xref: /linux/arch/arm/mm/alignment.c (revision f2ee442115c9b6219083c019939a9cc0c9abb2f8)
1 /*
2  *  linux/arch/arm/mm/alignment.c
3  *
4  *  Copyright (C) 1995  Linus Torvalds
5  *  Modifications for ARM processor (c) 1995-2001 Russell King
6  *  Thumb alignment fault fixups (c) 2004 MontaVista Software, Inc.
7  *  - Adapted from gdb/sim/arm/thumbemu.c -- Thumb instruction emulation.
8  *    Copyright (C) 1996, Cygnus Software Technologies Ltd.
9  *
10  * This program is free software; you can redistribute it and/or modify
11  * it under the terms of the GNU General Public License version 2 as
12  * published by the Free Software Foundation.
13  */
14 #include <linux/moduleparam.h>
15 #include <linux/compiler.h>
16 #include <linux/kernel.h>
17 #include <linux/errno.h>
18 #include <linux/string.h>
19 #include <linux/proc_fs.h>
20 #include <linux/seq_file.h>
21 #include <linux/init.h>
22 #include <linux/sched.h>
23 #include <linux/uaccess.h>
24 
25 #include <asm/system.h>
26 #include <asm/unaligned.h>
27 
28 #include "fault.h"
29 
30 /*
31  * 32-bit misaligned trap handler (c) 1998 San Mehat (CCC) -July 1998
32  * /proc/sys/debug/alignment, modified and integrated into
33  * Linux 2.1 by Russell King
34  *
35  * Speed optimisations and better fault handling by Russell King.
36  *
37  * *** NOTE ***
38  * This code is not portable to processors with late data abort handling.
39  */
40 #define CODING_BITS(i)	(i & 0x0e000000)
41 
42 #define LDST_I_BIT(i)	(i & (1 << 26))		/* Immediate constant	*/
43 #define LDST_P_BIT(i)	(i & (1 << 24))		/* Preindex		*/
44 #define LDST_U_BIT(i)	(i & (1 << 23))		/* Add offset		*/
45 #define LDST_W_BIT(i)	(i & (1 << 21))		/* Writeback		*/
46 #define LDST_L_BIT(i)	(i & (1 << 20))		/* Load			*/
47 
48 #define LDST_P_EQ_U(i)	((((i) ^ ((i) >> 1)) & (1 << 23)) == 0)
49 
50 #define LDSTHD_I_BIT(i)	(i & (1 << 22))		/* double/half-word immed */
51 #define LDM_S_BIT(i)	(i & (1 << 22))		/* write CPSR from SPSR	*/
52 
53 #define RN_BITS(i)	((i >> 16) & 15)	/* Rn			*/
54 #define RD_BITS(i)	((i >> 12) & 15)	/* Rd			*/
55 #define RM_BITS(i)	(i & 15)		/* Rm			*/
56 
57 #define REGMASK_BITS(i)	(i & 0xffff)
58 #define OFFSET_BITS(i)	(i & 0x0fff)
59 
60 #define IS_SHIFT(i)	(i & 0x0ff0)
61 #define SHIFT_BITS(i)	((i >> 7) & 0x1f)
62 #define SHIFT_TYPE(i)	(i & 0x60)
63 #define SHIFT_LSL	0x00
64 #define SHIFT_LSR	0x20
65 #define SHIFT_ASR	0x40
66 #define SHIFT_RORRRX	0x60
67 
68 #define BAD_INSTR 	0xdeadc0de
69 
70 /* Thumb-2 32 bit format per ARMv7 DDI0406A A6.3, either f800h,e800h,f800h */
71 #define IS_T32(hi16) \
72 	(((hi16) & 0xe000) == 0xe000 && ((hi16) & 0x1800))
73 
74 static unsigned long ai_user;
75 static unsigned long ai_sys;
76 static unsigned long ai_skipped;
77 static unsigned long ai_half;
78 static unsigned long ai_word;
79 static unsigned long ai_dword;
80 static unsigned long ai_multi;
81 static int ai_usermode;
82 
83 core_param(alignment, ai_usermode, int, 0600);
84 
85 #define UM_WARN		(1 << 0)
86 #define UM_FIXUP	(1 << 1)
87 #define UM_SIGNAL	(1 << 2)
88 
89 /* Return true if and only if the ARMv6 unaligned access model is in use. */
90 static bool cpu_is_v6_unaligned(void)
91 {
92 	return cpu_architecture() >= CPU_ARCH_ARMv6 && (cr_alignment & CR_U);
93 }
94 
95 static int safe_usermode(int new_usermode, bool warn)
96 {
97 	/*
98 	 * ARMv6 and later CPUs can perform unaligned accesses for
99 	 * most single load and store instructions up to word size.
100 	 * LDM, STM, LDRD and STRD still need to be handled.
101 	 *
102 	 * Ignoring the alignment fault is not an option on these
103 	 * CPUs since we spin re-faulting the instruction without
104 	 * making any progress.
105 	 */
106 	if (cpu_is_v6_unaligned() && !(new_usermode & (UM_FIXUP | UM_SIGNAL))) {
107 		new_usermode |= UM_FIXUP;
108 
109 		if (warn)
110 			printk(KERN_WARNING "alignment: ignoring faults is unsafe on this CPU.  Defaulting to fixup mode.\n");
111 	}
112 
113 	return new_usermode;
114 }
115 
116 #ifdef CONFIG_PROC_FS
117 static const char *usermode_action[] = {
118 	"ignored",
119 	"warn",
120 	"fixup",
121 	"fixup+warn",
122 	"signal",
123 	"signal+warn"
124 };
125 
126 static int alignment_proc_show(struct seq_file *m, void *v)
127 {
128 	seq_printf(m, "User:\t\t%lu\n", ai_user);
129 	seq_printf(m, "System:\t\t%lu\n", ai_sys);
130 	seq_printf(m, "Skipped:\t%lu\n", ai_skipped);
131 	seq_printf(m, "Half:\t\t%lu\n", ai_half);
132 	seq_printf(m, "Word:\t\t%lu\n", ai_word);
133 	if (cpu_architecture() >= CPU_ARCH_ARMv5TE)
134 		seq_printf(m, "DWord:\t\t%lu\n", ai_dword);
135 	seq_printf(m, "Multi:\t\t%lu\n", ai_multi);
136 	seq_printf(m, "User faults:\t%i (%s)\n", ai_usermode,
137 			usermode_action[ai_usermode]);
138 
139 	return 0;
140 }
141 
142 static int alignment_proc_open(struct inode *inode, struct file *file)
143 {
144 	return single_open(file, alignment_proc_show, NULL);
145 }
146 
147 static ssize_t alignment_proc_write(struct file *file, const char __user *buffer,
148 				    size_t count, loff_t *pos)
149 {
150 	char mode;
151 
152 	if (count > 0) {
153 		if (get_user(mode, buffer))
154 			return -EFAULT;
155 		if (mode >= '0' && mode <= '5')
156 			ai_usermode = safe_usermode(mode - '0', true);
157 	}
158 	return count;
159 }
160 
161 static const struct file_operations alignment_proc_fops = {
162 	.open		= alignment_proc_open,
163 	.read		= seq_read,
164 	.llseek		= seq_lseek,
165 	.release	= single_release,
166 	.write		= alignment_proc_write,
167 };
168 #endif /* CONFIG_PROC_FS */
169 
170 union offset_union {
171 	unsigned long un;
172 	  signed long sn;
173 };
174 
175 #define TYPE_ERROR	0
176 #define TYPE_FAULT	1
177 #define TYPE_LDST	2
178 #define TYPE_DONE	3
179 
180 #ifdef __ARMEB__
181 #define BE		1
182 #define FIRST_BYTE_16	"mov	%1, %1, ror #8\n"
183 #define FIRST_BYTE_32	"mov	%1, %1, ror #24\n"
184 #define NEXT_BYTE	"ror #24"
185 #else
186 #define BE		0
187 #define FIRST_BYTE_16
188 #define FIRST_BYTE_32
189 #define NEXT_BYTE	"lsr #8"
190 #endif
191 
192 #define __get8_unaligned_check(ins,val,addr,err)	\
193 	__asm__(					\
194  ARM(	"1:	"ins"	%1, [%2], #1\n"	)		\
195  THUMB(	"1:	"ins"	%1, [%2]\n"	)		\
196  THUMB(	"	add	%2, %2, #1\n"	)		\
197 	"2:\n"						\
198 	"	.pushsection .fixup,\"ax\"\n"		\
199 	"	.align	2\n"				\
200 	"3:	mov	%0, #1\n"			\
201 	"	b	2b\n"				\
202 	"	.popsection\n"				\
203 	"	.pushsection __ex_table,\"a\"\n"	\
204 	"	.align	3\n"				\
205 	"	.long	1b, 3b\n"			\
206 	"	.popsection\n"				\
207 	: "=r" (err), "=&r" (val), "=r" (addr)		\
208 	: "0" (err), "2" (addr))
209 
210 #define __get16_unaligned_check(ins,val,addr)			\
211 	do {							\
212 		unsigned int err = 0, v, a = addr;		\
213 		__get8_unaligned_check(ins,v,a,err);		\
214 		val =  v << ((BE) ? 8 : 0);			\
215 		__get8_unaligned_check(ins,v,a,err);		\
216 		val |= v << ((BE) ? 0 : 8);			\
217 		if (err)					\
218 			goto fault;				\
219 	} while (0)
220 
221 #define get16_unaligned_check(val,addr) \
222 	__get16_unaligned_check("ldrb",val,addr)
223 
224 #define get16t_unaligned_check(val,addr) \
225 	__get16_unaligned_check("ldrbt",val,addr)
226 
227 #define __get32_unaligned_check(ins,val,addr)			\
228 	do {							\
229 		unsigned int err = 0, v, a = addr;		\
230 		__get8_unaligned_check(ins,v,a,err);		\
231 		val =  v << ((BE) ? 24 :  0);			\
232 		__get8_unaligned_check(ins,v,a,err);		\
233 		val |= v << ((BE) ? 16 :  8);			\
234 		__get8_unaligned_check(ins,v,a,err);		\
235 		val |= v << ((BE) ?  8 : 16);			\
236 		__get8_unaligned_check(ins,v,a,err);		\
237 		val |= v << ((BE) ?  0 : 24);			\
238 		if (err)					\
239 			goto fault;				\
240 	} while (0)
241 
242 #define get32_unaligned_check(val,addr) \
243 	__get32_unaligned_check("ldrb",val,addr)
244 
245 #define get32t_unaligned_check(val,addr) \
246 	__get32_unaligned_check("ldrbt",val,addr)
247 
248 #define __put16_unaligned_check(ins,val,addr)			\
249 	do {							\
250 		unsigned int err = 0, v = val, a = addr;	\
251 		__asm__( FIRST_BYTE_16				\
252 	 ARM(	"1:	"ins"	%1, [%2], #1\n"	)		\
253 	 THUMB(	"1:	"ins"	%1, [%2]\n"	)		\
254 	 THUMB(	"	add	%2, %2, #1\n"	)		\
255 		"	mov	%1, %1, "NEXT_BYTE"\n"		\
256 		"2:	"ins"	%1, [%2]\n"			\
257 		"3:\n"						\
258 		"	.pushsection .fixup,\"ax\"\n"		\
259 		"	.align	2\n"				\
260 		"4:	mov	%0, #1\n"			\
261 		"	b	3b\n"				\
262 		"	.popsection\n"				\
263 		"	.pushsection __ex_table,\"a\"\n"	\
264 		"	.align	3\n"				\
265 		"	.long	1b, 4b\n"			\
266 		"	.long	2b, 4b\n"			\
267 		"	.popsection\n"				\
268 		: "=r" (err), "=&r" (v), "=&r" (a)		\
269 		: "0" (err), "1" (v), "2" (a));			\
270 		if (err)					\
271 			goto fault;				\
272 	} while (0)
273 
274 #define put16_unaligned_check(val,addr)  \
275 	__put16_unaligned_check("strb",val,addr)
276 
277 #define put16t_unaligned_check(val,addr) \
278 	__put16_unaligned_check("strbt",val,addr)
279 
280 #define __put32_unaligned_check(ins,val,addr)			\
281 	do {							\
282 		unsigned int err = 0, v = val, a = addr;	\
283 		__asm__( FIRST_BYTE_32				\
284 	 ARM(	"1:	"ins"	%1, [%2], #1\n"	)		\
285 	 THUMB(	"1:	"ins"	%1, [%2]\n"	)		\
286 	 THUMB(	"	add	%2, %2, #1\n"	)		\
287 		"	mov	%1, %1, "NEXT_BYTE"\n"		\
288 	 ARM(	"2:	"ins"	%1, [%2], #1\n"	)		\
289 	 THUMB(	"2:	"ins"	%1, [%2]\n"	)		\
290 	 THUMB(	"	add	%2, %2, #1\n"	)		\
291 		"	mov	%1, %1, "NEXT_BYTE"\n"		\
292 	 ARM(	"3:	"ins"	%1, [%2], #1\n"	)		\
293 	 THUMB(	"3:	"ins"	%1, [%2]\n"	)		\
294 	 THUMB(	"	add	%2, %2, #1\n"	)		\
295 		"	mov	%1, %1, "NEXT_BYTE"\n"		\
296 		"4:	"ins"	%1, [%2]\n"			\
297 		"5:\n"						\
298 		"	.pushsection .fixup,\"ax\"\n"		\
299 		"	.align	2\n"				\
300 		"6:	mov	%0, #1\n"			\
301 		"	b	5b\n"				\
302 		"	.popsection\n"				\
303 		"	.pushsection __ex_table,\"a\"\n"	\
304 		"	.align	3\n"				\
305 		"	.long	1b, 6b\n"			\
306 		"	.long	2b, 6b\n"			\
307 		"	.long	3b, 6b\n"			\
308 		"	.long	4b, 6b\n"			\
309 		"	.popsection\n"				\
310 		: "=r" (err), "=&r" (v), "=&r" (a)		\
311 		: "0" (err), "1" (v), "2" (a));			\
312 		if (err)					\
313 			goto fault;				\
314 	} while (0)
315 
316 #define put32_unaligned_check(val,addr) \
317 	__put32_unaligned_check("strb", val, addr)
318 
319 #define put32t_unaligned_check(val,addr) \
320 	__put32_unaligned_check("strbt", val, addr)
321 
322 static void
323 do_alignment_finish_ldst(unsigned long addr, unsigned long instr, struct pt_regs *regs, union offset_union offset)
324 {
325 	if (!LDST_U_BIT(instr))
326 		offset.un = -offset.un;
327 
328 	if (!LDST_P_BIT(instr))
329 		addr += offset.un;
330 
331 	if (!LDST_P_BIT(instr) || LDST_W_BIT(instr))
332 		regs->uregs[RN_BITS(instr)] = addr;
333 }
334 
335 static int
336 do_alignment_ldrhstrh(unsigned long addr, unsigned long instr, struct pt_regs *regs)
337 {
338 	unsigned int rd = RD_BITS(instr);
339 
340 	ai_half += 1;
341 
342 	if (user_mode(regs))
343 		goto user;
344 
345 	if (LDST_L_BIT(instr)) {
346 		unsigned long val;
347 		get16_unaligned_check(val, addr);
348 
349 		/* signed half-word? */
350 		if (instr & 0x40)
351 			val = (signed long)((signed short) val);
352 
353 		regs->uregs[rd] = val;
354 	} else
355 		put16_unaligned_check(regs->uregs[rd], addr);
356 
357 	return TYPE_LDST;
358 
359  user:
360 	if (LDST_L_BIT(instr)) {
361 		unsigned long val;
362 		get16t_unaligned_check(val, addr);
363 
364 		/* signed half-word? */
365 		if (instr & 0x40)
366 			val = (signed long)((signed short) val);
367 
368 		regs->uregs[rd] = val;
369 	} else
370 		put16t_unaligned_check(regs->uregs[rd], addr);
371 
372 	return TYPE_LDST;
373 
374  fault:
375 	return TYPE_FAULT;
376 }
377 
378 static int
379 do_alignment_ldrdstrd(unsigned long addr, unsigned long instr,
380 		      struct pt_regs *regs)
381 {
382 	unsigned int rd = RD_BITS(instr);
383 	unsigned int rd2;
384 	int load;
385 
386 	if ((instr & 0xfe000000) == 0xe8000000) {
387 		/* ARMv7 Thumb-2 32-bit LDRD/STRD */
388 		rd2 = (instr >> 8) & 0xf;
389 		load = !!(LDST_L_BIT(instr));
390 	} else if (((rd & 1) == 1) || (rd == 14))
391 		goto bad;
392 	else {
393 		load = ((instr & 0xf0) == 0xd0);
394 		rd2 = rd + 1;
395 	}
396 
397 	ai_dword += 1;
398 
399 	if (user_mode(regs))
400 		goto user;
401 
402 	if (load) {
403 		unsigned long val;
404 		get32_unaligned_check(val, addr);
405 		regs->uregs[rd] = val;
406 		get32_unaligned_check(val, addr + 4);
407 		regs->uregs[rd2] = val;
408 	} else {
409 		put32_unaligned_check(regs->uregs[rd], addr);
410 		put32_unaligned_check(regs->uregs[rd2], addr + 4);
411 	}
412 
413 	return TYPE_LDST;
414 
415  user:
416 	if (load) {
417 		unsigned long val;
418 		get32t_unaligned_check(val, addr);
419 		regs->uregs[rd] = val;
420 		get32t_unaligned_check(val, addr + 4);
421 		regs->uregs[rd2] = val;
422 	} else {
423 		put32t_unaligned_check(regs->uregs[rd], addr);
424 		put32t_unaligned_check(regs->uregs[rd2], addr + 4);
425 	}
426 
427 	return TYPE_LDST;
428  bad:
429 	return TYPE_ERROR;
430  fault:
431 	return TYPE_FAULT;
432 }
433 
434 static int
435 do_alignment_ldrstr(unsigned long addr, unsigned long instr, struct pt_regs *regs)
436 {
437 	unsigned int rd = RD_BITS(instr);
438 
439 	ai_word += 1;
440 
441 	if ((!LDST_P_BIT(instr) && LDST_W_BIT(instr)) || user_mode(regs))
442 		goto trans;
443 
444 	if (LDST_L_BIT(instr)) {
445 		unsigned int val;
446 		get32_unaligned_check(val, addr);
447 		regs->uregs[rd] = val;
448 	} else
449 		put32_unaligned_check(regs->uregs[rd], addr);
450 	return TYPE_LDST;
451 
452  trans:
453 	if (LDST_L_BIT(instr)) {
454 		unsigned int val;
455 		get32t_unaligned_check(val, addr);
456 		regs->uregs[rd] = val;
457 	} else
458 		put32t_unaligned_check(regs->uregs[rd], addr);
459 	return TYPE_LDST;
460 
461  fault:
462 	return TYPE_FAULT;
463 }
464 
465 /*
466  * LDM/STM alignment handler.
467  *
468  * There are 4 variants of this instruction:
469  *
470  * B = rn pointer before instruction, A = rn pointer after instruction
471  *              ------ increasing address ----->
472  *	        |    | r0 | r1 | ... | rx |    |
473  * PU = 01             B                    A
474  * PU = 11        B                    A
475  * PU = 00        A                    B
476  * PU = 10             A                    B
477  */
478 static int
479 do_alignment_ldmstm(unsigned long addr, unsigned long instr, struct pt_regs *regs)
480 {
481 	unsigned int rd, rn, correction, nr_regs, regbits;
482 	unsigned long eaddr, newaddr;
483 
484 	if (LDM_S_BIT(instr))
485 		goto bad;
486 
487 	correction = 4; /* processor implementation defined */
488 	regs->ARM_pc += correction;
489 
490 	ai_multi += 1;
491 
492 	/* count the number of registers in the mask to be transferred */
493 	nr_regs = hweight16(REGMASK_BITS(instr)) * 4;
494 
495 	rn = RN_BITS(instr);
496 	newaddr = eaddr = regs->uregs[rn];
497 
498 	if (!LDST_U_BIT(instr))
499 		nr_regs = -nr_regs;
500 	newaddr += nr_regs;
501 	if (!LDST_U_BIT(instr))
502 		eaddr = newaddr;
503 
504 	if (LDST_P_EQ_U(instr))	/* U = P */
505 		eaddr += 4;
506 
507 	/*
508 	 * For alignment faults on the ARM922T/ARM920T the MMU  makes
509 	 * the FSR (and hence addr) equal to the updated base address
510 	 * of the multiple access rather than the restored value.
511 	 * Switch this message off if we've got a ARM92[02], otherwise
512 	 * [ls]dm alignment faults are noisy!
513 	 */
514 #if !(defined CONFIG_CPU_ARM922T)  && !(defined CONFIG_CPU_ARM920T)
515 	/*
516 	 * This is a "hint" - we already have eaddr worked out by the
517 	 * processor for us.
518 	 */
519 	if (addr != eaddr) {
520 		printk(KERN_ERR "LDMSTM: PC = %08lx, instr = %08lx, "
521 			"addr = %08lx, eaddr = %08lx\n",
522 			 instruction_pointer(regs), instr, addr, eaddr);
523 		show_regs(regs);
524 	}
525 #endif
526 
527 	if (user_mode(regs)) {
528 		for (regbits = REGMASK_BITS(instr), rd = 0; regbits;
529 		     regbits >>= 1, rd += 1)
530 			if (regbits & 1) {
531 				if (LDST_L_BIT(instr)) {
532 					unsigned int val;
533 					get32t_unaligned_check(val, eaddr);
534 					regs->uregs[rd] = val;
535 				} else
536 					put32t_unaligned_check(regs->uregs[rd], eaddr);
537 				eaddr += 4;
538 			}
539 	} else {
540 		for (regbits = REGMASK_BITS(instr), rd = 0; regbits;
541 		     regbits >>= 1, rd += 1)
542 			if (regbits & 1) {
543 				if (LDST_L_BIT(instr)) {
544 					unsigned int val;
545 					get32_unaligned_check(val, eaddr);
546 					regs->uregs[rd] = val;
547 				} else
548 					put32_unaligned_check(regs->uregs[rd], eaddr);
549 				eaddr += 4;
550 			}
551 	}
552 
553 	if (LDST_W_BIT(instr))
554 		regs->uregs[rn] = newaddr;
555 	if (!LDST_L_BIT(instr) || !(REGMASK_BITS(instr) & (1 << 15)))
556 		regs->ARM_pc -= correction;
557 	return TYPE_DONE;
558 
559 fault:
560 	regs->ARM_pc -= correction;
561 	return TYPE_FAULT;
562 
563 bad:
564 	printk(KERN_ERR "Alignment trap: not handling ldm with s-bit set\n");
565 	return TYPE_ERROR;
566 }
567 
568 /*
569  * Convert Thumb ld/st instruction forms to equivalent ARM instructions so
570  * we can reuse ARM userland alignment fault fixups for Thumb.
571  *
572  * This implementation was initially based on the algorithm found in
573  * gdb/sim/arm/thumbemu.c. It is basically just a code reduction of same
574  * to convert only Thumb ld/st instruction forms to equivalent ARM forms.
575  *
576  * NOTES:
577  * 1. Comments below refer to ARM ARM DDI0100E Thumb Instruction sections.
578  * 2. If for some reason we're passed an non-ld/st Thumb instruction to
579  *    decode, we return 0xdeadc0de. This should never happen under normal
580  *    circumstances but if it does, we've got other problems to deal with
581  *    elsewhere and we obviously can't fix those problems here.
582  */
583 
584 static unsigned long
585 thumb2arm(u16 tinstr)
586 {
587 	u32 L = (tinstr & (1<<11)) >> 11;
588 
589 	switch ((tinstr & 0xf800) >> 11) {
590 	/* 6.5.1 Format 1: */
591 	case 0x6000 >> 11:				/* 7.1.52 STR(1) */
592 	case 0x6800 >> 11:				/* 7.1.26 LDR(1) */
593 	case 0x7000 >> 11:				/* 7.1.55 STRB(1) */
594 	case 0x7800 >> 11:				/* 7.1.30 LDRB(1) */
595 		return 0xe5800000 |
596 			((tinstr & (1<<12)) << (22-12)) |	/* fixup */
597 			(L<<20) |				/* L==1? */
598 			((tinstr & (7<<0)) << (12-0)) |		/* Rd */
599 			((tinstr & (7<<3)) << (16-3)) |		/* Rn */
600 			((tinstr & (31<<6)) >>			/* immed_5 */
601 				(6 - ((tinstr & (1<<12)) ? 0 : 2)));
602 	case 0x8000 >> 11:				/* 7.1.57 STRH(1) */
603 	case 0x8800 >> 11:				/* 7.1.32 LDRH(1) */
604 		return 0xe1c000b0 |
605 			(L<<20) |				/* L==1? */
606 			((tinstr & (7<<0)) << (12-0)) |		/* Rd */
607 			((tinstr & (7<<3)) << (16-3)) |		/* Rn */
608 			((tinstr & (7<<6)) >> (6-1)) |	 /* immed_5[2:0] */
609 			((tinstr & (3<<9)) >> (9-8));	 /* immed_5[4:3] */
610 
611 	/* 6.5.1 Format 2: */
612 	case 0x5000 >> 11:
613 	case 0x5800 >> 11:
614 		{
615 			static const u32 subset[8] = {
616 				0xe7800000,		/* 7.1.53 STR(2) */
617 				0xe18000b0,		/* 7.1.58 STRH(2) */
618 				0xe7c00000,		/* 7.1.56 STRB(2) */
619 				0xe19000d0,		/* 7.1.34 LDRSB */
620 				0xe7900000,		/* 7.1.27 LDR(2) */
621 				0xe19000b0,		/* 7.1.33 LDRH(2) */
622 				0xe7d00000,		/* 7.1.31 LDRB(2) */
623 				0xe19000f0		/* 7.1.35 LDRSH */
624 			};
625 			return subset[(tinstr & (7<<9)) >> 9] |
626 			    ((tinstr & (7<<0)) << (12-0)) |	/* Rd */
627 			    ((tinstr & (7<<3)) << (16-3)) |	/* Rn */
628 			    ((tinstr & (7<<6)) >> (6-0));	/* Rm */
629 		}
630 
631 	/* 6.5.1 Format 3: */
632 	case 0x4800 >> 11:				/* 7.1.28 LDR(3) */
633 		/* NOTE: This case is not technically possible. We're
634 		 *	 loading 32-bit memory data via PC relative
635 		 *	 addressing mode. So we can and should eliminate
636 		 *	 this case. But I'll leave it here for now.
637 		 */
638 		return 0xe59f0000 |
639 		    ((tinstr & (7<<8)) << (12-8)) |		/* Rd */
640 		    ((tinstr & 255) << (2-0));			/* immed_8 */
641 
642 	/* 6.5.1 Format 4: */
643 	case 0x9000 >> 11:				/* 7.1.54 STR(3) */
644 	case 0x9800 >> 11:				/* 7.1.29 LDR(4) */
645 		return 0xe58d0000 |
646 			(L<<20) |				/* L==1? */
647 			((tinstr & (7<<8)) << (12-8)) |		/* Rd */
648 			((tinstr & 255) << 2);			/* immed_8 */
649 
650 	/* 6.6.1 Format 1: */
651 	case 0xc000 >> 11:				/* 7.1.51 STMIA */
652 	case 0xc800 >> 11:				/* 7.1.25 LDMIA */
653 		{
654 			u32 Rn = (tinstr & (7<<8)) >> 8;
655 			u32 W = ((L<<Rn) & (tinstr&255)) ? 0 : 1<<21;
656 
657 			return 0xe8800000 | W | (L<<20) | (Rn<<16) |
658 				(tinstr&255);
659 		}
660 
661 	/* 6.6.1 Format 2: */
662 	case 0xb000 >> 11:				/* 7.1.48 PUSH */
663 	case 0xb800 >> 11:				/* 7.1.47 POP */
664 		if ((tinstr & (3 << 9)) == 0x0400) {
665 			static const u32 subset[4] = {
666 				0xe92d0000,	/* STMDB sp!,{registers} */
667 				0xe92d4000,	/* STMDB sp!,{registers,lr} */
668 				0xe8bd0000,	/* LDMIA sp!,{registers} */
669 				0xe8bd8000	/* LDMIA sp!,{registers,pc} */
670 			};
671 			return subset[(L<<1) | ((tinstr & (1<<8)) >> 8)] |
672 			    (tinstr & 255);		/* register_list */
673 		}
674 		/* Else fall through for illegal instruction case */
675 
676 	default:
677 		return BAD_INSTR;
678 	}
679 }
680 
681 /*
682  * Convert Thumb-2 32 bit LDM, STM, LDRD, STRD to equivalent instruction
683  * handlable by ARM alignment handler, also find the corresponding handler,
684  * so that we can reuse ARM userland alignment fault fixups for Thumb.
685  *
686  * @pinstr: original Thumb-2 instruction; returns new handlable instruction
687  * @regs: register context.
688  * @poffset: return offset from faulted addr for later writeback
689  *
690  * NOTES:
691  * 1. Comments below refer to ARMv7 DDI0406A Thumb Instruction sections.
692  * 2. Register name Rt from ARMv7 is same as Rd from ARMv6 (Rd is Rt)
693  */
694 static void *
695 do_alignment_t32_to_handler(unsigned long *pinstr, struct pt_regs *regs,
696 			    union offset_union *poffset)
697 {
698 	unsigned long instr = *pinstr;
699 	u16 tinst1 = (instr >> 16) & 0xffff;
700 	u16 tinst2 = instr & 0xffff;
701 	poffset->un = 0;
702 
703 	switch (tinst1 & 0xffe0) {
704 	/* A6.3.5 Load/Store multiple */
705 	case 0xe880:		/* STM/STMIA/STMEA,LDM/LDMIA, PUSH/POP T2 */
706 	case 0xe8a0:		/* ...above writeback version */
707 	case 0xe900:		/* STMDB/STMFD, LDMDB/LDMEA */
708 	case 0xe920:		/* ...above writeback version */
709 		/* no need offset decision since handler calculates it */
710 		return do_alignment_ldmstm;
711 
712 	case 0xf840:		/* POP/PUSH T3 (single register) */
713 		if (RN_BITS(instr) == 13 && (tinst2 & 0x09ff) == 0x0904) {
714 			u32 L = !!(LDST_L_BIT(instr));
715 			const u32 subset[2] = {
716 				0xe92d0000,	/* STMDB sp!,{registers} */
717 				0xe8bd0000,	/* LDMIA sp!,{registers} */
718 			};
719 			*pinstr = subset[L] | (1<<RD_BITS(instr));
720 			return do_alignment_ldmstm;
721 		}
722 		/* Else fall through for illegal instruction case */
723 		break;
724 
725 	/* A6.3.6 Load/store double, STRD/LDRD(immed, lit, reg) */
726 	case 0xe860:
727 	case 0xe960:
728 	case 0xe8e0:
729 	case 0xe9e0:
730 		poffset->un = (tinst2 & 0xff) << 2;
731 	case 0xe940:
732 	case 0xe9c0:
733 		return do_alignment_ldrdstrd;
734 
735 	/*
736 	 * No need to handle load/store instructions up to word size
737 	 * since ARMv6 and later CPUs can perform unaligned accesses.
738 	 */
739 	default:
740 		break;
741 	}
742 	return NULL;
743 }
744 
745 static int
746 do_alignment(unsigned long addr, unsigned int fsr, struct pt_regs *regs)
747 {
748 	union offset_union offset;
749 	unsigned long instr = 0, instrptr;
750 	int (*handler)(unsigned long addr, unsigned long instr, struct pt_regs *regs);
751 	unsigned int type;
752 	mm_segment_t fs;
753 	unsigned int fault;
754 	u16 tinstr = 0;
755 	int isize = 4;
756 	int thumb2_32b = 0;
757 
758 	if (interrupts_enabled(regs))
759 		local_irq_enable();
760 
761 	instrptr = instruction_pointer(regs);
762 
763 	fs = get_fs();
764 	set_fs(KERNEL_DS);
765 	if (thumb_mode(regs)) {
766 		fault = __get_user(tinstr, (u16 *)(instrptr & ~1));
767 		if (!fault) {
768 			if (cpu_architecture() >= CPU_ARCH_ARMv7 &&
769 			    IS_T32(tinstr)) {
770 				/* Thumb-2 32-bit */
771 				u16 tinst2 = 0;
772 				fault = __get_user(tinst2, (u16 *)(instrptr+2));
773 				instr = (tinstr << 16) | tinst2;
774 				thumb2_32b = 1;
775 			} else {
776 				isize = 2;
777 				instr = thumb2arm(tinstr);
778 			}
779 		}
780 	} else
781 		fault = __get_user(instr, (u32 *)instrptr);
782 	set_fs(fs);
783 
784 	if (fault) {
785 		type = TYPE_FAULT;
786 		goto bad_or_fault;
787 	}
788 
789 	if (user_mode(regs))
790 		goto user;
791 
792 	ai_sys += 1;
793 
794  fixup:
795 
796 	regs->ARM_pc += isize;
797 
798 	switch (CODING_BITS(instr)) {
799 	case 0x00000000:	/* 3.13.4 load/store instruction extensions */
800 		if (LDSTHD_I_BIT(instr))
801 			offset.un = (instr & 0xf00) >> 4 | (instr & 15);
802 		else
803 			offset.un = regs->uregs[RM_BITS(instr)];
804 
805 		if ((instr & 0x000000f0) == 0x000000b0 || /* LDRH, STRH */
806 		    (instr & 0x001000f0) == 0x001000f0)   /* LDRSH */
807 			handler = do_alignment_ldrhstrh;
808 		else if ((instr & 0x001000f0) == 0x000000d0 || /* LDRD */
809 			 (instr & 0x001000f0) == 0x000000f0)   /* STRD */
810 			handler = do_alignment_ldrdstrd;
811 		else if ((instr & 0x01f00ff0) == 0x01000090) /* SWP */
812 			goto swp;
813 		else
814 			goto bad;
815 		break;
816 
817 	case 0x04000000:	/* ldr or str immediate */
818 		offset.un = OFFSET_BITS(instr);
819 		handler = do_alignment_ldrstr;
820 		break;
821 
822 	case 0x06000000:	/* ldr or str register */
823 		offset.un = regs->uregs[RM_BITS(instr)];
824 
825 		if (IS_SHIFT(instr)) {
826 			unsigned int shiftval = SHIFT_BITS(instr);
827 
828 			switch(SHIFT_TYPE(instr)) {
829 			case SHIFT_LSL:
830 				offset.un <<= shiftval;
831 				break;
832 
833 			case SHIFT_LSR:
834 				offset.un >>= shiftval;
835 				break;
836 
837 			case SHIFT_ASR:
838 				offset.sn >>= shiftval;
839 				break;
840 
841 			case SHIFT_RORRRX:
842 				if (shiftval == 0) {
843 					offset.un >>= 1;
844 					if (regs->ARM_cpsr & PSR_C_BIT)
845 						offset.un |= 1 << 31;
846 				} else
847 					offset.un = offset.un >> shiftval |
848 							  offset.un << (32 - shiftval);
849 				break;
850 			}
851 		}
852 		handler = do_alignment_ldrstr;
853 		break;
854 
855 	case 0x08000000:	/* ldm or stm, or thumb-2 32bit instruction */
856 		if (thumb2_32b)
857 			handler = do_alignment_t32_to_handler(&instr, regs, &offset);
858 		else
859 			handler = do_alignment_ldmstm;
860 		break;
861 
862 	default:
863 		goto bad;
864 	}
865 
866 	if (!handler)
867 		goto bad;
868 	type = handler(addr, instr, regs);
869 
870 	if (type == TYPE_ERROR || type == TYPE_FAULT) {
871 		regs->ARM_pc -= isize;
872 		goto bad_or_fault;
873 	}
874 
875 	if (type == TYPE_LDST)
876 		do_alignment_finish_ldst(addr, instr, regs, offset);
877 
878 	return 0;
879 
880  bad_or_fault:
881 	if (type == TYPE_ERROR)
882 		goto bad;
883 	/*
884 	 * We got a fault - fix it up, or die.
885 	 */
886 	do_bad_area(addr, fsr, regs);
887 	return 0;
888 
889  swp:
890 	printk(KERN_ERR "Alignment trap: not handling swp instruction\n");
891 
892  bad:
893 	/*
894 	 * Oops, we didn't handle the instruction.
895 	 */
896 	printk(KERN_ERR "Alignment trap: not handling instruction "
897 		"%0*lx at [<%08lx>]\n",
898 		isize << 1,
899 		isize == 2 ? tinstr : instr, instrptr);
900 	ai_skipped += 1;
901 	return 1;
902 
903  user:
904 	ai_user += 1;
905 
906 	if (ai_usermode & UM_WARN)
907 		printk("Alignment trap: %s (%d) PC=0x%08lx Instr=0x%0*lx "
908 		       "Address=0x%08lx FSR 0x%03x\n", current->comm,
909 			task_pid_nr(current), instrptr,
910 			isize << 1,
911 			isize == 2 ? tinstr : instr,
912 		        addr, fsr);
913 
914 	if (ai_usermode & UM_FIXUP)
915 		goto fixup;
916 
917 	if (ai_usermode & UM_SIGNAL) {
918 		siginfo_t si;
919 
920 		si.si_signo = SIGBUS;
921 		si.si_errno = 0;
922 		si.si_code = BUS_ADRALN;
923 		si.si_addr = (void __user *)addr;
924 
925 		force_sig_info(si.si_signo, &si, current);
926 	} else {
927 		/*
928 		 * We're about to disable the alignment trap and return to
929 		 * user space.  But if an interrupt occurs before actually
930 		 * reaching user space, then the IRQ vector entry code will
931 		 * notice that we were still in kernel space and therefore
932 		 * the alignment trap won't be re-enabled in that case as it
933 		 * is presumed to be always on from kernel space.
934 		 * Let's prevent that race by disabling interrupts here (they
935 		 * are disabled on the way back to user space anyway in
936 		 * entry-common.S) and disable the alignment trap only if
937 		 * there is no work pending for this thread.
938 		 */
939 		raw_local_irq_disable();
940 		if (!(current_thread_info()->flags & _TIF_WORK_MASK))
941 			set_cr(cr_no_alignment);
942 	}
943 
944 	return 0;
945 }
946 
947 /*
948  * This needs to be done after sysctl_init, otherwise sys/ will be
949  * overwritten.  Actually, this shouldn't be in sys/ at all since
950  * it isn't a sysctl, and it doesn't contain sysctl information.
951  * We now locate it in /proc/cpu/alignment instead.
952  */
953 static int __init alignment_init(void)
954 {
955 #ifdef CONFIG_PROC_FS
956 	struct proc_dir_entry *res;
957 
958 	res = proc_create("cpu/alignment", S_IWUSR | S_IRUGO, NULL,
959 			  &alignment_proc_fops);
960 	if (!res)
961 		return -ENOMEM;
962 #endif
963 
964 	if (cpu_is_v6_unaligned()) {
965 		cr_alignment &= ~CR_A;
966 		cr_no_alignment &= ~CR_A;
967 		set_cr(cr_alignment);
968 		ai_usermode = safe_usermode(ai_usermode, false);
969 	}
970 
971 	hook_fault_code(1, do_alignment, SIGBUS, BUS_ADRALN,
972 			"alignment exception");
973 
974 	/*
975 	 * ARMv6K and ARMv7 use fault status 3 (0b00011) as Access Flag section
976 	 * fault, not as alignment error.
977 	 *
978 	 * TODO: handle ARMv6K properly. Runtime check for 'K' extension is
979 	 * needed.
980 	 */
981 	if (cpu_architecture() <= CPU_ARCH_ARMv6) {
982 		hook_fault_code(3, do_alignment, SIGBUS, BUS_ADRALN,
983 				"alignment exception");
984 	}
985 
986 	return 0;
987 }
988 
989 fs_initcall(alignment_init);
990