ftp://ftp.kernel.org/pub/linux/kernel/v2.6/linux-2.6.6.tar.bz2
[linux-2.6.git] / arch / alpha / lib / csum_partial_copy.c
1 /*
2  * csum_partial_copy - do IP checksumming and copy
3  *
4  * (C) Copyright 1996 Linus Torvalds
5  * accellerated versions (and 21264 assembly versions ) contributed by
6  *      Rick Gorton     <rick.gorton@alpha-processor.com>
7  *
8  * Don't look at this too closely - you'll go mad. The things
9  * we do for performance..
10  */
11
12 #include <linux/types.h>
13 #include <linux/string.h>
14 #include <asm/uaccess.h>
15
16
17 #define ldq_u(x,y) \
18 __asm__ __volatile__("ldq_u %0,%1":"=r" (x):"m" (*(const unsigned long *)(y)))
19
20 #define stq_u(x,y) \
21 __asm__ __volatile__("stq_u %1,%0":"=m" (*(unsigned long *)(y)):"r" (x))
22
23 #define extql(x,y,z) \
24 __asm__ __volatile__("extql %1,%2,%0":"=r" (z):"r" (x),"r" (y))
25
26 #define extqh(x,y,z) \
27 __asm__ __volatile__("extqh %1,%2,%0":"=r" (z):"r" (x),"r" (y))
28
29 #define mskql(x,y,z) \
30 __asm__ __volatile__("mskql %1,%2,%0":"=r" (z):"r" (x),"r" (y))
31
32 #define mskqh(x,y,z) \
33 __asm__ __volatile__("mskqh %1,%2,%0":"=r" (z):"r" (x),"r" (y))
34
35 #define insql(x,y,z) \
36 __asm__ __volatile__("insql %1,%2,%0":"=r" (z):"r" (x),"r" (y))
37
38 #define insqh(x,y,z) \
39 __asm__ __volatile__("insqh %1,%2,%0":"=r" (z):"r" (x),"r" (y))
40
41
42 #define __get_user_u(x,ptr)                             \
43 ({                                                      \
44         long __guu_err;                                 \
45         __asm__ __volatile__(                           \
46         "1:     ldq_u %0,%2\n"                          \
47         "2:\n"                                          \
48         ".section __ex_table,\"a\"\n"                   \
49         "       .long 1b - .\n"                         \
50         "       lda %0,2b-1b(%1)\n"                     \
51         ".previous"                                     \
52                 : "=r"(x), "=r"(__guu_err)              \
53                 : "m"(__m(ptr)), "1"(0));               \
54         __guu_err;                                      \
55 })
56
57 #define __put_user_u(x,ptr)                             \
58 ({                                                      \
59         long __puu_err;                                 \
60         __asm__ __volatile__(                           \
61         "1:     stq_u %2,%1\n"                          \
62         "2:\n"                                          \
63         ".section __ex_table,\"a\"\n"                   \
64         "       .long 1b - ."                           \
65         "       lda $31,2b-1b(%0)\n"                    \
66         ".previous"                                     \
67                 : "=r"(__puu_err)                       \
68                 : "m"(__m(addr)), "rJ"(x), "0"(0));     \
69         __puu_err;                                      \
70 })
71
72
73 static inline unsigned short from64to16(unsigned long x)
74 {
75         /* Using extract instructions is a bit more efficient
76            than the original shift/bitmask version.  */
77
78         union {
79                 unsigned long   ul;
80                 unsigned int    ui[2];
81                 unsigned short  us[4];
82         } in_v, tmp_v, out_v;
83
84         in_v.ul = x;
85         tmp_v.ul = (unsigned long) in_v.ui[0] + (unsigned long) in_v.ui[1];
86
87         /* Since the bits of tmp_v.sh[3] are going to always be zero,
88            we don't have to bother to add that in.  */
89         out_v.ul = (unsigned long) tmp_v.us[0] + (unsigned long) tmp_v.us[1]
90                         + (unsigned long) tmp_v.us[2];
91
92         /* Similarly, out_v.us[2] is always zero for the final add.  */
93         return out_v.us[0] + out_v.us[1];
94 }
95
96
97
98 /*
99  * Ok. This isn't fun, but this is the EASY case.
100  */
101 static inline unsigned long
102 csum_partial_cfu_aligned(const unsigned long *src, unsigned long *dst,
103                          long len, unsigned long checksum,
104                          int *errp)
105 {
106         unsigned long carry = 0;
107         int err = 0;
108
109         while (len >= 0) {
110                 unsigned long word;
111                 err |= __get_user(word, src);
112                 checksum += carry;
113                 src++;
114                 checksum += word;
115                 len -= 8;
116                 carry = checksum < word;
117                 *dst = word;
118                 dst++;
119         }
120         len += 8;
121         checksum += carry;
122         if (len) {
123                 unsigned long word, tmp;
124                 err |= __get_user(word, src);
125                 tmp = *dst;
126                 mskql(word, len, word);
127                 checksum += word;
128                 mskqh(tmp, len, tmp);
129                 carry = checksum < word;
130                 *dst = word | tmp;
131                 checksum += carry;
132         }
133         if (err) *errp = err;
134         return checksum;
135 }
136
137 /*
138  * This is even less fun, but this is still reasonably
139  * easy.
140  */
141 static inline unsigned long
142 csum_partial_cfu_dest_aligned(const unsigned long *src, unsigned long *dst,
143                               unsigned long soff,
144                               long len, unsigned long checksum,
145                               int *errp)
146 {
147         unsigned long first;
148         unsigned long word, carry;
149         unsigned long lastsrc = 7+len+(unsigned long)src;
150         int err = 0;
151
152         err |= __get_user_u(first,src);
153         carry = 0;
154         while (len >= 0) {
155                 unsigned long second;
156
157                 err |= __get_user_u(second, src+1);
158                 extql(first, soff, word);
159                 len -= 8;
160                 src++;
161                 extqh(second, soff, first);
162                 checksum += carry;
163                 word |= first;
164                 first = second;
165                 checksum += word;
166                 *dst = word;
167                 dst++;
168                 carry = checksum < word;
169         }
170         len += 8;
171         checksum += carry;
172         if (len) {
173                 unsigned long tmp;
174                 unsigned long second;
175                 err |= __get_user_u(second, lastsrc);
176                 tmp = *dst;
177                 extql(first, soff, word);
178                 extqh(second, soff, first);
179                 word |= first;
180                 mskql(word, len, word);
181                 checksum += word;
182                 mskqh(tmp, len, tmp);
183                 carry = checksum < word;
184                 *dst = word | tmp;
185                 checksum += carry;
186         }
187         if (err) *errp = err;
188         return checksum;
189 }
190
191 /*
192  * This is slightly less fun than the above..
193  */
194 static inline unsigned long
195 csum_partial_cfu_src_aligned(const unsigned long *src, unsigned long *dst,
196                              unsigned long doff,
197                              long len, unsigned long checksum,
198                              unsigned long partial_dest,
199                              int *errp)
200 {
201         unsigned long carry = 0;
202         unsigned long word;
203         unsigned long second_dest;
204         int err = 0;
205
206         mskql(partial_dest, doff, partial_dest);
207         while (len >= 0) {
208                 err |= __get_user(word, src);
209                 len -= 8;
210                 insql(word, doff, second_dest);
211                 checksum += carry;
212                 stq_u(partial_dest | second_dest, dst);
213                 src++;
214                 checksum += word;
215                 insqh(word, doff, partial_dest);
216                 carry = checksum < word;
217                 dst++;
218         }
219         len += 8;
220         if (len) {
221                 checksum += carry;
222                 err |= __get_user(word, src);
223                 mskql(word, len, word);
224                 len -= 8;
225                 checksum += word;
226                 insql(word, doff, second_dest);
227                 len += doff;
228                 carry = checksum < word;
229                 partial_dest |= second_dest;
230                 if (len >= 0) {
231                         stq_u(partial_dest, dst);
232                         if (!len) goto out;
233                         dst++;
234                         insqh(word, doff, partial_dest);
235                 }
236                 doff = len;
237         }
238         ldq_u(second_dest, dst);
239         mskqh(second_dest, doff, second_dest);
240         stq_u(partial_dest | second_dest, dst);
241 out:
242         checksum += carry;
243         if (err) *errp = err;
244         return checksum;
245 }
246
247 /*
248  * This is so totally un-fun that it's frightening. Don't
249  * look at this too closely, you'll go blind.
250  */
251 static inline unsigned long
252 csum_partial_cfu_unaligned(const unsigned long * src, unsigned long * dst,
253                            unsigned long soff, unsigned long doff,
254                            long len, unsigned long checksum,
255                            unsigned long partial_dest,
256                            int *errp)
257 {
258         unsigned long carry = 0;
259         unsigned long first;
260         unsigned long lastsrc;
261         int err = 0;
262
263         err |= __get_user_u(first, src);
264         lastsrc = 7+len+(unsigned long)src;
265         mskql(partial_dest, doff, partial_dest);
266         while (len >= 0) {
267                 unsigned long second, word;
268                 unsigned long second_dest;
269
270                 err |= __get_user_u(second, src+1);
271                 extql(first, soff, word);
272                 checksum += carry;
273                 len -= 8;
274                 extqh(second, soff, first);
275                 src++;
276                 word |= first;
277                 first = second;
278                 insql(word, doff, second_dest);
279                 checksum += word;
280                 stq_u(partial_dest | second_dest, dst);
281                 carry = checksum < word;
282                 insqh(word, doff, partial_dest);
283                 dst++;
284         }
285         len += doff;
286         checksum += carry;
287         if (len >= 0) {
288                 unsigned long second, word;
289                 unsigned long second_dest;
290
291                 err |= __get_user_u(second, lastsrc);
292                 extql(first, soff, word);
293                 extqh(second, soff, first);
294                 word |= first;
295                 first = second;
296                 mskql(word, len-doff, word);
297                 checksum += word;
298                 insql(word, doff, second_dest);
299                 carry = checksum < word;
300                 stq_u(partial_dest | second_dest, dst);
301                 if (len) {
302                         ldq_u(second_dest, dst+1);
303                         insqh(word, doff, partial_dest);
304                         mskqh(second_dest, len, second_dest);
305                         stq_u(partial_dest | second_dest, dst+1);
306                 }
307                 checksum += carry;
308         } else {
309                 unsigned long second, word;
310                 unsigned long second_dest;
311
312                 err |= __get_user_u(second, lastsrc);
313                 extql(first, soff, word);
314                 extqh(second, soff, first);
315                 word |= first;
316                 ldq_u(second_dest, dst);
317                 mskql(word, len-doff, word);
318                 checksum += word;
319                 mskqh(second_dest, len, second_dest);
320                 carry = checksum < word;
321                 insql(word, doff, word);
322                 stq_u(partial_dest | word | second_dest, dst);
323                 checksum += carry;
324         }
325         if (err) *errp = err;
326         return checksum;
327 }
328
329 static unsigned int
330 do_csum_partial_copy_from_user(const char *src, char *dst, int len,
331                                unsigned int sum, int *errp)
332 {
333         unsigned long checksum = (unsigned) sum;
334         unsigned long soff = 7 & (unsigned long) src;
335         unsigned long doff = 7 & (unsigned long) dst;
336
337         if (len) {
338                 if (!doff) {
339                         if (!soff)
340                                 checksum = csum_partial_cfu_aligned(
341                                         (const unsigned long *) src,
342                                         (unsigned long *) dst,
343                                         len-8, checksum, errp);
344                         else
345                                 checksum = csum_partial_cfu_dest_aligned(
346                                         (const unsigned long *) src,
347                                         (unsigned long *) dst,
348                                         soff, len-8, checksum, errp);
349                 } else {
350                         unsigned long partial_dest;
351                         ldq_u(partial_dest, dst);
352                         if (!soff)
353                                 checksum = csum_partial_cfu_src_aligned(
354                                         (const unsigned long *) src,
355                                         (unsigned long *) dst,
356                                         doff, len-8, checksum,
357                                         partial_dest, errp);
358                         else
359                                 checksum = csum_partial_cfu_unaligned(
360                                         (const unsigned long *) src,
361                                         (unsigned long *) dst,
362                                         soff, doff, len-8, checksum,
363                                         partial_dest, errp);
364                 }
365                 checksum = from64to16 (checksum);
366         }
367         return checksum;
368 }
369
370 unsigned int
371 csum_partial_copy_from_user(const char *src, char *dst, int len,
372                             unsigned int sum, int *errp)
373 {
374         if (!access_ok(src, len, VERIFY_READ)) {
375                 *errp = -EFAULT;
376                 memset(dst, 0, len);
377                 return sum;
378         }
379
380         return do_csum_partial_copy_from_user(src, dst, len, sum, errp);
381 }
382
383 unsigned int
384 csum_partial_copy_nocheck(const char *src, char *dst, int len, unsigned int sum)
385 {
386         return do_csum_partial_copy_from_user(src, dst, len, sum, NULL);
387 }