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/moduleloader.h>
44 #include <linux/objtool.h>
45 #include <linux/vmalloc.h>
46 #include <linux/pgtable.h>
47
48 #include <asm/text-patching.h>
49 #include <asm/cacheflush.h>
50 #include <asm/desc.h>
51 #include <linux/uaccess.h>
52 #include <asm/alternative.h>
53 #include <asm/insn.h>
54 #include <asm/debugreg.h>
55 #include <asm/set_memory.h>
56
57 #include "common.h"
58
59 DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
60 DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
61
62 #define stack_addr(regs) ((unsigned long *)regs->sp)
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
__synthesize_relative_insn(void * dest,void * from,void * to,u8 op)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'.*/
synthesize_reljump(void * dest,void * from,void * 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'.*/
synthesize_relcall(void * dest,void * from,void * 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 */
can_boost(struct insn * insn,void * addr)140 int can_boost(struct insn *insn, void *addr)
141 {
142 kprobe_opcode_t opcode;
143 insn_byte_t prefix;
144 int i;
145
146 if (search_exception_tables((unsigned long)addr))
147 return 0; /* Page fault may occur on this address. */
148
149 /* 2nd-byte opcode */
150 if (insn->opcode.nbytes == 2)
151 return test_bit(insn->opcode.bytes[1],
152 (unsigned long *)twobyte_is_boostable);
153
154 if (insn->opcode.nbytes != 1)
155 return 0;
156
157 for_each_insn_prefix(insn, i, prefix) {
158 insn_attr_t attr;
159
160 attr = inat_get_opcode_attribute(prefix);
161 /* Can't boost Address-size override prefix and CS override prefix */
162 if (prefix == 0x2e || inat_is_address_size_prefix(attr))
163 return 0;
164 }
165
166 opcode = insn->opcode.bytes[0];
167
168 switch (opcode) {
169 case 0x62: /* bound */
170 case 0x70 ... 0x7f: /* Conditional jumps */
171 case 0x9a: /* Call far */
172 case 0xc0 ... 0xc1: /* Grp2 */
173 case 0xcc ... 0xce: /* software exceptions */
174 case 0xd0 ... 0xd3: /* Grp2 */
175 case 0xd6: /* (UD) */
176 case 0xd8 ... 0xdf: /* ESC */
177 case 0xe0 ... 0xe3: /* LOOP*, JCXZ */
178 case 0xe8 ... 0xe9: /* near Call, JMP */
179 case 0xeb: /* Short JMP */
180 case 0xf0 ... 0xf4: /* LOCK/REP, HLT */
181 case 0xf6 ... 0xf7: /* Grp3 */
182 case 0xfe: /* Grp4 */
183 /* ... are not boostable */
184 return 0;
185 case 0xff: /* Grp5 */
186 /* Only indirect jmp is boostable */
187 return X86_MODRM_REG(insn->modrm.bytes[0]) == 4;
188 default:
189 return 1;
190 }
191 }
192
193 static unsigned long
__recover_probed_insn(kprobe_opcode_t * buf,unsigned long addr)194 __recover_probed_insn(kprobe_opcode_t *buf, unsigned long addr)
195 {
196 struct kprobe *kp;
197 unsigned long faddr;
198
199 kp = get_kprobe((void *)addr);
200 faddr = ftrace_location(addr);
201 /*
202 * Addresses inside the ftrace location are refused by
203 * arch_check_ftrace_location(). Something went terribly wrong
204 * if such an address is checked here.
205 */
206 if (WARN_ON(faddr && faddr != addr))
207 return 0UL;
208 /*
209 * Use the current code if it is not modified by Kprobe
210 * and it cannot be modified by ftrace.
211 */
212 if (!kp && !faddr)
213 return addr;
214
215 /*
216 * Basically, kp->ainsn.insn has an original instruction.
217 * However, RIP-relative instruction can not do single-stepping
218 * at different place, __copy_instruction() tweaks the displacement of
219 * that instruction. In that case, we can't recover the instruction
220 * from the kp->ainsn.insn.
221 *
222 * On the other hand, in case on normal Kprobe, kp->opcode has a copy
223 * of the first byte of the probed instruction, which is overwritten
224 * by int3. And the instruction at kp->addr is not modified by kprobes
225 * except for the first byte, we can recover the original instruction
226 * from it and kp->opcode.
227 *
228 * In case of Kprobes using ftrace, we do not have a copy of
229 * the original instruction. In fact, the ftrace location might
230 * be modified at anytime and even could be in an inconsistent state.
231 * Fortunately, we know that the original code is the ideal 5-byte
232 * long NOP.
233 */
234 if (copy_from_kernel_nofault(buf, (void *)addr,
235 MAX_INSN_SIZE * sizeof(kprobe_opcode_t)))
236 return 0UL;
237
238 if (faddr)
239 memcpy(buf, x86_nops[5], 5);
240 else
241 buf[0] = kp->opcode;
242 return (unsigned long)buf;
243 }
244
245 /*
246 * Recover the probed instruction at addr for further analysis.
247 * Caller must lock kprobes by kprobe_mutex, or disable preemption
248 * for preventing to release referencing kprobes.
249 * Returns zero if the instruction can not get recovered (or access failed).
250 */
recover_probed_instruction(kprobe_opcode_t * buf,unsigned long addr)251 unsigned long recover_probed_instruction(kprobe_opcode_t *buf, unsigned long addr)
252 {
253 unsigned long __addr;
254
255 __addr = __recover_optprobed_insn(buf, addr);
256 if (__addr != addr)
257 return __addr;
258
259 return __recover_probed_insn(buf, addr);
260 }
261
262 /* Check if paddr is at an instruction boundary */
can_probe(unsigned long paddr)263 static int can_probe(unsigned long paddr)
264 {
265 unsigned long addr, __addr, offset = 0;
266 struct insn insn;
267 kprobe_opcode_t buf[MAX_INSN_SIZE];
268
269 if (!kallsyms_lookup_size_offset(paddr, NULL, &offset))
270 return 0;
271
272 /* Decode instructions */
273 addr = paddr - offset;
274 while (addr < paddr) {
275 int ret;
276
277 /*
278 * Check if the instruction has been modified by another
279 * kprobe, in which case we replace the breakpoint by the
280 * original instruction in our buffer.
281 * Also, jump optimization will change the breakpoint to
282 * relative-jump. Since the relative-jump itself is
283 * normally used, we just go through if there is no kprobe.
284 */
285 __addr = recover_probed_instruction(buf, addr);
286 if (!__addr)
287 return 0;
288
289 ret = insn_decode_kernel(&insn, (void *)__addr);
290 if (ret < 0)
291 return 0;
292
293 #ifdef CONFIG_KGDB
294 /*
295 * If there is a dynamically installed kgdb sw breakpoint,
296 * this function should not be probed.
297 */
298 if (insn.opcode.bytes[0] == INT3_INSN_OPCODE &&
299 kgdb_has_hit_break(addr))
300 return 0;
301 #endif
302 addr += insn.length;
303 }
304
305 return (addr == paddr);
306 }
307
308 /*
309 * Copy an instruction with recovering modified instruction by kprobes
310 * and adjust the displacement if the instruction uses the %rip-relative
311 * addressing mode. Note that since @real will be the final place of copied
312 * instruction, displacement must be adjust by @real, not @dest.
313 * This returns the length of copied instruction, or 0 if it has an error.
314 */
__copy_instruction(u8 * dest,u8 * src,u8 * real,struct insn * insn)315 int __copy_instruction(u8 *dest, u8 *src, u8 *real, struct insn *insn)
316 {
317 kprobe_opcode_t buf[MAX_INSN_SIZE];
318 unsigned long recovered_insn = recover_probed_instruction(buf, (unsigned long)src);
319 int ret;
320
321 if (!recovered_insn || !insn)
322 return 0;
323
324 /* This can access kernel text if given address is not recovered */
325 if (copy_from_kernel_nofault(dest, (void *)recovered_insn,
326 MAX_INSN_SIZE))
327 return 0;
328
329 ret = insn_decode_kernel(insn, dest);
330 if (ret < 0)
331 return 0;
332
333 /* We can not probe force emulate prefixed instruction */
334 if (insn_has_emulate_prefix(insn))
335 return 0;
336
337 /* Another subsystem puts a breakpoint, failed to recover */
338 if (insn->opcode.bytes[0] == INT3_INSN_OPCODE)
339 return 0;
340
341 /* We should not singlestep on the exception masking instructions */
342 if (insn_masking_exception(insn))
343 return 0;
344
345 #ifdef CONFIG_X86_64
346 /* Only x86_64 has RIP relative instructions */
347 if (insn_rip_relative(insn)) {
348 s64 newdisp;
349 u8 *disp;
350 /*
351 * The copied instruction uses the %rip-relative addressing
352 * mode. Adjust the displacement for the difference between
353 * the original location of this instruction and the location
354 * of the copy that will actually be run. The tricky bit here
355 * is making sure that the sign extension happens correctly in
356 * this calculation, since we need a signed 32-bit result to
357 * be sign-extended to 64 bits when it's added to the %rip
358 * value and yield the same 64-bit result that the sign-
359 * extension of the original signed 32-bit displacement would
360 * have given.
361 */
362 newdisp = (u8 *) src + (s64) insn->displacement.value
363 - (u8 *) real;
364 if ((s64) (s32) newdisp != newdisp) {
365 pr_err("Kprobes error: new displacement does not fit into s32 (%llx)\n", newdisp);
366 return 0;
367 }
368 disp = (u8 *) dest + insn_offset_displacement(insn);
369 *(s32 *) disp = (s32) newdisp;
370 }
371 #endif
372 return insn->length;
373 }
374
375 /* Prepare reljump or int3 right after instruction */
prepare_singlestep(kprobe_opcode_t * buf,struct kprobe * p,struct insn * insn)376 static int prepare_singlestep(kprobe_opcode_t *buf, struct kprobe *p,
377 struct insn *insn)
378 {
379 int len = insn->length;
380
381 if (!IS_ENABLED(CONFIG_PREEMPTION) &&
382 !p->post_handler && can_boost(insn, p->addr) &&
383 MAX_INSN_SIZE - len >= JMP32_INSN_SIZE) {
384 /*
385 * These instructions can be executed directly if it
386 * jumps back to correct address.
387 */
388 synthesize_reljump(buf + len, p->ainsn.insn + len,
389 p->addr + insn->length);
390 len += JMP32_INSN_SIZE;
391 p->ainsn.boostable = 1;
392 } else {
393 /* Otherwise, put an int3 for trapping singlestep */
394 if (MAX_INSN_SIZE - len < INT3_INSN_SIZE)
395 return -ENOSPC;
396
397 buf[len] = INT3_INSN_OPCODE;
398 len += INT3_INSN_SIZE;
399 }
400
401 return len;
402 }
403
404 /* Make page to RO mode when allocate it */
alloc_insn_page(void)405 void *alloc_insn_page(void)
406 {
407 void *page;
408
409 page = module_alloc(PAGE_SIZE);
410 if (!page)
411 return NULL;
412
413 set_vm_flush_reset_perms(page);
414 /*
415 * First make the page read-only, and only then make it executable to
416 * prevent it from being W+X in between.
417 */
418 set_memory_ro((unsigned long)page, 1);
419
420 /*
421 * TODO: Once additional kernel code protection mechanisms are set, ensure
422 * that the page was not maliciously altered and it is still zeroed.
423 */
424 set_memory_x((unsigned long)page, 1);
425
426 return page;
427 }
428
429 /* Kprobe x86 instruction emulation - only regs->ip or IF flag modifiers */
430
kprobe_emulate_ifmodifiers(struct kprobe * p,struct pt_regs * regs)431 static void kprobe_emulate_ifmodifiers(struct kprobe *p, struct pt_regs *regs)
432 {
433 switch (p->ainsn.opcode) {
434 case 0xfa: /* cli */
435 regs->flags &= ~(X86_EFLAGS_IF);
436 break;
437 case 0xfb: /* sti */
438 regs->flags |= X86_EFLAGS_IF;
439 break;
440 case 0x9c: /* pushf */
441 int3_emulate_push(regs, regs->flags);
442 break;
443 case 0x9d: /* popf */
444 regs->flags = int3_emulate_pop(regs);
445 break;
446 }
447 regs->ip = regs->ip - INT3_INSN_SIZE + p->ainsn.size;
448 }
449 NOKPROBE_SYMBOL(kprobe_emulate_ifmodifiers);
450
kprobe_emulate_ret(struct kprobe * p,struct pt_regs * regs)451 static void kprobe_emulate_ret(struct kprobe *p, struct pt_regs *regs)
452 {
453 int3_emulate_ret(regs);
454 }
455 NOKPROBE_SYMBOL(kprobe_emulate_ret);
456
kprobe_emulate_call(struct kprobe * p,struct pt_regs * regs)457 static void kprobe_emulate_call(struct kprobe *p, struct pt_regs *regs)
458 {
459 unsigned long func = regs->ip - INT3_INSN_SIZE + p->ainsn.size;
460
461 func += p->ainsn.rel32;
462 int3_emulate_call(regs, func);
463 }
464 NOKPROBE_SYMBOL(kprobe_emulate_call);
465
466 static nokprobe_inline
__kprobe_emulate_jmp(struct kprobe * p,struct pt_regs * regs,bool cond)467 void __kprobe_emulate_jmp(struct kprobe *p, struct pt_regs *regs, bool cond)
468 {
469 unsigned long ip = regs->ip - INT3_INSN_SIZE + p->ainsn.size;
470
471 if (cond)
472 ip += p->ainsn.rel32;
473 int3_emulate_jmp(regs, ip);
474 }
475
kprobe_emulate_jmp(struct kprobe * p,struct pt_regs * regs)476 static void kprobe_emulate_jmp(struct kprobe *p, struct pt_regs *regs)
477 {
478 __kprobe_emulate_jmp(p, regs, true);
479 }
480 NOKPROBE_SYMBOL(kprobe_emulate_jmp);
481
482 static const unsigned long jcc_mask[6] = {
483 [0] = X86_EFLAGS_OF,
484 [1] = X86_EFLAGS_CF,
485 [2] = X86_EFLAGS_ZF,
486 [3] = X86_EFLAGS_CF | X86_EFLAGS_ZF,
487 [4] = X86_EFLAGS_SF,
488 [5] = X86_EFLAGS_PF,
489 };
490
kprobe_emulate_jcc(struct kprobe * p,struct pt_regs * regs)491 static void kprobe_emulate_jcc(struct kprobe *p, struct pt_regs *regs)
492 {
493 bool invert = p->ainsn.jcc.type & 1;
494 bool match;
495
496 if (p->ainsn.jcc.type < 0xc) {
497 match = regs->flags & jcc_mask[p->ainsn.jcc.type >> 1];
498 } else {
499 match = ((regs->flags & X86_EFLAGS_SF) >> X86_EFLAGS_SF_BIT) ^
500 ((regs->flags & X86_EFLAGS_OF) >> X86_EFLAGS_OF_BIT);
501 if (p->ainsn.jcc.type >= 0xe)
502 match = match || (regs->flags & X86_EFLAGS_ZF);
503 }
504 __kprobe_emulate_jmp(p, regs, (match && !invert) || (!match && invert));
505 }
506 NOKPROBE_SYMBOL(kprobe_emulate_jcc);
507
kprobe_emulate_loop(struct kprobe * p,struct pt_regs * regs)508 static void kprobe_emulate_loop(struct kprobe *p, struct pt_regs *regs)
509 {
510 bool match;
511
512 if (p->ainsn.loop.type != 3) { /* LOOP* */
513 if (p->ainsn.loop.asize == 32)
514 match = ((*(u32 *)®s->cx)--) != 0;
515 #ifdef CONFIG_X86_64
516 else if (p->ainsn.loop.asize == 64)
517 match = ((*(u64 *)®s->cx)--) != 0;
518 #endif
519 else
520 match = ((*(u16 *)®s->cx)--) != 0;
521 } else { /* JCXZ */
522 if (p->ainsn.loop.asize == 32)
523 match = *(u32 *)(®s->cx) == 0;
524 #ifdef CONFIG_X86_64
525 else if (p->ainsn.loop.asize == 64)
526 match = *(u64 *)(®s->cx) == 0;
527 #endif
528 else
529 match = *(u16 *)(®s->cx) == 0;
530 }
531
532 if (p->ainsn.loop.type == 0) /* LOOPNE */
533 match = match && !(regs->flags & X86_EFLAGS_ZF);
534 else if (p->ainsn.loop.type == 1) /* LOOPE */
535 match = match && (regs->flags & X86_EFLAGS_ZF);
536
537 __kprobe_emulate_jmp(p, regs, match);
538 }
539 NOKPROBE_SYMBOL(kprobe_emulate_loop);
540
541 static const int addrmode_regoffs[] = {
542 offsetof(struct pt_regs, ax),
543 offsetof(struct pt_regs, cx),
544 offsetof(struct pt_regs, dx),
545 offsetof(struct pt_regs, bx),
546 offsetof(struct pt_regs, sp),
547 offsetof(struct pt_regs, bp),
548 offsetof(struct pt_regs, si),
549 offsetof(struct pt_regs, di),
550 #ifdef CONFIG_X86_64
551 offsetof(struct pt_regs, r8),
552 offsetof(struct pt_regs, r9),
553 offsetof(struct pt_regs, r10),
554 offsetof(struct pt_regs, r11),
555 offsetof(struct pt_regs, r12),
556 offsetof(struct pt_regs, r13),
557 offsetof(struct pt_regs, r14),
558 offsetof(struct pt_regs, r15),
559 #endif
560 };
561
kprobe_emulate_call_indirect(struct kprobe * p,struct pt_regs * regs)562 static void kprobe_emulate_call_indirect(struct kprobe *p, struct pt_regs *regs)
563 {
564 unsigned long offs = addrmode_regoffs[p->ainsn.indirect.reg];
565
566 int3_emulate_push(regs, regs->ip - INT3_INSN_SIZE + p->ainsn.size);
567 int3_emulate_jmp(regs, regs_get_register(regs, offs));
568 }
569 NOKPROBE_SYMBOL(kprobe_emulate_call_indirect);
570
kprobe_emulate_jmp_indirect(struct kprobe * p,struct pt_regs * regs)571 static void kprobe_emulate_jmp_indirect(struct kprobe *p, struct pt_regs *regs)
572 {
573 unsigned long offs = addrmode_regoffs[p->ainsn.indirect.reg];
574
575 int3_emulate_jmp(regs, regs_get_register(regs, offs));
576 }
577 NOKPROBE_SYMBOL(kprobe_emulate_jmp_indirect);
578
prepare_emulation(struct kprobe * p,struct insn * insn)579 static int prepare_emulation(struct kprobe *p, struct insn *insn)
580 {
581 insn_byte_t opcode = insn->opcode.bytes[0];
582
583 switch (opcode) {
584 case 0xfa: /* cli */
585 case 0xfb: /* sti */
586 case 0x9c: /* pushfl */
587 case 0x9d: /* popf/popfd */
588 /*
589 * IF modifiers must be emulated since it will enable interrupt while
590 * int3 single stepping.
591 */
592 p->ainsn.emulate_op = kprobe_emulate_ifmodifiers;
593 p->ainsn.opcode = opcode;
594 break;
595 case 0xc2: /* ret/lret */
596 case 0xc3:
597 case 0xca:
598 case 0xcb:
599 p->ainsn.emulate_op = kprobe_emulate_ret;
600 break;
601 case 0x9a: /* far call absolute -- segment is not supported */
602 case 0xea: /* far jmp absolute -- segment is not supported */
603 case 0xcc: /* int3 */
604 case 0xcf: /* iret -- in-kernel IRET is not supported */
605 return -EOPNOTSUPP;
606 break;
607 case 0xe8: /* near call relative */
608 p->ainsn.emulate_op = kprobe_emulate_call;
609 if (insn->immediate.nbytes == 2)
610 p->ainsn.rel32 = *(s16 *)&insn->immediate.value;
611 else
612 p->ainsn.rel32 = *(s32 *)&insn->immediate.value;
613 break;
614 case 0xeb: /* short jump relative */
615 case 0xe9: /* near jump relative */
616 p->ainsn.emulate_op = kprobe_emulate_jmp;
617 if (insn->immediate.nbytes == 1)
618 p->ainsn.rel32 = *(s8 *)&insn->immediate.value;
619 else if (insn->immediate.nbytes == 2)
620 p->ainsn.rel32 = *(s16 *)&insn->immediate.value;
621 else
622 p->ainsn.rel32 = *(s32 *)&insn->immediate.value;
623 break;
624 case 0x70 ... 0x7f:
625 /* 1 byte conditional jump */
626 p->ainsn.emulate_op = kprobe_emulate_jcc;
627 p->ainsn.jcc.type = opcode & 0xf;
628 p->ainsn.rel32 = *(char *)insn->immediate.bytes;
629 break;
630 case 0x0f:
631 opcode = insn->opcode.bytes[1];
632 if ((opcode & 0xf0) == 0x80) {
633 /* 2 bytes Conditional Jump */
634 p->ainsn.emulate_op = kprobe_emulate_jcc;
635 p->ainsn.jcc.type = opcode & 0xf;
636 if (insn->immediate.nbytes == 2)
637 p->ainsn.rel32 = *(s16 *)&insn->immediate.value;
638 else
639 p->ainsn.rel32 = *(s32 *)&insn->immediate.value;
640 } else if (opcode == 0x01 &&
641 X86_MODRM_REG(insn->modrm.bytes[0]) == 0 &&
642 X86_MODRM_MOD(insn->modrm.bytes[0]) == 3) {
643 /* VM extensions - not supported */
644 return -EOPNOTSUPP;
645 }
646 break;
647 case 0xe0: /* Loop NZ */
648 case 0xe1: /* Loop */
649 case 0xe2: /* Loop */
650 case 0xe3: /* J*CXZ */
651 p->ainsn.emulate_op = kprobe_emulate_loop;
652 p->ainsn.loop.type = opcode & 0x3;
653 p->ainsn.loop.asize = insn->addr_bytes * 8;
654 p->ainsn.rel32 = *(s8 *)&insn->immediate.value;
655 break;
656 case 0xff:
657 /*
658 * Since the 0xff is an extended group opcode, the instruction
659 * is determined by the MOD/RM byte.
660 */
661 opcode = insn->modrm.bytes[0];
662 if ((opcode & 0x30) == 0x10) {
663 if ((opcode & 0x8) == 0x8)
664 return -EOPNOTSUPP; /* far call */
665 /* call absolute, indirect */
666 p->ainsn.emulate_op = kprobe_emulate_call_indirect;
667 } else if ((opcode & 0x30) == 0x20) {
668 if ((opcode & 0x8) == 0x8)
669 return -EOPNOTSUPP; /* far jmp */
670 /* jmp near absolute indirect */
671 p->ainsn.emulate_op = kprobe_emulate_jmp_indirect;
672 } else
673 break;
674
675 if (insn->addr_bytes != sizeof(unsigned long))
676 return -EOPNOTSUPP; /* Don't support different size */
677 if (X86_MODRM_MOD(opcode) != 3)
678 return -EOPNOTSUPP; /* TODO: support memory addressing */
679
680 p->ainsn.indirect.reg = X86_MODRM_RM(opcode);
681 #ifdef CONFIG_X86_64
682 if (X86_REX_B(insn->rex_prefix.value))
683 p->ainsn.indirect.reg += 8;
684 #endif
685 break;
686 default:
687 break;
688 }
689 p->ainsn.size = insn->length;
690
691 return 0;
692 }
693
arch_copy_kprobe(struct kprobe * p)694 static int arch_copy_kprobe(struct kprobe *p)
695 {
696 struct insn insn;
697 kprobe_opcode_t buf[MAX_INSN_SIZE];
698 int ret, len;
699
700 /* Copy an instruction with recovering if other optprobe modifies it.*/
701 len = __copy_instruction(buf, p->addr, p->ainsn.insn, &insn);
702 if (!len)
703 return -EINVAL;
704
705 /* Analyze the opcode and setup emulate functions */
706 ret = prepare_emulation(p, &insn);
707 if (ret < 0)
708 return ret;
709
710 /* Add int3 for single-step or booster jmp */
711 len = prepare_singlestep(buf, p, &insn);
712 if (len < 0)
713 return len;
714
715 /* Also, displacement change doesn't affect the first byte */
716 p->opcode = buf[0];
717
718 p->ainsn.tp_len = len;
719 perf_event_text_poke(p->ainsn.insn, NULL, 0, buf, len);
720
721 /* OK, write back the instruction(s) into ROX insn buffer */
722 text_poke(p->ainsn.insn, buf, len);
723
724 return 0;
725 }
726
arch_prepare_kprobe(struct kprobe * p)727 int arch_prepare_kprobe(struct kprobe *p)
728 {
729 int ret;
730
731 if (alternatives_text_reserved(p->addr, p->addr))
732 return -EINVAL;
733
734 if (!can_probe((unsigned long)p->addr))
735 return -EILSEQ;
736
737 memset(&p->ainsn, 0, sizeof(p->ainsn));
738
739 /* insn: must be on special executable page on x86. */
740 p->ainsn.insn = get_insn_slot();
741 if (!p->ainsn.insn)
742 return -ENOMEM;
743
744 ret = arch_copy_kprobe(p);
745 if (ret) {
746 free_insn_slot(p->ainsn.insn, 0);
747 p->ainsn.insn = NULL;
748 }
749
750 return ret;
751 }
752
arch_arm_kprobe(struct kprobe * p)753 void arch_arm_kprobe(struct kprobe *p)
754 {
755 u8 int3 = INT3_INSN_OPCODE;
756
757 text_poke(p->addr, &int3, 1);
758 text_poke_sync();
759 perf_event_text_poke(p->addr, &p->opcode, 1, &int3, 1);
760 }
761
arch_disarm_kprobe(struct kprobe * p)762 void arch_disarm_kprobe(struct kprobe *p)
763 {
764 u8 int3 = INT3_INSN_OPCODE;
765
766 perf_event_text_poke(p->addr, &int3, 1, &p->opcode, 1);
767 text_poke(p->addr, &p->opcode, 1);
768 text_poke_sync();
769 }
770
arch_remove_kprobe(struct kprobe * p)771 void arch_remove_kprobe(struct kprobe *p)
772 {
773 if (p->ainsn.insn) {
774 /* Record the perf event before freeing the slot */
775 perf_event_text_poke(p->ainsn.insn, p->ainsn.insn,
776 p->ainsn.tp_len, NULL, 0);
777 free_insn_slot(p->ainsn.insn, p->ainsn.boostable);
778 p->ainsn.insn = NULL;
779 }
780 }
781
782 static nokprobe_inline void
save_previous_kprobe(struct kprobe_ctlblk * kcb)783 save_previous_kprobe(struct kprobe_ctlblk *kcb)
784 {
785 kcb->prev_kprobe.kp = kprobe_running();
786 kcb->prev_kprobe.status = kcb->kprobe_status;
787 kcb->prev_kprobe.old_flags = kcb->kprobe_old_flags;
788 kcb->prev_kprobe.saved_flags = kcb->kprobe_saved_flags;
789 }
790
791 static nokprobe_inline void
restore_previous_kprobe(struct kprobe_ctlblk * kcb)792 restore_previous_kprobe(struct kprobe_ctlblk *kcb)
793 {
794 __this_cpu_write(current_kprobe, kcb->prev_kprobe.kp);
795 kcb->kprobe_status = kcb->prev_kprobe.status;
796 kcb->kprobe_old_flags = kcb->prev_kprobe.old_flags;
797 kcb->kprobe_saved_flags = kcb->prev_kprobe.saved_flags;
798 }
799
800 static nokprobe_inline void
set_current_kprobe(struct kprobe * p,struct pt_regs * regs,struct kprobe_ctlblk * kcb)801 set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
802 struct kprobe_ctlblk *kcb)
803 {
804 __this_cpu_write(current_kprobe, p);
805 kcb->kprobe_saved_flags = kcb->kprobe_old_flags
806 = (regs->flags & X86_EFLAGS_IF);
807 }
808
arch_prepare_kretprobe(struct kretprobe_instance * ri,struct pt_regs * regs)809 void arch_prepare_kretprobe(struct kretprobe_instance *ri, struct pt_regs *regs)
810 {
811 unsigned long *sara = stack_addr(regs);
812
813 ri->ret_addr = (kprobe_opcode_t *) *sara;
814 ri->fp = sara;
815
816 /* Replace the return addr with trampoline addr */
817 *sara = (unsigned long) &kretprobe_trampoline;
818 }
819 NOKPROBE_SYMBOL(arch_prepare_kretprobe);
820
kprobe_post_process(struct kprobe * cur,struct pt_regs * regs,struct kprobe_ctlblk * kcb)821 static void kprobe_post_process(struct kprobe *cur, struct pt_regs *regs,
822 struct kprobe_ctlblk *kcb)
823 {
824 /* Restore back the original saved kprobes variables and continue. */
825 if (kcb->kprobe_status == KPROBE_REENTER) {
826 /* This will restore both kcb and current_kprobe */
827 restore_previous_kprobe(kcb);
828 } else {
829 /*
830 * Always update the kcb status because
831 * reset_curent_kprobe() doesn't update kcb.
832 */
833 kcb->kprobe_status = KPROBE_HIT_SSDONE;
834 if (cur->post_handler)
835 cur->post_handler(cur, regs, 0);
836 reset_current_kprobe();
837 }
838 }
839 NOKPROBE_SYMBOL(kprobe_post_process);
840
setup_singlestep(struct kprobe * p,struct pt_regs * regs,struct kprobe_ctlblk * kcb,int reenter)841 static void setup_singlestep(struct kprobe *p, struct pt_regs *regs,
842 struct kprobe_ctlblk *kcb, int reenter)
843 {
844 if (setup_detour_execution(p, regs, reenter))
845 return;
846
847 #if !defined(CONFIG_PREEMPTION)
848 if (p->ainsn.boostable) {
849 /* Boost up -- we can execute copied instructions directly */
850 if (!reenter)
851 reset_current_kprobe();
852 /*
853 * Reentering boosted probe doesn't reset current_kprobe,
854 * nor set current_kprobe, because it doesn't use single
855 * stepping.
856 */
857 regs->ip = (unsigned long)p->ainsn.insn;
858 return;
859 }
860 #endif
861 if (reenter) {
862 save_previous_kprobe(kcb);
863 set_current_kprobe(p, regs, kcb);
864 kcb->kprobe_status = KPROBE_REENTER;
865 } else
866 kcb->kprobe_status = KPROBE_HIT_SS;
867
868 if (p->ainsn.emulate_op) {
869 p->ainsn.emulate_op(p, regs);
870 kprobe_post_process(p, regs, kcb);
871 return;
872 }
873
874 /* Disable interrupt, and set ip register on trampoline */
875 regs->flags &= ~X86_EFLAGS_IF;
876 regs->ip = (unsigned long)p->ainsn.insn;
877 }
878 NOKPROBE_SYMBOL(setup_singlestep);
879
880 /*
881 * Called after single-stepping. p->addr is the address of the
882 * instruction whose first byte has been replaced by the "int3"
883 * instruction. To avoid the SMP problems that can occur when we
884 * temporarily put back the original opcode to single-step, we
885 * single-stepped a copy of the instruction. The address of this
886 * copy is p->ainsn.insn. We also doesn't use trap, but "int3" again
887 * right after the copied instruction.
888 * Different from the trap single-step, "int3" single-step can not
889 * handle the instruction which changes the ip register, e.g. jmp,
890 * call, conditional jmp, and the instructions which changes the IF
891 * flags because interrupt must be disabled around the single-stepping.
892 * Such instructions are software emulated, but others are single-stepped
893 * using "int3".
894 *
895 * When the 2nd "int3" handled, the regs->ip and regs->flags needs to
896 * be adjusted, so that we can resume execution on correct code.
897 */
resume_singlestep(struct kprobe * p,struct pt_regs * regs,struct kprobe_ctlblk * kcb)898 static void resume_singlestep(struct kprobe *p, struct pt_regs *regs,
899 struct kprobe_ctlblk *kcb)
900 {
901 unsigned long copy_ip = (unsigned long)p->ainsn.insn;
902 unsigned long orig_ip = (unsigned long)p->addr;
903
904 /* Restore saved interrupt flag and ip register */
905 regs->flags |= kcb->kprobe_saved_flags;
906 /* Note that regs->ip is executed int3 so must be a step back */
907 regs->ip += (orig_ip - copy_ip) - INT3_INSN_SIZE;
908 }
909 NOKPROBE_SYMBOL(resume_singlestep);
910
911 /*
912 * We have reentered the kprobe_handler(), since another probe was hit while
913 * within the handler. We save the original kprobes variables and just single
914 * step on the instruction of the new probe without calling any user handlers.
915 */
reenter_kprobe(struct kprobe * p,struct pt_regs * regs,struct kprobe_ctlblk * kcb)916 static int reenter_kprobe(struct kprobe *p, struct pt_regs *regs,
917 struct kprobe_ctlblk *kcb)
918 {
919 switch (kcb->kprobe_status) {
920 case KPROBE_HIT_SSDONE:
921 case KPROBE_HIT_ACTIVE:
922 case KPROBE_HIT_SS:
923 kprobes_inc_nmissed_count(p);
924 setup_singlestep(p, regs, kcb, 1);
925 break;
926 case KPROBE_REENTER:
927 /* A probe has been hit in the codepath leading up to, or just
928 * after, single-stepping of a probed instruction. This entire
929 * codepath should strictly reside in .kprobes.text section.
930 * Raise a BUG or we'll continue in an endless reentering loop
931 * and eventually a stack overflow.
932 */
933 pr_err("Unrecoverable kprobe detected.\n");
934 dump_kprobe(p);
935 BUG();
936 default:
937 /* impossible cases */
938 WARN_ON(1);
939 return 0;
940 }
941
942 return 1;
943 }
944 NOKPROBE_SYMBOL(reenter_kprobe);
945
kprobe_is_ss(struct kprobe_ctlblk * kcb)946 static nokprobe_inline int kprobe_is_ss(struct kprobe_ctlblk *kcb)
947 {
948 return (kcb->kprobe_status == KPROBE_HIT_SS ||
949 kcb->kprobe_status == KPROBE_REENTER);
950 }
951
952 /*
953 * Interrupts are disabled on entry as trap3 is an interrupt gate and they
954 * remain disabled throughout this function.
955 */
kprobe_int3_handler(struct pt_regs * regs)956 int kprobe_int3_handler(struct pt_regs *regs)
957 {
958 kprobe_opcode_t *addr;
959 struct kprobe *p;
960 struct kprobe_ctlblk *kcb;
961
962 if (user_mode(regs))
963 return 0;
964
965 addr = (kprobe_opcode_t *)(regs->ip - sizeof(kprobe_opcode_t));
966 /*
967 * We don't want to be preempted for the entire duration of kprobe
968 * processing. Since int3 and debug trap disables irqs and we clear
969 * IF while singlestepping, it must be no preemptible.
970 */
971
972 kcb = get_kprobe_ctlblk();
973 p = get_kprobe(addr);
974
975 if (p) {
976 if (kprobe_running()) {
977 if (reenter_kprobe(p, regs, kcb))
978 return 1;
979 } else {
980 set_current_kprobe(p, regs, kcb);
981 kcb->kprobe_status = KPROBE_HIT_ACTIVE;
982
983 /*
984 * If we have no pre-handler or it returned 0, we
985 * continue with normal processing. If we have a
986 * pre-handler and it returned non-zero, that means
987 * user handler setup registers to exit to another
988 * instruction, we must skip the single stepping.
989 */
990 if (!p->pre_handler || !p->pre_handler(p, regs))
991 setup_singlestep(p, regs, kcb, 0);
992 else
993 reset_current_kprobe();
994 return 1;
995 }
996 } else if (kprobe_is_ss(kcb)) {
997 p = kprobe_running();
998 if ((unsigned long)p->ainsn.insn < regs->ip &&
999 (unsigned long)p->ainsn.insn + MAX_INSN_SIZE > regs->ip) {
1000 /* Most provably this is the second int3 for singlestep */
1001 resume_singlestep(p, regs, kcb);
1002 kprobe_post_process(p, regs, kcb);
1003 return 1;
1004 }
1005 }
1006
1007 if (*addr != INT3_INSN_OPCODE) {
1008 /*
1009 * The breakpoint instruction was removed right
1010 * after we hit it. Another cpu has removed
1011 * either a probepoint or a debugger breakpoint
1012 * at this address. In either case, no further
1013 * handling of this interrupt is appropriate.
1014 * Back up over the (now missing) int3 and run
1015 * the original instruction.
1016 */
1017 regs->ip = (unsigned long)addr;
1018 return 1;
1019 } /* else: not a kprobe fault; let the kernel handle it */
1020
1021 return 0;
1022 }
1023 NOKPROBE_SYMBOL(kprobe_int3_handler);
1024
1025 /*
1026 * When a retprobed function returns, this code saves registers and
1027 * calls trampoline_handler() runs, which calls the kretprobe's handler.
1028 */
1029 asm(
1030 ".text\n"
1031 ".global kretprobe_trampoline\n"
1032 ".type kretprobe_trampoline, @function\n"
1033 "kretprobe_trampoline:\n"
1034 /* We don't bother saving the ss register */
1035 #ifdef CONFIG_X86_64
1036 " pushq %rsp\n"
1037 " pushfq\n"
1038 SAVE_REGS_STRING
1039 " movq %rsp, %rdi\n"
1040 " call trampoline_handler\n"
1041 /* Replace saved sp with true return address. */
1042 " movq %rax, 19*8(%rsp)\n"
1043 RESTORE_REGS_STRING
1044 " popfq\n"
1045 #else
1046 " pushl %esp\n"
1047 " pushfl\n"
1048 SAVE_REGS_STRING
1049 " movl %esp, %eax\n"
1050 " call trampoline_handler\n"
1051 /* Replace saved sp with true return address. */
1052 " movl %eax, 15*4(%esp)\n"
1053 RESTORE_REGS_STRING
1054 " popfl\n"
1055 #endif
1056 ASM_RET
1057 ".size kretprobe_trampoline, .-kretprobe_trampoline\n"
1058 );
1059 NOKPROBE_SYMBOL(kretprobe_trampoline);
1060 STACK_FRAME_NON_STANDARD(kretprobe_trampoline);
1061
1062
1063 /*
1064 * Called from kretprobe_trampoline
1065 */
trampoline_handler(struct pt_regs * regs)1066 __used __visible void *trampoline_handler(struct pt_regs *regs)
1067 {
1068 /* fixup registers */
1069 regs->cs = __KERNEL_CS;
1070 #ifdef CONFIG_X86_32
1071 regs->gs = 0;
1072 #endif
1073 regs->ip = (unsigned long)&kretprobe_trampoline;
1074 regs->orig_ax = ~0UL;
1075
1076 return (void *)kretprobe_trampoline_handler(regs, &kretprobe_trampoline, ®s->sp);
1077 }
1078 NOKPROBE_SYMBOL(trampoline_handler);
1079
kprobe_fault_handler(struct pt_regs * regs,int trapnr)1080 int kprobe_fault_handler(struct pt_regs *regs, int trapnr)
1081 {
1082 struct kprobe *cur = kprobe_running();
1083 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1084
1085 if (unlikely(regs->ip == (unsigned long)cur->ainsn.insn)) {
1086 /* This must happen on single-stepping */
1087 WARN_ON(kcb->kprobe_status != KPROBE_HIT_SS &&
1088 kcb->kprobe_status != KPROBE_REENTER);
1089 /*
1090 * We are here because the instruction being single
1091 * stepped caused a page fault. We reset the current
1092 * kprobe and the ip points back to the probe address
1093 * and allow the page fault handler to continue as a
1094 * normal page fault.
1095 */
1096 regs->ip = (unsigned long)cur->addr;
1097
1098 /*
1099 * If the IF flag was set before the kprobe hit,
1100 * don't touch it:
1101 */
1102 regs->flags |= kcb->kprobe_old_flags;
1103
1104 if (kcb->kprobe_status == KPROBE_REENTER)
1105 restore_previous_kprobe(kcb);
1106 else
1107 reset_current_kprobe();
1108 }
1109
1110 return 0;
1111 }
1112 NOKPROBE_SYMBOL(kprobe_fault_handler);
1113
arch_populate_kprobe_blacklist(void)1114 int __init arch_populate_kprobe_blacklist(void)
1115 {
1116 return kprobe_add_area_blacklist((unsigned long)__entry_text_start,
1117 (unsigned long)__entry_text_end);
1118 }
1119
arch_init_kprobes(void)1120 int __init arch_init_kprobes(void)
1121 {
1122 return 0;
1123 }
1124
arch_trampoline_kprobe(struct kprobe * p)1125 int arch_trampoline_kprobe(struct kprobe *p)
1126 {
1127 return 0;
1128 }
1129