xref: /openbmc/linux/arch/arm/mm/alignment.c (revision 77d84ff8)
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/cp15.h>
26 #include <asm/system_info.h>
27 #include <asm/unaligned.h>
28 #include <asm/opcodes.h>
29 
30 #include "fault.h"
31 
32 /*
33  * 32-bit misaligned trap handler (c) 1998 San Mehat (CCC) -July 1998
34  * /proc/sys/debug/alignment, modified and integrated into
35  * Linux 2.1 by Russell King
36  *
37  * Speed optimisations and better fault handling by Russell King.
38  *
39  * *** NOTE ***
40  * This code is not portable to processors with late data abort handling.
41  */
42 #define CODING_BITS(i)	(i & 0x0e000000)
43 
44 #define LDST_I_BIT(i)	(i & (1 << 26))		/* Immediate constant	*/
45 #define LDST_P_BIT(i)	(i & (1 << 24))		/* Preindex		*/
46 #define LDST_U_BIT(i)	(i & (1 << 23))		/* Add offset		*/
47 #define LDST_W_BIT(i)	(i & (1 << 21))		/* Writeback		*/
48 #define LDST_L_BIT(i)	(i & (1 << 20))		/* Load			*/
49 
50 #define LDST_P_EQ_U(i)	((((i) ^ ((i) >> 1)) & (1 << 23)) == 0)
51 
52 #define LDSTHD_I_BIT(i)	(i & (1 << 22))		/* double/half-word immed */
53 #define LDM_S_BIT(i)	(i & (1 << 22))		/* write CPSR from SPSR	*/
54 
55 #define RN_BITS(i)	((i >> 16) & 15)	/* Rn			*/
56 #define RD_BITS(i)	((i >> 12) & 15)	/* Rd			*/
57 #define RM_BITS(i)	(i & 15)		/* Rm			*/
58 
59 #define REGMASK_BITS(i)	(i & 0xffff)
60 #define OFFSET_BITS(i)	(i & 0x0fff)
61 
62 #define IS_SHIFT(i)	(i & 0x0ff0)
63 #define SHIFT_BITS(i)	((i >> 7) & 0x1f)
64 #define SHIFT_TYPE(i)	(i & 0x60)
65 #define SHIFT_LSL	0x00
66 #define SHIFT_LSR	0x20
67 #define SHIFT_ASR	0x40
68 #define SHIFT_RORRRX	0x60
69 
70 #define BAD_INSTR 	0xdeadc0de
71 
72 /* Thumb-2 32 bit format per ARMv7 DDI0406A A6.3, either f800h,e800h,f800h */
73 #define IS_T32(hi16) \
74 	(((hi16) & 0xe000) == 0xe000 && ((hi16) & 0x1800))
75 
76 static unsigned long ai_user;
77 static unsigned long ai_sys;
78 static unsigned long ai_skipped;
79 static unsigned long ai_half;
80 static unsigned long ai_word;
81 static unsigned long ai_dword;
82 static unsigned long ai_multi;
83 static int ai_usermode;
84 
85 core_param(alignment, ai_usermode, int, 0600);
86 
87 #define UM_WARN		(1 << 0)
88 #define UM_FIXUP	(1 << 1)
89 #define UM_SIGNAL	(1 << 2)
90 
91 /* Return true if and only if the ARMv6 unaligned access model is in use. */
92 static bool cpu_is_v6_unaligned(void)
93 {
94 	return cpu_architecture() >= CPU_ARCH_ARMv6 && (cr_alignment & CR_U);
95 }
96 
97 static int safe_usermode(int new_usermode, bool warn)
98 {
99 	/*
100 	 * ARMv6 and later CPUs can perform unaligned accesses for
101 	 * most single load and store instructions up to word size.
102 	 * LDM, STM, LDRD and STRD still need to be handled.
103 	 *
104 	 * Ignoring the alignment fault is not an option on these
105 	 * CPUs since we spin re-faulting the instruction without
106 	 * making any progress.
107 	 */
108 	if (cpu_is_v6_unaligned() && !(new_usermode & (UM_FIXUP | UM_SIGNAL))) {
109 		new_usermode |= UM_FIXUP;
110 
111 		if (warn)
112 			printk(KERN_WARNING "alignment: ignoring faults is unsafe on this CPU.  Defaulting to fixup mode.\n");
113 	}
114 
115 	return new_usermode;
116 }
117 
118 #ifdef CONFIG_PROC_FS
119 static const char *usermode_action[] = {
120 	"ignored",
121 	"warn",
122 	"fixup",
123 	"fixup+warn",
124 	"signal",
125 	"signal+warn"
126 };
127 
128 static int alignment_proc_show(struct seq_file *m, void *v)
129 {
130 	seq_printf(m, "User:\t\t%lu\n", ai_user);
131 	seq_printf(m, "System:\t\t%lu\n", ai_sys);
132 	seq_printf(m, "Skipped:\t%lu\n", ai_skipped);
133 	seq_printf(m, "Half:\t\t%lu\n", ai_half);
134 	seq_printf(m, "Word:\t\t%lu\n", ai_word);
135 	if (cpu_architecture() >= CPU_ARCH_ARMv5TE)
136 		seq_printf(m, "DWord:\t\t%lu\n", ai_dword);
137 	seq_printf(m, "Multi:\t\t%lu\n", ai_multi);
138 	seq_printf(m, "User faults:\t%i (%s)\n", ai_usermode,
139 			usermode_action[ai_usermode]);
140 
141 	return 0;
142 }
143 
144 static int alignment_proc_open(struct inode *inode, struct file *file)
145 {
146 	return single_open(file, alignment_proc_show, NULL);
147 }
148 
149 static ssize_t alignment_proc_write(struct file *file, const char __user *buffer,
150 				    size_t count, loff_t *pos)
151 {
152 	char mode;
153 
154 	if (count > 0) {
155 		if (get_user(mode, buffer))
156 			return -EFAULT;
157 		if (mode >= '0' && mode <= '5')
158 			ai_usermode = safe_usermode(mode - '0', true);
159 	}
160 	return count;
161 }
162 
163 static const struct file_operations alignment_proc_fops = {
164 	.open		= alignment_proc_open,
165 	.read		= seq_read,
166 	.llseek		= seq_lseek,
167 	.release	= single_release,
168 	.write		= alignment_proc_write,
169 };
170 #endif /* CONFIG_PROC_FS */
171 
172 union offset_union {
173 	unsigned long un;
174 	  signed long sn;
175 };
176 
177 #define TYPE_ERROR	0
178 #define TYPE_FAULT	1
179 #define TYPE_LDST	2
180 #define TYPE_DONE	3
181 
182 #ifdef __ARMEB__
183 #define BE		1
184 #define FIRST_BYTE_16	"mov	%1, %1, ror #8\n"
185 #define FIRST_BYTE_32	"mov	%1, %1, ror #24\n"
186 #define NEXT_BYTE	"ror #24"
187 #else
188 #define BE		0
189 #define FIRST_BYTE_16
190 #define FIRST_BYTE_32
191 #define NEXT_BYTE	"lsr #8"
192 #endif
193 
194 #define __get8_unaligned_check(ins,val,addr,err)	\
195 	__asm__(					\
196  ARM(	"1:	"ins"	%1, [%2], #1\n"	)		\
197  THUMB(	"1:	"ins"	%1, [%2]\n"	)		\
198  THUMB(	"	add	%2, %2, #1\n"	)		\
199 	"2:\n"						\
200 	"	.pushsection .fixup,\"ax\"\n"		\
201 	"	.align	2\n"				\
202 	"3:	mov	%0, #1\n"			\
203 	"	b	2b\n"				\
204 	"	.popsection\n"				\
205 	"	.pushsection __ex_table,\"a\"\n"	\
206 	"	.align	3\n"				\
207 	"	.long	1b, 3b\n"			\
208 	"	.popsection\n"				\
209 	: "=r" (err), "=&r" (val), "=r" (addr)		\
210 	: "0" (err), "2" (addr))
211 
212 #define __get16_unaligned_check(ins,val,addr)			\
213 	do {							\
214 		unsigned int err = 0, v, a = addr;		\
215 		__get8_unaligned_check(ins,v,a,err);		\
216 		val =  v << ((BE) ? 8 : 0);			\
217 		__get8_unaligned_check(ins,v,a,err);		\
218 		val |= v << ((BE) ? 0 : 8);			\
219 		if (err)					\
220 			goto fault;				\
221 	} while (0)
222 
223 #define get16_unaligned_check(val,addr) \
224 	__get16_unaligned_check("ldrb",val,addr)
225 
226 #define get16t_unaligned_check(val,addr) \
227 	__get16_unaligned_check("ldrbt",val,addr)
228 
229 #define __get32_unaligned_check(ins,val,addr)			\
230 	do {							\
231 		unsigned int err = 0, v, a = addr;		\
232 		__get8_unaligned_check(ins,v,a,err);		\
233 		val =  v << ((BE) ? 24 :  0);			\
234 		__get8_unaligned_check(ins,v,a,err);		\
235 		val |= v << ((BE) ? 16 :  8);			\
236 		__get8_unaligned_check(ins,v,a,err);		\
237 		val |= v << ((BE) ?  8 : 16);			\
238 		__get8_unaligned_check(ins,v,a,err);		\
239 		val |= v << ((BE) ?  0 : 24);			\
240 		if (err)					\
241 			goto fault;				\
242 	} while (0)
243 
244 #define get32_unaligned_check(val,addr) \
245 	__get32_unaligned_check("ldrb",val,addr)
246 
247 #define get32t_unaligned_check(val,addr) \
248 	__get32_unaligned_check("ldrbt",val,addr)
249 
250 #define __put16_unaligned_check(ins,val,addr)			\
251 	do {							\
252 		unsigned int err = 0, v = val, a = addr;	\
253 		__asm__( FIRST_BYTE_16				\
254 	 ARM(	"1:	"ins"	%1, [%2], #1\n"	)		\
255 	 THUMB(	"1:	"ins"	%1, [%2]\n"	)		\
256 	 THUMB(	"	add	%2, %2, #1\n"	)		\
257 		"	mov	%1, %1, "NEXT_BYTE"\n"		\
258 		"2:	"ins"	%1, [%2]\n"			\
259 		"3:\n"						\
260 		"	.pushsection .fixup,\"ax\"\n"		\
261 		"	.align	2\n"				\
262 		"4:	mov	%0, #1\n"			\
263 		"	b	3b\n"				\
264 		"	.popsection\n"				\
265 		"	.pushsection __ex_table,\"a\"\n"	\
266 		"	.align	3\n"				\
267 		"	.long	1b, 4b\n"			\
268 		"	.long	2b, 4b\n"			\
269 		"	.popsection\n"				\
270 		: "=r" (err), "=&r" (v), "=&r" (a)		\
271 		: "0" (err), "1" (v), "2" (a));			\
272 		if (err)					\
273 			goto fault;				\
274 	} while (0)
275 
276 #define put16_unaligned_check(val,addr)  \
277 	__put16_unaligned_check("strb",val,addr)
278 
279 #define put16t_unaligned_check(val,addr) \
280 	__put16_unaligned_check("strbt",val,addr)
281 
282 #define __put32_unaligned_check(ins,val,addr)			\
283 	do {							\
284 		unsigned int err = 0, v = val, a = addr;	\
285 		__asm__( FIRST_BYTE_32				\
286 	 ARM(	"1:	"ins"	%1, [%2], #1\n"	)		\
287 	 THUMB(	"1:	"ins"	%1, [%2]\n"	)		\
288 	 THUMB(	"	add	%2, %2, #1\n"	)		\
289 		"	mov	%1, %1, "NEXT_BYTE"\n"		\
290 	 ARM(	"2:	"ins"	%1, [%2], #1\n"	)		\
291 	 THUMB(	"2:	"ins"	%1, [%2]\n"	)		\
292 	 THUMB(	"	add	%2, %2, #1\n"	)		\
293 		"	mov	%1, %1, "NEXT_BYTE"\n"		\
294 	 ARM(	"3:	"ins"	%1, [%2], #1\n"	)		\
295 	 THUMB(	"3:	"ins"	%1, [%2]\n"	)		\
296 	 THUMB(	"	add	%2, %2, #1\n"	)		\
297 		"	mov	%1, %1, "NEXT_BYTE"\n"		\
298 		"4:	"ins"	%1, [%2]\n"			\
299 		"5:\n"						\
300 		"	.pushsection .fixup,\"ax\"\n"		\
301 		"	.align	2\n"				\
302 		"6:	mov	%0, #1\n"			\
303 		"	b	5b\n"				\
304 		"	.popsection\n"				\
305 		"	.pushsection __ex_table,\"a\"\n"	\
306 		"	.align	3\n"				\
307 		"	.long	1b, 6b\n"			\
308 		"	.long	2b, 6b\n"			\
309 		"	.long	3b, 6b\n"			\
310 		"	.long	4b, 6b\n"			\
311 		"	.popsection\n"				\
312 		: "=r" (err), "=&r" (v), "=&r" (a)		\
313 		: "0" (err), "1" (v), "2" (a));			\
314 		if (err)					\
315 			goto fault;				\
316 	} while (0)
317 
318 #define put32_unaligned_check(val,addr) \
319 	__put32_unaligned_check("strb", val, addr)
320 
321 #define put32t_unaligned_check(val,addr) \
322 	__put32_unaligned_check("strbt", val, addr)
323 
324 static void
325 do_alignment_finish_ldst(unsigned long addr, unsigned long instr, struct pt_regs *regs, union offset_union offset)
326 {
327 	if (!LDST_U_BIT(instr))
328 		offset.un = -offset.un;
329 
330 	if (!LDST_P_BIT(instr))
331 		addr += offset.un;
332 
333 	if (!LDST_P_BIT(instr) || LDST_W_BIT(instr))
334 		regs->uregs[RN_BITS(instr)] = addr;
335 }
336 
337 static int
338 do_alignment_ldrhstrh(unsigned long addr, unsigned long instr, struct pt_regs *regs)
339 {
340 	unsigned int rd = RD_BITS(instr);
341 
342 	ai_half += 1;
343 
344 	if (user_mode(regs))
345 		goto user;
346 
347 	if (LDST_L_BIT(instr)) {
348 		unsigned long val;
349 		get16_unaligned_check(val, addr);
350 
351 		/* signed half-word? */
352 		if (instr & 0x40)
353 			val = (signed long)((signed short) val);
354 
355 		regs->uregs[rd] = val;
356 	} else
357 		put16_unaligned_check(regs->uregs[rd], addr);
358 
359 	return TYPE_LDST;
360 
361  user:
362 	if (LDST_L_BIT(instr)) {
363 		unsigned long val;
364 		get16t_unaligned_check(val, addr);
365 
366 		/* signed half-word? */
367 		if (instr & 0x40)
368 			val = (signed long)((signed short) val);
369 
370 		regs->uregs[rd] = val;
371 	} else
372 		put16t_unaligned_check(regs->uregs[rd], addr);
373 
374 	return TYPE_LDST;
375 
376  fault:
377 	return TYPE_FAULT;
378 }
379 
380 static int
381 do_alignment_ldrdstrd(unsigned long addr, unsigned long instr,
382 		      struct pt_regs *regs)
383 {
384 	unsigned int rd = RD_BITS(instr);
385 	unsigned int rd2;
386 	int load;
387 
388 	if ((instr & 0xfe000000) == 0xe8000000) {
389 		/* ARMv7 Thumb-2 32-bit LDRD/STRD */
390 		rd2 = (instr >> 8) & 0xf;
391 		load = !!(LDST_L_BIT(instr));
392 	} else if (((rd & 1) == 1) || (rd == 14))
393 		goto bad;
394 	else {
395 		load = ((instr & 0xf0) == 0xd0);
396 		rd2 = rd + 1;
397 	}
398 
399 	ai_dword += 1;
400 
401 	if (user_mode(regs))
402 		goto user;
403 
404 	if (load) {
405 		unsigned long val;
406 		get32_unaligned_check(val, addr);
407 		regs->uregs[rd] = val;
408 		get32_unaligned_check(val, addr + 4);
409 		regs->uregs[rd2] = val;
410 	} else {
411 		put32_unaligned_check(regs->uregs[rd], addr);
412 		put32_unaligned_check(regs->uregs[rd2], addr + 4);
413 	}
414 
415 	return TYPE_LDST;
416 
417  user:
418 	if (load) {
419 		unsigned long val;
420 		get32t_unaligned_check(val, addr);
421 		regs->uregs[rd] = val;
422 		get32t_unaligned_check(val, addr + 4);
423 		regs->uregs[rd2] = val;
424 	} else {
425 		put32t_unaligned_check(regs->uregs[rd], addr);
426 		put32t_unaligned_check(regs->uregs[rd2], addr + 4);
427 	}
428 
429 	return TYPE_LDST;
430  bad:
431 	return TYPE_ERROR;
432  fault:
433 	return TYPE_FAULT;
434 }
435 
436 static int
437 do_alignment_ldrstr(unsigned long addr, unsigned long instr, struct pt_regs *regs)
438 {
439 	unsigned int rd = RD_BITS(instr);
440 
441 	ai_word += 1;
442 
443 	if ((!LDST_P_BIT(instr) && LDST_W_BIT(instr)) || user_mode(regs))
444 		goto trans;
445 
446 	if (LDST_L_BIT(instr)) {
447 		unsigned int val;
448 		get32_unaligned_check(val, addr);
449 		regs->uregs[rd] = val;
450 	} else
451 		put32_unaligned_check(regs->uregs[rd], addr);
452 	return TYPE_LDST;
453 
454  trans:
455 	if (LDST_L_BIT(instr)) {
456 		unsigned int val;
457 		get32t_unaligned_check(val, addr);
458 		regs->uregs[rd] = val;
459 	} else
460 		put32t_unaligned_check(regs->uregs[rd], addr);
461 	return TYPE_LDST;
462 
463  fault:
464 	return TYPE_FAULT;
465 }
466 
467 /*
468  * LDM/STM alignment handler.
469  *
470  * There are 4 variants of this instruction:
471  *
472  * B = rn pointer before instruction, A = rn pointer after instruction
473  *              ------ increasing address ----->
474  *	        |    | r0 | r1 | ... | rx |    |
475  * PU = 01             B                    A
476  * PU = 11        B                    A
477  * PU = 00        A                    B
478  * PU = 10             A                    B
479  */
480 static int
481 do_alignment_ldmstm(unsigned long addr, unsigned long instr, struct pt_regs *regs)
482 {
483 	unsigned int rd, rn, correction, nr_regs, regbits;
484 	unsigned long eaddr, newaddr;
485 
486 	if (LDM_S_BIT(instr))
487 		goto bad;
488 
489 	correction = 4; /* processor implementation defined */
490 	regs->ARM_pc += correction;
491 
492 	ai_multi += 1;
493 
494 	/* count the number of registers in the mask to be transferred */
495 	nr_regs = hweight16(REGMASK_BITS(instr)) * 4;
496 
497 	rn = RN_BITS(instr);
498 	newaddr = eaddr = regs->uregs[rn];
499 
500 	if (!LDST_U_BIT(instr))
501 		nr_regs = -nr_regs;
502 	newaddr += nr_regs;
503 	if (!LDST_U_BIT(instr))
504 		eaddr = newaddr;
505 
506 	if (LDST_P_EQ_U(instr))	/* U = P */
507 		eaddr += 4;
508 
509 	/*
510 	 * For alignment faults on the ARM922T/ARM920T the MMU  makes
511 	 * the FSR (and hence addr) equal to the updated base address
512 	 * of the multiple access rather than the restored value.
513 	 * Switch this message off if we've got a ARM92[02], otherwise
514 	 * [ls]dm alignment faults are noisy!
515 	 */
516 #if !(defined CONFIG_CPU_ARM922T)  && !(defined CONFIG_CPU_ARM920T)
517 	/*
518 	 * This is a "hint" - we already have eaddr worked out by the
519 	 * processor for us.
520 	 */
521 	if (addr != eaddr) {
522 		printk(KERN_ERR "LDMSTM: PC = %08lx, instr = %08lx, "
523 			"addr = %08lx, eaddr = %08lx\n",
524 			 instruction_pointer(regs), instr, addr, eaddr);
525 		show_regs(regs);
526 	}
527 #endif
528 
529 	if (user_mode(regs)) {
530 		for (regbits = REGMASK_BITS(instr), rd = 0; regbits;
531 		     regbits >>= 1, rd += 1)
532 			if (regbits & 1) {
533 				if (LDST_L_BIT(instr)) {
534 					unsigned int val;
535 					get32t_unaligned_check(val, eaddr);
536 					regs->uregs[rd] = val;
537 				} else
538 					put32t_unaligned_check(regs->uregs[rd], eaddr);
539 				eaddr += 4;
540 			}
541 	} else {
542 		for (regbits = REGMASK_BITS(instr), rd = 0; regbits;
543 		     regbits >>= 1, rd += 1)
544 			if (regbits & 1) {
545 				if (LDST_L_BIT(instr)) {
546 					unsigned int val;
547 					get32_unaligned_check(val, eaddr);
548 					regs->uregs[rd] = val;
549 				} else
550 					put32_unaligned_check(regs->uregs[rd], eaddr);
551 				eaddr += 4;
552 			}
553 	}
554 
555 	if (LDST_W_BIT(instr))
556 		regs->uregs[rn] = newaddr;
557 	if (!LDST_L_BIT(instr) || !(REGMASK_BITS(instr) & (1 << 15)))
558 		regs->ARM_pc -= correction;
559 	return TYPE_DONE;
560 
561 fault:
562 	regs->ARM_pc -= correction;
563 	return TYPE_FAULT;
564 
565 bad:
566 	printk(KERN_ERR "Alignment trap: not handling ldm with s-bit set\n");
567 	return TYPE_ERROR;
568 }
569 
570 /*
571  * Convert Thumb ld/st instruction forms to equivalent ARM instructions so
572  * we can reuse ARM userland alignment fault fixups for Thumb.
573  *
574  * This implementation was initially based on the algorithm found in
575  * gdb/sim/arm/thumbemu.c. It is basically just a code reduction of same
576  * to convert only Thumb ld/st instruction forms to equivalent ARM forms.
577  *
578  * NOTES:
579  * 1. Comments below refer to ARM ARM DDI0100E Thumb Instruction sections.
580  * 2. If for some reason we're passed an non-ld/st Thumb instruction to
581  *    decode, we return 0xdeadc0de. This should never happen under normal
582  *    circumstances but if it does, we've got other problems to deal with
583  *    elsewhere and we obviously can't fix those problems here.
584  */
585 
586 static unsigned long
587 thumb2arm(u16 tinstr)
588 {
589 	u32 L = (tinstr & (1<<11)) >> 11;
590 
591 	switch ((tinstr & 0xf800) >> 11) {
592 	/* 6.5.1 Format 1: */
593 	case 0x6000 >> 11:				/* 7.1.52 STR(1) */
594 	case 0x6800 >> 11:				/* 7.1.26 LDR(1) */
595 	case 0x7000 >> 11:				/* 7.1.55 STRB(1) */
596 	case 0x7800 >> 11:				/* 7.1.30 LDRB(1) */
597 		return 0xe5800000 |
598 			((tinstr & (1<<12)) << (22-12)) |	/* fixup */
599 			(L<<20) |				/* L==1? */
600 			((tinstr & (7<<0)) << (12-0)) |		/* Rd */
601 			((tinstr & (7<<3)) << (16-3)) |		/* Rn */
602 			((tinstr & (31<<6)) >>			/* immed_5 */
603 				(6 - ((tinstr & (1<<12)) ? 0 : 2)));
604 	case 0x8000 >> 11:				/* 7.1.57 STRH(1) */
605 	case 0x8800 >> 11:				/* 7.1.32 LDRH(1) */
606 		return 0xe1c000b0 |
607 			(L<<20) |				/* L==1? */
608 			((tinstr & (7<<0)) << (12-0)) |		/* Rd */
609 			((tinstr & (7<<3)) << (16-3)) |		/* Rn */
610 			((tinstr & (7<<6)) >> (6-1)) |	 /* immed_5[2:0] */
611 			((tinstr & (3<<9)) >> (9-8));	 /* immed_5[4:3] */
612 
613 	/* 6.5.1 Format 2: */
614 	case 0x5000 >> 11:
615 	case 0x5800 >> 11:
616 		{
617 			static const u32 subset[8] = {
618 				0xe7800000,		/* 7.1.53 STR(2) */
619 				0xe18000b0,		/* 7.1.58 STRH(2) */
620 				0xe7c00000,		/* 7.1.56 STRB(2) */
621 				0xe19000d0,		/* 7.1.34 LDRSB */
622 				0xe7900000,		/* 7.1.27 LDR(2) */
623 				0xe19000b0,		/* 7.1.33 LDRH(2) */
624 				0xe7d00000,		/* 7.1.31 LDRB(2) */
625 				0xe19000f0		/* 7.1.35 LDRSH */
626 			};
627 			return subset[(tinstr & (7<<9)) >> 9] |
628 			    ((tinstr & (7<<0)) << (12-0)) |	/* Rd */
629 			    ((tinstr & (7<<3)) << (16-3)) |	/* Rn */
630 			    ((tinstr & (7<<6)) >> (6-0));	/* Rm */
631 		}
632 
633 	/* 6.5.1 Format 3: */
634 	case 0x4800 >> 11:				/* 7.1.28 LDR(3) */
635 		/* NOTE: This case is not technically possible. We're
636 		 *	 loading 32-bit memory data via PC relative
637 		 *	 addressing mode. So we can and should eliminate
638 		 *	 this case. But I'll leave it here for now.
639 		 */
640 		return 0xe59f0000 |
641 		    ((tinstr & (7<<8)) << (12-8)) |		/* Rd */
642 		    ((tinstr & 255) << (2-0));			/* immed_8 */
643 
644 	/* 6.5.1 Format 4: */
645 	case 0x9000 >> 11:				/* 7.1.54 STR(3) */
646 	case 0x9800 >> 11:				/* 7.1.29 LDR(4) */
647 		return 0xe58d0000 |
648 			(L<<20) |				/* L==1? */
649 			((tinstr & (7<<8)) << (12-8)) |		/* Rd */
650 			((tinstr & 255) << 2);			/* immed_8 */
651 
652 	/* 6.6.1 Format 1: */
653 	case 0xc000 >> 11:				/* 7.1.51 STMIA */
654 	case 0xc800 >> 11:				/* 7.1.25 LDMIA */
655 		{
656 			u32 Rn = (tinstr & (7<<8)) >> 8;
657 			u32 W = ((L<<Rn) & (tinstr&255)) ? 0 : 1<<21;
658 
659 			return 0xe8800000 | W | (L<<20) | (Rn<<16) |
660 				(tinstr&255);
661 		}
662 
663 	/* 6.6.1 Format 2: */
664 	case 0xb000 >> 11:				/* 7.1.48 PUSH */
665 	case 0xb800 >> 11:				/* 7.1.47 POP */
666 		if ((tinstr & (3 << 9)) == 0x0400) {
667 			static const u32 subset[4] = {
668 				0xe92d0000,	/* STMDB sp!,{registers} */
669 				0xe92d4000,	/* STMDB sp!,{registers,lr} */
670 				0xe8bd0000,	/* LDMIA sp!,{registers} */
671 				0xe8bd8000	/* LDMIA sp!,{registers,pc} */
672 			};
673 			return subset[(L<<1) | ((tinstr & (1<<8)) >> 8)] |
674 			    (tinstr & 255);		/* register_list */
675 		}
676 		/* Else fall through for illegal instruction case */
677 
678 	default:
679 		return BAD_INSTR;
680 	}
681 }
682 
683 /*
684  * Convert Thumb-2 32 bit LDM, STM, LDRD, STRD to equivalent instruction
685  * handlable by ARM alignment handler, also find the corresponding handler,
686  * so that we can reuse ARM userland alignment fault fixups for Thumb.
687  *
688  * @pinstr: original Thumb-2 instruction; returns new handlable instruction
689  * @regs: register context.
690  * @poffset: return offset from faulted addr for later writeback
691  *
692  * NOTES:
693  * 1. Comments below refer to ARMv7 DDI0406A Thumb Instruction sections.
694  * 2. Register name Rt from ARMv7 is same as Rd from ARMv6 (Rd is Rt)
695  */
696 static void *
697 do_alignment_t32_to_handler(unsigned long *pinstr, struct pt_regs *regs,
698 			    union offset_union *poffset)
699 {
700 	unsigned long instr = *pinstr;
701 	u16 tinst1 = (instr >> 16) & 0xffff;
702 	u16 tinst2 = instr & 0xffff;
703 
704 	switch (tinst1 & 0xffe0) {
705 	/* A6.3.5 Load/Store multiple */
706 	case 0xe880:		/* STM/STMIA/STMEA,LDM/LDMIA, PUSH/POP T2 */
707 	case 0xe8a0:		/* ...above writeback version */
708 	case 0xe900:		/* STMDB/STMFD, LDMDB/LDMEA */
709 	case 0xe920:		/* ...above writeback version */
710 		/* no need offset decision since handler calculates it */
711 		return do_alignment_ldmstm;
712 
713 	case 0xf840:		/* POP/PUSH T3 (single register) */
714 		if (RN_BITS(instr) == 13 && (tinst2 & 0x09ff) == 0x0904) {
715 			u32 L = !!(LDST_L_BIT(instr));
716 			const u32 subset[2] = {
717 				0xe92d0000,	/* STMDB sp!,{registers} */
718 				0xe8bd0000,	/* LDMIA sp!,{registers} */
719 			};
720 			*pinstr = subset[L] | (1<<RD_BITS(instr));
721 			return do_alignment_ldmstm;
722 		}
723 		/* Else fall through for illegal instruction case */
724 		break;
725 
726 	/* A6.3.6 Load/store double, STRD/LDRD(immed, lit, reg) */
727 	case 0xe860:
728 	case 0xe960:
729 	case 0xe8e0:
730 	case 0xe9e0:
731 		poffset->un = (tinst2 & 0xff) << 2;
732 	case 0xe940:
733 	case 0xe9c0:
734 		return do_alignment_ldrdstrd;
735 
736 	/*
737 	 * No need to handle load/store instructions up to word size
738 	 * since ARMv6 and later CPUs can perform unaligned accesses.
739 	 */
740 	default:
741 		break;
742 	}
743 	return NULL;
744 }
745 
746 static int
747 do_alignment(unsigned long addr, unsigned int fsr, struct pt_regs *regs)
748 {
749 	union offset_union uninitialized_var(offset);
750 	unsigned long instr = 0, instrptr;
751 	int (*handler)(unsigned long addr, unsigned long instr, struct pt_regs *regs);
752 	unsigned int type;
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 	if (thumb_mode(regs)) {
764 		u16 *ptr = (u16 *)(instrptr & ~1);
765 		fault = probe_kernel_address(ptr, tinstr);
766 		tinstr = __mem_to_opcode_thumb16(tinstr);
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 = probe_kernel_address(ptr + 1, tinst2);
773 				tinst2 = __mem_to_opcode_thumb16(tinst2);
774 				instr = __opcode_thumb32_compose(tinstr, tinst2);
775 				thumb2_32b = 1;
776 			} else {
777 				isize = 2;
778 				instr = thumb2arm(tinstr);
779 			}
780 		}
781 	} else {
782 		fault = probe_kernel_address(instrptr, instr);
783 		instr = __mem_to_opcode_arm(instr);
784 	}
785 
786 	if (fault) {
787 		type = TYPE_FAULT;
788 		goto bad_or_fault;
789 	}
790 
791 	if (user_mode(regs))
792 		goto user;
793 
794 	ai_sys += 1;
795 
796  fixup:
797 
798 	regs->ARM_pc += isize;
799 
800 	switch (CODING_BITS(instr)) {
801 	case 0x00000000:	/* 3.13.4 load/store instruction extensions */
802 		if (LDSTHD_I_BIT(instr))
803 			offset.un = (instr & 0xf00) >> 4 | (instr & 15);
804 		else
805 			offset.un = regs->uregs[RM_BITS(instr)];
806 
807 		if ((instr & 0x000000f0) == 0x000000b0 || /* LDRH, STRH */
808 		    (instr & 0x001000f0) == 0x001000f0)   /* LDRSH */
809 			handler = do_alignment_ldrhstrh;
810 		else if ((instr & 0x001000f0) == 0x000000d0 || /* LDRD */
811 			 (instr & 0x001000f0) == 0x000000f0)   /* STRD */
812 			handler = do_alignment_ldrdstrd;
813 		else if ((instr & 0x01f00ff0) == 0x01000090) /* SWP */
814 			goto swp;
815 		else
816 			goto bad;
817 		break;
818 
819 	case 0x04000000:	/* ldr or str immediate */
820 		offset.un = OFFSET_BITS(instr);
821 		handler = do_alignment_ldrstr;
822 		break;
823 
824 	case 0x06000000:	/* ldr or str register */
825 		offset.un = regs->uregs[RM_BITS(instr)];
826 
827 		if (IS_SHIFT(instr)) {
828 			unsigned int shiftval = SHIFT_BITS(instr);
829 
830 			switch(SHIFT_TYPE(instr)) {
831 			case SHIFT_LSL:
832 				offset.un <<= shiftval;
833 				break;
834 
835 			case SHIFT_LSR:
836 				offset.un >>= shiftval;
837 				break;
838 
839 			case SHIFT_ASR:
840 				offset.sn >>= shiftval;
841 				break;
842 
843 			case SHIFT_RORRRX:
844 				if (shiftval == 0) {
845 					offset.un >>= 1;
846 					if (regs->ARM_cpsr & PSR_C_BIT)
847 						offset.un |= 1 << 31;
848 				} else
849 					offset.un = offset.un >> shiftval |
850 							  offset.un << (32 - shiftval);
851 				break;
852 			}
853 		}
854 		handler = do_alignment_ldrstr;
855 		break;
856 
857 	case 0x08000000:	/* ldm or stm, or thumb-2 32bit instruction */
858 		if (thumb2_32b) {
859 			offset.un = 0;
860 			handler = do_alignment_t32_to_handler(&instr, regs, &offset);
861 		} else {
862 			offset.un = 0;
863 			handler = do_alignment_ldmstm;
864 		}
865 		break;
866 
867 	default:
868 		goto bad;
869 	}
870 
871 	if (!handler)
872 		goto bad;
873 	type = handler(addr, instr, regs);
874 
875 	if (type == TYPE_ERROR || type == TYPE_FAULT) {
876 		regs->ARM_pc -= isize;
877 		goto bad_or_fault;
878 	}
879 
880 	if (type == TYPE_LDST)
881 		do_alignment_finish_ldst(addr, instr, regs, offset);
882 
883 	return 0;
884 
885  bad_or_fault:
886 	if (type == TYPE_ERROR)
887 		goto bad;
888 	/*
889 	 * We got a fault - fix it up, or die.
890 	 */
891 	do_bad_area(addr, fsr, regs);
892 	return 0;
893 
894  swp:
895 	printk(KERN_ERR "Alignment trap: not handling swp instruction\n");
896 
897  bad:
898 	/*
899 	 * Oops, we didn't handle the instruction.
900 	 */
901 	printk(KERN_ERR "Alignment trap: not handling instruction "
902 		"%0*lx at [<%08lx>]\n",
903 		isize << 1,
904 		isize == 2 ? tinstr : instr, instrptr);
905 	ai_skipped += 1;
906 	return 1;
907 
908  user:
909 	ai_user += 1;
910 
911 	if (ai_usermode & UM_WARN)
912 		printk("Alignment trap: %s (%d) PC=0x%08lx Instr=0x%0*lx "
913 		       "Address=0x%08lx FSR 0x%03x\n", current->comm,
914 			task_pid_nr(current), instrptr,
915 			isize << 1,
916 			isize == 2 ? tinstr : instr,
917 		        addr, fsr);
918 
919 	if (ai_usermode & UM_FIXUP)
920 		goto fixup;
921 
922 	if (ai_usermode & UM_SIGNAL) {
923 		siginfo_t si;
924 
925 		si.si_signo = SIGBUS;
926 		si.si_errno = 0;
927 		si.si_code = BUS_ADRALN;
928 		si.si_addr = (void __user *)addr;
929 
930 		force_sig_info(si.si_signo, &si, current);
931 	} else {
932 		/*
933 		 * We're about to disable the alignment trap and return to
934 		 * user space.  But if an interrupt occurs before actually
935 		 * reaching user space, then the IRQ vector entry code will
936 		 * notice that we were still in kernel space and therefore
937 		 * the alignment trap won't be re-enabled in that case as it
938 		 * is presumed to be always on from kernel space.
939 		 * Let's prevent that race by disabling interrupts here (they
940 		 * are disabled on the way back to user space anyway in
941 		 * entry-common.S) and disable the alignment trap only if
942 		 * there is no work pending for this thread.
943 		 */
944 		raw_local_irq_disable();
945 		if (!(current_thread_info()->flags & _TIF_WORK_MASK))
946 			set_cr(cr_no_alignment);
947 	}
948 
949 	return 0;
950 }
951 
952 /*
953  * This needs to be done after sysctl_init, otherwise sys/ will be
954  * overwritten.  Actually, this shouldn't be in sys/ at all since
955  * it isn't a sysctl, and it doesn't contain sysctl information.
956  * We now locate it in /proc/cpu/alignment instead.
957  */
958 static int __init alignment_init(void)
959 {
960 #ifdef CONFIG_PROC_FS
961 	struct proc_dir_entry *res;
962 
963 	res = proc_create("cpu/alignment", S_IWUSR | S_IRUGO, NULL,
964 			  &alignment_proc_fops);
965 	if (!res)
966 		return -ENOMEM;
967 #endif
968 
969 #ifdef CONFIG_CPU_CP15
970 	if (cpu_is_v6_unaligned()) {
971 		cr_alignment &= ~CR_A;
972 		cr_no_alignment &= ~CR_A;
973 		set_cr(cr_alignment);
974 		ai_usermode = safe_usermode(ai_usermode, false);
975 	}
976 #endif
977 
978 	hook_fault_code(FAULT_CODE_ALIGNMENT, do_alignment, SIGBUS, BUS_ADRALN,
979 			"alignment exception");
980 
981 	/*
982 	 * ARMv6K and ARMv7 use fault status 3 (0b00011) as Access Flag section
983 	 * fault, not as alignment error.
984 	 *
985 	 * TODO: handle ARMv6K properly. Runtime check for 'K' extension is
986 	 * needed.
987 	 */
988 	if (cpu_architecture() <= CPU_ARCH_ARMv6) {
989 		hook_fault_code(3, do_alignment, SIGBUS, BUS_ADRALN,
990 				"alignment exception");
991 	}
992 
993 	return 0;
994 }
995 
996 fs_initcall(alignment_init);
997