Linux-2.6.12-rc2
[linux-2.6] / arch / alpha / lib / checksum.c
1 /*
2  * arch/alpha/lib/checksum.c
3  *
4  * This file contains network checksum routines that are better done
5  * in an architecture-specific manner due to speed..
6  * Comments in other versions indicate that the algorithms are from RFC1071
7  *
8  * accellerated versions (and 21264 assembly versions ) contributed by
9  *      Rick Gorton     <rick.gorton@alpha-processor.com>
10  */
11  
12 #include <linux/module.h>
13 #include <linux/string.h>
14
15 #include <asm/byteorder.h>
16
17 static inline unsigned short from64to16(unsigned long x)
18 {
19         /* Using extract instructions is a bit more efficient
20            than the original shift/bitmask version.  */
21
22         union {
23                 unsigned long   ul;
24                 unsigned int    ui[2];
25                 unsigned short  us[4];
26         } in_v, tmp_v, out_v;
27
28         in_v.ul = x;
29         tmp_v.ul = (unsigned long) in_v.ui[0] + (unsigned long) in_v.ui[1];
30
31         /* Since the bits of tmp_v.sh[3] are going to always be zero,
32            we don't have to bother to add that in.  */
33         out_v.ul = (unsigned long) tmp_v.us[0] + (unsigned long) tmp_v.us[1]
34                         + (unsigned long) tmp_v.us[2];
35
36         /* Similarly, out_v.us[2] is always zero for the final add.  */
37         return out_v.us[0] + out_v.us[1];
38 }
39
40 /*
41  * computes the checksum of the TCP/UDP pseudo-header
42  * returns a 16-bit checksum, already complemented.
43  */
44 unsigned short int csum_tcpudp_magic(unsigned long saddr,
45                                    unsigned long daddr,
46                                    unsigned short len,
47                                    unsigned short proto,
48                                    unsigned int sum)
49 {
50         return ~from64to16(saddr + daddr + sum +
51                 ((unsigned long) ntohs(len) << 16) +
52                 ((unsigned long) proto << 8));
53 }
54
55 unsigned int csum_tcpudp_nofold(unsigned long saddr,
56                                    unsigned long daddr,
57                                    unsigned short len,
58                                    unsigned short proto,
59                                    unsigned int sum)
60 {
61         unsigned long result;
62
63         result = (saddr + daddr + sum +
64                   ((unsigned long) ntohs(len) << 16) +
65                   ((unsigned long) proto << 8));
66
67         /* Fold down to 32-bits so we don't lose in the typedef-less 
68            network stack.  */
69         /* 64 to 33 */
70         result = (result & 0xffffffff) + (result >> 32);
71         /* 33 to 32 */
72         result = (result & 0xffffffff) + (result >> 32);
73         return result;
74 }
75
76 /*
77  * Do a 64-bit checksum on an arbitrary memory area..
78  *
79  * This isn't a great routine, but it's not _horrible_ either. The
80  * inner loop could be unrolled a bit further, and there are better
81  * ways to do the carry, but this is reasonable.
82  */
83 static inline unsigned long do_csum(const unsigned char * buff, int len)
84 {
85         int odd, count;
86         unsigned long result = 0;
87
88         if (len <= 0)
89                 goto out;
90         odd = 1 & (unsigned long) buff;
91         if (odd) {
92                 result = *buff << 8;
93                 len--;
94                 buff++;
95         }
96         count = len >> 1;               /* nr of 16-bit words.. */
97         if (count) {
98                 if (2 & (unsigned long) buff) {
99                         result += *(unsigned short *) buff;
100                         count--;
101                         len -= 2;
102                         buff += 2;
103                 }
104                 count >>= 1;            /* nr of 32-bit words.. */
105                 if (count) {
106                         if (4 & (unsigned long) buff) {
107                                 result += *(unsigned int *) buff;
108                                 count--;
109                                 len -= 4;
110                                 buff += 4;
111                         }
112                         count >>= 1;    /* nr of 64-bit words.. */
113                         if (count) {
114                                 unsigned long carry = 0;
115                                 do {
116                                         unsigned long w = *(unsigned long *) buff;
117                                         count--;
118                                         buff += 8;
119                                         result += carry;
120                                         result += w;
121                                         carry = (w > result);
122                                 } while (count);
123                                 result += carry;
124                                 result = (result & 0xffffffff) + (result >> 32);
125                         }
126                         if (len & 4) {
127                                 result += *(unsigned int *) buff;
128                                 buff += 4;
129                         }
130                 }
131                 if (len & 2) {
132                         result += *(unsigned short *) buff;
133                         buff += 2;
134                 }
135         }
136         if (len & 1)
137                 result += *buff;
138         result = from64to16(result);
139         if (odd)
140                 result = ((result >> 8) & 0xff) | ((result & 0xff) << 8);
141 out:
142         return result;
143 }
144
145 /*
146  *      This is a version of ip_compute_csum() optimized for IP headers,
147  *      which always checksum on 4 octet boundaries.
148  */
149 unsigned short ip_fast_csum(unsigned char * iph, unsigned int ihl)
150 {
151         return ~do_csum(iph,ihl*4);
152 }
153
154 /*
155  * computes the checksum of a memory block at buff, length len,
156  * and adds in "sum" (32-bit)
157  *
158  * returns a 32-bit number suitable for feeding into itself
159  * or csum_tcpudp_magic
160  *
161  * this function must be called with even lengths, except
162  * for the last fragment, which may be odd
163  *
164  * it's best to have buff aligned on a 32-bit boundary
165  */
166 unsigned int csum_partial(const unsigned char * buff, int len, unsigned int sum)
167 {
168         unsigned long result = do_csum(buff, len);
169
170         /* add in old sum, and carry.. */
171         result += sum;
172         /* 32+c bits -> 32 bits */
173         result = (result & 0xffffffff) + (result >> 32);
174         return result;
175 }
176
177 EXPORT_SYMBOL(csum_partial);
178
179 /*
180  * this routine is used for miscellaneous IP-like checksums, mainly
181  * in icmp.c
182  */
183 unsigned short ip_compute_csum(unsigned char * buff, int len)
184 {
185         return ~from64to16(do_csum(buff,len));
186 }