Linux Audio

Check our new training course

Loading...
v6.13.7
  1// SPDX-License-Identifier: GPL-2.0-or-later
  2/*
 
 
 
 
 
 
 
 
 
 
 
 
 
  3 *
  4 * Copyright (C) 2007 Alan Stern
  5 * Copyright (C) 2009 IBM Corporation
  6 * Copyright (C) 2009 Frederic Weisbecker <fweisbec@gmail.com>
  7 *
  8 * Authors: Alan Stern <stern@rowland.harvard.edu>
  9 *          K.Prasad <prasad@linux.vnet.ibm.com>
 10 *          Frederic Weisbecker <fweisbec@gmail.com>
 11 */
 12
 13/*
 14 * HW_breakpoint: a unified kernel/user-space hardware breakpoint facility,
 15 * using the CPU's debug registers.
 16 */
 17
 18#include <linux/perf_event.h>
 19#include <linux/hw_breakpoint.h>
 20#include <linux/irqflags.h>
 21#include <linux/notifier.h>
 22#include <linux/kallsyms.h>
 23#include <linux/kprobes.h>
 24#include <linux/percpu.h>
 25#include <linux/kdebug.h>
 26#include <linux/kernel.h>
 27#include <linux/export.h>
 28#include <linux/sched.h>
 
 29#include <linux/smp.h>
 30
 31#include <asm/hw_breakpoint.h>
 32#include <asm/processor.h>
 33#include <asm/debugreg.h>
 34#include <asm/user.h>
 35#include <asm/desc.h>
 36#include <asm/tlbflush.h>
 37
 38/* Per cpu debug control register value */
 39DEFINE_PER_CPU(unsigned long, cpu_dr7);
 40EXPORT_PER_CPU_SYMBOL(cpu_dr7);
 41
 42/* Per cpu debug address registers values */
 43static DEFINE_PER_CPU(unsigned long, cpu_debugreg[HBP_NUM]);
 44
 45/*
 46 * Stores the breakpoints currently in use on each breakpoint address
 47 * register for each cpus
 48 */
 49static DEFINE_PER_CPU(struct perf_event *, bp_per_reg[HBP_NUM]);
 50
 51
 52static inline unsigned long
 53__encode_dr7(int drnum, unsigned int len, unsigned int type)
 54{
 55	unsigned long bp_info;
 56
 57	bp_info = (len | type) & 0xf;
 58	bp_info <<= (DR_CONTROL_SHIFT + drnum * DR_CONTROL_SIZE);
 59	bp_info |= (DR_GLOBAL_ENABLE << (drnum * DR_ENABLE_SIZE));
 60
 61	return bp_info;
 62}
 63
 64/*
 65 * Encode the length, type, Exact, and Enable bits for a particular breakpoint
 66 * as stored in debug register 7.
 67 */
 68unsigned long encode_dr7(int drnum, unsigned int len, unsigned int type)
 69{
 70	return __encode_dr7(drnum, len, type) | DR_GLOBAL_SLOWDOWN;
 71}
 72
 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)
 78{
 79	int bp_info = dr7 >> (DR_CONTROL_SHIFT + bpnum * DR_CONTROL_SIZE);
 80
 81	*len = (bp_info & 0xc) | 0x40;
 82	*type = (bp_info & 0x3) | 0x80;
 83
 84	return (dr7 >> (bpnum * DR_ENABLE_SIZE)) & 0x3;
 85}
 86
 87/*
 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.
 95 */
 96int arch_install_hw_breakpoint(struct perf_event *bp)
 97{
 98	struct arch_hw_breakpoint *info = counter_arch_bp(bp);
 99	unsigned long *dr7;
100	int i;
101
102	lockdep_assert_irqs_disabled();
103
104	for (i = 0; i < HBP_NUM; i++) {
105		struct perf_event **slot = this_cpu_ptr(&bp_per_reg[i]);
106
107		if (!*slot) {
108			*slot = bp;
109			break;
110		}
111	}
112
113	if (WARN_ONCE(i == HBP_NUM, "Can't find any breakpoint slot"))
114		return -EBUSY;
115
116	set_debugreg(info->address, i);
117	__this_cpu_write(cpu_debugreg[i], info->address);
118
119	dr7 = this_cpu_ptr(&cpu_dr7);
120	*dr7 |= encode_dr7(i, info->len, info->type);
121
122	/*
123	 * Ensure we first write cpu_dr7 before we set the DR7 register.
124	 * This ensures an NMI never see cpu_dr7 0 when DR7 is not.
125	 */
126	barrier();
127
128	set_debugreg(*dr7, 7);
129	if (info->mask)
130		amd_set_dr_addr_mask(info->mask, i);
131
132	return 0;
133}
134
135/*
136 * Uninstall the breakpoint contained in the given counter.
137 *
138 * First we search the debug address register it uses and then we disable
139 * it.
140 *
141 * Atomic: we hold the counter->ctx->lock and we only handle variables
142 * and registers local to this cpu.
143 */
144void arch_uninstall_hw_breakpoint(struct perf_event *bp)
145{
146	struct arch_hw_breakpoint *info = counter_arch_bp(bp);
147	unsigned long dr7;
148	int i;
149
150	lockdep_assert_irqs_disabled();
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_read(cpu_dr7);
165	dr7 &= ~__encode_dr7(i, info->len, info->type);
166
167	set_debugreg(dr7, 7);
168	if (info->mask)
169		amd_set_dr_addr_mask(0, i);
170
171	/*
172	 * Ensure the write to cpu_dr7 is after we've set the DR7 register.
173	 * This ensures an NMI never see cpu_dr7 0 when DR7 is not.
174	 */
175	barrier();
176
177	this_cpu_write(cpu_dr7, dr7);
178}
179
180static int arch_bp_generic_len(int x86_len)
181{
182	switch (x86_len) {
 
 
183	case X86_BREAKPOINT_LEN_1:
184		return HW_BREAKPOINT_LEN_1;
 
185	case X86_BREAKPOINT_LEN_2:
186		return HW_BREAKPOINT_LEN_2;
 
187	case X86_BREAKPOINT_LEN_4:
188		return HW_BREAKPOINT_LEN_4;
 
189#ifdef CONFIG_X86_64
190	case X86_BREAKPOINT_LEN_8:
191		return HW_BREAKPOINT_LEN_8;
 
192#endif
193	default:
194		return -EINVAL;
195	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
196}
197
198int arch_bp_generic_fields(int x86_len, int x86_type,
199			   int *gen_len, int *gen_type)
200{
201	int len;
202
203	/* Type */
204	switch (x86_type) {
205	case X86_BREAKPOINT_EXECUTE:
206		if (x86_len != X86_BREAKPOINT_LEN_X)
207			return -EINVAL;
208
209		*gen_type = HW_BREAKPOINT_X;
210		*gen_len = sizeof(long);
211		return 0;
212	case X86_BREAKPOINT_WRITE:
213		*gen_type = HW_BREAKPOINT_W;
214		break;
215	case X86_BREAKPOINT_RW:
216		*gen_type = HW_BREAKPOINT_W | HW_BREAKPOINT_R;
217		break;
218	default:
219		return -EINVAL;
220	}
221
222	/* Len */
223	len = arch_bp_generic_len(x86_len);
224	if (len < 0)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
225		return -EINVAL;
226	*gen_len = len;
227
228	return 0;
229}
230
231/*
232 * Check for virtual address in kernel space.
233 */
234int arch_check_bp_in_kernelspace(struct arch_hw_breakpoint *hw)
235{
236	unsigned long va;
237	int len;
238
239	va = hw->address;
240	len = arch_bp_generic_len(hw->len);
241	WARN_ON_ONCE(len < 0);
242
243	/*
244	 * We don't need to worry about va + len - 1 overflowing:
245	 * we already require that va is aligned to a multiple of len.
246	 */
247	return (va >= TASK_SIZE_MAX) || ((va + len - 1) >= TASK_SIZE_MAX);
248}
249
250/*
251 * Checks whether the range [addr, end], overlaps the area [base, base + size).
252 */
253static inline bool within_area(unsigned long addr, unsigned long end,
254			       unsigned long base, unsigned long size)
255{
256	return end >= base && addr < (base + size);
257}
258
259/*
260 * Checks whether the range from addr to end, inclusive, overlaps the fixed
261 * mapped CPU entry area range or other ranges used for CPU entry.
262 */
263static inline bool within_cpu_entry(unsigned long addr, unsigned long end)
264{
265	int cpu;
266
267	/* CPU entry erea is always used for CPU entry */
268	if (within_area(addr, end, CPU_ENTRY_AREA_BASE,
269			CPU_ENTRY_AREA_MAP_SIZE))
270		return true;
271
272	/*
273	 * When FSGSBASE is enabled, paranoid_entry() fetches the per-CPU
274	 * GSBASE value via __per_cpu_offset or pcpu_unit_offsets.
275	 */
276#ifdef CONFIG_SMP
277	if (within_area(addr, end, (unsigned long)__per_cpu_offset,
278			sizeof(unsigned long) * nr_cpu_ids))
279		return true;
280#else
281	if (within_area(addr, end, (unsigned long)&pcpu_unit_offsets,
282			sizeof(pcpu_unit_offsets)))
283		return true;
284#endif
285
286	for_each_possible_cpu(cpu) {
287		/* The original rw GDT is being used after load_direct_gdt() */
288		if (within_area(addr, end, (unsigned long)get_cpu_gdt_rw(cpu),
289				GDT_SIZE))
290			return true;
291
292		/*
293		 * cpu_tss_rw is not directly referenced by hardware, but
294		 * cpu_tss_rw is also used in CPU entry code,
295		 */
296		if (within_area(addr, end,
297				(unsigned long)&per_cpu(cpu_tss_rw, cpu),
298				sizeof(struct tss_struct)))
299			return true;
300
301		/*
302		 * cpu_tlbstate.user_pcid_flush_mask is used for CPU entry.
303		 * If a data breakpoint on it, it will cause an unwanted #DB.
304		 * Protect the full cpu_tlbstate structure to be sure.
305		 */
306		if (within_area(addr, end,
307				(unsigned long)&per_cpu(cpu_tlbstate, cpu),
308				sizeof(struct tlb_state)))
309			return true;
310
311		/*
312		 * When in guest (X86_FEATURE_HYPERVISOR), local_db_save()
313		 * will read per-cpu cpu_dr7 before clear dr7 register.
314		 */
315		if (within_area(addr, end, (unsigned long)&per_cpu(cpu_dr7, cpu),
316				sizeof(cpu_dr7)))
317			return true;
318	}
319
320	return false;
321}
322
323static int arch_build_bp_info(struct perf_event *bp,
324			      const struct perf_event_attr *attr,
325			      struct arch_hw_breakpoint *hw)
326{
327	unsigned long bp_end;
328
329	bp_end = attr->bp_addr + attr->bp_len - 1;
330	if (bp_end < attr->bp_addr)
331		return -EINVAL;
332
333	/*
334	 * Prevent any breakpoint of any type that overlaps the CPU
335	 * entry area and data.  This protects the IST stacks and also
336	 * reduces the chance that we ever find out what happens if
337	 * there's a data breakpoint on the GDT, IDT, or TSS.
338	 */
339	if (within_cpu_entry(attr->bp_addr, bp_end))
340		return -EINVAL;
341
342	hw->address = attr->bp_addr;
343	hw->mask = 0;
344
345	/* Type */
346	switch (attr->bp_type) {
347	case HW_BREAKPOINT_W:
348		hw->type = X86_BREAKPOINT_WRITE;
349		break;
350	case HW_BREAKPOINT_W | HW_BREAKPOINT_R:
351		hw->type = X86_BREAKPOINT_RW;
352		break;
353	case HW_BREAKPOINT_X:
354		/*
355		 * We don't allow kernel breakpoints in places that are not
356		 * acceptable for kprobes.  On non-kprobes kernels, we don't
357		 * allow kernel breakpoints at all.
358		 */
359		if (attr->bp_addr >= TASK_SIZE_MAX) {
360			if (within_kprobe_blacklist(attr->bp_addr))
361				return -EINVAL;
362		}
363
364		hw->type = X86_BREAKPOINT_EXECUTE;
365		/*
366		 * x86 inst breakpoints need to have a specific undefined len.
367		 * But we still need to check userspace is not trying to setup
368		 * an unsupported length, to get a range breakpoint for example.
369		 */
370		if (attr->bp_len == sizeof(long)) {
371			hw->len = X86_BREAKPOINT_LEN_X;
372			return 0;
373		}
374		fallthrough;
375	default:
376		return -EINVAL;
377	}
378
379	/* Len */
380	switch (attr->bp_len) {
381	case HW_BREAKPOINT_LEN_1:
382		hw->len = X86_BREAKPOINT_LEN_1;
383		break;
384	case HW_BREAKPOINT_LEN_2:
385		hw->len = X86_BREAKPOINT_LEN_2;
386		break;
387	case HW_BREAKPOINT_LEN_4:
388		hw->len = X86_BREAKPOINT_LEN_4;
389		break;
390#ifdef CONFIG_X86_64
391	case HW_BREAKPOINT_LEN_8:
392		hw->len = X86_BREAKPOINT_LEN_8;
393		break;
394#endif
395	default:
396		/* AMD range breakpoint */
397		if (!is_power_of_2(attr->bp_len))
398			return -EINVAL;
399		if (attr->bp_addr & (attr->bp_len - 1))
400			return -EINVAL;
401
402		if (!boot_cpu_has(X86_FEATURE_BPEXT))
403			return -EOPNOTSUPP;
404
405		/*
406		 * It's impossible to use a range breakpoint to fake out
407		 * user vs kernel detection because bp_len - 1 can't
408		 * have the high bit set.  If we ever allow range instruction
409		 * breakpoints, then we'll have to check for kprobe-blacklisted
410		 * addresses anywhere in the range.
411		 */
412		hw->mask = attr->bp_len - 1;
413		hw->len = X86_BREAKPOINT_LEN_1;
414	}
415
416	return 0;
417}
418
419/*
420 * Validate the arch-specific HW Breakpoint register settings
421 */
422int hw_breakpoint_arch_parse(struct perf_event *bp,
423			     const struct perf_event_attr *attr,
424			     struct arch_hw_breakpoint *hw)
425{
 
426	unsigned int align;
427	int ret;
428
429
430	ret = arch_build_bp_info(bp, attr, hw);
431	if (ret)
432		return ret;
433
434	switch (hw->len) {
 
 
435	case X86_BREAKPOINT_LEN_1:
436		align = 0;
437		if (hw->mask)
438			align = hw->mask;
439		break;
440	case X86_BREAKPOINT_LEN_2:
441		align = 1;
442		break;
443	case X86_BREAKPOINT_LEN_4:
444		align = 3;
445		break;
446#ifdef CONFIG_X86_64
447	case X86_BREAKPOINT_LEN_8:
448		align = 7;
449		break;
450#endif
451	default:
452		WARN_ON_ONCE(1);
453		return -EINVAL;
454	}
455
456	/*
457	 * Check that the low-order bits of the address are appropriate
458	 * for the alignment implied by len.
459	 */
460	if (hw->address & align)
461		return -EINVAL;
462
463	return 0;
464}
465
466/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
467 * Release the user breakpoints used by ptrace
468 */
469void flush_ptrace_hw_breakpoint(struct task_struct *tsk)
470{
471	int i;
472	struct thread_struct *t = &tsk->thread;
473
474	for (i = 0; i < HBP_NUM; i++) {
475		unregister_hw_breakpoint(t->ptrace_bps[i]);
476		t->ptrace_bps[i] = NULL;
477	}
478
479	t->virtual_dr6 = 0;
480	t->ptrace_dr7 = 0;
481}
482
483void hw_breakpoint_restore(void)
484{
485	set_debugreg(__this_cpu_read(cpu_debugreg[0]), 0);
486	set_debugreg(__this_cpu_read(cpu_debugreg[1]), 1);
487	set_debugreg(__this_cpu_read(cpu_debugreg[2]), 2);
488	set_debugreg(__this_cpu_read(cpu_debugreg[3]), 3);
489	set_debugreg(DR6_RESERVED, 6);
490	set_debugreg(__this_cpu_read(cpu_dr7), 7);
491}
492EXPORT_SYMBOL_GPL(hw_breakpoint_restore);
493
494/*
495 * Handle debug exception notifications.
496 *
497 * Return value is either NOTIFY_STOP or NOTIFY_DONE as explained below.
498 *
499 * NOTIFY_DONE returned if one of the following conditions is true.
500 * i) When the causative address is from user-space and the exception
501 * is a valid one, i.e. not triggered as a result of lazy debug register
502 * switching
503 * ii) When there are more bits than trap<n> set in DR6 register (such
504 * as BD, BS or BT) indicating that more than one debug condition is
505 * met and requires some more action in do_debug().
506 *
507 * NOTIFY_STOP returned for all other cases
508 *
509 */
510static int hw_breakpoint_handler(struct die_args *args)
511{
512	int i, rc = NOTIFY_STOP;
513	struct perf_event *bp;
 
514	unsigned long *dr6_p;
515	unsigned long dr6;
516	bool bpx;
517
518	/* The DR6 value is pointed by args->err */
519	dr6_p = (unsigned long *)ERR_PTR(args->err);
520	dr6 = *dr6_p;
521
 
 
 
 
522	/* Do an early return if no trap bits are set in DR6 */
523	if ((dr6 & DR_TRAP_BITS) == 0)
524		return NOTIFY_DONE;
525
 
 
 
 
 
 
 
 
 
 
 
526	/* Handle all the breakpoints that were triggered */
527	for (i = 0; i < HBP_NUM; ++i) {
528		if (likely(!(dr6 & (DR_TRAP0 << i))))
529			continue;
530
531		bp = this_cpu_read(bp_per_reg[i]);
532		if (!bp)
533			continue;
534
535		bpx = bp->hw.info.type == X86_BREAKPOINT_EXECUTE;
536
537		/*
538		 * TF and data breakpoints are traps and can be merged, however
539		 * instruction breakpoints are faults and will be raised
540		 * separately.
541		 *
542		 * However DR6 can indicate both TF and instruction
543		 * breakpoints. In that case take TF as that has precedence and
544		 * delay the instruction breakpoint for the next exception.
545		 */
546		if (bpx && (dr6 & DR_STEP))
547			continue;
548
 
549		/*
550		 * Reset the 'i'th TRAP bit in dr6 to denote completion of
551		 * exception handling
552		 */
553		(*dr6_p) &= ~(DR_TRAP0 << i);
 
 
 
 
 
 
 
 
554
555		perf_bp_event(bp, args->regs);
556
557		/*
558		 * Set up resume flag to avoid breakpoint recursion when
559		 * returning back to origin.
560		 */
561		if (bpx)
562			args->regs->flags |= X86_EFLAGS_RF;
563	}
564
 
 
565	/*
566	 * Further processing in do_debug() is needed for a) user-space
567	 * breakpoints (to generate signals) and b) when the system has
568	 * taken exception due to multiple causes
569	 */
570	if ((current->thread.virtual_dr6 & DR_TRAP_BITS) ||
571	    (dr6 & (~DR_TRAP_BITS)))
572		rc = NOTIFY_DONE;
573
 
 
 
574	return rc;
575}
576
577/*
578 * Handle debug exception notifications.
579 */
580int hw_breakpoint_exceptions_notify(
581		struct notifier_block *unused, unsigned long val, void *data)
582{
583	if (val != DIE_DEBUG)
584		return NOTIFY_DONE;
585
586	return hw_breakpoint_handler(data);
587}
588
589void hw_breakpoint_pmu_read(struct perf_event *bp)
590{
591	/* TODO */
592}
v3.1
 
  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/init.h>
 42#include <linux/smp.h>
 43
 44#include <asm/hw_breakpoint.h>
 45#include <asm/processor.h>
 46#include <asm/debugreg.h>
 
 
 
 47
 48/* Per cpu debug control register value */
 49DEFINE_PER_CPU(unsigned long, cpu_dr7);
 50EXPORT_PER_CPU_SYMBOL(cpu_dr7);
 51
 52/* Per cpu debug address registers values */
 53static 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 */
 59static DEFINE_PER_CPU(struct perf_event *, bp_per_reg[HBP_NUM]);
 60
 61
 62static 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 */
 78unsigned 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 */
 87int 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 */
106int 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 = &__get_cpu_var(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 = &__get_cpu_var(cpu_dr7);
128	*dr7 |= encode_dr7(i, info->len, info->type);
129
 
 
 
 
 
 
130	set_debugreg(*dr7, 7);
 
 
131
132	return 0;
133}
134
135/*
136 * Uninstall the breakpoint contained in the given counter.
137 *
138 * First we search the debug address register it uses and then we disable
139 * it.
140 *
141 * Atomic: we hold the counter->ctx->lock and we only handle variables
142 * and registers local to this cpu.
143 */
144void arch_uninstall_hw_breakpoint(struct perf_event *bp)
145{
146	struct arch_hw_breakpoint *info = counter_arch_bp(bp);
147	unsigned long *dr7;
148	int i;
149
 
 
150	for (i = 0; i < HBP_NUM; i++) {
151		struct perf_event **slot = &__get_cpu_var(bp_per_reg[i]);
152
153		if (*slot == bp) {
154			*slot = NULL;
155			break;
156		}
157	}
158
159	if (WARN_ONCE(i == HBP_NUM, "Can't find any breakpoint slot"))
160		return;
161
162	dr7 = &__get_cpu_var(cpu_dr7);
163	*dr7 &= ~__encode_dr7(i, info->len, info->type);
 
 
 
 
164
165	set_debugreg(*dr7, 7);
 
 
 
 
 
 
166}
167
168static int get_hbp_len(u8 hbp_len)
169{
170	unsigned int len_in_bytes = 0;
171
172	switch (hbp_len) {
173	case X86_BREAKPOINT_LEN_1:
174		len_in_bytes = 1;
175		break;
176	case X86_BREAKPOINT_LEN_2:
177		len_in_bytes = 2;
178		break;
179	case X86_BREAKPOINT_LEN_4:
180		len_in_bytes = 4;
181		break;
182#ifdef CONFIG_X86_64
183	case X86_BREAKPOINT_LEN_8:
184		len_in_bytes = 8;
185		break;
186#endif
 
 
187	}
188	return len_in_bytes;
189}
190
191/*
192 * Check for virtual address in kernel space.
193 */
194int arch_check_bp_in_kernelspace(struct perf_event *bp)
195{
196	unsigned int len;
197	unsigned long va;
198	struct arch_hw_breakpoint *info = counter_arch_bp(bp);
199
200	va = info->address;
201	len = get_hbp_len(info->len);
202
203	return (va >= TASK_SIZE) && ((va + len - 1) >= TASK_SIZE);
204}
205
206int arch_bp_generic_fields(int x86_len, int x86_type,
207			   int *gen_len, int *gen_type)
208{
 
 
209	/* Type */
210	switch (x86_type) {
211	case X86_BREAKPOINT_EXECUTE:
212		if (x86_len != X86_BREAKPOINT_LEN_X)
213			return -EINVAL;
214
215		*gen_type = HW_BREAKPOINT_X;
216		*gen_len = sizeof(long);
217		return 0;
218	case X86_BREAKPOINT_WRITE:
219		*gen_type = HW_BREAKPOINT_W;
220		break;
221	case X86_BREAKPOINT_RW:
222		*gen_type = HW_BREAKPOINT_W | HW_BREAKPOINT_R;
223		break;
224	default:
225		return -EINVAL;
226	}
227
228	/* Len */
229	switch (x86_len) {
230	case X86_BREAKPOINT_LEN_1:
231		*gen_len = HW_BREAKPOINT_LEN_1;
232		break;
233	case X86_BREAKPOINT_LEN_2:
234		*gen_len = HW_BREAKPOINT_LEN_2;
235		break;
236	case X86_BREAKPOINT_LEN_4:
237		*gen_len = HW_BREAKPOINT_LEN_4;
238		break;
239#ifdef CONFIG_X86_64
240	case X86_BREAKPOINT_LEN_8:
241		*gen_len = HW_BREAKPOINT_LEN_8;
242		break;
243#endif
244	default:
245		return -EINVAL;
246	}
247
248	return 0;
249}
250
 
 
 
 
 
 
 
251
252static int arch_build_bp_info(struct perf_event *bp)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
253{
254	struct arch_hw_breakpoint *info = counter_arch_bp(bp);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
255
256	info->address = bp->attr.bp_addr;
 
257
258	/* Type */
259	switch (bp->attr.bp_type) {
260	case HW_BREAKPOINT_W:
261		info->type = X86_BREAKPOINT_WRITE;
262		break;
263	case HW_BREAKPOINT_W | HW_BREAKPOINT_R:
264		info->type = X86_BREAKPOINT_RW;
265		break;
266	case HW_BREAKPOINT_X:
267		info->type = X86_BREAKPOINT_EXECUTE;
 
 
 
 
 
 
 
 
 
 
268		/*
269		 * x86 inst breakpoints need to have a specific undefined len.
270		 * But we still need to check userspace is not trying to setup
271		 * an unsupported length, to get a range breakpoint for example.
272		 */
273		if (bp->attr.bp_len == sizeof(long)) {
274			info->len = X86_BREAKPOINT_LEN_X;
275			return 0;
276		}
 
277	default:
278		return -EINVAL;
279	}
280
281	/* Len */
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		return -EINVAL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
299	}
300
301	return 0;
302}
 
303/*
304 * Validate the arch-specific HW Breakpoint register settings
305 */
306int arch_validate_hwbkpt_settings(struct perf_event *bp)
 
 
307{
308	struct arch_hw_breakpoint *info = counter_arch_bp(bp);
309	unsigned int align;
310	int ret;
311
312
313	ret = arch_build_bp_info(bp);
314	if (ret)
315		return ret;
316
317	ret = -EINVAL;
318
319	switch (info->len) {
320	case X86_BREAKPOINT_LEN_1:
321		align = 0;
 
 
322		break;
323	case X86_BREAKPOINT_LEN_2:
324		align = 1;
325		break;
326	case X86_BREAKPOINT_LEN_4:
327		align = 3;
328		break;
329#ifdef CONFIG_X86_64
330	case X86_BREAKPOINT_LEN_8:
331		align = 7;
332		break;
333#endif
334	default:
335		return ret;
 
336	}
337
338	/*
339	 * Check that the low-order bits of the address are appropriate
340	 * for the alignment implied by len.
341	 */
342	if (info->address & align)
343		return -EINVAL;
344
345	return 0;
346}
347
348/*
349 * Dump the debug register contents to the user.
350 * We can't dump our per cpu values because it
351 * may contain cpu wide breakpoint, something that
352 * doesn't belong to the current task.
353 *
354 * TODO: include non-ptrace user breakpoints (perf)
355 */
356void aout_dump_debugregs(struct user *dump)
357{
358	int i;
359	int dr7 = 0;
360	struct perf_event *bp;
361	struct arch_hw_breakpoint *info;
362	struct thread_struct *thread = &current->thread;
363
364	for (i = 0; i < HBP_NUM; i++) {
365		bp = thread->ptrace_bps[i];
366
367		if (bp && !bp->attr.disabled) {
368			dump->u_debugreg[i] = bp->attr.bp_addr;
369			info = counter_arch_bp(bp);
370			dr7 |= encode_dr7(i, info->len, info->type);
371		} else {
372			dump->u_debugreg[i] = 0;
373		}
374	}
375
376	dump->u_debugreg[4] = 0;
377	dump->u_debugreg[5] = 0;
378	dump->u_debugreg[6] = current->thread.debugreg6;
379
380	dump->u_debugreg[7] = dr7;
381}
382EXPORT_SYMBOL_GPL(aout_dump_debugregs);
383
384/*
385 * Release the user breakpoints used by ptrace
386 */
387void flush_ptrace_hw_breakpoint(struct task_struct *tsk)
388{
389	int i;
390	struct thread_struct *t = &tsk->thread;
391
392	for (i = 0; i < HBP_NUM; i++) {
393		unregister_hw_breakpoint(t->ptrace_bps[i]);
394		t->ptrace_bps[i] = NULL;
395	}
 
 
 
396}
397
398void hw_breakpoint_restore(void)
399{
400	set_debugreg(__this_cpu_read(cpu_debugreg[0]), 0);
401	set_debugreg(__this_cpu_read(cpu_debugreg[1]), 1);
402	set_debugreg(__this_cpu_read(cpu_debugreg[2]), 2);
403	set_debugreg(__this_cpu_read(cpu_debugreg[3]), 3);
404	set_debugreg(current->thread.debugreg6, 6);
405	set_debugreg(__this_cpu_read(cpu_dr7), 7);
406}
407EXPORT_SYMBOL_GPL(hw_breakpoint_restore);
408
409/*
410 * Handle debug exception notifications.
411 *
412 * Return value is either NOTIFY_STOP or NOTIFY_DONE as explained below.
413 *
414 * NOTIFY_DONE returned if one of the following conditions is true.
415 * i) When the causative address is from user-space and the exception
416 * is a valid one, i.e. not triggered as a result of lazy debug register
417 * switching
418 * ii) When there are more bits than trap<n> set in DR6 register (such
419 * as BD, BS or BT) indicating that more than one debug condition is
420 * met and requires some more action in do_debug().
421 *
422 * NOTIFY_STOP returned for all other cases
423 *
424 */
425static int __kprobes hw_breakpoint_handler(struct die_args *args)
426{
427	int i, cpu, rc = NOTIFY_STOP;
428	struct perf_event *bp;
429	unsigned long dr7, dr6;
430	unsigned long *dr6_p;
 
 
431
432	/* The DR6 value is pointed by args->err */
433	dr6_p = (unsigned long *)ERR_PTR(args->err);
434	dr6 = *dr6_p;
435
436	/* If it's a single step, TRAP bits are random */
437	if (dr6 & DR_STEP)
438		return NOTIFY_DONE;
439
440	/* Do an early return if no trap bits are set in DR6 */
441	if ((dr6 & DR_TRAP_BITS) == 0)
442		return NOTIFY_DONE;
443
444	get_debugreg(dr7, 7);
445	/* Disable breakpoints during exception handling */
446	set_debugreg(0UL, 7);
447	/*
448	 * Assert that local interrupts are disabled
449	 * Reset the DRn bits in the virtualized register value.
450	 * The ptrace trigger routine will add in whatever is needed.
451	 */
452	current->thread.debugreg6 &= ~DR_TRAP_BITS;
453	cpu = get_cpu();
454
455	/* Handle all the breakpoints that were triggered */
456	for (i = 0; i < HBP_NUM; ++i) {
457		if (likely(!(dr6 & (DR_TRAP0 << i))))
458			continue;
459
 
 
 
 
 
 
460		/*
461		 * The counter may be concurrently released but that can only
462		 * occur from a call_rcu() path. We can then safely fetch
463		 * the breakpoint, use its callback, touch its counter
464		 * while we are in an rcu_read_lock() path.
 
 
 
465		 */
466		rcu_read_lock();
 
467
468		bp = per_cpu(bp_per_reg[i], cpu);
469		/*
470		 * Reset the 'i'th TRAP bit in dr6 to denote completion of
471		 * exception handling
472		 */
473		(*dr6_p) &= ~(DR_TRAP0 << i);
474		/*
475		 * bp can be NULL due to lazy debug register switching
476		 * or due to concurrent perf counter removing.
477		 */
478		if (!bp) {
479			rcu_read_unlock();
480			break;
481		}
482
483		perf_bp_event(bp, args->regs);
484
485		/*
486		 * Set up resume flag to avoid breakpoint recursion when
487		 * returning back to origin.
488		 */
489		if (bp->hw.info.type == X86_BREAKPOINT_EXECUTE)
490			args->regs->flags |= X86_EFLAGS_RF;
 
491
492		rcu_read_unlock();
493	}
494	/*
495	 * Further processing in do_debug() is needed for a) user-space
496	 * breakpoints (to generate signals) and b) when the system has
497	 * taken exception due to multiple causes
498	 */
499	if ((current->thread.debugreg6 & DR_TRAP_BITS) ||
500	    (dr6 & (~DR_TRAP_BITS)))
501		rc = NOTIFY_DONE;
502
503	set_debugreg(dr7, 7);
504	put_cpu();
505
506	return rc;
507}
508
509/*
510 * Handle debug exception notifications.
511 */
512int __kprobes hw_breakpoint_exceptions_notify(
513		struct notifier_block *unused, unsigned long val, void *data)
514{
515	if (val != DIE_DEBUG)
516		return NOTIFY_DONE;
517
518	return hw_breakpoint_handler(data);
519}
520
521void hw_breakpoint_pmu_read(struct perf_event *bp)
522{
523	/* TODO */
524}