xref: /linux/arch/powerpc/kernel/align.c (revision 0883c2c06fb5bcf5b9e008270827e63c09a88c1e)
1 /* align.c - handle alignment exceptions for the Power PC.
2  *
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
13  *
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.
18  */
19 
20 #include <linux/kernel.h>
21 #include <linux/mm.h>
22 #include <asm/processor.h>
23 #include <asm/uaccess.h>
24 #include <asm/cache.h>
25 #include <asm/cputable.h>
26 #include <asm/emulated_ops.h>
27 #include <asm/switch_to.h>
28 #include <asm/disassemble.h>
29 
30 struct aligninfo {
31 	unsigned char len;
32 	unsigned char flags;
33 };
34 
35 
36 #define INVALID	{ 0, 0 }
37 
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. */
49 #define E4	0x40	/* SPE endianness is word */
50 #define E8	0x80	/* SPE endianness is double word */
51 #define SPLT	0x80	/* VSX SPLAT load */
52 
53 /* DSISR bits reported for a DCBZ instruction: */
54 #define DCBZ	0x5f	/* 8xx/82xx dcbz faults when cache not enabled */
55 
56 /*
57  * The PowerPC stores certain bits of the instruction that caused the
58  * alignment exception in the DSISR register.  This array maps those
59  * bits to information about the operand length and what the
60  * instruction would do.
61  */
62 static struct aligninfo aligninfo[128] = {
63 	{ 4, LD },		/* 00 0 0000: lwz / lwarx */
64 	INVALID,		/* 00 0 0001 */
65 	{ 4, ST },		/* 00 0 0010: stw */
66 	INVALID,		/* 00 0 0011 */
67 	{ 2, LD },		/* 00 0 0100: lhz */
68 	{ 2, LD+SE },		/* 00 0 0101: lha */
69 	{ 2, ST },		/* 00 0 0110: sth */
70 	{ 4, LD+M },		/* 00 0 0111: lmw */
71 	{ 4, LD+F+S },		/* 00 0 1000: lfs */
72 	{ 8, LD+F },		/* 00 0 1001: lfd */
73 	{ 4, ST+F+S },		/* 00 0 1010: stfs */
74 	{ 8, ST+F },		/* 00 0 1011: stfd */
75 	{ 16, LD },		/* 00 0 1100: lq */
76 	{ 8, LD },		/* 00 0 1101: ld/ldu/lwa */
77 	INVALID,		/* 00 0 1110 */
78 	{ 8, ST },		/* 00 0 1111: std/stdu */
79 	{ 4, LD+U },		/* 00 1 0000: lwzu */
80 	INVALID,		/* 00 1 0001 */
81 	{ 4, ST+U },		/* 00 1 0010: stwu */
82 	INVALID,		/* 00 1 0011 */
83 	{ 2, LD+U },		/* 00 1 0100: lhzu */
84 	{ 2, LD+SE+U },		/* 00 1 0101: lhau */
85 	{ 2, ST+U },		/* 00 1 0110: sthu */
86 	{ 4, ST+M },		/* 00 1 0111: stmw */
87 	{ 4, LD+F+S+U },	/* 00 1 1000: lfsu */
88 	{ 8, LD+F+U },		/* 00 1 1001: lfdu */
89 	{ 4, ST+F+S+U },	/* 00 1 1010: stfsu */
90 	{ 8, ST+F+U },		/* 00 1 1011: stfdu */
91 	{ 16, LD+F },		/* 00 1 1100: lfdp */
92 	INVALID,		/* 00 1 1101 */
93 	{ 16, ST+F },		/* 00 1 1110: stfdp */
94 	INVALID,		/* 00 1 1111 */
95 	{ 8, LD },		/* 01 0 0000: ldx */
96 	INVALID,		/* 01 0 0001 */
97 	{ 8, ST },		/* 01 0 0010: stdx */
98 	INVALID,		/* 01 0 0011 */
99 	INVALID,		/* 01 0 0100 */
100 	{ 4, LD+SE },		/* 01 0 0101: lwax */
101 	INVALID,		/* 01 0 0110 */
102 	INVALID,		/* 01 0 0111 */
103 	{ 4, LD+M+HARD+SX },	/* 01 0 1000: lswx */
104 	{ 4, LD+M+HARD },	/* 01 0 1001: lswi */
105 	{ 4, ST+M+HARD+SX },	/* 01 0 1010: stswx */
106 	{ 4, ST+M+HARD },	/* 01 0 1011: stswi */
107 	INVALID,		/* 01 0 1100 */
108 	{ 8, LD+U },		/* 01 0 1101: ldu */
109 	INVALID,		/* 01 0 1110 */
110 	{ 8, ST+U },		/* 01 0 1111: stdu */
111 	{ 8, LD+U },		/* 01 1 0000: ldux */
112 	INVALID,		/* 01 1 0001 */
113 	{ 8, ST+U },		/* 01 1 0010: stdux */
114 	INVALID,		/* 01 1 0011 */
115 	INVALID,		/* 01 1 0100 */
116 	{ 4, LD+SE+U },		/* 01 1 0101: lwaux */
117 	INVALID,		/* 01 1 0110 */
118 	INVALID,		/* 01 1 0111 */
119 	INVALID,		/* 01 1 1000 */
120 	INVALID,		/* 01 1 1001 */
121 	INVALID,		/* 01 1 1010 */
122 	INVALID,		/* 01 1 1011 */
123 	INVALID,		/* 01 1 1100 */
124 	INVALID,		/* 01 1 1101 */
125 	INVALID,		/* 01 1 1110 */
126 	INVALID,		/* 01 1 1111 */
127 	INVALID,		/* 10 0 0000 */
128 	INVALID,		/* 10 0 0001 */
129 	INVALID,		/* 10 0 0010: stwcx. */
130 	INVALID,		/* 10 0 0011 */
131 	INVALID,		/* 10 0 0100 */
132 	INVALID,		/* 10 0 0101 */
133 	INVALID,		/* 10 0 0110 */
134 	INVALID,		/* 10 0 0111 */
135 	{ 4, LD+SW },		/* 10 0 1000: lwbrx */
136 	INVALID,		/* 10 0 1001 */
137 	{ 4, ST+SW },		/* 10 0 1010: stwbrx */
138 	INVALID,		/* 10 0 1011 */
139 	{ 2, LD+SW },		/* 10 0 1100: lhbrx */
140 	{ 4, LD+SE },		/* 10 0 1101  lwa */
141 	{ 2, ST+SW },		/* 10 0 1110: sthbrx */
142 	{ 16, ST },		/* 10 0 1111: stq */
143 	INVALID,		/* 10 1 0000 */
144 	INVALID,		/* 10 1 0001 */
145 	INVALID,		/* 10 1 0010 */
146 	INVALID,		/* 10 1 0011 */
147 	INVALID,		/* 10 1 0100 */
148 	INVALID,		/* 10 1 0101 */
149 	INVALID,		/* 10 1 0110 */
150 	INVALID,		/* 10 1 0111 */
151 	INVALID,		/* 10 1 1000 */
152 	INVALID,		/* 10 1 1001 */
153 	INVALID,		/* 10 1 1010 */
154 	INVALID,		/* 10 1 1011 */
155 	INVALID,		/* 10 1 1100 */
156 	INVALID,		/* 10 1 1101 */
157 	INVALID,		/* 10 1 1110 */
158 	{ 0, ST+HARD },		/* 10 1 1111: dcbz */
159 	{ 4, LD },		/* 11 0 0000: lwzx */
160 	INVALID,		/* 11 0 0001 */
161 	{ 4, ST },		/* 11 0 0010: stwx */
162 	INVALID,		/* 11 0 0011 */
163 	{ 2, LD },		/* 11 0 0100: lhzx */
164 	{ 2, LD+SE },		/* 11 0 0101: lhax */
165 	{ 2, ST },		/* 11 0 0110: sthx */
166 	INVALID,		/* 11 0 0111 */
167 	{ 4, LD+F+S },		/* 11 0 1000: lfsx */
168 	{ 8, LD+F },		/* 11 0 1001: lfdx */
169 	{ 4, ST+F+S },		/* 11 0 1010: stfsx */
170 	{ 8, ST+F },		/* 11 0 1011: stfdx */
171 	{ 16, LD+F },		/* 11 0 1100: lfdpx */
172 	{ 4, LD+F+SE },		/* 11 0 1101: lfiwax */
173 	{ 16, ST+F },		/* 11 0 1110: stfdpx */
174 	{ 4, ST+F },		/* 11 0 1111: stfiwx */
175 	{ 4, LD+U },		/* 11 1 0000: lwzux */
176 	INVALID,		/* 11 1 0001 */
177 	{ 4, ST+U },		/* 11 1 0010: stwux */
178 	INVALID,		/* 11 1 0011 */
179 	{ 2, LD+U },		/* 11 1 0100: lhzux */
180 	{ 2, LD+SE+U },		/* 11 1 0101: lhaux */
181 	{ 2, ST+U },		/* 11 1 0110: sthux */
182 	INVALID,		/* 11 1 0111 */
183 	{ 4, LD+F+S+U },	/* 11 1 1000: lfsux */
184 	{ 8, LD+F+U },		/* 11 1 1001: lfdux */
185 	{ 4, ST+F+S+U },	/* 11 1 1010: stfsux */
186 	{ 8, ST+F+U },		/* 11 1 1011: stfdux */
187 	INVALID,		/* 11 1 1100 */
188 	{ 4, LD+F },		/* 11 1 1101: lfiwzx */
189 	INVALID,		/* 11 1 1110 */
190 	INVALID,		/* 11 1 1111 */
191 };
192 
193 /*
194  * The dcbz (data cache block zero) instruction
195  * gives an alignment fault if used on non-cacheable
196  * memory.  We handle the fault mainly for the
197  * case when we are running with the cache disabled
198  * for debugging.
199  */
200 static int emulate_dcbz(struct pt_regs *regs, unsigned char __user *addr)
201 {
202 	long __user *p;
203 	int i, size;
204 
205 #ifdef __powerpc64__
206 	size = ppc64_caches.dline_size;
207 #else
208 	size = L1_CACHE_BYTES;
209 #endif
210 	p = (long __user *) (regs->dar & -size);
211 	if (user_mode(regs) && !access_ok(VERIFY_WRITE, p, size))
212 		return -EFAULT;
213 	for (i = 0; i < size / sizeof(long); ++i)
214 		if (__put_user_inatomic(0, p+i))
215 			return -EFAULT;
216 	return 1;
217 }
218 
219 /*
220  * Emulate load & store multiple instructions
221  * On 64-bit machines, these instructions only affect/use the
222  * bottom 4 bytes of each register, and the loads clear the
223  * top 4 bytes of the affected register.
224  */
225 #ifdef __BIG_ENDIAN__
226 #ifdef CONFIG_PPC64
227 #define REG_BYTE(rp, i)		*((u8 *)((rp) + ((i) >> 2)) + ((i) & 3) + 4)
228 #else
229 #define REG_BYTE(rp, i)		*((u8 *)(rp) + (i))
230 #endif
231 #else
232 #define REG_BYTE(rp, i)		(*(((u8 *)((rp) + ((i)>>2)) + ((i)&3))))
233 #endif
234 
235 #define SWIZ_PTR(p)		((unsigned char __user *)((p) ^ swiz))
236 
237 static int emulate_multiple(struct pt_regs *regs, unsigned char __user *addr,
238 			    unsigned int reg, unsigned int nb,
239 			    unsigned int flags, unsigned int instr,
240 			    unsigned long swiz)
241 {
242 	unsigned long *rptr;
243 	unsigned int nb0, i, bswiz;
244 	unsigned long p;
245 
246 	/*
247 	 * We do not try to emulate 8 bytes multiple as they aren't really
248 	 * available in our operating environments and we don't try to
249 	 * emulate multiples operations in kernel land as they should never
250 	 * be used/generated there at least not on unaligned boundaries
251 	 */
252 	if (unlikely((nb > 4) || !user_mode(regs)))
253 		return 0;
254 
255 	/* lmw, stmw, lswi/x, stswi/x */
256 	nb0 = 0;
257 	if (flags & HARD) {
258 		if (flags & SX) {
259 			nb = regs->xer & 127;
260 			if (nb == 0)
261 				return 1;
262 		} else {
263 			unsigned long pc = regs->nip ^ (swiz & 4);
264 
265 			if (__get_user_inatomic(instr,
266 						(unsigned int __user *)pc))
267 				return -EFAULT;
268 			if (swiz == 0 && (flags & SW))
269 				instr = cpu_to_le32(instr);
270 			nb = (instr >> 11) & 0x1f;
271 			if (nb == 0)
272 				nb = 32;
273 		}
274 		if (nb + reg * 4 > 128) {
275 			nb0 = nb + reg * 4 - 128;
276 			nb = 128 - reg * 4;
277 		}
278 #ifdef __LITTLE_ENDIAN__
279 		/*
280 		 *  String instructions are endian neutral but the code
281 		 *  below is not.  Force byte swapping on so that the
282 		 *  effects of swizzling are undone in the load/store
283 		 *  loops below.
284 		 */
285 		flags ^= SW;
286 #endif
287 	} else {
288 		/* lwm, stmw */
289 		nb = (32 - reg) * 4;
290 	}
291 
292 	if (!access_ok((flags & ST ? VERIFY_WRITE: VERIFY_READ), addr, nb+nb0))
293 		return -EFAULT;	/* bad address */
294 
295 	rptr = &regs->gpr[reg];
296 	p = (unsigned long) addr;
297 	bswiz = (flags & SW)? 3: 0;
298 
299 	if (!(flags & ST)) {
300 		/*
301 		 * This zeroes the top 4 bytes of the affected registers
302 		 * in 64-bit mode, and also zeroes out any remaining
303 		 * bytes of the last register for lsw*.
304 		 */
305 		memset(rptr, 0, ((nb + 3) / 4) * sizeof(unsigned long));
306 		if (nb0 > 0)
307 			memset(&regs->gpr[0], 0,
308 			       ((nb0 + 3) / 4) * sizeof(unsigned long));
309 
310 		for (i = 0; i < nb; ++i, ++p)
311 			if (__get_user_inatomic(REG_BYTE(rptr, i ^ bswiz),
312 						SWIZ_PTR(p)))
313 				return -EFAULT;
314 		if (nb0 > 0) {
315 			rptr = &regs->gpr[0];
316 			addr += nb;
317 			for (i = 0; i < nb0; ++i, ++p)
318 				if (__get_user_inatomic(REG_BYTE(rptr,
319 								 i ^ bswiz),
320 							SWIZ_PTR(p)))
321 					return -EFAULT;
322 		}
323 
324 	} else {
325 		for (i = 0; i < nb; ++i, ++p)
326 			if (__put_user_inatomic(REG_BYTE(rptr, i ^ bswiz),
327 						SWIZ_PTR(p)))
328 				return -EFAULT;
329 		if (nb0 > 0) {
330 			rptr = &regs->gpr[0];
331 			addr += nb;
332 			for (i = 0; i < nb0; ++i, ++p)
333 				if (__put_user_inatomic(REG_BYTE(rptr,
334 								 i ^ bswiz),
335 							SWIZ_PTR(p)))
336 					return -EFAULT;
337 		}
338 	}
339 	return 1;
340 }
341 
342 /*
343  * Emulate floating-point pair loads and stores.
344  * Only POWER6 has these instructions, and it does true little-endian,
345  * so we don't need the address swizzling.
346  */
347 static int emulate_fp_pair(unsigned char __user *addr, unsigned int reg,
348 			   unsigned int flags)
349 {
350 	char *ptr0 = (char *) &current->thread.TS_FPR(reg);
351 	char *ptr1 = (char *) &current->thread.TS_FPR(reg+1);
352 	int i, ret, sw = 0;
353 
354 	if (reg & 1)
355 		return 0;	/* invalid form: FRS/FRT must be even */
356 	if (flags & SW)
357 		sw = 7;
358 	ret = 0;
359 	for (i = 0; i < 8; ++i) {
360 		if (!(flags & ST)) {
361 			ret |= __get_user(ptr0[i^sw], addr + i);
362 			ret |= __get_user(ptr1[i^sw], addr + i + 8);
363 		} else {
364 			ret |= __put_user(ptr0[i^sw], addr + i);
365 			ret |= __put_user(ptr1[i^sw], addr + i + 8);
366 		}
367 	}
368 	if (ret)
369 		return -EFAULT;
370 	return 1;	/* exception handled and fixed up */
371 }
372 
373 #ifdef CONFIG_PPC64
374 static int emulate_lq_stq(struct pt_regs *regs, unsigned char __user *addr,
375 			  unsigned int reg, unsigned int flags)
376 {
377 	char *ptr0 = (char *)&regs->gpr[reg];
378 	char *ptr1 = (char *)&regs->gpr[reg+1];
379 	int i, ret, sw = 0;
380 
381 	if (reg & 1)
382 		return 0;	/* invalid form: GPR must be even */
383 	if (flags & SW)
384 		sw = 7;
385 	ret = 0;
386 	for (i = 0; i < 8; ++i) {
387 		if (!(flags & ST)) {
388 			ret |= __get_user(ptr0[i^sw], addr + i);
389 			ret |= __get_user(ptr1[i^sw], addr + i + 8);
390 		} else {
391 			ret |= __put_user(ptr0[i^sw], addr + i);
392 			ret |= __put_user(ptr1[i^sw], addr + i + 8);
393 		}
394 	}
395 	if (ret)
396 		return -EFAULT;
397 	return 1;	/* exception handled and fixed up */
398 }
399 #endif /* CONFIG_PPC64 */
400 
401 #ifdef CONFIG_SPE
402 
403 static struct aligninfo spe_aligninfo[32] = {
404 	{ 8, LD+E8 },		/* 0 00 00: evldd[x] */
405 	{ 8, LD+E4 },		/* 0 00 01: evldw[x] */
406 	{ 8, LD },		/* 0 00 10: evldh[x] */
407 	INVALID,		/* 0 00 11 */
408 	{ 2, LD },		/* 0 01 00: evlhhesplat[x] */
409 	INVALID,		/* 0 01 01 */
410 	{ 2, LD },		/* 0 01 10: evlhhousplat[x] */
411 	{ 2, LD+SE },		/* 0 01 11: evlhhossplat[x] */
412 	{ 4, LD },		/* 0 10 00: evlwhe[x] */
413 	INVALID,		/* 0 10 01 */
414 	{ 4, LD },		/* 0 10 10: evlwhou[x] */
415 	{ 4, LD+SE },		/* 0 10 11: evlwhos[x] */
416 	{ 4, LD+E4 },		/* 0 11 00: evlwwsplat[x] */
417 	INVALID,		/* 0 11 01 */
418 	{ 4, LD },		/* 0 11 10: evlwhsplat[x] */
419 	INVALID,		/* 0 11 11 */
420 
421 	{ 8, ST+E8 },		/* 1 00 00: evstdd[x] */
422 	{ 8, ST+E4 },		/* 1 00 01: evstdw[x] */
423 	{ 8, ST },		/* 1 00 10: evstdh[x] */
424 	INVALID,		/* 1 00 11 */
425 	INVALID,		/* 1 01 00 */
426 	INVALID,		/* 1 01 01 */
427 	INVALID,		/* 1 01 10 */
428 	INVALID,		/* 1 01 11 */
429 	{ 4, ST },		/* 1 10 00: evstwhe[x] */
430 	INVALID,		/* 1 10 01 */
431 	{ 4, ST },		/* 1 10 10: evstwho[x] */
432 	INVALID,		/* 1 10 11 */
433 	{ 4, ST+E4 },		/* 1 11 00: evstwwe[x] */
434 	INVALID,		/* 1 11 01 */
435 	{ 4, ST+E4 },		/* 1 11 10: evstwwo[x] */
436 	INVALID,		/* 1 11 11 */
437 };
438 
439 #define	EVLDD		0x00
440 #define	EVLDW		0x01
441 #define	EVLDH		0x02
442 #define	EVLHHESPLAT	0x04
443 #define	EVLHHOUSPLAT	0x06
444 #define	EVLHHOSSPLAT	0x07
445 #define	EVLWHE		0x08
446 #define	EVLWHOU		0x0A
447 #define	EVLWHOS		0x0B
448 #define	EVLWWSPLAT	0x0C
449 #define	EVLWHSPLAT	0x0E
450 #define	EVSTDD		0x10
451 #define	EVSTDW		0x11
452 #define	EVSTDH		0x12
453 #define	EVSTWHE		0x18
454 #define	EVSTWHO		0x1A
455 #define	EVSTWWE		0x1C
456 #define	EVSTWWO		0x1E
457 
458 /*
459  * Emulate SPE loads and stores.
460  * Only Book-E has these instructions, and it does true little-endian,
461  * so we don't need the address swizzling.
462  */
463 static int emulate_spe(struct pt_regs *regs, unsigned int reg,
464 		       unsigned int instr)
465 {
466 	int ret;
467 	union {
468 		u64 ll;
469 		u32 w[2];
470 		u16 h[4];
471 		u8 v[8];
472 	} data, temp;
473 	unsigned char __user *p, *addr;
474 	unsigned long *evr = &current->thread.evr[reg];
475 	unsigned int nb, flags;
476 
477 	instr = (instr >> 1) & 0x1f;
478 
479 	/* DAR has the operand effective address */
480 	addr = (unsigned char __user *)regs->dar;
481 
482 	nb = spe_aligninfo[instr].len;
483 	flags = spe_aligninfo[instr].flags;
484 
485 	/* Verify the address of the operand */
486 	if (unlikely(user_mode(regs) &&
487 		     !access_ok((flags & ST ? VERIFY_WRITE : VERIFY_READ),
488 				addr, nb)))
489 		return -EFAULT;
490 
491 	/* userland only */
492 	if (unlikely(!user_mode(regs)))
493 		return 0;
494 
495 	flush_spe_to_thread(current);
496 
497 	/* If we are loading, get the data from user space, else
498 	 * get it from register values
499 	 */
500 	if (flags & ST) {
501 		data.ll = 0;
502 		switch (instr) {
503 		case EVSTDD:
504 		case EVSTDW:
505 		case EVSTDH:
506 			data.w[0] = *evr;
507 			data.w[1] = regs->gpr[reg];
508 			break;
509 		case EVSTWHE:
510 			data.h[2] = *evr >> 16;
511 			data.h[3] = regs->gpr[reg] >> 16;
512 			break;
513 		case EVSTWHO:
514 			data.h[2] = *evr & 0xffff;
515 			data.h[3] = regs->gpr[reg] & 0xffff;
516 			break;
517 		case EVSTWWE:
518 			data.w[1] = *evr;
519 			break;
520 		case EVSTWWO:
521 			data.w[1] = regs->gpr[reg];
522 			break;
523 		default:
524 			return -EINVAL;
525 		}
526 	} else {
527 		temp.ll = data.ll = 0;
528 		ret = 0;
529 		p = addr;
530 
531 		switch (nb) {
532 		case 8:
533 			ret |= __get_user_inatomic(temp.v[0], p++);
534 			ret |= __get_user_inatomic(temp.v[1], p++);
535 			ret |= __get_user_inatomic(temp.v[2], p++);
536 			ret |= __get_user_inatomic(temp.v[3], p++);
537 		case 4:
538 			ret |= __get_user_inatomic(temp.v[4], p++);
539 			ret |= __get_user_inatomic(temp.v[5], p++);
540 		case 2:
541 			ret |= __get_user_inatomic(temp.v[6], p++);
542 			ret |= __get_user_inatomic(temp.v[7], p++);
543 			if (unlikely(ret))
544 				return -EFAULT;
545 		}
546 
547 		switch (instr) {
548 		case EVLDD:
549 		case EVLDW:
550 		case EVLDH:
551 			data.ll = temp.ll;
552 			break;
553 		case EVLHHESPLAT:
554 			data.h[0] = temp.h[3];
555 			data.h[2] = temp.h[3];
556 			break;
557 		case EVLHHOUSPLAT:
558 		case EVLHHOSSPLAT:
559 			data.h[1] = temp.h[3];
560 			data.h[3] = temp.h[3];
561 			break;
562 		case EVLWHE:
563 			data.h[0] = temp.h[2];
564 			data.h[2] = temp.h[3];
565 			break;
566 		case EVLWHOU:
567 		case EVLWHOS:
568 			data.h[1] = temp.h[2];
569 			data.h[3] = temp.h[3];
570 			break;
571 		case EVLWWSPLAT:
572 			data.w[0] = temp.w[1];
573 			data.w[1] = temp.w[1];
574 			break;
575 		case EVLWHSPLAT:
576 			data.h[0] = temp.h[2];
577 			data.h[1] = temp.h[2];
578 			data.h[2] = temp.h[3];
579 			data.h[3] = temp.h[3];
580 			break;
581 		default:
582 			return -EINVAL;
583 		}
584 	}
585 
586 	if (flags & SW) {
587 		switch (flags & 0xf0) {
588 		case E8:
589 			data.ll = swab64(data.ll);
590 			break;
591 		case E4:
592 			data.w[0] = swab32(data.w[0]);
593 			data.w[1] = swab32(data.w[1]);
594 			break;
595 		/* Its half word endian */
596 		default:
597 			data.h[0] = swab16(data.h[0]);
598 			data.h[1] = swab16(data.h[1]);
599 			data.h[2] = swab16(data.h[2]);
600 			data.h[3] = swab16(data.h[3]);
601 			break;
602 		}
603 	}
604 
605 	if (flags & SE) {
606 		data.w[0] = (s16)data.h[1];
607 		data.w[1] = (s16)data.h[3];
608 	}
609 
610 	/* Store result to memory or update registers */
611 	if (flags & ST) {
612 		ret = 0;
613 		p = addr;
614 		switch (nb) {
615 		case 8:
616 			ret |= __put_user_inatomic(data.v[0], p++);
617 			ret |= __put_user_inatomic(data.v[1], p++);
618 			ret |= __put_user_inatomic(data.v[2], p++);
619 			ret |= __put_user_inatomic(data.v[3], p++);
620 		case 4:
621 			ret |= __put_user_inatomic(data.v[4], p++);
622 			ret |= __put_user_inatomic(data.v[5], p++);
623 		case 2:
624 			ret |= __put_user_inatomic(data.v[6], p++);
625 			ret |= __put_user_inatomic(data.v[7], p++);
626 		}
627 		if (unlikely(ret))
628 			return -EFAULT;
629 	} else {
630 		*evr = data.w[0];
631 		regs->gpr[reg] = data.w[1];
632 	}
633 
634 	return 1;
635 }
636 #endif /* CONFIG_SPE */
637 
638 #ifdef CONFIG_VSX
639 /*
640  * Emulate VSX instructions...
641  */
642 static int emulate_vsx(unsigned char __user *addr, unsigned int reg,
643 		       unsigned int areg, struct pt_regs *regs,
644 		       unsigned int flags, unsigned int length,
645 		       unsigned int elsize)
646 {
647 	char *ptr;
648 	unsigned long *lptr;
649 	int ret = 0;
650 	int sw = 0;
651 	int i, j;
652 
653 	/* userland only */
654 	if (unlikely(!user_mode(regs)))
655 		return 0;
656 
657 	flush_vsx_to_thread(current);
658 
659 	if (reg < 32)
660 		ptr = (char *) &current->thread.fp_state.fpr[reg][0];
661 	else
662 		ptr = (char *) &current->thread.vr_state.vr[reg - 32];
663 
664 	lptr = (unsigned long *) ptr;
665 
666 #ifdef __LITTLE_ENDIAN__
667 	if (flags & SW) {
668 		elsize = length;
669 		sw = length-1;
670 	} else {
671 		/*
672 		 * The elements are BE ordered, even in LE mode, so process
673 		 * them in reverse order.
674 		 */
675 		addr += length - elsize;
676 
677 		/* 8 byte memory accesses go in the top 8 bytes of the VR */
678 		if (length == 8)
679 			ptr += 8;
680 	}
681 #else
682 	if (flags & SW)
683 		sw = elsize-1;
684 #endif
685 
686 	for (j = 0; j < length; j += elsize) {
687 		for (i = 0; i < elsize; ++i) {
688 			if (flags & ST)
689 				ret |= __put_user(ptr[i^sw], addr + i);
690 			else
691 				ret |= __get_user(ptr[i^sw], addr + i);
692 		}
693 		ptr  += elsize;
694 #ifdef __LITTLE_ENDIAN__
695 		addr -= elsize;
696 #else
697 		addr += elsize;
698 #endif
699 	}
700 
701 #ifdef __BIG_ENDIAN__
702 #define VSX_HI 0
703 #define VSX_LO 1
704 #else
705 #define VSX_HI 1
706 #define VSX_LO 0
707 #endif
708 
709 	if (!ret) {
710 		if (flags & U)
711 			regs->gpr[areg] = regs->dar;
712 
713 		/* Splat load copies the same data to top and bottom 8 bytes */
714 		if (flags & SPLT)
715 			lptr[VSX_LO] = lptr[VSX_HI];
716 		/* For 8 byte loads, zero the low 8 bytes */
717 		else if (!(flags & ST) && (8 == length))
718 			lptr[VSX_LO] = 0;
719 	} else
720 		return -EFAULT;
721 
722 	return 1;
723 }
724 #endif
725 
726 /*
727  * Called on alignment exception. Attempts to fixup
728  *
729  * Return 1 on success
730  * Return 0 if unable to handle the interrupt
731  * Return -EFAULT if data address is bad
732  */
733 
734 int fix_alignment(struct pt_regs *regs)
735 {
736 	unsigned int instr, nb, flags, instruction = 0;
737 	unsigned int reg, areg;
738 	unsigned int dsisr;
739 	unsigned char __user *addr;
740 	unsigned long p, swiz;
741 	int ret, i;
742 	union data {
743 		u64 ll;
744 		double dd;
745 		unsigned char v[8];
746 		struct {
747 #ifdef __LITTLE_ENDIAN__
748 			int	 low32;
749 			unsigned hi32;
750 #else
751 			unsigned hi32;
752 			int	 low32;
753 #endif
754 		} x32;
755 		struct {
756 #ifdef __LITTLE_ENDIAN__
757 			short	      low16;
758 			unsigned char hi48[6];
759 #else
760 			unsigned char hi48[6];
761 			short	      low16;
762 #endif
763 		} x16;
764 	} data;
765 
766 	/*
767 	 * We require a complete register set, if not, then our assembly
768 	 * is broken
769 	 */
770 	CHECK_FULL_REGS(regs);
771 
772 	dsisr = regs->dsisr;
773 
774 	/* Some processors don't provide us with a DSISR we can use here,
775 	 * let's make one up from the instruction
776 	 */
777 	if (cpu_has_feature(CPU_FTR_NODSISRALIGN)) {
778 		unsigned long pc = regs->nip;
779 
780 		if (cpu_has_feature(CPU_FTR_PPC_LE) && (regs->msr & MSR_LE))
781 			pc ^= 4;
782 		if (unlikely(__get_user_inatomic(instr,
783 						 (unsigned int __user *)pc)))
784 			return -EFAULT;
785 		if (cpu_has_feature(CPU_FTR_REAL_LE) && (regs->msr & MSR_LE))
786 			instr = cpu_to_le32(instr);
787 		dsisr = make_dsisr(instr);
788 		instruction = instr;
789 	}
790 
791 	/* extract the operation and registers from the dsisr */
792 	reg = (dsisr >> 5) & 0x1f;	/* source/dest register */
793 	areg = dsisr & 0x1f;		/* register to update */
794 
795 #ifdef CONFIG_SPE
796 	if ((instr >> 26) == 0x4) {
797 		PPC_WARN_ALIGNMENT(spe, regs);
798 		return emulate_spe(regs, reg, instr);
799 	}
800 #endif
801 
802 	instr = (dsisr >> 10) & 0x7f;
803 	instr |= (dsisr >> 13) & 0x60;
804 
805 	/* Lookup the operation in our table */
806 	nb = aligninfo[instr].len;
807 	flags = aligninfo[instr].flags;
808 
809 	/* ldbrx/stdbrx overlap lfs/stfs in the DSISR unfortunately */
810 	if (IS_XFORM(instruction) && ((instruction >> 1) & 0x3ff) == 532) {
811 		nb = 8;
812 		flags = LD+SW;
813 	} else if (IS_XFORM(instruction) &&
814 		   ((instruction >> 1) & 0x3ff) == 660) {
815 		nb = 8;
816 		flags = ST+SW;
817 	}
818 
819 	/* Byteswap little endian loads and stores */
820 	swiz = 0;
821 	if ((regs->msr & MSR_LE) != (MSR_KERNEL & MSR_LE)) {
822 		flags ^= SW;
823 #ifdef __BIG_ENDIAN__
824 		/*
825 		 * So-called "PowerPC little endian" mode works by
826 		 * swizzling addresses rather than by actually doing
827 		 * any byte-swapping.  To emulate this, we XOR each
828 		 * byte address with 7.  We also byte-swap, because
829 		 * the processor's address swizzling depends on the
830 		 * operand size (it xors the address with 7 for bytes,
831 		 * 6 for halfwords, 4 for words, 0 for doublewords) but
832 		 * we will xor with 7 and load/store each byte separately.
833 		 */
834 		if (cpu_has_feature(CPU_FTR_PPC_LE))
835 			swiz = 7;
836 #endif
837 	}
838 
839 	/* DAR has the operand effective address */
840 	addr = (unsigned char __user *)regs->dar;
841 
842 #ifdef CONFIG_VSX
843 	if ((instruction & 0xfc00003e) == 0x7c000018) {
844 		unsigned int elsize;
845 
846 		/* Additional register addressing bit (64 VSX vs 32 FPR/GPR) */
847 		reg |= (instruction & 0x1) << 5;
848 		/* Simple inline decoder instead of a table */
849 		/* VSX has only 8 and 16 byte memory accesses */
850 		nb = 8;
851 		if (instruction & 0x200)
852 			nb = 16;
853 
854 		/* Vector stores in little-endian mode swap individual
855 		   elements, so process them separately */
856 		elsize = 4;
857 		if (instruction & 0x80)
858 			elsize = 8;
859 
860 		flags = 0;
861 		if ((regs->msr & MSR_LE) != (MSR_KERNEL & MSR_LE))
862 			flags |= SW;
863 		if (instruction & 0x100)
864 			flags |= ST;
865 		if (instruction & 0x040)
866 			flags |= U;
867 		/* splat load needs a special decoder */
868 		if ((instruction & 0x400) == 0){
869 			flags |= SPLT;
870 			nb = 8;
871 		}
872 		PPC_WARN_ALIGNMENT(vsx, regs);
873 		return emulate_vsx(addr, reg, areg, regs, flags, nb, elsize);
874 	}
875 #endif
876 
877 	/*
878 	 * ISA 3.0 (such as P9) copy, copy_first, paste and paste_last alignment
879 	 * check.
880 	 *
881 	 * Send a SIGBUS to the process that caused the fault.
882 	 *
883 	 * We do not emulate these because paste may contain additional metadata
884 	 * when pasting to a co-processor. Furthermore, paste_last is the
885 	 * synchronisation point for preceding copy/paste sequences.
886 	 */
887 	if ((instruction & 0xfc0006fe) == PPC_INST_COPY)
888 		return -EIO;
889 
890 	/* A size of 0 indicates an instruction we don't support, with
891 	 * the exception of DCBZ which is handled as a special case here
892 	 */
893 	if (instr == DCBZ) {
894 		PPC_WARN_ALIGNMENT(dcbz, regs);
895 		return emulate_dcbz(regs, addr);
896 	}
897 	if (unlikely(nb == 0))
898 		return 0;
899 
900 	/* Load/Store Multiple instructions are handled in their own
901 	 * function
902 	 */
903 	if (flags & M) {
904 		PPC_WARN_ALIGNMENT(multiple, regs);
905 		return emulate_multiple(regs, addr, reg, nb,
906 					flags, instr, swiz);
907 	}
908 
909 	/* Verify the address of the operand */
910 	if (unlikely(user_mode(regs) &&
911 		     !access_ok((flags & ST ? VERIFY_WRITE : VERIFY_READ),
912 				addr, nb)))
913 		return -EFAULT;
914 
915 	/* Force the fprs into the save area so we can reference them */
916 	if (flags & F) {
917 		/* userland only */
918 		if (unlikely(!user_mode(regs)))
919 			return 0;
920 		flush_fp_to_thread(current);
921 	}
922 
923 	if (nb == 16) {
924 		if (flags & F) {
925 			/* Special case for 16-byte FP loads and stores */
926 			PPC_WARN_ALIGNMENT(fp_pair, regs);
927 			return emulate_fp_pair(addr, reg, flags);
928 		} else {
929 #ifdef CONFIG_PPC64
930 			/* Special case for 16-byte loads and stores */
931 			PPC_WARN_ALIGNMENT(lq_stq, regs);
932 			return emulate_lq_stq(regs, addr, reg, flags);
933 #else
934 			return 0;
935 #endif
936 		}
937 	}
938 
939 	PPC_WARN_ALIGNMENT(unaligned, regs);
940 
941 	/* If we are loading, get the data from user space, else
942 	 * get it from register values
943 	 */
944 	if (!(flags & ST)) {
945 		unsigned int start = 0;
946 
947 		switch (nb) {
948 		case 4:
949 			start = offsetof(union data, x32.low32);
950 			break;
951 		case 2:
952 			start = offsetof(union data, x16.low16);
953 			break;
954 		}
955 
956 		data.ll = 0;
957 		ret = 0;
958 		p = (unsigned long)addr;
959 
960 		for (i = 0; i < nb; i++)
961 			ret |= __get_user_inatomic(data.v[start + i],
962 						   SWIZ_PTR(p++));
963 
964 		if (unlikely(ret))
965 			return -EFAULT;
966 
967 	} else if (flags & F) {
968 		data.ll = current->thread.TS_FPR(reg);
969 		if (flags & S) {
970 			/* Single-precision FP store requires conversion... */
971 #ifdef CONFIG_PPC_FPU
972 			preempt_disable();
973 			enable_kernel_fp();
974 			cvt_df(&data.dd, (float *)&data.x32.low32);
975 			disable_kernel_fp();
976 			preempt_enable();
977 #else
978 			return 0;
979 #endif
980 		}
981 	} else
982 		data.ll = regs->gpr[reg];
983 
984 	if (flags & SW) {
985 		switch (nb) {
986 		case 8:
987 			data.ll = swab64(data.ll);
988 			break;
989 		case 4:
990 			data.x32.low32 = swab32(data.x32.low32);
991 			break;
992 		case 2:
993 			data.x16.low16 = swab16(data.x16.low16);
994 			break;
995 		}
996 	}
997 
998 	/* Perform other misc operations like sign extension
999 	 * or floating point single precision conversion
1000 	 */
1001 	switch (flags & ~(U|SW)) {
1002 	case LD+SE:	/* sign extending integer loads */
1003 	case LD+F+SE:	/* sign extend for lfiwax */
1004 		if ( nb == 2 )
1005 			data.ll = data.x16.low16;
1006 		else	/* nb must be 4 */
1007 			data.ll = data.x32.low32;
1008 		break;
1009 
1010 	/* Single-precision FP load requires conversion... */
1011 	case LD+F+S:
1012 #ifdef CONFIG_PPC_FPU
1013 		preempt_disable();
1014 		enable_kernel_fp();
1015 		cvt_fd((float *)&data.x32.low32, &data.dd);
1016 		disable_kernel_fp();
1017 		preempt_enable();
1018 #else
1019 		return 0;
1020 #endif
1021 		break;
1022 	}
1023 
1024 	/* Store result to memory or update registers */
1025 	if (flags & ST) {
1026 		unsigned int start = 0;
1027 
1028 		switch (nb) {
1029 		case 4:
1030 			start = offsetof(union data, x32.low32);
1031 			break;
1032 		case 2:
1033 			start = offsetof(union data, x16.low16);
1034 			break;
1035 		}
1036 
1037 		ret = 0;
1038 		p = (unsigned long)addr;
1039 
1040 		for (i = 0; i < nb; i++)
1041 			ret |= __put_user_inatomic(data.v[start + i],
1042 						   SWIZ_PTR(p++));
1043 
1044 		if (unlikely(ret))
1045 			return -EFAULT;
1046 	} else if (flags & F)
1047 		current->thread.TS_FPR(reg) = data.ll;
1048 	else
1049 		regs->gpr[reg] = data.ll;
1050 
1051 	/* Update RA as needed */
1052 	if (flags & U)
1053 		regs->gpr[areg] = regs->dar;
1054 
1055 	return 1;
1056 }
1057