2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * IP/TCP/UDP checksumming routines
8 * Authors: Jorge Cwik, <jorge@laser.satlink.net>
9 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
10 * Tom May, <ftom@netcom.com>
11 * Pentium Pro/II routines:
12 * Alexander Kjeldaas <astor@guardian.no>
13 * Finn Arne Gangstad <finnag@guardian.no>
14 * Lots of code moved from tcp.c and ip.c; see those files
17 * Changes: Ingo Molnar, converted csum_partial_copy() to 2.1 exception
19 * Andi Kleen, add zeroing on error
20 * converted to pure assembler
22 * This program is free software; you can redistribute it and/or
23 * modify it under the terms of the GNU General Public License
24 * as published by the Free Software Foundation; either version
25 * 2 of the License, or (at your option) any later version.
28 #include <asm/errno.h>
31 * computes a partial checksum, e.g. for TCP/UDP fragments
35 unsigned int csum_partial(const unsigned char * buff, int len, unsigned int sum)
42 #ifndef CONFIG_X86_USE_PPRO_CHECKSUM
45 * Experiments with Ethernet and SLIP connections show that buff
46 * is aligned on either a 2-byte or 4-byte boundary. We get at
47 * least a twofold speedup on 486 and Pentium if it is 4-byte aligned.
48 * Fortunately, it is easy to convert 2-byte alignment to 4-byte
49 * alignment for the unrolled loop.
54 movl 20(%esp),%eax # Function arg: unsigned int sum
55 movl 16(%esp),%ecx # Function arg: int len
56 movl 12(%esp),%esi # Function arg: unsigned char *buff
57 testl $2, %esi # Check alignment.
58 jz 2f # Jump if alignment is ok.
59 subl $2, %ecx # Alignment uses up two bytes.
60 jae 1f # Jump if we had at least two bytes.
61 addl $2, %ecx # ecx was < 2. Deal with it.
95 shrl $2, %edx # This clears CF
119 /* Version for PentiumII/PPro */
124 movl 20(%esp),%eax # Function arg: unsigned int sum
125 movl 16(%esp),%ecx # Function arg: int len
126 movl 12(%esp),%esi # Function arg: const unsigned char *buf
138 lea 45f(%ebx,%ebx,2), %ebx
142 # Handle 2-byte-aligned regions
151 movzbl (%esi),%ebx # csumming 1 byte, 2-aligned
156 addw (%esi), %ax # csumming 2 bytes, 2-aligned
161 addl -128(%esi), %eax
162 adcl -124(%esi), %eax
163 adcl -120(%esi), %eax
164 adcl -116(%esi), %eax
165 adcl -112(%esi), %eax
166 adcl -108(%esi), %eax
167 adcl -104(%esi), %eax
168 adcl -100(%esi), %eax
202 # Handle the last 1-3 bytes without jumping
203 notl %ecx # 1->2, 2->1, 3->0, higher bits are masked
204 movl $0xffffff,%ebx # by the shll and shrl instructions
207 andl -128(%esi),%ebx # esi is 4-aligned so should be ok
218 unsigned int csum_partial_copy_generic (const char *src, char *dst,
219 int len, int sum, int *src_err_ptr, int *dst_err_ptr)
223 * Copy from ds while checksumming, otherwise like csum_partial
225 * The macros SRC and DST specify the type of access for the instruction.
226 * thus we can call a custom exception handler for all access types.
228 * FIXME: could someone double-check whether I haven't mixed up some SRC and
229 * DST definitions? It's damn hard to trigger all cases. I hope I got
230 * them all but there's no guarantee.
235 .section __ex_table, "a"; \
236 .long 9999b, 6001f ; \
241 .section __ex_table, "a"; \
242 .long 9999b, 6002f ; \
246 .globl csum_partial_copy_generic_i386
248 #ifndef CONFIG_X86_USE_PPRO_CHECKSUM
253 csum_partial_copy_generic_i386:
258 movl ARGBASE+16(%esp),%eax # sum
259 movl ARGBASE+12(%esp),%ecx # len
260 movl ARGBASE+4(%esp),%esi # src
261 movl ARGBASE+8(%esp),%edi # dst
263 testl $2, %edi # Check alignment.
264 jz 2f # Jump if alignment is ok.
265 subl $2, %ecx # Alignment uses up two bytes.
266 jae 1f # Jump if we had at least two bytes.
267 addl $2, %ecx # ecx was < 2. Deal with it.
269 SRC(1: movw (%esi), %bx )
271 DST( movw %bx, (%edi) )
280 SRC(1: movl (%esi), %ebx )
281 SRC( movl 4(%esi), %edx )
283 DST( movl %ebx, (%edi) )
285 DST( movl %edx, 4(%edi) )
287 SRC( movl 8(%esi), %ebx )
288 SRC( movl 12(%esi), %edx )
290 DST( movl %ebx, 8(%edi) )
292 DST( movl %edx, 12(%edi) )
294 SRC( movl 16(%esi), %ebx )
295 SRC( movl 20(%esi), %edx )
297 DST( movl %ebx, 16(%edi) )
299 DST( movl %edx, 20(%edi) )
301 SRC( movl 24(%esi), %ebx )
302 SRC( movl 28(%esi), %edx )
304 DST( movl %ebx, 24(%edi) )
306 DST( movl %edx, 28(%edi) )
313 2: movl FP(%esp), %edx
317 shrl $2, %edx # This clears CF
318 SRC(3: movl (%esi), %ebx )
320 DST( movl %ebx, (%edi) )
330 SRC( movw (%esi), %cx )
332 DST( movw %cx, (%edi) )
336 SRC(5: movb (%esi), %cl )
337 DST( movb %cl, (%edi) )
344 .section .fixup, "ax"
347 movl ARGBASE+20(%esp), %ebx # src_err_ptr
348 movl $-EFAULT, (%ebx)
350 # zero the complete destination - computing the rest
352 movl ARGBASE+8(%esp), %edi # dst
353 movl ARGBASE+12(%esp), %ecx # len
360 movl ARGBASE+24(%esp), %ebx # dst_err_ptr
369 popl %ecx # equivalent to addl $4,%esp
374 /* Version for PentiumII/PPro */
377 SRC(movl x(%esi), %ebx ) ; \
379 DST(movl %ebx, x(%edi) ) ;
382 SRC(movl x(%esi), %ebx ) ; \
384 DST(movl %ebx, x(%edi) ) ;
388 csum_partial_copy_generic_i386:
392 movl ARGBASE+4(%esp),%esi #src
393 movl ARGBASE+8(%esp),%edi #dst
394 movl ARGBASE+12(%esp),%ecx #len
395 movl ARGBASE+16(%esp),%eax #sum
406 lea 3f(%ebx,%ebx), %ebx
411 SRC(movb -32(%edx),%bl) ; SRC(movb (%edx),%bl)
412 ROUND1(-64) ROUND(-60) ROUND(-56) ROUND(-52)
413 ROUND (-48) ROUND(-44) ROUND(-40) ROUND(-36)
414 ROUND (-32) ROUND(-28) ROUND(-24) ROUND(-20)
415 ROUND (-16) ROUND(-12) ROUND(-8) ROUND(-4)
420 4: movl ARGBASE+12(%esp),%edx #len
425 SRC( movw (%esi), %dx )
427 DST( movw %dx, (%edi) )
432 SRC( movb (%esi), %dl )
433 DST( movb %dl, (%edi) )
437 .section .fixup, "ax"
438 6001: movl ARGBASE+20(%esp), %ebx # src_err_ptr
439 movl $-EFAULT, (%ebx)
440 # zero the complete destination (computing the rest is too much work)
441 movl ARGBASE+8(%esp),%edi # dst
442 movl ARGBASE+12(%esp),%ecx # len
446 6002: movl ARGBASE+24(%esp), %ebx # dst_err_ptr
447 movl $-EFAULT, (%ebx)