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>
43 #include <linux/vmalloc.h>
45 #include <asm/text-patching.h>
46 #include <asm/cacheflush.h>
48 #include <asm/pgtable.h>
49 #include <linux/uaccess.h>
50 #include <asm/alternative.h>
52 #include <asm/debugreg.h>
53 #include <asm/set_memory.h>
57 DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
58 DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
60 #define stack_addr(regs) ((unsigned long *)regs->sp)
62 #define W(row, b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, ba, bb, bc, bd, be, bf)\
63 (((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) | \
64 (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) | \
65 (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) | \
66 (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf)) \
69 * Undefined/reserved opcodes, conditional jump, Opcode Extension
70 * Groups, and some special opcodes can not boost.
71 * This is non-const and volatile to keep gcc from statically
72 * optimizing it out, as variable_test_bit makes gcc think only
73 * *(unsigned long*) is used.
75 static volatile u32 twobyte_is_boostable[256 / 32] = {
76 /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */
77 /* ---------------------------------------------- */
78 W(0x00, 0, 0, 1, 1, 0, 0, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0) | /* 00 */
79 W(0x10, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1) , /* 10 */
80 W(0x20, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 20 */
81 W(0x30, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 30 */
82 W(0x40, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 40 */
83 W(0x50, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 50 */
84 W(0x60, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1) | /* 60 */
85 W(0x70, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1) , /* 70 */
86 W(0x80, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 80 */
87 W(0x90, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 90 */
88 W(0xa0, 1, 1, 0, 1, 1, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* a0 */
89 W(0xb0, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1) , /* b0 */
90 W(0xc0, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1) | /* c0 */
91 W(0xd0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) , /* d0 */
92 W(0xe0, 0, 1, 1, 0, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* e0 */
93 W(0xf0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 1, 0, 1, 1, 1, 0) /* f0 */
94 /* ----------------------------------------------- */
95 /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */
99 struct kretprobe_blackpoint kretprobe_blacklist[] = {
100 {"__switch_to", }, /* This function switches only current task, but
101 doesn't switch kernel stack.*/
102 {NULL, NULL} /* Terminator */
105 const int kretprobe_blacklist_size = ARRAY_SIZE(kretprobe_blacklist);
107 static nokprobe_inline void
108 __synthesize_relative_insn(void *dest, void *from, void *to, u8 op)
110 struct __arch_relative_insn {
115 insn = (struct __arch_relative_insn *)dest;
116 insn->raddr = (s32)((long)(to) - ((long)(from) + 5));
120 /* Insert a jump instruction at address 'from', which jumps to address 'to'.*/
121 void synthesize_reljump(void *dest, void *from, void *to)
123 __synthesize_relative_insn(dest, from, to, JMP32_INSN_OPCODE);
125 NOKPROBE_SYMBOL(synthesize_reljump);
127 /* Insert a call instruction at address 'from', which calls address 'to'.*/
128 void synthesize_relcall(void *dest, void *from, void *to)
130 __synthesize_relative_insn(dest, from, to, CALL_INSN_OPCODE);
132 NOKPROBE_SYMBOL(synthesize_relcall);
135 * Skip the prefixes of the instruction.
137 static kprobe_opcode_t *skip_prefixes(kprobe_opcode_t *insn)
141 attr = inat_get_opcode_attribute((insn_byte_t)*insn);
142 while (inat_is_legacy_prefix(attr)) {
144 attr = inat_get_opcode_attribute((insn_byte_t)*insn);
147 if (inat_is_rex_prefix(attr))
152 NOKPROBE_SYMBOL(skip_prefixes);
155 * Returns non-zero if INSN is boostable.
156 * RIP relative instructions are adjusted at copying time in 64 bits mode
158 int can_boost(struct insn *insn, void *addr)
160 kprobe_opcode_t opcode;
162 if (search_exception_tables((unsigned long)addr))
163 return 0; /* Page fault may occur on this address. */
165 /* 2nd-byte opcode */
166 if (insn->opcode.nbytes == 2)
167 return test_bit(insn->opcode.bytes[1],
168 (unsigned long *)twobyte_is_boostable);
170 if (insn->opcode.nbytes != 1)
173 /* Can't boost Address-size override prefix */
174 if (unlikely(inat_is_address_size_prefix(insn->attr)))
177 opcode = insn->opcode.bytes[0];
179 switch (opcode & 0xf0) {
181 /* can't boost "bound" */
182 return (opcode != 0x62);
184 return 0; /* can't boost conditional jump */
186 return opcode != 0x9a; /* can't boost call far */
188 /* can't boost software-interruptions */
189 return (0xc1 < opcode && opcode < 0xcc) || opcode == 0xcf;
191 /* can boost AA* and XLAT */
192 return (opcode == 0xd4 || opcode == 0xd5 || opcode == 0xd7);
194 /* can boost in/out and absolute jmps */
195 return ((opcode & 0x04) || opcode == 0xea);
197 /* clear and set flags are boostable */
198 return (opcode == 0xf5 || (0xf7 < opcode && opcode < 0xfe));
200 /* CS override prefix and call are not boostable */
201 return (opcode != 0x2e && opcode != 0x9a);
206 __recover_probed_insn(kprobe_opcode_t *buf, unsigned long addr)
211 kp = get_kprobe((void *)addr);
212 faddr = ftrace_location(addr);
214 * Addresses inside the ftrace location are refused by
215 * arch_check_ftrace_location(). Something went terribly wrong
216 * if such an address is checked here.
218 if (WARN_ON(faddr && faddr != addr))
221 * Use the current code if it is not modified by Kprobe
222 * and it cannot be modified by ftrace.
228 * Basically, kp->ainsn.insn has an original instruction.
229 * However, RIP-relative instruction can not do single-stepping
230 * at different place, __copy_instruction() tweaks the displacement of
231 * that instruction. In that case, we can't recover the instruction
232 * from the kp->ainsn.insn.
234 * On the other hand, in case on normal Kprobe, kp->opcode has a copy
235 * of the first byte of the probed instruction, which is overwritten
236 * by int3. And the instruction at kp->addr is not modified by kprobes
237 * except for the first byte, we can recover the original instruction
238 * from it and kp->opcode.
240 * In case of Kprobes using ftrace, we do not have a copy of
241 * the original instruction. In fact, the ftrace location might
242 * be modified at anytime and even could be in an inconsistent state.
243 * Fortunately, we know that the original code is the ideal 5-byte
246 if (probe_kernel_read(buf, (void *)addr,
247 MAX_INSN_SIZE * sizeof(kprobe_opcode_t)))
251 memcpy(buf, ideal_nops[NOP_ATOMIC5], 5);
254 return (unsigned long)buf;
258 * Recover the probed instruction at addr for further analysis.
259 * Caller must lock kprobes by kprobe_mutex, or disable preemption
260 * for preventing to release referencing kprobes.
261 * Returns zero if the instruction can not get recovered (or access failed).
263 unsigned long recover_probed_instruction(kprobe_opcode_t *buf, unsigned long addr)
265 unsigned long __addr;
267 __addr = __recover_optprobed_insn(buf, addr);
271 return __recover_probed_insn(buf, addr);
274 /* Check if paddr is at an instruction boundary */
275 static int can_probe(unsigned long paddr)
277 unsigned long addr, __addr, offset = 0;
279 kprobe_opcode_t buf[MAX_INSN_SIZE];
281 if (!kallsyms_lookup_size_offset(paddr, NULL, &offset))
284 /* Decode instructions */
285 addr = paddr - offset;
286 while (addr < paddr) {
288 * Check if the instruction has been modified by another
289 * kprobe, in which case we replace the breakpoint by the
290 * original instruction in our buffer.
291 * Also, jump optimization will change the breakpoint to
292 * relative-jump. Since the relative-jump itself is
293 * normally used, we just go through if there is no kprobe.
295 __addr = recover_probed_instruction(buf, addr);
298 kernel_insn_init(&insn, (void *)__addr, MAX_INSN_SIZE);
299 insn_get_length(&insn);
302 * Another debugging subsystem might insert this breakpoint.
303 * In that case, we can't recover it.
305 if (insn.opcode.bytes[0] == INT3_INSN_OPCODE)
310 return (addr == paddr);
314 * Returns non-zero if opcode modifies the interrupt flag.
316 static int is_IF_modifier(kprobe_opcode_t *insn)
319 insn = skip_prefixes(insn);
324 case 0xcf: /* iret/iretd */
325 case 0x9d: /* popf/popfd */
333 * Copy an instruction with recovering modified instruction by kprobes
334 * and adjust the displacement if the instruction uses the %rip-relative
335 * addressing mode. Note that since @real will be the final place of copied
336 * instruction, displacement must be adjust by @real, not @dest.
337 * This returns the length of copied instruction, or 0 if it has an error.
339 int __copy_instruction(u8 *dest, u8 *src, u8 *real, struct insn *insn)
341 kprobe_opcode_t buf[MAX_INSN_SIZE];
342 unsigned long recovered_insn =
343 recover_probed_instruction(buf, (unsigned long)src);
345 if (!recovered_insn || !insn)
348 /* This can access kernel text if given address is not recovered */
349 if (probe_kernel_read(dest, (void *)recovered_insn, MAX_INSN_SIZE))
352 kernel_insn_init(insn, dest, MAX_INSN_SIZE);
353 insn_get_length(insn);
355 /* We can not probe force emulate prefixed instruction */
356 if (insn_has_emulate_prefix(insn))
359 /* Another subsystem puts a breakpoint, failed to recover */
360 if (insn->opcode.bytes[0] == INT3_INSN_OPCODE)
363 /* We should not singlestep on the exception masking instructions */
364 if (insn_masking_exception(insn))
368 /* Only x86_64 has RIP relative instructions */
369 if (insn_rip_relative(insn)) {
373 * The copied instruction uses the %rip-relative addressing
374 * mode. Adjust the displacement for the difference between
375 * the original location of this instruction and the location
376 * of the copy that will actually be run. The tricky bit here
377 * is making sure that the sign extension happens correctly in
378 * this calculation, since we need a signed 32-bit result to
379 * be sign-extended to 64 bits when it's added to the %rip
380 * value and yield the same 64-bit result that the sign-
381 * extension of the original signed 32-bit displacement would
384 newdisp = (u8 *) src + (s64) insn->displacement.value
386 if ((s64) (s32) newdisp != newdisp) {
387 pr_err("Kprobes error: new displacement does not fit into s32 (%llx)\n", newdisp);
390 disp = (u8 *) dest + insn_offset_displacement(insn);
391 *(s32 *) disp = (s32) newdisp;
397 /* Prepare reljump right after instruction to boost */
398 static int prepare_boost(kprobe_opcode_t *buf, struct kprobe *p,
401 int len = insn->length;
403 if (can_boost(insn, p->addr) &&
404 MAX_INSN_SIZE - len >= JMP32_INSN_SIZE) {
406 * These instructions can be executed directly if it
407 * jumps back to correct address.
409 synthesize_reljump(buf + len, p->ainsn.insn + len,
410 p->addr + insn->length);
411 len += JMP32_INSN_SIZE;
412 p->ainsn.boostable = true;
414 p->ainsn.boostable = false;
420 /* Make page to RO mode when allocate it */
421 void *alloc_insn_page(void)
425 page = module_alloc(PAGE_SIZE);
429 set_vm_flush_reset_perms(page);
431 * First make the page read-only, and only then make it executable to
432 * prevent it from being W+X in between.
434 set_memory_ro((unsigned long)page, 1);
437 * TODO: Once additional kernel code protection mechanisms are set, ensure
438 * that the page was not maliciously altered and it is still zeroed.
440 set_memory_x((unsigned long)page, 1);
445 /* Recover page to RW mode before releasing it */
446 void free_insn_page(void *page)
448 module_memfree(page);
451 static int arch_copy_kprobe(struct kprobe *p)
454 kprobe_opcode_t buf[MAX_INSN_SIZE];
457 /* Copy an instruction with recovering if other optprobe modifies it.*/
458 len = __copy_instruction(buf, p->addr, p->ainsn.insn, &insn);
463 * __copy_instruction can modify the displacement of the instruction,
464 * but it doesn't affect boostable check.
466 len = prepare_boost(buf, p, &insn);
468 /* Check whether the instruction modifies Interrupt Flag or not */
469 p->ainsn.if_modifier = is_IF_modifier(buf);
471 /* Also, displacement change doesn't affect the first byte */
474 /* OK, write back the instruction(s) into ROX insn buffer */
475 text_poke(p->ainsn.insn, buf, len);
480 int arch_prepare_kprobe(struct kprobe *p)
484 if (alternatives_text_reserved(p->addr, p->addr))
487 if (!can_probe((unsigned long)p->addr))
489 /* insn: must be on special executable page on x86. */
490 p->ainsn.insn = get_insn_slot();
494 ret = arch_copy_kprobe(p);
496 free_insn_slot(p->ainsn.insn, 0);
497 p->ainsn.insn = NULL;
503 void arch_arm_kprobe(struct kprobe *p)
505 text_poke(p->addr, ((unsigned char []){INT3_INSN_OPCODE}), 1);
509 void arch_disarm_kprobe(struct kprobe *p)
511 text_poke(p->addr, &p->opcode, 1);
515 void arch_remove_kprobe(struct kprobe *p)
518 free_insn_slot(p->ainsn.insn, p->ainsn.boostable);
519 p->ainsn.insn = NULL;
523 static nokprobe_inline void
524 save_previous_kprobe(struct kprobe_ctlblk *kcb)
526 kcb->prev_kprobe.kp = kprobe_running();
527 kcb->prev_kprobe.status = kcb->kprobe_status;
528 kcb->prev_kprobe.old_flags = kcb->kprobe_old_flags;
529 kcb->prev_kprobe.saved_flags = kcb->kprobe_saved_flags;
532 static nokprobe_inline void
533 restore_previous_kprobe(struct kprobe_ctlblk *kcb)
535 __this_cpu_write(current_kprobe, kcb->prev_kprobe.kp);
536 kcb->kprobe_status = kcb->prev_kprobe.status;
537 kcb->kprobe_old_flags = kcb->prev_kprobe.old_flags;
538 kcb->kprobe_saved_flags = kcb->prev_kprobe.saved_flags;
541 static nokprobe_inline void
542 set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
543 struct kprobe_ctlblk *kcb)
545 __this_cpu_write(current_kprobe, p);
546 kcb->kprobe_saved_flags = kcb->kprobe_old_flags
547 = (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF));
548 if (p->ainsn.if_modifier)
549 kcb->kprobe_saved_flags &= ~X86_EFLAGS_IF;
552 static nokprobe_inline void clear_btf(void)
554 if (test_thread_flag(TIF_BLOCKSTEP)) {
555 unsigned long debugctl = get_debugctlmsr();
557 debugctl &= ~DEBUGCTLMSR_BTF;
558 update_debugctlmsr(debugctl);
562 static nokprobe_inline void restore_btf(void)
564 if (test_thread_flag(TIF_BLOCKSTEP)) {
565 unsigned long debugctl = get_debugctlmsr();
567 debugctl |= DEBUGCTLMSR_BTF;
568 update_debugctlmsr(debugctl);
572 void arch_prepare_kretprobe(struct kretprobe_instance *ri, struct pt_regs *regs)
574 unsigned long *sara = stack_addr(regs);
576 ri->ret_addr = (kprobe_opcode_t *) *sara;
579 /* Replace the return addr with trampoline addr */
580 *sara = (unsigned long) &kretprobe_trampoline;
582 NOKPROBE_SYMBOL(arch_prepare_kretprobe);
584 static void setup_singlestep(struct kprobe *p, struct pt_regs *regs,
585 struct kprobe_ctlblk *kcb, int reenter)
587 if (setup_detour_execution(p, regs, reenter))
590 #if !defined(CONFIG_PREEMPTION)
591 if (p->ainsn.boostable && !p->post_handler) {
592 /* Boost up -- we can execute copied instructions directly */
594 reset_current_kprobe();
596 * Reentering boosted probe doesn't reset current_kprobe,
597 * nor set current_kprobe, because it doesn't use single
600 regs->ip = (unsigned long)p->ainsn.insn;
605 save_previous_kprobe(kcb);
606 set_current_kprobe(p, regs, kcb);
607 kcb->kprobe_status = KPROBE_REENTER;
609 kcb->kprobe_status = KPROBE_HIT_SS;
610 /* Prepare real single stepping */
612 regs->flags |= X86_EFLAGS_TF;
613 regs->flags &= ~X86_EFLAGS_IF;
614 /* single step inline if the instruction is an int3 */
615 if (p->opcode == INT3_INSN_OPCODE)
616 regs->ip = (unsigned long)p->addr;
618 regs->ip = (unsigned long)p->ainsn.insn;
620 NOKPROBE_SYMBOL(setup_singlestep);
623 * We have reentered the kprobe_handler(), since another probe was hit while
624 * within the handler. We save the original kprobes variables and just single
625 * step on the instruction of the new probe without calling any user handlers.
627 static int reenter_kprobe(struct kprobe *p, struct pt_regs *regs,
628 struct kprobe_ctlblk *kcb)
630 switch (kcb->kprobe_status) {
631 case KPROBE_HIT_SSDONE:
632 case KPROBE_HIT_ACTIVE:
634 kprobes_inc_nmissed_count(p);
635 setup_singlestep(p, regs, kcb, 1);
638 /* A probe has been hit in the codepath leading up to, or just
639 * after, single-stepping of a probed instruction. This entire
640 * codepath should strictly reside in .kprobes.text section.
641 * Raise a BUG or we'll continue in an endless reentering loop
642 * and eventually a stack overflow.
644 pr_err("Unrecoverable kprobe detected.\n");
648 /* impossible cases */
655 NOKPROBE_SYMBOL(reenter_kprobe);
658 * Interrupts are disabled on entry as trap3 is an interrupt gate and they
659 * remain disabled throughout this function.
661 int kprobe_int3_handler(struct pt_regs *regs)
663 kprobe_opcode_t *addr;
665 struct kprobe_ctlblk *kcb;
670 addr = (kprobe_opcode_t *)(regs->ip - sizeof(kprobe_opcode_t));
672 * We don't want to be preempted for the entire duration of kprobe
673 * processing. Since int3 and debug trap disables irqs and we clear
674 * IF while singlestepping, it must be no preemptible.
677 kcb = get_kprobe_ctlblk();
678 p = get_kprobe(addr);
681 if (kprobe_running()) {
682 if (reenter_kprobe(p, regs, kcb))
685 set_current_kprobe(p, regs, kcb);
686 kcb->kprobe_status = KPROBE_HIT_ACTIVE;
689 * If we have no pre-handler or it returned 0, we
690 * continue with normal processing. If we have a
691 * pre-handler and it returned non-zero, that means
692 * user handler setup registers to exit to another
693 * instruction, we must skip the single stepping.
695 if (!p->pre_handler || !p->pre_handler(p, regs))
696 setup_singlestep(p, regs, kcb, 0);
698 reset_current_kprobe();
701 } else if (*addr != INT3_INSN_OPCODE) {
703 * The breakpoint instruction was removed right
704 * after we hit it. Another cpu has removed
705 * either a probepoint or a debugger breakpoint
706 * at this address. In either case, no further
707 * handling of this interrupt is appropriate.
708 * Back up over the (now missing) int3 and run
709 * the original instruction.
711 regs->ip = (unsigned long)addr;
713 } /* else: not a kprobe fault; let the kernel handle it */
717 NOKPROBE_SYMBOL(kprobe_int3_handler);
720 * When a retprobed function returns, this code saves registers and
721 * calls trampoline_handler() runs, which calls the kretprobe's handler.
725 ".global kretprobe_trampoline\n"
726 ".type kretprobe_trampoline, @function\n"
727 "kretprobe_trampoline:\n"
728 /* We don't bother saving the ss register */
734 " call trampoline_handler\n"
735 /* Replace saved sp with true return address. */
736 " movq %rax, 19*8(%rsp)\n"
744 " call trampoline_handler\n"
745 /* Replace saved sp with true return address. */
746 " movl %eax, 15*4(%esp)\n"
751 ".size kretprobe_trampoline, .-kretprobe_trampoline\n"
753 NOKPROBE_SYMBOL(kretprobe_trampoline);
754 STACK_FRAME_NON_STANDARD(kretprobe_trampoline);
756 static struct kprobe kretprobe_kprobe = {
757 .addr = (void *)kretprobe_trampoline,
761 * Called from kretprobe_trampoline
763 __used __visible void *trampoline_handler(struct pt_regs *regs)
765 struct kprobe_ctlblk *kcb;
766 struct kretprobe_instance *ri = NULL;
767 struct hlist_head *head, empty_rp;
768 struct hlist_node *tmp;
769 unsigned long flags, orig_ret_address = 0;
770 unsigned long trampoline_address = (unsigned long)&kretprobe_trampoline;
771 kprobe_opcode_t *correct_ret_addr = NULL;
773 bool skipped = false;
778 * Set a dummy kprobe for avoiding kretprobe recursion.
779 * Since kretprobe never run in kprobe handler, kprobe must not
780 * be running at this point.
782 kcb = get_kprobe_ctlblk();
783 __this_cpu_write(current_kprobe, &kretprobe_kprobe);
784 kcb->kprobe_status = KPROBE_HIT_ACTIVE;
786 INIT_HLIST_HEAD(&empty_rp);
787 kretprobe_hash_lock(current, &head, &flags);
788 /* fixup registers */
789 regs->cs = __KERNEL_CS;
791 regs->cs |= get_kernel_rpl();
794 /* We use pt_regs->sp for return address holder. */
795 frame_pointer = ®s->sp;
796 regs->ip = trampoline_address;
797 regs->orig_ax = ~0UL;
800 * It is possible to have multiple instances associated with a given
801 * task either because multiple functions in the call path have
802 * return probes installed on them, and/or more than one
803 * return probe was registered for a target function.
805 * We can handle this because:
806 * - instances are always pushed into the head of the list
807 * - when multiple return probes are registered for the same
808 * function, the (chronologically) first instance's ret_addr
809 * will be the real return address, and all the rest will
810 * point to kretprobe_trampoline.
812 hlist_for_each_entry(ri, head, hlist) {
813 if (ri->task != current)
814 /* another task is sharing our hash bucket */
817 * Return probes must be pushed on this hash list correct
818 * order (same as return order) so that it can be popped
819 * correctly. However, if we find it is pushed it incorrect
820 * order, this means we find a function which should not be
821 * probed, because the wrong order entry is pushed on the
822 * path of processing other kretprobe itself.
824 if (ri->fp != frame_pointer) {
826 pr_warn("kretprobe is stacked incorrectly. Trying to fixup.\n");
831 orig_ret_address = (unsigned long)ri->ret_addr;
833 pr_warn("%ps must be blacklisted because of incorrect kretprobe order\n",
836 if (orig_ret_address != trampoline_address)
838 * This is the real return address. Any other
839 * instances associated with this task are for
840 * other calls deeper on the call stack
845 kretprobe_assert(ri, orig_ret_address, trampoline_address);
847 correct_ret_addr = ri->ret_addr;
848 hlist_for_each_entry_safe(ri, tmp, head, hlist) {
849 if (ri->task != current)
850 /* another task is sharing our hash bucket */
852 if (ri->fp != frame_pointer)
855 orig_ret_address = (unsigned long)ri->ret_addr;
856 if (ri->rp && ri->rp->handler) {
857 __this_cpu_write(current_kprobe, &ri->rp->kp);
858 ri->ret_addr = correct_ret_addr;
859 ri->rp->handler(ri, regs);
860 __this_cpu_write(current_kprobe, &kretprobe_kprobe);
863 recycle_rp_inst(ri, &empty_rp);
865 if (orig_ret_address != trampoline_address)
867 * This is the real return address. Any other
868 * instances associated with this task are for
869 * other calls deeper on the call stack
874 kretprobe_hash_unlock(current, &flags);
876 __this_cpu_write(current_kprobe, NULL);
879 hlist_for_each_entry_safe(ri, tmp, &empty_rp, hlist) {
880 hlist_del(&ri->hlist);
883 return (void *)orig_ret_address;
885 NOKPROBE_SYMBOL(trampoline_handler);
888 * Called after single-stepping. p->addr is the address of the
889 * instruction whose first byte has been replaced by the "int 3"
890 * instruction. To avoid the SMP problems that can occur when we
891 * temporarily put back the original opcode to single-step, we
892 * single-stepped a copy of the instruction. The address of this
893 * copy is p->ainsn.insn.
895 * This function prepares to return from the post-single-step
896 * interrupt. We have to fix up the stack as follows:
898 * 0) Except in the case of absolute or indirect jump or call instructions,
899 * the new ip is relative to the copied instruction. We need to make
900 * it relative to the original instruction.
902 * 1) If the single-stepped instruction was pushfl, then the TF and IF
903 * flags are set in the just-pushed flags, and may need to be cleared.
905 * 2) If the single-stepped instruction was a call, the return address
906 * that is atop the stack is the address following the copied instruction.
907 * We need to make it the address following the original instruction.
909 * If this is the first time we've single-stepped the instruction at
910 * this probepoint, and the instruction is boostable, boost it: add a
911 * jump instruction after the copied instruction, that jumps to the next
912 * instruction after the probepoint.
914 static void resume_execution(struct kprobe *p, struct pt_regs *regs,
915 struct kprobe_ctlblk *kcb)
917 unsigned long *tos = stack_addr(regs);
918 unsigned long copy_ip = (unsigned long)p->ainsn.insn;
919 unsigned long orig_ip = (unsigned long)p->addr;
920 kprobe_opcode_t *insn = p->ainsn.insn;
923 insn = skip_prefixes(insn);
925 regs->flags &= ~X86_EFLAGS_TF;
927 case 0x9c: /* pushfl */
928 *tos &= ~(X86_EFLAGS_TF | X86_EFLAGS_IF);
929 *tos |= kcb->kprobe_old_flags;
931 case 0xc2: /* iret/ret/lret */
936 case 0xea: /* jmp absolute -- ip is correct */
937 /* ip is already adjusted, no more changes required */
938 p->ainsn.boostable = true;
940 case 0xe8: /* call relative - Fix return addr */
941 *tos = orig_ip + (*tos - copy_ip);
944 case 0x9a: /* call absolute -- same as call absolute, indirect */
945 *tos = orig_ip + (*tos - copy_ip);
949 if ((insn[1] & 0x30) == 0x10) {
951 * call absolute, indirect
952 * Fix return addr; ip is correct.
953 * But this is not boostable
955 *tos = orig_ip + (*tos - copy_ip);
957 } else if (((insn[1] & 0x31) == 0x20) ||
958 ((insn[1] & 0x31) == 0x21)) {
960 * jmp near and far, absolute indirect
961 * ip is correct. And this is boostable
963 p->ainsn.boostable = true;
970 regs->ip += orig_ip - copy_ip;
975 NOKPROBE_SYMBOL(resume_execution);
978 * Interrupts are disabled on entry as trap1 is an interrupt gate and they
979 * remain disabled throughout this function.
981 int kprobe_debug_handler(struct pt_regs *regs)
983 struct kprobe *cur = kprobe_running();
984 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
989 resume_execution(cur, regs, kcb);
990 regs->flags |= kcb->kprobe_saved_flags;
992 if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
993 kcb->kprobe_status = KPROBE_HIT_SSDONE;
994 cur->post_handler(cur, regs, 0);
997 /* Restore back the original saved kprobes variables and continue. */
998 if (kcb->kprobe_status == KPROBE_REENTER) {
999 restore_previous_kprobe(kcb);
1002 reset_current_kprobe();
1005 * if somebody else is singlestepping across a probe point, flags
1006 * will have TF set, in which case, continue the remaining processing
1007 * of do_debug, as if this is not a probe hit.
1009 if (regs->flags & X86_EFLAGS_TF)
1014 NOKPROBE_SYMBOL(kprobe_debug_handler);
1016 int kprobe_fault_handler(struct pt_regs *regs, int trapnr)
1018 struct kprobe *cur = kprobe_running();
1019 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1021 if (unlikely(regs->ip == (unsigned long)cur->ainsn.insn)) {
1022 /* This must happen on single-stepping */
1023 WARN_ON(kcb->kprobe_status != KPROBE_HIT_SS &&
1024 kcb->kprobe_status != KPROBE_REENTER);
1026 * We are here because the instruction being single
1027 * stepped caused a page fault. We reset the current
1028 * kprobe and the ip points back to the probe address
1029 * and allow the page fault handler to continue as a
1030 * normal page fault.
1032 regs->ip = (unsigned long)cur->addr;
1034 * Trap flag (TF) has been set here because this fault
1035 * happened where the single stepping will be done.
1036 * So clear it by resetting the current kprobe:
1038 regs->flags &= ~X86_EFLAGS_TF;
1041 * If the TF flag was set before the kprobe hit,
1044 regs->flags |= kcb->kprobe_old_flags;
1046 if (kcb->kprobe_status == KPROBE_REENTER)
1047 restore_previous_kprobe(kcb);
1049 reset_current_kprobe();
1050 } else if (kcb->kprobe_status == KPROBE_HIT_ACTIVE ||
1051 kcb->kprobe_status == KPROBE_HIT_SSDONE) {
1053 * We increment the nmissed count for accounting,
1054 * we can also use npre/npostfault count for accounting
1055 * these specific fault cases.
1057 kprobes_inc_nmissed_count(cur);
1060 * We come here because instructions in the pre/post
1061 * handler caused the page_fault, this could happen
1062 * if handler tries to access user space by
1063 * copy_from_user(), get_user() etc. Let the
1064 * user-specified handler try to fix it first.
1066 if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
1072 NOKPROBE_SYMBOL(kprobe_fault_handler);
1074 int __init arch_populate_kprobe_blacklist(void)
1078 ret = kprobe_add_area_blacklist((unsigned long)__irqentry_text_start,
1079 (unsigned long)__irqentry_text_end);
1083 return kprobe_add_area_blacklist((unsigned long)__entry_text_start,
1084 (unsigned long)__entry_text_end);
1087 int __init arch_init_kprobes(void)
1092 int arch_trampoline_kprobe(struct kprobe *p)