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 | #include <linux/set_memory.h> |
48 | #include <linux/cfi.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 | 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(add: (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(opcode: 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 |
194 | __recover_probed_insn(kprobe_opcode_t *buf, unsigned long addr) |
195 | { |
196 | struct kprobe *kp; |
197 | bool faddr; |
198 | |
199 | kp = get_kprobe(addr: (void *)addr); |
200 | faddr = ftrace_location(ip: addr) == addr; |
201 | /* |
202 | * Use the current code if it is not modified by Kprobe |
203 | * and it cannot be modified by ftrace. |
204 | */ |
205 | if (!kp && !faddr) |
206 | return addr; |
207 | |
208 | /* |
209 | * Basically, kp->ainsn.insn has an original instruction. |
210 | * However, RIP-relative instruction can not do single-stepping |
211 | * at different place, __copy_instruction() tweaks the displacement of |
212 | * that instruction. In that case, we can't recover the instruction |
213 | * from the kp->ainsn.insn. |
214 | * |
215 | * On the other hand, in case on normal Kprobe, kp->opcode has a copy |
216 | * of the first byte of the probed instruction, which is overwritten |
217 | * by int3. And the instruction at kp->addr is not modified by kprobes |
218 | * except for the first byte, we can recover the original instruction |
219 | * from it and kp->opcode. |
220 | * |
221 | * In case of Kprobes using ftrace, we do not have a copy of |
222 | * the original instruction. In fact, the ftrace location might |
223 | * be modified at anytime and even could be in an inconsistent state. |
224 | * Fortunately, we know that the original code is the ideal 5-byte |
225 | * long NOP. |
226 | */ |
227 | if (copy_from_kernel_nofault(dst: buf, src: (void *)addr, |
228 | MAX_INSN_SIZE * sizeof(kprobe_opcode_t))) |
229 | return 0UL; |
230 | |
231 | if (faddr) |
232 | memcpy(buf, x86_nops[5], 5); |
233 | else |
234 | buf[0] = kp->opcode; |
235 | return (unsigned long)buf; |
236 | } |
237 | |
238 | /* |
239 | * Recover the probed instruction at addr for further analysis. |
240 | * Caller must lock kprobes by kprobe_mutex, or disable preemption |
241 | * for preventing to release referencing kprobes. |
242 | * Returns zero if the instruction can not get recovered (or access failed). |
243 | */ |
244 | unsigned long recover_probed_instruction(kprobe_opcode_t *buf, unsigned long addr) |
245 | { |
246 | unsigned long __addr; |
247 | |
248 | __addr = __recover_optprobed_insn(buf, addr); |
249 | if (__addr != addr) |
250 | return __addr; |
251 | |
252 | return __recover_probed_insn(buf, addr); |
253 | } |
254 | |
255 | /* Check if paddr is at an instruction boundary */ |
256 | static int can_probe(unsigned long paddr) |
257 | { |
258 | unsigned long addr, __addr, offset = 0; |
259 | struct insn insn; |
260 | kprobe_opcode_t buf[MAX_INSN_SIZE]; |
261 | |
262 | if (!kallsyms_lookup_size_offset(addr: paddr, NULL, offset: &offset)) |
263 | return 0; |
264 | |
265 | /* Decode instructions */ |
266 | addr = paddr - offset; |
267 | while (addr < paddr) { |
268 | int ret; |
269 | |
270 | /* |
271 | * Check if the instruction has been modified by another |
272 | * kprobe, in which case we replace the breakpoint by the |
273 | * original instruction in our buffer. |
274 | * Also, jump optimization will change the breakpoint to |
275 | * relative-jump. Since the relative-jump itself is |
276 | * normally used, we just go through if there is no kprobe. |
277 | */ |
278 | __addr = recover_probed_instruction(buf, addr); |
279 | if (!__addr) |
280 | return 0; |
281 | |
282 | ret = insn_decode_kernel(&insn, (void *)__addr); |
283 | if (ret < 0) |
284 | return 0; |
285 | |
286 | #ifdef CONFIG_KGDB |
287 | /* |
288 | * If there is a dynamically installed kgdb sw breakpoint, |
289 | * this function should not be probed. |
290 | */ |
291 | if (insn.opcode.bytes[0] == INT3_INSN_OPCODE && |
292 | kgdb_has_hit_break(addr)) |
293 | return 0; |
294 | #endif |
295 | addr += insn.length; |
296 | } |
297 | if (IS_ENABLED(CONFIG_CFI_CLANG)) { |
298 | /* |
299 | * The compiler generates the following instruction sequence |
300 | * for indirect call checks and cfi.c decodes this; |
301 | * |
302 | * movl -<id>, %r10d ; 6 bytes |
303 | * addl -4(%reg), %r10d ; 4 bytes |
304 | * je .Ltmp1 ; 2 bytes |
305 | * ud2 ; <- regs->ip |
306 | * .Ltmp1: |
307 | * |
308 | * Also, these movl and addl are used for showing expected |
309 | * type. So those must not be touched. |
310 | */ |
311 | __addr = recover_probed_instruction(buf, addr); |
312 | if (!__addr) |
313 | return 0; |
314 | |
315 | if (insn_decode_kernel(&insn, (void *)__addr) < 0) |
316 | return 0; |
317 | |
318 | if (insn.opcode.value == 0xBA) |
319 | offset = 12; |
320 | else if (insn.opcode.value == 0x3) |
321 | offset = 6; |
322 | else |
323 | goto out; |
324 | |
325 | /* This movl/addl is used for decoding CFI. */ |
326 | if (is_cfi_trap(addr: addr + offset)) |
327 | return 0; |
328 | } |
329 | |
330 | out: |
331 | return (addr == paddr); |
332 | } |
333 | |
334 | /* If x86 supports IBT (ENDBR) it must be skipped. */ |
335 | kprobe_opcode_t *arch_adjust_kprobe_addr(unsigned long addr, unsigned long offset, |
336 | bool *on_func_entry) |
337 | { |
338 | if (is_endbr(val: *(u32 *)addr)) { |
339 | *on_func_entry = !offset || offset == 4; |
340 | if (*on_func_entry) |
341 | offset = 4; |
342 | |
343 | } else { |
344 | *on_func_entry = !offset; |
345 | } |
346 | |
347 | return (kprobe_opcode_t *)(addr + offset); |
348 | } |
349 | |
350 | /* |
351 | * Copy an instruction with recovering modified instruction by kprobes |
352 | * and adjust the displacement if the instruction uses the %rip-relative |
353 | * addressing mode. Note that since @real will be the final place of copied |
354 | * instruction, displacement must be adjust by @real, not @dest. |
355 | * This returns the length of copied instruction, or 0 if it has an error. |
356 | */ |
357 | int __copy_instruction(u8 *dest, u8 *src, u8 *real, struct insn *insn) |
358 | { |
359 | kprobe_opcode_t buf[MAX_INSN_SIZE]; |
360 | unsigned long recovered_insn = recover_probed_instruction(buf, addr: (unsigned long)src); |
361 | int ret; |
362 | |
363 | if (!recovered_insn || !insn) |
364 | return 0; |
365 | |
366 | /* This can access kernel text if given address is not recovered */ |
367 | if (copy_from_kernel_nofault(dst: dest, src: (void *)recovered_insn, |
368 | MAX_INSN_SIZE)) |
369 | return 0; |
370 | |
371 | ret = insn_decode_kernel(insn, dest); |
372 | if (ret < 0) |
373 | return 0; |
374 | |
375 | /* We can not probe force emulate prefixed instruction */ |
376 | if (insn_has_emulate_prefix(insn)) |
377 | return 0; |
378 | |
379 | /* Another subsystem puts a breakpoint, failed to recover */ |
380 | if (insn->opcode.bytes[0] == INT3_INSN_OPCODE) |
381 | return 0; |
382 | |
383 | /* We should not singlestep on the exception masking instructions */ |
384 | if (insn_masking_exception(insn)) |
385 | return 0; |
386 | |
387 | #ifdef CONFIG_X86_64 |
388 | /* Only x86_64 has RIP relative instructions */ |
389 | if (insn_rip_relative(insn)) { |
390 | s64 newdisp; |
391 | u8 *disp; |
392 | /* |
393 | * The copied instruction uses the %rip-relative addressing |
394 | * mode. Adjust the displacement for the difference between |
395 | * the original location of this instruction and the location |
396 | * of the copy that will actually be run. The tricky bit here |
397 | * is making sure that the sign extension happens correctly in |
398 | * this calculation, since we need a signed 32-bit result to |
399 | * be sign-extended to 64 bits when it's added to the %rip |
400 | * value and yield the same 64-bit result that the sign- |
401 | * extension of the original signed 32-bit displacement would |
402 | * have given. |
403 | */ |
404 | newdisp = (u8 *) src + (s64) insn->displacement.value |
405 | - (u8 *) real; |
406 | if ((s64) (s32) newdisp != newdisp) { |
407 | pr_err("Kprobes error: new displacement does not fit into s32 (%llx)\n" , newdisp); |
408 | return 0; |
409 | } |
410 | disp = (u8 *) dest + insn_offset_displacement(insn); |
411 | *(s32 *) disp = (s32) newdisp; |
412 | } |
413 | #endif |
414 | return insn->length; |
415 | } |
416 | |
417 | /* Prepare reljump or int3 right after instruction */ |
418 | static int prepare_singlestep(kprobe_opcode_t *buf, struct kprobe *p, |
419 | struct insn *insn) |
420 | { |
421 | int len = insn->length; |
422 | |
423 | if (!IS_ENABLED(CONFIG_PREEMPTION) && |
424 | !p->post_handler && can_boost(insn, addr: p->addr) && |
425 | MAX_INSN_SIZE - len >= JMP32_INSN_SIZE) { |
426 | /* |
427 | * These instructions can be executed directly if it |
428 | * jumps back to correct address. |
429 | */ |
430 | synthesize_reljump(dest: buf + len, from: p->ainsn.insn + len, |
431 | to: p->addr + insn->length); |
432 | len += JMP32_INSN_SIZE; |
433 | p->ainsn.boostable = 1; |
434 | } else { |
435 | /* Otherwise, put an int3 for trapping singlestep */ |
436 | if (MAX_INSN_SIZE - len < INT3_INSN_SIZE) |
437 | return -ENOSPC; |
438 | |
439 | buf[len] = INT3_INSN_OPCODE; |
440 | len += INT3_INSN_SIZE; |
441 | } |
442 | |
443 | return len; |
444 | } |
445 | |
446 | /* Make page to RO mode when allocate it */ |
447 | void *alloc_insn_page(void) |
448 | { |
449 | void *page; |
450 | |
451 | page = module_alloc(PAGE_SIZE); |
452 | if (!page) |
453 | return NULL; |
454 | |
455 | /* |
456 | * TODO: Once additional kernel code protection mechanisms are set, ensure |
457 | * that the page was not maliciously altered and it is still zeroed. |
458 | */ |
459 | set_memory_rox(addr: (unsigned long)page, numpages: 1); |
460 | |
461 | return page; |
462 | } |
463 | |
464 | /* Kprobe x86 instruction emulation - only regs->ip or IF flag modifiers */ |
465 | |
466 | static void kprobe_emulate_ifmodifiers(struct kprobe *p, struct pt_regs *regs) |
467 | { |
468 | switch (p->ainsn.opcode) { |
469 | case 0xfa: /* cli */ |
470 | regs->flags &= ~(X86_EFLAGS_IF); |
471 | break; |
472 | case 0xfb: /* sti */ |
473 | regs->flags |= X86_EFLAGS_IF; |
474 | break; |
475 | case 0x9c: /* pushf */ |
476 | int3_emulate_push(regs, val: regs->flags); |
477 | break; |
478 | case 0x9d: /* popf */ |
479 | regs->flags = int3_emulate_pop(regs); |
480 | break; |
481 | } |
482 | regs->ip = regs->ip - INT3_INSN_SIZE + p->ainsn.size; |
483 | } |
484 | NOKPROBE_SYMBOL(kprobe_emulate_ifmodifiers); |
485 | |
486 | static void kprobe_emulate_ret(struct kprobe *p, struct pt_regs *regs) |
487 | { |
488 | int3_emulate_ret(regs); |
489 | } |
490 | NOKPROBE_SYMBOL(kprobe_emulate_ret); |
491 | |
492 | static void kprobe_emulate_call(struct kprobe *p, struct pt_regs *regs) |
493 | { |
494 | unsigned long func = regs->ip - INT3_INSN_SIZE + p->ainsn.size; |
495 | |
496 | func += p->ainsn.rel32; |
497 | int3_emulate_call(regs, func); |
498 | } |
499 | NOKPROBE_SYMBOL(kprobe_emulate_call); |
500 | |
501 | static void kprobe_emulate_jmp(struct kprobe *p, struct pt_regs *regs) |
502 | { |
503 | unsigned long ip = regs->ip - INT3_INSN_SIZE + p->ainsn.size; |
504 | |
505 | ip += p->ainsn.rel32; |
506 | int3_emulate_jmp(regs, ip); |
507 | } |
508 | NOKPROBE_SYMBOL(kprobe_emulate_jmp); |
509 | |
510 | static void kprobe_emulate_jcc(struct kprobe *p, struct pt_regs *regs) |
511 | { |
512 | unsigned long ip = regs->ip - INT3_INSN_SIZE + p->ainsn.size; |
513 | |
514 | int3_emulate_jcc(regs, cc: p->ainsn.jcc.type, ip, disp: p->ainsn.rel32); |
515 | } |
516 | NOKPROBE_SYMBOL(kprobe_emulate_jcc); |
517 | |
518 | static void kprobe_emulate_loop(struct kprobe *p, struct pt_regs *regs) |
519 | { |
520 | unsigned long ip = regs->ip - INT3_INSN_SIZE + p->ainsn.size; |
521 | bool match; |
522 | |
523 | if (p->ainsn.loop.type != 3) { /* LOOP* */ |
524 | if (p->ainsn.loop.asize == 32) |
525 | match = ((*(u32 *)®s->cx)--) != 0; |
526 | #ifdef CONFIG_X86_64 |
527 | else if (p->ainsn.loop.asize == 64) |
528 | match = ((*(u64 *)®s->cx)--) != 0; |
529 | #endif |
530 | else |
531 | match = ((*(u16 *)®s->cx)--) != 0; |
532 | } else { /* JCXZ */ |
533 | if (p->ainsn.loop.asize == 32) |
534 | match = *(u32 *)(®s->cx) == 0; |
535 | #ifdef CONFIG_X86_64 |
536 | else if (p->ainsn.loop.asize == 64) |
537 | match = *(u64 *)(®s->cx) == 0; |
538 | #endif |
539 | else |
540 | match = *(u16 *)(®s->cx) == 0; |
541 | } |
542 | |
543 | if (p->ainsn.loop.type == 0) /* LOOPNE */ |
544 | match = match && !(regs->flags & X86_EFLAGS_ZF); |
545 | else if (p->ainsn.loop.type == 1) /* LOOPE */ |
546 | match = match && (regs->flags & X86_EFLAGS_ZF); |
547 | |
548 | if (match) |
549 | ip += p->ainsn.rel32; |
550 | int3_emulate_jmp(regs, ip); |
551 | } |
552 | NOKPROBE_SYMBOL(kprobe_emulate_loop); |
553 | |
554 | static const int addrmode_regoffs[] = { |
555 | offsetof(struct pt_regs, ax), |
556 | offsetof(struct pt_regs, cx), |
557 | offsetof(struct pt_regs, dx), |
558 | offsetof(struct pt_regs, bx), |
559 | offsetof(struct pt_regs, sp), |
560 | offsetof(struct pt_regs, bp), |
561 | offsetof(struct pt_regs, si), |
562 | offsetof(struct pt_regs, di), |
563 | #ifdef CONFIG_X86_64 |
564 | offsetof(struct pt_regs, r8), |
565 | offsetof(struct pt_regs, r9), |
566 | offsetof(struct pt_regs, r10), |
567 | offsetof(struct pt_regs, r11), |
568 | offsetof(struct pt_regs, r12), |
569 | offsetof(struct pt_regs, r13), |
570 | offsetof(struct pt_regs, r14), |
571 | offsetof(struct pt_regs, r15), |
572 | #endif |
573 | }; |
574 | |
575 | static void kprobe_emulate_call_indirect(struct kprobe *p, struct pt_regs *regs) |
576 | { |
577 | unsigned long offs = addrmode_regoffs[p->ainsn.indirect.reg]; |
578 | |
579 | int3_emulate_call(regs, func: regs_get_register(regs, offset: offs)); |
580 | } |
581 | NOKPROBE_SYMBOL(kprobe_emulate_call_indirect); |
582 | |
583 | static void kprobe_emulate_jmp_indirect(struct kprobe *p, struct pt_regs *regs) |
584 | { |
585 | unsigned long offs = addrmode_regoffs[p->ainsn.indirect.reg]; |
586 | |
587 | int3_emulate_jmp(regs, ip: regs_get_register(regs, offset: offs)); |
588 | } |
589 | NOKPROBE_SYMBOL(kprobe_emulate_jmp_indirect); |
590 | |
591 | static int prepare_emulation(struct kprobe *p, struct insn *insn) |
592 | { |
593 | insn_byte_t opcode = insn->opcode.bytes[0]; |
594 | |
595 | switch (opcode) { |
596 | case 0xfa: /* cli */ |
597 | case 0xfb: /* sti */ |
598 | case 0x9c: /* pushfl */ |
599 | case 0x9d: /* popf/popfd */ |
600 | /* |
601 | * IF modifiers must be emulated since it will enable interrupt while |
602 | * int3 single stepping. |
603 | */ |
604 | p->ainsn.emulate_op = kprobe_emulate_ifmodifiers; |
605 | p->ainsn.opcode = opcode; |
606 | break; |
607 | case 0xc2: /* ret/lret */ |
608 | case 0xc3: |
609 | case 0xca: |
610 | case 0xcb: |
611 | p->ainsn.emulate_op = kprobe_emulate_ret; |
612 | break; |
613 | case 0x9a: /* far call absolute -- segment is not supported */ |
614 | case 0xea: /* far jmp absolute -- segment is not supported */ |
615 | case 0xcc: /* int3 */ |
616 | case 0xcf: /* iret -- in-kernel IRET is not supported */ |
617 | return -EOPNOTSUPP; |
618 | break; |
619 | case 0xe8: /* near call relative */ |
620 | p->ainsn.emulate_op = kprobe_emulate_call; |
621 | if (insn->immediate.nbytes == 2) |
622 | p->ainsn.rel32 = *(s16 *)&insn->immediate.value; |
623 | else |
624 | p->ainsn.rel32 = *(s32 *)&insn->immediate.value; |
625 | break; |
626 | case 0xeb: /* short jump relative */ |
627 | case 0xe9: /* near jump relative */ |
628 | p->ainsn.emulate_op = kprobe_emulate_jmp; |
629 | if (insn->immediate.nbytes == 1) |
630 | p->ainsn.rel32 = *(s8 *)&insn->immediate.value; |
631 | else if (insn->immediate.nbytes == 2) |
632 | p->ainsn.rel32 = *(s16 *)&insn->immediate.value; |
633 | else |
634 | p->ainsn.rel32 = *(s32 *)&insn->immediate.value; |
635 | break; |
636 | case 0x70 ... 0x7f: |
637 | /* 1 byte conditional jump */ |
638 | p->ainsn.emulate_op = kprobe_emulate_jcc; |
639 | p->ainsn.jcc.type = opcode & 0xf; |
640 | p->ainsn.rel32 = insn->immediate.value; |
641 | break; |
642 | case 0x0f: |
643 | opcode = insn->opcode.bytes[1]; |
644 | if ((opcode & 0xf0) == 0x80) { |
645 | /* 2 bytes Conditional Jump */ |
646 | p->ainsn.emulate_op = kprobe_emulate_jcc; |
647 | p->ainsn.jcc.type = opcode & 0xf; |
648 | if (insn->immediate.nbytes == 2) |
649 | p->ainsn.rel32 = *(s16 *)&insn->immediate.value; |
650 | else |
651 | p->ainsn.rel32 = *(s32 *)&insn->immediate.value; |
652 | } else if (opcode == 0x01 && |
653 | X86_MODRM_REG(insn->modrm.bytes[0]) == 0 && |
654 | X86_MODRM_MOD(insn->modrm.bytes[0]) == 3) { |
655 | /* VM extensions - not supported */ |
656 | return -EOPNOTSUPP; |
657 | } |
658 | break; |
659 | case 0xe0: /* Loop NZ */ |
660 | case 0xe1: /* Loop */ |
661 | case 0xe2: /* Loop */ |
662 | case 0xe3: /* J*CXZ */ |
663 | p->ainsn.emulate_op = kprobe_emulate_loop; |
664 | p->ainsn.loop.type = opcode & 0x3; |
665 | p->ainsn.loop.asize = insn->addr_bytes * 8; |
666 | p->ainsn.rel32 = *(s8 *)&insn->immediate.value; |
667 | break; |
668 | case 0xff: |
669 | /* |
670 | * Since the 0xff is an extended group opcode, the instruction |
671 | * is determined by the MOD/RM byte. |
672 | */ |
673 | opcode = insn->modrm.bytes[0]; |
674 | switch (X86_MODRM_REG(opcode)) { |
675 | case 0b010: /* FF /2, call near, absolute indirect */ |
676 | p->ainsn.emulate_op = kprobe_emulate_call_indirect; |
677 | break; |
678 | case 0b100: /* FF /4, jmp near, absolute indirect */ |
679 | p->ainsn.emulate_op = kprobe_emulate_jmp_indirect; |
680 | break; |
681 | case 0b011: /* FF /3, call far, absolute indirect */ |
682 | case 0b101: /* FF /5, jmp far, absolute indirect */ |
683 | return -EOPNOTSUPP; |
684 | } |
685 | |
686 | if (!p->ainsn.emulate_op) |
687 | break; |
688 | |
689 | if (insn->addr_bytes != sizeof(unsigned long)) |
690 | return -EOPNOTSUPP; /* Don't support different size */ |
691 | if (X86_MODRM_MOD(opcode) != 3) |
692 | return -EOPNOTSUPP; /* TODO: support memory addressing */ |
693 | |
694 | p->ainsn.indirect.reg = X86_MODRM_RM(opcode); |
695 | #ifdef CONFIG_X86_64 |
696 | if (X86_REX_B(insn->rex_prefix.value)) |
697 | p->ainsn.indirect.reg += 8; |
698 | #endif |
699 | break; |
700 | default: |
701 | break; |
702 | } |
703 | p->ainsn.size = insn->length; |
704 | |
705 | return 0; |
706 | } |
707 | |
708 | static int arch_copy_kprobe(struct kprobe *p) |
709 | { |
710 | struct insn insn; |
711 | kprobe_opcode_t buf[MAX_INSN_SIZE]; |
712 | int ret, len; |
713 | |
714 | /* Copy an instruction with recovering if other optprobe modifies it.*/ |
715 | len = __copy_instruction(dest: buf, src: p->addr, real: p->ainsn.insn, insn: &insn); |
716 | if (!len) |
717 | return -EINVAL; |
718 | |
719 | /* Analyze the opcode and setup emulate functions */ |
720 | ret = prepare_emulation(p, insn: &insn); |
721 | if (ret < 0) |
722 | return ret; |
723 | |
724 | /* Add int3 for single-step or booster jmp */ |
725 | len = prepare_singlestep(buf, p, insn: &insn); |
726 | if (len < 0) |
727 | return len; |
728 | |
729 | /* Also, displacement change doesn't affect the first byte */ |
730 | p->opcode = buf[0]; |
731 | |
732 | p->ainsn.tp_len = len; |
733 | perf_event_text_poke(addr: p->ainsn.insn, NULL, old_len: 0, new_bytes: buf, new_len: len); |
734 | |
735 | /* OK, write back the instruction(s) into ROX insn buffer */ |
736 | text_poke(addr: p->ainsn.insn, opcode: buf, len); |
737 | |
738 | return 0; |
739 | } |
740 | |
741 | int arch_prepare_kprobe(struct kprobe *p) |
742 | { |
743 | int ret; |
744 | |
745 | if (alternatives_text_reserved(start: p->addr, end: p->addr)) |
746 | return -EINVAL; |
747 | |
748 | if (!can_probe(paddr: (unsigned long)p->addr)) |
749 | return -EILSEQ; |
750 | |
751 | memset(&p->ainsn, 0, sizeof(p->ainsn)); |
752 | |
753 | /* insn: must be on special executable page on x86. */ |
754 | p->ainsn.insn = get_insn_slot(); |
755 | if (!p->ainsn.insn) |
756 | return -ENOMEM; |
757 | |
758 | ret = arch_copy_kprobe(p); |
759 | if (ret) { |
760 | free_insn_slot(slot: p->ainsn.insn, dirty: 0); |
761 | p->ainsn.insn = NULL; |
762 | } |
763 | |
764 | return ret; |
765 | } |
766 | |
767 | void arch_arm_kprobe(struct kprobe *p) |
768 | { |
769 | u8 int3 = INT3_INSN_OPCODE; |
770 | |
771 | text_poke(addr: p->addr, opcode: &int3, len: 1); |
772 | text_poke_sync(); |
773 | perf_event_text_poke(addr: p->addr, old_bytes: &p->opcode, old_len: 1, new_bytes: &int3, new_len: 1); |
774 | } |
775 | |
776 | void arch_disarm_kprobe(struct kprobe *p) |
777 | { |
778 | u8 int3 = INT3_INSN_OPCODE; |
779 | |
780 | perf_event_text_poke(addr: p->addr, old_bytes: &int3, old_len: 1, new_bytes: &p->opcode, new_len: 1); |
781 | text_poke(addr: p->addr, opcode: &p->opcode, len: 1); |
782 | text_poke_sync(); |
783 | } |
784 | |
785 | void arch_remove_kprobe(struct kprobe *p) |
786 | { |
787 | if (p->ainsn.insn) { |
788 | /* Record the perf event before freeing the slot */ |
789 | perf_event_text_poke(addr: p->ainsn.insn, old_bytes: p->ainsn.insn, |
790 | old_len: p->ainsn.tp_len, NULL, new_len: 0); |
791 | free_insn_slot(slot: p->ainsn.insn, dirty: p->ainsn.boostable); |
792 | p->ainsn.insn = NULL; |
793 | } |
794 | } |
795 | |
796 | static nokprobe_inline void |
797 | save_previous_kprobe(struct kprobe_ctlblk *kcb) |
798 | { |
799 | kcb->prev_kprobe.kp = kprobe_running(); |
800 | kcb->prev_kprobe.status = kcb->kprobe_status; |
801 | kcb->prev_kprobe.old_flags = kcb->kprobe_old_flags; |
802 | kcb->prev_kprobe.saved_flags = kcb->kprobe_saved_flags; |
803 | } |
804 | |
805 | static nokprobe_inline void |
806 | restore_previous_kprobe(struct kprobe_ctlblk *kcb) |
807 | { |
808 | __this_cpu_write(current_kprobe, kcb->prev_kprobe.kp); |
809 | kcb->kprobe_status = kcb->prev_kprobe.status; |
810 | kcb->kprobe_old_flags = kcb->prev_kprobe.old_flags; |
811 | kcb->kprobe_saved_flags = kcb->prev_kprobe.saved_flags; |
812 | } |
813 | |
814 | static nokprobe_inline void |
815 | set_current_kprobe(struct kprobe *p, struct pt_regs *regs, |
816 | struct kprobe_ctlblk *kcb) |
817 | { |
818 | __this_cpu_write(current_kprobe, p); |
819 | kcb->kprobe_saved_flags = kcb->kprobe_old_flags |
820 | = (regs->flags & X86_EFLAGS_IF); |
821 | } |
822 | |
823 | static void kprobe_post_process(struct kprobe *cur, struct pt_regs *regs, |
824 | struct kprobe_ctlblk *kcb) |
825 | { |
826 | /* Restore back the original saved kprobes variables and continue. */ |
827 | if (kcb->kprobe_status == KPROBE_REENTER) { |
828 | /* This will restore both kcb and current_kprobe */ |
829 | restore_previous_kprobe(kcb); |
830 | } else { |
831 | /* |
832 | * Always update the kcb status because |
833 | * reset_curent_kprobe() doesn't update kcb. |
834 | */ |
835 | kcb->kprobe_status = KPROBE_HIT_SSDONE; |
836 | if (cur->post_handler) |
837 | cur->post_handler(cur, regs, 0); |
838 | reset_current_kprobe(); |
839 | } |
840 | } |
841 | NOKPROBE_SYMBOL(kprobe_post_process); |
842 | |
843 | static void setup_singlestep(struct kprobe *p, struct pt_regs *regs, |
844 | struct kprobe_ctlblk *kcb, int reenter) |
845 | { |
846 | if (setup_detour_execution(p, regs, reenter)) |
847 | return; |
848 | |
849 | #if !defined(CONFIG_PREEMPTION) |
850 | if (p->ainsn.boostable) { |
851 | /* Boost up -- we can execute copied instructions directly */ |
852 | if (!reenter) |
853 | reset_current_kprobe(); |
854 | /* |
855 | * Reentering boosted probe doesn't reset current_kprobe, |
856 | * nor set current_kprobe, because it doesn't use single |
857 | * stepping. |
858 | */ |
859 | regs->ip = (unsigned long)p->ainsn.insn; |
860 | return; |
861 | } |
862 | #endif |
863 | if (reenter) { |
864 | save_previous_kprobe(kcb); |
865 | set_current_kprobe(p, regs, kcb); |
866 | kcb->kprobe_status = KPROBE_REENTER; |
867 | } else |
868 | kcb->kprobe_status = KPROBE_HIT_SS; |
869 | |
870 | if (p->ainsn.emulate_op) { |
871 | p->ainsn.emulate_op(p, regs); |
872 | kprobe_post_process(cur: p, regs, kcb); |
873 | return; |
874 | } |
875 | |
876 | /* Disable interrupt, and set ip register on trampoline */ |
877 | regs->flags &= ~X86_EFLAGS_IF; |
878 | regs->ip = (unsigned long)p->ainsn.insn; |
879 | } |
880 | NOKPROBE_SYMBOL(setup_singlestep); |
881 | |
882 | /* |
883 | * Called after single-stepping. p->addr is the address of the |
884 | * instruction whose first byte has been replaced by the "int3" |
885 | * instruction. To avoid the SMP problems that can occur when we |
886 | * temporarily put back the original opcode to single-step, we |
887 | * single-stepped a copy of the instruction. The address of this |
888 | * copy is p->ainsn.insn. We also doesn't use trap, but "int3" again |
889 | * right after the copied instruction. |
890 | * Different from the trap single-step, "int3" single-step can not |
891 | * handle the instruction which changes the ip register, e.g. jmp, |
892 | * call, conditional jmp, and the instructions which changes the IF |
893 | * flags because interrupt must be disabled around the single-stepping. |
894 | * Such instructions are software emulated, but others are single-stepped |
895 | * using "int3". |
896 | * |
897 | * When the 2nd "int3" handled, the regs->ip and regs->flags needs to |
898 | * be adjusted, so that we can resume execution on correct code. |
899 | */ |
900 | static void resume_singlestep(struct kprobe *p, struct pt_regs *regs, |
901 | struct kprobe_ctlblk *kcb) |
902 | { |
903 | unsigned long copy_ip = (unsigned long)p->ainsn.insn; |
904 | unsigned long orig_ip = (unsigned long)p->addr; |
905 | |
906 | /* Restore saved interrupt flag and ip register */ |
907 | regs->flags |= kcb->kprobe_saved_flags; |
908 | /* Note that regs->ip is executed int3 so must be a step back */ |
909 | regs->ip += (orig_ip - copy_ip) - INT3_INSN_SIZE; |
910 | } |
911 | NOKPROBE_SYMBOL(resume_singlestep); |
912 | |
913 | /* |
914 | * We have reentered the kprobe_handler(), since another probe was hit while |
915 | * within the handler. We save the original kprobes variables and just single |
916 | * step on the instruction of the new probe without calling any user handlers. |
917 | */ |
918 | static int reenter_kprobe(struct kprobe *p, struct pt_regs *regs, |
919 | struct kprobe_ctlblk *kcb) |
920 | { |
921 | switch (kcb->kprobe_status) { |
922 | case KPROBE_HIT_SSDONE: |
923 | case KPROBE_HIT_ACTIVE: |
924 | case KPROBE_HIT_SS: |
925 | kprobes_inc_nmissed_count(p); |
926 | setup_singlestep(p, regs, kcb, reenter: 1); |
927 | break; |
928 | case KPROBE_REENTER: |
929 | /* A probe has been hit in the codepath leading up to, or just |
930 | * after, single-stepping of a probed instruction. This entire |
931 | * codepath should strictly reside in .kprobes.text section. |
932 | * Raise a BUG or we'll continue in an endless reentering loop |
933 | * and eventually a stack overflow. |
934 | */ |
935 | pr_err("Unrecoverable kprobe detected.\n" ); |
936 | dump_kprobe(kp: p); |
937 | BUG(); |
938 | default: |
939 | /* impossible cases */ |
940 | WARN_ON(1); |
941 | return 0; |
942 | } |
943 | |
944 | return 1; |
945 | } |
946 | NOKPROBE_SYMBOL(reenter_kprobe); |
947 | |
948 | static nokprobe_inline int kprobe_is_ss(struct kprobe_ctlblk *kcb) |
949 | { |
950 | return (kcb->kprobe_status == KPROBE_HIT_SS || |
951 | kcb->kprobe_status == KPROBE_REENTER); |
952 | } |
953 | |
954 | /* |
955 | * Interrupts are disabled on entry as trap3 is an interrupt gate and they |
956 | * remain disabled throughout this function. |
957 | */ |
958 | int kprobe_int3_handler(struct pt_regs *regs) |
959 | { |
960 | kprobe_opcode_t *addr; |
961 | struct kprobe *p; |
962 | struct kprobe_ctlblk *kcb; |
963 | |
964 | if (user_mode(regs)) |
965 | return 0; |
966 | |
967 | addr = (kprobe_opcode_t *)(regs->ip - sizeof(kprobe_opcode_t)); |
968 | /* |
969 | * We don't want to be preempted for the entire duration of kprobe |
970 | * processing. Since int3 and debug trap disables irqs and we clear |
971 | * IF while singlestepping, it must be no preemptible. |
972 | */ |
973 | |
974 | kcb = get_kprobe_ctlblk(); |
975 | p = get_kprobe(addr); |
976 | |
977 | if (p) { |
978 | if (kprobe_running()) { |
979 | if (reenter_kprobe(p, regs, kcb)) |
980 | return 1; |
981 | } else { |
982 | set_current_kprobe(p, regs, kcb); |
983 | kcb->kprobe_status = KPROBE_HIT_ACTIVE; |
984 | |
985 | /* |
986 | * If we have no pre-handler or it returned 0, we |
987 | * continue with normal processing. If we have a |
988 | * pre-handler and it returned non-zero, that means |
989 | * user handler setup registers to exit to another |
990 | * instruction, we must skip the single stepping. |
991 | */ |
992 | if (!p->pre_handler || !p->pre_handler(p, regs)) |
993 | setup_singlestep(p, regs, kcb, reenter: 0); |
994 | else |
995 | reset_current_kprobe(); |
996 | return 1; |
997 | } |
998 | } else if (kprobe_is_ss(kcb)) { |
999 | p = kprobe_running(); |
1000 | if ((unsigned long)p->ainsn.insn < regs->ip && |
1001 | (unsigned long)p->ainsn.insn + MAX_INSN_SIZE > regs->ip) { |
1002 | /* Most provably this is the second int3 for singlestep */ |
1003 | resume_singlestep(p, regs, kcb); |
1004 | kprobe_post_process(cur: p, regs, kcb); |
1005 | return 1; |
1006 | } |
1007 | } /* else: not a kprobe fault; let the kernel handle it */ |
1008 | |
1009 | return 0; |
1010 | } |
1011 | NOKPROBE_SYMBOL(kprobe_int3_handler); |
1012 | |
1013 | int kprobe_fault_handler(struct pt_regs *regs, int trapnr) |
1014 | { |
1015 | struct kprobe *cur = kprobe_running(); |
1016 | struct kprobe_ctlblk *kcb = get_kprobe_ctlblk(); |
1017 | |
1018 | if (unlikely(regs->ip == (unsigned long)cur->ainsn.insn)) { |
1019 | /* This must happen on single-stepping */ |
1020 | WARN_ON(kcb->kprobe_status != KPROBE_HIT_SS && |
1021 | kcb->kprobe_status != KPROBE_REENTER); |
1022 | /* |
1023 | * We are here because the instruction being single |
1024 | * stepped caused a page fault. We reset the current |
1025 | * kprobe and the ip points back to the probe address |
1026 | * and allow the page fault handler to continue as a |
1027 | * normal page fault. |
1028 | */ |
1029 | regs->ip = (unsigned long)cur->addr; |
1030 | |
1031 | /* |
1032 | * If the IF flag was set before the kprobe hit, |
1033 | * don't touch it: |
1034 | */ |
1035 | regs->flags |= kcb->kprobe_old_flags; |
1036 | |
1037 | if (kcb->kprobe_status == KPROBE_REENTER) |
1038 | restore_previous_kprobe(kcb); |
1039 | else |
1040 | reset_current_kprobe(); |
1041 | } |
1042 | |
1043 | return 0; |
1044 | } |
1045 | NOKPROBE_SYMBOL(kprobe_fault_handler); |
1046 | |
1047 | int __init arch_populate_kprobe_blacklist(void) |
1048 | { |
1049 | return kprobe_add_area_blacklist(start: (unsigned long)__entry_text_start, |
1050 | end: (unsigned long)__entry_text_end); |
1051 | } |
1052 | |
1053 | int __init arch_init_kprobes(void) |
1054 | { |
1055 | return 0; |
1056 | } |
1057 | |
1058 | int arch_trampoline_kprobe(struct kprobe *p) |
1059 | { |
1060 | return 0; |
1061 | } |
1062 | |