xref: /linux/arch/x86/kernel/hw_breakpoint.c (revision 93df8a1ed6231727c5db94a80b1a6bd5ee67cec3)
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
18  * Copyright (C) 2009 Frederic Weisbecker <fweisbec@gmail.com>
19  *
20  * Authors: Alan Stern <stern@rowland.harvard.edu>
21  *          K.Prasad <prasad@linux.vnet.ibm.com>
22  *          Frederic Weisbecker <fweisbec@gmail.com>
23  */
24 
25 /*
26  * HW_breakpoint: a unified kernel/user-space hardware breakpoint facility,
27  * using the CPU's debug registers.
28  */
29 
30 #include <linux/perf_event.h>
31 #include <linux/hw_breakpoint.h>
32 #include <linux/irqflags.h>
33 #include <linux/notifier.h>
34 #include <linux/kallsyms.h>
35 #include <linux/kprobes.h>
36 #include <linux/percpu.h>
37 #include <linux/kdebug.h>
38 #include <linux/kernel.h>
39 #include <linux/module.h>
40 #include <linux/sched.h>
41 #include <linux/smp.h>
42 
43 #include <asm/hw_breakpoint.h>
44 #include <asm/processor.h>
45 #include <asm/debugreg.h>
46 
47 /* Per cpu debug control register value */
48 DEFINE_PER_CPU(unsigned long, cpu_dr7);
49 EXPORT_PER_CPU_SYMBOL(cpu_dr7);
50 
51 /* Per cpu debug address registers values */
52 static DEFINE_PER_CPU(unsigned long, cpu_debugreg[HBP_NUM]);
53 
54 /*
55  * Stores the breakpoints currently in use on each breakpoint address
56  * register for each cpus
57  */
58 static DEFINE_PER_CPU(struct perf_event *, bp_per_reg[HBP_NUM]);
59 
60 
61 static inline unsigned long
62 __encode_dr7(int drnum, unsigned int len, unsigned int type)
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 
70 	return bp_info;
71 }
72 
73 /*
74  * Encode the length, type, Exact, and Enable bits for a particular breakpoint
75  * as stored in debug register 7.
76  */
77 unsigned long encode_dr7(int drnum, unsigned int len, unsigned int type)
78 {
79 	return __encode_dr7(drnum, len, type) | DR_GLOBAL_SLOWDOWN;
80 }
81 
82 /*
83  * Decode the length and type bits for a particular breakpoint as
84  * stored in debug register 7.  Return the "enabled" status.
85  */
86 int decode_dr7(unsigned long dr7, int bpnum, unsigned *len, unsigned *type)
87 {
88 	int bp_info = dr7 >> (DR_CONTROL_SHIFT + bpnum * DR_CONTROL_SIZE);
89 
90 	*len = (bp_info & 0xc) | 0x40;
91 	*type = (bp_info & 0x3) | 0x80;
92 
93 	return (dr7 >> (bpnum * DR_ENABLE_SIZE)) & 0x3;
94 }
95 
96 /*
97  * Install a perf counter breakpoint.
98  *
99  * We seek a free debug address register and use it for this
100  * breakpoint. Eventually we enable it in the debug control register.
101  *
102  * Atomic: we hold the counter->ctx->lock and we only handle variables
103  * and registers local to this cpu.
104  */
105 int arch_install_hw_breakpoint(struct perf_event *bp)
106 {
107 	struct arch_hw_breakpoint *info = counter_arch_bp(bp);
108 	unsigned long *dr7;
109 	int i;
110 
111 	for (i = 0; i < HBP_NUM; i++) {
112 		struct perf_event **slot = this_cpu_ptr(&bp_per_reg[i]);
113 
114 		if (!*slot) {
115 			*slot = bp;
116 			break;
117 		}
118 	}
119 
120 	if (WARN_ONCE(i == HBP_NUM, "Can't find any breakpoint slot"))
121 		return -EBUSY;
122 
123 	set_debugreg(info->address, i);
124 	__this_cpu_write(cpu_debugreg[i], info->address);
125 
126 	dr7 = this_cpu_ptr(&cpu_dr7);
127 	*dr7 |= encode_dr7(i, info->len, info->type);
128 
129 	set_debugreg(*dr7, 7);
130 	if (info->mask)
131 		set_dr_addr_mask(info->mask, i);
132 
133 	return 0;
134 }
135 
136 /*
137  * Uninstall the breakpoint contained in the given counter.
138  *
139  * First we search the debug address register it uses and then we disable
140  * it.
141  *
142  * Atomic: we hold the counter->ctx->lock and we only handle variables
143  * and registers local to this cpu.
144  */
145 void arch_uninstall_hw_breakpoint(struct perf_event *bp)
146 {
147 	struct arch_hw_breakpoint *info = counter_arch_bp(bp);
148 	unsigned long *dr7;
149 	int i;
150 
151 	for (i = 0; i < HBP_NUM; i++) {
152 		struct perf_event **slot = this_cpu_ptr(&bp_per_reg[i]);
153 
154 		if (*slot == bp) {
155 			*slot = NULL;
156 			break;
157 		}
158 	}
159 
160 	if (WARN_ONCE(i == HBP_NUM, "Can't find any breakpoint slot"))
161 		return;
162 
163 	dr7 = this_cpu_ptr(&cpu_dr7);
164 	*dr7 &= ~__encode_dr7(i, info->len, info->type);
165 
166 	set_debugreg(*dr7, 7);
167 	if (info->mask)
168 		set_dr_addr_mask(0, i);
169 }
170 
171 /*
172  * Check for virtual address in kernel space.
173  */
174 int arch_check_bp_in_kernelspace(struct perf_event *bp)
175 {
176 	unsigned int len;
177 	unsigned long va;
178 	struct arch_hw_breakpoint *info = counter_arch_bp(bp);
179 
180 	va = info->address;
181 	len = bp->attr.bp_len;
182 
183 	/*
184 	 * We don't need to worry about va + len - 1 overflowing:
185 	 * we already require that va is aligned to a multiple of len.
186 	 */
187 	return (va >= TASK_SIZE_MAX) || ((va + len - 1) >= TASK_SIZE_MAX);
188 }
189 
190 int arch_bp_generic_fields(int x86_len, int x86_type,
191 			   int *gen_len, int *gen_type)
192 {
193 	/* Type */
194 	switch (x86_type) {
195 	case X86_BREAKPOINT_EXECUTE:
196 		if (x86_len != X86_BREAKPOINT_LEN_X)
197 			return -EINVAL;
198 
199 		*gen_type = HW_BREAKPOINT_X;
200 		*gen_len = sizeof(long);
201 		return 0;
202 	case X86_BREAKPOINT_WRITE:
203 		*gen_type = HW_BREAKPOINT_W;
204 		break;
205 	case X86_BREAKPOINT_RW:
206 		*gen_type = HW_BREAKPOINT_W | HW_BREAKPOINT_R;
207 		break;
208 	default:
209 		return -EINVAL;
210 	}
211 
212 	/* Len */
213 	switch (x86_len) {
214 	case X86_BREAKPOINT_LEN_1:
215 		*gen_len = HW_BREAKPOINT_LEN_1;
216 		break;
217 	case X86_BREAKPOINT_LEN_2:
218 		*gen_len = HW_BREAKPOINT_LEN_2;
219 		break;
220 	case X86_BREAKPOINT_LEN_4:
221 		*gen_len = HW_BREAKPOINT_LEN_4;
222 		break;
223 #ifdef CONFIG_X86_64
224 	case X86_BREAKPOINT_LEN_8:
225 		*gen_len = HW_BREAKPOINT_LEN_8;
226 		break;
227 #endif
228 	default:
229 		return -EINVAL;
230 	}
231 
232 	return 0;
233 }
234 
235 
236 static int arch_build_bp_info(struct perf_event *bp)
237 {
238 	struct arch_hw_breakpoint *info = counter_arch_bp(bp);
239 
240 	info->address = bp->attr.bp_addr;
241 
242 	/* Type */
243 	switch (bp->attr.bp_type) {
244 	case HW_BREAKPOINT_W:
245 		info->type = X86_BREAKPOINT_WRITE;
246 		break;
247 	case HW_BREAKPOINT_W | HW_BREAKPOINT_R:
248 		info->type = X86_BREAKPOINT_RW;
249 		break;
250 	case HW_BREAKPOINT_X:
251 		/*
252 		 * We don't allow kernel breakpoints in places that are not
253 		 * acceptable for kprobes.  On non-kprobes kernels, we don't
254 		 * allow kernel breakpoints at all.
255 		 */
256 		if (bp->attr.bp_addr >= TASK_SIZE_MAX) {
257 #ifdef CONFIG_KPROBES
258 			if (within_kprobe_blacklist(bp->attr.bp_addr))
259 				return -EINVAL;
260 #else
261 			return -EINVAL;
262 #endif
263 		}
264 
265 		info->type = X86_BREAKPOINT_EXECUTE;
266 		/*
267 		 * x86 inst breakpoints need to have a specific undefined len.
268 		 * But we still need to check userspace is not trying to setup
269 		 * an unsupported length, to get a range breakpoint for example.
270 		 */
271 		if (bp->attr.bp_len == sizeof(long)) {
272 			info->len = X86_BREAKPOINT_LEN_X;
273 			return 0;
274 		}
275 	default:
276 		return -EINVAL;
277 	}
278 
279 	/* Len */
280 	info->mask = 0;
281 
282 	switch (bp->attr.bp_len) {
283 	case HW_BREAKPOINT_LEN_1:
284 		info->len = X86_BREAKPOINT_LEN_1;
285 		break;
286 	case HW_BREAKPOINT_LEN_2:
287 		info->len = X86_BREAKPOINT_LEN_2;
288 		break;
289 	case HW_BREAKPOINT_LEN_4:
290 		info->len = X86_BREAKPOINT_LEN_4;
291 		break;
292 #ifdef CONFIG_X86_64
293 	case HW_BREAKPOINT_LEN_8:
294 		info->len = X86_BREAKPOINT_LEN_8;
295 		break;
296 #endif
297 	default:
298 		/* AMD range breakpoint */
299 		if (!is_power_of_2(bp->attr.bp_len))
300 			return -EINVAL;
301 		if (bp->attr.bp_addr & (bp->attr.bp_len - 1))
302 			return -EINVAL;
303 		/*
304 		 * It's impossible to use a range breakpoint to fake out
305 		 * user vs kernel detection because bp_len - 1 can't
306 		 * have the high bit set.  If we ever allow range instruction
307 		 * breakpoints, then we'll have to check for kprobe-blacklisted
308 		 * addresses anywhere in the range.
309 		 */
310 		if (!cpu_has_bpext)
311 			return -EOPNOTSUPP;
312 		info->mask = bp->attr.bp_len - 1;
313 		info->len = X86_BREAKPOINT_LEN_1;
314 	}
315 
316 	return 0;
317 }
318 
319 /*
320  * Validate the arch-specific HW Breakpoint register settings
321  */
322 int arch_validate_hwbkpt_settings(struct perf_event *bp)
323 {
324 	struct arch_hw_breakpoint *info = counter_arch_bp(bp);
325 	unsigned int align;
326 	int ret;
327 
328 
329 	ret = arch_build_bp_info(bp);
330 	if (ret)
331 		return ret;
332 
333 	switch (info->len) {
334 	case X86_BREAKPOINT_LEN_1:
335 		align = 0;
336 		if (info->mask)
337 			align = info->mask;
338 		break;
339 	case X86_BREAKPOINT_LEN_2:
340 		align = 1;
341 		break;
342 	case X86_BREAKPOINT_LEN_4:
343 		align = 3;
344 		break;
345 #ifdef CONFIG_X86_64
346 	case X86_BREAKPOINT_LEN_8:
347 		align = 7;
348 		break;
349 #endif
350 	default:
351 		WARN_ON_ONCE(1);
352 	}
353 
354 	/*
355 	 * Check that the low-order bits of the address are appropriate
356 	 * for the alignment implied by len.
357 	 */
358 	if (info->address & align)
359 		return -EINVAL;
360 
361 	return 0;
362 }
363 
364 /*
365  * Dump the debug register contents to the user.
366  * We can't dump our per cpu values because it
367  * may contain cpu wide breakpoint, something that
368  * doesn't belong to the current task.
369  *
370  * TODO: include non-ptrace user breakpoints (perf)
371  */
372 void aout_dump_debugregs(struct user *dump)
373 {
374 	int i;
375 	int dr7 = 0;
376 	struct perf_event *bp;
377 	struct arch_hw_breakpoint *info;
378 	struct thread_struct *thread = &current->thread;
379 
380 	for (i = 0; i < HBP_NUM; i++) {
381 		bp = thread->ptrace_bps[i];
382 
383 		if (bp && !bp->attr.disabled) {
384 			dump->u_debugreg[i] = bp->attr.bp_addr;
385 			info = counter_arch_bp(bp);
386 			dr7 |= encode_dr7(i, info->len, info->type);
387 		} else {
388 			dump->u_debugreg[i] = 0;
389 		}
390 	}
391 
392 	dump->u_debugreg[4] = 0;
393 	dump->u_debugreg[5] = 0;
394 	dump->u_debugreg[6] = current->thread.debugreg6;
395 
396 	dump->u_debugreg[7] = dr7;
397 }
398 EXPORT_SYMBOL_GPL(aout_dump_debugregs);
399 
400 /*
401  * Release the user breakpoints used by ptrace
402  */
403 void flush_ptrace_hw_breakpoint(struct task_struct *tsk)
404 {
405 	int i;
406 	struct thread_struct *t = &tsk->thread;
407 
408 	for (i = 0; i < HBP_NUM; i++) {
409 		unregister_hw_breakpoint(t->ptrace_bps[i]);
410 		t->ptrace_bps[i] = NULL;
411 	}
412 
413 	t->debugreg6 = 0;
414 	t->ptrace_dr7 = 0;
415 }
416 
417 void hw_breakpoint_restore(void)
418 {
419 	set_debugreg(__this_cpu_read(cpu_debugreg[0]), 0);
420 	set_debugreg(__this_cpu_read(cpu_debugreg[1]), 1);
421 	set_debugreg(__this_cpu_read(cpu_debugreg[2]), 2);
422 	set_debugreg(__this_cpu_read(cpu_debugreg[3]), 3);
423 	set_debugreg(current->thread.debugreg6, 6);
424 	set_debugreg(__this_cpu_read(cpu_dr7), 7);
425 }
426 EXPORT_SYMBOL_GPL(hw_breakpoint_restore);
427 
428 /*
429  * Handle debug exception notifications.
430  *
431  * Return value is either NOTIFY_STOP or NOTIFY_DONE as explained below.
432  *
433  * NOTIFY_DONE returned if one of the following conditions is true.
434  * i) When the causative address is from user-space and the exception
435  * is a valid one, i.e. not triggered as a result of lazy debug register
436  * switching
437  * ii) When there are more bits than trap<n> set in DR6 register (such
438  * as BD, BS or BT) indicating that more than one debug condition is
439  * met and requires some more action in do_debug().
440  *
441  * NOTIFY_STOP returned for all other cases
442  *
443  */
444 static int hw_breakpoint_handler(struct die_args *args)
445 {
446 	int i, cpu, rc = NOTIFY_STOP;
447 	struct perf_event *bp;
448 	unsigned long dr7, dr6;
449 	unsigned long *dr6_p;
450 
451 	/* The DR6 value is pointed by args->err */
452 	dr6_p = (unsigned long *)ERR_PTR(args->err);
453 	dr6 = *dr6_p;
454 
455 	/* If it's a single step, TRAP bits are random */
456 	if (dr6 & DR_STEP)
457 		return NOTIFY_DONE;
458 
459 	/* Do an early return if no trap bits are set in DR6 */
460 	if ((dr6 & DR_TRAP_BITS) == 0)
461 		return NOTIFY_DONE;
462 
463 	get_debugreg(dr7, 7);
464 	/* Disable breakpoints during exception handling */
465 	set_debugreg(0UL, 7);
466 	/*
467 	 * Assert that local interrupts are disabled
468 	 * Reset the DRn bits in the virtualized register value.
469 	 * The ptrace trigger routine will add in whatever is needed.
470 	 */
471 	current->thread.debugreg6 &= ~DR_TRAP_BITS;
472 	cpu = get_cpu();
473 
474 	/* Handle all the breakpoints that were triggered */
475 	for (i = 0; i < HBP_NUM; ++i) {
476 		if (likely(!(dr6 & (DR_TRAP0 << i))))
477 			continue;
478 
479 		/*
480 		 * The counter may be concurrently released but that can only
481 		 * occur from a call_rcu() path. We can then safely fetch
482 		 * the breakpoint, use its callback, touch its counter
483 		 * while we are in an rcu_read_lock() path.
484 		 */
485 		rcu_read_lock();
486 
487 		bp = per_cpu(bp_per_reg[i], cpu);
488 		/*
489 		 * Reset the 'i'th TRAP bit in dr6 to denote completion of
490 		 * exception handling
491 		 */
492 		(*dr6_p) &= ~(DR_TRAP0 << i);
493 		/*
494 		 * bp can be NULL due to lazy debug register switching
495 		 * or due to concurrent perf counter removing.
496 		 */
497 		if (!bp) {
498 			rcu_read_unlock();
499 			break;
500 		}
501 
502 		perf_bp_event(bp, args->regs);
503 
504 		/*
505 		 * Set up resume flag to avoid breakpoint recursion when
506 		 * returning back to origin.
507 		 */
508 		if (bp->hw.info.type == X86_BREAKPOINT_EXECUTE)
509 			args->regs->flags |= X86_EFLAGS_RF;
510 
511 		rcu_read_unlock();
512 	}
513 	/*
514 	 * Further processing in do_debug() is needed for a) user-space
515 	 * breakpoints (to generate signals) and b) when the system has
516 	 * taken exception due to multiple causes
517 	 */
518 	if ((current->thread.debugreg6 & DR_TRAP_BITS) ||
519 	    (dr6 & (~DR_TRAP_BITS)))
520 		rc = NOTIFY_DONE;
521 
522 	set_debugreg(dr7, 7);
523 	put_cpu();
524 
525 	return rc;
526 }
527 
528 /*
529  * Handle debug exception notifications.
530  */
531 int hw_breakpoint_exceptions_notify(
532 		struct notifier_block *unused, unsigned long val, void *data)
533 {
534 	if (val != DIE_DEBUG)
535 		return NOTIFY_DONE;
536 
537 	return hw_breakpoint_handler(data);
538 }
539 
540 void hw_breakpoint_pmu_read(struct perf_event *bp)
541 {
542 	/* TODO */
543 }
544