xref: /xnu-11215.81.4/bsd/netinet/cpu_in_cksum_gen.c (revision d4514f0bc1d3f944c22d92e68b646ac3fb40d452)
1 /*
2  * Copyright (c) 2012-2021 Apple Inc. All rights reserved.
3  *
4  * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
5  *
6  * This file contains Original Code and/or Modifications of Original Code
7  * as defined in and that are subject to the Apple Public Source License
8  * Version 2.0 (the 'License'). You may not use this file except in
9  * compliance with the License. The rights granted to you under the License
10  * may not be used to create, or enable the creation or redistribution of,
11  * unlawful or unlicensed copies of an Apple operating system, or to
12  * circumvent, violate, or enable the circumvention or violation of, any
13  * terms of an Apple operating system software license agreement.
14  *
15  * Please obtain a copy of the License at
16  * http://www.opensource.apple.com/apsl/ and read it before using this file.
17  *
18  * The Original Code and all software distributed under the License are
19  * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
20  * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
21  * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
22  * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
23  * Please see the License for the specific language governing rights and
24  * limitations under the License.
25  *
26  * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
27  */
28 
29 /*-
30  * Copyright (c) 2008 Joerg Sonnenberger <[email protected]>.
31  * All rights reserved.
32  *
33  * Redistribution and use in source and binary forms, with or without
34  * modification, are permitted provided that the following conditions
35  * are met:
36  *
37  * 1. Redistributions of source code must retain the above copyright
38  *    notice, this list of conditions and the following disclaimer.
39  * 2. Redistributions in binary form must reproduce the above copyright
40  *    notice, this list of conditions and the following disclaimer in
41  *    the documentation and/or other materials provided with the
42  *    distribution.
43  *
44  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
45  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
46  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
47  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
48  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
49  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
50  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
51  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
52  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
53  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
54  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
55  * SUCH DAMAGE.
56  */
57 
58 #ifdef KERNEL
59 #include <sys/param.h>
60 #include <machine/endian.h>
61 #include <sys/mcache.h>
62 #include <sys/mbuf.h>
63 #include <kern/debug.h>
64 #include <libkern/libkern.h>
65 #include <mach/boolean.h>
66 #include <pexpert/pexpert.h>
67 #define CKSUM_ERR(fmt, args...) kprintf(fmt, ## args)
68 #else /* !KERNEL */
69 #ifndef LIBSYSCALL_INTERFACE
70 #error "LIBSYSCALL_INTERFACE not defined"
71 #endif /* !LIBSYSCALL_INTERFACE */
72 #include <stdlib.h>
73 #include <stddef.h>
74 #include <stdint.h>
75 #include <unistd.h>
76 #include <strings.h>
77 #include <mach/boolean.h>
78 #include <skywalk/os_skywalk_private.h>
79 #define CKSUM_ERR(fmt, args...) fprintf_stderr(fmt, ## args)
80 #endif /* !KERNEL */
81 
82 /* compile time assert */
83 #ifndef _CASSERT
84 #define _CASSERT(x)     _Static_assert(x, "compile-time assertion failed")
85 #endif /* !_CASSERT */
86 
87 #ifndef VERIFY
88 #define VERIFY(EX) ((void)0)
89 #endif /* !VERIFY */
90 
91 #ifndef CKSUM_ERR
92 #define CKSUM_ERR(fmt, args...) ((void)0)
93 #endif /* !CKSUM_ERR */
94 
95 #define PREDICT_TRUE(x)         __builtin_expect(!!((long)(x)), 1L)
96 #define PREDICT_FALSE(x)        __builtin_expect(!!((long)(x)), 0L)
97 
98 /* fake mbuf struct used only for calling os_cpu_in_cksum_mbuf() */
99 struct _mbuf {
100 	struct _mbuf    *_m_next;
101 	void            *_m_pad;
102 	uint8_t         *__sized_by(_m_len) _m_data;
103 	int32_t         _m_len;
104 };
105 
106 extern uint32_t os_cpu_in_cksum(const void *__sized_by(len), uint32_t len, uint32_t);
107 extern uint32_t os_cpu_in_cksum_mbuf(struct _mbuf *, int, int, uint32_t);
108 
109 uint32_t
os_cpu_in_cksum(const void * __sized_by (len)data,uint32_t len,uint32_t initial_sum)110 os_cpu_in_cksum(const void *__sized_by(len) data, uint32_t len, uint32_t initial_sum)
111 {
112 	/*
113 	 * If data is 4-bytes aligned (conditional), length is multiple
114 	 * of 4-bytes (required), and the amount to checksum is small,
115 	 * this would be quicker; this is suitable for IPv4/TCP header.
116 	 */
117 	if (
118 #if !defined(__arm64__) && !defined(__x86_64__)
119 		IS_P2ALIGNED(data, sizeof(uint32_t)) &&
120 #endif /* !__arm64__ && !__x86_64__ */
121 		len <= 64 && (len & 3) == 0) {
122 		uint32_t plen = len;
123 		uint8_t *__sized_by(plen) p = __DECONST(uint8_t *, data);
124 		uint64_t sum = initial_sum;
125 
126 		switch (len) {
127 		case 20:                /* simple IPv4 or TCP header */
128 			sum += *(uint32_t *)(void *)p;
129 			sum += *(uint32_t *)(void *)(p + 4);
130 			sum += *(uint32_t *)(void *)(p + 8);
131 			sum += *(uint32_t *)(void *)(p + 12);
132 			sum += *(uint32_t *)(void *)(p + 16);
133 			break;
134 
135 		case 32:                /* TCP header + timestamp option */
136 			sum += *(uint32_t *)(void *)p;
137 			sum += *(uint32_t *)(void *)(p + 4);
138 			sum += *(uint32_t *)(void *)(p + 8);
139 			sum += *(uint32_t *)(void *)(p + 12);
140 			sum += *(uint32_t *)(void *)(p + 16);
141 			sum += *(uint32_t *)(void *)(p + 20);
142 			sum += *(uint32_t *)(void *)(p + 24);
143 			sum += *(uint32_t *)(void *)(p + 28);
144 			break;
145 
146 		default:
147 			while (plen) {
148 				sum += *(uint32_t *)(void *)p;
149 				p += 4;
150 				plen -= 4;
151 			}
152 			break;
153 		}
154 
155 		/* fold 64-bit to 16-bit (deferred carries) */
156 		sum = (sum >> 32) + (sum & 0xffffffff); /* 33-bit */
157 		sum = (sum >> 16) + (sum & 0xffff);     /* 17-bit + carry */
158 		sum = (sum >> 16) + (sum & 0xffff);     /* 16-bit + carry */
159 		sum = (sum >> 16) + (sum & 0xffff);     /* final carry */
160 
161 		return sum & 0xffff;
162 	}
163 
164 	/*
165 	 * Otherwise, let os_cpu_in_cksum_mbuf() handle it; it only looks
166 	 * at 3 fields: {next,data,len}, and since it doesn't care about
167 	 * the authenticity of the mbuf, we use a fake one here.  Make
168 	 * sure the offsets are as expected.
169 	 */
170 #if defined(__LP64__)
171 	_CASSERT(offsetof(struct _mbuf, _m_next) == 0);
172 	_CASSERT(offsetof(struct _mbuf, _m_data) == 16);
173 	_CASSERT(offsetof(struct _mbuf, _m_len) == 24);
174 #else /* !__LP64__ */
175 	_CASSERT(offsetof(struct _mbuf, _m_next) == 0);
176 	_CASSERT(offsetof(struct _mbuf, _m_data) == 8);
177 	_CASSERT(offsetof(struct _mbuf, _m_len) == 12);
178 #endif /* !__LP64__ */
179 #ifdef KERNEL
180 	_CASSERT(offsetof(struct _mbuf, _m_next) ==
181 	    offsetof(struct mbuf, m_next));
182 	_CASSERT(offsetof(struct _mbuf, _m_data) ==
183 	    offsetof(struct mbuf, m_data));
184 	_CASSERT(offsetof(struct _mbuf, _m_len) ==
185 	    offsetof(struct mbuf, m_len));
186 #endif /* KERNEL */
187 	struct _mbuf m = {
188 		._m_next = NULL,
189 		._m_data = __DECONST(uint8_t *, data),
190 		._m_len = len,
191 	};
192 
193 	return os_cpu_in_cksum_mbuf(&m, len, 0, initial_sum);
194 }
195 
196 #if defined(__i386__) || defined(__x86_64__)
197 
198 /*
199  * Checksum routine for Internet Protocol family headers (Portable Version).
200  *
201  * This routine is very heavily used in the network
202  * code and should be modified for each CPU to be as fast as possible.
203  *
204  * A discussion of different implementation techniques can be found in
205  * RFC 1071.
206  *
207  * The default implementation for 32-bit architectures is using
208  * a 32-bit accumulator and operating on 16-bit operands.
209  *
210  * The default implementation for 64-bit architectures is using
211  * a 64-bit accumulator and operating on 32-bit operands.
212  *
213  * Both versions are unrolled to handle 32 Byte / 64 Byte fragments as core
214  * of the inner loop. After each iteration of the inner loop, a partial
215  * reduction is done to avoid carry in long packets.
216  */
217 
218 #if !defined(__LP64__)
219 /* 32-bit version */
220 uint32_t
os_cpu_in_cksum_mbuf(struct _mbuf * m,int len,int off,uint32_t initial_sum)221 os_cpu_in_cksum_mbuf(struct _mbuf *m, int len, int off, uint32_t initial_sum)
222 {
223 	int mlen;
224 	uint32_t sum, partial;
225 	unsigned int final_acc;
226 	uint8_t *data;
227 	boolean_t needs_swap, started_on_odd;
228 
229 	VERIFY(len >= 0);
230 	VERIFY(off >= 0);
231 
232 	needs_swap = FALSE;
233 	started_on_odd = FALSE;
234 	sum = (initial_sum >> 16) + (initial_sum & 0xffff);
235 
236 	for (;;) {
237 		if (PREDICT_FALSE(m == NULL)) {
238 			CKSUM_ERR("%s: out of data\n", __func__);
239 			return (uint32_t)-1;
240 		}
241 		mlen = m->_m_len;
242 		if (mlen > off) {
243 			mlen -= off;
244 			data = m->_m_data + off;
245 			goto post_initial_offset;
246 		}
247 		off -= mlen;
248 		if (len == 0) {
249 			break;
250 		}
251 		m = m->_m_next;
252 	}
253 
254 	for (; len > 0; m = m->_m_next) {
255 		if (PREDICT_FALSE(m == NULL)) {
256 			CKSUM_ERR("%s: out of data\n", __func__);
257 			return (uint32_t)-1;
258 		}
259 		mlen = m->_m_len;
260 		data = m->_m_data;
261 post_initial_offset:
262 		if (mlen == 0) {
263 			continue;
264 		}
265 		if (mlen > len) {
266 			mlen = len;
267 		}
268 		len -= mlen;
269 
270 		partial = 0;
271 		if ((uintptr_t)data & 1) {
272 			/* Align on word boundary */
273 			started_on_odd = !started_on_odd;
274 #if BYTE_ORDER == LITTLE_ENDIAN
275 			partial = *data << 8;
276 #else
277 			partial = *data;
278 #endif
279 			++data;
280 			--mlen;
281 		}
282 		needs_swap = started_on_odd;
283 		while (mlen >= 32) {
284 			__builtin_prefetch(data + 32);
285 			partial += *(uint16_t *)(void *)data;
286 			partial += *(uint16_t *)(void *)(data + 2);
287 			partial += *(uint16_t *)(void *)(data + 4);
288 			partial += *(uint16_t *)(void *)(data + 6);
289 			partial += *(uint16_t *)(void *)(data + 8);
290 			partial += *(uint16_t *)(void *)(data + 10);
291 			partial += *(uint16_t *)(void *)(data + 12);
292 			partial += *(uint16_t *)(void *)(data + 14);
293 			partial += *(uint16_t *)(void *)(data + 16);
294 			partial += *(uint16_t *)(void *)(data + 18);
295 			partial += *(uint16_t *)(void *)(data + 20);
296 			partial += *(uint16_t *)(void *)(data + 22);
297 			partial += *(uint16_t *)(void *)(data + 24);
298 			partial += *(uint16_t *)(void *)(data + 26);
299 			partial += *(uint16_t *)(void *)(data + 28);
300 			partial += *(uint16_t *)(void *)(data + 30);
301 			data += 32;
302 			mlen -= 32;
303 			if (PREDICT_FALSE(partial & 0xc0000000)) {
304 				if (needs_swap) {
305 					partial = (partial << 8) +
306 					    (partial >> 24);
307 				}
308 				sum += (partial >> 16);
309 				sum += (partial & 0xffff);
310 				partial = 0;
311 			}
312 		}
313 		if (mlen & 16) {
314 			partial += *(uint16_t *)(void *)data;
315 			partial += *(uint16_t *)(void *)(data + 2);
316 			partial += *(uint16_t *)(void *)(data + 4);
317 			partial += *(uint16_t *)(void *)(data + 6);
318 			partial += *(uint16_t *)(void *)(data + 8);
319 			partial += *(uint16_t *)(void *)(data + 10);
320 			partial += *(uint16_t *)(void *)(data + 12);
321 			partial += *(uint16_t *)(void *)(data + 14);
322 			data += 16;
323 			mlen -= 16;
324 		}
325 		/*
326 		 * mlen is not updated below as the remaining tests
327 		 * are using bit masks, which are not affected.
328 		 */
329 		if (mlen & 8) {
330 			partial += *(uint16_t *)(void *)data;
331 			partial += *(uint16_t *)(void *)(data + 2);
332 			partial += *(uint16_t *)(void *)(data + 4);
333 			partial += *(uint16_t *)(void *)(data + 6);
334 			data += 8;
335 		}
336 		if (mlen & 4) {
337 			partial += *(uint16_t *)(void *)data;
338 			partial += *(uint16_t *)(void *)(data + 2);
339 			data += 4;
340 		}
341 		if (mlen & 2) {
342 			partial += *(uint16_t *)(void *)data;
343 			data += 2;
344 		}
345 		if (mlen & 1) {
346 #if BYTE_ORDER == LITTLE_ENDIAN
347 			partial += *data;
348 #else
349 			partial += *data << 8;
350 #endif
351 			started_on_odd = !started_on_odd;
352 		}
353 
354 		if (needs_swap) {
355 			partial = (partial << 8) + (partial >> 24);
356 		}
357 		sum += (partial >> 16) + (partial & 0xffff);
358 		/*
359 		 * Reduce sum to allow potential byte swap
360 		 * in the next iteration without carry.
361 		 */
362 		sum = (sum >> 16) + (sum & 0xffff);
363 	}
364 	final_acc = ((sum >> 16) & 0xffff) + (sum & 0xffff);
365 	final_acc = (final_acc >> 16) + (final_acc & 0xffff);
366 	return final_acc & 0xffff;
367 }
368 
369 #else /* __LP64__ */
370 /* 64-bit version */
371 uint32_t
os_cpu_in_cksum_mbuf(struct _mbuf * m,int len,int off,uint32_t initial_sum)372 os_cpu_in_cksum_mbuf(struct _mbuf *m, int len, int off, uint32_t initial_sum)
373 {
374 	int mlen;
375 	uint64_t sum, partial;
376 	unsigned int final_acc;
377 	uint8_t *data;
378 	boolean_t needs_swap, started_on_odd;
379 
380 	VERIFY(len >= 0);
381 	VERIFY(off >= 0);
382 
383 	needs_swap = FALSE;
384 	started_on_odd = FALSE;
385 	sum = initial_sum;
386 
387 	for (;;) {
388 		if (PREDICT_FALSE(m == NULL)) {
389 			CKSUM_ERR("%s: out of data\n", __func__);
390 			return (uint32_t)-1;
391 		}
392 		mlen = m->_m_len;
393 		if (mlen > off) {
394 			mlen -= off;
395 			data = m->_m_data + off;
396 			goto post_initial_offset;
397 		}
398 		off -= mlen;
399 		if (len == 0) {
400 			break;
401 		}
402 		m = m->_m_next;
403 	}
404 
405 	for (; len > 0; m = m->_m_next) {
406 		if (PREDICT_FALSE(m == NULL)) {
407 			CKSUM_ERR("%s: out of data\n", __func__);
408 			return (uint32_t)-1;
409 		}
410 		mlen = m->_m_len;
411 		data = m->_m_data;
412 post_initial_offset:
413 		if (mlen == 0) {
414 			continue;
415 		}
416 		if (mlen > len) {
417 			mlen = len;
418 		}
419 		len -= mlen;
420 
421 		partial = 0;
422 		if ((uintptr_t)data & 1) {
423 			/* Align on word boundary */
424 			started_on_odd = !started_on_odd;
425 #if BYTE_ORDER == LITTLE_ENDIAN
426 			partial = *data << 8;
427 #else
428 			partial = *data;
429 #endif
430 			++data;
431 			--mlen;
432 		}
433 		needs_swap = started_on_odd;
434 		if ((uintptr_t)data & 2) {
435 			if (mlen < 2) {
436 				goto trailing_bytes;
437 			}
438 			partial += *(uint16_t *)(void *)data;
439 			data += 2;
440 			mlen -= 2;
441 		}
442 		while (mlen >= 64) {
443 			__builtin_prefetch(data + 32);
444 			__builtin_prefetch(data + 64);
445 			partial += *(uint32_t *)(void *)data;
446 			partial += *(uint32_t *)(void *)(data + 4);
447 			partial += *(uint32_t *)(void *)(data + 8);
448 			partial += *(uint32_t *)(void *)(data + 12);
449 			partial += *(uint32_t *)(void *)(data + 16);
450 			partial += *(uint32_t *)(void *)(data + 20);
451 			partial += *(uint32_t *)(void *)(data + 24);
452 			partial += *(uint32_t *)(void *)(data + 28);
453 			partial += *(uint32_t *)(void *)(data + 32);
454 			partial += *(uint32_t *)(void *)(data + 36);
455 			partial += *(uint32_t *)(void *)(data + 40);
456 			partial += *(uint32_t *)(void *)(data + 44);
457 			partial += *(uint32_t *)(void *)(data + 48);
458 			partial += *(uint32_t *)(void *)(data + 52);
459 			partial += *(uint32_t *)(void *)(data + 56);
460 			partial += *(uint32_t *)(void *)(data + 60);
461 			data += 64;
462 			mlen -= 64;
463 			if (PREDICT_FALSE(partial & (3ULL << 62))) {
464 				if (needs_swap) {
465 					partial = (partial << 8) +
466 					    (partial >> 56);
467 				}
468 				sum += (partial >> 32);
469 				sum += (partial & 0xffffffff);
470 				partial = 0;
471 			}
472 		}
473 		/*
474 		 * mlen is not updated below as the remaining tests
475 		 * are using bit masks, which are not affected.
476 		 */
477 		if (mlen & 32) {
478 			partial += *(uint32_t *)(void *)data;
479 			partial += *(uint32_t *)(void *)(data + 4);
480 			partial += *(uint32_t *)(void *)(data + 8);
481 			partial += *(uint32_t *)(void *)(data + 12);
482 			partial += *(uint32_t *)(void *)(data + 16);
483 			partial += *(uint32_t *)(void *)(data + 20);
484 			partial += *(uint32_t *)(void *)(data + 24);
485 			partial += *(uint32_t *)(void *)(data + 28);
486 			data += 32;
487 		}
488 		if (mlen & 16) {
489 			partial += *(uint32_t *)(void *)data;
490 			partial += *(uint32_t *)(void *)(data + 4);
491 			partial += *(uint32_t *)(void *)(data + 8);
492 			partial += *(uint32_t *)(void *)(data + 12);
493 			data += 16;
494 		}
495 		if (mlen & 8) {
496 			partial += *(uint32_t *)(void *)data;
497 			partial += *(uint32_t *)(void *)(data + 4);
498 			data += 8;
499 		}
500 		if (mlen & 4) {
501 			partial += *(uint32_t *)(void *)data;
502 			data += 4;
503 		}
504 		if (mlen & 2) {
505 			partial += *(uint16_t *)(void *)data;
506 			data += 2;
507 		}
508 trailing_bytes:
509 		if (mlen & 1) {
510 #if BYTE_ORDER == LITTLE_ENDIAN
511 			partial += *data;
512 #else
513 			partial += *data << 8;
514 #endif
515 			started_on_odd = !started_on_odd;
516 		}
517 
518 		if (needs_swap) {
519 			partial = (partial << 8) + (partial >> 56);
520 		}
521 		sum += (partial >> 32) + (partial & 0xffffffff);
522 		/*
523 		 * Reduce sum to allow potential byte swap
524 		 * in the next iteration without carry.
525 		 */
526 		sum = (sum >> 32) + (sum & 0xffffffff);
527 	}
528 	final_acc = (sum >> 48) + ((sum >> 32) & 0xffff) +
529 	    ((sum >> 16) & 0xffff) + (sum & 0xffff);
530 	final_acc = (final_acc >> 16) + (final_acc & 0xffff);
531 	final_acc = (final_acc >> 16) + (final_acc & 0xffff);
532 	return final_acc & 0xffff;
533 }
534 #endif /* __LP64 */
535 
536 #endif /* __i386__ || __x86_64__ */
537