perf: Start the restructuring
[linux-2.6.git] / kernel / events / hw_breakpoint.c
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
51 /*
52  * Constraints data
53  */
54
55 /* Number of pinned cpu breakpoints in a cpu */
56 static DEFINE_PER_CPU(unsigned int, nr_cpu_bp_pinned[TYPE_MAX]);
57
58 /* Number of pinned task breakpoints in a cpu */
59 static DEFINE_PER_CPU(unsigned int *, nr_task_bp_pinned[TYPE_MAX]);
60
61 /* Number of non-pinned cpu/task breakpoints in a cpu */
62 static DEFINE_PER_CPU(unsigned int, nr_bp_flexible[TYPE_MAX]);
63
64 static int nr_slots[TYPE_MAX];
65
66 /* Keep track of the breakpoints attached to tasks */
67 static LIST_HEAD(bp_task_head);
68
69 static int constraints_initialized;
70
71 /* Gather the number of total pinned and un-pinned bp in a cpuset */
72 struct bp_busy_slots {
73         unsigned int pinned;
74         unsigned int flexible;
75 };
76
77 /* Serialize accesses to the above constraints */
78 static DEFINE_MUTEX(nr_bp_mutex);
79
80 __weak int hw_breakpoint_weight(struct perf_event *bp)
81 {
82         return 1;
83 }
84
85 static inline enum bp_type_idx find_slot_idx(struct perf_event *bp)
86 {
87         if (bp->attr.bp_type & HW_BREAKPOINT_RW)
88                 return TYPE_DATA;
89
90         return TYPE_INST;
91 }
92
93 /*
94  * Report the maximum number of pinned breakpoints a task
95  * have in this cpu
96  */
97 static unsigned int max_task_bp_pinned(int cpu, enum bp_type_idx type)
98 {
99         int i;
100         unsigned int *tsk_pinned = per_cpu(nr_task_bp_pinned[type], cpu);
101
102         for (i = nr_slots[type] - 1; i >= 0; i--) {
103                 if (tsk_pinned[i] > 0)
104                         return i + 1;
105         }
106
107         return 0;
108 }
109
110 /*
111  * Count the number of breakpoints of the same type and same task.
112  * The given event must be not on the list.
113  */
114 static int task_bp_pinned(struct perf_event *bp, enum bp_type_idx type)
115 {
116         struct task_struct *tsk = bp->hw.bp_target;
117         struct perf_event *iter;
118         int count = 0;
119
120         list_for_each_entry(iter, &bp_task_head, hw.bp_list) {
121                 if (iter->hw.bp_target == tsk && find_slot_idx(iter) == type)
122                         count += hw_breakpoint_weight(iter);
123         }
124
125         return count;
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         int cpu = bp->cpu;
137         struct task_struct *tsk = bp->hw.bp_target;
138
139         if (cpu >= 0) {
140                 slots->pinned = per_cpu(nr_cpu_bp_pinned[type], cpu);
141                 if (!tsk)
142                         slots->pinned += max_task_bp_pinned(cpu, type);
143                 else
144                         slots->pinned += task_bp_pinned(bp, type);
145                 slots->flexible = per_cpu(nr_bp_flexible[type], cpu);
146
147                 return;
148         }
149
150         for_each_online_cpu(cpu) {
151                 unsigned int nr;
152
153                 nr = per_cpu(nr_cpu_bp_pinned[type], cpu);
154                 if (!tsk)
155                         nr += max_task_bp_pinned(cpu, type);
156                 else
157                         nr += task_bp_pinned(bp, type);
158
159                 if (nr > slots->pinned)
160                         slots->pinned = nr;
161
162                 nr = per_cpu(nr_bp_flexible[type], cpu);
163
164                 if (nr > slots->flexible)
165                         slots->flexible = nr;
166         }
167 }
168
169 /*
170  * For now, continue to consider flexible as pinned, until we can
171  * ensure no flexible event can ever be scheduled before a pinned event
172  * in a same cpu.
173  */
174 static void
175 fetch_this_slot(struct bp_busy_slots *slots, int weight)
176 {
177         slots->pinned += weight;
178 }
179
180 /*
181  * Add a pinned breakpoint for the given task in our constraint table
182  */
183 static void toggle_bp_task_slot(struct perf_event *bp, int cpu, bool enable,
184                                 enum bp_type_idx type, int weight)
185 {
186         unsigned int *tsk_pinned;
187         int old_count = 0;
188         int old_idx = 0;
189         int idx = 0;
190
191         old_count = task_bp_pinned(bp, type);
192         old_idx = old_count - 1;
193         idx = old_idx + weight;
194
195         /* tsk_pinned[n] is the number of tasks having n breakpoints */
196         tsk_pinned = per_cpu(nr_task_bp_pinned[type], cpu);
197         if (enable) {
198                 tsk_pinned[idx]++;
199                 if (old_count > 0)
200                         tsk_pinned[old_idx]--;
201         } else {
202                 tsk_pinned[idx]--;
203                 if (old_count > 0)
204                         tsk_pinned[old_idx]++;
205         }
206 }
207
208 /*
209  * Add/remove the given breakpoint in our constraint table
210  */
211 static void
212 toggle_bp_slot(struct perf_event *bp, bool enable, enum bp_type_idx type,
213                int weight)
214 {
215         int cpu = bp->cpu;
216         struct task_struct *tsk = bp->hw.bp_target;
217
218         /* Pinned counter cpu profiling */
219         if (!tsk) {
220
221                 if (enable)
222                         per_cpu(nr_cpu_bp_pinned[type], bp->cpu) += weight;
223                 else
224                         per_cpu(nr_cpu_bp_pinned[type], bp->cpu) -= weight;
225                 return;
226         }
227
228         /* Pinned counter task profiling */
229
230         if (!enable)
231                 list_del(&bp->hw.bp_list);
232
233         if (cpu >= 0) {
234                 toggle_bp_task_slot(bp, cpu, enable, type, weight);
235         } else {
236                 for_each_online_cpu(cpu)
237                         toggle_bp_task_slot(bp, cpu, enable, type, weight);
238         }
239
240         if (enable)
241                 list_add_tail(&bp->hw.bp_list, &bp_task_head);
242 }
243
244 /*
245  * Function to perform processor-specific cleanup during unregistration
246  */
247 __weak void arch_unregister_hw_breakpoint(struct perf_event *bp)
248 {
249         /*
250          * A weak stub function here for those archs that don't define
251          * it inside arch/.../kernel/hw_breakpoint.c
252          */
253 }
254
255 /*
256  * Contraints to check before allowing this new breakpoint counter:
257  *
258  *  == Non-pinned counter == (Considered as pinned for now)
259  *
260  *   - If attached to a single cpu, check:
261  *
262  *       (per_cpu(nr_bp_flexible, cpu) || (per_cpu(nr_cpu_bp_pinned, cpu)
263  *           + max(per_cpu(nr_task_bp_pinned, cpu)))) < HBP_NUM
264  *
265  *       -> If there are already non-pinned counters in this cpu, it means
266  *          there is already a free slot for them.
267  *          Otherwise, we check that the maximum number of per task
268  *          breakpoints (for this cpu) plus the number of per cpu breakpoint
269  *          (for this cpu) doesn't cover every registers.
270  *
271  *   - If attached to every cpus, check:
272  *
273  *       (per_cpu(nr_bp_flexible, *) || (max(per_cpu(nr_cpu_bp_pinned, *))
274  *           + max(per_cpu(nr_task_bp_pinned, *)))) < HBP_NUM
275  *
276  *       -> This is roughly the same, except we check the number of per cpu
277  *          bp for every cpu and we keep the max one. Same for the per tasks
278  *          breakpoints.
279  *
280  *
281  * == Pinned counter ==
282  *
283  *   - If attached to a single cpu, check:
284  *
285  *       ((per_cpu(nr_bp_flexible, cpu) > 1) + per_cpu(nr_cpu_bp_pinned, cpu)
286  *            + max(per_cpu(nr_task_bp_pinned, cpu))) < HBP_NUM
287  *
288  *       -> Same checks as before. But now the nr_bp_flexible, if any, must keep
289  *          one register at least (or they will never be fed).
290  *
291  *   - If attached to every cpus, check:
292  *
293  *       ((per_cpu(nr_bp_flexible, *) > 1) + max(per_cpu(nr_cpu_bp_pinned, *))
294  *            + max(per_cpu(nr_task_bp_pinned, *))) < HBP_NUM
295  */
296 static int __reserve_bp_slot(struct perf_event *bp)
297 {
298         struct bp_busy_slots slots = {0};
299         enum bp_type_idx type;
300         int weight;
301
302         /* We couldn't initialize breakpoint constraints on boot */
303         if (!constraints_initialized)
304                 return -ENOMEM;
305
306         /* Basic checks */
307         if (bp->attr.bp_type == HW_BREAKPOINT_EMPTY ||
308             bp->attr.bp_type == HW_BREAKPOINT_INVALID)
309                 return -EINVAL;
310
311         type = find_slot_idx(bp);
312         weight = hw_breakpoint_weight(bp);
313
314         fetch_bp_busy_slots(&slots, bp, type);
315         /*
316          * Simulate the addition of this breakpoint to the constraints
317          * and see the result.
318          */
319         fetch_this_slot(&slots, weight);
320
321         /* Flexible counters need to keep at least one slot */
322         if (slots.pinned + (!!slots.flexible) > nr_slots[type])
323                 return -ENOSPC;
324
325         toggle_bp_slot(bp, true, type, weight);
326
327         return 0;
328 }
329
330 int reserve_bp_slot(struct perf_event *bp)
331 {
332         int ret;
333
334         mutex_lock(&nr_bp_mutex);
335
336         ret = __reserve_bp_slot(bp);
337
338         mutex_unlock(&nr_bp_mutex);
339
340         return ret;
341 }
342
343 static void __release_bp_slot(struct perf_event *bp)
344 {
345         enum bp_type_idx type;
346         int weight;
347
348         type = find_slot_idx(bp);
349         weight = hw_breakpoint_weight(bp);
350         toggle_bp_slot(bp, false, type, weight);
351 }
352
353 void release_bp_slot(struct perf_event *bp)
354 {
355         mutex_lock(&nr_bp_mutex);
356
357         arch_unregister_hw_breakpoint(bp);
358         __release_bp_slot(bp);
359
360         mutex_unlock(&nr_bp_mutex);
361 }
362
363 /*
364  * Allow the kernel debugger to reserve breakpoint slots without
365  * taking a lock using the dbg_* variant of for the reserve and
366  * release breakpoint slots.
367  */
368 int dbg_reserve_bp_slot(struct perf_event *bp)
369 {
370         if (mutex_is_locked(&nr_bp_mutex))
371                 return -1;
372
373         return __reserve_bp_slot(bp);
374 }
375
376 int dbg_release_bp_slot(struct perf_event *bp)
377 {
378         if (mutex_is_locked(&nr_bp_mutex))
379                 return -1;
380
381         __release_bp_slot(bp);
382
383         return 0;
384 }
385
386 static int validate_hw_breakpoint(struct perf_event *bp)
387 {
388         int ret;
389
390         ret = arch_validate_hwbkpt_settings(bp);
391         if (ret)
392                 return ret;
393
394         if (arch_check_bp_in_kernelspace(bp)) {
395                 if (bp->attr.exclude_kernel)
396                         return -EINVAL;
397                 /*
398                  * Don't let unprivileged users set a breakpoint in the trap
399                  * path to avoid trap recursion attacks.
400                  */
401                 if (!capable(CAP_SYS_ADMIN))
402                         return -EPERM;
403         }
404
405         return 0;
406 }
407
408 int register_perf_hw_breakpoint(struct perf_event *bp)
409 {
410         int ret;
411
412         ret = reserve_bp_slot(bp);
413         if (ret)
414                 return ret;
415
416         ret = validate_hw_breakpoint(bp);
417
418         /* if arch_validate_hwbkpt_settings() fails then release bp slot */
419         if (ret)
420                 release_bp_slot(bp);
421
422         return ret;
423 }
424
425 /**
426  * register_user_hw_breakpoint - register a hardware breakpoint for user space
427  * @attr: breakpoint attributes
428  * @triggered: callback to trigger when we hit the breakpoint
429  * @tsk: pointer to 'task_struct' of the process to which the address belongs
430  */
431 struct perf_event *
432 register_user_hw_breakpoint(struct perf_event_attr *attr,
433                             perf_overflow_handler_t triggered,
434                             struct task_struct *tsk)
435 {
436         return perf_event_create_kernel_counter(attr, -1, tsk, triggered);
437 }
438 EXPORT_SYMBOL_GPL(register_user_hw_breakpoint);
439
440 /**
441  * modify_user_hw_breakpoint - modify a user-space hardware breakpoint
442  * @bp: the breakpoint structure to modify
443  * @attr: new breakpoint attributes
444  * @triggered: callback to trigger when we hit the breakpoint
445  * @tsk: pointer to 'task_struct' of the process to which the address belongs
446  */
447 int modify_user_hw_breakpoint(struct perf_event *bp, struct perf_event_attr *attr)
448 {
449         u64 old_addr = bp->attr.bp_addr;
450         u64 old_len = bp->attr.bp_len;
451         int old_type = bp->attr.bp_type;
452         int err = 0;
453
454         perf_event_disable(bp);
455
456         bp->attr.bp_addr = attr->bp_addr;
457         bp->attr.bp_type = attr->bp_type;
458         bp->attr.bp_len = attr->bp_len;
459
460         if (attr->disabled)
461                 goto end;
462
463         err = validate_hw_breakpoint(bp);
464         if (!err)
465                 perf_event_enable(bp);
466
467         if (err) {
468                 bp->attr.bp_addr = old_addr;
469                 bp->attr.bp_type = old_type;
470                 bp->attr.bp_len = old_len;
471                 if (!bp->attr.disabled)
472                         perf_event_enable(bp);
473
474                 return err;
475         }
476
477 end:
478         bp->attr.disabled = attr->disabled;
479
480         return 0;
481 }
482 EXPORT_SYMBOL_GPL(modify_user_hw_breakpoint);
483
484 /**
485  * unregister_hw_breakpoint - unregister a user-space hardware breakpoint
486  * @bp: the breakpoint structure to unregister
487  */
488 void unregister_hw_breakpoint(struct perf_event *bp)
489 {
490         if (!bp)
491                 return;
492         perf_event_release_kernel(bp);
493 }
494 EXPORT_SYMBOL_GPL(unregister_hw_breakpoint);
495
496 /**
497  * register_wide_hw_breakpoint - register a wide breakpoint in the kernel
498  * @attr: breakpoint attributes
499  * @triggered: callback to trigger when we hit the breakpoint
500  *
501  * @return a set of per_cpu pointers to perf events
502  */
503 struct perf_event * __percpu *
504 register_wide_hw_breakpoint(struct perf_event_attr *attr,
505                             perf_overflow_handler_t triggered)
506 {
507         struct perf_event * __percpu *cpu_events, **pevent, *bp;
508         long err;
509         int cpu;
510
511         cpu_events = alloc_percpu(typeof(*cpu_events));
512         if (!cpu_events)
513                 return (void __percpu __force *)ERR_PTR(-ENOMEM);
514
515         get_online_cpus();
516         for_each_online_cpu(cpu) {
517                 pevent = per_cpu_ptr(cpu_events, cpu);
518                 bp = perf_event_create_kernel_counter(attr, cpu, NULL, triggered);
519
520                 *pevent = bp;
521
522                 if (IS_ERR(bp)) {
523                         err = PTR_ERR(bp);
524                         goto fail;
525                 }
526         }
527         put_online_cpus();
528
529         return cpu_events;
530
531 fail:
532         for_each_online_cpu(cpu) {
533                 pevent = per_cpu_ptr(cpu_events, cpu);
534                 if (IS_ERR(*pevent))
535                         break;
536                 unregister_hw_breakpoint(*pevent);
537         }
538         put_online_cpus();
539
540         free_percpu(cpu_events);
541         return (void __percpu __force *)ERR_PTR(err);
542 }
543 EXPORT_SYMBOL_GPL(register_wide_hw_breakpoint);
544
545 /**
546  * unregister_wide_hw_breakpoint - unregister a wide breakpoint in the kernel
547  * @cpu_events: the per cpu set of events to unregister
548  */
549 void unregister_wide_hw_breakpoint(struct perf_event * __percpu *cpu_events)
550 {
551         int cpu;
552         struct perf_event **pevent;
553
554         for_each_possible_cpu(cpu) {
555                 pevent = per_cpu_ptr(cpu_events, cpu);
556                 unregister_hw_breakpoint(*pevent);
557         }
558         free_percpu(cpu_events);
559 }
560 EXPORT_SYMBOL_GPL(unregister_wide_hw_breakpoint);
561
562 static struct notifier_block hw_breakpoint_exceptions_nb = {
563         .notifier_call = hw_breakpoint_exceptions_notify,
564         /* we need to be notified first */
565         .priority = 0x7fffffff
566 };
567
568 static void bp_perf_event_destroy(struct perf_event *event)
569 {
570         release_bp_slot(event);
571 }
572
573 static int hw_breakpoint_event_init(struct perf_event *bp)
574 {
575         int err;
576
577         if (bp->attr.type != PERF_TYPE_BREAKPOINT)
578                 return -ENOENT;
579
580         err = register_perf_hw_breakpoint(bp);
581         if (err)
582                 return err;
583
584         bp->destroy = bp_perf_event_destroy;
585
586         return 0;
587 }
588
589 static int hw_breakpoint_add(struct perf_event *bp, int flags)
590 {
591         if (!(flags & PERF_EF_START))
592                 bp->hw.state = PERF_HES_STOPPED;
593
594         return arch_install_hw_breakpoint(bp);
595 }
596
597 static void hw_breakpoint_del(struct perf_event *bp, int flags)
598 {
599         arch_uninstall_hw_breakpoint(bp);
600 }
601
602 static void hw_breakpoint_start(struct perf_event *bp, int flags)
603 {
604         bp->hw.state = 0;
605 }
606
607 static void hw_breakpoint_stop(struct perf_event *bp, int flags)
608 {
609         bp->hw.state = PERF_HES_STOPPED;
610 }
611
612 static struct pmu perf_breakpoint = {
613         .task_ctx_nr    = perf_sw_context, /* could eventually get its own */
614
615         .event_init     = hw_breakpoint_event_init,
616         .add            = hw_breakpoint_add,
617         .del            = hw_breakpoint_del,
618         .start          = hw_breakpoint_start,
619         .stop           = hw_breakpoint_stop,
620         .read           = hw_breakpoint_pmu_read,
621 };
622
623 int __init init_hw_breakpoint(void)
624 {
625         unsigned int **task_bp_pinned;
626         int cpu, err_cpu;
627         int i;
628
629         for (i = 0; i < TYPE_MAX; i++)
630                 nr_slots[i] = hw_breakpoint_slots(i);
631
632         for_each_possible_cpu(cpu) {
633                 for (i = 0; i < TYPE_MAX; i++) {
634                         task_bp_pinned = &per_cpu(nr_task_bp_pinned[i], cpu);
635                         *task_bp_pinned = kzalloc(sizeof(int) * nr_slots[i],
636                                                   GFP_KERNEL);
637                         if (!*task_bp_pinned)
638                                 goto err_alloc;
639                 }
640         }
641
642         constraints_initialized = 1;
643
644         perf_pmu_register(&perf_breakpoint, "breakpoint", PERF_TYPE_BREAKPOINT);
645
646         return register_die_notifier(&hw_breakpoint_exceptions_nb);
647
648  err_alloc:
649         for_each_possible_cpu(err_cpu) {
650                 if (err_cpu == cpu)
651                         break;
652                 for (i = 0; i < TYPE_MAX; i++)
653                         kfree(per_cpu(nr_task_bp_pinned[i], cpu));
654         }
655
656         return -ENOMEM;
657 }
658
659