1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3 * Copyright (C) 2007 Alan Stern
4 * Copyright (C) IBM Corporation, 2009
5 * Copyright (C) 2009, Frederic Weisbecker <fweisbec@gmail.com>
6 *
7 * Thanks to Ingo Molnar for his many suggestions.
8 *
9 * Authors: Alan Stern <stern@rowland.harvard.edu>
10 * K.Prasad <prasad@linux.vnet.ibm.com>
11 * Frederic Weisbecker <fweisbec@gmail.com>
12 */
13
14 /*
15 * HW_breakpoint: a unified kernel/user-space hardware breakpoint facility,
16 * using the CPU's debug registers.
17 * This file contains the arch-independent routines.
18 */
19
20 #include <linux/irqflags.h>
21 #include <linux/kallsyms.h>
22 #include <linux/notifier.h>
23 #include <linux/kprobes.h>
24 #include <linux/kdebug.h>
25 #include <linux/kernel.h>
26 #include <linux/module.h>
27 #include <linux/percpu.h>
28 #include <linux/sched.h>
29 #include <linux/init.h>
30 #include <linux/slab.h>
31 #include <linux/list.h>
32 #include <linux/cpu.h>
33 #include <linux/smp.h>
34 #include <linux/bug.h>
35
36 #include <linux/hw_breakpoint.h>
37 /*
38 * Constraints data
39 */
40 struct bp_cpuinfo {
41 /* Number of pinned cpu breakpoints in a cpu */
42 unsigned int cpu_pinned;
43 /* tsk_pinned[n] is the number of tasks having n+1 breakpoints */
44 unsigned int *tsk_pinned;
45 /* Number of non-pinned cpu/task breakpoints in a cpu */
46 unsigned int flexible; /* XXX: placeholder, see fetch_this_slot() */
47 };
48
49 static DEFINE_PER_CPU(struct bp_cpuinfo, bp_cpuinfo[TYPE_MAX]);
50 static int nr_slots[TYPE_MAX];
51
52 static struct bp_cpuinfo *get_bp_info(int cpu, enum bp_type_idx type)
53 {
54 return per_cpu_ptr(bp_cpuinfo + type, cpu);
55 }
56
57 /* Keep track of the breakpoints attached to tasks */
58 static LIST_HEAD(bp_task_head);
59
60 static int constraints_initialized;
61
62 /* Gather the number of total pinned and un-pinned bp in a cpuset */
63 struct bp_busy_slots {
64 unsigned int pinned;
65 unsigned int flexible;
66 };
67
68 /* Serialize accesses to the above constraints */
69 static DEFINE_MUTEX(nr_bp_mutex);
70
71 __weak int hw_breakpoint_weight(struct perf_event *bp)
72 {
73 return 1;
74 }
75
76 static inline enum bp_type_idx find_slot_idx(u64 bp_type)
77 {
78 if (bp_type & HW_BREAKPOINT_RW)
79 return TYPE_DATA;
80
81 return TYPE_INST;
82 }
83
84 /*
85 * Report the maximum number of pinned breakpoints a task
86 * have in this cpu
87 */
88 static unsigned int max_task_bp_pinned(int cpu, enum bp_type_idx type)
89 {
90 unsigned int *tsk_pinned = get_bp_info(cpu, type)->tsk_pinned;
91 int i;
92
93 for (i = nr_slots[type] - 1; i >= 0; i--) {
94 if (tsk_pinned[i] > 0)
95 return i + 1;
96 }
97
98 return 0;
99 }
100
101 /*
102 * Count the number of breakpoints of the same type and same task.
103 * The given event must be not on the list.
104 */
105 static int task_bp_pinned(int cpu, struct perf_event *bp, enum bp_type_idx type)
106 {
107 struct task_struct *tsk = bp->hw.target;
108 struct perf_event *iter;
109 int count = 0;
110
111 list_for_each_entry(iter, &bp_task_head, hw.bp_list) {
112 if (iter->hw.target == tsk &&
113 find_slot_idx(iter->attr.bp_type) == type &&
114 (iter->cpu < 0 || cpu == iter->cpu))
115 count += hw_breakpoint_weight(iter);
116 }
117
118 return count;
119 }
120
121 static const struct cpumask *cpumask_of_bp(struct perf_event *bp)
122 {
123 if (bp->cpu >= 0)
124 return cpumask_of(bp->cpu);
125 return cpu_possible_mask;
126 }
127
128 /*
129 * Report the number of pinned/un-pinned breakpoints we have in
130 * a given cpu (cpu > -1) or in all of them (cpu = -1).
131 */
132 static void
133 fetch_bp_busy_slots(struct bp_busy_slots *slots, struct perf_event *bp,
134 enum bp_type_idx type)
135 {
136 const struct cpumask *cpumask = cpumask_of_bp(bp);
137 int cpu;
138
139 for_each_cpu(cpu, cpumask) {
140 struct bp_cpuinfo *info = get_bp_info(cpu, type);
141 int nr;
142
143 nr = info->cpu_pinned;
144 if (!bp->hw.target)
145 nr += max_task_bp_pinned(cpu, type);
146 else
147 nr += task_bp_pinned(cpu, bp, type);
148
149 if (nr > slots->pinned)
150 slots->pinned = nr;
151
152 nr = info->flexible;
153 if (nr > slots->flexible)
154 slots->flexible = nr;
155 }
156 }
157
158 /*
159 * For now, continue to consider flexible as pinned, until we can
160 * ensure no flexible event can ever be scheduled before a pinned event
161 * in a same cpu.
162 */
163 static void
164 fetch_this_slot(struct bp_busy_slots *slots, int weight)
165 {
166 slots->pinned += weight;
167 }
168
169 /*
170 * Add a pinned breakpoint for the given task in our constraint table
171 */
172 static void toggle_bp_task_slot(struct perf_event *bp, int cpu,
173 enum bp_type_idx type, int weight)
174 {
175 unsigned int *tsk_pinned = get_bp_info(cpu, type)->tsk_pinned;
176 int old_idx, new_idx;
177
178 old_idx = task_bp_pinned(cpu, bp, type) - 1;
179 new_idx = old_idx + weight;
180
181 if (old_idx >= 0)
182 tsk_pinned[old_idx]--;
183 if (new_idx >= 0)
184 tsk_pinned[new_idx]++;
185 }
186
187 /*
188 * Add/remove the given breakpoint in our constraint table
189 */
190 static void
191 toggle_bp_slot(struct perf_event *bp, bool enable, enum bp_type_idx type,
192 int weight)
193 {
194 const struct cpumask *cpumask = cpumask_of_bp(bp);
195 int cpu;
196
197 if (!enable)
198 weight = -weight;
199
200 /* Pinned counter cpu profiling */
201 if (!bp->hw.target) {
202 get_bp_info(bp->cpu, type)->cpu_pinned += weight;
203 return;
204 }
205
206 /* Pinned counter task profiling */
207 for_each_cpu(cpu, cpumask)
208 toggle_bp_task_slot(bp, cpu, type, weight);
209
210 if (enable)
211 list_add_tail(&bp->hw.bp_list, &bp_task_head);
212 else
213 list_del(&bp->hw.bp_list);
214 }
215
216 __weak int arch_reserve_bp_slot(struct perf_event *bp)
217 {
218 return 0;
219 }
220
221 __weak void arch_release_bp_slot(struct perf_event *bp)
222 {
223 }
224
225 /*
226 * Function to perform processor-specific cleanup during unregistration
227 */
228 __weak void arch_unregister_hw_breakpoint(struct perf_event *bp)
229 {
230 /*
231 * A weak stub function here for those archs that don't define
232 * it inside arch/.../kernel/hw_breakpoint.c
233 */
234 }
235
236 /*
237 * Constraints to check before allowing this new breakpoint counter:
238 *
239 * == Non-pinned counter == (Considered as pinned for now)
240 *
241 * - If attached to a single cpu, check:
242 *
243 * (per_cpu(info->flexible, cpu) || (per_cpu(info->cpu_pinned, cpu)
244 * + max(per_cpu(info->tsk_pinned, cpu)))) < HBP_NUM
245 *
246 * -> If there are already non-pinned counters in this cpu, it means
247 * there is already a free slot for them.
248 * Otherwise, we check that the maximum number of per task
249 * breakpoints (for this cpu) plus the number of per cpu breakpoint
250 * (for this cpu) doesn't cover every registers.
251 *
252 * - If attached to every cpus, check:
253 *
254 * (per_cpu(info->flexible, *) || (max(per_cpu(info->cpu_pinned, *))
255 * + max(per_cpu(info->tsk_pinned, *)))) < HBP_NUM
256 *
257 * -> This is roughly the same, except we check the number of per cpu
258 * bp for every cpu and we keep the max one. Same for the per tasks
259 * breakpoints.
260 *
261 *
262 * == Pinned counter ==
263 *
264 * - If attached to a single cpu, check:
265 *
266 * ((per_cpu(info->flexible, cpu) > 1) + per_cpu(info->cpu_pinned, cpu)
267 * + max(per_cpu(info->tsk_pinned, cpu))) < HBP_NUM
268 *
269 * -> Same checks as before. But now the info->flexible, if any, must keep
270 * one register at least (or they will never be fed).
271 *
272 * - If attached to every cpus, check:
273 *
274 * ((per_cpu(info->flexible, *) > 1) + max(per_cpu(info->cpu_pinned, *))
275 * + max(per_cpu(info->tsk_pinned, *))) < HBP_NUM
276 */
277 static int __reserve_bp_slot(struct perf_event *bp, u64 bp_type)
278 {
279 struct bp_busy_slots slots = {0};
280 enum bp_type_idx type;
281 int weight;
282 int ret;
283
284 /* We couldn't initialize breakpoint constraints on boot */
285 if (!constraints_initialized)
286 return -ENOMEM;
287
288 /* Basic checks */
289 if (bp_type == HW_BREAKPOINT_EMPTY ||
290 bp_type == HW_BREAKPOINT_INVALID)
291 return -EINVAL;
292
293 type = find_slot_idx(bp_type);
294 weight = hw_breakpoint_weight(bp);
295
296 fetch_bp_busy_slots(&slots, bp, type);
297 /*
298 * Simulate the addition of this breakpoint to the constraints
299 * and see the result.
300 */
301 fetch_this_slot(&slots, weight);
302
303 /* Flexible counters need to keep at least one slot */
304 if (slots.pinned + (!!slots.flexible) > nr_slots[type])
305 return -ENOSPC;
306
307 ret = arch_reserve_bp_slot(bp);
308 if (ret)
309 return ret;
310
311 toggle_bp_slot(bp, true, type, weight);
312
313 return 0;
314 }
315
316 int reserve_bp_slot(struct perf_event *bp)
317 {
318 int ret;
319
320 mutex_lock(&nr_bp_mutex);
321
322 ret = __reserve_bp_slot(bp, bp->attr.bp_type);
323
324 mutex_unlock(&nr_bp_mutex);
325
326 return ret;
327 }
328
329 static void __release_bp_slot(struct perf_event *bp, u64 bp_type)
330 {
331 enum bp_type_idx type;
332 int weight;
333
334 arch_release_bp_slot(bp);
335
336 type = find_slot_idx(bp_type);
337 weight = hw_breakpoint_weight(bp);
338 toggle_bp_slot(bp, false, type, weight);
339 }
340
341 void release_bp_slot(struct perf_event *bp)
342 {
343 mutex_lock(&nr_bp_mutex);
344
345 arch_unregister_hw_breakpoint(bp);
346 __release_bp_slot(bp, bp->attr.bp_type);
347
348 mutex_unlock(&nr_bp_mutex);
349 }
350
351 static int __modify_bp_slot(struct perf_event *bp, u64 old_type, u64 new_type)
352 {
353 int err;
354
355 __release_bp_slot(bp, old_type);
356
357 err = __reserve_bp_slot(bp, new_type);
358 if (err) {
359 /*
360 * Reserve the old_type slot back in case
361 * there's no space for the new type.
362 *
363 * This must succeed, because we just released
364 * the old_type slot in the __release_bp_slot
365 * call above. If not, something is broken.
366 */
367 WARN_ON(__reserve_bp_slot(bp, old_type));
368 }
369
370 return err;
371 }
372
373 static int modify_bp_slot(struct perf_event *bp, u64 old_type, u64 new_type)
374 {
375 int ret;
376
377 mutex_lock(&nr_bp_mutex);
378 ret = __modify_bp_slot(bp, old_type, new_type);
379 mutex_unlock(&nr_bp_mutex);
380 return ret;
381 }
382
383 /*
384 * Allow the kernel debugger to reserve breakpoint slots without
385 * taking a lock using the dbg_* variant of for the reserve and
386 * release breakpoint slots.
387 */
388 int dbg_reserve_bp_slot(struct perf_event *bp)
389 {
390 if (mutex_is_locked(&nr_bp_mutex))
391 return -1;
392
393 return __reserve_bp_slot(bp, bp->attr.bp_type);
394 }
395
396 int dbg_release_bp_slot(struct perf_event *bp)
397 {
398 if (mutex_is_locked(&nr_bp_mutex))
399 return -1;
400
401 __release_bp_slot(bp, bp->attr.bp_type);
402
403 return 0;
404 }
405
406 static int hw_breakpoint_parse(struct perf_event *bp,
407 const struct perf_event_attr *attr,
408 struct arch_hw_breakpoint *hw)
409 {
410 int err;
411
412 err = hw_breakpoint_arch_parse(bp, attr, hw);
413 if (err)
414 return err;
415
416 if (arch_check_bp_in_kernelspace(hw)) {
417 if (attr->exclude_kernel)
418 return -EINVAL;
419 /*
420 * Don't let unprivileged users set a breakpoint in the trap
421 * path to avoid trap recursion attacks.
422 */
423 if (!capable(CAP_SYS_ADMIN))
424 return -EPERM;
425 }
426
427 return 0;
428 }
429
430 int register_perf_hw_breakpoint(struct perf_event *bp)
431 {
432 struct arch_hw_breakpoint hw = { };
433 int err;
434
435 err = reserve_bp_slot(bp);
436 if (err)
437 return err;
438
439 err = hw_breakpoint_parse(bp, &bp->attr, &hw);
440 if (err) {
441 release_bp_slot(bp);
442 return err;
443 }
444
445 bp->hw.info = hw;
446
447 return 0;
448 }
449
450 /**
451 * register_user_hw_breakpoint - register a hardware breakpoint for user space
452 * @attr: breakpoint attributes
453 * @triggered: callback to trigger when we hit the breakpoint
454 * @tsk: pointer to 'task_struct' of the process to which the address belongs
455 */
456 struct perf_event *
457 register_user_hw_breakpoint(struct perf_event_attr *attr,
458 perf_overflow_handler_t triggered,
459 void *context,
460 struct task_struct *tsk)
461 {
462 return perf_event_create_kernel_counter(attr, -1, tsk, triggered,
463 context);
464 }
465 EXPORT_SYMBOL_GPL(register_user_hw_breakpoint);
466
467 static void hw_breakpoint_copy_attr(struct perf_event_attr *to,
468 struct perf_event_attr *from)
469 {
470 to->bp_addr = from->bp_addr;
471 to->bp_type = from->bp_type;
472 to->bp_len = from->bp_len;
473 to->disabled = from->disabled;
474 }
475
476 int
477 modify_user_hw_breakpoint_check(struct perf_event *bp, struct perf_event_attr *attr,
478 bool check)
479 {
480 struct arch_hw_breakpoint hw = { };
481 int err;
482
483 err = hw_breakpoint_parse(bp, attr, &hw);
484 if (err)
485 return err;
486
487 if (check) {
488 struct perf_event_attr old_attr;
489
490 old_attr = bp->attr;
491 hw_breakpoint_copy_attr(&old_attr, attr);
492 if (memcmp(&old_attr, attr, sizeof(*attr)))
493 return -EINVAL;
494 }
495
496 if (bp->attr.bp_type != attr->bp_type) {
497 err = modify_bp_slot(bp, bp->attr.bp_type, attr->bp_type);
498 if (err)
499 return err;
500 }
501
502 hw_breakpoint_copy_attr(&bp->attr, attr);
503 bp->hw.info = hw;
504
505 return 0;
506 }
507
508 /**
509 * modify_user_hw_breakpoint - modify a user-space hardware breakpoint
510 * @bp: the breakpoint structure to modify
511 * @attr: new breakpoint attributes
512 */
513 int modify_user_hw_breakpoint(struct perf_event *bp, struct perf_event_attr *attr)
514 {
515 int err;
516
517 /*
518 * modify_user_hw_breakpoint can be invoked with IRQs disabled and hence it
519 * will not be possible to raise IPIs that invoke __perf_event_disable.
520 * So call the function directly after making sure we are targeting the
521 * current task.
522 */
523 if (irqs_disabled() && bp->ctx && bp->ctx->task == current)
524 perf_event_disable_local(bp);
525 else
526 perf_event_disable(bp);
527
528 err = modify_user_hw_breakpoint_check(bp, attr, false);
529
530 if (!bp->attr.disabled)
531 perf_event_enable(bp);
532
533 return err;
534 }
535 EXPORT_SYMBOL_GPL(modify_user_hw_breakpoint);
536
537 /**
538 * unregister_hw_breakpoint - unregister a user-space hardware breakpoint
539 * @bp: the breakpoint structure to unregister
540 */
541 void unregister_hw_breakpoint(struct perf_event *bp)
542 {
543 if (!bp)
544 return;
545 perf_event_release_kernel(bp);
546 }
547 EXPORT_SYMBOL_GPL(unregister_hw_breakpoint);
548
549 /**
550 * register_wide_hw_breakpoint - register a wide breakpoint in the kernel
551 * @attr: breakpoint attributes
552 * @triggered: callback to trigger when we hit the breakpoint
553 *
554 * @return a set of per_cpu pointers to perf events
555 */
556 struct perf_event * __percpu *
557 register_wide_hw_breakpoint(struct perf_event_attr *attr,
558 perf_overflow_handler_t triggered,
559 void *context)
560 {
561 struct perf_event * __percpu *cpu_events, *bp;
562 long err = 0;
563 int cpu;
564
565 cpu_events = alloc_percpu(typeof(*cpu_events));
566 if (!cpu_events)
567 return (void __percpu __force *)ERR_PTR(-ENOMEM);
568
569 get_online_cpus();
570 for_each_online_cpu(cpu) {
571 bp = perf_event_create_kernel_counter(attr, cpu, NULL,
572 triggered, context);
573 if (IS_ERR(bp)) {
574 err = PTR_ERR(bp);
575 break;
576 }
577
578 per_cpu(*cpu_events, cpu) = bp;
579 }
580 put_online_cpus();
581
582 if (likely(!err))
583 return cpu_events;
584
585 unregister_wide_hw_breakpoint(cpu_events);
586 return (void __percpu __force *)ERR_PTR(err);
587 }
588 EXPORT_SYMBOL_GPL(register_wide_hw_breakpoint);
589
590 /**
591 * unregister_wide_hw_breakpoint - unregister a wide breakpoint in the kernel
592 * @cpu_events: the per cpu set of events to unregister
593 */
594 void unregister_wide_hw_breakpoint(struct perf_event * __percpu *cpu_events)
595 {
596 int cpu;
597
598 for_each_possible_cpu(cpu)
599 unregister_hw_breakpoint(per_cpu(*cpu_events, cpu));
600
601 free_percpu(cpu_events);
602 }
603 EXPORT_SYMBOL_GPL(unregister_wide_hw_breakpoint);
604
605 static struct notifier_block hw_breakpoint_exceptions_nb = {
606 .notifier_call = hw_breakpoint_exceptions_notify,
607 /* we need to be notified first */
608 .priority = 0x7fffffff
609 };
610
611 static void bp_perf_event_destroy(struct perf_event *event)
612 {
613 release_bp_slot(event);
614 }
615
616 static int hw_breakpoint_event_init(struct perf_event *bp)
617 {
618 int err;
619
620 if (bp->attr.type != PERF_TYPE_BREAKPOINT)
621 return -ENOENT;
622
623 /*
624 * no branch sampling for breakpoint events
625 */
626 if (has_branch_stack(bp))
627 return -EOPNOTSUPP;
628
629 err = register_perf_hw_breakpoint(bp);
630 if (err)
631 return err;
632
633 bp->destroy = bp_perf_event_destroy;
634
635 return 0;
636 }
637
638 static int hw_breakpoint_add(struct perf_event *bp, int flags)
639 {
640 if (!(flags & PERF_EF_START))
641 bp->hw.state = PERF_HES_STOPPED;
642
643 if (is_sampling_event(bp)) {
644 bp->hw.last_period = bp->hw.sample_period;
645 perf_swevent_set_period(bp);
646 }
647
648 return arch_install_hw_breakpoint(bp);
649 }
650
651 static void hw_breakpoint_del(struct perf_event *bp, int flags)
652 {
653 arch_uninstall_hw_breakpoint(bp);
654 }
655
656 static void hw_breakpoint_start(struct perf_event *bp, int flags)
657 {
658 bp->hw.state = 0;
659 }
660
661 static void hw_breakpoint_stop(struct perf_event *bp, int flags)
662 {
663 bp->hw.state = PERF_HES_STOPPED;
664 }
665
666 static struct pmu perf_breakpoint = {
667 .task_ctx_nr = perf_sw_context, /* could eventually get its own */
668
669 .event_init = hw_breakpoint_event_init,
670 .add = hw_breakpoint_add,
671 .del = hw_breakpoint_del,
672 .start = hw_breakpoint_start,
673 .stop = hw_breakpoint_stop,
674 .read = hw_breakpoint_pmu_read,
675 };
676
677 int __init init_hw_breakpoint(void)
678 {
679 int cpu, err_cpu;
680 int i;
681
682 for (i = 0; i < TYPE_MAX; i++)
683 nr_slots[i] = hw_breakpoint_slots(i);
684
685 for_each_possible_cpu(cpu) {
686 for (i = 0; i < TYPE_MAX; i++) {
687 struct bp_cpuinfo *info = get_bp_info(cpu, i);
688
689 info->tsk_pinned = kcalloc(nr_slots[i], sizeof(int),
690 GFP_KERNEL);
691 if (!info->tsk_pinned)
692 goto err_alloc;
693 }
694 }
695
696 constraints_initialized = 1;
697
698 perf_pmu_register(&perf_breakpoint, "breakpoint", PERF_TYPE_BREAKPOINT);
699
700 return register_die_notifier(&hw_breakpoint_exceptions_nb);
701
702 err_alloc:
703 for_each_possible_cpu(err_cpu) {
704 for (i = 0; i < TYPE_MAX; i++)
705 kfree(get_bp_info(err_cpu, i)->tsk_pinned);
706 if (err_cpu == cpu)
707 break;
708 }
709
710 return -ENOMEM;
711 }
712
713
714
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.