xref: /linux/arch/x86/kernel/kprobes/core.c (revision 7fc2cd2e4b398c57c9cf961cfea05eadbf34c05c)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  *  Kernel Probes (KProbes)
4  *
5  * Copyright (C) IBM Corporation, 2002, 2004
6  *
7  * 2002-Oct	Created by Vamsi Krishna S <vamsi_krishna@in.ibm.com> Kernel
8  *		Probes initial implementation ( includes contributions from
9  *		Rusty Russell).
10  * 2004-July	Suparna Bhattacharya <suparna@in.ibm.com> added jumper probes
11  *		interface to access function arguments.
12  * 2004-Oct	Jim Keniston <jkenisto@us.ibm.com> and Prasanna S Panchamukhi
13  *		<prasanna@in.ibm.com> adapted for x86_64 from i386.
14  * 2005-Mar	Roland McGrath <roland@redhat.com>
15  *		Fixed to handle %rip-relative addressing mode correctly.
16  * 2005-May	Hien Nguyen <hien@us.ibm.com>, Jim Keniston
17  *		<jkenisto@us.ibm.com> and Prasanna S Panchamukhi
18  *		<prasanna@in.ibm.com> added function-return probes.
19  * 2005-May	Rusty Lynch <rusty.lynch@intel.com>
20  *		Added function return probes functionality
21  * 2006-Feb	Masami Hiramatsu <hiramatu@sdl.hitachi.co.jp> added
22  *		kprobe-booster and kretprobe-booster for i386.
23  * 2007-Dec	Masami Hiramatsu <mhiramat@redhat.com> added kprobe-booster
24  *		and kretprobe-booster for x86-64
25  * 2007-Dec	Masami Hiramatsu <mhiramat@redhat.com>, Arjan van de Ven
26  *		<arjan@infradead.org> and Jim Keniston <jkenisto@us.ibm.com>
27  *		unified x86 kprobes code.
28  */
29 #include <linux/kprobes.h>
30 #include <linux/ptrace.h>
31 #include <linux/string.h>
32 #include <linux/slab.h>
33 #include <linux/hardirq.h>
34 #include <linux/preempt.h>
35 #include <linux/sched/debug.h>
36 #include <linux/perf_event.h>
37 #include <linux/extable.h>
38 #include <linux/kdebug.h>
39 #include <linux/kallsyms.h>
40 #include <linux/kgdb.h>
41 #include <linux/ftrace.h>
42 #include <linux/kasan.h>
43 #include <linux/objtool.h>
44 #include <linux/vmalloc.h>
45 #include <linux/pgtable.h>
46 #include <linux/set_memory.h>
47 #include <linux/cfi.h>
48 #include <linux/execmem.h>
49 
50 #include <asm/text-patching.h>
51 #include <asm/cacheflush.h>
52 #include <asm/desc.h>
53 #include <linux/uaccess.h>
54 #include <asm/alternative.h>
55 #include <asm/insn.h>
56 #include <asm/debugreg.h>
57 #include <asm/ibt.h>
58 
59 #include "common.h"
60 
61 DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
62 DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
63 
64 #define W(row, b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, ba, bb, bc, bd, be, bf)\
65 	(((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) |   \
66 	  (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) |   \
67 	  (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) |   \
68 	  (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf))    \
69 	 << (row % 32))
70 	/*
71 	 * Undefined/reserved opcodes, conditional jump, Opcode Extension
72 	 * Groups, and some special opcodes can not boost.
73 	 * This is non-const and volatile to keep gcc from statically
74 	 * optimizing it out, as variable_test_bit makes gcc think only
75 	 * *(unsigned long*) is used.
76 	 */
77 static volatile u32 twobyte_is_boostable[256 / 32] = {
78 	/*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f          */
79 	/*      ----------------------------------------------          */
80 	W(0x00, 0, 0, 1, 1, 0, 0, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0) | /* 00 */
81 	W(0x10, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1) , /* 10 */
82 	W(0x20, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 20 */
83 	W(0x30, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 30 */
84 	W(0x40, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 40 */
85 	W(0x50, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 50 */
86 	W(0x60, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1) | /* 60 */
87 	W(0x70, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1) , /* 70 */
88 	W(0x80, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 80 */
89 	W(0x90, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 90 */
90 	W(0xa0, 1, 1, 0, 1, 1, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* a0 */
91 	W(0xb0, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1) , /* b0 */
92 	W(0xc0, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1) | /* c0 */
93 	W(0xd0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) , /* d0 */
94 	W(0xe0, 0, 1, 1, 0, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* e0 */
95 	W(0xf0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 1, 0, 1, 1, 1, 0)   /* f0 */
96 	/*      -----------------------------------------------         */
97 	/*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f          */
98 };
99 #undef W
100 
101 struct kretprobe_blackpoint kretprobe_blacklist[] = {
102 	{"__switch_to", }, /* This function switches only current task, but
103 			      doesn't switch kernel stack.*/
104 	{NULL, NULL}	/* Terminator */
105 };
106 
107 const int kretprobe_blacklist_size = ARRAY_SIZE(kretprobe_blacklist);
108 
109 static nokprobe_inline void
110 __synthesize_relative_insn(void *dest, void *from, void *to, u8 op)
111 {
112 	struct __arch_relative_insn {
113 		u8 op;
114 		s32 raddr;
115 	} __packed *insn;
116 
117 	insn = (struct __arch_relative_insn *)dest;
118 	insn->raddr = (s32)((long)(to) - ((long)(from) + 5));
119 	insn->op = op;
120 }
121 
122 /* Insert a jump instruction at address 'from', which jumps to address 'to'.*/
123 void synthesize_reljump(void *dest, void *from, void *to)
124 {
125 	__synthesize_relative_insn(dest, from, to, JMP32_INSN_OPCODE);
126 }
127 NOKPROBE_SYMBOL(synthesize_reljump);
128 
129 /* Insert a call instruction at address 'from', which calls address 'to'.*/
130 void synthesize_relcall(void *dest, void *from, void *to)
131 {
132 	__synthesize_relative_insn(dest, from, to, CALL_INSN_OPCODE);
133 }
134 NOKPROBE_SYMBOL(synthesize_relcall);
135 
136 /*
137  * Returns non-zero if INSN is boostable.
138  * RIP relative instructions are adjusted at copying time in 64 bits mode
139  */
140 bool can_boost(struct insn *insn, void *addr)
141 {
142 	kprobe_opcode_t opcode;
143 	insn_byte_t prefix;
144 
145 	if (search_exception_tables((unsigned long)addr))
146 		return false;	/* Page fault may occur on this address. */
147 
148 	/* 2nd-byte opcode */
149 	if (insn->opcode.nbytes == 2)
150 		return test_bit(insn->opcode.bytes[1],
151 				(unsigned long *)twobyte_is_boostable);
152 
153 	if (insn->opcode.nbytes != 1)
154 		return false;
155 
156 	for_each_insn_prefix(insn, prefix) {
157 		insn_attr_t attr;
158 
159 		attr = inat_get_opcode_attribute(prefix);
160 		/* Can't boost Address-size override prefix and CS override prefix */
161 		if (prefix == 0x2e || inat_is_address_size_prefix(attr))
162 			return false;
163 	}
164 
165 	opcode = insn->opcode.bytes[0];
166 
167 	switch (opcode) {
168 	case 0x62:		/* bound */
169 	case 0x70 ... 0x7f:	/* Conditional jumps */
170 	case 0x9a:		/* Call far */
171 	case 0xcc ... 0xce:	/* software exceptions */
172 	case 0xd6:		/* (UD) */
173 	case 0xd8 ... 0xdf:	/* ESC */
174 	case 0xe0 ... 0xe3:	/* LOOP*, JCXZ */
175 	case 0xe8 ... 0xe9:	/* near Call, JMP */
176 	case 0xeb:		/* Short JMP */
177 	case 0xf0 ... 0xf4:	/* LOCK/REP, HLT */
178 		/* ... are not boostable */
179 		return false;
180 	case 0xc0 ... 0xc1:	/* Grp2 */
181 	case 0xd0 ... 0xd3:	/* Grp2 */
182 		/*
183 		 * AMD uses nnn == 110 as SHL/SAL, but Intel makes it reserved.
184 		 */
185 		return X86_MODRM_REG(insn->modrm.bytes[0]) != 0b110;
186 	case 0xf6 ... 0xf7:	/* Grp3 */
187 		/* AMD uses nnn == 001 as TEST, but Intel makes it reserved. */
188 		return X86_MODRM_REG(insn->modrm.bytes[0]) != 0b001;
189 	case 0xfe:		/* Grp4 */
190 		/* Only INC and DEC are boostable */
191 		return X86_MODRM_REG(insn->modrm.bytes[0]) == 0b000 ||
192 		       X86_MODRM_REG(insn->modrm.bytes[0]) == 0b001;
193 	case 0xff:		/* Grp5 */
194 		/* Only INC, DEC, and indirect JMP are boostable */
195 		return X86_MODRM_REG(insn->modrm.bytes[0]) == 0b000 ||
196 		       X86_MODRM_REG(insn->modrm.bytes[0]) == 0b001 ||
197 		       X86_MODRM_REG(insn->modrm.bytes[0]) == 0b100;
198 	default:
199 		return true;
200 	}
201 }
202 
203 static unsigned long
204 __recover_probed_insn(kprobe_opcode_t *buf, unsigned long addr)
205 {
206 	struct kprobe *kp;
207 	bool faddr;
208 
209 	kp = get_kprobe((void *)addr);
210 	faddr = ftrace_location(addr) == addr;
211 	/*
212 	 * Use the current code if it is not modified by Kprobe
213 	 * and it cannot be modified by ftrace.
214 	 */
215 	if (!kp && !faddr)
216 		return addr;
217 
218 	/*
219 	 * Basically, kp->ainsn.insn has an original instruction.
220 	 * However, RIP-relative instruction can not do single-stepping
221 	 * at different place, __copy_instruction() tweaks the displacement of
222 	 * that instruction. In that case, we can't recover the instruction
223 	 * from the kp->ainsn.insn.
224 	 *
225 	 * On the other hand, in case on normal Kprobe, kp->opcode has a copy
226 	 * of the first byte of the probed instruction, which is overwritten
227 	 * by int3. And the instruction at kp->addr is not modified by kprobes
228 	 * except for the first byte, we can recover the original instruction
229 	 * from it and kp->opcode.
230 	 *
231 	 * In case of Kprobes using ftrace, we do not have a copy of
232 	 * the original instruction. In fact, the ftrace location might
233 	 * be modified at anytime and even could be in an inconsistent state.
234 	 * Fortunately, we know that the original code is the ideal 5-byte
235 	 * long NOP.
236 	 */
237 	if (copy_from_kernel_nofault(buf, (void *)addr,
238 		MAX_INSN_SIZE * sizeof(kprobe_opcode_t)))
239 		return 0UL;
240 
241 	if (faddr)
242 		memcpy(buf, x86_nops[5], 5);
243 	else
244 		buf[0] = kp->opcode;
245 	return (unsigned long)buf;
246 }
247 
248 /*
249  * Recover the probed instruction at addr for further analysis.
250  * Caller must lock kprobes by kprobe_mutex, or disable preemption
251  * for preventing to release referencing kprobes.
252  * Returns zero if the instruction can not get recovered (or access failed).
253  */
254 unsigned long recover_probed_instruction(kprobe_opcode_t *buf, unsigned long addr)
255 {
256 	unsigned long __addr;
257 
258 	__addr = __recover_optprobed_insn(buf, addr);
259 	if (__addr != addr)
260 		return __addr;
261 
262 	return __recover_probed_insn(buf, addr);
263 }
264 
265 /* Check if insn is INT or UD */
266 static inline bool is_exception_insn(struct insn *insn)
267 {
268 	/* UD uses 0f escape */
269 	if (insn->opcode.bytes[0] == 0x0f) {
270 		/* UD0 / UD1 / UD2 */
271 		return insn->opcode.bytes[1] == 0xff ||
272 		       insn->opcode.bytes[1] == 0xb9 ||
273 		       insn->opcode.bytes[1] == 0x0b;
274 	}
275 
276 	/* INT3 / INT n / INTO / INT1 */
277 	return insn->opcode.bytes[0] == 0xcc ||
278 	       insn->opcode.bytes[0] == 0xcd ||
279 	       insn->opcode.bytes[0] == 0xce ||
280 	       insn->opcode.bytes[0] == 0xf1;
281 }
282 
283 /*
284  * Check if paddr is at an instruction boundary and that instruction can
285  * be probed
286  */
287 static bool can_probe(unsigned long paddr)
288 {
289 	unsigned long addr, __addr, offset = 0;
290 	struct insn insn;
291 	kprobe_opcode_t buf[MAX_INSN_SIZE];
292 
293 	if (!kallsyms_lookup_size_offset(paddr, NULL, &offset))
294 		return false;
295 
296 	/* Decode instructions */
297 	addr = paddr - offset;
298 	while (addr < paddr) {
299 		/*
300 		 * Check if the instruction has been modified by another
301 		 * kprobe, in which case we replace the breakpoint by the
302 		 * original instruction in our buffer.
303 		 * Also, jump optimization will change the breakpoint to
304 		 * relative-jump. Since the relative-jump itself is
305 		 * normally used, we just go through if there is no kprobe.
306 		 */
307 		__addr = recover_probed_instruction(buf, addr);
308 		if (!__addr)
309 			return false;
310 
311 		if (insn_decode_kernel(&insn, (void *)__addr) < 0)
312 			return false;
313 
314 #ifdef CONFIG_KGDB
315 		/*
316 		 * If there is a dynamically installed kgdb sw breakpoint,
317 		 * this function should not be probed.
318 		 */
319 		if (insn.opcode.bytes[0] == INT3_INSN_OPCODE &&
320 		    kgdb_has_hit_break(addr))
321 			return false;
322 #endif
323 		addr += insn.length;
324 	}
325 
326 	/* Check if paddr is at an instruction boundary */
327 	if (addr != paddr)
328 		return false;
329 
330 	__addr = recover_probed_instruction(buf, addr);
331 	if (!__addr)
332 		return false;
333 
334 	if (insn_decode_kernel(&insn, (void *)__addr) < 0)
335 		return false;
336 
337 	/* INT and UD are special and should not be kprobed */
338 	if (is_exception_insn(&insn))
339 		return false;
340 
341 	if (IS_ENABLED(CONFIG_CFI)) {
342 		/*
343 		 * The compiler generates the following instruction sequence
344 		 * for indirect call checks and cfi.c decodes this;
345 		 *
346 		 *   movl    -<id>, %r10d       ; 6 bytes
347 		 *   addl    -4(%reg), %r10d    ; 4 bytes
348 		 *   je      .Ltmp1             ; 2 bytes
349 		 *   ud2                        ; <- regs->ip
350 		 *   .Ltmp1:
351 		 *
352 		 * Also, these movl and addl are used for showing expected
353 		 * type. So those must not be touched.
354 		 */
355 		if (insn.opcode.value == 0xBA)
356 			offset = 12;
357 		else if (insn.opcode.value == 0x3)
358 			offset = 6;
359 		else
360 			goto out;
361 
362 		/* This movl/addl is used for decoding CFI. */
363 		if (is_cfi_trap(addr + offset))
364 			return false;
365 	}
366 
367 out:
368 	return true;
369 }
370 
371 /* If x86 supports IBT (ENDBR) it must be skipped. */
372 kprobe_opcode_t *arch_adjust_kprobe_addr(unsigned long addr, unsigned long offset,
373 					 bool *on_func_entry)
374 {
375 	if (is_endbr((u32 *)addr)) {
376 		*on_func_entry = !offset || offset == 4;
377 		if (*on_func_entry)
378 			offset = 4;
379 
380 	} else {
381 		*on_func_entry = !offset;
382 	}
383 
384 	return (kprobe_opcode_t *)(addr + offset);
385 }
386 
387 /*
388  * Copy an instruction with recovering modified instruction by kprobes
389  * and adjust the displacement if the instruction uses the %rip-relative
390  * addressing mode. Note that since @real will be the final place of copied
391  * instruction, displacement must be adjust by @real, not @dest.
392  * This returns the length of copied instruction, or 0 if it has an error.
393  */
394 int __copy_instruction(u8 *dest, u8 *src, u8 *real, struct insn *insn)
395 {
396 	kprobe_opcode_t buf[MAX_INSN_SIZE];
397 	unsigned long recovered_insn = recover_probed_instruction(buf, (unsigned long)src);
398 	int ret;
399 
400 	if (!recovered_insn || !insn)
401 		return 0;
402 
403 	/* This can access kernel text if given address is not recovered */
404 	if (copy_from_kernel_nofault(dest, (void *)recovered_insn,
405 			MAX_INSN_SIZE))
406 		return 0;
407 
408 	ret = insn_decode_kernel(insn, dest);
409 	if (ret < 0)
410 		return 0;
411 
412 	/* We can not probe force emulate prefixed instruction */
413 	if (insn_has_emulate_prefix(insn))
414 		return 0;
415 
416 	/* Another subsystem puts a breakpoint, failed to recover */
417 	if (insn->opcode.bytes[0] == INT3_INSN_OPCODE)
418 		return 0;
419 
420 	/* We should not singlestep on the exception masking instructions */
421 	if (insn_masking_exception(insn))
422 		return 0;
423 
424 #ifdef CONFIG_X86_64
425 	/* Only x86_64 has RIP relative instructions */
426 	if (insn_rip_relative(insn)) {
427 		s64 newdisp;
428 		u8 *disp;
429 		/*
430 		 * The copied instruction uses the %rip-relative addressing
431 		 * mode.  Adjust the displacement for the difference between
432 		 * the original location of this instruction and the location
433 		 * of the copy that will actually be run.  The tricky bit here
434 		 * is making sure that the sign extension happens correctly in
435 		 * this calculation, since we need a signed 32-bit result to
436 		 * be sign-extended to 64 bits when it's added to the %rip
437 		 * value and yield the same 64-bit result that the sign-
438 		 * extension of the original signed 32-bit displacement would
439 		 * have given.
440 		 */
441 		newdisp = (u8 *) src + (s64) insn->displacement.value
442 			  - (u8 *) real;
443 		if ((s64) (s32) newdisp != newdisp) {
444 			pr_err("Kprobes error: new displacement does not fit into s32 (%llx)\n", newdisp);
445 			return 0;
446 		}
447 		disp = (u8 *) dest + insn_offset_displacement(insn);
448 		*(s32 *) disp = (s32) newdisp;
449 	}
450 #endif
451 	return insn->length;
452 }
453 
454 /* Prepare reljump or int3 right after instruction */
455 static int prepare_singlestep(kprobe_opcode_t *buf, struct kprobe *p,
456 			      struct insn *insn)
457 {
458 	int len = insn->length;
459 
460 	if (!IS_ENABLED(CONFIG_PREEMPTION) &&
461 	    !p->post_handler && can_boost(insn, p->addr) &&
462 	    MAX_INSN_SIZE - len >= JMP32_INSN_SIZE) {
463 		/*
464 		 * These instructions can be executed directly if it
465 		 * jumps back to correct address.
466 		 */
467 		synthesize_reljump(buf + len, p->ainsn.insn + len,
468 				   p->addr + insn->length);
469 		len += JMP32_INSN_SIZE;
470 		p->ainsn.boostable = 1;
471 	} else {
472 		/* Otherwise, put an int3 for trapping singlestep */
473 		if (MAX_INSN_SIZE - len < INT3_INSN_SIZE)
474 			return -ENOSPC;
475 
476 		buf[len] = INT3_INSN_OPCODE;
477 		len += INT3_INSN_SIZE;
478 	}
479 
480 	return len;
481 }
482 
483 /* Kprobe x86 instruction emulation - only regs->ip or IF flag modifiers */
484 
485 static void kprobe_emulate_ifmodifiers(struct kprobe *p, struct pt_regs *regs)
486 {
487 	switch (p->ainsn.opcode) {
488 	case 0xfa:	/* cli */
489 		regs->flags &= ~(X86_EFLAGS_IF);
490 		break;
491 	case 0xfb:	/* sti */
492 		regs->flags |= X86_EFLAGS_IF;
493 		break;
494 	case 0x9c:	/* pushf */
495 		int3_emulate_push(regs, regs->flags);
496 		break;
497 	case 0x9d:	/* popf */
498 		regs->flags = int3_emulate_pop(regs);
499 		break;
500 	}
501 	regs->ip = regs->ip - INT3_INSN_SIZE + p->ainsn.size;
502 }
503 NOKPROBE_SYMBOL(kprobe_emulate_ifmodifiers);
504 
505 static void kprobe_emulate_ret(struct kprobe *p, struct pt_regs *regs)
506 {
507 	int3_emulate_ret(regs);
508 }
509 NOKPROBE_SYMBOL(kprobe_emulate_ret);
510 
511 static void kprobe_emulate_call(struct kprobe *p, struct pt_regs *regs)
512 {
513 	unsigned long func = regs->ip - INT3_INSN_SIZE + p->ainsn.size;
514 
515 	func += p->ainsn.rel32;
516 	int3_emulate_call(regs, func);
517 }
518 NOKPROBE_SYMBOL(kprobe_emulate_call);
519 
520 static void kprobe_emulate_jmp(struct kprobe *p, struct pt_regs *regs)
521 {
522 	unsigned long ip = regs->ip - INT3_INSN_SIZE + p->ainsn.size;
523 
524 	ip += p->ainsn.rel32;
525 	int3_emulate_jmp(regs, ip);
526 }
527 NOKPROBE_SYMBOL(kprobe_emulate_jmp);
528 
529 static void kprobe_emulate_jcc(struct kprobe *p, struct pt_regs *regs)
530 {
531 	unsigned long ip = regs->ip - INT3_INSN_SIZE + p->ainsn.size;
532 
533 	int3_emulate_jcc(regs, p->ainsn.jcc.type, ip, p->ainsn.rel32);
534 }
535 NOKPROBE_SYMBOL(kprobe_emulate_jcc);
536 
537 static void kprobe_emulate_loop(struct kprobe *p, struct pt_regs *regs)
538 {
539 	unsigned long ip = regs->ip - INT3_INSN_SIZE + p->ainsn.size;
540 	bool match;
541 
542 	if (p->ainsn.loop.type != 3) {	/* LOOP* */
543 		if (p->ainsn.loop.asize == 32)
544 			match = ((*(u32 *)&regs->cx)--) != 0;
545 #ifdef CONFIG_X86_64
546 		else if (p->ainsn.loop.asize == 64)
547 			match = ((*(u64 *)&regs->cx)--) != 0;
548 #endif
549 		else
550 			match = ((*(u16 *)&regs->cx)--) != 0;
551 	} else {			/* JCXZ */
552 		if (p->ainsn.loop.asize == 32)
553 			match = *(u32 *)(&regs->cx) == 0;
554 #ifdef CONFIG_X86_64
555 		else if (p->ainsn.loop.asize == 64)
556 			match = *(u64 *)(&regs->cx) == 0;
557 #endif
558 		else
559 			match = *(u16 *)(&regs->cx) == 0;
560 	}
561 
562 	if (p->ainsn.loop.type == 0)	/* LOOPNE */
563 		match = match && !(regs->flags & X86_EFLAGS_ZF);
564 	else if (p->ainsn.loop.type == 1)	/* LOOPE */
565 		match = match && (regs->flags & X86_EFLAGS_ZF);
566 
567 	if (match)
568 		ip += p->ainsn.rel32;
569 	int3_emulate_jmp(regs, ip);
570 }
571 NOKPROBE_SYMBOL(kprobe_emulate_loop);
572 
573 static const int addrmode_regoffs[] = {
574 	offsetof(struct pt_regs, ax),
575 	offsetof(struct pt_regs, cx),
576 	offsetof(struct pt_regs, dx),
577 	offsetof(struct pt_regs, bx),
578 	offsetof(struct pt_regs, sp),
579 	offsetof(struct pt_regs, bp),
580 	offsetof(struct pt_regs, si),
581 	offsetof(struct pt_regs, di),
582 #ifdef CONFIG_X86_64
583 	offsetof(struct pt_regs, r8),
584 	offsetof(struct pt_regs, r9),
585 	offsetof(struct pt_regs, r10),
586 	offsetof(struct pt_regs, r11),
587 	offsetof(struct pt_regs, r12),
588 	offsetof(struct pt_regs, r13),
589 	offsetof(struct pt_regs, r14),
590 	offsetof(struct pt_regs, r15),
591 #endif
592 };
593 
594 static void kprobe_emulate_call_indirect(struct kprobe *p, struct pt_regs *regs)
595 {
596 	unsigned long offs = addrmode_regoffs[p->ainsn.indirect.reg];
597 
598 	int3_emulate_push(regs, regs->ip - INT3_INSN_SIZE + p->ainsn.size);
599 	int3_emulate_jmp(regs, regs_get_register(regs, offs));
600 }
601 NOKPROBE_SYMBOL(kprobe_emulate_call_indirect);
602 
603 static void kprobe_emulate_jmp_indirect(struct kprobe *p, struct pt_regs *regs)
604 {
605 	unsigned long offs = addrmode_regoffs[p->ainsn.indirect.reg];
606 
607 	int3_emulate_jmp(regs, regs_get_register(regs, offs));
608 }
609 NOKPROBE_SYMBOL(kprobe_emulate_jmp_indirect);
610 
611 static int prepare_emulation(struct kprobe *p, struct insn *insn)
612 {
613 	insn_byte_t opcode = insn->opcode.bytes[0];
614 
615 	switch (opcode) {
616 	case 0xfa:		/* cli */
617 	case 0xfb:		/* sti */
618 	case 0x9c:		/* pushfl */
619 	case 0x9d:		/* popf/popfd */
620 		/*
621 		 * IF modifiers must be emulated since it will enable interrupt while
622 		 * int3 single stepping.
623 		 */
624 		p->ainsn.emulate_op = kprobe_emulate_ifmodifiers;
625 		p->ainsn.opcode = opcode;
626 		break;
627 	case 0xc2:	/* ret/lret */
628 	case 0xc3:
629 	case 0xca:
630 	case 0xcb:
631 		p->ainsn.emulate_op = kprobe_emulate_ret;
632 		break;
633 	case 0x9a:	/* far call absolute -- segment is not supported */
634 	case 0xea:	/* far jmp absolute -- segment is not supported */
635 	case 0xcc:	/* int3 */
636 	case 0xcf:	/* iret -- in-kernel IRET is not supported */
637 		return -EOPNOTSUPP;
638 		break;
639 	case 0xe8:	/* near call relative */
640 		p->ainsn.emulate_op = kprobe_emulate_call;
641 		if (insn->immediate.nbytes == 2)
642 			p->ainsn.rel32 = *(s16 *)&insn->immediate.value;
643 		else
644 			p->ainsn.rel32 = *(s32 *)&insn->immediate.value;
645 		break;
646 	case 0xeb:	/* short jump relative */
647 	case 0xe9:	/* near jump relative */
648 		p->ainsn.emulate_op = kprobe_emulate_jmp;
649 		if (insn->immediate.nbytes == 1)
650 			p->ainsn.rel32 = *(s8 *)&insn->immediate.value;
651 		else if (insn->immediate.nbytes == 2)
652 			p->ainsn.rel32 = *(s16 *)&insn->immediate.value;
653 		else
654 			p->ainsn.rel32 = *(s32 *)&insn->immediate.value;
655 		break;
656 	case 0x70 ... 0x7f:
657 		/* 1 byte conditional jump */
658 		p->ainsn.emulate_op = kprobe_emulate_jcc;
659 		p->ainsn.jcc.type = opcode & 0xf;
660 		p->ainsn.rel32 = insn->immediate.value;
661 		break;
662 	case 0x0f:
663 		opcode = insn->opcode.bytes[1];
664 		if ((opcode & 0xf0) == 0x80) {
665 			/* 2 bytes Conditional Jump */
666 			p->ainsn.emulate_op = kprobe_emulate_jcc;
667 			p->ainsn.jcc.type = opcode & 0xf;
668 			if (insn->immediate.nbytes == 2)
669 				p->ainsn.rel32 = *(s16 *)&insn->immediate.value;
670 			else
671 				p->ainsn.rel32 = *(s32 *)&insn->immediate.value;
672 		} else if (opcode == 0x01 &&
673 			   X86_MODRM_REG(insn->modrm.bytes[0]) == 0 &&
674 			   X86_MODRM_MOD(insn->modrm.bytes[0]) == 3) {
675 			/* VM extensions - not supported */
676 			return -EOPNOTSUPP;
677 		}
678 		break;
679 	case 0xe0:	/* Loop NZ */
680 	case 0xe1:	/* Loop */
681 	case 0xe2:	/* Loop */
682 	case 0xe3:	/* J*CXZ */
683 		p->ainsn.emulate_op = kprobe_emulate_loop;
684 		p->ainsn.loop.type = opcode & 0x3;
685 		p->ainsn.loop.asize = insn->addr_bytes * 8;
686 		p->ainsn.rel32 = *(s8 *)&insn->immediate.value;
687 		break;
688 	case 0xff:
689 		/*
690 		 * Since the 0xff is an extended group opcode, the instruction
691 		 * is determined by the MOD/RM byte.
692 		 */
693 		opcode = insn->modrm.bytes[0];
694 		switch (X86_MODRM_REG(opcode)) {
695 		case 0b010:	/* FF /2, call near, absolute indirect */
696 			p->ainsn.emulate_op = kprobe_emulate_call_indirect;
697 			break;
698 		case 0b100:	/* FF /4, jmp near, absolute indirect */
699 			p->ainsn.emulate_op = kprobe_emulate_jmp_indirect;
700 			break;
701 		case 0b011:	/* FF /3, call far, absolute indirect */
702 		case 0b101:	/* FF /5, jmp far, absolute indirect */
703 			return -EOPNOTSUPP;
704 		}
705 
706 		if (!p->ainsn.emulate_op)
707 			break;
708 
709 		if (insn->addr_bytes != sizeof(unsigned long))
710 			return -EOPNOTSUPP;	/* Don't support different size */
711 		if (X86_MODRM_MOD(opcode) != 3)
712 			return -EOPNOTSUPP;	/* TODO: support memory addressing */
713 
714 		p->ainsn.indirect.reg = X86_MODRM_RM(opcode);
715 #ifdef CONFIG_X86_64
716 		if (X86_REX_B(insn->rex_prefix.value))
717 			p->ainsn.indirect.reg += 8;
718 #endif
719 		break;
720 	default:
721 		break;
722 	}
723 	p->ainsn.size = insn->length;
724 
725 	return 0;
726 }
727 
728 static int arch_copy_kprobe(struct kprobe *p)
729 {
730 	struct insn insn;
731 	kprobe_opcode_t buf[MAX_INSN_SIZE];
732 	int ret, len;
733 
734 	/* Copy an instruction with recovering if other optprobe modifies it.*/
735 	len = __copy_instruction(buf, p->addr, p->ainsn.insn, &insn);
736 	if (!len)
737 		return -EINVAL;
738 
739 	/* Analyze the opcode and setup emulate functions */
740 	ret = prepare_emulation(p, &insn);
741 	if (ret < 0)
742 		return ret;
743 
744 	/* Add int3 for single-step or booster jmp */
745 	len = prepare_singlestep(buf, p, &insn);
746 	if (len < 0)
747 		return len;
748 
749 	/* Also, displacement change doesn't affect the first byte */
750 	p->opcode = buf[0];
751 
752 	p->ainsn.tp_len = len;
753 	perf_event_text_poke(p->ainsn.insn, NULL, 0, buf, len);
754 
755 	/* OK, write back the instruction(s) into ROX insn buffer */
756 	text_poke(p->ainsn.insn, buf, len);
757 
758 	return 0;
759 }
760 
761 int arch_prepare_kprobe(struct kprobe *p)
762 {
763 	int ret;
764 
765 	if (alternatives_text_reserved(p->addr, p->addr))
766 		return -EINVAL;
767 
768 	if (!can_probe((unsigned long)p->addr))
769 		return -EILSEQ;
770 
771 	memset(&p->ainsn, 0, sizeof(p->ainsn));
772 
773 	/* insn: must be on special executable page on x86. */
774 	p->ainsn.insn = get_insn_slot();
775 	if (!p->ainsn.insn)
776 		return -ENOMEM;
777 
778 	ret = arch_copy_kprobe(p);
779 	if (ret) {
780 		free_insn_slot(p->ainsn.insn, 0);
781 		p->ainsn.insn = NULL;
782 	}
783 
784 	return ret;
785 }
786 
787 void arch_arm_kprobe(struct kprobe *p)
788 {
789 	u8 int3 = INT3_INSN_OPCODE;
790 
791 	text_poke(p->addr, &int3, 1);
792 	smp_text_poke_sync_each_cpu();
793 	perf_event_text_poke(p->addr, &p->opcode, 1, &int3, 1);
794 }
795 
796 void arch_disarm_kprobe(struct kprobe *p)
797 {
798 	u8 int3 = INT3_INSN_OPCODE;
799 
800 	perf_event_text_poke(p->addr, &int3, 1, &p->opcode, 1);
801 	text_poke(p->addr, &p->opcode, 1);
802 	smp_text_poke_sync_each_cpu();
803 }
804 
805 void arch_remove_kprobe(struct kprobe *p)
806 {
807 	if (p->ainsn.insn) {
808 		/* Record the perf event before freeing the slot */
809 		perf_event_text_poke(p->ainsn.insn, p->ainsn.insn,
810 				     p->ainsn.tp_len, NULL, 0);
811 		free_insn_slot(p->ainsn.insn, p->ainsn.boostable);
812 		p->ainsn.insn = NULL;
813 	}
814 }
815 
816 static nokprobe_inline void
817 save_previous_kprobe(struct kprobe_ctlblk *kcb)
818 {
819 	kcb->prev_kprobe.kp = kprobe_running();
820 	kcb->prev_kprobe.status = kcb->kprobe_status;
821 	kcb->prev_kprobe.old_flags = kcb->kprobe_old_flags;
822 	kcb->prev_kprobe.saved_flags = kcb->kprobe_saved_flags;
823 }
824 
825 static nokprobe_inline void
826 restore_previous_kprobe(struct kprobe_ctlblk *kcb)
827 {
828 	__this_cpu_write(current_kprobe, kcb->prev_kprobe.kp);
829 	kcb->kprobe_status = kcb->prev_kprobe.status;
830 	kcb->kprobe_old_flags = kcb->prev_kprobe.old_flags;
831 	kcb->kprobe_saved_flags = kcb->prev_kprobe.saved_flags;
832 }
833 
834 static nokprobe_inline void
835 set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
836 		   struct kprobe_ctlblk *kcb)
837 {
838 	__this_cpu_write(current_kprobe, p);
839 	kcb->kprobe_saved_flags = kcb->kprobe_old_flags
840 		= (regs->flags & X86_EFLAGS_IF);
841 }
842 
843 static void kprobe_post_process(struct kprobe *cur, struct pt_regs *regs,
844 			       struct kprobe_ctlblk *kcb)
845 {
846 	/* Restore back the original saved kprobes variables and continue. */
847 	if (kcb->kprobe_status == KPROBE_REENTER) {
848 		/* This will restore both kcb and current_kprobe */
849 		restore_previous_kprobe(kcb);
850 	} else {
851 		/*
852 		 * Always update the kcb status because
853 		 * reset_curent_kprobe() doesn't update kcb.
854 		 */
855 		kcb->kprobe_status = KPROBE_HIT_SSDONE;
856 		if (cur->post_handler)
857 			cur->post_handler(cur, regs, 0);
858 		reset_current_kprobe();
859 	}
860 }
861 NOKPROBE_SYMBOL(kprobe_post_process);
862 
863 static void setup_singlestep(struct kprobe *p, struct pt_regs *regs,
864 			     struct kprobe_ctlblk *kcb, int reenter)
865 {
866 	if (setup_detour_execution(p, regs, reenter))
867 		return;
868 
869 #if !defined(CONFIG_PREEMPTION)
870 	if (p->ainsn.boostable) {
871 		/* Boost up -- we can execute copied instructions directly */
872 		if (!reenter)
873 			reset_current_kprobe();
874 		/*
875 		 * Reentering boosted probe doesn't reset current_kprobe,
876 		 * nor set current_kprobe, because it doesn't use single
877 		 * stepping.
878 		 */
879 		regs->ip = (unsigned long)p->ainsn.insn;
880 		return;
881 	}
882 #endif
883 	if (reenter) {
884 		save_previous_kprobe(kcb);
885 		set_current_kprobe(p, regs, kcb);
886 		kcb->kprobe_status = KPROBE_REENTER;
887 	} else
888 		kcb->kprobe_status = KPROBE_HIT_SS;
889 
890 	if (p->ainsn.emulate_op) {
891 		p->ainsn.emulate_op(p, regs);
892 		kprobe_post_process(p, regs, kcb);
893 		return;
894 	}
895 
896 	/* Disable interrupt, and set ip register on trampoline */
897 	regs->flags &= ~X86_EFLAGS_IF;
898 	regs->ip = (unsigned long)p->ainsn.insn;
899 }
900 NOKPROBE_SYMBOL(setup_singlestep);
901 
902 /*
903  * Called after single-stepping.  p->addr is the address of the
904  * instruction whose first byte has been replaced by the "int3"
905  * instruction.  To avoid the SMP problems that can occur when we
906  * temporarily put back the original opcode to single-step, we
907  * single-stepped a copy of the instruction.  The address of this
908  * copy is p->ainsn.insn. We also doesn't use trap, but "int3" again
909  * right after the copied instruction.
910  * Different from the trap single-step, "int3" single-step can not
911  * handle the instruction which changes the ip register, e.g. jmp,
912  * call, conditional jmp, and the instructions which changes the IF
913  * flags because interrupt must be disabled around the single-stepping.
914  * Such instructions are software emulated, but others are single-stepped
915  * using "int3".
916  *
917  * When the 2nd "int3" handled, the regs->ip and regs->flags needs to
918  * be adjusted, so that we can resume execution on correct code.
919  */
920 static void resume_singlestep(struct kprobe *p, struct pt_regs *regs,
921 			      struct kprobe_ctlblk *kcb)
922 {
923 	unsigned long copy_ip = (unsigned long)p->ainsn.insn;
924 	unsigned long orig_ip = (unsigned long)p->addr;
925 
926 	/* Restore saved interrupt flag and ip register */
927 	regs->flags |= kcb->kprobe_saved_flags;
928 	/* Note that regs->ip is executed int3 so must be a step back */
929 	regs->ip += (orig_ip - copy_ip) - INT3_INSN_SIZE;
930 }
931 NOKPROBE_SYMBOL(resume_singlestep);
932 
933 /*
934  * We have reentered the kprobe_handler(), since another probe was hit while
935  * within the handler. We save the original kprobes variables and just single
936  * step on the instruction of the new probe without calling any user handlers.
937  */
938 static int reenter_kprobe(struct kprobe *p, struct pt_regs *regs,
939 			  struct kprobe_ctlblk *kcb)
940 {
941 	switch (kcb->kprobe_status) {
942 	case KPROBE_HIT_SSDONE:
943 	case KPROBE_HIT_ACTIVE:
944 	case KPROBE_HIT_SS:
945 		kprobes_inc_nmissed_count(p);
946 		setup_singlestep(p, regs, kcb, 1);
947 		break;
948 	case KPROBE_REENTER:
949 		/* A probe has been hit in the codepath leading up to, or just
950 		 * after, single-stepping of a probed instruction. This entire
951 		 * codepath should strictly reside in .kprobes.text section.
952 		 * Raise a BUG or we'll continue in an endless reentering loop
953 		 * and eventually a stack overflow.
954 		 */
955 		pr_err("Unrecoverable kprobe detected.\n");
956 		dump_kprobe(p);
957 		BUG();
958 	default:
959 		/* impossible cases */
960 		WARN_ON(1);
961 		return 0;
962 	}
963 
964 	return 1;
965 }
966 NOKPROBE_SYMBOL(reenter_kprobe);
967 
968 static nokprobe_inline int kprobe_is_ss(struct kprobe_ctlblk *kcb)
969 {
970 	return (kcb->kprobe_status == KPROBE_HIT_SS ||
971 		kcb->kprobe_status == KPROBE_REENTER);
972 }
973 
974 /*
975  * Interrupts are disabled on entry as trap3 is an interrupt gate and they
976  * remain disabled throughout this function.
977  */
978 int kprobe_int3_handler(struct pt_regs *regs)
979 {
980 	kprobe_opcode_t *addr;
981 	struct kprobe *p;
982 	struct kprobe_ctlblk *kcb;
983 
984 	if (user_mode(regs))
985 		return 0;
986 
987 	addr = (kprobe_opcode_t *)(regs->ip - sizeof(kprobe_opcode_t));
988 	/*
989 	 * We don't want to be preempted for the entire duration of kprobe
990 	 * processing. Since int3 and debug trap disables irqs and we clear
991 	 * IF while singlestepping, it must be no preemptible.
992 	 */
993 
994 	kcb = get_kprobe_ctlblk();
995 	p = get_kprobe(addr);
996 
997 	if (p) {
998 		if (kprobe_running()) {
999 			if (reenter_kprobe(p, regs, kcb))
1000 				return 1;
1001 		} else {
1002 			set_current_kprobe(p, regs, kcb);
1003 			kcb->kprobe_status = KPROBE_HIT_ACTIVE;
1004 
1005 			/*
1006 			 * If we have no pre-handler or it returned 0, we
1007 			 * continue with normal processing.  If we have a
1008 			 * pre-handler and it returned non-zero, that means
1009 			 * user handler setup registers to exit to another
1010 			 * instruction, we must skip the single stepping.
1011 			 */
1012 			if (!p->pre_handler || !p->pre_handler(p, regs))
1013 				setup_singlestep(p, regs, kcb, 0);
1014 			else
1015 				reset_current_kprobe();
1016 			return 1;
1017 		}
1018 	} else if (kprobe_is_ss(kcb)) {
1019 		p = kprobe_running();
1020 		if ((unsigned long)p->ainsn.insn < regs->ip &&
1021 		    (unsigned long)p->ainsn.insn + MAX_INSN_SIZE > regs->ip) {
1022 			/* Most provably this is the second int3 for singlestep */
1023 			resume_singlestep(p, regs, kcb);
1024 			kprobe_post_process(p, regs, kcb);
1025 			return 1;
1026 		}
1027 	} /* else: not a kprobe fault; let the kernel handle it */
1028 
1029 	return 0;
1030 }
1031 NOKPROBE_SYMBOL(kprobe_int3_handler);
1032 
1033 int kprobe_fault_handler(struct pt_regs *regs, int trapnr)
1034 {
1035 	struct kprobe *cur = kprobe_running();
1036 	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1037 
1038 	if (unlikely(regs->ip == (unsigned long)cur->ainsn.insn)) {
1039 		/* This must happen on single-stepping */
1040 		WARN_ON(kcb->kprobe_status != KPROBE_HIT_SS &&
1041 			kcb->kprobe_status != KPROBE_REENTER);
1042 		/*
1043 		 * We are here because the instruction being single
1044 		 * stepped caused a page fault. We reset the current
1045 		 * kprobe and the ip points back to the probe address
1046 		 * and allow the page fault handler to continue as a
1047 		 * normal page fault.
1048 		 */
1049 		regs->ip = (unsigned long)cur->addr;
1050 
1051 		/*
1052 		 * If the IF flag was set before the kprobe hit,
1053 		 * don't touch it:
1054 		 */
1055 		regs->flags |= kcb->kprobe_old_flags;
1056 
1057 		if (kcb->kprobe_status == KPROBE_REENTER)
1058 			restore_previous_kprobe(kcb);
1059 		else
1060 			reset_current_kprobe();
1061 	}
1062 
1063 	return 0;
1064 }
1065 NOKPROBE_SYMBOL(kprobe_fault_handler);
1066 
1067 int __init arch_populate_kprobe_blacklist(void)
1068 {
1069 	return kprobe_add_area_blacklist((unsigned long)__entry_text_start,
1070 					 (unsigned long)__entry_text_end);
1071 }
1072 
1073 int __init arch_init_kprobes(void)
1074 {
1075 	return 0;
1076 }
1077 
1078 int arch_trampoline_kprobe(struct kprobe *p)
1079 {
1080 	return 0;
1081 }
1082