xref: /linux/arch/x86/kernel/hw_breakpoint.c (revision e58e871becec2d3b04ed91c0c16fe8deac9c9dfa)
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/export.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 #include <asm/user.h>
47 
48 /* Per cpu debug control register value */
49 DEFINE_PER_CPU(unsigned long, cpu_dr7);
50 EXPORT_PER_CPU_SYMBOL(cpu_dr7);
51 
52 /* Per cpu debug address registers values */
53 static DEFINE_PER_CPU(unsigned long, cpu_debugreg[HBP_NUM]);
54 
55 /*
56  * Stores the breakpoints currently in use on each breakpoint address
57  * register for each cpus
58  */
59 static DEFINE_PER_CPU(struct perf_event *, bp_per_reg[HBP_NUM]);
60 
61 
62 static inline unsigned long
63 __encode_dr7(int drnum, unsigned int len, unsigned int type)
64 {
65 	unsigned long bp_info;
66 
67 	bp_info = (len | type) & 0xf;
68 	bp_info <<= (DR_CONTROL_SHIFT + drnum * DR_CONTROL_SIZE);
69 	bp_info |= (DR_GLOBAL_ENABLE << (drnum * DR_ENABLE_SIZE));
70 
71 	return bp_info;
72 }
73 
74 /*
75  * Encode the length, type, Exact, and Enable bits for a particular breakpoint
76  * as stored in debug register 7.
77  */
78 unsigned long encode_dr7(int drnum, unsigned int len, unsigned int type)
79 {
80 	return __encode_dr7(drnum, len, type) | DR_GLOBAL_SLOWDOWN;
81 }
82 
83 /*
84  * Decode the length and type bits for a particular breakpoint as
85  * stored in debug register 7.  Return the "enabled" status.
86  */
87 int decode_dr7(unsigned long dr7, int bpnum, unsigned *len, unsigned *type)
88 {
89 	int bp_info = dr7 >> (DR_CONTROL_SHIFT + bpnum * DR_CONTROL_SIZE);
90 
91 	*len = (bp_info & 0xc) | 0x40;
92 	*type = (bp_info & 0x3) | 0x80;
93 
94 	return (dr7 >> (bpnum * DR_ENABLE_SIZE)) & 0x3;
95 }
96 
97 /*
98  * Install a perf counter breakpoint.
99  *
100  * We seek a free debug address register and use it for this
101  * breakpoint. Eventually we enable it in the debug control register.
102  *
103  * Atomic: we hold the counter->ctx->lock and we only handle variables
104  * and registers local to this cpu.
105  */
106 int arch_install_hw_breakpoint(struct perf_event *bp)
107 {
108 	struct arch_hw_breakpoint *info = counter_arch_bp(bp);
109 	unsigned long *dr7;
110 	int i;
111 
112 	for (i = 0; i < HBP_NUM; i++) {
113 		struct perf_event **slot = this_cpu_ptr(&bp_per_reg[i]);
114 
115 		if (!*slot) {
116 			*slot = bp;
117 			break;
118 		}
119 	}
120 
121 	if (WARN_ONCE(i == HBP_NUM, "Can't find any breakpoint slot"))
122 		return -EBUSY;
123 
124 	set_debugreg(info->address, i);
125 	__this_cpu_write(cpu_debugreg[i], info->address);
126 
127 	dr7 = this_cpu_ptr(&cpu_dr7);
128 	*dr7 |= encode_dr7(i, info->len, info->type);
129 
130 	set_debugreg(*dr7, 7);
131 	if (info->mask)
132 		set_dr_addr_mask(info->mask, i);
133 
134 	return 0;
135 }
136 
137 /*
138  * Uninstall the breakpoint contained in the given counter.
139  *
140  * First we search the debug address register it uses and then we disable
141  * it.
142  *
143  * Atomic: we hold the counter->ctx->lock and we only handle variables
144  * and registers local to this cpu.
145  */
146 void arch_uninstall_hw_breakpoint(struct perf_event *bp)
147 {
148 	struct arch_hw_breakpoint *info = counter_arch_bp(bp);
149 	unsigned long *dr7;
150 	int i;
151 
152 	for (i = 0; i < HBP_NUM; i++) {
153 		struct perf_event **slot = this_cpu_ptr(&bp_per_reg[i]);
154 
155 		if (*slot == bp) {
156 			*slot = NULL;
157 			break;
158 		}
159 	}
160 
161 	if (WARN_ONCE(i == HBP_NUM, "Can't find any breakpoint slot"))
162 		return;
163 
164 	dr7 = this_cpu_ptr(&cpu_dr7);
165 	*dr7 &= ~__encode_dr7(i, info->len, info->type);
166 
167 	set_debugreg(*dr7, 7);
168 	if (info->mask)
169 		set_dr_addr_mask(0, i);
170 }
171 
172 /*
173  * Check for virtual address in kernel space.
174  */
175 int arch_check_bp_in_kernelspace(struct perf_event *bp)
176 {
177 	unsigned int len;
178 	unsigned long va;
179 	struct arch_hw_breakpoint *info = counter_arch_bp(bp);
180 
181 	va = info->address;
182 	len = bp->attr.bp_len;
183 
184 	/*
185 	 * We don't need to worry about va + len - 1 overflowing:
186 	 * we already require that va is aligned to a multiple of len.
187 	 */
188 	return (va >= TASK_SIZE_MAX) || ((va + len - 1) >= TASK_SIZE_MAX);
189 }
190 
191 int arch_bp_generic_fields(int x86_len, int x86_type,
192 			   int *gen_len, int *gen_type)
193 {
194 	/* Type */
195 	switch (x86_type) {
196 	case X86_BREAKPOINT_EXECUTE:
197 		if (x86_len != X86_BREAKPOINT_LEN_X)
198 			return -EINVAL;
199 
200 		*gen_type = HW_BREAKPOINT_X;
201 		*gen_len = sizeof(long);
202 		return 0;
203 	case X86_BREAKPOINT_WRITE:
204 		*gen_type = HW_BREAKPOINT_W;
205 		break;
206 	case X86_BREAKPOINT_RW:
207 		*gen_type = HW_BREAKPOINT_W | HW_BREAKPOINT_R;
208 		break;
209 	default:
210 		return -EINVAL;
211 	}
212 
213 	/* Len */
214 	switch (x86_len) {
215 	case X86_BREAKPOINT_LEN_1:
216 		*gen_len = HW_BREAKPOINT_LEN_1;
217 		break;
218 	case X86_BREAKPOINT_LEN_2:
219 		*gen_len = HW_BREAKPOINT_LEN_2;
220 		break;
221 	case X86_BREAKPOINT_LEN_4:
222 		*gen_len = HW_BREAKPOINT_LEN_4;
223 		break;
224 #ifdef CONFIG_X86_64
225 	case X86_BREAKPOINT_LEN_8:
226 		*gen_len = HW_BREAKPOINT_LEN_8;
227 		break;
228 #endif
229 	default:
230 		return -EINVAL;
231 	}
232 
233 	return 0;
234 }
235 
236 
237 static int arch_build_bp_info(struct perf_event *bp)
238 {
239 	struct arch_hw_breakpoint *info = counter_arch_bp(bp);
240 
241 	info->address = bp->attr.bp_addr;
242 
243 	/* Type */
244 	switch (bp->attr.bp_type) {
245 	case HW_BREAKPOINT_W:
246 		info->type = X86_BREAKPOINT_WRITE;
247 		break;
248 	case HW_BREAKPOINT_W | HW_BREAKPOINT_R:
249 		info->type = X86_BREAKPOINT_RW;
250 		break;
251 	case HW_BREAKPOINT_X:
252 		/*
253 		 * We don't allow kernel breakpoints in places that are not
254 		 * acceptable for kprobes.  On non-kprobes kernels, we don't
255 		 * allow kernel breakpoints at all.
256 		 */
257 		if (bp->attr.bp_addr >= TASK_SIZE_MAX) {
258 #ifdef CONFIG_KPROBES
259 			if (within_kprobe_blacklist(bp->attr.bp_addr))
260 				return -EINVAL;
261 #else
262 			return -EINVAL;
263 #endif
264 		}
265 
266 		info->type = X86_BREAKPOINT_EXECUTE;
267 		/*
268 		 * x86 inst breakpoints need to have a specific undefined len.
269 		 * But we still need to check userspace is not trying to setup
270 		 * an unsupported length, to get a range breakpoint for example.
271 		 */
272 		if (bp->attr.bp_len == sizeof(long)) {
273 			info->len = X86_BREAKPOINT_LEN_X;
274 			return 0;
275 		}
276 	default:
277 		return -EINVAL;
278 	}
279 
280 	/* Len */
281 	info->mask = 0;
282 
283 	switch (bp->attr.bp_len) {
284 	case HW_BREAKPOINT_LEN_1:
285 		info->len = X86_BREAKPOINT_LEN_1;
286 		break;
287 	case HW_BREAKPOINT_LEN_2:
288 		info->len = X86_BREAKPOINT_LEN_2;
289 		break;
290 	case HW_BREAKPOINT_LEN_4:
291 		info->len = X86_BREAKPOINT_LEN_4;
292 		break;
293 #ifdef CONFIG_X86_64
294 	case HW_BREAKPOINT_LEN_8:
295 		info->len = X86_BREAKPOINT_LEN_8;
296 		break;
297 #endif
298 	default:
299 		/* AMD range breakpoint */
300 		if (!is_power_of_2(bp->attr.bp_len))
301 			return -EINVAL;
302 		if (bp->attr.bp_addr & (bp->attr.bp_len - 1))
303 			return -EINVAL;
304 
305 		if (!boot_cpu_has(X86_FEATURE_BPEXT))
306 			return -EOPNOTSUPP;
307 
308 		/*
309 		 * It's impossible to use a range breakpoint to fake out
310 		 * user vs kernel detection because bp_len - 1 can't
311 		 * have the high bit set.  If we ever allow range instruction
312 		 * breakpoints, then we'll have to check for kprobe-blacklisted
313 		 * addresses anywhere in the range.
314 		 */
315 		info->mask = bp->attr.bp_len - 1;
316 		info->len = X86_BREAKPOINT_LEN_1;
317 	}
318 
319 	return 0;
320 }
321 
322 /*
323  * Validate the arch-specific HW Breakpoint register settings
324  */
325 int arch_validate_hwbkpt_settings(struct perf_event *bp)
326 {
327 	struct arch_hw_breakpoint *info = counter_arch_bp(bp);
328 	unsigned int align;
329 	int ret;
330 
331 
332 	ret = arch_build_bp_info(bp);
333 	if (ret)
334 		return ret;
335 
336 	switch (info->len) {
337 	case X86_BREAKPOINT_LEN_1:
338 		align = 0;
339 		if (info->mask)
340 			align = info->mask;
341 		break;
342 	case X86_BREAKPOINT_LEN_2:
343 		align = 1;
344 		break;
345 	case X86_BREAKPOINT_LEN_4:
346 		align = 3;
347 		break;
348 #ifdef CONFIG_X86_64
349 	case X86_BREAKPOINT_LEN_8:
350 		align = 7;
351 		break;
352 #endif
353 	default:
354 		WARN_ON_ONCE(1);
355 	}
356 
357 	/*
358 	 * Check that the low-order bits of the address are appropriate
359 	 * for the alignment implied by len.
360 	 */
361 	if (info->address & align)
362 		return -EINVAL;
363 
364 	return 0;
365 }
366 
367 /*
368  * Dump the debug register contents to the user.
369  * We can't dump our per cpu values because it
370  * may contain cpu wide breakpoint, something that
371  * doesn't belong to the current task.
372  *
373  * TODO: include non-ptrace user breakpoints (perf)
374  */
375 void aout_dump_debugregs(struct user *dump)
376 {
377 	int i;
378 	int dr7 = 0;
379 	struct perf_event *bp;
380 	struct arch_hw_breakpoint *info;
381 	struct thread_struct *thread = &current->thread;
382 
383 	for (i = 0; i < HBP_NUM; i++) {
384 		bp = thread->ptrace_bps[i];
385 
386 		if (bp && !bp->attr.disabled) {
387 			dump->u_debugreg[i] = bp->attr.bp_addr;
388 			info = counter_arch_bp(bp);
389 			dr7 |= encode_dr7(i, info->len, info->type);
390 		} else {
391 			dump->u_debugreg[i] = 0;
392 		}
393 	}
394 
395 	dump->u_debugreg[4] = 0;
396 	dump->u_debugreg[5] = 0;
397 	dump->u_debugreg[6] = current->thread.debugreg6;
398 
399 	dump->u_debugreg[7] = dr7;
400 }
401 EXPORT_SYMBOL_GPL(aout_dump_debugregs);
402 
403 /*
404  * Release the user breakpoints used by ptrace
405  */
406 void flush_ptrace_hw_breakpoint(struct task_struct *tsk)
407 {
408 	int i;
409 	struct thread_struct *t = &tsk->thread;
410 
411 	for (i = 0; i < HBP_NUM; i++) {
412 		unregister_hw_breakpoint(t->ptrace_bps[i]);
413 		t->ptrace_bps[i] = NULL;
414 	}
415 
416 	t->debugreg6 = 0;
417 	t->ptrace_dr7 = 0;
418 }
419 
420 void hw_breakpoint_restore(void)
421 {
422 	set_debugreg(__this_cpu_read(cpu_debugreg[0]), 0);
423 	set_debugreg(__this_cpu_read(cpu_debugreg[1]), 1);
424 	set_debugreg(__this_cpu_read(cpu_debugreg[2]), 2);
425 	set_debugreg(__this_cpu_read(cpu_debugreg[3]), 3);
426 	set_debugreg(current->thread.debugreg6, 6);
427 	set_debugreg(__this_cpu_read(cpu_dr7), 7);
428 }
429 EXPORT_SYMBOL_GPL(hw_breakpoint_restore);
430 
431 /*
432  * Handle debug exception notifications.
433  *
434  * Return value is either NOTIFY_STOP or NOTIFY_DONE as explained below.
435  *
436  * NOTIFY_DONE returned if one of the following conditions is true.
437  * i) When the causative address is from user-space and the exception
438  * is a valid one, i.e. not triggered as a result of lazy debug register
439  * switching
440  * ii) When there are more bits than trap<n> set in DR6 register (such
441  * as BD, BS or BT) indicating that more than one debug condition is
442  * met and requires some more action in do_debug().
443  *
444  * NOTIFY_STOP returned for all other cases
445  *
446  */
447 static int hw_breakpoint_handler(struct die_args *args)
448 {
449 	int i, cpu, rc = NOTIFY_STOP;
450 	struct perf_event *bp;
451 	unsigned long dr7, dr6;
452 	unsigned long *dr6_p;
453 
454 	/* The DR6 value is pointed by args->err */
455 	dr6_p = (unsigned long *)ERR_PTR(args->err);
456 	dr6 = *dr6_p;
457 
458 	/* If it's a single step, TRAP bits are random */
459 	if (dr6 & DR_STEP)
460 		return NOTIFY_DONE;
461 
462 	/* Do an early return if no trap bits are set in DR6 */
463 	if ((dr6 & DR_TRAP_BITS) == 0)
464 		return NOTIFY_DONE;
465 
466 	get_debugreg(dr7, 7);
467 	/* Disable breakpoints during exception handling */
468 	set_debugreg(0UL, 7);
469 	/*
470 	 * Assert that local interrupts are disabled
471 	 * Reset the DRn bits in the virtualized register value.
472 	 * The ptrace trigger routine will add in whatever is needed.
473 	 */
474 	current->thread.debugreg6 &= ~DR_TRAP_BITS;
475 	cpu = get_cpu();
476 
477 	/* Handle all the breakpoints that were triggered */
478 	for (i = 0; i < HBP_NUM; ++i) {
479 		if (likely(!(dr6 & (DR_TRAP0 << i))))
480 			continue;
481 
482 		/*
483 		 * The counter may be concurrently released but that can only
484 		 * occur from a call_rcu() path. We can then safely fetch
485 		 * the breakpoint, use its callback, touch its counter
486 		 * while we are in an rcu_read_lock() path.
487 		 */
488 		rcu_read_lock();
489 
490 		bp = per_cpu(bp_per_reg[i], cpu);
491 		/*
492 		 * Reset the 'i'th TRAP bit in dr6 to denote completion of
493 		 * exception handling
494 		 */
495 		(*dr6_p) &= ~(DR_TRAP0 << i);
496 		/*
497 		 * bp can be NULL due to lazy debug register switching
498 		 * or due to concurrent perf counter removing.
499 		 */
500 		if (!bp) {
501 			rcu_read_unlock();
502 			break;
503 		}
504 
505 		perf_bp_event(bp, args->regs);
506 
507 		/*
508 		 * Set up resume flag to avoid breakpoint recursion when
509 		 * returning back to origin.
510 		 */
511 		if (bp->hw.info.type == X86_BREAKPOINT_EXECUTE)
512 			args->regs->flags |= X86_EFLAGS_RF;
513 
514 		rcu_read_unlock();
515 	}
516 	/*
517 	 * Further processing in do_debug() is needed for a) user-space
518 	 * breakpoints (to generate signals) and b) when the system has
519 	 * taken exception due to multiple causes
520 	 */
521 	if ((current->thread.debugreg6 & DR_TRAP_BITS) ||
522 	    (dr6 & (~DR_TRAP_BITS)))
523 		rc = NOTIFY_DONE;
524 
525 	set_debugreg(dr7, 7);
526 	put_cpu();
527 
528 	return rc;
529 }
530 
531 /*
532  * Handle debug exception notifications.
533  */
534 int hw_breakpoint_exceptions_notify(
535 		struct notifier_block *unused, unsigned long val, void *data)
536 {
537 	if (val != DIE_DEBUG)
538 		return NOTIFY_DONE;
539 
540 	return hw_breakpoint_handler(data);
541 }
542 
543 void hw_breakpoint_pmu_read(struct perf_event *bp)
544 {
545 	/* TODO */
546 }
547