xref: /openbmc/linux/arch/arm/mm/alignment.c (revision d2999e1b)
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 
45 #define LDST_I_BIT(i)	(i & (1 << 26))		/* Immediate constant	*/
46 #define LDST_P_BIT(i)	(i & (1 << 24))		/* Preindex		*/
47 #define LDST_U_BIT(i)	(i & (1 << 23))		/* Add offset		*/
48 #define LDST_W_BIT(i)	(i & (1 << 21))		/* Writeback		*/
49 #define LDST_L_BIT(i)	(i & (1 << 20))		/* Load			*/
50 
51 #define LDST_P_EQ_U(i)	((((i) ^ ((i) >> 1)) & (1 << 23)) == 0)
52 
53 #define LDSTHD_I_BIT(i)	(i & (1 << 22))		/* double/half-word immed */
54 #define LDM_S_BIT(i)	(i & (1 << 22))		/* write CPSR from SPSR	*/
55 
56 #define RN_BITS(i)	((i >> 16) & 15)	/* Rn			*/
57 #define RD_BITS(i)	((i >> 12) & 15)	/* Rd			*/
58 #define RM_BITS(i)	(i & 15)		/* Rm			*/
59 
60 #define REGMASK_BITS(i)	(i & 0xffff)
61 #define OFFSET_BITS(i)	(i & 0x0fff)
62 
63 #define IS_SHIFT(i)	(i & 0x0ff0)
64 #define SHIFT_BITS(i)	((i >> 7) & 0x1f)
65 #define SHIFT_TYPE(i)	(i & 0x60)
66 #define SHIFT_LSL	0x00
67 #define SHIFT_LSR	0x20
68 #define SHIFT_ASR	0x40
69 #define SHIFT_RORRRX	0x60
70 
71 #define BAD_INSTR 	0xdeadc0de
72 
73 /* Thumb-2 32 bit format per ARMv7 DDI0406A A6.3, either f800h,e800h,f800h */
74 #define IS_T32(hi16) \
75 	(((hi16) & 0xe000) == 0xe000 && ((hi16) & 0x1800))
76 
77 static unsigned long ai_user;
78 static unsigned long ai_sys;
79 static unsigned long ai_skipped;
80 static unsigned long ai_half;
81 static unsigned long ai_word;
82 static unsigned long ai_dword;
83 static unsigned long ai_multi;
84 static int ai_usermode;
85 static unsigned long cr_no_alignment;
86 
87 core_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. */
94 static bool cpu_is_v6_unaligned(void)
95 {
96 	return cpu_architecture() >= CPU_ARCH_ARMv6 && get_cr() & CR_U;
97 }
98 
99 static 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 			printk(KERN_WARNING "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
121 static const char *usermode_action[] = {
122 	"ignored",
123 	"warn",
124 	"fixup",
125 	"fixup+warn",
126 	"signal",
127 	"signal+warn"
128 };
129 
130 static 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\n", ai_sys);
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 
146 static int alignment_proc_open(struct inode *inode, struct file *file)
147 {
148 	return single_open(file, alignment_proc_show, NULL);
149 }
150 
151 static 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 
165 static const struct file_operations alignment_proc_fops = {
166 	.open		= alignment_proc_open,
167 	.read		= seq_read,
168 	.llseek		= seq_lseek,
169 	.release	= single_release,
170 	.write		= alignment_proc_write,
171 };
172 #endif /* CONFIG_PROC_FS */
173 
174 union 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 .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 .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 .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 
326 static void
327 do_alignment_finish_ldst(unsigned long addr, unsigned long 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 
339 static int
340 do_alignment_ldrhstrh(unsigned long addr, unsigned long 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 		get16t_unaligned_check(val, addr);
367 
368 		/* signed half-word? */
369 		if (instr & 0x40)
370 			val = (signed long)((signed short) val);
371 
372 		regs->uregs[rd] = val;
373 	} else
374 		put16t_unaligned_check(regs->uregs[rd], addr);
375 
376 	return TYPE_LDST;
377 
378  fault:
379 	return TYPE_FAULT;
380 }
381 
382 static int
383 do_alignment_ldrdstrd(unsigned long addr, unsigned long instr,
384 		      struct pt_regs *regs)
385 {
386 	unsigned int rd = RD_BITS(instr);
387 	unsigned int rd2;
388 	int load;
389 
390 	if ((instr & 0xfe000000) == 0xe8000000) {
391 		/* ARMv7 Thumb-2 32-bit LDRD/STRD */
392 		rd2 = (instr >> 8) & 0xf;
393 		load = !!(LDST_L_BIT(instr));
394 	} else if (((rd & 1) == 1) || (rd == 14))
395 		goto bad;
396 	else {
397 		load = ((instr & 0xf0) == 0xd0);
398 		rd2 = rd + 1;
399 	}
400 
401 	ai_dword += 1;
402 
403 	if (user_mode(regs))
404 		goto user;
405 
406 	if (load) {
407 		unsigned long val;
408 		get32_unaligned_check(val, addr);
409 		regs->uregs[rd] = val;
410 		get32_unaligned_check(val, addr + 4);
411 		regs->uregs[rd2] = val;
412 	} else {
413 		put32_unaligned_check(regs->uregs[rd], addr);
414 		put32_unaligned_check(regs->uregs[rd2], addr + 4);
415 	}
416 
417 	return TYPE_LDST;
418 
419  user:
420 	if (load) {
421 		unsigned long val;
422 		get32t_unaligned_check(val, addr);
423 		regs->uregs[rd] = val;
424 		get32t_unaligned_check(val, addr + 4);
425 		regs->uregs[rd2] = val;
426 	} else {
427 		put32t_unaligned_check(regs->uregs[rd], addr);
428 		put32t_unaligned_check(regs->uregs[rd2], addr + 4);
429 	}
430 
431 	return TYPE_LDST;
432  bad:
433 	return TYPE_ERROR;
434  fault:
435 	return TYPE_FAULT;
436 }
437 
438 static int
439 do_alignment_ldrstr(unsigned long addr, unsigned long instr, struct pt_regs *regs)
440 {
441 	unsigned int rd = RD_BITS(instr);
442 
443 	ai_word += 1;
444 
445 	if ((!LDST_P_BIT(instr) && LDST_W_BIT(instr)) || user_mode(regs))
446 		goto trans;
447 
448 	if (LDST_L_BIT(instr)) {
449 		unsigned int val;
450 		get32_unaligned_check(val, addr);
451 		regs->uregs[rd] = val;
452 	} else
453 		put32_unaligned_check(regs->uregs[rd], addr);
454 	return TYPE_LDST;
455 
456  trans:
457 	if (LDST_L_BIT(instr)) {
458 		unsigned int val;
459 		get32t_unaligned_check(val, addr);
460 		regs->uregs[rd] = val;
461 	} else
462 		put32t_unaligned_check(regs->uregs[rd], addr);
463 	return TYPE_LDST;
464 
465  fault:
466 	return TYPE_FAULT;
467 }
468 
469 /*
470  * LDM/STM alignment handler.
471  *
472  * There are 4 variants of this instruction:
473  *
474  * B = rn pointer before instruction, A = rn pointer after instruction
475  *              ------ increasing address ----->
476  *	        |    | r0 | r1 | ... | rx |    |
477  * PU = 01             B                    A
478  * PU = 11        B                    A
479  * PU = 00        A                    B
480  * PU = 10             A                    B
481  */
482 static int
483 do_alignment_ldmstm(unsigned long addr, unsigned long instr, struct pt_regs *regs)
484 {
485 	unsigned int rd, rn, correction, nr_regs, regbits;
486 	unsigned long eaddr, newaddr;
487 
488 	if (LDM_S_BIT(instr))
489 		goto bad;
490 
491 	correction = 4; /* processor implementation defined */
492 	regs->ARM_pc += correction;
493 
494 	ai_multi += 1;
495 
496 	/* count the number of registers in the mask to be transferred */
497 	nr_regs = hweight16(REGMASK_BITS(instr)) * 4;
498 
499 	rn = RN_BITS(instr);
500 	newaddr = eaddr = regs->uregs[rn];
501 
502 	if (!LDST_U_BIT(instr))
503 		nr_regs = -nr_regs;
504 	newaddr += nr_regs;
505 	if (!LDST_U_BIT(instr))
506 		eaddr = newaddr;
507 
508 	if (LDST_P_EQ_U(instr))	/* U = P */
509 		eaddr += 4;
510 
511 	/*
512 	 * For alignment faults on the ARM922T/ARM920T the MMU  makes
513 	 * the FSR (and hence addr) equal to the updated base address
514 	 * of the multiple access rather than the restored value.
515 	 * Switch this message off if we've got a ARM92[02], otherwise
516 	 * [ls]dm alignment faults are noisy!
517 	 */
518 #if !(defined CONFIG_CPU_ARM922T)  && !(defined CONFIG_CPU_ARM920T)
519 	/*
520 	 * This is a "hint" - we already have eaddr worked out by the
521 	 * processor for us.
522 	 */
523 	if (addr != eaddr) {
524 		printk(KERN_ERR "LDMSTM: PC = %08lx, instr = %08lx, "
525 			"addr = %08lx, eaddr = %08lx\n",
526 			 instruction_pointer(regs), instr, addr, eaddr);
527 		show_regs(regs);
528 	}
529 #endif
530 
531 	if (user_mode(regs)) {
532 		for (regbits = REGMASK_BITS(instr), rd = 0; regbits;
533 		     regbits >>= 1, rd += 1)
534 			if (regbits & 1) {
535 				if (LDST_L_BIT(instr)) {
536 					unsigned int val;
537 					get32t_unaligned_check(val, eaddr);
538 					regs->uregs[rd] = val;
539 				} else
540 					put32t_unaligned_check(regs->uregs[rd], eaddr);
541 				eaddr += 4;
542 			}
543 	} else {
544 		for (regbits = REGMASK_BITS(instr), rd = 0; regbits;
545 		     regbits >>= 1, rd += 1)
546 			if (regbits & 1) {
547 				if (LDST_L_BIT(instr)) {
548 					unsigned int val;
549 					get32_unaligned_check(val, eaddr);
550 					regs->uregs[rd] = val;
551 				} else
552 					put32_unaligned_check(regs->uregs[rd], eaddr);
553 				eaddr += 4;
554 			}
555 	}
556 
557 	if (LDST_W_BIT(instr))
558 		regs->uregs[rn] = newaddr;
559 	if (!LDST_L_BIT(instr) || !(REGMASK_BITS(instr) & (1 << 15)))
560 		regs->ARM_pc -= correction;
561 	return TYPE_DONE;
562 
563 fault:
564 	regs->ARM_pc -= correction;
565 	return TYPE_FAULT;
566 
567 bad:
568 	printk(KERN_ERR "Alignment trap: not handling ldm with s-bit set\n");
569 	return TYPE_ERROR;
570 }
571 
572 /*
573  * Convert Thumb ld/st instruction forms to equivalent ARM instructions so
574  * we can reuse ARM userland alignment fault fixups for Thumb.
575  *
576  * This implementation was initially based on the algorithm found in
577  * gdb/sim/arm/thumbemu.c. It is basically just a code reduction of same
578  * to convert only Thumb ld/st instruction forms to equivalent ARM forms.
579  *
580  * NOTES:
581  * 1. Comments below refer to ARM ARM DDI0100E Thumb Instruction sections.
582  * 2. If for some reason we're passed an non-ld/st Thumb instruction to
583  *    decode, we return 0xdeadc0de. This should never happen under normal
584  *    circumstances but if it does, we've got other problems to deal with
585  *    elsewhere and we obviously can't fix those problems here.
586  */
587 
588 static unsigned long
589 thumb2arm(u16 tinstr)
590 {
591 	u32 L = (tinstr & (1<<11)) >> 11;
592 
593 	switch ((tinstr & 0xf800) >> 11) {
594 	/* 6.5.1 Format 1: */
595 	case 0x6000 >> 11:				/* 7.1.52 STR(1) */
596 	case 0x6800 >> 11:				/* 7.1.26 LDR(1) */
597 	case 0x7000 >> 11:				/* 7.1.55 STRB(1) */
598 	case 0x7800 >> 11:				/* 7.1.30 LDRB(1) */
599 		return 0xe5800000 |
600 			((tinstr & (1<<12)) << (22-12)) |	/* fixup */
601 			(L<<20) |				/* L==1? */
602 			((tinstr & (7<<0)) << (12-0)) |		/* Rd */
603 			((tinstr & (7<<3)) << (16-3)) |		/* Rn */
604 			((tinstr & (31<<6)) >>			/* immed_5 */
605 				(6 - ((tinstr & (1<<12)) ? 0 : 2)));
606 	case 0x8000 >> 11:				/* 7.1.57 STRH(1) */
607 	case 0x8800 >> 11:				/* 7.1.32 LDRH(1) */
608 		return 0xe1c000b0 |
609 			(L<<20) |				/* L==1? */
610 			((tinstr & (7<<0)) << (12-0)) |		/* Rd */
611 			((tinstr & (7<<3)) << (16-3)) |		/* Rn */
612 			((tinstr & (7<<6)) >> (6-1)) |	 /* immed_5[2:0] */
613 			((tinstr & (3<<9)) >> (9-8));	 /* immed_5[4:3] */
614 
615 	/* 6.5.1 Format 2: */
616 	case 0x5000 >> 11:
617 	case 0x5800 >> 11:
618 		{
619 			static const u32 subset[8] = {
620 				0xe7800000,		/* 7.1.53 STR(2) */
621 				0xe18000b0,		/* 7.1.58 STRH(2) */
622 				0xe7c00000,		/* 7.1.56 STRB(2) */
623 				0xe19000d0,		/* 7.1.34 LDRSB */
624 				0xe7900000,		/* 7.1.27 LDR(2) */
625 				0xe19000b0,		/* 7.1.33 LDRH(2) */
626 				0xe7d00000,		/* 7.1.31 LDRB(2) */
627 				0xe19000f0		/* 7.1.35 LDRSH */
628 			};
629 			return subset[(tinstr & (7<<9)) >> 9] |
630 			    ((tinstr & (7<<0)) << (12-0)) |	/* Rd */
631 			    ((tinstr & (7<<3)) << (16-3)) |	/* Rn */
632 			    ((tinstr & (7<<6)) >> (6-0));	/* Rm */
633 		}
634 
635 	/* 6.5.1 Format 3: */
636 	case 0x4800 >> 11:				/* 7.1.28 LDR(3) */
637 		/* NOTE: This case is not technically possible. We're
638 		 *	 loading 32-bit memory data via PC relative
639 		 *	 addressing mode. So we can and should eliminate
640 		 *	 this case. But I'll leave it here for now.
641 		 */
642 		return 0xe59f0000 |
643 		    ((tinstr & (7<<8)) << (12-8)) |		/* Rd */
644 		    ((tinstr & 255) << (2-0));			/* immed_8 */
645 
646 	/* 6.5.1 Format 4: */
647 	case 0x9000 >> 11:				/* 7.1.54 STR(3) */
648 	case 0x9800 >> 11:				/* 7.1.29 LDR(4) */
649 		return 0xe58d0000 |
650 			(L<<20) |				/* L==1? */
651 			((tinstr & (7<<8)) << (12-8)) |		/* Rd */
652 			((tinstr & 255) << 2);			/* immed_8 */
653 
654 	/* 6.6.1 Format 1: */
655 	case 0xc000 >> 11:				/* 7.1.51 STMIA */
656 	case 0xc800 >> 11:				/* 7.1.25 LDMIA */
657 		{
658 			u32 Rn = (tinstr & (7<<8)) >> 8;
659 			u32 W = ((L<<Rn) & (tinstr&255)) ? 0 : 1<<21;
660 
661 			return 0xe8800000 | W | (L<<20) | (Rn<<16) |
662 				(tinstr&255);
663 		}
664 
665 	/* 6.6.1 Format 2: */
666 	case 0xb000 >> 11:				/* 7.1.48 PUSH */
667 	case 0xb800 >> 11:				/* 7.1.47 POP */
668 		if ((tinstr & (3 << 9)) == 0x0400) {
669 			static const u32 subset[4] = {
670 				0xe92d0000,	/* STMDB sp!,{registers} */
671 				0xe92d4000,	/* STMDB sp!,{registers,lr} */
672 				0xe8bd0000,	/* LDMIA sp!,{registers} */
673 				0xe8bd8000	/* LDMIA sp!,{registers,pc} */
674 			};
675 			return subset[(L<<1) | ((tinstr & (1<<8)) >> 8)] |
676 			    (tinstr & 255);		/* register_list */
677 		}
678 		/* Else fall through for illegal instruction case */
679 
680 	default:
681 		return BAD_INSTR;
682 	}
683 }
684 
685 /*
686  * Convert Thumb-2 32 bit LDM, STM, LDRD, STRD to equivalent instruction
687  * handlable by ARM alignment handler, also find the corresponding handler,
688  * so that we can reuse ARM userland alignment fault fixups for Thumb.
689  *
690  * @pinstr: original Thumb-2 instruction; returns new handlable instruction
691  * @regs: register context.
692  * @poffset: return offset from faulted addr for later writeback
693  *
694  * NOTES:
695  * 1. Comments below refer to ARMv7 DDI0406A Thumb Instruction sections.
696  * 2. Register name Rt from ARMv7 is same as Rd from ARMv6 (Rd is Rt)
697  */
698 static void *
699 do_alignment_t32_to_handler(unsigned long *pinstr, struct pt_regs *regs,
700 			    union offset_union *poffset)
701 {
702 	unsigned long instr = *pinstr;
703 	u16 tinst1 = (instr >> 16) & 0xffff;
704 	u16 tinst2 = instr & 0xffff;
705 
706 	switch (tinst1 & 0xffe0) {
707 	/* A6.3.5 Load/Store multiple */
708 	case 0xe880:		/* STM/STMIA/STMEA,LDM/LDMIA, PUSH/POP T2 */
709 	case 0xe8a0:		/* ...above writeback version */
710 	case 0xe900:		/* STMDB/STMFD, LDMDB/LDMEA */
711 	case 0xe920:		/* ...above writeback version */
712 		/* no need offset decision since handler calculates it */
713 		return do_alignment_ldmstm;
714 
715 	case 0xf840:		/* POP/PUSH T3 (single register) */
716 		if (RN_BITS(instr) == 13 && (tinst2 & 0x09ff) == 0x0904) {
717 			u32 L = !!(LDST_L_BIT(instr));
718 			const u32 subset[2] = {
719 				0xe92d0000,	/* STMDB sp!,{registers} */
720 				0xe8bd0000,	/* LDMIA sp!,{registers} */
721 			};
722 			*pinstr = subset[L] | (1<<RD_BITS(instr));
723 			return do_alignment_ldmstm;
724 		}
725 		/* Else fall through for illegal instruction case */
726 		break;
727 
728 	/* A6.3.6 Load/store double, STRD/LDRD(immed, lit, reg) */
729 	case 0xe860:
730 	case 0xe960:
731 	case 0xe8e0:
732 	case 0xe9e0:
733 		poffset->un = (tinst2 & 0xff) << 2;
734 	case 0xe940:
735 	case 0xe9c0:
736 		return do_alignment_ldrdstrd;
737 
738 	/*
739 	 * No need to handle load/store instructions up to word size
740 	 * since ARMv6 and later CPUs can perform unaligned accesses.
741 	 */
742 	default:
743 		break;
744 	}
745 	return NULL;
746 }
747 
748 static int
749 do_alignment(unsigned long addr, unsigned int fsr, struct pt_regs *regs)
750 {
751 	union offset_union uninitialized_var(offset);
752 	unsigned long instr = 0, instrptr;
753 	int (*handler)(unsigned long addr, unsigned long instr, struct pt_regs *regs);
754 	unsigned int type;
755 	unsigned int fault;
756 	u16 tinstr = 0;
757 	int isize = 4;
758 	int thumb2_32b = 0;
759 
760 	if (interrupts_enabled(regs))
761 		local_irq_enable();
762 
763 	instrptr = instruction_pointer(regs);
764 
765 	if (thumb_mode(regs)) {
766 		u16 *ptr = (u16 *)(instrptr & ~1);
767 		fault = probe_kernel_address(ptr, tinstr);
768 		tinstr = __mem_to_opcode_thumb16(tinstr);
769 		if (!fault) {
770 			if (cpu_architecture() >= CPU_ARCH_ARMv7 &&
771 			    IS_T32(tinstr)) {
772 				/* Thumb-2 32-bit */
773 				u16 tinst2 = 0;
774 				fault = probe_kernel_address(ptr + 1, tinst2);
775 				tinst2 = __mem_to_opcode_thumb16(tinst2);
776 				instr = __opcode_thumb32_compose(tinstr, tinst2);
777 				thumb2_32b = 1;
778 			} else {
779 				isize = 2;
780 				instr = thumb2arm(tinstr);
781 			}
782 		}
783 	} else {
784 		fault = probe_kernel_address(instrptr, instr);
785 		instr = __mem_to_opcode_arm(instr);
786 	}
787 
788 	if (fault) {
789 		type = TYPE_FAULT;
790 		goto bad_or_fault;
791 	}
792 
793 	if (user_mode(regs))
794 		goto user;
795 
796 	ai_sys += 1;
797 
798  fixup:
799 
800 	regs->ARM_pc += isize;
801 
802 	switch (CODING_BITS(instr)) {
803 	case 0x00000000:	/* 3.13.4 load/store instruction extensions */
804 		if (LDSTHD_I_BIT(instr))
805 			offset.un = (instr & 0xf00) >> 4 | (instr & 15);
806 		else
807 			offset.un = regs->uregs[RM_BITS(instr)];
808 
809 		if ((instr & 0x000000f0) == 0x000000b0 || /* LDRH, STRH */
810 		    (instr & 0x001000f0) == 0x001000f0)   /* LDRSH */
811 			handler = do_alignment_ldrhstrh;
812 		else if ((instr & 0x001000f0) == 0x000000d0 || /* LDRD */
813 			 (instr & 0x001000f0) == 0x000000f0)   /* STRD */
814 			handler = do_alignment_ldrdstrd;
815 		else if ((instr & 0x01f00ff0) == 0x01000090) /* SWP */
816 			goto swp;
817 		else
818 			goto bad;
819 		break;
820 
821 	case 0x04000000:	/* ldr or str immediate */
822 		offset.un = OFFSET_BITS(instr);
823 		handler = do_alignment_ldrstr;
824 		break;
825 
826 	case 0x06000000:	/* ldr or str register */
827 		offset.un = regs->uregs[RM_BITS(instr)];
828 
829 		if (IS_SHIFT(instr)) {
830 			unsigned int shiftval = SHIFT_BITS(instr);
831 
832 			switch(SHIFT_TYPE(instr)) {
833 			case SHIFT_LSL:
834 				offset.un <<= shiftval;
835 				break;
836 
837 			case SHIFT_LSR:
838 				offset.un >>= shiftval;
839 				break;
840 
841 			case SHIFT_ASR:
842 				offset.sn >>= shiftval;
843 				break;
844 
845 			case SHIFT_RORRRX:
846 				if (shiftval == 0) {
847 					offset.un >>= 1;
848 					if (regs->ARM_cpsr & PSR_C_BIT)
849 						offset.un |= 1 << 31;
850 				} else
851 					offset.un = offset.un >> shiftval |
852 							  offset.un << (32 - shiftval);
853 				break;
854 			}
855 		}
856 		handler = do_alignment_ldrstr;
857 		break;
858 
859 	case 0x08000000:	/* ldm or stm, or thumb-2 32bit instruction */
860 		if (thumb2_32b) {
861 			offset.un = 0;
862 			handler = do_alignment_t32_to_handler(&instr, regs, &offset);
863 		} else {
864 			offset.un = 0;
865 			handler = do_alignment_ldmstm;
866 		}
867 		break;
868 
869 	default:
870 		goto bad;
871 	}
872 
873 	if (!handler)
874 		goto bad;
875 	type = handler(addr, instr, regs);
876 
877 	if (type == TYPE_ERROR || type == TYPE_FAULT) {
878 		regs->ARM_pc -= isize;
879 		goto bad_or_fault;
880 	}
881 
882 	if (type == TYPE_LDST)
883 		do_alignment_finish_ldst(addr, instr, regs, offset);
884 
885 	return 0;
886 
887  bad_or_fault:
888 	if (type == TYPE_ERROR)
889 		goto bad;
890 	/*
891 	 * We got a fault - fix it up, or die.
892 	 */
893 	do_bad_area(addr, fsr, regs);
894 	return 0;
895 
896  swp:
897 	printk(KERN_ERR "Alignment trap: not handling swp instruction\n");
898 
899  bad:
900 	/*
901 	 * Oops, we didn't handle the instruction.
902 	 */
903 	printk(KERN_ERR "Alignment trap: not handling instruction "
904 		"%0*lx at [<%08lx>]\n",
905 		isize << 1,
906 		isize == 2 ? tinstr : instr, instrptr);
907 	ai_skipped += 1;
908 	return 1;
909 
910  user:
911 	ai_user += 1;
912 
913 	if (ai_usermode & UM_WARN)
914 		printk("Alignment trap: %s (%d) PC=0x%08lx Instr=0x%0*lx "
915 		       "Address=0x%08lx FSR 0x%03x\n", current->comm,
916 			task_pid_nr(current), instrptr,
917 			isize << 1,
918 			isize == 2 ? tinstr : instr,
919 		        addr, fsr);
920 
921 	if (ai_usermode & UM_FIXUP)
922 		goto fixup;
923 
924 	if (ai_usermode & UM_SIGNAL) {
925 		siginfo_t si;
926 
927 		si.si_signo = SIGBUS;
928 		si.si_errno = 0;
929 		si.si_code = BUS_ADRALN;
930 		si.si_addr = (void __user *)addr;
931 
932 		force_sig_info(si.si_signo, &si, current);
933 	} else {
934 		/*
935 		 * We're about to disable the alignment trap and return to
936 		 * user space.  But if an interrupt occurs before actually
937 		 * reaching user space, then the IRQ vector entry code will
938 		 * notice that we were still in kernel space and therefore
939 		 * the alignment trap won't be re-enabled in that case as it
940 		 * is presumed to be always on from kernel space.
941 		 * Let's prevent that race by disabling interrupts here (they
942 		 * are disabled on the way back to user space anyway in
943 		 * entry-common.S) and disable the alignment trap only if
944 		 * there is no work pending for this thread.
945 		 */
946 		raw_local_irq_disable();
947 		if (!(current_thread_info()->flags & _TIF_WORK_MASK))
948 			set_cr(cr_no_alignment);
949 	}
950 
951 	return 0;
952 }
953 
954 static int __init noalign_setup(char *__unused)
955 {
956 	set_cr(__clear_cr(CR_A));
957 	return 1;
958 }
959 __setup("noalign", noalign_setup);
960 
961 /*
962  * This needs to be done after sysctl_init, otherwise sys/ will be
963  * overwritten.  Actually, this shouldn't be in sys/ at all since
964  * it isn't a sysctl, and it doesn't contain sysctl information.
965  * We now locate it in /proc/cpu/alignment instead.
966  */
967 static int __init alignment_init(void)
968 {
969 #ifdef CONFIG_PROC_FS
970 	struct proc_dir_entry *res;
971 
972 	res = proc_create("cpu/alignment", S_IWUSR | S_IRUGO, NULL,
973 			  &alignment_proc_fops);
974 	if (!res)
975 		return -ENOMEM;
976 #endif
977 
978 	if (cpu_is_v6_unaligned()) {
979 		set_cr(__clear_cr(CR_A));
980 		ai_usermode = safe_usermode(ai_usermode, false);
981 	}
982 
983 	cr_no_alignment = get_cr() & ~CR_A;
984 
985 	hook_fault_code(FAULT_CODE_ALIGNMENT, do_alignment, SIGBUS, BUS_ADRALN,
986 			"alignment exception");
987 
988 	/*
989 	 * ARMv6K and ARMv7 use fault status 3 (0b00011) as Access Flag section
990 	 * fault, not as alignment error.
991 	 *
992 	 * TODO: handle ARMv6K properly. Runtime check for 'K' extension is
993 	 * needed.
994 	 */
995 	if (cpu_architecture() <= CPU_ARCH_ARMv6) {
996 		hook_fault_code(3, do_alignment, SIGBUS, BUS_ADRALN,
997 				"alignment exception");
998 	}
999 
1000 	return 0;
1001 }
1002 
1003 fs_initcall(alignment_init);
1004