2 * Kernel Probes (KProbes)
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18 * Copyright (C) IBM Corporation, 2002, 2004
20 * 2002-Oct Created by Vamsi Krishna S <vamsi_krishna@in.ibm.com> Kernel
21 * Probes initial implementation ( includes contributions from
23 * 2004-July Suparna Bhattacharya <suparna@in.ibm.com> added jumper probes
24 * interface to access function arguments.
25 * 2004-Oct Jim Keniston <jkenisto@us.ibm.com> and Prasanna S Panchamukhi
26 * <prasanna@in.ibm.com> adapted for x86_64 from i386.
27 * 2005-Mar Roland McGrath <roland@redhat.com>
28 * Fixed to handle %rip-relative addressing mode correctly.
29 * 2005-May Hien Nguyen <hien@us.ibm.com>, Jim Keniston
30 * <jkenisto@us.ibm.com> and Prasanna S Panchamukhi
31 * <prasanna@in.ibm.com> added function-return probes.
32 * 2005-May Rusty Lynch <rusty.lynch@intel.com>
33 * Added function return probes functionality
34 * 2006-Feb Masami Hiramatsu <hiramatu@sdl.hitachi.co.jp> added
35 * kprobe-booster and kretprobe-booster for i386.
36 * 2007-Dec Masami Hiramatsu <mhiramat@redhat.com> added kprobe-booster
37 * and kretprobe-booster for x86-64
38 * 2007-Dec Masami Hiramatsu <mhiramat@redhat.com>, Arjan van de Ven
39 * <arjan@infradead.org> and Jim Keniston <jkenisto@us.ibm.com>
40 * unified x86 kprobes code.
42 #include <linux/kprobes.h>
43 #include <linux/ptrace.h>
44 #include <linux/string.h>
45 #include <linux/slab.h>
46 #include <linux/hardirq.h>
47 #include <linux/preempt.h>
48 #include <linux/module.h>
49 #include <linux/kdebug.h>
50 #include <linux/kallsyms.h>
51 #include <linux/ftrace.h>
52 #include <linux/frame.h>
54 #include <asm/text-patching.h>
55 #include <asm/cacheflush.h>
57 #include <asm/pgtable.h>
58 #include <asm/uaccess.h>
59 #include <asm/alternative.h>
61 #include <asm/debugreg.h>
65 void jprobe_return_end(void);
67 DEFINE_PER_CPU(struct kprobe
*, current_kprobe
) = NULL
;
68 DEFINE_PER_CPU(struct kprobe_ctlblk
, kprobe_ctlblk
);
70 #define stack_addr(regs) ((unsigned long *)kernel_stack_pointer(regs))
72 #define W(row, b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, ba, bb, bc, bd, be, bf)\
73 (((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) | \
74 (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) | \
75 (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) | \
76 (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf)) \
79 * Undefined/reserved opcodes, conditional jump, Opcode Extension
80 * Groups, and some special opcodes can not boost.
81 * This is non-const and volatile to keep gcc from statically
82 * optimizing it out, as variable_test_bit makes gcc think only
83 * *(unsigned long*) is used.
85 static volatile u32 twobyte_is_boostable
[256 / 32] = {
86 /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */
87 /* ---------------------------------------------- */
88 W(0x00, 0, 0, 1, 1, 0, 0, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0) | /* 00 */
89 W(0x10, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1) , /* 10 */
90 W(0x20, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 20 */
91 W(0x30, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 30 */
92 W(0x40, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 40 */
93 W(0x50, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 50 */
94 W(0x60, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1) | /* 60 */
95 W(0x70, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1) , /* 70 */
96 W(0x80, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 80 */
97 W(0x90, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 90 */
98 W(0xa0, 1, 1, 0, 1, 1, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* a0 */
99 W(0xb0, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1) , /* b0 */
100 W(0xc0, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1) | /* c0 */
101 W(0xd0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) , /* d0 */
102 W(0xe0, 0, 1, 1, 0, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* e0 */
103 W(0xf0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 1, 0, 1, 1, 1, 0) /* f0 */
104 /* ----------------------------------------------- */
105 /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */
109 struct kretprobe_blackpoint kretprobe_blacklist
[] = {
110 {"__switch_to", }, /* This function switches only current task, but
111 doesn't switch kernel stack.*/
112 {NULL
, NULL
} /* Terminator */
115 const int kretprobe_blacklist_size
= ARRAY_SIZE(kretprobe_blacklist
);
117 static nokprobe_inline
void
118 __synthesize_relative_insn(void *from
, void *to
, u8 op
)
120 struct __arch_relative_insn
{
125 insn
= (struct __arch_relative_insn
*)from
;
126 insn
->raddr
= (s32
)((long)(to
) - ((long)(from
) + 5));
130 /* Insert a jump instruction at address 'from', which jumps to address 'to'.*/
131 void synthesize_reljump(void *from
, void *to
)
133 __synthesize_relative_insn(from
, to
, RELATIVEJUMP_OPCODE
);
135 NOKPROBE_SYMBOL(synthesize_reljump
);
137 /* Insert a call instruction at address 'from', which calls address 'to'.*/
138 void synthesize_relcall(void *from
, void *to
)
140 __synthesize_relative_insn(from
, to
, RELATIVECALL_OPCODE
);
142 NOKPROBE_SYMBOL(synthesize_relcall
);
145 * Skip the prefixes of the instruction.
147 static kprobe_opcode_t
*skip_prefixes(kprobe_opcode_t
*insn
)
151 attr
= inat_get_opcode_attribute((insn_byte_t
)*insn
);
152 while (inat_is_legacy_prefix(attr
)) {
154 attr
= inat_get_opcode_attribute((insn_byte_t
)*insn
);
157 if (inat_is_rex_prefix(attr
))
162 NOKPROBE_SYMBOL(skip_prefixes
);
165 * Returns non-zero if opcode is boostable.
166 * RIP relative instructions are adjusted at copying time in 64 bits mode
168 int can_boost(kprobe_opcode_t
*opcodes
)
170 kprobe_opcode_t opcode
;
171 kprobe_opcode_t
*orig_opcodes
= opcodes
;
173 if (search_exception_tables((unsigned long)opcodes
))
174 return 0; /* Page fault may occur on this address. */
177 if (opcodes
- orig_opcodes
> MAX_INSN_SIZE
- 1)
179 opcode
= *(opcodes
++);
181 /* 2nd-byte opcode */
182 if (opcode
== 0x0f) {
183 if (opcodes
- orig_opcodes
> MAX_INSN_SIZE
- 1)
185 return test_bit(*opcodes
,
186 (unsigned long *)twobyte_is_boostable
);
189 switch (opcode
& 0xf0) {
192 goto retry
; /* REX prefix is boostable */
195 if (0x63 < opcode
&& opcode
< 0x67)
196 goto retry
; /* prefixes */
197 /* can't boost Address-size override and bound */
198 return (opcode
!= 0x62 && opcode
!= 0x67);
200 return 0; /* can't boost conditional jump */
202 /* can't boost software-interruptions */
203 return (0xc1 < opcode
&& opcode
< 0xcc) || opcode
== 0xcf;
205 /* can boost AA* and XLAT */
206 return (opcode
== 0xd4 || opcode
== 0xd5 || opcode
== 0xd7);
208 /* can boost in/out and absolute jmps */
209 return ((opcode
& 0x04) || opcode
== 0xea);
211 if ((opcode
& 0x0c) == 0 && opcode
!= 0xf1)
212 goto retry
; /* lock/rep(ne) prefix */
213 /* clear and set flags are boostable */
214 return (opcode
== 0xf5 || (0xf7 < opcode
&& opcode
< 0xfe));
216 /* segment override prefixes are boostable */
217 if (opcode
== 0x26 || opcode
== 0x36 || opcode
== 0x3e)
218 goto retry
; /* prefixes */
219 /* CS override prefix and call are not boostable */
220 return (opcode
!= 0x2e && opcode
!= 0x9a);
225 __recover_probed_insn(kprobe_opcode_t
*buf
, unsigned long addr
)
230 kp
= get_kprobe((void *)addr
);
231 faddr
= ftrace_location(addr
);
233 * Addresses inside the ftrace location are refused by
234 * arch_check_ftrace_location(). Something went terribly wrong
235 * if such an address is checked here.
237 if (WARN_ON(faddr
&& faddr
!= addr
))
240 * Use the current code if it is not modified by Kprobe
241 * and it cannot be modified by ftrace.
247 * Basically, kp->ainsn.insn has an original instruction.
248 * However, RIP-relative instruction can not do single-stepping
249 * at different place, __copy_instruction() tweaks the displacement of
250 * that instruction. In that case, we can't recover the instruction
251 * from the kp->ainsn.insn.
253 * On the other hand, in case on normal Kprobe, kp->opcode has a copy
254 * of the first byte of the probed instruction, which is overwritten
255 * by int3. And the instruction at kp->addr is not modified by kprobes
256 * except for the first byte, we can recover the original instruction
257 * from it and kp->opcode.
259 * In case of Kprobes using ftrace, we do not have a copy of
260 * the original instruction. In fact, the ftrace location might
261 * be modified at anytime and even could be in an inconsistent state.
262 * Fortunately, we know that the original code is the ideal 5-byte
265 memcpy(buf
, (void *)addr
, MAX_INSN_SIZE
* sizeof(kprobe_opcode_t
));
267 memcpy(buf
, ideal_nops
[NOP_ATOMIC5
], 5);
270 return (unsigned long)buf
;
274 * Recover the probed instruction at addr for further analysis.
275 * Caller must lock kprobes by kprobe_mutex, or disable preemption
276 * for preventing to release referencing kprobes.
277 * Returns zero if the instruction can not get recovered.
279 unsigned long recover_probed_instruction(kprobe_opcode_t
*buf
, unsigned long addr
)
281 unsigned long __addr
;
283 __addr
= __recover_optprobed_insn(buf
, addr
);
287 return __recover_probed_insn(buf
, addr
);
290 /* Check if paddr is at an instruction boundary */
291 static int can_probe(unsigned long paddr
)
293 unsigned long addr
, __addr
, offset
= 0;
295 kprobe_opcode_t buf
[MAX_INSN_SIZE
];
297 if (!kallsyms_lookup_size_offset(paddr
, NULL
, &offset
))
300 /* Decode instructions */
301 addr
= paddr
- offset
;
302 while (addr
< paddr
) {
304 * Check if the instruction has been modified by another
305 * kprobe, in which case we replace the breakpoint by the
306 * original instruction in our buffer.
307 * Also, jump optimization will change the breakpoint to
308 * relative-jump. Since the relative-jump itself is
309 * normally used, we just go through if there is no kprobe.
311 __addr
= recover_probed_instruction(buf
, addr
);
314 kernel_insn_init(&insn
, (void *)__addr
, MAX_INSN_SIZE
);
315 insn_get_length(&insn
);
318 * Another debugging subsystem might insert this breakpoint.
319 * In that case, we can't recover it.
321 if (insn
.opcode
.bytes
[0] == BREAKPOINT_INSTRUCTION
)
326 return (addr
== paddr
);
330 * Returns non-zero if opcode modifies the interrupt flag.
332 static int is_IF_modifier(kprobe_opcode_t
*insn
)
335 insn
= skip_prefixes(insn
);
340 case 0xcf: /* iret/iretd */
341 case 0x9d: /* popf/popfd */
349 * Copy an instruction and adjust the displacement if the instruction
350 * uses the %rip-relative addressing mode.
351 * If it does, Return the address of the 32-bit displacement word.
352 * If not, return null.
353 * Only applicable to 64-bit x86.
355 int __copy_instruction(u8
*dest
, u8
*src
)
358 kprobe_opcode_t buf
[MAX_INSN_SIZE
];
360 unsigned long recovered_insn
=
361 recover_probed_instruction(buf
, (unsigned long)src
);
365 kernel_insn_init(&insn
, (void *)recovered_insn
, MAX_INSN_SIZE
);
366 insn_get_length(&insn
);
367 length
= insn
.length
;
369 /* Another subsystem puts a breakpoint, failed to recover */
370 if (insn
.opcode
.bytes
[0] == BREAKPOINT_INSTRUCTION
)
372 memcpy(dest
, insn
.kaddr
, length
);
375 if (insn_rip_relative(&insn
)) {
378 kernel_insn_init(&insn
, dest
, length
);
379 insn_get_displacement(&insn
);
381 * The copied instruction uses the %rip-relative addressing
382 * mode. Adjust the displacement for the difference between
383 * the original location of this instruction and the location
384 * of the copy that will actually be run. The tricky bit here
385 * is making sure that the sign extension happens correctly in
386 * this calculation, since we need a signed 32-bit result to
387 * be sign-extended to 64 bits when it's added to the %rip
388 * value and yield the same 64-bit result that the sign-
389 * extension of the original signed 32-bit displacement would
392 newdisp
= (u8
*) src
+ (s64
) insn
.displacement
.value
- (u8
*) dest
;
393 if ((s64
) (s32
) newdisp
!= newdisp
) {
394 pr_err("Kprobes error: new displacement does not fit into s32 (%llx)\n", newdisp
);
395 pr_err("\tSrc: %p, Dest: %p, old disp: %x\n", src
, dest
, insn
.displacement
.value
);
398 disp
= (u8
*) dest
+ insn_offset_displacement(&insn
);
399 *(s32
*) disp
= (s32
) newdisp
;
405 static int arch_copy_kprobe(struct kprobe
*p
)
409 /* Copy an instruction with recovering if other optprobe modifies it.*/
410 ret
= __copy_instruction(p
->ainsn
.insn
, p
->addr
);
415 * __copy_instruction can modify the displacement of the instruction,
416 * but it doesn't affect boostable check.
418 if (can_boost(p
->ainsn
.insn
))
419 p
->ainsn
.boostable
= 0;
421 p
->ainsn
.boostable
= -1;
423 /* Check whether the instruction modifies Interrupt Flag or not */
424 p
->ainsn
.if_modifier
= is_IF_modifier(p
->ainsn
.insn
);
426 /* Also, displacement change doesn't affect the first byte */
427 p
->opcode
= p
->ainsn
.insn
[0];
432 int arch_prepare_kprobe(struct kprobe
*p
)
434 if (alternatives_text_reserved(p
->addr
, p
->addr
))
437 if (!can_probe((unsigned long)p
->addr
))
439 /* insn: must be on special executable page on x86. */
440 p
->ainsn
.insn
= get_insn_slot();
444 return arch_copy_kprobe(p
);
447 void arch_arm_kprobe(struct kprobe
*p
)
449 text_poke(p
->addr
, ((unsigned char []){BREAKPOINT_INSTRUCTION
}), 1);
452 void arch_disarm_kprobe(struct kprobe
*p
)
454 text_poke(p
->addr
, &p
->opcode
, 1);
457 void arch_remove_kprobe(struct kprobe
*p
)
460 free_insn_slot(p
->ainsn
.insn
, (p
->ainsn
.boostable
== 1));
461 p
->ainsn
.insn
= NULL
;
465 static nokprobe_inline
void
466 save_previous_kprobe(struct kprobe_ctlblk
*kcb
)
468 kcb
->prev_kprobe
.kp
= kprobe_running();
469 kcb
->prev_kprobe
.status
= kcb
->kprobe_status
;
470 kcb
->prev_kprobe
.old_flags
= kcb
->kprobe_old_flags
;
471 kcb
->prev_kprobe
.saved_flags
= kcb
->kprobe_saved_flags
;
474 static nokprobe_inline
void
475 restore_previous_kprobe(struct kprobe_ctlblk
*kcb
)
477 __this_cpu_write(current_kprobe
, kcb
->prev_kprobe
.kp
);
478 kcb
->kprobe_status
= kcb
->prev_kprobe
.status
;
479 kcb
->kprobe_old_flags
= kcb
->prev_kprobe
.old_flags
;
480 kcb
->kprobe_saved_flags
= kcb
->prev_kprobe
.saved_flags
;
483 static nokprobe_inline
void
484 set_current_kprobe(struct kprobe
*p
, struct pt_regs
*regs
,
485 struct kprobe_ctlblk
*kcb
)
487 __this_cpu_write(current_kprobe
, p
);
488 kcb
->kprobe_saved_flags
= kcb
->kprobe_old_flags
489 = (regs
->flags
& (X86_EFLAGS_TF
| X86_EFLAGS_IF
));
490 if (p
->ainsn
.if_modifier
)
491 kcb
->kprobe_saved_flags
&= ~X86_EFLAGS_IF
;
494 static nokprobe_inline
void clear_btf(void)
496 if (test_thread_flag(TIF_BLOCKSTEP
)) {
497 unsigned long debugctl
= get_debugctlmsr();
499 debugctl
&= ~DEBUGCTLMSR_BTF
;
500 update_debugctlmsr(debugctl
);
504 static nokprobe_inline
void restore_btf(void)
506 if (test_thread_flag(TIF_BLOCKSTEP
)) {
507 unsigned long debugctl
= get_debugctlmsr();
509 debugctl
|= DEBUGCTLMSR_BTF
;
510 update_debugctlmsr(debugctl
);
514 void arch_prepare_kretprobe(struct kretprobe_instance
*ri
, struct pt_regs
*regs
)
516 unsigned long *sara
= stack_addr(regs
);
518 ri
->ret_addr
= (kprobe_opcode_t
*) *sara
;
520 /* Replace the return addr with trampoline addr */
521 *sara
= (unsigned long) &kretprobe_trampoline
;
523 NOKPROBE_SYMBOL(arch_prepare_kretprobe
);
525 static void setup_singlestep(struct kprobe
*p
, struct pt_regs
*regs
,
526 struct kprobe_ctlblk
*kcb
, int reenter
)
528 if (setup_detour_execution(p
, regs
, reenter
))
531 #if !defined(CONFIG_PREEMPT)
532 if (p
->ainsn
.boostable
== 1 && !p
->post_handler
) {
533 /* Boost up -- we can execute copied instructions directly */
535 reset_current_kprobe();
537 * Reentering boosted probe doesn't reset current_kprobe,
538 * nor set current_kprobe, because it doesn't use single
541 regs
->ip
= (unsigned long)p
->ainsn
.insn
;
542 preempt_enable_no_resched();
547 save_previous_kprobe(kcb
);
548 set_current_kprobe(p
, regs
, kcb
);
549 kcb
->kprobe_status
= KPROBE_REENTER
;
551 kcb
->kprobe_status
= KPROBE_HIT_SS
;
552 /* Prepare real single stepping */
554 regs
->flags
|= X86_EFLAGS_TF
;
555 regs
->flags
&= ~X86_EFLAGS_IF
;
556 /* single step inline if the instruction is an int3 */
557 if (p
->opcode
== BREAKPOINT_INSTRUCTION
)
558 regs
->ip
= (unsigned long)p
->addr
;
560 regs
->ip
= (unsigned long)p
->ainsn
.insn
;
562 NOKPROBE_SYMBOL(setup_singlestep
);
565 * We have reentered the kprobe_handler(), since another probe was hit while
566 * within the handler. We save the original kprobes variables and just single
567 * step on the instruction of the new probe without calling any user handlers.
569 static int reenter_kprobe(struct kprobe
*p
, struct pt_regs
*regs
,
570 struct kprobe_ctlblk
*kcb
)
572 switch (kcb
->kprobe_status
) {
573 case KPROBE_HIT_SSDONE
:
574 case KPROBE_HIT_ACTIVE
:
576 kprobes_inc_nmissed_count(p
);
577 setup_singlestep(p
, regs
, kcb
, 1);
580 /* A probe has been hit in the codepath leading up to, or just
581 * after, single-stepping of a probed instruction. This entire
582 * codepath should strictly reside in .kprobes.text section.
583 * Raise a BUG or we'll continue in an endless reentering loop
584 * and eventually a stack overflow.
586 printk(KERN_WARNING
"Unrecoverable kprobe detected at %p.\n",
591 /* impossible cases */
598 NOKPROBE_SYMBOL(reenter_kprobe
);
601 * Interrupts are disabled on entry as trap3 is an interrupt gate and they
602 * remain disabled throughout this function.
604 int kprobe_int3_handler(struct pt_regs
*regs
)
606 kprobe_opcode_t
*addr
;
608 struct kprobe_ctlblk
*kcb
;
613 addr
= (kprobe_opcode_t
*)(regs
->ip
- sizeof(kprobe_opcode_t
));
615 * We don't want to be preempted for the entire
616 * duration of kprobe processing. We conditionally
617 * re-enable preemption at the end of this function,
618 * and also in reenter_kprobe() and setup_singlestep().
622 kcb
= get_kprobe_ctlblk();
623 p
= get_kprobe(addr
);
626 if (kprobe_running()) {
627 if (reenter_kprobe(p
, regs
, kcb
))
630 set_current_kprobe(p
, regs
, kcb
);
631 kcb
->kprobe_status
= KPROBE_HIT_ACTIVE
;
634 * If we have no pre-handler or it returned 0, we
635 * continue with normal processing. If we have a
636 * pre-handler and it returned non-zero, it prepped
637 * for calling the break_handler below on re-entry
638 * for jprobe processing, so get out doing nothing
641 if (!p
->pre_handler
|| !p
->pre_handler(p
, regs
))
642 setup_singlestep(p
, regs
, kcb
, 0);
645 } else if (*addr
!= BREAKPOINT_INSTRUCTION
) {
647 * The breakpoint instruction was removed right
648 * after we hit it. Another cpu has removed
649 * either a probepoint or a debugger breakpoint
650 * at this address. In either case, no further
651 * handling of this interrupt is appropriate.
652 * Back up over the (now missing) int3 and run
653 * the original instruction.
655 regs
->ip
= (unsigned long)addr
;
656 preempt_enable_no_resched();
658 } else if (kprobe_running()) {
659 p
= __this_cpu_read(current_kprobe
);
660 if (p
->break_handler
&& p
->break_handler(p
, regs
)) {
661 if (!skip_singlestep(p
, regs
, kcb
))
662 setup_singlestep(p
, regs
, kcb
, 0);
665 } /* else: not a kprobe fault; let the kernel handle it */
667 preempt_enable_no_resched();
670 NOKPROBE_SYMBOL(kprobe_int3_handler
);
673 * When a retprobed function returns, this code saves registers and
674 * calls trampoline_handler() runs, which calls the kretprobe's handler.
677 ".global kretprobe_trampoline\n"
678 ".type kretprobe_trampoline, @function\n"
679 "kretprobe_trampoline:\n"
681 /* We don't bother saving the ss register */
686 " call trampoline_handler\n"
687 /* Replace saved sp with true return address. */
688 " movq %rax, 152(%rsp)\n"
695 " call trampoline_handler\n"
696 /* Move flags to cs */
697 " movl 56(%esp), %edx\n"
698 " movl %edx, 52(%esp)\n"
699 /* Replace saved flags with true return address. */
700 " movl %eax, 56(%esp)\n"
705 ".size kretprobe_trampoline, .-kretprobe_trampoline\n"
707 NOKPROBE_SYMBOL(kretprobe_trampoline
);
708 STACK_FRAME_NON_STANDARD(kretprobe_trampoline
);
711 * Called from kretprobe_trampoline
713 __visible __used
void *trampoline_handler(struct pt_regs
*regs
)
715 struct kretprobe_instance
*ri
= NULL
;
716 struct hlist_head
*head
, empty_rp
;
717 struct hlist_node
*tmp
;
718 unsigned long flags
, orig_ret_address
= 0;
719 unsigned long trampoline_address
= (unsigned long)&kretprobe_trampoline
;
720 kprobe_opcode_t
*correct_ret_addr
= NULL
;
722 INIT_HLIST_HEAD(&empty_rp
);
723 kretprobe_hash_lock(current
, &head
, &flags
);
724 /* fixup registers */
726 regs
->cs
= __KERNEL_CS
;
728 regs
->cs
= __KERNEL_CS
| get_kernel_rpl();
731 regs
->ip
= trampoline_address
;
732 regs
->orig_ax
= ~0UL;
735 * It is possible to have multiple instances associated with a given
736 * task either because multiple functions in the call path have
737 * return probes installed on them, and/or more than one
738 * return probe was registered for a target function.
740 * We can handle this because:
741 * - instances are always pushed into the head of the list
742 * - when multiple return probes are registered for the same
743 * function, the (chronologically) first instance's ret_addr
744 * will be the real return address, and all the rest will
745 * point to kretprobe_trampoline.
747 hlist_for_each_entry_safe(ri
, tmp
, head
, hlist
) {
748 if (ri
->task
!= current
)
749 /* another task is sharing our hash bucket */
752 orig_ret_address
= (unsigned long)ri
->ret_addr
;
754 if (orig_ret_address
!= trampoline_address
)
756 * This is the real return address. Any other
757 * instances associated with this task are for
758 * other calls deeper on the call stack
763 kretprobe_assert(ri
, orig_ret_address
, trampoline_address
);
765 correct_ret_addr
= ri
->ret_addr
;
766 hlist_for_each_entry_safe(ri
, tmp
, head
, hlist
) {
767 if (ri
->task
!= current
)
768 /* another task is sharing our hash bucket */
771 orig_ret_address
= (unsigned long)ri
->ret_addr
;
772 if (ri
->rp
&& ri
->rp
->handler
) {
773 __this_cpu_write(current_kprobe
, &ri
->rp
->kp
);
774 get_kprobe_ctlblk()->kprobe_status
= KPROBE_HIT_ACTIVE
;
775 ri
->ret_addr
= correct_ret_addr
;
776 ri
->rp
->handler(ri
, regs
);
777 __this_cpu_write(current_kprobe
, NULL
);
780 recycle_rp_inst(ri
, &empty_rp
);
782 if (orig_ret_address
!= trampoline_address
)
784 * This is the real return address. Any other
785 * instances associated with this task are for
786 * other calls deeper on the call stack
791 kretprobe_hash_unlock(current
, &flags
);
793 hlist_for_each_entry_safe(ri
, tmp
, &empty_rp
, hlist
) {
794 hlist_del(&ri
->hlist
);
797 return (void *)orig_ret_address
;
799 NOKPROBE_SYMBOL(trampoline_handler
);
802 * Called after single-stepping. p->addr is the address of the
803 * instruction whose first byte has been replaced by the "int 3"
804 * instruction. To avoid the SMP problems that can occur when we
805 * temporarily put back the original opcode to single-step, we
806 * single-stepped a copy of the instruction. The address of this
807 * copy is p->ainsn.insn.
809 * This function prepares to return from the post-single-step
810 * interrupt. We have to fix up the stack as follows:
812 * 0) Except in the case of absolute or indirect jump or call instructions,
813 * the new ip is relative to the copied instruction. We need to make
814 * it relative to the original instruction.
816 * 1) If the single-stepped instruction was pushfl, then the TF and IF
817 * flags are set in the just-pushed flags, and may need to be cleared.
819 * 2) If the single-stepped instruction was a call, the return address
820 * that is atop the stack is the address following the copied instruction.
821 * We need to make it the address following the original instruction.
823 * If this is the first time we've single-stepped the instruction at
824 * this probepoint, and the instruction is boostable, boost it: add a
825 * jump instruction after the copied instruction, that jumps to the next
826 * instruction after the probepoint.
828 static void resume_execution(struct kprobe
*p
, struct pt_regs
*regs
,
829 struct kprobe_ctlblk
*kcb
)
831 unsigned long *tos
= stack_addr(regs
);
832 unsigned long copy_ip
= (unsigned long)p
->ainsn
.insn
;
833 unsigned long orig_ip
= (unsigned long)p
->addr
;
834 kprobe_opcode_t
*insn
= p
->ainsn
.insn
;
837 insn
= skip_prefixes(insn
);
839 regs
->flags
&= ~X86_EFLAGS_TF
;
841 case 0x9c: /* pushfl */
842 *tos
&= ~(X86_EFLAGS_TF
| X86_EFLAGS_IF
);
843 *tos
|= kcb
->kprobe_old_flags
;
845 case 0xc2: /* iret/ret/lret */
850 case 0xea: /* jmp absolute -- ip is correct */
851 /* ip is already adjusted, no more changes required */
852 p
->ainsn
.boostable
= 1;
854 case 0xe8: /* call relative - Fix return addr */
855 *tos
= orig_ip
+ (*tos
- copy_ip
);
858 case 0x9a: /* call absolute -- same as call absolute, indirect */
859 *tos
= orig_ip
+ (*tos
- copy_ip
);
863 if ((insn
[1] & 0x30) == 0x10) {
865 * call absolute, indirect
866 * Fix return addr; ip is correct.
867 * But this is not boostable
869 *tos
= orig_ip
+ (*tos
- copy_ip
);
871 } else if (((insn
[1] & 0x31) == 0x20) ||
872 ((insn
[1] & 0x31) == 0x21)) {
874 * jmp near and far, absolute indirect
875 * ip is correct. And this is boostable
877 p
->ainsn
.boostable
= 1;
884 if (p
->ainsn
.boostable
== 0) {
885 if ((regs
->ip
> copy_ip
) &&
886 (regs
->ip
- copy_ip
) + 5 < MAX_INSN_SIZE
) {
888 * These instructions can be executed directly if it
889 * jumps back to correct address.
891 synthesize_reljump((void *)regs
->ip
,
892 (void *)orig_ip
+ (regs
->ip
- copy_ip
));
893 p
->ainsn
.boostable
= 1;
895 p
->ainsn
.boostable
= -1;
899 regs
->ip
+= orig_ip
- copy_ip
;
904 NOKPROBE_SYMBOL(resume_execution
);
907 * Interrupts are disabled on entry as trap1 is an interrupt gate and they
908 * remain disabled throughout this function.
910 int kprobe_debug_handler(struct pt_regs
*regs
)
912 struct kprobe
*cur
= kprobe_running();
913 struct kprobe_ctlblk
*kcb
= get_kprobe_ctlblk();
918 resume_execution(cur
, regs
, kcb
);
919 regs
->flags
|= kcb
->kprobe_saved_flags
;
921 if ((kcb
->kprobe_status
!= KPROBE_REENTER
) && cur
->post_handler
) {
922 kcb
->kprobe_status
= KPROBE_HIT_SSDONE
;
923 cur
->post_handler(cur
, regs
, 0);
926 /* Restore back the original saved kprobes variables and continue. */
927 if (kcb
->kprobe_status
== KPROBE_REENTER
) {
928 restore_previous_kprobe(kcb
);
931 reset_current_kprobe();
933 preempt_enable_no_resched();
936 * if somebody else is singlestepping across a probe point, flags
937 * will have TF set, in which case, continue the remaining processing
938 * of do_debug, as if this is not a probe hit.
940 if (regs
->flags
& X86_EFLAGS_TF
)
945 NOKPROBE_SYMBOL(kprobe_debug_handler
);
947 int kprobe_fault_handler(struct pt_regs
*regs
, int trapnr
)
949 struct kprobe
*cur
= kprobe_running();
950 struct kprobe_ctlblk
*kcb
= get_kprobe_ctlblk();
952 if (unlikely(regs
->ip
== (unsigned long)cur
->ainsn
.insn
)) {
953 /* This must happen on single-stepping */
954 WARN_ON(kcb
->kprobe_status
!= KPROBE_HIT_SS
&&
955 kcb
->kprobe_status
!= KPROBE_REENTER
);
957 * We are here because the instruction being single
958 * stepped caused a page fault. We reset the current
959 * kprobe and the ip points back to the probe address
960 * and allow the page fault handler to continue as a
963 regs
->ip
= (unsigned long)cur
->addr
;
964 regs
->flags
|= kcb
->kprobe_old_flags
;
965 if (kcb
->kprobe_status
== KPROBE_REENTER
)
966 restore_previous_kprobe(kcb
);
968 reset_current_kprobe();
969 preempt_enable_no_resched();
970 } else if (kcb
->kprobe_status
== KPROBE_HIT_ACTIVE
||
971 kcb
->kprobe_status
== KPROBE_HIT_SSDONE
) {
973 * We increment the nmissed count for accounting,
974 * we can also use npre/npostfault count for accounting
975 * these specific fault cases.
977 kprobes_inc_nmissed_count(cur
);
980 * We come here because instructions in the pre/post
981 * handler caused the page_fault, this could happen
982 * if handler tries to access user space by
983 * copy_from_user(), get_user() etc. Let the
984 * user-specified handler try to fix it first.
986 if (cur
->fault_handler
&& cur
->fault_handler(cur
, regs
, trapnr
))
990 * In case the user-specified fault handler returned
991 * zero, try to fix up.
993 if (fixup_exception(regs
, trapnr
))
997 * fixup routine could not handle it,
998 * Let do_page_fault() fix it.
1004 NOKPROBE_SYMBOL(kprobe_fault_handler
);
1007 * Wrapper routine for handling exceptions.
1009 int kprobe_exceptions_notify(struct notifier_block
*self
, unsigned long val
,
1012 struct die_args
*args
= data
;
1013 int ret
= NOTIFY_DONE
;
1015 if (args
->regs
&& user_mode(args
->regs
))
1018 if (val
== DIE_GPF
) {
1020 * To be potentially processing a kprobe fault and to
1021 * trust the result from kprobe_running(), we have
1022 * be non-preemptible.
1024 if (!preemptible() && kprobe_running() &&
1025 kprobe_fault_handler(args
->regs
, args
->trapnr
))
1030 NOKPROBE_SYMBOL(kprobe_exceptions_notify
);
1032 int setjmp_pre_handler(struct kprobe
*p
, struct pt_regs
*regs
)
1034 struct jprobe
*jp
= container_of(p
, struct jprobe
, kp
);
1036 struct kprobe_ctlblk
*kcb
= get_kprobe_ctlblk();
1038 kcb
->jprobe_saved_regs
= *regs
;
1039 kcb
->jprobe_saved_sp
= stack_addr(regs
);
1040 addr
= (unsigned long)(kcb
->jprobe_saved_sp
);
1043 * As Linus pointed out, gcc assumes that the callee
1044 * owns the argument space and could overwrite it, e.g.
1045 * tailcall optimization. So, to be absolutely safe
1046 * we also save and restore enough stack bytes to cover
1047 * the argument area.
1049 memcpy(kcb
->jprobes_stack
, (kprobe_opcode_t
*)addr
,
1050 MIN_STACK_SIZE(addr
));
1051 regs
->flags
&= ~X86_EFLAGS_IF
;
1052 trace_hardirqs_off();
1053 regs
->ip
= (unsigned long)(jp
->entry
);
1056 * jprobes use jprobe_return() which skips the normal return
1057 * path of the function, and this messes up the accounting of the
1058 * function graph tracer to get messed up.
1060 * Pause function graph tracing while performing the jprobe function.
1062 pause_graph_tracing();
1065 NOKPROBE_SYMBOL(setjmp_pre_handler
);
1067 void jprobe_return(void)
1069 struct kprobe_ctlblk
*kcb
= get_kprobe_ctlblk();
1072 #ifdef CONFIG_X86_64
1073 " xchg %%rbx,%%rsp \n"
1075 " xchgl %%ebx,%%esp \n"
1078 " .globl jprobe_return_end\n"
1079 " jprobe_return_end: \n"
1081 (kcb
->jprobe_saved_sp
):"memory");
1083 NOKPROBE_SYMBOL(jprobe_return
);
1084 NOKPROBE_SYMBOL(jprobe_return_end
);
1086 int longjmp_break_handler(struct kprobe
*p
, struct pt_regs
*regs
)
1088 struct kprobe_ctlblk
*kcb
= get_kprobe_ctlblk();
1089 u8
*addr
= (u8
*) (regs
->ip
- 1);
1090 struct jprobe
*jp
= container_of(p
, struct jprobe
, kp
);
1091 void *saved_sp
= kcb
->jprobe_saved_sp
;
1093 if ((addr
> (u8
*) jprobe_return
) &&
1094 (addr
< (u8
*) jprobe_return_end
)) {
1095 if (stack_addr(regs
) != saved_sp
) {
1096 struct pt_regs
*saved_regs
= &kcb
->jprobe_saved_regs
;
1098 "current sp %p does not match saved sp %p\n",
1099 stack_addr(regs
), saved_sp
);
1100 printk(KERN_ERR
"Saved registers for jprobe %p\n", jp
);
1101 show_regs(saved_regs
);
1102 printk(KERN_ERR
"Current registers\n");
1106 /* It's OK to start function graph tracing again */
1107 unpause_graph_tracing();
1108 *regs
= kcb
->jprobe_saved_regs
;
1109 memcpy(saved_sp
, kcb
->jprobes_stack
, MIN_STACK_SIZE(saved_sp
));
1110 preempt_enable_no_resched();
1115 NOKPROBE_SYMBOL(longjmp_break_handler
);
1117 bool arch_within_kprobe_blacklist(unsigned long addr
)
1119 return (addr
>= (unsigned long)__kprobes_text_start
&&
1120 addr
< (unsigned long)__kprobes_text_end
) ||
1121 (addr
>= (unsigned long)__entry_text_start
&&
1122 addr
< (unsigned long)__entry_text_end
);
1125 int __init
arch_init_kprobes(void)
1130 int arch_trampoline_kprobe(struct kprobe
*p
)