[POWERPC] Handle alignment faults on new FP load/store instructions
[linux-2.6/btrfs-unstable.git] / arch / powerpc / kernel / align.c
blob4c47f9cc0d9acbcadcdf561bb3a015e4ce4fd0e4
1 /* align.c - handle alignment exceptions for the Power PC.
3 * Copyright (c) 1996 Paul Mackerras <paulus@cs.anu.edu.au>
4 * Copyright (c) 1998-1999 TiVo, Inc.
5 * PowerPC 403GCX modifications.
6 * Copyright (c) 1999 Grant Erickson <grant@lcse.umn.edu>
7 * PowerPC 403GCX/405GP modifications.
8 * Copyright (c) 2001-2002 PPC64 team, IBM Corp
9 * 64-bit and Power4 support
10 * Copyright (c) 2005 Benjamin Herrenschmidt, IBM Corp
11 * <benh@kernel.crashing.org>
12 * Merge ppc32 and ppc64 implementations
14 * This program is free software; you can redistribute it and/or
15 * modify it under the terms of the GNU General Public License
16 * as published by the Free Software Foundation; either version
17 * 2 of the License, or (at your option) any later version.
20 #include <linux/kernel.h>
21 #include <linux/mm.h>
22 #include <asm/processor.h>
23 #include <asm/uaccess.h>
24 #include <asm/system.h>
25 #include <asm/cache.h>
26 #include <asm/cputable.h>
28 struct aligninfo {
29 unsigned char len;
30 unsigned char flags;
33 #define IS_XFORM(inst) (((inst) >> 26) == 31)
34 #define IS_DSFORM(inst) (((inst) >> 26) >= 56)
36 #define INVALID { 0, 0 }
38 /* Bits in the flags field */
39 #define LD 0 /* load */
40 #define ST 1 /* store */
41 #define SE 2 /* sign-extend value, or FP ld/st as word */
42 #define F 4 /* to/from fp regs */
43 #define U 8 /* update index register */
44 #define M 0x10 /* multiple load/store */
45 #define SW 0x20 /* byte swap */
46 #define S 0x40 /* single-precision fp or... */
47 #define SX 0x40 /* ... byte count in XER */
48 #define HARD 0x80 /* string, stwcx. */
50 /* DSISR bits reported for a DCBZ instruction: */
51 #define DCBZ 0x5f /* 8xx/82xx dcbz faults when cache not enabled */
53 #define SWAP(a, b) (t = (a), (a) = (b), (b) = t)
56 * The PowerPC stores certain bits of the instruction that caused the
57 * alignment exception in the DSISR register. This array maps those
58 * bits to information about the operand length and what the
59 * instruction would do.
61 static struct aligninfo aligninfo[128] = {
62 { 4, LD }, /* 00 0 0000: lwz / lwarx */
63 INVALID, /* 00 0 0001 */
64 { 4, ST }, /* 00 0 0010: stw */
65 INVALID, /* 00 0 0011 */
66 { 2, LD }, /* 00 0 0100: lhz */
67 { 2, LD+SE }, /* 00 0 0101: lha */
68 { 2, ST }, /* 00 0 0110: sth */
69 { 4, LD+M }, /* 00 0 0111: lmw */
70 { 4, LD+F+S }, /* 00 0 1000: lfs */
71 { 8, LD+F }, /* 00 0 1001: lfd */
72 { 4, ST+F+S }, /* 00 0 1010: stfs */
73 { 8, ST+F }, /* 00 0 1011: stfd */
74 INVALID, /* 00 0 1100 */
75 { 8, LD }, /* 00 0 1101: ld/ldu/lwa */
76 INVALID, /* 00 0 1110 */
77 { 8, ST }, /* 00 0 1111: std/stdu */
78 { 4, LD+U }, /* 00 1 0000: lwzu */
79 INVALID, /* 00 1 0001 */
80 { 4, ST+U }, /* 00 1 0010: stwu */
81 INVALID, /* 00 1 0011 */
82 { 2, LD+U }, /* 00 1 0100: lhzu */
83 { 2, LD+SE+U }, /* 00 1 0101: lhau */
84 { 2, ST+U }, /* 00 1 0110: sthu */
85 { 4, ST+M }, /* 00 1 0111: stmw */
86 { 4, LD+F+S+U }, /* 00 1 1000: lfsu */
87 { 8, LD+F+U }, /* 00 1 1001: lfdu */
88 { 4, ST+F+S+U }, /* 00 1 1010: stfsu */
89 { 8, ST+F+U }, /* 00 1 1011: stfdu */
90 { 16, LD+F }, /* 00 1 1100: lfdp */
91 INVALID, /* 00 1 1101 */
92 { 16, ST+F }, /* 00 1 1110: stfdp */
93 INVALID, /* 00 1 1111 */
94 { 8, LD }, /* 01 0 0000: ldx */
95 INVALID, /* 01 0 0001 */
96 { 8, ST }, /* 01 0 0010: stdx */
97 INVALID, /* 01 0 0011 */
98 INVALID, /* 01 0 0100 */
99 { 4, LD+SE }, /* 01 0 0101: lwax */
100 INVALID, /* 01 0 0110 */
101 INVALID, /* 01 0 0111 */
102 { 4, LD+M+HARD+SX }, /* 01 0 1000: lswx */
103 { 4, LD+M+HARD }, /* 01 0 1001: lswi */
104 { 4, ST+M+HARD+SX }, /* 01 0 1010: stswx */
105 { 4, ST+M+HARD }, /* 01 0 1011: stswi */
106 INVALID, /* 01 0 1100 */
107 { 8, LD+U }, /* 01 0 1101: ldu */
108 INVALID, /* 01 0 1110 */
109 { 8, ST+U }, /* 01 0 1111: stdu */
110 { 8, LD+U }, /* 01 1 0000: ldux */
111 INVALID, /* 01 1 0001 */
112 { 8, ST+U }, /* 01 1 0010: stdux */
113 INVALID, /* 01 1 0011 */
114 INVALID, /* 01 1 0100 */
115 { 4, LD+SE+U }, /* 01 1 0101: lwaux */
116 INVALID, /* 01 1 0110 */
117 INVALID, /* 01 1 0111 */
118 INVALID, /* 01 1 1000 */
119 INVALID, /* 01 1 1001 */
120 INVALID, /* 01 1 1010 */
121 INVALID, /* 01 1 1011 */
122 INVALID, /* 01 1 1100 */
123 INVALID, /* 01 1 1101 */
124 INVALID, /* 01 1 1110 */
125 INVALID, /* 01 1 1111 */
126 INVALID, /* 10 0 0000 */
127 INVALID, /* 10 0 0001 */
128 INVALID, /* 10 0 0010: stwcx. */
129 INVALID, /* 10 0 0011 */
130 INVALID, /* 10 0 0100 */
131 INVALID, /* 10 0 0101 */
132 INVALID, /* 10 0 0110 */
133 INVALID, /* 10 0 0111 */
134 { 4, LD+SW }, /* 10 0 1000: lwbrx */
135 INVALID, /* 10 0 1001 */
136 { 4, ST+SW }, /* 10 0 1010: stwbrx */
137 INVALID, /* 10 0 1011 */
138 { 2, LD+SW }, /* 10 0 1100: lhbrx */
139 { 4, LD+SE }, /* 10 0 1101 lwa */
140 { 2, ST+SW }, /* 10 0 1110: sthbrx */
141 INVALID, /* 10 0 1111 */
142 INVALID, /* 10 1 0000 */
143 INVALID, /* 10 1 0001 */
144 INVALID, /* 10 1 0010 */
145 INVALID, /* 10 1 0011 */
146 INVALID, /* 10 1 0100 */
147 INVALID, /* 10 1 0101 */
148 INVALID, /* 10 1 0110 */
149 INVALID, /* 10 1 0111 */
150 INVALID, /* 10 1 1000 */
151 INVALID, /* 10 1 1001 */
152 INVALID, /* 10 1 1010 */
153 INVALID, /* 10 1 1011 */
154 INVALID, /* 10 1 1100 */
155 INVALID, /* 10 1 1101 */
156 INVALID, /* 10 1 1110 */
157 { 0, ST+HARD }, /* 10 1 1111: dcbz */
158 { 4, LD }, /* 11 0 0000: lwzx */
159 INVALID, /* 11 0 0001 */
160 { 4, ST }, /* 11 0 0010: stwx */
161 INVALID, /* 11 0 0011 */
162 { 2, LD }, /* 11 0 0100: lhzx */
163 { 2, LD+SE }, /* 11 0 0101: lhax */
164 { 2, ST }, /* 11 0 0110: sthx */
165 INVALID, /* 11 0 0111 */
166 { 4, LD+F+S }, /* 11 0 1000: lfsx */
167 { 8, LD+F }, /* 11 0 1001: lfdx */
168 { 4, ST+F+S }, /* 11 0 1010: stfsx */
169 { 8, ST+F }, /* 11 0 1011: stfdx */
170 { 16, LD+F }, /* 11 0 1100: lfdpx */
171 { 4, LD+F+SE }, /* 11 0 1101: lfiwax */
172 { 16, ST+F }, /* 11 0 1110: stfdpx */
173 { 4, ST+F }, /* 11 0 1111: stfiwx */
174 { 4, LD+U }, /* 11 1 0000: lwzux */
175 INVALID, /* 11 1 0001 */
176 { 4, ST+U }, /* 11 1 0010: stwux */
177 INVALID, /* 11 1 0011 */
178 { 2, LD+U }, /* 11 1 0100: lhzux */
179 { 2, LD+SE+U }, /* 11 1 0101: lhaux */
180 { 2, ST+U }, /* 11 1 0110: sthux */
181 INVALID, /* 11 1 0111 */
182 { 4, LD+F+S+U }, /* 11 1 1000: lfsux */
183 { 8, LD+F+U }, /* 11 1 1001: lfdux */
184 { 4, ST+F+S+U }, /* 11 1 1010: stfsux */
185 { 8, ST+F+U }, /* 11 1 1011: stfdux */
186 INVALID, /* 11 1 1100 */
187 INVALID, /* 11 1 1101 */
188 INVALID, /* 11 1 1110 */
189 INVALID, /* 11 1 1111 */
193 * Create a DSISR value from the instruction
195 static inline unsigned make_dsisr(unsigned instr)
197 unsigned dsisr;
200 /* bits 6:15 --> 22:31 */
201 dsisr = (instr & 0x03ff0000) >> 16;
203 if (IS_XFORM(instr)) {
204 /* bits 29:30 --> 15:16 */
205 dsisr |= (instr & 0x00000006) << 14;
206 /* bit 25 --> 17 */
207 dsisr |= (instr & 0x00000040) << 8;
208 /* bits 21:24 --> 18:21 */
209 dsisr |= (instr & 0x00000780) << 3;
210 } else {
211 /* bit 5 --> 17 */
212 dsisr |= (instr & 0x04000000) >> 12;
213 /* bits 1: 4 --> 18:21 */
214 dsisr |= (instr & 0x78000000) >> 17;
215 /* bits 30:31 --> 12:13 */
216 if (IS_DSFORM(instr))
217 dsisr |= (instr & 0x00000003) << 18;
220 return dsisr;
224 * The dcbz (data cache block zero) instruction
225 * gives an alignment fault if used on non-cacheable
226 * memory. We handle the fault mainly for the
227 * case when we are running with the cache disabled
228 * for debugging.
230 static int emulate_dcbz(struct pt_regs *regs, unsigned char __user *addr)
232 long __user *p;
233 int i, size;
235 #ifdef __powerpc64__
236 size = ppc64_caches.dline_size;
237 #else
238 size = L1_CACHE_BYTES;
239 #endif
240 p = (long __user *) (regs->dar & -size);
241 if (user_mode(regs) && !access_ok(VERIFY_WRITE, p, size))
242 return -EFAULT;
243 for (i = 0; i < size / sizeof(long); ++i)
244 if (__put_user_inatomic(0, p+i))
245 return -EFAULT;
246 return 1;
250 * Emulate load & store multiple instructions
251 * On 64-bit machines, these instructions only affect/use the
252 * bottom 4 bytes of each register, and the loads clear the
253 * top 4 bytes of the affected register.
255 #ifdef CONFIG_PPC64
256 #define REG_BYTE(rp, i) *((u8 *)((rp) + ((i) >> 2)) + ((i) & 3) + 4)
257 #else
258 #define REG_BYTE(rp, i) *((u8 *)(rp) + (i))
259 #endif
261 #define SWIZ_PTR(p) ((unsigned char __user *)((p) ^ swiz))
263 static int emulate_multiple(struct pt_regs *regs, unsigned char __user *addr,
264 unsigned int reg, unsigned int nb,
265 unsigned int flags, unsigned int instr,
266 unsigned long swiz)
268 unsigned long *rptr;
269 unsigned int nb0, i, bswiz;
270 unsigned long p;
273 * We do not try to emulate 8 bytes multiple as they aren't really
274 * available in our operating environments and we don't try to
275 * emulate multiples operations in kernel land as they should never
276 * be used/generated there at least not on unaligned boundaries
278 if (unlikely((nb > 4) || !user_mode(regs)))
279 return 0;
281 /* lmw, stmw, lswi/x, stswi/x */
282 nb0 = 0;
283 if (flags & HARD) {
284 if (flags & SX) {
285 nb = regs->xer & 127;
286 if (nb == 0)
287 return 1;
288 } else {
289 unsigned long pc = regs->nip ^ (swiz & 4);
291 if (__get_user_inatomic(instr,
292 (unsigned int __user *)pc))
293 return -EFAULT;
294 if (swiz == 0 && (flags & SW))
295 instr = cpu_to_le32(instr);
296 nb = (instr >> 11) & 0x1f;
297 if (nb == 0)
298 nb = 32;
300 if (nb + reg * 4 > 128) {
301 nb0 = nb + reg * 4 - 128;
302 nb = 128 - reg * 4;
304 } else {
305 /* lwm, stmw */
306 nb = (32 - reg) * 4;
309 if (!access_ok((flags & ST ? VERIFY_WRITE: VERIFY_READ), addr, nb+nb0))
310 return -EFAULT; /* bad address */
312 rptr = &regs->gpr[reg];
313 p = (unsigned long) addr;
314 bswiz = (flags & SW)? 3: 0;
316 if (!(flags & ST)) {
318 * This zeroes the top 4 bytes of the affected registers
319 * in 64-bit mode, and also zeroes out any remaining
320 * bytes of the last register for lsw*.
322 memset(rptr, 0, ((nb + 3) / 4) * sizeof(unsigned long));
323 if (nb0 > 0)
324 memset(&regs->gpr[0], 0,
325 ((nb0 + 3) / 4) * sizeof(unsigned long));
327 for (i = 0; i < nb; ++i, ++p)
328 if (__get_user_inatomic(REG_BYTE(rptr, i ^ bswiz),
329 SWIZ_PTR(p)))
330 return -EFAULT;
331 if (nb0 > 0) {
332 rptr = &regs->gpr[0];
333 addr += nb;
334 for (i = 0; i < nb0; ++i, ++p)
335 if (__get_user_inatomic(REG_BYTE(rptr,
336 i ^ bswiz),
337 SWIZ_PTR(p)))
338 return -EFAULT;
341 } else {
342 for (i = 0; i < nb; ++i, ++p)
343 if (__put_user_inatomic(REG_BYTE(rptr, i ^ bswiz),
344 SWIZ_PTR(p)))
345 return -EFAULT;
346 if (nb0 > 0) {
347 rptr = &regs->gpr[0];
348 addr += nb;
349 for (i = 0; i < nb0; ++i, ++p)
350 if (__put_user_inatomic(REG_BYTE(rptr,
351 i ^ bswiz),
352 SWIZ_PTR(p)))
353 return -EFAULT;
356 return 1;
360 * Emulate floating-point pair loads and stores.
361 * Only POWER6 has these instructions, and it does true little-endian,
362 * so we don't need the address swizzling.
364 static int emulate_fp_pair(struct pt_regs *regs, unsigned char __user *addr,
365 unsigned int reg, unsigned int flags)
367 char *ptr = (char *) &current->thread.fpr[reg];
368 int i, ret;
370 if (!(flags & F))
371 return 0;
372 if (reg & 1)
373 return 0; /* invalid form: FRS/FRT must be even */
374 if (!(flags & SW)) {
375 /* not byte-swapped - easy */
376 if (!(flags & ST))
377 ret = __copy_from_user(ptr, addr, 16);
378 else
379 ret = __copy_to_user(addr, ptr, 16);
380 } else {
381 /* each FPR value is byte-swapped separately */
382 ret = 0;
383 for (i = 0; i < 16; ++i) {
384 if (!(flags & ST))
385 ret |= __get_user(ptr[i^7], addr + i);
386 else
387 ret |= __put_user(ptr[i^7], addr + i);
390 if (ret)
391 return -EFAULT;
392 return 1; /* exception handled and fixed up */
397 * Called on alignment exception. Attempts to fixup
399 * Return 1 on success
400 * Return 0 if unable to handle the interrupt
401 * Return -EFAULT if data address is bad
404 int fix_alignment(struct pt_regs *regs)
406 unsigned int instr, nb, flags;
407 unsigned int reg, areg;
408 unsigned int dsisr;
409 unsigned char __user *addr;
410 unsigned long p, swiz;
411 int ret, t;
412 union {
413 u64 ll;
414 double dd;
415 unsigned char v[8];
416 struct {
417 unsigned hi32;
418 int low32;
419 } x32;
420 struct {
421 unsigned char hi48[6];
422 short low16;
423 } x16;
424 } data;
427 * We require a complete register set, if not, then our assembly
428 * is broken
430 CHECK_FULL_REGS(regs);
432 dsisr = regs->dsisr;
434 /* Some processors don't provide us with a DSISR we can use here,
435 * let's make one up from the instruction
437 if (cpu_has_feature(CPU_FTR_NODSISRALIGN)) {
438 unsigned long pc = regs->nip;
440 if (cpu_has_feature(CPU_FTR_PPC_LE) && (regs->msr & MSR_LE))
441 pc ^= 4;
442 if (unlikely(__get_user_inatomic(instr,
443 (unsigned int __user *)pc)))
444 return -EFAULT;
445 if (cpu_has_feature(CPU_FTR_REAL_LE) && (regs->msr & MSR_LE))
446 instr = cpu_to_le32(instr);
447 dsisr = make_dsisr(instr);
450 /* extract the operation and registers from the dsisr */
451 reg = (dsisr >> 5) & 0x1f; /* source/dest register */
452 areg = dsisr & 0x1f; /* register to update */
453 instr = (dsisr >> 10) & 0x7f;
454 instr |= (dsisr >> 13) & 0x60;
456 /* Lookup the operation in our table */
457 nb = aligninfo[instr].len;
458 flags = aligninfo[instr].flags;
460 /* Byteswap little endian loads and stores */
461 swiz = 0;
462 if (regs->msr & MSR_LE) {
463 flags ^= SW;
465 * So-called "PowerPC little endian" mode works by
466 * swizzling addresses rather than by actually doing
467 * any byte-swapping. To emulate this, we XOR each
468 * byte address with 7. We also byte-swap, because
469 * the processor's address swizzling depends on the
470 * operand size (it xors the address with 7 for bytes,
471 * 6 for halfwords, 4 for words, 0 for doublewords) but
472 * we will xor with 7 and load/store each byte separately.
474 if (cpu_has_feature(CPU_FTR_PPC_LE))
475 swiz = 7;
478 /* DAR has the operand effective address */
479 addr = (unsigned char __user *)regs->dar;
481 /* A size of 0 indicates an instruction we don't support, with
482 * the exception of DCBZ which is handled as a special case here
484 if (instr == DCBZ)
485 return emulate_dcbz(regs, addr);
486 if (unlikely(nb == 0))
487 return 0;
489 /* Load/Store Multiple instructions are handled in their own
490 * function
492 if (flags & M)
493 return emulate_multiple(regs, addr, reg, nb,
494 flags, instr, swiz);
496 /* Verify the address of the operand */
497 if (unlikely(user_mode(regs) &&
498 !access_ok((flags & ST ? VERIFY_WRITE : VERIFY_READ),
499 addr, nb)))
500 return -EFAULT;
502 /* Force the fprs into the save area so we can reference them */
503 if (flags & F) {
504 /* userland only */
505 if (unlikely(!user_mode(regs)))
506 return 0;
507 flush_fp_to_thread(current);
510 /* Special case for 16-byte FP loads and stores */
511 if (nb == 16)
512 return emulate_fp_pair(regs, addr, reg, flags);
514 /* If we are loading, get the data from user space, else
515 * get it from register values
517 if (!(flags & ST)) {
518 data.ll = 0;
519 ret = 0;
520 p = (unsigned long) addr;
521 switch (nb) {
522 case 8:
523 ret |= __get_user_inatomic(data.v[0], SWIZ_PTR(p++));
524 ret |= __get_user_inatomic(data.v[1], SWIZ_PTR(p++));
525 ret |= __get_user_inatomic(data.v[2], SWIZ_PTR(p++));
526 ret |= __get_user_inatomic(data.v[3], SWIZ_PTR(p++));
527 case 4:
528 ret |= __get_user_inatomic(data.v[4], SWIZ_PTR(p++));
529 ret |= __get_user_inatomic(data.v[5], SWIZ_PTR(p++));
530 case 2:
531 ret |= __get_user_inatomic(data.v[6], SWIZ_PTR(p++));
532 ret |= __get_user_inatomic(data.v[7], SWIZ_PTR(p++));
533 if (unlikely(ret))
534 return -EFAULT;
536 } else if (flags & F) {
537 data.dd = current->thread.fpr[reg];
538 if (flags & S) {
539 /* Single-precision FP store requires conversion... */
540 #ifdef CONFIG_PPC_FPU
541 preempt_disable();
542 enable_kernel_fp();
543 cvt_df(&data.dd, (float *)&data.v[4], &current->thread);
544 preempt_enable();
545 #else
546 return 0;
547 #endif
549 } else
550 data.ll = regs->gpr[reg];
552 if (flags & SW) {
553 switch (nb) {
554 case 8:
555 SWAP(data.v[0], data.v[7]);
556 SWAP(data.v[1], data.v[6]);
557 SWAP(data.v[2], data.v[5]);
558 SWAP(data.v[3], data.v[4]);
559 break;
560 case 4:
561 SWAP(data.v[4], data.v[7]);
562 SWAP(data.v[5], data.v[6]);
563 break;
564 case 2:
565 SWAP(data.v[6], data.v[7]);
566 break;
570 /* Perform other misc operations like sign extension
571 * or floating point single precision conversion
573 switch (flags & ~(U|SW)) {
574 case LD+SE: /* sign extending integer loads */
575 case LD+F+SE: /* sign extend for lfiwax */
576 if ( nb == 2 )
577 data.ll = data.x16.low16;
578 else /* nb must be 4 */
579 data.ll = data.x32.low32;
580 break;
582 /* Single-precision FP load requires conversion... */
583 case LD+F+S:
584 #ifdef CONFIG_PPC_FPU
585 preempt_disable();
586 enable_kernel_fp();
587 cvt_fd((float *)&data.v[4], &data.dd, &current->thread);
588 preempt_enable();
589 #else
590 return 0;
591 #endif
592 break;
595 /* Store result to memory or update registers */
596 if (flags & ST) {
597 ret = 0;
598 p = (unsigned long) addr;
599 switch (nb) {
600 case 8:
601 ret |= __put_user_inatomic(data.v[0], SWIZ_PTR(p++));
602 ret |= __put_user_inatomic(data.v[1], SWIZ_PTR(p++));
603 ret |= __put_user_inatomic(data.v[2], SWIZ_PTR(p++));
604 ret |= __put_user_inatomic(data.v[3], SWIZ_PTR(p++));
605 case 4:
606 ret |= __put_user_inatomic(data.v[4], SWIZ_PTR(p++));
607 ret |= __put_user_inatomic(data.v[5], SWIZ_PTR(p++));
608 case 2:
609 ret |= __put_user_inatomic(data.v[6], SWIZ_PTR(p++));
610 ret |= __put_user_inatomic(data.v[7], SWIZ_PTR(p++));
612 if (unlikely(ret))
613 return -EFAULT;
614 } else if (flags & F)
615 current->thread.fpr[reg] = data.dd;
616 else
617 regs->gpr[reg] = data.ll;
619 /* Update RA as needed */
620 if (flags & U)
621 regs->gpr[areg] = regs->dar;
623 return 1;