xref: /linux/arch/x86/kernel/kprobes/core.c (revision b6ebbac51bedf9e98e837688bc838f400196da5e)
1 /*
2  *  Kernel Probes (KProbes)
3  *
4  * This program is free software; you can redistribute it and/or modify
5  * it under the terms of the GNU General Public License as published by
6  * the Free Software Foundation; either version 2 of the License, or
7  * (at your option) any later version.
8  *
9  * This program is distributed in the hope that it will be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write to the Free Software
16  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
17  *
18  * Copyright (C) IBM Corporation, 2002, 2004
19  *
20  * 2002-Oct	Created by Vamsi Krishna S <vamsi_krishna@in.ibm.com> Kernel
21  *		Probes initial implementation ( includes contributions from
22  *		Rusty Russell).
23  * 2004-July	Suparna Bhattacharya <suparna@in.ibm.com> added jumper probes
24  *		interface to access function arguments.
25  * 2004-Oct	Jim Keniston <jkenisto@us.ibm.com> and Prasanna S Panchamukhi
26  *		<prasanna@in.ibm.com> adapted for x86_64 from i386.
27  * 2005-Mar	Roland McGrath <roland@redhat.com>
28  *		Fixed to handle %rip-relative addressing mode correctly.
29  * 2005-May	Hien Nguyen <hien@us.ibm.com>, Jim Keniston
30  *		<jkenisto@us.ibm.com> and Prasanna S Panchamukhi
31  *		<prasanna@in.ibm.com> added function-return probes.
32  * 2005-May	Rusty Lynch <rusty.lynch@intel.com>
33  *		Added function return probes functionality
34  * 2006-Feb	Masami Hiramatsu <hiramatu@sdl.hitachi.co.jp> added
35  *		kprobe-booster and kretprobe-booster for i386.
36  * 2007-Dec	Masami Hiramatsu <mhiramat@redhat.com> added kprobe-booster
37  *		and kretprobe-booster for x86-64
38  * 2007-Dec	Masami Hiramatsu <mhiramat@redhat.com>, Arjan van de Ven
39  *		<arjan@infradead.org> and Jim Keniston <jkenisto@us.ibm.com>
40  *		unified x86 kprobes code.
41  */
42 #include <linux/kprobes.h>
43 #include <linux/ptrace.h>
44 #include <linux/string.h>
45 #include <linux/slab.h>
46 #include <linux/hardirq.h>
47 #include <linux/preempt.h>
48 #include <linux/module.h>
49 #include <linux/kdebug.h>
50 #include <linux/kallsyms.h>
51 #include <linux/ftrace.h>
52 #include <linux/frame.h>
53 
54 #include <asm/text-patching.h>
55 #include <asm/cacheflush.h>
56 #include <asm/desc.h>
57 #include <asm/pgtable.h>
58 #include <asm/uaccess.h>
59 #include <asm/alternative.h>
60 #include <asm/insn.h>
61 #include <asm/debugreg.h>
62 
63 #include "common.h"
64 
65 void jprobe_return_end(void);
66 
67 DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
68 DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
69 
70 #define stack_addr(regs) ((unsigned long *)kernel_stack_pointer(regs))
71 
72 #define W(row, b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, ba, bb, bc, bd, be, bf)\
73 	(((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) |   \
74 	  (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) |   \
75 	  (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) |   \
76 	  (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf))    \
77 	 << (row % 32))
78 	/*
79 	 * Undefined/reserved opcodes, conditional jump, Opcode Extension
80 	 * Groups, and some special opcodes can not boost.
81 	 * This is non-const and volatile to keep gcc from statically
82 	 * optimizing it out, as variable_test_bit makes gcc think only
83 	 * *(unsigned long*) is used.
84 	 */
85 static volatile u32 twobyte_is_boostable[256 / 32] = {
86 	/*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f          */
87 	/*      ----------------------------------------------          */
88 	W(0x00, 0, 0, 1, 1, 0, 0, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0) | /* 00 */
89 	W(0x10, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1) , /* 10 */
90 	W(0x20, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 20 */
91 	W(0x30, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 30 */
92 	W(0x40, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 40 */
93 	W(0x50, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 50 */
94 	W(0x60, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1) | /* 60 */
95 	W(0x70, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1) , /* 70 */
96 	W(0x80, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 80 */
97 	W(0x90, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 90 */
98 	W(0xa0, 1, 1, 0, 1, 1, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* a0 */
99 	W(0xb0, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1) , /* b0 */
100 	W(0xc0, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1) | /* c0 */
101 	W(0xd0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) , /* d0 */
102 	W(0xe0, 0, 1, 1, 0, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* e0 */
103 	W(0xf0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 1, 0, 1, 1, 1, 0)   /* f0 */
104 	/*      -----------------------------------------------         */
105 	/*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f          */
106 };
107 #undef W
108 
109 struct kretprobe_blackpoint kretprobe_blacklist[] = {
110 	{"__switch_to", }, /* This function switches only current task, but
111 			      doesn't switch kernel stack.*/
112 	{NULL, NULL}	/* Terminator */
113 };
114 
115 const int kretprobe_blacklist_size = ARRAY_SIZE(kretprobe_blacklist);
116 
117 static nokprobe_inline void
118 __synthesize_relative_insn(void *from, void *to, u8 op)
119 {
120 	struct __arch_relative_insn {
121 		u8 op;
122 		s32 raddr;
123 	} __packed *insn;
124 
125 	insn = (struct __arch_relative_insn *)from;
126 	insn->raddr = (s32)((long)(to) - ((long)(from) + 5));
127 	insn->op = op;
128 }
129 
130 /* Insert a jump instruction at address 'from', which jumps to address 'to'.*/
131 void synthesize_reljump(void *from, void *to)
132 {
133 	__synthesize_relative_insn(from, to, RELATIVEJUMP_OPCODE);
134 }
135 NOKPROBE_SYMBOL(synthesize_reljump);
136 
137 /* Insert a call instruction at address 'from', which calls address 'to'.*/
138 void synthesize_relcall(void *from, void *to)
139 {
140 	__synthesize_relative_insn(from, to, RELATIVECALL_OPCODE);
141 }
142 NOKPROBE_SYMBOL(synthesize_relcall);
143 
144 /*
145  * Skip the prefixes of the instruction.
146  */
147 static kprobe_opcode_t *skip_prefixes(kprobe_opcode_t *insn)
148 {
149 	insn_attr_t attr;
150 
151 	attr = inat_get_opcode_attribute((insn_byte_t)*insn);
152 	while (inat_is_legacy_prefix(attr)) {
153 		insn++;
154 		attr = inat_get_opcode_attribute((insn_byte_t)*insn);
155 	}
156 #ifdef CONFIG_X86_64
157 	if (inat_is_rex_prefix(attr))
158 		insn++;
159 #endif
160 	return insn;
161 }
162 NOKPROBE_SYMBOL(skip_prefixes);
163 
164 /*
165  * Returns non-zero if opcode is boostable.
166  * RIP relative instructions are adjusted at copying time in 64 bits mode
167  */
168 int can_boost(kprobe_opcode_t *opcodes)
169 {
170 	kprobe_opcode_t opcode;
171 	kprobe_opcode_t *orig_opcodes = opcodes;
172 
173 	if (search_exception_tables((unsigned long)opcodes))
174 		return 0;	/* Page fault may occur on this address. */
175 
176 retry:
177 	if (opcodes - orig_opcodes > MAX_INSN_SIZE - 1)
178 		return 0;
179 	opcode = *(opcodes++);
180 
181 	/* 2nd-byte opcode */
182 	if (opcode == 0x0f) {
183 		if (opcodes - orig_opcodes > MAX_INSN_SIZE - 1)
184 			return 0;
185 		return test_bit(*opcodes,
186 				(unsigned long *)twobyte_is_boostable);
187 	}
188 
189 	switch (opcode & 0xf0) {
190 #ifdef CONFIG_X86_64
191 	case 0x40:
192 		goto retry; /* REX prefix is boostable */
193 #endif
194 	case 0x60:
195 		if (0x63 < opcode && opcode < 0x67)
196 			goto retry; /* prefixes */
197 		/* can't boost Address-size override and bound */
198 		return (opcode != 0x62 && opcode != 0x67);
199 	case 0x70:
200 		return 0; /* can't boost conditional jump */
201 	case 0xc0:
202 		/* can't boost software-interruptions */
203 		return (0xc1 < opcode && opcode < 0xcc) || opcode == 0xcf;
204 	case 0xd0:
205 		/* can boost AA* and XLAT */
206 		return (opcode == 0xd4 || opcode == 0xd5 || opcode == 0xd7);
207 	case 0xe0:
208 		/* can boost in/out and absolute jmps */
209 		return ((opcode & 0x04) || opcode == 0xea);
210 	case 0xf0:
211 		if ((opcode & 0x0c) == 0 && opcode != 0xf1)
212 			goto retry; /* lock/rep(ne) prefix */
213 		/* clear and set flags are boostable */
214 		return (opcode == 0xf5 || (0xf7 < opcode && opcode < 0xfe));
215 	default:
216 		/* segment override prefixes are boostable */
217 		if (opcode == 0x26 || opcode == 0x36 || opcode == 0x3e)
218 			goto retry; /* prefixes */
219 		/* CS override prefix and call are not boostable */
220 		return (opcode != 0x2e && opcode != 0x9a);
221 	}
222 }
223 
224 static unsigned long
225 __recover_probed_insn(kprobe_opcode_t *buf, unsigned long addr)
226 {
227 	struct kprobe *kp;
228 	unsigned long faddr;
229 
230 	kp = get_kprobe((void *)addr);
231 	faddr = ftrace_location(addr);
232 	/*
233 	 * Addresses inside the ftrace location are refused by
234 	 * arch_check_ftrace_location(). Something went terribly wrong
235 	 * if such an address is checked here.
236 	 */
237 	if (WARN_ON(faddr && faddr != addr))
238 		return 0UL;
239 	/*
240 	 * Use the current code if it is not modified by Kprobe
241 	 * and it cannot be modified by ftrace.
242 	 */
243 	if (!kp && !faddr)
244 		return addr;
245 
246 	/*
247 	 * Basically, kp->ainsn.insn has an original instruction.
248 	 * However, RIP-relative instruction can not do single-stepping
249 	 * at different place, __copy_instruction() tweaks the displacement of
250 	 * that instruction. In that case, we can't recover the instruction
251 	 * from the kp->ainsn.insn.
252 	 *
253 	 * On the other hand, in case on normal Kprobe, kp->opcode has a copy
254 	 * of the first byte of the probed instruction, which is overwritten
255 	 * by int3. And the instruction at kp->addr is not modified by kprobes
256 	 * except for the first byte, we can recover the original instruction
257 	 * from it and kp->opcode.
258 	 *
259 	 * In case of Kprobes using ftrace, we do not have a copy of
260 	 * the original instruction. In fact, the ftrace location might
261 	 * be modified at anytime and even could be in an inconsistent state.
262 	 * Fortunately, we know that the original code is the ideal 5-byte
263 	 * long NOP.
264 	 */
265 	memcpy(buf, (void *)addr, MAX_INSN_SIZE * sizeof(kprobe_opcode_t));
266 	if (faddr)
267 		memcpy(buf, ideal_nops[NOP_ATOMIC5], 5);
268 	else
269 		buf[0] = kp->opcode;
270 	return (unsigned long)buf;
271 }
272 
273 /*
274  * Recover the probed instruction at addr for further analysis.
275  * Caller must lock kprobes by kprobe_mutex, or disable preemption
276  * for preventing to release referencing kprobes.
277  * Returns zero if the instruction can not get recovered.
278  */
279 unsigned long recover_probed_instruction(kprobe_opcode_t *buf, unsigned long addr)
280 {
281 	unsigned long __addr;
282 
283 	__addr = __recover_optprobed_insn(buf, addr);
284 	if (__addr != addr)
285 		return __addr;
286 
287 	return __recover_probed_insn(buf, addr);
288 }
289 
290 /* Check if paddr is at an instruction boundary */
291 static int can_probe(unsigned long paddr)
292 {
293 	unsigned long addr, __addr, offset = 0;
294 	struct insn insn;
295 	kprobe_opcode_t buf[MAX_INSN_SIZE];
296 
297 	if (!kallsyms_lookup_size_offset(paddr, NULL, &offset))
298 		return 0;
299 
300 	/* Decode instructions */
301 	addr = paddr - offset;
302 	while (addr < paddr) {
303 		/*
304 		 * Check if the instruction has been modified by another
305 		 * kprobe, in which case we replace the breakpoint by the
306 		 * original instruction in our buffer.
307 		 * Also, jump optimization will change the breakpoint to
308 		 * relative-jump. Since the relative-jump itself is
309 		 * normally used, we just go through if there is no kprobe.
310 		 */
311 		__addr = recover_probed_instruction(buf, addr);
312 		if (!__addr)
313 			return 0;
314 		kernel_insn_init(&insn, (void *)__addr, MAX_INSN_SIZE);
315 		insn_get_length(&insn);
316 
317 		/*
318 		 * Another debugging subsystem might insert this breakpoint.
319 		 * In that case, we can't recover it.
320 		 */
321 		if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION)
322 			return 0;
323 		addr += insn.length;
324 	}
325 
326 	return (addr == paddr);
327 }
328 
329 /*
330  * Returns non-zero if opcode modifies the interrupt flag.
331  */
332 static int is_IF_modifier(kprobe_opcode_t *insn)
333 {
334 	/* Skip prefixes */
335 	insn = skip_prefixes(insn);
336 
337 	switch (*insn) {
338 	case 0xfa:		/* cli */
339 	case 0xfb:		/* sti */
340 	case 0xcf:		/* iret/iretd */
341 	case 0x9d:		/* popf/popfd */
342 		return 1;
343 	}
344 
345 	return 0;
346 }
347 
348 /*
349  * Copy an instruction and adjust the displacement if the instruction
350  * uses the %rip-relative addressing mode.
351  * If it does, Return the address of the 32-bit displacement word.
352  * If not, return null.
353  * Only applicable to 64-bit x86.
354  */
355 int __copy_instruction(u8 *dest, u8 *src)
356 {
357 	struct insn insn;
358 	kprobe_opcode_t buf[MAX_INSN_SIZE];
359 	int length;
360 	unsigned long recovered_insn =
361 		recover_probed_instruction(buf, (unsigned long)src);
362 
363 	if (!recovered_insn)
364 		return 0;
365 	kernel_insn_init(&insn, (void *)recovered_insn, MAX_INSN_SIZE);
366 	insn_get_length(&insn);
367 	length = insn.length;
368 
369 	/* Another subsystem puts a breakpoint, failed to recover */
370 	if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION)
371 		return 0;
372 	memcpy(dest, insn.kaddr, length);
373 
374 #ifdef CONFIG_X86_64
375 	if (insn_rip_relative(&insn)) {
376 		s64 newdisp;
377 		u8 *disp;
378 		kernel_insn_init(&insn, dest, length);
379 		insn_get_displacement(&insn);
380 		/*
381 		 * The copied instruction uses the %rip-relative addressing
382 		 * mode.  Adjust the displacement for the difference between
383 		 * the original location of this instruction and the location
384 		 * of the copy that will actually be run.  The tricky bit here
385 		 * is making sure that the sign extension happens correctly in
386 		 * this calculation, since we need a signed 32-bit result to
387 		 * be sign-extended to 64 bits when it's added to the %rip
388 		 * value and yield the same 64-bit result that the sign-
389 		 * extension of the original signed 32-bit displacement would
390 		 * have given.
391 		 */
392 		newdisp = (u8 *) src + (s64) insn.displacement.value - (u8 *) dest;
393 		if ((s64) (s32) newdisp != newdisp) {
394 			pr_err("Kprobes error: new displacement does not fit into s32 (%llx)\n", newdisp);
395 			pr_err("\tSrc: %p, Dest: %p, old disp: %x\n", src, dest, insn.displacement.value);
396 			return 0;
397 		}
398 		disp = (u8 *) dest + insn_offset_displacement(&insn);
399 		*(s32 *) disp = (s32) newdisp;
400 	}
401 #endif
402 	return length;
403 }
404 
405 static int arch_copy_kprobe(struct kprobe *p)
406 {
407 	int ret;
408 
409 	/* Copy an instruction with recovering if other optprobe modifies it.*/
410 	ret = __copy_instruction(p->ainsn.insn, p->addr);
411 	if (!ret)
412 		return -EINVAL;
413 
414 	/*
415 	 * __copy_instruction can modify the displacement of the instruction,
416 	 * but it doesn't affect boostable check.
417 	 */
418 	if (can_boost(p->ainsn.insn))
419 		p->ainsn.boostable = 0;
420 	else
421 		p->ainsn.boostable = -1;
422 
423 	/* Check whether the instruction modifies Interrupt Flag or not */
424 	p->ainsn.if_modifier = is_IF_modifier(p->ainsn.insn);
425 
426 	/* Also, displacement change doesn't affect the first byte */
427 	p->opcode = p->ainsn.insn[0];
428 
429 	return 0;
430 }
431 
432 int arch_prepare_kprobe(struct kprobe *p)
433 {
434 	if (alternatives_text_reserved(p->addr, p->addr))
435 		return -EINVAL;
436 
437 	if (!can_probe((unsigned long)p->addr))
438 		return -EILSEQ;
439 	/* insn: must be on special executable page on x86. */
440 	p->ainsn.insn = get_insn_slot();
441 	if (!p->ainsn.insn)
442 		return -ENOMEM;
443 
444 	return arch_copy_kprobe(p);
445 }
446 
447 void arch_arm_kprobe(struct kprobe *p)
448 {
449 	text_poke(p->addr, ((unsigned char []){BREAKPOINT_INSTRUCTION}), 1);
450 }
451 
452 void arch_disarm_kprobe(struct kprobe *p)
453 {
454 	text_poke(p->addr, &p->opcode, 1);
455 }
456 
457 void arch_remove_kprobe(struct kprobe *p)
458 {
459 	if (p->ainsn.insn) {
460 		free_insn_slot(p->ainsn.insn, (p->ainsn.boostable == 1));
461 		p->ainsn.insn = NULL;
462 	}
463 }
464 
465 static nokprobe_inline void
466 save_previous_kprobe(struct kprobe_ctlblk *kcb)
467 {
468 	kcb->prev_kprobe.kp = kprobe_running();
469 	kcb->prev_kprobe.status = kcb->kprobe_status;
470 	kcb->prev_kprobe.old_flags = kcb->kprobe_old_flags;
471 	kcb->prev_kprobe.saved_flags = kcb->kprobe_saved_flags;
472 }
473 
474 static nokprobe_inline void
475 restore_previous_kprobe(struct kprobe_ctlblk *kcb)
476 {
477 	__this_cpu_write(current_kprobe, kcb->prev_kprobe.kp);
478 	kcb->kprobe_status = kcb->prev_kprobe.status;
479 	kcb->kprobe_old_flags = kcb->prev_kprobe.old_flags;
480 	kcb->kprobe_saved_flags = kcb->prev_kprobe.saved_flags;
481 }
482 
483 static nokprobe_inline void
484 set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
485 		   struct kprobe_ctlblk *kcb)
486 {
487 	__this_cpu_write(current_kprobe, p);
488 	kcb->kprobe_saved_flags = kcb->kprobe_old_flags
489 		= (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF));
490 	if (p->ainsn.if_modifier)
491 		kcb->kprobe_saved_flags &= ~X86_EFLAGS_IF;
492 }
493 
494 static nokprobe_inline void clear_btf(void)
495 {
496 	if (test_thread_flag(TIF_BLOCKSTEP)) {
497 		unsigned long debugctl = get_debugctlmsr();
498 
499 		debugctl &= ~DEBUGCTLMSR_BTF;
500 		update_debugctlmsr(debugctl);
501 	}
502 }
503 
504 static nokprobe_inline void restore_btf(void)
505 {
506 	if (test_thread_flag(TIF_BLOCKSTEP)) {
507 		unsigned long debugctl = get_debugctlmsr();
508 
509 		debugctl |= DEBUGCTLMSR_BTF;
510 		update_debugctlmsr(debugctl);
511 	}
512 }
513 
514 void arch_prepare_kretprobe(struct kretprobe_instance *ri, struct pt_regs *regs)
515 {
516 	unsigned long *sara = stack_addr(regs);
517 
518 	ri->ret_addr = (kprobe_opcode_t *) *sara;
519 
520 	/* Replace the return addr with trampoline addr */
521 	*sara = (unsigned long) &kretprobe_trampoline;
522 }
523 NOKPROBE_SYMBOL(arch_prepare_kretprobe);
524 
525 static void setup_singlestep(struct kprobe *p, struct pt_regs *regs,
526 			     struct kprobe_ctlblk *kcb, int reenter)
527 {
528 	if (setup_detour_execution(p, regs, reenter))
529 		return;
530 
531 #if !defined(CONFIG_PREEMPT)
532 	if (p->ainsn.boostable == 1 && !p->post_handler) {
533 		/* Boost up -- we can execute copied instructions directly */
534 		if (!reenter)
535 			reset_current_kprobe();
536 		/*
537 		 * Reentering boosted probe doesn't reset current_kprobe,
538 		 * nor set current_kprobe, because it doesn't use single
539 		 * stepping.
540 		 */
541 		regs->ip = (unsigned long)p->ainsn.insn;
542 		preempt_enable_no_resched();
543 		return;
544 	}
545 #endif
546 	if (reenter) {
547 		save_previous_kprobe(kcb);
548 		set_current_kprobe(p, regs, kcb);
549 		kcb->kprobe_status = KPROBE_REENTER;
550 	} else
551 		kcb->kprobe_status = KPROBE_HIT_SS;
552 	/* Prepare real single stepping */
553 	clear_btf();
554 	regs->flags |= X86_EFLAGS_TF;
555 	regs->flags &= ~X86_EFLAGS_IF;
556 	/* single step inline if the instruction is an int3 */
557 	if (p->opcode == BREAKPOINT_INSTRUCTION)
558 		regs->ip = (unsigned long)p->addr;
559 	else
560 		regs->ip = (unsigned long)p->ainsn.insn;
561 }
562 NOKPROBE_SYMBOL(setup_singlestep);
563 
564 /*
565  * We have reentered the kprobe_handler(), since another probe was hit while
566  * within the handler. We save the original kprobes variables and just single
567  * step on the instruction of the new probe without calling any user handlers.
568  */
569 static int reenter_kprobe(struct kprobe *p, struct pt_regs *regs,
570 			  struct kprobe_ctlblk *kcb)
571 {
572 	switch (kcb->kprobe_status) {
573 	case KPROBE_HIT_SSDONE:
574 	case KPROBE_HIT_ACTIVE:
575 	case KPROBE_HIT_SS:
576 		kprobes_inc_nmissed_count(p);
577 		setup_singlestep(p, regs, kcb, 1);
578 		break;
579 	case KPROBE_REENTER:
580 		/* A probe has been hit in the codepath leading up to, or just
581 		 * after, single-stepping of a probed instruction. This entire
582 		 * codepath should strictly reside in .kprobes.text section.
583 		 * Raise a BUG or we'll continue in an endless reentering loop
584 		 * and eventually a stack overflow.
585 		 */
586 		printk(KERN_WARNING "Unrecoverable kprobe detected at %p.\n",
587 		       p->addr);
588 		dump_kprobe(p);
589 		BUG();
590 	default:
591 		/* impossible cases */
592 		WARN_ON(1);
593 		return 0;
594 	}
595 
596 	return 1;
597 }
598 NOKPROBE_SYMBOL(reenter_kprobe);
599 
600 /*
601  * Interrupts are disabled on entry as trap3 is an interrupt gate and they
602  * remain disabled throughout this function.
603  */
604 int kprobe_int3_handler(struct pt_regs *regs)
605 {
606 	kprobe_opcode_t *addr;
607 	struct kprobe *p;
608 	struct kprobe_ctlblk *kcb;
609 
610 	if (user_mode(regs))
611 		return 0;
612 
613 	addr = (kprobe_opcode_t *)(regs->ip - sizeof(kprobe_opcode_t));
614 	/*
615 	 * We don't want to be preempted for the entire
616 	 * duration of kprobe processing. We conditionally
617 	 * re-enable preemption at the end of this function,
618 	 * and also in reenter_kprobe() and setup_singlestep().
619 	 */
620 	preempt_disable();
621 
622 	kcb = get_kprobe_ctlblk();
623 	p = get_kprobe(addr);
624 
625 	if (p) {
626 		if (kprobe_running()) {
627 			if (reenter_kprobe(p, regs, kcb))
628 				return 1;
629 		} else {
630 			set_current_kprobe(p, regs, kcb);
631 			kcb->kprobe_status = KPROBE_HIT_ACTIVE;
632 
633 			/*
634 			 * If we have no pre-handler or it returned 0, we
635 			 * continue with normal processing.  If we have a
636 			 * pre-handler and it returned non-zero, it prepped
637 			 * for calling the break_handler below on re-entry
638 			 * for jprobe processing, so get out doing nothing
639 			 * more here.
640 			 */
641 			if (!p->pre_handler || !p->pre_handler(p, regs))
642 				setup_singlestep(p, regs, kcb, 0);
643 			return 1;
644 		}
645 	} else if (*addr != BREAKPOINT_INSTRUCTION) {
646 		/*
647 		 * The breakpoint instruction was removed right
648 		 * after we hit it.  Another cpu has removed
649 		 * either a probepoint or a debugger breakpoint
650 		 * at this address.  In either case, no further
651 		 * handling of this interrupt is appropriate.
652 		 * Back up over the (now missing) int3 and run
653 		 * the original instruction.
654 		 */
655 		regs->ip = (unsigned long)addr;
656 		preempt_enable_no_resched();
657 		return 1;
658 	} else if (kprobe_running()) {
659 		p = __this_cpu_read(current_kprobe);
660 		if (p->break_handler && p->break_handler(p, regs)) {
661 			if (!skip_singlestep(p, regs, kcb))
662 				setup_singlestep(p, regs, kcb, 0);
663 			return 1;
664 		}
665 	} /* else: not a kprobe fault; let the kernel handle it */
666 
667 	preempt_enable_no_resched();
668 	return 0;
669 }
670 NOKPROBE_SYMBOL(kprobe_int3_handler);
671 
672 /*
673  * When a retprobed function returns, this code saves registers and
674  * calls trampoline_handler() runs, which calls the kretprobe's handler.
675  */
676 asm(
677 	".global kretprobe_trampoline\n"
678 	".type kretprobe_trampoline, @function\n"
679 	"kretprobe_trampoline:\n"
680 #ifdef CONFIG_X86_64
681 	/* We don't bother saving the ss register */
682 	"	pushq %rsp\n"
683 	"	pushfq\n"
684 	SAVE_REGS_STRING
685 	"	movq %rsp, %rdi\n"
686 	"	call trampoline_handler\n"
687 	/* Replace saved sp with true return address. */
688 	"	movq %rax, 152(%rsp)\n"
689 	RESTORE_REGS_STRING
690 	"	popfq\n"
691 #else
692 	"	pushf\n"
693 	SAVE_REGS_STRING
694 	"	movl %esp, %eax\n"
695 	"	call trampoline_handler\n"
696 	/* Move flags to cs */
697 	"	movl 56(%esp), %edx\n"
698 	"	movl %edx, 52(%esp)\n"
699 	/* Replace saved flags with true return address. */
700 	"	movl %eax, 56(%esp)\n"
701 	RESTORE_REGS_STRING
702 	"	popf\n"
703 #endif
704 	"	ret\n"
705 	".size kretprobe_trampoline, .-kretprobe_trampoline\n"
706 );
707 NOKPROBE_SYMBOL(kretprobe_trampoline);
708 STACK_FRAME_NON_STANDARD(kretprobe_trampoline);
709 
710 /*
711  * Called from kretprobe_trampoline
712  */
713 __visible __used void *trampoline_handler(struct pt_regs *regs)
714 {
715 	struct kretprobe_instance *ri = NULL;
716 	struct hlist_head *head, empty_rp;
717 	struct hlist_node *tmp;
718 	unsigned long flags, orig_ret_address = 0;
719 	unsigned long trampoline_address = (unsigned long)&kretprobe_trampoline;
720 	kprobe_opcode_t *correct_ret_addr = NULL;
721 
722 	INIT_HLIST_HEAD(&empty_rp);
723 	kretprobe_hash_lock(current, &head, &flags);
724 	/* fixup registers */
725 #ifdef CONFIG_X86_64
726 	regs->cs = __KERNEL_CS;
727 #else
728 	regs->cs = __KERNEL_CS | get_kernel_rpl();
729 	regs->gs = 0;
730 #endif
731 	regs->ip = trampoline_address;
732 	regs->orig_ax = ~0UL;
733 
734 	/*
735 	 * It is possible to have multiple instances associated with a given
736 	 * task either because multiple functions in the call path have
737 	 * return probes installed on them, and/or more than one
738 	 * return probe was registered for a target function.
739 	 *
740 	 * We can handle this because:
741 	 *     - instances are always pushed into the head of the list
742 	 *     - when multiple return probes are registered for the same
743 	 *	 function, the (chronologically) first instance's ret_addr
744 	 *	 will be the real return address, and all the rest will
745 	 *	 point to kretprobe_trampoline.
746 	 */
747 	hlist_for_each_entry_safe(ri, tmp, head, hlist) {
748 		if (ri->task != current)
749 			/* another task is sharing our hash bucket */
750 			continue;
751 
752 		orig_ret_address = (unsigned long)ri->ret_addr;
753 
754 		if (orig_ret_address != trampoline_address)
755 			/*
756 			 * This is the real return address. Any other
757 			 * instances associated with this task are for
758 			 * other calls deeper on the call stack
759 			 */
760 			break;
761 	}
762 
763 	kretprobe_assert(ri, orig_ret_address, trampoline_address);
764 
765 	correct_ret_addr = ri->ret_addr;
766 	hlist_for_each_entry_safe(ri, tmp, head, hlist) {
767 		if (ri->task != current)
768 			/* another task is sharing our hash bucket */
769 			continue;
770 
771 		orig_ret_address = (unsigned long)ri->ret_addr;
772 		if (ri->rp && ri->rp->handler) {
773 			__this_cpu_write(current_kprobe, &ri->rp->kp);
774 			get_kprobe_ctlblk()->kprobe_status = KPROBE_HIT_ACTIVE;
775 			ri->ret_addr = correct_ret_addr;
776 			ri->rp->handler(ri, regs);
777 			__this_cpu_write(current_kprobe, NULL);
778 		}
779 
780 		recycle_rp_inst(ri, &empty_rp);
781 
782 		if (orig_ret_address != trampoline_address)
783 			/*
784 			 * This is the real return address. Any other
785 			 * instances associated with this task are for
786 			 * other calls deeper on the call stack
787 			 */
788 			break;
789 	}
790 
791 	kretprobe_hash_unlock(current, &flags);
792 
793 	hlist_for_each_entry_safe(ri, tmp, &empty_rp, hlist) {
794 		hlist_del(&ri->hlist);
795 		kfree(ri);
796 	}
797 	return (void *)orig_ret_address;
798 }
799 NOKPROBE_SYMBOL(trampoline_handler);
800 
801 /*
802  * Called after single-stepping.  p->addr is the address of the
803  * instruction whose first byte has been replaced by the "int 3"
804  * instruction.  To avoid the SMP problems that can occur when we
805  * temporarily put back the original opcode to single-step, we
806  * single-stepped a copy of the instruction.  The address of this
807  * copy is p->ainsn.insn.
808  *
809  * This function prepares to return from the post-single-step
810  * interrupt.  We have to fix up the stack as follows:
811  *
812  * 0) Except in the case of absolute or indirect jump or call instructions,
813  * the new ip is relative to the copied instruction.  We need to make
814  * it relative to the original instruction.
815  *
816  * 1) If the single-stepped instruction was pushfl, then the TF and IF
817  * flags are set in the just-pushed flags, and may need to be cleared.
818  *
819  * 2) If the single-stepped instruction was a call, the return address
820  * that is atop the stack is the address following the copied instruction.
821  * We need to make it the address following the original instruction.
822  *
823  * If this is the first time we've single-stepped the instruction at
824  * this probepoint, and the instruction is boostable, boost it: add a
825  * jump instruction after the copied instruction, that jumps to the next
826  * instruction after the probepoint.
827  */
828 static void resume_execution(struct kprobe *p, struct pt_regs *regs,
829 			     struct kprobe_ctlblk *kcb)
830 {
831 	unsigned long *tos = stack_addr(regs);
832 	unsigned long copy_ip = (unsigned long)p->ainsn.insn;
833 	unsigned long orig_ip = (unsigned long)p->addr;
834 	kprobe_opcode_t *insn = p->ainsn.insn;
835 
836 	/* Skip prefixes */
837 	insn = skip_prefixes(insn);
838 
839 	regs->flags &= ~X86_EFLAGS_TF;
840 	switch (*insn) {
841 	case 0x9c:	/* pushfl */
842 		*tos &= ~(X86_EFLAGS_TF | X86_EFLAGS_IF);
843 		*tos |= kcb->kprobe_old_flags;
844 		break;
845 	case 0xc2:	/* iret/ret/lret */
846 	case 0xc3:
847 	case 0xca:
848 	case 0xcb:
849 	case 0xcf:
850 	case 0xea:	/* jmp absolute -- ip is correct */
851 		/* ip is already adjusted, no more changes required */
852 		p->ainsn.boostable = 1;
853 		goto no_change;
854 	case 0xe8:	/* call relative - Fix return addr */
855 		*tos = orig_ip + (*tos - copy_ip);
856 		break;
857 #ifdef CONFIG_X86_32
858 	case 0x9a:	/* call absolute -- same as call absolute, indirect */
859 		*tos = orig_ip + (*tos - copy_ip);
860 		goto no_change;
861 #endif
862 	case 0xff:
863 		if ((insn[1] & 0x30) == 0x10) {
864 			/*
865 			 * call absolute, indirect
866 			 * Fix return addr; ip is correct.
867 			 * But this is not boostable
868 			 */
869 			*tos = orig_ip + (*tos - copy_ip);
870 			goto no_change;
871 		} else if (((insn[1] & 0x31) == 0x20) ||
872 			   ((insn[1] & 0x31) == 0x21)) {
873 			/*
874 			 * jmp near and far, absolute indirect
875 			 * ip is correct. And this is boostable
876 			 */
877 			p->ainsn.boostable = 1;
878 			goto no_change;
879 		}
880 	default:
881 		break;
882 	}
883 
884 	if (p->ainsn.boostable == 0) {
885 		if ((regs->ip > copy_ip) &&
886 		    (regs->ip - copy_ip) + 5 < MAX_INSN_SIZE) {
887 			/*
888 			 * These instructions can be executed directly if it
889 			 * jumps back to correct address.
890 			 */
891 			synthesize_reljump((void *)regs->ip,
892 				(void *)orig_ip + (regs->ip - copy_ip));
893 			p->ainsn.boostable = 1;
894 		} else {
895 			p->ainsn.boostable = -1;
896 		}
897 	}
898 
899 	regs->ip += orig_ip - copy_ip;
900 
901 no_change:
902 	restore_btf();
903 }
904 NOKPROBE_SYMBOL(resume_execution);
905 
906 /*
907  * Interrupts are disabled on entry as trap1 is an interrupt gate and they
908  * remain disabled throughout this function.
909  */
910 int kprobe_debug_handler(struct pt_regs *regs)
911 {
912 	struct kprobe *cur = kprobe_running();
913 	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
914 
915 	if (!cur)
916 		return 0;
917 
918 	resume_execution(cur, regs, kcb);
919 	regs->flags |= kcb->kprobe_saved_flags;
920 
921 	if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
922 		kcb->kprobe_status = KPROBE_HIT_SSDONE;
923 		cur->post_handler(cur, regs, 0);
924 	}
925 
926 	/* Restore back the original saved kprobes variables and continue. */
927 	if (kcb->kprobe_status == KPROBE_REENTER) {
928 		restore_previous_kprobe(kcb);
929 		goto out;
930 	}
931 	reset_current_kprobe();
932 out:
933 	preempt_enable_no_resched();
934 
935 	/*
936 	 * if somebody else is singlestepping across a probe point, flags
937 	 * will have TF set, in which case, continue the remaining processing
938 	 * of do_debug, as if this is not a probe hit.
939 	 */
940 	if (regs->flags & X86_EFLAGS_TF)
941 		return 0;
942 
943 	return 1;
944 }
945 NOKPROBE_SYMBOL(kprobe_debug_handler);
946 
947 int kprobe_fault_handler(struct pt_regs *regs, int trapnr)
948 {
949 	struct kprobe *cur = kprobe_running();
950 	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
951 
952 	if (unlikely(regs->ip == (unsigned long)cur->ainsn.insn)) {
953 		/* This must happen on single-stepping */
954 		WARN_ON(kcb->kprobe_status != KPROBE_HIT_SS &&
955 			kcb->kprobe_status != KPROBE_REENTER);
956 		/*
957 		 * We are here because the instruction being single
958 		 * stepped caused a page fault. We reset the current
959 		 * kprobe and the ip points back to the probe address
960 		 * and allow the page fault handler to continue as a
961 		 * normal page fault.
962 		 */
963 		regs->ip = (unsigned long)cur->addr;
964 		/*
965 		 * Trap flag (TF) has been set here because this fault
966 		 * happened where the single stepping will be done.
967 		 * So clear it by resetting the current kprobe:
968 		 */
969 		regs->flags &= ~X86_EFLAGS_TF;
970 
971 		/*
972 		 * If the TF flag was set before the kprobe hit,
973 		 * don't touch it:
974 		 */
975 		regs->flags |= kcb->kprobe_old_flags;
976 
977 		if (kcb->kprobe_status == KPROBE_REENTER)
978 			restore_previous_kprobe(kcb);
979 		else
980 			reset_current_kprobe();
981 		preempt_enable_no_resched();
982 	} else if (kcb->kprobe_status == KPROBE_HIT_ACTIVE ||
983 		   kcb->kprobe_status == KPROBE_HIT_SSDONE) {
984 		/*
985 		 * We increment the nmissed count for accounting,
986 		 * we can also use npre/npostfault count for accounting
987 		 * these specific fault cases.
988 		 */
989 		kprobes_inc_nmissed_count(cur);
990 
991 		/*
992 		 * We come here because instructions in the pre/post
993 		 * handler caused the page_fault, this could happen
994 		 * if handler tries to access user space by
995 		 * copy_from_user(), get_user() etc. Let the
996 		 * user-specified handler try to fix it first.
997 		 */
998 		if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
999 			return 1;
1000 
1001 		/*
1002 		 * In case the user-specified fault handler returned
1003 		 * zero, try to fix up.
1004 		 */
1005 		if (fixup_exception(regs, trapnr))
1006 			return 1;
1007 
1008 		/*
1009 		 * fixup routine could not handle it,
1010 		 * Let do_page_fault() fix it.
1011 		 */
1012 	}
1013 
1014 	return 0;
1015 }
1016 NOKPROBE_SYMBOL(kprobe_fault_handler);
1017 
1018 /*
1019  * Wrapper routine for handling exceptions.
1020  */
1021 int kprobe_exceptions_notify(struct notifier_block *self, unsigned long val,
1022 			     void *data)
1023 {
1024 	struct die_args *args = data;
1025 	int ret = NOTIFY_DONE;
1026 
1027 	if (args->regs && user_mode(args->regs))
1028 		return ret;
1029 
1030 	if (val == DIE_GPF) {
1031 		/*
1032 		 * To be potentially processing a kprobe fault and to
1033 		 * trust the result from kprobe_running(), we have
1034 		 * be non-preemptible.
1035 		 */
1036 		if (!preemptible() && kprobe_running() &&
1037 		    kprobe_fault_handler(args->regs, args->trapnr))
1038 			ret = NOTIFY_STOP;
1039 	}
1040 	return ret;
1041 }
1042 NOKPROBE_SYMBOL(kprobe_exceptions_notify);
1043 
1044 int setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
1045 {
1046 	struct jprobe *jp = container_of(p, struct jprobe, kp);
1047 	unsigned long addr;
1048 	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1049 
1050 	kcb->jprobe_saved_regs = *regs;
1051 	kcb->jprobe_saved_sp = stack_addr(regs);
1052 	addr = (unsigned long)(kcb->jprobe_saved_sp);
1053 
1054 	/*
1055 	 * As Linus pointed out, gcc assumes that the callee
1056 	 * owns the argument space and could overwrite it, e.g.
1057 	 * tailcall optimization. So, to be absolutely safe
1058 	 * we also save and restore enough stack bytes to cover
1059 	 * the argument area.
1060 	 */
1061 	memcpy(kcb->jprobes_stack, (kprobe_opcode_t *)addr,
1062 	       MIN_STACK_SIZE(addr));
1063 	regs->flags &= ~X86_EFLAGS_IF;
1064 	trace_hardirqs_off();
1065 	regs->ip = (unsigned long)(jp->entry);
1066 
1067 	/*
1068 	 * jprobes use jprobe_return() which skips the normal return
1069 	 * path of the function, and this messes up the accounting of the
1070 	 * function graph tracer to get messed up.
1071 	 *
1072 	 * Pause function graph tracing while performing the jprobe function.
1073 	 */
1074 	pause_graph_tracing();
1075 	return 1;
1076 }
1077 NOKPROBE_SYMBOL(setjmp_pre_handler);
1078 
1079 void jprobe_return(void)
1080 {
1081 	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1082 
1083 	asm volatile (
1084 #ifdef CONFIG_X86_64
1085 			"       xchg   %%rbx,%%rsp	\n"
1086 #else
1087 			"       xchgl   %%ebx,%%esp	\n"
1088 #endif
1089 			"       int3			\n"
1090 			"       .globl jprobe_return_end\n"
1091 			"       jprobe_return_end:	\n"
1092 			"       nop			\n"::"b"
1093 			(kcb->jprobe_saved_sp):"memory");
1094 }
1095 NOKPROBE_SYMBOL(jprobe_return);
1096 NOKPROBE_SYMBOL(jprobe_return_end);
1097 
1098 int longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
1099 {
1100 	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1101 	u8 *addr = (u8 *) (regs->ip - 1);
1102 	struct jprobe *jp = container_of(p, struct jprobe, kp);
1103 	void *saved_sp = kcb->jprobe_saved_sp;
1104 
1105 	if ((addr > (u8 *) jprobe_return) &&
1106 	    (addr < (u8 *) jprobe_return_end)) {
1107 		if (stack_addr(regs) != saved_sp) {
1108 			struct pt_regs *saved_regs = &kcb->jprobe_saved_regs;
1109 			printk(KERN_ERR
1110 			       "current sp %p does not match saved sp %p\n",
1111 			       stack_addr(regs), saved_sp);
1112 			printk(KERN_ERR "Saved registers for jprobe %p\n", jp);
1113 			show_regs(saved_regs);
1114 			printk(KERN_ERR "Current registers\n");
1115 			show_regs(regs);
1116 			BUG();
1117 		}
1118 		/* It's OK to start function graph tracing again */
1119 		unpause_graph_tracing();
1120 		*regs = kcb->jprobe_saved_regs;
1121 		memcpy(saved_sp, kcb->jprobes_stack, MIN_STACK_SIZE(saved_sp));
1122 		preempt_enable_no_resched();
1123 		return 1;
1124 	}
1125 	return 0;
1126 }
1127 NOKPROBE_SYMBOL(longjmp_break_handler);
1128 
1129 bool arch_within_kprobe_blacklist(unsigned long addr)
1130 {
1131 	return  (addr >= (unsigned long)__kprobes_text_start &&
1132 		 addr < (unsigned long)__kprobes_text_end) ||
1133 		(addr >= (unsigned long)__entry_text_start &&
1134 		 addr < (unsigned long)__entry_text_end);
1135 }
1136 
1137 int __init arch_init_kprobes(void)
1138 {
1139 	return 0;
1140 }
1141 
1142 int arch_trampoline_kprobe(struct kprobe *p)
1143 {
1144 	return 0;
1145 }
1146