2 * arch/x86_64/lib/csum-partial.c
4 * This file contains network checksum routines that are better done
5 * in an architecture-specific manner due to speed.
8 #include <linux/compiler.h>
9 #include <linux/module.h>
10 #include <asm/checksum.h>
12 #define __force_inline inline __attribute__((always_inline))
14 static inline unsigned short from32to16(unsigned a
)
16 unsigned short b
= a
>> 16;
17 asm("addw %w2,%w0\n\t"
25 * Do a 64-bit checksum on an arbitrary memory area.
26 * Returns a 32bit checksum.
28 * This isn't as time critical as it used to be because many NICs
29 * do hardware checksumming these days.
31 * Things tried and found to not make it faster:
33 * Unrolling to an 128 bytes inner loop.
34 * Using interleaving with more registers to break the carry chains.
36 static __force_inline
unsigned do_csum(const unsigned char *buff
, unsigned len
)
39 unsigned long result
= 0;
41 if (unlikely(len
== 0))
43 odd
= 1 & (unsigned long) buff
;
49 count
= len
>> 1; /* nr of 16-bit words.. */
51 if (2 & (unsigned long) buff
) {
52 result
+= *(unsigned short *)buff
;
57 count
>>= 1; /* nr of 32-bit words.. */
61 if (4 & (unsigned long) buff
) {
62 result
+= *(unsigned int *) buff
;
67 count
>>= 1; /* nr of 64-bit words.. */
69 /* main loop using 64byte blocks */
73 asm("addq 0*8(%[src]),%[res]\n\t"
74 "adcq 1*8(%[src]),%[res]\n\t"
75 "adcq 2*8(%[src]),%[res]\n\t"
76 "adcq 3*8(%[src]),%[res]\n\t"
77 "adcq 4*8(%[src]),%[res]\n\t"
78 "adcq 5*8(%[src]),%[res]\n\t"
79 "adcq 6*8(%[src]),%[res]\n\t"
80 "adcq 7*8(%[src]),%[res]\n\t"
83 : [src
] "r" (buff
), [zero
] "r" (zero
),
89 /* last upto 7 8byte blocks */
95 : "m" (*(unsigned long *)buff
),
96 "r" (zero
), "0" (result
));
100 result
= add32_with_carry(result
>>32,
104 result
+= *(unsigned int *) buff
;
109 result
+= *(unsigned short *) buff
;
115 result
= add32_with_carry(result
>>32, result
& 0xffffffff);
117 result
= from32to16(result
);
118 result
= ((result
>> 8) & 0xff) | ((result
& 0xff) << 8);
124 * computes the checksum of a memory block at buff, length len,
125 * and adds in "sum" (32-bit)
127 * returns a 32-bit number suitable for feeding into itself
128 * or csum_tcpudp_magic
130 * this function must be called with even lengths, except
131 * for the last fragment, which may be odd
133 * it's best to have buff aligned on a 64-bit boundary
135 unsigned csum_partial(const unsigned char *buff
, unsigned len
, unsigned sum
)
137 return add32_with_carry(do_csum(buff
, len
), sum
);
140 EXPORT_SYMBOL(csum_partial
);
143 * this routine is used for miscellaneous IP-like checksums, mainly
146 unsigned short ip_compute_csum(unsigned char * buff
, int len
)
148 return csum_fold(csum_partial(buff
,len
,0));