xref: /linux/arch/arm64/kernel/hw_breakpoint.c (revision 69bfec7548f4c1595bac0e3ddfc0458a5af31f4c)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * HW_breakpoint: a unified kernel/user-space hardware breakpoint facility,
4  * using the CPU's debug registers.
5  *
6  * Copyright (C) 2012 ARM Limited
7  * Author: Will Deacon <will.deacon@arm.com>
8  */
9 
10 #define pr_fmt(fmt) "hw-breakpoint: " fmt
11 
12 #include <linux/compat.h>
13 #include <linux/cpu_pm.h>
14 #include <linux/errno.h>
15 #include <linux/hw_breakpoint.h>
16 #include <linux/kprobes.h>
17 #include <linux/perf_event.h>
18 #include <linux/ptrace.h>
19 #include <linux/smp.h>
20 #include <linux/uaccess.h>
21 
22 #include <asm/current.h>
23 #include <asm/debug-monitors.h>
24 #include <asm/hw_breakpoint.h>
25 #include <asm/traps.h>
26 #include <asm/cputype.h>
27 #include <asm/system_misc.h>
28 
29 /* Breakpoint currently in use for each BRP. */
30 static DEFINE_PER_CPU(struct perf_event *, bp_on_reg[ARM_MAX_BRP]);
31 
32 /* Watchpoint currently in use for each WRP. */
33 static DEFINE_PER_CPU(struct perf_event *, wp_on_reg[ARM_MAX_WRP]);
34 
35 /* Currently stepping a per-CPU kernel breakpoint. */
36 static DEFINE_PER_CPU(int, stepping_kernel_bp);
37 
38 /* Number of BRP/WRP registers on this CPU. */
39 static int core_num_brps;
40 static int core_num_wrps;
41 
42 int hw_breakpoint_slots(int type)
43 {
44 	/*
45 	 * We can be called early, so don't rely on
46 	 * our static variables being initialised.
47 	 */
48 	switch (type) {
49 	case TYPE_INST:
50 		return get_num_brps();
51 	case TYPE_DATA:
52 		return get_num_wrps();
53 	default:
54 		pr_warn("unknown slot type: %d\n", type);
55 		return 0;
56 	}
57 }
58 
59 #define READ_WB_REG_CASE(OFF, N, REG, VAL)	\
60 	case (OFF + N):				\
61 		AARCH64_DBG_READ(N, REG, VAL);	\
62 		break
63 
64 #define WRITE_WB_REG_CASE(OFF, N, REG, VAL)	\
65 	case (OFF + N):				\
66 		AARCH64_DBG_WRITE(N, REG, VAL);	\
67 		break
68 
69 #define GEN_READ_WB_REG_CASES(OFF, REG, VAL)	\
70 	READ_WB_REG_CASE(OFF,  0, REG, VAL);	\
71 	READ_WB_REG_CASE(OFF,  1, REG, VAL);	\
72 	READ_WB_REG_CASE(OFF,  2, REG, VAL);	\
73 	READ_WB_REG_CASE(OFF,  3, REG, VAL);	\
74 	READ_WB_REG_CASE(OFF,  4, REG, VAL);	\
75 	READ_WB_REG_CASE(OFF,  5, REG, VAL);	\
76 	READ_WB_REG_CASE(OFF,  6, REG, VAL);	\
77 	READ_WB_REG_CASE(OFF,  7, REG, VAL);	\
78 	READ_WB_REG_CASE(OFF,  8, REG, VAL);	\
79 	READ_WB_REG_CASE(OFF,  9, REG, VAL);	\
80 	READ_WB_REG_CASE(OFF, 10, REG, VAL);	\
81 	READ_WB_REG_CASE(OFF, 11, REG, VAL);	\
82 	READ_WB_REG_CASE(OFF, 12, REG, VAL);	\
83 	READ_WB_REG_CASE(OFF, 13, REG, VAL);	\
84 	READ_WB_REG_CASE(OFF, 14, REG, VAL);	\
85 	READ_WB_REG_CASE(OFF, 15, REG, VAL)
86 
87 #define GEN_WRITE_WB_REG_CASES(OFF, REG, VAL)	\
88 	WRITE_WB_REG_CASE(OFF,  0, REG, VAL);	\
89 	WRITE_WB_REG_CASE(OFF,  1, REG, VAL);	\
90 	WRITE_WB_REG_CASE(OFF,  2, REG, VAL);	\
91 	WRITE_WB_REG_CASE(OFF,  3, REG, VAL);	\
92 	WRITE_WB_REG_CASE(OFF,  4, REG, VAL);	\
93 	WRITE_WB_REG_CASE(OFF,  5, REG, VAL);	\
94 	WRITE_WB_REG_CASE(OFF,  6, REG, VAL);	\
95 	WRITE_WB_REG_CASE(OFF,  7, REG, VAL);	\
96 	WRITE_WB_REG_CASE(OFF,  8, REG, VAL);	\
97 	WRITE_WB_REG_CASE(OFF,  9, REG, VAL);	\
98 	WRITE_WB_REG_CASE(OFF, 10, REG, VAL);	\
99 	WRITE_WB_REG_CASE(OFF, 11, REG, VAL);	\
100 	WRITE_WB_REG_CASE(OFF, 12, REG, VAL);	\
101 	WRITE_WB_REG_CASE(OFF, 13, REG, VAL);	\
102 	WRITE_WB_REG_CASE(OFF, 14, REG, VAL);	\
103 	WRITE_WB_REG_CASE(OFF, 15, REG, VAL)
104 
105 static u64 read_wb_reg(int reg, int n)
106 {
107 	u64 val = 0;
108 
109 	switch (reg + n) {
110 	GEN_READ_WB_REG_CASES(AARCH64_DBG_REG_BVR, AARCH64_DBG_REG_NAME_BVR, val);
111 	GEN_READ_WB_REG_CASES(AARCH64_DBG_REG_BCR, AARCH64_DBG_REG_NAME_BCR, val);
112 	GEN_READ_WB_REG_CASES(AARCH64_DBG_REG_WVR, AARCH64_DBG_REG_NAME_WVR, val);
113 	GEN_READ_WB_REG_CASES(AARCH64_DBG_REG_WCR, AARCH64_DBG_REG_NAME_WCR, val);
114 	default:
115 		pr_warn("attempt to read from unknown breakpoint register %d\n", n);
116 	}
117 
118 	return val;
119 }
120 NOKPROBE_SYMBOL(read_wb_reg);
121 
122 static void write_wb_reg(int reg, int n, u64 val)
123 {
124 	switch (reg + n) {
125 	GEN_WRITE_WB_REG_CASES(AARCH64_DBG_REG_BVR, AARCH64_DBG_REG_NAME_BVR, val);
126 	GEN_WRITE_WB_REG_CASES(AARCH64_DBG_REG_BCR, AARCH64_DBG_REG_NAME_BCR, val);
127 	GEN_WRITE_WB_REG_CASES(AARCH64_DBG_REG_WVR, AARCH64_DBG_REG_NAME_WVR, val);
128 	GEN_WRITE_WB_REG_CASES(AARCH64_DBG_REG_WCR, AARCH64_DBG_REG_NAME_WCR, val);
129 	default:
130 		pr_warn("attempt to write to unknown breakpoint register %d\n", n);
131 	}
132 	isb();
133 }
134 NOKPROBE_SYMBOL(write_wb_reg);
135 
136 /*
137  * Convert a breakpoint privilege level to the corresponding exception
138  * level.
139  */
140 static enum dbg_active_el debug_exception_level(int privilege)
141 {
142 	switch (privilege) {
143 	case AARCH64_BREAKPOINT_EL0:
144 		return DBG_ACTIVE_EL0;
145 	case AARCH64_BREAKPOINT_EL1:
146 		return DBG_ACTIVE_EL1;
147 	default:
148 		pr_warn("invalid breakpoint privilege level %d\n", privilege);
149 		return -EINVAL;
150 	}
151 }
152 NOKPROBE_SYMBOL(debug_exception_level);
153 
154 enum hw_breakpoint_ops {
155 	HW_BREAKPOINT_INSTALL,
156 	HW_BREAKPOINT_UNINSTALL,
157 	HW_BREAKPOINT_RESTORE
158 };
159 
160 static int is_compat_bp(struct perf_event *bp)
161 {
162 	struct task_struct *tsk = bp->hw.target;
163 
164 	/*
165 	 * tsk can be NULL for per-cpu (non-ptrace) breakpoints.
166 	 * In this case, use the native interface, since we don't have
167 	 * the notion of a "compat CPU" and could end up relying on
168 	 * deprecated behaviour if we use unaligned watchpoints in
169 	 * AArch64 state.
170 	 */
171 	return tsk && is_compat_thread(task_thread_info(tsk));
172 }
173 
174 /**
175  * hw_breakpoint_slot_setup - Find and setup a perf slot according to
176  *			      operations
177  *
178  * @slots: pointer to array of slots
179  * @max_slots: max number of slots
180  * @bp: perf_event to setup
181  * @ops: operation to be carried out on the slot
182  *
183  * Return:
184  *	slot index on success
185  *	-ENOSPC if no slot is available/matches
186  *	-EINVAL on wrong operations parameter
187  */
188 static int hw_breakpoint_slot_setup(struct perf_event **slots, int max_slots,
189 				    struct perf_event *bp,
190 				    enum hw_breakpoint_ops ops)
191 {
192 	int i;
193 	struct perf_event **slot;
194 
195 	for (i = 0; i < max_slots; ++i) {
196 		slot = &slots[i];
197 		switch (ops) {
198 		case HW_BREAKPOINT_INSTALL:
199 			if (!*slot) {
200 				*slot = bp;
201 				return i;
202 			}
203 			break;
204 		case HW_BREAKPOINT_UNINSTALL:
205 			if (*slot == bp) {
206 				*slot = NULL;
207 				return i;
208 			}
209 			break;
210 		case HW_BREAKPOINT_RESTORE:
211 			if (*slot == bp)
212 				return i;
213 			break;
214 		default:
215 			pr_warn_once("Unhandled hw breakpoint ops %d\n", ops);
216 			return -EINVAL;
217 		}
218 	}
219 	return -ENOSPC;
220 }
221 
222 static int hw_breakpoint_control(struct perf_event *bp,
223 				 enum hw_breakpoint_ops ops)
224 {
225 	struct arch_hw_breakpoint *info = counter_arch_bp(bp);
226 	struct perf_event **slots;
227 	struct debug_info *debug_info = &current->thread.debug;
228 	int i, max_slots, ctrl_reg, val_reg, reg_enable;
229 	enum dbg_active_el dbg_el = debug_exception_level(info->ctrl.privilege);
230 	u32 ctrl;
231 
232 	if (info->ctrl.type == ARM_BREAKPOINT_EXECUTE) {
233 		/* Breakpoint */
234 		ctrl_reg = AARCH64_DBG_REG_BCR;
235 		val_reg = AARCH64_DBG_REG_BVR;
236 		slots = this_cpu_ptr(bp_on_reg);
237 		max_slots = core_num_brps;
238 		reg_enable = !debug_info->bps_disabled;
239 	} else {
240 		/* Watchpoint */
241 		ctrl_reg = AARCH64_DBG_REG_WCR;
242 		val_reg = AARCH64_DBG_REG_WVR;
243 		slots = this_cpu_ptr(wp_on_reg);
244 		max_slots = core_num_wrps;
245 		reg_enable = !debug_info->wps_disabled;
246 	}
247 
248 	i = hw_breakpoint_slot_setup(slots, max_slots, bp, ops);
249 
250 	if (WARN_ONCE(i < 0, "Can't find any breakpoint slot"))
251 		return i;
252 
253 	switch (ops) {
254 	case HW_BREAKPOINT_INSTALL:
255 		/*
256 		 * Ensure debug monitors are enabled at the correct exception
257 		 * level.
258 		 */
259 		enable_debug_monitors(dbg_el);
260 		fallthrough;
261 	case HW_BREAKPOINT_RESTORE:
262 		/* Setup the address register. */
263 		write_wb_reg(val_reg, i, info->address);
264 
265 		/* Setup the control register. */
266 		ctrl = encode_ctrl_reg(info->ctrl);
267 		write_wb_reg(ctrl_reg, i,
268 			     reg_enable ? ctrl | 0x1 : ctrl & ~0x1);
269 		break;
270 	case HW_BREAKPOINT_UNINSTALL:
271 		/* Reset the control register. */
272 		write_wb_reg(ctrl_reg, i, 0);
273 
274 		/*
275 		 * Release the debug monitors for the correct exception
276 		 * level.
277 		 */
278 		disable_debug_monitors(dbg_el);
279 		break;
280 	}
281 
282 	return 0;
283 }
284 
285 /*
286  * Install a perf counter breakpoint.
287  */
288 int arch_install_hw_breakpoint(struct perf_event *bp)
289 {
290 	return hw_breakpoint_control(bp, HW_BREAKPOINT_INSTALL);
291 }
292 
293 void arch_uninstall_hw_breakpoint(struct perf_event *bp)
294 {
295 	hw_breakpoint_control(bp, HW_BREAKPOINT_UNINSTALL);
296 }
297 
298 static int get_hbp_len(u8 hbp_len)
299 {
300 	unsigned int len_in_bytes = 0;
301 
302 	switch (hbp_len) {
303 	case ARM_BREAKPOINT_LEN_1:
304 		len_in_bytes = 1;
305 		break;
306 	case ARM_BREAKPOINT_LEN_2:
307 		len_in_bytes = 2;
308 		break;
309 	case ARM_BREAKPOINT_LEN_3:
310 		len_in_bytes = 3;
311 		break;
312 	case ARM_BREAKPOINT_LEN_4:
313 		len_in_bytes = 4;
314 		break;
315 	case ARM_BREAKPOINT_LEN_5:
316 		len_in_bytes = 5;
317 		break;
318 	case ARM_BREAKPOINT_LEN_6:
319 		len_in_bytes = 6;
320 		break;
321 	case ARM_BREAKPOINT_LEN_7:
322 		len_in_bytes = 7;
323 		break;
324 	case ARM_BREAKPOINT_LEN_8:
325 		len_in_bytes = 8;
326 		break;
327 	}
328 
329 	return len_in_bytes;
330 }
331 
332 /*
333  * Check whether bp virtual address is in kernel space.
334  */
335 int arch_check_bp_in_kernelspace(struct arch_hw_breakpoint *hw)
336 {
337 	unsigned int len;
338 	unsigned long va;
339 
340 	va = hw->address;
341 	len = get_hbp_len(hw->ctrl.len);
342 
343 	return (va >= TASK_SIZE) && ((va + len - 1) >= TASK_SIZE);
344 }
345 
346 /*
347  * Extract generic type and length encodings from an arch_hw_breakpoint_ctrl.
348  * Hopefully this will disappear when ptrace can bypass the conversion
349  * to generic breakpoint descriptions.
350  */
351 int arch_bp_generic_fields(struct arch_hw_breakpoint_ctrl ctrl,
352 			   int *gen_len, int *gen_type, int *offset)
353 {
354 	/* Type */
355 	switch (ctrl.type) {
356 	case ARM_BREAKPOINT_EXECUTE:
357 		*gen_type = HW_BREAKPOINT_X;
358 		break;
359 	case ARM_BREAKPOINT_LOAD:
360 		*gen_type = HW_BREAKPOINT_R;
361 		break;
362 	case ARM_BREAKPOINT_STORE:
363 		*gen_type = HW_BREAKPOINT_W;
364 		break;
365 	case ARM_BREAKPOINT_LOAD | ARM_BREAKPOINT_STORE:
366 		*gen_type = HW_BREAKPOINT_RW;
367 		break;
368 	default:
369 		return -EINVAL;
370 	}
371 
372 	if (!ctrl.len)
373 		return -EINVAL;
374 	*offset = __ffs(ctrl.len);
375 
376 	/* Len */
377 	switch (ctrl.len >> *offset) {
378 	case ARM_BREAKPOINT_LEN_1:
379 		*gen_len = HW_BREAKPOINT_LEN_1;
380 		break;
381 	case ARM_BREAKPOINT_LEN_2:
382 		*gen_len = HW_BREAKPOINT_LEN_2;
383 		break;
384 	case ARM_BREAKPOINT_LEN_3:
385 		*gen_len = HW_BREAKPOINT_LEN_3;
386 		break;
387 	case ARM_BREAKPOINT_LEN_4:
388 		*gen_len = HW_BREAKPOINT_LEN_4;
389 		break;
390 	case ARM_BREAKPOINT_LEN_5:
391 		*gen_len = HW_BREAKPOINT_LEN_5;
392 		break;
393 	case ARM_BREAKPOINT_LEN_6:
394 		*gen_len = HW_BREAKPOINT_LEN_6;
395 		break;
396 	case ARM_BREAKPOINT_LEN_7:
397 		*gen_len = HW_BREAKPOINT_LEN_7;
398 		break;
399 	case ARM_BREAKPOINT_LEN_8:
400 		*gen_len = HW_BREAKPOINT_LEN_8;
401 		break;
402 	default:
403 		return -EINVAL;
404 	}
405 
406 	return 0;
407 }
408 
409 /*
410  * Construct an arch_hw_breakpoint from a perf_event.
411  */
412 static int arch_build_bp_info(struct perf_event *bp,
413 			      const struct perf_event_attr *attr,
414 			      struct arch_hw_breakpoint *hw)
415 {
416 	/* Type */
417 	switch (attr->bp_type) {
418 	case HW_BREAKPOINT_X:
419 		hw->ctrl.type = ARM_BREAKPOINT_EXECUTE;
420 		break;
421 	case HW_BREAKPOINT_R:
422 		hw->ctrl.type = ARM_BREAKPOINT_LOAD;
423 		break;
424 	case HW_BREAKPOINT_W:
425 		hw->ctrl.type = ARM_BREAKPOINT_STORE;
426 		break;
427 	case HW_BREAKPOINT_RW:
428 		hw->ctrl.type = ARM_BREAKPOINT_LOAD | ARM_BREAKPOINT_STORE;
429 		break;
430 	default:
431 		return -EINVAL;
432 	}
433 
434 	/* Len */
435 	switch (attr->bp_len) {
436 	case HW_BREAKPOINT_LEN_1:
437 		hw->ctrl.len = ARM_BREAKPOINT_LEN_1;
438 		break;
439 	case HW_BREAKPOINT_LEN_2:
440 		hw->ctrl.len = ARM_BREAKPOINT_LEN_2;
441 		break;
442 	case HW_BREAKPOINT_LEN_3:
443 		hw->ctrl.len = ARM_BREAKPOINT_LEN_3;
444 		break;
445 	case HW_BREAKPOINT_LEN_4:
446 		hw->ctrl.len = ARM_BREAKPOINT_LEN_4;
447 		break;
448 	case HW_BREAKPOINT_LEN_5:
449 		hw->ctrl.len = ARM_BREAKPOINT_LEN_5;
450 		break;
451 	case HW_BREAKPOINT_LEN_6:
452 		hw->ctrl.len = ARM_BREAKPOINT_LEN_6;
453 		break;
454 	case HW_BREAKPOINT_LEN_7:
455 		hw->ctrl.len = ARM_BREAKPOINT_LEN_7;
456 		break;
457 	case HW_BREAKPOINT_LEN_8:
458 		hw->ctrl.len = ARM_BREAKPOINT_LEN_8;
459 		break;
460 	default:
461 		return -EINVAL;
462 	}
463 
464 	/*
465 	 * On AArch64, we only permit breakpoints of length 4, whereas
466 	 * AArch32 also requires breakpoints of length 2 for Thumb.
467 	 * Watchpoints can be of length 1, 2, 4 or 8 bytes.
468 	 */
469 	if (hw->ctrl.type == ARM_BREAKPOINT_EXECUTE) {
470 		if (is_compat_bp(bp)) {
471 			if (hw->ctrl.len != ARM_BREAKPOINT_LEN_2 &&
472 			    hw->ctrl.len != ARM_BREAKPOINT_LEN_4)
473 				return -EINVAL;
474 		} else if (hw->ctrl.len != ARM_BREAKPOINT_LEN_4) {
475 			/*
476 			 * FIXME: Some tools (I'm looking at you perf) assume
477 			 *	  that breakpoints should be sizeof(long). This
478 			 *	  is nonsense. For now, we fix up the parameter
479 			 *	  but we should probably return -EINVAL instead.
480 			 */
481 			hw->ctrl.len = ARM_BREAKPOINT_LEN_4;
482 		}
483 	}
484 
485 	/* Address */
486 	hw->address = attr->bp_addr;
487 
488 	/*
489 	 * Privilege
490 	 * Note that we disallow combined EL0/EL1 breakpoints because
491 	 * that would complicate the stepping code.
492 	 */
493 	if (arch_check_bp_in_kernelspace(hw))
494 		hw->ctrl.privilege = AARCH64_BREAKPOINT_EL1;
495 	else
496 		hw->ctrl.privilege = AARCH64_BREAKPOINT_EL0;
497 
498 	/* Enabled? */
499 	hw->ctrl.enabled = !attr->disabled;
500 
501 	return 0;
502 }
503 
504 /*
505  * Validate the arch-specific HW Breakpoint register settings.
506  */
507 int hw_breakpoint_arch_parse(struct perf_event *bp,
508 			     const struct perf_event_attr *attr,
509 			     struct arch_hw_breakpoint *hw)
510 {
511 	int ret;
512 	u64 alignment_mask, offset;
513 
514 	/* Build the arch_hw_breakpoint. */
515 	ret = arch_build_bp_info(bp, attr, hw);
516 	if (ret)
517 		return ret;
518 
519 	/*
520 	 * Check address alignment.
521 	 * We don't do any clever alignment correction for watchpoints
522 	 * because using 64-bit unaligned addresses is deprecated for
523 	 * AArch64.
524 	 *
525 	 * AArch32 tasks expect some simple alignment fixups, so emulate
526 	 * that here.
527 	 */
528 	if (is_compat_bp(bp)) {
529 		if (hw->ctrl.len == ARM_BREAKPOINT_LEN_8)
530 			alignment_mask = 0x7;
531 		else
532 			alignment_mask = 0x3;
533 		offset = hw->address & alignment_mask;
534 		switch (offset) {
535 		case 0:
536 			/* Aligned */
537 			break;
538 		case 1:
539 		case 2:
540 			/* Allow halfword watchpoints and breakpoints. */
541 			if (hw->ctrl.len == ARM_BREAKPOINT_LEN_2)
542 				break;
543 
544 			fallthrough;
545 		case 3:
546 			/* Allow single byte watchpoint. */
547 			if (hw->ctrl.len == ARM_BREAKPOINT_LEN_1)
548 				break;
549 
550 			fallthrough;
551 		default:
552 			return -EINVAL;
553 		}
554 	} else {
555 		if (hw->ctrl.type == ARM_BREAKPOINT_EXECUTE)
556 			alignment_mask = 0x3;
557 		else
558 			alignment_mask = 0x7;
559 		offset = hw->address & alignment_mask;
560 	}
561 
562 	hw->address &= ~alignment_mask;
563 	hw->ctrl.len <<= offset;
564 
565 	/*
566 	 * Disallow per-task kernel breakpoints since these would
567 	 * complicate the stepping code.
568 	 */
569 	if (hw->ctrl.privilege == AARCH64_BREAKPOINT_EL1 && bp->hw.target)
570 		return -EINVAL;
571 
572 	return 0;
573 }
574 
575 /*
576  * Enable/disable all of the breakpoints active at the specified
577  * exception level at the register level.
578  * This is used when single-stepping after a breakpoint exception.
579  */
580 static void toggle_bp_registers(int reg, enum dbg_active_el el, int enable)
581 {
582 	int i, max_slots, privilege;
583 	u32 ctrl;
584 	struct perf_event **slots;
585 
586 	switch (reg) {
587 	case AARCH64_DBG_REG_BCR:
588 		slots = this_cpu_ptr(bp_on_reg);
589 		max_slots = core_num_brps;
590 		break;
591 	case AARCH64_DBG_REG_WCR:
592 		slots = this_cpu_ptr(wp_on_reg);
593 		max_slots = core_num_wrps;
594 		break;
595 	default:
596 		return;
597 	}
598 
599 	for (i = 0; i < max_slots; ++i) {
600 		if (!slots[i])
601 			continue;
602 
603 		privilege = counter_arch_bp(slots[i])->ctrl.privilege;
604 		if (debug_exception_level(privilege) != el)
605 			continue;
606 
607 		ctrl = read_wb_reg(reg, i);
608 		if (enable)
609 			ctrl |= 0x1;
610 		else
611 			ctrl &= ~0x1;
612 		write_wb_reg(reg, i, ctrl);
613 	}
614 }
615 NOKPROBE_SYMBOL(toggle_bp_registers);
616 
617 /*
618  * Debug exception handlers.
619  */
620 static int breakpoint_handler(unsigned long unused, unsigned long esr,
621 			      struct pt_regs *regs)
622 {
623 	int i, step = 0, *kernel_step;
624 	u32 ctrl_reg;
625 	u64 addr, val;
626 	struct perf_event *bp, **slots;
627 	struct debug_info *debug_info;
628 	struct arch_hw_breakpoint_ctrl ctrl;
629 
630 	slots = this_cpu_ptr(bp_on_reg);
631 	addr = instruction_pointer(regs);
632 	debug_info = &current->thread.debug;
633 
634 	for (i = 0; i < core_num_brps; ++i) {
635 		rcu_read_lock();
636 
637 		bp = slots[i];
638 
639 		if (bp == NULL)
640 			goto unlock;
641 
642 		/* Check if the breakpoint value matches. */
643 		val = read_wb_reg(AARCH64_DBG_REG_BVR, i);
644 		if (val != (addr & ~0x3))
645 			goto unlock;
646 
647 		/* Possible match, check the byte address select to confirm. */
648 		ctrl_reg = read_wb_reg(AARCH64_DBG_REG_BCR, i);
649 		decode_ctrl_reg(ctrl_reg, &ctrl);
650 		if (!((1 << (addr & 0x3)) & ctrl.len))
651 			goto unlock;
652 
653 		counter_arch_bp(bp)->trigger = addr;
654 		perf_bp_event(bp, regs);
655 
656 		/* Do we need to handle the stepping? */
657 		if (is_default_overflow_handler(bp))
658 			step = 1;
659 unlock:
660 		rcu_read_unlock();
661 	}
662 
663 	if (!step)
664 		return 0;
665 
666 	if (user_mode(regs)) {
667 		debug_info->bps_disabled = 1;
668 		toggle_bp_registers(AARCH64_DBG_REG_BCR, DBG_ACTIVE_EL0, 0);
669 
670 		/* If we're already stepping a watchpoint, just return. */
671 		if (debug_info->wps_disabled)
672 			return 0;
673 
674 		if (test_thread_flag(TIF_SINGLESTEP))
675 			debug_info->suspended_step = 1;
676 		else
677 			user_enable_single_step(current);
678 	} else {
679 		toggle_bp_registers(AARCH64_DBG_REG_BCR, DBG_ACTIVE_EL1, 0);
680 		kernel_step = this_cpu_ptr(&stepping_kernel_bp);
681 
682 		if (*kernel_step != ARM_KERNEL_STEP_NONE)
683 			return 0;
684 
685 		if (kernel_active_single_step()) {
686 			*kernel_step = ARM_KERNEL_STEP_SUSPEND;
687 		} else {
688 			*kernel_step = ARM_KERNEL_STEP_ACTIVE;
689 			kernel_enable_single_step(regs);
690 		}
691 	}
692 
693 	return 0;
694 }
695 NOKPROBE_SYMBOL(breakpoint_handler);
696 
697 /*
698  * Arm64 hardware does not always report a watchpoint hit address that matches
699  * one of the watchpoints set. It can also report an address "near" the
700  * watchpoint if a single instruction access both watched and unwatched
701  * addresses. There is no straight-forward way, short of disassembling the
702  * offending instruction, to map that address back to the watchpoint. This
703  * function computes the distance of the memory access from the watchpoint as a
704  * heuristic for the likelihood that a given access triggered the watchpoint.
705  *
706  * See Section D2.10.5 "Determining the memory location that caused a Watchpoint
707  * exception" of ARMv8 Architecture Reference Manual for details.
708  *
709  * The function returns the distance of the address from the bytes watched by
710  * the watchpoint. In case of an exact match, it returns 0.
711  */
712 static u64 get_distance_from_watchpoint(unsigned long addr, u64 val,
713 					struct arch_hw_breakpoint_ctrl *ctrl)
714 {
715 	u64 wp_low, wp_high;
716 	u32 lens, lene;
717 
718 	addr = untagged_addr(addr);
719 
720 	lens = __ffs(ctrl->len);
721 	lene = __fls(ctrl->len);
722 
723 	wp_low = val + lens;
724 	wp_high = val + lene;
725 	if (addr < wp_low)
726 		return wp_low - addr;
727 	else if (addr > wp_high)
728 		return addr - wp_high;
729 	else
730 		return 0;
731 }
732 
733 static int watchpoint_report(struct perf_event *wp, unsigned long addr,
734 			     struct pt_regs *regs)
735 {
736 	int step = is_default_overflow_handler(wp);
737 	struct arch_hw_breakpoint *info = counter_arch_bp(wp);
738 
739 	info->trigger = addr;
740 
741 	/*
742 	 * If we triggered a user watchpoint from a uaccess routine, then
743 	 * handle the stepping ourselves since userspace really can't help
744 	 * us with this.
745 	 */
746 	if (!user_mode(regs) && info->ctrl.privilege == AARCH64_BREAKPOINT_EL0)
747 		step = 1;
748 	else
749 		perf_bp_event(wp, regs);
750 
751 	return step;
752 }
753 
754 static int watchpoint_handler(unsigned long addr, unsigned long esr,
755 			      struct pt_regs *regs)
756 {
757 	int i, step = 0, *kernel_step, access, closest_match = 0;
758 	u64 min_dist = -1, dist;
759 	u32 ctrl_reg;
760 	u64 val;
761 	struct perf_event *wp, **slots;
762 	struct debug_info *debug_info;
763 	struct arch_hw_breakpoint_ctrl ctrl;
764 
765 	slots = this_cpu_ptr(wp_on_reg);
766 	debug_info = &current->thread.debug;
767 
768 	/*
769 	 * Find all watchpoints that match the reported address. If no exact
770 	 * match is found. Attribute the hit to the closest watchpoint.
771 	 */
772 	rcu_read_lock();
773 	for (i = 0; i < core_num_wrps; ++i) {
774 		wp = slots[i];
775 		if (wp == NULL)
776 			continue;
777 
778 		/*
779 		 * Check that the access type matches.
780 		 * 0 => load, otherwise => store
781 		 */
782 		access = (esr & AARCH64_ESR_ACCESS_MASK) ? HW_BREAKPOINT_W :
783 			 HW_BREAKPOINT_R;
784 		if (!(access & hw_breakpoint_type(wp)))
785 			continue;
786 
787 		/* Check if the watchpoint value and byte select match. */
788 		val = read_wb_reg(AARCH64_DBG_REG_WVR, i);
789 		ctrl_reg = read_wb_reg(AARCH64_DBG_REG_WCR, i);
790 		decode_ctrl_reg(ctrl_reg, &ctrl);
791 		dist = get_distance_from_watchpoint(addr, val, &ctrl);
792 		if (dist < min_dist) {
793 			min_dist = dist;
794 			closest_match = i;
795 		}
796 		/* Is this an exact match? */
797 		if (dist != 0)
798 			continue;
799 
800 		step = watchpoint_report(wp, addr, regs);
801 	}
802 
803 	/* No exact match found? */
804 	if (min_dist > 0 && min_dist != -1)
805 		step = watchpoint_report(slots[closest_match], addr, regs);
806 
807 	rcu_read_unlock();
808 
809 	if (!step)
810 		return 0;
811 
812 	/*
813 	 * We always disable EL0 watchpoints because the kernel can
814 	 * cause these to fire via an unprivileged access.
815 	 */
816 	toggle_bp_registers(AARCH64_DBG_REG_WCR, DBG_ACTIVE_EL0, 0);
817 
818 	if (user_mode(regs)) {
819 		debug_info->wps_disabled = 1;
820 
821 		/* If we're already stepping a breakpoint, just return. */
822 		if (debug_info->bps_disabled)
823 			return 0;
824 
825 		if (test_thread_flag(TIF_SINGLESTEP))
826 			debug_info->suspended_step = 1;
827 		else
828 			user_enable_single_step(current);
829 	} else {
830 		toggle_bp_registers(AARCH64_DBG_REG_WCR, DBG_ACTIVE_EL1, 0);
831 		kernel_step = this_cpu_ptr(&stepping_kernel_bp);
832 
833 		if (*kernel_step != ARM_KERNEL_STEP_NONE)
834 			return 0;
835 
836 		if (kernel_active_single_step()) {
837 			*kernel_step = ARM_KERNEL_STEP_SUSPEND;
838 		} else {
839 			*kernel_step = ARM_KERNEL_STEP_ACTIVE;
840 			kernel_enable_single_step(regs);
841 		}
842 	}
843 
844 	return 0;
845 }
846 NOKPROBE_SYMBOL(watchpoint_handler);
847 
848 /*
849  * Handle single-step exception.
850  */
851 int reinstall_suspended_bps(struct pt_regs *regs)
852 {
853 	struct debug_info *debug_info = &current->thread.debug;
854 	int handled_exception = 0, *kernel_step;
855 
856 	kernel_step = this_cpu_ptr(&stepping_kernel_bp);
857 
858 	/*
859 	 * Called from single-step exception handler.
860 	 * Return 0 if execution can resume, 1 if a SIGTRAP should be
861 	 * reported.
862 	 */
863 	if (user_mode(regs)) {
864 		if (debug_info->bps_disabled) {
865 			debug_info->bps_disabled = 0;
866 			toggle_bp_registers(AARCH64_DBG_REG_BCR, DBG_ACTIVE_EL0, 1);
867 			handled_exception = 1;
868 		}
869 
870 		if (debug_info->wps_disabled) {
871 			debug_info->wps_disabled = 0;
872 			toggle_bp_registers(AARCH64_DBG_REG_WCR, DBG_ACTIVE_EL0, 1);
873 			handled_exception = 1;
874 		}
875 
876 		if (handled_exception) {
877 			if (debug_info->suspended_step) {
878 				debug_info->suspended_step = 0;
879 				/* Allow exception handling to fall-through. */
880 				handled_exception = 0;
881 			} else {
882 				user_disable_single_step(current);
883 			}
884 		}
885 	} else if (*kernel_step != ARM_KERNEL_STEP_NONE) {
886 		toggle_bp_registers(AARCH64_DBG_REG_BCR, DBG_ACTIVE_EL1, 1);
887 		toggle_bp_registers(AARCH64_DBG_REG_WCR, DBG_ACTIVE_EL1, 1);
888 
889 		if (!debug_info->wps_disabled)
890 			toggle_bp_registers(AARCH64_DBG_REG_WCR, DBG_ACTIVE_EL0, 1);
891 
892 		if (*kernel_step != ARM_KERNEL_STEP_SUSPEND) {
893 			kernel_disable_single_step();
894 			handled_exception = 1;
895 		} else {
896 			handled_exception = 0;
897 		}
898 
899 		*kernel_step = ARM_KERNEL_STEP_NONE;
900 	}
901 
902 	return !handled_exception;
903 }
904 NOKPROBE_SYMBOL(reinstall_suspended_bps);
905 
906 /*
907  * Context-switcher for restoring suspended breakpoints.
908  */
909 void hw_breakpoint_thread_switch(struct task_struct *next)
910 {
911 	/*
912 	 *           current        next
913 	 * disabled: 0              0     => The usual case, NOTIFY_DONE
914 	 *           0              1     => Disable the registers
915 	 *           1              0     => Enable the registers
916 	 *           1              1     => NOTIFY_DONE. per-task bps will
917 	 *                                   get taken care of by perf.
918 	 */
919 
920 	struct debug_info *current_debug_info, *next_debug_info;
921 
922 	current_debug_info = &current->thread.debug;
923 	next_debug_info = &next->thread.debug;
924 
925 	/* Update breakpoints. */
926 	if (current_debug_info->bps_disabled != next_debug_info->bps_disabled)
927 		toggle_bp_registers(AARCH64_DBG_REG_BCR,
928 				    DBG_ACTIVE_EL0,
929 				    !next_debug_info->bps_disabled);
930 
931 	/* Update watchpoints. */
932 	if (current_debug_info->wps_disabled != next_debug_info->wps_disabled)
933 		toggle_bp_registers(AARCH64_DBG_REG_WCR,
934 				    DBG_ACTIVE_EL0,
935 				    !next_debug_info->wps_disabled);
936 }
937 
938 /*
939  * CPU initialisation.
940  */
941 static int hw_breakpoint_reset(unsigned int cpu)
942 {
943 	int i;
944 	struct perf_event **slots;
945 	/*
946 	 * When a CPU goes through cold-boot, it does not have any installed
947 	 * slot, so it is safe to share the same function for restoring and
948 	 * resetting breakpoints; when a CPU is hotplugged in, it goes
949 	 * through the slots, which are all empty, hence it just resets control
950 	 * and value for debug registers.
951 	 * When this function is triggered on warm-boot through a CPU PM
952 	 * notifier some slots might be initialized; if so they are
953 	 * reprogrammed according to the debug slots content.
954 	 */
955 	for (slots = this_cpu_ptr(bp_on_reg), i = 0; i < core_num_brps; ++i) {
956 		if (slots[i]) {
957 			hw_breakpoint_control(slots[i], HW_BREAKPOINT_RESTORE);
958 		} else {
959 			write_wb_reg(AARCH64_DBG_REG_BCR, i, 0UL);
960 			write_wb_reg(AARCH64_DBG_REG_BVR, i, 0UL);
961 		}
962 	}
963 
964 	for (slots = this_cpu_ptr(wp_on_reg), i = 0; i < core_num_wrps; ++i) {
965 		if (slots[i]) {
966 			hw_breakpoint_control(slots[i], HW_BREAKPOINT_RESTORE);
967 		} else {
968 			write_wb_reg(AARCH64_DBG_REG_WCR, i, 0UL);
969 			write_wb_reg(AARCH64_DBG_REG_WVR, i, 0UL);
970 		}
971 	}
972 
973 	return 0;
974 }
975 
976 #ifdef CONFIG_CPU_PM
977 extern void cpu_suspend_set_dbg_restorer(int (*hw_bp_restore)(unsigned int));
978 #else
979 static inline void cpu_suspend_set_dbg_restorer(int (*hw_bp_restore)(unsigned int))
980 {
981 }
982 #endif
983 
984 /*
985  * One-time initialisation.
986  */
987 static int __init arch_hw_breakpoint_init(void)
988 {
989 	int ret;
990 
991 	core_num_brps = get_num_brps();
992 	core_num_wrps = get_num_wrps();
993 
994 	pr_info("found %d breakpoint and %d watchpoint registers.\n",
995 		core_num_brps, core_num_wrps);
996 
997 	/* Register debug fault handlers. */
998 	hook_debug_fault_code(DBG_ESR_EVT_HWBP, breakpoint_handler, SIGTRAP,
999 			      TRAP_HWBKPT, "hw-breakpoint handler");
1000 	hook_debug_fault_code(DBG_ESR_EVT_HWWP, watchpoint_handler, SIGTRAP,
1001 			      TRAP_HWBKPT, "hw-watchpoint handler");
1002 
1003 	/*
1004 	 * Reset the breakpoint resources. We assume that a halting
1005 	 * debugger will leave the world in a nice state for us.
1006 	 */
1007 	ret = cpuhp_setup_state(CPUHP_AP_PERF_ARM_HW_BREAKPOINT_STARTING,
1008 			  "perf/arm64/hw_breakpoint:starting",
1009 			  hw_breakpoint_reset, NULL);
1010 	if (ret)
1011 		pr_err("failed to register CPU hotplug notifier: %d\n", ret);
1012 
1013 	/* Register cpu_suspend hw breakpoint restore hook */
1014 	cpu_suspend_set_dbg_restorer(hw_breakpoint_reset);
1015 
1016 	return ret;
1017 }
1018 arch_initcall(arch_hw_breakpoint_init);
1019 
1020 void hw_breakpoint_pmu_read(struct perf_event *bp)
1021 {
1022 }
1023 
1024 /*
1025  * Dummy function to register with die_notifier.
1026  */
1027 int hw_breakpoint_exceptions_notify(struct notifier_block *unused,
1028 				    unsigned long val, void *data)
1029 {
1030 	return NOTIFY_DONE;
1031 }
1032