summaryrefslogtreecommitdiff
path: root/arch/cris/arch-v10/lib/usercopy.c
blob: b964c667acedfd6fbc2583a2205a2898cb70eb1c (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
/*
 * User address space access functions.
 * The non-inlined parts of asm-cris/uaccess.h are here.
 *
 * Copyright (C) 2000, Axis Communications AB.
 *
 * Written by Hans-Peter Nilsson.
 * Pieces used from memcpy, originally by Kenny Ranerup long time ago.
 */

#include <asm/uaccess.h>

/* Asm:s have been tweaked (within the domain of correctness) to give
   satisfactory results for "gcc version 2.96 20000427 (experimental)".

   Check regularly...

   Note that the PC saved at a bus-fault is the address *after* the
   faulting instruction, which means the branch-target for instructions in
   delay-slots for taken branches.  Note also that the postincrement in
   the instruction is performed regardless of bus-fault; the register is
   seen updated in fault handlers.

   Oh, and on the code formatting issue, to whomever feels like "fixing
   it" to Conformity: I'm too "lazy", but why don't you go ahead and "fix"
   string.c too.  I just don't think too many people will hack this file
   for the code format to be an issue.  */


/* Copy to userspace.  This is based on the memcpy used for
   kernel-to-kernel copying; see "string.c".  */

unsigned long __copy_user(void __user *pdst, const void *psrc, unsigned long pn)
{
  /* We want the parameters put in special registers.
     Make sure the compiler is able to make something useful of this.
     As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop).

     FIXME: Comment for old gcc version.  Check.
     If gcc was alright, it really would need no temporaries, and no
     stack space to save stuff on. */

  register char *dst __asm__ ("r13") = pdst;
  register const char *src __asm__ ("r11") = psrc;
  register int n __asm__ ("r12") = pn;
  register int retn __asm__ ("r10") = 0;


  /* When src is aligned but not dst, this makes a few extra needless
     cycles.  I believe it would take as many to check that the
     re-alignment was unnecessary.  */
  if (((unsigned long) dst & 3) != 0
      /* Don't align if we wouldn't copy more than a few bytes; so we
	 don't have to check further for overflows.  */
      && n >= 3)
  {
    if ((unsigned long) dst & 1)
    {
      __asm_copy_to_user_1 (dst, src, retn);
      n--;
    }

    if ((unsigned long) dst & 2)
    {
      __asm_copy_to_user_2 (dst, src, retn);
      n -= 2;
    }
  }

  /* Decide which copying method to use. */
  if (n >= 44*2)		/* Break even between movem and
				   move16 is at 38.7*2, but modulo 44. */
  {
    /* For large copies we use 'movem'.  */

    /* It is not optimal to tell the compiler about clobbering any
       registers; that will move the saving/restoring of those registers
       to the function prologue/epilogue, and make non-movem sizes
       suboptimal.

       This method is not foolproof; it assumes that the "asm reg"
       declarations at the beginning of the function really are used
       here (beware: they may be moved to temporary registers).
       This way, we do not have to save/move the registers around into
       temporaries; we can safely use them straight away.

       If you want to check that the allocation was right; then
       check the equalities in the first comment.  It should say
       "r13=r13, r11=r11, r12=r12".  */
    __asm__ volatile ("\
	.ifnc %0%1%2%3,$r13$r11$r12$r10					\n\
	.err								\n\
	.endif								\n\
									\n\
	;; Save the registers we'll use in the movem process		\n\
	;; on the stack.						\n\
	subq	11*4,$sp						\n\
	movem	$r10,[$sp]						\n\
									\n\
	;; Now we've got this:						\n\
	;; r11 - src							\n\
	;; r13 - dst							\n\
	;; r12 - n							\n\
									\n\
	;; Update n for the first loop					\n\
	subq	44,$r12							\n\
									\n\
; Since the noted PC of a faulting instruction in a delay-slot of a taken \n\
; branch, is that of the branch target, we actually point at the from-movem \n\
; for this case.  There is no ambiguity here; if there was a fault in that \n\
; instruction (meaning a kernel oops), the faulted PC would be the address \n\
; after *that* movem.							\n\
									\n\
0:									\n\
	movem	[$r11+],$r10						\n\
	subq   44,$r12							\n\
	bge	0b							\n\
	movem	$r10,[$r13+]						\n\
1:									\n\
	addq   44,$r12  ;; compensate for last loop underflowing n	\n\
									\n\
	;; Restore registers from stack					\n\
	movem [$sp+],$r10						\n\
2:									\n\
	.section .fixup,\"ax\"						\n\
									\n\
; To provide a correct count in r10 of bytes that failed to be copied,	\n\
; we jump back into the loop if the loop-branch was taken.  There is no	\n\
; performance penalty for sany use; the program will segfault soon enough.\n\
									\n\
3:									\n\
	move.d [$sp],$r10						\n\
	addq 44,$r10							\n\
	move.d $r10,[$sp]						\n\
	jump 0b								\n\
4:									\n\
	movem [$sp+],$r10						\n\
	addq 44,$r10							\n\
	addq 44,$r12							\n\
	jump 2b								\n\
									\n\
	.previous							\n\
	.section __ex_table,\"a\"					\n\
	.dword 0b,3b							\n\
	.dword 1b,4b							\n\
	.previous"

     /* Outputs */ : "=r" (dst), "=r" (src), "=r" (n), "=r" (retn)
     /* Inputs */ : "0" (dst), "1" (src), "2" (n), "3" (retn));

  }

  /* Either we directly start copying, using dword copying in a loop, or
     we copy as much as possible with 'movem' and then the last block (<44
     bytes) is copied here.  This will work since 'movem' will have
     updated SRC, DST and N.  */

  while (n >= 16)
  {
    __asm_copy_to_user_16 (dst, src, retn);
    n -= 16;
  }

  /* Having a separate by-four loops cuts down on cache footprint.
     FIXME:  Test with and without; increasing switch to be 0..15.  */
  while (n >= 4)
  {
    __asm_copy_to_user_4 (dst, src, retn);
    n -= 4;
  }

  switch (n)
  {
    case 0:
      break;
    case 1:
      __asm_copy_to_user_1 (dst, src, retn);
      break;
    case 2:
      __asm_copy_to_user_2 (dst, src, retn);
      break;
    case 3:
      __asm_copy_to_user_3 (dst, src, retn);
      break;
  }

  return retn;
}
EXPORT_SYMBOL(__copy_user);

/* Copy from user to kernel, zeroing the bytes that were inaccessible in
   userland.  The return-value is the number of bytes that were
   inaccessible.  */

unsigned long __copy_user_zeroing(void *pdst, const void __user *psrc,
				  unsigned long pn)
{
  /* We want the parameters put in special registers.
     Make sure the compiler is able to make something useful of this.
     As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop).

     FIXME: Comment for old gcc version.  Check.
     If gcc was alright, it really would need no temporaries, and no
     stack space to save stuff on.  */

  register char *dst __asm__ ("r13") = pdst;
  register const char *src __asm__ ("r11") = psrc;
  register int n __asm__ ("r12") = pn;
  register int retn __asm__ ("r10") = 0;

  /* The best reason to align src is that we then know that a read-fault
     was for aligned bytes; there's no 1..3 remaining good bytes to
     pickle.  */
  if (((unsigned long) src & 3) != 0)
  {
    if (((unsigned long) src & 1) && n != 0)
    {
      __asm_copy_from_user_1 (dst, src, retn);
      n--;
    }

    if (((unsigned long) src & 2) && n >= 2)
    {
      __asm_copy_from_user_2 (dst, src, retn);
      n -= 2;
    }

    /* We only need one check after the unalignment-adjustments, because
       if both adjustments were done, either both or neither reference
       had an exception.  */
    if (retn != 0)
      goto copy_exception_bytes;
  }

  /* Decide which copying method to use. */
  if (n >= 44*2)		/* Break even between movem and
				   move16 is at 38.7*2, but modulo 44.
				   FIXME: We use move4 now.  */
  {
    /* For large copies we use 'movem' */

    /* It is not optimal to tell the compiler about clobbering any
       registers; that will move the saving/restoring of those registers
       to the function prologue/epilogue, and make non-movem sizes
       suboptimal.

       This method is not foolproof; it assumes that the "asm reg"
       declarations at the beginning of the function really are used
       here (beware: they may be moved to temporary registers).
       This way, we do not have to save/move the registers around into
       temporaries; we can safely use them straight away.

       If you want to check that the allocation was right; then
       check the equalities in the first comment.  It should say
       "r13=r13, r11=r11, r12=r12" */
    __asm__ volatile ("\n\
	.ifnc %0%1%2%3,$r13$r11$r12$r10					\n\
	.err								\n\
	.endif								\n\
									\n\
	;; Save the registers we'll use in the movem process		\n\
	;; on the stack.						\n\
	subq	11*4,$sp						\n\
	movem	$r10,[$sp]						\n\
									\n\
	;; Now we've got this:						\n\
	;; r11 - src							\n\
	;; r13 - dst							\n\
	;; r12 - n							\n\
									\n\
	;; Update n for the first loop					\n\
	subq	44,$r12							\n\
0:									\n\
	movem	[$r11+],$r10						\n\
1:									\n\
	subq   44,$r12							\n\
	bge	0b							\n\
	movem	$r10,[$r13+]						\n\
									\n\
	addq   44,$r12  ;; compensate for last loop underflowing n	\n\
									\n\
	;; Restore registers from stack					\n\
	movem [$sp+],$r10						\n\
4:									\n\
	.section .fixup,\"ax\"						\n\
									\n\
;; Do not jump back into the loop if we fail.  For some uses, we get a	\n\
;; page fault somewhere on the line.  Without checking for page limits,	\n\
;; we don't know where, but we need to copy accurately and keep an	\n\
;; accurate count; not just clear the whole line.  To do that, we fall	\n\
;; down in the code below, proceeding with smaller amounts.  It should	\n\
;; be kept in mind that we have to cater to code like what at one time	\n\
;; was in fs/super.c:							\n\
;;  i = size - copy_from_user((void *)page, data, size);		\n\
;; which would cause repeated faults while clearing the remainder of	\n\
;; the SIZE bytes at PAGE after the first fault.			\n\
;; A caveat here is that we must not fall through from a failing page	\n\
;; to a valid page.							\n\
									\n\
3:									\n\
	movem  [$sp+],$r10						\n\
	addq	44,$r12 ;; Get back count before faulting point.	\n\
	subq	44,$r11 ;; Get back pointer to faulting movem-line.	\n\
	jump	4b	;; Fall through, pretending the fault didn't happen.\n\
									\n\
	.previous							\n\
	.section __ex_table,\"a\"					\n\
	.dword 1b,3b							\n\
	.previous"

     /* Outputs */ : "=r" (dst), "=r" (src), "=r" (n), "=r" (retn)
     /* Inputs */ : "0" (dst), "1" (src), "2" (n), "3" (retn));

  }

  /* Either we directly start copying here, using dword copying in a loop,
     or we copy as much as possible with 'movem' and then the last block
     (<44 bytes) is copied here.  This will work since 'movem' will have
     updated src, dst and n.  (Except with failing src.)

     Since we want to keep src accurate, we can't use
     __asm_copy_from_user_N with N != (1, 2, 4); it updates dst and
     retn, but not src (by design; it's value is ignored elsewhere).  */

  while (n >= 4)
  {
    __asm_copy_from_user_4 (dst, src, retn);
    n -= 4;

    if (retn)
      goto copy_exception_bytes;
  }

  /* If we get here, there were no memory read faults.  */
  switch (n)
  {
    /* These copies are at least "naturally aligned" (so we don't have
       to check each byte), due to the src alignment code before the
       movem loop.  The *_3 case *will* get the correct count for retn.  */
    case 0:
      /* This case deliberately left in (if you have doubts check the
	 generated assembly code).  */
      break;
    case 1:
      __asm_copy_from_user_1 (dst, src, retn);
      break;
    case 2:
      __asm_copy_from_user_2 (dst, src, retn);
      break;
    case 3:
      __asm_copy_from_user_3 (dst, src, retn);
      break;
  }

  /* If we get here, retn correctly reflects the number of failing
     bytes.  */
  return retn;

copy_exception_bytes:
  /* We already have "retn" bytes cleared, and need to clear the
     remaining "n" bytes.  A non-optimized simple byte-for-byte in-line
     memset is preferred here, since this isn't speed-critical code and
     we'd rather have this a leaf-function than calling memset.  */
  {
    char *endp;
    for (endp = dst + n; dst < endp; dst++)
      *dst = 0;
  }

  return retn + n;
}
EXPORT_SYMBOL(__copy_user_zeroing);

/* Zero userspace.  */
unsigned long __do_clear_user(void __user *pto, unsigned long pn)
{
  /* We want the parameters put in special registers.
     Make sure the compiler is able to make something useful of this.
      As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop).

     FIXME: Comment for old gcc version.  Check.
     If gcc was alright, it really would need no temporaries, and no
     stack space to save stuff on. */

  register char *dst __asm__ ("r13") = pto;
  register int n __asm__ ("r12") = pn;
  register int retn __asm__ ("r10") = 0;


  if (((unsigned long) dst & 3) != 0
     /* Don't align if we wouldn't copy more than a few bytes.  */
      && n >= 3)
  {
    if ((unsigned long) dst & 1)
    {
      __asm_clear_1 (dst, retn);
      n--;
    }

    if ((unsigned long) dst & 2)
    {
      __asm_clear_2 (dst, retn);
      n -= 2;
    }
  }

  /* Decide which copying method to use.
     FIXME: This number is from the "ordinary" kernel memset.  */
  if (n >= (1*48))
  {
    /* For large clears we use 'movem' */

    /* It is not optimal to tell the compiler about clobbering any
       call-saved registers; that will move the saving/restoring of
       those registers to the function prologue/epilogue, and make
       non-movem sizes suboptimal.

       This method is not foolproof; it assumes that the "asm reg"
       declarations at the beginning of the function really are used
       here (beware: they may be moved to temporary registers).
       This way, we do not have to save/move the registers around into
       temporaries; we can safely use them straight away.

      If you want to check that the allocation was right; then
      check the equalities in the first comment.  It should say
      something like "r13=r13, r11=r11, r12=r12". */
    __asm__ volatile ("\n\
	.ifnc %0%1%2,$r13$r12$r10					\n\
	.err								\n\
	.endif								\n\
									\n\
	;; Save the registers we'll clobber in the movem process	\n\
	;; on the stack.  Don't mention them to gcc, it will only be	\n\
	;; upset.							\n\
	subq	11*4,$sp						\n\
	movem	$r10,[$sp]						\n\
									\n\
	clear.d $r0							\n\
	clear.d $r1							\n\
	clear.d $r2							\n\
	clear.d $r3							\n\
	clear.d $r4							\n\
	clear.d $r5							\n\
	clear.d $r6							\n\
	clear.d $r7							\n\
	clear.d $r8							\n\
	clear.d $r9							\n\
	clear.d $r10							\n\
	clear.d $r11							\n\
									\n\
	;; Now we've got this:						\n\
	;; r13 - dst							\n\
	;; r12 - n							\n\
									\n\
	;; Update n for the first loop					\n\
	subq	12*4,$r12						\n\
0:									\n\
	subq   12*4,$r12						\n\
	bge	0b							\n\
	movem	$r11,[$r13+]						\n\
1:									\n\
	addq   12*4,$r12        ;; compensate for last loop underflowing n\n\
									\n\
	;; Restore registers from stack					\n\
	movem [$sp+],$r10						\n\
2:									\n\
	.section .fixup,\"ax\"						\n\
3:									\n\
	move.d [$sp],$r10						\n\
	addq 12*4,$r10							\n\
	move.d $r10,[$sp]						\n\
	clear.d $r10							\n\
	jump 0b								\n\
									\n\
4:									\n\
	movem [$sp+],$r10						\n\
	addq 12*4,$r10							\n\
	addq 12*4,$r12							\n\
	jump 2b								\n\
									\n\
	.previous							\n\
	.section __ex_table,\"a\"					\n\
	.dword 0b,3b							\n\
	.dword 1b,4b							\n\
	.previous"

     /* Outputs */ : "=r" (dst), "=r" (n), "=r" (retn)
     /* Inputs */ : "0" (dst), "1" (n), "2" (retn)
     /* Clobber */ : "r11");
  }

  while (n >= 16)
  {
    __asm_clear_16 (dst, retn);
    n -= 16;
  }

  /* Having a separate by-four loops cuts down on cache footprint.
     FIXME:  Test with and without; increasing switch to be 0..15.  */
  while (n >= 4)
  {
    __asm_clear_4 (dst, retn);
    n -= 4;
  }

  switch (n)
  {
    case 0:
      break;
    case 1:
      __asm_clear_1 (dst, retn);
      break;
    case 2:
      __asm_clear_2 (dst, retn);
      break;
    case 3:
      __asm_clear_3 (dst, retn);
      break;
  }

  return retn;
}
EXPORT_SYMBOL(__do_clear_user);