1 // SPDX-License-Identifier: GPL-2.0-only
3 /* PIPAPO: PIle PAcket POlicies: AVX2 packet lookup routines
5 * Copyright (c) 2019-2020 Red Hat GmbH
7 * Author: Stefano Brivio <sbrivio@redhat.com>
10 #include <linux/kernel.h>
11 #include <linux/init.h>
12 #include <linux/module.h>
13 #include <linux/netlink.h>
14 #include <linux/netfilter.h>
15 #include <linux/netfilter/nf_tables.h>
16 #include <net/netfilter/nf_tables_core.h>
17 #include <uapi/linux/netfilter/nf_tables.h>
18 #include <linux/bitmap.h>
19 #include <linux/bitops.h>
21 #include <linux/compiler.h>
22 #include <asm/fpu/api.h>
24 #include "nft_set_pipapo_avx2.h"
25 #include "nft_set_pipapo.h"
27 #define NFT_PIPAPO_LONGS_PER_M256 (XSAVE_YMM_SIZE / BITS_PER_LONG)
29 /* Load from memory into YMM register with non-temporal hint ("stream load"),
30 * that is, don't fetch lines from memory into the cache. This avoids pushing
31 * precious packet data out of the cache hierarchy, and is appropriate when:
33 * - loading buckets from lookup tables, as they are not going to be used
34 * again before packets are entirely classified
36 * - loading the result bitmap from the previous field, as it's never used
39 #define NFT_PIPAPO_AVX2_LOAD(reg, loc) \
40 asm volatile("vmovntdqa %0, %%ymm" #reg : : "m" (loc))
42 /* Stream a single lookup table bucket into YMM register given lookup table,
43 * group index, value of packet bits, bucket size.
45 #define NFT_PIPAPO_AVX2_BUCKET_LOAD4(reg, lt, group, v, bsize) \
46 NFT_PIPAPO_AVX2_LOAD(reg, \
47 lt[((group) * NFT_PIPAPO_BUCKETS(4) + \
49 #define NFT_PIPAPO_AVX2_BUCKET_LOAD8(reg, lt, group, v, bsize) \
50 NFT_PIPAPO_AVX2_LOAD(reg, \
51 lt[((group) * NFT_PIPAPO_BUCKETS(8) + \
54 /* Bitwise AND: the staple operation of this algorithm */
55 #define NFT_PIPAPO_AVX2_AND(dst, a, b) \
56 asm volatile("vpand %ymm" #a ", %ymm" #b ", %ymm" #dst)
58 /* Jump to label if @reg is zero */
59 #define NFT_PIPAPO_AVX2_NOMATCH_GOTO(reg, label) \
60 asm goto("vptest %%ymm" #reg ", %%ymm" #reg ";" \
61 "je %l[" #label "]" : : : : label)
63 /* Store 256 bits from YMM register into memory. Contrary to bucket load
64 * operation, we don't bypass the cache here, as stored matching results
65 * are always used shortly after.
67 #define NFT_PIPAPO_AVX2_STORE(loc, reg) \
68 asm volatile("vmovdqa %%ymm" #reg ", %0" : "=m" (loc))
70 /* Zero out a complete YMM register, @reg */
71 #define NFT_PIPAPO_AVX2_ZERO(reg) \
72 asm volatile("vpxor %ymm" #reg ", %ymm" #reg ", %ymm" #reg)
75 * nft_pipapo_avx2_prepare() - Prepare before main algorithm body
77 * This zeroes out ymm15, which is later used whenever we need to clear a
78 * memory location, by storing its content into memory.
80 static void nft_pipapo_avx2_prepare(void)
82 NFT_PIPAPO_AVX2_ZERO(15);
86 * nft_pipapo_avx2_fill() - Fill a bitmap region with ones
87 * @data: Base memory area
88 * @start: First bit to set
89 * @len: Count of bits to fill
91 * This is nothing else than a version of bitmap_set(), as used e.g. by
92 * pipapo_refill(), tailored for the microarchitectures using it and better
93 * suited for the specific usage: it's very likely that we'll set a small number
94 * of bits, not crossing a word boundary, and correct branch prediction is
97 * This function doesn't actually use any AVX2 instruction.
99 static void nft_pipapo_avx2_fill(unsigned long *data, int start, int len)
101 int offset = start % BITS_PER_LONG;
104 data += start / BITS_PER_LONG;
106 if (likely(len == 1)) {
107 *data |= BIT(offset);
111 if (likely(len < BITS_PER_LONG || offset)) {
112 if (likely(len + offset <= BITS_PER_LONG)) {
113 *data |= GENMASK(len - 1 + offset, offset);
117 *data |= ~0UL << offset;
118 len -= BITS_PER_LONG - offset;
121 if (len <= BITS_PER_LONG) {
122 mask = ~0UL >> (BITS_PER_LONG - len);
128 memset(data, 0xff, len / BITS_PER_BYTE);
129 data += len / BITS_PER_LONG;
131 len %= BITS_PER_LONG;
133 *data |= ~0UL >> (BITS_PER_LONG - len);
137 * nft_pipapo_avx2_refill() - Scan bitmap, select mapping table item, set bits
138 * @offset: Start from given bitmap (equivalent to bucket) offset, in longs
139 * @map: Bitmap to be scanned for set bits
140 * @dst: Destination bitmap
141 * @mt: Mapping table containing bit set specifiers
142 * @last: Return index of first set bit, if this is the last field
144 * This is an alternative implementation of pipapo_refill() suitable for usage
145 * with AVX2 lookup routines: we know there are four words to be scanned, at
146 * a given offset inside the map, for each matching iteration.
148 * This function doesn't actually use any AVX2 instruction.
150 * Return: first set bit index if @last, index of first filled word otherwise.
152 static int nft_pipapo_avx2_refill(int offset, unsigned long *map,
154 union nft_pipapo_map_bucket *mt, bool last)
158 #define NFT_PIPAPO_AVX2_REFILL_ONE_WORD(x) \
161 int r = __builtin_ctzl(map[(x)]); \
162 int i = (offset + (x)) * BITS_PER_LONG + r; \
167 nft_pipapo_avx2_fill(dst, mt[i].to, mt[i].n); \
172 map[(x)] &= ~(1UL << r); \
176 NFT_PIPAPO_AVX2_REFILL_ONE_WORD(0);
177 NFT_PIPAPO_AVX2_REFILL_ONE_WORD(1);
178 NFT_PIPAPO_AVX2_REFILL_ONE_WORD(2);
179 NFT_PIPAPO_AVX2_REFILL_ONE_WORD(3);
180 #undef NFT_PIPAPO_AVX2_REFILL_ONE_WORD
186 * nft_pipapo_avx2_lookup_4b_2() - AVX2-based lookup for 2 four-bit groups
187 * @map: Previous match result, used as initial bitmap
188 * @fill: Destination bitmap to be filled with current match result
189 * @f: Field, containing lookup and mapping tables
190 * @offset: Ignore buckets before the given index, no bits are filled there
191 * @pkt: Packet data, pointer to input nftables register
192 * @first: If this is the first field, don't source previous result
193 * @last: Last field: stop at the first match and return bit index
195 * Load buckets from lookup table corresponding to the values of each 4-bit
196 * group of packet bytes, and perform a bitwise intersection between them. If
197 * this is the first field in the set, simply AND the buckets together
198 * (equivalent to using an all-ones starting bitmap), use the provided starting
199 * bitmap otherwise. Then call nft_pipapo_avx2_refill() to generate the next
200 * working bitmap, @fill.
202 * This is used for 8-bit fields (i.e. protocol numbers).
204 * Out-of-order (and superscalar) execution is vital here, so it's critical to
205 * avoid false data dependencies. CPU and compiler could (mostly) take care of
206 * this on their own, but the operation ordering is explicitly given here with
207 * a likely execution order in mind, to highlight possible stalls. That's why
208 * a number of logically distinct operations (i.e. loading buckets, intersecting
209 * buckets) are interleaved.
211 * Return: -1 on no match, rule index of match if @last, otherwise first long
212 * word index to be checked next (i.e. first filled word).
214 static int nft_pipapo_avx2_lookup_4b_2(unsigned long *map, unsigned long *fill,
215 const struct nft_pipapo_field *f,
216 int offset, const u8 *pkt,
217 bool first, bool last)
219 int i, ret = -1, m256_size = f->bsize / NFT_PIPAPO_LONGS_PER_M256, b;
220 u8 pg[2] = { pkt[0] >> 4, pkt[0] & 0xf };
221 unsigned long *lt = f->lt, bsize = f->bsize;
223 lt += offset * NFT_PIPAPO_LONGS_PER_M256;
224 for (i = offset; i < m256_size; i++, lt += NFT_PIPAPO_LONGS_PER_M256) {
225 int i_ul = i * NFT_PIPAPO_LONGS_PER_M256;
228 NFT_PIPAPO_AVX2_BUCKET_LOAD4(0, lt, 0, pg[0], bsize);
229 NFT_PIPAPO_AVX2_BUCKET_LOAD4(1, lt, 1, pg[1], bsize);
230 NFT_PIPAPO_AVX2_AND(4, 0, 1);
232 NFT_PIPAPO_AVX2_BUCKET_LOAD4(0, lt, 0, pg[0], bsize);
233 NFT_PIPAPO_AVX2_LOAD(2, map[i_ul]);
234 NFT_PIPAPO_AVX2_BUCKET_LOAD4(1, lt, 1, pg[1], bsize);
235 NFT_PIPAPO_AVX2_NOMATCH_GOTO(2, nothing);
236 NFT_PIPAPO_AVX2_AND(3, 0, 1);
237 NFT_PIPAPO_AVX2_AND(4, 2, 3);
240 NFT_PIPAPO_AVX2_NOMATCH_GOTO(4, nomatch);
241 NFT_PIPAPO_AVX2_STORE(map[i_ul], 4);
243 b = nft_pipapo_avx2_refill(i_ul, &map[i_ul], fill, f->mt, last);
247 if (unlikely(ret == -1))
248 ret = b / XSAVE_YMM_SIZE;
252 NFT_PIPAPO_AVX2_STORE(map[i_ul], 15);
261 * nft_pipapo_avx2_lookup_4b_4() - AVX2-based lookup for 4 four-bit groups
262 * @map: Previous match result, used as initial bitmap
263 * @fill: Destination bitmap to be filled with current match result
264 * @f: Field, containing lookup and mapping tables
265 * @offset: Ignore buckets before the given index, no bits are filled there
266 * @pkt: Packet data, pointer to input nftables register
267 * @first: If this is the first field, don't source previous result
268 * @last: Last field: stop at the first match and return bit index
270 * See nft_pipapo_avx2_lookup_4b_2().
272 * This is used for 16-bit fields (i.e. ports).
274 * Return: -1 on no match, rule index of match if @last, otherwise first long
275 * word index to be checked next (i.e. first filled word).
277 static int nft_pipapo_avx2_lookup_4b_4(unsigned long *map, unsigned long *fill,
278 const struct nft_pipapo_field *f,
279 int offset, const u8 *pkt,
280 bool first, bool last)
282 int i, ret = -1, m256_size = f->bsize / NFT_PIPAPO_LONGS_PER_M256, b;
283 u8 pg[4] = { pkt[0] >> 4, pkt[0] & 0xf, pkt[1] >> 4, pkt[1] & 0xf };
284 unsigned long *lt = f->lt, bsize = f->bsize;
286 lt += offset * NFT_PIPAPO_LONGS_PER_M256;
287 for (i = offset; i < m256_size; i++, lt += NFT_PIPAPO_LONGS_PER_M256) {
288 int i_ul = i * NFT_PIPAPO_LONGS_PER_M256;
291 NFT_PIPAPO_AVX2_BUCKET_LOAD4(0, lt, 0, pg[0], bsize);
292 NFT_PIPAPO_AVX2_BUCKET_LOAD4(1, lt, 1, pg[1], bsize);
293 NFT_PIPAPO_AVX2_BUCKET_LOAD4(2, lt, 2, pg[2], bsize);
294 NFT_PIPAPO_AVX2_BUCKET_LOAD4(3, lt, 3, pg[3], bsize);
295 NFT_PIPAPO_AVX2_AND(4, 0, 1);
296 NFT_PIPAPO_AVX2_AND(5, 2, 3);
297 NFT_PIPAPO_AVX2_AND(7, 4, 5);
299 NFT_PIPAPO_AVX2_BUCKET_LOAD4(0, lt, 0, pg[0], bsize);
301 NFT_PIPAPO_AVX2_LOAD(1, map[i_ul]);
303 NFT_PIPAPO_AVX2_BUCKET_LOAD4(2, lt, 1, pg[1], bsize);
304 NFT_PIPAPO_AVX2_BUCKET_LOAD4(3, lt, 2, pg[2], bsize);
305 NFT_PIPAPO_AVX2_BUCKET_LOAD4(4, lt, 3, pg[3], bsize);
306 NFT_PIPAPO_AVX2_AND(5, 0, 1);
308 NFT_PIPAPO_AVX2_NOMATCH_GOTO(1, nothing);
310 NFT_PIPAPO_AVX2_AND(6, 2, 3);
311 NFT_PIPAPO_AVX2_AND(7, 4, 5);
313 NFT_PIPAPO_AVX2_AND(7, 6, 7);
317 NFT_PIPAPO_AVX2_NOMATCH_GOTO(7, nomatch);
318 NFT_PIPAPO_AVX2_STORE(map[i_ul], 7);
320 b = nft_pipapo_avx2_refill(i_ul, &map[i_ul], fill, f->mt, last);
324 if (unlikely(ret == -1))
325 ret = b / XSAVE_YMM_SIZE;
329 NFT_PIPAPO_AVX2_STORE(map[i_ul], 15);
338 * nft_pipapo_avx2_lookup_4b_8() - AVX2-based lookup for 8 four-bit groups
339 * @map: Previous match result, used as initial bitmap
340 * @fill: Destination bitmap to be filled with current match result
341 * @f: Field, containing lookup and mapping tables
342 * @offset: Ignore buckets before the given index, no bits are filled there
343 * @pkt: Packet data, pointer to input nftables register
344 * @first: If this is the first field, don't source previous result
345 * @last: Last field: stop at the first match and return bit index
347 * See nft_pipapo_avx2_lookup_4b_2().
349 * This is used for 32-bit fields (i.e. IPv4 addresses).
351 * Return: -1 on no match, rule index of match if @last, otherwise first long
352 * word index to be checked next (i.e. first filled word).
354 static int nft_pipapo_avx2_lookup_4b_8(unsigned long *map, unsigned long *fill,
355 const struct nft_pipapo_field *f,
356 int offset, const u8 *pkt,
357 bool first, bool last)
359 u8 pg[8] = { pkt[0] >> 4, pkt[0] & 0xf, pkt[1] >> 4, pkt[1] & 0xf,
360 pkt[2] >> 4, pkt[2] & 0xf, pkt[3] >> 4, pkt[3] & 0xf,
362 int i, ret = -1, m256_size = f->bsize / NFT_PIPAPO_LONGS_PER_M256, b;
363 unsigned long *lt = f->lt, bsize = f->bsize;
365 lt += offset * NFT_PIPAPO_LONGS_PER_M256;
366 for (i = offset; i < m256_size; i++, lt += NFT_PIPAPO_LONGS_PER_M256) {
367 int i_ul = i * NFT_PIPAPO_LONGS_PER_M256;
370 NFT_PIPAPO_AVX2_BUCKET_LOAD4(0, lt, 0, pg[0], bsize);
371 NFT_PIPAPO_AVX2_BUCKET_LOAD4(1, lt, 1, pg[1], bsize);
372 NFT_PIPAPO_AVX2_BUCKET_LOAD4(2, lt, 2, pg[2], bsize);
373 NFT_PIPAPO_AVX2_BUCKET_LOAD4(3, lt, 3, pg[3], bsize);
374 NFT_PIPAPO_AVX2_BUCKET_LOAD4(4, lt, 4, pg[4], bsize);
375 NFT_PIPAPO_AVX2_AND(5, 0, 1);
376 NFT_PIPAPO_AVX2_BUCKET_LOAD4(6, lt, 5, pg[5], bsize);
377 NFT_PIPAPO_AVX2_BUCKET_LOAD4(7, lt, 6, pg[6], bsize);
378 NFT_PIPAPO_AVX2_AND(8, 2, 3);
379 NFT_PIPAPO_AVX2_AND(9, 4, 5);
380 NFT_PIPAPO_AVX2_BUCKET_LOAD4(10, lt, 7, pg[7], bsize);
381 NFT_PIPAPO_AVX2_AND(11, 6, 7);
382 NFT_PIPAPO_AVX2_AND(12, 8, 9);
383 NFT_PIPAPO_AVX2_AND(13, 10, 11);
386 NFT_PIPAPO_AVX2_AND(1, 12, 13);
388 NFT_PIPAPO_AVX2_BUCKET_LOAD4(0, lt, 0, pg[0], bsize);
389 NFT_PIPAPO_AVX2_LOAD(1, map[i_ul]);
390 NFT_PIPAPO_AVX2_BUCKET_LOAD4(2, lt, 1, pg[1], bsize);
391 NFT_PIPAPO_AVX2_BUCKET_LOAD4(3, lt, 2, pg[2], bsize);
392 NFT_PIPAPO_AVX2_BUCKET_LOAD4(4, lt, 3, pg[3], bsize);
394 NFT_PIPAPO_AVX2_NOMATCH_GOTO(1, nothing);
396 NFT_PIPAPO_AVX2_AND(5, 0, 1);
397 NFT_PIPAPO_AVX2_BUCKET_LOAD4(6, lt, 4, pg[4], bsize);
398 NFT_PIPAPO_AVX2_BUCKET_LOAD4(7, lt, 5, pg[5], bsize);
399 NFT_PIPAPO_AVX2_AND(8, 2, 3);
400 NFT_PIPAPO_AVX2_BUCKET_LOAD4(9, lt, 6, pg[6], bsize);
401 NFT_PIPAPO_AVX2_AND(10, 4, 5);
402 NFT_PIPAPO_AVX2_BUCKET_LOAD4(11, lt, 7, pg[7], bsize);
403 NFT_PIPAPO_AVX2_AND(12, 6, 7);
404 NFT_PIPAPO_AVX2_AND(13, 8, 9);
405 NFT_PIPAPO_AVX2_AND(14, 10, 11);
408 NFT_PIPAPO_AVX2_AND(1, 12, 13);
409 NFT_PIPAPO_AVX2_AND(1, 1, 14);
412 NFT_PIPAPO_AVX2_NOMATCH_GOTO(1, nomatch);
413 NFT_PIPAPO_AVX2_STORE(map[i_ul], 1);
415 b = nft_pipapo_avx2_refill(i_ul, &map[i_ul], fill, f->mt, last);
419 if (unlikely(ret == -1))
420 ret = b / XSAVE_YMM_SIZE;
425 NFT_PIPAPO_AVX2_STORE(map[i_ul], 15);
434 * nft_pipapo_avx2_lookup_4b_12() - AVX2-based lookup for 12 four-bit groups
435 * @map: Previous match result, used as initial bitmap
436 * @fill: Destination bitmap to be filled with current match result
437 * @f: Field, containing lookup and mapping tables
438 * @offset: Ignore buckets before the given index, no bits are filled there
439 * @pkt: Packet data, pointer to input nftables register
440 * @first: If this is the first field, don't source previous result
441 * @last: Last field: stop at the first match and return bit index
443 * See nft_pipapo_avx2_lookup_4b_2().
445 * This is used for 48-bit fields (i.e. MAC addresses/EUI-48).
447 * Return: -1 on no match, rule index of match if @last, otherwise first long
448 * word index to be checked next (i.e. first filled word).
450 static int nft_pipapo_avx2_lookup_4b_12(unsigned long *map, unsigned long *fill,
451 const struct nft_pipapo_field *f,
452 int offset, const u8 *pkt,
453 bool first, bool last)
455 u8 pg[12] = { pkt[0] >> 4, pkt[0] & 0xf, pkt[1] >> 4, pkt[1] & 0xf,
456 pkt[2] >> 4, pkt[2] & 0xf, pkt[3] >> 4, pkt[3] & 0xf,
457 pkt[4] >> 4, pkt[4] & 0xf, pkt[5] >> 4, pkt[5] & 0xf,
459 int i, ret = -1, m256_size = f->bsize / NFT_PIPAPO_LONGS_PER_M256, b;
460 unsigned long *lt = f->lt, bsize = f->bsize;
462 lt += offset * NFT_PIPAPO_LONGS_PER_M256;
463 for (i = offset; i < m256_size; i++, lt += NFT_PIPAPO_LONGS_PER_M256) {
464 int i_ul = i * NFT_PIPAPO_LONGS_PER_M256;
467 NFT_PIPAPO_AVX2_LOAD(0, map[i_ul]);
469 NFT_PIPAPO_AVX2_BUCKET_LOAD4(1, lt, 0, pg[0], bsize);
470 NFT_PIPAPO_AVX2_BUCKET_LOAD4(2, lt, 1, pg[1], bsize);
471 NFT_PIPAPO_AVX2_BUCKET_LOAD4(3, lt, 2, pg[2], bsize);
474 NFT_PIPAPO_AVX2_NOMATCH_GOTO(0, nothing);
475 NFT_PIPAPO_AVX2_AND(1, 1, 0);
478 NFT_PIPAPO_AVX2_BUCKET_LOAD4(4, lt, 3, pg[3], bsize);
479 NFT_PIPAPO_AVX2_BUCKET_LOAD4(5, lt, 4, pg[4], bsize);
480 NFT_PIPAPO_AVX2_AND(6, 2, 3);
481 NFT_PIPAPO_AVX2_BUCKET_LOAD4(7, lt, 5, pg[5], bsize);
482 NFT_PIPAPO_AVX2_BUCKET_LOAD4(8, lt, 6, pg[6], bsize);
483 NFT_PIPAPO_AVX2_AND(9, 1, 4);
484 NFT_PIPAPO_AVX2_BUCKET_LOAD4(10, lt, 7, pg[7], bsize);
485 NFT_PIPAPO_AVX2_AND(11, 5, 6);
486 NFT_PIPAPO_AVX2_BUCKET_LOAD4(12, lt, 8, pg[8], bsize);
487 NFT_PIPAPO_AVX2_AND(13, 7, 8);
488 NFT_PIPAPO_AVX2_BUCKET_LOAD4(14, lt, 9, pg[9], bsize);
490 NFT_PIPAPO_AVX2_AND(0, 9, 10);
491 NFT_PIPAPO_AVX2_BUCKET_LOAD4(1, lt, 10, pg[10], bsize);
492 NFT_PIPAPO_AVX2_AND(2, 11, 12);
493 NFT_PIPAPO_AVX2_BUCKET_LOAD4(3, lt, 11, pg[11], bsize);
494 NFT_PIPAPO_AVX2_AND(4, 13, 14);
495 NFT_PIPAPO_AVX2_AND(5, 0, 1);
497 NFT_PIPAPO_AVX2_AND(6, 2, 3);
500 NFT_PIPAPO_AVX2_AND(7, 4, 5);
501 NFT_PIPAPO_AVX2_AND(8, 6, 7);
503 NFT_PIPAPO_AVX2_NOMATCH_GOTO(8, nomatch);
504 NFT_PIPAPO_AVX2_STORE(map[i_ul], 8);
506 b = nft_pipapo_avx2_refill(i_ul, &map[i_ul], fill, f->mt, last);
510 if (unlikely(ret == -1))
511 ret = b / XSAVE_YMM_SIZE;
515 NFT_PIPAPO_AVX2_STORE(map[i_ul], 15);
524 * nft_pipapo_avx2_lookup_4b_32() - AVX2-based lookup for 32 four-bit groups
525 * @map: Previous match result, used as initial bitmap
526 * @fill: Destination bitmap to be filled with current match result
527 * @f: Field, containing lookup and mapping tables
528 * @offset: Ignore buckets before the given index, no bits are filled there
529 * @pkt: Packet data, pointer to input nftables register
530 * @first: If this is the first field, don't source previous result
531 * @last: Last field: stop at the first match and return bit index
533 * See nft_pipapo_avx2_lookup_4b_2().
535 * This is used for 128-bit fields (i.e. IPv6 addresses).
537 * Return: -1 on no match, rule index of match if @last, otherwise first long
538 * word index to be checked next (i.e. first filled word).
540 static int nft_pipapo_avx2_lookup_4b_32(unsigned long *map, unsigned long *fill,
541 const struct nft_pipapo_field *f,
542 int offset, const u8 *pkt,
543 bool first, bool last)
545 u8 pg[32] = { pkt[0] >> 4, pkt[0] & 0xf, pkt[1] >> 4, pkt[1] & 0xf,
546 pkt[2] >> 4, pkt[2] & 0xf, pkt[3] >> 4, pkt[3] & 0xf,
547 pkt[4] >> 4, pkt[4] & 0xf, pkt[5] >> 4, pkt[5] & 0xf,
548 pkt[6] >> 4, pkt[6] & 0xf, pkt[7] >> 4, pkt[7] & 0xf,
549 pkt[8] >> 4, pkt[8] & 0xf, pkt[9] >> 4, pkt[9] & 0xf,
550 pkt[10] >> 4, pkt[10] & 0xf, pkt[11] >> 4, pkt[11] & 0xf,
551 pkt[12] >> 4, pkt[12] & 0xf, pkt[13] >> 4, pkt[13] & 0xf,
552 pkt[14] >> 4, pkt[14] & 0xf, pkt[15] >> 4, pkt[15] & 0xf,
554 int i, ret = -1, m256_size = f->bsize / NFT_PIPAPO_LONGS_PER_M256, b;
555 unsigned long *lt = f->lt, bsize = f->bsize;
557 lt += offset * NFT_PIPAPO_LONGS_PER_M256;
558 for (i = offset; i < m256_size; i++, lt += NFT_PIPAPO_LONGS_PER_M256) {
559 int i_ul = i * NFT_PIPAPO_LONGS_PER_M256;
562 NFT_PIPAPO_AVX2_LOAD(0, map[i_ul]);
564 NFT_PIPAPO_AVX2_BUCKET_LOAD4(1, lt, 0, pg[0], bsize);
565 NFT_PIPAPO_AVX2_BUCKET_LOAD4(2, lt, 1, pg[1], bsize);
566 NFT_PIPAPO_AVX2_BUCKET_LOAD4(3, lt, 2, pg[2], bsize);
567 NFT_PIPAPO_AVX2_BUCKET_LOAD4(4, lt, 3, pg[3], bsize);
569 NFT_PIPAPO_AVX2_NOMATCH_GOTO(0, nothing);
570 NFT_PIPAPO_AVX2_AND(1, 1, 0);
573 NFT_PIPAPO_AVX2_AND(5, 2, 3);
574 NFT_PIPAPO_AVX2_BUCKET_LOAD4(6, lt, 4, pg[4], bsize);
575 NFT_PIPAPO_AVX2_BUCKET_LOAD4(7, lt, 5, pg[5], bsize);
576 NFT_PIPAPO_AVX2_AND(8, 1, 4);
577 NFT_PIPAPO_AVX2_BUCKET_LOAD4(9, lt, 6, pg[6], bsize);
578 NFT_PIPAPO_AVX2_AND(10, 5, 6);
579 NFT_PIPAPO_AVX2_BUCKET_LOAD4(11, lt, 7, pg[7], bsize);
580 NFT_PIPAPO_AVX2_AND(12, 7, 8);
581 NFT_PIPAPO_AVX2_BUCKET_LOAD4(13, lt, 8, pg[8], bsize);
582 NFT_PIPAPO_AVX2_AND(14, 9, 10);
584 NFT_PIPAPO_AVX2_BUCKET_LOAD4(0, lt, 9, pg[9], bsize);
585 NFT_PIPAPO_AVX2_AND(1, 11, 12);
586 NFT_PIPAPO_AVX2_BUCKET_LOAD4(2, lt, 10, pg[10], bsize);
587 NFT_PIPAPO_AVX2_BUCKET_LOAD4(3, lt, 11, pg[11], bsize);
588 NFT_PIPAPO_AVX2_AND(4, 13, 14);
589 NFT_PIPAPO_AVX2_BUCKET_LOAD4(5, lt, 12, pg[12], bsize);
590 NFT_PIPAPO_AVX2_BUCKET_LOAD4(6, lt, 13, pg[13], bsize);
591 NFT_PIPAPO_AVX2_AND(7, 0, 1);
592 NFT_PIPAPO_AVX2_BUCKET_LOAD4(8, lt, 14, pg[14], bsize);
593 NFT_PIPAPO_AVX2_AND(9, 2, 3);
594 NFT_PIPAPO_AVX2_BUCKET_LOAD4(10, lt, 15, pg[15], bsize);
595 NFT_PIPAPO_AVX2_AND(11, 4, 5);
596 NFT_PIPAPO_AVX2_BUCKET_LOAD4(12, lt, 16, pg[16], bsize);
597 NFT_PIPAPO_AVX2_AND(13, 6, 7);
598 NFT_PIPAPO_AVX2_BUCKET_LOAD4(14, lt, 17, pg[17], bsize);
600 NFT_PIPAPO_AVX2_AND(0, 8, 9);
601 NFT_PIPAPO_AVX2_BUCKET_LOAD4(1, lt, 18, pg[18], bsize);
602 NFT_PIPAPO_AVX2_AND(2, 10, 11);
603 NFT_PIPAPO_AVX2_BUCKET_LOAD4(3, lt, 19, pg[19], bsize);
604 NFT_PIPAPO_AVX2_AND(4, 12, 13);
605 NFT_PIPAPO_AVX2_BUCKET_LOAD4(5, lt, 20, pg[20], bsize);
606 NFT_PIPAPO_AVX2_AND(6, 14, 0);
607 NFT_PIPAPO_AVX2_AND(7, 1, 2);
608 NFT_PIPAPO_AVX2_BUCKET_LOAD4(8, lt, 21, pg[21], bsize);
609 NFT_PIPAPO_AVX2_AND(9, 3, 4);
610 NFT_PIPAPO_AVX2_BUCKET_LOAD4(10, lt, 22, pg[22], bsize);
611 NFT_PIPAPO_AVX2_AND(11, 5, 6);
612 NFT_PIPAPO_AVX2_BUCKET_LOAD4(12, lt, 23, pg[23], bsize);
613 NFT_PIPAPO_AVX2_AND(13, 7, 8);
615 NFT_PIPAPO_AVX2_BUCKET_LOAD4(14, lt, 24, pg[24], bsize);
616 NFT_PIPAPO_AVX2_BUCKET_LOAD4(0, lt, 25, pg[25], bsize);
617 NFT_PIPAPO_AVX2_AND(1, 9, 10);
618 NFT_PIPAPO_AVX2_AND(2, 11, 12);
619 NFT_PIPAPO_AVX2_BUCKET_LOAD4(3, lt, 26, pg[26], bsize);
620 NFT_PIPAPO_AVX2_AND(4, 13, 14);
621 NFT_PIPAPO_AVX2_BUCKET_LOAD4(5, lt, 27, pg[27], bsize);
622 NFT_PIPAPO_AVX2_AND(6, 0, 1);
623 NFT_PIPAPO_AVX2_BUCKET_LOAD4(7, lt, 28, pg[28], bsize);
624 NFT_PIPAPO_AVX2_BUCKET_LOAD4(8, lt, 29, pg[29], bsize);
625 NFT_PIPAPO_AVX2_AND(9, 2, 3);
626 NFT_PIPAPO_AVX2_BUCKET_LOAD4(10, lt, 30, pg[30], bsize);
627 NFT_PIPAPO_AVX2_AND(11, 4, 5);
628 NFT_PIPAPO_AVX2_BUCKET_LOAD4(12, lt, 31, pg[31], bsize);
630 NFT_PIPAPO_AVX2_AND(0, 6, 7);
631 NFT_PIPAPO_AVX2_AND(1, 8, 9);
632 NFT_PIPAPO_AVX2_AND(2, 10, 11);
633 NFT_PIPAPO_AVX2_AND(3, 12, 0);
636 NFT_PIPAPO_AVX2_AND(4, 1, 2);
637 NFT_PIPAPO_AVX2_AND(5, 3, 4);
639 NFT_PIPAPO_AVX2_NOMATCH_GOTO(5, nomatch);
640 NFT_PIPAPO_AVX2_STORE(map[i_ul], 5);
642 b = nft_pipapo_avx2_refill(i_ul, &map[i_ul], fill, f->mt, last);
646 if (unlikely(ret == -1))
647 ret = b / XSAVE_YMM_SIZE;
651 NFT_PIPAPO_AVX2_STORE(map[i_ul], 15);
660 * nft_pipapo_avx2_lookup_8b_1() - AVX2-based lookup for one eight-bit group
661 * @map: Previous match result, used as initial bitmap
662 * @fill: Destination bitmap to be filled with current match result
663 * @f: Field, containing lookup and mapping tables
664 * @offset: Ignore buckets before the given index, no bits are filled there
665 * @pkt: Packet data, pointer to input nftables register
666 * @first: If this is the first field, don't source previous result
667 * @last: Last field: stop at the first match and return bit index
669 * See nft_pipapo_avx2_lookup_4b_2().
671 * This is used for 8-bit fields (i.e. protocol numbers).
673 * Return: -1 on no match, rule index of match if @last, otherwise first long
674 * word index to be checked next (i.e. first filled word).
676 static int nft_pipapo_avx2_lookup_8b_1(unsigned long *map, unsigned long *fill,
677 const struct nft_pipapo_field *f,
678 int offset, const u8 *pkt,
679 bool first, bool last)
681 int i, ret = -1, m256_size = f->bsize / NFT_PIPAPO_LONGS_PER_M256, b;
682 unsigned long *lt = f->lt, bsize = f->bsize;
684 lt += offset * NFT_PIPAPO_LONGS_PER_M256;
685 for (i = offset; i < m256_size; i++, lt += NFT_PIPAPO_LONGS_PER_M256) {
686 int i_ul = i * NFT_PIPAPO_LONGS_PER_M256;
689 NFT_PIPAPO_AVX2_BUCKET_LOAD8(2, lt, 0, pkt[0], bsize);
691 NFT_PIPAPO_AVX2_BUCKET_LOAD8(0, lt, 0, pkt[0], bsize);
692 NFT_PIPAPO_AVX2_LOAD(1, map[i_ul]);
693 NFT_PIPAPO_AVX2_AND(2, 0, 1);
694 NFT_PIPAPO_AVX2_NOMATCH_GOTO(1, nothing);
697 NFT_PIPAPO_AVX2_NOMATCH_GOTO(2, nomatch);
698 NFT_PIPAPO_AVX2_STORE(map[i_ul], 2);
700 b = nft_pipapo_avx2_refill(i_ul, &map[i_ul], fill, f->mt, last);
704 if (unlikely(ret == -1))
705 ret = b / XSAVE_YMM_SIZE;
709 NFT_PIPAPO_AVX2_STORE(map[i_ul], 15);
718 * nft_pipapo_avx2_lookup_8b_2() - AVX2-based lookup for 2 eight-bit groups
719 * @map: Previous match result, used as initial bitmap
720 * @fill: Destination bitmap to be filled with current match result
721 * @f: Field, containing lookup and mapping tables
722 * @offset: Ignore buckets before the given index, no bits are filled there
723 * @pkt: Packet data, pointer to input nftables register
724 * @first: If this is the first field, don't source previous result
725 * @last: Last field: stop at the first match and return bit index
727 * See nft_pipapo_avx2_lookup_4b_2().
729 * This is used for 16-bit fields (i.e. ports).
731 * Return: -1 on no match, rule index of match if @last, otherwise first long
732 * word index to be checked next (i.e. first filled word).
734 static int nft_pipapo_avx2_lookup_8b_2(unsigned long *map, unsigned long *fill,
735 const struct nft_pipapo_field *f,
736 int offset, const u8 *pkt,
737 bool first, bool last)
739 int i, ret = -1, m256_size = f->bsize / NFT_PIPAPO_LONGS_PER_M256, b;
740 unsigned long *lt = f->lt, bsize = f->bsize;
742 lt += offset * NFT_PIPAPO_LONGS_PER_M256;
743 for (i = offset; i < m256_size; i++, lt += NFT_PIPAPO_LONGS_PER_M256) {
744 int i_ul = i * NFT_PIPAPO_LONGS_PER_M256;
747 NFT_PIPAPO_AVX2_BUCKET_LOAD8(0, lt, 0, pkt[0], bsize);
748 NFT_PIPAPO_AVX2_BUCKET_LOAD8(1, lt, 1, pkt[1], bsize);
749 NFT_PIPAPO_AVX2_AND(4, 0, 1);
751 NFT_PIPAPO_AVX2_LOAD(0, map[i_ul]);
752 NFT_PIPAPO_AVX2_BUCKET_LOAD8(1, lt, 0, pkt[0], bsize);
753 NFT_PIPAPO_AVX2_BUCKET_LOAD8(2, lt, 1, pkt[1], bsize);
756 NFT_PIPAPO_AVX2_AND(3, 0, 1);
757 NFT_PIPAPO_AVX2_NOMATCH_GOTO(0, nothing);
758 NFT_PIPAPO_AVX2_AND(4, 3, 2);
762 NFT_PIPAPO_AVX2_NOMATCH_GOTO(4, nomatch);
763 NFT_PIPAPO_AVX2_STORE(map[i_ul], 4);
765 b = nft_pipapo_avx2_refill(i_ul, &map[i_ul], fill, f->mt, last);
769 if (unlikely(ret == -1))
770 ret = b / XSAVE_YMM_SIZE;
774 NFT_PIPAPO_AVX2_STORE(map[i_ul], 15);
783 * nft_pipapo_avx2_lookup_8b_4() - AVX2-based lookup for 4 eight-bit groups
784 * @map: Previous match result, used as initial bitmap
785 * @fill: Destination bitmap to be filled with current match result
786 * @f: Field, containing lookup and mapping tables
787 * @offset: Ignore buckets before the given index, no bits are filled there
788 * @pkt: Packet data, pointer to input nftables register
789 * @first: If this is the first field, don't source previous result
790 * @last: Last field: stop at the first match and return bit index
792 * See nft_pipapo_avx2_lookup_4b_2().
794 * This is used for 32-bit fields (i.e. IPv4 addresses).
796 * Return: -1 on no match, rule index of match if @last, otherwise first long
797 * word index to be checked next (i.e. first filled word).
799 static int nft_pipapo_avx2_lookup_8b_4(unsigned long *map, unsigned long *fill,
800 const struct nft_pipapo_field *f,
801 int offset, const u8 *pkt,
802 bool first, bool last)
804 int i, ret = -1, m256_size = f->bsize / NFT_PIPAPO_LONGS_PER_M256, b;
805 unsigned long *lt = f->lt, bsize = f->bsize;
807 lt += offset * NFT_PIPAPO_LONGS_PER_M256;
808 for (i = offset; i < m256_size; i++, lt += NFT_PIPAPO_LONGS_PER_M256) {
809 int i_ul = i * NFT_PIPAPO_LONGS_PER_M256;
812 NFT_PIPAPO_AVX2_BUCKET_LOAD8(0, lt, 0, pkt[0], bsize);
813 NFT_PIPAPO_AVX2_BUCKET_LOAD8(1, lt, 1, pkt[1], bsize);
814 NFT_PIPAPO_AVX2_BUCKET_LOAD8(2, lt, 2, pkt[2], bsize);
815 NFT_PIPAPO_AVX2_BUCKET_LOAD8(3, lt, 3, pkt[3], bsize);
818 NFT_PIPAPO_AVX2_AND(4, 0, 1);
819 NFT_PIPAPO_AVX2_AND(5, 2, 3);
820 NFT_PIPAPO_AVX2_AND(0, 4, 5);
822 NFT_PIPAPO_AVX2_BUCKET_LOAD8(0, lt, 0, pkt[0], bsize);
823 NFT_PIPAPO_AVX2_LOAD(1, map[i_ul]);
824 NFT_PIPAPO_AVX2_BUCKET_LOAD8(2, lt, 1, pkt[1], bsize);
825 NFT_PIPAPO_AVX2_BUCKET_LOAD8(3, lt, 2, pkt[2], bsize);
826 NFT_PIPAPO_AVX2_BUCKET_LOAD8(4, lt, 3, pkt[3], bsize);
828 NFT_PIPAPO_AVX2_AND(5, 0, 1);
829 NFT_PIPAPO_AVX2_NOMATCH_GOTO(1, nothing);
830 NFT_PIPAPO_AVX2_AND(6, 2, 3);
833 NFT_PIPAPO_AVX2_AND(7, 4, 5);
834 NFT_PIPAPO_AVX2_AND(0, 6, 7);
837 NFT_PIPAPO_AVX2_NOMATCH_GOTO(0, nomatch);
838 NFT_PIPAPO_AVX2_STORE(map[i_ul], 0);
840 b = nft_pipapo_avx2_refill(i_ul, &map[i_ul], fill, f->mt, last);
844 if (unlikely(ret == -1))
845 ret = b / XSAVE_YMM_SIZE;
850 NFT_PIPAPO_AVX2_STORE(map[i_ul], 15);
859 * nft_pipapo_avx2_lookup_8b_6() - AVX2-based lookup for 6 eight-bit groups
860 * @map: Previous match result, used as initial bitmap
861 * @fill: Destination bitmap to be filled with current match result
862 * @f: Field, containing lookup and mapping tables
863 * @offset: Ignore buckets before the given index, no bits are filled there
864 * @pkt: Packet data, pointer to input nftables register
865 * @first: If this is the first field, don't source previous result
866 * @last: Last field: stop at the first match and return bit index
868 * See nft_pipapo_avx2_lookup_4b_2().
870 * This is used for 48-bit fields (i.e. MAC addresses/EUI-48).
872 * Return: -1 on no match, rule index of match if @last, otherwise first long
873 * word index to be checked next (i.e. first filled word).
875 static int nft_pipapo_avx2_lookup_8b_6(unsigned long *map, unsigned long *fill,
876 const struct nft_pipapo_field *f,
877 int offset, const u8 *pkt,
878 bool first, bool last)
880 int i, ret = -1, m256_size = f->bsize / NFT_PIPAPO_LONGS_PER_M256, b;
881 unsigned long *lt = f->lt, bsize = f->bsize;
883 lt += offset * NFT_PIPAPO_LONGS_PER_M256;
884 for (i = offset; i < m256_size; i++, lt += NFT_PIPAPO_LONGS_PER_M256) {
885 int i_ul = i * NFT_PIPAPO_LONGS_PER_M256;
888 NFT_PIPAPO_AVX2_BUCKET_LOAD8(0, lt, 0, pkt[0], bsize);
889 NFT_PIPAPO_AVX2_BUCKET_LOAD8(1, lt, 1, pkt[1], bsize);
890 NFT_PIPAPO_AVX2_BUCKET_LOAD8(2, lt, 2, pkt[2], bsize);
891 NFT_PIPAPO_AVX2_BUCKET_LOAD8(3, lt, 3, pkt[3], bsize);
892 NFT_PIPAPO_AVX2_BUCKET_LOAD8(4, lt, 4, pkt[4], bsize);
894 NFT_PIPAPO_AVX2_AND(5, 0, 1);
895 NFT_PIPAPO_AVX2_BUCKET_LOAD8(6, lt, 5, pkt[5], bsize);
896 NFT_PIPAPO_AVX2_AND(7, 2, 3);
899 NFT_PIPAPO_AVX2_AND(0, 4, 5);
900 NFT_PIPAPO_AVX2_AND(1, 6, 7);
901 NFT_PIPAPO_AVX2_AND(4, 0, 1);
903 NFT_PIPAPO_AVX2_BUCKET_LOAD8(0, lt, 0, pkt[0], bsize);
904 NFT_PIPAPO_AVX2_LOAD(1, map[i_ul]);
905 NFT_PIPAPO_AVX2_BUCKET_LOAD8(2, lt, 1, pkt[1], bsize);
906 NFT_PIPAPO_AVX2_BUCKET_LOAD8(3, lt, 2, pkt[2], bsize);
907 NFT_PIPAPO_AVX2_BUCKET_LOAD8(4, lt, 3, pkt[3], bsize);
909 NFT_PIPAPO_AVX2_AND(5, 0, 1);
910 NFT_PIPAPO_AVX2_NOMATCH_GOTO(1, nothing);
912 NFT_PIPAPO_AVX2_AND(6, 2, 3);
913 NFT_PIPAPO_AVX2_BUCKET_LOAD8(7, lt, 4, pkt[4], bsize);
914 NFT_PIPAPO_AVX2_AND(0, 4, 5);
915 NFT_PIPAPO_AVX2_BUCKET_LOAD8(1, lt, 5, pkt[5], bsize);
916 NFT_PIPAPO_AVX2_AND(2, 6, 7);
919 NFT_PIPAPO_AVX2_AND(3, 0, 1);
920 NFT_PIPAPO_AVX2_AND(4, 2, 3);
923 NFT_PIPAPO_AVX2_NOMATCH_GOTO(4, nomatch);
924 NFT_PIPAPO_AVX2_STORE(map[i_ul], 4);
926 b = nft_pipapo_avx2_refill(i_ul, &map[i_ul], fill, f->mt, last);
930 if (unlikely(ret == -1))
931 ret = b / XSAVE_YMM_SIZE;
936 NFT_PIPAPO_AVX2_STORE(map[i_ul], 15);
945 * nft_pipapo_avx2_lookup_8b_16() - AVX2-based lookup for 16 eight-bit groups
946 * @map: Previous match result, used as initial bitmap
947 * @fill: Destination bitmap to be filled with current match result
948 * @f: Field, containing lookup and mapping tables
949 * @offset: Ignore buckets before the given index, no bits are filled there
950 * @pkt: Packet data, pointer to input nftables register
951 * @first: If this is the first field, don't source previous result
952 * @last: Last field: stop at the first match and return bit index
954 * See nft_pipapo_avx2_lookup_4b_2().
956 * This is used for 128-bit fields (i.e. IPv6 addresses).
958 * Return: -1 on no match, rule index of match if @last, otherwise first long
959 * word index to be checked next (i.e. first filled word).
961 static int nft_pipapo_avx2_lookup_8b_16(unsigned long *map, unsigned long *fill,
962 const struct nft_pipapo_field *f,
963 int offset, const u8 *pkt,
964 bool first, bool last)
966 int i, ret = -1, m256_size = f->bsize / NFT_PIPAPO_LONGS_PER_M256, b;
967 unsigned long *lt = f->lt, bsize = f->bsize;
969 lt += offset * NFT_PIPAPO_LONGS_PER_M256;
970 for (i = offset; i < m256_size; i++, lt += NFT_PIPAPO_LONGS_PER_M256) {
971 int i_ul = i * NFT_PIPAPO_LONGS_PER_M256;
974 NFT_PIPAPO_AVX2_LOAD(0, map[i_ul]);
976 NFT_PIPAPO_AVX2_BUCKET_LOAD8(1, lt, 0, pkt[0], bsize);
977 NFT_PIPAPO_AVX2_BUCKET_LOAD8(2, lt, 1, pkt[1], bsize);
978 NFT_PIPAPO_AVX2_BUCKET_LOAD8(3, lt, 2, pkt[2], bsize);
980 NFT_PIPAPO_AVX2_NOMATCH_GOTO(0, nothing);
981 NFT_PIPAPO_AVX2_AND(1, 1, 0);
983 NFT_PIPAPO_AVX2_BUCKET_LOAD8(4, lt, 3, pkt[3], bsize);
985 NFT_PIPAPO_AVX2_BUCKET_LOAD8(5, lt, 4, pkt[4], bsize);
986 NFT_PIPAPO_AVX2_AND(6, 1, 2);
987 NFT_PIPAPO_AVX2_BUCKET_LOAD8(7, lt, 5, pkt[5], bsize);
988 NFT_PIPAPO_AVX2_AND(0, 3, 4);
989 NFT_PIPAPO_AVX2_BUCKET_LOAD8(1, lt, 6, pkt[6], bsize);
991 NFT_PIPAPO_AVX2_BUCKET_LOAD8(2, lt, 7, pkt[7], bsize);
992 NFT_PIPAPO_AVX2_AND(3, 5, 6);
993 NFT_PIPAPO_AVX2_AND(4, 0, 1);
994 NFT_PIPAPO_AVX2_BUCKET_LOAD8(5, lt, 8, pkt[8], bsize);
996 NFT_PIPAPO_AVX2_AND(6, 2, 3);
997 NFT_PIPAPO_AVX2_BUCKET_LOAD8(7, lt, 9, pkt[9], bsize);
998 NFT_PIPAPO_AVX2_AND(0, 4, 5);
999 NFT_PIPAPO_AVX2_BUCKET_LOAD8(1, lt, 10, pkt[10], bsize);
1000 NFT_PIPAPO_AVX2_AND(2, 6, 7);
1001 NFT_PIPAPO_AVX2_BUCKET_LOAD8(3, lt, 11, pkt[11], bsize);
1002 NFT_PIPAPO_AVX2_AND(4, 0, 1);
1003 NFT_PIPAPO_AVX2_BUCKET_LOAD8(5, lt, 12, pkt[12], bsize);
1004 NFT_PIPAPO_AVX2_AND(6, 2, 3);
1005 NFT_PIPAPO_AVX2_BUCKET_LOAD8(7, lt, 13, pkt[13], bsize);
1006 NFT_PIPAPO_AVX2_AND(0, 4, 5);
1007 NFT_PIPAPO_AVX2_BUCKET_LOAD8(1, lt, 14, pkt[14], bsize);
1008 NFT_PIPAPO_AVX2_AND(2, 6, 7);
1009 NFT_PIPAPO_AVX2_BUCKET_LOAD8(3, lt, 15, pkt[15], bsize);
1010 NFT_PIPAPO_AVX2_AND(4, 0, 1);
1013 NFT_PIPAPO_AVX2_AND(5, 2, 3);
1014 NFT_PIPAPO_AVX2_AND(6, 4, 5);
1016 NFT_PIPAPO_AVX2_NOMATCH_GOTO(6, nomatch);
1017 NFT_PIPAPO_AVX2_STORE(map[i_ul], 6);
1019 b = nft_pipapo_avx2_refill(i_ul, &map[i_ul], fill, f->mt, last);
1023 if (unlikely(ret == -1))
1024 ret = b / XSAVE_YMM_SIZE;
1029 NFT_PIPAPO_AVX2_STORE(map[i_ul], 15);
1038 * nft_pipapo_avx2_lookup_slow() - Fallback function for uncommon field sizes
1039 * @map: Previous match result, used as initial bitmap
1040 * @fill: Destination bitmap to be filled with current match result
1041 * @f: Field, containing lookup and mapping tables
1042 * @offset: Ignore buckets before the given index, no bits are filled there
1043 * @pkt: Packet data, pointer to input nftables register
1044 * @first: If this is the first field, don't source previous result
1045 * @last: Last field: stop at the first match and return bit index
1047 * This function should never be called, but is provided for the case the field
1048 * size doesn't match any of the known data types. Matching rate is
1049 * substantially lower than AVX2 routines.
1051 * Return: -1 on no match, rule index of match if @last, otherwise first long
1052 * word index to be checked next (i.e. first filled word).
1054 static int nft_pipapo_avx2_lookup_slow(unsigned long *map, unsigned long *fill,
1055 const struct nft_pipapo_field *f,
1056 int offset, const u8 *pkt,
1057 bool first, bool last)
1059 unsigned long bsize = f->bsize;
1063 memset(map, 0xff, bsize * sizeof(*map));
1065 for (i = offset; i < bsize; i++) {
1067 pipapo_and_field_buckets_8bit(f, map, pkt);
1069 pipapo_and_field_buckets_4bit(f, map, pkt);
1070 NFT_PIPAPO_GROUP_BITS_ARE_8_OR_4;
1072 b = pipapo_refill(map, bsize, f->rules, fill, f->mt, last);
1078 ret = b / XSAVE_YMM_SIZE;
1085 * nft_pipapo_avx2_estimate() - Set size, space and lookup complexity
1086 * @desc: Set description, element count and field description used
1087 * @features: Flags: NFT_SET_INTERVAL needs to be there
1088 * @est: Storage for estimation data
1090 * Return: true if set is compatible and AVX2 available, false otherwise.
1092 bool nft_pipapo_avx2_estimate(const struct nft_set_desc *desc, u32 features,
1093 struct nft_set_estimate *est)
1095 if (!(features & NFT_SET_INTERVAL) ||
1096 desc->field_count < NFT_PIPAPO_MIN_FIELDS)
1099 if (!boot_cpu_has(X86_FEATURE_AVX2) || !boot_cpu_has(X86_FEATURE_AVX))
1102 est->size = pipapo_estimate_size(desc);
1106 est->lookup = NFT_SET_CLASS_O_LOG_N;
1108 est->space = NFT_SET_CLASS_O_N;
1114 * nft_pipapo_avx2_lookup() - Lookup function for AVX2 implementation
1115 * @net: Network namespace
1116 * @set: nftables API set representation
1117 * @key: nftables API element representation containing key data
1118 * @ext: nftables API extension pointer, filled with matching reference
1120 * For more details, see DOC: Theory of Operation in nft_set_pipapo.c.
1122 * This implementation exploits the repetitive characteristic of the algorithm
1123 * to provide a fast, vectorised version using the AVX2 SIMD instruction set.
1125 * Return: true on match, false otherwise.
1127 bool nft_pipapo_avx2_lookup(const struct net *net, const struct nft_set *set,
1128 const u32 *key, const struct nft_set_ext **ext)
1130 struct nft_pipapo *priv = nft_set_priv(set);
1131 struct nft_pipapo_scratch *scratch;
1132 u8 genmask = nft_genmask_cur(net);
1133 const struct nft_pipapo_match *m;
1134 const struct nft_pipapo_field *f;
1135 const u8 *rp = (const u8 *)key;
1136 unsigned long *res, *fill;
1140 if (unlikely(!irq_fpu_usable()))
1141 return nft_pipapo_lookup(net, set, key, ext);
1143 m = rcu_dereference(priv->match);
1145 /* This also protects access to all data related to scratch maps.
1147 * Note that we don't need a valid MXCSR state for any of the
1148 * operations we use here, so pass 0 as mask and spare a LDMXCSR
1151 kernel_fpu_begin_mask(0);
1153 scratch = *raw_cpu_ptr(m->scratch);
1154 if (unlikely(!scratch)) {
1159 map_index = scratch->map_index;
1161 res = scratch->map + (map_index ? m->bsize_max : 0);
1162 fill = scratch->map + (map_index ? 0 : m->bsize_max);
1164 /* Starting map doesn't need to be set for this implementation */
1166 nft_pipapo_avx2_prepare();
1169 nft_pipapo_for_each_field(f, i, m) {
1170 bool last = i == m->field_count - 1, first = !i;
1172 #define NFT_SET_PIPAPO_AVX2_LOOKUP(b, n) \
1173 (ret = nft_pipapo_avx2_lookup_##b##b_##n(res, fill, f, \
1177 if (likely(f->bb == 8)) {
1178 if (f->groups == 1) {
1179 NFT_SET_PIPAPO_AVX2_LOOKUP(8, 1);
1180 } else if (f->groups == 2) {
1181 NFT_SET_PIPAPO_AVX2_LOOKUP(8, 2);
1182 } else if (f->groups == 4) {
1183 NFT_SET_PIPAPO_AVX2_LOOKUP(8, 4);
1184 } else if (f->groups == 6) {
1185 NFT_SET_PIPAPO_AVX2_LOOKUP(8, 6);
1186 } else if (f->groups == 16) {
1187 NFT_SET_PIPAPO_AVX2_LOOKUP(8, 16);
1189 ret = nft_pipapo_avx2_lookup_slow(res, fill, f,
1194 if (f->groups == 2) {
1195 NFT_SET_PIPAPO_AVX2_LOOKUP(4, 2);
1196 } else if (f->groups == 4) {
1197 NFT_SET_PIPAPO_AVX2_LOOKUP(4, 4);
1198 } else if (f->groups == 8) {
1199 NFT_SET_PIPAPO_AVX2_LOOKUP(4, 8);
1200 } else if (f->groups == 12) {
1201 NFT_SET_PIPAPO_AVX2_LOOKUP(4, 12);
1202 } else if (f->groups == 32) {
1203 NFT_SET_PIPAPO_AVX2_LOOKUP(4, 32);
1205 ret = nft_pipapo_avx2_lookup_slow(res, fill, f,
1210 NFT_PIPAPO_GROUP_BITS_ARE_8_OR_4;
1212 #undef NFT_SET_PIPAPO_AVX2_LOOKUP
1218 *ext = &f->mt[ret].e->ext;
1219 if (unlikely(nft_set_elem_expired(*ext) ||
1220 !nft_set_elem_active(*ext, genmask))) {
1229 rp += NFT_PIPAPO_GROUPS_PADDED_SIZE(f);
1234 scratch->map_index = !map_index;