2
* User address space access functions.
3
* The non-inlined parts of asm-cris/uaccess.h are here.
5
* Copyright (C) 2000, Axis Communications AB.
7
* Written by Hans-Peter Nilsson.
8
* Pieces used from memcpy, originally by Kenny Ranerup long time ago.
11
#include <asm/uaccess.h>
13
/* Asm:s have been tweaked (within the domain of correctness) to give
14
satisfactory results for "gcc version 2.96 20000427 (experimental)".
18
Note that the PC saved at a bus-fault is the address *after* the
19
faulting instruction, which means the branch-target for instructions in
20
delay-slots for taken branches. Note also that the postincrement in
21
the instruction is performed regardless of bus-fault; the register is
22
seen updated in fault handlers.
24
Oh, and on the code formatting issue, to whomever feels like "fixing
25
it" to Conformity: I'm too "lazy", but why don't you go ahead and "fix"
26
string.c too. I just don't think too many people will hack this file
27
for the code format to be an issue. */
30
/* Copy to userspace. This is based on the memcpy used for
31
kernel-to-kernel copying; see "string.c". */
34
__copy_user (void __user *pdst, const void *psrc, unsigned long pn)
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).
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. */
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;
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. */
58
if ((unsigned long) dst & 1)
60
__asm_copy_to_user_1 (dst, src, retn);
64
if ((unsigned long) dst & 2)
66
__asm_copy_to_user_2 (dst, src, retn);
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. */
75
/* For large copies we use 'movem'. */
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
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.
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". */
92
.ifnc %0%1%2%3,$r13$r11$r12$r10 \n\
96
;; Save the registers we'll use in the movem process \n\
101
;; Now we've got this: \n\
106
;; Update n for the first loop \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\
116
movem [$r11+],$r10 \n\
119
movem $r10,[$r13+] \n\
121
addq 44,$r12 ;; compensate for last loop underflowing n \n\
123
;; Restore registers from stack \n\
124
movem [$sp+],$r10 \n\
126
.section .fixup,\"ax\" \n\
128
; To provide a correct count in r10 of bytes that failed to be copied, \n\
129
; we jump back into the loop if the loop-branch was taken. There is no \n\
130
; performance penalty for sany use; the program will segfault soon enough.\n\
133
move.d [$sp],$r10 \n\
135
move.d $r10,[$sp] \n\
138
movem [$sp+],$r10 \n\
144
.section __ex_table,\"a\" \n\
149
/* Outputs */ : "=r" (dst), "=r" (src), "=r" (n), "=r" (retn)
150
/* Inputs */ : "0" (dst), "1" (src), "2" (n), "3" (retn));
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. */
161
__asm_copy_to_user_16 (dst, src, retn);
165
/* Having a separate by-four loops cuts down on cache footprint.
166
FIXME: Test with and without; increasing switch to be 0..15. */
169
__asm_copy_to_user_4 (dst, src, retn);
178
__asm_copy_to_user_1 (dst, src, retn);
181
__asm_copy_to_user_2 (dst, src, retn);
184
__asm_copy_to_user_3 (dst, src, retn);
191
/* Copy from user to kernel, zeroing the bytes that were inaccessible in
192
userland. The return-value is the number of bytes that were
196
__copy_user_zeroing(void *pdst, const void __user *psrc, unsigned long pn)
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).
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. */
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;
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
214
if (((unsigned long) src & 3) != 0)
216
if (((unsigned long) src & 1) && n != 0)
218
__asm_copy_from_user_1 (dst, src, retn);
222
if (((unsigned long) src & 2) && n >= 2)
224
__asm_copy_from_user_2 (dst, src, retn);
228
/* We only need one check after the unalignment-adjustments, because
229
if both adjustments were done, either both or neither reference
232
goto copy_exception_bytes;
235
/* Decide which copying method to use. */
236
if (n >= 44*2) /* Break even between movem and
237
move16 is at 38.7*2, but modulo 44.
238
FIXME: We use move4 now. */
240
/* For large copies we use 'movem' */
242
/* It is not optimal to tell the compiler about clobbering any
243
registers; that will move the saving/restoring of those registers
244
to the function prologue/epilogue, and make non-movem sizes
247
This method is not foolproof; it assumes that the "asm reg"
248
declarations at the beginning of the function really are used
249
here (beware: they may be moved to temporary registers).
250
This way, we do not have to save/move the registers around into
251
temporaries; we can safely use them straight away.
253
If you want to check that the allocation was right; then
254
check the equalities in the first comment. It should say
255
"r13=r13, r11=r11, r12=r12" */
256
__asm__ volatile ("\n\
257
.ifnc %0%1%2%3,$r13$r11$r12$r10 \n\
261
;; Save the registers we'll use in the movem process \n\
266
;; Now we've got this: \n\
271
;; Update n for the first loop \n\
274
movem [$r11+],$r10 \n\
278
movem $r10,[$r13+] \n\
280
addq 44,$r12 ;; compensate for last loop underflowing n \n\
282
;; Restore registers from stack \n\
283
movem [$sp+],$r10 \n\
285
.section .fixup,\"ax\" \n\
287
;; Do not jump back into the loop if we fail. For some uses, we get a \n\
288
;; page fault somewhere on the line. Without checking for page limits, \n\
289
;; we don't know where, but we need to copy accurately and keep an \n\
290
;; accurate count; not just clear the whole line. To do that, we fall \n\
291
;; down in the code below, proceeding with smaller amounts. It should \n\
292
;; be kept in mind that we have to cater to code like what at one time \n\
293
;; was in fs/super.c: \n\
294
;; i = size - copy_from_user((void *)page, data, size); \n\
295
;; which would cause repeated faults while clearing the remainder of \n\
296
;; the SIZE bytes at PAGE after the first fault. \n\
297
;; A caveat here is that we must not fall through from a failing page \n\
298
;; to a valid page. \n\
301
movem [$sp+],$r10 \n\
302
addq 44,$r12 ;; Get back count before faulting point. \n\
303
subq 44,$r11 ;; Get back pointer to faulting movem-line. \n\
304
jump 4b ;; Fall through, pretending the fault didn't happen.\n\
307
.section __ex_table,\"a\" \n\
311
/* Outputs */ : "=r" (dst), "=r" (src), "=r" (n), "=r" (retn)
312
/* Inputs */ : "0" (dst), "1" (src), "2" (n), "3" (retn));
316
/* Either we directly start copying here, using dword copying in a loop,
317
or we copy as much as possible with 'movem' and then the last block
318
(<44 bytes) is copied here. This will work since 'movem' will have
319
updated src, dst and n. (Except with failing src.)
321
Since we want to keep src accurate, we can't use
322
__asm_copy_from_user_N with N != (1, 2, 4); it updates dst and
323
retn, but not src (by design; it's value is ignored elsewhere). */
327
__asm_copy_from_user_4 (dst, src, retn);
331
goto copy_exception_bytes;
334
/* If we get here, there were no memory read faults. */
337
/* These copies are at least "naturally aligned" (so we don't have
338
to check each byte), due to the src alignment code before the
339
movem loop. The *_3 case *will* get the correct count for retn. */
341
/* This case deliberately left in (if you have doubts check the
342
generated assembly code). */
345
__asm_copy_from_user_1 (dst, src, retn);
348
__asm_copy_from_user_2 (dst, src, retn);
351
__asm_copy_from_user_3 (dst, src, retn);
355
/* If we get here, retn correctly reflects the number of failing
359
copy_exception_bytes:
360
/* We already have "retn" bytes cleared, and need to clear the
361
remaining "n" bytes. A non-optimized simple byte-for-byte in-line
362
memset is preferred here, since this isn't speed-critical code and
363
we'd rather have this a leaf-function than calling memset. */
366
for (endp = dst + n; dst < endp; dst++)
373
/* Zero userspace. */
376
__do_clear_user (void __user *pto, unsigned long pn)
378
/* We want the parameters put in special registers.
379
Make sure the compiler is able to make something useful of this.
380
As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop).
382
FIXME: Comment for old gcc version. Check.
383
If gcc was alright, it really would need no temporaries, and no
384
stack space to save stuff on. */
386
register char *dst __asm__ ("r13") = pto;
387
register int n __asm__ ("r12") = pn;
388
register int retn __asm__ ("r10") = 0;
391
if (((unsigned long) dst & 3) != 0
392
/* Don't align if we wouldn't copy more than a few bytes. */
395
if ((unsigned long) dst & 1)
397
__asm_clear_1 (dst, retn);
401
if ((unsigned long) dst & 2)
403
__asm_clear_2 (dst, retn);
408
/* Decide which copying method to use.
409
FIXME: This number is from the "ordinary" kernel memset. */
412
/* For large clears we use 'movem' */
414
/* It is not optimal to tell the compiler about clobbering any
415
call-saved registers; that will move the saving/restoring of
416
those registers to the function prologue/epilogue, and make
417
non-movem sizes suboptimal.
419
This method is not foolproof; it assumes that the "asm reg"
420
declarations at the beginning of the function really are used
421
here (beware: they may be moved to temporary registers).
422
This way, we do not have to save/move the registers around into
423
temporaries; we can safely use them straight away.
425
If you want to check that the allocation was right; then
426
check the equalities in the first comment. It should say
427
something like "r13=r13, r11=r11, r12=r12". */
428
__asm__ volatile ("\n\
429
.ifnc %0%1%2,$r13$r12$r10 \n\
433
;; Save the registers we'll clobber in the movem process \n\
434
;; on the stack. Don't mention them to gcc, it will only be \n\
452
;; Now we've got this: \n\
456
;; Update n for the first loop \n\
461
movem $r11,[$r13+] \n\
463
addq 12*4,$r12 ;; compensate for last loop underflowing n\n\
465
;; Restore registers from stack \n\
466
movem [$sp+],$r10 \n\
468
.section .fixup,\"ax\" \n\
470
move.d [$sp],$r10 \n\
472
move.d $r10,[$sp] \n\
477
movem [$sp+],$r10 \n\
483
.section __ex_table,\"a\" \n\
488
/* Outputs */ : "=r" (dst), "=r" (n), "=r" (retn)
489
/* Inputs */ : "0" (dst), "1" (n), "2" (retn)
490
/* Clobber */ : "r11");
495
__asm_clear_16 (dst, retn);
499
/* Having a separate by-four loops cuts down on cache footprint.
500
FIXME: Test with and without; increasing switch to be 0..15. */
503
__asm_clear_4 (dst, retn);
512
__asm_clear_1 (dst, retn);
515
__asm_clear_2 (dst, retn);
518
__asm_clear_3 (dst, retn);