hw-breakpoints: Separate constraint space for data and instruction breakpoints
[linux-2.6.git] / kernel / 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/cpu.h>
44 #include <linux/smp.h>
45
46 #include <linux/hw_breakpoint.h>
47
48 enum bp_type_idx {
49         TYPE_INST       = 0,
50 #ifdef CONFIG_HAVE_MIXED_BREAKPOINTS_REGS
51         TYPE_DATA       = 0,
52 #else
53         TYPE_DATA       = 1,
54 #endif
55         TYPE_MAX
56 };
57
58 /*
59  * Constraints data
60  */
61
62 /* Number of pinned cpu breakpoints in a cpu */
63 static DEFINE_PER_CPU(unsigned int, nr_cpu_bp_pinned[TYPE_MAX]);
64
65 /* Number of pinned task breakpoints in a cpu */
66 static DEFINE_PER_CPU(unsigned int, nr_task_bp_pinned[TYPE_MAX][HBP_NUM]);
67
68 /* Number of non-pinned cpu/task breakpoints in a cpu */
69 static DEFINE_PER_CPU(unsigned int, nr_bp_flexible[TYPE_MAX]);
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 static inline enum bp_type_idx find_slot_idx(struct perf_event *bp)
81 {
82         if (bp->attr.bp_type & HW_BREAKPOINT_RW)
83                 return TYPE_DATA;
84
85         return TYPE_INST;
86 }
87
88 /*
89  * Report the maximum number of pinned breakpoints a task
90  * have in this cpu
91  */
92 static unsigned int max_task_bp_pinned(int cpu, enum bp_type_idx type)
93 {
94         int i;
95         unsigned int *tsk_pinned = per_cpu(nr_task_bp_pinned[type], cpu);
96
97         for (i = HBP_NUM -1; i >= 0; i--) {
98                 if (tsk_pinned[i] > 0)
99                         return i + 1;
100         }
101
102         return 0;
103 }
104
105 static int task_bp_pinned(struct task_struct *tsk, enum bp_type_idx type)
106 {
107         struct perf_event_context *ctx = tsk->perf_event_ctxp;
108         struct list_head *list;
109         struct perf_event *bp;
110         unsigned long flags;
111         int count = 0;
112
113         if (WARN_ONCE(!ctx, "No perf context for this task"))
114                 return 0;
115
116         list = &ctx->event_list;
117
118         raw_spin_lock_irqsave(&ctx->lock, flags);
119
120         /*
121          * The current breakpoint counter is not included in the list
122          * at the open() callback time
123          */
124         list_for_each_entry(bp, list, event_entry) {
125                 if (bp->attr.type == PERF_TYPE_BREAKPOINT)
126                         if (find_slot_idx(bp) == type)
127                                 count++;
128         }
129
130         raw_spin_unlock_irqrestore(&ctx->lock, flags);
131
132         return count;
133 }
134
135 /*
136  * Report the number of pinned/un-pinned breakpoints we have in
137  * a given cpu (cpu > -1) or in all of them (cpu = -1).
138  */
139 static void
140 fetch_bp_busy_slots(struct bp_busy_slots *slots, struct perf_event *bp,
141                     enum bp_type_idx type)
142 {
143         int cpu = bp->cpu;
144         struct task_struct *tsk = bp->ctx->task;
145
146         if (cpu >= 0) {
147                 slots->pinned = per_cpu(nr_cpu_bp_pinned[type], cpu);
148                 if (!tsk)
149                         slots->pinned += max_task_bp_pinned(cpu, type);
150                 else
151                         slots->pinned += task_bp_pinned(tsk, type);
152                 slots->flexible = per_cpu(nr_bp_flexible[type], cpu);
153
154                 return;
155         }
156
157         for_each_online_cpu(cpu) {
158                 unsigned int nr;
159
160                 nr = per_cpu(nr_cpu_bp_pinned[type], cpu);
161                 if (!tsk)
162                         nr += max_task_bp_pinned(cpu, type);
163                 else
164                         nr += task_bp_pinned(tsk, type);
165
166                 if (nr > slots->pinned)
167                         slots->pinned = nr;
168
169                 nr = per_cpu(nr_bp_flexible[type], cpu);
170
171                 if (nr > slots->flexible)
172                         slots->flexible = nr;
173         }
174 }
175
176 /*
177  * Add a pinned breakpoint for the given task in our constraint table
178  */
179 static void toggle_bp_task_slot(struct task_struct *tsk, int cpu, bool enable,
180                                 enum bp_type_idx type)
181 {
182         unsigned int *tsk_pinned;
183         int count = 0;
184
185         count = task_bp_pinned(tsk, type);
186
187         tsk_pinned = per_cpu(nr_task_bp_pinned[type], cpu);
188         if (enable) {
189                 tsk_pinned[count]++;
190                 if (count > 0)
191                         tsk_pinned[count-1]--;
192         } else {
193                 tsk_pinned[count]--;
194                 if (count > 0)
195                         tsk_pinned[count-1]++;
196         }
197 }
198
199 /*
200  * Add/remove the given breakpoint in our constraint table
201  */
202 static void
203 toggle_bp_slot(struct perf_event *bp, bool enable, enum bp_type_idx type)
204 {
205         int cpu = bp->cpu;
206         struct task_struct *tsk = bp->ctx->task;
207
208         /* Pinned counter task profiling */
209         if (tsk) {
210                 if (cpu >= 0) {
211                         toggle_bp_task_slot(tsk, cpu, enable, type);
212                         return;
213                 }
214
215                 for_each_online_cpu(cpu)
216                         toggle_bp_task_slot(tsk, cpu, enable, type);
217                 return;
218         }
219
220         /* Pinned counter cpu profiling */
221         if (enable)
222                 per_cpu(nr_cpu_bp_pinned[type], bp->cpu)++;
223         else
224                 per_cpu(nr_cpu_bp_pinned[type], bp->cpu)--;
225 }
226
227 /*
228  * Contraints to check before allowing this new breakpoint counter:
229  *
230  *  == Non-pinned counter == (Considered as pinned for now)
231  *
232  *   - If attached to a single cpu, check:
233  *
234  *       (per_cpu(nr_bp_flexible, cpu) || (per_cpu(nr_cpu_bp_pinned, cpu)
235  *           + max(per_cpu(nr_task_bp_pinned, cpu)))) < HBP_NUM
236  *
237  *       -> If there are already non-pinned counters in this cpu, it means
238  *          there is already a free slot for them.
239  *          Otherwise, we check that the maximum number of per task
240  *          breakpoints (for this cpu) plus the number of per cpu breakpoint
241  *          (for this cpu) doesn't cover every registers.
242  *
243  *   - If attached to every cpus, check:
244  *
245  *       (per_cpu(nr_bp_flexible, *) || (max(per_cpu(nr_cpu_bp_pinned, *))
246  *           + max(per_cpu(nr_task_bp_pinned, *)))) < HBP_NUM
247  *
248  *       -> This is roughly the same, except we check the number of per cpu
249  *          bp for every cpu and we keep the max one. Same for the per tasks
250  *          breakpoints.
251  *
252  *
253  * == Pinned counter ==
254  *
255  *   - If attached to a single cpu, check:
256  *
257  *       ((per_cpu(nr_bp_flexible, cpu) > 1) + per_cpu(nr_cpu_bp_pinned, cpu)
258  *            + max(per_cpu(nr_task_bp_pinned, cpu))) < HBP_NUM
259  *
260  *       -> Same checks as before. But now the nr_bp_flexible, if any, must keep
261  *          one register at least (or they will never be fed).
262  *
263  *   - If attached to every cpus, check:
264  *
265  *       ((per_cpu(nr_bp_flexible, *) > 1) + max(per_cpu(nr_cpu_bp_pinned, *))
266  *            + max(per_cpu(nr_task_bp_pinned, *))) < HBP_NUM
267  */
268 static int __reserve_bp_slot(struct perf_event *bp)
269 {
270         struct bp_busy_slots slots = {0};
271         enum bp_type_idx type;
272
273         /* Basic checks */
274         if (bp->attr.bp_type == HW_BREAKPOINT_EMPTY ||
275             bp->attr.bp_type == HW_BREAKPOINT_INVALID)
276                 return -EINVAL;
277
278         type = find_slot_idx(bp);
279         fetch_bp_busy_slots(&slots, bp, type);
280
281         /* Flexible counters need to keep at least one slot */
282         if (slots.pinned + (!!slots.flexible) == HBP_NUM)
283                 return -ENOSPC;
284
285         toggle_bp_slot(bp, true, type);
286
287         return 0;
288 }
289
290 int reserve_bp_slot(struct perf_event *bp)
291 {
292         int ret;
293
294         mutex_lock(&nr_bp_mutex);
295
296         ret = __reserve_bp_slot(bp);
297
298         mutex_unlock(&nr_bp_mutex);
299
300         return ret;
301 }
302
303 static void __release_bp_slot(struct perf_event *bp)
304 {
305         enum bp_type_idx type;
306
307         type = find_slot_idx(bp);
308         toggle_bp_slot(bp, false, type);
309 }
310
311 void release_bp_slot(struct perf_event *bp)
312 {
313         mutex_lock(&nr_bp_mutex);
314
315         __release_bp_slot(bp);
316
317         mutex_unlock(&nr_bp_mutex);
318 }
319
320 /*
321  * Allow the kernel debugger to reserve breakpoint slots without
322  * taking a lock using the dbg_* variant of for the reserve and
323  * release breakpoint slots.
324  */
325 int dbg_reserve_bp_slot(struct perf_event *bp)
326 {
327         if (mutex_is_locked(&nr_bp_mutex))
328                 return -1;
329
330         return __reserve_bp_slot(bp);
331 }
332
333 int dbg_release_bp_slot(struct perf_event *bp)
334 {
335         if (mutex_is_locked(&nr_bp_mutex))
336                 return -1;
337
338         __release_bp_slot(bp);
339
340         return 0;
341 }
342
343 static int validate_hw_breakpoint(struct perf_event *bp)
344 {
345         int ret;
346
347         ret = arch_validate_hwbkpt_settings(bp);
348         if (ret)
349                 return ret;
350
351         if (arch_check_bp_in_kernelspace(bp)) {
352                 if (bp->attr.exclude_kernel)
353                         return -EINVAL;
354                 /*
355                  * Don't let unprivileged users set a breakpoint in the trap
356                  * path to avoid trap recursion attacks.
357                  */
358                 if (!capable(CAP_SYS_ADMIN))
359                         return -EPERM;
360         }
361
362         return 0;
363 }
364
365 int register_perf_hw_breakpoint(struct perf_event *bp)
366 {
367         int ret;
368
369         ret = reserve_bp_slot(bp);
370         if (ret)
371                 return ret;
372
373         ret = validate_hw_breakpoint(bp);
374
375         /* if arch_validate_hwbkpt_settings() fails then release bp slot */
376         if (ret)
377                 release_bp_slot(bp);
378
379         return ret;
380 }
381
382 /**
383  * register_user_hw_breakpoint - register a hardware breakpoint for user space
384  * @attr: breakpoint attributes
385  * @triggered: callback to trigger when we hit the breakpoint
386  * @tsk: pointer to 'task_struct' of the process to which the address belongs
387  */
388 struct perf_event *
389 register_user_hw_breakpoint(struct perf_event_attr *attr,
390                             perf_overflow_handler_t triggered,
391                             struct task_struct *tsk)
392 {
393         return perf_event_create_kernel_counter(attr, -1, tsk->pid, triggered);
394 }
395 EXPORT_SYMBOL_GPL(register_user_hw_breakpoint);
396
397 /**
398  * modify_user_hw_breakpoint - modify a user-space hardware breakpoint
399  * @bp: the breakpoint structure to modify
400  * @attr: new breakpoint attributes
401  * @triggered: callback to trigger when we hit the breakpoint
402  * @tsk: pointer to 'task_struct' of the process to which the address belongs
403  */
404 int modify_user_hw_breakpoint(struct perf_event *bp, struct perf_event_attr *attr)
405 {
406         u64 old_addr = bp->attr.bp_addr;
407         u64 old_len = bp->attr.bp_len;
408         int old_type = bp->attr.bp_type;
409         int err = 0;
410
411         perf_event_disable(bp);
412
413         bp->attr.bp_addr = attr->bp_addr;
414         bp->attr.bp_type = attr->bp_type;
415         bp->attr.bp_len = attr->bp_len;
416
417         if (attr->disabled)
418                 goto end;
419
420         err = validate_hw_breakpoint(bp);
421         if (!err)
422                 perf_event_enable(bp);
423
424         if (err) {
425                 bp->attr.bp_addr = old_addr;
426                 bp->attr.bp_type = old_type;
427                 bp->attr.bp_len = old_len;
428                 if (!bp->attr.disabled)
429                         perf_event_enable(bp);
430
431                 return err;
432         }
433
434 end:
435         bp->attr.disabled = attr->disabled;
436
437         return 0;
438 }
439 EXPORT_SYMBOL_GPL(modify_user_hw_breakpoint);
440
441 /**
442  * unregister_hw_breakpoint - unregister a user-space hardware breakpoint
443  * @bp: the breakpoint structure to unregister
444  */
445 void unregister_hw_breakpoint(struct perf_event *bp)
446 {
447         if (!bp)
448                 return;
449         perf_event_release_kernel(bp);
450 }
451 EXPORT_SYMBOL_GPL(unregister_hw_breakpoint);
452
453 /**
454  * register_wide_hw_breakpoint - register a wide breakpoint in the kernel
455  * @attr: breakpoint attributes
456  * @triggered: callback to trigger when we hit the breakpoint
457  *
458  * @return a set of per_cpu pointers to perf events
459  */
460 struct perf_event * __percpu *
461 register_wide_hw_breakpoint(struct perf_event_attr *attr,
462                             perf_overflow_handler_t triggered)
463 {
464         struct perf_event * __percpu *cpu_events, **pevent, *bp;
465         long err;
466         int cpu;
467
468         cpu_events = alloc_percpu(typeof(*cpu_events));
469         if (!cpu_events)
470                 return (void __percpu __force *)ERR_PTR(-ENOMEM);
471
472         get_online_cpus();
473         for_each_online_cpu(cpu) {
474                 pevent = per_cpu_ptr(cpu_events, cpu);
475                 bp = perf_event_create_kernel_counter(attr, cpu, -1, triggered);
476
477                 *pevent = bp;
478
479                 if (IS_ERR(bp)) {
480                         err = PTR_ERR(bp);
481                         goto fail;
482                 }
483         }
484         put_online_cpus();
485
486         return cpu_events;
487
488 fail:
489         for_each_online_cpu(cpu) {
490                 pevent = per_cpu_ptr(cpu_events, cpu);
491                 if (IS_ERR(*pevent))
492                         break;
493                 unregister_hw_breakpoint(*pevent);
494         }
495         put_online_cpus();
496
497         free_percpu(cpu_events);
498         return (void __percpu __force *)ERR_PTR(err);
499 }
500 EXPORT_SYMBOL_GPL(register_wide_hw_breakpoint);
501
502 /**
503  * unregister_wide_hw_breakpoint - unregister a wide breakpoint in the kernel
504  * @cpu_events: the per cpu set of events to unregister
505  */
506 void unregister_wide_hw_breakpoint(struct perf_event * __percpu *cpu_events)
507 {
508         int cpu;
509         struct perf_event **pevent;
510
511         for_each_possible_cpu(cpu) {
512                 pevent = per_cpu_ptr(cpu_events, cpu);
513                 unregister_hw_breakpoint(*pevent);
514         }
515         free_percpu(cpu_events);
516 }
517 EXPORT_SYMBOL_GPL(unregister_wide_hw_breakpoint);
518
519 static struct notifier_block hw_breakpoint_exceptions_nb = {
520         .notifier_call = hw_breakpoint_exceptions_notify,
521         /* we need to be notified first */
522         .priority = 0x7fffffff
523 };
524
525 static int __init init_hw_breakpoint(void)
526 {
527         return register_die_notifier(&hw_breakpoint_exceptions_nb);
528 }
529 core_initcall(init_hw_breakpoint);
530
531
532 struct pmu perf_ops_bp = {
533         .enable         = arch_install_hw_breakpoint,
534         .disable        = arch_uninstall_hw_breakpoint,
535         .read           = hw_breakpoint_pmu_read,
536 };