1 /*
2  * User address space access functions.
3  * The non-inlined parts of asm-cris/uaccess.h are here.
4  *
5  * Copyright (C) 2000, Axis Communications AB.
6  *
7  * Written by Hans-Peter Nilsson.
8  * Pieces used from memcpy, originally by Kenny Ranerup long time ago.
9  */
10 
11 #include <asm/uaccess.h>
12 
13 /* Asm:s have been tweaked (within the domain of correctness) to give
14    satisfactory results for "gcc version 2.96 20000427 (experimental)".
15 
16    Check regularly...
17 
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.
23 
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.  */
28 
29 
30 /* Copy to userspace.  This is based on the memcpy used for
31    kernel-to-kernel copying; see "string.c".  */
32 
33 unsigned long
__copy_user(void __user * pdst,const void * psrc,unsigned long pn)34 __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 
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
193    inaccessible.  */
194 
195 unsigned long
__copy_user_zeroing(void * pdst,const void __user * psrc,unsigned long pn)196 __copy_user_zeroing(void *pdst, const void __user *psrc, 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     }
221 
222     if (((unsigned long) src & 2) && n >= 2)
223     {
224       __asm_copy_from_user_2 (dst, src, retn);
225       n -= 2;
226     }
227 
228     /* We only need one check after the unalignment-adjustments, because
229        if both adjustments were done, either both or neither reference
230        had an exception.  */
231     if (retn != 0)
232       goto copy_exception_bytes;
233   }
234 
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.  */
239   {
240     /* For large copies we use 'movem' */
241 
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
245        suboptimal.
246 
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.
252 
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\
258 	.err								\n\
259 	.endif								\n\
260 									\n\
261 	;; Save the registers we'll use in the movem process		\n\
262 	;; on the stack.						\n\
263 	subq	11*4,$sp						\n\
264 	movem	$r10,[$sp]						\n\
265 									\n\
266 	;; Now we've got this:						\n\
267 	;; r11 - src							\n\
268 	;; r13 - dst							\n\
269 	;; r12 - n							\n\
270 									\n\
271 	;; Update n for the first loop					\n\
272 	subq	44,$r12							\n\
273 0:									\n\
274 	movem	[$r11+],$r10						\n\
275 1:									\n\
276 	subq   44,$r12							\n\
277 	bge	0b							\n\
278 	movem	$r10,[$r13+]						\n\
279 									\n\
280 	addq   44,$r12  ;; compensate for last loop underflowing n	\n\
281 									\n\
282 	;; Restore registers from stack					\n\
283 	movem [$sp+],$r10						\n\
284 4:									\n\
285 	.section .fixup,\"ax\"						\n\
286 									\n\
287 ;; Do not jump back into the loop if we fail.  For some uses, we get a	\n\
288 ;; page fault somewhere on the line.  Without checking for page limits,	\n\
289 ;; we don't know where, but we need to copy accurately and keep an	\n\
290 ;; accurate count; not just clear the whole line.  To do that, we fall	\n\
291 ;; down in the code below, proceeding with smaller amounts.  It should	\n\
292 ;; be kept in mind that we have to cater to code like what at one time	\n\
293 ;; was in fs/super.c:							\n\
294 ;;  i = size - copy_from_user((void *)page, data, size);		\n\
295 ;; which would cause repeated faults while clearing the remainder of	\n\
296 ;; the SIZE bytes at PAGE after the first fault.			\n\
297 ;; A caveat here is that we must not fall through from a failing page	\n\
298 ;; to a valid page.							\n\
299 									\n\
300 3:									\n\
301 	movem  [$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\
305 									\n\
306 	.previous							\n\
307 	.section __ex_table,\"a\"					\n\
308 	.dword 1b,3b							\n\
309 	.previous"
310 
311      /* Outputs */ : "=r" (dst), "=r" (src), "=r" (n), "=r" (retn)
312      /* Inputs */ : "0" (dst), "1" (src), "2" (n), "3" (retn));
313 
314   }
315 
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.)
320 
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).  */
324 
325   while (n >= 4)
326   {
327     __asm_copy_from_user_4 (dst, src, retn);
328     n -= 4;
329 
330     if (retn)
331       goto copy_exception_bytes;
332   }
333 
334   /* If we get here, there were no memory read faults.  */
335   switch (n)
336   {
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.  */
340     case 0:
341       /* This case deliberately left in (if you have doubts check the
342 	 generated assembly code).  */
343       break;
344     case 1:
345       __asm_copy_from_user_1 (dst, src, retn);
346       break;
347     case 2:
348       __asm_copy_from_user_2 (dst, src, retn);
349       break;
350     case 3:
351       __asm_copy_from_user_3 (dst, src, retn);
352       break;
353   }
354 
355   /* If we get here, retn correctly reflects the number of failing
356      bytes.  */
357   return retn;
358 
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.  */
364   {
365     char *endp;
366     for (endp = dst + n; dst < endp; dst++)
367       *dst = 0;
368   }
369 
370   return retn + n;
371 }
372 
373 /* Zero userspace.  */
374 
375 unsigned long
__do_clear_user(void __user * pto,unsigned long pn)376 __do_clear_user (void __user *pto, unsigned long pn)
377 {
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).
381 
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. */
385 
386   register char *dst __asm__ ("r13") = pto;
387   register int n __asm__ ("r12") = pn;
388   register int retn __asm__ ("r10") = 0;
389 
390 
391   if (((unsigned long) dst & 3) != 0
392      /* Don't align if we wouldn't copy more than a few bytes.  */
393       && n >= 3)
394   {
395     if ((unsigned long) dst & 1)
396     {
397       __asm_clear_1 (dst, retn);
398       n--;
399     }
400 
401     if ((unsigned long) dst & 2)
402     {
403       __asm_clear_2 (dst, retn);
404       n -= 2;
405     }
406   }
407 
408   /* Decide which copying method to use.
409      FIXME: This number is from the "ordinary" kernel memset.  */
410   if (n >= (1*48))
411   {
412     /* For large clears we use 'movem' */
413 
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.
418 
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.
424 
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\
430 	.err								\n\
431 	.endif								\n\
432 									\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\
435 	;; upset.							\n\
436 	subq	11*4,$sp						\n\
437 	movem	$r10,[$sp]						\n\
438 									\n\
439 	clear.d $r0							\n\
440 	clear.d $r1							\n\
441 	clear.d $r2							\n\
442 	clear.d $r3							\n\
443 	clear.d $r4							\n\
444 	clear.d $r5							\n\
445 	clear.d $r6							\n\
446 	clear.d $r7							\n\
447 	clear.d $r8							\n\
448 	clear.d $r9							\n\
449 	clear.d $r10							\n\
450 	clear.d $r11							\n\
451 									\n\
452 	;; Now we've got this:						\n\
453 	;; r13 - dst							\n\
454 	;; r12 - n							\n\
455 									\n\
456 	;; Update n for the first loop					\n\
457 	subq	12*4,$r12						\n\
458 0:									\n\
459 	subq   12*4,$r12						\n\
460 	bge	0b							\n\
461 	movem	$r11,[$r13+]						\n\
462 1:									\n\
463 	addq   12*4,$r12        ;; compensate for last loop underflowing n\n\
464 									\n\
465 	;; Restore registers from stack					\n\
466 	movem [$sp+],$r10						\n\
467 2:									\n\
468 	.section .fixup,\"ax\"						\n\
469 3:									\n\
470 	move.d [$sp],$r10						\n\
471 	addq 12*4,$r10							\n\
472 	move.d $r10,[$sp]						\n\
473 	clear.d $r10							\n\
474 	jump 0b								\n\
475 									\n\
476 4:									\n\
477 	movem [$sp+],$r10						\n\
478 	addq 12*4,$r10							\n\
479 	addq 12*4,$r12							\n\
480 	jump 2b								\n\
481 									\n\
482 	.previous							\n\
483 	.section __ex_table,\"a\"					\n\
484 	.dword 0b,3b							\n\
485 	.dword 1b,4b							\n\
486 	.previous"
487 
488      /* Outputs */ : "=r" (dst), "=r" (n), "=r" (retn)
489      /* Inputs */ : "0" (dst), "1" (n), "2" (retn)
490      /* Clobber */ : "r11");
491   }
492 
493   while (n >= 16)
494   {
495     __asm_clear_16 (dst, retn);
496     n -= 16;
497   }
498 
499   /* Having a separate by-four loops cuts down on cache footprint.
500      FIXME:  Test with and without; increasing switch to be 0..15.  */
501   while (n >= 4)
502   {
503     __asm_clear_4 (dst, retn);
504     n -= 4;
505   }
506 
507   switch (n)
508   {
509     case 0:
510       break;
511     case 1:
512       __asm_clear_1 (dst, retn);
513       break;
514     case 2:
515       __asm_clear_2 (dst, retn);
516       break;
517     case 3:
518       __asm_clear_3 (dst, retn);
519       break;
520   }
521 
522   return retn;
523 }
524