2 * Implement fast CRC32C with PCLMULQDQ instructions. (x86_64)
4 * The white papers on CRC32C calculations with PCLMULQDQ instruction can be
6 * http://www.intel.com/content/dam/www/public/us/en/documents/white-papers/crc-iscsi-polynomial-crc32-instruction-paper.pdf
7 * http://www.intel.com/content/dam/www/public/us/en/documents/white-papers/fast-crc-computation-paper.pdf
9 * Copyright (C) 2012 Intel Corporation.
12 * Wajdi Feghali <wajdi.k.feghali@intel.com>
13 * James Guilford <james.guilford@intel.com>
14 * David Cote <david.m.cote@intel.com>
15 * Tim Chen <tim.c.chen@linux.intel.com>
17 * This software is available to you under a choice of one of two
18 * licenses. You may choose to be licensed under the terms of the GNU
19 * General Public License (GPL) Version 2, available from the file
20 * COPYING in the main directory of this source tree, or the
21 * OpenIB.org BSD license below:
23 * Redistribution and use in source and binary forms, with or
24 * without modification, are permitted provided that the following
27 * - Redistributions of source code must retain the above
28 * copyright notice, this list of conditions and the following
31 * - Redistributions in binary form must reproduce the above
32 * copyright notice, this list of conditions and the following
33 * disclaimer in the documentation and/or other materials
34 * provided with the distribution.
36 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
37 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
38 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
39 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
40 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
41 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
42 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
46 #include <linux/linkage.h>
47 #include <asm/nospec-branch.h>
49 ## ISCSI CRC 32 Implementation with crc32 and pclmulqdq Instruction
59 .macro JNC_LESS_THAN j
63 # Define threshold where buffers are considered "small" and routed to more
64 # efficient "by-1" code. This "by-1" code only handles up to 255 bytes, so
65 # SMALL_SIZE can be no larger than 255.
67 #define SMALL_SIZE 200
69 .if (SMALL_SIZE > 255)
70 .error "SMALL_ SIZE must be < 256"
73 # unsigned int crc_pcl(u8 *buffer, int len, unsigned int crc_init);
76 SYM_FUNC_START(crc_pcl)
89 #define crc_init_arg %rdx
92 #define crc_init_dw %r8d
100 ## Move crc_init for Linux to a different
101 mov crc_init_arg, crc_init
103 ################################################################
105 ################################################################
107 mov %bufp, bufptmp # rdi = *buf
109 and $7, %bufp # calculate the unalignment amount of
111 je proc_block # Skip if aligned
113 ## If len is less than 8 and we're unaligned, we need to jump
114 ## to special code to avoid reading beyond the end of the buffer
117 # less_than_8 expects length in upper 3 bits of len_dw
118 # less_than_8_post_shl1 expects length = carryflag * 8 + len_dw[31:30]
120 jmp less_than_8_post_shl1
123 #### Calculate CRC of unaligned bytes of the buffer (if any)
124 movq (bufptmp), tmp # load a quadward from the buffer
125 add %bufp, bufptmp # align buffer pointer for quadword
127 sub %bufp, len # update buffer length
129 crc32b %bl, crc_init_dw # compute crc32 of 1-byte
130 shr $8, tmp # get next byte
136 ################################################################
137 ## 2) PROCESS BLOCKS:
138 ################################################################
140 ## compute num of bytes to be processed
141 movq len, tmp # save num bytes in tmp
147 cmpq $SMALL_SIZE, len
151 movq $2731, %rax # 2731 = ceil(2^16 / 24)
155 ## eax contains floor(bytes / 24) = num 24-byte chunks to do
157 ## process rax 24-byte chunks (128 >= rax >= 0)
159 ## compute end address of each block
160 ## block 0 (base addr + RAX * 8)
161 ## block 1 (base addr + RAX * 16)
162 ## block 2 (base addr + RAX * 24)
163 lea (bufptmp, %rax, 8), block_0
164 lea (block_0, %rax, 8), block_1
165 lea (block_1, %rax, 8), block_2
171 mov jump_table(,%rax,8), %bufp
174 ################################################################
175 ## 2a) PROCESS FULL BLOCKS:
176 ################################################################
179 lea 128*8*2(block_0), block_1
180 lea 128*8*3(block_0), block_2
181 add $128*8*1, block_0
186 # Fall thruogh into top of crc array (crc_128)
188 ################################################################
190 ################################################################
198 crc32q -i*8(block_0), crc_init
199 crc32q -i*8(block_1), crc1
200 crc32q -i*8(block_2), crc2
207 crc32q -i*8(block_0), crc_init
208 crc32q -i*8(block_1), crc1
209 # SKIP crc32 -i*8(block_2), crc2 ; Don't do this one yet
213 ################################################################
214 ## 4) Combine three results:
215 ################################################################
217 lea (K_table-8)(%rip), %bufp # first entry is for idx 1
218 shlq $3, %rax # rax *= 8
219 pmovzxdq (%bufp,%rax), %xmm0 # 2 consts: K1:K2
220 leal (%eax,%eax,2), %eax # rax *= 3 (total *24)
221 subq %rax, tmp # tmp -= rax*24
223 movq crc_init, %xmm1 # CRC for block 1
224 pclmulqdq $0x00, %xmm0, %xmm1 # Multiply by K2
226 movq crc1, %xmm2 # CRC for block 2
227 pclmulqdq $0x10, %xmm0, %xmm2 # Multiply by K1
231 xor -i*8(block_2), %rax
235 ################################################################
237 ################################################################
247 shl $32-4, len_dw # less_than_16 expects length
248 # in upper 4 bits of len_dw
250 crc32q (bufptmp), crc_init
251 crc32q 8(bufptmp), crc_init
254 # len is less than 8 if we got here
255 # less_than_8 expects length in upper 3 bits of len_dw
256 # less_than_8_post_shl1 expects length = carryflag * 8 + len_dw[31:30]
258 jmp less_than_8_post_shl1
260 #######################################################################
261 ## 6) LESS THAN 256-bytes REMAIN AT THIS POINT (8-bits of len are full)
262 #######################################################################
264 shl $32-8, len_dw # Prepare len_dw for less_than_256
266 .rept 5 # j = {256, 128, 64, 32, 16}
268 LABEL less_than_ %j # less_than_j: Length should be in
269 # upper lg(j) bits of len_dw
271 shl $1, len_dw # Get next MSB
276 crc32q i(bufptmp), crc_init # Compute crc32 of 8-byte data
279 jz do_return # Return if remaining length is zero
280 add $j, bufptmp # Advance buf
283 less_than_8: # Length should be stored in
284 # upper 3 bits of len_dw
286 less_than_8_post_shl1:
288 crc32l (bufptmp), crc_init_dw # CRC of 4 bytes
289 jz do_return # return if remaining data is zero
291 less_than_4: # Length should be stored in
292 # upper 2 bits of len_dw
295 crc32w (bufptmp), crc_init_dw # CRC of 2 bytes
296 jz do_return # return if remaining data is zero
298 less_than_2: # Length should be stored in the MSB
302 crc32b (bufptmp), crc_init_dw # CRC of 1 byte
303 less_than_1: # Length should be zero
310 SYM_FUNC_END(crc_pcl)
312 .section .rodata, "a", @progbits
313 ################################################################
314 ## jump table Table is 129 entries x 2 bytes each
315 ################################################################
327 ################################################################
329 ## Table is 128 entries x 2 words (8 bytes) each
330 ################################################################
333 .long 0x493c7d27, 0x00000001
334 .long 0xba4fc28e, 0x493c7d27
335 .long 0xddc0152b, 0xf20c0dfe
336 .long 0x9e4addf8, 0xba4fc28e
337 .long 0x39d3b296, 0x3da6d0cb
338 .long 0x0715ce53, 0xddc0152b
339 .long 0x47db8317, 0x1c291d04
340 .long 0x0d3b6092, 0x9e4addf8
341 .long 0xc96cfdc0, 0x740eef02
342 .long 0x878a92a7, 0x39d3b296
343 .long 0xdaece73e, 0x083a6eec
344 .long 0xab7aff2a, 0x0715ce53
345 .long 0x2162d385, 0xc49f4f67
346 .long 0x83348832, 0x47db8317
347 .long 0x299847d5, 0x2ad91c30
348 .long 0xb9e02b86, 0x0d3b6092
349 .long 0x18b33a4e, 0x6992cea2
350 .long 0xb6dd949b, 0xc96cfdc0
351 .long 0x78d9ccb7, 0x7e908048
352 .long 0xbac2fd7b, 0x878a92a7
353 .long 0xa60ce07b, 0x1b3d8f29
354 .long 0xce7f39f4, 0xdaece73e
355 .long 0x61d82e56, 0xf1d0f55e
356 .long 0xd270f1a2, 0xab7aff2a
357 .long 0xc619809d, 0xa87ab8a8
358 .long 0x2b3cac5d, 0x2162d385
359 .long 0x65863b64, 0x8462d800
360 .long 0x1b03397f, 0x83348832
361 .long 0xebb883bd, 0x71d111a8
362 .long 0xb3e32c28, 0x299847d5
363 .long 0x064f7f26, 0xffd852c6
364 .long 0xdd7e3b0c, 0xb9e02b86
365 .long 0xf285651c, 0xdcb17aa4
366 .long 0x10746f3c, 0x18b33a4e
367 .long 0xc7a68855, 0xf37c5aee
368 .long 0x271d9844, 0xb6dd949b
369 .long 0x8e766a0c, 0x6051d5a2
370 .long 0x93a5f730, 0x78d9ccb7
371 .long 0x6cb08e5c, 0x18b0d4ff
372 .long 0x6b749fb2, 0xbac2fd7b
373 .long 0x1393e203, 0x21f3d99c
374 .long 0xcec3662e, 0xa60ce07b
375 .long 0x96c515bb, 0x8f158014
376 .long 0xe6fc4e6a, 0xce7f39f4
377 .long 0x8227bb8a, 0xa00457f7
378 .long 0xb0cd4768, 0x61d82e56
379 .long 0x39c7ff35, 0x8d6d2c43
380 .long 0xd7a4825c, 0xd270f1a2
381 .long 0x0ab3844b, 0x00ac29cf
382 .long 0x0167d312, 0xc619809d
383 .long 0xf6076544, 0xe9adf796
384 .long 0x26f6a60a, 0x2b3cac5d
385 .long 0xa741c1bf, 0x96638b34
386 .long 0x98d8d9cb, 0x65863b64
387 .long 0x49c3cc9c, 0xe0e9f351
388 .long 0x68bce87a, 0x1b03397f
389 .long 0x57a3d037, 0x9af01f2d
390 .long 0x6956fc3b, 0xebb883bd
391 .long 0x42d98888, 0x2cff42cf
392 .long 0x3771e98f, 0xb3e32c28
393 .long 0xb42ae3d9, 0x88f25a3a
394 .long 0x2178513a, 0x064f7f26
395 .long 0xe0ac139e, 0x4e36f0b0
396 .long 0x170076fa, 0xdd7e3b0c
397 .long 0x444dd413, 0xbd6f81f8
398 .long 0x6f345e45, 0xf285651c
399 .long 0x41d17b64, 0x91c9bd4b
400 .long 0xff0dba97, 0x10746f3c
401 .long 0xa2b73df1, 0x885f087b
402 .long 0xf872e54c, 0xc7a68855
403 .long 0x1e41e9fc, 0x4c144932
404 .long 0x86d8e4d2, 0x271d9844
405 .long 0x651bd98b, 0x52148f02
406 .long 0x5bb8f1bc, 0x8e766a0c
407 .long 0xa90fd27a, 0xa3c6f37a
408 .long 0xb3af077a, 0x93a5f730
409 .long 0x4984d782, 0xd7c0557f
410 .long 0xca6ef3ac, 0x6cb08e5c
411 .long 0x234e0b26, 0x63ded06a
412 .long 0xdd66cbbb, 0x6b749fb2
413 .long 0x4597456a, 0x4d56973c
414 .long 0xe9e28eb4, 0x1393e203
415 .long 0x7b3ff57a, 0x9669c9df
416 .long 0xc9c8b782, 0xcec3662e
417 .long 0x3f70cc6f, 0xe417f38a
418 .long 0x93e106a4, 0x96c515bb
419 .long 0x62ec6c6d, 0x4b9e0f71
420 .long 0xd813b325, 0xe6fc4e6a
421 .long 0x0df04680, 0xd104b8fc
422 .long 0x2342001e, 0x8227bb8a
423 .long 0x0a2a8d7e, 0x5b397730
424 .long 0x6d9a4957, 0xb0cd4768
425 .long 0xe8b6368b, 0xe78eb416
426 .long 0xd2c3ed1a, 0x39c7ff35
427 .long 0x995a5724, 0x61ff0e01
428 .long 0x9ef68d35, 0xd7a4825c
429 .long 0x0c139b31, 0x8d96551c
430 .long 0xf2271e60, 0x0ab3844b
431 .long 0x0b0bf8ca, 0x0bf80dd2
432 .long 0x2664fd8b, 0x0167d312
433 .long 0xed64812d, 0x8821abed
434 .long 0x02ee03b2, 0xf6076544
435 .long 0x8604ae0f, 0x6a45d2b2
436 .long 0x363bd6b3, 0x26f6a60a
437 .long 0x135c83fd, 0xd8d26619
438 .long 0x5fabe670, 0xa741c1bf
439 .long 0x35ec3279, 0xde87806c
440 .long 0x00bcf5f6, 0x98d8d9cb
441 .long 0x8ae00689, 0x14338754
442 .long 0x17f27698, 0x49c3cc9c
443 .long 0x58ca5f00, 0x5bd2011f
444 .long 0xaa7c7ad5, 0x68bce87a
445 .long 0xb5cfca28, 0xdd07448e
446 .long 0xded288f8, 0x57a3d037
447 .long 0x59f229bc, 0xdde8f5b9
448 .long 0x6d390dec, 0x6956fc3b
449 .long 0x37170390, 0xa3e3e02c
450 .long 0x6353c1cc, 0x42d98888
451 .long 0xc4584f5c, 0xd73c7bea
452 .long 0xf48642e9, 0x3771e98f
453 .long 0x531377e2, 0x80ff0093
454 .long 0xdd35bc8d, 0xb42ae3d9
455 .long 0xb25b29f2, 0x8fe4c34d
456 .long 0x9a5ede41, 0x2178513a
457 .long 0xa563905d, 0xdf99fc11
458 .long 0x45cddf4e, 0xe0ac139e
459 .long 0xacfa3103, 0x6c23e841
460 .long 0xa51b6135, 0x170076fa