xref: /openbmc/linux/arch/arm/kernel/hw_breakpoint.c (revision f220d3eb)
1 /*
2  * This program is free software; you can redistribute it and/or modify
3  * it under the terms of the GNU General Public License version 2 as
4  * published by the Free Software Foundation.
5  *
6  * This program is distributed in the hope that it will be useful,
7  * but WITHOUT ANY WARRANTY; without even the implied warranty of
8  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
9  * GNU General Public License for more details.
10  *
11  * You should have received a copy of the GNU General Public License
12  * along with this program; if not, write to the Free Software
13  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
14  *
15  * Copyright (C) 2009, 2010 ARM Limited
16  *
17  * Author: Will Deacon <will.deacon@arm.com>
18  */
19 
20 /*
21  * HW_breakpoint: a unified kernel/user-space hardware breakpoint facility,
22  * using the CPU's debug registers.
23  */
24 #define pr_fmt(fmt) "hw-breakpoint: " fmt
25 
26 #include <linux/errno.h>
27 #include <linux/hardirq.h>
28 #include <linux/perf_event.h>
29 #include <linux/hw_breakpoint.h>
30 #include <linux/smp.h>
31 #include <linux/cpu_pm.h>
32 #include <linux/coresight.h>
33 
34 #include <asm/cacheflush.h>
35 #include <asm/cputype.h>
36 #include <asm/current.h>
37 #include <asm/hw_breakpoint.h>
38 #include <asm/traps.h>
39 
40 /* Breakpoint currently in use for each BRP. */
41 static DEFINE_PER_CPU(struct perf_event *, bp_on_reg[ARM_MAX_BRP]);
42 
43 /* Watchpoint currently in use for each WRP. */
44 static DEFINE_PER_CPU(struct perf_event *, wp_on_reg[ARM_MAX_WRP]);
45 
46 /* Number of BRP/WRP registers on this CPU. */
47 static int core_num_brps __ro_after_init;
48 static int core_num_wrps __ro_after_init;
49 
50 /* Debug architecture version. */
51 static u8 debug_arch __ro_after_init;
52 
53 /* Does debug architecture support OS Save and Restore? */
54 static bool has_ossr __ro_after_init;
55 
56 /* Maximum supported watchpoint length. */
57 static u8 max_watchpoint_len __ro_after_init;
58 
59 #define READ_WB_REG_CASE(OP2, M, VAL)			\
60 	case ((OP2 << 4) + M):				\
61 		ARM_DBG_READ(c0, c ## M, OP2, VAL);	\
62 		break
63 
64 #define WRITE_WB_REG_CASE(OP2, M, VAL)			\
65 	case ((OP2 << 4) + M):				\
66 		ARM_DBG_WRITE(c0, c ## M, OP2, VAL);	\
67 		break
68 
69 #define GEN_READ_WB_REG_CASES(OP2, VAL)		\
70 	READ_WB_REG_CASE(OP2, 0, VAL);		\
71 	READ_WB_REG_CASE(OP2, 1, VAL);		\
72 	READ_WB_REG_CASE(OP2, 2, VAL);		\
73 	READ_WB_REG_CASE(OP2, 3, VAL);		\
74 	READ_WB_REG_CASE(OP2, 4, VAL);		\
75 	READ_WB_REG_CASE(OP2, 5, VAL);		\
76 	READ_WB_REG_CASE(OP2, 6, VAL);		\
77 	READ_WB_REG_CASE(OP2, 7, VAL);		\
78 	READ_WB_REG_CASE(OP2, 8, VAL);		\
79 	READ_WB_REG_CASE(OP2, 9, VAL);		\
80 	READ_WB_REG_CASE(OP2, 10, VAL);		\
81 	READ_WB_REG_CASE(OP2, 11, VAL);		\
82 	READ_WB_REG_CASE(OP2, 12, VAL);		\
83 	READ_WB_REG_CASE(OP2, 13, VAL);		\
84 	READ_WB_REG_CASE(OP2, 14, VAL);		\
85 	READ_WB_REG_CASE(OP2, 15, VAL)
86 
87 #define GEN_WRITE_WB_REG_CASES(OP2, VAL)	\
88 	WRITE_WB_REG_CASE(OP2, 0, VAL);		\
89 	WRITE_WB_REG_CASE(OP2, 1, VAL);		\
90 	WRITE_WB_REG_CASE(OP2, 2, VAL);		\
91 	WRITE_WB_REG_CASE(OP2, 3, VAL);		\
92 	WRITE_WB_REG_CASE(OP2, 4, VAL);		\
93 	WRITE_WB_REG_CASE(OP2, 5, VAL);		\
94 	WRITE_WB_REG_CASE(OP2, 6, VAL);		\
95 	WRITE_WB_REG_CASE(OP2, 7, VAL);		\
96 	WRITE_WB_REG_CASE(OP2, 8, VAL);		\
97 	WRITE_WB_REG_CASE(OP2, 9, VAL);		\
98 	WRITE_WB_REG_CASE(OP2, 10, VAL);	\
99 	WRITE_WB_REG_CASE(OP2, 11, VAL);	\
100 	WRITE_WB_REG_CASE(OP2, 12, VAL);	\
101 	WRITE_WB_REG_CASE(OP2, 13, VAL);	\
102 	WRITE_WB_REG_CASE(OP2, 14, VAL);	\
103 	WRITE_WB_REG_CASE(OP2, 15, VAL)
104 
105 static u32 read_wb_reg(int n)
106 {
107 	u32 val = 0;
108 
109 	switch (n) {
110 	GEN_READ_WB_REG_CASES(ARM_OP2_BVR, val);
111 	GEN_READ_WB_REG_CASES(ARM_OP2_BCR, val);
112 	GEN_READ_WB_REG_CASES(ARM_OP2_WVR, val);
113 	GEN_READ_WB_REG_CASES(ARM_OP2_WCR, val);
114 	default:
115 		pr_warn("attempt to read from unknown breakpoint register %d\n",
116 			n);
117 	}
118 
119 	return val;
120 }
121 
122 static void write_wb_reg(int n, u32 val)
123 {
124 	switch (n) {
125 	GEN_WRITE_WB_REG_CASES(ARM_OP2_BVR, val);
126 	GEN_WRITE_WB_REG_CASES(ARM_OP2_BCR, val);
127 	GEN_WRITE_WB_REG_CASES(ARM_OP2_WVR, val);
128 	GEN_WRITE_WB_REG_CASES(ARM_OP2_WCR, val);
129 	default:
130 		pr_warn("attempt to write to unknown breakpoint register %d\n",
131 			n);
132 	}
133 	isb();
134 }
135 
136 /* Determine debug architecture. */
137 static u8 get_debug_arch(void)
138 {
139 	u32 didr;
140 
141 	/* Do we implement the extended CPUID interface? */
142 	if (((read_cpuid_id() >> 16) & 0xf) != 0xf) {
143 		pr_warn_once("CPUID feature registers not supported. "
144 			     "Assuming v6 debug is present.\n");
145 		return ARM_DEBUG_ARCH_V6;
146 	}
147 
148 	ARM_DBG_READ(c0, c0, 0, didr);
149 	return (didr >> 16) & 0xf;
150 }
151 
152 u8 arch_get_debug_arch(void)
153 {
154 	return debug_arch;
155 }
156 
157 static int debug_arch_supported(void)
158 {
159 	u8 arch = get_debug_arch();
160 
161 	/* We don't support the memory-mapped interface. */
162 	return (arch >= ARM_DEBUG_ARCH_V6 && arch <= ARM_DEBUG_ARCH_V7_ECP14) ||
163 		arch >= ARM_DEBUG_ARCH_V7_1;
164 }
165 
166 /* Can we determine the watchpoint access type from the fsr? */
167 static int debug_exception_updates_fsr(void)
168 {
169 	return get_debug_arch() >= ARM_DEBUG_ARCH_V8;
170 }
171 
172 /* Determine number of WRP registers available. */
173 static int get_num_wrp_resources(void)
174 {
175 	u32 didr;
176 	ARM_DBG_READ(c0, c0, 0, didr);
177 	return ((didr >> 28) & 0xf) + 1;
178 }
179 
180 /* Determine number of BRP registers available. */
181 static int get_num_brp_resources(void)
182 {
183 	u32 didr;
184 	ARM_DBG_READ(c0, c0, 0, didr);
185 	return ((didr >> 24) & 0xf) + 1;
186 }
187 
188 /* Does this core support mismatch breakpoints? */
189 static int core_has_mismatch_brps(void)
190 {
191 	return (get_debug_arch() >= ARM_DEBUG_ARCH_V7_ECP14 &&
192 		get_num_brp_resources() > 1);
193 }
194 
195 /* Determine number of usable WRPs available. */
196 static int get_num_wrps(void)
197 {
198 	/*
199 	 * On debug architectures prior to 7.1, when a watchpoint fires, the
200 	 * only way to work out which watchpoint it was is by disassembling
201 	 * the faulting instruction and working out the address of the memory
202 	 * access.
203 	 *
204 	 * Furthermore, we can only do this if the watchpoint was precise
205 	 * since imprecise watchpoints prevent us from calculating register
206 	 * based addresses.
207 	 *
208 	 * Providing we have more than 1 breakpoint register, we only report
209 	 * a single watchpoint register for the time being. This way, we always
210 	 * know which watchpoint fired. In the future we can either add a
211 	 * disassembler and address generation emulator, or we can insert a
212 	 * check to see if the DFAR is set on watchpoint exception entry
213 	 * [the ARM ARM states that the DFAR is UNKNOWN, but experience shows
214 	 * that it is set on some implementations].
215 	 */
216 	if (get_debug_arch() < ARM_DEBUG_ARCH_V7_1)
217 		return 1;
218 
219 	return get_num_wrp_resources();
220 }
221 
222 /* Determine number of usable BRPs available. */
223 static int get_num_brps(void)
224 {
225 	int brps = get_num_brp_resources();
226 	return core_has_mismatch_brps() ? brps - 1 : brps;
227 }
228 
229 /*
230  * In order to access the breakpoint/watchpoint control registers,
231  * we must be running in debug monitor mode. Unfortunately, we can
232  * be put into halting debug mode at any time by an external debugger
233  * but there is nothing we can do to prevent that.
234  */
235 static int monitor_mode_enabled(void)
236 {
237 	u32 dscr;
238 	ARM_DBG_READ(c0, c1, 0, dscr);
239 	return !!(dscr & ARM_DSCR_MDBGEN);
240 }
241 
242 static int enable_monitor_mode(void)
243 {
244 	u32 dscr;
245 	ARM_DBG_READ(c0, c1, 0, dscr);
246 
247 	/* If monitor mode is already enabled, just return. */
248 	if (dscr & ARM_DSCR_MDBGEN)
249 		goto out;
250 
251 	/* Write to the corresponding DSCR. */
252 	switch (get_debug_arch()) {
253 	case ARM_DEBUG_ARCH_V6:
254 	case ARM_DEBUG_ARCH_V6_1:
255 		ARM_DBG_WRITE(c0, c1, 0, (dscr | ARM_DSCR_MDBGEN));
256 		break;
257 	case ARM_DEBUG_ARCH_V7_ECP14:
258 	case ARM_DEBUG_ARCH_V7_1:
259 	case ARM_DEBUG_ARCH_V8:
260 		ARM_DBG_WRITE(c0, c2, 2, (dscr | ARM_DSCR_MDBGEN));
261 		isb();
262 		break;
263 	default:
264 		return -ENODEV;
265 	}
266 
267 	/* Check that the write made it through. */
268 	ARM_DBG_READ(c0, c1, 0, dscr);
269 	if (!(dscr & ARM_DSCR_MDBGEN)) {
270 		pr_warn_once("Failed to enable monitor mode on CPU %d.\n",
271 				smp_processor_id());
272 		return -EPERM;
273 	}
274 
275 out:
276 	return 0;
277 }
278 
279 int hw_breakpoint_slots(int type)
280 {
281 	if (!debug_arch_supported())
282 		return 0;
283 
284 	/*
285 	 * We can be called early, so don't rely on
286 	 * our static variables being initialised.
287 	 */
288 	switch (type) {
289 	case TYPE_INST:
290 		return get_num_brps();
291 	case TYPE_DATA:
292 		return get_num_wrps();
293 	default:
294 		pr_warn("unknown slot type: %d\n", type);
295 		return 0;
296 	}
297 }
298 
299 /*
300  * Check if 8-bit byte-address select is available.
301  * This clobbers WRP 0.
302  */
303 static u8 get_max_wp_len(void)
304 {
305 	u32 ctrl_reg;
306 	struct arch_hw_breakpoint_ctrl ctrl;
307 	u8 size = 4;
308 
309 	if (debug_arch < ARM_DEBUG_ARCH_V7_ECP14)
310 		goto out;
311 
312 	memset(&ctrl, 0, sizeof(ctrl));
313 	ctrl.len = ARM_BREAKPOINT_LEN_8;
314 	ctrl_reg = encode_ctrl_reg(ctrl);
315 
316 	write_wb_reg(ARM_BASE_WVR, 0);
317 	write_wb_reg(ARM_BASE_WCR, ctrl_reg);
318 	if ((read_wb_reg(ARM_BASE_WCR) & ctrl_reg) == ctrl_reg)
319 		size = 8;
320 
321 out:
322 	return size;
323 }
324 
325 u8 arch_get_max_wp_len(void)
326 {
327 	return max_watchpoint_len;
328 }
329 
330 /*
331  * Install a perf counter breakpoint.
332  */
333 int arch_install_hw_breakpoint(struct perf_event *bp)
334 {
335 	struct arch_hw_breakpoint *info = counter_arch_bp(bp);
336 	struct perf_event **slot, **slots;
337 	int i, max_slots, ctrl_base, val_base;
338 	u32 addr, ctrl;
339 
340 	addr = info->address;
341 	ctrl = encode_ctrl_reg(info->ctrl) | 0x1;
342 
343 	if (info->ctrl.type == ARM_BREAKPOINT_EXECUTE) {
344 		/* Breakpoint */
345 		ctrl_base = ARM_BASE_BCR;
346 		val_base = ARM_BASE_BVR;
347 		slots = this_cpu_ptr(bp_on_reg);
348 		max_slots = core_num_brps;
349 	} else {
350 		/* Watchpoint */
351 		ctrl_base = ARM_BASE_WCR;
352 		val_base = ARM_BASE_WVR;
353 		slots = this_cpu_ptr(wp_on_reg);
354 		max_slots = core_num_wrps;
355 	}
356 
357 	for (i = 0; i < max_slots; ++i) {
358 		slot = &slots[i];
359 
360 		if (!*slot) {
361 			*slot = bp;
362 			break;
363 		}
364 	}
365 
366 	if (i == max_slots) {
367 		pr_warn("Can't find any breakpoint slot\n");
368 		return -EBUSY;
369 	}
370 
371 	/* Override the breakpoint data with the step data. */
372 	if (info->step_ctrl.enabled) {
373 		addr = info->trigger & ~0x3;
374 		ctrl = encode_ctrl_reg(info->step_ctrl);
375 		if (info->ctrl.type != ARM_BREAKPOINT_EXECUTE) {
376 			i = 0;
377 			ctrl_base = ARM_BASE_BCR + core_num_brps;
378 			val_base = ARM_BASE_BVR + core_num_brps;
379 		}
380 	}
381 
382 	/* Setup the address register. */
383 	write_wb_reg(val_base + i, addr);
384 
385 	/* Setup the control register. */
386 	write_wb_reg(ctrl_base + i, ctrl);
387 	return 0;
388 }
389 
390 void arch_uninstall_hw_breakpoint(struct perf_event *bp)
391 {
392 	struct arch_hw_breakpoint *info = counter_arch_bp(bp);
393 	struct perf_event **slot, **slots;
394 	int i, max_slots, base;
395 
396 	if (info->ctrl.type == ARM_BREAKPOINT_EXECUTE) {
397 		/* Breakpoint */
398 		base = ARM_BASE_BCR;
399 		slots = this_cpu_ptr(bp_on_reg);
400 		max_slots = core_num_brps;
401 	} else {
402 		/* Watchpoint */
403 		base = ARM_BASE_WCR;
404 		slots = this_cpu_ptr(wp_on_reg);
405 		max_slots = core_num_wrps;
406 	}
407 
408 	/* Remove the breakpoint. */
409 	for (i = 0; i < max_slots; ++i) {
410 		slot = &slots[i];
411 
412 		if (*slot == bp) {
413 			*slot = NULL;
414 			break;
415 		}
416 	}
417 
418 	if (i == max_slots) {
419 		pr_warn("Can't find any breakpoint slot\n");
420 		return;
421 	}
422 
423 	/* Ensure that we disable the mismatch breakpoint. */
424 	if (info->ctrl.type != ARM_BREAKPOINT_EXECUTE &&
425 	    info->step_ctrl.enabled) {
426 		i = 0;
427 		base = ARM_BASE_BCR + core_num_brps;
428 	}
429 
430 	/* Reset the control register. */
431 	write_wb_reg(base + i, 0);
432 }
433 
434 static int get_hbp_len(u8 hbp_len)
435 {
436 	unsigned int len_in_bytes = 0;
437 
438 	switch (hbp_len) {
439 	case ARM_BREAKPOINT_LEN_1:
440 		len_in_bytes = 1;
441 		break;
442 	case ARM_BREAKPOINT_LEN_2:
443 		len_in_bytes = 2;
444 		break;
445 	case ARM_BREAKPOINT_LEN_4:
446 		len_in_bytes = 4;
447 		break;
448 	case ARM_BREAKPOINT_LEN_8:
449 		len_in_bytes = 8;
450 		break;
451 	}
452 
453 	return len_in_bytes;
454 }
455 
456 /*
457  * Check whether bp virtual address is in kernel space.
458  */
459 int arch_check_bp_in_kernelspace(struct arch_hw_breakpoint *hw)
460 {
461 	unsigned int len;
462 	unsigned long va;
463 
464 	va = hw->address;
465 	len = get_hbp_len(hw->ctrl.len);
466 
467 	return (va >= TASK_SIZE) && ((va + len - 1) >= TASK_SIZE);
468 }
469 
470 /*
471  * Extract generic type and length encodings from an arch_hw_breakpoint_ctrl.
472  * Hopefully this will disappear when ptrace can bypass the conversion
473  * to generic breakpoint descriptions.
474  */
475 int arch_bp_generic_fields(struct arch_hw_breakpoint_ctrl ctrl,
476 			   int *gen_len, int *gen_type)
477 {
478 	/* Type */
479 	switch (ctrl.type) {
480 	case ARM_BREAKPOINT_EXECUTE:
481 		*gen_type = HW_BREAKPOINT_X;
482 		break;
483 	case ARM_BREAKPOINT_LOAD:
484 		*gen_type = HW_BREAKPOINT_R;
485 		break;
486 	case ARM_BREAKPOINT_STORE:
487 		*gen_type = HW_BREAKPOINT_W;
488 		break;
489 	case ARM_BREAKPOINT_LOAD | ARM_BREAKPOINT_STORE:
490 		*gen_type = HW_BREAKPOINT_RW;
491 		break;
492 	default:
493 		return -EINVAL;
494 	}
495 
496 	/* Len */
497 	switch (ctrl.len) {
498 	case ARM_BREAKPOINT_LEN_1:
499 		*gen_len = HW_BREAKPOINT_LEN_1;
500 		break;
501 	case ARM_BREAKPOINT_LEN_2:
502 		*gen_len = HW_BREAKPOINT_LEN_2;
503 		break;
504 	case ARM_BREAKPOINT_LEN_4:
505 		*gen_len = HW_BREAKPOINT_LEN_4;
506 		break;
507 	case ARM_BREAKPOINT_LEN_8:
508 		*gen_len = HW_BREAKPOINT_LEN_8;
509 		break;
510 	default:
511 		return -EINVAL;
512 	}
513 
514 	return 0;
515 }
516 
517 /*
518  * Construct an arch_hw_breakpoint from a perf_event.
519  */
520 static int arch_build_bp_info(struct perf_event *bp,
521 			      const struct perf_event_attr *attr,
522 			      struct arch_hw_breakpoint *hw)
523 {
524 	/* Type */
525 	switch (attr->bp_type) {
526 	case HW_BREAKPOINT_X:
527 		hw->ctrl.type = ARM_BREAKPOINT_EXECUTE;
528 		break;
529 	case HW_BREAKPOINT_R:
530 		hw->ctrl.type = ARM_BREAKPOINT_LOAD;
531 		break;
532 	case HW_BREAKPOINT_W:
533 		hw->ctrl.type = ARM_BREAKPOINT_STORE;
534 		break;
535 	case HW_BREAKPOINT_RW:
536 		hw->ctrl.type = ARM_BREAKPOINT_LOAD | ARM_BREAKPOINT_STORE;
537 		break;
538 	default:
539 		return -EINVAL;
540 	}
541 
542 	/* Len */
543 	switch (attr->bp_len) {
544 	case HW_BREAKPOINT_LEN_1:
545 		hw->ctrl.len = ARM_BREAKPOINT_LEN_1;
546 		break;
547 	case HW_BREAKPOINT_LEN_2:
548 		hw->ctrl.len = ARM_BREAKPOINT_LEN_2;
549 		break;
550 	case HW_BREAKPOINT_LEN_4:
551 		hw->ctrl.len = ARM_BREAKPOINT_LEN_4;
552 		break;
553 	case HW_BREAKPOINT_LEN_8:
554 		hw->ctrl.len = ARM_BREAKPOINT_LEN_8;
555 		if ((hw->ctrl.type != ARM_BREAKPOINT_EXECUTE)
556 			&& max_watchpoint_len >= 8)
557 			break;
558 	default:
559 		return -EINVAL;
560 	}
561 
562 	/*
563 	 * Breakpoints must be of length 2 (thumb) or 4 (ARM) bytes.
564 	 * Watchpoints can be of length 1, 2, 4 or 8 bytes if supported
565 	 * by the hardware and must be aligned to the appropriate number of
566 	 * bytes.
567 	 */
568 	if (hw->ctrl.type == ARM_BREAKPOINT_EXECUTE &&
569 	    hw->ctrl.len != ARM_BREAKPOINT_LEN_2 &&
570 	    hw->ctrl.len != ARM_BREAKPOINT_LEN_4)
571 		return -EINVAL;
572 
573 	/* Address */
574 	hw->address = attr->bp_addr;
575 
576 	/* Privilege */
577 	hw->ctrl.privilege = ARM_BREAKPOINT_USER;
578 	if (arch_check_bp_in_kernelspace(hw))
579 		hw->ctrl.privilege |= ARM_BREAKPOINT_PRIV;
580 
581 	/* Enabled? */
582 	hw->ctrl.enabled = !attr->disabled;
583 
584 	/* Mismatch */
585 	hw->ctrl.mismatch = 0;
586 
587 	return 0;
588 }
589 
590 /*
591  * Validate the arch-specific HW Breakpoint register settings.
592  */
593 int hw_breakpoint_arch_parse(struct perf_event *bp,
594 			     const struct perf_event_attr *attr,
595 			     struct arch_hw_breakpoint *hw)
596 {
597 	int ret = 0;
598 	u32 offset, alignment_mask = 0x3;
599 
600 	/* Ensure that we are in monitor debug mode. */
601 	if (!monitor_mode_enabled())
602 		return -ENODEV;
603 
604 	/* Build the arch_hw_breakpoint. */
605 	ret = arch_build_bp_info(bp, attr, hw);
606 	if (ret)
607 		goto out;
608 
609 	/* Check address alignment. */
610 	if (hw->ctrl.len == ARM_BREAKPOINT_LEN_8)
611 		alignment_mask = 0x7;
612 	offset = hw->address & alignment_mask;
613 	switch (offset) {
614 	case 0:
615 		/* Aligned */
616 		break;
617 	case 1:
618 	case 2:
619 		/* Allow halfword watchpoints and breakpoints. */
620 		if (hw->ctrl.len == ARM_BREAKPOINT_LEN_2)
621 			break;
622 	case 3:
623 		/* Allow single byte watchpoint. */
624 		if (hw->ctrl.len == ARM_BREAKPOINT_LEN_1)
625 			break;
626 	default:
627 		ret = -EINVAL;
628 		goto out;
629 	}
630 
631 	hw->address &= ~alignment_mask;
632 	hw->ctrl.len <<= offset;
633 
634 	if (is_default_overflow_handler(bp)) {
635 		/*
636 		 * Mismatch breakpoints are required for single-stepping
637 		 * breakpoints.
638 		 */
639 		if (!core_has_mismatch_brps())
640 			return -EINVAL;
641 
642 		/* We don't allow mismatch breakpoints in kernel space. */
643 		if (arch_check_bp_in_kernelspace(hw))
644 			return -EPERM;
645 
646 		/*
647 		 * Per-cpu breakpoints are not supported by our stepping
648 		 * mechanism.
649 		 */
650 		if (!bp->hw.target)
651 			return -EINVAL;
652 
653 		/*
654 		 * We only support specific access types if the fsr
655 		 * reports them.
656 		 */
657 		if (!debug_exception_updates_fsr() &&
658 		    (hw->ctrl.type == ARM_BREAKPOINT_LOAD ||
659 		     hw->ctrl.type == ARM_BREAKPOINT_STORE))
660 			return -EINVAL;
661 	}
662 
663 out:
664 	return ret;
665 }
666 
667 /*
668  * Enable/disable single-stepping over the breakpoint bp at address addr.
669  */
670 static void enable_single_step(struct perf_event *bp, u32 addr)
671 {
672 	struct arch_hw_breakpoint *info = counter_arch_bp(bp);
673 
674 	arch_uninstall_hw_breakpoint(bp);
675 	info->step_ctrl.mismatch  = 1;
676 	info->step_ctrl.len	  = ARM_BREAKPOINT_LEN_4;
677 	info->step_ctrl.type	  = ARM_BREAKPOINT_EXECUTE;
678 	info->step_ctrl.privilege = info->ctrl.privilege;
679 	info->step_ctrl.enabled	  = 1;
680 	info->trigger		  = addr;
681 	arch_install_hw_breakpoint(bp);
682 }
683 
684 static void disable_single_step(struct perf_event *bp)
685 {
686 	arch_uninstall_hw_breakpoint(bp);
687 	counter_arch_bp(bp)->step_ctrl.enabled = 0;
688 	arch_install_hw_breakpoint(bp);
689 }
690 
691 static void watchpoint_handler(unsigned long addr, unsigned int fsr,
692 			       struct pt_regs *regs)
693 {
694 	int i, access;
695 	u32 val, ctrl_reg, alignment_mask;
696 	struct perf_event *wp, **slots;
697 	struct arch_hw_breakpoint *info;
698 	struct arch_hw_breakpoint_ctrl ctrl;
699 
700 	slots = this_cpu_ptr(wp_on_reg);
701 
702 	for (i = 0; i < core_num_wrps; ++i) {
703 		rcu_read_lock();
704 
705 		wp = slots[i];
706 
707 		if (wp == NULL)
708 			goto unlock;
709 
710 		info = counter_arch_bp(wp);
711 		/*
712 		 * The DFAR is an unknown value on debug architectures prior
713 		 * to 7.1. Since we only allow a single watchpoint on these
714 		 * older CPUs, we can set the trigger to the lowest possible
715 		 * faulting address.
716 		 */
717 		if (debug_arch < ARM_DEBUG_ARCH_V7_1) {
718 			BUG_ON(i > 0);
719 			info->trigger = wp->attr.bp_addr;
720 		} else {
721 			if (info->ctrl.len == ARM_BREAKPOINT_LEN_8)
722 				alignment_mask = 0x7;
723 			else
724 				alignment_mask = 0x3;
725 
726 			/* Check if the watchpoint value matches. */
727 			val = read_wb_reg(ARM_BASE_WVR + i);
728 			if (val != (addr & ~alignment_mask))
729 				goto unlock;
730 
731 			/* Possible match, check the byte address select. */
732 			ctrl_reg = read_wb_reg(ARM_BASE_WCR + i);
733 			decode_ctrl_reg(ctrl_reg, &ctrl);
734 			if (!((1 << (addr & alignment_mask)) & ctrl.len))
735 				goto unlock;
736 
737 			/* Check that the access type matches. */
738 			if (debug_exception_updates_fsr()) {
739 				access = (fsr & ARM_FSR_ACCESS_MASK) ?
740 					  HW_BREAKPOINT_W : HW_BREAKPOINT_R;
741 				if (!(access & hw_breakpoint_type(wp)))
742 					goto unlock;
743 			}
744 
745 			/* We have a winner. */
746 			info->trigger = addr;
747 		}
748 
749 		pr_debug("watchpoint fired: address = 0x%x\n", info->trigger);
750 		perf_bp_event(wp, regs);
751 
752 		/*
753 		 * If no overflow handler is present, insert a temporary
754 		 * mismatch breakpoint so we can single-step over the
755 		 * watchpoint trigger.
756 		 */
757 		if (is_default_overflow_handler(wp))
758 			enable_single_step(wp, instruction_pointer(regs));
759 
760 unlock:
761 		rcu_read_unlock();
762 	}
763 }
764 
765 static void watchpoint_single_step_handler(unsigned long pc)
766 {
767 	int i;
768 	struct perf_event *wp, **slots;
769 	struct arch_hw_breakpoint *info;
770 
771 	slots = this_cpu_ptr(wp_on_reg);
772 
773 	for (i = 0; i < core_num_wrps; ++i) {
774 		rcu_read_lock();
775 
776 		wp = slots[i];
777 
778 		if (wp == NULL)
779 			goto unlock;
780 
781 		info = counter_arch_bp(wp);
782 		if (!info->step_ctrl.enabled)
783 			goto unlock;
784 
785 		/*
786 		 * Restore the original watchpoint if we've completed the
787 		 * single-step.
788 		 */
789 		if (info->trigger != pc)
790 			disable_single_step(wp);
791 
792 unlock:
793 		rcu_read_unlock();
794 	}
795 }
796 
797 static void breakpoint_handler(unsigned long unknown, struct pt_regs *regs)
798 {
799 	int i;
800 	u32 ctrl_reg, val, addr;
801 	struct perf_event *bp, **slots;
802 	struct arch_hw_breakpoint *info;
803 	struct arch_hw_breakpoint_ctrl ctrl;
804 
805 	slots = this_cpu_ptr(bp_on_reg);
806 
807 	/* The exception entry code places the amended lr in the PC. */
808 	addr = regs->ARM_pc;
809 
810 	/* Check the currently installed breakpoints first. */
811 	for (i = 0; i < core_num_brps; ++i) {
812 		rcu_read_lock();
813 
814 		bp = slots[i];
815 
816 		if (bp == NULL)
817 			goto unlock;
818 
819 		info = counter_arch_bp(bp);
820 
821 		/* Check if the breakpoint value matches. */
822 		val = read_wb_reg(ARM_BASE_BVR + i);
823 		if (val != (addr & ~0x3))
824 			goto mismatch;
825 
826 		/* Possible match, check the byte address select to confirm. */
827 		ctrl_reg = read_wb_reg(ARM_BASE_BCR + i);
828 		decode_ctrl_reg(ctrl_reg, &ctrl);
829 		if ((1 << (addr & 0x3)) & ctrl.len) {
830 			info->trigger = addr;
831 			pr_debug("breakpoint fired: address = 0x%x\n", addr);
832 			perf_bp_event(bp, regs);
833 			if (!bp->overflow_handler)
834 				enable_single_step(bp, addr);
835 			goto unlock;
836 		}
837 
838 mismatch:
839 		/* If we're stepping a breakpoint, it can now be restored. */
840 		if (info->step_ctrl.enabled)
841 			disable_single_step(bp);
842 unlock:
843 		rcu_read_unlock();
844 	}
845 
846 	/* Handle any pending watchpoint single-step breakpoints. */
847 	watchpoint_single_step_handler(addr);
848 }
849 
850 /*
851  * Called from either the Data Abort Handler [watchpoint] or the
852  * Prefetch Abort Handler [breakpoint] with interrupts disabled.
853  */
854 static int hw_breakpoint_pending(unsigned long addr, unsigned int fsr,
855 				 struct pt_regs *regs)
856 {
857 	int ret = 0;
858 	u32 dscr;
859 
860 	preempt_disable();
861 
862 	if (interrupts_enabled(regs))
863 		local_irq_enable();
864 
865 	/* We only handle watchpoints and hardware breakpoints. */
866 	ARM_DBG_READ(c0, c1, 0, dscr);
867 
868 	/* Perform perf callbacks. */
869 	switch (ARM_DSCR_MOE(dscr)) {
870 	case ARM_ENTRY_BREAKPOINT:
871 		breakpoint_handler(addr, regs);
872 		break;
873 	case ARM_ENTRY_ASYNC_WATCHPOINT:
874 		WARN(1, "Asynchronous watchpoint exception taken. Debugging results may be unreliable\n");
875 	case ARM_ENTRY_SYNC_WATCHPOINT:
876 		watchpoint_handler(addr, fsr, regs);
877 		break;
878 	default:
879 		ret = 1; /* Unhandled fault. */
880 	}
881 
882 	preempt_enable();
883 
884 	return ret;
885 }
886 
887 /*
888  * One-time initialisation.
889  */
890 static cpumask_t debug_err_mask;
891 
892 static int debug_reg_trap(struct pt_regs *regs, unsigned int instr)
893 {
894 	int cpu = smp_processor_id();
895 
896 	pr_warn("Debug register access (0x%x) caused undefined instruction on CPU %d\n",
897 		instr, cpu);
898 
899 	/* Set the error flag for this CPU and skip the faulting instruction. */
900 	cpumask_set_cpu(cpu, &debug_err_mask);
901 	instruction_pointer(regs) += 4;
902 	return 0;
903 }
904 
905 static struct undef_hook debug_reg_hook = {
906 	.instr_mask	= 0x0fe80f10,
907 	.instr_val	= 0x0e000e10,
908 	.fn		= debug_reg_trap,
909 };
910 
911 /* Does this core support OS Save and Restore? */
912 static bool core_has_os_save_restore(void)
913 {
914 	u32 oslsr;
915 
916 	switch (get_debug_arch()) {
917 	case ARM_DEBUG_ARCH_V7_1:
918 		return true;
919 	case ARM_DEBUG_ARCH_V7_ECP14:
920 		ARM_DBG_READ(c1, c1, 4, oslsr);
921 		if (oslsr & ARM_OSLSR_OSLM0)
922 			return true;
923 	default:
924 		return false;
925 	}
926 }
927 
928 static void reset_ctrl_regs(unsigned int cpu)
929 {
930 	int i, raw_num_brps, err = 0;
931 	u32 val;
932 
933 	/*
934 	 * v7 debug contains save and restore registers so that debug state
935 	 * can be maintained across low-power modes without leaving the debug
936 	 * logic powered up. It is IMPLEMENTATION DEFINED whether we can access
937 	 * the debug registers out of reset, so we must unlock the OS Lock
938 	 * Access Register to avoid taking undefined instruction exceptions
939 	 * later on.
940 	 */
941 	switch (debug_arch) {
942 	case ARM_DEBUG_ARCH_V6:
943 	case ARM_DEBUG_ARCH_V6_1:
944 		/* ARMv6 cores clear the registers out of reset. */
945 		goto out_mdbgen;
946 	case ARM_DEBUG_ARCH_V7_ECP14:
947 		/*
948 		 * Ensure sticky power-down is clear (i.e. debug logic is
949 		 * powered up).
950 		 */
951 		ARM_DBG_READ(c1, c5, 4, val);
952 		if ((val & 0x1) == 0)
953 			err = -EPERM;
954 
955 		if (!has_ossr)
956 			goto clear_vcr;
957 		break;
958 	case ARM_DEBUG_ARCH_V7_1:
959 		/*
960 		 * Ensure the OS double lock is clear.
961 		 */
962 		ARM_DBG_READ(c1, c3, 4, val);
963 		if ((val & 0x1) == 1)
964 			err = -EPERM;
965 		break;
966 	}
967 
968 	if (err) {
969 		pr_warn_once("CPU %d debug is powered down!\n", cpu);
970 		cpumask_or(&debug_err_mask, &debug_err_mask, cpumask_of(cpu));
971 		return;
972 	}
973 
974 	/*
975 	 * Unconditionally clear the OS lock by writing a value
976 	 * other than CS_LAR_KEY to the access register.
977 	 */
978 	ARM_DBG_WRITE(c1, c0, 4, ~CORESIGHT_UNLOCK);
979 	isb();
980 
981 	/*
982 	 * Clear any configured vector-catch events before
983 	 * enabling monitor mode.
984 	 */
985 clear_vcr:
986 	ARM_DBG_WRITE(c0, c7, 0, 0);
987 	isb();
988 
989 	if (cpumask_intersects(&debug_err_mask, cpumask_of(cpu))) {
990 		pr_warn_once("CPU %d failed to disable vector catch\n", cpu);
991 		return;
992 	}
993 
994 	/*
995 	 * The control/value register pairs are UNKNOWN out of reset so
996 	 * clear them to avoid spurious debug events.
997 	 */
998 	raw_num_brps = get_num_brp_resources();
999 	for (i = 0; i < raw_num_brps; ++i) {
1000 		write_wb_reg(ARM_BASE_BCR + i, 0UL);
1001 		write_wb_reg(ARM_BASE_BVR + i, 0UL);
1002 	}
1003 
1004 	for (i = 0; i < core_num_wrps; ++i) {
1005 		write_wb_reg(ARM_BASE_WCR + i, 0UL);
1006 		write_wb_reg(ARM_BASE_WVR + i, 0UL);
1007 	}
1008 
1009 	if (cpumask_intersects(&debug_err_mask, cpumask_of(cpu))) {
1010 		pr_warn_once("CPU %d failed to clear debug register pairs\n", cpu);
1011 		return;
1012 	}
1013 
1014 	/*
1015 	 * Have a crack at enabling monitor mode. We don't actually need
1016 	 * it yet, but reporting an error early is useful if it fails.
1017 	 */
1018 out_mdbgen:
1019 	if (enable_monitor_mode())
1020 		cpumask_or(&debug_err_mask, &debug_err_mask, cpumask_of(cpu));
1021 }
1022 
1023 static int dbg_reset_online(unsigned int cpu)
1024 {
1025 	local_irq_disable();
1026 	reset_ctrl_regs(cpu);
1027 	local_irq_enable();
1028 	return 0;
1029 }
1030 
1031 #ifdef CONFIG_CPU_PM
1032 static int dbg_cpu_pm_notify(struct notifier_block *self, unsigned long action,
1033 			     void *v)
1034 {
1035 	if (action == CPU_PM_EXIT)
1036 		reset_ctrl_regs(smp_processor_id());
1037 
1038 	return NOTIFY_OK;
1039 }
1040 
1041 static struct notifier_block dbg_cpu_pm_nb = {
1042 	.notifier_call = dbg_cpu_pm_notify,
1043 };
1044 
1045 static void __init pm_init(void)
1046 {
1047 	cpu_pm_register_notifier(&dbg_cpu_pm_nb);
1048 }
1049 #else
1050 static inline void pm_init(void)
1051 {
1052 }
1053 #endif
1054 
1055 static int __init arch_hw_breakpoint_init(void)
1056 {
1057 	int ret;
1058 
1059 	debug_arch = get_debug_arch();
1060 
1061 	if (!debug_arch_supported()) {
1062 		pr_info("debug architecture 0x%x unsupported.\n", debug_arch);
1063 		return 0;
1064 	}
1065 
1066 	/*
1067 	 * Scorpion CPUs (at least those in APQ8060) seem to set DBGPRSR.SPD
1068 	 * whenever a WFI is issued, even if the core is not powered down, in
1069 	 * violation of the architecture.  When DBGPRSR.SPD is set, accesses to
1070 	 * breakpoint and watchpoint registers are treated as undefined, so
1071 	 * this results in boot time and runtime failures when these are
1072 	 * accessed and we unexpectedly take a trap.
1073 	 *
1074 	 * It's not clear if/how this can be worked around, so we blacklist
1075 	 * Scorpion CPUs to avoid these issues.
1076 	*/
1077 	if (read_cpuid_part() == ARM_CPU_PART_SCORPION) {
1078 		pr_info("Scorpion CPU detected. Hardware breakpoints and watchpoints disabled\n");
1079 		return 0;
1080 	}
1081 
1082 	has_ossr = core_has_os_save_restore();
1083 
1084 	/* Determine how many BRPs/WRPs are available. */
1085 	core_num_brps = get_num_brps();
1086 	core_num_wrps = get_num_wrps();
1087 
1088 	/*
1089 	 * We need to tread carefully here because DBGSWENABLE may be
1090 	 * driven low on this core and there isn't an architected way to
1091 	 * determine that.
1092 	 */
1093 	cpus_read_lock();
1094 	register_undef_hook(&debug_reg_hook);
1095 
1096 	/*
1097 	 * Register CPU notifier which resets the breakpoint resources. We
1098 	 * assume that a halting debugger will leave the world in a nice state
1099 	 * for us.
1100 	 */
1101 	ret = cpuhp_setup_state_cpuslocked(CPUHP_AP_ONLINE_DYN,
1102 					   "arm/hw_breakpoint:online",
1103 					   dbg_reset_online, NULL);
1104 	unregister_undef_hook(&debug_reg_hook);
1105 	if (WARN_ON(ret < 0) || !cpumask_empty(&debug_err_mask)) {
1106 		core_num_brps = 0;
1107 		core_num_wrps = 0;
1108 		if (ret > 0)
1109 			cpuhp_remove_state_nocalls_cpuslocked(ret);
1110 		cpus_read_unlock();
1111 		return 0;
1112 	}
1113 
1114 	pr_info("found %d " "%s" "breakpoint and %d watchpoint registers.\n",
1115 		core_num_brps, core_has_mismatch_brps() ? "(+1 reserved) " :
1116 		"", core_num_wrps);
1117 
1118 	/* Work out the maximum supported watchpoint length. */
1119 	max_watchpoint_len = get_max_wp_len();
1120 	pr_info("maximum watchpoint size is %u bytes.\n",
1121 			max_watchpoint_len);
1122 
1123 	/* Register debug fault handler. */
1124 	hook_fault_code(FAULT_CODE_DEBUG, hw_breakpoint_pending, SIGTRAP,
1125 			TRAP_HWBKPT, "watchpoint debug exception");
1126 	hook_ifault_code(FAULT_CODE_DEBUG, hw_breakpoint_pending, SIGTRAP,
1127 			TRAP_HWBKPT, "breakpoint debug exception");
1128 	cpus_read_unlock();
1129 
1130 	/* Register PM notifiers. */
1131 	pm_init();
1132 	return 0;
1133 }
1134 arch_initcall(arch_hw_breakpoint_init);
1135 
1136 void hw_breakpoint_pmu_read(struct perf_event *bp)
1137 {
1138 }
1139 
1140 /*
1141  * Dummy function to register with die_notifier.
1142  */
1143 int hw_breakpoint_exceptions_notify(struct notifier_block *unused,
1144 					unsigned long val, void *data)
1145 {
1146 	return NOTIFY_DONE;
1147 }
1148