xref: /linux/arch/x86/lib/checksum_32.S (revision 37744feebc086908fd89760650f458ab19071750)
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, int sum, int *src_err_ptr, int *dst_err_ptr)
257 */
258
259/*
260 * Copy from ds while checksumming, otherwise like csum_partial
261 *
262 * The macros SRC and DST specify the type of access for the instruction.
263 * thus we can call a custom exception handler for all access types.
264 *
265 * FIXME: could someone double-check whether I haven't mixed up some SRC and
266 *	  DST definitions? It's damn hard to trigger all cases.  I hope I got
267 *	  them all but there's no guarantee.
268 */
269
270#define SRC(y...)			\
271	9999: y;			\
272	_ASM_EXTABLE_UA(9999b, 6001f)
273
274#define DST(y...)			\
275	9999: y;			\
276	_ASM_EXTABLE_UA(9999b, 6002f)
277
278#ifndef CONFIG_X86_USE_PPRO_CHECKSUM
279
280#define ARGBASE 16
281#define FP		12
282
283SYM_FUNC_START(csum_partial_copy_generic)
284	subl  $4,%esp
285	pushl %edi
286	pushl %esi
287	pushl %ebx
288	movl ARGBASE+16(%esp),%eax	# sum
289	movl ARGBASE+12(%esp),%ecx	# len
290	movl ARGBASE+4(%esp),%esi	# src
291	movl ARGBASE+8(%esp),%edi	# dst
292
293	testl $2, %edi			# Check alignment.
294	jz 2f				# Jump if alignment is ok.
295	subl $2, %ecx			# Alignment uses up two bytes.
296	jae 1f				# Jump if we had at least two bytes.
297	addl $2, %ecx			# ecx was < 2.  Deal with it.
298	jmp 4f
299SRC(1:	movw (%esi), %bx	)
300	addl $2, %esi
301DST(	movw %bx, (%edi)	)
302	addl $2, %edi
303	addw %bx, %ax
304	adcl $0, %eax
3052:
306	movl %ecx, FP(%esp)
307	shrl $5, %ecx
308	jz 2f
309	testl %esi, %esi
310SRC(1:	movl (%esi), %ebx	)
311SRC(	movl 4(%esi), %edx	)
312	adcl %ebx, %eax
313DST(	movl %ebx, (%edi)	)
314	adcl %edx, %eax
315DST(	movl %edx, 4(%edi)	)
316
317SRC(	movl 8(%esi), %ebx	)
318SRC(	movl 12(%esi), %edx	)
319	adcl %ebx, %eax
320DST(	movl %ebx, 8(%edi)	)
321	adcl %edx, %eax
322DST(	movl %edx, 12(%edi)	)
323
324SRC(	movl 16(%esi), %ebx 	)
325SRC(	movl 20(%esi), %edx	)
326	adcl %ebx, %eax
327DST(	movl %ebx, 16(%edi)	)
328	adcl %edx, %eax
329DST(	movl %edx, 20(%edi)	)
330
331SRC(	movl 24(%esi), %ebx	)
332SRC(	movl 28(%esi), %edx	)
333	adcl %ebx, %eax
334DST(	movl %ebx, 24(%edi)	)
335	adcl %edx, %eax
336DST(	movl %edx, 28(%edi)	)
337
338	lea 32(%esi), %esi
339	lea 32(%edi), %edi
340	dec %ecx
341	jne 1b
342	adcl $0, %eax
3432:	movl FP(%esp), %edx
344	movl %edx, %ecx
345	andl $0x1c, %edx
346	je 4f
347	shrl $2, %edx			# This clears CF
348SRC(3:	movl (%esi), %ebx	)
349	adcl %ebx, %eax
350DST(	movl %ebx, (%edi)	)
351	lea 4(%esi), %esi
352	lea 4(%edi), %edi
353	dec %edx
354	jne 3b
355	adcl $0, %eax
3564:	andl $3, %ecx
357	jz 7f
358	cmpl $2, %ecx
359	jb 5f
360SRC(	movw (%esi), %cx	)
361	leal 2(%esi), %esi
362DST(	movw %cx, (%edi)	)
363	leal 2(%edi), %edi
364	je 6f
365	shll $16,%ecx
366SRC(5:	movb (%esi), %cl	)
367DST(	movb %cl, (%edi)	)
3686:	addl %ecx, %eax
369	adcl $0, %eax
3707:
3715000:
372
373# Exception handler:
374.section .fixup, "ax"
375
3766001:
377	movl ARGBASE+20(%esp), %ebx	# src_err_ptr
378	movl $-EFAULT, (%ebx)
379
380	# zero the complete destination - computing the rest
381	# is too much work
382	movl ARGBASE+8(%esp), %edi	# dst
383	movl ARGBASE+12(%esp), %ecx	# len
384	xorl %eax,%eax
385	rep ; stosb
386
387	jmp 5000b
388
3896002:
390	movl ARGBASE+24(%esp), %ebx	# dst_err_ptr
391	movl $-EFAULT,(%ebx)
392	jmp 5000b
393
394.previous
395
396	popl %ebx
397	popl %esi
398	popl %edi
399	popl %ecx			# equivalent to addl $4,%esp
400	ret
401SYM_FUNC_END(csum_partial_copy_generic)
402
403#else
404
405/* Version for PentiumII/PPro */
406
407#define ROUND1(x) \
408	SRC(movl x(%esi), %ebx	)	;	\
409	addl %ebx, %eax			;	\
410	DST(movl %ebx, x(%edi)	)	;
411
412#define ROUND(x) \
413	SRC(movl x(%esi), %ebx	)	;	\
414	adcl %ebx, %eax			;	\
415	DST(movl %ebx, x(%edi)	)	;
416
417#define ARGBASE 12
418
419SYM_FUNC_START(csum_partial_copy_generic)
420	pushl %ebx
421	pushl %edi
422	pushl %esi
423	movl ARGBASE+4(%esp),%esi	#src
424	movl ARGBASE+8(%esp),%edi	#dst
425	movl ARGBASE+12(%esp),%ecx	#len
426	movl ARGBASE+16(%esp),%eax	#sum
427#	movl %ecx, %edx
428	movl %ecx, %ebx
429	movl %esi, %edx
430	shrl $6, %ecx
431	andl $0x3c, %ebx
432	negl %ebx
433	subl %ebx, %esi
434	subl %ebx, %edi
435	lea  -1(%esi),%edx
436	andl $-32,%edx
437	lea 3f(%ebx,%ebx), %ebx
438	testl %esi, %esi
439	JMP_NOSPEC %ebx
4401:	addl $64,%esi
441	addl $64,%edi
442	SRC(movb -32(%edx),%bl)	; SRC(movb (%edx),%bl)
443	ROUND1(-64) ROUND(-60) ROUND(-56) ROUND(-52)
444	ROUND (-48) ROUND(-44) ROUND(-40) ROUND(-36)
445	ROUND (-32) ROUND(-28) ROUND(-24) ROUND(-20)
446	ROUND (-16) ROUND(-12) ROUND(-8)  ROUND(-4)
4473:	adcl $0,%eax
448	addl $64, %edx
449	dec %ecx
450	jge 1b
4514:	movl ARGBASE+12(%esp),%edx	#len
452	andl $3, %edx
453	jz 7f
454	cmpl $2, %edx
455	jb 5f
456SRC(	movw (%esi), %dx         )
457	leal 2(%esi), %esi
458DST(	movw %dx, (%edi)         )
459	leal 2(%edi), %edi
460	je 6f
461	shll $16,%edx
4625:
463SRC(	movb (%esi), %dl         )
464DST(	movb %dl, (%edi)         )
4656:	addl %edx, %eax
466	adcl $0, %eax
4677:
468.section .fixup, "ax"
4696001:	movl	ARGBASE+20(%esp), %ebx	# src_err_ptr
470	movl $-EFAULT, (%ebx)
471	# zero the complete destination (computing the rest is too much work)
472	movl ARGBASE+8(%esp),%edi	# dst
473	movl ARGBASE+12(%esp),%ecx	# len
474	xorl %eax,%eax
475	rep; stosb
476	jmp 7b
4776002:	movl ARGBASE+24(%esp), %ebx	# dst_err_ptr
478	movl $-EFAULT, (%ebx)
479	jmp  7b
480.previous
481
482	popl %esi
483	popl %edi
484	popl %ebx
485	ret
486SYM_FUNC_END(csum_partial_copy_generic)
487
488#undef ROUND
489#undef ROUND1
490
491#endif
492EXPORT_SYMBOL(csum_partial_copy_generic)
493