projects/linux·blob master · refs

1025 lines · 26.7 KiB

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 <[email protected]>
6 *
7 * Thanks to Ingo Molnar for his many suggestions.
8 *
9 * Authors: Alan Stern <[email protected]>
10 * K.Prasad <[email protected]>
11 * Frederic Weisbecker <[email protected]>
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/hw_breakpoint.h>
21
22#include <linux/atomic.h>
23#include <linux/bug.h>
24#include <linux/cpu.h>
25#include <linux/export.h>
26#include <linux/init.h>
27#include <linux/irqflags.h>
28#include <linux/kdebug.h>
29#include <linux/kernel.h>
30#include <linux/mutex.h>
31#include <linux/notifier.h>
32#include <linux/percpu-rwsem.h>
33#include <linux/percpu.h>
34#include <linux/rhashtable.h>
35#include <linux/sched.h>
36#include <linux/slab.h>
37
38/*
39 * Datastructure to track the total uses of N slots across tasks or CPUs;
40 * bp_slots_histogram::count[N] is the number of assigned N+1 breakpoint slots.
41 */
42struct bp_slots_histogram {
43#ifdef hw_breakpoint_slots
44 atomic_t count[hw_breakpoint_slots(0)];
45#else
46 atomic_t *count;
47#endif
48};
49
50/*
51 * Per-CPU constraints data.
52 */
53struct bp_cpuinfo {
54 /* Number of pinned CPU breakpoints in a CPU. */
55 unsigned int cpu_pinned;
56 /* Histogram of pinned task breakpoints in a CPU. */
57 struct bp_slots_histogram tsk_pinned;
58};
59
60static DEFINE_PER_CPU(struct bp_cpuinfo, bp_cpuinfo[TYPE_MAX]);
61
62static struct bp_cpuinfo *get_bp_info(int cpu, enum bp_type_idx type)
63{
64 return per_cpu_ptr(bp_cpuinfo + type, cpu);
65}
66
67/* Number of pinned CPU breakpoints globally. */
68static struct bp_slots_histogram cpu_pinned[TYPE_MAX];
69/* Number of pinned CPU-independent task breakpoints. */
70static struct bp_slots_histogram tsk_pinned_all[TYPE_MAX];
71
72/* Keep track of the breakpoints attached to tasks */
73static struct rhltable task_bps_ht;
74static const struct rhashtable_params task_bps_ht_params = {
75 .head_offset = offsetof(struct hw_perf_event, bp_list),
76 .key_offset = offsetof(struct hw_perf_event, target),
77 .key_len = sizeof_field(struct hw_perf_event, target),
78 .automatic_shrinking = true,
79};
80
81static bool constraints_initialized __ro_after_init;
82
83/*
84 * Synchronizes accesses to the per-CPU constraints; the locking rules are:
85 *
86 * 1. Atomic updates to bp_cpuinfo::tsk_pinned only require a held read-lock
87 * (due to bp_slots_histogram::count being atomic, no update are lost).
88 *
89 * 2. Holding a write-lock is required for computations that require a
90 * stable snapshot of all bp_cpuinfo::tsk_pinned.
91 *
92 * 3. In all other cases, non-atomic accesses require the appropriately held
93 * lock (read-lock for read-only accesses; write-lock for reads/writes).
94 */
95DEFINE_STATIC_PERCPU_RWSEM(bp_cpuinfo_sem);
96
97/*
98 * Return mutex to serialize accesses to per-task lists in task_bps_ht. Since
99 * rhltable synchronizes concurrent insertions/deletions, independent tasks may
100 * insert/delete concurrently; therefore, a mutex per task is sufficient.
101 *
102 * Uses task_struct::perf_event_mutex, to avoid extending task_struct with a
103 * hw_breakpoint-only mutex, which may be infrequently used. The caveat here is
104 * that hw_breakpoint may contend with per-task perf event list management. The
105 * assumption is that perf usecases involving hw_breakpoints are very unlikely
106 * to result in unnecessary contention.
107 */
108static inline struct mutex *get_task_bps_mutex(struct perf_event *bp)
109{
110 struct task_struct *tsk = bp->hw.target;
111
112 return tsk ? &tsk->perf_event_mutex : NULL;
113}
114
115static struct mutex *bp_constraints_lock(struct perf_event *bp)
116{
117 struct mutex *tsk_mtx = get_task_bps_mutex(bp);
118
119 if (tsk_mtx) {
120 /*
121 * Fully analogous to the perf_try_init_event() nesting
122 * argument in the comment near perf_event_ctx_lock_nested();
123 * this child->perf_event_mutex cannot ever deadlock against
124 * the parent->perf_event_mutex usage from
125 * perf_event_task_{en,dis}able().
126 *
127 * Specifically, inherited events will never occur on
128 * ->perf_event_list.
129 */
130 mutex_lock_nested(tsk_mtx, SINGLE_DEPTH_NESTING);
131 percpu_down_read(&bp_cpuinfo_sem);
132 } else {
133 percpu_down_write(&bp_cpuinfo_sem);
134 }
135
136 return tsk_mtx;
137}
138
139static void bp_constraints_unlock(struct mutex *tsk_mtx)
140{
141 if (tsk_mtx) {
142 percpu_up_read(&bp_cpuinfo_sem);
143 mutex_unlock(tsk_mtx);
144 } else {
145 percpu_up_write(&bp_cpuinfo_sem);
146 }
147}
148
149static bool bp_constraints_is_locked(struct perf_event *bp)
150{
151 struct mutex *tsk_mtx = get_task_bps_mutex(bp);
152
153 return percpu_is_write_locked(&bp_cpuinfo_sem) ||
154 (tsk_mtx ? mutex_is_locked(tsk_mtx) :
155 percpu_is_read_locked(&bp_cpuinfo_sem));
156}
157
158static inline void assert_bp_constraints_lock_held(struct perf_event *bp)
159{
160 struct mutex *tsk_mtx = get_task_bps_mutex(bp);
161
162 if (tsk_mtx)
163 lockdep_assert_held(tsk_mtx);
164 lockdep_assert_held(&bp_cpuinfo_sem);
165}
166
167#ifdef hw_breakpoint_slots
168/*
169 * Number of breakpoint slots is constant, and the same for all types.
170 */
171static_assert(hw_breakpoint_slots(TYPE_INST) == hw_breakpoint_slots(TYPE_DATA));
172static inline int hw_breakpoint_slots_cached(int type) { return hw_breakpoint_slots(type); }
173static inline int init_breakpoint_slots(void) { return 0; }
174#else
175/*
176 * Dynamic number of breakpoint slots.
177 */
178static int __nr_bp_slots[TYPE_MAX] __ro_after_init;
179
180static inline int hw_breakpoint_slots_cached(int type)
181{
182 return __nr_bp_slots[type];
183}
184
185static __init bool
186bp_slots_histogram_alloc(struct bp_slots_histogram *hist, enum bp_type_idx type)
187{
188 hist->count = kzalloc_objs(*hist->count,
189 hw_breakpoint_slots_cached(type));
190 return hist->count;
191}
192
193static __init void bp_slots_histogram_free(struct bp_slots_histogram *hist)
194{
195 kfree(hist->count);
196}
197
198static __init int init_breakpoint_slots(void)
199{
200 int i, cpu, err_cpu;
201
202 for (i = 0; i < TYPE_MAX; i++)
203 __nr_bp_slots[i] = hw_breakpoint_slots(i);
204
205 for_each_possible_cpu(cpu) {
206 for (i = 0; i < TYPE_MAX; i++) {
207 struct bp_cpuinfo *info = get_bp_info(cpu, i);
208
209 if (!bp_slots_histogram_alloc(&info->tsk_pinned, i))
210 goto err;
211 }
212 }
213 for (i = 0; i < TYPE_MAX; i++) {
214 if (!bp_slots_histogram_alloc(&cpu_pinned[i], i))
215 goto err;
216 if (!bp_slots_histogram_alloc(&tsk_pinned_all[i], i))
217 goto err;
218 }
219
220 return 0;
221err:
222 for_each_possible_cpu(err_cpu) {
223 for (i = 0; i < TYPE_MAX; i++)
224 bp_slots_histogram_free(&get_bp_info(err_cpu, i)->tsk_pinned);
225 if (err_cpu == cpu)
226 break;
227 }
228 for (i = 0; i < TYPE_MAX; i++) {
229 bp_slots_histogram_free(&cpu_pinned[i]);
230 bp_slots_histogram_free(&tsk_pinned_all[i]);
231 }
232
233 return -ENOMEM;
234}
235#endif
236
237static inline void
238bp_slots_histogram_add(struct bp_slots_histogram *hist, int old, int val)
239{
240 const int old_idx = old - 1;
241 const int new_idx = old_idx + val;
242
243 if (old_idx >= 0)
244 WARN_ON(atomic_dec_return_relaxed(&hist->count[old_idx]) < 0);
245 if (new_idx >= 0)
246 WARN_ON(atomic_inc_return_relaxed(&hist->count[new_idx]) < 0);
247}
248
249static int
250bp_slots_histogram_max(struct bp_slots_histogram *hist, enum bp_type_idx type)
251{
252 for (int i = hw_breakpoint_slots_cached(type) - 1; i >= 0; i--) {
253 const int count = atomic_read(&hist->count[i]);
254
255 /* Catch unexpected writers; we want a stable snapshot. */
256 ASSERT_EXCLUSIVE_WRITER(hist->count[i]);
257 if (count > 0)
258 return i + 1;
259 WARN(count < 0, "inconsistent breakpoint slots histogram");
260 }
261
262 return 0;
263}
264
265static int
266bp_slots_histogram_max_merge(struct bp_slots_histogram *hist1, struct bp_slots_histogram *hist2,
267 enum bp_type_idx type)
268{
269 for (int i = hw_breakpoint_slots_cached(type) - 1; i >= 0; i--) {
270 const int count1 = atomic_read(&hist1->count[i]);
271 const int count2 = atomic_read(&hist2->count[i]);
272
273 /* Catch unexpected writers; we want a stable snapshot. */
274 ASSERT_EXCLUSIVE_WRITER(hist1->count[i]);
275 ASSERT_EXCLUSIVE_WRITER(hist2->count[i]);
276 if (count1 + count2 > 0)
277 return i + 1;
278 WARN(count1 < 0, "inconsistent breakpoint slots histogram");
279 WARN(count2 < 0, "inconsistent breakpoint slots histogram");
280 }
281
282 return 0;
283}
284
285#ifndef hw_breakpoint_weight
286static inline int hw_breakpoint_weight(struct perf_event *bp)
287{
288 return 1;
289}
290#endif
291
292static inline enum bp_type_idx find_slot_idx(u64 bp_type)
293{
294 if (bp_type & HW_BREAKPOINT_RW)
295 return TYPE_DATA;
296
297 return TYPE_INST;
298}
299
300/*
301 * Return the maximum number of pinned breakpoints a task has in this CPU.
302 */
303static unsigned int max_task_bp_pinned(int cpu, enum bp_type_idx type)
304{
305 struct bp_slots_histogram *tsk_pinned = &get_bp_info(cpu, type)->tsk_pinned;
306
307 /*
308 * At this point we want to have acquired the bp_cpuinfo_sem as a
309 * writer to ensure that there are no concurrent writers in
310 * toggle_bp_task_slot() to tsk_pinned, and we get a stable snapshot.
311 */
312 lockdep_assert_held_write(&bp_cpuinfo_sem);
313 return bp_slots_histogram_max_merge(tsk_pinned, &tsk_pinned_all[type], type);
314}
315
316/*
317 * Count the number of breakpoints of the same type and same task.
318 * The given event must be not on the list.
319 *
320 * If @cpu is -1, but the result of task_bp_pinned() is not CPU-independent,
321 * returns a negative value.
322 */
323static int task_bp_pinned(int cpu, struct perf_event *bp, enum bp_type_idx type)
324{
325 struct rhlist_head *head, *pos;
326 struct perf_event *iter;
327 int count = 0;
328
329 /*
330 * We need a stable snapshot of the per-task breakpoint list.
331 */
332 assert_bp_constraints_lock_held(bp);
333
334 rcu_read_lock();
335 head = rhltable_lookup(&task_bps_ht, &bp->hw.target, task_bps_ht_params);
336 if (!head)
337 goto out;
338
339 rhl_for_each_entry_rcu(iter, pos, head, hw.bp_list) {
340 if (find_slot_idx(iter->attr.bp_type) != type)
341 continue;
342
343 if (iter->cpu >= 0) {
344 if (cpu == -1) {
345 count = -1;
346 goto out;
347 } else if (cpu != iter->cpu)
348 continue;
349 }
350
351 count += hw_breakpoint_weight(iter);
352 }
353
354out:
355 rcu_read_unlock();
356 return count;
357}
358
359static const struct cpumask *cpumask_of_bp(struct perf_event *bp)
360{
361 if (bp->cpu >= 0)
362 return cpumask_of(bp->cpu);
363 return cpu_possible_mask;
364}
365
366/*
367 * Returns the max pinned breakpoint slots in a given
368 * CPU (cpu > -1) or across all of them (cpu = -1).
369 */
370static int
371max_bp_pinned_slots(struct perf_event *bp, enum bp_type_idx type)
372{
373 const struct cpumask *cpumask = cpumask_of_bp(bp);
374 int pinned_slots = 0;
375 int cpu;
376
377 if (bp->hw.target && bp->cpu < 0) {
378 int max_pinned = task_bp_pinned(-1, bp, type);
379
380 if (max_pinned >= 0) {
381 /*
382 * Fast path: task_bp_pinned() is CPU-independent and
383 * returns the same value for any CPU.
384 */
385 max_pinned += bp_slots_histogram_max(&cpu_pinned[type], type);
386 return max_pinned;
387 }
388 }
389
390 for_each_cpu(cpu, cpumask) {
391 struct bp_cpuinfo *info = get_bp_info(cpu, type);
392 int nr;
393
394 nr = info->cpu_pinned;
395 if (!bp->hw.target)
396 nr += max_task_bp_pinned(cpu, type);
397 else
398 nr += task_bp_pinned(cpu, bp, type);
399
400 pinned_slots = max(nr, pinned_slots);
401 }
402
403 return pinned_slots;
404}
405
406/*
407 * Add/remove the given breakpoint in our constraint table
408 */
409static int
410toggle_bp_slot(struct perf_event *bp, bool enable, enum bp_type_idx type, int weight)
411{
412 int cpu, next_tsk_pinned;
413
414 if (!enable)
415 weight = -weight;
416
417 if (!bp->hw.target) {
418 /*
419 * Update the pinned CPU slots, in per-CPU bp_cpuinfo and in the
420 * global histogram.
421 */
422 struct bp_cpuinfo *info = get_bp_info(bp->cpu, type);
423
424 lockdep_assert_held_write(&bp_cpuinfo_sem);
425 bp_slots_histogram_add(&cpu_pinned[type], info->cpu_pinned, weight);
426 info->cpu_pinned += weight;
427 return 0;
428 }
429
430 /*
431 * If bp->hw.target, tsk_pinned is only modified, but not used
432 * otherwise. We can permit concurrent updates as long as there are no
433 * other uses: having acquired bp_cpuinfo_sem as a reader allows
434 * concurrent updates here. Uses of tsk_pinned will require acquiring
435 * bp_cpuinfo_sem as a writer to stabilize tsk_pinned's value.
436 */
437 lockdep_assert_held_read(&bp_cpuinfo_sem);
438
439 /*
440 * Update the pinned task slots, in per-CPU bp_cpuinfo and in the global
441 * histogram. We need to take care of 4 cases:
442 *
443 * 1. This breakpoint targets all CPUs (cpu < 0), and there may only
444 * exist other task breakpoints targeting all CPUs. In this case we
445 * can simply update the global slots histogram.
446 *
447 * 2. This breakpoint targets a specific CPU (cpu >= 0), but there may
448 * only exist other task breakpoints targeting all CPUs.
449 *
450 * a. On enable: remove the existing breakpoints from the global
451 * slots histogram and use the per-CPU histogram.
452 *
453 * b. On disable: re-insert the existing breakpoints into the global
454 * slots histogram and remove from per-CPU histogram.
455 *
456 * 3. Some other existing task breakpoints target specific CPUs. Only
457 * update the per-CPU slots histogram.
458 */
459
460 if (!enable) {
461 /*
462 * Remove before updating histograms so we can determine if this
463 * was the last task breakpoint for a specific CPU.
464 */
465 int ret = rhltable_remove(&task_bps_ht, &bp->hw.bp_list, task_bps_ht_params);
466
467 if (ret)
468 return ret;
469 }
470 /*
471 * Note: If !enable, next_tsk_pinned will not count the to-be-removed breakpoint.
472 */
473 next_tsk_pinned = task_bp_pinned(-1, bp, type);
474
475 if (next_tsk_pinned >= 0) {
476 if (bp->cpu < 0) { /* Case 1: fast path */
477 if (!enable)
478 next_tsk_pinned += hw_breakpoint_weight(bp);
479 bp_slots_histogram_add(&tsk_pinned_all[type], next_tsk_pinned, weight);
480 } else if (enable) { /* Case 2.a: slow path */
481 /* Add existing to per-CPU histograms. */
482 for_each_possible_cpu(cpu) {
483 bp_slots_histogram_add(&get_bp_info(cpu, type)->tsk_pinned,
484 0, next_tsk_pinned);
485 }
486 /* Add this first CPU-pinned task breakpoint. */
487 bp_slots_histogram_add(&get_bp_info(bp->cpu, type)->tsk_pinned,
488 next_tsk_pinned, weight);
489 /* Rebalance global task pinned histogram. */
490 bp_slots_histogram_add(&tsk_pinned_all[type], next_tsk_pinned,
491 -next_tsk_pinned);
492 } else { /* Case 2.b: slow path */
493 /* Remove this last CPU-pinned task breakpoint. */
494 bp_slots_histogram_add(&get_bp_info(bp->cpu, type)->tsk_pinned,
495 next_tsk_pinned + hw_breakpoint_weight(bp), weight);
496 /* Remove all from per-CPU histograms. */
497 for_each_possible_cpu(cpu) {
498 bp_slots_histogram_add(&get_bp_info(cpu, type)->tsk_pinned,
499 next_tsk_pinned, -next_tsk_pinned);
500 }
501 /* Rebalance global task pinned histogram. */
502 bp_slots_histogram_add(&tsk_pinned_all[type], 0, next_tsk_pinned);
503 }
504 } else { /* Case 3: slow path */
505 const struct cpumask *cpumask = cpumask_of_bp(bp);
506
507 for_each_cpu(cpu, cpumask) {
508 next_tsk_pinned = task_bp_pinned(cpu, bp, type);
509 if (!enable)
510 next_tsk_pinned += hw_breakpoint_weight(bp);
511 bp_slots_histogram_add(&get_bp_info(cpu, type)->tsk_pinned,
512 next_tsk_pinned, weight);
513 }
514 }
515
516 /*
517 * Readers want a stable snapshot of the per-task breakpoint list.
518 */
519 assert_bp_constraints_lock_held(bp);
520
521 if (enable)
522 return rhltable_insert(&task_bps_ht, &bp->hw.bp_list, task_bps_ht_params);
523
524 return 0;
525}
526
527/*
528 * Constraints to check before allowing this new breakpoint counter.
529 *
530 * Note: Flexible breakpoints are currently unimplemented, but outlined in the
531 * below algorithm for completeness. The implementation treats flexible as
532 * pinned due to no guarantee that we currently always schedule flexible events
533 * before a pinned event in a same CPU.
534 *
535 * == Non-pinned counter == (Considered as pinned for now)
536 *
537 * - If attached to a single cpu, check:
538 *
539 * (per_cpu(info->flexible, cpu) || (per_cpu(info->cpu_pinned, cpu)
540 * + max(per_cpu(info->tsk_pinned, cpu)))) < HBP_NUM
541 *
542 * -> If there are already non-pinned counters in this cpu, it means
543 * there is already a free slot for them.
544 * Otherwise, we check that the maximum number of per task
545 * breakpoints (for this cpu) plus the number of per cpu breakpoint
546 * (for this cpu) doesn't cover every registers.
547 *
548 * - If attached to every cpus, check:
549 *
550 * (per_cpu(info->flexible, *) || (max(per_cpu(info->cpu_pinned, *))
551 * + max(per_cpu(info->tsk_pinned, *)))) < HBP_NUM
552 *
553 * -> This is roughly the same, except we check the number of per cpu
554 * bp for every cpu and we keep the max one. Same for the per tasks
555 * breakpoints.
556 *
557 *
558 * == Pinned counter ==
559 *
560 * - If attached to a single cpu, check:
561 *
562 * ((per_cpu(info->flexible, cpu) > 1) + per_cpu(info->cpu_pinned, cpu)
563 * + max(per_cpu(info->tsk_pinned, cpu))) < HBP_NUM
564 *
565 * -> Same checks as before. But now the info->flexible, if any, must keep
566 * one register at least (or they will never be fed).
567 *
568 * - If attached to every cpus, check:
569 *
570 * ((per_cpu(info->flexible, *) > 1) + max(per_cpu(info->cpu_pinned, *))
571 * + max(per_cpu(info->tsk_pinned, *))) < HBP_NUM
572 */
573static int __reserve_bp_slot(struct perf_event *bp, u64 bp_type)
574{
575 enum bp_type_idx type;
576 int max_pinned_slots;
577 int weight;
578
579 /* We couldn't initialize breakpoint constraints on boot */
580 if (!constraints_initialized)
581 return -ENOMEM;
582
583 /* Basic checks */
584 if (bp_type == HW_BREAKPOINT_EMPTY ||
585 bp_type == HW_BREAKPOINT_INVALID)
586 return -EINVAL;
587
588 type = find_slot_idx(bp_type);
589 weight = hw_breakpoint_weight(bp);
590
591 /* Check if this new breakpoint can be satisfied across all CPUs. */
592 max_pinned_slots = max_bp_pinned_slots(bp, type) + weight;
593 if (max_pinned_slots > hw_breakpoint_slots_cached(type))
594 return -ENOSPC;
595
596 return toggle_bp_slot(bp, true, type, weight);
597}
598
599int reserve_bp_slot(struct perf_event *bp)
600{
601 struct mutex *mtx = bp_constraints_lock(bp);
602 int ret = __reserve_bp_slot(bp, bp->attr.bp_type);
603
604 bp_constraints_unlock(mtx);
605 return ret;
606}
607
608static void __release_bp_slot(struct perf_event *bp, u64 bp_type)
609{
610 enum bp_type_idx type;
611 int weight;
612
613 type = find_slot_idx(bp_type);
614 weight = hw_breakpoint_weight(bp);
615 WARN_ON(toggle_bp_slot(bp, false, type, weight));
616}
617
618void release_bp_slot(struct perf_event *bp)
619{
620 struct mutex *mtx = bp_constraints_lock(bp);
621
622 __release_bp_slot(bp, bp->attr.bp_type);
623 bp_constraints_unlock(mtx);
624}
625
626static int __modify_bp_slot(struct perf_event *bp, u64 old_type, u64 new_type)
627{
628 int err;
629
630 __release_bp_slot(bp, old_type);
631
632 err = __reserve_bp_slot(bp, new_type);
633 if (err) {
634 /*
635 * Reserve the old_type slot back in case
636 * there's no space for the new type.
637 *
638 * This must succeed, because we just released
639 * the old_type slot in the __release_bp_slot
640 * call above. If not, something is broken.
641 */
642 WARN_ON(__reserve_bp_slot(bp, old_type));
643 }
644
645 return err;
646}
647
648static int modify_bp_slot(struct perf_event *bp, u64 old_type, u64 new_type)
649{
650 struct mutex *mtx = bp_constraints_lock(bp);
651 int ret = __modify_bp_slot(bp, old_type, new_type);
652
653 bp_constraints_unlock(mtx);
654 return ret;
655}
656
657/*
658 * Allow the kernel debugger to reserve breakpoint slots without
659 * taking a lock using the dbg_* variant of for the reserve and
660 * release breakpoint slots.
661 */
662int dbg_reserve_bp_slot(struct perf_event *bp)
663{
664 int ret;
665
666 if (bp_constraints_is_locked(bp))
667 return -1;
668
669 /* Locks aren't held; disable lockdep assert checking. */
670 lockdep_off();
671 ret = __reserve_bp_slot(bp, bp->attr.bp_type);
672 lockdep_on();
673
674 return ret;
675}
676
677int dbg_release_bp_slot(struct perf_event *bp)
678{
679 if (bp_constraints_is_locked(bp))
680 return -1;
681
682 /* Locks aren't held; disable lockdep assert checking. */
683 lockdep_off();
684 __release_bp_slot(bp, bp->attr.bp_type);
685 lockdep_on();
686
687 return 0;
688}
689
690static int hw_breakpoint_parse(struct perf_event *bp,
691 const struct perf_event_attr *attr,
692 struct arch_hw_breakpoint *hw)
693{
694 int err;
695
696 err = hw_breakpoint_arch_parse(bp, attr, hw);
697 if (err)
698 return err;
699
700 if (arch_check_bp_in_kernelspace(hw)) {
701 if (attr->exclude_kernel)
702 return -EINVAL;
703 /*
704 * Don't let unprivileged users set a breakpoint in the trap
705 * path to avoid trap recursion attacks.
706 */
707 if (!capable(CAP_SYS_ADMIN))
708 return -EPERM;
709 }
710
711 return 0;
712}
713
714int register_perf_hw_breakpoint(struct perf_event *bp)
715{
716 struct arch_hw_breakpoint hw = { };
717 int err;
718
719 err = reserve_bp_slot(bp);
720 if (err)
721 return err;
722
723 err = hw_breakpoint_parse(bp, &bp->attr, &hw);
724 if (err) {
725 release_bp_slot(bp);
726 return err;
727 }
728
729 bp->hw.info = hw;
730
731 return 0;
732}
733
734/**
735 * register_user_hw_breakpoint - register a hardware breakpoint for user space
736 * @attr: breakpoint attributes
737 * @triggered: callback to trigger when we hit the breakpoint
738 * @context: context data could be used in the triggered callback
739 * @tsk: pointer to 'task_struct' of the process to which the address belongs
740 */
741struct perf_event *
742register_user_hw_breakpoint(struct perf_event_attr *attr,
743 perf_overflow_handler_t triggered,
744 void *context,
745 struct task_struct *tsk)
746{
747 return perf_event_create_kernel_counter(attr, -1, tsk, triggered,
748 context);
749}
750EXPORT_SYMBOL_GPL(register_user_hw_breakpoint);
751
752static void hw_breakpoint_copy_attr(struct perf_event_attr *to,
753 struct perf_event_attr *from)
754{
755 to->bp_addr = from->bp_addr;
756 to->bp_type = from->bp_type;
757 to->bp_len = from->bp_len;
758 to->disabled = from->disabled;
759}
760
761int
762modify_user_hw_breakpoint_check(struct perf_event *bp, struct perf_event_attr *attr,
763 bool check)
764{
765 struct arch_hw_breakpoint hw = { };
766 int err;
767
768 err = hw_breakpoint_parse(bp, attr, &hw);
769 if (err)
770 return err;
771
772 if (check) {
773 struct perf_event_attr old_attr;
774
775 old_attr = bp->attr;
776 hw_breakpoint_copy_attr(&old_attr, attr);
777 if (memcmp(&old_attr, attr, sizeof(*attr)))
778 return -EINVAL;
779 }
780
781 if (bp->attr.bp_type != attr->bp_type) {
782 err = modify_bp_slot(bp, bp->attr.bp_type, attr->bp_type);
783 if (err)
784 return err;
785 }
786
787 hw_breakpoint_copy_attr(&bp->attr, attr);
788 bp->hw.info = hw;
789
790 return 0;
791}
792
793/**
794 * modify_user_hw_breakpoint - modify a user-space hardware breakpoint
795 * @bp: the breakpoint structure to modify
796 * @attr: new breakpoint attributes
797 */
798int modify_user_hw_breakpoint(struct perf_event *bp, struct perf_event_attr *attr)
799{
800 int err;
801
802 /*
803 * modify_user_hw_breakpoint can be invoked with IRQs disabled and hence it
804 * will not be possible to raise IPIs that invoke __perf_event_disable.
805 * So call the function directly after making sure we are targeting the
806 * current task.
807 */
808 if (irqs_disabled() && bp->ctx && bp->ctx->task == current)
809 perf_event_disable_local(bp);
810 else
811 perf_event_disable(bp);
812
813 err = modify_user_hw_breakpoint_check(bp, attr, false);
814
815 if (!bp->attr.disabled)
816 perf_event_enable(bp);
817
818 return err;
819}
820EXPORT_SYMBOL_GPL(modify_user_hw_breakpoint);
821
822/**
823 * unregister_hw_breakpoint - unregister a user-space hardware breakpoint
824 * @bp: the breakpoint structure to unregister
825 */
826void unregister_hw_breakpoint(struct perf_event *bp)
827{
828 if (!bp)
829 return;
830 perf_event_release_kernel(bp);
831}
832EXPORT_SYMBOL_GPL(unregister_hw_breakpoint);
833
834/**
835 * register_wide_hw_breakpoint - register a wide breakpoint in the kernel
836 * @attr: breakpoint attributes
837 * @triggered: callback to trigger when we hit the breakpoint
838 * @context: context data could be used in the triggered callback
839 *
840 * @return a set of per_cpu pointers to perf events
841 */
842struct perf_event * __percpu *
843register_wide_hw_breakpoint(struct perf_event_attr *attr,
844 perf_overflow_handler_t triggered,
845 void *context)
846{
847 struct perf_event * __percpu *cpu_events, *bp;
848 long err = 0;
849 int cpu;
850
851 cpu_events = alloc_percpu(typeof(*cpu_events));
852 if (!cpu_events)
853 return ERR_PTR_PCPU(-ENOMEM);
854
855 cpus_read_lock();
856 for_each_online_cpu(cpu) {
857 bp = perf_event_create_kernel_counter(attr, cpu, NULL,
858 triggered, context);
859 if (IS_ERR(bp)) {
860 err = PTR_ERR(bp);
861 break;
862 }
863
864 per_cpu(*cpu_events, cpu) = bp;
865 }
866 cpus_read_unlock();
867
868 if (likely(!err))
869 return cpu_events;
870
871 unregister_wide_hw_breakpoint(cpu_events);
872 return ERR_PTR_PCPU(err);
873}
874EXPORT_SYMBOL_GPL(register_wide_hw_breakpoint);
875
876/**
877 * unregister_wide_hw_breakpoint - unregister a wide breakpoint in the kernel
878 * @cpu_events: the per cpu set of events to unregister
879 */
880void unregister_wide_hw_breakpoint(struct perf_event * __percpu *cpu_events)
881{
882 int cpu;
883
884 for_each_possible_cpu(cpu)
885 unregister_hw_breakpoint(per_cpu(*cpu_events, cpu));
886
887 free_percpu(cpu_events);
888}
889EXPORT_SYMBOL_GPL(unregister_wide_hw_breakpoint);
890
891/**
892 * hw_breakpoint_is_used - check if breakpoints are currently used
893 *
894 * Returns: true if breakpoints are used, false otherwise.
895 */
896bool hw_breakpoint_is_used(void)
897{
898 int cpu;
899
900 if (!constraints_initialized)
901 return false;
902
903 for_each_possible_cpu(cpu) {
904 for (int type = 0; type < TYPE_MAX; ++type) {
905 struct bp_cpuinfo *info = get_bp_info(cpu, type);
906
907 if (info->cpu_pinned)
908 return true;
909
910 for (int slot = 0; slot < hw_breakpoint_slots_cached(type); ++slot) {
911 if (atomic_read(&info->tsk_pinned.count[slot]))
912 return true;
913 }
914 }
915 }
916
917 for (int type = 0; type < TYPE_MAX; ++type) {
918 for (int slot = 0; slot < hw_breakpoint_slots_cached(type); ++slot) {
919 /*
920 * Warn, because if there are CPU pinned counters,
921 * should never get here; bp_cpuinfo::cpu_pinned should
922 * be consistent with the global cpu_pinned histogram.
923 */
924 if (WARN_ON(atomic_read(&cpu_pinned[type].count[slot])))
925 return true;
926
927 if (atomic_read(&tsk_pinned_all[type].count[slot]))
928 return true;
929 }
930 }
931
932 return false;
933}
934
935static struct notifier_block hw_breakpoint_exceptions_nb = {
936 .notifier_call = hw_breakpoint_exceptions_notify,
937 /* we need to be notified first */
938 .priority = 0x7fffffff
939};
940
941static void bp_perf_event_destroy(struct perf_event *event)
942{
943 release_bp_slot(event);
944}
945
946static int hw_breakpoint_event_init(struct perf_event *bp)
947{
948 int err;
949
950 if (bp->attr.type != PERF_TYPE_BREAKPOINT)
951 return -ENOENT;
952
953 /*
954 * Check if breakpoint type is supported before proceeding.
955 * Also, no branch sampling for breakpoint events.
956 */
957 if (!hw_breakpoint_slots_cached(find_slot_idx(bp->attr.bp_type)) || has_branch_stack(bp))
958 return -EOPNOTSUPP;
959
960 err = register_perf_hw_breakpoint(bp);
961 if (err)
962 return err;
963
964 bp->destroy = bp_perf_event_destroy;
965
966 return 0;
967}
968
969static int hw_breakpoint_add(struct perf_event *bp, int flags)
970{
971 if (!(flags & PERF_EF_START))
972 bp->hw.state = PERF_HES_STOPPED;
973
974 if (is_sampling_event(bp)) {
975 bp->hw.last_period = bp->hw.sample_period;
976 perf_swevent_set_period(bp);
977 }
978
979 return arch_install_hw_breakpoint(bp);
980}
981
982static void hw_breakpoint_del(struct perf_event *bp, int flags)
983{
984 arch_uninstall_hw_breakpoint(bp);
985}
986
987static void hw_breakpoint_start(struct perf_event *bp, int flags)
988{
989 bp->hw.state = 0;
990}
991
992static void hw_breakpoint_stop(struct perf_event *bp, int flags)
993{
994 bp->hw.state = PERF_HES_STOPPED;
995}
996
997static struct pmu perf_breakpoint = {
998 .task_ctx_nr = perf_sw_context, /* could eventually get its own */
999
1000 .event_init = hw_breakpoint_event_init,
1001 .add = hw_breakpoint_add,
1002 .del = hw_breakpoint_del,
1003 .start = hw_breakpoint_start,
1004 .stop = hw_breakpoint_stop,
1005 .read = hw_breakpoint_pmu_read,
1006};
1007
1008int __init init_hw_breakpoint(void)
1009{
1010 int ret;
1011
1012 ret = rhltable_init(&task_bps_ht, &task_bps_ht_params);
1013 if (ret)
1014 return ret;
1015
1016 ret = init_breakpoint_slots();
1017 if (ret)
1018 return ret;
1019
1020 constraints_initialized = true;
1021
1022 perf_pmu_register(&perf_breakpoint, "breakpoint", PERF_TYPE_BREAKPOINT);
1023
1024 return register_die_notifier(&hw_breakpoint_exceptions_nb);
1025}