1 /*
2  * HW_breakpoint: a unified kernel/user-space hardware breakpoint facility,
3  * using the CPU's debug registers.
4  *
5  * Copyright (C) 2012 ARM Limited
6  * Author: Will Deacon <will.deacon@arm.com>
7  *
8  * This program is free software; you can redistribute it and/or modify
9  * it under the terms of the GNU General Public License version 2 as
10  * published by the Free Software Foundation.
11  *
12  * This program is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
15  * GNU General Public License for more details.
16  *
17  * You should have received a copy of the GNU General Public License
18  * along with this program.  If not, see <http://www.gnu.org/licenses/>.
19  */
20 
21 #define pr_fmt(fmt) "hw-breakpoint: " fmt
22 
23 #include <linux/compat.h>
24 #include <linux/cpu_pm.h>
25 #include <linux/errno.h>
26 #include <linux/hw_breakpoint.h>
27 #include <linux/perf_event.h>
28 #include <linux/ptrace.h>
29 #include <linux/smp.h>
30 
31 #include <asm/compat.h>
32 #include <asm/current.h>
33 #include <asm/debug-monitors.h>
34 #include <asm/hw_breakpoint.h>
35 #include <asm/traps.h>
36 #include <asm/cputype.h>
37 #include <asm/system_misc.h>
38 
39 /* Breakpoint currently in use for each BRP. */
40 static DEFINE_PER_CPU(struct perf_event *, bp_on_reg[ARM_MAX_BRP]);
41 
42 /* Watchpoint currently in use for each WRP. */
43 static DEFINE_PER_CPU(struct perf_event *, wp_on_reg[ARM_MAX_WRP]);
44 
45 /* Currently stepping a per-CPU kernel breakpoint. */
46 static DEFINE_PER_CPU(int, stepping_kernel_bp);
47 
48 /* Number of BRP/WRP registers on this CPU. */
49 static int core_num_brps;
50 static int core_num_wrps;
51 
52 int hw_breakpoint_slots(int type)
53 {
54 	/*
55 	 * We can be called early, so don't rely on
56 	 * our static variables being initialised.
57 	 */
58 	switch (type) {
59 	case TYPE_INST:
60 		return get_num_brps();
61 	case TYPE_DATA:
62 		return get_num_wrps();
63 	default:
64 		pr_warning("unknown slot type: %d\n", type);
65 		return 0;
66 	}
67 }
68 
69 #define READ_WB_REG_CASE(OFF, N, REG, VAL)	\
70 	case (OFF + N):				\
71 		AARCH64_DBG_READ(N, REG, VAL);	\
72 		break
73 
74 #define WRITE_WB_REG_CASE(OFF, N, REG, VAL)	\
75 	case (OFF + N):				\
76 		AARCH64_DBG_WRITE(N, REG, VAL);	\
77 		break
78 
79 #define GEN_READ_WB_REG_CASES(OFF, REG, VAL)	\
80 	READ_WB_REG_CASE(OFF,  0, REG, VAL);	\
81 	READ_WB_REG_CASE(OFF,  1, REG, VAL);	\
82 	READ_WB_REG_CASE(OFF,  2, REG, VAL);	\
83 	READ_WB_REG_CASE(OFF,  3, REG, VAL);	\
84 	READ_WB_REG_CASE(OFF,  4, REG, VAL);	\
85 	READ_WB_REG_CASE(OFF,  5, REG, VAL);	\
86 	READ_WB_REG_CASE(OFF,  6, REG, VAL);	\
87 	READ_WB_REG_CASE(OFF,  7, REG, VAL);	\
88 	READ_WB_REG_CASE(OFF,  8, REG, VAL);	\
89 	READ_WB_REG_CASE(OFF,  9, REG, VAL);	\
90 	READ_WB_REG_CASE(OFF, 10, REG, VAL);	\
91 	READ_WB_REG_CASE(OFF, 11, REG, VAL);	\
92 	READ_WB_REG_CASE(OFF, 12, REG, VAL);	\
93 	READ_WB_REG_CASE(OFF, 13, REG, VAL);	\
94 	READ_WB_REG_CASE(OFF, 14, REG, VAL);	\
95 	READ_WB_REG_CASE(OFF, 15, REG, VAL)
96 
97 #define GEN_WRITE_WB_REG_CASES(OFF, REG, VAL)	\
98 	WRITE_WB_REG_CASE(OFF,  0, REG, VAL);	\
99 	WRITE_WB_REG_CASE(OFF,  1, REG, VAL);	\
100 	WRITE_WB_REG_CASE(OFF,  2, REG, VAL);	\
101 	WRITE_WB_REG_CASE(OFF,  3, REG, VAL);	\
102 	WRITE_WB_REG_CASE(OFF,  4, REG, VAL);	\
103 	WRITE_WB_REG_CASE(OFF,  5, REG, VAL);	\
104 	WRITE_WB_REG_CASE(OFF,  6, REG, VAL);	\
105 	WRITE_WB_REG_CASE(OFF,  7, REG, VAL);	\
106 	WRITE_WB_REG_CASE(OFF,  8, REG, VAL);	\
107 	WRITE_WB_REG_CASE(OFF,  9, REG, VAL);	\
108 	WRITE_WB_REG_CASE(OFF, 10, REG, VAL);	\
109 	WRITE_WB_REG_CASE(OFF, 11, REG, VAL);	\
110 	WRITE_WB_REG_CASE(OFF, 12, REG, VAL);	\
111 	WRITE_WB_REG_CASE(OFF, 13, REG, VAL);	\
112 	WRITE_WB_REG_CASE(OFF, 14, REG, VAL);	\
113 	WRITE_WB_REG_CASE(OFF, 15, REG, VAL)
114 
115 static u64 read_wb_reg(int reg, int n)
116 {
117 	u64 val = 0;
118 
119 	switch (reg + n) {
120 	GEN_READ_WB_REG_CASES(AARCH64_DBG_REG_BVR, AARCH64_DBG_REG_NAME_BVR, val);
121 	GEN_READ_WB_REG_CASES(AARCH64_DBG_REG_BCR, AARCH64_DBG_REG_NAME_BCR, val);
122 	GEN_READ_WB_REG_CASES(AARCH64_DBG_REG_WVR, AARCH64_DBG_REG_NAME_WVR, val);
123 	GEN_READ_WB_REG_CASES(AARCH64_DBG_REG_WCR, AARCH64_DBG_REG_NAME_WCR, val);
124 	default:
125 		pr_warning("attempt to read from unknown breakpoint register %d\n", n);
126 	}
127 
128 	return val;
129 }
130 
131 static void write_wb_reg(int reg, int n, u64 val)
132 {
133 	switch (reg + n) {
134 	GEN_WRITE_WB_REG_CASES(AARCH64_DBG_REG_BVR, AARCH64_DBG_REG_NAME_BVR, val);
135 	GEN_WRITE_WB_REG_CASES(AARCH64_DBG_REG_BCR, AARCH64_DBG_REG_NAME_BCR, val);
136 	GEN_WRITE_WB_REG_CASES(AARCH64_DBG_REG_WVR, AARCH64_DBG_REG_NAME_WVR, val);
137 	GEN_WRITE_WB_REG_CASES(AARCH64_DBG_REG_WCR, AARCH64_DBG_REG_NAME_WCR, val);
138 	default:
139 		pr_warning("attempt to write to unknown breakpoint register %d\n", n);
140 	}
141 	isb();
142 }
143 
144 /*
145  * Convert a breakpoint privilege level to the corresponding exception
146  * level.
147  */
148 static enum dbg_active_el debug_exception_level(int privilege)
149 {
150 	switch (privilege) {
151 	case AARCH64_BREAKPOINT_EL0:
152 		return DBG_ACTIVE_EL0;
153 	case AARCH64_BREAKPOINT_EL1:
154 		return DBG_ACTIVE_EL1;
155 	default:
156 		pr_warning("invalid breakpoint privilege level %d\n", privilege);
157 		return -EINVAL;
158 	}
159 }
160 
161 enum hw_breakpoint_ops {
162 	HW_BREAKPOINT_INSTALL,
163 	HW_BREAKPOINT_UNINSTALL,
164 	HW_BREAKPOINT_RESTORE
165 };
166 
167 static int is_compat_bp(struct perf_event *bp)
168 {
169 	struct task_struct *tsk = bp->hw.target;
170 
171 	/*
172 	 * tsk can be NULL for per-cpu (non-ptrace) breakpoints.
173 	 * In this case, use the native interface, since we don't have
174 	 * the notion of a "compat CPU" and could end up relying on
175 	 * deprecated behaviour if we use unaligned watchpoints in
176 	 * AArch64 state.
177 	 */
178 	return tsk && is_compat_thread(task_thread_info(tsk));
179 }
180 
181 /**
182  * hw_breakpoint_slot_setup - Find and setup a perf slot according to
183  *			      operations
184  *
185  * @slots: pointer to array of slots
186  * @max_slots: max number of slots
187  * @bp: perf_event to setup
188  * @ops: operation to be carried out on the slot
189  *
190  * Return:
191  *	slot index on success
192  *	-ENOSPC if no slot is available/matches
193  *	-EINVAL on wrong operations parameter
194  */
195 static int hw_breakpoint_slot_setup(struct perf_event **slots, int max_slots,
196 				    struct perf_event *bp,
197 				    enum hw_breakpoint_ops ops)
198 {
199 	int i;
200 	struct perf_event **slot;
201 
202 	for (i = 0; i < max_slots; ++i) {
203 		slot = &slots[i];
204 		switch (ops) {
205 		case HW_BREAKPOINT_INSTALL:
206 			if (!*slot) {
207 				*slot = bp;
208 				return i;
209 			}
210 			break;
211 		case HW_BREAKPOINT_UNINSTALL:
212 			if (*slot == bp) {
213 				*slot = NULL;
214 				return i;
215 			}
216 			break;
217 		case HW_BREAKPOINT_RESTORE:
218 			if (*slot == bp)
219 				return i;
220 			break;
221 		default:
222 			pr_warn_once("Unhandled hw breakpoint ops %d\n", ops);
223 			return -EINVAL;
224 		}
225 	}
226 	return -ENOSPC;
227 }
228 
229 static int hw_breakpoint_control(struct perf_event *bp,
230 				 enum hw_breakpoint_ops ops)
231 {
232 	struct arch_hw_breakpoint *info = counter_arch_bp(bp);
233 	struct perf_event **slots;
234 	struct debug_info *debug_info = &current->thread.debug;
235 	int i, max_slots, ctrl_reg, val_reg, reg_enable;
236 	enum dbg_active_el dbg_el = debug_exception_level(info->ctrl.privilege);
237 	u32 ctrl;
238 
239 	if (info->ctrl.type == ARM_BREAKPOINT_EXECUTE) {
240 		/* Breakpoint */
241 		ctrl_reg = AARCH64_DBG_REG_BCR;
242 		val_reg = AARCH64_DBG_REG_BVR;
243 		slots = this_cpu_ptr(bp_on_reg);
244 		max_slots = core_num_brps;
245 		reg_enable = !debug_info->bps_disabled;
246 	} else {
247 		/* Watchpoint */
248 		ctrl_reg = AARCH64_DBG_REG_WCR;
249 		val_reg = AARCH64_DBG_REG_WVR;
250 		slots = this_cpu_ptr(wp_on_reg);
251 		max_slots = core_num_wrps;
252 		reg_enable = !debug_info->wps_disabled;
253 	}
254 
255 	i = hw_breakpoint_slot_setup(slots, max_slots, bp, ops);
256 
257 	if (WARN_ONCE(i < 0, "Can't find any breakpoint slot"))
258 		return i;
259 
260 	switch (ops) {
261 	case HW_BREAKPOINT_INSTALL:
262 		/*
263 		 * Ensure debug monitors are enabled at the correct exception
264 		 * level.
265 		 */
266 		enable_debug_monitors(dbg_el);
267 		/* Fall through */
268 	case HW_BREAKPOINT_RESTORE:
269 		/* Setup the address register. */
270 		write_wb_reg(val_reg, i, info->address);
271 
272 		/* Setup the control register. */
273 		ctrl = encode_ctrl_reg(info->ctrl);
274 		write_wb_reg(ctrl_reg, i,
275 			     reg_enable ? ctrl | 0x1 : ctrl & ~0x1);
276 		break;
277 	case HW_BREAKPOINT_UNINSTALL:
278 		/* Reset the control register. */
279 		write_wb_reg(ctrl_reg, i, 0);
280 
281 		/*
282 		 * Release the debug monitors for the correct exception
283 		 * level.
284 		 */
285 		disable_debug_monitors(dbg_el);
286 		break;
287 	}
288 
289 	return 0;
290 }
291 
292 /*
293  * Install a perf counter breakpoint.
294  */
295 int arch_install_hw_breakpoint(struct perf_event *bp)
296 {
297 	return hw_breakpoint_control(bp, HW_BREAKPOINT_INSTALL);
298 }
299 
300 void arch_uninstall_hw_breakpoint(struct perf_event *bp)
301 {
302 	hw_breakpoint_control(bp, HW_BREAKPOINT_UNINSTALL);
303 }
304 
305 static int get_hbp_len(u8 hbp_len)
306 {
307 	unsigned int len_in_bytes = 0;
308 
309 	switch (hbp_len) {
310 	case ARM_BREAKPOINT_LEN_1:
311 		len_in_bytes = 1;
312 		break;
313 	case ARM_BREAKPOINT_LEN_2:
314 		len_in_bytes = 2;
315 		break;
316 	case ARM_BREAKPOINT_LEN_4:
317 		len_in_bytes = 4;
318 		break;
319 	case ARM_BREAKPOINT_LEN_8:
320 		len_in_bytes = 8;
321 		break;
322 	}
323 
324 	return len_in_bytes;
325 }
326 
327 /*
328  * Check whether bp virtual address is in kernel space.
329  */
330 int arch_check_bp_in_kernelspace(struct perf_event *bp)
331 {
332 	unsigned int len;
333 	unsigned long va;
334 	struct arch_hw_breakpoint *info = counter_arch_bp(bp);
335 
336 	va = info->address;
337 	len = get_hbp_len(info->ctrl.len);
338 
339 	return (va >= TASK_SIZE) && ((va + len - 1) >= TASK_SIZE);
340 }
341 
342 /*
343  * Extract generic type and length encodings from an arch_hw_breakpoint_ctrl.
344  * Hopefully this will disappear when ptrace can bypass the conversion
345  * to generic breakpoint descriptions.
346  */
347 int arch_bp_generic_fields(struct arch_hw_breakpoint_ctrl ctrl,
348 			   int *gen_len, int *gen_type)
349 {
350 	/* Type */
351 	switch (ctrl.type) {
352 	case ARM_BREAKPOINT_EXECUTE:
353 		*gen_type = HW_BREAKPOINT_X;
354 		break;
355 	case ARM_BREAKPOINT_LOAD:
356 		*gen_type = HW_BREAKPOINT_R;
357 		break;
358 	case ARM_BREAKPOINT_STORE:
359 		*gen_type = HW_BREAKPOINT_W;
360 		break;
361 	case ARM_BREAKPOINT_LOAD | ARM_BREAKPOINT_STORE:
362 		*gen_type = HW_BREAKPOINT_RW;
363 		break;
364 	default:
365 		return -EINVAL;
366 	}
367 
368 	/* Len */
369 	switch (ctrl.len) {
370 	case ARM_BREAKPOINT_LEN_1:
371 		*gen_len = HW_BREAKPOINT_LEN_1;
372 		break;
373 	case ARM_BREAKPOINT_LEN_2:
374 		*gen_len = HW_BREAKPOINT_LEN_2;
375 		break;
376 	case ARM_BREAKPOINT_LEN_4:
377 		*gen_len = HW_BREAKPOINT_LEN_4;
378 		break;
379 	case ARM_BREAKPOINT_LEN_8:
380 		*gen_len = HW_BREAKPOINT_LEN_8;
381 		break;
382 	default:
383 		return -EINVAL;
384 	}
385 
386 	return 0;
387 }
388 
389 /*
390  * Construct an arch_hw_breakpoint from a perf_event.
391  */
392 static int arch_build_bp_info(struct perf_event *bp)
393 {
394 	struct arch_hw_breakpoint *info = counter_arch_bp(bp);
395 
396 	/* Type */
397 	switch (bp->attr.bp_type) {
398 	case HW_BREAKPOINT_X:
399 		info->ctrl.type = ARM_BREAKPOINT_EXECUTE;
400 		break;
401 	case HW_BREAKPOINT_R:
402 		info->ctrl.type = ARM_BREAKPOINT_LOAD;
403 		break;
404 	case HW_BREAKPOINT_W:
405 		info->ctrl.type = ARM_BREAKPOINT_STORE;
406 		break;
407 	case HW_BREAKPOINT_RW:
408 		info->ctrl.type = ARM_BREAKPOINT_LOAD | ARM_BREAKPOINT_STORE;
409 		break;
410 	default:
411 		return -EINVAL;
412 	}
413 
414 	/* Len */
415 	switch (bp->attr.bp_len) {
416 	case HW_BREAKPOINT_LEN_1:
417 		info->ctrl.len = ARM_BREAKPOINT_LEN_1;
418 		break;
419 	case HW_BREAKPOINT_LEN_2:
420 		info->ctrl.len = ARM_BREAKPOINT_LEN_2;
421 		break;
422 	case HW_BREAKPOINT_LEN_4:
423 		info->ctrl.len = ARM_BREAKPOINT_LEN_4;
424 		break;
425 	case HW_BREAKPOINT_LEN_8:
426 		info->ctrl.len = ARM_BREAKPOINT_LEN_8;
427 		break;
428 	default:
429 		return -EINVAL;
430 	}
431 
432 	/*
433 	 * On AArch64, we only permit breakpoints of length 4, whereas
434 	 * AArch32 also requires breakpoints of length 2 for Thumb.
435 	 * Watchpoints can be of length 1, 2, 4 or 8 bytes.
436 	 */
437 	if (info->ctrl.type == ARM_BREAKPOINT_EXECUTE) {
438 		if (is_compat_bp(bp)) {
439 			if (info->ctrl.len != ARM_BREAKPOINT_LEN_2 &&
440 			    info->ctrl.len != ARM_BREAKPOINT_LEN_4)
441 				return -EINVAL;
442 		} else if (info->ctrl.len != ARM_BREAKPOINT_LEN_4) {
443 			/*
444 			 * FIXME: Some tools (I'm looking at you perf) assume
445 			 *	  that breakpoints should be sizeof(long). This
446 			 *	  is nonsense. For now, we fix up the parameter
447 			 *	  but we should probably return -EINVAL instead.
448 			 */
449 			info->ctrl.len = ARM_BREAKPOINT_LEN_4;
450 		}
451 	}
452 
453 	/* Address */
454 	info->address = bp->attr.bp_addr;
455 
456 	/*
457 	 * Privilege
458 	 * Note that we disallow combined EL0/EL1 breakpoints because
459 	 * that would complicate the stepping code.
460 	 */
461 	if (arch_check_bp_in_kernelspace(bp))
462 		info->ctrl.privilege = AARCH64_BREAKPOINT_EL1;
463 	else
464 		info->ctrl.privilege = AARCH64_BREAKPOINT_EL0;
465 
466 	/* Enabled? */
467 	info->ctrl.enabled = !bp->attr.disabled;
468 
469 	return 0;
470 }
471 
472 /*
473  * Validate the arch-specific HW Breakpoint register settings.
474  */
475 int arch_validate_hwbkpt_settings(struct perf_event *bp)
476 {
477 	struct arch_hw_breakpoint *info = counter_arch_bp(bp);
478 	int ret;
479 	u64 alignment_mask, offset;
480 
481 	/* Build the arch_hw_breakpoint. */
482 	ret = arch_build_bp_info(bp);
483 	if (ret)
484 		return ret;
485 
486 	/*
487 	 * Check address alignment.
488 	 * We don't do any clever alignment correction for watchpoints
489 	 * because using 64-bit unaligned addresses is deprecated for
490 	 * AArch64.
491 	 *
492 	 * AArch32 tasks expect some simple alignment fixups, so emulate
493 	 * that here.
494 	 */
495 	if (is_compat_bp(bp)) {
496 		if (info->ctrl.len == ARM_BREAKPOINT_LEN_8)
497 			alignment_mask = 0x7;
498 		else
499 			alignment_mask = 0x3;
500 		offset = info->address & alignment_mask;
501 		switch (offset) {
502 		case 0:
503 			/* Aligned */
504 			break;
505 		case 1:
506 			/* Allow single byte watchpoint. */
507 			if (info->ctrl.len == ARM_BREAKPOINT_LEN_1)
508 				break;
509 		case 2:
510 			/* Allow halfword watchpoints and breakpoints. */
511 			if (info->ctrl.len == ARM_BREAKPOINT_LEN_2)
512 				break;
513 		default:
514 			return -EINVAL;
515 		}
516 
517 		info->address &= ~alignment_mask;
518 		info->ctrl.len <<= offset;
519 	} else {
520 		if (info->ctrl.type == ARM_BREAKPOINT_EXECUTE)
521 			alignment_mask = 0x3;
522 		else
523 			alignment_mask = 0x7;
524 		if (info->address & alignment_mask)
525 			return -EINVAL;
526 	}
527 
528 	/*
529 	 * Disallow per-task kernel breakpoints since these would
530 	 * complicate the stepping code.
531 	 */
532 	if (info->ctrl.privilege == AARCH64_BREAKPOINT_EL1 && bp->hw.target)
533 		return -EINVAL;
534 
535 	return 0;
536 }
537 
538 /*
539  * Enable/disable all of the breakpoints active at the specified
540  * exception level at the register level.
541  * This is used when single-stepping after a breakpoint exception.
542  */
543 static void toggle_bp_registers(int reg, enum dbg_active_el el, int enable)
544 {
545 	int i, max_slots, privilege;
546 	u32 ctrl;
547 	struct perf_event **slots;
548 
549 	switch (reg) {
550 	case AARCH64_DBG_REG_BCR:
551 		slots = this_cpu_ptr(bp_on_reg);
552 		max_slots = core_num_brps;
553 		break;
554 	case AARCH64_DBG_REG_WCR:
555 		slots = this_cpu_ptr(wp_on_reg);
556 		max_slots = core_num_wrps;
557 		break;
558 	default:
559 		return;
560 	}
561 
562 	for (i = 0; i < max_slots; ++i) {
563 		if (!slots[i])
564 			continue;
565 
566 		privilege = counter_arch_bp(slots[i])->ctrl.privilege;
567 		if (debug_exception_level(privilege) != el)
568 			continue;
569 
570 		ctrl = read_wb_reg(reg, i);
571 		if (enable)
572 			ctrl |= 0x1;
573 		else
574 			ctrl &= ~0x1;
575 		write_wb_reg(reg, i, ctrl);
576 	}
577 }
578 
579 /*
580  * Debug exception handlers.
581  */
582 static int breakpoint_handler(unsigned long unused, unsigned int esr,
583 			      struct pt_regs *regs)
584 {
585 	int i, step = 0, *kernel_step;
586 	u32 ctrl_reg;
587 	u64 addr, val;
588 	struct perf_event *bp, **slots;
589 	struct debug_info *debug_info;
590 	struct arch_hw_breakpoint_ctrl ctrl;
591 
592 	slots = this_cpu_ptr(bp_on_reg);
593 	addr = instruction_pointer(regs);
594 	debug_info = &current->thread.debug;
595 
596 	for (i = 0; i < core_num_brps; ++i) {
597 		rcu_read_lock();
598 
599 		bp = slots[i];
600 
601 		if (bp == NULL)
602 			goto unlock;
603 
604 		/* Check if the breakpoint value matches. */
605 		val = read_wb_reg(AARCH64_DBG_REG_BVR, i);
606 		if (val != (addr & ~0x3))
607 			goto unlock;
608 
609 		/* Possible match, check the byte address select to confirm. */
610 		ctrl_reg = read_wb_reg(AARCH64_DBG_REG_BCR, i);
611 		decode_ctrl_reg(ctrl_reg, &ctrl);
612 		if (!((1 << (addr & 0x3)) & ctrl.len))
613 			goto unlock;
614 
615 		counter_arch_bp(bp)->trigger = addr;
616 		perf_bp_event(bp, regs);
617 
618 		/* Do we need to handle the stepping? */
619 		if (!bp->overflow_handler)
620 			step = 1;
621 unlock:
622 		rcu_read_unlock();
623 	}
624 
625 	if (!step)
626 		return 0;
627 
628 	if (user_mode(regs)) {
629 		debug_info->bps_disabled = 1;
630 		toggle_bp_registers(AARCH64_DBG_REG_BCR, DBG_ACTIVE_EL0, 0);
631 
632 		/* If we're already stepping a watchpoint, just return. */
633 		if (debug_info->wps_disabled)
634 			return 0;
635 
636 		if (test_thread_flag(TIF_SINGLESTEP))
637 			debug_info->suspended_step = 1;
638 		else
639 			user_enable_single_step(current);
640 	} else {
641 		toggle_bp_registers(AARCH64_DBG_REG_BCR, DBG_ACTIVE_EL1, 0);
642 		kernel_step = this_cpu_ptr(&stepping_kernel_bp);
643 
644 		if (*kernel_step != ARM_KERNEL_STEP_NONE)
645 			return 0;
646 
647 		if (kernel_active_single_step()) {
648 			*kernel_step = ARM_KERNEL_STEP_SUSPEND;
649 		} else {
650 			*kernel_step = ARM_KERNEL_STEP_ACTIVE;
651 			kernel_enable_single_step(regs);
652 		}
653 	}
654 
655 	return 0;
656 }
657 
658 static int watchpoint_handler(unsigned long addr, unsigned int esr,
659 			      struct pt_regs *regs)
660 {
661 	int i, step = 0, *kernel_step, access;
662 	u32 ctrl_reg;
663 	u64 val, alignment_mask;
664 	struct perf_event *wp, **slots;
665 	struct debug_info *debug_info;
666 	struct arch_hw_breakpoint *info;
667 	struct arch_hw_breakpoint_ctrl ctrl;
668 
669 	slots = this_cpu_ptr(wp_on_reg);
670 	debug_info = &current->thread.debug;
671 
672 	for (i = 0; i < core_num_wrps; ++i) {
673 		rcu_read_lock();
674 
675 		wp = slots[i];
676 
677 		if (wp == NULL)
678 			goto unlock;
679 
680 		info = counter_arch_bp(wp);
681 		/* AArch32 watchpoints are either 4 or 8 bytes aligned. */
682 		if (is_compat_task()) {
683 			if (info->ctrl.len == ARM_BREAKPOINT_LEN_8)
684 				alignment_mask = 0x7;
685 			else
686 				alignment_mask = 0x3;
687 		} else {
688 			alignment_mask = 0x7;
689 		}
690 
691 		/* Check if the watchpoint value matches. */
692 		val = read_wb_reg(AARCH64_DBG_REG_WVR, i);
693 		if (val != (addr & ~alignment_mask))
694 			goto unlock;
695 
696 		/* Possible match, check the byte address select to confirm. */
697 		ctrl_reg = read_wb_reg(AARCH64_DBG_REG_WCR, i);
698 		decode_ctrl_reg(ctrl_reg, &ctrl);
699 		if (!((1 << (addr & alignment_mask)) & ctrl.len))
700 			goto unlock;
701 
702 		/*
703 		 * Check that the access type matches.
704 		 * 0 => load, otherwise => store
705 		 */
706 		access = (esr & AARCH64_ESR_ACCESS_MASK) ? HW_BREAKPOINT_W :
707 			 HW_BREAKPOINT_R;
708 		if (!(access & hw_breakpoint_type(wp)))
709 			goto unlock;
710 
711 		info->trigger = addr;
712 		perf_bp_event(wp, regs);
713 
714 		/* Do we need to handle the stepping? */
715 		if (!wp->overflow_handler)
716 			step = 1;
717 
718 unlock:
719 		rcu_read_unlock();
720 	}
721 
722 	if (!step)
723 		return 0;
724 
725 	/*
726 	 * We always disable EL0 watchpoints because the kernel can
727 	 * cause these to fire via an unprivileged access.
728 	 */
729 	toggle_bp_registers(AARCH64_DBG_REG_WCR, DBG_ACTIVE_EL0, 0);
730 
731 	if (user_mode(regs)) {
732 		debug_info->wps_disabled = 1;
733 
734 		/* If we're already stepping a breakpoint, just return. */
735 		if (debug_info->bps_disabled)
736 			return 0;
737 
738 		if (test_thread_flag(TIF_SINGLESTEP))
739 			debug_info->suspended_step = 1;
740 		else
741 			user_enable_single_step(current);
742 	} else {
743 		toggle_bp_registers(AARCH64_DBG_REG_WCR, DBG_ACTIVE_EL1, 0);
744 		kernel_step = this_cpu_ptr(&stepping_kernel_bp);
745 
746 		if (*kernel_step != ARM_KERNEL_STEP_NONE)
747 			return 0;
748 
749 		if (kernel_active_single_step()) {
750 			*kernel_step = ARM_KERNEL_STEP_SUSPEND;
751 		} else {
752 			*kernel_step = ARM_KERNEL_STEP_ACTIVE;
753 			kernel_enable_single_step(regs);
754 		}
755 	}
756 
757 	return 0;
758 }
759 
760 /*
761  * Handle single-step exception.
762  */
763 int reinstall_suspended_bps(struct pt_regs *regs)
764 {
765 	struct debug_info *debug_info = &current->thread.debug;
766 	int handled_exception = 0, *kernel_step;
767 
768 	kernel_step = this_cpu_ptr(&stepping_kernel_bp);
769 
770 	/*
771 	 * Called from single-step exception handler.
772 	 * Return 0 if execution can resume, 1 if a SIGTRAP should be
773 	 * reported.
774 	 */
775 	if (user_mode(regs)) {
776 		if (debug_info->bps_disabled) {
777 			debug_info->bps_disabled = 0;
778 			toggle_bp_registers(AARCH64_DBG_REG_BCR, DBG_ACTIVE_EL0, 1);
779 			handled_exception = 1;
780 		}
781 
782 		if (debug_info->wps_disabled) {
783 			debug_info->wps_disabled = 0;
784 			toggle_bp_registers(AARCH64_DBG_REG_WCR, DBG_ACTIVE_EL0, 1);
785 			handled_exception = 1;
786 		}
787 
788 		if (handled_exception) {
789 			if (debug_info->suspended_step) {
790 				debug_info->suspended_step = 0;
791 				/* Allow exception handling to fall-through. */
792 				handled_exception = 0;
793 			} else {
794 				user_disable_single_step(current);
795 			}
796 		}
797 	} else if (*kernel_step != ARM_KERNEL_STEP_NONE) {
798 		toggle_bp_registers(AARCH64_DBG_REG_BCR, DBG_ACTIVE_EL1, 1);
799 		toggle_bp_registers(AARCH64_DBG_REG_WCR, DBG_ACTIVE_EL1, 1);
800 
801 		if (!debug_info->wps_disabled)
802 			toggle_bp_registers(AARCH64_DBG_REG_WCR, DBG_ACTIVE_EL0, 1);
803 
804 		if (*kernel_step != ARM_KERNEL_STEP_SUSPEND) {
805 			kernel_disable_single_step();
806 			handled_exception = 1;
807 		} else {
808 			handled_exception = 0;
809 		}
810 
811 		*kernel_step = ARM_KERNEL_STEP_NONE;
812 	}
813 
814 	return !handled_exception;
815 }
816 
817 /*
818  * Context-switcher for restoring suspended breakpoints.
819  */
820 void hw_breakpoint_thread_switch(struct task_struct *next)
821 {
822 	/*
823 	 *           current        next
824 	 * disabled: 0              0     => The usual case, NOTIFY_DONE
825 	 *           0              1     => Disable the registers
826 	 *           1              0     => Enable the registers
827 	 *           1              1     => NOTIFY_DONE. per-task bps will
828 	 *                                   get taken care of by perf.
829 	 */
830 
831 	struct debug_info *current_debug_info, *next_debug_info;
832 
833 	current_debug_info = &current->thread.debug;
834 	next_debug_info = &next->thread.debug;
835 
836 	/* Update breakpoints. */
837 	if (current_debug_info->bps_disabled != next_debug_info->bps_disabled)
838 		toggle_bp_registers(AARCH64_DBG_REG_BCR,
839 				    DBG_ACTIVE_EL0,
840 				    !next_debug_info->bps_disabled);
841 
842 	/* Update watchpoints. */
843 	if (current_debug_info->wps_disabled != next_debug_info->wps_disabled)
844 		toggle_bp_registers(AARCH64_DBG_REG_WCR,
845 				    DBG_ACTIVE_EL0,
846 				    !next_debug_info->wps_disabled);
847 }
848 
849 /*
850  * CPU initialisation.
851  */
852 static void hw_breakpoint_reset(void *unused)
853 {
854 	int i;
855 	struct perf_event **slots;
856 	/*
857 	 * When a CPU goes through cold-boot, it does not have any installed
858 	 * slot, so it is safe to share the same function for restoring and
859 	 * resetting breakpoints; when a CPU is hotplugged in, it goes
860 	 * through the slots, which are all empty, hence it just resets control
861 	 * and value for debug registers.
862 	 * When this function is triggered on warm-boot through a CPU PM
863 	 * notifier some slots might be initialized; if so they are
864 	 * reprogrammed according to the debug slots content.
865 	 */
866 	for (slots = this_cpu_ptr(bp_on_reg), i = 0; i < core_num_brps; ++i) {
867 		if (slots[i]) {
868 			hw_breakpoint_control(slots[i], HW_BREAKPOINT_RESTORE);
869 		} else {
870 			write_wb_reg(AARCH64_DBG_REG_BCR, i, 0UL);
871 			write_wb_reg(AARCH64_DBG_REG_BVR, i, 0UL);
872 		}
873 	}
874 
875 	for (slots = this_cpu_ptr(wp_on_reg), i = 0; i < core_num_wrps; ++i) {
876 		if (slots[i]) {
877 			hw_breakpoint_control(slots[i], HW_BREAKPOINT_RESTORE);
878 		} else {
879 			write_wb_reg(AARCH64_DBG_REG_WCR, i, 0UL);
880 			write_wb_reg(AARCH64_DBG_REG_WVR, i, 0UL);
881 		}
882 	}
883 }
884 
885 static int hw_breakpoint_reset_notify(struct notifier_block *self,
886 						unsigned long action,
887 						void *hcpu)
888 {
889 	int cpu = (long)hcpu;
890 	if ((action & ~CPU_TASKS_FROZEN) == CPU_ONLINE)
891 		smp_call_function_single(cpu, hw_breakpoint_reset, NULL, 1);
892 	return NOTIFY_OK;
893 }
894 
895 static struct notifier_block hw_breakpoint_reset_nb = {
896 	.notifier_call = hw_breakpoint_reset_notify,
897 };
898 
899 #ifdef CONFIG_CPU_PM
900 extern void cpu_suspend_set_dbg_restorer(void (*hw_bp_restore)(void *));
901 #else
902 static inline void cpu_suspend_set_dbg_restorer(void (*hw_bp_restore)(void *))
903 {
904 }
905 #endif
906 
907 /*
908  * One-time initialisation.
909  */
910 static int __init arch_hw_breakpoint_init(void)
911 {
912 	core_num_brps = get_num_brps();
913 	core_num_wrps = get_num_wrps();
914 
915 	pr_info("found %d breakpoint and %d watchpoint registers.\n",
916 		core_num_brps, core_num_wrps);
917 
918 	cpu_notifier_register_begin();
919 
920 	/*
921 	 * Reset the breakpoint resources. We assume that a halting
922 	 * debugger will leave the world in a nice state for us.
923 	 */
924 	smp_call_function(hw_breakpoint_reset, NULL, 1);
925 	hw_breakpoint_reset(NULL);
926 
927 	/* Register debug fault handlers. */
928 	hook_debug_fault_code(DBG_ESR_EVT_HWBP, breakpoint_handler, SIGTRAP,
929 			      TRAP_HWBKPT, "hw-breakpoint handler");
930 	hook_debug_fault_code(DBG_ESR_EVT_HWWP, watchpoint_handler, SIGTRAP,
931 			      TRAP_HWBKPT, "hw-watchpoint handler");
932 
933 	/* Register hotplug notifier. */
934 	__register_cpu_notifier(&hw_breakpoint_reset_nb);
935 
936 	cpu_notifier_register_done();
937 
938 	/* Register cpu_suspend hw breakpoint restore hook */
939 	cpu_suspend_set_dbg_restorer(hw_breakpoint_reset);
940 
941 	return 0;
942 }
943 arch_initcall(arch_hw_breakpoint_init);
944 
945 void hw_breakpoint_pmu_read(struct perf_event *bp)
946 {
947 }
948 
949 /*
950  * Dummy function to register with die_notifier.
951  */
952 int hw_breakpoint_exceptions_notify(struct notifier_block *unused,
953 				    unsigned long val, void *data)
954 {
955 	return NOTIFY_DONE;
956 }
957