1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * User address space access functions.
4 * The non-inlined parts of asm-cris/uaccess.h are here.
5 *
6 * Copyright (C) 2000, Axis Communications AB.
7 *
8 * Written by Hans-Peter Nilsson.
9 * Pieces used from memcpy, originally by Kenny Ranerup long time ago.
10 */
11
12 #include <linux/uaccess.h>
13
14 /* Asm:s have been tweaked (within the domain of correctness) to give
15 satisfactory results for "gcc version 2.96 20000427 (experimental)".
16
17 Check regularly...
18
19 Note that the PC saved at a bus-fault is the address *after* the
20 faulting instruction, which means the branch-target for instructions in
21 delay-slots for taken branches. Note also that the postincrement in
22 the instruction is performed regardless of bus-fault; the register is
23 seen updated in fault handlers.
24
25 Oh, and on the code formatting issue, to whomever feels like "fixing
26 it" to Conformity: I'm too "lazy", but why don't you go ahead and "fix"
27 string.c too. I just don't think too many people will hack this file
28 for the code format to be an issue. */
29
30
31 /* Copy to userspace. This is based on the memcpy used for
32 kernel-to-kernel copying; see "string.c". */
33
__copy_user(void __user * pdst,const void * psrc,unsigned long pn)34 unsigned long __copy_user(void __user *pdst, const void *psrc, unsigned long pn)
35 {
36 /* We want the parameters put in special registers.
37 Make sure the compiler is able to make something useful of this.
38 As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop).
39
40 FIXME: Comment for old gcc version. Check.
41 If gcc was alright, it really would need no temporaries, and no
42 stack space to save stuff on. */
43
44 register char *dst __asm__ ("r13") = pdst;
45 register const char *src __asm__ ("r11") = psrc;
46 register int n __asm__ ("r12") = pn;
47 register int retn __asm__ ("r10") = 0;
48
49
50 /* When src is aligned but not dst, this makes a few extra needless
51 cycles. I believe it would take as many to check that the
52 re-alignment was unnecessary. */
53 if (((unsigned long) dst & 3) != 0
54 /* Don't align if we wouldn't copy more than a few bytes; so we
55 don't have to check further for overflows. */
56 && n >= 3)
57 {
58 if ((unsigned long) dst & 1)
59 {
60 __asm_copy_to_user_1 (dst, src, retn);
61 n--;
62 }
63
64 if ((unsigned long) dst & 2)
65 {
66 __asm_copy_to_user_2 (dst, src, retn);
67 n -= 2;
68 }
69 }
70
71 /* Decide which copying method to use. */
72 if (n >= 44*2) /* Break even between movem and
73 move16 is at 38.7*2, but modulo 44. */
74 {
75 /* For large copies we use 'movem'. */
76
77 /* It is not optimal to tell the compiler about clobbering any
78 registers; that will move the saving/restoring of those registers
79 to the function prologue/epilogue, and make non-movem sizes
80 suboptimal.
81
82 This method is not foolproof; it assumes that the "asm reg"
83 declarations at the beginning of the function really are used
84 here (beware: they may be moved to temporary registers).
85 This way, we do not have to save/move the registers around into
86 temporaries; we can safely use them straight away.
87
88 If you want to check that the allocation was right; then
89 check the equalities in the first comment. It should say
90 "r13=r13, r11=r11, r12=r12". */
91 __asm__ volatile ("\
92 .ifnc %0%1%2%3,$r13$r11$r12$r10 \n\
93 .err \n\
94 .endif \n\
95 \n\
96 ;; Save the registers we'll use in the movem process \n\
97 ;; on the stack. \n\
98 subq 11*4,$sp \n\
99 movem $r10,[$sp] \n\
100 \n\
101 ;; Now we've got this: \n\
102 ;; r11 - src \n\
103 ;; r13 - dst \n\
104 ;; r12 - n \n\
105 \n\
106 ;; Update n for the first loop \n\
107 subq 44,$r12 \n\
108 \n\
109 ; Since the noted PC of a faulting instruction in a delay-slot of a taken \n\
110 ; branch, is that of the branch target, we actually point at the from-movem \n\
111 ; for this case. There is no ambiguity here; if there was a fault in that \n\
112 ; instruction (meaning a kernel oops), the faulted PC would be the address \n\
113 ; after *that* movem. \n\
114 \n\
115 0: \n\
116 movem [$r11+],$r10 \n\
117 subq 44,$r12 \n\
118 bge 0b \n\
119 movem $r10,[$r13+] \n\
120 1: \n\
121 addq 44,$r12 ;; compensate for last loop underflowing n \n\
122 \n\
123 ;; Restore registers from stack \n\
124 movem [$sp+],$r10 \n\
125 2: \n\
126 .section .fixup,\"ax\" \n\
127 \n\
128 ; To provide a correct count in r10 of bytes that failed to be copied, \n\
129 ; we jump back into the loop if the loop-branch was taken. There is no \n\
130 ; performance penalty for sany use; the program will segfault soon enough.\n\
131 \n\
132 3: \n\
133 move.d [$sp],$r10 \n\
134 addq 44,$r10 \n\
135 move.d $r10,[$sp] \n\
136 jump 0b \n\
137 4: \n\
138 movem [$sp+],$r10 \n\
139 addq 44,$r10 \n\
140 addq 44,$r12 \n\
141 jump 2b \n\
142 \n\
143 .previous \n\
144 .section __ex_table,\"a\" \n\
145 .dword 0b,3b \n\
146 .dword 1b,4b \n\
147 .previous"
148
149 /* Outputs */ : "=r" (dst), "=r" (src), "=r" (n), "=r" (retn)
150 /* Inputs */ : "0" (dst), "1" (src), "2" (n), "3" (retn));
151
152 }
153
154 /* Either we directly start copying, using dword copying in a loop, or
155 we copy as much as possible with 'movem' and then the last block (<44
156 bytes) is copied here. This will work since 'movem' will have
157 updated SRC, DST and N. */
158
159 while (n >= 16)
160 {
161 __asm_copy_to_user_16 (dst, src, retn);
162 n -= 16;
163 }
164
165 /* Having a separate by-four loops cuts down on cache footprint.
166 FIXME: Test with and without; increasing switch to be 0..15. */
167 while (n >= 4)
168 {
169 __asm_copy_to_user_4 (dst, src, retn);
170 n -= 4;
171 }
172
173 switch (n)
174 {
175 case 0:
176 break;
177 case 1:
178 __asm_copy_to_user_1 (dst, src, retn);
179 break;
180 case 2:
181 __asm_copy_to_user_2 (dst, src, retn);
182 break;
183 case 3:
184 __asm_copy_to_user_3 (dst, src, retn);
185 break;
186 }
187
188 return retn;
189 }
190 EXPORT_SYMBOL(__copy_user);
191
192 /* Copy from user to kernel. The return-value is the number of bytes that were
193 inaccessible. */
194
__copy_user_in(void * pdst,const void __user * psrc,unsigned long pn)195 unsigned long __copy_user_in(void *pdst, const void __user *psrc,
196 unsigned long pn)
197 {
198 /* We want the parameters put in special registers.
199 Make sure the compiler is able to make something useful of this.
200 As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop).
201
202 FIXME: Comment for old gcc version. Check.
203 If gcc was alright, it really would need no temporaries, and no
204 stack space to save stuff on. */
205
206 register char *dst __asm__ ("r13") = pdst;
207 register const char *src __asm__ ("r11") = psrc;
208 register int n __asm__ ("r12") = pn;
209 register int retn __asm__ ("r10") = 0;
210
211 /* The best reason to align src is that we then know that a read-fault
212 was for aligned bytes; there's no 1..3 remaining good bytes to
213 pickle. */
214 if (((unsigned long) src & 3) != 0)
215 {
216 if (((unsigned long) src & 1) && n != 0)
217 {
218 __asm_copy_from_user_1 (dst, src, retn);
219 n--;
220 if (retn)
221 goto exception;
222 }
223
224 if (((unsigned long) src & 2) && n >= 2)
225 {
226 __asm_copy_from_user_2 (dst, src, retn);
227 n -= 2;
228 if (retn)
229 goto exception;
230 }
231 }
232
233 /* Decide which copying method to use. */
234 if (n >= 44*2) /* Break even between movem and
235 move16 is at 38.7*2, but modulo 44.
236 FIXME: We use move4 now. */
237 {
238 /* For large copies we use 'movem' */
239
240 /* It is not optimal to tell the compiler about clobbering any
241 registers; that will move the saving/restoring of those registers
242 to the function prologue/epilogue, and make non-movem sizes
243 suboptimal.
244
245 This method is not foolproof; it assumes that the "asm reg"
246 declarations at the beginning of the function really are used
247 here (beware: they may be moved to temporary registers).
248 This way, we do not have to save/move the registers around into
249 temporaries; we can safely use them straight away.
250
251 If you want to check that the allocation was right; then
252 check the equalities in the first comment. It should say
253 "r13=r13, r11=r11, r12=r12" */
254 __asm__ volatile ("\n\
255 .ifnc %0%1%2%3,$r13$r11$r12$r10 \n\
256 .err \n\
257 .endif \n\
258 \n\
259 ;; Save the registers we'll use in the movem process \n\
260 ;; on the stack. \n\
261 subq 11*4,$sp \n\
262 movem $r10,[$sp] \n\
263 \n\
264 ;; Now we've got this: \n\
265 ;; r11 - src \n\
266 ;; r13 - dst \n\
267 ;; r12 - n \n\
268 \n\
269 ;; Update n for the first loop \n\
270 subq 44,$r12 \n\
271 0: \n\
272 movem [$r11+],$r10 \n\
273 1: \n\
274 subq 44,$r12 \n\
275 bge 0b \n\
276 movem $r10,[$r13+] \n\
277 \n\
278 addq 44,$r12 ;; compensate for last loop underflowing n \n\
279 \n\
280 ;; Restore registers from stack \n\
281 movem [$sp+],$r10 \n\
282 4: \n\
283 .section .fixup,\"ax\" \n\
284 \n\
285 ;; Do not jump back into the loop if we fail. For some uses, we get a \n\
286 ;; page fault somewhere on the line. Without checking for page limits, \n\
287 ;; we don't know where, but we need to copy accurately and keep an \n\
288 ;; accurate count; not just clear the whole line. To do that, we fall \n\
289 ;; down in the code below, proceeding with smaller amounts. It should \n\
290 ;; be kept in mind that we have to cater to code like what at one time \n\
291 ;; was in fs/super.c: \n\
292 ;; i = size - copy_from_user((void *)page, data, size); \n\
293 ;; which would cause repeated faults while clearing the remainder of \n\
294 ;; the SIZE bytes at PAGE after the first fault. \n\
295 ;; A caveat here is that we must not fall through from a failing page \n\
296 ;; to a valid page. \n\
297 \n\
298 3: \n\
299 movem [$sp+],$r10 \n\
300 addq 44,$r12 ;; Get back count before faulting point. \n\
301 subq 44,$r11 ;; Get back pointer to faulting movem-line. \n\
302 jump 4b ;; Fall through, pretending the fault didn't happen.\n\
303 \n\
304 .previous \n\
305 .section __ex_table,\"a\" \n\
306 .dword 1b,3b \n\
307 .previous"
308
309 /* Outputs */ : "=r" (dst), "=r" (src), "=r" (n), "=r" (retn)
310 /* Inputs */ : "0" (dst), "1" (src), "2" (n), "3" (retn));
311
312 }
313
314 /* Either we directly start copying here, using dword copying in a loop,
315 or we copy as much as possible with 'movem' and then the last block
316 (<44 bytes) is copied here. This will work since 'movem' will have
317 updated src, dst and n. (Except with failing src.)
318
319 Since we want to keep src accurate, we can't use
320 __asm_copy_from_user_N with N != (1, 2, 4); it updates dst and
321 retn, but not src (by design; it's value is ignored elsewhere). */
322
323 while (n >= 4)
324 {
325 __asm_copy_from_user_4 (dst, src, retn);
326 n -= 4;
327
328 if (retn)
329 goto exception;
330 }
331
332 /* If we get here, there were no memory read faults. */
333 switch (n)
334 {
335 /* These copies are at least "naturally aligned" (so we don't have
336 to check each byte), due to the src alignment code before the
337 movem loop. The *_3 case *will* get the correct count for retn. */
338 case 0:
339 /* This case deliberately left in (if you have doubts check the
340 generated assembly code). */
341 break;
342 case 1:
343 __asm_copy_from_user_1 (dst, src, retn);
344 break;
345 case 2:
346 __asm_copy_from_user_2 (dst, src, retn);
347 break;
348 case 3:
349 __asm_copy_from_user_3 (dst, src, retn);
350 break;
351 }
352
353 /* If we get here, retn correctly reflects the number of failing
354 bytes. */
355 return retn;
356
357 exception:
358 return retn + n;
359 }
360 EXPORT_SYMBOL(__copy_user_in);
361
362 /* Zero userspace. */
__do_clear_user(void __user * pto,unsigned long pn)363 unsigned long __do_clear_user(void __user *pto, unsigned long pn)
364 {
365 /* We want the parameters put in special registers.
366 Make sure the compiler is able to make something useful of this.
367 As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop).
368
369 FIXME: Comment for old gcc version. Check.
370 If gcc was alright, it really would need no temporaries, and no
371 stack space to save stuff on. */
372
373 register char *dst __asm__ ("r13") = pto;
374 register int n __asm__ ("r12") = pn;
375 register int retn __asm__ ("r10") = 0;
376
377
378 if (((unsigned long) dst & 3) != 0
379 /* Don't align if we wouldn't copy more than a few bytes. */
380 && n >= 3)
381 {
382 if ((unsigned long) dst & 1)
383 {
384 __asm_clear_1 (dst, retn);
385 n--;
386 }
387
388 if ((unsigned long) dst & 2)
389 {
390 __asm_clear_2 (dst, retn);
391 n -= 2;
392 }
393 }
394
395 /* Decide which copying method to use.
396 FIXME: This number is from the "ordinary" kernel memset. */
397 if (n >= (1*48))
398 {
399 /* For large clears we use 'movem' */
400
401 /* It is not optimal to tell the compiler about clobbering any
402 call-saved registers; that will move the saving/restoring of
403 those registers to the function prologue/epilogue, and make
404 non-movem sizes suboptimal.
405
406 This method is not foolproof; it assumes that the "asm reg"
407 declarations at the beginning of the function really are used
408 here (beware: they may be moved to temporary registers).
409 This way, we do not have to save/move the registers around into
410 temporaries; we can safely use them straight away.
411
412 If you want to check that the allocation was right; then
413 check the equalities in the first comment. It should say
414 something like "r13=r13, r11=r11, r12=r12". */
415 __asm__ volatile ("\n\
416 .ifnc %0%1%2,$r13$r12$r10 \n\
417 .err \n\
418 .endif \n\
419 \n\
420 ;; Save the registers we'll clobber in the movem process \n\
421 ;; on the stack. Don't mention them to gcc, it will only be \n\
422 ;; upset. \n\
423 subq 11*4,$sp \n\
424 movem $r10,[$sp] \n\
425 \n\
426 clear.d $r0 \n\
427 clear.d $r1 \n\
428 clear.d $r2 \n\
429 clear.d $r3 \n\
430 clear.d $r4 \n\
431 clear.d $r5 \n\
432 clear.d $r6 \n\
433 clear.d $r7 \n\
434 clear.d $r8 \n\
435 clear.d $r9 \n\
436 clear.d $r10 \n\
437 clear.d $r11 \n\
438 \n\
439 ;; Now we've got this: \n\
440 ;; r13 - dst \n\
441 ;; r12 - n \n\
442 \n\
443 ;; Update n for the first loop \n\
444 subq 12*4,$r12 \n\
445 0: \n\
446 subq 12*4,$r12 \n\
447 bge 0b \n\
448 movem $r11,[$r13+] \n\
449 1: \n\
450 addq 12*4,$r12 ;; compensate for last loop underflowing n\n\
451 \n\
452 ;; Restore registers from stack \n\
453 movem [$sp+],$r10 \n\
454 2: \n\
455 .section .fixup,\"ax\" \n\
456 3: \n\
457 move.d [$sp],$r10 \n\
458 addq 12*4,$r10 \n\
459 move.d $r10,[$sp] \n\
460 clear.d $r10 \n\
461 jump 0b \n\
462 \n\
463 4: \n\
464 movem [$sp+],$r10 \n\
465 addq 12*4,$r10 \n\
466 addq 12*4,$r12 \n\
467 jump 2b \n\
468 \n\
469 .previous \n\
470 .section __ex_table,\"a\" \n\
471 .dword 0b,3b \n\
472 .dword 1b,4b \n\
473 .previous"
474
475 /* Outputs */ : "=r" (dst), "=r" (n), "=r" (retn)
476 /* Inputs */ : "0" (dst), "1" (n), "2" (retn)
477 /* Clobber */ : "r11");
478 }
479
480 while (n >= 16)
481 {
482 __asm_clear_16 (dst, retn);
483 n -= 16;
484 }
485
486 /* Having a separate by-four loops cuts down on cache footprint.
487 FIXME: Test with and without; increasing switch to be 0..15. */
488 while (n >= 4)
489 {
490 __asm_clear_4 (dst, retn);
491 n -= 4;
492 }
493
494 switch (n)
495 {
496 case 0:
497 break;
498 case 1:
499 __asm_clear_1 (dst, retn);
500 break;
501 case 2:
502 __asm_clear_2 (dst, retn);
503 break;
504 case 3:
505 __asm_clear_3 (dst, retn);
506 break;
507 }
508
509 return retn;
510 }
511 EXPORT_SYMBOL(__do_clear_user);
512