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;
163 if (search_exception_tables((unsigned long)addr))
164 return 0; /* Page fault may occur on this address. */
166 /* 2nd-byte opcode */
167 if (insn->opcode.nbytes == 2)
168 return test_bit(insn->opcode.bytes[1],
169 (unsigned long *)twobyte_is_boostable);
171 if (insn->opcode.nbytes != 1)
174 for_each_insn_prefix(insn, i, prefix) {
177 attr = inat_get_opcode_attribute(prefix);
178 /* Can't boost Address-size override prefix and CS override prefix */
179 if (prefix == 0x2e || inat_is_address_size_prefix(attr))
183 opcode = insn->opcode.bytes[0];
185 switch (opcode & 0xf0) {
187 /* can't boost "bound" */
188 return (opcode != 0x62);
190 return 0; /* can't boost conditional jump */
192 return opcode != 0x9a; /* can't boost call far */
194 /* can't boost software-interruptions */
195 return (0xc1 < opcode && opcode < 0xcc) || opcode == 0xcf;
197 /* can boost AA* and XLAT */
198 return (opcode == 0xd4 || opcode == 0xd5 || opcode == 0xd7);
200 /* can boost in/out and absolute jmps */
201 return ((opcode & 0x04) || opcode == 0xea);
203 /* clear and set flags are boostable */
204 return (opcode == 0xf5 || (0xf7 < opcode && opcode < 0xfe));
206 /* call is not boostable */
207 return opcode != 0x9a;
212 __recover_probed_insn(kprobe_opcode_t *buf, unsigned long addr)
217 kp = get_kprobe((void *)addr);
218 faddr = ftrace_location(addr);
220 * Addresses inside the ftrace location are refused by
221 * arch_check_ftrace_location(). Something went terribly wrong
222 * if such an address is checked here.
224 if (WARN_ON(faddr && faddr != addr))
227 * Use the current code if it is not modified by Kprobe
228 * and it cannot be modified by ftrace.
234 * Basically, kp->ainsn.insn has an original instruction.
235 * However, RIP-relative instruction can not do single-stepping
236 * at different place, __copy_instruction() tweaks the displacement of
237 * that instruction. In that case, we can't recover the instruction
238 * from the kp->ainsn.insn.
240 * On the other hand, in case on normal Kprobe, kp->opcode has a copy
241 * of the first byte of the probed instruction, which is overwritten
242 * by int3. And the instruction at kp->addr is not modified by kprobes
243 * except for the first byte, we can recover the original instruction
244 * from it and kp->opcode.
246 * In case of Kprobes using ftrace, we do not have a copy of
247 * the original instruction. In fact, the ftrace location might
248 * be modified at anytime and even could be in an inconsistent state.
249 * Fortunately, we know that the original code is the ideal 5-byte
252 if (probe_kernel_read(buf, (void *)addr,
253 MAX_INSN_SIZE * sizeof(kprobe_opcode_t)))
257 memcpy(buf, ideal_nops[NOP_ATOMIC5], 5);
260 return (unsigned long)buf;
264 * Recover the probed instruction at addr for further analysis.
265 * Caller must lock kprobes by kprobe_mutex, or disable preemption
266 * for preventing to release referencing kprobes.
267 * Returns zero if the instruction can not get recovered (or access failed).
269 unsigned long recover_probed_instruction(kprobe_opcode_t *buf, unsigned long addr)
271 unsigned long __addr;
273 __addr = __recover_optprobed_insn(buf, addr);
277 return __recover_probed_insn(buf, addr);
280 /* Check if paddr is at an instruction boundary */
281 static int can_probe(unsigned long paddr)
283 unsigned long addr, __addr, offset = 0;
285 kprobe_opcode_t buf[MAX_INSN_SIZE];
287 if (!kallsyms_lookup_size_offset(paddr, NULL, &offset))
290 /* Decode instructions */
291 addr = paddr - offset;
292 while (addr < paddr) {
294 * Check if the instruction has been modified by another
295 * kprobe, in which case we replace the breakpoint by the
296 * original instruction in our buffer.
297 * Also, jump optimization will change the breakpoint to
298 * relative-jump. Since the relative-jump itself is
299 * normally used, we just go through if there is no kprobe.
301 __addr = recover_probed_instruction(buf, addr);
304 kernel_insn_init(&insn, (void *)__addr, MAX_INSN_SIZE);
305 insn_get_length(&insn);
308 * Another debugging subsystem might insert this breakpoint.
309 * In that case, we can't recover it.
311 if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION)
316 return (addr == paddr);
320 * Returns non-zero if opcode modifies the interrupt flag.
322 static int is_IF_modifier(kprobe_opcode_t *insn)
325 insn = skip_prefixes(insn);
330 case 0xcf: /* iret/iretd */
331 case 0x9d: /* popf/popfd */
339 * Copy an instruction with recovering modified instruction by kprobes
340 * and adjust the displacement if the instruction uses the %rip-relative
341 * addressing mode. Note that since @real will be the final place of copied
342 * instruction, displacement must be adjust by @real, not @dest.
343 * This returns the length of copied instruction, or 0 if it has an error.
345 int __copy_instruction(u8 *dest, u8 *src, u8 *real, struct insn *insn)
347 kprobe_opcode_t buf[MAX_INSN_SIZE];
348 unsigned long recovered_insn =
349 recover_probed_instruction(buf, (unsigned long)src);
351 if (!recovered_insn || !insn)
354 /* This can access kernel text if given address is not recovered */
355 if (probe_kernel_read(dest, (void *)recovered_insn, MAX_INSN_SIZE))
358 kernel_insn_init(insn, dest, MAX_INSN_SIZE);
359 insn_get_length(insn);
361 /* We can not probe force emulate prefixed instruction */
362 if (insn_has_emulate_prefix(insn))
365 /* Another subsystem puts a breakpoint, failed to recover */
366 if (insn->opcode.bytes[0] == BREAKPOINT_INSTRUCTION)
369 /* We should not singlestep on the exception masking instructions */
370 if (insn_masking_exception(insn))
374 /* Only x86_64 has RIP relative instructions */
375 if (insn_rip_relative(insn)) {
379 * The copied instruction uses the %rip-relative addressing
380 * mode. Adjust the displacement for the difference between
381 * the original location of this instruction and the location
382 * of the copy that will actually be run. The tricky bit here
383 * is making sure that the sign extension happens correctly in
384 * this calculation, since we need a signed 32-bit result to
385 * be sign-extended to 64 bits when it's added to the %rip
386 * value and yield the same 64-bit result that the sign-
387 * extension of the original signed 32-bit displacement would
390 newdisp = (u8 *) src + (s64) insn->displacement.value
392 if ((s64) (s32) newdisp != newdisp) {
393 pr_err("Kprobes error: new displacement does not fit into s32 (%llx)\n", newdisp);
396 disp = (u8 *) dest + insn_offset_displacement(insn);
397 *(s32 *) disp = (s32) newdisp;
403 /* Prepare reljump right after instruction to boost */
404 static int prepare_boost(kprobe_opcode_t *buf, struct kprobe *p,
407 int len = insn->length;
409 if (can_boost(insn, p->addr) &&
410 MAX_INSN_SIZE - len >= RELATIVEJUMP_SIZE) {
412 * These instructions can be executed directly if it
413 * jumps back to correct address.
415 synthesize_reljump(buf + len, p->ainsn.insn + len,
416 p->addr + insn->length);
417 len += RELATIVEJUMP_SIZE;
418 p->ainsn.boostable = true;
420 p->ainsn.boostable = false;
426 /* Make page to RO mode when allocate it */
427 void *alloc_insn_page(void)
431 page = module_alloc(PAGE_SIZE);
435 set_vm_flush_reset_perms(page);
437 * First make the page read-only, and only then make it executable to
438 * prevent it from being W+X in between.
440 set_memory_ro((unsigned long)page, 1);
443 * TODO: Once additional kernel code protection mechanisms are set, ensure
444 * that the page was not maliciously altered and it is still zeroed.
446 set_memory_x((unsigned long)page, 1);
451 /* Recover page to RW mode before releasing it */
452 void free_insn_page(void *page)
454 module_memfree(page);
457 static int arch_copy_kprobe(struct kprobe *p)
460 kprobe_opcode_t buf[MAX_INSN_SIZE];
463 /* Copy an instruction with recovering if other optprobe modifies it.*/
464 len = __copy_instruction(buf, p->addr, p->ainsn.insn, &insn);
469 * __copy_instruction can modify the displacement of the instruction,
470 * but it doesn't affect boostable check.
472 len = prepare_boost(buf, p, &insn);
474 /* Check whether the instruction modifies Interrupt Flag or not */
475 p->ainsn.if_modifier = is_IF_modifier(buf);
477 /* Also, displacement change doesn't affect the first byte */
480 /* OK, write back the instruction(s) into ROX insn buffer */
481 text_poke(p->ainsn.insn, buf, len);
486 int arch_prepare_kprobe(struct kprobe *p)
490 if (alternatives_text_reserved(p->addr, p->addr))
493 if (!can_probe((unsigned long)p->addr))
495 /* insn: must be on special executable page on x86. */
496 p->ainsn.insn = get_insn_slot();
500 ret = arch_copy_kprobe(p);
502 free_insn_slot(p->ainsn.insn, 0);
503 p->ainsn.insn = NULL;
509 void arch_arm_kprobe(struct kprobe *p)
511 text_poke(p->addr, ((unsigned char []){BREAKPOINT_INSTRUCTION}), 1);
514 void arch_disarm_kprobe(struct kprobe *p)
516 text_poke(p->addr, &p->opcode, 1);
519 void arch_remove_kprobe(struct kprobe *p)
522 free_insn_slot(p->ainsn.insn, p->ainsn.boostable);
523 p->ainsn.insn = NULL;
527 static nokprobe_inline void
528 save_previous_kprobe(struct kprobe_ctlblk *kcb)
530 kcb->prev_kprobe.kp = kprobe_running();
531 kcb->prev_kprobe.status = kcb->kprobe_status;
532 kcb->prev_kprobe.old_flags = kcb->kprobe_old_flags;
533 kcb->prev_kprobe.saved_flags = kcb->kprobe_saved_flags;
536 static nokprobe_inline void
537 restore_previous_kprobe(struct kprobe_ctlblk *kcb)
539 __this_cpu_write(current_kprobe, kcb->prev_kprobe.kp);
540 kcb->kprobe_status = kcb->prev_kprobe.status;
541 kcb->kprobe_old_flags = kcb->prev_kprobe.old_flags;
542 kcb->kprobe_saved_flags = kcb->prev_kprobe.saved_flags;
545 static nokprobe_inline void
546 set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
547 struct kprobe_ctlblk *kcb)
549 __this_cpu_write(current_kprobe, p);
550 kcb->kprobe_saved_flags = kcb->kprobe_old_flags
551 = (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF));
552 if (p->ainsn.if_modifier)
553 kcb->kprobe_saved_flags &= ~X86_EFLAGS_IF;
556 static nokprobe_inline void clear_btf(void)
558 if (test_thread_flag(TIF_BLOCKSTEP)) {
559 unsigned long debugctl = get_debugctlmsr();
561 debugctl &= ~DEBUGCTLMSR_BTF;
562 update_debugctlmsr(debugctl);
566 static nokprobe_inline void restore_btf(void)
568 if (test_thread_flag(TIF_BLOCKSTEP)) {
569 unsigned long debugctl = get_debugctlmsr();
571 debugctl |= DEBUGCTLMSR_BTF;
572 update_debugctlmsr(debugctl);
576 void arch_prepare_kretprobe(struct kretprobe_instance *ri, struct pt_regs *regs)
578 unsigned long *sara = stack_addr(regs);
580 ri->ret_addr = (kprobe_opcode_t *) *sara;
583 /* Replace the return addr with trampoline addr */
584 *sara = (unsigned long) &kretprobe_trampoline;
586 NOKPROBE_SYMBOL(arch_prepare_kretprobe);
588 static void setup_singlestep(struct kprobe *p, struct pt_regs *regs,
589 struct kprobe_ctlblk *kcb, int reenter)
591 if (setup_detour_execution(p, regs, reenter))
594 #if !defined(CONFIG_PREEMPTION)
595 if (p->ainsn.boostable && !p->post_handler) {
596 /* Boost up -- we can execute copied instructions directly */
598 reset_current_kprobe();
600 * Reentering boosted probe doesn't reset current_kprobe,
601 * nor set current_kprobe, because it doesn't use single
604 regs->ip = (unsigned long)p->ainsn.insn;
609 save_previous_kprobe(kcb);
610 set_current_kprobe(p, regs, kcb);
611 kcb->kprobe_status = KPROBE_REENTER;
613 kcb->kprobe_status = KPROBE_HIT_SS;
614 /* Prepare real single stepping */
616 regs->flags |= X86_EFLAGS_TF;
617 regs->flags &= ~X86_EFLAGS_IF;
618 /* single step inline if the instruction is an int3 */
619 if (p->opcode == BREAKPOINT_INSTRUCTION)
620 regs->ip = (unsigned long)p->addr;
622 regs->ip = (unsigned long)p->ainsn.insn;
624 NOKPROBE_SYMBOL(setup_singlestep);
627 * We have reentered the kprobe_handler(), since another probe was hit while
628 * within the handler. We save the original kprobes variables and just single
629 * step on the instruction of the new probe without calling any user handlers.
631 static int reenter_kprobe(struct kprobe *p, struct pt_regs *regs,
632 struct kprobe_ctlblk *kcb)
634 switch (kcb->kprobe_status) {
635 case KPROBE_HIT_SSDONE:
636 case KPROBE_HIT_ACTIVE:
638 kprobes_inc_nmissed_count(p);
639 setup_singlestep(p, regs, kcb, 1);
642 /* A probe has been hit in the codepath leading up to, or just
643 * after, single-stepping of a probed instruction. This entire
644 * codepath should strictly reside in .kprobes.text section.
645 * Raise a BUG or we'll continue in an endless reentering loop
646 * and eventually a stack overflow.
648 pr_err("Unrecoverable kprobe detected.\n");
652 /* impossible cases */
659 NOKPROBE_SYMBOL(reenter_kprobe);
662 * Interrupts are disabled on entry as trap3 is an interrupt gate and they
663 * remain disabled throughout this function.
665 int kprobe_int3_handler(struct pt_regs *regs)
667 kprobe_opcode_t *addr;
669 struct kprobe_ctlblk *kcb;
674 addr = (kprobe_opcode_t *)(regs->ip - sizeof(kprobe_opcode_t));
676 * We don't want to be preempted for the entire duration of kprobe
677 * processing. Since int3 and debug trap disables irqs and we clear
678 * IF while singlestepping, it must be no preemptible.
681 kcb = get_kprobe_ctlblk();
682 p = get_kprobe(addr);
685 if (kprobe_running()) {
686 if (reenter_kprobe(p, regs, kcb))
689 set_current_kprobe(p, regs, kcb);
690 kcb->kprobe_status = KPROBE_HIT_ACTIVE;
693 * If we have no pre-handler or it returned 0, we
694 * continue with normal processing. If we have a
695 * pre-handler and it returned non-zero, that means
696 * user handler setup registers to exit to another
697 * instruction, we must skip the single stepping.
699 if (!p->pre_handler || !p->pre_handler(p, regs))
700 setup_singlestep(p, regs, kcb, 0);
702 reset_current_kprobe();
705 } else if (*addr != BREAKPOINT_INSTRUCTION) {
707 * The breakpoint instruction was removed right
708 * after we hit it. Another cpu has removed
709 * either a probepoint or a debugger breakpoint
710 * at this address. In either case, no further
711 * handling of this interrupt is appropriate.
712 * Back up over the (now missing) int3 and run
713 * the original instruction.
715 regs->ip = (unsigned long)addr;
717 } /* else: not a kprobe fault; let the kernel handle it */
721 NOKPROBE_SYMBOL(kprobe_int3_handler);
724 * When a retprobed function returns, this code saves registers and
725 * calls trampoline_handler() runs, which calls the kretprobe's handler.
729 ".global kretprobe_trampoline\n"
730 ".type kretprobe_trampoline, @function\n"
731 "kretprobe_trampoline:\n"
732 /* We don't bother saving the ss register */
738 " call trampoline_handler\n"
739 /* Replace saved sp with true return address. */
740 " movq %rax, 19*8(%rsp)\n"
748 " call trampoline_handler\n"
749 /* Replace saved sp with true return address. */
750 " movl %eax, 15*4(%esp)\n"
755 ".size kretprobe_trampoline, .-kretprobe_trampoline\n"
757 NOKPROBE_SYMBOL(kretprobe_trampoline);
758 STACK_FRAME_NON_STANDARD(kretprobe_trampoline);
761 * Called from kretprobe_trampoline
763 __used __visible void *trampoline_handler(struct pt_regs *regs)
765 struct kretprobe_instance *ri = NULL;
766 struct hlist_head *head, empty_rp;
767 struct hlist_node *tmp;
768 unsigned long flags, orig_ret_address = 0;
769 unsigned long trampoline_address = (unsigned long)&kretprobe_trampoline;
770 kprobe_opcode_t *correct_ret_addr = NULL;
772 bool skipped = false;
775 * Set a dummy kprobe for avoiding kretprobe recursion.
776 * Since kretprobe never run in kprobe handler, kprobe must not
777 * be running at this point.
781 INIT_HLIST_HEAD(&empty_rp);
782 kretprobe_hash_lock(current, &head, &flags);
783 /* fixup registers */
784 regs->cs = __KERNEL_CS;
786 regs->cs |= get_kernel_rpl();
789 /* We use pt_regs->sp for return address holder. */
790 frame_pointer = ®s->sp;
791 regs->ip = trampoline_address;
792 regs->orig_ax = ~0UL;
795 * It is possible to have multiple instances associated with a given
796 * task either because multiple functions in the call path have
797 * return probes installed on them, and/or more than one
798 * return probe was registered for a target function.
800 * We can handle this because:
801 * - instances are always pushed into the head of the list
802 * - when multiple return probes are registered for the same
803 * function, the (chronologically) first instance's ret_addr
804 * will be the real return address, and all the rest will
805 * point to kretprobe_trampoline.
807 hlist_for_each_entry(ri, head, hlist) {
808 if (ri->task != current)
809 /* another task is sharing our hash bucket */
812 * Return probes must be pushed on this hash list correct
813 * order (same as return order) so that it can be popped
814 * correctly. However, if we find it is pushed it incorrect
815 * order, this means we find a function which should not be
816 * probed, because the wrong order entry is pushed on the
817 * path of processing other kretprobe itself.
819 if (ri->fp != frame_pointer) {
821 pr_warn("kretprobe is stacked incorrectly. Trying to fixup.\n");
826 orig_ret_address = (unsigned long)ri->ret_addr;
828 pr_warn("%ps must be blacklisted because of incorrect kretprobe order\n",
831 if (orig_ret_address != trampoline_address)
833 * This is the real return address. Any other
834 * instances associated with this task are for
835 * other calls deeper on the call stack
840 kretprobe_assert(ri, orig_ret_address, trampoline_address);
842 correct_ret_addr = ri->ret_addr;
843 hlist_for_each_entry_safe(ri, tmp, head, hlist) {
844 if (ri->task != current)
845 /* another task is sharing our hash bucket */
847 if (ri->fp != frame_pointer)
850 orig_ret_address = (unsigned long)ri->ret_addr;
851 if (ri->rp && ri->rp->handler) {
852 __this_cpu_write(current_kprobe, &ri->rp->kp);
853 ri->ret_addr = correct_ret_addr;
854 ri->rp->handler(ri, regs);
855 __this_cpu_write(current_kprobe, &kprobe_busy);
858 recycle_rp_inst(ri, &empty_rp);
860 if (orig_ret_address != trampoline_address)
862 * This is the real return address. Any other
863 * instances associated with this task are for
864 * other calls deeper on the call stack
869 kretprobe_hash_unlock(current, &flags);
873 hlist_for_each_entry_safe(ri, tmp, &empty_rp, hlist) {
874 hlist_del(&ri->hlist);
877 return (void *)orig_ret_address;
879 NOKPROBE_SYMBOL(trampoline_handler);
882 * Called after single-stepping. p->addr is the address of the
883 * instruction whose first byte has been replaced by the "int 3"
884 * instruction. To avoid the SMP problems that can occur when we
885 * temporarily put back the original opcode to single-step, we
886 * single-stepped a copy of the instruction. The address of this
887 * copy is p->ainsn.insn.
889 * This function prepares to return from the post-single-step
890 * interrupt. We have to fix up the stack as follows:
892 * 0) Except in the case of absolute or indirect jump or call instructions,
893 * the new ip is relative to the copied instruction. We need to make
894 * it relative to the original instruction.
896 * 1) If the single-stepped instruction was pushfl, then the TF and IF
897 * flags are set in the just-pushed flags, and may need to be cleared.
899 * 2) If the single-stepped instruction was a call, the return address
900 * that is atop the stack is the address following the copied instruction.
901 * We need to make it the address following the original instruction.
903 * If this is the first time we've single-stepped the instruction at
904 * this probepoint, and the instruction is boostable, boost it: add a
905 * jump instruction after the copied instruction, that jumps to the next
906 * instruction after the probepoint.
908 static void resume_execution(struct kprobe *p, struct pt_regs *regs,
909 struct kprobe_ctlblk *kcb)
911 unsigned long *tos = stack_addr(regs);
912 unsigned long copy_ip = (unsigned long)p->ainsn.insn;
913 unsigned long orig_ip = (unsigned long)p->addr;
914 kprobe_opcode_t *insn = p->ainsn.insn;
917 insn = skip_prefixes(insn);
919 regs->flags &= ~X86_EFLAGS_TF;
921 case 0x9c: /* pushfl */
922 *tos &= ~(X86_EFLAGS_TF | X86_EFLAGS_IF);
923 *tos |= kcb->kprobe_old_flags;
925 case 0xc2: /* iret/ret/lret */
930 case 0xea: /* jmp absolute -- ip is correct */
931 /* ip is already adjusted, no more changes required */
932 p->ainsn.boostable = true;
934 case 0xe8: /* call relative - Fix return addr */
935 *tos = orig_ip + (*tos - copy_ip);
938 case 0x9a: /* call absolute -- same as call absolute, indirect */
939 *tos = orig_ip + (*tos - copy_ip);
943 if ((insn[1] & 0x30) == 0x10) {
945 * call absolute, indirect
946 * Fix return addr; ip is correct.
947 * But this is not boostable
949 *tos = orig_ip + (*tos - copy_ip);
951 } else if (((insn[1] & 0x31) == 0x20) ||
952 ((insn[1] & 0x31) == 0x21)) {
954 * jmp near and far, absolute indirect
955 * ip is correct. And this is boostable
957 p->ainsn.boostable = true;
964 regs->ip += orig_ip - copy_ip;
969 NOKPROBE_SYMBOL(resume_execution);
972 * Interrupts are disabled on entry as trap1 is an interrupt gate and they
973 * remain disabled throughout this function.
975 int kprobe_debug_handler(struct pt_regs *regs)
977 struct kprobe *cur = kprobe_running();
978 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
983 resume_execution(cur, regs, kcb);
984 regs->flags |= kcb->kprobe_saved_flags;
986 if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
987 kcb->kprobe_status = KPROBE_HIT_SSDONE;
988 cur->post_handler(cur, regs, 0);
991 /* Restore back the original saved kprobes variables and continue. */
992 if (kcb->kprobe_status == KPROBE_REENTER) {
993 restore_previous_kprobe(kcb);
996 reset_current_kprobe();
999 * if somebody else is singlestepping across a probe point, flags
1000 * will have TF set, in which case, continue the remaining processing
1001 * of do_debug, as if this is not a probe hit.
1003 if (regs->flags & X86_EFLAGS_TF)
1008 NOKPROBE_SYMBOL(kprobe_debug_handler);
1010 int kprobe_fault_handler(struct pt_regs *regs, int trapnr)
1012 struct kprobe *cur = kprobe_running();
1013 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1015 if (unlikely(regs->ip == (unsigned long)cur->ainsn.insn)) {
1016 /* This must happen on single-stepping */
1017 WARN_ON(kcb->kprobe_status != KPROBE_HIT_SS &&
1018 kcb->kprobe_status != KPROBE_REENTER);
1020 * We are here because the instruction being single
1021 * stepped caused a page fault. We reset the current
1022 * kprobe and the ip points back to the probe address
1023 * and allow the page fault handler to continue as a
1024 * normal page fault.
1026 regs->ip = (unsigned long)cur->addr;
1028 * Trap flag (TF) has been set here because this fault
1029 * happened where the single stepping will be done.
1030 * So clear it by resetting the current kprobe:
1032 regs->flags &= ~X86_EFLAGS_TF;
1034 * Since the single step (trap) has been cancelled,
1035 * we need to restore BTF here.
1040 * If the TF flag was set before the kprobe hit,
1043 regs->flags |= kcb->kprobe_old_flags;
1045 if (kcb->kprobe_status == KPROBE_REENTER)
1046 restore_previous_kprobe(kcb);
1048 reset_current_kprobe();
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))
1071 NOKPROBE_SYMBOL(kprobe_fault_handler);
1073 int __init arch_populate_kprobe_blacklist(void)
1077 ret = kprobe_add_area_blacklist((unsigned long)__irqentry_text_start,
1078 (unsigned long)__irqentry_text_end);
1082 return kprobe_add_area_blacklist((unsigned long)__entry_text_start,
1083 (unsigned long)__entry_text_end);
1086 int __init arch_init_kprobes(void)
1091 int arch_trampoline_kprobe(struct kprobe *p)