1########################################################################
2# Implement fast CRC-T10DIF computation with SSE and PCLMULQDQ instructions
3#
4# Copyright (c) 2013, Intel Corporation
5#
6# Authors:
7#     Erdinc Ozturk <erdinc.ozturk@intel.com>
8#     Vinodh Gopal <vinodh.gopal@intel.com>
9#     James Guilford <james.guilford@intel.com>
10#     Tim Chen <tim.c.chen@linux.intel.com>
11#
12# This software is available to you under a choice of one of two
13# licenses.  You may choose to be licensed under the terms of the GNU
14# General Public License (GPL) Version 2, available from the file
15# COPYING in the main directory of this source tree, or the
16# OpenIB.org BSD license below:
17#
18# Redistribution and use in source and binary forms, with or without
19# modification, are permitted provided that the following conditions are
20# met:
21#
22# * Redistributions of source code must retain the above copyright
23#   notice, this list of conditions and the following disclaimer.
24#
25# * Redistributions in binary form must reproduce the above copyright
26#   notice, this list of conditions and the following disclaimer in the
27#   documentation and/or other materials provided with the
28#   distribution.
29#
30# * Neither the name of the Intel Corporation nor the names of its
31#   contributors may be used to endorse or promote products derived from
32#   this software without specific prior written permission.
33#
34#
35# THIS SOFTWARE IS PROVIDED BY INTEL CORPORATION ""AS IS"" AND ANY
36# EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
37# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
38# PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL INTEL CORPORATION OR
39# CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
40# EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
41# PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
42# PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
43# LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
44# NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
45# SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
46#
47#       Reference paper titled "Fast CRC Computation for Generic
48#	Polynomials Using PCLMULQDQ Instruction"
49#       URL: http://www.intel.com/content/dam/www/public/us/en/documents
50#  /white-papers/fast-crc-computation-generic-polynomials-pclmulqdq-paper.pdf
51#
52
53#include <linux/linkage.h>
54
55.text
56
57#define		init_crc	%edi
58#define		buf		%rsi
59#define		len		%rdx
60
61#define		FOLD_CONSTS	%xmm10
62#define		BSWAP_MASK	%xmm11
63
64# Fold reg1, reg2 into the next 32 data bytes, storing the result back into
65# reg1, reg2.
66.macro	fold_32_bytes	offset, reg1, reg2
67	movdqu	\offset(buf), %xmm9
68	movdqu	\offset+16(buf), %xmm12
69	pshufb	BSWAP_MASK, %xmm9
70	pshufb	BSWAP_MASK, %xmm12
71	movdqa	\reg1, %xmm8
72	movdqa	\reg2, %xmm13
73	pclmulqdq	$0x00, FOLD_CONSTS, \reg1
74	pclmulqdq	$0x11, FOLD_CONSTS, %xmm8
75	pclmulqdq	$0x00, FOLD_CONSTS, \reg2
76	pclmulqdq	$0x11, FOLD_CONSTS, %xmm13
77	pxor	%xmm9 , \reg1
78	xorps	%xmm8 , \reg1
79	pxor	%xmm12, \reg2
80	xorps	%xmm13, \reg2
81.endm
82
83# Fold src_reg into dst_reg.
84.macro	fold_16_bytes	src_reg, dst_reg
85	movdqa	\src_reg, %xmm8
86	pclmulqdq	$0x11, FOLD_CONSTS, \src_reg
87	pclmulqdq	$0x00, FOLD_CONSTS, %xmm8
88	pxor	%xmm8, \dst_reg
89	xorps	\src_reg, \dst_reg
90.endm
91
92#
93# u16 crc_t10dif_pcl(u16 init_crc, const *u8 buf, size_t len);
94#
95# Assumes len >= 16.
96#
97.align 16
98SYM_FUNC_START(crc_t10dif_pcl)
99
100	movdqa	.Lbswap_mask(%rip), BSWAP_MASK
101
102	# For sizes less than 256 bytes, we can't fold 128 bytes at a time.
103	cmp	$256, len
104	jl	.Lless_than_256_bytes
105
106	# Load the first 128 data bytes.  Byte swapping is necessary to make the
107	# bit order match the polynomial coefficient order.
108	movdqu	16*0(buf), %xmm0
109	movdqu	16*1(buf), %xmm1
110	movdqu	16*2(buf), %xmm2
111	movdqu	16*3(buf), %xmm3
112	movdqu	16*4(buf), %xmm4
113	movdqu	16*5(buf), %xmm5
114	movdqu	16*6(buf), %xmm6
115	movdqu	16*7(buf), %xmm7
116	add	$128, buf
117	pshufb	BSWAP_MASK, %xmm0
118	pshufb	BSWAP_MASK, %xmm1
119	pshufb	BSWAP_MASK, %xmm2
120	pshufb	BSWAP_MASK, %xmm3
121	pshufb	BSWAP_MASK, %xmm4
122	pshufb	BSWAP_MASK, %xmm5
123	pshufb	BSWAP_MASK, %xmm6
124	pshufb	BSWAP_MASK, %xmm7
125
126	# XOR the first 16 data *bits* with the initial CRC value.
127	pxor	%xmm8, %xmm8
128	pinsrw	$7, init_crc, %xmm8
129	pxor	%xmm8, %xmm0
130
131	movdqa	.Lfold_across_128_bytes_consts(%rip), FOLD_CONSTS
132
133	# Subtract 128 for the 128 data bytes just consumed.  Subtract another
134	# 128 to simplify the termination condition of the following loop.
135	sub	$256, len
136
137	# While >= 128 data bytes remain (not counting xmm0-7), fold the 128
138	# bytes xmm0-7 into them, storing the result back into xmm0-7.
139.Lfold_128_bytes_loop:
140	fold_32_bytes	0, %xmm0, %xmm1
141	fold_32_bytes	32, %xmm2, %xmm3
142	fold_32_bytes	64, %xmm4, %xmm5
143	fold_32_bytes	96, %xmm6, %xmm7
144	add	$128, buf
145	sub	$128, len
146	jge	.Lfold_128_bytes_loop
147
148	# Now fold the 112 bytes in xmm0-xmm6 into the 16 bytes in xmm7.
149
150	# Fold across 64 bytes.
151	movdqa	.Lfold_across_64_bytes_consts(%rip), FOLD_CONSTS
152	fold_16_bytes	%xmm0, %xmm4
153	fold_16_bytes	%xmm1, %xmm5
154	fold_16_bytes	%xmm2, %xmm6
155	fold_16_bytes	%xmm3, %xmm7
156	# Fold across 32 bytes.
157	movdqa	.Lfold_across_32_bytes_consts(%rip), FOLD_CONSTS
158	fold_16_bytes	%xmm4, %xmm6
159	fold_16_bytes	%xmm5, %xmm7
160	# Fold across 16 bytes.
161	movdqa	.Lfold_across_16_bytes_consts(%rip), FOLD_CONSTS
162	fold_16_bytes	%xmm6, %xmm7
163
164	# Add 128 to get the correct number of data bytes remaining in 0...127
165	# (not counting xmm7), following the previous extra subtraction by 128.
166	# Then subtract 16 to simplify the termination condition of the
167	# following loop.
168	add	$128-16, len
169
170	# While >= 16 data bytes remain (not counting xmm7), fold the 16 bytes
171	# xmm7 into them, storing the result back into xmm7.
172	jl	.Lfold_16_bytes_loop_done
173.Lfold_16_bytes_loop:
174	movdqa	%xmm7, %xmm8
175	pclmulqdq	$0x11, FOLD_CONSTS, %xmm7
176	pclmulqdq	$0x00, FOLD_CONSTS, %xmm8
177	pxor	%xmm8, %xmm7
178	movdqu	(buf), %xmm0
179	pshufb	BSWAP_MASK, %xmm0
180	pxor	%xmm0 , %xmm7
181	add	$16, buf
182	sub	$16, len
183	jge	.Lfold_16_bytes_loop
184
185.Lfold_16_bytes_loop_done:
186	# Add 16 to get the correct number of data bytes remaining in 0...15
187	# (not counting xmm7), following the previous extra subtraction by 16.
188	add	$16, len
189	je	.Lreduce_final_16_bytes
190
191.Lhandle_partial_segment:
192	# Reduce the last '16 + len' bytes where 1 <= len <= 15 and the first 16
193	# bytes are in xmm7 and the rest are the remaining data in 'buf'.  To do
194	# this without needing a fold constant for each possible 'len', redivide
195	# the bytes into a first chunk of 'len' bytes and a second chunk of 16
196	# bytes, then fold the first chunk into the second.
197
198	movdqa	%xmm7, %xmm2
199
200	# xmm1 = last 16 original data bytes
201	movdqu	-16(buf, len), %xmm1
202	pshufb	BSWAP_MASK, %xmm1
203
204	# xmm2 = high order part of second chunk: xmm7 left-shifted by 'len' bytes.
205	lea	.Lbyteshift_table+16(%rip), %rax
206	sub	len, %rax
207	movdqu	(%rax), %xmm0
208	pshufb	%xmm0, %xmm2
209
210	# xmm7 = first chunk: xmm7 right-shifted by '16-len' bytes.
211	pxor	.Lmask1(%rip), %xmm0
212	pshufb	%xmm0, %xmm7
213
214	# xmm1 = second chunk: 'len' bytes from xmm1 (low-order bytes),
215	# then '16-len' bytes from xmm2 (high-order bytes).
216	pblendvb	%xmm2, %xmm1	#xmm0 is implicit
217
218	# Fold the first chunk into the second chunk, storing the result in xmm7.
219	movdqa	%xmm7, %xmm8
220	pclmulqdq	$0x11, FOLD_CONSTS, %xmm7
221	pclmulqdq	$0x00, FOLD_CONSTS, %xmm8
222	pxor	%xmm8, %xmm7
223	pxor	%xmm1, %xmm7
224
225.Lreduce_final_16_bytes:
226	# Reduce the 128-bit value M(x), stored in xmm7, to the final 16-bit CRC
227
228	# Load 'x^48 * (x^48 mod G(x))' and 'x^48 * (x^80 mod G(x))'.
229	movdqa	.Lfinal_fold_consts(%rip), FOLD_CONSTS
230
231	# Fold the high 64 bits into the low 64 bits, while also multiplying by
232	# x^64.  This produces a 128-bit value congruent to x^64 * M(x) and
233	# whose low 48 bits are 0.
234	movdqa	%xmm7, %xmm0
235	pclmulqdq	$0x11, FOLD_CONSTS, %xmm7 # high bits * x^48 * (x^80 mod G(x))
236	pslldq	$8, %xmm0
237	pxor	%xmm0, %xmm7			  # + low bits * x^64
238
239	# Fold the high 32 bits into the low 96 bits.  This produces a 96-bit
240	# value congruent to x^64 * M(x) and whose low 48 bits are 0.
241	movdqa	%xmm7, %xmm0
242	pand	.Lmask2(%rip), %xmm0		  # zero high 32 bits
243	psrldq	$12, %xmm7			  # extract high 32 bits
244	pclmulqdq	$0x00, FOLD_CONSTS, %xmm7 # high 32 bits * x^48 * (x^48 mod G(x))
245	pxor	%xmm0, %xmm7			  # + low bits
246
247	# Load G(x) and floor(x^48 / G(x)).
248	movdqa	.Lbarrett_reduction_consts(%rip), FOLD_CONSTS
249
250	# Use Barrett reduction to compute the final CRC value.
251	movdqa	%xmm7, %xmm0
252	pclmulqdq	$0x11, FOLD_CONSTS, %xmm7 # high 32 bits * floor(x^48 / G(x))
253	psrlq	$32, %xmm7			  # /= x^32
254	pclmulqdq	$0x00, FOLD_CONSTS, %xmm7 # *= G(x)
255	psrlq	$48, %xmm0
256	pxor	%xmm7, %xmm0		     # + low 16 nonzero bits
257	# Final CRC value (x^16 * M(x)) mod G(x) is in low 16 bits of xmm0.
258
259	pextrw	$0, %xmm0, %eax
260	ret
261
262.align 16
263.Lless_than_256_bytes:
264	# Checksumming a buffer of length 16...255 bytes
265
266	# Load the first 16 data bytes.
267	movdqu	(buf), %xmm7
268	pshufb	BSWAP_MASK, %xmm7
269	add	$16, buf
270
271	# XOR the first 16 data *bits* with the initial CRC value.
272	pxor	%xmm0, %xmm0
273	pinsrw	$7, init_crc, %xmm0
274	pxor	%xmm0, %xmm7
275
276	movdqa	.Lfold_across_16_bytes_consts(%rip), FOLD_CONSTS
277	cmp	$16, len
278	je	.Lreduce_final_16_bytes		# len == 16
279	sub	$32, len
280	jge	.Lfold_16_bytes_loop		# 32 <= len <= 255
281	add	$16, len
282	jmp	.Lhandle_partial_segment	# 17 <= len <= 31
283SYM_FUNC_END(crc_t10dif_pcl)
284
285.section	.rodata, "a", @progbits
286.align 16
287
288# Fold constants precomputed from the polynomial 0x18bb7
289# G(x) = x^16 + x^15 + x^11 + x^9 + x^8 + x^7 + x^5 + x^4 + x^2 + x^1 + x^0
290.Lfold_across_128_bytes_consts:
291	.quad		0x0000000000006123	# x^(8*128)	mod G(x)
292	.quad		0x0000000000002295	# x^(8*128+64)	mod G(x)
293.Lfold_across_64_bytes_consts:
294	.quad		0x0000000000001069	# x^(4*128)	mod G(x)
295	.quad		0x000000000000dd31	# x^(4*128+64)	mod G(x)
296.Lfold_across_32_bytes_consts:
297	.quad		0x000000000000857d	# x^(2*128)	mod G(x)
298	.quad		0x0000000000007acc	# x^(2*128+64)	mod G(x)
299.Lfold_across_16_bytes_consts:
300	.quad		0x000000000000a010	# x^(1*128)	mod G(x)
301	.quad		0x0000000000001faa	# x^(1*128+64)	mod G(x)
302.Lfinal_fold_consts:
303	.quad		0x1368000000000000	# x^48 * (x^48 mod G(x))
304	.quad		0x2d56000000000000	# x^48 * (x^80 mod G(x))
305.Lbarrett_reduction_consts:
306	.quad		0x0000000000018bb7	# G(x)
307	.quad		0x00000001f65a57f8	# floor(x^48 / G(x))
308
309.section	.rodata.cst16.mask1, "aM", @progbits, 16
310.align 16
311.Lmask1:
312	.octa	0x80808080808080808080808080808080
313
314.section	.rodata.cst16.mask2, "aM", @progbits, 16
315.align 16
316.Lmask2:
317	.octa	0x00000000FFFFFFFFFFFFFFFFFFFFFFFF
318
319.section	.rodata.cst16.bswap_mask, "aM", @progbits, 16
320.align 16
321.Lbswap_mask:
322	.octa	0x000102030405060708090A0B0C0D0E0F
323
324.section	.rodata.cst32.byteshift_table, "aM", @progbits, 32
325.align 16
326# For 1 <= len <= 15, the 16-byte vector beginning at &byteshift_table[16 - len]
327# is the index vector to shift left by 'len' bytes, and is also {0x80, ...,
328# 0x80} XOR the index vector to shift right by '16 - len' bytes.
329.Lbyteshift_table:
330	.byte		 0x0, 0x81, 0x82, 0x83, 0x84, 0x85, 0x86, 0x87
331	.byte		0x88, 0x89, 0x8a, 0x8b, 0x8c, 0x8d, 0x8e, 0x8f
332	.byte		 0x0,  0x1,  0x2,  0x3,  0x4,  0x5,  0x6,  0x7
333	.byte		 0x8,  0x9,  0xa,  0xb,  0xc,  0xd,  0xe , 0x0
334