1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Kernel Probes (KProbes)
5 * Copyright (C) IBM Corporation, 2002, 2004
7 * 2002-Oct Created by Vamsi Krishna S <vamsi_krishna@in.ibm.com> Kernel
8 * Probes initial implementation ( includes contributions from
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.
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/extable.h>
37 #include <linux/kdebug.h>
38 #include <linux/kallsyms.h>
39 #include <linux/ftrace.h>
40 #include <linux/frame.h>
41 #include <linux/kasan.h>
42 #include <linux/moduleloader.h>
44 #include <asm/text-patching.h>
45 #include <asm/cacheflush.h>
47 #include <asm/pgtable.h>
48 #include <linux/uaccess.h>
49 #include <asm/alternative.h>
51 #include <asm/debugreg.h>
52 #include <asm/set_memory.h>
56 DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
57 DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
59 #define stack_addr(regs) ((unsigned long *)regs->sp)
61 #define W(row, b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, ba, bb, bc, bd, be, bf)\
62 (((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) | \
63 (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) | \
64 (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) | \
65 (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf)) \
68 * Undefined/reserved opcodes, conditional jump, Opcode Extension
69 * Groups, and some special opcodes can not boost.
70 * This is non-const and volatile to keep gcc from statically
71 * optimizing it out, as variable_test_bit makes gcc think only
72 * *(unsigned long*) is used.
74 static volatile u32 twobyte_is_boostable[256 / 32] = {
75 /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */
76 /* ---------------------------------------------- */
77 W(0x00, 0, 0, 1, 1, 0, 0, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0) | /* 00 */
78 W(0x10, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1) , /* 10 */
79 W(0x20, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 20 */
80 W(0x30, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 30 */
81 W(0x40, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 40 */
82 W(0x50, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 50 */
83 W(0x60, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1) | /* 60 */
84 W(0x70, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1) , /* 70 */
85 W(0x80, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 80 */
86 W(0x90, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 90 */
87 W(0xa0, 1, 1, 0, 1, 1, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* a0 */
88 W(0xb0, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1) , /* b0 */
89 W(0xc0, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1) | /* c0 */
90 W(0xd0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) , /* d0 */
91 W(0xe0, 0, 1, 1, 0, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* e0 */
92 W(0xf0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 1, 0, 1, 1, 1, 0) /* f0 */
93 /* ----------------------------------------------- */
94 /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */
98 struct kretprobe_blackpoint kretprobe_blacklist[] = {
99 {"__switch_to", }, /* This function switches only current task, but
100 doesn't switch kernel stack.*/
101 {NULL, NULL} /* Terminator */
104 const int kretprobe_blacklist_size = ARRAY_SIZE(kretprobe_blacklist);
106 static nokprobe_inline void
107 __synthesize_relative_insn(void *dest, void *from, void *to, u8 op)
109 struct __arch_relative_insn {
114 insn = (struct __arch_relative_insn *)dest;
115 insn->raddr = (s32)((long)(to) - ((long)(from) + 5));
119 /* Insert a jump instruction at address 'from', which jumps to address 'to'.*/
120 void synthesize_reljump(void *dest, void *from, void *to)
122 __synthesize_relative_insn(dest, from, to, RELATIVEJUMP_OPCODE);
124 NOKPROBE_SYMBOL(synthesize_reljump);
126 /* Insert a call instruction at address 'from', which calls address 'to'.*/
127 void synthesize_relcall(void *dest, void *from, void *to)
129 __synthesize_relative_insn(dest, from, to, RELATIVECALL_OPCODE);
131 NOKPROBE_SYMBOL(synthesize_relcall);
134 * Skip the prefixes of the instruction.
136 static kprobe_opcode_t *skip_prefixes(kprobe_opcode_t *insn)
140 attr = inat_get_opcode_attribute((insn_byte_t)*insn);
141 while (inat_is_legacy_prefix(attr)) {
143 attr = inat_get_opcode_attribute((insn_byte_t)*insn);
146 if (inat_is_rex_prefix(attr))
151 NOKPROBE_SYMBOL(skip_prefixes);
154 * Returns non-zero if INSN is boostable.
155 * RIP relative instructions are adjusted at copying time in 64 bits mode
157 int can_boost(struct insn *insn, void *addr)
159 kprobe_opcode_t opcode;
161 if (search_exception_tables((unsigned long)addr))
162 return 0; /* Page fault may occur on this address. */
164 /* 2nd-byte opcode */
165 if (insn->opcode.nbytes == 2)
166 return test_bit(insn->opcode.bytes[1],
167 (unsigned long *)twobyte_is_boostable);
169 if (insn->opcode.nbytes != 1)
172 /* Can't boost Address-size override prefix */
173 if (unlikely(inat_is_address_size_prefix(insn->attr)))
176 opcode = insn->opcode.bytes[0];
178 switch (opcode & 0xf0) {
180 /* can't boost "bound" */
181 return (opcode != 0x62);
183 return 0; /* can't boost conditional jump */
185 return opcode != 0x9a; /* can't boost call far */
187 /* can't boost software-interruptions */
188 return (0xc1 < opcode && opcode < 0xcc) || opcode == 0xcf;
190 /* can boost AA* and XLAT */
191 return (opcode == 0xd4 || opcode == 0xd5 || opcode == 0xd7);
193 /* can boost in/out and absolute jmps */
194 return ((opcode & 0x04) || opcode == 0xea);
196 /* clear and set flags are boostable */
197 return (opcode == 0xf5 || (0xf7 < opcode && opcode < 0xfe));
199 /* CS override prefix and call are not boostable */
200 return (opcode != 0x2e && opcode != 0x9a);
205 __recover_probed_insn(kprobe_opcode_t *buf, unsigned long addr)
210 kp = get_kprobe((void *)addr);
211 faddr = ftrace_location(addr);
213 * Addresses inside the ftrace location are refused by
214 * arch_check_ftrace_location(). Something went terribly wrong
215 * if such an address is checked here.
217 if (WARN_ON(faddr && faddr != addr))
220 * Use the current code if it is not modified by Kprobe
221 * and it cannot be modified by ftrace.
227 * Basically, kp->ainsn.insn has an original instruction.
228 * However, RIP-relative instruction can not do single-stepping
229 * at different place, __copy_instruction() tweaks the displacement of
230 * that instruction. In that case, we can't recover the instruction
231 * from the kp->ainsn.insn.
233 * On the other hand, in case on normal Kprobe, kp->opcode has a copy
234 * of the first byte of the probed instruction, which is overwritten
235 * by int3. And the instruction at kp->addr is not modified by kprobes
236 * except for the first byte, we can recover the original instruction
237 * from it and kp->opcode.
239 * In case of Kprobes using ftrace, we do not have a copy of
240 * the original instruction. In fact, the ftrace location might
241 * be modified at anytime and even could be in an inconsistent state.
242 * Fortunately, we know that the original code is the ideal 5-byte
245 if (probe_kernel_read(buf, (void *)addr,
246 MAX_INSN_SIZE * sizeof(kprobe_opcode_t)))
250 memcpy(buf, ideal_nops[NOP_ATOMIC5], 5);
253 return (unsigned long)buf;
257 * Recover the probed instruction at addr for further analysis.
258 * Caller must lock kprobes by kprobe_mutex, or disable preemption
259 * for preventing to release referencing kprobes.
260 * Returns zero if the instruction can not get recovered (or access failed).
262 unsigned long recover_probed_instruction(kprobe_opcode_t *buf, unsigned long addr)
264 unsigned long __addr;
266 __addr = __recover_optprobed_insn(buf, addr);
270 return __recover_probed_insn(buf, addr);
273 /* Check if paddr is at an instruction boundary */
274 static int can_probe(unsigned long paddr)
276 unsigned long addr, __addr, offset = 0;
278 kprobe_opcode_t buf[MAX_INSN_SIZE];
280 if (!kallsyms_lookup_size_offset(paddr, NULL, &offset))
283 /* Decode instructions */
284 addr = paddr - offset;
285 while (addr < paddr) {
287 * Check if the instruction has been modified by another
288 * kprobe, in which case we replace the breakpoint by the
289 * original instruction in our buffer.
290 * Also, jump optimization will change the breakpoint to
291 * relative-jump. Since the relative-jump itself is
292 * normally used, we just go through if there is no kprobe.
294 __addr = recover_probed_instruction(buf, addr);
297 kernel_insn_init(&insn, (void *)__addr, MAX_INSN_SIZE);
298 insn_get_length(&insn);
301 * Another debugging subsystem might insert this breakpoint.
302 * In that case, we can't recover it.
304 if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION)
309 return (addr == paddr);
313 * Returns non-zero if opcode modifies the interrupt flag.
315 static int is_IF_modifier(kprobe_opcode_t *insn)
318 insn = skip_prefixes(insn);
323 case 0xcf: /* iret/iretd */
324 case 0x9d: /* popf/popfd */
332 * Copy an instruction with recovering modified instruction by kprobes
333 * and adjust the displacement if the instruction uses the %rip-relative
334 * addressing mode. Note that since @real will be the final place of copied
335 * instruction, displacement must be adjust by @real, not @dest.
336 * This returns the length of copied instruction, or 0 if it has an error.
338 int __copy_instruction(u8 *dest, u8 *src, u8 *real, struct insn *insn)
340 kprobe_opcode_t buf[MAX_INSN_SIZE];
341 unsigned long recovered_insn =
342 recover_probed_instruction(buf, (unsigned long)src);
344 if (!recovered_insn || !insn)
347 /* This can access kernel text if given address is not recovered */
348 if (probe_kernel_read(dest, (void *)recovered_insn, MAX_INSN_SIZE))
351 kernel_insn_init(insn, dest, MAX_INSN_SIZE);
352 insn_get_length(insn);
354 /* Another subsystem puts a breakpoint, failed to recover */
355 if (insn->opcode.bytes[0] == BREAKPOINT_INSTRUCTION)
358 /* We should not singlestep on the exception masking instructions */
359 if (insn_masking_exception(insn))
363 /* Only x86_64 has RIP relative instructions */
364 if (insn_rip_relative(insn)) {
368 * The copied instruction uses the %rip-relative addressing
369 * mode. Adjust the displacement for the difference between
370 * the original location of this instruction and the location
371 * of the copy that will actually be run. The tricky bit here
372 * is making sure that the sign extension happens correctly in
373 * this calculation, since we need a signed 32-bit result to
374 * be sign-extended to 64 bits when it's added to the %rip
375 * value and yield the same 64-bit result that the sign-
376 * extension of the original signed 32-bit displacement would
379 newdisp = (u8 *) src + (s64) insn->displacement.value
381 if ((s64) (s32) newdisp != newdisp) {
382 pr_err("Kprobes error: new displacement does not fit into s32 (%llx)\n", newdisp);
385 disp = (u8 *) dest + insn_offset_displacement(insn);
386 *(s32 *) disp = (s32) newdisp;
392 /* Prepare reljump right after instruction to boost */
393 static int prepare_boost(kprobe_opcode_t *buf, struct kprobe *p,
396 int len = insn->length;
398 if (can_boost(insn, p->addr) &&
399 MAX_INSN_SIZE - len >= RELATIVEJUMP_SIZE) {
401 * These instructions can be executed directly if it
402 * jumps back to correct address.
404 synthesize_reljump(buf + len, p->ainsn.insn + len,
405 p->addr + insn->length);
406 len += RELATIVEJUMP_SIZE;
407 p->ainsn.boostable = true;
409 p->ainsn.boostable = false;
415 /* Make page to RO mode when allocate it */
416 void *alloc_insn_page(void)
420 page = module_alloc(PAGE_SIZE);
424 set_vm_flush_reset_perms(page);
426 * First make the page read-only, and only then make it executable to
427 * prevent it from being W+X in between.
429 set_memory_ro((unsigned long)page, 1);
432 * TODO: Once additional kernel code protection mechanisms are set, ensure
433 * that the page was not maliciously altered and it is still zeroed.
435 set_memory_x((unsigned long)page, 1);
440 /* Recover page to RW mode before releasing it */
441 void free_insn_page(void *page)
443 module_memfree(page);
446 static int arch_copy_kprobe(struct kprobe *p)
449 kprobe_opcode_t buf[MAX_INSN_SIZE];
452 /* Copy an instruction with recovering if other optprobe modifies it.*/
453 len = __copy_instruction(buf, p->addr, p->ainsn.insn, &insn);
458 * __copy_instruction can modify the displacement of the instruction,
459 * but it doesn't affect boostable check.
461 len = prepare_boost(buf, p, &insn);
463 /* Check whether the instruction modifies Interrupt Flag or not */
464 p->ainsn.if_modifier = is_IF_modifier(buf);
466 /* Also, displacement change doesn't affect the first byte */
469 /* OK, write back the instruction(s) into ROX insn buffer */
470 text_poke(p->ainsn.insn, buf, len);
475 int arch_prepare_kprobe(struct kprobe *p)
479 if (alternatives_text_reserved(p->addr, p->addr))
482 if (!can_probe((unsigned long)p->addr))
484 /* insn: must be on special executable page on x86. */
485 p->ainsn.insn = get_insn_slot();
489 ret = arch_copy_kprobe(p);
491 free_insn_slot(p->ainsn.insn, 0);
492 p->ainsn.insn = NULL;
498 void arch_arm_kprobe(struct kprobe *p)
500 text_poke(p->addr, ((unsigned char []){BREAKPOINT_INSTRUCTION}), 1);
503 void arch_disarm_kprobe(struct kprobe *p)
505 text_poke(p->addr, &p->opcode, 1);
508 void arch_remove_kprobe(struct kprobe *p)
511 free_insn_slot(p->ainsn.insn, p->ainsn.boostable);
512 p->ainsn.insn = NULL;
516 static nokprobe_inline void
517 save_previous_kprobe(struct kprobe_ctlblk *kcb)
519 kcb->prev_kprobe.kp = kprobe_running();
520 kcb->prev_kprobe.status = kcb->kprobe_status;
521 kcb->prev_kprobe.old_flags = kcb->kprobe_old_flags;
522 kcb->prev_kprobe.saved_flags = kcb->kprobe_saved_flags;
525 static nokprobe_inline void
526 restore_previous_kprobe(struct kprobe_ctlblk *kcb)
528 __this_cpu_write(current_kprobe, kcb->prev_kprobe.kp);
529 kcb->kprobe_status = kcb->prev_kprobe.status;
530 kcb->kprobe_old_flags = kcb->prev_kprobe.old_flags;
531 kcb->kprobe_saved_flags = kcb->prev_kprobe.saved_flags;
534 static nokprobe_inline void
535 set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
536 struct kprobe_ctlblk *kcb)
538 __this_cpu_write(current_kprobe, p);
539 kcb->kprobe_saved_flags = kcb->kprobe_old_flags
540 = (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF));
541 if (p->ainsn.if_modifier)
542 kcb->kprobe_saved_flags &= ~X86_EFLAGS_IF;
545 static nokprobe_inline void clear_btf(void)
547 if (test_thread_flag(TIF_BLOCKSTEP)) {
548 unsigned long debugctl = get_debugctlmsr();
550 debugctl &= ~DEBUGCTLMSR_BTF;
551 update_debugctlmsr(debugctl);
555 static nokprobe_inline void restore_btf(void)
557 if (test_thread_flag(TIF_BLOCKSTEP)) {
558 unsigned long debugctl = get_debugctlmsr();
560 debugctl |= DEBUGCTLMSR_BTF;
561 update_debugctlmsr(debugctl);
565 void arch_prepare_kretprobe(struct kretprobe_instance *ri, struct pt_regs *regs)
567 unsigned long *sara = stack_addr(regs);
569 ri->ret_addr = (kprobe_opcode_t *) *sara;
572 /* Replace the return addr with trampoline addr */
573 *sara = (unsigned long) &kretprobe_trampoline;
575 NOKPROBE_SYMBOL(arch_prepare_kretprobe);
577 static void setup_singlestep(struct kprobe *p, struct pt_regs *regs,
578 struct kprobe_ctlblk *kcb, int reenter)
580 if (setup_detour_execution(p, regs, reenter))
583 #if !defined(CONFIG_PREEMPT)
584 if (p->ainsn.boostable && !p->post_handler) {
585 /* Boost up -- we can execute copied instructions directly */
587 reset_current_kprobe();
589 * Reentering boosted probe doesn't reset current_kprobe,
590 * nor set current_kprobe, because it doesn't use single
593 regs->ip = (unsigned long)p->ainsn.insn;
598 save_previous_kprobe(kcb);
599 set_current_kprobe(p, regs, kcb);
600 kcb->kprobe_status = KPROBE_REENTER;
602 kcb->kprobe_status = KPROBE_HIT_SS;
603 /* Prepare real single stepping */
605 regs->flags |= X86_EFLAGS_TF;
606 regs->flags &= ~X86_EFLAGS_IF;
607 /* single step inline if the instruction is an int3 */
608 if (p->opcode == BREAKPOINT_INSTRUCTION)
609 regs->ip = (unsigned long)p->addr;
611 regs->ip = (unsigned long)p->ainsn.insn;
613 NOKPROBE_SYMBOL(setup_singlestep);
616 * We have reentered the kprobe_handler(), since another probe was hit while
617 * within the handler. We save the original kprobes variables and just single
618 * step on the instruction of the new probe without calling any user handlers.
620 static int reenter_kprobe(struct kprobe *p, struct pt_regs *regs,
621 struct kprobe_ctlblk *kcb)
623 switch (kcb->kprobe_status) {
624 case KPROBE_HIT_SSDONE:
625 case KPROBE_HIT_ACTIVE:
627 kprobes_inc_nmissed_count(p);
628 setup_singlestep(p, regs, kcb, 1);
631 /* A probe has been hit in the codepath leading up to, or just
632 * after, single-stepping of a probed instruction. This entire
633 * codepath should strictly reside in .kprobes.text section.
634 * Raise a BUG or we'll continue in an endless reentering loop
635 * and eventually a stack overflow.
637 pr_err("Unrecoverable kprobe detected.\n");
641 /* impossible cases */
648 NOKPROBE_SYMBOL(reenter_kprobe);
651 * Interrupts are disabled on entry as trap3 is an interrupt gate and they
652 * remain disabled throughout this function.
654 int kprobe_int3_handler(struct pt_regs *regs)
656 kprobe_opcode_t *addr;
658 struct kprobe_ctlblk *kcb;
663 addr = (kprobe_opcode_t *)(regs->ip - sizeof(kprobe_opcode_t));
665 * We don't want to be preempted for the entire duration of kprobe
666 * processing. Since int3 and debug trap disables irqs and we clear
667 * IF while singlestepping, it must be no preemptible.
670 kcb = get_kprobe_ctlblk();
671 p = get_kprobe(addr);
674 if (kprobe_running()) {
675 if (reenter_kprobe(p, regs, kcb))
678 set_current_kprobe(p, regs, kcb);
679 kcb->kprobe_status = KPROBE_HIT_ACTIVE;
682 * If we have no pre-handler or it returned 0, we
683 * continue with normal processing. If we have a
684 * pre-handler and it returned non-zero, that means
685 * user handler setup registers to exit to another
686 * instruction, we must skip the single stepping.
688 if (!p->pre_handler || !p->pre_handler(p, regs))
689 setup_singlestep(p, regs, kcb, 0);
691 reset_current_kprobe();
694 } else if (*addr != BREAKPOINT_INSTRUCTION) {
696 * The breakpoint instruction was removed right
697 * after we hit it. Another cpu has removed
698 * either a probepoint or a debugger breakpoint
699 * at this address. In either case, no further
700 * handling of this interrupt is appropriate.
701 * Back up over the (now missing) int3 and run
702 * the original instruction.
704 regs->ip = (unsigned long)addr;
706 } /* else: not a kprobe fault; let the kernel handle it */
710 NOKPROBE_SYMBOL(kprobe_int3_handler);
713 * When a retprobed function returns, this code saves registers and
714 * calls trampoline_handler() runs, which calls the kretprobe's handler.
718 ".global kretprobe_trampoline\n"
719 ".type kretprobe_trampoline, @function\n"
720 "kretprobe_trampoline:\n"
721 /* We don't bother saving the ss register */
727 " call trampoline_handler\n"
728 /* Replace saved sp with true return address. */
729 " movq %rax, 19*8(%rsp)\n"
737 " call trampoline_handler\n"
738 /* Replace saved sp with true return address. */
739 " movl %eax, 15*4(%esp)\n"
744 ".size kretprobe_trampoline, .-kretprobe_trampoline\n"
746 NOKPROBE_SYMBOL(kretprobe_trampoline);
747 STACK_FRAME_NON_STANDARD(kretprobe_trampoline);
749 static struct kprobe kretprobe_kprobe = {
750 .addr = (void *)kretprobe_trampoline,
754 * Called from kretprobe_trampoline
756 __used __visible void *trampoline_handler(struct pt_regs *regs)
758 struct kprobe_ctlblk *kcb;
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;
771 * Set a dummy kprobe for avoiding kretprobe recursion.
772 * Since kretprobe never run in kprobe handler, kprobe must not
773 * be running at this point.
775 kcb = get_kprobe_ctlblk();
776 __this_cpu_write(current_kprobe, &kretprobe_kprobe);
777 kcb->kprobe_status = KPROBE_HIT_ACTIVE;
779 INIT_HLIST_HEAD(&empty_rp);
780 kretprobe_hash_lock(current, &head, &flags);
781 /* fixup registers */
782 regs->cs = __KERNEL_CS;
784 regs->cs |= get_kernel_rpl();
787 /* We use pt_regs->sp for return address holder. */
788 frame_pointer = ®s->sp;
789 regs->ip = trampoline_address;
790 regs->orig_ax = ~0UL;
793 * It is possible to have multiple instances associated with a given
794 * task either because multiple functions in the call path have
795 * return probes installed on them, and/or more than one
796 * return probe was registered for a target function.
798 * We can handle this because:
799 * - instances are always pushed into the head of the list
800 * - when multiple return probes are registered for the same
801 * function, the (chronologically) first instance's ret_addr
802 * will be the real return address, and all the rest will
803 * point to kretprobe_trampoline.
805 hlist_for_each_entry(ri, head, hlist) {
806 if (ri->task != current)
807 /* another task is sharing our hash bucket */
810 * Return probes must be pushed on this hash list correct
811 * order (same as return order) so that it can be popped
812 * correctly. However, if we find it is pushed it incorrect
813 * order, this means we find a function which should not be
814 * probed, because the wrong order entry is pushed on the
815 * path of processing other kretprobe itself.
817 if (ri->fp != frame_pointer) {
819 pr_warn("kretprobe is stacked incorrectly. Trying to fixup.\n");
824 orig_ret_address = (unsigned long)ri->ret_addr;
826 pr_warn("%ps must be blacklisted because of incorrect kretprobe order\n",
829 if (orig_ret_address != trampoline_address)
831 * This is the real return address. Any other
832 * instances associated with this task are for
833 * other calls deeper on the call stack
838 kretprobe_assert(ri, orig_ret_address, trampoline_address);
840 correct_ret_addr = ri->ret_addr;
841 hlist_for_each_entry_safe(ri, tmp, head, hlist) {
842 if (ri->task != current)
843 /* another task is sharing our hash bucket */
845 if (ri->fp != frame_pointer)
848 orig_ret_address = (unsigned long)ri->ret_addr;
849 if (ri->rp && ri->rp->handler) {
850 __this_cpu_write(current_kprobe, &ri->rp->kp);
851 ri->ret_addr = correct_ret_addr;
852 ri->rp->handler(ri, regs);
853 __this_cpu_write(current_kprobe, &kretprobe_kprobe);
856 recycle_rp_inst(ri, &empty_rp);
858 if (orig_ret_address != trampoline_address)
860 * This is the real return address. Any other
861 * instances associated with this task are for
862 * other calls deeper on the call stack
867 kretprobe_hash_unlock(current, &flags);
869 __this_cpu_write(current_kprobe, NULL);
872 hlist_for_each_entry_safe(ri, tmp, &empty_rp, hlist) {
873 hlist_del(&ri->hlist);
876 return (void *)orig_ret_address;
878 NOKPROBE_SYMBOL(trampoline_handler);
881 * Called after single-stepping. p->addr is the address of the
882 * instruction whose first byte has been replaced by the "int 3"
883 * instruction. To avoid the SMP problems that can occur when we
884 * temporarily put back the original opcode to single-step, we
885 * single-stepped a copy of the instruction. The address of this
886 * copy is p->ainsn.insn.
888 * This function prepares to return from the post-single-step
889 * interrupt. We have to fix up the stack as follows:
891 * 0) Except in the case of absolute or indirect jump or call instructions,
892 * the new ip is relative to the copied instruction. We need to make
893 * it relative to the original instruction.
895 * 1) If the single-stepped instruction was pushfl, then the TF and IF
896 * flags are set in the just-pushed flags, and may need to be cleared.
898 * 2) If the single-stepped instruction was a call, the return address
899 * that is atop the stack is the address following the copied instruction.
900 * We need to make it the address following the original instruction.
902 * If this is the first time we've single-stepped the instruction at
903 * this probepoint, and the instruction is boostable, boost it: add a
904 * jump instruction after the copied instruction, that jumps to the next
905 * instruction after the probepoint.
907 static void resume_execution(struct kprobe *p, struct pt_regs *regs,
908 struct kprobe_ctlblk *kcb)
910 unsigned long *tos = stack_addr(regs);
911 unsigned long copy_ip = (unsigned long)p->ainsn.insn;
912 unsigned long orig_ip = (unsigned long)p->addr;
913 kprobe_opcode_t *insn = p->ainsn.insn;
916 insn = skip_prefixes(insn);
918 regs->flags &= ~X86_EFLAGS_TF;
920 case 0x9c: /* pushfl */
921 *tos &= ~(X86_EFLAGS_TF | X86_EFLAGS_IF);
922 *tos |= kcb->kprobe_old_flags;
924 case 0xc2: /* iret/ret/lret */
929 case 0xea: /* jmp absolute -- ip is correct */
930 /* ip is already adjusted, no more changes required */
931 p->ainsn.boostable = true;
933 case 0xe8: /* call relative - Fix return addr */
934 *tos = orig_ip + (*tos - copy_ip);
937 case 0x9a: /* call absolute -- same as call absolute, indirect */
938 *tos = orig_ip + (*tos - copy_ip);
942 if ((insn[1] & 0x30) == 0x10) {
944 * call absolute, indirect
945 * Fix return addr; ip is correct.
946 * But this is not boostable
948 *tos = orig_ip + (*tos - copy_ip);
950 } else if (((insn[1] & 0x31) == 0x20) ||
951 ((insn[1] & 0x31) == 0x21)) {
953 * jmp near and far, absolute indirect
954 * ip is correct. And this is boostable
956 p->ainsn.boostable = true;
963 regs->ip += orig_ip - copy_ip;
968 NOKPROBE_SYMBOL(resume_execution);
971 * Interrupts are disabled on entry as trap1 is an interrupt gate and they
972 * remain disabled throughout this function.
974 int kprobe_debug_handler(struct pt_regs *regs)
976 struct kprobe *cur = kprobe_running();
977 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
982 resume_execution(cur, regs, kcb);
983 regs->flags |= kcb->kprobe_saved_flags;
985 if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
986 kcb->kprobe_status = KPROBE_HIT_SSDONE;
987 cur->post_handler(cur, regs, 0);
990 /* Restore back the original saved kprobes variables and continue. */
991 if (kcb->kprobe_status == KPROBE_REENTER) {
992 restore_previous_kprobe(kcb);
995 reset_current_kprobe();
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;
1034 * If the TF flag was set before the kprobe hit,
1037 regs->flags |= kcb->kprobe_old_flags;
1039 if (kcb->kprobe_status == KPROBE_REENTER)
1040 restore_previous_kprobe(kcb);
1042 reset_current_kprobe();
1043 } else if (kcb->kprobe_status == KPROBE_HIT_ACTIVE ||
1044 kcb->kprobe_status == KPROBE_HIT_SSDONE) {
1046 * We increment the nmissed count for accounting,
1047 * we can also use npre/npostfault count for accounting
1048 * these specific fault cases.
1050 kprobes_inc_nmissed_count(cur);
1053 * We come here because instructions in the pre/post
1054 * handler caused the page_fault, this could happen
1055 * if handler tries to access user space by
1056 * copy_from_user(), get_user() etc. Let the
1057 * user-specified handler try to fix it first.
1059 if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
1065 NOKPROBE_SYMBOL(kprobe_fault_handler);
1067 int __init arch_populate_kprobe_blacklist(void)
1071 ret = kprobe_add_area_blacklist((unsigned long)__irqentry_text_start,
1072 (unsigned long)__irqentry_text_end);
1076 return kprobe_add_area_blacklist((unsigned long)__entry_text_start,
1077 (unsigned long)__entry_text_end);
1080 int __init arch_init_kprobes(void)
1085 int arch_trampoline_kprobe(struct kprobe *p)