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/sched/debug.h>
49 #include <linux/extable.h>
50 #include <linux/kdebug.h>
51 #include <linux/kallsyms.h>
52 #include <linux/ftrace.h>
53 #include <linux/frame.h>
54 #include <linux/kasan.h>
55 #include <linux/moduleloader.h>
57 #include <asm/text-patching.h>
58 #include <asm/cacheflush.h>
60 #include <asm/pgtable.h>
61 #include <linux/uaccess.h>
62 #include <asm/alternative.h>
64 #include <asm/debugreg.h>
65 #include <asm/set_memory.h>
66 #include <asm/sections.h>
70 void jprobe_return_end(void);
72 DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
73 DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
75 #define stack_addr(regs) ((unsigned long *)kernel_stack_pointer(regs))
77 #define W(row, b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, ba, bb, bc, bd, be, bf)\
78 (((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) | \
79 (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) | \
80 (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) | \
81 (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf)) \
84 * Undefined/reserved opcodes, conditional jump, Opcode Extension
85 * Groups, and some special opcodes can not boost.
86 * This is non-const and volatile to keep gcc from statically
87 * optimizing it out, as variable_test_bit makes gcc think only
88 * *(unsigned long*) is used.
90 static volatile u32 twobyte_is_boostable[256 / 32] = {
91 /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */
92 /* ---------------------------------------------- */
93 W(0x00, 0, 0, 1, 1, 0, 0, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0) | /* 00 */
94 W(0x10, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1) , /* 10 */
95 W(0x20, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 20 */
96 W(0x30, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 30 */
97 W(0x40, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 40 */
98 W(0x50, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 50 */
99 W(0x60, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1) | /* 60 */
100 W(0x70, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1) , /* 70 */
101 W(0x80, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 80 */
102 W(0x90, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 90 */
103 W(0xa0, 1, 1, 0, 1, 1, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* a0 */
104 W(0xb0, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1) , /* b0 */
105 W(0xc0, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1) | /* c0 */
106 W(0xd0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) , /* d0 */
107 W(0xe0, 0, 1, 1, 0, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* e0 */
108 W(0xf0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 1, 0, 1, 1, 1, 0) /* f0 */
109 /* ----------------------------------------------- */
110 /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */
114 struct kretprobe_blackpoint kretprobe_blacklist[] = {
115 {"__switch_to", }, /* This function switches only current task, but
116 doesn't switch kernel stack.*/
117 {NULL, NULL} /* Terminator */
120 const int kretprobe_blacklist_size = ARRAY_SIZE(kretprobe_blacklist);
122 static nokprobe_inline void
123 __synthesize_relative_insn(void *from, void *to, u8 op)
125 struct __arch_relative_insn {
130 insn = (struct __arch_relative_insn *)from;
131 insn->raddr = (s32)((long)(to) - ((long)(from) + 5));
135 /* Insert a jump instruction at address 'from', which jumps to address 'to'.*/
136 void synthesize_reljump(void *from, void *to)
138 __synthesize_relative_insn(from, to, RELATIVEJUMP_OPCODE);
140 NOKPROBE_SYMBOL(synthesize_reljump);
142 /* Insert a call instruction at address 'from', which calls address 'to'.*/
143 void synthesize_relcall(void *from, void *to)
145 __synthesize_relative_insn(from, to, RELATIVECALL_OPCODE);
147 NOKPROBE_SYMBOL(synthesize_relcall);
150 * Skip the prefixes of the instruction.
152 static kprobe_opcode_t *skip_prefixes(kprobe_opcode_t *insn)
156 attr = inat_get_opcode_attribute((insn_byte_t)*insn);
157 while (inat_is_legacy_prefix(attr)) {
159 attr = inat_get_opcode_attribute((insn_byte_t)*insn);
162 if (inat_is_rex_prefix(attr))
167 NOKPROBE_SYMBOL(skip_prefixes);
170 * Returns non-zero if INSN is boostable.
171 * RIP relative instructions are adjusted at copying time in 64 bits mode
173 int can_boost(struct insn *insn, void *addr)
175 kprobe_opcode_t opcode;
179 if (search_exception_tables((unsigned long)addr))
180 return 0; /* Page fault may occur on this address. */
182 /* 2nd-byte opcode */
183 if (insn->opcode.nbytes == 2)
184 return test_bit(insn->opcode.bytes[1],
185 (unsigned long *)twobyte_is_boostable);
187 if (insn->opcode.nbytes != 1)
190 for_each_insn_prefix(insn, i, prefix) {
193 attr = inat_get_opcode_attribute(prefix);
194 /* Can't boost Address-size override prefix and CS override prefix */
195 if (prefix == 0x2e || inat_is_address_size_prefix(attr))
199 opcode = insn->opcode.bytes[0];
201 switch (opcode & 0xf0) {
203 /* can't boost "bound" */
204 return (opcode != 0x62);
206 return 0; /* can't boost conditional jump */
208 return opcode != 0x9a; /* can't boost call far */
210 /* can't boost software-interruptions */
211 return (0xc1 < opcode && opcode < 0xcc) || opcode == 0xcf;
213 /* can boost AA* and XLAT */
214 return (opcode == 0xd4 || opcode == 0xd5 || opcode == 0xd7);
216 /* can boost in/out and absolute jmps */
217 return ((opcode & 0x04) || opcode == 0xea);
219 /* clear and set flags are boostable */
220 return (opcode == 0xf5 || (0xf7 < opcode && opcode < 0xfe));
222 /* call is not boostable */
223 return opcode != 0x9a;
228 __recover_probed_insn(kprobe_opcode_t *buf, unsigned long addr)
233 kp = get_kprobe((void *)addr);
234 faddr = ftrace_location(addr);
236 * Addresses inside the ftrace location are refused by
237 * arch_check_ftrace_location(). Something went terribly wrong
238 * if such an address is checked here.
240 if (WARN_ON(faddr && faddr != addr))
243 * Use the current code if it is not modified by Kprobe
244 * and it cannot be modified by ftrace.
250 * Basically, kp->ainsn.insn has an original instruction.
251 * However, RIP-relative instruction can not do single-stepping
252 * at different place, __copy_instruction() tweaks the displacement of
253 * that instruction. In that case, we can't recover the instruction
254 * from the kp->ainsn.insn.
256 * On the other hand, in case on normal Kprobe, kp->opcode has a copy
257 * of the first byte of the probed instruction, which is overwritten
258 * by int3. And the instruction at kp->addr is not modified by kprobes
259 * except for the first byte, we can recover the original instruction
260 * from it and kp->opcode.
262 * In case of Kprobes using ftrace, we do not have a copy of
263 * the original instruction. In fact, the ftrace location might
264 * be modified at anytime and even could be in an inconsistent state.
265 * Fortunately, we know that the original code is the ideal 5-byte
268 if (probe_kernel_read(buf, (void *)addr,
269 MAX_INSN_SIZE * sizeof(kprobe_opcode_t)))
273 memcpy(buf, ideal_nops[NOP_ATOMIC5], 5);
276 return (unsigned long)buf;
280 * Recover the probed instruction at addr for further analysis.
281 * Caller must lock kprobes by kprobe_mutex, or disable preemption
282 * for preventing to release referencing kprobes.
283 * Returns zero if the instruction can not get recovered (or access failed).
285 unsigned long recover_probed_instruction(kprobe_opcode_t *buf, unsigned long addr)
287 unsigned long __addr;
289 __addr = __recover_optprobed_insn(buf, addr);
293 return __recover_probed_insn(buf, addr);
296 /* Check if paddr is at an instruction boundary */
297 static int can_probe(unsigned long paddr)
299 unsigned long addr, __addr, offset = 0;
301 kprobe_opcode_t buf[MAX_INSN_SIZE];
303 if (!kallsyms_lookup_size_offset(paddr, NULL, &offset))
306 /* Decode instructions */
307 addr = paddr - offset;
308 while (addr < paddr) {
310 * Check if the instruction has been modified by another
311 * kprobe, in which case we replace the breakpoint by the
312 * original instruction in our buffer.
313 * Also, jump optimization will change the breakpoint to
314 * relative-jump. Since the relative-jump itself is
315 * normally used, we just go through if there is no kprobe.
317 __addr = recover_probed_instruction(buf, addr);
320 kernel_insn_init(&insn, (void *)__addr, MAX_INSN_SIZE);
321 insn_get_length(&insn);
324 * Another debugging subsystem might insert this breakpoint.
325 * In that case, we can't recover it.
327 if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION)
332 return (addr == paddr);
336 * Returns non-zero if opcode modifies the interrupt flag.
338 static int is_IF_modifier(kprobe_opcode_t *insn)
341 insn = skip_prefixes(insn);
346 case 0xcf: /* iret/iretd */
347 case 0x9d: /* popf/popfd */
355 * Copy an instruction with recovering modified instruction by kprobes
356 * and adjust the displacement if the instruction uses the %rip-relative
358 * This returns the length of copied instruction, or 0 if it has an error.
360 int __copy_instruction(u8 *dest, u8 *src, struct insn *insn)
362 kprobe_opcode_t buf[MAX_INSN_SIZE];
363 unsigned long recovered_insn =
364 recover_probed_instruction(buf, (unsigned long)src);
366 if (!recovered_insn || !insn)
369 /* This can access kernel text if given address is not recovered */
370 if (probe_kernel_read(dest, (void *)recovered_insn, MAX_INSN_SIZE))
373 kernel_insn_init(insn, dest, MAX_INSN_SIZE);
374 insn_get_length(insn);
376 /* Another subsystem puts a breakpoint, failed to recover */
377 if (insn->opcode.bytes[0] == BREAKPOINT_INSTRUCTION)
380 /* We should not singlestep on the exception masking instructions */
381 if (insn_masking_exception(insn))
385 /* Only x86_64 has RIP relative instructions */
386 if (insn_rip_relative(insn)) {
390 * The copied instruction uses the %rip-relative addressing
391 * mode. Adjust the displacement for the difference between
392 * the original location of this instruction and the location
393 * of the copy that will actually be run. The tricky bit here
394 * is making sure that the sign extension happens correctly in
395 * this calculation, since we need a signed 32-bit result to
396 * be sign-extended to 64 bits when it's added to the %rip
397 * value and yield the same 64-bit result that the sign-
398 * extension of the original signed 32-bit displacement would
401 newdisp = (u8 *) src + (s64) insn->displacement.value
403 if ((s64) (s32) newdisp != newdisp) {
404 pr_err("Kprobes error: new displacement does not fit into s32 (%llx)\n", newdisp);
407 disp = (u8 *) dest + insn_offset_displacement(insn);
408 *(s32 *) disp = (s32) newdisp;
414 /* Prepare reljump right after instruction to boost */
415 static void prepare_boost(struct kprobe *p, struct insn *insn)
417 if (can_boost(insn, p->addr) &&
418 MAX_INSN_SIZE - insn->length >= RELATIVEJUMP_SIZE) {
420 * These instructions can be executed directly if it
421 * jumps back to correct address.
423 synthesize_reljump(p->ainsn.insn + insn->length,
424 p->addr + insn->length);
425 p->ainsn.boostable = true;
427 p->ainsn.boostable = false;
431 /* Recover page to RW mode before releasing it */
432 void free_insn_page(void *page)
434 set_memory_nx((unsigned long)page & PAGE_MASK, 1);
435 set_memory_rw((unsigned long)page & PAGE_MASK, 1);
436 module_memfree(page);
439 static int arch_copy_kprobe(struct kprobe *p)
444 set_memory_rw((unsigned long)p->ainsn.insn & PAGE_MASK, 1);
446 /* Copy an instruction with recovering if other optprobe modifies it.*/
447 len = __copy_instruction(p->ainsn.insn, p->addr, &insn);
452 * __copy_instruction can modify the displacement of the instruction,
453 * but it doesn't affect boostable check.
455 prepare_boost(p, &insn);
457 set_memory_ro((unsigned long)p->ainsn.insn & PAGE_MASK, 1);
459 /* Check whether the instruction modifies Interrupt Flag or not */
460 p->ainsn.if_modifier = is_IF_modifier(p->ainsn.insn);
462 /* Also, displacement change doesn't affect the first byte */
463 p->opcode = p->ainsn.insn[0];
468 int arch_prepare_kprobe(struct kprobe *p)
472 if (alternatives_text_reserved(p->addr, p->addr))
475 if (!can_probe((unsigned long)p->addr))
477 /* insn: must be on special executable page on x86. */
478 p->ainsn.insn = get_insn_slot();
482 ret = arch_copy_kprobe(p);
484 free_insn_slot(p->ainsn.insn, 0);
485 p->ainsn.insn = NULL;
491 void arch_arm_kprobe(struct kprobe *p)
493 text_poke(p->addr, ((unsigned char []){BREAKPOINT_INSTRUCTION}), 1);
496 void arch_disarm_kprobe(struct kprobe *p)
498 text_poke(p->addr, &p->opcode, 1);
501 void arch_remove_kprobe(struct kprobe *p)
504 free_insn_slot(p->ainsn.insn, p->ainsn.boostable);
505 p->ainsn.insn = NULL;
509 static nokprobe_inline void
510 save_previous_kprobe(struct kprobe_ctlblk *kcb)
512 kcb->prev_kprobe.kp = kprobe_running();
513 kcb->prev_kprobe.status = kcb->kprobe_status;
514 kcb->prev_kprobe.old_flags = kcb->kprobe_old_flags;
515 kcb->prev_kprobe.saved_flags = kcb->kprobe_saved_flags;
518 static nokprobe_inline void
519 restore_previous_kprobe(struct kprobe_ctlblk *kcb)
521 __this_cpu_write(current_kprobe, kcb->prev_kprobe.kp);
522 kcb->kprobe_status = kcb->prev_kprobe.status;
523 kcb->kprobe_old_flags = kcb->prev_kprobe.old_flags;
524 kcb->kprobe_saved_flags = kcb->prev_kprobe.saved_flags;
527 static nokprobe_inline void
528 set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
529 struct kprobe_ctlblk *kcb)
531 __this_cpu_write(current_kprobe, p);
532 kcb->kprobe_saved_flags = kcb->kprobe_old_flags
533 = (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF));
534 if (p->ainsn.if_modifier)
535 kcb->kprobe_saved_flags &= ~X86_EFLAGS_IF;
538 static nokprobe_inline void clear_btf(void)
540 if (test_thread_flag(TIF_BLOCKSTEP)) {
541 unsigned long debugctl = get_debugctlmsr();
543 debugctl &= ~DEBUGCTLMSR_BTF;
544 update_debugctlmsr(debugctl);
548 static nokprobe_inline void restore_btf(void)
550 if (test_thread_flag(TIF_BLOCKSTEP)) {
551 unsigned long debugctl = get_debugctlmsr();
553 debugctl |= DEBUGCTLMSR_BTF;
554 update_debugctlmsr(debugctl);
558 void arch_prepare_kretprobe(struct kretprobe_instance *ri, struct pt_regs *regs)
560 unsigned long *sara = stack_addr(regs);
562 ri->ret_addr = (kprobe_opcode_t *) *sara;
565 /* Replace the return addr with trampoline addr */
566 *sara = (unsigned long) &kretprobe_trampoline;
568 NOKPROBE_SYMBOL(arch_prepare_kretprobe);
570 static void setup_singlestep(struct kprobe *p, struct pt_regs *regs,
571 struct kprobe_ctlblk *kcb, int reenter)
573 if (setup_detour_execution(p, regs, reenter))
576 #if !defined(CONFIG_PREEMPT)
577 if (p->ainsn.boostable && !p->post_handler) {
578 /* Boost up -- we can execute copied instructions directly */
580 reset_current_kprobe();
582 * Reentering boosted probe doesn't reset current_kprobe,
583 * nor set current_kprobe, because it doesn't use single
586 regs->ip = (unsigned long)p->ainsn.insn;
587 preempt_enable_no_resched();
592 save_previous_kprobe(kcb);
593 set_current_kprobe(p, regs, kcb);
594 kcb->kprobe_status = KPROBE_REENTER;
596 kcb->kprobe_status = KPROBE_HIT_SS;
597 /* Prepare real single stepping */
599 regs->flags |= X86_EFLAGS_TF;
600 regs->flags &= ~X86_EFLAGS_IF;
601 /* single step inline if the instruction is an int3 */
602 if (p->opcode == BREAKPOINT_INSTRUCTION)
603 regs->ip = (unsigned long)p->addr;
605 regs->ip = (unsigned long)p->ainsn.insn;
607 NOKPROBE_SYMBOL(setup_singlestep);
610 * We have reentered the kprobe_handler(), since another probe was hit while
611 * within the handler. We save the original kprobes variables and just single
612 * step on the instruction of the new probe without calling any user handlers.
614 static int reenter_kprobe(struct kprobe *p, struct pt_regs *regs,
615 struct kprobe_ctlblk *kcb)
617 switch (kcb->kprobe_status) {
618 case KPROBE_HIT_SSDONE:
619 case KPROBE_HIT_ACTIVE:
621 kprobes_inc_nmissed_count(p);
622 setup_singlestep(p, regs, kcb, 1);
625 /* A probe has been hit in the codepath leading up to, or just
626 * after, single-stepping of a probed instruction. This entire
627 * codepath should strictly reside in .kprobes.text section.
628 * Raise a BUG or we'll continue in an endless reentering loop
629 * and eventually a stack overflow.
631 pr_err("Unrecoverable kprobe detected.\n");
635 /* impossible cases */
642 NOKPROBE_SYMBOL(reenter_kprobe);
645 * Interrupts are disabled on entry as trap3 is an interrupt gate and they
646 * remain disabled throughout this function.
648 int kprobe_int3_handler(struct pt_regs *regs)
650 kprobe_opcode_t *addr;
652 struct kprobe_ctlblk *kcb;
657 addr = (kprobe_opcode_t *)(regs->ip - sizeof(kprobe_opcode_t));
659 * We don't want to be preempted for the entire
660 * duration of kprobe processing. We conditionally
661 * re-enable preemption at the end of this function,
662 * and also in reenter_kprobe() and setup_singlestep().
666 kcb = get_kprobe_ctlblk();
667 p = get_kprobe(addr);
670 if (kprobe_running()) {
671 if (reenter_kprobe(p, regs, kcb))
674 set_current_kprobe(p, regs, kcb);
675 kcb->kprobe_status = KPROBE_HIT_ACTIVE;
678 * If we have no pre-handler or it returned 0, we
679 * continue with normal processing. If we have a
680 * pre-handler and it returned non-zero, it prepped
681 * for calling the break_handler below on re-entry
682 * for jprobe processing, so get out doing nothing
685 if (!p->pre_handler || !p->pre_handler(p, regs))
686 setup_singlestep(p, regs, kcb, 0);
689 } else if (*addr != BREAKPOINT_INSTRUCTION) {
691 * The breakpoint instruction was removed right
692 * after we hit it. Another cpu has removed
693 * either a probepoint or a debugger breakpoint
694 * at this address. In either case, no further
695 * handling of this interrupt is appropriate.
696 * Back up over the (now missing) int3 and run
697 * the original instruction.
699 regs->ip = (unsigned long)addr;
700 preempt_enable_no_resched();
702 } else if (kprobe_running()) {
703 p = __this_cpu_read(current_kprobe);
704 if (p->break_handler && p->break_handler(p, regs)) {
705 if (!skip_singlestep(p, regs, kcb))
706 setup_singlestep(p, regs, kcb, 0);
709 } /* else: not a kprobe fault; let the kernel handle it */
711 preempt_enable_no_resched();
714 NOKPROBE_SYMBOL(kprobe_int3_handler);
717 * When a retprobed function returns, this code saves registers and
718 * calls trampoline_handler() runs, which calls the kretprobe's handler.
721 ".global kretprobe_trampoline\n"
722 ".type kretprobe_trampoline, @function\n"
723 "kretprobe_trampoline:\n"
725 /* We don't bother saving the ss register */
730 " call trampoline_handler\n"
731 /* Replace saved sp with true return address. */
732 " movq %rax, 152(%rsp)\n"
739 " call trampoline_handler\n"
740 /* Move flags to cs */
741 " movl 56(%esp), %edx\n"
742 " movl %edx, 52(%esp)\n"
743 /* Replace saved flags with true return address. */
744 " movl %eax, 56(%esp)\n"
749 ".size kretprobe_trampoline, .-kretprobe_trampoline\n"
751 NOKPROBE_SYMBOL(kretprobe_trampoline);
752 STACK_FRAME_NON_STANDARD(kretprobe_trampoline);
755 * Called from kretprobe_trampoline
757 __visible __used void *trampoline_handler(struct pt_regs *regs)
759 struct kretprobe_instance *ri = NULL;
760 struct hlist_head *head, empty_rp;
761 struct hlist_node *tmp;
762 unsigned long flags, orig_ret_address = 0;
763 unsigned long trampoline_address = (unsigned long)&kretprobe_trampoline;
764 kprobe_opcode_t *correct_ret_addr = NULL;
766 bool skipped = false;
769 * Set a dummy kprobe for avoiding kretprobe recursion.
770 * Since kretprobe never run in kprobe handler, kprobe must not
771 * be running at this point.
775 INIT_HLIST_HEAD(&empty_rp);
776 kretprobe_hash_lock(current, &head, &flags);
777 /* fixup registers */
779 regs->cs = __KERNEL_CS;
780 /* On x86-64, we use pt_regs->sp for return address holder. */
781 frame_pointer = ®s->sp;
783 regs->cs = __KERNEL_CS | get_kernel_rpl();
785 /* On x86-32, we use pt_regs->flags for return address holder. */
786 frame_pointer = ®s->flags;
788 regs->ip = trampoline_address;
789 regs->orig_ax = ~0UL;
792 * It is possible to have multiple instances associated with a given
793 * task either because multiple functions in the call path have
794 * return probes installed on them, and/or more than one
795 * return probe was registered for a target function.
797 * We can handle this because:
798 * - instances are always pushed into the head of the list
799 * - when multiple return probes are registered for the same
800 * function, the (chronologically) first instance's ret_addr
801 * will be the real return address, and all the rest will
802 * point to kretprobe_trampoline.
804 hlist_for_each_entry(ri, head, hlist) {
805 if (ri->task != current)
806 /* another task is sharing our hash bucket */
809 * Return probes must be pushed on this hash list correct
810 * order (same as return order) so that it can be poped
811 * correctly. However, if we find it is pushed it incorrect
812 * order, this means we find a function which should not be
813 * probed, because the wrong order entry is pushed on the
814 * path of processing other kretprobe itself.
816 if (ri->fp != frame_pointer) {
818 pr_warn("kretprobe is stacked incorrectly. Trying to fixup.\n");
823 orig_ret_address = (unsigned long)ri->ret_addr;
825 pr_warn("%ps must be blacklisted because of incorrect kretprobe order\n",
828 if (orig_ret_address != trampoline_address)
830 * This is the real return address. Any other
831 * instances associated with this task are for
832 * other calls deeper on the call stack
837 kretprobe_assert(ri, orig_ret_address, trampoline_address);
839 correct_ret_addr = ri->ret_addr;
840 hlist_for_each_entry_safe(ri, tmp, head, hlist) {
841 if (ri->task != current)
842 /* another task is sharing our hash bucket */
844 if (ri->fp != frame_pointer)
847 orig_ret_address = (unsigned long)ri->ret_addr;
848 if (ri->rp && ri->rp->handler) {
849 __this_cpu_write(current_kprobe, &ri->rp->kp);
850 ri->ret_addr = correct_ret_addr;
851 ri->rp->handler(ri, regs);
852 __this_cpu_write(current_kprobe, &kprobe_busy);
855 recycle_rp_inst(ri, &empty_rp);
857 if (orig_ret_address != trampoline_address)
859 * This is the real return address. Any other
860 * instances associated with this task are for
861 * other calls deeper on the call stack
866 kretprobe_hash_unlock(current, &flags);
870 hlist_for_each_entry_safe(ri, tmp, &empty_rp, hlist) {
871 hlist_del(&ri->hlist);
874 return (void *)orig_ret_address;
876 NOKPROBE_SYMBOL(trampoline_handler);
879 * Called after single-stepping. p->addr is the address of the
880 * instruction whose first byte has been replaced by the "int 3"
881 * instruction. To avoid the SMP problems that can occur when we
882 * temporarily put back the original opcode to single-step, we
883 * single-stepped a copy of the instruction. The address of this
884 * copy is p->ainsn.insn.
886 * This function prepares to return from the post-single-step
887 * interrupt. We have to fix up the stack as follows:
889 * 0) Except in the case of absolute or indirect jump or call instructions,
890 * the new ip is relative to the copied instruction. We need to make
891 * it relative to the original instruction.
893 * 1) If the single-stepped instruction was pushfl, then the TF and IF
894 * flags are set in the just-pushed flags, and may need to be cleared.
896 * 2) If the single-stepped instruction was a call, the return address
897 * that is atop the stack is the address following the copied instruction.
898 * We need to make it the address following the original instruction.
900 * If this is the first time we've single-stepped the instruction at
901 * this probepoint, and the instruction is boostable, boost it: add a
902 * jump instruction after the copied instruction, that jumps to the next
903 * instruction after the probepoint.
905 static void resume_execution(struct kprobe *p, struct pt_regs *regs,
906 struct kprobe_ctlblk *kcb)
908 unsigned long *tos = stack_addr(regs);
909 unsigned long copy_ip = (unsigned long)p->ainsn.insn;
910 unsigned long orig_ip = (unsigned long)p->addr;
911 kprobe_opcode_t *insn = p->ainsn.insn;
914 insn = skip_prefixes(insn);
916 regs->flags &= ~X86_EFLAGS_TF;
918 case 0x9c: /* pushfl */
919 *tos &= ~(X86_EFLAGS_TF | X86_EFLAGS_IF);
920 *tos |= kcb->kprobe_old_flags;
922 case 0xc2: /* iret/ret/lret */
927 case 0xea: /* jmp absolute -- ip is correct */
928 /* ip is already adjusted, no more changes required */
929 p->ainsn.boostable = true;
931 case 0xe8: /* call relative - Fix return addr */
932 *tos = orig_ip + (*tos - copy_ip);
935 case 0x9a: /* call absolute -- same as call absolute, indirect */
936 *tos = orig_ip + (*tos - copy_ip);
940 if ((insn[1] & 0x30) == 0x10) {
942 * call absolute, indirect
943 * Fix return addr; ip is correct.
944 * But this is not boostable
946 *tos = orig_ip + (*tos - copy_ip);
948 } else if (((insn[1] & 0x31) == 0x20) ||
949 ((insn[1] & 0x31) == 0x21)) {
951 * jmp near and far, absolute indirect
952 * ip is correct. And this is boostable
954 p->ainsn.boostable = true;
961 regs->ip += orig_ip - copy_ip;
966 NOKPROBE_SYMBOL(resume_execution);
969 * Interrupts are disabled on entry as trap1 is an interrupt gate and they
970 * remain disabled throughout this function.
972 int kprobe_debug_handler(struct pt_regs *regs)
974 struct kprobe *cur = kprobe_running();
975 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
980 resume_execution(cur, regs, kcb);
981 regs->flags |= kcb->kprobe_saved_flags;
983 if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
984 kcb->kprobe_status = KPROBE_HIT_SSDONE;
985 cur->post_handler(cur, regs, 0);
988 /* Restore back the original saved kprobes variables and continue. */
989 if (kcb->kprobe_status == KPROBE_REENTER) {
990 restore_previous_kprobe(kcb);
993 reset_current_kprobe();
995 preempt_enable_no_resched();
998 * if somebody else is singlestepping across a probe point, flags
999 * will have TF set, in which case, continue the remaining processing
1000 * of do_debug, as if this is not a probe hit.
1002 if (regs->flags & X86_EFLAGS_TF)
1007 NOKPROBE_SYMBOL(kprobe_debug_handler);
1009 int kprobe_fault_handler(struct pt_regs *regs, int trapnr)
1011 struct kprobe *cur = kprobe_running();
1012 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1014 if (unlikely(regs->ip == (unsigned long)cur->ainsn.insn)) {
1015 /* This must happen on single-stepping */
1016 WARN_ON(kcb->kprobe_status != KPROBE_HIT_SS &&
1017 kcb->kprobe_status != KPROBE_REENTER);
1019 * We are here because the instruction being single
1020 * stepped caused a page fault. We reset the current
1021 * kprobe and the ip points back to the probe address
1022 * and allow the page fault handler to continue as a
1023 * normal page fault.
1025 regs->ip = (unsigned long)cur->addr;
1027 * Trap flag (TF) has been set here because this fault
1028 * happened where the single stepping will be done.
1029 * So clear it by resetting the current kprobe:
1031 regs->flags &= ~X86_EFLAGS_TF;
1033 * Since the single step (trap) has been cancelled,
1034 * we need to restore BTF here.
1039 * If the TF flag was set before the kprobe hit,
1042 regs->flags |= kcb->kprobe_old_flags;
1044 if (kcb->kprobe_status == KPROBE_REENTER)
1045 restore_previous_kprobe(kcb);
1047 reset_current_kprobe();
1048 preempt_enable_no_resched();
1049 } else if (kcb->kprobe_status == KPROBE_HIT_ACTIVE ||
1050 kcb->kprobe_status == KPROBE_HIT_SSDONE) {
1052 * We increment the nmissed count for accounting,
1053 * we can also use npre/npostfault count for accounting
1054 * these specific fault cases.
1056 kprobes_inc_nmissed_count(cur);
1059 * We come here because instructions in the pre/post
1060 * handler caused the page_fault, this could happen
1061 * if handler tries to access user space by
1062 * copy_from_user(), get_user() etc. Let the
1063 * user-specified handler try to fix it first.
1065 if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
1069 * In case the user-specified fault handler returned
1070 * zero, try to fix up.
1072 if (fixup_exception(regs, trapnr))
1076 * fixup routine could not handle it,
1077 * Let do_page_fault() fix it.
1083 NOKPROBE_SYMBOL(kprobe_fault_handler);
1086 * Wrapper routine for handling exceptions.
1088 int kprobe_exceptions_notify(struct notifier_block *self, unsigned long val,
1091 struct die_args *args = data;
1092 int ret = NOTIFY_DONE;
1094 if (args->regs && user_mode(args->regs))
1097 if (val == DIE_GPF) {
1099 * To be potentially processing a kprobe fault and to
1100 * trust the result from kprobe_running(), we have
1101 * be non-preemptible.
1103 if (!preemptible() && kprobe_running() &&
1104 kprobe_fault_handler(args->regs, args->trapnr))
1109 NOKPROBE_SYMBOL(kprobe_exceptions_notify);
1111 int setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
1113 struct jprobe *jp = container_of(p, struct jprobe, kp);
1115 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1117 kcb->jprobe_saved_regs = *regs;
1118 kcb->jprobe_saved_sp = stack_addr(regs);
1119 addr = (unsigned long)(kcb->jprobe_saved_sp);
1122 * As Linus pointed out, gcc assumes that the callee
1123 * owns the argument space and could overwrite it, e.g.
1124 * tailcall optimization. So, to be absolutely safe
1125 * we also save and restore enough stack bytes to cover
1126 * the argument area.
1127 * Use __memcpy() to avoid KASAN stack out-of-bounds reports as we copy
1128 * raw stack chunk with redzones:
1130 __memcpy(kcb->jprobes_stack, (kprobe_opcode_t *)addr, MIN_STACK_SIZE(addr));
1131 regs->ip = (unsigned long)(jp->entry);
1134 * jprobes use jprobe_return() which skips the normal return
1135 * path of the function, and this messes up the accounting of the
1136 * function graph tracer to get messed up.
1138 * Pause function graph tracing while performing the jprobe function.
1140 pause_graph_tracing();
1143 NOKPROBE_SYMBOL(setjmp_pre_handler);
1145 void jprobe_return(void)
1147 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1149 /* Unpoison stack redzones in the frames we are going to jump over. */
1150 kasan_unpoison_stack_above_sp_to(kcb->jprobe_saved_sp);
1153 #ifdef CONFIG_X86_64
1154 " xchg %%rbx,%%rsp \n"
1156 " xchgl %%ebx,%%esp \n"
1159 " .globl jprobe_return_end\n"
1160 " jprobe_return_end: \n"
1162 (kcb->jprobe_saved_sp):"memory");
1164 NOKPROBE_SYMBOL(jprobe_return);
1165 NOKPROBE_SYMBOL(jprobe_return_end);
1167 int longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
1169 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1170 u8 *addr = (u8 *) (regs->ip - 1);
1171 struct jprobe *jp = container_of(p, struct jprobe, kp);
1172 void *saved_sp = kcb->jprobe_saved_sp;
1174 if ((addr > (u8 *) jprobe_return) &&
1175 (addr < (u8 *) jprobe_return_end)) {
1176 if (stack_addr(regs) != saved_sp) {
1177 struct pt_regs *saved_regs = &kcb->jprobe_saved_regs;
1179 "current sp %p does not match saved sp %p\n",
1180 stack_addr(regs), saved_sp);
1181 printk(KERN_ERR "Saved registers for jprobe %p\n", jp);
1182 show_regs(saved_regs);
1183 printk(KERN_ERR "Current registers\n");
1187 /* It's OK to start function graph tracing again */
1188 unpause_graph_tracing();
1189 *regs = kcb->jprobe_saved_regs;
1190 __memcpy(saved_sp, kcb->jprobes_stack, MIN_STACK_SIZE(saved_sp));
1191 preempt_enable_no_resched();
1196 NOKPROBE_SYMBOL(longjmp_break_handler);
1198 bool arch_within_kprobe_blacklist(unsigned long addr)
1200 bool is_in_entry_trampoline_section = false;
1202 #ifdef CONFIG_X86_64
1203 is_in_entry_trampoline_section =
1204 (addr >= (unsigned long)__entry_trampoline_start &&
1205 addr < (unsigned long)__entry_trampoline_end);
1207 return (addr >= (unsigned long)__kprobes_text_start &&
1208 addr < (unsigned long)__kprobes_text_end) ||
1209 (addr >= (unsigned long)__entry_text_start &&
1210 addr < (unsigned long)__entry_text_end) ||
1211 is_in_entry_trampoline_section;
1214 int __init arch_init_kprobes(void)
1219 int arch_trampoline_kprobe(struct kprobe *p)