xref: /openbmc/linux/arch/powerpc/kernel/hw_breakpoint.c (revision 668a6ec6ed57f0248070c490aba75a9572e4b0a4)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * HW_breakpoint: a unified kernel/user-space hardware breakpoint facility,
4  * using the CPU's debug registers. Derived from
5  * "arch/x86/kernel/hw_breakpoint.c"
6  *
7  * Copyright 2010 IBM Corporation
8  * Author: K.Prasad <prasad@linux.vnet.ibm.com>
9  */
10 
11 #include <linux/hw_breakpoint.h>
12 #include <linux/notifier.h>
13 #include <linux/kprobes.h>
14 #include <linux/percpu.h>
15 #include <linux/kernel.h>
16 #include <linux/sched.h>
17 #include <linux/smp.h>
18 #include <linux/spinlock.h>
19 #include <linux/debugfs.h>
20 #include <linux/init.h>
21 
22 #include <asm/hw_breakpoint.h>
23 #include <asm/processor.h>
24 #include <asm/sstep.h>
25 #include <asm/debug.h>
26 #include <asm/hvcall.h>
27 #include <asm/inst.h>
28 #include <linux/uaccess.h>
29 
30 /*
31  * Stores the breakpoints currently in use on each breakpoint address
32  * register for every cpu
33  */
34 static DEFINE_PER_CPU(struct perf_event *, bp_per_reg[HBP_NUM_MAX]);
35 
36 /*
37  * Returns total number of data or instruction breakpoints available.
38  */
39 int hw_breakpoint_slots(int type)
40 {
41 	if (type == TYPE_DATA)
42 		return nr_wp_slots();
43 	return 0;		/* no instruction breakpoints available */
44 }
45 
46 static bool single_step_pending(void)
47 {
48 	int i;
49 
50 	for (i = 0; i < nr_wp_slots(); i++) {
51 		if (current->thread.last_hit_ubp[i])
52 			return true;
53 	}
54 	return false;
55 }
56 
57 /*
58  * Install a perf counter breakpoint.
59  *
60  * We seek a free debug address register and use it for this
61  * breakpoint.
62  *
63  * Atomic: we hold the counter->ctx->lock and we only handle variables
64  * and registers local to this cpu.
65  */
66 int arch_install_hw_breakpoint(struct perf_event *bp)
67 {
68 	struct arch_hw_breakpoint *info = counter_arch_bp(bp);
69 	struct perf_event **slot;
70 	int i;
71 
72 	for (i = 0; i < nr_wp_slots(); i++) {
73 		slot = this_cpu_ptr(&bp_per_reg[i]);
74 		if (!*slot) {
75 			*slot = bp;
76 			break;
77 		}
78 	}
79 
80 	if (WARN_ONCE(i == nr_wp_slots(), "Can't find any breakpoint slot"))
81 		return -EBUSY;
82 
83 	/*
84 	 * Do not install DABR values if the instruction must be single-stepped.
85 	 * If so, DABR will be populated in single_step_dabr_instruction().
86 	 */
87 	if (!single_step_pending())
88 		__set_breakpoint(i, info);
89 
90 	return 0;
91 }
92 
93 /*
94  * Uninstall the breakpoint contained in the given counter.
95  *
96  * First we search the debug address register it uses and then we disable
97  * it.
98  *
99  * Atomic: we hold the counter->ctx->lock and we only handle variables
100  * and registers local to this cpu.
101  */
102 void arch_uninstall_hw_breakpoint(struct perf_event *bp)
103 {
104 	struct arch_hw_breakpoint null_brk = {0};
105 	struct perf_event **slot;
106 	int i;
107 
108 	for (i = 0; i < nr_wp_slots(); i++) {
109 		slot = this_cpu_ptr(&bp_per_reg[i]);
110 		if (*slot == bp) {
111 			*slot = NULL;
112 			break;
113 		}
114 	}
115 
116 	if (WARN_ONCE(i == nr_wp_slots(), "Can't find any breakpoint slot"))
117 		return;
118 
119 	__set_breakpoint(i, &null_brk);
120 }
121 
122 static bool is_ptrace_bp(struct perf_event *bp)
123 {
124 	return bp->overflow_handler == ptrace_triggered;
125 }
126 
127 struct breakpoint {
128 	struct list_head list;
129 	struct perf_event *bp;
130 	bool ptrace_bp;
131 };
132 
133 /*
134  * While kernel/events/hw_breakpoint.c does its own synchronization, we cannot
135  * rely on it safely synchronizing internals here; however, we can rely on it
136  * not requesting more breakpoints than available.
137  */
138 static DEFINE_SPINLOCK(cpu_bps_lock);
139 static DEFINE_PER_CPU(struct breakpoint *, cpu_bps[HBP_NUM_MAX]);
140 static DEFINE_SPINLOCK(task_bps_lock);
141 static LIST_HEAD(task_bps);
142 
143 static struct breakpoint *alloc_breakpoint(struct perf_event *bp)
144 {
145 	struct breakpoint *tmp;
146 
147 	tmp = kzalloc(sizeof(*tmp), GFP_KERNEL);
148 	if (!tmp)
149 		return ERR_PTR(-ENOMEM);
150 	tmp->bp = bp;
151 	tmp->ptrace_bp = is_ptrace_bp(bp);
152 	return tmp;
153 }
154 
155 static bool bp_addr_range_overlap(struct perf_event *bp1, struct perf_event *bp2)
156 {
157 	__u64 bp1_saddr, bp1_eaddr, bp2_saddr, bp2_eaddr;
158 
159 	bp1_saddr = ALIGN_DOWN(bp1->attr.bp_addr, HW_BREAKPOINT_SIZE);
160 	bp1_eaddr = ALIGN(bp1->attr.bp_addr + bp1->attr.bp_len, HW_BREAKPOINT_SIZE);
161 	bp2_saddr = ALIGN_DOWN(bp2->attr.bp_addr, HW_BREAKPOINT_SIZE);
162 	bp2_eaddr = ALIGN(bp2->attr.bp_addr + bp2->attr.bp_len, HW_BREAKPOINT_SIZE);
163 
164 	return (bp1_saddr < bp2_eaddr && bp1_eaddr > bp2_saddr);
165 }
166 
167 static bool alternate_infra_bp(struct breakpoint *b, struct perf_event *bp)
168 {
169 	return is_ptrace_bp(bp) ? !b->ptrace_bp : b->ptrace_bp;
170 }
171 
172 static bool can_co_exist(struct breakpoint *b, struct perf_event *bp)
173 {
174 	return !(alternate_infra_bp(b, bp) && bp_addr_range_overlap(b->bp, bp));
175 }
176 
177 static int task_bps_add(struct perf_event *bp)
178 {
179 	struct breakpoint *tmp;
180 
181 	tmp = alloc_breakpoint(bp);
182 	if (IS_ERR(tmp))
183 		return PTR_ERR(tmp);
184 
185 	spin_lock(&task_bps_lock);
186 	list_add(&tmp->list, &task_bps);
187 	spin_unlock(&task_bps_lock);
188 	return 0;
189 }
190 
191 static void task_bps_remove(struct perf_event *bp)
192 {
193 	struct list_head *pos, *q;
194 
195 	spin_lock(&task_bps_lock);
196 	list_for_each_safe(pos, q, &task_bps) {
197 		struct breakpoint *tmp = list_entry(pos, struct breakpoint, list);
198 
199 		if (tmp->bp == bp) {
200 			list_del(&tmp->list);
201 			kfree(tmp);
202 			break;
203 		}
204 	}
205 	spin_unlock(&task_bps_lock);
206 }
207 
208 /*
209  * If any task has breakpoint from alternate infrastructure,
210  * return true. Otherwise return false.
211  */
212 static bool all_task_bps_check(struct perf_event *bp)
213 {
214 	struct breakpoint *tmp;
215 	bool ret = false;
216 
217 	spin_lock(&task_bps_lock);
218 	list_for_each_entry(tmp, &task_bps, list) {
219 		if (!can_co_exist(tmp, bp)) {
220 			ret = true;
221 			break;
222 		}
223 	}
224 	spin_unlock(&task_bps_lock);
225 	return ret;
226 }
227 
228 /*
229  * If same task has breakpoint from alternate infrastructure,
230  * return true. Otherwise return false.
231  */
232 static bool same_task_bps_check(struct perf_event *bp)
233 {
234 	struct breakpoint *tmp;
235 	bool ret = false;
236 
237 	spin_lock(&task_bps_lock);
238 	list_for_each_entry(tmp, &task_bps, list) {
239 		if (tmp->bp->hw.target == bp->hw.target &&
240 		    !can_co_exist(tmp, bp)) {
241 			ret = true;
242 			break;
243 		}
244 	}
245 	spin_unlock(&task_bps_lock);
246 	return ret;
247 }
248 
249 static int cpu_bps_add(struct perf_event *bp)
250 {
251 	struct breakpoint **cpu_bp;
252 	struct breakpoint *tmp;
253 	int i = 0;
254 
255 	tmp = alloc_breakpoint(bp);
256 	if (IS_ERR(tmp))
257 		return PTR_ERR(tmp);
258 
259 	spin_lock(&cpu_bps_lock);
260 	cpu_bp = per_cpu_ptr(cpu_bps, bp->cpu);
261 	for (i = 0; i < nr_wp_slots(); i++) {
262 		if (!cpu_bp[i]) {
263 			cpu_bp[i] = tmp;
264 			break;
265 		}
266 	}
267 	spin_unlock(&cpu_bps_lock);
268 	return 0;
269 }
270 
271 static void cpu_bps_remove(struct perf_event *bp)
272 {
273 	struct breakpoint **cpu_bp;
274 	int i = 0;
275 
276 	spin_lock(&cpu_bps_lock);
277 	cpu_bp = per_cpu_ptr(cpu_bps, bp->cpu);
278 	for (i = 0; i < nr_wp_slots(); i++) {
279 		if (!cpu_bp[i])
280 			continue;
281 
282 		if (cpu_bp[i]->bp == bp) {
283 			kfree(cpu_bp[i]);
284 			cpu_bp[i] = NULL;
285 			break;
286 		}
287 	}
288 	spin_unlock(&cpu_bps_lock);
289 }
290 
291 static bool cpu_bps_check(int cpu, struct perf_event *bp)
292 {
293 	struct breakpoint **cpu_bp;
294 	bool ret = false;
295 	int i;
296 
297 	spin_lock(&cpu_bps_lock);
298 	cpu_bp = per_cpu_ptr(cpu_bps, cpu);
299 	for (i = 0; i < nr_wp_slots(); i++) {
300 		if (cpu_bp[i] && !can_co_exist(cpu_bp[i], bp)) {
301 			ret = true;
302 			break;
303 		}
304 	}
305 	spin_unlock(&cpu_bps_lock);
306 	return ret;
307 }
308 
309 static bool all_cpu_bps_check(struct perf_event *bp)
310 {
311 	int cpu;
312 
313 	for_each_online_cpu(cpu) {
314 		if (cpu_bps_check(cpu, bp))
315 			return true;
316 	}
317 	return false;
318 }
319 
320 int arch_reserve_bp_slot(struct perf_event *bp)
321 {
322 	int ret;
323 
324 	/* ptrace breakpoint */
325 	if (is_ptrace_bp(bp)) {
326 		if (all_cpu_bps_check(bp))
327 			return -ENOSPC;
328 
329 		if (same_task_bps_check(bp))
330 			return -ENOSPC;
331 
332 		return task_bps_add(bp);
333 	}
334 
335 	/* perf breakpoint */
336 	if (is_kernel_addr(bp->attr.bp_addr))
337 		return 0;
338 
339 	if (bp->hw.target && bp->cpu == -1) {
340 		if (same_task_bps_check(bp))
341 			return -ENOSPC;
342 
343 		return task_bps_add(bp);
344 	} else if (!bp->hw.target && bp->cpu != -1) {
345 		if (all_task_bps_check(bp))
346 			return -ENOSPC;
347 
348 		return cpu_bps_add(bp);
349 	}
350 
351 	if (same_task_bps_check(bp))
352 		return -ENOSPC;
353 
354 	ret = cpu_bps_add(bp);
355 	if (ret)
356 		return ret;
357 	ret = task_bps_add(bp);
358 	if (ret)
359 		cpu_bps_remove(bp);
360 
361 	return ret;
362 }
363 
364 void arch_release_bp_slot(struct perf_event *bp)
365 {
366 	if (!is_kernel_addr(bp->attr.bp_addr)) {
367 		if (bp->hw.target)
368 			task_bps_remove(bp);
369 		if (bp->cpu != -1)
370 			cpu_bps_remove(bp);
371 	}
372 }
373 
374 /*
375  * Perform cleanup of arch-specific counters during unregistration
376  * of the perf-event
377  */
378 void arch_unregister_hw_breakpoint(struct perf_event *bp)
379 {
380 	/*
381 	 * If the breakpoint is unregistered between a hw_breakpoint_handler()
382 	 * and the single_step_dabr_instruction(), then cleanup the breakpoint
383 	 * restoration variables to prevent dangling pointers.
384 	 * FIXME, this should not be using bp->ctx at all! Sayeth peterz.
385 	 */
386 	if (bp->ctx && bp->ctx->task && bp->ctx->task != ((void *)-1L)) {
387 		int i;
388 
389 		for (i = 0; i < nr_wp_slots(); i++) {
390 			if (bp->ctx->task->thread.last_hit_ubp[i] == bp)
391 				bp->ctx->task->thread.last_hit_ubp[i] = NULL;
392 		}
393 	}
394 }
395 
396 /*
397  * Check for virtual address in kernel space.
398  */
399 int arch_check_bp_in_kernelspace(struct arch_hw_breakpoint *hw)
400 {
401 	return is_kernel_addr(hw->address);
402 }
403 
404 int arch_bp_generic_fields(int type, int *gen_bp_type)
405 {
406 	*gen_bp_type = 0;
407 	if (type & HW_BRK_TYPE_READ)
408 		*gen_bp_type |= HW_BREAKPOINT_R;
409 	if (type & HW_BRK_TYPE_WRITE)
410 		*gen_bp_type |= HW_BREAKPOINT_W;
411 	if (*gen_bp_type == 0)
412 		return -EINVAL;
413 	return 0;
414 }
415 
416 /*
417  * Watchpoint match range is always doubleword(8 bytes) aligned on
418  * powerpc. If the given range is crossing doubleword boundary, we
419  * need to increase the length such that next doubleword also get
420  * covered. Ex,
421  *
422  *          address   len = 6 bytes
423  *                |=========.
424  *   |------------v--|------v--------|
425  *   | | | | | | | | | | | | | | | | |
426  *   |---------------|---------------|
427  *    <---8 bytes--->
428  *
429  * In this case, we should configure hw as:
430  *   start_addr = address & ~(HW_BREAKPOINT_SIZE - 1)
431  *   len = 16 bytes
432  *
433  * @start_addr is inclusive but @end_addr is exclusive.
434  */
435 static int hw_breakpoint_validate_len(struct arch_hw_breakpoint *hw)
436 {
437 	u16 max_len = DABR_MAX_LEN;
438 	u16 hw_len;
439 	unsigned long start_addr, end_addr;
440 
441 	start_addr = ALIGN_DOWN(hw->address, HW_BREAKPOINT_SIZE);
442 	end_addr = ALIGN(hw->address + hw->len, HW_BREAKPOINT_SIZE);
443 	hw_len = end_addr - start_addr;
444 
445 	if (dawr_enabled()) {
446 		max_len = DAWR_MAX_LEN;
447 		/* DAWR region can't cross 512 bytes boundary on p10 predecessors */
448 		if (!cpu_has_feature(CPU_FTR_ARCH_31) &&
449 		    (ALIGN_DOWN(start_addr, SZ_512) != ALIGN_DOWN(end_addr - 1, SZ_512)))
450 			return -EINVAL;
451 	} else if (IS_ENABLED(CONFIG_PPC_8xx)) {
452 		/* 8xx can setup a range without limitation */
453 		max_len = U16_MAX;
454 	}
455 
456 	if (hw_len > max_len)
457 		return -EINVAL;
458 
459 	hw->hw_len = hw_len;
460 	return 0;
461 }
462 
463 /*
464  * Validate the arch-specific HW Breakpoint register settings
465  */
466 int hw_breakpoint_arch_parse(struct perf_event *bp,
467 			     const struct perf_event_attr *attr,
468 			     struct arch_hw_breakpoint *hw)
469 {
470 	int ret = -EINVAL;
471 
472 	if (!bp || !attr->bp_len)
473 		return ret;
474 
475 	hw->type = HW_BRK_TYPE_TRANSLATE;
476 	if (attr->bp_type & HW_BREAKPOINT_R)
477 		hw->type |= HW_BRK_TYPE_READ;
478 	if (attr->bp_type & HW_BREAKPOINT_W)
479 		hw->type |= HW_BRK_TYPE_WRITE;
480 	if (hw->type == HW_BRK_TYPE_TRANSLATE)
481 		/* must set alteast read or write */
482 		return ret;
483 	if (!attr->exclude_user)
484 		hw->type |= HW_BRK_TYPE_USER;
485 	if (!attr->exclude_kernel)
486 		hw->type |= HW_BRK_TYPE_KERNEL;
487 	if (!attr->exclude_hv)
488 		hw->type |= HW_BRK_TYPE_HYP;
489 	hw->address = attr->bp_addr;
490 	hw->len = attr->bp_len;
491 
492 	if (!ppc_breakpoint_available())
493 		return -ENODEV;
494 
495 	return hw_breakpoint_validate_len(hw);
496 }
497 
498 /*
499  * Restores the breakpoint on the debug registers.
500  * Invoke this function if it is known that the execution context is
501  * about to change to cause loss of MSR_SE settings.
502  *
503  * The perf watchpoint will simply re-trigger once the thread is started again,
504  * and the watchpoint handler will set up MSR_SE and perf_single_step as
505  * needed.
506  */
507 void thread_change_pc(struct task_struct *tsk, struct pt_regs *regs)
508 {
509 	struct arch_hw_breakpoint *info;
510 	int i;
511 
512 	for (i = 0; i < nr_wp_slots(); i++) {
513 		if (unlikely(tsk->thread.last_hit_ubp[i]))
514 			goto reset;
515 	}
516 	return;
517 
518 reset:
519 	regs_set_return_msr(regs, regs->msr & ~MSR_SE);
520 	for (i = 0; i < nr_wp_slots(); i++) {
521 		info = counter_arch_bp(__this_cpu_read(bp_per_reg[i]));
522 		__set_breakpoint(i, info);
523 		tsk->thread.last_hit_ubp[i] = NULL;
524 	}
525 }
526 
527 static bool is_larx_stcx_instr(int type)
528 {
529 	return type == LARX || type == STCX;
530 }
531 
532 static bool is_octword_vsx_instr(int type, int size)
533 {
534 	return ((type == LOAD_VSX || type == STORE_VSX) && size == 32);
535 }
536 
537 /*
538  * We've failed in reliably handling the hw-breakpoint. Unregister
539  * it and throw a warning message to let the user know about it.
540  */
541 static void handler_error(struct perf_event *bp)
542 {
543 	WARN(1, "Unable to handle hardware breakpoint. Breakpoint at 0x%lx will be disabled.",
544 	     counter_arch_bp(bp)->address);
545 	perf_event_disable_inatomic(bp);
546 }
547 
548 static void larx_stcx_err(struct perf_event *bp)
549 {
550 	printk_ratelimited("Breakpoint hit on instruction that can't be emulated. Breakpoint at 0x%lx will be disabled.\n",
551 			   counter_arch_bp(bp)->address);
552 	perf_event_disable_inatomic(bp);
553 }
554 
555 static bool stepping_handler(struct pt_regs *regs, struct perf_event **bp,
556 			     int *hit, ppc_inst_t instr)
557 {
558 	int i;
559 	int stepped;
560 
561 	/* Do not emulate user-space instructions, instead single-step them */
562 	if (user_mode(regs)) {
563 		for (i = 0; i < nr_wp_slots(); i++) {
564 			if (!hit[i])
565 				continue;
566 			current->thread.last_hit_ubp[i] = bp[i];
567 			bp[i] = NULL;
568 		}
569 		regs_set_return_msr(regs, regs->msr | MSR_SE);
570 		return false;
571 	}
572 
573 	stepped = emulate_step(regs, instr);
574 	if (!stepped) {
575 		for (i = 0; i < nr_wp_slots(); i++) {
576 			if (!hit[i])
577 				continue;
578 			handler_error(bp[i]);
579 			bp[i] = NULL;
580 		}
581 		return false;
582 	}
583 	return true;
584 }
585 
586 static void handle_p10dd1_spurious_exception(struct perf_event **bp,
587 					     int *hit, unsigned long ea)
588 {
589 	int i;
590 	unsigned long hw_end_addr;
591 
592 	/*
593 	 * Handle spurious exception only when any bp_per_reg is set.
594 	 * Otherwise this might be created by xmon and not actually a
595 	 * spurious exception.
596 	 */
597 	for (i = 0; i < nr_wp_slots(); i++) {
598 		struct arch_hw_breakpoint *info;
599 
600 		if (!bp[i])
601 			continue;
602 
603 		info = counter_arch_bp(bp[i]);
604 
605 		hw_end_addr = ALIGN(info->address + info->len, HW_BREAKPOINT_SIZE);
606 
607 		/*
608 		 * Ending address of DAWR range is less than starting
609 		 * address of op.
610 		 */
611 		if ((hw_end_addr - 1) >= ea)
612 			continue;
613 
614 		/*
615 		 * Those addresses need to be in the same or in two
616 		 * consecutive 512B blocks;
617 		 */
618 		if (((hw_end_addr - 1) >> 10) != (ea >> 10))
619 			continue;
620 
621 		/*
622 		 * 'op address + 64B' generates an address that has a
623 		 * carry into bit 52 (crosses 2K boundary).
624 		 */
625 		if ((ea & 0x800) == ((ea + 64) & 0x800))
626 			continue;
627 
628 		break;
629 	}
630 
631 	if (i == nr_wp_slots())
632 		return;
633 
634 	for (i = 0; i < nr_wp_slots(); i++) {
635 		if (bp[i]) {
636 			hit[i] = 1;
637 			counter_arch_bp(bp[i])->type |= HW_BRK_TYPE_EXTRANEOUS_IRQ;
638 		}
639 	}
640 }
641 
642 int hw_breakpoint_handler(struct die_args *args)
643 {
644 	bool err = false;
645 	int rc = NOTIFY_STOP;
646 	struct perf_event *bp[HBP_NUM_MAX] = { NULL };
647 	struct pt_regs *regs = args->regs;
648 	int i;
649 	int hit[HBP_NUM_MAX] = {0};
650 	int nr_hit = 0;
651 	bool ptrace_bp = false;
652 	ppc_inst_t instr = ppc_inst(0);
653 	int type = 0;
654 	int size = 0;
655 	unsigned long ea = 0;
656 
657 	/* Disable breakpoints during exception handling */
658 	hw_breakpoint_disable();
659 
660 	/*
661 	 * The counter may be concurrently released but that can only
662 	 * occur from a call_rcu() path. We can then safely fetch
663 	 * the breakpoint, use its callback, touch its counter
664 	 * while we are in an rcu_read_lock() path.
665 	 */
666 	rcu_read_lock();
667 
668 	if (!IS_ENABLED(CONFIG_PPC_8xx))
669 		wp_get_instr_detail(regs, &instr, &type, &size, &ea);
670 
671 	for (i = 0; i < nr_wp_slots(); i++) {
672 		struct arch_hw_breakpoint *info;
673 
674 		bp[i] = __this_cpu_read(bp_per_reg[i]);
675 		if (!bp[i])
676 			continue;
677 
678 		info = counter_arch_bp(bp[i]);
679 		info->type &= ~HW_BRK_TYPE_EXTRANEOUS_IRQ;
680 
681 		if (wp_check_constraints(regs, instr, ea, type, size, info)) {
682 			if (!IS_ENABLED(CONFIG_PPC_8xx) &&
683 			    ppc_inst_equal(instr, ppc_inst(0))) {
684 				handler_error(bp[i]);
685 				bp[i] = NULL;
686 				err = 1;
687 				continue;
688 			}
689 
690 			if (is_ptrace_bp(bp[i]))
691 				ptrace_bp = true;
692 			hit[i] = 1;
693 			nr_hit++;
694 		}
695 	}
696 
697 	if (err)
698 		goto reset;
699 
700 	if (!nr_hit) {
701 		/* Workaround for Power10 DD1 */
702 		if (!IS_ENABLED(CONFIG_PPC_8xx) && mfspr(SPRN_PVR) == 0x800100 &&
703 		    is_octword_vsx_instr(type, size)) {
704 			handle_p10dd1_spurious_exception(bp, hit, ea);
705 		} else {
706 			rc = NOTIFY_DONE;
707 			goto out;
708 		}
709 	}
710 
711 	/*
712 	 * Return early after invoking user-callback function without restoring
713 	 * DABR if the breakpoint is from ptrace which always operates in
714 	 * one-shot mode. The ptrace-ed process will receive the SIGTRAP signal
715 	 * generated in do_dabr().
716 	 */
717 	if (ptrace_bp) {
718 		for (i = 0; i < nr_wp_slots(); i++) {
719 			if (!hit[i])
720 				continue;
721 			perf_bp_event(bp[i], regs);
722 			bp[i] = NULL;
723 		}
724 		rc = NOTIFY_DONE;
725 		goto reset;
726 	}
727 
728 	if (!IS_ENABLED(CONFIG_PPC_8xx)) {
729 		if (is_larx_stcx_instr(type)) {
730 			for (i = 0; i < nr_wp_slots(); i++) {
731 				if (!hit[i])
732 					continue;
733 				larx_stcx_err(bp[i]);
734 				bp[i] = NULL;
735 			}
736 			goto reset;
737 		}
738 
739 		if (!stepping_handler(regs, bp, hit, instr))
740 			goto reset;
741 	}
742 
743 	/*
744 	 * As a policy, the callback is invoked in a 'trigger-after-execute'
745 	 * fashion
746 	 */
747 	for (i = 0; i < nr_wp_slots(); i++) {
748 		if (!hit[i])
749 			continue;
750 		if (!(counter_arch_bp(bp[i])->type & HW_BRK_TYPE_EXTRANEOUS_IRQ))
751 			perf_bp_event(bp[i], regs);
752 	}
753 
754 reset:
755 	for (i = 0; i < nr_wp_slots(); i++) {
756 		if (!bp[i])
757 			continue;
758 		__set_breakpoint(i, counter_arch_bp(bp[i]));
759 	}
760 
761 out:
762 	rcu_read_unlock();
763 	return rc;
764 }
765 NOKPROBE_SYMBOL(hw_breakpoint_handler);
766 
767 /*
768  * Handle single-step exceptions following a DABR hit.
769  */
770 static int single_step_dabr_instruction(struct die_args *args)
771 {
772 	struct pt_regs *regs = args->regs;
773 	struct perf_event *bp = NULL;
774 	struct arch_hw_breakpoint *info;
775 	int i;
776 	bool found = false;
777 
778 	/*
779 	 * Check if we are single-stepping as a result of a
780 	 * previous HW Breakpoint exception
781 	 */
782 	for (i = 0; i < nr_wp_slots(); i++) {
783 		bp = current->thread.last_hit_ubp[i];
784 
785 		if (!bp)
786 			continue;
787 
788 		found = true;
789 		info = counter_arch_bp(bp);
790 
791 		/*
792 		 * We shall invoke the user-defined callback function in the
793 		 * single stepping handler to confirm to 'trigger-after-execute'
794 		 * semantics
795 		 */
796 		if (!(info->type & HW_BRK_TYPE_EXTRANEOUS_IRQ))
797 			perf_bp_event(bp, regs);
798 		current->thread.last_hit_ubp[i] = NULL;
799 	}
800 
801 	if (!found)
802 		return NOTIFY_DONE;
803 
804 	for (i = 0; i < nr_wp_slots(); i++) {
805 		bp = __this_cpu_read(bp_per_reg[i]);
806 		if (!bp)
807 			continue;
808 
809 		info = counter_arch_bp(bp);
810 		__set_breakpoint(i, info);
811 	}
812 
813 	/*
814 	 * If the process was being single-stepped by ptrace, let the
815 	 * other single-step actions occur (e.g. generate SIGTRAP).
816 	 */
817 	if (test_thread_flag(TIF_SINGLESTEP))
818 		return NOTIFY_DONE;
819 
820 	return NOTIFY_STOP;
821 }
822 NOKPROBE_SYMBOL(single_step_dabr_instruction);
823 
824 /*
825  * Handle debug exception notifications.
826  */
827 int hw_breakpoint_exceptions_notify(
828 		struct notifier_block *unused, unsigned long val, void *data)
829 {
830 	int ret = NOTIFY_DONE;
831 
832 	switch (val) {
833 	case DIE_DABR_MATCH:
834 		ret = hw_breakpoint_handler(data);
835 		break;
836 	case DIE_SSTEP:
837 		ret = single_step_dabr_instruction(data);
838 		break;
839 	}
840 
841 	return ret;
842 }
843 NOKPROBE_SYMBOL(hw_breakpoint_exceptions_notify);
844 
845 /*
846  * Release the user breakpoints used by ptrace
847  */
848 void flush_ptrace_hw_breakpoint(struct task_struct *tsk)
849 {
850 	int i;
851 	struct thread_struct *t = &tsk->thread;
852 
853 	for (i = 0; i < nr_wp_slots(); i++) {
854 		unregister_hw_breakpoint(t->ptrace_bps[i]);
855 		t->ptrace_bps[i] = NULL;
856 	}
857 }
858 
859 void hw_breakpoint_pmu_read(struct perf_event *bp)
860 {
861 	/* TODO */
862 }
863 
864 void ptrace_triggered(struct perf_event *bp,
865 		      struct perf_sample_data *data, struct pt_regs *regs)
866 {
867 	struct perf_event_attr attr;
868 
869 	/*
870 	 * Disable the breakpoint request here since ptrace has defined a
871 	 * one-shot behaviour for breakpoint exceptions in PPC64.
872 	 * The SIGTRAP signal is generated automatically for us in do_dabr().
873 	 * We don't have to do anything about that here
874 	 */
875 	attr = bp->attr;
876 	attr.disabled = true;
877 	modify_user_hw_breakpoint(bp, &attr);
878 }
879