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 as published by
4 * the Free Software Foundation; either version 2 of the License, or
5 * (at your option) any later version.
6 *
7 * This program is distributed in the hope that it will be useful,
8 * but WITHOUT ANY WARRANTY; without even the implied warranty of
9 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
10 * GNU General Public License for more details.
11 *
12 * You should have received a copy of the GNU General Public License
13 * along with this program; if not, write to the Free Software
14 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
15 *
16 * Copyright (C) 2007 Alan Stern
17 * Copyright (C) IBM Corporation, 2009
18 * Copyright (C) 2009, Frederic Weisbecker <fweisbec@gmail.com>
19 *
20 * Thanks to Ingo Molnar for his many suggestions.
21 *
22 * Authors: Alan Stern <stern@rowland.harvard.edu>
23 *          K.Prasad <prasad@linux.vnet.ibm.com>
24 *          Frederic Weisbecker <fweisbec@gmail.com>
25 */
26
27/*
28 * HW_breakpoint: a unified kernel/user-space hardware breakpoint facility,
29 * using the CPU's debug registers.
30 * This file contains the arch-independent routines.
31 */
32
33#include <linux/irqflags.h>
34#include <linux/kallsyms.h>
35#include <linux/notifier.h>
36#include <linux/kprobes.h>
37#include <linux/kdebug.h>
38#include <linux/kernel.h>
39#include <linux/module.h>
40#include <linux/percpu.h>
41#include <linux/sched.h>
42#include <linux/init.h>
43#include <linux/slab.h>
44#include <linux/list.h>
45#include <linux/cpu.h>
46#include <linux/smp.h>
47
48#include <linux/hw_breakpoint.h>
49/*
50 * Constraints data
51 */
52struct bp_cpuinfo {
53	/* Number of pinned cpu breakpoints in a cpu */
54	unsigned int	cpu_pinned;
55	/* tsk_pinned[n] is the number of tasks having n+1 breakpoints */
56	unsigned int	*tsk_pinned;
57	/* Number of non-pinned cpu/task breakpoints in a cpu */
58	unsigned int	flexible; /* XXX: placeholder, see fetch_this_slot() */
59};
60
61static DEFINE_PER_CPU(struct bp_cpuinfo, bp_cpuinfo[TYPE_MAX]);
62static int nr_slots[TYPE_MAX];
63
64static struct bp_cpuinfo *get_bp_info(int cpu, enum bp_type_idx type)
65{
66	return per_cpu_ptr(bp_cpuinfo + type, cpu);
67}
68
69/* Keep track of the breakpoints attached to tasks */
70static LIST_HEAD(bp_task_head);
71
72static int constraints_initialized;
73
74/* Gather the number of total pinned and un-pinned bp in a cpuset */
75struct bp_busy_slots {
76	unsigned int pinned;
77	unsigned int flexible;
78};
79
80/* Serialize accesses to the above constraints */
81static DEFINE_MUTEX(nr_bp_mutex);
82
83__weak int hw_breakpoint_weight(struct perf_event *bp)
84{
85	return 1;
86}
87
88static inline enum bp_type_idx find_slot_idx(struct perf_event *bp)
89{
90	if (bp->attr.bp_type & HW_BREAKPOINT_RW)
91		return TYPE_DATA;
92
93	return TYPE_INST;
94}
95
96/*
97 * Report the maximum number of pinned breakpoints a task
98 * have in this cpu
99 */
100static unsigned int max_task_bp_pinned(int cpu, enum bp_type_idx type)
101{
102	unsigned int *tsk_pinned = get_bp_info(cpu, type)->tsk_pinned;
103	int i;
104
105	for (i = nr_slots[type] - 1; i >= 0; i--) {
106		if (tsk_pinned[i] > 0)
107			return i + 1;
108	}
109
110	return 0;
111}
112
113/*
114 * Count the number of breakpoints of the same type and same task.
115 * The given event must be not on the list.
116 */
117static int task_bp_pinned(int cpu, struct perf_event *bp, enum bp_type_idx type)
118{
119	struct task_struct *tsk = bp->hw.target;
120	struct perf_event *iter;
121	int count = 0;
122
123	list_for_each_entry(iter, &bp_task_head, hw.bp_list) {
124		if (iter->hw.target == tsk &&
125		    find_slot_idx(iter) == type &&
126		    (iter->cpu < 0 || cpu == iter->cpu))
127			count += hw_breakpoint_weight(iter);
128	}
129
130	return count;
131}
132
133static const struct cpumask *cpumask_of_bp(struct perf_event *bp)
134{
135	if (bp->cpu >= 0)
136		return cpumask_of(bp->cpu);
137	return cpu_possible_mask;
138}
139
140/*
141 * Report the number of pinned/un-pinned breakpoints we have in
142 * a given cpu (cpu > -1) or in all of them (cpu = -1).
143 */
144static void
145fetch_bp_busy_slots(struct bp_busy_slots *slots, struct perf_event *bp,
146		    enum bp_type_idx type)
147{
148	const struct cpumask *cpumask = cpumask_of_bp(bp);
149	int cpu;
150
151	for_each_cpu(cpu, cpumask) {
152		struct bp_cpuinfo *info = get_bp_info(cpu, type);
153		int nr;
154
155		nr = info->cpu_pinned;
156		if (!bp->hw.target)
157			nr += max_task_bp_pinned(cpu, type);
158		else
159			nr += task_bp_pinned(cpu, bp, type);
160
161		if (nr > slots->pinned)
162			slots->pinned = nr;
163
164		nr = info->flexible;
165		if (nr > slots->flexible)
166			slots->flexible = nr;
167	}
168}
169
170/*
171 * For now, continue to consider flexible as pinned, until we can
172 * ensure no flexible event can ever be scheduled before a pinned event
173 * in a same cpu.
174 */
175static void
176fetch_this_slot(struct bp_busy_slots *slots, int weight)
177{
178	slots->pinned += weight;
179}
180
181/*
182 * Add a pinned breakpoint for the given task in our constraint table
183 */
184static void toggle_bp_task_slot(struct perf_event *bp, int cpu,
185				enum bp_type_idx type, int weight)
186{
187	unsigned int *tsk_pinned = get_bp_info(cpu, type)->tsk_pinned;
188	int old_idx, new_idx;
189
190	old_idx = task_bp_pinned(cpu, bp, type) - 1;
191	new_idx = old_idx + weight;
192
193	if (old_idx >= 0)
194		tsk_pinned[old_idx]--;
195	if (new_idx >= 0)
196		tsk_pinned[new_idx]++;
197}
198
199/*
200 * Add/remove the given breakpoint in our constraint table
201 */
202static void
203toggle_bp_slot(struct perf_event *bp, bool enable, enum bp_type_idx type,
204	       int weight)
205{
206	const struct cpumask *cpumask = cpumask_of_bp(bp);
207	int cpu;
208
209	if (!enable)
210		weight = -weight;
211
212	/* Pinned counter cpu profiling */
213	if (!bp->hw.target) {
214		get_bp_info(bp->cpu, type)->cpu_pinned += weight;
215		return;
216	}
217
218	/* Pinned counter task profiling */
219	for_each_cpu(cpu, cpumask)
220		toggle_bp_task_slot(bp, cpu, type, weight);
221
222	if (enable)
223		list_add_tail(&bp->hw.bp_list, &bp_task_head);
224	else
225		list_del(&bp->hw.bp_list);
226}
227
228/*
229 * Function to perform processor-specific cleanup during unregistration
230 */
231__weak void arch_unregister_hw_breakpoint(struct perf_event *bp)
232{
233	/*
234	 * A weak stub function here for those archs that don't define
235	 * it inside arch/.../kernel/hw_breakpoint.c
236	 */
237}
238
239/*
240 * Contraints to check before allowing this new breakpoint counter:
241 *
242 *  == Non-pinned counter == (Considered as pinned for now)
243 *
244 *   - If attached to a single cpu, check:
245 *
246 *       (per_cpu(info->flexible, cpu) || (per_cpu(info->cpu_pinned, cpu)
247 *           + max(per_cpu(info->tsk_pinned, cpu)))) < HBP_NUM
248 *
249 *       -> If there are already non-pinned counters in this cpu, it means
250 *          there is already a free slot for them.
251 *          Otherwise, we check that the maximum number of per task
252 *          breakpoints (for this cpu) plus the number of per cpu breakpoint
253 *          (for this cpu) doesn't cover every registers.
254 *
255 *   - If attached to every cpus, check:
256 *
257 *       (per_cpu(info->flexible, *) || (max(per_cpu(info->cpu_pinned, *))
258 *           + max(per_cpu(info->tsk_pinned, *)))) < HBP_NUM
259 *
260 *       -> This is roughly the same, except we check the number of per cpu
261 *          bp for every cpu and we keep the max one. Same for the per tasks
262 *          breakpoints.
263 *
264 *
265 * == Pinned counter ==
266 *
267 *   - If attached to a single cpu, check:
268 *
269 *       ((per_cpu(info->flexible, cpu) > 1) + per_cpu(info->cpu_pinned, cpu)
270 *            + max(per_cpu(info->tsk_pinned, cpu))) < HBP_NUM
271 *
272 *       -> Same checks as before. But now the info->flexible, if any, must keep
273 *          one register at least (or they will never be fed).
274 *
275 *   - If attached to every cpus, check:
276 *
277 *       ((per_cpu(info->flexible, *) > 1) + max(per_cpu(info->cpu_pinned, *))
278 *            + max(per_cpu(info->tsk_pinned, *))) < HBP_NUM
279 */
280static int __reserve_bp_slot(struct perf_event *bp)
281{
282	struct bp_busy_slots slots = {0};
283	enum bp_type_idx type;
284	int weight;
285
286	/* We couldn't initialize breakpoint constraints on boot */
287	if (!constraints_initialized)
288		return -ENOMEM;
289
290	/* Basic checks */
291	if (bp->attr.bp_type == HW_BREAKPOINT_EMPTY ||
292	    bp->attr.bp_type == HW_BREAKPOINT_INVALID)
293		return -EINVAL;
294
295	type = find_slot_idx(bp);
296	weight = hw_breakpoint_weight(bp);
297
298	fetch_bp_busy_slots(&slots, bp, type);
299	/*
300	 * Simulate the addition of this breakpoint to the constraints
301	 * and see the result.
302	 */
303	fetch_this_slot(&slots, weight);
304
305	/* Flexible counters need to keep at least one slot */
306	if (slots.pinned + (!!slots.flexible) > nr_slots[type])
307		return -ENOSPC;
308
309	toggle_bp_slot(bp, true, type, weight);
310
311	return 0;
312}
313
314int reserve_bp_slot(struct perf_event *bp)
315{
316	int ret;
317
318	mutex_lock(&nr_bp_mutex);
319
320	ret = __reserve_bp_slot(bp);
321
322	mutex_unlock(&nr_bp_mutex);
323
324	return ret;
325}
326
327static void __release_bp_slot(struct perf_event *bp)
328{
329	enum bp_type_idx type;
330	int weight;
331
332	type = find_slot_idx(bp);
333	weight = hw_breakpoint_weight(bp);
334	toggle_bp_slot(bp, false, type, weight);
335}
336
337void release_bp_slot(struct perf_event *bp)
338{
339	mutex_lock(&nr_bp_mutex);
340
341	arch_unregister_hw_breakpoint(bp);
342	__release_bp_slot(bp);
343
344	mutex_unlock(&nr_bp_mutex);
345}
346
347/*
348 * Allow the kernel debugger to reserve breakpoint slots without
349 * taking a lock using the dbg_* variant of for the reserve and
350 * release breakpoint slots.
351 */
352int dbg_reserve_bp_slot(struct perf_event *bp)
353{
354	if (mutex_is_locked(&nr_bp_mutex))
355		return -1;
356
357	return __reserve_bp_slot(bp);
358}
359
360int dbg_release_bp_slot(struct perf_event *bp)
361{
362	if (mutex_is_locked(&nr_bp_mutex))
363		return -1;
364
365	__release_bp_slot(bp);
366
367	return 0;
368}
369
370static int validate_hw_breakpoint(struct perf_event *bp)
371{
372	int ret;
373
374	ret = arch_validate_hwbkpt_settings(bp);
375	if (ret)
376		return ret;
377
378	if (arch_check_bp_in_kernelspace(bp)) {
379		if (bp->attr.exclude_kernel)
380			return -EINVAL;
381		/*
382		 * Don't let unprivileged users set a breakpoint in the trap
383		 * path to avoid trap recursion attacks.
384		 */
385		if (!capable(CAP_SYS_ADMIN))
386			return -EPERM;
387	}
388
389	return 0;
390}
391
392int register_perf_hw_breakpoint(struct perf_event *bp)
393{
394	int ret;
395
396	ret = reserve_bp_slot(bp);
397	if (ret)
398		return ret;
399
400	ret = validate_hw_breakpoint(bp);
401
402	/* if arch_validate_hwbkpt_settings() fails then release bp slot */
403	if (ret)
404		release_bp_slot(bp);
405
406	return ret;
407}
408
409/**
410 * register_user_hw_breakpoint - register a hardware breakpoint for user space
411 * @attr: breakpoint attributes
412 * @triggered: callback to trigger when we hit the breakpoint
413 * @tsk: pointer to 'task_struct' of the process to which the address belongs
414 */
415struct perf_event *
416register_user_hw_breakpoint(struct perf_event_attr *attr,
417			    perf_overflow_handler_t triggered,
418			    void *context,
419			    struct task_struct *tsk)
420{
421	return perf_event_create_kernel_counter(attr, -1, tsk, triggered,
422						context);
423}
424EXPORT_SYMBOL_GPL(register_user_hw_breakpoint);
425
426/**
427 * modify_user_hw_breakpoint - modify a user-space hardware breakpoint
428 * @bp: the breakpoint structure to modify
429 * @attr: new breakpoint attributes
430 * @triggered: callback to trigger when we hit the breakpoint
431 * @tsk: pointer to 'task_struct' of the process to which the address belongs
432 */
433int modify_user_hw_breakpoint(struct perf_event *bp, struct perf_event_attr *attr)
434{
435	u64 old_addr = bp->attr.bp_addr;
436	u64 old_len = bp->attr.bp_len;
437	int old_type = bp->attr.bp_type;
438	int err = 0;
439
440	/*
441	 * modify_user_hw_breakpoint can be invoked with IRQs disabled and hence it
442	 * will not be possible to raise IPIs that invoke __perf_event_disable.
443	 * So call the function directly after making sure we are targeting the
444	 * current task.
445	 */
446	if (irqs_disabled() && bp->ctx && bp->ctx->task == current)
447		__perf_event_disable(bp);
448	else
449		perf_event_disable(bp);
450
451	bp->attr.bp_addr = attr->bp_addr;
452	bp->attr.bp_type = attr->bp_type;
453	bp->attr.bp_len = attr->bp_len;
454
455	if (attr->disabled)
456		goto end;
457
458	err = validate_hw_breakpoint(bp);
459	if (!err)
460		perf_event_enable(bp);
461
462	if (err) {
463		bp->attr.bp_addr = old_addr;
464		bp->attr.bp_type = old_type;
465		bp->attr.bp_len = old_len;
466		if (!bp->attr.disabled)
467			perf_event_enable(bp);
468
469		return err;
470	}
471
472end:
473	bp->attr.disabled = attr->disabled;
474
475	return 0;
476}
477EXPORT_SYMBOL_GPL(modify_user_hw_breakpoint);
478
479/**
480 * unregister_hw_breakpoint - unregister a user-space hardware breakpoint
481 * @bp: the breakpoint structure to unregister
482 */
483void unregister_hw_breakpoint(struct perf_event *bp)
484{
485	if (!bp)
486		return;
487	perf_event_release_kernel(bp);
488}
489EXPORT_SYMBOL_GPL(unregister_hw_breakpoint);
490
491/**
492 * register_wide_hw_breakpoint - register a wide breakpoint in the kernel
493 * @attr: breakpoint attributes
494 * @triggered: callback to trigger when we hit the breakpoint
495 *
496 * @return a set of per_cpu pointers to perf events
497 */
498struct perf_event * __percpu *
499register_wide_hw_breakpoint(struct perf_event_attr *attr,
500			    perf_overflow_handler_t triggered,
501			    void *context)
502{
503	struct perf_event * __percpu *cpu_events, *bp;
504	long err = 0;
505	int cpu;
506
507	cpu_events = alloc_percpu(typeof(*cpu_events));
508	if (!cpu_events)
509		return (void __percpu __force *)ERR_PTR(-ENOMEM);
510
511	get_online_cpus();
512	for_each_online_cpu(cpu) {
513		bp = perf_event_create_kernel_counter(attr, cpu, NULL,
514						      triggered, context);
515		if (IS_ERR(bp)) {
516			err = PTR_ERR(bp);
517			break;
518		}
519
520		per_cpu(*cpu_events, cpu) = bp;
521	}
522	put_online_cpus();
523
524	if (likely(!err))
525		return cpu_events;
526
527	unregister_wide_hw_breakpoint(cpu_events);
528	return (void __percpu __force *)ERR_PTR(err);
529}
530EXPORT_SYMBOL_GPL(register_wide_hw_breakpoint);
531
532/**
533 * unregister_wide_hw_breakpoint - unregister a wide breakpoint in the kernel
534 * @cpu_events: the per cpu set of events to unregister
535 */
536void unregister_wide_hw_breakpoint(struct perf_event * __percpu *cpu_events)
537{
538	int cpu;
539
540	for_each_possible_cpu(cpu)
541		unregister_hw_breakpoint(per_cpu(*cpu_events, cpu));
542
543	free_percpu(cpu_events);
544}
545EXPORT_SYMBOL_GPL(unregister_wide_hw_breakpoint);
546
547static struct notifier_block hw_breakpoint_exceptions_nb = {
548	.notifier_call = hw_breakpoint_exceptions_notify,
549	/* we need to be notified first */
550	.priority = 0x7fffffff
551};
552
553static void bp_perf_event_destroy(struct perf_event *event)
554{
555	release_bp_slot(event);
556}
557
558static int hw_breakpoint_event_init(struct perf_event *bp)
559{
560	int err;
561
562	if (bp->attr.type != PERF_TYPE_BREAKPOINT)
563		return -ENOENT;
564
565	/*
566	 * no branch sampling for breakpoint events
567	 */
568	if (has_branch_stack(bp))
569		return -EOPNOTSUPP;
570
571	err = register_perf_hw_breakpoint(bp);
572	if (err)
573		return err;
574
575	bp->destroy = bp_perf_event_destroy;
576
577	return 0;
578}
579
580static int hw_breakpoint_add(struct perf_event *bp, int flags)
581{
582	if (!(flags & PERF_EF_START))
583		bp->hw.state = PERF_HES_STOPPED;
584
585	if (is_sampling_event(bp)) {
586		bp->hw.last_period = bp->hw.sample_period;
587		perf_swevent_set_period(bp);
588	}
589
590	return arch_install_hw_breakpoint(bp);
591}
592
593static void hw_breakpoint_del(struct perf_event *bp, int flags)
594{
595	arch_uninstall_hw_breakpoint(bp);
596}
597
598static void hw_breakpoint_start(struct perf_event *bp, int flags)
599{
600	bp->hw.state = 0;
601}
602
603static void hw_breakpoint_stop(struct perf_event *bp, int flags)
604{
605	bp->hw.state = PERF_HES_STOPPED;
606}
607
608static struct pmu perf_breakpoint = {
609	.task_ctx_nr	= perf_sw_context, /* could eventually get its own */
610
611	.event_init	= hw_breakpoint_event_init,
612	.add		= hw_breakpoint_add,
613	.del		= hw_breakpoint_del,
614	.start		= hw_breakpoint_start,
615	.stop		= hw_breakpoint_stop,
616	.read		= hw_breakpoint_pmu_read,
617};
618
619int __init init_hw_breakpoint(void)
620{
621	int cpu, err_cpu;
622	int i;
623
624	for (i = 0; i < TYPE_MAX; i++)
625		nr_slots[i] = hw_breakpoint_slots(i);
626
627	for_each_possible_cpu(cpu) {
628		for (i = 0; i < TYPE_MAX; i++) {
629			struct bp_cpuinfo *info = get_bp_info(cpu, i);
630
631			info->tsk_pinned = kcalloc(nr_slots[i], sizeof(int),
632							GFP_KERNEL);
633			if (!info->tsk_pinned)
634				goto err_alloc;
635		}
636	}
637
638	constraints_initialized = 1;
639
640	perf_pmu_register(&perf_breakpoint, "breakpoint", PERF_TYPE_BREAKPOINT);
641
642	return register_die_notifier(&hw_breakpoint_exceptions_nb);
643
644 err_alloc:
645	for_each_possible_cpu(err_cpu) {
646		for (i = 0; i < TYPE_MAX; i++)
647			kfree(get_bp_info(err_cpu, i)->tsk_pinned);
648		if (err_cpu == cpu)
649			break;
650	}
651
652	return -ENOMEM;
653}
654
655
656