xref: /openbmc/linux/arch/arm/mm/alignment.c (revision c0e297dc)
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 
78 static unsigned long ai_user;
79 static unsigned long ai_sys;
80 static void *ai_sys_last_pc;
81 static unsigned long ai_skipped;
82 static unsigned long ai_half;
83 static unsigned long ai_word;
84 static unsigned long ai_dword;
85 static unsigned long ai_multi;
86 static int ai_usermode;
87 static unsigned long cr_no_alignment;
88 
89 core_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. */
96 static bool cpu_is_v6_unaligned(void)
97 {
98 	return cpu_architecture() >= CPU_ARCH_ARMv6 && get_cr() & CR_U;
99 }
100 
101 static 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
123 static const char *usermode_action[] = {
124 	"ignored",
125 	"warn",
126 	"fixup",
127 	"fixup+warn",
128 	"signal",
129 	"signal+warn"
130 };
131 
132 static 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 
148 static int alignment_proc_open(struct inode *inode, struct file *file)
149 {
150 	return single_open(file, alignment_proc_show, NULL);
151 }
152 
153 static 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 
167 static 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 
176 union 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 
328 static void
329 do_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 
341 static int
342 do_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 		get16t_unaligned_check(val, addr);
369 
370 		/* signed half-word? */
371 		if (instr & 0x40)
372 			val = (signed long)((signed short) val);
373 
374 		regs->uregs[rd] = val;
375 	} else
376 		put16t_unaligned_check(regs->uregs[rd], addr);
377 
378 	return TYPE_LDST;
379 
380  fault:
381 	return TYPE_FAULT;
382 }
383 
384 static int
385 do_alignment_ldrdstrd(unsigned long addr, unsigned long instr,
386 		      struct pt_regs *regs)
387 {
388 	unsigned int rd = RD_BITS(instr);
389 	unsigned int rd2;
390 	int load;
391 
392 	if ((instr & 0xfe000000) == 0xe8000000) {
393 		/* ARMv7 Thumb-2 32-bit LDRD/STRD */
394 		rd2 = (instr >> 8) & 0xf;
395 		load = !!(LDST_L_BIT(instr));
396 	} else if (((rd & 1) == 1) || (rd == 14))
397 		goto bad;
398 	else {
399 		load = ((instr & 0xf0) == 0xd0);
400 		rd2 = rd + 1;
401 	}
402 
403 	ai_dword += 1;
404 
405 	if (user_mode(regs))
406 		goto user;
407 
408 	if (load) {
409 		unsigned long val;
410 		get32_unaligned_check(val, addr);
411 		regs->uregs[rd] = val;
412 		get32_unaligned_check(val, addr + 4);
413 		regs->uregs[rd2] = val;
414 	} else {
415 		put32_unaligned_check(regs->uregs[rd], addr);
416 		put32_unaligned_check(regs->uregs[rd2], addr + 4);
417 	}
418 
419 	return TYPE_LDST;
420 
421  user:
422 	if (load) {
423 		unsigned long val;
424 		get32t_unaligned_check(val, addr);
425 		regs->uregs[rd] = val;
426 		get32t_unaligned_check(val, addr + 4);
427 		regs->uregs[rd2] = val;
428 	} else {
429 		put32t_unaligned_check(regs->uregs[rd], addr);
430 		put32t_unaligned_check(regs->uregs[rd2], addr + 4);
431 	}
432 
433 	return TYPE_LDST;
434  bad:
435 	return TYPE_ERROR;
436  fault:
437 	return TYPE_FAULT;
438 }
439 
440 static int
441 do_alignment_ldrstr(unsigned long addr, unsigned long instr, struct pt_regs *regs)
442 {
443 	unsigned int rd = RD_BITS(instr);
444 
445 	ai_word += 1;
446 
447 	if ((!LDST_P_BIT(instr) && LDST_W_BIT(instr)) || user_mode(regs))
448 		goto trans;
449 
450 	if (LDST_L_BIT(instr)) {
451 		unsigned int val;
452 		get32_unaligned_check(val, addr);
453 		regs->uregs[rd] = val;
454 	} else
455 		put32_unaligned_check(regs->uregs[rd], addr);
456 	return TYPE_LDST;
457 
458  trans:
459 	if (LDST_L_BIT(instr)) {
460 		unsigned int val;
461 		get32t_unaligned_check(val, addr);
462 		regs->uregs[rd] = val;
463 	} else
464 		put32t_unaligned_check(regs->uregs[rd], addr);
465 	return TYPE_LDST;
466 
467  fault:
468 	return TYPE_FAULT;
469 }
470 
471 /*
472  * LDM/STM alignment handler.
473  *
474  * There are 4 variants of this instruction:
475  *
476  * B = rn pointer before instruction, A = rn pointer after instruction
477  *              ------ increasing address ----->
478  *	        |    | r0 | r1 | ... | rx |    |
479  * PU = 01             B                    A
480  * PU = 11        B                    A
481  * PU = 00        A                    B
482  * PU = 10             A                    B
483  */
484 static int
485 do_alignment_ldmstm(unsigned long addr, unsigned long instr, struct pt_regs *regs)
486 {
487 	unsigned int rd, rn, correction, nr_regs, regbits;
488 	unsigned long eaddr, newaddr;
489 
490 	if (LDM_S_BIT(instr))
491 		goto bad;
492 
493 	correction = 4; /* processor implementation defined */
494 	regs->ARM_pc += correction;
495 
496 	ai_multi += 1;
497 
498 	/* count the number of registers in the mask to be transferred */
499 	nr_regs = hweight16(REGMASK_BITS(instr)) * 4;
500 
501 	rn = RN_BITS(instr);
502 	newaddr = eaddr = regs->uregs[rn];
503 
504 	if (!LDST_U_BIT(instr))
505 		nr_regs = -nr_regs;
506 	newaddr += nr_regs;
507 	if (!LDST_U_BIT(instr))
508 		eaddr = newaddr;
509 
510 	if (LDST_P_EQ_U(instr))	/* U = P */
511 		eaddr += 4;
512 
513 	/*
514 	 * For alignment faults on the ARM922T/ARM920T the MMU  makes
515 	 * the FSR (and hence addr) equal to the updated base address
516 	 * of the multiple access rather than the restored value.
517 	 * Switch this message off if we've got a ARM92[02], otherwise
518 	 * [ls]dm alignment faults are noisy!
519 	 */
520 #if !(defined CONFIG_CPU_ARM922T)  && !(defined CONFIG_CPU_ARM920T)
521 	/*
522 	 * This is a "hint" - we already have eaddr worked out by the
523 	 * processor for us.
524 	 */
525 	if (addr != eaddr) {
526 		pr_err("LDMSTM: PC = %08lx, instr = %08lx, "
527 			"addr = %08lx, eaddr = %08lx\n",
528 			 instruction_pointer(regs), instr, addr, eaddr);
529 		show_regs(regs);
530 	}
531 #endif
532 
533 	if (user_mode(regs)) {
534 		for (regbits = REGMASK_BITS(instr), rd = 0; regbits;
535 		     regbits >>= 1, rd += 1)
536 			if (regbits & 1) {
537 				if (LDST_L_BIT(instr)) {
538 					unsigned int val;
539 					get32t_unaligned_check(val, eaddr);
540 					regs->uregs[rd] = val;
541 				} else
542 					put32t_unaligned_check(regs->uregs[rd], eaddr);
543 				eaddr += 4;
544 			}
545 	} else {
546 		for (regbits = REGMASK_BITS(instr), rd = 0; regbits;
547 		     regbits >>= 1, rd += 1)
548 			if (regbits & 1) {
549 				if (LDST_L_BIT(instr)) {
550 					unsigned int val;
551 					get32_unaligned_check(val, eaddr);
552 					regs->uregs[rd] = val;
553 				} else
554 					put32_unaligned_check(regs->uregs[rd], eaddr);
555 				eaddr += 4;
556 			}
557 	}
558 
559 	if (LDST_W_BIT(instr))
560 		regs->uregs[rn] = newaddr;
561 	if (!LDST_L_BIT(instr) || !(REGMASK_BITS(instr) & (1 << 15)))
562 		regs->ARM_pc -= correction;
563 	return TYPE_DONE;
564 
565 fault:
566 	regs->ARM_pc -= correction;
567 	return TYPE_FAULT;
568 
569 bad:
570 	pr_err("Alignment trap: not handling ldm with s-bit set\n");
571 	return TYPE_ERROR;
572 }
573 
574 /*
575  * Convert Thumb ld/st instruction forms to equivalent ARM instructions so
576  * we can reuse ARM userland alignment fault fixups for Thumb.
577  *
578  * This implementation was initially based on the algorithm found in
579  * gdb/sim/arm/thumbemu.c. It is basically just a code reduction of same
580  * to convert only Thumb ld/st instruction forms to equivalent ARM forms.
581  *
582  * NOTES:
583  * 1. Comments below refer to ARM ARM DDI0100E Thumb Instruction sections.
584  * 2. If for some reason we're passed an non-ld/st Thumb instruction to
585  *    decode, we return 0xdeadc0de. This should never happen under normal
586  *    circumstances but if it does, we've got other problems to deal with
587  *    elsewhere and we obviously can't fix those problems here.
588  */
589 
590 static unsigned long
591 thumb2arm(u16 tinstr)
592 {
593 	u32 L = (tinstr & (1<<11)) >> 11;
594 
595 	switch ((tinstr & 0xf800) >> 11) {
596 	/* 6.5.1 Format 1: */
597 	case 0x6000 >> 11:				/* 7.1.52 STR(1) */
598 	case 0x6800 >> 11:				/* 7.1.26 LDR(1) */
599 	case 0x7000 >> 11:				/* 7.1.55 STRB(1) */
600 	case 0x7800 >> 11:				/* 7.1.30 LDRB(1) */
601 		return 0xe5800000 |
602 			((tinstr & (1<<12)) << (22-12)) |	/* fixup */
603 			(L<<20) |				/* L==1? */
604 			((tinstr & (7<<0)) << (12-0)) |		/* Rd */
605 			((tinstr & (7<<3)) << (16-3)) |		/* Rn */
606 			((tinstr & (31<<6)) >>			/* immed_5 */
607 				(6 - ((tinstr & (1<<12)) ? 0 : 2)));
608 	case 0x8000 >> 11:				/* 7.1.57 STRH(1) */
609 	case 0x8800 >> 11:				/* 7.1.32 LDRH(1) */
610 		return 0xe1c000b0 |
611 			(L<<20) |				/* L==1? */
612 			((tinstr & (7<<0)) << (12-0)) |		/* Rd */
613 			((tinstr & (7<<3)) << (16-3)) |		/* Rn */
614 			((tinstr & (7<<6)) >> (6-1)) |	 /* immed_5[2:0] */
615 			((tinstr & (3<<9)) >> (9-8));	 /* immed_5[4:3] */
616 
617 	/* 6.5.1 Format 2: */
618 	case 0x5000 >> 11:
619 	case 0x5800 >> 11:
620 		{
621 			static const u32 subset[8] = {
622 				0xe7800000,		/* 7.1.53 STR(2) */
623 				0xe18000b0,		/* 7.1.58 STRH(2) */
624 				0xe7c00000,		/* 7.1.56 STRB(2) */
625 				0xe19000d0,		/* 7.1.34 LDRSB */
626 				0xe7900000,		/* 7.1.27 LDR(2) */
627 				0xe19000b0,		/* 7.1.33 LDRH(2) */
628 				0xe7d00000,		/* 7.1.31 LDRB(2) */
629 				0xe19000f0		/* 7.1.35 LDRSH */
630 			};
631 			return subset[(tinstr & (7<<9)) >> 9] |
632 			    ((tinstr & (7<<0)) << (12-0)) |	/* Rd */
633 			    ((tinstr & (7<<3)) << (16-3)) |	/* Rn */
634 			    ((tinstr & (7<<6)) >> (6-0));	/* Rm */
635 		}
636 
637 	/* 6.5.1 Format 3: */
638 	case 0x4800 >> 11:				/* 7.1.28 LDR(3) */
639 		/* NOTE: This case is not technically possible. We're
640 		 *	 loading 32-bit memory data via PC relative
641 		 *	 addressing mode. So we can and should eliminate
642 		 *	 this case. But I'll leave it here for now.
643 		 */
644 		return 0xe59f0000 |
645 		    ((tinstr & (7<<8)) << (12-8)) |		/* Rd */
646 		    ((tinstr & 255) << (2-0));			/* immed_8 */
647 
648 	/* 6.5.1 Format 4: */
649 	case 0x9000 >> 11:				/* 7.1.54 STR(3) */
650 	case 0x9800 >> 11:				/* 7.1.29 LDR(4) */
651 		return 0xe58d0000 |
652 			(L<<20) |				/* L==1? */
653 			((tinstr & (7<<8)) << (12-8)) |		/* Rd */
654 			((tinstr & 255) << 2);			/* immed_8 */
655 
656 	/* 6.6.1 Format 1: */
657 	case 0xc000 >> 11:				/* 7.1.51 STMIA */
658 	case 0xc800 >> 11:				/* 7.1.25 LDMIA */
659 		{
660 			u32 Rn = (tinstr & (7<<8)) >> 8;
661 			u32 W = ((L<<Rn) & (tinstr&255)) ? 0 : 1<<21;
662 
663 			return 0xe8800000 | W | (L<<20) | (Rn<<16) |
664 				(tinstr&255);
665 		}
666 
667 	/* 6.6.1 Format 2: */
668 	case 0xb000 >> 11:				/* 7.1.48 PUSH */
669 	case 0xb800 >> 11:				/* 7.1.47 POP */
670 		if ((tinstr & (3 << 9)) == 0x0400) {
671 			static const u32 subset[4] = {
672 				0xe92d0000,	/* STMDB sp!,{registers} */
673 				0xe92d4000,	/* STMDB sp!,{registers,lr} */
674 				0xe8bd0000,	/* LDMIA sp!,{registers} */
675 				0xe8bd8000	/* LDMIA sp!,{registers,pc} */
676 			};
677 			return subset[(L<<1) | ((tinstr & (1<<8)) >> 8)] |
678 			    (tinstr & 255);		/* register_list */
679 		}
680 		/* Else fall through for illegal instruction case */
681 
682 	default:
683 		return BAD_INSTR;
684 	}
685 }
686 
687 /*
688  * Convert Thumb-2 32 bit LDM, STM, LDRD, STRD to equivalent instruction
689  * handlable by ARM alignment handler, also find the corresponding handler,
690  * so that we can reuse ARM userland alignment fault fixups for Thumb.
691  *
692  * @pinstr: original Thumb-2 instruction; returns new handlable instruction
693  * @regs: register context.
694  * @poffset: return offset from faulted addr for later writeback
695  *
696  * NOTES:
697  * 1. Comments below refer to ARMv7 DDI0406A Thumb Instruction sections.
698  * 2. Register name Rt from ARMv7 is same as Rd from ARMv6 (Rd is Rt)
699  */
700 static void *
701 do_alignment_t32_to_handler(unsigned long *pinstr, struct pt_regs *regs,
702 			    union offset_union *poffset)
703 {
704 	unsigned long instr = *pinstr;
705 	u16 tinst1 = (instr >> 16) & 0xffff;
706 	u16 tinst2 = instr & 0xffff;
707 
708 	switch (tinst1 & 0xffe0) {
709 	/* A6.3.5 Load/Store multiple */
710 	case 0xe880:		/* STM/STMIA/STMEA,LDM/LDMIA, PUSH/POP T2 */
711 	case 0xe8a0:		/* ...above writeback version */
712 	case 0xe900:		/* STMDB/STMFD, LDMDB/LDMEA */
713 	case 0xe920:		/* ...above writeback version */
714 		/* no need offset decision since handler calculates it */
715 		return do_alignment_ldmstm;
716 
717 	case 0xf840:		/* POP/PUSH T3 (single register) */
718 		if (RN_BITS(instr) == 13 && (tinst2 & 0x09ff) == 0x0904) {
719 			u32 L = !!(LDST_L_BIT(instr));
720 			const u32 subset[2] = {
721 				0xe92d0000,	/* STMDB sp!,{registers} */
722 				0xe8bd0000,	/* LDMIA sp!,{registers} */
723 			};
724 			*pinstr = subset[L] | (1<<RD_BITS(instr));
725 			return do_alignment_ldmstm;
726 		}
727 		/* Else fall through for illegal instruction case */
728 		break;
729 
730 	/* A6.3.6 Load/store double, STRD/LDRD(immed, lit, reg) */
731 	case 0xe860:
732 	case 0xe960:
733 	case 0xe8e0:
734 	case 0xe9e0:
735 		poffset->un = (tinst2 & 0xff) << 2;
736 	case 0xe940:
737 	case 0xe9c0:
738 		return do_alignment_ldrdstrd;
739 
740 	/*
741 	 * No need to handle load/store instructions up to word size
742 	 * since ARMv6 and later CPUs can perform unaligned accesses.
743 	 */
744 	default:
745 		break;
746 	}
747 	return NULL;
748 }
749 
750 static int
751 do_alignment(unsigned long addr, unsigned int fsr, struct pt_regs *regs)
752 {
753 	union offset_union uninitialized_var(offset);
754 	unsigned long instr = 0, instrptr;
755 	int (*handler)(unsigned long addr, unsigned long instr, struct pt_regs *regs);
756 	unsigned int type;
757 	unsigned int fault;
758 	u16 tinstr = 0;
759 	int isize = 4;
760 	int thumb2_32b = 0;
761 
762 	if (interrupts_enabled(regs))
763 		local_irq_enable();
764 
765 	instrptr = instruction_pointer(regs);
766 
767 	if (thumb_mode(regs)) {
768 		u16 *ptr = (u16 *)(instrptr & ~1);
769 		fault = probe_kernel_address(ptr, tinstr);
770 		tinstr = __mem_to_opcode_thumb16(tinstr);
771 		if (!fault) {
772 			if (cpu_architecture() >= CPU_ARCH_ARMv7 &&
773 			    IS_T32(tinstr)) {
774 				/* Thumb-2 32-bit */
775 				u16 tinst2 = 0;
776 				fault = probe_kernel_address(ptr + 1, tinst2);
777 				tinst2 = __mem_to_opcode_thumb16(tinst2);
778 				instr = __opcode_thumb32_compose(tinstr, tinst2);
779 				thumb2_32b = 1;
780 			} else {
781 				isize = 2;
782 				instr = thumb2arm(tinstr);
783 			}
784 		}
785 	} else {
786 		fault = probe_kernel_address(instrptr, instr);
787 		instr = __mem_to_opcode_arm(instr);
788 	}
789 
790 	if (fault) {
791 		type = TYPE_FAULT;
792 		goto bad_or_fault;
793 	}
794 
795 	if (user_mode(regs))
796 		goto user;
797 
798 	ai_sys += 1;
799 	ai_sys_last_pc = (void *)instruction_pointer(regs);
800 
801  fixup:
802 
803 	regs->ARM_pc += isize;
804 
805 	switch (CODING_BITS(instr)) {
806 	case 0x00000000:	/* 3.13.4 load/store instruction extensions */
807 		if (LDSTHD_I_BIT(instr))
808 			offset.un = (instr & 0xf00) >> 4 | (instr & 15);
809 		else
810 			offset.un = regs->uregs[RM_BITS(instr)];
811 
812 		if ((instr & 0x000000f0) == 0x000000b0 || /* LDRH, STRH */
813 		    (instr & 0x001000f0) == 0x001000f0)   /* LDRSH */
814 			handler = do_alignment_ldrhstrh;
815 		else if ((instr & 0x001000f0) == 0x000000d0 || /* LDRD */
816 			 (instr & 0x001000f0) == 0x000000f0)   /* STRD */
817 			handler = do_alignment_ldrdstrd;
818 		else if ((instr & 0x01f00ff0) == 0x01000090) /* SWP */
819 			goto swp;
820 		else
821 			goto bad;
822 		break;
823 
824 	case 0x04000000:	/* ldr or str immediate */
825 		if (COND_BITS(instr) == 0xf0000000) /* NEON VLDn, VSTn */
826 			goto bad;
827 		offset.un = OFFSET_BITS(instr);
828 		handler = do_alignment_ldrstr;
829 		break;
830 
831 	case 0x06000000:	/* ldr or str register */
832 		offset.un = regs->uregs[RM_BITS(instr)];
833 
834 		if (IS_SHIFT(instr)) {
835 			unsigned int shiftval = SHIFT_BITS(instr);
836 
837 			switch(SHIFT_TYPE(instr)) {
838 			case SHIFT_LSL:
839 				offset.un <<= shiftval;
840 				break;
841 
842 			case SHIFT_LSR:
843 				offset.un >>= shiftval;
844 				break;
845 
846 			case SHIFT_ASR:
847 				offset.sn >>= shiftval;
848 				break;
849 
850 			case SHIFT_RORRRX:
851 				if (shiftval == 0) {
852 					offset.un >>= 1;
853 					if (regs->ARM_cpsr & PSR_C_BIT)
854 						offset.un |= 1 << 31;
855 				} else
856 					offset.un = offset.un >> shiftval |
857 							  offset.un << (32 - shiftval);
858 				break;
859 			}
860 		}
861 		handler = do_alignment_ldrstr;
862 		break;
863 
864 	case 0x08000000:	/* ldm or stm, or thumb-2 32bit instruction */
865 		if (thumb2_32b) {
866 			offset.un = 0;
867 			handler = do_alignment_t32_to_handler(&instr, regs, &offset);
868 		} else {
869 			offset.un = 0;
870 			handler = do_alignment_ldmstm;
871 		}
872 		break;
873 
874 	default:
875 		goto bad;
876 	}
877 
878 	if (!handler)
879 		goto bad;
880 	type = handler(addr, instr, regs);
881 
882 	if (type == TYPE_ERROR || type == TYPE_FAULT) {
883 		regs->ARM_pc -= isize;
884 		goto bad_or_fault;
885 	}
886 
887 	if (type == TYPE_LDST)
888 		do_alignment_finish_ldst(addr, instr, regs, offset);
889 
890 	return 0;
891 
892  bad_or_fault:
893 	if (type == TYPE_ERROR)
894 		goto bad;
895 	/*
896 	 * We got a fault - fix it up, or die.
897 	 */
898 	do_bad_area(addr, fsr, regs);
899 	return 0;
900 
901  swp:
902 	pr_err("Alignment trap: not handling swp instruction\n");
903 
904  bad:
905 	/*
906 	 * Oops, we didn't handle the instruction.
907 	 */
908 	pr_err("Alignment trap: not handling instruction "
909 		"%0*lx at [<%08lx>]\n",
910 		isize << 1,
911 		isize == 2 ? tinstr : instr, instrptr);
912 	ai_skipped += 1;
913 	return 1;
914 
915  user:
916 	ai_user += 1;
917 
918 	if (ai_usermode & UM_WARN)
919 		printk("Alignment trap: %s (%d) PC=0x%08lx Instr=0x%0*lx "
920 		       "Address=0x%08lx FSR 0x%03x\n", current->comm,
921 			task_pid_nr(current), instrptr,
922 			isize << 1,
923 			isize == 2 ? tinstr : instr,
924 		        addr, fsr);
925 
926 	if (ai_usermode & UM_FIXUP)
927 		goto fixup;
928 
929 	if (ai_usermode & UM_SIGNAL) {
930 		siginfo_t si;
931 
932 		si.si_signo = SIGBUS;
933 		si.si_errno = 0;
934 		si.si_code = BUS_ADRALN;
935 		si.si_addr = (void __user *)addr;
936 
937 		force_sig_info(si.si_signo, &si, current);
938 	} else {
939 		/*
940 		 * We're about to disable the alignment trap and return to
941 		 * user space.  But if an interrupt occurs before actually
942 		 * reaching user space, then the IRQ vector entry code will
943 		 * notice that we were still in kernel space and therefore
944 		 * the alignment trap won't be re-enabled in that case as it
945 		 * is presumed to be always on from kernel space.
946 		 * Let's prevent that race by disabling interrupts here (they
947 		 * are disabled on the way back to user space anyway in
948 		 * entry-common.S) and disable the alignment trap only if
949 		 * there is no work pending for this thread.
950 		 */
951 		raw_local_irq_disable();
952 		if (!(current_thread_info()->flags & _TIF_WORK_MASK))
953 			set_cr(cr_no_alignment);
954 	}
955 
956 	return 0;
957 }
958 
959 static int __init noalign_setup(char *__unused)
960 {
961 	set_cr(__clear_cr(CR_A));
962 	return 1;
963 }
964 __setup("noalign", noalign_setup);
965 
966 /*
967  * This needs to be done after sysctl_init, otherwise sys/ will be
968  * overwritten.  Actually, this shouldn't be in sys/ at all since
969  * it isn't a sysctl, and it doesn't contain sysctl information.
970  * We now locate it in /proc/cpu/alignment instead.
971  */
972 static int __init alignment_init(void)
973 {
974 #ifdef CONFIG_PROC_FS
975 	struct proc_dir_entry *res;
976 
977 	res = proc_create("cpu/alignment", S_IWUSR | S_IRUGO, NULL,
978 			  &alignment_proc_fops);
979 	if (!res)
980 		return -ENOMEM;
981 #endif
982 
983 	if (cpu_is_v6_unaligned()) {
984 		set_cr(__clear_cr(CR_A));
985 		ai_usermode = safe_usermode(ai_usermode, false);
986 	}
987 
988 	cr_no_alignment = get_cr() & ~CR_A;
989 
990 	hook_fault_code(FAULT_CODE_ALIGNMENT, do_alignment, SIGBUS, BUS_ADRALN,
991 			"alignment exception");
992 
993 	/*
994 	 * ARMv6K and ARMv7 use fault status 3 (0b00011) as Access Flag section
995 	 * fault, not as alignment error.
996 	 *
997 	 * TODO: handle ARMv6K properly. Runtime check for 'K' extension is
998 	 * needed.
999 	 */
1000 	if (cpu_architecture() <= CPU_ARCH_ARMv6) {
1001 		hook_fault_code(3, do_alignment, SIGBUS, BUS_ADRALN,
1002 				"alignment exception");
1003 	}
1004 
1005 	return 0;
1006 }
1007 
1008 fs_initcall(alignment_init);
1009