xref: /openbmc/linux/arch/x86/lib/checksum_32.S (revision aac28965)
1/* SPDX-License-Identifier: GPL-2.0-or-later */
2/*
3 * INET		An implementation of the TCP/IP protocol suite for the LINUX
4 *		operating system.  INET is implemented using the  BSD Socket
5 *		interface as the means of communication with the user level.
6 *
7 *		IP/TCP/UDP checksumming routines
8 *
9 * Authors:	Jorge Cwik, <jorge@laser.satlink.net>
10 *		Arnt Gulbrandsen, <agulbra@nvg.unit.no>
11 *		Tom May, <ftom@netcom.com>
12 *              Pentium Pro/II routines:
13 *              Alexander Kjeldaas <astor@guardian.no>
14 *              Finn Arne Gangstad <finnag@guardian.no>
15 *		Lots of code moved from tcp.c and ip.c; see those files
16 *		for more names.
17 *
18 * Changes:     Ingo Molnar, converted csum_partial_copy() to 2.1 exception
19 *			     handling.
20 *		Andi Kleen,  add zeroing on error
21 *                   converted to pure assembler
22 */
23
24#include <linux/linkage.h>
25#include <asm/errno.h>
26#include <asm/asm.h>
27#include <asm/export.h>
28#include <asm/nospec-branch.h>
29
30/*
31 * computes a partial checksum, e.g. for TCP/UDP fragments
32 */
33
34/*
35unsigned int csum_partial(const unsigned char * buff, int len, unsigned int sum)
36 */
37
38.text
39
40#ifndef CONFIG_X86_USE_PPRO_CHECKSUM
41
42	  /*
43	   * Experiments with Ethernet and SLIP connections show that buff
44	   * is aligned on either a 2-byte or 4-byte boundary.  We get at
45	   * least a twofold speedup on 486 and Pentium if it is 4-byte aligned.
46	   * Fortunately, it is easy to convert 2-byte alignment to 4-byte
47	   * alignment for the unrolled loop.
48	   */
49SYM_FUNC_START(csum_partial)
50	pushl %esi
51	pushl %ebx
52	movl 20(%esp),%eax	# Function arg: unsigned int sum
53	movl 16(%esp),%ecx	# Function arg: int len
54	movl 12(%esp),%esi	# Function arg: unsigned char *buff
55	testl $3, %esi		# Check alignment.
56	jz 2f			# Jump if alignment is ok.
57	testl $1, %esi		# Check alignment.
58	jz 10f			# Jump if alignment is boundary of 2 bytes.
59
60	# buf is odd
61	dec %ecx
62	jl 8f
63	movzbl (%esi), %ebx
64	adcl %ebx, %eax
65	roll $8, %eax
66	inc %esi
67	testl $2, %esi
68	jz 2f
6910:
70	subl $2, %ecx		# Alignment uses up two bytes.
71	jae 1f			# Jump if we had at least two bytes.
72	addl $2, %ecx		# ecx was < 2.  Deal with it.
73	jmp 4f
741:	movw (%esi), %bx
75	addl $2, %esi
76	addw %bx, %ax
77	adcl $0, %eax
782:
79	movl %ecx, %edx
80	shrl $5, %ecx
81	jz 2f
82	testl %esi, %esi
831:	movl (%esi), %ebx
84	adcl %ebx, %eax
85	movl 4(%esi), %ebx
86	adcl %ebx, %eax
87	movl 8(%esi), %ebx
88	adcl %ebx, %eax
89	movl 12(%esi), %ebx
90	adcl %ebx, %eax
91	movl 16(%esi), %ebx
92	adcl %ebx, %eax
93	movl 20(%esi), %ebx
94	adcl %ebx, %eax
95	movl 24(%esi), %ebx
96	adcl %ebx, %eax
97	movl 28(%esi), %ebx
98	adcl %ebx, %eax
99	lea 32(%esi), %esi
100	dec %ecx
101	jne 1b
102	adcl $0, %eax
1032:	movl %edx, %ecx
104	andl $0x1c, %edx
105	je 4f
106	shrl $2, %edx		# This clears CF
1073:	adcl (%esi), %eax
108	lea 4(%esi), %esi
109	dec %edx
110	jne 3b
111	adcl $0, %eax
1124:	andl $3, %ecx
113	jz 7f
114	cmpl $2, %ecx
115	jb 5f
116	movw (%esi),%cx
117	leal 2(%esi),%esi
118	je 6f
119	shll $16,%ecx
1205:	movb (%esi),%cl
1216:	addl %ecx,%eax
122	adcl $0, %eax
1237:
124	testb $1, 12(%esp)
125	jz 8f
126	roll $8, %eax
1278:
128	popl %ebx
129	popl %esi
130	RET
131SYM_FUNC_END(csum_partial)
132
133#else
134
135/* Version for PentiumII/PPro */
136
137SYM_FUNC_START(csum_partial)
138	pushl %esi
139	pushl %ebx
140	movl 20(%esp),%eax	# Function arg: unsigned int sum
141	movl 16(%esp),%ecx	# Function arg: int len
142	movl 12(%esp),%esi	# Function arg:	const unsigned char *buf
143
144	testl $3, %esi
145	jnz 25f
14610:
147	movl %ecx, %edx
148	movl %ecx, %ebx
149	andl $0x7c, %ebx
150	shrl $7, %ecx
151	addl %ebx,%esi
152	shrl $2, %ebx
153	negl %ebx
154	lea 45f(%ebx,%ebx,2), %ebx
155	testl %esi, %esi
156	JMP_NOSPEC ebx
157
158	# Handle 2-byte-aligned regions
15920:	addw (%esi), %ax
160	lea 2(%esi), %esi
161	adcl $0, %eax
162	jmp 10b
16325:
164	testl $1, %esi
165	jz 30f
166	# buf is odd
167	dec %ecx
168	jl 90f
169	movzbl (%esi), %ebx
170	addl %ebx, %eax
171	adcl $0, %eax
172	roll $8, %eax
173	inc %esi
174	testl $2, %esi
175	jz 10b
176
17730:	subl $2, %ecx
178	ja 20b
179	je 32f
180	addl $2, %ecx
181	jz 80f
182	movzbl (%esi),%ebx	# csumming 1 byte, 2-aligned
183	addl %ebx, %eax
184	adcl $0, %eax
185	jmp 80f
18632:
187	addw (%esi), %ax	# csumming 2 bytes, 2-aligned
188	adcl $0, %eax
189	jmp 80f
190
19140:
192	addl -128(%esi), %eax
193	adcl -124(%esi), %eax
194	adcl -120(%esi), %eax
195	adcl -116(%esi), %eax
196	adcl -112(%esi), %eax
197	adcl -108(%esi), %eax
198	adcl -104(%esi), %eax
199	adcl -100(%esi), %eax
200	adcl -96(%esi), %eax
201	adcl -92(%esi), %eax
202	adcl -88(%esi), %eax
203	adcl -84(%esi), %eax
204	adcl -80(%esi), %eax
205	adcl -76(%esi), %eax
206	adcl -72(%esi), %eax
207	adcl -68(%esi), %eax
208	adcl -64(%esi), %eax
209	adcl -60(%esi), %eax
210	adcl -56(%esi), %eax
211	adcl -52(%esi), %eax
212	adcl -48(%esi), %eax
213	adcl -44(%esi), %eax
214	adcl -40(%esi), %eax
215	adcl -36(%esi), %eax
216	adcl -32(%esi), %eax
217	adcl -28(%esi), %eax
218	adcl -24(%esi), %eax
219	adcl -20(%esi), %eax
220	adcl -16(%esi), %eax
221	adcl -12(%esi), %eax
222	adcl -8(%esi), %eax
223	adcl -4(%esi), %eax
22445:
225	lea 128(%esi), %esi
226	adcl $0, %eax
227	dec %ecx
228	jge 40b
229	movl %edx, %ecx
23050:	andl $3, %ecx
231	jz 80f
232
233	# Handle the last 1-3 bytes without jumping
234	notl %ecx		# 1->2, 2->1, 3->0, higher bits are masked
235	movl $0xffffff,%ebx	# by the shll and shrl instructions
236	shll $3,%ecx
237	shrl %cl,%ebx
238	andl -128(%esi),%ebx	# esi is 4-aligned so should be ok
239	addl %ebx,%eax
240	adcl $0,%eax
24180:
242	testb $1, 12(%esp)
243	jz 90f
244	roll $8, %eax
24590:
246	popl %ebx
247	popl %esi
248	RET
249SYM_FUNC_END(csum_partial)
250
251#endif
252EXPORT_SYMBOL(csum_partial)
253
254/*
255unsigned int csum_partial_copy_generic (const char *src, char *dst,
256				  int len)
257 */
258
259/*
260 * Copy from ds while checksumming, otherwise like csum_partial
261 */
262
263#define EXC(y...)						\
264	9999: y;						\
265	_ASM_EXTABLE_TYPE(9999b, 7f, EX_TYPE_UACCESS | EX_FLAG_CLEAR_AX)
266
267#ifndef CONFIG_X86_USE_PPRO_CHECKSUM
268
269#define ARGBASE 16
270#define FP		12
271
272SYM_FUNC_START(csum_partial_copy_generic)
273	subl  $4,%esp
274	pushl %edi
275	pushl %esi
276	pushl %ebx
277	movl ARGBASE+12(%esp),%ecx	# len
278	movl ARGBASE+4(%esp),%esi	# src
279	movl ARGBASE+8(%esp),%edi	# dst
280
281	movl $-1, %eax			# sum
282	testl $2, %edi			# Check alignment.
283	jz 2f				# Jump if alignment is ok.
284	subl $2, %ecx			# Alignment uses up two bytes.
285	jae 1f				# Jump if we had at least two bytes.
286	addl $2, %ecx			# ecx was < 2.  Deal with it.
287	jmp 4f
288EXC(1:	movw (%esi), %bx	)
289	addl $2, %esi
290EXC(	movw %bx, (%edi)	)
291	addl $2, %edi
292	addw %bx, %ax
293	adcl $0, %eax
2942:
295	movl %ecx, FP(%esp)
296	shrl $5, %ecx
297	jz 2f
298	testl %esi, %esi		# what's wrong with clc?
299EXC(1:	movl (%esi), %ebx	)
300EXC(	movl 4(%esi), %edx	)
301	adcl %ebx, %eax
302EXC(	movl %ebx, (%edi)	)
303	adcl %edx, %eax
304EXC(	movl %edx, 4(%edi)	)
305
306EXC(	movl 8(%esi), %ebx	)
307EXC(	movl 12(%esi), %edx	)
308	adcl %ebx, %eax
309EXC(	movl %ebx, 8(%edi)	)
310	adcl %edx, %eax
311EXC(	movl %edx, 12(%edi)	)
312
313EXC(	movl 16(%esi), %ebx 	)
314EXC(	movl 20(%esi), %edx	)
315	adcl %ebx, %eax
316EXC(	movl %ebx, 16(%edi)	)
317	adcl %edx, %eax
318EXC(	movl %edx, 20(%edi)	)
319
320EXC(	movl 24(%esi), %ebx	)
321EXC(	movl 28(%esi), %edx	)
322	adcl %ebx, %eax
323EXC(	movl %ebx, 24(%edi)	)
324	adcl %edx, %eax
325EXC(	movl %edx, 28(%edi)	)
326
327	lea 32(%esi), %esi
328	lea 32(%edi), %edi
329	dec %ecx
330	jne 1b
331	adcl $0, %eax
3322:	movl FP(%esp), %edx
333	movl %edx, %ecx
334	andl $0x1c, %edx
335	je 4f
336	shrl $2, %edx			# This clears CF
337EXC(3:	movl (%esi), %ebx	)
338	adcl %ebx, %eax
339EXC(	movl %ebx, (%edi)	)
340	lea 4(%esi), %esi
341	lea 4(%edi), %edi
342	dec %edx
343	jne 3b
344	adcl $0, %eax
3454:	andl $3, %ecx
346	jz 7f
347	cmpl $2, %ecx
348	jb 5f
349EXC(	movw (%esi), %cx	)
350	leal 2(%esi), %esi
351EXC(	movw %cx, (%edi)	)
352	leal 2(%edi), %edi
353	je 6f
354	shll $16,%ecx
355EXC(5:	movb (%esi), %cl	)
356EXC(	movb %cl, (%edi)	)
3576:	addl %ecx, %eax
358	adcl $0, %eax
3597:
360
361	popl %ebx
362	popl %esi
363	popl %edi
364	popl %ecx			# equivalent to addl $4,%esp
365	RET
366SYM_FUNC_END(csum_partial_copy_generic)
367
368#else
369
370/* Version for PentiumII/PPro */
371
372#define ROUND1(x) \
373	EXC(movl x(%esi), %ebx	)	;	\
374	addl %ebx, %eax			;	\
375	EXC(movl %ebx, x(%edi)	)	;
376
377#define ROUND(x) \
378	EXC(movl x(%esi), %ebx	)	;	\
379	adcl %ebx, %eax			;	\
380	EXC(movl %ebx, x(%edi)	)	;
381
382#define ARGBASE 12
383
384SYM_FUNC_START(csum_partial_copy_generic)
385	pushl %ebx
386	pushl %edi
387	pushl %esi
388	movl ARGBASE+4(%esp),%esi	#src
389	movl ARGBASE+8(%esp),%edi	#dst
390	movl ARGBASE+12(%esp),%ecx	#len
391	movl $-1, %eax			#sum
392#	movl %ecx, %edx
393	movl %ecx, %ebx
394	movl %esi, %edx
395	shrl $6, %ecx
396	andl $0x3c, %ebx
397	negl %ebx
398	subl %ebx, %esi
399	subl %ebx, %edi
400	lea  -1(%esi),%edx
401	andl $-32,%edx
402	lea 3f(%ebx,%ebx), %ebx
403	testl %esi, %esi
404	JMP_NOSPEC ebx
4051:	addl $64,%esi
406	addl $64,%edi
407	EXC(movb -32(%edx),%bl)	; EXC(movb (%edx),%bl)
408	ROUND1(-64) ROUND(-60) ROUND(-56) ROUND(-52)
409	ROUND (-48) ROUND(-44) ROUND(-40) ROUND(-36)
410	ROUND (-32) ROUND(-28) ROUND(-24) ROUND(-20)
411	ROUND (-16) ROUND(-12) ROUND(-8)  ROUND(-4)
4123:	adcl $0,%eax
413	addl $64, %edx
414	dec %ecx
415	jge 1b
4164:	movl ARGBASE+12(%esp),%edx	#len
417	andl $3, %edx
418	jz 7f
419	cmpl $2, %edx
420	jb 5f
421EXC(	movw (%esi), %dx         )
422	leal 2(%esi), %esi
423EXC(	movw %dx, (%edi)         )
424	leal 2(%edi), %edi
425	je 6f
426	shll $16,%edx
4275:
428EXC(	movb (%esi), %dl         )
429EXC(	movb %dl, (%edi)         )
4306:	addl %edx, %eax
431	adcl $0, %eax
4327:
433
434	popl %esi
435	popl %edi
436	popl %ebx
437	RET
438SYM_FUNC_END(csum_partial_copy_generic)
439
440#undef ROUND
441#undef ROUND1
442
443#endif
444EXPORT_SYMBOL(csum_partial_copy_generic)
445