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/extable.h>
49 #include <linux/kdebug.h>
50 #include <linux/kallsyms.h>
51 #include <linux/ftrace.h>
52 #include <linux/frame.h>
53 #include <linux/kasan.h>
54 #include <linux/moduleloader.h>
56 #include <asm/text-patching.h>
57 #include <asm/cacheflush.h>
59 #include <asm/pgtable.h>
60 #include <asm/uaccess.h>
61 #include <asm/alternative.h>
63 #include <asm/debugreg.h>
64 #include <asm/sections.h>
68 void jprobe_return_end(void);
70 DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
71 DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
73 #define stack_addr(regs) ((unsigned long *)kernel_stack_pointer(regs))
75 #define W(row, b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, ba, bb, bc, bd, be, bf)\
76 (((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) | \
77 (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) | \
78 (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) | \
79 (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf)) \
82 * Undefined/reserved opcodes, conditional jump, Opcode Extension
83 * Groups, and some special opcodes can not boost.
84 * This is non-const and volatile to keep gcc from statically
85 * optimizing it out, as variable_test_bit makes gcc think only
86 * *(unsigned long*) is used.
88 static volatile u32 twobyte_is_boostable[256 / 32] = {
89 /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */
90 /* ---------------------------------------------- */
91 W(0x00, 0, 0, 1, 1, 0, 0, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0) | /* 00 */
92 W(0x10, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1) , /* 10 */
93 W(0x20, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 20 */
94 W(0x30, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 30 */
95 W(0x40, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 40 */
96 W(0x50, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 50 */
97 W(0x60, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1) | /* 60 */
98 W(0x70, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1) , /* 70 */
99 W(0x80, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 80 */
100 W(0x90, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 90 */
101 W(0xa0, 1, 1, 0, 1, 1, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* a0 */
102 W(0xb0, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1) , /* b0 */
103 W(0xc0, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1) | /* c0 */
104 W(0xd0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) , /* d0 */
105 W(0xe0, 0, 1, 1, 0, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* e0 */
106 W(0xf0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 1, 0, 1, 1, 1, 0) /* f0 */
107 /* ----------------------------------------------- */
108 /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */
112 struct kretprobe_blackpoint kretprobe_blacklist[] = {
113 {"__switch_to", }, /* This function switches only current task, but
114 doesn't switch kernel stack.*/
115 {NULL, NULL} /* Terminator */
118 const int kretprobe_blacklist_size = ARRAY_SIZE(kretprobe_blacklist);
120 static nokprobe_inline void
121 __synthesize_relative_insn(void *from, void *to, u8 op)
123 struct __arch_relative_insn {
128 insn = (struct __arch_relative_insn *)from;
129 insn->raddr = (s32)((long)(to) - ((long)(from) + 5));
133 /* Insert a jump instruction at address 'from', which jumps to address 'to'.*/
134 void synthesize_reljump(void *from, void *to)
136 __synthesize_relative_insn(from, to, RELATIVEJUMP_OPCODE);
138 NOKPROBE_SYMBOL(synthesize_reljump);
140 /* Insert a call instruction at address 'from', which calls address 'to'.*/
141 void synthesize_relcall(void *from, void *to)
143 __synthesize_relative_insn(from, to, RELATIVECALL_OPCODE);
145 NOKPROBE_SYMBOL(synthesize_relcall);
148 * Skip the prefixes of the instruction.
150 static kprobe_opcode_t *skip_prefixes(kprobe_opcode_t *insn)
154 attr = inat_get_opcode_attribute((insn_byte_t)*insn);
155 while (inat_is_legacy_prefix(attr)) {
157 attr = inat_get_opcode_attribute((insn_byte_t)*insn);
160 if (inat_is_rex_prefix(attr))
165 NOKPROBE_SYMBOL(skip_prefixes);
168 * Returns non-zero if opcode is boostable.
169 * RIP relative instructions are adjusted at copying time in 64 bits mode
171 int can_boost(kprobe_opcode_t *opcodes, void *addr)
173 kprobe_opcode_t opcode;
174 kprobe_opcode_t *orig_opcodes = opcodes;
176 if (search_exception_tables((unsigned long)addr))
177 return 0; /* Page fault may occur on this address. */
180 if (opcodes - orig_opcodes > MAX_INSN_SIZE - 1)
182 opcode = *(opcodes++);
184 /* 2nd-byte opcode */
185 if (opcode == 0x0f) {
186 if (opcodes - orig_opcodes > MAX_INSN_SIZE - 1)
188 return test_bit(*opcodes,
189 (unsigned long *)twobyte_is_boostable);
192 switch (opcode & 0xf0) {
195 goto retry; /* REX prefix is boostable */
198 if (0x63 < opcode && opcode < 0x67)
199 goto retry; /* prefixes */
200 /* can't boost Address-size override and bound */
201 return (opcode != 0x62 && opcode != 0x67);
203 return 0; /* can't boost conditional jump */
205 return opcode != 0x9a; /* can't boost call far */
207 /* can't boost software-interruptions */
208 return (0xc1 < opcode && opcode < 0xcc) || opcode == 0xcf;
210 /* can boost AA* and XLAT */
211 return (opcode == 0xd4 || opcode == 0xd5 || opcode == 0xd7);
213 /* can boost in/out and absolute jmps */
214 return ((opcode & 0x04) || opcode == 0xea);
216 if ((opcode & 0x0c) == 0 && opcode != 0xf1)
217 goto retry; /* lock/rep(ne) prefix */
218 /* clear and set flags are boostable */
219 return (opcode == 0xf5 || (0xf7 < opcode && opcode < 0xfe));
221 /* segment override prefixes are boostable */
222 if (opcode == 0x26 || opcode == 0x36 || opcode == 0x3e)
223 goto retry; /* prefixes */
224 /* CS override prefix and call are not boostable */
225 return (opcode != 0x2e && opcode != 0x9a);
230 __recover_probed_insn(kprobe_opcode_t *buf, unsigned long addr)
235 kp = get_kprobe((void *)addr);
236 faddr = ftrace_location(addr);
238 * Addresses inside the ftrace location are refused by
239 * arch_check_ftrace_location(). Something went terribly wrong
240 * if such an address is checked here.
242 if (WARN_ON(faddr && faddr != addr))
245 * Use the current code if it is not modified by Kprobe
246 * and it cannot be modified by ftrace.
252 * Basically, kp->ainsn.insn has an original instruction.
253 * However, RIP-relative instruction can not do single-stepping
254 * at different place, __copy_instruction() tweaks the displacement of
255 * that instruction. In that case, we can't recover the instruction
256 * from the kp->ainsn.insn.
258 * On the other hand, in case on normal Kprobe, kp->opcode has a copy
259 * of the first byte of the probed instruction, which is overwritten
260 * by int3. And the instruction at kp->addr is not modified by kprobes
261 * except for the first byte, we can recover the original instruction
262 * from it and kp->opcode.
264 * In case of Kprobes using ftrace, we do not have a copy of
265 * the original instruction. In fact, the ftrace location might
266 * be modified at anytime and even could be in an inconsistent state.
267 * Fortunately, we know that the original code is the ideal 5-byte
270 memcpy(buf, (void *)addr, MAX_INSN_SIZE * sizeof(kprobe_opcode_t));
272 memcpy(buf, ideal_nops[NOP_ATOMIC5], 5);
275 return (unsigned long)buf;
279 * Recover the probed instruction at addr for further analysis.
280 * Caller must lock kprobes by kprobe_mutex, or disable preemption
281 * for preventing to release referencing kprobes.
282 * Returns zero if the instruction can not get recovered.
284 unsigned long recover_probed_instruction(kprobe_opcode_t *buf, unsigned long addr)
286 unsigned long __addr;
288 __addr = __recover_optprobed_insn(buf, addr);
292 return __recover_probed_insn(buf, addr);
295 /* Check if paddr is at an instruction boundary */
296 static int can_probe(unsigned long paddr)
298 unsigned long addr, __addr, offset = 0;
300 kprobe_opcode_t buf[MAX_INSN_SIZE];
302 if (!kallsyms_lookup_size_offset(paddr, NULL, &offset))
305 /* Decode instructions */
306 addr = paddr - offset;
307 while (addr < paddr) {
309 * Check if the instruction has been modified by another
310 * kprobe, in which case we replace the breakpoint by the
311 * original instruction in our buffer.
312 * Also, jump optimization will change the breakpoint to
313 * relative-jump. Since the relative-jump itself is
314 * normally used, we just go through if there is no kprobe.
316 __addr = recover_probed_instruction(buf, addr);
319 kernel_insn_init(&insn, (void *)__addr, MAX_INSN_SIZE);
320 insn_get_length(&insn);
323 * Another debugging subsystem might insert this breakpoint.
324 * In that case, we can't recover it.
326 if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION)
331 return (addr == paddr);
335 * Returns non-zero if opcode modifies the interrupt flag.
337 static int is_IF_modifier(kprobe_opcode_t *insn)
340 insn = skip_prefixes(insn);
345 case 0xcf: /* iret/iretd */
346 case 0x9d: /* popf/popfd */
354 * Copy an instruction and adjust the displacement if the instruction
355 * uses the %rip-relative addressing mode.
356 * If it does, Return the address of the 32-bit displacement word.
357 * If not, return null.
358 * Only applicable to 64-bit x86.
360 int __copy_instruction(u8 *dest, u8 *src)
363 kprobe_opcode_t buf[MAX_INSN_SIZE];
365 unsigned long recovered_insn =
366 recover_probed_instruction(buf, (unsigned long)src);
370 kernel_insn_init(&insn, (void *)recovered_insn, MAX_INSN_SIZE);
371 insn_get_length(&insn);
372 length = insn.length;
374 /* Another subsystem puts a breakpoint, failed to recover */
375 if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION)
377 memcpy(dest, insn.kaddr, length);
379 /* We should not singlestep on the exception masking instructions */
380 if (insn_masking_exception(&insn))
384 if (insn_rip_relative(&insn)) {
387 kernel_insn_init(&insn, dest, length);
388 insn_get_displacement(&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 - (u8 *) dest;
402 if ((s64) (s32) newdisp != newdisp) {
403 pr_err("Kprobes error: new displacement does not fit into s32 (%llx)\n", newdisp);
406 disp = (u8 *) dest + insn_offset_displacement(&insn);
407 *(s32 *) disp = (s32) newdisp;
413 /* Recover page to RW mode before releasing it */
414 void free_insn_page(void *page)
416 set_memory_nx((unsigned long)page & PAGE_MASK, 1);
417 set_memory_rw((unsigned long)page & PAGE_MASK, 1);
418 module_memfree(page);
421 /* Prepare reljump right after instruction to boost */
422 static void prepare_boost(struct kprobe *p, int length)
424 if (can_boost(p->ainsn.insn, p->addr) &&
425 MAX_INSN_SIZE - length >= RELATIVEJUMP_SIZE) {
427 * These instructions can be executed directly if it
428 * jumps back to correct address.
430 synthesize_reljump(p->ainsn.insn + length, p->addr + length);
431 p->ainsn.boostable = 1;
433 p->ainsn.boostable = -1;
437 static int arch_copy_kprobe(struct kprobe *p)
441 set_memory_rw((unsigned long)p->ainsn.insn & PAGE_MASK, 1);
443 /* Copy an instruction with recovering if other optprobe modifies it.*/
444 len = __copy_instruction(p->ainsn.insn, p->addr);
449 * __copy_instruction can modify the displacement of the instruction,
450 * but it doesn't affect boostable check.
452 prepare_boost(p, len);
454 set_memory_ro((unsigned long)p->ainsn.insn & PAGE_MASK, 1);
456 /* Check whether the instruction modifies Interrupt Flag or not */
457 p->ainsn.if_modifier = is_IF_modifier(p->ainsn.insn);
459 /* Also, displacement change doesn't affect the first byte */
460 p->opcode = p->ainsn.insn[0];
465 int arch_prepare_kprobe(struct kprobe *p)
467 if (alternatives_text_reserved(p->addr, p->addr))
470 if (!can_probe((unsigned long)p->addr))
472 /* insn: must be on special executable page on x86. */
473 p->ainsn.insn = get_insn_slot();
477 return arch_copy_kprobe(p);
480 void arch_arm_kprobe(struct kprobe *p)
482 text_poke(p->addr, ((unsigned char []){BREAKPOINT_INSTRUCTION}), 1);
485 void arch_disarm_kprobe(struct kprobe *p)
487 text_poke(p->addr, &p->opcode, 1);
490 void arch_remove_kprobe(struct kprobe *p)
493 free_insn_slot(p->ainsn.insn, (p->ainsn.boostable == 1));
494 p->ainsn.insn = NULL;
498 static nokprobe_inline void
499 save_previous_kprobe(struct kprobe_ctlblk *kcb)
501 kcb->prev_kprobe.kp = kprobe_running();
502 kcb->prev_kprobe.status = kcb->kprobe_status;
503 kcb->prev_kprobe.old_flags = kcb->kprobe_old_flags;
504 kcb->prev_kprobe.saved_flags = kcb->kprobe_saved_flags;
507 static nokprobe_inline void
508 restore_previous_kprobe(struct kprobe_ctlblk *kcb)
510 __this_cpu_write(current_kprobe, kcb->prev_kprobe.kp);
511 kcb->kprobe_status = kcb->prev_kprobe.status;
512 kcb->kprobe_old_flags = kcb->prev_kprobe.old_flags;
513 kcb->kprobe_saved_flags = kcb->prev_kprobe.saved_flags;
516 static nokprobe_inline void
517 set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
518 struct kprobe_ctlblk *kcb)
520 __this_cpu_write(current_kprobe, p);
521 kcb->kprobe_saved_flags = kcb->kprobe_old_flags
522 = (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF));
523 if (p->ainsn.if_modifier)
524 kcb->kprobe_saved_flags &= ~X86_EFLAGS_IF;
527 static nokprobe_inline void clear_btf(void)
529 if (test_thread_flag(TIF_BLOCKSTEP)) {
530 unsigned long debugctl = get_debugctlmsr();
532 debugctl &= ~DEBUGCTLMSR_BTF;
533 update_debugctlmsr(debugctl);
537 static nokprobe_inline void restore_btf(void)
539 if (test_thread_flag(TIF_BLOCKSTEP)) {
540 unsigned long debugctl = get_debugctlmsr();
542 debugctl |= DEBUGCTLMSR_BTF;
543 update_debugctlmsr(debugctl);
547 void arch_prepare_kretprobe(struct kretprobe_instance *ri, struct pt_regs *regs)
549 unsigned long *sara = stack_addr(regs);
551 ri->ret_addr = (kprobe_opcode_t *) *sara;
554 /* Replace the return addr with trampoline addr */
555 *sara = (unsigned long) &kretprobe_trampoline;
557 NOKPROBE_SYMBOL(arch_prepare_kretprobe);
559 static void setup_singlestep(struct kprobe *p, struct pt_regs *regs,
560 struct kprobe_ctlblk *kcb, int reenter)
562 if (setup_detour_execution(p, regs, reenter))
565 #if !defined(CONFIG_PREEMPT)
566 if (p->ainsn.boostable == 1 && !p->post_handler) {
567 /* Boost up -- we can execute copied instructions directly */
569 reset_current_kprobe();
571 * Reentering boosted probe doesn't reset current_kprobe,
572 * nor set current_kprobe, because it doesn't use single
575 regs->ip = (unsigned long)p->ainsn.insn;
576 preempt_enable_no_resched();
581 save_previous_kprobe(kcb);
582 set_current_kprobe(p, regs, kcb);
583 kcb->kprobe_status = KPROBE_REENTER;
585 kcb->kprobe_status = KPROBE_HIT_SS;
586 /* Prepare real single stepping */
588 regs->flags |= X86_EFLAGS_TF;
589 regs->flags &= ~X86_EFLAGS_IF;
590 /* single step inline if the instruction is an int3 */
591 if (p->opcode == BREAKPOINT_INSTRUCTION)
592 regs->ip = (unsigned long)p->addr;
594 regs->ip = (unsigned long)p->ainsn.insn;
596 NOKPROBE_SYMBOL(setup_singlestep);
599 * We have reentered the kprobe_handler(), since another probe was hit while
600 * within the handler. We save the original kprobes variables and just single
601 * step on the instruction of the new probe without calling any user handlers.
603 static int reenter_kprobe(struct kprobe *p, struct pt_regs *regs,
604 struct kprobe_ctlblk *kcb)
606 switch (kcb->kprobe_status) {
607 case KPROBE_HIT_SSDONE:
608 case KPROBE_HIT_ACTIVE:
610 kprobes_inc_nmissed_count(p);
611 setup_singlestep(p, regs, kcb, 1);
614 /* A probe has been hit in the codepath leading up to, or just
615 * after, single-stepping of a probed instruction. This entire
616 * codepath should strictly reside in .kprobes.text section.
617 * Raise a BUG or we'll continue in an endless reentering loop
618 * and eventually a stack overflow.
620 pr_err("Unrecoverable kprobe detected.\n");
624 /* impossible cases */
631 NOKPROBE_SYMBOL(reenter_kprobe);
634 * Interrupts are disabled on entry as trap3 is an interrupt gate and they
635 * remain disabled throughout this function.
637 int kprobe_int3_handler(struct pt_regs *regs)
639 kprobe_opcode_t *addr;
641 struct kprobe_ctlblk *kcb;
646 addr = (kprobe_opcode_t *)(regs->ip - sizeof(kprobe_opcode_t));
648 * We don't want to be preempted for the entire
649 * duration of kprobe processing. We conditionally
650 * re-enable preemption at the end of this function,
651 * and also in reenter_kprobe() and setup_singlestep().
655 kcb = get_kprobe_ctlblk();
656 p = get_kprobe(addr);
659 if (kprobe_running()) {
660 if (reenter_kprobe(p, regs, kcb))
663 set_current_kprobe(p, regs, kcb);
664 kcb->kprobe_status = KPROBE_HIT_ACTIVE;
667 * If we have no pre-handler or it returned 0, we
668 * continue with normal processing. If we have a
669 * pre-handler and it returned non-zero, it prepped
670 * for calling the break_handler below on re-entry
671 * for jprobe processing, so get out doing nothing
674 if (!p->pre_handler || !p->pre_handler(p, regs))
675 setup_singlestep(p, regs, kcb, 0);
678 } else if (*addr != BREAKPOINT_INSTRUCTION) {
680 * The breakpoint instruction was removed right
681 * after we hit it. Another cpu has removed
682 * either a probepoint or a debugger breakpoint
683 * at this address. In either case, no further
684 * handling of this interrupt is appropriate.
685 * Back up over the (now missing) int3 and run
686 * the original instruction.
688 regs->ip = (unsigned long)addr;
689 preempt_enable_no_resched();
691 } else if (kprobe_running()) {
692 p = __this_cpu_read(current_kprobe);
693 if (p->break_handler && p->break_handler(p, regs)) {
694 if (!skip_singlestep(p, regs, kcb))
695 setup_singlestep(p, regs, kcb, 0);
698 } /* else: not a kprobe fault; let the kernel handle it */
700 preempt_enable_no_resched();
703 NOKPROBE_SYMBOL(kprobe_int3_handler);
706 * When a retprobed function returns, this code saves registers and
707 * calls trampoline_handler() runs, which calls the kretprobe's handler.
710 ".global kretprobe_trampoline\n"
711 ".type kretprobe_trampoline, @function\n"
712 "kretprobe_trampoline:\n"
714 /* We don't bother saving the ss register */
719 " call trampoline_handler\n"
720 /* Replace saved sp with true return address. */
721 " movq %rax, 152(%rsp)\n"
728 " call trampoline_handler\n"
729 /* Move flags to cs */
730 " movl 56(%esp), %edx\n"
731 " movl %edx, 52(%esp)\n"
732 /* Replace saved flags with true return address. */
733 " movl %eax, 56(%esp)\n"
738 ".size kretprobe_trampoline, .-kretprobe_trampoline\n"
740 NOKPROBE_SYMBOL(kretprobe_trampoline);
741 STACK_FRAME_NON_STANDARD(kretprobe_trampoline);
744 * Called from kretprobe_trampoline
746 __visible __used void *trampoline_handler(struct pt_regs *regs)
748 struct kretprobe_instance *ri = NULL;
749 struct hlist_head *head, empty_rp;
750 struct hlist_node *tmp;
751 unsigned long flags, orig_ret_address = 0;
752 unsigned long trampoline_address = (unsigned long)&kretprobe_trampoline;
753 kprobe_opcode_t *correct_ret_addr = NULL;
755 bool skipped = false;
758 * Set a dummy kprobe for avoiding kretprobe recursion.
759 * Since kretprobe never run in kprobe handler, kprobe must not
760 * be running at this point.
764 INIT_HLIST_HEAD(&empty_rp);
765 kretprobe_hash_lock(current, &head, &flags);
766 /* fixup registers */
768 regs->cs = __KERNEL_CS;
769 /* On x86-64, we use pt_regs->sp for return address holder. */
770 frame_pointer = ®s->sp;
772 regs->cs = __KERNEL_CS | get_kernel_rpl();
774 /* On x86-32, we use pt_regs->flags for return address holder. */
775 frame_pointer = ®s->flags;
777 regs->ip = trampoline_address;
778 regs->orig_ax = ~0UL;
781 * It is possible to have multiple instances associated with a given
782 * task either because multiple functions in the call path have
783 * return probes installed on them, and/or more than one
784 * return probe was registered for a target function.
786 * We can handle this because:
787 * - instances are always pushed into the head of the list
788 * - when multiple return probes are registered for the same
789 * function, the (chronologically) first instance's ret_addr
790 * will be the real return address, and all the rest will
791 * point to kretprobe_trampoline.
793 hlist_for_each_entry_safe(ri, tmp, head, hlist) {
794 if (ri->task != current)
795 /* another task is sharing our hash bucket */
798 * Return probes must be pushed on this hash list correct
799 * order (same as return order) so that it can be poped
800 * correctly. However, if we find it is pushed it incorrect
801 * order, this means we find a function which should not be
802 * probed, because the wrong order entry is pushed on the
803 * path of processing other kretprobe itself.
805 if (ri->fp != frame_pointer) {
807 pr_warn("kretprobe is stacked incorrectly. Trying to fixup.\n");
812 orig_ret_address = (unsigned long)ri->ret_addr;
814 pr_warn("%ps must be blacklisted because of incorrect kretprobe order\n",
817 if (orig_ret_address != trampoline_address)
819 * This is the real return address. Any other
820 * instances associated with this task are for
821 * other calls deeper on the call stack
826 kretprobe_assert(ri, orig_ret_address, trampoline_address);
828 correct_ret_addr = ri->ret_addr;
829 hlist_for_each_entry_safe(ri, tmp, head, hlist) {
830 if (ri->task != current)
831 /* another task is sharing our hash bucket */
833 if (ri->fp != frame_pointer)
836 orig_ret_address = (unsigned long)ri->ret_addr;
837 if (ri->rp && ri->rp->handler) {
838 __this_cpu_write(current_kprobe, &ri->rp->kp);
839 ri->ret_addr = correct_ret_addr;
840 ri->rp->handler(ri, regs);
841 __this_cpu_write(current_kprobe, &kprobe_busy);
844 recycle_rp_inst(ri, &empty_rp);
846 if (orig_ret_address != trampoline_address)
848 * This is the real return address. Any other
849 * instances associated with this task are for
850 * other calls deeper on the call stack
855 kretprobe_hash_unlock(current, &flags);
859 hlist_for_each_entry_safe(ri, tmp, &empty_rp, hlist) {
860 hlist_del(&ri->hlist);
863 return (void *)orig_ret_address;
865 NOKPROBE_SYMBOL(trampoline_handler);
868 * Called after single-stepping. p->addr is the address of the
869 * instruction whose first byte has been replaced by the "int 3"
870 * instruction. To avoid the SMP problems that can occur when we
871 * temporarily put back the original opcode to single-step, we
872 * single-stepped a copy of the instruction. The address of this
873 * copy is p->ainsn.insn.
875 * This function prepares to return from the post-single-step
876 * interrupt. We have to fix up the stack as follows:
878 * 0) Except in the case of absolute or indirect jump or call instructions,
879 * the new ip is relative to the copied instruction. We need to make
880 * it relative to the original instruction.
882 * 1) If the single-stepped instruction was pushfl, then the TF and IF
883 * flags are set in the just-pushed flags, and may need to be cleared.
885 * 2) If the single-stepped instruction was a call, the return address
886 * that is atop the stack is the address following the copied instruction.
887 * We need to make it the address following the original instruction.
889 * If this is the first time we've single-stepped the instruction at
890 * this probepoint, and the instruction is boostable, boost it: add a
891 * jump instruction after the copied instruction, that jumps to the next
892 * instruction after the probepoint.
894 static void resume_execution(struct kprobe *p, struct pt_regs *regs,
895 struct kprobe_ctlblk *kcb)
897 unsigned long *tos = stack_addr(regs);
898 unsigned long copy_ip = (unsigned long)p->ainsn.insn;
899 unsigned long orig_ip = (unsigned long)p->addr;
900 kprobe_opcode_t *insn = p->ainsn.insn;
903 insn = skip_prefixes(insn);
905 regs->flags &= ~X86_EFLAGS_TF;
907 case 0x9c: /* pushfl */
908 *tos &= ~(X86_EFLAGS_TF | X86_EFLAGS_IF);
909 *tos |= kcb->kprobe_old_flags;
911 case 0xc2: /* iret/ret/lret */
916 case 0xea: /* jmp absolute -- ip is correct */
917 /* ip is already adjusted, no more changes required */
918 p->ainsn.boostable = 1;
920 case 0xe8: /* call relative - Fix return addr */
921 *tos = orig_ip + (*tos - copy_ip);
924 case 0x9a: /* call absolute -- same as call absolute, indirect */
925 *tos = orig_ip + (*tos - copy_ip);
929 if ((insn[1] & 0x30) == 0x10) {
931 * call absolute, indirect
932 * Fix return addr; ip is correct.
933 * But this is not boostable
935 *tos = orig_ip + (*tos - copy_ip);
937 } else if (((insn[1] & 0x31) == 0x20) ||
938 ((insn[1] & 0x31) == 0x21)) {
940 * jmp near and far, absolute indirect
941 * ip is correct. And this is boostable
943 p->ainsn.boostable = 1;
950 regs->ip += orig_ip - copy_ip;
955 NOKPROBE_SYMBOL(resume_execution);
958 * Interrupts are disabled on entry as trap1 is an interrupt gate and they
959 * remain disabled throughout this function.
961 int kprobe_debug_handler(struct pt_regs *regs)
963 struct kprobe *cur = kprobe_running();
964 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
969 resume_execution(cur, regs, kcb);
970 regs->flags |= kcb->kprobe_saved_flags;
972 if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
973 kcb->kprobe_status = KPROBE_HIT_SSDONE;
974 cur->post_handler(cur, regs, 0);
977 /* Restore back the original saved kprobes variables and continue. */
978 if (kcb->kprobe_status == KPROBE_REENTER) {
979 restore_previous_kprobe(kcb);
982 reset_current_kprobe();
984 preempt_enable_no_resched();
987 * if somebody else is singlestepping across a probe point, flags
988 * will have TF set, in which case, continue the remaining processing
989 * of do_debug, as if this is not a probe hit.
991 if (regs->flags & X86_EFLAGS_TF)
996 NOKPROBE_SYMBOL(kprobe_debug_handler);
998 int kprobe_fault_handler(struct pt_regs *regs, int trapnr)
1000 struct kprobe *cur = kprobe_running();
1001 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1003 if (unlikely(regs->ip == (unsigned long)cur->ainsn.insn)) {
1004 /* This must happen on single-stepping */
1005 WARN_ON(kcb->kprobe_status != KPROBE_HIT_SS &&
1006 kcb->kprobe_status != KPROBE_REENTER);
1008 * We are here because the instruction being single
1009 * stepped caused a page fault. We reset the current
1010 * kprobe and the ip points back to the probe address
1011 * and allow the page fault handler to continue as a
1012 * normal page fault.
1014 regs->ip = (unsigned long)cur->addr;
1016 * Trap flag (TF) has been set here because this fault
1017 * happened where the single stepping will be done.
1018 * So clear it by resetting the current kprobe:
1020 regs->flags &= ~X86_EFLAGS_TF;
1022 * Since the single step (trap) has been cancelled,
1023 * we need to restore BTF here.
1028 * If the TF flag was set before the kprobe hit,
1031 regs->flags |= kcb->kprobe_old_flags;
1033 if (kcb->kprobe_status == KPROBE_REENTER)
1034 restore_previous_kprobe(kcb);
1036 reset_current_kprobe();
1037 preempt_enable_no_resched();
1038 } else if (kcb->kprobe_status == KPROBE_HIT_ACTIVE ||
1039 kcb->kprobe_status == KPROBE_HIT_SSDONE) {
1041 * We increment the nmissed count for accounting,
1042 * we can also use npre/npostfault count for accounting
1043 * these specific fault cases.
1045 kprobes_inc_nmissed_count(cur);
1048 * We come here because instructions in the pre/post
1049 * handler caused the page_fault, this could happen
1050 * if handler tries to access user space by
1051 * copy_from_user(), get_user() etc. Let the
1052 * user-specified handler try to fix it first.
1054 if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
1058 * In case the user-specified fault handler returned
1059 * zero, try to fix up.
1061 if (fixup_exception(regs, trapnr))
1065 * fixup routine could not handle it,
1066 * Let do_page_fault() fix it.
1072 NOKPROBE_SYMBOL(kprobe_fault_handler);
1075 * Wrapper routine for handling exceptions.
1077 int kprobe_exceptions_notify(struct notifier_block *self, unsigned long val,
1080 struct die_args *args = data;
1081 int ret = NOTIFY_DONE;
1083 if (args->regs && user_mode(args->regs))
1086 if (val == DIE_GPF) {
1088 * To be potentially processing a kprobe fault and to
1089 * trust the result from kprobe_running(), we have
1090 * be non-preemptible.
1092 if (!preemptible() && kprobe_running() &&
1093 kprobe_fault_handler(args->regs, args->trapnr))
1098 NOKPROBE_SYMBOL(kprobe_exceptions_notify);
1100 int setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
1102 struct jprobe *jp = container_of(p, struct jprobe, kp);
1104 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1106 kcb->jprobe_saved_regs = *regs;
1107 kcb->jprobe_saved_sp = stack_addr(regs);
1108 addr = (unsigned long)(kcb->jprobe_saved_sp);
1111 * As Linus pointed out, gcc assumes that the callee
1112 * owns the argument space and could overwrite it, e.g.
1113 * tailcall optimization. So, to be absolutely safe
1114 * we also save and restore enough stack bytes to cover
1115 * the argument area.
1116 * Use __memcpy() to avoid KASAN stack out-of-bounds reports as we copy
1117 * raw stack chunk with redzones:
1119 __memcpy(kcb->jprobes_stack, (kprobe_opcode_t *)addr, MIN_STACK_SIZE(addr));
1120 regs->flags &= ~X86_EFLAGS_IF;
1121 trace_hardirqs_off();
1122 regs->ip = (unsigned long)(jp->entry);
1125 * jprobes use jprobe_return() which skips the normal return
1126 * path of the function, and this messes up the accounting of the
1127 * function graph tracer to get messed up.
1129 * Pause function graph tracing while performing the jprobe function.
1131 pause_graph_tracing();
1134 NOKPROBE_SYMBOL(setjmp_pre_handler);
1136 void jprobe_return(void)
1138 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1140 /* Unpoison stack redzones in the frames we are going to jump over. */
1141 kasan_unpoison_stack_above_sp_to(kcb->jprobe_saved_sp);
1144 #ifdef CONFIG_X86_64
1145 " xchg %%rbx,%%rsp \n"
1147 " xchgl %%ebx,%%esp \n"
1150 " .globl jprobe_return_end\n"
1151 " jprobe_return_end: \n"
1153 (kcb->jprobe_saved_sp):"memory");
1155 NOKPROBE_SYMBOL(jprobe_return);
1156 NOKPROBE_SYMBOL(jprobe_return_end);
1158 int longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
1160 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1161 u8 *addr = (u8 *) (regs->ip - 1);
1162 struct jprobe *jp = container_of(p, struct jprobe, kp);
1163 void *saved_sp = kcb->jprobe_saved_sp;
1165 if ((addr > (u8 *) jprobe_return) &&
1166 (addr < (u8 *) jprobe_return_end)) {
1167 if (stack_addr(regs) != saved_sp) {
1168 struct pt_regs *saved_regs = &kcb->jprobe_saved_regs;
1170 "current sp %p does not match saved sp %p\n",
1171 stack_addr(regs), saved_sp);
1172 printk(KERN_ERR "Saved registers for jprobe %p\n", jp);
1173 show_regs(saved_regs);
1174 printk(KERN_ERR "Current registers\n");
1178 /* It's OK to start function graph tracing again */
1179 unpause_graph_tracing();
1180 *regs = kcb->jprobe_saved_regs;
1181 __memcpy(saved_sp, kcb->jprobes_stack, MIN_STACK_SIZE(saved_sp));
1182 preempt_enable_no_resched();
1187 NOKPROBE_SYMBOL(longjmp_break_handler);
1189 bool arch_within_kprobe_blacklist(unsigned long addr)
1191 return (addr >= (unsigned long)__kprobes_text_start &&
1192 addr < (unsigned long)__kprobes_text_end) ||
1193 (addr >= (unsigned long)__entry_text_start &&
1194 addr < (unsigned long)__entry_text_end);
1197 int __init arch_init_kprobes(void)
1202 int arch_trampoline_kprobe(struct kprobe *p)