2 * User address space access functions.
3 * The non-inlined parts of asm-cris/uaccess.h are here.
5 * Copyright (C) 2000, 2003 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 3.2.1 Axis release R53/1.53-v32".
18 Note that for CRISv32, the PC saved at a bus-fault is the address
19 *at* the faulting instruction, with a special case for instructions
20 in delay slots: then it's the address of the branch. Note also that
21 in contrast to v10, a postincrement in the instruction is *not*
22 performed at a bus-fault; the register is seen having the original
23 value in fault handlers. */
26 /* Copy to userspace. This is based on the memcpy used for
27 kernel-to-kernel copying; see "string.c". */
30 __copy_user (void __user *pdst, const void *psrc, unsigned long pn)
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).
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. */
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;
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. */
54 if ((unsigned long) dst & 1)
56 __asm_copy_to_user_1 (dst, src, retn);
60 if ((unsigned long) dst & 2)
62 __asm_copy_to_user_2 (dst, src, retn);
67 /* Movem is dirt cheap. The overheap is low enough to always use the
68 minimum possible block size as the threshold. */
71 /* For large copies we use 'movem'. */
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
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\
84 ;; Save the registers we'll use in the movem process \n\
89 ;; Now we've got this: \n\
94 ;; Update n for the first loop \n\
97 movem [$r11+],$r10 \n\
100 movem $r10,[$r13+] \n\
102 addq 44,$r12 ;; compensate for last loop underflowing n \n\
104 ;; Restore registers from stack \n\
105 movem [$sp+],$r10 \n\
107 .section .fixup,\"ax\" \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\
118 .section __ex_table,\"a\" \n\
122 /* Outputs */ : "=r" (dst), "=r" (src), "=r" (n), "=r" (retn)
123 /* Inputs */ : "0" (dst), "1" (src), "2" (n), "3" (retn));
129 __asm_copy_to_user_16 (dst, src, retn);
133 /* Having a separate by-four loops cuts down on cache footprint.
134 FIXME: Test with and without; increasing switch to be 0..15. */
137 __asm_copy_to_user_4 (dst, src, retn);
146 __asm_copy_to_user_1 (dst, src, retn);
149 __asm_copy_to_user_2 (dst, src, retn);
152 __asm_copy_to_user_3 (dst, src, retn);
159 /* Copy from user to kernel, zeroing the bytes that were inaccessible in
160 userland. The return-value is the number of bytes that were
164 __copy_user_zeroing(void *pdst, const void __user *psrc, unsigned long pn)
166 /* We want the parameters put in special registers.
167 Make sure the compiler is able to make something useful of this.
168 As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop).
170 FIXME: Comment for old gcc version. Check.
171 If gcc was alright, it really would need no temporaries, and no
172 stack space to save stuff on. */
174 register char *dst __asm__ ("r13") = pdst;
175 register const char *src __asm__ ("r11") = psrc;
176 register int n __asm__ ("r12") = pn;
177 register int retn __asm__ ("r10") = 0;
179 /* The best reason to align src is that we then know that a read-fault
180 was for aligned bytes; there's no 1..3 remaining good bytes to
182 if (((unsigned long) src & 3) != 0)
184 if (((unsigned long) src & 1) && n != 0)
186 __asm_copy_from_user_1 (dst, src, retn);
190 if (((unsigned long) src & 2) && n >= 2)
192 __asm_copy_from_user_2 (dst, src, retn);
196 /* We only need one check after the unalignment-adjustments, because
197 if both adjustments were done, either both or neither reference
200 goto copy_exception_bytes;
203 /* Movem is dirt cheap. The overheap is low enough to always use the
204 minimum possible block size as the threshold. */
207 /* It is not optimal to tell the compiler about clobbering any
208 registers; that will move the saving/restoring of those registers
209 to the function prologue/epilogue, and make non-movem sizes
212 .ifnc %0%1%2%3,$r13$r11$r12$r10 \n\
216 ;; Save the registers we'll use in the movem process \n\
221 ;; Now we've got this: \n\
226 ;; Update n for the first loop \n\
229 movem [$r11+],$r10 \n\
233 movem $r10,[$r13+] \n\
236 addq 44,$r12 ;; compensate for last loop underflowing n \n\
238 ;; Restore registers from stack \n\
239 movem [$sp+],$r10 \n\
240 .section .fixup,\"ax\" \n\
242 ;; Do not jump back into the loop if we fail. For some uses, we get a \n\
243 ;; page fault somewhere on the line. Without checking for page limits, \n\
244 ;; we don't know where, but we need to copy accurately and keep an \n\
245 ;; accurate count; not just clear the whole line. To do that, we fall \n\
246 ;; down in the code below, proceeding with smaller amounts. It should \n\
247 ;; be kept in mind that we have to cater to code like what at one time \n\
248 ;; was in fs/super.c: \n\
249 ;; i = size - copy_from_user((void *)page, data, size); \n\
250 ;; which would cause repeated faults while clearing the remainder of \n\
251 ;; the SIZE bytes at PAGE after the first fault. \n\
252 ;; A caveat here is that we must not fall through from a failing page \n\
253 ;; to a valid page. \n\
256 jump 4b ;; Fall through, pretending the fault didn't happen. \n\
260 .section __ex_table,\"a\" \n\
264 /* Outputs */ : "=r" (dst), "=r" (src), "=r" (n), "=r" (retn)
265 /* Inputs */ : "0" (dst), "1" (src), "2" (n), "3" (retn));
268 /* Either we directly start copying here, using dword copying in a loop,
269 or we copy as much as possible with 'movem' and then the last block
270 (<44 bytes) is copied here. This will work since 'movem' will have
271 updated src, dst and n. (Except with failing src.)
273 Since we want to keep src accurate, we can't use
274 __asm_copy_from_user_N with N != (1, 2, 4); it updates dst and
275 retn, but not src (by design; it's value is ignored elsewhere). */
279 __asm_copy_from_user_4 (dst, src, retn);
283 goto copy_exception_bytes;
286 /* If we get here, there were no memory read faults. */
289 /* These copies are at least "naturally aligned" (so we don't have
290 to check each byte), due to the src alignment code before the
291 movem loop. The *_3 case *will* get the correct count for retn. */
293 /* This case deliberately left in (if you have doubts check the
294 generated assembly code). */
297 __asm_copy_from_user_1 (dst, src, retn);
300 __asm_copy_from_user_2 (dst, src, retn);
303 __asm_copy_from_user_3 (dst, src, retn);
307 /* If we get here, retn correctly reflects the number of failing
311 copy_exception_bytes:
312 /* We already have "retn" bytes cleared, and need to clear the
313 remaining "n" bytes. A non-optimized simple byte-for-byte in-line
314 memset is preferred here, since this isn't speed-critical code and
315 we'd rather have this a leaf-function than calling memset. */
318 for (endp = dst + n; dst < endp; dst++)
325 /* Zero userspace. */
328 __do_clear_user (void __user *pto, unsigned long pn)
330 /* We want the parameters put in special registers.
331 Make sure the compiler is able to make something useful of this.
332 As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop).
334 FIXME: Comment for old gcc version. Check.
335 If gcc was alright, it really would need no temporaries, and no
336 stack space to save stuff on. */
338 register char *dst __asm__ ("r13") = pto;
339 register int n __asm__ ("r12") = pn;
340 register int retn __asm__ ("r10") = 0;
343 if (((unsigned long) dst & 3) != 0
344 /* Don't align if we wouldn't copy more than a few bytes. */
347 if ((unsigned long) dst & 1)
349 __asm_clear_1 (dst, retn);
353 if ((unsigned long) dst & 2)
355 __asm_clear_2 (dst, retn);
360 /* Decide which copying method to use.
361 FIXME: This number is from the "ordinary" kernel memset. */
364 /* For large clears we use 'movem' */
366 /* It is not optimal to tell the compiler about clobbering any
367 call-saved registers; that will move the saving/restoring of
368 those registers to the function prologue/epilogue, and make
369 non-movem sizes suboptimal.
371 This method is not foolproof; it assumes that the "asm reg"
372 declarations at the beginning of the function really are used
373 here (beware: they may be moved to temporary registers).
374 This way, we do not have to save/move the registers around into
375 temporaries; we can safely use them straight away.
377 If you want to check that the allocation was right; then
378 check the equalities in the first comment. It should say
379 something like "r13=r13, r11=r11, r12=r12". */
381 .ifnc %0%1%2,$r13$r12$r10 \n\
385 ;; Save the registers we'll clobber in the movem process \n\
386 ;; on the stack. Don't mention them to gcc, it will only be \n\
404 ;; Now we've got this: \n\
408 ;; Update n for the first loop \n\
414 movem $r11,[$r13+] \n\
416 addq 12*4,$r12 ;; compensate for last loop underflowing n \n\
418 ;; Restore registers from stack \n\
419 movem [$sp+],$r10 \n\
421 .section .fixup,\"ax\" \n\
431 .section __ex_table,\"a\" \n\
435 /* Outputs */ : "=r" (dst), "=r" (n), "=r" (retn)
436 /* Inputs */ : "0" (dst), "1" (n), "2" (retn)
437 /* Clobber */ : "r11");
442 __asm_clear_16 (dst, retn);
446 /* Having a separate by-four loops cuts down on cache footprint.
447 FIXME: Test with and without; increasing switch to be 0..15. */
450 __asm_clear_4 (dst, retn);
459 __asm_clear_1 (dst, retn);
462 __asm_clear_2 (dst, retn);
465 __asm_clear_3 (dst, retn);