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#include "mm.h"
32
33/*
34 * 32-bit misaligned trap handler (c) 1998 San Mehat (CCC) -July 1998
35 * /proc/sys/debug/alignment, modified and integrated into
36 * Linux 2.1 by Russell King
37 *
38 * Speed optimisations and better fault handling by Russell King.
39 *
40 * *** NOTE ***
41 * This code is not portable to processors with late data abort handling.
42 */
43#define CODING_BITS(i)	(i & 0x0e000000)
44#define COND_BITS(i)	(i & 0xf0000000)
45
46#define LDST_I_BIT(i)	(i & (1 << 26))		/* Immediate constant	*/
47#define LDST_P_BIT(i)	(i & (1 << 24))		/* Preindex		*/
48#define LDST_U_BIT(i)	(i & (1 << 23))		/* Add offset		*/
49#define LDST_W_BIT(i)	(i & (1 << 21))		/* Writeback		*/
50#define LDST_L_BIT(i)	(i & (1 << 20))		/* Load			*/
51
52#define LDST_P_EQ_U(i)	((((i) ^ ((i) >> 1)) & (1 << 23)) == 0)
53
54#define LDSTHD_I_BIT(i)	(i & (1 << 22))		/* double/half-word immed */
55#define LDM_S_BIT(i)	(i & (1 << 22))		/* write CPSR from SPSR	*/
56
57#define RN_BITS(i)	((i >> 16) & 15)	/* Rn			*/
58#define RD_BITS(i)	((i >> 12) & 15)	/* Rd			*/
59#define RM_BITS(i)	(i & 15)		/* Rm			*/
60
61#define REGMASK_BITS(i)	(i & 0xffff)
62#define OFFSET_BITS(i)	(i & 0x0fff)
63
64#define IS_SHIFT(i)	(i & 0x0ff0)
65#define SHIFT_BITS(i)	((i >> 7) & 0x1f)
66#define SHIFT_TYPE(i)	(i & 0x60)
67#define SHIFT_LSL	0x00
68#define SHIFT_LSR	0x20
69#define SHIFT_ASR	0x40
70#define SHIFT_RORRRX	0x60
71
72#define BAD_INSTR 	0xdeadc0de
73
74/* Thumb-2 32 bit format per ARMv7 DDI0406A A6.3, either f800h,e800h,f800h */
75#define IS_T32(hi16) \
76	(((hi16) & 0xe000) == 0xe000 && ((hi16) & 0x1800))
77
78static unsigned long ai_user;
79static unsigned long ai_sys;
80static void *ai_sys_last_pc;
81static unsigned long ai_skipped;
82static unsigned long ai_half;
83static unsigned long ai_word;
84static unsigned long ai_dword;
85static unsigned long ai_multi;
86static int ai_usermode;
87static unsigned long cr_no_alignment;
88
89core_param(alignment, ai_usermode, int, 0600);
90
91#define UM_WARN		(1 << 0)
92#define UM_FIXUP	(1 << 1)
93#define UM_SIGNAL	(1 << 2)
94
95/* Return true if and only if the ARMv6 unaligned access model is in use. */
96static bool cpu_is_v6_unaligned(void)
97{
98	return cpu_architecture() >= CPU_ARCH_ARMv6 && get_cr() & CR_U;
99}
100
101static int safe_usermode(int new_usermode, bool warn)
102{
103	/*
104	 * ARMv6 and later CPUs can perform unaligned accesses for
105	 * most single load and store instructions up to word size.
106	 * LDM, STM, LDRD and STRD still need to be handled.
107	 *
108	 * Ignoring the alignment fault is not an option on these
109	 * CPUs since we spin re-faulting the instruction without
110	 * making any progress.
111	 */
112	if (cpu_is_v6_unaligned() && !(new_usermode & (UM_FIXUP | UM_SIGNAL))) {
113		new_usermode |= UM_FIXUP;
114
115		if (warn)
116			pr_warn("alignment: ignoring faults is unsafe on this CPU.  Defaulting to fixup mode.\n");
117	}
118
119	return new_usermode;
120}
121
122#ifdef CONFIG_PROC_FS
123static const char *usermode_action[] = {
124	"ignored",
125	"warn",
126	"fixup",
127	"fixup+warn",
128	"signal",
129	"signal+warn"
130};
131
132static int alignment_proc_show(struct seq_file *m, void *v)
133{
134	seq_printf(m, "User:\t\t%lu\n", ai_user);
135	seq_printf(m, "System:\t\t%lu (%pF)\n", ai_sys, ai_sys_last_pc);
136	seq_printf(m, "Skipped:\t%lu\n", ai_skipped);
137	seq_printf(m, "Half:\t\t%lu\n", ai_half);
138	seq_printf(m, "Word:\t\t%lu\n", ai_word);
139	if (cpu_architecture() >= CPU_ARCH_ARMv5TE)
140		seq_printf(m, "DWord:\t\t%lu\n", ai_dword);
141	seq_printf(m, "Multi:\t\t%lu\n", ai_multi);
142	seq_printf(m, "User faults:\t%i (%s)\n", ai_usermode,
143			usermode_action[ai_usermode]);
144
145	return 0;
146}
147
148static int alignment_proc_open(struct inode *inode, struct file *file)
149{
150	return single_open(file, alignment_proc_show, NULL);
151}
152
153static ssize_t alignment_proc_write(struct file *file, const char __user *buffer,
154				    size_t count, loff_t *pos)
155{
156	char mode;
157
158	if (count > 0) {
159		if (get_user(mode, buffer))
160			return -EFAULT;
161		if (mode >= '0' && mode <= '5')
162			ai_usermode = safe_usermode(mode - '0', true);
163	}
164	return count;
165}
166
167static const struct file_operations alignment_proc_fops = {
168	.open		= alignment_proc_open,
169	.read		= seq_read,
170	.llseek		= seq_lseek,
171	.release	= single_release,
172	.write		= alignment_proc_write,
173};
174#endif /* CONFIG_PROC_FS */
175
176union offset_union {
177	unsigned long un;
178	  signed long sn;
179};
180
181#define TYPE_ERROR	0
182#define TYPE_FAULT	1
183#define TYPE_LDST	2
184#define TYPE_DONE	3
185
186#ifdef __ARMEB__
187#define BE		1
188#define FIRST_BYTE_16	"mov	%1, %1, ror #8\n"
189#define FIRST_BYTE_32	"mov	%1, %1, ror #24\n"
190#define NEXT_BYTE	"ror #24"
191#else
192#define BE		0
193#define FIRST_BYTE_16
194#define FIRST_BYTE_32
195#define NEXT_BYTE	"lsr #8"
196#endif
197
198#define __get8_unaligned_check(ins,val,addr,err)	\
199	__asm__(					\
200 ARM(	"1:	"ins"	%1, [%2], #1\n"	)		\
201 THUMB(	"1:	"ins"	%1, [%2]\n"	)		\
202 THUMB(	"	add	%2, %2, #1\n"	)		\
203	"2:\n"						\
204	"	.pushsection .text.fixup,\"ax\"\n"	\
205	"	.align	2\n"				\
206	"3:	mov	%0, #1\n"			\
207	"	b	2b\n"				\
208	"	.popsection\n"				\
209	"	.pushsection __ex_table,\"a\"\n"	\
210	"	.align	3\n"				\
211	"	.long	1b, 3b\n"			\
212	"	.popsection\n"				\
213	: "=r" (err), "=&r" (val), "=r" (addr)		\
214	: "0" (err), "2" (addr))
215
216#define __get16_unaligned_check(ins,val,addr)			\
217	do {							\
218		unsigned int err = 0, v, a = addr;		\
219		__get8_unaligned_check(ins,v,a,err);		\
220		val =  v << ((BE) ? 8 : 0);			\
221		__get8_unaligned_check(ins,v,a,err);		\
222		val |= v << ((BE) ? 0 : 8);			\
223		if (err)					\
224			goto fault;				\
225	} while (0)
226
227#define get16_unaligned_check(val,addr) \
228	__get16_unaligned_check("ldrb",val,addr)
229
230#define get16t_unaligned_check(val,addr) \
231	__get16_unaligned_check("ldrbt",val,addr)
232
233#define __get32_unaligned_check(ins,val,addr)			\
234	do {							\
235		unsigned int err = 0, v, a = addr;		\
236		__get8_unaligned_check(ins,v,a,err);		\
237		val =  v << ((BE) ? 24 :  0);			\
238		__get8_unaligned_check(ins,v,a,err);		\
239		val |= v << ((BE) ? 16 :  8);			\
240		__get8_unaligned_check(ins,v,a,err);		\
241		val |= v << ((BE) ?  8 : 16);			\
242		__get8_unaligned_check(ins,v,a,err);		\
243		val |= v << ((BE) ?  0 : 24);			\
244		if (err)					\
245			goto fault;				\
246	} while (0)
247
248#define get32_unaligned_check(val,addr) \
249	__get32_unaligned_check("ldrb",val,addr)
250
251#define get32t_unaligned_check(val,addr) \
252	__get32_unaligned_check("ldrbt",val,addr)
253
254#define __put16_unaligned_check(ins,val,addr)			\
255	do {							\
256		unsigned int err = 0, v = val, a = addr;	\
257		__asm__( FIRST_BYTE_16				\
258	 ARM(	"1:	"ins"	%1, [%2], #1\n"	)		\
259	 THUMB(	"1:	"ins"	%1, [%2]\n"	)		\
260	 THUMB(	"	add	%2, %2, #1\n"	)		\
261		"	mov	%1, %1, "NEXT_BYTE"\n"		\
262		"2:	"ins"	%1, [%2]\n"			\
263		"3:\n"						\
264		"	.pushsection .text.fixup,\"ax\"\n"	\
265		"	.align	2\n"				\
266		"4:	mov	%0, #1\n"			\
267		"	b	3b\n"				\
268		"	.popsection\n"				\
269		"	.pushsection __ex_table,\"a\"\n"	\
270		"	.align	3\n"				\
271		"	.long	1b, 4b\n"			\
272		"	.long	2b, 4b\n"			\
273		"	.popsection\n"				\
274		: "=r" (err), "=&r" (v), "=&r" (a)		\
275		: "0" (err), "1" (v), "2" (a));			\
276		if (err)					\
277			goto fault;				\
278	} while (0)
279
280#define put16_unaligned_check(val,addr)  \
281	__put16_unaligned_check("strb",val,addr)
282
283#define put16t_unaligned_check(val,addr) \
284	__put16_unaligned_check("strbt",val,addr)
285
286#define __put32_unaligned_check(ins,val,addr)			\
287	do {							\
288		unsigned int err = 0, v = val, a = addr;	\
289		__asm__( FIRST_BYTE_32				\
290	 ARM(	"1:	"ins"	%1, [%2], #1\n"	)		\
291	 THUMB(	"1:	"ins"	%1, [%2]\n"	)		\
292	 THUMB(	"	add	%2, %2, #1\n"	)		\
293		"	mov	%1, %1, "NEXT_BYTE"\n"		\
294	 ARM(	"2:	"ins"	%1, [%2], #1\n"	)		\
295	 THUMB(	"2:	"ins"	%1, [%2]\n"	)		\
296	 THUMB(	"	add	%2, %2, #1\n"	)		\
297		"	mov	%1, %1, "NEXT_BYTE"\n"		\
298	 ARM(	"3:	"ins"	%1, [%2], #1\n"	)		\
299	 THUMB(	"3:	"ins"	%1, [%2]\n"	)		\
300	 THUMB(	"	add	%2, %2, #1\n"	)		\
301		"	mov	%1, %1, "NEXT_BYTE"\n"		\
302		"4:	"ins"	%1, [%2]\n"			\
303		"5:\n"						\
304		"	.pushsection .text.fixup,\"ax\"\n"	\
305		"	.align	2\n"				\
306		"6:	mov	%0, #1\n"			\
307		"	b	5b\n"				\
308		"	.popsection\n"				\
309		"	.pushsection __ex_table,\"a\"\n"	\
310		"	.align	3\n"				\
311		"	.long	1b, 6b\n"			\
312		"	.long	2b, 6b\n"			\
313		"	.long	3b, 6b\n"			\
314		"	.long	4b, 6b\n"			\
315		"	.popsection\n"				\
316		: "=r" (err), "=&r" (v), "=&r" (a)		\
317		: "0" (err), "1" (v), "2" (a));			\
318		if (err)					\
319			goto fault;				\
320	} while (0)
321
322#define put32_unaligned_check(val,addr) \
323	__put32_unaligned_check("strb", val, addr)
324
325#define put32t_unaligned_check(val,addr) \
326	__put32_unaligned_check("strbt", val, addr)
327
328static void
329do_alignment_finish_ldst(unsigned long addr, unsigned long instr, struct pt_regs *regs, union offset_union offset)
330{
331	if (!LDST_U_BIT(instr))
332		offset.un = -offset.un;
333
334	if (!LDST_P_BIT(instr))
335		addr += offset.un;
336
337	if (!LDST_P_BIT(instr) || LDST_W_BIT(instr))
338		regs->uregs[RN_BITS(instr)] = addr;
339}
340
341static int
342do_alignment_ldrhstrh(unsigned long addr, unsigned long instr, struct pt_regs *regs)
343{
344	unsigned int rd = RD_BITS(instr);
345
346	ai_half += 1;
347
348	if (user_mode(regs))
349		goto user;
350
351	if (LDST_L_BIT(instr)) {
352		unsigned long val;
353		get16_unaligned_check(val, addr);
354
355		/* signed half-word? */
356		if (instr & 0x40)
357			val = (signed long)((signed short) val);
358
359		regs->uregs[rd] = val;
360	} else
361		put16_unaligned_check(regs->uregs[rd], addr);
362
363	return TYPE_LDST;
364
365 user:
366	if (LDST_L_BIT(instr)) {
367		unsigned long val;
368		unsigned int __ua_flags = uaccess_save_and_enable();
369
370		get16t_unaligned_check(val, addr);
371		uaccess_restore(__ua_flags);
372
373		/* signed half-word? */
374		if (instr & 0x40)
375			val = (signed long)((signed short) val);
376
377		regs->uregs[rd] = val;
378	} else {
379		unsigned int __ua_flags = uaccess_save_and_enable();
380		put16t_unaligned_check(regs->uregs[rd], addr);
381		uaccess_restore(__ua_flags);
382	}
383
384	return TYPE_LDST;
385
386 fault:
387	return TYPE_FAULT;
388}
389
390static int
391do_alignment_ldrdstrd(unsigned long addr, unsigned long instr,
392		      struct pt_regs *regs)
393{
394	unsigned int rd = RD_BITS(instr);
395	unsigned int rd2;
396	int load;
397
398	if ((instr & 0xfe000000) == 0xe8000000) {
399		/* ARMv7 Thumb-2 32-bit LDRD/STRD */
400		rd2 = (instr >> 8) & 0xf;
401		load = !!(LDST_L_BIT(instr));
402	} else if (((rd & 1) == 1) || (rd == 14))
403		goto bad;
404	else {
405		load = ((instr & 0xf0) == 0xd0);
406		rd2 = rd + 1;
407	}
408
409	ai_dword += 1;
410
411	if (user_mode(regs))
412		goto user;
413
414	if (load) {
415		unsigned long val;
416		get32_unaligned_check(val, addr);
417		regs->uregs[rd] = val;
418		get32_unaligned_check(val, addr + 4);
419		regs->uregs[rd2] = val;
420	} else {
421		put32_unaligned_check(regs->uregs[rd], addr);
422		put32_unaligned_check(regs->uregs[rd2], addr + 4);
423	}
424
425	return TYPE_LDST;
426
427 user:
428	if (load) {
429		unsigned long val, val2;
430		unsigned int __ua_flags = uaccess_save_and_enable();
431
432		get32t_unaligned_check(val, addr);
433		get32t_unaligned_check(val2, addr + 4);
434
435		uaccess_restore(__ua_flags);
436
437		regs->uregs[rd] = val;
438		regs->uregs[rd2] = val2;
439	} else {
440		unsigned int __ua_flags = uaccess_save_and_enable();
441		put32t_unaligned_check(regs->uregs[rd], addr);
442		put32t_unaligned_check(regs->uregs[rd2], addr + 4);
443		uaccess_restore(__ua_flags);
444	}
445
446	return TYPE_LDST;
447 bad:
448	return TYPE_ERROR;
449 fault:
450	return TYPE_FAULT;
451}
452
453static int
454do_alignment_ldrstr(unsigned long addr, unsigned long instr, struct pt_regs *regs)
455{
456	unsigned int rd = RD_BITS(instr);
457
458	ai_word += 1;
459
460	if ((!LDST_P_BIT(instr) && LDST_W_BIT(instr)) || user_mode(regs))
461		goto trans;
462
463	if (LDST_L_BIT(instr)) {
464		unsigned int val;
465		get32_unaligned_check(val, addr);
466		regs->uregs[rd] = val;
467	} else
468		put32_unaligned_check(regs->uregs[rd], addr);
469	return TYPE_LDST;
470
471 trans:
472	if (LDST_L_BIT(instr)) {
473		unsigned int val;
474		unsigned int __ua_flags = uaccess_save_and_enable();
475		get32t_unaligned_check(val, addr);
476		uaccess_restore(__ua_flags);
477		regs->uregs[rd] = val;
478	} else {
479		unsigned int __ua_flags = uaccess_save_and_enable();
480		put32t_unaligned_check(regs->uregs[rd], addr);
481		uaccess_restore(__ua_flags);
482	}
483	return TYPE_LDST;
484
485 fault:
486	return TYPE_FAULT;
487}
488
489/*
490 * LDM/STM alignment handler.
491 *
492 * There are 4 variants of this instruction:
493 *
494 * B = rn pointer before instruction, A = rn pointer after instruction
495 *              ------ increasing address ----->
496 *	        |    | r0 | r1 | ... | rx |    |
497 * PU = 01             B                    A
498 * PU = 11        B                    A
499 * PU = 00        A                    B
500 * PU = 10             A                    B
501 */
502static int
503do_alignment_ldmstm(unsigned long addr, unsigned long instr, struct pt_regs *regs)
504{
505	unsigned int rd, rn, correction, nr_regs, regbits;
506	unsigned long eaddr, newaddr;
507
508	if (LDM_S_BIT(instr))
509		goto bad;
510
511	correction = 4; /* processor implementation defined */
512	regs->ARM_pc += correction;
513
514	ai_multi += 1;
515
516	/* count the number of registers in the mask to be transferred */
517	nr_regs = hweight16(REGMASK_BITS(instr)) * 4;
518
519	rn = RN_BITS(instr);
520	newaddr = eaddr = regs->uregs[rn];
521
522	if (!LDST_U_BIT(instr))
523		nr_regs = -nr_regs;
524	newaddr += nr_regs;
525	if (!LDST_U_BIT(instr))
526		eaddr = newaddr;
527
528	if (LDST_P_EQ_U(instr))	/* U = P */
529		eaddr += 4;
530
531	/*
532	 * For alignment faults on the ARM922T/ARM920T the MMU  makes
533	 * the FSR (and hence addr) equal to the updated base address
534	 * of the multiple access rather than the restored value.
535	 * Switch this message off if we've got a ARM92[02], otherwise
536	 * [ls]dm alignment faults are noisy!
537	 */
538#if !(defined CONFIG_CPU_ARM922T)  && !(defined CONFIG_CPU_ARM920T)
539	/*
540	 * This is a "hint" - we already have eaddr worked out by the
541	 * processor for us.
542	 */
543	if (addr != eaddr) {
544		pr_err("LDMSTM: PC = %08lx, instr = %08lx, "
545			"addr = %08lx, eaddr = %08lx\n",
546			 instruction_pointer(regs), instr, addr, eaddr);
547		show_regs(regs);
548	}
549#endif
550
551	if (user_mode(regs)) {
552		unsigned int __ua_flags = uaccess_save_and_enable();
553		for (regbits = REGMASK_BITS(instr), rd = 0; regbits;
554		     regbits >>= 1, rd += 1)
555			if (regbits & 1) {
556				if (LDST_L_BIT(instr)) {
557					unsigned int val;
558					get32t_unaligned_check(val, eaddr);
559					regs->uregs[rd] = val;
560				} else
561					put32t_unaligned_check(regs->uregs[rd], eaddr);
562				eaddr += 4;
563			}
564		uaccess_restore(__ua_flags);
565	} else {
566		for (regbits = REGMASK_BITS(instr), rd = 0; regbits;
567		     regbits >>= 1, rd += 1)
568			if (regbits & 1) {
569				if (LDST_L_BIT(instr)) {
570					unsigned int val;
571					get32_unaligned_check(val, eaddr);
572					regs->uregs[rd] = val;
573				} else
574					put32_unaligned_check(regs->uregs[rd], eaddr);
575				eaddr += 4;
576			}
577	}
578
579	if (LDST_W_BIT(instr))
580		regs->uregs[rn] = newaddr;
581	if (!LDST_L_BIT(instr) || !(REGMASK_BITS(instr) & (1 << 15)))
582		regs->ARM_pc -= correction;
583	return TYPE_DONE;
584
585fault:
586	regs->ARM_pc -= correction;
587	return TYPE_FAULT;
588
589bad:
590	pr_err("Alignment trap: not handling ldm with s-bit set\n");
591	return TYPE_ERROR;
592}
593
594/*
595 * Convert Thumb ld/st instruction forms to equivalent ARM instructions so
596 * we can reuse ARM userland alignment fault fixups for Thumb.
597 *
598 * This implementation was initially based on the algorithm found in
599 * gdb/sim/arm/thumbemu.c. It is basically just a code reduction of same
600 * to convert only Thumb ld/st instruction forms to equivalent ARM forms.
601 *
602 * NOTES:
603 * 1. Comments below refer to ARM ARM DDI0100E Thumb Instruction sections.
604 * 2. If for some reason we're passed an non-ld/st Thumb instruction to
605 *    decode, we return 0xdeadc0de. This should never happen under normal
606 *    circumstances but if it does, we've got other problems to deal with
607 *    elsewhere and we obviously can't fix those problems here.
608 */
609
610static unsigned long
611thumb2arm(u16 tinstr)
612{
613	u32 L = (tinstr & (1<<11)) >> 11;
614
615	switch ((tinstr & 0xf800) >> 11) {
616	/* 6.5.1 Format 1: */
617	case 0x6000 >> 11:				/* 7.1.52 STR(1) */
618	case 0x6800 >> 11:				/* 7.1.26 LDR(1) */
619	case 0x7000 >> 11:				/* 7.1.55 STRB(1) */
620	case 0x7800 >> 11:				/* 7.1.30 LDRB(1) */
621		return 0xe5800000 |
622			((tinstr & (1<<12)) << (22-12)) |	/* fixup */
623			(L<<20) |				/* L==1? */
624			((tinstr & (7<<0)) << (12-0)) |		/* Rd */
625			((tinstr & (7<<3)) << (16-3)) |		/* Rn */
626			((tinstr & (31<<6)) >>			/* immed_5 */
627				(6 - ((tinstr & (1<<12)) ? 0 : 2)));
628	case 0x8000 >> 11:				/* 7.1.57 STRH(1) */
629	case 0x8800 >> 11:				/* 7.1.32 LDRH(1) */
630		return 0xe1c000b0 |
631			(L<<20) |				/* L==1? */
632			((tinstr & (7<<0)) << (12-0)) |		/* Rd */
633			((tinstr & (7<<3)) << (16-3)) |		/* Rn */
634			((tinstr & (7<<6)) >> (6-1)) |	 /* immed_5[2:0] */
635			((tinstr & (3<<9)) >> (9-8));	 /* immed_5[4:3] */
636
637	/* 6.5.1 Format 2: */
638	case 0x5000 >> 11:
639	case 0x5800 >> 11:
640		{
641			static const u32 subset[8] = {
642				0xe7800000,		/* 7.1.53 STR(2) */
643				0xe18000b0,		/* 7.1.58 STRH(2) */
644				0xe7c00000,		/* 7.1.56 STRB(2) */
645				0xe19000d0,		/* 7.1.34 LDRSB */
646				0xe7900000,		/* 7.1.27 LDR(2) */
647				0xe19000b0,		/* 7.1.33 LDRH(2) */
648				0xe7d00000,		/* 7.1.31 LDRB(2) */
649				0xe19000f0		/* 7.1.35 LDRSH */
650			};
651			return subset[(tinstr & (7<<9)) >> 9] |
652			    ((tinstr & (7<<0)) << (12-0)) |	/* Rd */
653			    ((tinstr & (7<<3)) << (16-3)) |	/* Rn */
654			    ((tinstr & (7<<6)) >> (6-0));	/* Rm */
655		}
656
657	/* 6.5.1 Format 3: */
658	case 0x4800 >> 11:				/* 7.1.28 LDR(3) */
659		/* NOTE: This case is not technically possible. We're
660		 *	 loading 32-bit memory data via PC relative
661		 *	 addressing mode. So we can and should eliminate
662		 *	 this case. But I'll leave it here for now.
663		 */
664		return 0xe59f0000 |
665		    ((tinstr & (7<<8)) << (12-8)) |		/* Rd */
666		    ((tinstr & 255) << (2-0));			/* immed_8 */
667
668	/* 6.5.1 Format 4: */
669	case 0x9000 >> 11:				/* 7.1.54 STR(3) */
670	case 0x9800 >> 11:				/* 7.1.29 LDR(4) */
671		return 0xe58d0000 |
672			(L<<20) |				/* L==1? */
673			((tinstr & (7<<8)) << (12-8)) |		/* Rd */
674			((tinstr & 255) << 2);			/* immed_8 */
675
676	/* 6.6.1 Format 1: */
677	case 0xc000 >> 11:				/* 7.1.51 STMIA */
678	case 0xc800 >> 11:				/* 7.1.25 LDMIA */
679		{
680			u32 Rn = (tinstr & (7<<8)) >> 8;
681			u32 W = ((L<<Rn) & (tinstr&255)) ? 0 : 1<<21;
682
683			return 0xe8800000 | W | (L<<20) | (Rn<<16) |
684				(tinstr&255);
685		}
686
687	/* 6.6.1 Format 2: */
688	case 0xb000 >> 11:				/* 7.1.48 PUSH */
689	case 0xb800 >> 11:				/* 7.1.47 POP */
690		if ((tinstr & (3 << 9)) == 0x0400) {
691			static const u32 subset[4] = {
692				0xe92d0000,	/* STMDB sp!,{registers} */
693				0xe92d4000,	/* STMDB sp!,{registers,lr} */
694				0xe8bd0000,	/* LDMIA sp!,{registers} */
695				0xe8bd8000	/* LDMIA sp!,{registers,pc} */
696			};
697			return subset[(L<<1) | ((tinstr & (1<<8)) >> 8)] |
698			    (tinstr & 255);		/* register_list */
699		}
700		/* Else fall through for illegal instruction case */
701
702	default:
703		return BAD_INSTR;
704	}
705}
706
707/*
708 * Convert Thumb-2 32 bit LDM, STM, LDRD, STRD to equivalent instruction
709 * handlable by ARM alignment handler, also find the corresponding handler,
710 * so that we can reuse ARM userland alignment fault fixups for Thumb.
711 *
712 * @pinstr: original Thumb-2 instruction; returns new handlable instruction
713 * @regs: register context.
714 * @poffset: return offset from faulted addr for later writeback
715 *
716 * NOTES:
717 * 1. Comments below refer to ARMv7 DDI0406A Thumb Instruction sections.
718 * 2. Register name Rt from ARMv7 is same as Rd from ARMv6 (Rd is Rt)
719 */
720static void *
721do_alignment_t32_to_handler(unsigned long *pinstr, struct pt_regs *regs,
722			    union offset_union *poffset)
723{
724	unsigned long instr = *pinstr;
725	u16 tinst1 = (instr >> 16) & 0xffff;
726	u16 tinst2 = instr & 0xffff;
727
728	switch (tinst1 & 0xffe0) {
729	/* A6.3.5 Load/Store multiple */
730	case 0xe880:		/* STM/STMIA/STMEA,LDM/LDMIA, PUSH/POP T2 */
731	case 0xe8a0:		/* ...above writeback version */
732	case 0xe900:		/* STMDB/STMFD, LDMDB/LDMEA */
733	case 0xe920:		/* ...above writeback version */
734		/* no need offset decision since handler calculates it */
735		return do_alignment_ldmstm;
736
737	case 0xf840:		/* POP/PUSH T3 (single register) */
738		if (RN_BITS(instr) == 13 && (tinst2 & 0x09ff) == 0x0904) {
739			u32 L = !!(LDST_L_BIT(instr));
740			const u32 subset[2] = {
741				0xe92d0000,	/* STMDB sp!,{registers} */
742				0xe8bd0000,	/* LDMIA sp!,{registers} */
743			};
744			*pinstr = subset[L] | (1<<RD_BITS(instr));
745			return do_alignment_ldmstm;
746		}
747		/* Else fall through for illegal instruction case */
748		break;
749
750	/* A6.3.6 Load/store double, STRD/LDRD(immed, lit, reg) */
751	case 0xe860:
752	case 0xe960:
753	case 0xe8e0:
754	case 0xe9e0:
755		poffset->un = (tinst2 & 0xff) << 2;
756	case 0xe940:
757	case 0xe9c0:
758		return do_alignment_ldrdstrd;
759
760	/*
761	 * No need to handle load/store instructions up to word size
762	 * since ARMv6 and later CPUs can perform unaligned accesses.
763	 */
764	default:
765		break;
766	}
767	return NULL;
768}
769
770static int
771do_alignment(unsigned long addr, unsigned int fsr, struct pt_regs *regs)
772{
773	union offset_union uninitialized_var(offset);
774	unsigned long instr = 0, instrptr;
775	int (*handler)(unsigned long addr, unsigned long instr, struct pt_regs *regs);
776	unsigned int type;
777	unsigned int fault;
778	u16 tinstr = 0;
779	int isize = 4;
780	int thumb2_32b = 0;
781
782	if (interrupts_enabled(regs))
783		local_irq_enable();
784
785	instrptr = instruction_pointer(regs);
786
787	if (thumb_mode(regs)) {
788		u16 *ptr = (u16 *)(instrptr & ~1);
789		fault = probe_kernel_address(ptr, tinstr);
790		tinstr = __mem_to_opcode_thumb16(tinstr);
791		if (!fault) {
792			if (cpu_architecture() >= CPU_ARCH_ARMv7 &&
793			    IS_T32(tinstr)) {
794				/* Thumb-2 32-bit */
795				u16 tinst2 = 0;
796				fault = probe_kernel_address(ptr + 1, tinst2);
797				tinst2 = __mem_to_opcode_thumb16(tinst2);
798				instr = __opcode_thumb32_compose(tinstr, tinst2);
799				thumb2_32b = 1;
800			} else {
801				isize = 2;
802				instr = thumb2arm(tinstr);
803			}
804		}
805	} else {
806		fault = probe_kernel_address((void *)instrptr, instr);
807		instr = __mem_to_opcode_arm(instr);
808	}
809
810	if (fault) {
811		type = TYPE_FAULT;
812		goto bad_or_fault;
813	}
814
815	if (user_mode(regs))
816		goto user;
817
818	ai_sys += 1;
819	ai_sys_last_pc = (void *)instruction_pointer(regs);
820
821 fixup:
822
823	regs->ARM_pc += isize;
824
825	switch (CODING_BITS(instr)) {
826	case 0x00000000:	/* 3.13.4 load/store instruction extensions */
827		if (LDSTHD_I_BIT(instr))
828			offset.un = (instr & 0xf00) >> 4 | (instr & 15);
829		else
830			offset.un = regs->uregs[RM_BITS(instr)];
831
832		if ((instr & 0x000000f0) == 0x000000b0 || /* LDRH, STRH */
833		    (instr & 0x001000f0) == 0x001000f0)   /* LDRSH */
834			handler = do_alignment_ldrhstrh;
835		else if ((instr & 0x001000f0) == 0x000000d0 || /* LDRD */
836			 (instr & 0x001000f0) == 0x000000f0)   /* STRD */
837			handler = do_alignment_ldrdstrd;
838		else if ((instr & 0x01f00ff0) == 0x01000090) /* SWP */
839			goto swp;
840		else
841			goto bad;
842		break;
843
844	case 0x04000000:	/* ldr or str immediate */
845		if (COND_BITS(instr) == 0xf0000000) /* NEON VLDn, VSTn */
846			goto bad;
847		offset.un = OFFSET_BITS(instr);
848		handler = do_alignment_ldrstr;
849		break;
850
851	case 0x06000000:	/* ldr or str register */
852		offset.un = regs->uregs[RM_BITS(instr)];
853
854		if (IS_SHIFT(instr)) {
855			unsigned int shiftval = SHIFT_BITS(instr);
856
857			switch(SHIFT_TYPE(instr)) {
858			case SHIFT_LSL:
859				offset.un <<= shiftval;
860				break;
861
862			case SHIFT_LSR:
863				offset.un >>= shiftval;
864				break;
865
866			case SHIFT_ASR:
867				offset.sn >>= shiftval;
868				break;
869
870			case SHIFT_RORRRX:
871				if (shiftval == 0) {
872					offset.un >>= 1;
873					if (regs->ARM_cpsr & PSR_C_BIT)
874						offset.un |= 1 << 31;
875				} else
876					offset.un = offset.un >> shiftval |
877							  offset.un << (32 - shiftval);
878				break;
879			}
880		}
881		handler = do_alignment_ldrstr;
882		break;
883
884	case 0x08000000:	/* ldm or stm, or thumb-2 32bit instruction */
885		if (thumb2_32b) {
886			offset.un = 0;
887			handler = do_alignment_t32_to_handler(&instr, regs, &offset);
888		} else {
889			offset.un = 0;
890			handler = do_alignment_ldmstm;
891		}
892		break;
893
894	default:
895		goto bad;
896	}
897
898	if (!handler)
899		goto bad;
900	type = handler(addr, instr, regs);
901
902	if (type == TYPE_ERROR || type == TYPE_FAULT) {
903		regs->ARM_pc -= isize;
904		goto bad_or_fault;
905	}
906
907	if (type == TYPE_LDST)
908		do_alignment_finish_ldst(addr, instr, regs, offset);
909
910	return 0;
911
912 bad_or_fault:
913	if (type == TYPE_ERROR)
914		goto bad;
915	/*
916	 * We got a fault - fix it up, or die.
917	 */
918	do_bad_area(addr, fsr, regs);
919	return 0;
920
921 swp:
922	pr_err("Alignment trap: not handling swp instruction\n");
923
924 bad:
925	/*
926	 * Oops, we didn't handle the instruction.
927	 */
928	pr_err("Alignment trap: not handling instruction "
929		"%0*lx at [<%08lx>]\n",
930		isize << 1,
931		isize == 2 ? tinstr : instr, instrptr);
932	ai_skipped += 1;
933	return 1;
934
935 user:
936	ai_user += 1;
937
938	if (ai_usermode & UM_WARN)
939		printk("Alignment trap: %s (%d) PC=0x%08lx Instr=0x%0*lx "
940		       "Address=0x%08lx FSR 0x%03x\n", current->comm,
941			task_pid_nr(current), instrptr,
942			isize << 1,
943			isize == 2 ? tinstr : instr,
944		        addr, fsr);
945
946	if (ai_usermode & UM_FIXUP)
947		goto fixup;
948
949	if (ai_usermode & UM_SIGNAL) {
950		siginfo_t si;
951
952		si.si_signo = SIGBUS;
953		si.si_errno = 0;
954		si.si_code = BUS_ADRALN;
955		si.si_addr = (void __user *)addr;
956
957		force_sig_info(si.si_signo, &si, current);
958	} else {
959		/*
960		 * We're about to disable the alignment trap and return to
961		 * user space.  But if an interrupt occurs before actually
962		 * reaching user space, then the IRQ vector entry code will
963		 * notice that we were still in kernel space and therefore
964		 * the alignment trap won't be re-enabled in that case as it
965		 * is presumed to be always on from kernel space.
966		 * Let's prevent that race by disabling interrupts here (they
967		 * are disabled on the way back to user space anyway in
968		 * entry-common.S) and disable the alignment trap only if
969		 * there is no work pending for this thread.
970		 */
971		raw_local_irq_disable();
972		if (!(current_thread_info()->flags & _TIF_WORK_MASK))
973			set_cr(cr_no_alignment);
974	}
975
976	return 0;
977}
978
979static int __init noalign_setup(char *__unused)
980{
981	set_cr(__clear_cr(CR_A));
982	return 1;
983}
984__setup("noalign", noalign_setup);
985
986/*
987 * This needs to be done after sysctl_init, otherwise sys/ will be
988 * overwritten.  Actually, this shouldn't be in sys/ at all since
989 * it isn't a sysctl, and it doesn't contain sysctl information.
990 * We now locate it in /proc/cpu/alignment instead.
991 */
992static int __init alignment_init(void)
993{
994#ifdef CONFIG_PROC_FS
995	struct proc_dir_entry *res;
996
997	res = proc_create("cpu/alignment", S_IWUSR | S_IRUGO, NULL,
998			  &alignment_proc_fops);
999	if (!res)
1000		return -ENOMEM;
1001#endif
1002
1003	if (cpu_is_v6_unaligned()) {
1004		set_cr(__clear_cr(CR_A));
1005		ai_usermode = safe_usermode(ai_usermode, false);
1006	}
1007
1008	cr_no_alignment = get_cr() & ~CR_A;
1009
1010	hook_fault_code(FAULT_CODE_ALIGNMENT, do_alignment, SIGBUS, BUS_ADRALN,
1011			"alignment exception");
1012
1013	/*
1014	 * ARMv6K and ARMv7 use fault status 3 (0b00011) as Access Flag section
1015	 * fault, not as alignment error.
1016	 *
1017	 * TODO: handle ARMv6K properly. Runtime check for 'K' extension is
1018	 * needed.
1019	 */
1020	if (cpu_architecture() <= CPU_ARCH_ARMv6) {
1021		hook_fault_code(3, do_alignment, SIGBUS, BUS_ADRALN,
1022				"alignment exception");
1023	}
1024
1025	return 0;
1026}
1027
1028fs_initcall(alignment_init);
1029