xref: /openbmc/linux/arch/powerpc/kernel/align.c (revision 726bd223)
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 <linux/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 #include <asm/cpu_has_feature.h>
30 
31 struct aligninfo {
32 	unsigned char len;
33 	unsigned char flags;
34 };
35 
36 
37 #define INVALID	{ 0, 0 }
38 
39 /* Bits in the flags field */
40 #define LD	0	/* load */
41 #define ST	1	/* store */
42 #define SE	2	/* sign-extend value, or FP ld/st as word */
43 #define F	4	/* to/from fp regs */
44 #define U	8	/* update index register */
45 #define M	0x10	/* multiple load/store */
46 #define SW	0x20	/* byte swap */
47 #define S	0x40	/* single-precision fp or... */
48 #define SX	0x40	/* ... byte count in XER */
49 #define HARD	0x80	/* string, stwcx. */
50 #define E4	0x40	/* SPE endianness is word */
51 #define E8	0x80	/* SPE endianness is double word */
52 #define SPLT	0x80	/* VSX SPLAT load */
53 
54 /* DSISR bits reported for a DCBZ instruction: */
55 #define DCBZ	0x5f	/* 8xx/82xx dcbz faults when cache not enabled */
56 
57 /*
58  * The PowerPC stores certain bits of the instruction that caused the
59  * alignment exception in the DSISR register.  This array maps those
60  * bits to information about the operand length and what the
61  * instruction would do.
62  */
63 static struct aligninfo aligninfo[128] = {
64 	{ 4, LD },		/* 00 0 0000: lwz / lwarx */
65 	INVALID,		/* 00 0 0001 */
66 	{ 4, ST },		/* 00 0 0010: stw */
67 	INVALID,		/* 00 0 0011 */
68 	{ 2, LD },		/* 00 0 0100: lhz */
69 	{ 2, LD+SE },		/* 00 0 0101: lha */
70 	{ 2, ST },		/* 00 0 0110: sth */
71 	{ 4, LD+M },		/* 00 0 0111: lmw */
72 	{ 4, LD+F+S },		/* 00 0 1000: lfs */
73 	{ 8, LD+F },		/* 00 0 1001: lfd */
74 	{ 4, ST+F+S },		/* 00 0 1010: stfs */
75 	{ 8, ST+F },		/* 00 0 1011: stfd */
76 	{ 16, LD },		/* 00 0 1100: lq */
77 	{ 8, LD },		/* 00 0 1101: ld/ldu/lwa */
78 	INVALID,		/* 00 0 1110 */
79 	{ 8, ST },		/* 00 0 1111: std/stdu */
80 	{ 4, LD+U },		/* 00 1 0000: lwzu */
81 	INVALID,		/* 00 1 0001 */
82 	{ 4, ST+U },		/* 00 1 0010: stwu */
83 	INVALID,		/* 00 1 0011 */
84 	{ 2, LD+U },		/* 00 1 0100: lhzu */
85 	{ 2, LD+SE+U },		/* 00 1 0101: lhau */
86 	{ 2, ST+U },		/* 00 1 0110: sthu */
87 	{ 4, ST+M },		/* 00 1 0111: stmw */
88 	{ 4, LD+F+S+U },	/* 00 1 1000: lfsu */
89 	{ 8, LD+F+U },		/* 00 1 1001: lfdu */
90 	{ 4, ST+F+S+U },	/* 00 1 1010: stfsu */
91 	{ 8, ST+F+U },		/* 00 1 1011: stfdu */
92 	{ 16, LD+F },		/* 00 1 1100: lfdp */
93 	INVALID,		/* 00 1 1101 */
94 	{ 16, ST+F },		/* 00 1 1110: stfdp */
95 	INVALID,		/* 00 1 1111 */
96 	{ 8, LD },		/* 01 0 0000: ldx */
97 	INVALID,		/* 01 0 0001 */
98 	{ 8, ST },		/* 01 0 0010: stdx */
99 	INVALID,		/* 01 0 0011 */
100 	INVALID,		/* 01 0 0100 */
101 	{ 4, LD+SE },		/* 01 0 0101: lwax */
102 	INVALID,		/* 01 0 0110 */
103 	INVALID,		/* 01 0 0111 */
104 	{ 4, LD+M+HARD+SX },	/* 01 0 1000: lswx */
105 	{ 4, LD+M+HARD },	/* 01 0 1001: lswi */
106 	{ 4, ST+M+HARD+SX },	/* 01 0 1010: stswx */
107 	{ 4, ST+M+HARD },	/* 01 0 1011: stswi */
108 	INVALID,		/* 01 0 1100 */
109 	{ 8, LD+U },		/* 01 0 1101: ldu */
110 	INVALID,		/* 01 0 1110 */
111 	{ 8, ST+U },		/* 01 0 1111: stdu */
112 	{ 8, LD+U },		/* 01 1 0000: ldux */
113 	INVALID,		/* 01 1 0001 */
114 	{ 8, ST+U },		/* 01 1 0010: stdux */
115 	INVALID,		/* 01 1 0011 */
116 	INVALID,		/* 01 1 0100 */
117 	{ 4, LD+SE+U },		/* 01 1 0101: lwaux */
118 	INVALID,		/* 01 1 0110 */
119 	INVALID,		/* 01 1 0111 */
120 	INVALID,		/* 01 1 1000 */
121 	INVALID,		/* 01 1 1001 */
122 	INVALID,		/* 01 1 1010 */
123 	INVALID,		/* 01 1 1011 */
124 	INVALID,		/* 01 1 1100 */
125 	INVALID,		/* 01 1 1101 */
126 	INVALID,		/* 01 1 1110 */
127 	INVALID,		/* 01 1 1111 */
128 	INVALID,		/* 10 0 0000 */
129 	INVALID,		/* 10 0 0001 */
130 	INVALID,		/* 10 0 0010: stwcx. */
131 	INVALID,		/* 10 0 0011 */
132 	INVALID,		/* 10 0 0100 */
133 	INVALID,		/* 10 0 0101 */
134 	INVALID,		/* 10 0 0110 */
135 	INVALID,		/* 10 0 0111 */
136 	{ 4, LD+SW },		/* 10 0 1000: lwbrx */
137 	INVALID,		/* 10 0 1001 */
138 	{ 4, ST+SW },		/* 10 0 1010: stwbrx */
139 	INVALID,		/* 10 0 1011 */
140 	{ 2, LD+SW },		/* 10 0 1100: lhbrx */
141 	{ 4, LD+SE },		/* 10 0 1101  lwa */
142 	{ 2, ST+SW },		/* 10 0 1110: sthbrx */
143 	{ 16, ST },		/* 10 0 1111: stq */
144 	INVALID,		/* 10 1 0000 */
145 	INVALID,		/* 10 1 0001 */
146 	INVALID,		/* 10 1 0010 */
147 	INVALID,		/* 10 1 0011 */
148 	INVALID,		/* 10 1 0100 */
149 	INVALID,		/* 10 1 0101 */
150 	INVALID,		/* 10 1 0110 */
151 	INVALID,		/* 10 1 0111 */
152 	INVALID,		/* 10 1 1000 */
153 	INVALID,		/* 10 1 1001 */
154 	INVALID,		/* 10 1 1010 */
155 	INVALID,		/* 10 1 1011 */
156 	INVALID,		/* 10 1 1100 */
157 	INVALID,		/* 10 1 1101 */
158 	INVALID,		/* 10 1 1110 */
159 	{ 0, ST+HARD },		/* 10 1 1111: dcbz */
160 	{ 4, LD },		/* 11 0 0000: lwzx */
161 	INVALID,		/* 11 0 0001 */
162 	{ 4, ST },		/* 11 0 0010: stwx */
163 	INVALID,		/* 11 0 0011 */
164 	{ 2, LD },		/* 11 0 0100: lhzx */
165 	{ 2, LD+SE },		/* 11 0 0101: lhax */
166 	{ 2, ST },		/* 11 0 0110: sthx */
167 	INVALID,		/* 11 0 0111 */
168 	{ 4, LD+F+S },		/* 11 0 1000: lfsx */
169 	{ 8, LD+F },		/* 11 0 1001: lfdx */
170 	{ 4, ST+F+S },		/* 11 0 1010: stfsx */
171 	{ 8, ST+F },		/* 11 0 1011: stfdx */
172 	{ 16, LD+F },		/* 11 0 1100: lfdpx */
173 	{ 4, LD+F+SE },		/* 11 0 1101: lfiwax */
174 	{ 16, ST+F },		/* 11 0 1110: stfdpx */
175 	{ 4, ST+F },		/* 11 0 1111: stfiwx */
176 	{ 4, LD+U },		/* 11 1 0000: lwzux */
177 	INVALID,		/* 11 1 0001 */
178 	{ 4, ST+U },		/* 11 1 0010: stwux */
179 	INVALID,		/* 11 1 0011 */
180 	{ 2, LD+U },		/* 11 1 0100: lhzux */
181 	{ 2, LD+SE+U },		/* 11 1 0101: lhaux */
182 	{ 2, ST+U },		/* 11 1 0110: sthux */
183 	INVALID,		/* 11 1 0111 */
184 	{ 4, LD+F+S+U },	/* 11 1 1000: lfsux */
185 	{ 8, LD+F+U },		/* 11 1 1001: lfdux */
186 	{ 4, ST+F+S+U },	/* 11 1 1010: stfsux */
187 	{ 8, ST+F+U },		/* 11 1 1011: stfdux */
188 	INVALID,		/* 11 1 1100 */
189 	{ 4, LD+F },		/* 11 1 1101: lfiwzx */
190 	INVALID,		/* 11 1 1110 */
191 	INVALID,		/* 11 1 1111 */
192 };
193 
194 /*
195  * The dcbz (data cache block zero) instruction
196  * gives an alignment fault if used on non-cacheable
197  * memory.  We handle the fault mainly for the
198  * case when we are running with the cache disabled
199  * for debugging.
200  */
201 static int emulate_dcbz(struct pt_regs *regs, unsigned char __user *addr)
202 {
203 	long __user *p;
204 	int i, size;
205 
206 #ifdef __powerpc64__
207 	size = ppc64_caches.l1d.block_size;
208 #else
209 	size = L1_CACHE_BYTES;
210 #endif
211 	p = (long __user *) (regs->dar & -size);
212 	if (user_mode(regs) && !access_ok(VERIFY_WRITE, p, size))
213 		return -EFAULT;
214 	for (i = 0; i < size / sizeof(long); ++i)
215 		if (__put_user_inatomic(0, p+i))
216 			return -EFAULT;
217 	return 1;
218 }
219 
220 /*
221  * Emulate load & store multiple instructions
222  * On 64-bit machines, these instructions only affect/use the
223  * bottom 4 bytes of each register, and the loads clear the
224  * top 4 bytes of the affected register.
225  */
226 #ifdef __BIG_ENDIAN__
227 #ifdef CONFIG_PPC64
228 #define REG_BYTE(rp, i)		*((u8 *)((rp) + ((i) >> 2)) + ((i) & 3) + 4)
229 #else
230 #define REG_BYTE(rp, i)		*((u8 *)(rp) + (i))
231 #endif
232 #else
233 #define REG_BYTE(rp, i)		(*(((u8 *)((rp) + ((i)>>2)) + ((i)&3))))
234 #endif
235 
236 #define SWIZ_PTR(p)		((unsigned char __user *)((p) ^ swiz))
237 
238 static int emulate_multiple(struct pt_regs *regs, unsigned char __user *addr,
239 			    unsigned int reg, unsigned int nb,
240 			    unsigned int flags, unsigned int instr,
241 			    unsigned long swiz)
242 {
243 	unsigned long *rptr;
244 	unsigned int nb0, i, bswiz;
245 	unsigned long p;
246 
247 	/*
248 	 * We do not try to emulate 8 bytes multiple as they aren't really
249 	 * available in our operating environments and we don't try to
250 	 * emulate multiples operations in kernel land as they should never
251 	 * be used/generated there at least not on unaligned boundaries
252 	 */
253 	if (unlikely((nb > 4) || !user_mode(regs)))
254 		return 0;
255 
256 	/* lmw, stmw, lswi/x, stswi/x */
257 	nb0 = 0;
258 	if (flags & HARD) {
259 		if (flags & SX) {
260 			nb = regs->xer & 127;
261 			if (nb == 0)
262 				return 1;
263 		} else {
264 			unsigned long pc = regs->nip ^ (swiz & 4);
265 
266 			if (__get_user_inatomic(instr,
267 						(unsigned int __user *)pc))
268 				return -EFAULT;
269 			if (swiz == 0 && (flags & SW))
270 				instr = cpu_to_le32(instr);
271 			nb = (instr >> 11) & 0x1f;
272 			if (nb == 0)
273 				nb = 32;
274 		}
275 		if (nb + reg * 4 > 128) {
276 			nb0 = nb + reg * 4 - 128;
277 			nb = 128 - reg * 4;
278 		}
279 #ifdef __LITTLE_ENDIAN__
280 		/*
281 		 *  String instructions are endian neutral but the code
282 		 *  below is not.  Force byte swapping on so that the
283 		 *  effects of swizzling are undone in the load/store
284 		 *  loops below.
285 		 */
286 		flags ^= SW;
287 #endif
288 	} else {
289 		/* lwm, stmw */
290 		nb = (32 - reg) * 4;
291 	}
292 
293 	if (!access_ok((flags & ST ? VERIFY_WRITE: VERIFY_READ), addr, nb+nb0))
294 		return -EFAULT;	/* bad address */
295 
296 	rptr = &regs->gpr[reg];
297 	p = (unsigned long) addr;
298 	bswiz = (flags & SW)? 3: 0;
299 
300 	if (!(flags & ST)) {
301 		/*
302 		 * This zeroes the top 4 bytes of the affected registers
303 		 * in 64-bit mode, and also zeroes out any remaining
304 		 * bytes of the last register for lsw*.
305 		 */
306 		memset(rptr, 0, ((nb + 3) / 4) * sizeof(unsigned long));
307 		if (nb0 > 0)
308 			memset(&regs->gpr[0], 0,
309 			       ((nb0 + 3) / 4) * sizeof(unsigned long));
310 
311 		for (i = 0; i < nb; ++i, ++p)
312 			if (__get_user_inatomic(REG_BYTE(rptr, i ^ bswiz),
313 						SWIZ_PTR(p)))
314 				return -EFAULT;
315 		if (nb0 > 0) {
316 			rptr = &regs->gpr[0];
317 			addr += nb;
318 			for (i = 0; i < nb0; ++i, ++p)
319 				if (__get_user_inatomic(REG_BYTE(rptr,
320 								 i ^ bswiz),
321 							SWIZ_PTR(p)))
322 					return -EFAULT;
323 		}
324 
325 	} else {
326 		for (i = 0; i < nb; ++i, ++p)
327 			if (__put_user_inatomic(REG_BYTE(rptr, i ^ bswiz),
328 						SWIZ_PTR(p)))
329 				return -EFAULT;
330 		if (nb0 > 0) {
331 			rptr = &regs->gpr[0];
332 			addr += nb;
333 			for (i = 0; i < nb0; ++i, ++p)
334 				if (__put_user_inatomic(REG_BYTE(rptr,
335 								 i ^ bswiz),
336 							SWIZ_PTR(p)))
337 					return -EFAULT;
338 		}
339 	}
340 	return 1;
341 }
342 
343 /*
344  * Emulate floating-point pair loads and stores.
345  * Only POWER6 has these instructions, and it does true little-endian,
346  * so we don't need the address swizzling.
347  */
348 static int emulate_fp_pair(unsigned char __user *addr, unsigned int reg,
349 			   unsigned int flags)
350 {
351 	char *ptr0 = (char *) &current->thread.TS_FPR(reg);
352 	char *ptr1 = (char *) &current->thread.TS_FPR(reg+1);
353 	int i, ret, sw = 0;
354 
355 	if (reg & 1)
356 		return 0;	/* invalid form: FRS/FRT must be even */
357 	if (flags & SW)
358 		sw = 7;
359 	ret = 0;
360 	for (i = 0; i < 8; ++i) {
361 		if (!(flags & ST)) {
362 			ret |= __get_user(ptr0[i^sw], addr + i);
363 			ret |= __get_user(ptr1[i^sw], addr + i + 8);
364 		} else {
365 			ret |= __put_user(ptr0[i^sw], addr + i);
366 			ret |= __put_user(ptr1[i^sw], addr + i + 8);
367 		}
368 	}
369 	if (ret)
370 		return -EFAULT;
371 	return 1;	/* exception handled and fixed up */
372 }
373 
374 #ifdef CONFIG_PPC64
375 static int emulate_lq_stq(struct pt_regs *regs, unsigned char __user *addr,
376 			  unsigned int reg, unsigned int flags)
377 {
378 	char *ptr0 = (char *)&regs->gpr[reg];
379 	char *ptr1 = (char *)&regs->gpr[reg+1];
380 	int i, ret, sw = 0;
381 
382 	if (reg & 1)
383 		return 0;	/* invalid form: GPR must be even */
384 	if (flags & SW)
385 		sw = 7;
386 	ret = 0;
387 	for (i = 0; i < 8; ++i) {
388 		if (!(flags & ST)) {
389 			ret |= __get_user(ptr0[i^sw], addr + i);
390 			ret |= __get_user(ptr1[i^sw], addr + i + 8);
391 		} else {
392 			ret |= __put_user(ptr0[i^sw], addr + i);
393 			ret |= __put_user(ptr1[i^sw], addr + i + 8);
394 		}
395 	}
396 	if (ret)
397 		return -EFAULT;
398 	return 1;	/* exception handled and fixed up */
399 }
400 #endif /* CONFIG_PPC64 */
401 
402 #ifdef CONFIG_SPE
403 
404 static struct aligninfo spe_aligninfo[32] = {
405 	{ 8, LD+E8 },		/* 0 00 00: evldd[x] */
406 	{ 8, LD+E4 },		/* 0 00 01: evldw[x] */
407 	{ 8, LD },		/* 0 00 10: evldh[x] */
408 	INVALID,		/* 0 00 11 */
409 	{ 2, LD },		/* 0 01 00: evlhhesplat[x] */
410 	INVALID,		/* 0 01 01 */
411 	{ 2, LD },		/* 0 01 10: evlhhousplat[x] */
412 	{ 2, LD+SE },		/* 0 01 11: evlhhossplat[x] */
413 	{ 4, LD },		/* 0 10 00: evlwhe[x] */
414 	INVALID,		/* 0 10 01 */
415 	{ 4, LD },		/* 0 10 10: evlwhou[x] */
416 	{ 4, LD+SE },		/* 0 10 11: evlwhos[x] */
417 	{ 4, LD+E4 },		/* 0 11 00: evlwwsplat[x] */
418 	INVALID,		/* 0 11 01 */
419 	{ 4, LD },		/* 0 11 10: evlwhsplat[x] */
420 	INVALID,		/* 0 11 11 */
421 
422 	{ 8, ST+E8 },		/* 1 00 00: evstdd[x] */
423 	{ 8, ST+E4 },		/* 1 00 01: evstdw[x] */
424 	{ 8, ST },		/* 1 00 10: evstdh[x] */
425 	INVALID,		/* 1 00 11 */
426 	INVALID,		/* 1 01 00 */
427 	INVALID,		/* 1 01 01 */
428 	INVALID,		/* 1 01 10 */
429 	INVALID,		/* 1 01 11 */
430 	{ 4, ST },		/* 1 10 00: evstwhe[x] */
431 	INVALID,		/* 1 10 01 */
432 	{ 4, ST },		/* 1 10 10: evstwho[x] */
433 	INVALID,		/* 1 10 11 */
434 	{ 4, ST+E4 },		/* 1 11 00: evstwwe[x] */
435 	INVALID,		/* 1 11 01 */
436 	{ 4, ST+E4 },		/* 1 11 10: evstwwo[x] */
437 	INVALID,		/* 1 11 11 */
438 };
439 
440 #define	EVLDD		0x00
441 #define	EVLDW		0x01
442 #define	EVLDH		0x02
443 #define	EVLHHESPLAT	0x04
444 #define	EVLHHOUSPLAT	0x06
445 #define	EVLHHOSSPLAT	0x07
446 #define	EVLWHE		0x08
447 #define	EVLWHOU		0x0A
448 #define	EVLWHOS		0x0B
449 #define	EVLWWSPLAT	0x0C
450 #define	EVLWHSPLAT	0x0E
451 #define	EVSTDD		0x10
452 #define	EVSTDW		0x11
453 #define	EVSTDH		0x12
454 #define	EVSTWHE		0x18
455 #define	EVSTWHO		0x1A
456 #define	EVSTWWE		0x1C
457 #define	EVSTWWO		0x1E
458 
459 /*
460  * Emulate SPE loads and stores.
461  * Only Book-E has these instructions, and it does true little-endian,
462  * so we don't need the address swizzling.
463  */
464 static int emulate_spe(struct pt_regs *regs, unsigned int reg,
465 		       unsigned int instr)
466 {
467 	int ret;
468 	union {
469 		u64 ll;
470 		u32 w[2];
471 		u16 h[4];
472 		u8 v[8];
473 	} data, temp;
474 	unsigned char __user *p, *addr;
475 	unsigned long *evr = &current->thread.evr[reg];
476 	unsigned int nb, flags;
477 
478 	instr = (instr >> 1) & 0x1f;
479 
480 	/* DAR has the operand effective address */
481 	addr = (unsigned char __user *)regs->dar;
482 
483 	nb = spe_aligninfo[instr].len;
484 	flags = spe_aligninfo[instr].flags;
485 
486 	/* Verify the address of the operand */
487 	if (unlikely(user_mode(regs) &&
488 		     !access_ok((flags & ST ? VERIFY_WRITE : VERIFY_READ),
489 				addr, nb)))
490 		return -EFAULT;
491 
492 	/* userland only */
493 	if (unlikely(!user_mode(regs)))
494 		return 0;
495 
496 	flush_spe_to_thread(current);
497 
498 	/* If we are loading, get the data from user space, else
499 	 * get it from register values
500 	 */
501 	if (flags & ST) {
502 		data.ll = 0;
503 		switch (instr) {
504 		case EVSTDD:
505 		case EVSTDW:
506 		case EVSTDH:
507 			data.w[0] = *evr;
508 			data.w[1] = regs->gpr[reg];
509 			break;
510 		case EVSTWHE:
511 			data.h[2] = *evr >> 16;
512 			data.h[3] = regs->gpr[reg] >> 16;
513 			break;
514 		case EVSTWHO:
515 			data.h[2] = *evr & 0xffff;
516 			data.h[3] = regs->gpr[reg] & 0xffff;
517 			break;
518 		case EVSTWWE:
519 			data.w[1] = *evr;
520 			break;
521 		case EVSTWWO:
522 			data.w[1] = regs->gpr[reg];
523 			break;
524 		default:
525 			return -EINVAL;
526 		}
527 	} else {
528 		temp.ll = data.ll = 0;
529 		ret = 0;
530 		p = addr;
531 
532 		switch (nb) {
533 		case 8:
534 			ret |= __get_user_inatomic(temp.v[0], p++);
535 			ret |= __get_user_inatomic(temp.v[1], p++);
536 			ret |= __get_user_inatomic(temp.v[2], p++);
537 			ret |= __get_user_inatomic(temp.v[3], p++);
538 		case 4:
539 			ret |= __get_user_inatomic(temp.v[4], p++);
540 			ret |= __get_user_inatomic(temp.v[5], p++);
541 		case 2:
542 			ret |= __get_user_inatomic(temp.v[6], p++);
543 			ret |= __get_user_inatomic(temp.v[7], p++);
544 			if (unlikely(ret))
545 				return -EFAULT;
546 		}
547 
548 		switch (instr) {
549 		case EVLDD:
550 		case EVLDW:
551 		case EVLDH:
552 			data.ll = temp.ll;
553 			break;
554 		case EVLHHESPLAT:
555 			data.h[0] = temp.h[3];
556 			data.h[2] = temp.h[3];
557 			break;
558 		case EVLHHOUSPLAT:
559 		case EVLHHOSSPLAT:
560 			data.h[1] = temp.h[3];
561 			data.h[3] = temp.h[3];
562 			break;
563 		case EVLWHE:
564 			data.h[0] = temp.h[2];
565 			data.h[2] = temp.h[3];
566 			break;
567 		case EVLWHOU:
568 		case EVLWHOS:
569 			data.h[1] = temp.h[2];
570 			data.h[3] = temp.h[3];
571 			break;
572 		case EVLWWSPLAT:
573 			data.w[0] = temp.w[1];
574 			data.w[1] = temp.w[1];
575 			break;
576 		case EVLWHSPLAT:
577 			data.h[0] = temp.h[2];
578 			data.h[1] = temp.h[2];
579 			data.h[2] = temp.h[3];
580 			data.h[3] = temp.h[3];
581 			break;
582 		default:
583 			return -EINVAL;
584 		}
585 	}
586 
587 	if (flags & SW) {
588 		switch (flags & 0xf0) {
589 		case E8:
590 			data.ll = swab64(data.ll);
591 			break;
592 		case E4:
593 			data.w[0] = swab32(data.w[0]);
594 			data.w[1] = swab32(data.w[1]);
595 			break;
596 		/* Its half word endian */
597 		default:
598 			data.h[0] = swab16(data.h[0]);
599 			data.h[1] = swab16(data.h[1]);
600 			data.h[2] = swab16(data.h[2]);
601 			data.h[3] = swab16(data.h[3]);
602 			break;
603 		}
604 	}
605 
606 	if (flags & SE) {
607 		data.w[0] = (s16)data.h[1];
608 		data.w[1] = (s16)data.h[3];
609 	}
610 
611 	/* Store result to memory or update registers */
612 	if (flags & ST) {
613 		ret = 0;
614 		p = addr;
615 		switch (nb) {
616 		case 8:
617 			ret |= __put_user_inatomic(data.v[0], p++);
618 			ret |= __put_user_inatomic(data.v[1], p++);
619 			ret |= __put_user_inatomic(data.v[2], p++);
620 			ret |= __put_user_inatomic(data.v[3], p++);
621 		case 4:
622 			ret |= __put_user_inatomic(data.v[4], p++);
623 			ret |= __put_user_inatomic(data.v[5], p++);
624 		case 2:
625 			ret |= __put_user_inatomic(data.v[6], p++);
626 			ret |= __put_user_inatomic(data.v[7], p++);
627 		}
628 		if (unlikely(ret))
629 			return -EFAULT;
630 	} else {
631 		*evr = data.w[0];
632 		regs->gpr[reg] = data.w[1];
633 	}
634 
635 	return 1;
636 }
637 #endif /* CONFIG_SPE */
638 
639 #ifdef CONFIG_VSX
640 /*
641  * Emulate VSX instructions...
642  */
643 static int emulate_vsx(unsigned char __user *addr, unsigned int reg,
644 		       unsigned int areg, struct pt_regs *regs,
645 		       unsigned int flags, unsigned int length,
646 		       unsigned int elsize)
647 {
648 	char *ptr;
649 	unsigned long *lptr;
650 	int ret = 0;
651 	int sw = 0;
652 	int i, j;
653 
654 	/* userland only */
655 	if (unlikely(!user_mode(regs)))
656 		return 0;
657 
658 	flush_vsx_to_thread(current);
659 
660 	if (reg < 32)
661 		ptr = (char *) &current->thread.fp_state.fpr[reg][0];
662 	else
663 		ptr = (char *) &current->thread.vr_state.vr[reg - 32];
664 
665 	lptr = (unsigned long *) ptr;
666 
667 #ifdef __LITTLE_ENDIAN__
668 	if (flags & SW) {
669 		elsize = length;
670 		sw = length-1;
671 	} else {
672 		/*
673 		 * The elements are BE ordered, even in LE mode, so process
674 		 * them in reverse order.
675 		 */
676 		addr += length - elsize;
677 
678 		/* 8 byte memory accesses go in the top 8 bytes of the VR */
679 		if (length == 8)
680 			ptr += 8;
681 	}
682 #else
683 	if (flags & SW)
684 		sw = elsize-1;
685 #endif
686 
687 	for (j = 0; j < length; j += elsize) {
688 		for (i = 0; i < elsize; ++i) {
689 			if (flags & ST)
690 				ret |= __put_user(ptr[i^sw], addr + i);
691 			else
692 				ret |= __get_user(ptr[i^sw], addr + i);
693 		}
694 		ptr  += elsize;
695 #ifdef __LITTLE_ENDIAN__
696 		addr -= elsize;
697 #else
698 		addr += elsize;
699 #endif
700 	}
701 
702 #ifdef __BIG_ENDIAN__
703 #define VSX_HI 0
704 #define VSX_LO 1
705 #else
706 #define VSX_HI 1
707 #define VSX_LO 0
708 #endif
709 
710 	if (!ret) {
711 		if (flags & U)
712 			regs->gpr[areg] = regs->dar;
713 
714 		/* Splat load copies the same data to top and bottom 8 bytes */
715 		if (flags & SPLT)
716 			lptr[VSX_LO] = lptr[VSX_HI];
717 		/* For 8 byte loads, zero the low 8 bytes */
718 		else if (!(flags & ST) && (8 == length))
719 			lptr[VSX_LO] = 0;
720 	} else
721 		return -EFAULT;
722 
723 	return 1;
724 }
725 #endif
726 
727 /*
728  * Called on alignment exception. Attempts to fixup
729  *
730  * Return 1 on success
731  * Return 0 if unable to handle the interrupt
732  * Return -EFAULT if data address is bad
733  */
734 
735 int fix_alignment(struct pt_regs *regs)
736 {
737 	unsigned int instr, nb, flags, instruction = 0;
738 	unsigned int reg, areg;
739 	unsigned int dsisr;
740 	unsigned char __user *addr;
741 	unsigned long p, swiz;
742 	int ret, i;
743 	union data {
744 		u64 ll;
745 		double dd;
746 		unsigned char v[8];
747 		struct {
748 #ifdef __LITTLE_ENDIAN__
749 			int	 low32;
750 			unsigned hi32;
751 #else
752 			unsigned hi32;
753 			int	 low32;
754 #endif
755 		} x32;
756 		struct {
757 #ifdef __LITTLE_ENDIAN__
758 			short	      low16;
759 			unsigned char hi48[6];
760 #else
761 			unsigned char hi48[6];
762 			short	      low16;
763 #endif
764 		} x16;
765 	} data;
766 
767 	/*
768 	 * We require a complete register set, if not, then our assembly
769 	 * is broken
770 	 */
771 	CHECK_FULL_REGS(regs);
772 
773 	dsisr = regs->dsisr;
774 
775 	/* Some processors don't provide us with a DSISR we can use here,
776 	 * let's make one up from the instruction
777 	 */
778 	if (cpu_has_feature(CPU_FTR_NODSISRALIGN)) {
779 		unsigned long pc = regs->nip;
780 
781 		if (cpu_has_feature(CPU_FTR_PPC_LE) && (regs->msr & MSR_LE))
782 			pc ^= 4;
783 		if (unlikely(__get_user_inatomic(instr,
784 						 (unsigned int __user *)pc)))
785 			return -EFAULT;
786 		if (cpu_has_feature(CPU_FTR_REAL_LE) && (regs->msr & MSR_LE))
787 			instr = cpu_to_le32(instr);
788 		dsisr = make_dsisr(instr);
789 		instruction = instr;
790 	}
791 
792 	/* extract the operation and registers from the dsisr */
793 	reg = (dsisr >> 5) & 0x1f;	/* source/dest register */
794 	areg = dsisr & 0x1f;		/* register to update */
795 
796 #ifdef CONFIG_SPE
797 	if ((instr >> 26) == 0x4) {
798 		PPC_WARN_ALIGNMENT(spe, regs);
799 		return emulate_spe(regs, reg, instr);
800 	}
801 #endif
802 
803 	instr = (dsisr >> 10) & 0x7f;
804 	instr |= (dsisr >> 13) & 0x60;
805 
806 	/* Lookup the operation in our table */
807 	nb = aligninfo[instr].len;
808 	flags = aligninfo[instr].flags;
809 
810 	/* ldbrx/stdbrx overlap lfs/stfs in the DSISR unfortunately */
811 	if (IS_XFORM(instruction) && ((instruction >> 1) & 0x3ff) == 532) {
812 		nb = 8;
813 		flags = LD+SW;
814 	} else if (IS_XFORM(instruction) &&
815 		   ((instruction >> 1) & 0x3ff) == 660) {
816 		nb = 8;
817 		flags = ST+SW;
818 	}
819 
820 	/* Byteswap little endian loads and stores */
821 	swiz = 0;
822 	if ((regs->msr & MSR_LE) != (MSR_KERNEL & MSR_LE)) {
823 		flags ^= SW;
824 #ifdef __BIG_ENDIAN__
825 		/*
826 		 * So-called "PowerPC little endian" mode works by
827 		 * swizzling addresses rather than by actually doing
828 		 * any byte-swapping.  To emulate this, we XOR each
829 		 * byte address with 7.  We also byte-swap, because
830 		 * the processor's address swizzling depends on the
831 		 * operand size (it xors the address with 7 for bytes,
832 		 * 6 for halfwords, 4 for words, 0 for doublewords) but
833 		 * we will xor with 7 and load/store each byte separately.
834 		 */
835 		if (cpu_has_feature(CPU_FTR_PPC_LE))
836 			swiz = 7;
837 #endif
838 	}
839 
840 	/* DAR has the operand effective address */
841 	addr = (unsigned char __user *)regs->dar;
842 
843 #ifdef CONFIG_VSX
844 	if ((instruction & 0xfc00003e) == 0x7c000018) {
845 		unsigned int elsize;
846 
847 		/* Additional register addressing bit (64 VSX vs 32 FPR/GPR) */
848 		reg |= (instruction & 0x1) << 5;
849 		/* Simple inline decoder instead of a table */
850 		/* VSX has only 8 and 16 byte memory accesses */
851 		nb = 8;
852 		if (instruction & 0x200)
853 			nb = 16;
854 
855 		/* Vector stores in little-endian mode swap individual
856 		   elements, so process them separately */
857 		elsize = 4;
858 		if (instruction & 0x80)
859 			elsize = 8;
860 
861 		flags = 0;
862 		if ((regs->msr & MSR_LE) != (MSR_KERNEL & MSR_LE))
863 			flags |= SW;
864 		if (instruction & 0x100)
865 			flags |= ST;
866 		if (instruction & 0x040)
867 			flags |= U;
868 		/* splat load needs a special decoder */
869 		if ((instruction & 0x400) == 0){
870 			flags |= SPLT;
871 			nb = 8;
872 		}
873 		PPC_WARN_ALIGNMENT(vsx, regs);
874 		return emulate_vsx(addr, reg, areg, regs, flags, nb, elsize);
875 	}
876 #endif
877 
878 	/*
879 	 * ISA 3.0 (such as P9) copy, copy_first, paste and paste_last alignment
880 	 * check.
881 	 *
882 	 * Send a SIGBUS to the process that caused the fault.
883 	 *
884 	 * We do not emulate these because paste may contain additional metadata
885 	 * when pasting to a co-processor. Furthermore, paste_last is the
886 	 * synchronisation point for preceding copy/paste sequences.
887 	 */
888 	if ((instruction & 0xfc0006fe) == PPC_INST_COPY)
889 		return -EIO;
890 
891 	/* A size of 0 indicates an instruction we don't support, with
892 	 * the exception of DCBZ which is handled as a special case here
893 	 */
894 	if (instr == DCBZ) {
895 		PPC_WARN_ALIGNMENT(dcbz, regs);
896 		return emulate_dcbz(regs, addr);
897 	}
898 	if (unlikely(nb == 0))
899 		return 0;
900 
901 	/* Load/Store Multiple instructions are handled in their own
902 	 * function
903 	 */
904 	if (flags & M) {
905 		PPC_WARN_ALIGNMENT(multiple, regs);
906 		return emulate_multiple(regs, addr, reg, nb,
907 					flags, instr, swiz);
908 	}
909 
910 	/* Verify the address of the operand */
911 	if (unlikely(user_mode(regs) &&
912 		     !access_ok((flags & ST ? VERIFY_WRITE : VERIFY_READ),
913 				addr, nb)))
914 		return -EFAULT;
915 
916 	/* Force the fprs into the save area so we can reference them */
917 	if (flags & F) {
918 		/* userland only */
919 		if (unlikely(!user_mode(regs)))
920 			return 0;
921 		flush_fp_to_thread(current);
922 	}
923 
924 	if (nb == 16) {
925 		if (flags & F) {
926 			/* Special case for 16-byte FP loads and stores */
927 			PPC_WARN_ALIGNMENT(fp_pair, regs);
928 			return emulate_fp_pair(addr, reg, flags);
929 		} else {
930 #ifdef CONFIG_PPC64
931 			/* Special case for 16-byte loads and stores */
932 			PPC_WARN_ALIGNMENT(lq_stq, regs);
933 			return emulate_lq_stq(regs, addr, reg, flags);
934 #else
935 			return 0;
936 #endif
937 		}
938 	}
939 
940 	PPC_WARN_ALIGNMENT(unaligned, regs);
941 
942 	/* If we are loading, get the data from user space, else
943 	 * get it from register values
944 	 */
945 	if (!(flags & ST)) {
946 		unsigned int start = 0;
947 
948 		switch (nb) {
949 		case 4:
950 			start = offsetof(union data, x32.low32);
951 			break;
952 		case 2:
953 			start = offsetof(union data, x16.low16);
954 			break;
955 		}
956 
957 		data.ll = 0;
958 		ret = 0;
959 		p = (unsigned long)addr;
960 
961 		for (i = 0; i < nb; i++)
962 			ret |= __get_user_inatomic(data.v[start + i],
963 						   SWIZ_PTR(p++));
964 
965 		if (unlikely(ret))
966 			return -EFAULT;
967 
968 	} else if (flags & F) {
969 		data.ll = current->thread.TS_FPR(reg);
970 		if (flags & S) {
971 			/* Single-precision FP store requires conversion... */
972 #ifdef CONFIG_PPC_FPU
973 			preempt_disable();
974 			enable_kernel_fp();
975 			cvt_df(&data.dd, (float *)&data.x32.low32);
976 			disable_kernel_fp();
977 			preempt_enable();
978 #else
979 			return 0;
980 #endif
981 		}
982 	} else
983 		data.ll = regs->gpr[reg];
984 
985 	if (flags & SW) {
986 		switch (nb) {
987 		case 8:
988 			data.ll = swab64(data.ll);
989 			break;
990 		case 4:
991 			data.x32.low32 = swab32(data.x32.low32);
992 			break;
993 		case 2:
994 			data.x16.low16 = swab16(data.x16.low16);
995 			break;
996 		}
997 	}
998 
999 	/* Perform other misc operations like sign extension
1000 	 * or floating point single precision conversion
1001 	 */
1002 	switch (flags & ~(U|SW)) {
1003 	case LD+SE:	/* sign extending integer loads */
1004 	case LD+F+SE:	/* sign extend for lfiwax */
1005 		if ( nb == 2 )
1006 			data.ll = data.x16.low16;
1007 		else	/* nb must be 4 */
1008 			data.ll = data.x32.low32;
1009 		break;
1010 
1011 	/* Single-precision FP load requires conversion... */
1012 	case LD+F+S:
1013 #ifdef CONFIG_PPC_FPU
1014 		preempt_disable();
1015 		enable_kernel_fp();
1016 		cvt_fd((float *)&data.x32.low32, &data.dd);
1017 		disable_kernel_fp();
1018 		preempt_enable();
1019 #else
1020 		return 0;
1021 #endif
1022 		break;
1023 	}
1024 
1025 	/* Store result to memory or update registers */
1026 	if (flags & ST) {
1027 		unsigned int start = 0;
1028 
1029 		switch (nb) {
1030 		case 4:
1031 			start = offsetof(union data, x32.low32);
1032 			break;
1033 		case 2:
1034 			start = offsetof(union data, x16.low16);
1035 			break;
1036 		}
1037 
1038 		ret = 0;
1039 		p = (unsigned long)addr;
1040 
1041 		for (i = 0; i < nb; i++)
1042 			ret |= __put_user_inatomic(data.v[start + i],
1043 						   SWIZ_PTR(p++));
1044 
1045 		if (unlikely(ret))
1046 			return -EFAULT;
1047 	} else if (flags & F)
1048 		current->thread.TS_FPR(reg) = data.ll;
1049 	else
1050 		regs->gpr[reg] = data.ll;
1051 
1052 	/* Update RA as needed */
1053 	if (flags & U)
1054 		regs->gpr[areg] = regs->dar;
1055 
1056 	return 1;
1057 }
1058