 0 };
2024 u8 left[NFT_PIPAPO_MAX_BYTES] = { 0 };
2025
2026 pipapo_get_boundaries(f, first_rule, rule_count, left, right);
2027
2028 return !memcmp(start, left,
2029 f->groups / NFT_PIPAPO_GROUPS_PER_BYTE(f)) &&
2030 !memcmp(end, right, f->groups / NFT_PIPAPO_GROUPS_PER_BYTE(f));
2031 }
2032
2033 /**
2034 * nft_pipapo_remove() - Remove element given key, commit
2035 * @net: Network namespace
2036 * @set: nftables API set representation
2037 * @elem_priv: nftables API element representation containing key data
2038 *
2039 * Similarly to nft_pipapo_activate(), this is used as commit operation by the
2040 * API, but it's called once per element in the pending transaction, so we can't
2041 * implement this as a single commit operation. Closest we can get is to remove
2042 * the matched element here, if any, and commit the updated matching data.
2043 */
2044 static void nft_pipapo_remove(const struct net *net, const struct nft_set *set,
2045 struct nft_elem_priv *elem_priv)
2046 {
2047 struct nft_pipapo *priv = nft_set_priv(set);
2048 struct nft_pipapo_match *m = priv->clone;
2049 unsigned int rules_f0, first_rule = 0;
2050 struct nft_pipapo_elem *e;
2051 const u8 *data;
2052
2053 e = nft_elem_priv_cast(elem_priv);
2054 data = (const u8 *)nft_set_ext_key(&e->ext);
2055
2056 while ((rules_f0 = pipapo_rules_same_key(m->f, first_rule))) {
2057 union nft_pipapo_map_bucket rulemap[NFT_PIPAPO_MAX_FIELDS];
2058 const u8 *match_start, *match_end;
2059 struct nft_pipapo_field *f;
2060 int i, start, rules_fx;
2061
2062 match_start = data;
2063
2064 if (nft_set_ext_exists(&e->ext, NFT_SET_EXT_KEY_END))
2065 match_end = (const u8 *)nft_set_ext_key_end(&e->ext)->data;
2066 else
2067 match_end = data;
2068
2069 start = first_rule;
2070 rules_fx = rules_f0;
2071
2072 nft_pipapo_for_each_field(f, i, m) {
2073 bool last = i == m->field_count - 1;
2074
2075 if (!pipapo_match_field(f, start, rules_fx,
2076 match_start, match_end))
2077 break;
2078
2079 rulemap[i].to = start;
2080 rulemap[i].n = rules_fx;
2081
2082 rules_fx = f->mt[start].n;
2083 start = f->mt[start].to;
2084
2085 match_start += NFT_PIPAPO_GROUPS_PADDED_SIZE(f);
2086 match_end += NFT_PIPAPO_GROUPS_PADDED_SIZE(f);
2087
2088 if (last && f->mt[rulemap[i].to].e == e) {
2089 pipapo_drop(m, rulemap);
2090 return;
2091 }
2092 }
2093
2094 first_rule += rules_f0;
2095 }
2096
2097 WARN_ON_ONCE(1); /* elem_priv not found */
2098 }
2099
2100 /**
2101 * nft_pipapo_do_walk() - Walk over elements in m
2102 * @ctx: nftables API context
2103 * @set: nftables API set representation
2104 * @m: matching data pointing to key mapping array
2105 * @iter: Iterator
2106 *
2107 * As elements are referenced in the mapping array for the last field, directly
2108 * scan that array: there's no need to follow rule mappings from the first
2109 * field. @m is protected either by RCU read lock or by transaction mutex.
2110 */
2111 static void nft_pipapo_do_walk(const struct nft_ctx *ctx, struct nft_set *set,
2112 const struct nft_pipapo_match *m,
2113 struct nft_set_iter *iter)
2114 {
2115 const struct nft_pipapo_field *f;
2116 unsigned int i, r;
2117
2118 for (i = 0, f = m->f; i < m->field_count - 1; i++, f++)
2119 ;
2120
2121 for (r = 0; r < f->rules; r++) {
2122 struct nft_pipapo_elem *e;
2123
2124 if (r < f->rules - 1 && f->mt[r + 1].e == f->mt[r].e)
2125 continue;
2126
2127 if (iter->count < iter->skip)
2128 goto cont;
2129
2130 e = f->mt[r].e;
2131
2132 iter->err = iter->fn(ctx, set, iter, &e->priv);
2133 if (iter->err < 0)
2134 return;
2135
2136 cont:
2137 iter->count++;
2138 }
2139 }
2140
2141 /**
2142 * nft_pipapo_walk() - Walk over elements
2143 * @ctx: nftables API context
2144 * @set: nftables API set representation
2145 * @iter: Iterator
2146 *
2147 * Test if destructive action is needed or not, clone active backend if needed
2148 * and call the real function to work on the data.
2149 */
2150 static void nft_pipapo_walk(const struct nft_ctx *ctx, struct nft_set *set,
2151 struct nft_set_iter *iter)
2152 {
2153 struct nft_pipapo *priv = nft_set_priv(set);
2154 const struct nft_pipapo_match *m;
2155
2156 switch (iter->type) {
2157 case NFT_ITER_UPDATE:
2158 m = pipapo_maybe_clone(set);
2159 if (!m) {
2160 iter->err = -ENOMEM;
2161 return;
2162 }
2163
2164 nft_pipapo_do_walk(ctx, set, m, iter);
2165 break;
2166 case NFT_ITER_READ:
2167 rcu_read_lock();
2168 m = rcu_dereference(priv->match);
2169 nft_pipapo_do_walk(ctx, set, m, iter);
2170 rcu_read_unlock();
2171 break;
2172 default:
2173 iter->err = -EINVAL;
2174 WARN_ON_ONCE(1);
2175 break;
2176 }
2177 }
2178
2179 /**
2180 * nft_pipapo_privsize() - Return the size of private data for the set
2181 * @nla: netlink attributes, ignored as size doesn't depend on them
2182 * @desc: Set description, ignored as size doesn't depend on it
2183 *
2184 * Return: size of private data for this set implementation, in bytes
2185 */
2186 static u64 nft_pipapo_privsize(const struct nlattr * const nla[],
2187 const struct nft_set_desc *desc)
2188 {
2189 return sizeof(struct nft_pipapo);
2190 }
2191
2192 /**
2193 * nft_pipapo_estimate() - Set size, space and lookup complexity
2194 * @desc: Set description, element count and field description used
2195 * @features: Flags: NFT_SET_INTERVAL needs to be there
2196 * @est: Storage for estimation data
2197 *
2198 * Return: true if set description is compatible, false otherwise
2199 */
2200 static bool nft_pipapo_estimate(const struct nft_set_desc *desc, u32 features,
2201 struct nft_set_estimate *est)
2202 {
2203 if (!(features & NFT_SET_INTERVAL) ||
2204 desc->field_count < NFT_PIPAPO_MIN_FIELDS)
2205 return false;
2206
2207 est->size = pipapo_estimate_size(desc);
2208 if (!est->size)
2209 return false;
2210
2211 est->lookup = NFT_SET_CLASS_O_LOG_N;
2212
2213 est->space = NFT_SET_CLASS_O_N;
2214
2215 return true;
2216 }
2217
2218 /**
2219 * nft_pipapo_init() - Initialise data for a set instance
2220 * @set: nftables API set representation
2221 * @desc: Set description
2222 * @nla: netlink attributes
2223 *
2224 * Validate number and size of fields passed as NFTA_SET_DESC_CONCAT netlink
2225 * attributes, initialise internal set parameters, current instance of matching
2226 * data and a copy for subsequent insertions.
2227 *
2228 * Return: 0 on success, negative error code on failure.
2229 */
2230 static int nft_pipapo_init(const struct nft_set *set,
2231 const struct nft_set_desc *desc,
2232 const struct nlattr * const nla[])
2233 {
2234 struct nft_pipapo *priv = nft_set_priv(set);
2235 struct nft_pipapo_match *m;
2236 struct nft_pipapo_field *f;
2237 int err, i, field_count;
2238
2239 BUILD_BUG_ON(offsetof(struct nft_pipapo_elem, priv) != 0);
2240
2241 field_count = desc->field_count ? : 1;
2242
2243 BUILD_BUG_ON(NFT_PIPAPO_MAX_FIELDS > 255);
2244 BUILD_BUG_ON(NFT_PIPAPO_MAX_FIELDS != NFT_REG32_COUNT);
2245
2246 if (field_count > NFT_PIPAPO_MAX_FIELDS)
2247 return -EINVAL;
2248
2249 m = kmalloc(struct_size(m, f, field_count), GFP_KERNEL);
2250 if (!m)
2251 return -ENOMEM;
2252
2253 m->field_count = field_count;
2254 m->bsize_max = 0;
2255
2256 m->scratch = alloc_percpu(struct nft_pipapo_scratch *);
2257 if (!m->scratch) {
2258 err = -ENOMEM;
2259 goto out_scratch;
2260 }
2261 for_each_possible_cpu(i)
2262 *per_cpu_ptr(m->scratch, i) = NULL;
2263
2264 rcu_head_init(&m->rcu);
2265
2266 nft_pipapo_for_each_field(f, i, m) {
2267 unsigned int len = desc->field_len[i] ? : set->klen;
2268
2269 /* f->groups is u8 */
2270 BUILD_BUG_ON((NFT_PIPAPO_MAX_BYTES *
2271 BITS_PER_BYTE / NFT_PIPAPO_GROUP_BITS_LARGE_SET) >= 256);
2272
2273 f->bb = NFT_PIPAPO_GROUP_BITS_INIT;
2274 f->groups = len * NFT_PIPAPO_GROUPS_PER_BYTE(f);
2275
2276 priv->width += round_up(len, sizeof(u32));
2277
2278 f->bsize = 0;
2279 f->rules = 0;
2280 f->rules_alloc = 0;
2281 f->lt = NULL;
2282 f->mt = NULL;
2283 }
2284
2285 rcu_assign_pointer(priv->match, m);
2286
2287 return 0;
2288
2289 out_scratch:
2290 kfree(m);
2291
2292 return err;
2293 }
2294
2295 /**
2296 * nft_set_pipapo_match_destroy() - Destroy elements from key mapping array
2297 * @ctx: context
2298 * @set: nftables API set representation
2299 * @m: matching data pointing to key mapping array
2300 */
2301 static void nft_set_pipapo_match_destroy(const struct nft_ctx *ctx,
2302 const struct nft_set *set,
2303 struct nft_pipapo_match *m)
2304 {
2305 struct nft_pipapo_field *f;
2306 unsigned int i, r;
2307
2308 for (i = 0, f = m->f; i < m->field_count - 1; i++, f++)
2309 ;
2310
2311 for (r = 0; r < f->rules; r++) {
2312 struct nft_pipapo_elem *e;
2313
2314 if (r < f->rules - 1 && f->mt[r + 1].e == f->mt[r].e)
2315 continue;
2316
2317 e = f->mt[r].e;
2318
2319 nf_tables_set_elem_destroy(ctx, set, &e->priv);
2320 }
2321 }
2322
2323 /**
2324 * nft_pipapo_destroy() - Free private data for set and all committed elements
2325 * @ctx: context
2326 * @set: nftables API set representation
2327 */
2328 static void nft_pipapo_destroy(const struct nft_ctx *ctx,
2329 const struct nft_set *set)
2330 {
2331 struct nft_pipapo *priv = nft_set_priv(set);
2332 struct nft_pipapo_match *m;
2333
2334 m = rcu_dereference_protected(priv->match, true);
2335
2336 if (priv->clone) {
2337 nft_set_pipapo_match_destroy(ctx, set, priv->clone);
2338 pipapo_free_match(priv->clone);
2339 priv->clone = NULL;
2340 } else {
2341 nft_set_pipapo_match_destroy(ctx, set, m);
2342 }
2343
2344 pipapo_free_match(m);
2345 }
2346
2347 /**
2348 * nft_pipapo_gc_init() - Initialise garbage collection
2349 * @set: nftables API set representation
2350 *
2351 * Instead of actually setting up a periodic work for garbage collection, as
2352 * this operation requires a swap of matching data with the working copy, we'll
2353 * do that opportunistically with other commit operations if the interval is
2354 * elapsed, so we just need to set the current jiffies timestamp here.
2355 */
2356 static void nft_pipapo_gc_init(const struct nft_set *set)
2357 {
2358 struct nft_pipapo *priv = nft_set_priv(set);
2359
2360 priv->last_gc = jiffies;
2361 }
2362
2363 const struct nft_set_type nft_set_pipapo_type = {
2364 .features = NFT_SET_INTERVAL | NFT_SET_MAP | NFT_SET_OBJECT |
2365 NFT_SET_TIMEOUT,
2366 .ops = {
2367 .lookup = nft_pipapo_lookup,
2368 .insert = nft_pipapo_insert,
2369 .activate = nft_pipapo_activate,
2370 .deactivate = nft_pipapo_deactivate,
2371 .flush = nft_pipapo_flush,
2372 .remove = nft_pipapo_remove,
2373 .walk = nft_pipapo_walk,
2374 .get = nft_pipapo_get,
2375 .privsize = nft_pipapo_privsize,
2376 .estimate = nft_pipapo_estimate,
2377 .init = nft_pipapo_init,
2378 .destroy = nft_pipapo_destroy,
2379 .gc_init = nft_pipapo_gc_init,
2380 .commit = nft_pipapo_commit,
2381 .abort = nft_pipapo_abort,
2382 .elemsize = offsetof(struct nft_pipapo_elem, ext),
2383 },
2384 };
2385
2386 #if defined(CONFIG_X86_64) && !defined(CONFIG_UML)
2387 const struct nft_set_type nft_set_pipapo_avx2_type = {
2388 .features = NFT_SET_INTERVAL | NFT_SET_MAP | NFT_SET_OBJECT |
2389 NFT_SET_TIMEOUT,
2390 .ops = {
2391 .lookup = nft_pipapo_avx2_lookup,
2392 .insert = nft_pipapo_insert,
2393 .activate = nft_pipapo_activate,
2394 .deactivate = nft_pipapo_deactivate,
2395 .flush = nft_pipapo_flush,
2396 .remove = nft_pipapo_remove,
2397 .walk = nft_pipapo_walk,
2398 .get = nft_pipapo_get,
2399 .privsize = nft_pipapo_privsize,
2400 .estimate = nft_pipapo_avx2_estimate,
2401 .init = nft_pipapo_init,
2402 .destroy = nft_pipapo_destroy,
2403 .gc_init = nft_pipapo_gc_init,
2404 .commit = nft_pipapo_commit,
2405 .abort = nft_pipapo_abort,
2406 .elemsize = offsetof(struct nft_pipapo_elem, ext),
2407 },
2408 };
2409 #endif
2410