1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 *  linux/arch/arm/mm/alignment.c
4 *
5 *  Copyright (C) 1995  Linus Torvalds
6 *  Modifications for ARM processor (c) 1995-2001 Russell King
7 *  Thumb alignment fault fixups (c) 2004 MontaVista Software, Inc.
8 *  - Adapted from gdb/sim/arm/thumbemu.c -- Thumb instruction emulation.
9 *    Copyright (C) 1996, Cygnus Software Technologies Ltd.
10 */
11#include <linux/moduleparam.h>
12#include <linux/compiler.h>
13#include <linux/kernel.h>
14#include <linux/sched/debug.h>
15#include <linux/errno.h>
16#include <linux/string.h>
17#include <linux/proc_fs.h>
18#include <linux/seq_file.h>
19#include <linux/init.h>
20#include <linux/sched/signal.h>
21#include <linux/uaccess.h>
22
23#include <asm/cp15.h>
24#include <asm/system_info.h>
25#include <asm/unaligned.h>
26#include <asm/opcodes.h>
27
28#include "fault.h"
29#include "mm.h"
30
31/*
32 * 32-bit misaligned trap handler (c) 1998 San Mehat (CCC) -July 1998
33 * /proc/sys/debug/alignment, modified and integrated into
34 * Linux 2.1 by Russell King
35 *
36 * Speed optimisations and better fault handling by Russell King.
37 *
38 * *** NOTE ***
39 * This code is not portable to processors with late data abort handling.
40 */
41#define CODING_BITS(i)	(i & 0x0e000000)
42#define COND_BITS(i)	(i & 0xf0000000)
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
76static unsigned long ai_user;
77static unsigned long ai_sys;
78static void *ai_sys_last_pc;
79static unsigned long ai_skipped;
80static unsigned long ai_half;
81static unsigned long ai_word;
82static unsigned long ai_dword;
83static unsigned long ai_multi;
84static int ai_usermode;
85static unsigned long cr_no_alignment;
86
87core_param(alignment, ai_usermode, int, 0600);
88
89#define UM_WARN		(1 << 0)
90#define UM_FIXUP	(1 << 1)
91#define UM_SIGNAL	(1 << 2)
92
93/* Return true if and only if the ARMv6 unaligned access model is in use. */
94static bool cpu_is_v6_unaligned(void)
95{
96	return cpu_architecture() >= CPU_ARCH_ARMv6 && get_cr() & CR_U;
97}
98
99static int safe_usermode(int new_usermode, bool warn)
100{
101	/*
102	 * ARMv6 and later CPUs can perform unaligned accesses for
103	 * most single load and store instructions up to word size.
104	 * LDM, STM, LDRD and STRD still need to be handled.
105	 *
106	 * Ignoring the alignment fault is not an option on these
107	 * CPUs since we spin re-faulting the instruction without
108	 * making any progress.
109	 */
110	if (cpu_is_v6_unaligned() && !(new_usermode & (UM_FIXUP | UM_SIGNAL))) {
111		new_usermode |= UM_FIXUP;
112
113		if (warn)
114			pr_warn("alignment: ignoring faults is unsafe on this CPU.  Defaulting to fixup mode.\n");
115	}
116
117	return new_usermode;
118}
119
120#ifdef CONFIG_PROC_FS
121static const char *usermode_action[] = {
122	"ignored",
123	"warn",
124	"fixup",
125	"fixup+warn",
126	"signal",
127	"signal+warn"
128};
129
130static int alignment_proc_show(struct seq_file *m, void *v)
131{
132	seq_printf(m, "User:\t\t%lu\n", ai_user);
133	seq_printf(m, "System:\t\t%lu (%pS)\n", ai_sys, ai_sys_last_pc);
134	seq_printf(m, "Skipped:\t%lu\n", ai_skipped);
135	seq_printf(m, "Half:\t\t%lu\n", ai_half);
136	seq_printf(m, "Word:\t\t%lu\n", ai_word);
137	if (cpu_architecture() >= CPU_ARCH_ARMv5TE)
138		seq_printf(m, "DWord:\t\t%lu\n", ai_dword);
139	seq_printf(m, "Multi:\t\t%lu\n", ai_multi);
140	seq_printf(m, "User faults:\t%i (%s)\n", ai_usermode,
141			usermode_action[ai_usermode]);
142
143	return 0;
144}
145
146static int alignment_proc_open(struct inode *inode, struct file *file)
147{
148	return single_open(file, alignment_proc_show, NULL);
149}
150
151static ssize_t alignment_proc_write(struct file *file, const char __user *buffer,
152				    size_t count, loff_t *pos)
153{
154	char mode;
155
156	if (count > 0) {
157		if (get_user(mode, buffer))
158			return -EFAULT;
159		if (mode >= '0' && mode <= '5')
160			ai_usermode = safe_usermode(mode - '0', true);
161	}
162	return count;
163}
164
165static const struct proc_ops alignment_proc_ops = {
166	.proc_open	= alignment_proc_open,
167	.proc_read	= seq_read,
168	.proc_lseek	= seq_lseek,
169	.proc_release	= single_release,
170	.proc_write	= alignment_proc_write,
171};
172#endif /* CONFIG_PROC_FS */
173
174union offset_union {
175	unsigned long un;
176	  signed long sn;
177};
178
179#define TYPE_ERROR	0
180#define TYPE_FAULT	1
181#define TYPE_LDST	2
182#define TYPE_DONE	3
183
184#ifdef __ARMEB__
185#define BE		1
186#define FIRST_BYTE_16	"mov	%1, %1, ror #8\n"
187#define FIRST_BYTE_32	"mov	%1, %1, ror #24\n"
188#define NEXT_BYTE	"ror #24"
189#else
190#define BE		0
191#define FIRST_BYTE_16
192#define FIRST_BYTE_32
193#define NEXT_BYTE	"lsr #8"
194#endif
195
196#define __get8_unaligned_check(ins,val,addr,err)	\
197	__asm__(					\
198 ARM(	"1:	"ins"	%1, [%2], #1\n"	)		\
199 THUMB(	"1:	"ins"	%1, [%2]\n"	)		\
200 THUMB(	"	add	%2, %2, #1\n"	)		\
201	"2:\n"						\
202	"	.pushsection .text.fixup,\"ax\"\n"	\
203	"	.align	2\n"				\
204	"3:	mov	%0, #1\n"			\
205	"	b	2b\n"				\
206	"	.popsection\n"				\
207	"	.pushsection __ex_table,\"a\"\n"	\
208	"	.align	3\n"				\
209	"	.long	1b, 3b\n"			\
210	"	.popsection\n"				\
211	: "=r" (err), "=&r" (val), "=r" (addr)		\
212	: "0" (err), "2" (addr))
213
214#define __get16_unaligned_check(ins,val,addr)			\
215	do {							\
216		unsigned int err = 0, v, a = addr;		\
217		__get8_unaligned_check(ins,v,a,err);		\
218		val =  v << ((BE) ? 8 : 0);			\
219		__get8_unaligned_check(ins,v,a,err);		\
220		val |= v << ((BE) ? 0 : 8);			\
221		if (err)					\
222			goto fault;				\
223	} while (0)
224
225#define get16_unaligned_check(val,addr) \
226	__get16_unaligned_check("ldrb",val,addr)
227
228#define get16t_unaligned_check(val,addr) \
229	__get16_unaligned_check("ldrbt",val,addr)
230
231#define __get32_unaligned_check(ins,val,addr)			\
232	do {							\
233		unsigned int err = 0, v, a = addr;		\
234		__get8_unaligned_check(ins,v,a,err);		\
235		val =  v << ((BE) ? 24 :  0);			\
236		__get8_unaligned_check(ins,v,a,err);		\
237		val |= v << ((BE) ? 16 :  8);			\
238		__get8_unaligned_check(ins,v,a,err);		\
239		val |= v << ((BE) ?  8 : 16);			\
240		__get8_unaligned_check(ins,v,a,err);		\
241		val |= v << ((BE) ?  0 : 24);			\
242		if (err)					\
243			goto fault;				\
244	} while (0)
245
246#define get32_unaligned_check(val,addr) \
247	__get32_unaligned_check("ldrb",val,addr)
248
249#define get32t_unaligned_check(val,addr) \
250	__get32_unaligned_check("ldrbt",val,addr)
251
252#define __put16_unaligned_check(ins,val,addr)			\
253	do {							\
254		unsigned int err = 0, v = val, a = addr;	\
255		__asm__( FIRST_BYTE_16				\
256	 ARM(	"1:	"ins"	%1, [%2], #1\n"	)		\
257	 THUMB(	"1:	"ins"	%1, [%2]\n"	)		\
258	 THUMB(	"	add	%2, %2, #1\n"	)		\
259		"	mov	%1, %1, "NEXT_BYTE"\n"		\
260		"2:	"ins"	%1, [%2]\n"			\
261		"3:\n"						\
262		"	.pushsection .text.fixup,\"ax\"\n"	\
263		"	.align	2\n"				\
264		"4:	mov	%0, #1\n"			\
265		"	b	3b\n"				\
266		"	.popsection\n"				\
267		"	.pushsection __ex_table,\"a\"\n"	\
268		"	.align	3\n"				\
269		"	.long	1b, 4b\n"			\
270		"	.long	2b, 4b\n"			\
271		"	.popsection\n"				\
272		: "=r" (err), "=&r" (v), "=&r" (a)		\
273		: "0" (err), "1" (v), "2" (a));			\
274		if (err)					\
275			goto fault;				\
276	} while (0)
277
278#define put16_unaligned_check(val,addr)  \
279	__put16_unaligned_check("strb",val,addr)
280
281#define put16t_unaligned_check(val,addr) \
282	__put16_unaligned_check("strbt",val,addr)
283
284#define __put32_unaligned_check(ins,val,addr)			\
285	do {							\
286		unsigned int err = 0, v = val, a = addr;	\
287		__asm__( FIRST_BYTE_32				\
288	 ARM(	"1:	"ins"	%1, [%2], #1\n"	)		\
289	 THUMB(	"1:	"ins"	%1, [%2]\n"	)		\
290	 THUMB(	"	add	%2, %2, #1\n"	)		\
291		"	mov	%1, %1, "NEXT_BYTE"\n"		\
292	 ARM(	"2:	"ins"	%1, [%2], #1\n"	)		\
293	 THUMB(	"2:	"ins"	%1, [%2]\n"	)		\
294	 THUMB(	"	add	%2, %2, #1\n"	)		\
295		"	mov	%1, %1, "NEXT_BYTE"\n"		\
296	 ARM(	"3:	"ins"	%1, [%2], #1\n"	)		\
297	 THUMB(	"3:	"ins"	%1, [%2]\n"	)		\
298	 THUMB(	"	add	%2, %2, #1\n"	)		\
299		"	mov	%1, %1, "NEXT_BYTE"\n"		\
300		"4:	"ins"	%1, [%2]\n"			\
301		"5:\n"						\
302		"	.pushsection .text.fixup,\"ax\"\n"	\
303		"	.align	2\n"				\
304		"6:	mov	%0, #1\n"			\
305		"	b	5b\n"				\
306		"	.popsection\n"				\
307		"	.pushsection __ex_table,\"a\"\n"	\
308		"	.align	3\n"				\
309		"	.long	1b, 6b\n"			\
310		"	.long	2b, 6b\n"			\
311		"	.long	3b, 6b\n"			\
312		"	.long	4b, 6b\n"			\
313		"	.popsection\n"				\
314		: "=r" (err), "=&r" (v), "=&r" (a)		\
315		: "0" (err), "1" (v), "2" (a));			\
316		if (err)					\
317			goto fault;				\
318	} while (0)
319
320#define put32_unaligned_check(val,addr) \
321	__put32_unaligned_check("strb", val, addr)
322
323#define put32t_unaligned_check(val,addr) \
324	__put32_unaligned_check("strbt", val, addr)
325
326static void
327do_alignment_finish_ldst(unsigned long addr, u32 instr, struct pt_regs *regs, union offset_union offset)
328{
329	if (!LDST_U_BIT(instr))
330		offset.un = -offset.un;
331
332	if (!LDST_P_BIT(instr))
333		addr += offset.un;
334
335	if (!LDST_P_BIT(instr) || LDST_W_BIT(instr))
336		regs->uregs[RN_BITS(instr)] = addr;
337}
338
339static int
340do_alignment_ldrhstrh(unsigned long addr, u32 instr, struct pt_regs *regs)
341{
342	unsigned int rd = RD_BITS(instr);
343
344	ai_half += 1;
345
346	if (user_mode(regs))
347		goto user;
348
349	if (LDST_L_BIT(instr)) {
350		unsigned long val;
351		get16_unaligned_check(val, addr);
352
353		/* signed half-word? */
354		if (instr & 0x40)
355			val = (signed long)((signed short) val);
356
357		regs->uregs[rd] = val;
358	} else
359		put16_unaligned_check(regs->uregs[rd], addr);
360
361	return TYPE_LDST;
362
363 user:
364	if (LDST_L_BIT(instr)) {
365		unsigned long val;
366		unsigned int __ua_flags = uaccess_save_and_enable();
367
368		get16t_unaligned_check(val, addr);
369		uaccess_restore(__ua_flags);
370
371		/* signed half-word? */
372		if (instr & 0x40)
373			val = (signed long)((signed short) val);
374
375		regs->uregs[rd] = val;
376	} else {
377		unsigned int __ua_flags = uaccess_save_and_enable();
378		put16t_unaligned_check(regs->uregs[rd], addr);
379		uaccess_restore(__ua_flags);
380	}
381
382	return TYPE_LDST;
383
384 fault:
385	return TYPE_FAULT;
386}
387
388static int
389do_alignment_ldrdstrd(unsigned long addr, u32 instr, struct pt_regs *regs)
390{
391	unsigned int rd = RD_BITS(instr);
392	unsigned int rd2;
393	int load;
394
395	if ((instr & 0xfe000000) == 0xe8000000) {
396		/* ARMv7 Thumb-2 32-bit LDRD/STRD */
397		rd2 = (instr >> 8) & 0xf;
398		load = !!(LDST_L_BIT(instr));
399	} else if (((rd & 1) == 1) || (rd == 14))
400		goto bad;
401	else {
402		load = ((instr & 0xf0) == 0xd0);
403		rd2 = rd + 1;
404	}
405
406	ai_dword += 1;
407
408	if (user_mode(regs))
409		goto user;
410
411	if (load) {
412		unsigned long val;
413		get32_unaligned_check(val, addr);
414		regs->uregs[rd] = val;
415		get32_unaligned_check(val, addr + 4);
416		regs->uregs[rd2] = val;
417	} else {
418		put32_unaligned_check(regs->uregs[rd], addr);
419		put32_unaligned_check(regs->uregs[rd2], addr + 4);
420	}
421
422	return TYPE_LDST;
423
424 user:
425	if (load) {
426		unsigned long val, val2;
427		unsigned int __ua_flags = uaccess_save_and_enable();
428
429		get32t_unaligned_check(val, addr);
430		get32t_unaligned_check(val2, addr + 4);
431
432		uaccess_restore(__ua_flags);
433
434		regs->uregs[rd] = val;
435		regs->uregs[rd2] = val2;
436	} else {
437		unsigned int __ua_flags = uaccess_save_and_enable();
438		put32t_unaligned_check(regs->uregs[rd], addr);
439		put32t_unaligned_check(regs->uregs[rd2], addr + 4);
440		uaccess_restore(__ua_flags);
441	}
442
443	return TYPE_LDST;
444 bad:
445	return TYPE_ERROR;
446 fault:
447	return TYPE_FAULT;
448}
449
450static int
451do_alignment_ldrstr(unsigned long addr, u32 instr, struct pt_regs *regs)
452{
453	unsigned int rd = RD_BITS(instr);
454
455	ai_word += 1;
456
457	if ((!LDST_P_BIT(instr) && LDST_W_BIT(instr)) || user_mode(regs))
458		goto trans;
459
460	if (LDST_L_BIT(instr)) {
461		unsigned int val;
462		get32_unaligned_check(val, addr);
463		regs->uregs[rd] = val;
464	} else
465		put32_unaligned_check(regs->uregs[rd], addr);
466	return TYPE_LDST;
467
468 trans:
469	if (LDST_L_BIT(instr)) {
470		unsigned int val;
471		unsigned int __ua_flags = uaccess_save_and_enable();
472		get32t_unaligned_check(val, addr);
473		uaccess_restore(__ua_flags);
474		regs->uregs[rd] = val;
475	} else {
476		unsigned int __ua_flags = uaccess_save_and_enable();
477		put32t_unaligned_check(regs->uregs[rd], addr);
478		uaccess_restore(__ua_flags);
479	}
480	return TYPE_LDST;
481
482 fault:
483	return TYPE_FAULT;
484}
485
486/*
487 * LDM/STM alignment handler.
488 *
489 * There are 4 variants of this instruction:
490 *
491 * B = rn pointer before instruction, A = rn pointer after instruction
492 *              ------ increasing address ----->
493 *	        |    | r0 | r1 | ... | rx |    |
494 * PU = 01             B                    A
495 * PU = 11        B                    A
496 * PU = 00        A                    B
497 * PU = 10             A                    B
498 */
499static int
500do_alignment_ldmstm(unsigned long addr, u32 instr, struct pt_regs *regs)
501{
502	unsigned int rd, rn, correction, nr_regs, regbits;
503	unsigned long eaddr, newaddr;
504
505	if (LDM_S_BIT(instr))
506		goto bad;
507
508	correction = 4; /* processor implementation defined */
509	regs->ARM_pc += correction;
510
511	ai_multi += 1;
512
513	/* count the number of registers in the mask to be transferred */
514	nr_regs = hweight16(REGMASK_BITS(instr)) * 4;
515
516	rn = RN_BITS(instr);
517	newaddr = eaddr = regs->uregs[rn];
518
519	if (!LDST_U_BIT(instr))
520		nr_regs = -nr_regs;
521	newaddr += nr_regs;
522	if (!LDST_U_BIT(instr))
523		eaddr = newaddr;
524
525	if (LDST_P_EQ_U(instr))	/* U = P */
526		eaddr += 4;
527
528	/*
529	 * For alignment faults on the ARM922T/ARM920T the MMU  makes
530	 * the FSR (and hence addr) equal to the updated base address
531	 * of the multiple access rather than the restored value.
532	 * Switch this message off if we've got a ARM92[02], otherwise
533	 * [ls]dm alignment faults are noisy!
534	 */
535#if !(defined CONFIG_CPU_ARM922T)  && !(defined CONFIG_CPU_ARM920T)
536	/*
537	 * This is a "hint" - we already have eaddr worked out by the
538	 * processor for us.
539	 */
540	if (addr != eaddr) {
541		pr_err("LDMSTM: PC = %08lx, instr = %08x, "
542			"addr = %08lx, eaddr = %08lx\n",
543			 instruction_pointer(regs), instr, addr, eaddr);
544		show_regs(regs);
545	}
546#endif
547
548	if (user_mode(regs)) {
549		unsigned int __ua_flags = uaccess_save_and_enable();
550		for (regbits = REGMASK_BITS(instr), rd = 0; regbits;
551		     regbits >>= 1, rd += 1)
552			if (regbits & 1) {
553				if (LDST_L_BIT(instr)) {
554					unsigned int val;
555					get32t_unaligned_check(val, eaddr);
556					regs->uregs[rd] = val;
557				} else
558					put32t_unaligned_check(regs->uregs[rd], eaddr);
559				eaddr += 4;
560			}
561		uaccess_restore(__ua_flags);
562	} else {
563		for (regbits = REGMASK_BITS(instr), rd = 0; regbits;
564		     regbits >>= 1, rd += 1)
565			if (regbits & 1) {
566				if (LDST_L_BIT(instr)) {
567					unsigned int val;
568					get32_unaligned_check(val, eaddr);
569					regs->uregs[rd] = val;
570				} else
571					put32_unaligned_check(regs->uregs[rd], eaddr);
572				eaddr += 4;
573			}
574	}
575
576	if (LDST_W_BIT(instr))
577		regs->uregs[rn] = newaddr;
578	if (!LDST_L_BIT(instr) || !(REGMASK_BITS(instr) & (1 << 15)))
579		regs->ARM_pc -= correction;
580	return TYPE_DONE;
581
582fault:
583	regs->ARM_pc -= correction;
584	return TYPE_FAULT;
585
586bad:
587	pr_err("Alignment trap: not handling ldm with s-bit set\n");
588	return TYPE_ERROR;
589}
590
591/*
592 * Convert Thumb ld/st instruction forms to equivalent ARM instructions so
593 * we can reuse ARM userland alignment fault fixups for Thumb.
594 *
595 * This implementation was initially based on the algorithm found in
596 * gdb/sim/arm/thumbemu.c. It is basically just a code reduction of same
597 * to convert only Thumb ld/st instruction forms to equivalent ARM forms.
598 *
599 * NOTES:
600 * 1. Comments below refer to ARM ARM DDI0100E Thumb Instruction sections.
601 * 2. If for some reason we're passed an non-ld/st Thumb instruction to
602 *    decode, we return 0xdeadc0de. This should never happen under normal
603 *    circumstances but if it does, we've got other problems to deal with
604 *    elsewhere and we obviously can't fix those problems here.
605 */
606
607static unsigned long
608thumb2arm(u16 tinstr)
609{
610	u32 L = (tinstr & (1<<11)) >> 11;
611
612	switch ((tinstr & 0xf800) >> 11) {
613	/* 6.5.1 Format 1: */
614	case 0x6000 >> 11:				/* 7.1.52 STR(1) */
615	case 0x6800 >> 11:				/* 7.1.26 LDR(1) */
616	case 0x7000 >> 11:				/* 7.1.55 STRB(1) */
617	case 0x7800 >> 11:				/* 7.1.30 LDRB(1) */
618		return 0xe5800000 |
619			((tinstr & (1<<12)) << (22-12)) |	/* fixup */
620			(L<<20) |				/* L==1? */
621			((tinstr & (7<<0)) << (12-0)) |		/* Rd */
622			((tinstr & (7<<3)) << (16-3)) |		/* Rn */
623			((tinstr & (31<<6)) >>			/* immed_5 */
624				(6 - ((tinstr & (1<<12)) ? 0 : 2)));
625	case 0x8000 >> 11:				/* 7.1.57 STRH(1) */
626	case 0x8800 >> 11:				/* 7.1.32 LDRH(1) */
627		return 0xe1c000b0 |
628			(L<<20) |				/* L==1? */
629			((tinstr & (7<<0)) << (12-0)) |		/* Rd */
630			((tinstr & (7<<3)) << (16-3)) |		/* Rn */
631			((tinstr & (7<<6)) >> (6-1)) |	 /* immed_5[2:0] */
632			((tinstr & (3<<9)) >> (9-8));	 /* immed_5[4:3] */
633
634	/* 6.5.1 Format 2: */
635	case 0x5000 >> 11:
636	case 0x5800 >> 11:
637		{
638			static const u32 subset[8] = {
639				0xe7800000,		/* 7.1.53 STR(2) */
640				0xe18000b0,		/* 7.1.58 STRH(2) */
641				0xe7c00000,		/* 7.1.56 STRB(2) */
642				0xe19000d0,		/* 7.1.34 LDRSB */
643				0xe7900000,		/* 7.1.27 LDR(2) */
644				0xe19000b0,		/* 7.1.33 LDRH(2) */
645				0xe7d00000,		/* 7.1.31 LDRB(2) */
646				0xe19000f0		/* 7.1.35 LDRSH */
647			};
648			return subset[(tinstr & (7<<9)) >> 9] |
649			    ((tinstr & (7<<0)) << (12-0)) |	/* Rd */
650			    ((tinstr & (7<<3)) << (16-3)) |	/* Rn */
651			    ((tinstr & (7<<6)) >> (6-0));	/* Rm */
652		}
653
654	/* 6.5.1 Format 3: */
655	case 0x4800 >> 11:				/* 7.1.28 LDR(3) */
656		/* NOTE: This case is not technically possible. We're
657		 *	 loading 32-bit memory data via PC relative
658		 *	 addressing mode. So we can and should eliminate
659		 *	 this case. But I'll leave it here for now.
660		 */
661		return 0xe59f0000 |
662		    ((tinstr & (7<<8)) << (12-8)) |		/* Rd */
663		    ((tinstr & 255) << (2-0));			/* immed_8 */
664
665	/* 6.5.1 Format 4: */
666	case 0x9000 >> 11:				/* 7.1.54 STR(3) */
667	case 0x9800 >> 11:				/* 7.1.29 LDR(4) */
668		return 0xe58d0000 |
669			(L<<20) |				/* L==1? */
670			((tinstr & (7<<8)) << (12-8)) |		/* Rd */
671			((tinstr & 255) << 2);			/* immed_8 */
672
673	/* 6.6.1 Format 1: */
674	case 0xc000 >> 11:				/* 7.1.51 STMIA */
675	case 0xc800 >> 11:				/* 7.1.25 LDMIA */
676		{
677			u32 Rn = (tinstr & (7<<8)) >> 8;
678			u32 W = ((L<<Rn) & (tinstr&255)) ? 0 : 1<<21;
679
680			return 0xe8800000 | W | (L<<20) | (Rn<<16) |
681				(tinstr&255);
682		}
683
684	/* 6.6.1 Format 2: */
685	case 0xb000 >> 11:				/* 7.1.48 PUSH */
686	case 0xb800 >> 11:				/* 7.1.47 POP */
687		if ((tinstr & (3 << 9)) == 0x0400) {
688			static const u32 subset[4] = {
689				0xe92d0000,	/* STMDB sp!,{registers} */
690				0xe92d4000,	/* STMDB sp!,{registers,lr} */
691				0xe8bd0000,	/* LDMIA sp!,{registers} */
692				0xe8bd8000	/* LDMIA sp!,{registers,pc} */
693			};
694			return subset[(L<<1) | ((tinstr & (1<<8)) >> 8)] |
695			    (tinstr & 255);		/* register_list */
696		}
697		fallthrough;	/* for illegal instruction case */
698
699	default:
700		return BAD_INSTR;
701	}
702}
703
704/*
705 * Convert Thumb-2 32 bit LDM, STM, LDRD, STRD to equivalent instruction
706 * handlable by ARM alignment handler, also find the corresponding handler,
707 * so that we can reuse ARM userland alignment fault fixups for Thumb.
708 *
709 * @pinstr: original Thumb-2 instruction; returns new handlable instruction
710 * @regs: register context.
711 * @poffset: return offset from faulted addr for later writeback
712 *
713 * NOTES:
714 * 1. Comments below refer to ARMv7 DDI0406A Thumb Instruction sections.
715 * 2. Register name Rt from ARMv7 is same as Rd from ARMv6 (Rd is Rt)
716 */
717static void *
718do_alignment_t32_to_handler(u32 *pinstr, struct pt_regs *regs,
719			    union offset_union *poffset)
720{
721	u32 instr = *pinstr;
722	u16 tinst1 = (instr >> 16) & 0xffff;
723	u16 tinst2 = instr & 0xffff;
724
725	switch (tinst1 & 0xffe0) {
726	/* A6.3.5 Load/Store multiple */
727	case 0xe880:		/* STM/STMIA/STMEA,LDM/LDMIA, PUSH/POP T2 */
728	case 0xe8a0:		/* ...above writeback version */
729	case 0xe900:		/* STMDB/STMFD, LDMDB/LDMEA */
730	case 0xe920:		/* ...above writeback version */
731		/* no need offset decision since handler calculates it */
732		return do_alignment_ldmstm;
733
734	case 0xf840:		/* POP/PUSH T3 (single register) */
735		if (RN_BITS(instr) == 13 && (tinst2 & 0x09ff) == 0x0904) {
736			u32 L = !!(LDST_L_BIT(instr));
737			const u32 subset[2] = {
738				0xe92d0000,	/* STMDB sp!,{registers} */
739				0xe8bd0000,	/* LDMIA sp!,{registers} */
740			};
741			*pinstr = subset[L] | (1<<RD_BITS(instr));
742			return do_alignment_ldmstm;
743		}
744		/* Else fall through for illegal instruction case */
745		break;
746
747	/* A6.3.6 Load/store double, STRD/LDRD(immed, lit, reg) */
748	case 0xe860:
749	case 0xe960:
750	case 0xe8e0:
751	case 0xe9e0:
752		poffset->un = (tinst2 & 0xff) << 2;
753		fallthrough;
754
755	case 0xe940:
756	case 0xe9c0:
757		return do_alignment_ldrdstrd;
758
759	/*
760	 * No need to handle load/store instructions up to word size
761	 * since ARMv6 and later CPUs can perform unaligned accesses.
762	 */
763	default:
764		break;
765	}
766	return NULL;
767}
768
769static int alignment_get_arm(struct pt_regs *regs, u32 *ip, u32 *inst)
770{
771	u32 instr = 0;
772	int fault;
773
774	if (user_mode(regs))
775		fault = get_user(instr, ip);
776	else
777		fault = get_kernel_nofault(instr, ip);
778
779	*inst = __mem_to_opcode_arm(instr);
780
781	return fault;
782}
783
784static int alignment_get_thumb(struct pt_regs *regs, u16 *ip, u16 *inst)
785{
786	u16 instr = 0;
787	int fault;
788
789	if (user_mode(regs))
790		fault = get_user(instr, ip);
791	else
792		fault = get_kernel_nofault(instr, ip);
793
794	*inst = __mem_to_opcode_thumb16(instr);
795
796	return fault;
797}
798
799static int
800do_alignment(unsigned long addr, unsigned int fsr, struct pt_regs *regs)
801{
802	union offset_union offset;
803	unsigned long instrptr;
804	int (*handler)(unsigned long addr, u32 instr, struct pt_regs *regs);
805	unsigned int type;
806	u32 instr = 0;
807	u16 tinstr = 0;
808	int isize = 4;
809	int thumb2_32b = 0;
810	int fault;
811
812	if (interrupts_enabled(regs))
813		local_irq_enable();
814
815	instrptr = instruction_pointer(regs);
816
817	if (thumb_mode(regs)) {
818		u16 *ptr = (u16 *)(instrptr & ~1);
819
820		fault = alignment_get_thumb(regs, ptr, &tinstr);
821		if (!fault) {
822			if (cpu_architecture() >= CPU_ARCH_ARMv7 &&
823			    IS_T32(tinstr)) {
824				/* Thumb-2 32-bit */
825				u16 tinst2;
826				fault = alignment_get_thumb(regs, ptr + 1, &tinst2);
827				instr = __opcode_thumb32_compose(tinstr, tinst2);
828				thumb2_32b = 1;
829			} else {
830				isize = 2;
831				instr = thumb2arm(tinstr);
832			}
833		}
834	} else {
835		fault = alignment_get_arm(regs, (void *)instrptr, &instr);
836	}
837
838	if (fault) {
839		type = TYPE_FAULT;
840		goto bad_or_fault;
841	}
842
843	if (user_mode(regs))
844		goto user;
845
846	ai_sys += 1;
847	ai_sys_last_pc = (void *)instruction_pointer(regs);
848
849 fixup:
850
851	regs->ARM_pc += isize;
852
853	switch (CODING_BITS(instr)) {
854	case 0x00000000:	/* 3.13.4 load/store instruction extensions */
855		if (LDSTHD_I_BIT(instr))
856			offset.un = (instr & 0xf00) >> 4 | (instr & 15);
857		else
858			offset.un = regs->uregs[RM_BITS(instr)];
859
860		if ((instr & 0x000000f0) == 0x000000b0 || /* LDRH, STRH */
861		    (instr & 0x001000f0) == 0x001000f0)   /* LDRSH */
862			handler = do_alignment_ldrhstrh;
863		else if ((instr & 0x001000f0) == 0x000000d0 || /* LDRD */
864			 (instr & 0x001000f0) == 0x000000f0)   /* STRD */
865			handler = do_alignment_ldrdstrd;
866		else if ((instr & 0x01f00ff0) == 0x01000090) /* SWP */
867			goto swp;
868		else
869			goto bad;
870		break;
871
872	case 0x04000000:	/* ldr or str immediate */
873		if (COND_BITS(instr) == 0xf0000000) /* NEON VLDn, VSTn */
874			goto bad;
875		offset.un = OFFSET_BITS(instr);
876		handler = do_alignment_ldrstr;
877		break;
878
879	case 0x06000000:	/* ldr or str register */
880		offset.un = regs->uregs[RM_BITS(instr)];
881
882		if (IS_SHIFT(instr)) {
883			unsigned int shiftval = SHIFT_BITS(instr);
884
885			switch(SHIFT_TYPE(instr)) {
886			case SHIFT_LSL:
887				offset.un <<= shiftval;
888				break;
889
890			case SHIFT_LSR:
891				offset.un >>= shiftval;
892				break;
893
894			case SHIFT_ASR:
895				offset.sn >>= shiftval;
896				break;
897
898			case SHIFT_RORRRX:
899				if (shiftval == 0) {
900					offset.un >>= 1;
901					if (regs->ARM_cpsr & PSR_C_BIT)
902						offset.un |= 1 << 31;
903				} else
904					offset.un = offset.un >> shiftval |
905							  offset.un << (32 - shiftval);
906				break;
907			}
908		}
909		handler = do_alignment_ldrstr;
910		break;
911
912	case 0x08000000:	/* ldm or stm, or thumb-2 32bit instruction */
913		if (thumb2_32b) {
914			offset.un = 0;
915			handler = do_alignment_t32_to_handler(&instr, regs, &offset);
916		} else {
917			offset.un = 0;
918			handler = do_alignment_ldmstm;
919		}
920		break;
921
922	default:
923		goto bad;
924	}
925
926	if (!handler)
927		goto bad;
928	type = handler(addr, instr, regs);
929
930	if (type == TYPE_ERROR || type == TYPE_FAULT) {
931		regs->ARM_pc -= isize;
932		goto bad_or_fault;
933	}
934
935	if (type == TYPE_LDST)
936		do_alignment_finish_ldst(addr, instr, regs, offset);
937
938	if (thumb_mode(regs))
939		regs->ARM_cpsr = it_advance(regs->ARM_cpsr);
940
941	return 0;
942
943 bad_or_fault:
944	if (type == TYPE_ERROR)
945		goto bad;
946	/*
947	 * We got a fault - fix it up, or die.
948	 */
949	do_bad_area(addr, fsr, regs);
950	return 0;
951
952 swp:
953	pr_err("Alignment trap: not handling swp instruction\n");
954
955 bad:
956	/*
957	 * Oops, we didn't handle the instruction.
958	 */
959	pr_err("Alignment trap: not handling instruction "
960		"%0*x at [<%08lx>]\n",
961		isize << 1,
962		isize == 2 ? tinstr : instr, instrptr);
963	ai_skipped += 1;
964	return 1;
965
966 user:
967	ai_user += 1;
968
969	if (ai_usermode & UM_WARN)
970		printk("Alignment trap: %s (%d) PC=0x%08lx Instr=0x%0*x "
971		       "Address=0x%08lx FSR 0x%03x\n", current->comm,
972			task_pid_nr(current), instrptr,
973			isize << 1,
974			isize == 2 ? tinstr : instr,
975		        addr, fsr);
976
977	if (ai_usermode & UM_FIXUP)
978		goto fixup;
979
980	if (ai_usermode & UM_SIGNAL) {
981		force_sig_fault(SIGBUS, BUS_ADRALN, (void __user *)addr);
982	} else {
983		/*
984		 * We're about to disable the alignment trap and return to
985		 * user space.  But if an interrupt occurs before actually
986		 * reaching user space, then the IRQ vector entry code will
987		 * notice that we were still in kernel space and therefore
988		 * the alignment trap won't be re-enabled in that case as it
989		 * is presumed to be always on from kernel space.
990		 * Let's prevent that race by disabling interrupts here (they
991		 * are disabled on the way back to user space anyway in
992		 * entry-common.S) and disable the alignment trap only if
993		 * there is no work pending for this thread.
994		 */
995		raw_local_irq_disable();
996		if (!(read_thread_flags() & _TIF_WORK_MASK))
997			set_cr(cr_no_alignment);
998	}
999
1000	return 0;
1001}
1002
1003static int __init noalign_setup(char *__unused)
1004{
1005	set_cr(__clear_cr(CR_A));
1006	return 1;
1007}
1008__setup("noalign", noalign_setup);
1009
1010/*
1011 * This needs to be done after sysctl_init_bases(), otherwise sys/ will be
1012 * overwritten.  Actually, this shouldn't be in sys/ at all since
1013 * it isn't a sysctl, and it doesn't contain sysctl information.
1014 * We now locate it in /proc/cpu/alignment instead.
1015 */
1016static int __init alignment_init(void)
1017{
1018#ifdef CONFIG_PROC_FS
1019	struct proc_dir_entry *res;
1020
1021	res = proc_create("cpu/alignment", S_IWUSR | S_IRUGO, NULL,
1022			  &alignment_proc_ops);
1023	if (!res)
1024		return -ENOMEM;
1025#endif
1026
1027	if (cpu_is_v6_unaligned()) {
1028		set_cr(__clear_cr(CR_A));
1029		ai_usermode = safe_usermode(ai_usermode, false);
1030	}
1031
1032	cr_no_alignment = get_cr() & ~CR_A;
1033
1034	hook_fault_code(FAULT_CODE_ALIGNMENT, do_alignment, SIGBUS, BUS_ADRALN,
1035			"alignment exception");
1036
1037	/*
1038	 * ARMv6K and ARMv7 use fault status 3 (0b00011) as Access Flag section
1039	 * fault, not as alignment error.
1040	 *
1041	 * TODO: handle ARMv6K properly. Runtime check for 'K' extension is
1042	 * needed.
1043	 */
1044	if (cpu_architecture() <= CPU_ARCH_ARMv6) {
1045		hook_fault_code(3, do_alignment, SIGBUS, BUS_ADRALN,
1046				"alignment exception");
1047	}
1048
1049	return 0;
1050}
1051
1052fs_initcall(alignment_init);
1053