perf_events: Restore sanity to scaling land
[deliverable/linux.git] / arch / x86 / kernel / hw_breakpoint.c
CommitLineData
0067f129
P
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) 2009 IBM Corporation
24f1e32c 18 * Copyright (C) 2009 Frederic Weisbecker <fweisbec@gmail.com>
0067f129
P
19 */
20
21/*
22 * HW_breakpoint: a unified kernel/user-space hardware breakpoint facility,
23 * using the CPU's debug registers.
24 */
25
24f1e32c
FW
26#include <linux/perf_event.h>
27#include <linux/hw_breakpoint.h>
0067f129
P
28#include <linux/irqflags.h>
29#include <linux/notifier.h>
30#include <linux/kallsyms.h>
31#include <linux/kprobes.h>
32#include <linux/percpu.h>
33#include <linux/kdebug.h>
34#include <linux/kernel.h>
35#include <linux/module.h>
36#include <linux/sched.h>
37#include <linux/init.h>
38#include <linux/smp.h>
39
40#include <asm/hw_breakpoint.h>
41#include <asm/processor.h>
42#include <asm/debugreg.h>
43
24f1e32c
FW
44/* Per cpu debug control register value */
45DEFINE_PER_CPU(unsigned long, dr7);
68efa37d 46EXPORT_PER_CPU_SYMBOL(dr7);
24f1e32c
FW
47
48/* Per cpu debug address registers values */
49static DEFINE_PER_CPU(unsigned long, cpu_debugreg[HBP_NUM]);
0067f129
P
50
51/*
24f1e32c
FW
52 * Stores the breakpoints currently in use on each breakpoint address
53 * register for each cpus
0067f129 54 */
24f1e32c 55static DEFINE_PER_CPU(struct perf_event *, bp_per_reg[HBP_NUM]);
0067f129
P
56
57
58/*
59 * Encode the length, type, Exact, and Enable bits for a particular breakpoint
60 * as stored in debug register 7.
61 */
24f1e32c 62unsigned long encode_dr7(int drnum, unsigned int len, unsigned int type)
0067f129
P
63{
64 unsigned long bp_info;
65
66 bp_info = (len | type) & 0xf;
67 bp_info <<= (DR_CONTROL_SHIFT + drnum * DR_CONTROL_SIZE);
68 bp_info |= (DR_GLOBAL_ENABLE << (drnum * DR_ENABLE_SIZE)) |
69 DR_GLOBAL_SLOWDOWN;
70 return bp_info;
71}
72
24f1e32c
FW
73/*
74 * Decode the length and type bits for a particular breakpoint as
75 * stored in debug register 7. Return the "enabled" status.
76 */
77int decode_dr7(unsigned long dr7, int bpnum, unsigned *len, unsigned *type)
0067f129 78{
24f1e32c 79 int bp_info = dr7 >> (DR_CONTROL_SHIFT + bpnum * DR_CONTROL_SIZE);
0067f129 80
24f1e32c
FW
81 *len = (bp_info & 0xc) | 0x40;
82 *type = (bp_info & 0x3) | 0x80;
0067f129 83
24f1e32c 84 return (dr7 >> (bpnum * DR_ENABLE_SIZE)) & 0x3;
0067f129
P
85}
86
87/*
24f1e32c
FW
88 * Install a perf counter breakpoint.
89 *
90 * We seek a free debug address register and use it for this
91 * breakpoint. Eventually we enable it in the debug control register.
92 *
93 * Atomic: we hold the counter->ctx->lock and we only handle variables
94 * and registers local to this cpu.
0067f129 95 */
24f1e32c 96int arch_install_hw_breakpoint(struct perf_event *bp)
0067f129 97{
24f1e32c
FW
98 struct arch_hw_breakpoint *info = counter_arch_bp(bp);
99 unsigned long *dr7;
100 int i;
101
102 for (i = 0; i < HBP_NUM; i++) {
103 struct perf_event **slot = &__get_cpu_var(bp_per_reg[i]);
104
105 if (!*slot) {
106 *slot = bp;
107 break;
108 }
0067f129
P
109 }
110
24f1e32c
FW
111 if (WARN_ONCE(i == HBP_NUM, "Can't find any breakpoint slot"))
112 return -EBUSY;
113
114 set_debugreg(info->address, i);
115 __get_cpu_var(cpu_debugreg[i]) = info->address;
116
117 dr7 = &__get_cpu_var(dr7);
118 *dr7 |= encode_dr7(i, info->len, info->type);
119
120 set_debugreg(*dr7, 7);
121
122 return 0;
0067f129
P
123}
124
125/*
24f1e32c
FW
126 * Uninstall the breakpoint contained in the given counter.
127 *
128 * First we search the debug address register it uses and then we disable
129 * it.
130 *
131 * Atomic: we hold the counter->ctx->lock and we only handle variables
132 * and registers local to this cpu.
0067f129 133 */
24f1e32c 134void arch_uninstall_hw_breakpoint(struct perf_event *bp)
0067f129 135{
24f1e32c
FW
136 struct arch_hw_breakpoint *info = counter_arch_bp(bp);
137 unsigned long *dr7;
138 int i;
139
140 for (i = 0; i < HBP_NUM; i++) {
141 struct perf_event **slot = &__get_cpu_var(bp_per_reg[i]);
142
143 if (*slot == bp) {
144 *slot = NULL;
145 break;
146 }
147 }
148
149 if (WARN_ONCE(i == HBP_NUM, "Can't find any breakpoint slot"))
150 return;
0067f129 151
24f1e32c
FW
152 dr7 = &__get_cpu_var(dr7);
153 *dr7 &= ~encode_dr7(i, info->len, info->type);
154
155 set_debugreg(*dr7, 7);
0067f129
P
156}
157
158static int get_hbp_len(u8 hbp_len)
159{
160 unsigned int len_in_bytes = 0;
161
162 switch (hbp_len) {
24f1e32c 163 case X86_BREAKPOINT_LEN_1:
0067f129
P
164 len_in_bytes = 1;
165 break;
24f1e32c 166 case X86_BREAKPOINT_LEN_2:
0067f129
P
167 len_in_bytes = 2;
168 break;
24f1e32c 169 case X86_BREAKPOINT_LEN_4:
0067f129
P
170 len_in_bytes = 4;
171 break;
172#ifdef CONFIG_X86_64
24f1e32c 173 case X86_BREAKPOINT_LEN_8:
0067f129
P
174 len_in_bytes = 8;
175 break;
176#endif
177 }
178 return len_in_bytes;
179}
180
181/*
182 * Check for virtual address in user space.
183 */
184int arch_check_va_in_userspace(unsigned long va, u8 hbp_len)
185{
186 unsigned int len;
187
188 len = get_hbp_len(hbp_len);
189
190 return (va <= TASK_SIZE - len);
191}
192
193/*
194 * Check for virtual address in kernel space.
195 */
4555835b 196static int arch_check_va_in_kernelspace(unsigned long va, u8 hbp_len)
0067f129
P
197{
198 unsigned int len;
199
200 len = get_hbp_len(hbp_len);
201
202 return (va >= TASK_SIZE) && ((va + len - 1) >= TASK_SIZE);
203}
204
205/*
206 * Store a breakpoint's encoded address, length, and type.
207 */
24f1e32c 208static int arch_store_info(struct perf_event *bp)
0067f129 209{
24f1e32c 210 struct arch_hw_breakpoint *info = counter_arch_bp(bp);
0067f129
P
211 /*
212 * For kernel-addresses, either the address or symbol name can be
213 * specified.
214 */
24f1e32c
FW
215 if (info->name)
216 info->address = (unsigned long)
217 kallsyms_lookup_name(info->name);
218 if (info->address)
0067f129 219 return 0;
24f1e32c 220
0067f129
P
221 return -EINVAL;
222}
223
24f1e32c
FW
224int arch_bp_generic_fields(int x86_len, int x86_type,
225 int *gen_len, int *gen_type)
0067f129 226{
24f1e32c
FW
227 /* Len */
228 switch (x86_len) {
229 case X86_BREAKPOINT_LEN_1:
230 *gen_len = HW_BREAKPOINT_LEN_1;
231 break;
232 case X86_BREAKPOINT_LEN_2:
233 *gen_len = HW_BREAKPOINT_LEN_2;
234 break;
235 case X86_BREAKPOINT_LEN_4:
236 *gen_len = HW_BREAKPOINT_LEN_4;
237 break;
238#ifdef CONFIG_X86_64
239 case X86_BREAKPOINT_LEN_8:
240 *gen_len = HW_BREAKPOINT_LEN_8;
241 break;
242#endif
243 default:
244 return -EINVAL;
245 }
0067f129 246
24f1e32c
FW
247 /* Type */
248 switch (x86_type) {
249 case X86_BREAKPOINT_EXECUTE:
250 *gen_type = HW_BREAKPOINT_X;
0067f129 251 break;
24f1e32c
FW
252 case X86_BREAKPOINT_WRITE:
253 *gen_type = HW_BREAKPOINT_W;
0067f129 254 break;
24f1e32c
FW
255 case X86_BREAKPOINT_RW:
256 *gen_type = HW_BREAKPOINT_W | HW_BREAKPOINT_R;
0067f129
P
257 break;
258 default:
24f1e32c 259 return -EINVAL;
0067f129
P
260 }
261
24f1e32c
FW
262 return 0;
263}
264
265
266static int arch_build_bp_info(struct perf_event *bp)
267{
268 struct arch_hw_breakpoint *info = counter_arch_bp(bp);
269
270 info->address = bp->attr.bp_addr;
271
272 /* Len */
273 switch (bp->attr.bp_len) {
0067f129 274 case HW_BREAKPOINT_LEN_1:
24f1e32c 275 info->len = X86_BREAKPOINT_LEN_1;
0067f129
P
276 break;
277 case HW_BREAKPOINT_LEN_2:
24f1e32c 278 info->len = X86_BREAKPOINT_LEN_2;
0067f129
P
279 break;
280 case HW_BREAKPOINT_LEN_4:
24f1e32c 281 info->len = X86_BREAKPOINT_LEN_4;
0067f129
P
282 break;
283#ifdef CONFIG_X86_64
284 case HW_BREAKPOINT_LEN_8:
24f1e32c
FW
285 info->len = X86_BREAKPOINT_LEN_8;
286 break;
287#endif
288 default:
289 return -EINVAL;
290 }
291
292 /* Type */
293 switch (bp->attr.bp_type) {
294 case HW_BREAKPOINT_W:
295 info->type = X86_BREAKPOINT_WRITE;
296 break;
297 case HW_BREAKPOINT_W | HW_BREAKPOINT_R:
298 info->type = X86_BREAKPOINT_RW;
299 break;
300 case HW_BREAKPOINT_X:
301 info->type = X86_BREAKPOINT_EXECUTE;
302 break;
303 default:
304 return -EINVAL;
305 }
306
307 return 0;
308}
309/*
310 * Validate the arch-specific HW Breakpoint register settings
311 */
312int arch_validate_hwbkpt_settings(struct perf_event *bp,
313 struct task_struct *tsk)
314{
315 struct arch_hw_breakpoint *info = counter_arch_bp(bp);
316 unsigned int align;
317 int ret;
318
319
320 ret = arch_build_bp_info(bp);
321 if (ret)
322 return ret;
323
324 ret = -EINVAL;
325
326 if (info->type == X86_BREAKPOINT_EXECUTE)
327 /*
328 * Ptrace-refactoring code
329 * For now, we'll allow instruction breakpoint only for user-space
330 * addresses
331 */
332 if ((!arch_check_va_in_userspace(info->address, info->len)) &&
333 info->len != X86_BREAKPOINT_EXECUTE)
334 return ret;
335
336 switch (info->len) {
337 case X86_BREAKPOINT_LEN_1:
338 align = 0;
339 break;
340 case X86_BREAKPOINT_LEN_2:
341 align = 1;
342 break;
343 case X86_BREAKPOINT_LEN_4:
344 align = 3;
345 break;
346#ifdef CONFIG_X86_64
347 case X86_BREAKPOINT_LEN_8:
0067f129
P
348 align = 7;
349 break;
350#endif
351 default:
352 return ret;
353 }
354
24f1e32c
FW
355 if (bp->callback)
356 ret = arch_store_info(bp);
0067f129
P
357
358 if (ret < 0)
359 return ret;
360 /*
361 * Check that the low-order bits of the address are appropriate
362 * for the alignment implied by len.
363 */
24f1e32c 364 if (info->address & align)
0067f129
P
365 return -EINVAL;
366
367 /* Check that the virtual address is in the proper range */
368 if (tsk) {
24f1e32c 369 if (!arch_check_va_in_userspace(info->address, info->len))
0067f129
P
370 return -EFAULT;
371 } else {
24f1e32c 372 if (!arch_check_va_in_kernelspace(info->address, info->len))
0067f129
P
373 return -EFAULT;
374 }
24f1e32c 375
0067f129
P
376 return 0;
377}
378
9f6b3c2c
FW
379/*
380 * Dump the debug register contents to the user.
381 * We can't dump our per cpu values because it
382 * may contain cpu wide breakpoint, something that
383 * doesn't belong to the current task.
384 *
385 * TODO: include non-ptrace user breakpoints (perf)
386 */
387void aout_dump_debugregs(struct user *dump)
388{
389 int i;
390 int dr7 = 0;
391 struct perf_event *bp;
392 struct arch_hw_breakpoint *info;
393 struct thread_struct *thread = &current->thread;
394
395 for (i = 0; i < HBP_NUM; i++) {
396 bp = thread->ptrace_bps[i];
397
398 if (bp && !bp->attr.disabled) {
399 dump->u_debugreg[i] = bp->attr.bp_addr;
400 info = counter_arch_bp(bp);
401 dr7 |= encode_dr7(i, info->len, info->type);
402 } else {
403 dump->u_debugreg[i] = 0;
404 }
405 }
406
407 dump->u_debugreg[4] = 0;
408 dump->u_debugreg[5] = 0;
409 dump->u_debugreg[6] = current->thread.debugreg6;
410
411 dump->u_debugreg[7] = dr7;
412}
68efa37d 413EXPORT_SYMBOL_GPL(aout_dump_debugregs);
9f6b3c2c 414
24f1e32c
FW
415/*
416 * Release the user breakpoints used by ptrace
417 */
418void flush_ptrace_hw_breakpoint(struct task_struct *tsk)
0067f129 419{
24f1e32c
FW
420 int i;
421 struct thread_struct *t = &tsk->thread;
422
423 for (i = 0; i < HBP_NUM; i++) {
424 unregister_hw_breakpoint(t->ptrace_bps[i]);
425 t->ptrace_bps[i] = NULL;
426 }
0067f129
P
427}
428
24f1e32c 429void hw_breakpoint_restore(void)
0067f129 430{
24f1e32c
FW
431 set_debugreg(__get_cpu_var(cpu_debugreg[0]), 0);
432 set_debugreg(__get_cpu_var(cpu_debugreg[1]), 1);
433 set_debugreg(__get_cpu_var(cpu_debugreg[2]), 2);
434 set_debugreg(__get_cpu_var(cpu_debugreg[3]), 3);
435 set_debugreg(current->thread.debugreg6, 6);
436 set_debugreg(__get_cpu_var(dr7), 7);
0067f129 437}
24f1e32c 438EXPORT_SYMBOL_GPL(hw_breakpoint_restore);
0067f129
P
439
440/*
441 * Handle debug exception notifications.
442 *
443 * Return value is either NOTIFY_STOP or NOTIFY_DONE as explained below.
444 *
445 * NOTIFY_DONE returned if one of the following conditions is true.
446 * i) When the causative address is from user-space and the exception
447 * is a valid one, i.e. not triggered as a result of lazy debug register
448 * switching
449 * ii) When there are more bits than trap<n> set in DR6 register (such
450 * as BD, BS or BT) indicating that more than one debug condition is
451 * met and requires some more action in do_debug().
452 *
453 * NOTIFY_STOP returned for all other cases
454 *
455 */
4555835b 456static int __kprobes hw_breakpoint_handler(struct die_args *args)
0067f129
P
457{
458 int i, cpu, rc = NOTIFY_STOP;
24f1e32c 459 struct perf_event *bp;
62edab90
P
460 unsigned long dr7, dr6;
461 unsigned long *dr6_p;
462
463 /* The DR6 value is pointed by args->err */
464 dr6_p = (unsigned long *)ERR_PTR(args->err);
465 dr6 = *dr6_p;
0067f129
P
466
467 /* Do an early return if no trap bits are set in DR6 */
468 if ((dr6 & DR_TRAP_BITS) == 0)
469 return NOTIFY_DONE;
470
0067f129
P
471 get_debugreg(dr7, 7);
472 /* Disable breakpoints during exception handling */
473 set_debugreg(0UL, 7);
474 /*
475 * Assert that local interrupts are disabled
476 * Reset the DRn bits in the virtualized register value.
477 * The ptrace trigger routine will add in whatever is needed.
478 */
479 current->thread.debugreg6 &= ~DR_TRAP_BITS;
480 cpu = get_cpu();
481
482 /* Handle all the breakpoints that were triggered */
483 for (i = 0; i < HBP_NUM; ++i) {
484 if (likely(!(dr6 & (DR_TRAP0 << i))))
485 continue;
24f1e32c 486
0067f129 487 /*
24f1e32c
FW
488 * The counter may be concurrently released but that can only
489 * occur from a call_rcu() path. We can then safely fetch
490 * the breakpoint, use its callback, touch its counter
491 * while we are in an rcu_read_lock() path.
0067f129 492 */
24f1e32c
FW
493 rcu_read_lock();
494
495 bp = per_cpu(bp_per_reg[i], cpu);
496 if (bp)
497 rc = NOTIFY_DONE;
62edab90
P
498 /*
499 * Reset the 'i'th TRAP bit in dr6 to denote completion of
500 * exception handling
501 */
502 (*dr6_p) &= ~(DR_TRAP0 << i);
0067f129
P
503 /*
504 * bp can be NULL due to lazy debug register switching
24f1e32c 505 * or due to concurrent perf counter removing.
0067f129 506 */
24f1e32c
FW
507 if (!bp) {
508 rcu_read_unlock();
509 break;
510 }
511
512 (bp->callback)(bp, args->regs);
0067f129 513
24f1e32c 514 rcu_read_unlock();
0067f129
P
515 }
516 if (dr6 & (~DR_TRAP_BITS))
517 rc = NOTIFY_DONE;
518
519 set_debugreg(dr7, 7);
eadb8a09 520 put_cpu();
24f1e32c 521
0067f129
P
522 return rc;
523}
524
525/*
526 * Handle debug exception notifications.
527 */
528int __kprobes hw_breakpoint_exceptions_notify(
529 struct notifier_block *unused, unsigned long val, void *data)
530{
531 if (val != DIE_DEBUG)
532 return NOTIFY_DONE;
533
534 return hw_breakpoint_handler(data);
535}
24f1e32c
FW
536
537void hw_breakpoint_pmu_read(struct perf_event *bp)
538{
539 /* TODO */
540}
541
542void hw_breakpoint_pmu_unthrottle(struct perf_event *bp)
543{
544 /* TODO */
545}
This page took 0.051843 seconds and 5 git commands to generate.