?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
mm: fix exec activate_mm vs TLB shootdown and lazy tlb switching race
[linux/fpc-iii.git] / arch / cris / arch-v32 / lib / usercopy.c
blob04e78b6ffa228a590778a795680bbe9311a9dede
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, 2003 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 3.2.1 Axis release R53/1.53-v32".
16
17 Check regularly...
18
19 Note that for CRISv32, the PC saved at a bus-fault is the address
20 *at* the faulting instruction, with a special case for instructions
21 in delay slots: then it's the address of the branch. Note also that
22 in contrast to v10, a postincrement in the instruction is *not*
23 performed at a bus-fault; the register is seen having the original
24 value in fault handlers. */
25
26
27 /* Copy to userspace. This is based on the memcpy used for
28 kernel-to-kernel copying; see "string.c". */
29
30 unsigned long __copy_user(void __user *pdst, const void *psrc, unsigned long pn)
31 {
32 /* We want the parameters put in special registers.
33 Make sure the compiler is able to make something useful of this.
34 As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop).
35
36 FIXME: Comment for old gcc version. Check.
37 If gcc was alright, it really would need no temporaries, and no
38 stack space to save stuff on. */
39
40 register char *dst __asm__ ("r13") = pdst;
41 register const char *src __asm__ ("r11") = psrc;
42 register int n __asm__ ("r12") = pn;
43 register int retn __asm__ ("r10") = 0;
44
45
46 /* When src is aligned but not dst, this makes a few extra needless
47 cycles. I believe it would take as many to check that the
48 re-alignment was unnecessary. */
49 if (((unsigned long) dst & 3) != 0
50 /* Don't align if we wouldn't copy more than a few bytes; so we
51 don't have to check further for overflows. */
52 && n >= 3)
53 {
54 if ((unsigned long) dst & 1)
55 {
56 __asm_copy_to_user_1 (dst, src, retn);
57 n--;
58 }
59
60 if ((unsigned long) dst & 2)
61 {
62 __asm_copy_to_user_2 (dst, src, retn);
63 n -= 2;
64 }
65 }
66
67 /* Movem is dirt cheap. The overheap is low enough to always use the
68 minimum possible block size as the threshold. */
69 if (n >= 44)
70 {
71 /* For large copies we use 'movem'. */
72
73 /* It is not optimal to tell the compiler about clobbering any
74 registers; that will move the saving/restoring of those registers
75 to the function prologue/epilogue, and make non-movem sizes
76 suboptimal. */
77 __asm__ volatile ("\
78 ;; Check that the register asm declaration got right. \n\
79 ;; The GCC manual explicitly says TRT will happen. \n\
80 .ifnc %0%1%2%3,$r13$r11$r12$r10 \n\
81 .err \n\
82 .endif \n\
83 \n\
84 ;; Save the registers we'll use in the movem process \n\
85 ;; on the stack. \n\
86 subq 11*4,$sp \n\
87 movem $r10,[$sp] \n\
88 \n\
89 ;; Now we've got this: \n\
90 ;; r11 - src \n\
91 ;; r13 - dst \n\
92 ;; r12 - n \n\
93 \n\
94 ;; Update n for the first loop \n\
95 subq 44,$r12 \n\
96 0: \n\
97 movem [$r11+],$r10 \n\
98 subq 44,$r12 \n\
99 1: bge 0b \n\
100 movem $r10,[$r13+] \n\
101 3: \n\
102 addq 44,$r12 ;; compensate for last loop underflowing n \n\
103 \n\
104 ;; Restore registers from stack \n\
105 movem [$sp+],$r10 \n\
106 2: \n\
107 .section .fixup,\"ax\" \n\
108 4: \n\
109 ; When failing on any of the 1..44 bytes in a chunk, we adjust back the \n\
110 ; source pointer and just drop through to the by-16 and by-4 loops to \n\
111 ; get the correct number of failing bytes. This necessarily means a \n\
112 ; few extra exceptions, but invalid user pointers shouldn't happen in \n\
113 ; time-critical code anyway. \n\
114 jump 3b \n\
115 subq 44,$r11 \n\
116 \n\
117 .previous \n\
118 .section __ex_table,\"a\" \n\
119 .dword 1b,4b \n\
120 .previous"
121
122 /* Outputs */ : "=r" (dst), "=r" (src), "=r" (n), "=r" (retn)
123 /* Inputs */ : "0" (dst), "1" (src), "2" (n), "3" (retn));
124
125 }
126
127 while (n >= 16)
128 {
129 __asm_copy_to_user_16 (dst, src, retn);
130 n -= 16;
131 }
132
133 /* Having a separate by-four loops cuts down on cache footprint.
134 FIXME: Test with and without; increasing switch to be 0..15. */
135 while (n >= 4)
136 {
137 __asm_copy_to_user_4 (dst, src, retn);
138 n -= 4;
139 }
140
141 switch (n)
142 {
143 case 0:
144 break;
145 case 1:
146 __asm_copy_to_user_1 (dst, src, retn);
147 break;
148 case 2:
149 __asm_copy_to_user_2 (dst, src, retn);
150 break;
151 case 3:
152 __asm_copy_to_user_3 (dst, src, retn);
153 break;
154 }
155
156 return retn;
157 }
158 EXPORT_SYMBOL(__copy_user);
159
160 /* Copy from user to kernel. The return-value is the number of bytes that were
161 inaccessible. */
162 unsigned long __copy_user_in(void *pdst, const void __user *psrc,
163 unsigned long pn)
164 {
165 /* We want the parameters put in special registers.
166 Make sure the compiler is able to make something useful of this.
167 As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop).
168
169 FIXME: Comment for old gcc version. Check.
170 If gcc was alright, it really would need no temporaries, and no
171 stack space to save stuff on. */
172
173 register char *dst __asm__ ("r13") = pdst;
174 register const char *src __asm__ ("r11") = psrc;
175 register int n __asm__ ("r12") = pn;
176 register int retn __asm__ ("r10") = 0;
177
178 /* The best reason to align src is that we then know that a read-fault
179 was for aligned bytes; there's no 1..3 remaining good bytes to
180 pickle. */
181 if (((unsigned long) src & 3) != 0)
182 {
183 if (((unsigned long) src & 1) && n != 0)
184 {
185 __asm_copy_from_user_1 (dst, src, retn);
186 n--;
187 if (retn != 0)
188 goto exception;
189 }
190
191 if (((unsigned long) src & 2) && n >= 2)
192 {
193 __asm_copy_from_user_2 (dst, src, retn);
194 n -= 2;
195 if (retn != 0)
196 goto exception;
197 }
198
199 }
200
201 /* Movem is dirt cheap. The overheap is low enough to always use the
202 minimum possible block size as the threshold. */
203 if (n >= 44)
204 {
205 /* It is not optimal to tell the compiler about clobbering any
206 registers; that will move the saving/restoring of those registers
207 to the function prologue/epilogue, and make non-movem sizes
208 suboptimal. */
209 __asm__ volatile ("\
210 .ifnc %0%1%2%3,$r13$r11$r12$r10 \n\
211 .err \n\
212 .endif \n\
213 \n\
214 ;; Save the registers we'll use in the movem process \n\
215 ;; on the stack. \n\
216 subq 11*4,$sp \n\
217 movem $r10,[$sp] \n\
218 \n\
219 ;; Now we've got this: \n\
220 ;; r11 - src \n\
221 ;; r13 - dst \n\
222 ;; r12 - n \n\
223 \n\
224 ;; Update n for the first loop \n\
225 subq 44,$r12 \n\
226 0: \n\
227 movem [$r11+],$r10 \n\
228 \n\
229 subq 44,$r12 \n\
230 bge 0b \n\
231 movem $r10,[$r13+] \n\
232 \n\
233 4: \n\
234 addq 44,$r12 ;; compensate for last loop underflowing n \n\
235 \n\
236 ;; Restore registers from stack \n\
237 movem [$sp+],$r10 \n\
238 .section .fixup,\"ax\" \n\
239 \n\
240 ;; Do not jump back into the loop if we fail. For some uses, we get a \n\
241 ;; page fault somewhere on the line. Without checking for page limits, \n\
242 ;; we don't know where, but we need to copy accurately and keep an \n\
243 ;; accurate count; not just clear the whole line. To do that, we fall \n\
244 ;; down in the code below, proceeding with smaller amounts. It should \n\
245 ;; be kept in mind that we have to cater to code like what at one time \n\
246 ;; was in fs/super.c: \n\
247 ;; i = size - copy_from_user((void *)page, data, size); \n\
248 ;; which would cause repeated faults while clearing the remainder of \n\
249 ;; the SIZE bytes at PAGE after the first fault. \n\
250 ;; A caveat here is that we must not fall through from a failing page \n\
251 ;; to a valid page. \n\
252 \n\
253 3: \n\
254 jump 4b ;; Fall through, pretending the fault didn't happen. \n\
255 nop \n\
256 \n\
257 .previous \n\
258 .section __ex_table,\"a\" \n\
259 .dword 0b,3b \n\
260 .previous"
261
262 /* Outputs */ : "=r" (dst), "=r" (src), "=r" (n), "=r" (retn)
263 /* Inputs */ : "0" (dst), "1" (src), "2" (n), "3" (retn));
264 }
265
266 /* Either we directly start copying here, using dword copying in a loop,
267 or we copy as much as possible with 'movem' and then the last block
268 (<44 bytes) is copied here. This will work since 'movem' will have
269 updated src, dst and n. (Except with failing src.)
270
271 Since we want to keep src accurate, we can't use
272 __asm_copy_from_user_N with N != (1, 2, 4); it updates dst and
273 retn, but not src (by design; it's value is ignored elsewhere). */
274
275 while (n >= 4)
276 {
277 __asm_copy_from_user_4 (dst, src, retn);
278 n -= 4;
279
280 if (retn)
281 goto exception;
282 }
283
284 /* If we get here, there were no memory read faults. */
285 switch (n)
286 {
287 /* These copies are at least "naturally aligned" (so we don't have
288 to check each byte), due to the src alignment code before the
289 movem loop. The *_3 case *will* get the correct count for retn. */
290 case 0:
291 /* This case deliberately left in (if you have doubts check the
292 generated assembly code). */
293 break;
294 case 1:
295 __asm_copy_from_user_1 (dst, src, retn);
296 break;
297 case 2:
298 __asm_copy_from_user_2 (dst, src, retn);
299 break;
300 case 3:
301 __asm_copy_from_user_3 (dst, src, retn);
302 break;
303 }
304
305 /* If we get here, retn correctly reflects the number of failing
306 bytes. */
307 return retn;
308
309 exception:
310 return retn + n;
311 }
312 EXPORT_SYMBOL(__copy_user_in);
313
314 /* Zero userspace. */
315 unsigned long __do_clear_user(void __user *pto, unsigned long pn)
316 {
317 /* We want the parameters put in special registers.
318 Make sure the compiler is able to make something useful of this.
319 As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop).
320
321 FIXME: Comment for old gcc version. Check.
322 If gcc was alright, it really would need no temporaries, and no
323 stack space to save stuff on. */
324
325 register char *dst __asm__ ("r13") = pto;
326 register int n __asm__ ("r12") = pn;
327 register int retn __asm__ ("r10") = 0;
328
329
330 if (((unsigned long) dst & 3) != 0
331 /* Don't align if we wouldn't copy more than a few bytes. */
332 && n >= 3)
333 {
334 if ((unsigned long) dst & 1)
335 {
336 __asm_clear_1 (dst, retn);
337 n--;
338 }
339
340 if ((unsigned long) dst & 2)
341 {
342 __asm_clear_2 (dst, retn);
343 n -= 2;
344 }
345 }
346
347 /* Decide which copying method to use.
348 FIXME: This number is from the "ordinary" kernel memset. */
349 if (n >= 48)
350 {
351 /* For large clears we use 'movem' */
352
353 /* It is not optimal to tell the compiler about clobbering any
354 call-saved registers; that will move the saving/restoring of
355 those registers to the function prologue/epilogue, and make
356 non-movem sizes suboptimal.
357
358 This method is not foolproof; it assumes that the "asm reg"
359 declarations at the beginning of the function really are used
360 here (beware: they may be moved to temporary registers).
361 This way, we do not have to save/move the registers around into
362 temporaries; we can safely use them straight away.
363
364 If you want to check that the allocation was right; then
365 check the equalities in the first comment. It should say
366 something like "r13=r13, r11=r11, r12=r12". */
367 __asm__ volatile ("\
368 .ifnc %0%1%2,$r13$r12$r10 \n\
369 .err \n\
370 .endif \n\
371 \n\
372 ;; Save the registers we'll clobber in the movem process \n\
373 ;; on the stack. Don't mention them to gcc, it will only be \n\
374 ;; upset. \n\
375 subq 11*4,$sp \n\
376 movem $r10,[$sp] \n\
377 \n\
378 clear.d $r0 \n\
379 clear.d $r1 \n\
380 clear.d $r2 \n\
381 clear.d $r3 \n\
382 clear.d $r4 \n\
383 clear.d $r5 \n\
384 clear.d $r6 \n\
385 clear.d $r7 \n\
386 clear.d $r8 \n\
387 clear.d $r9 \n\
388 clear.d $r10 \n\
389 clear.d $r11 \n\
390 \n\
391 ;; Now we've got this: \n\
392 ;; r13 - dst \n\
393 ;; r12 - n \n\
394 \n\
395 ;; Update n for the first loop \n\
396 subq 12*4,$r12 \n\
397 0: \n\
398 subq 12*4,$r12 \n\
399 1: \n\
400 bge 0b \n\
401 movem $r11,[$r13+] \n\
402 \n\
403 addq 12*4,$r12 ;; compensate for last loop underflowing n \n\
404 \n\
405 ;; Restore registers from stack \n\
406 movem [$sp+],$r10 \n\
407 2: \n\
408 .section .fixup,\"ax\" \n\
409 3: \n\
410 movem [$sp],$r10 \n\
411 addq 12*4,$r10 \n\
412 addq 12*4,$r13 \n\
413 movem $r10,[$sp] \n\
414 jump 0b \n\
415 clear.d $r10 \n\
416 \n\
417 .previous \n\
418 .section __ex_table,\"a\" \n\
419 .dword 1b,3b \n\
420 .previous"
421
422 /* Outputs */ : "=r" (dst), "=r" (n), "=r" (retn)
423 /* Inputs */ : "0" (dst), "1" (n), "2" (retn)
424 /* Clobber */ : "r11");
425 }
426
427 while (n >= 16)
428 {
429 __asm_clear_16 (dst, retn);
430 n -= 16;
431 }
432
433 /* Having a separate by-four loops cuts down on cache footprint.
434 FIXME: Test with and without; increasing switch to be 0..15. */
435 while (n >= 4)
436 {
437 __asm_clear_4 (dst, retn);
438 n -= 4;
439 }
440
441 switch (n)
442 {
443 case 0:
444 break;
445 case 1:
446 __asm_clear_1 (dst, retn);
447 break;
448 case 2:
449 __asm_clear_2 (dst, retn);
450 break;
451 case 3:
452 __asm_clear_3 (dst, retn);
453 break;
454 }
455
456 return retn;
457 }
458 EXPORT_SYMBOL(__do_clear_user);
Linux with added FPC-III support